diff --git a/.gitignore b/.gitignore
index e257658..1b958ab 100644
--- a/.gitignore
+++ b/.gitignore
@@ -1,6 +1,9 @@
 # ---> C++
 # Prerequisites
 *.d
+/.vscode
+/build
+build.sh
 
 # Compiled Object files
 *.slo
diff --git a/CMakeLists.txt b/CMakeLists.txt
index e69de29..d8ca905 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -0,0 +1,8 @@
+CMAKE_MINIMUM_REQUIRED(VERSION 3.0)
+project(FastDup)
+set(CMAKE_CXX_STANDARD 17)
+set(CMAKE_CXX_STANDARD_REQUIRED ON)
+# set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -pthread")
+# set(CMAKE_BUILD_TYPE Debug)
+# set(CMAKE_BUILD_TYPE Release)
+ADD_SUBDIRECTORY(src)
\ No newline at end of file
diff --git a/ext/argparse/argparse.hpp b/ext/argparse/argparse.hpp
new file mode 100644
index 0000000..cb4465b
--- /dev/null
+++ b/ext/argparse/argparse.hpp
@@ -0,0 +1,2555 @@
+/*
+  __ _ _ __ __ _ _ __   __ _ _ __ ___  ___
+ / _` | '__/ _` | '_ \ / _` | '__/ __|/ _ \ Argument Parser for Modern C++
+| (_| | | | (_| | |_) | (_| | |  \__ \  __/ http://github.com/p-ranav/argparse
+ \__,_|_|  \__, | .__/ \__,_|_|  |___/\___|
+           |___/|_|
+
+Licensed under the MIT License <http://opensource.org/licenses/MIT>.
+SPDX-License-Identifier: MIT
+Copyright (c) 2019-2022 Pranav Srinivas Kumar <pranav.srinivas.kumar@gmail.com>
+and other contributors.
+
+Permission is hereby  granted, free of charge, to any  person obtaining a copy
+of this software and associated  documentation files (the "Software"), to deal
+in the Software  without restriction, including without  limitation the rights
+to  use, copy,  modify, merge,  publish, distribute,  sublicense, and/or  sell
+copies  of  the Software,  and  to  permit persons  to  whom  the Software  is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE  IS PROVIDED "AS  IS", WITHOUT WARRANTY  OF ANY KIND,  EXPRESS OR
+IMPLIED,  INCLUDING BUT  NOT  LIMITED TO  THE  WARRANTIES OF  MERCHANTABILITY,
+FITNESS FOR  A PARTICULAR PURPOSE AND  NONINFRINGEMENT. IN NO EVENT  SHALL THE
+AUTHORS  OR COPYRIGHT  HOLDERS  BE  LIABLE FOR  ANY  CLAIM,  DAMAGES OR  OTHER
+LIABILITY, WHETHER IN AN ACTION OF  CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE  OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
+*/
+#pragma once
+
+#include <cerrno>
+
+#ifndef ARGPARSE_MODULE_USE_STD_MODULE
+#include <algorithm>
+#include <any>
+#include <array>
+#include <set>
+#include <charconv>
+#include <cstdlib>
+#include <functional>
+#include <iomanip>
+#include <iostream>
+#include <iterator>
+#include <limits>
+#include <list>
+#include <map>
+#include <numeric>
+#include <optional>
+#include <sstream>
+#include <stdexcept>
+#include <string>
+#include <string_view>
+#include <tuple>
+#include <type_traits>
+#include <utility>
+#include <variant>
+#include <vector>
+#endif
+
+#ifndef ARGPARSE_CUSTOM_STRTOF
+#define ARGPARSE_CUSTOM_STRTOF strtof
+#endif
+
+#ifndef ARGPARSE_CUSTOM_STRTOD
+#define ARGPARSE_CUSTOM_STRTOD strtod
+#endif
+
+#ifndef ARGPARSE_CUSTOM_STRTOLD
+#define ARGPARSE_CUSTOM_STRTOLD strtold
+#endif
+
+namespace argparse {
+
+namespace details { // namespace for helper methods
+
+template <typename T, typename = void>
+struct HasContainerTraits : std::false_type {};
+
+template <> struct HasContainerTraits<std::string> : std::false_type {};
+
+template <> struct HasContainerTraits<std::string_view> : std::false_type {};
+
+template <typename T>
+struct HasContainerTraits<
+    T, std::void_t<typename T::value_type, decltype(std::declval<T>().begin()),
+                   decltype(std::declval<T>().end()),
+                   decltype(std::declval<T>().size())>> : std::true_type {};
+
+template <typename T>
+inline constexpr bool IsContainer = HasContainerTraits<T>::value;
+
+template <typename T, typename = void>
+struct HasStreamableTraits : std::false_type {};
+
+template <typename T>
+struct HasStreamableTraits<
+    T,
+    std::void_t<decltype(std::declval<std::ostream &>() << std::declval<T>())>>
+    : std::true_type {};
+
+template <typename T>
+inline constexpr bool IsStreamable = HasStreamableTraits<T>::value;
+
+constexpr std::size_t repr_max_container_size = 5;
+
+template <typename T> std::string repr(T const &val) {
+  if constexpr (std::is_same_v<T, bool>) {
+    return val ? "true" : "false";
+  } else if constexpr (std::is_convertible_v<T, std::string_view>) {
+    return '"' + std::string{std::string_view{val}} + '"';
+  } else if constexpr (IsContainer<T>) {
+    std::stringstream out;
+    out << "{";
+    const auto size = val.size();
+    if (size > 1) {
+      out << repr(*val.begin());
+      std::for_each(
+          std::next(val.begin()),
+          std::next(
+              val.begin(),
+              static_cast<typename T::iterator::difference_type>(
+                  std::min<std::size_t>(size, repr_max_container_size) - 1)),
+          [&out](const auto &v) { out << " " << repr(v); });
+      if (size <= repr_max_container_size) {
+        out << " ";
+      } else {
+        out << "...";
+      }
+    }
+    if (size > 0) {
+      out << repr(*std::prev(val.end()));
+    }
+    out << "}";
+    return out.str();
+  } else if constexpr (IsStreamable<T>) {
+    std::stringstream out;
+    out << val;
+    return out.str();
+  } else {
+    return "<not representable>";
+  }
+}
+
+namespace {
+
+template <typename T> constexpr bool standard_signed_integer = false;
+template <> constexpr bool standard_signed_integer<signed char> = true;
+template <> constexpr bool standard_signed_integer<short int> = true;
+template <> constexpr bool standard_signed_integer<int> = true;
+template <> constexpr bool standard_signed_integer<long int> = true;
+template <> constexpr bool standard_signed_integer<long long int> = true;
+
+template <typename T> constexpr bool standard_unsigned_integer = false;
+template <> constexpr bool standard_unsigned_integer<unsigned char> = true;
+template <> constexpr bool standard_unsigned_integer<unsigned short int> = true;
+template <> constexpr bool standard_unsigned_integer<unsigned int> = true;
+template <> constexpr bool standard_unsigned_integer<unsigned long int> = true;
+template <>
+constexpr bool standard_unsigned_integer<unsigned long long int> = true;
+
+} // namespace
+
+constexpr int radix_2 = 2;
+constexpr int radix_8 = 8;
+constexpr int radix_10 = 10;
+constexpr int radix_16 = 16;
+
+template <typename T>
+constexpr bool standard_integer =
+    standard_signed_integer<T> || standard_unsigned_integer<T>;
+
+template <class F, class Tuple, class Extra, std::size_t... I>
+constexpr decltype(auto)
+apply_plus_one_impl(F &&f, Tuple &&t, Extra &&x,
+                    std::index_sequence<I...> /*unused*/) {
+  return std::invoke(std::forward<F>(f), std::get<I>(std::forward<Tuple>(t))...,
+                     std::forward<Extra>(x));
+}
+
+template <class F, class Tuple, class Extra>
+constexpr decltype(auto) apply_plus_one(F &&f, Tuple &&t, Extra &&x) {
+  return details::apply_plus_one_impl(
+      std::forward<F>(f), std::forward<Tuple>(t), std::forward<Extra>(x),
+      std::make_index_sequence<
+          std::tuple_size_v<std::remove_reference_t<Tuple>>>{});
+}
+
+constexpr auto pointer_range(std::string_view s) noexcept {
+  return std::tuple(s.data(), s.data() + s.size());
+}
+
+template <class CharT, class Traits>
+constexpr bool starts_with(std::basic_string_view<CharT, Traits> prefix,
+                           std::basic_string_view<CharT, Traits> s) noexcept {
+  return s.substr(0, prefix.size()) == prefix;
+}
+
+enum class chars_format {
+  scientific = 0xf1,
+  fixed = 0xf2,
+  hex = 0xf4,
+  binary = 0xf8,
+  general = fixed | scientific
+};
+
+struct ConsumeBinaryPrefixResult {
+  bool is_binary;
+  std::string_view rest;
+};
+
+constexpr auto consume_binary_prefix(std::string_view s)
+    -> ConsumeBinaryPrefixResult {
+  if (starts_with(std::string_view{"0b"}, s) ||
+      starts_with(std::string_view{"0B"}, s)) {
+    s.remove_prefix(2);
+    return {true, s};
+  }
+  return {false, s};
+}
+
+struct ConsumeHexPrefixResult {
+  bool is_hexadecimal;
+  std::string_view rest;
+};
+
+using namespace std::literals;
+
+constexpr auto consume_hex_prefix(std::string_view s)
+    -> ConsumeHexPrefixResult {
+  if (starts_with("0x"sv, s) || starts_with("0X"sv, s)) {
+    s.remove_prefix(2);
+    return {true, s};
+  }
+  return {false, s};
+}
+
+template <class T, auto Param>
+inline auto do_from_chars(std::string_view s) -> T {
+  T x{0};
+  auto [first, last] = pointer_range(s);
+  auto [ptr, ec] = std::from_chars(first, last, x, Param);
+  if (ec == std::errc()) {
+    if (ptr == last) {
+      return x;
+    }
+    throw std::invalid_argument{"pattern '" + std::string(s) +
+                                "' does not match to the end"};
+  }
+  if (ec == std::errc::invalid_argument) {
+    throw std::invalid_argument{"pattern '" + std::string(s) + "' not found"};
+  }
+  if (ec == std::errc::result_out_of_range) {
+    throw std::range_error{"'" + std::string(s) + "' not representable"};
+  }
+  return x; // unreachable
+}
+
+template <class T, auto Param = 0> struct parse_number {
+  auto operator()(std::string_view s) -> T {
+    return do_from_chars<T, Param>(s);
+  }
+};
+
+template <class T> struct parse_number<T, radix_2> {
+  auto operator()(std::string_view s) -> T {
+    if (auto [ok, rest] = consume_binary_prefix(s); ok) {
+      return do_from_chars<T, radix_2>(rest);
+    }
+    throw std::invalid_argument{"pattern not found"};
+  }
+};
+
+template <class T> struct parse_number<T, radix_16> {
+  auto operator()(std::string_view s) -> T {
+    if (starts_with("0x"sv, s) || starts_with("0X"sv, s)) {
+      if (auto [ok, rest] = consume_hex_prefix(s); ok) {
+        try {
+          return do_from_chars<T, radix_16>(rest);
+        } catch (const std::invalid_argument &err) {
+          throw std::invalid_argument("Failed to parse '" + std::string(s) +
+                                      "' as hexadecimal: " + err.what());
+        } catch (const std::range_error &err) {
+          throw std::range_error("Failed to parse '" + std::string(s) +
+                                 "' as hexadecimal: " + err.what());
+        }
+      }
+    } else {
+      // Allow passing hex numbers without prefix
+      // Shape 'x' already has to be specified
+      try {
+        return do_from_chars<T, radix_16>(s);
+      } catch (const std::invalid_argument &err) {
+        throw std::invalid_argument("Failed to parse '" + std::string(s) +
+                                    "' as hexadecimal: " + err.what());
+      } catch (const std::range_error &err) {
+        throw std::range_error("Failed to parse '" + std::string(s) +
+                               "' as hexadecimal: " + err.what());
+      }
+    }
+
+    throw std::invalid_argument{"pattern '" + std::string(s) +
+                                "' not identified as hexadecimal"};
+  }
+};
+
+template <class T> struct parse_number<T> {
+  auto operator()(std::string_view s) -> T {
+    auto [ok, rest] = consume_hex_prefix(s);
+    if (ok) {
+      try {
+        return do_from_chars<T, radix_16>(rest);
+      } catch (const std::invalid_argument &err) {
+        throw std::invalid_argument("Failed to parse '" + std::string(s) +
+                                    "' as hexadecimal: " + err.what());
+      } catch (const std::range_error &err) {
+        throw std::range_error("Failed to parse '" + std::string(s) +
+                               "' as hexadecimal: " + err.what());
+      }
+    }
+
+    auto [ok_binary, rest_binary] = consume_binary_prefix(s);
+    if (ok_binary) {
+      try {
+        return do_from_chars<T, radix_2>(rest_binary);
+      } catch (const std::invalid_argument &err) {
+        throw std::invalid_argument("Failed to parse '" + std::string(s) +
+                                    "' as binary: " + err.what());
+      } catch (const std::range_error &err) {
+        throw std::range_error("Failed to parse '" + std::string(s) +
+                               "' as binary: " + err.what());
+      }
+    }
+
+    if (starts_with("0"sv, s)) {
+      try {
+        return do_from_chars<T, radix_8>(rest);
+      } catch (const std::invalid_argument &err) {
+        throw std::invalid_argument("Failed to parse '" + std::string(s) +
+                                    "' as octal: " + err.what());
+      } catch (const std::range_error &err) {
+        throw std::range_error("Failed to parse '" + std::string(s) +
+                               "' as octal: " + err.what());
+      }
+    }
+
+    try {
+      return do_from_chars<T, radix_10>(rest);
+    } catch (const std::invalid_argument &err) {
+      throw std::invalid_argument("Failed to parse '" + std::string(s) +
+                                  "' as decimal integer: " + err.what());
+    } catch (const std::range_error &err) {
+      throw std::range_error("Failed to parse '" + std::string(s) +
+                             "' as decimal integer: " + err.what());
+    }
+  }
+};
+
+namespace {
+
+template <class T> inline const auto generic_strtod = nullptr;
+template <> inline const auto generic_strtod<float> = ARGPARSE_CUSTOM_STRTOF;
+template <> inline const auto generic_strtod<double> = ARGPARSE_CUSTOM_STRTOD;
+template <>
+inline const auto generic_strtod<long double> = ARGPARSE_CUSTOM_STRTOLD;
+
+} // namespace
+
+template <class T> inline auto do_strtod(std::string const &s) -> T {
+  if (isspace(static_cast<unsigned char>(s[0])) || s[0] == '+') {
+    throw std::invalid_argument{"pattern '" + s + "' not found"};
+  }
+
+  auto [first, last] = pointer_range(s);
+  char *ptr;
+
+  errno = 0;
+  auto x = generic_strtod<T>(first, &ptr);
+  if (errno == 0) {
+    if (ptr == last) {
+      return x;
+    }
+    throw std::invalid_argument{"pattern '" + s +
+                                "' does not match to the end"};
+  }
+  if (errno == ERANGE) {
+    throw std::range_error{"'" + s + "' not representable"};
+  }
+  return x; // unreachable
+}
+
+template <class T> struct parse_number<T, chars_format::general> {
+  auto operator()(std::string const &s) -> T {
+    if (auto r = consume_hex_prefix(s); r.is_hexadecimal) {
+      throw std::invalid_argument{
+          "chars_format::general does not parse hexfloat"};
+    }
+    if (auto r = consume_binary_prefix(s); r.is_binary) {
+      throw std::invalid_argument{
+          "chars_format::general does not parse binfloat"};
+    }
+
+    try {
+      return do_strtod<T>(s);
+    } catch (const std::invalid_argument &err) {
+      throw std::invalid_argument("Failed to parse '" + s +
+                                  "' as number: " + err.what());
+    } catch (const std::range_error &err) {
+      throw std::range_error("Failed to parse '" + s +
+                             "' as number: " + err.what());
+    }
+  }
+};
+
+template <class T> struct parse_number<T, chars_format::hex> {
+  auto operator()(std::string const &s) -> T {
+    if (auto r = consume_hex_prefix(s); !r.is_hexadecimal) {
+      throw std::invalid_argument{"chars_format::hex parses hexfloat"};
+    }
+    if (auto r = consume_binary_prefix(s); r.is_binary) {
+      throw std::invalid_argument{"chars_format::hex does not parse binfloat"};
+    }
+
+    try {
+      return do_strtod<T>(s);
+    } catch (const std::invalid_argument &err) {
+      throw std::invalid_argument("Failed to parse '" + s +
+                                  "' as hexadecimal: " + err.what());
+    } catch (const std::range_error &err) {
+      throw std::range_error("Failed to parse '" + s +
+                             "' as hexadecimal: " + err.what());
+    }
+  }
+};
+
+template <class T> struct parse_number<T, chars_format::binary> {
+  auto operator()(std::string const &s) -> T {
+    if (auto r = consume_hex_prefix(s); r.is_hexadecimal) {
+      throw std::invalid_argument{
+          "chars_format::binary does not parse hexfloat"};
+    }
+    if (auto r = consume_binary_prefix(s); !r.is_binary) {
+      throw std::invalid_argument{"chars_format::binary parses binfloat"};
+    }
+
+    return do_strtod<T>(s);
+  }
+};
+
+template <class T> struct parse_number<T, chars_format::scientific> {
+  auto operator()(std::string const &s) -> T {
+    if (auto r = consume_hex_prefix(s); r.is_hexadecimal) {
+      throw std::invalid_argument{
+          "chars_format::scientific does not parse hexfloat"};
+    }
+    if (auto r = consume_binary_prefix(s); r.is_binary) {
+      throw std::invalid_argument{
+          "chars_format::scientific does not parse binfloat"};
+    }
+    if (s.find_first_of("eE") == std::string::npos) {
+      throw std::invalid_argument{
+          "chars_format::scientific requires exponent part"};
+    }
+
+    try {
+      return do_strtod<T>(s);
+    } catch (const std::invalid_argument &err) {
+      throw std::invalid_argument("Failed to parse '" + s +
+                                  "' as scientific notation: " + err.what());
+    } catch (const std::range_error &err) {
+      throw std::range_error("Failed to parse '" + s +
+                             "' as scientific notation: " + err.what());
+    }
+  }
+};
+
+template <class T> struct parse_number<T, chars_format::fixed> {
+  auto operator()(std::string const &s) -> T {
+    if (auto r = consume_hex_prefix(s); r.is_hexadecimal) {
+      throw std::invalid_argument{
+          "chars_format::fixed does not parse hexfloat"};
+    }
+    if (auto r = consume_binary_prefix(s); r.is_binary) {
+      throw std::invalid_argument{
+          "chars_format::fixed does not parse binfloat"};
+    }
+    if (s.find_first_of("eE") != std::string::npos) {
+      throw std::invalid_argument{
+          "chars_format::fixed does not parse exponent part"};
+    }
+
+    try {
+      return do_strtod<T>(s);
+    } catch (const std::invalid_argument &err) {
+      throw std::invalid_argument("Failed to parse '" + s +
+                                  "' as fixed notation: " + err.what());
+    } catch (const std::range_error &err) {
+      throw std::range_error("Failed to parse '" + s +
+                             "' as fixed notation: " + err.what());
+    }
+  }
+};
+
+template <typename StrIt>
+std::string join(StrIt first, StrIt last, const std::string &separator) {
+  if (first == last) {
+    return "";
+  }
+  std::stringstream value;
+  value << *first;
+  ++first;
+  while (first != last) {
+    value << separator << *first;
+    ++first;
+  }
+  return value.str();
+}
+
+template <typename T> struct can_invoke_to_string {
+  template <typename U>
+  static auto test(int)
+      -> decltype(std::to_string(std::declval<U>()), std::true_type{});
+
+  template <typename U> static auto test(...) -> std::false_type;
+
+  static constexpr bool value = decltype(test<T>(0))::value;
+};
+
+template <typename T> struct IsChoiceTypeSupported {
+  using CleanType = typename std::decay<T>::type;
+  static const bool value = std::is_integral<CleanType>::value ||
+                            std::is_same<CleanType, std::string>::value ||
+                            std::is_same<CleanType, std::string_view>::value ||
+                            std::is_same<CleanType, const char *>::value;
+};
+
+template <typename StringType>
+std::size_t get_levenshtein_distance(const StringType &s1,
+                                     const StringType &s2) {
+  std::vector<std::vector<std::size_t>> dp(
+      s1.size() + 1, std::vector<std::size_t>(s2.size() + 1, 0));
+
+  for (std::size_t i = 0; i <= s1.size(); ++i) {
+    for (std::size_t j = 0; j <= s2.size(); ++j) {
+      if (i == 0) {
+        dp[i][j] = j;
+      } else if (j == 0) {
+        dp[i][j] = i;
+      } else if (s1[i - 1] == s2[j - 1]) {
+        dp[i][j] = dp[i - 1][j - 1];
+      } else {
+        dp[i][j] = 1 + std::min<std::size_t>({dp[i - 1][j], dp[i][j - 1], dp[i - 1][j - 1]});
+      }
+    }
+  }
+
+  return dp[s1.size()][s2.size()];
+}
+
+template <typename ValueType>
+std::string get_most_similar_string(const std::map<std::string, ValueType> &map,
+                                    const std::string &input) {
+  std::string most_similar{};
+  std::size_t min_distance = (std::numeric_limits<std::size_t>::max)();
+
+  for (const auto &entry : map) {
+    std::size_t distance = get_levenshtein_distance(entry.first, input);
+    if (distance < min_distance) {
+      min_distance = distance;
+      most_similar = entry.first;
+    }
+  }
+
+  return most_similar;
+}
+
+} // namespace details
+
+enum class nargs_pattern { optional, any, at_least_one };
+
+enum class default_arguments : unsigned int {
+  none = 0,
+  help = 1,
+  version = 2,
+  all = help | version,
+};
+
+inline default_arguments operator&(const default_arguments &a,
+                                   const default_arguments &b) {
+  return static_cast<default_arguments>(
+      static_cast<std::underlying_type<default_arguments>::type>(a) &
+      static_cast<std::underlying_type<default_arguments>::type>(b));
+}
+
+class ArgumentParser;
+
+class Argument {
+  friend class ArgumentParser;
+  friend auto operator<<(std::ostream &stream, const ArgumentParser &parser)
+      -> std::ostream &;
+
+  template <std::size_t N, std::size_t... I>
+  explicit Argument(std::string_view prefix_chars,
+                    std::array<std::string_view, N> &&a,
+                    std::index_sequence<I...> /*unused*/)
+      : m_accepts_optional_like_value(false),
+        m_is_optional((is_optional(a[I], prefix_chars) || ...)),
+        m_is_required(false), m_is_repeatable(false), m_is_used(false),
+        m_is_hidden(false), m_prefix_chars(prefix_chars) {
+    ((void)m_names.emplace_back(a[I]), ...);
+    std::sort(
+        m_names.begin(), m_names.end(), [](const auto &lhs, const auto &rhs) {
+          return lhs.size() == rhs.size() ? lhs < rhs : lhs.size() < rhs.size();
+        });
+  }
+
+public:
+  template <std::size_t N>
+  explicit Argument(std::string_view prefix_chars,
+                    std::array<std::string_view, N> &&a)
+      : Argument(prefix_chars, std::move(a), std::make_index_sequence<N>{}) {}
+
+  Argument &help(std::string help_text) {
+    m_help = std::move(help_text);
+    return *this;
+  }
+
+  Argument &metavar(std::string metavar) {
+    m_metavar = std::move(metavar);
+    return *this;
+  }
+
+  template <typename T> Argument &default_value(T &&value) {
+    m_num_args_range = NArgsRange{0, m_num_args_range.get_max()};
+    m_default_value_repr = details::repr(value);
+
+    if constexpr (std::is_convertible_v<T, std::string_view>) {
+      m_default_value_str = std::string{std::string_view{value}};
+    } else if constexpr (details::can_invoke_to_string<T>::value) {
+      m_default_value_str = std::to_string(value);
+    }
+
+    m_default_value = std::forward<T>(value);
+    return *this;
+  }
+
+  Argument &default_value(const char *value) {
+    return default_value(std::string(value));
+  }
+
+  Argument &required() {
+    m_is_required = true;
+    return *this;
+  }
+
+  Argument &implicit_value(std::any value) {
+    m_implicit_value = std::move(value);
+    m_num_args_range = NArgsRange{0, 0};
+    return *this;
+  }
+
+  // This is shorthand for:
+  //   program.add_argument("foo")
+  //     .default_value(false)
+  //     .implicit_value(true)
+  Argument &flag() {
+    default_value(false);
+    implicit_value(true);
+    return *this;
+  }
+
+  template <class F, class... Args>
+  auto action(F &&callable, Args &&... bound_args)
+      -> std::enable_if_t<std::is_invocable_v<F, Args..., std::string const>,
+                          Argument &> {
+    using action_type = std::conditional_t<
+        std::is_void_v<std::invoke_result_t<F, Args..., std::string const>>,
+        void_action, valued_action>;
+    if constexpr (sizeof...(Args) == 0) {
+      m_action.emplace<action_type>(std::forward<F>(callable));
+    } else {
+      m_action.emplace<action_type>(
+          [f = std::forward<F>(callable),
+           tup = std::make_tuple(std::forward<Args>(bound_args)...)](
+              std::string const &opt) mutable {
+            return details::apply_plus_one(f, tup, opt);
+          });
+    }
+    return *this;
+  }
+
+  auto &store_into(bool &var) {
+    if ((!m_default_value.has_value()) && (!m_implicit_value.has_value())) {
+      flag();
+    }
+    if (m_default_value.has_value()) {
+      var = std::any_cast<bool>(m_default_value);
+    }
+    action([&var](const auto & /*unused*/) { var = true; });
+    return *this;
+  }
+
+  template <typename T, typename std::enable_if<std::is_integral<T>::value>::type * = nullptr>
+  auto &store_into(T &var) {
+    if (m_default_value.has_value()) {
+      var = std::any_cast<T>(m_default_value);
+    }
+    action([&var](const auto &s) {
+      var = details::parse_number<T, details::radix_10>()(s);
+    });
+    return *this;
+  }
+
+  auto &store_into(double &var) {
+    if (m_default_value.has_value()) {
+      var = std::any_cast<double>(m_default_value);
+    }
+    action([&var](const auto &s) {
+      var = details::parse_number<double, details::chars_format::general>()(s);
+    });
+    return *this;
+  }
+
+  auto &store_into(std::string &var) {
+    if (m_default_value.has_value()) {
+      var = std::any_cast<std::string>(m_default_value);
+    }
+    action([&var](const std::string &s) { var = s; });
+    return *this;
+  }
+
+  auto &store_into(std::vector<std::string> &var) {
+    if (m_default_value.has_value()) {
+      var = std::any_cast<std::vector<std::string>>(m_default_value);
+    }
+    action([this, &var](const std::string &s) {
+      if (!m_is_used) {
+        var.clear();
+      }
+      m_is_used = true;
+      var.push_back(s);
+    });
+    return *this;
+  }
+
+  auto &store_into(std::vector<int> &var) {
+    if (m_default_value.has_value()) {
+      var = std::any_cast<std::vector<int>>(m_default_value);
+    }
+    action([this, &var](const std::string &s) {
+      if (!m_is_used) {
+        var.clear();
+      }
+      m_is_used = true;
+      var.push_back(details::parse_number<int, details::radix_10>()(s));
+    });
+    return *this;
+  }
+
+  auto &store_into(std::set<std::string> &var) {
+    if (m_default_value.has_value()) {
+      var = std::any_cast<std::set<std::string>>(m_default_value);
+    }
+    action([this, &var](const std::string &s) {
+      if (!m_is_used) {
+        var.clear();
+      }
+      m_is_used = true;
+      var.insert(s);
+    });
+    return *this;
+  }
+
+  auto &store_into(std::set<int> &var) {
+    if (m_default_value.has_value()) {
+      var = std::any_cast<std::set<int>>(m_default_value);
+    }
+    action([this, &var](const std::string &s) {
+      if (!m_is_used) {
+        var.clear();
+      }
+      m_is_used = true;
+      var.insert(details::parse_number<int, details::radix_10>()(s));
+    });
+    return *this;
+  }
+
+  auto &append() {
+    m_is_repeatable = true;
+    return *this;
+  }
+
+  // Cause the argument to be invisible in usage and help
+  auto &hidden() {
+    m_is_hidden = true;
+    return *this;
+  }
+
+  template <char Shape, typename T>
+  auto scan() -> std::enable_if_t<std::is_arithmetic_v<T>, Argument &> {
+    static_assert(!(std::is_const_v<T> || std::is_volatile_v<T>),
+                  "T should not be cv-qualified");
+    auto is_one_of = [](char c, auto... x) constexpr {
+      return ((c == x) || ...);
+    };
+
+    if constexpr (is_one_of(Shape, 'd') && details::standard_integer<T>) {
+      action(details::parse_number<T, details::radix_10>());
+    } else if constexpr (is_one_of(Shape, 'i') &&
+                         details::standard_integer<T>) {
+      action(details::parse_number<T>());
+    } else if constexpr (is_one_of(Shape, 'u') &&
+                         details::standard_unsigned_integer<T>) {
+      action(details::parse_number<T, details::radix_10>());
+    } else if constexpr (is_one_of(Shape, 'b') &&
+                         details::standard_unsigned_integer<T>) {
+      action(details::parse_number<T, details::radix_2>());
+    } else if constexpr (is_one_of(Shape, 'o') &&
+                         details::standard_unsigned_integer<T>) {
+      action(details::parse_number<T, details::radix_8>());
+    } else if constexpr (is_one_of(Shape, 'x', 'X') &&
+                         details::standard_unsigned_integer<T>) {
+      action(details::parse_number<T, details::radix_16>());
+    } else if constexpr (is_one_of(Shape, 'a', 'A') &&
+                         std::is_floating_point_v<T>) {
+      action(details::parse_number<T, details::chars_format::hex>());
+    } else if constexpr (is_one_of(Shape, 'e', 'E') &&
+                         std::is_floating_point_v<T>) {
+      action(details::parse_number<T, details::chars_format::scientific>());
+    } else if constexpr (is_one_of(Shape, 'f', 'F') &&
+                         std::is_floating_point_v<T>) {
+      action(details::parse_number<T, details::chars_format::fixed>());
+    } else if constexpr (is_one_of(Shape, 'g', 'G') &&
+                         std::is_floating_point_v<T>) {
+      action(details::parse_number<T, details::chars_format::general>());
+    } else {
+      static_assert(alignof(T) == 0, "No scan specification for T");
+    }
+
+    return *this;
+  }
+
+  Argument &nargs(std::size_t num_args) {
+    m_num_args_range = NArgsRange{num_args, num_args};
+    return *this;
+  }
+
+  Argument &nargs(std::size_t num_args_min, std::size_t num_args_max) {
+    m_num_args_range = NArgsRange{num_args_min, num_args_max};
+    return *this;
+  }
+
+  Argument &nargs(nargs_pattern pattern) {
+    switch (pattern) {
+    case nargs_pattern::optional:
+      m_num_args_range = NArgsRange{0, 1};
+      break;
+    case nargs_pattern::any:
+      m_num_args_range =
+          NArgsRange{0, (std::numeric_limits<std::size_t>::max)()};
+      break;
+    case nargs_pattern::at_least_one:
+      m_num_args_range =
+          NArgsRange{1, (std::numeric_limits<std::size_t>::max)()};
+      break;
+    }
+    return *this;
+  }
+
+  Argument &remaining() {
+    m_accepts_optional_like_value = true;
+    return nargs(nargs_pattern::any);
+  }
+
+  template <typename T> void add_choice(T &&choice) {
+    static_assert(details::IsChoiceTypeSupported<T>::value,
+                  "Only string or integer type supported for choice");
+    static_assert(std::is_convertible_v<T, std::string_view> ||
+                      details::can_invoke_to_string<T>::value,
+                  "Choice is not convertible to string_type");
+    if (!m_choices.has_value()) {
+      m_choices = std::vector<std::string>{};
+    }
+
+    if constexpr (std::is_convertible_v<T, std::string_view>) {
+      m_choices.value().push_back(
+          std::string{std::string_view{std::forward<T>(choice)}});
+    } else if constexpr (details::can_invoke_to_string<T>::value) {
+      m_choices.value().push_back(std::to_string(std::forward<T>(choice)));
+    }
+  }
+
+  Argument &choices() {
+    if (!m_choices.has_value()) {
+      throw std::runtime_error("Zero choices provided");
+    }
+    return *this;
+  }
+
+  template <typename T, typename... U>
+  Argument &choices(T &&first, U &&... rest) {
+    add_choice(std::forward<T>(first));
+    choices(std::forward<U>(rest)...);
+    return *this;
+  }
+
+  void find_default_value_in_choices_or_throw() const {
+
+    const auto &choices = m_choices.value();
+
+    if (m_default_value.has_value()) {
+      if (std::find(choices.begin(), choices.end(), m_default_value_str) ==
+          choices.end()) {
+        // provided arg not in list of allowed choices
+        // report error
+
+        std::string choices_as_csv =
+            std::accumulate(choices.begin(), choices.end(), std::string(),
+                            [](const std::string &a, const std::string &b) {
+                              return a + (a.empty() ? "" : ", ") + b;
+                            });
+
+        throw std::runtime_error(
+            std::string{"Invalid default value "} + m_default_value_repr +
+            " - allowed options: {" + choices_as_csv + "}");
+      }
+    }
+  }
+
+  template <typename Iterator>
+  bool is_value_in_choices(Iterator option_it) const {
+
+    const auto &choices = m_choices.value();
+
+    return (std::find(choices.begin(), choices.end(), *option_it) !=
+            choices.end());
+  }
+
+  template <typename Iterator>
+  void throw_invalid_arguments_error(Iterator option_it) const {
+    const auto &choices = m_choices.value();
+    const std::string choices_as_csv = std::accumulate(
+        choices.begin(), choices.end(), std::string(),
+        [](const std::string &option_a, const std::string &option_b) {
+          return option_a + (option_a.empty() ? "" : ", ") + option_b;
+        });
+
+    throw std::runtime_error(std::string{"Invalid argument "} +
+                             details::repr(*option_it) +
+                             " - allowed options: {" + choices_as_csv + "}");
+  }
+
+  /* The dry_run parameter can be set to true to avoid running the actions,
+   * and setting m_is_used. This may be used by a pre-processing step to do
+   * a first iteration over arguments.
+   */
+  template <typename Iterator>
+  Iterator consume(Iterator start, Iterator end,
+                   std::string_view used_name = {}, bool dry_run = false) {
+    if (!m_is_repeatable && m_is_used) {
+      throw std::runtime_error(
+          std::string("Duplicate argument ").append(used_name));
+    }
+    m_used_name = used_name;
+
+    std::size_t passed_options = 0;
+
+    if (m_choices.has_value()) {
+      // Check each value in (start, end) and make sure
+      // it is in the list of allowed choices/options
+      const auto max_number_of_args = m_num_args_range.get_max();
+      const auto min_number_of_args = m_num_args_range.get_min();
+      for (auto it = start; it != end; ++it) {
+        if (is_value_in_choices(it)) {
+          passed_options += 1;
+          continue;
+        }
+
+        if ((passed_options >= min_number_of_args) &&
+            (passed_options <= max_number_of_args)) {
+          break;
+        }
+
+        throw_invalid_arguments_error(it);
+      }
+    }
+
+    const auto num_args_max =
+        (m_choices.has_value()) ? passed_options : m_num_args_range.get_max();
+    const auto num_args_min = m_num_args_range.get_min();
+    std::size_t dist = 0;
+    if (num_args_max == 0) {
+      if (!dry_run) {
+        m_values.emplace_back(m_implicit_value);
+        std::visit([](const auto &f) { f({}); }, m_action);
+        m_is_used = true;
+      }
+      return start;
+    }
+    if ((dist = static_cast<std::size_t>(std::distance(start, end))) >=
+        num_args_min) {
+      if (num_args_max < dist) {
+        end = std::next(start, static_cast<typename Iterator::difference_type>(
+                                   num_args_max));
+      }
+      if (!m_accepts_optional_like_value) {
+        end = std::find_if(
+            start, end,
+            std::bind(is_optional, std::placeholders::_1, m_prefix_chars));
+        dist = static_cast<std::size_t>(std::distance(start, end));
+        if (dist < num_args_min) {
+          throw std::runtime_error("Too few arguments for '" +
+                                   std::string(m_used_name) + "'.");
+        }
+      }
+      struct ActionApply {
+        void operator()(valued_action &f) {
+          std::transform(first, last, std::back_inserter(self.m_values), f);
+        }
+
+        void operator()(void_action &f) {
+          std::for_each(first, last, f);
+          if (!self.m_default_value.has_value()) {
+            if (!self.m_accepts_optional_like_value) {
+              self.m_values.resize(
+                  static_cast<std::size_t>(std::distance(first, last)));
+            }
+          }
+        }
+
+        Iterator first, last;
+        Argument &self;
+      };
+      if (!dry_run) {
+        std::visit(ActionApply{start, end, *this}, m_action);
+        m_is_used = true;
+      }
+      return end;
+    }
+    if (m_default_value.has_value()) {
+      if (!dry_run) {
+        m_is_used = true;
+      }
+      return start;
+    }
+    throw std::runtime_error("Too few arguments for '" +
+                             std::string(m_used_name) + "'.");
+  }
+
+  /*
+   * @throws std::runtime_error if argument values are not valid
+   */
+  void validate() const {
+    if (m_is_optional) {
+      // TODO: check if an implicit value was programmed for this argument
+      if (!m_is_used && !m_default_value.has_value() && m_is_required) {
+        throw_required_arg_not_used_error();
+      }
+      if (m_is_used && m_is_required && m_values.empty()) {
+        throw_required_arg_no_value_provided_error();
+      }
+    } else {
+      if (!m_num_args_range.contains(m_values.size()) &&
+          !m_default_value.has_value()) {
+        throw_nargs_range_validation_error();
+      }
+    }
+
+    if (m_choices.has_value()) {
+      // Make sure the default value (if provided)
+      // is in the list of choices
+      find_default_value_in_choices_or_throw();
+    }
+  }
+
+  std::string get_names_csv(char separator = ',') const {
+    return std::accumulate(
+        m_names.begin(), m_names.end(), std::string{""},
+        [&](const std::string &result, const std::string &name) {
+          return result.empty() ? name : result + separator + name;
+        });
+  }
+
+  std::string get_usage_full() const {
+    std::stringstream usage;
+
+    usage << get_names_csv('/');
+    const std::string metavar = !m_metavar.empty() ? m_metavar : "VAR";
+    if (m_num_args_range.get_max() > 0) {
+      usage << " " << metavar;
+      if (m_num_args_range.get_max() > 1) {
+        usage << "...";
+      }
+    }
+    return usage.str();
+  }
+
+  std::string get_inline_usage() const {
+    std::stringstream usage;
+    // Find the longest variant to show in the usage string
+    std::string longest_name = m_names.front();
+    for (const auto &s : m_names) {
+      if (s.size() > longest_name.size()) {
+        longest_name = s;
+      }
+    }
+    if (!m_is_required) {
+      usage << "[";
+    }
+    usage << longest_name;
+    const std::string metavar = !m_metavar.empty() ? m_metavar : "VAR";
+    if (m_num_args_range.get_max() > 0) {
+      usage << " " << metavar;
+      if (m_num_args_range.get_max() > 1 &&
+          m_metavar.find("> <") == std::string::npos) {
+        usage << "...";
+      }
+    }
+    if (!m_is_required) {
+      usage << "]";
+    }
+    if (m_is_repeatable) {
+      usage << "...";
+    }
+    return usage.str();
+  }
+
+  std::size_t get_arguments_length() const {
+
+    std::size_t names_size = std::accumulate(
+        std::begin(m_names), std::end(m_names), std::size_t(0),
+        [](const auto &sum, const auto &s) { return sum + s.size(); });
+
+    if (is_positional(m_names.front(), m_prefix_chars)) {
+      // A set metavar means this replaces the names
+      if (!m_metavar.empty()) {
+        // Indent and metavar
+        return 2 + m_metavar.size();
+      }
+
+      // Indent and space-separated
+      return 2 + names_size + (m_names.size() - 1);
+    }
+    // Is an option - include both names _and_ metavar
+    // size = text + (", " between names)
+    std::size_t size = names_size + 2 * (m_names.size() - 1);
+    if (!m_metavar.empty() && m_num_args_range == NArgsRange{1, 1}) {
+      size += m_metavar.size() + 1;
+    }
+    return size + 2; // indent
+  }
+
+  friend std::ostream &operator<<(std::ostream &stream,
+                                  const Argument &argument) {
+    std::stringstream name_stream;
+    name_stream << "  "; // indent
+    if (argument.is_positional(argument.m_names.front(),
+                               argument.m_prefix_chars)) {
+      if (!argument.m_metavar.empty()) {
+        name_stream << argument.m_metavar;
+      } else {
+        name_stream << details::join(argument.m_names.begin(),
+                                     argument.m_names.end(), " ");
+      }
+    } else {
+      name_stream << details::join(argument.m_names.begin(),
+                                   argument.m_names.end(), ", ");
+      // If we have a metavar, and one narg - print the metavar
+      if (!argument.m_metavar.empty() &&
+          argument.m_num_args_range == NArgsRange{1, 1}) {
+        name_stream << " " << argument.m_metavar;
+      }
+      else if (!argument.m_metavar.empty() &&
+               argument.m_num_args_range.get_min() == argument.m_num_args_range.get_max() &&
+               argument.m_metavar.find("> <") != std::string::npos) {
+        name_stream << " " << argument.m_metavar;
+      }
+    }
+
+    // align multiline help message
+    auto stream_width = stream.width();
+    auto name_padding = std::string(name_stream.str().size(), ' ');
+    auto pos = std::string::size_type{};
+    auto prev = std::string::size_type{};
+    auto first_line = true;
+    auto hspace = "  "; // minimal space between name and help message
+    stream << name_stream.str();
+    std::string_view help_view(argument.m_help);
+    while ((pos = argument.m_help.find('\n', prev)) != std::string::npos) {
+      auto line = help_view.substr(prev, pos - prev + 1);
+      if (first_line) {
+        stream << hspace << line;
+        first_line = false;
+      } else {
+        stream.width(stream_width);
+        stream << name_padding << hspace << line;
+      }
+      prev += pos - prev + 1;
+    }
+    if (first_line) {
+      stream << hspace << argument.m_help;
+    } else {
+      auto leftover = help_view.substr(prev, argument.m_help.size() - prev);
+      if (!leftover.empty()) {
+        stream.width(stream_width);
+        stream << name_padding << hspace << leftover;
+      }
+    }
+
+    // print nargs spec
+    if (!argument.m_help.empty()) {
+      stream << " ";
+    }
+    stream << argument.m_num_args_range;
+
+    bool add_space = false;
+    if (argument.m_default_value.has_value() &&
+        argument.m_num_args_range != NArgsRange{0, 0}) {
+      stream << "[default: " << argument.m_default_value_repr << "]";
+      add_space = true;
+    } else if (argument.m_is_required) {
+      stream << "[required]";
+      add_space = true;
+    }
+    if (argument.m_is_repeatable) {
+      if (add_space) {
+        stream << " ";
+      }
+      stream << "[may be repeated]";
+    }
+    stream << "\n";
+    return stream;
+  }
+
+  template <typename T> bool operator!=(const T &rhs) const {
+    return !(*this == rhs);
+  }
+
+  /*
+   * Compare to an argument value of known type
+   * @throws std::logic_error in case of incompatible types
+   */
+  template <typename T> bool operator==(const T &rhs) const {
+    if constexpr (!details::IsContainer<T>) {
+      return get<T>() == rhs;
+    } else {
+      using ValueType = typename T::value_type;
+      auto lhs = get<T>();
+      return std::equal(std::begin(lhs), std::end(lhs), std::begin(rhs),
+                        std::end(rhs), [](const auto &a, const auto &b) {
+                          return std::any_cast<const ValueType &>(a) == b;
+                        });
+    }
+  }
+
+  /*
+   * positional:
+   *    _empty_
+   *    '-'
+   *    '-' decimal-literal
+   *    !'-' anything
+   */
+  static bool is_positional(std::string_view name,
+                            std::string_view prefix_chars) {
+    auto first = lookahead(name);
+
+    if (first == eof) {
+      return true;
+    }
+    if (prefix_chars.find(static_cast<char>(first)) !=
+                          std::string_view::npos) {
+      name.remove_prefix(1);
+      if (name.empty()) {
+        return true;
+      }
+      return is_decimal_literal(name);
+    }
+    return true;
+  }
+
+private:
+  class NArgsRange {
+    std::size_t m_min;
+    std::size_t m_max;
+
+  public:
+    NArgsRange(std::size_t minimum, std::size_t maximum)
+        : m_min(minimum), m_max(maximum) {
+      if (minimum > maximum) {
+        throw std::logic_error("Range of number of arguments is invalid");
+      }
+    }
+
+    bool contains(std::size_t value) const {
+      return value >= m_min && value <= m_max;
+    }
+
+    bool is_exact() const { return m_min == m_max; }
+
+    bool is_right_bounded() const {
+      return m_max < (std::numeric_limits<std::size_t>::max)();
+    }
+
+    std::size_t get_min() const { return m_min; }
+
+    std::size_t get_max() const { return m_max; }
+
+    // Print help message
+    friend auto operator<<(std::ostream &stream, const NArgsRange &range)
+        -> std::ostream & {
+      if (range.m_min == range.m_max) {
+        if (range.m_min != 0 && range.m_min != 1) {
+          stream << "[nargs: " << range.m_min << "] ";
+        }
+      } else {
+        if (range.m_max == (std::numeric_limits<std::size_t>::max)()) {
+          stream << "[nargs: " << range.m_min << " or more] ";
+        } else {
+          stream << "[nargs=" << range.m_min << ".." << range.m_max << "] ";
+        }
+      }
+      return stream;
+    }
+
+    bool operator==(const NArgsRange &rhs) const {
+      return rhs.m_min == m_min && rhs.m_max == m_max;
+    }
+
+    bool operator!=(const NArgsRange &rhs) const { return !(*this == rhs); }
+  };
+
+  void throw_nargs_range_validation_error() const {
+    std::stringstream stream;
+    if (!m_used_name.empty()) {
+      stream << m_used_name << ": ";
+    } else {
+      stream << m_names.front() << ": ";
+    }
+    if (m_num_args_range.is_exact()) {
+      stream << m_num_args_range.get_min();
+    } else if (m_num_args_range.is_right_bounded()) {
+      stream << m_num_args_range.get_min() << " to "
+             << m_num_args_range.get_max();
+    } else {
+      stream << m_num_args_range.get_min() << " or more";
+    }
+    stream << " argument(s) expected. " << m_values.size() << " provided.";
+    throw std::runtime_error(stream.str());
+  }
+
+  void throw_required_arg_not_used_error() const {
+    std::stringstream stream;
+    stream << m_names.front() << ": required.";
+    throw std::runtime_error(stream.str());
+  }
+
+  void throw_required_arg_no_value_provided_error() const {
+    std::stringstream stream;
+    stream << m_used_name << ": no value provided.";
+    throw std::runtime_error(stream.str());
+  }
+
+  static constexpr int eof = std::char_traits<char>::eof();
+
+  static auto lookahead(std::string_view s) -> int {
+    if (s.empty()) {
+      return eof;
+    }
+    return static_cast<int>(static_cast<unsigned char>(s[0]));
+  }
+
+  /*
+   * decimal-literal:
+   *    '0'
+   *    nonzero-digit digit-sequence_opt
+   *    integer-part fractional-part
+   *    fractional-part
+   *    integer-part '.' exponent-part_opt
+   *    integer-part exponent-part
+   *
+   * integer-part:
+   *    digit-sequence
+   *
+   * fractional-part:
+   *    '.' post-decimal-point
+   *
+   * post-decimal-point:
+   *    digit-sequence exponent-part_opt
+   *
+   * exponent-part:
+   *    'e' post-e
+   *    'E' post-e
+   *
+   * post-e:
+   *    sign_opt digit-sequence
+   *
+   * sign: one of
+   *    '+' '-'
+   */
+  static bool is_decimal_literal(std::string_view s) {
+    auto is_digit = [](auto c) constexpr {
+      switch (c) {
+      case '0':
+      case '1':
+      case '2':
+      case '3':
+      case '4':
+      case '5':
+      case '6':
+      case '7':
+      case '8':
+      case '9':
+        return true;
+      default:
+        return false;
+      }
+    };
+
+    // precondition: we have consumed or will consume at least one digit
+    auto consume_digits = [=](std::string_view sd) {
+      // NOLINTNEXTLINE(readability-qualified-auto)
+      auto it = std::find_if_not(std::begin(sd), std::end(sd), is_digit);
+      return sd.substr(static_cast<std::size_t>(it - std::begin(sd)));
+    };
+
+    switch (lookahead(s)) {
+    case '0': {
+      s.remove_prefix(1);
+      if (s.empty()) {
+        return true;
+      }
+      goto integer_part;
+    }
+    case '1':
+    case '2':
+    case '3':
+    case '4':
+    case '5':
+    case '6':
+    case '7':
+    case '8':
+    case '9': {
+      s = consume_digits(s);
+      if (s.empty()) {
+        return true;
+      }
+      goto integer_part_consumed;
+    }
+    case '.': {
+      s.remove_prefix(1);
+      goto post_decimal_point;
+    }
+    default:
+      return false;
+    }
+
+  integer_part:
+    s = consume_digits(s);
+  integer_part_consumed:
+    switch (lookahead(s)) {
+    case '.': {
+      s.remove_prefix(1);
+      if (is_digit(lookahead(s))) {
+        goto post_decimal_point;
+      } else {
+        goto exponent_part_opt;
+      }
+    }
+    case 'e':
+    case 'E': {
+      s.remove_prefix(1);
+      goto post_e;
+    }
+    default:
+      return false;
+    }
+
+  post_decimal_point:
+    if (is_digit(lookahead(s))) {
+      s = consume_digits(s);
+      goto exponent_part_opt;
+    }
+    return false;
+
+  exponent_part_opt:
+    switch (lookahead(s)) {
+    case eof:
+      return true;
+    case 'e':
+    case 'E': {
+      s.remove_prefix(1);
+      goto post_e;
+    }
+    default:
+      return false;
+    }
+
+  post_e:
+    switch (lookahead(s)) {
+    case '-':
+    case '+':
+      s.remove_prefix(1);
+    }
+    if (is_digit(lookahead(s))) {
+      s = consume_digits(s);
+      return s.empty();
+    }
+    return false;
+  }
+
+  static bool is_optional(std::string_view name,
+                          std::string_view prefix_chars) {
+    return !is_positional(name, prefix_chars);
+  }
+
+  /*
+   * Get argument value given a type
+   * @throws std::logic_error in case of incompatible types
+   */
+  template <typename T> T get() const {
+    if (!m_values.empty()) {
+      if constexpr (details::IsContainer<T>) {
+        return any_cast_container<T>(m_values);
+      } else {
+        return std::any_cast<T>(m_values.front());
+      }
+    }
+    if (m_default_value.has_value()) {
+      return std::any_cast<T>(m_default_value);
+    }
+    if constexpr (details::IsContainer<T>) {
+      if (!m_accepts_optional_like_value) {
+        return any_cast_container<T>(m_values);
+      }
+    }
+
+    throw std::logic_error("No value provided for '" + m_names.back() + "'.");
+  }
+
+  /*
+   * Get argument value given a type.
+   * @pre The object has no default value.
+   * @returns The stored value if any, std::nullopt otherwise.
+   */
+  template <typename T> auto present() const -> std::optional<T> {
+    if (m_default_value.has_value()) {
+      throw std::logic_error("Argument with default value always presents");
+    }
+    if (m_values.empty()) {
+      return std::nullopt;
+    }
+    if constexpr (details::IsContainer<T>) {
+      return any_cast_container<T>(m_values);
+    }
+    return std::any_cast<T>(m_values.front());
+  }
+
+  template <typename T>
+  static auto any_cast_container(const std::vector<std::any> &operand) -> T {
+    using ValueType = typename T::value_type;
+
+    T result;
+    std::transform(
+        std::begin(operand), std::end(operand), std::back_inserter(result),
+        [](const auto &value) { return std::any_cast<ValueType>(value); });
+    return result;
+  }
+
+  void set_usage_newline_counter(int i) { m_usage_newline_counter = i; }
+
+  void set_group_idx(std::size_t i) { m_group_idx = i; }
+
+  std::vector<std::string> m_names;
+  std::string_view m_used_name;
+  std::string m_help;
+  std::string m_metavar;
+  std::any m_default_value;
+  std::string m_default_value_repr;
+  std::optional<std::string>
+      m_default_value_str; // used for checking default_value against choices
+  std::any m_implicit_value;
+  std::optional<std::vector<std::string>> m_choices{std::nullopt};
+  using valued_action = std::function<std::any(const std::string &)>;
+  using void_action = std::function<void(const std::string &)>;
+  std::variant<valued_action, void_action> m_action{
+      std::in_place_type<valued_action>,
+      [](const std::string &value) { return value; }};
+  std::vector<std::any> m_values;
+  NArgsRange m_num_args_range{1, 1};
+  // Bit field of bool values. Set default value in ctor.
+  bool m_accepts_optional_like_value : 1;
+  bool m_is_optional : 1;
+  bool m_is_required : 1;
+  bool m_is_repeatable : 1;
+  bool m_is_used : 1;
+  bool m_is_hidden : 1;            // if set, does not appear in usage or help
+  std::string_view m_prefix_chars; // ArgumentParser has the prefix_chars
+  int m_usage_newline_counter = 0;
+  std::size_t m_group_idx = 0;
+};
+
+class ArgumentParser {
+public:
+  explicit ArgumentParser(std::string program_name = {},
+                          std::string version = "1.0",
+                          default_arguments add_args = default_arguments::all,
+                          bool exit_on_default_arguments = true,
+                          std::ostream &os = std::cout)
+      : m_program_name(std::move(program_name)), m_version(std::move(version)),
+        m_exit_on_default_arguments(exit_on_default_arguments),
+        m_parser_path(m_program_name) {
+    if ((add_args & default_arguments::help) == default_arguments::help) {
+      add_argument("-h", "--help")
+          .action([&](const auto & /*unused*/) {
+            os << help().str();
+            if (m_exit_on_default_arguments) {
+              std::exit(0);
+            }
+          })
+          .default_value(false)
+          .help("shows help message and exits")
+          .implicit_value(true)
+          .nargs(0);
+    }
+    if ((add_args & default_arguments::version) == default_arguments::version) {
+      add_argument("-v", "--version")
+          .action([&](const auto & /*unused*/) {
+            os << m_version << std::endl;
+            if (m_exit_on_default_arguments) {
+              std::exit(0);
+            }
+          })
+          .default_value(false)
+          .help("prints version information and exits")
+          .implicit_value(true)
+          .nargs(0);
+    }
+  }
+
+  ~ArgumentParser() = default;
+
+  // ArgumentParser is meant to be used in a single function.
+  // Setup everything and parse arguments in one place.
+  //
+  // ArgumentParser internally uses std::string_views,
+  // references, iterators, etc.
+  // Many of these elements become invalidated after a copy or move.
+  ArgumentParser(const ArgumentParser &other) = delete;
+  ArgumentParser &operator=(const ArgumentParser &other) = delete;
+  ArgumentParser(ArgumentParser &&) noexcept = delete;
+  ArgumentParser &operator=(ArgumentParser &&) = delete;
+
+  explicit operator bool() const {
+    auto arg_used = std::any_of(m_argument_map.cbegin(), m_argument_map.cend(),
+                                [](auto &it) { return it.second->m_is_used; });
+    auto subparser_used =
+        std::any_of(m_subparser_used.cbegin(), m_subparser_used.cend(),
+                    [](auto &it) { return it.second; });
+
+    return m_is_parsed && (arg_used || subparser_used);
+  }
+
+  // Parameter packing
+  // Call add_argument with variadic number of string arguments
+  template <typename... Targs> Argument &add_argument(Targs... f_args) {
+    using array_of_sv = std::array<std::string_view, sizeof...(Targs)>;
+    auto argument =
+        m_optional_arguments.emplace(std::cend(m_optional_arguments),
+                                     m_prefix_chars, array_of_sv{f_args...});
+
+    if (!argument->m_is_optional) {
+      m_positional_arguments.splice(std::cend(m_positional_arguments),
+                                    m_optional_arguments, argument);
+    }
+    argument->set_usage_newline_counter(m_usage_newline_counter);
+    argument->set_group_idx(m_group_names.size());
+
+    index_argument(argument);
+    return *argument;
+  }
+
+  class MutuallyExclusiveGroup {
+    friend class ArgumentParser;
+
+  public:
+    MutuallyExclusiveGroup() = delete;
+
+    explicit MutuallyExclusiveGroup(ArgumentParser &parent,
+                                    bool required = false)
+        : m_parent(parent), m_required(required), m_elements({}) {}
+
+    MutuallyExclusiveGroup(const MutuallyExclusiveGroup &other) = delete;
+    MutuallyExclusiveGroup &
+    operator=(const MutuallyExclusiveGroup &other) = delete;
+
+    MutuallyExclusiveGroup(MutuallyExclusiveGroup &&other) noexcept
+        : m_parent(other.m_parent), m_required(other.m_required),
+          m_elements(std::move(other.m_elements)) {
+      other.m_elements.clear();
+    }
+
+    template <typename... Targs> Argument &add_argument(Targs... f_args) {
+      auto &argument = m_parent.add_argument(std::forward<Targs>(f_args)...);
+      m_elements.push_back(&argument);
+      argument.set_usage_newline_counter(m_parent.m_usage_newline_counter);
+      argument.set_group_idx(m_parent.m_group_names.size());
+      return argument;
+    }
+
+  private:
+    ArgumentParser &m_parent;
+    bool m_required{false};
+    std::vector<Argument *> m_elements{};
+  };
+
+  MutuallyExclusiveGroup &add_mutually_exclusive_group(bool required = false) {
+    m_mutually_exclusive_groups.emplace_back(*this, required);
+    return m_mutually_exclusive_groups.back();
+  }
+
+  // Parameter packed add_parents method
+  // Accepts a variadic number of ArgumentParser objects
+  template <typename... Targs>
+  ArgumentParser &add_parents(const Targs &... f_args) {
+    for (const ArgumentParser &parent_parser : {std::ref(f_args)...}) {
+      for (const auto &argument : parent_parser.m_positional_arguments) {
+        auto it = m_positional_arguments.insert(
+            std::cend(m_positional_arguments), argument);
+        index_argument(it);
+      }
+      for (const auto &argument : parent_parser.m_optional_arguments) {
+        auto it = m_optional_arguments.insert(std::cend(m_optional_arguments),
+                                              argument);
+        index_argument(it);
+      }
+    }
+    return *this;
+  }
+
+  // Ask for the next optional arguments to be displayed on a separate
+  // line in usage() output. Only effective if set_usage_max_line_width() is
+  // also used.
+  ArgumentParser &add_usage_newline() {
+    ++m_usage_newline_counter;
+    return *this;
+  }
+
+  // Ask for the next optional arguments to be displayed in a separate section
+  // in usage() and help (<< *this) output.
+  // For usage(), this is only effective if set_usage_max_line_width() is
+  // also used.
+  ArgumentParser &add_group(std::string group_name) {
+    m_group_names.emplace_back(std::move(group_name));
+    return *this;
+  }
+
+  ArgumentParser &add_description(std::string description) {
+    m_description = std::move(description);
+    return *this;
+  }
+
+  ArgumentParser &add_epilog(std::string epilog) {
+    m_epilog = std::move(epilog);
+    return *this;
+  }
+
+  // Add a un-documented/hidden alias for an argument.
+  // Ideally we'd want this to be a method of Argument, but Argument
+  // does not own its owing ArgumentParser.
+  ArgumentParser &add_hidden_alias_for(Argument &arg, std::string_view alias) {
+    for (auto it = m_optional_arguments.begin();
+         it != m_optional_arguments.end(); ++it) {
+      if (&(*it) == &arg) {
+        m_argument_map.insert_or_assign(std::string(alias), it);
+        return *this;
+      }
+    }
+    throw std::logic_error(
+        "Argument is not an optional argument of this parser");
+  }
+
+  /* Getter for arguments and subparsers.
+   * @throws std::logic_error in case of an invalid argument or subparser name
+   */
+  template <typename T = Argument> T &at(std::string_view name) {
+    if constexpr (std::is_same_v<T, Argument>) {
+      return (*this)[name];
+    } else {
+      std::string str_name(name);
+      auto subparser_it = m_subparser_map.find(str_name);
+      if (subparser_it != m_subparser_map.end()) {
+        return subparser_it->second->get();
+      }
+      throw std::logic_error("No such subparser: " + str_name);
+    }
+  }
+
+  ArgumentParser &set_prefix_chars(std::string prefix_chars) {
+    m_prefix_chars = std::move(prefix_chars);
+    return *this;
+  }
+
+  ArgumentParser &set_assign_chars(std::string assign_chars) {
+    m_assign_chars = std::move(assign_chars);
+    return *this;
+  }
+
+  /* Call parse_args_internal - which does all the work
+   * Then, validate the parsed arguments
+   * This variant is used mainly for testing
+   * @throws std::runtime_error in case of any invalid argument
+   */
+  void parse_args(const std::vector<std::string> &arguments) {
+    parse_args_internal(arguments);
+    // Check if all arguments are parsed
+    for ([[maybe_unused]] const auto &[unused, argument] : m_argument_map) {
+      argument->validate();
+    }
+
+    // Check each mutually exclusive group and make sure
+    // there are no constraint violations
+    for (const auto &group : m_mutually_exclusive_groups) {
+      auto mutex_argument_used{false};
+      Argument *mutex_argument_it{nullptr};
+      for (Argument *arg : group.m_elements) {
+        if (!mutex_argument_used && arg->m_is_used) {
+          mutex_argument_used = true;
+          mutex_argument_it = arg;
+        } else if (mutex_argument_used && arg->m_is_used) {
+          // Violation
+          throw std::runtime_error("Argument '" + arg->get_usage_full() +
+                                   "' not allowed with '" +
+                                   mutex_argument_it->get_usage_full() + "'");
+        }
+      }
+
+      if (!mutex_argument_used && group.m_required) {
+        // at least one argument from the group is
+        // required
+        std::string argument_names{};
+        std::size_t i = 0;
+        std::size_t size = group.m_elements.size();
+        for (Argument *arg : group.m_elements) {
+          if (i + 1 == size) {
+            // last
+            argument_names += std::string("'") + arg->get_usage_full() + std::string("' ");
+          } else {
+            argument_names += std::string("'") + arg->get_usage_full() + std::string("' or ");
+          }
+          i += 1;
+        }
+        throw std::runtime_error("One of the arguments " + argument_names +
+                                 "is required");
+      }
+    }
+  }
+
+  /* Call parse_known_args_internal - which does all the work
+   * Then, validate the parsed arguments
+   * This variant is used mainly for testing
+   * @throws std::runtime_error in case of any invalid argument
+   */
+  std::vector<std::string>
+  parse_known_args(const std::vector<std::string> &arguments) {
+    auto unknown_arguments = parse_known_args_internal(arguments);
+    // Check if all arguments are parsed
+    for ([[maybe_unused]] const auto &[unused, argument] : m_argument_map) {
+      argument->validate();
+    }
+    return unknown_arguments;
+  }
+
+  /* Main entry point for parsing command-line arguments using this
+   * ArgumentParser
+   * @throws std::runtime_error in case of any invalid argument
+   */
+  // NOLINTNEXTLINE(cppcoreguidelines-avoid-c-arrays)
+  void parse_args(int argc, const char *const argv[]) {
+    parse_args({argv, argv + argc});
+  }
+
+  /* Main entry point for parsing command-line arguments using this
+   * ArgumentParser
+   * @throws std::runtime_error in case of any invalid argument
+   */
+  // NOLINTNEXTLINE(cppcoreguidelines-avoid-c-arrays)
+  auto parse_known_args(int argc, const char *const argv[]) {
+    return parse_known_args({argv, argv + argc});
+  }
+
+  /* Getter for options with default values.
+   * @throws std::logic_error if parse_args() has not been previously called
+   * @throws std::logic_error if there is no such option
+   * @throws std::logic_error if the option has no value
+   * @throws std::bad_any_cast if the option is not of type T
+   */
+  template <typename T = std::string> T get(std::string_view arg_name) const {
+    if (!m_is_parsed) {
+      throw std::logic_error("Nothing parsed, no arguments are available.");
+    }
+    return (*this)[arg_name].get<T>();
+  }
+
+  /* Getter for options without default values.
+   * @pre The option has no default value.
+   * @throws std::logic_error if there is no such option
+   * @throws std::bad_any_cast if the option is not of type T
+   */
+  template <typename T = std::string>
+  auto present(std::string_view arg_name) const -> std::optional<T> {
+    return (*this)[arg_name].present<T>();
+  }
+
+  /* Getter that returns true for user-supplied options. Returns false if not
+   * user-supplied, even with a default value.
+   */
+  auto is_used(std::string_view arg_name) const {
+    return (*this)[arg_name].m_is_used;
+  }
+
+  /* Getter that returns true if a subcommand is used.
+   */
+  auto is_subcommand_used(std::string_view subcommand_name) const {
+    return m_subparser_used.at(std::string(subcommand_name));
+  }
+
+  /* Getter that returns true if a subcommand is used.
+   */
+  auto is_subcommand_used(const ArgumentParser &subparser) const {
+    return is_subcommand_used(subparser.m_program_name);
+  }
+
+  /* Indexing operator. Return a reference to an Argument object
+   * Used in conjunction with Argument.operator== e.g., parser["foo"] == true
+   * @throws std::logic_error in case of an invalid argument name
+   */
+  Argument &operator[](std::string_view arg_name) const {
+    std::string name(arg_name);
+    auto it = m_argument_map.find(name);
+    if (it != m_argument_map.end()) {
+      return *(it->second);
+    }
+    if (!is_valid_prefix_char(arg_name.front())) {
+      const auto legal_prefix_char = get_any_valid_prefix_char();
+      const auto prefix = std::string(1, legal_prefix_char);
+
+      // "-" + arg_name
+      name = prefix + name;
+      it = m_argument_map.find(name);
+      if (it != m_argument_map.end()) {
+        return *(it->second);
+      }
+      // "--" + arg_name
+      name = prefix + name;
+      it = m_argument_map.find(name);
+      if (it != m_argument_map.end()) {
+        return *(it->second);
+      }
+    }
+    throw std::logic_error("No such argument: " + std::string(arg_name));
+  }
+
+  // Print help message
+  friend auto operator<<(std::ostream &stream, const ArgumentParser &parser)
+      -> std::ostream & {
+    stream.setf(std::ios_base::left);
+
+    auto longest_arg_length = parser.get_length_of_longest_argument();
+
+    stream << parser.usage() << "\n\n";
+
+    if (!parser.m_description.empty()) {
+      stream << parser.m_description << "\n\n";
+    }
+
+    const bool has_visible_positional_args = std::find_if(
+      parser.m_positional_arguments.begin(),
+      parser.m_positional_arguments.end(),
+      [](const auto &argument) {
+      return !argument.m_is_hidden; }) !=
+      parser.m_positional_arguments.end();
+    if (has_visible_positional_args) {
+      stream << "Positional arguments:\n";
+    }
+
+    for (const auto &argument : parser.m_positional_arguments) {
+      if (!argument.m_is_hidden) {
+        stream.width(static_cast<std::streamsize>(longest_arg_length));
+        stream << argument;
+      }
+    }
+
+    if (!parser.m_optional_arguments.empty()) {
+      stream << (!has_visible_positional_args ? "" : "\n")
+             << "Optional arguments:\n";
+    }
+
+    for (const auto &argument : parser.m_optional_arguments) {
+      if (argument.m_group_idx == 0 && !argument.m_is_hidden) {
+        stream.width(static_cast<std::streamsize>(longest_arg_length));
+        stream << argument;
+      }
+    }
+
+    for (size_t i_group = 0; i_group < parser.m_group_names.size(); ++i_group) {
+      stream << "\n" << parser.m_group_names[i_group] << " (detailed usage):\n";
+      for (const auto &argument : parser.m_optional_arguments) {
+        if (argument.m_group_idx == i_group + 1 && !argument.m_is_hidden) {
+          stream.width(static_cast<std::streamsize>(longest_arg_length));
+          stream << argument;
+        }
+      }
+    }
+
+    bool has_visible_subcommands = std::any_of(
+        parser.m_subparser_map.begin(), parser.m_subparser_map.end(),
+        [](auto &p) { return !p.second->get().m_suppress; });
+
+    if (has_visible_subcommands) {
+      stream << (parser.m_positional_arguments.empty()
+                     ? (parser.m_optional_arguments.empty() ? "" : "\n")
+                     : "\n")
+             << "Subcommands:\n";
+      for (const auto &[command, subparser] : parser.m_subparser_map) {
+        if (subparser->get().m_suppress) {
+          continue;
+        }
+
+        stream << std::setw(2) << " ";
+        stream << std::setw(static_cast<int>(longest_arg_length - 2))
+               << command;
+        stream << " " << subparser->get().m_description << "\n";
+      }
+    }
+
+    if (!parser.m_epilog.empty()) {
+      stream << '\n';
+      stream << parser.m_epilog << "\n\n";
+    }
+
+    return stream;
+  }
+
+  // Format help message
+  auto help() const -> std::stringstream {
+    std::stringstream out;
+    out << *this;
+    return out;
+  }
+
+  // Sets the maximum width for a line of the Usage message
+  ArgumentParser &set_usage_max_line_width(size_t w) {
+    this->m_usage_max_line_width = w;
+    return *this;
+  }
+
+  // Asks to display arguments of mutually exclusive group on separate lines in
+  // the Usage message
+  ArgumentParser &set_usage_break_on_mutex() {
+    this->m_usage_break_on_mutex = true;
+    return *this;
+  }
+
+  // Format usage part of help only
+  auto usage() const -> std::string {
+    std::stringstream stream;
+
+    std::string curline("Usage: ");
+    curline += this->m_parser_path;
+    const bool multiline_usage =
+        this->m_usage_max_line_width < (std::numeric_limits<std::size_t>::max)();
+    const size_t indent_size = curline.size();
+
+    const auto deal_with_options_of_group = [&](std::size_t group_idx) {
+      bool found_options = false;
+      // Add any options inline here
+      const MutuallyExclusiveGroup *cur_mutex = nullptr;
+      int usage_newline_counter = -1;
+      for (const auto &argument : this->m_optional_arguments) {
+        if (argument.m_is_hidden) {
+          continue;
+        }
+        if (multiline_usage) {
+          if (argument.m_group_idx != group_idx) {
+            continue;
+          }
+          if (usage_newline_counter != argument.m_usage_newline_counter) {
+            if (usage_newline_counter >= 0) {
+              if (curline.size() > indent_size) {
+                stream << curline << std::endl;
+                curline = std::string(indent_size, ' ');
+              }
+            }
+            usage_newline_counter = argument.m_usage_newline_counter;
+          }
+        }
+        found_options = true;
+        const std::string arg_inline_usage = argument.get_inline_usage();
+        const MutuallyExclusiveGroup *arg_mutex =
+            get_belonging_mutex(&argument);
+        if ((cur_mutex != nullptr) && (arg_mutex == nullptr)) {
+          curline += ']';
+          if (this->m_usage_break_on_mutex) {
+            stream << curline << std::endl;
+            curline = std::string(indent_size, ' ');
+          }
+        } else if ((cur_mutex == nullptr) && (arg_mutex != nullptr)) {
+          if ((this->m_usage_break_on_mutex && curline.size() > indent_size) ||
+              curline.size() + 3 + arg_inline_usage.size() >
+                  this->m_usage_max_line_width) {
+            stream << curline << std::endl;
+            curline = std::string(indent_size, ' ');
+          }
+          curline += " [";
+        } else if ((cur_mutex != nullptr) && (arg_mutex != nullptr)) {
+          if (cur_mutex != arg_mutex) {
+            curline += ']';
+            if (this->m_usage_break_on_mutex ||
+                curline.size() + 3 + arg_inline_usage.size() >
+                    this->m_usage_max_line_width) {
+              stream << curline << std::endl;
+              curline = std::string(indent_size, ' ');
+            }
+            curline += " [";
+          } else {
+            curline += '|';
+          }
+        }
+        cur_mutex = arg_mutex;
+        if (curline.size() + 1 + arg_inline_usage.size() >
+            this->m_usage_max_line_width) {
+          stream << curline << std::endl;
+          curline = std::string(indent_size, ' ');
+          curline += " ";
+        } else if (cur_mutex == nullptr) {
+          curline += " ";
+        }
+        curline += arg_inline_usage;
+      }
+      if (cur_mutex != nullptr) {
+        curline += ']';
+      }
+      return found_options;
+    };
+
+    const bool found_options = deal_with_options_of_group(0);
+
+    if (found_options && multiline_usage &&
+        !this->m_positional_arguments.empty()) {
+      stream << curline << std::endl;
+      curline = std::string(indent_size, ' ');
+    }
+    // Put positional arguments after the optionals
+    for (const auto &argument : this->m_positional_arguments) {
+      if (argument.m_is_hidden) {
+        continue;
+      }
+      const std::string pos_arg = !argument.m_metavar.empty()
+                                      ? argument.m_metavar
+                                      : argument.m_names.front();
+      if (curline.size() + 1 + pos_arg.size() > this->m_usage_max_line_width) {
+        stream << curline << std::endl;
+        curline = std::string(indent_size, ' ');
+      }
+      curline += " ";
+      if (argument.m_num_args_range.get_min() == 0 &&
+          !argument.m_num_args_range.is_right_bounded()) {
+        curline += "[";
+        curline += pos_arg;
+        curline += "]...";
+      } else if (argument.m_num_args_range.get_min() == 1 &&
+                 !argument.m_num_args_range.is_right_bounded()) {
+        curline += pos_arg;
+        curline += "...";
+      } else {
+        curline += pos_arg;
+      }
+    }
+
+    if (multiline_usage) {
+      // Display options of other groups
+      for (std::size_t i = 0; i < m_group_names.size(); ++i) {
+        stream << curline << std::endl << std::endl;
+        stream << m_group_names[i] << ":" << std::endl;
+        curline = std::string(indent_size, ' ');
+        deal_with_options_of_group(i + 1);
+      }
+    }
+
+    stream << curline;
+
+    // Put subcommands after positional arguments
+    if (!m_subparser_map.empty()) {
+      stream << " {";
+      std::size_t i{0};
+      for (const auto &[command, subparser] : m_subparser_map) {
+        if (subparser->get().m_suppress) {
+          continue;
+        }
+
+        if (i == 0) {
+          stream << command;
+        } else {
+          stream << "," << command;
+        }
+        ++i;
+      }
+      stream << "}";
+    }
+
+    return stream.str();
+  }
+
+  // Printing the one and only help message
+  // I've stuck with a simple message format, nothing fancy.
+  [[deprecated("Use cout << program; instead.  See also help().")]] std::string
+  print_help() const {
+    auto out = help();
+    std::cout << out.rdbuf();
+    return out.str();
+  }
+
+  void add_subparser(ArgumentParser &parser) {
+    parser.m_parser_path = m_program_name + " " + parser.m_program_name;
+    auto it = m_subparsers.emplace(std::cend(m_subparsers), parser);
+    m_subparser_map.insert_or_assign(parser.m_program_name, it);
+    m_subparser_used.insert_or_assign(parser.m_program_name, false);
+  }
+
+  void set_suppress(bool suppress) { m_suppress = suppress; }
+
+protected:
+  const MutuallyExclusiveGroup *get_belonging_mutex(const Argument *arg) const {
+    for (const auto &mutex : m_mutually_exclusive_groups) {
+      if (std::find(mutex.m_elements.begin(), mutex.m_elements.end(), arg) !=
+          mutex.m_elements.end()) {
+        return &mutex;
+      }
+    }
+    return nullptr;
+  }
+
+  bool is_valid_prefix_char(char c) const {
+    return m_prefix_chars.find(c) != std::string::npos;
+  }
+
+  char get_any_valid_prefix_char() const { return m_prefix_chars[0]; }
+
+  /*
+   * Pre-process this argument list. Anything starting with "--", that
+   * contains an =, where the prefix before the = has an entry in the
+   * options table, should be split.
+   */
+  std::vector<std::string>
+  preprocess_arguments(const std::vector<std::string> &raw_arguments) const {
+    std::vector<std::string> arguments{};
+    for (const auto &arg : raw_arguments) {
+      
+      const auto argument_starts_with_prefix_chars =
+          [this](const std::string &a) -> bool {
+        if (!a.empty()) {
+
+          const auto legal_prefix = [this](char c) -> bool {
+            return m_prefix_chars.find(c) != std::string::npos;
+          };
+
+          // Windows-style
+          // if '/' is a legal prefix char
+          // then allow single '/' followed by argument name, followed by an
+          // assign char, e.g., ':' e.g., 'test.exe /A:Foo'
+          const auto windows_style = legal_prefix('/');
+
+          if (windows_style) {
+            if (legal_prefix(a[0])) {
+              return true;
+            }
+          } else {
+            // Slash '/' is not a legal prefix char
+            // For all other characters, only support long arguments
+            // i.e., the argument must start with 2 prefix chars, e.g,
+            // '--foo' e,g, './test --foo=Bar -DARG=yes'
+            if (a.size() > 1) {
+              return (legal_prefix(a[0]) && legal_prefix(a[1]));
+            }
+          }
+        }
+        return false;
+      };
+
+      // Check that:
+      // - We don't have an argument named exactly this
+      // - The argument starts with a prefix char, e.g., "--"
+      // - The argument contains an assign char, e.g., "="
+      auto assign_char_pos = arg.find_first_of(m_assign_chars);
+
+      if (m_argument_map.find(arg) == m_argument_map.end() &&
+          argument_starts_with_prefix_chars(arg) &&
+          assign_char_pos != std::string::npos) {
+        // Get the name of the potential option, and check it exists
+        std::string opt_name = arg.substr(0, assign_char_pos);
+        if (m_argument_map.find(opt_name) != m_argument_map.end()) {
+          // This is the name of an option! Split it into two parts
+          arguments.push_back(std::move(opt_name));
+          arguments.push_back(arg.substr(assign_char_pos + 1));
+          continue;
+        }
+      }
+      // If we've fallen through to here, then it's a standard argument
+      arguments.push_back(arg);
+    }
+    return arguments;
+  }
+
+  /*
+   * @throws std::runtime_error in case of any invalid argument
+   */
+  void parse_args_internal(const std::vector<std::string> &raw_arguments) {
+    auto arguments = preprocess_arguments(raw_arguments);
+    if (m_program_name.empty() && !arguments.empty()) {
+      m_program_name = arguments.front();
+    }
+    auto end = std::end(arguments);
+    auto positional_argument_it = std::begin(m_positional_arguments);
+    for (auto it = std::next(std::begin(arguments)); it != end;) {
+      const auto &current_argument = *it;
+      if (Argument::is_positional(current_argument, m_prefix_chars)) {
+        if (positional_argument_it == std::end(m_positional_arguments)) {
+
+          // Check sub-parsers
+          auto subparser_it = m_subparser_map.find(current_argument);
+          if (subparser_it != m_subparser_map.end()) {
+
+            // build list of remaining args
+            const auto unprocessed_arguments =
+                std::vector<std::string>(it, end);
+
+            // invoke subparser
+            m_is_parsed = true;
+            m_subparser_used[current_argument] = true;
+            return subparser_it->second->get().parse_args(
+                unprocessed_arguments);
+          }
+
+          if (m_positional_arguments.empty()) {
+
+            // Ask the user if they argument they provided was a typo
+            // for some sub-parser,
+            // e.g., user provided `git totes` instead of `git notes`
+            if (!m_subparser_map.empty()) {
+              throw std::runtime_error(
+                  "Failed to parse '" + current_argument + "', did you mean '" +
+                  std::string{details::get_most_similar_string(
+                      m_subparser_map, current_argument)} +
+                  "'");
+            }
+
+            // Ask the user if they meant to use a specific optional argument
+            if (!m_optional_arguments.empty()) {
+              for (const auto &opt : m_optional_arguments) {
+                if (!opt.m_implicit_value.has_value()) {
+                  // not a flag, requires a value
+                  if (!opt.m_is_used) {
+                    throw std::runtime_error(
+                        "Zero positional arguments expected, did you mean " +
+                        opt.get_usage_full());
+                  }
+                }
+              }
+
+              throw std::runtime_error("Zero positional arguments expected");
+            } else {
+              throw std::runtime_error("Zero positional arguments expected");
+            }
+          } else {
+            throw std::runtime_error("Maximum number of positional arguments "
+                                     "exceeded, failed to parse '" +
+                                     current_argument + "'");
+          }
+        }
+        auto argument = positional_argument_it++;
+
+        // Deal with the situation of <positional_arg1>... <positional_arg2>
+        if (argument->m_num_args_range.get_min() == 1 &&
+            argument->m_num_args_range.get_max() == (std::numeric_limits<std::size_t>::max)() &&
+            positional_argument_it != std::end(m_positional_arguments) &&
+            std::next(positional_argument_it) == std::end(m_positional_arguments) &&
+            positional_argument_it->m_num_args_range.get_min() == 1 &&
+            positional_argument_it->m_num_args_range.get_max() == 1 ) {
+          if (std::next(it) != end) {
+            positional_argument_it->consume(std::prev(end), end);
+            end = std::prev(end);
+          } else {
+            throw std::runtime_error("Missing " + positional_argument_it->m_names.front());
+          }
+        }
+
+        it = argument->consume(it, end);
+        continue;
+      }
+
+      auto arg_map_it = m_argument_map.find(current_argument);
+      if (arg_map_it != m_argument_map.end()) {
+        auto argument = arg_map_it->second;
+        it = argument->consume(std::next(it), end, arg_map_it->first);
+      } else if (const auto &compound_arg = current_argument;
+                 compound_arg.size() > 1 &&
+                 is_valid_prefix_char(compound_arg[0]) &&
+                 !is_valid_prefix_char(compound_arg[1])) {
+        ++it;
+        for (std::size_t j = 1; j < compound_arg.size(); j++) {
+          auto hypothetical_arg = std::string{'-', compound_arg[j]};
+          auto arg_map_it2 = m_argument_map.find(hypothetical_arg);
+          if (arg_map_it2 != m_argument_map.end()) {
+            auto argument = arg_map_it2->second;
+            it = argument->consume(it, end, arg_map_it2->first);
+          } else {
+            throw std::runtime_error("Unknown argument: " + current_argument);
+          }
+        }
+      } else {
+        throw std::runtime_error("Unknown argument: " + current_argument);
+      }
+    }
+    m_is_parsed = true;
+  }
+
+  /*
+   * Like parse_args_internal but collects unused args into a vector<string>
+   */
+  std::vector<std::string>
+  parse_known_args_internal(const std::vector<std::string> &raw_arguments) {
+    auto arguments = preprocess_arguments(raw_arguments);
+
+    std::vector<std::string> unknown_arguments{};
+
+    if (m_program_name.empty() && !arguments.empty()) {
+      m_program_name = arguments.front();
+    }
+    auto end = std::end(arguments);
+    auto positional_argument_it = std::begin(m_positional_arguments);
+    for (auto it = std::next(std::begin(arguments)); it != end;) {
+      const auto &current_argument = *it;
+      if (Argument::is_positional(current_argument, m_prefix_chars)) {
+        if (positional_argument_it == std::end(m_positional_arguments)) {
+
+          // Check sub-parsers
+          auto subparser_it = m_subparser_map.find(current_argument);
+          if (subparser_it != m_subparser_map.end()) {
+
+            // build list of remaining args
+            const auto unprocessed_arguments =
+                std::vector<std::string>(it, end);
+
+            // invoke subparser
+            m_is_parsed = true;
+            m_subparser_used[current_argument] = true;
+            return subparser_it->second->get().parse_known_args_internal(
+                unprocessed_arguments);
+          }
+
+          // save current argument as unknown and go to next argument
+          unknown_arguments.push_back(current_argument);
+          ++it;
+        } else {
+          // current argument is the value of a positional argument
+          // consume it
+          auto argument = positional_argument_it++;
+          it = argument->consume(it, end);
+        }
+        continue;
+      }
+
+      auto arg_map_it = m_argument_map.find(current_argument);
+      if (arg_map_it != m_argument_map.end()) {
+        auto argument = arg_map_it->second;
+        it = argument->consume(std::next(it), end, arg_map_it->first);
+      } else if (const auto &compound_arg = current_argument;
+                 compound_arg.size() > 1 &&
+                 is_valid_prefix_char(compound_arg[0]) &&
+                 !is_valid_prefix_char(compound_arg[1])) {
+        ++it;
+        for (std::size_t j = 1; j < compound_arg.size(); j++) {
+          auto hypothetical_arg = std::string{'-', compound_arg[j]};
+          auto arg_map_it2 = m_argument_map.find(hypothetical_arg);
+          if (arg_map_it2 != m_argument_map.end()) {
+            auto argument = arg_map_it2->second;
+            it = argument->consume(it, end, arg_map_it2->first);
+          } else {
+            unknown_arguments.push_back(current_argument);
+            break;
+          }
+        }
+      } else {
+        // current argument is an optional-like argument that is unknown
+        // save it and move to next argument
+        unknown_arguments.push_back(current_argument);
+        ++it;
+      }
+    }
+    m_is_parsed = true;
+    return unknown_arguments;
+  }
+
+  // Used by print_help.
+  std::size_t get_length_of_longest_argument() const {
+    if (m_argument_map.empty()) {
+      return 0;
+    }
+    std::size_t max_size = 0;
+    for ([[maybe_unused]] const auto &[unused, argument] : m_argument_map) {
+      max_size =
+          std::max<std::size_t>(max_size, argument->get_arguments_length());
+    }
+    for ([[maybe_unused]] const auto &[command, unused] : m_subparser_map) {
+      max_size = std::max<std::size_t>(max_size, command.size());
+    }
+    return max_size;
+  }
+
+  using argument_it = std::list<Argument>::iterator;
+  using mutex_group_it = std::vector<MutuallyExclusiveGroup>::iterator;
+  using argument_parser_it =
+      std::list<std::reference_wrapper<ArgumentParser>>::iterator;
+
+  void index_argument(argument_it it) {
+    for (const auto &name : std::as_const(it->m_names)) {
+      m_argument_map.insert_or_assign(name, it);
+    }
+  }
+
+  std::string m_program_name;
+  std::string m_version;
+  std::string m_description;
+  std::string m_epilog;
+  bool m_exit_on_default_arguments = true;
+  std::string m_prefix_chars{"-"};
+  std::string m_assign_chars{"="};
+  bool m_is_parsed = false;
+  std::list<Argument> m_positional_arguments;
+  std::list<Argument> m_optional_arguments;
+  std::map<std::string, argument_it> m_argument_map;
+  std::string m_parser_path;
+  std::list<std::reference_wrapper<ArgumentParser>> m_subparsers;
+  std::map<std::string, argument_parser_it> m_subparser_map;
+  std::map<std::string, bool> m_subparser_used;
+  std::vector<MutuallyExclusiveGroup> m_mutually_exclusive_groups;
+  bool m_suppress = false;
+  std::size_t m_usage_max_line_width = (std::numeric_limits<std::size_t>::max)();
+  bool m_usage_break_on_mutex = false;
+  int m_usage_newline_counter = 0;
+  std::vector<std::string> m_group_names;
+};
+
+} // namespace argparse
\ No newline at end of file
diff --git a/ext/klib/.gitignore b/ext/klib/.gitignore
new file mode 100644
index 0000000..010a8eb
--- /dev/null
+++ b/ext/klib/.gitignore
@@ -0,0 +1,40 @@
+# General
+*.a
+*.dSYM/
+*.la
+*.lo
+*.o
+*.opensdf
+*.orig
+*.sdf
+*.suo
+*.swp
+*.tests
+*.vcxproj.filters
+*.vcxproj.user
+*~
+.git
+TAGS
+
+# Mac/Xcode-specfic
+xcuserdata
+contents.xcworkspacedata
+.DS_Store
+._*
+
+# Test byproducts
+test/kbtree_test
+test/khash_keith
+test/khash_keith2
+test/khash_test
+test/klist_test
+test/kmin_test
+test/kseq_bench
+test/kseq_bench2
+test/kseq_test
+test/ksort_test
+test/ksort_test-stl
+test/kstring_bench
+test/kstring_bench2
+test/kstring_test
+test/kvec_test
diff --git a/ext/klib/LICENSE.txt b/ext/klib/LICENSE.txt
new file mode 100644
index 0000000..0a996fa
--- /dev/null
+++ b/ext/klib/LICENSE.txt
@@ -0,0 +1,23 @@
+The MIT License
+
+Copyright (c) 2008-     Attractive Chaos <attractor@live.co.uk>
+
+Permission is hereby granted, free of charge, to any person obtaining
+a copy of this software and associated documentation files (the
+"Software"), to deal in the Software without restriction, including
+without limitation the rights to use, copy, modify, merge, publish,
+distribute, sublicense, and/or sell copies of the Software, and to
+permit persons to whom the Software is furnished to do so, subject to
+the following conditions:
+
+The above copyright notice and this permission notice shall be
+included in all copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
diff --git a/ext/klib/README.md b/ext/klib/README.md
new file mode 100644
index 0000000..7c7e216
--- /dev/null
+++ b/ext/klib/README.md
@@ -0,0 +1,243 @@
+# Klib: a Generic Library in C
+
+## <a name="overview"></a>Overview
+
+Klib is a standalone and lightweight C library distributed under [MIT/X11
+license][1]. Most components are independent of external libraries, except the
+standard C library, and independent of each other. To use a component of this
+library, you only need to copy a couple of files to your source code tree
+without worrying about library dependencies.
+
+Klib strives for efficiency and a small memory footprint. Some components, such
+as khash.h, kbtree.h, ksort.h and kvec.h, are among the most efficient
+implementations of similar algorithms or data structures in all programming
+languages, in terms of both speed and memory use.
+
+A new documentation is available [here](http://attractivechaos.github.io/klib/)
+which includes most information in this README file.
+
+#### Common components
+
+* [khash.h][khash]: generic [hash table][2] with open addressing.
+* [kbtree.h][kbtree]: generic search tree based on [B-tree][3].
+* [kavl.h][kavl]: generic intrusive [AVL tree][wiki-avl].
+* [ksort.h][ksort]: generic sort, including [introsort][4], [merge sort][5], [heap sort][6], [comb sort][7], [Knuth shuffle][8] and the [k-small][9] algorithm.
+* [kseq.h][kseq]: generic stream buffer and a [FASTA][10]/[FASTQ][11] format parser.
+* kvec.h: generic dynamic array.
+* klist.h: generic single-linked list and [memory pool][12].
+* kstring.{h,c}: basic string library.
+* kmath.{h,c}: numerical routines including [MT19937-64][13] [pseudorandom generator][14], basic [nonlinear programming][15] and a few special math functions.
+* [ketopt.h][ketopt]: portable command-line argument parser with getopt\_long-like API.
+
+#### Components for more specific use cases
+
+* ksa.c: constructing [suffix arrays][16] for strings with multiple sentinels, based on a revised [SAIS algorithm][17].
+* knetfile.{h,c}: random access to remote files on HTTP or FTP.
+* kopen.c: smart stream opening.
+* khmm.{h,c}: basic [HMM][18] library.
+* ksw.(h,c}: Striped [Smith-Waterman algorithm][19].
+* knhx.{h,c}: [Newick tree format][20] parser.
+
+
+## <a name="methodology"></a>Methodology
+
+For the implementation of generic [containers][21], klib extensively uses C
+macros. To use these data structures, we usually need to instantiate methods by
+expanding a long macro. This makes the source code look unusual or even ugly
+and adds difficulty to debugging. Unfortunately, for efficient generic
+programming in C that lacks [template][22], using macros is the only
+solution. Only with macros, we can write a generic container which, once
+instantiated, compete with a type-specific container in efficiency. Some
+generic libraries in C, such as [Glib][23], use the `void*` type to implement
+containers. These implementations are usually slower and use more memory than
+klib (see [this benchmark][31]).
+
+To effectively use klib, it is important to understand how it achieves generic
+programming. We will use the hash table library as an example:
+
+    #include "khash.h"
+    KHASH_MAP_INIT_INT(m32, char)        // instantiate structs and methods
+    int main() {
+        int ret, is_missing;
+        khint_t k;
+        khash_t(m32) *h = kh_init(m32);  // allocate a hash table
+        k = kh_put(m32, h, 5, &ret);     // insert a key to the hash table
+        if (!ret) kh_del(m32, h, k);
+        kh_value(h, k) = 10;             // set the value
+        k = kh_get(m32, h, 10);          // query the hash table
+        is_missing = (k == kh_end(h));   // test if the key is present
+        k = kh_get(m32, h, 5);
+        kh_del(m32, h, k);               // remove a key-value pair
+        for (k = kh_begin(h); k != kh_end(h); ++k)  // traverse
+            if (kh_exist(h, k))          // test if a bucket contains data
+    			kh_value(h, k) = 1;
+        kh_destroy(m32, h);              // deallocate the hash table
+        return 0;
+    }
+
+In this example, the second line instantiates a hash table with `unsigned` as
+the key type and `char` as the value type. `m32` names such a type of hash table.
+All types and functions associated with this name are macros, which will be
+explained later. Macro `kh_init()` initiates a hash table and `kh_destroy()`
+frees it. `kh_put()` inserts a key and returns the iterator (or the position)
+in the hash table. `kh_get()` and `kh_del()` get a key and delete an element,
+respectively. Macro `kh_exist()` tests if an iterator (or a position) is filled
+with data.
+
+An immediate question is this piece of code does not look like a valid C
+program (e.g. lacking semicolon, assignment to an _apparent_ function call and
+_apparent_ undefined `m32` 'variable'). To understand why the code is correct,
+let's go a bit further into the source code of `khash.h`, whose skeleton looks
+like:
+
+    #define KHASH_INIT(name, SCOPE, key_t, val_t, is_map, _hashf, _hasheq) \
+      typedef struct { \
+        int n_buckets, size, n_occupied, upper_bound; \
+        unsigned *flags; \
+        key_t *keys; \
+        val_t *vals; \
+      } kh_##name##_t; \
+      SCOPE inline kh_##name##_t *init_##name() { \
+        return (kh_##name##_t*)calloc(1, sizeof(kh_##name##_t)); \
+      } \
+      SCOPE inline int get_##name(kh_##name##_t *h, key_t k) \
+      ... \
+      SCOPE inline void destroy_##name(kh_##name##_t *h) { \
+        if (h) { \
+          free(h->keys); free(h->flags); free(h->vals); free(h); \
+        } \
+      }
+    
+    #define _int_hf(key) (unsigned)(key)
+    #define _int_heq(a, b) (a == b)
+    #define khash_t(name) kh_##name##_t
+    #define kh_value(h, k) ((h)->vals[k])
+    #define kh_begin(h, k) 0
+    #define kh_end(h) ((h)->n_buckets)
+    #define kh_init(name) init_##name()
+    #define kh_get(name, h, k) get_##name(h, k)
+    #define kh_destroy(name, h) destroy_##name(h)
+    ...
+    #define KHASH_MAP_INIT_INT(name, val_t) \
+    	KHASH_INIT(name, static, unsigned, val_t, is_map, _int_hf, _int_heq)
+
+`KHASH_INIT()` is a huge macro defining all the structs and methods. When this
+macro is called, all the code inside it will be inserted by the [C
+preprocess][37] to the place where it is called. If the macro is called
+multiple times, multiple copies of the code will be inserted. To avoid naming
+conflict of hash tables with different key-value types, the library uses [token
+concatenation][36], which is a preprocessor feature whereby we can substitute
+part of a symbol based on the parameter of the macro. In the end, the C
+preprocessor will generate the following code and feed it to the compiler
+(macro `kh_exist(h,k)` is a little complex and not expanded for simplicity):
+
+    typedef struct {
+      int n_buckets, size, n_occupied, upper_bound;
+      unsigned *flags;
+      unsigned *keys;
+      char *vals;
+    } kh_m32_t;
+    static inline kh_m32_t *init_m32() {
+      return (kh_m32_t*)calloc(1, sizeof(kh_m32_t));
+    }
+    static inline int get_m32(kh_m32_t *h, unsigned k)
+    ...
+    static inline void destroy_m32(kh_m32_t *h) {
+      if (h) {
+        free(h->keys); free(h->flags); free(h->vals); free(h);
+      }
+    }
+
+	int main() {
+		int ret, is_missing;
+		khint_t k;
+		kh_m32_t *h = init_m32();
+		k = put_m32(h, 5, &ret);
+		if (!ret) del_m32(h, k);
+		h->vals[k] = 10;
+		k = get_m32(h, 10);
+		is_missing = (k == h->n_buckets);
+		k = get_m32(h, 5);
+		del_m32(h, k);
+		for (k = 0; k != h->n_buckets; ++k)
+			if (kh_exist(h, k)) h->vals[k] = 1;
+		destroy_m32(h);
+		return 0;
+	}
+
+This is the C program we know.
+
+From this example, we can see that macros and the C preprocessor plays a key
+role in klib. Klib is fast partly because the compiler knows the key-value
+type at the compile time and is able to optimize the code to the same level
+as type-specific code. A generic library written with `void*` will not get such
+performance boost.
+
+Massively inserting code upon instantiation may remind us of C++'s slow
+compiling speed and huge binary size when STL/boost is in use. Klib is much
+better in this respect due to its small code size and component independency.
+Inserting several hundreds lines of code won't make compiling obviously slower.
+
+## <a name="resources"></a>Resources
+
+* Library documentation, if present, is available in the header files. Examples
+can be found in the [test/][24] directory.
+* **Obsolete** documentation of the hash table library can be found at
+[SourceForge][25]. This README is partly adapted from the old documentation.
+* [Blog post][26] describing the hash table library.
+* [Blog post][27] on why using `void*` for generic programming may be inefficient.
+* [Blog post][28] on the generic stream buffer.
+* [Blog post][29] evaluating the performance of `kvec.h`.
+* [Blog post][30] arguing B-tree may be a better data structure than a binary search tree.
+* [Blog post][31] evaluating the performance of `khash.h` and `kbtree.h` among many other implementations.
+[An older version][33] of the benchmark is also available.
+* [Blog post][34] benchmarking internal sorting algorithms and implementations.
+* [Blog post][32] on the k-small algorithm.
+* [Blog post][35] on the Hooke-Jeeve's algorithm for nonlinear programming.
+
+[1]: http://en.wikipedia.org/wiki/MIT_License
+[2]: https://en.wikipedia.org/wiki/Hash_table
+[3]: http://en.wikipedia.org/wiki/B-tree
+[4]: http://en.wikipedia.org/wiki/Introsort
+[5]: http://en.wikipedia.org/wiki/Merge_sort
+[6]: http://en.wikipedia.org/wiki/Heapsort
+[7]: http://en.wikipedia.org/wiki/Comb_sort
+[8]: http://en.wikipedia.org/wiki/Fisher-Yates_shuffle
+[9]: http://en.wikipedia.org/wiki/Selection_algorithm
+[10]: http://en.wikipedia.org/wiki/FASTA_format
+[11]: http://en.wikipedia.org/wiki/FASTQ_format
+[12]: http://en.wikipedia.org/wiki/Memory_pool
+[13]: http://en.wikipedia.org/wiki/Mersenne_twister
+[14]: http://en.wikipedia.org/wiki/Pseudorandom_generator
+[15]: http://en.wikipedia.org/wiki/Nonlinear_programming
+[16]: http://en.wikipedia.org/wiki/Suffix_array
+[17]: https://sites.google.com/site/yuta256/sais
+[18]: http://en.wikipedia.org/wiki/Hidden_Markov_model
+[19]: http://en.wikipedia.org/wiki/Smith-Waterman_algorithm
+[20]: http://en.wikipedia.org/wiki/Newick_format
+[21]: http://en.wikipedia.org/wiki/Container_(abstract_data_type)
+[22]: http://en.wikipedia.org/wiki/Template_(C%2B%2B)
+[23]: http://en.wikipedia.org/wiki/GLib
+[24]: https://github.com/attractivechaos/klib/tree/master/test
+[25]: http://klib.sourceforge.net/
+[26]: http://attractivechaos.wordpress.com/2008/09/02/implementing-generic-hash-library-in-c/
+[27]: http://attractivechaos.wordpress.com/2008/10/02/using-void-in-generic-c-programming-may-be-inefficient/
+[28]: http://attractivechaos.wordpress.com/2008/10/11/a-generic-buffered-stream-wrapper/
+[29]: http://attractivechaos.wordpress.com/2008/09/19/c-array-vs-c-vector/
+[30]: http://attractivechaos.wordpress.com/2008/09/24/b-tree-vs-binary-search-tree/
+[31]: http://attractivechaos.wordpress.com/2008/10/07/another-look-at-my-old-benchmark/
+[32]: http://attractivechaos.wordpress.com/2008/09/13/calculating-median/
+[33]: http://attractivechaos.wordpress.com/2008/08/28/comparison-of-hash-table-libraries/
+[34]: http://attractivechaos.wordpress.com/2008/08/28/comparison-of-internal-sorting-algorithms/
+[35]: http://attractivechaos.wordpress.com/2008/08/24/derivative-free-optimization-dfo/
+[36]: http://en.wikipedia.org/wiki/C_preprocessor#Token_concatenation
+[37]: http://en.wikipedia.org/wiki/C_preprocessor
+
+[wiki-avl]: https://en.wikipedia.org/wiki/AVL_tree
+
+[kbtree]: http://attractivechaos.github.io/klib/#KBtree%3A%20generic%20ordered%20map:%5B%5BKBtree%3A%20generic%20ordered%20map%5D%5D
+[khash]: http://attractivechaos.github.io/klib/#Khash%3A%20generic%20hash%20table:%5B%5BKhash%3A%20generic%20hash%20table%5D%5D
+[kseq]: http://attractivechaos.github.io/klib/#Kseq%3A%20stream%20buffer%20and%20FASTA%2FQ%20parser:%5B%5BKseq%3A%20stream%20buffer%20and%20FASTA%2FQ%20parser%5D%5D
+[ksort]: http://attractivechaos.github.io/klib/#Ksort%3A%20sorting%2C%20shuffling%2C%20heap%20and%20k-small:%5B%5BKsort%3A%20sorting%2C%20shuffling%2C%20heap%20and%20k-small%5D%5D
+[kavl]: http://attractivechaos.github.io/klib/#KAVL%3A%20generic%20intrusive%20AVL%20tree
+[ketopt]: http://attractivechaos.github.io/klib/#Ketopt%3A%20parsing%20command-line%20arguments
diff --git a/ext/klib/bgzf.c b/ext/klib/bgzf.c
new file mode 100644
index 0000000..9833414
--- /dev/null
+++ b/ext/klib/bgzf.c
@@ -0,0 +1,555 @@
+/* The MIT License
+
+   Copyright (c) 2008 Broad Institute / Massachusetts Institute of Technology
+                 2011 Attractive Chaos <attractor@live.co.uk>
+
+   Permission is hereby granted, free of charge, to any person obtaining a copy
+   of this software and associated documentation files (the "Software"), to deal
+   in the Software without restriction, including without limitation the rights
+   to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+   copies of the Software, and to permit persons to whom the Software is
+   furnished to do so, subject to the following conditions:
+
+   The above copyright notice and this permission notice shall be included in
+   all copies or substantial portions of the Software.
+
+   THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+   IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+   FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+   AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+   LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+   OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+   THE SOFTWARE.
+*/
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <unistd.h>
+#include <assert.h>
+#include <sys/types.h>
+#include "bgzf.h"
+
+#ifdef _USE_KNETFILE
+#include "knetfile.h"
+typedef knetFile *_bgzf_file_t;
+#define _bgzf_open(fn, mode) knet_open(fn, mode)
+#define _bgzf_dopen(fp, mode) knet_dopen(fp, mode)
+#define _bgzf_close(fp) knet_close(fp)
+#define _bgzf_fileno(fp) ((fp)->fd)
+#define _bgzf_tell(fp) knet_tell(fp)
+#define _bgzf_seek(fp, offset, whence) knet_seek(fp, offset, whence)
+#define _bgzf_read(fp, buf, len) knet_read(fp, buf, len)
+#define _bgzf_write(fp, buf, len) knet_write(fp, buf, len)
+#else // ~defined(_USE_KNETFILE)
+#if defined(_WIN32) || defined(_MSC_VER)
+#define ftello(fp) ftell(fp)
+#define fseeko(fp, offset, whence) fseek(fp, offset, whence)
+#else // ~defined(_WIN32)
+extern off_t ftello(FILE *stream);
+extern int fseeko(FILE *stream, off_t offset, int whence);
+#endif // ~defined(_WIN32)
+typedef FILE *_bgzf_file_t;
+#define _bgzf_open(fn, mode) fopen(fn, mode)
+#define _bgzf_dopen(fp, mode) fdopen(fp, mode)
+#define _bgzf_close(fp) fclose(fp)
+#define _bgzf_fileno(fp) fileno(fp)
+#define _bgzf_tell(fp) ftello(fp)
+#define _bgzf_seek(fp, offset, whence) fseeko(fp, offset, whence)
+#define _bgzf_read(fp, buf, len) fread(buf, 1, len, fp)
+#define _bgzf_write(fp, buf, len) fwrite(buf, 1, len, fp)
+#endif // ~define(_USE_KNETFILE)
+
+#define BLOCK_HEADER_LENGTH 18
+#define BLOCK_FOOTER_LENGTH 8
+
+/* BGZF/GZIP header (speciallized from RFC 1952; little endian):
+ +---+---+---+---+---+---+---+---+---+---+---+---+---+---+---+---+---+---+
+ | 31|139|  8|  4|              0|  0|255|      6| 66| 67|      2|BLK_LEN|
+ +---+---+---+---+---+---+---+---+---+---+---+---+---+---+---+---+---+---+
+*/
+static const uint8_t g_magic[19] = "\037\213\010\4\0\0\0\0\0\377\6\0\102\103\2\0\0\0";
+
+#ifdef BGZF_CACHE
+typedef struct {
+	int size;
+	uint8_t *block;
+	int64_t end_offset;
+} cache_t;
+#include "khash.h"
+KHASH_MAP_INIT_INT64(cache, cache_t)
+#endif
+
+static inline void packInt16(uint8_t *buffer, uint16_t value)
+{
+	buffer[0] = value;
+	buffer[1] = value >> 8;
+}
+
+static inline int unpackInt16(const uint8_t *buffer)
+{
+	return buffer[0] | buffer[1] << 8;
+}
+
+static inline void packInt32(uint8_t *buffer, uint32_t value)
+{
+	buffer[0] = value;
+	buffer[1] = value >> 8;
+	buffer[2] = value >> 16;
+	buffer[3] = value >> 24;
+}
+
+static BGZF *bgzf_read_init()
+{
+	BGZF *fp;
+	fp = calloc(1, sizeof(BGZF));
+	fp->open_mode = 'r';
+	fp->uncompressed_block = malloc(BGZF_MAX_BLOCK_SIZE);
+	fp->compressed_block = malloc(BGZF_MAX_BLOCK_SIZE);
+#ifdef BGZF_CACHE
+	fp->cache = kh_init(cache);
+#endif
+	return fp;
+}
+
+static BGZF *bgzf_write_init(int compress_level) // compress_level==-1 for the default level
+{
+	BGZF *fp;
+	fp = calloc(1, sizeof(BGZF));
+	fp->open_mode = 'w';
+	fp->uncompressed_block = malloc(BGZF_MAX_BLOCK_SIZE);
+	fp->compressed_block = malloc(BGZF_MAX_BLOCK_SIZE);
+	fp->compress_level = compress_level < 0? Z_DEFAULT_COMPRESSION : compress_level; // Z_DEFAULT_COMPRESSION==-1
+	if (fp->compress_level > 9) fp->compress_level = Z_DEFAULT_COMPRESSION;
+	return fp;
+}
+// get the compress level from the mode string
+static int mode2level(const char *__restrict mode)
+{
+	int i, compress_level = -1;
+	for (i = 0; mode[i]; ++i)
+		if (mode[i] >= '0' && mode[i] <= '9') break;
+	if (mode[i]) compress_level = (int)mode[i] - '0';
+	if (strchr(mode, 'u')) compress_level = 0;
+	return compress_level;
+}
+
+BGZF *bgzf_open(const char *path, const char *mode)
+{
+	BGZF *fp = 0;
+	if (strchr(mode, 'r') || strchr(mode, 'R')) {
+		_bgzf_file_t fpr;
+		if ((fpr = _bgzf_open(path, "r")) == 0) return 0;
+		fp = bgzf_read_init();
+		fp->fp = fpr;
+	} else if (strchr(mode, 'w') || strchr(mode, 'W')) {
+		FILE *fpw;
+		if ((fpw = fopen(path, "w")) == 0) return 0;
+		fp = bgzf_write_init(mode2level(mode));
+		fp->fp = fpw;
+	}
+	return fp;
+}
+
+BGZF *bgzf_dopen(int fd, const char *mode)
+{
+	BGZF *fp = 0;
+	if (strchr(mode, 'r') || strchr(mode, 'R')) {
+		_bgzf_file_t fpr;
+		if ((fpr = _bgzf_dopen(fd, "r")) == 0) return 0;
+		fp = bgzf_read_init();
+		fp->fp = fpr;
+	} else if (strchr(mode, 'w') || strchr(mode, 'W')) {
+		FILE *fpw;
+		if ((fpw = fdopen(fd, "w")) == 0) return 0;
+		fp = bgzf_write_init(mode2level(mode));
+		fp->fp = fpw;
+	}
+	return fp;
+}
+
+// Deflate the block in fp->uncompressed_block into fp->compressed_block. Also adds an extra field that stores the compressed block length.
+static int deflate_block(BGZF *fp, int block_length)
+{
+	uint8_t *buffer = fp->compressed_block;
+	int buffer_size = BGZF_BLOCK_SIZE;
+	int input_length = block_length;
+	int compressed_length = 0;
+	int remaining;
+	uint32_t crc;
+
+	assert(block_length <= BGZF_BLOCK_SIZE); // guaranteed by the caller
+	memcpy(buffer, g_magic, BLOCK_HEADER_LENGTH); // the last two bytes are a place holder for the length of the block
+	while (1) { // loop to retry for blocks that do not compress enough
+		int status;
+		z_stream zs;
+		zs.zalloc = NULL;
+		zs.zfree = NULL;
+		zs.next_in = fp->uncompressed_block;
+		zs.avail_in = input_length;
+		zs.next_out = (void*)&buffer[BLOCK_HEADER_LENGTH];
+		zs.avail_out = buffer_size - BLOCK_HEADER_LENGTH - BLOCK_FOOTER_LENGTH;
+		status = deflateInit2(&zs, fp->compress_level, Z_DEFLATED, -15, 8, Z_DEFAULT_STRATEGY); // -15 to disable zlib header/footer
+		if (status != Z_OK) {
+			fp->errcode |= BGZF_ERR_ZLIB;
+			return -1;
+		}
+		status = deflate(&zs, Z_FINISH);
+		if (status != Z_STREAM_END) { // not compressed enough
+			deflateEnd(&zs); // reset the stream
+			if (status == Z_OK) { // reduce the size and recompress
+				input_length -= 1024;
+				assert(input_length > 0); // logically, this should not happen
+				continue;
+			}
+			fp->errcode |= BGZF_ERR_ZLIB;
+			return -1;
+		}
+		if (deflateEnd(&zs) != Z_OK) {
+			fp->errcode |= BGZF_ERR_ZLIB;
+			return -1;
+		}
+		compressed_length = zs.total_out;
+		compressed_length += BLOCK_HEADER_LENGTH + BLOCK_FOOTER_LENGTH;
+		assert(compressed_length <= BGZF_BLOCK_SIZE);
+		break;
+	}
+
+	assert(compressed_length > 0);
+	packInt16((uint8_t*)&buffer[16], compressed_length - 1); // write the compressed_length; -1 to fit 2 bytes
+	crc = crc32(0L, NULL, 0L);
+	crc = crc32(crc, fp->uncompressed_block, input_length);
+	packInt32((uint8_t*)&buffer[compressed_length-8], crc);
+	packInt32((uint8_t*)&buffer[compressed_length-4], input_length);
+
+	remaining = block_length - input_length;
+	if (remaining > 0) {
+		assert(remaining <= input_length);
+		memcpy(fp->uncompressed_block, fp->uncompressed_block + input_length, remaining);
+	}
+	fp->block_offset = remaining;
+	return compressed_length;
+}
+
+// Inflate the block in fp->compressed_block into fp->uncompressed_block
+static int inflate_block(BGZF* fp, int block_length)
+{
+	z_stream zs;
+	zs.zalloc = NULL;
+	zs.zfree = NULL;
+	zs.next_in = fp->compressed_block + 18;
+	zs.avail_in = block_length - 16;
+	zs.next_out = fp->uncompressed_block;
+	zs.avail_out = BGZF_BLOCK_SIZE;
+
+	if (inflateInit2(&zs, -15) != Z_OK) {
+		fp->errcode |= BGZF_ERR_ZLIB;
+		return -1;
+	}
+	if (inflate(&zs, Z_FINISH) != Z_STREAM_END) {
+		inflateEnd(&zs);
+		fp->errcode |= BGZF_ERR_ZLIB;
+		return -1;
+	}
+	if (inflateEnd(&zs) != Z_OK) {
+		fp->errcode |= BGZF_ERR_ZLIB;
+		return -1;
+	}
+	return zs.total_out;
+}
+
+static int check_header(const uint8_t *header)
+{
+	return (header[0] == 31 && header[1] == 139 && header[2] == 8 && (header[3] & 4) != 0
+			&& unpackInt16((uint8_t*)&header[10]) == 6
+			&& header[12] == 'B' && header[13] == 'C'
+			&& unpackInt16((uint8_t*)&header[14]) == 2);
+}
+
+#ifdef BGZF_CACHE
+static void free_cache(BGZF *fp)
+{
+	khint_t k;
+	khash_t(cache) *h = (khash_t(cache)*)fp->cache;
+	if (fp->open_mode != 'r') return;
+	for (k = kh_begin(h); k < kh_end(h); ++k)
+		if (kh_exist(h, k)) free(kh_val(h, k).block);
+	kh_destroy(cache, h);
+}
+
+static int load_block_from_cache(BGZF *fp, int64_t block_address)
+{
+	khint_t k;
+	cache_t *p;
+	khash_t(cache) *h = (khash_t(cache)*)fp->cache;
+	k = kh_get(cache, h, block_address);
+	if (k == kh_end(h)) return 0;
+	p = &kh_val(h, k);
+	if (fp->block_length != 0) fp->block_offset = 0;
+	fp->block_address = block_address;
+	fp->block_length = p->size;
+	memcpy(fp->uncompressed_block, p->block, BGZF_BLOCK_SIZE);
+	_bgzf_seek((_bgzf_file_t)fp->fp, p->end_offset, SEEK_SET);
+	return p->size;
+}
+
+static void cache_block(BGZF *fp, int size)
+{
+	int ret;
+	khint_t k;
+	cache_t *p;
+	khash_t(cache) *h = (khash_t(cache)*)fp->cache;
+	if (BGZF_BLOCK_SIZE >= fp->cache_size) return;
+	if ((kh_size(h) + 1) * BGZF_BLOCK_SIZE > fp->cache_size) {
+		/* A better way would be to remove the oldest block in the
+		 * cache, but here we remove a random one for simplicity. This
+		 * should not have a big impact on performance. */
+		for (k = kh_begin(h); k < kh_end(h); ++k)
+			if (kh_exist(h, k)) break;
+		if (k < kh_end(h)) {
+			free(kh_val(h, k).block);
+			kh_del(cache, h, k);
+		}
+	}
+	k = kh_put(cache, h, fp->block_address, &ret);
+	if (ret == 0) return; // if this happens, a bug!
+	p = &kh_val(h, k);
+	p->size = fp->block_length;
+	p->end_offset = fp->block_address + size;
+	p->block = malloc(BGZF_BLOCK_SIZE);
+	memcpy(kh_val(h, k).block, fp->uncompressed_block, BGZF_BLOCK_SIZE);
+}
+#else
+static void free_cache(BGZF *fp) {}
+static int load_block_from_cache(BGZF *fp, int64_t block_address) {return 0;}
+static void cache_block(BGZF *fp, int size) {}
+#endif
+
+int bgzf_read_block(BGZF *fp)
+{
+	uint8_t header[BLOCK_HEADER_LENGTH], *compressed_block;
+	int count, size = 0, block_length, remaining;
+	int64_t block_address;
+	block_address = _bgzf_tell((_bgzf_file_t)fp->fp);
+	if (load_block_from_cache(fp, block_address)) return 0;
+	count = _bgzf_read(fp->fp, header, sizeof(header));
+	if (count == 0) { // no data read
+		fp->block_length = 0;
+		return 0;
+	}
+	if (count != sizeof(header) || !check_header(header)) {
+		fp->errcode |= BGZF_ERR_HEADER;
+		return -1;
+	}
+	size = count;
+	block_length = unpackInt16((uint8_t*)&header[16]) + 1; // +1 because when writing this number, we used "-1"
+	compressed_block = (uint8_t*)fp->compressed_block;
+	memcpy(compressed_block, header, BLOCK_HEADER_LENGTH);
+	remaining = block_length - BLOCK_HEADER_LENGTH;
+	count = _bgzf_read(fp->fp, &compressed_block[BLOCK_HEADER_LENGTH], remaining);
+	if (count != remaining) {
+		fp->errcode |= BGZF_ERR_IO;
+		return -1;
+	}
+	size += count;
+	if ((count = inflate_block(fp, block_length)) < 0) return -1;
+	if (fp->block_length != 0) fp->block_offset = 0; // Do not reset offset if this read follows a seek.
+	fp->block_address = block_address;
+	fp->block_length = count;
+	cache_block(fp, size);
+	return 0;
+}
+
+ssize_t bgzf_read(BGZF *fp, void *data, ssize_t length)
+{
+	ssize_t bytes_read = 0;
+	uint8_t *output = data;
+	if (length <= 0) return 0;
+	assert(fp->open_mode == 'r');
+	while (bytes_read < length) {
+		int copy_length, available = fp->block_length - fp->block_offset;
+		uint8_t *buffer;
+		if (available <= 0) {
+			if (bgzf_read_block(fp) != 0) return -1;
+			available = fp->block_length - fp->block_offset;
+			if (available <= 0) break;
+		}
+		copy_length = length - bytes_read < available? length - bytes_read : available;
+		buffer = fp->uncompressed_block;
+		memcpy(output, buffer + fp->block_offset, copy_length);
+		fp->block_offset += copy_length;
+		output += copy_length;
+		bytes_read += copy_length;
+	}
+	if (fp->block_offset == fp->block_length) {
+		fp->block_address = _bgzf_tell((_bgzf_file_t)fp->fp);
+		fp->block_offset = fp->block_length = 0;
+	}
+	return bytes_read;
+}
+
+int bgzf_flush(BGZF *fp)
+{
+	assert(fp->open_mode == 'w');
+	while (fp->block_offset > 0) {
+		int block_length;
+		block_length = deflate_block(fp, fp->block_offset);
+		if (block_length < 0) return -1;
+		if (fwrite(fp->compressed_block, 1, block_length, fp->fp) != block_length) {
+			fp->errcode |= BGZF_ERR_IO; // possibly truncated file
+			return -1;
+		}
+		fp->block_address += block_length;
+	}
+	return 0;
+}
+
+int bgzf_flush_try(BGZF *fp, ssize_t size)
+{
+	if (fp->block_offset + size > BGZF_BLOCK_SIZE)
+		return bgzf_flush(fp);
+	return -1;
+}
+
+ssize_t bgzf_write(BGZF *fp, const void *data, ssize_t length)
+{
+	const uint8_t *input = data;
+	int block_length = BGZF_BLOCK_SIZE, bytes_written;
+	assert(fp->open_mode == 'w');
+	input = data;
+	bytes_written = 0;
+	while (bytes_written < length) {
+		uint8_t* buffer = fp->uncompressed_block;
+		int copy_length = block_length - fp->block_offset < length - bytes_written? block_length - fp->block_offset : length - bytes_written;
+		memcpy(buffer + fp->block_offset, input, copy_length);
+		fp->block_offset += copy_length;
+		input += copy_length;
+		bytes_written += copy_length;
+		if (fp->block_offset == block_length && bgzf_flush(fp)) break;
+	}
+	return bytes_written;
+}
+
+int bgzf_close(BGZF* fp)
+{
+	int ret, count, block_length;
+	if (fp == 0) return -1;
+	if (fp->open_mode == 'w') {
+		if (bgzf_flush(fp) != 0) return -1;
+		block_length = deflate_block(fp, 0); // write an empty block
+		count = fwrite(fp->compressed_block, 1, block_length, fp->fp);
+		if (fflush(fp->fp) != 0) {
+			fp->errcode |= BGZF_ERR_IO;
+			return -1;
+		}
+	}
+	ret = fp->open_mode == 'w'? fclose(fp->fp) : _bgzf_close(fp->fp);
+	if (ret != 0) return -1;
+	free(fp->uncompressed_block);
+	free(fp->compressed_block);
+	free_cache(fp);
+	free(fp);
+	return 0;
+}
+
+void bgzf_set_cache_size(BGZF *fp, int cache_size)
+{
+	if (fp) fp->cache_size = cache_size;
+}
+
+int bgzf_check_EOF(BGZF *fp)
+{
+	static uint8_t magic[28] = "\037\213\010\4\0\0\0\0\0\377\6\0\102\103\2\0\033\0\3\0\0\0\0\0\0\0\0\0";
+	uint8_t buf[28];
+	off_t offset;
+	offset = _bgzf_tell((_bgzf_file_t)fp->fp);
+	if (_bgzf_seek(fp->fp, -28, SEEK_END) < 0) return 0;
+	_bgzf_read(fp->fp, buf, 28);
+	_bgzf_seek(fp->fp, offset, SEEK_SET);
+	return (memcmp(magic, buf, 28) == 0)? 1 : 0;
+}
+
+int64_t bgzf_seek(BGZF* fp, int64_t pos, int where)
+{
+	int block_offset;
+	int64_t block_address;
+
+	if (fp->open_mode != 'r' || where != SEEK_SET) {
+		fp->errcode |= BGZF_ERR_MISUSE;
+		return -1;
+	}
+	block_offset = pos & 0xFFFF;
+	block_address = pos >> 16;
+	if (_bgzf_seek(fp->fp, block_address, SEEK_SET) < 0) {
+		fp->errcode |= BGZF_ERR_IO;
+		return -1;
+	}
+	fp->block_length = 0;  // indicates current block has not been loaded
+	fp->block_address = block_address;
+	fp->block_offset = block_offset;
+	return 0;
+}
+
+int bgzf_is_bgzf(const char *fn)
+{
+	uint8_t buf[16];
+	int n;
+	_bgzf_file_t fp;
+	if ((fp = _bgzf_open(fn, "r")) == 0) return 0;
+	n = _bgzf_read(fp, buf, 16);
+	_bgzf_close(fp);
+	if (n != 16) return 0;
+	return memcmp(g_magic, buf, 16) == 0? 1 : 0;
+}
+
+int bgzf_getc(BGZF *fp)
+{
+	int c;
+	if (fp->block_offset >= fp->block_length) {
+		if (bgzf_read_block(fp) != 0) return -2; /* error */
+		if (fp->block_length == 0) return -1; /* end-of-file */
+	}
+	c = ((unsigned char*)fp->uncompressed_block)[fp->block_offset++];
+    if (fp->block_offset == fp->block_length) {
+        fp->block_address = _bgzf_tell((_bgzf_file_t)fp->fp);
+        fp->block_offset = 0;
+        fp->block_length = 0;
+    }
+	return c;
+}
+
+#ifndef kroundup32
+#define kroundup32(x) (--(x), (x)|=(x)>>1, (x)|=(x)>>2, (x)|=(x)>>4, (x)|=(x)>>8, (x)|=(x)>>16, ++(x))
+#endif
+
+int bgzf_getline(BGZF *fp, int delim, kstring_t *str)
+{
+	int l, state = 0;
+	unsigned char *buf = (unsigned char*)fp->uncompressed_block;
+	str->l = 0;
+	do {
+		if (fp->block_offset >= fp->block_length) {
+			if (bgzf_read_block(fp) != 0) { state = -2; break; }
+			if (fp->block_length == 0) { state = -1; break; }
+		}
+		for (l = fp->block_offset; l < fp->block_length && buf[l] != delim; ++l);
+		if (l < fp->block_length) state = 1;
+		l -= fp->block_offset;
+		if (str->l + l + 1 >= str->m) {
+			str->m = str->l + l + 2;
+			kroundup32(str->m);
+			str->s = (char*)realloc(str->s, str->m);
+		}
+		memcpy(str->s + str->l, buf + fp->block_offset, l);
+		str->l += l;
+		fp->block_offset += l + 1;
+		if (fp->block_offset >= fp->block_length) {
+			fp->block_address = _bgzf_tell((_bgzf_file_t)fp->fp);
+			fp->block_offset = 0;
+			fp->block_length = 0;
+		} 
+	} while (state == 0);
+	if (str->l == 0 && state < 0) return state;
+	str->s[str->l] = 0;
+	return str->l;
+}
diff --git a/ext/klib/bgzf.h b/ext/klib/bgzf.h
new file mode 100644
index 0000000..29fe0e5
--- /dev/null
+++ b/ext/klib/bgzf.h
@@ -0,0 +1,196 @@
+/* The MIT License
+
+   Copyright (c) 2008 Broad Institute / Massachusetts Institute of Technology
+                 2011 Attractive Chaos <attractor@live.co.uk>
+
+   Permission is hereby granted, free of charge, to any person obtaining a copy
+   of this software and associated documentation files (the "Software"), to deal
+   in the Software without restriction, including without limitation the rights
+   to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+   copies of the Software, and to permit persons to whom the Software is
+   furnished to do so, subject to the following conditions:
+
+   The above copyright notice and this permission notice shall be included in
+   all copies or substantial portions of the Software.
+
+   THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+   IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+   FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+   AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+   LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+   OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+   THE SOFTWARE.
+*/
+
+/* The BGZF library was originally written by Bob Handsaker from the Broad
+ * Institute. It was later improved by the SAMtools developers. */
+
+#ifndef __BGZF_H
+#define __BGZF_H
+
+#include <stdint.h>
+#include <stdio.h>
+#include <zlib.h>
+
+#define BGZF_BLOCK_SIZE     0x10000
+#define BGZF_MAX_BLOCK_SIZE 0x10000
+
+#define BGZF_ERR_ZLIB   1
+#define BGZF_ERR_HEADER 2
+#define BGZF_ERR_IO     4
+#define BGZF_ERR_MISUSE 8
+
+typedef struct {
+    int open_mode:8, compress_level:8, errcode:16;
+	int cache_size;
+    int block_length, block_offset;
+    int64_t block_address;
+    void *uncompressed_block, *compressed_block;
+	void *cache; // a pointer to a hash table
+	void *fp; // actual file handler; FILE* on writing; FILE* or knetFile* on reading
+} BGZF;
+
+#ifndef KSTRING_T
+#define KSTRING_T kstring_t
+typedef struct __kstring_t {
+	size_t l, m;
+	char *s;
+} kstring_t;
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+	/******************
+	 * Basic routines *
+	 ******************/
+
+	/**
+	 * Open an existing file descriptor for reading or writing.
+	 *
+	 * @param fd    file descriptor
+	 * @param mode  mode matching /[rwu0-9]+/: 'r' for reading, 'w' for writing and a digit specifies
+	 *              the zlib compression level; if both 'r' and 'w' are present, 'w' is ignored.
+	 * @return      BGZF file handler; 0 on error
+	 */
+	BGZF* bgzf_dopen(int fd, const char *mode);
+
+	#define bgzf_fdopen(fd, mode) bgzf_dopen((fd), (mode)) // for backward compatibility
+
+	/**
+	 * Open the specified file for reading or writing.
+	 */
+	BGZF* bgzf_open(const char* path, const char *mode);
+
+	/**
+	 * Close the BGZF and free all associated resources.
+	 *
+	 * @param fp    BGZF file handler
+	 * @return      0 on success and -1 on error
+	 */
+	int bgzf_close(BGZF *fp);
+
+	/**
+	 * Read up to _length_ bytes from the file storing into _data_.
+	 *
+	 * @param fp     BGZF file handler
+	 * @param data   data array to read into
+	 * @param length size of data to read
+	 * @return       number of bytes actually read; 0 on end-of-file and -1 on error
+	 */
+	ssize_t bgzf_read(BGZF *fp, void *data, ssize_t length);
+
+	/**
+	 * Write _length_ bytes from _data_ to the file.
+	 *
+	 * @param fp     BGZF file handler
+	 * @param data   data array to write
+	 * @param length size of data to write
+	 * @return       number of bytes actually written; -1 on error
+	 */
+	ssize_t bgzf_write(BGZF *fp, const void *data, ssize_t length);
+
+	/**
+	 * Write the data in the buffer to the file.
+	 */
+	int bgzf_flush(BGZF *fp);
+
+	/**
+	 * Return a virtual file pointer to the current location in the file.
+	 * No interpetation of the value should be made, other than a subsequent
+	 * call to bgzf_seek can be used to position the file at the same point.
+	 * Return value is non-negative on success.
+	 */
+	#define bgzf_tell(fp) ((fp->block_address << 16) | (fp->block_offset & 0xFFFF))
+
+	/**
+	 * Set the file to read from the location specified by _pos_.
+	 *
+	 * @param fp     BGZF file handler
+	 * @param pos    virtual file offset returned by bgzf_tell()
+	 * @param whence must be SEEK_SET
+	 * @return       0 on success and -1 on error
+	 */
+	int64_t bgzf_seek(BGZF *fp, int64_t pos, int whence);
+
+	/**
+	 * Check if the BGZF end-of-file (EOF) marker is present
+	 *
+	 * @param fp    BGZF file handler opened for reading
+	 * @return      1 if EOF is present; 0 if not or on I/O error
+	 */
+	int bgzf_check_EOF(BGZF *fp);
+
+	/**
+	 * Check if a file is in the BGZF format
+	 *
+	 * @param fn    file name
+	 * @return      1 if _fn_ is BGZF; 0 if not or on I/O error
+	 */
+	 int bgzf_is_bgzf(const char *fn);
+
+	/*********************
+	 * Advanced routines *
+	 *********************/
+
+	/**
+	 * Set the cache size. Only effective when compiled with -DBGZF_CACHE.
+	 *
+	 * @param fp    BGZF file handler
+	 * @param size  size of cache in bytes; 0 to disable caching (default)
+	 */
+	void bgzf_set_cache_size(BGZF *fp, int size);
+
+	/**
+	 * Flush the file if the remaining buffer size is smaller than _size_ 
+	 */
+	int bgzf_flush_try(BGZF *fp, ssize_t size);
+
+	/**
+	 * Read one byte from a BGZF file. It is faster than bgzf_read()
+	 * @param fp     BGZF file handler
+	 * @return       byte read; -1 on end-of-file or error
+	 */
+	int bgzf_getc(BGZF *fp);
+
+	/**
+	 * Read one line from a BGZF file. It is faster than bgzf_getc()
+	 *
+	 * @param fp     BGZF file handler
+	 * @param delim  delimitor
+	 * @param str    string to write to; must be initialized
+	 * @return       length of the string; 0 on end-of-file; negative on error
+	 */
+	int bgzf_getline(BGZF *fp, int delim, kstring_t *str);
+
+	/**
+	 * Read the next BGZF block.
+	 */
+	int bgzf_read_block(BGZF *fp);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif
diff --git a/ext/klib/cpp/kavl.hpp b/ext/klib/cpp/kavl.hpp
new file mode 100644
index 0000000..9fa497a
--- /dev/null
+++ b/ext/klib/cpp/kavl.hpp
@@ -0,0 +1,202 @@
+#ifndef KAVL_HPP
+#define KAVL_HPP
+
+#include <functional>
+
+namespace klib {
+
+template<class T, typename Less = std::less<T> >
+class Avl {
+	static const int MAX_DEPTH = 64;
+	struct Node {
+		T data;
+		signed char balance;
+		unsigned size;
+		Node *p[2];
+	};
+	Node *root;
+	inline int cmp_func(const T &x, const T &y) {
+		return Less()(y, x) - Less()(x, y);
+	}
+	inline unsigned child_size(Node *p, int dir) {
+		return p->p[dir]? p->p[dir]->size : 0;
+	};
+	// one rotation: (a,(b,c)q)p => ((a,b)p,c)q
+	inline Node *rotate1(Node *p, int dir) { // dir=0 to left; dir=1 to right
+		int opp = 1 - dir; // opposite direction
+		Node *q = p->p[opp];
+		unsigned size_p = p->size;
+		p->size -= q->size - child_size(q, dir);
+		q->size = size_p;
+		p->p[opp] = q->p[dir];
+		q->p[dir] = p;
+		return q;
+	};
+	// two consecutive rotations: (a,((b,c)r,d)q)p => ((a,b)p,(c,d)q)r
+	inline Node *rotate2(Node *p, int dir) {
+		int b1, opp = 1 - dir;
+		Node *q = p->p[opp], *r = q->p[dir];
+		unsigned size_x_dir = child_size(r, dir);
+		r->size = p->size;
+		p->size -= q->size - size_x_dir;
+		q->size -= size_x_dir + 1;
+		p->p[opp] = r->p[dir];
+		r->p[dir] = p;
+		q->p[dir] = r->p[opp];
+		r->p[opp] = q;
+		b1 = dir == 0? +1 : -1;
+		if (r->balance == b1) q->balance = 0, p->balance = -b1;
+		else if (r->balance == 0) q->balance = p->balance = 0;
+		else q->balance = b1, p->balance = 0;
+		r->balance = 0;
+		return r;
+	};
+	void destroy(Node *r) {
+		Node *p, *q;
+		for (p = r; p; p = q) {
+			if (p->p[0] == 0) {
+				q = p->p[1];
+				delete p;
+			} else {
+				q = p->p[0];
+				p->p[0] = q->p[1];
+				q->p[1] = p;
+			}
+		}
+	};
+public:
+	Avl() : root(NULL) {};
+	~Avl() { destroy(root); };
+	unsigned size() const { return root? root->size : 0; }
+	T *find(const T &data, unsigned *cnt_ = NULL) {
+		Node *p = root;
+		unsigned cnt = 0;
+		while (p != 0) {
+			int cmp = cmp_func(data, p->data);
+			if (cmp >= 0) cnt += child_size(p, 0) + 1;
+			if (cmp < 0) p = p->p[0];
+			else if (cmp > 0) p = p->p[1];
+			else break;
+		}
+		if (cnt_) *cnt_ = cnt;
+		return p? &p->data : NULL;
+	};
+	T *insert(const T &data, bool *is_new = NULL, unsigned *cnt_ = NULL) {
+		unsigned char stack[MAX_DEPTH];
+		Node *path[MAX_DEPTH];
+		Node *bp, *bq;
+		Node *x, *p, *q, *r = 0; // _r_ is potentially the new root
+		int i, which = 0, top, b1, path_len;
+		unsigned cnt = 0;
+		bp = root, bq = 0;
+		if (is_new) *is_new = false;
+		// find the insertion location
+		for (p = bp, q = bq, top = path_len = 0; p; q = p, p = p->p[which]) {
+			int cmp = cmp_func(data, p->data);
+			if (cmp >= 0) cnt += child_size(p, 0) + 1;
+			if (cmp == 0) {
+				if (cnt_) *cnt_ = cnt;
+				return &p->data;
+			}
+			if (p->balance != 0)
+				bq = q, bp = p, top = 0;
+			stack[top++] = which = (cmp > 0);
+			path[path_len++] = p;
+		}
+		if (cnt_) *cnt_ = cnt;
+		x = new Node;
+		x->data = data, x->balance = 0, x->size = 1, x->p[0] = x->p[1] = 0;
+		if (is_new) *is_new = true;
+		if (q == 0) root = x;
+		else q->p[which] = x;
+		if (bp == 0) return &x->data;
+		for (i = 0; i < path_len; ++i) ++path[i]->size;
+		for (p = bp, top = 0; p != x; p = p->p[stack[top]], ++top) /* update balance factors */
+			if (stack[top] == 0) --p->balance;
+			else ++p->balance;
+		if (bp->balance > -2 && bp->balance < 2) return &x->data; /* no re-balance needed */
+		// re-balance
+		which = (bp->balance < 0);
+		b1 = which == 0? +1 : -1;
+		q = bp->p[1 - which];
+		if (q->balance == b1) {
+			r = rotate1(bp, which);
+			q->balance = bp->balance = 0;
+		} else r = rotate2(bp, which);
+		if (bq == 0) root = r;
+		else bq->p[bp != bq->p[0]] = r;
+		return &x->data;
+	};
+	bool erase(const T &data) {
+		Node *p, *path[MAX_DEPTH], fake;
+		unsigned char dir[MAX_DEPTH];
+		int i, d = 0, cmp;
+		fake.p[0] = root, fake.p[1] = 0;
+		for (cmp = -1, p = &fake; cmp; cmp = cmp_func(data, p->data)) {
+			int which = (cmp > 0);
+			dir[d] = which;
+			path[d++] = p;
+			p = p->p[which];
+			if (p == 0) return false;
+		}
+		for (i = 1; i < d; ++i) --path[i]->size;
+		if (p->p[1] == 0) { // ((1,.)2,3)4 => (1,3)4; p=2
+			path[d-1]->p[dir[d-1]] = p->p[0];
+		} else {
+			Node *q = p->p[1];
+			if (q->p[0] == 0) { // ((1,2)3,4)5 => ((1)2,4)5; p=3
+				q->p[0] = p->p[0];
+				q->balance = p->balance;
+				path[d-1]->p[dir[d-1]] = q;
+				path[d] = q, dir[d++] = 1;
+				q->size = p->size - 1;
+			} else { // ((1,((.,2)3,4)5)6,7)8 => ((1,(2,4)5)3,7)8; p=6
+				Node *r;
+				int e = d++; // backup _d_
+				for (;;) {
+					dir[d] = 0;
+					path[d++] = q;
+					r = q->p[0];
+					if (r->p[0] == 0) break;
+					q = r;
+				}
+				r->p[0] = p->p[0];
+				q->p[0] = r->p[1];
+				r->p[1] = p->p[1];
+				r->balance = p->balance;
+				path[e-1]->p[dir[e-1]] = r;
+				path[e] = r, dir[e] = 1;
+				for (i = e + 1; i < d; ++i) --path[i]->size;
+				r->size = p->size - 1;
+			}
+		}
+		while (--d > 0) {
+			Node *q = path[d];
+			int which, other, b1 = 1, b2 = 2;
+			which = dir[d], other = 1 - which;
+			if (which) b1 = -b1, b2 = -b2;
+			q->balance += b1;
+			if (q->balance == b1) break;
+			else if (q->balance == b2) {
+				Node *r = q->p[other];
+				if (r->balance == -b1) {
+					path[d-1]->p[dir[d-1]] = rotate2(q, which);
+				} else {
+					path[d-1]->p[dir[d-1]] = rotate1(q, which);
+					if (r->balance == 0) {
+						r->balance = -b1;
+						q->balance = b1;
+						break;
+					} else r->balance = q->balance = 0;
+				}
+			}
+		}
+		root = fake.p[0];
+		delete p;
+		return true;
+	};
+};
+
+} // end of namespace klib
+
+#endif
diff --git a/ext/klib/cpp/khash.hpp b/ext/klib/cpp/khash.hpp
new file mode 100644
index 0000000..a57d6de
--- /dev/null
+++ b/ext/klib/cpp/khash.hpp
@@ -0,0 +1,163 @@
+#ifndef KHASH_HPP
+#define KHASH_HPP
+
+#include <memory>
+#include <functional>
+#include <cstdlib> // for malloc() etc
+#include <cstring> // for memset()
+
+#include <stdint.h> // for uint32_t
+
+namespace klib {
+
+#ifndef kroundup32 // FIXME: doesn't work for 64-bit integers
+#define kroundup32(x) (--(x), (x)|=(x)>>1, (x)|=(x)>>2, (x)|=(x)>>4, (x)|=(x)>>8, (x)|=(x)>>16, ++(x))
+#endif
+
+#define __ac_isempty(flag, i) ((flag[i>>4]>>((i&0xfU)<<1))&2)
+#define __ac_isdel(flag, i) ((flag[i>>4]>>((i&0xfU)<<1))&1)
+#define __ac_isempty(flag, i) ((flag[i>>4]>>((i&0xfU)<<1))&2)
+#define __ac_isdel(flag, i) ((flag[i>>4]>>((i&0xfU)<<1))&1)
+#define __ac_iseither(flag, i) ((flag[i>>4]>>((i&0xfU)<<1))&3)
+#define __ac_set_isdel_false(flag, i) (flag[i>>4]&=~(1ul<<((i&0xfU)<<1)))
+#define __ac_set_isempty_false(flag, i) (flag[i>>4]&=~(2ul<<((i&0xfU)<<1)))
+#define __ac_set_isboth_false(flag, i) (flag[i>>4]&=~(3ul<<((i&0xfU)<<1)))
+#define __ac_set_isdel_true(flag, i) (flag[i>>4]|=1ul<<((i&0xfU)<<1))
+
+#define __ac_fsize(m) ((m) < 16? 1 : (m)>>4)
+
+template<class T, class Hash, class Eq = std::equal_to<T>, typename khint_t = uint32_t>
+class KHash {
+	khint_t n_buckets, count, n_occupied, upper_bound;
+	uint32_t *flags;
+	T *keys;
+public:
+	KHash() : n_buckets(0), count(0), n_occupied(0), upper_bound(0), flags(NULL), keys(NULL) {};
+	~KHash() { std::free(flags); std::free(keys); };
+	khint_t capacity(void) const { return n_buckets; };
+	khint_t size(void) const { return count; };
+	khint_t begin(void) const { return 0; };
+	khint_t end(void) const { return n_buckets; };
+
+	void exist(khint_t x) const { return !__ac_iseither(flags, x); };
+	T &at(khint_t x) { return keys[x]; };
+
+	khint_t get(const T &key) const {
+		if (n_buckets) {
+			khint_t k, i, last, mask, step = 0;
+			mask = n_buckets - 1;
+			k = Hash()(key); i = k & mask;
+			last = i;
+			while (!__ac_isempty(flags, i) && (__ac_isdel(flags, i) || !Eq()(keys[i], key))) {
+				i = (i + (++step)) & mask;
+				if (i == last) return n_buckets;
+			}
+			return __ac_iseither(flags, i)? n_buckets : i;
+		} else return 0;
+	};
+
+	int resize(khint_t new_n_buckets) {
+		uint32_t *new_flags = 0;
+		khint_t j = 1;
+		{
+			kroundup32(new_n_buckets);
+			if (new_n_buckets < 4) new_n_buckets = 4;
+			if (count >= (new_n_buckets>>1) + (new_n_buckets>>2)) j = 0;	/* requested count is too small */
+			else { /* hash table count to be changed (shrink or expand); rehash */
+				new_flags = (uint32_t*)std::malloc(__ac_fsize(new_n_buckets) * sizeof(uint32_t));
+				if (!new_flags) return -1;
+				::memset(new_flags, 0xaa, __ac_fsize(new_n_buckets) * sizeof(uint32_t));
+				if (n_buckets < new_n_buckets) {	/* expand */
+					T *new_keys = (T*)std::realloc((void *)keys, new_n_buckets * sizeof(T));
+					if (!new_keys) { std::free(new_flags); return -1; }
+					keys = new_keys;
+				} /* otherwise shrink */
+			}
+		}
+		if (j) { /* rehashing is needed */
+			for (j = 0; j != n_buckets; ++j) {
+				if (__ac_iseither(flags, j) == 0) {
+					T key = keys[j];
+					khint_t new_mask;
+					new_mask = new_n_buckets - 1;
+					__ac_set_isdel_true(flags, j);
+					while (1) { /* kick-out process; sort of like in Cuckoo hashing */
+						khint_t k, i, step = 0;
+						k = Hash()(key);
+						i = k & new_mask;
+						while (!__ac_isempty(new_flags, i)) i = (i + (++step)) & new_mask;
+						__ac_set_isempty_false(new_flags, i);
+						if (i < n_buckets && __ac_iseither(flags, i) == 0) { /* kick out the existing element */
+							{ T tmp = keys[i]; keys[i] = key; key = tmp; }
+							__ac_set_isdel_true(flags, i); /* mark it as deleted in the old hash table */
+						} else { /* write the element and jump out of the loop */
+							keys[i] = key;
+							break;
+						}
+					}
+				}
+			}
+			if (n_buckets > new_n_buckets) /* shrink the hash table */
+				keys = (T*)std::realloc((void *)keys, new_n_buckets * sizeof(T));
+			std::free(flags); /* free the working space */
+			flags = new_flags;
+			n_buckets = new_n_buckets;
+			n_occupied = count;
+			upper_bound = (n_buckets>>1) + (n_buckets>>2);
+		}
+		return 0;
+	};
+
+	khint_t put(const T &key, int *ret) {
+		khint_t x;
+		if (n_occupied >= upper_bound) { /* update the hash table */
+			if (n_buckets > (count<<1)) {
+				if (resize(n_buckets - 1) < 0) { /* clear "deleted" elements */
+					*ret = -1; return n_buckets;
+				}
+			} else if (resize(n_buckets + 1) < 0) { /* expand the hash table */
+				*ret = -1; return n_buckets;
+			}
+		} /* TODO: to implement automatically shrinking; resize() already support shrinking */
+		{
+			khint_t k, i, site, last, mask = n_buckets - 1, step = 0;
+			x = site = n_buckets; k = Hash()(key); i = k & mask;
+			if (__ac_isempty(flags, i)) x = i; /* for speed up */
+			else {
+				last = i;
+				while (!__ac_isempty(flags, i) && (__ac_isdel(flags, i) || !Eq()(keys[i], key))) {
+					if (__ac_isdel(flags, i)) site = i;
+					i = (i + (++step)) & mask;
+					if (i == last) { x = site; break; }
+				}
+				if (x == n_buckets) {
+					if (__ac_isempty(flags, i) && site != n_buckets) x = site;
+					else x = i;
+				}
+			}
+		}
+		if (__ac_isempty(flags, x)) { /* not present at all */
+			keys[x] = key;
+			__ac_set_isboth_false(flags, x);
+			++count; ++n_occupied;
+			*ret = 1;
+		} else if (__ac_isdel(flags, x)) { /* deleted */
+			keys[x] = key;
+			__ac_set_isboth_false(flags, x);
+			++count;
+			*ret = 2;
+		} else *ret = 0; /* Don't touch keys[x] if present and not deleted */
+		return x;
+	};
+
+	void del(khint_t x) {
+		if (x != n_buckets && !__ac_iseither(flags, x)) {
+			__ac_set_isdel_true(flags, x);
+			--count;
+		}
+	};
+};
+
+} // end of namespace klib
+
+#endif
diff --git a/ext/klib/cpp/khashl.hpp b/ext/klib/cpp/khashl.hpp
new file mode 100644
index 0000000..8b870d6
--- /dev/null
+++ b/ext/klib/cpp/khashl.hpp
@@ -0,0 +1,258 @@
+#ifndef __AC_KHASHL_HPP
+#define __AC_KHASHL_HPP
+
+#include <functional> // for std::equal_to
+#include <cstdlib>    // for malloc() etc
+#include <cstring>    // for memset()
+#include <stdint.h>   // for uint32_t
+
+/* // ==> Code example <==
+#include <cstdio>
+#include "khashl.hpp"
+
+int main(void)
+{
+	klib::KHashMap<uint32_t, int, std::hash<uint32_t> > h; // NB: C++98 doesn't have std::hash
+	uint32_t k;
+	int absent;
+	h[43] = 1, h[53] = 2, h[63] = 3, h[73] = 4;       // one way to insert
+	k = h.put(53, &absent), h.value(k) = -2;          // another way to insert
+	if (!absent) printf("already in the table\n");    //   which allows to test presence
+	if (h.get(33) == h.end()) printf("not found!\n"); // test presence without insertion
+	h.del(h.get(43));               // deletion
+	for (k = 0; k != h.end(); ++k)  // traversal
+		if (h.occupied(k))          // some buckets are not occupied; skip them
+			printf("%u => %d\n", h.key(k), h.value(k));
+	return 0;
+}
+*/
+
+namespace klib {
+
+/***********
+ * HashSet *
+ ***********/
+
+template<class T, class Hash, class Eq = std::equal_to<T>, typename khint_t = uint32_t>
+class KHashSet {
+	khint_t bits, count;
+	uint32_t *used;
+	T *keys;
+	static inline uint32_t __kh_used(const uint32_t *flag, khint_t i) { return flag[i>>5] >> (i&0x1fU) & 1U; };
+	static inline void __kh_set_used(uint32_t *flag, khint_t i) { flag[i>>5] |= 1U<<(i&0x1fU); };
+	static inline void __kh_set_unused(uint32_t *flag, khint_t i) { flag[i>>5] &= ~(1U<<(i&0x1fU)); };
+	static inline khint_t __kh_fsize(khint_t m) { return m<32? 1 : m>>5; }
+	static inline khint_t __kh_h2b(uint32_t hash, khint_t bits) { return hash * 2654435769U >> (32 - bits); }
+	static inline khint_t __kh_h2b(uint64_t hash, khint_t bits) { return hash * 11400714819323198485ULL >> (64 - bits); }
+public:
+	KHashSet() : bits(0), count(0), used(0), keys(0) {};
+	~KHashSet() { std::free(used); std::free(keys); };
+	inline khint_t n_buckets() const { return used? khint_t(1) << bits : 0; }
+	inline khint_t end() const { return n_buckets(); }
+	inline khint_t size() const { return count; }
+	inline T &key(khint_t x) { return keys[x]; };
+	inline bool occupied(khint_t x) const { return (__kh_used(used, x) != 0); }
+	void clear(void) { 
+		if (!used) return;
+		memset(used, 0, __kh_fsize(n_buckets()) * sizeof(uint32_t));
+		count = 0;
+	}
+	khint_t get(const T &key) const {
+		khint_t i, last, mask, nb;
+		if (keys == 0) return 0;
+		nb = n_buckets();
+		mask = nb - khint_t(1);
+		i = last = __kh_h2b(Hash()(key), bits);
+		while (__kh_used(used, i) && !Eq()(keys[i], key)) {
+			i = (i + khint_t(1)) & mask;
+			if (i == last) return nb;
+		}
+		return !__kh_used(used, i)? nb : i;
+	}
+	int resize(khint_t new_nb) {
+		uint32_t *new_used = 0;
+		khint_t j = 0, x = new_nb, nb, new_bits, new_mask;
+		while ((x >>= khint_t(1)) != 0) ++j;
+		if (new_nb & (new_nb - 1)) ++j;
+		new_bits = j > 2? j : 2;
+		new_nb = khint_t(1) << new_bits;
+		if (count > (new_nb>>1) + (new_nb>>2)) return 0; /* requested size is too small */
+		new_used = (uint32_t*)std::malloc(__kh_fsize(new_nb) * sizeof(uint32_t));
+		memset(new_used, 0, __kh_fsize(new_nb) * sizeof(uint32_t));
+		if (!new_used) return -1; /* not enough memory */
+		nb = n_buckets();
+		if (nb < new_nb) { /* expand */
+			T *new_keys = (T*)std::realloc(keys, new_nb * sizeof(T));
+			if (!new_keys) { std::free(new_used); return -1; }
+			keys = new_keys;
+		} /* otherwise shrink */
+		new_mask = new_nb - 1;
+		for (j = 0; j != nb; ++j) {
+			if (!__kh_used(used, j)) continue;
+			T key = keys[j];
+			__kh_set_unused(used, j);
+			while (1) { /* kick-out process; sort of like in Cuckoo hashing */
+				khint_t i;
+				i = __kh_h2b(Hash()(key), new_bits);
+				while (__kh_used(new_used, i)) i = (i + khint_t(1)) & new_mask;
+				__kh_set_used(new_used, i);
+				if (i < nb && __kh_used(used, i)) { /* kick out the existing element */
+					{ T tmp = keys[i]; keys[i] = key; key = tmp; }
+					__kh_set_unused(used, i); /* mark it as deleted in the old hash table */
+				} else { /* write the element and jump out of the loop */
+					keys[i] = key;
+					break;
+				}
+			}
+		}
+		if (nb > new_nb) /* shrink the hash table */
+			keys = (T*)std::realloc(keys, new_nb * sizeof(T));
+		std::free(used); /* free the working space */
+		used = new_used, bits = new_bits;
+		return 0;
+	}
+	khint_t put(const T &key, int *absent_ = 0) {
+		khint_t nb, i, last, mask;
+		int absent = -1;
+		nb = n_buckets();
+		if (count >= (nb>>1) + (nb>>2)) { /* rehashing */
+			if (resize(nb + khint_t(1)) < 0) {
+				if (absent_) *absent_ = -1;
+				return nb;
+			}
+			nb = n_buckets();
+		} /* TODO: to implement automatically shrinking; resize() already support shrinking */
+		mask = nb - 1;
+		i = last = __kh_h2b(Hash()(key), bits);
+		while (__kh_used(used, i) && !Eq()(keys[i], key)) {
+			i = (i + 1U) & mask;
+			if (i == last) break;
+		}
+		if (!__kh_used(used, i)) { /* not present at all */
+			keys[i] = key;
+			__kh_set_used(used, i);
+			++count, absent = 1;
+		} else absent = 0; /* Don't touch keys[i] if present */
+		if (absent_) *absent_ = absent;
+		return i;
+	}
+	int del(khint_t i) {
+		khint_t j = i, k, mask, nb = n_buckets();
+		if (keys == 0 || i >= nb) return 0;
+		mask = nb - khint_t(1);
+		while (1) {
+			j = (j + khint_t(1)) & mask;
+			if (j == i || !__kh_used(used, j)) break; /* j==i only when the table is completely full */
+			k = __kh_h2b(Hash()(keys[j]), bits);
+			if ((j > i && (k <= i || k > j)) || (j < i && (k <= i && k > j)))
+				keys[i] = keys[j], i = j;
+		}
+		__kh_set_unused(used, i);
+		--count;
+		return 1;
+	}
+};
+
+/***********
+ * HashMap *
+ ***********/
+
+template<class KType, class VType>
+struct KHashMapBucket { KType key; VType val; };
+
+template<class T, class Hash, typename khint_t>
+struct KHashMapHash { khint_t operator() (const T &a) const { return Hash()(a.key); } };
+
+template<class T, class Eq>
+struct KHashMapEq { bool operator() (const T &a, const T &b) const { return Eq()(a.key, b.key); } };
+
+template<class KType, class VType, class Hash, class Eq=std::equal_to<KType>, typename khint_t=uint32_t>
+class KHashMap : public KHashSet<KHashMapBucket<KType, VType>,
+		KHashMapHash<KHashMapBucket<KType, VType>, Hash, khint_t>,
+		KHashMapEq<KHashMapBucket<KType, VType>, Eq>, khint_t>
+{
+	typedef KHashMapBucket<KType, VType> bucket_t;
+	typedef KHashSet<bucket_t, KHashMapHash<bucket_t, Hash, khint_t>, KHashMapEq<bucket_t, Eq>, khint_t> hashset_t;
+public:
+	khint_t get(const KType &key) const {
+		bucket_t t = { key, VType() };
+		return hashset_t::get(t);
+	}
+	khint_t put(const KType &key, int *absent) {
+		bucket_t t = { key, VType() };
+		return hashset_t::put(t, absent);
+	}
+	inline KType &key(khint_t i) { return hashset_t::key(i).key; }
+	inline VType &value(khint_t i) { return hashset_t::key(i).val; }
+	inline VType &operator[] (const KType &key) {
+		bucket_t t = { key, VType() };
+		return value(hashset_t::put(t));
+	}
+};
+
+/****************************
+ * HashSet with cached hash *
+ ****************************/
+
+template<class KType, typename khint_t>
+struct KHashSetCachedBucket { KType key; khint_t hash; };
+
+template<class T, typename khint_t>
+struct KHashCachedHash { khint_t operator() (const T &a) const { return a.hash; } };
+
+template<class T, class Eq>
+struct KHashCachedEq { bool operator() (const T &a, const T &b) const { return a.hash == b.hash && Eq()(a.key, b.key); } };
+
+template<class KType, class Hash, class Eq = std::equal_to<KType>, typename khint_t = uint32_t>
+class KHashSetCached : public KHashSet<KHashSetCachedBucket<KType, khint_t>,
+		KHashCachedHash<KHashSetCachedBucket<KType, khint_t>, khint_t>,
+		KHashCachedEq<KHashSetCachedBucket<KType, khint_t>, Eq>, khint_t>
+{
+	typedef KHashSetCachedBucket<KType, khint_t> bucket_t;
+	typedef KHashSet<bucket_t, KHashCachedHash<bucket_t, khint_t>, KHashCachedEq<bucket_t, Eq>, khint_t> hashset_t;
+public:
+	khint_t get(const KType &key) const {
+		bucket_t t = { key, Hash()(key) };
+		return hashset_t::get(t);
+	}
+	khint_t put(const KType &key, int *absent) {
+		bucket_t t = { key, Hash()(key) };
+		return hashset_t::put(t, absent);
+	}
+	inline KType &key(khint_t i) { return hashset_t::key(i).key; }
+};
+
+/****************************
+ * HashMap with cached hash *
+ ****************************/
+
+template<class KType, class VType, typename khint_t>
+struct KHashMapCachedBucket { KType key; VType val; khint_t hash; };
+
+template<class KType, class VType, class Hash, class Eq = std::equal_to<KType>, typename khint_t = uint32_t>
+class KHashMapCached : public KHashSet<KHashMapCachedBucket<KType, VType, khint_t>,
+		KHashCachedHash<KHashMapCachedBucket<KType, VType, khint_t>, khint_t>,
+		KHashCachedEq<KHashMapCachedBucket<KType, VType, khint_t>, Eq>, khint_t>
+{
+	typedef KHashMapCachedBucket<KType, VType, khint_t> bucket_t;
+	typedef KHashSet<bucket_t, KHashCachedHash<bucket_t, khint_t>, KHashCachedEq<bucket_t, Eq>, khint_t> hashset_t;
+public:
+	khint_t get(const KType &key) const {
+		bucket_t t = { key, VType(), Hash()(key) };
+		return hashset_t::get(t);
+	}
+	khint_t put(const KType &key, int *absent) {
+		bucket_t t = { key, VType(), Hash()(key) };
+		return hashset_t::put(t, absent);
+	}
+	inline KType &key(khint_t i) { return hashset_t::key(i).key; }
+	inline VType &value(khint_t i) { return hashset_t::key(i).val; }
+	inline VType &operator[] (const KType &key) {
+		bucket_t t = { key, VType(), Hash()(key) };
+		return value(hashset_t::put(t));
+	}
+};
+
+}
+
+#endif /* __AC_KHASHL_HPP */
diff --git a/ext/klib/kalloc.c b/ext/klib/kalloc.c
new file mode 100644
index 0000000..f5de41a
--- /dev/null
+++ b/ext/klib/kalloc.c
@@ -0,0 +1,224 @@
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include "kalloc.h"
+
+/* In kalloc, a *core* is a large chunk of contiguous memory. Each core is
+ * associated with a master header, which keeps the size of the current core
+ * and the pointer to next core. Kalloc allocates small *blocks* of memory from
+ * the cores and organizes free memory blocks in a circular single-linked list.
+ *
+ * In the following diagram, "@" stands for the header of a free block (of type
+ * header_t), "#" for the header of an allocated block (of type size_t), "-"
+ * for free memory, and "+" for allocated memory.
+ *
+ * master        This region is core 1.          master           This region is core 2.
+ *      |                                             |
+ *      *@-------#++++++#++++++++++++@--------        *@----------#++++++++++++#+++++++@------------
+ *       |                           |                 |                               |
+ *       p=p->ptr->ptr->ptr->ptr     p->ptr            p->ptr->ptr                     p->ptr->ptr->ptr
+ */
+typedef struct header_t {
+	size_t size;
+	struct header_t *ptr;
+} header_t;
+
+typedef struct {
+	void *par;
+	size_t min_core_size;
+	header_t base, *loop_head, *core_head; /* base is a zero-sized block always kept in the loop */
+} kmem_t;
+
+static void panic(const char *s)
+{
+	fprintf(stderr, "%s\n", s);
+	abort();
+}
+
+void *km_init2(void *km_par, size_t min_core_size)
+{
+	kmem_t *km;
+	km = (kmem_t*)kcalloc(km_par, 1, sizeof(kmem_t));
+	km->par = km_par;
+	if (km_par) km->min_core_size = min_core_size > 0? min_core_size : ((kmem_t*)km_par)->min_core_size - 2;
+	else km->min_core_size = min_core_size > 0? min_core_size : 0x80000;
+	return (void*)km;
+}
+
+void *km_init(void) { return km_init2(0, 0); }
+
+void km_destroy(void *_km)
+{
+	kmem_t *km = (kmem_t*)_km;
+	void *km_par;
+	header_t *p, *q;
+	if (km == NULL) return;
+	km_par = km->par;
+	for (p = km->core_head; p != NULL;) {
+		q = p->ptr;
+		kfree(km_par, p);
+		p = q;
+	}
+	kfree(km_par, km);
+}
+
+static header_t *morecore(kmem_t *km, size_t nu)
+{
+	header_t *q;
+	size_t bytes, *p;
+	nu = (nu + 1 + (km->min_core_size - 1)) / km->min_core_size * km->min_core_size; /* the first +1 for core header */
+	bytes = nu * sizeof(header_t);
+	q = (header_t*)kmalloc(km->par, bytes);
+	if (!q) panic("[morecore] insufficient memory");
+	q->ptr = km->core_head, q->size = nu, km->core_head = q;
+	p = (size_t*)(q + 1);
+	*p = nu - 1; /* the size of the free block; -1 because the first unit is used for the core header */
+	kfree(km, p + 1); /* initialize the new "core"; NB: the core header is not looped. */
+	return km->loop_head;
+}
+
+void kfree(void *_km, void *ap) /* kfree() also adds a new core to the circular list */
+{
+	header_t *p, *q;
+	kmem_t *km = (kmem_t*)_km;
+	
+	if (!ap) return;
+	if (km == NULL) {
+		free(ap);
+		return;
+	}
+	p = (header_t*)((size_t*)ap - 1);
+	p->size = *((size_t*)ap - 1);
+	/* Find the pointer that points to the block to be freed. The following loop can stop on two conditions:
+	 *
+	 * a) "p>q && p<q->ptr": @------#++++++++#+++++++@-------    @---------------#+++++++@-------
+	 *    (can also be in    |      |                |        -> |                       |
+	 *     two cores)        q      p           q->ptr           q                  q->ptr
+	 *
+	 *                       @--------    #+++++++++@--------    @--------    @------------------
+	 *                       |            |         |         -> |            |
+	 *                       q            p    q->ptr            q       q->ptr
+	 *
+	 * b) "q>=q->ptr && (p>q || p<q->ptr)":  @-------#+++++   @--------#+++++++     @-------#+++++   @----------------
+	 *                                       |                |        |         -> |                |
+	 *                                  q->ptr                q        p       q->ptr                q
+	 *
+	 *                                       #+++++++@-----   #++++++++@-------     @-------------   #++++++++@-------
+	 *                                       |       |                 |         -> |                         |
+	 *                                       p  q->ptr                 q       q->ptr                         q
+	 */
+	for (q = km->loop_head; !(p > q && p < q->ptr); q = q->ptr)
+		if (q >= q->ptr && (p > q || p < q->ptr)) break;
+	if (p + p->size == q->ptr) { /* two adjacent blocks, merge p and q->ptr (the 2nd and 4th cases) */
+		p->size += q->ptr->size;
+		p->ptr = q->ptr->ptr;
+	} else if (p + p->size > q->ptr && q->ptr >= p) {
+		panic("[kfree] The end of the allocated block enters a free block.");
+	} else p->ptr = q->ptr; /* backup q->ptr */
+
+	if (q + q->size == p) { /* two adjacent blocks, merge q and p (the other two cases) */
+		q->size += p->size;
+		q->ptr = p->ptr;
+		km->loop_head = q;
+	} else if (q + q->size > p && p >= q) {
+		panic("[kfree] The end of a free block enters the allocated block.");
+	} else km->loop_head = p, q->ptr = p; /* in two cores, cannot be merged; create a new block in the list */
+}
+
+void *kmalloc(void *_km, size_t n_bytes)
+{
+	kmem_t *km = (kmem_t*)_km;
+	size_t n_units;
+	header_t *p, *q;
+
+	if (n_bytes == 0) return 0;
+	if (km == NULL) return malloc(n_bytes);
+	n_units = (n_bytes + sizeof(size_t) + sizeof(header_t) - 1) / sizeof(header_t); /* header+n_bytes requires at least this number of units */
+
+	if (!(q = km->loop_head)) /* the first time when kmalloc() is called, intialize it */
+		q = km->loop_head = km->base.ptr = &km->base;
+	for (p = q->ptr;; q = p, p = p->ptr) { /* search for a suitable block */
+		if (p->size >= n_units) { /* p->size if the size of current block. This line means the current block is large enough. */
+			if (p->size == n_units) q->ptr = p->ptr; /* no need to split the block */
+			else { /* split the block. NB: memory is allocated at the end of the block! */
+				p->size -= n_units; /* reduce the size of the free block */
+				p += p->size; /* p points to the allocated block */
+				*(size_t*)p = n_units; /* set the size */
+			}
+			km->loop_head = q; /* set the end of chain */
+			return (size_t*)p + 1;
+		}
+		if (p == km->loop_head) { /* then ask for more "cores" */
+			if ((p = morecore(km, n_units)) == 0) return 0;
+		}
+	}
+}
+
+void *kcalloc(void *_km, size_t count, size_t size)
+{
+	kmem_t *km = (kmem_t*)_km;
+	void *p;
+	if (size == 0 || count == 0) return 0;
+	if (km == NULL) return calloc(count, size);
+	p = kmalloc(km, count * size);
+	memset(p, 0, count * size);
+	return p;
+}
+
+void *krealloc(void *_km, void *ap, size_t n_bytes) // TODO: this can be made more efficient in principle
+{
+	kmem_t *km = (kmem_t*)_km;
+	size_t cap, *p, *q;
+
+	if (n_bytes == 0) {
+		kfree(km, ap); return 0;
+	}
+	if (km == NULL) return realloc(ap, n_bytes);
+	if (ap == NULL) return kmalloc(km, n_bytes);
+	p = (size_t*)ap - 1;
+	cap = (*p) * sizeof(header_t) - sizeof(size_t);
+	if (cap >= n_bytes) return ap; /* TODO: this prevents shrinking */
+	q = (size_t*)kmalloc(km, n_bytes);
+	memcpy(q, ap, cap);
+	kfree(km, ap);
+	return q;
+}
+
+void *krelocate(void *km, void *ap, size_t n_bytes)
+{
+	void *p;
+	if (km == 0 || ap == 0) return ap;
+	p = kmalloc(km, n_bytes);
+	memcpy(p, ap, n_bytes);
+	kfree(km, ap);
+	return p;
+}
+
+void km_stat(const void *_km, km_stat_t *s)
+{
+	kmem_t *km = (kmem_t*)_km;
+	header_t *p;
+	memset(s, 0, sizeof(km_stat_t));
+	if (km == NULL || km->loop_head == NULL) return;
+	for (p = km->loop_head;; p = p->ptr) {
+		s->available += p->size * sizeof(header_t);
+		if (p->size != 0) ++s->n_blocks; /* &kmem_t::base is always one of the cores. It is zero-sized. */
+		if (p->ptr > p && p + p->size > p->ptr)
+			panic("[km_stat] The end of a free block enters another free block.");
+		if (p->ptr == km->loop_head) break;
+	}
+	for (p = km->core_head; p != NULL; p = p->ptr) {
+		size_t size = p->size * sizeof(header_t);
+		++s->n_cores;
+		s->capacity += size;
+		s->largest = s->largest > size? s->largest : size;
+	}
+}
+
+void km_stat_print(const void *km)
+{
+	km_stat_t st;
+	km_stat(km, &st);
+	fprintf(stderr, "[km_stat] cap=%ld, avail=%ld, largest=%ld, n_core=%ld, n_block=%ld\n",
+			st.capacity, st.available, st.largest, st.n_blocks, st.n_cores);
+}
diff --git a/ext/klib/kalloc.h b/ext/klib/kalloc.h
new file mode 100644
index 0000000..4378672
--- /dev/null
+++ b/ext/klib/kalloc.h
@@ -0,0 +1,87 @@
+#ifndef _KALLOC_H_
+#define _KALLOC_H_
+
+#include <stddef.h> /* for size_t */
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef struct {
+	size_t capacity, available, n_blocks, n_cores, largest;
+} km_stat_t;
+
+void *kmalloc(void *km, size_t size);
+void *krealloc(void *km, void *ptr, size_t size);
+void *krelocate(void *km, void *ap, size_t n_bytes);
+void *kcalloc(void *km, size_t count, size_t size);
+void kfree(void *km, void *ptr);
+
+void *km_init(void);
+void *km_init2(void *km_par, size_t min_core_size);
+void km_destroy(void *km);
+void km_stat(const void *_km, km_stat_t *s);
+void km_stat_print(const void *km);
+
+#ifdef __cplusplus
+}
+#endif
+
+#define Kmalloc(km, type, cnt)       ((type*)kmalloc((km), (cnt) * sizeof(type)))
+#define Kcalloc(km, type, cnt)       ((type*)kcalloc((km), (cnt), sizeof(type)))
+#define Krealloc(km, type, ptr, cnt) ((type*)krealloc((km), (ptr), (cnt) * sizeof(type)))
+
+#define Kexpand(km, type, a, m) do { \
+		(m) = (m) >= 4? (m) + ((m)>>1) : 16; \
+		(a) = Krealloc(km, type, (a), (m)); \
+	} while (0)
+
+#define KMALLOC(km, ptr, len) ((ptr) = (__typeof__(ptr))kmalloc((km), (len) * sizeof(*(ptr))))
+#define KCALLOC(km, ptr, len) ((ptr) = (__typeof__(ptr))kcalloc((km), (len), sizeof(*(ptr))))
+#define KREALLOC(km, ptr, len) ((ptr) = (__typeof__(ptr))krealloc((km), (ptr), (len) * sizeof(*(ptr))))
+
+#define KEXPAND(km, a, m) do { \
+		(m) = (m) >= 4? (m) + ((m)>>1) : 16; \
+		KREALLOC((km), (a), (m)); \
+	} while (0)
+
+#ifndef klib_unused
+#if (defined __clang__ && __clang_major__ >= 3) || (defined __GNUC__ && __GNUC__ >= 3)
+#define klib_unused __attribute__ ((__unused__))
+#else
+#define klib_unused
+#endif
+#endif /* klib_unused */
+
+#define KALLOC_POOL_INIT2(SCOPE, name, kmptype_t) \
+	typedef struct { \
+		size_t cnt, n, max; \
+		kmptype_t **buf; \
+		void *km; \
+	} kmp_##name##_t; \
+	SCOPE kmp_##name##_t *kmp_init_##name(void *km) { \
+		kmp_##name##_t *mp; \
+		mp = Kcalloc(km, kmp_##name##_t, 1); \
+		mp->km = km; \
+		return mp; \
+	} \
+	SCOPE void kmp_destroy_##name(kmp_##name##_t *mp) { \
+		size_t k; \
+		for (k = 0; k < mp->n; ++k) kfree(mp->km, mp->buf[k]); \
+		kfree(mp->km, mp->buf); kfree(mp->km, mp); \
+	} \
+	SCOPE kmptype_t *kmp_alloc_##name(kmp_##name##_t *mp) { \
+		++mp->cnt; \
+		if (mp->n == 0) return (kmptype_t*)kcalloc(mp->km, 1, sizeof(kmptype_t)); \
+		return mp->buf[--mp->n]; \
+	} \
+	SCOPE void kmp_free_##name(kmp_##name##_t *mp, kmptype_t *p) { \
+		--mp->cnt; \
+		if (mp->n == mp->max) Kexpand(mp->km, kmptype_t*, mp->buf, mp->max); \
+		mp->buf[mp->n++] = p; \
+	}
+
+#define KALLOC_POOL_INIT(name, kmptype_t) \
+	KALLOC_POOL_INIT2(static inline klib_unused, name, kmptype_t)
+
+#endif
diff --git a/ext/klib/kavl-lite.h b/ext/klib/kavl-lite.h
new file mode 100644
index 0000000..240ac8a
--- /dev/null
+++ b/ext/klib/kavl-lite.h
@@ -0,0 +1,306 @@
+/* The MIT License
+
+   Copyright (c) 2021 by Attractive Chaos <attractor@live.co.uk>
+
+   Permission is hereby granted, free of charge, to any person obtaining
+   a copy of this software and associated documentation files (the
+   "Software"), to deal in the Software without restriction, including
+   without limitation the rights to use, copy, modify, merge, publish,
+   distribute, sublicense, and/or sell copies of the Software, and to
+   permit persons to whom the Software is furnished to do so, subject to
+   the following conditions:
+
+   The above copyright notice and this permission notice shall be
+   included in all copies or substantial portions of the Software.
+
+   THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+   EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+   MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+   NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+   BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+   ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+   CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+   SOFTWARE.
+*/
+
+/* An example:
+
+#include <stdio.h>
+#include <string.h>
+#include <stdlib.h>
+#include "kavl-lite.h"
+
+struct my_node {
+  char key;
+  KAVLL_HEAD(struct my_node) head;
+};
+#define my_cmp(p, q) (((q)->key < (p)->key) - ((p)->key < (q)->key))
+KAVLL_INIT(my, struct my_node, head, my_cmp)
+
+int main(void) {
+  const char *str = "MNOLKQOPHIA"; // from wiki, except a duplicate
+  struct my_node *root = 0;
+  int i, l = strlen(str);
+  for (i = 0; i < l; ++i) {        // insert in the input order
+    struct my_node *q, *p = malloc(sizeof(*p));
+    p->key = str[i];
+    q = my_insert(&root, p);
+    if (p != q) free(p);           // if already present, free
+  }
+  my_itr_t itr;
+  my_itr_first(root, &itr);  // place at first
+  do {                             // traverse
+    const struct my_node *p = kavll_at(&itr);
+    putchar(p->key);
+    free((void*)p);                // free node
+  } while (my_itr_next(&itr));
+  putchar('\n');
+  return 0;
+}
+*/
+
+#ifndef KAVL_LITE_H
+#define KAVL_LITE_H
+
+#ifdef __STRICT_ANSI__
+#define inline __inline__
+#endif
+
+#define KAVLL_MAX_DEPTH 64
+
+#define KAVLL_HEAD(__type) \
+	struct { \
+		__type *p[2]; \
+		signed char balance; /* balance factor */ \
+	}
+
+#define __KAVLL_FIND(pre, __scope, __type, __head,  __cmp) \
+	__scope __type *pre##_find(const __type *root, const __type *x) { \
+		const __type *p = root; \
+		while (p != 0) { \
+			int cmp; \
+			cmp = __cmp(x, p); \
+			if (cmp < 0) p = p->__head.p[0]; \
+			else if (cmp > 0) p = p->__head.p[1]; \
+			else break; \
+		} \
+		return (__type*)p; \
+	}
+
+#define __KAVLL_ROTATE(pre, __type, __head) \
+	/* one rotation: (a,(b,c)q)p => ((a,b)p,c)q */ \
+	static inline __type *pre##_rotate1(__type *p, int dir) { /* dir=0 to left; dir=1 to right */ \
+		int opp = 1 - dir; /* opposite direction */ \
+		__type *q = p->__head.p[opp]; \
+		p->__head.p[opp] = q->__head.p[dir]; \
+		q->__head.p[dir] = p; \
+		return q; \
+	} \
+	/* two consecutive rotations: (a,((b,c)r,d)q)p => ((a,b)p,(c,d)q)r */ \
+	static inline __type *pre##_rotate2(__type *p, int dir) { \
+		int b1, opp = 1 - dir; \
+		__type *q = p->__head.p[opp], *r = q->__head.p[dir]; \
+		p->__head.p[opp] = r->__head.p[dir]; \
+		r->__head.p[dir] = p; \
+		q->__head.p[dir] = r->__head.p[opp]; \
+		r->__head.p[opp] = q; \
+		b1 = dir == 0? +1 : -1; \
+		if (r->__head.balance == b1) q->__head.balance = 0, p->__head.balance = -b1; \
+		else if (r->__head.balance == 0) q->__head.balance = p->__head.balance = 0; \
+		else q->__head.balance = b1, p->__head.balance = 0; \
+		r->__head.balance = 0; \
+		return r; \
+	}
+
+#define __KAVLL_INSERT(pre, __scope, __type, __head, __cmp) \
+	__scope __type *pre##_insert(__type **root_, __type *x) { \
+		unsigned char stack[KAVLL_MAX_DEPTH]; \
+		__type *path[KAVLL_MAX_DEPTH]; \
+		__type *bp, *bq; \
+		__type *p, *q, *r = 0; /* _r_ is potentially the new root */ \
+		int which = 0, top, b1, path_len; \
+		bp = *root_, bq = 0; \
+		/* find the insertion location */ \
+		for (p = bp, q = bq, top = path_len = 0; p; q = p, p = p->__head.p[which]) { \
+			int cmp; \
+			cmp = __cmp(x, p); \
+			if (cmp == 0) return p; \
+			if (p->__head.balance != 0) \
+				bq = q, bp = p, top = 0; \
+			stack[top++] = which = (cmp > 0); \
+			path[path_len++] = p; \
+		} \
+		x->__head.balance = 0, x->__head.p[0] = x->__head.p[1] = 0; \
+		if (q == 0) *root_ = x; \
+		else q->__head.p[which] = x; \
+		if (bp == 0) return x; \
+		for (p = bp, top = 0; p != x; p = p->__head.p[stack[top]], ++top) /* update balance factors */ \
+			if (stack[top] == 0) --p->__head.balance; \
+			else ++p->__head.balance; \
+		if (bp->__head.balance > -2 && bp->__head.balance < 2) return x; /* no re-balance needed */ \
+		/* re-balance */ \
+		which = (bp->__head.balance < 0); \
+		b1 = which == 0? +1 : -1; \
+		q = bp->__head.p[1 - which]; \
+		if (q->__head.balance == b1) { \
+			r = pre##_rotate1(bp, which); \
+			q->__head.balance = bp->__head.balance = 0; \
+		} else r = pre##_rotate2(bp, which); \
+		if (bq == 0) *root_ = r; \
+		else bq->__head.p[bp != bq->__head.p[0]] = r; \
+		return x; \
+	}
+
+#define __KAVLL_ERASE(pre, __scope, __type, __head, __cmp) \
+	__scope __type *pre##_erase(__type **root_, const __type *x) { \
+		__type *p, *path[KAVLL_MAX_DEPTH], fake; \
+		unsigned char dir[KAVLL_MAX_DEPTH]; \
+		int d = 0, cmp; \
+		fake.__head.p[0] = *root_, fake.__head.p[1] = 0; \
+		if (x) { \
+			for (cmp = -1, p = &fake; cmp; cmp = __cmp(x, p)) { \
+				int which = (cmp > 0); \
+				dir[d] = which; \
+				path[d++] = p; \
+				p = p->__head.p[which]; \
+				if (p == 0) return 0; \
+			} \
+		} else { \
+			for (p = &fake; p; p = p->__head.p[0]) \
+				dir[d] = 0, path[d++] = p; \
+			p = path[--d]; \
+		} \
+		if (p->__head.p[1] == 0) { /* ((1,.)2,3)4 => (1,3)4; p=2 */ \
+			path[d-1]->__head.p[dir[d-1]] = p->__head.p[0]; \
+		} else { \
+			__type *q = p->__head.p[1]; \
+			if (q->__head.p[0] == 0) { /* ((1,2)3,4)5 => ((1)2,4)5; p=3 */ \
+				q->__head.p[0] = p->__head.p[0]; \
+				q->__head.balance = p->__head.balance; \
+				path[d-1]->__head.p[dir[d-1]] = q; \
+				path[d] = q, dir[d++] = 1; \
+			} else { /* ((1,((.,2)3,4)5)6,7)8 => ((1,(2,4)5)3,7)8; p=6 */ \
+				__type *r; \
+				int e = d++; /* backup _d_ */\
+				for (;;) { \
+					dir[d] = 0; \
+					path[d++] = q; \
+					r = q->__head.p[0]; \
+					if (r->__head.p[0] == 0) break; \
+					q = r; \
+				} \
+				r->__head.p[0] = p->__head.p[0]; \
+				q->__head.p[0] = r->__head.p[1]; \
+				r->__head.p[1] = p->__head.p[1]; \
+				r->__head.balance = p->__head.balance; \
+				path[e-1]->__head.p[dir[e-1]] = r; \
+				path[e] = r, dir[e] = 1; \
+			} \
+		} \
+		while (--d > 0) { \
+			__type *q = path[d]; \
+			int which, other, b1 = 1, b2 = 2; \
+			which = dir[d], other = 1 - which; \
+			if (which) b1 = -b1, b2 = -b2; \
+			q->__head.balance += b1; \
+			if (q->__head.balance == b1) break; \
+			else if (q->__head.balance == b2) { \
+				__type *r = q->__head.p[other]; \
+				if (r->__head.balance == -b1) { \
+					path[d-1]->__head.p[dir[d-1]] = pre##_rotate2(q, which); \
+				} else { \
+					path[d-1]->__head.p[dir[d-1]] = pre##_rotate1(q, which); \
+					if (r->__head.balance == 0) { \
+						r->__head.balance = -b1; \
+						q->__head.balance = b1; \
+						break; \
+					} else r->__head.balance = q->__head.balance = 0; \
+				} \
+			} \
+		} \
+		*root_ = fake.__head.p[0]; \
+		return p; \
+	}
+
+#define kavll_free(__type, __head, __root, __free) do { \
+		__type *_p, *_q; \
+		for (_p = __root; _p; _p = _q) { \
+			if (_p->__head.p[0] == 0) { \
+				_q = _p->__head.p[1]; \
+				__free(_p); \
+			} else { \
+				_q = _p->__head.p[0]; \
+				_p->__head.p[0] = _q->__head.p[1]; \
+				_q->__head.p[1] = _p; \
+			} \
+		} \
+	} while (0)
+
+#define kavll_size(__type, __head, __root, __cnt) do { \
+		__type *_p, *_q; \
+		*(__cnt) = 0; \
+		for (_p = __root; _p; _p = _q) { \
+			if (_p->__head.p[0] == 0) { \
+				_q = _p->__head.p[1]; \
+				++*(__cnt); \
+			} else { \
+				_q = _p->__head.p[0]; \
+				_p->__head.p[0] = _q->__head.p[1]; \
+				_q->__head.p[1] = _p; \
+			} \
+		} \
+	} while (0)
+
+#define __KAVLL_ITR(pre, __scope, __type, __head, __cmp) \
+	typedef struct pre##_itr_t { \
+		const __type *stack[KAVLL_MAX_DEPTH], **top, *right; /* _right_ points to the right child of *top */ \
+	} pre##_itr_t; \
+	__scope void pre##_itr_first(const __type *root, struct pre##_itr_t *itr) { \
+		const __type *p; \
+		for (itr->top = itr->stack - 1, p = root; p; p = p->__head.p[0]) \
+			*++itr->top = p; \
+		itr->right = (*itr->top)->__head.p[1]; \
+	} \
+	__scope int pre##_itr_find(const __type *root, const __type *x, struct pre##_itr_t *itr) { \
+		const __type *p = root; \
+		itr->top = itr->stack - 1; \
+		while (p != 0) { \
+			int cmp; \
+			cmp = __cmp(x, p); \
+			if (cmp < 0) *++itr->top = p, p = p->__head.p[0]; \
+			else if (cmp > 0) p = p->__head.p[1]; \
+			else break; \
+		} \
+		if (p) { \
+			*++itr->top = p; \
+			itr->right = p->__head.p[1]; \
+			return 1; \
+		} else if (itr->top >= itr->stack) { \
+			itr->right = (*itr->top)->__head.p[1]; \
+			return 0; \
+		} else return 0; \
+	} \
+	__scope int pre##_itr_next(struct pre##_itr_t *itr) { \
+		for (;;) { \
+			const __type *p; \
+			for (p = itr->right, --itr->top; p; p = p->__head.p[0]) \
+				*++itr->top = p; \
+			if (itr->top < itr->stack) return 0; \
+			itr->right = (*itr->top)->__head.p[1]; \
+			return 1; \
+		} \
+	}
+
+#define kavll_at(itr) ((itr)->top < (itr)->stack? 0 : *(itr)->top)
+
+#define KAVLL_INIT2(pre, __scope, __type, __head, __cmp) \
+	__KAVLL_FIND(pre, __scope, __type, __head,  __cmp) \
+	__KAVLL_ROTATE(pre, __type, __head) \
+	__KAVLL_INSERT(pre, __scope, __type, __head, __cmp) \
+	__KAVLL_ERASE(pre, __scope, __type, __head, __cmp) \
+	__KAVLL_ITR(pre, __scope, __type, __head, __cmp)
+
+#define KAVLL_INIT(pre, __type, __head, __cmp) \
+	KAVLL_INIT2(pre,, __type, __head, __cmp)
+
+#endif
diff --git a/ext/klib/kavl.h b/ext/klib/kavl.h
new file mode 100644
index 0000000..f3520fd
--- /dev/null
+++ b/ext/klib/kavl.h
@@ -0,0 +1,400 @@
+/* The MIT License
+
+   Copyright (c) 2018 by Attractive Chaos <attractor@live.co.uk>
+
+   Permission is hereby granted, free of charge, to any person obtaining
+   a copy of this software and associated documentation files (the
+   "Software"), to deal in the Software without restriction, including
+   without limitation the rights to use, copy, modify, merge, publish,
+   distribute, sublicense, and/or sell copies of the Software, and to
+   permit persons to whom the Software is furnished to do so, subject to
+   the following conditions:
+
+   The above copyright notice and this permission notice shall be
+   included in all copies or substantial portions of the Software.
+
+   THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+   EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+   MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+   NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+   BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+   ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+   CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+   SOFTWARE.
+*/
+
+/* An example:
+
+#include <stdio.h>
+#include <string.h>
+#include <stdlib.h>
+#include "kavl.h"
+
+struct my_node {
+  char key;
+  KAVL_HEAD(struct my_node) head;
+};
+#define my_cmp(p, q) (((q)->key < (p)->key) - ((p)->key < (q)->key))
+KAVL_INIT(my, struct my_node, head, my_cmp)
+
+int main(void) {
+  const char *str = "MNOLKQOPHIA"; // from wiki, except a duplicate
+  struct my_node *root = 0;
+  int i, l = strlen(str);
+  for (i = 0; i < l; ++i) {        // insert in the input order
+    struct my_node *q, *p = malloc(sizeof(*p));
+    p->key = str[i];
+    q = kavl_insert(my, &root, p, 0);
+    if (p != q) free(p);           // if already present, free
+  }
+  kavl_itr_t(my) itr;
+  kavl_itr_first(my, root, &itr);  // place at first
+  do {                             // traverse
+    const struct my_node *p = kavl_at(&itr);
+    putchar(p->key);
+    free((void*)p);                // free node
+  } while (kavl_itr_next(my, &itr));
+  putchar('\n');
+  return 0;
+}
+*/
+
+#ifndef KAVL_H
+#define KAVL_H
+
+#ifdef __STRICT_ANSI__
+#define inline __inline__
+#endif
+
+#define KAVL_MAX_DEPTH 64
+
+#define kavl_size(head, p) ((p)? (p)->head.size : 0)
+#define kavl_size_child(head, q, i) ((q)->head.p[(i)]? (q)->head.p[(i)]->head.size : 0)
+
+#define KAVL_HEAD(__type) \
+	struct { \
+		__type *p[2]; \
+		signed char balance; /* balance factor */ \
+		unsigned size; /* #elements in subtree */ \
+	}
+
+#define __KAVL_FIND(suf, __scope, __type, __head,  __cmp) \
+	__scope __type *kavl_find_##suf(const __type *root, const __type *x, unsigned *cnt_) { \
+		const __type *p = root; \
+		unsigned cnt = 0; \
+		while (p != 0) { \
+			int cmp; \
+			cmp = __cmp(x, p); \
+			if (cmp >= 0) cnt += kavl_size_child(__head, p, 0) + 1; \
+			if (cmp < 0) p = p->__head.p[0]; \
+			else if (cmp > 0) p = p->__head.p[1]; \
+			else break; \
+		} \
+		if (cnt_) *cnt_ = cnt; \
+		return (__type*)p; \
+	}
+
+#define __KAVL_ROTATE(suf, __type, __head) \
+	/* one rotation: (a,(b,c)q)p => ((a,b)p,c)q */ \
+	static inline __type *kavl_rotate1_##suf(__type *p, int dir) { /* dir=0 to left; dir=1 to right */ \
+		int opp = 1 - dir; /* opposite direction */ \
+		__type *q = p->__head.p[opp]; \
+		unsigned size_p = p->__head.size; \
+		p->__head.size -= q->__head.size - kavl_size_child(__head, q, dir); \
+		q->__head.size = size_p; \
+		p->__head.p[opp] = q->__head.p[dir]; \
+		q->__head.p[dir] = p; \
+		return q; \
+	} \
+	/* two consecutive rotations: (a,((b,c)r,d)q)p => ((a,b)p,(c,d)q)r */ \
+	static inline __type *kavl_rotate2_##suf(__type *p, int dir) { \
+		int b1, opp = 1 - dir; \
+		__type *q = p->__head.p[opp], *r = q->__head.p[dir]; \
+		unsigned size_x_dir = kavl_size_child(__head, r, dir); \
+		r->__head.size = p->__head.size; \
+		p->__head.size -= q->__head.size - size_x_dir; \
+		q->__head.size -= size_x_dir + 1; \
+		p->__head.p[opp] = r->__head.p[dir]; \
+		r->__head.p[dir] = p; \
+		q->__head.p[dir] = r->__head.p[opp]; \
+		r->__head.p[opp] = q; \
+		b1 = dir == 0? +1 : -1; \
+		if (r->__head.balance == b1) q->__head.balance = 0, p->__head.balance = -b1; \
+		else if (r->__head.balance == 0) q->__head.balance = p->__head.balance = 0; \
+		else q->__head.balance = b1, p->__head.balance = 0; \
+		r->__head.balance = 0; \
+		return r; \
+	}
+
+#define __KAVL_INSERT(suf, __scope, __type, __head, __cmp) \
+	__scope __type *kavl_insert_##suf(__type **root_, __type *x, unsigned *cnt_) { \
+		unsigned char stack[KAVL_MAX_DEPTH]; \
+		__type *path[KAVL_MAX_DEPTH]; \
+		__type *bp, *bq; \
+		__type *p, *q, *r = 0; /* _r_ is potentially the new root */ \
+		int i, which = 0, top, b1, path_len; \
+		unsigned cnt = 0; \
+		bp = *root_, bq = 0; \
+		/* find the insertion location */ \
+		for (p = bp, q = bq, top = path_len = 0; p; q = p, p = p->__head.p[which]) { \
+			int cmp; \
+			cmp = __cmp(x, p); \
+			if (cmp >= 0) cnt += kavl_size_child(__head, p, 0) + 1; \
+			if (cmp == 0) { \
+				if (cnt_) *cnt_ = cnt; \
+				return p; \
+			} \
+			if (p->__head.balance != 0) \
+				bq = q, bp = p, top = 0; \
+			stack[top++] = which = (cmp > 0); \
+			path[path_len++] = p; \
+		} \
+		if (cnt_) *cnt_ = cnt; \
+		x->__head.balance = 0, x->__head.size = 1, x->__head.p[0] = x->__head.p[1] = 0; \
+		if (q == 0) *root_ = x; \
+		else q->__head.p[which] = x; \
+		if (bp == 0) return x; \
+		for (i = 0; i < path_len; ++i) ++path[i]->__head.size; \
+		for (p = bp, top = 0; p != x; p = p->__head.p[stack[top]], ++top) /* update balance factors */ \
+			if (stack[top] == 0) --p->__head.balance; \
+			else ++p->__head.balance; \
+		if (bp->__head.balance > -2 && bp->__head.balance < 2) return x; /* no re-balance needed */ \
+		/* re-balance */ \
+		which = (bp->__head.balance < 0); \
+		b1 = which == 0? +1 : -1; \
+		q = bp->__head.p[1 - which]; \
+		if (q->__head.balance == b1) { \
+			r = kavl_rotate1_##suf(bp, which); \
+			q->__head.balance = bp->__head.balance = 0; \
+		} else r = kavl_rotate2_##suf(bp, which); \
+		if (bq == 0) *root_ = r; \
+		else bq->__head.p[bp != bq->__head.p[0]] = r; \
+		return x; \
+	}
+
+#define __KAVL_ERASE(suf, __scope, __type, __head, __cmp) \
+	__scope __type *kavl_erase_##suf(__type **root_, const __type *x, unsigned *cnt_) { \
+		__type *p, *path[KAVL_MAX_DEPTH], fake; \
+		unsigned char dir[KAVL_MAX_DEPTH]; \
+		int i, d = 0, cmp; \
+		unsigned cnt = 0; \
+		fake.__head.p[0] = *root_, fake.__head.p[1] = 0; \
+		if (cnt_) *cnt_ = 0; \
+		if (x) { \
+			for (cmp = -1, p = &fake; cmp; cmp = __cmp(x, p)) { \
+				int which = (cmp > 0); \
+				if (cmp > 0) cnt += kavl_size_child(__head, p, 0) + 1; \
+				dir[d] = which; \
+				path[d++] = p; \
+				p = p->__head.p[which]; \
+				if (p == 0) { \
+					if (cnt_) *cnt_ = 0; \
+					return 0; \
+				} \
+			} \
+			cnt += kavl_size_child(__head, p, 0) + 1; /* because p==x is not counted */ \
+		} else { \
+			for (p = &fake, cnt = 1; p; p = p->__head.p[0]) \
+				dir[d] = 0, path[d++] = p; \
+			p = path[--d]; \
+		} \
+		if (cnt_) *cnt_ = cnt; \
+		for (i = 1; i < d; ++i) --path[i]->__head.size; \
+		if (p->__head.p[1] == 0) { /* ((1,.)2,3)4 => (1,3)4; p=2 */ \
+			path[d-1]->__head.p[dir[d-1]] = p->__head.p[0]; \
+		} else { \
+			__type *q = p->__head.p[1]; \
+			if (q->__head.p[0] == 0) { /* ((1,2)3,4)5 => ((1)2,4)5; p=3 */ \
+				q->__head.p[0] = p->__head.p[0]; \
+				q->__head.balance = p->__head.balance; \
+				path[d-1]->__head.p[dir[d-1]] = q; \
+				path[d] = q, dir[d++] = 1; \
+				q->__head.size = p->__head.size - 1; \
+			} else { /* ((1,((.,2)3,4)5)6,7)8 => ((1,(2,4)5)3,7)8; p=6 */ \
+				__type *r; \
+				int e = d++; /* backup _d_ */\
+				for (;;) { \
+					dir[d] = 0; \
+					path[d++] = q; \
+					r = q->__head.p[0]; \
+					if (r->__head.p[0] == 0) break; \
+					q = r; \
+				} \
+				r->__head.p[0] = p->__head.p[0]; \
+				q->__head.p[0] = r->__head.p[1]; \
+				r->__head.p[1] = p->__head.p[1]; \
+				r->__head.balance = p->__head.balance; \
+				path[e-1]->__head.p[dir[e-1]] = r; \
+				path[e] = r, dir[e] = 1; \
+				for (i = e + 1; i < d; ++i) --path[i]->__head.size; \
+				r->__head.size = p->__head.size - 1; \
+			} \
+		} \
+		while (--d > 0) { \
+			__type *q = path[d]; \
+			int which, other, b1 = 1, b2 = 2; \
+			which = dir[d], other = 1 - which; \
+			if (which) b1 = -b1, b2 = -b2; \
+			q->__head.balance += b1; \
+			if (q->__head.balance == b1) break; \
+			else if (q->__head.balance == b2) { \
+				__type *r = q->__head.p[other]; \
+				if (r->__head.balance == -b1) { \
+					path[d-1]->__head.p[dir[d-1]] = kavl_rotate2_##suf(q, which); \
+				} else { \
+					path[d-1]->__head.p[dir[d-1]] = kavl_rotate1_##suf(q, which); \
+					if (r->__head.balance == 0) { \
+						r->__head.balance = -b1; \
+						q->__head.balance = b1; \
+						break; \
+					} else r->__head.balance = q->__head.balance = 0; \
+				} \
+			} \
+		} \
+		*root_ = fake.__head.p[0]; \
+		return p; \
+	}
+
+#define kavl_free(__type, __head, __root, __free) do { \
+		__type *_p, *_q; \
+		for (_p = __root; _p; _p = _q) { \
+			if (_p->__head.p[0] == 0) { \
+				_q = _p->__head.p[1]; \
+				__free(_p); \
+			} else { \
+				_q = _p->__head.p[0]; \
+				_p->__head.p[0] = _q->__head.p[1]; \
+				_q->__head.p[1] = _p; \
+			} \
+		} \
+	} while (0)
+
+#define __KAVL_ITR(suf, __scope, __type, __head, __cmp) \
+	struct kavl_itr_##suf { \
+		const __type *stack[KAVL_MAX_DEPTH], **top, *right; /* _right_ points to the right child of *top */ \
+	}; \
+	__scope void kavl_itr_first_##suf(const __type *root, struct kavl_itr_##suf *itr) { \
+		const __type *p; \
+		for (itr->top = itr->stack - 1, p = root; p; p = p->__head.p[0]) \
+			*++itr->top = p; \
+		itr->right = (*itr->top)->__head.p[1]; \
+	} \
+	__scope int kavl_itr_find_##suf(const __type *root, const __type *x, struct kavl_itr_##suf *itr) { \
+		const __type *p = root; \
+		itr->top = itr->stack - 1; \
+		while (p != 0) { \
+			int cmp; \
+			cmp = __cmp(x, p); \
+			if (cmp < 0) *++itr->top = p, p = p->__head.p[0]; \
+			else if (cmp > 0) p = p->__head.p[1]; \
+			else break; \
+		} \
+		if (p) { \
+			*++itr->top = p; \
+			itr->right = p->__head.p[1]; \
+			return 1; \
+		} else if (itr->top >= itr->stack) { \
+			itr->right = (*itr->top)->__head.p[1]; \
+			return 0; \
+		} else return 0; \
+	} \
+	__scope int kavl_itr_next_##suf(struct kavl_itr_##suf *itr) { \
+		for (;;) { \
+			const __type *p; \
+			for (p = itr->right, --itr->top; p; p = p->__head.p[0]) \
+				*++itr->top = p; \
+			if (itr->top < itr->stack) return 0; \
+			itr->right = (*itr->top)->__head.p[1]; \
+			return 1; \
+		} \
+	}
+
+/**
+ * Insert a node to the tree
+ *
+ * @param suf     name suffix used in KAVL_INIT()
+ * @param proot   pointer to the root of the tree (in/out: root may change)
+ * @param x       node to insert (in)
+ * @param cnt     number of nodes smaller than or equal to _x_; can be NULL (out)
+ *
+ * @return _x_ if not present in the tree, or the node equal to x.
+ */
+#define kavl_insert(suf, proot, x, cnt) kavl_insert_##suf(proot, x, cnt)
+
+/**
+ * Find a node in the tree
+ *
+ * @param suf     name suffix used in KAVL_INIT()
+ * @param root    root of the tree
+ * @param x       node value to find (in)
+ * @param cnt     number of nodes smaller than or equal to _x_; can be NULL (out)
+ *
+ * @return node equal to _x_ if present, or NULL if absent
+ */
+#define kavl_find(suf, root, x, cnt) kavl_find_##suf(root, x, cnt)
+
+/**
+ * Delete a node from the tree
+ *
+ * @param suf     name suffix used in KAVL_INIT()
+ * @param proot   pointer to the root of the tree (in/out: root may change)
+ * @param x       node value to delete; if NULL, delete the first node (in)
+ *
+ * @return node removed from the tree if present, or NULL if absent
+ */
+#define kavl_erase(suf, proot, x, cnt) kavl_erase_##suf(proot, x, cnt)
+#define kavl_erase_first(suf, proot) kavl_erase_##suf(proot, 0, 0)
+
+#define kavl_itr_t(suf) struct kavl_itr_##suf
+
+/**
+ * Place the iterator at the smallest object
+ *
+ * @param suf     name suffix used in KAVL_INIT()
+ * @param root    root of the tree
+ * @param itr     iterator
+ */
+#define kavl_itr_first(suf, root, itr) kavl_itr_first_##suf(root, itr)
+
+/**
+ * Place the iterator at the object equal to or greater than the query
+ *
+ * @param suf     name suffix used in KAVL_INIT()
+ * @param root    root of the tree
+ * @param x       query (in)
+ * @param itr     iterator (out)
+ *
+ * @return 1 if find; 0 otherwise. kavl_at(itr) is NULL if and only if query is
+ *         larger than all objects in the tree
+ */
+#define kavl_itr_find(suf, root, x, itr) kavl_itr_find_##suf(root, x, itr)
+
+/**
+ * Move to the next object in order
+ *
+ * @param itr     iterator (modified)
+ *
+ * @return 1 if there is a next object; 0 otherwise
+ */
+#define kavl_itr_next(suf, itr) kavl_itr_next_##suf(itr)
+
+/**
+ * Return the pointer at the iterator
+ *
+ * @param itr     iterator
+ *
+ * @return pointer if present; NULL otherwise
+ */
+#define kavl_at(itr) ((itr)->top < (itr)->stack? 0 : *(itr)->top)
+
+#define KAVL_INIT2(suf, __scope, __type, __head, __cmp) \
+	__KAVL_FIND(suf, __scope, __type, __head,  __cmp) \
+	__KAVL_ROTATE(suf, __type, __head) \
+	__KAVL_INSERT(suf, __scope, __type, __head, __cmp) \
+	__KAVL_ERASE(suf, __scope, __type, __head, __cmp) \
+	__KAVL_ITR(suf, __scope, __type, __head, __cmp)
+
+#define KAVL_INIT(suf, __type, __head, __cmp) \
+	KAVL_INIT2(suf,, __type, __head, __cmp)
+
+#endif
diff --git a/ext/klib/kbit.h b/ext/klib/kbit.h
new file mode 100644
index 0000000..3793cf8
--- /dev/null
+++ b/ext/klib/kbit.h
@@ -0,0 +1,30 @@
+#ifndef KBIT_H
+#define KBIT_H
+
+#include <stdint.h>
+
+static inline uint64_t kbi_popcount64(uint64_t y) // standard popcount; from wikipedia
+{
+	y -= ((y >> 1) & 0x5555555555555555ull);
+	y = (y & 0x3333333333333333ull) + (y >> 2 & 0x3333333333333333ull);
+	return ((y + (y >> 4)) & 0xf0f0f0f0f0f0f0full) * 0x101010101010101ull >> 56;
+}
+
+static inline uint64_t kbi_DNAcount64(uint64_t y, int c) // count #A/C/G/T from a 2-bit encoded integer; from BWA
+{
+	// reduce nucleotide counting to bits counting
+	y = ((c&2)? y : ~y) >> 1 & ((c&1)? y : ~y) & 0x5555555555555555ull;
+	// count the number of 1s in y
+	y = (y & 0x3333333333333333ull) + (y >> 2 & 0x3333333333333333ull);
+	return ((y + (y >> 4)) & 0xf0f0f0f0f0f0f0full) * 0x101010101010101ull >> 56;
+}
+
+#ifndef kroundup32 // round a 32-bit integer to the next closet integer; from "bit twiddling hacks"
+#define kroundup32(x) (--(x), (x)|=(x)>>1, (x)|=(x)>>2, (x)|=(x)>>4, (x)|=(x)>>8, (x)|=(x)>>16, ++(x))
+#endif
+
+#ifndef kbi_swap
+#define kbi_swap(a, b) (((a) ^= (b)), ((b) ^= (a)), ((a) ^= (b))) // from "bit twiddling hacks"
+#endif
+
+#endif
diff --git a/ext/klib/kbtree.h b/ext/klib/kbtree.h
new file mode 100644
index 0000000..8b4f917
--- /dev/null
+++ b/ext/klib/kbtree.h
@@ -0,0 +1,437 @@
+/*-
+ * Copyright 1997-1999, 2001, John-Mark Gurney.
+ *           2008-2009, Attractive Chaos <attractor@live.co.uk>
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ *
+ * 1. Redistributions of source code must retain the above copyright
+ *    notice, this list of conditions and the following disclaimer.
+ * 2. Redistributions in binary form must reproduce the above copyright
+ *    notice, this list of conditions and the following disclaimer in the
+ *    documentation and/or other materials provided with the distribution.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
+ * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+ * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+ * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
+ * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
+ * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
+ * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
+ * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
+ * SUCH DAMAGE.
+ */
+
+#ifndef __AC_KBTREE_H
+#define __AC_KBTREE_H
+
+#include <stdlib.h>
+#include <string.h>
+#include <stdint.h>
+
+#define KB_MAX_DEPTH 64
+
+typedef struct {
+	int32_t is_internal:1, n:31;
+} kbnode_t;
+
+typedef struct {
+	kbnode_t *x;
+	int i;
+} kbpos_t;
+
+typedef struct {
+	kbpos_t stack[KB_MAX_DEPTH], *p;
+} kbitr_t;
+
+#define	__KB_KEY(type, x)	((type*)((char*)x + 4))
+#define __KB_PTR(btr, x)	((kbnode_t**)((char*)x + btr->off_ptr))
+
+#define __KB_TREE_T(name)						\
+	typedef struct {							\
+		kbnode_t *root;							\
+		int	off_key, off_ptr, ilen, elen;		\
+		int	n, t;								\
+		int	n_keys, n_nodes;					\
+	} kbtree_##name##_t;
+
+#define __KB_INIT(name, key_t)											\
+	kbtree_##name##_t *kb_init_##name(int size)							\
+	{																	\
+		kbtree_##name##_t *b;											\
+		b = (kbtree_##name##_t*)calloc(1, sizeof(kbtree_##name##_t));	\
+		b->t = ((size - 4 - sizeof(void*)) / (sizeof(void*) + sizeof(key_t)) + 1) >> 1; \
+		if (b->t < 2) {													\
+			free(b); return 0;											\
+		}																\
+		b->n = 2 * b->t - 1;											\
+		b->off_ptr = 4 + b->n * sizeof(key_t);							\
+		b->ilen = (4 + sizeof(void*) + b->n * (sizeof(void*) + sizeof(key_t)) + 3) >> 2 << 2; \
+		b->elen = (b->off_ptr + 3) >> 2 << 2;							\
+		b->root = (kbnode_t*)calloc(1, b->ilen);						\
+		++b->n_nodes;													\
+		return b;														\
+	}
+
+#define __kb_destroy(b) do {											\
+		int i, max = 8;													\
+		kbnode_t *x, **top, **stack = 0;								\
+		if (b) {														\
+			top = stack = (kbnode_t**)calloc(max, sizeof(kbnode_t*));	\
+			*top++ = (b)->root;											\
+			while (top != stack) {										\
+				x = *--top;												\
+				if (x->is_internal == 0) { free(x); continue; }			\
+				for (i = 0; i <= x->n; ++i)								\
+					if (__KB_PTR(b, x)[i]) {							\
+						if (top - stack == max) {						\
+							max <<= 1;									\
+							stack = (kbnode_t**)realloc(stack, max * sizeof(kbnode_t*)); \
+							top = stack + (max>>1);						\
+						}												\
+						*top++ = __KB_PTR(b, x)[i];						\
+					}													\
+				free(x);												\
+			}															\
+		}																\
+		free(b); free(stack);											\
+	} while (0)
+
+#define __KB_GET_AUX1(name, key_t, __cmp)								\
+	static inline int __kb_getp_aux_##name(const kbnode_t * __restrict x, const key_t * __restrict k, int *r) \
+	{																	\
+		int tr, *rr, begin = 0, end = x->n;								\
+		if (x->n == 0) return -1;										\
+		rr = r? r : &tr;												\
+		while (begin < end) {											\
+			int mid = (begin + end) >> 1;								\
+			if (__cmp(__KB_KEY(key_t, x)[mid], *k) < 0) begin = mid + 1; \
+			else end = mid;												\
+		}																\
+		if (begin == x->n) { *rr = 1; return x->n - 1; }				\
+		if ((*rr = __cmp(*k, __KB_KEY(key_t, x)[begin])) < 0) --begin;	\
+		return begin;													\
+	}
+
+#define __KB_GET(name, key_t)											\
+	static key_t *kb_getp_##name(kbtree_##name##_t *b, const key_t * __restrict k) \
+	{																	\
+		int i, r = 0;													\
+		kbnode_t *x = b->root;											\
+		while (x) {														\
+			i = __kb_getp_aux_##name(x, k, &r);							\
+			if (i >= 0 && r == 0) return &__KB_KEY(key_t, x)[i];		\
+			if (x->is_internal == 0) return 0;							\
+			x = __KB_PTR(b, x)[i + 1];									\
+		}																\
+		return 0;														\
+	}																	\
+	static inline key_t *kb_get_##name(kbtree_##name##_t *b, const key_t k) \
+	{																	\
+		return kb_getp_##name(b, &k);									\
+	}
+
+#define __KB_INTERVAL(name, key_t)										\
+	static void kb_intervalp_##name(kbtree_##name##_t *b, const key_t * __restrict k, key_t **lower, key_t **upper)	\
+	{																	\
+		int i, r = 0;													\
+		kbnode_t *x = b->root;											\
+		*lower = *upper = 0;											\
+		while (x) {														\
+			i = __kb_getp_aux_##name(x, k, &r);							\
+			if (i >= 0 && r == 0) {										\
+				*lower = *upper = &__KB_KEY(key_t, x)[i];				\
+				return;													\
+			}															\
+			if (i >= 0) *lower = &__KB_KEY(key_t, x)[i];				\
+			if (i < x->n - 1) *upper = &__KB_KEY(key_t, x)[i + 1];		\
+			if (x->is_internal == 0) return;							\
+			x = __KB_PTR(b, x)[i + 1];									\
+		}																\
+	}																	\
+	static inline void kb_interval_##name(kbtree_##name##_t *b, const key_t k, key_t **lower, key_t **upper) \
+	{																	\
+		kb_intervalp_##name(b, &k, lower, upper);						\
+	}
+
+#define __KB_PUT(name, key_t, __cmp)									\
+	/* x must be an internal node */									\
+	static void __kb_split_##name(kbtree_##name##_t *b, kbnode_t *x, int i, kbnode_t *y) \
+	{																	\
+		kbnode_t *z;													\
+		z = (kbnode_t*)calloc(1, y->is_internal? b->ilen : b->elen);	\
+		++b->n_nodes;													\
+		z->is_internal = y->is_internal;								\
+		z->n = b->t - 1;												\
+		memcpy(__KB_KEY(key_t, z), __KB_KEY(key_t, y) + b->t, sizeof(key_t) * (b->t - 1)); \
+		if (y->is_internal) memcpy(__KB_PTR(b, z), __KB_PTR(b, y) + b->t, sizeof(void*) * b->t); \
+		y->n = b->t - 1;												\
+		memmove(__KB_PTR(b, x) + i + 2, __KB_PTR(b, x) + i + 1, sizeof(void*) * (x->n - i)); \
+		__KB_PTR(b, x)[i + 1] = z;										\
+		memmove(__KB_KEY(key_t, x) + i + 1, __KB_KEY(key_t, x) + i, sizeof(key_t) * (x->n - i)); \
+		__KB_KEY(key_t, x)[i] = __KB_KEY(key_t, y)[b->t - 1];			\
+		++x->n;															\
+	}																	\
+	static key_t *__kb_putp_aux_##name(kbtree_##name##_t *b, kbnode_t *x, const key_t * __restrict k) \
+	{																	\
+		int i = x->n - 1;												\
+		key_t *ret;														\
+		if (x->is_internal == 0) {										\
+			i = __kb_getp_aux_##name(x, k, 0);							\
+			if (i != x->n - 1)											\
+				memmove(__KB_KEY(key_t, x) + i + 2, __KB_KEY(key_t, x) + i + 1, (x->n - i - 1) * sizeof(key_t)); \
+			ret = &__KB_KEY(key_t, x)[i + 1];							\
+			*ret = *k;													\
+			++x->n;														\
+		} else {														\
+			i = __kb_getp_aux_##name(x, k, 0) + 1;						\
+			if (__KB_PTR(b, x)[i]->n == 2 * b->t - 1) {					\
+				__kb_split_##name(b, x, i, __KB_PTR(b, x)[i]);			\
+				if (__cmp(*k, __KB_KEY(key_t, x)[i]) > 0) ++i;			\
+			}															\
+			ret = __kb_putp_aux_##name(b, __KB_PTR(b, x)[i], k);		\
+		}																\
+		return ret; 													\
+	}																	\
+	static key_t *kb_putp_##name(kbtree_##name##_t *b, const key_t * __restrict k) \
+	{																	\
+		kbnode_t *r, *s;												\
+		++b->n_keys;													\
+		r = b->root;													\
+		if (r->n == 2 * b->t - 1) {										\
+			++b->n_nodes;												\
+			s = (kbnode_t*)calloc(1, b->ilen);							\
+			b->root = s; s->is_internal = 1; s->n = 0;					\
+			__KB_PTR(b, s)[0] = r;										\
+			__kb_split_##name(b, s, 0, r);								\
+			r = s;														\
+		}																\
+		return __kb_putp_aux_##name(b, r, k);							\
+	}																	\
+	static inline void kb_put_##name(kbtree_##name##_t *b, const key_t k) \
+	{																	\
+		kb_putp_##name(b, &k);											\
+	}
+
+
+#define __KB_DEL(name, key_t)											\
+	static key_t __kb_delp_aux_##name(kbtree_##name##_t *b, kbnode_t *x, const key_t * __restrict k, int s) \
+	{																	\
+		int yn, zn, i, r = 0;											\
+		kbnode_t *xp, *y, *z;											\
+		key_t kp;														\
+		if (x == 0) return *k;											\
+		if (s) { /* s can only be 0, 1 or 2 */							\
+			r = x->is_internal == 0? 0 : s == 1? 1 : -1;				\
+			i = s == 1? x->n - 1 : -1;									\
+		} else i = __kb_getp_aux_##name(x, k, &r);						\
+		if (x->is_internal == 0) {										\
+			if (s == 2) ++i;											\
+			kp = __KB_KEY(key_t, x)[i];									\
+			memmove(__KB_KEY(key_t, x) + i, __KB_KEY(key_t, x) + i + 1, (x->n - i - 1) * sizeof(key_t)); \
+			--x->n;														\
+			return kp;													\
+		}																\
+		if (r == 0) {													\
+			if ((yn = __KB_PTR(b, x)[i]->n) >= b->t) {					\
+				xp = __KB_PTR(b, x)[i];									\
+				kp = __KB_KEY(key_t, x)[i];								\
+				__KB_KEY(key_t, x)[i] = __kb_delp_aux_##name(b, xp, 0, 1); \
+				return kp;												\
+			} else if ((zn = __KB_PTR(b, x)[i + 1]->n) >= b->t) {		\
+				xp = __KB_PTR(b, x)[i + 1];								\
+				kp = __KB_KEY(key_t, x)[i];								\
+				__KB_KEY(key_t, x)[i] = __kb_delp_aux_##name(b, xp, 0, 2); \
+				return kp;												\
+			} else if (yn == b->t - 1 && zn == b->t - 1) {				\
+				y = __KB_PTR(b, x)[i]; z = __KB_PTR(b, x)[i + 1];		\
+				__KB_KEY(key_t, y)[y->n++] = *k;						\
+				memmove(__KB_KEY(key_t, y) + y->n, __KB_KEY(key_t, z), z->n * sizeof(key_t)); \
+				if (y->is_internal) memmove(__KB_PTR(b, y) + y->n, __KB_PTR(b, z), (z->n + 1) * sizeof(void*)); \
+				y->n += z->n;											\
+				memmove(__KB_KEY(key_t, x) + i, __KB_KEY(key_t, x) + i + 1, (x->n - i - 1) * sizeof(key_t)); \
+				memmove(__KB_PTR(b, x) + i + 1, __KB_PTR(b, x) + i + 2, (x->n - i - 1) * sizeof(void*)); \
+				--x->n;													\
+				free(z);												\
+				return __kb_delp_aux_##name(b, y, k, s);				\
+			}															\
+		}																\
+		++i;															\
+		if ((xp = __KB_PTR(b, x)[i])->n == b->t - 1) {					\
+			if (i > 0 && (y = __KB_PTR(b, x)[i - 1])->n >= b->t) {		\
+				memmove(__KB_KEY(key_t, xp) + 1, __KB_KEY(key_t, xp), xp->n * sizeof(key_t)); \
+				if (xp->is_internal) memmove(__KB_PTR(b, xp) + 1, __KB_PTR(b, xp), (xp->n + 1) * sizeof(void*)); \
+				__KB_KEY(key_t, xp)[0] = __KB_KEY(key_t, x)[i - 1];		\
+				__KB_KEY(key_t, x)[i - 1] = __KB_KEY(key_t, y)[y->n - 1]; \
+				if (xp->is_internal) __KB_PTR(b, xp)[0] = __KB_PTR(b, y)[y->n]; \
+				--y->n; ++xp->n;										\
+			} else if (i < x->n && (y = __KB_PTR(b, x)[i + 1])->n >= b->t) { \
+				__KB_KEY(key_t, xp)[xp->n++] = __KB_KEY(key_t, x)[i];	\
+				__KB_KEY(key_t, x)[i] = __KB_KEY(key_t, y)[0];			\
+				if (xp->is_internal) __KB_PTR(b, xp)[xp->n] = __KB_PTR(b, y)[0]; \
+				--y->n;													\
+				memmove(__KB_KEY(key_t, y), __KB_KEY(key_t, y) + 1, y->n * sizeof(key_t)); \
+				if (y->is_internal) memmove(__KB_PTR(b, y), __KB_PTR(b, y) + 1, (y->n + 1) * sizeof(void*)); \
+			} else if (i > 0 && (y = __KB_PTR(b, x)[i - 1])->n == b->t - 1) { \
+				__KB_KEY(key_t, y)[y->n++] = __KB_KEY(key_t, x)[i - 1];	\
+				memmove(__KB_KEY(key_t, y) + y->n, __KB_KEY(key_t, xp), xp->n * sizeof(key_t));	\
+				if (y->is_internal) memmove(__KB_PTR(b, y) + y->n, __KB_PTR(b, xp), (xp->n + 1) * sizeof(void*)); \
+				y->n += xp->n;											\
+				memmove(__KB_KEY(key_t, x) + i - 1, __KB_KEY(key_t, x) + i, (x->n - i) * sizeof(key_t)); \
+				memmove(__KB_PTR(b, x) + i, __KB_PTR(b, x) + i + 1, (x->n - i) * sizeof(void*)); \
+				--x->n;													\
+				free(xp);												\
+				xp = y;													\
+			} else if (i < x->n && (y = __KB_PTR(b, x)[i + 1])->n == b->t - 1) { \
+				__KB_KEY(key_t, xp)[xp->n++] = __KB_KEY(key_t, x)[i];	\
+				memmove(__KB_KEY(key_t, xp) + xp->n, __KB_KEY(key_t, y), y->n * sizeof(key_t));	\
+				if (xp->is_internal) memmove(__KB_PTR(b, xp) + xp->n, __KB_PTR(b, y), (y->n + 1) * sizeof(void*)); \
+				xp->n += y->n;											\
+				memmove(__KB_KEY(key_t, x) + i, __KB_KEY(key_t, x) + i + 1, (x->n - i - 1) * sizeof(key_t)); \
+				memmove(__KB_PTR(b, x) + i + 1, __KB_PTR(b, x) + i + 2, (x->n - i - 1) * sizeof(void*)); \
+				--x->n;													\
+				free(y);												\
+			}															\
+		}																\
+		return __kb_delp_aux_##name(b, xp, k, s);						\
+	}																	\
+	static key_t kb_delp_##name(kbtree_##name##_t *b, const key_t * __restrict k) \
+	{																	\
+		kbnode_t *x;													\
+		key_t ret;														\
+		ret = __kb_delp_aux_##name(b, b->root, k, 0);					\
+		--b->n_keys;													\
+		if (b->root->n == 0 && b->root->is_internal) {					\
+			--b->n_nodes;												\
+			x = b->root;												\
+			b->root = __KB_PTR(b, x)[0];								\
+			free(x);													\
+		}																\
+		return ret;														\
+	}																	\
+	static inline key_t kb_del_##name(kbtree_##name##_t *b, const key_t k) \
+	{																	\
+		return kb_delp_##name(b, &k);									\
+	}
+
+#define __KB_ITR(name, key_t) \
+	static inline void kb_itr_first_##name(kbtree_##name##_t *b, kbitr_t *itr) \
+	{ \
+		itr->p = 0; \
+		if (b->n_keys == 0) return; \
+		itr->p = itr->stack; \
+		itr->p->x = b->root; itr->p->i = 0; \
+		while (itr->p->x->is_internal && __KB_PTR(b, itr->p->x)[0] != 0) { \
+			kbnode_t *x = itr->p->x; \
+			++itr->p; \
+			itr->p->x = __KB_PTR(b, x)[0]; itr->p->i = 0; \
+		} \
+	} \
+	static int kb_itr_get_##name(kbtree_##name##_t *b, const key_t * __restrict k, kbitr_t *itr) \
+	{ \
+		int i, r = 0; \
+		itr->p = itr->stack; \
+		itr->p->x = b->root; itr->p->i = 0; \
+		while (itr->p->x) { \
+			i = __kb_getp_aux_##name(itr->p->x, k, &r); \
+			if (i >= 0 && r == 0) return 0; \
+			if (itr->p->x->is_internal == 0) return -1; \
+			itr->p[1].x = __KB_PTR(b, itr->p->x)[i + 1]; \
+			itr->p[1].i = i; \
+			++itr->p; \
+		} \
+		return -1; \
+	} \
+	static inline int kb_itr_next_##name(kbtree_##name##_t *b, kbitr_t *itr) \
+	{ \
+		if (itr->p < itr->stack) return 0; \
+		for (;;) { \
+			++itr->p->i; \
+			while (itr->p->x && itr->p->i <= itr->p->x->n) { \
+				itr->p[1].i = 0; \
+				itr->p[1].x = itr->p->x->is_internal? __KB_PTR(b, itr->p->x)[itr->p->i] : 0; \
+				++itr->p; \
+			} \
+			--itr->p; \
+			if (itr->p < itr->stack) return 0; \
+			if (itr->p->x && itr->p->i < itr->p->x->n) return 1; \
+		} \
+	}
+
+#define KBTREE_INIT(name, key_t, __cmp)			\
+	__KB_TREE_T(name)							\
+	__KB_INIT(name, key_t)						\
+	__KB_GET_AUX1(name, key_t, __cmp)			\
+	__KB_GET(name, key_t)						\
+	__KB_INTERVAL(name, key_t)					\
+	__KB_PUT(name, key_t, __cmp)				\
+	__KB_DEL(name, key_t) \
+	__KB_ITR(name, key_t)
+
+#define KB_DEFAULT_SIZE 512
+
+#define kbtree_t(name) kbtree_##name##_t
+#define kb_init(name, s) kb_init_##name(s)
+#define kb_destroy(name, b) __kb_destroy(b)
+#define kb_get(name, b, k) kb_get_##name(b, k)
+#define kb_put(name, b, k) kb_put_##name(b, k)
+#define kb_del(name, b, k) kb_del_##name(b, k)
+#define kb_interval(name, b, k, l, u) kb_interval_##name(b, k, l, u)
+#define kb_getp(name, b, k) kb_getp_##name(b, k)
+#define kb_putp(name, b, k) kb_putp_##name(b, k)
+#define kb_delp(name, b, k) kb_delp_##name(b, k)
+#define kb_intervalp(name, b, k, l, u) kb_intervalp_##name(b, k, l, u)
+
+#define kb_itr_first(name, b, i) kb_itr_first_##name(b, i)
+#define kb_itr_get(name, b, k, i) kb_itr_get_##name(b, k, i)
+#define kb_itr_next(name, b, i) kb_itr_next_##name(b, i)
+#define kb_itr_key(type, itr) __KB_KEY(type, (itr)->p->x)[(itr)->p->i]
+#define kb_itr_valid(itr) ((itr)->p >= (itr)->stack)
+
+#define kb_size(b) ((b)->n_keys)
+
+#define kb_generic_cmp(a, b) (((b) < (a)) - ((a) < (b)))
+#define kb_str_cmp(a, b) strcmp(a, b)
+
+/* The following is *DEPRECATED*!!! Use the iterator interface instead! */
+
+typedef struct {
+	kbnode_t *x;
+	int i;
+} __kbstack_t;
+
+#define __kb_traverse(key_t, b, __func) do {							\
+		int __kmax = 8;													\
+		__kbstack_t *__kstack, *__kp;									\
+		__kp = __kstack = (__kbstack_t*)calloc(__kmax, sizeof(__kbstack_t)); \
+		__kp->x = (b)->root; __kp->i = 0;								\
+		for (;;) {														\
+			while (__kp->x && __kp->i <= __kp->x->n) {					\
+				if (__kp - __kstack == __kmax - 1) {					\
+					__kmax <<= 1;										\
+					__kstack = (__kbstack_t*)realloc(__kstack, __kmax * sizeof(__kbstack_t)); \
+					__kp = __kstack + (__kmax>>1) - 1;					\
+				}														\
+				(__kp+1)->i = 0; (__kp+1)->x = __kp->x->is_internal? __KB_PTR(b, __kp->x)[__kp->i] : 0; \
+				++__kp;													\
+			}															\
+			--__kp;														\
+			if (__kp >= __kstack) {										\
+				if (__kp->x && __kp->i < __kp->x->n) __func(&__KB_KEY(key_t, __kp->x)[__kp->i]); \
+				++__kp->i;												\
+			} else break;												\
+		}																\
+		free(__kstack);													\
+	} while (0)
+
+#define __kb_get_first(key_t, b, ret) do {	\
+		kbnode_t *__x = (b)->root;			\
+		while (__KB_PTR(b, __x)[0] != 0)	\
+			__x = __KB_PTR(b, __x)[0];		\
+		(ret) = __KB_KEY(key_t, __x)[0];	\
+	} while (0)
+
+#endif
diff --git a/ext/klib/kdq.h b/ext/klib/kdq.h
new file mode 100644
index 0000000..edd55b5
--- /dev/null
+++ b/ext/klib/kdq.h
@@ -0,0 +1,128 @@
+#ifndef __AC_KDQ_H
+#define __AC_KDQ_H
+
+#include <stdlib.h>
+#include <string.h>
+
+#define __KDQ_TYPE(type) \
+	typedef struct { \
+		size_t front:58, bits:6, count, mask; \
+		type *a; \
+	} kdq_##type##_t;
+
+#define kdq_t(type) kdq_##type##_t
+#define kdq_size(q) ((q)->count)
+#define kdq_first(q) ((q)->a[(q)->front])
+#define kdq_last(q) ((q)->a[((q)->front + (q)->count - 1) & (q)->mask])
+#define kdq_at(q, i) ((q)->a[((q)->front + (i)) & (q)->mask])
+
+#define __KDQ_IMPL(type, SCOPE) \
+	SCOPE kdq_##type##_t *kdq_init_##type() \
+	{ \
+		kdq_##type##_t *q; \
+		q = (kdq_##type##_t*)calloc(1, sizeof(kdq_##type##_t)); \
+		q->bits = 2, q->mask = (1ULL<<q->bits) - 1; \
+		q->a = (type*)malloc((1<<q->bits) * sizeof(type)); \
+		return q; \
+	} \
+	SCOPE void kdq_destroy_##type(kdq_##type##_t *q) \
+	{ \
+		if (q == 0) return; \
+		free(q->a); free(q); \
+	} \
+	SCOPE int kdq_resize_##type(kdq_##type##_t *q, int new_bits) \
+	{ \
+		size_t new_size = 1ULL<<new_bits, old_size = 1ULL<<q->bits; \
+		if (new_size < q->count) { /* not big enough */ \
+			int i; \
+			for (i = 0; i < 64; ++i) \
+				if (1ULL<<i > q->count) break; \
+			new_bits = i, new_size = 1ULL<<new_bits; \
+		} \
+		if (new_bits == q->bits) return q->bits; /* unchanged */ \
+		if (new_bits > q->bits) q->a = (type*)realloc(q->a, (1ULL<<new_bits) * sizeof(type)); \
+		if (q->front + q->count <= old_size) { /* unwrapped */ \
+			if (q->front + q->count > new_size) /* only happens for shrinking */ \
+				memmove(q->a, q->a + new_size, (q->front + q->count - new_size) * sizeof(type)); \
+		} else { /* wrapped */ \
+			memmove(q->a + (new_size - (old_size - q->front)), q->a + q->front, (old_size - q->front) * sizeof(type)); \
+			q->front = new_size - (old_size - q->front); \
+		} \
+		q->bits = new_bits, q->mask = (1ULL<<q->bits) - 1; \
+		if (new_bits < q->bits) q->a = (type*)realloc(q->a, (1ULL<<new_bits) * sizeof(type)); \
+		return q->bits; \
+	} \
+	SCOPE type *kdq_pushp_##type(kdq_##type##_t *q) \
+	{ \
+		if (q->count == 1ULL<<q->bits) kdq_resize_##type(q, q->bits + 1); \
+		return &q->a[((q->count++) + q->front) & (q)->mask]; \
+	} \
+	SCOPE void kdq_push_##type(kdq_##type##_t *q, type v) \
+	{ \
+		if (q->count == 1ULL<<q->bits) kdq_resize_##type(q, q->bits + 1); \
+		q->a[((q->count++) + q->front) & (q)->mask] = v; \
+	} \
+	SCOPE type *kdq_unshiftp_##type(kdq_##type##_t *q) \
+	{ \
+		if (q->count == 1ULL<<q->bits) kdq_resize_##type(q, q->bits + 1); \
+		++q->count; \
+		q->front = q->front? q->front - 1 : (1ULL<<q->bits) - 1; \
+		return &q->a[q->front]; \
+	} \
+	SCOPE void kdq_unshift_##type(kdq_##type##_t *q, type v) \
+	{ \
+		type *p; \
+		p = kdq_unshiftp_##type(q); \
+		*p = v; \
+	} \
+	SCOPE type *kdq_pop_##type(kdq_##type##_t *q) \
+	{ \
+		return q->count? &q->a[((--q->count) + q->front) & q->mask] : 0; \
+	} \
+	SCOPE type *kdq_shift_##type(kdq_##type##_t *q) \
+	{ \
+		type *d = 0; \
+		if (q->count == 0) return 0; \
+		d = &q->a[q->front++]; \
+		q->front &= q->mask; \
+		--q->count; \
+		return d; \
+	}
+
+#define KDQ_INIT2(type, SCOPE) \
+	__KDQ_TYPE(type) \
+	__KDQ_IMPL(type, SCOPE)
+
+#ifndef klib_unused
+#if (defined __clang__ && __clang_major__ >= 3) || (defined __GNUC__ && __GNUC__ >= 3)
+#define klib_unused __attribute__ ((__unused__))
+#else
+#define klib_unused
+#endif
+#endif /* klib_unused */
+
+#define KDQ_INIT(type) KDQ_INIT2(type, static inline klib_unused)
+
+#define KDQ_DECLARE(type) \
+	__KDQ_TYPE(type) \
+	kdq_##type##_t *kdq_init_##type(); \
+	void kdq_destroy_##type(kdq_##type##_t *q); \
+	int kdq_resize_##type(kdq_##type##_t *q, int new_bits); \
+	type *kdq_pushp_##type(kdq_##type##_t *q); \
+	void kdq_push_##type(kdq_##type##_t *q, type v); \
+	type *kdq_unshiftp_##type(kdq_##type##_t *q); \
+	void kdq_unshift_##type(kdq_##type##_t *q, type v); \
+	type *kdq_pop_##type(kdq_##type##_t *q); \
+	type *kdq_shift_##type(kdq_##type##_t *q);
+
+#define kdq_init(type) kdq_init_##type()
+#define kdq_destroy(type, q) kdq_destroy_##type(q)
+#define kdq_resize(type, q, new_bits) kdq_resize_##type(q, new_bits)
+#define kdq_pushp(type, q) kdq_pushp_##type(q)
+#define kdq_push(type, q, v) kdq_push_##type(q, v)
+#define kdq_pop(type, q) kdq_pop_##type(q)
+#define kdq_unshiftp(type, q) kdq_unshiftp_##type(q)
+#define kdq_unshift(type, q, v) kdq_unshift_##type(q, v)
+#define kdq_shift(type, q) kdq_shift_##type(q)
+
+#endif
diff --git a/ext/klib/keigen.c b/ext/klib/keigen.c
new file mode 100644
index 0000000..5dd2ddc
--- /dev/null
+++ b/ext/klib/keigen.c
@@ -0,0 +1,186 @@
+#include <math.h>
+#include <stdlib.h>
+#include "keigen.h"
+
+void ke_core_strq(int n, double *q, double *b, double *c)
+{
+	int	i, j, k, u, v;
+	double h, f, g, h2;
+	for (i = n - 1; i >= 1; i--) {
+		h = 0.0;
+		if (i > 1)
+			for (k = 0; k < i; k++) {
+				u = i * n + k;
+				h = h + q[u] * q[u];
+			}
+		if (h + 1.0 == 1.0) {
+			c[i] = 0.0;
+			if (i == 1)
+				c[i] = q[i * n + i - 1];
+			b[i] = 0.0;
+		} else {
+			c[i] = sqrt(h);
+			u = i * n + i - 1;
+			if (q[u] > 0.0)
+				c[i] = -c[i];
+			h = h - q[u] * c[i];
+			q[u] = q[u] - c[i];
+			f = 0.0;
+			for (j = 0; j < i; j++) {
+				q[j * n + i] = q[i * n + j] / h;
+				g = 0.0;
+				for (k = 0; k <= j; k++)
+					g = g + q[j * n + k] * q[i * n + k];
+				if (j + 1 < i)
+					for (k = j + 1; k <= i - 1; k++)
+						g = g + q[k * n + j] * q[i * n + k];
+				c[j] = g / h;
+				f = f + g * q[j * n + i];
+			}
+			h2 = f / (h + h);
+			for (j = 0; j < i; j++) {
+				f = q[i * n + j];
+				g = c[j] - h2 * f;
+				c[j] = g;
+				for (k = 0; k <= j; k++) {
+					u = j * n + k;
+					q[u] = q[u] - f * c[k] - g * q[i * n + k];
+				}
+			}
+			b[i] = h;
+		}
+	}
+	for (i = 0; i < n - 1; i++)
+		c[i] = c[i + 1];
+	c[n - 1] = 0.0;
+	b[0] = 0.0;
+	for (i = 0; i < n; i++) {
+		if (b[i] != 0.0 && i - 1 >= 0)
+			for (j = 0; j < i; j++) {
+				g = 0.0;
+				for (k = 0; k < i; k++)
+					g = g + q[i * n + k] * q[k * n + j];
+				for (k = 0; k < i; k++) {
+					u = k * n + j;
+					q[u] = q[u] - g * q[k * n + i];
+				}
+			}
+		u = i * n + i;
+		b[i] = q[u];
+		q[u] = 1.0;
+		if (i - 1 >= 0)
+			for (j = 0; j < i; j++) {
+				q[i * n + j] = 0.0;
+				q[j * n + i] = 0.0;
+			}
+	}
+}
+
+int ke_core_sstq(int n, double *b, double *c, double *q, int cal_ev, double eps, int l)
+{
+	int i, j, k, m, it, u, v;
+	double d, f, h, g, p, r, e, s;
+	c[n - 1] = 0.0;
+	d = 0.0;
+	f = 0.0;
+	for (j = 0; j < n; j++) {
+		it = 0;
+		h = eps * (fabs(b[j]) + fabs(c[j]));
+		if (h > d)
+			d = h;
+		m = j;
+		while (m < n && fabs(c[m]) > d)
+			m = m + 1;
+		if (m != j) {
+			do {
+				if (it == l) return KE_EXCESS_ITER;
+				it = it + 1;
+				g = b[j];
+				p = (b[j + 1] - g) / (2.0 * c[j]);
+				r = sqrt(p * p + 1.0);
+				if (p >= 0.0)
+					b[j] = c[j] / (p + r);
+				else
+					b[j] = c[j] / (p - r);
+				h = g - b[j];
+				for (i = j + 1; i < n; i++)
+					b[i] = b[i] - h;
+				f = f + h;
+				p = b[m];
+				e = 1.0;
+				s = 0.0;
+				for (i = m - 1; i >= j; i--) {
+					g = e * c[i];
+					h = e * p;
+					if (fabs(p) >= fabs(c[i])) {
+						e = c[i] / p;
+						r = sqrt(e * e + 1.0);
+						c[i + 1] = s * p * r;
+						s = e / r;
+						e = 1.0 / r;
+					} else {
+						e = p / c[i];
+						r = sqrt(e * e + 1.0);
+						c[i + 1] = s * c[i] * r;
+						s = 1.0 / r;
+						e = e / r;
+					}
+					p = e * b[i] - s * g;
+					b[i + 1] = h + s * (e * g + s * b[i]);
+					if (cal_ev) {
+						for (k = 0; k < n; k++) {
+							u = k * n + i + 1;
+							v = u - 1;
+							h = q[u];
+							q[u] = s * q[v] + e * h;
+							q[v] = e * q[v] - s * h;
+						}
+					}
+				}
+				c[j] = s * p;
+				b[j] = e * p;
+			}
+			while (fabs(c[j]) > d);
+		}
+		b[j] = b[j] + f;
+	}
+	for (i = 0; i < n; i++) {
+		k = i;
+		p = b[i];
+		if (i + 1 < n) {
+			j = i + 1;
+			while (j < n && b[j] <= p) {
+				k = j;
+				p = b[j];
+				j = j + 1;
+			}
+		}
+		if (k != i) {
+			b[k] = b[i];
+			b[i] = p;
+			for (j = 0; j < n; j++) {
+				u = j * n + i;
+				v = j * n + k;
+				p = q[u];
+				q[u] = q[v];
+				q[v] = p;
+			}
+		}
+	}
+	return 0;
+}
+
+#define MALLOC(type, size) ((type*)malloc(size * sizeof(type)))
+
+int ke_eigen_sd(int n, double *a, double *v, int cal_ev, double eps, int max_iter)
+{
+	double *c;
+	int r;
+	if (1.0 + eps <= 1.0) eps = 1e-7;
+	if (max_iter <= 0) max_iter = 50;
+	c = MALLOC(double, n);
+	ke_core_strq(n, a, v, c);
+	r = ke_core_sstq(n, v, c, a, cal_ev, eps, max_iter);
+	free(c);
+	return r;
+}
diff --git a/ext/klib/keigen.h b/ext/klib/keigen.h
new file mode 100644
index 0000000..14aa53d
--- /dev/null
+++ b/ext/klib/keigen.h
@@ -0,0 +1,53 @@
+#ifndef KEIGEN_H
+#define KEIGEN_H
+
+#define KE_EXCESS_ITER (-1)
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/**
+ * Compute eigenvalues/vectors for a dense symmetric matrix
+ *
+ * @param n       dimension
+ * @param a       input matrix and eigenvalues on return ([n*n]; in & out)
+ * @param v       eigenvalues ([n]; out)
+ * @param cal_ev  compute eigenvectos or not (faster without vectors)
+ * @param eps     precision (<=0 for default)
+ * @param max_itr max iteration (<=0 for detaul)
+ *
+ * @return 0 on success; KE_EXCESS_ITER if too many iterations
+ */
+int ke_eigen_sd(int n, double *a, double *v, int cal_ev, double eps, int max_iter);
+
+/**
+ * Transform a real symmetric matrix to a tridiagonal matrix
+ *
+ * @param n       dimension
+ * @param q       input matrix and transformation matrix ([n*n]; in & out)
+ * @param b       diagonal ([n]; out)
+ * @param c       subdiagonal ([n]; out)
+ */
+void ke_core_strq(int n, double *q, double *b, double *c);
+
+/**
+ * Compute eigenvalues and eigenvectors for a tridiagonal matrix
+ *
+ * @param n       dimension
+ * @param b       diagonal and eigenvalues on return ([n]; in & out)
+ * @param c       subdiagonal ([n]; in)
+ * @param q       transformation matrix and eigenvectors on return ([n*n]; in & out)
+ * @param cal_ev  compute eigenvectors or not (faster without vectors)
+ * @param eps     precision
+ * @param l       max iterations
+ *
+ * @return 0 on success; KE_EXCESS_ITER if too many iterations
+ */
+int ke_core_sstq(int n, double *b, double *c, double *q, int cal_ev, double eps, int l);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif
diff --git a/ext/klib/ketopt.h b/ext/klib/ketopt.h
new file mode 100644
index 0000000..73dd2dc
--- /dev/null
+++ b/ext/klib/ketopt.h
@@ -0,0 +1,120 @@
+#ifndef KETOPT_H
+#define KETOPT_H
+
+#include <string.h> /* for strchr() and strncmp() */
+
+#define ko_no_argument       0
+#define ko_required_argument 1
+#define ko_optional_argument 2
+
+typedef struct {
+	int ind;   /* equivalent to optind */
+	int opt;   /* equivalent to optopt */
+	char *arg; /* equivalent to optarg */
+	int longidx; /* index of a long option; or -1 if short */
+	/* private variables not intended for external uses */
+	int i, pos, n_args;
+} ketopt_t;
+
+typedef struct {
+	char *name;
+	int has_arg;
+	int val;
+} ko_longopt_t;
+
+static ketopt_t KETOPT_INIT = { 1, 0, 0, -1, 1, 0, 0 };
+
+static void ketopt_permute(char *argv[], int j, int n) /* move argv[j] over n elements to the left */
+{
+	int k;
+	char *p = argv[j];
+	for (k = 0; k < n; ++k)
+		argv[j - k] = argv[j - k - 1];
+	argv[j - k] = p;
+}
+
+/**
+ * Parse command-line options and arguments
+ *
+ * This fuction has a similar interface to GNU's getopt_long(). Each call
+ * parses one option and returns the option name.  s->arg points to the option
+ * argument if present. The function returns -1 when all command-line arguments
+ * are parsed. In this case, s->ind is the index of the first non-option
+ * argument.
