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

…functable

and remove SSE4.2 and ACLE optimizations. The functable overhead is higher
than the benefit from using optimized functions.

CMakeLists.txt

@@ -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

Makefile.in

@@ -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 \

README.md

@@ -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

arch/arm/Makefile.in

@@ -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 *~

arch/arm/arm_functions.h

@@ -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

arch/generic/Makefile.in

@@ -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
 

arch/generic/generic_functions.h

@@ -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

arch/x86/Makefile.in

@@ -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
 

arch/x86/x86_functions.h

@@ -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

configure

@@ -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
 

deflate.c

@@ -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;

deflate.h

@@ -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 */

deflate_fast.c

@@ -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.

deflate_medium.c

@@ -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;

deflate_quick.c

@@ -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) {