Split CPU features checks and CPU-specific function prototypes and re…

…duce include-dependencies.

Signed-off-by: Vladislav Shchapov <[email protected]>

CMakeLists.txt

@@ -560,7 +560,7 @@ endif()
 set(GENERIC_ARCHDIR "arch/generic")
 
 set(ZLIB_ARCH_SRCS)
-set(ZLIB_ARCH_HDRS ${GENERIC_ARCHDIR}/generic_features.h)
+set(ZLIB_ARCH_HDRS ${GENERIC_ARCHDIR}/generic_functions.h)
 
 if(BASEARCH_ARM_FOUND)
     set(ARCHDIR "arch/arm")
@@ -647,7 +647,7 @@ if(WITH_OPTIM)
                 endif()
             endif()
         endif()
-        list(APPEND ZLIB_ARCH_HDRS ${ARCHDIR}/arm_features.h)
+        list(APPEND ZLIB_ARCH_HDRS ${ARCHDIR}/arm_features.h ${ARCHDIR}/arm_functions.h)
         list(APPEND ZLIB_ARCH_SRCS ${ARCHDIR}/arm_features.c)
         if(WITH_ACLE)
             check_acle_compiler_flag()
@@ -719,7 +719,7 @@ if(WITH_OPTIM)
             add_definitions(-DPOWER_FEATURES)
         endif()
         if(HAVE_VMX OR HAVE_POWER8_INTRIN OR HAVE_POWER9_INTRIN)
-            list(APPEND ZLIB_ARCH_HDRS ${ARCHDIR}/power_features.h)
+            list(APPEND ZLIB_ARCH_HDRS ${ARCHDIR}/power_features.h ${ARCHDIR}/power_functions.h)
             list(APPEND ZLIB_ARCH_SRCS ${ARCHDIR}/power_features.c)
         endif()
         # VMX specific options and files
@@ -769,7 +769,7 @@ if(WITH_OPTIM)
             if(HAVE_RVV_INTRIN)
                 add_definitions(-DRISCV_FEATURES)
                 add_definitions(-DRISCV_RVV)
-                list(APPEND ZLIB_ARCH_HDRS ${ARCHDIR}/riscv_features.h)
+                list(APPEND ZLIB_ARCH_HDRS ${ARCHDIR}/riscv_features.h ${ARCHDIR}/riscv_functions.h)
                 list(APPEND ZLIB_ARCH_SRCS ${ARCHDIR}/riscv_features.c)
                 # FIXME: we will not set compile flags for riscv_features.c when
                 # the kernels update hwcap or hwprobe for riscv
@@ -784,7 +784,7 @@ if(WITH_OPTIM)
         check_s390_intrinsics()
         if(HAVE_S390_INTRIN)
             add_definitions(-DS390_FEATURES)
-            list(APPEND ZLIB_ARCH_HDRS ${ARCHDIR}/s390_features.h)
+            list(APPEND ZLIB_ARCH_HDRS ${ARCHDIR}/s390_features.h ${ARCHDIR}/s390_functions.h)
             list(APPEND ZLIB_ARCH_SRCS ${ARCHDIR}/s390_features.c)
         endif()
         if(WITH_DFLTCC_DEFLATE OR WITH_DFLTCC_INFLATE)
@@ -811,7 +811,7 @@ if(WITH_OPTIM)
         endif()
     elseif(BASEARCH_X86_FOUND)
         add_definitions(-DX86_FEATURES)
-        list(APPEND ZLIB_ARCH_HDRS ${ARCHDIR}/x86_features.h)
+        list(APPEND ZLIB_ARCH_HDRS ${ARCHDIR}/x86_features.h ${ARCHDIR}/x86_functions.h)
         list(APPEND ZLIB_ARCH_SRCS ${ARCHDIR}/x86_features.c)
         if(MSVC)
             list(APPEND ZLIB_ARCH_HDRS fallback_builtins.h)
@@ -1002,11 +1002,10 @@ set(ZLIB_PUBLIC_HDRS
     ${CMAKE_CURRENT_BINARY_DIR}/zlib${SUFFIX}.h
 )
 set(ZLIB_PRIVATE_HDRS
-    arch/generic/adler32_fold_c.h
-    arch/generic/crc32_fold_c.h
     adler32_p.h
     chunkset_tpl.h
     compare256_rle.h
+    cpu_functions.h
     cpu_features.h
     crc32_braid_p.h
     crc32_braid_comb_p.h

arch/arm/arm_features.h

@@ -2,8 +2,8 @@
  * For conditions of distribution and use, see copyright notice in zlib.h
  */
 