+ *
+ * @param s         status; shall be initialized to KETOPT_INIT on the first call
+ * @param argc      length of argv[]
+ * @param argv      list of command-line arguments; argv[0] is ignored
+ * @param permute   non-zero to move options ahead of non-option arguments
+ * @param ostr      option string
+ * @param longopts  long options
+ *
+ * @return ASCII for a short option; ko_longopt_t::val for a long option; -1 if
+ *         argv[] is fully processed; '?' for an unknown option or an ambiguous
+ *         long option; ':' if an option argument is missing
+ */
+static int ketopt(ketopt_t *s, int argc, char *argv[], int permute, const char *ostr, const ko_longopt_t *longopts)
+{
+	int opt = -1, i0, j;
+	if (permute) {
+		while (s->i < argc && (argv[s->i][0] != '-' || argv[s->i][1] == '\0'))
+			++s->i, ++s->n_args;
+	}
+	s->arg = 0, s->longidx = -1, i0 = s->i;
+	if (s->i >= argc || argv[s->i][0] != '-' || argv[s->i][1] == '\0') {
+		s->ind = s->i - s->n_args;
+		return -1;
+	}
+	if (argv[s->i][0] == '-' && argv[s->i][1] == '-') { /* "--" or a long option */
+		if (argv[s->i][2] == '\0') { /* a bare "--" */
+			ketopt_permute(argv, s->i, s->n_args);
+			++s->i, s->ind = s->i - s->n_args;
+			return -1;
+		}
+		s->opt = 0, opt = '?', s->pos = -1;
+		if (longopts) { /* parse long options */
+			int k, n_exact = 0, n_partial = 0;
+			const ko_longopt_t *o = 0, *o_exact = 0, *o_partial = 0;
+			for (j = 2; argv[s->i][j] != '\0' && argv[s->i][j] != '='; ++j) {} /* find the end of the option name */
+			for (k = 0; longopts[k].name != 0; ++k)
+				if (strncmp(&argv[s->i][2], longopts[k].name, j - 2) == 0) {
+					if (longopts[k].name[j - 2] == 0) ++n_exact, o_exact = &longopts[k];
+					else ++n_partial, o_partial = &longopts[k];
+				}
+			if (n_exact > 1 || (n_exact == 0 && n_partial > 1)) return '?';
+			o = n_exact == 1? o_exact : n_partial == 1? o_partial : 0;
+			if (o) {
+				s->opt = opt = o->val, s->longidx = o - longopts;
+				if (argv[s->i][j] == '=') s->arg = &argv[s->i][j + 1];
+				if (o->has_arg == 1 && argv[s->i][j] == '\0') {
+					if (s->i < argc - 1) s->arg = argv[++s->i];
+					else opt = ':'; /* missing option argument */
+				}
+			}
+		}
+	} else { /* a short option */
+		const char *p;
+		if (s->pos == 0) s->pos = 1;
+		opt = s->opt = argv[s->i][s->pos++];
+		p = strchr((char*)ostr, opt);
+		if (p == 0) {
+			opt = '?'; /* unknown option */
+		} else if (p[1] == ':') {
+			if (argv[s->i][s->pos] == 0) {
+				if (s->i < argc - 1) s->arg = argv[++s->i];
+				else opt = ':'; /* missing option argument */
+			} else s->arg = &argv[s->i][s->pos];
+			s->pos = -1;
+		}
+	}
+	if (s->pos < 0 || argv[s->i][s->pos] == 0) {
+		++s->i, s->pos = 0;
+		if (s->n_args > 0) /* permute */
+			for (j = i0; j < s->i; ++j)
+				ketopt_permute(argv, j, s->n_args);
+	}
+	s->ind = s->i - s->n_args;
+	return opt;
+}
+
+#endif
diff --git a/ext/klib/kexpr.c b/ext/klib/kexpr.c
new file mode 100644
index 0000000..79f110f
--- /dev/null
+++ b/ext/klib/kexpr.c
@@ -0,0 +1,586 @@
+#include <string.h>
+#include <stdlib.h>
+#include <stdint.h>
+#include <assert.h>
+#include <stdio.h>
+#include <ctype.h>
+#include <math.h>
+#include "kexpr.h"
+
+/***************
+ * Definitions *
+ ***************/
+
+#define KEO_NULL  0
+#define KEO_POS   1
+#define KEO_NEG   2
+#define KEO_BNOT  3
+#define KEO_LNOT  4
+#define KEO_POW   5
+#define KEO_MUL   6
+#define KEO_DIV   7
+#define KEO_IDIV  8
+#define KEO_MOD   9
+#define KEO_ADD  10
+#define KEO_SUB  11
+#define KEO_LSH  12
+#define KEO_RSH  13
+#define KEO_LT   14
+#define KEO_LE   15
+#define KEO_GT   16
+#define KEO_GE   17
+#define KEO_EQ   18
+#define KEO_NE   19
+#define KEO_BAND 20
+#define KEO_BXOR 21
+#define KEO_BOR  22
+#define KEO_LAND 23
+#define KEO_LOR  24
+
+#define KET_NULL  0
+#define KET_VAL   1
+#define KET_OP    2
+#define KET_FUNC  3
+
+#define KEF_NULL  0
+#define KEF_REAL  1
+
+struct ke1_s;
+
+typedef struct ke1_s {
+	uint32_t ttype:16, vtype:10, assigned:1, user_func:5; // ttype: token type; vtype: value type
+	int32_t op:8, n_args:24; // op: operator, n_args: number of arguments
+	char *name; // variable name or function name
+	union {
+		void (*builtin)(struct ke1_s *a, struct ke1_s *b); // execution function
+		double (*real_func1)(double);
+		double (*real_func2)(double, double);
+	} f;
+	double r;
+	int64_t i;
+	char *s;
+} ke1_t;
+
+static int ke_op[25] = {
+	0,
+	1<<1|1, 1<<1|1, 1<<1|1, 1<<1|1, // unary operators
+	2<<1|1, // pow()
+	3<<1, 3<<1, 3<<1, 3<<1, // * / // %
+	4<<1, 4<<1, // + and -
+	5<<1, 5<<1, // << and >>
+	6<<1, 6<<1, 6<<1, 6<<1, // < > <= >=
+	7<<1, 7<<1, // == !=
+	8<<1, // &
+	9<<1, // ^
+	10<<1,// |
+	11<<1,// &&
+	12<<1 // ||
+};
+
+static const char *ke_opstr[] = {
+	"",
+	"+(1)", "-(1)", "~", "!",
+	"**",
+	"*", "/", "//", "%",
+	"+", "-",
+	"<<", ">>",
+	"<", "<=", ">", ">=",
+	"==", "!=",
+	"&",
+	"^",
+	"|",
+	"&&",
+	"||"
+};
+
+struct kexpr_s {
+	int n;
+	ke1_t *e;
+};
+
+/**********************
+ * Operator functions *
+ **********************/
+
+#define KE_GEN_CMP(_type, _op) \
+	static void ke_op_##_type(ke1_t *p, ke1_t *q) { \
+		if (p->vtype == KEV_STR && q->vtype == KEV_STR) p->i = (strcmp(p->s, q->s) _op 0); \
+		else p->i = p->vtype == KEV_REAL || q->vtype == KEV_REAL? (p->r _op q->r) : (p->i _op q->i); \
+		p->r = (double)p->i; \
+		p->vtype = KEV_INT; \
+	}
+
+KE_GEN_CMP(KEO_LT, <)
+KE_GEN_CMP(KEO_LE, <=)
+KE_GEN_CMP(KEO_GT, >)
+KE_GEN_CMP(KEO_GE, >=)
+KE_GEN_CMP(KEO_EQ, ==)
+KE_GEN_CMP(KEO_NE, !=)
+
+#define KE_GEN_BIN_INT(_type, _op) \
+	static void ke_op_##_type(ke1_t *p, ke1_t *q) { \
+		p->i _op q->i; p->r = (double)p->i; \
+		p->vtype = KEV_INT; \
+	}
+
+KE_GEN_BIN_INT(KEO_BAND, &=)
+KE_GEN_BIN_INT(KEO_BOR, |=)
+KE_GEN_BIN_INT(KEO_BXOR, ^=)
+KE_GEN_BIN_INT(KEO_LSH, <<=)
+KE_GEN_BIN_INT(KEO_RSH, >>=)
+KE_GEN_BIN_INT(KEO_MOD, %=)
+KE_GEN_BIN_INT(KEO_IDIV, /=)
+
+#define KE_GEN_BIN_BOTH(_type, _op) \
+	static void ke_op_##_type(ke1_t *p, ke1_t *q) { \
+		p->i _op q->i; p->r _op q->r; \
+		p->vtype = p->vtype == KEV_REAL || q->vtype == KEV_REAL? KEV_REAL : KEV_INT; \
+	}
+
+KE_GEN_BIN_BOTH(KEO_ADD, +=)
+KE_GEN_BIN_BOTH(KEO_SUB, -=)
+KE_GEN_BIN_BOTH(KEO_MUL, *=)
+
+static void ke_op_KEO_DIV(ke1_t *p, ke1_t *q)  { p->r /= q->r, p->i = (int64_t)(p->r + .5); p->vtype = KEV_REAL; }
+static void ke_op_KEO_LAND(ke1_t *p, ke1_t *q) { p->i = (p->i && q->i); p->r = p->i; p->vtype = KEV_INT; }
+static void ke_op_KEO_LOR(ke1_t *p, ke1_t *q)  { p->i = (p->i || q->i); p->r = p->i; p->vtype = KEV_INT; }
+static void ke_op_KEO_POW(ke1_t *p, ke1_t *q)  { p->r = pow(p->r, q->r), p->i = (int64_t)(p->r + .5); p->vtype = p->vtype == KEV_REAL || q->vtype == KEV_REAL? KEV_REAL : KEV_INT; }
+static void ke_op_KEO_BNOT(ke1_t *p, ke1_t *q) { p->i = ~p->i; p->r = (double)p->i; p->vtype = KEV_INT; }
+static void ke_op_KEO_LNOT(ke1_t *p, ke1_t *q) { p->i = !p->i; p->r = (double)p->i; p->vtype = KEV_INT; }
+static void ke_op_KEO_POS(ke1_t *p, ke1_t *q)  { } // do nothing
+static void ke_op_KEO_NEG(ke1_t *p, ke1_t *q)  { p->i = -p->i, p->r = -p->r; }
+
+static void ke_func1_abs(ke1_t *p, ke1_t *q) { if (p->vtype == KEV_INT) p->i = abs(p->i), p->r = (double)p->i; else p->r = fabs(p->r), p->i = (int64_t)(p->r + .5); }
+
+/**********
+ * Parser *
+ **********/
+
+static inline char *mystrndup(const char *src, int n)
+{
+	char *dst;
+	dst = (char*)calloc(n + 1, 1);
+	strncpy(dst, src, n);
+	return dst;
+}
+
+// parse a token except "(", ")" and ","
+static ke1_t ke_read_token(char *p, char **r, int *err, int last_is_val) // it doesn't parse parentheses
+{
+	char *q = p;
+	ke1_t e;
+	memset(&e, 0, sizeof(ke1_t));
+	if (isalpha(*p) || *p == '_') { // a variable or a function
+		for (; *p && (*p == '_' || isalnum(*p)); ++p);
+		if (*p == '(') e.ttype = KET_FUNC, e.n_args = 1;
+		else e.ttype = KET_VAL, e.vtype = KEV_REAL;
+		e.name = mystrndup(q, p - q);
+		e.i = 0, e.r = 0.;
+		*r = p;
+	} else if (isdigit(*p) || *p == '.') { // a number
+		long x;
+		double y;
+		char *pp;
+		e.ttype = KET_VAL;
+		y = strtod(q, &p);
+		x = strtol(q, &pp, 0); // FIXME: check int/double parsing errors
+		if (q == p && q == pp) { // parse error
+			*err |= KEE_NUM;
+		} else if (p > pp) { // has "." or "[eE]"; then it is a real number
+			e.vtype = KEV_REAL;
+			e.i = (int64_t)(y + .5), e.r = y;
+			*r = p;
+		} else {
+			e.vtype = KEV_INT;
+			e.i = x, e.r = y;
+			*r = pp;
+		}
+	} else if (*p == '"' || *p == '\'') { // a string value
+		int c = *p;
+		for (++p; *p && *p != c; ++p)
+			if (*p == '\\') ++p; // escaping
+		if (*p == c) {
+			e.ttype = KET_VAL, e.vtype = KEV_STR;
+			e.s = mystrndup(q + 1, p - q - 1);
+			*r = p + 1;
+		} else *err |= KEE_UNQU, *r = p;
+	} else { // an operator
+		e.ttype = KET_OP;
+		if (*p == '*' && p[1] == '*') e.op = KEO_POW, e.f.builtin = ke_op_KEO_POW, e.n_args = 2, *r = q + 2;
+		else if (*p == '*') e.op = KEO_MUL, e.f.builtin = ke_op_KEO_MUL, e.n_args = 2, *r = q + 1; // FIXME: NOT working for unary operators
+		else if (*p == '/' && p[1] == '/') e.op = KEO_IDIV, e.f.builtin = ke_op_KEO_IDIV, e.n_args = 2, *r = q + 2;
+		else if (*p == '/') e.op = KEO_DIV, e.f.builtin = ke_op_KEO_DIV, e.n_args = 2, *r = q + 1;
+		else if (*p == '%') e.op = KEO_MOD, e.f.builtin = ke_op_KEO_MOD, e.n_args = 2, *r = q + 1;
+		else if (*p == '+') {
+			if (last_is_val) e.op = KEO_ADD, e.f.builtin = ke_op_KEO_ADD, e.n_args = 2;
+			else e.op = KEO_POS, e.f.builtin = ke_op_KEO_POS, e.n_args = 1;
+			*r = q + 1;
+		} else if (*p == '-') {
+			if (last_is_val) e.op = KEO_SUB, e.f.builtin = ke_op_KEO_SUB, e.n_args = 2;
+			else e.op = KEO_NEG, e.f.builtin = ke_op_KEO_NEG, e.n_args = 1;
+			*r = q + 1;
+		} else if (*p == '=' && p[1] == '=') e.op = KEO_EQ, e.f.builtin = ke_op_KEO_EQ, e.n_args = 2, *r = q + 2;
+		else if (*p == '!' && p[1] == '=') e.op = KEO_NE, e.f.builtin = ke_op_KEO_NE, e.n_args = 2, *r = q + 2;
+		else if (*p == '<' && p[1] == '>') e.op = KEO_NE, e.f.builtin = ke_op_KEO_NE, e.n_args = 2, *r = q + 2;
+		else if (*p == '>' && p[1] == '=') e.op = KEO_GE, e.f.builtin = ke_op_KEO_GE, e.n_args = 2, *r = q + 2;
+		else if (*p == '<' && p[1] == '=') e.op = KEO_LE, e.f.builtin = ke_op_KEO_LE, e.n_args = 2, *r = q + 2;
+		else if (*p == '>' && p[1] == '>') e.op = KEO_RSH, e.f.builtin = ke_op_KEO_RSH, e.n_args = 2, *r = q + 2;
+		else if (*p == '<' && p[1] == '<') e.op = KEO_LSH, e.f.builtin = ke_op_KEO_LSH, e.n_args = 2, *r = q + 2;
+		else if (*p == '>') e.op = KEO_GT, e.f.builtin = ke_op_KEO_GT, e.n_args = 2, *r = q + 1;
+		else if (*p == '<') e.op = KEO_LT, e.f.builtin = ke_op_KEO_LT, e.n_args = 2, *r = q + 1;
+		else if (*p == '|' && p[1] == '|') e.op = KEO_LOR, e.f.builtin = ke_op_KEO_LOR, e.n_args = 2, *r = q + 2;
+		else if (*p == '&' && p[1] == '&') e.op = KEO_LAND, e.f.builtin = ke_op_KEO_LAND, e.n_args = 2, *r = q + 2;
+		else if (*p == '|') e.op = KEO_BOR, e.f.builtin = ke_op_KEO_BOR, e.n_args = 2, *r = q + 1;
+		else if (*p == '&') e.op = KEO_BAND, e.f.builtin = ke_op_KEO_BAND, e.n_args = 2, *r = q + 1;
+		else if (*p == '^') e.op = KEO_BXOR, e.f.builtin = ke_op_KEO_BXOR, e.n_args = 2, *r = q + 1;
+		else if (*p == '~') e.op = KEO_BNOT, e.f.builtin = ke_op_KEO_BNOT, e.n_args = 1, *r = q + 1;
+		else if (*p == '!') e.op = KEO_LNOT, e.f.builtin = ke_op_KEO_LNOT, e.n_args = 1, *r = q + 1;
+		else e.ttype = KET_NULL, *err |= KEE_UNOP;
+	}
+	return e;
+}
+
+static inline ke1_t *push_back(ke1_t **a, int *n, int *m)
+{
+	if (*n == *m) {
+		int old_m = *m;
+		*m = *m? *m<<1 : 8;
+		*a = (ke1_t*)realloc(*a, *m * sizeof(ke1_t));
+		memset(*a + old_m, 0, (*m - old_m) * sizeof(ke1_t));
+	}
+	return &(*a)[(*n)++];
+}
+
+static ke1_t *ke_parse_core(const char *_s, int *_n, int *err)
+{
+	char *s, *p, *q;
+	int n_out, m_out, n_op, m_op, last_is_val = 0;
+	ke1_t *out, *op, *t, *u;
+
+	*err = 0; *_n = 0;
+	s = strdup(_s); // make a copy
+	for (p = q = s; *p; ++p) // squeeze out spaces
+		if (!isspace(*p)) *q++ = *p;
+	*q++ = 0;
+
+	out = op = 0;
+	n_out = m_out = n_op = m_op = 0;
+	p = s;
+	while (*p) {
+		if (*p == '(') {
+			t = push_back(&op, &n_op, &m_op); // push to the operator stack
+			t->op = -1, t->ttype = KET_NULL; // ->op < 0 for a left parenthsis
+			++p;
+		} else if (*p == ')') {
+			while (n_op > 0 && op[n_op-1].op >= 0) { // move operators to the output until we see a left parenthesis
+				u = push_back(&out, &n_out, &m_out);
+				*u = op[--n_op];
+			}
+			if (n_op == 0) { // error: extra right parenthesis
+				*err |= KEE_UNRP;
+				break;
+			} else --n_op; // pop out '('
+			if (n_op > 0 && op[n_op-1].ttype == KET_FUNC) { // the top of the operator stack is a function
+				u = push_back(&out, &n_out, &m_out); // move it to the output
+				*u = op[--n_op];
+				if (u->n_args == 1 && strcmp(u->name, "abs") == 0) u->f.builtin = ke_func1_abs;
+			}
+			++p;
+		} else if (*p == ',') { // function arguments separator
+			while (n_op > 0 && op[n_op-1].op >= 0) {
+				u = push_back(&out, &n_out, &m_out);
+				*u = op[--n_op];
+			}
+			if (n_op < 2 || op[n_op-2].ttype != KET_FUNC) { // we should at least see a function and a left parenthesis
+				*err |= KEE_FUNC;
+				break;
+			}
+			++op[n_op-2].n_args;
+			++p;
+		} else { // output-able token
+			ke1_t v;
+			v = ke_read_token(p, &p, err, last_is_val);
+			if (*err) break;
+			if (v.ttype == KET_VAL) {
+				u = push_back(&out, &n_out, &m_out);
+				*u = v;
+				last_is_val = 1;
+			} else if (v.ttype == KET_FUNC) {
+				t = push_back(&op, &n_op, &m_op);
+				*t = v;
+				last_is_val = 0;
+			} else if (v.ttype == KET_OP) {
+				int oi = ke_op[v.op];
+				while (n_op > 0 && op[n_op-1].ttype == KET_OP) {
+					int pre = ke_op[op[n_op-1].op]>>1;
+					if (((oi&1) && oi>>1 <= pre) || (!(oi&1) && oi>>1 < pre)) break;
+					u = push_back(&out, &n_out, &m_out);
+					*u = op[--n_op];
+				}
+				t = push_back(&op, &n_op, &m_op);
+				*t = v;
+				last_is_val = 0;
+			}
+		}
+	}
+
+	if (*err == 0) {
+		while (n_op > 0 && op[n_op-1].op >= 0) {
+			u = push_back(&out, &n_out, &m_out);
+			*u = op[--n_op];
+		}
+		if (n_op > 0) *err |= KEE_UNLP;
+	}
+	
+	if (*err == 0) { // then check if the number of args is correct
+		int i, n;
+		for (i = n = 0; i < n_out; ++i) {
+			ke1_t *e = &out[i];
+			if (e->ttype == KET_VAL) ++n;
+			else n -= e->n_args - 1;
+		}
+		if (n != 1) *err |= KEE_ARG;
+	}
+
+	free(op); free(s);
+	if (*err) {
+		free(out);
+		return 0;
+	}
+	*_n = n_out;
+	return out;
+}
+
+kexpr_t *ke_parse(const char *_s, int *err)
+{
+	int n;
+	ke1_t *e;
+	kexpr_t *ke;
+	e = ke_parse_core(_s, &n, err);
+	if (*err) return 0;
+	ke = (kexpr_t*)calloc(1, sizeof(kexpr_t));
+	ke->n = n, ke->e = e;
+	return ke;
+}
+
+int ke_eval(const kexpr_t *ke, int64_t *_i, double *_r, const char **_p, int *ret_type)
+{
+	ke1_t *stack, *p, *q;
+	int i, top = 0, err = 0;
+	*_i = 0, *_r = 0., *ret_type = 0;
+	for (i = 0; i < ke->n; ++i) {
+		ke1_t *e = &ke->e[i];
+		if ((e->ttype == KET_OP || e->ttype == KET_FUNC) && e->f.builtin == 0) err |= KEE_UNFUNC;
+		else if (e->ttype == KET_VAL && e->name && e->assigned == 0) err |= KEE_UNVAR;
+	}
+	stack = (ke1_t*)malloc(ke->n * sizeof(ke1_t));
+	for (i = 0; i < ke->n; ++i) {
+		ke1_t *e = &ke->e[i];
+		if (e->ttype == KET_OP || e->ttype == KET_FUNC) {
+			if (e->n_args == 2 && e->f.builtin) {
+				q = &stack[--top], p = &stack[top-1];
+				if (e->user_func) {
+					if (e->user_func == KEF_REAL)
+						p->r = e->f.real_func2(p->r, q->r), p->i = (int64_t)(p->r + .5), p->vtype = KEV_REAL;
+				} else e->f.builtin(p, q);
+			} else if (e->n_args == 1 && e->f.builtin) {
+				p = &stack[top-1];
+				if (e->user_func) {
+					if (e->user_func == KEF_REAL)
+						p->r = e->f.real_func1(p->r), p->i = (int64_t)(p->r + .5), p->vtype = KEV_REAL;
+				} else e->f.builtin(&stack[top-1], 0);
+			} else top -= e->n_args - 1;
+		} else stack[top++] = *e;
+	}
+	*ret_type = stack->vtype;
+	*_i = stack->i, *_r = stack->r, *_p = stack->s;
+	free(stack);
+	return err;
+}	
+
+int64_t ke_eval_int(const kexpr_t *ke, int *err)
+{
+	int int_ret;
+	int64_t i;
+	double r;
+	const char *s;
+	*err = ke_eval(ke, &i, &r, &s, &int_ret);
+	return i;
+}
+
+double ke_eval_real(const kexpr_t *ke, int *err)
+{
+	int int_ret;
+	int64_t i;
+	double r;
+	const char *s;
+	*err = ke_eval(ke, &i, &r, &s, &int_ret);
+	return r;
+}
+
+void ke_destroy(kexpr_t *ke)
+{
+	int i;
+	if (ke == 0) return;
+	for (i = 0; i < ke->n; ++i) {
+		free(ke->e[i].name);
+		free(ke->e[i].s);
+	}
+	free(ke->e); free(ke);
+}
+
+int ke_set_int(kexpr_t *ke, const char *var, int64_t y)
+{
+	int i, n = 0;
+	double yy = (double)y;
+	for (i = 0; i < ke->n; ++i) {
+		ke1_t *e = &ke->e[i];
+		if (e->ttype == KET_VAL && e->name && strcmp(e->name, var) == 0)
+			e->i = y, e->r = yy, e->vtype = KEV_INT, e->assigned = 1, ++n;
+	}
+	return n;
+}
+
+int ke_set_real(kexpr_t *ke, const char *var, double x)
+{
+	int i, n = 0;
+	int64_t xx = (int64_t)(x + .5);
+	for (i = 0; i < ke->n; ++i) {
+		ke1_t *e = &ke->e[i];
+		if (e->ttype == KET_VAL && e->name && strcmp(e->name, var) == 0)
+			e->r = x, e->i = xx, e->vtype = KEV_REAL, e->assigned = 1, ++n;
+	}
+	return n;
+}
+
+int ke_set_str(kexpr_t *ke, const char *var, const char *x)
+{
+	int i, n = 0;
+	for (i = 0; i < ke->n; ++i) {
+		ke1_t *e = &ke->e[i];
+		if (e->ttype == KET_VAL && e->name && strcmp(e->name, var) == 0) {
+			if (e->vtype == KEV_STR) free(e->s);
+			e->s = strdup(x);
+			e->i = 0, e->r = 0., e->assigned = 1;
+			e->vtype = KEV_STR;
+			++n;
+		}
+	}
+	return n;
+}
+
+int ke_set_real_func1(kexpr_t *ke, const char *name, double (*func)(double))
+{
+	int i, n = 0;
+	for (i = 0; i < ke->n; ++i) {
+		ke1_t *e = &ke->e[i];
+		if (e->ttype == KET_FUNC && e->n_args == 1 && strcmp(e->name, name) == 0)
+			e->f.real_func1 = func, e->user_func = KEF_REAL, ++n;
+	}
+	return n;
+}
+
+int ke_set_real_func2(kexpr_t *ke, const char *name, double (*func)(double, double))
+{
+	int i, n = 0;
+	for (i = 0; i < ke->n; ++i) {
+		ke1_t *e = &ke->e[i];
+		if (e->ttype == KET_FUNC && e->n_args == 2 && strcmp(e->name, name) == 0)
+			e->f.real_func2 = func, e->user_func = KEF_REAL, ++n;
+	}
+	return n;
+}
+
+int ke_set_default_func(kexpr_t *ke)
+{
+	int n = 0;
+	n += ke_set_real_func1(ke, "exp", exp);
+	n += ke_set_real_func1(ke, "log", log);
+	n += ke_set_real_func1(ke, "log10", log10);
+	n += ke_set_real_func1(ke, "sqrt", sqrt);
+	n += ke_set_real_func1(ke, "sin", sin);
+	n += ke_set_real_func1(ke, "cos", cos);
+	n += ke_set_real_func1(ke, "tan", tan);
+	n += ke_set_real_func2(ke, "pow", pow);
+	return n;
+}
+
+void ke_unset(kexpr_t *ke)
+{
+	int i;
+	for (i = 0; i < ke->n; ++i) {
+		ke1_t *e = &ke->e[i];
+		if (e->ttype == KET_VAL && e->name) e->assigned = 0;
+	}
+}
+
+void ke_print(const kexpr_t *ke)
+{
+	int i;
+	if (ke == 0) return;
+	for (i = 0; i < ke->n; ++i) {
+		const ke1_t *u = &ke->e[i];
+		if (i) putchar(' ');
+		if (u->ttype == KET_VAL) {
+			if (u->name) printf("%s", u->name);
+			else if (u->vtype == KEV_REAL) printf("%g", u->r);
+			else if (u->vtype == KEV_INT) printf("%lld", (long long)u->i);
+			else if (u->vtype == KEV_STR) printf("\"%s\"", u->s);
+		} else if (u->ttype == KET_OP) {
+			printf("%s", ke_opstr[u->op]);
+		} else if (u->ttype == KET_FUNC) {
+			printf("%s(%d)", u->name, u->n_args);
+		}
+	}
+	putchar('\n');
+}
+
+
+#ifdef KE_MAIN
+#include <unistd.h>
+
+int main(int argc, char *argv[])
+{
+	int c, err, to_print = 0, is_int = 0;
+	kexpr_t *ke;
+
+	while ((c = getopt(argc, argv, "pi")) >= 0) {
+		if (c == 'p') to_print = 1;
+		else if (c == 'i') is_int = 1;
+	}
+	if (optind == argc) {
+		fprintf(stderr, "Usage: %s [-pi] <expr>\n", argv[0]);
+		return 1;
+	}
+	ke = ke_parse(argv[optind], &err);
+	ke_set_default_func(ke);
+	if (err) {
+		fprintf(stderr, "Parse error: 0x%x\n", err);
+		return 1;
+	}
+	if (!to_print) {
+		int64_t vi;
+		double vr;
+		const char *vs;
+		int i, ret_type;
+		if (argc - optind > 1) {
+			for (i = optind + 1; i < argc; ++i) {
+				char *p, *s = argv[i];
+				for (p = s; *p && *p != '='; ++p);
+				if (*p == 0) continue; // not an assignment
+				*p = 0;
+				ke_set_real(ke, s, strtod(p+1, &p));
+			}
+		}
+		err |= ke_eval(ke, &vi, &vr, &vs, &ret_type);
+		if (err & KEE_UNFUNC)
+			fprintf(stderr, "Evaluation warning: an undefined function returns the first function argument.\n");
+		if (err & KEE_UNVAR) fprintf(stderr, "Evaluation warning: unassigned variables are set to 0.\n");
+		if (ret_type == KEV_INT) printf("%lld\n", (long long)vi);
+		else if (ret_type == KEV_REAL) printf("%g\n", vr);
+		else printf("%s\n", vs);
+	} else ke_print(ke);
+	ke_destroy(ke);
+	return 0;
+}
+#endif
diff --git a/ext/klib/kexpr.h b/ext/klib/kexpr.h
new file mode 100644
index 0000000..8023d4c
--- /dev/null
+++ b/ext/klib/kexpr.h
@@ -0,0 +1,68 @@
+#ifndef KEXPR_H
+#define KEXPR_H
+
+#include <stdint.h>
+
+struct kexpr_s;
+typedef struct kexpr_s kexpr_t;
+
+// Parse errors
+#define KEE_UNQU    0x01 // unmatched quotation marks
+#define KEE_UNLP    0x02 // unmatched left parentheses
+#define KEE_UNRP    0x04 // unmatched right parentheses
+#define KEE_UNOP    0x08 // unknown operators
+#define KEE_FUNC    0x10 // wrong function syntax
+#define KEE_ARG     0x20
+#define KEE_NUM     0x40 // fail to parse a number
+
+// Evaluation errors
+#define KEE_UNFUNC  0x40 // undefined function
+#define KEE_UNVAR   0x80 // unassigned variable
+
+// Return type
+#define KEV_REAL  1
+#define KEV_INT   2
+#define KEV_STR   3
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+	// parse an expression and return errors in $err
+	kexpr_t *ke_parse(const char *_s, int *err);
+
+	// free memory allocated during parsing
+	void ke_destroy(kexpr_t *ke);
+
+	// set a variable to integer value and return the occurrence of the variable
+	int ke_set_int(kexpr_t *ke, const char *var, int64_t x);
+
+	// set a variable to real value and return the occurrence of the variable
+	int ke_set_real(kexpr_t *ke, const char *var, double x);
+
+	// set a variable to string value and return the occurrence of the variable
+	int ke_set_str(kexpr_t *ke, const char *var, const char *x);
+
+	// set a user-defined function
+	int ke_set_real_func1(kexpr_t *ke, const char *name, double (*func)(double));
+	int ke_set_real_func2(kexpr_t *ke, const char *name, double (*func)(double, double));
+
+	// set default math functions
+	int ke_set_default_func(kexpr_t *ke);
+
+	// mark all variable as unset
+	void ke_unset(kexpr_t *e);
+
+	// evaluate expression; return error code; final value is returned via pointers
+	int ke_eval(const kexpr_t *ke, int64_t *_i, double *_r, const char **_s, int *ret_type);
+	int64_t ke_eval_int(const kexpr_t *ke, int *err);
+	double ke_eval_real(const kexpr_t *ke, int *err);
+
+	// print the expression in Reverse Polish notation (RPN)
+	void ke_print(const kexpr_t *ke);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif
diff --git a/ext/klib/kgraph.h b/ext/klib/kgraph.h
new file mode 100644
index 0000000..af008ef
--- /dev/null
+++ b/ext/klib/kgraph.h
@@ -0,0 +1,79 @@
+#ifndef AC_KGRAPH_H
+#define AC_KGRAPH_H
+
+#include <stdint.h>
+#include <stdlib.h>
+#include "khash.h"
+#include "kbtree.h"
+
+typedef unsigned kgint_t;
+
+#define kgraph_t(name) kh_##name##_t
+
+#define __KG_BASIC(name, SCOPE, vertex_t, arc_t, ehn) \
+	SCOPE kgraph_t(name) *kg_init_##name(void) { return kh_init(name); } \
+	SCOPE void kg_destroy_##name(kgraph_t(name) *g) { \
+		khint_t k; \
+		if (g == 0) return; \
+		for (k = kh_begin(g); k != kh_end(g); ++k) \
+			if (kh_exist(g, k)) kh_destroy(ehn, kh_val(g, k)._arc); \
+		kh_destroy(name, g); \
+	} \
+	SCOPE vertex_t *kg_get_v_##name(kgraph_t(name) *g, kgint_t v) { \
+		khint_t k = kh_get(name, g, v); \
+		return k == kh_end(g)? 0 : &kh_val(g, k); \
+	} \
+	SCOPE vertex_t *kg_put_v_##name(kgraph_t(name) *g, kgint_t v, int *absent) { \
+		khint_t k; \
+		k = kh_put(name, g, v, absent); \
+		if (*absent) kh_val(g, k)._arc = kh_init(ehn); \
+		return &kh_val(g, k); \
+	} \
+	SCOPE void kg_put_a_##name(kgraph_t(name) *g, kgint_t vbeg, kgint_t vend, int dir, arc_t **pb, arc_t **pe) { \
+		vertex_t *p; \
+		khint_t k; \
+		int absent; \
+		p = kg_put_v_##name(g, vbeg, &absent); \
+		k = kh_put(ehn, p->_arc, vend<<2|dir, &absent); \
+		*pb = &kh_val(p->_arc, k); \
+		p = kg_put_v_##name(g, vend, &absent); \
+		k = kh_put(ehn, p->_arc, vbeg<<2|(~dir&3), &absent); \
+		*pe = &kh_val(p->_arc, k); \
+	} \
+	SCOPE vertex_t *kg_del_v_##name(kgraph_t(name) *g, kgint_t v) { \
+		khint_t k, k0, k2, k3; \
+		khash_t(ehn) *h; \
+		k0 = k = kh_get(name, g, v); \
+		if (k == kh_end(g)) return 0; /* not present in the graph */ \
+		h = kh_val(g, k)._arc; \
+		for (k = kh_begin(h); k != kh_end(h); ++k) /* remove v from its neighbors */ \
+			if (kh_exist(h, k)) { \
+				k2 = kh_get(name, g, kh_key(h, k)>>2); \
+				/* assert(k2 != kh_end(g)); */ \
+				k3 = kh_get(ehn, kh_val(g, k2)._arc, v<<2|(~kh_key(h, k)&3)); \
+				/* assert(k3 != kh_end(kh_val(g, k2)._arc)); */ \
+				kh_del(ehn, kh_val(g, k2)._arc, k3); \
+			} \
+		kh_destroy(ehn, h); \
+		kh_del(name, g, k0); \
+		return &kh_val(g, k0); \
+	}
+
+#define KGRAPH_PRINT(name, SCOPE) \
+	SCOPE void kg_print_##name(kgraph_t(name) *g) { \
+		khint_t k, k2; \
+		for (k = kh_begin(g); k != kh_end(g); ++k) \
+			if (kh_exist(g, k)) { \
+				printf("v %u\n", kh_key(g, k)); \
+				for (k2 = kh_begin(kh_val(g, k)._arc); k2 != kh_end(kh_val(g, k)._arc); ++k2) \
+					if (kh_exist(kh_val(g, k)._arc, k2) && kh_key(g, k) < kh_key(kh_val(g, k)._arc, k2)>>2) \
+						printf("a %u%c%c%u\n", kh_key(g, k), "><"[kh_key(kh_val(g, k)._arc, k2)>>1&1], \
+								"><"[kh_key(kh_val(g, k)._arc, k2)&1], kh_key(kh_val(g, k)._arc, k2)>>2); \
+			} \
+	}
+
+#define KGRAPH_INIT(name, SCOPE, vertex_t, arc_t, ehn) \
+	KHASH_INIT2(name, SCOPE, kgint_t, vertex_t, 1, kh_int_hash_func, kh_int_hash_equal) \
+	__KG_BASIC(name, SCOPE, vertex_t, arc_t, ehn)
+
+#endif
diff --git a/ext/klib/khash.h b/ext/klib/khash.h
new file mode 100644
index 0000000..f75f347
--- /dev/null
+++ b/ext/klib/khash.h
@@ -0,0 +1,627 @@
+/* The MIT License
+
+   Copyright (c) 2008, 2009, 2011 by Attractive Chaos <attractor@live.co.uk>
+
+   Permission is hereby granted, free of charge, to any person obtaining
+   a copy of this software and associated documentation files (the
+   "Software"), to deal in the Software without restriction, including
+   without limitation the rights to use, copy, modify, merge, publish,
+   distribute, sublicense, and/or sell copies of the Software, and to
+   permit persons to whom the Software is furnished to do so, subject to
+   the following conditions:
+
+   The above copyright notice and this permission notice shall be
+   included in all copies or substantial portions of the Software.
+
+   THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+   EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+   MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+   NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+   BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+   ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+   CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+   SOFTWARE.
+*/
+
+/*
+  An example:
+
+#include "khash.h"
+KHASH_MAP_INIT_INT(32, char)
+int main() {
+	int ret, is_missing;
+	khiter_t k;
+	khash_t(32) *h = kh_init(32);
+	k = kh_put(32, h, 5, &ret);
+	kh_value(h, k) = 10;
+	k = kh_get(32, h, 10);
+	is_missing = (k == kh_end(h));
+	k = kh_get(32, h, 5);
+	kh_del(32, h, k);
+	for (k = kh_begin(h); k != kh_end(h); ++k)
+		if (kh_exist(h, k)) kh_value(h, k) = 1;
+	kh_destroy(32, h);
+	return 0;
+}
+*/
+
+/*
+  2013-05-02 (0.2.8):
+
+	* Use quadratic probing. When the capacity is power of 2, stepping function
+	  i*(i+1)/2 guarantees to traverse each bucket. It is better than double
+	  hashing on cache performance and is more robust than linear probing.
+
+	  In theory, double hashing should be more robust than quadratic probing.
+	  However, my implementation is probably not for large hash tables, because
+	  the second hash function is closely tied to the first hash function,
+	  which reduce the effectiveness of double hashing.
+
+	Reference: http://research.cs.vt.edu/AVresearch/hashing/quadratic.php
+
+  2011-12-29 (0.2.7):
+
+    * Minor code clean up; no actual effect.
+
+  2011-09-16 (0.2.6):
+
+	* The capacity is a power of 2. This seems to dramatically improve the
+	  speed for simple keys. Thank Zilong Tan for the suggestion. Reference:
+
+	   - http://code.google.com/p/ulib/
+	   - http://nothings.org/computer/judy/
+
+	* Allow to optionally use linear probing which usually has better
+	  performance for random input. Double hashing is still the default as it
+	  is more robust to certain non-random input.
+
+	* Added Wang's integer hash function (not used by default). This hash
+	  function is more robust to certain non-random input.
+
+  2011-02-14 (0.2.5):
+
+    * Allow to declare global functions.
+
+  2009-09-26 (0.2.4):
+
+    * Improve portability
+
+  2008-09-19 (0.2.3):
+
+	* Corrected the example
+	* Improved interfaces
+
+  2008-09-11 (0.2.2):
+
+	* Improved speed a little in kh_put()
+
+  2008-09-10 (0.2.1):
+
+	* Added kh_clear()
+	* Fixed a compiling error
+
+  2008-09-02 (0.2.0):
+
+	* Changed to token concatenation which increases flexibility.
+
+  2008-08-31 (0.1.2):
+
+	* Fixed a bug in kh_get(), which has not been tested previously.
+
+  2008-08-31 (0.1.1):
+
+	* Added destructor
+*/
+
+
+#ifndef __AC_KHASH_H
+#define __AC_KHASH_H
+
+/*!
+  @header
+
+  Generic hash table library.
+ */
+
+#define AC_VERSION_KHASH_H "0.2.8"
+
+#include <stdlib.h>
+#include <string.h>
+#include <limits.h>
+
+/* compiler specific configuration */
+
+#if UINT_MAX == 0xffffffffu
+typedef unsigned int khint32_t;
+#elif ULONG_MAX == 0xffffffffu
+typedef unsigned long khint32_t;
+#endif
+
+#if ULONG_MAX == ULLONG_MAX
+typedef unsigned long khint64_t;
+#else
+typedef unsigned long long khint64_t;
+#endif
+
+#ifndef kh_inline
+#ifdef _MSC_VER
+#define kh_inline __inline
+#else
+#define kh_inline inline
+#endif
+#endif /* kh_inline */
+
+#ifndef klib_unused
+#if (defined __clang__ && __clang_major__ >= 3) || (defined __GNUC__ && __GNUC__ >= 3)
+#define klib_unused __attribute__ ((__unused__))
+#else
+#define klib_unused
+#endif
+#endif /* klib_unused */
+
+typedef khint32_t khint_t;
+typedef khint_t khiter_t;
+
+#define __ac_isempty(flag, i) ((flag[i>>4]>>((i&0xfU)<<1))&2)
+#define __ac_isdel(flag, i) ((flag[i>>4]>>((i&0xfU)<<1))&1)
+#define __ac_iseither(flag, i) ((flag[i>>4]>>((i&0xfU)<<1))&3)
+#define __ac_set_isdel_false(flag, i) (flag[i>>4]&=~(1ul<<((i&0xfU)<<1)))
+#define __ac_set_isempty_false(flag, i) (flag[i>>4]&=~(2ul<<((i&0xfU)<<1)))
+#define __ac_set_isboth_false(flag, i) (flag[i>>4]&=~(3ul<<((i&0xfU)<<1)))
+#define __ac_set_isdel_true(flag, i) (flag[i>>4]|=1ul<<((i&0xfU)<<1))
+
+#define __ac_fsize(m) ((m) < 16? 1 : (m)>>4)
+
+#ifndef kroundup32
+#define kroundup32(x) (--(x), (x)|=(x)>>1, (x)|=(x)>>2, (x)|=(x)>>4, (x)|=(x)>>8, (x)|=(x)>>16, ++(x))
+#endif
+
+#ifndef kcalloc
+#define kcalloc(N,Z) calloc(N,Z)
+#endif
+#ifndef kmalloc
+#define kmalloc(Z) malloc(Z)
+#endif
+#ifndef krealloc
+#define krealloc(P,Z) realloc(P,Z)
+#endif
+#ifndef kfree
+#define kfree(P) free(P)
+#endif
+
+static const double __ac_HASH_UPPER = 0.77;
+
+#define __KHASH_TYPE(name, khkey_t, khval_t) \
+	typedef struct kh_##name##_s { \
+		khint_t n_buckets, size, n_occupied, upper_bound; \
+		khint32_t *flags; \
+		khkey_t *keys; \
+		khval_t *vals; \
+	} kh_##name##_t;
+
+#define __KHASH_PROTOTYPES(name, khkey_t, khval_t)	 					\
+	extern kh_##name##_t *kh_init_##name(void);							\
+	extern void kh_destroy_##name(kh_##name##_t *h);					\
+	extern void kh_clear_##name(kh_##name##_t *h);						\
+	extern khint_t kh_get_##name(const kh_##name##_t *h, khkey_t key); 	\
+	extern int kh_resize_##name(kh_##name##_t *h, khint_t new_n_buckets); \
+	extern khint_t kh_put_##name(kh_##name##_t *h, khkey_t key, int *ret); \
+	extern void kh_del_##name(kh_##name##_t *h, khint_t x);
+
+#define __KHASH_IMPL(name, SCOPE, khkey_t, khval_t, kh_is_map, __hash_func, __hash_equal) \
+	SCOPE kh_##name##_t *kh_init_##name(void) {							\
+		return (kh_##name##_t*)kcalloc(1, sizeof(kh_##name##_t));		\
+	}																	\
+	SCOPE void kh_destroy_##name(kh_##name##_t *h)						\
+	{																	\
+		if (h) {														\
+			kfree((void *)h->keys); kfree(h->flags);					\
+			kfree((void *)h->vals);										\
+			kfree(h);													\
+		}																\
+	}																	\
+	SCOPE void kh_clear_##name(kh_##name##_t *h)						\
+	{																	\
+		if (h && h->flags) {											\
+			memset(h->flags, 0xaa, __ac_fsize(h->n_buckets) * sizeof(khint32_t)); \
+			h->size = h->n_occupied = 0;								\
+		}																\
+	}																	\
+	SCOPE khint_t kh_get_##name(const kh_##name##_t *h, khkey_t key) 	\
+	{																	\
+		if (h->n_buckets) {												\
+			khint_t k, i, last, mask, step = 0; \
+			mask = h->n_buckets - 1;									\
+			k = __hash_func(key); i = k & mask;							\
+			last = i; \
+			while (!__ac_isempty(h->flags, i) && (__ac_isdel(h->flags, i) || !__hash_equal(h->keys[i], key))) { \
+				i = (i + (++step)) & mask; \
+				if (i == last) return h->n_buckets;						\
+			}															\
+			return __ac_iseither(h->flags, i)? h->n_buckets : i;		\
+		} else return 0;												\
+	}																	\
+	SCOPE int kh_resize_##name(kh_##name##_t *h, khint_t new_n_buckets) \
+	{ /* This function uses 0.25*n_buckets bytes of working space instead of [sizeof(key_t+val_t)+.25]*n_buckets. */ \
+		khint32_t *new_flags = 0;										\
+		khint_t j = 1;													\
+		{																\
+			kroundup32(new_n_buckets); 									\
+			if (new_n_buckets < 4) new_n_buckets = 4;					\
+			if (h->size >= (khint_t)(new_n_buckets * __ac_HASH_UPPER + 0.5)) j = 0;	/* requested size is too small */ \
+			else { /* hash table size to be changed (shrink or expand); rehash */ \
+				new_flags = (khint32_t*)kmalloc(__ac_fsize(new_n_buckets) * sizeof(khint32_t));	\
+				if (!new_flags) return -1;								\
+				memset(new_flags, 0xaa, __ac_fsize(new_n_buckets) * sizeof(khint32_t)); \
+				if (h->n_buckets < new_n_buckets) {	/* expand */		\
+					khkey_t *new_keys = (khkey_t*)krealloc((void *)h->keys, new_n_buckets * sizeof(khkey_t)); \
+					if (!new_keys) { kfree(new_flags); return -1; }		\
+					h->keys = new_keys;									\
+					if (kh_is_map) {									\
+						khval_t *new_vals = (khval_t*)krealloc((void *)h->vals, new_n_buckets * sizeof(khval_t)); \
+						if (!new_vals) { kfree(new_flags); return -1; }	\
+						h->vals = new_vals;								\
+					}													\
+				} /* otherwise shrink */								\
+			}															\
+		}																\
+		if (j) { /* rehashing is needed */								\
+			for (j = 0; j != h->n_buckets; ++j) {						\
+				if (__ac_iseither(h->flags, j) == 0) {					\
+					khkey_t key = h->keys[j];							\
+					khval_t val;										\
+					khint_t new_mask;									\
+					new_mask = new_n_buckets - 1; 						\
+					if (kh_is_map) val = h->vals[j];					\
+					__ac_set_isdel_true(h->flags, j);					\
+					while (1) { /* kick-out process; sort of like in Cuckoo hashing */ \
+						khint_t k, i, step = 0; \
+						k = __hash_func(key);							\
+						i = k & new_mask;								\
+						while (!__ac_isempty(new_flags, i)) i = (i + (++step)) & new_mask; \
+						__ac_set_isempty_false(new_flags, i);			\
+						if (i < h->n_buckets && __ac_iseither(h->flags, i) == 0) { /* kick out the existing element */ \
+							{ khkey_t tmp = h->keys[i]; h->keys[i] = key; key = tmp; } \
+							if (kh_is_map) { khval_t tmp = h->vals[i]; h->vals[i] = val; val = tmp; } \
+							__ac_set_isdel_true(h->flags, i); /* mark it as deleted in the old hash table */ \
+						} else { /* write the element and jump out of the loop */ \
+							h->keys[i] = key;							\
+							if (kh_is_map) h->vals[i] = val;			\
+							break;										\
+						}												\
+					}													\
+				}														\
+			}															\
+			if (h->n_buckets > new_n_buckets) { /* shrink the hash table */ \
+				h->keys = (khkey_t*)krealloc((void *)h->keys, new_n_buckets * sizeof(khkey_t)); \
+				if (kh_is_map) h->vals = (khval_t*)krealloc((void *)h->vals, new_n_buckets * sizeof(khval_t)); \
+			}															\
+			kfree(h->flags); /* free the working space */				\
+			h->flags = new_flags;										\
+			h->n_buckets = new_n_buckets;								\
+			h->n_occupied = h->size;									\
+			h->upper_bound = (khint_t)(h->n_buckets * __ac_HASH_UPPER + 0.5); \
+		}																\
+		return 0;														\
+	}																	\
+	SCOPE khint_t kh_put_##name(kh_##name##_t *h, khkey_t key, int *ret) \
+	{																	\
+		khint_t x;														\
+		if (h->n_occupied >= h->upper_bound) { /* update the hash table */ \
+			if (h->n_buckets > (h->size<<1)) {							\
+				if (kh_resize_##name(h, h->n_buckets - 1) < 0) { /* clear "deleted" elements */ \
+					*ret = -1; return h->n_buckets;						\
+				}														\
+			} else if (kh_resize_##name(h, h->n_buckets + 1) < 0) { /* expand the hash table */ \
+				*ret = -1; return h->n_buckets;							\
+			}															\
+		} /* TODO: to implement automatically shrinking; resize() already support shrinking */ \
+		{																\
+			khint_t k, i, site, last, mask = h->n_buckets - 1, step = 0; \
+			x = site = h->n_buckets; k = __hash_func(key); i = k & mask; \
+			if (__ac_isempty(h->flags, i)) x = i; /* for speed up */	\
+			else {														\
+				last = i; \
+				while (!__ac_isempty(h->flags, i) && (__ac_isdel(h->flags, i) || !__hash_equal(h->keys[i], key))) { \
+					if (__ac_isdel(h->flags, i)) site = i;				\
+					i = (i + (++step)) & mask; \
+					if (i == last) { x = site; break; }					\
+				}														\
+				if (x == h->n_buckets) {								\
+					if (__ac_isempty(h->flags, i) && site != h->n_buckets) x = site; \
+					else x = i;											\
+				}														\
+			}															\
+		}																\
+		if (__ac_isempty(h->flags, x)) { /* not present at all */		\
+			h->keys[x] = key;											\
+			__ac_set_isboth_false(h->flags, x);							\
+			++h->size; ++h->n_occupied;									\
+			*ret = 1;													\
+		} else if (__ac_isdel(h->flags, x)) { /* deleted */				\
+			h->keys[x] = key;											\
+			__ac_set_isboth_false(h->flags, x);							\
+			++h->size;													\
+			*ret = 2;													\
+		} else *ret = 0; /* Don't touch h->keys[x] if present and not deleted */ \
+		return x;														\
+	}																	\
+	SCOPE void kh_del_##name(kh_##name##_t *h, khint_t x)				\
+	{																	\
+		if (x != h->n_buckets && !__ac_iseither(h->flags, x)) {			\
+			__ac_set_isdel_true(h->flags, x);							\
+			--h->size;													\
+		}																\
+	}
+
+#define KHASH_DECLARE(name, khkey_t, khval_t)		 					\
+	__KHASH_TYPE(name, khkey_t, khval_t) 								\
+	__KHASH_PROTOTYPES(name, khkey_t, khval_t)
+
+#define KHASH_INIT2(name, SCOPE, khkey_t, khval_t, kh_is_map, __hash_func, __hash_equal) \
+	__KHASH_TYPE(name, khkey_t, khval_t) 								\
+	__KHASH_IMPL(name, SCOPE, khkey_t, khval_t, kh_is_map, __hash_func, __hash_equal)
+
+#define KHASH_INIT(name, khkey_t, khval_t, kh_is_map, __hash_func, __hash_equal) \
+	KHASH_INIT2(name, static kh_inline klib_unused, khkey_t, khval_t, kh_is_map, __hash_func, __hash_equal)
+
+/* --- BEGIN OF HASH FUNCTIONS --- */
+
+/*! @function
+  @abstract     Integer hash function
+  @param  key   The integer [khint32_t]
+  @return       The hash value [khint_t]
+ */
+#define kh_int_hash_func(key) (khint32_t)(key)
+/*! @function
+  @abstract     Integer comparison function
+ */
+#define kh_int_hash_equal(a, b) ((a) == (b))
+/*! @function
+  @abstract     64-bit integer hash function
+  @param  key   The integer [khint64_t]
+  @return       The hash value [khint_t]
+ */
+#define kh_int64_hash_func(key) (khint32_t)((key)>>33^(key)^(key)<<11)
+/*! @function
+  @abstract     64-bit integer comparison function
+ */
+#define kh_int64_hash_equal(a, b) ((a) == (b))
+/*! @function
+  @abstract     const char* hash function
+  @param  s     Pointer to a null terminated string
+  @return       The hash value
+ */
+static kh_inline khint_t __ac_X31_hash_string(const char *s)
+{
+	khint_t h = (khint_t)*s;
+	if (h) for (++s ; *s; ++s) h = (h << 5) - h + (khint_t)*s;
+	return h;
+}
+/*! @function
+  @abstract     Another interface to const char* hash function
+  @param  key   Pointer to a null terminated string [const char*]
+  @return       The hash value [khint_t]
+ */
+#define kh_str_hash_func(key) __ac_X31_hash_string(key)
+/*! @function
+  @abstract     Const char* comparison function
+ */
+#define kh_str_hash_equal(a, b) (strcmp(a, b) == 0)
+
+static kh_inline khint_t __ac_Wang_hash(khint_t key)
+{
+    key += ~(key << 15);
+    key ^=  (key >> 10);
+    key +=  (key << 3);
+    key ^=  (key >> 6);
+    key += ~(key << 11);
+    key ^=  (key >> 16);
+    return key;
+}
+#define kh_int_hash_func2(key) __ac_Wang_hash((khint_t)key)
+
+/* --- END OF HASH FUNCTIONS --- */
+
+/* Other convenient macros... */
+
+/*!
+  @abstract Type of the hash table.
+  @param  name  Name of the hash table [symbol]
+ */
+#define khash_t(name) kh_##name##_t
+
+/*! @function
+  @abstract     Initiate a hash table.
+  @param  name  Name of the hash table [symbol]
+  @return       Pointer to the hash table [khash_t(name)*]
+ */
+#define kh_init(name) kh_init_##name()
+
+/*! @function
+  @abstract     Destroy a hash table.
+  @param  name  Name of the hash table [symbol]
+  @param  h     Pointer to the hash table [khash_t(name)*]
+ */
+#define kh_destroy(name, h) kh_destroy_##name(h)
+
+/*! @function
+  @abstract     Reset a hash table without deallocating memory.
+  @param  name  Name of the hash table [symbol]
+  @param  h     Pointer to the hash table [khash_t(name)*]
+ */
+#define kh_clear(name, h) kh_clear_##name(h)
+
+/*! @function
+  @abstract     Resize a hash table.
+  @param  name  Name of the hash table [symbol]
+  @param  h     Pointer to the hash table [khash_t(name)*]
+  @param  s     New size [khint_t]
+ */
+#define kh_resize(name, h, s) kh_resize_##name(h, s)
+
+/*! @function
+  @abstract     Insert a key to the hash table.
+  @param  name  Name of the hash table [symbol]
+  @param  h     Pointer to the hash table [khash_t(name)*]
+  @param  k     Key [type of keys]
+  @param  r     Extra return code: -1 if the operation failed;
+                0 if the key is present in the hash table;
+                1 if the bucket is empty (never used); 2 if the element in
+				the bucket has been deleted [int*]
+  @return       Iterator to the inserted element [khint_t]
+ */
+#define kh_put(name, h, k, r) kh_put_##name(h, k, r)
+
+/*! @function
+  @abstract     Retrieve a key from the hash table.
+  @param  name  Name of the hash table [symbol]
+  @param  h     Pointer to the hash table [khash_t(name)*]
+  @param  k     Key [type of keys]
+  @return       Iterator to the found element, or kh_end(h) if the element is absent [khint_t]
+ */
+#define kh_get(name, h, k) kh_get_##name(h, k)
+
+/*! @function
+  @abstract     Remove a key from the hash table.
+  @param  name  Name of the hash table [symbol]
+  @param  h     Pointer to the hash table [khash_t(name)*]
+  @param  k     Iterator to the element to be deleted [khint_t]
+ */
+#define kh_del(name, h, k) kh_del_##name(h, k)
+
+/*! @function
+  @abstract     Test whether a bucket contains data.
+  @param  h     Pointer to the hash table [khash_t(name)*]
+  @param  x     Iterator to the bucket [khint_t]
+  @return       1 if containing data; 0 otherwise [int]
+ */
+#define kh_exist(h, x) (!__ac_iseither((h)->flags, (x)))
+
+/*! @function
+  @abstract     Get key given an iterator
+  @param  h     Pointer to the hash table [khash_t(name)*]
+  @param  x     Iterator to the bucket [khint_t]
+  @return       Key [type of keys]
+ */
+#define kh_key(h, x) ((h)->keys[x])
+
+/*! @function
+  @abstract     Get value given an iterator
+  @param  h     Pointer to the hash table [khash_t(name)*]
+  @param  x     Iterator to the bucket [khint_t]
+  @return       Value [type of values]
+  @discussion   For hash sets, calling this results in segfault.
+ */
+#define kh_val(h, x) ((h)->vals[x])
+
+/*! @function
+  @abstract     Alias of kh_val()
+ */
+#define kh_value(h, x) ((h)->vals[x])
+
+/*! @function
+  @abstract     Get the start iterator
+  @param  h     Pointer to the hash table [khash_t(name)*]
+  @return       The start iterator [khint_t]
+ */
+#define kh_begin(h) (khint_t)(0)
+
+/*! @function
+  @abstract     Get the end iterator
+  @param  h     Pointer to the hash table [khash_t(name)*]
+  @return       The end iterator [khint_t]
+ */
+#define kh_end(h) ((h)->n_buckets)
+
+/*! @function
+  @abstract     Get the number of elements in the hash table
+  @param  h     Pointer to the hash table [khash_t(name)*]
+  @return       Number of elements in the hash table [khint_t]
+ */
+#define kh_size(h) ((h)->size)
+
+/*! @function
+  @abstract     Get the number of buckets in the hash table
+  @param  h     Pointer to the hash table [khash_t(name)*]
+  @return       Number of buckets in the hash table [khint_t]
+ */
+#define kh_n_buckets(h) ((h)->n_buckets)
+
+/*! @function
+  @abstract     Iterate over the entries in the hash table
+  @param  h     Pointer to the hash table [khash_t(name)*]
+  @param  kvar  Variable to which key will be assigned
+  @param  vvar  Variable to which value will be assigned
+  @param  code  Block of code to execute
+ */
+#define kh_foreach(h, kvar, vvar, code) { khint_t __i;		\
+	for (__i = kh_begin(h); __i != kh_end(h); ++__i) {		\
+		if (!kh_exist(h,__i)) continue;						\
+		(kvar) = kh_key(h,__i);								\
+		(vvar) = kh_val(h,__i);								\
+		code;												\
+	} }
+
+/*! @function
+  @abstract     Iterate over the values in the hash table
+  @param  h     Pointer to the hash table [khash_t(name)*]
+  @param  vvar  Variable to which value will be assigned
+  @param  code  Block of code to execute
+ */
+#define kh_foreach_value(h, vvar, code) { khint_t __i;		\
+	for (__i = kh_begin(h); __i != kh_end(h); ++__i) {		\
+		if (!kh_exist(h,__i)) continue;						\
+		(vvar) = kh_val(h,__i);								\
+		code;												\
+	} }
+
+/* More convenient interfaces */
+
+/*! @function
+  @abstract     Instantiate a hash set containing integer keys
+  @param  name  Name of the hash table [symbol]
+ */
+#define KHASH_SET_INIT_INT(name)										\
+	KHASH_INIT(name, khint32_t, char, 0, kh_int_hash_func, kh_int_hash_equal)
+
+/*! @function
+  @abstract     Instantiate a hash map containing integer keys
+  @param  name  Name of the hash table [symbol]
+  @param  khval_t  Type of values [type]
+ */
+#define KHASH_MAP_INIT_INT(name, khval_t)								\
+	KHASH_INIT(name, khint32_t, khval_t, 1, kh_int_hash_func, kh_int_hash_equal)
+
+/*! @function
+  @abstract     Instantiate a hash set containing 64-bit integer keys
+  @param  name  Name of the hash table [symbol]
+ */
+#define KHASH_SET_INIT_INT64(name)										\
+	KHASH_INIT(name, khint64_t, char, 0, kh_int64_hash_func, kh_int64_hash_equal)
+
+/*! @function
+  @abstract     Instantiate a hash map containing 64-bit integer keys
+  @param  name  Name of the hash table [symbol]
+  @param  khval_t  Type of values [type]
+ */
+#define KHASH_MAP_INIT_INT64(name, khval_t)								\
+	KHASH_INIT(name, khint64_t, khval_t, 1, kh_int64_hash_func, kh_int64_hash_equal)
+
+typedef const char *kh_cstr_t;
+/*! @function
+  @abstract     Instantiate a hash map containing const char* keys
+  @param  name  Name of the hash table [symbol]
+ */
+#define KHASH_SET_INIT_STR(name)										\
+	KHASH_INIT(name, kh_cstr_t, char, 0, kh_str_hash_func, kh_str_hash_equal)
+
+/*! @function
+  @abstract     Instantiate a hash map containing const char* keys
+  @param  name  Name of the hash table [symbol]
+  @param  khval_t  Type of values [type]
+ */
+#define KHASH_MAP_INIT_STR(name, khval_t)								\
+	KHASH_INIT(name, kh_cstr_t, khval_t, 1, kh_str_hash_func, kh_str_hash_equal)
+
+#endif /* __AC_KHASH_H */
diff --git a/ext/klib/khashl.h b/ext/klib/khashl.h
new file mode 100644
index 0000000..7f2a81a
--- /dev/null
+++ b/ext/klib/khashl.h
@@ -0,0 +1,446 @@
+/* The MIT License
+
+   Copyright (c) 2019-2024 by Attractive Chaos <attractor@live.co.uk>
+
+   Permission is hereby granted, free of charge, to any person obtaining
+   a copy of this software and associated documentation files (the
+   "Software"), to deal in the Software without restriction, including
+   without limitation the rights to use, copy, modify, merge, publish,
+   distribute, sublicense, and/or sell copies of the Software, and to
+   permit persons to whom the Software is furnished to do so, subject to
+   the following conditions:
+
+   The above copyright notice and this permission notice shall be
+   included in all copies or substantial portions of the Software.
+
+   THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+   EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+   MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+   NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+   BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+   ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+   CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+   SOFTWARE.
+*/
+
+#ifndef __AC_KHASHL_H
+#define __AC_KHASHL_H
+
+#define AC_VERSION_KHASHL_H "r20"
+
+#include <stdlib.h>
+#include <string.h>
+#include <limits.h>
+
+/************************************
+ * Compiler specific configurations *
+ ************************************/
+
+#if UINT_MAX == 0xffffffffu
+typedef unsigned int khint32_t;
+#elif ULONG_MAX == 0xffffffffu
+typedef unsigned long khint32_t;
+#endif
+
+#if ULONG_MAX == ULLONG_MAX
+typedef unsigned long khint64_t;
+#else
+typedef unsigned long long khint64_t;
+#endif
+
+#ifndef kh_inline
+#ifdef _MSC_VER
+#define kh_inline __inline
+#else
+#define kh_inline inline
+#endif
+#endif /* kh_inline */
+
+#ifndef klib_unused
+#if (defined __clang__ && __clang_major__ >= 3) || (defined __GNUC__ && __GNUC__ >= 3)
+#define klib_unused __attribute__ ((__unused__))
+#else
+#define klib_unused
+#endif
+#endif /* klib_unused */
+
+#define KH_LOCAL static kh_inline klib_unused
+
+typedef khint32_t khint_t;
+typedef const char *kh_cstr_t;
+
+/***********************
+ * Configurable macros *
+ ***********************/
+
+#ifndef kh_max_count
+#define kh_max_count(cap) (((cap)>>1) + ((cap)>>2)) /* default load factor: 75% */
+#endif
+
+#ifndef kh_packed
+#define kh_packed __attribute__ ((__packed__))
+#endif
+
+#ifndef kcalloc
+#define kcalloc(N,Z) calloc(N,Z)
+#endif
+#ifndef kmalloc
+#define kmalloc(Z) malloc(Z)
+#endif
+#ifndef krealloc
+#define krealloc(P,Z) realloc(P,Z)
+#endif
+#ifndef kfree
+#define kfree(P) free(P)
+#endif
+
+/****************************
+ * Simple private functions *
+ ****************************/
+
+#define __kh_used(flag, i)       (flag[i>>5] >> (i&0x1fU) & 1U)
+#define __kh_set_used(flag, i)   (flag[i>>5] |= 1U<<(i&0x1fU))
+#define __kh_set_unused(flag, i) (flag[i>>5] &= ~(1U<<(i&0x1fU)))
+
+#define __kh_fsize(m) ((m) < 32? 1 : (m)>>5)
+
+static kh_inline khint_t __kh_h2b(khint_t hash, khint_t bits) { return hash * 2654435769U >> (32 - bits); }
+
+/*******************
+ * Hash table base *
+ *******************/
+
+#define __KHASHL_TYPE(HType, khkey_t) \
+	typedef struct HType { \
+		khint_t bits, count; \
+		khint32_t *used; \
+		khkey_t *keys; \
+	} HType;
+
+#define __KHASHL_PROTOTYPES(HType, prefix, khkey_t) \
+	extern HType *prefix##_init(void); \
+	extern void prefix##_destroy(HType *h); \
+	extern void prefix##_clear(HType *h); \
+	extern khint_t prefix##_getp(const HType *h, const khkey_t *key); \
+	extern int prefix##_resize(HType *h, khint_t new_n_buckets); \
+	extern khint_t prefix##_putp(HType *h, const khkey_t *key, int *absent); \
+	extern void prefix##_del(HType *h, khint_t k);
+
+#define __KHASHL_IMPL_BASIC(SCOPE, HType, prefix) \
+	SCOPE HType *prefix##_init(void) { \
+		return (HType*)kcalloc(1, sizeof(HType)); \
+	} \
+	SCOPE void prefix##_destroy(HType *h) { \
+		if (!h) return; \
+		kfree((void *)h->keys); kfree(h->used); \
+		kfree(h); \
+	} \
+	SCOPE void prefix##_clear(HType *h) { \
+		if (h && h->used) { \
+			khint_t n_buckets = (khint_t)1U << h->bits; \
+			memset(h->used, 0, __kh_fsize(n_buckets) * sizeof(khint32_t)); \
+			h->count = 0; \
+		} \
+	}
+
+#define __KHASHL_IMPL_GET(SCOPE, HType, prefix, khkey_t, __hash_fn, __hash_eq) \
+	SCOPE khint_t prefix##_getp_core(const HType *h, const khkey_t *key, khint_t hash) { \
+		khint_t i, last, n_buckets, mask; \
+		if (h->keys == 0) return 0; \
+		n_buckets = (khint_t)1U << h->bits; \
+		mask = n_buckets - 1U; \
+		i = last = __kh_h2b(hash, h->bits); \
+		while (__kh_used(h->used, i) && !__hash_eq(h->keys[i], *key)) { \
+			i = (i + 1U) & mask; \
+			if (i == last) return n_buckets; \
+		} \
+		return !__kh_used(h->used, i)? n_buckets : i; \
+	} \
+	SCOPE khint_t prefix##_getp(const HType *h, const khkey_t *key) { return prefix##_getp_core(h, key, __hash_fn(*key)); } \
+	SCOPE khint_t prefix##_get(const HType *h, khkey_t key) { return prefix##_getp_core(h, &key, __hash_fn(key)); }
+
+#define __KHASHL_IMPL_RESIZE(SCOPE, HType, prefix, khkey_t, __hash_fn, __hash_eq) \
+	SCOPE int prefix##_resize(HType *h, khint_t new_n_buckets) { \
+		khint32_t *new_used = 0; \
+		khint_t j = 0, x = new_n_buckets, n_buckets, new_bits, new_mask; \
+		while ((x >>= 1) != 0) ++j; \
+		if (new_n_buckets & (new_n_buckets - 1)) ++j; \
+		new_bits = j > 2? j : 2; \
+		new_n_buckets = (khint_t)1U << new_bits; \
+		if (h->count > kh_max_count(new_n_buckets)) return 0; /* requested size is too small */ \
+		new_used = (khint32_t*)kmalloc(__kh_fsize(new_n_buckets) * sizeof(khint32_t)); \
+		memset(new_used, 0, __kh_fsize(new_n_buckets) * sizeof(khint32_t)); \
+		if (!new_used) return -1; /* not enough memory */ \
+		n_buckets = h->keys? (khint_t)1U<<h->bits : 0U; \
+		if (n_buckets < new_n_buckets) { /* expand */ \
+			khkey_t *new_keys = (khkey_t*)krealloc((void*)h->keys, new_n_buckets * sizeof(khkey_t)); \
+			if (!new_keys) { kfree(new_used); return -1; } \
+			h->keys = new_keys; \
+		} /* otherwise shrink */ \
+		new_mask = new_n_buckets - 1; \
+		for (j = 0; j != n_buckets; ++j) { \
+			khkey_t key; \
+			if (!__kh_used(h->used, j)) continue; \
+			key = h->keys[j]; \
+			__kh_set_unused(h->used, j); \
+			while (1) { /* kick-out process; sort of like in Cuckoo hashing */ \
+				khint_t i; \
+				i = __kh_h2b(__hash_fn(key), new_bits); \
+				while (__kh_used(new_used, i)) i = (i + 1) & new_mask; \
+				__kh_set_used(new_used, i); \
+				if (i < n_buckets && __kh_used(h->used, i)) { /* kick out the existing element */ \
+					{ khkey_t tmp = h->keys[i]; h->keys[i] = key; key = tmp; } \
+					__kh_set_unused(h->used, i); /* mark it as deleted in the old hash table */ \
+				} else { /* write the element and jump out of the loop */ \
+					h->keys[i] = key; \
+					break; \
+				} \
+			} \
+		} \
+		if (n_buckets > new_n_buckets) /* shrink the hash table */ \
+			h->keys = (khkey_t*)krealloc((void *)h->keys, new_n_buckets * sizeof(khkey_t)); \
+		kfree(h->used); /* free the working space */ \
+		h->used = new_used, h->bits = new_bits; \
+		return 0; \
+	}
+
+#define __KHASHL_IMPL_PUT(SCOPE, HType, prefix, khkey_t, __hash_fn, __hash_eq) \
+	SCOPE khint_t prefix##_putp_core(HType *h, const khkey_t *key, khint_t hash, int *absent) { \
+		khint_t n_buckets, i, last, mask; \
+		n_buckets = h->keys? (khint_t)1U<<h->bits : 0U; \
+		*absent = -1; \
+		if (h->count >= kh_max_count(n_buckets)) { /* rehashing */ \
+			if (prefix##_resize(h, n_buckets + 1U) < 0) \
+				return n_buckets; \
+			n_buckets = (khint_t)1U<<h->bits; \
+		} /* TODO: to implement automatically shrinking; resize() already support shrinking */ \
+		mask = n_buckets - 1; \
+		i = last = __kh_h2b(hash, h->bits); \
+		while (__kh_used(h->used, i) && !__hash_eq(h->keys[i], *key)) { \
+			i = (i + 1U) & mask; \
+			if (i == last) break; \
+		} \
+		if (!__kh_used(h->used, i)) { /* not present at all */ \
+			h->keys[i] = *key; \
+			__kh_set_used(h->used, i); \
+			++h->count; \
+			*absent = 1; \
+		} else *absent = 0; /* Don't touch h->keys[i] if present */ \
+		return i; \
+	} \
+	SCOPE khint_t prefix##_putp(HType *h, const khkey_t *key, int *absent) { return prefix##_putp_core(h, key, __hash_fn(*key), absent); } \
+	SCOPE khint_t prefix##_put(HType *h, khkey_t key, int *absent) { return prefix##_putp_core(h, &key, __hash_fn(key), absent); }
+
+#define __KHASHL_IMPL_DEL(SCOPE, HType, prefix, khkey_t, __hash_fn) \
+	SCOPE int prefix##_del(HType *h, khint_t i) { \
+		khint_t j = i, k, mask, n_buckets; \
+		if (h->keys == 0) return 0; \
+		n_buckets = (khint_t)1U<<h->bits; \
+		mask = n_buckets - 1U; \
+		while (1) { \
+			j = (j + 1U) & mask; \
+			if (j == i || !__kh_used(h->used, j)) break; /* j==i only when the table is completely full */ \
+			k = __kh_h2b(__hash_fn(h->keys[j]), h->bits); \
+			if ((j > i && (k <= i || k > j)) || (j < i && (k <= i && k > j))) \
+				h->keys[i] = h->keys[j], i = j; \
+		} \
+		__kh_set_unused(h->used, i); \
+		--h->count; \
+		return 1; \
+	}
+
+#define KHASHL_DECLARE(HType, prefix, khkey_t) \
+	__KHASHL_TYPE(HType, khkey_t) \
+	__KHASHL_PROTOTYPES(HType, prefix, khkey_t)
+
+#define KHASHL_INIT(SCOPE, HType, prefix, khkey_t, __hash_fn, __hash_eq) \
+	__KHASHL_TYPE(HType, khkey_t) \
+	__KHASHL_IMPL_BASIC(SCOPE, HType, prefix) \
+	__KHASHL_IMPL_GET(SCOPE, HType, prefix, khkey_t, __hash_fn, __hash_eq) \
+	__KHASHL_IMPL_RESIZE(SCOPE, HType, prefix, khkey_t, __hash_fn, __hash_eq) \
+	__KHASHL_IMPL_PUT(SCOPE, HType, prefix, khkey_t, __hash_fn, __hash_eq) \
+	__KHASHL_IMPL_DEL(SCOPE, HType, prefix, khkey_t, __hash_fn)
+
+/***************************
+ * Ensemble of hash tables *
+ ***************************/
+
+typedef struct {
+	khint_t sub, pos;
+} kh_ensitr_t;
+
+#define KHASHE_INIT(SCOPE, HType, prefix, khkey_t, __hash_fn, __hash_eq) \
+	KHASHL_INIT(KH_LOCAL, HType##_sub, prefix##_sub, khkey_t, __hash_fn, __hash_eq) \
+	typedef struct HType { \
+		khint64_t count:54, bits:8; \
+		HType##_sub *sub; \
+	} HType; \
+	SCOPE HType *prefix##_init(int bits) { \
+		HType *g; \
+		g = (HType*)kcalloc(1, sizeof(*g)); \
+		g->bits = bits; \
+		g->sub = (HType##_sub*)kcalloc(1U<<bits, sizeof(*g->sub)); \
+		return g; \
+	} \
+	SCOPE void prefix##_destroy(HType *g) { \
+		int t; \
+		if (!g) return; \
+		for (t = 0; t < 1<<g->bits; ++t) { kfree((void*)g->sub[t].keys); kfree(g->sub[t].used); } \
+		kfree(g->sub); kfree(g); \
+	} \
+	SCOPE kh_ensitr_t prefix##_getp(const HType *g, const khkey_t *key) { \
+		khint_t hash, low, ret; \
+		kh_ensitr_t r; \
+		HType##_sub *h; \
+		hash = __hash_fn(*key); \
+		low = hash & ((1U<<g->bits) - 1); \
+		h = &g->sub[low]; \
+		ret = prefix##_sub_getp_core(h, key, hash); \
+		if (ret == kh_end(h)) r.sub = low, r.pos = (khint_t)-1; \
+		else r.sub = low, r.pos = ret; \
+		return r; \
+	} \
+	SCOPE kh_ensitr_t prefix##_get(const HType *g, const khkey_t key) { return prefix##_getp(g, &key); } \
+	SCOPE kh_ensitr_t prefix##_putp(HType *g, const khkey_t *key, int *absent) { \
+		khint_t hash, low, ret; \
+		kh_ensitr_t r; \
+		HType##_sub *h; \
+		hash = __hash_fn(*key); \
+		low = hash & ((1U<<g->bits) - 1); \
+		h = &g->sub[low]; \
+		ret = prefix##_sub_putp_core(h, key, hash, absent); \
+		if (*absent) ++g->count; \
+		r.sub = low, r.pos = ret; \
+		return r; \
+	} \
+	SCOPE kh_ensitr_t prefix##_put(HType *g, const khkey_t key, int *absent) { return prefix##_putp(g, &key, absent); } \
+	SCOPE int prefix##_del(HType *g, kh_ensitr_t itr) { \
+		HType##_sub *h = &g->sub[itr.sub]; \
+		int ret; \
+		ret = prefix##_sub_del(h, itr.pos); \
+		if (ret) --g->count; \
+		return ret; \
+	}
+
+/*****************************
+ * More convenient interface *
+ *****************************/
+
+#define __kh_cached_hash(x) ((x).hash)
+
+#define KHASHL_SET_INIT(SCOPE, HType, prefix, khkey_t, __hash_fn, __hash_eq) \
+	KHASHL_INIT(SCOPE, HType, prefix, khkey_t, __hash_fn, __hash_eq)
+
+#define KHASHL_MAP_INIT(SCOPE, HType, prefix, khkey_t, kh_val_t, __hash_fn, __hash_eq) \
+	typedef struct { khkey_t key; kh_val_t val; } kh_packed HType##_m_bucket_t; \
+	static kh_inline khint_t prefix##_m_hash(HType##_m_bucket_t x) { return __hash_fn(x.key); } \
+	static kh_inline int prefix##_m_eq(HType##_m_bucket_t x, HType##_m_bucket_t y) { return __hash_eq(x.key, y.key); } \
+	KHASHL_INIT(KH_LOCAL, HType, prefix##_m, HType##_m_bucket_t, prefix##_m_hash, prefix##_m_eq) \
+	SCOPE HType *prefix##_init(void) { return prefix##_m_init(); } \
+	SCOPE void prefix##_destroy(HType *h) { prefix##_m_destroy(h); } \
+	SCOPE khint_t prefix##_get(const HType *h, khkey_t key) { HType##_m_bucket_t t; t.key = key; return prefix##_m_getp(h, &t); } \
+	SCOPE int prefix##_del(HType *h, khint_t k) { return prefix##_m_del(h, k); } \
+	SCOPE khint_t prefix##_put(HType *h, khkey_t key, int *absent) { HType##_m_bucket_t t; t.key = key; return prefix##_m_putp(h, &t, absent); }
+
+#define KHASHL_CSET_INIT(SCOPE, HType, prefix, khkey_t, __hash_fn, __hash_eq) \
+	typedef struct { khkey_t key; khint_t hash; } kh_packed HType##_cs_bucket_t; \
+	static kh_inline int prefix##_cs_eq(HType##_cs_bucket_t x, HType##_cs_bucket_t y) { return x.hash == y.hash && __hash_eq(x.key, y.key); } \
+	KHASHL_INIT(KH_LOCAL, HType, prefix##_cs, HType##_cs_bucket_t, __kh_cached_hash, prefix##_cs_eq) \
+	SCOPE HType *prefix##_init(void) { return prefix##_cs_init(); } \
+	SCOPE void prefix##_destroy(HType *h) { prefix##_cs_destroy(h); } \
+	SCOPE khint_t prefix##_get(const HType *h, khkey_t key) { HType##_cs_bucket_t t; t.key = key; t.hash = __hash_fn(key); return prefix##_cs_getp(h, &t); } \
+	SCOPE int prefix##_del(HType *h, khint_t k) { return prefix##_cs_del(h, k); } \
+	SCOPE khint_t prefix##_put(HType *h, khkey_t key, int *absent) { HType##_cs_bucket_t t; t.key = key, t.hash = __hash_fn(key); return prefix##_cs_putp(h, &t, absent); }
+
+#define KHASHL_CMAP_INIT(SCOPE, HType, prefix, khkey_t, kh_val_t, __hash_fn, __hash_eq) \
+	typedef struct { khkey_t key; kh_val_t val; khint_t hash; } kh_packed HType##_cm_bucket_t; \
+	static kh_inline int prefix##_cm_eq(HType##_cm_bucket_t x, HType##_cm_bucket_t y) { return x.hash == y.hash && __hash_eq(x.key, y.key); } \
+	KHASHL_INIT(KH_LOCAL, HType, prefix##_cm, HType##_cm_bucket_t, __kh_cached_hash, prefix##_cm_eq) \
+	SCOPE HType *prefix##_init(void) { return prefix##_cm_init(); } \
+	SCOPE void prefix##_destroy(HType *h) { prefix##_cm_destroy(h); } \
+	SCOPE khint_t prefix##_get(const HType *h, khkey_t key) { HType##_cm_bucket_t t; t.key = key; t.hash = __hash_fn(key); return prefix##_cm_getp(h, &t); } \
+	SCOPE int prefix##_del(HType *h, khint_t k) { return prefix##_cm_del(h, k); } \
+	SCOPE khint_t prefix##_put(HType *h, khkey_t key, int *absent) { HType##_cm_bucket_t t; t.key = key, t.hash = __hash_fn(key); return prefix##_cm_putp(h, &t, absent); }
+
+#define KHASHE_SET_INIT(SCOPE, HType, prefix, khkey_t, __hash_fn, __hash_eq) \
+	KHASHE_INIT(SCOPE, HType, prefix, khkey_t, __hash_fn, __hash_eq)
+
+#define KHASHE_MAP_INIT(SCOPE, HType, prefix, khkey_t, kh_val_t, __hash_fn, __hash_eq) \
+	typedef struct { khkey_t key; kh_val_t val; } kh_packed HType##_m_bucket_t; \
+	static kh_inline khint_t prefix##_m_hash(HType##_m_bucket_t x) { return __hash_fn(x.key); } \
+	static kh_inline int prefix##_m_eq(HType##_m_bucket_t x, HType##_m_bucket_t y) { return __hash_eq(x.key, y.key); } \
+	KHASHE_INIT(KH_LOCAL, HType, prefix##_m, HType##_m_bucket_t, prefix##_m_hash, prefix##_m_eq) \
+	SCOPE HType *prefix##_init(int bits) { return prefix##_m_init(bits); } \
+	SCOPE void prefix##_destroy(HType *h) { prefix##_m_destroy(h); } \
+	SCOPE kh_ensitr_t prefix##_get(const HType *h, khkey_t key) { HType##_m_bucket_t t; t.key = key; return prefix##_m_getp(h, &t); } \
+	SCOPE int prefix##_del(HType *h, kh_ensitr_t k) { return prefix##_m_del(h, k); } \
+	SCOPE kh_ensitr_t prefix##_put(HType *h, khkey_t key, int *absent) { HType##_m_bucket_t t; t.key = key; return prefix##_m_putp(h, &t, absent); }
+
+/**************************
+ * Public macro functions *
+ **************************/
+
+#define kh_bucket(h, x) ((h)->keys[x])
+#define kh_size(h) ((h)->count)
+#define kh_capacity(h) ((h)->keys? 1U<<(h)->bits : 0U)
+#define kh_end(h) kh_capacity(h)
+
+#define kh_key(h, x) ((h)->keys[x].key)
+#define kh_val(h, x) ((h)->keys[x].val)
+#define kh_exist(h, x) __kh_used((h)->used, (x))
+
+#define kh_foreach(h, x) for ((x) = 0; (x) != kh_end(h); ++(x)) if (kh_exist((h), (x)))
+
+#define kh_ens_key(g, x) kh_key(&(g)->sub[(x).sub], (x).pos)
+#define kh_ens_val(g, x) kh_val(&(g)->sub[(x).sub], (x).pos)
+#define kh_ens_exist(g, x) kh_exist(&(g)->sub[(x).sub], (x).pos)
+#define kh_ens_is_end(x) ((x).pos == (khint_t)-1)
+#define kh_ens_size(g) ((g)->count)
+
+#define kh_ens_foreach(g, x) for ((x).sub = 0; (x).sub != 1<<(g)->bits; ++(x).sub) for ((x).pos = 0; (x).pos != kh_end(&(g)->sub[(x).sub]); ++(x).pos) if (kh_ens_exist((g), (x)))
+
+/**************************************
+ * Common hash and equality functions *
+ **************************************/
+
+#define kh_eq_generic(a, b) ((a) == (b))
+#define kh_eq_str(a, b) (strcmp((a), (b)) == 0)
+#define kh_hash_dummy(x) ((khint_t)(x))
+
+static kh_inline khint_t kh_hash_uint32(khint_t x) { /* murmur finishing */
+	x ^= x >> 16;
+	x *= 0x85ebca6bU;
+	x ^= x >> 13;
+	x *= 0xc2b2ae35U;
+	x ^= x >> 16;
+	return x;
+}
+
+static kh_inline khint_t kh_hash_uint64(khint64_t x) { /* splitmix64; see https://nullprogram.com/blog/2018/07/31/ for inversion */
+	x ^= x >> 30;
+	x *= 0xbf58476d1ce4e5b9ULL;
+	x ^= x >> 27;
+	x *= 0x94d049bb133111ebULL;
+	x ^= x >> 31;
+	return (khint_t)x;
+}
+
+#define KH_FNV_SEED 11
+
+static kh_inline khint_t kh_hash_str(kh_cstr_t s) { /* FNV1a */
+	khint_t h = KH_FNV_SEED ^ 2166136261U;
+	const unsigned char *t = (const unsigned char*)s;
+	for (; *t; ++t)
+		h ^= *t, h *= 16777619;
+	return h;
+}
+
+static kh_inline khint_t kh_hash_bytes(int len, const unsigned char *s) {
+	khint_t h = KH_FNV_SEED ^ 2166136261U;
+	int i;
+	for (i = 0; i < len; ++i)
+		h ^= s[i], h *= 16777619;
+	return h;
+}
+
+#endif /* __AC_KHASHL_H */
diff --git a/ext/klib/khmm.c b/ext/klib/khmm.c
new file mode 100644
index 0000000..711ade5
--- /dev/null
+++ b/ext/klib/khmm.c
@@ -0,0 +1,423 @@
+#include <math.h>
+#include <stdio.h>
+#include <assert.h>
+#include <string.h>
+#include <stdlib.h>
+#include "khmm.h"
+
+// new/delete hmm_par_t
+
+hmm_par_t *hmm_new_par(int m, int n)
+{
+	hmm_par_t *hp;
+	int i;
+	assert(m > 0 && n > 0);
+	hp = (hmm_par_t*)calloc(1, sizeof(hmm_par_t));
+	hp->m = m; hp->n = n;
+	hp->a0 = (FLOAT*)calloc(n, sizeof(FLOAT));
+	hp->a = (FLOAT**)calloc2(n, n, sizeof(FLOAT));
+	hp->e = (FLOAT**)calloc2(m + 1, n, sizeof(FLOAT));
+	hp->ae = (FLOAT**)calloc2((m + 1) * n, n, sizeof(FLOAT));
+	for (i = 0; i != n; ++i) hp->e[m][i] = 1.0;
+	return hp;
+}
+void hmm_delete_par(hmm_par_t *hp)
+{
+	int i;
+	if (hp == 0) return;
+	for (i = 0; i != hp->n; ++i) free(hp->a[i]);
+	for (i = 0; i <= hp->m; ++i) free(hp->e[i]);
+	for (i = 0; i < (hp->m + 1) * hp->n; ++i) free(hp->ae[i]);
+	free(hp->a); free(hp->e); free(hp->a0); free(hp->ae);
+	free(hp);
+}
+
+// new/delete hmm_data_t
+
+hmm_data_t *hmm_new_data(int L, const char *seq, const hmm_par_t *hp)
+{
+	hmm_data_t *hd;
+	hd = (hmm_data_t*)calloc(1, sizeof(hmm_data_t));
+	hd->L = L;
+	hd->seq = (char*)malloc(L + 1);
+	memcpy(hd->seq + 1, seq, L);
+	return hd;
+}
+void hmm_delete_data(hmm_data_t *hd)
+{
+	int i;
+	if (hd == 0) return;
+	for (i = 0; i <= hd->L; ++i) {
+		if (hd->f) free(hd->f[i]);
+		if (hd->b) free(hd->b[i]);
+	}
+	free(hd->f); free(hd->b); free(hd->s); free(hd->v); free(hd->p); free(hd->seq);
+	free(hd);
+}
+
+// new/delete hmm_exp_t
+
+hmm_exp_t *hmm_new_exp(const hmm_par_t *hp)
+{
+	hmm_exp_t *he;
+	assert(hp);
+	he = (hmm_exp_t*)calloc(1, sizeof(hmm_exp_t));
+	he->m = hp->m; he->n = hp->n;
+	he->A0 = (FLOAT*)calloc(hp->n, sizeof(FLOAT));
+	he->A = (FLOAT**)calloc2(hp->n, hp->n, sizeof(FLOAT));
+	he->E = (FLOAT**)calloc2(hp->m + 1, hp->n, sizeof(FLOAT));
+	return he;
+}
+void hmm_delete_exp(hmm_exp_t *he)
+{
+	int i;
+	if (he == 0) return;
+	for (i = 0; i != he->n; ++i) free(he->A[i]);
+	for (i = 0; i <= he->m; ++i) free(he->E[i]);
+	free(he->A); free(he->E); free(he->A0);
+	free(he);
+}
+
+// Viterbi algorithm
+
+FLOAT hmm_Viterbi(const hmm_par_t *hp, hmm_data_t *hd)
+{
+	FLOAT **la, **le, *preV, *curV, max;
+	int **Vmax, max_l; // backtrace matrix
+	int k, l, b, u;
+	
+	if (hd->v) free(hd->v);
+	hd->v = (int*)calloc(hd->L+1, sizeof(int));
+	la = (FLOAT**)calloc2(hp->n, hp->n, sizeof(FLOAT));
+	le = (FLOAT**)calloc2(hp->m + 1, hp->n, sizeof(FLOAT));
+	Vmax = (int**)calloc2(hd->L+1, hp->n, sizeof(int));
+	preV = (FLOAT*)malloc(sizeof(FLOAT) * hp->n);
+	curV = (FLOAT*)malloc(sizeof(FLOAT) * hp->n);
+	for (k = 0; k != hp->n; ++k)
+		for (l = 0; l != hp->n; ++l)
+			la[k][l] = log(hp->a[l][k]); // this is not a bug
+	for (b = 0; b != hp->m; ++b)
+		for (k = 0; k != hp->n; ++k)
+			le[b][k] = log(hp->e[b][k]);
+	for (k = 0; k != hp->n; ++k) le[hp->m][k] = 0.0;
+	// V_k(1)
+	for (k = 0; k != hp->n; ++k) {
+		preV[k] = le[(int)hd->seq[1]][k] + log(hp->a0[k]);
+		Vmax[1][k] = 0;
+	}
+	// all the rest
+	for (u = 2; u <= hd->L; ++u) {
+		FLOAT *tmp, *leu = le[(int)hd->seq[u]];
+		for (k = 0; k != hp->n; ++k) {
+			FLOAT *laa = la[k];
+			for (l = 0, max = -HMM_INF, max_l = -1; l != hp->n; ++l) {
+				if (max < preV[l] + laa[l]) {
+					max = preV[l] + laa[l];
+					max_l = l;
+				}
+			}
+			assert(max_l >= 0); // cannot be zero
+			curV[k] = leu[k] + max;
+			Vmax[u][k] = max_l;
+		}
+		tmp = curV; curV = preV; preV = tmp; // swap
+	}
+	// backtrace
+	for (k = 0, max_l = -1, max = -HMM_INF; k != hp->n; ++k) {
+		if (max < preV[k]) {
+			max = preV[k]; max_l = k;
+		}
+	}
+	assert(max_l >= 0); // cannot be zero
+	hd->v[hd->L] = max_l;
+	for (u = hd->L; u >= 1; --u)
+		hd->v[u-1] = Vmax[u][hd->v[u]];
+	for (k = 0; k != hp->n; ++k) free(la[k]);
+	for (b = 0; b < hp->m; ++b) free(le[b]);
+	for (u = 0; u <= hd->L; ++u) free(Vmax[u]);
+	free(la); free(le); free(Vmax); free(preV); free(curV);
+	hd->status |= HMM_VITERBI;
+	return max;
+}
+
+// forward algorithm
+
+void hmm_forward(const hmm_par_t *hp, hmm_data_t *hd)
+{
+	FLOAT sum, tmp, **at;
+	int u, k, l;
+	int n, m, L;
+	assert(hp && hd);
+	// allocate memory for hd->f and hd->s
+	n = hp->n; m = hp->m; L = hd->L;
+	if (hd->s) free(hd->s);
+	if (hd->f) { 
+		for (k = 0; k <= hd->L; ++k) free(hd->f[k]);
+		free(hd->f);
+	}
+	hd->f = (FLOAT**)calloc2(hd->L+1, hp->n, sizeof(FLOAT));
+	hd->s = (FLOAT*)calloc(hd->L+1, sizeof(FLOAT));
+	hd->status &= ~(unsigned)HMM_FORWARD;
+	// at[][] array helps to improve the cache efficiency
+	at = (FLOAT**)calloc2(n, n, sizeof(FLOAT));
+	// transpose a[][]
+	for (k = 0; k != n; ++k)
+		for (l = 0; l != n; ++l)
+			at[k][l] = hp->a[l][k];
+	// f[0], but it should never be used
+	hd->s[0] = 1.0;
+	for (k = 0; k != n; ++k) hd->f[0][k] = 0.0;
+	// f[1]
+	for (k = 0, sum = 0.0; k != n; ++k)
+		sum += (hd->f[1][k] = hp->a0[k] * hp->e[(int)hd->seq[1]][k]);
+	for (k = 0; k != n; ++k) hd->f[1][k] /= sum;
+	hd->s[1] = sum;
+	// f[2..hmmL], the core loop
+	for (u = 2; u <= L; ++u) {
+		FLOAT *fu = hd->f[u], *fu1 = hd->f[u-1], *eu = hp->e[(int)hd->seq[u]];
+		for (k = 0, sum = 0.0; k != n; ++k) {
+			FLOAT *aa = at[k];
+			for (l = 0, tmp = 0.0; l != n; ++l) tmp += fu1[l] * aa[l];
+			sum += (fu[k] = eu[k] * tmp);
+		}
+		for (k = 0; k != n; ++k) fu[k] /= sum;
+		hd->s[u] = sum;
+	}
+	// free at array
+	for (k = 0; k != hp->n; ++k) free(at[k]);
+	free(at);
+	hd->status |= HMM_FORWARD;
+}
+
+//  precalculate hp->ae
+
+void hmm_pre_backward(hmm_par_t *hp)
+{
+	int m, n, b, k, l;
+	assert(hp);
+	m = hp->m; n = hp->n;
+	for (b = 0; b <= m; ++b) {
+		for (k = 0; k != n; ++k) {
+			FLOAT *p = hp->ae[b * hp->n + k];
+			for (l = 0; l != n; ++l)
+				p[l] = hp->e[b][l] * hp->a[k][l];
+		}
+	}
+}
+
+// backward algorithm
+
+void hmm_backward(const hmm_par_t *hp, hmm_data_t *hd)
+{
+	FLOAT tmp;
+	int k, l, u;
+	int m, n, L;
+	assert(hp && hd);
+	assert(hd->status & HMM_FORWARD);
+	// allocate memory for hd->b
+	m = hp->m; n = hp->n; L = hd->L;
+	if (hd->b) { 
+		for (k = 0; k <= hd->L; ++k) free(hd->b[k]);
+		free(hd->b);
+	}
+	hd->status &= ~(unsigned)HMM_BACKWARD;
+	hd->b = (FLOAT**)calloc2(L+1, hp->n, sizeof(FLOAT));
+	// b[L]
+	for (k = 0; k != hp->n; ++k) hd->b[L][k] = 1.0 / hd->s[L];
+	// b[1..L-1], the core loop
+	for (u = L-1; u >= 1; --u) {
+		FLOAT *bu1 = hd->b[u+1], **p = hp->ae + (int)hd->seq[u+1] * n;
+		for (k = 0; k != n; ++k) {
+			FLOAT *q = p[k];
+			for (l = 0, tmp = 0.0; l != n; ++l) tmp += q[l] * bu1[l];
+			hd->b[u][k] = tmp / hd->s[u];
+		}
+	}
+	hd->status |= HMM_BACKWARD;
+	for (l = 0, tmp = 0.0; l != n; ++l)
+		tmp += hp->a0[l] * hd->b[1][l] * hp->e[(int)hd->seq[1]][l];
+	if (tmp > 1.0 + 1e-6 || tmp < 1.0 - 1e-6) // in theory, tmp should always equal to 1
+		fprintf(stderr, "++ Underflow may have happened (%lg).\n", tmp);
+}
+
+// log-likelihood of the observation
+
+FLOAT hmm_lk(const hmm_data_t *hd)
+{
+    FLOAT sum = 0.0, prod = 1.0;
+	int u, L;
+	L = hd->L;
+	assert(hd->status & HMM_FORWARD);
+	for (u = 1; u <= L; ++u) {
+		prod *= hd->s[u];
+		if (prod < HMM_TINY || prod >= 1.0/HMM_TINY) { // reset
+			sum += log(prod);
+			prod = 1.0;
+		}
+	}
+	sum += log(prod);
+	return sum;
+}
+
+// posterior decoding
+
+void hmm_post_decode(const hmm_par_t *hp, hmm_data_t *hd)
+{
+	int u, k;
+	assert(hd->status && HMM_BACKWARD);
+	if (hd->p) free(hd->p);
+	hd->p = (int*)calloc(hd->L + 1, sizeof(int));
+	for (u = 1; u <= hd->L; ++u) {
+		FLOAT prob, max, *fu = hd->f[u], *bu = hd->b[u], su = hd->s[u];
+		int max_k;
+		for (k = 0, max = -1.0, max_k = -1; k != hp->n; ++k) {
+			if (max < (prob = fu[k] * bu[k] * su)) {
+				max = prob; max_k = k;
+			}
+		}
+		assert(max_k >= 0);
+		hd->p[u] = max_k;
+	}
+	hd->status |= HMM_POSTDEC;
+}
+
+// posterior probability of states
+
+FLOAT hmm_post_state(const hmm_par_t *hp, const hmm_data_t *hd, int u, FLOAT *prob)
+{
+	FLOAT sum = 0.0, ss = hd->s[u], *fu = hd->f[u], *bu = hd->b[u];
+	int k;
+	for (k = 0; k != hp->n; ++k)
+		sum += (prob[k] = fu[k] * bu[k] * ss);
+	return sum; // in theory, this should always equal to 1.0
+}
+
+// expected counts
+
+hmm_exp_t *hmm_expect(const hmm_par_t *hp, const hmm_data_t *hd)
+{
+	int k, l, u, b, m, n;
+	hmm_exp_t *he;
+	assert(hd->status & HMM_BACKWARD);
+	he = hmm_new_exp(hp);
+	// initialization
+	m = hp->m; n = hp->n;
+	for (k = 0; k != n; ++k)
+		for (l = 0; l != n; ++l) he->A[k][l] = HMM_TINY;
+	for (b = 0; b <= m; ++b)
+		for (l = 0; l != n; ++l) he->E[b][l] = HMM_TINY;
+	// calculate A_{kl} and E_k(b), k,l\in[0,n)
+	for (u = 1; u < hd->L; ++u) {
+		FLOAT *fu = hd->f[u], *bu = hd->b[u], *bu1 = hd->b[u+1], ss = hd->s[u];
+		FLOAT *Ec = he->E[(int)hd->seq[u]], **p = hp->ae + (int)hd->seq[u+1] * n;
+		for (k = 0; k != n; ++k) {
+			FLOAT *q = p[k], *AA = he->A[k], fuk = fu[k];
+			for (l = 0; l != n; ++l) // this is cache-efficient
+				AA[l] += fuk * q[l] * bu1[l];
+			Ec[k] += fuk * bu[k] * ss;
+		}
+	}
+	// calculate A0_l
+	for (l = 0; l != n; ++l)
+		he->A0[l] += hp->a0[l] * hp->e[(int)hd->seq[1]][l] * hd->b[1][l];
+	return he;
+}
+
+FLOAT hmm_Q0(const hmm_par_t *hp, hmm_exp_t *he)
+{
+	int k, l, b;
+	FLOAT sum = 0.0;
+	for (k = 0; k != hp->n; ++k) {
+		FLOAT tmp;
+		for (b = 0, tmp = 0.0; b != hp->m; ++b) tmp += he->E[b][k];
+		for (b = 0; b != hp->m; ++b)
+			sum += he->E[b][k] * log(he->E[b][k] / tmp);
+	}
+	for (k = 0; k != hp->n; ++k) {
+		FLOAT tmp, *A = he->A[k];
+		for (l = 0, tmp = 0.0; l != hp->n; ++l) tmp += A[l];
+		for (l = 0; l != hp->n; ++l) sum += A[l] * log(A[l] / tmp);
+	}
+	return (he->Q0 = sum);
+}
+
+// add he0 to he1
+
+void hmm_add_expect(const hmm_exp_t *he0, hmm_exp_t *he1)
+{
+	int b, k, l;
+	assert(he0->m == he1->m && he0->n == he1->n);
+	for (k = 0; k != he1->n; ++k) {
+		he1->A0[k] += he0->A0[k];
+		for (l = 0; l != he1->n; ++l)
+			he1->A[k][l] += he0->A[k][l];
+	}
+	for (b = 0; b != he1->m; ++b) {
+		for (l = 0; l != he1->n; ++l)
+			he1->E[b][l] += he0->E[b][l];
+	}
+}
+
+// the EM-Q function
+
+FLOAT hmm_Q(const hmm_par_t *hp, const hmm_exp_t *he)
+{
+	FLOAT sum = 0.0;
+	int bb, k, l;
+	for (bb = 0; bb != he->m; ++bb) {
+		FLOAT *eb = hp->e[bb], *Eb = he->E[bb];
+		for (k = 0; k != hp->n; ++k) {
+			if (eb[k] <= 0.0) return -HMM_INF;
+			sum += Eb[k] * log(eb[k]);
+		}
+	}
+	for (k = 0; k != he->n; ++k) {
+		FLOAT *Ak = he->A[k], *ak = hp->a[k];
+		for (l = 0; l != he->n; ++l) {
+			if (ak[l] <= 0.0) return -HMM_INF;
+			sum += Ak[l] * log(ak[l]);
+		}
+	}
+	return (sum -= he->Q0);
+}
+
+// simulate sequence
+
+char *hmm_simulate(const hmm_par_t *hp, int L)
+{
+	int i, k, l, b;
+	FLOAT x, y, **et;
+	char *seq;
+	seq = (char*)calloc(L+1, 1);
+	// calculate the transpose of hp->e[][]
+	et = (FLOAT**)calloc2(hp->n, hp->m, sizeof(FLOAT));
+	for (k = 0; k != hp->n; ++k)
+		for (b = 0; b != hp->m; ++b)
+			et[k][b] = hp->e[b][k];
+	// the initial state, drawn from a0[]
+	x = drand48();
+	for (k = 0, y = 0.0; k != hp->n; ++k) {
+		y += hp->a0[k];
+		if (y >= x) break;
+	}
+	// main loop
+	for (i = 0; i != L; ++i) {
+		FLOAT *el, *ak = hp->a[k];
+		x = drand48();
+		for (l = 0, y = 0.0; l != hp->n; ++l) {
+			y += ak[l];
+			if (y >= x) break;
+		}
+		el = et[l];
+		x = drand48();
+		for (b = 0, y = 0.0; b != hp->m; ++b) {
+			y += el[b];
+			if (y >= x) break;
+		} 
+		seq[i] = b;
+		k = l;
+	}
+	for (k = 0; k != hp->n; ++k) free(et[k]);
+	free(et);
+	return seq;
+}
diff --git a/ext/klib/khmm.h b/ext/klib/khmm.h
new file mode 100644
index 0000000..d87673b
--- /dev/null
+++ b/ext/klib/khmm.h
@@ -0,0 +1,107 @@
+#ifndef AC_SCHMM_H_
+#define AC_SCHMM_H_
+
+/*
+ * Last Modified: 2008-03-10
+ * Version: 0.1.0-8
+ *
+ * 2008-03-10, 0.1.0-8: make icc report two more "VECTORIZED"
+ * 2008-03-10, 0.1.0-7: accelerate for some CPU
+ * 2008-02-07, 0.1.0-6: simulate sequences
+ * 2008-01-15, 0.1.0-5: goodness of fit
+ * 2007-11-20, 0.1.0-4: add function declaration of hmm_post_decode()
+ * 2007-11-09: fix a memory leak
+ */
+
+#include <stdlib.h>
+
+#define HMM_VERSION "0.1.0-7"
+
+#define HMM_FORWARD  0x02
+#define HMM_BACKWARD 0x04
+#define HMM_VITERBI  0x40
+#define HMM_POSTDEC  0x80
+
+#ifndef FLOAT
+#define FLOAT double
+#endif
+#define HMM_TINY     1e-25
+#define HMM_INF      1e300
+
+typedef struct
+{
+	int m, n; // number of symbols, number of states
+	FLOAT **a, **e; // transition matrix and emitting probilities
+	FLOAT **ae; // auxiliary array for acceleration, should be calculated by hmm_pre_backward()
+	FLOAT *a0; // trasition matrix from the start state
+} hmm_par_t;
+
+typedef struct
+{
+	int L;
+	unsigned status;
+	char *seq;
+	FLOAT **f, **b, *s;
+	int *v; // Viterbi path
+	int *p; // posterior decoding
+} hmm_data_t;
+
+typedef struct
+{
+	int m, n;
+	FLOAT Q0, **A, **E, *A0;
+} hmm_exp_t;
+
+typedef struct
+{
+	int l, *obs;
+	FLOAT *thr;
+} hmm_gof_t;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+	/* initialize and destroy hmm_par_t */
+	hmm_par_t *hmm_new_par(int m, int n);
+	void hmm_delete_par(hmm_par_t *hp);
+	/* initialize and destroy hmm_data_t */
+	hmm_data_t *hmm_new_data(int L, const char *seq, const hmm_par_t *hp);
+	void hmm_delete_data(hmm_data_t *hd);
+	/* initialize and destroy hmm_exp_t */
+	hmm_exp_t *hmm_new_exp(const hmm_par_t *hp);
+	void hmm_delete_exp(hmm_exp_t *he);
+	/* Viterbi, forward and backward algorithms */
+	FLOAT hmm_Viterbi(const hmm_par_t *hp, hmm_data_t *hd);
+	void hmm_pre_backward(hmm_par_t *hp);
+	void hmm_forward(const hmm_par_t *hp, hmm_data_t *hd);
+	void hmm_backward(const hmm_par_t *hp, hmm_data_t *hd);
+	/* log-likelihood of the observations (natural based) */
+	FLOAT hmm_lk(const hmm_data_t *hd);
+	/* posterior probability at the position on the sequence */
+	FLOAT hmm_post_state(const hmm_par_t *hp, const hmm_data_t *hd, int u, FLOAT *prob);
+	/* posterior decoding */
+	void hmm_post_decode(const hmm_par_t *hp, hmm_data_t *hd);
+	/* expected counts of transitions and emissions */
+	hmm_exp_t *hmm_expect(const hmm_par_t *hp, const hmm_data_t *hd);
+	/* add he0 counts to he1 counts*/
+	void hmm_add_expect(const hmm_exp_t *he0, hmm_exp_t *he1);
+	/* the Q function that should be maximized in EM */
+	FLOAT hmm_Q(const hmm_par_t *hp, const hmm_exp_t *he);
+	FLOAT hmm_Q0(const hmm_par_t *hp, hmm_exp_t *he);
+	/* simulate sequences */
+	char *hmm_simulate(const hmm_par_t *hp, int L);
+#ifdef __cplusplus
+}
+#endif
+
+static inline void **calloc2(int n_row, int n_col, int size)
+{
+	char **p;
+	int k;
+	p = (char**)malloc(sizeof(char*) * n_row);
+	for (k = 0; k != n_row; ++k)
+		p[k] = (char*)calloc(n_col, size);
+	return (void**)p;
+}
+
+#endif
diff --git a/ext/klib/klist.h b/ext/klib/klist.h
new file mode 100644
index 0000000..adc3db1
--- /dev/null
+++ b/ext/klib/klist.h
@@ -0,0 +1,135 @@
+/* The MIT License
+
+   Copyright (c) 2008-2009, by Attractive Chaos <attractor@live.co.uk>
+
+   Permission is hereby granted, free of charge, to any person obtaining
+   a copy of this software and associated documentation files (the
+   "Software"), to deal in the Software without restriction, including
+   without limitation the rights to use, copy, modify, merge, publish,
+   distribute, sublicense, and/or sell copies of the Software, and to
+   permit persons to whom the Software is furnished to do so, subject to
+   the following conditions:
+
+   The above copyright notice and this permission notice shall be
+   included in all copies or substantial portions of the Software.
+
+   THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+   EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+   MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+   NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+   BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+   ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+   CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+   SOFTWARE.
+*/
+
+#ifndef _AC_KLIST_H
+#define _AC_KLIST_H
+
+#include <stdlib.h>
+
+#ifndef klib_unused
+#if (defined __clang__ && __clang_major__ >= 3) || (defined __GNUC__ && __GNUC__ >= 3)
+#define klib_unused __attribute__ ((__unused__))
+#else
+#define klib_unused
+#endif
+#endif /* klib_unused */
+
+#define KMEMPOOL_INIT2(SCOPE, name, kmptype_t, kmpfree_f)				\
+	typedef struct {													\
+		size_t cnt, n, max;												\
+		kmptype_t **buf;												\
+	} kmp_##name##_t;													\
+	SCOPE kmp_##name##_t *kmp_init_##name(void) {						\
+		return calloc(1, sizeof(kmp_##name##_t));						\
+	}																	\
+	SCOPE void kmp_destroy_##name(kmp_##name##_t *mp) {					\
+		size_t k;														\
+		for (k = 0; k < mp->n; ++k) {									\
+			kmpfree_f(mp->buf[k]); free(mp->buf[k]);					\
+		}																\
+		free(mp->buf); free(mp);										\
+	}																	\
+	SCOPE kmptype_t *kmp_alloc_##name(kmp_##name##_t *mp) {				\
+		++mp->cnt;														\
+		if (mp->n == 0) return calloc(1, sizeof(kmptype_t));			\
+		return mp->buf[--mp->n];										\
+	}																	\
+	SCOPE void kmp_free_##name(kmp_##name##_t *mp, kmptype_t *p) {		\
+		--mp->cnt;														\
+		if (mp->n == mp->max) {											\
+			mp->max = mp->max? mp->max<<1 : 16;							\
+			mp->buf = realloc(mp->buf, sizeof(kmptype_t *) * mp->max);	\
+		}																\
+		mp->buf[mp->n++] = p;											\
+	}
+
+#define KMEMPOOL_INIT(name, kmptype_t, kmpfree_f)						\
+	KMEMPOOL_INIT2(static inline klib_unused, name, kmptype_t, kmpfree_f)
+
+#define kmempool_t(name) kmp_##name##_t
+#define kmp_init(name) kmp_init_##name()
+#define kmp_destroy(name, mp) kmp_destroy_##name(mp)
+#define kmp_alloc(name, mp) kmp_alloc_##name(mp)
+#define kmp_free(name, mp, p) kmp_free_##name(mp, p)
+
+#define KLIST_INIT2(SCOPE, name, kltype_t, kmpfree_t)					\
+	struct __kl1_##name {												\
+		kltype_t data;													\
+		struct __kl1_##name *next;										\
+	};																	\
+	typedef struct __kl1_##name kl1_##name;								\
+	KMEMPOOL_INIT2(SCOPE, name, kl1_##name, kmpfree_t)					\
+	typedef struct {													\
+		kl1_##name *head, *tail;										\
+		kmp_##name##_t *mp;												\
+		size_t size;													\
+	} kl_##name##_t;													\
+	SCOPE kl_##name##_t *kl_init_##name(void) {							\
+		kl_##name##_t *kl = calloc(1, sizeof(kl_##name##_t));			\
+		kl->mp = kmp_init(name);										\
+		kl->head = kl->tail = kmp_alloc(name, kl->mp);					\
+		kl->head->next = 0;												\
+		return kl;														\
+	}																	\
+	SCOPE void kl_destroy_##name(kl_##name##_t *kl) {					\
+		kl1_##name *p;													\
+		for (p = kl->head; p != kl->tail; p = p->next)					\
+			kmp_free(name, kl->mp, p);									\
+		kmp_free(name, kl->mp, p);										\
+		kmp_destroy(name, kl->mp);										\
+		free(kl);														\
+	}																	\
+	SCOPE kltype_t *kl_pushp_##name(kl_##name##_t *kl) {				\
+		kl1_##name *q, *p = kmp_alloc(name, kl->mp);					\
+		q = kl->tail; p->next = 0; kl->tail->next = p; kl->tail = p;	\
+		++kl->size;														\
+		return &q->data;												\
+	}																	\
+	SCOPE int kl_shift_##name(kl_##name##_t *kl, kltype_t *d) {			\
+		kl1_##name *p;													\
+		if (kl->head->next == 0) return -1;								\
+		--kl->size;														\
+		p = kl->head; kl->head = kl->head->next;						\
+		if (d) *d = p->data;											\
+		kmp_free(name, kl->mp, p);										\
+		return 0;														\
+	}
+
+#define KLIST_INIT(name, kltype_t, kmpfree_t)							\
+	KLIST_INIT2(static inline klib_unused, name, kltype_t, kmpfree_t)
+
+#define kliter_t(name) kl1_##name
+#define klist_t(name) kl_##name##_t
+#define kl_val(iter) ((iter)->data)
+#define kl_next(iter) ((iter)->next)
+#define kl_begin(kl) ((kl)->head)
+#define kl_end(kl) ((kl)->tail)
+
+#define kl_init(name) kl_init_##name()
+#define kl_destroy(name, kl) kl_destroy_##name(kl)
+#define kl_pushp(name, kl) kl_pushp_##name(kl)
+#define kl_shift(name, kl, d) kl_shift_##name(kl, d)
+
+#endif
diff --git a/ext/klib/kmath.c b/ext/klib/kmath.c
new file mode 100644
index 0000000..9368b2c
--- /dev/null
+++ b/ext/klib/kmath.c
@@ -0,0 +1,447 @@
+#include <stdlib.h>
+#include <string.h>
+#include <math.h>
+#include "kmath.h"
+
+/******************************
+ *** Non-linear programming ***
+ ******************************/
+
+/* Hooke-Jeeves algorithm for nonlinear minimization
+ 
+   Based on the pseudocodes by Bell and Pike (CACM 9(9):684-685), and
+   the revision by Tomlin and Smith (CACM 12(11):637-638). Both of the
+   papers are comments on Kaupe's Algorithm 178 "Direct Search" (ACM
+   6(6):313-314). The original algorithm was designed by Hooke and
+   Jeeves (ACM 8:212-229). This program is further revised according to
+   Johnson's implementation at Netlib (opt/hooke.c).
+ 
+   Hooke-Jeeves algorithm is very simple and it works quite well on a
+   few examples. However, it might fail to converge due to its heuristic
+   nature. A possible improvement, as is suggested by Johnson, may be to
+   choose a small r at the beginning to quickly approach to the minimum
+   and a large r at later step to hit the minimum.
+ */
+
+static double __kmin_hj_aux(kmin_f func, int n, double *x1, void *data, double fx1, double *dx, int *n_calls)
+{
+	int k, j = *n_calls;
+	double ftmp;
+	for (k = 0; k != n; ++k) {
+		x1[k] += dx[k];
+		ftmp = func(n, x1, data); ++j;
+		if (ftmp < fx1) fx1 = ftmp;
+		else { /* search the opposite direction */
+			dx[k] = 0.0 - dx[k];
+			x1[k] += dx[k] + dx[k];
+			ftmp = func(n, x1, data); ++j;
+			if (ftmp < fx1) fx1 = ftmp;
+			else x1[k] -= dx[k]; /* back to the original x[k] */
+		}
+	}
+	*n_calls = j;
+	return fx1; /* here: fx1=f(n,x1) */
+}
+
+double kmin_hj(kmin_f func, int n, double *x, void *data, double r, double eps, int max_calls)
+{
+	double fx, fx1, *x1, *dx, radius;
+	int k, n_calls = 0;
+	x1 = (double*)calloc(n, sizeof(double));
+	dx = (double*)calloc(n, sizeof(double));
+	for (k = 0; k != n; ++k) { /* initial directions, based on MGJ */
+		dx[k] = fabs(x[k]) * r;
+		if (dx[k] == 0) dx[k] = r;
+	}
+	radius = r;
+	fx1 = fx = func(n, x, data); ++n_calls;
+	for (;;) {
+		memcpy(x1, x, n * sizeof(double)); /* x1 = x */
+		fx1 = __kmin_hj_aux(func, n, x1, data, fx, dx, &n_calls);
+		while (fx1 < fx) {
+			for (k = 0; k != n; ++k) {
+				double t = x[k];
+				dx[k] = x1[k] > x[k]? fabs(dx[k]) : 0.0 - fabs(dx[k]);
+				x[k] = x1[k];
+				x1[k] = x1[k] + x1[k] - t;
+			}
+			fx = fx1;
+			if (n_calls >= max_calls) break;
+			fx1 = func(n, x1, data); ++n_calls;
+			fx1 = __kmin_hj_aux(func, n, x1, data, fx1, dx, &n_calls);
+			if (fx1 >= fx) break;
+			for (k = 0; k != n; ++k)
+				if (fabs(x1[k] - x[k]) > .5 * fabs(dx[k])) break;
+			if (k == n) break;
+		}
+		if (radius >= eps) {
+			if (n_calls >= max_calls) break;
+			radius *= r;
+			for (k = 0; k != n; ++k) dx[k] *= r;
+		} else break; /* converge */
+	}
+	free(x1); free(dx);
+	return fx1;
+}
+
+// I copied this function somewhere several years ago with some of my modifications, but I forgot the source.
+double kmin_brent(kmin1_f func, double a, double b, void *data, double tol, double *xmin)
+{
+	double bound, u, r, q, fu, tmp, fa, fb, fc, c;
+	const double gold1 = 1.6180339887;
+	const double gold2 = 0.3819660113;
+	const double tiny = 1e-20;
+	const int max_iter = 100;
+
+	double e, d, w, v, mid, tol1, tol2, p, eold, fv, fw;
+	int iter;
+
+	fa = func(a, data); fb = func(b, data);
+	if (fb > fa) { // swap, such that f(a) > f(b)
+		tmp = a; a = b; b = tmp;
+		tmp = fa; fa = fb; fb = tmp;
+	}
+	c = b + gold1 * (b - a), fc = func(c, data); // golden section extrapolation
+	while (fb > fc) {
+		bound = b + 100.0 * (c - b); // the farthest point where we want to go
+		r = (b - a) * (fb - fc);
+		q = (b - c) * (fb - fa);
+		if (fabs(q - r) < tiny) { // avoid 0 denominator
+			tmp = q > r? tiny : 0.0 - tiny;
+		} else tmp = q - r;
+		u = b - ((b - c) * q - (b - a) * r) / (2.0 * tmp); // u is the parabolic extrapolation point
+		if ((b > u && u > c) || (b < u && u < c)) { // u lies between b and c
+			fu = func(u, data);
+			if (fu < fc) { // (b,u,c) bracket the minimum
+				a = b; b = u; fa = fb; fb = fu;
+				break;
+			} else if (fu > fb) { // (a,b,u) bracket the minimum
+				c = u; fc = fu;
+				break;
+			}
+			u = c + gold1 * (c - b); fu = func(u, data); // golden section extrapolation
+		} else if ((c > u && u > bound) || (c < u && u < bound)) { // u lies between c and bound
+			fu = func(u, data);
+			if (fu < fc) { // fb > fc > fu
+				b = c; c = u; u = c + gold1 * (c - b);
+				fb = fc; fc = fu; fu = func(u, data);
+			} else { // (b,c,u) bracket the minimum
+				a = b; b = c; c = u;
+				fa = fb; fb = fc; fc = fu;
+				break;
+			}
+		} else if ((u > bound && bound > c) || (u < bound && bound < c)) { // u goes beyond the bound
+			u = bound; fu = func(u, data);
+		} else { // u goes the other way around, use golden section extrapolation
+			u = c + gold1 * (c - b); fu = func(u, data);
+		}
+		a = b; b = c; c = u;
+		fa = fb; fb = fc; fc = fu;
+	}
+	if (a > c) u = a, a = c, c = u; // swap
+
+	// now, a<b<c, fa>fb and fb<fc, move on to Brent's algorithm
+	e = d = 0.0;
+	w = v = b; fv = fw = fb;
+	for (iter = 0; iter != max_iter; ++iter) {
+		mid = 0.5 * (a + c);
+		tol2 = 2.0 * (tol1 = tol * fabs(b) + tiny);
+		if (fabs(b - mid) <= (tol2 - 0.5 * (c - a))) {
+			*xmin = b; return fb; // found
+		}
+		if (fabs(e) > tol1) {
+			// related to parabolic interpolation
+			r = (b - w) * (fb - fv);
+			q = (b - v) * (fb - fw);
+			p = (b - v) * q - (b - w) * r;
+			q = 2.0 * (q - r);
+			if (q > 0.0) p = 0.0 - p;
+			else q = 0.0 - q;
+			eold = e; e = d;
+			if (fabs(p) >= fabs(0.5 * q * eold) || p <= q * (a - b) || p >= q * (c - b)) {
+				d = gold2 * (e = (b >= mid ? a - b : c - b));
+			} else {
+				d = p / q; u = b + d; // actual parabolic interpolation happens here
+				if (u - a < tol2 || c - u < tol2)
+					d = (mid > b)? tol1 : 0.0 - tol1;
+			}
+		} else d = gold2 * (e = (b >= mid ? a - b : c - b)); // golden section interpolation
+		u = fabs(d) >= tol1 ? b + d : b + (d > 0.0? tol1 : -tol1);
+		fu = func(u, data);
+		if (fu <= fb) { // u is the minimum point so far
+			if (u >= b) a = b;
+			else c = b;
+			v = w; w = b; b = u; fv = fw; fw = fb; fb = fu;
+		} else { // adjust (a,c) and (u,v,w)
+			if (u < b) a = u;
+			else c = u;
+			if (fu <= fw || w == b) {
+				v = w; w = u;
+				fv = fw; fw = fu;
+			} else if (fu <= fv || v == b || v == w) {
+				v = u; fv = fu;
+			}
+		}
+	}
+	*xmin = b;
+	return fb;
+}
+
+static inline float SIGN(float a, float b)
+{
+	return b >= 0 ? (a >= 0 ? a : -a) : (a >= 0 ? -a : a);
+}
+
+double krf_brent(double x1, double x2, double tol, double (*func)(double, void*), void *data, int *err)
+{
+	const int max_iter = 100;
+	const double eps = 3e-8f;
+	int i;
+	double a = x1, b = x2, c = x2, d, e, min1, min2;
+	double fa, fb, fc, p, q, r, s, tol1, xm;
+
+	*err = 0;
+	fa = func(a, data), fb = func(b, data);
+	if ((fa > 0.0f && fb > 0.0f) || (fa < 0.0f && fb < 0.0f)) {
+		*err = -1;
+		return 0.0f;
+	}
+	fc = fb;
+	for (i = 0; i < max_iter; ++i) {
+		if ((fb > 0.0f && fc > 0.0f) || (fb < 0.0f && fc < 0.0f)) {
+			c = a;
+			fc = fa;
+			e = d = b - a;
+		}
+		if (fabs(fc) < fabs(fb)) {
+			a = b, b = c, c = a;
+			fa = fb, fb = fc, fc = fa;
+		}
+		tol1 = 2.0f * eps * fabs(b) + 0.5f * tol;
+		xm = 0.5f * (c - b);
+		if (fabs(xm) <= tol1 || fb == 0.0f)
+			return b;
+		if (fabs(e) >= tol1 && fabs(fa) > fabs(fb)) {
+			s = fb / fa;
+			if (a == c) {
+				p = 2.0f * xm * s;
+				q = 1.0f - s;
+			} else {
+				q = fa / fc;
+				r = fb / fc;
+				p = s * (2.0f * xm * q * (q - r) - (b - a) * (r - 1.0f));
+				q = (q - 1.0f) * (r - 1.0f) * (s - 1.0f);
+			}
+			if (p > 0.0f) q = -q;
+			p = fabs(p);
+			min1 = 3.0f * xm * q - fabs(tol1 * q);
+			min2 = fabs(e * q);
+			if (2.0f * p < (min1 < min2 ? min1 : min2)) {
+				e = d;
+				d = p / q;
+			} else {
+				d = xm;
+				e = d;
+			}
+		} else {
+			d = xm;
+			e = d;
+		}
+		a = b;
+		fa = fb;
+		if (fabs(d) > tol1) b += d;
+		else b += SIGN(tol1, xm);
+		fb = func(b, data);
+	}
+	*err = -2;
+	return 0.0;
+}
+
+/*************************
+ *** Special functions ***
+ *************************/
+
+/* Log gamma function
+ * \log{\Gamma(z)}
+ * AS245, 2nd algorithm, http://lib.stat.cmu.edu/apstat/245
+ */
+double kf_lgamma(double z)
+{
+	double x = 0;
+	x += 0.1659470187408462e-06 / (z+7);
+	x += 0.9934937113930748e-05 / (z+6);
+	x -= 0.1385710331296526     / (z+5);
+	x += 12.50734324009056      / (z+4);
+	x -= 176.6150291498386      / (z+3);
+	x += 771.3234287757674      / (z+2);
+	x -= 1259.139216722289      / (z+1);
+	x += 676.5203681218835      / z;
+	x += 0.9999999999995183;
+	return log(x) - 5.58106146679532777 - z + (z-0.5) * log(z+6.5);
+}
+
+/* complementary error function
+ * \frac{2}{\sqrt{\pi}} \int_x^{\infty} e^{-t^2} dt
+ * AS66, 2nd algorithm, http://lib.stat.cmu.edu/apstat/66
+ */
+double kf_erfc(double x)
+{
+	const double p0 = 220.2068679123761;
+	const double p1 = 221.2135961699311;
+	const double p2 = 112.0792914978709;
+	const double p3 = 33.912866078383;
+	const double p4 = 6.37396220353165;
+	const double p5 = .7003830644436881;
+	const double p6 = .03526249659989109;
+	const double q0 = 440.4137358247522;
+	const double q1 = 793.8265125199484;
+	const double q2 = 637.3336333788311;
+	const double q3 = 296.5642487796737;
+	const double q4 = 86.78073220294608;
+	const double q5 = 16.06417757920695;
+	const double q6 = 1.755667163182642;
+	const double q7 = .08838834764831844;
+	double expntl, z, p;
+	z = fabs(x) * M_SQRT2;
+	if (z > 37.) return x > 0.? 0. : 2.;
+	expntl = exp(z * z * - .5);
+	if (z < 10. / M_SQRT2) // for small z
+	    p = expntl * ((((((p6 * z + p5) * z + p4) * z + p3) * z + p2) * z + p1) * z + p0)
+			/ (((((((q7 * z + q6) * z + q5) * z + q4) * z + q3) * z + q2) * z + q1) * z + q0);
+	else p = expntl / 2.506628274631001 / (z + 1. / (z + 2. / (z + 3. / (z + 4. / (z + .65)))));
+	return x > 0.? 2. * p : 2. * (1. - p);
+}
+
+/* The following computes regularized incomplete gamma functions.
+ * Formulas are taken from Wiki, with additional input from Numerical
+ * Recipes in C (for modified Lentz's algorithm) and AS245
+ * (http://lib.stat.cmu.edu/apstat/245).
+ *
+ * A good online calculator is available at:
+ *
+ *   http://www.danielsoper.com/statcalc/calc23.aspx
+ *
+ * It calculates upper incomplete gamma function, which equals
+ * kf_gammaq(s,z)*tgamma(s).
+ */
+
+#define KF_GAMMA_EPS 1e-14
+#define KF_TINY 1e-290
+
+// regularized lower incomplete gamma function, by series expansion
+static double _kf_gammap(double s, double z)
+{
+	double sum, x;
+	int k;
+	for (k = 1, sum = x = 1.; k < 100; ++k) {
+		sum += (x *= z / (s + k));
+		if (x / sum < KF_GAMMA_EPS) break;
+	}
+	return exp(s * log(z) - z - kf_lgamma(s + 1.) + log(sum));
+}
+// regularized upper incomplete gamma function, by continued fraction
+static double _kf_gammaq(double s, double z)
+{
+	int j;
+	double C, D, f;
+	f = 1. + z - s; C = f; D = 0.;
+	// Modified Lentz's algorithm for computing continued fraction
+	// See Numerical Recipes in C, 2nd edition, section 5.2
+	for (j = 1; j < 100; ++j) {
+		double a = j * (s - j), b = (j<<1) + 1 + z - s, d;
+		D = b + a * D;
+		if (D < KF_TINY) D = KF_TINY;
+		C = b + a / C;
+		if (C < KF_TINY) C = KF_TINY;
+		D = 1. / D;
+		d = C * D;
+		f *= d;
+		if (fabs(d - 1.) < KF_GAMMA_EPS) break;
+	}
+	return exp(s * log(z) - z - kf_lgamma(s) - log(f));
+}
+
+double kf_gammap(double s, double z)
+{
+	return z <= 1. || z < s? _kf_gammap(s, z) : 1. - _kf_gammaq(s, z);
+}
+
+double kf_gammaq(double s, double z)
+{
+	return z <= 1. || z < s? 1. - _kf_gammap(s, z) : _kf_gammaq(s, z);
+}
+
+/* Regularized incomplete beta function. The method is taken from
+ * Numerical Recipe in C, 2nd edition, section 6.4. The following web
+ * page calculates the incomplete beta function, which equals
+ * kf_betai(a,b,x) * gamma(a) * gamma(b) / gamma(a+b):
+ *
+ *   http://www.danielsoper.com/statcalc/calc36.aspx
+ */
+static double kf_betai_aux(double a, double b, double x)
+{
+	double C, D, f;
+	int j;
+	if (x == 0.) return 0.;
+	if (x == 1.) return 1.;
+	f = 1.; C = f; D = 0.;
+	// Modified Lentz's algorithm for computing continued fraction
+	for (j = 1; j < 200; ++j) {
+		double aa, d;
+		int m = j>>1;
+		aa = (j&1)? -(a + m) * (a + b + m) * x / ((a + 2*m) * (a + 2*m + 1))
+			: m * (b - m) * x / ((a + 2*m - 1) * (a + 2*m));
+		D = 1. + aa * D;
+		if (D < KF_TINY) D = KF_TINY;
+		C = 1. + aa / C;
+		if (C < KF_TINY) C = KF_TINY;
+		D = 1. / D;
+		d = C * D;
+		f *= d;
+		if (fabs(d - 1.) < KF_GAMMA_EPS) break;
+	}
+	return exp(kf_lgamma(a+b) - kf_lgamma(a) - kf_lgamma(b) + a * log(x) + b * log(1.-x)) / a / f;
+}
+double kf_betai(double a, double b, double x)
+{
+	return x < (a + 1.) / (a + b + 2.)? kf_betai_aux(a, b, x) : 1. - kf_betai_aux(b, a, 1. - x);
+}
+
+/******************
+ *** Statistics ***
+ ******************/
+
+double km_ks_dist(int na, const double a[], int nb, const double b[]) // a[] and b[] MUST BE sorted
+{
+	int ia = 0, ib = 0;
+	double fa = 0, fb = 0, sup = 0, na1 = 1. / na, nb1 = 1. / nb;
+	while (ia < na || ib < nb) {
+		if (ia == na) fb += nb1, ++ib;
+		else if (ib == nb) fa += na1, ++ia;
+		else if (a[ia] < b[ib]) fa += na1, ++ia;
+		else if (a[ia] > b[ib]) fb += nb1, ++ib;
+		else fa += na1, fb += nb1, ++ia, ++ib;
+		if (sup < fabs(fa - fb)) sup = fabs(fa - fb);
+	}
+	return sup;
+}
+
+#ifdef KF_MAIN
+#include <stdio.h>
+#include "ksort.h"
+KSORT_INIT_GENERIC(double)
+int main(int argc, char *argv[])
+{
+	double x = 5.5, y = 3;
+	double a, b;
+	double xx[] = {0.22, -0.87, -2.39, -1.79, 0.37, -1.54, 1.28, -0.31, -0.74, 1.72, 0.38, -0.17, -0.62, -1.10, 0.30, 0.15, 2.30, 0.19, -0.50, -0.09};
+	double yy[] = {-5.13, -2.19, -2.43, -3.83, 0.50, -3.25, 4.32, 1.63, 5.18, -0.43, 7.11, 4.87, -3.10, -5.81, 3.76, 6.31, 2.58, 0.07, 5.76, 3.50};
+	ks_introsort(double, 20, xx); ks_introsort(double, 20, yy);
+	printf("K-S distance: %f\n", km_ks_dist(20, xx, 20, yy));
+	printf("erfc(%lg): %lg, %lg\n", x, erfc(x), kf_erfc(x));
+	printf("upper-gamma(%lg,%lg): %lg\n", x, y, kf_gammaq(y, x)*tgamma(y));
+	a = 2; b = 2; x = 0.5;
+	printf("incomplete-beta(%lg,%lg,%lg): %lg\n", a, b, x, kf_betai(a, b, x) / exp(kf_lgamma(a+b) - kf_lgamma(a) - kf_lgamma(b)));
+	return 0;
+}
+#endif
diff --git a/ext/klib/kmath.h b/ext/klib/kmath.h
new file mode 100644
index 0000000..1815a14
--- /dev/null
+++ b/ext/klib/kmath.h
@@ -0,0 +1,38 @@
+#ifndef AC_KMATH_H
+#define AC_KMATH_H
+
+#include <stdint.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+	/**************************
+	 * Non-linear programming *
+	 **************************/
+
+	#define KMIN_RADIUS  0.5
+	#define KMIN_EPS     1e-7
+	#define KMIN_MAXCALL 50000
+
+	typedef double (*kmin_f)(int, double*, void*);
+	typedef double (*kmin1_f)(double, void*);
+
+	double kmin_hj(kmin_f func, int n, double *x, void *data, double r, double eps, int max_calls); // Hooke-Jeeves'
+	double kmin_brent(kmin1_f func, double a, double b, void *data, double tol, double *xmin); // Brent's 1-dimenssion
+
+	/*********************
+	 * Special functions *
+	 *********************/
+
+	double kf_lgamma(double z); // log gamma function
+	double kf_erfc(double x); // complementary error function
+	double kf_gammap(double s, double z); // regularized lower incomplete gamma function
+	double kf_gammaq(double s, double z); // regularized upper incomplete gamma function
+	double kf_betai(double a, double b, double x); // regularized incomplete beta function
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif
diff --git a/ext/klib/knetfile.c b/ext/klib/knetfile.c
new file mode 100644
index 0000000..43ccb77
--- /dev/null
+++ b/ext/klib/knetfile.c
@@ -0,0 +1,628 @@
+/* The MIT License
+
+   Copyright (c) 2008 by Genome Research Ltd (GRL).
+                 2010 by Attractive Chaos <attractor@live.co.uk>
+
+   Permission is hereby granted, free of charge, to any person obtaining
+   a copy of this software and associated documentation files (the
+   "Software"), to deal in the Software without restriction, including
+   without limitation the rights to use, copy, modify, merge, publish,
+   distribute, sublicense, and/or sell copies of the Software, and to
+   permit persons to whom the Software is furnished to do so, subject to
+   the following conditions:
+
+   The above copyright notice and this permission notice shall be
+   included in all copies or substantial portions of the Software.
+
+   THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+   EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+   MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+   NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+   BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+   ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+   CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+   SOFTWARE.
+*/
+
+/* Probably I will not do socket programming in the next few years and
+   therefore I decide to heavily annotate this file, for Linux and
+   Windows as well.  -ac */
+
+#include <time.h>
+#include <stdio.h>
+#include <ctype.h>
+#include <stdlib.h>
+#include <string.h>
+#include <errno.h>
+#include <unistd.h>
+#include <sys/types.h>
+
+#ifndef _WIN32
+#include <netdb.h>
+#include <arpa/inet.h>
+#include <sys/socket.h>
+#endif
+
+#include "knetfile.h"
+
+/* In winsock.h, the type of a socket is SOCKET, which is: "typedef
+ * u_int SOCKET". An invalid SOCKET is: "(SOCKET)(~0)", or signed
+ * integer -1. In knetfile.c, I use "int" for socket type
+ * throughout. This should be improved to avoid confusion.
+ *
+ * In Linux/Mac, recv() and read() do almost the same thing. You can see
+ * in the header file that netread() is simply an alias of read(). In
+ * Windows, however, they are different and using recv() is mandatory.
+ */
+
+/* This function tests if the file handler is ready for reading (or
+ * writing if is_read==0). */
+static int socket_wait(int fd, int is_read)
+{
+	fd_set fds, *fdr = 0, *fdw = 0;
+	struct timeval tv;
+	int ret;
+	tv.tv_sec = 5; tv.tv_usec = 0; // 5 seconds time out
+	FD_ZERO(&fds);
+	FD_SET(fd, &fds);
+	if (is_read) fdr = &fds;
+	else fdw = &fds;
+	ret = select(fd+1, fdr, fdw, 0, &tv);
+#ifndef _WIN32
+	if (ret == -1) perror("select");
+#else
+	if (ret == 0)
+		fprintf(stderr, "select time-out\n");
+	else if (ret == SOCKET_ERROR)
+		fprintf(stderr, "select: %d\n", WSAGetLastError());
+#endif
+	return ret;
+}
+
+#ifndef _WIN32
+/* This function does not work with Windows due to the lack of
+ * getaddrinfo() in winsock. It is addapted from an example in "Beej's
+ * Guide to Network Programming" (http://beej.us/guide/bgnet/). */
+static int socket_connect(const char *host, const char *port)
+{
+#define __err_connect(func) do { perror(func); freeaddrinfo(res); return -1; } while (0)
+
+	int ai_err, on = 1, fd;
+	struct linger lng = { 0, 0 };
+	struct addrinfo hints, *res = 0;
+	memset(&hints, 0, sizeof(struct addrinfo));
+	hints.ai_family = AF_UNSPEC;
+	hints.ai_socktype = SOCK_STREAM;
+	/* In Unix/Mac, getaddrinfo() is the most convenient way to get
+	 * server information. */
+	if ((ai_err = getaddrinfo(host, port, &hints, &res)) != 0) { fprintf(stderr, "can't resolve %s:%s: %s\n", host, port, gai_strerror(ai_err)); return -1; }
+	if ((fd = socket(res->ai_family, res->ai_socktype, res->ai_protocol)) == -1) __err_connect("socket");
+	/* The following two setsockopt() are used by ftplib
+	 * (http://nbpfaus.net/~pfau/ftplib/). I am not sure if they
+	 * necessary. */
+	if (setsockopt(fd, SOL_SOCKET, SO_REUSEADDR, &on, sizeof(on)) == -1) __err_connect("setsockopt");
+	if (setsockopt(fd, SOL_SOCKET, SO_LINGER, &lng, sizeof(lng)) == -1) __err_connect("setsockopt");
+	if (connect(fd, res->ai_addr, res->ai_addrlen) != 0) __err_connect("connect");
+	freeaddrinfo(res);
+	return fd;
+}
+#else
+/* MinGW's printf has problem with "%lld" */
+char *int64tostr(char *buf, int64_t x)
+{
+	int cnt;
+	int i = 0;
+	do {
+		buf[i++] = '0' + x % 10;
+		x /= 10;
+	} while (x);
+	buf[i] = 0;
+	for (cnt = i, i = 0; i < cnt/2; ++i) {
+		int c = buf[i]; buf[i] = buf[cnt-i-1]; buf[cnt-i-1] = c;
+	}
+	return buf;
+}
+
+int64_t strtoint64(const char *buf)
+{
+	int64_t x;
+	for (x = 0; *buf != '\0'; ++buf)
+		x = x * 10 + ((int64_t) *buf - 48);
+	return x;
+}
+/* In windows, the first thing is to establish the TCP connection. */
+int knet_win32_init()
+{
+	WSADATA wsaData;
+	return WSAStartup(MAKEWORD(2, 2), &wsaData);
+}
+void knet_win32_destroy()
+{
+	WSACleanup();
+}
+/* A slightly modfied version of the following function also works on
+ * Mac (and presummably Linux). However, this function is not stable on
+ * my Mac. It sometimes works fine but sometimes does not. Therefore for
+ * non-Windows OS, I do not use this one. */
+static SOCKET socket_connect(const char *host, const char *port)
+{
+#define __err_connect(func)										\
+	do {														\
+		fprintf(stderr, "%s: %d\n", func, WSAGetLastError());	\
+		return -1;												\
+	} while (0)
+
+	int on = 1;
+	SOCKET fd;
+	struct linger lng = { 0, 0 };
+	struct sockaddr_in server;
+	struct hostent *hp = 0;
+	// open socket
+	if ((fd = socket(AF_INET, SOCK_STREAM, IPPROTO_TCP)) == INVALID_SOCKET) __err_connect("socket");
+	if (setsockopt(fd, SOL_SOCKET, SO_REUSEADDR, (char*)&on, sizeof(on)) == -1) __err_connect("setsockopt");
+	if (setsockopt(fd, SOL_SOCKET, SO_LINGER, (char*)&lng, sizeof(lng)) == -1) __err_connect("setsockopt");
+	// get host info
+	if (isalpha(host[0])) hp = gethostbyname(host);
+	else {
+		struct in_addr addr;
+		addr.s_addr = inet_addr(host);
+		hp = gethostbyaddr((char*)&addr, 4, AF_INET);
+	}
+	if (hp == 0) __err_connect("gethost");
+	// connect
+	server.sin_addr.s_addr = *((unsigned long*)hp->h_addr);
+	server.sin_family= AF_INET;
+	server.sin_port = htons(atoi(port));
+	if (connect(fd, (struct sockaddr*)&server, sizeof(server)) != 0) __err_connect("connect");
+	// freehostent(hp); // strangely in MSDN, hp is NOT freed (memory leak?!)
+	return fd;
+}
+#endif
+
+static off_t my_netread(int fd, void *buf, off_t len)
+{
+	off_t rest = len, curr, l = 0;
+	/* recv() and read() may not read the required length of data with
+	 * one call. They have to be called repeatedly. */
+	while (rest) {
+		if (socket_wait(fd, 1) <= 0) break; // socket is not ready for reading
+		curr = netread(fd, buf + l, rest);
+		/* According to the glibc manual, section 13.2, a zero returned
+		 * value indicates end-of-file (EOF), which should mean that
+		 * read() will not return zero if EOF has not been met but data
+		 * are not immediately available. */
+		if (curr == 0) break;
+		l += curr; rest -= curr;
+	}
+	return l;
+}
+
+/*************************
+ * FTP specific routines *
+ *************************/
+
+static int kftp_get_response(knetFile *ftp)
+{
+#ifndef _WIN32
+	unsigned char c;
+#else
+	char c;
+#endif
+	int n = 0;
+	char *p;
+	if (socket_wait(ftp->ctrl_fd, 1) <= 0) return 0;
+	while (netread(ftp->ctrl_fd, &c, 1)) { // FIXME: this is *VERY BAD* for unbuffered I/O
+		//fputc(c, stderr);
+		if (n >= ftp->max_response) {
+			ftp->max_response = ftp->max_response? ftp->max_response<<1 : 256;
+			ftp->response = (char*)realloc(ftp->response, ftp->max_response);
+		}
+		ftp->response[n++] = c;
+		if (c == '\n') {
+			if (n >= 4 && isdigit(ftp->response[0]) && isdigit(ftp->response[1]) && isdigit(ftp->response[2])
+				&& ftp->response[3] != '-') break;
+			n = 0;
+			continue;
+		}
+	}
+	if (n < 2) return -1;
+	ftp->response[n-2] = 0;
+	return strtol(ftp->response, &p, 0);
+}
+
+static int kftp_send_cmd(knetFile *ftp, const char *cmd, int is_get)
+{
+	if (socket_wait(ftp->ctrl_fd, 0) <= 0) return -1; // socket is not ready for writing
+	netwrite(ftp->ctrl_fd, cmd, strlen(cmd));
+	return is_get? kftp_get_response(ftp) : 0;
+}
+
+static int kftp_pasv_prep(knetFile *ftp)
+{
+	char *p;
+	int v[6];
+	kftp_send_cmd(ftp, "PASV\r\n", 1);
+	for (p = ftp->response; *p && *p != '('; ++p);
+	if (*p != '(') return -1;
+	++p;
+	sscanf(p, "%d,%d,%d,%d,%d,%d", &v[0], &v[1], &v[2], &v[3], &v[4], &v[5]);
+	memcpy(ftp->pasv_ip, v, 4 * sizeof(int));
+	ftp->pasv_port = (v[4]<<8&0xff00) + v[5];
+	return 0;
+}
+
+
+static int kftp_pasv_connect(knetFile *ftp)
+{
+	char host[80], port[10];
+	if (ftp->pasv_port == 0) {
+		fprintf(stderr, "[kftp_pasv_connect] kftp_pasv_prep() is not called before hand.\n");
+		return -1;
+	}
+	sprintf(host, "%d.%d.%d.%d", ftp->pasv_ip[0], ftp->pasv_ip[1], ftp->pasv_ip[2], ftp->pasv_ip[3]);
+	sprintf(port, "%d", ftp->pasv_port);
+	ftp->fd = socket_connect(host, port);
+	if (ftp->fd == -1) return -1;
+	return 0;
+}
+
+int kftp_connect(knetFile *ftp)
+{
+	ftp->ctrl_fd = socket_connect(ftp->host, ftp->port);
+	if (ftp->ctrl_fd == -1) return -1;
+	kftp_get_response(ftp);
+	kftp_send_cmd(ftp, "USER anonymous\r\n", 1);
+	kftp_send_cmd(ftp, "PASS kftp@\r\n", 1);
+	kftp_send_cmd(ftp, "TYPE I\r\n", 1);
+	return 0;
+}
+
+int kftp_reconnect(knetFile *ftp)
+{
+	if (ftp->ctrl_fd != -1) {
+		netclose(ftp->ctrl_fd);
+		ftp->ctrl_fd = -1;
+	}
+	netclose(ftp->fd);
+	ftp->fd = -1;
+	return kftp_connect(ftp);
+}
+
+// initialize ->type, ->host, ->retr and ->size
+knetFile *kftp_parse_url(const char *fn, const char *mode)
+{
+	knetFile *fp;
+	char *p;
+	int l;
+	if (strstr(fn, "ftp://") != fn) return 0;
+	for (p = (char*)fn + 6; *p && *p != '/'; ++p);
+	if (*p != '/') return 0;
+	l = p - fn - 6;
+	fp = (knetFile*)calloc(1, sizeof(knetFile));
+	fp->type = KNF_TYPE_FTP;
+	fp->fd = -1;
+	/* the Linux/Mac version of socket_connect() also recognizes a port
+	 * like "ftp", but the Windows version does not. */
+	fp->port = strdup("21");
+	fp->host = (char*)calloc(l + 1, 1);
+	if (strchr(mode, 'c')) fp->no_reconnect = 1;
+	strncpy(fp->host, fn + 6, l);
+	fp->retr = (char*)calloc(strlen(p) + 8, 1);
+	sprintf(fp->retr, "RETR %s\r\n", p);
+    fp->size_cmd = (char*)calloc(strlen(p) + 8, 1);
+    sprintf(fp->size_cmd, "SIZE %s\r\n", p);
+	fp->seek_offset = 0;
+	return fp;
+}
+// place ->fd at offset off
+int kftp_connect_file(knetFile *fp)
+{
+	int ret;
+	long long file_size;
+	if (fp->fd != -1) {
+		netclose(fp->fd);
+		if (fp->no_reconnect) kftp_get_response(fp);
+	}
+	kftp_pasv_prep(fp);
+    kftp_send_cmd(fp, fp->size_cmd, 1);
+#ifndef _WIN32
+    if ( sscanf(fp->response,"%*d %lld", &file_size) != 1 )
+    {
+        fprintf(stderr,"[kftp_connect_file] %s\n", fp->response);
+        return -1;
+    }
+#else
+	const char *p = fp->response;
+	while (*p != ' ') ++p;
+	while (*p < '0' || *p > '9') ++p;
+	file_size = strtoint64(p);
+#endif
+	fp->file_size = file_size;
+	if (fp->offset>=0) {
+		char tmp[32];
+#ifndef _WIN32
+		sprintf(tmp, "REST %lld\r\n", (long long)fp->offset);
+#else
+		strcpy(tmp, "REST ");
+		int64tostr(tmp + 5, fp->offset);
+		strcat(tmp, "\r\n");
+#endif
+		kftp_send_cmd(fp, tmp, 1);
+	}
+	kftp_send_cmd(fp, fp->retr, 0);
+	kftp_pasv_connect(fp);
+	ret = kftp_get_response(fp);
+	if (ret != 150) {
+		fprintf(stderr, "[kftp_connect_file] %s\n", fp->response);
+		netclose(fp->fd);
+		fp->fd = -1;
+		return -1;
+	}
+	fp->is_ready = 1;
+	return 0;
+}
+
+
+/**************************
+ * HTTP specific routines *
+ **************************/
+
+knetFile *khttp_parse_url(const char *fn, const char *mode)
+{
+	knetFile *fp;
+	char *p, *proxy, *q;
+	int l;
+	if (strstr(fn, "http://") != fn) return 0;
+	// set ->http_host
+	for (p = (char*)fn + 7; *p && *p != '/'; ++p);
+	l = p - fn - 7;
+	fp = (knetFile*)calloc(1, sizeof(knetFile));
+	fp->http_host = (char*)calloc(l + 1, 1);
+	strncpy(fp->http_host, fn + 7, l);
+	fp->http_host[l] = 0;
+	for (q = fp->http_host; *q && *q != ':'; ++q);
+	if (*q == ':') *q++ = 0;
+	// get http_proxy
+	proxy = getenv("http_proxy");
+	// set ->host, ->port and ->path
+	if (proxy == 0) {
+		fp->host = strdup(fp->http_host); // when there is no proxy, server name is identical to http_host name.
+		fp->port = strdup(*q? q : "80");
+		fp->path = strdup(*p? p : "/");
+	} else {
+		fp->host = (strstr(proxy, "http://") == proxy)? strdup(proxy + 7) : strdup(proxy);
+		for (q = fp->host; *q && *q != ':'; ++q);
+		if (*q == ':') *q++ = 0; 
+		fp->port = strdup(*q? q : "80");
+		fp->path = strdup(fn);
+	}
+	fp->type = KNF_TYPE_HTTP;
+	fp->ctrl_fd = fp->fd = -1;
+	fp->seek_offset = 0;
+	return fp;
+}
+
+int khttp_connect_file(knetFile *fp)
+{
+	int ret, l = 0;
+	char *buf, *p;
+	if (fp->fd != -1) netclose(fp->fd);
+	fp->fd = socket_connect(fp->host, fp->port);
+	buf = (char*)calloc(0x10000, 1); // FIXME: I am lazy... But in principle, 64KB should be large enough.
+	l += sprintf(buf + l, "GET %s HTTP/1.0\r\nHost: %s\r\n", fp->path, fp->http_host);
+    l += sprintf(buf + l, "Range: bytes=%lld-\r\n", (long long)fp->offset);
+	l += sprintf(buf + l, "\r\n");
+	netwrite(fp->fd, buf, l);
+	l = 0;
+	while (netread(fp->fd, buf + l, 1)) { // read HTTP header; FIXME: bad efficiency
+		if (buf[l] == '\n' && l >= 3)
+			if (strncmp(buf + l - 3, "\r\n\r\n", 4) == 0) break;
+		++l;
+	}
+	buf[l] = 0;
+	if (l < 14) { // prematured header
+		netclose(fp->fd);
+		fp->fd = -1;
+		return -1;
+	}
+	ret = strtol(buf + 8, &p, 0); // HTTP return code
+	if (ret == 200 && fp->offset>0) { // 200 (complete result); then skip beginning of the file
+		off_t rest = fp->offset;
+		while (rest) {
+			off_t l = rest < 0x10000? rest : 0x10000;
+			rest -= my_netread(fp->fd, buf, l);
+		}
+	} else if (ret != 206 && ret != 200) {
+		free(buf);
+		fprintf(stderr, "[khttp_connect_file] fail to open file (HTTP code: %d).\n", ret);
+		netclose(fp->fd);
+		fp->fd = -1;
+		return -1;
+	}
+	free(buf);
+	fp->is_ready = 1;
+	return 0;
+}
+
+/********************
+ * Generic routines *
+ ********************/
+
+knetFile *knet_open(const char *fn, const char *mode)
+{
+	knetFile *fp = 0;
+	if (mode[0] != 'r') {
+		fprintf(stderr, "[kftp_open] only mode \"r\" is supported.\n");
+		return 0;
+	}
+	if (strstr(fn, "ftp://") == fn) {
+		fp = kftp_parse_url(fn, mode);
+		if (fp == 0) return 0;
+		if (kftp_connect(fp) == -1) {
+			knet_close(fp);
+			return 0;
+		}
+		kftp_connect_file(fp);
+	} else if (strstr(fn, "http://") == fn) {
+		fp = khttp_parse_url(fn, mode);
+		if (fp == 0) return 0;
+		khttp_connect_file(fp);
+	} else { // local file
+#ifdef _WIN32
+		/* In windows, O_BINARY is necessary. In Linux/Mac, O_BINARY may
+		 * be undefined on some systems, although it is defined on my
+		 * Mac and the Linux I have tested on. */
+		int fd = open(fn, O_RDONLY | O_BINARY);
+#else		
+		int fd = open(fn, O_RDONLY);
+#endif
+		if (fd == -1) {
+			perror("open");
+			return 0;
+		}
+		fp = (knetFile*)calloc(1, sizeof(knetFile));
+		fp->type = KNF_TYPE_LOCAL;
+		fp->fd = fd;
+		fp->ctrl_fd = -1;
+	}
+	if (fp && fp->fd == -1) {
+		knet_close(fp);
+		return 0;
+	}
+	return fp;
+}
+
+knetFile *knet_dopen(int fd, const char *mode)
+{
+	knetFile *fp = (knetFile*)calloc(1, sizeof(knetFile));
+	fp->type = KNF_TYPE_LOCAL;
+	fp->fd = fd;
+	return fp;
+}
+
+off_t knet_read(knetFile *fp, void *buf, off_t len)
+{
+	off_t l = 0;
+	if (fp->fd == -1) return 0;
+	if (fp->type == KNF_TYPE_FTP) {
+		if (fp->is_ready == 0) {
+			if (!fp->no_reconnect) kftp_reconnect(fp);
+			kftp_connect_file(fp);
+		}
+	} else if (fp->type == KNF_TYPE_HTTP) {
+		if (fp->is_ready == 0)
+			khttp_connect_file(fp);
+	}
+	if (fp->type == KNF_TYPE_LOCAL) { // on Windows, the following block is necessary; not on UNIX
+		off_t rest = len, curr;
+		while (rest) {
+			do {
+				curr = read(fp->fd, buf + l, rest);
+			} while (curr < 0 && EINTR == errno);
+			if (curr < 0) return -1;
+			if (curr == 0) break;
+			l += curr; rest -= curr;
+		}
+	} else l = my_netread(fp->fd, buf, len);
+	fp->offset += l;
+	return l;
+}
+
+off_t knet_seek(knetFile *fp, int64_t off, int whence)
+{
+	if (whence == SEEK_SET && off == fp->offset) return 0;
+	if (fp->type == KNF_TYPE_LOCAL) {
+		/* Be aware that lseek() returns the offset after seeking,
+		 * while fseek() returns zero on success. */
+		off_t offset = lseek(fp->fd, off, whence);
+		if (offset == -1) {
+            // Be silent, it is OK for knet_seek to fail when the file is streamed
+            // fprintf(stderr,"[knet_seek] %s\n", strerror(errno));
+			return -1;
+		}
+		fp->offset = offset;
+		return off;
+	} else if (fp->type == KNF_TYPE_FTP) {
+        if (whence==SEEK_CUR)
+            fp->offset += off;
+        else if (whence==SEEK_SET)
+            fp->offset = off;
+        else if ( whence==SEEK_END)
+            fp->offset = fp->file_size+off;
+		fp->is_ready = 0;
+		return off;
+	} else if (fp->type == KNF_TYPE_HTTP) {
+		if (whence == SEEK_END) { // FIXME: can we allow SEEK_END in future?
+			fprintf(stderr, "[knet_seek] SEEK_END is not supported for HTTP. Offset is unchanged.\n");
+			errno = ESPIPE;
+			return -1;
+		}
+        if (whence==SEEK_CUR)
+            fp->offset += off;
+        else if (whence==SEEK_SET)
+            fp->offset = off;
+		fp->is_ready = 0;
+		return off;
+	}
+	errno = EINVAL;
+    fprintf(stderr,"[knet_seek] %s\n", strerror(errno));
+	return -1;
+}
+
+int knet_close(knetFile *fp)
+{
+	if (fp == 0) return 0;
+	if (fp->ctrl_fd != -1) netclose(fp->ctrl_fd); // FTP specific
+	if (fp->fd != -1) {
+		/* On Linux/Mac, netclose() is an alias of close(), but on
+		 * Windows, it is an alias of closesocket(). */
+		if (fp->type == KNF_TYPE_LOCAL) close(fp->fd);
+		else netclose(fp->fd);
+	}
+	free(fp->host); free(fp->port);
+	free(fp->response); free(fp->retr); // FTP specific
+	free(fp->path); free(fp->http_host); // HTTP specific
+	free(fp);
+	return 0;
+}
+
+#ifdef KNETFILE_MAIN
+int main(void)
+{
+	char *buf;
+	knetFile *fp;
+	int type = 4, l;
+#ifdef _WIN32
+	knet_win32_init();
+#endif
+	buf = calloc(0x100000, 1);
+	if (type == 0) {
+		fp = knet_open("knetfile.c", "r");
+		knet_seek(fp, 1000, SEEK_SET);
+	} else if (type == 1) { // NCBI FTP, large file
+		fp = knet_open("ftp://ftp.ncbi.nih.gov/1000genomes/ftp/data/NA12878/alignment/NA12878.chrom6.SLX.SRP000032.2009_06.bam", "r");
+		knet_seek(fp, 2500000000ll, SEEK_SET);
+		l = knet_read(fp, buf, 255);
+	} else if (type == 2) {
+		fp = knet_open("ftp://ftp.sanger.ac.uk/pub4/treefam/tmp/index.shtml", "r");
+		knet_seek(fp, 1000, SEEK_SET);
+	} else if (type == 3) {
+		fp = knet_open("http://www.sanger.ac.uk/Users/lh3/index.shtml", "r");
+		knet_seek(fp, 1000, SEEK_SET);
+	} else if (type == 4) {
+		fp = knet_open("http://www.sanger.ac.uk/Users/lh3/ex1.bam", "r");
+		knet_read(fp, buf, 10000);
+		knet_seek(fp, 20000, SEEK_SET);
+		knet_seek(fp, 10000, SEEK_SET);
+		l = knet_read(fp, buf+10000, 10000000) + 10000;
+	}
+	if (type != 4 && type != 1) {
+		knet_read(fp, buf, 255);
+		buf[255] = 0;
+		printf("%s\n", buf);
+	} else write(fileno(stdout), buf, l);
+	knet_close(fp);
+	free(buf);
+	return 0;
+}
+#endif
diff --git a/ext/klib/knetfile.h b/ext/klib/knetfile.h
new file mode 100644
index 0000000..0a0e66f
--- /dev/null
+++ b/ext/klib/knetfile.h
@@ -0,0 +1,75 @@
+#ifndef KNETFILE_H
+#define KNETFILE_H
+
+#include <stdint.h>
+#include <fcntl.h>
+
+#ifndef _WIN32
+#define netread(fd, ptr, len) read(fd, ptr, len)
+#define netwrite(fd, ptr, len) write(fd, ptr, len)
+#define netclose(fd) close(fd)
+#else
+#include <winsock2.h>
+#define netread(fd, ptr, len) recv(fd, ptr, len, 0)
+#define netwrite(fd, ptr, len) send(fd, ptr, len, 0)
+#define netclose(fd) closesocket(fd)
+#endif
+
+// FIXME: currently I/O is unbuffered
+
+#define KNF_TYPE_LOCAL 1
+#define KNF_TYPE_FTP   2
+#define KNF_TYPE_HTTP  3
+
+typedef struct knetFile_s {
+	int type, fd;
+	int64_t offset;
+	char *host, *port;
+
+	// the following are for FTP only
+	int ctrl_fd, pasv_ip[4], pasv_port, max_response, no_reconnect, is_ready;
+	char *response, *retr, *size_cmd;
+	int64_t seek_offset; // for lazy seek
+    int64_t file_size;
+
+	// the following are for HTTP only
+	char *path, *http_host;
+} knetFile;
+
+#define knet_tell(fp) ((fp)->offset)
+#define knet_fileno(fp) ((fp)->fd)
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#ifdef _WIN32
+	int knet_win32_init();
+	void knet_win32_destroy();
+#endif
+
+	knetFile *knet_open(const char *fn, const char *mode);
+
+	/* 
+	   This only works with local files.
+	 */
+	knetFile *knet_dopen(int fd, const char *mode);
+
+	/*
+	  If ->is_ready==0, this routine updates ->fd; otherwise, it simply
+	  reads from ->fd.
+	 */
+	off_t knet_read(knetFile *fp, void *buf, off_t len);
+
+	/*
+	  This routine only sets ->offset and ->is_ready=0. It does not
+	  communicate with the FTP server.
+	 */
+	off_t knet_seek(knetFile *fp, int64_t off, int whence);
+	int knet_close(knetFile *fp);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif
diff --git a/ext/klib/knhx.c b/ext/klib/knhx.c
new file mode 100644
index 0000000..7448e46
--- /dev/null
+++ b/ext/klib/knhx.c
@@ -0,0 +1,172 @@
+#include <ctype.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include "knhx.h"
+
+typedef struct {
+	int error, n, max;
+	knhx1_t *node;
+} knaux_t;
+
+static inline char *add_node(const char *s, knaux_t *aux, int x)
+{
+	char *p, *nbeg, *nend = 0;
+	knhx1_t *r;
+	if (aux->n == aux->max) {
+		aux->max = aux->max? aux->max<<1 : 8;
+		aux->node = (knhx1_t*)realloc(aux->node, sizeof(knhx1_t) * aux->max);
+	}
+	r = aux->node + (aux->n++);
+	r->n = x; r->parent = -1;
+	for (p = (char*)s, nbeg = p, r->d = -1.0; *p && *p != ',' && *p != ')'; ++p) {
+		if (*p == '[') {
+			if (nend == 0) nend = p;
+			do ++p; while (*p && *p != ']');
+			if (*p == 0) {
+				aux->error |= KNERR_BRACKET;
+				break;
+			}
+		} else if (*p == ':') {
+			if (nend == 0) nend = p;
+			r->d = strtod(p + 1, &p);
+			--p;
+		} else if (!isgraph(*p)) if (nend == 0) nend = p;
+	}
+	if (nend == 0) nend = p;
+	if (nend != nbeg) {
+		r->name = (char*)calloc(nend - nbeg + 1, 1);
+		strncpy(r->name, nbeg, nend - nbeg);
+	} else r->name = strdup("");
+	return p;
+}
+
+knhx1_t *kn_parse(const char *nhx, int *_n, int *_error)
+{
+	char *p;
+	int *stack, top, max;
+	knaux_t *aux;
+	knhx1_t *ret;
+
+#define __push_back(y) do {										\
+		if (top == max) {										\
+			max = max? max<<1 : 16;								\
+			stack = (int*)realloc(stack, sizeof(int) * max);	\
+		}														\
+		stack[top++] = (y);										\
+	} while (0)													\
+
+	stack = 0; top = max = 0;
+	p = (char*)nhx;
+	aux = (knaux_t*)calloc(1, sizeof(knaux_t));
+	while (*p) {
+		while (*p && !isgraph(*p)) ++p;
+		if (*p == 0) break;
+		if (*p == ',') ++p;
+		else if (*p == '(') {
+			__push_back(-1);
+			++p;
+		} else if (*p == ')') {
+			int x = aux->n, m, i;
+			for (i = top - 1; i >= 0; --i)
+				if (stack[i] < 0) break;
+			m = top - 1 - i;
+			p = add_node(p + 1, aux, m);
+			aux->node[x].child = (int*)calloc(m, sizeof(int));
+			for (i = top - 1, m = m - 1; m >= 0; --m, --i) {
+				aux->node[x].child[m] = stack[i];
+				aux->node[stack[i]].parent = x;
+			}
+			top = i;
+			__push_back(x);
+		} else {
+			__push_back(aux->n);
+			p = add_node(p, aux, 0);
+		}
+	}
+	*_n = aux->n;
+	*_error = aux->error;
+	ret = aux->node;
+	free(aux); free(stack);
+	return ret;
+}
+
+#ifndef kroundup32
+#define kroundup32(x) (--(x), (x)|=(x)>>1, (x)|=(x)>>2, (x)|=(x)>>4, (x)|=(x)>>8, (x)|=(x)>>16, ++(x))
+#endif
+
+static inline int kputsn(const char *p, int l, kstring_t *s)
+{
+	if (s->l + l + 1 >= s->m) {
+		s->m = s->l + l + 2;
+		kroundup32(s->m);
+		s->s = (char*)realloc(s->s, s->m);
+	}
+	memcpy(s->s + s->l, p, l);
+	s->l += l; s->s[s->l] = 0;
+	return l;
+}
+
+static inline int kputc(int c, kstring_t *s)
+{
+	if (s->l + 1 >= s->m) {
+		s->m = s->l + 2;
+		kroundup32(s->m);
+		s->s = (char*)realloc(s->s, s->m);
+	}
+	s->s[s->l++] = c; s->s[s->l] = 0;
+	return c;
+}
+
+static void format_node_recur(const knhx1_t *node, const knhx1_t *p, kstring_t *s, char *numbuf)
+{
+	if (p->n) {
+		int i;
+		kputc('(', s);
+		for (i = 0; i < p->n; ++i) {
+			if (i) kputc(',', s);
+			format_node_recur(node, &node[p->child[i]], s, numbuf);
+		}
+		kputc(')', s);
+		if (p->name) kputsn(p->name, strlen(p->name), s);
+		if (p->d >= 0) {
+			sprintf(numbuf, ":%g", p->d);
+			kputsn(numbuf, strlen(numbuf), s);
+		}
+	} else {
+	  kputsn(p->name, strlen(p->name), s);
+	  if (p->d >= 0) {
+	    sprintf(numbuf, ":%g", p->d);
+	    kputsn(numbuf, strlen(numbuf), s);
+	  }
+	}
+}
+
+void kn_format(const knhx1_t *node, int root, kstring_t *s) // TODO: get rid of recursion
+{
+	char numbuf[128];
+	format_node_recur(node, &node[root], s, numbuf);
+}
+
+#ifdef KNHX_MAIN
+int main(int argc, char *argv[])
+{
+	char *s = "((a[abc],d1)x:0.5,((b[&&NHX:S=MOUSE],h2)[&&NHX:S=HUMAN:B=99][blabla][&&NHX:K=foo],c))";
+	knhx1_t *node;
+	int i, j, n, error;
+	kstring_t str;
+	node = kn_parse(s, &n, &error);
+	for (i = 0; i < n; ++i) {
+		knhx1_t *p = node + i;
+		printf("[%d] %s\t%d\t%d\t%g", i, p->name, p->parent, p->n, p->d);
+		for (j = 0; j < p->n; ++j)
+			printf("\t%d", p->child[j]);
+		putchar('\n');
+	}
+	str.l = str.m = 0; str.s = 0;
+	kn_format(node, n-1, &str);
+	puts(str.s);
+	free(str.s);
+	return 0;
+}
+#endif
diff --git a/ext/klib/knhx.h b/ext/klib/knhx.h
new file mode 100644
index 0000000..dbad7dd
--- /dev/null
+++ b/ext/klib/knhx.h
@@ -0,0 +1,35 @@
+#ifndef KNHX_H_
+#define KNHX_H_
+
+#define KNERR_MISSING_LEFT   0x01
+#define KNERR_MISSING_RGHT   0x02
+#define KNERR_BRACKET        0x04
+#define KNERR_COLON          0x08
+
+typedef struct {
+	int parent, n;
+	int *child;
+	char *name;
+	double d;
+} knhx1_t;
+
+#ifndef KSTRING_T
+#define KSTRING_T kstring_t
+typedef struct __kstring_t {
+	size_t l, m;
+	char *s;
+} kstring_t;
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+	knhx1_t *kn_parse(const char *nhx, int *_n, int *_error);
+	void kn_format(const knhx1_t *node, int root, kstring_t *s);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif
diff --git a/ext/klib/kopen.c b/ext/klib/kopen.c
new file mode 100644
index 0000000..ad825da
--- /dev/null
+++ b/ext/klib/kopen.c
@@ -0,0 +1,343 @@
+#include <stdio.h>
+#include <fcntl.h>
+#include <errno.h>
+#include <ctype.h>
+#include <unistd.h>
+#include <string.h>
+#include <stdlib.h>
+#include <signal.h>
+#include <sys/types.h>
+#ifndef _WIN32
+#include <netdb.h>
+#include <arpa/inet.h>
+#include <sys/socket.h>
+#endif
+
+#ifdef _WIN32
+#define _KO_NO_NET
+#endif
+
+#ifndef _KO_NO_NET
+static int socket_wait(int fd, int is_read)
+{
+	fd_set fds, *fdr = 0, *fdw = 0;
+	struct timeval tv;
+	int ret;
+	tv.tv_sec = 5; tv.tv_usec = 0; // 5 seconds time out
+	FD_ZERO(&fds);
+	FD_SET(fd, &fds);
+	if (is_read) fdr = &fds;
+	else fdw = &fds;
+	ret = select(fd+1, fdr, fdw, 0, &tv);
+	if (ret == -1) perror("select");
+	return ret;
+}
+
+static int socket_connect(const char *host, const char *port)
+{
+#define __err_connect(func) do { perror(func); freeaddrinfo(res); return -1; } while (0)
+
+	int ai_err, on = 1, fd;
+	struct linger lng = { 0, 0 };
+	struct addrinfo hints, *res = 0;
+	memset(&hints, 0, sizeof(struct addrinfo));
+	hints.ai_family = AF_UNSPEC;
+	hints.ai_socktype = SOCK_STREAM;
+	if ((ai_err = getaddrinfo(host, port, &hints, &res)) != 0) { fprintf(stderr, "can't resolve %s:%s: %s\n", host, port, gai_strerror(ai_err)); return -1; }
+	if ((fd = socket(res->ai_family, res->ai_socktype, res->ai_protocol)) == -1) __err_connect("socket");
+	if (setsockopt(fd, SOL_SOCKET, SO_REUSEADDR, &on, sizeof(on)) == -1) __err_connect("setsockopt");
+	if (setsockopt(fd, SOL_SOCKET, SO_LINGER, &lng, sizeof(lng)) == -1) __err_connect("setsockopt");
+	if (connect(fd, res->ai_addr, res->ai_addrlen) != 0) __err_connect("connect");
+	freeaddrinfo(res);
+	return fd;
+#undef __err_connect
+}
+
+static int http_open(const char *fn)
+{
+	char *p, *proxy, *q, *http_host, *host, *port, *path, *buf;
+	int fd, ret, l;
+
+	/* parse URL; adapted from khttp_parse_url() in knetfile.c */
+	if (strstr(fn, "http://") != fn) return 0;
+	// set ->http_host
+	for (p = (char*)fn + 7; *p && *p != '/'; ++p);
+	l = p - fn - 7;
+	http_host = calloc(l + 1, 1);
+	strncpy(http_host, fn + 7, l);
+	http_host[l] = 0;
+	for (q = http_host; *q && *q != ':'; ++q);
+	if (*q == ':') *q++ = 0;
+	// get http_proxy
+	proxy = getenv("http_proxy");
+	// set host, port and path
+	if (proxy == 0) {
+		host = strdup(http_host); // when there is no proxy, server name is identical to http_host name.
+		port = strdup(*q? q : "80");
+		path = strdup(*p? p : "/");
+	} else {
+		host = (strstr(proxy, "http://") == proxy)? strdup(proxy + 7) : strdup(proxy);
+		for (q = host; *q && *q != ':'; ++q);
+		if (*q == ':') *q++ = 0; 
+		port = strdup(*q? q : "80");
+		path = strdup(fn);
+	}
+
+	/* connect; adapted from khttp_connect() in knetfile.c */
+	l = 0;
+	fd = socket_connect(host, port);
+	buf = calloc(0x10000, 1); // FIXME: I am lazy... But in principle, 64KB should be large enough.
+	l += sprintf(buf + l, "GET %s HTTP/1.0\r\nHost: %s\r\n", path, http_host);
+	l += sprintf(buf + l, "\r\n");
+	write(fd, buf, l);
+	l = 0;
+	while (read(fd, buf + l, 1)) { // read HTTP header; FIXME: bad efficiency
+		if (buf[l] == '\n' && l >= 3)
+			if (strncmp(buf + l - 3, "\r\n\r\n", 4) == 0) break;
+		++l;
+	}
+	buf[l] = 0;
+	if (l < 14) { // prematured header
+		close(fd);
+		fd = -1;
+	}
+	ret = strtol(buf + 8, &p, 0); // HTTP return code
+	if (ret != 200) {
+		close(fd);
+		fd = -1;
+	}
+	free(buf); free(http_host); free(host); free(port); free(path);
+	return fd;
+}
+
+typedef struct {
+	int max_response, ctrl_fd;
+	char *response;
+} ftpaux_t;
+
+static int kftp_get_response(ftpaux_t *aux)
+{
+	unsigned char c;
+	int n = 0;
+	char *p;
+	if (socket_wait(aux->ctrl_fd, 1) <= 0) return 0;
+	while (read(aux->ctrl_fd, &c, 1)) { // FIXME: this is *VERY BAD* for unbuffered I/O
+		if (n >= aux->max_response) {
+			aux->max_response = aux->max_response? aux->max_response<<1 : 256;
+			aux->response = realloc(aux->response, aux->max_response);
+		}
+		aux->response[n++] = c;
+		if (c == '\n') {
+			if (n >= 4 && isdigit(aux->response[0]) && isdigit(aux->response[1]) && isdigit(aux->response[2])
+				&& aux->response[3] != '-') break;
+			n = 0;
+			continue;
+		}
+	}
+	if (n < 2) return -1;
+	aux->response[n-2] = 0;
+	return strtol(aux->response, &p, 0);
+}
+
+static int kftp_send_cmd(ftpaux_t *aux, const char *cmd, int is_get)
+{
+	if (socket_wait(aux->ctrl_fd, 0) <= 0) return -1; // socket is not ready for writing
+	write(aux->ctrl_fd, cmd, strlen(cmd));
+	return is_get? kftp_get_response(aux) : 0;
+}
+
+static int ftp_open(const char *fn)
+{
+	char *p, *host = 0, *port = 0, *retr = 0;
+	char host2[80], port2[10];
+	int v[6], l, fd = -1, ret, pasv_port, pasv_ip[4];
+	ftpaux_t aux;
+	
+	/* parse URL */
+	if (strstr(fn, "ftp://") != fn) return 0;
+	for (p = (char*)fn + 6; *p && *p != '/'; ++p);
+	if (*p != '/') return 0;
+	l = p - fn - 6;
+	port = strdup("21");
+	host = calloc(l + 1, 1);
+	strncpy(host, fn + 6, l);
+	retr = calloc(strlen(p) + 8, 1);
+	sprintf(retr, "RETR %s\r\n", p);
+	
+	/* connect to ctrl */
+	memset(&aux, 0, sizeof(ftpaux_t));
+	aux.ctrl_fd = socket_connect(host, port);
+	if (aux.ctrl_fd == -1) goto ftp_open_end; /* fail to connect ctrl */
+
+	/* connect to the data stream */
+	kftp_get_response(&aux);
+	kftp_send_cmd(&aux, "USER anonymous\r\n", 1);
+	kftp_send_cmd(&aux, "PASS kopen@\r\n", 1);
+	kftp_send_cmd(&aux, "TYPE I\r\n", 1);
+	kftp_send_cmd(&aux, "PASV\r\n", 1);
+	for (p = aux.response; *p && *p != '('; ++p);
+	if (*p != '(') goto ftp_open_end;
+	++p;
+	sscanf(p, "%d,%d,%d,%d,%d,%d", &v[0], &v[1], &v[2], &v[3], &v[4], &v[5]);
+	memcpy(pasv_ip, v, 4 * sizeof(int));
+	pasv_port = (v[4]<<8&0xff00) + v[5];
+	kftp_send_cmd(&aux, retr, 0);
+	sprintf(host2, "%d.%d.%d.%d", pasv_ip[0], pasv_ip[1], pasv_ip[2], pasv_ip[3]);
+	sprintf(port2, "%d", pasv_port);
+	fd = socket_connect(host2, port2);
+	if (fd == -1) goto ftp_open_end;
+	ret = kftp_get_response(&aux);
+	if (ret != 150) {
+		close(fd);
+		fd = -1;
+	}
+	close(aux.ctrl_fd);
+
+ftp_open_end:
+	free(host); free(port); free(retr); free(aux.response);
+	return fd;
+}
+#endif /* !defined(_KO_NO_NET) */
+
+static char **cmd2argv(const char *cmd)
+{
+	int i, beg, end, argc;
+	char **argv, *p, *q, *str;
+	end = strlen(cmd);
+	for (i = end - 1; i >= 0; --i)
+		if (!isspace(cmd[i])) break;
+	end = i + 1;
+	for (beg = 0; beg < end; ++beg)
+		if (!isspace(cmd[beg])) break;
+	if (beg == end) return 0;
+	for (i = beg + 1, argc = 0; i < end; ++i)
+		if (isspace(cmd[i]) && !isspace(cmd[i-1]))
+			++argc;
+	argv = (char**)calloc(argc + 2, sizeof(void*));
+	argv[0] = str = (char*)calloc(end - beg + 1, 1);
+	strncpy(argv[0], cmd + beg, end - beg);
+	for (i = argc = 1, q = p = str; i < end - beg; ++i)
+		if (isspace(str[i])) str[i] = 0;
+		else if (str[i] && str[i-1] == 0) argv[argc++] = &str[i];
+	return argv;
+}
+
+#define KO_STDIN    1
+#define KO_FILE     2
+#define KO_PIPE     3
+#define KO_HTTP     4
+#define KO_FTP      5
+
+typedef struct {
+	int type, fd;
+	pid_t pid;
+} koaux_t;
+
+void *kopen(const char *fn, int *_fd)
+{
+	koaux_t *aux = 0;
+	*_fd = -1;
+	if (strstr(fn, "http://") == fn) {
+		aux = calloc(1, sizeof(koaux_t));
+		aux->type = KO_HTTP;
+		aux->fd = http_open(fn);
+	} else if (strstr(fn, "ftp://") == fn) {
+		aux = calloc(1, sizeof(koaux_t));
+		aux->type = KO_FTP;
+		aux->fd = ftp_open(fn);
+	} else if (strcmp(fn, "-") == 0) {
+		aux = calloc(1, sizeof(koaux_t));
+		aux->type = KO_STDIN;
+		aux->fd = STDIN_FILENO;
+	} else {
+		const char *p, *q;
+		for (p = fn; *p; ++p)
+			if (!isspace(*p)) break;
+		if (*p == '<') { // pipe open
+			int need_shell, pfd[2];
+			pid_t pid;
+			// a simple check to see if we need to invoke a shell; not always working
+			for (q = p + 1; *q; ++q)
+				if (ispunct(*q) && *q != '.' && *q != '_' && *q != '-' && *q != ':')
+					break;
+			need_shell = (*q != 0);
+			pipe(pfd);
+			pid = vfork();
+			if (pid == -1) { /* vfork() error */
+				close(pfd[0]); close(pfd[1]);
+				return 0;
+			}
+			if (pid == 0) { /* the child process */
+				char **argv; /* FIXME: I do not know if this will lead to a memory leak */
+				close(pfd[0]);
+				dup2(pfd[1], STDOUT_FILENO);
+				close(pfd[1]);
+				if (!need_shell) {
+					argv = cmd2argv(p + 1);
+					execvp(argv[0], argv);
+					free(argv[0]); free(argv);
+				} else execl("/bin/sh", "sh", "-c", p + 1, NULL);
+				exit(1);
+			} else { /* parent process */
+				close(pfd[1]);
+				aux = calloc(1, sizeof(koaux_t));
+				aux->type = KO_PIPE;
+				aux->fd = pfd[0];
+				aux->pid = pid;
+			}
+		} else {
+#ifdef _WIN32
+			*_fd = open(fn, O_RDONLY | O_BINARY);
+#else
+			*_fd = open(fn, O_RDONLY);
+#endif
+			if (*_fd) {
+				aux = calloc(1, sizeof(koaux_t));
+				aux->type = KO_FILE;
+				aux->fd = *_fd;
+			}
+		}
+	}
+	*_fd = aux->fd;
+	return aux;
+}
+
+int kclose(void *a)
+{
+	koaux_t *aux = (koaux_t*)a;
+	if (aux->type == KO_PIPE) {
+		int status;
+		pid_t pid;
+		pid = waitpid(aux->pid, &status, WNOHANG);
+		if (pid != aux->pid) kill(aux->pid, 15);
+	}
+	return 0;
+}
+
+#ifdef _KO_MAIN
+#define BUF_SIZE 0x10000
+int main(int argc, char *argv[])
+{
+	void *x;
+	int l, fd;
+	unsigned char buf[BUF_SIZE];
+	FILE *fp;
+	if (argc == 1) {
+		fprintf(stderr, "Usage: kopen <file>\n");
+		return 1;
+	}
+	x = kopen(argv[1], &fd);
+	fp = fdopen(fd, "r");
+	if (fp == 0) {
+		fprintf(stderr, "ERROR: fail to open the input\n");
+		return 1;
+	}
+	do {
+		if ((l = fread(buf, 1, BUF_SIZE, fp)) != 0)
+			fwrite(buf, 1, l, stdout);
+	} while (l == BUF_SIZE);
+	fclose(fp);
+	kclose(x);
+	return 0;
+}
+#endif
diff --git a/ext/klib/krmq.h b/ext/klib/krmq.h
new file mode 100644
index 0000000..4f4ea75
--- /dev/null
+++ b/ext/klib/krmq.h
@@ -0,0 +1,474 @@
+/* The MIT License
+
+   Copyright (c) 2019 by Attractive Chaos <attractor@live.co.uk>
+
+   Permission is hereby granted, free of charge, to any person obtaining
+   a copy of this software and associated documentation files (the
+   "Software"), to deal in the Software without restriction, including
+   without limitation the rights to use, copy, modify, merge, publish,
+   distribute, sublicense, and/or sell copies of the Software, and to
+   permit persons to whom the Software is furnished to do so, subject to
+   the following conditions:
+
+   The above copyright notice and this permission notice shall be
+   included in all copies or substantial portions of the Software.
+
+   THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+   EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+   MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+   NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+   BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+   ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+   CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+   SOFTWARE.
+*/
+
+/* An example:
+
+#include <stdio.h>
+#include <string.h>
+#include <stdlib.h>
+#include "krmq.h"
+
+struct my_node {
+  char key;
+  KRMQ_HEAD(struct my_node) head;
+};
+#define my_cmp(p, q) (((q)->key < (p)->key) - ((p)->key < (q)->key))
+KRMQ_INIT(my, struct my_node, head, my_cmp)
+
+int main(void) {
+  const char *str = "MNOLKQOPHIA"; // from wiki, except a duplicate
+  struct my_node *root = 0;
+  int i, l = strlen(str);
+  for (i = 0; i < l; ++i) {        // insert in the input order
+    struct my_node *q, *p = malloc(sizeof(*p));
+    p->key = str[i];
+    q = krmq_insert(my, &root, p, 0);
+    if (p != q) free(p);           // if already present, free
+  }
+  krmq_itr_t(my) itr;
+  krmq_itr_first(my, root, &itr);  // place at first
+  do {                             // traverse
+    const struct my_node *p = krmq_at(&itr);
+    putchar(p->key);
+    free((void*)p);                // free node
+  } while (krmq_itr_next(my, &itr));
+  putchar('\n');
+  return 0;
+}
+*/
+
+#ifndef KRMQ_H
+#define KRMQ_H
+
+#ifdef __STRICT_ANSI__
+#define inline __inline__
+#endif
+
+#define KRMQ_MAX_DEPTH 64
+
+#define krmq_size(head, p) ((p)? (p)->head.size : 0)
+#define krmq_size_child(head, q, i) ((q)->head.p[(i)]? (q)->head.p[(i)]->head.size : 0)
+
+#define KRMQ_HEAD(__type) \
+	struct { \
+		__type *p[2], *s; \
+		signed char balance; /* balance factor */ \
+		unsigned size; /* #elements in subtree */ \
+	}
+
+#define __KRMQ_FIND(suf, __scope, __type, __head,  __cmp) \
+	__scope __type *krmq_find_##suf(const __type *root, const __type *x, unsigned *cnt_) { \
+		const __type *p = root; \
+		unsigned cnt = 0; \
+		while (p != 0) { \
+			int cmp; \
+			cmp = __cmp(x, p); \
+			if (cmp >= 0) cnt += krmq_size_child(__head, p, 0) + 1; \
+			if (cmp < 0) p = p->__head.p[0]; \
+			else if (cmp > 0) p = p->__head.p[1]; \
+			else break; \
+		} \
+		if (cnt_) *cnt_ = cnt; \
+		return (__type*)p; \
+	} \
+	__scope __type *krmq_interval_##suf(const __type *root, const __type *x, __type **lower, __type **upper) { \
+		const __type *p = root, *l = 0, *u = 0; \
+		while (p != 0) { \
+			int cmp; \
+			cmp = __cmp(x, p); \
+			if (cmp < 0) u = p, p = p->__head.p[0]; \
+			else if (cmp > 0) l = p, p = p->__head.p[1]; \
+			else { l = u = p; break; } \
+		} \
+		if (lower) *lower = (__type*)l; \
+		if (upper) *upper = (__type*)u; \
+		return (__type*)p; \
+	}
+
+#define __KRMQ_RMQ(suf, __scope, __type, __head,  __cmp, __lt2) \
+	__scope __type *krmq_rmq_##suf(const __type *root, const __type *lo, const __type *up) { /* CLOSED interval */ \
+		const __type *p = root, *path[2][KRMQ_MAX_DEPTH], *min; \
+		int plen[2] = {0, 0}, pcmp[2][KRMQ_MAX_DEPTH], i, cmp, lca; \
+		if (root == 0) return 0; \
+		while (p) { \
+			cmp = __cmp(lo, p); \
+			path[0][plen[0]] = p, pcmp[0][plen[0]++] = cmp; \
+			if (cmp < 0) p = p->__head.p[0]; \
+			else if (cmp > 0) p = p->__head.p[1]; \
+			else break; \
+		} \
+		p = root; \
+		while (p) { \
+			cmp = __cmp(up, p); \
+			path[1][plen[1]] = p, pcmp[1][plen[1]++] = cmp; \
+			if (cmp < 0) p = p->__head.p[0]; \
+			else if (cmp > 0) p = p->__head.p[1]; \
+			else break; \
+		} \
+		for (i = 0; i < plen[0] && i < plen[1]; ++i) /* find the LCA */ \
+			if (path[0][i] == path[1][i] && pcmp[0][i] <= 0 && pcmp[1][i] >= 0) \
+				break; \
+		if (i == plen[0] || i == plen[1]) return 0; /* no elements in the closed interval */ \
+		lca = i, min = path[0][lca]; \
+		for (i = lca + 1; i < plen[0]; ++i) { \
+			if (pcmp[0][i] <= 0) { \
+				if (__lt2(path[0][i], min)) min = path[0][i]; \
+				if (path[0][i]->__head.p[1] && __lt2(path[0][i]->__head.p[1]->__head.s, min)) \
+					min = path[0][i]->__head.p[1]->__head.s; \
+			} \
+		} \
+		for (i = lca + 1; i < plen[1]; ++i) { \
+			if (pcmp[1][i] >= 0) { \
+				if (__lt2(path[1][i], min)) min = path[1][i]; \
+				if (path[1][i]->__head.p[0] && __lt2(path[1][i]->__head.p[0]->__head.s, min)) \
+					min = path[1][i]->__head.p[0]->__head.s; \
+			} \
+		} \
+		return (__type*)min; \
+	}
+
+#define __KRMQ_ROTATE(suf, __type, __head, __lt2) \
+	/* */ \
+	static inline void krmq_update_min_##suf(__type *p, const __type *q, const __type *r) { \
+		p->__head.s = !q || __lt2(p, q->__head.s)? p : q->__head.s; \
+		p->__head.s = !r || __lt2(p->__head.s, r->__head.s)? p->__head.s : r->__head.s; \
+	} \
+	/* one rotation: (a,(b,c)q)p => ((a,b)p,c)q */ \
+	static inline __type *krmq_rotate1_##suf(__type *p, int dir) { /* dir=0 to left; dir=1 to right */ \
+		int opp = 1 - dir; /* opposite direction */ \
+		__type *q = p->__head.p[opp], *s = p->__head.s; \
+		unsigned size_p = p->__head.size; \
+		p->__head.size -= q->__head.size - krmq_size_child(__head, q, dir); \
+		q->__head.size = size_p; \
+		krmq_update_min_##suf(p, p->__head.p[dir], q->__head.p[dir]); \
+		q->__head.s = s; \
+		p->__head.p[opp] = q->__head.p[dir]; \
+		q->__head.p[dir] = p; \
+		return q; \
+	} \
+	/* two consecutive rotations: (a,((b,c)r,d)q)p => ((a,b)p,(c,d)q)r */ \
+	static inline __type *krmq_rotate2_##suf(__type *p, int dir) { \
+		int b1, opp = 1 - dir; \
+		__type *q = p->__head.p[opp], *r = q->__head.p[dir], *s = p->__head.s; \
+		unsigned size_x_dir = krmq_size_child(__head, r, dir); \
+		r->__head.size = p->__head.size; \
+		p->__head.size -= q->__head.size - size_x_dir; \
+		q->__head.size -= size_x_dir + 1; \
+		krmq_update_min_##suf(p, p->__head.p[dir], r->__head.p[dir]); \
+		krmq_update_min_##suf(q, q->__head.p[opp], r->__head.p[opp]); \
+		r->__head.s = s; \
+		p->__head.p[opp] = r->__head.p[dir]; \
+		r->__head.p[dir] = p; \
+		q->__head.p[dir] = r->__head.p[opp]; \
+		r->__head.p[opp] = q; \
+		b1 = dir == 0? +1 : -1; \
+		if (r->__head.balance == b1) q->__head.balance = 0, p->__head.balance = -b1; \
+		else if (r->__head.balance == 0) q->__head.balance = p->__head.balance = 0; \
+		else q->__head.balance = b1, p->__head.balance = 0; \
+		r->__head.balance = 0; \
+		return r; \
+	}
+
+#define __KRMQ_INSERT(suf, __scope, __type, __head, __cmp, __lt2) \
+	__scope __type *krmq_insert_##suf(__type **root_, __type *x, unsigned *cnt_) { \
+		unsigned char stack[KRMQ_MAX_DEPTH]; \
+		__type *path[KRMQ_MAX_DEPTH]; \
+		__type *bp, *bq; \
+		__type *p, *q, *r = 0; /* _r_ is potentially the new root */ \
+		int i, which = 0, top, b1, path_len; \
+		unsigned cnt = 0; \
+		bp = *root_, bq = 0; \
+		/* find the insertion location */ \
+		for (p = bp, q = bq, top = path_len = 0; p; q = p, p = p->__head.p[which]) { \
+			int cmp; \
+			cmp = __cmp(x, p); \
+			if (cmp >= 0) cnt += krmq_size_child(__head, p, 0) + 1; \
+			if (cmp == 0) { \
+				if (cnt_) *cnt_ = cnt; \
+				return p; \
+			} \
+			if (p->__head.balance != 0) \
+				bq = q, bp = p, top = 0; \
+			stack[top++] = which = (cmp > 0); \
+			path[path_len++] = p; \
+		} \
+		if (cnt_) *cnt_ = cnt; \
+		x->__head.balance = 0, x->__head.size = 1, x->__head.p[0] = x->__head.p[1] = 0, x->__head.s = x; \
+		if (q == 0) *root_ = x; \
+		else q->__head.p[which] = x; \
+		if (bp == 0) return x; \
+		for (i = 0; i < path_len; ++i) ++path[i]->__head.size; \
+		for (i = path_len - 1; i >= 0; --i) { \
+			krmq_update_min_##suf(path[i], path[i]->__head.p[0], path[i]->__head.p[1]); \
+			if (path[i]->__head.s != x) break; \
+		} \
+		for (p = bp, top = 0; p != x; p = p->__head.p[stack[top]], ++top) /* update balance factors */ \
+			if (stack[top] == 0) --p->__head.balance; \
+			else ++p->__head.balance; \
+		if (bp->__head.balance > -2 && bp->__head.balance < 2) return x; /* no re-balance needed */ \
+		/* re-balance */ \
+		which = (bp->__head.balance < 0); \
+		b1 = which == 0? +1 : -1; \
+		q = bp->__head.p[1 - which]; \
+		if (q->__head.balance == b1) { \
+			r = krmq_rotate1_##suf(bp, which); \
+			q->__head.balance = bp->__head.balance = 0; \
+		} else r = krmq_rotate2_##suf(bp, which); \
+		if (bq == 0) *root_ = r; \
+		else bq->__head.p[bp != bq->__head.p[0]] = r; \
+		return x; \
+	}
+
+#define __KRMQ_ERASE(suf, __scope, __type, __head, __cmp, __lt2) \
+	__scope __type *krmq_erase_##suf(__type **root_, const __type *x, unsigned *cnt_) { \
+		__type *p, *path[KRMQ_MAX_DEPTH], fake; \
+		unsigned char dir[KRMQ_MAX_DEPTH]; \
+		int i, d = 0, cmp; \
+		unsigned cnt = 0; \
+		fake.__head.p[0] = *root_, fake.__head.p[1] = 0; \
+		if (cnt_) *cnt_ = 0; \
+		if (x) { \
+			for (cmp = -1, p = &fake; cmp; cmp = __cmp(x, p)) { \
+				int which = (cmp > 0); \
+				if (cmp > 0) cnt += krmq_size_child(__head, p, 0) + 1; \
+				dir[d] = which; \
+				path[d++] = p; \
+				p = p->__head.p[which]; \
+				if (p == 0) { \
+					if (cnt_) *cnt_ = 0; \
+					return 0; \
+				} \
+			} \
+			cnt += krmq_size_child(__head, p, 0) + 1; /* because p==x is not counted */ \
+		} else { \
+			for (p = &fake, cnt = 1; p; p = p->__head.p[0]) \
+				dir[d] = 0, path[d++] = p; \
+			p = path[--d]; \
+		} \
+		if (cnt_) *cnt_ = cnt; \
+		for (i = 1; i < d; ++i) --path[i]->__head.size; \
+		if (p->__head.p[1] == 0) { /* ((1,.)2,3)4 => (1,3)4; p=2 */ \
+			path[d-1]->__head.p[dir[d-1]] = p->__head.p[0]; \
+		} else { \
+			__type *q = p->__head.p[1]; \
+			if (q->__head.p[0] == 0) { /* ((1,2)3,4)5 => ((1)2,4)5; p=3,q=2 */ \
+				q->__head.p[0] = p->__head.p[0]; \
+				q->__head.balance = p->__head.balance; \
+				path[d-1]->__head.p[dir[d-1]] = q; \
+				path[d] = q, dir[d++] = 1; \
+				q->__head.size = p->__head.size - 1; \
+			} else { /* ((1,((.,2)3,4)5)6,7)8 => ((1,(2,4)5)3,7)8; p=6 */ \
+				__type *r; \
+				int e = d++; /* backup _d_ */\
+				for (;;) { \
+					dir[d] = 0; \
+					path[d++] = q; \
+					r = q->__head.p[0]; \
+					if (r->__head.p[0] == 0) break; \
+					q = r; \
+				} \
+				r->__head.p[0] = p->__head.p[0]; \
+				q->__head.p[0] = r->__head.p[1]; \
+				r->__head.p[1] = p->__head.p[1]; \
+				r->__head.balance = p->__head.balance; \
+				path[e-1]->__head.p[dir[e-1]] = r; \
+				path[e] = r, dir[e] = 1; \
+				for (i = e + 1; i < d; ++i) --path[i]->__head.size; \
+				r->__head.size = p->__head.size - 1; \
+			} \
+		} \
+		for (i = d - 1; i >= 0; --i) /* not sure why adding condition "path[i]->__head.s==p" doesn't work */ \
+			krmq_update_min_##suf(path[i], path[i]->__head.p[0], path[i]->__head.p[1]); \
+		while (--d > 0) { \
+			__type *q = path[d]; \
+			int which, other, b1 = 1, b2 = 2; \
+			which = dir[d], other = 1 - which; \
+			if (which) b1 = -b1, b2 = -b2; \
+			q->__head.balance += b1; \
+			if (q->__head.balance == b1) break; \
+			else if (q->__head.balance == b2) { \
+				__type *r = q->__head.p[other]; \
+				if (r->__head.balance == -b1) { \
+					path[d-1]->__head.p[dir[d-1]] = krmq_rotate2_##suf(q, which); \
+				} else { \
+					path[d-1]->__head.p[dir[d-1]] = krmq_rotate1_##suf(q, which); \
+					if (r->__head.balance == 0) { \
+						r->__head.balance = -b1; \
+						q->__head.balance = b1; \
+						break; \
+					} else r->__head.balance = q->__head.balance = 0; \
+				} \
+			} \
+		} \
+		*root_ = fake.__head.p[0]; \
+		return p; \
+	}
+
+#define krmq_free(__type, __head, __root, __free) do { \
+		__type *_p, *_q; \
+		for (_p = __root; _p; _p = _q) { \
+			if (_p->__head.p[0] == 0) { \
+				_q = _p->__head.p[1]; \
+				__free(_p); \
+			} else { \
+				_q = _p->__head.p[0]; \
+				_p->__head.p[0] = _q->__head.p[1]; \
+				_q->__head.p[1] = _p; \
+			} \
+		} \
+	} while (0)
+
+#define __KRMQ_ITR(suf, __scope, __type, __head, __cmp) \
+	struct krmq_itr_##suf { \
+		const __type *stack[KRMQ_MAX_DEPTH], **top; \
+	}; \
+	__scope void krmq_itr_first_##suf(const __type *root, struct krmq_itr_##suf *itr) { \
+		const __type *p; \
+		for (itr->top = itr->stack - 1, p = root; p; p = p->__head.p[0]) \
+			*++itr->top = p; \
+	} \
+	__scope int krmq_itr_find_##suf(const __type *root, const __type *x, struct krmq_itr_##suf *itr) { \
+		const __type *p = root; \
+		itr->top = itr->stack - 1; \
+		while (p != 0) { \
+			int cmp; \
+			*++itr->top = p; \
+			cmp = __cmp(x, p); \
+			if (cmp < 0) p = p->__head.p[0]; \
+			else if (cmp > 0) p = p->__head.p[1]; \
+			else break; \
+		} \
+		return p? 1 : 0; \
+	} \
+	__scope int krmq_itr_next_bidir_##suf(struct krmq_itr_##suf *itr, int dir) { \
+		const __type *p; \
+		if (itr->top < itr->stack) return 0; \
+		dir = !!dir; \
+		p = (*itr->top)->__head.p[dir]; \
+		if (p) { /* go down */ \
+			for (; p; p = p->__head.p[!dir]) \
+				*++itr->top = p; \
+			return 1; \
+		} else { /* go up */ \
+			do { \
+				p = *itr->top--; \
+			} while (itr->top >= itr->stack && p == (*itr->top)->__head.p[dir]); \
+			return itr->top < itr->stack? 0 : 1; \
+		} \
+	} \
+
+/**
+ * Insert a node to the tree
+ *
+ * @param suf     name suffix used in KRMQ_INIT()
+ * @param proot   pointer to the root of the tree (in/out: root may change)
+ * @param x       node to insert (in)
+ * @param cnt     number of nodes smaller than or equal to _x_; can be NULL (out)
+ *
+ * @return _x_ if not present in the tree, or the node equal to x.
+ */
+#define krmq_insert(suf, proot, x, cnt) krmq_insert_##suf(proot, x, cnt)
+
+/**
+ * Find a node in the tree
+ *
+ * @param suf     name suffix used in KRMQ_INIT()
+ * @param root    root of the tree
+ * @param x       node value to find (in)
+ * @param cnt     number of nodes smaller than or equal to _x_; can be NULL (out)
+ *
+ * @return node equal to _x_ if present, or NULL if absent
+ */
+#define krmq_find(suf, root, x, cnt) krmq_find_##suf(root, x, cnt)
+#define krmq_interval(suf, root, x, lower, upper) krmq_interval_##suf(root, x, lower, upper)
+#define krmq_rmq(suf, root, lo, up) krmq_rmq_##suf(root, lo, up)
+
+/**
+ * Delete a node from the tree
+ *
+ * @param suf     name suffix used in KRMQ_INIT()
+ * @param proot   pointer to the root of the tree (in/out: root may change)
+ * @param x       node value to delete; if NULL, delete the first node (in)
+ *
+ * @return node removed from the tree if present, or NULL if absent
+ */
+#define krmq_erase(suf, proot, x, cnt) krmq_erase_##suf(proot, x, cnt)
+#define krmq_erase_first(suf, proot) krmq_erase_##suf(proot, 0, 0)
+
+#define krmq_itr_t(suf) struct krmq_itr_##suf
+
+/**
+ * Place the iterator at the smallest object
+ *
+ * @param suf     name suffix used in KRMQ_INIT()
+ * @param root    root of the tree
+ * @param itr     iterator
+ */
+#define krmq_itr_first(suf, root, itr) krmq_itr_first_##suf(root, itr)
+
+/**
+ * Place the iterator at the object equal to or greater than the query
+ *
+ * @param suf     name suffix used in KRMQ_INIT()
+ * @param root    root of the tree
+ * @param x       query (in)
+ * @param itr     iterator (out)
+ *
+ * @return 1 if find; 0 otherwise. krmq_at(itr) is NULL if and only if query is
+ *         larger than all objects in the tree
+ */
+#define krmq_itr_find(suf, root, x, itr) krmq_itr_find_##suf(root, x, itr)
+
+/**
+ * Move to the next object in order
+ *
+ * @param itr     iterator (modified)
+ *
+ * @return 1 if there is a next object; 0 otherwise
+ */
+#define krmq_itr_next(suf, itr) krmq_itr_next_bidir_##suf(itr, 1)
+#define krmq_itr_prev(suf, itr) krmq_itr_next_bidir_##suf(itr, 0)
+
+/**
+ * Return the pointer at the iterator
+ *
+ * @param itr     iterator
+ *
+ * @return pointer if present; NULL otherwise
+ */
+#define krmq_at(itr) ((itr)->top < (itr)->stack? 0 : *(itr)->top)
+
+#define KRMQ_INIT2(suf, __scope, __type, __head, __cmp, __lt2) \
+	__KRMQ_FIND(suf, __scope, __type, __head,  __cmp) \
+	__KRMQ_RMQ(suf, __scope, __type, __head,  __cmp, __lt2) \
+	__KRMQ_ROTATE(suf, __type, __head, __lt2) \
+	__KRMQ_INSERT(suf, __scope, __type, __head, __cmp, __lt2) \
+	__KRMQ_ERASE(suf, __scope, __type, __head, __cmp, __lt2) \
+	__KRMQ_ITR(suf, __scope, __type, __head, __cmp)
+
+#define KRMQ_INIT(suf, __type, __head, __cmp, __lt2) \
+	KRMQ_INIT2(suf,, __type, __head, __cmp, __lt2)
+
+#endif
diff --git a/ext/klib/krng.h b/ext/klib/krng.h
new file mode 100644
index 0000000..7423a85
--- /dev/null
+++ b/ext/klib/krng.h
@@ -0,0 +1,54 @@
+#ifndef KRNG_H
+#define KRNG_H
+
+typedef struct {
+	uint64_t s[2];
+} krng_t;
+
+static inline uint64_t kr_splitmix64(uint64_t x)
+{
+	uint64_t z = (x += 0x9E3779B97F4A7C15ULL);
+	z = (z ^ (z >> 30)) * 0xBF58476D1CE4E5B9ULL;
+	z = (z ^ (z >> 27)) * 0x94D049BB133111EBULL;
+	return z ^ (z >> 31);
+}
+
+static inline uint64_t kr_rand_r(krng_t *r)
+{
+	const uint64_t s0 = r->s[0];
+	uint64_t s1 = r->s[1];
+	const uint64_t result = s0 + s1;
+	s1 ^= s0;
+	r->s[0] = (s0 << 55 | s0 >> 9) ^ s1 ^ (s1 << 14);
+	r->s[1] = s0 << 36 | s0 >> 28;
+	return result;
+}
+
+static inline void kr_jump_r(krng_t *r)
+{
+	static const uint64_t JUMP[] = { 0xbeac0467eba5facbULL, 0xd86b048b86aa9922ULL };
+	uint64_t s0 = 0, s1 = 0;
+	int i, b;
+	for (i = 0; i < 2; ++i)
+		for (b = 0; b < 64; b++) {
+			if (JUMP[i] & 1ULL << b)
+				s0 ^= r->s[0], s1 ^= r->s[1];
+			kr_rand_r(r);
+		}
+	r->s[0] = s0, r->s[1] = s1;
+}
+
+static inline void kr_srand_r(krng_t *r, uint64_t seed)
+{
+	r->s[0] = kr_splitmix64(seed);
+	r->s[1] = kr_splitmix64(r->s[0]);
+}
+
+static inline double kr_drand_r(krng_t *r)
+{
+	union { uint64_t i; double d; } u;
+	u.i = 0x3FFULL << 52 | kr_rand_r(r) >> 12;
+	return u.d - 1.0;
+}
+
+#endif
diff --git a/ext/klib/ksa.c b/ext/klib/ksa.c
new file mode 100644
index 0000000..18f686d
--- /dev/null
+++ b/ext/klib/ksa.c
@@ -0,0 +1,242 @@
+/*
+ * Copyright (c) 2008 Yuta Mori    All Rights Reserved.
+ *               2011 Attractive Chaos <attractor@live.co.uk>
+ *
+ * Permission is hereby granted, free of charge, to any person
+ * obtaining a copy of this software and associated documentation
+ * files (the "Software"), to deal in the Software without
+ * restriction, including without limitation the rights to use,
+ * copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following
+ * conditions:
+ *
+ * The above copyright notice and this permission notice shall be
+ * included in all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES
+ * OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT
+ * HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
+ * WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
+ * OTHER DEALINGS IN THE SOFTWARE.
+ */
+
+/* This is a library for constructing the suffix array for a string containing
+ * multiple sentinels with sentinels all represented by 0. The last symbol in
+ * the string must be a sentinel. The library is modified from an early version
+ * of Yuta Mori's SAIS library, but is slower than the lastest SAIS by about
+ * 30%, partly due to the recent optimization Yuta has applied and partly due
+ * to the extra comparisons between sentinels. This is not the first effort in
+ * supporting multi-sentinel strings, but is probably the easiest to use. */
+
+#include <stdlib.h>
+
+#ifdef _KSA64
+#include <stdint.h>
+typedef int64_t saint_t;
+#define SAINT_MAX INT64_MAX
+#define SAIS_CORE ksa_core64
+#define SAIS_BWT  ksa_bwt64
+#define SAIS_MAIN ksa_sa64
+#else
+#include <limits.h>
+typedef int saint_t;
+#define SAINT_MAX INT_MAX
+#define SAIS_CORE ksa_core
+#define SAIS_BWT  ksa_bwt
+#define SAIS_MAIN ksa_sa
+#endif
+
+/* T is of type "const unsigned char*". If T[i] is a sentinel, chr(i) takes a negative value */
+#define chr(i) (cs == sizeof(saint_t) ? ((const saint_t *)T)[i] : (T[i]? (saint_t)T[i] : i - SAINT_MAX))
+
+/** Count the occurrences of each symbol */
+static void getCounts(const unsigned char *T, saint_t *C, saint_t n, saint_t k, int cs)
+{
+	saint_t i;
+	for (i = 0; i < k; ++i) C[i] = 0;
+	for (i = 0; i < n; ++i) {
+		saint_t c = chr(i);
+		++C[c > 0? c : 0];
+	}
+}
+
+/**
+ * Find the end of each bucket
+ *
+ * @param C   occurrences computed by getCounts(); input
+ * @param B   start/end of each bucket; output
+ * @param k   size of alphabet
+ * @param end compute the end of bucket if true; otherwise compute the end
+ */
+static inline void getBuckets(const saint_t *C, saint_t *B, saint_t k, saint_t end)
+{
+	saint_t i, sum = 0;
+	if (end) for (i = 0; i < k; ++i) sum += C[i], B[i] = sum;
+	else for (i = 0; i < k; ++i) sum += C[i], B[i] = sum - C[i];
+}
+
+/** Induced sort */
+static void induceSA(const unsigned char *T, saint_t *SA, saint_t *C, saint_t *B, saint_t n, saint_t k, saint_t cs)
+{
+	saint_t *b, i, j;
+	saint_t  c0, c1;
+	/* left-to-right induced sort (for L-type) */
+	if (C == B) getCounts(T, C, n, k, cs);
+	getBuckets(C, B, k, 0);	/* find starts of buckets */
+	for (i = 0, b = 0, c1 = -1; i < n; ++i) {
+		j = SA[i], SA[i] = ~j;
+		if (0 < j) { /* >0 if j-1 is L-type; <0 if S-type; ==0 undefined */
+			--j;
+			if ((c0 = chr(j)) != c1) {
+				B[c1 > 0? c1 : 0] = b - SA;
+				c1 = c0;
+				b = SA + B[c1 > 0? c1 : 0];
+			}
+			*b++ = (0 < j && chr(j - 1) < c1) ? ~j : j;
+		}
+	}
+	/* right-to-left induced sort (for S-type) */
+	if (C == B) getCounts(T, C, n, k, cs);
+	getBuckets(C, B, k, 1);	/* find ends of buckets */
+	for (i = n - 1, b = 0, c1 = -1; 0 <= i; --i) {
+		if (0 < (j = SA[i])) { /* the prefix is S-type */
+			--j;
+			if ((c0 = chr(j)) != c1) {
+				B[c1 > 0? c1 : 0] = b - SA;
+				c1 = c0;
+				b = SA + B[c1 > 0? c1 : 0];
+			}
+			if (c0 > 0) *--b = (j == 0 || chr(j - 1) > c1) ? ~j : j;
+		} else SA[i] = ~j; /* if L-type, change the sign */
+	}
+}
+
+/**
+ * Recursively construct the suffix array for a string containing multiple
+ * sentinels. NULL is taken as the sentinel.
+ *
+ * @param T   NULL terminated input string (there can be multiple NULLs)
+ * @param SA  output suffix array
+ * @param fs  working space available in SA (typically 0 when first called)
+ * @param n   length of T, including the trailing NULL
+ * @param k   size of the alphabet (typically 256 when first called)
+ * @param cs  # bytes per element in T; 1 or sizeof(saint_t) (typically 1 when first called)
+ *
+ * @return    0 upon success
+ */
+int SAIS_CORE(const unsigned char *T, saint_t *SA, saint_t fs, saint_t n, saint_t k, int cs)
+{
+	saint_t *C, *B;
+	saint_t  i, j, c, m, q, qlen, name;
+	saint_t  c0, c1;
+
+	/* STAGE I: reduce the problem by at least 1/2 sort all the S-substrings */
+	if (k <= fs) C = SA + n, B = (k <= fs - k) ? C + k : C;
+	else {
+		if ((C = (saint_t*)malloc(k * (1 + (cs == 1)) * sizeof(saint_t))) == NULL) return -2;
+		B = cs == 1? C + k : C;
+	}
+	getCounts(T, C, n, k, cs);
+	getBuckets(C, B, k, 1);	/* find ends of buckets */
+	for (i = 0; i < n; ++i) SA[i] = 0;
+	/* mark L and S (the t array in Nong et al.), and keep the positions of LMS in the buckets */
+	for (i = n - 2, c = 1, c1 = chr(n - 1); 0 <= i; --i, c1 = c0) {
+		if ((c0 = chr(i)) < c1 + c) c = 1; /* c1 = chr(i+1); c==1 if in an S run */
+		else if (c) SA[--B[c1 > 0? c1 : 0]] = i + 1, c = 0;
+	}
+	induceSA(T, SA, C, B, n, k, cs);
+	if (fs < k) free(C);
+	/* pack all the sorted LMS into the first m items of SA 
+	   2*m must be not larger than n (see Nong et al. for the proof) */
+	for (i = 0, m = 0; i < n; ++i) {
+		saint_t p = SA[i];
+		if (p == n - 1) SA[m++] = p;
+		else if (0 < p && chr(p - 1) > (c0 = chr(p))) {
+			for (j = p + 1; j < n && c0 == (c1 = chr(j)); ++j);
+			if (j < n && c0 < c1) SA[m++] = p;
+		}
+	}
+	for (i = m; i < n; ++i) SA[i] = 0;	/* init the name array buffer */
+	/* store the length of all substrings */
+	for (i = n - 2, j = n, c = 1, c1 = chr(n - 1); 0 <= i; --i, c1 = c0) {
+		if ((c0 = chr(i)) < c1 + c) c = 1; /* c1 = chr(i+1) */
+		else if (c) SA[m + ((i + 1) >> 1)] = j - i - 1, j = i + 1, c = 0;
+	}
+	/* find the lexicographic names of all substrings */
+	for (i = 0, name = 0, q = n, qlen = 0; i < m; ++i) {
+		saint_t p = SA[i], plen = SA[m + (p >> 1)], diff = 1;
+		if (plen == qlen) {
+			for (j = 0; j < plen && chr(p + j) == chr(q + j); j++);
+			if (j == plen) diff = 0;
+		}
+		if (diff) ++name, q = p, qlen = plen;
+		SA[m + (p >> 1)] = name;
+	}
+
+	/* STAGE II: solve the reduced problem; recurse if names are not yet unique */
+	if (name < m) {
+		saint_t *RA = SA + n + fs - m - 1;
+		for (i = n - 1, j = m - 1; m <= i; --i)
+			if (SA[i] != 0) RA[j--] = SA[i];
+		RA[m] = 0; // add a sentinel; in the resulting SA, SA[0]==m always stands
+		if (SAIS_CORE((unsigned char *)RA, SA, fs + n - m * 2 - 2, m + 1, name + 1, sizeof(saint_t)) != 0) return -2;
+		for (i = n - 2, j = m - 1, c = 1, c1 = chr(n - 1); 0 <= i; --i, c1 = c0) {
+			if ((c0 = chr(i)) < c1 + c) c = 1;
+			else if (c) RA[j--] = i + 1, c = 0; /* get p1 */
+		}
+		for (i = 0; i < m; ++i) SA[i] = RA[SA[i+1]]; /* get index  */
+	}
+
+	/* STAGE III: induce the result for the original problem */
+	if (k <= fs) C = SA + n, B = (k <= fs - k) ? C + k : C;
+	else {
+		if ((C = (saint_t*)malloc(k * (1 + (cs == 1)) * sizeof(saint_t))) == NULL) return -2;
+		B = cs == 1? C + k : C;
+	}
+	/* put all LMS characters into their buckets */
+	getCounts(T, C, n, k, cs);
+	getBuckets(C, B, k, 1);	/* find ends of buckets */
+	for (i = m; i < n; ++i) SA[i] = 0; /* init SA[m..n-1] */
+	for (i = m - 1; 0 <= i; --i) {
+		j = SA[i], SA[i] = 0;
+		c = chr(j);
+		SA[--B[c > 0? c : 0]] = j;
+	}
+	induceSA(T, SA, C, B, n, k, cs);
+	if (fs < k) free(C);
+	return 0;
+}
+
+/**
+ * Construct the suffix array for a NULL terminated string possibly containing
+ * multiple sentinels (NULLs).
+ *
+ * @param T[0..n-1]  NULL terminated input string
+ * @param SA[0..n-1] output suffix array
+ * @param n          length of the given string, including NULL
+ * @param k          size of the alphabet including the sentinel; no more than 256
+ * @return           0 upon success
+ */
+int SAIS_MAIN(const unsigned char *T, saint_t *SA, saint_t n, int k)
+{
+	if (T == NULL || SA == NULL || T[n - 1] != '\0' || n <= 0) return -1;
+	if (k < 0 || k > 256) k = 256;
+	return SAIS_CORE(T, SA, 0, n, (saint_t)k, 1);
+}
+
+int SAIS_BWT(unsigned char *T, saint_t n, int k)
+{
+	saint_t *SA, i;
+	int ret;
+	if ((SA = malloc(n * sizeof(saint_t))) == 0) return -1;
+	if ((ret = SAIS_MAIN(T, SA, n, k)) != 0) return ret;
+	for (i = 0; i < n; ++i)
+		if (SA[i]) SA[i] = T[SA[i] - 1]; // if SA[i]==0, SA[i]=0
+	for (i = 0; i < n; ++i) T[i] = SA[i];
+	free(SA);
+	return 0;
+}
diff --git a/ext/klib/kseq.h b/ext/klib/kseq.h
new file mode 100644
index 0000000..2959622
--- /dev/null
+++ b/ext/klib/kseq.h
@@ -0,0 +1,242 @@
+/* The MIT License
+
+   Copyright (c) 2008, 2009, 2011 Attractive Chaos <attractor@live.co.uk>
+
+   Permission is hereby granted, free of charge, to any person obtaining
+   a copy of this software and associated documentation files (the
+   "Software"), to deal in the Software without restriction, including
+   without limitation the rights to use, copy, modify, merge, publish,
+   distribute, sublicense, and/or sell copies of the Software, and to
+   permit persons to whom the Software is furnished to do so, subject to
+   the following conditions:
+
+   The above copyright notice and this permission notice shall be
+   included in all copies or substantial portions of the Software.
+
+   THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+   EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+   MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+   NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+   BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+   ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+   CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+   SOFTWARE.
+*/
+
+/* Last Modified: 05MAR2012 */
+
+#ifndef AC_KSEQ_H
+#define AC_KSEQ_H
+
+#include <ctype.h>
+#include <string.h>
+#include <stdlib.h>
+
+#define KS_SEP_SPACE 0 // isspace(): \t, \n, \v, \f, \r
+#define KS_SEP_TAB   1 // isspace() && !' '
+#define KS_SEP_LINE  2 // line separator: "\n" (Unix) or "\r\n" (Windows)
+#define KS_SEP_MAX   2
+
+#define __KS_TYPE(type_t)						\
+	typedef struct __kstream_t {				\
+		unsigned char *buf;						\
+		int begin, end, is_eof;					\
+		type_t f;								\
+	} kstream_t;
+
+#define ks_err(ks) ((ks)->end == -1)
+#define ks_eof(ks) ((ks)->is_eof && (ks)->begin >= (ks)->end)
+#define ks_rewind(ks) ((ks)->is_eof = (ks)->begin = (ks)->end = 0)
+
+#define __KS_BASIC(type_t, __bufsize)								\
+	static inline kstream_t *ks_init(type_t f)						\
+	{																\
+		kstream_t *ks = (kstream_t*)calloc(1, sizeof(kstream_t));	\
+		ks->f = f;													\
+		ks->buf = (unsigned char*)malloc(__bufsize);				\
+		return ks;													\
+	}																\
+	static inline void ks_destroy(kstream_t *ks)					\
+	{																\
+		if (ks) {													\
+			free(ks->buf);											\
+			free(ks);												\
+		}															\
+	}
+
+#define __KS_GETC(__read, __bufsize)						\
+	static inline int ks_getc(kstream_t *ks)				\
+	{														\
+		if (ks_err(ks)) return -3;							\
+		if (ks->is_eof && ks->begin >= ks->end) return -1;	\
+		if (ks->begin >= ks->end) {							\
+			ks->begin = 0;									\
+			ks->end = __read(ks->f, ks->buf, __bufsize);	\
+			if (ks->end == 0) { ks->is_eof = 1; return -1;}	\
+			if (ks->end == -1) { ks->is_eof = 1; return -3;}\
+		}													\
+		return (int)ks->buf[ks->begin++];					\
+	}
+
+#ifndef KSTRING_T
+#define KSTRING_T kstring_t
+typedef struct __kstring_t {
+	size_t l, m;
+	char *s;
+} kstring_t;
+#endif
+
+#ifndef kroundup32
+#define kroundup32(x) (--(x), (x)|=(x)>>1, (x)|=(x)>>2, (x)|=(x)>>4, (x)|=(x)>>8, (x)|=(x)>>16, ++(x))
+#endif
+
+#define __KS_GETUNTIL(__read, __bufsize)								\
+	static int ks_getuntil2(kstream_t *ks, int delimiter, kstring_t *str, int *dret, int append) \
+	{																	\
+		int gotany = 0;													\
+		if (dret) *dret = 0;											\
+		str->l = append? str->l : 0;									\
+		for (;;) {														\
+			int i;														\
+			if (ks_err(ks)) return -3;									\
+			if (ks->begin >= ks->end) {									\
+				if (!ks->is_eof) {										\
+					ks->begin = 0;										\
+					ks->end = __read(ks->f, ks->buf, __bufsize);		\
+					if (ks->end == 0) { ks->is_eof = 1; break; }		\
+					if (ks->end == -1) { ks->is_eof = 1; return -3; }	\
+				} else break;											\
+			}															\
+			if (delimiter == KS_SEP_LINE) {								\
+				unsigned char *sep = memchr(ks->buf + ks->begin, '\n', ks->end - ks->begin); \
+				i = sep != NULL ? sep - ks->buf : ks->end;				\
+			} else if (delimiter > KS_SEP_MAX) {						\
+				unsigned char *sep = memchr(ks->buf + ks->begin, delimiter, ks->end - ks->begin); \
+				i = sep != NULL ? sep - ks->buf : ks->end;				\
+			} else if (delimiter == KS_SEP_SPACE) {						\
+				for (i = ks->begin; i < ks->end; ++i)					\
+					if (isspace(ks->buf[i])) break;						\
+			} else if (delimiter == KS_SEP_TAB) {						\
+				for (i = ks->begin; i < ks->end; ++i)					\
+					if (isspace(ks->buf[i]) && ks->buf[i] != ' ') break; \
+			} else i = 0; /* never come to here! */						\
+			if (str->m - str->l < (size_t)(i - ks->begin + 1)) {		\
+				str->m = str->l + (i - ks->begin) + 1;					\
+				kroundup32(str->m);										\
+				str->s = (char*)realloc(str->s, str->m);				\
+			}															\
+			gotany = 1;													\
+			memcpy(str->s + str->l, ks->buf + ks->begin, i - ks->begin); \
+			str->l = str->l + (i - ks->begin);							\
+			ks->begin = i + 1;											\
+			if (i < ks->end) {											\
+				if (dret) *dret = ks->buf[i];							\
+				break;													\
+			}															\
+		}																\
+		if (!gotany && ks_eof(ks)) return -1;							\
+		if (str->s == 0) {												\
+			str->m = 1;													\
+			str->s = (char*)calloc(1, 1);								\
+		} else if (delimiter == KS_SEP_LINE && str->l > 1 && str->s[str->l-1] == '\r') --str->l; \
+		str->s[str->l] = '\0';											\
+		return str->l;													\
+	} \
+	static inline int ks_getuntil(kstream_t *ks, int delimiter, kstring_t *str, int *dret) \
+	{ return ks_getuntil2(ks, delimiter, str, dret, 0); }
+
+#define KSTREAM_INIT(type_t, __read, __bufsize) \
+	__KS_TYPE(type_t)							\
+	__KS_BASIC(type_t, __bufsize)				\
+	__KS_GETC(__read, __bufsize)				\
+	__KS_GETUNTIL(__read, __bufsize)
+
+#define kseq_rewind(ks) ((ks)->last_char = (ks)->f->is_eof = (ks)->f->begin = (ks)->f->end = 0)
+
+#define __KSEQ_BASIC(SCOPE, type_t)										\
+	SCOPE kseq_t *kseq_init(type_t fd)									\
+	{																	\
+		kseq_t *s = (kseq_t*)calloc(1, sizeof(kseq_t));					\
+		s->f = ks_init(fd);												\
+		return s;														\
+	}																	\
+	SCOPE void kseq_destroy(kseq_t *ks)									\
+	{																	\
+		if (!ks) return;												\
+		free(ks->name.s); free(ks->comment.s); free(ks->seq.s);	free(ks->qual.s); \
+		ks_destroy(ks->f);												\
+		free(ks);														\
+	}
+
+/* Return value:
+   >=0  length of the sequence (normal)
+   -1   end-of-file
+   -2   truncated quality string
+   -3   error reading stream
+ */
+#define __KSEQ_READ(SCOPE) \
+	SCOPE int kseq_read(kseq_t *seq) \
+	{ \
+		int c,r; \
+		kstream_t *ks = seq->f; \
+		if (seq->last_char == 0) { /* then jump to the next header line */ \
+			while ((c = ks_getc(ks)) >= 0 && c != '>' && c != '@'); \
+			if (c < 0) return c; /* end of file or error*/ \
+			seq->last_char = c; \
+		} /* else: the first header char has been read in the previous call */ \
+		seq->comment.l = seq->seq.l = seq->qual.l = 0; /* reset all members */ \
+		if ((r=ks_getuntil(ks, 0, &seq->name, &c)) < 0) return r;  /* normal exit: EOF or error */ \
+		if (c != '\n') ks_getuntil(ks, KS_SEP_LINE, &seq->comment, 0); /* read FASTA/Q comment */ \
+		if (seq->seq.s == 0) { /* we can do this in the loop below, but that is slower */ \
+			seq->seq.m = 256; \
+			seq->seq.s = (char*)malloc(seq->seq.m); \
+		} \
+		while ((c = ks_getc(ks)) >= 0 && c != '>' && c != '+' && c != '@') { \
+			if (c == '\n') continue; /* skip empty lines */ \
+			seq->seq.s[seq->seq.l++] = c; /* this is safe: we always have enough space for 1 char */ \
+			ks_getuntil2(ks, KS_SEP_LINE, &seq->seq, 0, 1); /* read the rest of the line */ \
+		} \
+		if (c == '>' || c == '@') seq->last_char = c; /* the first header char has been read */	\
+		if (seq->seq.l + 1 >= seq->seq.m) { /* seq->seq.s[seq->seq.l] below may be out of boundary */ \
+			seq->seq.m = seq->seq.l + 2; \
+			kroundup32(seq->seq.m); /* rounded to the next closest 2^k */ \
+			seq->seq.s = (char*)realloc(seq->seq.s, seq->seq.m); \
+		} \
+		seq->seq.s[seq->seq.l] = 0;	/* null terminated string */ \
+		if (c != '+') return seq->seq.l; /* FASTA */ \
+		if (seq->qual.m < seq->seq.m) {	/* allocate memory for qual in case insufficient */ \
+			seq->qual.m = seq->seq.m; \
+			seq->qual.s = (char*)realloc(seq->qual.s, seq->qual.m); \
+		} \
+		while ((c = ks_getc(ks)) >= 0 && c != '\n'); /* skip the rest of '+' line */ \
+		if (c == -1) return -2; /* error: no quality string */ \
+		while ((c = ks_getuntil2(ks, KS_SEP_LINE, &seq->qual, 0, 1) >= 0 && seq->qual.l < seq->seq.l)); \
+		if (c == -3) return -3; /* stream error */ \
+		seq->last_char = 0;	/* we have not come to the next header line */ \
+		if (seq->seq.l != seq->qual.l) return -2; /* error: qual string is of a different length */ \
+		return seq->seq.l; \
+	}
+
+#define __KSEQ_TYPE(type_t)						\
+	typedef struct {							\
+		kstring_t name, comment, seq, qual;		\
+		int last_char;							\
+		kstream_t *f;							\
+	} kseq_t;
+
+#define KSEQ_INIT2(SCOPE, type_t, __read)		\
+	KSTREAM_INIT(type_t, __read, 16384)			\
+	__KSEQ_TYPE(type_t)							\
+	__KSEQ_BASIC(SCOPE, type_t)					\
+	__KSEQ_READ(SCOPE)
+
+#define KSEQ_INIT(type_t, __read) KSEQ_INIT2(static, type_t, __read)
+
+#define KSEQ_DECLARE(type_t) \
+	__KS_TYPE(type_t) \
+	__KSEQ_TYPE(type_t) \
+	extern kseq_t *kseq_init(type_t fd); \
+	void kseq_destroy(kseq_t *ks); \
+	int kseq_read(kseq_t *seq);
+
+#endif
diff --git a/ext/klib/kson.c b/ext/klib/kson.c
new file mode 100644
index 0000000..a8bf160
--- /dev/null
+++ b/ext/klib/kson.c
@@ -0,0 +1,253 @@
+#include <string.h>
+#include <stdlib.h>
+#include <stdarg.h>
+#include <assert.h>
+#include <ctype.h>
+#include <stdio.h>
+#include "kson.h"
+
+/*************
+ *** Parse ***
+ *************/
+
+kson_node_t *kson_parse_core(const char *json, long *_n, int *error, long *parsed_len)
+{
+	long *stack = 0, top = 0, max = 0, n_a = 0, m_a = 0, i, j;
+	kson_node_t *a = 0, *u;
+	const char *p, *q;
+	size_t *tmp;
+
+#define __push_back(y) do { \
+		if (top == max) { \
+			max = max? max<<1 : 4; \
+			stack = (long*)realloc(stack, sizeof(long) * max); \
+		} \
+		stack[top++] = (y); \
+	} while (0)
+
+#define __new_node(z) do { \
+		if (n_a == m_a) { \
+			long old_m = m_a; \
+			m_a = m_a? m_a<<1 : 4; \
+			a = (kson_node_t*)realloc(a, sizeof(kson_node_t) * m_a); \
+			memset(a + old_m, 0, sizeof(kson_node_t) * (m_a - old_m)); \
+		} \
+		*(z) = &a[n_a++]; \
+	} while (0)
+
+	assert(sizeof(size_t) == sizeof(kson_node_t*));
+	*error = KSON_OK;
+	for (p = json; *p; ++p) {
+		while (*p && isspace(*p)) ++p;
+		if (*p == 0) break;
+		if (*p == ',') { // comma is somewhat redundant
+		} else if (*p == '[' || *p == '{') {
+			int t = *p == '['? -1 : -2;
+			if (top < 2 || stack[top-1] != -3) { // unnamed internal node
+				__push_back(n_a);
+				__new_node(&u);
+				__push_back(t);
+			} else stack[top-1] = t; // named internal node
+		} else if (*p == ']' || *p == '}') {
+			long i, start, t = *p == ']'? -1 : -2;
+			for (i = top - 1; i >= 0 && stack[i] != t; --i);
+			if (i < 0) { // error: an extra right bracket
+				*error = KSON_ERR_EXTRA_RIGHT;
+				break;
+			}
+			start = i;
+			u = &a[stack[start-1]];
+			u->key = u->v.str;
+			u->n = top - 1 - start;
+			u->v.child = (kson_node_t**)malloc(u->n * sizeof(kson_node_t*));
+			tmp = (size_t*)u->v.child;
+			for (i = start + 1; i < top; ++i)
+				tmp[i - start - 1] = stack[i];
+			u->type = *p == ']'? KSON_TYPE_BRACKET : KSON_TYPE_BRACE;
+			if ((top = start) == 1) break; // completed one object; remaining characters discarded
+		} else if (*p == ':') {
+			if (top == 0 || stack[top-1] == -3) {
+				*error = KSON_ERR_NO_KEY;
+				break;
+			}
+			__push_back(-3);
+		} else {
+			int c = *p;
+			// get the node to modify
+			if (top >= 2 && stack[top-1] == -3) { // we have a key:value pair here
+				--top;
+				u = &a[stack[top-1]];
+				u->key = u->v.str; // move old value to key
+			} else { // don't know if this is a bare value or a key:value pair; keep it as a value for now
+				__push_back(n_a);
+				__new_node(&u);
+			}
+			// parse string
+			if (c == '\'' || c == '"') {
+				for (q = ++p; *q && *q != c; ++q)
+					if (*q == '\\') ++q;
+			} else {
+				for (q = p; *q && *q != ']' && *q != '}' && *q != ',' && *q != ':' && *q != '\n'; ++q)
+					if (*q == '\\') ++q;
+			}
+			u->v.str = (char*)malloc(q - p + 1); strncpy(u->v.str, p, q - p); u->v.str[q-p] = 0; // equivalent to u->v.str=strndup(p, q-p)
+			u->type = c == '\''? KSON_TYPE_SGL_QUOTE : c == '"'? KSON_TYPE_DBL_QUOTE : KSON_TYPE_NO_QUOTE;
+			p = c == '\'' || c == '"'? q : q - 1;
+		}
+	}
+	while (*p && isspace(*p)) ++p; // skip trailing blanks
+	if (parsed_len) *parsed_len = p - json;
+	if (top != 1) *error = KSON_ERR_EXTRA_LEFT;
+
+	for (i = 0; i < n_a; ++i)
+		for (j = 0, u = &a[i], tmp = (size_t*)u->v.child; j < (long)u->n; ++j)
+			u->v.child[j] = &a[tmp[j]];
+
+	free(stack);
+	*_n = n_a;
+	return a;
+}
+
+void kson_destroy(kson_t *kson)
+{
+	long i;
+	if (kson == 0) return;
+	for (i = 0; i < kson->n_nodes; ++i) {
+		free(kson->root[i].key); free(kson->root[i].v.str);
+	}
+	free(kson->root); free(kson);
+}
+
+kson_t *kson_parse(const char *json)
+{
+	kson_t *kson;
+	int error;
+	kson = (kson_t*)calloc(1, sizeof(kson_t));
+	kson->root = kson_parse_core(json, &kson->n_nodes, &error, 0);
+	if (error) {
+		kson_destroy(kson);
+		return 0;
+	}
+	return kson;
+}
+
+/*************
+ *** Query ***
+ *************/
+
+const kson_node_t *kson_by_path(const kson_node_t *p, int depth, ...)
+{
+	va_list ap;
+	va_start(ap, depth);
+	while (p && depth > 0) {
+		if (p->type == KSON_TYPE_BRACE) {
+			p = kson_by_key(p, va_arg(ap, const char*));
+		} else if (p->type == KSON_TYPE_BRACKET) {
+			p = kson_by_index(p, va_arg(ap, long));
+		} else break;
+		--depth;
+	}
+	va_end(ap);
+	return p;
+}
+
+/**************
+ *** Fromat ***
+ **************/
+
+void kson_format_recur(const kson_node_t *p, int depth)
+{
+	long i;
+	if (p->key) printf("\"%s\":", p->key);
+	if (p->type == KSON_TYPE_BRACKET || p->type == KSON_TYPE_BRACE) {
+		putchar(p->type == KSON_TYPE_BRACKET? '[' : '{');
+		if (p->n) {
+			putchar('\n'); for (i = 0; i <= depth; ++i) fputs("  ", stdout);
+			for (i = 0; i < (long)p->n; ++i) {
+				if (i) {
+					int i;
+					putchar(',');
+					putchar('\n'); for (i = 0; i <= depth; ++i) fputs("  ", stdout);
+				}
+				kson_format_recur(p->v.child[i], depth + 1);
+			}
+			putchar('\n'); for (i = 0; i < depth; ++i) fputs("  ", stdout);
+		}
+		putchar(p->type == KSON_TYPE_BRACKET? ']' : '}');
+	} else {
+		if (p->type != KSON_TYPE_NO_QUOTE)
+			putchar(p->type == KSON_TYPE_SGL_QUOTE? '\'' : '"');
+		fputs(p->v.str, stdout);
+		if (p->type != KSON_TYPE_NO_QUOTE)
+			putchar(p->type == KSON_TYPE_SGL_QUOTE? '\'' : '"');
+	}
+}
+
+void kson_format(const kson_node_t *root)
+{
+	kson_format_recur(root, 0);
+	putchar('\n');
+}
+
+/*********************
+ *** Main function ***
+ *********************/
+
+#ifdef KSON_MAIN
+#define kroundup32(x) (--(x), (x)|=(x)>>1, (x)|=(x)>>2, (x)|=(x)>>4, (x)|=(x)>>8, (x)|=(x)>>16, ++(x))
+int main(int argc, char *argv[])
+{
+	kson_t *kson = 0;
+	if (argc > 1) {
+		FILE *fp;
+		int len = 0, max = 0, tmp, i;
+		char *json = 0, buf[0x10000];
+		if ((fp = fopen(argv[1], "rb")) != 0) {
+			// read the entire file into a string
+			while ((tmp = fread(buf, 1, 0x10000, fp)) != 0) {
+				if (len + tmp + 1 > max) {
+					max = len + tmp + 1;
+					kroundup32(max);
+					json = (char*)realloc(json, max);
+				}
+				memcpy(json + len, buf, tmp);
+				len += tmp;
+			}
+			fclose(fp);
+			// parse
+			kson = kson_parse(json);
+			free(json);
+			if (kson) {
+				kson_format(kson->root);
+				if (argc > 2) {
+					// path finding
+					const kson_node_t *p = kson->root;
+					for (i = 2; i < argc && p; ++i) {
+						if (p->type == KSON_TYPE_BRACKET)
+							p = kson_by_index(p, atoi(argv[i]));
+						else if (p->type == KSON_TYPE_BRACE)
+							p = kson_by_key(p, argv[i]);
+						else p = 0;
+					}
+					if (p) {
+						if (kson_is_internal(p)) printf("Reached an internal node\n");
+						else printf("Value: %s\n", p->v.str);
+					} else printf("Failed to find the slot\n");
+				}
+			} else printf("Failed to parse\n");
+		}
+	} else {
+		kson = kson_parse("{'a' : 1,'b':[0,'isn\\'t',true],'d':[{\n\n\n}]}");
+		if (kson) {
+			const kson_node_t *p = kson_by_path(kson->root, 2, "b", 1);
+			if (p) printf("*** %s\n", p->v.str);
+			else printf("!!! not found\n");
+			kson_format(kson->root);
+		} else {
+			printf("Failed to parse\n");
+		}
+	}
+	kson_destroy(kson);
+	return 0;
+}
+#endif
diff --git a/ext/klib/kson.h b/ext/klib/kson.h
new file mode 100644
index 0000000..a03eb52
--- /dev/null
+++ b/ext/klib/kson.h
@@ -0,0 +1,64 @@
+#ifndef KSON_H
+#define KSON_H
+
+#include <string.h>
+
+#define KSON_TYPE_NO_QUOTE  1
+#define KSON_TYPE_SGL_QUOTE 2
+#define KSON_TYPE_DBL_QUOTE 3
+#define KSON_TYPE_BRACKET   4
+#define KSON_TYPE_BRACE     5
+
+#define KSON_OK              0
+#define KSON_ERR_EXTRA_LEFT  1
+#define KSON_ERR_EXTRA_RIGHT 2
+#define KSON_ERR_NO_KEY      3
+
+typedef struct kson_node_s {
+	unsigned long long type:3, n:61;
+	char *key;
+	union {
+		struct kson_node_s **child;
+		char *str;
+	} v;
+} kson_node_t;
+
+typedef struct {
+	long n_nodes;
+	kson_node_t *root;
+} kson_t;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+	kson_t *kson_parse(const char *json);
+	void kson_destroy(kson_t *kson);
+	const kson_node_t *kson_by_path(const kson_node_t *root, int path_len, ...);
+	void kson_format(const kson_node_t *root);
+
+#ifdef __cplusplus
+}
+#endif
+
+#define kson_is_internal(p) ((p)->type == KSON_TYPE_BRACKET || (p)->type == KSON_TYPE_BRACE)
+
+static inline const kson_node_t *kson_by_key(const kson_node_t *p, const char *key)
+{
+	long i;
+	if (!kson_is_internal(p)) return 0;
+	for (i = 0; i < (long)p->n; ++i) {
+		const kson_node_t *q = p->v.child[i];
+		if (q->key && strcmp(q->key, key) == 0)
+			return q;
+	}
+	return 0;
+}
+
+static inline const kson_node_t *kson_by_index(const kson_node_t *p, long i)
+{
+	if (!kson_is_internal(p)) return 0;
+	return 0 <= i && i < (long)p->n? p->v.child[i] : 0;
+}
+
+#endif
diff --git a/ext/klib/ksort.h b/ext/klib/ksort.h
new file mode 100644
index 0000000..5a2a1a6
--- /dev/null
+++ b/ext/klib/ksort.h
@@ -0,0 +1,353 @@
+/* The MIT License
+
+   Copyright (c) 2008, 2011 Attractive Chaos <attractor@live.co.uk>
+
+   Permission is hereby granted, free of charge, to any person obtaining
+   a copy of this software and associated documentation files (the
+   "Software"), to deal in the Software without restriction, including
+   without limitation the rights to use, copy, modify, merge, publish,
+   distribute, sublicense, and/or sell copies of the Software, and to
+   permit persons to whom the Software is furnished to do so, subject to
+   the following conditions:
+
+   The above copyright notice and this permission notice shall be
+   included in all copies or substantial portions of the Software.
+
+   THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+   EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+   MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+   NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+   BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+   ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+   CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+   SOFTWARE.
+*/
+
+/*
+  2011-04-10 (0.1.6):
+
+  	* Added sample
+
+  2011-03 (0.1.5):
+
+	* Added shuffle/permutation
+
+  2008-11-16 (0.1.4):
+
+    * Fixed a bug in introsort() that happens in rare cases.
+
+  2008-11-05 (0.1.3):
+
+    * Fixed a bug in introsort() for complex comparisons.
+
+	* Fixed a bug in mergesort(). The previous version is not stable.
+
+  2008-09-15 (0.1.2):
+
+	* Accelerated introsort. On my Mac (not on another Linux machine),
+	  my implementation is as fast as std::sort on random input.
+
+	* Added combsort and in introsort, switch to combsort if the
+	  recursion is too deep.
