diff --git a/Makefile b/Makefile
index 32966c2..58cb711 100644
--- a/Makefile
+++ b/Makefile
@@ -14,7 +14,7 @@ OBJS=		utils.o bwt.o bwtio.o bwtaln.o bwtgap.o is.o \
 			bwtsw2_core.o bwtsw2_main.o bwtsw2_aux.o bwt_lite.o \
 			bwtsw2_chain.o bamlite.o bam2bam.o bgzf.o \
 			insert_size.o
-PROG=		bwa
+PROG=		network-aware-bwa
 INCLUDES=	
 LIBS=		-lm -lz -lpthread -Lbwt_gen -lbwtgen
 LIBS+=		`pkg-config --libs libzmq`
@@ -33,9 +33,9 @@ all:$(PROG)
 
 install: all
 	install -d         $(prefix)/bin
-	install -m 775 bwa $(prefix)/bin/
+	install -m 775 network-aware-bwa $(prefix)/bin/
 	install -d           $(prefix)/share/man/man1
-	install -m 644 bwa.1 $(prefix)/share/man/man1/
+	install -m 644 network-aware-bwa.1 $(prefix)/share/man/man1/
 
 lib-recur all-recur clean-recur cleanlocal-recur install-recur:
 		@target=`echo $@ | sed s/-recur//`; \
@@ -49,7 +49,7 @@ lib-recur all-recur clean-recur cleanlocal-recur install-recur:
 
 lib:
 
-bwa:lib-recur $(OBJS) main.o
+network-aware-bwa:lib-recur $(OBJS) main.o
 		$(CC) $(CFLAGS) $(DFLAGS) $(OBJS) main.o -o $@ $(LIBS)
 
 
diff --git a/bwt.c b/bwt.c
index 44a86b2..3e568d5 100644
--- a/bwt.c
+++ b/bwt.c
@@ -89,7 +89,7 @@ static inline int __occ_aux(uint64_t y, int c)
 	return ((y + (y >> 4)) & 0xf0f0f0f0f0f0f0full) * 0x101010101010101ull >> 56;
 }
 
-inline bwtint_t bwt_occ(const bwt_t *bwt, bwtint_t k, ubyte_t c)
+bwtint_t bwt_occ(const bwt_t *bwt, bwtint_t k, ubyte_t c)
 {
 	bwtint_t n, l, j;
 	uint32_t *p;
@@ -115,7 +115,7 @@ inline bwtint_t bwt_occ(const bwt_t *bwt, bwtint_t k, ubyte_t c)
 }
 
 // an analogy to bwt_occ() but more efficient, requiring k <= l
-inline void bwt_2occ(const bwt_t *bwt, bwtint_t k, bwtint_t l, ubyte_t c, bwtint_t *ok, bwtint_t *ol)
+void bwt_2occ(const bwt_t *bwt, bwtint_t k, bwtint_t l, ubyte_t c, bwtint_t *ok, bwtint_t *ol)
 {
 	bwtint_t _k, _l;
 	if (k == l) {
@@ -156,7 +156,7 @@ inline void bwt_2occ(const bwt_t *bwt, bwtint_t k, bwtint_t l, ubyte_t c, bwtint
 	((bwt)->cnt_table[(b)&0xff] + (bwt)->cnt_table[(b)>>8&0xff]		\
 	 + (bwt)->cnt_table[(b)>>16&0xff] + (bwt)->cnt_table[(b)>>24])
 
-inline void bwt_occ4(const bwt_t *bwt, bwtint_t k, bwtint_t cnt[4])
+void bwt_occ4(const bwt_t *bwt, bwtint_t k, bwtint_t cnt[4])
 {
 	bwtint_t l, j, x;
 	uint32_t *p;
@@ -176,7 +176,7 @@ inline void bwt_occ4(const bwt_t *bwt, bwtint_t k, bwtint_t cnt[4])
 }
 
 // an analogy to bwt_occ4() but more efficient, requiring k <= l
-inline void bwt_2occ4(const bwt_t *bwt, bwtint_t k, bwtint_t l, bwtint_t cntk[4], bwtint_t cntl[4])
+void bwt_2occ4(const bwt_t *bwt, bwtint_t k, bwtint_t l, bwtint_t cntk[4], bwtint_t cntl[4])
 {
 	bwtint_t _k, _l;
 	if (k == l) {
diff --git a/bwt.h b/bwt.h
index c9cc4ed..b99b951 100644
--- a/bwt.h
+++ b/bwt.h
@@ -95,14 +95,14 @@ extern "C" {
 
 	void bwt_bwtupdate_core(bwt_t *bwt);
 
