GerbilSoft/zlib-ng
1
Fork 0
mirror of https://github.com/GerbilSoft/zlib-ng.git synced 2025-06-18 11:35:35 -04:00
Code Issues Actions Packages Projects Releases Wiki Activity

Move update_hash(), insert_string() and quick_insert_string() out of functable

Browse Source 
and remove SSE4.2 and ACLE optimizations. The functable overhead is higher
than the benefit from using optimized functions.
This commit is contained in:
Hans Kristian Rosbach 2024-02-21 11:44:18 +01:00 committed by Hans Kristian Rosbach
parent ef2f8d528c
commit 9953f12e21
23 changed files with 44 additions and 155 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
CMakeLists.txt
View File

@ -653,7 +653,7 @@ if(WITH_OPTIM)
check_acle_compiler_flag()
if(HAVE_ACLE_FLAG)
add_definitions(-DARM_ACLE)
set(ACLE_SRCS ${ARCHDIR}/crc32_acle.c ${ARCHDIR}/insert_string_acle.c)
set(ACLE_SRCS ${ARCHDIR}/crc32_acle.c)
set_property(SOURCE ${ACLE_SRCS} PROPERTY COMPILE_FLAGS "${ACLEFLAG} ${NOLTOFLAG}")
list(APPEND ZLIB_ARCH_SRCS ${ACLE_SRCS})
add_feature_info(ACLE_CRC 1 "Support ACLE optimized CRC hash generation, using \"${ACLEFLAG}\"")
@ -857,8 +857,8 @@ if(WITH_OPTIM)
check_sse42_intrinsics()
if(HAVE_SSE42_INTRIN AND WITH_SSSE3)
add_definitions(-DX86_SSE42)
set(SSE42_SRCS ${ARCHDIR}/adler32_sse42.c ${ARCHDIR}/insert_string_sse42.c)
add_feature_info(SSE42_CRC 1 "Support SSE4.2 optimized CRC hash generation, using \"${SSE42FLAG}\"")
set(SSE42_SRCS ${ARCHDIR}/adler32_sse42.c)
add_feature_info(SSE42_CRC 1 "Support SSE4.2 optimized adler32 hash generation, using \"${SSE42FLAG}\"")
list(APPEND ZLIB_ARCH_SRCS ${SSE42_SRCS})
set_property(SOURCE ${SSE42_SRCS} PROPERTY COMPILE_FLAGS "${SSE42FLAG} ${NOLTOFLAG}")
else()
@ -1034,7 +1034,6 @@ set(ZLIB_SRCS
arch/generic/compare256_c.c
arch/generic/crc32_braid_c.c
arch/generic/crc32_fold_c.c
arch/generic/insert_string_c.c
arch/generic/slide_hash_c.c
adler32.c
compress.c
@ -1053,6 +1052,7 @@ set(ZLIB_SRCS
infback.c
inflate.c
inftrees.c
insert_string.c
insert_string_roll.c
trees.c
uncompr.c

4
Makefile.in
View File

@ -80,7 +80,6 @@ OBJZ = \
arch/generic/compare256_c.o \
arch/generic/crc32_braid_c.o \
arch/generic/crc32_fold_c.o \
arch/generic/insert_string_c.o \
arch/generic/slide_hash_c.o \
adler32.o \
compress.o \
@ -99,6 +98,7 @@ OBJZ = \
infback.o \
inflate.o \
inftrees.o \
insert_string.o \
insert_string_roll.o \
trees.o \
uncompr.o \
@ -120,7 +120,6 @@ PIC_OBJZ = \
arch/generic/compare256_c.lo \
arch/generic/crc32_braid_c.lo \
arch/generic/crc32_fold_c.lo \
arch/generic/insert_string_c.lo \
arch/generic/slide_hash_c.lo \
adler32.lo \
compress.lo \
@ -139,6 +138,7 @@ PIC_OBJZ = \
infback.lo \
inflate.lo \
inftrees.lo \
insert_string.lo \
insert_string_roll.lo \
trees.lo \
uncompr.lo \