+
+  2008-09-13 (0.1.1):
+
+	* Added k-small algorithm
+
+  2008-09-05 (0.1.0):
+
+	* Initial version
+
+*/
+
+#ifndef AC_KSORT_H
+#define AC_KSORT_H
+
+#include <stdlib.h>
+#include <string.h>
+
+typedef struct {
+	void *left, *right;
+	int depth;
+} ks_isort_stack_t;
+
+#define KSORT_SWAP(type_t, a, b) { register type_t t=(a); (a)=(b); (b)=t; }
+
+#define KSORT_INIT(name, type_t, __sort_lt)								\
+	void ks_mergesort_##name(size_t n, type_t array[], type_t temp[])	\
+	{																	\
+		type_t *a2[2], *a, *b;											\
+		int curr, shift;												\
+																		\
+		a2[0] = array;													\
+		a2[1] = temp? temp : (type_t*)malloc(sizeof(type_t) * n);		\
+		for (curr = 0, shift = 0; (1ul<<shift) < n; ++shift) {			\
+			a = a2[curr]; b = a2[1-curr];								\
+			if (shift == 0) {											\
+				type_t *p = b, *i, *eb = a + n;							\
+				for (i = a; i < eb; i += 2) {							\
+					if (i == eb - 1) *p++ = *i;							\
+					else {												\
+						if (__sort_lt(*(i+1), *i)) {					\
+							*p++ = *(i+1); *p++ = *i;					\
+						} else {										\
+							*p++ = *i; *p++ = *(i+1);					\
+						}												\
+					}													\
+				}														\
+			} else {													\
+				size_t i, step = 1ul<<shift;							\
+				for (i = 0; i < n; i += step<<1) {						\
+					type_t *p, *j, *k, *ea, *eb;						\
+					if (n < i + step) {									\
+						ea = a + n; eb = a;								\
+					} else {											\
+						ea = a + i + step;								\
+						eb = a + (n < i + (step<<1)? n : i + (step<<1)); \
+					}													\
+					j = a + i; k = a + i + step; p = b + i;				\
+					while (j < ea && k < eb) {							\
+						if (__sort_lt(*k, *j)) *p++ = *k++;				\
+						else *p++ = *j++;								\
+					}													\
+					while (j < ea) *p++ = *j++;							\
+					while (k < eb) *p++ = *k++;							\
+				}														\
+			}															\
+			curr = 1 - curr;											\
+		}																\
+		if (curr == 1) {												\
+			type_t *p = a2[0], *i = a2[1], *eb = array + n;				\
+			for (; p < eb; ++i) *p++ = *i;								\
+		}																\
+		if (temp == 0) free(a2[1]);										\
+	}																	\
+	void ks_heapadjust_##name(size_t i, size_t n, type_t l[])			\
+	{																	\
+		size_t k = i;													\
+		type_t tmp = l[i];												\
+		while ((k = (k << 1) + 1) < n) {								\
+			if (k != n - 1 && __sort_lt(l[k], l[k+1])) ++k;				\
+			if (__sort_lt(l[k], tmp)) break;							\
+			l[i] = l[k]; i = k;											\
+		}																\
+		l[i] = tmp;														\
+	}																	\
+	void ks_heapmake_##name(size_t lsize, type_t l[])					\
+	{																	\
+		size_t i;														\
+		for (i = (lsize >> 1) - 1; i != (size_t)(-1); --i)				\
+			ks_heapadjust_##name(i, lsize, l);							\
+	}																	\
+	void ks_heapsort_##name(size_t lsize, type_t l[])					\
+	{																	\
+		size_t i;														\
+		for (i = lsize - 1; i > 0; --i) {								\
+			type_t tmp;													\
+			tmp = *l; *l = l[i]; l[i] = tmp; ks_heapadjust_##name(0, i, l); \
+		}																\
+	}																	\
+	static inline void __ks_insertsort_##name(type_t *s, type_t *t)			\
+	{																	\
+		type_t *i, *j, swap_tmp;										\
+		for (i = s + 1; i < t; ++i)										\
+			for (j = i; j > s && __sort_lt(*j, *(j-1)); --j) {			\
+				swap_tmp = *j; *j = *(j-1); *(j-1) = swap_tmp;			\
+			}															\
+	}																	\
+	void ks_combsort_##name(size_t n, type_t a[])						\
+	{																	\
+		const double shrink_factor = 1.2473309501039786540366528676643; \
+		int do_swap;													\
+		size_t gap = n;													\
+		type_t tmp, *i, *j;												\
+		do {															\
+			if (gap > 2) {												\
+				gap = (size_t)(gap / shrink_factor);					\
+				if (gap == 9 || gap == 10) gap = 11;					\
+			}															\
+			do_swap = 0;												\
+			for (i = a; i < a + n - gap; ++i) {							\
+				j = i + gap;											\
+				if (__sort_lt(*j, *i)) {								\
+					tmp = *i; *i = *j; *j = tmp;						\
+					do_swap = 1;										\
+				}														\
+			}															\
+		} while (do_swap || gap > 2);									\
+		if (gap != 1) __ks_insertsort_##name(a, a + n);					\
+	}																	\
+	void ks_introsort_##name(size_t n, type_t a[])						\
+	{																	\
+		int d;															\
+		ks_isort_stack_t *top, *stack;									\
+		type_t rp, swap_tmp;											\
+		type_t *s, *t, *i, *j, *k;										\
+																		\
+		if (n < 1) return;												\
+		else if (n == 2) {												\
+			if (__sort_lt(a[1], a[0])) { swap_tmp = a[0]; a[0] = a[1]; a[1] = swap_tmp; } \
+			return;														\
+		}																\
+		for (d = 2; 1ul<<d < n; ++d);									\
+		stack = (ks_isort_stack_t*)malloc(sizeof(ks_isort_stack_t) * ((sizeof(size_t)*d)+2)); \
+		top = stack; s = a; t = a + (n-1); d <<= 1;						\
+		while (1) {														\
+			if (s < t) {												\
+				if (--d == 0) {											\
+					ks_combsort_##name(t - s + 1, s);					\
+					t = s;												\
+					continue;											\
+				}														\
+				i = s; j = t; k = i + ((j-i)>>1) + 1;					\
+				if (__sort_lt(*k, *i)) {								\
+					if (__sort_lt(*k, *j)) k = j;						\
+				} else k = __sort_lt(*j, *i)? i : j;					\
+				rp = *k;												\
+				if (k != t) { swap_tmp = *k; *k = *t; *t = swap_tmp; }	\
+				for (;;) {												\
+					do ++i; while (__sort_lt(*i, rp));					\
+					do --j; while (i <= j && __sort_lt(rp, *j));		\
+					if (j <= i) break;									\
+					swap_tmp = *i; *i = *j; *j = swap_tmp;				\
+				}														\
+				swap_tmp = *i; *i = *t; *t = swap_tmp;					\
+				if (i-s > t-i) {										\
+					if (i-s > 16) { top->left = s; top->right = i-1; top->depth = d; ++top; } \
+					s = t-i > 16? i+1 : t;								\
+				} else {												\
+					if (t-i > 16) { top->left = i+1; top->right = t; top->depth = d; ++top; } \
+					t = i-s > 16? i-1 : s;								\
+				}														\
+			} else {													\
+				if (top == stack) {										\
+					free(stack);										\
+					__ks_insertsort_##name(a, a+n);						\
+					return;												\
+				} else { --top; s = (type_t*)top->left; t = (type_t*)top->right; d = top->depth; } \
+			}															\
+		}																\
+	}																	\
+	/* This function is adapted from: http://ndevilla.free.fr/median/ */ \
+	/* 0 <= kk < n */													\
+	type_t ks_ksmall_##name(size_t n, type_t arr[], size_t kk)			\
+	{																	\
+		type_t *low, *high, *k, *ll, *hh, *mid;							\
+		low = arr; high = arr + n - 1; k = arr + kk;					\
+		for (;;) {														\
+			if (high <= low) return *k;									\
+			if (high == low + 1) {										\
+				if (__sort_lt(*high, *low)) KSORT_SWAP(type_t, *low, *high); \
+				return *k;												\
+			}															\
+			mid = low + (high - low) / 2;								\
+			if (__sort_lt(*high, *mid)) KSORT_SWAP(type_t, *mid, *high); \
+			if (__sort_lt(*high, *low)) KSORT_SWAP(type_t, *low, *high); \
+			if (__sort_lt(*low, *mid)) KSORT_SWAP(type_t, *mid, *low);	\
+			KSORT_SWAP(type_t, *mid, *(low+1));							\
+			ll = low + 1; hh = high;									\
+			for (;;) {													\
+				do ++ll; while (__sort_lt(*ll, *low));					\
+				do --hh; while (__sort_lt(*low, *hh));					\
+				if (hh < ll) break;										\
+				KSORT_SWAP(type_t, *ll, *hh);							\
+			}															\
+			KSORT_SWAP(type_t, *low, *hh);								\
+			if (hh <= k) low = ll;										\
+			if (hh >= k) high = hh - 1;									\
+		}																\
+	}																	\
+	void ks_shuffle_##name(size_t n, type_t a[])						\
+	{																	\
+		int i, j;														\
+		for (i = n; i > 1; --i) {										\
+			type_t tmp;													\
+			j = (int)(drand48() * i);									\
+			tmp = a[j]; a[j] = a[i-1]; a[i-1] = tmp;					\
+		}																\
+	}																	\
+	void ks_sample_##name(size_t n, size_t r, type_t a[]) /* FIXME: NOT TESTED!!! */ \
+	{ /* reference: http://code.activestate.com/recipes/272884/ */ \
+		int i, k, pop = n; \
+		for (i = (int)r, k = 0; i >= 0; --i) { \
+			double z = 1., x = drand48(); \
+			type_t tmp; \
+			while (x < z) z -= z * i / (pop--); \
+			if (k != n - pop - 1) tmp = a[k], a[k] = a[n-pop-1], a[n-pop-1] = tmp; \
+			++k; \
+		} \
+	}
+
+#define ks_mergesort(name, n, a, t) ks_mergesort_##name(n, a, t)
+#define ks_introsort(name, n, a) ks_introsort_##name(n, a)
+#define ks_combsort(name, n, a) ks_combsort_##name(n, a)
+#define ks_heapsort(name, n, a) ks_heapsort_##name(n, a)
+#define ks_heapmake(name, n, a) ks_heapmake_##name(n, a)
+#define ks_heapadjust(name, i, n, a) ks_heapadjust_##name(i, n, a)
+#define ks_ksmall(name, n, a, k) ks_ksmall_##name(n, a, k)
+#define ks_shuffle(name, n, a) ks_shuffle_##name(n, a)
+
+#define ks_lt_generic(a, b) ((a) < (b))
+#define ks_lt_str(a, b) (strcmp((a), (b)) < 0)
+
+typedef const char *ksstr_t;
+
+#define KSORT_INIT_GENERIC(type_t) KSORT_INIT(type_t, type_t, ks_lt_generic)
+#define KSORT_INIT_STR KSORT_INIT(str, ksstr_t, ks_lt_str)
+
+#define RS_MIN_SIZE 64
+#define RS_MAX_BITS 8
+
+#define KRADIX_SORT_INIT(name, rstype_t, rskey, sizeof_key) \
+	typedef struct { \
+		rstype_t *b, *e; \
+	} rsbucket_##name##_t; \
+	void rs_insertsort_##name(rstype_t *beg, rstype_t *end) \
+	{ \
+		rstype_t *i; \
+		for (i = beg + 1; i < end; ++i) \
+			if (rskey(*i) < rskey(*(i - 1))) { \
+				rstype_t *j, tmp = *i; \
+				for (j = i; j > beg && rskey(tmp) < rskey(*(j-1)); --j) \
+					*j = *(j - 1); \
+				*j = tmp; \
+			} \
+	} \
+	void rs_sort_##name(rstype_t *beg, rstype_t *end, int n_bits, int s) \
+	{ \
+		rstype_t *i; \
+		int size = 1<<n_bits, m = size - 1; \
+		rsbucket_##name##_t *k, b[1<<RS_MAX_BITS], *be = b + size; \
+		assert(n_bits <= RS_MAX_BITS); \
+		for (k = b; k != be; ++k) k->b = k->e = beg; \
+		for (i = beg; i != end; ++i) ++b[rskey(*i)>>s&m].e; \
+		for (k = b + 1; k != be; ++k) \
+			k->e += (k-1)->e - beg, k->b = (k-1)->e; \
+		for (k = b; k != be;) { \
+			if (k->b != k->e) { \
+				rsbucket_##name##_t *l; \
+				if ((l = b + (rskey(*k->b)>>s&m)) != k) { \
+					rstype_t tmp = *k->b, swap; \
+					do { \
+						swap = tmp; tmp = *l->b; *l->b++ = swap; \
+						l = b + (rskey(tmp)>>s&m); \
+					} while (l != k); \
+					*k->b++ = tmp; \
+				} else ++k->b; \
+			} else ++k; \
+		} \
+		for (b->b = beg, k = b + 1; k != be; ++k) k->b = (k-1)->e; \
+		if (s) { \
+			s = s > n_bits? s - n_bits : 0; \
+			for (k = b; k != be; ++k) \
+				if (k->e - k->b > RS_MIN_SIZE) rs_sort_##name(k->b, k->e, n_bits, s); \
+				else if (k->e - k->b > 1) rs_insertsort_##name(k->b, k->e); \
+		} \
+	} \
+	void radix_sort_##name(rstype_t *beg, rstype_t *end) \
+	{ \
+		if (end - beg <= RS_MIN_SIZE) rs_insertsort_##name(beg, end); \
+		else rs_sort_##name(beg, end, RS_MAX_BITS, (sizeof_key - 1) * RS_MAX_BITS); \
+	}
+
+#endif
diff --git a/ext/klib/kstring.c b/ext/klib/kstring.c
new file mode 100644
index 0000000..aa6c2ee
--- /dev/null
+++ b/ext/klib/kstring.c
@@ -0,0 +1,250 @@
+#include <stdarg.h>
+#include <stdio.h>
+#include <ctype.h>
+#include <string.h>
+#include <stdint.h>
+#include "kstring.h"
+
+int kvsprintf(kstring_t *s, const char *fmt, va_list ap)
+{
+	va_list args;
+	int l;
+	va_copy(args, ap);
+	l = vsnprintf(s->s + s->l, s->m - s->l, fmt, args); // This line does not work with glibc 2.0. See `man snprintf'.
+	va_end(args);
+	if (l + 1 > s->m - s->l) {
+		s->m = s->l + l + 2;
+		kroundup32(s->m);
+		s->s = (char*)realloc(s->s, s->m);
+		va_copy(args, ap);
+		l = vsnprintf(s->s + s->l, s->m - s->l, fmt, args);
+		va_end(args);
+	}
+	s->l += l;
+	return l;
+}
+
+int ksprintf(kstring_t *s, const char *fmt, ...)
+{
+	va_list ap;
+	int l;
+	va_start(ap, fmt);
+	l = kvsprintf(s, fmt, ap);
+	va_end(ap);
+	return l;
+}
+
+char *kstrtok(const char *str, const char *sep_in, ks_tokaux_t *aux)
+{
+	const unsigned char *p, *start, *sep = (unsigned char *) sep_in;
+	if (sep) { // set up the table
+		if (str == 0 && aux->finished) return 0; // no need to set up if we have finished
+		aux->finished = 0;
+		if (sep[0] && sep[1]) {
+			aux->sep = -1;
+			aux->tab[0] = aux->tab[1] = aux->tab[2] = aux->tab[3] = 0;
+			for (p = sep; *p; ++p) aux->tab[*p>>6] |= 1ull<<(*p&0x3f);
+		} else aux->sep = sep[0];
+	}
+	if (aux->finished) return 0;
+	else if (str) start = (unsigned char *) str, aux->finished = 0;
+	else start = (unsigned char *) aux->p + 1;
+	if (aux->sep < 0) {
+		for (p = start; *p; ++p)
+			if (aux->tab[*p>>6]>>(*p&0x3f)&1) break;
+	} else {
+		for (p = start; *p; ++p)
+			if (*p == aux->sep) break;
+	}
+	aux->p = (const char *) p; // end of token
+	if (*p == 0) aux->finished = 1; // no more tokens
+	return (char*)start;
+}
+
+// s MUST BE a null terminated string; l = strlen(s)
+int ksplit_core(char *s, int delimiter, int *_max, int **_offsets)
+{
+	int i, n, max, last_char, last_start, *offsets, l;
+	n = 0; max = *_max; offsets = *_offsets;
+	l = strlen(s);
+	
+#define __ksplit_aux do {						\
+		if (_offsets) {						\
+			s[i] = 0;					\
+			if (n == max) {					\
+				int *tmp;				\
+				max = max? max<<1 : 2;			\
+				if ((tmp = (int*)realloc(offsets, sizeof(int) * max))) {  \
+					offsets = tmp;			\
+				} else	{				\
+					free(offsets);			\
+					*_offsets = NULL;		\
+					return 0;			\
+				}					\
+			}						\
+			offsets[n++] = last_start;			\
+		} else ++n;						\
+	} while (0)
+
+	for (i = 0, last_char = last_start = 0; i <= l; ++i) {
+		if (delimiter == 0) {
+			if (isspace(s[i]) || s[i] == 0) {
+				if (isgraph(last_char)) __ksplit_aux; // the end of a field
+			} else {
+				if (isspace(last_char) || last_char == 0) last_start = i;
+			}
+		} else {
+			if (s[i] == delimiter || s[i] == 0) {
+				if (last_char != 0 && last_char != delimiter) __ksplit_aux; // the end of a field
+			} else {
+				if (last_char == delimiter || last_char == 0) last_start = i;
+			}
+		}
+		last_char = s[i];
+	}
+	*_max = max; *_offsets = offsets;
+	return n;
+}
+
+int kgetline(kstring_t *s, kgets_func *fgets_fn, void *fp)
+{
+	size_t l0 = s->l;
+
+	while (s->l == l0 || s->s[s->l-1] != '\n') {
+		if (s->m - s->l < 200) ks_resize(s, s->m + 200);
+		if (fgets_fn(s->s + s->l, s->m - s->l, fp) == NULL) break;
+		s->l += strlen(s->s + s->l);
+	}
+
+	if (s->l == l0) return EOF;
+
+	if (s->l > l0 && s->s[s->l-1] == '\n') {
+		s->l--;
+		if (s->l > l0 && s->s[s->l-1] == '\r') s->l--;
+	}
+	s->s[s->l] = '\0';
+	return 0;
+}
+
+/**********************
+ * Boyer-Moore search *
+ **********************/
+
+typedef unsigned char ubyte_t;
+
+// reference: http://www-igm.univ-mlv.fr/~lecroq/string/node14.html
+static int *ksBM_prep(const ubyte_t *pat, int m)
+{
+	int i, *suff, *prep, *bmGs, *bmBc;
+	prep = (int*)calloc(m + 256, sizeof(int));
+	bmGs = prep; bmBc = prep + m;
+	{ // preBmBc()
+		for (i = 0; i < 256; ++i) bmBc[i] = m;
+		for (i = 0; i < m - 1; ++i) bmBc[pat[i]] = m - i - 1;
+	}
+	suff = (int*)calloc(m, sizeof(int));
+	{ // suffixes()
+		int f = 0, g;
+		suff[m - 1] = m;
+		g = m - 1;
+		for (i = m - 2; i >= 0; --i) {
+			if (i > g && suff[i + m - 1 - f] < i - g)
+				suff[i] = suff[i + m - 1 - f];
+			else {
+				if (i < g) g = i;
+				f = i;
+				while (g >= 0 && pat[g] == pat[g + m - 1 - f]) --g;
+				suff[i] = f - g;
+			}
+		}
+	}
+	{ // preBmGs()
+		int j = 0;
+		for (i = 0; i < m; ++i) bmGs[i] = m;
+		for (i = m - 1; i >= 0; --i)
+			if (suff[i] == i + 1)
+				for (; j < m - 1 - i; ++j)
+					if (bmGs[j] == m)
+						bmGs[j] = m - 1 - i;
+		for (i = 0; i <= m - 2; ++i)
+			bmGs[m - 1 - suff[i]] = m - 1 - i;
+	}
+	free(suff);
+	return prep;
+}
+
+void *kmemmem(const void *_str, int n, const void *_pat, int m, int **_prep)
+{
+	int i, j, *prep = 0, *bmGs, *bmBc;
+	const ubyte_t *str, *pat;
+	str = (const ubyte_t*)_str; pat = (const ubyte_t*)_pat;
+	prep = (_prep == 0 || *_prep == 0)? ksBM_prep(pat, m) : *_prep;
+	if (_prep && *_prep == 0) *_prep = prep;
+	bmGs = prep; bmBc = prep + m;
+	j = 0;
+	while (j <= n - m) {
+		for (i = m - 1; i >= 0 && pat[i] == str[i+j]; --i);
+		if (i >= 0) {
+			int max = bmBc[str[i+j]] - m + 1 + i;
+			if (max < bmGs[i]) max = bmGs[i];
+			j += max;
+		} else return (void*)(str + j);
+	}
+	if (_prep == 0) free(prep);
+	return 0;
+}
+
+char *kstrstr(const char *str, const char *pat, int **_prep)
+{
+	return (char*)kmemmem(str, strlen(str), pat, strlen(pat), _prep);
+}
+
+char *kstrnstr(const char *str, const char *pat, int n, int **_prep)
+{
+	return (char*)kmemmem(str, n, pat, strlen(pat), _prep);
+}
+
+/***********************
+ * The main() function *
+ ***********************/
+
+#ifdef KSTRING_MAIN
+#include <stdio.h>
+int main()
+{
+	kstring_t *s;
+	int *fields, n, i;
+	ks_tokaux_t aux;
+	char *p;
+	s = (kstring_t*)calloc(1, sizeof(kstring_t));
+	// test ksprintf()
+	ksprintf(s, " abcdefg:    %d ", 100);
+	printf("'%s'\n", s->s);
+	// test ksplit()
+	fields = ksplit(s, 0, &n);
+	for (i = 0; i < n; ++i)
+		printf("field[%d] = '%s'\n", i, s->s + fields[i]);
+	// test kstrtok()
+	s->l = 0;
+	for (p = kstrtok("ab:cde:fg/hij::k", ":/", &aux); p; p = kstrtok(0, 0, &aux)) {
+		kputsn(p, aux.p - p, s);
+		kputc('\n', s);
+	}
+	printf("%s", s->s);
+	// free
+	free(s->s); free(s); free(fields);
+
+	{
+		static char *str = "abcdefgcdgcagtcakcdcd";
+		static char *pat = "cd";
+		char *ret, *s = str;
+		int *prep = 0;
+		while ((ret = kstrstr(s, pat, &prep)) != 0) {
+			printf("match: %s\n", ret);
+			s = ret + prep[0];
+		}
+		free(prep);
+	}
+	return 0;
+}
+#endif
diff --git a/ext/klib/kstring.h b/ext/klib/kstring.h
new file mode 100644
index 0000000..f13fcd9
--- /dev/null
+++ b/ext/klib/kstring.h
@@ -0,0 +1,277 @@
+/* The MIT License
+
+   Copyright (c) by Attractive Chaos <attractor@live.co.uk> 
+
+   Permission is hereby granted, free of charge, to any person obtaining
+   a copy of this software and associated documentation files (the
+   "Software"), to deal in the Software without restriction, including
+   without limitation the rights to use, copy, modify, merge, publish,
+   distribute, sublicense, and/or sell copies of the Software, and to
+   permit persons to whom the Software is furnished to do so, subject to
+   the following conditions:
+
+   The above copyright notice and this permission notice shall be
+   included in all copies or substantial portions of the Software.
+
+   THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+   EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+   MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+   NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+   BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+   ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+   CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+   SOFTWARE.
+*/
+
+#ifndef KSTRING_H
+#define KSTRING_H
+
+#include <stdlib.h>
+#include <string.h>
+#include <stdarg.h>
+#include <stdint.h>
+#include <stdio.h>
+
+#ifndef kroundup32
+#define kroundup32(x) (--(x), (x)|=(x)>>1, (x)|=(x)>>2, (x)|=(x)>>4, (x)|=(x)>>8, (x)|=(x)>>16, ++(x))
+#endif
+
+#if __GNUC__ > 2 || (__GNUC__ == 2 && __GNUC_MINOR__ > 4)
+#define KS_ATTR_PRINTF(fmt, arg) __attribute__((__format__ (__printf__, fmt, arg)))
+#else
+#define KS_ATTR_PRINTF(fmt, arg)
+#endif
+
+
+/* kstring_t is a simple non-opaque type whose fields are likely to be
+ * used directly by user code (but see also ks_str() and ks_len() below).
+ * A kstring_t object is initialised by either of
+ *       kstring_t str = { 0, 0, NULL };
+ *       kstring_t str; ...; str.l = str.m = 0; str.s = NULL;
+ * and either ownership of the underlying buffer should be given away before
+ * the object disappears (see ks_release() below) or the kstring_t should be
+ * destroyed with  free(str.s);  */
+#ifndef KSTRING_T
+#define KSTRING_T kstring_t
+typedef struct __kstring_t {
+	size_t l, m;
+	char *s;
+} kstring_t;
+#endif
+
+typedef struct {
+	uint64_t tab[4];
+	int sep, finished;
+	const char *p; // end of the current token
+} ks_tokaux_t;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+	int kvsprintf(kstring_t *s, const char *fmt, va_list ap) KS_ATTR_PRINTF(2,0);
+	int ksprintf(kstring_t *s, const char *fmt, ...) KS_ATTR_PRINTF(2,3);
+	int ksplit_core(char *s, int delimiter, int *_max, int **_offsets);
+	char *kstrstr(const char *str, const char *pat, int **_prep);
+	char *kstrnstr(const char *str, const char *pat, int n, int **_prep);
+	void *kmemmem(const void *_str, int n, const void *_pat, int m, int **_prep);
+
+	/* kstrtok() is similar to strtok_r() except that str is not
+	 * modified and both str and sep can be NULL. For efficiency, it is
+	 * actually recommended to set both to NULL in the subsequent calls
+	 * if sep is not changed. */
+	char *kstrtok(const char *str, const char *sep, ks_tokaux_t *aux);
+
+	/* kgetline() uses the supplied fgets()-like function to read a "\n"-
+	 * or "\r\n"-terminated line from fp.  The line read is appended to the
+	 * kstring without its terminator and 0 is returned; EOF is returned at
+	 * EOF or on error (determined by querying fp, as per fgets()). */
+	typedef char *kgets_func(char *, int, void *);
+	int kgetline(kstring_t *s, kgets_func *fgets, void *fp);
+
+#ifdef __cplusplus
+}
+#endif
+
+static inline int ks_resize(kstring_t *s, size_t size)
+{
+	if (s->m < size) {
+		char *tmp;
+		s->m = size;
+		kroundup32(s->m);
+		if ((tmp = (char*)realloc(s->s, s->m)))
+			s->s = tmp;
+		else
+			return -1;
+	}
+	return 0;
+}
+
+static inline char *ks_str(kstring_t *s)
+{
+	return s->s;
+}
+
+static inline size_t ks_len(kstring_t *s)
+{
+	return s->l;
+}
+
+// Give ownership of the underlying buffer away to something else (making
+// that something else responsible for freeing it), leaving the kstring_t
+// empty and ready to be used again, or ready to go out of scope without
+// needing  free(str.s)  to prevent a memory leak.
+static inline char *ks_release(kstring_t *s)
+{
+	char *ss = s->s;
+	s->l = s->m = 0;
+	s->s = NULL;
+	return ss;
+}
+
+static inline int kputsn(const char *p, int l, kstring_t *s)
+{
+	if (s->l + l + 1 >= s->m) {
+		char *tmp;
+		s->m = s->l + l + 2;
+		kroundup32(s->m);
+		if ((tmp = (char*)realloc(s->s, s->m)))
+			s->s = tmp;
+		else
+			return EOF;
+	}
+	memcpy(s->s + s->l, p, l);
+	s->l += l;
+	s->s[s->l] = 0;
+	return l;
+}
+
+static inline int kputs(const char *p, kstring_t *s)
+{
+	return kputsn(p, strlen(p), s);
+}
+
+static inline int kputc(int c, kstring_t *s)
+{
+	if (s->l + 1 >= s->m) {
+		char *tmp;
+		s->m = s->l + 2;
+		kroundup32(s->m);
+		if ((tmp = (char*)realloc(s->s, s->m)))
+			s->s = tmp;
+		else
+			return EOF;
+	}
+	s->s[s->l++] = c;
+	s->s[s->l] = 0;
+	return c;
+}
+
+static inline int kputc_(int c, kstring_t *s)
+{
+	if (s->l + 1 > s->m) {
+		char *tmp;
+		s->m = s->l + 1;
+		kroundup32(s->m);
+		if ((tmp = (char*)realloc(s->s, s->m)))
+			s->s = tmp;
+		else
+			return EOF;
+	}
+	s->s[s->l++] = c;
+	return 1;
+}
+
+static inline int kputsn_(const void *p, int l, kstring_t *s)
+{
+	if (s->l + l > s->m) {
+		char *tmp;
+		s->m = s->l + l;
+		kroundup32(s->m);
+		if ((tmp = (char*)realloc(s->s, s->m)))
+			s->s = tmp;
+		else
+			return EOF;
+	}
+	memcpy(s->s + s->l, p, l);
+	s->l += l;
+	return l;
+}
+
+static inline int kputw(int c, kstring_t *s)
+{
+	char buf[16];
+	int i, l = 0;
+	unsigned int x = c;
+	if (c < 0) x = -x;
+	do { buf[l++] = x%10 + '0'; x /= 10; } while (x > 0);
+	if (c < 0) buf[l++] = '-';
+	if (s->l + l + 1 >= s->m) {
+		char *tmp;
+		s->m = s->l + l + 2;
+		kroundup32(s->m);
+		if ((tmp = (char*)realloc(s->s, s->m)))
+			s->s = tmp;
+		else
+			return EOF;
+	}
+	for (i = l - 1; i >= 0; --i) s->s[s->l++] = buf[i];
+	s->s[s->l] = 0;
+	return 0;
+}
+
+static inline int kputuw(unsigned c, kstring_t *s)
+{
+	char buf[16];
+	int l, i;
+	unsigned x;
+	if (c == 0) return kputc('0', s);
+	for (l = 0, x = c; x > 0; x /= 10) buf[l++] = x%10 + '0';
+	if (s->l + l + 1 >= s->m) {
+		char *tmp;
+		s->m = s->l + l + 2;
+		kroundup32(s->m);
+		if ((tmp = (char*)realloc(s->s, s->m)))
+			s->s = tmp;
+		else
+			return EOF;
+	}
+	for (i = l - 1; i >= 0; --i) s->s[s->l++] = buf[i];
+	s->s[s->l] = 0;
+	return 0;
+}
+
+static inline int kputl(long c, kstring_t *s)
+{
+	char buf[32];
+	int i, l = 0;
+	unsigned long x = c;
+	if (c < 0) x = -x;
+	do { buf[l++] = x%10 + '0'; x /= 10; } while (x > 0);
+	if (c < 0) buf[l++] = '-';
+	if (s->l + l + 1 >= s->m) {
+		char *tmp;
+		s->m = s->l + l + 2;
+		kroundup32(s->m);
+		if ((tmp = (char*)realloc(s->s, s->m)))
+			s->s = tmp;
+		else
+			return EOF;
+	}
+	for (i = l - 1; i >= 0; --i) s->s[s->l++] = buf[i];
+	s->s[s->l] = 0;
+	return 0;
+}
+
+/*
+ * Returns 's' split by delimiter, with *n being the number of components;
+ *         NULL on failue.
+ */
+static inline int *ksplit(kstring_t *s, int delimiter, int *n)
+{
+	int max = 0, *offsets = 0;
+	*n = ksplit_core(s->s, delimiter, &max, &offsets);
+	return offsets;
+}
+
+#endif
diff --git a/ext/klib/ksw.c b/ext/klib/ksw.c
new file mode 100644
index 0000000..742fec9
--- /dev/null
+++ b/ext/klib/ksw.c
@@ -0,0 +1,633 @@
+/* The MIT License
+
+   Copyright (c) 2011 by Attractive Chaos <attractor@live.co.uk>
+
+   Permission is hereby granted, free of charge, to any person obtaining
+   a copy of this software and associated documentation files (the
+   "Software"), to deal in the Software without restriction, including
+   without limitation the rights to use, copy, modify, merge, publish,
+   distribute, sublicense, and/or sell copies of the Software, and to
+   permit persons to whom the Software is furnished to do so, subject to
+   the following conditions:
+
+   The above copyright notice and this permission notice shall be
+   included in all copies or substantial portions of the Software.
+
+   THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+   EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+   MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+   NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+   BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+   ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+   CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+   SOFTWARE.
+*/
+
+#include <stdlib.h>
+#include <stdint.h>
+#include <emmintrin.h>
+#include "ksw.h"
+
+#ifdef __GNUC__
+#define LIKELY(x) __builtin_expect((x),1)
+#define UNLIKELY(x) __builtin_expect((x),0)
+#else
+#define LIKELY(x) (x)
+#define UNLIKELY(x) (x)
+#endif
+
+const kswr_t g_defr = { 0, -1, -1, -1, -1, -1, -1 };
+
+struct _kswq_t {
+	int qlen, slen;
+	uint8_t shift, mdiff, max, size;
+	__m128i *qp, *H0, *H1, *E, *Hmax;
+};
+
+/**
+ * Initialize the query data structure
+ *
+ * @param size   Number of bytes used to store a score; valid valures are 1 or 2
+ * @param qlen   Length of the query sequence
+ * @param query  Query sequence
+ * @param m      Size of the alphabet
+ * @param mat    Scoring matrix in a one-dimension array
+ *
+ * @return       Query data structure
+ */
+kswq_t *ksw_qinit(int size, int qlen, const uint8_t *query, int m, const int8_t *mat)
+{
+	kswq_t *q;
+	int slen, a, tmp, p;
+
+	size = size > 1? 2 : 1;
+	p = 8 * (3 - size); // # values per __m128i
+	slen = (qlen + p - 1) / p; // segmented length
+	q = (kswq_t*)malloc(sizeof(kswq_t) + 256 + 16 * slen * (m + 4)); // a single block of memory
+	q->qp = (__m128i*)(((size_t)q + sizeof(kswq_t) + 15) >> 4 << 4); // align memory
+	q->H0 = q->qp + slen * m;
+	q->H1 = q->H0 + slen;
+	q->E  = q->H1 + slen;
+	q->Hmax = q->E + slen;
+	q->slen = slen; q->qlen = qlen; q->size = size;
+	// compute shift
+	tmp = m * m;
+	for (a = 0, q->shift = 127, q->mdiff = 0; a < tmp; ++a) { // find the minimum and maximum score
+		if (mat[a] < (int8_t)q->shift) q->shift = mat[a];
+		if (mat[a] > (int8_t)q->mdiff) q->mdiff = mat[a];
+	}
+	q->max = q->mdiff;
+	q->shift = 256 - q->shift; // NB: q->shift is uint8_t
+	q->mdiff += q->shift; // this is the difference between the min and max scores
+	// An example: p=8, qlen=19, slen=3 and segmentation:
+	//  {{0,3,6,9,12,15,18,-1},{1,4,7,10,13,16,-1,-1},{2,5,8,11,14,17,-1,-1}}
+	if (size == 1) {
+		int8_t *t = (int8_t*)q->qp;
+		for (a = 0; a < m; ++a) {
+			int i, k, nlen = slen * p;
+			const int8_t *ma = mat + a * m;
+			for (i = 0; i < slen; ++i)
+				for (k = i; k < nlen; k += slen) // p iterations
+					*t++ = (k >= qlen? 0 : ma[query[k]]) + q->shift;
+		}
+	} else {
+		int16_t *t = (int16_t*)q->qp;
+		for (a = 0; a < m; ++a) {
+			int i, k, nlen = slen * p;
+			const int8_t *ma = mat + a * m;
+			for (i = 0; i < slen; ++i)
+				for (k = i; k < nlen; k += slen) // p iterations
+					*t++ = (k >= qlen? 0 : ma[query[k]]);
+		}
+	}
+	return q;
+}
+
+kswr_t ksw_u8(kswq_t *q, int tlen, const uint8_t *target, int _gapo, int _gape, int xtra) // the first gap costs -(_o+_e)
+{
+	int slen, i, m_b, n_b, te = -1, gmax = 0, minsc, endsc;
+	uint64_t *b;
+	__m128i zero, gapoe, gape, shift, *H0, *H1, *E, *Hmax;
+	kswr_t r;
+
+#define __max_16(ret, xx) do { \
+		(xx) = _mm_max_epu8((xx), _mm_srli_si128((xx), 8)); \
+		(xx) = _mm_max_epu8((xx), _mm_srli_si128((xx), 4)); \
+		(xx) = _mm_max_epu8((xx), _mm_srli_si128((xx), 2)); \
+		(xx) = _mm_max_epu8((xx), _mm_srli_si128((xx), 1)); \
+    	(ret) = _mm_extract_epi16((xx), 0) & 0x00ff; \
+	} while (0)
+
+	// initialization
+	r = g_defr;
+	minsc = (xtra&KSW_XSUBO)? xtra&0xffff : 0x10000;
+	endsc = (xtra&KSW_XSTOP)? xtra&0xffff : 0x10000;
+	m_b = n_b = 0; b = 0;
+	zero = _mm_set1_epi32(0);
+	gapoe = _mm_set1_epi8(_gapo + _gape);
+	gape = _mm_set1_epi8(_gape);
+	shift = _mm_set1_epi8(q->shift);
+	H0 = q->H0; H1 = q->H1; E = q->E; Hmax = q->Hmax;
+	slen = q->slen;
+	for (i = 0; i < slen; ++i) {
+		_mm_store_si128(E + i, zero);
+		_mm_store_si128(H0 + i, zero);
+		_mm_store_si128(Hmax + i, zero);
+	}
+	// the core loop
+	for (i = 0; i < tlen; ++i) {
+		int j, k, cmp, imax;
+		__m128i e, h, f = zero, max = zero, *S = q->qp + target[i] * slen; // s is the 1st score vector
+		h = _mm_load_si128(H0 + slen - 1); // h={2,5,8,11,14,17,-1,-1} in the above example
+		h = _mm_slli_si128(h, 1); // h=H(i-1,-1); << instead of >> because x64 is little-endian
+		for (j = 0; LIKELY(j < slen); ++j) {
+			/* SW cells are computed in the following order:
+			 *   H(i,j)   = max{H(i-1,j-1)+S(i,j), E(i,j), F(i,j)}
+			 *   E(i+1,j) = max{H(i,j)-q, E(i,j)-r}
+			 *   F(i,j+1) = max{H(i,j)-q, F(i,j)-r}
+			 */
+			// compute H'(i,j); note that at the beginning, h=H'(i-1,j-1)
+			h = _mm_adds_epu8(h, _mm_load_si128(S + j));
+			h = _mm_subs_epu8(h, shift); // h=H'(i-1,j-1)+S(i,j)
+			e = _mm_load_si128(E + j); // e=E'(i,j)
+			h = _mm_max_epu8(h, e);
+			h = _mm_max_epu8(h, f); // h=H'(i,j)
+			max = _mm_max_epu8(max, h); // set max
+			_mm_store_si128(H1 + j, h); // save to H'(i,j)
+			// now compute E'(i+1,j)
+			h = _mm_subs_epu8(h, gapoe); // h=H'(i,j)-gapo
+			e = _mm_subs_epu8(e, gape); // e=E'(i,j)-gape
+			e = _mm_max_epu8(e, h); // e=E'(i+1,j)
+			_mm_store_si128(E + j, e); // save to E'(i+1,j)
+			// now compute F'(i,j+1)
+			f = _mm_subs_epu8(f, gape);
+			f = _mm_max_epu8(f, h);
+			// get H'(i-1,j) and prepare for the next j
+			h = _mm_load_si128(H0 + j); // h=H'(i-1,j)
+		}
+		// NB: we do not need to set E(i,j) as we disallow adjecent insertion and then deletion
+		for (k = 0; LIKELY(k < 16); ++k) { // this block mimics SWPS3; NB: H(i,j) updated in the lazy-F loop cannot exceed max
+			f = _mm_slli_si128(f, 1);
+			for (j = 0; LIKELY(j < slen); ++j) {
+				h = _mm_load_si128(H1 + j);
+				h = _mm_max_epu8(h, f); // h=H'(i,j)
+				_mm_store_si128(H1 + j, h);
+				h = _mm_subs_epu8(h, gapoe);
+				f = _mm_subs_epu8(f, gape);
+				cmp = _mm_movemask_epi8(_mm_cmpeq_epi8(_mm_subs_epu8(f, h), zero));
+				if (UNLIKELY(cmp == 0xffff)) goto end_loop16;
+			}
+		}
+end_loop16:
+		//int k;for (k=0;k<16;++k)printf("%d ", ((uint8_t*)&max)[k]);printf("\n");
+		__max_16(imax, max); // imax is the maximum number in max
+		if (imax >= minsc) { // write the b array; this condition adds branching unfornately
+			if (n_b == 0 || (int32_t)b[n_b-1] + 1 != i) { // then append
+				if (n_b == m_b) {
+					m_b = m_b? m_b<<1 : 8;
+					b = (uint64_t*)realloc(b, 8 * m_b);
+				}
+				b[n_b++] = (uint64_t)imax<<32 | i;
+			} else if ((int)(b[n_b-1]>>32) < imax) b[n_b-1] = (uint64_t)imax<<32 | i; // modify the last
+		}
+		if (imax > gmax) {
+			gmax = imax; te = i; // te is the end position on the target
+			for (j = 0; LIKELY(j < slen); ++j) // keep the H1 vector
+				_mm_store_si128(Hmax + j, _mm_load_si128(H1 + j));
+			if (gmax + q->shift >= 255 || gmax >= endsc) break;
+		}
+		S = H1; H1 = H0; H0 = S; // swap H0 and H1
+	}
+	r.score = gmax + q->shift < 255? gmax : 255;
+	r.te = te;
+	if (r.score != 255) { // get a->qe, the end of query match; find the 2nd best score
+		int max = -1, low, high, qlen = slen * 16;
+		uint8_t *t = (uint8_t*)Hmax;
+		for (i = 0; i < qlen; ++i, ++t)
+			if ((int)*t > max) max = *t, r.qe = i / 16 + i % 16 * slen;
+		//printf("%d,%d\n", max, gmax);
+		if (b) {
+			i = (r.score + q->max - 1) / q->max;
+			low = te - i; high = te + i;
+			for (i = 0; i < n_b; ++i) {
+				int e = (int32_t)b[i];
+				if ((e < low || e > high) && (int)(b[i]>>32) > r.score2)
+					r.score2 = b[i]>>32, r.te2 = e;
+			}
+		}
+	}
+	free(b);
+	return r;
+}
+
+kswr_t ksw_i16(kswq_t *q, int tlen, const uint8_t *target, int _gapo, int _gape, int xtra) // the first gap costs -(_o+_e)
+{
+	int slen, i, m_b, n_b, te = -1, gmax = 0, minsc, endsc;
+	uint64_t *b;
+	__m128i zero, gapoe, gape, *H0, *H1, *E, *Hmax;
+	kswr_t r;
+
+#define __max_8(ret, xx) do { \
+		(xx) = _mm_max_epi16((xx), _mm_srli_si128((xx), 8)); \
+		(xx) = _mm_max_epi16((xx), _mm_srli_si128((xx), 4)); \
+		(xx) = _mm_max_epi16((xx), _mm_srli_si128((xx), 2)); \
+    	(ret) = _mm_extract_epi16((xx), 0); \
+	} while (0)
+
+	// initialization
+	r = g_defr;
+	minsc = (xtra&KSW_XSUBO)? xtra&0xffff : 0x10000;
+	endsc = (xtra&KSW_XSTOP)? xtra&0xffff : 0x10000;
+	m_b = n_b = 0; b = 0;
+	zero = _mm_set1_epi32(0);
+	gapoe = _mm_set1_epi16(_gapo + _gape);
+	gape = _mm_set1_epi16(_gape);
+	H0 = q->H0; H1 = q->H1; E = q->E; Hmax = q->Hmax;
+	slen = q->slen;
+	for (i = 0; i < slen; ++i) {
+		_mm_store_si128(E + i, zero);
+		_mm_store_si128(H0 + i, zero);
+		_mm_store_si128(Hmax + i, zero);
+	}
+	// the core loop
+	for (i = 0; i < tlen; ++i) {
+		int j, k, imax;
+		__m128i e, h, f = zero, max = zero, *S = q->qp + target[i] * slen; // s is the 1st score vector
+		h = _mm_load_si128(H0 + slen - 1); // h={2,5,8,11,14,17,-1,-1} in the above example
+		h = _mm_slli_si128(h, 2);
+		for (j = 0; LIKELY(j < slen); ++j) {
+			h = _mm_adds_epi16(h, *S++);
+			e = _mm_load_si128(E + j);
+			h = _mm_max_epi16(h, e);
+			h = _mm_max_epi16(h, f);
+			max = _mm_max_epi16(max, h);
+			_mm_store_si128(H1 + j, h);
+			h = _mm_subs_epu16(h, gapoe);
+			e = _mm_subs_epu16(e, gape);
+			e = _mm_max_epi16(e, h);
+			_mm_store_si128(E + j, e);
+			f = _mm_subs_epu16(f, gape);
+			f = _mm_max_epi16(f, h);
+			h = _mm_load_si128(H0 + j);
+		}
+		for (k = 0; LIKELY(k < 16); ++k) {
+			f = _mm_slli_si128(f, 2);
+			for (j = 0; LIKELY(j < slen); ++j) {
+				h = _mm_load_si128(H1 + j);
+				h = _mm_max_epi16(h, f);
+				_mm_store_si128(H1 + j, h);
+				h = _mm_subs_epu16(h, gapoe);
+				f = _mm_subs_epu16(f, gape);
+				if(UNLIKELY(!_mm_movemask_epi8(_mm_cmpgt_epi16(f, h)))) goto end_loop8;
+			}
+		}
+end_loop8:
+		__max_8(imax, max);
+		if (imax >= minsc) {
+			if (n_b == 0 || (int32_t)b[n_b-1] + 1 != i) {
+				if (n_b == m_b) {
+					m_b = m_b? m_b<<1 : 8;
+					b = (uint64_t*)realloc(b, 8 * m_b);
+				}
+				b[n_b++] = (uint64_t)imax<<32 | i;
+			} else if ((int)(b[n_b-1]>>32) < imax) b[n_b-1] = (uint64_t)imax<<32 | i; // modify the last
+		}
+		if (imax > gmax) {
+			gmax = imax; te = i;
+			for (j = 0; LIKELY(j < slen); ++j)
+				_mm_store_si128(Hmax + j, _mm_load_si128(H1 + j));
+			if (gmax >= endsc) break;
+		}
+		S = H1; H1 = H0; H0 = S;
+	}
+	r.score = gmax; r.te = te;
+	{
+		int max = -1, low, high, qlen = slen * 8;
+		uint16_t *t = (uint16_t*)Hmax;
+		for (i = 0, r.qe = -1; i < qlen; ++i, ++t)
+			if ((int)*t > max) max = *t, r.qe = i / 8 + i % 8 * slen;
+		if (b) {
+			i = (r.score + q->max - 1) / q->max;
+			low = te - i; high = te + i;
+			for (i = 0; i < n_b; ++i) {
+				int e = (int32_t)b[i];
+				if ((e < low || e > high) && (int)(b[i]>>32) > r.score2)
+					r.score2 = b[i]>>32, r.te2 = e;
+			}
+		}
+	}
+	free(b);
+	return r;
+}
+
+static void revseq(int l, uint8_t *s)
+{
+	int i, t;
+	for (i = 0; i < l>>1; ++i)
+		t = s[i], s[i] = s[l - 1 - i], s[l - 1 - i] = t;
+}
+
+kswr_t ksw_align(int qlen, uint8_t *query, int tlen, uint8_t *target, int m, const int8_t *mat, int gapo, int gape, int xtra, kswq_t **qry)
+{
+	int size;
+	kswq_t *q;
+	kswr_t r, rr;
+	kswr_t (*func)(kswq_t*, int, const uint8_t*, int, int, int);
+
+	q = (qry && *qry)? *qry : ksw_qinit((xtra&KSW_XBYTE)? 1 : 2, qlen, query, m, mat);
+	if (qry && *qry == 0) *qry = q;
+	func = q->size == 2? ksw_i16 : ksw_u8;
+	size = q->size;
+	r = func(q, tlen, target, gapo, gape, xtra);
+	if (qry == 0) free(q);
+	if ((xtra&KSW_XSTART) == 0 || ((xtra&KSW_XSUBO) && r.score < (xtra&0xffff))) return r;
+	revseq(r.qe + 1, query); revseq(r.te + 1, target); // +1 because qe/te points to the exact end, not the position after the end
+	q = ksw_qinit(size, r.qe + 1, query, m, mat);
+	rr = func(q, tlen, target, gapo, gape, KSW_XSTOP | r.score);
+	revseq(r.qe + 1, query); revseq(r.te + 1, target);
+	free(q);
+	if (r.score == rr.score)
+		r.tb = r.te - rr.te, r.qb = r.qe - rr.qe;
+	return r;
+}
+
+/********************
+ *** SW extension ***
+ ********************/
+
+typedef struct {
+	int32_t h, e;
+} eh_t;
+
+int ksw_extend(int qlen, const uint8_t *query, int tlen, const uint8_t *target, int m, const int8_t *mat, int gapo, int gape, int w, int h0, int *_qle, int *_tle)
+{
+	eh_t *eh; // score array
+	int8_t *qp; // query profile
+	int i, j, k, gapoe = gapo + gape, beg, end, max, max_i, max_j, max_gap;
+	if (h0 < 0) h0 = 0;
+	// allocate memory
+	qp = malloc(qlen * m);
+	eh = calloc(qlen + 1, 8);
+	// generate the query profile
+	for (k = i = 0; k < m; ++k) {
+		const int8_t *p = &mat[k * m];
+		for (j = 0; j < qlen; ++j) qp[i++] = p[query[j]];
+	}
+	// fill the first row
+	eh[0].h = h0; eh[1].h = h0 > gapoe? h0 - gapoe : 0;
+	for (j = 2; j <= qlen && eh[j-1].h > gape; ++j)
+		eh[j].h = eh[j-1].h - gape;
+	// adjust $w if it is too large
+	k = m * m;
+	for (i = 0, max = 0; i < k; ++i) // get the max score
+		max = max > mat[i]? max : mat[i];
+	max_gap = (int)((double)(qlen * max - gapo) / gape + 1.);
+	max_gap = max_gap > 1? max_gap : 1;
+	w = w < max_gap? w : max_gap;
+	// DP loop
+	max = h0, max_i = max_j = -1;
+	beg = 0, end = qlen;
+	for (i = 0; LIKELY(i < tlen); ++i) {
+		int f = 0, h1, m = 0, mj = -1;
+		int8_t *q = &qp[target[i] * qlen];
+		// compute the first column
+		h1 = h0 - (gapo + gape * (i + 1));
+		if (h1 < 0) h1 = 0;
+		// apply the band and the constraint (if provided)
+		if (beg < i - w) beg = i - w;
+		if (end > i + w + 1) end = i + w + 1;
+		if (end > qlen) end = qlen;
+		for (j = beg; LIKELY(j < end); ++j) {
+			// At the beginning of the loop: eh[j] = { H(i-1,j-1), E(i,j) }, f = F(i,j) and h1 = H(i,j-1)
+			// Similar to SSE2-SW, cells are computed in the following order:
+			//   H(i,j)   = max{H(i-1,j-1)+S(i,j), E(i,j), F(i,j)}
+			//   E(i+1,j) = max{H(i,j)-gapo, E(i,j)} - gape
+			//   F(i,j+1) = max{H(i,j)-gapo, F(i,j)} - gape
+			eh_t *p = &eh[j];
+			int h = p->h, e = p->e; // get H(i-1,j-1) and E(i-1,j)
+			p->h = h1;          // set H(i,j-1) for the next row
+			h += q[j];
+			h = h > e? h : e;
+			h = h > f? h : f;
+			h1 = h;             // save H(i,j) to h1 for the next column
+			mj = m > h? mj : j;
+			m = m > h? m : h;   // m is stored at eh[mj+1]
+			h -= gapoe;
+			h = h > 0? h : 0;
+			e -= gape;
+			e = e > h? e : h;   // computed E(i+1,j)
+			p->e = e;           // save E(i+1,j) for the next row
+			f -= gape;
+			f = f > h? f : h;   // computed F(i,j+1)
+		}
+		eh[end].h = h1; eh[end].e = 0;
+		if (m == 0) break;
+		if (m > max) max = m, max_i = i, max_j = mj;
+		// update beg and end for the next round
+		for (j = mj; j >= beg && eh[j].h; --j);
+		beg = j + 1;
+		for (j = mj + 2; j <= end && eh[j].h; ++j);
+		end = j;
+		//beg = 0; end = qlen; // uncomment this line for debugging
+	}
+	free(eh); free(qp);
+	if (_qle) *_qle = max_j + 1;
+	if (_tle) *_tle = max_i + 1;
+	return max;
+}
+
+/********************
+ * Global alignment *
+ ********************/
+
+#define MINUS_INF -0x40000000
+
+static inline uint32_t *push_cigar(int *n_cigar, int *m_cigar, uint32_t *cigar, int op, int len)
+{
+	if (*n_cigar == 0 || op != (cigar[(*n_cigar) - 1]&0xf)) {
+		if (*n_cigar == *m_cigar) {
+			*m_cigar = *m_cigar? (*m_cigar)<<1 : 4;
+			cigar = realloc(cigar, (*m_cigar) << 2);
+		}
+		cigar[(*n_cigar)++] = len<<4 | op;
+	} else cigar[(*n_cigar)-1] += len<<4;
+	return cigar;
+}
+
+int ksw_global(int qlen, const uint8_t *query, int tlen, const uint8_t *target, int m, const int8_t *mat, int gapo, int gape, int w, int *n_cigar_, uint32_t **cigar_)
+{
+	eh_t *eh;
+	int8_t *qp; // query profile
+	int i, j, k, gapoe = gapo + gape, score, n_col;
+	uint8_t *z; // backtrack matrix; in each cell: f<<4|e<<2|h; in principle, we can halve the memory, but backtrack will be a little more complex
+	if (n_cigar_) *n_cigar_ = 0;
+	// allocate memory
+	n_col = qlen < 2*w+1? qlen : 2*w+1; // maximum #columns of the backtrack matrix
+	z = malloc(n_col * tlen);
+	qp = malloc(qlen * m);
+	eh = calloc(qlen + 1, 8);
+	// generate the query profile
+	for (k = i = 0; k < m; ++k) {
+		const int8_t *p = &mat[k * m];
+		for (j = 0; j < qlen; ++j) qp[i++] = p[query[j]];
+	}
+	// fill the first row
+	eh[0].h = 0; eh[0].e = MINUS_INF;
+	for (j = 1; j <= qlen && j <= w; ++j)
+		eh[j].h = -(gapo + gape * j), eh[j].e = MINUS_INF;
+	for (; j <= qlen; ++j) eh[j].h = eh[j].e = MINUS_INF; // everything is -inf outside the band
+	// DP loop
+	for (i = 0; LIKELY(i < tlen); ++i) { // target sequence is in the outer loop
+		int32_t f = MINUS_INF, h1, beg, end;
+		int8_t *q = &qp[target[i] * qlen];
+		uint8_t *zi = &z[i * n_col];
+		beg = i > w? i - w : 0;
+		end = i + w + 1 < qlen? i + w + 1 : qlen; // only loop through [beg,end) of the query sequence
+		h1 = beg == 0? -(gapo + gape * (i + 1)) : MINUS_INF;
+		for (j = beg; LIKELY(j < end); ++j) {
+			// This loop is organized in a similar way to ksw_extend() and ksw_sse2(), except:
+			// 1) not checking h>0; 2) recording direction for backtracking
+			eh_t *p = &eh[j];
+			int32_t h = p->h, e = p->e;
+			uint8_t d; // direction
+			p->h = h1;
+			h += q[j];
+			d = h > e? 0 : 1;
+			h = h > e? h : e;
+			d = h > f? d : 2;
+			h = h > f? h : f;
+			h1 = h;
+			h -= gapoe;
+			e -= gape;
+			d |= e > h? 1<<2 : 0;
+			e = e > h? e : h;
+			p->e = e;
+			f -= gape;
+			d |= f > h? 2<<4 : 0; // if we want to halve the memory, use one bit only, instead of two
+			f = f > h? f : h;
+			zi[j - beg] = d; // z[i,j] keeps h for the current cell and e/f for the next cell
+		}
+		eh[end].h = h1; eh[end].e = MINUS_INF;
+	}
+	score = eh[qlen].h;
+	if (n_cigar_ && cigar_) { // backtrack
+		int n_cigar = 0, m_cigar = 0, which = 0;
+		uint32_t *cigar = 0, tmp;
+		i = tlen - 1; k = (i + w + 1 < qlen? i + w + 1 : qlen) - 1; // (i,k) points to the last cell
+		while (i >= 0 && k >= 0) {
+			which = z[i * n_col + (k - (i > w? i - w : 0))] >> (which<<1) & 3;
+			if (which == 0)      cigar = push_cigar(&n_cigar, &m_cigar, cigar, 0, 1), --i, --k;
+			else if (which == 1) cigar = push_cigar(&n_cigar, &m_cigar, cigar, 2, 1), --i;
+			else                 cigar = push_cigar(&n_cigar, &m_cigar, cigar, 1, 1), --k;
+		}
+		if (i >= 0) cigar = push_cigar(&n_cigar, &m_cigar, cigar, 2, i + 1);
+		if (k >= 0) cigar = push_cigar(&n_cigar, &m_cigar, cigar, 1, k + 1);
+		for (i = 0; i < n_cigar>>1; ++i) // reverse CIGAR
+			tmp = cigar[i], cigar[i] = cigar[n_cigar-1-i], cigar[n_cigar-1-i] = tmp;
+		*n_cigar_ = n_cigar, *cigar_ = cigar;
+	}
+	free(eh); free(qp); free(z);
+	return score;
+}
+
+/*******************************************
+ * Main function (not compiled by default) *
+ *******************************************/
+
+#ifdef _KSW_MAIN
+
+#include <unistd.h>
+#include <stdio.h>
+#include <zlib.h>
+#include "kseq.h"
+KSEQ_INIT(gzFile, gzread)
+
+unsigned char seq_nt4_table[256] = {
+	4, 4, 4, 4,  4, 4, 4, 4,  4, 4, 4, 4,  4, 4, 4, 4, 
+	4, 4, 4, 4,  4, 4, 4, 4,  4, 4, 4, 4,  4, 4, 4, 4, 
+	4, 4, 4, 4,  4, 4, 4, 4,  4, 4, 4, 4,  4, 4, 4, 4,
+	4, 4, 4, 4,  4, 4, 4, 4,  4, 4, 4, 4,  4, 4, 4, 4, 
+	4, 0, 4, 1,  4, 4, 4, 2,  4, 4, 4, 4,  4, 4, 4, 4, 
+	4, 4, 4, 4,  3, 4, 4, 4,  4, 4, 4, 4,  4, 4, 4, 4, 
+	4, 0, 4, 1,  4, 4, 4, 2,  4, 4, 4, 4,  4, 4, 4, 4, 
+	4, 4, 4, 4,  3, 4, 4, 4,  4, 4, 4, 4,  4, 4, 4, 4, 
+	4, 4, 4, 4,  4, 4, 4, 4,  4, 4, 4, 4,  4, 4, 4, 4, 
+	4, 4, 4, 4,  4, 4, 4, 4,  4, 4, 4, 4,  4, 4, 4, 4, 
+	4, 4, 4, 4,  4, 4, 4, 4,  4, 4, 4, 4,  4, 4, 4, 4, 
+	4, 4, 4, 4,  4, 4, 4, 4,  4, 4, 4, 4,  4, 4, 4, 4, 
+	4, 4, 4, 4,  4, 4, 4, 4,  4, 4, 4, 4,  4, 4, 4, 4, 
+	4, 4, 4, 4,  4, 4, 4, 4,  4, 4, 4, 4,  4, 4, 4, 4, 
+	4, 4, 4, 4,  4, 4, 4, 4,  4, 4, 4, 4,  4, 4, 4, 4, 
+	4, 4, 4, 4,  4, 4, 4, 4,  4, 4, 4, 4,  4, 4, 4, 4
+};
+
+int main(int argc, char *argv[])
+{
+	int c, sa = 1, sb = 3, i, j, k, forward_only = 0, max_rseq = 0;
+	int8_t mat[25];
+	int gapo = 5, gape = 2, minsc = 0, xtra = KSW_XSTART;
+	uint8_t *rseq = 0;
+	gzFile fpt, fpq;
+	kseq_t *kst, *ksq;
+
+	// parse command line
+	while ((c = getopt(argc, argv, "a:b:q:r:ft:1")) >= 0) {
+		switch (c) {
+			case 'a': sa = atoi(optarg); break;
+			case 'b': sb = atoi(optarg); break;
+			case 'q': gapo = atoi(optarg); break;
+			case 'r': gape = atoi(optarg); break;
+			case 't': minsc = atoi(optarg); break;
+			case 'f': forward_only = 1; break;
+			case '1': xtra |= KSW_XBYTE; break;
+		}
+	}
+	if (optind + 2 > argc) {
+		fprintf(stderr, "Usage: ksw [-1] [-f] [-a%d] [-b%d] [-q%d] [-r%d] [-t%d] <target.fa> <query.fa>\n", sa, sb, gapo, gape, minsc);
+		return 1;
+	}
+	if (minsc > 0xffff) minsc = 0xffff;
+	xtra |= KSW_XSUBO | minsc;
+	// initialize scoring matrix
+	for (i = k = 0; i < 4; ++i) {
+		for (j = 0; j < 4; ++j)
+			mat[k++] = i == j? sa : -sb;
+		mat[k++] = 0; // ambiguous base
+	}
+	for (j = 0; j < 5; ++j) mat[k++] = 0;
+	// open file
+	fpt = gzopen(argv[optind],   "r"); kst = kseq_init(fpt);
+	fpq = gzopen(argv[optind+1], "r"); ksq = kseq_init(fpq);
+	// all-pair alignment
+	while (kseq_read(ksq) > 0) {
+		kswq_t *q[2] = {0, 0};
+		kswr_t r;
+		for (i = 0; i < (int)ksq->seq.l; ++i) ksq->seq.s[i] = seq_nt4_table[(int)ksq->seq.s[i]];
+		if (!forward_only) { // reverse
+			if ((int)ksq->seq.m > max_rseq) {
+				max_rseq = ksq->seq.m;
+				rseq = (uint8_t*)realloc(rseq, max_rseq);
+			}
+			for (i = 0, j = ksq->seq.l - 1; i < (int)ksq->seq.l; ++i, --j)
+				rseq[j] = ksq->seq.s[i] == 4? 4 : 3 - ksq->seq.s[i];
+		}
+		gzrewind(fpt); kseq_rewind(kst);
+		while (kseq_read(kst) > 0) {
+			for (i = 0; i < (int)kst->seq.l; ++i) kst->seq.s[i] = seq_nt4_table[(int)kst->seq.s[i]];
+			r = ksw_align(ksq->seq.l, (uint8_t*)ksq->seq.s, kst->seq.l, (uint8_t*)kst->seq.s, 5, mat, gapo, gape, xtra, &q[0]);
+			if (r.score >= minsc)
+				printf("%s\t%d\t%d\t%s\t%d\t%d\t%d\t%d\t%d\n", kst->name.s, r.tb, r.te+1, ksq->name.s, r.qb, r.qe+1, r.score, r.score2, r.te2);
+			if (rseq) {
+				r = ksw_align(ksq->seq.l, rseq, kst->seq.l, (uint8_t*)kst->seq.s, 5, mat, gapo, gape, xtra, &q[1]);
+				if (r.score >= minsc)
+					printf("%s\t%d\t%d\t%s\t%d\t%d\t%d\t%d\t%d\n", kst->name.s, r.tb, r.te+1, ksq->name.s, (int)ksq->seq.l - r.qb, (int)ksq->seq.l - 1 - r.qe, r.score, r.score2, r.te2);
+			}
+		}
+		free(q[0]); free(q[1]);
+	}
+	free(rseq);
+	kseq_destroy(kst); gzclose(fpt);
+	kseq_destroy(ksq); gzclose(fpq);
+	return 0;
+}
+#endif
diff --git a/ext/klib/ksw.h b/ext/klib/ksw.h
new file mode 100644
index 0000000..5162dc0
--- /dev/null
+++ b/ext/klib/ksw.h
@@ -0,0 +1,72 @@
+#ifndef __AC_KSW_H
+#define __AC_KSW_H
+
+#include <stdint.h>
+
+#define KSW_XBYTE  0x10000
+#define KSW_XSTOP  0x20000
+#define KSW_XSUBO  0x40000
+#define KSW_XSTART 0x80000
+
+struct _kswq_t;
+typedef struct _kswq_t kswq_t;
+
+typedef struct {
+	int score; // best score
+	int te, qe; // target end and query end
+	int score2, te2; // second best score and ending position on the target
+	int tb, qb; // target start and query start
+} kswr_t;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+	/**
+	 * Aligning two sequences
+	 *
+	 * @param qlen    length of the query sequence (typically <tlen)
+	 * @param query   query sequence with 0 <= query[i] < m
+	 * @param tlen    length of the target sequence
+	 * @param target  target sequence
+	 * @param m       number of residue types
+	 * @param mat     m*m scoring matrix in one-dimention array
+	 * @param gapo    gap open penalty; a gap of length l cost "-(gapo+l*gape)"
+	 * @param gape    gap extension penalty
+	 * @param xtra    extra information (see below)
+	 * @param qry     query profile (see below)
+	 *
+	 * @return        alignment information in a struct; unset values to -1
+	 *
+	 * When xtra==0, ksw_align() uses a signed two-byte integer to store a
+	 * score and only finds the best score and the end positions. The 2nd best
+	 * score or the start positions are not attempted. The default behavior can
+	 * be tuned by setting KSW_X* flags:
+	 *
+	 *   KSW_XBYTE:  use an unsigned byte to store a score. If overflow occurs,
+	 *               kswr_t::score will be set to 255
+	 *
+	 *   KSW_XSUBO:  track the 2nd best score and the ending position on the
+	 *               target if the 2nd best is higher than (xtra&0xffff)
+	 *
+	 *   KSW_XSTOP:  stop if the maximum score is above (xtra&0xffff)
+	 *
+	 *   KSW_XSTART: find the start positions
+	 *
+	 * When *qry==NULL, ksw_align() will compute and allocate the query profile
+	 * and when the function returns, *qry will point to the profile, which can
+	 * be deallocated simply by free(). If one query is aligned against multiple
+	 * target sequences, *qry should be set to NULL during the first call and
+	 * freed after the last call. Note that qry can equal 0. In this case, the
+	 * query profile will be deallocated in ksw_align().
+	 */
+	kswr_t ksw_align(int qlen, uint8_t *query, int tlen, uint8_t *target, int m, const int8_t *mat, int gapo, int gape, int xtra, kswq_t **qry);
+
+	int ksw_extend(int qlen, const uint8_t *query, int tlen, const uint8_t *target, int m, const int8_t *mat, int gapo, int gape, int w, int h0, int *_qle, int *_tle);
+	int ksw_global(int qlen, const uint8_t *query, int tlen, const uint8_t *target, int m, const int8_t *mat, int gapo, int gape, int w, int *_n_cigar, uint32_t **_cigar);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif
diff --git a/ext/klib/kthread.c b/ext/klib/kthread.c
new file mode 100644
index 0000000..986f695
--- /dev/null
+++ b/ext/klib/kthread.c
@@ -0,0 +1,256 @@
+#include <pthread.h>
+#include <stdlib.h>
+#include <limits.h>
+
+/************
+ * kt_for() *
+ ************/
+
+struct kt_for_t;
+
+typedef struct {
+	struct kt_for_t *t;
+	long i;
+} ktf_worker_t;
+
+typedef struct kt_for_t {
+	int n_threads;
+	long n;
+	ktf_worker_t *w;
+	void (*func)(void*,long,int);
+	void *data;
+} kt_for_t;
+
+static inline long steal_work(kt_for_t *t)
+{
+	int i, min_i = -1;
+	long k, min = LONG_MAX;
+	for (i = 0; i < t->n_threads; ++i)
+		if (min > t->w[i].i) min = t->w[i].i, min_i = i;
+	k = __sync_fetch_and_add(&t->w[min_i].i, t->n_threads);
+	return k >= t->n? -1 : k;
+}
+
+static void *ktf_worker(void *data)
+{
+	ktf_worker_t *w = (ktf_worker_t*)data;
+	long i;
+	for (;;) {
+		i = __sync_fetch_and_add(&w->i, w->t->n_threads);
+		if (i >= w->t->n) break;
+		w->t->func(w->t->data, i, w - w->t->w);
+	}
+	while ((i = steal_work(w->t)) >= 0)
+		w->t->func(w->t->data, i, w - w->t->w);
+	pthread_exit(0);
+}
+
+void kt_for(int n_threads, void (*func)(void*,long,int), void *data, long n)
+{
+	if (n_threads > 1) {
+		int i;
+		kt_for_t t;
+		pthread_t *tid;
+		t.func = func, t.data = data, t.n_threads = n_threads, t.n = n;
+		t.w = (ktf_worker_t*)alloca(n_threads * sizeof(ktf_worker_t));
+		tid = (pthread_t*)alloca(n_threads * sizeof(pthread_t));
+		for (i = 0; i < n_threads; ++i)
+			t.w[i].t = &t, t.w[i].i = i;
+		for (i = 0; i < n_threads; ++i) pthread_create(&tid[i], 0, ktf_worker, &t.w[i]);
+		for (i = 0; i < n_threads; ++i) pthread_join(tid[i], 0);
+	} else {
+		long j;
+		for (j = 0; j < n; ++j) func(data, j, 0);
+	}
+}
+
+/***************************
+ * kt_for with thread pool *
+ ***************************/
+
+struct kt_forpool_t;
+
+typedef struct {
+	struct kt_forpool_t *t;
+	long i;
+	int action;
+} kto_worker_t;
+
+typedef struct kt_forpool_t {
+	int n_threads, n_pending;
+	long n;
+	pthread_t *tid;
+	kto_worker_t *w;
+	void (*func)(void*,long,int);
+	void *data;
+	pthread_mutex_t mutex;
+	pthread_cond_t cv_m, cv_s;
+} kt_forpool_t;
+
+static inline long kt_fp_steal_work(kt_forpool_t *t)
+{
+	int i, min_i = -1;
+	long k, min = LONG_MAX;
+	for (i = 0; i < t->n_threads; ++i)
+		if (min > t->w[i].i) min = t->w[i].i, min_i = i;
+	k = __sync_fetch_and_add(&t->w[min_i].i, t->n_threads);
+	return k >= t->n? -1 : k;
+}
+
+static void *kt_fp_worker(void *data)
+{
+	kto_worker_t *w = (kto_worker_t*)data;
+	kt_forpool_t *fp = w->t;
+	for (;;) {
+		long i;
+		int action;
+		pthread_mutex_lock(&fp->mutex);
+		if (--fp->n_pending == 0)
+			pthread_cond_signal(&fp->cv_m);
+		w->action = 0;
+		while (w->action == 0) pthread_cond_wait(&fp->cv_s, &fp->mutex);
+		action = w->action;
+		pthread_mutex_unlock(&fp->mutex);
+		if (action < 0) break;
+		for (;;) { // process jobs allocated to this worker
+			i = __sync_fetch_and_add(&w->i, fp->n_threads);
+			if (i >= fp->n) break;
+			fp->func(fp->data, i, w - fp->w);
+		}
+		while ((i = kt_fp_steal_work(fp)) >= 0) // steal jobs allocated to other workers
+			fp->func(fp->data, i, w - fp->w);
+	}
+	pthread_exit(0);
+}
+
+void *kt_forpool_init(int n_threads)
+{
+	kt_forpool_t *fp;
+	int i;
+	fp = (kt_forpool_t*)calloc(1, sizeof(kt_forpool_t));
+	fp->n_threads = fp->n_pending = n_threads;
+	fp->tid = (pthread_t*)calloc(fp->n_threads, sizeof(pthread_t));
+	fp->w = (kto_worker_t*)calloc(fp->n_threads, sizeof(kto_worker_t));
+	for (i = 0; i < fp->n_threads; ++i) fp->w[i].t = fp;
+	pthread_mutex_init(&fp->mutex, 0);
+	pthread_cond_init(&fp->cv_m, 0);
+	pthread_cond_init(&fp->cv_s, 0);
+	for (i = 0; i < fp->n_threads; ++i) pthread_create(&fp->tid[i], 0, kt_fp_worker, &fp->w[i]);
+	pthread_mutex_lock(&fp->mutex);
+	while (fp->n_pending) pthread_cond_wait(&fp->cv_m, &fp->mutex);
+	pthread_mutex_unlock(&fp->mutex);
+	return fp;
+}
+
+void kt_forpool_destroy(void *_fp)
+{
+	kt_forpool_t *fp = (kt_forpool_t*)_fp;
+	int i;
+	for (i = 0; i < fp->n_threads; ++i) fp->w[i].action = -1;
+	pthread_cond_broadcast(&fp->cv_s);
+	for (i = 0; i < fp->n_threads; ++i) pthread_join(fp->tid[i], 0);
+	pthread_cond_destroy(&fp->cv_s);
+	pthread_cond_destroy(&fp->cv_m);
+	pthread_mutex_destroy(&fp->mutex);
+	free(fp->w); free(fp->tid); free(fp);
+}
+
+void kt_forpool(void *_fp, void (*func)(void*,long,int), void *data, long n)
+{
+	kt_forpool_t *fp = (kt_forpool_t*)_fp;
+	long i;
+	if (fp && fp->n_threads > 1) {
+		fp->n = n, fp->func = func, fp->data = data, fp->n_pending = fp->n_threads;
+		for (i = 0; i < fp->n_threads; ++i) fp->w[i].i = i, fp->w[i].action = 1;
+		pthread_mutex_lock(&fp->mutex);
+		pthread_cond_broadcast(&fp->cv_s);
+		while (fp->n_pending) pthread_cond_wait(&fp->cv_m, &fp->mutex);
+		pthread_mutex_unlock(&fp->mutex);
+	} else for (i = 0; i < n; ++i) func(data, i, 0);
+}
+
+/*****************
+ * kt_pipeline() *
+ *****************/
+
+struct ktp_t;
+
+typedef struct {
+	struct ktp_t *pl;
+	int64_t index;
+	int step;
+	void *data;
+} ktp_worker_t;
+
+typedef struct ktp_t {
+	void *shared;
+	void *(*func)(void*, int, void*);
+	int64_t index;
+	int n_workers, n_steps;
+	ktp_worker_t *workers;
+	pthread_mutex_t mutex;
+	pthread_cond_t cv;
+} ktp_t;
+
+static void *ktp_worker(void *data)
+{
+	ktp_worker_t *w = (ktp_worker_t*)data;
+	ktp_t *p = w->pl;
+	while (w->step < p->n_steps) {
+		// test whether we can kick off the job with this worker
+		pthread_mutex_lock(&p->mutex);
+		for (;;) {
+			int i;
+			// test whether another worker is doing the same step
+			for (i = 0; i < p->n_workers; ++i) {
+				if (w == &p->workers[i]) continue; // ignore itself
+				if (p->workers[i].step <= w->step && p->workers[i].index < w->index)
+					break;
+			}
+			if (i == p->n_workers) break; // no workers with smaller indices are doing w->step or the previous steps
+			pthread_cond_wait(&p->cv, &p->mutex);
+		}
+		pthread_mutex_unlock(&p->mutex);
+
+		// working on w->step
+		w->data = p->func(p->shared, w->step, w->step? w->data : 0); // for the first step, input is NULL
+
+		// update step and let other workers know
+		pthread_mutex_lock(&p->mutex);
+		w->step = w->step == p->n_steps - 1 || w->data? (w->step + 1) % p->n_steps : p->n_steps;
+		if (w->step == 0) w->index = p->index++;
+		pthread_cond_broadcast(&p->cv);
+		pthread_mutex_unlock(&p->mutex);
+	}
+	pthread_exit(0);
+}
+
+void kt_pipeline(int n_threads, void *(*func)(void*, int, void*), void *shared_data, int n_steps)
+{
+	ktp_t aux;
+	pthread_t *tid;
+	int i;
+
+	if (n_threads < 1) n_threads = 1;
+	aux.n_workers = n_threads;
+	aux.n_steps = n_steps;
+	aux.func = func;
+	aux.shared = shared_data;
+	aux.index = 0;
+	pthread_mutex_init(&aux.mutex, 0);
+	pthread_cond_init(&aux.cv, 0);
+
+	aux.workers = (ktp_worker_t*)alloca(n_threads * sizeof(ktp_worker_t));
+	for (i = 0; i < n_threads; ++i) {
+		ktp_worker_t *w = &aux.workers[i];
+		w->step = 0; w->pl = &aux; w->data = 0;
+		w->index = aux.index++;
+	}
+
+	tid = (pthread_t*)alloca(n_threads * sizeof(pthread_t));
+	for (i = 0; i < n_threads; ++i) pthread_create(&tid[i], 0, ktp_worker, &aux.workers[i]);
+	for (i = 0; i < n_threads; ++i) pthread_join(tid[i], 0);
+
+	pthread_mutex_destroy(&aux.mutex);
+	pthread_cond_destroy(&aux.cv);
+}
diff --git a/ext/klib/kthread.h b/ext/klib/kthread.h
new file mode 100644
index 0000000..8325c3f
--- /dev/null
+++ b/ext/klib/kthread.h
@@ -0,0 +1,19 @@
+#ifndef KTHREAD_H
+#define KTHREAD_H
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void kt_for(int n_threads, void (*func)(void*,long,int), void *data, long n);
+void kt_pipeline(int n_threads, void *(*func)(void*, int, void*), void *shared_data, int n_steps);
+
+void *kt_forpool_init(int n_threads);
+void kt_forpool_destroy(void *_fp);
+void kt_forpool(void *_fp, void (*func)(void*,long,int), void *data, long n);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif
diff --git a/ext/klib/kurl.c b/ext/klib/kurl.c
new file mode 100644
index 0000000..3bf9290
--- /dev/null
+++ b/ext/klib/kurl.c
@@ -0,0 +1,583 @@
+#include <stdio.h>
+#include <fcntl.h>
+#include <ctype.h>
+#include <assert.h>
+#include <stdint.h>
+#include <stdlib.h>
+#include <unistd.h>
+#include <string.h>
+#include <curl/curl.h>
+#include "kurl.h"
+
+/**********************
+ *** Core kurl APIs ***
+ **********************/
+
+#define KU_DEF_BUFLEN   0x8000
+#define KU_MAX_SKIP     (KU_DEF_BUFLEN<<1) // if seek step is smaller than this, skip
+
+#define kurl_isfile(u) ((u)->fd >= 0)
+
+#ifndef kroundup32
+#define kroundup32(x) (--(x), (x)|=(x)>>1, (x)|=(x)>>2, (x)|=(x)>>4, (x)|=(x)>>8, (x)|=(x)>>16, ++(x))
+#endif
+
+struct kurl_t {
+	CURLM *multi; // cURL multi handler
+	CURL *curl;   // cURL easy handle
+	uint8_t *buf; // buffer
+	off_t off0;   // offset of the first byte in the buffer; the actual file offset equals off0 + p_buf
+	int fd;       // file descriptor for a normal file; <0 for a remote file
+	int m_buf;    // max buffer size; for a remote file, CURL_MAX_WRITE_SIZE*2 is recommended
+	int l_buf;    // length of the buffer; l_buf == 0 iff the input read entirely; l_buf <= m_buf
+	int p_buf;    // file position in the buffer; p_buf <= l_buf
+	int done_reading; // true if we can read nothing from the file; buffer may not be empty even if done_reading is set
+	int err;      // error code
+	struct curl_slist *hdr;
+};
+
+typedef struct {
+	char *url, *date, *auth;
+} s3aux_t;
+
+int kurl_init(void) // required for SSL and win32 socket; NOT thread safe
+{
+	return curl_global_init(CURL_GLOBAL_DEFAULT);
+}
+
+void kurl_destroy(void)
+{
+	curl_global_cleanup();
+}
+
+static int prepare(kurl_t *ku, int do_seek)
+{
+	if (kurl_isfile(ku)) {
+		if (do_seek && lseek(ku->fd, ku->off0, SEEK_SET) != ku->off0)
+			return -1;
+	} else { // FIXME: for S3, we need to re-authorize
+		int rc;
+		rc = curl_multi_remove_handle(ku->multi, ku->curl);
+		rc = curl_easy_setopt(ku->curl, CURLOPT_RESUME_FROM, ku->off0);
+		rc = curl_multi_add_handle(ku->multi, ku->curl);
+	}
+	ku->p_buf = ku->l_buf = 0; // empty the buffer
+	return 0;
+}
+
+static size_t write_cb(char *ptr, size_t size, size_t nmemb, void *data) // callback required by cURL
+{
+	kurl_t *ku = (kurl_t*)data;
+	ssize_t nbytes = size * nmemb;
+	if (nbytes + ku->l_buf > ku->m_buf)
+		return CURL_WRITEFUNC_PAUSE;
+	memcpy(ku->buf + ku->l_buf, ptr, nbytes);
+	ku->l_buf += nbytes;
+	return nbytes;
+}
+
+static int fill_buffer(kurl_t *ku) // fill the buffer
+{
+	assert(ku->p_buf == ku->l_buf); // buffer is always used up when fill_buffer() is called; otherwise a bug
+	ku->off0 += ku->l_buf;
+	ku->p_buf = ku->l_buf = 0;
+	if (ku->done_reading) return 0;
+	if (kurl_isfile(ku)) {
+		// The following block is equivalent to "ku->l_buf = read(ku->fd, ku->buf, ku->m_buf)" on Mac.
+		// On Linux, the man page does not specify whether read() guarantees to read ku->m_buf bytes
+		// even if ->fd references a normal file with sufficient remaining bytes.
+		while (ku->l_buf < ku->m_buf) {
+			int l;
+			l = read(ku->fd, ku->buf + ku->l_buf, ku->m_buf - ku->l_buf);
+			if (l == 0) break;
+			ku->l_buf += l;
+		}
+		if (ku->l_buf < ku->m_buf) ku->done_reading = 1;
+	} else {
+		int n_running, rc;
+		fd_set fdr, fdw, fde;
+		do {
+			int maxfd = -1;
+			long curl_to = -1;
+			struct timeval to;
+			// the following is adaped from docs/examples/fopen.c 
+			to.tv_sec = 10, to.tv_usec = 0; // 10 seconds
+			curl_multi_timeout(ku->multi, &curl_to);
+			if (curl_to >= 0) {
+				to.tv_sec = curl_to / 1000;
+				if (to.tv_sec > 1) to.tv_sec = 1;
+				else to.tv_usec = (curl_to % 1000) * 1000;
+			}
+			FD_ZERO(&fdr); FD_ZERO(&fdw); FD_ZERO(&fde);
+			curl_multi_fdset(ku->multi, &fdr, &fdw, &fde, &maxfd); // FIXME: check return code
+			if (maxfd >= 0 && (rc = select(maxfd+1, &fdr, &fdw, &fde, &to)) < 0) break;
+			if (maxfd < 0) { // check curl_multi_fdset.3 about why we wait for 100ms here
+				struct timespec req, rem;
+				req.tv_sec = 0; req.tv_nsec = 100000000; // this is 100ms
+				nanosleep(&req, &rem);
+			}
+			curl_easy_pause(ku->curl, CURLPAUSE_CONT);
+			rc = curl_multi_perform(ku->multi, &n_running); // FIXME: check return code
+		} while (n_running && ku->l_buf < ku->m_buf - CURL_MAX_WRITE_SIZE);
+		if (ku->l_buf < ku->m_buf - CURL_MAX_WRITE_SIZE) ku->done_reading = 1;
+	}
+	return ku->l_buf;
+}
+
+int kurl_close(kurl_t *ku)
+{
+	if (ku == 0) return 0;
+	if (ku->fd < 0) {
+		curl_multi_remove_handle(ku->multi, ku->curl);
+		curl_easy_cleanup(ku->curl);
+		curl_multi_cleanup(ku->multi);
+		if (ku->hdr) curl_slist_free_all(ku->hdr);
+	} else close(ku->fd);
+	free(ku->buf);
+	free(ku);
+	return 0;
+}
+
+kurl_t *kurl_open(const char *url, kurl_opt_t *opt)
+{
+	extern s3aux_t s3_parse(const char *url, const char *_id, const char *_secret, const char *fn);
+	const char *p, *q;
+	kurl_t *ku;
+	int fd = -1, is_file = 1, failed = 0;
+
+	p = strstr(url, "://");
+	if (p && *p) {
+		for (q = url; q != p; ++q)
+			if (!isalnum(*q)) break;
+		if (q == p) is_file = 0;
+	}
+	if (is_file && (fd = open(url, O_RDONLY)) < 0) return 0;
+
+	ku = (kurl_t*)calloc(1, sizeof(kurl_t));
+	ku->fd = is_file? fd : -1;
+	if (!kurl_isfile(ku)) {
+		ku->multi = curl_multi_init();
+		ku->curl  = curl_easy_init();
+		if (strstr(url, "s3://") == url) {
+			s3aux_t a;
+			a = s3_parse(url, (opt? opt->s3keyid : 0), (opt? opt->s3secretkey : 0), (opt? opt->s3key_fn : 0));
+			if (a.url == 0 || a.date == 0 || a.auth == 0) {
+				kurl_close(ku);
+				return 0;
+			}
+			ku->hdr = curl_slist_append(ku->hdr, a.date);
+			ku->hdr = curl_slist_append(ku->hdr, a.auth);
+			curl_easy_setopt(ku->curl, CURLOPT_URL, a.url);
+			curl_easy_setopt(ku->curl, CURLOPT_HTTPHEADER, ku->hdr);
+			free(a.date); free(a.auth); free(a.url);
+		} else curl_easy_setopt(ku->curl, CURLOPT_URL, url);
+		curl_easy_setopt(ku->curl, CURLOPT_WRITEDATA, ku);
+		curl_easy_setopt(ku->curl, CURLOPT_VERBOSE, 0L);
+		curl_easy_setopt(ku->curl, CURLOPT_NOSIGNAL, 1L);
+		curl_easy_setopt(ku->curl, CURLOPT_WRITEFUNCTION, write_cb);
+		curl_easy_setopt(ku->curl, CURLOPT_SSL_VERIFYPEER, 0L);
+		curl_easy_setopt(ku->curl, CURLOPT_SSL_VERIFYHOST, 0L);
+		curl_easy_setopt(ku->curl, CURLOPT_FOLLOWLOCATION, 1L);
+	}
+	ku->m_buf = KU_DEF_BUFLEN;
+	if (!kurl_isfile(ku) && ku->m_buf < CURL_MAX_WRITE_SIZE * 2)
+		ku->m_buf = CURL_MAX_WRITE_SIZE * 2; // for remote files, the buffer set to 2*CURL_MAX_WRITE_SIZE
+	ku->buf = (uint8_t*)calloc(ku->m_buf, 1);
+	if (kurl_isfile(ku)) failed = (fill_buffer(ku) <= 0);
+	else failed = (prepare(ku, 0) < 0 || fill_buffer(ku) <= 0);
+	if (failed) {
+		kurl_close(ku);
+		return 0;
+	}
+	return ku;
+}
+
+kurl_t *kurl_dopen(int fd)
+{
+	kurl_t *ku;
+	ku = (kurl_t*)calloc(1, sizeof(kurl_t));
+	ku->fd = fd;
+	ku->m_buf = KU_DEF_BUFLEN;
+	ku->buf = (uint8_t*)calloc(ku->m_buf, 1);
+	if (prepare(ku, 0) < 0 || fill_buffer(ku) <= 0) {
+		kurl_close(ku);
+		return 0;
+	}
+	return ku;
+}
+
+int kurl_buflen(kurl_t *ku, int len)
+{
+	if (len <= 0 || len < ku->l_buf) return ku->m_buf;
+	if (!kurl_isfile(ku) && len < CURL_MAX_WRITE_SIZE * 2) return ku->m_buf;
+	ku->m_buf = len;
+	kroundup32(ku->m_buf);
+	ku->buf = (uint8_t*)realloc(ku->buf, ku->m_buf);
+	return ku->m_buf;
+}
+
+ssize_t kurl_read(kurl_t *ku, void *buf, size_t nbytes)
+{
+	ssize_t rest = nbytes;
+	if (ku->l_buf == 0) return 0; // end-of-file
+	while (rest) {
+		if (ku->l_buf - ku->p_buf >= rest) {
+			if (buf) memcpy((uint8_t*)buf + (nbytes - rest), ku->buf + ku->p_buf, rest);
+			ku->p_buf += rest;
+			rest = 0;
+		} else {
+			int ret;
+			if (buf && ku->l_buf > ku->p_buf)
+				memcpy((uint8_t*)buf + (nbytes - rest), ku->buf + ku->p_buf, ku->l_buf - ku->p_buf);
+			rest -= ku->l_buf - ku->p_buf;
+			ku->p_buf = ku->l_buf;
+			ret = fill_buffer(ku);
+			if (ret <= 0) break;
+		}
+	}
+	return nbytes - rest;
+}
+
+off_t kurl_seek(kurl_t *ku, off_t offset, int whence) // FIXME: sometimes when seek() fails, read() will fail as well.
+{
+	off_t new_off = -1, cur_off;
+	int failed = 0, seek_end = 0;
+	if (ku == 0) return -1;
+	cur_off = ku->off0 + ku->p_buf;
+	if (whence == SEEK_SET) new_off = offset;
+	else if (whence == SEEK_CUR) new_off += cur_off + offset;
+	else if (whence == SEEK_END && kurl_isfile(ku)) new_off = lseek(ku->fd, offset, SEEK_END), seek_end = 1;
+	else { // not supported whence
+		ku->err = KURL_INV_WHENCE;
+		return -1;
+	}
+	if (new_off < 0) { // negtive absolute offset
+		ku->err = KURL_SEEK_OUT;
+		return -1;
+	}
+	if (!seek_end && new_off >= cur_off && new_off - cur_off + ku->p_buf < ku->l_buf) {
+		ku->p_buf += new_off - cur_off;
+		return ku->off0 + ku->p_buf;
+	}
+	if (seek_end || new_off < cur_off || new_off - cur_off > KU_MAX_SKIP) { // if jump is large, do actual seek
+		ku->off0 = new_off;
+		ku->done_reading = 0;
+		if (prepare(ku, 1) < 0 || fill_buffer(ku) <= 0) failed = 1;
+	} else { // if jump is small, read through
+		off_t r;
+		r = kurl_read(ku, 0, new_off - cur_off);
+		if (r + cur_off != new_off) failed = 1; // out of range
+	}
+	if (failed) ku->err = KURL_SEEK_OUT, ku->l_buf = ku->p_buf = 0, new_off = -1;
+	return new_off;
+}
+
+off_t kurl_tell(const kurl_t *ku)
+{
+	if (ku == 0) return -1;
+	return ku->off0 + ku->p_buf;
+}
+
+int kurl_eof(const kurl_t *ku)
+{
+	if (ku == 0) return 1;
+	return (ku->l_buf == 0); // unless file end, buffer should never be empty
+}
+
+int kurl_fileno(const kurl_t *ku)
+{
+	if (ku == 0) return -1;
+	return ku->fd;
+}
+
+int kurl_error(const kurl_t *ku)
+{
+	if (ku == 0) return KURL_NULL;
+	return ku->err;
+}
+
+/*****************
+ *** HMAC-SHA1 ***
+ *****************/
+
+/* This code is public-domain - it is based on libcrypt placed in the public domain by Wei Dai and other contributors. */
+
+#define HASH_LENGTH 20
+#define BLOCK_LENGTH 64
+
+typedef struct sha1nfo {
+	union { uint8_t b[BLOCK_LENGTH]; uint32_t w[BLOCK_LENGTH/4]; } buf;
+	uint8_t bufOffset;
+	union { uint8_t b[HASH_LENGTH]; uint32_t w[HASH_LENGTH/4]; } state;
+	uint32_t byteCount;
+	uint8_t keyBuffer[BLOCK_LENGTH];
+	uint8_t innerHash[HASH_LENGTH];
+} sha1nfo;
+
+void sha1_init(sha1nfo *s)
+{
+	const uint8_t table[] = { 0x01,0x23,0x45,0x67, 0x89,0xab,0xcd,0xef, 0xfe,0xdc,0xba,0x98, 0x76,0x54,0x32,0x10, 0xf0,0xe1,0xd2,0xc3 };
+	memcpy(s->state.b, table, HASH_LENGTH);
+	s->byteCount = 0;
+	s->bufOffset = 0;
+}
+
+#define rol32(value, bits) (((value) << (bits)) | ((value) >> (32 - (bits))))
+
+static void sha1_hashBlock(sha1nfo *s)
+{
+	uint32_t i, t, a = s->state.w[0], b = s->state.w[1], c = s->state.w[2], d = s->state.w[3], e = s->state.w[4];
+	for (i = 0; i < 80; i++) {
+		if (i >= 16) {
+			t = s->buf.w[(i+13)&15] ^ s->buf.w[(i+8)&15] ^ s->buf.w[(i+2)&15] ^ s->buf.w[i&15];
+			s->buf.w[i&15] = rol32(t, 1);
+		}
+		if (i < 20)      t = 0x5a827999 + (d ^ (b & (c ^ d)));
+		else if (i < 40) t = 0x6ed9eba1 + (b ^ c ^ d);
+		else if (i < 60) t = 0x8f1bbcdc + ((b & c) | (d & (b | c)));
+		else             t = 0xca62c1d6 + (b ^ c ^ d);
+		t += rol32(a, 5) + e + s->buf.w[i&15];
+		e = d; d = c; c = rol32(b, 30); b = a; a = t;
+	}
+	s->state.w[0] += a; s->state.w[1] += b; s->state.w[2] += c; s->state.w[3] += d; s->state.w[4] += e;
+}
+
+static inline void sha1_add(sha1nfo *s, uint8_t data)
+{
+	s->buf.b[s->bufOffset ^ 3] = data;
+	if (++s->bufOffset == BLOCK_LENGTH) {
+		sha1_hashBlock(s);
+		s->bufOffset = 0;
+	}
+}
+
+void sha1_write1(sha1nfo *s, uint8_t data)
+{
+	++s->byteCount;
+	sha1_add(s, data);
+}
+
+void sha1_write(sha1nfo *s, const char *data, size_t len)
+{
+	while (len--) sha1_write1(s, (uint8_t)*data++);
+}
+
+const uint8_t *sha1_final(sha1nfo *s)
+{
+	int i;
+	sha1_add(s, 0x80);
+	while (s->bufOffset != 56) sha1_add(s, 0);
+	sha1_add(s, 0);
+	sha1_add(s, 0);
+	sha1_add(s, 0);
+	sha1_add(s, s->byteCount >> 29);
+	sha1_add(s, s->byteCount >> 21);
+	sha1_add(s, s->byteCount >> 13);
+	sha1_add(s, s->byteCount >> 5);
+	sha1_add(s, s->byteCount << 3);
+	for (i = 0; i < 5; ++i) {
+		uint32_t a = s->state.w[i];
+		s->state.w[i] = a<<24 | (a<<8&0x00ff0000) | (a>>8&0x0000ff00) | a>>24;
+	}
+	return s->state.b;
+}
+
+#define HMAC_IPAD 0x36
+#define HMAC_OPAD 0x5c
+
+void sha1_init_hmac(sha1nfo *s, const uint8_t* key, int l_key)
+{
+	uint8_t i;
+	memset(s->keyBuffer, 0, BLOCK_LENGTH);
+	if (l_key > BLOCK_LENGTH) {
+		sha1_init(s);
+		while (l_key--) sha1_write1(s, *key++);
+		memcpy(s->keyBuffer, sha1_final(s), HASH_LENGTH);
+	} else memcpy(s->keyBuffer, key, l_key);
+	sha1_init(s);
+	for (i = 0; i < BLOCK_LENGTH; ++i)
+		sha1_write1(s, s->keyBuffer[i] ^ HMAC_IPAD);
+}
+
+const uint8_t *sha1_final_hmac(sha1nfo *s)
+{
+	uint8_t i;
+	memcpy(s->innerHash, sha1_final(s), HASH_LENGTH);
+	sha1_init(s);
+	for (i = 0; i < BLOCK_LENGTH; ++i) sha1_write1(s, s->keyBuffer[i] ^ HMAC_OPAD);
+	for (i = 0; i < HASH_LENGTH; ++i) sha1_write1(s, s->innerHash[i]);
+	return sha1_final(s);
+}
+
+/*******************
+ *** S3 protocol ***
+ *******************/
+
+#include <time.h>
+#include <ctype.h>
+
+static void s3_sign(const char *key, const char *data, char out[29])
+{
+	const char *b64tab = "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789+/";
+	const uint8_t *digest;
+	int i, j, rest;
+	sha1nfo s;
+	sha1_init_hmac(&s, (uint8_t*)key, strlen(key));
+	sha1_write(&s, data, strlen(data));
+	digest = sha1_final_hmac(&s);
+	for (j = i = 0, rest = 8; i < 20; ++j) { // base64 encoding
+		if (rest <= 6) {
+			int next = i < 19? digest[i+1] : 0;
+			out[j] = b64tab[(int)(digest[i] << (6-rest) & 0x3f) | next >> (rest+2)], ++i, rest += 2;
+		} else out[j] = b64tab[(int)digest[i] >> (rest-6) & 0x3f], rest -= 6;
+	}
+	out[j++] = '='; out[j] = 0; // SHA1 digest always has 160 bits, or 20 bytes. We need one '=' at the end.
+}
+
+static char *s3_read_awssecret(const char *fn)
+{
+	char *p, *secret, buf[128], *path;
+	FILE *fp;
+	int l;
+	if (fn == 0) {
+		char *home;
+		home = getenv("HOME");
+		if (home == 0) return 0;
+		l = strlen(home) + 12;
+		path = (char*)malloc(strlen(home) + 12);
+		strcat(strcpy(path, home), "/.awssecret");
+	} else path = (char*)fn;
+	fp = fopen(path, "r");
+	if (path != fn) free(path);
+	if (fp == 0) return 0;
+	l = fread(buf, 1, 127, fp);
+	fclose(fp);
+	buf[l] = 0;
+	for (p = buf; *p != 0 && *p != '\n'; ++p);
+	if (*p == 0) return 0;
+	*p = 0; secret = p + 1;
+	for (++p; *p != 0 && *p != '\n'; ++p);
+	*p = 0;
+	l = p - buf + 1;
+	p = (char*)malloc(l);
+	memcpy(p, buf, l);
+	return p;
+}
+
+typedef struct { int l, m; char *s; } kstring_t;
+
+static inline int kputsn(const char *p, int l, kstring_t *s)
+{
+	if (s->l + l + 1 >= s->m) {
+		s->m = s->l + l + 2;
+		kroundup32(s->m);
+		s->s = (char*)realloc(s->s, s->m);
+	}
+	memcpy(s->s + s->l, p, l);
+	s->l += l;
+	s->s[s->l] = 0;
+	return l;
+}
+
+s3aux_t s3_parse(const char *url, const char *_id, const char *_secret, const char *fn_secret)
+{
+	const char *id, *secret, *bucket, *obj;
+	char *id_secret = 0, date[64], sig[29];
+	time_t t;
+	struct tm tmt;
+	s3aux_t a = {0,0};
+	kstring_t str = {0,0,0};
+	// parse URL
+	if (strstr(url, "s3://") != url) return a;
+	bucket = url + 5;
+	for (obj = bucket; *obj && *obj != '/'; ++obj);
+	if (*obj == 0) return a; // no object
+	// acquire AWS credential and time
+	if (_id == 0 || _secret == 0) {
+		id_secret = s3_read_awssecret(fn_secret);
+		if (id_secret == 0) return a; // fail to read the AWS credential
+		id = id_secret;
+		secret = id_secret + strlen(id) + 1;
+	} else id = _id, secret = _secret;
+	// compose URL for curl
+	kputsn("https://", 8, &str);
+	kputsn(bucket, obj - bucket, &str);
+	kputsn(".s3.amazonaws.com", 17, &str);
+	kputsn(obj, strlen(obj), &str);
+	a.url = str.s;
+	// compose the Date line
+	str.l = str.m = 0; str.s = 0;
+	t = time(0);
+	strftime(date, 64, "%a, %d %b %Y %H:%M:%S +0000", gmtime_r(&t, &tmt));
+	kputsn("Date: ", 6, &str);
+	kputsn(date, strlen(date), &str);
+	a.date = str.s;
+	// compose the string to sign and sign it
+	str.l = str.m = 0; str.s = 0;
+	kputsn("GET\n\n\n", 6, &str);
+	kputsn(date, strlen(date), &str);
+	kputsn("\n", 1, &str);
+	kputsn(bucket-1, strlen(bucket-1), &str);
+	s3_sign(secret, str.s, sig);
+	// compose the Authorization line
+	str.l = 0;
+	kputsn("Authorization: AWS ", 19, &str);
+	kputsn(id, strlen(id), &str);
+	kputsn(":", 1, &str);
+	kputsn(sig, strlen(sig), &str);
+	a.auth = str.s;
+//	printf("curl -H '%s' -H '%s' %s\n", a.date, a.auth, a.url);
+	return a;
+}
+
+/*********************
+ *** Main function ***
+ *********************/
+
+#ifdef KURL_MAIN
+int main(int argc, char *argv[])
+{
+	kurl_t *f;
+	int c, l, l_buf = 0x10000;
+	off_t start = 0, rest = -1;
+	uint8_t *buf;
+	char *p;
+	kurl_opt_t opt;
+
+	memset(&opt, 0, sizeof(kurl_opt_t));
+	while ((c = getopt(argc, argv, "c:l:a:")) >= 0) {
+		if (c == 'c') start = strtol(optarg, &p, 0);
+		else if (c == 'l') rest = strtol(optarg, &p, 0);
+		else if (c == 'a') opt.s3key_fn = optarg;
+	}
+	if (optind == argc) {
+		fprintf(stderr, "Usage: kurl [-c start] [-l length] <url>\n");
+		return 1;
+	}
+	kurl_init();
+	f = kurl_open(argv[optind], &opt);
+	if (f == 0) {
+		fprintf(stderr, "ERROR: fail to open URL\n");
+		return 2;
+	}
+	if (start > 0) {
+		if (kurl_seek(f, start, SEEK_SET) < 0) {
+			kurl_close(f);
+			fprintf(stderr, "ERROR: fail to seek\n");
+			return 3;
+		}
+	}
+	buf = (uint8_t*)calloc(l_buf, 1);
+	while (rest != 0) {
+		int to_read = rest > 0 && rest < l_buf? rest : l_buf;
+		l = kurl_read(f, buf, to_read);
+		if (l == 0) break;
+		fwrite(buf, 1, l, stdout);
+		rest -= l;
+	}
+	free(buf);
+	kurl_close(f);
+	kurl_destroy();
+	return 0;
+}
+#endif
diff --git a/ext/klib/kurl.h b/ext/klib/kurl.h
new file mode 100644
index 0000000..f07f641
--- /dev/null
+++ b/ext/klib/kurl.h
@@ -0,0 +1,57 @@
+#ifndef KURL_H
+#define KURL_H
+
+#include <sys/types.h>
+
+#define KURL_NULL       1
+#define KURL_INV_WHENCE 2
+#define KURL_SEEK_OUT   3
+#define KURL_NO_AUTH    4
+
+struct kurl_t;
+typedef struct kurl_t kurl_t;
+
+typedef struct {
+	const char *s3keyid;
+	const char *s3secretkey;
+	const char *s3key_fn;
+} kurl_opt_t;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+int kurl_init(void);
+void kurl_destroy(void);
+
+kurl_t *kurl_open(const char *url, kurl_opt_t *opt);
+kurl_t *kurl_dopen(int fd);
+int kurl_close(kurl_t *ku);
+ssize_t kurl_read(kurl_t *ku, void *buf, size_t nbytes);
+off_t kurl_seek(kurl_t *ku, off_t offset, int whence);
+int kurl_buflen(kurl_t *ku, int len);
+
+off_t kurl_tell(const kurl_t *ku);
+int kurl_eof(const kurl_t *ku);
+int kurl_fileno(const kurl_t *ku);
+int kurl_error(const kurl_t *ku);
+
+#ifdef __cplusplus
+}
+#endif
+
+#ifndef KNETFILE_H
+#define KNETFILE_H
+typedef kurl_t knetFile;
+#define knet_open(fn, mode) kurl_open(fn, 0)
+#define knet_dopen(fd, mode) kurl_dopen(fd)
+#define knet_close(fp) kurl_close(fp)
+#define knet_read(fp, buf, len) kurl_read(fp, buf, len)
+#define knet_seek(fp, off, whence) kurl_seek(fp, off, whence)
+#define knet_tell(fp) kurl_tell(fp)
+#define knet_fileno(fp) kurl_fileno(fp)
+#define knet_win32_init() kurl_init()
+#define knet_win32_destroy() kurl_destroy()
+#endif
+
+#endif
diff --git a/ext/klib/kvec.h b/ext/klib/kvec.h
new file mode 100644
index 0000000..676be8b
--- /dev/null
+++ b/ext/klib/kvec.h
@@ -0,0 +1,90 @@
+/* The MIT License
+
+   Copyright (c) 2008, by Attractive Chaos <attractor@live.co.uk>
+
+   Permission is hereby granted, free of charge, to any person obtaining
+   a copy of this software and associated documentation files (the
+   "Software"), to deal in the Software without restriction, including
+   without limitation the rights to use, copy, modify, merge, publish,
+   distribute, sublicense, and/or sell copies of the Software, and to
+   permit persons to whom the Software is furnished to do so, subject to
+   the following conditions:
+
+   The above copyright notice and this permission notice shall be
+   included in all copies or substantial portions of the Software.
+
+   THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+   EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+   MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+   NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+   BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+   ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+   CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+   SOFTWARE.
+*/
+
+/*
+  An example:
+
+#include "kvec.h"
+int main() {
+	kvec_t(int) array;
+	kv_init(array);
+	kv_push(int, array, 10); // append
+	kv_a(int, array, 20) = 5; // dynamic
+	kv_A(array, 20) = 4; // static
+	kv_destroy(array);
+	return 0;
+}
+*/
+
+/*
+  2008-09-22 (0.1.0):
+
+	* The initial version.
+
+*/
+
+#ifndef AC_KVEC_H
+#define AC_KVEC_H
+
+#include <stdlib.h>
+
+#define kv_roundup32(x) (--(x), (x)|=(x)>>1, (x)|=(x)>>2, (x)|=(x)>>4, (x)|=(x)>>8, (x)|=(x)>>16, ++(x))
+
+#define kvec_t(type) struct { size_t n, m; type *a; }
+#define kv_init(v) ((v).n = (v).m = 0, (v).a = 0)
+#define kv_destroy(v) free((v).a)
+#define kv_A(v, i) ((v).a[(i)])
+#define kv_pop(v) ((v).a[--(v).n])
+#define kv_size(v) ((v).n)
+#define kv_max(v) ((v).m)
+
+#define kv_resize(type, v, s)  ((v).m = (s), (v).a = (type*)realloc((v).a, sizeof(type) * (v).m))
+
+#define kv_copy(type, v1, v0) do {							\
+		if ((v1).m < (v0).n) kv_resize(type, v1, (v0).n);	\
+		(v1).n = (v0).n;									\
+		memcpy((v1).a, (v0).a, sizeof(type) * (v0).n);		\
+	} while (0)												\
+
+#define kv_push(type, v, x) do {									\
+		if ((v).n == (v).m) {										\
+			(v).m = (v).m? (v).m<<1 : 2;							\
+			(v).a = (type*)realloc((v).a, sizeof(type) * (v).m);	\
+		}															\
+		(v).a[(v).n++] = (x);										\
+	} while (0)
+
+#define kv_pushp(type, v) (((v).n == (v).m)?							\
+						   ((v).m = ((v).m? (v).m<<1 : 2),				\
+							(v).a = (type*)realloc((v).a, sizeof(type) * (v).m), 0)	\
+						   : 0), ((v).a + ((v).n++))
+
+#define kv_a(type, v, i) (((v).m <= (size_t)(i)? \
+						  ((v).m = (v).n = (i) + 1, kv_roundup32((v).m), \
+						   (v).a = (type*)realloc((v).a, sizeof(type) * (v).m), 0) \
+						  : (v).n <= (size_t)(i)? (v).n = (i) + 1 \
+						  : 0), (v).a[(i)])
+
+#endif
diff --git a/ext/klib/lua/bio.lua b/ext/klib/lua/bio.lua
new file mode 100644
index 0000000..c9f2200
--- /dev/null
+++ b/ext/klib/lua/bio.lua
@@ -0,0 +1,149 @@
+-- bioinformatics routines
+
+-- Description: read a fasta/fastq file
+local function readseq(fp)
+	local finished, last = false, nil;
+	return function()
+		local match;
+		if finished then return nil end
+		if (last == nil) then -- the first record or a record following a fastq
+			for l in fp:lines() do
+				if l:byte(1) == 62 or l:byte(1) == 64 then -- ">" || "@"
+					last = l;
+					break;
+				end
+			end
+			if last == nil then
+				finished = true;
+				return nil;
+			end
+		end
+		local tmp = last:find("%s");
+		name = (tmp and last:sub(2, tmp-1)) or last:sub(2); -- sequence name
+		local seqs = {};
+		local c; -- the first character of the last line
+		last = nil;
+		for l in fp:lines() do -- read sequence
+			c = l:byte(1);
+			if c == 62 or c == 64 or c == 43 then
+				last = l;
+				break;
+			end
+			table.insert(seqs, l);
+		end
+		if last == nil then finished = true end -- end of file
+		if c ~= 43 then return name, table.concat(seqs) end -- a fasta record
+		local seq, len = table.concat(seqs), 0; -- prepare to parse quality
+		seqs = {};
+		for l in fp:lines() do -- read quality
+			table.insert(seqs, l);
+			len = len + #l;
+			if len >= #seq then
+				last = nil;
+				return name, seq, table.concat(seqs);
+			end
+		end
+		finished = true;
+		return name, seq;
+	end
+end
+
+-- extract subsequence from a fasta file indexe by samtools faidx
+local function faidxsub(fn)
+	local fpidx = io.open(fn .. ".fai");
+	if fpidx == nil then
+		io.stderr:write("[faidxsub] fail to open the FASTA index file.\n");
+		return nil
+	end
+	local idx = {};
+	for l in fpidx:lines() do
+		local name, len, offset, line_blen, line_len = l:match("(%S+)%s(%d+)%s(%d+)%s(%d+)%s(%d+)");
+		if name then
+			idx[name] = {tonumber(len), offset, line_blen, line_len};
+		end
+	end
+	fpidx:close();
+	local fp = io.open(fn);
+	return function(name, beg_, end_) -- 0-based coordinate
+		if name == nil then fp:close(); return nil; end
+		if idx[name] then
+			local a = idx[name];
+			beg_ = beg_ or 0;
+			end_ = end_ or a[1];
+			end_ = (end_ <= a[1] and end_) or a[1];
+			local fb, fe = math.floor(beg_ / a[3]), math.floor(end_ / a[3]);
+			local qb, qe = beg_ - fb * a[3], end_ - fe * a[3];
+			fp:seek("set", a[2] + fb * a[4] + qb);
+			local s = fp:read((fe - fb) * a[4] + (qe - qb)):gsub("%s", "");
+			return s;
+		end
+	end
+end
+
+--Description: Index a list of intervals and test if a given interval overlaps with the list
+--Example: lua -lbio -e 'a={{100,201},{200,300},{400,600}};f=bio.intvovlp(a);print(f(600,700))'
+--[[
+  By default, we keep for each tiling 8192 window the interval overlaping the
+  window while having the smallest start position. This method may not work
+  well when most intervals are small but few intervals span a long distance.
+]]--
+local function intvovlp(intv, bits)
+	bits = bits or 13 -- the default bin size is 8192 = 1<<13
+	table.sort(intv, function(a,b) return a[1] < b[1] end) -- sort by the start
+	-- merge intervals; the step speeds up testing, but can be skipped
+	local b, e, k = -1, -1, 1
+	for i = 1, #intv do
+		if e < intv[i][1] then
+			if e >= 0 then intv[k], k = {b, e}, k + 1 end
+			b, e = intv[i][1], intv[i][2]
+		else e = intv[i][2] end
+	end
+	if e >= 0 then intv[k] = {b, e} end
+	while #a > k do table.remove(a) end -- truncate the interval list
+	-- build the index for the list of intervals
+	local idx, size, max = {}, math.pow(2, bits), 0
+	for i = 1, #a do
+		b = math.modf(intv[i][1] / size)
+		e = math.modf(intv[i][2] / size)
+		if b == e then idx[b] = idx[b] or i
+		else for j = b, e do idx[j] = idx[j] or i end end
+		max = (max > e and max) or e
+	end
+	-- return a function (closure)
+	return function(_beg, _end)
+		local x = math.modf(_beg / size)
+		if x > max then return false end
+		local off = idx[x]; -- the start bin
+		if off == nil then -- the following is not the best in efficiency
+			for i = x - 1, 0, -1 do -- find the minimum bin with a value
+				if idx[i] ~= nil then off = idx[i]; break; end
+			end
+			if off == nil then return false end
+		end
+		for i = off, #intv do -- start from off and search for overlaps
+			if intv[i][1] >= _end then return false
+			elseif intv[i][2] > _beg then return true end 
+		end
+		return false
+	end
+end
+
+bio = {
+	readseq = readseq,
+	faidxsub = faidxsub,
+	intvovlp = intvovlp
+}
+
+bio.nt16 = {
+[0]=15,15,15,15, 15,15,15,15, 15,15,15,15, 15,15,15,15, 15,15,15,15, 15,15,15,15, 15,15,15,15, 15,15,15,15,
+	15,15,15,15, 15,15,15,15, 15,15,15,15, 15,15,15,15, 15,15,15,15, 15,15,15,15, 15,15,15,15, 15,15,15,15,
+	15, 1,14, 2, 13,15,15, 4, 11,15,15,12, 15, 3,15,15, 15,15, 5, 6,  8,15, 7, 9,  0,10,15,15, 15,15,15,15,
+	15, 1,14, 2, 13,15,15, 4, 11,15,15,12, 15, 3,15,15, 15,15, 5, 6,  8,15, 7, 9,  0,10,15,15, 15,15,15,15,
+	15,15,15,15, 15,15,15,15, 15,15,15,15, 15,15,15,15, 15,15,15,15, 15,15,15,15, 15,15,15,15, 15,15,15,15,
+	15,15,15,15, 15,15,15,15, 15,15,15,15, 15,15,15,15, 15,15,15,15, 15,15,15,15, 15,15,15,15, 15,15,15,15,
+	15,15,15,15, 15,15,15,15, 15,15,15,15, 15,15,15,15, 15,15,15,15, 15,15,15,15, 15,15,15,15, 15,15,15,15,
+	15,15,15,15, 15,15,15,15, 15,15,15,15, 15,15,15,15, 15,15,15,15, 15,15,15,15, 15,15,15,15, 15,15,15,15
+}
+bio.ntcnt  = { [0]=4, 1, 1,  2, 1,  2, 2,  3, 1, 2, 2,  3, 2,  3, 3,  4 }
+bio.ntcomp = { [0]=0, 8, 4, 12, 2, 10, 9, 14, 1, 6, 5, 13, 3, 11, 7, 15 }
+bio.ntrev  = 'XACMGRSVTWYHKDBN'
diff --git a/ext/klib/lua/klib.lua b/ext/klib/lua/klib.lua
new file mode 100644
index 0000000..bfe52f7
--- /dev/null
+++ b/ext/klib/lua/klib.lua
@@ -0,0 +1,677 @@
+--[[
+  The MIT License
+  
+  Copyright (c) 2011, Attractive Chaos <attractor@live.co.uk>
+  
+  Permission is hereby granted, free of charge, to any person obtaining a copy
+  of this software and associated documentation files (the "Software"), to deal
+  in the Software without restriction, including without limitation the rights
+  to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+  copies of the Software, and to permit persons to whom the Software is
+  furnished to do so, subject to the following conditions:
+  
+  The above copyright notice and this permission notice shall be included in
+  all copies or substantial portions of the Software.
+  
+  THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+  IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+  FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+  AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+  LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+  OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+  SOFTWARE.
+]]--
+
+--[[
+  This is a Lua library, more exactly a collection of Lua snippets, covering
+  utilities (e.g. getopt), string operations (e.g. split), statistics (e.g.
+  Fisher's exact test), special functions (e.g. logarithm gamma) and matrix
+  operations (e.g. Gauss-Jordan elimination). The routines are designed to be
+  as independent as possible, such that one can copy-paste relevant pieces of
+  code without worrying about additional library dependencies.
+
+  If you use routines from this library, please include the licensing
+  information above where appropriate.
+]]--
+
+--[[
+  Library functions and dependencies. "a>b" means "a is required by b"; "b<a"
+  means "b depends on a".
+
+  os.getopt()
+  string:split()
+  io.xopen()
+  table.ksmall()
+  table.shuffle()
+  math.lgamma() >math.lbinom() >math.igamma()
+  math.igamma() <math.lgamma() >matrix.chi2()
+  math.erfc()
+  math.lbinom() <math.lgamma() >math.fisher_exact()
+  math.bernstein_poly() <math.lbinom()
+  math.fisher_exact() <math.lbinom()
+  math.jackknife()
+  math.pearson()
+  math.spearman()
+  math.fmin()
+  matrix
+  matrix.add()
+  matrix.T() >matrix.mul()
+  matrix.mul() <matrix.T()
+  matrix.tostring()
+  matrix.chi2() <math.igamma()
+  matrix.solve()
+]]--
+
+-- Description: getopt() translated from the BSD getopt(); compatible with the default Unix getopt()
+--[[ Example:
+	for o, a in os.getopt(arg, 'a:b') do
+		print(o, a)
+	end
+]]--
+function os.getopt(args, ostr)
+	local arg, place = nil, 0;
+	return function ()
+		if place == 0 then -- update scanning pointer
+			place = 1
+			if #args == 0 or args[1]:sub(1, 1) ~= '-' then place = 0; return nil end
+			if #args[1] >= 2 then
+				place = place + 1
+				if args[1]:sub(2, 2) == '-' then -- found "--"
+					place = 0
+					table.remove(args, 1);
+					return nil;
+				end
+			end
+		end
+		local optopt = args[1]:sub(place, place);
+		place = place + 1;
+		local oli = ostr:find(optopt);
+		if optopt == ':' or oli == nil then -- unknown option
+			if optopt == '-' then return nil end
+			if place > #args[1] then
+				table.remove(args, 1);
+				place = 0;
+			end
+			return '?';
+		end
+		oli = oli + 1;
+		if ostr:sub(oli, oli) ~= ':' then -- do not need argument
+			arg = nil;
+			if place > #args[1] then
+				table.remove(args, 1);
+				place = 0;
+			end
+		else -- need an argument
+			if place <= #args[1] then  -- no white space
+				arg = args[1]:sub(place);
+			else
+				table.remove(args, 1);
+				if #args == 0 then -- an option requiring argument is the last one
+					place = 0;
+					if ostr:sub(1, 1) == ':' then return ':' end
+					return '?';
+				else arg = args[1] end
+			end
+			table.remove(args, 1);
+			place = 0;
+		end
+		return optopt, arg;
+	end
+end
+
+-- Description: string split
+function string:split(sep, n)
+	local a, start = {}, 1;
+	sep = sep or "%s+";
+	repeat
+		local b, e = self:find(sep, start);
+		if b == nil then
+			table.insert(a, self:sub(start));
+			break
+		end
+		a[#a+1] = self:sub(start, b - 1);
+		start = e + 1;
+		if n and #a == n then
+			table.insert(a, self:sub(start));
+			break
+		end
+	until start > #self;
+	return a;
+end
+
+-- Description: smart file open
+function io.xopen(fn, mode)
+	mode = mode or 'r';
+	if fn == nil then return io.stdin;
+	elseif fn == '-' then return (mode == 'r' and io.stdin) or io.stdout;
+	elseif fn:sub(-3) == '.gz' then return (mode == 'r' and io.popen('gzip -dc ' .. fn, 'r')) or io.popen('gzip > ' .. fn, 'w');
+	elseif fn:sub(-4) == '.bz2' then return (mode == 'r' and io.popen('bzip2 -dc ' .. fn, 'r')) or io.popen('bgzip2 > ' .. fn, 'w');
+	else return io.open(fn, mode) end
+end
+
+-- Description: find the k-th smallest element in an array (Ref. http://ndevilla.free.fr/median/)
+function table.ksmall(arr, k)
+	local low, high = 1, #arr;
+	while true do
+		if high <= low then return arr[k] end
+		if high == low + 1 then
+			if arr[high] < arr[low] then arr[high], arr[low] = arr[low], arr[high] end;
+			return arr[k];
+		end
+		local mid = math.floor((high + low) / 2);
+		if arr[high] < arr[mid] then arr[mid], arr[high] = arr[high], arr[mid] end
+		if arr[high] < arr[low] then arr[low], arr[high] = arr[high], arr[low] end
+		if arr[low]  < arr[mid] then arr[low], arr[mid]  = arr[mid],  arr[low] end
+		arr[mid], arr[low+1] = arr[low+1], arr[mid];
+		local ll, hh = low + 1, high;
+		while true do
+			repeat ll = ll + 1 until arr[ll] >= arr[low]
+			repeat hh = hh - 1 until arr[low] >= arr[hh]
+			if hh < ll then break end
+			arr[ll], arr[hh] = arr[hh], arr[ll];
+		end
+		arr[low], arr[hh] = arr[hh], arr[low];
+		if hh <= k then low = ll end
+		if hh >= k then high = hh - 1 end
+	end
+end
+
+-- Description: shuffle/permutate an array
+function table.shuffle(a)
+	for i = #a, 1, -1 do
+		local j = math.random(i)
+		a[j], a[i] = a[i], a[j]
+	end
+end
+
+--
+-- Mathematics
+--
+
+-- Description: log gamma function
+-- Required by: math.lbinom()
+-- Reference: AS245, 2nd algorithm, http://lib.stat.cmu.edu/apstat/245
+function math.lgamma(z)
+	local x;
+	x = 0.1659470187408462e-06     / (z+7);
+	x = x + 0.9934937113930748e-05 / (z+6);
+	x = x - 0.1385710331296526     / (z+5);
+	x = x + 12.50734324009056      / (z+4);
+	x = x - 176.6150291498386      / (z+3);
+	x = x + 771.3234287757674      / (z+2);
+	x = x - 1259.139216722289      / (z+1);
+	x = x + 676.5203681218835      / z;
+	x = x + 0.9999999999995183;
+	return math.log(x) - 5.58106146679532777 - z + (z-0.5) * math.log(z+6.5);
+end
+
+-- Description: regularized incomplete gamma function
+-- Dependent on: math.lgamma()
+--[[
+  Formulas are taken from Wiki, with additional input from Numerical
+  Recipes in C (for modified Lentz's algorithm) and AS245
+  (http://lib.stat.cmu.edu/apstat/245).
+ 
+  A good online calculator is available at:
+ 
+    http://www.danielsoper.com/statcalc/calc23.aspx
+ 
+  It calculates upper incomplete gamma function, which equals
+  math.igamma(s,z,true)*math.exp(math.lgamma(s))
+]]--
+function math.igamma(s, z, complement)
+
+	local function _kf_gammap(s, z)
+		local sum, x = 1, 1;
+		for k = 1, 100 do
+			x = x * z / (s + k);
+			sum = sum + x;
+			if x / sum < 1e-14 then break end
+		end
+		return math.exp(s * math.log(z) - z - math.lgamma(s + 1.) + math.log(sum));
+	end
+
+	local function _kf_gammaq(s, z)
+		local C, D, f, TINY;
+		f = 1. + z - s; C = f; D = 0.; TINY = 1e-290;
+		-- Modified Lentz's algorithm for computing continued fraction. See Numerical Recipes in C, 2nd edition, section 5.2
+		for j = 1, 100 do
+			local d;
+			local a, b = j * (s - j), j*2 + 1 + z - s;
+			D = b + a * D;
+			if D < TINY then D = TINY end
+			C = b + a / C;
+			if C < TINY then C = TINY end
+			D = 1. / D;
+			d = C * D;
+			f = f * d;
+			if math.abs(d - 1) < 1e-14 then break end
+		end
+		return math.exp(s * math.log(z) - z - math.lgamma(s) - math.log(f));
+	end
+
+	if complement then
+		return ((z <= 1 or z < s) and 1 - _kf_gammap(s, z)) or _kf_gammaq(s, z);
+	else 
+		return ((z <= 1 or z < s) and _kf_gammap(s, z)) or (1 - _kf_gammaq(s, z));
+	end
+end
+
+math.M_SQRT2   = 1.41421356237309504880  -- sqrt(2)
+math.M_SQRT1_2 = 0.70710678118654752440  -- 1/sqrt(2)
+
+-- Description: complement error function erfc(x): \Phi(x) = 0.5 * erfc(-x/M_SQRT2)
+function math.erfc(x)
+	local z = math.abs(x) * math.M_SQRT2
+	if z > 37 then return (x > 0 and 0) or 2 end
+	local expntl = math.exp(-0.5 * z * z)
+	local p
+	if z < 10. / math.M_SQRT2 then -- for small z
+	    p = expntl * ((((((.03526249659989109 * z + .7003830644436881) * z + 6.37396220353165) * z + 33.912866078383)
+				* z + 112.0792914978709) * z + 221.2135961699311) * z + 220.2068679123761)
+			/ (((((((.08838834764831844 * z + 1.755667163182642) * z + 16.06417757920695) * z + 86.78073220294608)
+				* z + 296.5642487796737) * z + 637.3336333788311) * z + 793.8265125199484) * z + 440.4137358247522);
+	else p = expntl / 2.506628274631001 / (z + 1. / (z + 2. / (z + 3. / (z + 4. / (z + .65))))) end
+	return (x > 0 and 2 * p) or 2 * (1 - p)
+end
+
+-- Description: log binomial coefficient
+-- Dependent on: math.lgamma()
+-- Required by: math.fisher_exact()
+function math.lbinom(n, m)
+	if m == nil then
+		local a = {};
+		a[0], a[n] = 0, 0;
+		local t = math.lgamma(n+1);
+		for m = 1, n-1 do a[m] = t - math.lgamma(m+1) - math.lgamma(n-m+1) end
+		return a;
+	else return math.lgamma(n+1) - math.lgamma(m+1) - math.lgamma(n-m+1) end
+end
+
+-- Description: Berstein polynomials (mainly for Bezier curves)
+-- Dependent on: math.lbinom()
+-- Note: to compute derivative: let beta_new[i]=beta[i+1]-beta[i]
+function math.bernstein_poly(beta)
+	local n = #beta - 1;
+	local lbc = math.lbinom(n); -- log binomial coefficients
+	return function (t)
+		assert(t >= 0 and t <= 1);
+		if t == 0 then return beta[1] end
+		if t == 1 then return beta[n+1] end
+		local sum, logt, logt1 = 0, math.log(t), math.log(1-t);
+		for i = 0, n do sum = sum + beta[i+1] * math.exp(lbc[i] + i * logt + (n-i) * logt1) end
+		return sum;
+	end
+end
+
+-- Description: Fisher's exact test
+-- Dependent on: math.lbinom()
+-- Return: left-, right- and two-tail P-values
+--[[
+  Fisher's exact test for 2x2 congintency tables:
+
+    n11  n12  | n1_
+    n21  n22  | n2_
+   -----------+----
+    n_1  n_2  | n
+
+  Reference: http://www.langsrud.com/fisher.htm
+]]--
+function math.fisher_exact(n11, n12, n21, n22)
+	local aux; -- keep the states of n* for acceleration
+
+	-- Description: hypergeometric function
+	local function hypergeo(n11, n1_, n_1, n)
+		return math.exp(math.lbinom(n1_, n11) + math.lbinom(n-n1_, n_1-n11) - math.lbinom(n, n_1));
+	end
+
+	-- Description: incremental hypergeometric function
+	-- Note: aux = {n11, n1_, n_1, n, p}
+	local function hypergeo_inc(n11, n1_, n_1, n)
+		if n1_ ~= 0 or n_1 ~= 0 or n ~= 0 then
+			aux = {n11, n1_, n_1, n, 1};
+		else -- then only n11 is changed
+			local mod;
+			_, mod = math.modf(n11 / 11);
+			if mod ~= 0 and n11 + aux[4] - aux[2] - aux[3] ~= 0 then
+				if n11 == aux[1] + 1 then -- increase by 1
+					aux[5] = aux[5] * (aux[2] - aux[1]) / n11 * (aux[3] - aux[1]) / (n11 + aux[4] - aux[2] - aux[3]);
+					aux[1] = n11;
+					return aux[5];
+				end
+				if n11 == aux[1] - 1 then -- descrease by 1
+					aux[5] = aux[5] * aux[1] / (aux[2] - n11) * (aux[1] + aux[4] - aux[2] - aux[3]) / (aux[3] - n11);
+					aux[1] = n11;
+					return aux[5];
+				end
+			end
+			aux[1] = n11;
+		end
+		aux[5] = hypergeo(aux[1], aux[2], aux[3], aux[4]);
+		return aux[5];
+	end
+	
+	-- Description: computing the P-value by Fisher's exact test
+	local max, min, left, right, n1_, n_1, n, two, p, q, i, j;
+	n1_, n_1, n = n11 + n12, n11 + n21, n11 + n12 + n21 + n22;
+	max = (n_1 < n1_ and n_1) or n1_; -- max n11, for the right tail
+	min = n1_ + n_1 - n;
+	if min < 0 then min = 0 end -- min n11, for the left tail
+	two, left, right = 1, 1, 1;
+	if min == max then return 1 end -- no need to do test
+	q = hypergeo_inc(n11, n1_, n_1, n); -- the probability of the current table
+	-- left tail
+	i, left, p = min + 1, 0, hypergeo_inc(min, 0, 0, 0);
+	while p < 0.99999999 * q do
+		left, p, i = left + p, hypergeo_inc(i, 0, 0, 0), i + 1;
+	end
+	i = i - 1;
+	if p < 1.00000001 * q then left = left + p;
+	else i = i - 1 end
+	-- right tail
+	j, right, p = max - 1, 0, hypergeo_inc(max, 0, 0, 0);
+	while p < 0.99999999 * q do
+		right, p, j = right + p, hypergeo_inc(j, 0, 0, 0), j - 1;
+	end
+	j = j + 1;
+	if p < 1.00000001 * q then right = right + p;
+	else j = j + 1 end
+	-- two-tail
+	two = left + right;
+	if two > 1 then two = 1 end
+	-- adjust left and right
+	if math.abs(i - n11) < math.abs(j - n11) then right = 1 - left + q;
+	else left = 1 - right + q end
+	return left, right, two;
+end
+
+-- Description: Delete-m Jackknife
+--[[
+  Given g groups of values with a statistics estimated from m[i] samples in
+  i-th group being t[i], compute the mean and the variance. t0 below is the
+  estimate from all samples. Reference:
+
+     Busing et al. (1999) Delete-m Jackknife for unequal m. Statistics and Computing, 9:3-8.
+]]--
+function math.jackknife(g, m, t, t0)
+	local h, n, sum = {}, 0, 0;
+	for j = 1, g do n = n + m[j] end
+	if t0 == nil then -- When t0 is absent, estimate it in a naive way
+		t0 = 0;
+		for j = 1, g do t0 = t0 + m[j] * t[j] end
+		t0 = t0 / n;
+	end
+	local mean, var = 0, 0;
+	for j = 1, g do
+		h[j] = n / m[j];
+		mean = mean + (1 - m[j] / n) * t[j];
+	end
+	mean = g * t0 - mean; -- Eq. (8)
+	for j = 1, g do
+		local x = h[j] * t0 - (h[j] - 1) * t[j] - mean;
+		var = var + 1 / (h[j] - 1) * x * x;
+	end
+	var = var / g;
+	return mean, var;
+end
+
+-- Description: Pearson correlation coefficient
+-- Input: a is an n*2 table
+function math.pearson(a)
+	-- compute the mean
+	local x1, y1 = 0, 0
+	for _, v in pairs(a) do
+		x1, y1 = x1 + v[1], y1 + v[2]
+	end
+	-- compute the coefficient
+	x1, y1 = x1 / #a, y1 / #a
+	local x2, y2, xy = 0, 0, 0
+	for _, v in pairs(a) do
+		local tx, ty = v[1] - x1, v[2] - y1
+		xy, x2, y2 = xy + tx * ty, x2 + tx * tx, y2 + ty * ty
+	end
+	return xy / math.sqrt(x2) / math.sqrt(y2)
+end
+
+-- Description: Spearman correlation coefficient
+function math.spearman(a)
+	local function aux_func(t) -- auxiliary function
+		return (t == 1 and 0) or (t*t - 1) * t / 12
+	end
+
+	for _, v in pairs(a) do v.r = {} end
+	local T, S = {}, {}
+	-- compute the rank
+	for k = 1, 2 do
+		table.sort(a, function(u,v) return u[k]<v[k] end)
+		local same = 1
+		T[k] = 0
+		for i = 2, #a + 1 do
+			if i <= #a and a[i-1][k] == a[i][k] then same = same + 1
+			else
+				local rank = (i-1) * 2 - same + 1
+				for j = i - same, i - 1 do a[j].r[k] = rank end
+				if same > 1 then T[k], same = T[k] + aux_func(same), 1 end
+			end
+		end
+		S[k] = aux_func(#a) - T[k]
+	end
+	-- compute the coefficient
+	local sum = 0
+	for _, v in pairs(a) do -- TODO: use nested loops to reduce loss of precision
+		local t = (v.r[1] - v.r[2]) / 2
+		sum = sum + t * t
+	end
+	return (S[1] + S[2] - sum) / 2 / math.sqrt(S[1] * S[2])
+end
+
+-- Description: Hooke-Jeeves derivative-free optimization
+function math.fmin(func, x, data, r, eps, max_calls)
+	local n, n_calls = #x, 0;
+	r = r or 0.5;
+	eps = eps or 1e-7;
+	max_calls = max_calls or 50000
+
+	function fmin_aux(x1, data, fx1, dx) -- auxiliary function
+		local ftmp;
+		for k = 1, n do
+			x1[k] = x1[k] + dx[k];
+			local ftmp = func(x1, data); n_calls = n_calls + 1;
+			if ftmp < fx1 then fx1 = ftmp;
+			else -- search the opposite direction
+				dx[k] = -dx[k];
+				x1[k] = x1[k] + dx[k] + dx[k];
+				ftmp = func(x1, data); n_calls = n_calls + 1;
+				if ftmp < fx1 then fx1 = ftmp
+				else x1[k] = x1[k] - dx[k] end -- back to the original x[k]
+			end
+		end
+		return fx1; -- here: fx1=f(n,x1)
+	end
+
+	local dx, x1 = {}, {};
+	for k = 1, n do -- initial directions, based on MGJ
+		dx[k] = math.abs(x[k]) * r;
+		if dx[k] == 0 then dx[k] = r end;
+	end
+	local radius = r;
+	local fx1, fx;
+	fx = func(x, data); fx1 = fx; n_calls = n_calls + 1;
+	while true do
+		for i = 1, n do x1[i] = x[i] end; -- x1 = x
+		fx1 = fmin_aux(x1, data, fx, dx);
+		while fx1 < fx do
+			for k = 1, n do
+				local t = x[k];
+				dx[k] = (x1[k] > x[k] and math.abs(dx[k])) or -math.abs(dx[k]);
+				x[k] = x1[k];
+				x1[k] = x1[k] + x1[k] - t;
+			end
+			fx = fx1;
+			if n_calls >= max_calls then break end
+			fx1 = func(x1, data); n_calls = n_calls + 1;
+			fx1 = fmin_aux(x1, data, fx1, dx);
+			if fx1 >= fx then break end
+			local kk = n;
+			for k = 1, n do
+				if math.abs(x1[k] - x[k]) > .5 * math.abs(dx[k]) then
+					kk = k;
+					break;
+				end
+			end
+			if kk == n then break end
+		end
+		if radius >= eps then
+			if n_calls >= max_calls then break end
+			radius = radius * r;
+			for k = 1, n do dx[k] = dx[k] * r end
+		else break end
+	end
+	return fx1, n_calls;
+end
+
+--
+-- Matrix
+--
+
+matrix = {}
+
+-- Description: matrix transpose
+-- Required by: matrix.mul()
+function matrix.T(a)
+	local m, n, x = #a, #a[1], {};
+	for i = 1, n do
+		x[i] = {};
+		for j = 1, m do x[i][j] = a[j][i] end
+	end
+	return x;
+end
+
+-- Description: matrix add
+function matrix.add(a, b)
+	assert(#a == #b and #a[1] == #b[1]);
+	local m, n, x = #a, #a[1], {};
+	for i = 1, m do
+		x[i] = {};
+		local ai, bi, xi = a[i], b[i], x[i];
+		for j = 1, n do xi[j] = ai[j] + bi[j] end
+	end
+	return x;
+end
+
+-- Description: matrix mul
+-- Dependent on: matrix.T()
+-- Note: much slower without transpose
+function matrix.mul(a, b)
+	assert(#a[1] == #b);
+	local m, n, p, x = #a, #a[1], #b[1], {};
+	local c = matrix.T(b); -- transpose for efficiency
+	for i = 1, m do
+		x[i] = {}
+		local xi = x[i];
+		for j = 1, p do
+			local sum, ai, cj = 0, a[i], c[j];
+			for k = 1, n do sum = sum + ai[k] * cj[k] end
+			xi[j] = sum;
+		end
+	end
+	return x;
+end
+
+-- Description: matrix print
+function matrix.tostring(a)
+	local z = {};
+	for i = 1, #a do
+		z[i] = table.concat(a[i], "\t");
+	end
+	return table.concat(z, "\n");
+end
+
+-- Description: chi^2 test for contingency tables
+-- Dependent on: math.igamma()
+function matrix.chi2(a)
+	if #a == 2 and #a[1] == 2 then -- 2x2 table
+		local x, z
+		x = (a[1][1] + a[1][2]) * (a[2][1] + a[2][2]) * (a[1][1] + a[2][1]) * (a[1][2] + a[2][2])
+		if x == 0 then return 0, 1, false end
+		z = a[1][1] * a[2][2] - a[1][2] * a[2][1]
+		z = (a[1][1] + a[1][2] + a[2][1] + a[2][2]) * z * z / x
+		return z, math.igamma(.5, .5 * z, true), true
+	else -- generic table
+		local rs, cs, n, m, N, z = {}, {}, #a, #a[1], 0, 0
+		for i = 1, n do rs[i] = 0 end
+		for j = 1, m do cs[j] = 0 end
+		for i = 1, n do -- compute column sum and row sum
+			for j = 1, m do cs[j], rs[i] = cs[j] + a[i][j], rs[i] + a[i][j] end
+		end
+		for i = 1, n do N = N + rs[i] end
+		for i = 1, n do -- compute the chi^2 statistics
+			for j = 1, m do
+				local E = rs[i] * cs[j] / N;
+				z = z + (a[i][j] - E) * (a[i][j] - E) / E
+			end
+		end
+		return z, math.igamma(.5 * (n-1) * (m-1), .5 * z, true), true;
+	end
+end
+
+-- Description: Gauss-Jordan elimination (solving equations; computing inverse)
+-- Note: on return, a[n][n] is the inverse; b[n][m] is the solution
+-- Reference: Section 2.1, Numerical Recipes in C, 2nd edition
+function matrix.solve(a, b)
+	assert(#a == #a[1]);
+	local n, m = #a, (b and #b[1]) or 0;
+	local xc, xr, ipiv = {}, {}, {};
+	local ic, ir;
+
+	for j = 1, n do ipiv[j] = 0 end
+	for i = 1, n do
+		local big = 0;
+		for j = 1, n do
+			local aj = a[j];
+			if ipiv[j] ~= 1 then
+				for k = 1, n do
+					if ipiv[k] == 0 then
+						if math.abs(aj[k]) >= big then
+							big = math.abs(aj[k]);
+							ir, ic = j, k;
+						end
+					elseif ipiv[k] > 1 then return -2 end -- singular matrix
+				end
+			end
+		end
+		ipiv[ic] = ipiv[ic] + 1;
+		if ir ~= ic then
+			for l = 1, n do a[ir][l], a[ic][l] = a[ic][l], a[ir][l] end
+			if b then
+				for l = 1, m do b[ir][l], b[ic][l] = b[ic][l], b[ir][l] end
+			end
+		end
+		xr[i], xc[i] = ir, ic;
+		if a[ic][ic] == 0 then return -3 end -- singular matrix
+		local pivinv = 1 / a[ic][ic];
+		a[ic][ic] = 1;
+		for l = 1, n do a[ic][l] = a[ic][l] * pivinv end
+		if b then
+			for l = 1, n do b[ic][l] = b[ic][l] * pivinv end
+		end
+		for ll = 1, n do
+			if ll ~= ic then
+				local tmp = a[ll][ic];
+				a[ll][ic] = 0;
+				local all, aic = a[ll], a[ic];
+				for l = 1, n do all[l] = all[l] - aic[l] * tmp end
+				if b then
+					local bll, bic = b[ll], b[ic];
+					for l = 1, m do bll[l] = bll[l] - bic[l] * tmp end
+				end
+			end
+		end
+	end
+	for l = n, 1, -1 do
+		if xr[l] ~= xc[l] then
+			for k = 1, n do a[k][xr[l]], a[k][xc[l]] = a[k][xc[l]], a[k][xr[l]] end
+		end
+	end
+	return 0;
+end
diff --git a/ext/klib/test/Makefile b/ext/klib/test/Makefile
new file mode 100644
index 0000000..c27d721
--- /dev/null
+++ b/ext/klib/test/Makefile
@@ -0,0 +1,72 @@
+CC=gcc
+CXX=g++
+CFLAGS=-g -Wall -O2 -I..
+CXXFLAGS=$(CFLAGS)
+PROGS=kbtree_test khash_keith khash_keith2 khash_test klist_test kseq_test kseq_bench \
+		kseq_bench2 ksort_test ksort_test-stl kvec_test kmin_test kstring_bench kstring_bench2 kstring_test \
+		kavl_test kavl-lite_test kthread_test2
+
+all:$(PROGS)
+
+clean:
+		rm -fr $(PROGS) *.dSYM a.out *.o
+
+kavl_test:kavl_test.c ../kavl.h
+		$(CC) $(CFLAGS) -o $@ kavl_test.c
+
+kavl-lite_test:kavl-lite_test.c ../kavl-lite.h
+		$(CC) $(CFLAGS) -o $@ kavl-lite_test.c
+
+kbtree_test:kbtree_test.c ../kbtree.h
+		$(CC) $(CFLAGS) -o $@ kbtree_test.c
+
+khash_keith:khash_keith.c ../khash.h
+		$(CC) $(CFLAGS) -o $@ khash_keith.c
+
+khash_keith2:khash_keith2.c ../khash.h
+		$(CC) $(CFLAGS) -o $@ khash_keith2.c
+
+khash_test:khash_test.c ../khash.h
+		$(CC) $(CFLAGS) -o $@ khash_test.c
+
+klist_test:klist_test.c ../klist.h
+		$(CC) $(CFLAGS) -o $@ klist_test.c
+
+kseq_test:kseq_test.c ../kseq.h
+		$(CC) $(CFLAGS) -o $@ kseq_test.c -lz
+
+kseq_bench:kseq_bench.c ../kseq.h
+		$(CC) $(CFLAGS) -o $@ kseq_bench.c -lz
+
+kseq_bench2:kseq_bench2.c ../kseq.h
+		$(CC) $(CFLAGS) -o $@ kseq_bench2.c -lz
+
+ksort_test:ksort_test.c ../ksort.h
+		$(CC) $(CFLAGS) -o $@ ksort_test.c
+
+ksort_test-stl:ksort_test.cc ../ksort.h
+		$(CXX) $(CXXFLAGS) -o $@ ksort_test.cc
+
+kvec_test:kvec_test.cc ../kvec.h
+		$(CXX) $(CXXFLAGS) -o $@ kvec_test.cc
+
+kmin_test:kmin_test.c ../kmath.h ../kmath.c
+		$(CC) $(CFLAGS) -o $@ kmin_test.c ../kmath.c
+
+kstring_bench:kstring_bench.c ../kstring.h ../kstring.c
+		$(CC) $(CFLAGS) -o $@ kstring_bench.c ../kstring.c
+
+kstring_bench2:kstring_bench2.c ../kstring.h ../kstring.c
+		$(CC) $(CFLAGS) -o $@ kstring_bench2.c ../kstring.c
+
+kstring_test:kstring_test.c ../kstring.h ../kstring.c
+		$(CC) $(CFLAGS) -o $@ kstring_test.c ../kstring.c
+
+kthread_test:kthread_test.c ../kthread.c
+		$(CC) $(CFLAGS) -fopenmp -o $@ kthread_test.c ../kthread.c
+
+kthread_test2:kthread_test2.c ../kthread.c
+		$(CC) $(CFLAGS) -o $@ kthread_test2.c ../kthread.c
+
+ketopt_test:ketopt_test.c ../ketopt.h
+		$(CC) $(CFLAGS) -o $@ ketopt_test.c
diff --git a/ext/klib/test/kavl-lite_test.c b/ext/klib/test/kavl-lite_test.c
new file mode 100644
index 0000000..375d5d3
--- /dev/null
+++ b/ext/klib/test/kavl-lite_test.c
@@ -0,0 +1,60 @@
+#include <stdio.h>
+#include <stdlib.h>
+#include <stdint.h>
+#include <string.h>
+#include "kavl-lite.h"
+
+#define CALLOC(type, num) ((type*)calloc(num, sizeof(type)))
+
+struct my_node {
+	int key;
+	KAVLL_HEAD(struct my_node) head;
+};
+
+#define my_cmp(p, q) (((p)->key > (q)->key) - ((p)->key < (q)->key))
+KAVLL_INIT(my, struct my_node, head, my_cmp)
+
+void shuffle(int n, char a[])
+{
+	int i, j;
+	for (i = n; i > 1; --i) {
+		char tmp;
+		j = (int)(drand48() * i);
+		tmp = a[j]; a[j] = a[i-1]; a[i-1] = tmp;
+	}
+}
+
+int main(void)
+{
+	char buf[256];
+	int i, n;
+	struct my_node *root = 0;
+	struct my_node *p, *q, t;
+	my_itr_t itr;
+
+	for (i = 33, n = 0; i <= 126; ++i)
+		if (i != '(' && i != ')' && i != '.' && i != ';')
+			buf[n++] = i;
+	shuffle(n, buf);
+	for (i = 0; i < n; ++i) {
+		p = CALLOC(struct my_node, 1);
+		p->key = buf[i];
+		q = my_insert(&root, p);
+		if (p != q) free(p);
+	}
+	shuffle(n, buf);
+	for (i = 0; i < n/2; ++i) {
+		t.key = buf[i];
+		q = my_erase(&root, &t);
+		if (q) free(q);
+	}
+
+	my_itr_first(root, &itr);
+	do {
+		const struct my_node *r = kavll_at(&itr);
+		putchar(r->key);
+		free((void*)r);
+	} while (my_itr_next(&itr));
+	putchar('\n');
+	return 0;
+}
diff --git a/ext/klib/test/kavl_test.c b/ext/klib/test/kavl_test.c
new file mode 100644
index 0000000..15ce395
--- /dev/null
+++ b/ext/klib/test/kavl_test.c
@@ -0,0 +1,104 @@
+#include <stdio.h>
+#include <stdlib.h>
+#include <stdint.h>
+#include <string.h>
+#include "kavl.h"
+
+#define CALLOC(type, num) ((type*)calloc(num, sizeof(type)))
+
+struct my_node {
+	int key;
+	KAVL_HEAD(struct my_node) head;
+};
+
+#define my_cmp(p, q) (((p)->key > (q)->key) - ((p)->key < (q)->key))
+KAVL_INIT(my, struct my_node, head, my_cmp)
+
+int check(struct my_node *p, int *hh)
+{
+	int c = 1, h[2] = {0, 0};
+	*hh = 0;
+	if (p) {
+		if (p->head.p[0]) c += check(p->head.p[0], &h[0]);
+		if (p->head.p[1]) c += check(p->head.p[1], &h[1]);
+		*hh = (h[0] > h[1]? h[0] : h[1]) + 1;
+		if (h[1] - h[0] != (int)p->head.balance)
+			fprintf(stderr, "%d - %d != %d at %c\n", h[1], h[0], p->head.balance, p->key);
+		if (c != (int)p->head.size)
+			fprintf(stderr, "%d != %d at %c\n", p->head.size, c, p->key);
+		return c;
+	} else return 0;
+}
+/*
+int print_tree(const struct my_node *p)
+{
+	int c = 1;
+	if (p == 0) return 0;
+	if (p->head.p[0] || p->head.p[1]) {
+		putchar('(');
+		if (p->head.p[0]) c += print_tree(p->head.p[0]);
+		else putchar('.');
+		putchar(',');
+		if (p->head.p[1]) c += print_tree(p->head.p[1]);
+		else putchar('.');
+		putchar(')');
+	}
+	putchar(p->key);
+	return c;
+}
+
+void check_and_print(struct my_node *root)
+{
+	int h;
+	check(root, &h);
+	print_tree(root);
+	putchar('\n');
+}
+*/
+void shuffle(int n, char a[])
+{
+	int i, j;
+	for (i = n; i > 1; --i) {
+		char tmp;
+		j = (int)(drand48() * i);
+		tmp = a[j]; a[j] = a[i-1]; a[i-1] = tmp;
+	}
+}
+
+int main(void)
+{
+	char buf[256];
+	int i, n, h;
+	struct my_node *root = 0;
+	struct my_node *p, *q, t;
+	kavl_itr_t(my) itr;
+	unsigned cnt;
+
+	for (i = 33, n = 0; i <= 126; ++i)
+		if (i != '(' && i != ')' && i != '.' && i != ';')
+			buf[n++] = i;
+	shuffle(n, buf);
+	for (i = 0; i < n; ++i) {
+		p = CALLOC(struct my_node, 1);
+		p->key = buf[i];
+		q = kavl_insert(my, &root, p, &cnt);
+		if (p != q) free(p);
+		check(root, &h);
+	}
+	shuffle(n, buf);
+	for (i = 0; i < n/2; ++i) {
+		t.key = buf[i];
+		q = kavl_erase(my, &root, &t, 0);
+		if (q) free(q);
+		check(root, &h);
+	}
+
+	kavl_itr_first(my, root, &itr);
+	do {
+		const struct my_node *r = kavl_at(&itr);
+		putchar(r->key);
+		free((void*)r);
+	} while (kavl_itr_next(my, &itr));
+	putchar('\n');
+	return 0;
+}
diff --git a/ext/klib/test/kbit_test.c b/ext/klib/test/kbit_test.c
new file mode 100644
index 0000000..3ae3bd3
--- /dev/null
+++ b/ext/klib/test/kbit_test.c
@@ -0,0 +1,137 @@
+#include <stdlib.h>
+#include <stdio.h>
+#include <time.h>
+#include <emmintrin.h>
+#include "kbit.h"
+
+// from bowtie-0.9.8.1
+inline static int bt1_pop64(uint64_t x) // the kbi_popcount64() equivalence; similar to popcount_2() in wiki
+{
+   x -= ((x >> 1) & 0x5555555555555555llu);
+   x = (x & 0x3333333333333333llu) + ((x >> 2) & 0x3333333333333333llu);
+   x = (x + (x >> 4)) & 0x0F0F0F0F0F0F0F0Fllu;
+   x = x + (x >> 8);
+   x = x + (x >> 16);
+   x = x + (x >> 32);
+   return x & 0x3F;
+}
+
+inline static int bt1_countInU64(uint64_t dw, int c) // the kbi_DNAcount64() equivalence
+{
+	uint64_t dwA  = dw &  0xAAAAAAAAAAAAAAAAllu;
+	uint64_t dwNA = dw & ~0xAAAAAAAAAAAAAAAAllu;
+	uint64_t tmp;
+	switch (c) {
+	case 0: tmp = (dwA >> 1) | dwNA; break;
+	case 1: tmp = ~(dwA >> 1) & dwNA; break;
+	case 2: tmp = (dwA >> 1) & ~dwNA; break;
+	default: tmp = (dwA >> 1) & dwNA;
+	}
+	tmp = bt1_pop64(tmp);
+	if (c == 0) tmp = 32 - tmp;
+	return (int)tmp;
+}
+
+// from bigmagic
+static uint32_t sse2_bit_count32(const __m128i* block, const __m128i* block_end)
+{
+    const unsigned mu1 = 0x55555555;
+    const unsigned mu2 = 0x33333333;
+    const unsigned mu3 = 0x0F0F0F0F;
+    const unsigned mu4 = 0x0000003F;
+
+	uint32_t tcnt[4];
+
+    // Loading masks
+    __m128i m1 = _mm_set_epi32 (mu1, mu1, mu1, mu1);
+    __m128i m2 = _mm_set_epi32 (mu2, mu2, mu2, mu2);
+    __m128i m3 = _mm_set_epi32 (mu3, mu3, mu3, mu3);
+    __m128i m4 = _mm_set_epi32 (mu4, mu4, mu4, mu4);
+    __m128i mcnt;
+    mcnt = _mm_xor_si128(m1, m1); // cnt = 0
+
+    __m128i tmp1, tmp2;
+    do
+    {        
+        __m128i b = _mm_load_si128(block);
+        ++block;
+
+        // b = (b & 0x55555555) + (b >> 1 & 0x55555555);
+        tmp1 = _mm_srli_epi32(b, 1);                    // tmp1 = (b >> 1 & 0x55555555)
+        tmp1 = _mm_and_si128(tmp1, m1); 
+        tmp2 = _mm_and_si128(b, m1);                    // tmp2 = (b & 0x55555555)
+        b    = _mm_add_epi32(tmp1, tmp2);               //  b = tmp1 + tmp2
+
+        // b = (b & 0x33333333) + (b >> 2 & 0x33333333);
+        tmp1 = _mm_srli_epi32(b, 2);                    // (b >> 2 & 0x33333333)
+        tmp1 = _mm_and_si128(tmp1, m2); 
+        tmp2 = _mm_and_si128(b, m2);                    // (b & 0x33333333)
+        b    = _mm_add_epi32(tmp1, tmp2);               // b = tmp1 + tmp2
+
+        // b = (b + (b >> 4)) & 0x0F0F0F0F;
+        tmp1 = _mm_srli_epi32(b, 4);                    // tmp1 = b >> 4
+        b = _mm_add_epi32(b, tmp1);                     // b = b + (b >> 4)
+        b = _mm_and_si128(b, m3);                       //           & 0x0F0F0F0F
+
+        // b = b + (b >> 8);
+        tmp1 = _mm_srli_epi32 (b, 8);                   // tmp1 = b >> 8
+        b = _mm_add_epi32(b, tmp1);                     // b = b + (b >> 8)
+
+        // b = (b + (b >> 16)) & 0x0000003F;
+        tmp1 = _mm_srli_epi32 (b, 16);                  // b >> 16
+        b = _mm_add_epi32(b, tmp1);                     // b + (b >> 16)
+        b = _mm_and_si128(b, m4);                       // (b >> 16) & 0x0000003F;
+
+        mcnt = _mm_add_epi32(mcnt, b);                  // mcnt += b
+
+    } while (block < block_end);
+
+    _mm_store_si128((__m128i*)tcnt, mcnt);
+
+    return tcnt[0] + tcnt[1] + tcnt[2] + tcnt[3];
+}
+
+int main(void)
+{
+	int i, N = 100000000;
+	uint64_t *x, cnt;
+	clock_t t;
+	int c = 1;
+
+	x = (uint64_t*)calloc(N, 8);
+	srand48(11);
+	for (i = 0; i < N; ++i)
+		x[i] = (uint64_t)lrand48() << 32 ^ lrand48();
+
+	fprintf(stderr, "\n===> Calculate # of 1 in an integer (popcount) <===\n");
+
+	t = clock(); cnt = 0;
+	for (i = 0; i < N; ++i) cnt += kbi_popcount64(x[i]);
+	fprintf(stderr, "%20s\t%20ld\t%10.6f\n", "kbit", (long)cnt, (double)(clock() - t) / CLOCKS_PER_SEC);
+
+	t = clock(); cnt = 0;
+	for (i = 0; i < N; ++i) cnt += bt1_pop64(x[i]);
+	fprintf(stderr, "%20s\t%20ld\t%10.6f\n", "wiki-popcount_2", (long)cnt, (double)(clock() - t) / CLOCKS_PER_SEC);
+
+	t = clock(); cnt = 0;
+	for (i = 0; i < N; ++i) cnt += __builtin_popcountl(x[i]);
+	fprintf(stderr, "%20s\t%20ld\t%10.6f\n", "__builtin_popcountl", (long)cnt, (double)(clock() - t) / CLOCKS_PER_SEC);
+
+	t = clock(); cnt = 0;
+	cnt += sse2_bit_count32((__m128i*)x, (__m128i*)(x+N));
+	fprintf(stderr, "%20s\t%20ld\t%10.6f\n", "SSE2-32bit", (long)cnt, (double)(clock() - t) / CLOCKS_PER_SEC);
+
+	fprintf(stderr, "\n===> Count '%c' in 2-bit encoded integers <===\n", "ACGT"[c]);
+
+	t = clock(); cnt = 0;
+	for (i = 0; i < N; ++i) cnt += kbi_DNAcount64(x[i], c);
+	fprintf(stderr, "%20s\t%20ld\t%10.6f\n", "kbit", (long)cnt, (double)(clock() - t) / CLOCKS_PER_SEC);
+
+	t = clock(); cnt = 0;
+	for (i = 0; i < N; ++i) cnt += bt1_countInU64(x[i], c);
+	fprintf(stderr, "%20s\t%20ld\t%10.6f\n", "bowtie1", (long)cnt, (double)(clock() - t) / CLOCKS_PER_SEC);
+
+	fprintf(stderr, "\n");
+	free(x);
+	return 0;
+}
diff --git a/ext/klib/test/kbtree_test.c b/ext/klib/test/kbtree_test.c
new file mode 100644
index 0000000..8e10687
--- /dev/null
+++ b/ext/klib/test/kbtree_test.c
@@ -0,0 +1,94 @@
+#include <stdio.h>
+#include <assert.h>
+#include <time.h>
+#include <stdlib.h>
+#include <stdint.h>
+#include <string.h>
+
+typedef const char *str_t;
+
+#include "kbtree.h"
+KBTREE_INIT(int, uint32_t, kb_generic_cmp)
+KBTREE_INIT(str, str_t, kb_str_cmp)
+
+static int data_size = 5000000;
+static unsigned *int_data;
+static char **str_data;
+
+void ht_init_data()
+{
+	int i;
+	char buf[256];
+	printf("--- generating data... ");
+	srand48(11);
+	int_data = (unsigned*)calloc(data_size, sizeof(unsigned));
+	str_data = (char**)calloc(data_size, sizeof(char*));
+	for (i = 0; i < data_size; ++i) {
+		int_data[i] = (unsigned)(data_size * drand48() / 4) * 271828183u;
+		sprintf(buf, "%x", int_data[i]);
+		str_data[i] = strdup(buf);
+	}
+	printf("done!\n");
+}
+void ht_destroy_data()
+{
+	int i;
+	for (i = 0; i < data_size; ++i) free(str_data[i]);
+	free(str_data); free(int_data);
+}
+
+void ht_khash_int()
+{
+	int i;
+	unsigned *data = int_data;
+	uint32_t *l, *u;
+	kbtree_t(int) *h;
+
+	h = kb_init(int, KB_DEFAULT_SIZE);
+	for (i = 0; i < data_size; ++i) {
+		if (kb_get(int, h, data[i]) == 0) kb_put(int, h, data[i]);
+		else kb_del(int, h, data[i]);
+	}
+	printf("[ht_khash_int] size: %d\n", kb_size(h));
+	if (1) {
+		int cnt = 0;
+		uint32_t x, y;
+		kb_interval(int, h, 2174625464u, &l, &u);
+		printf("interval for 2174625464: (%u, %u)\n", l? *l : 0, u? *u : 0);
+#define traverse_f(p) { if (cnt == 0) y = *p; ++cnt; }
+		__kb_traverse(uint32_t, h, traverse_f);
+		__kb_get_first(uint32_t, h, x);
+		printf("# of elements from traversal: %d\n", cnt);
+		printf("first element: %d == %d\n", x, y);
+	}
+	__kb_destroy(h);
+}
+void ht_khash_str()
+{
+	int i;
+	char **data = str_data;
+	kbtree_t(str) *h;
+
+	h = kb_init(str, KB_DEFAULT_SIZE);
+	for (i = 0; i < data_size; ++i) {
+		if (kb_get(str, h, data[i]) == 0) kb_put(str, h, data[i]);
+		else kb_del(str, h, data[i]);
+	}
+	printf("[ht_khash_int] size: %d\n", kb_size(h));
+	__kb_destroy(h);
+}
+void ht_timing(void (*f)(void))
+{
+	clock_t t = clock();
+	(*f)();
+	printf("[ht_timing] %.3lf sec\n", (double)(clock() - t) / CLOCKS_PER_SEC);
+}
+int main(int argc, char *argv[])
+{
+	if (argc > 1) data_size = atoi(argv[1]);
+	ht_init_data();
+	ht_timing(ht_khash_int);
+	ht_timing(ht_khash_str);
+	ht_destroy_data();
+	return 0;
+}
diff --git a/ext/klib/test/ketopt_test.c b/ext/klib/test/ketopt_test.c
new file mode 100644
index 0000000..4009e24
--- /dev/null
+++ b/ext/klib/test/ketopt_test.c
@@ -0,0 +1,89 @@
+#include <stdio.h>
+#include <stdlib.h>
+#include <getopt.h>
+#include "ketopt.h"
+
+static void test_opt(int c, int opt, const char *arg)
+{
+	if (c == 'x') fprintf(stderr, "-x\n");
+	else if (c == 'y') fprintf(stderr, "-y %s\n", arg);
+	else if (c == 301) fprintf(stderr, "--foo\n");
+	else if (c == 302) fprintf(stderr, "--bar %s\n", arg? arg : "(null)");
+	else if (c == 303) fprintf(stderr, "--opt %s\n", arg? arg : "(null)");
+	else if (c == '?') fprintf(stderr, "unknown option -%c\n", opt? opt : ':');
+	else if (c == ':') fprintf(stderr, "missing option argument: -%c\n", opt? opt : ':');
+}
+
+static void print_cmd(int argc, char *argv[], int ind)
+{
+	int i;
+	fprintf(stderr, "CMD: %s", argv[0]);
+	if (ind > 1) {
+		fputs(" [", stderr);
+		for (i = 1; i < ind; ++i) {
+			if (i != 1) fputc(' ', stderr);
+			fputs(argv[i], stderr);
+		}
+		fputc(']', stderr);
+	}
+	for (i = ind; i < argc; ++i)
+		fprintf(stderr, " %s", argv[i]);
+	fputc('\n', stderr);
+}
+
+static void test_ketopt(int argc, char *argv[])
+{
+	static ko_longopt_t longopts[] = {
+		{ "foo", ko_no_argument,       301 },
+		{ "bar", ko_required_argument, 302 },
+		{ "opt", ko_optional_argument, 303 },
+		{ NULL, 0, 0 }
+	};
+	ketopt_t opt = KETOPT_INIT;
+	int c;
+	fprintf(stderr, "===> ketopt() <===\n");
+	while ((c = ketopt(&opt, argc, argv, 1, "xy:", longopts)) >= 0)
+		test_opt(c, opt.opt, opt.arg);
+	print_cmd(argc, argv, opt.ind);
+}
+
+static void test_getopt(int argc, char *argv[])
+{
+	static struct option long_options[] = {
+		{ "foo", no_argument,       0, 301 },
+		{ "bar", required_argument, 0, 302 },
+		{ "opt", optional_argument, 0, 303 },
+		{0, 0, 0, 0}
+	};
+	int c, option_index;
+	fprintf(stderr, "===> getopt() <===\n");
+	while ((c = getopt_long(argc, argv, ":xy:", long_options, &option_index)) >= 0)
+		test_opt(c, optopt, optarg);
+	print_cmd(argc, argv, optind);
+}
+
+int main(int argc, char *argv[])
+{
+	int i;
+	char **argv2;
+	if (argc == 1) {
+		fprintf(stderr, "Usage: ketopt_test [options] <argument> [...]\n");
+		fprintf(stderr, "Options:\n");
+		fprintf(stderr, "  -x           no argument\n");
+		fprintf(stderr, "  -y STR       required argument\n");
+		fprintf(stderr, "  --foo        no argument\n");
+		fprintf(stderr, "  --bar=STR    required argument\n");
+		fprintf(stderr, "  --opt[=STR]  optional argument\n");
+		fprintf(stderr, "\nExamples:\n");
+		fprintf(stderr, "  ketopt_test -xy1 -x arg1 -y -x -- arg2 -x\n");
+		fprintf(stderr, "  ketopt_test --foo --bar=1 --bar 2 --opt arg1 --opt=3\n");
+		fprintf(stderr, "  ketopt_test arg1 -y\n");
+		return 1;
+	}
+	argv2 = (char**)malloc(sizeof(char*) * argc);
+	for (i = 0; i < argc; ++i) argv2[i] = argv[i];
+	test_ketopt(argc, argv);
+	test_getopt(argc, argv2);
+	free(argv2);
+	return 0;
+}
diff --git a/ext/klib/test/kgraph_test.c b/ext/klib/test/kgraph_test.c
new file mode 100644
index 0000000..3da1cd7
--- /dev/null
+++ b/ext/klib/test/kgraph_test.c
@@ -0,0 +1,26 @@
+#include <stdio.h>
+#include "kgraph.h"
+
+KHASH_INIT2(e32, extern, uint32_t, int, 1, kh_int_hash_func, kh_int_hash_equal)
+
+typedef struct {
+	int i;
+	khash_t(e32) *_arc;
+} vertex_t;
+
+KGRAPH_INIT(g, extern, vertex_t, int, e32)
+KGRAPH_PRINT(g, extern)
+
+int main()
+{
+	int *pb, *pe;
+	kgraph_t(g) *g;
+	g = kg_init_g();
+	kg_put_a_g(g, 10, 20, 0, &pb, &pe);
+	kg_put_a_g(g, 20, 30, 0, &pb, &pe);
+	kg_put_a_g(g, 30, 10, 1, &pb, &pe);
+	kg_del_v_g(g, 20);
+	kg_print_g(g);
+	kg_destroy_g(g);
+	return 0;
+}
diff --git a/ext/klib/test/khash_keith.c b/ext/klib/test/khash_keith.c
new file mode 100644
index 0000000..ddd755a
--- /dev/null
+++ b/ext/klib/test/khash_keith.c
@@ -0,0 +1,95 @@
+/*
+ * This is an optimized version of the following C++ program:
+ *
+ *   http://keithlea.com/javabench/src/cpp/hash.cpp
+ *
+ * Keith in his benchmark (http://keithlea.com/javabench/data) showed that the
+ * Java implementation is twice as fast as the C++ version. In fact, this is
+ * only because the C++ implementation is substandard. Most importantly, Keith
+ * is using "sprintf()" to convert an integer to a string, which is known to be
+ * extremely inefficient.
+ */
+#include <stdio.h>
+#include "khash.h"
+KHASH_MAP_INIT_STR(str, int)
+
+inline void int2str(int c, int base, char *ret)
+{
+	const char *tab = "0123456789abcdef";
+	if (c == 0) ret[0] = '0', ret[1] = 0;
+	else {
+		int l, x, y;
+		char buf[16];
+		for (l = 0, x = c < 0? -c : c; x > 0; x /= base) buf[l++] = tab[x%base];
+		if (c < 0) buf[l++] = '-';
+		for (x = l - 1, y = 0; x >= 0; --x) ret[y++] = buf[x];
+		ret[y] = 0;
+	}
+}
+
+#ifndef _USE_STRDUP
+#define BLOCK_SIZE 0x100000
+int main(int argc, char *argv[])
+{
+	char **mem = 0;
+	int i, l, n = 1000000, ret, block_end = 0, curr = 0, c = 0;
+	khash_t(str) *h;
+	h = kh_init(str);
+	if (argc > 1) n = atoi(argv[1]);
+	mem = malloc(sizeof(void*));
+	mem[0] = malloc(BLOCK_SIZE); // memory buffer to avoid memory fragmentation
+	curr = block_end = 0;
+	for (i = 1; i <= n; ++i) {
+		char buf[16];
+		int2str(i, 16, buf);
+		khint_t k = kh_put(str, h, buf, &ret);
+		l = strlen(buf) + 1;
+		if (block_end + l > BLOCK_SIZE) {
+			++curr; block_end = 0;
+			mem = realloc(mem, (curr + 1) * sizeof(void*));
+			mem[curr] = malloc(BLOCK_SIZE);
+		}
+		memcpy(mem[curr] + block_end, buf, l);
+		kh_key(h, k) = mem[curr] + block_end;
+		block_end += l;
+		kh_val(h, k) = i;
+	}
+	for (i = 1; i <= n; ++i) {
+		char buf[16];
+		int2str(i, 10, buf);
+		khint_t k = kh_get(str, h, buf);
+		if (k != kh_end(h)) ++c;
+	}
+	printf("%d\n", c);
+	for (ret = 0; ret <= curr; ++ret) free(mem[ret]);
+	free(mem);
+	kh_destroy(str, h);
+	return 0;
+}
+#else // _USE_STRDUP
+int main(int argc, char *argv[])
+{
+	int i, l, n = 1000000, ret, c = 0;
+	khash_t(str) *h;
+	khint_t k;
+	h = kh_init(str);
+	if (argc > 1) n = atoi(argv[1]);
+	for (i = 1; i <= n; ++i) {
+		char buf[16];
+		int2str(i, 16, buf);
+		k = kh_put(str, h, strdup(buf), &ret);
+		kh_val(h, k) = i;
+	}
+	for (i = 1; i <= n; ++i) {
+		char buf[16];
+		int2str(i, 10, buf);
+		k = kh_get(str, h, buf);
+		if (k != kh_end(h)) ++c;
+	}
+	for (k = kh_begin(h); k != kh_end(h); ++k) // explicitly freeing memory takes 10-20% CPU time.
+		if (kh_exist(h, k)) free((char*)kh_key(h, k));
+	printf("%d\n", c);
+	kh_destroy(str, h);
+	return 0;
+}
+#endif
diff --git a/ext/klib/test/khash_keith2.c b/ext/klib/test/khash_keith2.c
new file mode 100644
index 0000000..b9df9b7
--- /dev/null
+++ b/ext/klib/test/khash_keith2.c
@@ -0,0 +1,67 @@
+/*
+ * This is an optimized version of the following C++ program:
+ *
+ *   http://keithlea.com/javabench/src/cpp/hash.cpp
+ *
+ * Keith in his benchmark (http://keithlea.com/javabench/data) showed that the
+ * Java implementation is twice as fast as the C++ version. In fact, this is
+ * only because the C++ implementation is substandard. Most importantly, Keith
+ * is using "sprintf()" to convert an integer to a string, which is known to be
+ * extremely inefficient.
+ */
+#include <stdio.h>
+#include "khash.h"
+KHASH_MAP_INIT_STR(str, int)
+
+inline void int2str(int c, int base, char *ret)
+{
+	const char *tab = "0123456789abcdef";
+	if (c == 0) ret[0] = '0', ret[1] = 0;
+	else {
+		int l, x, y;
+		char buf[16];
+		for (l = 0, x = c < 0? -c : c; x > 0; x /= base) buf[l++] = tab[x%base];
+		if (c < 0) buf[l++] = '-';
+		for (x = l - 1, y = 0; x >= 0; --x) ret[y++] = buf[x];
+		ret[y] = 0;
+	}
+}
+
+int main(int argc, char *argv[])
+{
+	int i, l, n = 1000, ret;
+	khash_t(str) *h, *h2;
+	khint_t k;
+	h = kh_init(str);
+	h2 = kh_init(str);
+	if (argc > 1) n = atoi(argv[1]);
+	for (i = 0; i < 10000; ++i) {
+		char buf[32];
+		strcpy(buf, "foo_");
+		int2str(i, 10, buf+4);
+		k = kh_put(str, h, strdup(buf), &ret);
+		kh_val(h, k) = i;
+	}
+	for (i = 0; i < n; ++i) {
+		for (k = kh_begin(h); k != kh_end(h); ++k) {
+			if (kh_exist(h, k)) {
+				khint_t k2 = kh_put(str, h2, kh_key(h, k), &ret);
+				if (ret) { // absent
+					kh_key(h2, k2) = strdup(kh_key(h, k));
+					kh_val(h2, k2) = kh_val(h, k);
+				} else kh_val(h2, k2) += kh_val(h, k);
+			}
+		}
+	}
+	k = kh_get(str, h, "foo_1"); printf("%d", kh_val(h, k));
+	k = kh_get(str, h, "foo_9999"); printf(" %d", kh_val(h, k));
+	k = kh_get(str, h2, "foo_1"); printf(" %d", kh_val(h2, k));
+	k = kh_get(str, h2, "foo_9999"); printf(" %d\n", kh_val(h2, k));
+	for (k = kh_begin(h); k != kh_end(h); ++k)
+		if (kh_exist(h, k)) free((char*)kh_key(h, k));
+	for (k = kh_begin(h2); k != kh_end(h2); ++k)
+		if (kh_exist(h2, k)) free((char*)kh_key(h2, k));
+	kh_destroy(str, h);
+	kh_destroy(str, h2);
+	return 0;
+}
diff --git a/ext/klib/test/khash_test.c b/ext/klib/test/khash_test.c
new file mode 100644
index 0000000..8d6687f
--- /dev/null
+++ b/ext/klib/test/khash_test.c
@@ -0,0 +1,141 @@
+#include <stdio.h>
+#include <assert.h>
+#include <time.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "khash.h"
+KHASH_SET_INIT_STR(str)
+KHASH_MAP_INIT_INT(int, unsigned char)
+
+typedef struct {
+	unsigned key;
+	unsigned char val;
+} int_unpack_t;
+
+typedef struct {
+	unsigned key;
+	unsigned char val;
+} __attribute__ ((__packed__)) int_packed_t;
+
+#define hash_eq(a, b) ((a).key == (b).key)
+#define hash_func(a) ((a).key)
+
+KHASH_INIT(iun, int_unpack_t, char, 0, hash_func, hash_eq)
+KHASH_INIT(ipk, int_packed_t, char, 0, hash_func, hash_eq)
+
+static int data_size = 5000000;
+static unsigned *int_data;
+static char **str_data;
+
+void ht_init_data()
+{
+	int i;
+	char buf[256];
+	khint32_t x = 11;
+	printf("--- generating data... ");
+	int_data = (unsigned*)calloc(data_size, sizeof(unsigned));
+	str_data = (char**)calloc(data_size, sizeof(char*));
+	for (i = 0; i < data_size; ++i) {
+		int_data[i] = (unsigned)(data_size * ((double)x / UINT_MAX) / 4) * 271828183u;
+		sprintf(buf, "%x", int_data[i]);
+		str_data[i] = strdup(buf);
+		x = 1664525L * x + 1013904223L;
+	}
+	printf("done!\n");
+}
+
+void ht_destroy_data()
+{
+	int i;
+	for (i = 0; i < data_size; ++i) free(str_data[i]);
+	free(str_data); free(int_data);
+}
+
+void ht_khash_int()
+{
+	int i, ret;
+	unsigned *data = int_data;
+	khash_t(int) *h;
+	unsigned k;
+
+	h = kh_init(int);
+	for (i = 0; i < data_size; ++i) {
+		k = kh_put(int, h, data[i], &ret);
+		kh_val(h, k) = i&0xff;
+		if (!ret) kh_del(int, h, k);
+	}
+	printf("[ht_khash_int] size: %u\n", kh_size(h));
+	kh_destroy(int, h);
+}
+
+void ht_khash_str()
+{
+	int i, ret;
+	char **data = str_data;
+	khash_t(str) *h;
+	unsigned k;
+
+	h = kh_init(str);
+	for (i = 0; i < data_size; ++i) {
+		k = kh_put(str, h, data[i], &ret);
+		if (!ret) kh_del(str, h, k);
+	}
+	printf("[ht_khash_int] size: %u\n", kh_size(h));
+	kh_destroy(str, h);
+}
+
+void ht_khash_unpack()
+{
+	int i, ret;
+	unsigned *data = int_data;
+	khash_t(iun) *h;
+	unsigned k;
+
+	h = kh_init(iun);
+	for (i = 0; i < data_size; ++i) {
+		int_unpack_t x;
+		x.key = data[i]; x.val = i&0xff;
+		k = kh_put(iun, h, x, &ret);
+		if (!ret) kh_del(iun, h, k);
+	}
+	printf("[ht_khash_unpack] size: %u (sizeof=%ld)\n", kh_size(h), sizeof(int_unpack_t));
+	kh_destroy(iun, h);
+}
+
+void ht_khash_packed()
+{
+	int i, ret;
+	unsigned *data = int_data;
+	khash_t(ipk) *h;
+	unsigned k;
+
+	h = kh_init(ipk);
+	for (i = 0; i < data_size; ++i) {
+		int_packed_t x;
+		x.key = data[i]; x.val = i&0xff;
+		k = kh_put(ipk, h, x, &ret);
+		if (!ret) kh_del(ipk, h, k);
+	}
+	printf("[ht_khash_packed] size: %u (sizeof=%ld)\n", kh_size(h), sizeof(int_packed_t));
+	kh_destroy(ipk, h);
+}
+
+void ht_timing(void (*f)(void))
+{
+	clock_t t = clock();
+	(*f)();
+	printf("[ht_timing] %.3lf sec\n", (double)(clock() - t) / CLOCKS_PER_SEC);
+}
+
+int main(int argc, char *argv[])
+{
+	if (argc > 1) data_size = atoi(argv[1]);
+	ht_init_data();
+	ht_timing(ht_khash_int);
+	ht_timing(ht_khash_str);
+	ht_timing(ht_khash_unpack);
+	ht_timing(ht_khash_packed);
+	ht_destroy_data();
+	return 0;
+}
diff --git a/ext/klib/test/klist_test.c b/ext/klib/test/klist_test.c
new file mode 100644
index 0000000..cd13813
--- /dev/null
+++ b/ext/klib/test/klist_test.c
@@ -0,0 +1,19 @@
+#include <stdio.h>
+#include "klist.h"
+
+#define __int_free(x)
+KLIST_INIT(32, int, __int_free)
+
+int main()
+{
+	klist_t(32) *kl;
+	kliter_t(32) *p;
+	kl = kl_init(32);
+	*kl_pushp(32, kl) = 1;
+	*kl_pushp(32, kl) = 10;
+	kl_shift(32, kl, 0);
+	for (p = kl_begin(kl); p != kl_end(kl); p = kl_next(p))
+		printf("%d\n", kl_val(p));
+	kl_destroy(32, kl);
+	return 0;
+}
diff --git a/ext/klib/test/kmin_test.c b/ext/klib/test/kmin_test.c
new file mode 100644
index 0000000..33ccd1c
--- /dev/null
+++ b/ext/klib/test/kmin_test.c
@@ -0,0 +1,48 @@
+#include <stdio.h>
+#include <math.h>
+#include "kmath.h"
+
+static int n_evals;
+
+double f_Chebyquad(int n, double *x, void *data)
+{
+    int i, j;
+    double y[20][20], f;
+    int np, iw;
+    double sum;
+    for (j = 0; j != n; ++j) {
+		y[0][j] = 1.;
+		y[1][j] = 2. * x[j] - 1.;
+    }
+    for (i = 1; i != n; ++i)
+		for (j = 0; j != n; ++j)
+			y[i+1][j] = 2. * y[1][j] * y[i][j] - y[i-1][j];
+    f = 0.;
+    np = n + 1;
+    iw = 1;
+    for (i = 0; i != np; ++i) {
+		sum = 0.;
+		for (j = 0; j != n; ++j) sum += y[i][j];
+		sum /= n;
+		if (iw > 0) sum += 1. / ((i - 1) * (i + 1));
+		iw = -iw;
+		f += sum * sum;
+    }
+	++n_evals;
+    return f;
+}
+
+int main()
+{
+	double x[20], y;
+	int n, i;
+	printf("\nMinimizer: Hooke-Jeeves\n");
+	for (n = 2; n <= 8; n += 2) {
+		for (i = 0; i != n; ++i) x[i] = (double)(i + 1) / n;
+		n_evals = 0;
+		y = kmin_hj(f_Chebyquad, n, x, 0, KMIN_RADIUS, KMIN_EPS, KMIN_MAXCALL);
+		printf("n=%d,min=%.8lg,n_evals=%d\n", n, y, n_evals);
+	}
+	printf("\n");
+	return 0;
+}
diff --git a/ext/klib/test/krmq_test.c b/ext/klib/test/krmq_test.c
new file mode 100644
index 0000000..69aff4c
--- /dev/null
+++ b/ext/klib/test/krmq_test.c
@@ -0,0 +1,151 @@
+#include <stdio.h>
+#include <assert.h>
+#include <stdlib.h>
+#include <stdint.h>
+#include <string.h>
+#include <limits.h>
+#include "krmq.h"
+
+#define CALLOC(type, num) ((type*)calloc(num, sizeof(type)))
+
+struct my_node {
+	int key;
+	int val;
+	KRMQ_HEAD(struct my_node) head;
+};
+
+#define my_cmp(p, q) (((p)->key > (q)->key) - ((p)->key < (q)->key))
+#define my_lt2(p, q) ((p)->val < (q)->val)
+KRMQ_INIT(my, struct my_node, head, my_cmp, my_lt2)
+
+int check(struct my_node *p, int *hh, int *min)
+{
+	*hh = 0, *min = INT_MAX;
+	if (p) {
+		int c = 1, h[2] = {0, 0}, m[2] = {INT_MAX, INT_MAX};
+		*min = p->val;
+		if (p->head.p[0]) c += check(p->head.p[0], &h[0], &m[0]);
+		if (p->head.p[1]) c += check(p->head.p[1], &h[1], &m[1]);
+		*min = *min < m[0]? *min : m[0];
+		*min = *min < m[1]? *min : m[1];
+		*hh = (h[0] > h[1]? h[0] : h[1]) + 1;
+		if (*min != p->head.s->val)
+			fprintf(stderr, "min %d != %d at %c\n", *min, p->head.s->val, p->key);
+		if (h[1] - h[0] != (int)p->head.balance)
+			fprintf(stderr, "%d - %d != %d at %c\n", h[1], h[0], p->head.balance, p->key);
+		if (c != (int)p->head.size)
+			fprintf(stderr, "%d != %d at %c\n", p->head.size, c, p->key);
+		return c;
+	} else return 0;
+}
+
+int check_rmq(const struct my_node *root, int lo, int hi)
+{
+	struct my_node s, t, *p, *q;
+	krmq_itr_t(my) itr;
+	int min = INT_MAX;
+	s.key = lo, t.key = hi;
+	p = krmq_rmq(my, root, &s, &t);
+	krmq_interval(my, root, &s, 0, &q);
+	if (p == 0) return -1;
+	krmq_itr_find(my, root, q, &itr);
+	do {
+		const struct my_node *r = krmq_at(&itr);
+		if (r->key > hi) break;
+		//fprintf(stderr, "%c\t%d\n", r->key, r->val);
+		if (r->val < min) min = r->val;
+	} while (krmq_itr_next(my, &itr));
+	assert((min == INT_MAX && p == 0) || (min < INT_MAX && p));
+	if (min != p->val) fprintf(stderr, "rmq_min %d != %d\n", p->val, min);
+	return min;
+}
+
+int print_tree(const struct my_node *p)
+{
+	int c = 1;
+	if (p == 0) return 0;
+	if (p->head.p[0] || p->head.p[1]) {
+		putchar('(');
+		if (p->head.p[0]) c += print_tree(p->head.p[0]);
+		else putchar('.');
+		putchar(',');
+		if (p->head.p[1]) c += print_tree(p->head.p[1]);
+		else putchar('.');
+		putchar(')');
+	}
+	printf("%c:%d/%d", p->key, p->val, p->head.s->val);
+	return c;
+}
+
+void check_and_print(struct my_node *root)
+{
+	int h, min;
+	check(root, &h, &min);
+	print_tree(root);
+	putchar('\n');
+}
+
+void shuffle(int n, char a[])
+{
+	int i, j;
+	for (i = n; i > 1; --i) {
+		char tmp;
+		j = (int)(drand48() * i);
+		tmp = a[j]; a[j] = a[i-1]; a[i-1] = tmp;
+	}
+}
+
+int main(void)
+{
+	char buf[256];
+	int i, n, h, min;
+	struct my_node *root = 0;
+	struct my_node *p, *q, t;
+	krmq_itr_t(my) itr;
+	unsigned cnt;
+
+	srand48(123);
+	for (i = 33, n = 0; i <= 126; ++i)
+		if (i != '(' && i != ')' && i != '.' && i != ';')
+			buf[n++] = i;
+	shuffle(n, buf);
+	for (i = 0; i < n; ++i) {
+		p = CALLOC(struct my_node, 1);
+		p->key = buf[i];
+		p->val = i;
+		q = krmq_insert(my, &root, p, &cnt);
+		if (p != q) free(p);
+		check(root, &h, &min);
+	}
+
+	shuffle(n, buf);
+	for (i = 0; i < n/2; ++i) {
+		t.key = buf[i];
+		//fprintf(stderr, "i=%d, key=%c, n/2=%d\n", i, t.key, n/2);
+		q = krmq_erase(my, &root, &t, 0);
+		if (q) free(q);
+		check(root, &h, &min);
+	}
+	check_and_print(root);
+
+	check_rmq(root, '0', '9');
+	check_rmq(root, '!', '~');
+	check_rmq(root, 'A', 'Z');
+	check_rmq(root, 'F', 'G');
+	check_rmq(root, 'a', 'z');
+	for (i = 0; i < n; ++i) {
+		int lo, hi;
+		lo = (int)(drand48() * n);
+		hi = (int)(drand48() * n);
+		check_rmq(root, lo, hi);
+	}
+
+	krmq_itr_first(my, root, &itr);
+	do {
+		const struct my_node *r = krmq_at(&itr);
+		putchar(r->key);
+	} while (krmq_itr_next(my, &itr));
+	putchar('\n');
+	krmq_free(struct my_node, head, root, free);
+	return 0;
+}
diff --git a/ext/klib/test/kseq_bench.c b/ext/klib/test/kseq_bench.c
new file mode 100644
index 0000000..eeda13f
--- /dev/null
+++ b/ext/klib/test/kseq_bench.c
@@ -0,0 +1,69 @@
+#include <zlib.h>
+#include <stdio.h>
+#include <time.h>
+#include <stdint.h>
+#include <stdlib.h>
+#include "kseq.h"
+
+#define BUF_SIZE 4096
+KSTREAM_INIT(gzFile, gzread, BUF_SIZE)
+
+int main(int argc, char *argv[])
+{
+	gzFile fp;
+	clock_t t;
+	if (argc == 1) {
+		fprintf(stderr, "Usage: kseq_bench <in.gz>\n");
+		return 1;
+	}
+	{
+		uint8_t *buf = malloc(BUF_SIZE);
+		fp = gzopen(argv[1], "r");
+		t = clock();
+		while (gzread(fp, buf, BUF_SIZE) > 0);
+		fprintf(stderr, "[gzread] %.2f sec\n", (float)(clock() - t) / CLOCKS_PER_SEC);
+		gzclose(fp);
+		free(buf);
+	}
+	{
+		kstream_t *ks;
+		fp = gzopen(argv[1], "r");
+		ks = ks_init(fp);
+		t = clock();
+		while (ks_getc(ks) >= 0);
+		fprintf(stderr, "[ks_getc] %.2f sec\n", (float)(clock() - t) / CLOCKS_PER_SEC);
+		ks_destroy(ks);
+		gzclose(fp);
+	}
+	{
+		kstream_t *ks;
+		kstring_t *s;
+		int dret;
+		s = calloc(1, sizeof(kstring_t));
+		fp = gzopen(argv[1], "r");
+		ks = ks_init(fp);
+		t = clock();
+		while (ks_getuntil(ks, '\n', s, &dret) >= 0);
+		fprintf(stderr, "[ks_getuntil] %.2f sec\n", (float)(clock() - t) / CLOCKS_PER_SEC);
+		ks_destroy(ks);
+		gzclose(fp);
+		free(s->s); free(s);
+	}
+	if (argc == 2) {
+		fp = gzopen(argv[1], "r");
+		t = clock();
+		while (gzgetc(fp) >= 0);
+		fprintf(stderr, "[gzgetc] %.2f sec\n", (float)(clock() - t) / CLOCKS_PER_SEC);
+		gzclose(fp);
+	}
+	if (argc == 2) {
+		char *buf = malloc(BUF_SIZE);
+		fp = gzopen(argv[1], "r");
+		t = clock();
+		while (gzgets(fp, buf, BUF_SIZE) > 0);
+		fprintf(stderr, "[gzgets] %.2f sec\n", (float)(clock() - t) / CLOCKS_PER_SEC);
+		gzclose(fp);
+		free(buf);
+	}
+	return 0;
+}
diff --git a/ext/klib/test/kseq_bench2.c b/ext/klib/test/kseq_bench2.c
new file mode 100644
index 0000000..3267777
--- /dev/null
+++ b/ext/klib/test/kseq_bench2.c
@@ -0,0 +1,44 @@
+#include <stdio.h>
+#include <time.h>
+#include <stdint.h>
+#include <stdlib.h>
+#include <fcntl.h>
+#include <unistd.h>
+#include "kseq.h"
+KSTREAM_INIT(int, read, 4096)
+
+#define BUF_SIZE 65536
+
+int main(int argc, char *argv[])
+{
+	clock_t t;
+	if (argc == 1) {
+		fprintf(stderr, "Usage: %s <in.txt>\n", argv[0]);
+		return 1;
+	}
+	{
+		FILE *fp;
+		char *s;
+		t = clock();
+		s = malloc(BUF_SIZE);
+		fp = fopen(argv[1], "r");
+		while (fgets(s, BUF_SIZE, fp));
+		fclose(fp);
+		fprintf(stderr, "[fgets] %.2f sec\n", (float)(clock() - t) / CLOCKS_PER_SEC);
+	}
+	{
+		int fd, dret;
+		kstream_t *ks;
+		kstring_t s;
+		t = clock();
+		s.l = s.m = 0; s.s = 0;
+		fd = open(argv[1], O_RDONLY);
+		ks = ks_init(fd);
+		while (ks_getuntil(ks, '\n', &s, &dret) >= 0);
+		free(s.s);
+		ks_destroy(ks);
+		close(fd);
+		fprintf(stderr, "[kstream] %.2f sec\n", (float)(clock() - t) / CLOCKS_PER_SEC);
+	}
+	return 0;
+}
diff --git a/ext/klib/test/kseq_test.c b/ext/klib/test/kseq_test.c
new file mode 100644
index 0000000..0304dea
--- /dev/null
+++ b/ext/klib/test/kseq_test.c
@@ -0,0 +1,27 @@
+#include <zlib.h>
+#include <stdio.h>
+#include "kseq.h"
+KSEQ_INIT(gzFile, gzread)
+
+int main(int argc, char *argv[])
+{
+	gzFile fp;
+	kseq_t *seq;
+	int l;
+	if (argc == 1) {
+		fprintf(stderr, "Usage: %s <in.fasta>\n", argv[0]);
+		return 1;
+	}
+	fp = gzopen(argv[1], "r");
+	seq = kseq_init(fp);
+	while ((l = kseq_read(seq)) >= 0) {
+		printf("name: %s\n", seq->name.s);
+		if (seq->comment.l) printf("comment: %s\n", seq->comment.s);
+		printf("seq: %s\n", seq->seq.s);
+		if (seq->qual.l) printf("qual: %s\n", seq->qual.s);
+	}
+	printf("return value: %d\n", l);
+	kseq_destroy(seq);
+	gzclose(fp);
+	return 0;
+}
diff --git a/ext/klib/test/kseq_test.dat b/ext/klib/test/kseq_test.dat
new file mode 100644
index 0000000..b774ae2
--- /dev/null
+++ b/ext/klib/test/kseq_test.dat
@@ -0,0 +1,12 @@
+>1
+acgtacgtacgtagc
+>2 test
+acgatcgatc
+@3 test2
+cgctagcatagc
+cgatatgactta
++
+78wo82usd980
+d88fau
+
+238ud8
diff --git a/ext/klib/test/ksort_test.c b/ext/klib/test/ksort_test.c
new file mode 100644
index 0000000..92c7d3d
--- /dev/null
+++ b/ext/klib/test/ksort_test.c
@@ -0,0 +1,104 @@
+#include <stdlib.h>
+#include <string.h>
+#include <stdio.h>
+#include <time.h>
+#include "ksort.h"
+
+KSORT_INIT_GENERIC(int)
+
+int main(int argc, char *argv[])
+{
+	int i, N = 10000000;
+	int *array, x;
+	clock_t t1, t2;
+	if (argc > 1) N = atoi(argv[1]);
+	array = (int*)malloc(sizeof(int) * N);
+
+	srand48(11);
+	for (i = 0; i < N; ++i) array[i] = (int)lrand48();
+	t1 = clock();
+	x = ks_ksmall(int, N, array, 10500);
+	t2 = clock();
+	fprintf(stderr, "ksmall [%d]: %.3lf\n", x, (double)(t2-t1)/CLOCKS_PER_SEC);
+
+	srand48(11);
+	for (i = 0; i < N; ++i) array[i] = (int)lrand48();
+	t1 = clock();
+	ks_introsort(int, N, array);
+	t2 = clock();
+	fprintf(stderr, "introsort [%d]: %.3lf\n", array[10500], (double)(t2-t1)/CLOCKS_PER_SEC);
+	for (i = 0; i < N-1; ++i) {
+		if (array[i] > array[i+1]) {
+			fprintf(stderr, "Bug in introsort!\n");
+			exit(1);
+		}
+	}
+
+#ifndef _ALIGNED_ONLY
+	{ // test unaligned ksmall
+		srand48(11);
+		unsigned char *a;
+		int *b;
+		a = malloc(N * sizeof(int) + 1);
+		b = (int*)(a + 1);
+		for (i = 0; i < N; ++i) b[i] = (int)lrand48();
+		t1 = clock();
+		ks_introsort(int, N, b);
+		t2 = clock();
+		fprintf(stderr, "introsort [%d]: %.3lf (unaligned: 0x%lx) \n", b[10500], (double)(t2-t1)/CLOCKS_PER_SEC, (size_t)b);
+	}
+#endif
+
+	t1 = clock();
+	ks_introsort(int, N, array);
+	t2 = clock();
+	fprintf(stderr, "introsort (sorted): %.3lf\n", (double)(t2-t1)/CLOCKS_PER_SEC);
+
+	srand48(11);
+	for (i = 0; i < N; ++i) array[i] = (int)lrand48();
+	t1 = clock();
+	ks_combsort(int, N, array);
+	t2 = clock();
+	fprintf(stderr, "combsort: %.3lf\n", (double)(t2-t1)/CLOCKS_PER_SEC);
+	for (i = 0; i < N-1; ++i) {
+		if (array[i] > array[i+1]) {
+			fprintf(stderr, "Bug in combsort!\n");
+			exit(1);
+		}
+	}
+
+	srand48(11);
+	for (i = 0; i < N; ++i) array[i] = (int)lrand48();
+	t1 = clock();
+	ks_mergesort(int, N, array, 0);
+	t2 = clock();
+	fprintf(stderr, "mergesort: %.3lf\n", (double)(t2-t1)/CLOCKS_PER_SEC);
+	for (i = 0; i < N-1; ++i) {
+		if (array[i] > array[i+1]) {
+			fprintf(stderr, "Bug in mergesort!\n");
+			exit(1);
+		}
+	}
+
+	t1 = clock();
+	ks_mergesort(int, N, array, 0);
+	t2 = clock();
+	fprintf(stderr, "mergesort (sorted): %.3lf\n", (double)(t2-t1)/CLOCKS_PER_SEC);
+
+	srand48(11);
+	for (i = 0; i < N; ++i) array[i] = (int)lrand48();
+	t1 = clock();
+	ks_heapmake(int, N, array);
+	ks_heapsort(int, N, array);
+	t2 = clock();
+	fprintf(stderr, "heapsort: %.3lf\n", (double)(t2-t1)/CLOCKS_PER_SEC);
+	for (i = 0; i < N-1; ++i) {
+		if (array[i] > array[i+1]) {
+			fprintf(stderr, "Bug in heapsort!\n");
+			exit(1);
+		}
+	}
+
+	free(array);
+	return 0;
+}
diff --git a/ext/klib/test/ksort_test.cc b/ext/klib/test/ksort_test.cc
new file mode 100644
index 0000000..8950d80
--- /dev/null
+++ b/ext/klib/test/ksort_test.cc
@@ -0,0 +1,997 @@
+#include <stdlib.h>
+#include <string.h>
+#include <stdio.h>
+#include <time.h>
+#include <algorithm>
+
+#include "ksort.h"
+KSORT_INIT_GENERIC(int)
+
+using namespace std;
+
+/**********************************
+ * BEGIN OF PAUL'S IMPLEMENTATION *
+ **********************************/
+
+/* Attractive Chaos: I have added inline where necessary. */
+
+/*
+Copyright (c) 2004 Paul Hsieh
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+
+    Redistributions of source code must retain the above copyright notice,
+    this list of conditions and the following disclaimer.
+
+    Redistributions in binary form must reproduce the above copyright notice,
+    this list of conditions and the following disclaimer in the documentation
+    and/or other materials provided with the distribution.
+
+    Neither the name of sorttest nor the names of its contributors may be
+    used to endorse or promote products derived from this software without
+    specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+POSSIBILITY OF SUCH DAMAGE.
+*/
+
+/*
+
+Recommended flags:
+------------------
+
+Intel C/C++:
+icl /O2 /G6 /Qaxi /Qxi /Qip sorttest.c
+
+WATCOM C/C++:
+wcl386 /otexan /6r sorttest.c
+
+GCC:
+gcc -O3 -mcpu=athlon-xp -march=athlon-xp sorttest.c
+
+MSVC:
+cl /O2 /Ot /Og /G6 sorttest.c
+
+*/
+
+static inline void sort2 (int * numbers) {
+int tmp;
+
+    if (numbers[0] <= numbers[1]) return;
+    tmp = numbers[0];
+    numbers[0] = numbers[1];
+    numbers[1] = tmp;
+}
+
+static inline void sort3 (int * numbers) {
+int tmp;
+
+    if (numbers[0] <= numbers[1]) {
+        if (numbers[1] <= numbers[2]) return;
+        if (numbers[2] <= numbers[0]) {
+            tmp = numbers[0];
+            numbers[0] = numbers[2];
+            numbers[2] = numbers[1];
+            numbers[1] = tmp;
+            return;
+        }
+        tmp = numbers[1];
+    } else {
+        tmp = numbers[0];
+        if (numbers[0] <= numbers[2]) {
+            numbers[0] = numbers[1];
+            numbers[1] = tmp;
+            return;
+        }
+        if (numbers[2] <= numbers[1]) {
+            numbers[0] = numbers[2];
+            numbers[2] = tmp;
+            return;
+        }
+        numbers[0] = numbers[1];
+    }
+    numbers[1] = numbers[2];
+    numbers[2] = tmp;
+}
+
+static inline void sort4 (int * num) {
+int tmp;
+  if (num[0] < num[1]) {
+    if (num[1] < num[2]) {
+      if (num[1] < num[3]) {
+        if (num[2] >= num[3]) {
+          tmp = num[2];
+          num[2] = num[3];
+          num[3] = tmp;
+        }
+      } else {
+        tmp = num[1];
+        if (num[0] < num[3]) {
+          num[1] = num[3];
+        } else {
+          num[1] = num[0];
+          num[0] = num[3];
+        }
+        num[3] = num[2];
+        num[2] = tmp;
+      }
+    } else {
+      if (num[0] < num[2]) {
+        if (num[2] < num[3]) {
+          if (num[1] < num[3]) {
+            tmp = num[1];
+          } else {
+            tmp = num[3];
+            num[3] = num[1];
+          }
+          num[1] = num[2];
+          num[2] = tmp;
+        } else {
+          if (num[0] < num[3]) {
+            tmp = num[3];
+          } else {
+            tmp = num[0];
+            num[0] = num[3];
+          }
+          num[3] = num[1];
+          num[1] = tmp;
+        }
+      } else {
+        if (num[0] < num[3]) {
+          tmp = num[0];
+          num[0] = num[2];
+          if (num[1] < num[3]) {
+            num[2] = num[1];
+          } else {
+            num[2] = num[3];
+            num[3] = num[1];
+          }
+          num[1] = tmp;
+        } else {
+          if (num[2] < num[3]) {
+            tmp = num[0];
+            num[0] = num[2];
+            num[2] = tmp;
+            tmp = num[1];
+            num[1] = num[3];
+          } else {
+            tmp = num[1];
+            num[1] = num[2];
+            num[2] = num[0];
+            num[0] = num[3];
+          }
+          num[3] = tmp;
+        }
+      }
+    }
+  } else {
+    tmp = num[0];
+    if (tmp < num[2]) {
+      if (tmp < num[3]) {
+        num[0] = num[1];
+        num[1] = tmp;
+        if (num[2] >= num[3]) {
+          tmp = num[2];
+          num[2] = num[3];
+          num[3] = tmp;
+        }
+      } else {
+        if (num[1] < num[3]) {
+          num[0] = num[1];
+          num[1] = num[3];
+        } else {
+          num[0] = num[3];
+        }
+        num[3] = num[2];
+        num[2] = tmp;
+      }
+    } else {
+      if (num[1] < num[2]) {
+        if (num[2] < num[3]) {
+          num[0] = num[1];
+          num[1] = num[2];
+          if (tmp < num[3]) {
+            num[2] = tmp;
+          } else {
+            num[2] = num[3];
+            num[3] = tmp;
+          }
+        } else {
+          if (num[1] < num[3]) {
+            num[0] = num[1];
+            num[1] = num[3];
+          } else {
+            num[0] = num[3];
+          }
+          num[3] = tmp;
+        }
+      } else {
+        if (num[1] < num[3]) {
+          num[0] = num[2];
+          if (tmp < num[3]) {
+            num[2] = tmp;
+          } else {
+            num[2] = num[3];
+            num[3] = tmp;
+          }
+        } else {
+          if (num[2] < num[3]) {
+            num[0] = num[2];
+            num[2] = num[1];
+            num[1] = num[3];
+            num[3] = tmp;
+          } else {
+            num[0] = num[3];
+            num[3] = tmp;
+            tmp = num[1];
+            num[1] = num[2];
+            num[2] = tmp;
+          }
+        }
+      }
+    }
+  }
+}
+
+static inline void sortAlt2 (int * numbers, int * altNumbers) {
+    if (numbers[0] <= numbers[1]) {
+        altNumbers[0] = numbers[0];
+        altNumbers[1] = numbers[1];
+    } else {
+        altNumbers[0] = numbers[1];
+        altNumbers[1] = numbers[0];
+    }
+}
+
+static inline void sortAlt3 (int * numbers, int * altNumbers) {
+    if (numbers[0] <= numbers[1]) {
+        if (numbers[1] <= numbers[2]) {
+            altNumbers[0] = numbers[0];
+            altNumbers[1] = numbers[1];
+            altNumbers[2] = numbers[2];
+        } else if (numbers[2] <= numbers[0]) {
+            altNumbers[0] = numbers[2];
+            altNumbers[1] = numbers[0];
+            altNumbers[2] = numbers[1];
+        } else {
+            altNumbers[0] = numbers[0];
+            altNumbers[1] = numbers[2];
+            altNumbers[2] = numbers[1];
+        }
+    } else {
+        if (numbers[0] <= numbers[2]) {
+            altNumbers[0] = numbers[1];
+            altNumbers[1] = numbers[0];
+            altNumbers[2] = numbers[2];
+        } else if (numbers[2] <= numbers[1]) {
+            altNumbers[0] = numbers[2];
+            altNumbers[1] = numbers[1];
+            altNumbers[2] = numbers[0];
+        } else {
+            altNumbers[0] = numbers[1];
+            altNumbers[1] = numbers[2];
+            altNumbers[2] = numbers[0];
+        }
+    }
+}
+
+/*
+ *  Insert Sort
+ */
+
+inline void insertSort (int numbers[], int qty) {
+int i, j, idx, q4;
+int tmp;
+
+    if (qty <= 4) {
+        if (qty == 4) sort4 (numbers);
+        else if (qty == 3) sort3 (numbers);
+        else if (qty == 2) sort2 (numbers);
+        return;
+    }
+
+    q4 = qty - 4;
+
+    for (i=0; i < q4; i++) {
+        idx = i;
+        for (j=i+1; j < qty; j++) {
+            if (numbers[j] < numbers[idx]) idx = j;
+        }
+        if (idx != i) {
+            tmp = numbers[idx];
+            numbers[idx] = numbers[i];
+            numbers[i] = tmp;
+        }
+    }
+
+    sort4 (numbers + q4);
+}
+
+/*
+ *  Heap Sort
+ */
+
+/* Assure the heap property for entries from top to last */
+static void siftDown (int numbers[], int top, int last) {
+int tmp = numbers[top];
+int maxIdx = top;
+
+    while (last >= (maxIdx += maxIdx)) {
+
+        /* This is where the comparison occurrs and where a sufficiently
+           good compiler can use a computed conditional result rather
+           than using control logic. */
+        if (maxIdx != last && numbers[maxIdx] < numbers[maxIdx + 1]) maxIdx++;
+
+        if (tmp >= numbers[maxIdx]) break;
+        numbers[top] = numbers[maxIdx];
+        top = maxIdx;
+    }
+    numbers[top] = tmp;
+}
+
+/* Peel off the top siftDown operation since its parameters are trivial to
+   fill in directly (and this saves us some moves.) */
+static void siftDown0 (int numbers[], int last) {
+int tmp;
+
+    if (numbers[0] < numbers[1]) {
+        tmp = numbers[1];
+        numbers[1] = numbers[0];
+        siftDown (numbers, 1, last);
+    } else {
+        tmp = numbers[0];
+    }
+    numbers[0] = numbers[last];
+    numbers[last] = tmp;
+}
+
+void heapSort (int numbers[], int qty) {
+int i;
+
+    if (qty <= 4) {
+        if (qty == 4) sort4 (numbers);
+        else if (qty == 3) sort3 (numbers);
+        else if (qty == 2) sort2 (numbers);
+        return;
+    }
+
+    i = qty / 2;
+    /* Enforce the heap property for each position in the tree */
+    for (  qty--; i >  0; i--) siftDown  (numbers, i, qty);
+    for (i = qty; i > 0; i--) siftDown0 (numbers, i);
+}
+
+/*
+ *  Quick Sort
+ */
+
+static int medianOf3 (int * numbers, int i, int j) {
+int tmp;
+
+    if (numbers[0] <= numbers[i]) {
+        if (numbers[j] <= numbers[0]) return numbers[0]; /* j 0 i */
+        if (numbers[i] <= numbers[j]) j = i;             /* 0 i j */
+                                                         /* 0 j i */
+    } else {
+        if (numbers[0] <= numbers[j]) return numbers[0]; /* i 0 j */
+        if (numbers[j] <= numbers[i]) j = i;             /* j i 0 */
+                                                         /* i j 0 */
+    }
+    tmp = numbers[j];
+    numbers[j] = numbers[0];
+    numbers[0] = tmp;
+    return tmp;
+}
+
+static void quickSortRecurse (int * numbers, int left, int right) {
+int pivot, lTmp, rTmp;
+
+    qsrStart:;
+
+#if defined(__GNUC__)
+    if (right <= left + 8) {
+        insertSort (numbers + left, right - left + 1);
+        return;
+    }
+#else
+    if (right <= left + 3) {
+        if (right == left + 1) {
+            sort2 (numbers + left);
+        } else if (right == left + 2) {
+            sort3 (numbers + left);
+        } else if (right == left + 3) {
+            sort4 (numbers + left);
+        }
+        return;
+    }
+#endif
+
+    lTmp = left;
+    rTmp = right;
+
+    pivot = medianOf3 (numbers + left, (right-left) >> 1, right-1-left);
+
+    goto QStart;
+    while (1) {
+        do {
+            right--;
+            if (left >= right) goto QEnd;
+            QStart:;
+        } while (numbers[right] > pivot);
+        numbers[left] = numbers[right];
+        do { 
+            left++;
+            if (left >= right) {
+                left = right;
+                goto QEnd;
+            }
+        } while (numbers[ left] < pivot);
+        numbers[right] = numbers[left];
+    }
+    QEnd:;
+    numbers[left] = pivot;
+
+    /* Only recurse the smaller partition */
+
+    if (left-1 - lTmp <= rTmp - left - 1) {
+        if (lTmp < left) quickSortRecurse (numbers,   lTmp, left-1);
+
+        /* Set up for larger partition */
+        left++;
+        right = rTmp;
+    } else {
+        if (rTmp > left) quickSortRecurse (numbers, left+1,   rTmp);
+
+        /* Set up for larger partition */
+        right = left - 1;
+        left = lTmp;
+    }
+
+    /* Rerun with larger partition (recursion not required.) */
+    goto qsrStart;
+}
+
+void quickSort (int numbers[], int qty) {
+    if (qty < 2) return;
+    quickSortRecurse (numbers, 0, qty - 1);
+}
+
+/*
+ *  Merge Sort
+ */
+
+static void mergesortInPlace (int * numbers, int * altNumbers, int qty);
+
+/* Perform mergesort, but store results in altNumbers */
+
+static void mergesortExchange (int * numbers, int * altNumbers, int qty) {
+int half, i0, i1, i;
+
+    if (qty == 2) {
+        sortAlt2 (numbers, altNumbers);
+        return;
+    }
+    if (qty == 3) {
+        sortAlt3 (numbers, altNumbers);
+        return;
+    }
+
+    half = (qty + 1)/2;
+
+    mergesortInPlace (numbers, altNumbers, half);
+    mergesortInPlace (numbers + half, altNumbers, qty - half);
+
+    i0 = 0; i1 = half;
+
+    for (i=0; i < qty; i++) {
+        if (i1 >= qty || (i0 < half && numbers[i0] < numbers[i1])) {
+            altNumbers[i] = numbers[i0];
+            i0++;
+        } else {
+            altNumbers[i] = numbers[i1];
+            i1++;
+        }
+    }
+}
+
+/* Perform mergesort and store results in numbers */
+
+static void mergesortInPlace (int * numbers, int * altNumbers, int qty) {
+int half, i0, i1, i;
+
+#if 0
+    if (qty == 2) {
+        sort2 (numbers);
+        return;
+    }
+    if (qty == 3) {
+        sort3 (numbers);
+        return;
+    }
+    if (qty == 4) {
+        sort4 (numbers);
+        return;
+    }
+#else
+    if (qty <= 12) {
+        insertSort (numbers, qty);
+        return;
+    }
+#endif
+
+    half = (qty + 1)/2;
+
+    mergesortExchange (numbers, altNumbers, half);
+    mergesortExchange (numbers + half, altNumbers + half, qty - half);
+
+    i0 = 0; i1 = half;
+
+    for (i=0; i < qty; i++) {
+        if (i1 >= qty || (i0 < half && altNumbers[i0] < altNumbers[i1])) {
+            numbers[i] = altNumbers[i0];
+            i0++;
+        } else {
+            numbers[i] = altNumbers[i1];
+            i1++;
+        }
+    }
+}
+
+#include <stdlib.h>
+
+void mergeSort (int numbers[], int qty) {
+int * tmpArray;
+
+    if (qty <= 12) {
+        insertSort (numbers, qty);
+        return;
+    }
+
+    tmpArray = (int *) malloc (qty * sizeof (int));
+    mergesortInPlace (numbers, tmpArray, qty);
+    free (tmpArray);
+}
+
+/********************************
+ * END OF PAUL'S IMPLEMENTATION *
+ ********************************/
+
+/*************************************************
+ *** Implementation 1: faster on sorted arrays ***
+ *************************************************/
+
+#define rstype_t unsigned
+#define rskey(x) (x)
+
+#define RS_MIN_SIZE 64
+
+typedef struct {
+	rstype_t *b, *e;
+} rsbucket_t;
+
+void rs_sort(rstype_t *beg, rstype_t *end, int n_bits, int s)
+{
+	rstype_t *i;
+	int size = 1<<n_bits, m = size - 1;
+	rsbucket_t *k, b[size], *be = b + size;
+
+	for (k = b; k != be; ++k) k->b = k->e = beg;
+	for (i = beg; i != end; ++i) ++b[rskey(*i)>>s&m].e;
+	for (k = b + 1; k != be; ++k)
+		k->e += (k-1)->e - beg, k->b = (k-1)->e;
+	for (k = b; k != be;) {
+		if (k->b != k->e) {
+			rsbucket_t *l;
+			if ((l = b + (rskey(*k->b)>>s&m)) != k) {
+				rstype_t tmp = *k->b, swap;
+				do {
+					swap = tmp; tmp = *l->b; *l->b++ = swap;
+					l = b + (rskey(tmp)>>s&m);
+				} while (l != k);
+				*k->b++ = tmp;
+			} else ++k->b;
+		} else ++k;
+	}
+	for (b->b = beg, k = b + 1; k != be; ++k) k->b = (k-1)->e;
+	if (s) {
+		s = s > n_bits? s - n_bits : 0;
+		for (k = b; k != be; ++k)
+			if (k->e - k->b > RS_MIN_SIZE) rs_sort(k->b, k->e, n_bits, s);
+			else if (k->e - k->b > 1)
+				for (i = k->b + 1; i < k->e; ++i)
+					if (rskey(*i) < rskey(*(i - 1))) {
+						rstype_t *j, tmp = *i;
+						for (j = i; j > k->b && rskey(tmp) < rskey(*(j-1)); --j)
+							*j = *(j - 1);
+						*j = tmp;
+					}
+	}
+}
+
+/*************************************************
+ *** Implementation 2: faster on random arrays ***
+ *************************************************/
+
+static inline void rs_insertsort(rstype_t *s, rstype_t *t)
+{
+	rstype_t *i;
+	for (i = s + 1; i < t; ++i) {
+		if (rskey(*i) < rskey(*(i - 1))) {
+			rstype_t *j, tmp = *i;
+			for (j = i; j > s && rskey(tmp) < rskey(*(j-1)); --j)
+				*j = *(j - 1);
+			*j = tmp;
+		}
+	}
+}
+/*
+void rs_sort2(rstype_t *beg, rstype_t *end, int n_bits, int s)
+{
+	int j, size = 1<<n_bits, m = size - 1;
+	unsigned long c[size];
+	rstype_t *i, *b[size], *e[size];
+
+	for (j = 0; j < size; ++j) c[j] = 0;
+	for (i = beg; i != end; ++i) ++c[rskey(*i)>>s&m];
+	b[0] = e[0] = beg;
+	for (j = 1; j != size; ++j) b[j] = e[j] = b[j - 1] + c[j - 1];
+	for (i = beg, j = 0; i != end;) {
+		rstype_t tmp = *i, swap;
+		int x;
+		for (;;) {
+			x = rskey(tmp)>>s&m;
+			if (e[x] == i) break;
+			swap = tmp; tmp = *e[x]; *e[x]++ = swap;
+		}
+		*i++ = tmp;
+		++e[x];
+		while (j != size && i >= b[j]) ++j;
+		while (j != size && e[j-1] == b[j]) ++j;
+		if (i < e[j-1]) i = e[j-1];
+	}
+	if (s) {
+		s = s > n_bits? s - n_bits : 0;
+		for (j = 0; j < size; ++j) {
+			if (c[j] >= RS_MIN_SIZE) rs_sort2(b[j], e[j], n_bits, s);
+			else if (c[j] >= 2) rs_insertsort(b[j], e[j]);
+		}
+	}
+}
+*/
+void radix_sort(unsigned *array, int offset, int end, int shift) {
+    int x, y, value, temp;
+    int last[256] = { 0 }, pointer[256];
+
+    for (x=offset; x<end; ++x) {
+        ++last[(array[x] >> shift) & 0xFF];
+    }
+
+    last[0] += offset;
+    pointer[0] = offset;
+    for (x=1; x<256; ++x) {
+        pointer[x] = last[x-1];
+        last[x] += last[x-1];
+    }
+
+    for (x=0; x<256; ++x) {
+        while (pointer[x] != last[x]) {
+            value = array[pointer[x]];
+            y = (value >> shift) & 0xFF;
+            while (x != y) {
+                temp = array[pointer[y]];
+                array[pointer[y]++] = value;
+                value = temp;
+                y = (value >> shift) & 0xFF;
+            }
+            array[pointer[x]++] = value;
+        }
+    }
+
+    if (shift > 0) {
+        shift -= 8;
+        for (x=0; x<256; ++x) {
+            temp = x > 0 ? pointer[x] - pointer[x-1] : pointer[0] - offset;
+            if (temp > 64) {
+                radix_sort(array, pointer[x] - temp, pointer[x], shift);
+            } else if (temp > 1) rs_insertsort(array + pointer[x] - temp, array + pointer[x]);
+        }
+    }
+}
+/*************************
+ *** END OF RADIX SORT ***
+ *************************/
+
+template< class _Type, unsigned long PowerOfTwoRadix, unsigned long Log2ofPowerOfTwoRadix, long Threshold >
+inline void _RadixSort_Unsigned_PowerOf2Radix_1( _Type* a, long last, _Type bitMask, unsigned long shiftRightAmount )
+{
+	const unsigned long numberOfBins = PowerOfTwoRadix;
+	unsigned long count[ numberOfBins ];
+	for( unsigned long i = 0; i < numberOfBins; i++ )
+		count[ i ] = 0;
+	for ( long _current = 0; _current <= last; _current++ ) // Scan the array and count the number of times each value appears
+	{
+		unsigned long digit = (unsigned long)(( a[ _current ] & bitMask ) >> shiftRightAmount ); // extract the digit we are sorting based on
+		count[ digit ]++;
+	}
+	long startOfBin[ numberOfBins ], endOfBin[ numberOfBins ], nextBin;
+	startOfBin[ 0 ] = endOfBin[ 0 ] = nextBin = 0;
+	for( unsigned long i = 1; i < numberOfBins; i++ )
+		startOfBin[ i ] = endOfBin[ i ] = startOfBin[ i - 1 ] + count[ i - 1 ];
+	for ( long _current = 0; _current <= last; )
+	{
+		unsigned long digit;
+		_Type tmp = a[ _current ];  // get the compiler to recognize that a register can be used for the loop instead of a[_current] memory location
+		while ( true ) {
+			digit = (unsigned long)(( tmp & bitMask ) >> shiftRightAmount );   // extract the digit we are sorting based on
+			if ( endOfBin[ digit ] == _current )
+				break;
+			_Type tmp2;
+			//_swap( tmp, a[ endOfBin[ digit ] ] );
+			tmp2 = a[endOfBin[digit]]; a[endOfBin[digit]] = tmp; tmp = tmp2;
+			endOfBin[ digit ]++;
+		}
+		a[ _current ] = tmp;
+		endOfBin[ digit ]++;   // leave the element at its location and grow the bin
+		_current++;  // advance the current pointer to the next element
+		while( _current >= startOfBin[ nextBin ] && nextBin < numberOfBins )
+			nextBin++;
+		while( endOfBin[ nextBin - 1 ] == startOfBin[ nextBin ] && nextBin < numberOfBins )
+			nextBin++;
+		if ( _current < endOfBin[ nextBin - 1 ] )
+			_current = endOfBin[ nextBin - 1 ];
+	}
+	bitMask >>= Log2ofPowerOfTwoRadix;
+	if ( bitMask != 0 )   // end recursion when all the bits have been processes
+	{
+		if ( shiftRightAmount >= Log2ofPowerOfTwoRadix ) shiftRightAmount -= Log2ofPowerOfTwoRadix;
+		else shiftRightAmount  = 0;
+		for( unsigned long i = 0; i < numberOfBins; i++ )
+		{
+			long numberOfElements = endOfBin[ i ] - startOfBin[ i ];
+			if ( numberOfElements >= Threshold )  // endOfBin actually points to one beyond the bin
+				_RadixSort_Unsigned_PowerOf2Radix_1< _Type, PowerOfTwoRadix, Log2ofPowerOfTwoRadix, Threshold >( &a[ startOfBin[ i ]], numberOfElements - 1, bitMask, shiftRightAmount );
+			else if ( numberOfElements >= 2 )
+				rs_insertsort(&a[ startOfBin[ i ]], &a[ endOfBin[ i ]]);
+		}
+	}
+}
+inline void RadixSortInPlace_HybridUnsigned_Radix256( unsigned* a, unsigned long a_size )
+{
+	if ( a_size < 2 ) return;
+	unsigned long bitMask = 0xFF000000; // bitMask controls how many bits we process at a time
+	unsigned long shiftRightAmount = 24;
+	if ( a_size >= 32 )
+		_RadixSort_Unsigned_PowerOf2Radix_1<unsigned, 256, 8, 32>(a, a_size - 1, bitMask, shiftRightAmount );
+	else
+		rs_insertsort(a, a + a_size);
+}
+
+struct intcmp_t {
+	inline int operator() (int a, int b) const {
+		return a < b? -1 : a > b? 1 : 0;
+	}
+};
+
+int compare_int(int a, int b)
+{
+	return a < b? -1 : a > b? 1 : 0;
+}
+int compare(const void *a, const void *b)
+{
+	return *((int*)a) - *((int*)b);
+}
+
+int main(int argc, char *argv[])
+{
+	int i, N = 50000000;
+	int *array, *temp;
+	clock_t t1, t2;
+	if (argc == 1) fprintf(stderr, "Usage: %s [%d]\n", argv[0], N);
+	if (argc > 1) N = atoi(argv[1]);
+	temp = (int*)malloc(sizeof(int) * N);
+	array = (int*)malloc(sizeof(int) * N);
+
+	srand48(11);
+	for (i = 0; i < N; ++i) array[i] = (int)lrand48();
+	t1 = clock();
+	rs_sort((unsigned*)array, (unsigned*)array + N, 8, 24);
+	t2 = clock();
+	fprintf(stderr, "radix sort: %.3lf\n", (double)(t2-t1)/CLOCKS_PER_SEC);
+	for (i = 0; i < N-1; ++i) {
+		if (array[i] > array[i+1]) {
+			fprintf(stderr, "Bug in radix sort!\n");
+			exit(1);
+		}
+	}
+	t1 = clock();
+	rs_sort((unsigned*)array, (unsigned*)array + N, 8, 24);
+	t2 = clock();
+	fprintf(stderr, "radix sort (sorted): %.3lf\n", (double)(t2-t1)/CLOCKS_PER_SEC);
+
+	srand48(11);
+	for (i = 0; i < N; ++i) array[i] = (int)lrand48();
+	t1 = clock();
+	RadixSortInPlace_HybridUnsigned_Radix256((unsigned*)array, N);
+//	radix_sort((unsigned*)array, 0, N, 24);
+	t2 = clock();
+	fprintf(stderr, "vd's radix sort: %.3lf\n", (double)(t2-t1)/CLOCKS_PER_SEC);
+	for (i = 0; i < N-1; ++i) {
+		if (array[i] > array[i+1]) {
+			fprintf(stderr, "Bug in radix sort!\n");
+			exit(1);
+		}
+	}
+	t1 = clock();
+	RadixSortInPlace_HybridUnsigned_Radix256((unsigned*)array, N);
+//	radix_sort((unsigned*)array, 0, N, 24);
+	t2 = clock();
+	fprintf(stderr, "vd's radix sort (sorted): %.3lf\n", (double)(t2-t1)/CLOCKS_PER_SEC);
+
+	srand48(11);
+	for (i = 0; i < N; ++i) array[i] = (int)lrand48();
+	t1 = clock();
+	sort(array, array+N);
+	t2 = clock();
+	fprintf(stderr, "STL introsort: %.3lf\n", (double)(t2-t1)/CLOCKS_PER_SEC);
+	t1 = clock();
+	sort(array, array+N);
+	t2 = clock();
+	fprintf(stderr, "STL introsort (sorted): %.3lf\n", (double)(t2-t1)/CLOCKS_PER_SEC);
+
+	srand48(11);
+	for (i = 0; i < N; ++i) array[i] = (int)lrand48();
+	t1 = clock();
+	stable_sort(array, array+N);
+	t2 = clock();
+	fprintf(stderr, "STL stablesort: %.3lf\n", (double)(t2-t1)/CLOCKS_PER_SEC);
+	t1 = clock();
+	stable_sort(array, array+N);
+	t2 = clock();
+	fprintf(stderr, "STL stablesort (sorted): %.3lf\n", (double)(t2-t1)/CLOCKS_PER_SEC);
+
+	srand48(11);
+	for (i = 0; i < N; ++i) array[i] = (int)lrand48();
+	t1 = clock();
+	make_heap(array, array+N);
+	sort_heap(array, array+N);
+	t2 = clock();
+	fprintf(stderr, "STL heapsort: %.3lf\n", (double)(t2-t1)/CLOCKS_PER_SEC);
+	for (i = 0; i < N-1; ++i) {
+		if (array[i] > array[i+1]) {
+			fprintf(stderr, "Bug in heap_sort!\n");
+			exit(1);
+		}
+	}
+	t1 = clock();
+	make_heap(array, array+N);
+	sort_heap(array, array+N);
+	t2 = clock();
+	fprintf(stderr, "STL heapsort (sorted): %.3lf\n", (double)(t2-t1)/CLOCKS_PER_SEC);
+
+	srand48(11);
+	for (i = 0; i < N; ++i) array[i] = (int)lrand48();
+	t1 = clock();
+	ks_combsort(int, N, array);
+	t2 = clock();
+	fprintf(stderr, "combsort: %.3lf\n", (double)(t2-t1)/CLOCKS_PER_SEC);
+	for (i = 0; i < N-1; ++i) {
+		if (array[i] > array[i+1]) {
+			fprintf(stderr, "Bug in combsort!\n");
+			exit(1);
+		}
+	}
+
+	srand48(11);
+	for (i = 0; i < N; ++i) array[i] = (int)lrand48();
+	t1 = clock();
+	qsort(array, N, sizeof(int), compare);
+	t2 = clock();
+	fprintf(stderr, "libc qsort: %.3lf\n", (double)(t2-t1)/CLOCKS_PER_SEC);
+
+	srand48(11);
+	for (i = 0; i < N; ++i) array[i] = (int)lrand48();
+	t1 = clock();
+	ks_introsort(int, N, array);
+	t2 = clock();
+	fprintf(stderr, "my introsort: %.3lf\n", (double)(t2-t1)/CLOCKS_PER_SEC);
+	for (i = 0; i < N-1; ++i) {
+		if (array[i] > array[i+1]) {
+			fprintf(stderr, "Bug in intro_sort!\n");
+			exit(1);
+		}
+	}
+	t1 = clock();
+	ks_introsort(int, N, array);
+	t2 = clock();
+	fprintf(stderr, "introsort (sorted): %.3lf\n", (double)(t2-t1)/CLOCKS_PER_SEC);
+
+	srand48(11);
+	for (i = 0; i < N; ++i) array[i] = (int)lrand48();
+	t1 = clock();
+	ks_mergesort(int, N, array, 0);
+	t2 = clock();
+	fprintf(stderr, "iterative mergesort: %.3lf\n", (double)(t2-t1)/CLOCKS_PER_SEC);
+	for (i = 0; i < N-1; ++i) {
+		if (array[i] > array[i+1]) {
+			fprintf(stderr, "Bug in merge_sort!\n");
+			exit(1);
+		}
+	}
+	t1 = clock();
+	ks_mergesort(int, N, array, 0);
+	t2 = clock();
+	fprintf(stderr, "iterative mergesort (sorted): %.3lf\n", (double)(t2-t1)/CLOCKS_PER_SEC);
+
+	srand48(11);
+	for (i = 0; i < N; ++i) array[i] = (int)lrand48();
+	t1 = clock();
+	ks_heapmake(int, N, array);
+	ks_heapsort(int, N, array);
+	t2 = clock();
+	fprintf(stderr, "my heapsort: %.3lf\n", (double)(t2-t1)/CLOCKS_PER_SEC);
+	for (i = 0; i < N-1; ++i) {
+		if (array[i] > array[i+1]) {
+			fprintf(stderr, "Bug in heap_sort!\n");
+			exit(1);
+		}
+	}
+	t1 = clock();
+	ks_heapmake(int, N, array);
+	ks_heapsort(int, N, array);
+	t2 = clock();
+	fprintf(stderr, "heapsort (sorted): %.3lf\n", (double)(t2-t1)/CLOCKS_PER_SEC);
+
+	srand48(11);
+	for (i = 0; i < N; ++i) array[i] = (int)lrand48();
+	t1 = clock();
+	heapSort(array, N);
+	t2 = clock();
+	fprintf(stderr, "Paul's heapsort: %.3lf\n", (double)(t2-t1)/CLOCKS_PER_SEC);
+	for (i = 0; i < N-1; ++i) {
+		if (array[i] > array[i+1]) {
+			fprintf(stderr, "Bug in intro_sort!\n");
+			exit(1);
+		}
+	}
+
+	srand48(11);
+	for (i = 0; i < N; ++i) array[i] = (int)lrand48();
+	t1 = clock();
+	quickSort(array, N);
+	t2 = clock();
+	fprintf(stderr, "Paul's quicksort: %.3lf\n", (double)(t2-t1)/CLOCKS_PER_SEC);
+	for (i = 0; i < N-1; ++i) {
+		if (array[i] > array[i+1]) {
+			fprintf(stderr, "Bug in intro_sort!\n");
+			exit(1);
+		}
+	}
+
+	srand48(11);
+	for (i = 0; i < N; ++i) array[i] = (int)lrand48();
+	t1 = clock();
+	mergeSort(array, N);
+	t2 = clock();
+	fprintf(stderr, "Paul's mergesort: %.3lf\n", (double)(t2-t1)/CLOCKS_PER_SEC);
+	for (i = 0; i < N-1; ++i) {
+		if (array[i] > array[i+1]) {
+			fprintf(stderr, "Bug in intro_sort!\n");
+			exit(1);
+		}
+	}
+
+	free(array); free(temp);
+	return 0;
+}
diff --git a/ext/klib/test/kstring_bench.c b/ext/klib/test/kstring_bench.c
new file mode 100644
index 0000000..82598e8
--- /dev/null
+++ b/ext/klib/test/kstring_bench.c
@@ -0,0 +1,51 @@
+#include <stdio.h>
+#include <stdlib.h>
+#include <time.h>
+#include "kstring.h"
+
+#define N 10000000
+
+int main()
+{
+	int i;
+	clock_t t;
+	kstring_t s, s2;
+	srand48(11);
+	s.l = s.m = 0; s.s = 0;
+	t = clock();
+	for (i = 0; i < N; ++i) {
+		int x = lrand48();
+		s.l = 0;
+		kputw(x, &s);
+	}
+	fprintf(stderr, "kputw: %lf\n", (double)(clock() - t) / CLOCKS_PER_SEC);
+	srand48(11);
+	t = clock();
+	for (i = 0; i < N; ++i) {
+		int x = lrand48();
+		s.l = 0;
+		ksprintf(&s, "%d", x);
+	}
+	fprintf(stderr, "ksprintf: %lf\n", (double)(clock() - t) / CLOCKS_PER_SEC);
+
+	srand48(11);
+	s2.l = s2.m = 0; s2.s = 0;
+	t = clock();
+	for (i = 0; i < N; ++i) {
+		int x = lrand48();
+		s2.l = s.l = 0;
+		kputw(x, &s2);
+		kputs(s2.s, &s);
+	}
+	fprintf(stderr, "kputw+kputs: %lf\n", (double)(clock() - t) / CLOCKS_PER_SEC);
+	srand48(11);
+	t = clock();
+	for (i = 0; i < N; ++i) {
+		int x = lrand48();
+		s2.l = s.l = 0;
+		kputw(x, &s2);
+		ksprintf(&s, "%s", s2.s);
+	}
+	fprintf(stderr, "kputw+ksprintf: %lf\n", (double)(clock() - t) / CLOCKS_PER_SEC);
+	return 0;
+}
diff --git a/ext/klib/test/kstring_bench2.c b/ext/klib/test/kstring_bench2.c
new file mode 100644
index 0000000..b7707a8
--- /dev/null
+++ b/ext/klib/test/kstring_bench2.c
@@ -0,0 +1,131 @@
+#include <string.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <time.h>
+#include "kstring.h"
+
+#ifdef __APPLE__
+#define HAVE_STRNSTR
+#endif
+
+#ifdef __linux__
+#define HAVE_MEMMEM
+#endif
+
+static int str_len = 1024*1024*128;
+static int pat_len = 30;
+static int alphabet = 2;
+static int repeat = 50;
+
+char *gen_data(int len, int a)
+{
+	char *data;
+	int i;
+	long x;
+	srand48(11);
+	data = malloc(len);
+	for (i = 0; i < len; ++i)
+		data[i] = (int)(a * drand48()) + '!';
+	data[str_len - 1] = 0;
+	return data;
+}
+// http://srcvault.scali.eu.org/cgi-bin/Syntax/c/BoyerMoore.c
+char *BoyerMoore( unsigned char *data, unsigned int dataLength, unsigned char *string, unsigned int strLength )
+{
+	unsigned int skipTable[256], i;
+	unsigned char *search;
+	register unsigned char lastChar;
+
+	if (strLength == 0)
+		return NULL;
+
+	for (i = 0; i < 256; i++)
+		skipTable[i] = strLength;
+	search = string;
+	i = --strLength;
+	do {
+		skipTable[*search++] = i;
+	} while (i--);
+	lastChar = *--search;
+	search = data + strLength;
+	dataLength -= strLength+(strLength-1);
+	while ((int)dataLength > 0 ) {
+		unsigned int skip;
+		skip = skipTable[*search];
+		search += skip;
+		dataLength -= skip;
+		skip = skipTable[*search];
+		search += skip;
+		dataLength -= skip;
+		skip = skipTable[*search];
+		if (*search != lastChar) {
+			search += skip;
+			dataLength -= skip;
+			continue;
+		}
+		i = strLength;
+		do {
+			if (i-- == 0) return search;
+		} while (*--search == string[i]);
+		search += (strLength - i + 1);
+		dataLength--;
+	}
+	return NULL;
+}
+
+int main()
+{
+	char *data;
+	int i;
+	clock_t t;
+	t = clock();
+	data = gen_data(str_len, alphabet);
+	fprintf(stderr, "Generate data in %.3f sec\n", (float)(clock() - t) / CLOCKS_PER_SEC);
+	{
+		t = clock(); srand48(1331);
+		for (i = 0; i < repeat; ++i) {
+			int y = lrand48() % (str_len - pat_len);
+			char *ret;
+			ret = kmemmem(data, str_len, data + y, pat_len, 0);
+//			printf("%d, %d\n", (int)(ret - data), y);
+		}
+		fprintf(stderr, "Search patterns in %.3f sec\n", (float)(clock() - t) / CLOCKS_PER_SEC);
+	}
+	if (1) {
+		t = clock(); srand48(1331);
+		for (i = 0; i < repeat; ++i) {
+			int y = lrand48() % (str_len - pat_len);
+			char *ret;
+			ret = BoyerMoore(data, str_len, data + y, pat_len);
+//			printf("%d, %d\n", (int)(ret - data), y);
+		}
+		fprintf(stderr, "Search patterns in %.3f sec\n", (float)(clock() - t) / CLOCKS_PER_SEC);
+	}
+#ifdef HAVE_STRNSTR
+	if (1) {
+		char *tmp;
+		t = clock(); srand48(1331);
+		tmp = calloc(pat_len+1, 1);
+		for (i = 0; i < repeat; ++i) {
+			int y = lrand48() % (str_len - pat_len);
+			char *ret;
+			memcpy(tmp, data + y, pat_len);
+			ret = strnstr(data, tmp, str_len);
+		}
+		fprintf(stderr, "Search patterns in %.3f sec\n", (float)(clock() - t) / CLOCKS_PER_SEC);		
+	}
+#endif
+#ifdef HAVE_MEMMEM
+	if (1) {
+		t = clock(); srand48(1331);
+		for (i = 0; i < repeat; ++i) {
+			int y = lrand48() % (str_len - pat_len);
+			char *ret;
+			ret = memmem(data, str_len, data + y, pat_len);
+//			printf("%d, %d\n", (int)(ret - data), y);
+		}
+		fprintf(stderr, "Search patterns in %.3f sec\n", (float)(clock() - t) / CLOCKS_PER_SEC);
+	}
+#endif
+	return 0;
+}
diff --git a/ext/klib/test/kstring_test.c b/ext/klib/test/kstring_test.c
new file mode 100644
index 0000000..9f9b6e6
--- /dev/null
+++ b/ext/klib/test/kstring_test.c
@@ -0,0 +1,132 @@
+#include <limits.h>
+#include <stdarg.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "kstring.h"
+
+int nfail = 0;
+
+void check(const char *what, const kstring_t *ks, const char *correct)
+{
+	if (ks->l != strlen(correct) || strcmp(ks->s, correct) != 0) {
+		fprintf(stderr, "%s produced \"%.*s\" (\"%s\" is correct)\tFAIL\n", what, (int)(ks->l), ks->s, correct);
+		nfail++;
+	}
+}
+
+void test_kputw(kstring_t *ks, int n)
+{
+	char buf[16];
+
+	ks->l = 0;
+	kputw(n, ks);
+
+	sprintf(buf, "%d", n);
+	check("kputw()", ks, buf);
+}
+
+void test_kputl(kstring_t *ks, long n)
+{
+	char buf[24];
+
+	ks->l = 0;
+	kputl(n, ks);
+
+	sprintf(buf, "%ld", n);
+	check("kputl()", ks, buf);
+}
+
+static char *mem_gets(char *buf, int buflen, void *vtextp)
+{
+	const char **textp = (const char **) vtextp;
+
+	const char *nl = strchr(*textp, '\n');
+	size_t n = nl? nl - *textp + 1 : strlen(*textp);
+
+	if (n == 0) return NULL;
+
+	if (n > buflen-1) n = buflen-1;
+	memcpy(buf, *textp, n);
+	buf[n] = '\0';
+	*textp += n;
+	return buf;
+}
+
+void test_kgetline(kstring_t *ks, const char *text, ...)
+{
+	const char *exp;
+	va_list arg;
+
+	va_start(arg, text);
+	while ((exp = va_arg(arg, const char *)) != NULL) {
+		ks->l = 0;
+		if (kgetline(ks, mem_gets, &text) != 0) kputs("EOF", ks);
+		check("kgetline()", ks, exp);
+	}
+	va_end(arg);
+
+	ks->l = 0;
+	if (kgetline(ks, mem_gets, &text) == 0) check("kgetline()", ks, "EOF");
+}
+
+int main(int argc, char **argv)
+{
+	kstring_t ks;
+
+	ks.l = ks.m = 0;
+	ks.s = NULL;
+
+	test_kputw(&ks, 0);
+	test_kputw(&ks, 1);
+	test_kputw(&ks, 37);
+	test_kputw(&ks, 12345);
+	test_kputw(&ks, -12345);
+	test_kputw(&ks, INT_MAX);
+	test_kputw(&ks, -INT_MAX);
+	test_kputw(&ks, INT_MIN);
+
+	test_kputl(&ks, 0);
+	test_kputl(&ks, 1);
+	test_kputl(&ks, 37);
+	test_kputl(&ks, 12345);
+	test_kputl(&ks, -12345);
+	test_kputl(&ks, INT_MAX);
+	test_kputl(&ks, -INT_MAX);
+	test_kputl(&ks, INT_MIN);
+	test_kputl(&ks, LONG_MAX);
+	test_kputl(&ks, -LONG_MAX);
+	test_kputl(&ks, LONG_MIN);
+
+	test_kgetline(&ks, "", NULL);
+	test_kgetline(&ks, "apple", "apple", NULL);
+	test_kgetline(&ks, "banana\n", "banana", NULL);
+	test_kgetline(&ks, "carrot\r\n", "carrot", NULL);
+	test_kgetline(&ks, "\n", "", NULL);
+	test_kgetline(&ks, "\n\n", "", "", NULL);
+	test_kgetline(&ks, "foo\nbar", "foo", "bar", NULL);
+	test_kgetline(&ks, "foo\nbar\n", "foo", "bar", NULL);
+	test_kgetline(&ks,
+		"abcdefghijklmnopqrstuvwxyz0123456789\nABCDEFGHIJKLMNOPQRSTUVWXYZ\n",
+		"abcdefghijklmnopqrstuvwxyz0123456789",
+		"ABCDEFGHIJKLMNOPQRSTUVWXYZ", NULL);
+
+	if (argc > 1) {
+		FILE *f = fopen(argv[1], "r");
+		if (f) {
+			for (ks.l = 0; kgetline(&ks, (kgets_func *)fgets, f) == 0; ks.l = 0)
+				puts(ks.s);
+			fclose(f);
+		}
+	}
+
+	free(ks.s);
+
+	if (nfail > 0) {
+		fprintf(stderr, "Total failures: %d\n", nfail);
+		return EXIT_FAILURE;
+	}
+
+	return EXIT_SUCCESS;
+}
diff --git a/ext/klib/test/kthread_test.c b/ext/klib/test/kthread_test.c
new file mode 100644
index 0000000..1b67ed4
--- /dev/null
+++ b/ext/klib/test/kthread_test.c
@@ -0,0 +1,69 @@
+#include <stdio.h>
+#include <assert.h>
+#include <stdlib.h>
+#include <pthread.h>
+#if HAVE_CILK
+#include <cilk/cilk.h>
+#include <cilk/cilk_api.h>
+#endif
+
+typedef struct {
+	int max_iter, w, h;
+	double xmin, xmax, ymin, ymax;
+	int *k;
+} global_t;
+
+static void compute(void *_g, int i, int tid)
+{
+	global_t *g = (global_t*)_g;
+	double x, x0 = g->xmin + (g->xmax - g->xmin) * (i%g->w) / g->w;
+	double y, y0 = g->ymin + (g->ymax - g->ymin) * (i/g->w) / g->h;
+	int k;
+
+	assert(g->k[i] < 0);
+	x = x0, y = y0;
+	for (k = 0; k < g->max_iter; ++k) {
+		double z = x * y;
+		x *= x; y *= y;
+		if (x + y >= 4) break;
+		x = x - y + x0;
+		y = z + z + y0; 
+	}
+	g->k[i] = k;
+}
+
+void kt_for(int n_threads, int n_items, void (*func)(void*,int,int), void *data);
+
+int main(int argc, char *argv[])
+{
+	int i, tmp, tot, type = 0, n_threads = 2;
+	global_t global = { 10240*100, 800, 600, -2., -1.2, -1.2, 1.2, 0 };
+//	global_t global = { 10240*1, 8, 6, -2., -1.2, -1.2, 1.2, 0 };
+
+	if (argc > 1) {
+		type = argv[1][0] == 'o'? 2 : argv[1][0] == 'c'? 3 : argv[1][0] == 'n'? 1 : 0;
+		if (argv[1][0] >= '0' && argv[1][0] <= '9')
+			n_threads = atoi(argv[1]);
+	} else {
+		fprintf(stderr, "Usage: ./a.out [openmp | cilk | #threads]\n");
+	}
+	tot = global.w * global.h;
+	global.k = calloc(tot, sizeof(int));
+	for (i = 0; i < tot; ++i) global.k[i] = -1;
+	if (type == 0) {
+		kt_for(n_threads, tot, compute, &global);
+	} else if (type == 2) {
+		#pragma omp parallel for
+		for (i = 0; i < tot; ++i)
+			compute(&global, i, 0);
+	} else if (type == 3) {
+		#if HAVE_CILK
+		cilk_for (i = 0; i < tot; ++i)
+			compute(&global, i, 0);
+		#endif
+	}
+	for (i = tmp = 0; i < tot; ++i) tmp += (global.k[i] < 0);
+	free(global.k);
+	assert(tmp == 0);
+	return 0;
+}
diff --git a/ext/klib/test/kthread_test2.c b/ext/klib/test/kthread_test2.c
new file mode 100644
index 0000000..68663a5
--- /dev/null
+++ b/ext/klib/test/kthread_test2.c
@@ -0,0 +1,80 @@
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <assert.h>
+
+void kt_for(int n_threads, void (*func)(void*,long,int), void *data, long n);
+void kt_pipeline(int n_threads, void *(*func)(void*, int, void*), void *shared_data, int n_steps);
+
+typedef struct {
+	FILE *fp;
+	int max_lines, buf_size, n_threads;
+	char *buf;
+} pipeline_t;
+
+typedef struct {
+	int n_lines;
+	char **lines;
+} step_t;
+
+static void worker_for(void *_data, long i, int tid) // kt_for() callback
+{
+	step_t *step = (step_t*)_data;
+	char *s = step->lines[i];
+	int t, l, j;
+	l = strlen(s) - 1;
+	assert(s[l] == '\n'); // not supporting long lines
+	for (j = 0; j < l>>1; ++j)
+		t = s[j], s[j] = s[l - 1 - j], s[l - 1 - j] = t;
+}
+
+static void *worker_pipeline(void *shared, int step, void *in) // kt_pipeline() callback
+{
+	pipeline_t *p = (pipeline_t*)shared;
+	if (step == 0) { // step 0: read lines into the buffer
+		step_t *s;
+		s = calloc(1, sizeof(step_t));
+		s->lines = calloc(p->max_lines, sizeof(char*));
+		while (fgets(p->buf, p->buf_size, p->fp) != 0) {
+			s->lines[s->n_lines] = strdup(p->buf);
+			if (++s->n_lines >= p->max_lines)
+				break;
+		}
+		if (s->n_lines) return s;
+	} else if (step == 1) { // step 1: reverse lines
+		kt_for(p->n_threads, worker_for, in, ((step_t*)in)->n_lines);
+		return in;
+	} else if (step == 2) { // step 3: write the buffer to output
+		step_t *s = (step_t*)in;
+		while (s->n_lines > 0) {
+			fputs(s->lines[--s->n_lines], stdout);
+			free(s->lines[s->n_lines]);
+		}
+		free(s->lines); free(s);
+	}
+	return 0;
+}
+
+int main(int argc, char *argv[])
+{
+	pipeline_t pl;
+	int pl_threads;
+	if (argc == 1) {
+		fprintf(stderr, "Usage: reverse <in.txt> [pipeline_threads [for_threads]]\n");
+		return 1;
+	}
+	pl.fp = strcmp(argv[1], "-")? fopen(argv[1], "r") : stdin;
+	if (pl.fp == 0) {
+		fprintf(stderr, "ERROR: failed to open the input file.\n");
+		return 1;
+	}
+	pl_threads = argc > 2? atoi(argv[2]) : 3;
+	pl.max_lines = 4096;
+	pl.buf_size = 0x10000;
+	pl.n_threads = argc > 3? atoi(argv[3]) : 1;
+	pl.buf = calloc(pl.buf_size, 1);
+	kt_pipeline(pl_threads, worker_pipeline, &pl, 3);
+	free(pl.buf);
+	if (pl.fp != stdin) fclose(pl.fp);
+	return 0;
+}
diff --git a/ext/klib/test/kvec_test.cc b/ext/klib/test/kvec_test.cc
new file mode 100644
index 0000000..1015574
--- /dev/null
+++ b/ext/klib/test/kvec_test.cc
@@ -0,0 +1,69 @@
+#include <vector>
+#include <time.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include "kvec.h"
+
+int main()
+{
+	int M = 10, N = 20000000, i, j;
+	clock_t t;
+	t = clock();
+	for (i = 0; i < M; ++i) {
+		int *array = (int*)malloc(N * sizeof(int));
+		for (j = 0; j < N; ++j) array[j] = j;
+		free(array);
+	}
+	printf("C array, preallocated: %.3f sec\n",
+		   (float)(clock() - t) / CLOCKS_PER_SEC);
+	t = clock();
+	for (i = 0; i < M; ++i) {
+		int *array = 0, max = 0;
+		for (j = 0; j < N; ++j) {
+			if (j == max) {
+				max = !max? 1 : max << 1;
+				array = (int*)realloc(array, sizeof(int)*max);
+			}
+			array[j] = j;
+		}
+		free(array);
+	}
+	printf("C array, dynamic: %.3f sec\n",
+		   (float)(clock() - t) / CLOCKS_PER_SEC);
+	t = clock();
+	for (i = 0; i < M; ++i) {
+		kvec_t(int) array;
+		kv_init(array);
+		kv_resize(int, array, N);
+		for (j = 0; j < N; ++j) kv_a(int, array, j) = j;
+		kv_destroy(array);
+	}
+	printf("C vector, dynamic(kv_a): %.3f sec\n",
+		   (float)(clock() - t) / CLOCKS_PER_SEC);
+	t = clock();
+	for (i = 0; i < M; ++i) {
+		kvec_t(int) array;
+		kv_init(array);
+		for (j = 0; j < N; ++j)
+			kv_push(int, array, j);
+		kv_destroy(array);
+	}
+	printf("C vector, dynamic(kv_push): %.3f sec\n",
+		   (float)(clock() - t) / CLOCKS_PER_SEC);
+	t = clock();
+	for (i = 0; i < M; ++i) {
+		std::vector<int> array;
+		array.reserve(N);
+		for (j = 0; j < N; ++j) array[j] = j;
+	}
+	printf("C++ vector, preallocated: %.3f sec\n",
+		   (float)(clock() - t) / CLOCKS_PER_SEC);
+	t = clock();
+	for (i = 0; i < M; ++i) {
+		std::vector<int> array;
+		for (j = 0; j < N; ++j) array.push_back(j);
+	}
+	printf("C++ vector, dynamic: %.3f sec\n",
+		   (float)(clock() - t) / CLOCKS_PER_SEC);
+	return 0;
+}
diff --git a/ext/robin-map/robin_growth_policy.h b/ext/robin-map/robin_growth_policy.h
new file mode 100644
index 0000000..b100bfc
--- /dev/null
+++ b/ext/robin-map/robin_growth_policy.h
@@ -0,0 +1,416 @@
+/**
+ * MIT License
+ *
+ * Copyright (c) 2017 Thibaut Goetghebuer-Planchon <tessil@gmx.com>
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#ifndef TSL_ROBIN_GROWTH_POLICY_H
+#define TSL_ROBIN_GROWTH_POLICY_H
+
+#include <algorithm>
+#include <array>
+#include <climits>
+#include <cmath>
+#include <cstddef>
+#include <cstdint>
+#include <iterator>
+#include <limits>
+#include <ratio>
+#include <stdexcept>
+
+// A change of the major version indicates an API and/or ABI break (change of
+// in-memory layout of the data structure)
+#define TSL_RH_VERSION_MAJOR 1
+// A change of the minor version indicates the addition of a feature without
+// impact on the API/ABI
+#define TSL_RH_VERSION_MINOR 3
+// A change of the patch version indicates a bugfix without additional
+// functionality
+#define TSL_RH_VERSION_PATCH 0
+
+#ifdef TSL_DEBUG
+#define tsl_rh_assert(expr) assert(expr)
+#else
+#define tsl_rh_assert(expr) (static_cast<void>(0))
+#endif
+
+/**
+ * If exceptions are enabled, throw the exception passed in parameter, otherwise
+ * call std::terminate.
