diff --git a/bseq.c b/bseq.c
index 54a25f6..d20b983 100644
--- a/bseq.c
+++ b/bseq.c
@@ -41,15 +41,15 @@ bseq1_t *bseq_read(int chunk_size, int *n_, void *ks1_, void *ks2_)
 		size += seqs[n++].l_seq;
 		if (ks2) {
 			trim_readno(&ks2->name);
-			kseq2bseq1(ks2, &seqs[n++]);
+			kseq2bseq1(ks2, &seqs[n]);
 			size += seqs[n++].l_seq;
 		}
 		if (size >= chunk_size) break;
 	}
-	*n_ = n;
-	if (size < chunk_size) { // test if the 2nd file is finished
+	if (size == 0) { // test if the 2nd file is finished
 		if (ks2 && kseq_read(ks2) >= 0)
 			fprintf(stderr, "[W::%s] the 1st file has fewer sequences.\n", __func__);
 	}
+	*n_ = n;
 	return seqs;
 }
diff --git a/bwamem.c b/bwamem.c
index b17d23b..f617fcf 100644
--- a/bwamem.c
+++ b/bwamem.c
@@ -644,21 +644,21 @@ int mem_process_seqs(const mem_opt_t *opt, const bwt_t *bwt, const bntseq_t *bns
 #ifdef HAVE_PTHREAD
 	if (opt->n_threads == 1) {
 		worker1(w);
-		mem_pestat(opt, bns->l_pac, n, regs, pes);
+		if (opt->is_pe) mem_pestat(opt, bns->l_pac, n, regs, pes);
 		worker2(w);
 	} else {
 		pthread_t *tid;
 		tid = (pthread_t*)calloc(opt->n_threads, sizeof(pthread_t));
 		for (i = 0; i < opt->n_threads; ++i) pthread_create(&tid[i], 0, worker1, &w[i]);
 		for (i = 0; i < opt->n_threads; ++i) pthread_join(tid[i], 0);
-		mem_pestat(opt, bns->l_pac, n, regs, pes);
+		if (opt->is_pe) mem_pestat(opt, bns->l_pac, n, regs, pes);
 		for (i = 0; i < opt->n_threads; ++i) pthread_create(&tid[i], 0, worker2, &w[i]);
 		for (i = 0; i < opt->n_threads; ++i) pthread_join(tid[i], 0);
 		free(tid);
 	}
 #else
 	worker1(w);
-	mem_pestat(opt, bns->l_pac, n, regs, pes);
+	if (opt->is_pe) mem_pestat(opt, bns->l_pac, n, regs, pes);
 	worker2(w);
 #endif
 	for (i = 0; i < n; ++i) {
diff --git a/fastmap.c b/fastmap.c
index 698b3e1..56674f9 100644
--- a/fastmap.c
+++ b/fastmap.c
@@ -58,7 +58,7 @@ int main_mem(int argc, char *argv[])
 	ks = kseq_init(fp);
 	if (optind + 2 < argc) {
 		fp2 = gzopen(argv[optind + 2], "r");
-		ks2 = kseq_init(fp);
+		ks2 = kseq_init(fp2);
 		opt->is_pe = 1;
 	}
 	while ((seqs = bseq_read(opt->n_threads * opt->chunk_size, &n, ks, ks2)) != 0) {