1
README.md
View File

@ -21,7 +21,6 @@ Features
* Support for CPU intrinsics when available
* Adler32 implementation using SSSE3, AVX2, AVX512, AVX512-VNNI, Neon, VMX & VSX
* CRC32-B implementation using PCLMULQDQ, VPCLMULQDQ, ACLE, & IBM Z
* Hash table implementation using CRC32-C intrinsics on x86 and ARM
* Slide hash implementations using SSE2, AVX2, ARMv6, Neon, VMX & VSX
* Compare256 implementations using SSE2, AVX2, Neon, POWER9 & RVV
* Inflate chunk copying using SSE2, SSSE3, AVX, Neon & VSX

7
arch/arm/Makefile.in
View File

@ -25,7 +25,6 @@ all: \
crc32_acle.o crc32_acle.lo \
slide_hash_neon.o slide_hash_neon.lo \
slide_hash_armv6.o slide_hash_armv6.lo \
insert_string_acle.o insert_string_acle.lo
adler32_neon.o:
$(CC) $(CFLAGS) $(NEONFLAG) $(NOLTOFLAG) $(INCLUDES) -c -o $@ $(SRCDIR)/adler32_neon.c
@ -69,12 +68,6 @@ slide_hash_armv6.o:
slide_hash_armv6.lo:
$(CC) $(SFLAGS) $(ARMV6FLAG) $(NOLTOFLAG) $(INCLUDES) -c -o $@ $(SRCDIR)/slide_hash_armv6.c
insert_string_acle.o:
$(CC) $(CFLAGS) $(ACLEFLAG) $(NOLTOFLAG) $(INCLUDES) -c -o $@ $(SRCDIR)/insert_string_acle.c
insert_string_acle.lo:
$(CC) $(SFLAGS) $(ACLEFLAG) $(NOLTOFLAG) $(INCLUDES) -c -o $@ $(SRCDIR)/insert_string_acle.c
mostlyclean: clean
clean:
rm -f *.o *.lo *~

4
arch/arm/arm_functions.h
View File

@ -22,10 +22,6 @@ void inflate_fast_neon(PREFIX3(stream) *strm, uint32_t start);
#ifdef ARM_ACLE
uint32_t crc32_acle(uint32_t crc, const uint8_t *buf, size_t len);
void insert_string_acle(deflate_state *const s, const uint32_t str, uint32_t count);
Pos quick_insert_string_acle(deflate_state *const s, const uint32_t str);
uint32_t update_hash_acle(uint32_t h, uint32_t val);
#endif
#ifdef ARM_SIMD

24
arch/arm/insert_string_acle.c
View File

@ -1,24 +0,0 @@
/* insert_string_acle.c -- insert_string integer hash variant using ACLE's CRC instructions
*
* Copyright (C) 1995-2024 Jean-loup Gailly and Mark Adler
* For conditions of distribution and use, see copyright notice in zlib.h
*
*/
#ifdef ARM_ACLE
#include "acle_intrins.h"
#include "zbuild.h"
#include "deflate.h"
#define HASH_CALC(s, h, val) \
h = __crc32w(0, val)
#define HASH_CALC_VAR h
#define HASH_CALC_VAR_INIT uint32_t h = 0
#define UPDATE_HASH Z_TARGET_CRC update_hash_acle
#define INSERT_STRING Z_TARGET_CRC insert_string_acle
#define QUICK_INSERT_STRING Z_TARGET_CRC quick_insert_string_acle
#include "insert_string_tpl.h"
#endif

7
arch/generic/Makefile.in
View File

@ -19,7 +19,6 @@ all: \
compare256_c.o compare256_c.lo \
crc32_braid_c.o crc32_braid_c.lo \
crc32_fold_c.o crc32_fold_c.lo \
insert_string_c.o insert_string_c.lo \
slide_hash_c.o slide_hash_c.lo
@ -59,12 +58,6 @@ crc32_fold_c.o: $(SRCDIR)/crc32_fold_c.c $(SRCTOP)/zbuild.h $(SRCTOP)/functable
crc32_fold_c.lo: $(SRCDIR)/crc32_fold_c.c $(SRCTOP)/zbuild.h $(SRCTOP)/functable.h
$(CC) $(SFLAGS) $(INCLUDES) -c -o $@ $(SRCDIR)/crc32_fold_c.c
insert_string_c.o: $(SRCDIR)/insert_string_c.c $(SRCTOP)/zbuild.h $(SRCTOP)/deflate.h $(SRCTOP)/insert_string_tpl.h
$(CC) $(CFLAGS) $(INCLUDES) -c -o $@ $(SRCDIR)/insert_string_c.c
insert_string_c.lo: $(SRCDIR)/insert_string_c.c $(SRCTOP)/zbuild.h $(SRCTOP)/deflate.h $(SRCTOP)/insert_string_tpl.h
$(CC) $(SFLAGS) $(INCLUDES) -c -o $@ $(SRCDIR)/insert_string_c.c
slide_hash_c.o: $(SRCDIR)/slide_hash_c.c $(SRCTOP)/zbuild.h $(SRCTOP)/deflate.h
$(CC) $(CFLAGS) $(INCLUDES) -c -o $@ $(SRCDIR)/slide_hash_c.c