+ */
+#if (defined(__cpp_exceptions) || defined(__EXCEPTIONS) || \
+     (defined(_MSC_VER) && defined(_CPPUNWIND))) &&        \
+    !defined(TSL_NO_EXCEPTIONS)
+#define TSL_RH_THROW_OR_TERMINATE(ex, msg) throw ex(msg)
+#else
+#define TSL_RH_NO_EXCEPTIONS
+#ifdef TSL_DEBUG
+#include <iostream>
+#define TSL_RH_THROW_OR_TERMINATE(ex, msg) \
+  do {                                     \
+    std::cerr << msg << std::endl;         \
+    std::terminate();                      \
+  } while (0)
+#else
+#define TSL_RH_THROW_OR_TERMINATE(ex, msg) std::terminate()
+#endif
+#endif
+
+#if defined(__GNUC__) || defined(__clang__)
+#define TSL_RH_LIKELY(exp) (__builtin_expect(!!(exp), true))
+#else
+#define TSL_RH_LIKELY(exp) (exp)
+#endif
+
+#define TSL_RH_UNUSED(x) static_cast<void>(x)
+
+namespace tsl {
+namespace rh {
+
+/**
+ * Grow the hash table by a factor of GrowthFactor keeping the bucket count to a
+ * power of two. It allows the table to use a mask operation instead of a modulo
+ * operation to map a hash to a bucket.
+ *
+ * GrowthFactor must be a power of two >= 2.
+ */
+template <std::size_t GrowthFactor>
+class power_of_two_growth_policy {
+ public:
+  /**
+   * Called on the hash table creation and on rehash. The number of buckets for
+   * the table is passed in parameter. This number is a minimum, the policy may
+   * update this value with a higher value if needed (but not lower).
+   *
+   * If 0 is given, min_bucket_count_in_out must still be 0 after the policy
+   * creation and bucket_for_hash must always return 0 in this case.
+   */
+  explicit power_of_two_growth_policy(std::size_t& min_bucket_count_in_out) {
+    if (min_bucket_count_in_out > max_bucket_count()) {
+      TSL_RH_THROW_OR_TERMINATE(std::length_error,
+                                "The hash table exceeds its maximum size.");
+    }
+
+    if (min_bucket_count_in_out > 0) {
+      min_bucket_count_in_out =
+          round_up_to_power_of_two(min_bucket_count_in_out);
+      m_mask = min_bucket_count_in_out - 1;
+    } else {
+      m_mask = 0;
+    }
+  }
+
+  /**
+   * Return the bucket [0, bucket_count()) to which the hash belongs.
+   * If bucket_count() is 0, it must always return 0.
+   */
+  std::size_t bucket_for_hash(std::size_t hash) const noexcept {
+    return hash & m_mask;
+  }
+
+  /**
+   * Return the number of buckets that should be used on next growth.
+   */
+  std::size_t next_bucket_count() const {
+    if ((m_mask + 1) > max_bucket_count() / GrowthFactor) {
+      TSL_RH_THROW_OR_TERMINATE(std::length_error,
+                                "The hash table exceeds its maximum size.");
+    }
+
+    return (m_mask + 1) * GrowthFactor;
+  }
+
+  /**
+   * Return the maximum number of buckets supported by the policy.
+   */
+  std::size_t max_bucket_count() const {
+    // Largest power of two.
+    return (std::numeric_limits<std::size_t>::max() / 2) + 1;
+  }
+
+  /**
+   * Reset the growth policy as if it was created with a bucket count of 0.
+   * After a clear, the policy must always return 0 when bucket_for_hash is
+   * called.
+   */
+  void clear() noexcept { m_mask = 0; }
+
+ private:
+  static std::size_t round_up_to_power_of_two(std::size_t value) {
+    if (is_power_of_two(value)) {
+      return value;
+    }
+
+    if (value == 0) {
+      return 1;
+    }
+
+    --value;
+    for (std::size_t i = 1; i < sizeof(std::size_t) * CHAR_BIT; i *= 2) {
+      value |= value >> i;
+    }
+
+    return value + 1;
+  }
+
+  static constexpr bool is_power_of_two(std::size_t value) {
+    return value != 0 && (value & (value - 1)) == 0;
+  }
+
+ protected:
+  static_assert(is_power_of_two(GrowthFactor) && GrowthFactor >= 2,
+                "GrowthFactor must be a power of two >= 2.");
+
+  std::size_t m_mask;
+};
+
+/**
+ * Grow the hash table by GrowthFactor::num / GrowthFactor::den and use a modulo
+ * to map a hash to a bucket. Slower but it can be useful if you want a slower
+ * growth.
+ */
+template <class GrowthFactor = std::ratio<3, 2>>
+class mod_growth_policy {
+ public:
+  explicit mod_growth_policy(std::size_t& min_bucket_count_in_out) {
+    if (min_bucket_count_in_out > max_bucket_count()) {
+      TSL_RH_THROW_OR_TERMINATE(std::length_error,
+                                "The hash table exceeds its maximum size.");
+    }
+
+    if (min_bucket_count_in_out > 0) {
+      m_mod = min_bucket_count_in_out;
+    } else {
+      m_mod = 1;
+    }
+  }
+
+  std::size_t bucket_for_hash(std::size_t hash) const noexcept {
+    return hash % m_mod;
+  }
+
+  std::size_t next_bucket_count() const {
+    if (m_mod == max_bucket_count()) {
+      TSL_RH_THROW_OR_TERMINATE(std::length_error,
+                                "The hash table exceeds its maximum size.");
+    }
+
+    const double next_bucket_count =
+        std::ceil(double(m_mod) * REHASH_SIZE_MULTIPLICATION_FACTOR);
+    if (!std::isnormal(next_bucket_count)) {
+      TSL_RH_THROW_OR_TERMINATE(std::length_error,
+                                "The hash table exceeds its maximum size.");
+    }
+
+    if (next_bucket_count > double(max_bucket_count())) {
+      return max_bucket_count();
+    } else {
+      return std::size_t(next_bucket_count);
+    }
+  }
+
+  std::size_t max_bucket_count() const { return MAX_BUCKET_COUNT; }
+
+  void clear() noexcept { m_mod = 1; }
+
+ private:
+  static constexpr double REHASH_SIZE_MULTIPLICATION_FACTOR =
+      1.0 * GrowthFactor::num / GrowthFactor::den;
+  static const std::size_t MAX_BUCKET_COUNT =
+      std::size_t(double(std::numeric_limits<std::size_t>::max() /
+                         REHASH_SIZE_MULTIPLICATION_FACTOR));
+
+  static_assert(REHASH_SIZE_MULTIPLICATION_FACTOR >= 1.1,
+                "Growth factor should be >= 1.1.");
+
+  std::size_t m_mod;
+};
+
+namespace detail {
+
+#if SIZE_MAX >= ULLONG_MAX
+#define TSL_RH_NB_PRIMES 51
+#elif SIZE_MAX >= ULONG_MAX
+#define TSL_RH_NB_PRIMES 40
+#else
+#define TSL_RH_NB_PRIMES 23
+#endif
+
+static constexpr const std::array<std::size_t, TSL_RH_NB_PRIMES> PRIMES = {{
+    1u,
+    5u,
+    17u,
+    29u,
+    37u,
+    53u,
+    67u,
+    79u,
+    97u,
+    131u,
+    193u,
+    257u,
+    389u,
+    521u,
+    769u,
+    1031u,
+    1543u,
+    2053u,
+    3079u,
+    6151u,
+    12289u,
+    24593u,
+    49157u,
+#if SIZE_MAX >= ULONG_MAX
+    98317ul,
+    196613ul,
+    393241ul,
+    786433ul,
+    1572869ul,
+    3145739ul,
+    6291469ul,
+    12582917ul,
+    25165843ul,
+    50331653ul,
+    100663319ul,
+    201326611ul,
+    402653189ul,
+    805306457ul,
+    1610612741ul,
+    3221225473ul,
+    4294967291ul,
+#endif
+#if SIZE_MAX >= ULLONG_MAX
+    6442450939ull,
+    12884901893ull,
+    25769803751ull,
+    51539607551ull,
+    103079215111ull,
+    206158430209ull,
+    412316860441ull,
+    824633720831ull,
+    1649267441651ull,
+    3298534883309ull,
+    6597069766657ull,
+#endif
+}};
+
+template <unsigned int IPrime>
+static constexpr std::size_t mod(std::size_t hash) {
+  return hash % PRIMES[IPrime];
+}
+
+// MOD_PRIME[iprime](hash) returns hash % PRIMES[iprime]. This table allows for
+// faster modulo as the compiler can optimize the modulo code better with a
+// constant known at the compilation.
+static constexpr const std::array<std::size_t (*)(std::size_t),
+                                  TSL_RH_NB_PRIMES>
+    MOD_PRIME = {{
+        &mod<0>,  &mod<1>,  &mod<2>,  &mod<3>,  &mod<4>,  &mod<5>,
+        &mod<6>,  &mod<7>,  &mod<8>,  &mod<9>,  &mod<10>, &mod<11>,
+        &mod<12>, &mod<13>, &mod<14>, &mod<15>, &mod<16>, &mod<17>,
+        &mod<18>, &mod<19>, &mod<20>, &mod<21>, &mod<22>,
+#if SIZE_MAX >= ULONG_MAX
+        &mod<23>, &mod<24>, &mod<25>, &mod<26>, &mod<27>, &mod<28>,
+        &mod<29>, &mod<30>, &mod<31>, &mod<32>, &mod<33>, &mod<34>,
+        &mod<35>, &mod<36>, &mod<37>, &mod<38>, &mod<39>,
+#endif
+#if SIZE_MAX >= ULLONG_MAX
+        &mod<40>, &mod<41>, &mod<42>, &mod<43>, &mod<44>, &mod<45>,
+        &mod<46>, &mod<47>, &mod<48>, &mod<49>, &mod<50>,
+#endif
+    }};
+
+}  // namespace detail
+
+/**
+ * Grow the hash table by using prime numbers as bucket count. Slower than
+ * tsl::rh::power_of_two_growth_policy in general but will probably distribute
+ * the values around better in the buckets with a poor hash function.
+ *
+ * To allow the compiler to optimize the modulo operation, a lookup table is
+ * used with constant primes numbers.
+ *
+ * With a switch the code would look like:
+ * \code
+ * switch(iprime) { // iprime is the current prime of the hash table
+ *     case 0: hash % 5ul;
+ *             break;
+ *     case 1: hash % 17ul;
+ *             break;
+ *     case 2: hash % 29ul;
+ *             break;
+ *     ...
+ * }
+ * \endcode
+ *
+ * Due to the constant variable in the modulo the compiler is able to optimize
+ * the operation by a series of multiplications, substractions and shifts.
+ *
+ * The 'hash % 5' could become something like 'hash - (hash * 0xCCCCCCCD) >> 34)
+ * * 5' in a 64 bits environment.
+ */
+class prime_growth_policy {
+ public:
+  explicit prime_growth_policy(std::size_t& min_bucket_count_in_out) {
+    auto it_prime = std::lower_bound(
+        detail::PRIMES.begin(), detail::PRIMES.end(), min_bucket_count_in_out);
+    if (it_prime == detail::PRIMES.end()) {
+      TSL_RH_THROW_OR_TERMINATE(std::length_error,
+                                "The hash table exceeds its maximum size.");
+    }
+
+    m_iprime = static_cast<unsigned int>(
+        std::distance(detail::PRIMES.begin(), it_prime));
+    if (min_bucket_count_in_out > 0) {
+      min_bucket_count_in_out = *it_prime;
+    } else {
+      min_bucket_count_in_out = 0;
+    }
+  }
+
+  std::size_t bucket_for_hash(std::size_t hash) const noexcept {
+    return detail::MOD_PRIME[m_iprime](hash);
+  }
+
+  std::size_t next_bucket_count() const {
+    if (m_iprime + 1 >= detail::PRIMES.size()) {
+      TSL_RH_THROW_OR_TERMINATE(std::length_error,
+                                "The hash table exceeds its maximum size.");
+    }
+
+    return detail::PRIMES[m_iprime + 1];
+  }
+
+  std::size_t max_bucket_count() const { return detail::PRIMES.back(); }
+
+  void clear() noexcept { m_iprime = 0; }
+
+ private:
+  unsigned int m_iprime;
+
+  static_assert(std::numeric_limits<decltype(m_iprime)>::max() >=
+                    detail::PRIMES.size(),
+                "The type of m_iprime is not big enough.");
+};
+
+}  // namespace rh
+}  // namespace tsl
+
+#endif
diff --git a/ext/robin-map/robin_hash.h b/ext/robin-map/robin_hash.h
new file mode 100644
index 0000000..03d23a9
--- /dev/null
+++ b/ext/robin-map/robin_hash.h
@@ -0,0 +1,1642 @@
+/**
+ * MIT License
+ *
+ * Copyright (c) 2017 Thibaut Goetghebuer-Planchon <tessil@gmx.com>
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#ifndef TSL_ROBIN_HASH_H
+#define TSL_ROBIN_HASH_H
+
+#include <algorithm>
+#include <cassert>
+#include <cmath>
+#include <cstddef>
+#include <cstdint>
+#include <exception>
+#include <iterator>
+#include <limits>
+#include <memory>
+#include <new>
+#include <stdexcept>
+#include <tuple>
+#include <type_traits>
+#include <utility>
+#include <vector>
+
+#include "robin_growth_policy.h"
+
+namespace tsl {
+
+namespace detail_robin_hash {
+
+template <typename T>
+struct make_void {
+  using type = void;
+};
+
+template <typename T, typename = void>
+struct has_is_transparent : std::false_type {};
+
+template <typename T>
+struct has_is_transparent<T,
+                          typename make_void<typename T::is_transparent>::type>
+    : std::true_type {};
+
+template <typename U>
+struct is_power_of_two_policy : std::false_type {};
+
+template <std::size_t GrowthFactor>
+struct is_power_of_two_policy<tsl::rh::power_of_two_growth_policy<GrowthFactor>>
+    : std::true_type {};
+
+// Only available in C++17, we need to be compatible with C++11
+template <class T>
+const T& clamp(const T& v, const T& lo, const T& hi) {
+  return std::min(hi, std::max(lo, v));
+}
+
+template <typename T, typename U>
+static T numeric_cast(U value,
+                      const char* error_message = "numeric_cast() failed.") {
+  T ret = static_cast<T>(value);
+  if (static_cast<U>(ret) != value) {
+    TSL_RH_THROW_OR_TERMINATE(std::runtime_error, error_message);
+  }
+
+  const bool is_same_signedness =
+      (std::is_unsigned<T>::value && std::is_unsigned<U>::value) ||
+      (std::is_signed<T>::value && std::is_signed<U>::value);
+  if (!is_same_signedness && (ret < T{}) != (value < U{})) {
+    TSL_RH_THROW_OR_TERMINATE(std::runtime_error, error_message);
+  }
+
+  TSL_RH_UNUSED(error_message);
+
+  return ret;
+}
+
+template <class T, class Deserializer>
+static T deserialize_value(Deserializer& deserializer) {
+  // MSVC < 2017 is not conformant, circumvent the problem by removing the
+  // template keyword
+#if defined(_MSC_VER) && _MSC_VER < 1910
+  return deserializer.Deserializer::operator()<T>();
+#else
+  return deserializer.Deserializer::template operator()<T>();
+#endif
+}
+
+/**
+ * Fixed size type used to represent size_type values on serialization. Need to
+ * be big enough to represent a std::size_t on 32 and 64 bits platforms, and
+ * must be the same size on both platforms.
+ */
+using slz_size_type = std::uint64_t;
+static_assert(std::numeric_limits<slz_size_type>::max() >=
+                  std::numeric_limits<std::size_t>::max(),
+              "slz_size_type must be >= std::size_t");
+
+using truncated_hash_type = std::uint32_t;
+
+/**
+ * Helper class that stores a truncated hash if StoreHash is true and nothing
+ * otherwise.
+ */
+template <bool StoreHash>
+class bucket_entry_hash {
+ public:
+  bool bucket_hash_equal(std::size_t /*hash*/) const noexcept { return true; }
+
+  truncated_hash_type truncated_hash() const noexcept { return 0; }
+
+ protected:
+  void set_hash(truncated_hash_type /*hash*/) noexcept {}
+};
+
+template <>
+class bucket_entry_hash<true> {
+ public:
+  bool bucket_hash_equal(std::size_t hash) const noexcept {
+    return m_hash == truncated_hash_type(hash);
+  }
+
+  truncated_hash_type truncated_hash() const noexcept { return m_hash; }
+
+ protected:
+  void set_hash(truncated_hash_type hash) noexcept {
+    m_hash = truncated_hash_type(hash);
+  }
+
+ private:
+  truncated_hash_type m_hash;
+};
+
+/**
+ * Each bucket entry has:
+ * - A value of type `ValueType`.
+ * - An integer to store how far the value of the bucket, if any, is from its
+ * ideal bucket (ex: if the current bucket 5 has the value 'foo' and
+ * `hash('foo') % nb_buckets` == 3, `dist_from_ideal_bucket()` will return 2 as
+ * the current value of the bucket is two buckets away from its ideal bucket) If
+ * there is no value in the bucket (i.e. `empty()` is true)
+ * `dist_from_ideal_bucket()` will be < 0.
+ * - A marker which tells us if the bucket is the last bucket of the bucket
+ * array (useful for the iterator of the hash table).
+ * - If `StoreHash` is true, 32 bits of the hash of the value, if any, are also
+ * stored in the bucket. If the size of the hash is more than 32 bits, it is
+ * truncated. We don't store the full hash as storing the hash is a potential
+ * opportunity to use the unused space due to the alignment of the bucket_entry
+ * structure. We can thus potentially store the hash without any extra space
+ *   (which would not be possible with 64 bits of the hash).
+ */
+template <typename ValueType, bool StoreHash>
+class bucket_entry : public bucket_entry_hash<StoreHash> {
+  using bucket_hash = bucket_entry_hash<StoreHash>;
+
+ public:
+  using value_type = ValueType;
+  using distance_type = std::int16_t;
+
+  bucket_entry() noexcept
+      : bucket_hash(),
+        m_dist_from_ideal_bucket(EMPTY_MARKER_DIST_FROM_IDEAL_BUCKET),
+        m_last_bucket(false) {
+    tsl_rh_assert(empty());
+  }
+
+  bucket_entry(bool last_bucket) noexcept
+      : bucket_hash(),
+        m_dist_from_ideal_bucket(EMPTY_MARKER_DIST_FROM_IDEAL_BUCKET),
+        m_last_bucket(last_bucket) {
+    tsl_rh_assert(empty());
+  }
+
+  bucket_entry(const bucket_entry& other) noexcept(
+      std::is_nothrow_copy_constructible<value_type>::value)
+      : bucket_hash(other),
+        m_dist_from_ideal_bucket(EMPTY_MARKER_DIST_FROM_IDEAL_BUCKET),
+        m_last_bucket(other.m_last_bucket) {
+    if (!other.empty()) {
+      ::new (static_cast<void*>(std::addressof(m_value)))
+          value_type(other.value());
+      m_dist_from_ideal_bucket = other.m_dist_from_ideal_bucket;
+    }
+    tsl_rh_assert(empty() == other.empty());
+  }
+
+  /**
+   * Never really used, but still necessary as we must call resize on an empty
+   * `std::vector<bucket_entry>`. and we need to support move-only types. See
+   * robin_hash constructor for details.
+   */
+  bucket_entry(bucket_entry&& other) noexcept(
+      std::is_nothrow_move_constructible<value_type>::value)
+      : bucket_hash(std::move(other)),
+        m_dist_from_ideal_bucket(EMPTY_MARKER_DIST_FROM_IDEAL_BUCKET),
+        m_last_bucket(other.m_last_bucket) {
+    if (!other.empty()) {
+      ::new (static_cast<void*>(std::addressof(m_value)))
+          value_type(std::move(other.value()));
+      m_dist_from_ideal_bucket = other.m_dist_from_ideal_bucket;
+    }
+    tsl_rh_assert(empty() == other.empty());
+  }
+
+  bucket_entry& operator=(const bucket_entry& other) noexcept(
+      std::is_nothrow_copy_constructible<value_type>::value) {
+    if (this != &other) {
+      clear();
+
+      bucket_hash::operator=(other);
+      if (!other.empty()) {
+        ::new (static_cast<void*>(std::addressof(m_value)))
+            value_type(other.value());
+      }
+
+      m_dist_from_ideal_bucket = other.m_dist_from_ideal_bucket;
+      m_last_bucket = other.m_last_bucket;
+    }
+
+    return *this;
+  }
+
+  bucket_entry& operator=(bucket_entry&&) = delete;
+
+  ~bucket_entry() noexcept { clear(); }
+
+  void clear() noexcept {
+    if (!empty()) {
+      destroy_value();
+      m_dist_from_ideal_bucket = EMPTY_MARKER_DIST_FROM_IDEAL_BUCKET;
+    }
+  }
+
+  bool empty() const noexcept {
+    return m_dist_from_ideal_bucket == EMPTY_MARKER_DIST_FROM_IDEAL_BUCKET;
+  }
+
+  value_type& value() noexcept {
+    tsl_rh_assert(!empty());
+#if defined(__cplusplus) && __cplusplus >= 201703L
+    return *std::launder(
+        reinterpret_cast<value_type*>(std::addressof(m_value)));
+#else
+    return *reinterpret_cast<value_type*>(std::addressof(m_value));
+#endif
+  }
+
+  const value_type& value() const noexcept {
+    tsl_rh_assert(!empty());
+#if defined(__cplusplus) && __cplusplus >= 201703L
+    return *std::launder(
+        reinterpret_cast<const value_type*>(std::addressof(m_value)));
+#else
+    return *reinterpret_cast<const value_type*>(std::addressof(m_value));
+#endif
+  }
+
+  distance_type dist_from_ideal_bucket() const noexcept {
+    return m_dist_from_ideal_bucket;
+  }
+
+  bool last_bucket() const noexcept { return m_last_bucket; }
+
+  void set_as_last_bucket() noexcept { m_last_bucket = true; }
+
+  template <typename... Args>
+  void set_value_of_empty_bucket(distance_type dist_from_ideal_bucket,
+                                 truncated_hash_type hash,
+                                 Args&&... value_type_args) {
+    tsl_rh_assert(dist_from_ideal_bucket >= 0);
+    tsl_rh_assert(empty());
+
+    ::new (static_cast<void*>(std::addressof(m_value)))
+        value_type(std::forward<Args>(value_type_args)...);
+    this->set_hash(hash);
+    m_dist_from_ideal_bucket = dist_from_ideal_bucket;
+
+    tsl_rh_assert(!empty());
+  }
+
+  void swap_with_value_in_bucket(distance_type& dist_from_ideal_bucket,
+                                 truncated_hash_type& hash, value_type& value) {
+    tsl_rh_assert(!empty());
+    tsl_rh_assert(dist_from_ideal_bucket > m_dist_from_ideal_bucket);
+
+    using std::swap;
+    swap(value, this->value());
+    swap(dist_from_ideal_bucket, m_dist_from_ideal_bucket);
+
+    if (StoreHash) {
+      const truncated_hash_type tmp_hash = this->truncated_hash();
+      this->set_hash(hash);
+      hash = tmp_hash;
+    } else {
+      // Avoid warning of unused variable if StoreHash is false
+      TSL_RH_UNUSED(hash);
+    }
+  }
+
+  static truncated_hash_type truncate_hash(std::size_t hash) noexcept {
+    return truncated_hash_type(hash);
+  }
+
+ private:
+  void destroy_value() noexcept {
+    tsl_rh_assert(!empty());
+    value().~value_type();
+  }
+
+ public:
+  static const distance_type EMPTY_MARKER_DIST_FROM_IDEAL_BUCKET = -1;
+  static const distance_type DIST_FROM_IDEAL_BUCKET_LIMIT = 8192;
+  static_assert(DIST_FROM_IDEAL_BUCKET_LIMIT <=
+                    std::numeric_limits<distance_type>::max() - 1,
+                "DIST_FROM_IDEAL_BUCKET_LIMIT must be <= "
+                "std::numeric_limits<distance_type>::max() - 1.");
+
+ private:
+  distance_type m_dist_from_ideal_bucket;
+  bool m_last_bucket;
+  alignas(value_type) unsigned char m_value[sizeof(value_type)];
+};
+
+/**
+ * Internal common class used by `robin_map` and `robin_set`.
+ *
+ * ValueType is what will be stored by `robin_hash` (usually `std::pair<Key, T>`
+ * for map and `Key` for set).
+ *
+ * `KeySelect` should be a `FunctionObject` which takes a `ValueType` in
+ * parameter and returns a reference to the key.
+ *
+ * `ValueSelect` should be a `FunctionObject` which takes a `ValueType` in
+ * parameter and returns a reference to the value. `ValueSelect` should be void
+ * if there is no value (in a set for example).
+ *
+ * The strong exception guarantee only holds if the expression
+ * `std::is_nothrow_swappable<ValueType>::value &&
+ * std::is_nothrow_move_constructible<ValueType>::value` is true.
+ *
+ * Behaviour is undefined if the destructor of `ValueType` throws.
+ */
+template <class ValueType, class KeySelect, class ValueSelect, class Hash,
+          class KeyEqual, class Allocator, bool StoreHash, class GrowthPolicy>
+class robin_hash : private Hash, private KeyEqual, private GrowthPolicy {
+ private:
+  template <typename U>
+  using has_mapped_type =
+      typename std::integral_constant<bool, !std::is_same<U, void>::value>;
+
+  static_assert(
+      noexcept(std::declval<GrowthPolicy>().bucket_for_hash(std::size_t(0))),
+      "GrowthPolicy::bucket_for_hash must be noexcept.");
+  static_assert(noexcept(std::declval<GrowthPolicy>().clear()),
+                "GrowthPolicy::clear must be noexcept.");
+
+ public:
+  template <bool IsConst>
+  class robin_iterator;
+
+  using key_type = typename KeySelect::key_type;
+  using value_type = ValueType;
+  using size_type = std::size_t;
+  using difference_type = std::ptrdiff_t;
+  using hasher = Hash;
+  using key_equal = KeyEqual;
+  using allocator_type = Allocator;
+  using reference = value_type&;
+  using const_reference = const value_type&;
+  using pointer = value_type*;
+  using const_pointer = const value_type*;
+  using iterator = robin_iterator<false>;
+  using const_iterator = robin_iterator<true>;
+
+ private:
+  /**
+   * Either store the hash because we are asked by the `StoreHash` template
+   * parameter or store the hash because it doesn't cost us anything in size and
+   * can be used to speed up rehash.
+   */
+  static constexpr bool STORE_HASH =
+      StoreHash ||
+      ((sizeof(tsl::detail_robin_hash::bucket_entry<value_type, true>) ==
+        sizeof(tsl::detail_robin_hash::bucket_entry<value_type, false>)) &&
+       (sizeof(std::size_t) == sizeof(truncated_hash_type) ||
+        is_power_of_two_policy<GrowthPolicy>::value) &&
+       // Don't store the hash for primitive types with default hash.
+       (!std::is_arithmetic<key_type>::value ||
+        !std::is_same<Hash, std::hash<key_type>>::value));
+
+  /**
+   * Only use the stored hash on lookup if we are explicitly asked. We are not
+   * sure how slow the KeyEqual operation is. An extra comparison may slow
+   * things down with a fast KeyEqual.
+   */
+  static constexpr bool USE_STORED_HASH_ON_LOOKUP = StoreHash;
+
+  /**
+   * We can only use the hash on rehash if the size of the hash type is the same
+   * as the stored one or if we use a power of two modulo. In the case of the
+   * power of two modulo, we just mask the least significant bytes, we just have
+   * to check that the truncated_hash_type didn't truncated more bytes.
+   */
+  static bool USE_STORED_HASH_ON_REHASH(size_type bucket_count) {
+    if (STORE_HASH && sizeof(std::size_t) == sizeof(truncated_hash_type)) {
+      TSL_RH_UNUSED(bucket_count);
+      return true;
+    } else if (STORE_HASH && is_power_of_two_policy<GrowthPolicy>::value) {
+      return bucket_count == 0 ||
+             (bucket_count - 1) <=
+                 std::numeric_limits<truncated_hash_type>::max();
+    } else {
+      TSL_RH_UNUSED(bucket_count);
+      return false;
+    }
+  }
+
+  using bucket_entry =
+      tsl::detail_robin_hash::bucket_entry<value_type, STORE_HASH>;
+  using distance_type = typename bucket_entry::distance_type;
+
+  using buckets_allocator = typename std::allocator_traits<
+      allocator_type>::template rebind_alloc<bucket_entry>;
+  using buckets_container_type = std::vector<bucket_entry, buckets_allocator>;
+
+ public:
+  /**
+   * The 'operator*()' and 'operator->()' methods return a const reference and
+   * const pointer respectively to the stored value type.
+   *
+   * In case of a map, to get a mutable reference to the value associated to a
+   * key (the '.second' in the stored pair), you have to call 'value()'.
+   *
+   * The main reason for this is that if we returned a `std::pair<Key, T>&`
+   * instead of a `const std::pair<Key, T>&`, the user may modify the key which
+   * will put the map in a undefined state.
+   */
+  template <bool IsConst>
+  class robin_iterator {
+    friend class robin_hash;
+
+   private:
+    using bucket_entry_ptr =
+        typename std::conditional<IsConst, const bucket_entry*,
+                                  bucket_entry*>::type;
+
+    robin_iterator(bucket_entry_ptr bucket) noexcept : m_bucket(bucket) {}
+
+   public:
+    using iterator_category = std::forward_iterator_tag;
+    using value_type = const typename robin_hash::value_type;
+    using difference_type = std::ptrdiff_t;
+    using reference = value_type&;
+    using pointer = value_type*;
+
+    robin_iterator() noexcept {}
+
+    // Copy constructor from iterator to const_iterator.
+    template <bool TIsConst = IsConst,
+              typename std::enable_if<TIsConst>::type* = nullptr>
+    robin_iterator(const robin_iterator<!TIsConst>& other) noexcept
+        : m_bucket(other.m_bucket) {}
+
+    robin_iterator(const robin_iterator& other) = default;
+    robin_iterator(robin_iterator&& other) = default;
+    robin_iterator& operator=(const robin_iterator& other) = default;
+    robin_iterator& operator=(robin_iterator&& other) = default;
+
+    const typename robin_hash::key_type& key() const {
+      return KeySelect()(m_bucket->value());
+    }
+
+    template <class U = ValueSelect,
+              typename std::enable_if<has_mapped_type<U>::value &&
+                                      IsConst>::type* = nullptr>
+    const typename U::value_type& value() const {
+      return U()(m_bucket->value());
+    }
+
+    template <class U = ValueSelect,
+              typename std::enable_if<has_mapped_type<U>::value &&
+                                      !IsConst>::type* = nullptr>
+    typename U::value_type& value() const {
+      return U()(m_bucket->value());
+    }
+
+    reference operator*() const { return m_bucket->value(); }
+
+    pointer operator->() const { return std::addressof(m_bucket->value()); }
+
+    robin_iterator& operator++() {
+      while (true) {
+        if (m_bucket->last_bucket()) {
+          ++m_bucket;
+          return *this;
+        }
+
+        ++m_bucket;
+        if (!m_bucket->empty()) {
+          return *this;
+        }
+      }
+    }
+
+    robin_iterator operator++(int) {
+      robin_iterator tmp(*this);
+      ++*this;
+
+      return tmp;
+    }
+
+    friend bool operator==(const robin_iterator& lhs,
+                           const robin_iterator& rhs) {
+      return lhs.m_bucket == rhs.m_bucket;
+    }
+
+    friend bool operator!=(const robin_iterator& lhs,
+                           const robin_iterator& rhs) {
+      return !(lhs == rhs);
+    }
+
+   private:
+    bucket_entry_ptr m_bucket;
+  };
+
+ public:
+#if defined(__cplusplus) && __cplusplus >= 201402L
+  robin_hash(size_type bucket_count, const Hash& hash, const KeyEqual& equal,
+             const Allocator& alloc,
+             float min_load_factor = DEFAULT_MIN_LOAD_FACTOR,
+             float max_load_factor = DEFAULT_MAX_LOAD_FACTOR)
+      : Hash(hash),
+        KeyEqual(equal),
+        GrowthPolicy(bucket_count),
+        m_buckets_data(bucket_count, alloc),
+        m_buckets(m_buckets_data.empty() ? static_empty_bucket_ptr()
+                                         : m_buckets_data.data()),
+        m_bucket_count(bucket_count),
+        m_nb_elements(0),
+        m_grow_on_next_insert(false),
+        m_try_shrink_on_next_insert(false) {
+    if (bucket_count > max_bucket_count()) {
+      TSL_RH_THROW_OR_TERMINATE(std::length_error,
+                                "The map exceeds its maximum bucket count.");
+    }
+
+    if (m_bucket_count > 0) {
+      tsl_rh_assert(!m_buckets_data.empty());
+      m_buckets_data.back().set_as_last_bucket();
+    }
+
+    this->min_load_factor(min_load_factor);
+    this->max_load_factor(max_load_factor);
+  }
+#else
+  /**
+   * C++11 doesn't support the creation of a std::vector with a custom allocator
+   * and 'count' default-inserted elements. The needed contructor `explicit
+   * vector(size_type count, const Allocator& alloc = Allocator());` is only
+   * available in C++14 and later. We thus must resize after using the
+   * `vector(const Allocator& alloc)` constructor.
+   *
+   * We can't use `vector(size_type count, const T& value, const Allocator&
+   * alloc)` as it requires the value T to be copyable.
+   */
+  robin_hash(size_type bucket_count, const Hash& hash, const KeyEqual& equal,
+             const Allocator& alloc,
+             float min_load_factor = DEFAULT_MIN_LOAD_FACTOR,
+             float max_load_factor = DEFAULT_MAX_LOAD_FACTOR)
+      : Hash(hash),
+        KeyEqual(equal),
+        GrowthPolicy(bucket_count),
+        m_buckets_data(alloc),
+        m_buckets(static_empty_bucket_ptr()),
+        m_bucket_count(bucket_count),
+        m_nb_elements(0),
+        m_grow_on_next_insert(false),
+        m_try_shrink_on_next_insert(false) {
+    if (bucket_count > max_bucket_count()) {
+      TSL_RH_THROW_OR_TERMINATE(std::length_error,
+                                "The map exceeds its maximum bucket count.");
+    }
+
+    if (m_bucket_count > 0) {
+      m_buckets_data.resize(m_bucket_count);
+      m_buckets = m_buckets_data.data();
+
+      tsl_rh_assert(!m_buckets_data.empty());
+      m_buckets_data.back().set_as_last_bucket();
+    }
+
+    this->min_load_factor(min_load_factor);
+    this->max_load_factor(max_load_factor);
+  }
+#endif
+
+  robin_hash(const robin_hash& other)
+      : Hash(other),
+        KeyEqual(other),
+        GrowthPolicy(other),
+        m_buckets_data(other.m_buckets_data),
+        m_buckets(m_buckets_data.empty() ? static_empty_bucket_ptr()
+                                         : m_buckets_data.data()),
+        m_bucket_count(other.m_bucket_count),
+        m_nb_elements(other.m_nb_elements),
+        m_load_threshold(other.m_load_threshold),
+        m_min_load_factor(other.m_min_load_factor),
+        m_max_load_factor(other.m_max_load_factor),
+        m_grow_on_next_insert(other.m_grow_on_next_insert),
+        m_try_shrink_on_next_insert(other.m_try_shrink_on_next_insert) {}
+
+  robin_hash(robin_hash&& other) noexcept(
+      std::is_nothrow_move_constructible<
+          Hash>::value&& std::is_nothrow_move_constructible<KeyEqual>::value&&
+          std::is_nothrow_move_constructible<GrowthPolicy>::value&&
+              std::is_nothrow_move_constructible<buckets_container_type>::value)
+      : Hash(std::move(static_cast<Hash&>(other))),
+        KeyEqual(std::move(static_cast<KeyEqual&>(other))),
+        GrowthPolicy(std::move(static_cast<GrowthPolicy&>(other))),
+        m_buckets_data(std::move(other.m_buckets_data)),
+        m_buckets(m_buckets_data.empty() ? static_empty_bucket_ptr()
+                                         : m_buckets_data.data()),
+        m_bucket_count(other.m_bucket_count),
+        m_nb_elements(other.m_nb_elements),
+        m_load_threshold(other.m_load_threshold),
+        m_min_load_factor(other.m_min_load_factor),
+        m_max_load_factor(other.m_max_load_factor),
+        m_grow_on_next_insert(other.m_grow_on_next_insert),
+        m_try_shrink_on_next_insert(other.m_try_shrink_on_next_insert) {
+    other.clear_and_shrink();
+  }
+
+  robin_hash& operator=(const robin_hash& other) {
+    if (&other != this) {
+      Hash::operator=(other);
+      KeyEqual::operator=(other);
+      GrowthPolicy::operator=(other);
+
+      m_buckets_data = other.m_buckets_data;
+      m_buckets = m_buckets_data.empty() ? static_empty_bucket_ptr()
+                                         : m_buckets_data.data();
+      m_bucket_count = other.m_bucket_count;
+      m_nb_elements = other.m_nb_elements;
+
+      m_load_threshold = other.m_load_threshold;
+      m_min_load_factor = other.m_min_load_factor;
+      m_max_load_factor = other.m_max_load_factor;
+
+      m_grow_on_next_insert = other.m_grow_on_next_insert;
+      m_try_shrink_on_next_insert = other.m_try_shrink_on_next_insert;
+    }
+
+    return *this;
+  }
+
+  robin_hash& operator=(robin_hash&& other) {
+    other.swap(*this);
+    other.clear_and_shrink();
+
+    return *this;
+  }
+
+  allocator_type get_allocator() const {
+    return m_buckets_data.get_allocator();
+  }
+
+  /*
+   * Iterators
+   */
+  iterator begin() noexcept {
+    std::size_t i = 0;
+    while (i < m_bucket_count && m_buckets[i].empty()) {
+      i++;
+    }
+
+    return iterator(m_buckets + i);
+  }
+
+  const_iterator begin() const noexcept { return cbegin(); }
+
+  const_iterator cbegin() const noexcept {
+    std::size_t i = 0;
+    while (i < m_bucket_count && m_buckets[i].empty()) {
+      i++;
+    }
+
+    return const_iterator(m_buckets + i);
+  }
+
+  iterator end() noexcept { return iterator(m_buckets + m_bucket_count); }
+
+  const_iterator end() const noexcept { return cend(); }
+
+  const_iterator cend() const noexcept {
+    return const_iterator(m_buckets + m_bucket_count);
+  }
+
+  /*
+   * Capacity
+   */
+  bool empty() const noexcept { return m_nb_elements == 0; }
+
+  size_type size() const noexcept { return m_nb_elements; }
+
+  size_type max_size() const noexcept { return m_buckets_data.max_size(); }
+
+  /*
+   * Modifiers
+   */
+  void clear() noexcept {
+    if (m_min_load_factor > 0.0f) {
+      clear_and_shrink();
+    } else {
+      for (auto& bucket : m_buckets_data) {
+        bucket.clear();
+      }
+
+      m_nb_elements = 0;
+      m_grow_on_next_insert = false;
+    }
+  }
+
+  template <typename P>
+  std::pair<iterator, bool> insert(P&& value) {
+    return insert_impl(KeySelect()(value), std::forward<P>(value));
+  }
+
+  template <typename P>
+  iterator insert_hint(const_iterator hint, P&& value) {
+    if (hint != cend() &&
+        compare_keys(KeySelect()(*hint), KeySelect()(value))) {
+      return mutable_iterator(hint);
+    }
+
+    return insert(std::forward<P>(value)).first;
+  }
+
+  template <class InputIt>
+  void insert(InputIt first, InputIt last) {
+    if (std::is_base_of<
+            std::forward_iterator_tag,
+            typename std::iterator_traits<InputIt>::iterator_category>::value) {
+      const auto nb_elements_insert = std::distance(first, last);
+      const size_type nb_free_buckets = m_load_threshold - size();
+      tsl_rh_assert(m_load_threshold >= size());
+
+      if (nb_elements_insert > 0 &&
+          nb_free_buckets < size_type(nb_elements_insert)) {
+        reserve(size() + size_type(nb_elements_insert));
+      }
+    }
+
+    for (; first != last; ++first) {
+      insert(*first);
+    }
+  }
+
+  template <class K, class M>
+  std::pair<iterator, bool> insert_or_assign(K&& key, M&& obj) {
+    auto it = try_emplace(std::forward<K>(key), std::forward<M>(obj));
+    if (!it.second) {
+      it.first.value() = std::forward<M>(obj);
+    }
+
+    return it;
+  }
+
+  template <class K, class M>
+  iterator insert_or_assign(const_iterator hint, K&& key, M&& obj) {
+    if (hint != cend() && compare_keys(KeySelect()(*hint), key)) {
+      auto it = mutable_iterator(hint);
+      it.value() = std::forward<M>(obj);
+
+      return it;
+    }
+
+    return insert_or_assign(std::forward<K>(key), std::forward<M>(obj)).first;
+  }
+
+  template <class... Args>
+  std::pair<iterator, bool> emplace(Args&&... args) {
+    return insert(value_type(std::forward<Args>(args)...));
+  }
+
+  template <class... Args>
+  iterator emplace_hint(const_iterator hint, Args&&... args) {
+    return insert_hint(hint, value_type(std::forward<Args>(args)...));
+  }
+
+  template <class K, class... Args>
+  std::pair<iterator, bool> try_emplace(K&& key, Args&&... args) {
+    return insert_impl(key, std::piecewise_construct,
+                       std::forward_as_tuple(std::forward<K>(key)),
+                       std::forward_as_tuple(std::forward<Args>(args)...));
+  }
+
+  template <class K, class... Args>
+  iterator try_emplace_hint(const_iterator hint, K&& key, Args&&... args) {
+    if (hint != cend() && compare_keys(KeySelect()(*hint), key)) {
+      return mutable_iterator(hint);
+    }
+
+    return try_emplace(std::forward<K>(key), std::forward<Args>(args)...).first;
+  }
+
+  void erase_fast(iterator pos) {
+    erase_from_bucket(pos);
+  }
+
+  /**
+   * Here to avoid `template<class K> size_type erase(const K& key)` being used
+   * when we use an `iterator` instead of a `const_iterator`.
+   */
+  iterator erase(iterator pos) {
+    erase_from_bucket(pos);
+
+    /**
+     * Erase bucket used a backward shift after clearing the bucket.
+     * Check if there is a new value in the bucket, if not get the next
+     * non-empty.
+     */
+    if (pos.m_bucket->empty()) {
+      ++pos;
+    }
+
+    return pos;
+  }
+
+  iterator erase(const_iterator pos) { return erase(mutable_iterator(pos)); }
+
+  iterator erase(const_iterator first, const_iterator last) {
+    if (first == last) {
+      return mutable_iterator(first);
+    }
+
+    auto first_mutable = mutable_iterator(first);
+    auto last_mutable = mutable_iterator(last);
+    for (auto it = first_mutable.m_bucket; it != last_mutable.m_bucket; ++it) {
+      if (!it->empty()) {
+        it->clear();
+        m_nb_elements--;
+      }
+    }
+
+    if (last_mutable == end()) {
+      m_try_shrink_on_next_insert = true;
+      return end();
+    }
+
+    /*
+     * Backward shift on the values which come after the deleted values.
+     * We try to move the values closer to their ideal bucket.
+     */
+    std::size_t icloser_bucket =
+        static_cast<std::size_t>(first_mutable.m_bucket - m_buckets);
+    std::size_t ito_move_closer_value =
+        static_cast<std::size_t>(last_mutable.m_bucket - m_buckets);
+    tsl_rh_assert(ito_move_closer_value > icloser_bucket);
+
+    const std::size_t ireturn_bucket =
+        ito_move_closer_value -
+        std::min(
+            ito_move_closer_value - icloser_bucket,
+            std::size_t(
+                m_buckets[ito_move_closer_value].dist_from_ideal_bucket()));
+
+    while (ito_move_closer_value < m_bucket_count &&
+           m_buckets[ito_move_closer_value].dist_from_ideal_bucket() > 0) {
+      icloser_bucket =
+          ito_move_closer_value -
+          std::min(
+              ito_move_closer_value - icloser_bucket,
+              std::size_t(
+                  m_buckets[ito_move_closer_value].dist_from_ideal_bucket()));
+
+      tsl_rh_assert(m_buckets[icloser_bucket].empty());
+      const distance_type new_distance = distance_type(
+          m_buckets[ito_move_closer_value].dist_from_ideal_bucket() -
+          (ito_move_closer_value - icloser_bucket));
+      m_buckets[icloser_bucket].set_value_of_empty_bucket(
+          new_distance, m_buckets[ito_move_closer_value].truncated_hash(),
+          std::move(m_buckets[ito_move_closer_value].value()));
+      m_buckets[ito_move_closer_value].clear();
+
+      ++icloser_bucket;
+      ++ito_move_closer_value;
+    }
+
+    m_try_shrink_on_next_insert = true;
+
+    return iterator(m_buckets + ireturn_bucket);
+  }
+
+  template <class K>
+  size_type erase(const K& key) {
+    return erase(key, hash_key(key));
+  }
+
+  template <class K>
+  size_type erase(const K& key, std::size_t hash) {
+    auto it = find(key, hash);
+    if (it != end()) {
+      erase_from_bucket(it);
+      return 1;
+    } else {
+      return 0;
+    }
+  }
+
+  void swap(robin_hash& other) {
+    using std::swap;
+
+    swap(static_cast<Hash&>(*this), static_cast<Hash&>(other));
+    swap(static_cast<KeyEqual&>(*this), static_cast<KeyEqual&>(other));
+    swap(static_cast<GrowthPolicy&>(*this), static_cast<GrowthPolicy&>(other));
+    swap(m_buckets_data, other.m_buckets_data);
+    swap(m_buckets, other.m_buckets);
+    swap(m_bucket_count, other.m_bucket_count);
+    swap(m_nb_elements, other.m_nb_elements);
+    swap(m_load_threshold, other.m_load_threshold);
+    swap(m_min_load_factor, other.m_min_load_factor);
+    swap(m_max_load_factor, other.m_max_load_factor);
+    swap(m_grow_on_next_insert, other.m_grow_on_next_insert);
+    swap(m_try_shrink_on_next_insert, other.m_try_shrink_on_next_insert);
+  }
+
+  /*
+   * Lookup
+   */
+  template <class K, class U = ValueSelect,
+            typename std::enable_if<has_mapped_type<U>::value>::type* = nullptr>
+  typename U::value_type& at(const K& key) {
+    return at(key, hash_key(key));
+  }
+
+  template <class K, class U = ValueSelect,
+            typename std::enable_if<has_mapped_type<U>::value>::type* = nullptr>
+  typename U::value_type& at(const K& key, std::size_t hash) {
+    return const_cast<typename U::value_type&>(
+        static_cast<const robin_hash*>(this)->at(key, hash));
+  }
+
+  template <class K, class U = ValueSelect,
+            typename std::enable_if<has_mapped_type<U>::value>::type* = nullptr>
+  const typename U::value_type& at(const K& key) const {
+    return at(key, hash_key(key));
+  }
+
+  template <class K, class U = ValueSelect,
+            typename std::enable_if<has_mapped_type<U>::value>::type* = nullptr>
+  const typename U::value_type& at(const K& key, std::size_t hash) const {
+    auto it = find(key, hash);
+    if (it != cend()) {
+      return it.value();
+    } else {
+      TSL_RH_THROW_OR_TERMINATE(std::out_of_range, "Couldn't find key.");
+    }
+  }
+
+  template <class K, class U = ValueSelect,
+            typename std::enable_if<has_mapped_type<U>::value>::type* = nullptr>
+  typename U::value_type& operator[](K&& key) {
+    return try_emplace(std::forward<K>(key)).first.value();
+  }
+
+  template <class K>
+  size_type count(const K& key) const {
+    return count(key, hash_key(key));
+  }
+
+  template <class K>
+  size_type count(const K& key, std::size_t hash) const {
+    if (find(key, hash) != cend()) {
+      return 1;
+    } else {
+      return 0;
+    }
+  }
+
+  template <class K>
+  iterator find(const K& key) {
+    return find_impl(key, hash_key(key));
+  }
+
+  template <class K>
+  iterator find(const K& key, std::size_t hash) {
+    return find_impl(key, hash);
+  }
+
+  template <class K>
+  const_iterator find(const K& key) const {
+    return find_impl(key, hash_key(key));
+  }
+
+  template <class K>
+  const_iterator find(const K& key, std::size_t hash) const {
+    return find_impl(key, hash);
+  }
+
+  template <class K>
+  bool contains(const K& key) const {
+    return contains(key, hash_key(key));
+  }
+
+  template <class K>
+  bool contains(const K& key, std::size_t hash) const {
+    return count(key, hash) != 0;
+  }
+
+  template <class K>
+  std::pair<iterator, iterator> equal_range(const K& key) {
+    return equal_range(key, hash_key(key));
+  }
+
+  template <class K>
+  std::pair<iterator, iterator> equal_range(const K& key, std::size_t hash) {
+    iterator it = find(key, hash);
+    return std::make_pair(it, (it == end()) ? it : std::next(it));
+  }
+
+  template <class K>
+  std::pair<const_iterator, const_iterator> equal_range(const K& key) const {
+    return equal_range(key, hash_key(key));
+  }
+
+  template <class K>
+  std::pair<const_iterator, const_iterator> equal_range(
+      const K& key, std::size_t hash) const {
+    const_iterator it = find(key, hash);
+    return std::make_pair(it, (it == cend()) ? it : std::next(it));
+  }
+
+  /*
+   * Bucket interface
+   */
+  size_type bucket_count() const { return m_bucket_count; }
+
+  size_type max_bucket_count() const {
+    return std::min(GrowthPolicy::max_bucket_count(),
+                    m_buckets_data.max_size());
+  }
+
+  /*
+   * Hash policy
+   */
+  float load_factor() const {
+    if (bucket_count() == 0) {
+      return 0;
+    }
+
+    return float(m_nb_elements) / float(bucket_count());
+  }
+
+  float min_load_factor() const { return m_min_load_factor; }
+
+  float max_load_factor() const { return m_max_load_factor; }
+
+  void min_load_factor(float ml) {
+    m_min_load_factor = clamp(ml, float(MINIMUM_MIN_LOAD_FACTOR),
+                              float(MAXIMUM_MIN_LOAD_FACTOR));
+  }
+
+  void max_load_factor(float ml) {
+    m_max_load_factor = clamp(ml, float(MINIMUM_MAX_LOAD_FACTOR),
+                              float(MAXIMUM_MAX_LOAD_FACTOR));
+    m_load_threshold = size_type(float(bucket_count()) * m_max_load_factor);
+    tsl_rh_assert(bucket_count() == 0 || m_load_threshold < bucket_count());
+  }
+
+  void rehash(size_type count_) {
+    count_ = std::max(count_,
+                      size_type(std::ceil(float(size()) / max_load_factor())));
+    rehash_impl(count_);
+  }
+
+  void reserve(size_type count_) {
+    rehash(size_type(std::ceil(float(count_) / max_load_factor())));
+  }
+
+  /*
+   * Observers
+   */
+  hasher hash_function() const { return static_cast<const Hash&>(*this); }
+
+  key_equal key_eq() const { return static_cast<const KeyEqual&>(*this); }
+
+  /*
+   * Other
+   */
+  iterator mutable_iterator(const_iterator pos) {
+    return iterator(const_cast<bucket_entry*>(pos.m_bucket));
+  }
+
+  template <class Serializer>
+  void serialize(Serializer& serializer) const {
+    serialize_impl(serializer);
+  }
+
+  template <class Deserializer>
+  void deserialize(Deserializer& deserializer, bool hash_compatible) {
+    deserialize_impl(deserializer, hash_compatible);
+  }
+
+ private:
+  template <class K>
+  std::size_t hash_key(const K& key) const {
+    return Hash::operator()(key);
+  }
+
+  template <class K1, class K2>
+  bool compare_keys(const K1& key1, const K2& key2) const {
+    return KeyEqual::operator()(key1, key2);
+  }
+
+  std::size_t bucket_for_hash(std::size_t hash) const {
+    const std::size_t bucket = GrowthPolicy::bucket_for_hash(hash);
+    tsl_rh_assert(bucket < m_bucket_count ||
+                  (bucket == 0 && m_bucket_count == 0));
+
+    return bucket;
+  }
+
+  template <class U = GrowthPolicy,
+            typename std::enable_if<is_power_of_two_policy<U>::value>::type* =
+                nullptr>
+  std::size_t next_bucket(std::size_t index) const noexcept {
+    tsl_rh_assert(index < bucket_count());
+
+    return (index + 1) & this->m_mask;
+  }
+
+  template <class U = GrowthPolicy,
+            typename std::enable_if<!is_power_of_two_policy<U>::value>::type* =
+                nullptr>
+  std::size_t next_bucket(std::size_t index) const noexcept {
+    tsl_rh_assert(index < bucket_count());
+
+    index++;
+    return (index != bucket_count()) ? index : 0;
+  }
+
+  template <class K>
+  iterator find_impl(const K& key, std::size_t hash) {
+    return mutable_iterator(
+        static_cast<const robin_hash*>(this)->find(key, hash));
+  }
+
+  template <class K>
+  const_iterator find_impl(const K& key, std::size_t hash) const {
+    std::size_t ibucket = bucket_for_hash(hash);
+    distance_type dist_from_ideal_bucket = 0;
+
+    while (dist_from_ideal_bucket <=
+           m_buckets[ibucket].dist_from_ideal_bucket()) {
+      if (TSL_RH_LIKELY(
+              (!USE_STORED_HASH_ON_LOOKUP ||
+               m_buckets[ibucket].bucket_hash_equal(hash)) &&
+              compare_keys(KeySelect()(m_buckets[ibucket].value()), key))) {
+        return const_iterator(m_buckets + ibucket);
+      }
+
+      ibucket = next_bucket(ibucket);
+      dist_from_ideal_bucket++;
+    }
+
+    return cend();
+  }
+
+  void erase_from_bucket(iterator pos) {
+    pos.m_bucket->clear();
+    m_nb_elements--;
+
+    /**
+     * Backward shift, swap the empty bucket, previous_ibucket, with the values
+     * on its right, ibucket, until we cross another empty bucket or if the
+     * other bucket has a distance_from_ideal_bucket == 0.
+     *
+     * We try to move the values closer to their ideal bucket.
+     */
+    std::size_t previous_ibucket =
+        static_cast<std::size_t>(pos.m_bucket - m_buckets);
+    std::size_t ibucket = next_bucket(previous_ibucket);
+
+    while (m_buckets[ibucket].dist_from_ideal_bucket() > 0) {
+      tsl_rh_assert(m_buckets[previous_ibucket].empty());
+
+      const distance_type new_distance =
+          distance_type(m_buckets[ibucket].dist_from_ideal_bucket() - 1);
+      m_buckets[previous_ibucket].set_value_of_empty_bucket(
+          new_distance, m_buckets[ibucket].truncated_hash(),
+          std::move(m_buckets[ibucket].value()));
+      m_buckets[ibucket].clear();
+
+      previous_ibucket = ibucket;
+      ibucket = next_bucket(ibucket);
+    }
+    m_try_shrink_on_next_insert = true;
+  }
+
+  template <class K, class... Args>
+  std::pair<iterator, bool> insert_impl(const K& key,
+                                        Args&&... value_type_args) {
+    const std::size_t hash = hash_key(key);
+
+    std::size_t ibucket = bucket_for_hash(hash);
+    distance_type dist_from_ideal_bucket = 0;
+
+    while (dist_from_ideal_bucket <=
+           m_buckets[ibucket].dist_from_ideal_bucket()) {
+      if ((!USE_STORED_HASH_ON_LOOKUP ||
+           m_buckets[ibucket].bucket_hash_equal(hash)) &&
+          compare_keys(KeySelect()(m_buckets[ibucket].value()), key)) {
+        return std::make_pair(iterator(m_buckets + ibucket), false);
+      }
+
+      ibucket = next_bucket(ibucket);
+      dist_from_ideal_bucket++;
+    }
+
+    while (rehash_on_extreme_load(dist_from_ideal_bucket)) {
+      ibucket = bucket_for_hash(hash);
+      dist_from_ideal_bucket = 0;
+
+      while (dist_from_ideal_bucket <=
+             m_buckets[ibucket].dist_from_ideal_bucket()) {
+        ibucket = next_bucket(ibucket);
+        dist_from_ideal_bucket++;
+      }
+    }
+
+    if (m_buckets[ibucket].empty()) {
+      m_buckets[ibucket].set_value_of_empty_bucket(
+          dist_from_ideal_bucket, bucket_entry::truncate_hash(hash),
+          std::forward<Args>(value_type_args)...);
+    } else {
+      insert_value(ibucket, dist_from_ideal_bucket,
+                   bucket_entry::truncate_hash(hash),
+                   std::forward<Args>(value_type_args)...);
+    }
+
+    m_nb_elements++;
+    /*
+     * The value will be inserted in ibucket in any case, either because it was
+     * empty or by stealing the bucket (robin hood).
+     */
+    return std::make_pair(iterator(m_buckets + ibucket), true);
+  }
+
+  template <class... Args>
+  void insert_value(std::size_t ibucket, distance_type dist_from_ideal_bucket,
+                    truncated_hash_type hash, Args&&... value_type_args) {
+    value_type value(std::forward<Args>(value_type_args)...);
+    insert_value_impl(ibucket, dist_from_ideal_bucket, hash, value);
+  }
+
+  void insert_value(std::size_t ibucket, distance_type dist_from_ideal_bucket,
+                    truncated_hash_type hash, value_type&& value) {
+    insert_value_impl(ibucket, dist_from_ideal_bucket, hash, value);
+  }
+
+  /*
+   * We don't use `value_type&& value` as last argument due to a bug in MSVC
+   * when `value_type` is a pointer, The compiler is not able to see the
+   * difference between `std::string*` and `std::string*&&` resulting in a
+   * compilation error.
+   *
+   * The `value` will be in a moved state at the end of the function.
+   */
+  void insert_value_impl(std::size_t ibucket,
+                         distance_type dist_from_ideal_bucket,
+                         truncated_hash_type hash, value_type& value) {
+    tsl_rh_assert(dist_from_ideal_bucket >
+                  m_buckets[ibucket].dist_from_ideal_bucket());
+    m_buckets[ibucket].swap_with_value_in_bucket(dist_from_ideal_bucket, hash,
+                                                 value);
+    ibucket = next_bucket(ibucket);
+    dist_from_ideal_bucket++;
+
+    while (!m_buckets[ibucket].empty()) {
+      if (dist_from_ideal_bucket >
+          m_buckets[ibucket].dist_from_ideal_bucket()) {
+        if (dist_from_ideal_bucket >
+            bucket_entry::DIST_FROM_IDEAL_BUCKET_LIMIT) {
+          /**
+           * The number of probes is really high, rehash the map on the next
+           * insert. Difficult to do now as rehash may throw an exception.
+           */
+          m_grow_on_next_insert = true;
+        }
+
+        m_buckets[ibucket].swap_with_value_in_bucket(dist_from_ideal_bucket,
+                                                     hash, value);
+      }
+
+      ibucket = next_bucket(ibucket);
+      dist_from_ideal_bucket++;
+    }
+
+    m_buckets[ibucket].set_value_of_empty_bucket(dist_from_ideal_bucket, hash,
+                                                 std::move(value));
+  }
+
+  void rehash_impl(size_type count_) {
+    robin_hash new_table(count_, static_cast<Hash&>(*this),
+                         static_cast<KeyEqual&>(*this), get_allocator(),
+                         m_min_load_factor, m_max_load_factor);
+    tsl_rh_assert(size() <= new_table.m_load_threshold);
+
+    const bool use_stored_hash =
+        USE_STORED_HASH_ON_REHASH(new_table.bucket_count());
+    for (auto& bucket : m_buckets_data) {
+      if (bucket.empty()) {
+        continue;
+      }
+
+      const std::size_t hash =
+          use_stored_hash ? bucket.truncated_hash()
+                          : new_table.hash_key(KeySelect()(bucket.value()));
+
+      new_table.insert_value_on_rehash(new_table.bucket_for_hash(hash), 0,
+                                       bucket_entry::truncate_hash(hash),
+                                       std::move(bucket.value()));
+    }
+
+    new_table.m_nb_elements = m_nb_elements;
+    new_table.swap(*this);
+  }
+
+  void clear_and_shrink() noexcept {
+    GrowthPolicy::clear();
+    m_buckets_data.clear();
+    m_buckets = static_empty_bucket_ptr();
+    m_bucket_count = 0;
+    m_nb_elements = 0;
+    m_load_threshold = 0;
+    m_grow_on_next_insert = false;
+    m_try_shrink_on_next_insert = false;
+  }
+
+  void insert_value_on_rehash(std::size_t ibucket,
+                              distance_type dist_from_ideal_bucket,
+                              truncated_hash_type hash, value_type&& value) {
+    while (true) {
+      if (dist_from_ideal_bucket >
+          m_buckets[ibucket].dist_from_ideal_bucket()) {
+        if (m_buckets[ibucket].empty()) {
+          m_buckets[ibucket].set_value_of_empty_bucket(dist_from_ideal_bucket,
+                                                       hash, std::move(value));
+          return;
+        } else {
+          m_buckets[ibucket].swap_with_value_in_bucket(dist_from_ideal_bucket,
+                                                       hash, value);
+        }
+      }
+
+      dist_from_ideal_bucket++;
+      ibucket = next_bucket(ibucket);
+    }
+  }
+
+  /**
+   * Grow the table if m_grow_on_next_insert is true or we reached the
+   * max_load_factor. Shrink the table if m_try_shrink_on_next_insert is true
+   * (an erase occurred) and we're below the min_load_factor.
+   *
+   * Return true if the table has been rehashed.
+   */
+  bool rehash_on_extreme_load(distance_type curr_dist_from_ideal_bucket) {
+    if (m_grow_on_next_insert ||
+        curr_dist_from_ideal_bucket >
+            bucket_entry::DIST_FROM_IDEAL_BUCKET_LIMIT ||
+        size() >= m_load_threshold) {
+      rehash_impl(GrowthPolicy::next_bucket_count());
+      m_grow_on_next_insert = false;
+
+      return true;
+    }
+
+    if (m_try_shrink_on_next_insert) {
+      m_try_shrink_on_next_insert = false;
+      if (m_min_load_factor != 0.0f && load_factor() < m_min_load_factor) {
+        reserve(size() + 1);
+
+        return true;
+      }
+    }
+
+    return false;
+  }
+
+  template <class Serializer>
+  void serialize_impl(Serializer& serializer) const {
+    const slz_size_type version = SERIALIZATION_PROTOCOL_VERSION;
+    serializer(version);
+
+    // Indicate if the truncated hash of each bucket is stored. Use a
+    // std::int16_t instead of a bool to avoid the need for the serializer to
+    // support an extra 'bool' type.
+    const std::int16_t hash_stored_for_bucket =
+        static_cast<std::int16_t>(STORE_HASH);
+    serializer(hash_stored_for_bucket);
+
+    const slz_size_type nb_elements = m_nb_elements;
+    serializer(nb_elements);
+
+    const slz_size_type bucket_count = m_buckets_data.size();
+    serializer(bucket_count);
+
+    const float min_load_factor = m_min_load_factor;
+    serializer(min_load_factor);
+
+    const float max_load_factor = m_max_load_factor;
+    serializer(max_load_factor);
+
+    for (const bucket_entry& bucket : m_buckets_data) {
+      if (bucket.empty()) {
+        const std::int16_t empty_bucket =
+            bucket_entry::EMPTY_MARKER_DIST_FROM_IDEAL_BUCKET;
+        serializer(empty_bucket);
+      } else {
+        const std::int16_t dist_from_ideal_bucket =
+            bucket.dist_from_ideal_bucket();
+        serializer(dist_from_ideal_bucket);
+        if (STORE_HASH) {
+          const std::uint32_t truncated_hash = bucket.truncated_hash();
+          serializer(truncated_hash);
+        }
+        serializer(bucket.value());
+      }
+    }
+  }
+
+  template <class Deserializer>
+  void deserialize_impl(Deserializer& deserializer, bool hash_compatible) {
+    tsl_rh_assert(m_buckets_data.empty());  // Current hash table must be empty
+
+    const slz_size_type version =
+        deserialize_value<slz_size_type>(deserializer);
+    // For now we only have one version of the serialization protocol.
+    // If it doesn't match there is a problem with the file.
+    if (version != SERIALIZATION_PROTOCOL_VERSION) {
+      TSL_RH_THROW_OR_TERMINATE(std::runtime_error,
+                                "Can't deserialize the ordered_map/set. "
+                                "The protocol version header is invalid.");
+    }
+
+    const bool hash_stored_for_bucket =
+        deserialize_value<std::int16_t>(deserializer) ? true : false;
+    if (hash_compatible && STORE_HASH != hash_stored_for_bucket) {
+      TSL_RH_THROW_OR_TERMINATE(
+          std::runtime_error,
+          "Can't deserialize a map with a different StoreHash "
+          "than the one used during the serialization when "
+          "hash compatibility is used");
+    }
+
+    const slz_size_type nb_elements =
+        deserialize_value<slz_size_type>(deserializer);
+    const slz_size_type bucket_count_ds =
+        deserialize_value<slz_size_type>(deserializer);
+    const float min_load_factor = deserialize_value<float>(deserializer);
+    const float max_load_factor = deserialize_value<float>(deserializer);
+
+    if (min_load_factor < MINIMUM_MIN_LOAD_FACTOR ||
+        min_load_factor > MAXIMUM_MIN_LOAD_FACTOR) {
+      TSL_RH_THROW_OR_TERMINATE(
+          std::runtime_error,
+          "Invalid min_load_factor. Check that the serializer "
+          "and deserializer support floats correctly as they "
+          "can be converted implicitly to ints.");
+    }
+
+    if (max_load_factor < MINIMUM_MAX_LOAD_FACTOR ||
+        max_load_factor > MAXIMUM_MAX_LOAD_FACTOR) {
+      TSL_RH_THROW_OR_TERMINATE(
+          std::runtime_error,
+          "Invalid max_load_factor. Check that the serializer "
+          "and deserializer support floats correctly as they "
+          "can be converted implicitly to ints.");
+    }
+
+    this->min_load_factor(min_load_factor);
+    this->max_load_factor(max_load_factor);
+
+    if (bucket_count_ds == 0) {
+      tsl_rh_assert(nb_elements == 0);
+      return;
+    }
+
+    if (!hash_compatible) {
+      reserve(numeric_cast<size_type>(nb_elements,
+                                      "Deserialized nb_elements is too big."));
+      for (slz_size_type ibucket = 0; ibucket < bucket_count_ds; ibucket++) {
+        const distance_type dist_from_ideal_bucket =
+            deserialize_value<std::int16_t>(deserializer);
+        if (dist_from_ideal_bucket !=
+            bucket_entry::EMPTY_MARKER_DIST_FROM_IDEAL_BUCKET) {
+          if (hash_stored_for_bucket) {
+            TSL_RH_UNUSED(deserialize_value<std::uint32_t>(deserializer));
+          }
+
+          insert(deserialize_value<value_type>(deserializer));
+        }
+      }
+
+      tsl_rh_assert(nb_elements == size());
+    } else {
+      m_bucket_count = numeric_cast<size_type>(
+          bucket_count_ds, "Deserialized bucket_count is too big.");
+
+      GrowthPolicy::operator=(GrowthPolicy(m_bucket_count));
+      // GrowthPolicy should not modify the bucket count we got from
+      // deserialization
+      if (m_bucket_count != bucket_count_ds) {
+        TSL_RH_THROW_OR_TERMINATE(std::runtime_error,
+                                  "The GrowthPolicy is not the same even "
+                                  "though hash_compatible is true.");
+      }
+
+      m_nb_elements = numeric_cast<size_type>(
+          nb_elements, "Deserialized nb_elements is too big.");
+      m_buckets_data.resize(m_bucket_count);
+      m_buckets = m_buckets_data.data();
+
+      for (bucket_entry& bucket : m_buckets_data) {
+        const distance_type dist_from_ideal_bucket =
+            deserialize_value<std::int16_t>(deserializer);
+        if (dist_from_ideal_bucket !=
+            bucket_entry::EMPTY_MARKER_DIST_FROM_IDEAL_BUCKET) {
+          truncated_hash_type truncated_hash = 0;
+          if (hash_stored_for_bucket) {
+            tsl_rh_assert(hash_stored_for_bucket);
+            truncated_hash = deserialize_value<std::uint32_t>(deserializer);
+          }
+
+          bucket.set_value_of_empty_bucket(
+              dist_from_ideal_bucket, truncated_hash,
+              deserialize_value<value_type>(deserializer));
+        }
+      }
+
+      if (!m_buckets_data.empty()) {
+        m_buckets_data.back().set_as_last_bucket();
+      }
+    }
+  }
+
+ public:
+  static const size_type DEFAULT_INIT_BUCKETS_SIZE = 0;
+
+  static constexpr float DEFAULT_MAX_LOAD_FACTOR = 0.5f;
+  static constexpr float MINIMUM_MAX_LOAD_FACTOR = 0.2f;
+  static constexpr float MAXIMUM_MAX_LOAD_FACTOR = 0.95f;
+
+  static constexpr float DEFAULT_MIN_LOAD_FACTOR = 0.0f;
+  static constexpr float MINIMUM_MIN_LOAD_FACTOR = 0.0f;
+  static constexpr float MAXIMUM_MIN_LOAD_FACTOR = 0.15f;
+
+  static_assert(MINIMUM_MAX_LOAD_FACTOR < MAXIMUM_MAX_LOAD_FACTOR,
+                "MINIMUM_MAX_LOAD_FACTOR should be < MAXIMUM_MAX_LOAD_FACTOR");
+  static_assert(MINIMUM_MIN_LOAD_FACTOR < MAXIMUM_MIN_LOAD_FACTOR,
+                "MINIMUM_MIN_LOAD_FACTOR should be < MAXIMUM_MIN_LOAD_FACTOR");
+  static_assert(MAXIMUM_MIN_LOAD_FACTOR < MINIMUM_MAX_LOAD_FACTOR,
+                "MAXIMUM_MIN_LOAD_FACTOR should be < MINIMUM_MAX_LOAD_FACTOR");
+
+ private:
+  /**
+   * Protocol version currenlty used for serialization.
+   */
+  static const slz_size_type SERIALIZATION_PROTOCOL_VERSION = 1;
+
+  /**
+   * Return an always valid pointer to an static empty bucket_entry with
+   * last_bucket() == true.
+   */
+  bucket_entry* static_empty_bucket_ptr() noexcept {
+    static bucket_entry empty_bucket(true);
+    tsl_rh_assert(empty_bucket.empty());
+    return &empty_bucket;
+  }
+
+ private:
+  buckets_container_type m_buckets_data;
+
+  /**
+   * Points to m_buckets_data.data() if !m_buckets_data.empty() otherwise points
+   * to static_empty_bucket_ptr. This variable is useful to avoid the cost of
+   * checking if m_buckets_data is empty when trying to find an element.
+   *
+   * TODO Remove m_buckets_data and only use a pointer instead of a
+   * pointer+vector to save some space in the robin_hash object. Manage the
+   * Allocator manually.
+   */
+  bucket_entry* m_buckets;
+
+  /**
+   * Used a lot in find, avoid the call to m_buckets_data.size() which is a bit
+   * slower.
+   */
+  size_type m_bucket_count;
+
+  size_type m_nb_elements;
+
+  size_type m_load_threshold;
+
+  float m_min_load_factor;
+  float m_max_load_factor;
+
+  bool m_grow_on_next_insert;
+
+  /**
+   * We can't shrink down the map on erase operations as the erase methods need
+   * to return the next iterator. Shrinking the map would invalidate all the
+   * iterators and we could not return the next iterator in a meaningful way, On
+   * erase, we thus just indicate on erase that we should try to shrink the hash
+   * table on the next insert if we go below the min_load_factor.
+   */
+  bool m_try_shrink_on_next_insert;
+};
+
+}  // namespace detail_robin_hash
+
+}  // namespace tsl
+
+#endif
diff --git a/ext/robin-map/robin_map.h b/ext/robin-map/robin_map.h
new file mode 100644
index 0000000..b594810
--- /dev/null
+++ b/ext/robin-map/robin_map.h
@@ -0,0 +1,815 @@
+/**
+ * MIT License
+ *
+ * Copyright (c) 2017 Thibaut Goetghebuer-Planchon <tessil@gmx.com>
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#ifndef TSL_ROBIN_MAP_H
+#define TSL_ROBIN_MAP_H
+
+#include <cstddef>
+#include <functional>
+#include <initializer_list>
+#include <memory>
+#include <type_traits>
+#include <utility>
+
+#include "robin_hash.h"
+
+namespace tsl {
+
+/**
+ * Implementation of a hash map using open-addressing and the robin hood hashing
+ * algorithm with backward shift deletion.
+ *
+ * For operations modifying the hash map (insert, erase, rehash, ...), the
+ * strong exception guarantee is only guaranteed when the expression
+ * `std::is_nothrow_swappable<std::pair<Key, T>>::value &&
+ * std::is_nothrow_move_constructible<std::pair<Key, T>>::value` is true,
+ * otherwise if an exception is thrown during the swap or the move, the hash map
+ * may end up in a undefined state. Per the standard a `Key` or `T` with a
+ * noexcept copy constructor and no move constructor also satisfies the
+ * `std::is_nothrow_move_constructible<std::pair<Key, T>>::value` criterion (and
+ * will thus guarantee the strong exception for the map).
+ *
+ * When `StoreHash` is true, 32 bits of the hash are stored alongside the
+ * values. It can improve the performance during lookups if the `KeyEqual`
+ * function takes time (if it engenders a cache-miss for example) as we then
+ * compare the stored hashes before comparing the keys. When
+ * `tsl::rh::power_of_two_growth_policy` is used as `GrowthPolicy`, it may also
+ * speed-up the rehash process as we can avoid to recalculate the hash. When it
+ * is detected that storing the hash will not incur any memory penalty due to
+ * alignment (i.e. `sizeof(tsl::detail_robin_hash::bucket_entry<ValueType,
+ * true>) == sizeof(tsl::detail_robin_hash::bucket_entry<ValueType, false>)`)
+ * and `tsl::rh::power_of_two_growth_policy` is used, the hash will be stored
+ * even if `StoreHash` is false so that we can speed-up the rehash (but it will
+ * not be used on lookups unless `StoreHash` is true).
+ *
+ * `GrowthPolicy` defines how the map grows and consequently how a hash value is
+ * mapped to a bucket. By default the map uses
+ * `tsl::rh::power_of_two_growth_policy`. This policy keeps the number of
+ * buckets to a power of two and uses a mask to map the hash to a bucket instead
+ * of the slow modulo. Other growth policies are available and you may define
+ * your own growth policy, check `tsl::rh::power_of_two_growth_policy` for the
+ * interface.
+ *
+ * `std::pair<Key, T>` must be swappable.
+ *
+ * `Key` and `T` must be copy and/or move constructible.
+ *
+ * If the destructor of `Key` or `T` throws an exception, the behaviour of the
+ * class is undefined.
+ *
+ * Iterators invalidation:
+ *  - clear, operator=, reserve, rehash: always invalidate the iterators.
+ *  - insert, emplace, emplace_hint, operator[]: if there is an effective
+ * insert, invalidate the iterators.
+ *  - erase: always invalidate the iterators.
+ */
+template <class Key, class T, class Hash = std::hash<Key>,
+          class KeyEqual = std::equal_to<Key>,
+          class Allocator = std::allocator<std::pair<Key, T>>,
+          bool StoreHash = false,
+          class GrowthPolicy = tsl::rh::power_of_two_growth_policy<2>>
+class robin_map {
+ private:
+  template <typename U>
+  using has_is_transparent = tsl::detail_robin_hash::has_is_transparent<U>;
+
+  class KeySelect {
+   public:
+    using key_type = Key;
+
+    const key_type& operator()(
+        const std::pair<Key, T>& key_value) const noexcept {
+      return key_value.first;
+    }
+
+    key_type& operator()(std::pair<Key, T>& key_value) noexcept {
+      return key_value.first;
+    }
+  };
+
+  class ValueSelect {
+   public:
+    using value_type = T;
+
+    const value_type& operator()(
+        const std::pair<Key, T>& key_value) const noexcept {
+      return key_value.second;
+    }
+
+    value_type& operator()(std::pair<Key, T>& key_value) noexcept {
+      return key_value.second;
+    }
+  };
+
+  using ht = detail_robin_hash::robin_hash<std::pair<Key, T>, KeySelect,
+                                           ValueSelect, Hash, KeyEqual,
+                                           Allocator, StoreHash, GrowthPolicy>;
+
+ public:
+  using key_type = typename ht::key_type;
+  using mapped_type = T;
+  using value_type = typename ht::value_type;
+  using size_type = typename ht::size_type;
+  using difference_type = typename ht::difference_type;
+  using hasher = typename ht::hasher;
+  using key_equal = typename ht::key_equal;
+  using allocator_type = typename ht::allocator_type;
+  using reference = typename ht::reference;
+  using const_reference = typename ht::const_reference;
+  using pointer = typename ht::pointer;
+  using const_pointer = typename ht::const_pointer;
+  using iterator = typename ht::iterator;
+  using const_iterator = typename ht::const_iterator;
+
+ public:
+  /*
+   * Constructors
+   */
+  robin_map() : robin_map(ht::DEFAULT_INIT_BUCKETS_SIZE) {}
+
+  explicit robin_map(size_type bucket_count, const Hash& hash = Hash(),
+                     const KeyEqual& equal = KeyEqual(),
+                     const Allocator& alloc = Allocator())
+      : m_ht(bucket_count, hash, equal, alloc) {}
+
+  robin_map(size_type bucket_count, const Allocator& alloc)
+      : robin_map(bucket_count, Hash(), KeyEqual(), alloc) {}
+
+  robin_map(size_type bucket_count, const Hash& hash, const Allocator& alloc)
+      : robin_map(bucket_count, hash, KeyEqual(), alloc) {}
+
+  explicit robin_map(const Allocator& alloc)
+      : robin_map(ht::DEFAULT_INIT_BUCKETS_SIZE, alloc) {}
+
+  template <class InputIt>
+  robin_map(InputIt first, InputIt last,
+            size_type bucket_count = ht::DEFAULT_INIT_BUCKETS_SIZE,
+            const Hash& hash = Hash(), const KeyEqual& equal = KeyEqual(),
+            const Allocator& alloc = Allocator())
+      : robin_map(bucket_count, hash, equal, alloc) {
+    insert(first, last);
+  }
+
+  template <class InputIt>
+  robin_map(InputIt first, InputIt last, size_type bucket_count,
+            const Allocator& alloc)
+      : robin_map(first, last, bucket_count, Hash(), KeyEqual(), alloc) {}
+
+  template <class InputIt>
+  robin_map(InputIt first, InputIt last, size_type bucket_count,
+            const Hash& hash, const Allocator& alloc)
+      : robin_map(first, last, bucket_count, hash, KeyEqual(), alloc) {}
+
+  robin_map(std::initializer_list<value_type> init,
+            size_type bucket_count = ht::DEFAULT_INIT_BUCKETS_SIZE,
+            const Hash& hash = Hash(), const KeyEqual& equal = KeyEqual(),
+            const Allocator& alloc = Allocator())
+      : robin_map(init.begin(), init.end(), bucket_count, hash, equal, alloc) {}
+
+  robin_map(std::initializer_list<value_type> init, size_type bucket_count,
+            const Allocator& alloc)
+      : robin_map(init.begin(), init.end(), bucket_count, Hash(), KeyEqual(),
+                  alloc) {}
+
+  robin_map(std::initializer_list<value_type> init, size_type bucket_count,
+            const Hash& hash, const Allocator& alloc)
+      : robin_map(init.begin(), init.end(), bucket_count, hash, KeyEqual(),
+                  alloc) {}
+
+  robin_map& operator=(std::initializer_list<value_type> ilist) {
+    m_ht.clear();
+
+    m_ht.reserve(ilist.size());
+    m_ht.insert(ilist.begin(), ilist.end());
+
+    return *this;
+  }
+
+  allocator_type get_allocator() const { return m_ht.get_allocator(); }
+
+  /*
+   * Iterators
+   */
+  iterator begin() noexcept { return m_ht.begin(); }
+  const_iterator begin() const noexcept { return m_ht.begin(); }
+  const_iterator cbegin() const noexcept { return m_ht.cbegin(); }
+
+  iterator end() noexcept { return m_ht.end(); }
+  const_iterator end() const noexcept { return m_ht.end(); }
+  const_iterator cend() const noexcept { return m_ht.cend(); }
+
+  /*
+   * Capacity
+   */
+  bool empty() const noexcept { return m_ht.empty(); }
+  size_type size() const noexcept { return m_ht.size(); }
+  size_type max_size() const noexcept { return m_ht.max_size(); }
+
+  /*
+   * Modifiers
+   */
+  void clear() noexcept { m_ht.clear(); }
+
+  std::pair<iterator, bool> insert(const value_type& value) {
+    return m_ht.insert(value);
+  }
+
+  template <class P, typename std::enable_if<std::is_constructible<
+                         value_type, P&&>::value>::type* = nullptr>
+  std::pair<iterator, bool> insert(P&& value) {
+    return m_ht.emplace(std::forward<P>(value));
+  }
+
+  std::pair<iterator, bool> insert(value_type&& value) {
+    return m_ht.insert(std::move(value));
+  }
+
+  iterator insert(const_iterator hint, const value_type& value) {
+    return m_ht.insert_hint(hint, value);
+  }
+
+  template <class P, typename std::enable_if<std::is_constructible<
+                         value_type, P&&>::value>::type* = nullptr>
+  iterator insert(const_iterator hint, P&& value) {
+    return m_ht.emplace_hint(hint, std::forward<P>(value));
+  }
+
+  iterator insert(const_iterator hint, value_type&& value) {
+    return m_ht.insert_hint(hint, std::move(value));
+  }
+
+  template <class InputIt>
+  void insert(InputIt first, InputIt last) {
+    m_ht.insert(first, last);
+  }
+
+  void insert(std::initializer_list<value_type> ilist) {
+    m_ht.insert(ilist.begin(), ilist.end());
+  }
+
+  template <class M>
+  std::pair<iterator, bool> insert_or_assign(const key_type& k, M&& obj) {
+    return m_ht.insert_or_assign(k, std::forward<M>(obj));
+  }
+
+  template <class M>
+  std::pair<iterator, bool> insert_or_assign(key_type&& k, M&& obj) {
+    return m_ht.insert_or_assign(std::move(k), std::forward<M>(obj));
+  }
+
+  template <class M>
+  iterator insert_or_assign(const_iterator hint, const key_type& k, M&& obj) {
+    return m_ht.insert_or_assign(hint, k, std::forward<M>(obj));
+  }
+
+  template <class M>
+  iterator insert_or_assign(const_iterator hint, key_type&& k, M&& obj) {
+    return m_ht.insert_or_assign(hint, std::move(k), std::forward<M>(obj));
+  }
+
+  /**
+   * Due to the way elements are stored, emplace will need to move or copy the
+   * key-value once. The method is equivalent to
+   * insert(value_type(std::forward<Args>(args)...));
+   *
+   * Mainly here for compatibility with the std::unordered_map interface.
+   */
+  template <class... Args>
+  std::pair<iterator, bool> emplace(Args&&... args) {
+    return m_ht.emplace(std::forward<Args>(args)...);
+  }
+
+  /**
+   * Due to the way elements are stored, emplace_hint will need to move or copy
+   * the key-value once. The method is equivalent to insert(hint,
+   * value_type(std::forward<Args>(args)...));
+   *
+   * Mainly here for compatibility with the std::unordered_map interface.
+   */
+  template <class... Args>
+  iterator emplace_hint(const_iterator hint, Args&&... args) {
+    return m_ht.emplace_hint(hint, std::forward<Args>(args)...);
+  }
+
+  template <class... Args>
+  std::pair<iterator, bool> try_emplace(const key_type& k, Args&&... args) {
+    return m_ht.try_emplace(k, std::forward<Args>(args)...);
+  }
+
+  template <class... Args>
+  std::pair<iterator, bool> try_emplace(key_type&& k, Args&&... args) {
+    return m_ht.try_emplace(std::move(k), std::forward<Args>(args)...);
+  }
+
+  template <class... Args>
+  iterator try_emplace(const_iterator hint, const key_type& k, Args&&... args) {
+    return m_ht.try_emplace_hint(hint, k, std::forward<Args>(args)...);
+  }
+
+  template <class... Args>
+  iterator try_emplace(const_iterator hint, key_type&& k, Args&&... args) {
+    return m_ht.try_emplace_hint(hint, std::move(k),
+                                 std::forward<Args>(args)...);
+  }
+
+  iterator erase(iterator pos) { return m_ht.erase(pos); }
+  iterator erase(const_iterator pos) { return m_ht.erase(pos); }
+  iterator erase(const_iterator first, const_iterator last) {
+    return m_ht.erase(first, last);
+  }
+  size_type erase(const key_type& key) { return m_ht.erase(key); }
+
+  /**
+   * Erase the element at position 'pos'. In contrast to the regular erase()
+   * function, erase_fast() does not return an iterator. This allows it to be
+   * faster especially in hash tables with a low load factor, where finding the
+   * next nonempty bucket would be costly.
+   */
+  void erase_fast(iterator pos) { return m_ht.erase_fast(pos); }
+
+  /**
+   * Use the hash value 'precalculated_hash' instead of hashing the key. The
+   * hash value should be the same as hash_function()(key). Useful to speed-up
+   * the lookup to the value if you already have the hash.
+   */
+  size_type erase(const key_type& key, std::size_t precalculated_hash) {
+    return m_ht.erase(key, precalculated_hash);
+  }
+
+  /**
+   * This overload only participates in the overload resolution if the typedef
+   * KeyEqual::is_transparent exists. If so, K must be hashable and comparable
+   * to Key.
+   */
+  template <
+      class K, class KE = KeyEqual,
+      typename std::enable_if<has_is_transparent<KE>::value>::type* = nullptr>
+  size_type erase(const K& key) {
+    return m_ht.erase(key);
+  }
+
+  /**
+   * @copydoc erase(const K& key)
+   *
+   * Use the hash value 'precalculated_hash' instead of hashing the key. The
+   * hash value should be the same as hash_function()(key). Useful to speed-up
+   * the lookup to the value if you already have the hash.
+   */
+  template <
+      class K, class KE = KeyEqual,
+      typename std::enable_if<has_is_transparent<KE>::value>::type* = nullptr>
+  size_type erase(const K& key, std::size_t precalculated_hash) {
+    return m_ht.erase(key, precalculated_hash);
+  }
+
+  void swap(robin_map& other) { other.m_ht.swap(m_ht); }
+
+  /*
+   * Lookup
+   */
+  T& at(const Key& key) { return m_ht.at(key); }
+
+  /**
+   * Use the hash value 'precalculated_hash' instead of hashing the key. The
+   * hash value should be the same as hash_function()(key). Useful to speed-up
+   * the lookup if you already have the hash.
+   */
+  T& at(const Key& key, std::size_t precalculated_hash) {
+    return m_ht.at(key, precalculated_hash);
+  }
+
+  const T& at(const Key& key) const { return m_ht.at(key); }
+
+  /**
+   * @copydoc at(const Key& key, std::size_t precalculated_hash)
+   */
+  const T& at(const Key& key, std::size_t precalculated_hash) const {
+    return m_ht.at(key, precalculated_hash);
+  }
+
+  /**
+   * This overload only participates in the overload resolution if the typedef
+   * KeyEqual::is_transparent exists. If so, K must be hashable and comparable
+   * to Key.
+   */
+  template <
+      class K, class KE = KeyEqual,
+      typename std::enable_if<has_is_transparent<KE>::value>::type* = nullptr>
+  T& at(const K& key) {
+    return m_ht.at(key);
+  }
+
+  /**
+   * @copydoc at(const K& key)
+   *
+   * Use the hash value 'precalculated_hash' instead of hashing the key. The
+   * hash value should be the same as hash_function()(key). Useful to speed-up
+   * the lookup if you already have the hash.
+   */
+  template <
+      class K, class KE = KeyEqual,
+      typename std::enable_if<has_is_transparent<KE>::value>::type* = nullptr>
+  T& at(const K& key, std::size_t precalculated_hash) {
+    return m_ht.at(key, precalculated_hash);
+  }
+
+  /**
+   * @copydoc at(const K& key)
+   */
+  template <
+      class K, class KE = KeyEqual,
+      typename std::enable_if<has_is_transparent<KE>::value>::type* = nullptr>
+  const T& at(const K& key) const {
+    return m_ht.at(key);
+  }
+
+  /**
+   * @copydoc at(const K& key, std::size_t precalculated_hash)
+   */
+  template <
+      class K, class KE = KeyEqual,
+      typename std::enable_if<has_is_transparent<KE>::value>::type* = nullptr>
+  const T& at(const K& key, std::size_t precalculated_hash) const {
+    return m_ht.at(key, precalculated_hash);
+  }
+
+  T& operator[](const Key& key) { return m_ht[key]; }
+  T& operator[](Key&& key) { return m_ht[std::move(key)]; }
+
+  size_type count(const Key& key) const { return m_ht.count(key); }
+
+  /**
+   * Use the hash value 'precalculated_hash' instead of hashing the key. The
+   * hash value should be the same as hash_function()(key). Useful to speed-up
+   * the lookup if you already have the hash.
+   */
+  size_type count(const Key& key, std::size_t precalculated_hash) const {
+    return m_ht.count(key, precalculated_hash);
+  }
+
+  /**
+   * This overload only participates in the overload resolution if the typedef
+   * KeyEqual::is_transparent exists. If so, K must be hashable and comparable
+   * to Key.
+   */
+  template <
+      class K, class KE = KeyEqual,
+      typename std::enable_if<has_is_transparent<KE>::value>::type* = nullptr>
+  size_type count(const K& key) const {
+    return m_ht.count(key);
+  }
+
+  /**
+   * @copydoc count(const K& key) const
+   *
+   * Use the hash value 'precalculated_hash' instead of hashing the key. The
+   * hash value should be the same as hash_function()(key). Useful to speed-up
+   * the lookup if you already have the hash.
+   */
+  template <
+      class K, class KE = KeyEqual,
+      typename std::enable_if<has_is_transparent<KE>::value>::type* = nullptr>
+  size_type count(const K& key, std::size_t precalculated_hash) const {
+    return m_ht.count(key, precalculated_hash);
+  }
+
+  iterator find(const Key& key) { return m_ht.find(key); }
+
+  /**
+   * Use the hash value 'precalculated_hash' instead of hashing the key. The
+   * hash value should be the same as hash_function()(key). Useful to speed-up
+   * the lookup if you already have the hash.
+   */
+  iterator find(const Key& key, std::size_t precalculated_hash) {
+    return m_ht.find(key, precalculated_hash);
+  }
+
+  const_iterator find(const Key& key) const { return m_ht.find(key); }
+
+  /**
+   * @copydoc find(const Key& key, std::size_t precalculated_hash)
+   */
+  const_iterator find(const Key& key, std::size_t precalculated_hash) const {
+    return m_ht.find(key, precalculated_hash);
+  }
+
+  /**
+   * This overload only participates in the overload resolution if the typedef
+   * KeyEqual::is_transparent exists. If so, K must be hashable and comparable
+   * to Key.
+   */
+  template <
+      class K, class KE = KeyEqual,
+      typename std::enable_if<has_is_transparent<KE>::value>::type* = nullptr>
+  iterator find(const K& key) {
+    return m_ht.find(key);
+  }
+
+  /**
+   * @copydoc find(const K& key)
+   *
+   * Use the hash value 'precalculated_hash' instead of hashing the key. The
+   * hash value should be the same as hash_function()(key). Useful to speed-up
+   * the lookup if you already have the hash.
+   */
+  template <
+      class K, class KE = KeyEqual,
+      typename std::enable_if<has_is_transparent<KE>::value>::type* = nullptr>
+  iterator find(const K& key, std::size_t precalculated_hash) {
+    return m_ht.find(key, precalculated_hash);
+  }
+
+  /**
+   * @copydoc find(const K& key)
+   */
+  template <
+      class K, class KE = KeyEqual,
+      typename std::enable_if<has_is_transparent<KE>::value>::type* = nullptr>
+  const_iterator find(const K& key) const {
+    return m_ht.find(key);
+  }
+
+  /**
+   * @copydoc find(const K& key)
+   *
+   * Use the hash value 'precalculated_hash' instead of hashing the key. The
+   * hash value should be the same as hash_function()(key). Useful to speed-up
+   * the lookup if you already have the hash.
+   */
+  template <
+      class K, class KE = KeyEqual,
+      typename std::enable_if<has_is_transparent<KE>::value>::type* = nullptr>
+  const_iterator find(const K& key, std::size_t precalculated_hash) const {
+    return m_ht.find(key, precalculated_hash);
+  }
+
+  bool contains(const Key& key) const { return m_ht.contains(key); }
+
+  /**
+   * Use the hash value 'precalculated_hash' instead of hashing the key. The
+   * hash value should be the same as hash_function()(key). Useful to speed-up
+   * the lookup if you already have the hash.
+   */
+  bool contains(const Key& key, std::size_t precalculated_hash) const {
+    return m_ht.contains(key, precalculated_hash);
+  }
+
+  /**
+   * This overload only participates in the overload resolution if the typedef
+   * KeyEqual::is_transparent exists. If so, K must be hashable and comparable
+   * to Key.
+   */
+  template <
+      class K, class KE = KeyEqual,
+      typename std::enable_if<has_is_transparent<KE>::value>::type* = nullptr>
+  bool contains(const K& key) const {
+    return m_ht.contains(key);
+  }
+
+  /**
+   * @copydoc contains(const K& key) const
+   *
+   * Use the hash value 'precalculated_hash' instead of hashing the key. The
+   * hash value should be the same as hash_function()(key). Useful to speed-up
+   * the lookup if you already have the hash.
+   */
+  template <
+      class K, class KE = KeyEqual,
+      typename std::enable_if<has_is_transparent<KE>::value>::type* = nullptr>
+  bool contains(const K& key, std::size_t precalculated_hash) const {
+    return m_ht.contains(key, precalculated_hash);
+  }
+
+  std::pair<iterator, iterator> equal_range(const Key& key) {
+    return m_ht.equal_range(key);
+  }
+
+  /**
+   * Use the hash value 'precalculated_hash' instead of hashing the key. The
+   * hash value should be the same as hash_function()(key). Useful to speed-up
+   * the lookup if you already have the hash.
+   */
+  std::pair<iterator, iterator> equal_range(const Key& key,
+                                            std::size_t precalculated_hash) {
+    return m_ht.equal_range(key, precalculated_hash);
+  }
+
+  std::pair<const_iterator, const_iterator> equal_range(const Key& key) const {
+    return m_ht.equal_range(key);
+  }
+
+  /**
+   * @copydoc equal_range(const Key& key, std::size_t precalculated_hash)
+   */
+  std::pair<const_iterator, const_iterator> equal_range(
+      const Key& key, std::size_t precalculated_hash) const {
+    return m_ht.equal_range(key, precalculated_hash);
+  }
+
+  /**
+   * This overload only participates in the overload resolution if the typedef
+   * KeyEqual::is_transparent exists. If so, K must be hashable and comparable
+   * to Key.
+   */
+  template <
+      class K, class KE = KeyEqual,
+      typename std::enable_if<has_is_transparent<KE>::value>::type* = nullptr>
+  std::pair<iterator, iterator> equal_range(const K& key) {
+    return m_ht.equal_range(key);
+  }
+
+  /**
+   * @copydoc equal_range(const K& key)
+   *
+   * Use the hash value 'precalculated_hash' instead of hashing the key. The
+   * hash value should be the same as hash_function()(key). Useful to speed-up
+   * the lookup if you already have the hash.
+   */
+  template <
+      class K, class KE = KeyEqual,
+      typename std::enable_if<has_is_transparent<KE>::value>::type* = nullptr>
+  std::pair<iterator, iterator> equal_range(const K& key,
+                                            std::size_t precalculated_hash) {
+    return m_ht.equal_range(key, precalculated_hash);
+  }
+
+  /**
+   * @copydoc equal_range(const K& key)
+   */
+  template <
+      class K, class KE = KeyEqual,
+      typename std::enable_if<has_is_transparent<KE>::value>::type* = nullptr>
+  std::pair<const_iterator, const_iterator> equal_range(const K& key) const {
+    return m_ht.equal_range(key);
+  }
+
+  /**
+   * @copydoc equal_range(const K& key, std::size_t precalculated_hash)
+   */
+  template <
+      class K, class KE = KeyEqual,
+      typename std::enable_if<has_is_transparent<KE>::value>::type* = nullptr>
+  std::pair<const_iterator, const_iterator> equal_range(
+      const K& key, std::size_t precalculated_hash) const {
+    return m_ht.equal_range(key, precalculated_hash);
+  }
+
+  /*
+   * Bucket interface
+   */
+  size_type bucket_count() const { return m_ht.bucket_count(); }
+  size_type max_bucket_count() const { return m_ht.max_bucket_count(); }
+
+  /*
+   *  Hash policy
+   */
+  float load_factor() const { return m_ht.load_factor(); }
+
+  float min_load_factor() const { return m_ht.min_load_factor(); }
+  float max_load_factor() const { return m_ht.max_load_factor(); }
+
+  /**
+   * Set the `min_load_factor` to `ml`. When the `load_factor` of the map goes
+   * below `min_load_factor` after some erase operations, the map will be
+   * shrunk when an insertion occurs. The erase method itself never shrinks
+   * the map.
+   *
+   * The default value of `min_load_factor` is 0.0f, the map never shrinks by
+   * default.
+   */
+  void min_load_factor(float ml) { m_ht.min_load_factor(ml); }
+  void max_load_factor(float ml) { m_ht.max_load_factor(ml); }
+
+  void rehash(size_type count_) { m_ht.rehash(count_); }
+  void reserve(size_type count_) { m_ht.reserve(count_); }
+
+  /*
+   * Observers
+   */
+  hasher hash_function() const { return m_ht.hash_function(); }
+  key_equal key_eq() const { return m_ht.key_eq(); }
+
+  /*
+   * Other
+   */
+
+  /**
+   * Convert a const_iterator to an iterator.
+   */
+  iterator mutable_iterator(const_iterator pos) {
+    return m_ht.mutable_iterator(pos);
+  }
+
+  /**
+   * Serialize the map through the `serializer` parameter.
+   *
+   * The `serializer` parameter must be a function object that supports the
+   * following call:
+   *  - `template<typename U> void operator()(const U& value);` where the types
+   * `std::int16_t`, `std::uint32_t`, `std::uint64_t`, `float` and
+   * `std::pair<Key, T>` must be supported for U.
+   *
+   * The implementation leaves binary compatibility (endianness, IEEE 754 for
+   * floats, ...) of the types it serializes in the hands of the `Serializer`
+   * function object if compatibility is required.
+   */
+  template <class Serializer>
+  void serialize(Serializer& serializer) const {
+    m_ht.serialize(serializer);
+  }
+
+  /**
+   * Deserialize a previously serialized map through the `deserializer`
+   * parameter.
+   *
+   * The `deserializer` parameter must be a function object that supports the
+   * following call:
+   *  - `template<typename U> U operator()();` where the types `std::int16_t`,
+   * `std::uint32_t`, `std::uint64_t`, `float` and `std::pair<Key, T>` must be
+   * supported for U.
+   *
+   * If the deserialized hash map type is hash compatible with the serialized
+   * map, the deserialization process can be sped up by setting
+   * `hash_compatible` to true. To be hash compatible, the Hash, KeyEqual and
+   * GrowthPolicy must behave the same way than the ones used on the serialized
+   * map and the StoreHash must have the same value. The `std::size_t` must also
+   * be of the same size as the one on the platform used to serialize the map.
+   * If these criteria are not met, the behaviour is undefined with
+   * `hash_compatible` sets to true.
+   *
+   * The behaviour is undefined if the type `Key` and `T` of the `robin_map` are
+   * not the same as the types used during serialization.
+   *
+   * The implementation leaves binary compatibility (endianness, IEEE 754 for
+   * floats, size of int, ...) of the types it deserializes in the hands of the
+   * `Deserializer` function object if compatibility is required.
+   */
+  template <class Deserializer>
+  static robin_map deserialize(Deserializer& deserializer,
+                               bool hash_compatible = false) {
+    robin_map map(0);
+    map.m_ht.deserialize(deserializer, hash_compatible);
+
+    return map;
+  }
+
+  friend bool operator==(const robin_map& lhs, const robin_map& rhs) {
+    if (lhs.size() != rhs.size()) {
+      return false;
+    }
+
+    for (const auto& element_lhs : lhs) {
+      const auto it_element_rhs = rhs.find(element_lhs.first);
+      if (it_element_rhs == rhs.cend() ||
+          element_lhs.second != it_element_rhs->second) {
+        return false;
+      }
+    }
+
+    return true;
+  }
+
+  friend bool operator!=(const robin_map& lhs, const robin_map& rhs) {
+    return !operator==(lhs, rhs);
+  }
+
+  friend void swap(robin_map& lhs, robin_map& rhs) { lhs.swap(rhs); }
+
+ private:
+  ht m_ht;
+};
+
+/**
+ * Same as `tsl::robin_map<Key, T, Hash, KeyEqual, Allocator, StoreHash,
+ * tsl::rh::prime_growth_policy>`.
+ */
+template <class Key, class T, class Hash = std::hash<Key>,
+          class KeyEqual = std::equal_to<Key>,
+          class Allocator = std::allocator<std::pair<Key, T>>,
+          bool StoreHash = false>
+using robin_pg_map = robin_map<Key, T, Hash, KeyEqual, Allocator, StoreHash,
+                               tsl::rh::prime_growth_policy>;
+
+}  // end namespace tsl
+
+#endif
diff --git a/ext/robin-map/robin_set.h b/ext/robin-map/robin_set.h
new file mode 100644
index 0000000..e115007
--- /dev/null
+++ b/ext/robin-map/robin_set.h
@@ -0,0 +1,668 @@
+/**
+ * MIT License
+ *
+ * Copyright (c) 2017 Thibaut Goetghebuer-Planchon <tessil@gmx.com>
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#ifndef TSL_ROBIN_SET_H
+#define TSL_ROBIN_SET_H
+
+#include <cstddef>
+#include <functional>
+#include <initializer_list>
+#include <memory>
+#include <type_traits>
+#include <utility>
+
+#include "robin_hash.h"
+
+namespace tsl {
+
+/**
+ * Implementation of a hash set using open-addressing and the robin hood hashing
+ * algorithm with backward shift deletion.
+ *
+ * For operations modifying the hash set (insert, erase, rehash, ...), the
+ * strong exception guarantee is only guaranteed when the expression
+ * `std::is_nothrow_swappable<Key>::value &&
+ * std::is_nothrow_move_constructible<Key>::value` is true, otherwise if an
+ * exception is thrown during the swap or the move, the hash set may end up in a
+ * undefined state. Per the standard a `Key` with a noexcept copy constructor
+ * and no move constructor also satisfies the
+ * `std::is_nothrow_move_constructible<Key>::value` criterion (and will thus
+ * guarantee the strong exception for the set).
+ *
+ * When `StoreHash` is true, 32 bits of the hash are stored alongside the
+ * values. It can improve the performance during lookups if the `KeyEqual`
+ * function takes time (or engenders a cache-miss for example) as we then
+ * compare the stored hashes before comparing the keys. When
+ * `tsl::rh::power_of_two_growth_policy` is used as `GrowthPolicy`, it may also
+ * speed-up the rehash process as we can avoid to recalculate the hash. When it
+ * is detected that storing the hash will not incur any memory penalty due to
+ * alignment (i.e. `sizeof(tsl::detail_robin_hash::bucket_entry<ValueType,
+ * true>) == sizeof(tsl::detail_robin_hash::bucket_entry<ValueType, false>)`)
+ * and `tsl::rh::power_of_two_growth_policy` is used, the hash will be stored
+ * even if `StoreHash` is false so that we can speed-up the rehash (but it will
+ * not be used on lookups unless `StoreHash` is true).
+ *
+ * `GrowthPolicy` defines how the set grows and consequently how a hash value is
+ * mapped to a bucket. By default the set uses
+ * `tsl::rh::power_of_two_growth_policy`. This policy keeps the number of
+ * buckets to a power of two and uses a mask to set the hash to a bucket instead
+ * of the slow modulo. Other growth policies are available and you may define
+ * your own growth policy, check `tsl::rh::power_of_two_growth_policy` for the
+ * interface.
+ *
+ * `Key` must be swappable.
+ *
+ * `Key` must be copy and/or move constructible.
+ *
+ * If the destructor of `Key` throws an exception, the behaviour of the class is
+ * undefined.
+ *
+ * Iterators invalidation:
+ *  - clear, operator=, reserve, rehash: always invalidate the iterators.
+ *  - insert, emplace, emplace_hint, operator[]: if there is an effective
+ * insert, invalidate the iterators.
+ *  - erase: always invalidate the iterators.
+ */
+template <class Key, class Hash = std::hash<Key>,
+          class KeyEqual = std::equal_to<Key>,
+          class Allocator = std::allocator<Key>, bool StoreHash = false,
+          class GrowthPolicy = tsl::rh::power_of_two_growth_policy<2>>
+class robin_set {
+ private:
+  template <typename U>
+  using has_is_transparent = tsl::detail_robin_hash::has_is_transparent<U>;
+
+  class KeySelect {
+   public:
+    using key_type = Key;
+
+    const key_type& operator()(const Key& key) const noexcept { return key; }
+
+    key_type& operator()(Key& key) noexcept { return key; }
+  };
+
+  using ht = detail_robin_hash::robin_hash<Key, KeySelect, void, Hash, KeyEqual,
+                                           Allocator, StoreHash, GrowthPolicy>;
+
+ public:
+  using key_type = typename ht::key_type;
+  using value_type = typename ht::value_type;
+  using size_type = typename ht::size_type;
+  using difference_type = typename ht::difference_type;
+  using hasher = typename ht::hasher;
+  using key_equal = typename ht::key_equal;
+  using allocator_type = typename ht::allocator_type;
+  using reference = typename ht::reference;
+  using const_reference = typename ht::const_reference;
+  using pointer = typename ht::pointer;
+  using const_pointer = typename ht::const_pointer;
+  using iterator = typename ht::iterator;
+  using const_iterator = typename ht::const_iterator;
+
+  /*
+   * Constructors
+   */
+  robin_set() : robin_set(ht::DEFAULT_INIT_BUCKETS_SIZE) {}
+
+  explicit robin_set(size_type bucket_count, const Hash& hash = Hash(),
+                     const KeyEqual& equal = KeyEqual(),
+                     const Allocator& alloc = Allocator())
+      : m_ht(bucket_count, hash, equal, alloc) {}
+
+  robin_set(size_type bucket_count, const Allocator& alloc)
+      : robin_set(bucket_count, Hash(), KeyEqual(), alloc) {}
+
+  robin_set(size_type bucket_count, const Hash& hash, const Allocator& alloc)
+      : robin_set(bucket_count, hash, KeyEqual(), alloc) {}
+
+  explicit robin_set(const Allocator& alloc)
+      : robin_set(ht::DEFAULT_INIT_BUCKETS_SIZE, alloc) {}
+
+  template <class InputIt>
+  robin_set(InputIt first, InputIt last,
+            size_type bucket_count = ht::DEFAULT_INIT_BUCKETS_SIZE,
+            const Hash& hash = Hash(), const KeyEqual& equal = KeyEqual(),
+            const Allocator& alloc = Allocator())
+      : robin_set(bucket_count, hash, equal, alloc) {
+    insert(first, last);
+  }
+
+  template <class InputIt>
+  robin_set(InputIt first, InputIt last, size_type bucket_count,
+            const Allocator& alloc)
+      : robin_set(first, last, bucket_count, Hash(), KeyEqual(), alloc) {}
+
+  template <class InputIt>
+  robin_set(InputIt first, InputIt last, size_type bucket_count,
+            const Hash& hash, const Allocator& alloc)
+      : robin_set(first, last, bucket_count, hash, KeyEqual(), alloc) {}
+
+  robin_set(std::initializer_list<value_type> init,
+            size_type bucket_count = ht::DEFAULT_INIT_BUCKETS_SIZE,
+            const Hash& hash = Hash(), const KeyEqual& equal = KeyEqual(),
+            const Allocator& alloc = Allocator())
+      : robin_set(init.begin(), init.end(), bucket_count, hash, equal, alloc) {}
+
+  robin_set(std::initializer_list<value_type> init, size_type bucket_count,
+            const Allocator& alloc)
+      : robin_set(init.begin(), init.end(), bucket_count, Hash(), KeyEqual(),
+                  alloc) {}
+
+  robin_set(std::initializer_list<value_type> init, size_type bucket_count,
+            const Hash& hash, const Allocator& alloc)
+      : robin_set(init.begin(), init.end(), bucket_count, hash, KeyEqual(),
+                  alloc) {}
+
+  robin_set& operator=(std::initializer_list<value_type> ilist) {
+    m_ht.clear();
+
+    m_ht.reserve(ilist.size());
+    m_ht.insert(ilist.begin(), ilist.end());
+
+    return *this;
+  }
+
+  allocator_type get_allocator() const { return m_ht.get_allocator(); }
+
+  /*
+   * Iterators
+   */
+  iterator begin() noexcept { return m_ht.begin(); }
+  const_iterator begin() const noexcept { return m_ht.begin(); }
+  const_iterator cbegin() const noexcept { return m_ht.cbegin(); }
+
+  iterator end() noexcept { return m_ht.end(); }
+  const_iterator end() const noexcept { return m_ht.end(); }
+  const_iterator cend() const noexcept { return m_ht.cend(); }
+
+  /*
+   * Capacity
+   */
+  bool empty() const noexcept { return m_ht.empty(); }
+  size_type size() const noexcept { return m_ht.size(); }
+  size_type max_size() const noexcept { return m_ht.max_size(); }
+
+  /*
+   * Modifiers
+   */
+  void clear() noexcept { m_ht.clear(); }
+
+  std::pair<iterator, bool> insert(const value_type& value) {
+    return m_ht.insert(value);
+  }
+
+  std::pair<iterator, bool> insert(value_type&& value) {
+    return m_ht.insert(std::move(value));
+  }
+
+  iterator insert(const_iterator hint, const value_type& value) {
+    return m_ht.insert_hint(hint, value);
+  }
+
+  iterator insert(const_iterator hint, value_type&& value) {
+    return m_ht.insert_hint(hint, std::move(value));
+  }
+
+  template <class InputIt>
+  void insert(InputIt first, InputIt last) {
+    m_ht.insert(first, last);
+  }
+
+  void insert(std::initializer_list<value_type> ilist) {
+    m_ht.insert(ilist.begin(), ilist.end());
+  }
+
+  /**
+   * Due to the way elements are stored, emplace will need to move or copy the
+   * key-value once. The method is equivalent to
+   * insert(value_type(std::forward<Args>(args)...));
+   *
+   * Mainly here for compatibility with the std::unordered_map interface.
+   */
+  template <class... Args>
+  std::pair<iterator, bool> emplace(Args&&... args) {
+    return m_ht.emplace(std::forward<Args>(args)...);
+  }
+
+  /**
+   * Due to the way elements are stored, emplace_hint will need to move or copy
+   * the key-value once. The method is equivalent to insert(hint,
+   * value_type(std::forward<Args>(args)...));
+   *
+   * Mainly here for compatibility with the std::unordered_map interface.
+   */
+  template <class... Args>
+  iterator emplace_hint(const_iterator hint, Args&&... args) {
+    return m_ht.emplace_hint(hint, std::forward<Args>(args)...);
+  }
+
+  iterator erase(iterator pos) { return m_ht.erase(pos); }
+  iterator erase(const_iterator pos) { return m_ht.erase(pos); }
+  iterator erase(const_iterator first, const_iterator last) {
+    return m_ht.erase(first, last);
+  }
+  size_type erase(const key_type& key) { return m_ht.erase(key); }
+
+  /**
+   * Erase the element at position 'pos'. In contrast to the regular erase()
+   * function, erase_fast() does not return an iterator. This allows it to be
+   * faster especially in hash sets with a low load factor, where finding the
+   * next nonempty bucket would be costly.
+   */
+  void erase_fast(iterator pos) { return m_ht.erase_fast(pos); }
+
+  /**
+   * Use the hash value 'precalculated_hash' instead of hashing the key. The
+   * hash value should be the same as hash_function()(key). Useful to speed-up
+   * the lookup to the value if you already have the hash.
+   */
+  size_type erase(const key_type& key, std::size_t precalculated_hash) {
+    return m_ht.erase(key, precalculated_hash);
+  }
+
+  /**
+   * This overload only participates in the overload resolution if the typedef
+   * KeyEqual::is_transparent exists. If so, K must be hashable and comparable
+   * to Key.
+   */
+  template <
+      class K, class KE = KeyEqual,
+      typename std::enable_if<has_is_transparent<KE>::value>::type* = nullptr>
+  size_type erase(const K& key) {
+    return m_ht.erase(key);
+  }
+
+  /**
+   * @copydoc erase(const K& key)
+   *
+   * Use the hash value 'precalculated_hash' instead of hashing the key. The
+   * hash value should be the same as hash_function()(key). Useful to speed-up
+   * the lookup to the value if you already have the hash.
+   */
+  template <
+      class K, class KE = KeyEqual,
+      typename std::enable_if<has_is_transparent<KE>::value>::type* = nullptr>
+  size_type erase(const K& key, std::size_t precalculated_hash) {
+    return m_ht.erase(key, precalculated_hash);
+  }
+
+  void swap(robin_set& other) { other.m_ht.swap(m_ht); }
+
+  /*
+   * Lookup
+   */
+  size_type count(const Key& key) const { return m_ht.count(key); }
+
+  /**
+   * Use the hash value 'precalculated_hash' instead of hashing the key. The
+   * hash value should be the same as hash_function()(key). Useful to speed-up
+   * the lookup if you already have the hash.
+   */
+  size_type count(const Key& key, std::size_t precalculated_hash) const {
+    return m_ht.count(key, precalculated_hash);
+  }
+
+  /**
+   * This overload only participates in the overload resolution if the typedef
+   * KeyEqual::is_transparent exists. If so, K must be hashable and comparable
+   * to Key.
+   */
+  template <
+      class K, class KE = KeyEqual,
+      typename std::enable_if<has_is_transparent<KE>::value>::type* = nullptr>
+  size_type count(const K& key) const {
+    return m_ht.count(key);
+  }
+
+  /**
+   * @copydoc count(const K& key) const
+   *
+   * Use the hash value 'precalculated_hash' instead of hashing the key. The
+   * hash value should be the same as hash_function()(key). Useful to speed-up
+   * the lookup if you already have the hash.
+   */
+  template <
+      class K, class KE = KeyEqual,
+      typename std::enable_if<has_is_transparent<KE>::value>::type* = nullptr>
+  size_type count(const K& key, std::size_t precalculated_hash) const {
+    return m_ht.count(key, precalculated_hash);
+  }
+
+  iterator find(const Key& key) { return m_ht.find(key); }
+
+  /**
+   * Use the hash value 'precalculated_hash' instead of hashing the key. The
+   * hash value should be the same as hash_function()(key). Useful to speed-up
+   * the lookup if you already have the hash.
+   */
+  iterator find(const Key& key, std::size_t precalculated_hash) {
+    return m_ht.find(key, precalculated_hash);
+  }
+
+  const_iterator find(const Key& key) const { return m_ht.find(key); }
+
+  /**
+   * @copydoc find(const Key& key, std::size_t precalculated_hash)
+   */
+  const_iterator find(const Key& key, std::size_t precalculated_hash) const {
+    return m_ht.find(key, precalculated_hash);
+  }
+
+  /**
+   * This overload only participates in the overload resolution if the typedef
+   * KeyEqual::is_transparent exists. If so, K must be hashable and comparable
+   * to Key.
+   */
+  template <
+      class K, class KE = KeyEqual,
+      typename std::enable_if<has_is_transparent<KE>::value>::type* = nullptr>
+  iterator find(const K& key) {
+    return m_ht.find(key);
+  }
+
+  /**
+   * @copydoc find(const K& key)
+   *
+   * Use the hash value 'precalculated_hash' instead of hashing the key. The
+   * hash value should be the same as hash_function()(key). Useful to speed-up
+   * the lookup if you already have the hash.
+   */
+  template <
+      class K, class KE = KeyEqual,
+      typename std::enable_if<has_is_transparent<KE>::value>::type* = nullptr>
+  iterator find(const K& key, std::size_t precalculated_hash) {
+    return m_ht.find(key, precalculated_hash);
+  }
+
+  /**
+   * @copydoc find(const K& key)
+   */
+  template <
+      class K, class KE = KeyEqual,
+      typename std::enable_if<has_is_transparent<KE>::value>::type* = nullptr>
+  const_iterator find(const K& key) const {
+    return m_ht.find(key);
+  }
+
+  /**
+   * @copydoc find(const K& key)
+   *
+   * Use the hash value 'precalculated_hash' instead of hashing the key. The
+   * hash value should be the same as hash_function()(key). Useful to speed-up
+   * the lookup if you already have the hash.
+   */
+  template <
+      class K, class KE = KeyEqual,
+      typename std::enable_if<has_is_transparent<KE>::value>::type* = nullptr>
+  const_iterator find(const K& key, std::size_t precalculated_hash) const {
+    return m_ht.find(key, precalculated_hash);
+  }
+
+  bool contains(const Key& key) const { return m_ht.contains(key); }
+
+  /**
+   * Use the hash value 'precalculated_hash' instead of hashing the key. The
+   * hash value should be the same as hash_function()(key). Useful to speed-up
+   * the lookup if you already have the hash.
+   */
+  bool contains(const Key& key, std::size_t precalculated_hash) const {
+    return m_ht.contains(key, precalculated_hash);
+  }
+
+  /**
+   * This overload only participates in the overload resolution if the typedef
+   * KeyEqual::is_transparent exists. If so, K must be hashable and comparable
+   * to Key.
+   */
+  template <
+      class K, class KE = KeyEqual,
+      typename std::enable_if<has_is_transparent<KE>::value>::type* = nullptr>
+  bool contains(const K& key) const {
+    return m_ht.contains(key);
+  }
+
+  /**
+   * @copydoc contains(const K& key) const
+   *
+   * Use the hash value 'precalculated_hash' instead of hashing the key. The
+   * hash value should be the same as hash_function()(key). Useful to speed-up
+   * the lookup if you already have the hash.
+   */
+  template <
+      class K, class KE = KeyEqual,
+      typename std::enable_if<has_is_transparent<KE>::value>::type* = nullptr>
+  bool contains(const K& key, std::size_t precalculated_hash) const {
+    return m_ht.contains(key, precalculated_hash);
+  }
+
+  std::pair<iterator, iterator> equal_range(const Key& key) {
+    return m_ht.equal_range(key);
+  }
+
+  /**
+   * Use the hash value 'precalculated_hash' instead of hashing the key. The
+   * hash value should be the same as hash_function()(key). Useful to speed-up
+   * the lookup if you already have the hash.
+   */
+  std::pair<iterator, iterator> equal_range(const Key& key,
+                                            std::size_t precalculated_hash) {
+    return m_ht.equal_range(key, precalculated_hash);
+  }
+
+  std::pair<const_iterator, const_iterator> equal_range(const Key& key) const {
+    return m_ht.equal_range(key);
+  }
+
+  /**
+   * @copydoc equal_range(const Key& key, std::size_t precalculated_hash)
+   */
+  std::pair<const_iterator, const_iterator> equal_range(
+      const Key& key, std::size_t precalculated_hash) const {
+    return m_ht.equal_range(key, precalculated_hash);
+  }
+
+  /**
+   * This overload only participates in the overload resolution if the typedef
+   * KeyEqual::is_transparent exists. If so, K must be hashable and comparable
+   * to Key.
+   */
+  template <
+      class K, class KE = KeyEqual,
+      typename std::enable_if<has_is_transparent<KE>::value>::type* = nullptr>
+  std::pair<iterator, iterator> equal_range(const K& key) {
+    return m_ht.equal_range(key);
+  }
+
+  /**
+   * @copydoc equal_range(const K& key)
+   *
+   * Use the hash value 'precalculated_hash' instead of hashing the key. The
+   * hash value should be the same as hash_function()(key). Useful to speed-up
+   * the lookup if you already have the hash.
+   */
+  template <
+      class K, class KE = KeyEqual,
+      typename std::enable_if<has_is_transparent<KE>::value>::type* = nullptr>
+  std::pair<iterator, iterator> equal_range(const K& key,
+                                            std::size_t precalculated_hash) {
+    return m_ht.equal_range(key, precalculated_hash);
+  }
+
+  /**
+   * @copydoc equal_range(const K& key)
+   */
+  template <
+      class K, class KE = KeyEqual,
+      typename std::enable_if<has_is_transparent<KE>::value>::type* = nullptr>
+  std::pair<const_iterator, const_iterator> equal_range(const K& key) const {
+    return m_ht.equal_range(key);
+  }
+
+  /**
+   * @copydoc equal_range(const K& key, std::size_t precalculated_hash)
+   */
+  template <
+      class K, class KE = KeyEqual,
+      typename std::enable_if<has_is_transparent<KE>::value>::type* = nullptr>
+  std::pair<const_iterator, const_iterator> equal_range(
+      const K& key, std::size_t precalculated_hash) const {
+    return m_ht.equal_range(key, precalculated_hash);
+  }
+
+  /*
+   * Bucket interface
+   */
+  size_type bucket_count() const { return m_ht.bucket_count(); }
+  size_type max_bucket_count() const { return m_ht.max_bucket_count(); }
+
+  /*
+   *  Hash policy
+   */
+  float load_factor() const { return m_ht.load_factor(); }
+
+  float min_load_factor() const { return m_ht.min_load_factor(); }
+  float max_load_factor() const { return m_ht.max_load_factor(); }
+
+  /**
+   * Set the `min_load_factor` to `ml`. When the `load_factor` of the set goes
+   * below `min_load_factor` after some erase operations, the set will be
+   * shrunk when an insertion occurs. The erase method itself never shrinks
+   * the set.
+   *
+   * The default value of `min_load_factor` is 0.0f, the set never shrinks by
+   * default.
+   */
+  void min_load_factor(float ml) { m_ht.min_load_factor(ml); }
+  void max_load_factor(float ml) { m_ht.max_load_factor(ml); }
+
+  void rehash(size_type count_) { m_ht.rehash(count_); }
+  void reserve(size_type count_) { m_ht.reserve(count_); }
+
+  /*
+   * Observers
+   */
+  hasher hash_function() const { return m_ht.hash_function(); }
+  key_equal key_eq() const { return m_ht.key_eq(); }
+
+  /*
+   * Other
+   */
+
+  /**
+   * Convert a const_iterator to an iterator.
+   */
+  iterator mutable_iterator(const_iterator pos) {
+    return m_ht.mutable_iterator(pos);
+  }
+
+  friend bool operator==(const robin_set& lhs, const robin_set& rhs) {
+    if (lhs.size() != rhs.size()) {
+      return false;
+    }
+
+    for (const auto& element_lhs : lhs) {
+      const auto it_element_rhs = rhs.find(element_lhs);
+      if (it_element_rhs == rhs.cend()) {
+        return false;
+      }
+    }
+
+    return true;
+  }
+
+  /**
+   * Serialize the set through the `serializer` parameter.
+   *
+   * The `serializer` parameter must be a function object that supports the
+   * following call:
+   *  - `template<typename U> void operator()(const U& value);` where the types
+   * `std::int16_t`, `std::uint32_t`, `std::uint64_t`, `float` and `Key` must be
+   * supported for U.
+   *
+   * The implementation leaves binary compatibility (endianness, IEEE 754 for
+   * floats, ...) of the types it serializes in the hands of the `Serializer`
+   * function object if compatibility is required.
+   */
+  template <class Serializer>
+  void serialize(Serializer& serializer) const {
+    m_ht.serialize(serializer);
+  }
+
+  /**
+   * Deserialize a previously serialized set through the `deserializer`
+   * parameter.
+   *
+   * The `deserializer` parameter must be a function object that supports the
+   * following call:
+   *  - `template<typename U> U operator()();` where the types `std::int16_t`,
+   * `std::uint32_t`, `std::uint64_t`, `float` and `Key` must be supported for
+   * U.
+   *
+   * If the deserialized hash set type is hash compatible with the serialized
+   * set, the deserialization process can be sped up by setting
+   * `hash_compatible` to true. To be hash compatible, the Hash, KeyEqual and
+   * GrowthPolicy must behave the same way than the ones used on the serialized
+   * set and the StoreHash must have the same value. The `std::size_t` must also
+   * be of the same size as the one on the platform used to serialize the set.
+   * If these criteria are not met, the behaviour is undefined with
+   * `hash_compatible` sets to true.
+   *
+   * The behaviour is undefined if the type `Key` of the `robin_set` is not the
+   * same as the type used during serialization.
+   *
+   * The implementation leaves binary compatibility (endianness, IEEE 754 for
+   * floats, size of int, ...) of the types it deserializes in the hands of the
+   * `Deserializer` function object if compatibility is required.
+   */
+  template <class Deserializer>
+  static robin_set deserialize(Deserializer& deserializer,
+                               bool hash_compatible = false) {
+    robin_set set(0);
+    set.m_ht.deserialize(deserializer, hash_compatible);
+
+    return set;
+  }
+
+  friend bool operator!=(const robin_set& lhs, const robin_set& rhs) {
+    return !operator==(lhs, rhs);
+  }
+
+  friend void swap(robin_set& lhs, robin_set& rhs) { lhs.swap(rhs); }
+
+ private:
+  ht m_ht;
+};
+
+/**
+ * Same as `tsl::robin_set<Key, Hash, KeyEqual, Allocator, StoreHash,
+ * tsl::rh::prime_growth_policy>`.
+ */
+template <class Key, class Hash = std::hash<Key>,
+          class KeyEqual = std::equal_to<Key>,
+          class Allocator = std::allocator<Key>, bool StoreHash = false>
+using robin_pg_set = robin_set<Key, Hash, KeyEqual, Allocator, StoreHash,
+                               tsl::rh::prime_growth_policy>;
+
+}  // end namespace tsl
+
+#endif
diff --git a/ext/spdlog/async.h b/ext/spdlog/async.h
new file mode 100644
index 0000000..e96abd1
--- /dev/null
+++ b/ext/spdlog/async.h
@@ -0,0 +1,100 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+//
+// Async logging using global thread pool
+// All loggers created here share same global thread pool.
+// Each log message is pushed to a queue along with a shared pointer to the
+// logger.
+// If a logger deleted while having pending messages in the queue, it's actual
+// destruction will defer
+// until all its messages are processed by the thread pool.
+// This is because each message in the queue holds a shared_ptr to the
+// originating logger.
+
+#include <spdlog/async_logger.h>
+#include <spdlog/details/registry.h>
+#include <spdlog/details/thread_pool.h>
+
+#include <functional>
+#include <memory>
+#include <mutex>
+
+namespace spdlog {
+
+namespace details {
+static const size_t default_async_q_size = 8192;
+}
+
+// async logger factory - creates async loggers backed with thread pool.
+// if a global thread pool doesn't already exist, create it with default queue
+// size of 8192 items and single thread.
+template <async_overflow_policy OverflowPolicy = async_overflow_policy::block>
+struct async_factory_impl {
+    template <typename Sink, typename... SinkArgs>
+    static std::shared_ptr<async_logger> create(std::string logger_name, SinkArgs &&...args) {
+        auto &registry_inst = details::registry::instance();
+
+        // create global thread pool if not already exists..
+
+        auto &mutex = registry_inst.tp_mutex();
+        std::lock_guard<std::recursive_mutex> tp_lock(mutex);
+        auto tp = registry_inst.get_tp();
+        if (tp == nullptr) {
+            tp = std::make_shared<details::thread_pool>(details::default_async_q_size, 1U);
+            registry_inst.set_tp(tp);
+        }
+
+        auto sink = std::make_shared<Sink>(std::forward<SinkArgs>(args)...);
+        auto new_logger = std::make_shared<async_logger>(std::move(logger_name), std::move(sink),
+                                                         std::move(tp), OverflowPolicy);
+        registry_inst.initialize_logger(new_logger);
+        return new_logger;
+    }
+};
+
+using async_factory = async_factory_impl<async_overflow_policy::block>;
+using async_factory_nonblock = async_factory_impl<async_overflow_policy::overrun_oldest>;
+
+template <typename Sink, typename... SinkArgs>
+inline std::shared_ptr<spdlog::logger> create_async(std::string logger_name,
+                                                    SinkArgs &&...sink_args) {
+    return async_factory::create<Sink>(std::move(logger_name),
+                                       std::forward<SinkArgs>(sink_args)...);
+}
+
+template <typename Sink, typename... SinkArgs>
+inline std::shared_ptr<spdlog::logger> create_async_nb(std::string logger_name,
+                                                       SinkArgs &&...sink_args) {
+    return async_factory_nonblock::create<Sink>(std::move(logger_name),
+                                                std::forward<SinkArgs>(sink_args)...);
+}
+
+// set global thread pool.
+inline void init_thread_pool(size_t q_size,
+                             size_t thread_count,
+                             std::function<void()> on_thread_start,
+                             std::function<void()> on_thread_stop) {
+    auto tp = std::make_shared<details::thread_pool>(q_size, thread_count, on_thread_start,
+                                                     on_thread_stop);
+    details::registry::instance().set_tp(std::move(tp));
+}
+
+inline void init_thread_pool(size_t q_size,
+                             size_t thread_count,
+                             std::function<void()> on_thread_start) {
+    init_thread_pool(q_size, thread_count, on_thread_start, [] {});
+}
+
+inline void init_thread_pool(size_t q_size, size_t thread_count) {
+    init_thread_pool(
+        q_size, thread_count, [] {}, [] {});
+}
+
+// get the global thread pool.
+inline std::shared_ptr<spdlog::details::thread_pool> thread_pool() {
+    return details::registry::instance().get_tp();
+}
+}  // namespace spdlog
diff --git a/ext/spdlog/async_logger-inl.h b/ext/spdlog/async_logger-inl.h
new file mode 100644
index 0000000..1e79479
--- /dev/null
+++ b/ext/spdlog/async_logger-inl.h
@@ -0,0 +1,84 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#ifndef SPDLOG_HEADER_ONLY
+    #include <spdlog/async_logger.h>
+#endif
+
+#include <spdlog/details/thread_pool.h>
+#include <spdlog/sinks/sink.h>
+
+#include <memory>
+#include <string>
+
+SPDLOG_INLINE spdlog::async_logger::async_logger(std::string logger_name,
+                                                 sinks_init_list sinks_list,
+                                                 std::weak_ptr<details::thread_pool> tp,
+                                                 async_overflow_policy overflow_policy)
+    : async_logger(std::move(logger_name),
+                   sinks_list.begin(),
+                   sinks_list.end(),
+                   std::move(tp),
+                   overflow_policy) {}
+
+SPDLOG_INLINE spdlog::async_logger::async_logger(std::string logger_name,
+                                                 sink_ptr single_sink,
+                                                 std::weak_ptr<details::thread_pool> tp,
+                                                 async_overflow_policy overflow_policy)
+    : async_logger(
+          std::move(logger_name), {std::move(single_sink)}, std::move(tp), overflow_policy) {}
+
+// send the log message to the thread pool
+SPDLOG_INLINE void spdlog::async_logger::sink_it_(const details::log_msg &msg){
+    SPDLOG_TRY{if (auto pool_ptr = thread_pool_.lock()){
+        pool_ptr->post_log(shared_from_this(), msg, overflow_policy_);
+}
+else {
+    throw_spdlog_ex("async log: thread pool doesn't exist anymore");
+}
+}
+SPDLOG_LOGGER_CATCH(msg.source)
+}
+
+// send flush request to the thread pool
+SPDLOG_INLINE void spdlog::async_logger::flush_(){
+    SPDLOG_TRY{if (auto pool_ptr = thread_pool_.lock()){
+        pool_ptr->post_flush(shared_from_this(), overflow_policy_);
+}
+else {
+    throw_spdlog_ex("async flush: thread pool doesn't exist anymore");
+}
+}
+SPDLOG_LOGGER_CATCH(source_loc())
+}
+
+//
+// backend functions - called from the thread pool to do the actual job
+//
+SPDLOG_INLINE void spdlog::async_logger::backend_sink_it_(const details::log_msg &msg) {
+    for (auto &sink : sinks_) {
+        if (sink->should_log(msg.level)) {
+            SPDLOG_TRY { sink->log(msg); }
+            SPDLOG_LOGGER_CATCH(msg.source)
+        }
+    }
+
+    if (should_flush_(msg)) {
+        backend_flush_();
+    }
+}
+
+SPDLOG_INLINE void spdlog::async_logger::backend_flush_() {
+    for (auto &sink : sinks_) {
+        SPDLOG_TRY { sink->flush(); }
+        SPDLOG_LOGGER_CATCH(source_loc())
+    }
+}
+
+SPDLOG_INLINE std::shared_ptr<spdlog::logger> spdlog::async_logger::clone(std::string new_name) {
+    auto cloned = std::make_shared<spdlog::async_logger>(*this);
+    cloned->name_ = std::move(new_name);
+    return cloned;
+}
diff --git a/ext/spdlog/async_logger.h b/ext/spdlog/async_logger.h
new file mode 100644
index 0000000..846c4c6
--- /dev/null
+++ b/ext/spdlog/async_logger.h
@@ -0,0 +1,74 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+// Fast asynchronous logger.
+// Uses pre allocated queue.
+// Creates a single back thread to pop messages from the queue and log them.
+//
+// Upon each log write the logger:
+//    1. Checks if its log level is enough to log the message
+//    2. Push a new copy of the message to a queue (or block the caller until
+//    space is available in the queue)
+// Upon destruction, logs all remaining messages in the queue before
+// destructing..
+
+#include <spdlog/logger.h>
+
+namespace spdlog {
+
+// Async overflow policy - block by default.
+enum class async_overflow_policy {
+    block,           // Block until message can be enqueued
+    overrun_oldest,  // Discard oldest message in the queue if full when trying to
+                     // add new item.
+    discard_new      // Discard new message if the queue is full when trying to add new item.
+};
+
+namespace details {
+class thread_pool;
+}
+
+class SPDLOG_API async_logger final : public std::enable_shared_from_this<async_logger>,
+                                      public logger {
+    friend class details::thread_pool;
+
+public:
+    template <typename It>
+    async_logger(std::string logger_name,
+                 It begin,
+                 It end,
+                 std::weak_ptr<details::thread_pool> tp,
+                 async_overflow_policy overflow_policy = async_overflow_policy::block)
+        : logger(std::move(logger_name), begin, end),
+          thread_pool_(std::move(tp)),
+          overflow_policy_(overflow_policy) {}
+
+    async_logger(std::string logger_name,
+                 sinks_init_list sinks_list,
+                 std::weak_ptr<details::thread_pool> tp,
+                 async_overflow_policy overflow_policy = async_overflow_policy::block);
+
+    async_logger(std::string logger_name,
+                 sink_ptr single_sink,
+                 std::weak_ptr<details::thread_pool> tp,
+                 async_overflow_policy overflow_policy = async_overflow_policy::block);
+
+    std::shared_ptr<logger> clone(std::string new_name) override;
+
+protected:
+    void sink_it_(const details::log_msg &msg) override;
+    void flush_() override;
+    void backend_sink_it_(const details::log_msg &incoming_log_msg);
+    void backend_flush_();
+
+private:
+    std::weak_ptr<details::thread_pool> thread_pool_;
+    async_overflow_policy overflow_policy_;
+};
+}  // namespace spdlog
+
+#ifdef SPDLOG_HEADER_ONLY
+    #include "async_logger-inl.h"
+#endif
diff --git a/ext/spdlog/cfg/argv.h b/ext/spdlog/cfg/argv.h
new file mode 100644
index 0000000..7de2f83
--- /dev/null
+++ b/ext/spdlog/cfg/argv.h
@@ -0,0 +1,40 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+#include <spdlog/cfg/helpers.h>
+#include <spdlog/details/registry.h>
+
+//
+// Init log levels using each argv entry that starts with "SPDLOG_LEVEL="
+//
+// set all loggers to debug level:
+// example.exe "SPDLOG_LEVEL=debug"
+
+// set logger1 to trace level
+// example.exe "SPDLOG_LEVEL=logger1=trace"
+
+// turn off all logging except for logger1 and logger2:
+// example.exe "SPDLOG_LEVEL=off,logger1=debug,logger2=info"
+
+namespace spdlog {
+namespace cfg {
+
+// search for SPDLOG_LEVEL= in the args and use it to init the levels
+inline void load_argv_levels(int argc, const char **argv) {
+    const std::string spdlog_level_prefix = "SPDLOG_LEVEL=";
+    for (int i = 1; i < argc; i++) {
+        std::string arg = argv[i];
+        if (arg.find(spdlog_level_prefix) == 0) {
+            auto levels_string = arg.substr(spdlog_level_prefix.size());
+            helpers::load_levels(levels_string);
+        }
+    }
+}
+
+inline void load_argv_levels(int argc, char **argv) {
+    load_argv_levels(argc, const_cast<const char **>(argv));
+}
+
+}  // namespace cfg
+}  // namespace spdlog
diff --git a/ext/spdlog/cfg/env.h b/ext/spdlog/cfg/env.h
new file mode 100644
index 0000000..6e55414
--- /dev/null
+++ b/ext/spdlog/cfg/env.h
@@ -0,0 +1,36 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+#include <spdlog/cfg/helpers.h>
+#include <spdlog/details/os.h>
+#include <spdlog/details/registry.h>
+
+//
+// Init levels and patterns from env variables SPDLOG_LEVEL
+// Inspired from Rust's "env_logger" crate (https://crates.io/crates/env_logger).
+// Note - fallback to "info" level on unrecognized levels
+//
+// Examples:
+//
+// set global level to debug:
+// export SPDLOG_LEVEL=debug
+//
+// turn off all logging except for logger1:
+// export SPDLOG_LEVEL="*=off,logger1=debug"
+//
+
+// turn off all logging except for logger1 and logger2:
+// export SPDLOG_LEVEL="off,logger1=debug,logger2=info"
+
+namespace spdlog {
+namespace cfg {
+inline void load_env_levels() {
+    auto env_val = details::os::getenv("SPDLOG_LEVEL");
+    if (!env_val.empty()) {
+        helpers::load_levels(env_val);
+    }
+}
+
+}  // namespace cfg
+}  // namespace spdlog
diff --git a/ext/spdlog/cfg/helpers-inl.h b/ext/spdlog/cfg/helpers-inl.h
new file mode 100644
index 0000000..93650a2
--- /dev/null
+++ b/ext/spdlog/cfg/helpers-inl.h
@@ -0,0 +1,107 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#ifndef SPDLOG_HEADER_ONLY
+    #include <spdlog/cfg/helpers.h>
+#endif
+
+#include <spdlog/details/os.h>
+#include <spdlog/details/registry.h>
+#include <spdlog/spdlog.h>
+
+#include <algorithm>
+#include <sstream>
+#include <string>
+#include <utility>
+
+namespace spdlog {
+namespace cfg {
+namespace helpers {
+
+// inplace convert to lowercase
+inline std::string &to_lower_(std::string &str) {
+    std::transform(str.begin(), str.end(), str.begin(), [](char ch) {
+        return static_cast<char>((ch >= 'A' && ch <= 'Z') ? ch + ('a' - 'A') : ch);
+    });
+    return str;
+}
+
+// inplace trim spaces
+inline std::string &trim_(std::string &str) {
+    const char *spaces = " \n\r\t";
+    str.erase(str.find_last_not_of(spaces) + 1);
+    str.erase(0, str.find_first_not_of(spaces));
+    return str;
+}
+
+// return (name,value) trimmed pair from given "name=value" string.
+// return empty string on missing parts
+// "key=val" => ("key", "val")
+// " key  =  val " => ("key", "val")
+// "key=" => ("key", "")
+// "val" => ("", "val")
+
+inline std::pair<std::string, std::string> extract_kv_(char sep, const std::string &str) {
+    auto n = str.find(sep);
+    std::string k, v;
+    if (n == std::string::npos) {
+        v = str;
+    } else {
+        k = str.substr(0, n);
+        v = str.substr(n + 1);
+    }
+    return std::make_pair(trim_(k), trim_(v));
+}
+
+// return vector of key/value pairs from sequence of "K1=V1,K2=V2,.."
+// "a=AAA,b=BBB,c=CCC,.." => {("a","AAA"),("b","BBB"),("c", "CCC"),...}
+inline std::unordered_map<std::string, std::string> extract_key_vals_(const std::string &str) {
+    std::string token;
+    std::istringstream token_stream(str);
+    std::unordered_map<std::string, std::string> rv{};
+    while (std::getline(token_stream, token, ',')) {
+        if (token.empty()) {
+            continue;
+        }
+        auto kv = extract_kv_('=', token);
+        rv[kv.first] = kv.second;
+    }
+    return rv;
+}
+
+SPDLOG_INLINE void load_levels(const std::string &input) {
+    if (input.empty() || input.size() > 512) {
+        return;
+    }
+
+    auto key_vals = extract_key_vals_(input);
+    std::unordered_map<std::string, level::level_enum> levels;
+    level::level_enum global_level = level::info;
+    bool global_level_found = false;
+
+    for (auto &name_level : key_vals) {
+        auto &logger_name = name_level.first;
+        auto level_name = to_lower_(name_level.second);
+        auto level = level::from_str(level_name);
+        // ignore unrecognized level names
+        if (level == level::off && level_name != "off") {
+            continue;
+        }
+        if (logger_name.empty())  // no logger name indicate global level
+        {
+            global_level_found = true;
+            global_level = level;
+        } else {
+            levels[logger_name] = level;
+        }
+    }
+
+    details::registry::instance().set_levels(std::move(levels),
+                                             global_level_found ? &global_level : nullptr);
+}
+
+}  // namespace helpers
+}  // namespace cfg
+}  // namespace spdlog
diff --git a/ext/spdlog/cfg/helpers.h b/ext/spdlog/cfg/helpers.h
new file mode 100644
index 0000000..c023818
--- /dev/null
+++ b/ext/spdlog/cfg/helpers.h
@@ -0,0 +1,29 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#include <spdlog/common.h>
+#include <unordered_map>
+
+namespace spdlog {
+namespace cfg {
+namespace helpers {
+//
+// Init levels from given string
+//
+// Examples:
+//
+// set global level to debug: "debug"
+// turn off all logging except for logger1: "off,logger1=debug"
+// turn off all logging except for logger1 and logger2: "off,logger1=debug,logger2=info"
+//
+SPDLOG_API void load_levels(const std::string &txt);
+}  // namespace helpers
+
+}  // namespace cfg
+}  // namespace spdlog
+
+#ifdef SPDLOG_HEADER_ONLY
+    #include "helpers-inl.h"
+#endif  // SPDLOG_HEADER_ONLY
diff --git a/ext/spdlog/common-inl.h b/ext/spdlog/common-inl.h
new file mode 100644
index 0000000..a8a0453
--- /dev/null
+++ b/ext/spdlog/common-inl.h
@@ -0,0 +1,68 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#ifndef SPDLOG_HEADER_ONLY
+    #include <spdlog/common.h>
+#endif
+
+#include <algorithm>
+#include <iterator>
+
+namespace spdlog {
+namespace level {
+
+#if __cplusplus >= 201703L
+constexpr
+#endif
+    static string_view_t level_string_views[] SPDLOG_LEVEL_NAMES;
+
+static const char *short_level_names[] SPDLOG_SHORT_LEVEL_NAMES;
+
+SPDLOG_INLINE const string_view_t &to_string_view(spdlog::level::level_enum l) SPDLOG_NOEXCEPT {
+    return level_string_views[l];
+}
+
+SPDLOG_INLINE const char *to_short_c_str(spdlog::level::level_enum l) SPDLOG_NOEXCEPT {
+    return short_level_names[l];
+}
+
+SPDLOG_INLINE spdlog::level::level_enum from_str(const std::string &name) SPDLOG_NOEXCEPT {
+    auto it = std::find(std::begin(level_string_views), std::end(level_string_views), name);
+    if (it != std::end(level_string_views))
+        return static_cast<level::level_enum>(std::distance(std::begin(level_string_views), it));
+
+    // check also for "warn" and "err" before giving up..
+    if (name == "warn") {
+        return level::warn;
+    }
+    if (name == "err") {
+        return level::err;
+    }
+    return level::off;
+}
+}  // namespace level
+
+SPDLOG_INLINE spdlog_ex::spdlog_ex(std::string msg)
+    : msg_(std::move(msg)) {}
+
+SPDLOG_INLINE spdlog_ex::spdlog_ex(const std::string &msg, int last_errno) {
+#ifdef SPDLOG_USE_STD_FORMAT
+    msg_ = std::system_error(std::error_code(last_errno, std::generic_category()), msg).what();
+#else
+    memory_buf_t outbuf;
+    fmt::format_system_error(outbuf, last_errno, msg.c_str());
+    msg_ = fmt::to_string(outbuf);
+#endif
+}
+
+SPDLOG_INLINE const char *spdlog_ex::what() const SPDLOG_NOEXCEPT { return msg_.c_str(); }
+
+SPDLOG_INLINE void throw_spdlog_ex(const std::string &msg, int last_errno) {
+    SPDLOG_THROW(spdlog_ex(msg, last_errno));
+}
+
+SPDLOG_INLINE void throw_spdlog_ex(std::string msg) { SPDLOG_THROW(spdlog_ex(std::move(msg))); }
+
+}  // namespace spdlog
diff --git a/ext/spdlog/common.h b/ext/spdlog/common.h
new file mode 100644
index 0000000..aca483c
--- /dev/null
+++ b/ext/spdlog/common.h
@@ -0,0 +1,411 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#include <spdlog/details/null_mutex.h>
+#include <spdlog/tweakme.h>
+
+#include <atomic>
+#include <chrono>
+#include <cstdio>
+#include <exception>
+#include <functional>
+#include <initializer_list>
+#include <memory>
+#include <string>
+#include <type_traits>
+
+#ifdef SPDLOG_USE_STD_FORMAT
+    #include <version>
+    #if __cpp_lib_format >= 202207L
+        #include <format>
+    #else
+        #include <string_view>
+    #endif
+#endif
+
+#ifdef SPDLOG_COMPILED_LIB
+    #undef SPDLOG_HEADER_ONLY
+    #if defined(SPDLOG_SHARED_LIB)
+        #if defined(_WIN32)
+            #ifdef spdlog_EXPORTS
+                #define SPDLOG_API __declspec(dllexport)
+            #else  // !spdlog_EXPORTS
+                #define SPDLOG_API __declspec(dllimport)
+            #endif
+        #else  // !defined(_WIN32)
+            #define SPDLOG_API __attribute__((visibility("default")))
+        #endif
+    #else  // !defined(SPDLOG_SHARED_LIB)
+        #define SPDLOG_API
+    #endif
+    #define SPDLOG_INLINE
+#else  // !defined(SPDLOG_COMPILED_LIB)
+    #define SPDLOG_API
+    #define SPDLOG_HEADER_ONLY
+    #define SPDLOG_INLINE inline
+#endif  // #ifdef SPDLOG_COMPILED_LIB
+
+#include <spdlog/fmt/fmt.h>
+
+#if !defined(SPDLOG_USE_STD_FORMAT) && \
+    FMT_VERSION >= 80000  // backward compatibility with fmt versions older than 8
+    #define SPDLOG_FMT_RUNTIME(format_string) fmt::runtime(format_string)
+    #define SPDLOG_FMT_STRING(format_string) FMT_STRING(format_string)
+    #if defined(SPDLOG_WCHAR_FILENAMES) || defined(SPDLOG_WCHAR_TO_UTF8_SUPPORT)
+        #include <spdlog/fmt/xchar.h>
+    #endif
+#else
+    #define SPDLOG_FMT_RUNTIME(format_string) format_string
+    #define SPDLOG_FMT_STRING(format_string) format_string
+#endif
+
+// visual studio up to 2013 does not support noexcept nor constexpr
+#if defined(_MSC_VER) && (_MSC_VER < 1900)
+    #define SPDLOG_NOEXCEPT _NOEXCEPT
+    #define SPDLOG_CONSTEXPR
+#else
+    #define SPDLOG_NOEXCEPT noexcept
+    #define SPDLOG_CONSTEXPR constexpr
+#endif
+
+// If building with std::format, can just use constexpr, otherwise if building with fmt
+// SPDLOG_CONSTEXPR_FUNC needs to be set the same as FMT_CONSTEXPR to avoid situations where
+// a constexpr function in spdlog could end up calling a non-constexpr function in fmt
+// depending on the compiler
+// If fmt determines it can't use constexpr, we should inline the function instead
+#ifdef SPDLOG_USE_STD_FORMAT
+    #define SPDLOG_CONSTEXPR_FUNC constexpr
+#else  // Being built with fmt
+    #if FMT_USE_CONSTEXPR
+        #define SPDLOG_CONSTEXPR_FUNC FMT_CONSTEXPR
+    #else
+        #define SPDLOG_CONSTEXPR_FUNC inline
+    #endif
+#endif
+
+#if defined(__GNUC__) || defined(__clang__)
+    #define SPDLOG_DEPRECATED __attribute__((deprecated))
+#elif defined(_MSC_VER)
+    #define SPDLOG_DEPRECATED __declspec(deprecated)
+#else
+    #define SPDLOG_DEPRECATED
+#endif
+
+// disable thread local on msvc 2013
+#ifndef SPDLOG_NO_TLS
+    #if (defined(_MSC_VER) && (_MSC_VER < 1900)) || defined(__cplusplus_winrt)
+        #define SPDLOG_NO_TLS 1
+    #endif
+#endif
+
+#ifndef SPDLOG_FUNCTION
+    #define SPDLOG_FUNCTION static_cast<const char *>(__FUNCTION__)
+#endif
+
+#ifdef SPDLOG_NO_EXCEPTIONS
+    #define SPDLOG_TRY
+    #define SPDLOG_THROW(ex)                               \
+        do {                                               \
+            printf("spdlog fatal error: %s\n", ex.what()); \
+            std::abort();                                  \
+        } while (0)
+    #define SPDLOG_CATCH_STD
+#else
+    #define SPDLOG_TRY try
+    #define SPDLOG_THROW(ex) throw(ex)
+    #define SPDLOG_CATCH_STD             \
+        catch (const std::exception &) { \
+        }
+#endif
+
+namespace spdlog {
+
+class formatter;
+
+namespace sinks {
+class sink;
+}
+
+#if defined(_WIN32) && defined(SPDLOG_WCHAR_FILENAMES)
+using filename_t = std::wstring;
+    // allow macro expansion to occur in SPDLOG_FILENAME_T
+    #define SPDLOG_FILENAME_T_INNER(s) L##s
+    #define SPDLOG_FILENAME_T(s) SPDLOG_FILENAME_T_INNER(s)
+#else
+using filename_t = std::string;
+    #define SPDLOG_FILENAME_T(s) s
+#endif
+
+using log_clock = std::chrono::system_clock;
+using sink_ptr = std::shared_ptr<sinks::sink>;
+using sinks_init_list = std::initializer_list<sink_ptr>;
+using err_handler = std::function<void(const std::string &err_msg)>;
+#ifdef SPDLOG_USE_STD_FORMAT
+namespace fmt_lib = std;
+
+using string_view_t = std::string_view;
+using memory_buf_t = std::string;
+
+template <typename... Args>
+    #if __cpp_lib_format >= 202207L
+using format_string_t = std::format_string<Args...>;
+    #else
+using format_string_t = std::string_view;
+    #endif
+
+template <class T, class Char = char>
+struct is_convertible_to_basic_format_string
+    : std::integral_constant<bool, std::is_convertible<T, std::basic_string_view<Char>>::value> {};
+
+    #if defined(SPDLOG_WCHAR_FILENAMES) || defined(SPDLOG_WCHAR_TO_UTF8_SUPPORT)
+using wstring_view_t = std::wstring_view;
+using wmemory_buf_t = std::wstring;
+
+template <typename... Args>
+        #if __cpp_lib_format >= 202207L
+using wformat_string_t = std::wformat_string<Args...>;
+        #else
+using wformat_string_t = std::wstring_view;
+        #endif
+    #endif
+    #define SPDLOG_BUF_TO_STRING(x) x
+#else  // use fmt lib instead of std::format
+namespace fmt_lib = fmt;
+
+using string_view_t = fmt::basic_string_view<char>;
+using memory_buf_t = fmt::basic_memory_buffer<char, 250>;
+
+template <typename... Args>
+using format_string_t = fmt::format_string<Args...>;
+
+template <class T>
+using remove_cvref_t = typename std::remove_cv<typename std::remove_reference<T>::type>::type;
+
+template <typename Char>
+    #if FMT_VERSION >= 90101
+using fmt_runtime_string = fmt::runtime_format_string<Char>;
+    #else
+using fmt_runtime_string = fmt::basic_runtime<Char>;
+    #endif
+
+// clang doesn't like SFINAE disabled constructor in std::is_convertible<> so have to repeat the
+// condition from basic_format_string here, in addition, fmt::basic_runtime<Char> is only
+// convertible to basic_format_string<Char> but not basic_string_view<Char>
+template <class T, class Char = char>
+struct is_convertible_to_basic_format_string
+    : std::integral_constant<bool,
+                             std::is_convertible<T, fmt::basic_string_view<Char>>::value ||
+                                 std::is_same<remove_cvref_t<T>, fmt_runtime_string<Char>>::value> {
+};
+
+    #if defined(SPDLOG_WCHAR_FILENAMES) || defined(SPDLOG_WCHAR_TO_UTF8_SUPPORT)
+using wstring_view_t = fmt::basic_string_view<wchar_t>;
+using wmemory_buf_t = fmt::basic_memory_buffer<wchar_t, 250>;
+
+template <typename... Args>
+using wformat_string_t = fmt::wformat_string<Args...>;
+    #endif
+    #define SPDLOG_BUF_TO_STRING(x) fmt::to_string(x)
+#endif
+
+#ifdef SPDLOG_WCHAR_TO_UTF8_SUPPORT
+    #ifndef _WIN32
+        #error SPDLOG_WCHAR_TO_UTF8_SUPPORT only supported on windows
+    #endif  // _WIN32
+#endif      // SPDLOG_WCHAR_TO_UTF8_SUPPORT
+
+template <class T>
+struct is_convertible_to_any_format_string
+    : std::integral_constant<bool,
+                             is_convertible_to_basic_format_string<T, char>::value ||
+                                 is_convertible_to_basic_format_string<T, wchar_t>::value> {};
+
+#if defined(SPDLOG_NO_ATOMIC_LEVELS)
+using level_t = details::null_atomic_int;
+#else
+using level_t = std::atomic<int>;
+#endif
+
+#define SPDLOG_LEVEL_TRACE 0
+#define SPDLOG_LEVEL_DEBUG 1
+#define SPDLOG_LEVEL_INFO 2
+#define SPDLOG_LEVEL_WARN 3
+#define SPDLOG_LEVEL_ERROR 4
+#define SPDLOG_LEVEL_CRITICAL 5
+#define SPDLOG_LEVEL_OFF 6
+
+#if !defined(SPDLOG_ACTIVE_LEVEL)
+    #define SPDLOG_ACTIVE_LEVEL SPDLOG_LEVEL_INFO
+#endif
+
+// Log level enum
+namespace level {
+enum level_enum : int {
+    trace = SPDLOG_LEVEL_TRACE,
+    debug = SPDLOG_LEVEL_DEBUG,
+    info = SPDLOG_LEVEL_INFO,
+    warn = SPDLOG_LEVEL_WARN,
+    err = SPDLOG_LEVEL_ERROR,
+    critical = SPDLOG_LEVEL_CRITICAL,
+    off = SPDLOG_LEVEL_OFF,
+    n_levels
+};
+
+#define SPDLOG_LEVEL_NAME_TRACE spdlog::string_view_t("trace", 5)
+#define SPDLOG_LEVEL_NAME_DEBUG spdlog::string_view_t("debug", 5)
+#define SPDLOG_LEVEL_NAME_INFO spdlog::string_view_t("info", 4)
+#define SPDLOG_LEVEL_NAME_WARNING spdlog::string_view_t("warning", 7)
+#define SPDLOG_LEVEL_NAME_ERROR spdlog::string_view_t("error", 5)
+#define SPDLOG_LEVEL_NAME_CRITICAL spdlog::string_view_t("critical", 8)
+#define SPDLOG_LEVEL_NAME_OFF spdlog::string_view_t("off", 3)
+
+#if !defined(SPDLOG_LEVEL_NAMES)
+    #define SPDLOG_LEVEL_NAMES                                                                  \
+        {                                                                                       \
+            SPDLOG_LEVEL_NAME_TRACE, SPDLOG_LEVEL_NAME_DEBUG, SPDLOG_LEVEL_NAME_INFO,           \
+                SPDLOG_LEVEL_NAME_WARNING, SPDLOG_LEVEL_NAME_ERROR, SPDLOG_LEVEL_NAME_CRITICAL, \
+                SPDLOG_LEVEL_NAME_OFF                                                           \
+        }
+#endif
+
+#if !defined(SPDLOG_SHORT_LEVEL_NAMES)
+
+    #define SPDLOG_SHORT_LEVEL_NAMES \
+        { "T", "D", "I", "W", "E", "C", "O" }
+#endif
+
+SPDLOG_API const string_view_t &to_string_view(spdlog::level::level_enum l) SPDLOG_NOEXCEPT;
+SPDLOG_API const char *to_short_c_str(spdlog::level::level_enum l) SPDLOG_NOEXCEPT;
+SPDLOG_API spdlog::level::level_enum from_str(const std::string &name) SPDLOG_NOEXCEPT;
+
+}  // namespace level
+
+//
+// Color mode used by sinks with color support.
+//
+enum class color_mode { always, automatic, never };
+
+//
+// Pattern time - specific time getting to use for pattern_formatter.
+// local time by default
+//
+enum class pattern_time_type {
+    local,  // log localtime
+    utc     // log utc
+};
+
+//
+// Log exception
+//
+class SPDLOG_API spdlog_ex : public std::exception {
+public:
+    explicit spdlog_ex(std::string msg);
+    spdlog_ex(const std::string &msg, int last_errno);
+    const char *what() const SPDLOG_NOEXCEPT override;
+
+private:
+    std::string msg_;
+};
+
+[[noreturn]] SPDLOG_API void throw_spdlog_ex(const std::string &msg, int last_errno);
+[[noreturn]] SPDLOG_API void throw_spdlog_ex(std::string msg);
+
+struct source_loc {
+    SPDLOG_CONSTEXPR source_loc() = default;
+    SPDLOG_CONSTEXPR source_loc(const char *filename_in, int line_in, const char *funcname_in)
+        : filename{filename_in},
+          line{line_in},
+          funcname{funcname_in} {}
+
+    SPDLOG_CONSTEXPR bool empty() const SPDLOG_NOEXCEPT { return line <= 0; }
+    const char *filename{nullptr};
+    int line{0};
+    const char *funcname{nullptr};
+};
+
+struct file_event_handlers {
+    file_event_handlers()
+        : before_open(nullptr),
+          after_open(nullptr),
+          before_close(nullptr),
+          after_close(nullptr) {}
+
+    std::function<void(const filename_t &filename)> before_open;
+    std::function<void(const filename_t &filename, std::FILE *file_stream)> after_open;
+    std::function<void(const filename_t &filename, std::FILE *file_stream)> before_close;
+    std::function<void(const filename_t &filename)> after_close;
+};
+
+namespace details {
+
+// to_string_view
+
+SPDLOG_CONSTEXPR_FUNC spdlog::string_view_t to_string_view(const memory_buf_t &buf)
+    SPDLOG_NOEXCEPT {
+    return spdlog::string_view_t{buf.data(), buf.size()};
+}
+
+SPDLOG_CONSTEXPR_FUNC spdlog::string_view_t to_string_view(spdlog::string_view_t str)
+    SPDLOG_NOEXCEPT {
+    return str;
+}
+
+#if defined(SPDLOG_WCHAR_FILENAMES) || defined(SPDLOG_WCHAR_TO_UTF8_SUPPORT)
+SPDLOG_CONSTEXPR_FUNC spdlog::wstring_view_t to_string_view(const wmemory_buf_t &buf)
+    SPDLOG_NOEXCEPT {
+    return spdlog::wstring_view_t{buf.data(), buf.size()};
+}
+
+SPDLOG_CONSTEXPR_FUNC spdlog::wstring_view_t to_string_view(spdlog::wstring_view_t str)
+    SPDLOG_NOEXCEPT {
+    return str;
+}
+#endif
+
+#ifndef SPDLOG_USE_STD_FORMAT
+template <typename T, typename... Args>
+inline fmt::basic_string_view<T> to_string_view(fmt::basic_format_string<T, Args...> fmt) {
+    return fmt;
+}
+#elif __cpp_lib_format >= 202207L
+template <typename T, typename... Args>
+SPDLOG_CONSTEXPR_FUNC std::basic_string_view<T> to_string_view(
+    std::basic_format_string<T, Args...> fmt) SPDLOG_NOEXCEPT {
+    return fmt.get();
+}
+#endif
+
+// make_unique support for pre c++14
+#if __cplusplus >= 201402L  // C++14 and beyond
+using std::enable_if_t;
+using std::make_unique;
+#else
+template <bool B, class T = void>
+using enable_if_t = typename std::enable_if<B, T>::type;
+
+template <typename T, typename... Args>
+std::unique_ptr<T> make_unique(Args &&...args) {
+    static_assert(!std::is_array<T>::value, "arrays not supported");
+    return std::unique_ptr<T>(new T(std::forward<Args>(args)...));
+}
+#endif
+
+// to avoid useless casts (see https://github.com/nlohmann/json/issues/2893#issuecomment-889152324)
+template <typename T, typename U, enable_if_t<!std::is_same<T, U>::value, int> = 0>
+constexpr T conditional_static_cast(U value) {
+    return static_cast<T>(value);
+}
+
+template <typename T, typename U, enable_if_t<std::is_same<T, U>::value, int> = 0>
+constexpr T conditional_static_cast(U value) {
+    return value;
+}
+
+}  // namespace details
+}  // namespace spdlog
+
+#ifdef SPDLOG_HEADER_ONLY
+    #include "common-inl.h"
+#endif
diff --git a/ext/spdlog/details/backtracer-inl.h b/ext/spdlog/details/backtracer-inl.h
new file mode 100644
index 0000000..43d1002
--- /dev/null
+++ b/ext/spdlog/details/backtracer-inl.h
@@ -0,0 +1,63 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#ifndef SPDLOG_HEADER_ONLY
+    #include <spdlog/details/backtracer.h>
+#endif
+namespace spdlog {
+namespace details {
+SPDLOG_INLINE backtracer::backtracer(const backtracer &other) {
+    std::lock_guard<std::mutex> lock(other.mutex_);
+    enabled_ = other.enabled();
+    messages_ = other.messages_;
+}
+
+SPDLOG_INLINE backtracer::backtracer(backtracer &&other) SPDLOG_NOEXCEPT {
+    std::lock_guard<std::mutex> lock(other.mutex_);
+    enabled_ = other.enabled();
+    messages_ = std::move(other.messages_);
+}
+
+SPDLOG_INLINE backtracer &backtracer::operator=(backtracer other) {
+    std::lock_guard<std::mutex> lock(mutex_);
+    enabled_ = other.enabled();
+    messages_ = std::move(other.messages_);
+    return *this;
+}
+
+SPDLOG_INLINE void backtracer::enable(size_t size) {
+    std::lock_guard<std::mutex> lock{mutex_};
+    enabled_.store(true, std::memory_order_relaxed);
+    messages_ = circular_q<log_msg_buffer>{size};
+}
+
+SPDLOG_INLINE void backtracer::disable() {
+    std::lock_guard<std::mutex> lock{mutex_};
+    enabled_.store(false, std::memory_order_relaxed);
+}
+
+SPDLOG_INLINE bool backtracer::enabled() const { return enabled_.load(std::memory_order_relaxed); }
+
+SPDLOG_INLINE void backtracer::push_back(const log_msg &msg) {
+    std::lock_guard<std::mutex> lock{mutex_};
+    messages_.push_back(log_msg_buffer{msg});
+}
+
+SPDLOG_INLINE bool backtracer::empty() const {
+    std::lock_guard<std::mutex> lock{mutex_};
+    return messages_.empty();
+}
+
+// pop all items in the q and apply the given fun on each of them.
+SPDLOG_INLINE void backtracer::foreach_pop(std::function<void(const details::log_msg &)> fun) {
+    std::lock_guard<std::mutex> lock{mutex_};
+    while (!messages_.empty()) {
+        auto &front_msg = messages_.front();
+        fun(front_msg);
+        messages_.pop_front();
+    }
+}
+}  // namespace details
+}  // namespace spdlog
diff --git a/ext/spdlog/details/backtracer.h b/ext/spdlog/details/backtracer.h
new file mode 100644
index 0000000..541339c
--- /dev/null
+++ b/ext/spdlog/details/backtracer.h
@@ -0,0 +1,45 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#include <spdlog/details/circular_q.h>
+#include <spdlog/details/log_msg_buffer.h>
+
+#include <atomic>
+#include <functional>
+#include <mutex>
+
+// Store log messages in circular buffer.
+// Useful for storing debug data in case of error/warning happens.
+
+namespace spdlog {
+namespace details {
+class SPDLOG_API backtracer {
+    mutable std::mutex mutex_;
+    std::atomic<bool> enabled_{false};
+    circular_q<log_msg_buffer> messages_;
+
+public:
+    backtracer() = default;
+    backtracer(const backtracer &other);
+
+    backtracer(backtracer &&other) SPDLOG_NOEXCEPT;
+    backtracer &operator=(backtracer other);
+
+    void enable(size_t size);
+    void disable();
+    bool enabled() const;
+    void push_back(const log_msg &msg);
+    bool empty() const;
+
+    // pop all items in the q and apply the given fun on each of them.
+    void foreach_pop(std::function<void(const details::log_msg &)> fun);
+};
+
+}  // namespace details
+}  // namespace spdlog
+
+#ifdef SPDLOG_HEADER_ONLY
+    #include "backtracer-inl.h"
+#endif
diff --git a/ext/spdlog/details/circular_q.h b/ext/spdlog/details/circular_q.h
new file mode 100644
index 0000000..29e9d25
--- /dev/null
+++ b/ext/spdlog/details/circular_q.h
@@ -0,0 +1,115 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+// circular q view of std::vector.
+#pragma once
+
+#include <cassert>
+#include <vector>
+
+#include "spdlog/common.h"
+
+namespace spdlog {
+namespace details {
+template <typename T>
+class circular_q {
+    size_t max_items_ = 0;
+    typename std::vector<T>::size_type head_ = 0;
+    typename std::vector<T>::size_type tail_ = 0;
+    size_t overrun_counter_ = 0;
+    std::vector<T> v_;
+
+public:
+    using value_type = T;
+
+    // empty ctor - create a disabled queue with no elements allocated at all
+    circular_q() = default;
+
+    explicit circular_q(size_t max_items)
+        : max_items_(max_items + 1)  // one item is reserved as marker for full q
+          ,
+          v_(max_items_) {}
+
+    circular_q(const circular_q &) = default;
+    circular_q &operator=(const circular_q &) = default;
+
+    // move cannot be default,
+    // since we need to reset head_, tail_, etc to zero in the moved object
+    circular_q(circular_q &&other) SPDLOG_NOEXCEPT { copy_moveable(std::move(other)); }
+
+    circular_q &operator=(circular_q &&other) SPDLOG_NOEXCEPT {
+        copy_moveable(std::move(other));
+        return *this;
+    }
+
+    // push back, overrun (oldest) item if no room left
+    void push_back(T &&item) {
+        if (max_items_ > 0) {
+            v_[tail_] = std::move(item);
+            tail_ = (tail_ + 1) % max_items_;
+
+            if (tail_ == head_)  // overrun last item if full
+            {
+                head_ = (head_ + 1) % max_items_;
+                ++overrun_counter_;
+            }
+        }
+    }
+
+    // Return reference to the front item.
+    // If there are no elements in the container, the behavior is undefined.
+    const T &front() const { return v_[head_]; }
+
+    T &front() { return v_[head_]; }
+
+    // Return number of elements actually stored
+    size_t size() const {
+        if (tail_ >= head_) {
+            return tail_ - head_;
+        } else {
+            return max_items_ - (head_ - tail_);
+        }
+    }
+
+    // Return const reference to item by index.
+    // If index is out of range 0…size()-1, the behavior is undefined.
+    const T &at(size_t i) const {
+        assert(i < size());
+        return v_[(head_ + i) % max_items_];
+    }
+
+    // Pop item from front.
+    // If there are no elements in the container, the behavior is undefined.
+    void pop_front() { head_ = (head_ + 1) % max_items_; }
+
+    bool empty() const { return tail_ == head_; }
+
+    bool full() const {
+        // head is ahead of the tail by 1
+        if (max_items_ > 0) {
+            return ((tail_ + 1) % max_items_) == head_;
+        }
+        return false;
+    }
+
+    size_t overrun_counter() const { return overrun_counter_; }
+
+    void reset_overrun_counter() { overrun_counter_ = 0; }
+
+private:
+    // copy from other&& and reset it to disabled state
+    void copy_moveable(circular_q &&other) SPDLOG_NOEXCEPT {
+        max_items_ = other.max_items_;
+        head_ = other.head_;
+        tail_ = other.tail_;
+        overrun_counter_ = other.overrun_counter_;
+        v_ = std::move(other.v_);
+
+        // put &&other in disabled, but valid state
+        other.max_items_ = 0;
+        other.head_ = other.tail_ = 0;
+        other.overrun_counter_ = 0;
+    }
+};
+}  // namespace details
+}  // namespace spdlog
diff --git a/ext/spdlog/details/console_globals.h b/ext/spdlog/details/console_globals.h
new file mode 100644
index 0000000..9c55210
--- /dev/null
+++ b/ext/spdlog/details/console_globals.h
@@ -0,0 +1,28 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#include <mutex>
+#include <spdlog/details/null_mutex.h>
+
+namespace spdlog {
+namespace details {
+
+struct console_mutex {
+    using mutex_t = std::mutex;
+    static mutex_t &mutex() {
+        static mutex_t s_mutex;
+        return s_mutex;
+    }
+};
+
+struct console_nullmutex {
+    using mutex_t = null_mutex;
+    static mutex_t &mutex() {
+        static mutex_t s_mutex;
+        return s_mutex;
+    }
+};
+}  // namespace details
+}  // namespace spdlog
diff --git a/ext/spdlog/details/file_helper-inl.h b/ext/spdlog/details/file_helper-inl.h
new file mode 100644
index 0000000..8742b96
--- /dev/null
+++ b/ext/spdlog/details/file_helper-inl.h
@@ -0,0 +1,153 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#ifndef SPDLOG_HEADER_ONLY
+    #include <spdlog/details/file_helper.h>
+#endif
+
+#include <spdlog/common.h>
+#include <spdlog/details/os.h>
+
+#include <cerrno>
+#include <chrono>
+#include <cstdio>
+#include <string>
+#include <thread>
+#include <tuple>
+
+namespace spdlog {
+namespace details {
+
+SPDLOG_INLINE file_helper::file_helper(const file_event_handlers &event_handlers)
+    : event_handlers_(event_handlers) {}
+
+SPDLOG_INLINE file_helper::~file_helper() { close(); }
+
+SPDLOG_INLINE void file_helper::open(const filename_t &fname, bool truncate) {
+    close();
+    filename_ = fname;
+
+    auto *mode = SPDLOG_FILENAME_T("ab");
+    auto *trunc_mode = SPDLOG_FILENAME_T("wb");
+
+    if (event_handlers_.before_open) {
+        event_handlers_.before_open(filename_);
+    }
+    for (int tries = 0; tries < open_tries_; ++tries) {
+        // create containing folder if not exists already.
+        os::create_dir(os::dir_name(fname));
+        if (truncate) {
+            // Truncate by opening-and-closing a tmp file in "wb" mode, always
+            // opening the actual log-we-write-to in "ab" mode, since that
+            // interacts more politely with eternal processes that might
+            // rotate/truncate the file underneath us.
+            std::FILE *tmp;
+            if (os::fopen_s(&tmp, fname, trunc_mode)) {
+                continue;
+            }
+            std::fclose(tmp);
+        }
+        if (!os::fopen_s(&fd_, fname, mode)) {
+            if (event_handlers_.after_open) {
+                event_handlers_.after_open(filename_, fd_);
+            }
+            return;
+        }
+
+        details::os::sleep_for_millis(open_interval_);
+    }
+
+    throw_spdlog_ex("Failed opening file " + os::filename_to_str(filename_) + " for writing",
+                    errno);
+}
+
+SPDLOG_INLINE void file_helper::reopen(bool truncate) {
+    if (filename_.empty()) {
+        throw_spdlog_ex("Failed re opening file - was not opened before");
+    }
+    this->open(filename_, truncate);
+}
+
+SPDLOG_INLINE void file_helper::flush() {
+    if (std::fflush(fd_) != 0) {
+        throw_spdlog_ex("Failed flush to file " + os::filename_to_str(filename_), errno);
+    }
+}
+
+SPDLOG_INLINE void file_helper::sync() {
+    if (!os::fsync(fd_)) {
+        throw_spdlog_ex("Failed to fsync file " + os::filename_to_str(filename_), errno);
+    }
+}
+
+SPDLOG_INLINE void file_helper::close() {
+    if (fd_ != nullptr) {
+        if (event_handlers_.before_close) {
+            event_handlers_.before_close(filename_, fd_);
+        }
+
+        std::fclose(fd_);
+        fd_ = nullptr;
+
+        if (event_handlers_.after_close) {
+            event_handlers_.after_close(filename_);
+        }
+    }
+}
+
+SPDLOG_INLINE void file_helper::write(const memory_buf_t &buf) {
+    if (fd_ == nullptr) return;
+    size_t msg_size = buf.size();
+    auto data = buf.data();
+
+    if (!details::os::fwrite_bytes(data, msg_size, fd_)) {
+        throw_spdlog_ex("Failed writing to file " + os::filename_to_str(filename_), errno);
+    }
+}
+
+SPDLOG_INLINE size_t file_helper::size() const {
+    if (fd_ == nullptr) {
+        throw_spdlog_ex("Cannot use size() on closed file " + os::filename_to_str(filename_));
+    }
+    return os::filesize(fd_);
+}
+
+SPDLOG_INLINE const filename_t &file_helper::filename() const { return filename_; }
+
+//
+// return file path and its extension:
+//
+// "mylog.txt" => ("mylog", ".txt")
+// "mylog" => ("mylog", "")
+// "mylog." => ("mylog.", "")
+// "/dir1/dir2/mylog.txt" => ("/dir1/dir2/mylog", ".txt")
+//
+// the starting dot in filenames is ignored (hidden files):
+//
+// ".mylog" => (".mylog". "")
+// "my_folder/.mylog" => ("my_folder/.mylog", "")
+// "my_folder/.mylog.txt" => ("my_folder/.mylog", ".txt")
+SPDLOG_INLINE std::tuple<filename_t, filename_t> file_helper::split_by_extension(
+    const filename_t &fname) {
+    auto ext_index = fname.rfind('.');
+
+    // no valid extension found - return whole path and empty string as
+    // extension
+    if (ext_index == filename_t::npos || ext_index == 0 || ext_index == fname.size() - 1) {
+        return std::make_tuple(fname, filename_t());
+    }
+
+    // treat cases like "/etc/rc.d/somelogfile or "/abc/.hiddenfile"
+    auto folder_index = fname.find_last_of(details::os::folder_seps_filename);
+    if (folder_index != filename_t::npos && folder_index >= ext_index - 1) {
+        return std::make_tuple(fname, filename_t());
+    }
+
+    // finally - return a valid base and extension tuple
+    return std::make_tuple(fname.substr(0, ext_index), fname.substr(ext_index));
+}
+
+}  // namespace details
+}  // namespace spdlog
diff --git a/ext/spdlog/details/file_helper.h b/ext/spdlog/details/file_helper.h
new file mode 100644
index 0000000..f0e5d18
--- /dev/null
+++ b/ext/spdlog/details/file_helper.h
@@ -0,0 +1,61 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#include <spdlog/common.h>
+#include <tuple>
+
+namespace spdlog {
+namespace details {
+
+// Helper class for file sinks.
+// When failing to open a file, retry several times(5) with a delay interval(10 ms).
+// Throw spdlog_ex exception on errors.
+
+class SPDLOG_API file_helper {
+public:
+    file_helper() = default;
+    explicit file_helper(const file_event_handlers &event_handlers);
+
+    file_helper(const file_helper &) = delete;
+    file_helper &operator=(const file_helper &) = delete;
+    ~file_helper();
+
+    void open(const filename_t &fname, bool truncate = false);
+    void reopen(bool truncate);
+    void flush();
+    void sync();
+    void close();
+    void write(const memory_buf_t &buf);
+    size_t size() const;
+    const filename_t &filename() const;
+
+    //
+    // return file path and its extension:
+    //
+    // "mylog.txt" => ("mylog", ".txt")
+    // "mylog" => ("mylog", "")
+    // "mylog." => ("mylog.", "")
+    // "/dir1/dir2/mylog.txt" => ("/dir1/dir2/mylog", ".txt")
+    //
+    // the starting dot in filenames is ignored (hidden files):
+    //
+    // ".mylog" => (".mylog". "")
+    // "my_folder/.mylog" => ("my_folder/.mylog", "")
+    // "my_folder/.mylog.txt" => ("my_folder/.mylog", ".txt")
+    static std::tuple<filename_t, filename_t> split_by_extension(const filename_t &fname);
+
+private:
+    const int open_tries_ = 5;
+    const unsigned int open_interval_ = 10;
+    std::FILE *fd_{nullptr};
+    filename_t filename_;
+    file_event_handlers event_handlers_;
+};
+}  // namespace details
+}  // namespace spdlog
+
+#ifdef SPDLOG_HEADER_ONLY
+    #include "file_helper-inl.h"
+#endif
diff --git a/ext/spdlog/details/fmt_helper.h b/ext/spdlog/details/fmt_helper.h
new file mode 100644
index 0000000..6130600
--- /dev/null
+++ b/ext/spdlog/details/fmt_helper.h
@@ -0,0 +1,141 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+#pragma once
+
+#include <chrono>
+#include <iterator>
+#include <spdlog/common.h>
+#include <spdlog/fmt/fmt.h>
+#include <type_traits>
+
+#ifdef SPDLOG_USE_STD_FORMAT
+    #include <charconv>
+    #include <limits>
+#endif
+
+// Some fmt helpers to efficiently format and pad ints and strings
+namespace spdlog {
+namespace details {
+namespace fmt_helper {
+
+inline void append_string_view(spdlog::string_view_t view, memory_buf_t &dest) {
+    auto *buf_ptr = view.data();
+    dest.append(buf_ptr, buf_ptr + view.size());
+}
+
+#ifdef SPDLOG_USE_STD_FORMAT
+template <typename T>
+inline void append_int(T n, memory_buf_t &dest) {
+    // Buffer should be large enough to hold all digits (digits10 + 1) and a sign
+    SPDLOG_CONSTEXPR const auto BUF_SIZE = std::numeric_limits<T>::digits10 + 2;
+    char buf[BUF_SIZE];
+
+    auto [ptr, ec] = std::to_chars(buf, buf + BUF_SIZE, n, 10);
+    if (ec == std::errc()) {
+        dest.append(buf, ptr);
+    } else {
+        throw_spdlog_ex("Failed to format int", static_cast<int>(ec));
+    }
+}
+#else
+template <typename T>
+inline void append_int(T n, memory_buf_t &dest) {
+    fmt::format_int i(n);
+    dest.append(i.data(), i.data() + i.size());
+}
+#endif
+
+template <typename T>
+SPDLOG_CONSTEXPR_FUNC unsigned int count_digits_fallback(T n) {
+    // taken from fmt: https://github.com/fmtlib/fmt/blob/8.0.1/include/fmt/format.h#L899-L912
+    unsigned int count = 1;
+    for (;;) {
+        // Integer division is slow so do it for a group of four digits instead
+        // of for every digit. The idea comes from the talk by Alexandrescu
+        // "Three Optimization Tips for C++". See speed-test for a comparison.
+        if (n < 10) return count;
+        if (n < 100) return count + 1;
+        if (n < 1000) return count + 2;
+        if (n < 10000) return count + 3;
+        n /= 10000u;
+        count += 4;
+    }
+}
+
+template <typename T>
+inline unsigned int count_digits(T n) {
+    using count_type =
+        typename std::conditional<(sizeof(T) > sizeof(uint32_t)), uint64_t, uint32_t>::type;
+#ifdef SPDLOG_USE_STD_FORMAT
+    return count_digits_fallback(static_cast<count_type>(n));
+#else
+    return static_cast<unsigned int>(fmt::
+    // fmt 7.0.0 renamed the internal namespace to detail.
+    // See: https://github.com/fmtlib/fmt/issues/1538
+    #if FMT_VERSION < 70000
+                                         internal
+    #else
+                                         detail
+    #endif
+                                     ::count_digits(static_cast<count_type>(n)));
+#endif
+}
+
+inline void pad2(int n, memory_buf_t &dest) {
+    if (n >= 0 && n < 100)  // 0-99
+    {
+        dest.push_back(static_cast<char>('0' + n / 10));
+        dest.push_back(static_cast<char>('0' + n % 10));
+    } else  // unlikely, but just in case, let fmt deal with it
+    {
+        fmt_lib::format_to(std::back_inserter(dest), SPDLOG_FMT_STRING("{:02}"), n);
+    }
+}
+
+template <typename T>
+inline void pad_uint(T n, unsigned int width, memory_buf_t &dest) {
+    static_assert(std::is_unsigned<T>::value, "pad_uint must get unsigned T");
+    for (auto digits = count_digits(n); digits < width; digits++) {
+        dest.push_back('0');
+    }
+    append_int(n, dest);
+}
+
+template <typename T>
+inline void pad3(T n, memory_buf_t &dest) {
+    static_assert(std::is_unsigned<T>::value, "pad3 must get unsigned T");
+    if (n < 1000) {
+        dest.push_back(static_cast<char>(n / 100 + '0'));
+        n = n % 100;
+        dest.push_back(static_cast<char>((n / 10) + '0'));
+        dest.push_back(static_cast<char>((n % 10) + '0'));
+    } else {
+        append_int(n, dest);
+    }
+}
+
+template <typename T>
+inline void pad6(T n, memory_buf_t &dest) {
+    pad_uint(n, 6, dest);
+}
+
+template <typename T>
+inline void pad9(T n, memory_buf_t &dest) {
+    pad_uint(n, 9, dest);
+}
+
+// return fraction of a second of the given time_point.
+// e.g.
+// fraction<std::milliseconds>(tp) -> will return the millis part of the second
+template <typename ToDuration>
+inline ToDuration time_fraction(log_clock::time_point tp) {
+    using std::chrono::duration_cast;
+    using std::chrono::seconds;
+    auto duration = tp.time_since_epoch();
+    auto secs = duration_cast<seconds>(duration);
+    return duration_cast<ToDuration>(duration) - duration_cast<ToDuration>(secs);
+}
+
+}  // namespace fmt_helper
+}  // namespace details
+}  // namespace spdlog
diff --git a/ext/spdlog/details/log_msg-inl.h b/ext/spdlog/details/log_msg-inl.h
new file mode 100644
index 0000000..aa3a957
--- /dev/null
+++ b/ext/spdlog/details/log_msg-inl.h
@@ -0,0 +1,44 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#ifndef SPDLOG_HEADER_ONLY
+    #include <spdlog/details/log_msg.h>
+#endif
+
+#include <spdlog/details/os.h>
+
+namespace spdlog {
+namespace details {
+
+SPDLOG_INLINE log_msg::log_msg(spdlog::log_clock::time_point log_time,
+                               spdlog::source_loc loc,
+                               string_view_t a_logger_name,
+                               spdlog::level::level_enum lvl,
+                               spdlog::string_view_t msg)
+    : logger_name(a_logger_name),
+      level(lvl),
+      time(log_time)
+#ifndef SPDLOG_NO_THREAD_ID
+      ,
+      thread_id(os::thread_id())
+#endif
+      ,
+      source(loc),
+      payload(msg) {
+}
+
+SPDLOG_INLINE log_msg::log_msg(spdlog::source_loc loc,
+                               string_view_t a_logger_name,
+                               spdlog::level::level_enum lvl,
+                               spdlog::string_view_t msg)
+    : log_msg(os::now(), loc, a_logger_name, lvl, msg) {}
+
+SPDLOG_INLINE log_msg::log_msg(string_view_t a_logger_name,
+                               spdlog::level::level_enum lvl,
+                               spdlog::string_view_t msg)
+    : log_msg(os::now(), source_loc{}, a_logger_name, lvl, msg) {}
+
+}  // namespace details
+}  // namespace spdlog
diff --git a/ext/spdlog/details/log_msg.h b/ext/spdlog/details/log_msg.h
new file mode 100644
index 0000000..87df1e8
--- /dev/null
+++ b/ext/spdlog/details/log_msg.h
@@ -0,0 +1,40 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#include <spdlog/common.h>
+#include <string>
+
+namespace spdlog {
+namespace details {
+struct SPDLOG_API log_msg {
+    log_msg() = default;
+    log_msg(log_clock::time_point log_time,
+            source_loc loc,
+            string_view_t logger_name,
+            level::level_enum lvl,
+            string_view_t msg);
+    log_msg(source_loc loc, string_view_t logger_name, level::level_enum lvl, string_view_t msg);
+    log_msg(string_view_t logger_name, level::level_enum lvl, string_view_t msg);
+    log_msg(const log_msg &other) = default;
+    log_msg &operator=(const log_msg &other) = default;
+
+    string_view_t logger_name;
+    level::level_enum level{level::off};
+    log_clock::time_point time;
+    size_t thread_id{0};
+
+    // wrapping the formatted text with color (updated by pattern_formatter).
+    mutable size_t color_range_start{0};
+    mutable size_t color_range_end{0};
+
+    source_loc source;
+    string_view_t payload;
+};
+}  // namespace details
+}  // namespace spdlog
+
+#ifdef SPDLOG_HEADER_ONLY
+    #include "log_msg-inl.h"
+#endif
diff --git a/ext/spdlog/details/log_msg_buffer-inl.h b/ext/spdlog/details/log_msg_buffer-inl.h
new file mode 100644
index 0000000..2eb2428
--- /dev/null
+++ b/ext/spdlog/details/log_msg_buffer-inl.h
@@ -0,0 +1,54 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#ifndef SPDLOG_HEADER_ONLY
+    #include <spdlog/details/log_msg_buffer.h>
+#endif
+
+namespace spdlog {
+namespace details {
+
+SPDLOG_INLINE log_msg_buffer::log_msg_buffer(const log_msg &orig_msg)
+    : log_msg{orig_msg} {
+    buffer.append(logger_name.begin(), logger_name.end());
+    buffer.append(payload.begin(), payload.end());
+    update_string_views();
+}
+
+SPDLOG_INLINE log_msg_buffer::log_msg_buffer(const log_msg_buffer &other)
+    : log_msg{other} {
+    buffer.append(logger_name.begin(), logger_name.end());
+    buffer.append(payload.begin(), payload.end());
+    update_string_views();
+}
+
+SPDLOG_INLINE log_msg_buffer::log_msg_buffer(log_msg_buffer &&other) SPDLOG_NOEXCEPT
+    : log_msg{other},
+      buffer{std::move(other.buffer)} {
+    update_string_views();
+}
+
+SPDLOG_INLINE log_msg_buffer &log_msg_buffer::operator=(const log_msg_buffer &other) {
+    log_msg::operator=(other);
+    buffer.clear();
+    buffer.append(other.buffer.data(), other.buffer.data() + other.buffer.size());
+    update_string_views();
+    return *this;
+}
+
+SPDLOG_INLINE log_msg_buffer &log_msg_buffer::operator=(log_msg_buffer &&other) SPDLOG_NOEXCEPT {
+    log_msg::operator=(other);
+    buffer = std::move(other.buffer);
+    update_string_views();
+    return *this;
+}
+
+SPDLOG_INLINE void log_msg_buffer::update_string_views() {
+    logger_name = string_view_t{buffer.data(), logger_name.size()};
+    payload = string_view_t{buffer.data() + logger_name.size(), payload.size()};
+}
+
+}  // namespace details
+}  // namespace spdlog
diff --git a/ext/spdlog/details/log_msg_buffer.h b/ext/spdlog/details/log_msg_buffer.h
new file mode 100644
index 0000000..1143b3b
--- /dev/null
+++ b/ext/spdlog/details/log_msg_buffer.h
@@ -0,0 +1,32 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#include <spdlog/details/log_msg.h>
+
+namespace spdlog {
+namespace details {
+
+// Extend log_msg with internal buffer to store its payload.
+// This is needed since log_msg holds string_views that points to stack data.
+
+class SPDLOG_API log_msg_buffer : public log_msg {
+    memory_buf_t buffer;
+    void update_string_views();
+
+public:
+    log_msg_buffer() = default;
+    explicit log_msg_buffer(const log_msg &orig_msg);
+    log_msg_buffer(const log_msg_buffer &other);
+    log_msg_buffer(log_msg_buffer &&other) SPDLOG_NOEXCEPT;
+    log_msg_buffer &operator=(const log_msg_buffer &other);
+    log_msg_buffer &operator=(log_msg_buffer &&other) SPDLOG_NOEXCEPT;
+};
+
+}  // namespace details
+}  // namespace spdlog
+
+#ifdef SPDLOG_HEADER_ONLY
+    #include "log_msg_buffer-inl.h"
+#endif
diff --git a/ext/spdlog/details/mpmc_blocking_q.h b/ext/spdlog/details/mpmc_blocking_q.h
new file mode 100644
index 0000000..5848cca
--- /dev/null
+++ b/ext/spdlog/details/mpmc_blocking_q.h
@@ -0,0 +1,177 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+// multi producer-multi consumer blocking queue.
+// enqueue(..) - will block until room found to put the new message.
+// enqueue_nowait(..) - will return immediately with false if no room left in
+// the queue.
+// dequeue_for(..) - will block until the queue is not empty or timeout have
+// passed.
+
+#include <spdlog/details/circular_q.h>
+
+#include <atomic>
+#include <condition_variable>
+#include <mutex>
+
+namespace spdlog {
+namespace details {
+
+template <typename T>
+class mpmc_blocking_queue {
+public:
+    using item_type = T;
+    explicit mpmc_blocking_queue(size_t max_items)
+        : q_(max_items) {}
+
+#ifndef __MINGW32__
+    // try to enqueue and block if no room left
+    void enqueue(T &&item) {
+        {
+            std::unique_lock<std::mutex> lock(queue_mutex_);
+            pop_cv_.wait(lock, [this] { return !this->q_.full(); });
+            q_.push_back(std::move(item));
+        }
+        push_cv_.notify_one();
+    }
+
+    // enqueue immediately. overrun oldest message in the queue if no room left.
+    void enqueue_nowait(T &&item) {
+        {
+            std::unique_lock<std::mutex> lock(queue_mutex_);
+            q_.push_back(std::move(item));
+        }
+        push_cv_.notify_one();
+    }
+
+    void enqueue_if_have_room(T &&item) {
+        bool pushed = false;
+        {
+            std::unique_lock<std::mutex> lock(queue_mutex_);
+            if (!q_.full()) {
+                q_.push_back(std::move(item));
+                pushed = true;
+            }
+        }
+
+        if (pushed) {
+            push_cv_.notify_one();
+        } else {
+            ++discard_counter_;
+        }
+    }
+
+    // dequeue with a timeout.
+    // Return true, if succeeded dequeue item, false otherwise
+    bool dequeue_for(T &popped_item, std::chrono::milliseconds wait_duration) {
+        {
+            std::unique_lock<std::mutex> lock(queue_mutex_);
+            if (!push_cv_.wait_for(lock, wait_duration, [this] { return !this->q_.empty(); })) {
+                return false;
+            }
+            popped_item = std::move(q_.front());
+            q_.pop_front();
+        }
+        pop_cv_.notify_one();
+        return true;
+    }
+
+    // blocking dequeue without a timeout.
+    void dequeue(T &popped_item) {
+        {
+            std::unique_lock<std::mutex> lock(queue_mutex_);
+            push_cv_.wait(lock, [this] { return !this->q_.empty(); });
+            popped_item = std::move(q_.front());
+            q_.pop_front();
+        }
+        pop_cv_.notify_one();
+    }
+
+#else
+    // apparently mingw deadlocks if the mutex is released before cv.notify_one(),
+    // so release the mutex at the very end each function.
+
+    // try to enqueue and block if no room left
+    void enqueue(T &&item) {
+        std::unique_lock<std::mutex> lock(queue_mutex_);
+        pop_cv_.wait(lock, [this] { return !this->q_.full(); });
+        q_.push_back(std::move(item));
+        push_cv_.notify_one();
+    }
+
+    // enqueue immediately. overrun oldest message in the queue if no room left.
+    void enqueue_nowait(T &&item) {
+        std::unique_lock<std::mutex> lock(queue_mutex_);
+        q_.push_back(std::move(item));
+        push_cv_.notify_one();
+    }
+
+    void enqueue_if_have_room(T &&item) {
+        bool pushed = false;
+        std::unique_lock<std::mutex> lock(queue_mutex_);
+        if (!q_.full()) {
+            q_.push_back(std::move(item));
+            pushed = true;
+        }
+
+        if (pushed) {
+            push_cv_.notify_one();
+        } else {
+            ++discard_counter_;
+        }
+    }
+
+    // dequeue with a timeout.
+    // Return true, if succeeded dequeue item, false otherwise
+    bool dequeue_for(T &popped_item, std::chrono::milliseconds wait_duration) {
+        std::unique_lock<std::mutex> lock(queue_mutex_);
+        if (!push_cv_.wait_for(lock, wait_duration, [this] { return !this->q_.empty(); })) {
+            return false;
+        }
+        popped_item = std::move(q_.front());
+        q_.pop_front();
+        pop_cv_.notify_one();
+        return true;
+    }
+
+    // blocking dequeue without a timeout.
+    void dequeue(T &popped_item) {
+        std::unique_lock<std::mutex> lock(queue_mutex_);
+        push_cv_.wait(lock, [this] { return !this->q_.empty(); });
+        popped_item = std::move(q_.front());
+        q_.pop_front();
+        pop_cv_.notify_one();
+    }
+
+#endif
+
+    size_t overrun_counter() {
+        std::lock_guard<std::mutex> lock(queue_mutex_);
+        return q_.overrun_counter();
+    }
+
+    size_t discard_counter() { return discard_counter_.load(std::memory_order_relaxed); }
+
+    size_t size() {
+        std::lock_guard<std::mutex> lock(queue_mutex_);
+        return q_.size();
+    }
+
+    void reset_overrun_counter() {
+        std::lock_guard<std::mutex> lock(queue_mutex_);
+        q_.reset_overrun_counter();
+    }
+
+    void reset_discard_counter() { discard_counter_.store(0, std::memory_order_relaxed); }
+
+private:
+    std::mutex queue_mutex_;
+    std::condition_variable push_cv_;
+    std::condition_variable pop_cv_;
+    spdlog::details::circular_q<T> q_;
+    std::atomic<size_t> discard_counter_{0};
+};
+}  // namespace details
+}  // namespace spdlog
diff --git a/ext/spdlog/details/null_mutex.h b/ext/spdlog/details/null_mutex.h
new file mode 100644
index 0000000..e3b3220
--- /dev/null
+++ b/ext/spdlog/details/null_mutex.h
@@ -0,0 +1,35 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#include <atomic>
+#include <utility>
+// null, no cost dummy "mutex" and dummy "atomic" int
+
+namespace spdlog {
+namespace details {
+struct null_mutex {
+    void lock() const {}
+    void unlock() const {}
+};
+
+struct null_atomic_int {
+    int value;
+    null_atomic_int() = default;
+
+    explicit null_atomic_int(int new_value)
+        : value(new_value) {}
+
+    int load(std::memory_order = std::memory_order_relaxed) const { return value; }
+
+    void store(int new_value, std::memory_order = std::memory_order_relaxed) { value = new_value; }
+
+    int exchange(int new_value, std::memory_order = std::memory_order_relaxed) {
+        std::swap(new_value, value);
+        return new_value;  // return value before the call
+    }
+};
+
+}  // namespace details
+}  // namespace spdlog
diff --git a/ext/spdlog/details/os-inl.h b/ext/spdlog/details/os-inl.h
new file mode 100644
index 0000000..8ee4230
--- /dev/null
+++ b/ext/spdlog/details/os-inl.h
@@ -0,0 +1,606 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#ifndef SPDLOG_HEADER_ONLY
+    #include <spdlog/details/os.h>
+#endif
+
+#include <spdlog/common.h>
+
+#include <algorithm>
+#include <array>
+#include <chrono>
+#include <cstdio>
+#include <cstdlib>
+#include <cstring>
+#include <ctime>
+#include <string>
+#include <sys/stat.h>
+#include <sys/types.h>
+#include <thread>
+
+#ifdef _WIN32
+    #include <spdlog/details/windows_include.h>
+    #include <fileapi.h>  // for FlushFileBuffers
+    #include <io.h>       // for _get_osfhandle, _isatty, _fileno
+    #include <process.h>  // for _get_pid
+
+    #ifdef __MINGW32__
+        #include <share.h>
+    #endif
+
+    #if defined(SPDLOG_WCHAR_TO_UTF8_SUPPORT) || defined(SPDLOG_WCHAR_FILENAMES)
+        #include <cassert>
+        #include <limits>
+    #endif
+
+    #include <direct.h>  // for _mkdir/_wmkdir
+
+#else  // unix
+
+    #include <fcntl.h>
+    #include <unistd.h>
+
+    #ifdef __linux__
+        #include <sys/syscall.h>  //Use gettid() syscall under linux to get thread id
+
+    #elif defined(_AIX)
+        #include <pthread.h>  // for pthread_getthrds_np
+
+    #elif defined(__DragonFly__) || defined(__FreeBSD__)
+        #include <pthread_np.h>  // for pthread_getthreadid_np
+
+    #elif defined(__NetBSD__)
+        #include <lwp.h>  // for _lwp_self
+
+    #elif defined(__sun)
+        #include <thread.h>  // for thr_self
+    #endif
+
+#endif  // unix
+
+#if defined __APPLE__
+    #include <AvailabilityMacros.h>
+#endif
+
+#ifndef __has_feature           // Clang - feature checking macros.
+    #define __has_feature(x) 0  // Compatibility with non-clang compilers.
+#endif
+
+namespace spdlog {
+namespace details {
+namespace os {
+
+SPDLOG_INLINE spdlog::log_clock::time_point now() SPDLOG_NOEXCEPT {
+#if defined __linux__ && defined SPDLOG_CLOCK_COARSE
+    timespec ts;
+    ::clock_gettime(CLOCK_REALTIME_COARSE, &ts);
+    return std::chrono::time_point<log_clock, typename log_clock::duration>(
+        std::chrono::duration_cast<typename log_clock::duration>(
+            std::chrono::seconds(ts.tv_sec) + std::chrono::nanoseconds(ts.tv_nsec)));
+
+#else
+    return log_clock::now();
+#endif
+}
+SPDLOG_INLINE std::tm localtime(const std::time_t &time_tt) SPDLOG_NOEXCEPT {
+#ifdef _WIN32
+    std::tm tm;
+    ::localtime_s(&tm, &time_tt);
+#else
+    std::tm tm;
+    ::localtime_r(&time_tt, &tm);
+#endif
+    return tm;
+}
+
+SPDLOG_INLINE std::tm localtime() SPDLOG_NOEXCEPT {
+    std::time_t now_t = ::time(nullptr);
+    return localtime(now_t);
+}
+
+SPDLOG_INLINE std::tm gmtime(const std::time_t &time_tt) SPDLOG_NOEXCEPT {
+#ifdef _WIN32
+    std::tm tm;
+    ::gmtime_s(&tm, &time_tt);
+#else
+    std::tm tm;
+    ::gmtime_r(&time_tt, &tm);
+#endif
+    return tm;
+}
+
+SPDLOG_INLINE std::tm gmtime() SPDLOG_NOEXCEPT {
+    std::time_t now_t = ::time(nullptr);
+    return gmtime(now_t);
+}
+
+// fopen_s on non windows for writing
+SPDLOG_INLINE bool fopen_s(FILE **fp, const filename_t &filename, const filename_t &mode) {
+#ifdef _WIN32
+    #ifdef SPDLOG_WCHAR_FILENAMES
+    *fp = ::_wfsopen((filename.c_str()), mode.c_str(), _SH_DENYNO);
+    #else
+    *fp = ::_fsopen((filename.c_str()), mode.c_str(), _SH_DENYNO);
+    #endif
+    #if defined(SPDLOG_PREVENT_CHILD_FD)
+    if (*fp != nullptr) {
+        auto file_handle = reinterpret_cast<HANDLE>(_get_osfhandle(::_fileno(*fp)));
+        if (!::SetHandleInformation(file_handle, HANDLE_FLAG_INHERIT, 0)) {
+            ::fclose(*fp);
+            *fp = nullptr;
+        }
+    }
+    #endif
+#else  // unix
+    #if defined(SPDLOG_PREVENT_CHILD_FD)
+    const int mode_flag = mode == SPDLOG_FILENAME_T("ab") ? O_APPEND : O_TRUNC;
+    const int fd =
+        ::open((filename.c_str()), O_CREAT | O_WRONLY | O_CLOEXEC | mode_flag, mode_t(0644));
+    if (fd == -1) {
+        return true;
+    }
+    *fp = ::fdopen(fd, mode.c_str());
+    if (*fp == nullptr) {
+        ::close(fd);
+    }
+    #else
+    *fp = ::fopen((filename.c_str()), mode.c_str());
+    #endif
+#endif
+
+    return *fp == nullptr;
+}
+
+SPDLOG_INLINE int remove(const filename_t &filename) SPDLOG_NOEXCEPT {
+#if defined(_WIN32) && defined(SPDLOG_WCHAR_FILENAMES)
+    return ::_wremove(filename.c_str());
+#else
+    return std::remove(filename.c_str());
+#endif
+}
+
+SPDLOG_INLINE int remove_if_exists(const filename_t &filename) SPDLOG_NOEXCEPT {
+    return path_exists(filename) ? remove(filename) : 0;
+}
+
+SPDLOG_INLINE int rename(const filename_t &filename1, const filename_t &filename2) SPDLOG_NOEXCEPT {
+#if defined(_WIN32) && defined(SPDLOG_WCHAR_FILENAMES)
+    return ::_wrename(filename1.c_str(), filename2.c_str());
+#else
+    return std::rename(filename1.c_str(), filename2.c_str());
+#endif
+}
+
+// Return true if path exists (file or directory)
+SPDLOG_INLINE bool path_exists(const filename_t &filename) SPDLOG_NOEXCEPT {
+#ifdef _WIN32
+    struct _stat buffer;
+    #ifdef SPDLOG_WCHAR_FILENAMES
+    return (::_wstat(filename.c_str(), &buffer) == 0);
+    #else
+    return (::_stat(filename.c_str(), &buffer) == 0);
+    #endif
+#else  // common linux/unix all have the stat system call
+    struct stat buffer;
+    return (::stat(filename.c_str(), &buffer) == 0);
+#endif
+}
+
+#ifdef _MSC_VER
+    // avoid warning about unreachable statement at the end of filesize()
+    #pragma warning(push)
+    #pragma warning(disable : 4702)
+#endif
+
+// Return file size according to open FILE* object
+SPDLOG_INLINE size_t filesize(FILE *f) {
+    if (f == nullptr) {
+        throw_spdlog_ex("Failed getting file size. fd is null");
+    }
+#if defined(_WIN32) && !defined(__CYGWIN__)
+    int fd = ::_fileno(f);
+    #if defined(_WIN64)  // 64 bits
+    __int64 ret = ::_filelengthi64(fd);
+    if (ret >= 0) {
+        return static_cast<size_t>(ret);
+    }
+
+    #else  // windows 32 bits
+    long ret = ::_filelength(fd);
+    if (ret >= 0) {
+        return static_cast<size_t>(ret);
+    }
+    #endif
+
+#else  // unix
+    // OpenBSD and AIX doesn't compile with :: before the fileno(..)
+    #if defined(__OpenBSD__) || defined(_AIX)
+    int fd = fileno(f);
+    #else
+    int fd = ::fileno(f);
+    #endif
+    // 64 bits(but not in osx, linux/musl or cygwin, where fstat64 is deprecated)
+    #if ((defined(__linux__) && defined(__GLIBC__)) || defined(__sun) || defined(_AIX)) && \
+        (defined(__LP64__) || defined(_LP64))
+    struct stat64 st;
+    if (::fstat64(fd, &st) == 0) {
+        return static_cast<size_t>(st.st_size);
+    }
+    #else  // other unix or linux 32 bits or cygwin
+    struct stat st;
+    if (::fstat(fd, &st) == 0) {
+        return static_cast<size_t>(st.st_size);
+    }
+    #endif
+#endif
+    throw_spdlog_ex("Failed getting file size from fd", errno);
+    return 0;  // will not be reached.
+}
+
+#ifdef _MSC_VER
+    #pragma warning(pop)
+#endif
+
+// Return utc offset in minutes or throw spdlog_ex on failure
+SPDLOG_INLINE int utc_minutes_offset(const std::tm &tm) {
+#ifdef _WIN32
+    #if _WIN32_WINNT < _WIN32_WINNT_WS08
+    TIME_ZONE_INFORMATION tzinfo;
+    auto rv = ::GetTimeZoneInformation(&tzinfo);
+    #else
+    DYNAMIC_TIME_ZONE_INFORMATION tzinfo;
+    auto rv = ::GetDynamicTimeZoneInformation(&tzinfo);
+    #endif
+    if (rv == TIME_ZONE_ID_INVALID) throw_spdlog_ex("Failed getting timezone info. ", errno);
+
+    int offset = -tzinfo.Bias;
+    if (tm.tm_isdst) {
+        offset -= tzinfo.DaylightBias;
+    } else {
+        offset -= tzinfo.StandardBias;
+    }
+    return offset;
+#else
+
+    #if defined(sun) || defined(__sun) || defined(_AIX) || \
+        (defined(__NEWLIB__) && !defined(__TM_GMTOFF)) ||  \
+        (!defined(_BSD_SOURCE) && !defined(_GNU_SOURCE))
+    // 'tm_gmtoff' field is BSD extension and it's missing on SunOS/Solaris
+    struct helper {
+        static long int calculate_gmt_offset(const std::tm &localtm = details::os::localtime(),
+                                             const std::tm &gmtm = details::os::gmtime()) {
+            int local_year = localtm.tm_year + (1900 - 1);
+            int gmt_year = gmtm.tm_year + (1900 - 1);
+
+            long int days = (
+                // difference in day of year
+                localtm.tm_yday -
+                gmtm.tm_yday
+
+                // + intervening leap days
+                + ((local_year >> 2) - (gmt_year >> 2)) - (local_year / 100 - gmt_year / 100) +
+                ((local_year / 100 >> 2) - (gmt_year / 100 >> 2))
+
+                // + difference in years * 365 */
+                + static_cast<long int>(local_year - gmt_year) * 365);
+
+            long int hours = (24 * days) + (localtm.tm_hour - gmtm.tm_hour);
+            long int mins = (60 * hours) + (localtm.tm_min - gmtm.tm_min);
+            long int secs = (60 * mins) + (localtm.tm_sec - gmtm.tm_sec);
+
+            return secs;
+        }
+    };
+
+    auto offset_seconds = helper::calculate_gmt_offset(tm);
+    #else
+    auto offset_seconds = tm.tm_gmtoff;
+    #endif
+
+    return static_cast<int>(offset_seconds / 60);
+#endif
+}
+
+// Return current thread id as size_t
+// It exists because the std::this_thread::get_id() is much slower(especially
+// under VS 2013)
+SPDLOG_INLINE size_t _thread_id() SPDLOG_NOEXCEPT {
+#ifdef _WIN32
+    return static_cast<size_t>(::GetCurrentThreadId());
+#elif defined(__linux__)
+    #if defined(__ANDROID__) && defined(__ANDROID_API__) && (__ANDROID_API__ < 21)
+        #define SYS_gettid __NR_gettid
+    #endif
+    return static_cast<size_t>(::syscall(SYS_gettid));
+#elif defined(_AIX)
+    struct __pthrdsinfo buf;
+    int reg_size = 0;
+    pthread_t pt = pthread_self();
+    int retval = pthread_getthrds_np(&pt, PTHRDSINFO_QUERY_TID, &buf, sizeof(buf), NULL, &reg_size);
+    int tid = (!retval) ? buf.__pi_tid : 0;
+    return static_cast<size_t>(tid);
+#elif defined(__DragonFly__) || defined(__FreeBSD__)
+    return static_cast<size_t>(::pthread_getthreadid_np());
+#elif defined(__NetBSD__)
+    return static_cast<size_t>(::_lwp_self());
+#elif defined(__OpenBSD__)
+    return static_cast<size_t>(::getthrid());
+#elif defined(__sun)
+    return static_cast<size_t>(::thr_self());
+#elif __APPLE__
+    uint64_t tid;
+    // There is no pthread_threadid_np prior to Mac OS X 10.6, and it is not supported on any PPC,
+    // including 10.6.8 Rosetta. __POWERPC__ is Apple-specific define encompassing ppc and ppc64.
+    #ifdef MAC_OS_X_VERSION_MAX_ALLOWED
+    {
+        #if (MAC_OS_X_VERSION_MAX_ALLOWED < 1060) || defined(__POWERPC__)
+        tid = pthread_mach_thread_np(pthread_self());
+        #elif MAC_OS_X_VERSION_MIN_REQUIRED < 1060
+        if (&pthread_threadid_np) {
+            pthread_threadid_np(nullptr, &tid);
+        } else {
+            tid = pthread_mach_thread_np(pthread_self());
+        }
+        #else
+        pthread_threadid_np(nullptr, &tid);
+        #endif
+    }
+    #else
+    pthread_threadid_np(nullptr, &tid);
+    #endif
+    return static_cast<size_t>(tid);
+#else  // Default to standard C++11 (other Unix)
+    return static_cast<size_t>(std::hash<std::thread::id>()(std::this_thread::get_id()));
+#endif
+}
+
+// Return current thread id as size_t (from thread local storage)
+SPDLOG_INLINE size_t thread_id() SPDLOG_NOEXCEPT {
+#if defined(SPDLOG_NO_TLS)
+    return _thread_id();
+#else  // cache thread id in tls
+    static thread_local const size_t tid = _thread_id();
+    return tid;
+#endif
+}
+
+// This is avoid msvc issue in sleep_for that happens if the clock changes.
+// See https://github.com/gabime/spdlog/issues/609
+SPDLOG_INLINE void sleep_for_millis(unsigned int milliseconds) SPDLOG_NOEXCEPT {
+#if defined(_WIN32)
+    ::Sleep(milliseconds);
+#else
+    std::this_thread::sleep_for(std::chrono::milliseconds(milliseconds));
+#endif
+}
+
+// wchar support for windows file names (SPDLOG_WCHAR_FILENAMES must be defined)
+#if defined(_WIN32) && defined(SPDLOG_WCHAR_FILENAMES)
+SPDLOG_INLINE std::string filename_to_str(const filename_t &filename) {
+    memory_buf_t buf;
+    wstr_to_utf8buf(filename, buf);
+    return SPDLOG_BUF_TO_STRING(buf);
+}
+#else
+SPDLOG_INLINE std::string filename_to_str(const filename_t &filename) { return filename; }
+#endif
+
+SPDLOG_INLINE int pid() SPDLOG_NOEXCEPT {
+#ifdef _WIN32
+    return conditional_static_cast<int>(::GetCurrentProcessId());
+#else
+    return conditional_static_cast<int>(::getpid());
+#endif
+}
+
+// Determine if the terminal supports colors
+// Based on: https://github.com/agauniyal/rang/
+SPDLOG_INLINE bool is_color_terminal() SPDLOG_NOEXCEPT {
+#ifdef _WIN32
+    return true;
+#else
+
+    static const bool result = []() {
+        const char *env_colorterm_p = std::getenv("COLORTERM");
+        if (env_colorterm_p != nullptr) {
+            return true;
+        }
+
+        static constexpr std::array<const char *, 16> terms = {
+            {"ansi", "color", "console", "cygwin", "gnome", "konsole", "kterm", "linux", "msys",
+             "putty", "rxvt", "screen", "vt100", "xterm", "alacritty", "vt102"}};
+
+        const char *env_term_p = std::getenv("TERM");
+        if (env_term_p == nullptr) {
+            return false;
+        }
+
+        return std::any_of(terms.begin(), terms.end(), [&](const char *term) {
+            return std::strstr(env_term_p, term) != nullptr;
+        });
+    }();
+
+    return result;
+#endif
+}
+
+// Determine if the terminal attached
+// Source: https://github.com/agauniyal/rang/
+SPDLOG_INLINE bool in_terminal(FILE *file) SPDLOG_NOEXCEPT {
+#ifdef _WIN32
+    return ::_isatty(_fileno(file)) != 0;
+#else
+    return ::isatty(fileno(file)) != 0;
+#endif
+}
+
+#if (defined(SPDLOG_WCHAR_TO_UTF8_SUPPORT) || defined(SPDLOG_WCHAR_FILENAMES)) && defined(_WIN32)
+SPDLOG_INLINE void wstr_to_utf8buf(wstring_view_t wstr, memory_buf_t &target) {
+    if (wstr.size() > static_cast<size_t>((std::numeric_limits<int>::max)()) / 4 - 1) {
+        throw_spdlog_ex("UTF-16 string is too big to be converted to UTF-8");
+    }
+
+    int wstr_size = static_cast<int>(wstr.size());
+    if (wstr_size == 0) {
+        target.resize(0);
+        return;
+    }
+
+    int result_size = static_cast<int>(target.capacity());
+    if ((wstr_size + 1) * 4 > result_size) {
+        result_size =
+            ::WideCharToMultiByte(CP_UTF8, 0, wstr.data(), wstr_size, NULL, 0, NULL, NULL);
+    }
+
+    if (result_size > 0) {
+        target.resize(result_size);
+        result_size = ::WideCharToMultiByte(CP_UTF8, 0, wstr.data(), wstr_size, target.data(),
+                                            result_size, NULL, NULL);
+
+        if (result_size > 0) {
+            target.resize(result_size);
+            return;
+        }
+    }
+
+    throw_spdlog_ex(
+        fmt_lib::format("WideCharToMultiByte failed. Last error: {}", ::GetLastError()));
+}
+
+SPDLOG_INLINE void utf8_to_wstrbuf(string_view_t str, wmemory_buf_t &target) {
+    if (str.size() > static_cast<size_t>((std::numeric_limits<int>::max)()) - 1) {
+        throw_spdlog_ex("UTF-8 string is too big to be converted to UTF-16");
+    }
+
+    int str_size = static_cast<int>(str.size());
+    if (str_size == 0) {
+        target.resize(0);
+        return;
+    }
+
+    // find the size to allocate for the result buffer
+    int result_size =
+        ::MultiByteToWideChar(CP_UTF8, 0, str.data(), str_size, NULL, 0);
+
+    if (result_size > 0) {
+        target.resize(result_size);
+        result_size = ::MultiByteToWideChar(CP_UTF8, 0, str.data(), str_size, target.data(),
+                                            result_size);
+        if (result_size > 0) {
+            assert(result_size == target.size());
+            return;
+        }
+    }
+
+    throw_spdlog_ex(
+        fmt_lib::format("MultiByteToWideChar failed. Last error: {}", ::GetLastError()));
+}
+#endif  // (defined(SPDLOG_WCHAR_TO_UTF8_SUPPORT) || defined(SPDLOG_WCHAR_FILENAMES)) &&
+        // defined(_WIN32)
+
+// return true on success
+static SPDLOG_INLINE bool mkdir_(const filename_t &path) {
+#ifdef _WIN32
+    #ifdef SPDLOG_WCHAR_FILENAMES
+    return ::_wmkdir(path.c_str()) == 0;
+    #else
+    return ::_mkdir(path.c_str()) == 0;
+    #endif
+#else
+    return ::mkdir(path.c_str(), mode_t(0755)) == 0;
+#endif
+}
+
+// create the given directory - and all directories leading to it
+// return true on success or if the directory already exists
+SPDLOG_INLINE bool create_dir(const filename_t &path) {
+    if (path_exists(path)) {
+        return true;
+    }
+
+    if (path.empty()) {
+        return false;
+    }
+
+    size_t search_offset = 0;
+    do {
+        auto token_pos = path.find_first_of(folder_seps_filename, search_offset);
+        // treat the entire path as a folder if no folder separator not found
+        if (token_pos == filename_t::npos) {
+            token_pos = path.size();
+        }
+
+        auto subdir = path.substr(0, token_pos);
+#ifdef _WIN32
+        // if subdir is just a drive letter, add a slash e.g. "c:"=>"c:\",
+        // otherwise path_exists(subdir) returns false (issue #3079)
+        const bool is_drive = subdir.length() == 2 && subdir[1] == ':';
+        if (is_drive) {
+            subdir += '\\';
+            token_pos++;
+        }
+#endif
+
+        if (!subdir.empty() && !path_exists(subdir) && !mkdir_(subdir)) {
+            return false;  // return error if failed creating dir
+        }
+        search_offset = token_pos + 1;
+    } while (search_offset < path.size());
+
+    return true;
+}
+
+// Return directory name from given path or empty string
+// "abc/file" => "abc"
+// "abc/" => "abc"
+// "abc" => ""
+// "abc///" => "abc//"
+SPDLOG_INLINE filename_t dir_name(const filename_t &path) {
+    auto pos = path.find_last_of(folder_seps_filename);
+    return pos != filename_t::npos ? path.substr(0, pos) : filename_t{};
+}
+
+std::string SPDLOG_INLINE getenv(const char *field) {
+#if defined(_MSC_VER)
+    #if defined(__cplusplus_winrt)
+    return std::string{};  // not supported under uwp
+    #else
+    size_t len = 0;
+    char buf[128];
+    bool ok = ::getenv_s(&len, buf, sizeof(buf), field) == 0;
+    return ok ? buf : std::string{};
+    #endif
+#else  // revert to getenv
+    char *buf = ::getenv(field);
+    return buf ? buf : std::string{};
+#endif
+}
+
+// Do fsync by FILE handlerpointer
+// Return true on success
+SPDLOG_INLINE bool fsync(FILE *fp) {
+#ifdef _WIN32
+    return FlushFileBuffers(reinterpret_cast<HANDLE>(_get_osfhandle(_fileno(fp)))) != 0;
+#else
+    return ::fsync(fileno(fp)) == 0;
+#endif
+}
+
+// Do non-locking fwrite if possible by the os or use the regular locking fwrite
+// Return true on success.
+SPDLOG_INLINE bool fwrite_bytes(const void *ptr, const size_t n_bytes, FILE *fp) {
+    #if defined(_WIN32) && defined(SPDLOG_FWRITE_UNLOCKED)
+    return _fwrite_nolock(ptr, 1, n_bytes, fp) == n_bytes;
+    #elif defined(SPDLOG_FWRITE_UNLOCKED)
+    return ::fwrite_unlocked(ptr, 1, n_bytes, fp) == n_bytes;
+    #else
+    return std::fwrite(ptr, 1, n_bytes, fp) == n_bytes;
+    #endif
+}
+
+}  // namespace os
+}  // namespace details
+}  // namespace spdlog
diff --git a/ext/spdlog/details/os.h b/ext/spdlog/details/os.h
new file mode 100644
index 0000000..5fd12ba
--- /dev/null
+++ b/ext/spdlog/details/os.h
@@ -0,0 +1,127 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#include <ctime>  // std::time_t
+#include <spdlog/common.h>
+
+namespace spdlog {
+namespace details {
+namespace os {
+
+SPDLOG_API spdlog::log_clock::time_point now() SPDLOG_NOEXCEPT;
+
+SPDLOG_API std::tm localtime(const std::time_t &time_tt) SPDLOG_NOEXCEPT;
+
+SPDLOG_API std::tm localtime() SPDLOG_NOEXCEPT;
+
+SPDLOG_API std::tm gmtime(const std::time_t &time_tt) SPDLOG_NOEXCEPT;
+
+SPDLOG_API std::tm gmtime() SPDLOG_NOEXCEPT;
+
+// eol definition
+#if !defined(SPDLOG_EOL)
+    #ifdef _WIN32
+        #define SPDLOG_EOL "\r\n"
+    #else
+        #define SPDLOG_EOL "\n"
+    #endif
+#endif
+
+SPDLOG_CONSTEXPR static const char *default_eol = SPDLOG_EOL;
+
+// folder separator
+#if !defined(SPDLOG_FOLDER_SEPS)
+    #ifdef _WIN32
+        #define SPDLOG_FOLDER_SEPS "\\/"
+    #else
+        #define SPDLOG_FOLDER_SEPS "/"
+    #endif
+#endif
+
+SPDLOG_CONSTEXPR static const char folder_seps[] = SPDLOG_FOLDER_SEPS;
+SPDLOG_CONSTEXPR static const filename_t::value_type folder_seps_filename[] =
+    SPDLOG_FILENAME_T(SPDLOG_FOLDER_SEPS);
+
+// fopen_s on non windows for writing
+SPDLOG_API bool fopen_s(FILE **fp, const filename_t &filename, const filename_t &mode);
+
+// Remove filename. return 0 on success
+SPDLOG_API int remove(const filename_t &filename) SPDLOG_NOEXCEPT;
+
+// Remove file if exists. return 0 on success
+// Note: Non atomic (might return failure to delete if concurrently deleted by other process/thread)
+SPDLOG_API int remove_if_exists(const filename_t &filename) SPDLOG_NOEXCEPT;
+
+SPDLOG_API int rename(const filename_t &filename1, const filename_t &filename2) SPDLOG_NOEXCEPT;
+
+// Return if file exists.
+SPDLOG_API bool path_exists(const filename_t &filename) SPDLOG_NOEXCEPT;
+
+// Return file size according to open FILE* object
+SPDLOG_API size_t filesize(FILE *f);
+
+// Return utc offset in minutes or throw spdlog_ex on failure
+SPDLOG_API int utc_minutes_offset(const std::tm &tm = details::os::localtime());
+
+// Return current thread id as size_t
+// It exists because the std::this_thread::get_id() is much slower(especially
+// under VS 2013)
+SPDLOG_API size_t _thread_id() SPDLOG_NOEXCEPT;
+
+// Return current thread id as size_t (from thread local storage)
+SPDLOG_API size_t thread_id() SPDLOG_NOEXCEPT;
+
+// This is avoid msvc issue in sleep_for that happens if the clock changes.
+// See https://github.com/gabime/spdlog/issues/609
+SPDLOG_API void sleep_for_millis(unsigned int milliseconds) SPDLOG_NOEXCEPT;
+
+SPDLOG_API std::string filename_to_str(const filename_t &filename);
+
+SPDLOG_API int pid() SPDLOG_NOEXCEPT;
+
+// Determine if the terminal supports colors
+// Source: https://github.com/agauniyal/rang/
+SPDLOG_API bool is_color_terminal() SPDLOG_NOEXCEPT;
+
+// Determine if the terminal attached
+// Source: https://github.com/agauniyal/rang/
+SPDLOG_API bool in_terminal(FILE *file) SPDLOG_NOEXCEPT;
+
+#if (defined(SPDLOG_WCHAR_TO_UTF8_SUPPORT) || defined(SPDLOG_WCHAR_FILENAMES)) && defined(_WIN32)
+SPDLOG_API void wstr_to_utf8buf(wstring_view_t wstr, memory_buf_t &target);
+
+SPDLOG_API void utf8_to_wstrbuf(string_view_t str, wmemory_buf_t &target);
+#endif
+
+// Return directory name from given path or empty string
+// "abc/file" => "abc"
+// "abc/" => "abc"
+// "abc" => ""
+// "abc///" => "abc//"
+SPDLOG_API filename_t dir_name(const filename_t &path);
+
+// Create a dir from the given path.
+// Return true if succeeded or if this dir already exists.
+SPDLOG_API bool create_dir(const filename_t &path);
+
+// non thread safe, cross platform getenv/getenv_s
+// return empty string if field not found
+SPDLOG_API std::string getenv(const char *field);
+
+// Do fsync by FILE objectpointer.
+// Return true on success.
+SPDLOG_API bool fsync(FILE *fp);
+
+// Do non-locking fwrite if possible by the os or use the regular locking fwrite
+// Return true on success.
+SPDLOG_API bool fwrite_bytes(const void *ptr, const size_t n_bytes, FILE *fp);
+
+}  // namespace os
+}  // namespace details
+}  // namespace spdlog
+
+#ifdef SPDLOG_HEADER_ONLY
+    #include "os-inl.h"
+#endif
diff --git a/ext/spdlog/details/periodic_worker-inl.h b/ext/spdlog/details/periodic_worker-inl.h
new file mode 100644
index 0000000..18f11fb
--- /dev/null
+++ b/ext/spdlog/details/periodic_worker-inl.h
@@ -0,0 +1,26 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#ifndef SPDLOG_HEADER_ONLY
+    #include <spdlog/details/periodic_worker.h>
+#endif
+
+namespace spdlog {
+namespace details {
+
+// stop the worker thread and join it
+SPDLOG_INLINE periodic_worker::~periodic_worker() {
+    if (worker_thread_.joinable()) {
+        {
+            std::lock_guard<std::mutex> lock(mutex_);
+            active_ = false;
+        }
+        cv_.notify_one();
+        worker_thread_.join();
+    }
+}
+
+}  // namespace details
+}  // namespace spdlog
diff --git a/ext/spdlog/details/periodic_worker.h b/ext/spdlog/details/periodic_worker.h
new file mode 100644
index 0000000..d647b66
--- /dev/null
+++ b/ext/spdlog/details/periodic_worker.h
@@ -0,0 +1,58 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+// periodic worker thread - periodically executes the given callback function.
+//
+// RAII over the owned thread:
+//    creates the thread on construction.
+//    stops and joins the thread on destruction (if the thread is executing a callback, wait for it
+//    to finish first).
+
+#include <chrono>
+#include <condition_variable>
+#include <functional>
+#include <mutex>
+#include <thread>
+namespace spdlog {
+namespace details {
+
+class SPDLOG_API periodic_worker {
+public:
+    template <typename Rep, typename Period>
+    periodic_worker(const std::function<void()> &callback_fun,
+                    std::chrono::duration<Rep, Period> interval) {
+        active_ = (interval > std::chrono::duration<Rep, Period>::zero());
+        if (!active_) {
+            return;
+        }
+
+        worker_thread_ = std::thread([this, callback_fun, interval]() {
+            for (;;) {
+                std::unique_lock<std::mutex> lock(this->mutex_);
+                if (this->cv_.wait_for(lock, interval, [this] { return !this->active_; })) {
+                    return;  // active_ == false, so exit this thread
+                }
+                callback_fun();
+            }
+        });
+    }
+    std::thread &get_thread() { return worker_thread_; }
+    periodic_worker(const periodic_worker &) = delete;
+    periodic_worker &operator=(const periodic_worker &) = delete;
+    // stop the worker thread and join it
+    ~periodic_worker();
+
+private:
+    bool active_;
+    std::thread worker_thread_;
+    std::mutex mutex_;
+    std::condition_variable cv_;
+};
+}  // namespace details
+}  // namespace spdlog
+
+#ifdef SPDLOG_HEADER_ONLY
+    #include "periodic_worker-inl.h"
+#endif
diff --git a/ext/spdlog/details/registry-inl.h b/ext/spdlog/details/registry-inl.h
new file mode 100644
index 0000000..f447848
--- /dev/null
+++ b/ext/spdlog/details/registry-inl.h
@@ -0,0 +1,261 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#ifndef SPDLOG_HEADER_ONLY
+    #include <spdlog/details/registry.h>
+#endif
+
+#include <spdlog/common.h>
+#include <spdlog/details/periodic_worker.h>
+#include <spdlog/logger.h>
+#include <spdlog/pattern_formatter.h>
+
+#ifndef SPDLOG_DISABLE_DEFAULT_LOGGER
+    // support for the default stdout color logger
+    #ifdef _WIN32
+        #include <spdlog/sinks/wincolor_sink.h>
+    #else
+        #include <spdlog/sinks/ansicolor_sink.h>
+    #endif
+#endif  // SPDLOG_DISABLE_DEFAULT_LOGGER
+
+#include <chrono>
+#include <functional>
+#include <memory>
+#include <string>
+#include <unordered_map>
+
+namespace spdlog {
+namespace details {
+
+SPDLOG_INLINE registry::registry()
+    : formatter_(new pattern_formatter()) {
+#ifndef SPDLOG_DISABLE_DEFAULT_LOGGER
+    // create default logger (ansicolor_stdout_sink_mt or wincolor_stdout_sink_mt in windows).
+    #ifdef _WIN32
+    auto color_sink = std::make_shared<sinks::wincolor_stdout_sink_mt>();
+    #else
+    auto color_sink = std::make_shared<sinks::ansicolor_stdout_sink_mt>();
+    #endif
+
+    const char *default_logger_name = "";
+    default_logger_ = std::make_shared<spdlog::logger>(default_logger_name, std::move(color_sink));
+    loggers_[default_logger_name] = default_logger_;
+
+#endif  // SPDLOG_DISABLE_DEFAULT_LOGGER
+}
+
+SPDLOG_INLINE registry::~registry() = default;
+
+SPDLOG_INLINE void registry::register_logger(std::shared_ptr<logger> new_logger) {
+    std::lock_guard<std::mutex> lock(logger_map_mutex_);
+    register_logger_(std::move(new_logger));
+}
+
+SPDLOG_INLINE void registry::initialize_logger(std::shared_ptr<logger> new_logger) {
+    std::lock_guard<std::mutex> lock(logger_map_mutex_);
+    new_logger->set_formatter(formatter_->clone());
+
+    if (err_handler_) {
+        new_logger->set_error_handler(err_handler_);
+    }
+
+    // set new level according to previously configured level or default level
+    auto it = log_levels_.find(new_logger->name());
+    auto new_level = it != log_levels_.end() ? it->second : global_log_level_;
+    new_logger->set_level(new_level);
+
+    new_logger->flush_on(flush_level_);
+
+    if (backtrace_n_messages_ > 0) {
+        new_logger->enable_backtrace(backtrace_n_messages_);
+    }
+
+    if (automatic_registration_) {
+        register_logger_(std::move(new_logger));
+    }
+}
+
+SPDLOG_INLINE std::shared_ptr<logger> registry::get(const std::string &logger_name) {
+    std::lock_guard<std::mutex> lock(logger_map_mutex_);
+    auto found = loggers_.find(logger_name);
+    return found == loggers_.end() ? nullptr : found->second;
+}
+
+SPDLOG_INLINE std::shared_ptr<logger> registry::default_logger() {
+    std::lock_guard<std::mutex> lock(logger_map_mutex_);
+    return default_logger_;
+}
+
+// Return raw ptr to the default logger.
+// To be used directly by the spdlog default api (e.g. spdlog::info)
+// This make the default API faster, but cannot be used concurrently with set_default_logger().
+// e.g do not call set_default_logger() from one thread while calling spdlog::info() from another.
+SPDLOG_INLINE logger *registry::get_default_raw() { return default_logger_.get(); }
+
+// set default logger.
+// default logger is stored in default_logger_ (for faster retrieval) and in the loggers_ map.
+SPDLOG_INLINE void registry::set_default_logger(std::shared_ptr<logger> new_default_logger) {
+    std::lock_guard<std::mutex> lock(logger_map_mutex_);
+    if (new_default_logger != nullptr) {
+        loggers_[new_default_logger->name()] = new_default_logger;
+    }
+    default_logger_ = std::move(new_default_logger);
+}
+
+SPDLOG_INLINE void registry::set_tp(std::shared_ptr<thread_pool> tp) {
+    std::lock_guard<std::recursive_mutex> lock(tp_mutex_);
+    tp_ = std::move(tp);
+}
+
+SPDLOG_INLINE std::shared_ptr<thread_pool> registry::get_tp() {
+    std::lock_guard<std::recursive_mutex> lock(tp_mutex_);
+    return tp_;
+}
+
+// Set global formatter. Each sink in each logger will get a clone of this object
+SPDLOG_INLINE void registry::set_formatter(std::unique_ptr<formatter> formatter) {
+    std::lock_guard<std::mutex> lock(logger_map_mutex_);
+    formatter_ = std::move(formatter);
+    for (auto &l : loggers_) {
+        l.second->set_formatter(formatter_->clone());
+    }
+}
+
+SPDLOG_INLINE void registry::enable_backtrace(size_t n_messages) {
+    std::lock_guard<std::mutex> lock(logger_map_mutex_);
+    backtrace_n_messages_ = n_messages;
+
+    for (auto &l : loggers_) {
+        l.second->enable_backtrace(n_messages);
+    }
+}
+
+SPDLOG_INLINE void registry::disable_backtrace() {
+    std::lock_guard<std::mutex> lock(logger_map_mutex_);
+    backtrace_n_messages_ = 0;
+    for (auto &l : loggers_) {
+        l.second->disable_backtrace();
+    }
+}
+
+SPDLOG_INLINE void registry::set_level(level::level_enum log_level) {
+    std::lock_guard<std::mutex> lock(logger_map_mutex_);
+    for (auto &l : loggers_) {
+        l.second->set_level(log_level);
+    }
+    global_log_level_ = log_level;
+}
+
+SPDLOG_INLINE void registry::flush_on(level::level_enum log_level) {
+    std::lock_guard<std::mutex> lock(logger_map_mutex_);
+    for (auto &l : loggers_) {
+        l.second->flush_on(log_level);
+    }
+    flush_level_ = log_level;
+}
+
+SPDLOG_INLINE void registry::set_error_handler(err_handler handler) {
+    std::lock_guard<std::mutex> lock(logger_map_mutex_);
+    for (auto &l : loggers_) {
+        l.second->set_error_handler(handler);
+    }
+    err_handler_ = std::move(handler);
+}
+
+SPDLOG_INLINE void registry::apply_all(
+    const std::function<void(const std::shared_ptr<logger>)> &fun) {
+    std::lock_guard<std::mutex> lock(logger_map_mutex_);
+    for (auto &l : loggers_) {
+        fun(l.second);
+    }
+}
+
+SPDLOG_INLINE void registry::flush_all() {
+    std::lock_guard<std::mutex> lock(logger_map_mutex_);
+    for (auto &l : loggers_) {
+        l.second->flush();
+    }
+}
+
+SPDLOG_INLINE void registry::drop(const std::string &logger_name) {
+    std::lock_guard<std::mutex> lock(logger_map_mutex_);
+    auto is_default_logger = default_logger_ && default_logger_->name() == logger_name;
+    loggers_.erase(logger_name);
+    if (is_default_logger) {
+        default_logger_.reset();
+    }
+}
+
+SPDLOG_INLINE void registry::drop_all() {
+    std::lock_guard<std::mutex> lock(logger_map_mutex_);
+    loggers_.clear();
+    default_logger_.reset();
+}
+
+// clean all resources and threads started by the registry
+SPDLOG_INLINE void registry::shutdown() {
+    {
+        std::lock_guard<std::mutex> lock(flusher_mutex_);
+        periodic_flusher_.reset();
+    }
+
+    drop_all();
+
+    {
+        std::lock_guard<std::recursive_mutex> lock(tp_mutex_);
+        tp_.reset();
+    }
+}
+
+SPDLOG_INLINE std::recursive_mutex &registry::tp_mutex() { return tp_mutex_; }
+
+SPDLOG_INLINE void registry::set_automatic_registration(bool automatic_registration) {
+    std::lock_guard<std::mutex> lock(logger_map_mutex_);
+    automatic_registration_ = automatic_registration;
+}
+
+SPDLOG_INLINE void registry::set_levels(log_levels levels, level::level_enum *global_level) {
+    std::lock_guard<std::mutex> lock(logger_map_mutex_);
+    log_levels_ = std::move(levels);
+    auto global_level_requested = global_level != nullptr;
+    global_log_level_ = global_level_requested ? *global_level : global_log_level_;
+
+    for (auto &logger : loggers_) {
+        auto logger_entry = log_levels_.find(logger.first);
+        if (logger_entry != log_levels_.end()) {
+            logger.second->set_level(logger_entry->second);
+        } else if (global_level_requested) {
+            logger.second->set_level(*global_level);
+        }
+    }
+}
+
+SPDLOG_INLINE registry &registry::instance() {
+    static registry s_instance;
+    return s_instance;
+}
+
+SPDLOG_INLINE void registry::apply_logger_env_levels(std::shared_ptr<logger> new_logger) {
+    std::lock_guard<std::mutex> lock(logger_map_mutex_);
+    auto it = log_levels_.find(new_logger->name());
+    auto new_level = it != log_levels_.end() ? it->second : global_log_level_;
+    new_logger->set_level(new_level);
+}
+
+SPDLOG_INLINE void registry::throw_if_exists_(const std::string &logger_name) {
+    if (loggers_.find(logger_name) != loggers_.end()) {
+        throw_spdlog_ex("logger with name '" + logger_name + "' already exists");
+    }
+}
+
+SPDLOG_INLINE void registry::register_logger_(std::shared_ptr<logger> new_logger) {
+    auto logger_name = new_logger->name();
+    throw_if_exists_(logger_name);
+    loggers_[logger_name] = std::move(new_logger);
+}
+
+}  // namespace details
+}  // namespace spdlog
diff --git a/ext/spdlog/details/registry.h b/ext/spdlog/details/registry.h
new file mode 100644
index 0000000..8afcbd6
--- /dev/null
+++ b/ext/spdlog/details/registry.h
@@ -0,0 +1,129 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+// Loggers registry of unique name->logger pointer
+// An attempt to create a logger with an already existing name will result with spdlog_ex exception.
+// If user requests a non existing logger, nullptr will be returned
+// This class is thread safe
+
+#include <spdlog/common.h>
+#include <spdlog/details/periodic_worker.h>
+
+#include <chrono>
+#include <functional>
+#include <memory>
+#include <mutex>
+#include <string>
+#include <unordered_map>
+
+namespace spdlog {
+class logger;
+
+namespace details {
+class thread_pool;
+
+class SPDLOG_API registry {
+public:
+    using log_levels = std::unordered_map<std::string, level::level_enum>;
+    registry(const registry &) = delete;
+    registry &operator=(const registry &) = delete;
+
+    void register_logger(std::shared_ptr<logger> new_logger);
+    void initialize_logger(std::shared_ptr<logger> new_logger);
+    std::shared_ptr<logger> get(const std::string &logger_name);
+    std::shared_ptr<logger> default_logger();
+
+    // Return raw ptr to the default logger.
+    // To be used directly by the spdlog default api (e.g. spdlog::info)
+    // This make the default API faster, but cannot be used concurrently with set_default_logger().
+    // e.g do not call set_default_logger() from one thread while calling spdlog::info() from
+    // another.
+    logger *get_default_raw();
+
+    // set default logger and add it to the registry if not registered already.
+    // default logger is stored in default_logger_ (for faster retrieval) and in the loggers_ map.
+    // Note: Make sure to unregister it when no longer needed or before calling again with a new
+    // logger.
+    void set_default_logger(std::shared_ptr<logger> new_default_logger);
+
+    void set_tp(std::shared_ptr<thread_pool> tp);
+
+    std::shared_ptr<thread_pool> get_tp();
+
+    // Set global formatter. Each sink in each logger will get a clone of this object
+    void set_formatter(std::unique_ptr<formatter> formatter);
+
+    void enable_backtrace(size_t n_messages);
+
+    void disable_backtrace();
+
+    void set_level(level::level_enum log_level);
+
+    void flush_on(level::level_enum log_level);
+
+    template <typename Rep, typename Period>
+    void flush_every(std::chrono::duration<Rep, Period> interval) {
+        std::lock_guard<std::mutex> lock(flusher_mutex_);
+        auto clbk = [this]() { this->flush_all(); };
+        periodic_flusher_ = details::make_unique<periodic_worker>(clbk, interval);
+    }
+
+    std::unique_ptr<periodic_worker> &get_flusher() {
+        std::lock_guard<std::mutex> lock(flusher_mutex_);
+        return periodic_flusher_;
+    }
+
+    void set_error_handler(err_handler handler);
+
+    void apply_all(const std::function<void(const std::shared_ptr<logger>)> &fun);
+
+    void flush_all();
+
+    void drop(const std::string &logger_name);
+
+    void drop_all();
+
+    // clean all resources and threads started by the registry
+    void shutdown();
+
+    std::recursive_mutex &tp_mutex();
+
+    void set_automatic_registration(bool automatic_registration);
+
+    // set levels for all existing/future loggers. global_level can be null if should not set.
+    void set_levels(log_levels levels, level::level_enum *global_level);
+
+    static registry &instance();
+
+    void apply_logger_env_levels(std::shared_ptr<logger> new_logger);
+
+private:
+    registry();
+    ~registry();
+
+    void throw_if_exists_(const std::string &logger_name);
+    void register_logger_(std::shared_ptr<logger> new_logger);
+    bool set_level_from_cfg_(logger *logger);
+    std::mutex logger_map_mutex_, flusher_mutex_;
+    std::recursive_mutex tp_mutex_;
+    std::unordered_map<std::string, std::shared_ptr<logger>> loggers_;
+    log_levels log_levels_;
+    std::unique_ptr<formatter> formatter_;
+    spdlog::level::level_enum global_log_level_ = level::info;
+    level::level_enum flush_level_ = level::off;
+    err_handler err_handler_;
+    std::shared_ptr<thread_pool> tp_;
+    std::unique_ptr<periodic_worker> periodic_flusher_;
+    std::shared_ptr<logger> default_logger_;
+    bool automatic_registration_ = true;
+    size_t backtrace_n_messages_ = 0;
+};
+
+}  // namespace details
+}  // namespace spdlog
+
+#ifdef SPDLOG_HEADER_ONLY
+    #include "registry-inl.h"
+#endif
diff --git a/ext/spdlog/details/synchronous_factory.h b/ext/spdlog/details/synchronous_factory.h
new file mode 100644
index 0000000..4bd5a51
--- /dev/null
+++ b/ext/spdlog/details/synchronous_factory.h
@@ -0,0 +1,22 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#include "registry.h"
+
+namespace spdlog {
+
+// Default logger factory-  creates synchronous loggers
+class logger;
+
+struct synchronous_factory {
+    template <typename Sink, typename... SinkArgs>
+    static std::shared_ptr<spdlog::logger> create(std::string logger_name, SinkArgs &&...args) {
+        auto sink = std::make_shared<Sink>(std::forward<SinkArgs>(args)...);
+        auto new_logger = std::make_shared<spdlog::logger>(std::move(logger_name), std::move(sink));
+        details::registry::instance().initialize_logger(new_logger);
+        return new_logger;
+    }
+};
+}  // namespace spdlog
diff --git a/ext/spdlog/details/tcp_client-windows.h b/ext/spdlog/details/tcp_client-windows.h
new file mode 100644
index 0000000..bf8f7b8
--- /dev/null
+++ b/ext/spdlog/details/tcp_client-windows.h
@@ -0,0 +1,135 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#define WIN32_LEAN_AND_MEAN
+// tcp client helper
+#include <spdlog/common.h>
+#include <spdlog/details/os.h>
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <string>
+#include <windows.h>
+#include <winsock2.h>
+#include <ws2tcpip.h>
+
+#pragma comment(lib, "Ws2_32.lib")
+#pragma comment(lib, "Mswsock.lib")
+#pragma comment(lib, "AdvApi32.lib")
+
+namespace spdlog {
+namespace details {
+class tcp_client {
+    SOCKET socket_ = INVALID_SOCKET;
+
+    static void init_winsock_() {
+        WSADATA wsaData;
+        auto rv = WSAStartup(MAKEWORD(2, 2), &wsaData);
+        if (rv != 0) {
+            throw_winsock_error_("WSAStartup failed", ::WSAGetLastError());
+        }
+    }
+
+    static void throw_winsock_error_(const std::string &msg, int last_error) {
+        char buf[512];
+        ::FormatMessageA(FORMAT_MESSAGE_FROM_SYSTEM | FORMAT_MESSAGE_IGNORE_INSERTS, NULL,
+                         last_error, MAKELANGID(LANG_NEUTRAL, SUBLANG_DEFAULT), buf,
+                         (sizeof(buf) / sizeof(char)), NULL);
+
+        throw_spdlog_ex(fmt_lib::format("tcp_sink - {}: {}", msg, buf));
+    }
+
+public:
+    tcp_client() { init_winsock_(); }
+
+    ~tcp_client() {
+        close();
+        ::WSACleanup();
+    }
+
+    bool is_connected() const { return socket_ != INVALID_SOCKET; }
+
+    void close() {
+        ::closesocket(socket_);
+        socket_ = INVALID_SOCKET;
+    }
+
+    SOCKET fd() const { return socket_; }
+
+    // try to connect or throw on failure
+    void connect(const std::string &host, int port) {
+        if (is_connected()) {
+            close();
+        }
+        struct addrinfo hints {};
+        ZeroMemory(&hints, sizeof(hints));
+
+        hints.ai_family = AF_UNSPEC;      // To work with IPv4, IPv6, and so on
+        hints.ai_socktype = SOCK_STREAM;  // TCP
+        hints.ai_flags = AI_NUMERICSERV;  // port passed as as numeric value
+        hints.ai_protocol = 0;
+
+        auto port_str = std::to_string(port);
+        struct addrinfo *addrinfo_result;
+        auto rv = ::getaddrinfo(host.c_str(), port_str.c_str(), &hints, &addrinfo_result);
+        int last_error = 0;
+        if (rv != 0) {
+            last_error = ::WSAGetLastError();
+            WSACleanup();
+            throw_winsock_error_("getaddrinfo failed", last_error);
+        }
+
+        // Try each address until we successfully connect(2).