-	inline bwtint_t bwt_occ(const bwt_t *bwt, bwtint_t k, ubyte_t c);
-	inline void bwt_occ4(const bwt_t *bwt, bwtint_t k, bwtint_t cnt[4]);
+	bwtint_t bwt_occ(const bwt_t *bwt, bwtint_t k, ubyte_t c);
+	void bwt_occ4(const bwt_t *bwt, bwtint_t k, bwtint_t cnt[4]);
 	bwtint_t bwt_sa(const bwt_t *bwt, bwtint_t k);
 
 	// more efficient version of bwt_occ/bwt_occ4 for retrieving two close Occ values
 	void bwt_gen_cnt_table(bwt_t *bwt);
-	inline void bwt_2occ(const bwt_t *bwt, bwtint_t k, bwtint_t l, ubyte_t c, bwtint_t *ok, bwtint_t *ol);
-	inline void bwt_2occ4(const bwt_t *bwt, bwtint_t k, bwtint_t l, bwtint_t cntk[4], bwtint_t cntl[4]);
+	void bwt_2occ(const bwt_t *bwt, bwtint_t k, bwtint_t l, ubyte_t c, bwtint_t *ok, bwtint_t *ol);
+	void bwt_2occ4(const bwt_t *bwt, bwtint_t k, bwtint_t l, bwtint_t cntk[4], bwtint_t cntl[4]);
 
 	int bwt_match_exact(const bwt_t *bwt, int len, const ubyte_t *str, bwtint_t *sa_begin, bwtint_t *sa_end);
 	int bwt_match_exact_alt(const bwt_t *bwt, int len, const ubyte_t *str, bwtint_t *k0, bwtint_t *l0);
diff --git a/bwt_lite.c b/bwt_lite.c
index dd411e1..3355f8d 100644
--- a/bwt_lite.c
+++ b/bwt_lite.c
@@ -52,7 +52,7 @@ bwtl_t *bwtl_seq2bwtl(int len, const uint8_t *seq)
 	}
 	return b;
 }
-inline uint32_t bwtl_occ(const bwtl_t *bwt, uint32_t k, uint8_t c)
+uint32_t bwtl_occ(const bwtl_t *bwt, uint32_t k, uint8_t c)
 {
 	uint32_t n, b;
 	if (k == bwt->seq_len) return bwt->L2[c+1] - bwt->L2[c];
@@ -65,7 +65,7 @@ inline uint32_t bwtl_occ(const bwtl_t *bwt, uint32_t k, uint8_t c)
 	if (c == 0) n -= 15 - (k&15); // corrected for the masked bits
 	return n;
 }
-inline void bwtl_occ4(const bwtl_t *bwt, uint32_t k, uint32_t cnt[4])
+void bwtl_occ4(const bwtl_t *bwt, uint32_t k, uint32_t cnt[4])
 {
 	uint32_t x, b;
 	if (k == (uint32_t)(-1)) {
@@ -80,7 +80,7 @@ inline void bwtl_occ4(const bwtl_t *bwt, uint32_t k, uint32_t cnt[4])
 	x -= 15 - (k&15);
 	cnt[0] += x&0xff; cnt[1] += x>>8&0xff; cnt[2] += x>>16&0xff; cnt[3] += x>>24;
 }
-inline void bwtl_2occ4(const bwtl_t *bwt, uint32_t k, uint32_t l, uint32_t cntk[4], uint32_t cntl[4])
+void bwtl_2occ4(const bwtl_t *bwt, uint32_t k, uint32_t l, uint32_t cntk[4], uint32_t cntl[4])
 {
 	bwtl_occ4(bwt, k, cntk);
 	bwtl_occ4(bwt, l, cntl);
diff --git a/bwt_lite.h b/bwt_lite.h
index 0096b93..4fadcce 100644
--- a/bwt_lite.h
+++ b/bwt_lite.h
@@ -17,9 +17,9 @@ extern "C" {
 #endif
 
 	bwtl_t *bwtl_seq2bwtl(int len, const uint8_t *seq);
-	inline uint32_t bwtl_occ(const bwtl_t *bwt, uint32_t k, uint8_t c);
-	inline void bwtl_occ4(const bwtl_t *bwt, uint32_t k, uint32_t cnt[4]);
-	inline void bwtl_2occ4(const bwtl_t *bwt, uint32_t k, uint32_t l, uint32_t cntk[4], uint32_t cntl[4]);
+	uint32_t bwtl_occ(const bwtl_t *bwt, uint32_t k, uint8_t c);
+	void bwtl_occ4(const bwtl_t *bwt, uint32_t k, uint32_t cnt[4]);
+	void bwtl_2occ4(const bwtl_t *bwt, uint32_t k, uint32_t l, uint32_t cntk[4], uint32_t cntl[4]);
 	void bwtl_destroy(bwtl_t *bwt);
 
 #ifdef __cplusplus