3
arch/generic/generic_functions.h
View File

@ -40,10 +40,7 @@ uint32_t compare256_unaligned_16(const uint8_t *src0, const uint8_t *src1);
typedef void (*slide_hash_func)(deflate_state *s);
void insert_string_c(deflate_state *const s, const uint32_t str, uint32_t count);
Pos quick_insert_string_c(deflate_state *const s, const uint32_t str);
void slide_hash_c(deflate_state *s);
uint32_t update_hash_c(deflate_state *const s, uint32_t h, uint32_t val);
uint32_t longest_match_c(deflate_state *const s, Pos cur_match);
# if defined(UNALIGNED_OK) && BYTE_ORDER == LITTLE_ENDIAN

7
arch/x86/Makefile.in
View File

@ -35,7 +35,6 @@ all: \
chunkset_ssse3.o chunkset_ssse3.lo \
compare256_avx2.o compare256_avx2.lo \
compare256_sse2.o compare256_sse2.lo \
insert_string_sse42.o insert_string_sse42.lo \
crc32_pclmulqdq.o crc32_pclmulqdq.lo \
crc32_vpclmulqdq.o crc32_vpclmulqdq.lo \
slide_hash_avx2.o slide_hash_avx2.lo \
@ -77,12 +76,6 @@ compare256_sse2.o:
compare256_sse2.lo:
$(CC) $(SFLAGS) $(SSE2FLAG) $(NOLTOFLAG) -DPIC $(INCLUDES) -c -o $@ $(SRCDIR)/compare256_sse2.c
insert_string_sse42.o:
$(CC) $(CFLAGS) $(SSE42FLAG) $(NOLTOFLAG) $(INCLUDES) -c -o $@ $(SRCDIR)/insert_string_sse42.c
insert_string_sse42.lo:
$(CC) $(SFLAGS) $(SSE42FLAG) $(NOLTOFLAG) -DPIC $(INCLUDES) -c -o $@ $(SRCDIR)/insert_string_sse42.c
crc32_pclmulqdq.o:
$(CC) $(CFLAGS) $(PCLMULFLAG) $(SSE42FLAG) $(NOLTOFLAG) $(INCLUDES) -c -o $@ $(SRCDIR)/crc32_pclmulqdq.c

24
arch/x86/insert_string_sse42.c
View File

@ -1,24 +0,0 @@
/* insert_string_sse42.c -- insert_string integer hash variant using SSE4.2's CRC instructions
*
* Copyright (C) 1995-2024 Jean-loup Gailly and Mark Adler
* For conditions of distribution and use, see copyright notice in zlib.h
*
*/
#ifdef X86_SSE42
#include "zbuild.h"
#include <nmmintrin.h>
#include "deflate.h"
#define HASH_CALC(s, h, val)\
h = _mm_crc32_u32(h, val)
#define HASH_CALC_VAR h
#define HASH_CALC_VAR_INIT uint32_t h = 0
#define UPDATE_HASH update_hash_sse42
#define INSERT_STRING insert_string_sse42
#define QUICK_INSERT_STRING quick_insert_string_sse42
#include "insert_string_tpl.h"
#endif

3
arch/x86/x86_functions.h
View File

@ -27,9 +27,6 @@ void inflate_fast_ssse3(PREFIX3(stream) *strm, uint32_t start);
#ifdef X86_SSE42
uint32_t adler32_fold_copy_sse42(uint32_t adler, uint8_t *dst, const uint8_t *src, size_t len);
void insert_string_sse42(deflate_state *const s, const uint32_t str, uint32_t count);
Pos quick_insert_string_sse42(deflate_state *const s, const uint32_t str);
uint32_t update_hash_sse42(uint32_t h, uint32_t val);
#endif
#ifdef X86_AVX2