+
+        for (auto *rp = addrinfo_result; rp != nullptr; rp = rp->ai_next) {
+            socket_ = socket(rp->ai_family, rp->ai_socktype, rp->ai_protocol);
+            if (socket_ == INVALID_SOCKET) {
+                last_error = ::WSAGetLastError();
+                WSACleanup();
+                continue;
+            }
+            if (::connect(socket_, rp->ai_addr, (int)rp->ai_addrlen) == 0) {
+                break;
+            } else {
+                last_error = ::WSAGetLastError();
+                close();
+            }
+        }
+        ::freeaddrinfo(addrinfo_result);
+        if (socket_ == INVALID_SOCKET) {
+            WSACleanup();
+            throw_winsock_error_("connect failed", last_error);
+        }
+
+        // set TCP_NODELAY
+        int enable_flag = 1;
+        ::setsockopt(socket_, IPPROTO_TCP, TCP_NODELAY, reinterpret_cast<char *>(&enable_flag),
+                     sizeof(enable_flag));
+    }
+
+    // Send exactly n_bytes of the given data.
+    // On error close the connection and throw.
+    void send(const char *data, size_t n_bytes) {
+        size_t bytes_sent = 0;
+        while (bytes_sent < n_bytes) {
+            const int send_flags = 0;
+            auto write_result =
+                ::send(socket_, data + bytes_sent, (int)(n_bytes - bytes_sent), send_flags);
+            if (write_result == SOCKET_ERROR) {
+                int last_error = ::WSAGetLastError();
+                close();
+                throw_winsock_error_("send failed", last_error);
+            }
+
+            if (write_result == 0)  // (probably should not happen but in any case..)
+            {
+                break;
+            }
+            bytes_sent += static_cast<size_t>(write_result);
+        }
+    }
+};
+}  // namespace details
+}  // namespace spdlog
diff --git a/ext/spdlog/details/tcp_client.h b/ext/spdlog/details/tcp_client.h
new file mode 100644
index 0000000..9d3c40d
--- /dev/null
+++ b/ext/spdlog/details/tcp_client.h
@@ -0,0 +1,127 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#ifdef _WIN32
+    #error include tcp_client-windows.h instead
+#endif
+
+// tcp client helper
+#include <spdlog/common.h>
+#include <spdlog/details/os.h>
+
+#include <arpa/inet.h>
+#include <netdb.h>
+#include <netinet/in.h>
+#include <netinet/tcp.h>
+#include <sys/socket.h>
+#include <unistd.h>
+
+#include <string>
+
+namespace spdlog {
+namespace details {
+class tcp_client {
+    int socket_ = -1;
+
+public:
+    bool is_connected() const { return socket_ != -1; }
+
+    void close() {
+        if (is_connected()) {
+            ::close(socket_);
+            socket_ = -1;
+        }
+    }
+
+    int fd() const { return socket_; }
+
+    ~tcp_client() { close(); }
+
+    // try to connect or throw on failure
+    void connect(const std::string &host, int port) {
+        close();
+        struct addrinfo hints {};
+        memset(&hints, 0, sizeof(struct addrinfo));
+        hints.ai_family = AF_UNSPEC;      // To work with IPv4, IPv6, and so on
+        hints.ai_socktype = SOCK_STREAM;  // TCP
+        hints.ai_flags = AI_NUMERICSERV;  // port passed as as numeric value
+        hints.ai_protocol = 0;
+
+        auto port_str = std::to_string(port);
+        struct addrinfo *addrinfo_result;
+        auto rv = ::getaddrinfo(host.c_str(), port_str.c_str(), &hints, &addrinfo_result);
+        if (rv != 0) {
+            throw_spdlog_ex(fmt_lib::format("::getaddrinfo failed: {}", gai_strerror(rv)));
+        }
+
+        // Try each address until we successfully connect(2).
+        int last_errno = 0;
+        for (auto *rp = addrinfo_result; rp != nullptr; rp = rp->ai_next) {
+#if defined(SOCK_CLOEXEC)
+            const int flags = SOCK_CLOEXEC;
+#else
+            const int flags = 0;
+#endif
+            socket_ = ::socket(rp->ai_family, rp->ai_socktype | flags, rp->ai_protocol);
+            if (socket_ == -1) {
+                last_errno = errno;
+                continue;
+            }
+            rv = ::connect(socket_, rp->ai_addr, rp->ai_addrlen);
+            if (rv == 0) {
+                break;
+            }
+            last_errno = errno;
+            ::close(socket_);
+            socket_ = -1;
+        }
+        ::freeaddrinfo(addrinfo_result);
+        if (socket_ == -1) {
+            throw_spdlog_ex("::connect failed", last_errno);
+        }
+
+        // set TCP_NODELAY
+        int enable_flag = 1;
+        ::setsockopt(socket_, IPPROTO_TCP, TCP_NODELAY, reinterpret_cast<char *>(&enable_flag),
+                     sizeof(enable_flag));
+
+        // prevent sigpipe on systems where MSG_NOSIGNAL is not available
+#if defined(SO_NOSIGPIPE) && !defined(MSG_NOSIGNAL)
+        ::setsockopt(socket_, SOL_SOCKET, SO_NOSIGPIPE, reinterpret_cast<char *>(&enable_flag),
+                     sizeof(enable_flag));
+#endif
+
+#if !defined(SO_NOSIGPIPE) && !defined(MSG_NOSIGNAL)
+    #error "tcp_sink would raise SIGPIPE since neither SO_NOSIGPIPE nor MSG_NOSIGNAL are available"
+#endif
+    }
+
+    // Send exactly n_bytes of the given data.
+    // On error close the connection and throw.
+    void send(const char *data, size_t n_bytes) {
+        size_t bytes_sent = 0;
+        while (bytes_sent < n_bytes) {
+#if defined(MSG_NOSIGNAL)
+            const int send_flags = MSG_NOSIGNAL;
+#else
+            const int send_flags = 0;
+#endif
+            auto write_result =
+                ::send(socket_, data + bytes_sent, n_bytes - bytes_sent, send_flags);
+            if (write_result < 0) {
+                close();
+                throw_spdlog_ex("write(2) failed", errno);
+            }
+
+            if (write_result == 0)  // (probably should not happen but in any case..)
+            {
+                break;
+            }
+            bytes_sent += static_cast<size_t>(write_result);
+        }
+    }
+};
+}  // namespace details
+}  // namespace spdlog
diff --git a/ext/spdlog/details/thread_pool-inl.h b/ext/spdlog/details/thread_pool-inl.h
new file mode 100644
index 0000000..17e01c0
--- /dev/null
+++ b/ext/spdlog/details/thread_pool-inl.h
@@ -0,0 +1,127 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#ifndef SPDLOG_HEADER_ONLY
+    #include <spdlog/details/thread_pool.h>
+#endif
+
+#include <cassert>
+#include <spdlog/common.h>
+
+namespace spdlog {
+namespace details {
+
+SPDLOG_INLINE thread_pool::thread_pool(size_t q_max_items,
+                                       size_t threads_n,
+                                       std::function<void()> on_thread_start,
+                                       std::function<void()> on_thread_stop)
+    : q_(q_max_items) {
+    if (threads_n == 0 || threads_n > 1000) {
+        throw_spdlog_ex(
+            "spdlog::thread_pool(): invalid threads_n param (valid "
+            "range is 1-1000)");
+    }
+    for (size_t i = 0; i < threads_n; i++) {
+        threads_.emplace_back([this, on_thread_start, on_thread_stop] {
+            on_thread_start();
+            this->thread_pool::worker_loop_();
+            on_thread_stop();
+        });
+    }
+}
+
+SPDLOG_INLINE thread_pool::thread_pool(size_t q_max_items,
+                                       size_t threads_n,
+                                       std::function<void()> on_thread_start)
+    : thread_pool(q_max_items, threads_n, on_thread_start, [] {}) {}
+
+SPDLOG_INLINE thread_pool::thread_pool(size_t q_max_items, size_t threads_n)
+    : thread_pool(
+          q_max_items, threads_n, [] {}, [] {}) {}
+
+// message all threads to terminate gracefully join them
+SPDLOG_INLINE thread_pool::~thread_pool() {
+    SPDLOG_TRY {
+        for (size_t i = 0; i < threads_.size(); i++) {
+            post_async_msg_(async_msg(async_msg_type::terminate), async_overflow_policy::block);
+        }
+
+        for (auto &t : threads_) {
+            t.join();
+        }
+    }
+    SPDLOG_CATCH_STD
+}
+
+void SPDLOG_INLINE thread_pool::post_log(async_logger_ptr &&worker_ptr,
+                                         const details::log_msg &msg,
+                                         async_overflow_policy overflow_policy) {
+    async_msg async_m(std::move(worker_ptr), async_msg_type::log, msg);
+    post_async_msg_(std::move(async_m), overflow_policy);
+}
+
+void SPDLOG_INLINE thread_pool::post_flush(async_logger_ptr &&worker_ptr,
+                                           async_overflow_policy overflow_policy) {
+    post_async_msg_(async_msg(std::move(worker_ptr), async_msg_type::flush), overflow_policy);
+}
+
+size_t SPDLOG_INLINE thread_pool::overrun_counter() { return q_.overrun_counter(); }
+
+void SPDLOG_INLINE thread_pool::reset_overrun_counter() { q_.reset_overrun_counter(); }
+
+size_t SPDLOG_INLINE thread_pool::discard_counter() { return q_.discard_counter(); }
+
+void SPDLOG_INLINE thread_pool::reset_discard_counter() { q_.reset_discard_counter(); }
+
+size_t SPDLOG_INLINE thread_pool::queue_size() { return q_.size(); }
+
+void SPDLOG_INLINE thread_pool::post_async_msg_(async_msg &&new_msg,
+                                                async_overflow_policy overflow_policy) {
+    if (overflow_policy == async_overflow_policy::block) {
+        q_.enqueue(std::move(new_msg));
+    } else if (overflow_policy == async_overflow_policy::overrun_oldest) {
+        q_.enqueue_nowait(std::move(new_msg));
+    } else {
+        assert(overflow_policy == async_overflow_policy::discard_new);
+        q_.enqueue_if_have_room(std::move(new_msg));
+    }
+}
+
+void SPDLOG_INLINE thread_pool::worker_loop_() {
+    while (process_next_msg_()) {
+    }
+}
+
+// process next message in the queue
+// return true if this thread should still be active (while no terminate msg
+// was received)
+bool SPDLOG_INLINE thread_pool::process_next_msg_() {
+    async_msg incoming_async_msg;
+    q_.dequeue(incoming_async_msg);
+
+    switch (incoming_async_msg.msg_type) {
+        case async_msg_type::log: {
+            incoming_async_msg.worker_ptr->backend_sink_it_(incoming_async_msg);
+            return true;
+        }
+        case async_msg_type::flush: {
+            incoming_async_msg.worker_ptr->backend_flush_();
+            return true;
+        }
+
+        case async_msg_type::terminate: {
+            return false;
+        }
+
+        default: {
+            assert(false);
+        }
+    }
+
+    return true;
+}
+
+}  // namespace details
+}  // namespace spdlog
diff --git a/ext/spdlog/details/thread_pool.h b/ext/spdlog/details/thread_pool.h
new file mode 100644
index 0000000..f22b078
--- /dev/null
+++ b/ext/spdlog/details/thread_pool.h
@@ -0,0 +1,117 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#include <spdlog/details/log_msg_buffer.h>
+#include <spdlog/details/mpmc_blocking_q.h>
+#include <spdlog/details/os.h>
+
+#include <chrono>
+#include <functional>
+#include <memory>
+#include <thread>
+#include <vector>
+
+namespace spdlog {
+class async_logger;
+
+namespace details {
+
+using async_logger_ptr = std::shared_ptr<spdlog::async_logger>;
+
+enum class async_msg_type { log, flush, terminate };
+
+// Async msg to move to/from the queue
+// Movable only. should never be copied
+struct async_msg : log_msg_buffer {
+    async_msg_type msg_type{async_msg_type::log};
+    async_logger_ptr worker_ptr;
+
+    async_msg() = default;
+    ~async_msg() = default;
+
+    // should only be moved in or out of the queue..
+    async_msg(const async_msg &) = delete;
+
+// support for vs2013 move
+#if defined(_MSC_VER) && _MSC_VER <= 1800
+    async_msg(async_msg &&other)
+        : log_msg_buffer(std::move(other)),
+          msg_type(other.msg_type),
+          worker_ptr(std::move(other.worker_ptr)) {}
+
+    async_msg &operator=(async_msg &&other) {
+        *static_cast<log_msg_buffer *>(this) = std::move(other);
+        msg_type = other.msg_type;
+        worker_ptr = std::move(other.worker_ptr);
+        return *this;
+    }
+#else  // (_MSC_VER) && _MSC_VER <= 1800
+    async_msg(async_msg &&) = default;
+    async_msg &operator=(async_msg &&) = default;
+#endif
+
+    // construct from log_msg with given type
+    async_msg(async_logger_ptr &&worker, async_msg_type the_type, const details::log_msg &m)
+        : log_msg_buffer{m},
+          msg_type{the_type},
+          worker_ptr{std::move(worker)} {}
+
+    async_msg(async_logger_ptr &&worker, async_msg_type the_type)
+        : log_msg_buffer{},
+          msg_type{the_type},
+          worker_ptr{std::move(worker)} {}
+
+    explicit async_msg(async_msg_type the_type)
+        : async_msg{nullptr, the_type} {}
+};
+
+class SPDLOG_API thread_pool {
+public:
+    using item_type = async_msg;
+    using q_type = details::mpmc_blocking_queue<item_type>;
+
+    thread_pool(size_t q_max_items,
+                size_t threads_n,
+                std::function<void()> on_thread_start,
+                std::function<void()> on_thread_stop);
+    thread_pool(size_t q_max_items, size_t threads_n, std::function<void()> on_thread_start);
+    thread_pool(size_t q_max_items, size_t threads_n);
+
+    // message all threads to terminate gracefully and join them
+    ~thread_pool();
+
+    thread_pool(const thread_pool &) = delete;
+    thread_pool &operator=(thread_pool &&) = delete;
+
+    void post_log(async_logger_ptr &&worker_ptr,
+                  const details::log_msg &msg,
+                  async_overflow_policy overflow_policy);
+    void post_flush(async_logger_ptr &&worker_ptr, async_overflow_policy overflow_policy);
+    size_t overrun_counter();
+    void reset_overrun_counter();
+    size_t discard_counter();
+    void reset_discard_counter();
+    size_t queue_size();
+
+private:
+    q_type q_;
+
+    std::vector<std::thread> threads_;
+
+    void post_async_msg_(async_msg &&new_msg, async_overflow_policy overflow_policy);
+    void worker_loop_();
+
+    // process next message in the queue
+    // return true if this thread should still be active (while no terminate msg
+    // was received)
+    bool process_next_msg_();
+};
+
+}  // namespace details
+}  // namespace spdlog
+
+#ifdef SPDLOG_HEADER_ONLY
+    #include "thread_pool-inl.h"
+#endif
diff --git a/ext/spdlog/details/udp_client-windows.h b/ext/spdlog/details/udp_client-windows.h
new file mode 100644
index 0000000..8b7c223
--- /dev/null
+++ b/ext/spdlog/details/udp_client-windows.h
@@ -0,0 +1,98 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+// Helper RAII over winsock udp client socket.
+// Will throw on construction if socket creation failed.
+
+#include <spdlog/common.h>
+#include <spdlog/details/os.h>
+#include <spdlog/details/windows_include.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string>
+#include <winsock2.h>
+#include <ws2tcpip.h>
+
+#if defined(_MSC_VER)
+    #pragma comment(lib, "Ws2_32.lib")
+    #pragma comment(lib, "Mswsock.lib")
+    #pragma comment(lib, "AdvApi32.lib")
+#endif
+
+namespace spdlog {
+namespace details {
+class udp_client {
+    static constexpr int TX_BUFFER_SIZE = 1024 * 10;
+    SOCKET socket_ = INVALID_SOCKET;
+    sockaddr_in addr_ = {};
+
+    static void init_winsock_() {
+        WSADATA wsaData;
+        auto rv = ::WSAStartup(MAKEWORD(2, 2), &wsaData);
+        if (rv != 0) {
+            throw_winsock_error_("WSAStartup failed", ::WSAGetLastError());
+        }
+    }
+
+    static void throw_winsock_error_(const std::string &msg, int last_error) {
+        char buf[512];
+        ::FormatMessageA(FORMAT_MESSAGE_FROM_SYSTEM | FORMAT_MESSAGE_IGNORE_INSERTS, NULL,
+                         last_error, MAKELANGID(LANG_NEUTRAL, SUBLANG_DEFAULT), buf,
+                         (sizeof(buf) / sizeof(char)), NULL);
+
+        throw_spdlog_ex(fmt_lib::format("udp_sink - {}: {}", msg, buf));
+    }
+
+    void cleanup_() {
+        if (socket_ != INVALID_SOCKET) {
+            ::closesocket(socket_);
+        }
+        socket_ = INVALID_SOCKET;
+        ::WSACleanup();
+    }
+
+public:
+    udp_client(const std::string &host, uint16_t port) {
+        init_winsock_();
+
+        addr_.sin_family = PF_INET;
+        addr_.sin_port = htons(port);
+        addr_.sin_addr.s_addr = INADDR_ANY;
+        if (InetPtonA(PF_INET, host.c_str(), &addr_.sin_addr.s_addr) != 1) {
+            int last_error = ::WSAGetLastError();
+            ::WSACleanup();
+            throw_winsock_error_("error: Invalid address!", last_error);
+        }
+
+        socket_ = ::socket(PF_INET, SOCK_DGRAM, 0);
+        if (socket_ == INVALID_SOCKET) {
+            int last_error = ::WSAGetLastError();
+            ::WSACleanup();
+            throw_winsock_error_("error: Create Socket failed", last_error);
+        }
+
+        int option_value = TX_BUFFER_SIZE;
+        if (::setsockopt(socket_, SOL_SOCKET, SO_SNDBUF,
+                         reinterpret_cast<const char *>(&option_value), sizeof(option_value)) < 0) {
+            int last_error = ::WSAGetLastError();
+            cleanup_();
+            throw_winsock_error_("error: setsockopt(SO_SNDBUF) Failed!", last_error);
+        }
+    }
+
+    ~udp_client() { cleanup_(); }
+
+    SOCKET fd() const { return socket_; }
+
+    void send(const char *data, size_t n_bytes) {
+        socklen_t tolen = sizeof(struct sockaddr);
+        if (::sendto(socket_, data, static_cast<int>(n_bytes), 0, (struct sockaddr *)&addr_,
+                     tolen) == -1) {
+            throw_spdlog_ex("sendto(2) failed", errno);
+        }
+    }
+};
+}  // namespace details
+}  // namespace spdlog
diff --git a/ext/spdlog/details/udp_client.h b/ext/spdlog/details/udp_client.h
new file mode 100644
index 0000000..95826f5
--- /dev/null
+++ b/ext/spdlog/details/udp_client.h
@@ -0,0 +1,81 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+// Helper RAII over unix udp client socket.
+// Will throw on construction if the socket creation failed.
+
+#ifdef _WIN32
+    #error "include udp_client-windows.h instead"
+#endif
+
+#include <arpa/inet.h>
+#include <cstring>
+#include <netdb.h>
+#include <netinet/in.h>
+#include <netinet/udp.h>
+#include <spdlog/common.h>
+#include <spdlog/details/os.h>
+#include <sys/socket.h>
+#include <unistd.h>
+
+#include <string>
+
+namespace spdlog {
+namespace details {
+
+class udp_client {
+    static constexpr int TX_BUFFER_SIZE = 1024 * 10;
+    int socket_ = -1;
+    struct sockaddr_in sockAddr_;
+
+    void cleanup_() {
+        if (socket_ != -1) {
+            ::close(socket_);
+            socket_ = -1;
+        }
+    }
+
+public:
+    udp_client(const std::string &host, uint16_t port) {
+        socket_ = ::socket(PF_INET, SOCK_DGRAM, 0);
+        if (socket_ < 0) {
+            throw_spdlog_ex("error: Create Socket Failed!");
+        }
+
+        int option_value = TX_BUFFER_SIZE;
+        if (::setsockopt(socket_, SOL_SOCKET, SO_SNDBUF,
+                         reinterpret_cast<const char *>(&option_value), sizeof(option_value)) < 0) {
+            cleanup_();
+            throw_spdlog_ex("error: setsockopt(SO_SNDBUF) Failed!");
+        }
+
+        sockAddr_.sin_family = AF_INET;
+        sockAddr_.sin_port = htons(port);
+
+        if (::inet_aton(host.c_str(), &sockAddr_.sin_addr) == 0) {
+            cleanup_();
+            throw_spdlog_ex("error: Invalid address!");
+        }
+
+        ::memset(sockAddr_.sin_zero, 0x00, sizeof(sockAddr_.sin_zero));
+    }
+
+    ~udp_client() { cleanup_(); }
+
+    int fd() const { return socket_; }
+
+    // Send exactly n_bytes of the given data.
+    // On error close the connection and throw.
+    void send(const char *data, size_t n_bytes) {
+        ssize_t toslen = 0;
+        socklen_t tolen = sizeof(struct sockaddr);
+        if ((toslen = ::sendto(socket_, data, n_bytes, 0, (struct sockaddr *)&sockAddr_, tolen)) ==
+            -1) {
+            throw_spdlog_ex("sendto(2) failed", errno);
+        }
+    }
+};
+}  // namespace details
+}  // namespace spdlog
diff --git a/ext/spdlog/details/windows_include.h b/ext/spdlog/details/windows_include.h
new file mode 100644
index 0000000..bbab59b
--- /dev/null
+++ b/ext/spdlog/details/windows_include.h
@@ -0,0 +1,11 @@
+#pragma once
+
+#ifndef NOMINMAX
+    #define NOMINMAX  // prevent windows redefining min/max
+#endif
+
+#ifndef WIN32_LEAN_AND_MEAN
+    #define WIN32_LEAN_AND_MEAN
+#endif
+
+#include <windows.h>
diff --git a/ext/spdlog/fmt/bin_to_hex.h b/ext/spdlog/fmt/bin_to_hex.h
new file mode 100644
index 0000000..c2998d5
--- /dev/null
+++ b/ext/spdlog/fmt/bin_to_hex.h
@@ -0,0 +1,224 @@
+//
+// Copyright(c) 2015 Gabi Melman.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+//
+
+#pragma once
+
+#include <cctype>
+#include <spdlog/common.h>
+
+#if defined(__has_include)
+    #if __has_include(<version>)
+        #include <version>
+    #endif
+#endif
+
+#if __cpp_lib_span >= 202002L
+    #include <span>
+#endif
+
+//
+// Support for logging binary data as hex
+// format flags, any combination of the following:
+// {:X} - print in uppercase.
+// {:s} - don't separate each byte with space.
+// {:p} - don't print the position on each line start.
+// {:n} - don't split the output to lines.
+// {:a} - show ASCII if :n is not set
+
+//
+// Examples:
+//
+// std::vector<char> v(200, 0x0b);
+// logger->info("Some buffer {}", spdlog::to_hex(v));
+// char buf[128];
+// logger->info("Some buffer {:X}", spdlog::to_hex(std::begin(buf), std::end(buf)));
+// logger->info("Some buffer {:X}", spdlog::to_hex(std::begin(buf), std::end(buf), 16));
+
+namespace spdlog {
+namespace details {
+
+template <typename It>
+class dump_info {
+public:
+    dump_info(It range_begin, It range_end, size_t size_per_line)
+        : begin_(range_begin),
+          end_(range_end),
+          size_per_line_(size_per_line) {}
+
+    // do not use begin() and end() to avoid collision with fmt/ranges
+    It get_begin() const { return begin_; }
+    It get_end() const { return end_; }
+    size_t size_per_line() const { return size_per_line_; }
+
+private:
+    It begin_, end_;
+    size_t size_per_line_;
+};
+}  // namespace details
+
+// create a dump_info that wraps the given container
+template <typename Container>
+inline details::dump_info<typename Container::const_iterator> to_hex(const Container &container,
+                                                                     size_t size_per_line = 32) {
+    static_assert(sizeof(typename Container::value_type) == 1,
+                  "sizeof(Container::value_type) != 1");
+    using Iter = typename Container::const_iterator;
+    return details::dump_info<Iter>(std::begin(container), std::end(container), size_per_line);
+}
+
+#if __cpp_lib_span >= 202002L
+
+template <typename Value, size_t Extent>
+inline details::dump_info<typename std::span<Value, Extent>::iterator> to_hex(
+    const std::span<Value, Extent> &container, size_t size_per_line = 32) {
+    using Container = std::span<Value, Extent>;
+    static_assert(sizeof(typename Container::value_type) == 1,
+                  "sizeof(Container::value_type) != 1");
+    using Iter = typename Container::iterator;
+    return details::dump_info<Iter>(std::begin(container), std::end(container), size_per_line);
+}
+
+#endif
+
+// create dump_info from ranges
+template <typename It>
+inline details::dump_info<It> to_hex(const It range_begin,
+                                     const It range_end,
+                                     size_t size_per_line = 32) {
+    return details::dump_info<It>(range_begin, range_end, size_per_line);
+}
+
+}  // namespace spdlog
+
+namespace
+#ifdef SPDLOG_USE_STD_FORMAT
+    std
+#else
+    fmt
+#endif
+{
+
+template <typename T>
+struct formatter<spdlog::details::dump_info<T>, char> {
+    const char delimiter = ' ';
+    bool put_newlines = true;
+    bool put_delimiters = true;
+    bool use_uppercase = false;
+    bool put_positions = true;  // position on start of each line
+    bool show_ascii = false;
+
+    // parse the format string flags
+    template <typename ParseContext>
+    SPDLOG_CONSTEXPR_FUNC auto parse(ParseContext &ctx) -> decltype(ctx.begin()) {
+        auto it = ctx.begin();
+        while (it != ctx.end() && *it != '}') {
+            switch (*it) {
+                case 'X':
+                    use_uppercase = true;
+                    break;
+                case 's':
+                    put_delimiters = false;
+                    break;
+                case 'p':
+                    put_positions = false;
+                    break;
+                case 'n':
+                    put_newlines = false;
+                    show_ascii = false;
+                    break;
+                case 'a':
+                    if (put_newlines) {
+                        show_ascii = true;
+                    }
+                    break;
+            }
+
+            ++it;
+        }
+        return it;
+    }
+
+    // format the given bytes range as hex
+    template <typename FormatContext, typename Container>
+    auto format(const spdlog::details::dump_info<Container> &the_range, FormatContext &ctx) const
+        -> decltype(ctx.out()) {
+        SPDLOG_CONSTEXPR const char *hex_upper = "0123456789ABCDEF";
+        SPDLOG_CONSTEXPR const char *hex_lower = "0123456789abcdef";
+        const char *hex_chars = use_uppercase ? hex_upper : hex_lower;
+
+#if !defined(SPDLOG_USE_STD_FORMAT) && FMT_VERSION < 60000
+        auto inserter = ctx.begin();
+#else
+        auto inserter = ctx.out();
+#endif
+
+        int size_per_line = static_cast<int>(the_range.size_per_line());
+        auto start_of_line = the_range.get_begin();
+        for (auto i = the_range.get_begin(); i != the_range.get_end(); i++) {
+            auto ch = static_cast<unsigned char>(*i);
+
+            if (put_newlines &&
+                (i == the_range.get_begin() || i - start_of_line >= size_per_line)) {
+                if (show_ascii && i != the_range.get_begin()) {
+                    *inserter++ = delimiter;
+                    *inserter++ = delimiter;
+                    for (auto j = start_of_line; j < i; j++) {
+                        auto pc = static_cast<unsigned char>(*j);
+                        *inserter++ = std::isprint(pc) ? static_cast<char>(*j) : '.';
+                    }
+                }
+
+                put_newline(inserter, static_cast<size_t>(i - the_range.get_begin()));
+
+                // put first byte without delimiter in front of it
+                *inserter++ = hex_chars[(ch >> 4) & 0x0f];
+                *inserter++ = hex_chars[ch & 0x0f];
+                start_of_line = i;
+                continue;
+            }
+
+            if (put_delimiters && i != the_range.get_begin()) {
+                *inserter++ = delimiter;
+            }
+
+            *inserter++ = hex_chars[(ch >> 4) & 0x0f];
+            *inserter++ = hex_chars[ch & 0x0f];
+        }
+        if (show_ascii)  // add ascii to last line
+        {
+            if (the_range.get_end() - the_range.get_begin() > size_per_line) {
+                auto blank_num = size_per_line - (the_range.get_end() - start_of_line);
+                while (blank_num-- > 0) {
+                    *inserter++ = delimiter;
+                    *inserter++ = delimiter;
+                    if (put_delimiters) {
+                        *inserter++ = delimiter;
+                    }
+                }
+            }
+            *inserter++ = delimiter;
+            *inserter++ = delimiter;
+            for (auto j = start_of_line; j != the_range.get_end(); j++) {
+                auto pc = static_cast<unsigned char>(*j);
+                *inserter++ = std::isprint(pc) ? static_cast<char>(*j) : '.';
+            }
+        }
+        return inserter;
+    }
+
+    // put newline(and position header)
+    template <typename It>
+    void put_newline(It inserter, std::size_t pos) const {
+#ifdef _WIN32
+        *inserter++ = '\r';
+#endif
+        *inserter++ = '\n';
+
+        if (put_positions) {
+            spdlog::fmt_lib::format_to(inserter, SPDLOG_FMT_STRING("{:04X}: "), pos);
+        }
+    }
+};
+}  // namespace std
diff --git a/ext/spdlog/fmt/bundled/args.h b/ext/spdlog/fmt/bundled/args.h
new file mode 100644
index 0000000..31a60e8
--- /dev/null
+++ b/ext/spdlog/fmt/bundled/args.h
@@ -0,0 +1,228 @@
+// Formatting library for C++ - dynamic argument lists
+//
+// Copyright (c) 2012 - present, Victor Zverovich
+// All rights reserved.
+//
+// For the license information refer to format.h.
+
+#ifndef FMT_ARGS_H_
+#define FMT_ARGS_H_
+
+#ifndef FMT_MODULE
+#  include <functional>  // std::reference_wrapper
+#  include <memory>      // std::unique_ptr
+#  include <vector>
+#endif
+
+#include "format.h"  // std_string_view
+
+FMT_BEGIN_NAMESPACE
+
+namespace detail {
+
+template <typename T> struct is_reference_wrapper : std::false_type {};
+template <typename T>
+struct is_reference_wrapper<std::reference_wrapper<T>> : std::true_type {};
+
+template <typename T> auto unwrap(const T& v) -> const T& { return v; }
+template <typename T>
+auto unwrap(const std::reference_wrapper<T>& v) -> const T& {
+  return static_cast<const T&>(v);
+}
+
+// node is defined outside dynamic_arg_list to workaround a C2504 bug in MSVC
+// 2022 (v17.10.0).
+//
+// Workaround for clang's -Wweak-vtables. Unlike for regular classes, for
+// templates it doesn't complain about inability to deduce single translation
+// unit for placing vtable. So node is made a fake template.
+template <typename = void> struct node {
+  virtual ~node() = default;
+  std::unique_ptr<node<>> next;
+};
+
+class dynamic_arg_list {
+  template <typename T> struct typed_node : node<> {
+    T value;
+
+    template <typename Arg>
+    FMT_CONSTEXPR typed_node(const Arg& arg) : value(arg) {}
+
+    template <typename Char>
+    FMT_CONSTEXPR typed_node(const basic_string_view<Char>& arg)
+        : value(arg.data(), arg.size()) {}
+  };
+
+  std::unique_ptr<node<>> head_;
+
+ public:
+  template <typename T, typename Arg> auto push(const Arg& arg) -> const T& {
+    auto new_node = std::unique_ptr<typed_node<T>>(new typed_node<T>(arg));
+    auto& value = new_node->value;
+    new_node->next = std::move(head_);
+    head_ = std::move(new_node);
+    return value;
+  }
+};
+}  // namespace detail
+
+/**
+ * A dynamic list of formatting arguments with storage.
+ *
+ * It can be implicitly converted into `fmt::basic_format_args` for passing
+ * into type-erased formatting functions such as `fmt::vformat`.
+ */
+template <typename Context>
+class dynamic_format_arg_store
+#if FMT_GCC_VERSION && FMT_GCC_VERSION < 409
+    // Workaround a GCC template argument substitution bug.
+    : public basic_format_args<Context>
+#endif
+{
+ private:
+  using char_type = typename Context::char_type;
+
+  template <typename T> struct need_copy {
+    static constexpr detail::type mapped_type =
+        detail::mapped_type_constant<T, Context>::value;
+
+    enum {
+      value = !(detail::is_reference_wrapper<T>::value ||
+                std::is_same<T, basic_string_view<char_type>>::value ||
+                std::is_same<T, detail::std_string_view<char_type>>::value ||
+                (mapped_type != detail::type::cstring_type &&
+                 mapped_type != detail::type::string_type &&
+                 mapped_type != detail::type::custom_type))
+    };
+  };
+
+  template <typename T>
+  using stored_type = conditional_t<
+      std::is_convertible<T, std::basic_string<char_type>>::value &&
+          !detail::is_reference_wrapper<T>::value,
+      std::basic_string<char_type>, T>;
+
+  // Storage of basic_format_arg must be contiguous.
+  std::vector<basic_format_arg<Context>> data_;
+  std::vector<detail::named_arg_info<char_type>> named_info_;
+
+  // Storage of arguments not fitting into basic_format_arg must grow
+  // without relocation because items in data_ refer to it.
+  detail::dynamic_arg_list dynamic_args_;
+
+  friend class basic_format_args<Context>;
+
+  auto get_types() const -> unsigned long long {
+    return detail::is_unpacked_bit | data_.size() |
+           (named_info_.empty()
+                ? 0ULL
+                : static_cast<unsigned long long>(detail::has_named_args_bit));
+  }
+
+  auto data() const -> const basic_format_arg<Context>* {
+    return named_info_.empty() ? data_.data() : data_.data() + 1;
+  }
+
+  template <typename T> void emplace_arg(const T& arg) {
+    data_.emplace_back(detail::make_arg<Context>(arg));
+  }
+
+  template <typename T>
+  void emplace_arg(const detail::named_arg<char_type, T>& arg) {
+    if (named_info_.empty()) {
+      constexpr const detail::named_arg_info<char_type>* zero_ptr{nullptr};
+      data_.insert(data_.begin(), {zero_ptr, 0});
+    }
+    data_.emplace_back(detail::make_arg<Context>(detail::unwrap(arg.value)));
+    auto pop_one = [](std::vector<basic_format_arg<Context>>* data) {
+      data->pop_back();
+    };
+    std::unique_ptr<std::vector<basic_format_arg<Context>>, decltype(pop_one)>
+        guard{&data_, pop_one};
+    named_info_.push_back({arg.name, static_cast<int>(data_.size() - 2u)});
+    data_[0].value_.named_args = {named_info_.data(), named_info_.size()};
+    guard.release();
+  }
+
+ public:
+  constexpr dynamic_format_arg_store() = default;
+
+  /**
+   * Adds an argument into the dynamic store for later passing to a formatting
+   * function.
+   *
+   * Note that custom types and string types (but not string views) are copied
+   * into the store dynamically allocating memory if necessary.
+   *
+   * **Example**:
+   *
+   *     fmt::dynamic_format_arg_store<fmt::format_context> store;
+   *     store.push_back(42);
+   *     store.push_back("abc");
+   *     store.push_back(1.5f);
+   *     std::string result = fmt::vformat("{} and {} and {}", store);
+   */
+  template <typename T> void push_back(const T& arg) {
+    if (detail::const_check(need_copy<T>::value))
+      emplace_arg(dynamic_args_.push<stored_type<T>>(arg));
+    else
+      emplace_arg(detail::unwrap(arg));
+  }
+
+  /**
+   * Adds a reference to the argument into the dynamic store for later passing
+   * to a formatting function.
+   *
+   * **Example**:
+   *
+   *     fmt::dynamic_format_arg_store<fmt::format_context> store;
+   *     char band[] = "Rolling Stones";
+   *     store.push_back(std::cref(band));
+   *     band[9] = 'c'; // Changing str affects the output.
+   *     std::string result = fmt::vformat("{}", store);
+   *     // result == "Rolling Scones"
+   */
+  template <typename T> void push_back(std::reference_wrapper<T> arg) {
+    static_assert(
+        need_copy<T>::value,
+        "objects of built-in types and string views are always copied");
+    emplace_arg(arg.get());
+  }
+
+  /**
+   * Adds named argument into the dynamic store for later passing to a
+   * formatting function. `std::reference_wrapper` is supported to avoid
+   * copying of the argument. The name is always copied into the store.
+   */
+  template <typename T>
+  void push_back(const detail::named_arg<char_type, T>& arg) {
+    const char_type* arg_name =
+        dynamic_args_.push<std::basic_string<char_type>>(arg.name).c_str();
+    if (detail::const_check(need_copy<T>::value)) {
+      emplace_arg(
+          fmt::arg(arg_name, dynamic_args_.push<stored_type<T>>(arg.value)));
+    } else {
+      emplace_arg(fmt::arg(arg_name, arg.value));
+    }
+  }
+
+  /// Erase all elements from the store.
+  void clear() {
+    data_.clear();
+    named_info_.clear();
+    dynamic_args_ = detail::dynamic_arg_list();
+  }
+
+  /// Reserves space to store at least `new_cap` arguments including
+  /// `new_cap_named` named arguments.
+  void reserve(size_t new_cap, size_t new_cap_named) {
+    FMT_ASSERT(new_cap >= new_cap_named,
+               "Set of arguments includes set of named arguments");
+    data_.reserve(new_cap);
+    named_info_.reserve(new_cap_named);
+  }
+};
+
+FMT_END_NAMESPACE
+
+#endif  // FMT_ARGS_H_
diff --git a/ext/spdlog/fmt/bundled/base.h b/ext/spdlog/fmt/bundled/base.h
new file mode 100644
index 0000000..6276494
--- /dev/null
+++ b/ext/spdlog/fmt/bundled/base.h
@@ -0,0 +1,3077 @@
+// Formatting library for C++ - the base API for char/UTF-8
+//
+// Copyright (c) 2012 - present, Victor Zverovich
+// All rights reserved.
+//
+// For the license information refer to format.h.
+
+#ifndef FMT_BASE_H_
+#define FMT_BASE_H_
+
+#if defined(FMT_IMPORT_STD) && !defined(FMT_MODULE)
+#  define FMT_MODULE
+#endif
+
+#ifndef FMT_MODULE
+#  include <limits.h>  // CHAR_BIT
+#  include <stdio.h>   // FILE
+#  include <string.h>  // strlen
+
+// <cstddef> is also included transitively from <type_traits>.
+#  include <cstddef>      // std::byte
+#  include <type_traits>  // std::enable_if
+#endif
+
+// The fmt library version in the form major * 10000 + minor * 100 + patch.
+#define FMT_VERSION 110002
+
+// Detect compiler versions.
+#if defined(__clang__) && !defined(__ibmxl__)
+#  define FMT_CLANG_VERSION (__clang_major__ * 100 + __clang_minor__)
+#else
+#  define FMT_CLANG_VERSION 0
+#endif
+#if defined(__GNUC__) && !defined(__clang__) && !defined(__INTEL_COMPILER)
+#  define FMT_GCC_VERSION (__GNUC__ * 100 + __GNUC_MINOR__)
+#else
+#  define FMT_GCC_VERSION 0
+#endif
+#if defined(__ICL)
+#  define FMT_ICC_VERSION __ICL
+#elif defined(__INTEL_COMPILER)
+#  define FMT_ICC_VERSION __INTEL_COMPILER
+#else
+#  define FMT_ICC_VERSION 0
+#endif
+#if defined(_MSC_VER)
+#  define FMT_MSC_VERSION _MSC_VER
+#else
+#  define FMT_MSC_VERSION 0
+#endif
+
+// Detect standard library versions.
+#ifdef _GLIBCXX_RELEASE
+#  define FMT_GLIBCXX_RELEASE _GLIBCXX_RELEASE
+#else
+#  define FMT_GLIBCXX_RELEASE 0
+#endif
+#ifdef _LIBCPP_VERSION
+#  define FMT_LIBCPP_VERSION _LIBCPP_VERSION
+#else
+#  define FMT_LIBCPP_VERSION 0
+#endif
+
+#ifdef _MSVC_LANG
+#  define FMT_CPLUSPLUS _MSVC_LANG
+#else
+#  define FMT_CPLUSPLUS __cplusplus
+#endif
+
+// Detect __has_*.
+#ifdef __has_feature
+#  define FMT_HAS_FEATURE(x) __has_feature(x)
+#else
+#  define FMT_HAS_FEATURE(x) 0
+#endif
+#ifdef __has_include
+#  define FMT_HAS_INCLUDE(x) __has_include(x)
+#else
+#  define FMT_HAS_INCLUDE(x) 0
+#endif
+#ifdef __has_cpp_attribute
+#  define FMT_HAS_CPP_ATTRIBUTE(x) __has_cpp_attribute(x)
+#else
+#  define FMT_HAS_CPP_ATTRIBUTE(x) 0
+#endif
+
+#define FMT_HAS_CPP14_ATTRIBUTE(attribute) \
+  (FMT_CPLUSPLUS >= 201402L && FMT_HAS_CPP_ATTRIBUTE(attribute))
+
+#define FMT_HAS_CPP17_ATTRIBUTE(attribute) \
+  (FMT_CPLUSPLUS >= 201703L && FMT_HAS_CPP_ATTRIBUTE(attribute))
+
+// Detect C++14 relaxed constexpr.
+#ifdef FMT_USE_CONSTEXPR
+// Use the provided definition.
+#elif FMT_GCC_VERSION >= 600 && FMT_CPLUSPLUS >= 201402L
+// GCC only allows throw in constexpr since version 6:
+// https://gcc.gnu.org/bugzilla/show_bug.cgi?id=67371.
+#  define FMT_USE_CONSTEXPR 1
+#elif FMT_ICC_VERSION
+#  define FMT_USE_CONSTEXPR 0  // https://github.com/fmtlib/fmt/issues/1628
+#elif FMT_HAS_FEATURE(cxx_relaxed_constexpr) || FMT_MSC_VERSION >= 1912
+#  define FMT_USE_CONSTEXPR 1
+#else
+#  define FMT_USE_CONSTEXPR 0
+#endif
+#if FMT_USE_CONSTEXPR
+#  define FMT_CONSTEXPR constexpr
+#else
+#  define FMT_CONSTEXPR
+#endif
+
+// Detect consteval, C++20 constexpr extensions and std::is_constant_evaluated.
+#if !defined(__cpp_lib_is_constant_evaluated)
+#  define FMT_USE_CONSTEVAL 0
+#elif FMT_CPLUSPLUS < 201709L
+#  define FMT_USE_CONSTEVAL 0
+#elif FMT_GLIBCXX_RELEASE && FMT_GLIBCXX_RELEASE < 10
+#  define FMT_USE_CONSTEVAL 0
+#elif FMT_LIBCPP_VERSION && FMT_LIBCPP_VERSION < 10000
+#  define FMT_USE_CONSTEVAL 0
+#elif defined(__apple_build_version__) && __apple_build_version__ < 14000029L
+#  define FMT_USE_CONSTEVAL 0  // consteval is broken in Apple clang < 14.
+#elif FMT_MSC_VERSION && FMT_MSC_VERSION < 1929
+#  define FMT_USE_CONSTEVAL 0  // consteval is broken in MSVC VS2019 < 16.10.
+#elif defined(__cpp_consteval)
+#  define FMT_USE_CONSTEVAL 1
+#elif FMT_GCC_VERSION >= 1002 || FMT_CLANG_VERSION >= 1101
+#  define FMT_USE_CONSTEVAL 1
+#else
+#  define FMT_USE_CONSTEVAL 0
+#endif
+#if FMT_USE_CONSTEVAL
+#  define FMT_CONSTEVAL consteval
+#  define FMT_CONSTEXPR20 constexpr
+#else
+#  define FMT_CONSTEVAL
+#  define FMT_CONSTEXPR20
+#endif
+
+#if defined(FMT_USE_NONTYPE_TEMPLATE_ARGS)
+// Use the provided definition.
+#elif defined(__NVCOMPILER)
+#  define FMT_USE_NONTYPE_TEMPLATE_ARGS 0
+#elif FMT_GCC_VERSION >= 903 && FMT_CPLUSPLUS >= 201709L
+#  define FMT_USE_NONTYPE_TEMPLATE_ARGS 1
+#elif defined(__cpp_nontype_template_args) && \
+    __cpp_nontype_template_args >= 201911L
+#  define FMT_USE_NONTYPE_TEMPLATE_ARGS 1
+#elif FMT_CLANG_VERSION >= 1200 && FMT_CPLUSPLUS >= 202002L
+#  define FMT_USE_NONTYPE_TEMPLATE_ARGS 1
+#else
+#  define FMT_USE_NONTYPE_TEMPLATE_ARGS 0
+#endif
+
+#ifdef FMT_USE_CONCEPTS
+// Use the provided definition.
+#elif defined(__cpp_concepts)
+#  define FMT_USE_CONCEPTS 1
+#else
+#  define FMT_USE_CONCEPTS 0
+#endif
+
+// Check if exceptions are disabled.
+#ifdef FMT_EXCEPTIONS
+// Use the provided definition.
+#elif defined(__GNUC__) && !defined(__EXCEPTIONS)
+#  define FMT_EXCEPTIONS 0
+#elif FMT_MSC_VERSION && !_HAS_EXCEPTIONS
+#  define FMT_EXCEPTIONS 0
+#else
+#  define FMT_EXCEPTIONS 1
+#endif
+#if FMT_EXCEPTIONS
+#  define FMT_TRY try
+#  define FMT_CATCH(x) catch (x)
+#else
+#  define FMT_TRY if (true)
+#  define FMT_CATCH(x) if (false)
+#endif
+
+#if FMT_HAS_CPP17_ATTRIBUTE(fallthrough)
+#  define FMT_FALLTHROUGH [[fallthrough]]
+#elif defined(__clang__)
+#  define FMT_FALLTHROUGH [[clang::fallthrough]]
+#elif FMT_GCC_VERSION >= 700 && \
+    (!defined(__EDG_VERSION__) || __EDG_VERSION__ >= 520)
+#  define FMT_FALLTHROUGH [[gnu::fallthrough]]
+#else
+#  define FMT_FALLTHROUGH
+#endif
+
+// Disable [[noreturn]] on MSVC/NVCC because of bogus unreachable code warnings.
+#if FMT_HAS_CPP_ATTRIBUTE(noreturn) && !FMT_MSC_VERSION && !defined(__NVCC__)
+#  define FMT_NORETURN [[noreturn]]
+#else
+#  define FMT_NORETURN
+#endif
+
+#ifndef FMT_NODISCARD
+#  if FMT_HAS_CPP17_ATTRIBUTE(nodiscard)
+#    define FMT_NODISCARD [[nodiscard]]
+#  else
+#    define FMT_NODISCARD
+#  endif
+#endif
+
+#ifdef FMT_DEPRECATED
+// Use the provided definition.
+#elif FMT_HAS_CPP14_ATTRIBUTE(deprecated)
+#  define FMT_DEPRECATED [[deprecated]]
+#else
+#  define FMT_DEPRECATED /* deprecated */
+#endif
+
+#ifdef FMT_INLINE
+// Use the provided definition.
+#elif FMT_GCC_VERSION || FMT_CLANG_VERSION
+#  define FMT_ALWAYS_INLINE inline __attribute__((always_inline))
+#else
+#  define FMT_ALWAYS_INLINE inline
+#endif
+// A version of FMT_INLINE to prevent code bloat in debug mode.
+#ifdef NDEBUG
+#  define FMT_INLINE FMT_ALWAYS_INLINE
+#else
+#  define FMT_INLINE inline
+#endif
+
+#if FMT_GCC_VERSION || FMT_CLANG_VERSION
+#  define FMT_VISIBILITY(value) __attribute__((visibility(value)))
+#else
+#  define FMT_VISIBILITY(value)
+#endif
+
+#ifndef FMT_GCC_PRAGMA
+// Workaround a _Pragma bug https://gcc.gnu.org/bugzilla/show_bug.cgi?id=59884
+// and an nvhpc warning: https://github.com/fmtlib/fmt/pull/2582.
+#  if FMT_GCC_VERSION >= 504 && !defined(__NVCOMPILER)
+#    define FMT_GCC_PRAGMA(arg) _Pragma(arg)
+#  else
+#    define FMT_GCC_PRAGMA(arg)
+#  endif
+#endif
+
+// GCC < 5 requires this-> in decltype.
+#if FMT_GCC_VERSION && FMT_GCC_VERSION < 500
+#  define FMT_DECLTYPE_THIS this->
+#else
+#  define FMT_DECLTYPE_THIS
+#endif
+
+#if FMT_MSC_VERSION
+#  define FMT_MSC_WARNING(...) __pragma(warning(__VA_ARGS__))
+#  define FMT_UNCHECKED_ITERATOR(It) \
+    using _Unchecked_type = It  // Mark iterator as checked.
+#else
+#  define FMT_MSC_WARNING(...)
+#  define FMT_UNCHECKED_ITERATOR(It) using unchecked_type = It
+#endif
+
+#ifndef FMT_BEGIN_NAMESPACE
+#  define FMT_BEGIN_NAMESPACE \
+    namespace fmt {           \
+    inline namespace v11 {
+#  define FMT_END_NAMESPACE \
+    }                       \
+    }
+#endif
+
+#ifndef FMT_EXPORT
+#  define FMT_EXPORT
+#  define FMT_BEGIN_EXPORT
+#  define FMT_END_EXPORT
+#endif
+
+#if !defined(FMT_HEADER_ONLY) && defined(_WIN32)
+#  if defined(FMT_LIB_EXPORT)
+#    define FMT_API __declspec(dllexport)
+#  elif defined(FMT_SHARED)
+#    define FMT_API __declspec(dllimport)
+#  endif
+#elif defined(FMT_LIB_EXPORT) || defined(FMT_SHARED)
+#  define FMT_API FMT_VISIBILITY("default")
+#endif
+#ifndef FMT_API
+#  define FMT_API
+#endif
+
+#ifndef FMT_UNICODE
+#  define FMT_UNICODE 1
+#endif
+
+// Check if rtti is available.
+#ifndef FMT_USE_RTTI
+// __RTTI is for EDG compilers. _CPPRTTI is for MSVC.
+#  if defined(__GXX_RTTI) || FMT_HAS_FEATURE(cxx_rtti) || defined(_CPPRTTI) || \
+      defined(__INTEL_RTTI__) || defined(__RTTI)
+#    define FMT_USE_RTTI 1
+#  else
+#    define FMT_USE_RTTI 0
+#  endif
+#endif
+
+#define FMT_FWD(...) static_cast<decltype(__VA_ARGS__)&&>(__VA_ARGS__)
+
+// Enable minimal optimizations for more compact code in debug mode.
+FMT_GCC_PRAGMA("GCC push_options")
+#if !defined(__OPTIMIZE__) && !defined(__CUDACC__)
+FMT_GCC_PRAGMA("GCC optimize(\"Og\")")
+#endif
+
+FMT_BEGIN_NAMESPACE
+
+// Implementations of enable_if_t and other metafunctions for older systems.
+template <bool B, typename T = void>
+using enable_if_t = typename std::enable_if<B, T>::type;
+template <bool B, typename T, typename F>
+using conditional_t = typename std::conditional<B, T, F>::type;
+template <bool B> using bool_constant = std::integral_constant<bool, B>;
+template <typename T>
+using remove_reference_t = typename std::remove_reference<T>::type;
+template <typename T>
+using remove_const_t = typename std::remove_const<T>::type;
+template <typename T>
+using remove_cvref_t = typename std::remove_cv<remove_reference_t<T>>::type;
+template <typename T> struct type_identity {
+  using type = T;
+};
+template <typename T> using type_identity_t = typename type_identity<T>::type;
+template <typename T>
+using make_unsigned_t = typename std::make_unsigned<T>::type;
+template <typename T>
+using underlying_t = typename std::underlying_type<T>::type;
+
+#if FMT_GCC_VERSION && FMT_GCC_VERSION < 500
+// A workaround for gcc 4.8 to make void_t work in a SFINAE context.
+template <typename...> struct void_t_impl {
+  using type = void;
+};
+template <typename... T> using void_t = typename void_t_impl<T...>::type;
+#else
+template <typename...> using void_t = void;
+#endif
+
+struct monostate {
+  constexpr monostate() {}
+};
+
+// An enable_if helper to be used in template parameters which results in much
+// shorter symbols: https://godbolt.org/z/sWw4vP. Extra parentheses are needed
+// to workaround a bug in MSVC 2019 (see #1140 and #1186).
+#ifdef FMT_DOC
+#  define FMT_ENABLE_IF(...)
+#else
+#  define FMT_ENABLE_IF(...) fmt::enable_if_t<(__VA_ARGS__), int> = 0
+#endif
+
+// This is defined in base.h instead of format.h to avoid injecting in std.
+// It is a template to avoid undesirable implicit conversions to std::byte.
+#ifdef __cpp_lib_byte
+template <typename T, FMT_ENABLE_IF(std::is_same<T, std::byte>::value)>
+inline auto format_as(T b) -> unsigned char {
+  return static_cast<unsigned char>(b);
+}
+#endif
+
+namespace detail {
+// Suppresses "unused variable" warnings with the method described in
+// https://herbsutter.com/2009/10/18/mailbag-shutting-up-compiler-warnings/.
+// (void)var does not work on many Intel compilers.
+template <typename... T> FMT_CONSTEXPR void ignore_unused(const T&...) {}
+
+constexpr auto is_constant_evaluated(bool default_value = false) noexcept
+    -> bool {
+// Workaround for incompatibility between libstdc++ consteval-based
+// std::is_constant_evaluated() implementation and clang-14:
+// https://github.com/fmtlib/fmt/issues/3247.
+#if FMT_CPLUSPLUS >= 202002L && FMT_GLIBCXX_RELEASE >= 12 && \
+    (FMT_CLANG_VERSION >= 1400 && FMT_CLANG_VERSION < 1500)
+  ignore_unused(default_value);
+  return __builtin_is_constant_evaluated();
+#elif defined(__cpp_lib_is_constant_evaluated)
+  ignore_unused(default_value);
+  return std::is_constant_evaluated();
+#else
+  return default_value;
+#endif
+}
+
+// Suppresses "conditional expression is constant" warnings.
+template <typename T> constexpr auto const_check(T value) -> T { return value; }
+
+FMT_NORETURN FMT_API void assert_fail(const char* file, int line,
+                                      const char* message);
+
+#if defined(FMT_ASSERT)
+// Use the provided definition.
+#elif defined(NDEBUG)
+// FMT_ASSERT is not empty to avoid -Wempty-body.
+#  define FMT_ASSERT(condition, message) \
+    fmt::detail::ignore_unused((condition), (message))
+#else
+#  define FMT_ASSERT(condition, message)                                    \
+    ((condition) /* void() fails with -Winvalid-constexpr on clang 4.0.1 */ \
+         ? (void)0                                                          \
+         : fmt::detail::assert_fail(__FILE__, __LINE__, (message)))
+#endif
+
+#ifdef FMT_USE_INT128
+// Do nothing.
+#elif defined(__SIZEOF_INT128__) && !defined(__NVCC__) && \
+    !(FMT_CLANG_VERSION && FMT_MSC_VERSION)
+#  define FMT_USE_INT128 1
+using int128_opt = __int128_t;  // An optional native 128-bit integer.
+using uint128_opt = __uint128_t;
+template <typename T> inline auto convert_for_visit(T value) -> T {
+  return value;
+}
+#else
+#  define FMT_USE_INT128 0
+#endif
+#if !FMT_USE_INT128
+enum class int128_opt {};
+enum class uint128_opt {};
+// Reduce template instantiations.
+template <typename T> auto convert_for_visit(T) -> monostate { return {}; }
+#endif
+
+// Casts a nonnegative integer to unsigned.
+template <typename Int>
+FMT_CONSTEXPR auto to_unsigned(Int value) -> make_unsigned_t<Int> {
+  FMT_ASSERT(std::is_unsigned<Int>::value || value >= 0, "negative value");
+  return static_cast<make_unsigned_t<Int>>(value);
+}
+
+// A heuristic to detect std::string and std::[experimental::]string_view.
+// It is mainly used to avoid dependency on <[experimental/]string_view>.
+template <typename T, typename Enable = void>
+struct is_std_string_like : std::false_type {};
+template <typename T>
+struct is_std_string_like<T, void_t<decltype(std::declval<T>().find_first_of(
+                                 typename T::value_type(), 0))>>
+    : std::is_convertible<decltype(std::declval<T>().data()),
+                          const typename T::value_type*> {};
+
+// Returns true iff the literal encoding is UTF-8.
+constexpr auto is_utf8_enabled() -> bool {
+  // Avoid an MSVC sign extension bug: https://github.com/fmtlib/fmt/pull/2297.
+  using uchar = unsigned char;
+  return sizeof("\u00A7") == 3 && uchar("\u00A7"[0]) == 0xC2 &&
+         uchar("\u00A7"[1]) == 0xA7;
+}
+constexpr auto use_utf8() -> bool {
+  return !FMT_MSC_VERSION || is_utf8_enabled();
+}
+
+static_assert(!FMT_UNICODE || use_utf8(),
+              "Unicode support requires compiling with /utf-8");
+
+template <typename Char> FMT_CONSTEXPR auto length(const Char* s) -> size_t {
+  size_t len = 0;
+  while (*s++) ++len;
+  return len;
+}
+
+template <typename Char>
+FMT_CONSTEXPR auto compare(const Char* s1, const Char* s2, std::size_t n)
+    -> int {
+  if (!is_constant_evaluated() && sizeof(Char) == 1) return memcmp(s1, s2, n);
+  for (; n != 0; ++s1, ++s2, --n) {
+    if (*s1 < *s2) return -1;
+    if (*s1 > *s2) return 1;
+  }
+  return 0;
+}
+
+namespace adl {
+using namespace std;
+
+template <typename Container>
+auto invoke_back_inserter()
+    -> decltype(back_inserter(std::declval<Container&>()));
+}  // namespace adl
+
+template <typename It, typename Enable = std::true_type>
+struct is_back_insert_iterator : std::false_type {};
+
+template <typename It>
+struct is_back_insert_iterator<
+    It, bool_constant<std::is_same<
+            decltype(adl::invoke_back_inserter<typename It::container_type>()),
+            It>::value>> : std::true_type {};
+
+// Extracts a reference to the container from *insert_iterator.
+template <typename OutputIt>
+inline auto get_container(OutputIt it) -> typename OutputIt::container_type& {
+  struct accessor : OutputIt {
+    accessor(OutputIt base) : OutputIt(base) {}
+    using OutputIt::container;
+  };
+  return *accessor(it).container;
+}
+}  // namespace detail
+
+// Checks whether T is a container with contiguous storage.
+template <typename T> struct is_contiguous : std::false_type {};
+
+/**
+ * An implementation of `std::basic_string_view` for pre-C++17. It provides a
+ * subset of the API. `fmt::basic_string_view` is used for format strings even
+ * if `std::basic_string_view` is available to prevent issues when a library is
+ * compiled with a different `-std` option than the client code (which is not
+ * recommended).
+ */
+FMT_EXPORT
+template <typename Char> class basic_string_view {
+ private:
+  const Char* data_;
+  size_t size_;
+
+ public:
+  using value_type = Char;
+  using iterator = const Char*;
+
+  constexpr basic_string_view() noexcept : data_(nullptr), size_(0) {}
+
+  /// Constructs a string reference object from a C string and a size.
+  constexpr basic_string_view(const Char* s, size_t count) noexcept
+      : data_(s), size_(count) {}
+
+  constexpr basic_string_view(std::nullptr_t) = delete;
+
+  /// Constructs a string reference object from a C string.
+  FMT_CONSTEXPR20
+  basic_string_view(const Char* s)
+      : data_(s),
+        size_(detail::const_check(std::is_same<Char, char>::value &&
+                                  !detail::is_constant_evaluated(false))
+                  ? strlen(reinterpret_cast<const char*>(s))
+                  : detail::length(s)) {}
+
+  /// Constructs a string reference from a `std::basic_string` or a
+  /// `std::basic_string_view` object.
+  template <typename S,
+            FMT_ENABLE_IF(detail::is_std_string_like<S>::value&& std::is_same<
+                          typename S::value_type, Char>::value)>
+  FMT_CONSTEXPR basic_string_view(const S& s) noexcept
+      : data_(s.data()), size_(s.size()) {}
+
+  /// Returns a pointer to the string data.
+  constexpr auto data() const noexcept -> const Char* { return data_; }
+
+  /// Returns the string size.
+  constexpr auto size() const noexcept -> size_t { return size_; }
+
+  constexpr auto begin() const noexcept -> iterator { return data_; }
+  constexpr auto end() const noexcept -> iterator { return data_ + size_; }
+
+  constexpr auto operator[](size_t pos) const noexcept -> const Char& {
+    return data_[pos];
+  }
+
+  FMT_CONSTEXPR void remove_prefix(size_t n) noexcept {
+    data_ += n;
+    size_ -= n;
+  }
+
+  FMT_CONSTEXPR auto starts_with(basic_string_view<Char> sv) const noexcept
+      -> bool {
+    return size_ >= sv.size_ && detail::compare(data_, sv.data_, sv.size_) == 0;
+  }
+  FMT_CONSTEXPR auto starts_with(Char c) const noexcept -> bool {
+    return size_ >= 1 && *data_ == c;
+  }
+  FMT_CONSTEXPR auto starts_with(const Char* s) const -> bool {
+    return starts_with(basic_string_view<Char>(s));
+  }
+
+  // Lexicographically compare this string reference to other.
+  FMT_CONSTEXPR auto compare(basic_string_view other) const -> int {
+    size_t str_size = size_ < other.size_ ? size_ : other.size_;
+    int result = detail::compare(data_, other.data_, str_size);
+    if (result == 0)
+      result = size_ == other.size_ ? 0 : (size_ < other.size_ ? -1 : 1);
+    return result;
+  }
+
+  FMT_CONSTEXPR friend auto operator==(basic_string_view lhs,
+                                       basic_string_view rhs) -> bool {
+    return lhs.compare(rhs) == 0;
+  }
+  friend auto operator!=(basic_string_view lhs, basic_string_view rhs) -> bool {
+    return lhs.compare(rhs) != 0;
+  }
+  friend auto operator<(basic_string_view lhs, basic_string_view rhs) -> bool {
+    return lhs.compare(rhs) < 0;
+  }
+  friend auto operator<=(basic_string_view lhs, basic_string_view rhs) -> bool {
+    return lhs.compare(rhs) <= 0;
+  }
+  friend auto operator>(basic_string_view lhs, basic_string_view rhs) -> bool {
+    return lhs.compare(rhs) > 0;
+  }
+  friend auto operator>=(basic_string_view lhs, basic_string_view rhs) -> bool {
+    return lhs.compare(rhs) >= 0;
+  }
+};
+
+FMT_EXPORT
+using string_view = basic_string_view<char>;
+
+/// Specifies if `T` is a character type. Can be specialized by users.
+FMT_EXPORT
+template <typename T> struct is_char : std::false_type {};
+template <> struct is_char<char> : std::true_type {};
+
+namespace detail {
+
+// Constructs fmt::basic_string_view<Char> from types implicitly convertible
+// to it, deducing Char. Explicitly convertible types such as the ones returned
+// from FMT_STRING are intentionally excluded.
+template <typename Char, FMT_ENABLE_IF(is_char<Char>::value)>
+constexpr auto to_string_view(const Char* s) -> basic_string_view<Char> {
+  return s;
+}
+template <typename T, FMT_ENABLE_IF(is_std_string_like<T>::value)>
+constexpr auto to_string_view(const T& s)
+    -> basic_string_view<typename T::value_type> {
+  return s;
+}
+template <typename Char>
+constexpr auto to_string_view(basic_string_view<Char> s)
+    -> basic_string_view<Char> {
+  return s;
+}
+
+template <typename T, typename Enable = void>
+struct has_to_string_view : std::false_type {};
+// detail:: is intentional since to_string_view is not an extension point.
+template <typename T>
+struct has_to_string_view<
+    T, void_t<decltype(detail::to_string_view(std::declval<T>()))>>
+    : std::true_type {};
+
+template <typename Char, Char... C> struct string_literal {
+  static constexpr Char value[sizeof...(C)] = {C...};
+  constexpr operator basic_string_view<Char>() const {
+    return {value, sizeof...(C)};
+  }
+};
+#if FMT_CPLUSPLUS < 201703L
+template <typename Char, Char... C>
+constexpr Char string_literal<Char, C...>::value[sizeof...(C)];
+#endif
+
+enum class type {
+  none_type,
+  // Integer types should go first,
+  int_type,
+  uint_type,
+  long_long_type,
+  ulong_long_type,
+  int128_type,
+  uint128_type,
+  bool_type,
+  char_type,
+  last_integer_type = char_type,
+  // followed by floating-point types.
+  float_type,
+  double_type,
+  long_double_type,
+  last_numeric_type = long_double_type,
+  cstring_type,
+  string_type,
+  pointer_type,
+  custom_type
+};
+
+// Maps core type T to the corresponding type enum constant.
+template <typename T, typename Char>
+struct type_constant : std::integral_constant<type, type::custom_type> {};
+
+#define FMT_TYPE_CONSTANT(Type, constant) \
+  template <typename Char>                \
+  struct type_constant<Type, Char>        \
+      : std::integral_constant<type, type::constant> {}
+
+FMT_TYPE_CONSTANT(int, int_type);
+FMT_TYPE_CONSTANT(unsigned, uint_type);
+FMT_TYPE_CONSTANT(long long, long_long_type);
+FMT_TYPE_CONSTANT(unsigned long long, ulong_long_type);
+FMT_TYPE_CONSTANT(int128_opt, int128_type);
+FMT_TYPE_CONSTANT(uint128_opt, uint128_type);
+FMT_TYPE_CONSTANT(bool, bool_type);
+FMT_TYPE_CONSTANT(Char, char_type);
+FMT_TYPE_CONSTANT(float, float_type);
+FMT_TYPE_CONSTANT(double, double_type);
+FMT_TYPE_CONSTANT(long double, long_double_type);
+FMT_TYPE_CONSTANT(const Char*, cstring_type);
+FMT_TYPE_CONSTANT(basic_string_view<Char>, string_type);
+FMT_TYPE_CONSTANT(const void*, pointer_type);
+
+constexpr auto is_integral_type(type t) -> bool {
+  return t > type::none_type && t <= type::last_integer_type;
+}
+constexpr auto is_arithmetic_type(type t) -> bool {
+  return t > type::none_type && t <= type::last_numeric_type;
+}
+
+constexpr auto set(type rhs) -> int { return 1 << static_cast<int>(rhs); }
+constexpr auto in(type t, int set) -> bool {
+  return ((set >> static_cast<int>(t)) & 1) != 0;
+}
+
+// Bitsets of types.
+enum {
+  sint_set =
+      set(type::int_type) | set(type::long_long_type) | set(type::int128_type),
+  uint_set = set(type::uint_type) | set(type::ulong_long_type) |
+             set(type::uint128_type),
+  bool_set = set(type::bool_type),
+  char_set = set(type::char_type),
+  float_set = set(type::float_type) | set(type::double_type) |
+              set(type::long_double_type),
+  string_set = set(type::string_type),
+  cstring_set = set(type::cstring_type),
+  pointer_set = set(type::pointer_type)
+};
+}  // namespace detail
+
+/// Reports a format error at compile time or, via a `format_error` exception,
+/// at runtime.
+// This function is intentionally not constexpr to give a compile-time error.
+FMT_NORETURN FMT_API void report_error(const char* message);
+
+FMT_DEPRECATED FMT_NORETURN inline void throw_format_error(
+    const char* message) {
+  report_error(message);
+}
+
+/// String's character (code unit) type.
+template <typename S,
+          typename V = decltype(detail::to_string_view(std::declval<S>()))>
+using char_t = typename V::value_type;
+
+/**
+ * Parsing context consisting of a format string range being parsed and an
+ * argument counter for automatic indexing.
+ * You can use the `format_parse_context` type alias for `char` instead.
+ */
+FMT_EXPORT
+template <typename Char> class basic_format_parse_context {
+ private:
+  basic_string_view<Char> format_str_;
+  int next_arg_id_;
+
+  FMT_CONSTEXPR void do_check_arg_id(int id);
+
+ public:
+  using char_type = Char;
+  using iterator = const Char*;
+
+  explicit constexpr basic_format_parse_context(
+      basic_string_view<Char> format_str, int next_arg_id = 0)
+      : format_str_(format_str), next_arg_id_(next_arg_id) {}
+
+  /// Returns an iterator to the beginning of the format string range being
+  /// parsed.
+  constexpr auto begin() const noexcept -> iterator {
+    return format_str_.begin();
+  }
+
+  /// Returns an iterator past the end of the format string range being parsed.
+  constexpr auto end() const noexcept -> iterator { return format_str_.end(); }
+
+  /// Advances the begin iterator to `it`.
+  FMT_CONSTEXPR void advance_to(iterator it) {
+    format_str_.remove_prefix(detail::to_unsigned(it - begin()));
+  }
+
+  /// Reports an error if using the manual argument indexing; otherwise returns
+  /// the next argument index and switches to the automatic indexing.
+  FMT_CONSTEXPR auto next_arg_id() -> int {
+    if (next_arg_id_ < 0) {
+      report_error("cannot switch from manual to automatic argument indexing");
+      return 0;
+    }
+    int id = next_arg_id_++;
+    do_check_arg_id(id);
+    return id;
+  }
+
+  /// Reports an error if using the automatic argument indexing; otherwise
+  /// switches to the manual indexing.
+  FMT_CONSTEXPR void check_arg_id(int id) {
+    if (next_arg_id_ > 0) {
+      report_error("cannot switch from automatic to manual argument indexing");
+      return;
+    }
+    next_arg_id_ = -1;
+    do_check_arg_id(id);
+  }
+  FMT_CONSTEXPR void check_arg_id(basic_string_view<Char>) {
+    next_arg_id_ = -1;
+  }
+  FMT_CONSTEXPR void check_dynamic_spec(int arg_id);
+};
+
+FMT_EXPORT
+using format_parse_context = basic_format_parse_context<char>;
+
+namespace detail {
+// A parse context with extra data used only in compile-time checks.
+template <typename Char>
+class compile_parse_context : public basic_format_parse_context<Char> {
+ private:
+  int num_args_;
+  const type* types_;
+  using base = basic_format_parse_context<Char>;
+
+ public:
+  explicit FMT_CONSTEXPR compile_parse_context(
+      basic_string_view<Char> format_str, int num_args, const type* types,
+      int next_arg_id = 0)
+      : base(format_str, next_arg_id), num_args_(num_args), types_(types) {}
+
+  constexpr auto num_args() const -> int { return num_args_; }
+  constexpr auto arg_type(int id) const -> type { return types_[id]; }
+
+  FMT_CONSTEXPR auto next_arg_id() -> int {
+    int id = base::next_arg_id();
+    if (id >= num_args_) report_error("argument not found");
+    return id;
+  }
+
+  FMT_CONSTEXPR void check_arg_id(int id) {
+    base::check_arg_id(id);
+    if (id >= num_args_) report_error("argument not found");
+  }
+  using base::check_arg_id;
+
+  FMT_CONSTEXPR void check_dynamic_spec(int arg_id) {
+    detail::ignore_unused(arg_id);
+    if (arg_id < num_args_ && types_ && !is_integral_type(types_[arg_id]))
+      report_error("width/precision is not integer");
+  }
+};
+
+/// A contiguous memory buffer with an optional growing ability. It is an
+/// internal class and shouldn't be used directly, only via `memory_buffer`.
+template <typename T> class buffer {
+ private:
+  T* ptr_;
+  size_t size_;
+  size_t capacity_;
+
+  using grow_fun = void (*)(buffer& buf, size_t capacity);
+  grow_fun grow_;
+
+ protected:
+  // Don't initialize ptr_ since it is not accessed to save a few cycles.
+  FMT_MSC_WARNING(suppress : 26495)
+  FMT_CONSTEXPR20 buffer(grow_fun grow, size_t sz) noexcept
+      : size_(sz), capacity_(sz), grow_(grow) {}
+
+  constexpr buffer(grow_fun grow, T* p = nullptr, size_t sz = 0,
+                   size_t cap = 0) noexcept
+      : ptr_(p), size_(sz), capacity_(cap), grow_(grow) {}
+
+  FMT_CONSTEXPR20 ~buffer() = default;
+  buffer(buffer&&) = default;
+
+  /// Sets the buffer data and capacity.
+  FMT_CONSTEXPR void set(T* buf_data, size_t buf_capacity) noexcept {
+    ptr_ = buf_data;
+    capacity_ = buf_capacity;
+  }
+
+ public:
+  using value_type = T;
+  using const_reference = const T&;
+
+  buffer(const buffer&) = delete;
+  void operator=(const buffer&) = delete;
+
+  auto begin() noexcept -> T* { return ptr_; }
+  auto end() noexcept -> T* { return ptr_ + size_; }
+
+  auto begin() const noexcept -> const T* { return ptr_; }
+  auto end() const noexcept -> const T* { return ptr_ + size_; }
+
+  /// Returns the size of this buffer.
+  constexpr auto size() const noexcept -> size_t { return size_; }
+
+  /// Returns the capacity of this buffer.
+  constexpr auto capacity() const noexcept -> size_t { return capacity_; }
+
+  /// Returns a pointer to the buffer data (not null-terminated).
+  FMT_CONSTEXPR auto data() noexcept -> T* { return ptr_; }
+  FMT_CONSTEXPR auto data() const noexcept -> const T* { return ptr_; }
+
+  /// Clears this buffer.
+  void clear() { size_ = 0; }
+
+  // Tries resizing the buffer to contain `count` elements. If T is a POD type
+  // the new elements may not be initialized.
+  FMT_CONSTEXPR void try_resize(size_t count) {
+    try_reserve(count);
+    size_ = count <= capacity_ ? count : capacity_;
+  }
+
+  // Tries increasing the buffer capacity to `new_capacity`. It can increase the
+  // capacity by a smaller amount than requested but guarantees there is space
+  // for at least one additional element either by increasing the capacity or by
+  // flushing the buffer if it is full.
+  FMT_CONSTEXPR void try_reserve(size_t new_capacity) {
+    if (new_capacity > capacity_) grow_(*this, new_capacity);
+  }
+
+  FMT_CONSTEXPR void push_back(const T& value) {
+    try_reserve(size_ + 1);
+    ptr_[size_++] = value;
+  }
+
+  /// Appends data to the end of the buffer.
+  template <typename U> void append(const U* begin, const U* end) {
+    while (begin != end) {
+      auto count = to_unsigned(end - begin);
+      try_reserve(size_ + count);
+      auto free_cap = capacity_ - size_;
+      if (free_cap < count) count = free_cap;
+      // A loop is faster than memcpy on small sizes.
+      T* out = ptr_ + size_;
+      for (size_t i = 0; i < count; ++i) out[i] = begin[i];
+      size_ += count;
+      begin += count;
+    }
+  }
+
+  template <typename Idx> FMT_CONSTEXPR auto operator[](Idx index) -> T& {
+    return ptr_[index];
+  }
+  template <typename Idx>
+  FMT_CONSTEXPR auto operator[](Idx index) const -> const T& {
+    return ptr_[index];
+  }
+};
+
+struct buffer_traits {
+  explicit buffer_traits(size_t) {}
+  auto count() const -> size_t { return 0; }
+  auto limit(size_t size) -> size_t { return size; }
+};
+
+class fixed_buffer_traits {
+ private:
+  size_t count_ = 0;
+  size_t limit_;
+
+ public:
+  explicit fixed_buffer_traits(size_t limit) : limit_(limit) {}
+  auto count() const -> size_t { return count_; }
+  auto limit(size_t size) -> size_t {
+    size_t n = limit_ > count_ ? limit_ - count_ : 0;
+    count_ += size;
+    return size < n ? size : n;
+  }
+};
+
+// A buffer that writes to an output iterator when flushed.
+template <typename OutputIt, typename T, typename Traits = buffer_traits>
+class iterator_buffer : public Traits, public buffer<T> {
+ private:
+  OutputIt out_;
+  enum { buffer_size = 256 };
+  T data_[buffer_size];
+
+  static FMT_CONSTEXPR void grow(buffer<T>& buf, size_t) {
+    if (buf.size() == buffer_size) static_cast<iterator_buffer&>(buf).flush();
+  }
+
+  void flush() {
+    auto size = this->size();
+    this->clear();
+    const T* begin = data_;
+    const T* end = begin + this->limit(size);
+    while (begin != end) *out_++ = *begin++;
+  }
+
+ public:
+  explicit iterator_buffer(OutputIt out, size_t n = buffer_size)
+      : Traits(n), buffer<T>(grow, data_, 0, buffer_size), out_(out) {}
+  iterator_buffer(iterator_buffer&& other) noexcept
+      : Traits(other),
+        buffer<T>(grow, data_, 0, buffer_size),
+        out_(other.out_) {}
+  ~iterator_buffer() {
+    // Don't crash if flush fails during unwinding.
+    FMT_TRY { flush(); }
+    FMT_CATCH(...) {}
+  }
+
+  auto out() -> OutputIt {
+    flush();
+    return out_;
+  }
+  auto count() const -> size_t { return Traits::count() + this->size(); }
+};
+
+template <typename T>
+class iterator_buffer<T*, T, fixed_buffer_traits> : public fixed_buffer_traits,
+                                                    public buffer<T> {
+ private:
+  T* out_;
+  enum { buffer_size = 256 };
+  T data_[buffer_size];
+
+  static FMT_CONSTEXPR void grow(buffer<T>& buf, size_t) {
+    if (buf.size() == buf.capacity())
+      static_cast<iterator_buffer&>(buf).flush();
+  }
+
+  void flush() {
+    size_t n = this->limit(this->size());
+    if (this->data() == out_) {
+      out_ += n;
+      this->set(data_, buffer_size);
+    }
+    this->clear();
+  }
+
+ public:
+  explicit iterator_buffer(T* out, size_t n = buffer_size)
+      : fixed_buffer_traits(n), buffer<T>(grow, out, 0, n), out_(out) {}
+  iterator_buffer(iterator_buffer&& other) noexcept
+      : fixed_buffer_traits(other),
+        buffer<T>(static_cast<iterator_buffer&&>(other)),
+        out_(other.out_) {
+    if (this->data() != out_) {
+      this->set(data_, buffer_size);
+      this->clear();
+    }
+  }
+  ~iterator_buffer() { flush(); }
+
+  auto out() -> T* {
+    flush();
+    return out_;
+  }
+  auto count() const -> size_t {
+    return fixed_buffer_traits::count() + this->size();
+  }
+};
+
+template <typename T> class iterator_buffer<T*, T> : public buffer<T> {
+ public:
+  explicit iterator_buffer(T* out, size_t = 0)
+      : buffer<T>([](buffer<T>&, size_t) {}, out, 0, ~size_t()) {}
+
+  auto out() -> T* { return &*this->end(); }
+};
+
+// A buffer that writes to a container with the contiguous storage.
+template <typename OutputIt>
+class iterator_buffer<
+    OutputIt,
+    enable_if_t<detail::is_back_insert_iterator<OutputIt>::value &&
+                    is_contiguous<typename OutputIt::container_type>::value,
+                typename OutputIt::container_type::value_type>>
+    : public buffer<typename OutputIt::container_type::value_type> {
+ private:
+  using container_type = typename OutputIt::container_type;
+  using value_type = typename container_type::value_type;
+  container_type& container_;
+
+  static FMT_CONSTEXPR void grow(buffer<value_type>& buf, size_t capacity) {
+    auto& self = static_cast<iterator_buffer&>(buf);
+    self.container_.resize(capacity);
+    self.set(&self.container_[0], capacity);
+  }
+
+ public:
+  explicit iterator_buffer(container_type& c)
+      : buffer<value_type>(grow, c.size()), container_(c) {}
+  explicit iterator_buffer(OutputIt out, size_t = 0)
+      : iterator_buffer(get_container(out)) {}
+
+  auto out() -> OutputIt { return back_inserter(container_); }
+};
+
+// A buffer that counts the number of code units written discarding the output.
+template <typename T = char> class counting_buffer : public buffer<T> {
+ private:
+  enum { buffer_size = 256 };
+  T data_[buffer_size];
+  size_t count_ = 0;
+
+  static FMT_CONSTEXPR void grow(buffer<T>& buf, size_t) {
+    if (buf.size() != buffer_size) return;
+    static_cast<counting_buffer&>(buf).count_ += buf.size();
+    buf.clear();
+  }
+
+ public:
+  counting_buffer() : buffer<T>(grow, data_, 0, buffer_size) {}
+
+  auto count() -> size_t { return count_ + this->size(); }
+};
+}  // namespace detail
+
+template <typename Char>
+FMT_CONSTEXPR void basic_format_parse_context<Char>::do_check_arg_id(int id) {
+  // Argument id is only checked at compile-time during parsing because
+  // formatting has its own validation.
+  if (detail::is_constant_evaluated() &&
+      (!FMT_GCC_VERSION || FMT_GCC_VERSION >= 1200)) {
+    using context = detail::compile_parse_context<Char>;
+    if (id >= static_cast<context*>(this)->num_args())
+      report_error("argument not found");
+  }
+}
+
+template <typename Char>
+FMT_CONSTEXPR void basic_format_parse_context<Char>::check_dynamic_spec(
+    int arg_id) {
+  if (detail::is_constant_evaluated() &&
+      (!FMT_GCC_VERSION || FMT_GCC_VERSION >= 1200)) {
+    using context = detail::compile_parse_context<Char>;
+    static_cast<context*>(this)->check_dynamic_spec(arg_id);
+  }
+}
+
+FMT_EXPORT template <typename Context> class basic_format_arg;
+FMT_EXPORT template <typename Context> class basic_format_args;
+FMT_EXPORT template <typename Context> class dynamic_format_arg_store;
+
+// A formatter for objects of type T.
+FMT_EXPORT
+template <typename T, typename Char = char, typename Enable = void>
+struct formatter {
+  // A deleted default constructor indicates a disabled formatter.
+  formatter() = delete;
+};
+
+// Specifies if T has an enabled formatter specialization. A type can be
+// formattable even if it doesn't have a formatter e.g. via a conversion.
+template <typename T, typename Context>
+using has_formatter =
+    std::is_constructible<typename Context::template formatter_type<T>>;
+
+// An output iterator that appends to a buffer. It is used instead of
+// back_insert_iterator to reduce symbol sizes and avoid <iterator> dependency.
+template <typename T> class basic_appender {
+ private:
+  detail::buffer<T>* buffer_;
+
+  friend auto get_container(basic_appender app) -> detail::buffer<T>& {
+    return *app.buffer_;
+  }
+
+ public:
+  using iterator_category = int;
+  using value_type = T;
+  using difference_type = ptrdiff_t;
+  using pointer = T*;
+  using reference = T&;
+  using container_type = detail::buffer<T>;
+  FMT_UNCHECKED_ITERATOR(basic_appender);
+
+  FMT_CONSTEXPR basic_appender(detail::buffer<T>& buf) : buffer_(&buf) {}
+
+  auto operator=(T c) -> basic_appender& {
+    buffer_->push_back(c);
+    return *this;
+  }
+  auto operator*() -> basic_appender& { return *this; }
+  auto operator++() -> basic_appender& { return *this; }
+  auto operator++(int) -> basic_appender { return *this; }
+};
+
+using appender = basic_appender<char>;
+
+namespace detail {
+template <typename T>
+struct is_back_insert_iterator<basic_appender<T>> : std::true_type {};
+
+template <typename T, typename Enable = void>
+struct locking : std::true_type {};
+template <typename T>
+struct locking<T, void_t<typename formatter<remove_cvref_t<T>>::nonlocking>>
+    : std::false_type {};
+
+template <typename T = int> FMT_CONSTEXPR inline auto is_locking() -> bool {
+  return locking<T>::value;
+}
+template <typename T1, typename T2, typename... Tail>
+FMT_CONSTEXPR inline auto is_locking() -> bool {
+  return locking<T1>::value || is_locking<T2, Tail...>();
+}
+
+// An optimized version of std::copy with the output value type (T).
+template <typename T, typename InputIt, typename OutputIt,
+          FMT_ENABLE_IF(is_back_insert_iterator<OutputIt>::value)>
+auto copy(InputIt begin, InputIt end, OutputIt out) -> OutputIt {
+  get_container(out).append(begin, end);
+  return out;
+}
+
+template <typename T, typename InputIt, typename OutputIt,
+          FMT_ENABLE_IF(!is_back_insert_iterator<OutputIt>::value)>
+FMT_CONSTEXPR auto copy(InputIt begin, InputIt end, OutputIt out) -> OutputIt {
+  while (begin != end) *out++ = static_cast<T>(*begin++);
+  return out;
+}
+
+template <typename T, typename V, typename OutputIt>
+FMT_CONSTEXPR auto copy(basic_string_view<V> s, OutputIt out) -> OutputIt {
+  return copy<T>(s.begin(), s.end(), out);
+}
+
+template <typename Context, typename T>
+constexpr auto has_const_formatter_impl(T*)
+    -> decltype(typename Context::template formatter_type<T>().format(
+                    std::declval<const T&>(), std::declval<Context&>()),
+                true) {
+  return true;
+}
+template <typename Context>
+constexpr auto has_const_formatter_impl(...) -> bool {
+  return false;
+}
+template <typename T, typename Context>
+constexpr auto has_const_formatter() -> bool {
+  return has_const_formatter_impl<Context>(static_cast<T*>(nullptr));
+}
+
+template <typename It, typename Enable = std::true_type>
+struct is_buffer_appender : std::false_type {};
+template <typename It>
+struct is_buffer_appender<
+    It, bool_constant<
+            is_back_insert_iterator<It>::value &&
+            std::is_base_of<buffer<typename It::container_type::value_type>,
+                            typename It::container_type>::value>>
+    : std::true_type {};
+
+// Maps an output iterator to a buffer.
+template <typename T, typename OutputIt,
+          FMT_ENABLE_IF(!is_buffer_appender<OutputIt>::value)>
+auto get_buffer(OutputIt out) -> iterator_buffer<OutputIt, T> {
+  return iterator_buffer<OutputIt, T>(out);
+}
+template <typename T, typename OutputIt,
+          FMT_ENABLE_IF(is_buffer_appender<OutputIt>::value)>
+auto get_buffer(OutputIt out) -> buffer<T>& {
+  return get_container(out);
+}
+
+template <typename Buf, typename OutputIt>
+auto get_iterator(Buf& buf, OutputIt) -> decltype(buf.out()) {
+  return buf.out();
+}
+template <typename T, typename OutputIt>
+auto get_iterator(buffer<T>&, OutputIt out) -> OutputIt {
+  return out;
+}
+
+struct view {};
+
+template <typename Char, typename T> struct named_arg : view {
+  const Char* name;
+  const T& value;
+  named_arg(const Char* n, const T& v) : name(n), value(v) {}
+};
+
+template <typename Char> struct named_arg_info {
+  const Char* name;
+  int id;
+};
+
+template <typename T> struct is_named_arg : std::false_type {};
+template <typename T> struct is_statically_named_arg : std::false_type {};
+
+template <typename T, typename Char>
+struct is_named_arg<named_arg<Char, T>> : std::true_type {};
+
+template <bool B = false> constexpr auto count() -> size_t { return B ? 1 : 0; }
+template <bool B1, bool B2, bool... Tail> constexpr auto count() -> size_t {
+  return (B1 ? 1 : 0) + count<B2, Tail...>();
+}
+
+template <typename... Args> constexpr auto count_named_args() -> size_t {
+  return count<is_named_arg<Args>::value...>();
+}
+
+template <typename... Args>
+constexpr auto count_statically_named_args() -> size_t {
+  return count<is_statically_named_arg<Args>::value...>();
+}
+
+struct unformattable {};
+struct unformattable_char : unformattable {};
+struct unformattable_pointer : unformattable {};
+
+template <typename Char> struct string_value {
+  const Char* data;
+  size_t size;
+};
+
+template <typename Char> struct named_arg_value {
+  const named_arg_info<Char>* data;
+  size_t size;
+};
+
+template <typename Context> struct custom_value {
+  using parse_context = typename Context::parse_context_type;
+  void* value;
+  void (*format)(void* arg, parse_context& parse_ctx, Context& ctx);
+};
+
+// A formatting argument value.
+template <typename Context> class value {
+ public:
+  using char_type = typename Context::char_type;
+
+  union {
+    monostate no_value;
+    int int_value;
+    unsigned uint_value;
+    long long long_long_value;
+    unsigned long long ulong_long_value;
+    int128_opt int128_value;
+    uint128_opt uint128_value;
+    bool bool_value;
+    char_type char_value;
+    float float_value;
+    double double_value;
+    long double long_double_value;
+    const void* pointer;
+    string_value<char_type> string;
+    custom_value<Context> custom;
+    named_arg_value<char_type> named_args;
+  };
+
+  constexpr FMT_ALWAYS_INLINE value() : no_value() {}
+  constexpr FMT_ALWAYS_INLINE value(int val) : int_value(val) {}
+  constexpr FMT_ALWAYS_INLINE value(unsigned val) : uint_value(val) {}
+  constexpr FMT_ALWAYS_INLINE value(long long val) : long_long_value(val) {}
+  constexpr FMT_ALWAYS_INLINE value(unsigned long long val)
+      : ulong_long_value(val) {}
+  FMT_ALWAYS_INLINE value(int128_opt val) : int128_value(val) {}
+  FMT_ALWAYS_INLINE value(uint128_opt val) : uint128_value(val) {}
+  constexpr FMT_ALWAYS_INLINE value(float val) : float_value(val) {}
+  constexpr FMT_ALWAYS_INLINE value(double val) : double_value(val) {}
+  FMT_ALWAYS_INLINE value(long double val) : long_double_value(val) {}
+  constexpr FMT_ALWAYS_INLINE value(bool val) : bool_value(val) {}
+  constexpr FMT_ALWAYS_INLINE value(char_type val) : char_value(val) {}
+  FMT_CONSTEXPR FMT_ALWAYS_INLINE value(const char_type* val) {
+    string.data = val;
+    if (is_constant_evaluated()) string.size = {};
+  }
+  FMT_CONSTEXPR FMT_ALWAYS_INLINE value(basic_string_view<char_type> val) {
+    string.data = val.data();
+    string.size = val.size();
+  }
+  FMT_ALWAYS_INLINE value(const void* val) : pointer(val) {}
+  FMT_ALWAYS_INLINE value(const named_arg_info<char_type>* args, size_t size)
+      : named_args{args, size} {}
+
+  template <typename T> FMT_CONSTEXPR20 FMT_ALWAYS_INLINE value(T& val) {
+    using value_type = remove_const_t<T>;
+    // T may overload operator& e.g. std::vector<bool>::reference in libc++.
+#if defined(__cpp_if_constexpr)
+    if constexpr (std::is_same<decltype(&val), T*>::value)
+      custom.value = const_cast<value_type*>(&val);
+#endif
+    if (!is_constant_evaluated())
+      custom.value = const_cast<char*>(&reinterpret_cast<const char&>(val));
+    // Get the formatter type through the context to allow different contexts
+    // have different extension points, e.g. `formatter<T>` for `format` and
+    // `printf_formatter<T>` for `printf`.
+    custom.format = format_custom_arg<
+        value_type, typename Context::template formatter_type<value_type>>;
+  }
+  value(unformattable);
+  value(unformattable_char);
+  value(unformattable_pointer);
+
+ private:
+  // Formats an argument of a custom type, such as a user-defined class.
+  template <typename T, typename Formatter>
+  static void format_custom_arg(void* arg,
+                                typename Context::parse_context_type& parse_ctx,
+                                Context& ctx) {
+    auto f = Formatter();
+    parse_ctx.advance_to(f.parse(parse_ctx));
+    using qualified_type =
+        conditional_t<has_const_formatter<T, Context>(), const T, T>;
+    // format must be const for compatibility with std::format and compilation.
+    const auto& cf = f;
+    ctx.advance_to(cf.format(*static_cast<qualified_type*>(arg), ctx));
+  }
+};
+
+// To minimize the number of types we need to deal with, long is translated
+// either to int or to long long depending on its size.
+enum { long_short = sizeof(long) == sizeof(int) };
+using long_type = conditional_t<long_short, int, long long>;
+using ulong_type = conditional_t<long_short, unsigned, unsigned long long>;
+
+template <typename T> struct format_as_result {
+  template <typename U,
+            FMT_ENABLE_IF(std::is_enum<U>::value || std::is_class<U>::value)>
+  static auto map(U*) -> remove_cvref_t<decltype(format_as(std::declval<U>()))>;
+  static auto map(...) -> void;
+
+  using type = decltype(map(static_cast<T*>(nullptr)));
+};
+template <typename T> using format_as_t = typename format_as_result<T>::type;
+
+template <typename T>
+struct has_format_as
+    : bool_constant<!std::is_same<format_as_t<T>, void>::value> {};
+
+#define FMT_MAP_API FMT_CONSTEXPR FMT_ALWAYS_INLINE
+
+// Maps formatting arguments to core types.
+// arg_mapper reports errors by returning unformattable instead of using
+// static_assert because it's used in the is_formattable trait.
+template <typename Context> struct arg_mapper {
+  using char_type = typename Context::char_type;
+
+  FMT_MAP_API auto map(signed char val) -> int { return val; }
+  FMT_MAP_API auto map(unsigned char val) -> unsigned { return val; }
+  FMT_MAP_API auto map(short val) -> int { return val; }
+  FMT_MAP_API auto map(unsigned short val) -> unsigned { return val; }
+  FMT_MAP_API auto map(int val) -> int { return val; }
+  FMT_MAP_API auto map(unsigned val) -> unsigned { return val; }
+  FMT_MAP_API auto map(long val) -> long_type { return val; }
+  FMT_MAP_API auto map(unsigned long val) -> ulong_type { return val; }
+  FMT_MAP_API auto map(long long val) -> long long { return val; }
+  FMT_MAP_API auto map(unsigned long long val) -> unsigned long long {
+    return val;
+  }
+  FMT_MAP_API auto map(int128_opt val) -> int128_opt { return val; }
+  FMT_MAP_API auto map(uint128_opt val) -> uint128_opt { return val; }
+  FMT_MAP_API auto map(bool val) -> bool { return val; }
+
+  template <typename T, FMT_ENABLE_IF(std::is_same<T, char>::value ||
+                                      std::is_same<T, char_type>::value)>
+  FMT_MAP_API auto map(T val) -> char_type {
+    return val;
+  }
+  template <typename T, enable_if_t<(std::is_same<T, wchar_t>::value ||
+#ifdef __cpp_char8_t
+                                     std::is_same<T, char8_t>::value ||
+#endif
+                                     std::is_same<T, char16_t>::value ||
+                                     std::is_same<T, char32_t>::value) &&
+                                        !std::is_same<T, char_type>::value,
+                                    int> = 0>
+  FMT_MAP_API auto map(T) -> unformattable_char {
+    return {};
+  }
+
+  FMT_MAP_API auto map(float val) -> float { return val; }
+  FMT_MAP_API auto map(double val) -> double { return val; }
+  FMT_MAP_API auto map(long double val) -> long double { return val; }
+
+  FMT_MAP_API auto map(char_type* val) -> const char_type* { return val; }
+  FMT_MAP_API auto map(const char_type* val) -> const char_type* { return val; }
+  template <typename T, typename Char = char_t<T>,
+            FMT_ENABLE_IF(std::is_same<Char, char_type>::value &&
+                          !std::is_pointer<T>::value)>
+  FMT_MAP_API auto map(const T& val) -> basic_string_view<Char> {
+    return to_string_view(val);
+  }
+  template <typename T, typename Char = char_t<T>,
+            FMT_ENABLE_IF(!std::is_same<Char, char_type>::value &&
+                          !std::is_pointer<T>::value)>
+  FMT_MAP_API auto map(const T&) -> unformattable_char {
+    return {};
+  }
+
+  FMT_MAP_API auto map(void* val) -> const void* { return val; }
+  FMT_MAP_API auto map(const void* val) -> const void* { return val; }
+  FMT_MAP_API auto map(volatile void* val) -> const void* {
+    return const_cast<const void*>(val);
+  }
+  FMT_MAP_API auto map(const volatile void* val) -> const void* {
+    return const_cast<const void*>(val);
+  }
+  FMT_MAP_API auto map(std::nullptr_t val) -> const void* { return val; }
+
+  // Use SFINAE instead of a const T* parameter to avoid a conflict with the
+  // array overload.
+  template <
+      typename T,
+      FMT_ENABLE_IF(
+          std::is_pointer<T>::value || std::is_member_pointer<T>::value ||
+          std::is_function<typename std::remove_pointer<T>::type>::value ||
+          (std::is_array<T>::value &&
+           !std::is_convertible<T, const char_type*>::value))>
+  FMT_CONSTEXPR auto map(const T&) -> unformattable_pointer {
+    return {};
+  }
+
+  template <typename T, std::size_t N,
+            FMT_ENABLE_IF(!std::is_same<T, wchar_t>::value)>
+  FMT_MAP_API auto map(const T (&values)[N]) -> const T (&)[N] {
+    return values;
+  }
+
+  // Only map owning types because mapping views can be unsafe.
+  template <typename T, typename U = format_as_t<T>,
+            FMT_ENABLE_IF(std::is_arithmetic<U>::value)>
+  FMT_MAP_API auto map(const T& val) -> decltype(FMT_DECLTYPE_THIS map(U())) {
+    return map(format_as(val));
+  }
+
+  template <typename T, typename U = remove_const_t<T>>
+  struct formattable : bool_constant<has_const_formatter<U, Context>() ||
+                                     (has_formatter<U, Context>::value &&
+                                      !std::is_const<T>::value)> {};
+
+  template <typename T, FMT_ENABLE_IF(formattable<T>::value)>
+  FMT_MAP_API auto do_map(T& val) -> T& {
+    return val;
+  }
+  template <typename T, FMT_ENABLE_IF(!formattable<T>::value)>
+  FMT_MAP_API auto do_map(T&) -> unformattable {
+    return {};
+  }
+
+  // is_fundamental is used to allow formatters for extended FP types.
+  template <typename T, typename U = remove_const_t<T>,
+            FMT_ENABLE_IF(
+                (std::is_class<U>::value || std::is_enum<U>::value ||
+                 std::is_union<U>::value || std::is_fundamental<U>::value) &&
+                !has_to_string_view<U>::value && !is_char<U>::value &&
+                !is_named_arg<U>::value && !std::is_integral<U>::value &&
+                !std::is_arithmetic<format_as_t<U>>::value)>
+  FMT_MAP_API auto map(T& val) -> decltype(FMT_DECLTYPE_THIS do_map(val)) {
+    return do_map(val);
+  }
+
+  template <typename T, FMT_ENABLE_IF(is_named_arg<T>::value)>
+  FMT_MAP_API auto map(const T& named_arg)
+      -> decltype(FMT_DECLTYPE_THIS map(named_arg.value)) {
+    return map(named_arg.value);
+  }
+
+  auto map(...) -> unformattable { return {}; }
+};
+
+// A type constant after applying arg_mapper<Context>.
+template <typename T, typename Context>
+using mapped_type_constant =
+    type_constant<decltype(arg_mapper<Context>().map(std::declval<const T&>())),
+                  typename Context::char_type>;
+
+enum { packed_arg_bits = 4 };
+// Maximum number of arguments with packed types.
+enum { max_packed_args = 62 / packed_arg_bits };
+enum : unsigned long long { is_unpacked_bit = 1ULL << 63 };
+enum : unsigned long long { has_named_args_bit = 1ULL << 62 };
+
+template <typename It, typename T, typename Enable = void>
+struct is_output_iterator : std::false_type {};
+
+template <> struct is_output_iterator<appender, char> : std::true_type {};
+
+template <typename It, typename T>
+struct is_output_iterator<
+    It, T, void_t<decltype(*std::declval<It&>()++ = std::declval<T>())>>
+    : std::true_type {};
+
+// A type-erased reference to an std::locale to avoid a heavy <locale> include.
+class locale_ref {
+ private:
+  const void* locale_;  // A type-erased pointer to std::locale.
+
+ public:
+  constexpr locale_ref() : locale_(nullptr) {}
+  template <typename Locale> explicit locale_ref(const Locale& loc);
+
+  explicit operator bool() const noexcept { return locale_ != nullptr; }
+
+  template <typename Locale> auto get() const -> Locale;
+};
+
+template <typename> constexpr auto encode_types() -> unsigned long long {
+  return 0;
+}
+
+template <typename Context, typename Arg, typename... Args>
+constexpr auto encode_types() -> unsigned long long {
+  return static_cast<unsigned>(mapped_type_constant<Arg, Context>::value) |
+         (encode_types<Context, Args...>() << packed_arg_bits);
+}
+
+template <typename Context, typename... T, size_t NUM_ARGS = sizeof...(T)>
+constexpr unsigned long long make_descriptor() {
+  return NUM_ARGS <= max_packed_args ? encode_types<Context, T...>()
+                                     : is_unpacked_bit | NUM_ARGS;
+}
+
+// This type is intentionally undefined, only used for errors.
+template <typename T, typename Char>
+#if FMT_CLANG_VERSION && FMT_CLANG_VERSION <= 1500
+// https://github.com/fmtlib/fmt/issues/3796
+struct type_is_unformattable_for {
+};
+#else
+struct type_is_unformattable_for;
+#endif
+
+template <bool PACKED, typename Context, typename T, FMT_ENABLE_IF(PACKED)>
+FMT_CONSTEXPR auto make_arg(T& val) -> value<Context> {
+  using arg_type = remove_cvref_t<decltype(arg_mapper<Context>().map(val))>;
+
+  // Use enum instead of constexpr because the latter may generate code.
+  enum {
+    formattable_char = !std::is_same<arg_type, unformattable_char>::value
+  };
+  static_assert(formattable_char, "Mixing character types is disallowed.");
+
+  // Formatting of arbitrary pointers is disallowed. If you want to format a
+  // pointer cast it to `void*` or `const void*`. In particular, this forbids
+  // formatting of `[const] volatile char*` printed as bool by iostreams.
+  enum {
+    formattable_pointer = !std::is_same<arg_type, unformattable_pointer>::value
+  };
+  static_assert(formattable_pointer,
+                "Formatting of non-void pointers is disallowed.");
+
+  enum { formattable = !std::is_same<arg_type, unformattable>::value };
+#if defined(__cpp_if_constexpr)
+  if constexpr (!formattable)
+    type_is_unformattable_for<T, typename Context::char_type> _;
+#endif
+  static_assert(
+      formattable,
+      "Cannot format an argument. To make type T formattable provide a "
+      "formatter<T> specialization: https://fmt.dev/latest/api.html#udt");
+  return {arg_mapper<Context>().map(val)};
+}
+
+template <typename Context, typename T>
+FMT_CONSTEXPR auto make_arg(T& val) -> basic_format_arg<Context> {
+  auto arg = basic_format_arg<Context>();
+  arg.type_ = mapped_type_constant<T, Context>::value;
+  arg.value_ = make_arg<true, Context>(val);
+  return arg;
+}
+
+template <bool PACKED, typename Context, typename T, FMT_ENABLE_IF(!PACKED)>
+FMT_CONSTEXPR inline auto make_arg(T& val) -> basic_format_arg<Context> {
+  return make_arg<Context>(val);
+}
+
+template <typename Context, size_t NUM_ARGS>
+using arg_t = conditional_t<NUM_ARGS <= max_packed_args, value<Context>,
+                            basic_format_arg<Context>>;
+
+template <typename Char, typename T, FMT_ENABLE_IF(!is_named_arg<T>::value)>
+void init_named_arg(named_arg_info<Char>*, int& arg_index, int&, const T&) {
+  ++arg_index;
+}
+template <typename Char, typename T, FMT_ENABLE_IF(is_named_arg<T>::value)>
+void init_named_arg(named_arg_info<Char>* named_args, int& arg_index,
+                    int& named_arg_index, const T& arg) {
+  named_args[named_arg_index++] = {arg.name, arg_index++};
+}
+
+// An array of references to arguments. It can be implicitly converted to
+// `fmt::basic_format_args` for passing into type-erased formatting functions
+// such as `fmt::vformat`.
+template <typename Context, size_t NUM_ARGS, size_t NUM_NAMED_ARGS,
+          unsigned long long DESC>
+struct format_arg_store {
+  // args_[0].named_args points to named_args to avoid bloating format_args.
+  // +1 to workaround a bug in gcc 7.5 that causes duplicated-branches warning.
+  static constexpr size_t ARGS_ARR_SIZE = 1 + (NUM_ARGS != 0 ? NUM_ARGS : +1);
+
+  arg_t<Context, NUM_ARGS> args[ARGS_ARR_SIZE];
+  named_arg_info<typename Context::char_type> named_args[NUM_NAMED_ARGS];
+
+  template <typename... T>
+  FMT_MAP_API format_arg_store(T&... values)
+      : args{{named_args, NUM_NAMED_ARGS},
+             make_arg<NUM_ARGS <= max_packed_args, Context>(values)...} {
+    using dummy = int[];
+    int arg_index = 0, named_arg_index = 0;
+    (void)dummy{
+        0,
+        (init_named_arg(named_args, arg_index, named_arg_index, values), 0)...};
+  }
+
+  format_arg_store(format_arg_store&& rhs) {
+    args[0] = {named_args, NUM_NAMED_ARGS};
+    for (size_t i = 1; i < ARGS_ARR_SIZE; ++i) args[i] = rhs.args[i];
+    for (size_t i = 0; i < NUM_NAMED_ARGS; ++i)
+      named_args[i] = rhs.named_args[i];
+  }
+
+  format_arg_store(const format_arg_store& rhs) = delete;
+  format_arg_store& operator=(const format_arg_store& rhs) = delete;
+  format_arg_store& operator=(format_arg_store&& rhs) = delete;
+};
+
+// A specialization of format_arg_store without named arguments.
+// It is a plain struct to reduce binary size in debug mode.
+template <typename Context, size_t NUM_ARGS, unsigned long long DESC>
+struct format_arg_store<Context, NUM_ARGS, 0, DESC> {
+  // +1 to workaround a bug in gcc 7.5 that causes duplicated-branches warning.
+  arg_t<Context, NUM_ARGS> args[NUM_ARGS != 0 ? NUM_ARGS : +1];
+};
+
+}  // namespace detail
+FMT_BEGIN_EXPORT
+
+// A formatting argument. Context is a template parameter for the compiled API
+// where output can be unbuffered.
+template <typename Context> class basic_format_arg {
+ private:
+  detail::value<Context> value_;
+  detail::type type_;
+
+  template <typename ContextType, typename T>
+  friend FMT_CONSTEXPR auto detail::make_arg(T& value)
+      -> basic_format_arg<ContextType>;
+
+  friend class basic_format_args<Context>;
+  friend class dynamic_format_arg_store<Context>;
+
+  using char_type = typename Context::char_type;
+
+  template <typename, size_t, size_t, unsigned long long>
+  friend struct detail::format_arg_store;
+
+  basic_format_arg(const detail::named_arg_info<char_type>* args, size_t size)
+      : value_(args, size) {}
+
+ public:
+  class handle {
+   public:
+    explicit handle(detail::custom_value<Context> custom) : custom_(custom) {}
+
+    void format(typename Context::parse_context_type& parse_ctx,
+                Context& ctx) const {
+      custom_.format(custom_.value, parse_ctx, ctx);
+    }
+
+   private:
+    detail::custom_value<Context> custom_;
+  };
+
+  constexpr basic_format_arg() : type_(detail::type::none_type) {}
+
+  constexpr explicit operator bool() const noexcept {
+    return type_ != detail::type::none_type;
+  }
+
+  auto type() const -> detail::type { return type_; }
+
+  auto is_integral() const -> bool { return detail::is_integral_type(type_); }
+  auto is_arithmetic() const -> bool {
+    return detail::is_arithmetic_type(type_);
+  }
+
+  /**
+   * Visits an argument dispatching to the appropriate visit method based on
+   * the argument type. For example, if the argument type is `double` then
+   * `vis(value)` will be called with the value of type `double`.
+   */
+  template <typename Visitor>
+  FMT_CONSTEXPR FMT_INLINE auto visit(Visitor&& vis) const -> decltype(vis(0)) {
+    switch (type_) {
+    case detail::type::none_type:
+      break;
+    case detail::type::int_type:
+      return vis(value_.int_value);
+    case detail::type::uint_type:
+      return vis(value_.uint_value);
+    case detail::type::long_long_type:
+      return vis(value_.long_long_value);
+    case detail::type::ulong_long_type:
+      return vis(value_.ulong_long_value);
+    case detail::type::int128_type:
+      return vis(detail::convert_for_visit(value_.int128_value));
+    case detail::type::uint128_type:
+      return vis(detail::convert_for_visit(value_.uint128_value));
+    case detail::type::bool_type:
+      return vis(value_.bool_value);
+    case detail::type::char_type:
+      return vis(value_.char_value);
+    case detail::type::float_type:
+      return vis(value_.float_value);
+    case detail::type::double_type:
+      return vis(value_.double_value);
+    case detail::type::long_double_type:
+      return vis(value_.long_double_value);
+    case detail::type::cstring_type:
+      return vis(value_.string.data);
+    case detail::type::string_type:
+      using sv = basic_string_view<typename Context::char_type>;
+      return vis(sv(value_.string.data, value_.string.size));
+    case detail::type::pointer_type:
+      return vis(value_.pointer);
+    case detail::type::custom_type:
+      return vis(typename basic_format_arg<Context>::handle(value_.custom));
+    }
+    return vis(monostate());
+  }
+
+  auto format_custom(const char_type* parse_begin,
+                     typename Context::parse_context_type& parse_ctx,
+                     Context& ctx) -> bool {
+    if (type_ != detail::type::custom_type) return false;
+    parse_ctx.advance_to(parse_begin);
+    value_.custom.format(value_.custom.value, parse_ctx, ctx);
+    return true;
+  }
+};
+
+template <typename Visitor, typename Context>
+FMT_DEPRECATED FMT_CONSTEXPR auto visit_format_arg(
+    Visitor&& vis, const basic_format_arg<Context>& arg) -> decltype(vis(0)) {
+  return arg.visit(static_cast<Visitor&&>(vis));
+}
+
+/**
+ * A view of a collection of formatting arguments. To avoid lifetime issues it
+ * should only be used as a parameter type in type-erased functions such as
+ * `vformat`:
+ *
+ *     void vlog(fmt::string_view fmt, fmt::format_args args);  // OK
+ *     fmt::format_args args = fmt::make_format_args();  // Dangling reference
+ */
+template <typename Context> class basic_format_args {
+ public:
+  using size_type = int;
+  using format_arg = basic_format_arg<Context>;
+
+ private:
+  // A descriptor that contains information about formatting arguments.
+  // If the number of arguments is less or equal to max_packed_args then
+  // argument types are passed in the descriptor. This reduces binary code size
+  // per formatting function call.
+  unsigned long long desc_;
+  union {
+    // If is_packed() returns true then argument values are stored in values_;
+    // otherwise they are stored in args_. This is done to improve cache
+    // locality and reduce compiled code size since storing larger objects
+    // may require more code (at least on x86-64) even if the same amount of
+    // data is actually copied to stack. It saves ~10% on the bloat test.
+    const detail::value<Context>* values_;
+    const format_arg* args_;
+  };
+
+  constexpr auto is_packed() const -> bool {
+    return (desc_ & detail::is_unpacked_bit) == 0;
+  }
+  constexpr auto has_named_args() const -> bool {
+    return (desc_ & detail::has_named_args_bit) != 0;
+  }
+
+  FMT_CONSTEXPR auto type(int index) const -> detail::type {
+    int shift = index * detail::packed_arg_bits;
+    unsigned int mask = (1 << detail::packed_arg_bits) - 1;
+    return static_cast<detail::type>((desc_ >> shift) & mask);
+  }
+
+ public:
+  constexpr basic_format_args() : desc_(0), args_(nullptr) {}
+
+  /// Constructs a `basic_format_args` object from `format_arg_store`.
+  template <size_t NUM_ARGS, size_t NUM_NAMED_ARGS, unsigned long long DESC,
+            FMT_ENABLE_IF(NUM_ARGS <= detail::max_packed_args)>
+  constexpr FMT_ALWAYS_INLINE basic_format_args(
+      const detail::format_arg_store<Context, NUM_ARGS, NUM_NAMED_ARGS, DESC>&
+          store)
+      : desc_(DESC), values_(store.args + (NUM_NAMED_ARGS != 0 ? 1 : 0)) {}
+
+  template <size_t NUM_ARGS, size_t NUM_NAMED_ARGS, unsigned long long DESC,
+            FMT_ENABLE_IF(NUM_ARGS > detail::max_packed_args)>
+  constexpr basic_format_args(
+      const detail::format_arg_store<Context, NUM_ARGS, NUM_NAMED_ARGS, DESC>&
+          store)
+      : desc_(DESC), args_(store.args + (NUM_NAMED_ARGS != 0 ? 1 : 0)) {}
+
+  /// Constructs a `basic_format_args` object from `dynamic_format_arg_store`.
+  constexpr basic_format_args(const dynamic_format_arg_store<Context>& store)
+      : desc_(store.get_types()), args_(store.data()) {}
+
+  /// Constructs a `basic_format_args` object from a dynamic list of arguments.
+  constexpr basic_format_args(const format_arg* args, int count)
+      : desc_(detail::is_unpacked_bit | detail::to_unsigned(count)),
+        args_(args) {}
+
+  /// Returns the argument with the specified id.
+  FMT_CONSTEXPR auto get(int id) const -> format_arg {
+    format_arg arg;
+    if (!is_packed()) {
+      if (id < max_size()) arg = args_[id];
+      return arg;
+    }
+    if (static_cast<unsigned>(id) >= detail::max_packed_args) return arg;
+    arg.type_ = type(id);
+    if (arg.type_ == detail::type::none_type) return arg;
+    arg.value_ = values_[id];
+    return arg;
+  }
+
+  template <typename Char>
+  auto get(basic_string_view<Char> name) const -> format_arg {
+    int id = get_id(name);
+    return id >= 0 ? get(id) : format_arg();
+  }
+
+  template <typename Char>
+  FMT_CONSTEXPR auto get_id(basic_string_view<Char> name) const -> int {
+    if (!has_named_args()) return -1;
+    const auto& named_args =
+        (is_packed() ? values_[-1] : args_[-1].value_).named_args;
+    for (size_t i = 0; i < named_args.size; ++i) {
+      if (named_args.data[i].name == name) return named_args.data[i].id;
+    }
+    return -1;
+  }
+
+  auto max_size() const -> int {
+    unsigned long long max_packed = detail::max_packed_args;
+    return static_cast<int>(is_packed() ? max_packed
+                                        : desc_ & ~detail::is_unpacked_bit);
+  }
+};
+
+// A formatting context.
+class context {
+ private:
+  appender out_;
+  basic_format_args<context> args_;
+  detail::locale_ref loc_;
+
+ public:
+  /// The character type for the output.
+  using char_type = char;
+
+  using iterator = appender;
+  using format_arg = basic_format_arg<context>;
+  using parse_context_type = basic_format_parse_context<char>;
+  template <typename T> using formatter_type = formatter<T, char>;
+
+  /// Constructs a `basic_format_context` object. References to the arguments
+  /// are stored in the object so make sure they have appropriate lifetimes.
+  FMT_CONSTEXPR context(iterator out, basic_format_args<context> ctx_args,
+                        detail::locale_ref loc = {})
+      : out_(out), args_(ctx_args), loc_(loc) {}
+  context(context&&) = default;
+  context(const context&) = delete;
+  void operator=(const context&) = delete;
+
+  FMT_CONSTEXPR auto arg(int id) const -> format_arg { return args_.get(id); }
+  auto arg(string_view name) -> format_arg { return args_.get(name); }
+  FMT_CONSTEXPR auto arg_id(string_view name) -> int {
+    return args_.get_id(name);
+  }
+  auto args() const -> const basic_format_args<context>& { return args_; }
+
+  // Returns an iterator to the beginning of the output range.
+  FMT_CONSTEXPR auto out() -> iterator { return out_; }
+
+  // Advances the begin iterator to `it`.
+  void advance_to(iterator) {}
+
+  FMT_CONSTEXPR auto locale() -> detail::locale_ref { return loc_; }
+};
+
+template <typename OutputIt, typename Char> class generic_context;
+
+// Longer aliases for C++20 compatibility.
+template <typename OutputIt, typename Char>
+using basic_format_context =
+    conditional_t<std::is_same<OutputIt, appender>::value, context,
+                  generic_context<OutputIt, Char>>;
+using format_context = context;
+
+template <typename Char>
+using buffered_context = basic_format_context<basic_appender<Char>, Char>;
+
+template <typename T, typename Char = char>
+using is_formattable = bool_constant<!std::is_base_of<
+    detail::unformattable, decltype(detail::arg_mapper<buffered_context<Char>>()
+                                        .map(std::declval<T&>()))>::value>;
+
+#if FMT_USE_CONCEPTS
+template <typename T, typename Char = char>
+concept formattable = is_formattable<remove_reference_t<T>, Char>::value;
+#endif
+
+/**
+ * Constructs an object that stores references to arguments and can be
+ * implicitly converted to `format_args`. `Context` can be omitted in which case
+ * it defaults to `format_context`. See `arg` for lifetime considerations.
+ */
+// Take arguments by lvalue references to avoid some lifetime issues, e.g.
+//   auto args = make_format_args(std::string());
+template <typename Context = format_context, typename... T,
+          size_t NUM_ARGS = sizeof...(T),
+          size_t NUM_NAMED_ARGS = detail::count_named_args<T...>(),
+          unsigned long long DESC = detail::make_descriptor<Context, T...>(),
+          FMT_ENABLE_IF(NUM_NAMED_ARGS == 0)>
+constexpr FMT_ALWAYS_INLINE auto make_format_args(T&... args)
+    -> detail::format_arg_store<Context, NUM_ARGS, 0, DESC> {
+  return {{detail::make_arg<NUM_ARGS <= detail::max_packed_args, Context>(
+      args)...}};
+}
+
+#ifndef FMT_DOC
+template <typename Context = format_context, typename... T,
+          size_t NUM_NAMED_ARGS = detail::count_named_args<T...>(),
+          unsigned long long DESC =
+              detail::make_descriptor<Context, T...>() |
+              static_cast<unsigned long long>(detail::has_named_args_bit),
+          FMT_ENABLE_IF(NUM_NAMED_ARGS != 0)>
+constexpr auto make_format_args(T&... args)
+    -> detail::format_arg_store<Context, sizeof...(T), NUM_NAMED_ARGS, DESC> {
+  return {args...};
+}
+#endif
+
+/**
+ * Returns a named argument to be used in a formatting function.
+ * It should only be used in a call to a formatting function or
+ * `dynamic_format_arg_store::push_back`.
+ *
+ * **Example**:
+ *
+ *     fmt::print("The answer is {answer}.", fmt::arg("answer", 42));
+ */
+template <typename Char, typename T>
+inline auto arg(const Char* name, const T& arg) -> detail::named_arg<Char, T> {
+  static_assert(!detail::is_named_arg<T>(), "nested named arguments");
+  return {name, arg};
+}
+FMT_END_EXPORT
+
+/// An alias for `basic_format_args<format_context>`.
+// A separate type would result in shorter symbols but break ABI compatibility
+// between clang and gcc on ARM (#1919).
+FMT_EXPORT using format_args = basic_format_args<format_context>;
+
+// We cannot use enum classes as bit fields because of a gcc bug, so we put them
+// in namespaces instead (https://gcc.gnu.org/bugzilla/show_bug.cgi?id=61414).
+// Additionally, if an underlying type is specified, older gcc incorrectly warns
+// that the type is too small. Both bugs are fixed in gcc 9.3.
+#if FMT_GCC_VERSION && FMT_GCC_VERSION < 903
+#  define FMT_ENUM_UNDERLYING_TYPE(type)
+#else
+#  define FMT_ENUM_UNDERLYING_TYPE(type) : type
+#endif
+namespace align {
+enum type FMT_ENUM_UNDERLYING_TYPE(unsigned char){none, left, right, center,
+                                                  numeric};
+}
+using align_t = align::type;
+namespace sign {
+enum type FMT_ENUM_UNDERLYING_TYPE(unsigned char){none, minus, plus, space};
+}
+using sign_t = sign::type;
+
+namespace detail {
+
+template <typename Char>
+using unsigned_char = typename conditional_t<std::is_integral<Char>::value,
+                                             std::make_unsigned<Char>,
+                                             type_identity<unsigned>>::type;
+
+// Character (code unit) type is erased to prevent template bloat.
+struct fill_t {
+ private:
+  enum { max_size = 4 };
+  char data_[max_size] = {' '};
+  unsigned char size_ = 1;
+
+ public:
+  template <typename Char>
+  FMT_CONSTEXPR void operator=(basic_string_view<Char> s) {
+    auto size = s.size();
+    size_ = static_cast<unsigned char>(size);
+    if (size == 1) {
+      unsigned uchar = static_cast<unsigned_char<Char>>(s[0]);
+      data_[0] = static_cast<char>(uchar);
+      data_[1] = static_cast<char>(uchar >> 8);
+      return;
+    }
+    FMT_ASSERT(size <= max_size, "invalid fill");
+    for (size_t i = 0; i < size; ++i) data_[i] = static_cast<char>(s[i]);
+  }
+
+  FMT_CONSTEXPR void operator=(char c) {
+    data_[0] = c;
+    size_ = 1;
+  }
+
+  constexpr auto size() const -> size_t { return size_; }
+
+  template <typename Char> constexpr auto get() const -> Char {
+    using uchar = unsigned char;
+    return static_cast<Char>(static_cast<uchar>(data_[0]) |
+                             (static_cast<uchar>(data_[1]) << 8));
+  }
+
+  template <typename Char, FMT_ENABLE_IF(std::is_same<Char, char>::value)>
+  constexpr auto data() const -> const Char* {
+    return data_;
+  }
+  template <typename Char, FMT_ENABLE_IF(!std::is_same<Char, char>::value)>
+  constexpr auto data() const -> const Char* {
+    return nullptr;
+  }
+};
+}  // namespace detail
+
+enum class presentation_type : unsigned char {
+  // Common specifiers:
+  none = 0,
+  debug = 1,   // '?'
+  string = 2,  // 's' (string, bool)
+
+  // Integral, bool and character specifiers:
+  dec = 3,  // 'd'
+  hex,      // 'x' or 'X'
+  oct,      // 'o'
+  bin,      // 'b' or 'B'
+  chr,      // 'c'
+
+  // String and pointer specifiers:
+  pointer = 3,  // 'p'
+
+  // Floating-point specifiers:
+  exp = 1,  // 'e' or 'E' (1 since there is no FP debug presentation)
+  fixed,    // 'f' or 'F'
+  general,  // 'g' or 'G'
+  hexfloat  // 'a' or 'A'
+};
+
+// Format specifiers for built-in and string types.
+struct format_specs {
+  int width;
+  int precision;
+  presentation_type type;
+  align_t align : 4;
+  sign_t sign : 3;
+  bool upper : 1;  // An uppercase version e.g. 'X' for 'x'.
+  bool alt : 1;    // Alternate form ('#').
+  bool localized : 1;
+  detail::fill_t fill;
+
+  constexpr format_specs()
+      : width(0),
+        precision(-1),
+        type(presentation_type::none),
+        align(align::none),
+        sign(sign::none),
+        upper(false),
+        alt(false),
+        localized(false) {}
+};
+
+namespace detail {
+
+enum class arg_id_kind { none, index, name };
+
+// An argument reference.
+template <typename Char> struct arg_ref {
+  FMT_CONSTEXPR arg_ref() : kind(arg_id_kind::none), val() {}
+
+  FMT_CONSTEXPR explicit arg_ref(int index)
+      : kind(arg_id_kind::index), val(index) {}
+  FMT_CONSTEXPR explicit arg_ref(basic_string_view<Char> name)
+      : kind(arg_id_kind::name), val(name) {}
+
+  FMT_CONSTEXPR auto operator=(int idx) -> arg_ref& {
+    kind = arg_id_kind::index;
+    val.index = idx;
+    return *this;
+  }
+
+  arg_id_kind kind;
+  union value {
+    FMT_CONSTEXPR value(int idx = 0) : index(idx) {}
+    FMT_CONSTEXPR value(basic_string_view<Char> n) : name(n) {}
+
+    int index;
+    basic_string_view<Char> name;
+  } val;
+};
+
+// Format specifiers with width and precision resolved at formatting rather
+// than parsing time to allow reusing the same parsed specifiers with
+// different sets of arguments (precompilation of format strings).
+template <typename Char = char> struct dynamic_format_specs : format_specs {
+  arg_ref<Char> width_ref;
+  arg_ref<Char> precision_ref;
+};
+
+// Converts a character to ASCII. Returns '\0' on conversion failure.
+template <typename Char, FMT_ENABLE_IF(std::is_integral<Char>::value)>
+constexpr auto to_ascii(Char c) -> char {
+  return c <= 0xff ? static_cast<char>(c) : '\0';
+}
+
+// Returns the number of code units in a code point or 1 on error.
+template <typename Char>
+FMT_CONSTEXPR auto code_point_length(const Char* begin) -> int {
+  if (const_check(sizeof(Char) != 1)) return 1;
+  auto c = static_cast<unsigned char>(*begin);
+  return static_cast<int>((0x3a55000000000000ull >> (2 * (c >> 3))) & 0x3) + 1;
+}
+
+// Return the result via the out param to workaround gcc bug 77539.
+template <bool IS_CONSTEXPR, typename T, typename Ptr = const T*>
+FMT_CONSTEXPR auto find(Ptr first, Ptr last, T value, Ptr& out) -> bool {
+  for (out = first; out != last; ++out) {
+    if (*out == value) return true;
+  }
+  return false;
+}
+
+template <>
+inline auto find<false, char>(const char* first, const char* last, char value,
+                              const char*& out) -> bool {
+  out =
+      static_cast<const char*>(memchr(first, value, to_unsigned(last - first)));
+  return out != nullptr;
+}
+
+// Parses the range [begin, end) as an unsigned integer. This function assumes
+// that the range is non-empty and the first character is a digit.
+template <typename Char>
+FMT_CONSTEXPR auto parse_nonnegative_int(const Char*& begin, const Char* end,
+                                         int error_value) noexcept -> int {
+  FMT_ASSERT(begin != end && '0' <= *begin && *begin <= '9', "");
+  unsigned value = 0, prev = 0;
+  auto p = begin;
+  do {
+    prev = value;
+    value = value * 10 + unsigned(*p - '0');
+    ++p;
+  } while (p != end && '0' <= *p && *p <= '9');
+  auto num_digits = p - begin;
+  begin = p;
+  int digits10 = static_cast<int>(sizeof(int) * CHAR_BIT * 3 / 10);
+  if (num_digits <= digits10) return static_cast<int>(value);
+  // Check for overflow.
+  unsigned max = INT_MAX;
+  return num_digits == digits10 + 1 &&
+                 prev * 10ull + unsigned(p[-1] - '0') <= max
+             ? static_cast<int>(value)
+             : error_value;
+}
+
+FMT_CONSTEXPR inline auto parse_align(char c) -> align_t {
+  switch (c) {
+  case '<':
+    return align::left;
+  case '>':
+    return align::right;
+  case '^':
+    return align::center;
+  }
+  return align::none;
+}
+
+template <typename Char> constexpr auto is_name_start(Char c) -> bool {
+  return ('a' <= c && c <= 'z') || ('A' <= c && c <= 'Z') || c == '_';
+}
+
+template <typename Char, typename Handler>
+FMT_CONSTEXPR auto do_parse_arg_id(const Char* begin, const Char* end,
+                                   Handler&& handler) -> const Char* {
+  Char c = *begin;
+  if (c >= '0' && c <= '9') {
+    int index = 0;
+    if (c != '0')
+      index = parse_nonnegative_int(begin, end, INT_MAX);
+    else
+      ++begin;
+    if (begin == end || (*begin != '}' && *begin != ':'))
+      report_error("invalid format string");
+    else
+      handler.on_index(index);
+    return begin;
+  }
+  if (!is_name_start(c)) {
+    report_error("invalid format string");
+    return begin;
+  }
+  auto it = begin;
+  do {
+    ++it;
+  } while (it != end && (is_name_start(*it) || ('0' <= *it && *it <= '9')));
+  handler.on_name({begin, to_unsigned(it - begin)});
+  return it;
+}
+
+template <typename Char, typename Handler>
+FMT_CONSTEXPR auto parse_arg_id(const Char* begin, const Char* end,
+                                Handler&& handler) -> const Char* {
+  FMT_ASSERT(begin != end, "");
+  Char c = *begin;
+  if (c != '}' && c != ':') return do_parse_arg_id(begin, end, handler);
+  handler.on_auto();
+  return begin;
+}
+
+template <typename Char> struct dynamic_spec_id_handler {
+  basic_format_parse_context<Char>& ctx;
+  arg_ref<Char>& ref;
+
+  FMT_CONSTEXPR void on_auto() {
+    int id = ctx.next_arg_id();
+    ref = arg_ref<Char>(id);
+    ctx.check_dynamic_spec(id);
+  }
+  FMT_CONSTEXPR void on_index(int id) {
+    ref = arg_ref<Char>(id);
+    ctx.check_arg_id(id);
+    ctx.check_dynamic_spec(id);
+  }
+  FMT_CONSTEXPR void on_name(basic_string_view<Char> id) {
+    ref = arg_ref<Char>(id);
+    ctx.check_arg_id(id);
+  }
+};
+
+// Parses [integer | "{" [arg_id] "}"].
+template <typename Char>
+FMT_CONSTEXPR auto parse_dynamic_spec(const Char* begin, const Char* end,
+                                      int& value, arg_ref<Char>& ref,
+                                      basic_format_parse_context<Char>& ctx)
+    -> const Char* {
+  FMT_ASSERT(begin != end, "");
+  if ('0' <= *begin && *begin <= '9') {
+    int val = parse_nonnegative_int(begin, end, -1);
+    if (val != -1)
+      value = val;
+    else
+      report_error("number is too big");
+  } else if (*begin == '{') {
+    ++begin;
+    auto handler = dynamic_spec_id_handler<Char>{ctx, ref};
+    if (begin != end) begin = parse_arg_id(begin, end, handler);
+    if (begin != end && *begin == '}') return ++begin;
+    report_error("invalid format string");
+  }
+  return begin;
+}
+
+template <typename Char>
+FMT_CONSTEXPR auto parse_precision(const Char* begin, const Char* end,
+                                   int& value, arg_ref<Char>& ref,
+                                   basic_format_parse_context<Char>& ctx)
+    -> const Char* {
+  ++begin;
+  if (begin == end || *begin == '}') {
+    report_error("invalid precision");
+    return begin;
+  }
+  return parse_dynamic_spec(begin, end, value, ref, ctx);
+}
+
+enum class state { start, align, sign, hash, zero, width, precision, locale };
+
+// Parses standard format specifiers.
+template <typename Char>
+FMT_CONSTEXPR auto parse_format_specs(const Char* begin, const Char* end,
+                                      dynamic_format_specs<Char>& specs,
+                                      basic_format_parse_context<Char>& ctx,
+                                      type arg_type) -> const Char* {
+  auto c = '\0';
+  if (end - begin > 1) {
+    auto next = to_ascii(begin[1]);
+    c = parse_align(next) == align::none ? to_ascii(*begin) : '\0';
+  } else {
+    if (begin == end) return begin;
+    c = to_ascii(*begin);
+  }
+
+  struct {
+    state current_state = state::start;
+    FMT_CONSTEXPR void operator()(state s, bool valid = true) {
+      if (current_state >= s || !valid)
+        report_error("invalid format specifier");
+      current_state = s;
+    }
+  } enter_state;
+
+  using pres = presentation_type;
+  constexpr auto integral_set = sint_set | uint_set | bool_set | char_set;
+  struct {
+    const Char*& begin;
+    dynamic_format_specs<Char>& specs;
+    type arg_type;
+
+    FMT_CONSTEXPR auto operator()(pres pres_type, int set) -> const Char* {
+      if (!in(arg_type, set)) {
+        if (arg_type == type::none_type) return begin;
+        report_error("invalid format specifier");
+      }
+      specs.type = pres_type;
+      return begin + 1;
+    }
+  } parse_presentation_type{begin, specs, arg_type};
+
+  for (;;) {
+    switch (c) {
+    case '<':
+    case '>':
+    case '^':
+      enter_state(state::align);
+      specs.align = parse_align(c);
+      ++begin;
+      break;
+    case '+':
+    case '-':
+    case ' ':
+      if (arg_type == type::none_type) return begin;
+      enter_state(state::sign, in(arg_type, sint_set | float_set));
+      switch (c) {
+      case '+':
+        specs.sign = sign::plus;
+        break;
+      case '-':
+        specs.sign = sign::minus;
+        break;
+      case ' ':
+        specs.sign = sign::space;
+        break;
+      }
+      ++begin;
+      break;
+    case '#':
+      if (arg_type == type::none_type) return begin;
+      enter_state(state::hash, is_arithmetic_type(arg_type));
+      specs.alt = true;
+      ++begin;
+      break;
+    case '0':
+      enter_state(state::zero);
+      if (!is_arithmetic_type(arg_type)) {
+        if (arg_type == type::none_type) return begin;
+        report_error("format specifier requires numeric argument");
+      }
+      if (specs.align == align::none) {
+        // Ignore 0 if align is specified for compatibility with std::format.
+        specs.align = align::numeric;
+        specs.fill = '0';
+      }
+      ++begin;
+      break;
+    case '1':
+    case '2':
+    case '3':
+    case '4':
+    case '5':
+    case '6':
+    case '7':
+    case '8':
+    case '9':
+    case '{':
+      enter_state(state::width);
+      begin = parse_dynamic_spec(begin, end, specs.width, specs.width_ref, ctx);
+      break;
+    case '.':
+      if (arg_type == type::none_type) return begin;
+      enter_state(state::precision,
+                  in(arg_type, float_set | string_set | cstring_set));
+      begin = parse_precision(begin, end, specs.precision, specs.precision_ref,
+                              ctx);
+      break;
+    case 'L':
+      if (arg_type == type::none_type) return begin;
+      enter_state(state::locale, is_arithmetic_type(arg_type));
+      specs.localized = true;
+      ++begin;
+      break;
+    case 'd':
+      return parse_presentation_type(pres::dec, integral_set);
+    case 'X':
+      specs.upper = true;
+      FMT_FALLTHROUGH;
+    case 'x':
+      return parse_presentation_type(pres::hex, integral_set);
+    case 'o':
+      return parse_presentation_type(pres::oct, integral_set);
+    case 'B':
+      specs.upper = true;
+      FMT_FALLTHROUGH;
+    case 'b':
+      return parse_presentation_type(pres::bin, integral_set);
+    case 'E':
+      specs.upper = true;
+      FMT_FALLTHROUGH;
+    case 'e':
+      return parse_presentation_type(pres::exp, float_set);
+    case 'F':
+      specs.upper = true;
+      FMT_FALLTHROUGH;
+    case 'f':
+      return parse_presentation_type(pres::fixed, float_set);
+    case 'G':
+      specs.upper = true;
+      FMT_FALLTHROUGH;
+    case 'g':
+      return parse_presentation_type(pres::general, float_set);
+    case 'A':
+      specs.upper = true;
+      FMT_FALLTHROUGH;
+    case 'a':
+      return parse_presentation_type(pres::hexfloat, float_set);
+    case 'c':
+      if (arg_type == type::bool_type) report_error("invalid format specifier");
+      return parse_presentation_type(pres::chr, integral_set);
+    case 's':
+      return parse_presentation_type(pres::string,
+                                     bool_set | string_set | cstring_set);
+    case 'p':
+      return parse_presentation_type(pres::pointer, pointer_set | cstring_set);
+    case '?':
+      return parse_presentation_type(pres::debug,
+                                     char_set | string_set | cstring_set);
+    case '}':
+      return begin;
+    default: {
+      if (*begin == '}') return begin;
+      // Parse fill and alignment.
+      auto fill_end = begin + code_point_length(begin);
+      if (end - fill_end <= 0) {
+        report_error("invalid format specifier");
+        return begin;
+      }
+      if (*begin == '{') {
+        report_error("invalid fill character '{'");
+        return begin;
+      }
+      auto align = parse_align(to_ascii(*fill_end));
+      enter_state(state::align, align != align::none);
+      specs.fill =
+          basic_string_view<Char>(begin, to_unsigned(fill_end - begin));
+      specs.align = align;
+      begin = fill_end + 1;
+    }
+    }
+    if (begin == end) return begin;
+    c = to_ascii(*begin);
+  }
+}
+
+template <typename Char, typename Handler>
+FMT_CONSTEXPR auto parse_replacement_field(const Char* begin, const Char* end,
+                                           Handler&& handler) -> const Char* {
+  struct id_adapter {
+    Handler& handler;
+    int arg_id;
+
+    FMT_CONSTEXPR void on_auto() { arg_id = handler.on_arg_id(); }
+    FMT_CONSTEXPR void on_index(int id) { arg_id = handler.on_arg_id(id); }
+    FMT_CONSTEXPR void on_name(basic_string_view<Char> id) {
+      arg_id = handler.on_arg_id(id);
+    }
+  };
+
+  ++begin;
+  if (begin == end) return handler.on_error("invalid format string"), end;
+  if (*begin == '}') {
+    handler.on_replacement_field(handler.on_arg_id(), begin);
+  } else if (*begin == '{') {
+    handler.on_text(begin, begin + 1);
+  } else {
+    auto adapter = id_adapter{handler, 0};
+    begin = parse_arg_id(begin, end, adapter);
+    Char c = begin != end ? *begin : Char();
+    if (c == '}') {
+      handler.on_replacement_field(adapter.arg_id, begin);
+    } else if (c == ':') {
+      begin = handler.on_format_specs(adapter.arg_id, begin + 1, end);
+      if (begin == end || *begin != '}')
+        return handler.on_error("unknown format specifier"), end;
+    } else {
+      return handler.on_error("missing '}' in format string"), end;
+    }
+  }
+  return begin + 1;
+}
+
+template <bool IS_CONSTEXPR, typename Char, typename Handler>
+FMT_CONSTEXPR void parse_format_string(basic_string_view<Char> format_str,
+                                       Handler&& handler) {
+  auto begin = format_str.data();
+  auto end = begin + format_str.size();
+  if (end - begin < 32) {
+    // Use a simple loop instead of memchr for small strings.
+    const Char* p = begin;
+    while (p != end) {
+      auto c = *p++;
+      if (c == '{') {
+        handler.on_text(begin, p - 1);
+        begin = p = parse_replacement_field(p - 1, end, handler);
+      } else if (c == '}') {
+        if (p == end || *p != '}')
+          return handler.on_error("unmatched '}' in format string");
+        handler.on_text(begin, p);
+        begin = ++p;
+      }
+    }
+    handler.on_text(begin, end);
+    return;
+  }
+  struct writer {
+    FMT_CONSTEXPR void operator()(const Char* from, const Char* to) {
+      if (from == to) return;
+      for (;;) {
+        const Char* p = nullptr;
+        if (!find<IS_CONSTEXPR>(from, to, Char('}'), p))
+          return handler_.on_text(from, to);
+        ++p;
+        if (p == to || *p != '}')
+          return handler_.on_error("unmatched '}' in format string");
+        handler_.on_text(from, p);
+        from = p + 1;
+      }
+    }
+    Handler& handler_;
+  } write = {handler};
+  while (begin != end) {
+    // Doing two passes with memchr (one for '{' and another for '}') is up to
+    // 2.5x faster than the naive one-pass implementation on big format strings.
+    const Char* p = begin;
+    if (*begin != '{' && !find<IS_CONSTEXPR>(begin + 1, end, Char('{'), p))
+      return write(begin, end);
+    write(begin, p);
+    begin = parse_replacement_field(p, end, handler);
+  }
+}
+
+template <typename T, bool = is_named_arg<T>::value> struct strip_named_arg {
+  using type = T;
+};
+template <typename T> struct strip_named_arg<T, true> {
+  using type = remove_cvref_t<decltype(T::value)>;
+};
+
+template <typename T, typename ParseContext>
+FMT_VISIBILITY("hidden")  // Suppress an ld warning on macOS (#3769).
+FMT_CONSTEXPR auto parse_format_specs(ParseContext& ctx)
+    -> decltype(ctx.begin()) {
+  using char_type = typename ParseContext::char_type;
+  using context = buffered_context<char_type>;
+  using mapped_type = conditional_t<
+      mapped_type_constant<T, context>::value != type::custom_type,
+      decltype(arg_mapper<context>().map(std::declval<const T&>())),
+      typename strip_named_arg<T>::type>;
+#if defined(__cpp_if_constexpr)
+  if constexpr (std::is_default_constructible<
+                    formatter<mapped_type, char_type>>::value) {
+    return formatter<mapped_type, char_type>().parse(ctx);
+  } else {
+    type_is_unformattable_for<T, char_type> _;
+    return ctx.begin();
+  }
+#else
+  return formatter<mapped_type, char_type>().parse(ctx);
+#endif
+}
+
+// Checks char specs and returns true iff the presentation type is char-like.
+FMT_CONSTEXPR inline auto check_char_specs(const format_specs& specs) -> bool {
+  if (specs.type != presentation_type::none &&
+      specs.type != presentation_type::chr &&
+      specs.type != presentation_type::debug) {
+    return false;
+  }
+  if (specs.align == align::numeric || specs.sign != sign::none || specs.alt)
+    report_error("invalid format specifier for char");
+  return true;
+}
+
+#if FMT_USE_NONTYPE_TEMPLATE_ARGS
+template <int N, typename T, typename... Args, typename Char>
+constexpr auto get_arg_index_by_name(basic_string_view<Char> name) -> int {
+  if constexpr (is_statically_named_arg<T>()) {
+    if (name == T::name) return N;
+  }
+  if constexpr (sizeof...(Args) > 0)
+    return get_arg_index_by_name<N + 1, Args...>(name);
+  (void)name;  // Workaround an MSVC bug about "unused" parameter.
+  return -1;
+}
+#endif
+
+template <typename... Args, typename Char>
+FMT_CONSTEXPR auto get_arg_index_by_name(basic_string_view<Char> name) -> int {
+#if FMT_USE_NONTYPE_TEMPLATE_ARGS
+  if constexpr (sizeof...(Args) > 0)
+    return get_arg_index_by_name<0, Args...>(name);
+#endif
+  (void)name;
+  return -1;
+}
+
+template <typename Char, typename... Args> class format_string_checker {
+ private:
+  using parse_context_type = compile_parse_context<Char>;
+  static constexpr int num_args = sizeof...(Args);
+
+  // Format specifier parsing function.
+  // In the future basic_format_parse_context will replace compile_parse_context
+  // here and will use is_constant_evaluated and downcasting to access the data
+  // needed for compile-time checks: https://godbolt.org/z/GvWzcTjh1.
+  using parse_func = const Char* (*)(parse_context_type&);
+
+  type types_[num_args > 0 ? static_cast<size_t>(num_args) : 1];
+  parse_context_type context_;
+  parse_func parse_funcs_[num_args > 0 ? static_cast<size_t>(num_args) : 1];
+
+ public:
+  explicit FMT_CONSTEXPR format_string_checker(basic_string_view<Char> fmt)
+      : types_{mapped_type_constant<Args, buffered_context<Char>>::value...},
+        context_(fmt, num_args, types_),
+        parse_funcs_{&parse_format_specs<Args, parse_context_type>...} {}
+
+  FMT_CONSTEXPR void on_text(const Char*, const Char*) {}
+
+  FMT_CONSTEXPR auto on_arg_id() -> int { return context_.next_arg_id(); }
+  FMT_CONSTEXPR auto on_arg_id(int id) -> int {
+    return context_.check_arg_id(id), id;
+  }
+  FMT_CONSTEXPR auto on_arg_id(basic_string_view<Char> id) -> int {
+#if FMT_USE_NONTYPE_TEMPLATE_ARGS
+    auto index = get_arg_index_by_name<Args...>(id);
+    if (index < 0) on_error("named argument is not found");
+    return index;
+#else
+    (void)id;
+    on_error("compile-time checks for named arguments require C++20 support");
+    return 0;
+#endif
+  }
+
+  FMT_CONSTEXPR void on_replacement_field(int id, const Char* begin) {
+    on_format_specs(id, begin, begin);  // Call parse() on empty specs.
+  }
+
+  FMT_CONSTEXPR auto on_format_specs(int id, const Char* begin, const Char*)
+      -> const Char* {
+    context_.advance_to(begin);
+    // id >= 0 check is a workaround for gcc 10 bug (#2065).
+    return id >= 0 && id < num_args ? parse_funcs_[id](context_) : begin;
+  }
+
+  FMT_NORETURN FMT_CONSTEXPR void on_error(const char* message) {
+    report_error(message);
+  }
+};
+
+// A base class for compile-time strings.
+struct compile_string {};
+
+template <typename S>
+using is_compile_string = std::is_base_of<compile_string, S>;
+
+// Reports a compile-time error if S is not a valid format string.
+template <typename..., typename S, FMT_ENABLE_IF(!is_compile_string<S>::value)>
+FMT_ALWAYS_INLINE void check_format_string(const S&) {
+#ifdef FMT_ENFORCE_COMPILE_STRING
+  static_assert(is_compile_string<S>::value,
+                "FMT_ENFORCE_COMPILE_STRING requires all format strings to use "
+                "FMT_STRING.");
+#endif
+}
+template <typename... Args, typename S,
+          FMT_ENABLE_IF(is_compile_string<S>::value)>
+void check_format_string(S format_str) {
+  using char_t = typename S::char_type;
+  FMT_CONSTEXPR auto s = basic_string_view<char_t>(format_str);
+  using checker = format_string_checker<char_t, remove_cvref_t<Args>...>;
+  FMT_CONSTEXPR bool error = (parse_format_string<true>(s, checker(s)), true);
+  ignore_unused(error);
+}
+
+// Report truncation to prevent silent data loss.
+inline void report_truncation(bool truncated) {
+  if (truncated) report_error("output is truncated");
+}
+
+// Use vformat_args and avoid type_identity to keep symbols short and workaround
+// a GCC <= 4.8 bug.
+template <typename Char = char> struct vformat_args {
+  using type = basic_format_args<buffered_context<Char>>;
+};
+template <> struct vformat_args<char> {
+  using type = format_args;
+};
+
+template <typename Char>
+void vformat_to(buffer<Char>& buf, basic_string_view<Char> fmt,
+                typename vformat_args<Char>::type args, locale_ref loc = {});
+
+FMT_API void vprint_mojibake(FILE*, string_view, format_args, bool = false);
+#ifndef _WIN32
+inline void vprint_mojibake(FILE*, string_view, format_args, bool) {}
+#endif
+
+template <typename T, typename Char, type TYPE> struct native_formatter {
+ private:
+  dynamic_format_specs<Char> specs_;
+
+ public:
+  using nonlocking = void;
+
+  template <typename ParseContext>
+  FMT_CONSTEXPR auto parse(ParseContext& ctx) -> const Char* {
+    if (ctx.begin() == ctx.end() || *ctx.begin() == '}') return ctx.begin();
+    auto end = parse_format_specs(ctx.begin(), ctx.end(), specs_, ctx, TYPE);
+    if (const_check(TYPE == type::char_type)) check_char_specs(specs_);
+    return end;
+  }
+
+  template <type U = TYPE,
+            FMT_ENABLE_IF(U == type::string_type || U == type::cstring_type ||
+                          U == type::char_type)>
+  FMT_CONSTEXPR void set_debug_format(bool set = true) {
+    specs_.type = set ? presentation_type::debug : presentation_type::none;
+  }
+
+  template <typename FormatContext>
+  FMT_CONSTEXPR auto format(const T& val, FormatContext& ctx) const
+      -> decltype(ctx.out());
+};
+}  // namespace detail
+
+FMT_BEGIN_EXPORT
+
+// A formatter specialization for natively supported types.
+template <typename T, typename Char>
+struct formatter<T, Char,
+                 enable_if_t<detail::type_constant<T, Char>::value !=
+                             detail::type::custom_type>>
+    : detail::native_formatter<T, Char, detail::type_constant<T, Char>::value> {
+};
+
+template <typename Char = char> struct runtime_format_string {
+  basic_string_view<Char> str;
+};
+
+/// A compile-time format string.
+template <typename Char, typename... Args> class basic_format_string {
+ private:
+  basic_string_view<Char> str_;
+
+ public:
+  template <
+      typename S,
+      FMT_ENABLE_IF(
+          std::is_convertible<const S&, basic_string_view<Char>>::value ||
+          (detail::is_compile_string<S>::value &&
+           std::is_constructible<basic_string_view<Char>, const S&>::value))>
+  FMT_CONSTEVAL FMT_ALWAYS_INLINE basic_format_string(const S& s) : str_(s) {
+    static_assert(
+        detail::count<
+            (std::is_base_of<detail::view, remove_reference_t<Args>>::value &&
+             std::is_reference<Args>::value)...>() == 0,
+        "passing views as lvalues is disallowed");
+#if FMT_USE_CONSTEVAL
+    if constexpr (detail::count_named_args<Args...>() ==
+                  detail::count_statically_named_args<Args...>()) {
+      using checker =
+          detail::format_string_checker<Char, remove_cvref_t<Args>...>;
+      detail::parse_format_string<true>(str_, checker(s));
+    }
+#else
+    detail::check_format_string<Args...>(s);
+#endif
+  }
+  basic_format_string(runtime_format_string<Char> fmt) : str_(fmt.str) {}
+
+  FMT_ALWAYS_INLINE operator basic_string_view<Char>() const { return str_; }
+  auto get() const -> basic_string_view<Char> { return str_; }
+};
+
+#if FMT_GCC_VERSION && FMT_GCC_VERSION < 409
+// Workaround broken conversion on older gcc.
+template <typename...> using format_string = string_view;
+inline auto runtime(string_view s) -> string_view { return s; }
+#else
+template <typename... Args>
+using format_string = basic_format_string<char, type_identity_t<Args>...>;
+/**
+ * Creates a runtime format string.
+ *
+ * **Example**:
+ *
+ *     // Check format string at runtime instead of compile-time.
+ *     fmt::print(fmt::runtime("{:d}"), "I am not a number");
+ */
+inline auto runtime(string_view s) -> runtime_format_string<> { return {{s}}; }
+#endif
+
+/// Formats a string and writes the output to `out`.
+template <typename OutputIt,
+          FMT_ENABLE_IF(detail::is_output_iterator<remove_cvref_t<OutputIt>,
+                                                   char>::value)>
+auto vformat_to(OutputIt&& out, string_view fmt, format_args args)
+    -> remove_cvref_t<OutputIt> {
+  auto&& buf = detail::get_buffer<char>(out);
+  detail::vformat_to(buf, fmt, args, {});
+  return detail::get_iterator(buf, out);
+}
+
+/**
+ * Formats `args` according to specifications in `fmt`, writes the result to
+ * the output iterator `out` and returns the iterator past the end of the output
+ * range. `format_to` does not append a terminating null character.
+ *
+ * **Example**:
+ *
+ *     auto out = std::vector<char>();
+ *     fmt::format_to(std::back_inserter(out), "{}", 42);
+ */
+template <typename OutputIt, typename... T,
+          FMT_ENABLE_IF(detail::is_output_iterator<remove_cvref_t<OutputIt>,
+                                                   char>::value)>
+FMT_INLINE auto format_to(OutputIt&& out, format_string<T...> fmt, T&&... args)
+    -> remove_cvref_t<OutputIt> {
+  return vformat_to(FMT_FWD(out), fmt, fmt::make_format_args(args...));
+}
+
+template <typename OutputIt> struct format_to_n_result {
+  /// Iterator past the end of the output range.
+  OutputIt out;
+  /// Total (not truncated) output size.
+  size_t size;
+};
+
+template <typename OutputIt, typename... T,
+          FMT_ENABLE_IF(detail::is_output_iterator<OutputIt, char>::value)>
+auto vformat_to_n(OutputIt out, size_t n, string_view fmt, format_args args)
+    -> format_to_n_result<OutputIt> {
+  using traits = detail::fixed_buffer_traits;
+  auto buf = detail::iterator_buffer<OutputIt, char, traits>(out, n);
+  detail::vformat_to(buf, fmt, args, {});
+  return {buf.out(), buf.count()};
+}
+
+/**
+ * Formats `args` according to specifications in `fmt`, writes up to `n`
+ * characters of the result to the output iterator `out` and returns the total
+ * (not truncated) output size and the iterator past the end of the output
+ * range. `format_to_n` does not append a terminating null character.
+ */
+template <typename OutputIt, typename... T,
+          FMT_ENABLE_IF(detail::is_output_iterator<OutputIt, char>::value)>
+FMT_INLINE auto format_to_n(OutputIt out, size_t n, format_string<T...> fmt,
+                            T&&... args) -> format_to_n_result<OutputIt> {
+  return vformat_to_n(out, n, fmt, fmt::make_format_args(args...));
+}
+
+template <typename OutputIt, typename Sentinel = OutputIt>
+struct format_to_result {
+  /// Iterator pointing to just after the last successful write in the range.
+  OutputIt out;
+  /// Specifies if the output was truncated.
+  bool truncated;
+
+  FMT_CONSTEXPR operator OutputIt&() & {
+    detail::report_truncation(truncated);
+    return out;
+  }
+  FMT_CONSTEXPR operator const OutputIt&() const& {
+    detail::report_truncation(truncated);
+    return out;
+  }
+  FMT_CONSTEXPR operator OutputIt&&() && {
+    detail::report_truncation(truncated);
+    return static_cast<OutputIt&&>(out);
+  }
+};
+
+template <size_t N>
+auto vformat_to(char (&out)[N], string_view fmt, format_args args)
+    -> format_to_result<char*> {
+  auto result = vformat_to_n(out, N, fmt, args);
+  return {result.out, result.size > N};
+}
+
+template <size_t N, typename... T>
+FMT_INLINE auto format_to(char (&out)[N], format_string<T...> fmt, T&&... args)
+    -> format_to_result<char*> {
+  auto result = fmt::format_to_n(out, N, fmt, static_cast<T&&>(args)...);
+  return {result.out, result.size > N};
+}
+
+/// Returns the number of chars in the output of `format(fmt, args...)`.
+template <typename... T>
+FMT_NODISCARD FMT_INLINE auto formatted_size(format_string<T...> fmt,
+                                             T&&... args) -> size_t {
+  auto buf = detail::counting_buffer<>();
+  detail::vformat_to<char>(buf, fmt, fmt::make_format_args(args...), {});
+  return buf.count();
+}
+
+FMT_API void vprint(string_view fmt, format_args args);
+FMT_API void vprint(FILE* f, string_view fmt, format_args args);
+FMT_API void vprint_buffered(FILE* f, string_view fmt, format_args args);
+FMT_API void vprintln(FILE* f, string_view fmt, format_args args);
+
+/**
+ * Formats `args` according to specifications in `fmt` and writes the output
+ * to `stdout`.
+ *
+ * **Example**:
+ *
+ *     fmt::print("The answer is {}.", 42);
+ */
+template <typename... T>
+FMT_INLINE void print(format_string<T...> fmt, T&&... args) {
+  const auto& vargs = fmt::make_format_args(args...);
+  if (!detail::use_utf8()) return detail::vprint_mojibake(stdout, fmt, vargs);
+  return detail::is_locking<T...>() ? vprint_buffered(stdout, fmt, vargs)
+                                    : vprint(fmt, vargs);
+}
+
+/**
+ * Formats `args` according to specifications in `fmt` and writes the
+ * output to the file `f`.
+ *
+ * **Example**:
+ *
+ *     fmt::print(stderr, "Don't {}!", "panic");
+ */
+template <typename... T>
+FMT_INLINE void print(FILE* f, format_string<T...> fmt, T&&... args) {
+  const auto& vargs = fmt::make_format_args(args...);
+  if (!detail::use_utf8()) return detail::vprint_mojibake(f, fmt, vargs);
+  return detail::is_locking<T...>() ? vprint_buffered(f, fmt, vargs)
+                                    : vprint(f, fmt, vargs);
+}
+
+/// Formats `args` according to specifications in `fmt` and writes the output
+/// to the file `f` followed by a newline.
+template <typename... T>
+FMT_INLINE void println(FILE* f, format_string<T...> fmt, T&&... args) {
+  const auto& vargs = fmt::make_format_args(args...);
+  return detail::use_utf8() ? vprintln(f, fmt, vargs)
+                            : detail::vprint_mojibake(f, fmt, vargs, true);
+}
+
+/// Formats `args` according to specifications in `fmt` and writes the output
+/// to `stdout` followed by a newline.
+template <typename... T>
+FMT_INLINE void println(format_string<T...> fmt, T&&... args) {
+  return fmt::println(stdout, fmt, static_cast<T&&>(args)...);
+}
+
+FMT_END_EXPORT
+FMT_GCC_PRAGMA("GCC pop_options")
+FMT_END_NAMESPACE
+
+#ifdef FMT_HEADER_ONLY
+#  include "format.h"
+#endif
+#endif  // FMT_BASE_H_
diff --git a/ext/spdlog/fmt/bundled/chrono.h b/ext/spdlog/fmt/bundled/chrono.h
new file mode 100644
index 0000000..c93123f
--- /dev/null
+++ b/ext/spdlog/fmt/bundled/chrono.h
@@ -0,0 +1,2432 @@
+// Formatting library for C++ - chrono support
+//
+// Copyright (c) 2012 - present, Victor Zverovich
+// All rights reserved.
+//
+// For the license information refer to format.h.
+
+#ifndef FMT_CHRONO_H_
+#define FMT_CHRONO_H_
+
+#ifndef FMT_MODULE
+#  include <algorithm>
+#  include <chrono>
+#  include <cmath>    // std::isfinite
+#  include <cstring>  // std::memcpy
+#  include <ctime>
+#  include <iterator>
+#  include <locale>
+#  include <ostream>
+#  include <type_traits>
+#endif
+
+#include "format.h"
+
+FMT_BEGIN_NAMESPACE
+
+// Check if std::chrono::local_t is available.
+#ifndef FMT_USE_LOCAL_TIME
+#  ifdef __cpp_lib_chrono
+#    define FMT_USE_LOCAL_TIME (__cpp_lib_chrono >= 201907L)
+#  else
+#    define FMT_USE_LOCAL_TIME 0
+#  endif
+#endif
+
+// Check if std::chrono::utc_timestamp is available.
+#ifndef FMT_USE_UTC_TIME
+#  ifdef __cpp_lib_chrono
+#    define FMT_USE_UTC_TIME (__cpp_lib_chrono >= 201907L)
+#  else
+#    define FMT_USE_UTC_TIME 0
+#  endif
+#endif
+
+// Enable tzset.
+#ifndef FMT_USE_TZSET
+// UWP doesn't provide _tzset.
+#  if FMT_HAS_INCLUDE("winapifamily.h")
+#    include <winapifamily.h>
+#  endif
+#  if defined(_WIN32) && (!defined(WINAPI_FAMILY) || \
+                          (WINAPI_FAMILY == WINAPI_FAMILY_DESKTOP_APP))
+#    define FMT_USE_TZSET 1
+#  else
+#    define FMT_USE_TZSET 0
+#  endif
+#endif
+
+// Enable safe chrono durations, unless explicitly disabled.
+#ifndef FMT_SAFE_DURATION_CAST
+#  define FMT_SAFE_DURATION_CAST 1
+#endif
+#if FMT_SAFE_DURATION_CAST
+
+// For conversion between std::chrono::durations without undefined
+// behaviour or erroneous results.
+// This is a stripped down version of duration_cast, for inclusion in fmt.
+// See https://github.com/pauldreik/safe_duration_cast
+//
+// Copyright Paul Dreik 2019
+namespace safe_duration_cast {
+
+template <typename To, typename From,
+          FMT_ENABLE_IF(!std::is_same<From, To>::value &&
+                        std::numeric_limits<From>::is_signed ==
+                            std::numeric_limits<To>::is_signed)>
+FMT_CONSTEXPR auto lossless_integral_conversion(const From from, int& ec)
+    -> To {
+  ec = 0;
+  using F = std::numeric_limits<From>;
+  using T = std::numeric_limits<To>;
+  static_assert(F::is_integer, "From must be integral");
+  static_assert(T::is_integer, "To must be integral");
+
+  // A and B are both signed, or both unsigned.
+  if (detail::const_check(F::digits <= T::digits)) {
+    // From fits in To without any problem.
+  } else {
+    // From does not always fit in To, resort to a dynamic check.
+    if (from < (T::min)() || from > (T::max)()) {
+      // outside range.
+      ec = 1;
+      return {};
+    }
+  }
+  return static_cast<To>(from);
+}
+
+/// Converts From to To, without loss. If the dynamic value of from
+/// can't be converted to To without loss, ec is set.
+template <typename To, typename From,
+          FMT_ENABLE_IF(!std::is_same<From, To>::value &&
+                        std::numeric_limits<From>::is_signed !=
+                            std::numeric_limits<To>::is_signed)>
+FMT_CONSTEXPR auto lossless_integral_conversion(const From from, int& ec)
+    -> To {
+  ec = 0;
+  using F = std::numeric_limits<From>;
+  using T = std::numeric_limits<To>;
+  static_assert(F::is_integer, "From must be integral");
+  static_assert(T::is_integer, "To must be integral");
+
+  if (detail::const_check(F::is_signed && !T::is_signed)) {
+    // From may be negative, not allowed!
+    if (fmt::detail::is_negative(from)) {
+      ec = 1;
+      return {};
+    }
+    // From is positive. Can it always fit in To?
+    if (detail::const_check(F::digits > T::digits) &&
+        from > static_cast<From>(detail::max_value<To>())) {
+      ec = 1;
+      return {};
+    }
+  }
+
+  if (detail::const_check(!F::is_signed && T::is_signed &&
+                          F::digits >= T::digits) &&
+      from > static_cast<From>(detail::max_value<To>())) {
+    ec = 1;
+    return {};
+  }
+  return static_cast<To>(from);  // Lossless conversion.
+}
+
+template <typename To, typename From,
+          FMT_ENABLE_IF(std::is_same<From, To>::value)>
+FMT_CONSTEXPR auto lossless_integral_conversion(const From from, int& ec)
+    -> To {
+  ec = 0;
+  return from;
+}  // function
+
+// clang-format off
+/**
+ * converts From to To if possible, otherwise ec is set.
+ *
+ * input                            |    output
+ * ---------------------------------|---------------
+ * NaN                              | NaN
+ * Inf                              | Inf
+ * normal, fits in output           | converted (possibly lossy)
+ * normal, does not fit in output   | ec is set
+ * subnormal                        | best effort
+ * -Inf                             | -Inf
+ */
+// clang-format on
+template <typename To, typename From,
+          FMT_ENABLE_IF(!std::is_same<From, To>::value)>
+FMT_CONSTEXPR auto safe_float_conversion(const From from, int& ec) -> To {
+  ec = 0;
+  using T = std::numeric_limits<To>;
+  static_assert(std::is_floating_point<From>::value, "From must be floating");
+  static_assert(std::is_floating_point<To>::value, "To must be floating");
+
+  // catch the only happy case
+  if (std::isfinite(from)) {
+    if (from >= T::lowest() && from <= (T::max)()) {
+      return static_cast<To>(from);
+    }
+    // not within range.
+    ec = 1;
+    return {};
+  }
+
+  // nan and inf will be preserved
+  return static_cast<To>(from);
+}  // function
+
+template <typename To, typename From,
+          FMT_ENABLE_IF(std::is_same<From, To>::value)>
+FMT_CONSTEXPR auto safe_float_conversion(const From from, int& ec) -> To {
+  ec = 0;
+  static_assert(std::is_floating_point<From>::value, "From must be floating");
+  return from;
+}
+
+/// Safe duration cast between integral durations
+template <typename To, typename FromRep, typename FromPeriod,
+          FMT_ENABLE_IF(std::is_integral<FromRep>::value),
+          FMT_ENABLE_IF(std::is_integral<typename To::rep>::value)>
+auto safe_duration_cast(std::chrono::duration<FromRep, FromPeriod> from,
+                        int& ec) -> To {
+  using From = std::chrono::duration<FromRep, FromPeriod>;
+  ec = 0;
+  // the basic idea is that we need to convert from count() in the from type
+  // to count() in the To type, by multiplying it with this:
+  struct Factor
+      : std::ratio_divide<typename From::period, typename To::period> {};
+
+  static_assert(Factor::num > 0, "num must be positive");
+  static_assert(Factor::den > 0, "den must be positive");
+
+  // the conversion is like this: multiply from.count() with Factor::num
+  // /Factor::den and convert it to To::rep, all this without
+  // overflow/underflow. let's start by finding a suitable type that can hold
+  // both To, From and Factor::num
+  using IntermediateRep =
+      typename std::common_type<typename From::rep, typename To::rep,
+                                decltype(Factor::num)>::type;
+
+  // safe conversion to IntermediateRep
+  IntermediateRep count =
+      lossless_integral_conversion<IntermediateRep>(from.count(), ec);
+  if (ec) return {};
+  // multiply with Factor::num without overflow or underflow
+  if (detail::const_check(Factor::num != 1)) {
+    const auto max1 = detail::max_value<IntermediateRep>() / Factor::num;
+    if (count > max1) {
+      ec = 1;
+      return {};
+    }
+    const auto min1 =
+        (std::numeric_limits<IntermediateRep>::min)() / Factor::num;
+    if (detail::const_check(!std::is_unsigned<IntermediateRep>::value) &&
+        count < min1) {
+      ec = 1;
+      return {};
+    }
+    count *= Factor::num;
+  }
+
+  if (detail::const_check(Factor::den != 1)) count /= Factor::den;
+  auto tocount = lossless_integral_conversion<typename To::rep>(count, ec);
+  return ec ? To() : To(tocount);
+}
+
+/// Safe duration_cast between floating point durations
+template <typename To, typename FromRep, typename FromPeriod,
+          FMT_ENABLE_IF(std::is_floating_point<FromRep>::value),
+          FMT_ENABLE_IF(std::is_floating_point<typename To::rep>::value)>
+auto safe_duration_cast(std::chrono::duration<FromRep, FromPeriod> from,
+                        int& ec) -> To {
+  using From = std::chrono::duration<FromRep, FromPeriod>;
+  ec = 0;
+  if (std::isnan(from.count())) {
+    // nan in, gives nan out. easy.
+    return To{std::numeric_limits<typename To::rep>::quiet_NaN()};
+  }
+  // maybe we should also check if from is denormal, and decide what to do about
+  // it.
+
+  // +-inf should be preserved.
+  if (std::isinf(from.count())) {
+    return To{from.count()};
+  }
+
+  // the basic idea is that we need to convert from count() in the from type
+  // to count() in the To type, by multiplying it with this:
+  struct Factor
+      : std::ratio_divide<typename From::period, typename To::period> {};
+
+  static_assert(Factor::num > 0, "num must be positive");
+  static_assert(Factor::den > 0, "den must be positive");
+
+  // the conversion is like this: multiply from.count() with Factor::num
+  // /Factor::den and convert it to To::rep, all this without
+  // overflow/underflow. let's start by finding a suitable type that can hold
+  // both To, From and Factor::num
+  using IntermediateRep =
+      typename std::common_type<typename From::rep, typename To::rep,
+                                decltype(Factor::num)>::type;
+
+  // force conversion of From::rep -> IntermediateRep to be safe,
+  // even if it will never happen be narrowing in this context.
+  IntermediateRep count =
+      safe_float_conversion<IntermediateRep>(from.count(), ec);
+  if (ec) {
+    return {};
+  }
+
+  // multiply with Factor::num without overflow or underflow
+  if (detail::const_check(Factor::num != 1)) {
+    constexpr auto max1 = detail::max_value<IntermediateRep>() /
+                          static_cast<IntermediateRep>(Factor::num);
+    if (count > max1) {
+      ec = 1;
+      return {};
+    }
+    constexpr auto min1 = std::numeric_limits<IntermediateRep>::lowest() /
+                          static_cast<IntermediateRep>(Factor::num);
+    if (count < min1) {
+      ec = 1;
+      return {};
+    }
+    count *= static_cast<IntermediateRep>(Factor::num);
+  }
+
+  // this can't go wrong, right? den>0 is checked earlier.
+  if (detail::const_check(Factor::den != 1)) {
+    using common_t = typename std::common_type<IntermediateRep, intmax_t>::type;
+    count /= static_cast<common_t>(Factor::den);
+  }
+
+  // convert to the to type, safely
+  using ToRep = typename To::rep;
+
+  const ToRep tocount = safe_float_conversion<ToRep>(count, ec);
+  if (ec) {
+    return {};
+  }
+  return To{tocount};
+}
+}  // namespace safe_duration_cast
+#endif
+
+// Prevents expansion of a preceding token as a function-style macro.
+// Usage: f FMT_NOMACRO()
+#define FMT_NOMACRO
+
+namespace detail {
+template <typename T = void> struct null {};
+inline auto localtime_r FMT_NOMACRO(...) -> null<> { return null<>(); }
+inline auto localtime_s(...) -> null<> { return null<>(); }
+inline auto gmtime_r(...) -> null<> { return null<>(); }
+inline auto gmtime_s(...) -> null<> { return null<>(); }
+
+// It is defined here and not in ostream.h because the latter has expensive
+// includes.
+template <typename Streambuf> class formatbuf : public Streambuf {
+ private:
+  using char_type = typename Streambuf::char_type;
+  using streamsize = decltype(std::declval<Streambuf>().sputn(nullptr, 0));
+  using int_type = typename Streambuf::int_type;
+  using traits_type = typename Streambuf::traits_type;
+
+  buffer<char_type>& buffer_;
+
+ public:
+  explicit formatbuf(buffer<char_type>& buf) : buffer_(buf) {}
+
+ protected:
+  // The put area is always empty. This makes the implementation simpler and has
+  // the advantage that the streambuf and the buffer are always in sync and
+  // sputc never writes into uninitialized memory. A disadvantage is that each
+  // call to sputc always results in a (virtual) call to overflow. There is no
+  // disadvantage here for sputn since this always results in a call to xsputn.
+
+  auto overflow(int_type ch) -> int_type override {
+    if (!traits_type::eq_int_type(ch, traits_type::eof()))
+      buffer_.push_back(static_cast<char_type>(ch));
+    return ch;
+  }
+
+  auto xsputn(const char_type* s, streamsize count) -> streamsize override {
+    buffer_.append(s, s + count);
+    return count;
+  }
+};
+
+inline auto get_classic_locale() -> const std::locale& {
+  static const auto& locale = std::locale::classic();
+  return locale;
+}
+
+template <typename CodeUnit> struct codecvt_result {
+  static constexpr const size_t max_size = 32;
+  CodeUnit buf[max_size];
+  CodeUnit* end;
+};
+
+template <typename CodeUnit>
+void write_codecvt(codecvt_result<CodeUnit>& out, string_view in_buf,
+                   const std::locale& loc) {
+#if FMT_CLANG_VERSION
+#  pragma clang diagnostic push
+#  pragma clang diagnostic ignored "-Wdeprecated"
+  auto& f = std::use_facet<std::codecvt<CodeUnit, char, std::mbstate_t>>(loc);
+#  pragma clang diagnostic pop
+#else
+  auto& f = std::use_facet<std::codecvt<CodeUnit, char, std::mbstate_t>>(loc);
+#endif
+  auto mb = std::mbstate_t();
+  const char* from_next = nullptr;
+  auto result = f.in(mb, in_buf.begin(), in_buf.end(), from_next,
+                     std::begin(out.buf), std::end(out.buf), out.end);
+  if (result != std::codecvt_base::ok)
+    FMT_THROW(format_error("failed to format time"));
+}
+
+template <typename OutputIt>
+auto write_encoded_tm_str(OutputIt out, string_view in, const std::locale& loc)
+    -> OutputIt {
+  if (detail::use_utf8() && loc != get_classic_locale()) {
+    // char16_t and char32_t codecvts are broken in MSVC (linkage errors) and
+    // gcc-4.
+#if FMT_MSC_VERSION != 0 ||  \
+    (defined(__GLIBCXX__) && \
+     (!defined(_GLIBCXX_USE_DUAL_ABI) || _GLIBCXX_USE_DUAL_ABI == 0))
+    // The _GLIBCXX_USE_DUAL_ABI macro is always defined in libstdc++ from gcc-5
+    // and newer.
+    using code_unit = wchar_t;
+#else
+    using code_unit = char32_t;
+#endif
+
+    using unit_t = codecvt_result<code_unit>;
+    unit_t unit;
+    write_codecvt(unit, in, loc);
+    // In UTF-8 is used one to four one-byte code units.
+    auto u =
+        to_utf8<code_unit, basic_memory_buffer<char, unit_t::max_size * 4>>();
+    if (!u.convert({unit.buf, to_unsigned(unit.end - unit.buf)}))
+      FMT_THROW(format_error("failed to format time"));
+    return copy<char>(u.c_str(), u.c_str() + u.size(), out);
+  }
+  return copy<char>(in.data(), in.data() + in.size(), out);
+}
+
+template <typename Char, typename OutputIt,
+          FMT_ENABLE_IF(!std::is_same<Char, char>::value)>
+auto write_tm_str(OutputIt out, string_view sv, const std::locale& loc)
+    -> OutputIt {
+  codecvt_result<Char> unit;
+  write_codecvt(unit, sv, loc);
+  return copy<Char>(unit.buf, unit.end, out);
+}
+
+template <typename Char, typename OutputIt,
+          FMT_ENABLE_IF(std::is_same<Char, char>::value)>
+auto write_tm_str(OutputIt out, string_view sv, const std::locale& loc)
+    -> OutputIt {
+  return write_encoded_tm_str(out, sv, loc);
+}
+
+template <typename Char>
+inline void do_write(buffer<Char>& buf, const std::tm& time,
+                     const std::locale& loc, char format, char modifier) {
+  auto&& format_buf = formatbuf<std::basic_streambuf<Char>>(buf);
+  auto&& os = std::basic_ostream<Char>(&format_buf);
+  os.imbue(loc);
+  const auto& facet = std::use_facet<std::time_put<Char>>(loc);
+  auto end = facet.put(os, os, Char(' '), &time, format, modifier);
+  if (end.failed()) FMT_THROW(format_error("failed to format time"));
+}
+
+template <typename Char, typename OutputIt,
+          FMT_ENABLE_IF(!std::is_same<Char, char>::value)>
+auto write(OutputIt out, const std::tm& time, const std::locale& loc,
+           char format, char modifier = 0) -> OutputIt {
+  auto&& buf = get_buffer<Char>(out);
+  do_write<Char>(buf, time, loc, format, modifier);
+  return get_iterator(buf, out);
+}
+
+template <typename Char, typename OutputIt,
+          FMT_ENABLE_IF(std::is_same<Char, char>::value)>
+auto write(OutputIt out, const std::tm& time, const std::locale& loc,
+           char format, char modifier = 0) -> OutputIt {
+  auto&& buf = basic_memory_buffer<Char>();
+  do_write<char>(buf, time, loc, format, modifier);
+  return write_encoded_tm_str(out, string_view(buf.data(), buf.size()), loc);
+}
+
+template <typename Rep1, typename Rep2>
+struct is_same_arithmetic_type
+    : public std::integral_constant<bool,
+                                    (std::is_integral<Rep1>::value &&
+                                     std::is_integral<Rep2>::value) ||
+                                        (std::is_floating_point<Rep1>::value &&
+                                         std::is_floating_point<Rep2>::value)> {
+};
+
+template <
+    typename To, typename FromRep, typename FromPeriod,
+    FMT_ENABLE_IF(is_same_arithmetic_type<FromRep, typename To::rep>::value)>
+auto fmt_duration_cast(std::chrono::duration<FromRep, FromPeriod> from) -> To {
+#if FMT_SAFE_DURATION_CAST
+  // Throwing version of safe_duration_cast is only available for
+  // integer to integer or float to float casts.
+  int ec;
+  To to = safe_duration_cast::safe_duration_cast<To>(from, ec);
+  if (ec) FMT_THROW(format_error("cannot format duration"));
+  return to;
+#else
+  // Standard duration cast, may overflow.
+  return std::chrono::duration_cast<To>(from);
+#endif
+}
+
+template <
+    typename To, typename FromRep, typename FromPeriod,
+    FMT_ENABLE_IF(!is_same_arithmetic_type<FromRep, typename To::rep>::value)>
+auto fmt_duration_cast(std::chrono::duration<FromRep, FromPeriod> from) -> To {
+  // Mixed integer <-> float cast is not supported by safe_duration_cast.
+  return std::chrono::duration_cast<To>(from);
+}
+
+template <typename Duration>
+auto to_time_t(
+    std::chrono::time_point<std::chrono::system_clock, Duration> time_point)
+    -> std::time_t {
+  // Cannot use std::chrono::system_clock::to_time_t since this would first
+  // require a cast to std::chrono::system_clock::time_point, which could
+  // overflow.
+  return fmt_duration_cast<std::chrono::duration<std::time_t>>(
+             time_point.time_since_epoch())
+      .count();
+}
+}  // namespace detail
+
+FMT_BEGIN_EXPORT
+
+/**
+ * Converts given time since epoch as `std::time_t` value into calendar time,
+ * expressed in local time. Unlike `std::localtime`, this function is
+ * thread-safe on most platforms.
+ */
+inline auto localtime(std::time_t time) -> std::tm {
+  struct dispatcher {
+    std::time_t time_;
+    std::tm tm_;
+
+    dispatcher(std::time_t t) : time_(t) {}
+
+    auto run() -> bool {
+      using namespace fmt::detail;
+      return handle(localtime_r(&time_, &tm_));
+    }
+
+    auto handle(std::tm* tm) -> bool { return tm != nullptr; }
+
+    auto handle(detail::null<>) -> bool {
+      using namespace fmt::detail;
+      return fallback(localtime_s(&tm_, &time_));
+    }
+
+    auto fallback(int res) -> bool { return res == 0; }
+
+#if !FMT_MSC_VERSION
+    auto fallback(detail::null<>) -> bool {
+      using namespace fmt::detail;
+      std::tm* tm = std::localtime(&time_);
+      if (tm) tm_ = *tm;
+      return tm != nullptr;
+    }
+#endif
+  };
+  dispatcher lt(time);
+  // Too big time values may be unsupported.
+  if (!lt.run()) FMT_THROW(format_error("time_t value out of range"));
+  return lt.tm_;
+}
+
+#if FMT_USE_LOCAL_TIME
+template <typename Duration>
+inline auto localtime(std::chrono::local_time<Duration> time) -> std::tm {
+  return localtime(
+      detail::to_time_t(std::chrono::current_zone()->to_sys(time)));
+}
+#endif
+
+/**
+ * Converts given time since epoch as `std::time_t` value into calendar time,
+ * expressed in Coordinated Universal Time (UTC). Unlike `std::gmtime`, this
+ * function is thread-safe on most platforms.
+ */
+inline auto gmtime(std::time_t time) -> std::tm {
+  struct dispatcher {
+    std::time_t time_;
+    std::tm tm_;
+
+    dispatcher(std::time_t t) : time_(t) {}
+
+    auto run() -> bool {
+      using namespace fmt::detail;
+      return handle(gmtime_r(&time_, &tm_));
+    }
+
+    auto handle(std::tm* tm) -> bool { return tm != nullptr; }
+
+    auto handle(detail::null<>) -> bool {
+      using namespace fmt::detail;
+      return fallback(gmtime_s(&tm_, &time_));
+    }
+
+    auto fallback(int res) -> bool { return res == 0; }
+
+#if !FMT_MSC_VERSION
+    auto fallback(detail::null<>) -> bool {
+      std::tm* tm = std::gmtime(&time_);
+      if (tm) tm_ = *tm;
+      return tm != nullptr;
+    }
+#endif
+  };
+  auto gt = dispatcher(time);
+  // Too big time values may be unsupported.
+  if (!gt.run()) FMT_THROW(format_error("time_t value out of range"));
+  return gt.tm_;
+}
+
+template <typename Duration>
+inline auto gmtime(
+    std::chrono::time_point<std::chrono::system_clock, Duration> time_point)
+    -> std::tm {
+  return gmtime(detail::to_time_t(time_point));
+}
+
+namespace detail {
+
+// Writes two-digit numbers a, b and c separated by sep to buf.
+// The method by Pavel Novikov based on
+// https://johnnylee-sde.github.io/Fast-unsigned-integer-to-time-string/.
+inline void write_digit2_separated(char* buf, unsigned a, unsigned b,
+                                   unsigned c, char sep) {
+  unsigned long long digits =
+      a | (b << 24) | (static_cast<unsigned long long>(c) << 48);
+  // Convert each value to BCD.
+  // We have x = a * 10 + b and we want to convert it to BCD y = a * 16 + b.
+  // The difference is
+  //   y - x = a * 6
+  // a can be found from x:
+  //   a = floor(x / 10)
+  // then
+  //   y = x + a * 6 = x + floor(x / 10) * 6
+  // floor(x / 10) is (x * 205) >> 11 (needs 16 bits).
+  digits += (((digits * 205) >> 11) & 0x000f00000f00000f) * 6;
+  // Put low nibbles to high bytes and high nibbles to low bytes.
+  digits = ((digits & 0x00f00000f00000f0) >> 4) |
+           ((digits & 0x000f00000f00000f) << 8);
+  auto usep = static_cast<unsigned long long>(sep);
+  // Add ASCII '0' to each digit byte and insert separators.
+  digits |= 0x3030003030003030 | (usep << 16) | (usep << 40);
+
+  constexpr const size_t len = 8;
+  if (const_check(is_big_endian())) {
+    char tmp[len];
+    std::memcpy(tmp, &digits, len);
+    std::reverse_copy(tmp, tmp + len, buf);
+  } else {
+    std::memcpy(buf, &digits, len);
+  }
+}
+
+template <typename Period>
+FMT_CONSTEXPR inline auto get_units() -> const char* {
+  if (std::is_same<Period, std::atto>::value) return "as";
+  if (std::is_same<Period, std::femto>::value) return "fs";
+  if (std::is_same<Period, std::pico>::value) return "ps";
+  if (std::is_same<Period, std::nano>::value) return "ns";
+  if (std::is_same<Period, std::micro>::value) return "µs";
+  if (std::is_same<Period, std::milli>::value) return "ms";
+  if (std::is_same<Period, std::centi>::value) return "cs";
+  if (std::is_same<Period, std::deci>::value) return "ds";
+  if (std::is_same<Period, std::ratio<1>>::value) return "s";
+  if (std::is_same<Period, std::deca>::value) return "das";
+  if (std::is_same<Period, std::hecto>::value) return "hs";
+  if (std::is_same<Period, std::kilo>::value) return "ks";
+  if (std::is_same<Period, std::mega>::value) return "Ms";
+  if (std::is_same<Period, std::giga>::value) return "Gs";
+  if (std::is_same<Period, std::tera>::value) return "Ts";
+  if (std::is_same<Period, std::peta>::value) return "Ps";
+  if (std::is_same<Period, std::exa>::value) return "Es";
+  if (std::is_same<Period, std::ratio<60>>::value) return "min";
+  if (std::is_same<Period, std::ratio<3600>>::value) return "h";
+  if (std::is_same<Period, std::ratio<86400>>::value) return "d";
+  return nullptr;
+}
+
+enum class numeric_system {
+  standard,
+  // Alternative numeric system, e.g. 十二 instead of 12 in ja_JP locale.
+  alternative
+};
+
+// Glibc extensions for formatting numeric values.
+enum class pad_type {
+  // Pad a numeric result string with zeros (the default).
+  zero,
+  // Do not pad a numeric result string.
+  none,
+  // Pad a numeric result string with spaces.
+  space,
+};
+
+template <typename OutputIt>
+auto write_padding(OutputIt out, pad_type pad, int width) -> OutputIt {
+  if (pad == pad_type::none) return out;
+  return detail::fill_n(out, width, pad == pad_type::space ? ' ' : '0');
+}
+
+template <typename OutputIt>
+auto write_padding(OutputIt out, pad_type pad) -> OutputIt {
+  if (pad != pad_type::none) *out++ = pad == pad_type::space ? ' ' : '0';
+  return out;
+}
+
+// Parses a put_time-like format string and invokes handler actions.
+template <typename Char, typename Handler>
+FMT_CONSTEXPR auto parse_chrono_format(const Char* begin, const Char* end,
+                                       Handler&& handler) -> const Char* {
+  if (begin == end || *begin == '}') return begin;
+  if (*begin != '%') FMT_THROW(format_error("invalid format"));
+  auto ptr = begin;
+  while (ptr != end) {
+    pad_type pad = pad_type::zero;
+    auto c = *ptr;
+    if (c == '}') break;
+    if (c != '%') {
+      ++ptr;
+      continue;
+    }
+    if (begin != ptr) handler.on_text(begin, ptr);
+    ++ptr;  // consume '%'
+    if (ptr == end) FMT_THROW(format_error("invalid format"));
+    c = *ptr;
+    switch (c) {
+    case '_':
+      pad = pad_type::space;
+      ++ptr;
+      break;
+    case '-':
+      pad = pad_type::none;
+      ++ptr;
+      break;
+    }
+    if (ptr == end) FMT_THROW(format_error("invalid format"));
+    c = *ptr++;
+    switch (c) {
+    case '%':
+      handler.on_text(ptr - 1, ptr);
+      break;
+    case 'n': {
+      const Char newline[] = {'\n'};
+      handler.on_text(newline, newline + 1);
+      break;
+    }
+    case 't': {
+      const Char tab[] = {'\t'};
+      handler.on_text(tab, tab + 1);
+      break;
+    }
+    // Year:
+    case 'Y':
+      handler.on_year(numeric_system::standard);
+      break;
+    case 'y':
+      handler.on_short_year(numeric_system::standard);
+      break;
+    case 'C':
+      handler.on_century(numeric_system::standard);
+      break;
+    case 'G':
+      handler.on_iso_week_based_year();
+      break;
+    case 'g':
+      handler.on_iso_week_based_short_year();
+      break;
+    // Day of the week:
+    case 'a':
+      handler.on_abbr_weekday();
+      break;
+    case 'A':
+      handler.on_full_weekday();
+      break;
+    case 'w':
+      handler.on_dec0_weekday(numeric_system::standard);
+      break;
+    case 'u':
+      handler.on_dec1_weekday(numeric_system::standard);
+      break;
+    // Month:
+    case 'b':
+    case 'h':
+      handler.on_abbr_month();
+      break;
+    case 'B':
+      handler.on_full_month();
+      break;
+    case 'm':
+      handler.on_dec_month(numeric_system::standard);
+      break;
+    // Day of the year/month:
+    case 'U':
+      handler.on_dec0_week_of_year(numeric_system::standard, pad);
+      break;
+    case 'W':
+      handler.on_dec1_week_of_year(numeric_system::standard, pad);
+      break;
+    case 'V':
+      handler.on_iso_week_of_year(numeric_system::standard, pad);
+      break;
+    case 'j':
+      handler.on_day_of_year();
+      break;
+    case 'd':
+      handler.on_day_of_month(numeric_system::standard, pad);
+      break;
+    case 'e':
+      handler.on_day_of_month(numeric_system::standard, pad_type::space);
+      break;
+    // Hour, minute, second:
+    case 'H':
+      handler.on_24_hour(numeric_system::standard, pad);
+      break;
+    case 'I':
+      handler.on_12_hour(numeric_system::standard, pad);
+      break;
+    case 'M':
+      handler.on_minute(numeric_system::standard, pad);
+      break;
+    case 'S':
+      handler.on_second(numeric_system::standard, pad);
+      break;
+    // Other:
+    case 'c':
+      handler.on_datetime(numeric_system::standard);
+      break;
+    case 'x':
+      handler.on_loc_date(numeric_system::standard);
+      break;
+    case 'X':
+      handler.on_loc_time(numeric_system::standard);
+      break;
+    case 'D':
+      handler.on_us_date();
+      break;
+    case 'F':
+      handler.on_iso_date();
+      break;
+    case 'r':
+      handler.on_12_hour_time();
+      break;
+    case 'R':
+      handler.on_24_hour_time();
+      break;
+    case 'T':
+      handler.on_iso_time();
+      break;
+    case 'p':
+      handler.on_am_pm();
+      break;
+    case 'Q':
+      handler.on_duration_value();
+      break;
+    case 'q':
+      handler.on_duration_unit();
+      break;
+    case 'z':
+      handler.on_utc_offset(numeric_system::standard);
+      break;
+    case 'Z':
+      handler.on_tz_name();
+      break;
+    // Alternative representation:
+    case 'E': {
+      if (ptr == end) FMT_THROW(format_error("invalid format"));
+      c = *ptr++;
+      switch (c) {
+      case 'Y':
+        handler.on_year(numeric_system::alternative);
+        break;
+      case 'y':
+        handler.on_offset_year();
+        break;
+      case 'C':
+        handler.on_century(numeric_system::alternative);
+        break;
+      case 'c':
+        handler.on_datetime(numeric_system::alternative);
+        break;
+      case 'x':
+        handler.on_loc_date(numeric_system::alternative);
+        break;
+      case 'X':
+        handler.on_loc_time(numeric_system::alternative);
+        break;
+      case 'z':
+        handler.on_utc_offset(numeric_system::alternative);
+        break;
+      default:
+        FMT_THROW(format_error("invalid format"));
+      }
+      break;
+    }
+    case 'O':
+      if (ptr == end) FMT_THROW(format_error("invalid format"));
+      c = *ptr++;
+      switch (c) {
+      case 'y':
+        handler.on_short_year(numeric_system::alternative);
+        break;
+      case 'm':
+        handler.on_dec_month(numeric_system::alternative);
+        break;
+      case 'U':
+        handler.on_dec0_week_of_year(numeric_system::alternative, pad);
+        break;
+      case 'W':
+        handler.on_dec1_week_of_year(numeric_system::alternative, pad);
+        break;
+      case 'V':
+        handler.on_iso_week_of_year(numeric_system::alternative, pad);
+        break;
+      case 'd':
+        handler.on_day_of_month(numeric_system::alternative, pad);
+        break;
+      case 'e':
+        handler.on_day_of_month(numeric_system::alternative, pad_type::space);
+        break;
+      case 'w':
+        handler.on_dec0_weekday(numeric_system::alternative);
+        break;
+      case 'u':
+        handler.on_dec1_weekday(numeric_system::alternative);
+        break;
+      case 'H':
+        handler.on_24_hour(numeric_system::alternative, pad);
+        break;
+      case 'I':
+        handler.on_12_hour(numeric_system::alternative, pad);
+        break;
+      case 'M':
+        handler.on_minute(numeric_system::alternative, pad);
+        break;
+      case 'S':
+        handler.on_second(numeric_system::alternative, pad);
+        break;
+      case 'z':
+        handler.on_utc_offset(numeric_system::alternative);
+        break;
+      default:
+        FMT_THROW(format_error("invalid format"));
+      }
+      break;
+    default:
+      FMT_THROW(format_error("invalid format"));
+    }
+    begin = ptr;
+  }
+  if (begin != ptr) handler.on_text(begin, ptr);
+  return ptr;
+}
+
+template <typename Derived> struct null_chrono_spec_handler {
+  FMT_CONSTEXPR void unsupported() {
+    static_cast<Derived*>(this)->unsupported();
+  }
+  FMT_CONSTEXPR void on_year(numeric_system) { unsupported(); }
+  FMT_CONSTEXPR void on_short_year(numeric_system) { unsupported(); }
+  FMT_CONSTEXPR void on_offset_year() { unsupported(); }
+  FMT_CONSTEXPR void on_century(numeric_system) { unsupported(); }
+  FMT_CONSTEXPR void on_iso_week_based_year() { unsupported(); }
+  FMT_CONSTEXPR void on_iso_week_based_short_year() { unsupported(); }
+  FMT_CONSTEXPR void on_abbr_weekday() { unsupported(); }
+  FMT_CONSTEXPR void on_full_weekday() { unsupported(); }
+  FMT_CONSTEXPR void on_dec0_weekday(numeric_system) { unsupported(); }
+  FMT_CONSTEXPR void on_dec1_weekday(numeric_system) { unsupported(); }
+  FMT_CONSTEXPR void on_abbr_month() { unsupported(); }
+  FMT_CONSTEXPR void on_full_month() { unsupported(); }
+  FMT_CONSTEXPR void on_dec_month(numeric_system) { unsupported(); }
+  FMT_CONSTEXPR void on_dec0_week_of_year(numeric_system, pad_type) {
+    unsupported();
+  }
+  FMT_CONSTEXPR void on_dec1_week_of_year(numeric_system, pad_type) {
+    unsupported();
+  }
+  FMT_CONSTEXPR void on_iso_week_of_year(numeric_system, pad_type) {
+    unsupported();
+  }
+  FMT_CONSTEXPR void on_day_of_year() { unsupported(); }
+  FMT_CONSTEXPR void on_day_of_month(numeric_system, pad_type) {
+    unsupported();
+  }
+  FMT_CONSTEXPR void on_24_hour(numeric_system) { unsupported(); }
+  FMT_CONSTEXPR void on_12_hour(numeric_system) { unsupported(); }
+  FMT_CONSTEXPR void on_minute(numeric_system) { unsupported(); }
+  FMT_CONSTEXPR void on_second(numeric_system) { unsupported(); }
+  FMT_CONSTEXPR void on_datetime(numeric_system) { unsupported(); }
+  FMT_CONSTEXPR void on_loc_date(numeric_system) { unsupported(); }
+  FMT_CONSTEXPR void on_loc_time(numeric_system) { unsupported(); }
+  FMT_CONSTEXPR void on_us_date() { unsupported(); }
+  FMT_CONSTEXPR void on_iso_date() { unsupported(); }
+  FMT_CONSTEXPR void on_12_hour_time() { unsupported(); }
+  FMT_CONSTEXPR void on_24_hour_time() { unsupported(); }
+  FMT_CONSTEXPR void on_iso_time() { unsupported(); }
+  FMT_CONSTEXPR void on_am_pm() { unsupported(); }
+  FMT_CONSTEXPR void on_duration_value() { unsupported(); }
+  FMT_CONSTEXPR void on_duration_unit() { unsupported(); }
+  FMT_CONSTEXPR void on_utc_offset(numeric_system) { unsupported(); }
+  FMT_CONSTEXPR void on_tz_name() { unsupported(); }
+};
+
+struct tm_format_checker : null_chrono_spec_handler<tm_format_checker> {
+  FMT_NORETURN void unsupported() { FMT_THROW(format_error("no format")); }
+
+  template <typename Char>
+  FMT_CONSTEXPR void on_text(const Char*, const Char*) {}
+  FMT_CONSTEXPR void on_year(numeric_system) {}
+  FMT_CONSTEXPR void on_short_year(numeric_system) {}
+  FMT_CONSTEXPR void on_offset_year() {}
+  FMT_CONSTEXPR void on_century(numeric_system) {}
+  FMT_CONSTEXPR void on_iso_week_based_year() {}
+  FMT_CONSTEXPR void on_iso_week_based_short_year() {}
+  FMT_CONSTEXPR void on_abbr_weekday() {}
+  FMT_CONSTEXPR void on_full_weekday() {}
+  FMT_CONSTEXPR void on_dec0_weekday(numeric_system) {}
+  FMT_CONSTEXPR void on_dec1_weekday(numeric_system) {}
+  FMT_CONSTEXPR void on_abbr_month() {}
+  FMT_CONSTEXPR void on_full_month() {}
+  FMT_CONSTEXPR void on_dec_month(numeric_system) {}
+  FMT_CONSTEXPR void on_dec0_week_of_year(numeric_system, pad_type) {}
+  FMT_CONSTEXPR void on_dec1_week_of_year(numeric_system, pad_type) {}
+  FMT_CONSTEXPR void on_iso_week_of_year(numeric_system, pad_type) {}
+  FMT_CONSTEXPR void on_day_of_year() {}
+  FMT_CONSTEXPR void on_day_of_month(numeric_system, pad_type) {}
+  FMT_CONSTEXPR void on_24_hour(numeric_system, pad_type) {}
+  FMT_CONSTEXPR void on_12_hour(numeric_system, pad_type) {}
+  FMT_CONSTEXPR void on_minute(numeric_system, pad_type) {}
+  FMT_CONSTEXPR void on_second(numeric_system, pad_type) {}
+  FMT_CONSTEXPR void on_datetime(numeric_system) {}
+  FMT_CONSTEXPR void on_loc_date(numeric_system) {}
+  FMT_CONSTEXPR void on_loc_time(numeric_system) {}
+  FMT_CONSTEXPR void on_us_date() {}
+  FMT_CONSTEXPR void on_iso_date() {}
+  FMT_CONSTEXPR void on_12_hour_time() {}
+  FMT_CONSTEXPR void on_24_hour_time() {}
+  FMT_CONSTEXPR void on_iso_time() {}
+  FMT_CONSTEXPR void on_am_pm() {}
+  FMT_CONSTEXPR void on_utc_offset(numeric_system) {}
+  FMT_CONSTEXPR void on_tz_name() {}
+};
+
+inline auto tm_wday_full_name(int wday) -> const char* {
+  static constexpr const char* full_name_list[] = {
+      "Sunday",   "Monday", "Tuesday", "Wednesday",
+      "Thursday", "Friday", "Saturday"};
+  return wday >= 0 && wday <= 6 ? full_name_list[wday] : "?";
+}
+inline auto tm_wday_short_name(int wday) -> const char* {
+  static constexpr const char* short_name_list[] = {"Sun", "Mon", "Tue", "Wed",
+                                                    "Thu", "Fri", "Sat"};
+  return wday >= 0 && wday <= 6 ? short_name_list[wday] : "???";
+}
+
+inline auto tm_mon_full_name(int mon) -> const char* {
+  static constexpr const char* full_name_list[] = {
+      "January", "February", "March",     "April",   "May",      "June",
+      "July",    "August",   "September", "October", "November", "December"};
+  return mon >= 0 && mon <= 11 ? full_name_list[mon] : "?";
+}
+inline auto tm_mon_short_name(int mon) -> const char* {
+  static constexpr const char* short_name_list[] = {
+      "Jan", "Feb", "Mar", "Apr", "May", "Jun",
+      "Jul", "Aug", "Sep", "Oct", "Nov", "Dec",
+  };
+  return mon >= 0 && mon <= 11 ? short_name_list[mon] : "???";
+}
+
+template <typename T, typename = void>
+struct has_member_data_tm_gmtoff : std::false_type {};
+template <typename T>
+struct has_member_data_tm_gmtoff<T, void_t<decltype(T::tm_gmtoff)>>
+    : std::true_type {};
+
+template <typename T, typename = void>
+struct has_member_data_tm_zone : std::false_type {};
+template <typename T>
+struct has_member_data_tm_zone<T, void_t<decltype(T::tm_zone)>>
+    : std::true_type {};
+
+#if FMT_USE_TZSET
+inline void tzset_once() {
+  static bool init = []() -> bool {
+    _tzset();
+    return true;
+  }();
+  ignore_unused(init);
+}
+#endif
+
+// Converts value to Int and checks that it's in the range [0, upper).
+template <typename T, typename Int, FMT_ENABLE_IF(std::is_integral<T>::value)>
+inline auto to_nonnegative_int(T value, Int upper) -> Int {
+  if (!std::is_unsigned<Int>::value &&
+      (value < 0 || to_unsigned(value) > to_unsigned(upper))) {
+    FMT_THROW(fmt::format_error("chrono value is out of range"));
+  }
+  return static_cast<Int>(value);
+}
+template <typename T, typename Int, FMT_ENABLE_IF(!std::is_integral<T>::value)>
+inline auto to_nonnegative_int(T value, Int upper) -> Int {
+  auto int_value = static_cast<Int>(value);
+  if (int_value < 0 || value > static_cast<T>(upper))
+    FMT_THROW(format_error("invalid value"));
+  return int_value;
+}
+
+constexpr auto pow10(std::uint32_t n) -> long long {
+  return n == 0 ? 1 : 10 * pow10(n - 1);
+}
+
+// Counts the number of fractional digits in the range [0, 18] according to the
+// C++20 spec. If more than 18 fractional digits are required then returns 6 for
+// microseconds precision.
+template <long long Num, long long Den, int N = 0,
+          bool Enabled = (N < 19) && (Num <= max_value<long long>() / 10)>
+struct count_fractional_digits {
+  static constexpr int value =
+      Num % Den == 0 ? N : count_fractional_digits<Num * 10, Den, N + 1>::value;
+};
+
+// Base case that doesn't instantiate any more templates
+// in order to avoid overflow.
+template <long long Num, long long Den, int N>
+struct count_fractional_digits<Num, Den, N, false> {
+  static constexpr int value = (Num % Den == 0) ? N : 6;
+};
+
+// Format subseconds which are given as an integer type with an appropriate
+// number of digits.
+template <typename Char, typename OutputIt, typename Duration>
+void write_fractional_seconds(OutputIt& out, Duration d, int precision = -1) {
+  constexpr auto num_fractional_digits =
+      count_fractional_digits<Duration::period::num,
+                              Duration::period::den>::value;
+
+  using subsecond_precision = std::chrono::duration<
+      typename std::common_type<typename Duration::rep,
+                                std::chrono::seconds::rep>::type,
+      std::ratio<1, detail::pow10(num_fractional_digits)>>;
+
+  const auto fractional = d - fmt_duration_cast<std::chrono::seconds>(d);
+  const auto subseconds =
+      std::chrono::treat_as_floating_point<
+          typename subsecond_precision::rep>::value
+          ? fractional.count()
+          : fmt_duration_cast<subsecond_precision>(fractional).count();
+  auto n = static_cast<uint32_or_64_or_128_t<long long>>(subseconds);
+  const int num_digits = detail::count_digits(n);
+
+  int leading_zeroes = (std::max)(0, num_fractional_digits - num_digits);
+  if (precision < 0) {
+    FMT_ASSERT(!std::is_floating_point<typename Duration::rep>::value, "");
+    if (std::ratio_less<typename subsecond_precision::period,
+                        std::chrono::seconds::period>::value) {
+      *out++ = '.';
+      out = detail::fill_n(out, leading_zeroes, '0');
+      out = format_decimal<Char>(out, n, num_digits).end;
+    }
+  } else if (precision > 0) {
+    *out++ = '.';
+    leading_zeroes = (std::min)(leading_zeroes, precision);
+    int remaining = precision - leading_zeroes;
+    out = detail::fill_n(out, leading_zeroes, '0');
+    if (remaining < num_digits) {
+      int num_truncated_digits = num_digits - remaining;
+      n /= to_unsigned(detail::pow10(to_unsigned(num_truncated_digits)));
+      if (n) {
+        out = format_decimal<Char>(out, n, remaining).end;
+      }
+      return;
+    }
+    if (n) {
+      out = format_decimal<Char>(out, n, num_digits).end;
+      remaining -= num_digits;
+    }
+    out = detail::fill_n(out, remaining, '0');
+  }
+}
+
+// Format subseconds which are given as a floating point type with an
+// appropriate number of digits. We cannot pass the Duration here, as we
+// explicitly need to pass the Rep value in the chrono_formatter.
+template <typename Duration>
+void write_floating_seconds(memory_buffer& buf, Duration duration,
+                            int num_fractional_digits = -1) {
+  using rep = typename Duration::rep;
+  FMT_ASSERT(std::is_floating_point<rep>::value, "");
+
+  auto val = duration.count();
+
+  if (num_fractional_digits < 0) {
+    // For `std::round` with fallback to `round`:
+    // On some toolchains `std::round` is not available (e.g. GCC 6).
+    using namespace std;
+    num_fractional_digits =
+        count_fractional_digits<Duration::period::num,
+                                Duration::period::den>::value;
+    if (num_fractional_digits < 6 && static_cast<rep>(round(val)) != val)
+      num_fractional_digits = 6;
+  }
+
+  fmt::format_to(std::back_inserter(buf), FMT_STRING("{:.{}f}"),
+                 std::fmod(val * static_cast<rep>(Duration::period::num) /
+                               static_cast<rep>(Duration::period::den),
+                           static_cast<rep>(60)),
+                 num_fractional_digits);
+}
+
+template <typename OutputIt, typename Char,
+          typename Duration = std::chrono::seconds>
+class tm_writer {
+ private:
+  static constexpr int days_per_week = 7;
+
+  const std::locale& loc_;
+  const bool is_classic_;
+  OutputIt out_;
+  const Duration* subsecs_;
+  const std::tm& tm_;
+
+  auto tm_sec() const noexcept -> int {
+    FMT_ASSERT(tm_.tm_sec >= 0 && tm_.tm_sec <= 61, "");
+    return tm_.tm_sec;
+  }
+  auto tm_min() const noexcept -> int {
+    FMT_ASSERT(tm_.tm_min >= 0 && tm_.tm_min <= 59, "");
+    return tm_.tm_min;
+  }
+  auto tm_hour() const noexcept -> int {
+    FMT_ASSERT(tm_.tm_hour >= 0 && tm_.tm_hour <= 23, "");
+    return tm_.tm_hour;
+  }
+  auto tm_mday() const noexcept -> int {
+    FMT_ASSERT(tm_.tm_mday >= 1 && tm_.tm_mday <= 31, "");
+    return tm_.tm_mday;
+  }
+  auto tm_mon() const noexcept -> int {
+    FMT_ASSERT(tm_.tm_mon >= 0 && tm_.tm_mon <= 11, "");
+    return tm_.tm_mon;
+  }
+  auto tm_year() const noexcept -> long long { return 1900ll + tm_.tm_year; }
+  auto tm_wday() const noexcept -> int {
+    FMT_ASSERT(tm_.tm_wday >= 0 && tm_.tm_wday <= 6, "");
+    return tm_.tm_wday;
+  }
+  auto tm_yday() const noexcept -> int {
+    FMT_ASSERT(tm_.tm_yday >= 0 && tm_.tm_yday <= 365, "");
+    return tm_.tm_yday;
+  }
+
+  auto tm_hour12() const noexcept -> int {
+    const auto h = tm_hour();
+    const auto z = h < 12 ? h : h - 12;
+    return z == 0 ? 12 : z;
+  }
+
+  // POSIX and the C Standard are unclear or inconsistent about what %C and %y
+  // do if the year is negative or exceeds 9999. Use the convention that %C
+  // concatenated with %y yields the same output as %Y, and that %Y contains at
+  // least 4 characters, with more only if necessary.
+  auto split_year_lower(long long year) const noexcept -> int {
+    auto l = year % 100;
+    if (l < 0) l = -l;  // l in [0, 99]
+    return static_cast<int>(l);
+  }
+
+  // Algorithm: https://en.wikipedia.org/wiki/ISO_week_date.
+  auto iso_year_weeks(long long curr_year) const noexcept -> int {
+    const auto prev_year = curr_year - 1;
+    const auto curr_p =
+        (curr_year + curr_year / 4 - curr_year / 100 + curr_year / 400) %
+        days_per_week;
+    const auto prev_p =
+        (prev_year + prev_year / 4 - prev_year / 100 + prev_year / 400) %
+        days_per_week;
+    return 52 + ((curr_p == 4 || prev_p == 3) ? 1 : 0);
+  }
+  auto iso_week_num(int tm_yday, int tm_wday) const noexcept -> int {
+    return (tm_yday + 11 - (tm_wday == 0 ? days_per_week : tm_wday)) /
+           days_per_week;
+  }
+  auto tm_iso_week_year() const noexcept -> long long {
+    const auto year = tm_year();
+    const auto w = iso_week_num(tm_yday(), tm_wday());
+    if (w < 1) return year - 1;
+    if (w > iso_year_weeks(year)) return year + 1;
+    return year;
+  }
+  auto tm_iso_week_of_year() const noexcept -> int {
+    const auto year = tm_year();
+    const auto w = iso_week_num(tm_yday(), tm_wday());
+    if (w < 1) return iso_year_weeks(year - 1);
+    if (w > iso_year_weeks(year)) return 1;
+    return w;
+  }
+
+  void write1(int value) {
+    *out_++ = static_cast<char>('0' + to_unsigned(value) % 10);
+  }
+  void write2(int value) {
+    const char* d = digits2(to_unsigned(value) % 100);
+    *out_++ = *d++;
+    *out_++ = *d;
+  }
+  void write2(int value, pad_type pad) {
+    unsigned int v = to_unsigned(value) % 100;
+    if (v >= 10) {
+      const char* d = digits2(v);
+      *out_++ = *d++;
+      *out_++ = *d;
+    } else {
+      out_ = detail::write_padding(out_, pad);
+      *out_++ = static_cast<char>('0' + v);
+    }
+  }
+
+  void write_year_extended(long long year) {
+    // At least 4 characters.
+    int width = 4;
+    if (year < 0) {
+      *out_++ = '-';
+      year = 0 - year;
+      --width;
+    }
+    uint32_or_64_or_128_t<long long> n = to_unsigned(year);
+    const int num_digits = count_digits(n);
+    if (width > num_digits)
+      out_ = detail::fill_n(out_, width - num_digits, '0');
+    out_ = format_decimal<Char>(out_, n, num_digits).end;
+  }
+  void write_year(long long year) {
+    if (year >= 0 && year < 10000) {
+      write2(static_cast<int>(year / 100));
+      write2(static_cast<int>(year % 100));
+    } else {
+      write_year_extended(year);
+    }
+  }
+
+  void write_utc_offset(long offset, numeric_system ns) {
+    if (offset < 0) {
+      *out_++ = '-';
+      offset = -offset;
+    } else {
+      *out_++ = '+';
+    }
+    offset /= 60;
+    write2(static_cast<int>(offset / 60));
+    if (ns != numeric_system::standard) *out_++ = ':';
+    write2(static_cast<int>(offset % 60));
+  }
+  template <typename T, FMT_ENABLE_IF(has_member_data_tm_gmtoff<T>::value)>
+  void format_utc_offset_impl(const T& tm, numeric_system ns) {
+    write_utc_offset(tm.tm_gmtoff, ns);
+  }
+  template <typename T, FMT_ENABLE_IF(!has_member_data_tm_gmtoff<T>::value)>
+  void format_utc_offset_impl(const T& tm, numeric_system ns) {
+#if defined(_WIN32) && defined(_UCRT)
+#  if FMT_USE_TZSET
+    tzset_once();
+#  endif
+    long offset = 0;
+    _get_timezone(&offset);
+    if (tm.tm_isdst) {
+      long dstbias = 0;
+      _get_dstbias(&dstbias);
+      offset += dstbias;
+    }
+    write_utc_offset(-offset, ns);
+#else
+    if (ns == numeric_system::standard) return format_localized('z');
+
+    // Extract timezone offset from timezone conversion functions.
+    std::tm gtm = tm;
+    std::time_t gt = std::mktime(&gtm);
+    std::tm ltm = gmtime(gt);
+    std::time_t lt = std::mktime(&ltm);
+    long offset = gt - lt;
+    write_utc_offset(offset, ns);
+#endif
+  }
+
+  template <typename T, FMT_ENABLE_IF(has_member_data_tm_zone<T>::value)>
+  void format_tz_name_impl(const T& tm) {
+    if (is_classic_)
+      out_ = write_tm_str<Char>(out_, tm.tm_zone, loc_);
+    else
+      format_localized('Z');
+  }
+  template <typename T, FMT_ENABLE_IF(!has_member_data_tm_zone<T>::value)>
+  void format_tz_name_impl(const T&) {
+    format_localized('Z');
+  }
+
+  void format_localized(char format, char modifier = 0) {
+    out_ = write<Char>(out_, tm_, loc_, format, modifier);
+  }
+
+ public:
+  tm_writer(const std::locale& loc, OutputIt out, const std::tm& tm,
+            const Duration* subsecs = nullptr)
+      : loc_(loc),
+        is_classic_(loc_ == get_classic_locale()),
+        out_(out),
+        subsecs_(subsecs),
+        tm_(tm) {}
+
+  auto out() const -> OutputIt { return out_; }
+
+  FMT_CONSTEXPR void on_text(const Char* begin, const Char* end) {
+    out_ = copy<Char>(begin, end, out_);
+  }
+
+  void on_abbr_weekday() {
+    if (is_classic_)
+      out_ = write(out_, tm_wday_short_name(tm_wday()));
+    else
+      format_localized('a');
+  }
+  void on_full_weekday() {
+    if (is_classic_)
+      out_ = write(out_, tm_wday_full_name(tm_wday()));
+    else
+      format_localized('A');
+  }
+  void on_dec0_weekday(numeric_system ns) {
+    if (is_classic_ || ns == numeric_system::standard) return write1(tm_wday());
+    format_localized('w', 'O');
+  }
+  void on_dec1_weekday(numeric_system ns) {
+    if (is_classic_ || ns == numeric_system::standard) {
+      auto wday = tm_wday();
+      write1(wday == 0 ? days_per_week : wday);
+    } else {
+      format_localized('u', 'O');
+    }
+  }
+
+  void on_abbr_month() {
+    if (is_classic_)
+      out_ = write(out_, tm_mon_short_name(tm_mon()));
+    else
+      format_localized('b');
+  }
+  void on_full_month() {
+    if (is_classic_)
+      out_ = write(out_, tm_mon_full_name(tm_mon()));
+    else
+      format_localized('B');
+  }
+
+  void on_datetime(numeric_system ns) {
+    if (is_classic_) {
+      on_abbr_weekday();
+      *out_++ = ' ';
+      on_abbr_month();
+      *out_++ = ' ';
+      on_day_of_month(numeric_system::standard, pad_type::space);
+      *out_++ = ' ';
+      on_iso_time();
+      *out_++ = ' ';
+      on_year(numeric_system::standard);
+    } else {
+      format_localized('c', ns == numeric_system::standard ? '\0' : 'E');
+    }
+  }
+  void on_loc_date(numeric_system ns) {
+    if (is_classic_)
+      on_us_date();
+    else
+      format_localized('x', ns == numeric_system::standard ? '\0' : 'E');
+  }
+  void on_loc_time(numeric_system ns) {
+    if (is_classic_)
+      on_iso_time();
+    else
+      format_localized('X', ns == numeric_system::standard ? '\0' : 'E');
+  }
+  void on_us_date() {
+    char buf[8];
+    write_digit2_separated(buf, to_unsigned(tm_mon() + 1),
+                           to_unsigned(tm_mday()),
+                           to_unsigned(split_year_lower(tm_year())), '/');
+    out_ = copy<Char>(std::begin(buf), std::end(buf), out_);
+  }
+  void on_iso_date() {
+    auto year = tm_year();
+    char buf[10];
+    size_t offset = 0;
+    if (year >= 0 && year < 10000) {
+      copy2(buf, digits2(static_cast<size_t>(year / 100)));
+    } else {
+      offset = 4;
+      write_year_extended(year);
+      year = 0;
+    }
+    write_digit2_separated(buf + 2, static_cast<unsigned>(year % 100),
+                           to_unsigned(tm_mon() + 1), to_unsigned(tm_mday()),
+                           '-');
+    out_ = copy<Char>(std::begin(buf) + offset, std::end(buf), out_);
+  }
+
+  void on_utc_offset(numeric_system ns) { format_utc_offset_impl(tm_, ns); }
+  void on_tz_name() { format_tz_name_impl(tm_); }
+
+  void on_year(numeric_system ns) {
+    if (is_classic_ || ns == numeric_system::standard)
+      return write_year(tm_year());
+    format_localized('Y', 'E');
+  }
+  void on_short_year(numeric_system ns) {
+    if (is_classic_ || ns == numeric_system::standard)
+      return write2(split_year_lower(tm_year()));
+    format_localized('y', 'O');
+  }
+  void on_offset_year() {
+    if (is_classic_) return write2(split_year_lower(tm_year()));
+    format_localized('y', 'E');
+  }
+
+  void on_century(numeric_system ns) {
+    if (is_classic_ || ns == numeric_system::standard) {
+      auto year = tm_year();
+      auto upper = year / 100;
+      if (year >= -99 && year < 0) {
+        // Zero upper on negative year.
+        *out_++ = '-';
+        *out_++ = '0';
+      } else if (upper >= 0 && upper < 100) {
+        write2(static_cast<int>(upper));
+      } else {
+        out_ = write<Char>(out_, upper);
+      }
+    } else {
+      format_localized('C', 'E');
+    }
+  }
+
+  void on_dec_month(numeric_system ns) {
+    if (is_classic_ || ns == numeric_system::standard)
+      return write2(tm_mon() + 1);
+    format_localized('m', 'O');
+  }
+
+  void on_dec0_week_of_year(numeric_system ns, pad_type pad) {
+    if (is_classic_ || ns == numeric_system::standard)
+      return write2((tm_yday() + days_per_week - tm_wday()) / days_per_week,
+                    pad);
+    format_localized('U', 'O');
+  }
+  void on_dec1_week_of_year(numeric_system ns, pad_type pad) {
+    if (is_classic_ || ns == numeric_system::standard) {
+      auto wday = tm_wday();
+      write2((tm_yday() + days_per_week -
+              (wday == 0 ? (days_per_week - 1) : (wday - 1))) /
+                 days_per_week,
+             pad);
+    } else {
+      format_localized('W', 'O');
+    }
+  }
+  void on_iso_week_of_year(numeric_system ns, pad_type pad) {
+    if (is_classic_ || ns == numeric_system::standard)
+      return write2(tm_iso_week_of_year(), pad);
+    format_localized('V', 'O');
+  }
+
+  void on_iso_week_based_year() { write_year(tm_iso_week_year()); }
+  void on_iso_week_based_short_year() {
+    write2(split_year_lower(tm_iso_week_year()));
+  }
+
+  void on_day_of_year() {
+    auto yday = tm_yday() + 1;
+    write1(yday / 100);
+    write2(yday % 100);
+  }
+  void on_day_of_month(numeric_system ns, pad_type pad) {
+    if (is_classic_ || ns == numeric_system::standard)
+      return write2(tm_mday(), pad);
+    format_localized('d', 'O');
+  }
+
+  void on_24_hour(numeric_system ns, pad_type pad) {
+    if (is_classic_ || ns == numeric_system::standard)
+      return write2(tm_hour(), pad);
+    format_localized('H', 'O');
+  }
+  void on_12_hour(numeric_system ns, pad_type pad) {
+    if (is_classic_ || ns == numeric_system::standard)
+      return write2(tm_hour12(), pad);
+    format_localized('I', 'O');
+  }
+  void on_minute(numeric_system ns, pad_type pad) {
+    if (is_classic_ || ns == numeric_system::standard)
+      return write2(tm_min(), pad);
+    format_localized('M', 'O');
+  }
+
+  void on_second(numeric_system ns, pad_type pad) {
+    if (is_classic_ || ns == numeric_system::standard) {
+      write2(tm_sec(), pad);
+      if (subsecs_) {
+        if (std::is_floating_point<typename Duration::rep>::value) {
+          auto buf = memory_buffer();
+          write_floating_seconds(buf, *subsecs_);
+          if (buf.size() > 1) {
+            // Remove the leading "0", write something like ".123".
+            out_ = std::copy(buf.begin() + 1, buf.end(), out_);
+          }
+        } else {
+          write_fractional_seconds<Char>(out_, *subsecs_);
+        }
+      }
+    } else {
+      // Currently no formatting of subseconds when a locale is set.
+      format_localized('S', 'O');
+    }
+  }
+
+  void on_12_hour_time() {
+    if (is_classic_) {
+      char buf[8];
+      write_digit2_separated(buf, to_unsigned(tm_hour12()),
+                             to_unsigned(tm_min()), to_unsigned(tm_sec()), ':');
+      out_ = copy<Char>(std::begin(buf), std::end(buf), out_);
+      *out_++ = ' ';
+      on_am_pm();
+    } else {
+      format_localized('r');
+    }
+  }
+  void on_24_hour_time() {
+    write2(tm_hour());
+    *out_++ = ':';
+    write2(tm_min());
+  }
+  void on_iso_time() {
+    on_24_hour_time();
+    *out_++ = ':';
+    on_second(numeric_system::standard, pad_type::zero);
+  }
+
+  void on_am_pm() {
+    if (is_classic_) {
+      *out_++ = tm_hour() < 12 ? 'A' : 'P';
+      *out_++ = 'M';
+    } else {
+      format_localized('p');
+    }
+  }
+
+  // These apply to chrono durations but not tm.
+  void on_duration_value() {}
+  void on_duration_unit() {}
+};
+
+struct chrono_format_checker : null_chrono_spec_handler<chrono_format_checker> {
+  bool has_precision_integral = false;
+
+  FMT_NORETURN void unsupported() { FMT_THROW(format_error("no date")); }
+
+  template <typename Char>
+  FMT_CONSTEXPR void on_text(const Char*, const Char*) {}
+  FMT_CONSTEXPR void on_day_of_year() {}
+  FMT_CONSTEXPR void on_24_hour(numeric_system, pad_type) {}
+  FMT_CONSTEXPR void on_12_hour(numeric_system, pad_type) {}
+  FMT_CONSTEXPR void on_minute(numeric_system, pad_type) {}
+  FMT_CONSTEXPR void on_second(numeric_system, pad_type) {}
+  FMT_CONSTEXPR void on_12_hour_time() {}
+  FMT_CONSTEXPR void on_24_hour_time() {}
+  FMT_CONSTEXPR void on_iso_time() {}
+  FMT_CONSTEXPR void on_am_pm() {}
+  FMT_CONSTEXPR void on_duration_value() const {
+    if (has_precision_integral) {
+      FMT_THROW(format_error("precision not allowed for this argument type"));
+    }
+  }
+  FMT_CONSTEXPR void on_duration_unit() {}
+};
+
+template <typename T,
+          FMT_ENABLE_IF(std::is_integral<T>::value&& has_isfinite<T>::value)>
+inline auto isfinite(T) -> bool {
+  return true;
+}
+
+template <typename T, FMT_ENABLE_IF(std::is_integral<T>::value)>
+inline auto mod(T x, int y) -> T {
+  return x % static_cast<T>(y);
+}
+template <typename T, FMT_ENABLE_IF(std::is_floating_point<T>::value)>
+inline auto mod(T x, int y) -> T {
+  return std::fmod(x, static_cast<T>(y));
+}
+
+// If T is an integral type, maps T to its unsigned counterpart, otherwise
+// leaves it unchanged (unlike std::make_unsigned).
+template <typename T, bool INTEGRAL = std::is_integral<T>::value>
+struct make_unsigned_or_unchanged {
+  using type = T;
+};
+
+template <typename T> struct make_unsigned_or_unchanged<T, true> {
+  using type = typename std::make_unsigned<T>::type;
+};
+
+template <typename Rep, typename Period,
+          FMT_ENABLE_IF(std::is_integral<Rep>::value)>
+inline auto get_milliseconds(std::chrono::duration<Rep, Period> d)
+    -> std::chrono::duration<Rep, std::milli> {
+  // this may overflow and/or the result may not fit in the
+  // target type.
+#if FMT_SAFE_DURATION_CAST
+  using CommonSecondsType =
+      typename std::common_type<decltype(d), std::chrono::seconds>::type;
+  const auto d_as_common = fmt_duration_cast<CommonSecondsType>(d);
+  const auto d_as_whole_seconds =
+      fmt_duration_cast<std::chrono::seconds>(d_as_common);
+  // this conversion should be nonproblematic
+  const auto diff = d_as_common - d_as_whole_seconds;
+  const auto ms =
+      fmt_duration_cast<std::chrono::duration<Rep, std::milli>>(diff);
+  return ms;
+#else
+  auto s = fmt_duration_cast<std::chrono::seconds>(d);
+  return fmt_duration_cast<std::chrono::milliseconds>(d - s);
+#endif
+}
+
+template <typename Char, typename Rep, typename OutputIt,
+          FMT_ENABLE_IF(std::is_integral<Rep>::value)>
+auto format_duration_value(OutputIt out, Rep val, int) -> OutputIt {
+  return write<Char>(out, val);
+}
+
+template <typename Char, typename Rep, typename OutputIt,
+          FMT_ENABLE_IF(std::is_floating_point<Rep>::value)>
+auto format_duration_value(OutputIt out, Rep val, int precision) -> OutputIt {
+  auto specs = format_specs();
+  specs.precision = precision;
+  specs.type =
+      precision >= 0 ? presentation_type::fixed : presentation_type::general;
+  return write<Char>(out, val, specs);
+}
+
+template <typename Char, typename OutputIt>
+auto copy_unit(string_view unit, OutputIt out, Char) -> OutputIt {
+  return std::copy(unit.begin(), unit.end(), out);
+}
+
+template <typename OutputIt>
+auto copy_unit(string_view unit, OutputIt out, wchar_t) -> OutputIt {
+  // This works when wchar_t is UTF-32 because units only contain characters
+  // that have the same representation in UTF-16 and UTF-32.
+  utf8_to_utf16 u(unit);
+  return std::copy(u.c_str(), u.c_str() + u.size(), out);
+}
+
+template <typename Char, typename Period, typename OutputIt>
+auto format_duration_unit(OutputIt out) -> OutputIt {
+  if (const char* unit = get_units<Period>())
+    return copy_unit(string_view(unit), out, Char());
+  *out++ = '[';
+  out = write<Char>(out, Period::num);
+  if (const_check(Period::den != 1)) {
+    *out++ = '/';
+    out = write<Char>(out, Period::den);
+  }
+  *out++ = ']';
+  *out++ = 's';
+  return out;
+}
+
+class get_locale {
+ private:
+  union {
+    std::locale locale_;
+  };
+  bool has_locale_ = false;
+
+ public:
+  get_locale(bool localized, locale_ref loc) : has_locale_(localized) {
+#ifndef FMT_STATIC_THOUSANDS_SEPARATOR
+    if (localized)
+      ::new (&locale_) std::locale(loc.template get<std::locale>());
+#endif
+  }
+  ~get_locale() {
+    if (has_locale_) locale_.~locale();
+  }
+  operator const std::locale&() const {
+    return has_locale_ ? locale_ : get_classic_locale();
+  }
+};
+
+template <typename FormatContext, typename OutputIt, typename Rep,
+          typename Period>
+struct chrono_formatter {
+  FormatContext& context;
+  OutputIt out;
+  int precision;
+  bool localized = false;
+  // rep is unsigned to avoid overflow.
+  using rep =
+      conditional_t<std::is_integral<Rep>::value && sizeof(Rep) < sizeof(int),
+                    unsigned, typename make_unsigned_or_unchanged<Rep>::type>;
+  rep val;
+  using seconds = std::chrono::duration<rep>;
+  seconds s;
+  using milliseconds = std::chrono::duration<rep, std::milli>;
+  bool negative;
+
+  using char_type = typename FormatContext::char_type;
+  using tm_writer_type = tm_writer<OutputIt, char_type>;
+
+  chrono_formatter(FormatContext& ctx, OutputIt o,
+                   std::chrono::duration<Rep, Period> d)
+      : context(ctx),
+        out(o),
+        val(static_cast<rep>(d.count())),
+        negative(false) {
+    if (d.count() < 0) {
+      val = 0 - val;
+      negative = true;
+    }
+
+    // this may overflow and/or the result may not fit in the
+    // target type.
+    // might need checked conversion (rep!=Rep)
+    s = fmt_duration_cast<seconds>(std::chrono::duration<rep, Period>(val));
+  }
+
+  // returns true if nan or inf, writes to out.
+  auto handle_nan_inf() -> bool {
+    if (isfinite(val)) {
+      return false;
+    }
+    if (isnan(val)) {
+      write_nan();
+      return true;
+    }
+    // must be +-inf
+    if (val > 0) {
+      write_pinf();
+    } else {
+      write_ninf();
+    }
+    return true;
+  }
+
+  auto days() const -> Rep { return static_cast<Rep>(s.count() / 86400); }
+  auto hour() const -> Rep {
+    return static_cast<Rep>(mod((s.count() / 3600), 24));
+  }
+
+  auto hour12() const -> Rep {
+    Rep hour = static_cast<Rep>(mod((s.count() / 3600), 12));
+    return hour <= 0 ? 12 : hour;
+  }
+
+  auto minute() const -> Rep {
+    return static_cast<Rep>(mod((s.count() / 60), 60));
+  }
+  auto second() const -> Rep { return static_cast<Rep>(mod(s.count(), 60)); }
+
+  auto time() const -> std::tm {
+    auto time = std::tm();
+    time.tm_hour = to_nonnegative_int(hour(), 24);
+    time.tm_min = to_nonnegative_int(minute(), 60);
+    time.tm_sec = to_nonnegative_int(second(), 60);
+    return time;
+  }
+
+  void write_sign() {
+    if (negative) {
+      *out++ = '-';
+      negative = false;
+    }
+  }
+
+  void write(Rep value, int width, pad_type pad = pad_type::zero) {
+    write_sign();
+    if (isnan(value)) return write_nan();
+    uint32_or_64_or_128_t<int> n =
+        to_unsigned(to_nonnegative_int(value, max_value<int>()));
+    int num_digits = detail::count_digits(n);
+    if (width > num_digits) {
+      out = detail::write_padding(out, pad, width - num_digits);
+    }
+    out = format_decimal<char_type>(out, n, num_digits).end;
+  }
+
+  void write_nan() { std::copy_n("nan", 3, out); }
+  void write_pinf() { std::copy_n("inf", 3, out); }
+  void write_ninf() { std::copy_n("-inf", 4, out); }
+
+  template <typename Callback, typename... Args>
+  void format_tm(const tm& time, Callback cb, Args... args) {
+    if (isnan(val)) return write_nan();
+    get_locale loc(localized, context.locale());
+    auto w = tm_writer_type(loc, out, time);
+    (w.*cb)(args...);
+    out = w.out();
+  }
+
+  void on_text(const char_type* begin, const char_type* end) {
+    std::copy(begin, end, out);
+  }
+
+  // These are not implemented because durations don't have date information.
+  void on_abbr_weekday() {}
+  void on_full_weekday() {}
+  void on_dec0_weekday(numeric_system) {}
+  void on_dec1_weekday(numeric_system) {}
+  void on_abbr_month() {}
+  void on_full_month() {}
+  void on_datetime(numeric_system) {}
+  void on_loc_date(numeric_system) {}
+  void on_loc_time(numeric_system) {}
+  void on_us_date() {}
+  void on_iso_date() {}
+  void on_utc_offset(numeric_system) {}
+  void on_tz_name() {}
+  void on_year(numeric_system) {}
+  void on_short_year(numeric_system) {}
+  void on_offset_year() {}
+  void on_century(numeric_system) {}
+  void on_iso_week_based_year() {}
+  void on_iso_week_based_short_year() {}
+  void on_dec_month(numeric_system) {}
+  void on_dec0_week_of_year(numeric_system, pad_type) {}
+  void on_dec1_week_of_year(numeric_system, pad_type) {}
+  void on_iso_week_of_year(numeric_system, pad_type) {}
+  void on_day_of_month(numeric_system, pad_type) {}
+
+  void on_day_of_year() {
+    if (handle_nan_inf()) return;
+    write(days(), 0);
+  }
+
+  void on_24_hour(numeric_system ns, pad_type pad) {
+    if (handle_nan_inf()) return;
+
+    if (ns == numeric_system::standard) return write(hour(), 2, pad);
+    auto time = tm();
+    time.tm_hour = to_nonnegative_int(hour(), 24);
+    format_tm(time, &tm_writer_type::on_24_hour, ns, pad);
+  }
+
+  void on_12_hour(numeric_system ns, pad_type pad) {
+    if (handle_nan_inf()) return;
+
+    if (ns == numeric_system::standard) return write(hour12(), 2, pad);
+    auto time = tm();
+    time.tm_hour = to_nonnegative_int(hour12(), 12);
+    format_tm(time, &tm_writer_type::on_12_hour, ns, pad);
+  }
+
+  void on_minute(numeric_system ns, pad_type pad) {
+    if (handle_nan_inf()) return;
+
+    if (ns == numeric_system::standard) return write(minute(), 2, pad);
+    auto time = tm();
+    time.tm_min = to_nonnegative_int(minute(), 60);
+    format_tm(time, &tm_writer_type::on_minute, ns, pad);
+  }
+
+  void on_second(numeric_system ns, pad_type pad) {
+    if (handle_nan_inf()) return;
+
+    if (ns == numeric_system::standard) {
+      if (std::is_floating_point<rep>::value) {
+        auto buf = memory_buffer();
+        write_floating_seconds(buf, std::chrono::duration<rep, Period>(val),
+                               precision);
+        if (negative) *out++ = '-';
+        if (buf.size() < 2 || buf[1] == '.') {
+          out = detail::write_padding(out, pad);
+        }
+        out = std::copy(buf.begin(), buf.end(), out);
+      } else {
+        write(second(), 2, pad);
+        write_fractional_seconds<char_type>(
+            out, std::chrono::duration<rep, Period>(val), precision);
+      }
+      return;
+    }
+    auto time = tm();
+    time.tm_sec = to_nonnegative_int(second(), 60);
+    format_tm(time, &tm_writer_type::on_second, ns, pad);
+  }
+
+  void on_12_hour_time() {
+    if (handle_nan_inf()) return;
+    format_tm(time(), &tm_writer_type::on_12_hour_time);
+  }
+
+  void on_24_hour_time() {
+    if (handle_nan_inf()) {
+      *out++ = ':';
+      handle_nan_inf();
+      return;
+    }
+
+    write(hour(), 2);
+    *out++ = ':';
+    write(minute(), 2);
+  }
+
+  void on_iso_time() {
+    on_24_hour_time();
+    *out++ = ':';
+    if (handle_nan_inf()) return;
+    on_second(numeric_system::standard, pad_type::zero);
+  }
+
+  void on_am_pm() {
+    if (handle_nan_inf()) return;
+    format_tm(time(), &tm_writer_type::on_am_pm);
+  }
+
+  void on_duration_value() {
+    if (handle_nan_inf()) return;
+    write_sign();
+    out = format_duration_value<char_type>(out, val, precision);
+  }
+
+  void on_duration_unit() {
+    out = format_duration_unit<char_type, Period>(out);
+  }
+};
+
+}  // namespace detail
+
+#if defined(__cpp_lib_chrono) && __cpp_lib_chrono >= 201907
+using weekday = std::chrono::weekday;
+using day = std::chrono::day;
+using month = std::chrono::month;
+using year = std::chrono::year;
+using year_month_day = std::chrono::year_month_day;
+#else
+// A fallback version of weekday.
+class weekday {
+ private:
+  unsigned char value_;
+
+ public:
+  weekday() = default;
+  constexpr explicit weekday(unsigned wd) noexcept
+      : value_(static_cast<unsigned char>(wd != 7 ? wd : 0)) {}
+  constexpr auto c_encoding() const noexcept -> unsigned { return value_; }
+};
+
+class day {
+ private:
+  unsigned char value_;
+
+ public:
+  day() = default;
+  constexpr explicit day(unsigned d) noexcept
+      : value_(static_cast<unsigned char>(d)) {}
+  constexpr explicit operator unsigned() const noexcept { return value_; }
+};
+
+class month {
+ private:
+  unsigned char value_;
+
+ public:
+  month() = default;
+  constexpr explicit month(unsigned m) noexcept
+      : value_(static_cast<unsigned char>(m)) {}
+  constexpr explicit operator unsigned() const noexcept { return value_; }
+};
+
+class year {
+ private:
+  int value_;
+
+ public:
+  year() = default;
+  constexpr explicit year(int y) noexcept : value_(y) {}
+  constexpr explicit operator int() const noexcept { return value_; }
+};
+
+class year_month_day {
+ private:
+  fmt::year year_;
+  fmt::month month_;
+  fmt::day day_;
+
+ public:
+  year_month_day() = default;
+  constexpr year_month_day(const year& y, const month& m, const day& d) noexcept
+      : year_(y), month_(m), day_(d) {}
+  constexpr auto year() const noexcept -> fmt::year { return year_; }
+  constexpr auto month() const noexcept -> fmt::month { return month_; }
+  constexpr auto day() const noexcept -> fmt::day { return day_; }
+};
+#endif
+
+template <typename Char>
+struct formatter<weekday, Char> : private formatter<std::tm, Char> {
+ private:
+  bool localized_ = false;
+  bool use_tm_formatter_ = false;
+
+ public:
+  FMT_CONSTEXPR auto parse(basic_format_parse_context<Char>& ctx)
+      -> decltype(ctx.begin()) {
+    auto it = ctx.begin(), end = ctx.end();
+    if (it != end && *it == 'L') {
+      ++it;
+      localized_ = true;
+      return it;
+    }
+    use_tm_formatter_ = it != end && *it != '}';
+    return use_tm_formatter_ ? formatter<std::tm, Char>::parse(ctx) : it;
+  }
+
+  template <typename FormatContext>
+  auto format(weekday wd, FormatContext& ctx) const -> decltype(ctx.out()) {
+    auto time = std::tm();
+    time.tm_wday = static_cast<int>(wd.c_encoding());
+    if (use_tm_formatter_) return formatter<std::tm, Char>::format(time, ctx);
+    detail::get_locale loc(localized_, ctx.locale());
+    auto w = detail::tm_writer<decltype(ctx.out()), Char>(loc, ctx.out(), time);
+    w.on_abbr_weekday();
+    return w.out();
+  }
+};
+
+template <typename Char>
+struct formatter<day, Char> : private formatter<std::tm, Char> {
+ private:
+  bool use_tm_formatter_ = false;
+
+ public:
+  FMT_CONSTEXPR auto parse(basic_format_parse_context<Char>& ctx)
+      -> decltype(ctx.begin()) {
+    auto it = ctx.begin(), end = ctx.end();
+    use_tm_formatter_ = it != end && *it != '}';
+    return use_tm_formatter_ ? formatter<std::tm, Char>::parse(ctx) : it;
+  }
+
+  template <typename FormatContext>
+  auto format(day d, FormatContext& ctx) const -> decltype(ctx.out()) {
+    auto time = std::tm();
+    time.tm_mday = static_cast<int>(static_cast<unsigned>(d));
+    if (use_tm_formatter_) return formatter<std::tm, Char>::format(time, ctx);
+    detail::get_locale loc(false, ctx.locale());
+    auto w = detail::tm_writer<decltype(ctx.out()), Char>(loc, ctx.out(), time);
+    w.on_day_of_month(detail::numeric_system::standard, detail::pad_type::zero);
+    return w.out();
+  }
+};
+
+template <typename Char>
+struct formatter<month, Char> : private formatter<std::tm, Char> {
+ private:
+  bool localized_ = false;
+  bool use_tm_formatter_ = false;
+
+ public:
+  FMT_CONSTEXPR auto parse(basic_format_parse_context<Char>& ctx)
+      -> decltype(ctx.begin()) {
+    auto it = ctx.begin(), end = ctx.end();
+    if (it != end && *it == 'L') {
+      ++it;
+      localized_ = true;
+      return it;
+    }
+    use_tm_formatter_ = it != end && *it != '}';
+    return use_tm_formatter_ ? formatter<std::tm, Char>::parse(ctx) : it;
+  }
+
+  template <typename FormatContext>
+  auto format(month m, FormatContext& ctx) const -> decltype(ctx.out()) {
+    auto time = std::tm();
+    time.tm_mon = static_cast<int>(static_cast<unsigned>(m)) - 1;
+    if (use_tm_formatter_) return formatter<std::tm, Char>::format(time, ctx);
+    detail::get_locale loc(localized_, ctx.locale());
+    auto w = detail::tm_writer<decltype(ctx.out()), Char>(loc, ctx.out(), time);
+    w.on_abbr_month();
+    return w.out();
+  }
+};
+
+template <typename Char>
+struct formatter<year, Char> : private formatter<std::tm, Char> {
+ private:
+  bool use_tm_formatter_ = false;
+
+ public:
+  FMT_CONSTEXPR auto parse(basic_format_parse_context<Char>& ctx)
+      -> decltype(ctx.begin()) {
+    auto it = ctx.begin(), end = ctx.end();
+    use_tm_formatter_ = it != end && *it != '}';
+    return use_tm_formatter_ ? formatter<std::tm, Char>::parse(ctx) : it;
+  }
+
+  template <typename FormatContext>
+  auto format(year y, FormatContext& ctx) const -> decltype(ctx.out()) {
+    auto time = std::tm();
+    time.tm_year = static_cast<int>(y) - 1900;
+    if (use_tm_formatter_) return formatter<std::tm, Char>::format(time, ctx);
+    detail::get_locale loc(false, ctx.locale());
+    auto w = detail::tm_writer<decltype(ctx.out()), Char>(loc, ctx.out(), time);
+    w.on_year(detail::numeric_system::standard);
+    return w.out();
+  }
+};
+
+template <typename Char>
+struct formatter<year_month_day, Char> : private formatter<std::tm, Char> {
+ private:
+  bool use_tm_formatter_ = false;
+
+ public:
+  FMT_CONSTEXPR auto parse(basic_format_parse_context<Char>& ctx)
+      -> decltype(ctx.begin()) {
+    auto it = ctx.begin(), end = ctx.end();
+    use_tm_formatter_ = it != end && *it != '}';
+    return use_tm_formatter_ ? formatter<std::tm, Char>::parse(ctx) : it;
+  }
+
+  template <typename FormatContext>
+  auto format(year_month_day val, FormatContext& ctx) const
+      -> decltype(ctx.out()) {
+    auto time = std::tm();
+    time.tm_year = static_cast<int>(val.year()) - 1900;
+    time.tm_mon = static_cast<int>(static_cast<unsigned>(val.month())) - 1;
+    time.tm_mday = static_cast<int>(static_cast<unsigned>(val.day()));
+    if (use_tm_formatter_) return formatter<std::tm, Char>::format(time, ctx);
+    detail::get_locale loc(true, ctx.locale());
+    auto w = detail::tm_writer<decltype(ctx.out()), Char>(loc, ctx.out(), time);
+    w.on_iso_date();
+    return w.out();
+  }
+};
+
+template <typename Rep, typename Period, typename Char>
+struct formatter<std::chrono::duration<Rep, Period>, Char> {
+ private:
+  format_specs specs_;
+  detail::arg_ref<Char> width_ref_;
+  detail::arg_ref<Char> precision_ref_;
+  bool localized_ = false;
+  basic_string_view<Char> format_str_;
+
+ public:
+  FMT_CONSTEXPR auto parse(basic_format_parse_context<Char>& ctx)
+      -> decltype(ctx.begin()) {
+    auto it = ctx.begin(), end = ctx.end();
+    if (it == end || *it == '}') return it;
+
+    it = detail::parse_align(it, end, specs_);
+    if (it == end) return it;
+
+    it = detail::parse_dynamic_spec(it, end, specs_.width, width_ref_, ctx);
+    if (it == end) return it;
+
+    auto checker = detail::chrono_format_checker();
+    if (*it == '.') {
+      checker.has_precision_integral = !std::is_floating_point<Rep>::value;
+      it = detail::parse_precision(it, end, specs_.precision, precision_ref_,
+                                   ctx);
+    }
+    if (it != end && *it == 'L') {
+      localized_ = true;
+      ++it;
+    }
+    end = detail::parse_chrono_format(it, end, checker);
+    format_str_ = {it, detail::to_unsigned(end - it)};
+    return end;
+  }
+
+  template <typename FormatContext>
+  auto format(std::chrono::duration<Rep, Period> d, FormatContext& ctx) const
+      -> decltype(ctx.out()) {
+    auto specs = specs_;
+    auto precision = specs.precision;
+    specs.precision = -1;
+    auto begin = format_str_.begin(), end = format_str_.end();
+    // As a possible future optimization, we could avoid extra copying if width
+    // is not specified.
+    auto buf = basic_memory_buffer<Char>();
+    auto out = std::back_inserter(buf);
+    detail::handle_dynamic_spec<detail::width_checker>(specs.width, width_ref_,
+                                                       ctx);
+    detail::handle_dynamic_spec<detail::precision_checker>(precision,
+                                                           precision_ref_, ctx);
+    if (begin == end || *begin == '}') {
+      out = detail::format_duration_value<Char>(out, d.count(), precision);
+      detail::format_duration_unit<Char, Period>(out);
+    } else {
+      using chrono_formatter =
+          detail::chrono_formatter<FormatContext, decltype(out), Rep, Period>;
+      auto f = chrono_formatter(ctx, out, d);
+      f.precision = precision;
+      f.localized = localized_;
+      detail::parse_chrono_format(begin, end, f);
+    }
+    return detail::write(
+        ctx.out(), basic_string_view<Char>(buf.data(), buf.size()), specs);
+  }
+};
+
+template <typename Char, typename Duration>
+struct formatter<std::chrono::time_point<std::chrono::system_clock, Duration>,
+                 Char> : formatter<std::tm, Char> {
+  FMT_CONSTEXPR formatter() {
+    this->format_str_ = detail::string_literal<Char, '%', 'F', ' ', '%', 'T'>{};
+  }
+
+  template <typename FormatContext>
+  auto format(std::chrono::time_point<std::chrono::system_clock, Duration> val,
+              FormatContext& ctx) const -> decltype(ctx.out()) {
+    std::tm tm = gmtime(val);
+    using period = typename Duration::period;
+    if (detail::const_check(
+            period::num == 1 && period::den == 1 &&
+            !std::is_floating_point<typename Duration::rep>::value)) {
+      return formatter<std::tm, Char>::format(tm, ctx);
+    }
+    Duration epoch = val.time_since_epoch();
+    Duration subsecs = detail::fmt_duration_cast<Duration>(
+        epoch - detail::fmt_duration_cast<std::chrono::seconds>(epoch));
+    if (subsecs.count() < 0) {
+      auto second =
+          detail::fmt_duration_cast<Duration>(std::chrono::seconds(1));
+      if (tm.tm_sec != 0)
+        --tm.tm_sec;
+      else
+        tm = gmtime(val - second);
+      subsecs += detail::fmt_duration_cast<Duration>(std::chrono::seconds(1));
+    }
+    return formatter<std::tm, Char>::do_format(tm, ctx, &subsecs);
+  }
+};
+
+#if FMT_USE_LOCAL_TIME
+template <typename Char, typename Duration>
+struct formatter<std::chrono::local_time<Duration>, Char>
+    : formatter<std::tm, Char> {
+  FMT_CONSTEXPR formatter() {
+    this->format_str_ = detail::string_literal<Char, '%', 'F', ' ', '%', 'T'>{};
+  }
+
+  template <typename FormatContext>
+  auto format(std::chrono::local_time<Duration> val, FormatContext& ctx) const
+      -> decltype(ctx.out()) {
+    using period = typename Duration::period;
+    if (period::num != 1 || period::den != 1 ||
+        std::is_floating_point<typename Duration::rep>::value) {
+      const auto epoch = val.time_since_epoch();
+      const auto subsecs = detail::fmt_duration_cast<Duration>(
+          epoch - detail::fmt_duration_cast<std::chrono::seconds>(epoch));
+
+      return formatter<std::tm, Char>::do_format(localtime(val), ctx, &subsecs);
+    }
+
+    return formatter<std::tm, Char>::format(localtime(val), ctx);
+  }
+};
+#endif
+
+#if FMT_USE_UTC_TIME
+template <typename Char, typename Duration>
+struct formatter<std::chrono::time_point<std::chrono::utc_clock, Duration>,
+                 Char>
+    : formatter<std::chrono::time_point<std::chrono::system_clock, Duration>,
+                Char> {
+  template <typename FormatContext>
+  auto format(std::chrono::time_point<std::chrono::utc_clock, Duration> val,
+              FormatContext& ctx) const -> decltype(ctx.out()) {
+    return formatter<
+        std::chrono::time_point<std::chrono::system_clock, Duration>,
+        Char>::format(std::chrono::utc_clock::to_sys(val), ctx);
+  }
+};
+#endif
+
+template <typename Char> struct formatter<std::tm, Char> {
+ private:
+  format_specs specs_;
+  detail::arg_ref<Char> width_ref_;
+
+ protected:
+  basic_string_view<Char> format_str_;
+
+  template <typename FormatContext, typename Duration>
+  auto do_format(const std::tm& tm, FormatContext& ctx,
+                 const Duration* subsecs) const -> decltype(ctx.out()) {
+    auto specs = specs_;
+    auto buf = basic_memory_buffer<Char>();
+    auto out = std::back_inserter(buf);
+    detail::handle_dynamic_spec<detail::width_checker>(specs.width, width_ref_,
+                                                       ctx);
+
+    auto loc_ref = ctx.locale();
+    detail::get_locale loc(static_cast<bool>(loc_ref), loc_ref);
+    auto w =
+        detail::tm_writer<decltype(out), Char, Duration>(loc, out, tm, subsecs);
+    detail::parse_chrono_format(format_str_.begin(), format_str_.end(), w);
+    return detail::write(
+        ctx.out(), basic_string_view<Char>(buf.data(), buf.size()), specs);
+  }
+
+ public:
+  FMT_CONSTEXPR auto parse(basic_format_parse_context<Char>& ctx)
+      -> decltype(ctx.begin()) {
+    auto it = ctx.begin(), end = ctx.end();
+    if (it == end || *it == '}') return it;
+
+    it = detail::parse_align(it, end, specs_);
+    if (it == end) return it;
+
+    it = detail::parse_dynamic_spec(it, end, specs_.width, width_ref_, ctx);
+    if (it == end) return it;
+
+    end = detail::parse_chrono_format(it, end, detail::tm_format_checker());
+    // Replace the default format_str only if the new spec is not empty.
+    if (end != it) format_str_ = {it, detail::to_unsigned(end - it)};
+    return end;
+  }
+
+  template <typename FormatContext>
+  auto format(const std::tm& tm, FormatContext& ctx) const
+      -> decltype(ctx.out()) {
+    return do_format<FormatContext, std::chrono::seconds>(tm, ctx, nullptr);
+  }
+};
+
+FMT_END_EXPORT
+FMT_END_NAMESPACE
+
+#endif  // FMT_CHRONO_H_
diff --git a/ext/spdlog/fmt/bundled/color.h b/ext/spdlog/fmt/bundled/color.h
new file mode 100644
index 0000000..f0e9dd9
--- /dev/null
+++ b/ext/spdlog/fmt/bundled/color.h
@@ -0,0 +1,612 @@
+// Formatting library for C++ - color support
+//
+// Copyright (c) 2018 - present, Victor Zverovich and fmt contributors
+// All rights reserved.
+//
+// For the license information refer to format.h.
+
+#ifndef FMT_COLOR_H_
+#define FMT_COLOR_H_
+
+#include "format.h"
+
+FMT_BEGIN_NAMESPACE
+FMT_BEGIN_EXPORT
+
+enum class color : uint32_t {
+  alice_blue = 0xF0F8FF,               // rgb(240,248,255)
+  antique_white = 0xFAEBD7,            // rgb(250,235,215)
+  aqua = 0x00FFFF,                     // rgb(0,255,255)
+  aquamarine = 0x7FFFD4,               // rgb(127,255,212)
+  azure = 0xF0FFFF,                    // rgb(240,255,255)
+  beige = 0xF5F5DC,                    // rgb(245,245,220)
+  bisque = 0xFFE4C4,                   // rgb(255,228,196)
+  black = 0x000000,                    // rgb(0,0,0)
+  blanched_almond = 0xFFEBCD,          // rgb(255,235,205)
+  blue = 0x0000FF,                     // rgb(0,0,255)
+  blue_violet = 0x8A2BE2,              // rgb(138,43,226)
+  brown = 0xA52A2A,                    // rgb(165,42,42)
+  burly_wood = 0xDEB887,               // rgb(222,184,135)
+  cadet_blue = 0x5F9EA0,               // rgb(95,158,160)
+  chartreuse = 0x7FFF00,               // rgb(127,255,0)
+  chocolate = 0xD2691E,                // rgb(210,105,30)
+  coral = 0xFF7F50,                    // rgb(255,127,80)
+  cornflower_blue = 0x6495ED,          // rgb(100,149,237)
+  cornsilk = 0xFFF8DC,                 // rgb(255,248,220)
+  crimson = 0xDC143C,                  // rgb(220,20,60)
+  cyan = 0x00FFFF,                     // rgb(0,255,255)
+  dark_blue = 0x00008B,                // rgb(0,0,139)
+  dark_cyan = 0x008B8B,                // rgb(0,139,139)
+  dark_golden_rod = 0xB8860B,          // rgb(184,134,11)
+  dark_gray = 0xA9A9A9,                // rgb(169,169,169)
+  dark_green = 0x006400,               // rgb(0,100,0)
+  dark_khaki = 0xBDB76B,               // rgb(189,183,107)
+  dark_magenta = 0x8B008B,             // rgb(139,0,139)
+  dark_olive_green = 0x556B2F,         // rgb(85,107,47)
+  dark_orange = 0xFF8C00,              // rgb(255,140,0)
+  dark_orchid = 0x9932CC,              // rgb(153,50,204)
+  dark_red = 0x8B0000,                 // rgb(139,0,0)
+  dark_salmon = 0xE9967A,              // rgb(233,150,122)
+  dark_sea_green = 0x8FBC8F,           // rgb(143,188,143)
+  dark_slate_blue = 0x483D8B,          // rgb(72,61,139)
+  dark_slate_gray = 0x2F4F4F,          // rgb(47,79,79)
+  dark_turquoise = 0x00CED1,           // rgb(0,206,209)
+  dark_violet = 0x9400D3,              // rgb(148,0,211)
+  deep_pink = 0xFF1493,                // rgb(255,20,147)
+  deep_sky_blue = 0x00BFFF,            // rgb(0,191,255)
+  dim_gray = 0x696969,                 // rgb(105,105,105)
+  dodger_blue = 0x1E90FF,              // rgb(30,144,255)
+  fire_brick = 0xB22222,               // rgb(178,34,34)
+  floral_white = 0xFFFAF0,             // rgb(255,250,240)
+  forest_green = 0x228B22,             // rgb(34,139,34)
+  fuchsia = 0xFF00FF,                  // rgb(255,0,255)
+  gainsboro = 0xDCDCDC,                // rgb(220,220,220)
+  ghost_white = 0xF8F8FF,              // rgb(248,248,255)
+  gold = 0xFFD700,                     // rgb(255,215,0)
+  golden_rod = 0xDAA520,               // rgb(218,165,32)
+  gray = 0x808080,                     // rgb(128,128,128)
+  green = 0x008000,                    // rgb(0,128,0)
+  green_yellow = 0xADFF2F,             // rgb(173,255,47)
+  honey_dew = 0xF0FFF0,                // rgb(240,255,240)
+  hot_pink = 0xFF69B4,                 // rgb(255,105,180)
+  indian_red = 0xCD5C5C,               // rgb(205,92,92)
+  indigo = 0x4B0082,                   // rgb(75,0,130)
+  ivory = 0xFFFFF0,                    // rgb(255,255,240)
+  khaki = 0xF0E68C,                    // rgb(240,230,140)
+  lavender = 0xE6E6FA,                 // rgb(230,230,250)
+  lavender_blush = 0xFFF0F5,           // rgb(255,240,245)
+  lawn_green = 0x7CFC00,               // rgb(124,252,0)
+  lemon_chiffon = 0xFFFACD,            // rgb(255,250,205)
+  light_blue = 0xADD8E6,               // rgb(173,216,230)
+  light_coral = 0xF08080,              // rgb(240,128,128)
+  light_cyan = 0xE0FFFF,               // rgb(224,255,255)
+  light_golden_rod_yellow = 0xFAFAD2,  // rgb(250,250,210)
+  light_gray = 0xD3D3D3,               // rgb(211,211,211)
+  light_green = 0x90EE90,              // rgb(144,238,144)
+  light_pink = 0xFFB6C1,               // rgb(255,182,193)
+  light_salmon = 0xFFA07A,             // rgb(255,160,122)
+  light_sea_green = 0x20B2AA,          // rgb(32,178,170)
+  light_sky_blue = 0x87CEFA,           // rgb(135,206,250)
+  light_slate_gray = 0x778899,         // rgb(119,136,153)
+  light_steel_blue = 0xB0C4DE,         // rgb(176,196,222)
+  light_yellow = 0xFFFFE0,             // rgb(255,255,224)
+  lime = 0x00FF00,                     // rgb(0,255,0)
+  lime_green = 0x32CD32,               // rgb(50,205,50)
+  linen = 0xFAF0E6,                    // rgb(250,240,230)
+  magenta = 0xFF00FF,                  // rgb(255,0,255)
+  maroon = 0x800000,                   // rgb(128,0,0)
+  medium_aquamarine = 0x66CDAA,        // rgb(102,205,170)
+  medium_blue = 0x0000CD,              // rgb(0,0,205)
+  medium_orchid = 0xBA55D3,            // rgb(186,85,211)
+  medium_purple = 0x9370DB,            // rgb(147,112,219)
+  medium_sea_green = 0x3CB371,         // rgb(60,179,113)
+  medium_slate_blue = 0x7B68EE,        // rgb(123,104,238)
+  medium_spring_green = 0x00FA9A,      // rgb(0,250,154)
+  medium_turquoise = 0x48D1CC,         // rgb(72,209,204)
+  medium_violet_red = 0xC71585,        // rgb(199,21,133)
+  midnight_blue = 0x191970,            // rgb(25,25,112)
+  mint_cream = 0xF5FFFA,               // rgb(245,255,250)
+  misty_rose = 0xFFE4E1,               // rgb(255,228,225)
+  moccasin = 0xFFE4B5,                 // rgb(255,228,181)
+  navajo_white = 0xFFDEAD,             // rgb(255,222,173)
+  navy = 0x000080,                     // rgb(0,0,128)
+  old_lace = 0xFDF5E6,                 // rgb(253,245,230)
+  olive = 0x808000,                    // rgb(128,128,0)
+  olive_drab = 0x6B8E23,               // rgb(107,142,35)
+  orange = 0xFFA500,                   // rgb(255,165,0)
+  orange_red = 0xFF4500,               // rgb(255,69,0)
+  orchid = 0xDA70D6,                   // rgb(218,112,214)
+  pale_golden_rod = 0xEEE8AA,          // rgb(238,232,170)
+  pale_green = 0x98FB98,               // rgb(152,251,152)
+  pale_turquoise = 0xAFEEEE,           // rgb(175,238,238)
+  pale_violet_red = 0xDB7093,          // rgb(219,112,147)
+  papaya_whip = 0xFFEFD5,              // rgb(255,239,213)
+  peach_puff = 0xFFDAB9,               // rgb(255,218,185)
+  peru = 0xCD853F,                     // rgb(205,133,63)
+  pink = 0xFFC0CB,                     // rgb(255,192,203)
+  plum = 0xDDA0DD,                     // rgb(221,160,221)
+  powder_blue = 0xB0E0E6,              // rgb(176,224,230)
+  purple = 0x800080,                   // rgb(128,0,128)
+  rebecca_purple = 0x663399,           // rgb(102,51,153)
+  red = 0xFF0000,                      // rgb(255,0,0)
+  rosy_brown = 0xBC8F8F,               // rgb(188,143,143)
+  royal_blue = 0x4169E1,               // rgb(65,105,225)
+  saddle_brown = 0x8B4513,             // rgb(139,69,19)
+  salmon = 0xFA8072,                   // rgb(250,128,114)
+  sandy_brown = 0xF4A460,              // rgb(244,164,96)
+  sea_green = 0x2E8B57,                // rgb(46,139,87)
+  sea_shell = 0xFFF5EE,                // rgb(255,245,238)
+  sienna = 0xA0522D,                   // rgb(160,82,45)
+  silver = 0xC0C0C0,                   // rgb(192,192,192)
+  sky_blue = 0x87CEEB,                 // rgb(135,206,235)
+  slate_blue = 0x6A5ACD,               // rgb(106,90,205)
+  slate_gray = 0x708090,               // rgb(112,128,144)
+  snow = 0xFFFAFA,                     // rgb(255,250,250)
+  spring_green = 0x00FF7F,             // rgb(0,255,127)
+  steel_blue = 0x4682B4,               // rgb(70,130,180)
+  tan = 0xD2B48C,                      // rgb(210,180,140)
+  teal = 0x008080,                     // rgb(0,128,128)
+  thistle = 0xD8BFD8,                  // rgb(216,191,216)
+  tomato = 0xFF6347,                   // rgb(255,99,71)
+  turquoise = 0x40E0D0,                // rgb(64,224,208)
+  violet = 0xEE82EE,                   // rgb(238,130,238)
+  wheat = 0xF5DEB3,                    // rgb(245,222,179)
+  white = 0xFFFFFF,                    // rgb(255,255,255)
+  white_smoke = 0xF5F5F5,              // rgb(245,245,245)
+  yellow = 0xFFFF00,                   // rgb(255,255,0)
+  yellow_green = 0x9ACD32              // rgb(154,205,50)
+};                                     // enum class color
+
+enum class terminal_color : uint8_t {
+  black = 30,
+  red,
+  green,
+  yellow,
+  blue,
+  magenta,
+  cyan,
+  white,
+  bright_black = 90,
+  bright_red,
+  bright_green,
+  bright_yellow,
+  bright_blue,
+  bright_magenta,
+  bright_cyan,
+  bright_white
+};
+
+enum class emphasis : uint8_t {
+  bold = 1,
+  faint = 1 << 1,
+  italic = 1 << 2,
+  underline = 1 << 3,
+  blink = 1 << 4,
+  reverse = 1 << 5,
+  conceal = 1 << 6,
+  strikethrough = 1 << 7,
+};
+
+// rgb is a struct for red, green and blue colors.
+// Using the name "rgb" makes some editors show the color in a tooltip.
+struct rgb {
+  FMT_CONSTEXPR rgb() : r(0), g(0), b(0) {}
+  FMT_CONSTEXPR rgb(uint8_t r_, uint8_t g_, uint8_t b_) : r(r_), g(g_), b(b_) {}
+  FMT_CONSTEXPR rgb(uint32_t hex)
+      : r((hex >> 16) & 0xFF), g((hex >> 8) & 0xFF), b(hex & 0xFF) {}
+  FMT_CONSTEXPR rgb(color hex)
+      : r((uint32_t(hex) >> 16) & 0xFF),
+        g((uint32_t(hex) >> 8) & 0xFF),
+        b(uint32_t(hex) & 0xFF) {}
+  uint8_t r;
+  uint8_t g;
+  uint8_t b;
+};
+
+namespace detail {
+
+// color is a struct of either a rgb color or a terminal color.
+struct color_type {
+  FMT_CONSTEXPR color_type() noexcept : is_rgb(), value{} {}
+  FMT_CONSTEXPR color_type(color rgb_color) noexcept : is_rgb(true), value{} {
+    value.rgb_color = static_cast<uint32_t>(rgb_color);
+  }
+  FMT_CONSTEXPR color_type(rgb rgb_color) noexcept : is_rgb(true), value{} {
+    value.rgb_color = (static_cast<uint32_t>(rgb_color.r) << 16) |
+                      (static_cast<uint32_t>(rgb_color.g) << 8) | rgb_color.b;
+  }
+  FMT_CONSTEXPR color_type(terminal_color term_color) noexcept
+      : is_rgb(), value{} {
+    value.term_color = static_cast<uint8_t>(term_color);
+  }
+  bool is_rgb;
+  union color_union {
+    uint8_t term_color;
+    uint32_t rgb_color;
+  } value;
+};
+}  // namespace detail
+
+/// A text style consisting of foreground and background colors and emphasis.
+class text_style {
+ public:
+  FMT_CONSTEXPR text_style(emphasis em = emphasis()) noexcept
+      : set_foreground_color(), set_background_color(), ems(em) {}
+
+  FMT_CONSTEXPR auto operator|=(const text_style& rhs) -> text_style& {
+    if (!set_foreground_color) {
+      set_foreground_color = rhs.set_foreground_color;
+      foreground_color = rhs.foreground_color;
+    } else if (rhs.set_foreground_color) {
+      if (!foreground_color.is_rgb || !rhs.foreground_color.is_rgb)
+        report_error("can't OR a terminal color");
+      foreground_color.value.rgb_color |= rhs.foreground_color.value.rgb_color;
+    }
+
+    if (!set_background_color) {
+      set_background_color = rhs.set_background_color;
+      background_color = rhs.background_color;
+    } else if (rhs.set_background_color) {
+      if (!background_color.is_rgb || !rhs.background_color.is_rgb)
+        report_error("can't OR a terminal color");
+      background_color.value.rgb_color |= rhs.background_color.value.rgb_color;
+    }
+
+    ems = static_cast<emphasis>(static_cast<uint8_t>(ems) |
+                                static_cast<uint8_t>(rhs.ems));
+    return *this;
+  }
+
+  friend FMT_CONSTEXPR auto operator|(text_style lhs, const text_style& rhs)
+      -> text_style {
+    return lhs |= rhs;
+  }
+
+  FMT_CONSTEXPR auto has_foreground() const noexcept -> bool {
+    return set_foreground_color;
+  }
+  FMT_CONSTEXPR auto has_background() const noexcept -> bool {
+    return set_background_color;
+  }
+  FMT_CONSTEXPR auto has_emphasis() const noexcept -> bool {
+    return static_cast<uint8_t>(ems) != 0;
+  }
+  FMT_CONSTEXPR auto get_foreground() const noexcept -> detail::color_type {
+    FMT_ASSERT(has_foreground(), "no foreground specified for this style");
+    return foreground_color;
+  }
+  FMT_CONSTEXPR auto get_background() const noexcept -> detail::color_type {
+    FMT_ASSERT(has_background(), "no background specified for this style");
+    return background_color;
+  }
+  FMT_CONSTEXPR auto get_emphasis() const noexcept -> emphasis {
+    FMT_ASSERT(has_emphasis(), "no emphasis specified for this style");
+    return ems;
+  }
+
+ private:
+  FMT_CONSTEXPR text_style(bool is_foreground,
+                           detail::color_type text_color) noexcept
+      : set_foreground_color(), set_background_color(), ems() {
+    if (is_foreground) {
+      foreground_color = text_color;
+      set_foreground_color = true;
+    } else {
+      background_color = text_color;
+      set_background_color = true;
+    }
+  }
+
+  friend FMT_CONSTEXPR auto fg(detail::color_type foreground) noexcept
+      -> text_style;
+
+  friend FMT_CONSTEXPR auto bg(detail::color_type background) noexcept
+      -> text_style;
+
+  detail::color_type foreground_color;
+  detail::color_type background_color;
+  bool set_foreground_color;
+  bool set_background_color;
+  emphasis ems;
+};
+
+/// Creates a text style from the foreground (text) color.
+FMT_CONSTEXPR inline auto fg(detail::color_type foreground) noexcept
+    -> text_style {
+  return text_style(true, foreground);
+}
+
+/// Creates a text style from the background color.
+FMT_CONSTEXPR inline auto bg(detail::color_type background) noexcept
+    -> text_style {
+  return text_style(false, background);
+}
+
+FMT_CONSTEXPR inline auto operator|(emphasis lhs, emphasis rhs) noexcept
+    -> text_style {
+  return text_style(lhs) | rhs;
+}
+
+namespace detail {
+
+template <typename Char> struct ansi_color_escape {
+  FMT_CONSTEXPR ansi_color_escape(detail::color_type text_color,
+                                  const char* esc) noexcept {
+    // If we have a terminal color, we need to output another escape code
+    // sequence.
+    if (!text_color.is_rgb) {
+      bool is_background = esc == string_view("\x1b[48;2;");
+      uint32_t value = text_color.value.term_color;
+      // Background ASCII codes are the same as the foreground ones but with
+      // 10 more.
+      if (is_background) value += 10u;
+
+      size_t index = 0;
+      buffer[index++] = static_cast<Char>('\x1b');
+      buffer[index++] = static_cast<Char>('[');
+
+      if (value >= 100u) {
+        buffer[index++] = static_cast<Char>('1');
+        value %= 100u;
+      }
+      buffer[index++] = static_cast<Char>('0' + value / 10u);
+      buffer[index++] = static_cast<Char>('0' + value % 10u);
+
+      buffer[index++] = static_cast<Char>('m');
+      buffer[index++] = static_cast<Char>('\0');
+      return;
+    }
+
+    for (int i = 0; i < 7; i++) {
+      buffer[i] = static_cast<Char>(esc[i]);
+    }
+    rgb color(text_color.value.rgb_color);
+    to_esc(color.r, buffer + 7, ';');
+    to_esc(color.g, buffer + 11, ';');
+    to_esc(color.b, buffer + 15, 'm');
+    buffer[19] = static_cast<Char>(0);
+  }
+  FMT_CONSTEXPR ansi_color_escape(emphasis em) noexcept {
+    uint8_t em_codes[num_emphases] = {};
+    if (has_emphasis(em, emphasis::bold)) em_codes[0] = 1;
+    if (has_emphasis(em, emphasis::faint)) em_codes[1] = 2;
+    if (has_emphasis(em, emphasis::italic)) em_codes[2] = 3;
+    if (has_emphasis(em, emphasis::underline)) em_codes[3] = 4;
+    if (has_emphasis(em, emphasis::blink)) em_codes[4] = 5;
+    if (has_emphasis(em, emphasis::reverse)) em_codes[5] = 7;
+    if (has_emphasis(em, emphasis::conceal)) em_codes[6] = 8;
+    if (has_emphasis(em, emphasis::strikethrough)) em_codes[7] = 9;
+
+    size_t index = 0;
+    for (size_t i = 0; i < num_emphases; ++i) {
+      if (!em_codes[i]) continue;
+      buffer[index++] = static_cast<Char>('\x1b');
+      buffer[index++] = static_cast<Char>('[');
+      buffer[index++] = static_cast<Char>('0' + em_codes[i]);
+      buffer[index++] = static_cast<Char>('m');
+    }
+    buffer[index++] = static_cast<Char>(0);
+  }
+  FMT_CONSTEXPR operator const Char*() const noexcept { return buffer; }
+
+  FMT_CONSTEXPR auto begin() const noexcept -> const Char* { return buffer; }
+  FMT_CONSTEXPR20 auto end() const noexcept -> const Char* {
+    return buffer + basic_string_view<Char>(buffer).size();
+  }
+
+ private:
+  static constexpr size_t num_emphases = 8;
+  Char buffer[7u + 3u * num_emphases + 1u];
+
+  static FMT_CONSTEXPR void to_esc(uint8_t c, Char* out,
+                                   char delimiter) noexcept {
+    out[0] = static_cast<Char>('0' + c / 100);
+    out[1] = static_cast<Char>('0' + c / 10 % 10);
+    out[2] = static_cast<Char>('0' + c % 10);
+    out[3] = static_cast<Char>(delimiter);
+  }
+  static FMT_CONSTEXPR auto has_emphasis(emphasis em, emphasis mask) noexcept
+      -> bool {
+    return static_cast<uint8_t>(em) & static_cast<uint8_t>(mask);
+  }
+};
+
+template <typename Char>
+FMT_CONSTEXPR auto make_foreground_color(detail::color_type foreground) noexcept
+    -> ansi_color_escape<Char> {
+  return ansi_color_escape<Char>(foreground, "\x1b[38;2;");
+}
+
+template <typename Char>
+FMT_CONSTEXPR auto make_background_color(detail::color_type background) noexcept
+    -> ansi_color_escape<Char> {
+  return ansi_color_escape<Char>(background, "\x1b[48;2;");
+}
+
+template <typename Char>
+FMT_CONSTEXPR auto make_emphasis(emphasis em) noexcept
+    -> ansi_color_escape<Char> {
+  return ansi_color_escape<Char>(em);
+}
+
+template <typename Char> inline void reset_color(buffer<Char>& buffer) {
+  auto reset_color = string_view("\x1b[0m");
+  buffer.append(reset_color.begin(), reset_color.end());
+}
+
+template <typename T> struct styled_arg : detail::view {
+  const T& value;
+  text_style style;
+  styled_arg(const T& v, text_style s) : value(v), style(s) {}
+};
+
+template <typename Char>
+void vformat_to(
+    buffer<Char>& buf, const text_style& ts, basic_string_view<Char> format_str,
+    basic_format_args<buffered_context<type_identity_t<Char>>> args) {
+  bool has_style = false;
+  if (ts.has_emphasis()) {
+    has_style = true;
+    auto emphasis = detail::make_emphasis<Char>(ts.get_emphasis());
+    buf.append(emphasis.begin(), emphasis.end());
+  }
+  if (ts.has_foreground()) {
+    has_style = true;
+    auto foreground = detail::make_foreground_color<Char>(ts.get_foreground());
+    buf.append(foreground.begin(), foreground.end());
+  }
+  if (ts.has_background()) {
+    has_style = true;
+    auto background = detail::make_background_color<Char>(ts.get_background());
+    buf.append(background.begin(), background.end());
+  }
+  detail::vformat_to(buf, format_str, args, {});
+  if (has_style) detail::reset_color<Char>(buf);
+}
+
+}  // namespace detail
+
+inline void vprint(FILE* f, const text_style& ts, string_view fmt,
+                   format_args args) {
+  auto buf = memory_buffer();
+  detail::vformat_to(buf, ts, fmt, args);
+  print(f, FMT_STRING("{}"), string_view(buf.begin(), buf.size()));
+}
+
+/**
+ * Formats a string and prints it to the specified file stream using ANSI
+ * escape sequences to specify text formatting.
+ *
+ * **Example**:
+ *
+ *     fmt::print(fmt::emphasis::bold | fg(fmt::color::red),
+ *                "Elapsed time: {0:.2f} seconds", 1.23);
+ */
+template <typename... T>
+void print(FILE* f, const text_style& ts, format_string<T...> fmt,
+           T&&... args) {
+  vprint(f, ts, fmt, fmt::make_format_args(args...));
+}
+
+/**
+ * Formats a string and prints it to stdout using ANSI escape sequences to
+ * specify text formatting.
+ *
+ * **Example**:
+ *
+ *     fmt::print(fmt::emphasis::bold | fg(fmt::color::red),
+ *                "Elapsed time: {0:.2f} seconds", 1.23);
+ */
+template <typename... T>
+void print(const text_style& ts, format_string<T...> fmt, T&&... args) {
+  return print(stdout, ts, fmt, std::forward<T>(args)...);
+}
+
+inline auto vformat(const text_style& ts, string_view fmt, format_args args)
+    -> std::string {
+  auto buf = memory_buffer();
+  detail::vformat_to(buf, ts, fmt, args);
+  return fmt::to_string(buf);
+}
+
+/**
+ * Formats arguments and returns the result as a string using ANSI escape
+ * sequences to specify text formatting.
+ *
+ * **Example**:
+ *
+ * ```
+ * #include <fmt/color.h>
+ * std::string message = fmt::format(fmt::emphasis::bold | fg(fmt::color::red),
+ *                                   "The answer is {}", 42);
+ * ```
+ */
+template <typename... T>
+inline auto format(const text_style& ts, format_string<T...> fmt, T&&... args)
+    -> std::string {
+  return fmt::vformat(ts, fmt, fmt::make_format_args(args...));
+}
+
+/// Formats a string with the given text_style and writes the output to `out`.
+template <typename OutputIt,
+          FMT_ENABLE_IF(detail::is_output_iterator<OutputIt, char>::value)>
+auto vformat_to(OutputIt out, const text_style& ts, string_view fmt,
+                format_args args) -> OutputIt {
+  auto&& buf = detail::get_buffer<char>(out);
+  detail::vformat_to(buf, ts, fmt, args);
+  return detail::get_iterator(buf, out);
+}
+
+/**
+ * Formats arguments with the given text style, writes the result to the output
+ * iterator `out` and returns the iterator past the end of the output range.
+ *
+ * **Example**:
+ *
+ *     std::vector<char> out;
+ *     fmt::format_to(std::back_inserter(out),
+ *                    fmt::emphasis::bold | fg(fmt::color::red), "{}", 42);
+ */
+template <typename OutputIt, typename... T,
+          FMT_ENABLE_IF(detail::is_output_iterator<OutputIt, char>::value)>
+inline auto format_to(OutputIt out, const text_style& ts,
+                      format_string<T...> fmt, T&&... args) -> OutputIt {
+  return vformat_to(out, ts, fmt, fmt::make_format_args(args...));
+}
+
+template <typename T, typename Char>
+struct formatter<detail::styled_arg<T>, Char> : formatter<T, Char> {
+  template <typename FormatContext>
+  auto format(const detail::styled_arg<T>& arg, FormatContext& ctx) const
+      -> decltype(ctx.out()) {
+    const auto& ts = arg.style;
+    const auto& value = arg.value;
+    auto out = ctx.out();
+
+    bool has_style = false;
+    if (ts.has_emphasis()) {
+      has_style = true;
+      auto emphasis = detail::make_emphasis<Char>(ts.get_emphasis());
+      out = std::copy(emphasis.begin(), emphasis.end(), out);
+    }
+    if (ts.has_foreground()) {
+      has_style = true;
+      auto foreground =
+          detail::make_foreground_color<Char>(ts.get_foreground());
+      out = std::copy(foreground.begin(), foreground.end(), out);
+    }
+    if (ts.has_background()) {
+      has_style = true;
+      auto background =
+          detail::make_background_color<Char>(ts.get_background());
+      out = std::copy(background.begin(), background.end(), out);
+    }
+    out = formatter<T, Char>::format(value, ctx);
+    if (has_style) {
+      auto reset_color = string_view("\x1b[0m");
+      out = std::copy(reset_color.begin(), reset_color.end(), out);
+    }
+    return out;
+  }
+};
+
+/**
+ * Returns an argument that will be formatted using ANSI escape sequences,
+ * to be used in a formatting function.
+ *
+ * **Example**:
+ *
+ *     fmt::print("Elapsed time: {0:.2f} seconds",
+ *                fmt::styled(1.23, fmt::fg(fmt::color::green) |
+ *                                  fmt::bg(fmt::color::blue)));
+ */
+template <typename T>
+FMT_CONSTEXPR auto styled(const T& value, text_style ts)
+    -> detail::styled_arg<remove_cvref_t<T>> {
+  return detail::styled_arg<remove_cvref_t<T>>{value, ts};
+}
+
+FMT_END_EXPORT
+FMT_END_NAMESPACE
+
+#endif  // FMT_COLOR_H_
diff --git a/ext/spdlog/fmt/bundled/compile.h b/ext/spdlog/fmt/bundled/compile.h
new file mode 100644
index 0000000..b2afc2c
--- /dev/null
+++ b/ext/spdlog/fmt/bundled/compile.h
@@ -0,0 +1,529 @@
+// Formatting library for C++ - experimental format string compilation
+//
+// Copyright (c) 2012 - present, Victor Zverovich and fmt contributors
+// All rights reserved.
+//
+// For the license information refer to format.h.
+
+#ifndef FMT_COMPILE_H_
+#define FMT_COMPILE_H_
+
+#ifndef FMT_MODULE
+#  include <iterator>  // std::back_inserter
+#endif
+
+#include "format.h"
+
+FMT_BEGIN_NAMESPACE
+
+// A compile-time string which is compiled into fast formatting code.
+FMT_EXPORT class compiled_string {};
+
+namespace detail {
+
+template <typename T, typename InputIt>
+FMT_CONSTEXPR inline auto copy(InputIt begin, InputIt end, counting_iterator it)
+    -> counting_iterator {
+  return it + (end - begin);
+}
+
+template <typename S>
+struct is_compiled_string : std::is_base_of<compiled_string, S> {};
+
+/**
+ * Converts a string literal `s` into a format string that will be parsed at
+ * compile time and converted into efficient formatting code. Requires C++17
+ * `constexpr if` compiler support.
+ *
+ * **Example**:
+ *
+ *     // Converts 42 into std::string using the most efficient method and no
+ *     // runtime format string processing.
+ *     std::string s = fmt::format(FMT_COMPILE("{}"), 42);
+ */
+#if defined(__cpp_if_constexpr) && defined(__cpp_return_type_deduction)
+#  define FMT_COMPILE(s) FMT_STRING_IMPL(s, fmt::compiled_string, explicit)
+#else
+#  define FMT_COMPILE(s) FMT_STRING(s)
+#endif
+
+#if FMT_USE_NONTYPE_TEMPLATE_ARGS
+template <typename Char, size_t N,
+          fmt::detail_exported::fixed_string<Char, N> Str>
+struct udl_compiled_string : compiled_string {
+  using char_type = Char;
+  explicit constexpr operator basic_string_view<char_type>() const {
+    return {Str.data, N - 1};
+  }
+};
+#endif
+
+template <typename T, typename... Tail>
+auto first(const T& value, const Tail&...) -> const T& {
+  return value;
+}
+
+#if defined(__cpp_if_constexpr) && defined(__cpp_return_type_deduction)
+template <typename... Args> struct type_list {};
+
+// Returns a reference to the argument at index N from [first, rest...].
+template <int N, typename T, typename... Args>
+constexpr const auto& get([[maybe_unused]] const T& first,
+                          [[maybe_unused]] const Args&... rest) {
+  static_assert(N < 1 + sizeof...(Args), "index is out of bounds");
+  if constexpr (N == 0)
+    return first;
+  else
+    return detail::get<N - 1>(rest...);
+}
+
+template <typename Char, typename... Args>
+constexpr int get_arg_index_by_name(basic_string_view<Char> name,
+                                    type_list<Args...>) {
+  return get_arg_index_by_name<Args...>(name);
+}
+
+template <int N, typename> struct get_type_impl;
+
+template <int N, typename... Args> struct get_type_impl<N, type_list<Args...>> {
+  using type =
+      remove_cvref_t<decltype(detail::get<N>(std::declval<Args>()...))>;
+};
+
+template <int N, typename T>
+using get_type = typename get_type_impl<N, T>::type;
+
+template <typename T> struct is_compiled_format : std::false_type {};
+
+template <typename Char> struct text {
+  basic_string_view<Char> data;
+  using char_type = Char;
+
+  template <typename OutputIt, typename... Args>
+  constexpr OutputIt format(OutputIt out, const Args&...) const {
+    return write<Char>(out, data);
+  }
+};
+
+template <typename Char>
+struct is_compiled_format<text<Char>> : std::true_type {};
+
+template <typename Char>
+constexpr text<Char> make_text(basic_string_view<Char> s, size_t pos,
+                               size_t size) {
+  return {{&s[pos], size}};
+}
+
+template <typename Char> struct code_unit {
+  Char value;
+  using char_type = Char;
+
+  template <typename OutputIt, typename... Args>
+  constexpr OutputIt format(OutputIt out, const Args&...) const {
+    *out++ = value;
+    return out;
+  }
+};
+
+// This ensures that the argument type is convertible to `const T&`.
+template <typename T, int N, typename... Args>
+constexpr const T& get_arg_checked(const Args&... args) {
+  const auto& arg = detail::get<N>(args...);
+  if constexpr (detail::is_named_arg<remove_cvref_t<decltype(arg)>>()) {
+    return arg.value;
+  } else {
+    return arg;
+  }
+}
+
+template <typename Char>
+struct is_compiled_format<code_unit<Char>> : std::true_type {};
+
+// A replacement field that refers to argument N.
+template <typename Char, typename T, int N> struct field {
+  using char_type = Char;
+
+  template <typename OutputIt, typename... Args>
+  constexpr OutputIt format(OutputIt out, const Args&... args) const {
+    const T& arg = get_arg_checked<T, N>(args...);
+    if constexpr (std::is_convertible<T, basic_string_view<Char>>::value) {
+      auto s = basic_string_view<Char>(arg);
+      return copy<Char>(s.begin(), s.end(), out);
+    }
+    return write<Char>(out, arg);
+  }
+};
+
+template <typename Char, typename T, int N>
+struct is_compiled_format<field<Char, T, N>> : std::true_type {};
+
+// A replacement field that refers to argument with name.
+template <typename Char> struct runtime_named_field {
+  using char_type = Char;
+  basic_string_view<Char> name;
+
+  template <typename OutputIt, typename T>
+  constexpr static bool try_format_argument(
+      OutputIt& out,
+      // [[maybe_unused]] due to unused-but-set-parameter warning in GCC 7,8,9
+      [[maybe_unused]] basic_string_view<Char> arg_name, const T& arg) {
+    if constexpr (is_named_arg<typename std::remove_cv<T>::type>::value) {
+      if (arg_name == arg.name) {
+        out = write<Char>(out, arg.value);
+        return true;
+      }
+    }
+    return false;
+  }
+
+  template <typename OutputIt, typename... Args>
+  constexpr OutputIt format(OutputIt out, const Args&... args) const {
+    bool found = (try_format_argument(out, name, args) || ...);
+    if (!found) {
+      FMT_THROW(format_error("argument with specified name is not found"));
+    }
+    return out;
+  }
+};
+
+template <typename Char>
+struct is_compiled_format<runtime_named_field<Char>> : std::true_type {};
+
+// A replacement field that refers to argument N and has format specifiers.
+template <typename Char, typename T, int N> struct spec_field {
+  using char_type = Char;
+  formatter<T, Char> fmt;
+
+  template <typename OutputIt, typename... Args>
+  constexpr FMT_INLINE OutputIt format(OutputIt out,
+                                       const Args&... args) const {
+    const auto& vargs =
+        fmt::make_format_args<basic_format_context<OutputIt, Char>>(args...);
+    basic_format_context<OutputIt, Char> ctx(out, vargs);
+    return fmt.format(get_arg_checked<T, N>(args...), ctx);
+  }
+};
+
+template <typename Char, typename T, int N>
+struct is_compiled_format<spec_field<Char, T, N>> : std::true_type {};
+
+template <typename L, typename R> struct concat {
+  L lhs;
+  R rhs;
+  using char_type = typename L::char_type;
+
+  template <typename OutputIt, typename... Args>
+  constexpr OutputIt format(OutputIt out, const Args&... args) const {
+    out = lhs.format(out, args...);
+    return rhs.format(out, args...);
+  }
+};
+
+template <typename L, typename R>
+struct is_compiled_format<concat<L, R>> : std::true_type {};
+
+template <typename L, typename R>
+constexpr concat<L, R> make_concat(L lhs, R rhs) {
+  return {lhs, rhs};
+}
+
+struct unknown_format {};
+
+template <typename Char>
+constexpr size_t parse_text(basic_string_view<Char> str, size_t pos) {
+  for (size_t size = str.size(); pos != size; ++pos) {
+    if (str[pos] == '{' || str[pos] == '}') break;
+  }
+  return pos;
+}
+
+template <typename Args, size_t POS, int ID, typename S>
+constexpr auto compile_format_string(S fmt);
+
+template <typename Args, size_t POS, int ID, typename T, typename S>
+constexpr auto parse_tail(T head, S fmt) {
+  if constexpr (POS != basic_string_view<typename S::char_type>(fmt).size()) {
+    constexpr auto tail = compile_format_string<Args, POS, ID>(fmt);
+    if constexpr (std::is_same<remove_cvref_t<decltype(tail)>,
+                               unknown_format>())
+      return tail;
+    else
+      return make_concat(head, tail);
+  } else {
+    return head;
+  }
+}
+
+template <typename T, typename Char> struct parse_specs_result {
+  formatter<T, Char> fmt;
+  size_t end;
+  int next_arg_id;
+};
+
+enum { manual_indexing_id = -1 };
+
+template <typename T, typename Char>
+constexpr parse_specs_result<T, Char> parse_specs(basic_string_view<Char> str,
+                                                  size_t pos, int next_arg_id) {
+  str.remove_prefix(pos);
+  auto ctx =
+      compile_parse_context<Char>(str, max_value<int>(), nullptr, next_arg_id);
+  auto f = formatter<T, Char>();
+  auto end = f.parse(ctx);
+  return {f, pos + fmt::detail::to_unsigned(end - str.data()),
+          next_arg_id == 0 ? manual_indexing_id : ctx.next_arg_id()};
+}
+
+template <typename Char> struct arg_id_handler {
+  arg_ref<Char> arg_id;
+
+  constexpr int on_auto() {
+    FMT_ASSERT(false, "handler cannot be used with automatic indexing");
+    return 0;
+  }
+  constexpr int on_index(int id) {
+    arg_id = arg_ref<Char>(id);
+    return 0;
+  }
+  constexpr int on_name(basic_string_view<Char> id) {
+    arg_id = arg_ref<Char>(id);
+    return 0;
+  }
+};
+
+template <typename Char> struct parse_arg_id_result {
+  arg_ref<Char> arg_id;
+  const Char* arg_id_end;
+};
+
+template <int ID, typename Char>
+constexpr auto parse_arg_id(const Char* begin, const Char* end) {
+  auto handler = arg_id_handler<Char>{arg_ref<Char>{}};
+  auto arg_id_end = parse_arg_id(begin, end, handler);
+  return parse_arg_id_result<Char>{handler.arg_id, arg_id_end};
+}
+
+template <typename T, typename Enable = void> struct field_type {
+  using type = remove_cvref_t<T>;
+};
+
+template <typename T>
+struct field_type<T, enable_if_t<detail::is_named_arg<T>::value>> {
+  using type = remove_cvref_t<decltype(T::value)>;
+};
+
+template <typename T, typename Args, size_t END_POS, int ARG_INDEX, int NEXT_ID,
+          typename S>
+constexpr auto parse_replacement_field_then_tail(S fmt) {
+  using char_type = typename S::char_type;
+  constexpr auto str = basic_string_view<char_type>(fmt);
+  constexpr char_type c = END_POS != str.size() ? str[END_POS] : char_type();
+  if constexpr (c == '}') {
+    return parse_tail<Args, END_POS + 1, NEXT_ID>(
+        field<char_type, typename field_type<T>::type, ARG_INDEX>(), fmt);
+  } else if constexpr (c != ':') {
+    FMT_THROW(format_error("expected ':'"));
+  } else {
+    constexpr auto result = parse_specs<typename field_type<T>::type>(
+        str, END_POS + 1, NEXT_ID == manual_indexing_id ? 0 : NEXT_ID);
+    if constexpr (result.end >= str.size() || str[result.end] != '}') {
+      FMT_THROW(format_error("expected '}'"));
+      return 0;
+    } else {
+      return parse_tail<Args, result.end + 1, result.next_arg_id>(
+          spec_field<char_type, typename field_type<T>::type, ARG_INDEX>{
+              result.fmt},
+          fmt);
+    }
+  }
+}
+
+// Compiles a non-empty format string and returns the compiled representation
+// or unknown_format() on unrecognized input.
+template <typename Args, size_t POS, int ID, typename S>
+constexpr auto compile_format_string(S fmt) {
+  using char_type = typename S::char_type;
+  constexpr auto str = basic_string_view<char_type>(fmt);
+  if constexpr (str[POS] == '{') {
+    if constexpr (POS + 1 == str.size())
+      FMT_THROW(format_error("unmatched '{' in format string"));
+    if constexpr (str[POS + 1] == '{') {
+      return parse_tail<Args, POS + 2, ID>(make_text(str, POS, 1), fmt);
+    } else if constexpr (str[POS + 1] == '}' || str[POS + 1] == ':') {
+      static_assert(ID != manual_indexing_id,
+                    "cannot switch from manual to automatic argument indexing");
+      constexpr auto next_id =
+          ID != manual_indexing_id ? ID + 1 : manual_indexing_id;
+      return parse_replacement_field_then_tail<get_type<ID, Args>, Args,
+                                               POS + 1, ID, next_id>(fmt);
+    } else {
+      constexpr auto arg_id_result =
+          parse_arg_id<ID>(str.data() + POS + 1, str.data() + str.size());
+      constexpr auto arg_id_end_pos = arg_id_result.arg_id_end - str.data();
+      constexpr char_type c =
+          arg_id_end_pos != str.size() ? str[arg_id_end_pos] : char_type();
+      static_assert(c == '}' || c == ':', "missing '}' in format string");
+      if constexpr (arg_id_result.arg_id.kind == arg_id_kind::index) {
+        static_assert(
+            ID == manual_indexing_id || ID == 0,
+            "cannot switch from automatic to manual argument indexing");
+        constexpr auto arg_index = arg_id_result.arg_id.val.index;
+        return parse_replacement_field_then_tail<get_type<arg_index, Args>,
+                                                 Args, arg_id_end_pos,
+                                                 arg_index, manual_indexing_id>(
+            fmt);
+      } else if constexpr (arg_id_result.arg_id.kind == arg_id_kind::name) {
+        constexpr auto arg_index =
+            get_arg_index_by_name(arg_id_result.arg_id.val.name, Args{});
+        if constexpr (arg_index >= 0) {
+          constexpr auto next_id =
+              ID != manual_indexing_id ? ID + 1 : manual_indexing_id;
+          return parse_replacement_field_then_tail<
+              decltype(get_type<arg_index, Args>::value), Args, arg_id_end_pos,
+              arg_index, next_id>(fmt);
+        } else if constexpr (c == '}') {
+          return parse_tail<Args, arg_id_end_pos + 1, ID>(
+              runtime_named_field<char_type>{arg_id_result.arg_id.val.name},
+              fmt);
+        } else if constexpr (c == ':') {
+          return unknown_format();  // no type info for specs parsing
+        }
+      }
+    }
+  } else if constexpr (str[POS] == '}') {
+    if constexpr (POS + 1 == str.size())
+      FMT_THROW(format_error("unmatched '}' in format string"));
+    return parse_tail<Args, POS + 2, ID>(make_text(str, POS, 1), fmt);
+  } else {
+    constexpr auto end = parse_text(str, POS + 1);
+    if constexpr (end - POS > 1) {
+      return parse_tail<Args, end, ID>(make_text(str, POS, end - POS), fmt);
+    } else {
+      return parse_tail<Args, end, ID>(code_unit<char_type>{str[POS]}, fmt);
+    }
+  }
+}
+
+template <typename... Args, typename S,
+          FMT_ENABLE_IF(detail::is_compiled_string<S>::value)>
+constexpr auto compile(S fmt) {
+  constexpr auto str = basic_string_view<typename S::char_type>(fmt);
+  if constexpr (str.size() == 0) {
+    return detail::make_text(str, 0, 0);
+  } else {
+    constexpr auto result =
+        detail::compile_format_string<detail::type_list<Args...>, 0, 0>(fmt);
+    return result;
+  }
+}
+#endif  // defined(__cpp_if_constexpr) && defined(__cpp_return_type_deduction)
+}  // namespace detail
+
+FMT_BEGIN_EXPORT
+
+#if defined(__cpp_if_constexpr) && defined(__cpp_return_type_deduction)
+
+template <typename CompiledFormat, typename... Args,
+          typename Char = typename CompiledFormat::char_type,
+          FMT_ENABLE_IF(detail::is_compiled_format<CompiledFormat>::value)>
+FMT_INLINE std::basic_string<Char> format(const CompiledFormat& cf,
+                                          const Args&... args) {
+  auto s = std::basic_string<Char>();
+  cf.format(std::back_inserter(s), args...);
+  return s;
+}
+
+template <typename OutputIt, typename CompiledFormat, typename... Args,
+          FMT_ENABLE_IF(detail::is_compiled_format<CompiledFormat>::value)>
+constexpr FMT_INLINE OutputIt format_to(OutputIt out, const CompiledFormat& cf,
+                                        const Args&... args) {
+  return cf.format(out, args...);
+}
+
+template <typename S, typename... Args,
+          FMT_ENABLE_IF(detail::is_compiled_string<S>::value)>
+FMT_INLINE std::basic_string<typename S::char_type> format(const S&,
+                                                           Args&&... args) {
+  if constexpr (std::is_same<typename S::char_type, char>::value) {
+    constexpr auto str = basic_string_view<typename S::char_type>(S());
+    if constexpr (str.size() == 2 && str[0] == '{' && str[1] == '}') {
+      const auto& first = detail::first(args...);
+      if constexpr (detail::is_named_arg<
+                        remove_cvref_t<decltype(first)>>::value) {
+        return fmt::to_string(first.value);
+      } else {
+        return fmt::to_string(first);
+      }
+    }
+  }
+  constexpr auto compiled = detail::compile<Args...>(S());
+  if constexpr (std::is_same<remove_cvref_t<decltype(compiled)>,
+                             detail::unknown_format>()) {
+    return fmt::format(
+        static_cast<basic_string_view<typename S::char_type>>(S()),
+        std::forward<Args>(args)...);
+  } else {
+    return fmt::format(compiled, std::forward<Args>(args)...);
+  }
+}
+
+template <typename OutputIt, typename S, typename... Args,
+          FMT_ENABLE_IF(detail::is_compiled_string<S>::value)>
+FMT_CONSTEXPR OutputIt format_to(OutputIt out, const S&, Args&&... args) {
+  constexpr auto compiled = detail::compile<Args...>(S());
+  if constexpr (std::is_same<remove_cvref_t<decltype(compiled)>,
+                             detail::unknown_format>()) {
+    return fmt::format_to(
+        out, static_cast<basic_string_view<typename S::char_type>>(S()),
+        std::forward<Args>(args)...);
+  } else {
+    return fmt::format_to(out, compiled, std::forward<Args>(args)...);
+  }
+}
+#endif
+
+template <typename OutputIt, typename S, typename... Args,
+          FMT_ENABLE_IF(detail::is_compiled_string<S>::value)>
+auto format_to_n(OutputIt out, size_t n, const S& fmt, Args&&... args)
+    -> format_to_n_result<OutputIt> {
+  using traits = detail::fixed_buffer_traits;
+  auto buf = detail::iterator_buffer<OutputIt, char, traits>(out, n);
+  fmt::format_to(std::back_inserter(buf), fmt, std::forward<Args>(args)...);
+  return {buf.out(), buf.count()};
+}
+
+template <typename S, typename... Args,
+          FMT_ENABLE_IF(detail::is_compiled_string<S>::value)>
+FMT_CONSTEXPR20 auto formatted_size(const S& fmt, const Args&... args)
+    -> size_t {
+  return fmt::format_to(detail::counting_iterator(), fmt, args...).count();
+}
+
+template <typename S, typename... Args,
+          FMT_ENABLE_IF(detail::is_compiled_string<S>::value)>
+void print(std::FILE* f, const S& fmt, const Args&... args) {
+  memory_buffer buffer;
+  fmt::format_to(std::back_inserter(buffer), fmt, args...);
+  detail::print(f, {buffer.data(), buffer.size()});
+}
+
+template <typename S, typename... Args,
+          FMT_ENABLE_IF(detail::is_compiled_string<S>::value)>
+void print(const S& fmt, const Args&... args) {
+  print(stdout, fmt, args...);
+}
+
+#if FMT_USE_NONTYPE_TEMPLATE_ARGS
+inline namespace literals {
+template <detail_exported::fixed_string Str> constexpr auto operator""_cf() {
+  using char_t = remove_cvref_t<decltype(Str.data[0])>;
+  return detail::udl_compiled_string<char_t, sizeof(Str.data) / sizeof(char_t),
+                                     Str>();
+}
+}  // namespace literals
+#endif
+
+FMT_END_EXPORT
+FMT_END_NAMESPACE
+
+#endif  // FMT_COMPILE_H_
diff --git a/ext/spdlog/fmt/bundled/core.h b/ext/spdlog/fmt/bundled/core.h
new file mode 100644
index 0000000..8ca735f
--- /dev/null
+++ b/ext/spdlog/fmt/bundled/core.h
@@ -0,0 +1,5 @@
+// This file is only provided for compatibility and may be removed in future
+// versions. Use fmt/base.h if you don't need fmt::format and fmt/format.h
+// otherwise.
+
+#include "format.h"
diff --git a/ext/spdlog/fmt/bundled/fmt.license.rst b/ext/spdlog/fmt/bundled/fmt.license.rst
new file mode 100644
index 0000000..f0ec3db
--- /dev/null
+++ b/ext/spdlog/fmt/bundled/fmt.license.rst
@@ -0,0 +1,27 @@
+Copyright (c) 2012 - present, Victor Zverovich
+
+Permission is hereby granted, free of charge, to any person obtaining
+a copy of this software and associated documentation files (the
+"Software"), to deal in the Software without restriction, including
+without limitation the rights to use, copy, modify, merge, publish,
+distribute, sublicense, and/or sell copies of the Software, and to
+permit persons to whom the Software is furnished to do so, subject to
+the following conditions:
+
+The above copyright notice and this permission notice shall be
+included in all copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE
+LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
+OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
+WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+
+--- Optional exception to the license ---
+
+As an exception, if, as a result of your compiling your source code, portions
+of this Software are embedded into a machine-executable object form of such
+source code, you may redistribute such embedded portions in such object form
+without including the above copyright and permission notices.
diff --git a/ext/spdlog/fmt/bundled/format-inl.h b/ext/spdlog/fmt/bundled/format-inl.h
new file mode 100644
index 0000000..a887483
--- /dev/null
+++ b/ext/spdlog/fmt/bundled/format-inl.h
@@ -0,0 +1,1928 @@
+// Formatting library for C++ - implementation
+//
+// Copyright (c) 2012 - 2016, Victor Zverovich
+// All rights reserved.
+//
+// For the license information refer to format.h.
+
+#ifndef FMT_FORMAT_INL_H_
+#define FMT_FORMAT_INL_H_
+
+#ifndef FMT_MODULE
+#  include <algorithm>
+#  include <cerrno>  // errno
+#  include <climits>
+#  include <cmath>
+#  include <exception>
+
+#  if !defined(FMT_STATIC_THOUSANDS_SEPARATOR)
+#    include <locale>
+#  endif
+#endif
+
+#if defined(_WIN32) && !defined(FMT_USE_WRITE_CONSOLE)
+#  include <io.h>  // _isatty
+#endif
+
+#include "format.h"
+
+FMT_BEGIN_NAMESPACE
+namespace detail {
+
+FMT_FUNC void assert_fail(const char* file, int line, const char* message) {
+  // Use unchecked std::fprintf to avoid triggering another assertion when
+  // writing to stderr fails
+  std::fprintf(stderr, "%s:%d: assertion failed: %s", file, line, message);
+  // Chosen instead of std::abort to satisfy Clang in CUDA mode during device
+  // code pass.
+  std::terminate();
+}
+
+FMT_FUNC void format_error_code(detail::buffer<char>& out, int error_code,
+                                string_view message) noexcept {
+  // Report error code making sure that the output fits into
+  // inline_buffer_size to avoid dynamic memory allocation and potential
+  // bad_alloc.
+  out.try_resize(0);
+  static const char SEP[] = ": ";
+  static const char ERROR_STR[] = "error ";
+  // Subtract 2 to account for terminating null characters in SEP and ERROR_STR.
+  size_t error_code_size = sizeof(SEP) + sizeof(ERROR_STR) - 2;
+  auto abs_value = static_cast<uint32_or_64_or_128_t<int>>(error_code);
+  if (detail::is_negative(error_code)) {
+    abs_value = 0 - abs_value;
+    ++error_code_size;
+  }
+  error_code_size += detail::to_unsigned(detail::count_digits(abs_value));
+  auto it = appender(out);
+  if (message.size() <= inline_buffer_size - error_code_size)
+    fmt::format_to(it, FMT_STRING("{}{}"), message, SEP);
+  fmt::format_to(it, FMT_STRING("{}{}"), ERROR_STR, error_code);
+  FMT_ASSERT(out.size() <= inline_buffer_size, "");
+}
+
+FMT_FUNC void report_error(format_func func, int error_code,
+                           const char* message) noexcept {
+  memory_buffer full_message;
+  func(full_message, error_code, message);
+  // Don't use fwrite_fully because the latter may throw.
+  if (std::fwrite(full_message.data(), full_message.size(), 1, stderr) > 0)
+    std::fputc('\n', stderr);
+}
+
+// A wrapper around fwrite that throws on error.
+inline void fwrite_fully(const void* ptr, size_t count, FILE* stream) {
+  size_t written = std::fwrite(ptr, 1, count, stream);
+  if (written < count)
+    FMT_THROW(system_error(errno, FMT_STRING("cannot write to file")));
+}
+
+#ifndef FMT_STATIC_THOUSANDS_SEPARATOR
+template <typename Locale>
+locale_ref::locale_ref(const Locale& loc) : locale_(&loc) {
+  static_assert(std::is_same<Locale, std::locale>::value, "");
+}
+
+template <typename Locale> auto locale_ref::get() const -> Locale {
+  static_assert(std::is_same<Locale, std::locale>::value, "");
+  return locale_ ? *static_cast<const std::locale*>(locale_) : std::locale();
+}
+
+template <typename Char>
+FMT_FUNC auto thousands_sep_impl(locale_ref loc) -> thousands_sep_result<Char> {
+  auto& facet = std::use_facet<std::numpunct<Char>>(loc.get<std::locale>());
+  auto grouping = facet.grouping();
+  auto thousands_sep = grouping.empty() ? Char() : facet.thousands_sep();
+  return {std::move(grouping), thousands_sep};
+}
+template <typename Char>
+FMT_FUNC auto decimal_point_impl(locale_ref loc) -> Char {
+  return std::use_facet<std::numpunct<Char>>(loc.get<std::locale>())
+      .decimal_point();
+}
+#else
+template <typename Char>
+FMT_FUNC auto thousands_sep_impl(locale_ref) -> thousands_sep_result<Char> {
+  return {"\03", FMT_STATIC_THOUSANDS_SEPARATOR};
+}
+template <typename Char> FMT_FUNC Char decimal_point_impl(locale_ref) {
+  return '.';
+}
+#endif
+
+FMT_FUNC auto write_loc(appender out, loc_value value,
+                        const format_specs& specs, locale_ref loc) -> bool {
+#ifdef FMT_STATIC_THOUSANDS_SEPARATOR
+  value.visit(loc_writer<>{
+      out, specs, std::string(1, FMT_STATIC_THOUSANDS_SEPARATOR), "\3", "."});
+  return true;
+#else
+  auto locale = loc.get<std::locale>();
+  // We cannot use the num_put<char> facet because it may produce output in
+  // a wrong encoding.
+  using facet = format_facet<std::locale>;
+  if (std::has_facet<facet>(locale))
+    return std::use_facet<facet>(locale).put(out, value, specs);
+  return facet(locale).put(out, value, specs);
+#endif
+}
+}  // namespace detail
+
+FMT_FUNC void report_error(const char* message) {
+  FMT_THROW(format_error(message));
+}
+
+template <typename Locale> typename Locale::id format_facet<Locale>::id;
+
+#ifndef FMT_STATIC_THOUSANDS_SEPARATOR
+template <typename Locale> format_facet<Locale>::format_facet(Locale& loc) {
+  auto& numpunct = std::use_facet<std::numpunct<char>>(loc);
+  grouping_ = numpunct.grouping();
+  if (!grouping_.empty()) separator_ = std::string(1, numpunct.thousands_sep());
+}
+
+template <>
+FMT_API FMT_FUNC auto format_facet<std::locale>::do_put(
+    appender out, loc_value val, const format_specs& specs) const -> bool {
+  return val.visit(
+      detail::loc_writer<>{out, specs, separator_, grouping_, decimal_point_});
+}
+#endif
+
+FMT_FUNC auto vsystem_error(int error_code, string_view fmt, format_args args)
+    -> std::system_error {
+  auto ec = std::error_code(error_code, std::generic_category());
+  return std::system_error(ec, vformat(fmt, args));
+}
+
+namespace detail {
+
+template <typename F>
+inline auto operator==(basic_fp<F> x, basic_fp<F> y) -> bool {
+  return x.f == y.f && x.e == y.e;
+}
+
+// Compilers should be able to optimize this into the ror instruction.
+FMT_CONSTEXPR inline auto rotr(uint32_t n, uint32_t r) noexcept -> uint32_t {
+  r &= 31;
+  return (n >> r) | (n << (32 - r));
+}
+FMT_CONSTEXPR inline auto rotr(uint64_t n, uint32_t r) noexcept -> uint64_t {
+  r &= 63;
+  return (n >> r) | (n << (64 - r));
+}
+
+// Implementation of Dragonbox algorithm: https://github.com/jk-jeon/dragonbox.
+namespace dragonbox {
+// Computes upper 64 bits of multiplication of a 32-bit unsigned integer and a
+// 64-bit unsigned integer.
+inline auto umul96_upper64(uint32_t x, uint64_t y) noexcept -> uint64_t {
+  return umul128_upper64(static_cast<uint64_t>(x) << 32, y);
+}
+
+// Computes lower 128 bits of multiplication of a 64-bit unsigned integer and a
+// 128-bit unsigned integer.
+inline auto umul192_lower128(uint64_t x, uint128_fallback y) noexcept
+    -> uint128_fallback {
+  uint64_t high = x * y.high();
+  uint128_fallback high_low = umul128(x, y.low());
+  return {high + high_low.high(), high_low.low()};
+}
+
+// Computes lower 64 bits of multiplication of a 32-bit unsigned integer and a
+// 64-bit unsigned integer.
+inline auto umul96_lower64(uint32_t x, uint64_t y) noexcept -> uint64_t {
+  return x * y;
+}
+
+// Various fast log computations.
+inline auto floor_log10_pow2_minus_log10_4_over_3(int e) noexcept -> int {
+  FMT_ASSERT(e <= 2936 && e >= -2985, "too large exponent");
+  return (e * 631305 - 261663) >> 21;
+}
+
+FMT_INLINE_VARIABLE constexpr struct {
+  uint32_t divisor;
+  int shift_amount;
+} div_small_pow10_infos[] = {{10, 16}, {100, 16}};
+
+// Replaces n by floor(n / pow(10, N)) returning true if and only if n is
+// divisible by pow(10, N).
+// Precondition: n <= pow(10, N + 1).
+template <int N>
+auto check_divisibility_and_divide_by_pow10(uint32_t& n) noexcept -> bool {
+  // The numbers below are chosen such that:
+  //   1. floor(n/d) = floor(nm / 2^k) where d=10 or d=100,
+  //   2. nm mod 2^k < m if and only if n is divisible by d,
+  // where m is magic_number, k is shift_amount
+  // and d is divisor.
+  //
+  // Item 1 is a common technique of replacing division by a constant with
+  // multiplication, see e.g. "Division by Invariant Integers Using
+  // Multiplication" by Granlund and Montgomery (1994). magic_number (m) is set
+  // to ceil(2^k/d) for large enough k.
+  // The idea for item 2 originates from Schubfach.
+  constexpr auto info = div_small_pow10_infos[N - 1];
+  FMT_ASSERT(n <= info.divisor * 10, "n is too large");
+  constexpr uint32_t magic_number =
+      (1u << info.shift_amount) / info.divisor + 1;
+  n *= magic_number;
+  const uint32_t comparison_mask = (1u << info.shift_amount) - 1;
+  bool result = (n & comparison_mask) < magic_number;
+  n >>= info.shift_amount;
+  return result;
+}
+
+// Computes floor(n / pow(10, N)) for small n and N.
+// Precondition: n <= pow(10, N + 1).
+template <int N> auto small_division_by_pow10(uint32_t n) noexcept -> uint32_t {
+  constexpr auto info = div_small_pow10_infos[N - 1];
+  FMT_ASSERT(n <= info.divisor * 10, "n is too large");
+  constexpr uint32_t magic_number =
+      (1u << info.shift_amount) / info.divisor + 1;
+  return (n * magic_number) >> info.shift_amount;
+}
+
+// Computes floor(n / 10^(kappa + 1)) (float)
+inline auto divide_by_10_to_kappa_plus_1(uint32_t n) noexcept -> uint32_t {
+  // 1374389535 = ceil(2^37/100)
+  return static_cast<uint32_t>((static_cast<uint64_t>(n) * 1374389535) >> 37);
+}
+// Computes floor(n / 10^(kappa + 1)) (double)
+inline auto divide_by_10_to_kappa_plus_1(uint64_t n) noexcept -> uint64_t {
+  // 2361183241434822607 = ceil(2^(64+7)/1000)
+  return umul128_upper64(n, 2361183241434822607ull) >> 7;
+}
+
+// Various subroutines using pow10 cache
+template <typename T> struct cache_accessor;
+
+template <> struct cache_accessor<float> {
+  using carrier_uint = float_info<float>::carrier_uint;
+  using cache_entry_type = uint64_t;
+
+  static auto get_cached_power(int k) noexcept -> uint64_t {
+    FMT_ASSERT(k >= float_info<float>::min_k && k <= float_info<float>::max_k,
+               "k is out of range");
+    static constexpr const uint64_t pow10_significands[] = {
+        0x81ceb32c4b43fcf5, 0xa2425ff75e14fc32, 0xcad2f7f5359a3b3f,
+        0xfd87b5f28300ca0e, 0x9e74d1b791e07e49, 0xc612062576589ddb,
+        0xf79687aed3eec552, 0x9abe14cd44753b53, 0xc16d9a0095928a28,
+        0xf1c90080baf72cb2, 0x971da05074da7bef, 0xbce5086492111aeb,
+        0xec1e4a7db69561a6, 0x9392ee8e921d5d08, 0xb877aa3236a4b44a,
+        0xe69594bec44de15c, 0x901d7cf73ab0acda, 0xb424dc35095cd810,
+        0xe12e13424bb40e14, 0x8cbccc096f5088cc, 0xafebff0bcb24aaff,
+        0xdbe6fecebdedd5bf, 0x89705f4136b4a598, 0xabcc77118461cefd,
+        0xd6bf94d5e57a42bd, 0x8637bd05af6c69b6, 0xa7c5ac471b478424,
+        0xd1b71758e219652c, 0x83126e978d4fdf3c, 0xa3d70a3d70a3d70b,
+        0xcccccccccccccccd, 0x8000000000000000, 0xa000000000000000,
+        0xc800000000000000, 0xfa00000000000000, 0x9c40000000000000,
+        0xc350000000000000, 0xf424000000000000, 0x9896800000000000,
+        0xbebc200000000000, 0xee6b280000000000, 0x9502f90000000000,
+        0xba43b74000000000, 0xe8d4a51000000000, 0x9184e72a00000000,
+        0xb5e620f480000000, 0xe35fa931a0000000, 0x8e1bc9bf04000000,
+        0xb1a2bc2ec5000000, 0xde0b6b3a76400000, 0x8ac7230489e80000,
+        0xad78ebc5ac620000, 0xd8d726b7177a8000, 0x878678326eac9000,
+        0xa968163f0a57b400, 0xd3c21bcecceda100, 0x84595161401484a0,
+        0xa56fa5b99019a5c8, 0xcecb8f27f4200f3a, 0x813f3978f8940985,
+        0xa18f07d736b90be6, 0xc9f2c9cd04674edf, 0xfc6f7c4045812297,
+        0x9dc5ada82b70b59e, 0xc5371912364ce306, 0xf684df56c3e01bc7,
+        0x9a130b963a6c115d, 0xc097ce7bc90715b4, 0xf0bdc21abb48db21,
+        0x96769950b50d88f5, 0xbc143fa4e250eb32, 0xeb194f8e1ae525fe,
+        0x92efd1b8d0cf37bf, 0xb7abc627050305ae, 0xe596b7b0c643c71a,
+        0x8f7e32ce7bea5c70, 0xb35dbf821ae4f38c, 0xe0352f62a19e306f};
+    return pow10_significands[k - float_info<float>::min_k];
+  }
+
+  struct compute_mul_result {
+    carrier_uint result;
+    bool is_integer;
+  };
+  struct compute_mul_parity_result {
+    bool parity;
+    bool is_integer;
+  };
+
+  static auto compute_mul(carrier_uint u,
+                          const cache_entry_type& cache) noexcept
+      -> compute_mul_result {
+    auto r = umul96_upper64(u, cache);
+    return {static_cast<carrier_uint>(r >> 32),
+            static_cast<carrier_uint>(r) == 0};
+  }
+
+  static auto compute_delta(const cache_entry_type& cache, int beta) noexcept
+      -> uint32_t {
+    return static_cast<uint32_t>(cache >> (64 - 1 - beta));
+  }
+
+  static auto compute_mul_parity(carrier_uint two_f,
+                                 const cache_entry_type& cache,
+                                 int beta) noexcept
+      -> compute_mul_parity_result {
+    FMT_ASSERT(beta >= 1, "");
+    FMT_ASSERT(beta < 64, "");
+
+    auto r = umul96_lower64(two_f, cache);
+    return {((r >> (64 - beta)) & 1) != 0,
+            static_cast<uint32_t>(r >> (32 - beta)) == 0};
+  }
+
+  static auto compute_left_endpoint_for_shorter_interval_case(
+      const cache_entry_type& cache, int beta) noexcept -> carrier_uint {
+    return static_cast<carrier_uint>(
+        (cache - (cache >> (num_significand_bits<float>() + 2))) >>
+        (64 - num_significand_bits<float>() - 1 - beta));
+  }
+
+  static auto compute_right_endpoint_for_shorter_interval_case(
+      const cache_entry_type& cache, int beta) noexcept -> carrier_uint {
+    return static_cast<carrier_uint>(
+        (cache + (cache >> (num_significand_bits<float>() + 1))) >>
+        (64 - num_significand_bits<float>() - 1 - beta));
+  }
+
+  static auto compute_round_up_for_shorter_interval_case(
+      const cache_entry_type& cache, int beta) noexcept -> carrier_uint {
+    return (static_cast<carrier_uint>(
+                cache >> (64 - num_significand_bits<float>() - 2 - beta)) +
+            1) /
+           2;
+  }
+};
+
+template <> struct cache_accessor<double> {
+  using carrier_uint = float_info<double>::carrier_uint;
+  using cache_entry_type = uint128_fallback;
+
+  static auto get_cached_power(int k) noexcept -> uint128_fallback {
+    FMT_ASSERT(k >= float_info<double>::min_k && k <= float_info<double>::max_k,
+               "k is out of range");
+
+    static constexpr const uint128_fallback pow10_significands[] = {
+#if FMT_USE_FULL_CACHE_DRAGONBOX
+      {0xff77b1fcbebcdc4f, 0x25e8e89c13bb0f7b},
+      {0x9faacf3df73609b1, 0x77b191618c54e9ad},
+      {0xc795830d75038c1d, 0xd59df5b9ef6a2418},
+      {0xf97ae3d0d2446f25, 0x4b0573286b44ad1e},
+      {0x9becce62836ac577, 0x4ee367f9430aec33},
+      {0xc2e801fb244576d5, 0x229c41f793cda740},
+      {0xf3a20279ed56d48a, 0x6b43527578c11110},
+      {0x9845418c345644d6, 0x830a13896b78aaaa},
+      {0xbe5691ef416bd60c, 0x23cc986bc656d554},
+      {0xedec366b11c6cb8f, 0x2cbfbe86b7ec8aa9},
+      {0x94b3a202eb1c3f39, 0x7bf7d71432f3d6aa},
+      {0xb9e08a83a5e34f07, 0xdaf5ccd93fb0cc54},
+      {0xe858ad248f5c22c9, 0xd1b3400f8f9cff69},
+      {0x91376c36d99995be, 0x23100809b9c21fa2},
+      {0xb58547448ffffb2d, 0xabd40a0c2832a78b},
+      {0xe2e69915b3fff9f9, 0x16c90c8f323f516d},
+      {0x8dd01fad907ffc3b, 0xae3da7d97f6792e4},
+      {0xb1442798f49ffb4a, 0x99cd11cfdf41779d},
+      {0xdd95317f31c7fa1d, 0x40405643d711d584},
+      {0x8a7d3eef7f1cfc52, 0x482835ea666b2573},
+      {0xad1c8eab5ee43b66, 0xda3243650005eed0},
+      {0xd863b256369d4a40, 0x90bed43e40076a83},
+      {0x873e4f75e2224e68, 0x5a7744a6e804a292},
+      {0xa90de3535aaae202, 0x711515d0a205cb37},
+      {0xd3515c2831559a83, 0x0d5a5b44ca873e04},
+      {0x8412d9991ed58091, 0xe858790afe9486c3},
+      {0xa5178fff668ae0b6, 0x626e974dbe39a873},
+      {0xce5d73ff402d98e3, 0xfb0a3d212dc81290},
+      {0x80fa687f881c7f8e, 0x7ce66634bc9d0b9a},
+      {0xa139029f6a239f72, 0x1c1fffc1ebc44e81},
+      {0xc987434744ac874e, 0xa327ffb266b56221},
+      {0xfbe9141915d7a922, 0x4bf1ff9f0062baa9},
+      {0x9d71ac8fada6c9b5, 0x6f773fc3603db4aa},
+      {0xc4ce17b399107c22, 0xcb550fb4384d21d4},
+      {0xf6019da07f549b2b, 0x7e2a53a146606a49},
+      {0x99c102844f94e0fb, 0x2eda7444cbfc426e},
+      {0xc0314325637a1939, 0xfa911155fefb5309},
+      {0xf03d93eebc589f88, 0x793555ab7eba27cb},
+      {0x96267c7535b763b5, 0x4bc1558b2f3458df},
+      {0xbbb01b9283253ca2, 0x9eb1aaedfb016f17},
+      {0xea9c227723ee8bcb, 0x465e15a979c1cadd},
+      {0x92a1958a7675175f, 0x0bfacd89ec191eca},
+      {0xb749faed14125d36, 0xcef980ec671f667c},
+      {0xe51c79a85916f484, 0x82b7e12780e7401b},
+      {0x8f31cc0937ae58d2, 0xd1b2ecb8b0908811},
+      {0xb2fe3f0b8599ef07, 0x861fa7e6dcb4aa16},
+      {0xdfbdcece67006ac9, 0x67a791e093e1d49b},
+      {0x8bd6a141006042bd, 0xe0c8bb2c5c6d24e1},
+      {0xaecc49914078536d, 0x58fae9f773886e19},
+      {0xda7f5bf590966848, 0xaf39a475506a899f},
+      {0x888f99797a5e012d, 0x6d8406c952429604},
+      {0xaab37fd7d8f58178, 0xc8e5087ba6d33b84},
+      {0xd5605fcdcf32e1d6, 0xfb1e4a9a90880a65},
+      {0x855c3be0a17fcd26, 0x5cf2eea09a550680},
+      {0xa6b34ad8c9dfc06f, 0xf42faa48c0ea481f},
+      {0xd0601d8efc57b08b, 0xf13b94daf124da27},
+      {0x823c12795db6ce57, 0x76c53d08d6b70859},
+      {0xa2cb1717b52481ed, 0x54768c4b0c64ca6f},
+      {0xcb7ddcdda26da268, 0xa9942f5dcf7dfd0a},
+      {0xfe5d54150b090b02, 0xd3f93b35435d7c4d},
+      {0x9efa548d26e5a6e1, 0xc47bc5014a1a6db0},
+      {0xc6b8e9b0709f109a, 0x359ab6419ca1091c},
+      {0xf867241c8cc6d4c0, 0xc30163d203c94b63},
+      {0x9b407691d7fc44f8, 0x79e0de63425dcf1e},
+      {0xc21094364dfb5636, 0x985915fc12f542e5},
+      {0xf294b943e17a2bc4, 0x3e6f5b7b17b2939e},
+      {0x979cf3ca6cec5b5a, 0xa705992ceecf9c43},
+      {0xbd8430bd08277231, 0x50c6ff782a838354},
+      {0xece53cec4a314ebd, 0xa4f8bf5635246429},
+      {0x940f4613ae5ed136, 0x871b7795e136be9a},
+      {0xb913179899f68584, 0x28e2557b59846e40},
+      {0xe757dd7ec07426e5, 0x331aeada2fe589d0},
+      {0x9096ea6f3848984f, 0x3ff0d2c85def7622},
+      {0xb4bca50b065abe63, 0x0fed077a756b53aa},
+      {0xe1ebce4dc7f16dfb, 0xd3e8495912c62895},
+      {0x8d3360f09cf6e4bd, 0x64712dd7abbbd95d},
+      {0xb080392cc4349dec, 0xbd8d794d96aacfb4},
+      {0xdca04777f541c567, 0xecf0d7a0fc5583a1},
+      {0x89e42caaf9491b60, 0xf41686c49db57245},
+      {0xac5d37d5b79b6239, 0x311c2875c522ced6},
+      {0xd77485cb25823ac7, 0x7d633293366b828c},
+      {0x86a8d39ef77164bc, 0xae5dff9c02033198},
+      {0xa8530886b54dbdeb, 0xd9f57f830283fdfd},
+      {0xd267caa862a12d66, 0xd072df63c324fd7c},
+      {0x8380dea93da4bc60, 0x4247cb9e59f71e6e},
+      {0xa46116538d0deb78, 0x52d9be85f074e609},
+      {0xcd795be870516656, 0x67902e276c921f8c},
+      {0x806bd9714632dff6, 0x00ba1cd8a3db53b7},
+      {0xa086cfcd97bf97f3, 0x80e8a40eccd228a5},
+      {0xc8a883c0fdaf7df0, 0x6122cd128006b2ce},
+      {0xfad2a4b13d1b5d6c, 0x796b805720085f82},
+      {0x9cc3a6eec6311a63, 0xcbe3303674053bb1},
+      {0xc3f490aa77bd60fc, 0xbedbfc4411068a9d},
+      {0xf4f1b4d515acb93b, 0xee92fb5515482d45},
+      {0x991711052d8bf3c5, 0x751bdd152d4d1c4b},
+      {0xbf5cd54678eef0b6, 0xd262d45a78a0635e},
+      {0xef340a98172aace4, 0x86fb897116c87c35},
+      {0x9580869f0e7aac0e, 0xd45d35e6ae3d4da1},
+      {0xbae0a846d2195712, 0x8974836059cca10a},
+      {0xe998d258869facd7, 0x2bd1a438703fc94c},
+      {0x91ff83775423cc06, 0x7b6306a34627ddd0},
+      {0xb67f6455292cbf08, 0x1a3bc84c17b1d543},
+      {0xe41f3d6a7377eeca, 0x20caba5f1d9e4a94},
+      {0x8e938662882af53e, 0x547eb47b7282ee9d},
+      {0xb23867fb2a35b28d, 0xe99e619a4f23aa44},
+      {0xdec681f9f4c31f31, 0x6405fa00e2ec94d5},
+      {0x8b3c113c38f9f37e, 0xde83bc408dd3dd05},
+      {0xae0b158b4738705e, 0x9624ab50b148d446},
+      {0xd98ddaee19068c76, 0x3badd624dd9b0958},
+      {0x87f8a8d4cfa417c9, 0xe54ca5d70a80e5d7},
+      {0xa9f6d30a038d1dbc, 0x5e9fcf4ccd211f4d},
+      {0xd47487cc8470652b, 0x7647c32000696720},
+      {0x84c8d4dfd2c63f3b, 0x29ecd9f40041e074},
+      {0xa5fb0a17c777cf09, 0xf468107100525891},
+      {0xcf79cc9db955c2cc, 0x7182148d4066eeb5},
+      {0x81ac1fe293d599bf, 0xc6f14cd848405531},
+      {0xa21727db38cb002f, 0xb8ada00e5a506a7d},
+      {0xca9cf1d206fdc03b, 0xa6d90811f0e4851d},
+      {0xfd442e4688bd304a, 0x908f4a166d1da664},
+      {0x9e4a9cec15763e2e, 0x9a598e4e043287ff},
+      {0xc5dd44271ad3cdba, 0x40eff1e1853f29fe},
+      {0xf7549530e188c128, 0xd12bee59e68ef47d},
+      {0x9a94dd3e8cf578b9, 0x82bb74f8301958cf},
+      {0xc13a148e3032d6e7, 0xe36a52363c1faf02},
+      {0xf18899b1bc3f8ca1, 0xdc44e6c3cb279ac2},
+      {0x96f5600f15a7b7e5, 0x29ab103a5ef8c0ba},
+      {0xbcb2b812db11a5de, 0x7415d448f6b6f0e8},
+      {0xebdf661791d60f56, 0x111b495b3464ad22},
+      {0x936b9fcebb25c995, 0xcab10dd900beec35},
+      {0xb84687c269ef3bfb, 0x3d5d514f40eea743},
+      {0xe65829b3046b0afa, 0x0cb4a5a3112a5113},
+      {0x8ff71a0fe2c2e6dc, 0x47f0e785eaba72ac},
+      {0xb3f4e093db73a093, 0x59ed216765690f57},
+      {0xe0f218b8d25088b8, 0x306869c13ec3532d},
+      {0x8c974f7383725573, 0x1e414218c73a13fc},
+      {0xafbd2350644eeacf, 0xe5d1929ef90898fb},
+      {0xdbac6c247d62a583, 0xdf45f746b74abf3a},
+      {0x894bc396ce5da772, 0x6b8bba8c328eb784},
+      {0xab9eb47c81f5114f, 0x066ea92f3f326565},
+      {0xd686619ba27255a2, 0xc80a537b0efefebe},
+      {0x8613fd0145877585, 0xbd06742ce95f5f37},
+      {0xa798fc4196e952e7, 0x2c48113823b73705},
+      {0xd17f3b51fca3a7a0, 0xf75a15862ca504c6},
+      {0x82ef85133de648c4, 0x9a984d73dbe722fc},
+      {0xa3ab66580d5fdaf5, 0xc13e60d0d2e0ebbb},
+      {0xcc963fee10b7d1b3, 0x318df905079926a9},
+      {0xffbbcfe994e5c61f, 0xfdf17746497f7053},
+      {0x9fd561f1fd0f9bd3, 0xfeb6ea8bedefa634},
+      {0xc7caba6e7c5382c8, 0xfe64a52ee96b8fc1},
+      {0xf9bd690a1b68637b, 0x3dfdce7aa3c673b1},
+      {0x9c1661a651213e2d, 0x06bea10ca65c084f},
+      {0xc31bfa0fe5698db8, 0x486e494fcff30a63},
+      {0xf3e2f893dec3f126, 0x5a89dba3c3efccfb},
+      {0x986ddb5c6b3a76b7, 0xf89629465a75e01d},
+      {0xbe89523386091465, 0xf6bbb397f1135824},
+      {0xee2ba6c0678b597f, 0x746aa07ded582e2d},
+      {0x94db483840b717ef, 0xa8c2a44eb4571cdd},
+      {0xba121a4650e4ddeb, 0x92f34d62616ce414},
+      {0xe896a0d7e51e1566, 0x77b020baf9c81d18},
+      {0x915e2486ef32cd60, 0x0ace1474dc1d122f},
+      {0xb5b5ada8aaff80b8, 0x0d819992132456bb},
+      {0xe3231912d5bf60e6, 0x10e1fff697ed6c6a},
+      {0x8df5efabc5979c8f, 0xca8d3ffa1ef463c2},
+      {0xb1736b96b6fd83b3, 0xbd308ff8a6b17cb3},
+      {0xddd0467c64bce4a0, 0xac7cb3f6d05ddbdf},
+      {0x8aa22c0dbef60ee4, 0x6bcdf07a423aa96c},
+      {0xad4ab7112eb3929d, 0x86c16c98d2c953c7},
+      {0xd89d64d57a607744, 0xe871c7bf077ba8b8},
+      {0x87625f056c7c4a8b, 0x11471cd764ad4973},
+      {0xa93af6c6c79b5d2d, 0xd598e40d3dd89bd0},
+      {0xd389b47879823479, 0x4aff1d108d4ec2c4},
+      {0x843610cb4bf160cb, 0xcedf722a585139bb},
+      {0xa54394fe1eedb8fe, 0xc2974eb4ee658829},
+      {0xce947a3da6a9273e, 0x733d226229feea33},
+      {0x811ccc668829b887, 0x0806357d5a3f5260},
+      {0xa163ff802a3426a8, 0xca07c2dcb0cf26f8},
+      {0xc9bcff6034c13052, 0xfc89b393dd02f0b6},
+      {0xfc2c3f3841f17c67, 0xbbac2078d443ace3},
+      {0x9d9ba7832936edc0, 0xd54b944b84aa4c0e},
+      {0xc5029163f384a931, 0x0a9e795e65d4df12},
+      {0xf64335bcf065d37d, 0x4d4617b5ff4a16d6},
+      {0x99ea0196163fa42e, 0x504bced1bf8e4e46},
+      {0xc06481fb9bcf8d39, 0xe45ec2862f71e1d7},
+      {0xf07da27a82c37088, 0x5d767327bb4e5a4d},
+      {0x964e858c91ba2655, 0x3a6a07f8d510f870},
+      {0xbbe226efb628afea, 0x890489f70a55368c},
+      {0xeadab0aba3b2dbe5, 0x2b45ac74ccea842f},
+      {0x92c8ae6b464fc96f, 0x3b0b8bc90012929e},
+      {0xb77ada0617e3bbcb, 0x09ce6ebb40173745},
+      {0xe55990879ddcaabd, 0xcc420a6a101d0516},
+      {0x8f57fa54c2a9eab6, 0x9fa946824a12232e},
+      {0xb32df8e9f3546564, 0x47939822dc96abfa},
+      {0xdff9772470297ebd, 0x59787e2b93bc56f8},
+      {0x8bfbea76c619ef36, 0x57eb4edb3c55b65b},
+      {0xaefae51477a06b03, 0xede622920b6b23f2},
+      {0xdab99e59958885c4, 0xe95fab368e45ecee},
+      {0x88b402f7fd75539b, 0x11dbcb0218ebb415},
+      {0xaae103b5fcd2a881, 0xd652bdc29f26a11a},
+      {0xd59944a37c0752a2, 0x4be76d3346f04960},
+      {0x857fcae62d8493a5, 0x6f70a4400c562ddc},
+      {0xa6dfbd9fb8e5b88e, 0xcb4ccd500f6bb953},
+      {0xd097ad07a71f26b2, 0x7e2000a41346a7a8},
+      {0x825ecc24c873782f, 0x8ed400668c0c28c9},
+      {0xa2f67f2dfa90563b, 0x728900802f0f32fb},
+      {0xcbb41ef979346bca, 0x4f2b40a03ad2ffba},
+      {0xfea126b7d78186bc, 0xe2f610c84987bfa9},
+      {0x9f24b832e6b0f436, 0x0dd9ca7d2df4d7ca},
+      {0xc6ede63fa05d3143, 0x91503d1c79720dbc},
+      {0xf8a95fcf88747d94, 0x75a44c6397ce912b},
+      {0x9b69dbe1b548ce7c, 0xc986afbe3ee11abb},
+      {0xc24452da229b021b, 0xfbe85badce996169},
+      {0xf2d56790ab41c2a2, 0xfae27299423fb9c4},
+      {0x97c560ba6b0919a5, 0xdccd879fc967d41b},
+      {0xbdb6b8e905cb600f, 0x5400e987bbc1c921},
+      {0xed246723473e3813, 0x290123e9aab23b69},
+      {0x9436c0760c86e30b, 0xf9a0b6720aaf6522},
+      {0xb94470938fa89bce, 0xf808e40e8d5b3e6a},
+      {0xe7958cb87392c2c2, 0xb60b1d1230b20e05},
+      {0x90bd77f3483bb9b9, 0xb1c6f22b5e6f48c3},
+      {0xb4ecd5f01a4aa828, 0x1e38aeb6360b1af4},
+      {0xe2280b6c20dd5232, 0x25c6da63c38de1b1},
+      {0x8d590723948a535f, 0x579c487e5a38ad0f},
+      {0xb0af48ec79ace837, 0x2d835a9df0c6d852},
+      {0xdcdb1b2798182244, 0xf8e431456cf88e66},
+      {0x8a08f0f8bf0f156b, 0x1b8e9ecb641b5900},
+      {0xac8b2d36eed2dac5, 0xe272467e3d222f40},
+      {0xd7adf884aa879177, 0x5b0ed81dcc6abb10},
+      {0x86ccbb52ea94baea, 0x98e947129fc2b4ea},
+      {0xa87fea27a539e9a5, 0x3f2398d747b36225},
+      {0xd29fe4b18e88640e, 0x8eec7f0d19a03aae},
+      {0x83a3eeeef9153e89, 0x1953cf68300424ad},
+      {0xa48ceaaab75a8e2b, 0x5fa8c3423c052dd8},
+      {0xcdb02555653131b6, 0x3792f412cb06794e},
+      {0x808e17555f3ebf11, 0xe2bbd88bbee40bd1},
+      {0xa0b19d2ab70e6ed6, 0x5b6aceaeae9d0ec5},
+      {0xc8de047564d20a8b, 0xf245825a5a445276},
+      {0xfb158592be068d2e, 0xeed6e2f0f0d56713},
+      {0x9ced737bb6c4183d, 0x55464dd69685606c},
+      {0xc428d05aa4751e4c, 0xaa97e14c3c26b887},
+      {0xf53304714d9265df, 0xd53dd99f4b3066a9},
+      {0x993fe2c6d07b7fab, 0xe546a8038efe402a},
+      {0xbf8fdb78849a5f96, 0xde98520472bdd034},
+      {0xef73d256a5c0f77c, 0x963e66858f6d4441},
+      {0x95a8637627989aad, 0xdde7001379a44aa9},
+      {0xbb127c53b17ec159, 0x5560c018580d5d53},
+      {0xe9d71b689dde71af, 0xaab8f01e6e10b4a7},
+      {0x9226712162ab070d, 0xcab3961304ca70e9},
+      {0xb6b00d69bb55c8d1, 0x3d607b97c5fd0d23},
+      {0xe45c10c42a2b3b05, 0x8cb89a7db77c506b},
+      {0x8eb98a7a9a5b04e3, 0x77f3608e92adb243},
+      {0xb267ed1940f1c61c, 0x55f038b237591ed4},
+      {0xdf01e85f912e37a3, 0x6b6c46dec52f6689},
+      {0x8b61313bbabce2c6, 0x2323ac4b3b3da016},
+      {0xae397d8aa96c1b77, 0xabec975e0a0d081b},
+      {0xd9c7dced53c72255, 0x96e7bd358c904a22},
+      {0x881cea14545c7575, 0x7e50d64177da2e55},
+      {0xaa242499697392d2, 0xdde50bd1d5d0b9ea},
+      {0xd4ad2dbfc3d07787, 0x955e4ec64b44e865},
+      {0x84ec3c97da624ab4, 0xbd5af13bef0b113f},
+      {0xa6274bbdd0fadd61, 0xecb1ad8aeacdd58f},
+      {0xcfb11ead453994ba, 0x67de18eda5814af3},
+      {0x81ceb32c4b43fcf4, 0x80eacf948770ced8},
+      {0xa2425ff75e14fc31, 0xa1258379a94d028e},
+      {0xcad2f7f5359a3b3e, 0x096ee45813a04331},
+      {0xfd87b5f28300ca0d, 0x8bca9d6e188853fd},
+      {0x9e74d1b791e07e48, 0x775ea264cf55347e},
+      {0xc612062576589dda, 0x95364afe032a819e},
+      {0xf79687aed3eec551, 0x3a83ddbd83f52205},
+      {0x9abe14cd44753b52, 0xc4926a9672793543},
+      {0xc16d9a0095928a27, 0x75b7053c0f178294},
+      {0xf1c90080baf72cb1, 0x5324c68b12dd6339},
+      {0x971da05074da7bee, 0xd3f6fc16ebca5e04},
+      {0xbce5086492111aea, 0x88f4bb1ca6bcf585},
+      {0xec1e4a7db69561a5, 0x2b31e9e3d06c32e6},
+      {0x9392ee8e921d5d07, 0x3aff322e62439fd0},
+      {0xb877aa3236a4b449, 0x09befeb9fad487c3},
+      {0xe69594bec44de15b, 0x4c2ebe687989a9b4},
+      {0x901d7cf73ab0acd9, 0x0f9d37014bf60a11},
+      {0xb424dc35095cd80f, 0x538484c19ef38c95},
+      {0xe12e13424bb40e13, 0x2865a5f206b06fba},
+      {0x8cbccc096f5088cb, 0xf93f87b7442e45d4},
+      {0xafebff0bcb24aafe, 0xf78f69a51539d749},
+      {0xdbe6fecebdedd5be, 0xb573440e5a884d1c},
+      {0x89705f4136b4a597, 0x31680a88f8953031},
+      {0xabcc77118461cefc, 0xfdc20d2b36ba7c3e},
+      {0xd6bf94d5e57a42bc, 0x3d32907604691b4d},
+      {0x8637bd05af6c69b5, 0xa63f9a49c2c1b110},
+      {0xa7c5ac471b478423, 0x0fcf80dc33721d54},
+      {0xd1b71758e219652b, 0xd3c36113404ea4a9},
+      {0x83126e978d4fdf3b, 0x645a1cac083126ea},
+      {0xa3d70a3d70a3d70a, 0x3d70a3d70a3d70a4},
+      {0xcccccccccccccccc, 0xcccccccccccccccd},
+      {0x8000000000000000, 0x0000000000000000},
+      {0xa000000000000000, 0x0000000000000000},
+      {0xc800000000000000, 0x0000000000000000},
+      {0xfa00000000000000, 0x0000000000000000},
+      {0x9c40000000000000, 0x0000000000000000},
+      {0xc350000000000000, 0x0000000000000000},
+      {0xf424000000000000, 0x0000000000000000},
+      {0x9896800000000000, 0x0000000000000000},
+      {0xbebc200000000000, 0x0000000000000000},
+      {0xee6b280000000000, 0x0000000000000000},
+      {0x9502f90000000000, 0x0000000000000000},
+      {0xba43b74000000000, 0x0000000000000000},
+      {0xe8d4a51000000000, 0x0000000000000000},
+      {0x9184e72a00000000, 0x0000000000000000},
+      {0xb5e620f480000000, 0x0000000000000000},
+      {0xe35fa931a0000000, 0x0000000000000000},
+      {0x8e1bc9bf04000000, 0x0000000000000000},
+      {0xb1a2bc2ec5000000, 0x0000000000000000},
+      {0xde0b6b3a76400000, 0x0000000000000000},
+      {0x8ac7230489e80000, 0x0000000000000000},
+      {0xad78ebc5ac620000, 0x0000000000000000},
+      {0xd8d726b7177a8000, 0x0000000000000000},
+      {0x878678326eac9000, 0x0000000000000000},
+      {0xa968163f0a57b400, 0x0000000000000000},
+      {0xd3c21bcecceda100, 0x0000000000000000},
+      {0x84595161401484a0, 0x0000000000000000},
+      {0xa56fa5b99019a5c8, 0x0000000000000000},
+      {0xcecb8f27f4200f3a, 0x0000000000000000},
+      {0x813f3978f8940984, 0x4000000000000000},
+      {0xa18f07d736b90be5, 0x5000000000000000},
+      {0xc9f2c9cd04674ede, 0xa400000000000000},
+      {0xfc6f7c4045812296, 0x4d00000000000000},
+      {0x9dc5ada82b70b59d, 0xf020000000000000},
+      {0xc5371912364ce305, 0x6c28000000000000},
+      {0xf684df56c3e01bc6, 0xc732000000000000},
+      {0x9a130b963a6c115c, 0x3c7f400000000000},
+      {0xc097ce7bc90715b3, 0x4b9f100000000000},
+      {0xf0bdc21abb48db20, 0x1e86d40000000000},
+      {0x96769950b50d88f4, 0x1314448000000000},
+      {0xbc143fa4e250eb31, 0x17d955a000000000},
+      {0xeb194f8e1ae525fd, 0x5dcfab0800000000},
+      {0x92efd1b8d0cf37be, 0x5aa1cae500000000},
+      {0xb7abc627050305ad, 0xf14a3d9e40000000},
+      {0xe596b7b0c643c719, 0x6d9ccd05d0000000},
+      {0x8f7e32ce7bea5c6f, 0xe4820023a2000000},
+      {0xb35dbf821ae4f38b, 0xdda2802c8a800000},
+      {0xe0352f62a19e306e, 0xd50b2037ad200000},
+      {0x8c213d9da502de45, 0x4526f422cc340000},
+      {0xaf298d050e4395d6, 0x9670b12b7f410000},
+      {0xdaf3f04651d47b4c, 0x3c0cdd765f114000},
+      {0x88d8762bf324cd0f, 0xa5880a69fb6ac800},
+      {0xab0e93b6efee0053, 0x8eea0d047a457a00},
+      {0xd5d238a4abe98068, 0x72a4904598d6d880},
+      {0x85a36366eb71f041, 0x47a6da2b7f864750},
+      {0xa70c3c40a64e6c51, 0x999090b65f67d924},
+      {0xd0cf4b50cfe20765, 0xfff4b4e3f741cf6d},
+      {0x82818f1281ed449f, 0xbff8f10e7a8921a5},
+      {0xa321f2d7226895c7, 0xaff72d52192b6a0e},
+      {0xcbea6f8ceb02bb39, 0x9bf4f8a69f764491},
+      {0xfee50b7025c36a08, 0x02f236d04753d5b5},
+      {0x9f4f2726179a2245, 0x01d762422c946591},
+      {0xc722f0ef9d80aad6, 0x424d3ad2b7b97ef6},
+      {0xf8ebad2b84e0d58b, 0xd2e0898765a7deb3},
+      {0x9b934c3b330c8577, 0x63cc55f49f88eb30},
+      {0xc2781f49ffcfa6d5, 0x3cbf6b71c76b25fc},
+      {0xf316271c7fc3908a, 0x8bef464e3945ef7b},
+      {0x97edd871cfda3a56, 0x97758bf0e3cbb5ad},
+      {0xbde94e8e43d0c8ec, 0x3d52eeed1cbea318},
+      {0xed63a231d4c4fb27, 0x4ca7aaa863ee4bde},
+      {0x945e455f24fb1cf8, 0x8fe8caa93e74ef6b},
+      {0xb975d6b6ee39e436, 0xb3e2fd538e122b45},
+      {0xe7d34c64a9c85d44, 0x60dbbca87196b617},
+      {0x90e40fbeea1d3a4a, 0xbc8955e946fe31ce},
+      {0xb51d13aea4a488dd, 0x6babab6398bdbe42},
+      {0xe264589a4dcdab14, 0xc696963c7eed2dd2},
+      {0x8d7eb76070a08aec, 0xfc1e1de5cf543ca3},
+      {0xb0de65388cc8ada8, 0x3b25a55f43294bcc},
+      {0xdd15fe86affad912, 0x49ef0eb713f39ebf},
+      {0x8a2dbf142dfcc7ab, 0x6e3569326c784338},
+      {0xacb92ed9397bf996, 0x49c2c37f07965405},
+      {0xd7e77a8f87daf7fb, 0xdc33745ec97be907},
+      {0x86f0ac99b4e8dafd, 0x69a028bb3ded71a4},
+      {0xa8acd7c0222311bc, 0xc40832ea0d68ce0d},
+      {0xd2d80db02aabd62b, 0xf50a3fa490c30191},
+      {0x83c7088e1aab65db, 0x792667c6da79e0fb},
+      {0xa4b8cab1a1563f52, 0x577001b891185939},
+      {0xcde6fd5e09abcf26, 0xed4c0226b55e6f87},
+      {0x80b05e5ac60b6178, 0x544f8158315b05b5},
+      {0xa0dc75f1778e39d6, 0x696361ae3db1c722},
+      {0xc913936dd571c84c, 0x03bc3a19cd1e38ea},
+      {0xfb5878494ace3a5f, 0x04ab48a04065c724},
+      {0x9d174b2dcec0e47b, 0x62eb0d64283f9c77},
+      {0xc45d1df942711d9a, 0x3ba5d0bd324f8395},
+      {0xf5746577930d6500, 0xca8f44ec7ee3647a},
+      {0x9968bf6abbe85f20, 0x7e998b13cf4e1ecc},
+      {0xbfc2ef456ae276e8, 0x9e3fedd8c321a67f},
+      {0xefb3ab16c59b14a2, 0xc5cfe94ef3ea101f},
+      {0x95d04aee3b80ece5, 0xbba1f1d158724a13},
+      {0xbb445da9ca61281f, 0x2a8a6e45ae8edc98},
+      {0xea1575143cf97226, 0xf52d09d71a3293be},
+      {0x924d692ca61be758, 0x593c2626705f9c57},
+      {0xb6e0c377cfa2e12e, 0x6f8b2fb00c77836d},
+      {0xe498f455c38b997a, 0x0b6dfb9c0f956448},
+      {0x8edf98b59a373fec, 0x4724bd4189bd5ead},
+      {0xb2977ee300c50fe7, 0x58edec91ec2cb658},
+      {0xdf3d5e9bc0f653e1, 0x2f2967b66737e3ee},
+      {0x8b865b215899f46c, 0xbd79e0d20082ee75},
+      {0xae67f1e9aec07187, 0xecd8590680a3aa12},
+      {0xda01ee641a708de9, 0xe80e6f4820cc9496},
+      {0x884134fe908658b2, 0x3109058d147fdcde},
+      {0xaa51823e34a7eede, 0xbd4b46f0599fd416},
+      {0xd4e5e2cdc1d1ea96, 0x6c9e18ac7007c91b},
+      {0x850fadc09923329e, 0x03e2cf6bc604ddb1},
+      {0xa6539930bf6bff45, 0x84db8346b786151d},
+      {0xcfe87f7cef46ff16, 0xe612641865679a64},
+      {0x81f14fae158c5f6e, 0x4fcb7e8f3f60c07f},
+      {0xa26da3999aef7749, 0xe3be5e330f38f09e},
+      {0xcb090c8001ab551c, 0x5cadf5bfd3072cc6},
+      {0xfdcb4fa002162a63, 0x73d9732fc7c8f7f7},
+      {0x9e9f11c4014dda7e, 0x2867e7fddcdd9afb},
+      {0xc646d63501a1511d, 0xb281e1fd541501b9},
+      {0xf7d88bc24209a565, 0x1f225a7ca91a4227},
+      {0x9ae757596946075f, 0x3375788de9b06959},
+      {0xc1a12d2fc3978937, 0x0052d6b1641c83af},
+      {0xf209787bb47d6b84, 0xc0678c5dbd23a49b},
+      {0x9745eb4d50ce6332, 0xf840b7ba963646e1},
+      {0xbd176620a501fbff, 0xb650e5a93bc3d899},
+      {0xec5d3fa8ce427aff, 0xa3e51f138ab4cebf},
+      {0x93ba47c980e98cdf, 0xc66f336c36b10138},
+      {0xb8a8d9bbe123f017, 0xb80b0047445d4185},
+      {0xe6d3102ad96cec1d, 0xa60dc059157491e6},
+      {0x9043ea1ac7e41392, 0x87c89837ad68db30},
+      {0xb454e4a179dd1877, 0x29babe4598c311fc},
+      {0xe16a1dc9d8545e94, 0xf4296dd6fef3d67b},
+      {0x8ce2529e2734bb1d, 0x1899e4a65f58660d},
+      {0xb01ae745b101e9e4, 0x5ec05dcff72e7f90},
+      {0xdc21a1171d42645d, 0x76707543f4fa1f74},
+      {0x899504ae72497eba, 0x6a06494a791c53a9},
+      {0xabfa45da0edbde69, 0x0487db9d17636893},
+      {0xd6f8d7509292d603, 0x45a9d2845d3c42b7},
+      {0x865b86925b9bc5c2, 0x0b8a2392ba45a9b3},
+      {0xa7f26836f282b732, 0x8e6cac7768d7141f},
+      {0xd1ef0244af2364ff, 0x3207d795430cd927},
+      {0x8335616aed761f1f, 0x7f44e6bd49e807b9},
+      {0xa402b9c5a8d3a6e7, 0x5f16206c9c6209a7},
+      {0xcd036837130890a1, 0x36dba887c37a8c10},
+      {0x802221226be55a64, 0xc2494954da2c978a},
+      {0xa02aa96b06deb0fd, 0xf2db9baa10b7bd6d},
+      {0xc83553c5c8965d3d, 0x6f92829494e5acc8},
+      {0xfa42a8b73abbf48c, 0xcb772339ba1f17fa},
+      {0x9c69a97284b578d7, 0xff2a760414536efc},
+      {0xc38413cf25e2d70d, 0xfef5138519684abb},
+      {0xf46518c2ef5b8cd1, 0x7eb258665fc25d6a},
+      {0x98bf2f79d5993802, 0xef2f773ffbd97a62},
+      {0xbeeefb584aff8603, 0xaafb550ffacfd8fb},
+      {0xeeaaba2e5dbf6784, 0x95ba2a53f983cf39},
+      {0x952ab45cfa97a0b2, 0xdd945a747bf26184},
+      {0xba756174393d88df, 0x94f971119aeef9e5},
+      {0xe912b9d1478ceb17, 0x7a37cd5601aab85e},
+      {0x91abb422ccb812ee, 0xac62e055c10ab33b},
+      {0xb616a12b7fe617aa, 0x577b986b314d600a},
+      {0xe39c49765fdf9d94, 0xed5a7e85fda0b80c},
+      {0x8e41ade9fbebc27d, 0x14588f13be847308},
+      {0xb1d219647ae6b31c, 0x596eb2d8ae258fc9},
+      {0xde469fbd99a05fe3, 0x6fca5f8ed9aef3bc},
+      {0x8aec23d680043bee, 0x25de7bb9480d5855},
+      {0xada72ccc20054ae9, 0xaf561aa79a10ae6b},
+      {0xd910f7ff28069da4, 0x1b2ba1518094da05},
+      {0x87aa9aff79042286, 0x90fb44d2f05d0843},
+      {0xa99541bf57452b28, 0x353a1607ac744a54},
+      {0xd3fa922f2d1675f2, 0x42889b8997915ce9},
+      {0x847c9b5d7c2e09b7, 0x69956135febada12},
+      {0xa59bc234db398c25, 0x43fab9837e699096},
+      {0xcf02b2c21207ef2e, 0x94f967e45e03f4bc},
+      {0x8161afb94b44f57d, 0x1d1be0eebac278f6},
+      {0xa1ba1ba79e1632dc, 0x6462d92a69731733},
+      {0xca28a291859bbf93, 0x7d7b8f7503cfdcff},
+      {0xfcb2cb35e702af78, 0x5cda735244c3d43f},
+      {0x9defbf01b061adab, 0x3a0888136afa64a8},
+      {0xc56baec21c7a1916, 0x088aaa1845b8fdd1},
+      {0xf6c69a72a3989f5b, 0x8aad549e57273d46},
+      {0x9a3c2087a63f6399, 0x36ac54e2f678864c},
+      {0xc0cb28a98fcf3c7f, 0x84576a1bb416a7de},
+      {0xf0fdf2d3f3c30b9f, 0x656d44a2a11c51d6},
+      {0x969eb7c47859e743, 0x9f644ae5a4b1b326},
+      {0xbc4665b596706114, 0x873d5d9f0dde1fef},
+      {0xeb57ff22fc0c7959, 0xa90cb506d155a7eb},
+      {0x9316ff75dd87cbd8, 0x09a7f12442d588f3},
+      {0xb7dcbf5354e9bece, 0x0c11ed6d538aeb30},
+      {0xe5d3ef282a242e81, 0x8f1668c8a86da5fb},
+      {0x8fa475791a569d10, 0xf96e017d694487bd},
+      {0xb38d92d760ec4455, 0x37c981dcc395a9ad},
+      {0xe070f78d3927556a, 0x85bbe253f47b1418},
+      {0x8c469ab843b89562, 0x93956d7478ccec8f},
+      {0xaf58416654a6babb, 0x387ac8d1970027b3},
+      {0xdb2e51bfe9d0696a, 0x06997b05fcc0319f},
+      {0x88fcf317f22241e2, 0x441fece3bdf81f04},
+      {0xab3c2fddeeaad25a, 0xd527e81cad7626c4},
+      {0xd60b3bd56a5586f1, 0x8a71e223d8d3b075},
+      {0x85c7056562757456, 0xf6872d5667844e4a},
+      {0xa738c6bebb12d16c, 0xb428f8ac016561dc},
+      {0xd106f86e69d785c7, 0xe13336d701beba53},
+      {0x82a45b450226b39c, 0xecc0024661173474},
+      {0xa34d721642b06084, 0x27f002d7f95d0191},
+      {0xcc20ce9bd35c78a5, 0x31ec038df7b441f5},
+      {0xff290242c83396ce, 0x7e67047175a15272},
+      {0x9f79a169bd203e41, 0x0f0062c6e984d387},
+      {0xc75809c42c684dd1, 0x52c07b78a3e60869},
+      {0xf92e0c3537826145, 0xa7709a56ccdf8a83},
+      {0x9bbcc7a142b17ccb, 0x88a66076400bb692},
+      {0xc2abf989935ddbfe, 0x6acff893d00ea436},
+      {0xf356f7ebf83552fe, 0x0583f6b8c4124d44},
+      {0x98165af37b2153de, 0xc3727a337a8b704b},
+      {0xbe1bf1b059e9a8d6, 0x744f18c0592e4c5d},
+      {0xeda2ee1c7064130c, 0x1162def06f79df74},
+      {0x9485d4d1c63e8be7, 0x8addcb5645ac2ba9},
+      {0xb9a74a0637ce2ee1, 0x6d953e2bd7173693},
+      {0xe8111c87c5c1ba99, 0xc8fa8db6ccdd0438},
+      {0x910ab1d4db9914a0, 0x1d9c9892400a22a3},
+      {0xb54d5e4a127f59c8, 0x2503beb6d00cab4c},
+      {0xe2a0b5dc971f303a, 0x2e44ae64840fd61e},
+      {0x8da471a9de737e24, 0x5ceaecfed289e5d3},
+      {0xb10d8e1456105dad, 0x7425a83e872c5f48},
+      {0xdd50f1996b947518, 0xd12f124e28f7771a},
+      {0x8a5296ffe33cc92f, 0x82bd6b70d99aaa70},
+      {0xace73cbfdc0bfb7b, 0x636cc64d1001550c},
+      {0xd8210befd30efa5a, 0x3c47f7e05401aa4f},
+      {0x8714a775e3e95c78, 0x65acfaec34810a72},
+      {0xa8d9d1535ce3b396, 0x7f1839a741a14d0e},
+      {0xd31045a8341ca07c, 0x1ede48111209a051},
+      {0x83ea2b892091e44d, 0x934aed0aab460433},
+      {0xa4e4b66b68b65d60, 0xf81da84d56178540},
+      {0xce1de40642e3f4b9, 0x36251260ab9d668f},
+      {0x80d2ae83e9ce78f3, 0xc1d72b7c6b42601a},
+      {0xa1075a24e4421730, 0xb24cf65b8612f820},
+      {0xc94930ae1d529cfc, 0xdee033f26797b628},
+      {0xfb9b7cd9a4a7443c, 0x169840ef017da3b2},
+      {0x9d412e0806e88aa5, 0x8e1f289560ee864f},
+      {0xc491798a08a2ad4e, 0xf1a6f2bab92a27e3},
+      {0xf5b5d7ec8acb58a2, 0xae10af696774b1dc},
+      {0x9991a6f3d6bf1765, 0xacca6da1e0a8ef2a},
+      {0xbff610b0cc6edd3f, 0x17fd090a58d32af4},
+      {0xeff394dcff8a948e, 0xddfc4b4cef07f5b1},
+      {0x95f83d0a1fb69cd9, 0x4abdaf101564f98f},
+      {0xbb764c4ca7a4440f, 0x9d6d1ad41abe37f2},
+      {0xea53df5fd18d5513, 0x84c86189216dc5ee},
+      {0x92746b9be2f8552c, 0x32fd3cf5b4e49bb5},
+      {0xb7118682dbb66a77, 0x3fbc8c33221dc2a2},
+      {0xe4d5e82392a40515, 0x0fabaf3feaa5334b},
+      {0x8f05b1163ba6832d, 0x29cb4d87f2a7400f},
+      {0xb2c71d5bca9023f8, 0x743e20e9ef511013},
+      {0xdf78e4b2bd342cf6, 0x914da9246b255417},
+      {0x8bab8eefb6409c1a, 0x1ad089b6c2f7548f},
+      {0xae9672aba3d0c320, 0xa184ac2473b529b2},
+      {0xda3c0f568cc4f3e8, 0xc9e5d72d90a2741f},
+      {0x8865899617fb1871, 0x7e2fa67c7a658893},
+      {0xaa7eebfb9df9de8d, 0xddbb901b98feeab8},
+      {0xd51ea6fa85785631, 0x552a74227f3ea566},
+      {0x8533285c936b35de, 0xd53a88958f872760},
+      {0xa67ff273b8460356, 0x8a892abaf368f138},
+      {0xd01fef10a657842c, 0x2d2b7569b0432d86},
+      {0x8213f56a67f6b29b, 0x9c3b29620e29fc74},
+      {0xa298f2c501f45f42, 0x8349f3ba91b47b90},
+      {0xcb3f2f7642717713, 0x241c70a936219a74},
+      {0xfe0efb53d30dd4d7, 0xed238cd383aa0111},
+      {0x9ec95d1463e8a506, 0xf4363804324a40ab},
+      {0xc67bb4597ce2ce48, 0xb143c6053edcd0d6},
+      {0xf81aa16fdc1b81da, 0xdd94b7868e94050b},
+      {0x9b10a4e5e9913128, 0xca7cf2b4191c8327},
+      {0xc1d4ce1f63f57d72, 0xfd1c2f611f63a3f1},
+      {0xf24a01a73cf2dccf, 0xbc633b39673c8ced},
+      {0x976e41088617ca01, 0xd5be0503e085d814},
+      {0xbd49d14aa79dbc82, 0x4b2d8644d8a74e19},
+      {0xec9c459d51852ba2, 0xddf8e7d60ed1219f},
+      {0x93e1ab8252f33b45, 0xcabb90e5c942b504},
+      {0xb8da1662e7b00a17, 0x3d6a751f3b936244},
+      {0xe7109bfba19c0c9d, 0x0cc512670a783ad5},
+      {0x906a617d450187e2, 0x27fb2b80668b24c6},
+      {0xb484f9dc9641e9da, 0xb1f9f660802dedf7},
+      {0xe1a63853bbd26451, 0x5e7873f8a0396974},
+      {0x8d07e33455637eb2, 0xdb0b487b6423e1e9},
+      {0xb049dc016abc5e5f, 0x91ce1a9a3d2cda63},
+      {0xdc5c5301c56b75f7, 0x7641a140cc7810fc},
+      {0x89b9b3e11b6329ba, 0xa9e904c87fcb0a9e},
+      {0xac2820d9623bf429, 0x546345fa9fbdcd45},
+      {0xd732290fbacaf133, 0xa97c177947ad4096},
+      {0x867f59a9d4bed6c0, 0x49ed8eabcccc485e},
+      {0xa81f301449ee8c70, 0x5c68f256bfff5a75},
+      {0xd226fc195c6a2f8c, 0x73832eec6fff3112},
+      {0x83585d8fd9c25db7, 0xc831fd53c5ff7eac},
+      {0xa42e74f3d032f525, 0xba3e7ca8b77f5e56},
+      {0xcd3a1230c43fb26f, 0x28ce1bd2e55f35ec},
+      {0x80444b5e7aa7cf85, 0x7980d163cf5b81b4},
+      {0xa0555e361951c366, 0xd7e105bcc3326220},
+      {0xc86ab5c39fa63440, 0x8dd9472bf3fefaa8},
+      {0xfa856334878fc150, 0xb14f98f6f0feb952},
+      {0x9c935e00d4b9d8d2, 0x6ed1bf9a569f33d4},
+      {0xc3b8358109e84f07, 0x0a862f80ec4700c9},
+      {0xf4a642e14c6262c8, 0xcd27bb612758c0fb},
+      {0x98e7e9cccfbd7dbd, 0x8038d51cb897789d},
+      {0xbf21e44003acdd2c, 0xe0470a63e6bd56c4},
+      {0xeeea5d5004981478, 0x1858ccfce06cac75},
+      {0x95527a5202df0ccb, 0x0f37801e0c43ebc9},
+      {0xbaa718e68396cffd, 0xd30560258f54e6bb},
+      {0xe950df20247c83fd, 0x47c6b82ef32a206a},
+      {0x91d28b7416cdd27e, 0x4cdc331d57fa5442},
+      {0xb6472e511c81471d, 0xe0133fe4adf8e953},
+      {0xe3d8f9e563a198e5, 0x58180fddd97723a7},
+      {0x8e679c2f5e44ff8f, 0x570f09eaa7ea7649},
+      {0xb201833b35d63f73, 0x2cd2cc6551e513db},
+      {0xde81e40a034bcf4f, 0xf8077f7ea65e58d2},
+      {0x8b112e86420f6191, 0xfb04afaf27faf783},
+      {0xadd57a27d29339f6, 0x79c5db9af1f9b564},
+      {0xd94ad8b1c7380874, 0x18375281ae7822bd},
+      {0x87cec76f1c830548, 0x8f2293910d0b15b6},
+      {0xa9c2794ae3a3c69a, 0xb2eb3875504ddb23},
+      {0xd433179d9c8cb841, 0x5fa60692a46151ec},
+      {0x849feec281d7f328, 0xdbc7c41ba6bcd334},
+      {0xa5c7ea73224deff3, 0x12b9b522906c0801},
+      {0xcf39e50feae16bef, 0xd768226b34870a01},
+      {0x81842f29f2cce375, 0xe6a1158300d46641},
+      {0xa1e53af46f801c53, 0x60495ae3c1097fd1},
+      {0xca5e89b18b602368, 0x385bb19cb14bdfc5},
+      {0xfcf62c1dee382c42, 0x46729e03dd9ed7b6},
+      {0x9e19db92b4e31ba9, 0x6c07a2c26a8346d2},
+      {0xc5a05277621be293, 0xc7098b7305241886},
+      {0xf70867153aa2db38, 0xb8cbee4fc66d1ea8},
+      {0x9a65406d44a5c903, 0x737f74f1dc043329},
+      {0xc0fe908895cf3b44, 0x505f522e53053ff3},
+      {0xf13e34aabb430a15, 0x647726b9e7c68ff0},
+      {0x96c6e0eab509e64d, 0x5eca783430dc19f6},
+      {0xbc789925624c5fe0, 0xb67d16413d132073},
+      {0xeb96bf6ebadf77d8, 0xe41c5bd18c57e890},
+      {0x933e37a534cbaae7, 0x8e91b962f7b6f15a},
+      {0xb80dc58e81fe95a1, 0x723627bbb5a4adb1},
+      {0xe61136f2227e3b09, 0xcec3b1aaa30dd91d},
+      {0x8fcac257558ee4e6, 0x213a4f0aa5e8a7b2},
+      {0xb3bd72ed2af29e1f, 0xa988e2cd4f62d19e},
+      {0xe0accfa875af45a7, 0x93eb1b80a33b8606},
+      {0x8c6c01c9498d8b88, 0xbc72f130660533c4},
+      {0xaf87023b9bf0ee6a, 0xeb8fad7c7f8680b5},
+      {0xdb68c2ca82ed2a05, 0xa67398db9f6820e2},
+#else
+      {0xff77b1fcbebcdc4f, 0x25e8e89c13bb0f7b},
+      {0xce5d73ff402d98e3, 0xfb0a3d212dc81290},
+      {0xa6b34ad8c9dfc06f, 0xf42faa48c0ea481f},
+      {0x86a8d39ef77164bc, 0xae5dff9c02033198},
+      {0xd98ddaee19068c76, 0x3badd624dd9b0958},
+      {0xafbd2350644eeacf, 0xe5d1929ef90898fb},
+      {0x8df5efabc5979c8f, 0xca8d3ffa1ef463c2},
+      {0xe55990879ddcaabd, 0xcc420a6a101d0516},
+      {0xb94470938fa89bce, 0xf808e40e8d5b3e6a},
+      {0x95a8637627989aad, 0xdde7001379a44aa9},
+      {0xf1c90080baf72cb1, 0x5324c68b12dd6339},
+      {0xc350000000000000, 0x0000000000000000},
+      {0x9dc5ada82b70b59d, 0xf020000000000000},
+      {0xfee50b7025c36a08, 0x02f236d04753d5b5},
+      {0xcde6fd5e09abcf26, 0xed4c0226b55e6f87},
+      {0xa6539930bf6bff45, 0x84db8346b786151d},
+      {0x865b86925b9bc5c2, 0x0b8a2392ba45a9b3},
+      {0xd910f7ff28069da4, 0x1b2ba1518094da05},
+      {0xaf58416654a6babb, 0x387ac8d1970027b3},
+      {0x8da471a9de737e24, 0x5ceaecfed289e5d3},
+      {0xe4d5e82392a40515, 0x0fabaf3feaa5334b},
+      {0xb8da1662e7b00a17, 0x3d6a751f3b936244},
+      {0x95527a5202df0ccb, 0x0f37801e0c43ebc9},
+      {0xf13e34aabb430a15, 0x647726b9e7c68ff0}
+#endif
+    };
+
+#if FMT_USE_FULL_CACHE_DRAGONBOX
+    return pow10_significands[k - float_info<double>::min_k];
+#else
+    static constexpr const uint64_t powers_of_5_64[] = {
+        0x0000000000000001, 0x0000000000000005, 0x0000000000000019,
+        0x000000000000007d, 0x0000000000000271, 0x0000000000000c35,
+        0x0000000000003d09, 0x000000000001312d, 0x000000000005f5e1,
+        0x00000000001dcd65, 0x00000000009502f9, 0x0000000002e90edd,
+        0x000000000e8d4a51, 0x0000000048c27395, 0x000000016bcc41e9,
+        0x000000071afd498d, 0x0000002386f26fc1, 0x000000b1a2bc2ec5,
+        0x000003782dace9d9, 0x00001158e460913d, 0x000056bc75e2d631,
+        0x0001b1ae4d6e2ef5, 0x000878678326eac9, 0x002a5a058fc295ed,
+        0x00d3c21bcecceda1, 0x0422ca8b0a00a425, 0x14adf4b7320334b9};
+
+    static const int compression_ratio = 27;
+
+    // Compute base index.
+    int cache_index = (k - float_info<double>::min_k) / compression_ratio;
+    int kb = cache_index * compression_ratio + float_info<double>::min_k;
+    int offset = k - kb;
+
+    // Get base cache.
+    uint128_fallback base_cache = pow10_significands[cache_index];
+    if (offset == 0) return base_cache;
+
+    // Compute the required amount of bit-shift.
+    int alpha = floor_log2_pow10(kb + offset) - floor_log2_pow10(kb) - offset;
+    FMT_ASSERT(alpha > 0 && alpha < 64, "shifting error detected");
+
+    // Try to recover the real cache.
+    uint64_t pow5 = powers_of_5_64[offset];
+    uint128_fallback recovered_cache = umul128(base_cache.high(), pow5);
+    uint128_fallback middle_low = umul128(base_cache.low(), pow5);
+
+    recovered_cache += middle_low.high();
+
+    uint64_t high_to_middle = recovered_cache.high() << (64 - alpha);
+    uint64_t middle_to_low = recovered_cache.low() << (64 - alpha);
+
+    recovered_cache =
+        uint128_fallback{(recovered_cache.low() >> alpha) | high_to_middle,
+                         ((middle_low.low() >> alpha) | middle_to_low)};
+    FMT_ASSERT(recovered_cache.low() + 1 != 0, "");
+    return {recovered_cache.high(), recovered_cache.low() + 1};
+#endif
+  }
+
+  struct compute_mul_result {
+    carrier_uint result;
+    bool is_integer;
+  };
+  struct compute_mul_parity_result {
+    bool parity;
+    bool is_integer;
+  };
+
+  static auto compute_mul(carrier_uint u,
+                          const cache_entry_type& cache) noexcept
+      -> compute_mul_result {
+    auto r = umul192_upper128(u, cache);
+    return {r.high(), r.low() == 0};
+  }
+
+  static auto compute_delta(cache_entry_type const& cache, int beta) noexcept
+      -> uint32_t {
+    return static_cast<uint32_t>(cache.high() >> (64 - 1 - beta));
+  }
+
+  static auto compute_mul_parity(carrier_uint two_f,
+                                 const cache_entry_type& cache,
+                                 int beta) noexcept
+      -> compute_mul_parity_result {
+    FMT_ASSERT(beta >= 1, "");
+    FMT_ASSERT(beta < 64, "");
+
+    auto r = umul192_lower128(two_f, cache);
+    return {((r.high() >> (64 - beta)) & 1) != 0,
+            ((r.high() << beta) | (r.low() >> (64 - beta))) == 0};
+  }
+
+  static auto compute_left_endpoint_for_shorter_interval_case(
+      const cache_entry_type& cache, int beta) noexcept -> carrier_uint {
+    return (cache.high() -
+            (cache.high() >> (num_significand_bits<double>() + 2))) >>
+           (64 - num_significand_bits<double>() - 1 - beta);
+  }
+
+  static auto compute_right_endpoint_for_shorter_interval_case(
+      const cache_entry_type& cache, int beta) noexcept -> carrier_uint {
+    return (cache.high() +
+            (cache.high() >> (num_significand_bits<double>() + 1))) >>
+           (64 - num_significand_bits<double>() - 1 - beta);
+  }
+
+  static auto compute_round_up_for_shorter_interval_case(
+      const cache_entry_type& cache, int beta) noexcept -> carrier_uint {
+    return ((cache.high() >> (64 - num_significand_bits<double>() - 2 - beta)) +
+            1) /
+           2;
+  }
+};
+
+FMT_FUNC auto get_cached_power(int k) noexcept -> uint128_fallback {
+  return cache_accessor<double>::get_cached_power(k);
+}
+
+// Various integer checks
+template <typename T>
+auto is_left_endpoint_integer_shorter_interval(int exponent) noexcept -> bool {
+  const int case_shorter_interval_left_endpoint_lower_threshold = 2;
+  const int case_shorter_interval_left_endpoint_upper_threshold = 3;
+  return exponent >= case_shorter_interval_left_endpoint_lower_threshold &&
+         exponent <= case_shorter_interval_left_endpoint_upper_threshold;
+}
+
+// Remove trailing zeros from n and return the number of zeros removed (float)
+FMT_INLINE int remove_trailing_zeros(uint32_t& n, int s = 0) noexcept {
+  FMT_ASSERT(n != 0, "");
+  // Modular inverse of 5 (mod 2^32): (mod_inv_5 * 5) mod 2^32 = 1.
+  constexpr uint32_t mod_inv_5 = 0xcccccccd;
+  constexpr uint32_t mod_inv_25 = 0xc28f5c29;  // = mod_inv_5 * mod_inv_5
+
+  while (true) {
+    auto q = rotr(n * mod_inv_25, 2);
+    if (q > max_value<uint32_t>() / 100) break;
+    n = q;
+    s += 2;
+  }
+  auto q = rotr(n * mod_inv_5, 1);
+  if (q <= max_value<uint32_t>() / 10) {
+    n = q;
+    s |= 1;
+  }
+  return s;
+}
+
+// Removes trailing zeros and returns the number of zeros removed (double)
+FMT_INLINE int remove_trailing_zeros(uint64_t& n) noexcept {
+  FMT_ASSERT(n != 0, "");
+
+  // This magic number is ceil(2^90 / 10^8).
+  constexpr uint64_t magic_number = 12379400392853802749ull;
+  auto nm = umul128(n, magic_number);
+
+  // Is n is divisible by 10^8?
+  if ((nm.high() & ((1ull << (90 - 64)) - 1)) == 0 && nm.low() < magic_number) {
+    // If yes, work with the quotient...
+    auto n32 = static_cast<uint32_t>(nm.high() >> (90 - 64));
+    // ... and use the 32 bit variant of the function
+    int s = remove_trailing_zeros(n32, 8);
+    n = n32;
+    return s;
+  }
+
+  // If n is not divisible by 10^8, work with n itself.
+  constexpr uint64_t mod_inv_5 = 0xcccccccccccccccd;
+  constexpr uint64_t mod_inv_25 = 0x8f5c28f5c28f5c29;  // mod_inv_5 * mod_inv_5
+
+  int s = 0;
+  while (true) {
+    auto q = rotr(n * mod_inv_25, 2);
+    if (q > max_value<uint64_t>() / 100) break;
+    n = q;
+    s += 2;
+  }
+  auto q = rotr(n * mod_inv_5, 1);
+  if (q <= max_value<uint64_t>() / 10) {
+    n = q;
+    s |= 1;
+  }
+
+  return s;
+}
+
+// The main algorithm for shorter interval case
+template <typename T>
+FMT_INLINE decimal_fp<T> shorter_interval_case(int exponent) noexcept {
+  decimal_fp<T> ret_value;
+  // Compute k and beta
+  const int minus_k = floor_log10_pow2_minus_log10_4_over_3(exponent);
+  const int beta = exponent + floor_log2_pow10(-minus_k);
+
+  // Compute xi and zi
+  using cache_entry_type = typename cache_accessor<T>::cache_entry_type;
+  const cache_entry_type cache = cache_accessor<T>::get_cached_power(-minus_k);
+
+  auto xi = cache_accessor<T>::compute_left_endpoint_for_shorter_interval_case(
+      cache, beta);
+  auto zi = cache_accessor<T>::compute_right_endpoint_for_shorter_interval_case(
+      cache, beta);
+
+  // If the left endpoint is not an integer, increase it
+  if (!is_left_endpoint_integer_shorter_interval<T>(exponent)) ++xi;
+
+  // Try bigger divisor
+  ret_value.significand = zi / 10;
+
+  // If succeed, remove trailing zeros if necessary and return
+  if (ret_value.significand * 10 >= xi) {
+    ret_value.exponent = minus_k + 1;
+    ret_value.exponent += remove_trailing_zeros(ret_value.significand);
+    return ret_value;
+  }
+
+  // Otherwise, compute the round-up of y
+  ret_value.significand =
+      cache_accessor<T>::compute_round_up_for_shorter_interval_case(cache,
+                                                                    beta);
+  ret_value.exponent = minus_k;
+
+  // When tie occurs, choose one of them according to the rule
+  if (exponent >= float_info<T>::shorter_interval_tie_lower_threshold &&
+      exponent <= float_info<T>::shorter_interval_tie_upper_threshold) {
+    ret_value.significand = ret_value.significand % 2 == 0
+                                ? ret_value.significand
+                                : ret_value.significand - 1;
+  } else if (ret_value.significand < xi) {
+    ++ret_value.significand;
+  }
+  return ret_value;
+}
+
+template <typename T> auto to_decimal(T x) noexcept -> decimal_fp<T> {
+  // Step 1: integer promotion & Schubfach multiplier calculation.
+
+  using carrier_uint = typename float_info<T>::carrier_uint;
+  using cache_entry_type = typename cache_accessor<T>::cache_entry_type;
+  auto br = bit_cast<carrier_uint>(x);
+
+  // Extract significand bits and exponent bits.
+  const carrier_uint significand_mask =
+      (static_cast<carrier_uint>(1) << num_significand_bits<T>()) - 1;
+  carrier_uint significand = (br & significand_mask);
+  int exponent =
+      static_cast<int>((br & exponent_mask<T>()) >> num_significand_bits<T>());
+
+  if (exponent != 0) {  // Check if normal.
+    exponent -= exponent_bias<T>() + num_significand_bits<T>();
+
+    // Shorter interval case; proceed like Schubfach.
+    // In fact, when exponent == 1 and significand == 0, the interval is
+    // regular. However, it can be shown that the end-results are anyway same.
+    if (significand == 0) return shorter_interval_case<T>(exponent);
+
+    significand |= (static_cast<carrier_uint>(1) << num_significand_bits<T>());
+  } else {
+    // Subnormal case; the interval is always regular.
+    if (significand == 0) return {0, 0};
+    exponent =
+        std::numeric_limits<T>::min_exponent - num_significand_bits<T>() - 1;
+  }
+
+  const bool include_left_endpoint = (significand % 2 == 0);
+  const bool include_right_endpoint = include_left_endpoint;
+
+  // Compute k and beta.
+  const int minus_k = floor_log10_pow2(exponent) - float_info<T>::kappa;
+  const cache_entry_type cache = cache_accessor<T>::get_cached_power(-minus_k);
+  const int beta = exponent + floor_log2_pow10(-minus_k);
+
+  // Compute zi and deltai.
+  // 10^kappa <= deltai < 10^(kappa + 1)
+  const uint32_t deltai = cache_accessor<T>::compute_delta(cache, beta);
+  const carrier_uint two_fc = significand << 1;
+
+  // For the case of binary32, the result of integer check is not correct for
+  // 29711844 * 2^-82
+  // = 6.1442653300000000008655037797566933477355632930994033813476... * 10^-18
+  // and 29711844 * 2^-81
+  // = 1.2288530660000000001731007559513386695471126586198806762695... * 10^-17,
+  // and they are the unique counterexamples. However, since 29711844 is even,
+  // this does not cause any problem for the endpoints calculations; it can only
+  // cause a problem when we need to perform integer check for the center.
+  // Fortunately, with these inputs, that branch is never executed, so we are
+  // fine.
+  const typename cache_accessor<T>::compute_mul_result z_mul =
+      cache_accessor<T>::compute_mul((two_fc | 1) << beta, cache);
+
+  // Step 2: Try larger divisor; remove trailing zeros if necessary.
+
+  // Using an upper bound on zi, we might be able to optimize the division
+  // better than the compiler; we are computing zi / big_divisor here.
+  decimal_fp<T> ret_value;
+  ret_value.significand = divide_by_10_to_kappa_plus_1(z_mul.result);
+  uint32_t r = static_cast<uint32_t>(z_mul.result - float_info<T>::big_divisor *
+                                                        ret_value.significand);
+
+  if (r < deltai) {
+    // Exclude the right endpoint if necessary.
+    if (r == 0 && (z_mul.is_integer & !include_right_endpoint)) {
+      --ret_value.significand;
+      r = float_info<T>::big_divisor;
+      goto small_divisor_case_label;
+    }
+  } else if (r > deltai) {
+    goto small_divisor_case_label;
+  } else {
+    // r == deltai; compare fractional parts.
+    const typename cache_accessor<T>::compute_mul_parity_result x_mul =
+        cache_accessor<T>::compute_mul_parity(two_fc - 1, cache, beta);
+
+    if (!(x_mul.parity | (x_mul.is_integer & include_left_endpoint)))
+      goto small_divisor_case_label;
+  }
+  ret_value.exponent = minus_k + float_info<T>::kappa + 1;
+
+  // We may need to remove trailing zeros.
+  ret_value.exponent += remove_trailing_zeros(ret_value.significand);
+  return ret_value;
+
+  // Step 3: Find the significand with the smaller divisor.
+
+small_divisor_case_label:
+  ret_value.significand *= 10;
+  ret_value.exponent = minus_k + float_info<T>::kappa;
+
+  uint32_t dist = r - (deltai / 2) + (float_info<T>::small_divisor / 2);
+  const bool approx_y_parity =
+      ((dist ^ (float_info<T>::small_divisor / 2)) & 1) != 0;
+
+  // Is dist divisible by 10^kappa?
+  const bool divisible_by_small_divisor =
+      check_divisibility_and_divide_by_pow10<float_info<T>::kappa>(dist);
+
+  // Add dist / 10^kappa to the significand.
+  ret_value.significand += dist;
+
+  if (!divisible_by_small_divisor) return ret_value;
+
+  // Check z^(f) >= epsilon^(f).
+  // We have either yi == zi - epsiloni or yi == (zi - epsiloni) - 1,
+  // where yi == zi - epsiloni if and only if z^(f) >= epsilon^(f).
+  // Since there are only 2 possibilities, we only need to care about the
+  // parity. Also, zi and r should have the same parity since the divisor
+  // is an even number.
+  const auto y_mul = cache_accessor<T>::compute_mul_parity(two_fc, cache, beta);
+
+  // If z^(f) >= epsilon^(f), we might have a tie when z^(f) == epsilon^(f),
+  // or equivalently, when y is an integer.
+  if (y_mul.parity != approx_y_parity)
+    --ret_value.significand;
+  else if (y_mul.is_integer & (ret_value.significand % 2 != 0))
+    --ret_value.significand;
+  return ret_value;
+}
+}  // namespace dragonbox
+}  // namespace detail
+
+template <> struct formatter<detail::bigint> {
+  FMT_CONSTEXPR auto parse(format_parse_context& ctx)
+      -> format_parse_context::iterator {
+    return ctx.begin();
+  }
+
+  auto format(const detail::bigint& n, format_context& ctx) const
+      -> format_context::iterator {
+    auto out = ctx.out();
+    bool first = true;
+    for (auto i = n.bigits_.size(); i > 0; --i) {
+      auto value = n.bigits_[i - 1u];
+      if (first) {
+        out = fmt::format_to(out, FMT_STRING("{:x}"), value);
+        first = false;
+        continue;
+      }
+      out = fmt::format_to(out, FMT_STRING("{:08x}"), value);
+    }
+    if (n.exp_ > 0)
+      out = fmt::format_to(out, FMT_STRING("p{}"),
+                           n.exp_ * detail::bigint::bigit_bits);
+    return out;
+  }
+};
+
+FMT_FUNC detail::utf8_to_utf16::utf8_to_utf16(string_view s) {
+  for_each_codepoint(s, [this](uint32_t cp, string_view) {
+    if (cp == invalid_code_point) FMT_THROW(std::runtime_error("invalid utf8"));
+    if (cp <= 0xFFFF) {
+      buffer_.push_back(static_cast<wchar_t>(cp));
+    } else {
+      cp -= 0x10000;
+      buffer_.push_back(static_cast<wchar_t>(0xD800 + (cp >> 10)));
+      buffer_.push_back(static_cast<wchar_t>(0xDC00 + (cp & 0x3FF)));
+    }
+    return true;
+  });
+  buffer_.push_back(0);
+}
+
+FMT_FUNC void format_system_error(detail::buffer<char>& out, int error_code,
+                                  const char* message) noexcept {
+  FMT_TRY {
+    auto ec = std::error_code(error_code, std::generic_category());
+    detail::write(appender(out), std::system_error(ec, message).what());
+    return;
+  }
+  FMT_CATCH(...) {}
+  format_error_code(out, error_code, message);
+}
+
+FMT_FUNC void report_system_error(int error_code,
+                                  const char* message) noexcept {
+  report_error(format_system_error, error_code, message);
+}
+
+FMT_FUNC auto vformat(string_view fmt, format_args args) -> std::string {
+  // Don't optimize the "{}" case to keep the binary size small and because it
+  // can be better optimized in fmt::format anyway.
+  auto buffer = memory_buffer();
+  detail::vformat_to(buffer, fmt, args);
+  return to_string(buffer);
+}
+
+namespace detail {
+
+template <typename T> struct span {
+  T* data;
+  size_t size;
+};
+
+template <typename F> auto flockfile(F* f) -> decltype(_lock_file(f)) {
+  _lock_file(f);
+}
+template <typename F> auto funlockfile(F* f) -> decltype(_unlock_file(f)) {
+  _unlock_file(f);
+}
+
+#ifndef getc_unlocked
+template <typename F> auto getc_unlocked(F* f) -> decltype(_fgetc_nolock(f)) {
+  return _fgetc_nolock(f);
+}
+#endif
+
+template <typename F = FILE, typename Enable = void>
+struct has_flockfile : std::false_type {};
+
+template <typename F>
+struct has_flockfile<F, void_t<decltype(flockfile(&std::declval<F&>()))>>
+    : std::true_type {};
+
+// A FILE wrapper. F is FILE defined as a template parameter to make system API
+// detection work.
+template <typename F> class file_base {
+ public:
+  F* file_;
+
+ public:
+  file_base(F* file) : file_(file) {}
+  operator F*() const { return file_; }
+
+  // Reads a code unit from the stream.
+  auto get() -> int {
+    int result = getc_unlocked(file_);
+    if (result == EOF && ferror(file_) != 0)
+      FMT_THROW(system_error(errno, FMT_STRING("getc failed")));
+    return result;
+  }
+
+  // Puts the code unit back into the stream buffer.
+  void unget(char c) {
+    if (ungetc(c, file_) == EOF)
+      FMT_THROW(system_error(errno, FMT_STRING("ungetc failed")));
+  }
+
+  void flush() { fflush(this->file_); }
+};
+
+// A FILE wrapper for glibc.
+template <typename F> class glibc_file : public file_base<F> {
+ private:
+  enum {
+    line_buffered = 0x200,  // _IO_LINE_BUF
+    unbuffered = 2          // _IO_UNBUFFERED
+  };
+
+ public:
+  using file_base<F>::file_base;
+
+  auto is_buffered() const -> bool {
+    return (this->file_->_flags & unbuffered) == 0;
+  }
+
+  void init_buffer() {
+    if (this->file_->_IO_write_ptr) return;
+    // Force buffer initialization by placing and removing a char in a buffer.
+    putc_unlocked(0, this->file_);
+    --this->file_->_IO_write_ptr;
+  }
+
+  // Returns the file's read buffer.
+  auto get_read_buffer() const -> span<const char> {
+    auto ptr = this->file_->_IO_read_ptr;
+    return {ptr, to_unsigned(this->file_->_IO_read_end - ptr)};
+  }
+
+  // Returns the file's write buffer.
+  auto get_write_buffer() const -> span<char> {
+    auto ptr = this->file_->_IO_write_ptr;
+    return {ptr, to_unsigned(this->file_->_IO_buf_end - ptr)};
+  }
+
+  void advance_write_buffer(size_t size) { this->file_->_IO_write_ptr += size; }
+
+  bool needs_flush() const {
+    if ((this->file_->_flags & line_buffered) == 0) return false;
+    char* end = this->file_->_IO_write_end;
+    return memchr(end, '\n', to_unsigned(this->file_->_IO_write_ptr - end));
+  }
+
+  void flush() { fflush_unlocked(this->file_); }
+};
+
+// A FILE wrapper for Apple's libc.
+template <typename F> class apple_file : public file_base<F> {
+ private:
+  enum {
+    line_buffered = 1,  // __SNBF
+    unbuffered = 2      // __SLBF
+  };
+
+ public:
+  using file_base<F>::file_base;
+
+  auto is_buffered() const -> bool {
+    return (this->file_->_flags & unbuffered) == 0;
+  }
+
+  void init_buffer() {
+    if (this->file_->_p) return;
+    // Force buffer initialization by placing and removing a char in a buffer.
+    putc_unlocked(0, this->file_);
+    --this->file_->_p;
+    ++this->file_->_w;
+  }
+
+  auto get_read_buffer() const -> span<const char> {
+    return {reinterpret_cast<char*>(this->file_->_p),
+            to_unsigned(this->file_->_r)};
+  }
+
+  auto get_write_buffer() const -> span<char> {
+    return {reinterpret_cast<char*>(this->file_->_p),
+            to_unsigned(this->file_->_bf._base + this->file_->_bf._size -
+                        this->file_->_p)};
+  }
+
+  void advance_write_buffer(size_t size) {
+    this->file_->_p += size;
+    this->file_->_w -= size;
+  }
+
+  bool needs_flush() const {
+    if ((this->file_->_flags & line_buffered) == 0) return false;
+    return memchr(this->file_->_p + this->file_->_w, '\n',
+                  to_unsigned(-this->file_->_w));
+  }
+};
+
+// A fallback FILE wrapper.
+template <typename F> class fallback_file : public file_base<F> {
+ private:
+  char next_;  // The next unconsumed character in the buffer.
+  bool has_next_ = false;
+
+ public:
+  using file_base<F>::file_base;
+
+  auto is_buffered() const -> bool { return false; }
+  auto needs_flush() const -> bool { return false; }
+  void init_buffer() {}
+
+  auto get_read_buffer() const -> span<const char> {
+    return {&next_, has_next_ ? 1u : 0u};
+  }
+
+  auto get_write_buffer() const -> span<char> { return {nullptr, 0}; }
+
+  void advance_write_buffer(size_t) {}
+
+  auto get() -> int {
+    has_next_ = false;
+    return file_base<F>::get();
+  }
+
+  void unget(char c) {
+    file_base<F>::unget(c);
+    next_ = c;
+    has_next_ = true;
+  }
+};
+
+#ifndef FMT_USE_FALLBACK_FILE
+#  define FMT_USE_FALLBACK_FILE 1
+#endif
+
+template <typename F,
+          FMT_ENABLE_IF(sizeof(F::_p) != 0 && !FMT_USE_FALLBACK_FILE)>
+auto get_file(F* f, int) -> apple_file<F> {
+  return f;
+}
+template <typename F,
+          FMT_ENABLE_IF(sizeof(F::_IO_read_ptr) != 0 && !FMT_USE_FALLBACK_FILE)>
+inline auto get_file(F* f, int) -> glibc_file<F> {
+  return f;
+}
+
+inline auto get_file(FILE* f, ...) -> fallback_file<FILE> { return f; }
+
+using file_ref = decltype(get_file(static_cast<FILE*>(nullptr), 0));
+
+template <typename F = FILE, typename Enable = void>
+class file_print_buffer : public buffer<char> {
+ public:
+  explicit file_print_buffer(F*) : buffer(nullptr, size_t()) {}
+};
+
+template <typename F>
+class file_print_buffer<F, enable_if_t<has_flockfile<F>::value>>
+    : public buffer<char> {
+ private:
+  file_ref file_;
+
+  static void grow(buffer<char>& base, size_t) {
+    auto& self = static_cast<file_print_buffer&>(base);
+    self.file_.advance_write_buffer(self.size());
+    if (self.file_.get_write_buffer().size == 0) self.file_.flush();
+    auto buf = self.file_.get_write_buffer();
+    FMT_ASSERT(buf.size > 0, "");
+    self.set(buf.data, buf.size);
+    self.clear();
+  }
+
+ public:
+  explicit file_print_buffer(F* f) : buffer(grow, size_t()), file_(f) {
+    flockfile(f);
+    file_.init_buffer();
+    auto buf = file_.get_write_buffer();
+    set(buf.data, buf.size);
+  }
+  ~file_print_buffer() {
+    file_.advance_write_buffer(size());
+    bool flush = file_.needs_flush();
+    F* f = file_;    // Make funlockfile depend on the template parameter F
+    funlockfile(f);  // for the system API detection to work.
+    if (flush) fflush(file_);
+  }
+};
+
+#if !defined(_WIN32) || defined(FMT_USE_WRITE_CONSOLE)
+FMT_FUNC auto write_console(int, string_view) -> bool { return false; }
+#else
+using dword = conditional_t<sizeof(long) == 4, unsigned long, unsigned>;
+extern "C" __declspec(dllimport) int __stdcall WriteConsoleW(  //
+    void*, const void*, dword, dword*, void*);
+
+FMT_FUNC bool write_console(int fd, string_view text) {
+  auto u16 = utf8_to_utf16(text);
+  return WriteConsoleW(reinterpret_cast<void*>(_get_osfhandle(fd)), u16.c_str(),
+                       static_cast<dword>(u16.size()), nullptr, nullptr) != 0;
+}
+#endif
+
+#ifdef _WIN32
+// Print assuming legacy (non-Unicode) encoding.
+FMT_FUNC void vprint_mojibake(std::FILE* f, string_view fmt, format_args args,
+                              bool newline) {
+  auto buffer = memory_buffer();
+  detail::vformat_to(buffer, fmt, args);
+  if (newline) buffer.push_back('\n');
+  fwrite_fully(buffer.data(), buffer.size(), f);
+}
+#endif
+
+FMT_FUNC void print(std::FILE* f, string_view text) {
+#if defined(_WIN32) && !defined(FMT_USE_WRITE_CONSOLE)
+  int fd = _fileno(f);
+  if (_isatty(fd)) {
+    std::fflush(f);
+    if (write_console(fd, text)) return;
+  }
+#endif
+  fwrite_fully(text.data(), text.size(), f);
+}
+}  // namespace detail
+
+FMT_FUNC void vprint_buffered(std::FILE* f, string_view fmt, format_args args) {
+  auto buffer = memory_buffer();
+  detail::vformat_to(buffer, fmt, args);
+  detail::print(f, {buffer.data(), buffer.size()});
+}
+
+FMT_FUNC void vprint(std::FILE* f, string_view fmt, format_args args) {
+  if (!detail::file_ref(f).is_buffered() || !detail::has_flockfile<>())
+    return vprint_buffered(f, fmt, args);
+  auto&& buffer = detail::file_print_buffer<>(f);
+  return detail::vformat_to(buffer, fmt, args);
+}
+
+FMT_FUNC void vprintln(std::FILE* f, string_view fmt, format_args args) {
+  auto buffer = memory_buffer();
+  detail::vformat_to(buffer, fmt, args);
+  buffer.push_back('\n');
+  detail::print(f, {buffer.data(), buffer.size()});
+}
+
+FMT_FUNC void vprint(string_view fmt, format_args args) {
+  vprint(stdout, fmt, args);
+}
+
+namespace detail {
+
+struct singleton {
+  unsigned char upper;
+  unsigned char lower_count;
+};
+
+inline auto is_printable(uint16_t x, const singleton* singletons,
+                         size_t singletons_size,
+                         const unsigned char* singleton_lowers,
+                         const unsigned char* normal, size_t normal_size)
+    -> bool {
+  auto upper = x >> 8;
+  auto lower_start = 0;
+  for (size_t i = 0; i < singletons_size; ++i) {
+    auto s = singletons[i];
+    auto lower_end = lower_start + s.lower_count;
+    if (upper < s.upper) break;
+    if (upper == s.upper) {
+      for (auto j = lower_start; j < lower_end; ++j) {
+        if (singleton_lowers[j] == (x & 0xff)) return false;
+      }
+    }
+    lower_start = lower_end;
+  }
+
+  auto xsigned = static_cast<int>(x);
+  auto current = true;
+  for (size_t i = 0; i < normal_size; ++i) {
+    auto v = static_cast<int>(normal[i]);
+    auto len = (v & 0x80) != 0 ? (v & 0x7f) << 8 | normal[++i] : v;
+    xsigned -= len;
+    if (xsigned < 0) break;
+    current = !current;
+  }
+  return current;
+}
+
+// This code is generated by support/printable.py.
+FMT_FUNC auto is_printable(uint32_t cp) -> bool {
+  static constexpr singleton singletons0[] = {
+      {0x00, 1},  {0x03, 5},  {0x05, 6},  {0x06, 3},  {0x07, 6},  {0x08, 8},
+      {0x09, 17}, {0x0a, 28}, {0x0b, 25}, {0x0c, 20}, {0x0d, 16}, {0x0e, 13},
+      {0x0f, 4},  {0x10, 3},  {0x12, 18}, {0x13, 9},  {0x16, 1},  {0x17, 5},
+      {0x18, 2},  {0x19, 3},  {0x1a, 7},  {0x1c, 2},  {0x1d, 1},  {0x1f, 22},
+      {0x20, 3},  {0x2b, 3},  {0x2c, 2},  {0x2d, 11}, {0x2e, 1},  {0x30, 3},
+      {0x31, 2},  {0x32, 1},  {0xa7, 2},  {0xa9, 2},  {0xaa, 4},  {0xab, 8},
+      {0xfa, 2},  {0xfb, 5},  {0xfd, 4},  {0xfe, 3},  {0xff, 9},
+  };
+  static constexpr unsigned char singletons0_lower[] = {
+      0xad, 0x78, 0x79, 0x8b, 0x8d, 0xa2, 0x30, 0x57, 0x58, 0x8b, 0x8c, 0x90,
+      0x1c, 0x1d, 0xdd, 0x0e, 0x0f, 0x4b, 0x4c, 0xfb, 0xfc, 0x2e, 0x2f, 0x3f,
+      0x5c, 0x5d, 0x5f, 0xb5, 0xe2, 0x84, 0x8d, 0x8e, 0x91, 0x92, 0xa9, 0xb1,
+      0xba, 0xbb, 0xc5, 0xc6, 0xc9, 0xca, 0xde, 0xe4, 0xe5, 0xff, 0x00, 0x04,
+      0x11, 0x12, 0x29, 0x31, 0x34, 0x37, 0x3a, 0x3b, 0x3d, 0x49, 0x4a, 0x5d,
+      0x84, 0x8e, 0x92, 0xa9, 0xb1, 0xb4, 0xba, 0xbb, 0xc6, 0xca, 0xce, 0xcf,
+      0xe4, 0xe5, 0x00, 0x04, 0x0d, 0x0e, 0x11, 0x12, 0x29, 0x31, 0x34, 0x3a,
+      0x3b, 0x45, 0x46, 0x49, 0x4a, 0x5e, 0x64, 0x65, 0x84, 0x91, 0x9b, 0x9d,
+      0xc9, 0xce, 0xcf, 0x0d, 0x11, 0x29, 0x45, 0x49, 0x57, 0x64, 0x65, 0x8d,
+      0x91, 0xa9, 0xb4, 0xba, 0xbb, 0xc5, 0xc9, 0xdf, 0xe4, 0xe5, 0xf0, 0x0d,
+      0x11, 0x45, 0x49, 0x64, 0x65, 0x80, 0x84, 0xb2, 0xbc, 0xbe, 0xbf, 0xd5,
+      0xd7, 0xf0, 0xf1, 0x83, 0x85, 0x8b, 0xa4, 0xa6, 0xbe, 0xbf, 0xc5, 0xc7,
+      0xce, 0xcf, 0xda, 0xdb, 0x48, 0x98, 0xbd, 0xcd, 0xc6, 0xce, 0xcf, 0x49,
+      0x4e, 0x4f, 0x57, 0x59, 0x5e, 0x5f, 0x89, 0x8e, 0x8f, 0xb1, 0xb6, 0xb7,
+      0xbf, 0xc1, 0xc6, 0xc7, 0xd7, 0x11, 0x16, 0x17, 0x5b, 0x5c, 0xf6, 0xf7,
+      0xfe, 0xff, 0x80, 0x0d, 0x6d, 0x71, 0xde, 0xdf, 0x0e, 0x0f, 0x1f, 0x6e,
+      0x6f, 0x1c, 0x1d, 0x5f, 0x7d, 0x7e, 0xae, 0xaf, 0xbb, 0xbc, 0xfa, 0x16,
+      0x17, 0x1e, 0x1f, 0x46, 0x47, 0x4e, 0x4f, 0x58, 0x5a, 0x5c, 0x5e, 0x7e,
+      0x7f, 0xb5, 0xc5, 0xd4, 0xd5, 0xdc, 0xf0, 0xf1, 0xf5, 0x72, 0x73, 0x8f,
+      0x74, 0x75, 0x96, 0x2f, 0x5f, 0x26, 0x2e, 0x2f, 0xa7, 0xaf, 0xb7, 0xbf,
+      0xc7, 0xcf, 0xd7, 0xdf, 0x9a, 0x40, 0x97, 0x98, 0x30, 0x8f, 0x1f, 0xc0,
+      0xc1, 0xce, 0xff, 0x4e, 0x4f, 0x5a, 0x5b, 0x07, 0x08, 0x0f, 0x10, 0x27,
+      0x2f, 0xee, 0xef, 0x6e, 0x6f, 0x37, 0x3d, 0x3f, 0x42, 0x45, 0x90, 0x91,
+      0xfe, 0xff, 0x53, 0x67, 0x75, 0xc8, 0xc9, 0xd0, 0xd1, 0xd8, 0xd9, 0xe7,
+      0xfe, 0xff,
+  };
+  static constexpr singleton singletons1[] = {
+      {0x00, 6},  {0x01, 1}, {0x03, 1},  {0x04, 2}, {0x08, 8},  {0x09, 2},
+      {0x0a, 5},  {0x0b, 2}, {0x0e, 4},  {0x10, 1}, {0x11, 2},  {0x12, 5},
+      {0x13, 17}, {0x14, 1}, {0x15, 2},  {0x17, 2}, {0x19, 13}, {0x1c, 5},
+      {0x1d, 8},  {0x24, 1}, {0x6a, 3},  {0x6b, 2}, {0xbc, 2},  {0xd1, 2},
+      {0xd4, 12}, {0xd5, 9}, {0xd6, 2},  {0xd7, 2}, {0xda, 1},  {0xe0, 5},
+      {0xe1, 2},  {0xe8, 2}, {0xee, 32}, {0xf0, 4}, {0xf8, 2},  {0xf9, 2},
+      {0xfa, 2},  {0xfb, 1},
+  };
+  static constexpr unsigned char singletons1_lower[] = {
+      0x0c, 0x27, 0x3b, 0x3e, 0x4e, 0x4f, 0x8f, 0x9e, 0x9e, 0x9f, 0x06, 0x07,
+      0x09, 0x36, 0x3d, 0x3e, 0x56, 0xf3, 0xd0, 0xd1, 0x04, 0x14, 0x18, 0x36,
+      0x37, 0x56, 0x57, 0x7f, 0xaa, 0xae, 0xaf, 0xbd, 0x35, 0xe0, 0x12, 0x87,
+      0x89, 0x8e, 0x9e, 0x04, 0x0d, 0x0e, 0x11, 0x12, 0x29, 0x31, 0x34, 0x3a,
+      0x45, 0x46, 0x49, 0x4a, 0x4e, 0x4f, 0x64, 0x65, 0x5c, 0xb6, 0xb7, 0x1b,
+      0x1c, 0x07, 0x08, 0x0a, 0x0b, 0x14, 0x17, 0x36, 0x39, 0x3a, 0xa8, 0xa9,
+      0xd8, 0xd9, 0x09, 0x37, 0x90, 0x91, 0xa8, 0x07, 0x0a, 0x3b, 0x3e, 0x66,
+      0x69, 0x8f, 0x92, 0x6f, 0x5f, 0xee, 0xef, 0x5a, 0x62, 0x9a, 0x9b, 0x27,
+      0x28, 0x55, 0x9d, 0xa0, 0xa1, 0xa3, 0xa4, 0xa7, 0xa8, 0xad, 0xba, 0xbc,
+      0xc4, 0x06, 0x0b, 0x0c, 0x15, 0x1d, 0x3a, 0x3f, 0x45, 0x51, 0xa6, 0xa7,
+      0xcc, 0xcd, 0xa0, 0x07, 0x19, 0x1a, 0x22, 0x25, 0x3e, 0x3f, 0xc5, 0xc6,
+      0x04, 0x20, 0x23, 0x25, 0x26, 0x28, 0x33, 0x38, 0x3a, 0x48, 0x4a, 0x4c,
+      0x50, 0x53, 0x55, 0x56, 0x58, 0x5a, 0x5c, 0x5e, 0x60, 0x63, 0x65, 0x66,
+      0x6b, 0x73, 0x78, 0x7d, 0x7f, 0x8a, 0xa4, 0xaa, 0xaf, 0xb0, 0xc0, 0xd0,
+      0xae, 0xaf, 0x79, 0xcc, 0x6e, 0x6f, 0x93,
+  };
+  static constexpr unsigned char normal0[] = {
+      0x00, 0x20, 0x5f, 0x22, 0x82, 0xdf, 0x04, 0x82, 0x44, 0x08, 0x1b, 0x04,
+      0x06, 0x11, 0x81, 0xac, 0x0e, 0x80, 0xab, 0x35, 0x28, 0x0b, 0x80, 0xe0,
+      0x03, 0x19, 0x08, 0x01, 0x04, 0x2f, 0x04, 0x34, 0x04, 0x07, 0x03, 0x01,
+      0x07, 0x06, 0x07, 0x11, 0x0a, 0x50, 0x0f, 0x12, 0x07, 0x55, 0x07, 0x03,
+      0x04, 0x1c, 0x0a, 0x09, 0x03, 0x08, 0x03, 0x07, 0x03, 0x02, 0x03, 0x03,
+      0x03, 0x0c, 0x04, 0x05, 0x03, 0x0b, 0x06, 0x01, 0x0e, 0x15, 0x05, 0x3a,
+      0x03, 0x11, 0x07, 0x06, 0x05, 0x10, 0x07, 0x57, 0x07, 0x02, 0x07, 0x15,
+      0x0d, 0x50, 0x04, 0x43, 0x03, 0x2d, 0x03, 0x01, 0x04, 0x11, 0x06, 0x0f,
+      0x0c, 0x3a, 0x04, 0x1d, 0x25, 0x5f, 0x20, 0x6d, 0x04, 0x6a, 0x25, 0x80,
+      0xc8, 0x05, 0x82, 0xb0, 0x03, 0x1a, 0x06, 0x82, 0xfd, 0x03, 0x59, 0x07,
+      0x15, 0x0b, 0x17, 0x09, 0x14, 0x0c, 0x14, 0x0c, 0x6a, 0x06, 0x0a, 0x06,
+      0x1a, 0x06, 0x59, 0x07, 0x2b, 0x05, 0x46, 0x0a, 0x2c, 0x04, 0x0c, 0x04,
+      0x01, 0x03, 0x31, 0x0b, 0x2c, 0x04, 0x1a, 0x06, 0x0b, 0x03, 0x80, 0xac,
+      0x06, 0x0a, 0x06, 0x21, 0x3f, 0x4c, 0x04, 0x2d, 0x03, 0x74, 0x08, 0x3c,
+      0x03, 0x0f, 0x03, 0x3c, 0x07, 0x38, 0x08, 0x2b, 0x05, 0x82, 0xff, 0x11,
+      0x18, 0x08, 0x2f, 0x11, 0x2d, 0x03, 0x20, 0x10, 0x21, 0x0f, 0x80, 0x8c,
+      0x04, 0x82, 0x97, 0x19, 0x0b, 0x15, 0x88, 0x94, 0x05, 0x2f, 0x05, 0x3b,
+      0x07, 0x02, 0x0e, 0x18, 0x09, 0x80, 0xb3, 0x2d, 0x74, 0x0c, 0x80, 0xd6,
+      0x1a, 0x0c, 0x05, 0x80, 0xff, 0x05, 0x80, 0xdf, 0x0c, 0xee, 0x0d, 0x03,
+      0x84, 0x8d, 0x03, 0x37, 0x09, 0x81, 0x5c, 0x14, 0x80, 0xb8, 0x08, 0x80,
+      0xcb, 0x2a, 0x38, 0x03, 0x0a, 0x06, 0x38, 0x08, 0x46, 0x08, 0x0c, 0x06,
+      0x74, 0x0b, 0x1e, 0x03, 0x5a, 0x04, 0x59, 0x09, 0x80, 0x83, 0x18, 0x1c,
+      0x0a, 0x16, 0x09, 0x4c, 0x04, 0x80, 0x8a, 0x06, 0xab, 0xa4, 0x0c, 0x17,
+      0x04, 0x31, 0xa1, 0x04, 0x81, 0xda, 0x26, 0x07, 0x0c, 0x05, 0x05, 0x80,
+      0xa5, 0x11, 0x81, 0x6d, 0x10, 0x78, 0x28, 0x2a, 0x06, 0x4c, 0x04, 0x80,
+      0x8d, 0x04, 0x80, 0xbe, 0x03, 0x1b, 0x03, 0x0f, 0x0d,
+  };
+  static constexpr unsigned char normal1[] = {
+      0x5e, 0x22, 0x7b, 0x05, 0x03, 0x04, 0x2d, 0x03, 0x66, 0x03, 0x01, 0x2f,
+      0x2e, 0x80, 0x82, 0x1d, 0x03, 0x31, 0x0f, 0x1c, 0x04, 0x24, 0x09, 0x1e,
+      0x05, 0x2b, 0x05, 0x44, 0x04, 0x0e, 0x2a, 0x80, 0xaa, 0x06, 0x24, 0x04,
+      0x24, 0x04, 0x28, 0x08, 0x34, 0x0b, 0x01, 0x80, 0x90, 0x81, 0x37, 0x09,
+      0x16, 0x0a, 0x08, 0x80, 0x98, 0x39, 0x03, 0x63, 0x08, 0x09, 0x30, 0x16,
+      0x05, 0x21, 0x03, 0x1b, 0x05, 0x01, 0x40, 0x38, 0x04, 0x4b, 0x05, 0x2f,
+      0x04, 0x0a, 0x07, 0x09, 0x07, 0x40, 0x20, 0x27, 0x04, 0x0c, 0x09, 0x36,
+      0x03, 0x3a, 0x05, 0x1a, 0x07, 0x04, 0x0c, 0x07, 0x50, 0x49, 0x37, 0x33,
+      0x0d, 0x33, 0x07, 0x2e, 0x08, 0x0a, 0x81, 0x26, 0x52, 0x4e, 0x28, 0x08,
+      0x2a, 0x56, 0x1c, 0x14, 0x17, 0x09, 0x4e, 0x04, 0x1e, 0x0f, 0x43, 0x0e,
+      0x19, 0x07, 0x0a, 0x06, 0x48, 0x08, 0x27, 0x09, 0x75, 0x0b, 0x3f, 0x41,
+      0x2a, 0x06, 0x3b, 0x05, 0x0a, 0x06, 0x51, 0x06, 0x01, 0x05, 0x10, 0x03,
+      0x05, 0x80, 0x8b, 0x62, 0x1e, 0x48, 0x08, 0x0a, 0x80, 0xa6, 0x5e, 0x22,
+      0x45, 0x0b, 0x0a, 0x06, 0x0d, 0x13, 0x39, 0x07, 0x0a, 0x36, 0x2c, 0x04,
+      0x10, 0x80, 0xc0, 0x3c, 0x64, 0x53, 0x0c, 0x48, 0x09, 0x0a, 0x46, 0x45,
+      0x1b, 0x48, 0x08, 0x53, 0x1d, 0x39, 0x81, 0x07, 0x46, 0x0a, 0x1d, 0x03,
+      0x47, 0x49, 0x37, 0x03, 0x0e, 0x08, 0x0a, 0x06, 0x39, 0x07, 0x0a, 0x81,
+      0x36, 0x19, 0x80, 0xb7, 0x01, 0x0f, 0x32, 0x0d, 0x83, 0x9b, 0x66, 0x75,
+      0x0b, 0x80, 0xc4, 0x8a, 0xbc, 0x84, 0x2f, 0x8f, 0xd1, 0x82, 0x47, 0xa1,
+      0xb9, 0x82, 0x39, 0x07, 0x2a, 0x04, 0x02, 0x60, 0x26, 0x0a, 0x46, 0x0a,
+      0x28, 0x05, 0x13, 0x82, 0xb0, 0x5b, 0x65, 0x4b, 0x04, 0x39, 0x07, 0x11,
+      0x40, 0x05, 0x0b, 0x02, 0x0e, 0x97, 0xf8, 0x08, 0x84, 0xd6, 0x2a, 0x09,
+      0xa2, 0xf7, 0x81, 0x1f, 0x31, 0x03, 0x11, 0x04, 0x08, 0x81, 0x8c, 0x89,
+      0x04, 0x6b, 0x05, 0x0d, 0x03, 0x09, 0x07, 0x10, 0x93, 0x60, 0x80, 0xf6,
+      0x0a, 0x73, 0x08, 0x6e, 0x17, 0x46, 0x80, 0x9a, 0x14, 0x0c, 0x57, 0x09,
+      0x19, 0x80, 0x87, 0x81, 0x47, 0x03, 0x85, 0x42, 0x0f, 0x15, 0x85, 0x50,
+      0x2b, 0x80, 0xd5, 0x2d, 0x03, 0x1a, 0x04, 0x02, 0x81, 0x70, 0x3a, 0x05,
+      0x01, 0x85, 0x00, 0x80, 0xd7, 0x29, 0x4c, 0x04, 0x0a, 0x04, 0x02, 0x83,
+      0x11, 0x44, 0x4c, 0x3d, 0x80, 0xc2, 0x3c, 0x06, 0x01, 0x04, 0x55, 0x05,
+      0x1b, 0x34, 0x02, 0x81, 0x0e, 0x2c, 0x04, 0x64, 0x0c, 0x56, 0x0a, 0x80,
+      0xae, 0x38, 0x1d, 0x0d, 0x2c, 0x04, 0x09, 0x07, 0x02, 0x0e, 0x06, 0x80,
+      0x9a, 0x83, 0xd8, 0x08, 0x0d, 0x03, 0x0d, 0x03, 0x74, 0x0c, 0x59, 0x07,
+      0x0c, 0x14, 0x0c, 0x04, 0x38, 0x08, 0x0a, 0x06, 0x28, 0x08, 0x22, 0x4e,
+      0x81, 0x54, 0x0c, 0x15, 0x03, 0x03, 0x05, 0x07, 0x09, 0x19, 0x07, 0x07,
+      0x09, 0x03, 0x0d, 0x07, 0x29, 0x80, 0xcb, 0x25, 0x0a, 0x84, 0x06,
+  };
+  auto lower = static_cast<uint16_t>(cp);
+  if (cp < 0x10000) {
+    return is_printable(lower, singletons0,
+                        sizeof(singletons0) / sizeof(*singletons0),
+                        singletons0_lower, normal0, sizeof(normal0));
+  }
+  if (cp < 0x20000) {
+    return is_printable(lower, singletons1,
+                        sizeof(singletons1) / sizeof(*singletons1),
+                        singletons1_lower, normal1, sizeof(normal1));
+  }
+  if (0x2a6de <= cp && cp < 0x2a700) return false;
+  if (0x2b735 <= cp && cp < 0x2b740) return false;
+  if (0x2b81e <= cp && cp < 0x2b820) return false;
+  if (0x2cea2 <= cp && cp < 0x2ceb0) return false;
+  if (0x2ebe1 <= cp && cp < 0x2f800) return false;
+  if (0x2fa1e <= cp && cp < 0x30000) return false;
+  if (0x3134b <= cp && cp < 0xe0100) return false;
+  if (0xe01f0 <= cp && cp < 0x110000) return false;
+  return cp < 0x110000;
+}
+
+}  // namespace detail
+
+FMT_END_NAMESPACE
+
+#endif  // FMT_FORMAT_INL_H_
diff --git a/ext/spdlog/fmt/bundled/format.h b/ext/spdlog/fmt/bundled/format.h
new file mode 100644
index 0000000..67f0ab7
--- /dev/null
+++ b/ext/spdlog/fmt/bundled/format.h
@@ -0,0 +1,4427 @@
+/*
+  Formatting library for C++
+
+  Copyright (c) 2012 - present, Victor Zverovich
+
+  Permission is hereby granted, free of charge, to any person obtaining
+  a copy of this software and associated documentation files (the
+  "Software"), to deal in the Software without restriction, including
+  without limitation the rights to use, copy, modify, merge, publish,
+  distribute, sublicense, and/or sell copies of the Software, and to
+  permit persons to whom the Software is furnished to do so, subject to
+  the following conditions:
+
+  The above copyright notice and this permission notice shall be
+  included in all copies or substantial portions of the Software.
+
+  THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+  EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+  MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+  NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE
+  LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
+  OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
+  WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+
+  --- Optional exception to the license ---
+
+  As an exception, if, as a result of your compiling your source code, portions
+  of this Software are embedded into a machine-executable object form of such
+  source code, you may redistribute such embedded portions in such object form
+  without including the above copyright and permission notices.
+ */
+
+#ifndef FMT_FORMAT_H_
+#define FMT_FORMAT_H_
+
+#ifndef _LIBCPP_REMOVE_TRANSITIVE_INCLUDES
+#  define _LIBCPP_REMOVE_TRANSITIVE_INCLUDES
+#  define FMT_REMOVE_TRANSITIVE_INCLUDES
+#endif
+
+#include "base.h"
+
+#ifndef FMT_MODULE
+#  include <cmath>             // std::signbit
+#  include <cstdint>           // uint32_t
+#  include <cstring>           // std::memcpy
+#  include <initializer_list>  // std::initializer_list
+#  include <limits>            // std::numeric_limits
+#  if defined(__GLIBCXX__) && !defined(_GLIBCXX_USE_DUAL_ABI)
+// Workaround for pre gcc 5 libstdc++.
+#    include <memory>  // std::allocator_traits
+#  endif
+#  include <stdexcept>     // std::runtime_error
+#  include <string>        // std::string
+#  include <system_error>  // std::system_error
+
+// Checking FMT_CPLUSPLUS for warning suppression in MSVC.
+#  if FMT_HAS_INCLUDE(<bit>) && FMT_CPLUSPLUS > 201703L
+#    include <bit>  // std::bit_cast
+#  endif
+
+// libc++ supports string_view in pre-c++17.
+#  if FMT_HAS_INCLUDE(<string_view>) && \
+      (FMT_CPLUSPLUS >= 201703L || defined(_LIBCPP_VERSION))
+#    include <string_view>
+#    define FMT_USE_STRING_VIEW
+#  endif
+#endif  // FMT_MODULE
+
+#if defined __cpp_inline_variables && __cpp_inline_variables >= 201606L
+#  define FMT_INLINE_VARIABLE inline
+#else
+#  define FMT_INLINE_VARIABLE
+#endif
+
+#ifndef FMT_NO_UNIQUE_ADDRESS
+#  if FMT_CPLUSPLUS >= 202002L
+#    if FMT_HAS_CPP_ATTRIBUTE(no_unique_address)
+#      define FMT_NO_UNIQUE_ADDRESS [[no_unique_address]]
+// VS2019 v16.10 and later except clang-cl (https://reviews.llvm.org/D110485).
+#    elif (FMT_MSC_VERSION >= 1929) && !FMT_CLANG_VERSION
+#      define FMT_NO_UNIQUE_ADDRESS [[msvc::no_unique_address]]
+#    endif
+#  endif
+#endif
+#ifndef FMT_NO_UNIQUE_ADDRESS
+#  define FMT_NO_UNIQUE_ADDRESS
+#endif
+
+// Visibility when compiled as a shared library/object.
+#if defined(FMT_LIB_EXPORT) || defined(FMT_SHARED)
+#  define FMT_SO_VISIBILITY(value) FMT_VISIBILITY(value)
+#else
+#  define FMT_SO_VISIBILITY(value)
+#endif
+
+#ifdef __has_builtin
+#  define FMT_HAS_BUILTIN(x) __has_builtin(x)
+#else
+#  define FMT_HAS_BUILTIN(x) 0
+#endif
+
+#if FMT_GCC_VERSION || FMT_CLANG_VERSION
+#  define FMT_NOINLINE __attribute__((noinline))
+#else
+#  define FMT_NOINLINE
+#endif
+
+namespace std {
+template <> struct iterator_traits<fmt::appender> {
+  using iterator_category = output_iterator_tag;
+  using value_type = char;
+};
+}  // namespace std
+
+#ifndef FMT_THROW
+#  if FMT_EXCEPTIONS
+#    if FMT_MSC_VERSION || defined(__NVCC__)
+FMT_BEGIN_NAMESPACE
+namespace detail {
+template <typename Exception> inline void do_throw(const Exception& x) {
+  // Silence unreachable code warnings in MSVC and NVCC because these
+  // are nearly impossible to fix in a generic code.
+  volatile bool b = true;
+  if (b) throw x;
+}
+}  // namespace detail
+FMT_END_NAMESPACE
+#      define FMT_THROW(x) detail::do_throw(x)
+#    else
+#      define FMT_THROW(x) throw x
+#    endif
+#  else
+#    define FMT_THROW(x) \
+      ::fmt::detail::assert_fail(__FILE__, __LINE__, (x).what())
+#  endif
+#endif
+
+#ifndef FMT_MAYBE_UNUSED
+#  if FMT_HAS_CPP17_ATTRIBUTE(maybe_unused)
+#    define FMT_MAYBE_UNUSED [[maybe_unused]]
+#  else
+#    define FMT_MAYBE_UNUSED
+#  endif
+#endif
+
+#ifndef FMT_USE_USER_DEFINED_LITERALS
+// EDG based compilers (Intel, NVIDIA, Elbrus, etc), GCC and MSVC support UDLs.
+//
+// GCC before 4.9 requires a space in `operator"" _a` which is invalid in later
+// compiler versions.
+#  if (FMT_HAS_FEATURE(cxx_user_literals) || FMT_GCC_VERSION >= 409 || \
+       FMT_MSC_VERSION >= 1900) &&                                     \
+      (!defined(__EDG_VERSION__) || __EDG_VERSION__ >= /* UDL feature */ 480)
+#    define FMT_USE_USER_DEFINED_LITERALS 1
+#  else
+#    define FMT_USE_USER_DEFINED_LITERALS 0
+#  endif
+#endif
+
+// Defining FMT_REDUCE_INT_INSTANTIATIONS to 1, will reduce the number of
+// integer formatter template instantiations to just one by only using the
+// largest integer type. This results in a reduction in binary size but will
+// cause a decrease in integer formatting performance.
+#if !defined(FMT_REDUCE_INT_INSTANTIATIONS)
+#  define FMT_REDUCE_INT_INSTANTIATIONS 0
+#endif
+
+// __builtin_clz is broken in clang with Microsoft CodeGen:
+// https://github.com/fmtlib/fmt/issues/519.
+#if !FMT_MSC_VERSION
+#  if FMT_HAS_BUILTIN(__builtin_clz) || FMT_GCC_VERSION || FMT_ICC_VERSION
+#    define FMT_BUILTIN_CLZ(n) __builtin_clz(n)
+#  endif
+#  if FMT_HAS_BUILTIN(__builtin_clzll) || FMT_GCC_VERSION || FMT_ICC_VERSION
+#    define FMT_BUILTIN_CLZLL(n) __builtin_clzll(n)
+#  endif
+#endif
+
+// __builtin_ctz is broken in Intel Compiler Classic on Windows:
+// https://github.com/fmtlib/fmt/issues/2510.
+#ifndef __ICL
+#  if FMT_HAS_BUILTIN(__builtin_ctz) || FMT_GCC_VERSION || FMT_ICC_VERSION || \
+      defined(__NVCOMPILER)
+#    define FMT_BUILTIN_CTZ(n) __builtin_ctz(n)
+#  endif
+#  if FMT_HAS_BUILTIN(__builtin_ctzll) || FMT_GCC_VERSION || \
+      FMT_ICC_VERSION || defined(__NVCOMPILER)
+#    define FMT_BUILTIN_CTZLL(n) __builtin_ctzll(n)
+#  endif
+#endif
+
+#if FMT_MSC_VERSION
+#  include <intrin.h>  // _BitScanReverse[64], _BitScanForward[64], _umul128
+#endif
+
+// Some compilers masquerade as both MSVC and GCC-likes or otherwise support
+// __builtin_clz and __builtin_clzll, so only define FMT_BUILTIN_CLZ using the
+// MSVC intrinsics if the clz and clzll builtins are not available.
+#if FMT_MSC_VERSION && !defined(FMT_BUILTIN_CLZLL) && \
+    !defined(FMT_BUILTIN_CTZLL)
+FMT_BEGIN_NAMESPACE
+namespace detail {
+// Avoid Clang with Microsoft CodeGen's -Wunknown-pragmas warning.
+#  if !defined(__clang__)
+#    pragma intrinsic(_BitScanForward)
+#    pragma intrinsic(_BitScanReverse)
+#    if defined(_WIN64)
+#      pragma intrinsic(_BitScanForward64)
+#      pragma intrinsic(_BitScanReverse64)
+#    endif
+#  endif
+
+inline auto clz(uint32_t x) -> int {
+  unsigned long r = 0;
+  _BitScanReverse(&r, x);
+  FMT_ASSERT(x != 0, "");
+  // Static analysis complains about using uninitialized data
+  // "r", but the only way that can happen is if "x" is 0,
+  // which the callers guarantee to not happen.
+  FMT_MSC_WARNING(suppress : 6102)
+  return 31 ^ static_cast<int>(r);
+}
+#  define FMT_BUILTIN_CLZ(n) detail::clz(n)
+
+inline auto clzll(uint64_t x) -> int {
+  unsigned long r = 0;
+#  ifdef _WIN64
+  _BitScanReverse64(&r, x);
+#  else
+  // Scan the high 32 bits.
+  if (_BitScanReverse(&r, static_cast<uint32_t>(x >> 32)))
+    return 63 ^ static_cast<int>(r + 32);
+  // Scan the low 32 bits.
+  _BitScanReverse(&r, static_cast<uint32_t>(x));
+#  endif
+  FMT_ASSERT(x != 0, "");
+  FMT_MSC_WARNING(suppress : 6102)  // Suppress a bogus static analysis warning.
+  return 63 ^ static_cast<int>(r);
+}
+#  define FMT_BUILTIN_CLZLL(n) detail::clzll(n)
+
+inline auto ctz(uint32_t x) -> int {
+  unsigned long r = 0;
+  _BitScanForward(&r, x);
+  FMT_ASSERT(x != 0, "");
+  FMT_MSC_WARNING(suppress : 6102)  // Suppress a bogus static analysis warning.
+  return static_cast<int>(r);
+}
+#  define FMT_BUILTIN_CTZ(n) detail::ctz(n)
+
+inline auto ctzll(uint64_t x) -> int {
+  unsigned long r = 0;
+  FMT_ASSERT(x != 0, "");
+  FMT_MSC_WARNING(suppress : 6102)  // Suppress a bogus static analysis warning.
+#  ifdef _WIN64
+  _BitScanForward64(&r, x);
+#  else
+  // Scan the low 32 bits.
+  if (_BitScanForward(&r, static_cast<uint32_t>(x))) return static_cast<int>(r);
+  // Scan the high 32 bits.
+  _BitScanForward(&r, static_cast<uint32_t>(x >> 32));
+  r += 32;
+#  endif
+  return static_cast<int>(r);
+}
+#  define FMT_BUILTIN_CTZLL(n) detail::ctzll(n)
+}  // namespace detail
+FMT_END_NAMESPACE
+#endif
+
+FMT_BEGIN_NAMESPACE
+
+template <typename Char, typename Traits, typename Allocator>
+struct is_contiguous<std::basic_string<Char, Traits, Allocator>>
+    : std::true_type {};
+
+namespace detail {
+
+FMT_CONSTEXPR inline void abort_fuzzing_if(bool condition) {
+  ignore_unused(condition);
+#ifdef FMT_FUZZ
+  if (condition) throw std::runtime_error("fuzzing limit reached");
+#endif
+}
+
+#if defined(FMT_USE_STRING_VIEW)
+template <typename Char> using std_string_view = std::basic_string_view<Char>;
+#else
+template <typename T> struct std_string_view {};
+#endif
+
+// Implementation of std::bit_cast for pre-C++20.
+template <typename To, typename From, FMT_ENABLE_IF(sizeof(To) == sizeof(From))>
+FMT_CONSTEXPR20 auto bit_cast(const From& from) -> To {
+#ifdef __cpp_lib_bit_cast
+  if (is_constant_evaluated()) return std::bit_cast<To>(from);
+#endif
+  auto to = To();
+  // The cast suppresses a bogus -Wclass-memaccess on GCC.
+  std::memcpy(static_cast<void*>(&to), &from, sizeof(to));
+  return to;
+}
+
+inline auto is_big_endian() -> bool {
+#ifdef _WIN32
+  return false;
+#elif defined(__BIG_ENDIAN__)
+  return true;
+#elif defined(__BYTE_ORDER__) && defined(__ORDER_BIG_ENDIAN__)
+  return __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__;
+#else
+  struct bytes {
+    char data[sizeof(int)];
+  };
+  return bit_cast<bytes>(1).data[0] == 0;
+#endif
+}
+
+class uint128_fallback {
+ private:
+  uint64_t lo_, hi_;
+
+ public:
+  constexpr uint128_fallback(uint64_t hi, uint64_t lo) : lo_(lo), hi_(hi) {}
+  constexpr uint128_fallback(uint64_t value = 0) : lo_(value), hi_(0) {}
+
+  constexpr auto high() const noexcept -> uint64_t { return hi_; }
+  constexpr auto low() const noexcept -> uint64_t { return lo_; }
+
+  template <typename T, FMT_ENABLE_IF(std::is_integral<T>::value)>
+  constexpr explicit operator T() const {
+    return static_cast<T>(lo_);
+  }
+
+  friend constexpr auto operator==(const uint128_fallback& lhs,
+                                   const uint128_fallback& rhs) -> bool {
+    return lhs.hi_ == rhs.hi_ && lhs.lo_ == rhs.lo_;
+  }
+  friend constexpr auto operator!=(const uint128_fallback& lhs,
+                                   const uint128_fallback& rhs) -> bool {
+    return !(lhs == rhs);
+  }
+  friend constexpr auto operator>(const uint128_fallback& lhs,
+                                  const uint128_fallback& rhs) -> bool {
+    return lhs.hi_ != rhs.hi_ ? lhs.hi_ > rhs.hi_ : lhs.lo_ > rhs.lo_;
+  }
+  friend constexpr auto operator|(const uint128_fallback& lhs,
+                                  const uint128_fallback& rhs)
+      -> uint128_fallback {
+    return {lhs.hi_ | rhs.hi_, lhs.lo_ | rhs.lo_};
+  }
+  friend constexpr auto operator&(const uint128_fallback& lhs,
+                                  const uint128_fallback& rhs)
+      -> uint128_fallback {
+    return {lhs.hi_ & rhs.hi_, lhs.lo_ & rhs.lo_};
+  }
+  friend constexpr auto operator~(const uint128_fallback& n)
+      -> uint128_fallback {
+    return {~n.hi_, ~n.lo_};
+  }
+  friend auto operator+(const uint128_fallback& lhs,
+                        const uint128_fallback& rhs) -> uint128_fallback {
+    auto result = uint128_fallback(lhs);
+    result += rhs;
+    return result;
+  }
+  friend auto operator*(const uint128_fallback& lhs, uint32_t rhs)
+      -> uint128_fallback {
+    FMT_ASSERT(lhs.hi_ == 0, "");
+    uint64_t hi = (lhs.lo_ >> 32) * rhs;
+    uint64_t lo = (lhs.lo_ & ~uint32_t()) * rhs;
+    uint64_t new_lo = (hi << 32) + lo;
+    return {(hi >> 32) + (new_lo < lo ? 1 : 0), new_lo};
+  }
+  friend auto operator-(const uint128_fallback& lhs, uint64_t rhs)
+      -> uint128_fallback {
+    return {lhs.hi_ - (lhs.lo_ < rhs ? 1 : 0), lhs.lo_ - rhs};
+  }
+  FMT_CONSTEXPR auto operator>>(int shift) const -> uint128_fallback {
+    if (shift == 64) return {0, hi_};
+    if (shift > 64) return uint128_fallback(0, hi_) >> (shift - 64);
+    return {hi_ >> shift, (hi_ << (64 - shift)) | (lo_ >> shift)};
+  }
+  FMT_CONSTEXPR auto operator<<(int shift) const -> uint128_fallback {
+    if (shift == 64) return {lo_, 0};
+    if (shift > 64) return uint128_fallback(lo_, 0) << (shift - 64);
+    return {hi_ << shift | (lo_ >> (64 - shift)), (lo_ << shift)};
+  }
+  FMT_CONSTEXPR auto operator>>=(int shift) -> uint128_fallback& {
+    return *this = *this >> shift;
+  }
+  FMT_CONSTEXPR void operator+=(uint128_fallback n) {
+    uint64_t new_lo = lo_ + n.lo_;
+    uint64_t new_hi = hi_ + n.hi_ + (new_lo < lo_ ? 1 : 0);
+    FMT_ASSERT(new_hi >= hi_, "");
+    lo_ = new_lo;
+    hi_ = new_hi;
+  }
+  FMT_CONSTEXPR void operator&=(uint128_fallback n) {
+    lo_ &= n.lo_;
+    hi_ &= n.hi_;
+  }
+
+  FMT_CONSTEXPR20 auto operator+=(uint64_t n) noexcept -> uint128_fallback& {
+    if (is_constant_evaluated()) {
+      lo_ += n;
+      hi_ += (lo_ < n ? 1 : 0);
+      return *this;
+    }
+#if FMT_HAS_BUILTIN(__builtin_addcll) && !defined(__ibmxl__)
+    unsigned long long carry;
+    lo_ = __builtin_addcll(lo_, n, 0, &carry);
+    hi_ += carry;
+#elif FMT_HAS_BUILTIN(__builtin_ia32_addcarryx_u64) && !defined(__ibmxl__)
+    unsigned long long result;
+    auto carry = __builtin_ia32_addcarryx_u64(0, lo_, n, &result);
+    lo_ = result;
+    hi_ += carry;
+#elif defined(_MSC_VER) && defined(_M_X64)
+    auto carry = _addcarry_u64(0, lo_, n, &lo_);
+    _addcarry_u64(carry, hi_, 0, &hi_);
+#else
+    lo_ += n;
+    hi_ += (lo_ < n ? 1 : 0);
+#endif
+    return *this;
+  }
+};
+
+using uint128_t = conditional_t<FMT_USE_INT128, uint128_opt, uint128_fallback>;
+
+#ifdef UINTPTR_MAX
+using uintptr_t = ::uintptr_t;
+#else
+using uintptr_t = uint128_t;
+#endif
+
+// Returns the largest possible value for type T. Same as
+// std::numeric_limits<T>::max() but shorter and not affected by the max macro.
+template <typename T> constexpr auto max_value() -> T {
+  return (std::numeric_limits<T>::max)();
+}
+template <typename T> constexpr auto num_bits() -> int {
+  return std::numeric_limits<T>::digits;
+}
+// std::numeric_limits<T>::digits may return 0 for 128-bit ints.
+template <> constexpr auto num_bits<int128_opt>() -> int { return 128; }
+template <> constexpr auto num_bits<uint128_opt>() -> int { return 128; }
+template <> constexpr auto num_bits<uint128_fallback>() -> int { return 128; }
+
+// A heterogeneous bit_cast used for converting 96-bit long double to uint128_t
+// and 128-bit pointers to uint128_fallback.
+template <typename To, typename From, FMT_ENABLE_IF(sizeof(To) > sizeof(From))>
+inline auto bit_cast(const From& from) -> To {
+  constexpr auto size = static_cast<int>(sizeof(From) / sizeof(unsigned));
+  struct data_t {
+    unsigned value[static_cast<unsigned>(size)];
+  } data = bit_cast<data_t>(from);
+  auto result = To();
+  if (const_check(is_big_endian())) {
+    for (int i = 0; i < size; ++i)
+      result = (result << num_bits<unsigned>()) | data.value[i];
+  } else {
+    for (int i = size - 1; i >= 0; --i)
+      result = (result << num_bits<unsigned>()) | data.value[i];
+  }
+  return result;
+}
+
+template <typename UInt>
+FMT_CONSTEXPR20 inline auto countl_zero_fallback(UInt n) -> int {
+  int lz = 0;
+  constexpr UInt msb_mask = static_cast<UInt>(1) << (num_bits<UInt>() - 1);
+  for (; (n & msb_mask) == 0; n <<= 1) lz++;
+  return lz;
+}
+
+FMT_CONSTEXPR20 inline auto countl_zero(uint32_t n) -> int {
+#ifdef FMT_BUILTIN_CLZ
+  if (!is_constant_evaluated()) return FMT_BUILTIN_CLZ(n);
+#endif
+  return countl_zero_fallback(n);
+}
+
+FMT_CONSTEXPR20 inline auto countl_zero(uint64_t n) -> int {
+#ifdef FMT_BUILTIN_CLZLL
+  if (!is_constant_evaluated()) return FMT_BUILTIN_CLZLL(n);
+#endif
+  return countl_zero_fallback(n);
+}
+
+FMT_INLINE void assume(bool condition) {
+  (void)condition;
+#if FMT_HAS_BUILTIN(__builtin_assume) && !FMT_ICC_VERSION
+  __builtin_assume(condition);
+#elif FMT_GCC_VERSION
+  if (!condition) __builtin_unreachable();
+#endif
+}
+
+// An approximation of iterator_t for pre-C++20 systems.
+template <typename T>
+using iterator_t = decltype(std::begin(std::declval<T&>()));
+template <typename T> using sentinel_t = decltype(std::end(std::declval<T&>()));
+
+// A workaround for std::string not having mutable data() until C++17.
+template <typename Char>
+inline auto get_data(std::basic_string<Char>& s) -> Char* {
+  return &s[0];
+}
+template <typename Container>
+inline auto get_data(Container& c) -> typename Container::value_type* {
+  return c.data();
+}
+
+// Attempts to reserve space for n extra characters in the output range.
+// Returns a pointer to the reserved range or a reference to it.
+template <typename OutputIt,
+          FMT_ENABLE_IF(is_back_insert_iterator<OutputIt>::value&&
+                            is_contiguous<typename OutputIt::container>::value)>
+#if FMT_CLANG_VERSION >= 307 && !FMT_ICC_VERSION
+__attribute__((no_sanitize("undefined")))
+#endif
+inline auto
+reserve(OutputIt it, size_t n) -> typename OutputIt::value_type* {
+  auto& c = get_container(it);
+  size_t size = c.size();
+  c.resize(size + n);
+  return get_data(c) + size;
+}
+
+template <typename T>
+inline auto reserve(basic_appender<T> it, size_t n) -> basic_appender<T> {
+  buffer<T>& buf = get_container(it);
+  buf.try_reserve(buf.size() + n);
+  return it;
+}
+
+template <typename Iterator>
+constexpr auto reserve(Iterator& it, size_t) -> Iterator& {
+  return it;
+}
+
+template <typename OutputIt>
+using reserve_iterator =
+    remove_reference_t<decltype(reserve(std::declval<OutputIt&>(), 0))>;
+
+template <typename T, typename OutputIt>
+constexpr auto to_pointer(OutputIt, size_t) -> T* {
+  return nullptr;
+}
+template <typename T> auto to_pointer(basic_appender<T> it, size_t n) -> T* {
+  buffer<T>& buf = get_container(it);
+  auto size = buf.size();
+  buf.try_reserve(size + n);
+  if (buf.capacity() < size + n) return nullptr;
+  buf.try_resize(size + n);
+  return buf.data() + size;
+}
+
+template <typename OutputIt,
+          FMT_ENABLE_IF(is_back_insert_iterator<OutputIt>::value&&
+                            is_contiguous<typename OutputIt::container>::value)>
+inline auto base_iterator(OutputIt it,
+                          typename OutputIt::container_type::value_type*)
+    -> OutputIt {
+  return it;
+}
+
+template <typename Iterator>
+constexpr auto base_iterator(Iterator, Iterator it) -> Iterator {
+  return it;
+}
+
+// <algorithm> is spectacularly slow to compile in C++20 so use a simple fill_n
+// instead (#1998).
+template <typename OutputIt, typename Size, typename T>
+FMT_CONSTEXPR auto fill_n(OutputIt out, Size count, const T& value)
+    -> OutputIt {
+  for (Size i = 0; i < count; ++i) *out++ = value;
+  return out;
+}
+template <typename T, typename Size>
+FMT_CONSTEXPR20 auto fill_n(T* out, Size count, char value) -> T* {
+  if (is_constant_evaluated()) {
+    return fill_n<T*, Size, T>(out, count, value);
+  }
+  std::memset(out, value, to_unsigned(count));
+  return out + count;
+}
+
+template <typename OutChar, typename InputIt, typename OutputIt>
+FMT_CONSTEXPR FMT_NOINLINE auto copy_noinline(InputIt begin, InputIt end,
+                                              OutputIt out) -> OutputIt {
+  return copy<OutChar>(begin, end, out);
+}
+
+// A public domain branchless UTF-8 decoder by Christopher Wellons:
+// https://github.com/skeeto/branchless-utf8
+/* Decode the next character, c, from s, reporting errors in e.
+ *
+ * Since this is a branchless decoder, four bytes will be read from the
+ * buffer regardless of the actual length of the next character. This
+ * means the buffer _must_ have at least three bytes of zero padding
+ * following the end of the data stream.
+ *
+ * Errors are reported in e, which will be non-zero if the parsed
+ * character was somehow invalid: invalid byte sequence, non-canonical
+ * encoding, or a surrogate half.
+ *
+ * The function returns a pointer to the next character. When an error
+ * occurs, this pointer will be a guess that depends on the particular
+ * error, but it will always advance at least one byte.
+ */
+FMT_CONSTEXPR inline auto utf8_decode(const char* s, uint32_t* c, int* e)
+    -> const char* {
+  constexpr const int masks[] = {0x00, 0x7f, 0x1f, 0x0f, 0x07};
+  constexpr const uint32_t mins[] = {4194304, 0, 128, 2048, 65536};
+  constexpr const int shiftc[] = {0, 18, 12, 6, 0};
+  constexpr const int shifte[] = {0, 6, 4, 2, 0};
+
+  int len = "\1\1\1\1\1\1\1\1\1\1\1\1\1\1\1\1\0\0\0\0\0\0\0\0\2\2\2\2\3\3\4"
+      [static_cast<unsigned char>(*s) >> 3];
+  // Compute the pointer to the next character early so that the next
+  // iteration can start working on the next character. Neither Clang
+  // nor GCC figure out this reordering on their own.
+  const char* next = s + len + !len;
+
+  using uchar = unsigned char;
+
+  // Assume a four-byte character and load four bytes. Unused bits are
+  // shifted out.
+  *c = uint32_t(uchar(s[0]) & masks[len]) << 18;
+  *c |= uint32_t(uchar(s[1]) & 0x3f) << 12;
+  *c |= uint32_t(uchar(s[2]) & 0x3f) << 6;
+  *c |= uint32_t(uchar(s[3]) & 0x3f) << 0;
+  *c >>= shiftc[len];
+
+  // Accumulate the various error conditions.
+  *e = (*c < mins[len]) << 6;       // non-canonical encoding
+  *e |= ((*c >> 11) == 0x1b) << 7;  // surrogate half?
+  *e |= (*c > 0x10FFFF) << 8;       // out of range?
+  *e |= (uchar(s[1]) & 0xc0) >> 2;
+  *e |= (uchar(s[2]) & 0xc0) >> 4;
+  *e |= uchar(s[3]) >> 6;
+  *e ^= 0x2a;  // top two bits of each tail byte correct?
+  *e >>= shifte[len];
+
+  return next;
+}
+
+constexpr FMT_INLINE_VARIABLE uint32_t invalid_code_point = ~uint32_t();
+
+// Invokes f(cp, sv) for every code point cp in s with sv being the string view
+// corresponding to the code point. cp is invalid_code_point on error.
+template <typename F>
+FMT_CONSTEXPR void for_each_codepoint(string_view s, F f) {
+  auto decode = [f](const char* buf_ptr, const char* ptr) {
+    auto cp = uint32_t();
+    auto error = 0;
+    auto end = utf8_decode(buf_ptr, &cp, &error);
+    bool result = f(error ? invalid_code_point : cp,
+                    string_view(ptr, error ? 1 : to_unsigned(end - buf_ptr)));
+    return result ? (error ? buf_ptr + 1 : end) : nullptr;
+  };
+  auto p = s.data();
+  const size_t block_size = 4;  // utf8_decode always reads blocks of 4 chars.
+  if (s.size() >= block_size) {
+    for (auto end = p + s.size() - block_size + 1; p < end;) {
+      p = decode(p, p);
+      if (!p) return;
+    }
+  }
+  if (auto num_chars_left = s.data() + s.size() - p) {
+    char buf[2 * block_size - 1] = {};
+    copy<char>(p, p + num_chars_left, buf);
+    const char* buf_ptr = buf;
+    do {
+      auto end = decode(buf_ptr, p);
+      if (!end) return;
+      p += end - buf_ptr;
+      buf_ptr = end;
+    } while (buf_ptr - buf < num_chars_left);
+  }
+}
+
+template <typename Char>
+inline auto compute_width(basic_string_view<Char> s) -> size_t {
+  return s.size();
+}
+
+// Computes approximate display width of a UTF-8 string.
+FMT_CONSTEXPR inline auto compute_width(string_view s) -> size_t {
+  size_t num_code_points = 0;
+  // It is not a lambda for compatibility with C++14.
+  struct count_code_points {
+    size_t* count;
+    FMT_CONSTEXPR auto operator()(uint32_t cp, string_view) const -> bool {
+      *count += detail::to_unsigned(
+          1 +
+          (cp >= 0x1100 &&
+           (cp <= 0x115f ||  // Hangul Jamo init. consonants
+            cp == 0x2329 ||  // LEFT-POINTING ANGLE BRACKET
+            cp == 0x232a ||  // RIGHT-POINTING ANGLE BRACKET
+            // CJK ... Yi except IDEOGRAPHIC HALF FILL SPACE:
+            (cp >= 0x2e80 && cp <= 0xa4cf && cp != 0x303f) ||
+            (cp >= 0xac00 && cp <= 0xd7a3) ||    // Hangul Syllables
+            (cp >= 0xf900 && cp <= 0xfaff) ||    // CJK Compatibility Ideographs
+            (cp >= 0xfe10 && cp <= 0xfe19) ||    // Vertical Forms
+            (cp >= 0xfe30 && cp <= 0xfe6f) ||    // CJK Compatibility Forms
+            (cp >= 0xff00 && cp <= 0xff60) ||    // Fullwidth Forms
+            (cp >= 0xffe0 && cp <= 0xffe6) ||    // Fullwidth Forms
+            (cp >= 0x20000 && cp <= 0x2fffd) ||  // CJK
+            (cp >= 0x30000 && cp <= 0x3fffd) ||
+            // Miscellaneous Symbols and Pictographs + Emoticons:
+            (cp >= 0x1f300 && cp <= 0x1f64f) ||
+            // Supplemental Symbols and Pictographs:
+            (cp >= 0x1f900 && cp <= 0x1f9ff))));
+      return true;
+    }
+  };
+  // We could avoid branches by using utf8_decode directly.
+  for_each_codepoint(s, count_code_points{&num_code_points});
+  return num_code_points;
+}
+
+template <typename Char>
+inline auto code_point_index(basic_string_view<Char> s, size_t n) -> size_t {
+  size_t size = s.size();
+  return n < size ? n : size;
+}
+
+// Calculates the index of the nth code point in a UTF-8 string.
+inline auto code_point_index(string_view s, size_t n) -> size_t {
+  size_t result = s.size();
+  const char* begin = s.begin();
+  for_each_codepoint(s, [begin, &n, &result](uint32_t, string_view sv) {
+    if (n != 0) {
+      --n;
+      return true;
+    }
+    result = to_unsigned(sv.begin() - begin);
+    return false;
+  });
+  return result;
+}
+
+template <typename T> struct is_integral : std::is_integral<T> {};
+template <> struct is_integral<int128_opt> : std::true_type {};
+template <> struct is_integral<uint128_t> : std::true_type {};
+
+template <typename T>
+using is_signed =
+    std::integral_constant<bool, std::numeric_limits<T>::is_signed ||
+                                     std::is_same<T, int128_opt>::value>;
+
+template <typename T>
+using is_integer =
+    bool_constant<is_integral<T>::value && !std::is_same<T, bool>::value &&
+                  !std::is_same<T, char>::value &&
+                  !std::is_same<T, wchar_t>::value>;
+
+#ifndef FMT_USE_FLOAT
+#  define FMT_USE_FLOAT 1
+#endif
+#ifndef FMT_USE_DOUBLE
+#  define FMT_USE_DOUBLE 1
+#endif
+#ifndef FMT_USE_LONG_DOUBLE
+#  define FMT_USE_LONG_DOUBLE 1
+#endif
+
+#if defined(FMT_USE_FLOAT128)
+// Use the provided definition.
+#elif FMT_CLANG_VERSION && FMT_HAS_INCLUDE(<quadmath.h>)
+#  define FMT_USE_FLOAT128 1
+#elif FMT_GCC_VERSION && defined(_GLIBCXX_USE_FLOAT128) && \
+    !defined(__STRICT_ANSI__)
+#  define FMT_USE_FLOAT128 1
+#else
+#  define FMT_USE_FLOAT128 0
+#endif
+#if FMT_USE_FLOAT128
+using float128 = __float128;
+#else
+using float128 = void;
+#endif
+
+template <typename T> using is_float128 = std::is_same<T, float128>;
+
+template <typename T>
+using is_floating_point =
+    bool_constant<std::is_floating_point<T>::value || is_float128<T>::value>;
+
+template <typename T, bool = std::is_floating_point<T>::value>
+struct is_fast_float : bool_constant<std::numeric_limits<T>::is_iec559 &&
+                                     sizeof(T) <= sizeof(double)> {};
+template <typename T> struct is_fast_float<T, false> : std::false_type {};
+
+template <typename T>
+using is_double_double = bool_constant<std::numeric_limits<T>::digits == 106>;
+
+#ifndef FMT_USE_FULL_CACHE_DRAGONBOX
+#  define FMT_USE_FULL_CACHE_DRAGONBOX 0
+#endif
+
+template <typename T, typename Enable = void>
+struct is_locale : std::false_type {};
+template <typename T>
+struct is_locale<T, void_t<decltype(T::classic())>> : std::true_type {};
+}  // namespace detail
+
+FMT_BEGIN_EXPORT
+
+// The number of characters to store in the basic_memory_buffer object itself
+// to avoid dynamic memory allocation.
+enum { inline_buffer_size = 500 };
+
+/**
+ * A dynamically growing memory buffer for trivially copyable/constructible
+ * types with the first `SIZE` elements stored in the object itself. Most
+ * commonly used via the `memory_buffer` alias for `char`.
+ *
+ * **Example**:
+ *
+ *     auto out = fmt::memory_buffer();
+ *     fmt::format_to(std::back_inserter(out), "The answer is {}.", 42);
+ *
+ * This will append "The answer is 42." to `out`. The buffer content can be
+ * converted to `std::string` with `to_string(out)`.
+ */
+template <typename T, size_t SIZE = inline_buffer_size,
+          typename Allocator = std::allocator<T>>
+class basic_memory_buffer : public detail::buffer<T> {
+ private:
+  T store_[SIZE];
+
+  // Don't inherit from Allocator to avoid generating type_info for it.
+  FMT_NO_UNIQUE_ADDRESS Allocator alloc_;
+
+  // Deallocate memory allocated by the buffer.
+  FMT_CONSTEXPR20 void deallocate() {
+    T* data = this->data();
+    if (data != store_) alloc_.deallocate(data, this->capacity());
+  }
+
+  static FMT_CONSTEXPR20 void grow(detail::buffer<T>& buf, size_t size) {
+    detail::abort_fuzzing_if(size > 5000);
+    auto& self = static_cast<basic_memory_buffer&>(buf);
+    const size_t max_size =
+        std::allocator_traits<Allocator>::max_size(self.alloc_);
+    size_t old_capacity = buf.capacity();
+    size_t new_capacity = old_capacity + old_capacity / 2;
+    if (size > new_capacity)
+      new_capacity = size;
+    else if (new_capacity > max_size)
+      new_capacity = size > max_size ? size : max_size;
+    T* old_data = buf.data();
+    T* new_data = self.alloc_.allocate(new_capacity);
+    // Suppress a bogus -Wstringop-overflow in gcc 13.1 (#3481).
+    detail::assume(buf.size() <= new_capacity);
+    // The following code doesn't throw, so the raw pointer above doesn't leak.
+    memcpy(new_data, old_data, buf.size() * sizeof(T));
+    self.set(new_data, new_capacity);
+    // deallocate must not throw according to the standard, but even if it does,
+    // the buffer already uses the new storage and will deallocate it in
+    // destructor.
+    if (old_data != self.store_) self.alloc_.deallocate(old_data, old_capacity);
+  }
+
+ public:
+  using value_type = T;
+  using const_reference = const T&;
+
+  FMT_CONSTEXPR20 explicit basic_memory_buffer(
+      const Allocator& alloc = Allocator())
+      : detail::buffer<T>(grow), alloc_(alloc) {
+    this->set(store_, SIZE);
+    if (detail::is_constant_evaluated()) detail::fill_n(store_, SIZE, T());
+  }
+  FMT_CONSTEXPR20 ~basic_memory_buffer() { deallocate(); }
+
+ private:
+  // Move data from other to this buffer.
+  FMT_CONSTEXPR20 void move(basic_memory_buffer& other) {
+    alloc_ = std::move(other.alloc_);
+    T* data = other.data();
+    size_t size = other.size(), capacity = other.capacity();
+    if (data == other.store_) {
+      this->set(store_, capacity);
+      detail::copy<T>(other.store_, other.store_ + size, store_);
+    } else {
+      this->set(data, capacity);
+      // Set pointer to the inline array so that delete is not called
+      // when deallocating.
+      other.set(other.store_, 0);
+      other.clear();
+    }
+    this->resize(size);
+  }
+
+ public:
+  /// Constructs a `basic_memory_buffer` object moving the content of the other
+  /// object to it.
+  FMT_CONSTEXPR20 basic_memory_buffer(basic_memory_buffer&& other) noexcept
+      : detail::buffer<T>(grow) {
+    move(other);
+  }
+
+  /// Moves the content of the other `basic_memory_buffer` object to this one.
+  auto operator=(basic_memory_buffer&& other) noexcept -> basic_memory_buffer& {
+    FMT_ASSERT(this != &other, "");
+    deallocate();
+    move(other);
+    return *this;
+  }
+
+  // Returns a copy of the allocator associated with this buffer.
+  auto get_allocator() const -> Allocator { return alloc_; }
+
+  /// Resizes the buffer to contain `count` elements. If T is a POD type new
+  /// elements may not be initialized.
+  FMT_CONSTEXPR20 void resize(size_t count) { this->try_resize(count); }
+
+  /// Increases the buffer capacity to `new_capacity`.
+  void reserve(size_t new_capacity) { this->try_reserve(new_capacity); }
+
+  using detail::buffer<T>::append;
+  template <typename ContiguousRange>
+  void append(const ContiguousRange& range) {
+    append(range.data(), range.data() + range.size());
+  }
+};
+
+using memory_buffer = basic_memory_buffer<char>;
+
+template <typename T, size_t SIZE, typename Allocator>
+struct is_contiguous<basic_memory_buffer<T, SIZE, Allocator>> : std::true_type {
+};
+
+FMT_END_EXPORT
+namespace detail {
+FMT_API auto write_console(int fd, string_view text) -> bool;
+FMT_API void print(std::FILE*, string_view);
+}  // namespace detail
+
+FMT_BEGIN_EXPORT
+
+// Suppress a misleading warning in older versions of clang.
+#if FMT_CLANG_VERSION
+#  pragma clang diagnostic ignored "-Wweak-vtables"
+#endif
+
+/// An error reported from a formatting function.
+class FMT_SO_VISIBILITY("default") format_error : public std::runtime_error {
+ public:
+  using std::runtime_error::runtime_error;
+};
+
+namespace detail_exported {
+#if FMT_USE_NONTYPE_TEMPLATE_ARGS
+template <typename Char, size_t N> struct fixed_string {
+  constexpr fixed_string(const Char (&str)[N]) {
+    detail::copy<Char, const Char*, Char*>(static_cast<const Char*>(str),
+                                           str + N, data);
+  }
+  Char data[N] = {};
+};
+#endif
+
+// Converts a compile-time string to basic_string_view.
+template <typename Char, size_t N>
+constexpr auto compile_string_to_view(const Char (&s)[N])
+    -> basic_string_view<Char> {
+  // Remove trailing NUL character if needed. Won't be present if this is used
+  // with a raw character array (i.e. not defined as a string).
+  return {s, N - (std::char_traits<Char>::to_int_type(s[N - 1]) == 0 ? 1 : 0)};
+}
+template <typename Char>
+constexpr auto compile_string_to_view(basic_string_view<Char> s)
+    -> basic_string_view<Char> {
+  return s;
+}
+}  // namespace detail_exported
+
+// A generic formatting context with custom output iterator and character
+// (code unit) support. Char is the format string code unit type which can be
+// different from OutputIt::value_type.
+template <typename OutputIt, typename Char> class generic_context {
+ private:
+  OutputIt out_;
+  basic_format_args<generic_context> args_;
+  detail::locale_ref loc_;
+
+ public:
+  using char_type = Char;
+  using iterator = OutputIt;
+  using parse_context_type = basic_format_parse_context<Char>;
+  template <typename T> using formatter_type = formatter<T, Char>;
+
+  constexpr generic_context(OutputIt out,
+                            basic_format_args<generic_context> ctx_args,
+                            detail::locale_ref loc = {})
+      : out_(out), args_(ctx_args), loc_(loc) {}
+  generic_context(generic_context&&) = default;
+  generic_context(const generic_context&) = delete;
+  void operator=(const generic_context&) = delete;
+
+  constexpr auto arg(int id) const -> basic_format_arg<generic_context> {
+    return args_.get(id);
+  }
+  auto arg(basic_string_view<Char> name) -> basic_format_arg<generic_context> {
+    return args_.get(name);
+  }
+  FMT_CONSTEXPR auto arg_id(basic_string_view<Char> name) -> int {
+    return args_.get_id(name);
+  }
+  auto args() const -> const basic_format_args<generic_context>& {
+    return args_;
+  }
+
+  FMT_CONSTEXPR auto out() -> iterator { return out_; }
+
+  void advance_to(iterator it) {
+    if (!detail::is_back_insert_iterator<iterator>()) out_ = it;
+  }
+
+  FMT_CONSTEXPR auto locale() -> detail::locale_ref { return loc_; }
+};
+
+class loc_value {
+ private:
+  basic_format_arg<format_context> value_;
+
+ public:
+  template <typename T, FMT_ENABLE_IF(!detail::is_float128<T>::value)>
+  loc_value(T value) : value_(detail::make_arg<format_context>(value)) {}
+
+  template <typename T, FMT_ENABLE_IF(detail::is_float128<T>::value)>
+  loc_value(T) {}
+
+  template <typename Visitor> auto visit(Visitor&& vis) -> decltype(vis(0)) {
+    return value_.visit(vis);
+  }
+};
+
+// A locale facet that formats values in UTF-8.
+// It is parameterized on the locale to avoid the heavy <locale> include.
+template <typename Locale> class format_facet : public Locale::facet {
+ private:
+  std::string separator_;
+  std::string grouping_;
+  std::string decimal_point_;
+
+ protected:
+  virtual auto do_put(appender out, loc_value val,
+                      const format_specs& specs) const -> bool;
+
+ public:
+  static FMT_API typename Locale::id id;
+
+  explicit format_facet(Locale& loc);
+  explicit format_facet(string_view sep = "",
+                        std::initializer_list<unsigned char> g = {3},
+                        std::string decimal_point = ".")
+      : separator_(sep.data(), sep.size()),
+        grouping_(g.begin(), g.end()),
+        decimal_point_(decimal_point) {}
+
+  auto put(appender out, loc_value val, const format_specs& specs) const
+      -> bool {
+    return do_put(out, val, specs);
+  }
+};
+
+FMT_END_EXPORT
+
+namespace detail {
+
+// Returns true if value is negative, false otherwise.
+// Same as `value < 0` but doesn't produce warnings if T is an unsigned type.
+template <typename T, FMT_ENABLE_IF(is_signed<T>::value)>
+constexpr auto is_negative(T value) -> bool {
+  return value < 0;
+}
+template <typename T, FMT_ENABLE_IF(!is_signed<T>::value)>
+constexpr auto is_negative(T) -> bool {
+  return false;
+}
+
+template <typename T>
+FMT_CONSTEXPR auto is_supported_floating_point(T) -> bool {
+  if (std::is_same<T, float>()) return FMT_USE_FLOAT;
+  if (std::is_same<T, double>()) return FMT_USE_DOUBLE;
+  if (std::is_same<T, long double>()) return FMT_USE_LONG_DOUBLE;
+  return true;
+}
+
+// Smallest of uint32_t, uint64_t, uint128_t that is large enough to
+// represent all values of an integral type T.
+template <typename T>
+using uint32_or_64_or_128_t =
+    conditional_t<num_bits<T>() <= 32 && !FMT_REDUCE_INT_INSTANTIATIONS,
+                  uint32_t,
+                  conditional_t<num_bits<T>() <= 64, uint64_t, uint128_t>>;
+template <typename T>
+using uint64_or_128_t = conditional_t<num_bits<T>() <= 64, uint64_t, uint128_t>;
+
+#define FMT_POWERS_OF_10(factor)                                  \
+  factor * 10, (factor) * 100, (factor) * 1000, (factor) * 10000, \
+      (factor) * 100000, (factor) * 1000000, (factor) * 10000000, \
+      (factor) * 100000000, (factor) * 1000000000
+
+// Converts value in the range [0, 100) to a string.
+constexpr auto digits2(size_t value) -> const char* {
+  // GCC generates slightly better code when value is pointer-size.
+  return &"0001020304050607080910111213141516171819"
+         "2021222324252627282930313233343536373839"
+         "4041424344454647484950515253545556575859"
+         "6061626364656667686970717273747576777879"
+         "8081828384858687888990919293949596979899"[value * 2];
+}
+
+// Sign is a template parameter to workaround a bug in gcc 4.8.
+template <typename Char, typename Sign> constexpr auto sign(Sign s) -> Char {
+#if !FMT_GCC_VERSION || FMT_GCC_VERSION >= 604
+  static_assert(std::is_same<Sign, sign_t>::value, "");
+#endif
+  return static_cast<char>(((' ' << 24) | ('+' << 16) | ('-' << 8)) >> (s * 8));
+}
+
+template <typename T> FMT_CONSTEXPR auto count_digits_fallback(T n) -> int {
+  int count = 1;
+  for (;;) {
+    // Integer division is slow so do it for a group of four digits instead
+    // of for every digit. The idea comes from the talk by Alexandrescu
+    // "Three Optimization Tips for C++". See speed-test for a comparison.
+    if (n < 10) return count;
+    if (n < 100) return count + 1;
+    if (n < 1000) return count + 2;
+    if (n < 10000) return count + 3;
+    n /= 10000u;
+    count += 4;
+  }
+}
+#if FMT_USE_INT128
+FMT_CONSTEXPR inline auto count_digits(uint128_opt n) -> int {
+  return count_digits_fallback(n);
+}
+#endif
+
+#ifdef FMT_BUILTIN_CLZLL
+// It is a separate function rather than a part of count_digits to workaround
+// the lack of static constexpr in constexpr functions.
+inline auto do_count_digits(uint64_t n) -> int {
+  // This has comparable performance to the version by Kendall Willets
+  // (https://github.com/fmtlib/format-benchmark/blob/master/digits10)
+  // but uses smaller tables.
+  // Maps bsr(n) to ceil(log10(pow(2, bsr(n) + 1) - 1)).
+  static constexpr uint8_t bsr2log10[] = {
+      1,  1,  1,  2,  2,  2,  3,  3,  3,  4,  4,  4,  4,  5,  5,  5,
+      6,  6,  6,  7,  7,  7,  7,  8,  8,  8,  9,  9,  9,  10, 10, 10,
+      10, 11, 11, 11, 12, 12, 12, 13, 13, 13, 13, 14, 14, 14, 15, 15,
+      15, 16, 16, 16, 16, 17, 17, 17, 18, 18, 18, 19, 19, 19, 19, 20};
+  auto t = bsr2log10[FMT_BUILTIN_CLZLL(n | 1) ^ 63];
+  static constexpr const uint64_t zero_or_powers_of_10[] = {
+      0, 0, FMT_POWERS_OF_10(1U), FMT_POWERS_OF_10(1000000000ULL),
+      10000000000000000000ULL};
+  return t - (n < zero_or_powers_of_10[t]);
+}
+#endif
+
+// Returns the number of decimal digits in n. Leading zeros are not counted
+// except for n == 0 in which case count_digits returns 1.
+FMT_CONSTEXPR20 inline auto count_digits(uint64_t n) -> int {
+#ifdef FMT_BUILTIN_CLZLL
+  if (!is_constant_evaluated()) return do_count_digits(n);
+#endif
+  return count_digits_fallback(n);
+}
+
+// Counts the number of digits in n. BITS = log2(radix).
+template <int BITS, typename UInt>
+FMT_CONSTEXPR auto count_digits(UInt n) -> int {
+#ifdef FMT_BUILTIN_CLZ
+  if (!is_constant_evaluated() && num_bits<UInt>() == 32)
+    return (FMT_BUILTIN_CLZ(static_cast<uint32_t>(n) | 1) ^ 31) / BITS + 1;
+#endif
+  // Lambda avoids unreachable code warnings from NVHPC.
+  return [](UInt m) {
+    int num_digits = 0;
+    do {
+      ++num_digits;
+    } while ((m >>= BITS) != 0);
+    return num_digits;
+  }(n);
+}
+
+#ifdef FMT_BUILTIN_CLZ
+// It is a separate function rather than a part of count_digits to workaround
+// the lack of static constexpr in constexpr functions.
+FMT_INLINE auto do_count_digits(uint32_t n) -> int {
+// An optimization by Kendall Willets from https://bit.ly/3uOIQrB.
+// This increments the upper 32 bits (log10(T) - 1) when >= T is added.
+#  define FMT_INC(T) (((sizeof(#T) - 1ull) << 32) - T)
+  static constexpr uint64_t table[] = {
+      FMT_INC(0),          FMT_INC(0),          FMT_INC(0),           // 8
+      FMT_INC(10),         FMT_INC(10),         FMT_INC(10),          // 64
+      FMT_INC(100),        FMT_INC(100),        FMT_INC(100),         // 512
+      FMT_INC(1000),       FMT_INC(1000),       FMT_INC(1000),        // 4096
+      FMT_INC(10000),      FMT_INC(10000),      FMT_INC(10000),       // 32k
+      FMT_INC(100000),     FMT_INC(100000),     FMT_INC(100000),      // 256k
+      FMT_INC(1000000),    FMT_INC(1000000),    FMT_INC(1000000),     // 2048k
+      FMT_INC(10000000),   FMT_INC(10000000),   FMT_INC(10000000),    // 16M
+      FMT_INC(100000000),  FMT_INC(100000000),  FMT_INC(100000000),   // 128M
+      FMT_INC(1000000000), FMT_INC(1000000000), FMT_INC(1000000000),  // 1024M
+      FMT_INC(1000000000), FMT_INC(1000000000)                        // 4B
+  };
+  auto inc = table[FMT_BUILTIN_CLZ(n | 1) ^ 31];
+  return static_cast<int>((n + inc) >> 32);
+}
+#endif
+
+// Optional version of count_digits for better performance on 32-bit platforms.
+FMT_CONSTEXPR20 inline auto count_digits(uint32_t n) -> int {
+#ifdef FMT_BUILTIN_CLZ
+  if (!is_constant_evaluated()) {
+    return do_count_digits(n);
+  }
+#endif
+  return count_digits_fallback(n);
+}
+
+template <typename Int> constexpr auto digits10() noexcept -> int {
+  return std::numeric_limits<Int>::digits10;
+}
+template <> constexpr auto digits10<int128_opt>() noexcept -> int { return 38; }
+template <> constexpr auto digits10<uint128_t>() noexcept -> int { return 38; }
+
+template <typename Char> struct thousands_sep_result {
+  std::string grouping;
+  Char thousands_sep;
+};
+
+template <typename Char>
+FMT_API auto thousands_sep_impl(locale_ref loc) -> thousands_sep_result<Char>;
+template <typename Char>
+inline auto thousands_sep(locale_ref loc) -> thousands_sep_result<Char> {
+  auto result = thousands_sep_impl<char>(loc);
+  return {result.grouping, Char(result.thousands_sep)};
+}
+template <>
+inline auto thousands_sep(locale_ref loc) -> thousands_sep_result<wchar_t> {
+  return thousands_sep_impl<wchar_t>(loc);
+}
+
+template <typename Char>
+FMT_API auto decimal_point_impl(locale_ref loc) -> Char;
+template <typename Char> inline auto decimal_point(locale_ref loc) -> Char {
+  return Char(decimal_point_impl<char>(loc));
+}
+template <> inline auto decimal_point(locale_ref loc) -> wchar_t {
+  return decimal_point_impl<wchar_t>(loc);
+}
+
+// Compares two characters for equality.
+template <typename Char> auto equal2(const Char* lhs, const char* rhs) -> bool {
+  return lhs[0] == Char(rhs[0]) && lhs[1] == Char(rhs[1]);
+}
+inline auto equal2(const char* lhs, const char* rhs) -> bool {
+  return memcmp(lhs, rhs, 2) == 0;
+}
+
+// Copies two characters from src to dst.
+template <typename Char>
+FMT_CONSTEXPR20 FMT_INLINE void copy2(Char* dst, const char* src) {
+  if (!is_constant_evaluated() && sizeof(Char) == sizeof(char)) {
+    memcpy(dst, src, 2);
+    return;
+  }
+  *dst++ = static_cast<Char>(*src++);
+  *dst = static_cast<Char>(*src);
+}
+
+template <typename Iterator> struct format_decimal_result {
+  Iterator begin;
+  Iterator end;
+};
+
+// Formats a decimal unsigned integer value writing into out pointing to a
+// buffer of specified size. The caller must ensure that the buffer is large
+// enough.
+template <typename Char, typename UInt>
+FMT_CONSTEXPR20 auto format_decimal(Char* out, UInt value, int size)
+    -> format_decimal_result<Char*> {
+  FMT_ASSERT(size >= count_digits(value), "invalid digit count");
+  out += size;
+  Char* end = out;
+  while (value >= 100) {
+    // Integer division is slow so do it for a group of two digits instead
+    // of for every digit. The idea comes from the talk by Alexandrescu
+    // "Three Optimization Tips for C++". See speed-test for a comparison.
+    out -= 2;
+    copy2(out, digits2(static_cast<size_t>(value % 100)));
+    value /= 100;
+  }
+  if (value < 10) {
+    *--out = static_cast<Char>('0' + value);
+    return {out, end};
+  }
+  out -= 2;
+  copy2(out, digits2(static_cast<size_t>(value)));
+  return {out, end};
+}
+
+template <typename Char, typename UInt, typename Iterator,
+          FMT_ENABLE_IF(!std::is_pointer<remove_cvref_t<Iterator>>::value)>
+FMT_CONSTEXPR inline auto format_decimal(Iterator out, UInt value, int size)
+    -> format_decimal_result<Iterator> {
+  // Buffer is large enough to hold all digits (digits10 + 1).
+  Char buffer[digits10<UInt>() + 1] = {};
+  auto end = format_decimal(buffer, value, size).end;
+  return {out, detail::copy_noinline<Char>(buffer, end, out)};
+}
+
+template <unsigned BASE_BITS, typename Char, typename UInt>
+FMT_CONSTEXPR auto format_uint(Char* buffer, UInt value, int num_digits,
+                               bool upper = false) -> Char* {
+  buffer += num_digits;
+  Char* end = buffer;
+  do {
+    const char* digits = upper ? "0123456789ABCDEF" : "0123456789abcdef";
+    unsigned digit = static_cast<unsigned>(value & ((1 << BASE_BITS) - 1));
+    *--buffer = static_cast<Char>(BASE_BITS < 4 ? static_cast<char>('0' + digit)
+                                                : digits[digit]);
+  } while ((value >>= BASE_BITS) != 0);
+  return end;
+}
+
+template <unsigned BASE_BITS, typename Char, typename It, typename UInt>
+FMT_CONSTEXPR inline auto format_uint(It out, UInt value, int num_digits,
+                                      bool upper = false) -> It {
+  if (auto ptr = to_pointer<Char>(out, to_unsigned(num_digits))) {
+    format_uint<BASE_BITS>(ptr, value, num_digits, upper);
+    return out;
+  }
+  // Buffer should be large enough to hold all digits (digits / BASE_BITS + 1).
+  char buffer[num_bits<UInt>() / BASE_BITS + 1] = {};
+  format_uint<BASE_BITS>(buffer, value, num_digits, upper);
+  return detail::copy_noinline<Char>(buffer, buffer + num_digits, out);
+}
+
+// A converter from UTF-8 to UTF-16.
+class utf8_to_utf16 {
+ private:
+  basic_memory_buffer<wchar_t> buffer_;
+
+ public:
+  FMT_API explicit utf8_to_utf16(string_view s);
+  operator basic_string_view<wchar_t>() const { return {&buffer_[0], size()}; }
+  auto size() const -> size_t { return buffer_.size() - 1; }
+  auto c_str() const -> const wchar_t* { return &buffer_[0]; }
+  auto str() const -> std::wstring { return {&buffer_[0], size()}; }
+};
+
+enum class to_utf8_error_policy { abort, replace };
+
+// A converter from UTF-16/UTF-32 (host endian) to UTF-8.
+template <typename WChar, typename Buffer = memory_buffer> class to_utf8 {
+ private:
+  Buffer buffer_;
+
+ public:
+  to_utf8() {}
+  explicit to_utf8(basic_string_view<WChar> s,
+                   to_utf8_error_policy policy = to_utf8_error_policy::abort) {
+    static_assert(sizeof(WChar) == 2 || sizeof(WChar) == 4,
+                  "Expect utf16 or utf32");
+    if (!convert(s, policy))
+      FMT_THROW(std::runtime_error(sizeof(WChar) == 2 ? "invalid utf16"
+                                                      : "invalid utf32"));
+  }
+  operator string_view() const { return string_view(&buffer_[0], size()); }
+  auto size() const -> size_t { return buffer_.size() - 1; }
+  auto c_str() const -> const char* { return &buffer_[0]; }
+  auto str() const -> std::string { return std::string(&buffer_[0], size()); }
+
+  // Performs conversion returning a bool instead of throwing exception on
+  // conversion error. This method may still throw in case of memory allocation
+  // error.
+  auto convert(basic_string_view<WChar> s,
+               to_utf8_error_policy policy = to_utf8_error_policy::abort)
+      -> bool {
+    if (!convert(buffer_, s, policy)) return false;
+    buffer_.push_back(0);
+    return true;
+  }
+  static auto convert(Buffer& buf, basic_string_view<WChar> s,
+                      to_utf8_error_policy policy = to_utf8_error_policy::abort)
+      -> bool {
+    for (auto p = s.begin(); p != s.end(); ++p) {
+      uint32_t c = static_cast<uint32_t>(*p);
+      if (sizeof(WChar) == 2 && c >= 0xd800 && c <= 0xdfff) {
+        // Handle a surrogate pair.
+        ++p;
+        if (p == s.end() || (c & 0xfc00) != 0xd800 || (*p & 0xfc00) != 0xdc00) {
+          if (policy == to_utf8_error_policy::abort) return false;
+          buf.append(string_view("\xEF\xBF\xBD"));
+          --p;
+        } else {
+          c = (c << 10) + static_cast<uint32_t>(*p) - 0x35fdc00;
+        }
+      } else if (c < 0x80) {
+        buf.push_back(static_cast<char>(c));
+      } else if (c < 0x800) {
+        buf.push_back(static_cast<char>(0xc0 | (c >> 6)));
+        buf.push_back(static_cast<char>(0x80 | (c & 0x3f)));
+      } else if ((c >= 0x800 && c <= 0xd7ff) || (c >= 0xe000 && c <= 0xffff)) {
+        buf.push_back(static_cast<char>(0xe0 | (c >> 12)));
+        buf.push_back(static_cast<char>(0x80 | ((c & 0xfff) >> 6)));
+        buf.push_back(static_cast<char>(0x80 | (c & 0x3f)));
+      } else if (c >= 0x10000 && c <= 0x10ffff) {
+        buf.push_back(static_cast<char>(0xf0 | (c >> 18)));
+        buf.push_back(static_cast<char>(0x80 | ((c & 0x3ffff) >> 12)));
+        buf.push_back(static_cast<char>(0x80 | ((c & 0xfff) >> 6)));
+        buf.push_back(static_cast<char>(0x80 | (c & 0x3f)));
+      } else {
+        return false;
+      }
+    }
+    return true;
+  }
+};
+
+// Computes 128-bit result of multiplication of two 64-bit unsigned integers.
+inline auto umul128(uint64_t x, uint64_t y) noexcept -> uint128_fallback {
+#if FMT_USE_INT128
+  auto p = static_cast<uint128_opt>(x) * static_cast<uint128_opt>(y);
+  return {static_cast<uint64_t>(p >> 64), static_cast<uint64_t>(p)};
+#elif defined(_MSC_VER) && defined(_M_X64)
+  auto hi = uint64_t();
+  auto lo = _umul128(x, y, &hi);
+  return {hi, lo};
+#else
+  const uint64_t mask = static_cast<uint64_t>(max_value<uint32_t>());
+
+  uint64_t a = x >> 32;
+  uint64_t b = x & mask;
+  uint64_t c = y >> 32;
+  uint64_t d = y & mask;
+
+  uint64_t ac = a * c;
+  uint64_t bc = b * c;
+  uint64_t ad = a * d;
+  uint64_t bd = b * d;
+
+  uint64_t intermediate = (bd >> 32) + (ad & mask) + (bc & mask);
+
+  return {ac + (intermediate >> 32) + (ad >> 32) + (bc >> 32),
+          (intermediate << 32) + (bd & mask)};
+#endif
+}
+
+namespace dragonbox {
+// Computes floor(log10(pow(2, e))) for e in [-2620, 2620] using the method from
+// https://fmt.dev/papers/Dragonbox.pdf#page=28, section 6.1.
+inline auto floor_log10_pow2(int e) noexcept -> int {
+  FMT_ASSERT(e <= 2620 && e >= -2620, "too large exponent");
+  static_assert((-1 >> 1) == -1, "right shift is not arithmetic");
+  return (e * 315653) >> 20;
+}
+
+inline auto floor_log2_pow10(int e) noexcept -> int {
+  FMT_ASSERT(e <= 1233 && e >= -1233, "too large exponent");
+  return (e * 1741647) >> 19;
+}
+
+// Computes upper 64 bits of multiplication of two 64-bit unsigned integers.
+inline auto umul128_upper64(uint64_t x, uint64_t y) noexcept -> uint64_t {
+#if FMT_USE_INT128
+  auto p = static_cast<uint128_opt>(x) * static_cast<uint128_opt>(y);
+  return static_cast<uint64_t>(p >> 64);
+#elif defined(_MSC_VER) && defined(_M_X64)
+  return __umulh(x, y);
+#else
+  return umul128(x, y).high();
+#endif
+}
+
+// Computes upper 128 bits of multiplication of a 64-bit unsigned integer and a
+// 128-bit unsigned integer.
+inline auto umul192_upper128(uint64_t x, uint128_fallback y) noexcept
+    -> uint128_fallback {
+  uint128_fallback r = umul128(x, y.high());
+  r += umul128_upper64(x, y.low());
+  return r;
+}
+
+FMT_API auto get_cached_power(int k) noexcept -> uint128_fallback;
+
+// Type-specific information that Dragonbox uses.
+template <typename T, typename Enable = void> struct float_info;
+
+template <> struct float_info<float> {
+  using carrier_uint = uint32_t;
+  static const int exponent_bits = 8;
+  static const int kappa = 1;
+  static const int big_divisor = 100;
+  static const int small_divisor = 10;
+  static const int min_k = -31;
+  static const int max_k = 46;
+  static const int shorter_interval_tie_lower_threshold = -35;
+  static const int shorter_interval_tie_upper_threshold = -35;
+};
+
+template <> struct float_info<double> {
+  using carrier_uint = uint64_t;
+  static const int exponent_bits = 11;
+  static const int kappa = 2;
+  static const int big_divisor = 1000;
+  static const int small_divisor = 100;
+  static const int min_k = -292;
+  static const int max_k = 341;
+  static const int shorter_interval_tie_lower_threshold = -77;
+  static const int shorter_interval_tie_upper_threshold = -77;
+};
+
+// An 80- or 128-bit floating point number.
+template <typename T>
+struct float_info<T, enable_if_t<std::numeric_limits<T>::digits == 64 ||
+                                 std::numeric_limits<T>::digits == 113 ||
+                                 is_float128<T>::value>> {
+  using carrier_uint = detail::uint128_t;
+  static const int exponent_bits = 15;
+};
+
+// A double-double floating point number.
+template <typename T>
+struct float_info<T, enable_if_t<is_double_double<T>::value>> {
+  using carrier_uint = detail::uint128_t;
+};
+
+template <typename T> struct decimal_fp {
+  using significand_type = typename float_info<T>::carrier_uint;
+  significand_type significand;
+  int exponent;
+};
+
+template <typename T> FMT_API auto to_decimal(T x) noexcept -> decimal_fp<T>;
+}  // namespace dragonbox
+
+// Returns true iff Float has the implicit bit which is not stored.
+template <typename Float> constexpr auto has_implicit_bit() -> bool {
+  // An 80-bit FP number has a 64-bit significand an no implicit bit.
+  return std::numeric_limits<Float>::digits != 64;
+}
+
+// Returns the number of significand bits stored in Float. The implicit bit is
+// not counted since it is not stored.
+template <typename Float> constexpr auto num_significand_bits() -> int {
+  // std::numeric_limits may not support __float128.
+  return is_float128<Float>() ? 112
+                              : (std::numeric_limits<Float>::digits -
+                                 (has_implicit_bit<Float>() ? 1 : 0));
+}
+
+template <typename Float>
+constexpr auto exponent_mask() ->
+    typename dragonbox::float_info<Float>::carrier_uint {
+  using float_uint = typename dragonbox::float_info<Float>::carrier_uint;
+  return ((float_uint(1) << dragonbox::float_info<Float>::exponent_bits) - 1)
+         << num_significand_bits<Float>();
+}
+template <typename Float> constexpr auto exponent_bias() -> int {
+  // std::numeric_limits may not support __float128.
+  return is_float128<Float>() ? 16383
+                              : std::numeric_limits<Float>::max_exponent - 1;
+}
+
+// Writes the exponent exp in the form "[+-]d{2,3}" to buffer.
+template <typename Char, typename It>
+FMT_CONSTEXPR auto write_exponent(int exp, It it) -> It {
+  FMT_ASSERT(-10000 < exp && exp < 10000, "exponent out of range");
+  if (exp < 0) {
+    *it++ = static_cast<Char>('-');
+    exp = -exp;
+  } else {
+    *it++ = static_cast<Char>('+');
+  }
+  if (exp >= 100) {
+    const char* top = digits2(to_unsigned(exp / 100));
+    if (exp >= 1000) *it++ = static_cast<Char>(top[0]);
+    *it++ = static_cast<Char>(top[1]);
+    exp %= 100;
+  }
+  const char* d = digits2(to_unsigned(exp));
+  *it++ = static_cast<Char>(d[0]);
+  *it++ = static_cast<Char>(d[1]);
+  return it;
+}
+
+// A floating-point number f * pow(2, e) where F is an unsigned type.
+template <typename F> struct basic_fp {
+  F f;
+  int e;
+
+  static constexpr const int num_significand_bits =
+      static_cast<int>(sizeof(F) * num_bits<unsigned char>());
+
+  constexpr basic_fp() : f(0), e(0) {}
+  constexpr basic_fp(uint64_t f_val, int e_val) : f(f_val), e(e_val) {}
+
+  // Constructs fp from an IEEE754 floating-point number.
+  template <typename Float> FMT_CONSTEXPR basic_fp(Float n) { assign(n); }
+
+  // Assigns n to this and return true iff predecessor is closer than successor.
+  template <typename Float, FMT_ENABLE_IF(!is_double_double<Float>::value)>
+  FMT_CONSTEXPR auto assign(Float n) -> bool {
+    static_assert(std::numeric_limits<Float>::digits <= 113, "unsupported FP");
+    // Assume Float is in the format [sign][exponent][significand].
+    using carrier_uint = typename dragonbox::float_info<Float>::carrier_uint;
+    const auto num_float_significand_bits =
+        detail::num_significand_bits<Float>();
+    const auto implicit_bit = carrier_uint(1) << num_float_significand_bits;
+    const auto significand_mask = implicit_bit - 1;
+    auto u = bit_cast<carrier_uint>(n);
+    f = static_cast<F>(u & significand_mask);
+    auto biased_e = static_cast<int>((u & exponent_mask<Float>()) >>
+                                     num_float_significand_bits);
+    // The predecessor is closer if n is a normalized power of 2 (f == 0)
+    // other than the smallest normalized number (biased_e > 1).
+    auto is_predecessor_closer = f == 0 && biased_e > 1;
+    if (biased_e == 0)
+      biased_e = 1;  // Subnormals use biased exponent 1 (min exponent).
+    else if (has_implicit_bit<Float>())
+      f += static_cast<F>(implicit_bit);
+    e = biased_e - exponent_bias<Float>() - num_float_significand_bits;
+    if (!has_implicit_bit<Float>()) ++e;
+    return is_predecessor_closer;
+  }
+
+  template <typename Float, FMT_ENABLE_IF(is_double_double<Float>::value)>
+  FMT_CONSTEXPR auto assign(Float n) -> bool {
+    static_assert(std::numeric_limits<double>::is_iec559, "unsupported FP");
+    return assign(static_cast<double>(n));
+  }
+};
+
+using fp = basic_fp<unsigned long long>;
+
+// Normalizes the value converted from double and multiplied by (1 << SHIFT).
+template <int SHIFT = 0, typename F>
+FMT_CONSTEXPR auto normalize(basic_fp<F> value) -> basic_fp<F> {
+  // Handle subnormals.
+  const auto implicit_bit = F(1) << num_significand_bits<double>();
+  const auto shifted_implicit_bit = implicit_bit << SHIFT;
+  while ((value.f & shifted_implicit_bit) == 0) {
+    value.f <<= 1;
+    --value.e;
+  }
+  // Subtract 1 to account for hidden bit.
+  const auto offset = basic_fp<F>::num_significand_bits -
+                      num_significand_bits<double>() - SHIFT - 1;
+  value.f <<= offset;
+  value.e -= offset;
+  return value;
+}
+
+// Computes lhs * rhs / pow(2, 64) rounded to nearest with half-up tie breaking.
+FMT_CONSTEXPR inline auto multiply(uint64_t lhs, uint64_t rhs) -> uint64_t {
+#if FMT_USE_INT128
+  auto product = static_cast<__uint128_t>(lhs) * rhs;
+  auto f = static_cast<uint64_t>(product >> 64);
+  return (static_cast<uint64_t>(product) & (1ULL << 63)) != 0 ? f + 1 : f;
+#else
+  // Multiply 32-bit parts of significands.
+  uint64_t mask = (1ULL << 32) - 1;
+  uint64_t a = lhs >> 32, b = lhs & mask;
+  uint64_t c = rhs >> 32, d = rhs & mask;
+  uint64_t ac = a * c, bc = b * c, ad = a * d, bd = b * d;
+  // Compute mid 64-bit of result and round.
+  uint64_t mid = (bd >> 32) + (ad & mask) + (bc & mask) + (1U << 31);
+  return ac + (ad >> 32) + (bc >> 32) + (mid >> 32);
+#endif
+}
+
+FMT_CONSTEXPR inline auto operator*(fp x, fp y) -> fp {
+  return {multiply(x.f, y.f), x.e + y.e + 64};
+}
+
+template <typename T, bool doublish = num_bits<T>() == num_bits<double>()>
+using convert_float_result =
+    conditional_t<std::is_same<T, float>::value || doublish, double, T>;
+
+template <typename T>
+constexpr auto convert_float(T value) -> convert_float_result<T> {
+  return static_cast<convert_float_result<T>>(value);
+}
+
+template <typename Char, typename OutputIt>
+FMT_NOINLINE FMT_CONSTEXPR auto fill(OutputIt it, size_t n, const fill_t& fill)
+    -> OutputIt {
+  auto fill_size = fill.size();
+  if (fill_size == 1) return detail::fill_n(it, n, fill.template get<Char>());
+  if (const Char* data = fill.template data<Char>()) {
+    for (size_t i = 0; i < n; ++i) it = copy<Char>(data, data + fill_size, it);
+  }
+  return it;
+}
+
+// Writes the output of f, padded according to format specifications in specs.
+// size: output size in code units.
+// width: output display width in (terminal) column positions.
+template <typename Char, align::type align = align::left, typename OutputIt,
+          typename F>
+FMT_CONSTEXPR auto write_padded(OutputIt out, const format_specs& specs,
+                                size_t size, size_t width, F&& f) -> OutputIt {
+  static_assert(align == align::left || align == align::right, "");
+  unsigned spec_width = to_unsigned(specs.width);
+  size_t padding = spec_width > width ? spec_width - width : 0;
+  // Shifts are encoded as string literals because static constexpr is not
+  // supported in constexpr functions.