8
configure vendored
View File

@ -1503,8 +1503,8 @@ case "${ARCH}" in
if test ${HAVE_SSE42_INTRIN} -eq 1; then
CFLAGS="${CFLAGS} -DX86_SSE42"
SFLAGS="${SFLAGS} -DX86_SSE42"
ARCH_STATIC_OBJS="${ARCH_STATIC_OBJS} adler32_sse42.o insert_string_sse42.o"
ARCH_SHARED_OBJS="${ARCH_SHARED_OBJS} adler32_sse42.lo insert_string_sse42.lo"
ARCH_STATIC_OBJS="${ARCH_STATIC_OBJS} adler32_sse42.o"
ARCH_SHARED_OBJS="${ARCH_SHARED_OBJS} adler32_sse42.lo"
fi
check_pclmulqdq_intrinsics
@ -1695,8 +1695,8 @@ EOF
CFLAGS="${CFLAGS} -DARM_ACLE"
SFLAGS="${SFLAGS} -DARM_ACLE"
ARCH_STATIC_OBJS="${ARCH_STATIC_OBJS} crc32_acle.o insert_string_acle.o"
ARCH_SHARED_OBJS="${ARCH_SHARED_OBJS} crc32_acle.lo insert_string_acle.lo"
ARCH_STATIC_OBJS="${ARCH_STATIC_OBJS} crc32_acle.o"
ARCH_SHARED_OBJS="${ARCH_SHARED_OBJS} crc32_acle.lo"
fi
fi

10
deflate.c
View File

@ -120,10 +120,6 @@ static void lm_set_level (deflate_state *s, int level);
static void lm_init (deflate_state *s);
Z_INTERNAL unsigned read_buf (PREFIX3(stream) *strm, unsigned char *buf, unsigned size);
extern uint32_t update_hash_roll (uint32_t h, uint32_t val);
extern void insert_string_roll (deflate_state *const s, uint32_t str, uint32_t count);
extern Pos quick_insert_string_roll(deflate_state *const s, uint32_t str);
/* ===========================================================================
* Local data
*/
@ -1144,9 +1140,9 @@ static void lm_set_level(deflate_state *s, int level) {
s->insert_string = &insert_string_roll;
s->quick_insert_string = &quick_insert_string_roll;
} else {
s->update_hash = functable.update_hash;
s->insert_string = functable.insert_string;
s->quick_insert_string = functable.quick_insert_string;
s->update_hash = update_hash;
s->insert_string = insert_string;
s->quick_insert_string = quick_insert_string;
}
s->level = level;

8
deflate.h
View File

@ -117,6 +117,14 @@ typedef uint32_t (* update_hash_cb) (uint32_t h, uint32_t val);
typedef void (* insert_string_cb) (deflate_state *const s, uint32_t str, uint32_t count);
typedef Pos (* quick_insert_string_cb)(deflate_state *const s, uint32_t str);
uint32_t update_hash (uint32_t h, uint32_t val);
void insert_string (deflate_state *const s, uint32_t str, uint32_t count);
Pos quick_insert_string (deflate_state *const s, uint32_t str);
uint32_t update_hash_roll (uint32_t h, uint32_t val);
void insert_string_roll (deflate_state *const s, uint32_t str, uint32_t count);
Pos quick_insert_string_roll(deflate_state *const s, uint32_t str);
struct ALIGNED_(16) internal_state {
PREFIX3(stream) *strm; /* pointer back to this zlib stream */
unsigned char *pending_buf; /* output still pending */

6
deflate_fast.c
View File

@ -41,7 +41,7 @@ Z_INTERNAL block_state deflate_fast(deflate_state *s, int flush) {
* dictionary, and set hash_head to the head of the hash chain:
*/
if (s->lookahead >= WANT_MIN_MATCH) {
hash_head = functable.quick_insert_string(s, s->strstart);
hash_head = quick_insert_string(s, s->strstart);
dist = (int64_t)s->strstart - hash_head;
/* Find the longest match, discarding those <= prev_length.
@ -71,11 +71,11 @@ Z_INTERNAL block_state deflate_fast(deflate_state *s, int flush) {
match_len--; /* string at strstart already in table */
s->strstart++;
functable.insert_string(s, s->strstart, match_len);
insert_string(s, s->strstart, match_len);
s->strstart += match_len;
} else {
s->strstart += match_len;
functable.quick_insert_string(s, s->strstart + 2 - STD_MIN_MATCH);
quick_insert_string(s, s->strstart + 2 - STD_MIN_MATCH);
/* If lookahead < STD_MIN_MATCH, ins_h is garbage, but it does not
* matter since it will be recomputed at next deflate call.

16
deflate_medium.c
View File

@ -52,9 +52,9 @@ static void insert_match(deflate_state *s, struct match match) {
if (UNLIKELY(match.match_length > 0)) {
if (match.strstart >= match.orgstart) {
if (match.strstart + match.match_length - 1 >= match.orgstart) {
functable.insert_string(s, match.strstart, match.match_length);
insert_string(s, match.strstart, match.match_length);
} else {
functable.insert_string(s, match.strstart, match.orgstart - match.strstart + 1);
insert_string(s, match.strstart, match.orgstart - match.strstart + 1);
}
match.strstart += match.match_length;
match.match_length = 0;
@ -72,12 +72,12 @@ static void insert_match(deflate_state *s, struct match match) {
if (LIKELY(match.strstart >= match.orgstart)) {
if (LIKELY(match.strstart + match.match_length - 1 >= match.orgstart)) {
functable.insert_string(s, match.strstart, match.match_length);
insert_string(s, match.strstart, match.match_length);
} else {
functable.insert_string(s, match.strstart, match.orgstart - match.strstart + 1);
insert_string(s, match.strstart, match.orgstart - match.strstart + 1);
}
} else if (match.orgstart < match.strstart + match.match_length) {
functable.insert_string(s, match.orgstart, match.strstart + match.match_length - match.orgstart);
insert_string(s, match.orgstart, match.strstart + match.match_length - match.orgstart);
}
match.strstart += match.match_length;
match.match_length = 0;
@ -86,7 +86,7 @@ static void insert_match(deflate_state *s, struct match match) {
match.match_length = 0;
if (match.strstart >= (STD_MIN_MATCH - 2))
functable.quick_insert_string(s, match.strstart + 2 - STD_MIN_MATCH);
quick_insert_string(s, match.strstart + 2 - STD_MIN_MATCH);
/* If lookahead < WANT_MIN_MATCH, ins_h is garbage, but it does not
* matter since it will be recomputed at next deflate call.
@ -199,7 +199,7 @@ Z_INTERNAL block_state deflate_medium(deflate_state *s, int flush) {
} else {
hash_head = 0;
if (s->lookahead >= WANT_MIN_MATCH) {
hash_head = functable.quick_insert_string(s, s->strstart);
hash_head = quick_insert_string(s, s->strstart);
}
current_match.strstart = (uint16_t)s->strstart;
@ -235,7 +235,7 @@ Z_INTERNAL block_state deflate_medium(deflate_state *s, int flush) {
/* now, look ahead one */
if (LIKELY(!early_exit && s->lookahead > MIN_LOOKAHEAD && (uint32_t)(current_match.strstart + current_match.match_length) < (s->window_size - MIN_LOOKAHEAD))) {
s->strstart = current_match.strstart + current_match.match_length;
hash_head = functable.quick_insert_string(s, s->strstart);
hash_head = quick_insert_string(s, s->strstart);
next_match.strstart = (uint16_t)s->strstart;
next_match.orgstart = next_match.strstart;

2
deflate_quick.c
View File

@ -86,7 +86,7 @@ Z_INTERNAL block_state deflate_quick(deflate_state *s, int flush) {
}
if (LIKELY(s->lookahead >= WANT_MIN_MATCH)) {
hash_head = functable.quick_insert_string(s, s->strstart);
hash_head = quick_insert_string(s, s->strstart);
dist = (int64_t)s->strstart - hash_head;
if (dist <= MAX_DIST(s) && dist > 0) {