+  auto* shifts = align == align::left ? "\x1f\x1f\x00\x01" : "\x00\x1f\x00\x01";
+  size_t left_padding = padding >> shifts[specs.align];
+  size_t right_padding = padding - left_padding;
+  auto it = reserve(out, size + padding * specs.fill.size());
+  if (left_padding != 0) it = fill<Char>(it, left_padding, specs.fill);
+  it = f(it);
+  if (right_padding != 0) it = fill<Char>(it, right_padding, specs.fill);
+  return base_iterator(out, it);
+}
+
+template <typename Char, align::type align = align::left, typename OutputIt,
+          typename F>
+constexpr auto write_padded(OutputIt out, const format_specs& specs,
+                            size_t size, F&& f) -> OutputIt {
+  return write_padded<Char, align>(out, specs, size, size, f);
+}
+
+template <typename Char, align::type align = align::left, typename OutputIt>
+FMT_CONSTEXPR auto write_bytes(OutputIt out, string_view bytes,
+                               const format_specs& specs = {}) -> OutputIt {
+  return write_padded<Char, align>(
+      out, specs, bytes.size(), [bytes](reserve_iterator<OutputIt> it) {
+        const char* data = bytes.data();
+        return copy<Char>(data, data + bytes.size(), it);
+      });
+}
+
+template <typename Char, typename OutputIt, typename UIntPtr>
+auto write_ptr(OutputIt out, UIntPtr value, const format_specs* specs)
+    -> OutputIt {
+  int num_digits = count_digits<4>(value);
+  auto size = to_unsigned(num_digits) + size_t(2);
+  auto write = [=](reserve_iterator<OutputIt> it) {
+    *it++ = static_cast<Char>('0');
+    *it++ = static_cast<Char>('x');
+    return format_uint<4, Char>(it, value, num_digits);
+  };
+  return specs ? write_padded<Char, align::right>(out, *specs, size, write)
+               : base_iterator(out, write(reserve(out, size)));
+}
+
+// Returns true iff the code point cp is printable.
+FMT_API auto is_printable(uint32_t cp) -> bool;
+
+inline auto needs_escape(uint32_t cp) -> bool {
+  return cp < 0x20 || cp == 0x7f || cp == '"' || cp == '\\' ||
+         !is_printable(cp);
+}
+
+template <typename Char> struct find_escape_result {
+  const Char* begin;
+  const Char* end;
+  uint32_t cp;
+};
+
+template <typename Char>
+auto find_escape(const Char* begin, const Char* end)
+    -> find_escape_result<Char> {
+  for (; begin != end; ++begin) {
+    uint32_t cp = static_cast<unsigned_char<Char>>(*begin);
+    if (const_check(sizeof(Char) == 1) && cp >= 0x80) continue;
+    if (needs_escape(cp)) return {begin, begin + 1, cp};
+  }
+  return {begin, nullptr, 0};
+}
+
+inline auto find_escape(const char* begin, const char* end)
+    -> find_escape_result<char> {
+  if (!use_utf8()) return find_escape<char>(begin, end);
+  auto result = find_escape_result<char>{end, nullptr, 0};
+  for_each_codepoint(string_view(begin, to_unsigned(end - begin)),
+                     [&](uint32_t cp, string_view sv) {
+                       if (needs_escape(cp)) {
+                         result = {sv.begin(), sv.end(), cp};
+                         return false;
+                       }
+                       return true;
+                     });
+  return result;
+}
+
+#define FMT_STRING_IMPL(s, base, explicit)                                    \
+  [] {                                                                        \
+    /* Use the hidden visibility as a workaround for a GCC bug (#1973). */    \
+    /* Use a macro-like name to avoid shadowing warnings. */                  \
+    struct FMT_VISIBILITY("hidden") FMT_COMPILE_STRING : base {               \
+      using char_type FMT_MAYBE_UNUSED = fmt::remove_cvref_t<decltype(s[0])>; \
+      FMT_MAYBE_UNUSED FMT_CONSTEXPR explicit                                 \
+      operator fmt::basic_string_view<char_type>() const {                    \
+        return fmt::detail_exported::compile_string_to_view<char_type>(s);    \
+      }                                                                       \
+    };                                                                        \
+    return FMT_COMPILE_STRING();                                              \
+  }()
+
+/**
+ * Constructs a compile-time format string from a string literal `s`.
+ *
+ * **Example**:
+ *
+ *     // A compile-time error because 'd' is an invalid specifier for strings.
+ *     std::string s = fmt::format(FMT_STRING("{:d}"), "foo");
+ */
+#define FMT_STRING(s) FMT_STRING_IMPL(s, fmt::detail::compile_string, )
+
+template <size_t width, typename Char, typename OutputIt>
+auto write_codepoint(OutputIt out, char prefix, uint32_t cp) -> OutputIt {
+  *out++ = static_cast<Char>('\\');
+  *out++ = static_cast<Char>(prefix);
+  Char buf[width];
+  fill_n(buf, width, static_cast<Char>('0'));
+  format_uint<4>(buf, cp, width);
+  return copy<Char>(buf, buf + width, out);
+}
+
+template <typename OutputIt, typename Char>
+auto write_escaped_cp(OutputIt out, const find_escape_result<Char>& escape)
+    -> OutputIt {
+  auto c = static_cast<Char>(escape.cp);
+  switch (escape.cp) {
+  case '\n':
+    *out++ = static_cast<Char>('\\');
+    c = static_cast<Char>('n');
+    break;
+  case '\r':
+    *out++ = static_cast<Char>('\\');
+    c = static_cast<Char>('r');
+    break;
+  case '\t':
+    *out++ = static_cast<Char>('\\');
+    c = static_cast<Char>('t');
+    break;
+  case '"':
+    FMT_FALLTHROUGH;
+  case '\'':
+    FMT_FALLTHROUGH;
+  case '\\':
+    *out++ = static_cast<Char>('\\');
+    break;
+  default:
+    if (escape.cp < 0x100) return write_codepoint<2, Char>(out, 'x', escape.cp);
+    if (escape.cp < 0x10000)
+      return write_codepoint<4, Char>(out, 'u', escape.cp);
+    if (escape.cp < 0x110000)
+      return write_codepoint<8, Char>(out, 'U', escape.cp);
+    for (Char escape_char : basic_string_view<Char>(
+             escape.begin, to_unsigned(escape.end - escape.begin))) {
+      out = write_codepoint<2, Char>(out, 'x',
+                                     static_cast<uint32_t>(escape_char) & 0xFF);
+    }
+    return out;
+  }
+  *out++ = c;
+  return out;
+}
+
+template <typename Char, typename OutputIt>
+auto write_escaped_string(OutputIt out, basic_string_view<Char> str)
+    -> OutputIt {
+  *out++ = static_cast<Char>('"');
+  auto begin = str.begin(), end = str.end();
+  do {
+    auto escape = find_escape(begin, end);
+    out = copy<Char>(begin, escape.begin, out);
+    begin = escape.end;
+    if (!begin) break;
+    out = write_escaped_cp<OutputIt, Char>(out, escape);
+  } while (begin != end);
+  *out++ = static_cast<Char>('"');
+  return out;
+}
+
+template <typename Char, typename OutputIt>
+auto write_escaped_char(OutputIt out, Char v) -> OutputIt {
+  Char v_array[1] = {v};
+  *out++ = static_cast<Char>('\'');
+  if ((needs_escape(static_cast<uint32_t>(v)) && v != static_cast<Char>('"')) ||
+      v == static_cast<Char>('\'')) {
+    out = write_escaped_cp(out,
+                           find_escape_result<Char>{v_array, v_array + 1,
+                                                    static_cast<uint32_t>(v)});
+  } else {
+    *out++ = v;
+  }
+  *out++ = static_cast<Char>('\'');
+  return out;
+}
+
+template <typename Char, typename OutputIt>
+FMT_CONSTEXPR auto write_char(OutputIt out, Char value,
+                              const format_specs& specs) -> OutputIt {
+  bool is_debug = specs.type == presentation_type::debug;
+  return write_padded<Char>(out, specs, 1, [=](reserve_iterator<OutputIt> it) {
+    if (is_debug) return write_escaped_char(it, value);
+    *it++ = value;
+    return it;
+  });
+}
+template <typename Char, typename OutputIt>
+FMT_CONSTEXPR auto write(OutputIt out, Char value, const format_specs& specs,
+                         locale_ref loc = {}) -> OutputIt {
+  // char is formatted as unsigned char for consistency across platforms.
+  using unsigned_type =
+      conditional_t<std::is_same<Char, char>::value, unsigned char, unsigned>;
+  return check_char_specs(specs)
+             ? write_char<Char>(out, value, specs)
+             : write<Char>(out, static_cast<unsigned_type>(value), specs, loc);
+}
+
+// Data for write_int that doesn't depend on output iterator type. It is used to
+// avoid template code bloat.
+template <typename Char> struct write_int_data {
+  size_t size;
+  size_t padding;
+
+  FMT_CONSTEXPR write_int_data(int num_digits, unsigned prefix,
+                               const format_specs& specs)
+      : size((prefix >> 24) + to_unsigned(num_digits)), padding(0) {
+    if (specs.align == align::numeric) {
+      auto width = to_unsigned(specs.width);
+      if (width > size) {
+        padding = width - size;
+        size = width;
+      }
+    } else if (specs.precision > num_digits) {
+      size = (prefix >> 24) + to_unsigned(specs.precision);
+      padding = to_unsigned(specs.precision - num_digits);
+    }
+  }
+};
+
+// Writes an integer in the format
+//   <left-padding><prefix><numeric-padding><digits><right-padding>
+// where <digits> are written by write_digits(it).
+// prefix contains chars in three lower bytes and the size in the fourth byte.
+template <typename Char, typename OutputIt, typename W>
+FMT_CONSTEXPR FMT_INLINE auto write_int(OutputIt out, int num_digits,
+                                        unsigned prefix,
+                                        const format_specs& specs,
+                                        W write_digits) -> OutputIt {
+  // Slightly faster check for specs.width == 0 && specs.precision == -1.
+  if ((specs.width | (specs.precision + 1)) == 0) {
+    auto it = reserve(out, to_unsigned(num_digits) + (prefix >> 24));
+    if (prefix != 0) {
+      for (unsigned p = prefix & 0xffffff; p != 0; p >>= 8)
+        *it++ = static_cast<Char>(p & 0xff);
+    }
+    return base_iterator(out, write_digits(it));
+  }
+  auto data = write_int_data<Char>(num_digits, prefix, specs);
+  return write_padded<Char, align::right>(
+      out, specs, data.size, [=](reserve_iterator<OutputIt> it) {
+        for (unsigned p = prefix & 0xffffff; p != 0; p >>= 8)
+          *it++ = static_cast<Char>(p & 0xff);
+        it = detail::fill_n(it, data.padding, static_cast<Char>('0'));
+        return write_digits(it);
+      });
+}
+
+template <typename Char> class digit_grouping {
+ private:
+  std::string grouping_;
+  std::basic_string<Char> thousands_sep_;
+
+  struct next_state {
+    std::string::const_iterator group;
+    int pos;
+  };
+  auto initial_state() const -> next_state { return {grouping_.begin(), 0}; }
+
+  // Returns the next digit group separator position.
+  auto next(next_state& state) const -> int {
+    if (thousands_sep_.empty()) return max_value<int>();
+    if (state.group == grouping_.end()) return state.pos += grouping_.back();
+    if (*state.group <= 0 || *state.group == max_value<char>())
+      return max_value<int>();
+    state.pos += *state.group++;
+    return state.pos;
+  }
+
+ public:
+  explicit digit_grouping(locale_ref loc, bool localized = true) {
+    if (!localized) return;
+    auto sep = thousands_sep<Char>(loc);
+    grouping_ = sep.grouping;
+    if (sep.thousands_sep) thousands_sep_.assign(1, sep.thousands_sep);
+  }
+  digit_grouping(std::string grouping, std::basic_string<Char> sep)
+      : grouping_(std::move(grouping)), thousands_sep_(std::move(sep)) {}
+
+  auto has_separator() const -> bool { return !thousands_sep_.empty(); }
+
+  auto count_separators(int num_digits) const -> int {
+    int count = 0;
+    auto state = initial_state();
+    while (num_digits > next(state)) ++count;
+    return count;
+  }
+
+  // Applies grouping to digits and write the output to out.
+  template <typename Out, typename C>
+  auto apply(Out out, basic_string_view<C> digits) const -> Out {
+    auto num_digits = static_cast<int>(digits.size());
+    auto separators = basic_memory_buffer<int>();
+    separators.push_back(0);
+    auto state = initial_state();
+    while (int i = next(state)) {
+      if (i >= num_digits) break;
+      separators.push_back(i);
+    }
+    for (int i = 0, sep_index = static_cast<int>(separators.size() - 1);
+         i < num_digits; ++i) {
+      if (num_digits - i == separators[sep_index]) {
+        out = copy<Char>(thousands_sep_.data(),
+                         thousands_sep_.data() + thousands_sep_.size(), out);
+        --sep_index;
+      }
+      *out++ = static_cast<Char>(digits[to_unsigned(i)]);
+    }
+    return out;
+  }
+};
+
+FMT_CONSTEXPR inline void prefix_append(unsigned& prefix, unsigned value) {
+  prefix |= prefix != 0 ? value << 8 : value;
+  prefix += (1u + (value > 0xff ? 1 : 0)) << 24;
+}
+
+// Writes a decimal integer with digit grouping.
+template <typename OutputIt, typename UInt, typename Char>
+auto write_int(OutputIt out, UInt value, unsigned prefix,
+               const format_specs& specs, const digit_grouping<Char>& grouping)
+    -> OutputIt {
+  static_assert(std::is_same<uint64_or_128_t<UInt>, UInt>::value, "");
+  int num_digits = 0;
+  auto buffer = memory_buffer();
+  switch (specs.type) {
+  default:
+    FMT_ASSERT(false, "");
+    FMT_FALLTHROUGH;
+  case presentation_type::none:
+  case presentation_type::dec:
+    num_digits = count_digits(value);
+    format_decimal<char>(appender(buffer), value, num_digits);
+    break;
+  case presentation_type::hex:
+    if (specs.alt)
+      prefix_append(prefix, unsigned(specs.upper ? 'X' : 'x') << 8 | '0');
+    num_digits = count_digits<4>(value);
+    format_uint<4, char>(appender(buffer), value, num_digits, specs.upper);
+    break;
+  case presentation_type::oct:
+    num_digits = count_digits<3>(value);
+    // Octal prefix '0' is counted as a digit, so only add it if precision
+    // is not greater than the number of digits.
+    if (specs.alt && specs.precision <= num_digits && value != 0)
+      prefix_append(prefix, '0');
+    format_uint<3, char>(appender(buffer), value, num_digits);
+    break;
+  case presentation_type::bin:
+    if (specs.alt)
+      prefix_append(prefix, unsigned(specs.upper ? 'B' : 'b') << 8 | '0');
+    num_digits = count_digits<1>(value);
+    format_uint<1, char>(appender(buffer), value, num_digits);
+    break;
+  case presentation_type::chr:
+    return write_char<Char>(out, static_cast<Char>(value), specs);
+  }
+
+  unsigned size = (prefix != 0 ? prefix >> 24 : 0) + to_unsigned(num_digits) +
+                  to_unsigned(grouping.count_separators(num_digits));
+  return write_padded<Char, align::right>(
+      out, specs, size, size, [&](reserve_iterator<OutputIt> it) {
+        for (unsigned p = prefix & 0xffffff; p != 0; p >>= 8)
+          *it++ = static_cast<Char>(p & 0xff);
+        return grouping.apply(it, string_view(buffer.data(), buffer.size()));
+      });
+}
+
+// Writes a localized value.
+FMT_API auto write_loc(appender out, loc_value value, const format_specs& specs,
+                       locale_ref loc) -> bool;
+template <typename OutputIt>
+inline auto write_loc(OutputIt, loc_value, const format_specs&, locale_ref)
+    -> bool {
+  return false;
+}
+
+template <typename UInt> struct write_int_arg {
+  UInt abs_value;
+  unsigned prefix;
+};
+
+template <typename T>
+FMT_CONSTEXPR auto make_write_int_arg(T value, sign_t sign)
+    -> write_int_arg<uint32_or_64_or_128_t<T>> {
+  auto prefix = 0u;
+  auto abs_value = static_cast<uint32_or_64_or_128_t<T>>(value);
+  if (is_negative(value)) {
+    prefix = 0x01000000 | '-';
+    abs_value = 0 - abs_value;
+  } else {
+    constexpr const unsigned prefixes[4] = {0, 0, 0x1000000u | '+',
+                                            0x1000000u | ' '};
+    prefix = prefixes[sign];
+  }
+  return {abs_value, prefix};
+}
+
+template <typename Char = char> struct loc_writer {
+  basic_appender<Char> out;
+  const format_specs& specs;
+  std::basic_string<Char> sep;
+  std::string grouping;
+  std::basic_string<Char> decimal_point;
+
+  template <typename T, FMT_ENABLE_IF(is_integer<T>::value)>
+  auto operator()(T value) -> bool {
+    auto arg = make_write_int_arg(value, specs.sign);
+    write_int(out, static_cast<uint64_or_128_t<T>>(arg.abs_value), arg.prefix,
+              specs, digit_grouping<Char>(grouping, sep));
+    return true;
+  }
+
+  template <typename T, FMT_ENABLE_IF(!is_integer<T>::value)>
+  auto operator()(T) -> bool {
+    return false;
+  }
+};
+
+template <typename Char, typename OutputIt, typename T>
+FMT_CONSTEXPR FMT_INLINE auto write_int(OutputIt out, write_int_arg<T> arg,
+                                        const format_specs& specs, locale_ref)
+    -> OutputIt {
+  static_assert(std::is_same<T, uint32_or_64_or_128_t<T>>::value, "");
+  auto abs_value = arg.abs_value;
+  auto prefix = arg.prefix;
+  switch (specs.type) {
+  default:
+    FMT_ASSERT(false, "");
+    FMT_FALLTHROUGH;
+  case presentation_type::none:
+  case presentation_type::dec: {
+    int num_digits = count_digits(abs_value);
+    return write_int<Char>(
+        out, num_digits, prefix, specs, [=](reserve_iterator<OutputIt> it) {
+          return format_decimal<Char>(it, abs_value, num_digits).end;
+        });
+  }
+  case presentation_type::hex: {
+    if (specs.alt)
+      prefix_append(prefix, unsigned(specs.upper ? 'X' : 'x') << 8 | '0');
+    int num_digits = count_digits<4>(abs_value);
+    return write_int<Char>(
+        out, num_digits, prefix, specs, [=](reserve_iterator<OutputIt> it) {
+          return format_uint<4, Char>(it, abs_value, num_digits, specs.upper);
+        });
+  }
+  case presentation_type::oct: {
+    int num_digits = count_digits<3>(abs_value);
+    // Octal prefix '0' is counted as a digit, so only add it if precision
+    // is not greater than the number of digits.
+    if (specs.alt && specs.precision <= num_digits && abs_value != 0)
+      prefix_append(prefix, '0');
+    return write_int<Char>(
+        out, num_digits, prefix, specs, [=](reserve_iterator<OutputIt> it) {
+          return format_uint<3, Char>(it, abs_value, num_digits);
+        });
+  }
+  case presentation_type::bin: {
+    if (specs.alt)
+      prefix_append(prefix, unsigned(specs.upper ? 'B' : 'b') << 8 | '0');
+    int num_digits = count_digits<1>(abs_value);
+    return write_int<Char>(
+        out, num_digits, prefix, specs, [=](reserve_iterator<OutputIt> it) {
+          return format_uint<1, Char>(it, abs_value, num_digits);
+        });
+  }
+  case presentation_type::chr:
+    return write_char<Char>(out, static_cast<Char>(abs_value), specs);
+  }
+}
+template <typename Char, typename OutputIt, typename T>
+FMT_CONSTEXPR FMT_NOINLINE auto write_int_noinline(OutputIt out,
+                                                   write_int_arg<T> arg,
+                                                   const format_specs& specs,
+                                                   locale_ref loc) -> OutputIt {
+  return write_int<Char>(out, arg, specs, loc);
+}
+template <typename Char, typename T,
+          FMT_ENABLE_IF(is_integral<T>::value &&
+                        !std::is_same<T, bool>::value &&
+                        !std::is_same<T, Char>::value)>
+FMT_CONSTEXPR FMT_INLINE auto write(basic_appender<Char> out, T value,
+                                    const format_specs& specs, locale_ref loc)
+    -> basic_appender<Char> {
+  if (specs.localized && write_loc(out, value, specs, loc)) return out;
+  return write_int_noinline<Char>(out, make_write_int_arg(value, specs.sign),
+                                  specs, loc);
+}
+// An inlined version of write used in format string compilation.
+template <typename Char, typename OutputIt, typename T,
+          FMT_ENABLE_IF(is_integral<T>::value &&
+                        !std::is_same<T, bool>::value &&
+                        !std::is_same<T, Char>::value &&
+                        !std::is_same<OutputIt, basic_appender<Char>>::value)>
+FMT_CONSTEXPR FMT_INLINE auto write(OutputIt out, T value,
+                                    const format_specs& specs, locale_ref loc)
+    -> OutputIt {
+  if (specs.localized && write_loc(out, value, specs, loc)) return out;
+  return write_int<Char>(out, make_write_int_arg(value, specs.sign), specs,
+                         loc);
+}
+
+// An output iterator that counts the number of objects written to it and
+// discards them.
+class counting_iterator {
+ private:
+  size_t count_;
+
+ public:
+  using iterator_category = std::output_iterator_tag;
+  using difference_type = std::ptrdiff_t;
+  using pointer = void;
+  using reference = void;
+  FMT_UNCHECKED_ITERATOR(counting_iterator);
+
+  struct value_type {
+    template <typename T> FMT_CONSTEXPR void operator=(const T&) {}
+  };
+
+  FMT_CONSTEXPR counting_iterator() : count_(0) {}
+
+  FMT_CONSTEXPR auto count() const -> size_t { return count_; }
+
+  FMT_CONSTEXPR auto operator++() -> counting_iterator& {
+    ++count_;
+    return *this;
+  }
+  FMT_CONSTEXPR auto operator++(int) -> counting_iterator {
+    auto it = *this;
+    ++*this;
+    return it;
+  }
+
+  FMT_CONSTEXPR friend auto operator+(counting_iterator it, difference_type n)
+      -> counting_iterator {
+    it.count_ += static_cast<size_t>(n);
+    return it;
+  }
+
+  FMT_CONSTEXPR auto operator*() const -> value_type { return {}; }
+};
+
+template <typename Char, typename OutputIt>
+FMT_CONSTEXPR auto write(OutputIt out, basic_string_view<Char> s,
+                         const format_specs& specs) -> OutputIt {
+  auto data = s.data();
+  auto size = s.size();
+  if (specs.precision >= 0 && to_unsigned(specs.precision) < size)
+    size = code_point_index(s, to_unsigned(specs.precision));
+  bool is_debug = specs.type == presentation_type::debug;
+  size_t width = 0;
+
+  if (is_debug) size = write_escaped_string(counting_iterator{}, s).count();
+
+  if (specs.width != 0) {
+    if (is_debug)
+      width = size;
+    else
+      width = compute_width(basic_string_view<Char>(data, size));
+  }
+  return write_padded<Char>(out, specs, size, width,
+                            [=](reserve_iterator<OutputIt> it) {
+                              if (is_debug) return write_escaped_string(it, s);
+                              return copy<Char>(data, data + size, it);
+                            });
+}
+template <typename Char, typename OutputIt>
+FMT_CONSTEXPR auto write(OutputIt out,
+                         basic_string_view<type_identity_t<Char>> s,
+                         const format_specs& specs, locale_ref) -> OutputIt {
+  return write<Char>(out, s, specs);
+}
+template <typename Char, typename OutputIt>
+FMT_CONSTEXPR auto write(OutputIt out, const Char* s, const format_specs& specs,
+                         locale_ref) -> OutputIt {
+  if (specs.type == presentation_type::pointer)
+    return write_ptr<Char>(out, bit_cast<uintptr_t>(s), &specs);
+  if (!s) report_error("string pointer is null");
+  return write<Char>(out, basic_string_view<Char>(s), specs, {});
+}
+
+template <typename Char, typename OutputIt, typename T,
+          FMT_ENABLE_IF(is_integral<T>::value &&
+                        !std::is_same<T, bool>::value &&
+                        !std::is_same<T, Char>::value)>
+FMT_CONSTEXPR auto write(OutputIt out, T value) -> OutputIt {
+  auto abs_value = static_cast<uint32_or_64_or_128_t<T>>(value);
+  bool negative = is_negative(value);
+  // Don't do -abs_value since it trips unsigned-integer-overflow sanitizer.
+  if (negative) abs_value = ~abs_value + 1;
+  int num_digits = count_digits(abs_value);
+  auto size = (negative ? 1 : 0) + static_cast<size_t>(num_digits);
+  if (auto ptr = to_pointer<Char>(out, size)) {
+    if (negative) *ptr++ = static_cast<Char>('-');
+    format_decimal<Char>(ptr, abs_value, num_digits);
+    return out;
+  }
+  if (negative) *out++ = static_cast<Char>('-');
+  return format_decimal<Char>(out, abs_value, num_digits).end;
+}
+
+// DEPRECATED!
+template <typename Char>
+FMT_CONSTEXPR auto parse_align(const Char* begin, const Char* end,
+                               format_specs& specs) -> const Char* {
+  FMT_ASSERT(begin != end, "");
+  auto align = align::none;
+  auto p = begin + code_point_length(begin);
+  if (end - p <= 0) p = begin;
+  for (;;) {
+    switch (to_ascii(*p)) {
+    case '<':
+      align = align::left;
+      break;
+    case '>':
+      align = align::right;
+      break;
+    case '^':
+      align = align::center;
+      break;
+    }
+    if (align != align::none) {
+      if (p != begin) {
+        auto c = *begin;
+        if (c == '}') return begin;
+        if (c == '{') {
+          report_error("invalid fill character '{'");
+          return begin;
+        }
+        specs.fill = basic_string_view<Char>(begin, to_unsigned(p - begin));
+        begin = p + 1;
+      } else {
+        ++begin;
+      }
+      break;
+    } else if (p == begin) {
+      break;
+    }
+    p = begin;
+  }
+  specs.align = align;
+  return begin;
+}
+
+// A floating-point presentation format.
+enum class float_format : unsigned char {
+  general,  // General: exponent notation or fixed point based on magnitude.
+  exp,      // Exponent notation with the default precision of 6, e.g. 1.2e-3.
+  fixed     // Fixed point with the default precision of 6, e.g. 0.0012.
+};
+
+struct float_specs {
+  int precision;
+  float_format format : 8;
+  sign_t sign : 8;
+  bool locale : 1;
+  bool binary32 : 1;
+  bool showpoint : 1;
+};
+
+// DEPRECATED!
+FMT_CONSTEXPR inline auto parse_float_type_spec(const format_specs& specs)
+    -> float_specs {
+  auto result = float_specs();
+  result.showpoint = specs.alt;
+  result.locale = specs.localized;
+  switch (specs.type) {
+  default:
+    FMT_FALLTHROUGH;
+  case presentation_type::none:
+    result.format = float_format::general;
+    break;
+  case presentation_type::exp:
+    result.format = float_format::exp;
+    result.showpoint |= specs.precision != 0;
+    break;
+  case presentation_type::fixed:
+    result.format = float_format::fixed;
+    result.showpoint |= specs.precision != 0;
+    break;
+  case presentation_type::general:
+    result.format = float_format::general;
+    break;
+  }
+  return result;
+}
+
+template <typename Char, typename OutputIt>
+FMT_CONSTEXPR20 auto write_nonfinite(OutputIt out, bool isnan,
+                                     format_specs specs, sign_t sign)
+    -> OutputIt {
+  auto str =
+      isnan ? (specs.upper ? "NAN" : "nan") : (specs.upper ? "INF" : "inf");
+  constexpr size_t str_size = 3;
+  auto size = str_size + (sign ? 1 : 0);
+  // Replace '0'-padding with space for non-finite values.
+  const bool is_zero_fill =
+      specs.fill.size() == 1 && specs.fill.template get<Char>() == '0';
+  if (is_zero_fill) specs.fill = ' ';
+  return write_padded<Char>(out, specs, size,
+                            [=](reserve_iterator<OutputIt> it) {
+                              if (sign) *it++ = detail::sign<Char>(sign);
+                              return copy<Char>(str, str + str_size, it);
+                            });
+}
+
+// A decimal floating-point number significand * pow(10, exp).
+struct big_decimal_fp {
+  const char* significand;
+  int significand_size;
+  int exponent;
+};
+
+constexpr auto get_significand_size(const big_decimal_fp& f) -> int {
+  return f.significand_size;
+}
+template <typename T>
+inline auto get_significand_size(const dragonbox::decimal_fp<T>& f) -> int {
+  return count_digits(f.significand);
+}
+
+template <typename Char, typename OutputIt>
+constexpr auto write_significand(OutputIt out, const char* significand,
+                                 int significand_size) -> OutputIt {
+  return copy<Char>(significand, significand + significand_size, out);
+}
+template <typename Char, typename OutputIt, typename UInt>
+inline auto write_significand(OutputIt out, UInt significand,
+                              int significand_size) -> OutputIt {
+  return format_decimal<Char>(out, significand, significand_size).end;
+}
+template <typename Char, typename OutputIt, typename T, typename Grouping>
+FMT_CONSTEXPR20 auto write_significand(OutputIt out, T significand,
+                                       int significand_size, int exponent,
+                                       const Grouping& grouping) -> OutputIt {
+  if (!grouping.has_separator()) {
+    out = write_significand<Char>(out, significand, significand_size);
+    return detail::fill_n(out, exponent, static_cast<Char>('0'));
+  }
+  auto buffer = memory_buffer();
+  write_significand<char>(appender(buffer), significand, significand_size);
+  detail::fill_n(appender(buffer), exponent, '0');
+  return grouping.apply(out, string_view(buffer.data(), buffer.size()));
+}
+
+template <typename Char, typename UInt,
+          FMT_ENABLE_IF(std::is_integral<UInt>::value)>
+inline auto write_significand(Char* out, UInt significand, int significand_size,
+                              int integral_size, Char decimal_point) -> Char* {
+  if (!decimal_point)
+    return format_decimal(out, significand, significand_size).end;
+  out += significand_size + 1;
+  Char* end = out;
+  int floating_size = significand_size - integral_size;
+  for (int i = floating_size / 2; i > 0; --i) {
+    out -= 2;
+    copy2(out, digits2(static_cast<std::size_t>(significand % 100)));
+    significand /= 100;
+  }
+  if (floating_size % 2 != 0) {
+    *--out = static_cast<Char>('0' + significand % 10);
+    significand /= 10;
+  }
+  *--out = decimal_point;
+  format_decimal(out - integral_size, significand, integral_size);
+  return end;
+}
+
+template <typename OutputIt, typename UInt, typename Char,
+          FMT_ENABLE_IF(!std::is_pointer<remove_cvref_t<OutputIt>>::value)>
+inline auto write_significand(OutputIt out, UInt significand,
+                              int significand_size, int integral_size,
+                              Char decimal_point) -> OutputIt {
+  // Buffer is large enough to hold digits (digits10 + 1) and a decimal point.
+  Char buffer[digits10<UInt>() + 2];
+  auto end = write_significand(buffer, significand, significand_size,
+                               integral_size, decimal_point);
+  return detail::copy_noinline<Char>(buffer, end, out);
+}
+
+template <typename OutputIt, typename Char>
+FMT_CONSTEXPR auto write_significand(OutputIt out, const char* significand,
+                                     int significand_size, int integral_size,
+                                     Char decimal_point) -> OutputIt {
+  out = detail::copy_noinline<Char>(significand, significand + integral_size,
+                                    out);
+  if (!decimal_point) return out;
+  *out++ = decimal_point;
+  return detail::copy_noinline<Char>(significand + integral_size,
+                                     significand + significand_size, out);
+}
+
+template <typename OutputIt, typename Char, typename T, typename Grouping>
+FMT_CONSTEXPR20 auto write_significand(OutputIt out, T significand,
+                                       int significand_size, int integral_size,
+                                       Char decimal_point,
+                                       const Grouping& grouping) -> OutputIt {
+  if (!grouping.has_separator()) {
+    return write_significand(out, significand, significand_size, integral_size,
+                             decimal_point);
+  }
+  auto buffer = basic_memory_buffer<Char>();
+  write_significand(basic_appender<Char>(buffer), significand, significand_size,
+                    integral_size, decimal_point);
+  grouping.apply(
+      out, basic_string_view<Char>(buffer.data(), to_unsigned(integral_size)));
+  return detail::copy_noinline<Char>(buffer.data() + integral_size,
+                                     buffer.end(), out);
+}
+
+template <typename Char, typename OutputIt, typename DecimalFP,
+          typename Grouping = digit_grouping<Char>>
+FMT_CONSTEXPR20 auto do_write_float(OutputIt out, const DecimalFP& f,
+                                    const format_specs& specs,
+                                    float_specs fspecs, locale_ref loc)
+    -> OutputIt {
+  auto significand = f.significand;
+  int significand_size = get_significand_size(f);
+  const Char zero = static_cast<Char>('0');
+  auto sign = fspecs.sign;
+  size_t size = to_unsigned(significand_size) + (sign ? 1 : 0);
+  using iterator = reserve_iterator<OutputIt>;
+
+  Char decimal_point =
+      fspecs.locale ? detail::decimal_point<Char>(loc) : static_cast<Char>('.');
+
+  int output_exp = f.exponent + significand_size - 1;
+  auto use_exp_format = [=]() {
+    if (fspecs.format == float_format::exp) return true;
+    if (fspecs.format != float_format::general) return false;
+    // Use the fixed notation if the exponent is in [exp_lower, exp_upper),
+    // e.g. 0.0001 instead of 1e-04. Otherwise use the exponent notation.
+    const int exp_lower = -4, exp_upper = 16;
+    return output_exp < exp_lower ||
+           output_exp >= (fspecs.precision > 0 ? fspecs.precision : exp_upper);
+  };
+  if (use_exp_format()) {
+    int num_zeros = 0;
+    if (fspecs.showpoint) {
+      num_zeros = fspecs.precision - significand_size;
+      if (num_zeros < 0) num_zeros = 0;
+      size += to_unsigned(num_zeros);
+    } else if (significand_size == 1) {
+      decimal_point = Char();
+    }
+    auto abs_output_exp = output_exp >= 0 ? output_exp : -output_exp;
+    int exp_digits = 2;
+    if (abs_output_exp >= 100) exp_digits = abs_output_exp >= 1000 ? 4 : 3;
+
+    size += to_unsigned((decimal_point ? 1 : 0) + 2 + exp_digits);
+    char exp_char = specs.upper ? 'E' : 'e';
+    auto write = [=](iterator it) {
+      if (sign) *it++ = detail::sign<Char>(sign);
+      // Insert a decimal point after the first digit and add an exponent.
+      it = write_significand(it, significand, significand_size, 1,
+                             decimal_point);
+      if (num_zeros > 0) it = detail::fill_n(it, num_zeros, zero);
+      *it++ = static_cast<Char>(exp_char);
+      return write_exponent<Char>(output_exp, it);
+    };
+    return specs.width > 0
+               ? write_padded<Char, align::right>(out, specs, size, write)
+               : base_iterator(out, write(reserve(out, size)));
+  }
+
+  int exp = f.exponent + significand_size;
+  if (f.exponent >= 0) {
+    // 1234e5 -> 123400000[.0+]
+    size += to_unsigned(f.exponent);
+    int num_zeros = fspecs.precision - exp;
+    abort_fuzzing_if(num_zeros > 5000);
+    if (fspecs.showpoint) {
+      ++size;
+      if (num_zeros <= 0 && fspecs.format != float_format::fixed) num_zeros = 0;
+      if (num_zeros > 0) size += to_unsigned(num_zeros);
+    }
+    auto grouping = Grouping(loc, fspecs.locale);
+    size += to_unsigned(grouping.count_separators(exp));
+    return write_padded<Char, align::right>(out, specs, size, [&](iterator it) {
+      if (sign) *it++ = detail::sign<Char>(sign);
+      it = write_significand<Char>(it, significand, significand_size,
+                                   f.exponent, grouping);
+      if (!fspecs.showpoint) return it;
+      *it++ = decimal_point;
+      return num_zeros > 0 ? detail::fill_n(it, num_zeros, zero) : it;
+    });
+  } else if (exp > 0) {
+    // 1234e-2 -> 12.34[0+]
+    int num_zeros = fspecs.showpoint ? fspecs.precision - significand_size : 0;
+    size += 1 + to_unsigned(num_zeros > 0 ? num_zeros : 0);
+    auto grouping = Grouping(loc, fspecs.locale);
+    size += to_unsigned(grouping.count_separators(exp));
+    return write_padded<Char, align::right>(out, specs, size, [&](iterator it) {
+      if (sign) *it++ = detail::sign<Char>(sign);
+      it = write_significand(it, significand, significand_size, exp,
+                             decimal_point, grouping);
+      return num_zeros > 0 ? detail::fill_n(it, num_zeros, zero) : it;
+    });
+  }
+  // 1234e-6 -> 0.001234
+  int num_zeros = -exp;
+  if (significand_size == 0 && fspecs.precision >= 0 &&
+      fspecs.precision < num_zeros) {
+    num_zeros = fspecs.precision;
+  }
+  bool pointy = num_zeros != 0 || significand_size != 0 || fspecs.showpoint;
+  size += 1 + (pointy ? 1 : 0) + to_unsigned(num_zeros);
+  return write_padded<Char, align::right>(out, specs, size, [&](iterator it) {
+    if (sign) *it++ = detail::sign<Char>(sign);
+    *it++ = zero;
+    if (!pointy) return it;
+    *it++ = decimal_point;
+    it = detail::fill_n(it, num_zeros, zero);
+    return write_significand<Char>(it, significand, significand_size);
+  });
+}
+
+template <typename Char> class fallback_digit_grouping {
+ public:
+  constexpr fallback_digit_grouping(locale_ref, bool) {}
+
+  constexpr auto has_separator() const -> bool { return false; }
+
+  constexpr auto count_separators(int) const -> int { return 0; }
+
+  template <typename Out, typename C>
+  constexpr auto apply(Out out, basic_string_view<C>) const -> Out {
+    return out;
+  }
+};
+
+template <typename Char, typename OutputIt, typename DecimalFP>
+FMT_CONSTEXPR20 auto write_float(OutputIt out, const DecimalFP& f,
+                                 const format_specs& specs, float_specs fspecs,
+                                 locale_ref loc) -> OutputIt {
+  if (is_constant_evaluated()) {
+    return do_write_float<Char, OutputIt, DecimalFP,
+                          fallback_digit_grouping<Char>>(out, f, specs, fspecs,
+                                                         loc);
+  } else {
+    return do_write_float<Char>(out, f, specs, fspecs, loc);
+  }
+}
+
+template <typename T> constexpr auto isnan(T value) -> bool {
+  return value != value;  // std::isnan doesn't support __float128.
+}
+
+template <typename T, typename Enable = void>
+struct has_isfinite : std::false_type {};
+
+template <typename T>
+struct has_isfinite<T, enable_if_t<sizeof(std::isfinite(T())) != 0>>
+    : std::true_type {};
+
+template <typename T, FMT_ENABLE_IF(std::is_floating_point<T>::value&&
+                                        has_isfinite<T>::value)>
+FMT_CONSTEXPR20 auto isfinite(T value) -> bool {
+  constexpr T inf = T(std::numeric_limits<double>::infinity());
+  if (is_constant_evaluated())
+    return !detail::isnan(value) && value < inf && value > -inf;
+  return std::isfinite(value);
+}
+template <typename T, FMT_ENABLE_IF(!has_isfinite<T>::value)>
+FMT_CONSTEXPR auto isfinite(T value) -> bool {
+  T inf = T(std::numeric_limits<double>::infinity());
+  // std::isfinite doesn't support __float128.
+  return !detail::isnan(value) && value < inf && value > -inf;
+}
+
+template <typename T, FMT_ENABLE_IF(is_floating_point<T>::value)>
+FMT_INLINE FMT_CONSTEXPR bool signbit(T value) {
+  if (is_constant_evaluated()) {
+#ifdef __cpp_if_constexpr
+    if constexpr (std::numeric_limits<double>::is_iec559) {
+      auto bits = detail::bit_cast<uint64_t>(static_cast<double>(value));
+      return (bits >> (num_bits<uint64_t>() - 1)) != 0;
+    }
+#endif
+  }
+  return std::signbit(static_cast<double>(value));
+}
+
+inline FMT_CONSTEXPR20 void adjust_precision(int& precision, int exp10) {
+  // Adjust fixed precision by exponent because it is relative to decimal
+  // point.
+  if (exp10 > 0 && precision > max_value<int>() - exp10)
+    FMT_THROW(format_error("number is too big"));
+  precision += exp10;
+}
+
+class bigint {
+ private:
+  // A bigint is stored as an array of bigits (big digits), with bigit at index
+  // 0 being the least significant one.
+  using bigit = uint32_t;
+  using double_bigit = uint64_t;
+  enum { bigits_capacity = 32 };
+  basic_memory_buffer<bigit, bigits_capacity> bigits_;
+  int exp_;
+
+  FMT_CONSTEXPR20 auto operator[](int index) const -> bigit {
+    return bigits_[to_unsigned(index)];
+  }
+  FMT_CONSTEXPR20 auto operator[](int index) -> bigit& {
+    return bigits_[to_unsigned(index)];
+  }
+
+  static constexpr const int bigit_bits = num_bits<bigit>();
+
+  friend struct formatter<bigint>;
+
+  FMT_CONSTEXPR20 void subtract_bigits(int index, bigit other, bigit& borrow) {
+    auto result = static_cast<double_bigit>((*this)[index]) - other - borrow;
+    (*this)[index] = static_cast<bigit>(result);
+    borrow = static_cast<bigit>(result >> (bigit_bits * 2 - 1));
+  }
+
+  FMT_CONSTEXPR20 void remove_leading_zeros() {
+    int num_bigits = static_cast<int>(bigits_.size()) - 1;
+    while (num_bigits > 0 && (*this)[num_bigits] == 0) --num_bigits;
+    bigits_.resize(to_unsigned(num_bigits + 1));
+  }
+
+  // Computes *this -= other assuming aligned bigints and *this >= other.
+  FMT_CONSTEXPR20 void subtract_aligned(const bigint& other) {
+    FMT_ASSERT(other.exp_ >= exp_, "unaligned bigints");
+    FMT_ASSERT(compare(*this, other) >= 0, "");
+    bigit borrow = 0;
+    int i = other.exp_ - exp_;
+    for (size_t j = 0, n = other.bigits_.size(); j != n; ++i, ++j)
+      subtract_bigits(i, other.bigits_[j], borrow);
+    while (borrow > 0) subtract_bigits(i, 0, borrow);
+    remove_leading_zeros();
+  }
+
+  FMT_CONSTEXPR20 void multiply(uint32_t value) {
+    const double_bigit wide_value = value;
+    bigit carry = 0;
+    for (size_t i = 0, n = bigits_.size(); i < n; ++i) {
+      double_bigit result = bigits_[i] * wide_value + carry;
+      bigits_[i] = static_cast<bigit>(result);
+      carry = static_cast<bigit>(result >> bigit_bits);
+    }
+    if (carry != 0) bigits_.push_back(carry);
+  }
+
+  template <typename UInt, FMT_ENABLE_IF(std::is_same<UInt, uint64_t>::value ||
+                                         std::is_same<UInt, uint128_t>::value)>
+  FMT_CONSTEXPR20 void multiply(UInt value) {
+    using half_uint =
+        conditional_t<std::is_same<UInt, uint128_t>::value, uint64_t, uint32_t>;
+    const int shift = num_bits<half_uint>() - bigit_bits;
+    const UInt lower = static_cast<half_uint>(value);
+    const UInt upper = value >> num_bits<half_uint>();
+    UInt carry = 0;
+    for (size_t i = 0, n = bigits_.size(); i < n; ++i) {
+      UInt result = lower * bigits_[i] + static_cast<bigit>(carry);
+      carry = (upper * bigits_[i] << shift) + (result >> bigit_bits) +
+              (carry >> bigit_bits);
+      bigits_[i] = static_cast<bigit>(result);
+    }
+    while (carry != 0) {
+      bigits_.push_back(static_cast<bigit>(carry));
+      carry >>= bigit_bits;
+    }
+  }
+
+  template <typename UInt, FMT_ENABLE_IF(std::is_same<UInt, uint64_t>::value ||
+                                         std::is_same<UInt, uint128_t>::value)>
+  FMT_CONSTEXPR20 void assign(UInt n) {
+    size_t num_bigits = 0;
+    do {
+      bigits_[num_bigits++] = static_cast<bigit>(n);
+      n >>= bigit_bits;
+    } while (n != 0);
+    bigits_.resize(num_bigits);
+    exp_ = 0;
+  }
+
+ public:
+  FMT_CONSTEXPR20 bigint() : exp_(0) {}
+  explicit bigint(uint64_t n) { assign(n); }
+
+  bigint(const bigint&) = delete;
+  void operator=(const bigint&) = delete;
+
+  FMT_CONSTEXPR20 void assign(const bigint& other) {
+    auto size = other.bigits_.size();
+    bigits_.resize(size);
+    auto data = other.bigits_.data();
+    copy<bigit>(data, data + size, bigits_.data());
+    exp_ = other.exp_;
+  }
+
+  template <typename Int> FMT_CONSTEXPR20 void operator=(Int n) {
+    FMT_ASSERT(n > 0, "");
+    assign(uint64_or_128_t<Int>(n));
+  }
+
+  FMT_CONSTEXPR20 auto num_bigits() const -> int {
+    return static_cast<int>(bigits_.size()) + exp_;
+  }
+
+  FMT_NOINLINE FMT_CONSTEXPR20 auto operator<<=(int shift) -> bigint& {
+    FMT_ASSERT(shift >= 0, "");
+    exp_ += shift / bigit_bits;
+    shift %= bigit_bits;
+    if (shift == 0) return *this;
+    bigit carry = 0;
+    for (size_t i = 0, n = bigits_.size(); i < n; ++i) {
+      bigit c = bigits_[i] >> (bigit_bits - shift);
+      bigits_[i] = (bigits_[i] << shift) + carry;
+      carry = c;
+    }
+    if (carry != 0) bigits_.push_back(carry);
+    return *this;
+  }
+
+  template <typename Int>
+  FMT_CONSTEXPR20 auto operator*=(Int value) -> bigint& {
+    FMT_ASSERT(value > 0, "");
+    multiply(uint32_or_64_or_128_t<Int>(value));
+    return *this;
+  }
+
+  friend FMT_CONSTEXPR20 auto compare(const bigint& lhs, const bigint& rhs)
+      -> int {
+    int num_lhs_bigits = lhs.num_bigits(), num_rhs_bigits = rhs.num_bigits();
+    if (num_lhs_bigits != num_rhs_bigits)
+      return num_lhs_bigits > num_rhs_bigits ? 1 : -1;
+    int i = static_cast<int>(lhs.bigits_.size()) - 1;
+    int j = static_cast<int>(rhs.bigits_.size()) - 1;
+    int end = i - j;
+    if (end < 0) end = 0;
+    for (; i >= end; --i, --j) {
+      bigit lhs_bigit = lhs[i], rhs_bigit = rhs[j];
+      if (lhs_bigit != rhs_bigit) return lhs_bigit > rhs_bigit ? 1 : -1;
+    }
+    if (i != j) return i > j ? 1 : -1;
+    return 0;
+  }
+
+  // Returns compare(lhs1 + lhs2, rhs).
+  friend FMT_CONSTEXPR20 auto add_compare(const bigint& lhs1,
+                                          const bigint& lhs2, const bigint& rhs)
+      -> int {
+    auto minimum = [](int a, int b) { return a < b ? a : b; };
+    auto maximum = [](int a, int b) { return a > b ? a : b; };
+    int max_lhs_bigits = maximum(lhs1.num_bigits(), lhs2.num_bigits());
+    int num_rhs_bigits = rhs.num_bigits();
+    if (max_lhs_bigits + 1 < num_rhs_bigits) return -1;
+    if (max_lhs_bigits > num_rhs_bigits) return 1;
+    auto get_bigit = [](const bigint& n, int i) -> bigit {
+      return i >= n.exp_ && i < n.num_bigits() ? n[i - n.exp_] : 0;
+    };
+    double_bigit borrow = 0;
+    int min_exp = minimum(minimum(lhs1.exp_, lhs2.exp_), rhs.exp_);
+    for (int i = num_rhs_bigits - 1; i >= min_exp; --i) {
+      double_bigit sum =
+          static_cast<double_bigit>(get_bigit(lhs1, i)) + get_bigit(lhs2, i);
+      bigit rhs_bigit = get_bigit(rhs, i);
+      if (sum > rhs_bigit + borrow) return 1;
+      borrow = rhs_bigit + borrow - sum;
+      if (borrow > 1) return -1;
+      borrow <<= bigit_bits;
+    }
+    return borrow != 0 ? -1 : 0;
+  }
+
+  // Assigns pow(10, exp) to this bigint.
+  FMT_CONSTEXPR20 void assign_pow10(int exp) {
+    FMT_ASSERT(exp >= 0, "");
+    if (exp == 0) return *this = 1;
+    // Find the top bit.
+    int bitmask = 1;
+    while (exp >= bitmask) bitmask <<= 1;
+    bitmask >>= 1;
+    // pow(10, exp) = pow(5, exp) * pow(2, exp). First compute pow(5, exp) by
+    // repeated squaring and multiplication.
+    *this = 5;
+    bitmask >>= 1;
+    while (bitmask != 0) {
+      square();
+      if ((exp & bitmask) != 0) *this *= 5;
+      bitmask >>= 1;
+    }
+    *this <<= exp;  // Multiply by pow(2, exp) by shifting.
+  }
+
+  FMT_CONSTEXPR20 void square() {
+    int num_bigits = static_cast<int>(bigits_.size());
+    int num_result_bigits = 2 * num_bigits;
+    basic_memory_buffer<bigit, bigits_capacity> n(std::move(bigits_));
+    bigits_.resize(to_unsigned(num_result_bigits));
+    auto sum = uint128_t();
+    for (int bigit_index = 0; bigit_index < num_bigits; ++bigit_index) {
+      // Compute bigit at position bigit_index of the result by adding
+      // cross-product terms n[i] * n[j] such that i + j == bigit_index.
+      for (int i = 0, j = bigit_index; j >= 0; ++i, --j) {
+        // Most terms are multiplied twice which can be optimized in the future.
+        sum += static_cast<double_bigit>(n[i]) * n[j];
+      }
+      (*this)[bigit_index] = static_cast<bigit>(sum);
+      sum >>= num_bits<bigit>();  // Compute the carry.
+    }
+    // Do the same for the top half.
+    for (int bigit_index = num_bigits; bigit_index < num_result_bigits;
+         ++bigit_index) {
+      for (int j = num_bigits - 1, i = bigit_index - j; i < num_bigits;)
+        sum += static_cast<double_bigit>(n[i++]) * n[j--];
+      (*this)[bigit_index] = static_cast<bigit>(sum);
+      sum >>= num_bits<bigit>();
+    }
+    remove_leading_zeros();
+    exp_ *= 2;
+  }
+
+  // If this bigint has a bigger exponent than other, adds trailing zero to make
+  // exponents equal. This simplifies some operations such as subtraction.
+  FMT_CONSTEXPR20 void align(const bigint& other) {
+    int exp_difference = exp_ - other.exp_;
+    if (exp_difference <= 0) return;
+    int num_bigits = static_cast<int>(bigits_.size());
+    bigits_.resize(to_unsigned(num_bigits + exp_difference));
+    for (int i = num_bigits - 1, j = i + exp_difference; i >= 0; --i, --j)
+      bigits_[j] = bigits_[i];
+    memset(bigits_.data(), 0, to_unsigned(exp_difference) * sizeof(bigit));
+    exp_ -= exp_difference;
+  }
+
+  // Divides this bignum by divisor, assigning the remainder to this and
+  // returning the quotient.
+  FMT_CONSTEXPR20 auto divmod_assign(const bigint& divisor) -> int {
+    FMT_ASSERT(this != &divisor, "");
+    if (compare(*this, divisor) < 0) return 0;
+    FMT_ASSERT(divisor.bigits_[divisor.bigits_.size() - 1u] != 0, "");
+    align(divisor);
+    int quotient = 0;
+    do {
+      subtract_aligned(divisor);
+      ++quotient;
+    } while (compare(*this, divisor) >= 0);
+    return quotient;
+  }
+};
+
+// format_dragon flags.
+enum dragon {
+  predecessor_closer = 1,
+  fixup = 2,  // Run fixup to correct exp10 which can be off by one.
+  fixed = 4,
+};
+
+// Formats a floating-point number using a variation of the Fixed-Precision
+// Positive Floating-Point Printout ((FPP)^2) algorithm by Steele & White:
+// https://fmt.dev/papers/p372-steele.pdf.
+FMT_CONSTEXPR20 inline void format_dragon(basic_fp<uint128_t> value,
+                                          unsigned flags, int num_digits,
+                                          buffer<char>& buf, int& exp10) {
+  bigint numerator;    // 2 * R in (FPP)^2.
+  bigint denominator;  // 2 * S in (FPP)^2.
+  // lower and upper are differences between value and corresponding boundaries.
+  bigint lower;             // (M^- in (FPP)^2).
+  bigint upper_store;       // upper's value if different from lower.
+  bigint* upper = nullptr;  // (M^+ in (FPP)^2).
+  // Shift numerator and denominator by an extra bit or two (if lower boundary
+  // is closer) to make lower and upper integers. This eliminates multiplication
+  // by 2 during later computations.
+  bool is_predecessor_closer = (flags & dragon::predecessor_closer) != 0;
+  int shift = is_predecessor_closer ? 2 : 1;
+  if (value.e >= 0) {
+    numerator = value.f;
+    numerator <<= value.e + shift;
+    lower = 1;
+    lower <<= value.e;
+    if (is_predecessor_closer) {
+      upper_store = 1;
+      upper_store <<= value.e + 1;
+      upper = &upper_store;
+    }
+    denominator.assign_pow10(exp10);
+    denominator <<= shift;
+  } else if (exp10 < 0) {
+    numerator.assign_pow10(-exp10);
+    lower.assign(numerator);
+    if (is_predecessor_closer) {
+      upper_store.assign(numerator);
+      upper_store <<= 1;
+      upper = &upper_store;
+    }
+    numerator *= value.f;
+    numerator <<= shift;
+    denominator = 1;
+    denominator <<= shift - value.e;
+  } else {
+    numerator = value.f;
+    numerator <<= shift;
+    denominator.assign_pow10(exp10);
+    denominator <<= shift - value.e;
+    lower = 1;
+    if (is_predecessor_closer) {
+      upper_store = 1ULL << 1;
+      upper = &upper_store;
+    }
+  }
+  int even = static_cast<int>((value.f & 1) == 0);
+  if (!upper) upper = &lower;
+  bool shortest = num_digits < 0;
+  if ((flags & dragon::fixup) != 0) {
+    if (add_compare(numerator, *upper, denominator) + even <= 0) {
+      --exp10;
+      numerator *= 10;
+      if (num_digits < 0) {
+        lower *= 10;
+        if (upper != &lower) *upper *= 10;
+      }
+    }
+    if ((flags & dragon::fixed) != 0) adjust_precision(num_digits, exp10 + 1);
+  }
+  // Invariant: value == (numerator / denominator) * pow(10, exp10).
+  if (shortest) {
+    // Generate the shortest representation.
+    num_digits = 0;
+    char* data = buf.data();
+    for (;;) {
+      int digit = numerator.divmod_assign(denominator);
+      bool low = compare(numerator, lower) - even < 0;  // numerator <[=] lower.
+      // numerator + upper >[=] pow10:
+      bool high = add_compare(numerator, *upper, denominator) + even > 0;
+      data[num_digits++] = static_cast<char>('0' + digit);
+      if (low || high) {
+        if (!low) {
+          ++data[num_digits - 1];
+        } else if (high) {
+          int result = add_compare(numerator, numerator, denominator);
+          // Round half to even.
+          if (result > 0 || (result == 0 && (digit % 2) != 0))
+            ++data[num_digits - 1];
+        }
+        buf.try_resize(to_unsigned(num_digits));
+        exp10 -= num_digits - 1;
+        return;
+      }
+      numerator *= 10;
+      lower *= 10;
+      if (upper != &lower) *upper *= 10;
+    }
+  }
+  // Generate the given number of digits.
+  exp10 -= num_digits - 1;
+  if (num_digits <= 0) {
+    auto digit = '0';
+    if (num_digits == 0) {
+      denominator *= 10;
+      digit = add_compare(numerator, numerator, denominator) > 0 ? '1' : '0';
+    }
+    buf.push_back(digit);
+    return;
+  }
+  buf.try_resize(to_unsigned(num_digits));
+  for (int i = 0; i < num_digits - 1; ++i) {
+    int digit = numerator.divmod_assign(denominator);
+    buf[i] = static_cast<char>('0' + digit);
+    numerator *= 10;
+  }
+  int digit = numerator.divmod_assign(denominator);
+  auto result = add_compare(numerator, numerator, denominator);
+  if (result > 0 || (result == 0 && (digit % 2) != 0)) {
+    if (digit == 9) {
+      const auto overflow = '0' + 10;
+      buf[num_digits - 1] = overflow;
+      // Propagate the carry.
+      for (int i = num_digits - 1; i > 0 && buf[i] == overflow; --i) {
+        buf[i] = '0';
+        ++buf[i - 1];
+      }
+      if (buf[0] == overflow) {
+        buf[0] = '1';
+        if ((flags & dragon::fixed) != 0)
+          buf.push_back('0');
+        else
+          ++exp10;
+      }
+      return;
+    }
+    ++digit;
+  }
+  buf[num_digits - 1] = static_cast<char>('0' + digit);
+}
+
+// Formats a floating-point number using the hexfloat format.
+template <typename Float, FMT_ENABLE_IF(!is_double_double<Float>::value)>
+FMT_CONSTEXPR20 void format_hexfloat(Float value, format_specs specs,
+                                     buffer<char>& buf) {
+  // float is passed as double to reduce the number of instantiations and to
+  // simplify implementation.
+  static_assert(!std::is_same<Float, float>::value, "");
+
+  using info = dragonbox::float_info<Float>;
+
+  // Assume Float is in the format [sign][exponent][significand].
+  using carrier_uint = typename info::carrier_uint;
+
+  constexpr auto num_float_significand_bits =
+      detail::num_significand_bits<Float>();
+
+  basic_fp<carrier_uint> f(value);
+  f.e += num_float_significand_bits;
+  if (!has_implicit_bit<Float>()) --f.e;
+
+  constexpr auto num_fraction_bits =
+      num_float_significand_bits + (has_implicit_bit<Float>() ? 1 : 0);
+  constexpr auto num_xdigits = (num_fraction_bits + 3) / 4;
+
+  constexpr auto leading_shift = ((num_xdigits - 1) * 4);
+  const auto leading_mask = carrier_uint(0xF) << leading_shift;
+  const auto leading_xdigit =
+      static_cast<uint32_t>((f.f & leading_mask) >> leading_shift);
+  if (leading_xdigit > 1) f.e -= (32 - countl_zero(leading_xdigit) - 1);
+
+  int print_xdigits = num_xdigits - 1;
+  if (specs.precision >= 0 && print_xdigits > specs.precision) {
+    const int shift = ((print_xdigits - specs.precision - 1) * 4);
+    const auto mask = carrier_uint(0xF) << shift;
+    const auto v = static_cast<uint32_t>((f.f & mask) >> shift);
+
+    if (v >= 8) {
+      const auto inc = carrier_uint(1) << (shift + 4);
+      f.f += inc;
+      f.f &= ~(inc - 1);
+    }
+
+    // Check long double overflow
+    if (!has_implicit_bit<Float>()) {
+      const auto implicit_bit = carrier_uint(1) << num_float_significand_bits;
+      if ((f.f & implicit_bit) == implicit_bit) {
+        f.f >>= 4;
+        f.e += 4;
+      }
+    }
+
+    print_xdigits = specs.precision;
+  }
+
+  char xdigits[num_bits<carrier_uint>() / 4];
+  detail::fill_n(xdigits, sizeof(xdigits), '0');
+  format_uint<4>(xdigits, f.f, num_xdigits, specs.upper);
+
+  // Remove zero tail
+  while (print_xdigits > 0 && xdigits[print_xdigits] == '0') --print_xdigits;
+
+  buf.push_back('0');
+  buf.push_back(specs.upper ? 'X' : 'x');
+  buf.push_back(xdigits[0]);
+  if (specs.alt || print_xdigits > 0 || print_xdigits < specs.precision)
+    buf.push_back('.');
+  buf.append(xdigits + 1, xdigits + 1 + print_xdigits);
+  for (; print_xdigits < specs.precision; ++print_xdigits) buf.push_back('0');
+
+  buf.push_back(specs.upper ? 'P' : 'p');
+
+  uint32_t abs_e;
+  if (f.e < 0) {
+    buf.push_back('-');
+    abs_e = static_cast<uint32_t>(-f.e);
+  } else {
+    buf.push_back('+');
+    abs_e = static_cast<uint32_t>(f.e);
+  }
+  format_decimal<char>(appender(buf), abs_e, detail::count_digits(abs_e));
+}
+
+template <typename Float, FMT_ENABLE_IF(is_double_double<Float>::value)>
+FMT_CONSTEXPR20 void format_hexfloat(Float value, format_specs specs,
+                                     buffer<char>& buf) {
+  format_hexfloat(static_cast<double>(value), specs, buf);
+}
+
+constexpr auto fractional_part_rounding_thresholds(int index) -> uint32_t {
+  // For checking rounding thresholds.
+  // The kth entry is chosen to be the smallest integer such that the
+  // upper 32-bits of 10^(k+1) times it is strictly bigger than 5 * 10^k.
+  // It is equal to ceil(2^31 + 2^32/10^(k + 1)).
+  // These are stored in a string literal because we cannot have static arrays
+  // in constexpr functions and non-static ones are poorly optimized.
+  return U"\x9999999a\x828f5c29\x80418938\x80068db9\x8000a7c6\x800010c7"
+         U"\x800001ae\x8000002b"[index];
+}
+
+template <typename Float>
+FMT_CONSTEXPR20 auto format_float(Float value, int precision, float_specs specs,
+                                  buffer<char>& buf) -> int {
+  // float is passed as double to reduce the number of instantiations.
+  static_assert(!std::is_same<Float, float>::value, "");
+  FMT_ASSERT(value >= 0, "value is negative");
+  auto converted_value = convert_float(value);
+
+  const bool fixed = specs.format == float_format::fixed;
+  if (value <= 0) {  // <= instead of == to silence a warning.
+    if (precision <= 0 || !fixed) {
+      buf.push_back('0');
+      return 0;
+    }
+    buf.try_resize(to_unsigned(precision));
+    fill_n(buf.data(), precision, '0');
+    return -precision;
+  }
+
+  int exp = 0;
+  bool use_dragon = true;
+  unsigned dragon_flags = 0;
+  if (!is_fast_float<Float>() || is_constant_evaluated()) {
+    const auto inv_log2_10 = 0.3010299956639812;  // 1 / log2(10)
+    using info = dragonbox::float_info<decltype(converted_value)>;
+    const auto f = basic_fp<typename info::carrier_uint>(converted_value);
+    // Compute exp, an approximate power of 10, such that
+    //   10^(exp - 1) <= value < 10^exp or 10^exp <= value < 10^(exp + 1).
+    // This is based on log10(value) == log2(value) / log2(10) and approximation
+    // of log2(value) by e + num_fraction_bits idea from double-conversion.
+    auto e = (f.e + count_digits<1>(f.f) - 1) * inv_log2_10 - 1e-10;
+    exp = static_cast<int>(e);
+    if (e > exp) ++exp;  // Compute ceil.
+    dragon_flags = dragon::fixup;
+  } else if (precision < 0) {
+    // Use Dragonbox for the shortest format.
+    if (specs.binary32) {
+      auto dec = dragonbox::to_decimal(static_cast<float>(value));
+      write<char>(appender(buf), dec.significand);
+      return dec.exponent;
+    }
+    auto dec = dragonbox::to_decimal(static_cast<double>(value));
+    write<char>(appender(buf), dec.significand);
+    return dec.exponent;
+  } else {
+    // Extract significand bits and exponent bits.
+    using info = dragonbox::float_info<double>;
+    auto br = bit_cast<uint64_t>(static_cast<double>(value));
+
+    const uint64_t significand_mask =
+        (static_cast<uint64_t>(1) << num_significand_bits<double>()) - 1;
+    uint64_t significand = (br & significand_mask);
+    int exponent = static_cast<int>((br & exponent_mask<double>()) >>
+                                    num_significand_bits<double>());
+
+    if (exponent != 0) {  // Check if normal.
+      exponent -= exponent_bias<double>() + num_significand_bits<double>();
+      significand |=
+          (static_cast<uint64_t>(1) << num_significand_bits<double>());
+      significand <<= 1;
+    } else {
+      // Normalize subnormal inputs.
+      FMT_ASSERT(significand != 0, "zeros should not appear here");
+      int shift = countl_zero(significand);
+      FMT_ASSERT(shift >= num_bits<uint64_t>() - num_significand_bits<double>(),
+                 "");
+      shift -= (num_bits<uint64_t>() - num_significand_bits<double>() - 2);
+      exponent = (std::numeric_limits<double>::min_exponent -
+                  num_significand_bits<double>()) -
+                 shift;
+      significand <<= shift;
+    }
+
+    // Compute the first several nonzero decimal significand digits.
+    // We call the number we get the first segment.
+    const int k = info::kappa - dragonbox::floor_log10_pow2(exponent);
+    exp = -k;
+    const int beta = exponent + dragonbox::floor_log2_pow10(k);
+    uint64_t first_segment;
+    bool has_more_segments;
+    int digits_in_the_first_segment;
+    {
+      const auto r = dragonbox::umul192_upper128(
+          significand << beta, dragonbox::get_cached_power(k));
+      first_segment = r.high();
+      has_more_segments = r.low() != 0;
+
+      // The first segment can have 18 ~ 19 digits.
+      if (first_segment >= 1000000000000000000ULL) {
+        digits_in_the_first_segment = 19;
+      } else {
+        // When it is of 18-digits, we align it to 19-digits by adding a bogus
+        // zero at the end.
+        digits_in_the_first_segment = 18;
+        first_segment *= 10;
+      }
+    }
+
+    // Compute the actual number of decimal digits to print.
+    if (fixed) adjust_precision(precision, exp + digits_in_the_first_segment);
+
+    // Use Dragon4 only when there might be not enough digits in the first
+    // segment.
+    if (digits_in_the_first_segment > precision) {
+      use_dragon = false;
+
+      if (precision <= 0) {
+        exp += digits_in_the_first_segment;
+
+        if (precision < 0) {
+          // Nothing to do, since all we have are just leading zeros.
+          buf.try_resize(0);
+        } else {
+          // We may need to round-up.
+          buf.try_resize(1);
+          if ((first_segment | static_cast<uint64_t>(has_more_segments)) >
+              5000000000000000000ULL) {
+            buf[0] = '1';
+          } else {
+            buf[0] = '0';
+          }
+        }
+      }  // precision <= 0
+      else {
+        exp += digits_in_the_first_segment - precision;
+
+        // When precision > 0, we divide the first segment into three
+        // subsegments, each with 9, 9, and 0 ~ 1 digits so that each fits
+        // in 32-bits which usually allows faster calculation than in
+        // 64-bits. Since some compiler (e.g. MSVC) doesn't know how to optimize
+        // division-by-constant for large 64-bit divisors, we do it here
+        // manually. The magic number 7922816251426433760 below is equal to
+        // ceil(2^(64+32) / 10^10).
+        const uint32_t first_subsegment = static_cast<uint32_t>(
+            dragonbox::umul128_upper64(first_segment, 7922816251426433760ULL) >>
+            32);
+        const uint64_t second_third_subsegments =
+            first_segment - first_subsegment * 10000000000ULL;
+
+        uint64_t prod;
+        uint32_t digits;
+        bool should_round_up;
+        int number_of_digits_to_print = precision > 9 ? 9 : precision;
+
+        // Print a 9-digits subsegment, either the first or the second.
+        auto print_subsegment = [&](uint32_t subsegment, char* buffer) {
+          int number_of_digits_printed = 0;
+
+          // If we want to print an odd number of digits from the subsegment,
+          if ((number_of_digits_to_print & 1) != 0) {
+            // Convert to 64-bit fixed-point fractional form with 1-digit
+            // integer part. The magic number 720575941 is a good enough
+            // approximation of 2^(32 + 24) / 10^8; see
+            // https://jk-jeon.github.io/posts/2022/12/fixed-precision-formatting/#fixed-length-case
+            // for details.
+            prod = ((subsegment * static_cast<uint64_t>(720575941)) >> 24) + 1;
+            digits = static_cast<uint32_t>(prod >> 32);
+            *buffer = static_cast<char>('0' + digits);
+            number_of_digits_printed++;
+          }
+          // If we want to print an even number of digits from the
+          // first_subsegment,
+          else {
+            // Convert to 64-bit fixed-point fractional form with 2-digits
+            // integer part. The magic number 450359963 is a good enough
+            // approximation of 2^(32 + 20) / 10^7; see
+            // https://jk-jeon.github.io/posts/2022/12/fixed-precision-formatting/#fixed-length-case
+            // for details.
+            prod = ((subsegment * static_cast<uint64_t>(450359963)) >> 20) + 1;
+            digits = static_cast<uint32_t>(prod >> 32);
+            copy2(buffer, digits2(digits));
+            number_of_digits_printed += 2;
+          }
+
+          // Print all digit pairs.
+          while (number_of_digits_printed < number_of_digits_to_print) {
+            prod = static_cast<uint32_t>(prod) * static_cast<uint64_t>(100);
+            digits = static_cast<uint32_t>(prod >> 32);
+            copy2(buffer + number_of_digits_printed, digits2(digits));
+            number_of_digits_printed += 2;
+          }
+        };
+
+        // Print first subsegment.
+        print_subsegment(first_subsegment, buf.data());
+
+        // Perform rounding if the first subsegment is the last subsegment to
+        // print.
+        if (precision <= 9) {
+          // Rounding inside the subsegment.
+          // We round-up if:
+          //  - either the fractional part is strictly larger than 1/2, or
+          //  - the fractional part is exactly 1/2 and the last digit is odd.
+          // We rely on the following observations:
+          //  - If fractional_part >= threshold, then the fractional part is
+          //    strictly larger than 1/2.
+          //  - If the MSB of fractional_part is set, then the fractional part
+          //    must be at least 1/2.
+          //  - When the MSB of fractional_part is set, either
+          //    second_third_subsegments being nonzero or has_more_segments
+          //    being true means there are further digits not printed, so the
+          //    fractional part is strictly larger than 1/2.
+          if (precision < 9) {
+            uint32_t fractional_part = static_cast<uint32_t>(prod);
+            should_round_up =
+                fractional_part >= fractional_part_rounding_thresholds(
+                                       8 - number_of_digits_to_print) ||
+                ((fractional_part >> 31) &
+                 ((digits & 1) | (second_third_subsegments != 0) |
+                  has_more_segments)) != 0;
+          }
+          // Rounding at the subsegment boundary.
+          // In this case, the fractional part is at least 1/2 if and only if
+          // second_third_subsegments >= 5000000000ULL, and is strictly larger
+          // than 1/2 if we further have either second_third_subsegments >
+          // 5000000000ULL or has_more_segments == true.
+          else {
+            should_round_up = second_third_subsegments > 5000000000ULL ||
+                              (second_third_subsegments == 5000000000ULL &&
+                               ((digits & 1) != 0 || has_more_segments));
+          }
+        }
+        // Otherwise, print the second subsegment.
+        else {
+          // Compilers are not aware of how to leverage the maximum value of
+          // second_third_subsegments to find out a better magic number which
+          // allows us to eliminate an additional shift. 1844674407370955162 =
+          // ceil(2^64/10) < ceil(2^64*(10^9/(10^10 - 1))).
+          const uint32_t second_subsegment =
+              static_cast<uint32_t>(dragonbox::umul128_upper64(
+                  second_third_subsegments, 1844674407370955162ULL));
+          const uint32_t third_subsegment =
+              static_cast<uint32_t>(second_third_subsegments) -
+              second_subsegment * 10;
+
+          number_of_digits_to_print = precision - 9;
+          print_subsegment(second_subsegment, buf.data() + 9);
+
+          // Rounding inside the subsegment.
+          if (precision < 18) {
+            // The condition third_subsegment != 0 implies that the segment was
+            // of 19 digits, so in this case the third segment should be
+            // consisting of a genuine digit from the input.
+            uint32_t fractional_part = static_cast<uint32_t>(prod);
+            should_round_up =
+                fractional_part >= fractional_part_rounding_thresholds(
+                                       8 - number_of_digits_to_print) ||
+                ((fractional_part >> 31) &
+                 ((digits & 1) | (third_subsegment != 0) |
+                  has_more_segments)) != 0;
+          }
+          // Rounding at the subsegment boundary.
+          else {
+            // In this case, the segment must be of 19 digits, thus
+            // the third subsegment should be consisting of a genuine digit from
+            // the input.
+            should_round_up = third_subsegment > 5 ||
+                              (third_subsegment == 5 &&
+                               ((digits & 1) != 0 || has_more_segments));
+          }
+        }
+
+        // Round-up if necessary.
+        if (should_round_up) {
+          ++buf[precision - 1];
+          for (int i = precision - 1; i > 0 && buf[i] > '9'; --i) {
+            buf[i] = '0';
+            ++buf[i - 1];
+          }
+          if (buf[0] > '9') {
+            buf[0] = '1';
+            if (fixed)
+              buf[precision++] = '0';
+            else
+              ++exp;
+          }
+        }
+        buf.try_resize(to_unsigned(precision));
+      }
+    }  // if (digits_in_the_first_segment > precision)
+    else {
+      // Adjust the exponent for its use in Dragon4.
+      exp += digits_in_the_first_segment - 1;
+    }
+  }
+  if (use_dragon) {
+    auto f = basic_fp<uint128_t>();
+    bool is_predecessor_closer = specs.binary32
+                                     ? f.assign(static_cast<float>(value))
+                                     : f.assign(converted_value);
+    if (is_predecessor_closer) dragon_flags |= dragon::predecessor_closer;
+    if (fixed) dragon_flags |= dragon::fixed;
+    // Limit precision to the maximum possible number of significant digits in
+    // an IEEE754 double because we don't need to generate zeros.
+    const int max_double_digits = 767;
+    if (precision > max_double_digits) precision = max_double_digits;
+    format_dragon(f, dragon_flags, precision, buf, exp);
+  }
+  if (!fixed && !specs.showpoint) {
+    // Remove trailing zeros.
+    auto num_digits = buf.size();
+    while (num_digits > 0 && buf[num_digits - 1] == '0') {
+      --num_digits;
+      ++exp;
+    }
+    buf.try_resize(num_digits);
+  }
+  return exp;
+}
+
+template <typename Char, typename OutputIt, typename T>
+FMT_CONSTEXPR20 auto write_float(OutputIt out, T value, format_specs specs,
+                                 locale_ref loc) -> OutputIt {
+  sign_t sign = specs.sign;
+  if (detail::signbit(value)) {  // value < 0 is false for NaN so use signbit.
+    sign = sign::minus;
+    value = -value;
+  } else if (sign == sign::minus) {
+    sign = sign::none;
+  }
+
+  if (!detail::isfinite(value))
+    return write_nonfinite<Char>(out, detail::isnan(value), specs, sign);
+
+  if (specs.align == align::numeric && sign) {
+    auto it = reserve(out, 1);
+    *it++ = detail::sign<Char>(sign);
+    out = base_iterator(out, it);
+    sign = sign::none;
+    if (specs.width != 0) --specs.width;
+  }
+
+  memory_buffer buffer;
+  if (specs.type == presentation_type::hexfloat) {
+    if (sign) buffer.push_back(detail::sign<char>(sign));
+    format_hexfloat(convert_float(value), specs, buffer);
+    return write_bytes<Char, align::right>(out, {buffer.data(), buffer.size()},
+                                           specs);
+  }
+
+  int precision = specs.precision >= 0 || specs.type == presentation_type::none
+                      ? specs.precision
+                      : 6;
+  if (specs.type == presentation_type::exp) {
+    if (precision == max_value<int>())
+      report_error("number is too big");
+    else
+      ++precision;
+  } else if (specs.type != presentation_type::fixed && precision == 0) {
+    precision = 1;
+  }
+  float_specs fspecs = parse_float_type_spec(specs);
+  fspecs.sign = sign;
+  if (const_check(std::is_same<T, float>())) fspecs.binary32 = true;
+  int exp = format_float(convert_float(value), precision, fspecs, buffer);
+  fspecs.precision = precision;
+  auto f = big_decimal_fp{buffer.data(), static_cast<int>(buffer.size()), exp};
+  return write_float<Char>(out, f, specs, fspecs, loc);
+}
+
+template <typename Char, typename OutputIt, typename T,
+          FMT_ENABLE_IF(is_floating_point<T>::value)>
+FMT_CONSTEXPR20 auto write(OutputIt out, T value, format_specs specs,
+                           locale_ref loc = {}) -> OutputIt {
+  if (const_check(!is_supported_floating_point(value))) return out;
+  return specs.localized && write_loc(out, value, specs, loc)
+             ? out
+             : write_float<Char>(out, value, specs, loc);
+}
+
+template <typename Char, typename OutputIt, typename T,
+          FMT_ENABLE_IF(is_fast_float<T>::value)>
+FMT_CONSTEXPR20 auto write(OutputIt out, T value) -> OutputIt {
+  if (is_constant_evaluated()) return write<Char>(out, value, format_specs());
+  if (const_check(!is_supported_floating_point(value))) return out;
+
+  auto sign = sign_t::none;
+  if (detail::signbit(value)) {
+    sign = sign::minus;
+    value = -value;
+  }
+
+  constexpr auto specs = format_specs();
+  using floaty = conditional_t<std::is_same<T, long double>::value, double, T>;
+  using floaty_uint = typename dragonbox::float_info<floaty>::carrier_uint;
+  floaty_uint mask = exponent_mask<floaty>();
+  if ((bit_cast<floaty_uint>(value) & mask) == mask)
+    return write_nonfinite<Char>(out, std::isnan(value), specs, sign);
+
+  auto fspecs = float_specs();
+  fspecs.sign = sign;
+  auto dec = dragonbox::to_decimal(static_cast<floaty>(value));
+  return write_float<Char>(out, dec, specs, fspecs, {});
+}
+
+template <typename Char, typename OutputIt, typename T,
+          FMT_ENABLE_IF(is_floating_point<T>::value &&
+                        !is_fast_float<T>::value)>
+inline auto write(OutputIt out, T value) -> OutputIt {
+  return write<Char>(out, value, format_specs());
+}
+
+template <typename Char, typename OutputIt>
+auto write(OutputIt out, monostate, format_specs = {}, locale_ref = {})
+    -> OutputIt {
+  FMT_ASSERT(false, "");
+  return out;
+}
+
+template <typename Char, typename OutputIt>
+FMT_CONSTEXPR auto write(OutputIt out, basic_string_view<Char> value)
+    -> OutputIt {
+  return copy_noinline<Char>(value.begin(), value.end(), out);
+}
+
+template <typename Char, typename OutputIt, typename T,
+          FMT_ENABLE_IF(has_to_string_view<T>::value)>
+constexpr auto write(OutputIt out, const T& value) -> OutputIt {
+  return write<Char>(out, to_string_view(value));
+}
+
+// FMT_ENABLE_IF() condition separated to workaround an MSVC bug.
+template <
+    typename Char, typename OutputIt, typename T,
+    bool check =
+        std::is_enum<T>::value && !std::is_same<T, Char>::value &&
+        mapped_type_constant<T, basic_format_context<OutputIt, Char>>::value !=
+            type::custom_type,
+    FMT_ENABLE_IF(check)>
+FMT_CONSTEXPR auto write(OutputIt out, T value) -> OutputIt {
+  return write<Char>(out, static_cast<underlying_t<T>>(value));
+}
+
+template <typename Char, typename OutputIt, typename T,
+          FMT_ENABLE_IF(std::is_same<T, bool>::value)>
+FMT_CONSTEXPR auto write(OutputIt out, T value, const format_specs& specs = {},
+                         locale_ref = {}) -> OutputIt {
+  return specs.type != presentation_type::none &&
+                 specs.type != presentation_type::string
+             ? write<Char>(out, value ? 1 : 0, specs, {})
+             : write_bytes<Char>(out, value ? "true" : "false", specs);
+}
+
+template <typename Char, typename OutputIt>
+FMT_CONSTEXPR auto write(OutputIt out, Char value) -> OutputIt {
+  auto it = reserve(out, 1);
+  *it++ = value;
+  return base_iterator(out, it);
+}
+
+template <typename Char, typename OutputIt>
+FMT_CONSTEXPR20 auto write(OutputIt out, const Char* value) -> OutputIt {
+  if (value) return write(out, basic_string_view<Char>(value));
+  report_error("string pointer is null");
+  return out;
+}
+
+template <typename Char, typename OutputIt, typename T,
+          FMT_ENABLE_IF(std::is_same<T, void>::value)>
+auto write(OutputIt out, const T* value, const format_specs& specs = {},
+           locale_ref = {}) -> OutputIt {
+  return write_ptr<Char>(out, bit_cast<uintptr_t>(value), &specs);
+}
+
+// A write overload that handles implicit conversions.
+template <typename Char, typename OutputIt, typename T,
+          typename Context = basic_format_context<OutputIt, Char>>
+FMT_CONSTEXPR auto write(OutputIt out, const T& value) -> enable_if_t<
+    std::is_class<T>::value && !has_to_string_view<T>::value &&
+        !is_floating_point<T>::value && !std::is_same<T, Char>::value &&
+        !std::is_same<T, remove_cvref_t<decltype(arg_mapper<Context>().map(
+                             value))>>::value,
+    OutputIt> {
+  return write<Char>(out, arg_mapper<Context>().map(value));
+}
+
+template <typename Char, typename OutputIt, typename T,
+          typename Context = basic_format_context<OutputIt, Char>>
+FMT_CONSTEXPR auto write(OutputIt out, const T& value)
+    -> enable_if_t<mapped_type_constant<T, Context>::value ==
+                           type::custom_type &&
+                       !std::is_fundamental<T>::value,
+                   OutputIt> {
+  auto formatter = typename Context::template formatter_type<T>();
+  auto parse_ctx = typename Context::parse_context_type({});
+  formatter.parse(parse_ctx);
+  auto ctx = Context(out, {}, {});
+  return formatter.format(value, ctx);
+}
+
+// An argument visitor that formats the argument and writes it via the output
+// iterator. It's a class and not a generic lambda for compatibility with C++11.
+template <typename Char> struct default_arg_formatter {
+  using iterator = basic_appender<Char>;
+  using context = buffered_context<Char>;
+
+  iterator out;
+  basic_format_args<context> args;
+  locale_ref loc;
+
+  template <typename T> auto operator()(T value) -> iterator {
+    return write<Char>(out, value);
+  }
+  auto operator()(typename basic_format_arg<context>::handle h) -> iterator {
+    basic_format_parse_context<Char> parse_ctx({});
+    context format_ctx(out, args, loc);
+    h.format(parse_ctx, format_ctx);
+    return format_ctx.out();
+  }
+};
+
+template <typename Char> struct arg_formatter {
+  using iterator = basic_appender<Char>;
+  using context = buffered_context<Char>;
+
+  iterator out;
+  const format_specs& specs;
+  locale_ref locale;
+
+  template <typename T>
+  FMT_CONSTEXPR FMT_INLINE auto operator()(T value) -> iterator {
+    return detail::write<Char>(out, value, specs, locale);
+  }
+  auto operator()(typename basic_format_arg<context>::handle) -> iterator {
+    // User-defined types are handled separately because they require access
+    // to the parse context.
+    return out;
+  }
+};
+
+struct width_checker {
+  template <typename T, FMT_ENABLE_IF(is_integer<T>::value)>
+  FMT_CONSTEXPR auto operator()(T value) -> unsigned long long {
+    if (is_negative(value)) report_error("negative width");
+    return static_cast<unsigned long long>(value);
+  }
+
+  template <typename T, FMT_ENABLE_IF(!is_integer<T>::value)>
+  FMT_CONSTEXPR auto operator()(T) -> unsigned long long {
+    report_error("width is not integer");
+    return 0;
+  }
+};
+
+struct precision_checker {
+  template <typename T, FMT_ENABLE_IF(is_integer<T>::value)>
+  FMT_CONSTEXPR auto operator()(T value) -> unsigned long long {
+    if (is_negative(value)) report_error("negative precision");
+    return static_cast<unsigned long long>(value);
+  }
+
+  template <typename T, FMT_ENABLE_IF(!is_integer<T>::value)>
+  FMT_CONSTEXPR auto operator()(T) -> unsigned long long {
+    report_error("precision is not integer");
+    return 0;
+  }
+};
+
+template <typename Handler, typename FormatArg>
+FMT_CONSTEXPR auto get_dynamic_spec(FormatArg arg) -> int {
+  unsigned long long value = arg.visit(Handler());
+  if (value > to_unsigned(max_value<int>())) report_error("number is too big");
+  return static_cast<int>(value);
+}
+
+template <typename Context, typename ID>
+FMT_CONSTEXPR auto get_arg(Context& ctx, ID id) -> decltype(ctx.arg(id)) {
+  auto arg = ctx.arg(id);
+  if (!arg) report_error("argument not found");
+  return arg;
+}
+
+template <typename Handler, typename Context>
+FMT_CONSTEXPR void handle_dynamic_spec(int& value,
+                                       arg_ref<typename Context::char_type> ref,
+                                       Context& ctx) {
+  switch (ref.kind) {
+  case arg_id_kind::none:
+    break;
+  case arg_id_kind::index:
+    value = detail::get_dynamic_spec<Handler>(get_arg(ctx, ref.val.index));
+    break;
+  case arg_id_kind::name:
+    value = detail::get_dynamic_spec<Handler>(get_arg(ctx, ref.val.name));
+    break;
+  }
+}
+
+#if FMT_USE_USER_DEFINED_LITERALS
+#  if FMT_USE_NONTYPE_TEMPLATE_ARGS
+template <typename T, typename Char, size_t N,
+          fmt::detail_exported::fixed_string<Char, N> Str>
+struct statically_named_arg : view {
+  static constexpr auto name = Str.data;
+
+  const T& value;
+  statically_named_arg(const T& v) : value(v) {}
+};
+
+template <typename T, typename Char, size_t N,
+          fmt::detail_exported::fixed_string<Char, N> Str>
+struct is_named_arg<statically_named_arg<T, Char, N, Str>> : std::true_type {};
+
+template <typename T, typename Char, size_t N,
+          fmt::detail_exported::fixed_string<Char, N> Str>
+struct is_statically_named_arg<statically_named_arg<T, Char, N, Str>>
+    : std::true_type {};
+
+template <typename Char, size_t N,
+          fmt::detail_exported::fixed_string<Char, N> Str>
+struct udl_arg {
+  template <typename T> auto operator=(T&& value) const {
+    return statically_named_arg<T, Char, N, Str>(std::forward<T>(value));
+  }
+};
+#  else
+template <typename Char> struct udl_arg {
+  const Char* str;
+
+  template <typename T> auto operator=(T&& value) const -> named_arg<Char, T> {
+    return {str, std::forward<T>(value)};
+  }
+};
+#  endif
+#endif  // FMT_USE_USER_DEFINED_LITERALS
+
+template <typename Locale, typename Char>
+auto vformat(const Locale& loc, basic_string_view<Char> fmt,
+             typename detail::vformat_args<Char>::type args)
+    -> std::basic_string<Char> {
+  auto buf = basic_memory_buffer<Char>();
+  detail::vformat_to(buf, fmt, args, detail::locale_ref(loc));
+  return {buf.data(), buf.size()};
+}
+
+using format_func = void (*)(detail::buffer<char>&, int, const char*);
+
+FMT_API void format_error_code(buffer<char>& out, int error_code,
+                               string_view message) noexcept;
+
+using fmt::report_error;
+FMT_API void report_error(format_func func, int error_code,
+                          const char* message) noexcept;
+}  // namespace detail
+
+FMT_BEGIN_EXPORT
+FMT_API auto vsystem_error(int error_code, string_view format_str,
+                           format_args args) -> std::system_error;
+
+/**
+ * Constructs `std::system_error` with a message formatted with
+ * `fmt::format(fmt, args...)`.
+ * `error_code` is a system error code as given by `errno`.
+ *
+ * **Example**:
+ *
+ *     // This throws std::system_error with the description
+ *     //   cannot open file 'madeup': No such file or directory
+ *     // or similar (system message may vary).
+ *     const char* filename = "madeup";
+ *     std::FILE* file = std::fopen(filename, "r");
+ *     if (!file)
+ *       throw fmt::system_error(errno, "cannot open file '{}'", filename);
+ */
+template <typename... T>
+auto system_error(int error_code, format_string<T...> fmt, T&&... args)
+    -> std::system_error {
+  return vsystem_error(error_code, fmt, fmt::make_format_args(args...));
+}
+
+/**
+ * Formats an error message for an error returned by an operating system or a
+ * language runtime, for example a file opening error, and writes it to `out`.
+ * The format is the same as the one used by `std::system_error(ec, message)`
+ * where `ec` is `std::error_code(error_code, std::generic_category())`.
+ * It is implementation-defined but normally looks like:
+ *
+ *     <message>: <system-message>
+ *
+ * where `<message>` is the passed message and `<system-message>` is the system
+ * message corresponding to the error code.
+ * `error_code` is a system error code as given by `errno`.
+ */
+FMT_API void format_system_error(detail::buffer<char>& out, int error_code,
+                                 const char* message) noexcept;
+
+// Reports a system error without throwing an exception.
+// Can be used to report errors from destructors.
+FMT_API void report_system_error(int error_code, const char* message) noexcept;
+
+/// A fast integer formatter.
+class format_int {
+ private:
+  // Buffer should be large enough to hold all digits (digits10 + 1),
+  // a sign and a null character.
+  enum { buffer_size = std::numeric_limits<unsigned long long>::digits10 + 3 };
+  mutable char buffer_[buffer_size];
+  char* str_;
+
+  template <typename UInt>
+  FMT_CONSTEXPR20 auto format_unsigned(UInt value) -> char* {
+    auto n = static_cast<detail::uint32_or_64_or_128_t<UInt>>(value);
+    return detail::format_decimal(buffer_, n, buffer_size - 1).begin;
+  }
+
+  template <typename Int>
+  FMT_CONSTEXPR20 auto format_signed(Int value) -> char* {
+    auto abs_value = static_cast<detail::uint32_or_64_or_128_t<Int>>(value);
+    bool negative = value < 0;
+    if (negative) abs_value = 0 - abs_value;
+    auto begin = format_unsigned(abs_value);
+    if (negative) *--begin = '-';
+    return begin;
+  }
+
+ public:
+  explicit FMT_CONSTEXPR20 format_int(int value) : str_(format_signed(value)) {}
+  explicit FMT_CONSTEXPR20 format_int(long value)
+      : str_(format_signed(value)) {}
+  explicit FMT_CONSTEXPR20 format_int(long long value)
+      : str_(format_signed(value)) {}
+  explicit FMT_CONSTEXPR20 format_int(unsigned value)
+      : str_(format_unsigned(value)) {}
+  explicit FMT_CONSTEXPR20 format_int(unsigned long value)
+      : str_(format_unsigned(value)) {}
+  explicit FMT_CONSTEXPR20 format_int(unsigned long long value)
+      : str_(format_unsigned(value)) {}
+
+  /// Returns the number of characters written to the output buffer.
+  FMT_CONSTEXPR20 auto size() const -> size_t {
+    return detail::to_unsigned(buffer_ - str_ + buffer_size - 1);
+  }
+
+  /// Returns a pointer to the output buffer content. No terminating null
+  /// character is appended.
+  FMT_CONSTEXPR20 auto data() const -> const char* { return str_; }
+
+  /// Returns a pointer to the output buffer content with terminating null
+  /// character appended.
+  FMT_CONSTEXPR20 auto c_str() const -> const char* {
+    buffer_[buffer_size - 1] = '\0';
+    return str_;
+  }
+
+  /// Returns the content of the output buffer as an `std::string`.
+  auto str() const -> std::string { return std::string(str_, size()); }
+};
+
+template <typename T, typename Char>
+struct formatter<T, Char, enable_if_t<detail::has_format_as<T>::value>>
+    : formatter<detail::format_as_t<T>, Char> {
+  template <typename FormatContext>
+  auto format(const T& value, FormatContext& ctx) const -> decltype(ctx.out()) {
+    auto&& val = format_as(value);  // Make an lvalue reference for format.
+    return formatter<detail::format_as_t<T>, Char>::format(val, ctx);
+  }
+};
+
+#define FMT_FORMAT_AS(Type, Base)                                              \
+  template <typename Char>                                                     \
+  struct formatter<Type, Char> : formatter<Base, Char> {                       \
+    template <typename FormatContext>                                          \
+    auto format(Type value, FormatContext& ctx) const -> decltype(ctx.out()) { \
+      return formatter<Base, Char>::format(value, ctx);                        \
+    }                                                                          \
+  }
+
+FMT_FORMAT_AS(signed char, int);
+FMT_FORMAT_AS(unsigned char, unsigned);
+FMT_FORMAT_AS(short, int);
+FMT_FORMAT_AS(unsigned short, unsigned);
+FMT_FORMAT_AS(long, detail::long_type);
+FMT_FORMAT_AS(unsigned long, detail::ulong_type);
+FMT_FORMAT_AS(Char*, const Char*);
+FMT_FORMAT_AS(std::nullptr_t, const void*);
+FMT_FORMAT_AS(detail::std_string_view<Char>, basic_string_view<Char>);
+FMT_FORMAT_AS(void*, const void*);
+
+template <typename Char, typename Traits, typename Allocator>
+class formatter<std::basic_string<Char, Traits, Allocator>, Char>
+    : public formatter<basic_string_view<Char>, Char> {};
+
+template <typename Char, size_t N>
+struct formatter<Char[N], Char> : formatter<basic_string_view<Char>, Char> {};
+
+/**
+ * Converts `p` to `const void*` for pointer formatting.
+ *
+ * **Example**:
+ *
+ *     auto s = fmt::format("{}", fmt::ptr(p));
+ */
+template <typename T> auto ptr(T p) -> const void* {
+  static_assert(std::is_pointer<T>::value, "");
+  return detail::bit_cast<const void*>(p);
+}
+
+/**
+ * Converts `e` to the underlying type.
+ *
+ * **Example**:
+ *
+ *     enum class color { red, green, blue };
+ *     auto s = fmt::format("{}", fmt::underlying(color::red));
+ */
+template <typename Enum>
+constexpr auto underlying(Enum e) noexcept -> underlying_t<Enum> {
+  return static_cast<underlying_t<Enum>>(e);
+}
+
+namespace enums {
+template <typename Enum, FMT_ENABLE_IF(std::is_enum<Enum>::value)>
+constexpr auto format_as(Enum e) noexcept -> underlying_t<Enum> {
+  return static_cast<underlying_t<Enum>>(e);
+}
+}  // namespace enums
+
+class bytes {
+ private:
+  string_view data_;
+  friend struct formatter<bytes>;
+
+ public:
+  explicit bytes(string_view data) : data_(data) {}
+};
+
+template <> struct formatter<bytes> {
+ private:
+  detail::dynamic_format_specs<> specs_;
+
+ public:
+  template <typename ParseContext>
+  FMT_CONSTEXPR auto parse(ParseContext& ctx) -> const char* {
+    return parse_format_specs(ctx.begin(), ctx.end(), specs_, ctx,
+                              detail::type::string_type);
+  }
+
+  template <typename FormatContext>
+  auto format(bytes b, FormatContext& ctx) const -> decltype(ctx.out()) {
+    auto specs = specs_;
+    detail::handle_dynamic_spec<detail::width_checker>(specs.width,
+                                                       specs.width_ref, ctx);
+    detail::handle_dynamic_spec<detail::precision_checker>(
+        specs.precision, specs.precision_ref, ctx);
+    return detail::write_bytes<char>(ctx.out(), b.data_, specs);
+  }
+};
+
+// group_digits_view is not derived from view because it copies the argument.
+template <typename T> struct group_digits_view {
+  T value;
+};
+
+/**
+ * Returns a view that formats an integer value using ',' as a
+ * locale-independent thousands separator.
+ *
+ * **Example**:
+ *
+ *     fmt::print("{}", fmt::group_digits(12345));
+ *     // Output: "12,345"
+ */
+template <typename T> auto group_digits(T value) -> group_digits_view<T> {
+  return {value};
+}
+
+template <typename T> struct formatter<group_digits_view<T>> : formatter<T> {
+ private:
+  detail::dynamic_format_specs<> specs_;
+
+ public:
+  template <typename ParseContext>
+  FMT_CONSTEXPR auto parse(ParseContext& ctx) -> const char* {
+    return parse_format_specs(ctx.begin(), ctx.end(), specs_, ctx,
+                              detail::type::int_type);
+  }
+
+  template <typename FormatContext>
+  auto format(group_digits_view<T> t, FormatContext& ctx) const
+      -> decltype(ctx.out()) {
+    auto specs = specs_;
+    detail::handle_dynamic_spec<detail::width_checker>(specs.width,
+                                                       specs.width_ref, ctx);
+    detail::handle_dynamic_spec<detail::precision_checker>(
+        specs.precision, specs.precision_ref, ctx);
+    auto arg = detail::make_write_int_arg(t.value, specs.sign);
+    return detail::write_int(
+        ctx.out(), static_cast<detail::uint64_or_128_t<T>>(arg.abs_value),
+        arg.prefix, specs, detail::digit_grouping<char>("\3", ","));
+  }
+};
+
+template <typename T, typename Char> struct nested_view {
+  const formatter<T, Char>* fmt;
+  const T* value;
+};
+
+template <typename T, typename Char>
+struct formatter<nested_view<T, Char>, Char> {
+  template <typename ParseContext>
+  FMT_CONSTEXPR auto parse(ParseContext& ctx) -> decltype(ctx.begin()) {
+    return ctx.begin();
+  }
+  template <typename FormatContext>
+  auto format(nested_view<T, Char> view, FormatContext& ctx) const
+      -> decltype(ctx.out()) {
+    return view.fmt->format(*view.value, ctx);
+  }
+};
+
+template <typename T, typename Char = char> struct nested_formatter {
+ private:
+  int width_;
+  detail::fill_t fill_;
+  align_t align_ : 4;
+  formatter<T, Char> formatter_;
+
+ public:
+  constexpr nested_formatter() : width_(0), align_(align_t::none) {}
+
+  FMT_CONSTEXPR auto parse(basic_format_parse_context<Char>& ctx)
+      -> decltype(ctx.begin()) {
+    auto specs = detail::dynamic_format_specs<Char>();
+    auto it = parse_format_specs(ctx.begin(), ctx.end(), specs, ctx,
+                                 detail::type::none_type);
+    width_ = specs.width;
+    fill_ = specs.fill;
+    align_ = specs.align;
+    ctx.advance_to(it);
+    return formatter_.parse(ctx);
+  }
+
+  template <typename FormatContext, typename F>
+  auto write_padded(FormatContext& ctx, F write) const -> decltype(ctx.out()) {
+    if (width_ == 0) return write(ctx.out());
+    auto buf = basic_memory_buffer<Char>();
+    write(basic_appender<Char>(buf));
+    auto specs = format_specs();
+    specs.width = width_;
+    specs.fill = fill_;
+    specs.align = align_;
+    return detail::write<Char>(
+        ctx.out(), basic_string_view<Char>(buf.data(), buf.size()), specs);
+  }
+
+  auto nested(const T& value) const -> nested_view<T, Char> {
+    return nested_view<T, Char>{&formatter_, &value};
+  }
+};
+
+/**
+ * Converts `value` to `std::string` using the default format for type `T`.
+ *
+ * **Example**:
+ *
+ *     std::string answer = fmt::to_string(42);
+ */
+template <typename T, FMT_ENABLE_IF(!std::is_integral<T>::value &&
+                                    !detail::has_format_as<T>::value)>
+inline auto to_string(const T& value) -> std::string {
+  auto buffer = memory_buffer();
+  detail::write<char>(appender(buffer), value);
+  return {buffer.data(), buffer.size()};
+}
+
+template <typename T, FMT_ENABLE_IF(std::is_integral<T>::value)>
+FMT_NODISCARD inline auto to_string(T value) -> std::string {
+  // The buffer should be large enough to store the number including the sign
+  // or "false" for bool.
+  constexpr int max_size = detail::digits10<T>() + 2;
+  char buffer[max_size > 5 ? static_cast<unsigned>(max_size) : 5];
+  char* begin = buffer;
+  return std::string(begin, detail::write<char>(begin, value));
+}
+
+template <typename Char, size_t SIZE>
+FMT_NODISCARD auto to_string(const basic_memory_buffer<Char, SIZE>& buf)
+    -> std::basic_string<Char> {
+  auto size = buf.size();
+  detail::assume(size < std::basic_string<Char>().max_size());
+  return std::basic_string<Char>(buf.data(), size);
+}
+
+template <typename T, FMT_ENABLE_IF(!std::is_integral<T>::value &&
+                                    detail::has_format_as<T>::value)>
+inline auto to_string(const T& value) -> std::string {
+  return to_string(format_as(value));
+}
+
+FMT_END_EXPORT
+
+namespace detail {
+
+template <typename Char>
+void vformat_to(buffer<Char>& buf, basic_string_view<Char> fmt,
+                typename vformat_args<Char>::type args, locale_ref loc) {
+  auto out = basic_appender<Char>(buf);
+  if (fmt.size() == 2 && equal2(fmt.data(), "{}")) {
+    auto arg = args.get(0);
+    if (!arg) report_error("argument not found");
+    arg.visit(default_arg_formatter<Char>{out, args, loc});
+    return;
+  }
+
+  struct format_handler {
+    basic_format_parse_context<Char> parse_context;
+    buffered_context<Char> context;
+
+    format_handler(basic_appender<Char> p_out, basic_string_view<Char> str,
+                   basic_format_args<buffered_context<Char>> p_args,
+                   locale_ref p_loc)
+        : parse_context(str), context(p_out, p_args, p_loc) {}
+
+    void on_text(const Char* begin, const Char* end) {
+      auto text = basic_string_view<Char>(begin, to_unsigned(end - begin));
+      context.advance_to(write<Char>(context.out(), text));
+    }
+
+    FMT_CONSTEXPR auto on_arg_id() -> int {
+      return parse_context.next_arg_id();
+    }
+    FMT_CONSTEXPR auto on_arg_id(int id) -> int {
+      parse_context.check_arg_id(id);
+      return id;
+    }
+    FMT_CONSTEXPR auto on_arg_id(basic_string_view<Char> id) -> int {
+      parse_context.check_arg_id(id);
+      int arg_id = context.arg_id(id);
+      if (arg_id < 0) report_error("argument not found");
+      return arg_id;
+    }
+
+    FMT_INLINE void on_replacement_field(int id, const Char*) {
+      auto arg = get_arg(context, id);
+      context.advance_to(arg.visit(default_arg_formatter<Char>{
+          context.out(), context.args(), context.locale()}));
+    }
+
+    auto on_format_specs(int id, const Char* begin, const Char* end)
+        -> const Char* {
+      auto arg = get_arg(context, id);
+      // Not using a visitor for custom types gives better codegen.
+      if (arg.format_custom(begin, parse_context, context))
+        return parse_context.begin();
+      auto specs = detail::dynamic_format_specs<Char>();
+      begin = parse_format_specs(begin, end, specs, parse_context, arg.type());
+      detail::handle_dynamic_spec<detail::width_checker>(
+          specs.width, specs.width_ref, context);
+      detail::handle_dynamic_spec<detail::precision_checker>(
+          specs.precision, specs.precision_ref, context);
+      if (begin == end || *begin != '}')
+        report_error("missing '}' in format string");
+      context.advance_to(arg.visit(
+          arg_formatter<Char>{context.out(), specs, context.locale()}));
+      return begin;
+    }
+
+    FMT_NORETURN void on_error(const char* message) { report_error(message); }
+  };
+  detail::parse_format_string<false>(fmt, format_handler(out, fmt, args, loc));
+}
+
+FMT_BEGIN_EXPORT
+
+#ifndef FMT_HEADER_ONLY
+extern template FMT_API void vformat_to(buffer<char>&, string_view,
+                                        typename vformat_args<>::type,
+                                        locale_ref);
+extern template FMT_API auto thousands_sep_impl<char>(locale_ref)
+    -> thousands_sep_result<char>;
+extern template FMT_API auto thousands_sep_impl<wchar_t>(locale_ref)
+    -> thousands_sep_result<wchar_t>;
+extern template FMT_API auto decimal_point_impl(locale_ref) -> char;
+extern template FMT_API auto decimal_point_impl(locale_ref) -> wchar_t;
+#endif  // FMT_HEADER_ONLY
+
+FMT_END_EXPORT
+
+template <typename T, typename Char, type TYPE>
+template <typename FormatContext>
+FMT_CONSTEXPR FMT_INLINE auto native_formatter<T, Char, TYPE>::format(
+    const T& val, FormatContext& ctx) const -> decltype(ctx.out()) {
+  if (specs_.width_ref.kind == arg_id_kind::none &&
+      specs_.precision_ref.kind == arg_id_kind::none) {
+    return write<Char>(ctx.out(), val, specs_, ctx.locale());
+  }
+  auto specs = specs_;
+  handle_dynamic_spec<width_checker>(specs.width, specs.width_ref, ctx);
+  handle_dynamic_spec<precision_checker>(specs.precision, specs.precision_ref,
+                                         ctx);
+  return write<Char>(ctx.out(), val, specs, ctx.locale());
+}
+
+}  // namespace detail
+
+FMT_BEGIN_EXPORT
+
+template <typename Char>
+struct formatter<detail::float128, Char>
+    : detail::native_formatter<detail::float128, Char,
+                               detail::type::float_type> {};
+
+#if FMT_USE_USER_DEFINED_LITERALS
+inline namespace literals {
+/**
+ * User-defined literal equivalent of `fmt::arg`.
+ *
+ * **Example**:
+ *
+ *     using namespace fmt::literals;
+ *     fmt::print("The answer is {answer}.", "answer"_a=42);
+ */
+#  if FMT_USE_NONTYPE_TEMPLATE_ARGS
+template <detail_exported::fixed_string Str> constexpr auto operator""_a() {
+  using char_t = remove_cvref_t<decltype(Str.data[0])>;
+  return detail::udl_arg<char_t, sizeof(Str.data) / sizeof(char_t), Str>();
+}
+#  else
+constexpr auto operator""_a(const char* s, size_t) -> detail::udl_arg<char> {
+  return {s};
+}
+#  endif
+}  // namespace literals
+#endif  // FMT_USE_USER_DEFINED_LITERALS
+
+FMT_API auto vformat(string_view fmt, format_args args) -> std::string;
+
+/**
+ * Formats `args` according to specifications in `fmt` and returns the result
+ * as a string.
+ *
+ * **Example**:
+ *
+ *     #include <fmt/format.h>
+ *     std::string message = fmt::format("The answer is {}.", 42);
+ */
+template <typename... T>
+FMT_NODISCARD FMT_INLINE auto format(format_string<T...> fmt, T&&... args)
+    -> std::string {
+  return vformat(fmt, fmt::make_format_args(args...));
+}
+
+template <typename Locale, FMT_ENABLE_IF(detail::is_locale<Locale>::value)>
+inline auto vformat(const Locale& loc, string_view fmt, format_args args)
+    -> std::string {
+  return detail::vformat(loc, fmt, args);
+}
+
+template <typename Locale, typename... T,
+          FMT_ENABLE_IF(detail::is_locale<Locale>::value)>
+inline auto format(const Locale& loc, format_string<T...> fmt, T&&... args)
+    -> std::string {
+  return fmt::vformat(loc, string_view(fmt), fmt::make_format_args(args...));
+}
+
+template <typename OutputIt, typename Locale,
+          FMT_ENABLE_IF(detail::is_output_iterator<OutputIt, char>::value&&
+                            detail::is_locale<Locale>::value)>
+auto vformat_to(OutputIt out, const Locale& loc, string_view fmt,
+                format_args args) -> OutputIt {
+  using detail::get_buffer;
+  auto&& buf = get_buffer<char>(out);
+  detail::vformat_to(buf, fmt, args, detail::locale_ref(loc));
+  return detail::get_iterator(buf, out);
+}
+
+template <typename OutputIt, typename Locale, typename... T,
+          FMT_ENABLE_IF(detail::is_output_iterator<OutputIt, char>::value&&
+                            detail::is_locale<Locale>::value)>
+FMT_INLINE auto format_to(OutputIt out, const Locale& loc,
+                          format_string<T...> fmt, T&&... args) -> OutputIt {
+  return vformat_to(out, loc, fmt, fmt::make_format_args(args...));
+}
+
+template <typename Locale, typename... T,
+          FMT_ENABLE_IF(detail::is_locale<Locale>::value)>
+FMT_NODISCARD FMT_INLINE auto formatted_size(const Locale& loc,
+                                             format_string<T...> fmt,
+                                             T&&... args) -> size_t {
+  auto buf = detail::counting_buffer<>();
+  detail::vformat_to<char>(buf, fmt, fmt::make_format_args(args...),
+                           detail::locale_ref(loc));
+  return buf.count();
+}
+
+FMT_END_EXPORT
+
+FMT_END_NAMESPACE
+
+#ifdef FMT_HEADER_ONLY
+#  define FMT_FUNC inline
+#  include "format-inl.h"
+#else
+#  define FMT_FUNC
+#endif
+
+// Restore _LIBCPP_REMOVE_TRANSITIVE_INCLUDES.
+#ifdef FMT_REMOVE_TRANSITIVE_INCLUDES
+#  undef _LIBCPP_REMOVE_TRANSITIVE_INCLUDES
+#endif
+
+#endif  // FMT_FORMAT_H_
diff --git a/ext/spdlog/fmt/bundled/locale.h b/ext/spdlog/fmt/bundled/locale.h
new file mode 100644
index 0000000..7571b52
--- /dev/null
+++ b/ext/spdlog/fmt/bundled/locale.h
@@ -0,0 +1,2 @@
+#include "xchar.h"
+#warning fmt/locale.h is deprecated, include fmt/format.h or fmt/xchar.h instead
diff --git a/ext/spdlog/fmt/bundled/os.h b/ext/spdlog/fmt/bundled/os.h
new file mode 100644
index 0000000..5c85ea0
--- /dev/null
+++ b/ext/spdlog/fmt/bundled/os.h
@@ -0,0 +1,439 @@
+// Formatting library for C++ - optional OS-specific functionality
+//
+// Copyright (c) 2012 - present, Victor Zverovich
+// All rights reserved.
+//
+// For the license information refer to format.h.
+
+#ifndef FMT_OS_H_
+#define FMT_OS_H_
+
+#include "format.h"
+
+#ifndef FMT_MODULE
+#  include <cerrno>
+#  include <cstddef>
+#  include <cstdio>
+#  include <system_error>  // std::system_error
+
+#  if FMT_HAS_INCLUDE(<xlocale.h>)
+#    include <xlocale.h>  // LC_NUMERIC_MASK on macOS
+#  endif
+#endif  // FMT_MODULE
+
+#ifndef FMT_USE_FCNTL
+// UWP doesn't provide _pipe.
+#  if FMT_HAS_INCLUDE("winapifamily.h")
+#    include <winapifamily.h>
+#  endif
+#  if (FMT_HAS_INCLUDE(<fcntl.h>) || defined(__APPLE__) || \
+       defined(__linux__)) &&                              \
+      (!defined(WINAPI_FAMILY) ||                          \
+       (WINAPI_FAMILY == WINAPI_FAMILY_DESKTOP_APP))
+#    include <fcntl.h>  // for O_RDONLY
+#    define FMT_USE_FCNTL 1
+#  else
+#    define FMT_USE_FCNTL 0
+#  endif
+#endif
+
+#ifndef FMT_POSIX
+#  if defined(_WIN32) && !defined(__MINGW32__)
+// Fix warnings about deprecated symbols.
+#    define FMT_POSIX(call) _##call
+#  else
+#    define FMT_POSIX(call) call
+#  endif
+#endif
+
+// Calls to system functions are wrapped in FMT_SYSTEM for testability.
+#ifdef FMT_SYSTEM
+#  define FMT_HAS_SYSTEM
+#  define FMT_POSIX_CALL(call) FMT_SYSTEM(call)
+#else
+#  define FMT_SYSTEM(call) ::call
+#  ifdef _WIN32
+// Fix warnings about deprecated symbols.
+#    define FMT_POSIX_CALL(call) ::_##call
+#  else
+#    define FMT_POSIX_CALL(call) ::call
+#  endif
+#endif
+
+// Retries the expression while it evaluates to error_result and errno
+// equals to EINTR.
+#ifndef _WIN32
+#  define FMT_RETRY_VAL(result, expression, error_result) \
+    do {                                                  \
+      (result) = (expression);                            \
+    } while ((result) == (error_result) && errno == EINTR)
+#else
+#  define FMT_RETRY_VAL(result, expression, error_result) result = (expression)
+#endif
+
+#define FMT_RETRY(result, expression) FMT_RETRY_VAL(result, expression, -1)
+
+FMT_BEGIN_NAMESPACE
+FMT_BEGIN_EXPORT
+
+/**
+ * A reference to a null-terminated string. It can be constructed from a C
+ * string or `std::string`.
+ *
+ * You can use one of the following type aliases for common character types:
+ *
+ * +---------------+-----------------------------+
+ * | Type          | Definition                  |
+ * +===============+=============================+
+ * | cstring_view  | basic_cstring_view<char>    |
+ * +---------------+-----------------------------+
+ * | wcstring_view | basic_cstring_view<wchar_t> |
+ * +---------------+-----------------------------+
+ *
+ * This class is most useful as a parameter type for functions that wrap C APIs.
+ */
+template <typename Char> class basic_cstring_view {
+ private:
+  const Char* data_;
+
+ public:
+  /// Constructs a string reference object from a C string.
+  basic_cstring_view(const Char* s) : data_(s) {}
+
+  /// Constructs a string reference from an `std::string` object.
+  basic_cstring_view(const std::basic_string<Char>& s) : data_(s.c_str()) {}
+
+  /// Returns the pointer to a C string.
+  auto c_str() const -> const Char* { return data_; }
+};
+
+using cstring_view = basic_cstring_view<char>;
+using wcstring_view = basic_cstring_view<wchar_t>;
+
+#ifdef _WIN32
+FMT_API const std::error_category& system_category() noexcept;
+
+namespace detail {
+FMT_API void format_windows_error(buffer<char>& out, int error_code,
+                                  const char* message) noexcept;
+}
+
+FMT_API std::system_error vwindows_error(int error_code, string_view format_str,
+                                         format_args args);
+
+/**
+ * Constructs a `std::system_error` object with the description of the form
+ *
+ *     <message>: <system-message>
+ *
+ * where `<message>` is the formatted message and `<system-message>` is the
+ * system message corresponding to the error code.
+ * `error_code` is a Windows error code as given by `GetLastError`.
+ * If `error_code` is not a valid error code such as -1, the system message
+ * will look like "error -1".
+ *
+ * **Example**:
+ *
+ *     // This throws a system_error with the description
+ *     //   cannot open file 'madeup': The system cannot find the file
+ * specified.
+ *     // or similar (system message may vary).
+ *     const char *filename = "madeup";
+ *     LPOFSTRUCT of = LPOFSTRUCT();
+ *     HFILE file = OpenFile(filename, &of, OF_READ);
+ *     if (file == HFILE_ERROR) {
+ *       throw fmt::windows_error(GetLastError(),
+ *                                "cannot open file '{}'", filename);
+ *     }
+ */
+template <typename... Args>
+std::system_error windows_error(int error_code, string_view message,
+                                const Args&... args) {
+  return vwindows_error(error_code, message, fmt::make_format_args(args...));
+}
+
+// Reports a Windows error without throwing an exception.
+// Can be used to report errors from destructors.
+FMT_API void report_windows_error(int error_code, const char* message) noexcept;
+#else
+inline auto system_category() noexcept -> const std::error_category& {
+  return std::system_category();
+}
+#endif  // _WIN32
+
+// std::system is not available on some platforms such as iOS (#2248).
+#ifdef __OSX__
+template <typename S, typename... Args, typename Char = char_t<S>>
+void say(const S& format_str, Args&&... args) {
+  std::system(format("say \"{}\"", format(format_str, args...)).c_str());
+}
+#endif
+
+// A buffered file.
+class buffered_file {
+ private:
+  FILE* file_;
+
+  friend class file;
+
+  explicit buffered_file(FILE* f) : file_(f) {}
+
+ public:
+  buffered_file(const buffered_file&) = delete;
+  void operator=(const buffered_file&) = delete;
+
+  // Constructs a buffered_file object which doesn't represent any file.
+  buffered_file() noexcept : file_(nullptr) {}
+
+  // Destroys the object closing the file it represents if any.
+  FMT_API ~buffered_file() noexcept;
+
+ public:
+  buffered_file(buffered_file&& other) noexcept : file_(other.file_) {
+    other.file_ = nullptr;
+  }
+
+  auto operator=(buffered_file&& other) -> buffered_file& {
+    close();
+    file_ = other.file_;
+    other.file_ = nullptr;
+    return *this;
+  }
+
+  // Opens a file.
+  FMT_API buffered_file(cstring_view filename, cstring_view mode);
+
+  // Closes the file.
+  FMT_API void close();
+
+  // Returns the pointer to a FILE object representing this file.
+  auto get() const noexcept -> FILE* { return file_; }
+
+  FMT_API auto descriptor() const -> int;
+
+  template <typename... T>
+  inline void print(string_view fmt, const T&... args) {
+    const auto& vargs = fmt::make_format_args(args...);
+    detail::is_locking<T...>() ? fmt::vprint_buffered(file_, fmt, vargs)
+                               : fmt::vprint(file_, fmt, vargs);
+  }
+};
+
+#if FMT_USE_FCNTL
+
+// A file. Closed file is represented by a file object with descriptor -1.
+// Methods that are not declared with noexcept may throw
+// fmt::system_error in case of failure. Note that some errors such as
+// closing the file multiple times will cause a crash on Windows rather
+// than an exception. You can get standard behavior by overriding the
+// invalid parameter handler with _set_invalid_parameter_handler.
+class FMT_API file {
+ private:
+  int fd_;  // File descriptor.
+
+  // Constructs a file object with a given descriptor.
+  explicit file(int fd) : fd_(fd) {}
+
+  friend struct pipe;
+
+ public:
+  // Possible values for the oflag argument to the constructor.
+  enum {
+    RDONLY = FMT_POSIX(O_RDONLY),  // Open for reading only.
+    WRONLY = FMT_POSIX(O_WRONLY),  // Open for writing only.
+    RDWR = FMT_POSIX(O_RDWR),      // Open for reading and writing.
+    CREATE = FMT_POSIX(O_CREAT),   // Create if the file doesn't exist.
+    APPEND = FMT_POSIX(O_APPEND),  // Open in append mode.
+    TRUNC = FMT_POSIX(O_TRUNC)     // Truncate the content of the file.
+  };
+
+  // Constructs a file object which doesn't represent any file.
+  file() noexcept : fd_(-1) {}
+
+  // Opens a file and constructs a file object representing this file.
+  file(cstring_view path, int oflag);
+
+ public:
+  file(const file&) = delete;
+  void operator=(const file&) = delete;
+
+  file(file&& other) noexcept : fd_(other.fd_) { other.fd_ = -1; }
+
+  // Move assignment is not noexcept because close may throw.
+  auto operator=(file&& other) -> file& {
+    close();
+    fd_ = other.fd_;
+    other.fd_ = -1;
+    return *this;
+  }
+
+  // Destroys the object closing the file it represents if any.
+  ~file() noexcept;
+
+  // Returns the file descriptor.
+  auto descriptor() const noexcept -> int { return fd_; }
+
+  // Closes the file.
+  void close();
+
+  // Returns the file size. The size has signed type for consistency with
+  // stat::st_size.
+  auto size() const -> long long;
+
+  // Attempts to read count bytes from the file into the specified buffer.
+  auto read(void* buffer, size_t count) -> size_t;
+
+  // Attempts to write count bytes from the specified buffer to the file.
+  auto write(const void* buffer, size_t count) -> size_t;
+
+  // Duplicates a file descriptor with the dup function and returns
+  // the duplicate as a file object.
+  static auto dup(int fd) -> file;
+
+  // Makes fd be the copy of this file descriptor, closing fd first if
+  // necessary.
+  void dup2(int fd);
+
+  // Makes fd be the copy of this file descriptor, closing fd first if
+  // necessary.
+  void dup2(int fd, std::error_code& ec) noexcept;
+
+  // Creates a buffered_file object associated with this file and detaches
+  // this file object from the file.
+  auto fdopen(const char* mode) -> buffered_file;
+
+#  if defined(_WIN32) && !defined(__MINGW32__)
+  // Opens a file and constructs a file object representing this file by
+  // wcstring_view filename. Windows only.
+  static file open_windows_file(wcstring_view path, int oflag);
+#  endif
+};
+
+struct FMT_API pipe {
+  file read_end;
+  file write_end;
+
+  // Creates a pipe setting up read_end and write_end file objects for reading
+  // and writing respectively.
+  pipe();
+};
+
+// Returns the memory page size.
+auto getpagesize() -> long;
+
+namespace detail {
+
+struct buffer_size {
+  buffer_size() = default;
+  size_t value = 0;
+  auto operator=(size_t val) const -> buffer_size {
+    auto bs = buffer_size();
+    bs.value = val;
+    return bs;
+  }
+};
+
+struct ostream_params {
+  int oflag = file::WRONLY | file::CREATE | file::TRUNC;
+  size_t buffer_size = BUFSIZ > 32768 ? BUFSIZ : 32768;
+
+  ostream_params() {}
+
+  template <typename... T>
+  ostream_params(T... params, int new_oflag) : ostream_params(params...) {
+    oflag = new_oflag;
+  }
+
+  template <typename... T>
+  ostream_params(T... params, detail::buffer_size bs)
+      : ostream_params(params...) {
+    this->buffer_size = bs.value;
+  }
+
+// Intel has a bug that results in failure to deduce a constructor
+// for empty parameter packs.
+#  if defined(__INTEL_COMPILER) && __INTEL_COMPILER < 2000
+  ostream_params(int new_oflag) : oflag(new_oflag) {}
+  ostream_params(detail::buffer_size bs) : buffer_size(bs.value) {}
+#  endif
+};
+
+class file_buffer final : public buffer<char> {
+ private:
+  file file_;
+
+  FMT_API static void grow(buffer<char>& buf, size_t);
+
+ public:
+  FMT_API file_buffer(cstring_view path, const ostream_params& params);
+  FMT_API file_buffer(file_buffer&& other) noexcept;
+  FMT_API ~file_buffer();
+
+  void flush() {
+    if (size() == 0) return;
+    file_.write(data(), size() * sizeof(data()[0]));
+    clear();
+  }
+
+  void close() {
+    flush();
+    file_.close();
+  }
+};
+
+}  // namespace detail
+
+constexpr auto buffer_size = detail::buffer_size();
+
+/// A fast output stream for writing from a single thread. Writing from
+/// multiple threads without external synchronization may result in a data race.
+class FMT_API ostream {
+ private:
+  FMT_MSC_WARNING(suppress : 4251)
+  detail::file_buffer buffer_;
+
+  ostream(cstring_view path, const detail::ostream_params& params)
+      : buffer_(path, params) {}
+
+ public:
+  ostream(ostream&& other) : buffer_(std::move(other.buffer_)) {}
+
+  ~ostream();
+
+  void flush() { buffer_.flush(); }
+
+  template <typename... T>
+  friend auto output_file(cstring_view path, T... params) -> ostream;
+
+  void close() { buffer_.close(); }
+
+  /// Formats `args` according to specifications in `fmt` and writes the
+  /// output to the file.
+  template <typename... T> void print(format_string<T...> fmt, T&&... args) {
+    vformat_to(appender(buffer_), fmt, fmt::make_format_args(args...));
+  }
+};
+
+/**
+ * Opens a file for writing. Supported parameters passed in `params`:
+ *
+ * - `<integer>`: Flags passed to [open](
+ *   https://pubs.opengroup.org/onlinepubs/007904875/functions/open.html)
+ *   (`file::WRONLY | file::CREATE | file::TRUNC` by default)
+ * - `buffer_size=<integer>`: Output buffer size
+ *
+ * **Example**:
+ *
+ *     auto out = fmt::output_file("guide.txt");
+ *     out.print("Don't {}", "Panic");
+ */
+template <typename... T>
+inline auto output_file(cstring_view path, T... params) -> ostream {
+  return {path, detail::ostream_params(params...)};
+}
+#endif  // FMT_USE_FCNTL
+
+FMT_END_EXPORT
+FMT_END_NAMESPACE
+
+#endif  // FMT_OS_H_
diff --git a/ext/spdlog/fmt/bundled/ostream.h b/ext/spdlog/fmt/bundled/ostream.h
new file mode 100644
index 0000000..98faef6
--- /dev/null
+++ b/ext/spdlog/fmt/bundled/ostream.h
@@ -0,0 +1,211 @@
+// Formatting library for C++ - std::ostream support
+//
+// Copyright (c) 2012 - present, Victor Zverovich
+// All rights reserved.
+//
+// For the license information refer to format.h.
+
+#ifndef FMT_OSTREAM_H_
+#define FMT_OSTREAM_H_
+
+#ifndef FMT_MODULE
+#  include <fstream>  // std::filebuf
+#endif
+
+#ifdef _WIN32
+#  ifdef __GLIBCXX__
+#    include <ext/stdio_filebuf.h>
+#    include <ext/stdio_sync_filebuf.h>
+#  endif
+#  include <io.h>
+#endif
+
+#include "chrono.h"  // formatbuf
+
+FMT_BEGIN_NAMESPACE
+namespace detail {
+
+// Generate a unique explicit instantion in every translation unit using a tag
+// type in an anonymous namespace.
+namespace {
+struct file_access_tag {};
+}  // namespace
+template <typename Tag, typename BufType, FILE* BufType::*FileMemberPtr>
+class file_access {
+  friend auto get_file(BufType& obj) -> FILE* { return obj.*FileMemberPtr; }
+};
+
+#if FMT_MSC_VERSION
+template class file_access<file_access_tag, std::filebuf,
+                           &std::filebuf::_Myfile>;
+auto get_file(std::filebuf&) -> FILE*;
+#endif
+
+inline auto write_ostream_unicode(std::ostream& os, fmt::string_view data)
+    -> bool {
+  FILE* f = nullptr;
+#if FMT_MSC_VERSION && FMT_USE_RTTI
+  if (auto* buf = dynamic_cast<std::filebuf*>(os.rdbuf()))
+    f = get_file(*buf);
+  else
+    return false;
+#elif defined(_WIN32) && defined(__GLIBCXX__) && FMT_USE_RTTI
+  auto* rdbuf = os.rdbuf();
+  if (auto* sfbuf = dynamic_cast<__gnu_cxx::stdio_sync_filebuf<char>*>(rdbuf))
+    f = sfbuf->file();
+  else if (auto* fbuf = dynamic_cast<__gnu_cxx::stdio_filebuf<char>*>(rdbuf))
+    f = fbuf->file();
+  else
+    return false;
+#else
+  ignore_unused(os, data, f);
+#endif
+#ifdef _WIN32
+  if (f) {
+    int fd = _fileno(f);
+    if (_isatty(fd)) {
+      os.flush();
+      return write_console(fd, data);
+    }
+  }
+#endif
+  return false;
+}
+inline auto write_ostream_unicode(std::wostream&,
+                                  fmt::basic_string_view<wchar_t>) -> bool {
+  return false;
+}
+
+// Write the content of buf to os.
+// It is a separate function rather than a part of vprint to simplify testing.
+template <typename Char>
+void write_buffer(std::basic_ostream<Char>& os, buffer<Char>& buf) {
+  const Char* buf_data = buf.data();
+  using unsigned_streamsize = std::make_unsigned<std::streamsize>::type;
+  unsigned_streamsize size = buf.size();
+  unsigned_streamsize max_size = to_unsigned(max_value<std::streamsize>());
+  do {
+    unsigned_streamsize n = size <= max_size ? size : max_size;
+    os.write(buf_data, static_cast<std::streamsize>(n));
+    buf_data += n;
+    size -= n;
+  } while (size != 0);
+}
+
+template <typename Char, typename T>
+void format_value(buffer<Char>& buf, const T& value) {
+  auto&& format_buf = formatbuf<std::basic_streambuf<Char>>(buf);
+  auto&& output = std::basic_ostream<Char>(&format_buf);
+#if !defined(FMT_STATIC_THOUSANDS_SEPARATOR)
+  output.imbue(std::locale::classic());  // The default is always unlocalized.
+#endif
+  output << value;
+  output.exceptions(std::ios_base::failbit | std::ios_base::badbit);
+}
+
+template <typename T> struct streamed_view {
+  const T& value;
+};
+
+}  // namespace detail
+
+// Formats an object of type T that has an overloaded ostream operator<<.
+template <typename Char>
+struct basic_ostream_formatter : formatter<basic_string_view<Char>, Char> {
+  void set_debug_format() = delete;
+
+  template <typename T, typename Context>
+  auto format(const T& value, Context& ctx) const -> decltype(ctx.out()) {
+    auto buffer = basic_memory_buffer<Char>();
+    detail::format_value(buffer, value);
+    return formatter<basic_string_view<Char>, Char>::format(
+        {buffer.data(), buffer.size()}, ctx);
+  }
+};
+
+using ostream_formatter = basic_ostream_formatter<char>;
+
+template <typename T, typename Char>
+struct formatter<detail::streamed_view<T>, Char>
+    : basic_ostream_formatter<Char> {
+  template <typename Context>
+  auto format(detail::streamed_view<T> view, Context& ctx) const
+      -> decltype(ctx.out()) {
+    return basic_ostream_formatter<Char>::format(view.value, ctx);
+  }
+};
+
+/**
+ * Returns a view that formats `value` via an ostream `operator<<`.
+ *
+ * **Example**:
+ *
+ *     fmt::print("Current thread id: {}\n",
+ *                fmt::streamed(std::this_thread::get_id()));
+ */
+template <typename T>
+constexpr auto streamed(const T& value) -> detail::streamed_view<T> {
+  return {value};
+}
+
+namespace detail {
+
+inline void vprint_directly(std::ostream& os, string_view format_str,
+                            format_args args) {
+  auto buffer = memory_buffer();
+  detail::vformat_to(buffer, format_str, args);
+  detail::write_buffer(os, buffer);
+}
+
+}  // namespace detail
+
+FMT_EXPORT template <typename Char>
+void vprint(std::basic_ostream<Char>& os,
+            basic_string_view<type_identity_t<Char>> format_str,
+            typename detail::vformat_args<Char>::type args) {
+  auto buffer = basic_memory_buffer<Char>();
+  detail::vformat_to(buffer, format_str, args);
+  if (detail::write_ostream_unicode(os, {buffer.data(), buffer.size()})) return;
+  detail::write_buffer(os, buffer);
+}
+
+/**
+ * Prints formatted data to the stream `os`.
+ *
+ * **Example**:
+ *
+ *     fmt::print(cerr, "Don't {}!", "panic");
+ */
+FMT_EXPORT template <typename... T>
+void print(std::ostream& os, format_string<T...> fmt, T&&... args) {
+  const auto& vargs = fmt::make_format_args(args...);
+  if (detail::use_utf8())
+    vprint(os, fmt, vargs);
+  else
+    detail::vprint_directly(os, fmt, vargs);
+}
+
+FMT_EXPORT
+template <typename... Args>
+void print(std::wostream& os,
+           basic_format_string<wchar_t, type_identity_t<Args>...> fmt,
+           Args&&... args) {
+  vprint(os, fmt, fmt::make_format_args<buffered_context<wchar_t>>(args...));
+}
+
+FMT_EXPORT template <typename... T>
+void println(std::ostream& os, format_string<T...> fmt, T&&... args) {
+  fmt::print(os, "{}\n", fmt::format(fmt, std::forward<T>(args)...));
+}
+
+FMT_EXPORT
+template <typename... Args>
+void println(std::wostream& os,
+             basic_format_string<wchar_t, type_identity_t<Args>...> fmt,
+             Args&&... args) {
+  print(os, L"{}\n", fmt::format(fmt, std::forward<Args>(args)...));
+}
+
+FMT_END_NAMESPACE
+
+#endif  // FMT_OSTREAM_H_
diff --git a/ext/spdlog/fmt/bundled/printf.h b/ext/spdlog/fmt/bundled/printf.h
new file mode 100644
index 0000000..072cc6b
--- /dev/null
+++ b/ext/spdlog/fmt/bundled/printf.h
@@ -0,0 +1,656 @@
+// Formatting library for C++ - legacy printf implementation
+//
+// Copyright (c) 2012 - 2016, Victor Zverovich
+// All rights reserved.
+//
+// For the license information refer to format.h.
+
+#ifndef FMT_PRINTF_H_
+#define FMT_PRINTF_H_
+
+#ifndef FMT_MODULE
+#  include <algorithm>  // std::max
+#  include <limits>     // std::numeric_limits
+#endif
+
+#include "format.h"
+
+FMT_BEGIN_NAMESPACE
+FMT_BEGIN_EXPORT
+
+template <typename T> struct printf_formatter {
+  printf_formatter() = delete;
+};
+
+template <typename Char> class basic_printf_context {
+ private:
+  basic_appender<Char> out_;
+  basic_format_args<basic_printf_context> args_;
+
+  static_assert(std::is_same<Char, char>::value ||
+                    std::is_same<Char, wchar_t>::value,
+                "Unsupported code unit type.");
+
+ public:
+  using char_type = Char;
+  using parse_context_type = basic_format_parse_context<Char>;
+  template <typename T> using formatter_type = printf_formatter<T>;
+
+  /// Constructs a `printf_context` object. References to the arguments are
+  /// stored in the context object so make sure they have appropriate lifetimes.
+  basic_printf_context(basic_appender<Char> out,
+                       basic_format_args<basic_printf_context> args)
+      : out_(out), args_(args) {}
+
+  auto out() -> basic_appender<Char> { return out_; }
+  void advance_to(basic_appender<Char>) {}
+
+  auto locale() -> detail::locale_ref { return {}; }
+
+  auto arg(int id) const -> basic_format_arg<basic_printf_context> {
+    return args_.get(id);
+  }
+};
+
+namespace detail {
+
+// Checks if a value fits in int - used to avoid warnings about comparing
+// signed and unsigned integers.
+template <bool IsSigned> struct int_checker {
+  template <typename T> static auto fits_in_int(T value) -> bool {
+    unsigned max = to_unsigned(max_value<int>());
+    return value <= max;
+  }
+  static auto fits_in_int(bool) -> bool { return true; }
+};
+
+template <> struct int_checker<true> {
+  template <typename T> static auto fits_in_int(T value) -> bool {
+    return value >= (std::numeric_limits<int>::min)() &&
+           value <= max_value<int>();
+  }
+  static auto fits_in_int(int) -> bool { return true; }
+};
+
+struct printf_precision_handler {
+  template <typename T, FMT_ENABLE_IF(std::is_integral<T>::value)>
+  auto operator()(T value) -> int {
+    if (!int_checker<std::numeric_limits<T>::is_signed>::fits_in_int(value))
+      report_error("number is too big");
+    return (std::max)(static_cast<int>(value), 0);
+  }
+
+  template <typename T, FMT_ENABLE_IF(!std::is_integral<T>::value)>
+  auto operator()(T) -> int {
+    report_error("precision is not integer");
+    return 0;
+  }
+};
+
+// An argument visitor that returns true iff arg is a zero integer.
+struct is_zero_int {
+  template <typename T, FMT_ENABLE_IF(std::is_integral<T>::value)>
+  auto operator()(T value) -> bool {
+    return value == 0;
+  }
+
+  template <typename T, FMT_ENABLE_IF(!std::is_integral<T>::value)>
+  auto operator()(T) -> bool {
+    return false;
+  }
+};
+
+template <typename T> struct make_unsigned_or_bool : std::make_unsigned<T> {};
+
+template <> struct make_unsigned_or_bool<bool> {
+  using type = bool;
+};
+
+template <typename T, typename Context> class arg_converter {
+ private:
+  using char_type = typename Context::char_type;
+
+  basic_format_arg<Context>& arg_;
+  char_type type_;
+
+ public:
+  arg_converter(basic_format_arg<Context>& arg, char_type type)
+      : arg_(arg), type_(type) {}
+
+  void operator()(bool value) {
+    if (type_ != 's') operator()<bool>(value);
+  }
+
+  template <typename U, FMT_ENABLE_IF(std::is_integral<U>::value)>
+  void operator()(U value) {
+    bool is_signed = type_ == 'd' || type_ == 'i';
+    using target_type = conditional_t<std::is_same<T, void>::value, U, T>;
+    if (const_check(sizeof(target_type) <= sizeof(int))) {
+      // Extra casts are used to silence warnings.
+      if (is_signed) {
+        auto n = static_cast<int>(static_cast<target_type>(value));
+        arg_ = detail::make_arg<Context>(n);
+      } else {
+        using unsigned_type = typename make_unsigned_or_bool<target_type>::type;
+        auto n = static_cast<unsigned>(static_cast<unsigned_type>(value));
+        arg_ = detail::make_arg<Context>(n);
+      }
+    } else {
+      if (is_signed) {
+        // glibc's printf doesn't sign extend arguments of smaller types:
+        //   std::printf("%lld", -42);  // prints "4294967254"
+        // but we don't have to do the same because it's a UB.
+        auto n = static_cast<long long>(value);
+        arg_ = detail::make_arg<Context>(n);
+      } else {
+        auto n = static_cast<typename make_unsigned_or_bool<U>::type>(value);
+        arg_ = detail::make_arg<Context>(n);
+      }
+    }
+  }
+
+  template <typename U, FMT_ENABLE_IF(!std::is_integral<U>::value)>
+  void operator()(U) {}  // No conversion needed for non-integral types.
+};
+
+// Converts an integer argument to T for printf, if T is an integral type.
+// If T is void, the argument is converted to corresponding signed or unsigned
+// type depending on the type specifier: 'd' and 'i' - signed, other -
+// unsigned).
+template <typename T, typename Context, typename Char>
+void convert_arg(basic_format_arg<Context>& arg, Char type) {
+  arg.visit(arg_converter<T, Context>(arg, type));
+}
+
+// Converts an integer argument to char for printf.
+template <typename Context> class char_converter {
+ private:
+  basic_format_arg<Context>& arg_;
+
+ public:
+  explicit char_converter(basic_format_arg<Context>& arg) : arg_(arg) {}
+
+  template <typename T, FMT_ENABLE_IF(std::is_integral<T>::value)>
+  void operator()(T value) {
+    auto c = static_cast<typename Context::char_type>(value);
+    arg_ = detail::make_arg<Context>(c);
+  }
+
+  template <typename T, FMT_ENABLE_IF(!std::is_integral<T>::value)>
+  void operator()(T) {}  // No conversion needed for non-integral types.
+};
+
+// An argument visitor that return a pointer to a C string if argument is a
+// string or null otherwise.
+template <typename Char> struct get_cstring {
+  template <typename T> auto operator()(T) -> const Char* { return nullptr; }
+  auto operator()(const Char* s) -> const Char* { return s; }
+};
+
+// Checks if an argument is a valid printf width specifier and sets
+// left alignment if it is negative.
+class printf_width_handler {
+ private:
+  format_specs& specs_;
+
+ public:
+  explicit printf_width_handler(format_specs& specs) : specs_(specs) {}
+
+  template <typename T, FMT_ENABLE_IF(std::is_integral<T>::value)>
+  auto operator()(T value) -> unsigned {
+    auto width = static_cast<uint32_or_64_or_128_t<T>>(value);
+    if (detail::is_negative(value)) {
+      specs_.align = align::left;
+      width = 0 - width;
+    }
+    unsigned int_max = to_unsigned(max_value<int>());
+    if (width > int_max) report_error("number is too big");
+    return static_cast<unsigned>(width);
+  }
+
+  template <typename T, FMT_ENABLE_IF(!std::is_integral<T>::value)>
+  auto operator()(T) -> unsigned {
+    report_error("width is not integer");
+    return 0;
+  }
+};
+
+// Workaround for a bug with the XL compiler when initializing
+// printf_arg_formatter's base class.
+template <typename Char>
+auto make_arg_formatter(basic_appender<Char> iter, format_specs& s)
+    -> arg_formatter<Char> {
+  return {iter, s, locale_ref()};
+}
+
+// The `printf` argument formatter.
+template <typename Char>
+class printf_arg_formatter : public arg_formatter<Char> {
+ private:
+  using base = arg_formatter<Char>;
+  using context_type = basic_printf_context<Char>;
+
+  context_type& context_;
+
+  void write_null_pointer(bool is_string = false) {
+    auto s = this->specs;
+    s.type = presentation_type::none;
+    write_bytes<Char>(this->out, is_string ? "(null)" : "(nil)", s);
+  }
+
+ public:
+  printf_arg_formatter(basic_appender<Char> iter, format_specs& s,
+                       context_type& ctx)
+      : base(make_arg_formatter(iter, s)), context_(ctx) {}
+
+  void operator()(monostate value) { base::operator()(value); }
+
+  template <typename T, FMT_ENABLE_IF(detail::is_integral<T>::value)>
+  void operator()(T value) {
+    // MSVC2013 fails to compile separate overloads for bool and Char so use
+    // std::is_same instead.
+    if (!std::is_same<T, Char>::value) {
+      base::operator()(value);
+      return;
+    }
+    format_specs s = this->specs;
+    if (s.type != presentation_type::none && s.type != presentation_type::chr) {
+      return (*this)(static_cast<int>(value));
+    }
+    s.sign = sign::none;
+    s.alt = false;
+    s.fill = ' ';  // Ignore '0' flag for char types.
+    // align::numeric needs to be overwritten here since the '0' flag is
+    // ignored for non-numeric types
+    if (s.align == align::none || s.align == align::numeric)
+      s.align = align::right;
+    write<Char>(this->out, static_cast<Char>(value), s);
+  }
+
+  template <typename T, FMT_ENABLE_IF(std::is_floating_point<T>::value)>
+  void operator()(T value) {
+    base::operator()(value);
+  }
+
+  void operator()(const char* value) {
+    if (value)
+      base::operator()(value);
+    else
+      write_null_pointer(this->specs.type != presentation_type::pointer);
+  }
+
+  void operator()(const wchar_t* value) {
+    if (value)
+      base::operator()(value);
+    else
+      write_null_pointer(this->specs.type != presentation_type::pointer);
+  }
+
+  void operator()(basic_string_view<Char> value) { base::operator()(value); }
+
+  void operator()(const void* value) {
+    if (value)
+      base::operator()(value);
+    else
+      write_null_pointer();
+  }
+
+  void operator()(typename basic_format_arg<context_type>::handle handle) {
+    auto parse_ctx = basic_format_parse_context<Char>({});
+    handle.format(parse_ctx, context_);
+  }
+};
+
+template <typename Char>
+void parse_flags(format_specs& specs, const Char*& it, const Char* end) {
+  for (; it != end; ++it) {
+    switch (*it) {
+    case '-':
+      specs.align = align::left;
+      break;
+    case '+':
+      specs.sign = sign::plus;
+      break;
+    case '0':
+      specs.fill = '0';
+      break;
+    case ' ':
+      if (specs.sign != sign::plus) specs.sign = sign::space;
+      break;
+    case '#':
+      specs.alt = true;
+      break;
+    default:
+      return;
+    }
+  }
+}
+
+template <typename Char, typename GetArg>
+auto parse_header(const Char*& it, const Char* end, format_specs& specs,
+                  GetArg get_arg) -> int {
+  int arg_index = -1;
+  Char c = *it;
+  if (c >= '0' && c <= '9') {
+    // Parse an argument index (if followed by '$') or a width possibly
+    // preceded with '0' flag(s).
+    int value = parse_nonnegative_int(it, end, -1);
+    if (it != end && *it == '$') {  // value is an argument index
+      ++it;
+      arg_index = value != -1 ? value : max_value<int>();
+    } else {
+      if (c == '0') specs.fill = '0';
+      if (value != 0) {
+        // Nonzero value means that we parsed width and don't need to
+        // parse it or flags again, so return now.
+        if (value == -1) report_error("number is too big");
+        specs.width = value;
+        return arg_index;
+      }
+    }
+  }
+  parse_flags(specs, it, end);
+  // Parse width.
+  if (it != end) {
+    if (*it >= '0' && *it <= '9') {
+      specs.width = parse_nonnegative_int(it, end, -1);
+      if (specs.width == -1) report_error("number is too big");
+    } else if (*it == '*') {
+      ++it;
+      specs.width = static_cast<int>(
+          get_arg(-1).visit(detail::printf_width_handler(specs)));
+    }
+  }
+  return arg_index;
+}
+
+inline auto parse_printf_presentation_type(char c, type t, bool& upper)
+    -> presentation_type {
+  using pt = presentation_type;
+  constexpr auto integral_set = sint_set | uint_set | bool_set | char_set;
+  switch (c) {
+  case 'd':
+    return in(t, integral_set) ? pt::dec : pt::none;
+  case 'o':
+    return in(t, integral_set) ? pt::oct : pt::none;
+  case 'X':
+    upper = true;
+    FMT_FALLTHROUGH;
+  case 'x':
+    return in(t, integral_set) ? pt::hex : pt::none;
+  case 'E':
+    upper = true;
+    FMT_FALLTHROUGH;
+  case 'e':
+    return in(t, float_set) ? pt::exp : pt::none;
+  case 'F':
+    upper = true;
+    FMT_FALLTHROUGH;
+  case 'f':
+    return in(t, float_set) ? pt::fixed : pt::none;
+  case 'G':
+    upper = true;
+    FMT_FALLTHROUGH;
+  case 'g':
+    return in(t, float_set) ? pt::general : pt::none;
+  case 'A':
+    upper = true;
+    FMT_FALLTHROUGH;
+  case 'a':
+    return in(t, float_set) ? pt::hexfloat : pt::none;
+  case 'c':
+    return in(t, integral_set) ? pt::chr : pt::none;
+  case 's':
+    return in(t, string_set | cstring_set) ? pt::string : pt::none;
+  case 'p':
+    return in(t, pointer_set | cstring_set) ? pt::pointer : pt::none;
+  default:
+    return pt::none;
+  }
+}
+
+template <typename Char, typename Context>
+void vprintf(buffer<Char>& buf, basic_string_view<Char> format,
+             basic_format_args<Context> args) {
+  using iterator = basic_appender<Char>;
+  auto out = iterator(buf);
+  auto context = basic_printf_context<Char>(out, args);
+  auto parse_ctx = basic_format_parse_context<Char>(format);
+
+  // Returns the argument with specified index or, if arg_index is -1, the next
+  // argument.
+  auto get_arg = [&](int arg_index) {
+    if (arg_index < 0)
+      arg_index = parse_ctx.next_arg_id();
+    else
+      parse_ctx.check_arg_id(--arg_index);
+    return detail::get_arg(context, arg_index);
+  };
+
+  const Char* start = parse_ctx.begin();
+  const Char* end = parse_ctx.end();
+  auto it = start;
+  while (it != end) {
+    if (!find<false, Char>(it, end, '%', it)) {
+      it = end;  // find leaves it == nullptr if it doesn't find '%'.
+      break;
+    }
+    Char c = *it++;
+    if (it != end && *it == c) {
+      write(out, basic_string_view<Char>(start, to_unsigned(it - start)));
+      start = ++it;
+      continue;
+    }
+    write(out, basic_string_view<Char>(start, to_unsigned(it - 1 - start)));
+
+    auto specs = format_specs();
+    specs.align = align::right;
+
+    // Parse argument index, flags and width.
+    int arg_index = parse_header(it, end, specs, get_arg);
+    if (arg_index == 0) report_error("argument not found");
+
+    // Parse precision.
+    if (it != end && *it == '.') {
+      ++it;
+      c = it != end ? *it : 0;
+      if ('0' <= c && c <= '9') {
+        specs.precision = parse_nonnegative_int(it, end, 0);
+      } else if (c == '*') {
+        ++it;
+        specs.precision =
+            static_cast<int>(get_arg(-1).visit(printf_precision_handler()));
+      } else {
+        specs.precision = 0;
+      }
+    }
+
+    auto arg = get_arg(arg_index);
+    // For d, i, o, u, x, and X conversion specifiers, if a precision is
+    // specified, the '0' flag is ignored
+    if (specs.precision >= 0 && arg.is_integral()) {
+      // Ignore '0' for non-numeric types or if '-' present.
+      specs.fill = ' ';
+    }
+    if (specs.precision >= 0 && arg.type() == type::cstring_type) {
+      auto str = arg.visit(get_cstring<Char>());
+      auto str_end = str + specs.precision;
+      auto nul = std::find(str, str_end, Char());
+      auto sv = basic_string_view<Char>(
+          str, to_unsigned(nul != str_end ? nul - str : specs.precision));
+      arg = make_arg<basic_printf_context<Char>>(sv);
+    }
+    if (specs.alt && arg.visit(is_zero_int())) specs.alt = false;
+    if (specs.fill.template get<Char>() == '0') {
+      if (arg.is_arithmetic() && specs.align != align::left)
+        specs.align = align::numeric;
+      else
+        specs.fill = ' ';  // Ignore '0' flag for non-numeric types or if '-'
+                           // flag is also present.
+    }
+
+    // Parse length and convert the argument to the required type.
+    c = it != end ? *it++ : 0;
+    Char t = it != end ? *it : 0;
+    switch (c) {
+    case 'h':
+      if (t == 'h') {
+        ++it;
+        t = it != end ? *it : 0;
+        convert_arg<signed char>(arg, t);
+      } else {
+        convert_arg<short>(arg, t);
+      }
+      break;
+    case 'l':
+      if (t == 'l') {
+        ++it;
+        t = it != end ? *it : 0;
+        convert_arg<long long>(arg, t);
+      } else {
+        convert_arg<long>(arg, t);
+      }
+      break;
+    case 'j':
+      convert_arg<intmax_t>(arg, t);
+      break;
+    case 'z':
+      convert_arg<size_t>(arg, t);
+      break;
+    case 't':
+      convert_arg<std::ptrdiff_t>(arg, t);
+      break;
+    case 'L':
+      // printf produces garbage when 'L' is omitted for long double, no
+      // need to do the same.
+      break;
+    default:
+      --it;
+      convert_arg<void>(arg, c);
+    }
+
+    // Parse type.
+    if (it == end) report_error("invalid format string");
+    char type = static_cast<char>(*it++);
+    if (arg.is_integral()) {
+      // Normalize type.
+      switch (type) {
+      case 'i':
+      case 'u':
+        type = 'd';
+        break;
+      case 'c':
+        arg.visit(char_converter<basic_printf_context<Char>>(arg));
+        break;
+      }
+    }
+    bool upper = false;
+    specs.type = parse_printf_presentation_type(type, arg.type(), upper);
+    if (specs.type == presentation_type::none)
+      report_error("invalid format specifier");
+    specs.upper = upper;
+
+    start = it;
+
+    // Format argument.
+    arg.visit(printf_arg_formatter<Char>(out, specs, context));
+  }
+  write(out, basic_string_view<Char>(start, to_unsigned(it - start)));
+}
+}  // namespace detail
+
+using printf_context = basic_printf_context<char>;
+using wprintf_context = basic_printf_context<wchar_t>;
+
+using printf_args = basic_format_args<printf_context>;
+using wprintf_args = basic_format_args<wprintf_context>;
+
+/// Constructs an `format_arg_store` object that contains references to
+/// arguments and can be implicitly converted to `printf_args`.
+template <typename Char = char, typename... T>
+inline auto make_printf_args(T&... args)
+    -> decltype(fmt::make_format_args<basic_printf_context<Char>>(args...)) {
+  return fmt::make_format_args<basic_printf_context<Char>>(args...);
+}
+
+template <typename Char> struct vprintf_args {
+  using type = basic_format_args<basic_printf_context<Char>>;
+};
+
+template <typename Char>
+inline auto vsprintf(basic_string_view<Char> fmt,
+                     typename vprintf_args<Char>::type args)
+    -> std::basic_string<Char> {
+  auto buf = basic_memory_buffer<Char>();
+  detail::vprintf(buf, fmt, args);
+  return to_string(buf);
+}
+
+/**
+ * Formats `args` according to specifications in `fmt` and returns the result
+ * as as string.
+ *
+ * **Example**:
+ *
+ *     std::string message = fmt::sprintf("The answer is %d", 42);
+ */
+template <typename S, typename... T, typename Char = char_t<S>>
+inline auto sprintf(const S& fmt, const T&... args) -> std::basic_string<Char> {
+  return vsprintf(detail::to_string_view(fmt),
+                  fmt::make_format_args<basic_printf_context<Char>>(args...));
+}
+
+template <typename Char>
+inline auto vfprintf(std::FILE* f, basic_string_view<Char> fmt,
+                     typename vprintf_args<Char>::type args) -> int {
+  auto buf = basic_memory_buffer<Char>();
+  detail::vprintf(buf, fmt, args);
+  size_t size = buf.size();
+  return std::fwrite(buf.data(), sizeof(Char), size, f) < size
+             ? -1
+             : static_cast<int>(size);
+}
+
+/**
+ * Formats `args` according to specifications in `fmt` and writes the output
+ * to `f`.
+ *
+ * **Example**:
+ *
+ *     fmt::fprintf(stderr, "Don't %s!", "panic");
+ */
+template <typename S, typename... T, typename Char = char_t<S>>
+inline auto fprintf(std::FILE* f, const S& fmt, const T&... args) -> int {
+  return vfprintf(f, detail::to_string_view(fmt),
+                  make_printf_args<Char>(args...));
+}
+
+template <typename Char>
+FMT_DEPRECATED inline auto vprintf(basic_string_view<Char> fmt,
+                                   typename vprintf_args<Char>::type args)
+    -> int {
+  return vfprintf(stdout, fmt, args);
+}
+
+/**
+ * Formats `args` according to specifications in `fmt` and writes the output
+ * to `stdout`.
+ *
+ * **Example**:
+ *
+ *   fmt::printf("Elapsed time: %.2f seconds", 1.23);
+ */
+template <typename... T>
+inline auto printf(string_view fmt, const T&... args) -> int {
+  return vfprintf(stdout, fmt, make_printf_args(args...));
+}
+template <typename... T>
+FMT_DEPRECATED inline auto printf(basic_string_view<wchar_t> fmt,
+                                  const T&... args) -> int {
+  return vfprintf(stdout, fmt, make_printf_args<wchar_t>(args...));
+}
+
+FMT_END_EXPORT
+FMT_END_NAMESPACE
+
+#endif  // FMT_PRINTF_H_
diff --git a/ext/spdlog/fmt/bundled/ranges.h b/ext/spdlog/fmt/bundled/ranges.h
new file mode 100644
index 0000000..0d3dfbd
--- /dev/null
+++ b/ext/spdlog/fmt/bundled/ranges.h
@@ -0,0 +1,882 @@
+// Formatting library for C++ - range and tuple support
+//
+// Copyright (c) 2012 - present, Victor Zverovich and {fmt} contributors
+// All rights reserved.
+//
+// For the license information refer to format.h.
+
+#ifndef FMT_RANGES_H_
+#define FMT_RANGES_H_
+
+#ifndef FMT_MODULE
+#  include <initializer_list>
+#  include <iterator>
+#  include <string>
+#  include <tuple>
+#  include <type_traits>
+#  include <utility>
+#endif
+
+#include "format.h"
+
+FMT_BEGIN_NAMESPACE
+
+FMT_EXPORT
+enum class range_format { disabled, map, set, sequence, string, debug_string };
+
+namespace detail {
+
+template <typename T> class is_map {
+  template <typename U> static auto check(U*) -> typename U::mapped_type;
+  template <typename> static void check(...);
+
+ public:
+  static constexpr const bool value =
+      !std::is_void<decltype(check<T>(nullptr))>::value;
+};
+
+template <typename T> class is_set {
+  template <typename U> static auto check(U*) -> typename U::key_type;
+  template <typename> static void check(...);
+
+ public:
+  static constexpr const bool value =
+      !std::is_void<decltype(check<T>(nullptr))>::value && !is_map<T>::value;
+};
+
+template <typename... Ts> struct conditional_helper {};
+
+template <typename T, typename _ = void> struct is_range_ : std::false_type {};
+
+#if !FMT_MSC_VERSION || FMT_MSC_VERSION > 1800
+
+#  define FMT_DECLTYPE_RETURN(val)  \
+    ->decltype(val) { return val; } \
+    static_assert(                  \
+        true, "")  // This makes it so that a semicolon is required after the
+                   // macro, which helps clang-format handle the formatting.
+
+// C array overload
+template <typename T, std::size_t N>
+auto range_begin(const T (&arr)[N]) -> const T* {
+  return arr;
+}
+template <typename T, std::size_t N>
+auto range_end(const T (&arr)[N]) -> const T* {
+  return arr + N;
+}
+
+template <typename T, typename Enable = void>
+struct has_member_fn_begin_end_t : std::false_type {};
+
+template <typename T>
+struct has_member_fn_begin_end_t<T, void_t<decltype(*std::declval<T>().begin()),
+                                           decltype(std::declval<T>().end())>>
+    : std::true_type {};
+
+// Member function overloads.
+template <typename T>
+auto range_begin(T&& rng) FMT_DECLTYPE_RETURN(static_cast<T&&>(rng).begin());
+template <typename T>
+auto range_end(T&& rng) FMT_DECLTYPE_RETURN(static_cast<T&&>(rng).end());
+
+// ADL overloads. Only participate in overload resolution if member functions
+// are not found.
+template <typename T>
+auto range_begin(T&& rng)
+    -> enable_if_t<!has_member_fn_begin_end_t<T&&>::value,
+                   decltype(begin(static_cast<T&&>(rng)))> {
+  return begin(static_cast<T&&>(rng));
+}
+template <typename T>
+auto range_end(T&& rng) -> enable_if_t<!has_member_fn_begin_end_t<T&&>::value,
+                                       decltype(end(static_cast<T&&>(rng)))> {
+  return end(static_cast<T&&>(rng));
+}
+
+template <typename T, typename Enable = void>
+struct has_const_begin_end : std::false_type {};
+template <typename T, typename Enable = void>
+struct has_mutable_begin_end : std::false_type {};
+
+template <typename T>
+struct has_const_begin_end<
+    T, void_t<decltype(*detail::range_begin(
+                  std::declval<const remove_cvref_t<T>&>())),
+              decltype(detail::range_end(
+                  std::declval<const remove_cvref_t<T>&>()))>>
+    : std::true_type {};
+
+template <typename T>
+struct has_mutable_begin_end<
+    T, void_t<decltype(*detail::range_begin(std::declval<T&>())),
+              decltype(detail::range_end(std::declval<T&>())),
+              // the extra int here is because older versions of MSVC don't
+              // SFINAE properly unless there are distinct types
+              int>> : std::true_type {};
+
+template <typename T>
+struct is_range_<T, void>
+    : std::integral_constant<bool, (has_const_begin_end<T>::value ||
+                                    has_mutable_begin_end<T>::value)> {};
+#  undef FMT_DECLTYPE_RETURN
+#endif
+
+// tuple_size and tuple_element check.
+template <typename T> class is_tuple_like_ {
+  template <typename U>
+  static auto check(U* p) -> decltype(std::tuple_size<U>::value, int());
+  template <typename> static void check(...);
+
+ public:
+  static constexpr const bool value =
+      !std::is_void<decltype(check<T>(nullptr))>::value;
+};
+
+// Check for integer_sequence
+#if defined(__cpp_lib_integer_sequence) || FMT_MSC_VERSION >= 1900
+template <typename T, T... N>
+using integer_sequence = std::integer_sequence<T, N...>;
+template <size_t... N> using index_sequence = std::index_sequence<N...>;
+template <size_t N> using make_index_sequence = std::make_index_sequence<N>;
+#else
+template <typename T, T... N> struct integer_sequence {
+  using value_type = T;
+
+  static FMT_CONSTEXPR auto size() -> size_t { return sizeof...(N); }
+};
+
+template <size_t... N> using index_sequence = integer_sequence<size_t, N...>;
+
+template <typename T, size_t N, T... Ns>
+struct make_integer_sequence : make_integer_sequence<T, N - 1, N - 1, Ns...> {};
+template <typename T, T... Ns>
+struct make_integer_sequence<T, 0, Ns...> : integer_sequence<T, Ns...> {};
+
+template <size_t N>
+using make_index_sequence = make_integer_sequence<size_t, N>;
+#endif
+
+template <typename T>
+using tuple_index_sequence = make_index_sequence<std::tuple_size<T>::value>;
+
+template <typename T, typename C, bool = is_tuple_like_<T>::value>
+class is_tuple_formattable_ {
+ public:
+  static constexpr const bool value = false;
+};
+template <typename T, typename C> class is_tuple_formattable_<T, C, true> {
+  template <size_t... Is>
+  static auto all_true(index_sequence<Is...>,
+                       integer_sequence<bool, (Is >= 0)...>) -> std::true_type;
+  static auto all_true(...) -> std::false_type;
+
+  template <size_t... Is>
+  static auto check(index_sequence<Is...>) -> decltype(all_true(
+      index_sequence<Is...>{},
+      integer_sequence<bool,
+                       (is_formattable<typename std::tuple_element<Is, T>::type,
+                                       C>::value)...>{}));
+
+ public:
+  static constexpr const bool value =
+      decltype(check(tuple_index_sequence<T>{}))::value;
+};
+
+template <typename Tuple, typename F, size_t... Is>
+FMT_CONSTEXPR void for_each(index_sequence<Is...>, Tuple&& t, F&& f) {
+  using std::get;
+  // Using a free function get<Is>(Tuple) now.
+  const int unused[] = {0, ((void)f(get<Is>(t)), 0)...};
+  ignore_unused(unused);
+}
+
+template <typename Tuple, typename F>
+FMT_CONSTEXPR void for_each(Tuple&& t, F&& f) {
+  for_each(tuple_index_sequence<remove_cvref_t<Tuple>>(),
+           std::forward<Tuple>(t), std::forward<F>(f));
+}
+
+template <typename Tuple1, typename Tuple2, typename F, size_t... Is>
+void for_each2(index_sequence<Is...>, Tuple1&& t1, Tuple2&& t2, F&& f) {
+  using std::get;
+  const int unused[] = {0, ((void)f(get<Is>(t1), get<Is>(t2)), 0)...};
+  ignore_unused(unused);
+}
+
+template <typename Tuple1, typename Tuple2, typename F>
+void for_each2(Tuple1&& t1, Tuple2&& t2, F&& f) {
+  for_each2(tuple_index_sequence<remove_cvref_t<Tuple1>>(),
+            std::forward<Tuple1>(t1), std::forward<Tuple2>(t2),
+            std::forward<F>(f));
+}
+
+namespace tuple {
+// Workaround a bug in MSVC 2019 (v140).
+template <typename Char, typename... T>
+using result_t = std::tuple<formatter<remove_cvref_t<T>, Char>...>;
+
+using std::get;
+template <typename Tuple, typename Char, std::size_t... Is>
+auto get_formatters(index_sequence<Is...>)
+    -> result_t<Char, decltype(get<Is>(std::declval<Tuple>()))...>;
+}  // namespace tuple
+
+#if FMT_MSC_VERSION && FMT_MSC_VERSION < 1920
+// Older MSVC doesn't get the reference type correctly for arrays.
+template <typename R> struct range_reference_type_impl {
+  using type = decltype(*detail::range_begin(std::declval<R&>()));
+};
+
+template <typename T, std::size_t N> struct range_reference_type_impl<T[N]> {
+  using type = T&;
+};
+
+template <typename T>
+using range_reference_type = typename range_reference_type_impl<T>::type;
+#else
+template <typename Range>
+using range_reference_type =
+    decltype(*detail::range_begin(std::declval<Range&>()));
+#endif
+
+// We don't use the Range's value_type for anything, but we do need the Range's
+// reference type, with cv-ref stripped.
+template <typename Range>
+using uncvref_type = remove_cvref_t<range_reference_type<Range>>;
+
+template <typename Formatter>
+FMT_CONSTEXPR auto maybe_set_debug_format(Formatter& f, bool set)
+    -> decltype(f.set_debug_format(set)) {
+  f.set_debug_format(set);
+}
+template <typename Formatter>
+FMT_CONSTEXPR void maybe_set_debug_format(Formatter&, ...) {}
+
+template <typename T>
+struct range_format_kind_
+    : std::integral_constant<range_format,
+                             std::is_same<uncvref_type<T>, T>::value
+                                 ? range_format::disabled
+                             : is_map<T>::value ? range_format::map
+                             : is_set<T>::value ? range_format::set
+                                                : range_format::sequence> {};
+
+template <range_format K>
+using range_format_constant = std::integral_constant<range_format, K>;
+
+// These are not generic lambdas for compatibility with C++11.
+template <typename ParseContext> struct parse_empty_specs {
+  template <typename Formatter> FMT_CONSTEXPR void operator()(Formatter& f) {
+    f.parse(ctx);
+    detail::maybe_set_debug_format(f, true);
+  }
+  ParseContext& ctx;
+};
+template <typename FormatContext> struct format_tuple_element {
+  using char_type = typename FormatContext::char_type;
+
+  template <typename T>
+  void operator()(const formatter<T, char_type>& f, const T& v) {
+    if (i > 0) ctx.advance_to(detail::copy<char_type>(separator, ctx.out()));
+    ctx.advance_to(f.format(v, ctx));
+    ++i;
+  }
+
+  int i;
+  FormatContext& ctx;
+  basic_string_view<char_type> separator;
+};
+
+}  // namespace detail
+
+template <typename T> struct is_tuple_like {
+  static constexpr const bool value =
+      detail::is_tuple_like_<T>::value && !detail::is_range_<T>::value;
+};
+
+template <typename T, typename C> struct is_tuple_formattable {
+  static constexpr const bool value =
+      detail::is_tuple_formattable_<T, C>::value;
+};
+
+template <typename Tuple, typename Char>
+struct formatter<Tuple, Char,
+                 enable_if_t<fmt::is_tuple_like<Tuple>::value &&
+                             fmt::is_tuple_formattable<Tuple, Char>::value>> {
+ private:
+  decltype(detail::tuple::get_formatters<Tuple, Char>(
+      detail::tuple_index_sequence<Tuple>())) formatters_;
+
+  basic_string_view<Char> separator_ = detail::string_literal<Char, ',', ' '>{};
+  basic_string_view<Char> opening_bracket_ =
+      detail::string_literal<Char, '('>{};
+  basic_string_view<Char> closing_bracket_ =
+      detail::string_literal<Char, ')'>{};
+
+ public:
+  FMT_CONSTEXPR formatter() {}
+
+  FMT_CONSTEXPR void set_separator(basic_string_view<Char> sep) {
+    separator_ = sep;
+  }
+
+  FMT_CONSTEXPR void set_brackets(basic_string_view<Char> open,
+                                  basic_string_view<Char> close) {
+    opening_bracket_ = open;
+    closing_bracket_ = close;
+  }
+
+  template <typename ParseContext>
+  FMT_CONSTEXPR auto parse(ParseContext& ctx) -> decltype(ctx.begin()) {
+    auto it = ctx.begin();
+    if (it != ctx.end() && *it != '}') report_error("invalid format specifier");
+    detail::for_each(formatters_, detail::parse_empty_specs<ParseContext>{ctx});
+    return it;
+  }
+
+  template <typename FormatContext>
+  auto format(const Tuple& value, FormatContext& ctx) const
+      -> decltype(ctx.out()) {
+    ctx.advance_to(detail::copy<Char>(opening_bracket_, ctx.out()));
+    detail::for_each2(
+        formatters_, value,
+        detail::format_tuple_element<FormatContext>{0, ctx, separator_});
+    return detail::copy<Char>(closing_bracket_, ctx.out());
+  }
+};
+
+template <typename T, typename Char> struct is_range {
+  static constexpr const bool value =
+      detail::is_range_<T>::value && !detail::has_to_string_view<T>::value;
+};
+
+namespace detail {
+template <typename Context> struct range_mapper {
+  using mapper = arg_mapper<Context>;
+
+  template <typename T,
+            FMT_ENABLE_IF(has_formatter<remove_cvref_t<T>, Context>::value)>
+  static auto map(T&& value) -> T&& {
+    return static_cast<T&&>(value);
+  }
+  template <typename T,
+            FMT_ENABLE_IF(!has_formatter<remove_cvref_t<T>, Context>::value)>
+  static auto map(T&& value)
+      -> decltype(mapper().map(static_cast<T&&>(value))) {
+    return mapper().map(static_cast<T&&>(value));
+  }
+};
+
+template <typename Char, typename Element>
+using range_formatter_type =
+    formatter<remove_cvref_t<decltype(range_mapper<buffered_context<Char>>{}
+                                          .map(std::declval<Element>()))>,
+              Char>;
+
+template <typename R>
+using maybe_const_range =
+    conditional_t<has_const_begin_end<R>::value, const R, R>;
+
+// Workaround a bug in MSVC 2015 and earlier.
+#if !FMT_MSC_VERSION || FMT_MSC_VERSION >= 1910
+template <typename R, typename Char>
+struct is_formattable_delayed
+    : is_formattable<uncvref_type<maybe_const_range<R>>, Char> {};
+#endif
+}  // namespace detail
+
+template <typename...> struct conjunction : std::true_type {};
+template <typename P> struct conjunction<P> : P {};
+template <typename P1, typename... Pn>
+struct conjunction<P1, Pn...>
+    : conditional_t<bool(P1::value), conjunction<Pn...>, P1> {};
+
+template <typename T, typename Char, typename Enable = void>
+struct range_formatter;
+
+template <typename T, typename Char>
+struct range_formatter<
+    T, Char,
+    enable_if_t<conjunction<std::is_same<T, remove_cvref_t<T>>,
+                            is_formattable<T, Char>>::value>> {
+ private:
+  detail::range_formatter_type<Char, T> underlying_;
+  basic_string_view<Char> separator_ = detail::string_literal<Char, ',', ' '>{};
+  basic_string_view<Char> opening_bracket_ =
+      detail::string_literal<Char, '['>{};
+  basic_string_view<Char> closing_bracket_ =
+      detail::string_literal<Char, ']'>{};
+  bool is_debug = false;
+
+  template <typename Output, typename It, typename Sentinel, typename U = T,
+            FMT_ENABLE_IF(std::is_same<U, Char>::value)>
+  auto write_debug_string(Output& out, It it, Sentinel end) const -> Output {
+    auto buf = basic_memory_buffer<Char>();
+    for (; it != end; ++it) buf.push_back(*it);
+    auto specs = format_specs();
+    specs.type = presentation_type::debug;
+    return detail::write<Char>(
+        out, basic_string_view<Char>(buf.data(), buf.size()), specs);
+  }
+
+  template <typename Output, typename It, typename Sentinel, typename U = T,
+            FMT_ENABLE_IF(!std::is_same<U, Char>::value)>
+  auto write_debug_string(Output& out, It, Sentinel) const -> Output {
+    return out;
+  }
+
+ public:
+  FMT_CONSTEXPR range_formatter() {}
+
+  FMT_CONSTEXPR auto underlying() -> detail::range_formatter_type<Char, T>& {
+    return underlying_;
+  }
+
+  FMT_CONSTEXPR void set_separator(basic_string_view<Char> sep) {
+    separator_ = sep;
+  }
+
+  FMT_CONSTEXPR void set_brackets(basic_string_view<Char> open,
+                                  basic_string_view<Char> close) {
+    opening_bracket_ = open;
+    closing_bracket_ = close;
+  }
+
+  template <typename ParseContext>
+  FMT_CONSTEXPR auto parse(ParseContext& ctx) -> decltype(ctx.begin()) {
+    auto it = ctx.begin();
+    auto end = ctx.end();
+    detail::maybe_set_debug_format(underlying_, true);
+    if (it == end) return underlying_.parse(ctx);
+
+    switch (detail::to_ascii(*it)) {
+    case 'n':
+      set_brackets({}, {});
+      ++it;
+      break;
+    case '?':
+      is_debug = true;
+      set_brackets({}, {});
+      ++it;
+      if (it == end || *it != 's') report_error("invalid format specifier");
+      FMT_FALLTHROUGH;
+    case 's':
+      if (!std::is_same<T, Char>::value)
+        report_error("invalid format specifier");
+      if (!is_debug) {
+        set_brackets(detail::string_literal<Char, '"'>{},
+                     detail::string_literal<Char, '"'>{});
+        set_separator({});
+        detail::maybe_set_debug_format(underlying_, false);
+      }
+      ++it;
+      return it;
+    }
+
+    if (it != end && *it != '}') {
+      if (*it != ':') report_error("invalid format specifier");
+      detail::maybe_set_debug_format(underlying_, false);
+      ++it;
+    }
+
+    ctx.advance_to(it);
+    return underlying_.parse(ctx);
+  }
+
+  template <typename R, typename FormatContext>
+  auto format(R&& range, FormatContext& ctx) const -> decltype(ctx.out()) {
+    auto mapper = detail::range_mapper<buffered_context<Char>>();
+    auto out = ctx.out();
+    auto it = detail::range_begin(range);
+    auto end = detail::range_end(range);
+    if (is_debug) return write_debug_string(out, std::move(it), end);
+
+    out = detail::copy<Char>(opening_bracket_, out);
+    int i = 0;
+    for (; it != end; ++it) {
+      if (i > 0) out = detail::copy<Char>(separator_, out);
+      ctx.advance_to(out);
+      auto&& item = *it;  // Need an lvalue
+      out = underlying_.format(mapper.map(item), ctx);
+      ++i;
+    }
+    out = detail::copy<Char>(closing_bracket_, out);
+    return out;
+  }
+};
+
+FMT_EXPORT
+template <typename T, typename Char, typename Enable = void>
+struct range_format_kind
+    : conditional_t<
+          is_range<T, Char>::value, detail::range_format_kind_<T>,
+          std::integral_constant<range_format, range_format::disabled>> {};
+
+template <typename R, typename Char>
+struct formatter<
+    R, Char,
+    enable_if_t<conjunction<
+        bool_constant<
+            range_format_kind<R, Char>::value != range_format::disabled &&
+            range_format_kind<R, Char>::value != range_format::map &&
+            range_format_kind<R, Char>::value != range_format::string &&
+            range_format_kind<R, Char>::value != range_format::debug_string>
+// Workaround a bug in MSVC 2015 and earlier.
+#if !FMT_MSC_VERSION || FMT_MSC_VERSION >= 1910
+        ,
+        detail::is_formattable_delayed<R, Char>
+#endif
+        >::value>> {
+ private:
+  using range_type = detail::maybe_const_range<R>;
+  range_formatter<detail::uncvref_type<range_type>, Char> range_formatter_;
+
+ public:
+  using nonlocking = void;
+
+  FMT_CONSTEXPR formatter() {
+    if (detail::const_check(range_format_kind<R, Char>::value !=
+                            range_format::set))
+      return;
+    range_formatter_.set_brackets(detail::string_literal<Char, '{'>{},
+                                  detail::string_literal<Char, '}'>{});
+  }
+
+  template <typename ParseContext>
+  FMT_CONSTEXPR auto parse(ParseContext& ctx) -> decltype(ctx.begin()) {
+    return range_formatter_.parse(ctx);
+  }
+
+  template <typename FormatContext>
+  auto format(range_type& range, FormatContext& ctx) const
+      -> decltype(ctx.out()) {
+    return range_formatter_.format(range, ctx);
+  }
+};
+
+// A map formatter.
+template <typename R, typename Char>
+struct formatter<
+    R, Char,
+    enable_if_t<range_format_kind<R, Char>::value == range_format::map>> {
+ private:
+  using map_type = detail::maybe_const_range<R>;
+  using element_type = detail::uncvref_type<map_type>;
+
+  decltype(detail::tuple::get_formatters<element_type, Char>(
+      detail::tuple_index_sequence<element_type>())) formatters_;
+  bool no_delimiters_ = false;
+
+ public:
+  FMT_CONSTEXPR formatter() {}
+
+  template <typename ParseContext>
+  FMT_CONSTEXPR auto parse(ParseContext& ctx) -> decltype(ctx.begin()) {
+    auto it = ctx.begin();
+    auto end = ctx.end();
+    if (it != end) {
+      if (detail::to_ascii(*it) == 'n') {
+        no_delimiters_ = true;
+        ++it;
+      }
+      if (it != end && *it != '}') {
+        if (*it != ':') report_error("invalid format specifier");
+        ++it;
+      }
+      ctx.advance_to(it);
+    }
+    detail::for_each(formatters_, detail::parse_empty_specs<ParseContext>{ctx});
+    return it;
+  }
+
+  template <typename FormatContext>
+  auto format(map_type& map, FormatContext& ctx) const -> decltype(ctx.out()) {
+    auto out = ctx.out();
+    basic_string_view<Char> open = detail::string_literal<Char, '{'>{};
+    if (!no_delimiters_) out = detail::copy<Char>(open, out);
+    int i = 0;
+    auto mapper = detail::range_mapper<buffered_context<Char>>();
+    basic_string_view<Char> sep = detail::string_literal<Char, ',', ' '>{};
+    for (auto&& value : map) {
+      if (i > 0) out = detail::copy<Char>(sep, out);
+      ctx.advance_to(out);
+      detail::for_each2(formatters_, mapper.map(value),
+                        detail::format_tuple_element<FormatContext>{
+                            0, ctx, detail::string_literal<Char, ':', ' '>{}});
+      ++i;
+    }
+    basic_string_view<Char> close = detail::string_literal<Char, '}'>{};
+    if (!no_delimiters_) out = detail::copy<Char>(close, out);
+    return out;
+  }
+};
+
+// A (debug_)string formatter.
+template <typename R, typename Char>
+struct formatter<
+    R, Char,
+    enable_if_t<range_format_kind<R, Char>::value == range_format::string ||
+                range_format_kind<R, Char>::value ==
+                    range_format::debug_string>> {
+ private:
+  using range_type = detail::maybe_const_range<R>;
+  using string_type =
+      conditional_t<std::is_constructible<
+                        detail::std_string_view<Char>,
+                        decltype(detail::range_begin(std::declval<R>())),
+                        decltype(detail::range_end(std::declval<R>()))>::value,
+                    detail::std_string_view<Char>, std::basic_string<Char>>;
+
+  formatter<string_type, Char> underlying_;
+
+ public:
+  template <typename ParseContext>
+  FMT_CONSTEXPR auto parse(ParseContext& ctx) -> decltype(ctx.begin()) {
+    return underlying_.parse(ctx);
+  }
+
+  template <typename FormatContext>
+  auto format(range_type& range, FormatContext& ctx) const
+      -> decltype(ctx.out()) {
+    auto out = ctx.out();
+    if (detail::const_check(range_format_kind<R, Char>::value ==
+                            range_format::debug_string))
+      *out++ = '"';
+    out = underlying_.format(
+        string_type{detail::range_begin(range), detail::range_end(range)}, ctx);
+    if (detail::const_check(range_format_kind<R, Char>::value ==
+                            range_format::debug_string))
+      *out++ = '"';
+    return out;
+  }
+};
+
+template <typename It, typename Sentinel, typename Char = char>
+struct join_view : detail::view {
+  It begin;
+  Sentinel end;
+  basic_string_view<Char> sep;
+
+  join_view(It b, Sentinel e, basic_string_view<Char> s)
+      : begin(std::move(b)), end(e), sep(s) {}
+};
+
+template <typename It, typename Sentinel, typename Char>
+struct formatter<join_view<It, Sentinel, Char>, Char> {
+ private:
+  using value_type =
+#ifdef __cpp_lib_ranges
+      std::iter_value_t<It>;
+#else
+      typename std::iterator_traits<It>::value_type;
+#endif
+  formatter<remove_cvref_t<value_type>, Char> value_formatter_;
+
+  using view_ref = conditional_t<std::is_copy_constructible<It>::value,
+                                 const join_view<It, Sentinel, Char>&,
+                                 join_view<It, Sentinel, Char>&&>;
+
+ public:
+  using nonlocking = void;
+
+  template <typename ParseContext>
+  FMT_CONSTEXPR auto parse(ParseContext& ctx) -> const Char* {
+    return value_formatter_.parse(ctx);
+  }
+
+  template <typename FormatContext>
+  auto format(view_ref& value, FormatContext& ctx) const
+      -> decltype(ctx.out()) {
+    auto it = std::forward<view_ref>(value).begin;
+    auto out = ctx.out();
+    if (it == value.end) return out;
+    out = value_formatter_.format(*it, ctx);
+    ++it;
+    while (it != value.end) {
+      out = detail::copy<Char>(value.sep.begin(), value.sep.end(), out);
+      ctx.advance_to(out);
+      out = value_formatter_.format(*it, ctx);
+      ++it;
+    }
+    return out;
+  }
+};
+
+/// Returns a view that formats the iterator range `[begin, end)` with elements
+/// separated by `sep`.
+template <typename It, typename Sentinel>
+auto join(It begin, Sentinel end, string_view sep) -> join_view<It, Sentinel> {
+  return {std::move(begin), end, sep};
+}
+
+/**
+ * Returns a view that formats `range` with elements separated by `sep`.
+ *
+ * **Example**:
+ *
+ *     auto v = std::vector<int>{1, 2, 3};
+ *     fmt::print("{}", fmt::join(v, ", "));
+ *     // Output: 1, 2, 3
+ *
+ * `fmt::join` applies passed format specifiers to the range elements:
+ *
+ *     fmt::print("{:02}", fmt::join(v, ", "));
+ *     // Output: 01, 02, 03
+ */
+template <typename Range>
+auto join(Range&& r, string_view sep)
+    -> join_view<decltype(detail::range_begin(r)),
+                 decltype(detail::range_end(r))> {
+  return {detail::range_begin(r), detail::range_end(r), sep};
+}
+
+template <typename Char, typename... T> struct tuple_join_view : detail::view {
+  const std::tuple<T...>& tuple;
+  basic_string_view<Char> sep;
+
+  tuple_join_view(const std::tuple<T...>& t, basic_string_view<Char> s)
+      : tuple(t), sep{s} {}
+};
+
+// Define FMT_TUPLE_JOIN_SPECIFIERS to enable experimental format specifiers
+// support in tuple_join. It is disabled by default because of issues with
+// the dynamic width and precision.
+#ifndef FMT_TUPLE_JOIN_SPECIFIERS
+#  define FMT_TUPLE_JOIN_SPECIFIERS 0
+#endif
+
+template <typename Char, typename... T>
+struct formatter<tuple_join_view<Char, T...>, Char> {
+  template <typename ParseContext>
+  FMT_CONSTEXPR auto parse(ParseContext& ctx) -> decltype(ctx.begin()) {
+    return do_parse(ctx, std::integral_constant<size_t, sizeof...(T)>());
+  }
+
+  template <typename FormatContext>
+  auto format(const tuple_join_view<Char, T...>& value,
+              FormatContext& ctx) const -> typename FormatContext::iterator {
+    return do_format(value, ctx,
+                     std::integral_constant<size_t, sizeof...(T)>());
+  }
+
+ private:
+  std::tuple<formatter<typename std::decay<T>::type, Char>...> formatters_;
+
+  template <typename ParseContext>
+  FMT_CONSTEXPR auto do_parse(ParseContext& ctx,
+                              std::integral_constant<size_t, 0>)
+      -> decltype(ctx.begin()) {
+    return ctx.begin();
+  }
+
+  template <typename ParseContext, size_t N>
+  FMT_CONSTEXPR auto do_parse(ParseContext& ctx,
+                              std::integral_constant<size_t, N>)
+      -> decltype(ctx.begin()) {
+    auto end = ctx.begin();
+#if FMT_TUPLE_JOIN_SPECIFIERS
+    end = std::get<sizeof...(T) - N>(formatters_).parse(ctx);
+    if (N > 1) {
+      auto end1 = do_parse(ctx, std::integral_constant<size_t, N - 1>());
+      if (end != end1)
+        report_error("incompatible format specs for tuple elements");
+    }
+#endif
+    return end;
+  }
+
+  template <typename FormatContext>
+  auto do_format(const tuple_join_view<Char, T...>&, FormatContext& ctx,
+                 std::integral_constant<size_t, 0>) const ->
+      typename FormatContext::iterator {
+    return ctx.out();
+  }
+
+  template <typename FormatContext, size_t N>
+  auto do_format(const tuple_join_view<Char, T...>& value, FormatContext& ctx,
+                 std::integral_constant<size_t, N>) const ->
+      typename FormatContext::iterator {
+    auto out = std::get<sizeof...(T) - N>(formatters_)
+                   .format(std::get<sizeof...(T) - N>(value.tuple), ctx);
+    if (N <= 1) return out;
+    out = detail::copy<Char>(value.sep, out);
+    ctx.advance_to(out);
+    return do_format(value, ctx, std::integral_constant<size_t, N - 1>());
+  }
+};
+
+namespace detail {
+// Check if T has an interface like a container adaptor (e.g. std::stack,
+// std::queue, std::priority_queue).
+template <typename T> class is_container_adaptor_like {
+  template <typename U> static auto check(U* p) -> typename U::container_type;
+  template <typename> static void check(...);
+
+ public:
+  static constexpr const bool value =
+      !std::is_void<decltype(check<T>(nullptr))>::value;
+};
+
+template <typename Container> struct all {
+  const Container& c;
+  auto begin() const -> typename Container::const_iterator { return c.begin(); }
+  auto end() const -> typename Container::const_iterator { return c.end(); }
+};
+}  // namespace detail
+
+template <typename T, typename Char>
+struct formatter<
+    T, Char,
+    enable_if_t<conjunction<detail::is_container_adaptor_like<T>,
+                            bool_constant<range_format_kind<T, Char>::value ==
+                                          range_format::disabled>>::value>>
+    : formatter<detail::all<typename T::container_type>, Char> {
+  using all = detail::all<typename T::container_type>;
+  template <typename FormatContext>
+  auto format(const T& t, FormatContext& ctx) const -> decltype(ctx.out()) {
+    struct getter : T {
+      static auto get(const T& t) -> all {
+        return {t.*(&getter::c)};  // Access c through the derived class.
+      }
+    };
+    return formatter<all>::format(getter::get(t), ctx);
+  }
+};
+
+FMT_BEGIN_EXPORT
+
+/**
+ * Returns an object that formats `std::tuple` with elements separated by `sep`.
+ *
+ * **Example**:
+ *
+ *     auto t = std::tuple<int, char>{1, 'a'};
+ *     fmt::print("{}", fmt::join(t, ", "));
+ *     // Output: 1, a
+ */
+template <typename... T>
+FMT_CONSTEXPR auto join(const std::tuple<T...>& tuple, string_view sep)
+    -> tuple_join_view<char, T...> {
+  return {tuple, sep};
+}
+
+/**
+ * Returns an object that formats `std::initializer_list` with elements
+ * separated by `sep`.
+ *
+ * **Example**:
+ *
+ *     fmt::print("{}", fmt::join({1, 2, 3}, ", "));
+ *     // Output: "1, 2, 3"
+ */
+template <typename T>
+auto join(std::initializer_list<T> list, string_view sep)
+    -> join_view<const T*, const T*> {
+  return join(std::begin(list), std::end(list), sep);
+}
+
+FMT_END_EXPORT
+FMT_END_NAMESPACE
+
+#endif  // FMT_RANGES_H_
diff --git a/ext/spdlog/fmt/bundled/std.h b/ext/spdlog/fmt/bundled/std.h
new file mode 100644
index 0000000..fb43940
--- /dev/null
+++ b/ext/spdlog/fmt/bundled/std.h
@@ -0,0 +1,699 @@
+// Formatting library for C++ - formatters for standard library types
+//
+// Copyright (c) 2012 - present, Victor Zverovich
+// All rights reserved.
+//
+// For the license information refer to format.h.
+
+#ifndef FMT_STD_H_
+#define FMT_STD_H_
+
+#include "format.h"
+#include "ostream.h"
+
+#ifndef FMT_MODULE
+#  include <atomic>
+#  include <bitset>
+#  include <complex>
+#  include <cstdlib>
+#  include <exception>
+#  include <memory>
+#  include <thread>
+#  include <type_traits>
+#  include <typeinfo>
+#  include <utility>
+#  include <vector>
+
+// Check FMT_CPLUSPLUS to suppress a bogus warning in MSVC.
+#  if FMT_CPLUSPLUS >= 201703L
+#    if FMT_HAS_INCLUDE(<filesystem>)
+#      include <filesystem>
+#    endif
+#    if FMT_HAS_INCLUDE(<variant>)
+#      include <variant>
+#    endif
+#    if FMT_HAS_INCLUDE(<optional>)
+#      include <optional>
+#    endif
+#  endif
+// Use > instead of >= in the version check because <source_location> may be
+// available after C++17 but before C++20 is marked as implemented.
+#  if FMT_CPLUSPLUS > 201703L && FMT_HAS_INCLUDE(<source_location>)
+#    include <source_location>
+#  endif
+#  if FMT_CPLUSPLUS > 202002L && FMT_HAS_INCLUDE(<expected>)
+#    include <expected>
+#  endif
+#endif  // FMT_MODULE
+
+#if FMT_HAS_INCLUDE(<version>)
+#  include <version>
+#endif
+
+// GCC 4 does not support FMT_HAS_INCLUDE.
+#if FMT_HAS_INCLUDE(<cxxabi.h>) || defined(__GLIBCXX__)
+#  include <cxxabi.h>
+// Android NDK with gabi++ library on some architectures does not implement
+// abi::__cxa_demangle().
+#  ifndef __GABIXX_CXXABI_H__
+#    define FMT_HAS_ABI_CXA_DEMANGLE
+#  endif
+#endif
+
+// For older Xcode versions, __cpp_lib_xxx flags are inaccurately defined.
+#ifndef FMT_CPP_LIB_FILESYSTEM
+#  ifdef __cpp_lib_filesystem
+#    define FMT_CPP_LIB_FILESYSTEM __cpp_lib_filesystem
+#  else
+#    define FMT_CPP_LIB_FILESYSTEM 0
+#  endif
+#endif
+
+#ifndef FMT_CPP_LIB_VARIANT
+#  ifdef __cpp_lib_variant
+#    define FMT_CPP_LIB_VARIANT __cpp_lib_variant
+#  else
+#    define FMT_CPP_LIB_VARIANT 0
+#  endif
+#endif
+
+#if FMT_CPP_LIB_FILESYSTEM
+FMT_BEGIN_NAMESPACE
+
+namespace detail {
+
+template <typename Char, typename PathChar>
+auto get_path_string(const std::filesystem::path& p,
+                     const std::basic_string<PathChar>& native) {
+  if constexpr (std::is_same_v<Char, char> && std::is_same_v<PathChar, wchar_t>)
+    return to_utf8<wchar_t>(native, to_utf8_error_policy::replace);
+  else
+    return p.string<Char>();
+}
+
+template <typename Char, typename PathChar>
+void write_escaped_path(basic_memory_buffer<Char>& quoted,
+                        const std::filesystem::path& p,
+                        const std::basic_string<PathChar>& native) {
+  if constexpr (std::is_same_v<Char, char> &&
+                std::is_same_v<PathChar, wchar_t>) {
+    auto buf = basic_memory_buffer<wchar_t>();
+    write_escaped_string<wchar_t>(std::back_inserter(buf), native);
+    bool valid = to_utf8<wchar_t>::convert(quoted, {buf.data(), buf.size()});
+    FMT_ASSERT(valid, "invalid utf16");
+  } else if constexpr (std::is_same_v<Char, PathChar>) {
+    write_escaped_string<std::filesystem::path::value_type>(
+        std::back_inserter(quoted), native);
+  } else {
+    write_escaped_string<Char>(std::back_inserter(quoted), p.string<Char>());
+  }
+}
+
+}  // namespace detail
+
+FMT_EXPORT
+template <typename Char> struct formatter<std::filesystem::path, Char> {
+ private:
+  format_specs specs_;
+  detail::arg_ref<Char> width_ref_;
+  bool debug_ = false;
+  char path_type_ = 0;
+
+ public:
+  FMT_CONSTEXPR void set_debug_format(bool set = true) { debug_ = set; }
+
+  template <typename ParseContext> FMT_CONSTEXPR auto parse(ParseContext& ctx) {
+    auto it = ctx.begin(), end = ctx.end();
+    if (it == end) return it;
+
+    it = detail::parse_align(it, end, specs_);
+    if (it == end) return it;
+
+    it = detail::parse_dynamic_spec(it, end, specs_.width, width_ref_, ctx);
+    if (it != end && *it == '?') {
+      debug_ = true;
+      ++it;
+    }
+    if (it != end && (*it == 'g')) path_type_ = detail::to_ascii(*it++);
+    return it;
+  }
+
+  template <typename FormatContext>
+  auto format(const std::filesystem::path& p, FormatContext& ctx) const {
+    auto specs = specs_;
+    auto path_string =
+        !path_type_ ? p.native()
+                    : p.generic_string<std::filesystem::path::value_type>();
+
+    detail::handle_dynamic_spec<detail::width_checker>(specs.width, width_ref_,
+                                                       ctx);
+    if (!debug_) {
+      auto s = detail::get_path_string<Char>(p, path_string);
+      return detail::write(ctx.out(), basic_string_view<Char>(s), specs);
+    }
+    auto quoted = basic_memory_buffer<Char>();
+    detail::write_escaped_path(quoted, p, path_string);
+    return detail::write(ctx.out(),
+                         basic_string_view<Char>(quoted.data(), quoted.size()),
+                         specs);
+  }
+};
+
+class path : public std::filesystem::path {
+ public:
+  auto display_string() const -> std::string {
+    const std::filesystem::path& base = *this;
+    return fmt::format(FMT_STRING("{}"), base);
+  }
+  auto system_string() const -> std::string { return string(); }
+
+  auto generic_display_string() const -> std::string {
+    const std::filesystem::path& base = *this;
+    return fmt::format(FMT_STRING("{:g}"), base);
+  }
+  auto generic_system_string() const -> std::string { return generic_string(); }
+};
+
+FMT_END_NAMESPACE
+#endif  // FMT_CPP_LIB_FILESYSTEM
+
+FMT_BEGIN_NAMESPACE
+FMT_EXPORT
+template <std::size_t N, typename Char>
+struct formatter<std::bitset<N>, Char> : nested_formatter<string_view> {
+ private:
+  // Functor because C++11 doesn't support generic lambdas.
+  struct writer {
+    const std::bitset<N>& bs;
+
+    template <typename OutputIt>
+    FMT_CONSTEXPR auto operator()(OutputIt out) -> OutputIt {
+      for (auto pos = N; pos > 0; --pos) {
+        out = detail::write<Char>(out, bs[pos - 1] ? Char('1') : Char('0'));
+      }
+
+      return out;
+    }
+  };
+
+ public:
+  template <typename FormatContext>
+  auto format(const std::bitset<N>& bs, FormatContext& ctx) const
+      -> decltype(ctx.out()) {
+    return write_padded(ctx, writer{bs});
+  }
+};
+
+FMT_EXPORT
+template <typename Char>
+struct formatter<std::thread::id, Char> : basic_ostream_formatter<Char> {};
+FMT_END_NAMESPACE
+
+#ifdef __cpp_lib_optional
+FMT_BEGIN_NAMESPACE
+FMT_EXPORT
+template <typename T, typename Char>
+struct formatter<std::optional<T>, Char,
+                 std::enable_if_t<is_formattable<T, Char>::value>> {
+ private:
+  formatter<T, Char> underlying_;
+  static constexpr basic_string_view<Char> optional =
+      detail::string_literal<Char, 'o', 'p', 't', 'i', 'o', 'n', 'a', 'l',
+                             '('>{};
+  static constexpr basic_string_view<Char> none =
+      detail::string_literal<Char, 'n', 'o', 'n', 'e'>{};
+
+  template <class U>
+  FMT_CONSTEXPR static auto maybe_set_debug_format(U& u, bool set)
+      -> decltype(u.set_debug_format(set)) {
+    u.set_debug_format(set);
+  }
+
+  template <class U>
+  FMT_CONSTEXPR static void maybe_set_debug_format(U&, ...) {}
+
+ public:
+  template <typename ParseContext> FMT_CONSTEXPR auto parse(ParseContext& ctx) {
+    maybe_set_debug_format(underlying_, true);
+    return underlying_.parse(ctx);
+  }
+
+  template <typename FormatContext>
+  auto format(const std::optional<T>& opt, FormatContext& ctx) const
+      -> decltype(ctx.out()) {
+    if (!opt) return detail::write<Char>(ctx.out(), none);
+
+    auto out = ctx.out();
+    out = detail::write<Char>(out, optional);
+    ctx.advance_to(out);
+    out = underlying_.format(*opt, ctx);
+    return detail::write(out, ')');
+  }
+};
+FMT_END_NAMESPACE
+#endif  // __cpp_lib_optional
+
+#if defined(__cpp_lib_expected) || FMT_CPP_LIB_VARIANT
+
+FMT_BEGIN_NAMESPACE
+namespace detail {
+
+template <typename Char, typename OutputIt, typename T>
+auto write_escaped_alternative(OutputIt out, const T& v) -> OutputIt {
+  if constexpr (has_to_string_view<T>::value)
+    return write_escaped_string<Char>(out, detail::to_string_view(v));
+  if constexpr (std::is_same_v<T, Char>) return write_escaped_char(out, v);
+  return write<Char>(out, v);
+}
+
+}  // namespace detail
+
+FMT_END_NAMESPACE
+#endif
+
+#ifdef __cpp_lib_expected
+FMT_BEGIN_NAMESPACE
+
+FMT_EXPORT
+template <typename T, typename E, typename Char>
+struct formatter<std::expected<T, E>, Char,
+                 std::enable_if_t<is_formattable<T, Char>::value &&
+                                  is_formattable<E, Char>::value>> {
+  template <typename ParseContext>
+  FMT_CONSTEXPR auto parse(ParseContext& ctx) -> decltype(ctx.begin()) {
+    return ctx.begin();
+  }
+
+  template <typename FormatContext>
+  auto format(const std::expected<T, E>& value, FormatContext& ctx) const
+      -> decltype(ctx.out()) {
+    auto out = ctx.out();
+
+    if (value.has_value()) {
+      out = detail::write<Char>(out, "expected(");
+      out = detail::write_escaped_alternative<Char>(out, *value);
+    } else {
+      out = detail::write<Char>(out, "unexpected(");
+      out = detail::write_escaped_alternative<Char>(out, value.error());
+    }
+    *out++ = ')';
+    return out;
+  }
+};
+FMT_END_NAMESPACE
+#endif  // __cpp_lib_expected
+
+#ifdef __cpp_lib_source_location
+FMT_BEGIN_NAMESPACE
+FMT_EXPORT
+template <> struct formatter<std::source_location> {
+  template <typename ParseContext> FMT_CONSTEXPR auto parse(ParseContext& ctx) {
+    return ctx.begin();
+  }
+
+  template <typename FormatContext>
+  auto format(const std::source_location& loc, FormatContext& ctx) const
+      -> decltype(ctx.out()) {
+    auto out = ctx.out();
+    out = detail::write(out, loc.file_name());
+    out = detail::write(out, ':');
+    out = detail::write<char>(out, loc.line());
+    out = detail::write(out, ':');
+    out = detail::write<char>(out, loc.column());
+    out = detail::write(out, ": ");
+    out = detail::write(out, loc.function_name());
+    return out;
+  }
+};
+FMT_END_NAMESPACE
+#endif
+
+#if FMT_CPP_LIB_VARIANT
+FMT_BEGIN_NAMESPACE
+namespace detail {
+
+template <typename T>
+using variant_index_sequence =
+    std::make_index_sequence<std::variant_size<T>::value>;
+
+template <typename> struct is_variant_like_ : std::false_type {};
+template <typename... Types>
+struct is_variant_like_<std::variant<Types...>> : std::true_type {};
+
+// formattable element check.
+template <typename T, typename C> class is_variant_formattable_ {
+  template <std::size_t... Is>
+  static std::conjunction<
+      is_formattable<std::variant_alternative_t<Is, T>, C>...>
+      check(std::index_sequence<Is...>);
+
+ public:
+  static constexpr const bool value =
+      decltype(check(variant_index_sequence<T>{}))::value;
+};
+
+}  // namespace detail
+
+template <typename T> struct is_variant_like {
+  static constexpr const bool value = detail::is_variant_like_<T>::value;
+};
+
+template <typename T, typename C> struct is_variant_formattable {
+  static constexpr const bool value =
+      detail::is_variant_formattable_<T, C>::value;
+};
+
+FMT_EXPORT
+template <typename Char> struct formatter<std::monostate, Char> {
+  template <typename ParseContext>
+  FMT_CONSTEXPR auto parse(ParseContext& ctx) -> decltype(ctx.begin()) {
+    return ctx.begin();
+  }
+
+  template <typename FormatContext>
+  auto format(const std::monostate&, FormatContext& ctx) const
+      -> decltype(ctx.out()) {
+    return detail::write<Char>(ctx.out(), "monostate");
+  }
+};
+
+FMT_EXPORT
+template <typename Variant, typename Char>
+struct formatter<
+    Variant, Char,
+    std::enable_if_t<std::conjunction_v<
+        is_variant_like<Variant>, is_variant_formattable<Variant, Char>>>> {
+  template <typename ParseContext>
+  FMT_CONSTEXPR auto parse(ParseContext& ctx) -> decltype(ctx.begin()) {
+    return ctx.begin();
+  }
+
+  template <typename FormatContext>
+  auto format(const Variant& value, FormatContext& ctx) const
+      -> decltype(ctx.out()) {
+    auto out = ctx.out();
+
+    out = detail::write<Char>(out, "variant(");
+    FMT_TRY {
+      std::visit(
+          [&](const auto& v) {
+            out = detail::write_escaped_alternative<Char>(out, v);
+          },
+          value);
+    }
+    FMT_CATCH(const std::bad_variant_access&) {
+      detail::write<Char>(out, "valueless by exception");
+    }
+    *out++ = ')';
+    return out;
+  }
+};
+FMT_END_NAMESPACE
+#endif  // FMT_CPP_LIB_VARIANT
+
+FMT_BEGIN_NAMESPACE
+FMT_EXPORT
+template <typename Char> struct formatter<std::error_code, Char> {
+  template <typename ParseContext>
+  FMT_CONSTEXPR auto parse(ParseContext& ctx) -> decltype(ctx.begin()) {
+    return ctx.begin();
+  }
+
+  template <typename FormatContext>
+  FMT_CONSTEXPR auto format(const std::error_code& ec, FormatContext& ctx) const
+      -> decltype(ctx.out()) {
+    auto out = ctx.out();
+    out = detail::write_bytes<Char>(out, ec.category().name(), format_specs());
+    out = detail::write<Char>(out, Char(':'));
+    out = detail::write<Char>(out, ec.value());
+    return out;
+  }
+};
+
+#if FMT_USE_RTTI
+namespace detail {
+
+template <typename Char, typename OutputIt>
+auto write_demangled_name(OutputIt out, const std::type_info& ti) -> OutputIt {
+#  ifdef FMT_HAS_ABI_CXA_DEMANGLE
+  int status = 0;
+  std::size_t size = 0;
+  std::unique_ptr<char, void (*)(void*)> demangled_name_ptr(
+      abi::__cxa_demangle(ti.name(), nullptr, &size, &status), &std::free);
+
+  string_view demangled_name_view;
+  if (demangled_name_ptr) {
+    demangled_name_view = demangled_name_ptr.get();
+
+    // Normalization of stdlib inline namespace names.
+    // libc++ inline namespaces.
+    //  std::__1::*       -> std::*
+    //  std::__1::__fs::* -> std::*
+    // libstdc++ inline namespaces.
+    //  std::__cxx11::*             -> std::*
+    //  std::filesystem::__cxx11::* -> std::filesystem::*
+    if (demangled_name_view.starts_with("std::")) {
+      char* begin = demangled_name_ptr.get();
+      char* to = begin + 5;  // std::
+      for (char *from = to, *end = begin + demangled_name_view.size();
+           from < end;) {
+        // This is safe, because demangled_name is NUL-terminated.
+        if (from[0] == '_' && from[1] == '_') {
+          char* next = from + 1;
+          while (next < end && *next != ':') next++;
+          if (next[0] == ':' && next[1] == ':') {
+            from = next + 2;
+            continue;
+          }
+        }
+        *to++ = *from++;
+      }
+      demangled_name_view = {begin, detail::to_unsigned(to - begin)};
+    }
+  } else {
+    demangled_name_view = string_view(ti.name());
+  }
+  return detail::write_bytes<Char>(out, demangled_name_view);
+#  elif FMT_MSC_VERSION
+  const string_view demangled_name(ti.name());
+  for (std::size_t i = 0; i < demangled_name.size(); ++i) {
+    auto sub = demangled_name;
+    sub.remove_prefix(i);
+    if (sub.starts_with("enum ")) {
+      i += 4;
+      continue;
+    }
+    if (sub.starts_with("class ") || sub.starts_with("union ")) {
+      i += 5;
+      continue;
+    }
+    if (sub.starts_with("struct ")) {
+      i += 6;
+      continue;
+    }
+    if (*sub.begin() != ' ') *out++ = *sub.begin();
+  }
+  return out;
+#  else
+  return detail::write_bytes<Char>(out, string_view(ti.name()));
+#  endif
+}
+
+}  // namespace detail
+
+FMT_EXPORT
+template <typename Char>
+struct formatter<std::type_info, Char  // DEPRECATED! Mixing code unit types.
+                 > {
+ public:
+  FMT_CONSTEXPR auto parse(basic_format_parse_context<Char>& ctx)
+      -> decltype(ctx.begin()) {
+    return ctx.begin();
+  }
+
+  template <typename Context>
+  auto format(const std::type_info& ti, Context& ctx) const
+      -> decltype(ctx.out()) {
+    return detail::write_demangled_name<Char>(ctx.out(), ti);
+  }
+};
+#endif
+
+FMT_EXPORT
+template <typename T, typename Char>
+struct formatter<
+    T, Char,  // DEPRECATED! Mixing code unit types.
+    typename std::enable_if<std::is_base_of<std::exception, T>::value>::type> {
+ private:
+  bool with_typename_ = false;
+
+ public:
+  FMT_CONSTEXPR auto parse(basic_format_parse_context<Char>& ctx)
+      -> decltype(ctx.begin()) {
+    auto it = ctx.begin();
+    auto end = ctx.end();
+    if (it == end || *it == '}') return it;
+    if (*it == 't') {
+      ++it;
+      with_typename_ = FMT_USE_RTTI != 0;
+    }
+    return it;
+  }
+
+  template <typename Context>
+  auto format(const std::exception& ex, Context& ctx) const
+      -> decltype(ctx.out()) {
+    auto out = ctx.out();
+#if FMT_USE_RTTI
+    if (with_typename_) {
+      out = detail::write_demangled_name<Char>(out, typeid(ex));
+      *out++ = ':';
+      *out++ = ' ';
+    }
+#endif
+    return detail::write_bytes<Char>(out, string_view(ex.what()));
+  }
+};
+
+namespace detail {
+
+template <typename T, typename Enable = void>
+struct has_flip : std::false_type {};
+
+template <typename T>
+struct has_flip<T, void_t<decltype(std::declval<T>().flip())>>
+    : std::true_type {};
+
+template <typename T> struct is_bit_reference_like {
+  static constexpr const bool value =
+      std::is_convertible<T, bool>::value &&
+      std::is_nothrow_assignable<T, bool>::value && has_flip<T>::value;
+};
+
+#ifdef _LIBCPP_VERSION
+
+// Workaround for libc++ incompatibility with C++ standard.
+// According to the Standard, `bitset::operator[] const` returns bool.
+template <typename C>
+struct is_bit_reference_like<std::__bit_const_reference<C>> {
+  static constexpr const bool value = true;
+};
+
+#endif
+
+}  // namespace detail
+
+// We can't use std::vector<bool, Allocator>::reference and
+// std::bitset<N>::reference because the compiler can't deduce Allocator and N
+// in partial specialization.
+FMT_EXPORT
+template <typename BitRef, typename Char>
+struct formatter<BitRef, Char,
+                 enable_if_t<detail::is_bit_reference_like<BitRef>::value>>
+    : formatter<bool, Char> {
+  template <typename FormatContext>
+  FMT_CONSTEXPR auto format(const BitRef& v, FormatContext& ctx) const
+      -> decltype(ctx.out()) {
+    return formatter<bool, Char>::format(v, ctx);
+  }
+};
+
+template <typename T, typename Deleter>
+auto ptr(const std::unique_ptr<T, Deleter>& p) -> const void* {
+  return p.get();
+}
+template <typename T> auto ptr(const std::shared_ptr<T>& p) -> const void* {
+  return p.get();
+}
+
+FMT_EXPORT
+template <typename T, typename Char>
+struct formatter<std::atomic<T>, Char,
+                 enable_if_t<is_formattable<T, Char>::value>>
+    : formatter<T, Char> {
+  template <typename FormatContext>
+  auto format(const std::atomic<T>& v, FormatContext& ctx) const
+      -> decltype(ctx.out()) {
+    return formatter<T, Char>::format(v.load(), ctx);
+  }
+};
+
+#ifdef __cpp_lib_atomic_flag_test
+FMT_EXPORT
+template <typename Char>
+struct formatter<std::atomic_flag, Char> : formatter<bool, Char> {
+  template <typename FormatContext>
+  auto format(const std::atomic_flag& v, FormatContext& ctx) const
+      -> decltype(ctx.out()) {
+    return formatter<bool, Char>::format(v.test(), ctx);
+  }
+};
+#endif  // __cpp_lib_atomic_flag_test
+
+FMT_EXPORT
+template <typename T, typename Char> struct formatter<std::complex<T>, Char> {
+ private:
+  detail::dynamic_format_specs<Char> specs_;
+
+  template <typename FormatContext, typename OutputIt>
+  FMT_CONSTEXPR auto do_format(const std::complex<T>& c,
+                               detail::dynamic_format_specs<Char>& specs,
+                               FormatContext& ctx, OutputIt out) const
+      -> OutputIt {
+    if (c.real() != 0) {
+      *out++ = Char('(');
+      out = detail::write<Char>(out, c.real(), specs, ctx.locale());
+      specs.sign = sign::plus;
+      out = detail::write<Char>(out, c.imag(), specs, ctx.locale());
+      if (!detail::isfinite(c.imag())) *out++ = Char(' ');
+      *out++ = Char('i');
+      *out++ = Char(')');
+      return out;
+    }
+    out = detail::write<Char>(out, c.imag(), specs, ctx.locale());
+    if (!detail::isfinite(c.imag())) *out++ = Char(' ');
+    *out++ = Char('i');
+    return out;
+  }
+
+ public:
+  FMT_CONSTEXPR auto parse(basic_format_parse_context<Char>& ctx)
+      -> decltype(ctx.begin()) {
+    if (ctx.begin() == ctx.end() || *ctx.begin() == '}') return ctx.begin();
+    return parse_format_specs(ctx.begin(), ctx.end(), specs_, ctx,
+                              detail::type_constant<T, Char>::value);
+  }
+
+  template <typename FormatContext>
+  auto format(const std::complex<T>& c, FormatContext& ctx) const
+      -> decltype(ctx.out()) {
+    auto specs = specs_;
+    if (specs.width_ref.kind != detail::arg_id_kind::none ||
+        specs.precision_ref.kind != detail::arg_id_kind::none) {
+      detail::handle_dynamic_spec<detail::width_checker>(specs.width,
+                                                         specs.width_ref, ctx);
+      detail::handle_dynamic_spec<detail::precision_checker>(
+          specs.precision, specs.precision_ref, ctx);
+    }
+
+    if (specs.width == 0) return do_format(c, specs, ctx, ctx.out());
+    auto buf = basic_memory_buffer<Char>();
+
+    auto outer_specs = format_specs();
+    outer_specs.width = specs.width;
+    outer_specs.fill = specs.fill;
+    outer_specs.align = specs.align;
+
+    specs.width = 0;
+    specs.fill = {};
+    specs.align = align::none;
+
+    do_format(c, specs, ctx, basic_appender<Char>(buf));
+    return detail::write<Char>(ctx.out(),
+                               basic_string_view<Char>(buf.data(), buf.size()),
+                               outer_specs);
+  }
+};
+
+FMT_END_NAMESPACE
+#endif  // FMT_STD_H_
diff --git a/ext/spdlog/fmt/bundled/xchar.h b/ext/spdlog/fmt/bundled/xchar.h
new file mode 100644
index 0000000..b1f39ed
--- /dev/null
+++ b/ext/spdlog/fmt/bundled/xchar.h
@@ -0,0 +1,322 @@
+// Formatting library for C++ - optional wchar_t and exotic character support
+//
+// Copyright (c) 2012 - present, Victor Zverovich
+// All rights reserved.
+//
+// For the license information refer to format.h.
+
+#ifndef FMT_XCHAR_H_
+#define FMT_XCHAR_H_
+
+#include "color.h"
+#include "format.h"
+#include "ranges.h"
+
+#ifndef FMT_MODULE
+#  include <cwchar>
+#  if !defined(FMT_STATIC_THOUSANDS_SEPARATOR)
+#    include <locale>
+#  endif
+#endif
+
+FMT_BEGIN_NAMESPACE
+namespace detail {
+
+template <typename T>
+using is_exotic_char = bool_constant<!std::is_same<T, char>::value>;
+
+template <typename S, typename = void> struct format_string_char {};
+
+template <typename S>
+struct format_string_char<
+    S, void_t<decltype(sizeof(detail::to_string_view(std::declval<S>())))>> {
+  using type = char_t<S>;
+};
+
+template <typename S>
+struct format_string_char<S, enable_if_t<is_compile_string<S>::value>> {
+  using type = typename S::char_type;
+};
+
+template <typename S>
+using format_string_char_t = typename format_string_char<S>::type;
+
+inline auto write_loc(basic_appender<wchar_t> out, loc_value value,
+                      const format_specs& specs, locale_ref loc) -> bool {
+#ifndef FMT_STATIC_THOUSANDS_SEPARATOR
+  auto& numpunct =
+      std::use_facet<std::numpunct<wchar_t>>(loc.get<std::locale>());
+  auto separator = std::wstring();
+  auto grouping = numpunct.grouping();
+  if (!grouping.empty()) separator = std::wstring(1, numpunct.thousands_sep());
+  return value.visit(loc_writer<wchar_t>{out, specs, separator, grouping, {}});
+#endif
+  return false;
+}
+}  // namespace detail
+
+FMT_BEGIN_EXPORT
+
+using wstring_view = basic_string_view<wchar_t>;
+using wformat_parse_context = basic_format_parse_context<wchar_t>;
+using wformat_context = buffered_context<wchar_t>;
+using wformat_args = basic_format_args<wformat_context>;
+using wmemory_buffer = basic_memory_buffer<wchar_t>;
+
+#if FMT_GCC_VERSION && FMT_GCC_VERSION < 409
+// Workaround broken conversion on older gcc.
+template <typename... Args> using wformat_string = wstring_view;
+inline auto runtime(wstring_view s) -> wstring_view { return s; }
+#else
+template <typename... Args>
+using wformat_string = basic_format_string<wchar_t, type_identity_t<Args>...>;
+inline auto runtime(wstring_view s) -> runtime_format_string<wchar_t> {
+  return {{s}};
+}
+#endif
+
+template <> struct is_char<wchar_t> : std::true_type {};
+template <> struct is_char<char16_t> : std::true_type {};
+template <> struct is_char<char32_t> : std::true_type {};
+
+#ifdef __cpp_char8_t
+template <>
+struct is_char<char8_t> : bool_constant<detail::is_utf8_enabled()> {};
+#endif
+
+template <typename... T>
+constexpr auto make_wformat_args(T&... args)
+    -> decltype(fmt::make_format_args<wformat_context>(args...)) {
+  return fmt::make_format_args<wformat_context>(args...);
+}
+
+inline namespace literals {
+#if FMT_USE_USER_DEFINED_LITERALS && !FMT_USE_NONTYPE_TEMPLATE_ARGS
+constexpr auto operator""_a(const wchar_t* s, size_t)
+    -> detail::udl_arg<wchar_t> {
+  return {s};
+}
+#endif
+}  // namespace literals
+
+template <typename It, typename Sentinel>
+auto join(It begin, Sentinel end, wstring_view sep)
+    -> join_view<It, Sentinel, wchar_t> {
+  return {begin, end, sep};
+}
+
+template <typename Range>
+auto join(Range&& range, wstring_view sep)
+    -> join_view<detail::iterator_t<Range>, detail::sentinel_t<Range>,
+                 wchar_t> {
+  return join(std::begin(range), std::end(range), sep);
+}
+
+template <typename T>
+auto join(std::initializer_list<T> list, wstring_view sep)
+    -> join_view<const T*, const T*, wchar_t> {
+  return join(std::begin(list), std::end(list), sep);
+}
+
+template <typename... T>
+auto join(const std::tuple<T...>& tuple, basic_string_view<wchar_t> sep)
+    -> tuple_join_view<wchar_t, T...> {
+  return {tuple, sep};
+}
+
+template <typename Char, FMT_ENABLE_IF(!std::is_same<Char, char>::value)>
+auto vformat(basic_string_view<Char> format_str,
+             typename detail::vformat_args<Char>::type args)
+    -> std::basic_string<Char> {
+  auto buf = basic_memory_buffer<Char>();
+  detail::vformat_to(buf, format_str, args);
+  return to_string(buf);
+}
+
+template <typename... T>
+auto format(wformat_string<T...> fmt, T&&... args) -> std::wstring {
+  return vformat(fmt::wstring_view(fmt), fmt::make_wformat_args(args...));
+}
+
+template <typename OutputIt, typename... T>
+auto format_to(OutputIt out, wformat_string<T...> fmt, T&&... args)
+    -> OutputIt {
+  return vformat_to(out, fmt::wstring_view(fmt),
+                    fmt::make_wformat_args(args...));
+}
+
+// Pass char_t as a default template parameter instead of using
+// std::basic_string<char_t<S>> to reduce the symbol size.
+template <typename S, typename... T,
+          typename Char = detail::format_string_char_t<S>,
+          FMT_ENABLE_IF(!std::is_same<Char, char>::value &&
+                        !std::is_same<Char, wchar_t>::value)>
+auto format(const S& format_str, T&&... args) -> std::basic_string<Char> {
+  return vformat(detail::to_string_view(format_str),
+                 fmt::make_format_args<buffered_context<Char>>(args...));
+}
+
+template <typename Locale, typename S,
+          typename Char = detail::format_string_char_t<S>,
+          FMT_ENABLE_IF(detail::is_locale<Locale>::value&&
+                            detail::is_exotic_char<Char>::value)>
+inline auto vformat(const Locale& loc, const S& format_str,
+                    typename detail::vformat_args<Char>::type args)
+    -> std::basic_string<Char> {
+  return detail::vformat(loc, detail::to_string_view(format_str), args);
+}
+
+template <typename Locale, typename S, typename... T,
+          typename Char = detail::format_string_char_t<S>,
+          FMT_ENABLE_IF(detail::is_locale<Locale>::value&&
+                            detail::is_exotic_char<Char>::value)>
+inline auto format(const Locale& loc, const S& format_str, T&&... args)
+    -> std::basic_string<Char> {
+  return detail::vformat(
+      loc, detail::to_string_view(format_str),
+      fmt::make_format_args<buffered_context<Char>>(args...));
+}
+
+template <typename OutputIt, typename S,
+          typename Char = detail::format_string_char_t<S>,
+          FMT_ENABLE_IF(detail::is_output_iterator<OutputIt, Char>::value&&
+                            detail::is_exotic_char<Char>::value)>
+auto vformat_to(OutputIt out, const S& format_str,
+                typename detail::vformat_args<Char>::type args) -> OutputIt {
+  auto&& buf = detail::get_buffer<Char>(out);
+  detail::vformat_to(buf, detail::to_string_view(format_str), args);
+  return detail::get_iterator(buf, out);
+}
+
+template <typename OutputIt, typename S, typename... T,
+          typename Char = detail::format_string_char_t<S>,
+          FMT_ENABLE_IF(detail::is_output_iterator<OutputIt, Char>::value &&
+                        !std::is_same<Char, char>::value &&
+                        !std::is_same<Char, wchar_t>::value)>
+inline auto format_to(OutputIt out, const S& fmt, T&&... args) -> OutputIt {
+  return vformat_to(out, detail::to_string_view(fmt),
+                    fmt::make_format_args<buffered_context<Char>>(args...));
+}
+
+template <typename Locale, typename S, typename OutputIt, typename... Args,
+          typename Char = detail::format_string_char_t<S>,
+          FMT_ENABLE_IF(detail::is_output_iterator<OutputIt, Char>::value&&
+                            detail::is_locale<Locale>::value&&
+                                detail::is_exotic_char<Char>::value)>
+inline auto vformat_to(OutputIt out, const Locale& loc, const S& format_str,
+                       typename detail::vformat_args<Char>::type args)
+    -> OutputIt {
+  auto&& buf = detail::get_buffer<Char>(out);
+  vformat_to(buf, detail::to_string_view(format_str), args,
+             detail::locale_ref(loc));
+  return detail::get_iterator(buf, out);
+}
+
+template <typename OutputIt, typename Locale, typename S, typename... T,
+          typename Char = detail::format_string_char_t<S>,
+          bool enable = detail::is_output_iterator<OutputIt, Char>::value &&
+                        detail::is_locale<Locale>::value &&
+                        detail::is_exotic_char<Char>::value>
+inline auto format_to(OutputIt out, const Locale& loc, const S& format_str,
+                      T&&... args) ->
+    typename std::enable_if<enable, OutputIt>::type {
+  return vformat_to(out, loc, detail::to_string_view(format_str),
+                    fmt::make_format_args<buffered_context<Char>>(args...));
+}
+
+template <typename OutputIt, typename Char, typename... Args,
+          FMT_ENABLE_IF(detail::is_output_iterator<OutputIt, Char>::value&&
+                            detail::is_exotic_char<Char>::value)>
+inline auto vformat_to_n(OutputIt out, size_t n,
+                         basic_string_view<Char> format_str,
+                         typename detail::vformat_args<Char>::type args)
+    -> format_to_n_result<OutputIt> {
+  using traits = detail::fixed_buffer_traits;
+  auto buf = detail::iterator_buffer<OutputIt, Char, traits>(out, n);
+  detail::vformat_to(buf, format_str, args);
+  return {buf.out(), buf.count()};
+}
+
+template <typename OutputIt, typename S, typename... T,
+          typename Char = detail::format_string_char_t<S>,
+          FMT_ENABLE_IF(detail::is_output_iterator<OutputIt, Char>::value&&
+                            detail::is_exotic_char<Char>::value)>
+inline auto format_to_n(OutputIt out, size_t n, const S& fmt, T&&... args)
+    -> format_to_n_result<OutputIt> {
+  return vformat_to_n(out, n, fmt::basic_string_view<Char>(fmt),
+                      fmt::make_format_args<buffered_context<Char>>(args...));
+}
+
+template <typename S, typename... T,
+          typename Char = detail::format_string_char_t<S>,
+          FMT_ENABLE_IF(detail::is_exotic_char<Char>::value)>
+inline auto formatted_size(const S& fmt, T&&... args) -> size_t {
+  auto buf = detail::counting_buffer<Char>();
+  detail::vformat_to(buf, detail::to_string_view(fmt),
+                     fmt::make_format_args<buffered_context<Char>>(args...));
+  return buf.count();
+}
+
+inline void vprint(std::FILE* f, wstring_view fmt, wformat_args args) {
+  auto buf = wmemory_buffer();
+  detail::vformat_to(buf, fmt, args);
+  buf.push_back(L'\0');
+  if (std::fputws(buf.data(), f) == -1)
+    FMT_THROW(system_error(errno, FMT_STRING("cannot write to file")));
+}
+
+inline void vprint(wstring_view fmt, wformat_args args) {
+  vprint(stdout, fmt, args);
+}
+
+template <typename... T>
+void print(std::FILE* f, wformat_string<T...> fmt, T&&... args) {
+  return vprint(f, wstring_view(fmt), fmt::make_wformat_args(args...));
+}
+
+template <typename... T> void print(wformat_string<T...> fmt, T&&... args) {
+  return vprint(wstring_view(fmt), fmt::make_wformat_args(args...));
+}
+
+template <typename... T>
+void println(std::FILE* f, wformat_string<T...> fmt, T&&... args) {
+  return print(f, L"{}\n", fmt::format(fmt, std::forward<T>(args)...));
+}
+
+template <typename... T> void println(wformat_string<T...> fmt, T&&... args) {
+  return print(L"{}\n", fmt::format(fmt, std::forward<T>(args)...));
+}
+
+inline auto vformat(const text_style& ts, wstring_view fmt, wformat_args args)
+    -> std::wstring {
+  auto buf = wmemory_buffer();
+  detail::vformat_to(buf, ts, fmt, args);
+  return fmt::to_string(buf);
+}
+
+template <typename... T>
+inline auto format(const text_style& ts, wformat_string<T...> fmt, T&&... args)
+    -> std::wstring {
+  return fmt::vformat(ts, fmt, fmt::make_wformat_args(args...));
+}
+
+template <typename... T>
+FMT_DEPRECATED void print(std::FILE* f, const text_style& ts,
+                          wformat_string<T...> fmt, const T&... args) {
+  vprint(f, ts, fmt, fmt::make_wformat_args(args...));
+}
+
+template <typename... T>
+FMT_DEPRECATED void print(const text_style& ts, wformat_string<T...> fmt,
+                          const T&... args) {
+  return print(stdout, ts, fmt, args...);
+}
+
+/// Converts `value` to `std::wstring` using the default format for type `T`.
+template <typename T> inline auto to_wstring(const T& value) -> std::wstring {
+  return format(FMT_STRING(L"{}"), value);
+}
+FMT_END_EXPORT
+FMT_END_NAMESPACE
+
+#endif  // FMT_XCHAR_H_
diff --git a/ext/spdlog/fmt/chrono.h b/ext/spdlog/fmt/chrono.h
new file mode 100644
index 0000000..a72a5bd
--- /dev/null
+++ b/ext/spdlog/fmt/chrono.h
@@ -0,0 +1,23 @@
+//
+// Copyright(c) 2016 Gabi Melman.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+//
+
+#pragma once
+//
+// include bundled or external copy of fmtlib's chrono support
+//
+#include <spdlog/tweakme.h>
+
+#if !defined(SPDLOG_USE_STD_FORMAT)
+    #if !defined(SPDLOG_FMT_EXTERNAL)
+        #ifdef SPDLOG_HEADER_ONLY
+            #ifndef FMT_HEADER_ONLY
+                #define FMT_HEADER_ONLY
+            #endif
+        #endif
+        #include <spdlog/fmt/bundled/chrono.h>
+    #else
+        #include <fmt/chrono.h>
+    #endif
+#endif
diff --git a/ext/spdlog/fmt/compile.h b/ext/spdlog/fmt/compile.h
new file mode 100644
index 0000000..3c9c25d
--- /dev/null
+++ b/ext/spdlog/fmt/compile.h
@@ -0,0 +1,23 @@
+//
+// Copyright(c) 2016 Gabi Melman.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+//
+
+#pragma once
+//
+// include bundled or external copy of fmtlib's compile-time support
+//
+#include <spdlog/tweakme.h>
+
+#if !defined(SPDLOG_USE_STD_FORMAT)
+    #if !defined(SPDLOG_FMT_EXTERNAL)
+        #ifdef SPDLOG_HEADER_ONLY
+            #ifndef FMT_HEADER_ONLY
+                #define FMT_HEADER_ONLY
+            #endif
+        #endif
+        #include <spdlog/fmt/bundled/compile.h>
+    #else
+        #include <fmt/compile.h>
+    #endif
+#endif
diff --git a/ext/spdlog/fmt/fmt.h b/ext/spdlog/fmt/fmt.h
new file mode 100644
index 0000000..7fa6b09
--- /dev/null
+++ b/ext/spdlog/fmt/fmt.h
@@ -0,0 +1,30 @@
+//
+// Copyright(c) 2016-2018 Gabi Melman.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+//
+
+#pragma once
+
+//
+// Include a bundled header-only copy of fmtlib or an external one.
+// By default spdlog include its own copy.
+//
+#include <spdlog/tweakme.h>
+
+#if defined(SPDLOG_USE_STD_FORMAT)  // SPDLOG_USE_STD_FORMAT is defined - use std::format
+    #include <format>
+#elif !defined(SPDLOG_FMT_EXTERNAL)
+    #if !defined(SPDLOG_COMPILED_LIB) && !defined(FMT_HEADER_ONLY)
+        #define FMT_HEADER_ONLY
+    #endif
+    #ifndef FMT_USE_WINDOWS_H
+        #define FMT_USE_WINDOWS_H 0
+    #endif
+
+    #include <spdlog/fmt/bundled/core.h>
+    #include <spdlog/fmt/bundled/format.h>
+
+#else  // SPDLOG_FMT_EXTERNAL is defined - use external fmtlib
+    #include <fmt/core.h>
+    #include <fmt/format.h>
+#endif
diff --git a/ext/spdlog/fmt/ostr.h b/ext/spdlog/fmt/ostr.h
new file mode 100644
index 0000000..2b90105
--- /dev/null
+++ b/ext/spdlog/fmt/ostr.h
@@ -0,0 +1,23 @@
+//
+// Copyright(c) 2016 Gabi Melman.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+//
+
+#pragma once
+//
+// include bundled or external copy of fmtlib's ostream support
+//
+#include <spdlog/tweakme.h>
+
+#if !defined(SPDLOG_USE_STD_FORMAT)
+    #if !defined(SPDLOG_FMT_EXTERNAL)
+        #ifdef SPDLOG_HEADER_ONLY
+            #ifndef FMT_HEADER_ONLY
+                #define FMT_HEADER_ONLY
+            #endif
+        #endif
+        #include <spdlog/fmt/bundled/ostream.h>
+    #else
+        #include <fmt/ostream.h>
+    #endif
+#endif
diff --git a/ext/spdlog/fmt/ranges.h b/ext/spdlog/fmt/ranges.h
new file mode 100644
index 0000000..5bb91e9
--- /dev/null
+++ b/ext/spdlog/fmt/ranges.h
@@ -0,0 +1,23 @@
+//
+// Copyright(c) 2016 Gabi Melman.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+//
+
+#pragma once
+//
+// include bundled or external copy of fmtlib's ranges support
+//
+#include <spdlog/tweakme.h>
+
+#if !defined(SPDLOG_USE_STD_FORMAT)
+    #if !defined(SPDLOG_FMT_EXTERNAL)
+        #ifdef SPDLOG_HEADER_ONLY
+            #ifndef FMT_HEADER_ONLY
+                #define FMT_HEADER_ONLY
+            #endif
+        #endif
+        #include <spdlog/fmt/bundled/ranges.h>
+    #else
+        #include <fmt/ranges.h>
+    #endif
+#endif
diff --git a/ext/spdlog/fmt/std.h b/ext/spdlog/fmt/std.h
new file mode 100644
index 0000000..dabe6f6
--- /dev/null
+++ b/ext/spdlog/fmt/std.h
@@ -0,0 +1,24 @@
+//
+// Copyright(c) 2016 Gabi Melman.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+//
+
+#pragma once
+//
+// include bundled or external copy of fmtlib's std support (for formatting e.g.
+// std::filesystem::path, std::thread::id, std::monostate, std::variant, ...)
+//
+#include <spdlog/tweakme.h>
+
+#if !defined(SPDLOG_USE_STD_FORMAT)
+    #if !defined(SPDLOG_FMT_EXTERNAL)
+        #ifdef SPDLOG_HEADER_ONLY
+            #ifndef FMT_HEADER_ONLY
+                #define FMT_HEADER_ONLY
+            #endif
+        #endif
+        #include <spdlog/fmt/bundled/std.h>
+    #else
+        #include <fmt/std.h>
+    #endif
+#endif
diff --git a/ext/spdlog/fmt/xchar.h b/ext/spdlog/fmt/xchar.h
new file mode 100644
index 0000000..2525f05
--- /dev/null
+++ b/ext/spdlog/fmt/xchar.h
@@ -0,0 +1,23 @@
+//
+// Copyright(c) 2016 Gabi Melman.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+//
+
+#pragma once
+//
+// include bundled or external copy of fmtlib's xchar support
+//
+#include <spdlog/tweakme.h>
+
+#if !defined(SPDLOG_USE_STD_FORMAT)
+    #if !defined(SPDLOG_FMT_EXTERNAL)
+        #ifdef SPDLOG_HEADER_ONLY
+            #ifndef FMT_HEADER_ONLY
+                #define FMT_HEADER_ONLY
+            #endif
+        #endif
+        #include <spdlog/fmt/bundled/xchar.h>
+    #else
+        #include <fmt/xchar.h>
+    #endif
+#endif
diff --git a/ext/spdlog/formatter.h b/ext/spdlog/formatter.h
new file mode 100644
index 0000000..4d482f8
--- /dev/null
+++ b/ext/spdlog/formatter.h
@@ -0,0 +1,17 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#include <spdlog/details/log_msg.h>
+#include <spdlog/fmt/fmt.h>
+
+namespace spdlog {
+
+class formatter {
+public:
+    virtual ~formatter() = default;
+    virtual void format(const details::log_msg &msg, memory_buf_t &dest) = 0;
+    virtual std::unique_ptr<formatter> clone() const = 0;
+};
+}  // namespace spdlog
diff --git a/ext/spdlog/fwd.h b/ext/spdlog/fwd.h
new file mode 100644
index 0000000..647b16b
--- /dev/null
+++ b/ext/spdlog/fwd.h
@@ -0,0 +1,18 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+namespace spdlog {
+class logger;
+class formatter;
+
+namespace sinks {
+class sink;
+}
+
+namespace level {
+enum level_enum : int;
+}
+
+}  // namespace spdlog
diff --git a/ext/spdlog/logger-inl.h b/ext/spdlog/logger-inl.h
new file mode 100644
index 0000000..5218fe4
--- /dev/null
+++ b/ext/spdlog/logger-inl.h
@@ -0,0 +1,198 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#ifndef SPDLOG_HEADER_ONLY
+    #include <spdlog/logger.h>
+#endif
+
+#include <spdlog/details/backtracer.h>
+#include <spdlog/pattern_formatter.h>
+#include <spdlog/sinks/sink.h>
+
+#include <cstdio>
+
+namespace spdlog {
+
+// public methods
+SPDLOG_INLINE logger::logger(const logger &other)
+    : name_(other.name_),
+      sinks_(other.sinks_),
+      level_(other.level_.load(std::memory_order_relaxed)),
+      flush_level_(other.flush_level_.load(std::memory_order_relaxed)),
+      custom_err_handler_(other.custom_err_handler_),
+      tracer_(other.tracer_) {}
+
+SPDLOG_INLINE logger::logger(logger &&other) SPDLOG_NOEXCEPT
+    : name_(std::move(other.name_)),
+      sinks_(std::move(other.sinks_)),
+      level_(other.level_.load(std::memory_order_relaxed)),
+      flush_level_(other.flush_level_.load(std::memory_order_relaxed)),
+      custom_err_handler_(std::move(other.custom_err_handler_)),
+      tracer_(std::move(other.tracer_))
+
+{}
+
+SPDLOG_INLINE logger &logger::operator=(logger other) SPDLOG_NOEXCEPT {
+    this->swap(other);
+    return *this;
+}
+
+SPDLOG_INLINE void logger::swap(spdlog::logger &other) SPDLOG_NOEXCEPT {
+    name_.swap(other.name_);
+    sinks_.swap(other.sinks_);
+
+    // swap level_
+    auto other_level = other.level_.load();
+    auto my_level = level_.exchange(other_level);
+    other.level_.store(my_level);
+
+    // swap flush level_
+    other_level = other.flush_level_.load();
+    my_level = flush_level_.exchange(other_level);
+    other.flush_level_.store(my_level);
+
+    custom_err_handler_.swap(other.custom_err_handler_);
+    std::swap(tracer_, other.tracer_);
+}
+
+SPDLOG_INLINE void swap(logger &a, logger &b) { a.swap(b); }
+
+SPDLOG_INLINE void logger::set_level(level::level_enum log_level) { level_.store(log_level); }
+
+SPDLOG_INLINE level::level_enum logger::level() const {
+    return static_cast<level::level_enum>(level_.load(std::memory_order_relaxed));
+}
+
+SPDLOG_INLINE const std::string &logger::name() const { return name_; }
+
+// set formatting for the sinks in this logger.
+// each sink will get a separate instance of the formatter object.
+SPDLOG_INLINE void logger::set_formatter(std::unique_ptr<formatter> f) {
+    for (auto it = sinks_.begin(); it != sinks_.end(); ++it) {
+        if (std::next(it) == sinks_.end()) {
+            // last element - we can be move it.
+            (*it)->set_formatter(std::move(f));
+            break;  // to prevent clang-tidy warning
+        } else {
+            (*it)->set_formatter(f->clone());
+        }
+    }
+}
+
+SPDLOG_INLINE void logger::set_pattern(std::string pattern, pattern_time_type time_type) {
+    auto new_formatter = details::make_unique<pattern_formatter>(std::move(pattern), time_type);
+    set_formatter(std::move(new_formatter));
+}
+
+// create new backtrace sink and move to it all our child sinks
+SPDLOG_INLINE void logger::enable_backtrace(size_t n_messages) { tracer_.enable(n_messages); }
+
+// restore orig sinks and level and delete the backtrace sink
+SPDLOG_INLINE void logger::disable_backtrace() { tracer_.disable(); }
+
+SPDLOG_INLINE void logger::dump_backtrace() { dump_backtrace_(); }
+
+// flush functions
+SPDLOG_INLINE void logger::flush() { flush_(); }
+
+SPDLOG_INLINE void logger::flush_on(level::level_enum log_level) { flush_level_.store(log_level); }
+
+SPDLOG_INLINE level::level_enum logger::flush_level() const {
+    return static_cast<level::level_enum>(flush_level_.load(std::memory_order_relaxed));
+}
+
+// sinks
+SPDLOG_INLINE const std::vector<sink_ptr> &logger::sinks() const { return sinks_; }
+
+SPDLOG_INLINE std::vector<sink_ptr> &logger::sinks() { return sinks_; }
+
+// error handler
+SPDLOG_INLINE void logger::set_error_handler(err_handler handler) {
+    custom_err_handler_ = std::move(handler);
+}
+
+// create new logger with same sinks and configuration.
+SPDLOG_INLINE std::shared_ptr<logger> logger::clone(std::string logger_name) {
+    auto cloned = std::make_shared<logger>(*this);
+    cloned->name_ = std::move(logger_name);
+    return cloned;
+}
+
+// protected methods
+SPDLOG_INLINE void logger::log_it_(const spdlog::details::log_msg &log_msg,
+                                   bool log_enabled,
+                                   bool traceback_enabled) {
+    if (log_enabled) {
+        sink_it_(log_msg);
+    }
+    if (traceback_enabled) {
+        tracer_.push_back(log_msg);
+    }
+}
+
+SPDLOG_INLINE void logger::sink_it_(const details::log_msg &msg) {
+    for (auto &sink : sinks_) {
+        if (sink->should_log(msg.level)) {
+            SPDLOG_TRY { sink->log(msg); }
+            SPDLOG_LOGGER_CATCH(msg.source)
+        }
+    }
+
+    if (should_flush_(msg)) {
+        flush_();
+    }
+}
+
+SPDLOG_INLINE void logger::flush_() {
+    for (auto &sink : sinks_) {
+        SPDLOG_TRY { sink->flush(); }
+        SPDLOG_LOGGER_CATCH(source_loc())
+    }
+}
+
+SPDLOG_INLINE void logger::dump_backtrace_() {
+    using details::log_msg;
+    if (tracer_.enabled() && !tracer_.empty()) {
+        sink_it_(
+            log_msg{name(), level::info, "****************** Backtrace Start ******************"});
+        tracer_.foreach_pop([this](const log_msg &msg) { this->sink_it_(msg); });
+        sink_it_(
+            log_msg{name(), level::info, "****************** Backtrace End ********************"});
+    }
+}
+
+SPDLOG_INLINE bool logger::should_flush_(const details::log_msg &msg) {
+    auto flush_level = flush_level_.load(std::memory_order_relaxed);
+    return (msg.level >= flush_level) && (msg.level != level::off);
+}
+
+SPDLOG_INLINE void logger::err_handler_(const std::string &msg) {
+    if (custom_err_handler_) {
+        custom_err_handler_(msg);
+    } else {
+        using std::chrono::system_clock;
+        static std::mutex mutex;
+        static std::chrono::system_clock::time_point last_report_time;
+        static size_t err_counter = 0;
+        std::lock_guard<std::mutex> lk{mutex};
+        auto now = system_clock::now();
+        err_counter++;
+        if (now - last_report_time < std::chrono::seconds(1)) {
+            return;
+        }
+        last_report_time = now;
+        auto tm_time = details::os::localtime(system_clock::to_time_t(now));
+        char date_buf[64];
+        std::strftime(date_buf, sizeof(date_buf), "%Y-%m-%d %H:%M:%S", &tm_time);
+#if defined(USING_R) && defined(R_R_H)  // if in R environment
+        REprintf("[*** LOG ERROR #%04zu ***] [%s] [%s] %s\n", err_counter, date_buf, name().c_str(),
+                 msg.c_str());
+#else
+        std::fprintf(stderr, "[*** LOG ERROR #%04zu ***] [%s] [%s] %s\n", err_counter, date_buf,
+                     name().c_str(), msg.c_str());
+#endif
+    }
+}
+}  // namespace spdlog
diff --git a/ext/spdlog/logger.h b/ext/spdlog/logger.h
new file mode 100644
index 0000000..f49bdc0
--- /dev/null
+++ b/ext/spdlog/logger.h
@@ -0,0 +1,379 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+// Thread safe logger (except for set_error_handler())
+// Has name, log level, vector of std::shared sink pointers and formatter
+// Upon each log write the logger:
+// 1. Checks if its log level is enough to log the message and if yes:
+// 2. Call the underlying sinks to do the job.
+// 3. Each sink use its own private copy of a formatter to format the message
+// and send to its destination.
+//
+// The use of private formatter per sink provides the opportunity to cache some
+// formatted data, and support for different format per sink.