30
functable.c
View File

@ -58,10 +58,7 @@ static void init_functable(void) {
ft.crc32_fold_final = &crc32_fold_final_c;
ft.crc32_fold_reset = &crc32_fold_reset_c;
ft.inflate_fast = &inflate_fast_c;
ft.insert_string = &insert_string_c;
ft.quick_insert_string = &quick_insert_string_c;
ft.slide_hash = &slide_hash_c;
ft.update_hash = &update_hash_c;
ft.longest_match = &longest_match_generic;
ft.longest_match_slow = &longest_match_slow_generic;
ft.compare256 = &compare256_generic;
@ -97,9 +94,6 @@ static void init_functable(void) {
#ifdef X86_SSE42
if (cf.x86.has_sse42) {
ft.adler32_fold_copy = &adler32_fold_copy_sse42;
ft.insert_string = &insert_string_sse42;
ft.quick_insert_string = &quick_insert_string_sse42;
ft.update_hash = &update_hash_sse42;
}
#endif
// X86 - PCLMUL
@ -183,9 +177,6 @@ static void init_functable(void) {
#ifdef ARM_ACLE
if (cf.arm.has_crc32) {
ft.crc32 = &crc32_acle;
ft.insert_string = &insert_string_acle;
ft.quick_insert_string = &quick_insert_string_acle;
ft.update_hash = &update_hash_acle;
}
#endif
@ -256,12 +247,9 @@ static void init_functable(void) {
FUNCTABLE_ASSIGN(ft, crc32_fold_final);
FUNCTABLE_ASSIGN(ft, crc32_fold_reset);
FUNCTABLE_ASSIGN(ft, inflate_fast);
FUNCTABLE_ASSIGN(ft, insert_string);
FUNCTABLE_ASSIGN(ft, longest_match);
FUNCTABLE_ASSIGN(ft, longest_match_slow);
FUNCTABLE_ASSIGN(ft, quick_insert_string);
FUNCTABLE_ASSIGN(ft, slide_hash);
FUNCTABLE_ASSIGN(ft, update_hash);
// Memory barrier for weak memory order CPUs
FUNCTABLE_BARRIER();
@ -327,11 +315,6 @@ static void inflate_fast_stub(PREFIX3(stream) *strm, uint32_t start) {
functable.inflate_fast(strm, start);
}
static void insert_string_stub(deflate_state* const s, uint32_t str, uint32_t count) {
init_functable();
functable.insert_string(s, str, count);
}
static uint32_t longest_match_stub(deflate_state* const s, Pos cur_match) {
init_functable();
return functable.longest_match(s, cur_match);
@ -342,21 +325,11 @@ static uint32_t longest_match_slow_stub(deflate_state* const s, Pos cur_match) {
return functable.longest_match_slow(s, cur_match);
}
static Pos quick_insert_string_stub(deflate_state* const s, const uint32_t str) {
init_functable();
return functable.quick_insert_string(s, str);
}
static void slide_hash_stub(deflate_state* s) {
init_functable();
functable.slide_hash(s);
}
static uint32_t update_hash_stub(uint32_t h, uint32_t val) {
init_functable();
return functable.update_hash(h, val);
}
/* functable init */
Z_INTERNAL struct functable_s functable = {
force_init_stub,
@ -371,10 +344,7 @@ Z_INTERNAL struct functable_s functable = {
crc32_fold_final_stub,
crc32_fold_reset_stub,
inflate_fast_stub,
insert_string_stub,
longest_match_stub,
longest_match_slow_stub,
quick_insert_string_stub,
slide_hash_stub,
update_hash_stub
};

3
functable.h
View File

@ -22,12 +22,9 @@ struct functable_s {
uint32_t (* crc32_fold_final) (struct crc32_fold_s *crc);
uint32_t (* crc32_fold_reset) (struct crc32_fold_s *crc);
void (* inflate_fast) (PREFIX3(stream) *strm, uint32_t start);
void (* insert_string) (deflate_state *const s, uint32_t str, uint32_t count);
uint32_t (* longest_match) (deflate_state *const s, Pos cur_match);
uint32_t (* longest_match_slow) (deflate_state *const s, Pos cur_match);
Pos (* quick_insert_string)(deflate_state *const s, uint32_t str);
void (* slide_hash) (deflate_state *s);
uint32_t (* update_hash) (uint32_t h, uint32_t val);
};
Z_INTERNAL extern struct functable_s functable;

6
arch/generic/insert_string_c.c → insert_string.c
View File

@ -14,8 +14,8 @@
#define HASH_CALC_VAR h
#define HASH_CALC_VAR_INIT uint32_t h = 0
#define UPDATE_HASH update_hash_c
#define INSERT_STRING insert_string_c
#define QUICK_INSERT_STRING quick_insert_string_c
#define UPDATE_HASH update_hash
#define INSERT_STRING insert_string
#define QUICK_INSERT_STRING quick_insert_string
#include "insert_string_tpl.h"

6
win32/Makefile.a64
View File

@ -67,7 +67,7 @@ OBJS = \
infback.obj \
inflate.obj \
inftrees.obj \
insert_string_c.obj \
insert_string.obj \
insert_string_roll.obj \
slide_hash_c.obj \
trees.obj \
@ -99,7 +99,7 @@ WFLAGS = $(WFLAGS) \
-DARM_NEON \
-DARM_NOCHECK_NEON \
#
OBJS = $(OBJS) crc32_acle.obj insert_string_acle.obj adler32_neon.obj chunkset_neon.obj compare256_neon.obj slide_hash_neon.obj
OBJS = $(OBJS) crc32_acle.obj adler32_neon.obj chunkset_neon.obj compare256_neon.obj slide_hash_neon.obj
# targets
all: $(STATICLIB) $(SHAREDLIB) $(IMPLIB) \
@ -208,7 +208,7 @@ deflate_stored.obj: $(SRCDIR)/deflate_stored.c $(SRCDIR)/zbuild.h $(SRCDIR)/defl
infback.obj: $(SRCDIR)/infback.c $(SRCDIR)/zbuild.h $(SRCDIR)/zutil.h $(SRCDIR)/inftrees.h $(SRCDIR)/inflate.h $(SRCDIR)/inflate_p.h $(SRCDIR)/functable.h
inflate.obj: $(SRCDIR)/inflate.c $(SRCDIR)/zbuild.h $(SRCDIR)/zutil.h $(SRCDIR)/inftrees.h $(SRCDIR)/inflate.h $(SRCDIR)/inflate_p.h $(SRCDIR)/functable.h $(SRCDIR)/functable.h
inftrees.obj: $(SRCDIR)/inftrees.c $(SRCDIR)/zbuild.h $(SRCDIR)/zutil.h $(SRCDIR)/inftrees.h
insert_string_c.obj: $(SRCDIR)/arch/generic/insert_string_c.c $(SRCDIR)/zbuild.h $(SRCDIR)/deflate.h $(SRCDIR)/insert_string_tpl.h
insert_string.obj: $(SRCDIR)/insert_string.c $(SRCDIR)/zbuild.h $(SRCDIR)/deflate.h $(SRCDIR)/insert_string_tpl.h
slide_hash_c.obj: $(SRCDIR)/arch/generic/slide_hash_c.c $(SRCDIR)/zbuild.h $(SRCDIR)/deflate.h
slide_hash_neon.obj: $(SRCDIR)/arch/arm/slide_hash_neon.c $(SRCDIR)/zbuild.h $(SRCDIR)/deflate.h
trees.obj: $(SRCDIR)/zbuild.h $(SRCDIR)/deflate.h $(SRCDIR)/trees_tbl.h