+
+#include <spdlog/common.h>
+#include <spdlog/details/backtracer.h>
+#include <spdlog/details/log_msg.h>
+
+#ifdef SPDLOG_WCHAR_TO_UTF8_SUPPORT
+    #ifndef _WIN32
+        #error SPDLOG_WCHAR_TO_UTF8_SUPPORT only supported on windows
+    #endif
+    #include <spdlog/details/os.h>
+#endif
+
+#include <vector>
+
+#ifndef SPDLOG_NO_EXCEPTIONS
+    #define SPDLOG_LOGGER_CATCH(location)                                                 \
+        catch (const std::exception &ex) {                                                \
+            if (location.filename) {                                                      \
+                err_handler_(fmt_lib::format(SPDLOG_FMT_STRING("{} [{}({})]"), ex.what(), \
+                                             location.filename, location.line));          \
+            } else {                                                                      \
+                err_handler_(ex.what());                                                  \
+            }                                                                             \
+        }                                                                                 \
+        catch (...) {                                                                     \
+            err_handler_("Rethrowing unknown exception in logger");                       \
+            throw;                                                                        \
+        }
+#else
+    #define SPDLOG_LOGGER_CATCH(location)
+#endif
+
+namespace spdlog {
+
+class SPDLOG_API logger {
+public:
+    // Empty logger
+    explicit logger(std::string name)
+        : name_(std::move(name)),
+          sinks_() {}
+
+    // Logger with range on sinks
+    template <typename It>
+    logger(std::string name, It begin, It end)
+        : name_(std::move(name)),
+          sinks_(begin, end) {}
+
+    // Logger with single sink
+    logger(std::string name, sink_ptr single_sink)
+        : logger(std::move(name), {std::move(single_sink)}) {}
+
+    // Logger with sinks init list
+    logger(std::string name, sinks_init_list sinks)
+        : logger(std::move(name), sinks.begin(), sinks.end()) {}
+
+    virtual ~logger() = default;
+
+    logger(const logger &other);
+    logger(logger &&other) SPDLOG_NOEXCEPT;
+    logger &operator=(logger other) SPDLOG_NOEXCEPT;
+    void swap(spdlog::logger &other) SPDLOG_NOEXCEPT;
+
+    template <typename... Args>
+    void log(source_loc loc, level::level_enum lvl, format_string_t<Args...> fmt, Args &&...args) {
+        log_(loc, lvl, details::to_string_view(fmt), std::forward<Args>(args)...);
+    }
+
+    template <typename... Args>
+    void log(level::level_enum lvl, format_string_t<Args...> fmt, Args &&...args) {
+        log(source_loc{}, lvl, fmt, std::forward<Args>(args)...);
+    }
+
+    template <typename T>
+    void log(level::level_enum lvl, const T &msg) {
+        log(source_loc{}, lvl, msg);
+    }
+
+    // T cannot be statically converted to format string (including string_view/wstring_view)
+    template <class T,
+              typename std::enable_if<!is_convertible_to_any_format_string<const T &>::value,
+                                      int>::type = 0>
+    void log(source_loc loc, level::level_enum lvl, const T &msg) {
+        log(loc, lvl, "{}", msg);
+    }
+
+    void log(log_clock::time_point log_time,
+             source_loc loc,
+             level::level_enum lvl,
+             string_view_t msg) {
+        bool log_enabled = should_log(lvl);
+        bool traceback_enabled = tracer_.enabled();
+        if (!log_enabled && !traceback_enabled) {
+            return;
+        }
+
+        details::log_msg log_msg(log_time, loc, name_, lvl, msg);
+        log_it_(log_msg, log_enabled, traceback_enabled);
+    }
+
+    void log(source_loc loc, level::level_enum lvl, string_view_t msg) {
+        bool log_enabled = should_log(lvl);
+        bool traceback_enabled = tracer_.enabled();
+        if (!log_enabled && !traceback_enabled) {
+            return;
+        }
+
+        details::log_msg log_msg(loc, name_, lvl, msg);
+        log_it_(log_msg, log_enabled, traceback_enabled);
+    }
+
+    void log(level::level_enum lvl, string_view_t msg) { log(source_loc{}, lvl, msg); }
+
+    template <typename... Args>
+    void trace(format_string_t<Args...> fmt, Args &&...args) {
+        log(level::trace, fmt, std::forward<Args>(args)...);
+    }
+
+    template <typename... Args>
+    void debug(format_string_t<Args...> fmt, Args &&...args) {
+        log(level::debug, fmt, std::forward<Args>(args)...);
+    }
+
+    template <typename... Args>
+    void info(format_string_t<Args...> fmt, Args &&...args) {
+        log(level::info, fmt, std::forward<Args>(args)...);
+    }
+
+    template <typename... Args>
+    void warn(format_string_t<Args...> fmt, Args &&...args) {
+        log(level::warn, fmt, std::forward<Args>(args)...);
+    }
+
+    template <typename... Args>
+    void error(format_string_t<Args...> fmt, Args &&...args) {
+        log(level::err, fmt, std::forward<Args>(args)...);
+    }
+
+    template <typename... Args>
+    void critical(format_string_t<Args...> fmt, Args &&...args) {
+        log(level::critical, fmt, std::forward<Args>(args)...);
+    }
+
+#ifdef SPDLOG_WCHAR_TO_UTF8_SUPPORT
+    template <typename... Args>
+    void log(source_loc loc, level::level_enum lvl, wformat_string_t<Args...> fmt, Args &&...args) {
+        log_(loc, lvl, details::to_string_view(fmt), std::forward<Args>(args)...);
+    }
+
+    template <typename... Args>
+    void log(level::level_enum lvl, wformat_string_t<Args...> fmt, Args &&...args) {
+        log(source_loc{}, lvl, fmt, std::forward<Args>(args)...);
+    }
+
+    void log(log_clock::time_point log_time,
+             source_loc loc,
+             level::level_enum lvl,
+             wstring_view_t msg) {
+        bool log_enabled = should_log(lvl);
+        bool traceback_enabled = tracer_.enabled();
+        if (!log_enabled && !traceback_enabled) {
+            return;
+        }
+
+        memory_buf_t buf;
+        details::os::wstr_to_utf8buf(wstring_view_t(msg.data(), msg.size()), buf);
+        details::log_msg log_msg(log_time, loc, name_, lvl, string_view_t(buf.data(), buf.size()));
+        log_it_(log_msg, log_enabled, traceback_enabled);
+    }
+
+    void log(source_loc loc, level::level_enum lvl, wstring_view_t msg) {
+        bool log_enabled = should_log(lvl);
+        bool traceback_enabled = tracer_.enabled();
+        if (!log_enabled && !traceback_enabled) {
+            return;
+        }
+
+        memory_buf_t buf;
+        details::os::wstr_to_utf8buf(wstring_view_t(msg.data(), msg.size()), buf);
+        details::log_msg log_msg(loc, name_, lvl, string_view_t(buf.data(), buf.size()));
+        log_it_(log_msg, log_enabled, traceback_enabled);
+    }
+
+    void log(level::level_enum lvl, wstring_view_t msg) { log(source_loc{}, lvl, msg); }
+
+    template <typename... Args>
+    void trace(wformat_string_t<Args...> fmt, Args &&...args) {
+        log(level::trace, fmt, std::forward<Args>(args)...);
+    }
+
+    template <typename... Args>
+    void debug(wformat_string_t<Args...> fmt, Args &&...args) {
+        log(level::debug, fmt, std::forward<Args>(args)...);
+    }
+
+    template <typename... Args>
+    void info(wformat_string_t<Args...> fmt, Args &&...args) {
+        log(level::info, fmt, std::forward<Args>(args)...);
+    }
+
+    template <typename... Args>
+    void warn(wformat_string_t<Args...> fmt, Args &&...args) {
+        log(level::warn, fmt, std::forward<Args>(args)...);
+    }
+
+    template <typename... Args>
+    void error(wformat_string_t<Args...> fmt, Args &&...args) {
+        log(level::err, fmt, std::forward<Args>(args)...);
+    }
+
+    template <typename... Args>
+    void critical(wformat_string_t<Args...> fmt, Args &&...args) {
+        log(level::critical, fmt, std::forward<Args>(args)...);
+    }
+#endif
+
+    template <typename T>
+    void trace(const T &msg) {
+        log(level::trace, msg);
+    }
+
+    template <typename T>
+    void debug(const T &msg) {
+        log(level::debug, msg);
+    }
+
+    template <typename T>
+    void info(const T &msg) {
+        log(level::info, msg);
+    }
+
+    template <typename T>
+    void warn(const T &msg) {
+        log(level::warn, msg);
+    }
+
+    template <typename T>
+    void error(const T &msg) {
+        log(level::err, msg);
+    }
+
+    template <typename T>
+    void critical(const T &msg) {
+        log(level::critical, msg);
+    }
+
+    // return true logging is enabled for the given level.
+    bool should_log(level::level_enum msg_level) const {
+        return msg_level >= level_.load(std::memory_order_relaxed);
+    }
+
+    // return true if backtrace logging is enabled.
+    bool should_backtrace() const { return tracer_.enabled(); }
+
+    void set_level(level::level_enum log_level);
+
+    level::level_enum level() const;
+
+    const std::string &name() const;
+
+    // set formatting for the sinks in this logger.
+    // each sink will get a separate instance of the formatter object.
+    void set_formatter(std::unique_ptr<formatter> f);
+
+    // set formatting for the sinks in this logger.
+    // equivalent to
+    //     set_formatter(make_unique<pattern_formatter>(pattern, time_type))
+    // Note: each sink will get a new instance of a formatter object, replacing the old one.
+    void set_pattern(std::string pattern, pattern_time_type time_type = pattern_time_type::local);
+
+    // backtrace support.
+    // efficiently store all debug/trace messages in a circular buffer until needed for debugging.
+    void enable_backtrace(size_t n_messages);
+    void disable_backtrace();
+    void dump_backtrace();
+
+    // flush functions
+    void flush();
+    void flush_on(level::level_enum log_level);
+    level::level_enum flush_level() const;
+
+    // sinks
+    const std::vector<sink_ptr> &sinks() const;
+
+    std::vector<sink_ptr> &sinks();
+
+    // error handler
+    void set_error_handler(err_handler);
+
+    // create new logger with same sinks and configuration.
+    virtual std::shared_ptr<logger> clone(std::string logger_name);
+
+protected:
+    std::string name_;
+    std::vector<sink_ptr> sinks_;
+    spdlog::level_t level_{level::info};
+    spdlog::level_t flush_level_{level::off};
+    err_handler custom_err_handler_{nullptr};
+    details::backtracer tracer_;
+
+    // common implementation for after templated public api has been resolved
+    template <typename... Args>
+    void log_(source_loc loc, level::level_enum lvl, string_view_t fmt, Args &&...args) {
+        bool log_enabled = should_log(lvl);
+        bool traceback_enabled = tracer_.enabled();
+        if (!log_enabled && !traceback_enabled) {
+            return;
+        }
+        SPDLOG_TRY {
+            memory_buf_t buf;
+#ifdef SPDLOG_USE_STD_FORMAT
+            fmt_lib::vformat_to(std::back_inserter(buf), fmt, fmt_lib::make_format_args(args...));
+#else
+            fmt::vformat_to(fmt::appender(buf), fmt, fmt::make_format_args(args...));
+#endif
+
+            details::log_msg log_msg(loc, name_, lvl, string_view_t(buf.data(), buf.size()));
+            log_it_(log_msg, log_enabled, traceback_enabled);
+        }
+        SPDLOG_LOGGER_CATCH(loc)
+    }
+
+#ifdef SPDLOG_WCHAR_TO_UTF8_SUPPORT
+    template <typename... Args>
+    void log_(source_loc loc, level::level_enum lvl, wstring_view_t fmt, Args &&...args) {
+        bool log_enabled = should_log(lvl);
+        bool traceback_enabled = tracer_.enabled();
+        if (!log_enabled && !traceback_enabled) {
+            return;
+        }
+        SPDLOG_TRY {
+            // format to wmemory_buffer and convert to utf8
+            wmemory_buf_t wbuf;
+            fmt_lib::vformat_to(std::back_inserter(wbuf), fmt,
+                                fmt_lib::make_format_args<fmt_lib::wformat_context>(args...));
+
+            memory_buf_t buf;
+            details::os::wstr_to_utf8buf(wstring_view_t(wbuf.data(), wbuf.size()), buf);
+            details::log_msg log_msg(loc, name_, lvl, string_view_t(buf.data(), buf.size()));
+            log_it_(log_msg, log_enabled, traceback_enabled);
+        }
+        SPDLOG_LOGGER_CATCH(loc)
+    }
+#endif  // SPDLOG_WCHAR_TO_UTF8_SUPPORT
+
+    // log the given message (if the given log level is high enough),
+    // and save backtrace (if backtrace is enabled).
+    void log_it_(const details::log_msg &log_msg, bool log_enabled, bool traceback_enabled);
+    virtual void sink_it_(const details::log_msg &msg);
+    virtual void flush_();
+    void dump_backtrace_();
+    bool should_flush_(const details::log_msg &msg);
+
+    // handle errors during logging.
+    // default handler prints the error to stderr at max rate of 1 message/sec.
+    void err_handler_(const std::string &msg);
+};
+
+void swap(logger &a, logger &b);
+
+}  // namespace spdlog
+
+#ifdef SPDLOG_HEADER_ONLY
+    #include "logger-inl.h"
+#endif
diff --git a/ext/spdlog/mdc.h b/ext/spdlog/mdc.h
new file mode 100644
index 0000000..80b6f25
--- /dev/null
+++ b/ext/spdlog/mdc.h
@@ -0,0 +1,50 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#if defined(SPDLOG_NO_TLS)
+    #error "This header requires thread local storage support, but SPDLOG_NO_TLS is defined."
+#endif
+
+#include <map>
+#include <string>
+
+#include <spdlog/common.h>
+
+// MDC is a simple map of key->string values stored in thread local storage whose content will be printed by the loggers.
+// Note: Not supported in async mode (thread local storage - so the async thread pool have different copy).
+//
+// Usage example:
+// spdlog::mdc::put("mdc_key_1", "mdc_value_1");
+// spdlog::info("Hello, {}", "World!");  // => [2024-04-26 02:08:05.040] [info] [mdc_key_1:mdc_value_1] Hello, World!
+
+namespace spdlog {
+class SPDLOG_API mdc {
+public:
+    using mdc_map_t = std::map<std::string, std::string>;
+
+    static void put(const std::string &key, const std::string &value) {
+        get_context()[key] = value;
+    }
+
+    static std::string get(const std::string &key) {
+        auto &context = get_context();
+        auto it = context.find(key);
+        if (it != context.end()) {
+            return it->second;
+        }
+        return "";
+    }
+
+    static void remove(const std::string &key) { get_context().erase(key); }
+
+    static void clear() { get_context().clear(); }
+
+    static mdc_map_t &get_context() {
+        static thread_local mdc_map_t context;
+        return context;
+    }
+};
+
+}  // namespace spdlog
diff --git a/ext/spdlog/pattern_formatter-inl.h b/ext/spdlog/pattern_formatter-inl.h
new file mode 100644
index 0000000..b53d805
--- /dev/null
+++ b/ext/spdlog/pattern_formatter-inl.h
@@ -0,0 +1,1338 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#ifndef SPDLOG_HEADER_ONLY
+    #include <spdlog/pattern_formatter.h>
+#endif
+
+#include <spdlog/details/fmt_helper.h>
+#include <spdlog/details/log_msg.h>
+#include <spdlog/details/os.h>
+
+#ifndef SPDLOG_NO_TLS
+    #include <spdlog/mdc.h>
+#endif
+
+#include <spdlog/fmt/fmt.h>
+#include <spdlog/formatter.h>
+
+#include <algorithm>
+#include <array>
+#include <cctype>
+#include <chrono>
+#include <cstring>
+#include <ctime>
+#include <iterator>
+#include <memory>
+#include <mutex>
+#include <string>
+#include <thread>
+#include <utility>
+#include <vector>
+
+namespace spdlog {
+namespace details {
+
+///////////////////////////////////////////////////////////////////////
+// name & level pattern appender
+///////////////////////////////////////////////////////////////////////
+
+class scoped_padder {
+public:
+    scoped_padder(size_t wrapped_size, const padding_info &padinfo, memory_buf_t &dest)
+        : padinfo_(padinfo),
+          dest_(dest) {
+        remaining_pad_ = static_cast<long>(padinfo.width_) - static_cast<long>(wrapped_size);
+        if (remaining_pad_ <= 0) {
+            return;
+        }
+
+        if (padinfo_.side_ == padding_info::pad_side::left) {
+            pad_it(remaining_pad_);
+            remaining_pad_ = 0;
+        } else if (padinfo_.side_ == padding_info::pad_side::center) {
+            auto half_pad = remaining_pad_ / 2;
+            auto reminder = remaining_pad_ & 1;
+            pad_it(half_pad);
+            remaining_pad_ = half_pad + reminder;  // for the right side
+        }
+    }
+
+    template <typename T>
+    static unsigned int count_digits(T n) {
+        return fmt_helper::count_digits(n);
+    }
+
+    ~scoped_padder() {
+        if (remaining_pad_ >= 0) {
+            pad_it(remaining_pad_);
+        } else if (padinfo_.truncate_) {
+            long new_size = static_cast<long>(dest_.size()) + remaining_pad_;
+            dest_.resize(static_cast<size_t>(new_size));
+        }
+    }
+
+private:
+    void pad_it(long count) {
+        fmt_helper::append_string_view(string_view_t(spaces_.data(), static_cast<size_t>(count)),
+                                       dest_);
+    }
+
+    const padding_info &padinfo_;
+    memory_buf_t &dest_;
+    long remaining_pad_;
+    string_view_t spaces_{"                                                                ", 64};
+};
+
+struct null_scoped_padder {
+    null_scoped_padder(size_t /*wrapped_size*/,
+                       const padding_info & /*padinfo*/,
+                       memory_buf_t & /*dest*/) {}
+
+    template <typename T>
+    static unsigned int count_digits(T /* number */) {
+        return 0;
+    }
+};
+
+template <typename ScopedPadder>
+class name_formatter final : public flag_formatter {
+public:
+    explicit name_formatter(padding_info padinfo)
+        : flag_formatter(padinfo) {}
+
+    void format(const details::log_msg &msg, const std::tm &, memory_buf_t &dest) override {
+        ScopedPadder p(msg.logger_name.size(), padinfo_, dest);
+        fmt_helper::append_string_view(msg.logger_name, dest);
+    }
+};
+
+// log level appender
+template <typename ScopedPadder>
+class level_formatter final : public flag_formatter {
+public:
+    explicit level_formatter(padding_info padinfo)
+        : flag_formatter(padinfo) {}
+
+    void format(const details::log_msg &msg, const std::tm &, memory_buf_t &dest) override {
+        const string_view_t &level_name = level::to_string_view(msg.level);
+        ScopedPadder p(level_name.size(), padinfo_, dest);
+        fmt_helper::append_string_view(level_name, dest);
+    }
+};
+
+// short log level appender
+template <typename ScopedPadder>
+class short_level_formatter final : public flag_formatter {
+public:
+    explicit short_level_formatter(padding_info padinfo)
+        : flag_formatter(padinfo) {}
+
+    void format(const details::log_msg &msg, const std::tm &, memory_buf_t &dest) override {
+        string_view_t level_name{level::to_short_c_str(msg.level)};
+        ScopedPadder p(level_name.size(), padinfo_, dest);
+        fmt_helper::append_string_view(level_name, dest);
+    }
+};
+
+///////////////////////////////////////////////////////////////////////
+// Date time pattern appenders
+///////////////////////////////////////////////////////////////////////
+
+static const char *ampm(const tm &t) { return t.tm_hour >= 12 ? "PM" : "AM"; }
+
+static int to12h(const tm &t) { return t.tm_hour > 12 ? t.tm_hour - 12 : t.tm_hour; }
+
+// Abbreviated weekday name
+static std::array<const char *, 7> days{{"Sun", "Mon", "Tue", "Wed", "Thu", "Fri", "Sat"}};
+
+template <typename ScopedPadder>
+class a_formatter final : public flag_formatter {
+public:
+    explicit a_formatter(padding_info padinfo)
+        : flag_formatter(padinfo) {}
+
+    void format(const details::log_msg &, const std::tm &tm_time, memory_buf_t &dest) override {
+        string_view_t field_value{days[static_cast<size_t>(tm_time.tm_wday)]};
+        ScopedPadder p(field_value.size(), padinfo_, dest);
+        fmt_helper::append_string_view(field_value, dest);
+    }
+};
+
+// Full weekday name
+static std::array<const char *, 7> full_days{
+    {"Sunday", "Monday", "Tuesday", "Wednesday", "Thursday", "Friday", "Saturday"}};
+
+template <typename ScopedPadder>
+class A_formatter : public flag_formatter {
+public:
+    explicit A_formatter(padding_info padinfo)
+        : flag_formatter(padinfo) {}
+
+    void format(const details::log_msg &, const std::tm &tm_time, memory_buf_t &dest) override {
+        string_view_t field_value{full_days[static_cast<size_t>(tm_time.tm_wday)]};
+        ScopedPadder p(field_value.size(), padinfo_, dest);
+        fmt_helper::append_string_view(field_value, dest);
+    }
+};
+
+// Abbreviated month
+static const std::array<const char *, 12> months{
+    {"Jan", "Feb", "Mar", "Apr", "May", "Jun", "Jul", "Aug", "Sep", "Oct", "Nov", "Dec"}};
+
+template <typename ScopedPadder>
+class b_formatter final : public flag_formatter {
+public:
+    explicit b_formatter(padding_info padinfo)
+        : flag_formatter(padinfo) {}
+
+    void format(const details::log_msg &, const std::tm &tm_time, memory_buf_t &dest) override {
+        string_view_t field_value{months[static_cast<size_t>(tm_time.tm_mon)]};
+        ScopedPadder p(field_value.size(), padinfo_, dest);
+        fmt_helper::append_string_view(field_value, dest);
+    }
+};
+
+// Full month name
+static const std::array<const char *, 12> full_months{{"January", "February", "March", "April",
+                                                       "May", "June", "July", "August", "September",
+                                                       "October", "November", "December"}};
+
+template <typename ScopedPadder>
+class B_formatter final : public flag_formatter {
+public:
+    explicit B_formatter(padding_info padinfo)
+        : flag_formatter(padinfo) {}
+
+    void format(const details::log_msg &, const std::tm &tm_time, memory_buf_t &dest) override {
+        string_view_t field_value{full_months[static_cast<size_t>(tm_time.tm_mon)]};
+        ScopedPadder p(field_value.size(), padinfo_, dest);
+        fmt_helper::append_string_view(field_value, dest);
+    }
+};
+
+// Date and time representation (Thu Aug 23 15:35:46 2014)
+template <typename ScopedPadder>
+class c_formatter final : public flag_formatter {
+public:
+    explicit c_formatter(padding_info padinfo)
+        : flag_formatter(padinfo) {}
+
+    void format(const details::log_msg &, const std::tm &tm_time, memory_buf_t &dest) override {
+        const size_t field_size = 24;
+        ScopedPadder p(field_size, padinfo_, dest);
+
+        fmt_helper::append_string_view(days[static_cast<size_t>(tm_time.tm_wday)], dest);
+        dest.push_back(' ');
+        fmt_helper::append_string_view(months[static_cast<size_t>(tm_time.tm_mon)], dest);
+        dest.push_back(' ');
+        fmt_helper::append_int(tm_time.tm_mday, dest);
+        dest.push_back(' ');
+        // time
+
+        fmt_helper::pad2(tm_time.tm_hour, dest);
+        dest.push_back(':');
+        fmt_helper::pad2(tm_time.tm_min, dest);
+        dest.push_back(':');
+        fmt_helper::pad2(tm_time.tm_sec, dest);
+        dest.push_back(' ');
+        fmt_helper::append_int(tm_time.tm_year + 1900, dest);
+    }
+};
+
+// year - 2 digit
+template <typename ScopedPadder>
+class C_formatter final : public flag_formatter {
+public:
+    explicit C_formatter(padding_info padinfo)
+        : flag_formatter(padinfo) {}
+
+    void format(const details::log_msg &, const std::tm &tm_time, memory_buf_t &dest) override {
+        const size_t field_size = 2;
+        ScopedPadder p(field_size, padinfo_, dest);
+        fmt_helper::pad2(tm_time.tm_year % 100, dest);
+    }
+};
+
+// Short MM/DD/YY date, equivalent to %m/%d/%y 08/23/01
+template <typename ScopedPadder>
+class D_formatter final : public flag_formatter {
+public:
+    explicit D_formatter(padding_info padinfo)
+        : flag_formatter(padinfo) {}
+
+    void format(const details::log_msg &, const std::tm &tm_time, memory_buf_t &dest) override {
+        const size_t field_size = 10;
+        ScopedPadder p(field_size, padinfo_, dest);
+
+        fmt_helper::pad2(tm_time.tm_mon + 1, dest);
+        dest.push_back('/');
+        fmt_helper::pad2(tm_time.tm_mday, dest);
+        dest.push_back('/');
+        fmt_helper::pad2(tm_time.tm_year % 100, dest);
+    }
+};
+
+// year - 4 digit
+template <typename ScopedPadder>
+class Y_formatter final : public flag_formatter {
+public:
+    explicit Y_formatter(padding_info padinfo)
+        : flag_formatter(padinfo) {}
+
+    void format(const details::log_msg &, const std::tm &tm_time, memory_buf_t &dest) override {
+        const size_t field_size = 4;
+        ScopedPadder p(field_size, padinfo_, dest);
+        fmt_helper::append_int(tm_time.tm_year + 1900, dest);
+    }
+};
+
+// month 1-12
+template <typename ScopedPadder>
+class m_formatter final : public flag_formatter {
+public:
+    explicit m_formatter(padding_info padinfo)
+        : flag_formatter(padinfo) {}
+
+    void format(const details::log_msg &, const std::tm &tm_time, memory_buf_t &dest) override {
+        const size_t field_size = 2;
+        ScopedPadder p(field_size, padinfo_, dest);
+        fmt_helper::pad2(tm_time.tm_mon + 1, dest);
+    }
+};
+
+// day of month 1-31
+template <typename ScopedPadder>
+class d_formatter final : public flag_formatter {
+public:
+    explicit d_formatter(padding_info padinfo)
+        : flag_formatter(padinfo) {}
+
+    void format(const details::log_msg &, const std::tm &tm_time, memory_buf_t &dest) override {
+        const size_t field_size = 2;
+        ScopedPadder p(field_size, padinfo_, dest);
+        fmt_helper::pad2(tm_time.tm_mday, dest);
+    }
+};
+
+// hours in 24 format 0-23
+template <typename ScopedPadder>
+class H_formatter final : public flag_formatter {
+public:
+    explicit H_formatter(padding_info padinfo)
+        : flag_formatter(padinfo) {}
+
+    void format(const details::log_msg &, const std::tm &tm_time, memory_buf_t &dest) override {
+        const size_t field_size = 2;
+        ScopedPadder p(field_size, padinfo_, dest);
+        fmt_helper::pad2(tm_time.tm_hour, dest);
+    }
+};
+
+// hours in 12 format 1-12
+template <typename ScopedPadder>
+class I_formatter final : public flag_formatter {
+public:
+    explicit I_formatter(padding_info padinfo)
+        : flag_formatter(padinfo) {}
+
+    void format(const details::log_msg &, const std::tm &tm_time, memory_buf_t &dest) override {
+        const size_t field_size = 2;
+        ScopedPadder p(field_size, padinfo_, dest);
+        fmt_helper::pad2(to12h(tm_time), dest);
+    }
+};
+
+// minutes 0-59
+template <typename ScopedPadder>
+class M_formatter final : public flag_formatter {
+public:
+    explicit M_formatter(padding_info padinfo)
+        : flag_formatter(padinfo) {}
+
+    void format(const details::log_msg &, const std::tm &tm_time, memory_buf_t &dest) override {
+        const size_t field_size = 2;
+        ScopedPadder p(field_size, padinfo_, dest);
+        fmt_helper::pad2(tm_time.tm_min, dest);
+    }
+};
+
+// seconds 0-59
+template <typename ScopedPadder>
+class S_formatter final : public flag_formatter {
+public:
+    explicit S_formatter(padding_info padinfo)
+        : flag_formatter(padinfo) {}
+
+    void format(const details::log_msg &, const std::tm &tm_time, memory_buf_t &dest) override {
+        const size_t field_size = 2;
+        ScopedPadder p(field_size, padinfo_, dest);
+        fmt_helper::pad2(tm_time.tm_sec, dest);
+    }
+};
+
+// milliseconds
+template <typename ScopedPadder>
+class e_formatter final : public flag_formatter {
+public:
+    explicit e_formatter(padding_info padinfo)
+        : flag_formatter(padinfo) {}
+
+    void format(const details::log_msg &msg, const std::tm &, memory_buf_t &dest) override {
+        auto millis = fmt_helper::time_fraction<std::chrono::milliseconds>(msg.time);
+        const size_t field_size = 3;
+        ScopedPadder p(field_size, padinfo_, dest);
+        fmt_helper::pad3(static_cast<uint32_t>(millis.count()), dest);
+    }
+};
+
+// microseconds
+template <typename ScopedPadder>
+class f_formatter final : public flag_formatter {
+public:
+    explicit f_formatter(padding_info padinfo)
+        : flag_formatter(padinfo) {}
+
+    void format(const details::log_msg &msg, const std::tm &, memory_buf_t &dest) override {
+        auto micros = fmt_helper::time_fraction<std::chrono::microseconds>(msg.time);
+
+        const size_t field_size = 6;
+        ScopedPadder p(field_size, padinfo_, dest);
+        fmt_helper::pad6(static_cast<size_t>(micros.count()), dest);
+    }
+};
+
+// nanoseconds
+template <typename ScopedPadder>
+class F_formatter final : public flag_formatter {
+public:
+    explicit F_formatter(padding_info padinfo)
+        : flag_formatter(padinfo) {}
+
+    void format(const details::log_msg &msg, const std::tm &, memory_buf_t &dest) override {
+        auto ns = fmt_helper::time_fraction<std::chrono::nanoseconds>(msg.time);
+        const size_t field_size = 9;
+        ScopedPadder p(field_size, padinfo_, dest);
+        fmt_helper::pad9(static_cast<size_t>(ns.count()), dest);
+    }
+};
+
+// seconds since epoch
+template <typename ScopedPadder>
+class E_formatter final : public flag_formatter {
+public:
+    explicit E_formatter(padding_info padinfo)
+        : flag_formatter(padinfo) {}
+
+    void format(const details::log_msg &msg, const std::tm &, memory_buf_t &dest) override {
+        const size_t field_size = 10;
+        ScopedPadder p(field_size, padinfo_, dest);
+        auto duration = msg.time.time_since_epoch();
+        auto seconds = std::chrono::duration_cast<std::chrono::seconds>(duration).count();
+        fmt_helper::append_int(seconds, dest);
+    }
+};
+
+// AM/PM
+template <typename ScopedPadder>
+class p_formatter final : public flag_formatter {
+public:
+    explicit p_formatter(padding_info padinfo)
+        : flag_formatter(padinfo) {}
+
+    void format(const details::log_msg &, const std::tm &tm_time, memory_buf_t &dest) override {
+        const size_t field_size = 2;
+        ScopedPadder p(field_size, padinfo_, dest);
+        fmt_helper::append_string_view(ampm(tm_time), dest);
+    }
+};
+
+// 12 hour clock 02:55:02 pm
+template <typename ScopedPadder>
+class r_formatter final : public flag_formatter {
+public:
+    explicit r_formatter(padding_info padinfo)
+        : flag_formatter(padinfo) {}
+
+    void format(const details::log_msg &, const std::tm &tm_time, memory_buf_t &dest) override {
+        const size_t field_size = 11;
+        ScopedPadder p(field_size, padinfo_, dest);
+
+        fmt_helper::pad2(to12h(tm_time), dest);
+        dest.push_back(':');
+        fmt_helper::pad2(tm_time.tm_min, dest);
+        dest.push_back(':');
+        fmt_helper::pad2(tm_time.tm_sec, dest);
+        dest.push_back(' ');
+        fmt_helper::append_string_view(ampm(tm_time), dest);
+    }
+};
+
+// 24-hour HH:MM time, equivalent to %H:%M
+template <typename ScopedPadder>
+class R_formatter final : public flag_formatter {
+public:
+    explicit R_formatter(padding_info padinfo)
+        : flag_formatter(padinfo) {}
+
+    void format(const details::log_msg &, const std::tm &tm_time, memory_buf_t &dest) override {
+        const size_t field_size = 5;
+        ScopedPadder p(field_size, padinfo_, dest);
+
+        fmt_helper::pad2(tm_time.tm_hour, dest);
+        dest.push_back(':');
+        fmt_helper::pad2(tm_time.tm_min, dest);
+    }
+};
+
+// ISO 8601 time format (HH:MM:SS), equivalent to %H:%M:%S
+template <typename ScopedPadder>
+class T_formatter final : public flag_formatter {
+public:
+    explicit T_formatter(padding_info padinfo)
+        : flag_formatter(padinfo) {}
+
+    void format(const details::log_msg &, const std::tm &tm_time, memory_buf_t &dest) override {
+        const size_t field_size = 8;
+        ScopedPadder p(field_size, padinfo_, dest);
+
+        fmt_helper::pad2(tm_time.tm_hour, dest);
+        dest.push_back(':');
+        fmt_helper::pad2(tm_time.tm_min, dest);
+        dest.push_back(':');
+        fmt_helper::pad2(tm_time.tm_sec, dest);
+    }
+};
+
+// ISO 8601 offset from UTC in timezone (+-HH:MM)
+template <typename ScopedPadder>
+class z_formatter final : public flag_formatter {
+public:
+    explicit z_formatter(padding_info padinfo)
+        : flag_formatter(padinfo) {}
+
+    z_formatter() = default;
+    z_formatter(const z_formatter &) = delete;
+    z_formatter &operator=(const z_formatter &) = delete;
+
+    void format(const details::log_msg &msg, const std::tm &tm_time, memory_buf_t &dest) override {
+        const size_t field_size = 6;
+        ScopedPadder p(field_size, padinfo_, dest);
+
+        auto total_minutes = get_cached_offset(msg, tm_time);
+        bool is_negative = total_minutes < 0;
+        if (is_negative) {
+            total_minutes = -total_minutes;
+            dest.push_back('-');
+        } else {
+            dest.push_back('+');
+        }
+
+        fmt_helper::pad2(total_minutes / 60, dest);  // hours
+        dest.push_back(':');
+        fmt_helper::pad2(total_minutes % 60, dest);  // minutes
+    }
+
+private:
+    log_clock::time_point last_update_{std::chrono::seconds(0)};
+    int offset_minutes_{0};
+
+    int get_cached_offset(const log_msg &msg, const std::tm &tm_time) {
+        // refresh every 10 seconds
+        if (msg.time - last_update_ >= std::chrono::seconds(10)) {
+            offset_minutes_ = os::utc_minutes_offset(tm_time);
+            last_update_ = msg.time;
+        }
+        return offset_minutes_;
+    }
+};
+
+// Thread id
+template <typename ScopedPadder>
+class t_formatter final : public flag_formatter {
+public:
+    explicit t_formatter(padding_info padinfo)
+        : flag_formatter(padinfo) {}
+
+    void format(const details::log_msg &msg, const std::tm &, memory_buf_t &dest) override {
+        const auto field_size = ScopedPadder::count_digits(msg.thread_id);
+        ScopedPadder p(field_size, padinfo_, dest);
+        fmt_helper::append_int(msg.thread_id, dest);
+    }
+};
+
+// Current pid
+template <typename ScopedPadder>
+class pid_formatter final : public flag_formatter {
+public:
+    explicit pid_formatter(padding_info padinfo)
+        : flag_formatter(padinfo) {}
+
+    void format(const details::log_msg &, const std::tm &, memory_buf_t &dest) override {
+        const auto pid = static_cast<uint32_t>(details::os::pid());
+        auto field_size = ScopedPadder::count_digits(pid);
+        ScopedPadder p(field_size, padinfo_, dest);
+        fmt_helper::append_int(pid, dest);
+    }
+};
+
+template <typename ScopedPadder>
+class v_formatter final : public flag_formatter {
+public:
+    explicit v_formatter(padding_info padinfo)
+        : flag_formatter(padinfo) {}
+
+    void format(const details::log_msg &msg, const std::tm &, memory_buf_t &dest) override {
+        ScopedPadder p(msg.payload.size(), padinfo_, dest);
+        fmt_helper::append_string_view(msg.payload, dest);
+    }
+};
+
+class ch_formatter final : public flag_formatter {
+public:
+    explicit ch_formatter(char ch)
+        : ch_(ch) {}
+
+    void format(const details::log_msg &, const std::tm &, memory_buf_t &dest) override {
+        dest.push_back(ch_);
+    }
+
+private:
+    char ch_;
+};
+
+// aggregate user chars to display as is
+class aggregate_formatter final : public flag_formatter {
+public:
+    aggregate_formatter() = default;
+
+    void add_ch(char ch) { str_ += ch; }
+    void format(const details::log_msg &, const std::tm &, memory_buf_t &dest) override {
+        fmt_helper::append_string_view(str_, dest);
+    }
+
+private:
+    std::string str_;
+};
+
+// mark the color range. expect it to be in the form of "%^colored text%$"
+class color_start_formatter final : public flag_formatter {
+public:
+    explicit color_start_formatter(padding_info padinfo)
+        : flag_formatter(padinfo) {}
+
+    void format(const details::log_msg &msg, const std::tm &, memory_buf_t &dest) override {
+        msg.color_range_start = dest.size();
+    }
+};
+
+class color_stop_formatter final : public flag_formatter {
+public:
+    explicit color_stop_formatter(padding_info padinfo)
+        : flag_formatter(padinfo) {}
+
+    void format(const details::log_msg &msg, const std::tm &, memory_buf_t &dest) override {
+        msg.color_range_end = dest.size();
+    }
+};
+
+// print source location
+template <typename ScopedPadder>
+class source_location_formatter final : public flag_formatter {
+public:
+    explicit source_location_formatter(padding_info padinfo)
+        : flag_formatter(padinfo) {}
+
+    void format(const details::log_msg &msg, const std::tm &, memory_buf_t &dest) override {
+        if (msg.source.empty()) {
+            ScopedPadder p(0, padinfo_, dest);
+            return;
+        }
+
+        size_t text_size;
+        if (padinfo_.enabled()) {
+            // calc text size for padding based on "filename:line"
+            text_size = std::char_traits<char>::length(msg.source.filename) +
+                        ScopedPadder::count_digits(msg.source.line) + 1;
+        } else {
+            text_size = 0;
+        }
+
+        ScopedPadder p(text_size, padinfo_, dest);
+        fmt_helper::append_string_view(msg.source.filename, dest);
+        dest.push_back(':');
+        fmt_helper::append_int(msg.source.line, dest);
+    }
+};
+
+// print source filename
+template <typename ScopedPadder>
+class source_filename_formatter final : public flag_formatter {
+public:
+    explicit source_filename_formatter(padding_info padinfo)
+        : flag_formatter(padinfo) {}
+
+    void format(const details::log_msg &msg, const std::tm &, memory_buf_t &dest) override {
+        if (msg.source.empty()) {
+            ScopedPadder p(0, padinfo_, dest);
+            return;
+        }
+        size_t text_size =
+            padinfo_.enabled() ? std::char_traits<char>::length(msg.source.filename) : 0;
+        ScopedPadder p(text_size, padinfo_, dest);
+        fmt_helper::append_string_view(msg.source.filename, dest);
+    }
+};
+
+template <typename ScopedPadder>
+class short_filename_formatter final : public flag_formatter {
+public:
+    explicit short_filename_formatter(padding_info padinfo)
+        : flag_formatter(padinfo) {}
+
+#ifdef _MSC_VER
+    #pragma warning(push)
+    #pragma warning(disable : 4127)  // consider using 'if constexpr' instead
+#endif                               // _MSC_VER
+    static const char *basename(const char *filename) {
+        // if the size is 2 (1 character + null terminator) we can use the more efficient strrchr
+        // the branch will be elided by optimizations
+        if (sizeof(os::folder_seps) == 2) {
+            const char *rv = std::strrchr(filename, os::folder_seps[0]);
+            return rv != nullptr ? rv + 1 : filename;
+        } else {
+            const std::reverse_iterator<const char *> begin(filename + std::strlen(filename));
+            const std::reverse_iterator<const char *> end(filename);
+
+            const auto it = std::find_first_of(begin, end, std::begin(os::folder_seps),
+                                               std::end(os::folder_seps) - 1);
+            return it != end ? it.base() : filename;
+        }
+    }
+#ifdef _MSC_VER
+    #pragma warning(pop)
+#endif  // _MSC_VER
+
+    void format(const details::log_msg &msg, const std::tm &, memory_buf_t &dest) override {
+        if (msg.source.empty()) {
+            ScopedPadder p(0, padinfo_, dest);
+            return;
+        }
+        auto filename = basename(msg.source.filename);
+        size_t text_size = padinfo_.enabled() ? std::char_traits<char>::length(filename) : 0;
+        ScopedPadder p(text_size, padinfo_, dest);
+        fmt_helper::append_string_view(filename, dest);
+    }
+};
+
+template <typename ScopedPadder>
+class source_linenum_formatter final : public flag_formatter {
+public:
+    explicit source_linenum_formatter(padding_info padinfo)
+        : flag_formatter(padinfo) {}
+
+    void format(const details::log_msg &msg, const std::tm &, memory_buf_t &dest) override {
+        if (msg.source.empty()) {
+            ScopedPadder p(0, padinfo_, dest);
+            return;
+        }
+
+        auto field_size = ScopedPadder::count_digits(msg.source.line);
+        ScopedPadder p(field_size, padinfo_, dest);
+        fmt_helper::append_int(msg.source.line, dest);
+    }
+};
+
+// print source funcname
+template <typename ScopedPadder>
+class source_funcname_formatter final : public flag_formatter {
+public:
+    explicit source_funcname_formatter(padding_info padinfo)
+        : flag_formatter(padinfo) {}
+
+    void format(const details::log_msg &msg, const std::tm &, memory_buf_t &dest) override {
+        if (msg.source.empty()) {
+            ScopedPadder p(0, padinfo_, dest);
+            return;
+        }
+        size_t text_size =
+            padinfo_.enabled() ? std::char_traits<char>::length(msg.source.funcname) : 0;
+        ScopedPadder p(text_size, padinfo_, dest);
+        fmt_helper::append_string_view(msg.source.funcname, dest);
+    }
+};
+
+// print elapsed time since last message
+template <typename ScopedPadder, typename Units>
+class elapsed_formatter final : public flag_formatter {
+public:
+    using DurationUnits = Units;
+
+    explicit elapsed_formatter(padding_info padinfo)
+        : flag_formatter(padinfo),
+          last_message_time_(log_clock::now()) {}
+
+    void format(const details::log_msg &msg, const std::tm &, memory_buf_t &dest) override {
+        auto delta = (std::max)(msg.time - last_message_time_, log_clock::duration::zero());
+        auto delta_units = std::chrono::duration_cast<DurationUnits>(delta);
+        last_message_time_ = msg.time;
+        auto delta_count = static_cast<size_t>(delta_units.count());
+        auto n_digits = static_cast<size_t>(ScopedPadder::count_digits(delta_count));
+        ScopedPadder p(n_digits, padinfo_, dest);
+        fmt_helper::append_int(delta_count, dest);
+    }
+
+private:
+    log_clock::time_point last_message_time_;
+};
+
+// Class for formatting Mapped Diagnostic Context (MDC) in log messages.
+// Example: [logger-name] [info] [mdc_key_1:mdc_value_1 mdc_key_2:mdc_value_2] some message
+#ifndef SPDLOG_NO_TLS
+template <typename ScopedPadder>
+class mdc_formatter : public flag_formatter {
+public:
+    explicit mdc_formatter(padding_info padinfo)
+        : flag_formatter(padinfo) {}
+
+    void format(const details::log_msg &, const std::tm &, memory_buf_t &dest) override {
+        auto &mdc_map = mdc::get_context();
+        if (mdc_map.empty()) {
+            ScopedPadder p(0, padinfo_, dest);
+            return;
+        } else {
+            format_mdc(mdc_map, dest);
+        }
+    }
+
+    void format_mdc(const mdc::mdc_map_t &mdc_map, memory_buf_t &dest) {
+        auto last_element = --mdc_map.end();
+        for (auto it = mdc_map.begin(); it != mdc_map.end(); ++it) {
+            auto &pair = *it;
+            const auto &key = pair.first;
+            const auto &value = pair.second;
+            size_t content_size = key.size() + value.size() + 1;  // 1 for ':'
+
+            if (it != last_element) {
+                content_size++;  // 1 for ' '
+            }
+
+            ScopedPadder p(content_size, padinfo_, dest);
+            fmt_helper::append_string_view(key, dest);
+            fmt_helper::append_string_view(":", dest);
+            fmt_helper::append_string_view(value, dest);
+            if (it != last_element) {
+                fmt_helper::append_string_view(" ", dest);
+            }
+        }
+    }
+};
+#endif
+
+// Full info formatter
+// pattern: [%Y-%m-%d %H:%M:%S.%e] [%n] [%l] [%s:%#] %v
+class full_formatter final : public flag_formatter {
+public:
+    explicit full_formatter(padding_info padinfo)
+        : flag_formatter(padinfo) {}
+
+    void format(const details::log_msg &msg, const std::tm &tm_time, memory_buf_t &dest) override {
+        using std::chrono::duration_cast;
+        using std::chrono::milliseconds;
+        using std::chrono::seconds;
+
+        // cache the date/time part for the next second.
+        auto duration = msg.time.time_since_epoch();
+        auto secs = duration_cast<seconds>(duration);
+
+        if (cache_timestamp_ != secs || cached_datetime_.size() == 0) {
+            cached_datetime_.clear();
+            cached_datetime_.push_back('[');
+            fmt_helper::append_int(tm_time.tm_year + 1900, cached_datetime_);
+            cached_datetime_.push_back('-');
+
+            fmt_helper::pad2(tm_time.tm_mon + 1, cached_datetime_);
+            cached_datetime_.push_back('-');
+
+            fmt_helper::pad2(tm_time.tm_mday, cached_datetime_);
+            cached_datetime_.push_back(' ');
+
+            fmt_helper::pad2(tm_time.tm_hour, cached_datetime_);
+            cached_datetime_.push_back(':');
+
+            fmt_helper::pad2(tm_time.tm_min, cached_datetime_);
+            cached_datetime_.push_back(':');
+
+            fmt_helper::pad2(tm_time.tm_sec, cached_datetime_);
+            cached_datetime_.push_back('.');
+
+            cache_timestamp_ = secs;
+        }
+        dest.append(cached_datetime_.begin(), cached_datetime_.end());
+
+        auto millis = fmt_helper::time_fraction<milliseconds>(msg.time);
+        fmt_helper::pad3(static_cast<uint32_t>(millis.count()), dest);
+        dest.push_back(']');
+        dest.push_back(' ');
+
+        // append logger name if exists
+        if (msg.logger_name.size() > 0) {
+            dest.push_back('[');
+            fmt_helper::append_string_view(msg.logger_name, dest);
+            dest.push_back(']');
+            dest.push_back(' ');
+        }
+
+        dest.push_back('[');
+        // wrap the level name with color
+        msg.color_range_start = dest.size();
+        // fmt_helper::append_string_view(level::to_c_str(msg.level), dest);
+        fmt_helper::append_string_view(level::to_string_view(msg.level), dest);
+        msg.color_range_end = dest.size();
+        dest.push_back(']');
+        dest.push_back(' ');
+
+        // add source location if present
+        if (!msg.source.empty()) {
+            dest.push_back('[');
+            const char *filename =
+                details::short_filename_formatter<details::null_scoped_padder>::basename(
+                    msg.source.filename);
+            fmt_helper::append_string_view(filename, dest);
+            dest.push_back(':');
+            fmt_helper::append_int(msg.source.line, dest);
+            dest.push_back(']');
+            dest.push_back(' ');
+        }
+
+#ifndef SPDLOG_NO_TLS
+        // add mdc if present
+        auto &mdc_map = mdc::get_context();
+        if (!mdc_map.empty()) {
+            dest.push_back('[');
+            mdc_formatter_.format_mdc(mdc_map, dest);
+            dest.push_back(']');
+            dest.push_back(' ');
+        }
+#endif
+        // fmt_helper::append_string_view(msg.msg(), dest);
+        fmt_helper::append_string_view(msg.payload, dest);
+    }
+
+private:
+    std::chrono::seconds cache_timestamp_{0};
+    memory_buf_t cached_datetime_;
+
+#ifndef SPDLOG_NO_TLS
+    mdc_formatter<null_scoped_padder> mdc_formatter_{padding_info{}};
+#endif
+
+};
+
+}  // namespace details
+
+SPDLOG_INLINE pattern_formatter::pattern_formatter(std::string pattern,
+                                                   pattern_time_type time_type,
+                                                   std::string eol,
+                                                   custom_flags custom_user_flags)
+    : pattern_(std::move(pattern)),
+      eol_(std::move(eol)),
+      pattern_time_type_(time_type),
+      need_localtime_(false),
+      last_log_secs_(0),
+      custom_handlers_(std::move(custom_user_flags)) {
+    std::memset(&cached_tm_, 0, sizeof(cached_tm_));
+    compile_pattern_(pattern_);
+}
+
+// use by default full formatter for if pattern is not given
+SPDLOG_INLINE pattern_formatter::pattern_formatter(pattern_time_type time_type, std::string eol)
+    : pattern_("%+"),
+      eol_(std::move(eol)),
+      pattern_time_type_(time_type),
+      need_localtime_(true),
+      last_log_secs_(0) {
+    std::memset(&cached_tm_, 0, sizeof(cached_tm_));
+    formatters_.push_back(details::make_unique<details::full_formatter>(details::padding_info{}));
+}
+
+SPDLOG_INLINE std::unique_ptr<formatter> pattern_formatter::clone() const {
+    custom_flags cloned_custom_formatters;
+    for (auto &it : custom_handlers_) {
+        cloned_custom_formatters[it.first] = it.second->clone();
+    }
+    auto cloned = details::make_unique<pattern_formatter>(pattern_, pattern_time_type_, eol_,
+                                                          std::move(cloned_custom_formatters));
+    cloned->need_localtime(need_localtime_);
+#if defined(__GNUC__) && __GNUC__ < 5
+    return std::move(cloned);
+#else
+    return cloned;
+#endif
+}
+
+SPDLOG_INLINE void pattern_formatter::format(const details::log_msg &msg, memory_buf_t &dest) {
+    if (need_localtime_) {
+        const auto secs =
+            std::chrono::duration_cast<std::chrono::seconds>(msg.time.time_since_epoch());
+        if (secs != last_log_secs_) {
+            cached_tm_ = get_time_(msg);
+            last_log_secs_ = secs;
+        }
+    }
+
+    for (auto &f : formatters_) {
+        f->format(msg, cached_tm_, dest);
+    }
+    // write eol
+    details::fmt_helper::append_string_view(eol_, dest);
+}
+
+SPDLOG_INLINE void pattern_formatter::set_pattern(std::string pattern) {
+    pattern_ = std::move(pattern);
+    need_localtime_ = false;
+    compile_pattern_(pattern_);
+}
+
+SPDLOG_INLINE void pattern_formatter::need_localtime(bool need) { need_localtime_ = need; }
+
+SPDLOG_INLINE std::tm pattern_formatter::get_time_(const details::log_msg &msg) {
+    if (pattern_time_type_ == pattern_time_type::local) {
+        return details::os::localtime(log_clock::to_time_t(msg.time));
+    }
+    return details::os::gmtime(log_clock::to_time_t(msg.time));
+}
+
+template <typename Padder>
+SPDLOG_INLINE void pattern_formatter::handle_flag_(char flag, details::padding_info padding) {
+    // process custom flags
+    auto it = custom_handlers_.find(flag);
+    if (it != custom_handlers_.end()) {
+        auto custom_handler = it->second->clone();
+        custom_handler->set_padding_info(padding);
+        formatters_.push_back(std::move(custom_handler));
+        return;
+    }
+
+    // process built-in flags
+    switch (flag) {
+        case ('+'):  // default formatter
+            formatters_.push_back(details::make_unique<details::full_formatter>(padding));
+            need_localtime_ = true;
+            break;
+
+        case 'n':  // logger name
+            formatters_.push_back(details::make_unique<details::name_formatter<Padder>>(padding));
+            break;
+
+        case 'l':  // level
+            formatters_.push_back(details::make_unique<details::level_formatter<Padder>>(padding));
+            break;
+
+        case 'L':  // short level
+            formatters_.push_back(
+                details::make_unique<details::short_level_formatter<Padder>>(padding));
+            break;
+
+        case ('t'):  // thread id
+            formatters_.push_back(details::make_unique<details::t_formatter<Padder>>(padding));
+            break;
+
+        case ('v'):  // the message text
+            formatters_.push_back(details::make_unique<details::v_formatter<Padder>>(padding));
+            break;
+
+        case ('a'):  // weekday
+            formatters_.push_back(details::make_unique<details::a_formatter<Padder>>(padding));
+            need_localtime_ = true;
+            break;
+
+        case ('A'):  // short weekday
+            formatters_.push_back(details::make_unique<details::A_formatter<Padder>>(padding));
+            need_localtime_ = true;
+            break;
+
+        case ('b'):
+        case ('h'):  // month
+            formatters_.push_back(details::make_unique<details::b_formatter<Padder>>(padding));
+            need_localtime_ = true;
+            break;
+
+        case ('B'):  // short month
+            formatters_.push_back(details::make_unique<details::B_formatter<Padder>>(padding));
+            need_localtime_ = true;
+            break;
+
+        case ('c'):  // datetime
+            formatters_.push_back(details::make_unique<details::c_formatter<Padder>>(padding));
+            need_localtime_ = true;
+            break;
+
+        case ('C'):  // year 2 digits
+            formatters_.push_back(details::make_unique<details::C_formatter<Padder>>(padding));
+            need_localtime_ = true;
+            break;
+
+        case ('Y'):  // year 4 digits
+            formatters_.push_back(details::make_unique<details::Y_formatter<Padder>>(padding));
+            need_localtime_ = true;
+            break;
+
+        case ('D'):
+        case ('x'):  // datetime MM/DD/YY
+            formatters_.push_back(details::make_unique<details::D_formatter<Padder>>(padding));
+            need_localtime_ = true;
+            break;
+
+        case ('m'):  // month 1-12
+            formatters_.push_back(details::make_unique<details::m_formatter<Padder>>(padding));
+            need_localtime_ = true;
+            break;
+
+        case ('d'):  // day of month 1-31
+            formatters_.push_back(details::make_unique<details::d_formatter<Padder>>(padding));
+            need_localtime_ = true;
+            break;
+
+        case ('H'):  // hours 24
+            formatters_.push_back(details::make_unique<details::H_formatter<Padder>>(padding));
+            need_localtime_ = true;
+            break;
+
+        case ('I'):  // hours 12
+            formatters_.push_back(details::make_unique<details::I_formatter<Padder>>(padding));
+            need_localtime_ = true;
+            break;
+
+        case ('M'):  // minutes
+            formatters_.push_back(details::make_unique<details::M_formatter<Padder>>(padding));
+            need_localtime_ = true;
+            break;
+
+        case ('S'):  // seconds
+            formatters_.push_back(details::make_unique<details::S_formatter<Padder>>(padding));
+            need_localtime_ = true;
+            break;
+
+        case ('e'):  // milliseconds
+            formatters_.push_back(details::make_unique<details::e_formatter<Padder>>(padding));
+            break;
+
+        case ('f'):  // microseconds
+            formatters_.push_back(details::make_unique<details::f_formatter<Padder>>(padding));
+            break;
+
+        case ('F'):  // nanoseconds
+            formatters_.push_back(details::make_unique<details::F_formatter<Padder>>(padding));
+            break;
+
+        case ('E'):  // seconds since epoch
+            formatters_.push_back(details::make_unique<details::E_formatter<Padder>>(padding));
+            break;
+
+        case ('p'):  // am/pm
+            formatters_.push_back(details::make_unique<details::p_formatter<Padder>>(padding));
+            need_localtime_ = true;
+            break;
+
+        case ('r'):  // 12 hour clock 02:55:02 pm
+            formatters_.push_back(details::make_unique<details::r_formatter<Padder>>(padding));
+            need_localtime_ = true;
+            break;
+
+        case ('R'):  // 24-hour HH:MM time
+            formatters_.push_back(details::make_unique<details::R_formatter<Padder>>(padding));
+            need_localtime_ = true;
+            break;
+
+        case ('T'):
+        case ('X'):  // ISO 8601 time format (HH:MM:SS)
+            formatters_.push_back(details::make_unique<details::T_formatter<Padder>>(padding));
+            need_localtime_ = true;
+            break;
+
+        case ('z'):  // timezone
+            formatters_.push_back(details::make_unique<details::z_formatter<Padder>>(padding));
+            need_localtime_ = true;
+            break;
+
+        case ('P'):  // pid
+            formatters_.push_back(details::make_unique<details::pid_formatter<Padder>>(padding));
+            break;
+
+        case ('^'):  // color range start
+            formatters_.push_back(details::make_unique<details::color_start_formatter>(padding));
+            break;
+
+        case ('$'):  // color range end
+            formatters_.push_back(details::make_unique<details::color_stop_formatter>(padding));
+            break;
+
+        case ('@'):  // source location (filename:filenumber)
+            formatters_.push_back(
+                details::make_unique<details::source_location_formatter<Padder>>(padding));
+            break;
+
+        case ('s'):  // short source filename - without directory name
+            formatters_.push_back(
+                details::make_unique<details::short_filename_formatter<Padder>>(padding));
+            break;
+
+        case ('g'):  // full source filename
+            formatters_.push_back(
+                details::make_unique<details::source_filename_formatter<Padder>>(padding));
+            break;
+
+        case ('#'):  // source line number
+            formatters_.push_back(
+                details::make_unique<details::source_linenum_formatter<Padder>>(padding));
+            break;
+
+        case ('!'):  // source funcname
+            formatters_.push_back(
+                details::make_unique<details::source_funcname_formatter<Padder>>(padding));
+            break;
+
+        case ('%'):  // % char
+            formatters_.push_back(details::make_unique<details::ch_formatter>('%'));
+            break;
+
+        case ('u'):  // elapsed time since last log message in nanos
+            formatters_.push_back(
+                details::make_unique<details::elapsed_formatter<Padder, std::chrono::nanoseconds>>(
+                    padding));
+            break;
+
+        case ('i'):  // elapsed time since last log message in micros
+            formatters_.push_back(
+                details::make_unique<details::elapsed_formatter<Padder, std::chrono::microseconds>>(
+                    padding));
+            break;
+
+        case ('o'):  // elapsed time since last log message in millis
+            formatters_.push_back(
+                details::make_unique<details::elapsed_formatter<Padder, std::chrono::milliseconds>>(
+                    padding));
+            break;
+
+        case ('O'):  // elapsed time since last log message in seconds
+            formatters_.push_back(
+                details::make_unique<details::elapsed_formatter<Padder, std::chrono::seconds>>(
+                    padding));
+            break;
+
+#ifndef SPDLOG_NO_TLS  // mdc formatter requires TLS support
+        case ('&'):
+            formatters_.push_back(details::make_unique<details::mdc_formatter<Padder>>(padding));
+            break;
+#endif
+
+        default:  // Unknown flag appears as is
+            auto unknown_flag = details::make_unique<details::aggregate_formatter>();
+
+            if (!padding.truncate_) {
+                unknown_flag->add_ch('%');
+                unknown_flag->add_ch(flag);
+                formatters_.push_back((std::move(unknown_flag)));
+            }
+            // fix issue #1617 (prev char was '!' and should have been treated as funcname flag
+            // instead of truncating flag) spdlog::set_pattern("[%10!] %v") => "[      main] some
+            // message" spdlog::set_pattern("[%3!!] %v") => "[mai] some message"
+            else {
+                padding.truncate_ = false;
+                formatters_.push_back(
+                    details::make_unique<details::source_funcname_formatter<Padder>>(padding));
+                unknown_flag->add_ch(flag);
+                formatters_.push_back((std::move(unknown_flag)));
+            }
+
+            break;
+    }
+}
+
+// Extract given pad spec (e.g. %8X, %=8X, %-8!X, %8!X, %=8!X, %-8!X, %+8!X)
+// Advance the given it pass the end of the padding spec found (if any)
+// Return padding.
+SPDLOG_INLINE details::padding_info pattern_formatter::handle_padspec_(
+    std::string::const_iterator &it, std::string::const_iterator end) {
+    using details::padding_info;
+    using details::scoped_padder;
+    const size_t max_width = 64;
+    if (it == end) {
+        return padding_info{};
+    }
+
+    padding_info::pad_side side;
+    switch (*it) {
+        case '-':
+            side = padding_info::pad_side::right;
+            ++it;
+            break;
+        case '=':
+            side = padding_info::pad_side::center;
+            ++it;
+            break;
+        default:
+            side = details::padding_info::pad_side::left;
+            break;
+    }
+
+    if (it == end || !std::isdigit(static_cast<unsigned char>(*it))) {
+        return padding_info{};  // no padding if no digit found here
+    }
+
+    auto width = static_cast<size_t>(*it) - '0';
+    for (++it; it != end && std::isdigit(static_cast<unsigned char>(*it)); ++it) {
+        auto digit = static_cast<size_t>(*it) - '0';
+        width = width * 10 + digit;
+    }
+
+    // search for the optional truncate marker '!'
+    bool truncate;
+    if (it != end && *it == '!') {
+        truncate = true;
+        ++it;
+    } else {
+        truncate = false;
+    }
+    return details::padding_info{std::min<size_t>(width, max_width), side, truncate};
+}
+
+SPDLOG_INLINE void pattern_formatter::compile_pattern_(const std::string &pattern) {
+    auto end = pattern.end();
+    std::unique_ptr<details::aggregate_formatter> user_chars;
+    formatters_.clear();
+    for (auto it = pattern.begin(); it != end; ++it) {
+        if (*it == '%') {
+            if (user_chars)  // append user chars found so far
+            {
+                formatters_.push_back(std::move(user_chars));
+            }
+
+            auto padding = handle_padspec_(++it, end);
+
+            if (it != end) {
+                if (padding.enabled()) {
+                    handle_flag_<details::scoped_padder>(*it, padding);
+                } else {
+                    handle_flag_<details::null_scoped_padder>(*it, padding);
+                }
+            } else {
+                break;
+            }
+        } else  // chars not following the % sign should be displayed as is
+        {
+            if (!user_chars) {
+                user_chars = details::make_unique<details::aggregate_formatter>();
+            }
+            user_chars->add_ch(*it);
+        }
+    }
+    if (user_chars)  // append raw chars found so far
+    {
+        formatters_.push_back(std::move(user_chars));
+    }
+}
+}  // namespace spdlog
diff --git a/ext/spdlog/pattern_formatter.h b/ext/spdlog/pattern_formatter.h
new file mode 100644
index 0000000..ececd67
--- /dev/null
+++ b/ext/spdlog/pattern_formatter.h
@@ -0,0 +1,118 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#include <spdlog/common.h>
+#include <spdlog/details/log_msg.h>
+#include <spdlog/details/os.h>
+#include <spdlog/formatter.h>
+
+#include <chrono>
+#include <ctime>
+#include <memory>
+
+#include <string>
+#include <unordered_map>
+#include <vector>
+
+namespace spdlog {
+namespace details {
+
+// padding information.
+struct padding_info {
+    enum class pad_side { left, right, center };
+
+    padding_info() = default;
+    padding_info(size_t width, padding_info::pad_side side, bool truncate)
+        : width_(width),
+          side_(side),
+          truncate_(truncate),
+          enabled_(true) {}
+
+    bool enabled() const { return enabled_; }
+    size_t width_ = 0;
+    pad_side side_ = pad_side::left;
+    bool truncate_ = false;
+    bool enabled_ = false;
+};
+
+class SPDLOG_API flag_formatter {
+public:
+    explicit flag_formatter(padding_info padinfo)
+        : padinfo_(padinfo) {}
+    flag_formatter() = default;
+    virtual ~flag_formatter() = default;
+    virtual void format(const details::log_msg &msg,
+                        const std::tm &tm_time,
+                        memory_buf_t &dest) = 0;
+
+protected:
+    padding_info padinfo_;
+};
+
+}  // namespace details
+
+class SPDLOG_API custom_flag_formatter : public details::flag_formatter {
+public:
+    virtual std::unique_ptr<custom_flag_formatter> clone() const = 0;
+
+    void set_padding_info(const details::padding_info &padding) {
+        flag_formatter::padinfo_ = padding;
+    }
+};
+
+class SPDLOG_API pattern_formatter final : public formatter {
+public:
+    using custom_flags = std::unordered_map<char, std::unique_ptr<custom_flag_formatter>>;
+
+    explicit pattern_formatter(std::string pattern,
+                               pattern_time_type time_type = pattern_time_type::local,
+                               std::string eol = spdlog::details::os::default_eol,
+                               custom_flags custom_user_flags = custom_flags());
+
+    // use default pattern is not given
+    explicit pattern_formatter(pattern_time_type time_type = pattern_time_type::local,
+                               std::string eol = spdlog::details::os::default_eol);
+
+    pattern_formatter(const pattern_formatter &other) = delete;
+    pattern_formatter &operator=(const pattern_formatter &other) = delete;
+
+    std::unique_ptr<formatter> clone() const override;
+    void format(const details::log_msg &msg, memory_buf_t &dest) override;
+
+    template <typename T, typename... Args>
+    pattern_formatter &add_flag(char flag, Args &&...args) {
+        custom_handlers_[flag] = details::make_unique<T>(std::forward<Args>(args)...);
+        return *this;
+    }
+    void set_pattern(std::string pattern);
+    void need_localtime(bool need = true);
+
+private:
+    std::string pattern_;
+    std::string eol_;
+    pattern_time_type pattern_time_type_;
+    bool need_localtime_;
+    std::tm cached_tm_;
+    std::chrono::seconds last_log_secs_;
+    std::vector<std::unique_ptr<details::flag_formatter>> formatters_;
+    custom_flags custom_handlers_;
+
+    std::tm get_time_(const details::log_msg &msg);
+    template <typename Padder>
+    void handle_flag_(char flag, details::padding_info padding);
+
+    // Extract given pad spec (e.g. %8X)
+    // Advance the given it pass the end of the padding spec found (if any)
+    // Return padding.
+    static details::padding_info handle_padspec_(std::string::const_iterator &it,
+                                                 std::string::const_iterator end);
+
+    void compile_pattern_(const std::string &pattern);
+};
+}  // namespace spdlog
+
+#ifdef SPDLOG_HEADER_ONLY
+    #include "pattern_formatter-inl.h"
+#endif
diff --git a/ext/spdlog/sinks/android_sink.h b/ext/spdlog/sinks/android_sink.h
new file mode 100644
index 0000000..4435a56
--- /dev/null
+++ b/ext/spdlog/sinks/android_sink.h
@@ -0,0 +1,137 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#ifdef __ANDROID__
+
+    #include <spdlog/details/fmt_helper.h>
+    #include <spdlog/details/null_mutex.h>
+    #include <spdlog/details/os.h>
+    #include <spdlog/details/synchronous_factory.h>
+    #include <spdlog/sinks/base_sink.h>
+
+    #include <android/log.h>
+    #include <chrono>
+    #include <mutex>
+    #include <string>
+    #include <thread>
+    #include <type_traits>
+
+    #if !defined(SPDLOG_ANDROID_RETRIES)
+        #define SPDLOG_ANDROID_RETRIES 2
+    #endif
+
+namespace spdlog {
+namespace sinks {
+
+/*
+ * Android sink
+ * (logging using __android_log_write or __android_log_buf_write depending on the specified
+ * BufferID)
+ */
+template <typename Mutex, int BufferID = log_id::LOG_ID_MAIN>
+class android_sink final : public base_sink<Mutex> {
+public:
+    explicit android_sink(std::string tag = "spdlog", bool use_raw_msg = false)
+        : tag_(std::move(tag)),
+          use_raw_msg_(use_raw_msg) {}
+
+protected:
+    void sink_it_(const details::log_msg &msg) override {
+        const android_LogPriority priority = convert_to_android_(msg.level);
+        memory_buf_t formatted;
+        if (use_raw_msg_) {
+            details::fmt_helper::append_string_view(msg.payload, formatted);
+        } else {
+            base_sink<Mutex>::formatter_->format(msg, formatted);
+        }
+        formatted.push_back('\0');
+        const char *msg_output = formatted.data();
+
+        // See system/core/liblog/logger_write.c for explanation of return value
+        int ret = android_log(priority, tag_.c_str(), msg_output);
+        if (ret == -EPERM) {
+            return;  // !__android_log_is_loggable
+        }
+        int retry_count = 0;
+        while ((ret == -11 /*EAGAIN*/) && (retry_count < SPDLOG_ANDROID_RETRIES)) {
+            details::os::sleep_for_millis(5);
+            ret = android_log(priority, tag_.c_str(), msg_output);
+            retry_count++;
+        }
+
+        if (ret < 0) {
+            throw_spdlog_ex("logging to Android failed", ret);
+        }
+    }
+
+    void flush_() override {}
+
+private:
+    // There might be liblog versions used, that do not support __android_log_buf_write. So we only
+    // compile and link against
+    // __android_log_buf_write, if user explicitly provides a non-default log buffer. Otherwise,
+    // when using the default log buffer, always log via __android_log_write.
+    template <int ID = BufferID>
+    typename std::enable_if<ID == static_cast<int>(log_id::LOG_ID_MAIN), int>::type android_log(
+        int prio, const char *tag, const char *text) {
+        return __android_log_write(prio, tag, text);
+    }
+
+    template <int ID = BufferID>
+    typename std::enable_if<ID != static_cast<int>(log_id::LOG_ID_MAIN), int>::type android_log(
+        int prio, const char *tag, const char *text) {
+        return __android_log_buf_write(ID, prio, tag, text);
+    }
+
+    static android_LogPriority convert_to_android_(spdlog::level::level_enum level) {
+        switch (level) {
+            case spdlog::level::trace:
+                return ANDROID_LOG_VERBOSE;
+            case spdlog::level::debug:
+                return ANDROID_LOG_DEBUG;
+            case spdlog::level::info:
+                return ANDROID_LOG_INFO;
+            case spdlog::level::warn:
+                return ANDROID_LOG_WARN;
+            case spdlog::level::err:
+                return ANDROID_LOG_ERROR;
+            case spdlog::level::critical:
+                return ANDROID_LOG_FATAL;
+            default:
+                return ANDROID_LOG_DEFAULT;
+        }
+    }
+
+    std::string tag_;
+    bool use_raw_msg_;
+};
+
+using android_sink_mt = android_sink<std::mutex>;
+using android_sink_st = android_sink<details::null_mutex>;
+
+template <int BufferId = log_id::LOG_ID_MAIN>
+using android_sink_buf_mt = android_sink<std::mutex, BufferId>;
+template <int BufferId = log_id::LOG_ID_MAIN>
+using android_sink_buf_st = android_sink<details::null_mutex, BufferId>;
+
+}  // namespace sinks
+
+// Create and register android syslog logger
+
+template <typename Factory = spdlog::synchronous_factory>
+inline std::shared_ptr<logger> android_logger_mt(const std::string &logger_name,
+                                                 const std::string &tag = "spdlog") {
+    return Factory::template create<sinks::android_sink_mt>(logger_name, tag);
+}
+
+template <typename Factory = spdlog::synchronous_factory>
+inline std::shared_ptr<logger> android_logger_st(const std::string &logger_name,
+                                                 const std::string &tag = "spdlog") {
+    return Factory::template create<sinks::android_sink_st>(logger_name, tag);
+}
+
+}  // namespace spdlog
+
+#endif  // __ANDROID__
diff --git a/ext/spdlog/sinks/ansicolor_sink-inl.h b/ext/spdlog/sinks/ansicolor_sink-inl.h
new file mode 100644
index 0000000..8fd3078
--- /dev/null
+++ b/ext/spdlog/sinks/ansicolor_sink-inl.h
@@ -0,0 +1,135 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#ifndef SPDLOG_HEADER_ONLY
+    #include <spdlog/sinks/ansicolor_sink.h>
+#endif
+
+#include <spdlog/details/os.h>
+#include <spdlog/pattern_formatter.h>
+
+namespace spdlog {
+namespace sinks {
+
+template <typename ConsoleMutex>
+SPDLOG_INLINE ansicolor_sink<ConsoleMutex>::ansicolor_sink(FILE *target_file, color_mode mode)
+    : target_file_(target_file),
+      mutex_(ConsoleMutex::mutex()),
+      formatter_(details::make_unique<spdlog::pattern_formatter>())
+
+{
+    set_color_mode(mode);
+    colors_.at(level::trace) = to_string_(white);
+    colors_.at(level::debug) = to_string_(cyan);
+    colors_.at(level::info) = to_string_(green);
+    colors_.at(level::warn) = to_string_(yellow_bold);
+    colors_.at(level::err) = to_string_(red_bold);
+    colors_.at(level::critical) = to_string_(bold_on_red);
+    colors_.at(level::off) = to_string_(reset);
+}
+
+template <typename ConsoleMutex>
+SPDLOG_INLINE void ansicolor_sink<ConsoleMutex>::set_color(level::level_enum color_level,
+                                                           string_view_t color) {
+    std::lock_guard<mutex_t> lock(mutex_);
+    colors_.at(static_cast<size_t>(color_level)) = to_string_(color);
+}
+
+template <typename ConsoleMutex>
+SPDLOG_INLINE void ansicolor_sink<ConsoleMutex>::log(const details::log_msg &msg) {
+    // Wrap the originally formatted message in color codes.
+    // If color is not supported in the terminal, log as is instead.
+    std::lock_guard<mutex_t> lock(mutex_);
+    msg.color_range_start = 0;
+    msg.color_range_end = 0;
+    memory_buf_t formatted;
+    formatter_->format(msg, formatted);
+    if (should_do_colors_ && msg.color_range_end > msg.color_range_start) {
+        // before color range
+        print_range_(formatted, 0, msg.color_range_start);
+        // in color range
+        print_ccode_(colors_.at(static_cast<size_t>(msg.level)));
+        print_range_(formatted, msg.color_range_start, msg.color_range_end);
+        print_ccode_(reset);
+        // after color range
+        print_range_(formatted, msg.color_range_end, formatted.size());
+    } else  // no color
+    {
+        print_range_(formatted, 0, formatted.size());
+    }
+    fflush(target_file_);
+}
+
+template <typename ConsoleMutex>
+SPDLOG_INLINE void ansicolor_sink<ConsoleMutex>::flush() {
+    std::lock_guard<mutex_t> lock(mutex_);
+    fflush(target_file_);
+}
+
+template <typename ConsoleMutex>
+SPDLOG_INLINE void ansicolor_sink<ConsoleMutex>::set_pattern(const std::string &pattern) {
+    std::lock_guard<mutex_t> lock(mutex_);
+    formatter_ = std::unique_ptr<spdlog::formatter>(new pattern_formatter(pattern));
+}
+
+template <typename ConsoleMutex>
+SPDLOG_INLINE void ansicolor_sink<ConsoleMutex>::set_formatter(
+    std::unique_ptr<spdlog::formatter> sink_formatter) {
+    std::lock_guard<mutex_t> lock(mutex_);
+    formatter_ = std::move(sink_formatter);
+}
+
+template <typename ConsoleMutex>
+SPDLOG_INLINE bool ansicolor_sink<ConsoleMutex>::should_color() {
+    return should_do_colors_;
+}
+
+template <typename ConsoleMutex>
+SPDLOG_INLINE void ansicolor_sink<ConsoleMutex>::set_color_mode(color_mode mode) {
+    switch (mode) {
+        case color_mode::always:
+            should_do_colors_ = true;
+            return;
+        case color_mode::automatic:
+            should_do_colors_ =
+                details::os::in_terminal(target_file_) && details::os::is_color_terminal();
+            return;
+        case color_mode::never:
+            should_do_colors_ = false;
+            return;
+        default:
+            should_do_colors_ = false;
+    }
+}
+
+template <typename ConsoleMutex>
+SPDLOG_INLINE void ansicolor_sink<ConsoleMutex>::print_ccode_(const string_view_t &color_code) {
+    details::os::fwrite_bytes(color_code.data(), color_code.size(), target_file_);
+}
+
+template <typename ConsoleMutex>
+SPDLOG_INLINE void ansicolor_sink<ConsoleMutex>::print_range_(const memory_buf_t &formatted,
+                                                              size_t start,
+                                                              size_t end) {
+    details::os::fwrite_bytes(formatted.data() + start, end - start, target_file_);
+}
+
+template <typename ConsoleMutex>
+SPDLOG_INLINE std::string ansicolor_sink<ConsoleMutex>::to_string_(const string_view_t &sv) {
+    return std::string(sv.data(), sv.size());
+}
+
+// ansicolor_stdout_sink
+template <typename ConsoleMutex>
+SPDLOG_INLINE ansicolor_stdout_sink<ConsoleMutex>::ansicolor_stdout_sink(color_mode mode)
+    : ansicolor_sink<ConsoleMutex>(stdout, mode) {}
+
+// ansicolor_stderr_sink
+template <typename ConsoleMutex>
+SPDLOG_INLINE ansicolor_stderr_sink<ConsoleMutex>::ansicolor_stderr_sink(color_mode mode)
+    : ansicolor_sink<ConsoleMutex>(stderr, mode) {}
+
+}  // namespace sinks
+}  // namespace spdlog
diff --git a/ext/spdlog/sinks/ansicolor_sink.h b/ext/spdlog/sinks/ansicolor_sink.h
new file mode 100644
index 0000000..d0dadd7
--- /dev/null
+++ b/ext/spdlog/sinks/ansicolor_sink.h
@@ -0,0 +1,115 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#include <array>
+#include <memory>
+#include <mutex>
+#include <spdlog/details/console_globals.h>
+#include <spdlog/details/null_mutex.h>
+#include <spdlog/sinks/sink.h>
+#include <string>
+
+namespace spdlog {
+namespace sinks {
+
+/**
+ * This sink prefixes the output with an ANSI escape sequence color code
+ * depending on the severity
+ * of the message.
+ * If no color terminal detected, omit the escape codes.
+ */
+
+template <typename ConsoleMutex>
+class ansicolor_sink : public sink {
+public:
+    using mutex_t = typename ConsoleMutex::mutex_t;
+    ansicolor_sink(FILE *target_file, color_mode mode);
+    ~ansicolor_sink() override = default;
+
+    ansicolor_sink(const ansicolor_sink &other) = delete;
+    ansicolor_sink(ansicolor_sink &&other) = delete;
+
+    ansicolor_sink &operator=(const ansicolor_sink &other) = delete;
+    ansicolor_sink &operator=(ansicolor_sink &&other) = delete;
+
+    void set_color(level::level_enum color_level, string_view_t color);
+    void set_color_mode(color_mode mode);
+    bool should_color();
+
+    void log(const details::log_msg &msg) override;
+    void flush() override;
+    void set_pattern(const std::string &pattern) final override;
+    void set_formatter(std::unique_ptr<spdlog::formatter> sink_formatter) override;
+
+    // Formatting codes
+    const string_view_t reset = "\033[m";
+    const string_view_t bold = "\033[1m";
+    const string_view_t dark = "\033[2m";
+    const string_view_t underline = "\033[4m";
+    const string_view_t blink = "\033[5m";
+    const string_view_t reverse = "\033[7m";
+    const string_view_t concealed = "\033[8m";
+    const string_view_t clear_line = "\033[K";
+
+    // Foreground colors
+    const string_view_t black = "\033[30m";
+    const string_view_t red = "\033[31m";
+    const string_view_t green = "\033[32m";
+    const string_view_t yellow = "\033[33m";
+    const string_view_t blue = "\033[34m";
+    const string_view_t magenta = "\033[35m";
+    const string_view_t cyan = "\033[36m";
+    const string_view_t white = "\033[37m";
+
+    /// Background colors
+    const string_view_t on_black = "\033[40m";
+    const string_view_t on_red = "\033[41m";
+    const string_view_t on_green = "\033[42m";
+    const string_view_t on_yellow = "\033[43m";
+    const string_view_t on_blue = "\033[44m";
+    const string_view_t on_magenta = "\033[45m";
+    const string_view_t on_cyan = "\033[46m";
+    const string_view_t on_white = "\033[47m";
+
+    /// Bold colors
+    const string_view_t yellow_bold = "\033[33m\033[1m";
+    const string_view_t red_bold = "\033[31m\033[1m";
+    const string_view_t bold_on_red = "\033[1m\033[41m";
+
+private:
+    FILE *target_file_;
+    mutex_t &mutex_;
+    bool should_do_colors_;
+    std::unique_ptr<spdlog::formatter> formatter_;
+    std::array<std::string, level::n_levels> colors_;
+    void print_ccode_(const string_view_t &color_code);
+    void print_range_(const memory_buf_t &formatted, size_t start, size_t end);
+    static std::string to_string_(const string_view_t &sv);
+};
+
+template <typename ConsoleMutex>
+class ansicolor_stdout_sink : public ansicolor_sink<ConsoleMutex> {
+public:
+    explicit ansicolor_stdout_sink(color_mode mode = color_mode::automatic);
+};
+
+template <typename ConsoleMutex>
+class ansicolor_stderr_sink : public ansicolor_sink<ConsoleMutex> {
+public:
+    explicit ansicolor_stderr_sink(color_mode mode = color_mode::automatic);
+};
+
+using ansicolor_stdout_sink_mt = ansicolor_stdout_sink<details::console_mutex>;
+using ansicolor_stdout_sink_st = ansicolor_stdout_sink<details::console_nullmutex>;
+
+using ansicolor_stderr_sink_mt = ansicolor_stderr_sink<details::console_mutex>;
+using ansicolor_stderr_sink_st = ansicolor_stderr_sink<details::console_nullmutex>;
+
+}  // namespace sinks
+}  // namespace spdlog
+
+#ifdef SPDLOG_HEADER_ONLY
+    #include "ansicolor_sink-inl.h"
+#endif
diff --git a/ext/spdlog/sinks/base_sink-inl.h b/ext/spdlog/sinks/base_sink-inl.h
new file mode 100644
index 0000000..ada161b
--- /dev/null
+++ b/ext/spdlog/sinks/base_sink-inl.h
@@ -0,0 +1,59 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#ifndef SPDLOG_HEADER_ONLY
+    #include <spdlog/sinks/base_sink.h>
+#endif
+
+#include <spdlog/common.h>
+#include <spdlog/pattern_formatter.h>
+
+#include <memory>
+#include <mutex>
+
+template <typename Mutex>
+SPDLOG_INLINE spdlog::sinks::base_sink<Mutex>::base_sink()
+    : formatter_{details::make_unique<spdlog::pattern_formatter>()} {}
+
+template <typename Mutex>
+SPDLOG_INLINE spdlog::sinks::base_sink<Mutex>::base_sink(
+    std::unique_ptr<spdlog::formatter> formatter)
+    : formatter_{std::move(formatter)} {}
+
+template <typename Mutex>
+void SPDLOG_INLINE spdlog::sinks::base_sink<Mutex>::log(const details::log_msg &msg) {
+    std::lock_guard<Mutex> lock(mutex_);
+    sink_it_(msg);
+}
+
+template <typename Mutex>
+void SPDLOG_INLINE spdlog::sinks::base_sink<Mutex>::flush() {
+    std::lock_guard<Mutex> lock(mutex_);
+    flush_();
+}
+
+template <typename Mutex>
+void SPDLOG_INLINE spdlog::sinks::base_sink<Mutex>::set_pattern(const std::string &pattern) {
+    std::lock_guard<Mutex> lock(mutex_);
+    set_pattern_(pattern);
+}
+
+template <typename Mutex>
+void SPDLOG_INLINE
+spdlog::sinks::base_sink<Mutex>::set_formatter(std::unique_ptr<spdlog::formatter> sink_formatter) {
+    std::lock_guard<Mutex> lock(mutex_);
+    set_formatter_(std::move(sink_formatter));
+}
+
+template <typename Mutex>
+void SPDLOG_INLINE spdlog::sinks::base_sink<Mutex>::set_pattern_(const std::string &pattern) {
+    set_formatter_(details::make_unique<spdlog::pattern_formatter>(pattern));
+}
+
+template <typename Mutex>
+void SPDLOG_INLINE
+spdlog::sinks::base_sink<Mutex>::set_formatter_(std::unique_ptr<spdlog::formatter> sink_formatter) {
+    formatter_ = std::move(sink_formatter);
+}
diff --git a/ext/spdlog/sinks/base_sink.h b/ext/spdlog/sinks/base_sink.h
new file mode 100644
index 0000000..1b4bb06
--- /dev/null
+++ b/ext/spdlog/sinks/base_sink.h
@@ -0,0 +1,51 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+//
+// base sink templated over a mutex (either dummy or real)
+// concrete implementation should override the sink_it_() and flush_()  methods.
+// locking is taken care of in this class - no locking needed by the
+// implementers..
+//
+
+#include <spdlog/common.h>
+#include <spdlog/details/log_msg.h>
+#include <spdlog/sinks/sink.h>
+
+namespace spdlog {
+namespace sinks {
+template <typename Mutex>
+class SPDLOG_API base_sink : public sink {
+public:
+    base_sink();
+    explicit base_sink(std::unique_ptr<spdlog::formatter> formatter);
+    ~base_sink() override = default;
+
+    base_sink(const base_sink &) = delete;
+    base_sink(base_sink &&) = delete;
+
+    base_sink &operator=(const base_sink &) = delete;
+    base_sink &operator=(base_sink &&) = delete;
+
+    void log(const details::log_msg &msg) final override;
+    void flush() final override;
+    void set_pattern(const std::string &pattern) final override;
+    void set_formatter(std::unique_ptr<spdlog::formatter> sink_formatter) final override;
+
+protected:
+    // sink formatter
+    std::unique_ptr<spdlog::formatter> formatter_;
+    Mutex mutex_;
+
+    virtual void sink_it_(const details::log_msg &msg) = 0;
+    virtual void flush_() = 0;
+    virtual void set_pattern_(const std::string &pattern);
+    virtual void set_formatter_(std::unique_ptr<spdlog::formatter> sink_formatter);
+};
+}  // namespace sinks
+}  // namespace spdlog
+
+#ifdef SPDLOG_HEADER_ONLY
+    #include "base_sink-inl.h"
+#endif
diff --git a/ext/spdlog/sinks/basic_file_sink-inl.h b/ext/spdlog/sinks/basic_file_sink-inl.h
new file mode 100644
index 0000000..ce0ddad
--- /dev/null
+++ b/ext/spdlog/sinks/basic_file_sink-inl.h
@@ -0,0 +1,48 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#ifndef SPDLOG_HEADER_ONLY
+    #include <spdlog/sinks/basic_file_sink.h>
+#endif
+
+#include <spdlog/common.h>
+#include <spdlog/details/os.h>
+
+namespace spdlog {
+namespace sinks {
+
+template <typename Mutex>
+SPDLOG_INLINE basic_file_sink<Mutex>::basic_file_sink(const filename_t &filename,
+                                                      bool truncate,
+                                                      const file_event_handlers &event_handlers)
+    : file_helper_{event_handlers} {
+    file_helper_.open(filename, truncate);
+}
+
+template <typename Mutex>
+SPDLOG_INLINE const filename_t &basic_file_sink<Mutex>::filename() const {
+    return file_helper_.filename();
+}
+
+template <typename Mutex>
+SPDLOG_INLINE void basic_file_sink<Mutex>::truncate() {
+    std::lock_guard<Mutex> lock(base_sink<Mutex>::mutex_);
+    file_helper_.reopen(true);
+}
+
+template <typename Mutex>
+SPDLOG_INLINE void basic_file_sink<Mutex>::sink_it_(const details::log_msg &msg) {
+    memory_buf_t formatted;
+    base_sink<Mutex>::formatter_->format(msg, formatted);
+    file_helper_.write(formatted);
+}
+
+template <typename Mutex>
+SPDLOG_INLINE void basic_file_sink<Mutex>::flush_() {
+    file_helper_.flush();
+}
+
+}  // namespace sinks
+}  // namespace spdlog
diff --git a/ext/spdlog/sinks/basic_file_sink.h b/ext/spdlog/sinks/basic_file_sink.h
new file mode 100644
index 0000000..48c0767
--- /dev/null
+++ b/ext/spdlog/sinks/basic_file_sink.h
@@ -0,0 +1,66 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#include <spdlog/details/file_helper.h>
+#include <spdlog/details/null_mutex.h>
+#include <spdlog/details/synchronous_factory.h>
+#include <spdlog/sinks/base_sink.h>
+
+#include <mutex>
+#include <string>
+
+namespace spdlog {
+namespace sinks {
+/*
+ * Trivial file sink with single file as target
+ */
+template <typename Mutex>
+class basic_file_sink final : public base_sink<Mutex> {
+public:
+    explicit basic_file_sink(const filename_t &filename,
+                             bool truncate = false,
+                             const file_event_handlers &event_handlers = {});
+    const filename_t &filename() const;
+    void truncate();
+
+protected:
+    void sink_it_(const details::log_msg &msg) override;
+    void flush_() override;
+
+private:
+    details::file_helper file_helper_;
+};
+
+using basic_file_sink_mt = basic_file_sink<std::mutex>;
+using basic_file_sink_st = basic_file_sink<details::null_mutex>;
+
+}  // namespace sinks
+
+//
+// factory functions
+//
+template <typename Factory = spdlog::synchronous_factory>
+inline std::shared_ptr<logger> basic_logger_mt(const std::string &logger_name,
+                                               const filename_t &filename,
+                                               bool truncate = false,
+                                               const file_event_handlers &event_handlers = {}) {
+    return Factory::template create<sinks::basic_file_sink_mt>(logger_name, filename, truncate,
+                                                               event_handlers);
+}
+
+template <typename Factory = spdlog::synchronous_factory>
+inline std::shared_ptr<logger> basic_logger_st(const std::string &logger_name,
+                                               const filename_t &filename,
+                                               bool truncate = false,
+                                               const file_event_handlers &event_handlers = {}) {
+    return Factory::template create<sinks::basic_file_sink_st>(logger_name, filename, truncate,
+                                                               event_handlers);
+}
+
+}  // namespace spdlog
+
+#ifdef SPDLOG_HEADER_ONLY
+    #include "basic_file_sink-inl.h"
+#endif
diff --git a/ext/spdlog/sinks/callback_sink.h b/ext/spdlog/sinks/callback_sink.h
new file mode 100644
index 0000000..5f8b6bc
--- /dev/null
+++ b/ext/spdlog/sinks/callback_sink.h
@@ -0,0 +1,56 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#include <spdlog/details/null_mutex.h>
+#include <spdlog/details/synchronous_factory.h>
+#include <spdlog/sinks/base_sink.h>
+
+#include <mutex>
+#include <string>
+
+namespace spdlog {
+
+// callbacks type
+typedef std::function<void(const details::log_msg &msg)> custom_log_callback;
+
+namespace sinks {
+/*
+ * Trivial callback sink, gets a callback function and calls it on each log
+ */
+template <typename Mutex>
+class callback_sink final : public base_sink<Mutex> {
+public:
+    explicit callback_sink(const custom_log_callback &callback)
+        : callback_{callback} {}
+
+protected:
+    void sink_it_(const details::log_msg &msg) override { callback_(msg); }
+    void flush_() override{}
+
+private:
+    custom_log_callback callback_;
+};
+
+using callback_sink_mt = callback_sink<std::mutex>;
+using callback_sink_st = callback_sink<details::null_mutex>;
+
+}  // namespace sinks
+
+//
+// factory functions
+//
+template <typename Factory = spdlog::synchronous_factory>
+inline std::shared_ptr<logger> callback_logger_mt(const std::string &logger_name,
+                                                  const custom_log_callback &callback) {
+    return Factory::template create<sinks::callback_sink_mt>(logger_name, callback);
+}
+
+template <typename Factory = spdlog::synchronous_factory>
+inline std::shared_ptr<logger> callback_logger_st(const std::string &logger_name,
+                                                  const custom_log_callback &callback) {
+    return Factory::template create<sinks::callback_sink_st>(logger_name, callback);
+}
+
+}  // namespace spdlog
diff --git a/ext/spdlog/sinks/daily_file_sink.h b/ext/spdlog/sinks/daily_file_sink.h
new file mode 100644
index 0000000..615c9f7
--- /dev/null
+++ b/ext/spdlog/sinks/daily_file_sink.h
@@ -0,0 +1,254 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#include <spdlog/common.h>
+#include <spdlog/details/circular_q.h>
+#include <spdlog/details/file_helper.h>
+#include <spdlog/details/null_mutex.h>
+#include <spdlog/details/os.h>
+#include <spdlog/details/synchronous_factory.h>
+#include <spdlog/fmt/chrono.h>
+#include <spdlog/fmt/fmt.h>
+#include <spdlog/sinks/base_sink.h>
+
+#include <chrono>
+#include <cstdio>
+#include <iomanip>
+#include <mutex>
+#include <sstream>
+#include <string>
+
+namespace spdlog {
+namespace sinks {
+
+/*
+ * Generator of daily log file names in format basename.YYYY-MM-DD.ext
+ */
+struct daily_filename_calculator {
+    // Create filename for the form basename.YYYY-MM-DD
+    static filename_t calc_filename(const filename_t &filename, const tm &now_tm) {
+        filename_t basename, ext;
+        std::tie(basename, ext) = details::file_helper::split_by_extension(filename);
+        return fmt_lib::format(SPDLOG_FMT_STRING(SPDLOG_FILENAME_T("{}_{:04d}-{:02d}-{:02d}{}")),
+                               basename, now_tm.tm_year + 1900, now_tm.tm_mon + 1, now_tm.tm_mday,
+                               ext);
+    }
+};
+
+/*
+ * Generator of daily log file names with strftime format.
+ * Usages:
+ *    auto sink =
+ * std::make_shared<spdlog::sinks::daily_file_format_sink_mt>("myapp-%Y-%m-%d:%H:%M:%S.log", hour,
+ * minute);" auto logger = spdlog::daily_logger_format_mt("loggername, "myapp-%Y-%m-%d:%X.log",
+ * hour,  minute)"
+ *
+ */
+struct daily_filename_format_calculator {
+    static filename_t calc_filename(const filename_t &file_path, const tm &now_tm) {
+#if defined(_WIN32) && defined(SPDLOG_WCHAR_FILENAMES)
+        std::wstringstream stream;
+#else
+        std::stringstream stream;
+#endif
+        stream << std::put_time(&now_tm, file_path.c_str());
+        return stream.str();
+    }
+};
+
+/*
+ * Rotating file sink based on date.
+ * If truncate != false , the created file will be truncated.
+ * If max_files > 0, retain only the last max_files and delete previous.
+ * Note that old log files from previous executions will not be deleted by this class,
+ * rotation and deletion is only applied while the program is running.
+ */
+template <typename Mutex, typename FileNameCalc = daily_filename_calculator>
+class daily_file_sink final : public base_sink<Mutex> {
+public:
+    // create daily file sink which rotates on given time
+    daily_file_sink(filename_t base_filename,
+                    int rotation_hour,
+                    int rotation_minute,
+                    bool truncate = false,
+                    uint16_t max_files = 0,
+                    const file_event_handlers &event_handlers = {})
+        : base_filename_(std::move(base_filename)),
+          rotation_h_(rotation_hour),
+          rotation_m_(rotation_minute),
+          file_helper_{event_handlers},
+          truncate_(truncate),
+          max_files_(max_files),
+          filenames_q_() {
+        if (rotation_hour < 0 || rotation_hour > 23 || rotation_minute < 0 ||
+            rotation_minute > 59) {
+            throw_spdlog_ex("daily_file_sink: Invalid rotation time in ctor");
+        }
+
+        auto now = log_clock::now();
+        auto filename = FileNameCalc::calc_filename(base_filename_, now_tm(now));
+        file_helper_.open(filename, truncate_);
+        rotation_tp_ = next_rotation_tp_();
+
+        if (max_files_ > 0) {
+            init_filenames_q_();
+        }
+    }
+
+    filename_t filename() {
+        std::lock_guard<Mutex> lock(base_sink<Mutex>::mutex_);
+        return file_helper_.filename();
+    }
+
+protected:
+    void sink_it_(const details::log_msg &msg) override {
+        auto time = msg.time;
+        bool should_rotate = time >= rotation_tp_;
+        if (should_rotate) {
+            auto filename = FileNameCalc::calc_filename(base_filename_, now_tm(time));
+            file_helper_.open(filename, truncate_);
+            rotation_tp_ = next_rotation_tp_();
+        }
+        memory_buf_t formatted;
+        base_sink<Mutex>::formatter_->format(msg, formatted);
+        file_helper_.write(formatted);
+
+        // Do the cleaning only at the end because it might throw on failure.
+        if (should_rotate && max_files_ > 0) {
+            delete_old_();
+        }
+    }
+
+    void flush_() override { file_helper_.flush(); }
+
+private:
+    void init_filenames_q_() {
+        using details::os::path_exists;
+
+        filenames_q_ = details::circular_q<filename_t>(static_cast<size_t>(max_files_));
+        std::vector<filename_t> filenames;
+        auto now = log_clock::now();
+        while (filenames.size() < max_files_) {
+            auto filename = FileNameCalc::calc_filename(base_filename_, now_tm(now));
+            if (!path_exists(filename)) {
+                break;
+            }
+            filenames.emplace_back(filename);
+            now -= std::chrono::hours(24);
+        }
+        for (auto iter = filenames.rbegin(); iter != filenames.rend(); ++iter) {
+            filenames_q_.push_back(std::move(*iter));
+        }
+    }
+
+    tm now_tm(log_clock::time_point tp) {
+        time_t tnow = log_clock::to_time_t(tp);
+        return spdlog::details::os::localtime(tnow);
+    }
+
+    log_clock::time_point next_rotation_tp_() {
+        auto now = log_clock::now();
+        tm date = now_tm(now);
+        date.tm_hour = rotation_h_;
+        date.tm_min = rotation_m_;
+        date.tm_sec = 0;
+        auto rotation_time = log_clock::from_time_t(std::mktime(&date));
+        if (rotation_time > now) {
+            return rotation_time;
+        }
+        return {rotation_time + std::chrono::hours(24)};
+    }
+
+    // Delete the file N rotations ago.
+    // Throw spdlog_ex on failure to delete the old file.
+    void delete_old_() {
+        using details::os::filename_to_str;
+        using details::os::remove_if_exists;
+
+        filename_t current_file = file_helper_.filename();
+        if (filenames_q_.full()) {
+            auto old_filename = std::move(filenames_q_.front());
+            filenames_q_.pop_front();
+            bool ok = remove_if_exists(old_filename) == 0;
+            if (!ok) {
+                filenames_q_.push_back(std::move(current_file));
+                throw_spdlog_ex("Failed removing daily file " + filename_to_str(old_filename),
+                                errno);
+            }
+        }
+        filenames_q_.push_back(std::move(current_file));
+    }
+
+    filename_t base_filename_;
+    int rotation_h_;
+    int rotation_m_;
+    log_clock::time_point rotation_tp_;
+    details::file_helper file_helper_;
+    bool truncate_;
+    uint16_t max_files_;
+    details::circular_q<filename_t> filenames_q_;
+};
+
+using daily_file_sink_mt = daily_file_sink<std::mutex>;
+using daily_file_sink_st = daily_file_sink<details::null_mutex>;
+using daily_file_format_sink_mt = daily_file_sink<std::mutex, daily_filename_format_calculator>;
+using daily_file_format_sink_st =
+    daily_file_sink<details::null_mutex, daily_filename_format_calculator>;
+
+}  // namespace sinks
+
+//
+// factory functions
+//
+template <typename Factory = spdlog::synchronous_factory>
+inline std::shared_ptr<logger> daily_logger_mt(const std::string &logger_name,
+                                               const filename_t &filename,
+                                               int hour = 0,
+                                               int minute = 0,
+                                               bool truncate = false,
+                                               uint16_t max_files = 0,
+                                               const file_event_handlers &event_handlers = {}) {
+    return Factory::template create<sinks::daily_file_sink_mt>(logger_name, filename, hour, minute,
+                                                               truncate, max_files, event_handlers);
+}
+
+template <typename Factory = spdlog::synchronous_factory>
+inline std::shared_ptr<logger> daily_logger_format_mt(
+    const std::string &logger_name,
+    const filename_t &filename,
+    int hour = 0,
+    int minute = 0,
+    bool truncate = false,
+    uint16_t max_files = 0,
+    const file_event_handlers &event_handlers = {}) {
+    return Factory::template create<sinks::daily_file_format_sink_mt>(
+        logger_name, filename, hour, minute, truncate, max_files, event_handlers);
+}
+
+template <typename Factory = spdlog::synchronous_factory>
+inline std::shared_ptr<logger> daily_logger_st(const std::string &logger_name,
+                                               const filename_t &filename,
+                                               int hour = 0,
+                                               int minute = 0,
+                                               bool truncate = false,
+                                               uint16_t max_files = 0,
+                                               const file_event_handlers &event_handlers = {}) {
+    return Factory::template create<sinks::daily_file_sink_st>(logger_name, filename, hour, minute,
+                                                               truncate, max_files, event_handlers);
+}
+
+template <typename Factory = spdlog::synchronous_factory>
+inline std::shared_ptr<logger> daily_logger_format_st(
+    const std::string &logger_name,
+    const filename_t &filename,
+    int hour = 0,
+    int minute = 0,
+    bool truncate = false,
+    uint16_t max_files = 0,
+    const file_event_handlers &event_handlers = {}) {
+    return Factory::template create<sinks::daily_file_format_sink_st>(
+        logger_name, filename, hour, minute, truncate, max_files, event_handlers);
+}
+}  // namespace spdlog
diff --git a/ext/spdlog/sinks/dist_sink.h b/ext/spdlog/sinks/dist_sink.h
new file mode 100644
index 0000000..69c4971
--- /dev/null
+++ b/ext/spdlog/sinks/dist_sink.h
@@ -0,0 +1,81 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#include "base_sink.h"
+#include <spdlog/details/log_msg.h>
+#include <spdlog/details/null_mutex.h>
+#include <spdlog/pattern_formatter.h>
+
+#include <algorithm>
+#include <memory>
+#include <mutex>
+#include <vector>
+
+// Distribution sink (mux). Stores a vector of sinks which get called when log
+// is called
+
+namespace spdlog {
+namespace sinks {
+
+template <typename Mutex>
+class dist_sink : public base_sink<Mutex> {
+public:
+    dist_sink() = default;
+    explicit dist_sink(std::vector<std::shared_ptr<sink>> sinks)
+        : sinks_(sinks) {}
+
+    dist_sink(const dist_sink &) = delete;
+    dist_sink &operator=(const dist_sink &) = delete;
+
+    void add_sink(std::shared_ptr<sink> sub_sink) {
+        std::lock_guard<Mutex> lock(base_sink<Mutex>::mutex_);
+        sinks_.push_back(sub_sink);
+    }
+
+    void remove_sink(std::shared_ptr<sink> sub_sink) {
+        std::lock_guard<Mutex> lock(base_sink<Mutex>::mutex_);
+        sinks_.erase(std::remove(sinks_.begin(), sinks_.end(), sub_sink), sinks_.end());
+    }
+
+    void set_sinks(std::vector<std::shared_ptr<sink>> sinks) {
+        std::lock_guard<Mutex> lock(base_sink<Mutex>::mutex_);
+        sinks_ = std::move(sinks);
+    }
+
+    std::vector<std::shared_ptr<sink>> &sinks() { return sinks_; }
+
+protected:
+    void sink_it_(const details::log_msg &msg) override {
+        for (auto &sub_sink : sinks_) {
+            if (sub_sink->should_log(msg.level)) {
+                sub_sink->log(msg);
+            }
+        }
+    }
+
+    void flush_() override {
+        for (auto &sub_sink : sinks_) {
+            sub_sink->flush();
+        }
+    }
+
+    void set_pattern_(const std::string &pattern) override {
+        set_formatter_(details::make_unique<spdlog::pattern_formatter>(pattern));
+    }
+
+    void set_formatter_(std::unique_ptr<spdlog::formatter> sink_formatter) override {
+        base_sink<Mutex>::formatter_ = std::move(sink_formatter);
+        for (auto &sub_sink : sinks_) {
+            sub_sink->set_formatter(base_sink<Mutex>::formatter_->clone());
+        }
+    }
+    std::vector<std::shared_ptr<sink>> sinks_;
+};
+
+using dist_sink_mt = dist_sink<std::mutex>;
+using dist_sink_st = dist_sink<details::null_mutex>;
+
+}  // namespace sinks
+}  // namespace spdlog
diff --git a/ext/spdlog/sinks/dup_filter_sink.h b/ext/spdlog/sinks/dup_filter_sink.h
new file mode 100644
index 0000000..1498142
--- /dev/null
+++ b/ext/spdlog/sinks/dup_filter_sink.h
@@ -0,0 +1,92 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#include "dist_sink.h"
+#include <spdlog/details/log_msg.h>
+#include <spdlog/details/null_mutex.h>
+
+#include <chrono>
+#include <cstdio>
+#include <mutex>
+#include <string>
+
+// Duplicate message removal sink.
+// Skip the message if previous one is identical and less than "max_skip_duration" have passed
+//
+// Example:
+//
+//     #include <spdlog/sinks/dup_filter_sink.h>
+//
+//     int main() {
+//         auto dup_filter = std::make_shared<dup_filter_sink_st>(std::chrono::seconds(5),
+//         level::info); dup_filter->add_sink(std::make_shared<stdout_color_sink_mt>());
+//         spdlog::logger l("logger", dup_filter);
+//         l.info("Hello");
+//         l.info("Hello");
+//         l.info("Hello");
+//         l.info("Different Hello");
+//     }
+//
+// Will produce:
+//       [2019-06-25 17:50:56.511] [logger] [info] Hello
+//       [2019-06-25 17:50:56.512] [logger] [info] Skipped 3 duplicate messages..
+//       [2019-06-25 17:50:56.512] [logger] [info] Different Hello
+
+namespace spdlog {
+namespace sinks {
+template <typename Mutex>
+class dup_filter_sink : public dist_sink<Mutex> {
+public:
+    template <class Rep, class Period>
+    explicit dup_filter_sink(std::chrono::duration<Rep, Period> max_skip_duration,
+                             level::level_enum notification_level = level::info)
+        : max_skip_duration_{max_skip_duration},
+          log_level_{notification_level} {}
+
+protected:
+    std::chrono::microseconds max_skip_duration_;
+    log_clock::time_point last_msg_time_;
+    std::string last_msg_payload_;
+    size_t skip_counter_ = 0;
+    level::level_enum log_level_;
+
+    void sink_it_(const details::log_msg &msg) override {
+        bool filtered = filter_(msg);
+        if (!filtered) {
+            skip_counter_ += 1;
+            return;
+        }
+
+        // log the "skipped.." message
+        if (skip_counter_ > 0) {
+            char buf[64];
+            auto msg_size = ::snprintf(buf, sizeof(buf), "Skipped %u duplicate messages..",
+                                       static_cast<unsigned>(skip_counter_));
+            if (msg_size > 0 && static_cast<size_t>(msg_size) < sizeof(buf)) {
+                details::log_msg skipped_msg{msg.source, msg.logger_name, log_level_,
+                                             string_view_t{buf, static_cast<size_t>(msg_size)}};
+                dist_sink<Mutex>::sink_it_(skipped_msg);
+            }
+        }
+
+        // log current message
+        dist_sink<Mutex>::sink_it_(msg);
+        last_msg_time_ = msg.time;
+        skip_counter_ = 0;
+        last_msg_payload_.assign(msg.payload.data(), msg.payload.data() + msg.payload.size());
+    }
+
+    // return whether the log msg should be displayed (true) or skipped (false)
+    bool filter_(const details::log_msg &msg) {
+        auto filter_duration = msg.time - last_msg_time_;
+        return (filter_duration > max_skip_duration_) || (msg.payload != last_msg_payload_);
+    }
+};
+
+using dup_filter_sink_mt = dup_filter_sink<std::mutex>;
+using dup_filter_sink_st = dup_filter_sink<details::null_mutex>;
+
+}  // namespace sinks
+}  // namespace spdlog
diff --git a/ext/spdlog/sinks/hourly_file_sink.h b/ext/spdlog/sinks/hourly_file_sink.h
new file mode 100644
index 0000000..3e61872
--- /dev/null
+++ b/ext/spdlog/sinks/hourly_file_sink.h
@@ -0,0 +1,193 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#include <spdlog/common.h>
+#include <spdlog/details/circular_q.h>
+#include <spdlog/details/file_helper.h>
+#include <spdlog/details/null_mutex.h>
+#include <spdlog/details/os.h>
+#include <spdlog/details/synchronous_factory.h>
+#include <spdlog/fmt/fmt.h>
+#include <spdlog/sinks/base_sink.h>
+
+#include <chrono>
+#include <cstdio>
+#include <ctime>
+#include <mutex>
+#include <string>
+
+namespace spdlog {
+namespace sinks {
+
+/*
+ * Generator of Hourly log file names in format basename.YYYY-MM-DD-HH.ext
+ */
+struct hourly_filename_calculator {
+    // Create filename for the form basename.YYYY-MM-DD-H
+    static filename_t calc_filename(const filename_t &filename, const tm &now_tm) {
+        filename_t basename, ext;
+        std::tie(basename, ext) = details::file_helper::split_by_extension(filename);
+        return fmt_lib::format(SPDLOG_FILENAME_T("{}_{:04d}-{:02d}-{:02d}_{:02d}{}"), basename,
+                               now_tm.tm_year + 1900, now_tm.tm_mon + 1, now_tm.tm_mday,
+                               now_tm.tm_hour, ext);
+    }
+};
+
+/*
+ * Rotating file sink based on time.
+ * If truncate != false , the created file will be truncated.
+ * If max_files > 0, retain only the last max_files and delete previous.
+ * Note that old log files from previous executions will not be deleted by this class,
+ * rotation and deletion is only applied while the program is running.
+ */
+template <typename Mutex, typename FileNameCalc = hourly_filename_calculator>
+class hourly_file_sink final : public base_sink<Mutex> {
+public:
+    // create hourly file sink which rotates on given time
+    hourly_file_sink(filename_t base_filename,
+                     bool truncate = false,
+                     uint16_t max_files = 0,
+                     const file_event_handlers &event_handlers = {})
+        : base_filename_(std::move(base_filename)),
+          file_helper_{event_handlers},
+          truncate_(truncate),
+          max_files_(max_files),
+          filenames_q_() {
+        auto now = log_clock::now();
+        auto filename = FileNameCalc::calc_filename(base_filename_, now_tm(now));
+        file_helper_.open(filename, truncate_);
+        remove_init_file_ = file_helper_.size() == 0;
+        rotation_tp_ = next_rotation_tp_();
+
+        if (max_files_ > 0) {
+            init_filenames_q_();
+        }
+    }
+
+    filename_t filename() {
+        std::lock_guard<Mutex> lock(base_sink<Mutex>::mutex_);
+        return file_helper_.filename();
+    }
+
+protected:
+    void sink_it_(const details::log_msg &msg) override {
+        auto time = msg.time;
+        bool should_rotate = time >= rotation_tp_;
+        if (should_rotate) {
+            if (remove_init_file_) {
+                file_helper_.close();
+                details::os::remove(file_helper_.filename());
+            }
+            auto filename = FileNameCalc::calc_filename(base_filename_, now_tm(time));
+            file_helper_.open(filename, truncate_);
+            rotation_tp_ = next_rotation_tp_();
+        }
+        remove_init_file_ = false;
+        memory_buf_t formatted;
+        base_sink<Mutex>::formatter_->format(msg, formatted);
+        file_helper_.write(formatted);
+
+        // Do the cleaning only at the end because it might throw on failure.
+        if (should_rotate && max_files_ > 0) {
+            delete_old_();
+        }
+    }
+
+    void flush_() override { file_helper_.flush(); }
+
+private:
+    void init_filenames_q_() {
+        using details::os::path_exists;
+
+        filenames_q_ = details::circular_q<filename_t>(static_cast<size_t>(max_files_));
+        std::vector<filename_t> filenames;
+        auto now = log_clock::now();
+        while (filenames.size() < max_files_) {
+            auto filename = FileNameCalc::calc_filename(base_filename_, now_tm(now));
+            if (!path_exists(filename)) {
+                break;
+            }
+            filenames.emplace_back(filename);
+            now -= std::chrono::hours(1);
+        }
+        for (auto iter = filenames.rbegin(); iter != filenames.rend(); ++iter) {
+            filenames_q_.push_back(std::move(*iter));
+        }
+    }
+
+    tm now_tm(log_clock::time_point tp) {
+        time_t tnow = log_clock::to_time_t(tp);
+        return spdlog::details::os::localtime(tnow);
+    }
+
+    log_clock::time_point next_rotation_tp_() {
+        auto now = log_clock::now();
+        tm date = now_tm(now);
+        date.tm_min = 0;
+        date.tm_sec = 0;
+        auto rotation_time = log_clock::from_time_t(std::mktime(&date));
+        if (rotation_time > now) {
+            return rotation_time;
+        }
+        return {rotation_time + std::chrono::hours(1)};
+    }
+
+    // Delete the file N rotations ago.
+    // Throw spdlog_ex on failure to delete the old file.
+    void delete_old_() {
+        using details::os::filename_to_str;
+        using details::os::remove_if_exists;
+
+        filename_t current_file = file_helper_.filename();
+        if (filenames_q_.full()) {
+            auto old_filename = std::move(filenames_q_.front());
+            filenames_q_.pop_front();
+            bool ok = remove_if_exists(old_filename) == 0;
+            if (!ok) {
+                filenames_q_.push_back(std::move(current_file));
+                SPDLOG_THROW(spdlog_ex(
+                    "Failed removing hourly file " + filename_to_str(old_filename), errno));
+            }
+        }
+        filenames_q_.push_back(std::move(current_file));
+    }
+
+    filename_t base_filename_;
+    log_clock::time_point rotation_tp_;
+    details::file_helper file_helper_;
+    bool truncate_;
+    uint16_t max_files_;
+    details::circular_q<filename_t> filenames_q_;
+    bool remove_init_file_;
+};
+
+using hourly_file_sink_mt = hourly_file_sink<std::mutex>;
+using hourly_file_sink_st = hourly_file_sink<details::null_mutex>;
+
+}  // namespace sinks
+
+//
+// factory functions
+//
+template <typename Factory = spdlog::synchronous_factory>
+inline std::shared_ptr<logger> hourly_logger_mt(const std::string &logger_name,
+                                                const filename_t &filename,
+                                                bool truncate = false,
+                                                uint16_t max_files = 0,
+                                                const file_event_handlers &event_handlers = {}) {
+    return Factory::template create<sinks::hourly_file_sink_mt>(logger_name, filename, truncate,
+                                                                max_files, event_handlers);
+}
+
+template <typename Factory = spdlog::synchronous_factory>
+inline std::shared_ptr<logger> hourly_logger_st(const std::string &logger_name,
+                                                const filename_t &filename,
+                                                bool truncate = false,
+                                                uint16_t max_files = 0,
+                                                const file_event_handlers &event_handlers = {}) {
+    return Factory::template create<sinks::hourly_file_sink_st>(logger_name, filename, truncate,
+                                                                max_files, event_handlers);
+}
+}  // namespace spdlog
diff --git a/ext/spdlog/sinks/kafka_sink.h b/ext/spdlog/sinks/kafka_sink.h
new file mode 100644
index 0000000..91e9878
--- /dev/null
+++ b/ext/spdlog/sinks/kafka_sink.h
@@ -0,0 +1,119 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+//
+// Custom sink for kafka
+// Building and using requires librdkafka library.
+// For building librdkafka library check the url below
+// https://github.com/confluentinc/librdkafka
+//
+
+#include "spdlog/async.h"
+#include "spdlog/details/log_msg.h"
+#include "spdlog/details/null_mutex.h"
+#include "spdlog/details/synchronous_factory.h"
+#include "spdlog/sinks/base_sink.h"
+#include <mutex>
+#include <spdlog/common.h>
+
+// kafka header
+#include <librdkafka/rdkafkacpp.h>
+
+namespace spdlog {
+namespace sinks {
+
+struct kafka_sink_config {
+    std::string server_addr;
+    std::string produce_topic;
+    int32_t flush_timeout_ms = 1000;
+
+    kafka_sink_config(std::string addr, std::string topic, int flush_timeout_ms = 1000)
+        : server_addr{std::move(addr)},
+          produce_topic{std::move(topic)},
+          flush_timeout_ms(flush_timeout_ms) {}
+};
+
+template <typename Mutex>
+class kafka_sink : public base_sink<Mutex> {
+public:
+    kafka_sink(kafka_sink_config config)
+        : config_{std::move(config)} {
+        try {
+            std::string errstr;
+            conf_.reset(RdKafka::Conf::create(RdKafka::Conf::CONF_GLOBAL));
+            RdKafka::Conf::ConfResult confRes =
+                conf_->set("bootstrap.servers", config_.server_addr, errstr);
+            if (confRes != RdKafka::Conf::CONF_OK) {
+                throw_spdlog_ex(
+                    fmt_lib::format("conf set bootstrap.servers failed err:{}", errstr));
+            }
+
+            tconf_.reset(RdKafka::Conf::create(RdKafka::Conf::CONF_TOPIC));
+            if (tconf_ == nullptr) {
+                throw_spdlog_ex(fmt_lib::format("create topic config failed"));
+            }
+
+            producer_.reset(RdKafka::Producer::create(conf_.get(), errstr));
+            if (producer_ == nullptr) {
+                throw_spdlog_ex(fmt_lib::format("create producer failed err:{}", errstr));
+            }
+            topic_.reset(RdKafka::Topic::create(producer_.get(), config_.produce_topic,
+                                                tconf_.get(), errstr));
+            if (topic_ == nullptr) {
+                throw_spdlog_ex(fmt_lib::format("create topic failed err:{}", errstr));
+            }
+        } catch (const std::exception &e) {
+            throw_spdlog_ex(fmt_lib::format("error create kafka instance: {}", e.what()));
+        }
+    }
+
+    ~kafka_sink() { producer_->flush(config_.flush_timeout_ms); }
+
+protected:
+    void sink_it_(const details::log_msg &msg) override {
+        producer_->produce(topic_.get(), 0, RdKafka::Producer::RK_MSG_COPY,
+                           (void *)msg.payload.data(), msg.payload.size(), NULL, NULL);
+    }
+
+    void flush_() override { producer_->flush(config_.flush_timeout_ms); }
+
+private:
+    kafka_sink_config config_;
+    std::unique_ptr<RdKafka::Producer> producer_ = nullptr;
+    std::unique_ptr<RdKafka::Conf> conf_ = nullptr;
+    std::unique_ptr<RdKafka::Conf> tconf_ = nullptr;
+    std::unique_ptr<RdKafka::Topic> topic_ = nullptr;
+};
+
+using kafka_sink_mt = kafka_sink<std::mutex>;
+using kafka_sink_st = kafka_sink<spdlog::details::null_mutex>;
+
+}  // namespace sinks
+
+template <typename Factory = spdlog::synchronous_factory>
+inline std::shared_ptr<logger> kafka_logger_mt(const std::string &logger_name,
+                                               spdlog::sinks::kafka_sink_config config) {
+    return Factory::template create<sinks::kafka_sink_mt>(logger_name, config);
+}
+
+template <typename Factory = spdlog::synchronous_factory>
+inline std::shared_ptr<logger> kafka_logger_st(const std::string &logger_name,
+                                               spdlog::sinks::kafka_sink_config config) {
+    return Factory::template create<sinks::kafka_sink_st>(logger_name, config);
+}
+
+template <typename Factory = spdlog::async_factory>
+inline std::shared_ptr<spdlog::logger> kafka_logger_async_mt(
+    std::string logger_name, spdlog::sinks::kafka_sink_config config) {
+    return Factory::template create<sinks::kafka_sink_mt>(logger_name, config);
+}
+
+template <typename Factory = spdlog::async_factory>
+inline std::shared_ptr<spdlog::logger> kafka_logger_async_st(
+    std::string logger_name, spdlog::sinks::kafka_sink_config config) {
+    return Factory::template create<sinks::kafka_sink_st>(logger_name, config);
+}
+
+}  // namespace spdlog
diff --git a/ext/spdlog/sinks/mongo_sink.h b/ext/spdlog/sinks/mongo_sink.h
new file mode 100644
index 0000000..c5b38ab
--- /dev/null
+++ b/ext/spdlog/sinks/mongo_sink.h
@@ -0,0 +1,108 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+//
+// Custom sink for mongodb
+// Building and using requires mongocxx library.
+// For building mongocxx library check the url below
+// http://mongocxx.org/mongocxx-v3/installation/
+//
+
+#include "spdlog/common.h"
+#include "spdlog/details/log_msg.h"
+#include "spdlog/sinks/base_sink.h"
+#include <spdlog/details/synchronous_factory.h>
+
+#include <bsoncxx/builder/stream/document.hpp>
+#include <bsoncxx/types.hpp>
+#include <bsoncxx/view_or_value.hpp>
+
+#include <mongocxx/client.hpp>
+#include <mongocxx/instance.hpp>
+#include <mongocxx/uri.hpp>
+
+namespace spdlog {
+namespace sinks {
+template <typename Mutex>
+class mongo_sink : public base_sink<Mutex> {
+public:
+    mongo_sink(const std::string &db_name,
+               const std::string &collection_name,
+               const std::string &uri = "mongodb://localhost:27017") try
+        : mongo_sink(std::make_shared<mongocxx::instance>(), db_name, collection_name, uri) {
+    } catch (const std::exception &e) {
+        throw_spdlog_ex(fmt_lib::format("Error opening database: {}", e.what()));
+    }
+
+    mongo_sink(std::shared_ptr<mongocxx::instance> instance,
+               const std::string &db_name,
+               const std::string &collection_name,
+               const std::string &uri = "mongodb://localhost:27017")
+        : instance_(std::move(instance)),
+          db_name_(db_name),
+          coll_name_(collection_name) {
+        try {
+            client_ = spdlog::details::make_unique<mongocxx::client>(mongocxx::uri{uri});
+        } catch (const std::exception &e) {
+            throw_spdlog_ex(fmt_lib::format("Error opening database: {}", e.what()));
+        }
+    }
+
+    ~mongo_sink() { flush_(); }
+
+protected:
+    void sink_it_(const details::log_msg &msg) override {
+        using bsoncxx::builder::stream::document;
+        using bsoncxx::builder::stream::finalize;
+
+        if (client_ != nullptr) {
+            auto doc = document{} << "timestamp" << bsoncxx::types::b_date(msg.time) << "level"
+                                  << level::to_string_view(msg.level).data() << "level_num"
+                                  << msg.level << "message"
+                                  << std::string(msg.payload.begin(), msg.payload.end())
+                                  << "logger_name"
+                                  << std::string(msg.logger_name.begin(), msg.logger_name.end())
+                                  << "thread_id" << static_cast<int>(msg.thread_id) << finalize;
+            client_->database(db_name_).collection(coll_name_).insert_one(doc.view());
+        }
+    }
+
+    void flush_() override {}
+
+private:
+    std::shared_ptr<mongocxx::instance> instance_;
+    std::string db_name_;
+    std::string coll_name_;
+    std::unique_ptr<mongocxx::client> client_ = nullptr;
+};
+
+#include "spdlog/details/null_mutex.h"
+#include <mutex>
+using mongo_sink_mt = mongo_sink<std::mutex>;
+using mongo_sink_st = mongo_sink<spdlog::details::null_mutex>;
+
+}  // namespace sinks
+
+template <typename Factory = spdlog::synchronous_factory>
+inline std::shared_ptr<logger> mongo_logger_mt(
+    const std::string &logger_name,
+    const std::string &db_name,
+    const std::string &collection_name,
+    const std::string &uri = "mongodb://localhost:27017") {
+    return Factory::template create<sinks::mongo_sink_mt>(logger_name, db_name, collection_name,
+                                                          uri);
+}
+
+template <typename Factory = spdlog::synchronous_factory>
+inline std::shared_ptr<logger> mongo_logger_st(
+    const std::string &logger_name,
+    const std::string &db_name,
+    const std::string &collection_name,
+    const std::string &uri = "mongodb://localhost:27017") {
+    return Factory::template create<sinks::mongo_sink_st>(logger_name, db_name, collection_name,
+                                                          uri);
+}
+
+}  // namespace spdlog
diff --git a/ext/spdlog/sinks/msvc_sink.h b/ext/spdlog/sinks/msvc_sink.h
new file mode 100644
index 0000000..c28d6eb
--- /dev/null
+++ b/ext/spdlog/sinks/msvc_sink.h
@@ -0,0 +1,68 @@
+// Copyright(c) 2016 Alexander Dalshov & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#if defined(_WIN32)
+
+    #include <spdlog/details/null_mutex.h>
+    #if defined(SPDLOG_WCHAR_TO_UTF8_SUPPORT)
+        #include <spdlog/details/os.h>
+    #endif
+    #include <spdlog/sinks/base_sink.h>
+
+    #include <mutex>
+    #include <string>
+
+    // Avoid including windows.h (https://stackoverflow.com/a/30741042)
+    #if defined(SPDLOG_WCHAR_TO_UTF8_SUPPORT)
+extern "C" __declspec(dllimport) void __stdcall OutputDebugStringW(const wchar_t *lpOutputString);
+    #else
+extern "C" __declspec(dllimport) void __stdcall OutputDebugStringA(const char *lpOutputString);
+    #endif
+extern "C" __declspec(dllimport) int __stdcall IsDebuggerPresent();
+
+namespace spdlog {
+namespace sinks {
+/*
+ * MSVC sink (logging using OutputDebugStringA)
+ */
+template <typename Mutex>
+class msvc_sink : public base_sink<Mutex> {
+public:
+    msvc_sink() = default;
+    msvc_sink(bool check_debugger_present)
+        : check_debugger_present_{check_debugger_present} {}
+
+protected:
+    void sink_it_(const details::log_msg &msg) override {
+        if (check_debugger_present_ && !IsDebuggerPresent()) {
+            return;
+        }
+        memory_buf_t formatted;
+        base_sink<Mutex>::formatter_->format(msg, formatted);
+        formatted.push_back('\0');  // add a null terminator for OutputDebugString
+    #if defined(SPDLOG_WCHAR_TO_UTF8_SUPPORT)
+        wmemory_buf_t wformatted;
+        details::os::utf8_to_wstrbuf(string_view_t(formatted.data(), formatted.size()), wformatted);
+        OutputDebugStringW(wformatted.data());
+    #else
+        OutputDebugStringA(formatted.data());
+    #endif
+    }
+
+    void flush_() override {}
+
+    bool check_debugger_present_ = true;
+};
+
+using msvc_sink_mt = msvc_sink<std::mutex>;
+using msvc_sink_st = msvc_sink<details::null_mutex>;
+
+using windebug_sink_mt = msvc_sink_mt;
+using windebug_sink_st = msvc_sink_st;
+
+}  // namespace sinks
+}  // namespace spdlog
+
+#endif
diff --git a/ext/spdlog/sinks/null_sink.h b/ext/spdlog/sinks/null_sink.h
new file mode 100644
index 0000000..74530b5
--- /dev/null
+++ b/ext/spdlog/sinks/null_sink.h
@@ -0,0 +1,41 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#include <spdlog/details/null_mutex.h>
+#include <spdlog/details/synchronous_factory.h>
+#include <spdlog/sinks/base_sink.h>
+
+#include <mutex>
+
+namespace spdlog {
+namespace sinks {
+
+template <typename Mutex>
+class null_sink final : public base_sink<Mutex> {
+protected:
+    void sink_it_(const details::log_msg &) override {}
+    void flush_() override {}
+};
+
+using null_sink_mt = null_sink<details::null_mutex>;
+using null_sink_st = null_sink<details::null_mutex>;
+
+}  // namespace sinks
+
+template <typename Factory = spdlog::synchronous_factory>
+inline std::shared_ptr<logger> null_logger_mt(const std::string &logger_name) {
+    auto null_logger = Factory::template create<sinks::null_sink_mt>(logger_name);
+    null_logger->set_level(level::off);
+    return null_logger;
+}
+
+template <typename Factory = spdlog::synchronous_factory>
+inline std::shared_ptr<logger> null_logger_st(const std::string &logger_name) {
+    auto null_logger = Factory::template create<sinks::null_sink_st>(logger_name);
+    null_logger->set_level(level::off);
+    return null_logger;
+}
+
+}  // namespace spdlog
diff --git a/ext/spdlog/sinks/ostream_sink.h b/ext/spdlog/sinks/ostream_sink.h
new file mode 100644
index 0000000..6af9dd0
--- /dev/null
+++ b/ext/spdlog/sinks/ostream_sink.h
@@ -0,0 +1,43 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#include <spdlog/details/null_mutex.h>
+#include <spdlog/sinks/base_sink.h>
+
+#include <mutex>
+#include <ostream>
+
+namespace spdlog {
+namespace sinks {
+template <typename Mutex>
+class ostream_sink final : public base_sink<Mutex> {
+public:
+    explicit ostream_sink(std::ostream &os, bool force_flush = false)
+        : ostream_(os),
+          force_flush_(force_flush) {}
+    ostream_sink(const ostream_sink &) = delete;
+    ostream_sink &operator=(const ostream_sink &) = delete;
+
+protected:
+    void sink_it_(const details::log_msg &msg) override {
+        memory_buf_t formatted;
+        base_sink<Mutex>::formatter_->format(msg, formatted);
+        ostream_.write(formatted.data(), static_cast<std::streamsize>(formatted.size()));
+        if (force_flush_) {
+            ostream_.flush();
+        }
+    }
+
+    void flush_() override { ostream_.flush(); }
+
+    std::ostream &ostream_;
+    bool force_flush_;
+};
+
+using ostream_sink_mt = ostream_sink<std::mutex>;
+using ostream_sink_st = ostream_sink<details::null_mutex>;
+
+}  // namespace sinks
+}  // namespace spdlog
diff --git a/ext/spdlog/sinks/qt_sinks.h b/ext/spdlog/sinks/qt_sinks.h
new file mode 100644
index 0000000..d319e84
--- /dev/null
+++ b/ext/spdlog/sinks/qt_sinks.h
@@ -0,0 +1,304 @@
+// Copyright(c) 2015-present, Gabi Melman, mguludag and spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+//
+// Custom sink for QPlainTextEdit or QTextEdit and its children (QTextBrowser...
+// etc) Building and using requires Qt library.
+//
+// Warning: the qt_sink won't be notified if the target widget is destroyed.
+// If the widget's lifetime can be shorter than the logger's one, you should provide some permanent
+// QObject, and then use a standard signal/slot.
+//
+
+#include "spdlog/common.h"
+#include "spdlog/details/log_msg.h"
+#include "spdlog/details/synchronous_factory.h"
+#include "spdlog/sinks/base_sink.h"
+#include <array>
+
+#include <QPlainTextEdit>
+#include <QTextEdit>
+
+//
+// qt_sink class
+//
+namespace spdlog {
+namespace sinks {
+template <typename Mutex>
+class qt_sink : public base_sink<Mutex> {
+public:
+    qt_sink(QObject *qt_object, std::string meta_method)
+        : qt_object_(qt_object),
+          meta_method_(std::move(meta_method)) {
+        if (!qt_object_) {
+            throw_spdlog_ex("qt_sink: qt_object is null");
+        }
+    }
+
+    ~qt_sink() { flush_(); }
+
+protected:
+    void sink_it_(const details::log_msg &msg) override {
+        memory_buf_t formatted;
+        base_sink<Mutex>::formatter_->format(msg, formatted);
+        const string_view_t str = string_view_t(formatted.data(), formatted.size());
+        QMetaObject::invokeMethod(
+            qt_object_, meta_method_.c_str(), Qt::AutoConnection,
+            Q_ARG(QString, QString::fromUtf8(str.data(), static_cast<int>(str.size())).trimmed()));
+    }
+
+    void flush_() override {}
+
+private:
+    QObject *qt_object_ = nullptr;
+    std::string meta_method_;
+};
+
+// Qt color sink to QTextEdit.
+// Color location is determined by the sink log pattern like in the rest of spdlog sinks.
+// Colors can be modified if needed using sink->set_color(level, qtTextCharFormat).
+// max_lines is the maximum number of lines that the sink will hold before removing the oldest
+// lines. By default, only ascii (latin1) is supported by this sink. Set is_utf8 to true if utf8
+// support is needed.
+template <typename Mutex>
+class qt_color_sink : public base_sink<Mutex> {
+public:
+    qt_color_sink(QTextEdit *qt_text_edit,
+                  int max_lines,
+                  bool dark_colors = false,
+                  bool is_utf8 = false)
+        : qt_text_edit_(qt_text_edit),
+          max_lines_(max_lines),
+          is_utf8_(is_utf8) {
+        if (!qt_text_edit_) {
+            throw_spdlog_ex("qt_color_text_sink: text_edit is null");
+        }
+
+        default_color_ = qt_text_edit_->currentCharFormat();
+        // set colors
+        QTextCharFormat format;
+        // trace
+        format.setForeground(dark_colors ? Qt::darkGray : Qt::gray);
+        colors_.at(level::trace) = format;
+        // debug
+        format.setForeground(dark_colors ? Qt::darkCyan : Qt::cyan);
+        colors_.at(level::debug) = format;
+        // info
+        format.setForeground(dark_colors ? Qt::darkGreen : Qt::green);
+        colors_.at(level::info) = format;
+        // warn
+        format.setForeground(dark_colors ? Qt::darkYellow : Qt::yellow);
+        colors_.at(level::warn) = format;
+        // err
+        format.setForeground(Qt::red);
+        colors_.at(level::err) = format;
+        // critical
+        format.setForeground(Qt::white);
+        format.setBackground(Qt::red);
+        colors_.at(level::critical) = format;
+    }
+
+    ~qt_color_sink() { flush_(); }
+
+    void set_default_color(QTextCharFormat format) {
+        // std::lock_guard<Mutex> lock(base_sink<Mutex>::mutex_);
+        default_color_ = format;
+    }
+
+    void set_level_color(level::level_enum color_level, QTextCharFormat format) {
+        // std::lock_guard<Mutex> lock(base_sink<Mutex>::mutex_);
+        colors_.at(static_cast<size_t>(color_level)) = format;
+    }
+
+    QTextCharFormat &get_level_color(level::level_enum color_level) {
+        std::lock_guard<Mutex> lock(base_sink<Mutex>::mutex_);
+        return colors_.at(static_cast<size_t>(color_level));
+    }
+
+    QTextCharFormat &get_default_color() {
+        std::lock_guard<Mutex> lock(base_sink<Mutex>::mutex_);
+        return default_color_;
+    }
+
+protected:
+    struct invoke_params {
+        invoke_params(int max_lines,
+                      QTextEdit *q_text_edit,
+                      QString payload,
+                      QTextCharFormat default_color,
+                      QTextCharFormat level_color,
+                      int color_range_start,
+                      int color_range_end)
+            : max_lines(max_lines),
+              q_text_edit(q_text_edit),
+              payload(std::move(payload)),
+              default_color(default_color),
+              level_color(level_color),
+              color_range_start(color_range_start),
+              color_range_end(color_range_end) {}
+        int max_lines;
+        QTextEdit *q_text_edit;
+        QString payload;
+        QTextCharFormat default_color;
+        QTextCharFormat level_color;
+        int color_range_start;
+        int color_range_end;
+    };
+
+    void sink_it_(const details::log_msg &msg) override {
+        memory_buf_t formatted;
+        base_sink<Mutex>::formatter_->format(msg, formatted);
+
+        const string_view_t str = string_view_t(formatted.data(), formatted.size());
+        // apply the color to the color range in the formatted message.
+        QString payload;
+        int color_range_start = static_cast<int>(msg.color_range_start);
+        int color_range_end = static_cast<int>(msg.color_range_end);
+        if (is_utf8_) {
+            payload = QString::fromUtf8(str.data(), static_cast<int>(str.size()));
+            // convert color ranges from byte index to character index.
+            if (msg.color_range_start < msg.color_range_end) {
+                color_range_start = QString::fromUtf8(str.data(), msg.color_range_start).size();
+                color_range_end = QString::fromUtf8(str.data(), msg.color_range_end).size();
+            }
+        } else {
+            payload = QString::fromLatin1(str.data(), static_cast<int>(str.size()));
+        }
+
+        invoke_params params{max_lines_,             // max lines
+                             qt_text_edit_,          // text edit to append to
+                             std::move(payload),     // text to append
+                             default_color_,         // default color
+                             colors_.at(msg.level),  // color to apply
+                             color_range_start,      // color range start
+                             color_range_end};       // color range end
+
+        QMetaObject::invokeMethod(
+            qt_text_edit_, [params]() { invoke_method_(params); }, Qt::AutoConnection);
+    }
+
+    void flush_() override {}
+
+    // Add colored text to the text edit widget. This method is invoked in the GUI thread.
+    // It is a static method to ensure that it is handled correctly even if the sink is destroyed
+    // prematurely before it is invoked.
+
+    static void invoke_method_(invoke_params params) {
+        auto *document = params.q_text_edit->document();
+        QTextCursor cursor(document);
+
+        // remove first blocks if number of blocks exceeds max_lines
+        while (document->blockCount() > params.max_lines) {
+            cursor.select(QTextCursor::BlockUnderCursor);
+            cursor.removeSelectedText();
+            cursor.deleteChar();  // delete the newline after the block
+        }
+
+        cursor.movePosition(QTextCursor::End);
+        cursor.setCharFormat(params.default_color);
+
+        // if color range not specified or not not valid, just append the text with default color
+        if (params.color_range_end <= params.color_range_start) {
+            cursor.insertText(params.payload);
+            return;
+        }
+
+        // insert the text before the color range
+        cursor.insertText(params.payload.left(params.color_range_start));
+
+        // insert the colorized text
+        cursor.setCharFormat(params.level_color);
+        cursor.insertText(params.payload.mid(params.color_range_start,
+                                             params.color_range_end - params.color_range_start));
+
+        // insert the text after the color range with default format
+        cursor.setCharFormat(params.default_color);
+        cursor.insertText(params.payload.mid(params.color_range_end));
+    }
+
+    QTextEdit *qt_text_edit_;
+    int max_lines_;
+    bool is_utf8_;
+    QTextCharFormat default_color_;
+    std::array<QTextCharFormat, level::n_levels> colors_;
+};
+
+#include "spdlog/details/null_mutex.h"
+#include <mutex>
+
+using qt_sink_mt = qt_sink<std::mutex>;
+using qt_sink_st = qt_sink<details::null_mutex>;
+using qt_color_sink_mt = qt_color_sink<std::mutex>;
+using qt_color_sink_st = qt_color_sink<details::null_mutex>;
+}  // namespace sinks
+
+//
+// Factory functions
+//
+
+// log to QTextEdit
+template <typename Factory = spdlog::synchronous_factory>
+inline std::shared_ptr<logger> qt_logger_mt(const std::string &logger_name,
+                                            QTextEdit *qt_object,
+                                            const std::string &meta_method = "append") {
+    return Factory::template create<sinks::qt_sink_mt>(logger_name, qt_object, meta_method);
+}
+
+template <typename Factory = spdlog::synchronous_factory>
+inline std::shared_ptr<logger> qt_logger_st(const std::string &logger_name,
+                                            QTextEdit *qt_object,
+                                            const std::string &meta_method = "append") {
+    return Factory::template create<sinks::qt_sink_st>(logger_name, qt_object, meta_method);
+}
+
+// log to QPlainTextEdit
+template <typename Factory = spdlog::synchronous_factory>
+inline std::shared_ptr<logger> qt_logger_mt(const std::string &logger_name,
+                                            QPlainTextEdit *qt_object,
+                                            const std::string &meta_method = "appendPlainText") {
+    return Factory::template create<sinks::qt_sink_mt>(logger_name, qt_object, meta_method);
+}
+
+template <typename Factory = spdlog::synchronous_factory>
+inline std::shared_ptr<logger> qt_logger_st(const std::string &logger_name,
+                                            QPlainTextEdit *qt_object,
+                                            const std::string &meta_method = "appendPlainText") {
+    return Factory::template create<sinks::qt_sink_st>(logger_name, qt_object, meta_method);
+}
+// log to QObject
+template <typename Factory = spdlog::synchronous_factory>
+inline std::shared_ptr<logger> qt_logger_mt(const std::string &logger_name,
+                                            QObject *qt_object,
+                                            const std::string &meta_method) {
+    return Factory::template create<sinks::qt_sink_mt>(logger_name, qt_object, meta_method);
+}
+
+template <typename Factory = spdlog::synchronous_factory>
+inline std::shared_ptr<logger> qt_logger_st(const std::string &logger_name,
+                                            QObject *qt_object,
+                                            const std::string &meta_method) {
+    return Factory::template create<sinks::qt_sink_st>(logger_name, qt_object, meta_method);
+}
+
+// log to QTextEdit with colorized output
+template <typename Factory = spdlog::synchronous_factory>
+inline std::shared_ptr<logger> qt_color_logger_mt(const std::string &logger_name,
+                                                  QTextEdit *qt_text_edit,
+                                                  int max_lines,
+                                                  bool is_utf8 = false) {
+    return Factory::template create<sinks::qt_color_sink_mt>(logger_name, qt_text_edit, max_lines,
+                                                             false, is_utf8);
+}
+
+template <typename Factory = spdlog::synchronous_factory>
+inline std::shared_ptr<logger> qt_color_logger_st(const std::string &logger_name,
+                                                  QTextEdit *qt_text_edit,
+                                                  int max_lines,
+                                                  bool is_utf8 = false) {
+    return Factory::template create<sinks::qt_color_sink_st>(logger_name, qt_text_edit, max_lines,
+                                                             false, is_utf8);
+}
+
+}  // namespace spdlog
diff --git a/ext/spdlog/sinks/ringbuffer_sink.h b/ext/spdlog/sinks/ringbuffer_sink.h
new file mode 100644
index 0000000..6156c6a
--- /dev/null
+++ b/ext/spdlog/sinks/ringbuffer_sink.h
@@ -0,0 +1,67 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#include "spdlog/details/circular_q.h"
+#include "spdlog/details/log_msg_buffer.h"
+#include "spdlog/details/null_mutex.h"
+#include "spdlog/sinks/base_sink.h"
+
+#include <mutex>
+#include <string>
+#include <vector>
+
+namespace spdlog {
+namespace sinks {
+/*
+ * Ring buffer sink
+ */
+template <typename Mutex>
+class ringbuffer_sink final : public base_sink<Mutex> {
+public:
+    explicit ringbuffer_sink(size_t n_items)
+        : q_{n_items} {}
+
+    std::vector<details::log_msg_buffer> last_raw(size_t lim = 0) {
+        std::lock_guard<Mutex> lock(base_sink<Mutex>::mutex_);
+        auto items_available = q_.size();
+        auto n_items = lim > 0 ? (std::min)(lim, items_available) : items_available;
+        std::vector<details::log_msg_buffer> ret;
+        ret.reserve(n_items);
+        for (size_t i = (items_available - n_items); i < items_available; i++) {
+            ret.push_back(q_.at(i));
+        }
+        return ret;
+    }
+
+    std::vector<std::string> last_formatted(size_t lim = 0) {
+        std::lock_guard<Mutex> lock(base_sink<Mutex>::mutex_);
+        auto items_available = q_.size();
+        auto n_items = lim > 0 ? (std::min)(lim, items_available) : items_available;
+        std::vector<std::string> ret;
+        ret.reserve(n_items);
+        for (size_t i = (items_available - n_items); i < items_available; i++) {
+            memory_buf_t formatted;
+            base_sink<Mutex>::formatter_->format(q_.at(i), formatted);
+            ret.push_back(SPDLOG_BUF_TO_STRING(formatted));
+        }
+        return ret;
+    }
+
+protected:
+    void sink_it_(const details::log_msg &msg) override {
+        q_.push_back(details::log_msg_buffer{msg});
+    }
+    void flush_() override {}
+
+private:
+    details::circular_q<details::log_msg_buffer> q_;
+};
+
+using ringbuffer_sink_mt = ringbuffer_sink<std::mutex>;
+using ringbuffer_sink_st = ringbuffer_sink<details::null_mutex>;
+
+}  // namespace sinks
+
+}  // namespace spdlog
diff --git a/ext/spdlog/sinks/rotating_file_sink-inl.h b/ext/spdlog/sinks/rotating_file_sink-inl.h
new file mode 100644
index 0000000..420bafb
--- /dev/null
+++ b/ext/spdlog/sinks/rotating_file_sink-inl.h
@@ -0,0 +1,150 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#ifndef SPDLOG_HEADER_ONLY
+    #include <spdlog/sinks/rotating_file_sink.h>
+#endif
+
+#include <spdlog/common.h>
+
+#include <spdlog/details/file_helper.h>
+#include <spdlog/details/null_mutex.h>
+#include <spdlog/fmt/fmt.h>
+
+#include <cerrno>
+#include <chrono>
+#include <ctime>
+#include <mutex>
+#include <string>
+#include <tuple>
+
+namespace spdlog {
+namespace sinks {
+
+template <typename Mutex>
+SPDLOG_INLINE rotating_file_sink<Mutex>::rotating_file_sink(
+    filename_t base_filename,
+    std::size_t max_size,
+    std::size_t max_files,
+    bool rotate_on_open,
+    const file_event_handlers &event_handlers)
+    : base_filename_(std::move(base_filename)),
+      max_size_(max_size),
+      max_files_(max_files),
+      file_helper_{event_handlers} {
+    if (max_size == 0) {
+        throw_spdlog_ex("rotating sink constructor: max_size arg cannot be zero");
+    }
+
+    if (max_files > 200000) {
+        throw_spdlog_ex("rotating sink constructor: max_files arg cannot exceed 200000");
+    }
+    file_helper_.open(calc_filename(base_filename_, 0));
+    current_size_ = file_helper_.size();  // expensive. called only once
+    if (rotate_on_open && current_size_ > 0) {
+        rotate_();
+        current_size_ = 0;
+    }
+}
+
+// calc filename according to index and file extension if exists.
+// e.g. calc_filename("logs/mylog.txt, 3) => "logs/mylog.3.txt".
+template <typename Mutex>
+SPDLOG_INLINE filename_t rotating_file_sink<Mutex>::calc_filename(const filename_t &filename,
+                                                                  std::size_t index) {
+    if (index == 0u) {
+        return filename;
+    }
+
+    filename_t basename, ext;
+    std::tie(basename, ext) = details::file_helper::split_by_extension(filename);
+    return fmt_lib::format(SPDLOG_FMT_STRING(SPDLOG_FILENAME_T("{}.{}{}")), basename, index, ext);
+}
+
+template <typename Mutex>
+SPDLOG_INLINE filename_t rotating_file_sink<Mutex>::filename() {
+    std::lock_guard<Mutex> lock(base_sink<Mutex>::mutex_);
+    return file_helper_.filename();
+}
+
+template <typename Mutex>
+SPDLOG_INLINE void rotating_file_sink<Mutex>::rotate_now() {
+    std::lock_guard<Mutex> lock(base_sink<Mutex>::mutex_);
+    rotate_();
+}
+
+template <typename Mutex>
+SPDLOG_INLINE void rotating_file_sink<Mutex>::sink_it_(const details::log_msg &msg) {
+    memory_buf_t formatted;
+    base_sink<Mutex>::formatter_->format(msg, formatted);
+    auto new_size = current_size_ + formatted.size();
+
+    // rotate if the new estimated file size exceeds max size.
+    // rotate only if the real size > 0 to better deal with full disk (see issue #2261).
+    // we only check the real size when new_size > max_size_ because it is relatively expensive.
+    if (new_size > max_size_) {
+        file_helper_.flush();
+        if (file_helper_.size() > 0) {
+            rotate_();
+            new_size = formatted.size();
+        }
+    }
+    file_helper_.write(formatted);
+    current_size_ = new_size;
+}
+
+template <typename Mutex>
+SPDLOG_INLINE void rotating_file_sink<Mutex>::flush_() {
+    file_helper_.flush();
+}
+
+// Rotate files:
+// log.txt -> log.1.txt
+// log.1.txt -> log.2.txt
+// log.2.txt -> log.3.txt
+// log.3.txt -> delete
+template <typename Mutex>
+SPDLOG_INLINE void rotating_file_sink<Mutex>::rotate_() {
+    using details::os::filename_to_str;
+    using details::os::path_exists;
+
+    file_helper_.close();
+    for (auto i = max_files_; i > 0; --i) {
+        filename_t src = calc_filename(base_filename_, i - 1);
+        if (!path_exists(src)) {
+            continue;
+        }
+        filename_t target = calc_filename(base_filename_, i);
+
+        if (!rename_file_(src, target)) {
+            // if failed try again after a small delay.
+            // this is a workaround to a windows issue, where very high rotation
+            // rates can cause the rename to fail with permission denied (because of antivirus?).
+            details::os::sleep_for_millis(100);
+            if (!rename_file_(src, target)) {
+                file_helper_.reopen(
+                    true);  // truncate the log file anyway to prevent it to grow beyond its limit!
+                current_size_ = 0;
+                throw_spdlog_ex("rotating_file_sink: failed renaming " + filename_to_str(src) +
+                                    " to " + filename_to_str(target),
+                                errno);
+            }
+        }
+    }
+    file_helper_.reopen(true);
+}
+
+// delete the target if exists, and rename the src file  to target
+// return true on success, false otherwise.
+template <typename Mutex>
+SPDLOG_INLINE bool rotating_file_sink<Mutex>::rename_file_(const filename_t &src_filename,
+                                                           const filename_t &target_filename) {
+    // try to delete the target file in case it already exists.
+    (void)details::os::remove(target_filename);
+    return details::os::rename(src_filename, target_filename) == 0;
+}
+
+}  // namespace sinks
+}  // namespace spdlog
diff --git a/ext/spdlog/sinks/rotating_file_sink.h b/ext/spdlog/sinks/rotating_file_sink.h
new file mode 100644
index 0000000..42bd376
--- /dev/null
+++ b/ext/spdlog/sinks/rotating_file_sink.h
@@ -0,0 +1,90 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#include <spdlog/details/file_helper.h>
+#include <spdlog/details/null_mutex.h>
+#include <spdlog/details/synchronous_factory.h>
+#include <spdlog/sinks/base_sink.h>
+
+#include <chrono>
+#include <mutex>
+#include <string>
+
+namespace spdlog {
+namespace sinks {
+
+//
+// Rotating file sink based on size
+//
+template <typename Mutex>
+class rotating_file_sink final : public base_sink<Mutex> {
+public:
+    rotating_file_sink(filename_t base_filename,
+                       std::size_t max_size,
+                       std::size_t max_files,
+                       bool rotate_on_open = false,
+                       const file_event_handlers &event_handlers = {});
+    static filename_t calc_filename(const filename_t &filename, std::size_t index);
+    filename_t filename();
+    void rotate_now();
+
+protected:
+    void sink_it_(const details::log_msg &msg) override;
+    void flush_() override;
+
+private:
+    // Rotate files:
+    // log.txt -> log.1.txt
+    // log.1.txt -> log.2.txt
+    // log.2.txt -> log.3.txt
+    // log.3.txt -> delete
+    void rotate_();
+
+    // delete the target if exists, and rename the src file  to target
+    // return true on success, false otherwise.
+    bool rename_file_(const filename_t &src_filename, const filename_t &target_filename);
+
+    filename_t base_filename_;
+    std::size_t max_size_;
+    std::size_t max_files_;
+    std::size_t current_size_;
+    details::file_helper file_helper_;
+};
+
+using rotating_file_sink_mt = rotating_file_sink<std::mutex>;
+using rotating_file_sink_st = rotating_file_sink<details::null_mutex>;
+
+}  // namespace sinks
+
+//
+// factory functions
+//
+
+template <typename Factory = spdlog::synchronous_factory>
+inline std::shared_ptr<logger> rotating_logger_mt(const std::string &logger_name,
+                                                  const filename_t &filename,
+                                                  size_t max_file_size,
+                                                  size_t max_files,
+                                                  bool rotate_on_open = false,
+                                                  const file_event_handlers &event_handlers = {}) {
+    return Factory::template create<sinks::rotating_file_sink_mt>(
+        logger_name, filename, max_file_size, max_files, rotate_on_open, event_handlers);
+}
+
+template <typename Factory = spdlog::synchronous_factory>
+inline std::shared_ptr<logger> rotating_logger_st(const std::string &logger_name,
+                                                  const filename_t &filename,
+                                                  size_t max_file_size,
+                                                  size_t max_files,
+                                                  bool rotate_on_open = false,
+                                                  const file_event_handlers &event_handlers = {}) {
+    return Factory::template create<sinks::rotating_file_sink_st>(
+        logger_name, filename, max_file_size, max_files, rotate_on_open, event_handlers);
+}
+}  // namespace spdlog
+
+#ifdef SPDLOG_HEADER_ONLY
+    #include "rotating_file_sink-inl.h"
+#endif
diff --git a/ext/spdlog/sinks/sink-inl.h b/ext/spdlog/sinks/sink-inl.h
new file mode 100644
index 0000000..e4b2714
--- /dev/null
+++ b/ext/spdlog/sinks/sink-inl.h
@@ -0,0 +1,22 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#ifndef SPDLOG_HEADER_ONLY
+    #include <spdlog/sinks/sink.h>
+#endif
+
+#include <spdlog/common.h>
+
+SPDLOG_INLINE bool spdlog::sinks::sink::should_log(spdlog::level::level_enum msg_level) const {
+    return msg_level >= level_.load(std::memory_order_relaxed);
+}
+
+SPDLOG_INLINE void spdlog::sinks::sink::set_level(level::level_enum log_level) {
+    level_.store(log_level, std::memory_order_relaxed);
+}
+
+SPDLOG_INLINE spdlog::level::level_enum spdlog::sinks::sink::level() const {
+    return static_cast<spdlog::level::level_enum>(level_.load(std::memory_order_relaxed));
+}
diff --git a/ext/spdlog/sinks/sink.h b/ext/spdlog/sinks/sink.h
new file mode 100644
index 0000000..5850685
--- /dev/null
+++ b/ext/spdlog/sinks/sink.h
@@ -0,0 +1,34 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#include <spdlog/details/log_msg.h>
+#include <spdlog/formatter.h>
+
+namespace spdlog {
+
+namespace sinks {
+class SPDLOG_API sink {
+public:
+    virtual ~sink() = default;
+    virtual void log(const details::log_msg &msg) = 0;
+    virtual void flush() = 0;
+    virtual void set_pattern(const std::string &pattern) = 0;
+    virtual void set_formatter(std::unique_ptr<spdlog::formatter> sink_formatter) = 0;
+
+    void set_level(level::level_enum log_level);
+    level::level_enum level() const;
+    bool should_log(level::level_enum msg_level) const;
+
+protected:
+    // sink log level - default is all
+    level_t level_{level::trace};
+};
+
+}  // namespace sinks
+}  // namespace spdlog
+
+#ifdef SPDLOG_HEADER_ONLY
+    #include "sink-inl.h"
+#endif
diff --git a/ext/spdlog/sinks/stdout_color_sinks-inl.h b/ext/spdlog/sinks/stdout_color_sinks-inl.h
new file mode 100644
index 0000000..166e386
--- /dev/null
+++ b/ext/spdlog/sinks/stdout_color_sinks-inl.h
@@ -0,0 +1,38 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#ifndef SPDLOG_HEADER_ONLY
+    #include <spdlog/sinks/stdout_color_sinks.h>
+#endif
+
+#include <spdlog/common.h>
+#include <spdlog/logger.h>
+
+namespace spdlog {
+
+template <typename Factory>
+SPDLOG_INLINE std::shared_ptr<logger> stdout_color_mt(const std::string &logger_name,
+                                                      color_mode mode) {
+    return Factory::template create<sinks::stdout_color_sink_mt>(logger_name, mode);
+}
+
+template <typename Factory>
+SPDLOG_INLINE std::shared_ptr<logger> stdout_color_st(const std::string &logger_name,
+                                                      color_mode mode) {
+    return Factory::template create<sinks::stdout_color_sink_st>(logger_name, mode);
+}
+
+template <typename Factory>
+SPDLOG_INLINE std::shared_ptr<logger> stderr_color_mt(const std::string &logger_name,
+                                                      color_mode mode) {
+    return Factory::template create<sinks::stderr_color_sink_mt>(logger_name, mode);
+}
+
+template <typename Factory>
+SPDLOG_INLINE std::shared_ptr<logger> stderr_color_st(const std::string &logger_name,
+                                                      color_mode mode) {
+    return Factory::template create<sinks::stderr_color_sink_st>(logger_name, mode);
+}
+}  // namespace spdlog
diff --git a/ext/spdlog/sinks/stdout_color_sinks.h b/ext/spdlog/sinks/stdout_color_sinks.h
new file mode 100644
index 0000000..72991fe
--- /dev/null
+++ b/ext/spdlog/sinks/stdout_color_sinks.h
@@ -0,0 +1,49 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#ifdef _WIN32
+    #include <spdlog/sinks/wincolor_sink.h>
+#else
+    #include <spdlog/sinks/ansicolor_sink.h>
+#endif
+
+#include <spdlog/details/synchronous_factory.h>
+
+namespace spdlog {
+namespace sinks {
+#ifdef _WIN32
+using stdout_color_sink_mt = wincolor_stdout_sink_mt;
+using stdout_color_sink_st = wincolor_stdout_sink_st;
+using stderr_color_sink_mt = wincolor_stderr_sink_mt;
+using stderr_color_sink_st = wincolor_stderr_sink_st;
+#else
+using stdout_color_sink_mt = ansicolor_stdout_sink_mt;
+using stdout_color_sink_st = ansicolor_stdout_sink_st;
+using stderr_color_sink_mt = ansicolor_stderr_sink_mt;
+using stderr_color_sink_st = ansicolor_stderr_sink_st;
+#endif
+}  // namespace sinks
+
+template <typename Factory = spdlog::synchronous_factory>
+std::shared_ptr<logger> stdout_color_mt(const std::string &logger_name,
+                                        color_mode mode = color_mode::automatic);
+
+template <typename Factory = spdlog::synchronous_factory>
+std::shared_ptr<logger> stdout_color_st(const std::string &logger_name,
+                                        color_mode mode = color_mode::automatic);
+
+template <typename Factory = spdlog::synchronous_factory>
+std::shared_ptr<logger> stderr_color_mt(const std::string &logger_name,
+                                        color_mode mode = color_mode::automatic);
+
+template <typename Factory = spdlog::synchronous_factory>
+std::shared_ptr<logger> stderr_color_st(const std::string &logger_name,
+                                        color_mode mode = color_mode::automatic);
+
+}  // namespace spdlog
+
+#ifdef SPDLOG_HEADER_ONLY
+    #include "stdout_color_sinks-inl.h"
+#endif
diff --git a/ext/spdlog/sinks/stdout_sinks-inl.h b/ext/spdlog/sinks/stdout_sinks-inl.h
new file mode 100644
index 0000000..dcb21d8
--- /dev/null
+++ b/ext/spdlog/sinks/stdout_sinks-inl.h
@@ -0,0 +1,127 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#ifndef SPDLOG_HEADER_ONLY
+    #include <spdlog/sinks/stdout_sinks.h>
+#endif
+
+#include <memory>
+#include <spdlog/details/console_globals.h>
+#include <spdlog/pattern_formatter.h>
+#include <spdlog/details/os.h>
+
+#ifdef _WIN32
+    // under windows using fwrite to non-binary stream results in \r\r\n (see issue #1675)
+    // so instead we use ::FileWrite
+    #include <spdlog/details/windows_include.h>
+
+    #ifndef _USING_V110_SDK71_  // fileapi.h doesn't exist in winxp
+        #include <fileapi.h>    // WriteFile (..)
+    #endif
+
+    #include <io.h>     // _get_osfhandle(..)
+    #include <stdio.h>  // _fileno(..)
+#endif                  // _WIN32
+
+namespace spdlog {
+
+namespace sinks {
+
+template <typename ConsoleMutex>
+SPDLOG_INLINE stdout_sink_base<ConsoleMutex>::stdout_sink_base(FILE *file)
+    : mutex_(ConsoleMutex::mutex()),
+      file_(file),
+      formatter_(details::make_unique<spdlog::pattern_formatter>()) {
+#ifdef _WIN32
+    // get windows handle from the FILE* object
+
+    handle_ = reinterpret_cast<HANDLE>(::_get_osfhandle(::_fileno(file_)));
+
+    // don't throw to support cases where no console is attached,
+    // and let the log method to do nothing if (handle_ == INVALID_HANDLE_VALUE).
+    // throw only if non stdout/stderr target is requested (probably regular file and not console).
+    if (handle_ == INVALID_HANDLE_VALUE && file != stdout && file != stderr) {
+        throw_spdlog_ex("spdlog::stdout_sink_base: _get_osfhandle() failed", errno);
+    }
+#endif  // _WIN32
+}
+
+template <typename ConsoleMutex>
+SPDLOG_INLINE void stdout_sink_base<ConsoleMutex>::log(const details::log_msg &msg) {
+#ifdef _WIN32
+    if (handle_ == INVALID_HANDLE_VALUE) {
+        return;
+    }
+    std::lock_guard<mutex_t> lock(mutex_);
+    memory_buf_t formatted;
+    formatter_->format(msg, formatted);
+    auto size = static_cast<DWORD>(formatted.size());
+    DWORD bytes_written = 0;
+    bool ok = ::WriteFile(handle_, formatted.data(), size, &bytes_written, nullptr) != 0;
+    if (!ok) {
+        throw_spdlog_ex("stdout_sink_base: WriteFile() failed. GetLastError(): " +
+                        std::to_string(::GetLastError()));
+    }
+#else
+    std::lock_guard<mutex_t> lock(mutex_);
+    memory_buf_t formatted;
+    formatter_->format(msg, formatted);
+    details::os::fwrite_bytes(formatted.data(), formatted.size(), file_);
+#endif                // _WIN32
+    ::fflush(file_);  // flush every line to terminal
+}
+
+template <typename ConsoleMutex>
+SPDLOG_INLINE void stdout_sink_base<ConsoleMutex>::flush() {
+    std::lock_guard<mutex_t> lock(mutex_);
+    fflush(file_);
+}
+
+template <typename ConsoleMutex>
+SPDLOG_INLINE void stdout_sink_base<ConsoleMutex>::set_pattern(const std::string &pattern) {
+    std::lock_guard<mutex_t> lock(mutex_);
+    formatter_ = std::unique_ptr<spdlog::formatter>(new pattern_formatter(pattern));
+}
+
+template <typename ConsoleMutex>
+SPDLOG_INLINE void stdout_sink_base<ConsoleMutex>::set_formatter(
+    std::unique_ptr<spdlog::formatter> sink_formatter) {
+    std::lock_guard<mutex_t> lock(mutex_);
+    formatter_ = std::move(sink_formatter);
+}
+
+// stdout sink
+template <typename ConsoleMutex>
+SPDLOG_INLINE stdout_sink<ConsoleMutex>::stdout_sink()
+    : stdout_sink_base<ConsoleMutex>(stdout) {}
+
+// stderr sink
+template <typename ConsoleMutex>
+SPDLOG_INLINE stderr_sink<ConsoleMutex>::stderr_sink()
+    : stdout_sink_base<ConsoleMutex>(stderr) {}
+
+}  // namespace sinks
+
+// factory methods
+template <typename Factory>
+SPDLOG_INLINE std::shared_ptr<logger> stdout_logger_mt(const std::string &logger_name) {
+    return Factory::template create<sinks::stdout_sink_mt>(logger_name);
+}
+
+template <typename Factory>
+SPDLOG_INLINE std::shared_ptr<logger> stdout_logger_st(const std::string &logger_name) {
+    return Factory::template create<sinks::stdout_sink_st>(logger_name);
+}
+
+template <typename Factory>
+SPDLOG_INLINE std::shared_ptr<logger> stderr_logger_mt(const std::string &logger_name) {
+    return Factory::template create<sinks::stderr_sink_mt>(logger_name);
+}
+
+template <typename Factory>
+SPDLOG_INLINE std::shared_ptr<logger> stderr_logger_st(const std::string &logger_name) {
+    return Factory::template create<sinks::stderr_sink_st>(logger_name);
+}
+}  // namespace spdlog
diff --git a/ext/spdlog/sinks/stdout_sinks.h b/ext/spdlog/sinks/stdout_sinks.h
new file mode 100644
index 0000000..6ef0996
--- /dev/null
+++ b/ext/spdlog/sinks/stdout_sinks.h
@@ -0,0 +1,84 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#include <cstdio>
+#include <spdlog/details/console_globals.h>
+#include <spdlog/details/synchronous_factory.h>
+#include <spdlog/sinks/sink.h>
+
+#ifdef _WIN32
+    #include <spdlog/details/windows_include.h>
+#endif
+
+namespace spdlog {
+
+namespace sinks {
+
+template <typename ConsoleMutex>
+class stdout_sink_base : public sink {
+public:
+    using mutex_t = typename ConsoleMutex::mutex_t;
+    explicit stdout_sink_base(FILE *file);
+    ~stdout_sink_base() override = default;
+
+    stdout_sink_base(const stdout_sink_base &other) = delete;
+    stdout_sink_base(stdout_sink_base &&other) = delete;
+
+    stdout_sink_base &operator=(const stdout_sink_base &other) = delete;
+    stdout_sink_base &operator=(stdout_sink_base &&other) = delete;
+
+    void log(const details::log_msg &msg) override;
+    void flush() override;
+    void set_pattern(const std::string &pattern) override;
+
+    void set_formatter(std::unique_ptr<spdlog::formatter> sink_formatter) override;
+
+protected:
+    mutex_t &mutex_;
+    FILE *file_;
+    std::unique_ptr<spdlog::formatter> formatter_;
+#ifdef _WIN32
+    HANDLE handle_;
+#endif  // WIN32
+};
+
+template <typename ConsoleMutex>
+class stdout_sink : public stdout_sink_base<ConsoleMutex> {
+public:
+    stdout_sink();
+};
+
+template <typename ConsoleMutex>
+class stderr_sink : public stdout_sink_base<ConsoleMutex> {
+public:
+    stderr_sink();
+};
+
+using stdout_sink_mt = stdout_sink<details::console_mutex>;
+using stdout_sink_st = stdout_sink<details::console_nullmutex>;
+
+using stderr_sink_mt = stderr_sink<details::console_mutex>;
+using stderr_sink_st = stderr_sink<details::console_nullmutex>;
+
+}  // namespace sinks
+
+// factory methods
+template <typename Factory = spdlog::synchronous_factory>
+std::shared_ptr<logger> stdout_logger_mt(const std::string &logger_name);
+
+template <typename Factory = spdlog::synchronous_factory>
+std::shared_ptr<logger> stdout_logger_st(const std::string &logger_name);
+
+template <typename Factory = spdlog::synchronous_factory>
+std::shared_ptr<logger> stderr_logger_mt(const std::string &logger_name);
+
+template <typename Factory = spdlog::synchronous_factory>
+std::shared_ptr<logger> stderr_logger_st(const std::string &logger_name);
+
+}  // namespace spdlog
+
+#ifdef SPDLOG_HEADER_ONLY
+    #include "stdout_sinks-inl.h"
+#endif
diff --git a/ext/spdlog/sinks/syslog_sink.h b/ext/spdlog/sinks/syslog_sink.h
new file mode 100644
index 0000000..913d41b
--- /dev/null
+++ b/ext/spdlog/sinks/syslog_sink.h
@@ -0,0 +1,104 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#include <spdlog/details/null_mutex.h>
+#include <spdlog/details/synchronous_factory.h>
+#include <spdlog/sinks/base_sink.h>
+
+#include <array>
+#include <string>
+#include <syslog.h>
+
+namespace spdlog {
+namespace sinks {
+/**
+ * Sink that write to syslog using the `syscall()` library call.
+ */
+template <typename Mutex>
+class syslog_sink : public base_sink<Mutex> {
+public:
+    syslog_sink(std::string ident, int syslog_option, int syslog_facility, bool enable_formatting)
+        : enable_formatting_{enable_formatting},
+          syslog_levels_{{/* spdlog::level::trace      */ LOG_DEBUG,
+                          /* spdlog::level::debug      */ LOG_DEBUG,
+                          /* spdlog::level::info       */ LOG_INFO,
+                          /* spdlog::level::warn       */ LOG_WARNING,
+                          /* spdlog::level::err        */ LOG_ERR,
+                          /* spdlog::level::critical   */ LOG_CRIT,
+                          /* spdlog::level::off        */ LOG_INFO}},
+          ident_{std::move(ident)} {
+        // set ident to be program name if empty
+        ::openlog(ident_.empty() ? nullptr : ident_.c_str(), syslog_option, syslog_facility);
+    }
+
+    ~syslog_sink() override { ::closelog(); }
+
+    syslog_sink(const syslog_sink &) = delete;
+    syslog_sink &operator=(const syslog_sink &) = delete;
+
+protected:
+    void sink_it_(const details::log_msg &msg) override {
+        string_view_t payload;
+        memory_buf_t formatted;
+        if (enable_formatting_) {
+            base_sink<Mutex>::formatter_->format(msg, formatted);
+            payload = string_view_t(formatted.data(), formatted.size());
+        } else {
+            payload = msg.payload;
+        }
+
+        size_t length = payload.size();
+        // limit to max int
+        if (length > static_cast<size_t>(std::numeric_limits<int>::max())) {
+            length = static_cast<size_t>(std::numeric_limits<int>::max());
+        }
+
+        ::syslog(syslog_prio_from_level(msg), "%.*s", static_cast<int>(length), payload.data());
+    }
+
+    void flush_() override {}
+    bool enable_formatting_ = false;
+
+    //
+    // Simply maps spdlog's log level to syslog priority level.
+    //
+    virtual int syslog_prio_from_level(const details::log_msg &msg) const {
+        return syslog_levels_.at(static_cast<levels_array::size_type>(msg.level));
+    }
+
+    using levels_array = std::array<int, 7>;
+    levels_array syslog_levels_;
+
+private:
+    // must store the ident because the man says openlog might use the pointer as
+    // is and not a string copy
+    const std::string ident_;
+};
+
+using syslog_sink_mt = syslog_sink<std::mutex>;
+using syslog_sink_st = syslog_sink<details::null_mutex>;
+}  // namespace sinks
+
+// Create and register a syslog logger
+template <typename Factory = spdlog::synchronous_factory>
+inline std::shared_ptr<logger> syslog_logger_mt(const std::string &logger_name,
+                                                const std::string &syslog_ident = "",
+                                                int syslog_option = 0,
+                                                int syslog_facility = LOG_USER,
+                                                bool enable_formatting = false) {
+    return Factory::template create<sinks::syslog_sink_mt>(logger_name, syslog_ident, syslog_option,
+                                                           syslog_facility, enable_formatting);
+}
+
+template <typename Factory = spdlog::synchronous_factory>
+inline std::shared_ptr<logger> syslog_logger_st(const std::string &logger_name,
+                                                const std::string &syslog_ident = "",
+                                                int syslog_option = 0,
+                                                int syslog_facility = LOG_USER,
+                                                bool enable_formatting = false) {
+    return Factory::template create<sinks::syslog_sink_st>(logger_name, syslog_ident, syslog_option,
+                                                           syslog_facility, enable_formatting);
+}
+}  // namespace spdlog
diff --git a/ext/spdlog/sinks/systemd_sink.h b/ext/spdlog/sinks/systemd_sink.h
new file mode 100644
index 0000000..d2cd55f
--- /dev/null
+++ b/ext/spdlog/sinks/systemd_sink.h
@@ -0,0 +1,121 @@
+// Copyright(c) 2019 ZVYAGIN.Alexander@gmail.com
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#include <spdlog/details/null_mutex.h>
+#include <spdlog/details/os.h>
+#include <spdlog/details/synchronous_factory.h>
+#include <spdlog/sinks/base_sink.h>
+
+#include <array>
+#ifndef SD_JOURNAL_SUPPRESS_LOCATION
+    #define SD_JOURNAL_SUPPRESS_LOCATION
+#endif
+#include <systemd/sd-journal.h>
+
+namespace spdlog {
+namespace sinks {
+
+/**
+ * Sink that write to systemd journal using the `sd_journal_send()` library call.
+ */
+template <typename Mutex>
+class systemd_sink : public base_sink<Mutex> {
+public:
+    systemd_sink(std::string ident = "", bool enable_formatting = false)
+        : ident_{std::move(ident)},
+          enable_formatting_{enable_formatting},
+          syslog_levels_{{/* spdlog::level::trace      */ LOG_DEBUG,
+                          /* spdlog::level::debug      */ LOG_DEBUG,
+                          /* spdlog::level::info       */ LOG_INFO,
+                          /* spdlog::level::warn       */ LOG_WARNING,
+                          /* spdlog::level::err        */ LOG_ERR,
+                          /* spdlog::level::critical   */ LOG_CRIT,
+                          /* spdlog::level::off        */ LOG_INFO}} {}
+
+    ~systemd_sink() override {}
+
+    systemd_sink(const systemd_sink &) = delete;
+    systemd_sink &operator=(const systemd_sink &) = delete;
+
+protected:
+    const std::string ident_;
+    bool enable_formatting_ = false;
+    using levels_array = std::array<int, 7>;
+    levels_array syslog_levels_;
+
+    void sink_it_(const details::log_msg &msg) override {
+        int err;
+        string_view_t payload;
+        memory_buf_t formatted;
+        if (enable_formatting_) {
+            base_sink<Mutex>::formatter_->format(msg, formatted);
+            payload = string_view_t(formatted.data(), formatted.size());
+        } else {
+            payload = msg.payload;
+        }
+
+        size_t length = payload.size();
+        // limit to max int
+        if (length > static_cast<size_t>(std::numeric_limits<int>::max())) {
+            length = static_cast<size_t>(std::numeric_limits<int>::max());
+        }
+
+        const string_view_t syslog_identifier = ident_.empty() ? msg.logger_name : ident_;
+
+        // Do not send source location if not available
+        if (msg.source.empty()) {
+            // Note: function call inside '()' to avoid macro expansion
+            err = (sd_journal_send)("MESSAGE=%.*s", static_cast<int>(length), payload.data(),
+                                    "PRIORITY=%d", syslog_level(msg.level),
+#ifndef SPDLOG_NO_THREAD_ID
+                                    "TID=%zu", msg.thread_id,
+#endif
+                                    "SYSLOG_IDENTIFIER=%.*s",
+                                    static_cast<int>(syslog_identifier.size()),
+                                    syslog_identifier.data(), nullptr);
+        } else {
+            err = (sd_journal_send)("MESSAGE=%.*s", static_cast<int>(length), payload.data(),
+                                    "PRIORITY=%d", syslog_level(msg.level),
+#ifndef SPDLOG_NO_THREAD_ID
+                                    "TID=%zu", msg.thread_id,
+#endif
+                                    "SYSLOG_IDENTIFIER=%.*s",
+                                    static_cast<int>(syslog_identifier.size()),
+                                    syslog_identifier.data(), "CODE_FILE=%s", msg.source.filename,
+                                    "CODE_LINE=%d", msg.source.line, "CODE_FUNC=%s",
+                                    msg.source.funcname, nullptr);
+        }
+
+        if (err) {
+            throw_spdlog_ex("Failed writing to systemd", errno);
+        }
+    }
+
+    int syslog_level(level::level_enum l) {
+        return syslog_levels_.at(static_cast<levels_array::size_type>(l));
+    }
+
+    void flush_() override {}
+};
+
+using systemd_sink_mt = systemd_sink<std::mutex>;
+using systemd_sink_st = systemd_sink<details::null_mutex>;
+}  // namespace sinks
+
+// Create and register a syslog logger
+template <typename Factory = spdlog::synchronous_factory>
+inline std::shared_ptr<logger> systemd_logger_mt(const std::string &logger_name,
+                                                 const std::string &ident = "",
+                                                 bool enable_formatting = false) {
+    return Factory::template create<sinks::systemd_sink_mt>(logger_name, ident, enable_formatting);
+}
+
+template <typename Factory = spdlog::synchronous_factory>
+inline std::shared_ptr<logger> systemd_logger_st(const std::string &logger_name,
+                                                 const std::string &ident = "",
+                                                 bool enable_formatting = false) {
+    return Factory::template create<sinks::systemd_sink_st>(logger_name, ident, enable_formatting);
+}
+}  // namespace spdlog
diff --git a/ext/spdlog/sinks/tcp_sink.h b/ext/spdlog/sinks/tcp_sink.h
new file mode 100644
index 0000000..2534964
--- /dev/null
+++ b/ext/spdlog/sinks/tcp_sink.h
@@ -0,0 +1,75 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#include <spdlog/common.h>
+#include <spdlog/details/null_mutex.h>
+#include <spdlog/sinks/base_sink.h>
+#ifdef _WIN32
+    #include <spdlog/details/tcp_client-windows.h>
+#else
+    #include <spdlog/details/tcp_client.h>
+#endif
+
+#include <chrono>
+#include <functional>
+#include <mutex>
+#include <string>
+
+#pragma once
+
+// Simple tcp client sink
+// Connects to remote address and send the formatted log.
+// Will attempt to reconnect if connection drops.
+// If more complicated behaviour is needed (i.e get responses), you can inherit it and override the
+// sink_it_ method.
+
+namespace spdlog {
+namespace sinks {
+
+struct tcp_sink_config {
+    std::string server_host;
+    int server_port;
+    bool lazy_connect = false;  // if true connect on first log call instead of on construction
+
+    tcp_sink_config(std::string host, int port)
+        : server_host{std::move(host)},
+          server_port{port} {}
+};
+
+template <typename Mutex>
+class tcp_sink : public spdlog::sinks::base_sink<Mutex> {
+public:
+    // connect to tcp host/port or throw if failed
+    // host can be hostname or ip address
+
+    explicit tcp_sink(tcp_sink_config sink_config)
+        : config_{std::move(sink_config)} {
+        if (!config_.lazy_connect) {
+            this->client_.connect(config_.server_host, config_.server_port);
+        }
+    }
+
+    ~tcp_sink() override = default;
+
+protected:
+    void sink_it_(const spdlog::details::log_msg &msg) override {
+        spdlog::memory_buf_t formatted;
+        spdlog::sinks::base_sink<Mutex>::formatter_->format(msg, formatted);
+        if (!client_.is_connected()) {
+            client_.connect(config_.server_host, config_.server_port);
+        }
+        client_.send(formatted.data(), formatted.size());
+    }
+
+    void flush_() override {}
+    tcp_sink_config config_;
+    details::tcp_client client_;
+};
+
+using tcp_sink_mt = tcp_sink<std::mutex>;
+using tcp_sink_st = tcp_sink<spdlog::details::null_mutex>;
+
+}  // namespace sinks
+}  // namespace spdlog
diff --git a/ext/spdlog/sinks/udp_sink.h b/ext/spdlog/sinks/udp_sink.h
new file mode 100644
index 0000000..4bff0fd
--- /dev/null
+++ b/ext/spdlog/sinks/udp_sink.h
@@ -0,0 +1,69 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#include <spdlog/common.h>
+#include <spdlog/details/null_mutex.h>
+#include <spdlog/sinks/base_sink.h>
+#ifdef _WIN32
+    #include <spdlog/details/udp_client-windows.h>
+#else
+    #include <spdlog/details/udp_client.h>
+#endif
+
+#include <chrono>
+#include <functional>
+#include <mutex>
+#include <string>
+
+// Simple udp client sink
+// Sends formatted log via udp
+
+namespace spdlog {
+namespace sinks {
+
+struct udp_sink_config {
+    std::string server_host;
+    uint16_t server_port;
+
+    udp_sink_config(std::string host, uint16_t port)
+        : server_host{std::move(host)},
+          server_port{port} {}
+};
+
+template <typename Mutex>
+class udp_sink : public spdlog::sinks::base_sink<Mutex> {
+public:
+    // host can be hostname or ip address
+    explicit udp_sink(udp_sink_config sink_config)
+        : client_{sink_config.server_host, sink_config.server_port} {}
+
+    ~udp_sink() override = default;
+
+protected:
+    void sink_it_(const spdlog::details::log_msg &msg) override {
+        spdlog::memory_buf_t formatted;
+        spdlog::sinks::base_sink<Mutex>::formatter_->format(msg, formatted);
+        client_.send(formatted.data(), formatted.size());
+    }
+
+    void flush_() override {}
+    details::udp_client client_;
+};
+
+using udp_sink_mt = udp_sink<std::mutex>;
+using udp_sink_st = udp_sink<spdlog::details::null_mutex>;
+
+}  // namespace sinks
+
+//
+// factory functions
+//
+template <typename Factory = spdlog::synchronous_factory>
+inline std::shared_ptr<logger> udp_logger_mt(const std::string &logger_name,
+                                             sinks::udp_sink_config skin_config) {
+    return Factory::template create<sinks::udp_sink_mt>(logger_name, skin_config);
+}
+
+}  // namespace spdlog
diff --git a/ext/spdlog/sinks/win_eventlog_sink.h b/ext/spdlog/sinks/win_eventlog_sink.h
new file mode 100644
index 0000000..2c9b582
--- /dev/null
+++ b/ext/spdlog/sinks/win_eventlog_sink.h
@@ -0,0 +1,260 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+// Writing to Windows Event Log requires the registry entries below to be present, with the
+// following modifications:
+// 1. <log_name>    should be replaced with your log name (e.g. your application name)
+// 2. <source_name> should be replaced with the specific source name and the key should be
+// duplicated for
+//                  each source used in the application
+//
+// Since typically modifications of this kind require elevation, it's better to do it as a part of
+// setup procedure. The snippet below uses mscoree.dll as the message file as it exists on most of
+// the Windows systems anyway and happens to contain the needed resource.
+//
+// You can also specify a custom message file if needed.
+// Please refer to Event Log functions descriptions in MSDN for more details on custom message
+// files.
+
+/*---------------------------------------------------------------------------------------
+
+Windows Registry Editor Version 5.00
+
+[HKEY_LOCAL_MACHINE\SYSTEM\CurrentControlSet\Services\EventLog\<log_name>]
+
+[HKEY_LOCAL_MACHINE\SYSTEM\CurrentControlSet\Services\EventLog\<log_name>\<source_name>]
+"TypesSupported"=dword:00000007
+"EventMessageFile"=hex(2):25,00,73,00,79,00,73,00,74,00,65,00,6d,00,72,00,6f,\
+  00,6f,00,74,00,25,00,5c,00,53,00,79,00,73,00,74,00,65,00,6d,00,33,00,32,00,\
+  5c,00,6d,00,73,00,63,00,6f,00,72,00,65,00,65,00,2e,00,64,00,6c,00,6c,00,00,\
+  00
+
+-----------------------------------------------------------------------------------------*/
+
+#pragma once
+
+#include <spdlog/details/null_mutex.h>
+#include <spdlog/sinks/base_sink.h>
+
+#include <spdlog/details/windows_include.h>
+#include <winbase.h>
+
+#include <mutex>
+#include <string>
+#include <vector>
+
+namespace spdlog {
+namespace sinks {
+
+namespace win_eventlog {
+
+namespace internal {
+
+struct local_alloc_t {
+    HLOCAL hlocal_;
+
+    SPDLOG_CONSTEXPR local_alloc_t() SPDLOG_NOEXCEPT : hlocal_(nullptr) {}
+
+    local_alloc_t(local_alloc_t const &) = delete;
+    local_alloc_t &operator=(local_alloc_t const &) = delete;
+
+    ~local_alloc_t() SPDLOG_NOEXCEPT {
+        if (hlocal_) {
+            LocalFree(hlocal_);
+        }
+    }
+};
+
+/** Windows error */
+struct win32_error : public spdlog_ex {
+    /** Formats an error report line: "user-message: error-code (system message)" */
+    static std::string format(std::string const &user_message, DWORD error_code = GetLastError()) {
+        std::string system_message;
+
+        local_alloc_t format_message_result{};
+        auto format_message_succeeded =
+            ::FormatMessageA(FORMAT_MESSAGE_ALLOCATE_BUFFER | FORMAT_MESSAGE_FROM_SYSTEM |
+                                 FORMAT_MESSAGE_IGNORE_INSERTS,
+                             nullptr, error_code, MAKELANGID(LANG_NEUTRAL, SUBLANG_DEFAULT),
+                             (LPSTR)&format_message_result.hlocal_, 0, nullptr);
+
+        if (format_message_succeeded && format_message_result.hlocal_) {
+            system_message = fmt_lib::format(" ({})", (LPSTR)format_message_result.hlocal_);
+        }
+
+        return fmt_lib::format("{}: {}{}", user_message, error_code, system_message);
+    }
+
+    explicit win32_error(std::string const &func_name, DWORD error = GetLastError())
+        : spdlog_ex(format(func_name, error)) {}
+};
+
+/** Wrapper for security identifiers (SID) on Windows */
+struct sid_t {
+    std::vector<char> buffer_;
+
+public:
+    sid_t() {}
+
+    /** creates a wrapped SID copy */
+    static sid_t duplicate_sid(PSID psid) {
+        if (!::IsValidSid(psid)) {
+            throw_spdlog_ex("sid_t::sid_t(): invalid SID received");
+        }
+
+        auto const sid_length{::GetLengthSid(psid)};
+
+        sid_t result;
+        result.buffer_.resize(sid_length);
+        if (!::CopySid(sid_length, (PSID)result.as_sid(), psid)) {
+            SPDLOG_THROW(win32_error("CopySid"));
+        }
+
+        return result;
+    }
+
+    /** Retrieves pointer to the internal buffer contents as SID* */
+    SID *as_sid() const { return buffer_.empty() ? nullptr : (SID *)buffer_.data(); }
+
+    /** Get SID for the current user */
+    static sid_t get_current_user_sid() {
+        /* create and init RAII holder for process token */
+        struct process_token_t {
+            HANDLE token_handle_ = INVALID_HANDLE_VALUE;
+            explicit process_token_t(HANDLE process) {
+                if (!::OpenProcessToken(process, TOKEN_QUERY, &token_handle_)) {
+                    SPDLOG_THROW(win32_error("OpenProcessToken"));
+                }
+            }
+
+            ~process_token_t() { ::CloseHandle(token_handle_); }
+
+        } current_process_token(
+            ::GetCurrentProcess());  // GetCurrentProcess returns pseudohandle, no leak here!
+
+        // Get the required size, this is expected to fail with ERROR_INSUFFICIENT_BUFFER and return
+        // the token size
+        DWORD tusize = 0;
+        if (::GetTokenInformation(current_process_token.token_handle_, TokenUser, NULL, 0,
+                                  &tusize)) {
+            SPDLOG_THROW(win32_error("GetTokenInformation should fail"));
+        }
+
+        // get user token
+        std::vector<unsigned char> buffer(static_cast<size_t>(tusize));
+        if (!::GetTokenInformation(current_process_token.token_handle_, TokenUser,
+                                   (LPVOID)buffer.data(), tusize, &tusize)) {
+            SPDLOG_THROW(win32_error("GetTokenInformation"));
+        }
+
+        // create a wrapper of the SID data as stored in the user token
+        return sid_t::duplicate_sid(((TOKEN_USER *)buffer.data())->User.Sid);
+    }
+};
+
+struct eventlog {
+    static WORD get_event_type(details::log_msg const &msg) {
+        switch (msg.level) {
+            case level::trace:
+            case level::debug:
+                return EVENTLOG_SUCCESS;
+
+            case level::info:
+                return EVENTLOG_INFORMATION_TYPE;
+
+            case level::warn:
+                return EVENTLOG_WARNING_TYPE;
+
+            case level::err:
+            case level::critical:
+            case level::off:
+                return EVENTLOG_ERROR_TYPE;
+
+            default:
+                return EVENTLOG_INFORMATION_TYPE;
+        }
+    }
+
+    static WORD get_event_category(details::log_msg const &msg) { return (WORD)msg.level; }
+};
+
+}  // namespace internal
+
+/*
+ * Windows Event Log sink
+ */
+template <typename Mutex>
+class win_eventlog_sink : public base_sink<Mutex> {
+private:
+    HANDLE hEventLog_{NULL};
+    internal::sid_t current_user_sid_;
+    std::string source_;
+    DWORD event_id_;
+
+    HANDLE event_log_handle() {
+        if (!hEventLog_) {
+            hEventLog_ = ::RegisterEventSourceA(nullptr, source_.c_str());
+            if (!hEventLog_ || hEventLog_ == (HANDLE)ERROR_ACCESS_DENIED) {
+                SPDLOG_THROW(internal::win32_error("RegisterEventSource"));
+            }
+        }
+
+        return hEventLog_;
+    }
+
+protected:
+    void sink_it_(const details::log_msg &msg) override {
+        using namespace internal;
+
+        bool succeeded;
+        memory_buf_t formatted;
+        base_sink<Mutex>::formatter_->format(msg, formatted);
+        formatted.push_back('\0');
+
+#ifdef SPDLOG_WCHAR_TO_UTF8_SUPPORT
+        wmemory_buf_t buf;
+        details::os::utf8_to_wstrbuf(string_view_t(formatted.data(), formatted.size()), buf);
+
+        LPCWSTR lp_wstr = buf.data();
+        succeeded = static_cast<bool>(::ReportEventW(
+            event_log_handle(), eventlog::get_event_type(msg), eventlog::get_event_category(msg),
+            event_id_, current_user_sid_.as_sid(), 1, 0, &lp_wstr, nullptr));
+#else
+        LPCSTR lp_str = formatted.data();
+        succeeded = static_cast<bool>(::ReportEventA(
+            event_log_handle(), eventlog::get_event_type(msg), eventlog::get_event_category(msg),
+            event_id_, current_user_sid_.as_sid(), 1, 0, &lp_str, nullptr));
+#endif
+
+        if (!succeeded) {
+            SPDLOG_THROW(win32_error("ReportEvent"));
+        }
+    }
+
+    void flush_() override {}
+
+public:
+    win_eventlog_sink(std::string const &source,
+                      DWORD event_id = 1000 /* according to mscoree.dll */)
+        : source_(source),
+          event_id_(event_id) {
+        try {
+            current_user_sid_ = internal::sid_t::get_current_user_sid();
+        } catch (...) {
+            // get_current_user_sid() is unlikely to fail and if it does, we can still proceed
+            // without current_user_sid but in the event log the record will have no user name
+        }
+    }
+
+    ~win_eventlog_sink() {
+        if (hEventLog_) DeregisterEventSource(hEventLog_);
+    }
+};
+
+}  // namespace win_eventlog
+
+using win_eventlog_sink_mt = win_eventlog::win_eventlog_sink<std::mutex>;
+using win_eventlog_sink_st = win_eventlog::win_eventlog_sink<details::null_mutex>;
+
+}  // namespace sinks
+}  // namespace spdlog
diff --git a/ext/spdlog/sinks/wincolor_sink-inl.h b/ext/spdlog/sinks/wincolor_sink-inl.h
new file mode 100644
index 0000000..696db56
--- /dev/null
+++ b/ext/spdlog/sinks/wincolor_sink-inl.h
@@ -0,0 +1,172 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#ifndef SPDLOG_HEADER_ONLY
+    #include <spdlog/sinks/wincolor_sink.h>
+#endif
+
+#include <spdlog/details/windows_include.h>
+#include <wincon.h>
+
+#include <spdlog/common.h>
+#include <spdlog/pattern_formatter.h>
+
+namespace spdlog {
+namespace sinks {
+template <typename ConsoleMutex>
+SPDLOG_INLINE wincolor_sink<ConsoleMutex>::wincolor_sink(void *out_handle, color_mode mode)
+    : out_handle_(out_handle),
+      mutex_(ConsoleMutex::mutex()),
+      formatter_(details::make_unique<spdlog::pattern_formatter>()) {
+    set_color_mode_impl(mode);
+    // set level colors
+    colors_[level::trace] = FOREGROUND_RED | FOREGROUND_GREEN | FOREGROUND_BLUE;  // white
+    colors_[level::debug] = FOREGROUND_GREEN | FOREGROUND_BLUE;                   // cyan
+    colors_[level::info] = FOREGROUND_GREEN;                                      // green
+    colors_[level::warn] =
+        FOREGROUND_RED | FOREGROUND_GREEN | FOREGROUND_INTENSITY;  // intense yellow
+    colors_[level::err] = FOREGROUND_RED | FOREGROUND_INTENSITY;   // intense red
+    colors_[level::critical] = BACKGROUND_RED | FOREGROUND_RED | FOREGROUND_GREEN |
+                               FOREGROUND_BLUE |
+                               FOREGROUND_INTENSITY;  // intense white on red background
+    colors_[level::off] = 0;
+}
+
+template <typename ConsoleMutex>
+SPDLOG_INLINE wincolor_sink<ConsoleMutex>::~wincolor_sink() {
+    this->flush();
+}
+
+// change the color for the given level
+template <typename ConsoleMutex>
+void SPDLOG_INLINE wincolor_sink<ConsoleMutex>::set_color(level::level_enum level,
+                                                          std::uint16_t color) {
+    std::lock_guard<mutex_t> lock(mutex_);
+    colors_[static_cast<size_t>(level)] = color;
+}
+
+template <typename ConsoleMutex>
+void SPDLOG_INLINE wincolor_sink<ConsoleMutex>::log(const details::log_msg &msg) {
+    if (out_handle_ == nullptr || out_handle_ == INVALID_HANDLE_VALUE) {
+        return;
+    }
+
+    std::lock_guard<mutex_t> lock(mutex_);
+    msg.color_range_start = 0;
+    msg.color_range_end = 0;
+    memory_buf_t formatted;
+    formatter_->format(msg, formatted);
+    if (should_do_colors_ && msg.color_range_end > msg.color_range_start) {
+        // before color range
+        print_range_(formatted, 0, msg.color_range_start);
+        // in color range
+        auto orig_attribs =
+            static_cast<WORD>(set_foreground_color_(colors_[static_cast<size_t>(msg.level)]));
+        print_range_(formatted, msg.color_range_start, msg.color_range_end);
+        // reset to orig colors
+        ::SetConsoleTextAttribute(static_cast<HANDLE>(out_handle_), orig_attribs);
+        print_range_(formatted, msg.color_range_end, formatted.size());
+    } else  // print without colors if color range is invalid (or color is disabled)
+    {
+        write_to_file_(formatted);
+    }
+}
+
+template <typename ConsoleMutex>
+void SPDLOG_INLINE wincolor_sink<ConsoleMutex>::flush() {
+    // windows console always flushed?
+}
+
+template <typename ConsoleMutex>
+void SPDLOG_INLINE wincolor_sink<ConsoleMutex>::set_pattern(const std::string &pattern) {
+    std::lock_guard<mutex_t> lock(mutex_);
+    formatter_ = std::unique_ptr<spdlog::formatter>(new pattern_formatter(pattern));
+}
+
+template <typename ConsoleMutex>
+void SPDLOG_INLINE
+wincolor_sink<ConsoleMutex>::set_formatter(std::unique_ptr<spdlog::formatter> sink_formatter) {
+    std::lock_guard<mutex_t> lock(mutex_);
+    formatter_ = std::move(sink_formatter);
+}
+
+template <typename ConsoleMutex>
+void SPDLOG_INLINE wincolor_sink<ConsoleMutex>::set_color_mode(color_mode mode) {
+    std::lock_guard<mutex_t> lock(mutex_);
+    set_color_mode_impl(mode);
+}
+
+template <typename ConsoleMutex>
+void SPDLOG_INLINE wincolor_sink<ConsoleMutex>::set_color_mode_impl(color_mode mode) {
+    if (mode == color_mode::automatic) {
+        // should do colors only if out_handle_  points to actual console.
+        DWORD console_mode;
+        bool in_console = ::GetConsoleMode(static_cast<HANDLE>(out_handle_), &console_mode) != 0;
+        should_do_colors_ = in_console;
+    } else {
+        should_do_colors_ = mode == color_mode::always ? true : false;
+    }
+}
+
+// set foreground color and return the orig console attributes (for resetting later)
+template <typename ConsoleMutex>
+std::uint16_t SPDLOG_INLINE
+wincolor_sink<ConsoleMutex>::set_foreground_color_(std::uint16_t attribs) {
+    CONSOLE_SCREEN_BUFFER_INFO orig_buffer_info;
+    if (!::GetConsoleScreenBufferInfo(static_cast<HANDLE>(out_handle_), &orig_buffer_info)) {
+        // just return white if failed getting console info
+        return FOREGROUND_RED | FOREGROUND_GREEN | FOREGROUND_BLUE;
+    }
+
+    // change only the foreground bits (lowest 4 bits)
+    auto new_attribs = static_cast<WORD>(attribs) | (orig_buffer_info.wAttributes & 0xfff0);
+    auto ignored =
+        ::SetConsoleTextAttribute(static_cast<HANDLE>(out_handle_), static_cast<WORD>(new_attribs));
+    (void)(ignored);
+    return static_cast<std::uint16_t>(orig_buffer_info.wAttributes);  // return orig attribs
+}
+
+// print a range of formatted message to console
+template <typename ConsoleMutex>
+void SPDLOG_INLINE wincolor_sink<ConsoleMutex>::print_range_(const memory_buf_t &formatted,
+                                                             size_t start,
+                                                             size_t end) {
+    if (end > start) {
+#if defined(SPDLOG_UTF8_TO_WCHAR_CONSOLE)
+        wmemory_buf_t wformatted;
+        details::os::utf8_to_wstrbuf(string_view_t(formatted.data() + start, end - start),
+            wformatted);
+        auto size = static_cast<DWORD>(wformatted.size());
+        auto ignored = ::WriteConsoleW(static_cast<HANDLE>(out_handle_), wformatted.data(), size,
+            nullptr, nullptr);
+#else
+        auto size = static_cast<DWORD>(end - start);
+        auto ignored = ::WriteConsoleA(static_cast<HANDLE>(out_handle_), formatted.data() + start,
+                                       size, nullptr, nullptr);
+#endif
+        (void)(ignored);
+    }
+}
+
+template <typename ConsoleMutex>
+void SPDLOG_INLINE wincolor_sink<ConsoleMutex>::write_to_file_(const memory_buf_t &formatted) {
+    auto size = static_cast<DWORD>(formatted.size());
+    DWORD bytes_written = 0;
+    auto ignored = ::WriteFile(static_cast<HANDLE>(out_handle_), formatted.data(), size,
+                               &bytes_written, nullptr);
+    (void)(ignored);
+}
+
+// wincolor_stdout_sink
+template <typename ConsoleMutex>
+SPDLOG_INLINE wincolor_stdout_sink<ConsoleMutex>::wincolor_stdout_sink(color_mode mode)
+    : wincolor_sink<ConsoleMutex>(::GetStdHandle(STD_OUTPUT_HANDLE), mode) {}
+
+// wincolor_stderr_sink
+template <typename ConsoleMutex>
+SPDLOG_INLINE wincolor_stderr_sink<ConsoleMutex>::wincolor_stderr_sink(color_mode mode)
+    : wincolor_sink<ConsoleMutex>(::GetStdHandle(STD_ERROR_HANDLE), mode) {}
+}  // namespace sinks
+}  // namespace spdlog
diff --git a/ext/spdlog/sinks/wincolor_sink.h b/ext/spdlog/sinks/wincolor_sink.h
new file mode 100644
index 0000000..8ba594c
--- /dev/null
+++ b/ext/spdlog/sinks/wincolor_sink.h
@@ -0,0 +1,82 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#include <spdlog/common.h>
+#include <spdlog/details/console_globals.h>
+#include <spdlog/details/null_mutex.h>
+#include <spdlog/sinks/sink.h>
+
+#include <array>
+#include <cstdint>
+#include <memory>
+#include <mutex>
+#include <string>
+
+namespace spdlog {
+namespace sinks {
+/*
+ * Windows color console sink. Uses WriteConsoleA to write to the console with
+ * colors
+ */
+template <typename ConsoleMutex>
+class wincolor_sink : public sink {
+public:
+    wincolor_sink(void *out_handle, color_mode mode);
+    ~wincolor_sink() override;
+
+    wincolor_sink(const wincolor_sink &other) = delete;
+    wincolor_sink &operator=(const wincolor_sink &other) = delete;
+
+    // change the color for the given level
+    void set_color(level::level_enum level, std::uint16_t color);
+    void log(const details::log_msg &msg) final override;
+    void flush() final override;
+    void set_pattern(const std::string &pattern) override final;
+    void set_formatter(std::unique_ptr<spdlog::formatter> sink_formatter) override final;
+    void set_color_mode(color_mode mode);
+
+protected:
+    using mutex_t = typename ConsoleMutex::mutex_t;
+    void *out_handle_;
+    mutex_t &mutex_;
+    bool should_do_colors_;
+    std::unique_ptr<spdlog::formatter> formatter_;
+    std::array<std::uint16_t, level::n_levels> colors_;
+
+    // set foreground color and return the orig console attributes (for resetting later)
+    std::uint16_t set_foreground_color_(std::uint16_t attribs);
+
+    // print a range of formatted message to console
+    void print_range_(const memory_buf_t &formatted, size_t start, size_t end);
+
+    // in case we are redirected to file (not in console mode)
+    void write_to_file_(const memory_buf_t &formatted);
+
+    void set_color_mode_impl(color_mode mode);
+};
+
+template <typename ConsoleMutex>
+class wincolor_stdout_sink : public wincolor_sink<ConsoleMutex> {
+public:
+    explicit wincolor_stdout_sink(color_mode mode = color_mode::automatic);
+};
+
+template <typename ConsoleMutex>
+class wincolor_stderr_sink : public wincolor_sink<ConsoleMutex> {
+public:
+    explicit wincolor_stderr_sink(color_mode mode = color_mode::automatic);
+};
+
+using wincolor_stdout_sink_mt = wincolor_stdout_sink<details::console_mutex>;
+using wincolor_stdout_sink_st = wincolor_stdout_sink<details::console_nullmutex>;
+
+using wincolor_stderr_sink_mt = wincolor_stderr_sink<details::console_mutex>;
+using wincolor_stderr_sink_st = wincolor_stderr_sink<details::console_nullmutex>;
+}  // namespace sinks
+}  // namespace spdlog
+
+#ifdef SPDLOG_HEADER_ONLY
+    #include "wincolor_sink-inl.h"
+#endif
diff --git a/ext/spdlog/spdlog-inl.h b/ext/spdlog/spdlog-inl.h
new file mode 100644
index 0000000..97c3622
--- /dev/null
+++ b/ext/spdlog/spdlog-inl.h
@@ -0,0 +1,92 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#ifndef SPDLOG_HEADER_ONLY
+    #include <spdlog/spdlog.h>
+#endif
+
+#include <spdlog/common.h>
+#include <spdlog/pattern_formatter.h>
+
+namespace spdlog {
+
+SPDLOG_INLINE void initialize_logger(std::shared_ptr<logger> logger) {
+    details::registry::instance().initialize_logger(std::move(logger));
+}
+
+SPDLOG_INLINE std::shared_ptr<logger> get(const std::string &name) {
+    return details::registry::instance().get(name);
+}
+
+SPDLOG_INLINE void set_formatter(std::unique_ptr<spdlog::formatter> formatter) {
+    details::registry::instance().set_formatter(std::move(formatter));
+}
+
+SPDLOG_INLINE void set_pattern(std::string pattern, pattern_time_type time_type) {
+    set_formatter(
+        std::unique_ptr<spdlog::formatter>(new pattern_formatter(std::move(pattern), time_type)));
+}
+
+SPDLOG_INLINE void enable_backtrace(size_t n_messages) {
+    details::registry::instance().enable_backtrace(n_messages);
+}
+
+SPDLOG_INLINE void disable_backtrace() { details::registry::instance().disable_backtrace(); }
+
+SPDLOG_INLINE void dump_backtrace() { default_logger_raw()->dump_backtrace(); }
+
+SPDLOG_INLINE level::level_enum get_level() { return default_logger_raw()->level(); }
+
+SPDLOG_INLINE bool should_log(level::level_enum log_level) {
+    return default_logger_raw()->should_log(log_level);
+}
+
+SPDLOG_INLINE void set_level(level::level_enum log_level) {
+    details::registry::instance().set_level(log_level);
+}
+
+SPDLOG_INLINE void flush_on(level::level_enum log_level) {
+    details::registry::instance().flush_on(log_level);
+}
+
+SPDLOG_INLINE void set_error_handler(void (*handler)(const std::string &msg)) {
+    details::registry::instance().set_error_handler(handler);
+}
+
+SPDLOG_INLINE void register_logger(std::shared_ptr<logger> logger) {
+    details::registry::instance().register_logger(std::move(logger));
+}
+
+SPDLOG_INLINE void apply_all(const std::function<void(std::shared_ptr<logger>)> &fun) {
+    details::registry::instance().apply_all(fun);
+}
+
+SPDLOG_INLINE void drop(const std::string &name) { details::registry::instance().drop(name); }
+
+SPDLOG_INLINE void drop_all() { details::registry::instance().drop_all(); }
+
+SPDLOG_INLINE void shutdown() { details::registry::instance().shutdown(); }
+
+SPDLOG_INLINE void set_automatic_registration(bool automatic_registration) {
+    details::registry::instance().set_automatic_registration(automatic_registration);
+}
+
+SPDLOG_INLINE std::shared_ptr<spdlog::logger> default_logger() {
+    return details::registry::instance().default_logger();
+}
+
+SPDLOG_INLINE spdlog::logger *default_logger_raw() {
+    return details::registry::instance().get_default_raw();
+}
+
+SPDLOG_INLINE void set_default_logger(std::shared_ptr<spdlog::logger> default_logger) {
+    details::registry::instance().set_default_logger(std::move(default_logger));
+}
+
+SPDLOG_INLINE void apply_logger_env_levels(std::shared_ptr<logger> logger) {
+    details::registry::instance().apply_logger_env_levels(std::move(logger));
+}
+
+}  // namespace spdlog
diff --git a/ext/spdlog/spdlog.h b/ext/spdlog/spdlog.h
new file mode 100644
index 0000000..a8afbce
--- /dev/null
+++ b/ext/spdlog/spdlog.h
@@ -0,0 +1,352 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+// spdlog main header file.
+// see example.cpp for usage example
+
+#ifndef SPDLOG_H
+#define SPDLOG_H
+
+#pragma once
+
+#include <spdlog/common.h>
+#include <spdlog/details/registry.h>
+#include <spdlog/details/synchronous_factory.h>
+#include <spdlog/logger.h>
+#include <spdlog/version.h>
+
+#include <chrono>
+#include <functional>
+#include <memory>
+#include <string>
+
+namespace spdlog {
+
+using default_factory = synchronous_factory;
+
+// Create and register a logger with a templated sink type
+// The logger's level, formatter and flush level will be set according the
+// global settings.
+//
+// Example:
+//   spdlog::create<daily_file_sink_st>("logger_name", "dailylog_filename", 11, 59);
+template <typename Sink, typename... SinkArgs>
+inline std::shared_ptr<spdlog::logger> create(std::string logger_name, SinkArgs &&...sink_args) {
+    return default_factory::create<Sink>(std::move(logger_name),
+                                         std::forward<SinkArgs>(sink_args)...);
+}
+
+// Initialize and register a logger,
+// formatter and flush level will be set according the global settings.
+//
+// Useful for initializing manually created loggers with the global settings.
+//
+// Example:
+//   auto mylogger = std::make_shared<spdlog::logger>("mylogger", ...);
+//   spdlog::initialize_logger(mylogger);
+SPDLOG_API void initialize_logger(std::shared_ptr<logger> logger);
+
+// Return an existing logger or nullptr if a logger with such name doesn't
+// exist.
+// example: spdlog::get("my_logger")->info("hello {}", "world");
+SPDLOG_API std::shared_ptr<logger> get(const std::string &name);
+
+// Set global formatter. Each sink in each logger will get a clone of this object
+SPDLOG_API void set_formatter(std::unique_ptr<spdlog::formatter> formatter);
+
+// Set global format string.
+// example: spdlog::set_pattern("%Y-%m-%d %H:%M:%S.%e %l : %v");
+SPDLOG_API void set_pattern(std::string pattern,
+                            pattern_time_type time_type = pattern_time_type::local);
+
+// enable global backtrace support
+SPDLOG_API void enable_backtrace(size_t n_messages);
+
+// disable global backtrace support
+SPDLOG_API void disable_backtrace();
+
+// call dump backtrace on default logger
+SPDLOG_API void dump_backtrace();
+
+// Get global logging level
+SPDLOG_API level::level_enum get_level();
+
+// Set global logging level
+SPDLOG_API void set_level(level::level_enum log_level);
+
+// Determine whether the default logger should log messages with a certain level
+SPDLOG_API bool should_log(level::level_enum lvl);
+
+// Set global flush level
+SPDLOG_API void flush_on(level::level_enum log_level);
+
+// Start/Restart a periodic flusher thread
+// Warning: Use only if all your loggers are thread safe!
+template <typename Rep, typename Period>
+inline void flush_every(std::chrono::duration<Rep, Period> interval) {
+    details::registry::instance().flush_every(interval);
+}
+
+// Set global error handler
+SPDLOG_API void set_error_handler(void (*handler)(const std::string &msg));
+
+// Register the given logger with the given name
+SPDLOG_API void register_logger(std::shared_ptr<logger> logger);
+
+// Apply a user defined function on all registered loggers
+// Example:
+// spdlog::apply_all([&](std::shared_ptr<spdlog::logger> l) {l->flush();});
+SPDLOG_API void apply_all(const std::function<void(std::shared_ptr<logger>)> &fun);
+
+// Drop the reference to the given logger
+SPDLOG_API void drop(const std::string &name);
+
+// Drop all references from the registry
+SPDLOG_API void drop_all();
+
+// stop any running threads started by spdlog and clean registry loggers
+SPDLOG_API void shutdown();
+
+// Automatic registration of loggers when using spdlog::create() or spdlog::create_async
+SPDLOG_API void set_automatic_registration(bool automatic_registration);
+
+// API for using default logger (stdout_color_mt),
+// e.g: spdlog::info("Message {}", 1);
+//
+// The default logger object can be accessed using the spdlog::default_logger():
+// For example, to add another sink to it:
+// spdlog::default_logger()->sinks().push_back(some_sink);
+//
+// The default logger can replaced using spdlog::set_default_logger(new_logger).
+// For example, to replace it with a file logger.
+//
+// IMPORTANT:
+// The default API is thread safe (for _mt loggers), but:
+// set_default_logger() *should not* be used concurrently with the default API.
+// e.g do not call set_default_logger() from one thread while calling spdlog::info() from another.
+
+SPDLOG_API std::shared_ptr<spdlog::logger> default_logger();
+
+SPDLOG_API spdlog::logger *default_logger_raw();
+
+SPDLOG_API void set_default_logger(std::shared_ptr<spdlog::logger> default_logger);
+
+// Initialize logger level based on environment configs.
+//
+// Useful for applying SPDLOG_LEVEL to manually created loggers.
+//
+// Example:
+//   auto mylogger = std::make_shared<spdlog::logger>("mylogger", ...);
+//   spdlog::apply_logger_env_levels(mylogger);
+SPDLOG_API void apply_logger_env_levels(std::shared_ptr<logger> logger);
+
+template <typename... Args>
+inline void log(source_loc source,
+                level::level_enum lvl,
+                format_string_t<Args...> fmt,
+                Args &&...args) {
+    default_logger_raw()->log(source, lvl, fmt, std::forward<Args>(args)...);
+}
+
+template <typename... Args>
+inline void log(level::level_enum lvl, format_string_t<Args...> fmt, Args &&...args) {
+    default_logger_raw()->log(source_loc{}, lvl, fmt, std::forward<Args>(args)...);
+}
+
+template <typename... Args>
+inline void trace(format_string_t<Args...> fmt, Args &&...args) {
+    default_logger_raw()->trace(fmt, std::forward<Args>(args)...);
+}
+
+template <typename... Args>
+inline void debug(format_string_t<Args...> fmt, Args &&...args) {
+    default_logger_raw()->debug(fmt, std::forward<Args>(args)...);
+}
+
+template <typename... Args>
+inline void info(format_string_t<Args...> fmt, Args &&...args) {
+    default_logger_raw()->info(fmt, std::forward<Args>(args)...);
+}
+
+template <typename... Args>
+inline void warn(format_string_t<Args...> fmt, Args &&...args) {
+    default_logger_raw()->warn(fmt, std::forward<Args>(args)...);
+}
+
+template <typename... Args>
+inline void error(format_string_t<Args...> fmt, Args &&...args) {
+    default_logger_raw()->error(fmt, std::forward<Args>(args)...);
+}
+
+template <typename... Args>
+inline void critical(format_string_t<Args...> fmt, Args &&...args) {
+    default_logger_raw()->critical(fmt, std::forward<Args>(args)...);
+}
+
+template <typename T>
+inline void log(source_loc source, level::level_enum lvl, const T &msg) {
+    default_logger_raw()->log(source, lvl, msg);
+}
+
+template <typename T>
+inline void log(level::level_enum lvl, const T &msg) {
+    default_logger_raw()->log(lvl, msg);
+}
+
+#ifdef SPDLOG_WCHAR_TO_UTF8_SUPPORT
+template <typename... Args>
+inline void log(source_loc source,
+                level::level_enum lvl,
+                wformat_string_t<Args...> fmt,
+                Args &&...args) {
+    default_logger_raw()->log(source, lvl, fmt, std::forward<Args>(args)...);
+}
+
+template <typename... Args>
+inline void log(level::level_enum lvl, wformat_string_t<Args...> fmt, Args &&...args) {
+    default_logger_raw()->log(source_loc{}, lvl, fmt, std::forward<Args>(args)...);
+}
+
+template <typename... Args>
+inline void trace(wformat_string_t<Args...> fmt, Args &&...args) {
+    default_logger_raw()->trace(fmt, std::forward<Args>(args)...);
+}
+
+template <typename... Args>
+inline void debug(wformat_string_t<Args...> fmt, Args &&...args) {
+    default_logger_raw()->debug(fmt, std::forward<Args>(args)...);
+}
+
+template <typename... Args>
+inline void info(wformat_string_t<Args...> fmt, Args &&...args) {
+    default_logger_raw()->info(fmt, std::forward<Args>(args)...);
+}
+
+template <typename... Args>
+inline void warn(wformat_string_t<Args...> fmt, Args &&...args) {
+    default_logger_raw()->warn(fmt, std::forward<Args>(args)...);
+}
+
+template <typename... Args>
+inline void error(wformat_string_t<Args...> fmt, Args &&...args) {
+    default_logger_raw()->error(fmt, std::forward<Args>(args)...);
+}
+
+template <typename... Args>
+inline void critical(wformat_string_t<Args...> fmt, Args &&...args) {
+    default_logger_raw()->critical(fmt, std::forward<Args>(args)...);
+}
+#endif
+
+template <typename T>
+inline void trace(const T &msg) {
+    default_logger_raw()->trace(msg);
+}
+
+template <typename T>
+inline void debug(const T &msg) {
+    default_logger_raw()->debug(msg);
+}
+
+template <typename T>
+inline void info(const T &msg) {
+    default_logger_raw()->info(msg);
+}
+
+template <typename T>
+inline void warn(const T &msg) {
+    default_logger_raw()->warn(msg);
+}
+
+template <typename T>
+inline void error(const T &msg) {
+    default_logger_raw()->error(msg);
+}
+
+template <typename T>
+inline void critical(const T &msg) {
+    default_logger_raw()->critical(msg);
+}
+
+}  // namespace spdlog
+
+//
+// enable/disable log calls at compile time according to global level.
+//
+// define SPDLOG_ACTIVE_LEVEL to one of those (before including spdlog.h):
+// SPDLOG_LEVEL_TRACE,
+// SPDLOG_LEVEL_DEBUG,
+// SPDLOG_LEVEL_INFO,
+// SPDLOG_LEVEL_WARN,
+// SPDLOG_LEVEL_ERROR,
+// SPDLOG_LEVEL_CRITICAL,
+// SPDLOG_LEVEL_OFF
+//
+
+#ifndef SPDLOG_NO_SOURCE_LOC
+    #define SPDLOG_LOGGER_CALL(logger, level, ...) \
+        (logger)->log(spdlog::source_loc{__FILE__, __LINE__, SPDLOG_FUNCTION}, level, __VA_ARGS__)
+#else
+    #define SPDLOG_LOGGER_CALL(logger, level, ...) \
+        (logger)->log(spdlog::source_loc{}, level, __VA_ARGS__)
+#endif
+
+#if SPDLOG_ACTIVE_LEVEL <= SPDLOG_LEVEL_TRACE
+    #define SPDLOG_LOGGER_TRACE(logger, ...) \
+        SPDLOG_LOGGER_CALL(logger, spdlog::level::trace, __VA_ARGS__)
+    #define SPDLOG_TRACE(...) SPDLOG_LOGGER_TRACE(spdlog::default_logger_raw(), __VA_ARGS__)
+#else
+    #define SPDLOG_LOGGER_TRACE(logger, ...) (void)0
+    #define SPDLOG_TRACE(...) (void)0
+#endif
+
+#if SPDLOG_ACTIVE_LEVEL <= SPDLOG_LEVEL_DEBUG
+    #define SPDLOG_LOGGER_DEBUG(logger, ...) \
+        SPDLOG_LOGGER_CALL(logger, spdlog::level::debug, __VA_ARGS__)
+    #define SPDLOG_DEBUG(...) SPDLOG_LOGGER_DEBUG(spdlog::default_logger_raw(), __VA_ARGS__)
+#else
+    #define SPDLOG_LOGGER_DEBUG(logger, ...) (void)0
+    #define SPDLOG_DEBUG(...) (void)0
+#endif
+
+#if SPDLOG_ACTIVE_LEVEL <= SPDLOG_LEVEL_INFO
+    #define SPDLOG_LOGGER_INFO(logger, ...) \
+        SPDLOG_LOGGER_CALL(logger, spdlog::level::info, __VA_ARGS__)
+    #define SPDLOG_INFO(...) SPDLOG_LOGGER_INFO(spdlog::default_logger_raw(), __VA_ARGS__)
+#else
+    #define SPDLOG_LOGGER_INFO(logger, ...) (void)0
+    #define SPDLOG_INFO(...) (void)0
+#endif
+
+#if SPDLOG_ACTIVE_LEVEL <= SPDLOG_LEVEL_WARN
+    #define SPDLOG_LOGGER_WARN(logger, ...) \
+        SPDLOG_LOGGER_CALL(logger, spdlog::level::warn, __VA_ARGS__)
+    #define SPDLOG_WARN(...) SPDLOG_LOGGER_WARN(spdlog::default_logger_raw(), __VA_ARGS__)
+#else
+    #define SPDLOG_LOGGER_WARN(logger, ...) (void)0
+    #define SPDLOG_WARN(...) (void)0
+#endif
+
+#if SPDLOG_ACTIVE_LEVEL <= SPDLOG_LEVEL_ERROR
+    #define SPDLOG_LOGGER_ERROR(logger, ...) \
+        SPDLOG_LOGGER_CALL(logger, spdlog::level::err, __VA_ARGS__)
+    #define SPDLOG_ERROR(...) SPDLOG_LOGGER_ERROR(spdlog::default_logger_raw(), __VA_ARGS__)
+#else
+    #define SPDLOG_LOGGER_ERROR(logger, ...) (void)0
+    #define SPDLOG_ERROR(...) (void)0
+#endif
+
+#if SPDLOG_ACTIVE_LEVEL <= SPDLOG_LEVEL_CRITICAL
+    #define SPDLOG_LOGGER_CRITICAL(logger, ...) \
+        SPDLOG_LOGGER_CALL(logger, spdlog::level::critical, __VA_ARGS__)
+    #define SPDLOG_CRITICAL(...) SPDLOG_LOGGER_CRITICAL(spdlog::default_logger_raw(), __VA_ARGS__)
+#else
+    #define SPDLOG_LOGGER_CRITICAL(logger, ...) (void)0
+    #define SPDLOG_CRITICAL(...) (void)0
+#endif
+
+#ifdef SPDLOG_HEADER_ONLY
+    #include "spdlog-inl.h"
+#endif
+
+#endif  // SPDLOG_H
diff --git a/ext/spdlog/stopwatch.h b/ext/spdlog/stopwatch.h
new file mode 100644
index 0000000..54ab3d3
--- /dev/null
+++ b/ext/spdlog/stopwatch.h
@@ -0,0 +1,66 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#include <chrono>
+#include <spdlog/fmt/fmt.h>
+
+// Stopwatch support for spdlog  (using std::chrono::steady_clock).
+// Displays elapsed seconds since construction as double.
+//
+// Usage:
+//
+// spdlog::stopwatch sw;
+// ...
+// spdlog::debug("Elapsed: {} seconds", sw);    =>  "Elapsed 0.005116733 seconds"
+// spdlog::info("Elapsed: {:.6} seconds", sw);  =>  "Elapsed 0.005163 seconds"
+//
+//
+// If other units are needed (e.g. millis instead of double), include "fmt/chrono.h" and use
+// "duration_cast<..>(sw.elapsed())":
+//
+// #include <spdlog/fmt/chrono.h>
+//..
+// using std::chrono::duration_cast;
+// using std::chrono::milliseconds;
+// spdlog::info("Elapsed {}", duration_cast<milliseconds>(sw.elapsed())); => "Elapsed 5ms"
+
+namespace spdlog {
+class stopwatch {
+    using clock = std::chrono::steady_clock;
+    std::chrono::time_point<clock> start_tp_;
+
+public:
+    stopwatch()
+        : start_tp_{clock::now()} {}
+
+    std::chrono::duration<double> elapsed() const {
+        return std::chrono::duration<double>(clock::now() - start_tp_);
+    }
+
+    std::chrono::milliseconds elapsed_ms() const {
+        return std::chrono::duration_cast<std::chrono::milliseconds>(clock::now() - start_tp_);
+    }
+
+    void reset() { start_tp_ = clock::now(); }
+};
+}  // namespace spdlog
+
+// Support for fmt formatting  (e.g. "{:012.9}" or just "{}")
+namespace
+#ifdef SPDLOG_USE_STD_FORMAT
+    std
+#else
+    fmt
+#endif
+{
+
+template <>
+struct formatter<spdlog::stopwatch> : formatter<double> {
+    template <typename FormatContext>
+    auto format(const spdlog::stopwatch &sw, FormatContext &ctx) const -> decltype(ctx.out()) {
+        return formatter<double>::format(sw.elapsed().count(), ctx);
+    }
+};
+}  // namespace std
diff --git a/ext/spdlog/tweakme.h b/ext/spdlog/tweakme.h
new file mode 100644
index 0000000..a47a907
--- /dev/null
+++ b/ext/spdlog/tweakme.h
@@ -0,0 +1,141 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+///////////////////////////////////////////////////////////////////////////////
+//
+// Edit this file to squeeze more performance, and to customize supported
+// features
+//
+///////////////////////////////////////////////////////////////////////////////
+
+///////////////////////////////////////////////////////////////////////////////
+// Under Linux, the much faster CLOCK_REALTIME_COARSE clock can be used.
+// This clock is less accurate - can be off by dozens of millis - depending on
+// the kernel HZ.
+// Uncomment to use it instead of the regular clock.
+//
+// #define SPDLOG_CLOCK_COARSE
+///////////////////////////////////////////////////////////////////////////////
+
+///////////////////////////////////////////////////////////////////////////////
+// Uncomment if source location logging is not needed.
+// This will prevent spdlog from using __FILE__, __LINE__ and SPDLOG_FUNCTION
+//
+// #define SPDLOG_NO_SOURCE_LOC
+///////////////////////////////////////////////////////////////////////////////
+
+///////////////////////////////////////////////////////////////////////////////
+// Uncomment if thread id logging is not needed (i.e. no %t in the log pattern).
+// This will prevent spdlog from querying the thread id on each log call.
+//
+// WARNING: If the log pattern contains thread id (i.e, %t) while this flag is
+// on, zero will be logged as thread id.
+//
+// #define SPDLOG_NO_THREAD_ID
+///////////////////////////////////////////////////////////////////////////////
+
+///////////////////////////////////////////////////////////////////////////////
+// Uncomment to prevent spdlog from using thread local storage.
+//
+// WARNING: if your program forks, UNCOMMENT this flag to prevent undefined
+// thread ids in the children logs.
+//
+// #define SPDLOG_NO_TLS
+///////////////////////////////////////////////////////////////////////////////
+
+///////////////////////////////////////////////////////////////////////////////
+// Uncomment to avoid spdlog's usage of atomic log levels
+// Use only if your code never modifies a logger's log levels concurrently by
+// different threads.
+//
+// #define SPDLOG_NO_ATOMIC_LEVELS
+///////////////////////////////////////////////////////////////////////////////
+
+///////////////////////////////////////////////////////////////////////////////
+// Uncomment to enable usage of wchar_t for file names on Windows.
+//
+// #define SPDLOG_WCHAR_FILENAMES
+///////////////////////////////////////////////////////////////////////////////
+
+///////////////////////////////////////////////////////////////////////////////
+// Uncomment to override default eol ("\n" or "\r\n" under Linux/Windows)
+//
+// #define SPDLOG_EOL ";-)\n"
+///////////////////////////////////////////////////////////////////////////////
+
+///////////////////////////////////////////////////////////////////////////////
+// Uncomment to override default folder separators ("/" or "\\/" under
+// Linux/Windows). Each character in the string is treated as a different
+// separator.
+//
+// #define SPDLOG_FOLDER_SEPS "\\"
+///////////////////////////////////////////////////////////////////////////////
+
+///////////////////////////////////////////////////////////////////////////////
+// Uncomment to use your own copy of the fmt library instead of spdlog's copy.
+// In this case spdlog will try to include <fmt/format.h> so set your -I flag
+// accordingly.
+//
+// #define SPDLOG_FMT_EXTERNAL
+///////////////////////////////////////////////////////////////////////////////
+
+///////////////////////////////////////////////////////////////////////////////
+// Uncomment to use C++20 std::format instead of fmt.
+//
+// #define SPDLOG_USE_STD_FORMAT
+///////////////////////////////////////////////////////////////////////////////
+
+///////////////////////////////////////////////////////////////////////////////
+// Uncomment to enable wchar_t support (convert to utf8)
+//
+// #define SPDLOG_WCHAR_TO_UTF8_SUPPORT
+///////////////////////////////////////////////////////////////////////////////
+
+///////////////////////////////////////////////////////////////////////////////
+// Uncomment to prevent child processes from inheriting log file descriptors
+//
+// #define SPDLOG_PREVENT_CHILD_FD
+///////////////////////////////////////////////////////////////////////////////
+
+///////////////////////////////////////////////////////////////////////////////
+// Uncomment to customize level names (e.g. "MY TRACE")
+//
+// #define SPDLOG_LEVEL_NAMES { "MY TRACE", "MY DEBUG", "MY INFO", "MY WARNING", "MY ERROR", "MY
+// CRITICAL", "OFF" }
+///////////////////////////////////////////////////////////////////////////////
+
+///////////////////////////////////////////////////////////////////////////////
+// Uncomment to customize short level names (e.g. "MT")
+// These can be longer than one character.
+//
+// #define SPDLOG_SHORT_LEVEL_NAMES { "T", "D", "I", "W", "E", "C", "O" }
+///////////////////////////////////////////////////////////////////////////////
+
+///////////////////////////////////////////////////////////////////////////////
+// Uncomment to disable default logger creation.
+// This might save some (very) small initialization time if no default logger is needed.
+//
+// #define SPDLOG_DISABLE_DEFAULT_LOGGER
+///////////////////////////////////////////////////////////////////////////////
+
+///////////////////////////////////////////////////////////////////////////////
+// Uncomment and set to compile time level with zero cost (default is INFO).
+// Macros like SPDLOG_DEBUG(..), SPDLOG_INFO(..)  will expand to empty statements if not enabled
+//
+// #define SPDLOG_ACTIVE_LEVEL SPDLOG_LEVEL_INFO
+///////////////////////////////////////////////////////////////////////////////
+
+///////////////////////////////////////////////////////////////////////////////
+// Uncomment (and change if desired) macro to use for function names.
+// This is compiler dependent.
+// __PRETTY_FUNCTION__ might be nicer in clang/gcc, and __FUNCTION__ in msvc.
+// Defaults to __FUNCTION__ (should work on all compilers) if not defined.
+//
+// #ifdef __PRETTY_FUNCTION__
+// # define SPDLOG_FUNCTION __PRETTY_FUNCTION__
+// #else
+// # define SPDLOG_FUNCTION __FUNCTION__
+// #endif
+///////////////////////////////////////////////////////////////////////////////
diff --git a/ext/spdlog/version.h b/ext/spdlog/version.h
new file mode 100644
index 0000000..7c5e129
--- /dev/null
+++ b/ext/spdlog/version.h
@@ -0,0 +1,11 @@
+// Copyright(c) 2015-present, Gabi Melman & spdlog contributors.
+// Distributed under the MIT License (http://opensource.org/licenses/MIT)
+
+#pragma once
+
+#define SPDLOG_VER_MAJOR 1
+#define SPDLOG_VER_MINOR 15
+#define SPDLOG_VER_PATCH 0
+
+#define SPDLOG_TO_VERSION(major, minor, patch) (major * 10000 + minor * 100 + patch)
+#define SPDLOG_VERSION SPDLOG_TO_VERSION(SPDLOG_VER_MAJOR, SPDLOG_VER_MINOR, SPDLOG_VER_PATCH)
diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt
index e69de29..4a3a06e 100644
--- a/src/CMakeLists.txt
+++ b/src/CMakeLists.txt
@@ -0,0 +1,84 @@
+# exe path
+set(EXECUTABLE_OUTPUT_PATH "${PROJECT_BINARY_DIR}/bin")
+
+# source codes path
+aux_source_directory(${PROJECT_SOURCE_DIR}/src MAIN_SRC)
+aux_source_directory(${PROJECT_SOURCE_DIR}/src/util UTIL_SRC)
+aux_source_directory(${PROJECT_SOURCE_DIR}/src/markdup MARKDUP_SRC)
+
+set(KTHREAD_FILE ${PROJECT_SOURCE_DIR}/ext/klib/kthread.c)
+
+# including path
+include_directories("${PROJECT_SOURCE_DIR}/src")
+include_directories("${PROJECT_SOURCE_DIR}/ext")
+include_directories("${PROJECT_SOURCE_DIR}/ext/htslib")
+
+# linking path
+link_directories("${PROJECT_SOURCE_DIR}/ext/htslib")
+
+# program name
+set(PG_NAME "fastdup")
+
+# dependency files
+add_executable(${PG_NAME} ${MAIN_SRC} ${UTIL_SRC} ${MARKDUP_SRC} ${KTHREAD_FILE})
+
+# link htslib
+target_link_libraries(${PG_NAME} libhts.a)
+
+# lib pthread
+find_package(Threads REQUIRED)
+if(THREADS_HAVE_PTHREAD_ARG)
+    set_property(TARGET ${PG_NAME} PROPERTY COMPILE_OPTIONS "-pthread")
+    set_property(TARGET ${PG_NAME} PROPERTY INTERFACE_COMPILE_OPTIONS "-pthread")
+endif()
+if(CMAKE_THREAD_LIBS_INIT)
+    target_link_libraries(${PG_NAME} "${CMAKE_THREAD_LIBS_INIT}")
+endif()
+
+# lib bzip2
+find_package(BZip2 REQUIRED)
+if(BZip2_FOUND)
+    include_directories(${BZIP2_INCLUDE_DIR})
+    target_link_libraries(${PG_NAME} ${BZIP2_LIBRARIES})
+else()
+    message(FATAL_ERROR "BZIP2 is not found")
+endif()
+
+# lib deflate
+find_library(DeflateLib dlibdeflate)
+if(DeflateLib)
+    target_link_libraries(${PG_NAME} ${DeflateLib})
+else()
+    message(WARNING "libdeflate is not found")
+endif()
+
+# lib lzma
+find_package(LibLZMA MODULE)
+if (LibLZMA_FOUND)
+    include_directories(${LIBLZMA_INCLUDE_DIR})
+    target_link_libraries(${PG_NAME} ${LIBLZMA_LIBRARY})
+else()
+    message(FATAL_ERROR "lzma is not found")
+endif()
+
+# lz
+find_library(Z_LIBRARY z)
+if(Z_LIBRARY)
+    target_link_libraries(${PG_NAME} ${Z_LIBRARY})
+else()
+    message(FATAL_ERROR "lz is not found")
+endif()
+
+# lm
+find_library(MATH_LIBRARY m)
+if(MATH_LIBRARY)
+    target_link_libraries(${PG_NAME} ${MATH_LIBRARY})
+else()
+    message(FATAL_ERROR "lm is not found")
+endif()
+
+# install path
+install(TARGETS ${PG_NAME}
+    RUNTIME DESTINATION bin
+    LIBRARY DESTINATION lib
+    ARCHIVE DESTINATION libstatic)
\ No newline at end of file
diff --git a/src/main.cpp b/src/main.cpp
index e69de29..3b7b1fe 100644
--- a/src/main.cpp
+++ b/src/main.cpp
@@ -0,0 +1,25 @@
+#include <spdlog/cfg/env.h>
+#include <spdlog/sinks/stdout_color_sinks.h>
+#include <spdlog/spdlog.h>
+
+#include <argparse/argparse.hpp>
+
+#include "version.h"
+#include "markdup/md_args.h"
+
+using namespace std;
+
+extern MarkDupsArg gMdArg;
+
+int main(int argc, char *argv[]) {
+    // init log
+    spdlog::set_default_logger(spdlog::stderr_color_st("fastdup"));
+    spdlog::cfg::load_env_levels();
+    // init arg parser
+    argparse::ArgumentParser cliArgs()
+
+
+    spdlog::info("fast markduplicates start");
+
+    return 0;
+}
\ No newline at end of file
diff --git a/src/markdup/args.cpp b/src/markdup/args.cpp
deleted file mode 100644
index e69de29..0000000
diff --git a/src/markdup/args.h b/src/markdup/args.h
deleted file mode 100644
index e69de29..0000000
diff --git a/src/markdup/global_vars.h b/src/markdup/global_vars.h
index e69de29..7b9637e 100644
--- a/src/markdup/global_vars.h
+++ b/src/markdup/global_vars.h
@@ -0,0 +1 @@
+#pragma once
\ No newline at end of file
diff --git a/src/markdup/markdup.cpp b/src/markdup/markdup.cpp
index e69de29..c1af2d9 100644
--- a/src/markdup/markdup.cpp
+++ b/src/markdup/markdup.cpp
@@ -0,0 +1,10 @@
+/*
+Description:
+标记bam文件中的冗余信息，只处理按照坐标排序后的bam，且bam为单一样本数据
+
+Copyright : All right reserved by ICT
+
+Author : Zhang Zhonghai
+Date : 2023/10/23
+*/
+
diff --git a/src/markdup/md_args.cpp b/src/markdup/md_args.cpp
new file mode 100644
index 0000000..736fb84
--- /dev/null
+++ b/src/markdup/md_args.cpp
@@ -0,0 +1,26 @@
+/*
+Description: Markduplicate需要用到的一些参数，读取命令行给的参数，并做一些初始化
+
+Copyright : All right reserved by ICT
+
+Author : Zhang Zhonghai
+Date : 2023/10/27
+*/
+
+#include <cstring>
+#include <iostream>
+#include <sstream>
+#include <string>
+#include <vector>
+
+#include "md_args.h"
+using std::cout, std::endl;
+
+using std::ostringstream;
+using std::stod;
+using std::stoi;
+using std::stol;
+using std::string;
+using std::vector;
+
+using namespace ns_md;
\ No newline at end of file
diff --git a/src/markdup/md_args.h b/src/markdup/md_args.h
new file mode 100644
index 0000000..26b64d7
--- /dev/null
+++ b/src/markdup/md_args.h
@@ -0,0 +1,258 @@
+/*
+Description: Markduplicate需要用到的一些参数
+
+Copyright : All right reserved by ICT
+
+Author : Zhang Zhonghai
+Date : 2023/10/23
+*/
+#pragma once
+
+#include <string>
+#include <vector>
+
+using std::string;
+using std::vector;
+
+namespace ns_md {
+/* How strict to be when reading a SAM or BAM, beyond bare minimum validation.
+ */
+enum ValidationStringency {
+    /**
+     * Do the right thing, throw an exception if something looks wrong.
+     */
+    STRICT,
+    /**
+     * Emit warnings but keep going if possible.
+     */
+    LENIENT,
+    /**
+     * Like LENIENT, only don't emit warning messages.
+     */
+    SILENT,
+
+    DEFAULT_STRINGENCY = SILENT
+};
+
+/**
+ * Enum used to control how duplicates are flagged in the DT optional tag on
+ * each read.
+ */
+enum DuplicateTaggingPolicy { DontTag, OpticalOnly, All };
+
+/* 排序的方式 */
+enum SortOrder {
+    unsorted,
+    queryname,
+    coordinate,
+    duplicate,  // NB: this is not in the SAM spec!
+    unknown
+};
+
+/* 计算reads分数的方式（比那个read得分更高） */
+enum ScoringStrategy { SUM_OF_BASE_QUALITIES, TOTAL_MAPPED_REFERENCE_LENGTH, RANDOM };
+
+/* 索引文件的格式 （bai或者csi） */
+enum IndexFormat { BAI, CSI };
+}  // namespace ns_md
+
+/* markduplicate 需要的参数*/
+struct MarkDupsArg {
+    /**
+     * The optional attribute in SAM/BAM/CRAM files used to store the duplicate type.
+     */
+    string DUPLICATE_TYPE_TAG = "DT";
+    /**
+     * The duplicate type tag value for duplicate type: library.
+     */
+    string DUPLICATE_TYPE_LIBRARY = "LB";
+    /**
+     * The duplicate type tag value for duplicate type: sequencing (optical & pad-hopping, or "co-localized").
+     */
+    string DUPLICATE_TYPE_SEQUENCING = "SQ";
+    /**
+     * The attribute in the SAM/BAM file used to store which read was selected as representative out of a duplicate set
+     */
+    string DUPLICATE_SET_INDEX_TAG = "DI";
+    /**
+     * The attribute in the SAM/BAM file used to store the size of a duplicate set
+     */
+    string DUPLICATE_SET_SIZE_TAG = "DS";
+
+    /* OpticalDuplicateFinder */
+    int DEFAULT_OPTICAL_DUPLICATE_DISTANCE = 100;
+    int DEFAULT_BIG_DUPLICATE_SET_SIZE = 1000;
+    int DEFAULT_MAX_DUPLICATE_SET_SIZE =
+        300000;  // larger than this number will generate over 100 billion comparisons in the n^2 algorithm below
+
+    /**
+     * If more than this many sequences in SAM file, don't spill to disk because there will not
+     * be enough file handles.
+     */
+
+    /* "This option is obsolete. ReadEnds will always be spilled to disk." */
+    int MAX_SEQUENCES_FOR_DISK_READ_ENDS_MAP = 50000;
+
+    /* "Maximum number of file handles to keep open when spilling read ends to disk. " +
+                    "Set this number a little lower than the per-process maximum number of file that may be open. " +
+                    "This number can be found by executing the 'ulimit -n' command on a Unix system." */
+    int MAX_FILE_HANDLES_FOR_READ_ENDS_MAP = 8000;
+
+    /* "This number, plus the maximum RAM available to the JVM, determine the memory footprint used by " +
+                    "some of the sorting collections.  If you are running out of memory, try reducing this number." */
+    double SORTING_COLLECTION_SIZE_RATIO = 0.25;
+
+    /* "Barcode SAM tag (ex. BC for 10X Genomics)", optional = true */
+    string BARCODE_TAG = "";
+
+    /* "Read one barcode SAM tag (ex. BX for 10X Genomics)", optional = true */
+    string READ_ONE_BARCODE_TAG = "";
+
+    /* "Read two barcode SAM tag (ex. BX for 10X Genomics)", optional = true */
+    string READ_TWO_BARCODE_TAG = "";
+
+    /* "If a read appears in a duplicate set, add two tags. The first tag, DUPLICATE_SET_SIZE_TAG (DS), " +
+                     "indicates the size of the duplicate set. The smallest possible DS value is 2 which occurs when two
+       " + "reads map to the same portion of the reference only one of which is marked as duplicate. The second " +
+                     "tag, DUPLICATE_SET_INDEX_TAG (DI), represents a unique identifier for the duplicate set to which
+       the " + "record belongs. This identifier is the index-in-file of the representative read that was selected out "
+       + "of the duplicate set.", optional = true) */
+    bool TAG_DUPLICATE_SET_MEMBERS = false;
+
+    /* "If true remove 'optical' duplicates and other duplicates that appear to have arisen from the " +
+                    "sequencing process instead of the library preparation process, even if REMOVE_DUPLICATES is false.
+       " + "If REMOVE_DUPLICATES is true, all duplicates are removed and this option is ignored.") */
+    bool REMOVE_SEQUENCING_DUPLICATES = false;
+
+    /* "Determines how duplicate types are recorded in the DT optional attribute.") */
+    ns_md::DuplicateTaggingPolicy TAGGING_POLICY = ns_md::DuplicateTaggingPolicy::DontTag;
+
+    /* "Clear DT tag from input SAM records. Should be set to false if input SAM doesn't have this tag.  Default true")
+     */
+    bool CLEAR_DT = true;
+
+    /* "Treat UMIs as being duplex stranded.  This option requires that the UMI consist of two equal length " +
+                    "strings that are separated by a hyphen (e.g. 'ATC-GTC'). Reads are considered duplicates if, in
+       addition to standard " + "definition, have identical normalized UMIs.  A UMI from the 'bottom' strand is
+       normalized by swapping its content " + "around the hyphen (eg. ATC-GTC becomes GTC-ATC).  A UMI from the 'top'
+       strand is already normalized as it is. " + "Both reads from a read pair considered top strand if the read 1
+       unclipped 5' coordinate is less than the read " + "2 unclipped 5' coordinate. All chimeric reads and read
+       fragments are treated as having come from the top strand. " + "With this option is it required that the
+       BARCODE_TAG hold non-normalized UMIs. Default false.") */
+    bool DUPLEX_UMI = false;
+
+    /* "SAM tag to uniquely identify the molecule from which a read was derived.  Use of this option requires that " +
+                    "the BARCODE_TAG option be set to a non null value.  Default null.",
+              optional = true) */
+    string MOLECULAR_IDENTIFIER_TAG = "";
+
+    /* 继承自 AbstractMarkDuplicatesCommandLineProgram 的参数*/
+    /* "File to write duplication metrics to" */
+    string METRICS_FILE;
+
+    /* "If true do not write duplicates to the output file instead of writing them with appropriate flags set." */
+    bool REMOVE_DUPLICATES = false;
+
+    /* "If true, assume that the input file is coordinate sorted even if the header says otherwise. " +
+                    "Deprecated, used ASSUME_SORT_ORDER=coordinate instead." mutex = {"ASSUME_SORT_ORDER"} */
+    bool ASSUME_SORTED = false;
+
+    /* "If not null, assume that the input file has this order even if the header says otherwise.",
+              optional = true, mutex = {"ASSUME_SORTED"} */
+    ns_md::SortOrder ASSUME_SORT_ORDER = ns_md::SortOrder::unsorted;
+
+    /* "The scoring strategy for choosing the non-duplicate among candidates." */
+    ns_md::ScoringStrategy DUPLICATE_SCORING_STRATEGY = ns_md::ScoringStrategy::SUM_OF_BASE_QUALITIES;
+    // ns_md::ScoringStrategy DUPLICATE_SCORING_STRATEGY = ns_md::ScoringStrategy::TOTAL_MAPPED_REFERENCE_LENGTH;
+
+    /* "The program record ID for the @PG record(s) created by this program. Set to null to disable " +
+                    "PG record creation.  This string may have a suffix appended to avoid collision with other " +
+                    "program record IDs.",
+              optional = true */
+    string PROGRAM_RECORD_ID = "FastDup";
+
+    /* "Value of VN tag of PG record to be created. If not specified, the version will be detected automatically.",
+              optional = true */
+    string PROGRAM_GROUP_VERSION;
+
+    /* "Value of CL tag of PG record to be created. If not supplied the command line will be detected automatically.",
+              optional = true */
+    string PROGRAM_GROUP_COMMAND_LINE;
+
+    /* "Value of PN tag of PG record to be created." */
+    string PROGRAM_GROUP_NAME = "FastDup";
+
+    /* "Comment(s) to include in the output file's header.",
+              optional = true */
+    vector<string> COMMENT;
+
+    /* 继承自 AbstractOpticalDuplicateFinderCommandLineProgram 的参数 */
+
+    /* "MarkDuplicates can use the tile and cluster positions to estimate the rate of optical duplication " +
+            "in addition to the dominant source of duplication, PCR, to provide a more accurate estimation of library
+       size. " + "By default (with no READ_NAME_REGEX specified), MarkDuplicates will attempt to extract coordinates " +
+            "using a split on ':' (see Note below).  " +
+            "Set READ_NAME_REGEX to 'null' to disable optical duplicate detection. " +
+            "Note that without optical duplicate counts, library size estimation will be less accurate. " +
+            "If the read name does not follow a standard Illumina colon-separation convention, but does contain tile and
+       x,y coordinates, " + "a regular expression can be specified to extract three variables: tile/region, x coordinate
+       and y coordinate from a read name. " + "The regular expression must contain three capture groups for the three
+       variables, in order. " + "It must match the entire read name. " + "  e.g. if field names were separated by
+       semi-colon (';') this example regex could be specified " + " (?:.*;)?([0-9]+)[^;]*;([0-9]+)[^;]*;([0-9]+)[^;]*$ "
+       + "Note that if no READ_NAME_REGEX is specified, the read name is split on ':'. " + "  For 5 element names, the
+       3rd, 4th and 5th elements are assumed to be tile, x and y values. " + "  For 7 element names (CASAVA 1.8), the
+       5th, 6th, and 7th elements are assumed to be tile, x and y values.", optional = true */
+    string READ_NAME_REGEX = "(?:.*:)?([0-9]+)[^:]*:([0-9]+)[^:]*:([0-9]+)[^:]*$";
+
+    /* "The maximum offset between two duplicate clusters in order to consider them optical duplicates. The default " +
+                    "is appropriate for unpatterned versions of the Illumina platform. For the patterned flowcell
+       models, 2500 is more" + "appropriate. For other platforms and models, users should experiment to find what works
+       best." */
+    int OPTICAL_DUPLICATE_PIXEL_DISTANCE = DEFAULT_OPTICAL_DUPLICATE_DISTANCE;
+
+    /* "This number is the maximum size of a set of duplicate reads for which we will attempt to determine " +
+                    "which are optical duplicates.  Please be aware that if you raise this value too high and do
+       encounter a very " + "large set of duplicate reads, it will severely affect the runtime of this tool.  To
+       completely disable this check, " + "set the value to -1." */
+    long MAX_OPTICAL_DUPLICATE_SET_SIZE = DEFAULT_MAX_DUPLICATE_SET_SIZE;
+
+    /* 继承自 CommandLineProgram 的参数*/
+
+    /* "Whether to suppress job-summary info on System.err.", common = true */
+    bool QUIET = false;
+
+    /* "Validation stringency for all SAM files read by this program.  Setting stringency to SILENT " +
+            "can improve performance when processing a BAM file in which variable-length data (read, qualities, tags) "
+       + "do not otherwise need to be decoded.", common=true */
+    ns_md::ValidationStringency VALIDATION_STRINGENCY = ns_md::ValidationStringency::DEFAULT_STRINGENCY;
+
+    /* "Compression level for all compressed files created (e.g. BAM and VCF).", common = true */
+    int COMPRESSION_LEVEL = 5;
+
+    /* "When writing files that need to be sorted, this will specify the number of records stored in RAM before spilling
+       to disk. " + "Increasing this number reduces the number of file handles needed to sort the file, and increases
+       the amount of RAM needed.", optional = true, common = true */
+    int MAX_RECORDS_IN_RAM = 500000;
+
+    /* "Whether to create an index when writing VCF or coordinate sorted BAM output.", common = true */
+    bool CREATE_INDEX = false;
+
+    ns_md::IndexFormat INDEX_FORMAT = ns_md::IndexFormat::BAI;
+
+    /* "Whether to create an MD5 digest for any BAM or FASTQ files created.  ", common = true */
+    bool CREATE_MD5_FILE = false;
+
+    /* Add PG tag to each read in a SAM or BAM (PGTagArgumentCollection)*/
+    bool ADD_PG_TAG_TO_READS = true;
+
+    // 解析参数
+    void ParseArgument(int argc, char **argv);
+
+    void PrintArgValue();
+    string GetArgValueString();
+
+    static void PrintHelp();
+
+    static void PrintVersion();
+
+};
\ No newline at end of file
diff --git a/src/version.h b/src/version.h
new file mode 100644
index 0000000..c821e55
--- /dev/null
+++ b/src/version.h
@@ -0,0 +1,3 @@
+#pragma once
+
+#define FASTDUP_VERSION "1.0.0"
\ No newline at end of file