6
win32/Makefile.arm
View File

@ -72,7 +72,7 @@ OBJS = \
infback.obj \
inflate.obj \
inftrees.obj \
insert_string_c.obj \
insert_string.obj \
insert_string_roll.obj \
slide_hash_c.obj \
trees.obj \
@ -100,7 +100,7 @@ OBJS = $(OBJS) gzlib.obj gzread.obj gzwrite.obj
!if "$(WITH_ACLE)" != ""
WFLAGS = $(WFLAGS) -DARM_ACLE
OBJS = $(OBJS) crc32_acle.obj insert_string_acle.obj
OBJS = $(OBJS) crc32_acle.obj
!endif
!if "$(WITH_VFPV3)" != ""
NEON_ARCH = /arch:VFPv3
@ -229,7 +229,7 @@ deflate_stored.obj: $(SRCDIR)/deflate_stored.c $(SRCDIR)/zbuild.h $(SRCDIR)/defl
infback.obj: $(SRCDIR)/infback.c $(SRCDIR)/zbuild.h $(SRCDIR)/zutil.h $(SRCDIR)/inftrees.h $(SRCDIR)/inflate.h $(SRCDIR)/inflate_p.h $(SRCDIR)/functable.h
inflate.obj: $(SRCDIR)/inflate.c $(SRCDIR)/zbuild.h $(SRCDIR)/zutil.h $(SRCDIR)/inftrees.h $(SRCDIR)/inflate.h $(SRCDIR)/inflate_p.h $(SRCDIR)/functable.h $(SRCDIR)/functable.h
inftrees.obj: $(SRCDIR)/inftrees.c $(SRCDIR)/zbuild.h $(SRCDIR)/zutil.h $(SRCDIR)/inftrees.h
insert_string_c.obj: $(SRCDIR)/arch/generic/insert_string_c.c $(SRCDIR)/zbuild.h $(SRCDIR)/deflate.h $(SRCDIR)/insert_string_tpl.h
insert_string.obj: $(SRCDIR)/insert_string.c $(SRCDIR)/zbuild.h $(SRCDIR)/deflate.h $(SRCDIR)/insert_string_tpl.h
slide_hash_c.obj: $(SRCDIR)/arch/generic/slide_hash_c.c $(SRCDIR)/zbuild.h $(SRCDIR)/deflate.h
trees.obj: $(SRCDIR)/zbuild.h $(SRCDIR)/deflate.h $(SRCDIR)/trees_tbl.h
zutil.obj: $(SRCDIR)/zbuild.h $(SRCDIR)/zutil.h $(SRCDIR)/zutil_p.h

6
win32/Makefile.msc
View File

@ -80,9 +80,8 @@ OBJS = \
infback.obj \
inflate.obj \
inftrees.obj \
insert_string_c.obj \
insert_string.obj \
insert_string_roll.obj \
insert_string_sse42.obj \
slide_hash_c.obj \
slide_hash_avx2.obj \
slide_hash_sse2.obj \
@ -231,9 +230,8 @@ deflate_stored.obj: $(SRCDIR)/deflate_stored.c $(SRCDIR)/zbuild.h $(SRCDIR)/defl
infback.obj: $(SRCDIR)/infback.c $(SRCDIR)/zbuild.h $(SRCDIR)/zutil.h $(SRCDIR)/inftrees.h $(SRCDIR)/inflate.h $(SRCDIR)/inflate_p.h $(SRCDIR)/functable.h
inflate.obj: $(SRCDIR)/inflate.c $(SRCDIR)/zbuild.h $(SRCDIR)/zutil.h $(SRCDIR)/inftrees.h $(SRCDIR)/inflate.h $(SRCDIR)/inflate_p.h $(SRCDIR)/functable.h $(SRCDIR)/functable.h $(SRCDIR)/functable.h
inftrees.obj: $(SRCDIR)/inftrees.c $(SRCDIR)/zbuild.h $(SRCDIR)/zutil.h $(SRCDIR)/inftrees.h
insert_string_c.obj: $(SRCDIR)/arch/generic/insert_string_c.c $(SRCDIR)/zbuild.h $(SRCDIR)/deflate.h $(SRCDIR)/insert_string_tpl.h
insert_string.obj: $(SRCDIR)/insert_string.c $(SRCDIR)/zbuild.h $(SRCDIR)/deflate.h $(SRCDIR)/insert_string_tpl.h
insert_string_roll.obj: $(SRCDIR)/insert_string_roll.c $(SRCDIR)/zbuild.h $(SRCDIR)/deflate.h $(SRCDIR)/insert_string_tpl.h
insert_string_sse42.obj: $(SRCDIR)/arch/x86/insert_string_sse42.c $(SRCDIR)/zbuild.h $(SRCDIR)/deflate.h $(SRCDIR)/insert_string_tpl.h
slide_hash_c.obj: $(SRCDIR)/arch/generic/slide_hash_c.c $(SRCDIR)/zbuild.h $(SRCDIR)/deflate.h
slide_hash_avx2.obj: $(SRCDIR)/arch/x86/slide_hash_avx2.c $(SRCDIR)/zbuild.h $(SRCDIR)/deflate.h
slide_hash_sse2.obj: $(SRCDIR)/arch/x86/slide_hash_sse2.c $(SRCDIR)/zbuild.h $(SRCDIR)/deflate.h