-#ifndef ARM_H_
-#define ARM_H_
+#ifndef ARM_FEATURES_H_
+#define ARM_FEATURES_H_
 
 struct arm_cpu_features {
     int has_simd;
@@ -13,44 +13,4 @@ struct arm_cpu_features {
 
 void Z_INTERNAL arm_check_features(struct arm_cpu_features *features);
 
-#ifdef CPU_FEATURES_H_
-
-#ifdef ARM_NEON
-uint32_t adler32_neon(uint32_t adler, const uint8_t *buf, size_t len);
-uint32_t chunksize_neon(void);
-uint8_t* chunkmemset_safe_neon(uint8_t *out, unsigned dist, unsigned len, unsigned left);
-
-#  ifdef HAVE_BUILTIN_CTZLL
-    uint32_t compare256_neon(const uint8_t *src0, const uint8_t *src1);
-#    ifdef DEFLATE_H_
-        uint32_t longest_match_neon(deflate_state *const s, Pos cur_match);
-        uint32_t longest_match_slow_neon(deflate_state *const s, Pos cur_match);
-#    endif
-#  endif
-#  ifdef DEFLATE_H_
-    void slide_hash_neon(deflate_state *s);
-#  endif
-#  ifdef INFLATE_H_
-    void inflate_fast_neon(PREFIX3(stream) *strm, uint32_t start);
-#  endif
-#endif
-
-#ifdef ARM_ACLE
-uint32_t crc32_acle(uint32_t crc, const uint8_t *buf, size_t len);
-
-#  ifdef DEFLATE_H_
-    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(deflate_state *const s, uint32_t h, uint32_t val);
-#  endif
-#endif
-
-#ifdef ARM_SIMD
-#  ifdef DEFLATE_H_
-    void slide_hash_armv6(deflate_state *s);
-#  endif
-#endif
-
-#endif
-
-#endif /* ARM_H_ */
+#endif /* ARM_FEATURES_H_ */

arch/arm/arm_functions.h

@@ -0,0 +1,35 @@
+/* arm_functions.h -- ARM implementations for arch-specific functions.
+ * For conditions of distribution and use, see copyright notice in zlib.h
+ */
+
+#ifndef ARM_FUNCTIONS_H_
+#define ARM_FUNCTIONS_H_
+
+
+#ifdef ARM_NEON
+uint32_t adler32_neon(uint32_t adler, const uint8_t *buf, size_t len);
+uint32_t chunksize_neon(void);
+uint8_t* chunkmemset_safe_neon(uint8_t *out, unsigned dist, unsigned len, unsigned left);
+
+#  ifdef HAVE_BUILTIN_CTZLL
+uint32_t compare256_neon(const uint8_t *src0, const uint8_t *src1);
+uint32_t longest_match_neon(deflate_state *const s, Pos cur_match);
+uint32_t longest_match_slow_neon(deflate_state *const s, Pos cur_match);
+#  endif
+void slide_hash_neon(deflate_state *s);
+void inflate_fast_neon(PREFIX3(stream) *strm, uint32_t start);
+#endif
+
+#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(deflate_state *const s, uint32_t h, uint32_t val);
+#endif
+
+#ifdef ARM_SIMD
+void slide_hash_armv6(deflate_state *s);
+#endif
+
+#endif /* ARM_FUNCTIONS_H_ */

arch/generic/Makefile.in

@@ -29,10 +29,10 @@ adler32_c.o: $(SRCDIR)/adler32_c.c  $(SRCTOP)/zbuild.h $(SRCTOP)/adler32_p.h
 adler32_c.lo: $(SRCDIR)/adler32_c.c  $(SRCTOP)/zbuild.h $(SRCTOP)/adler32_p.h
 	$(CC) $(SFLAGS) $(INCLUDES) -c -o $@ $(SRCDIR)/adler32_c.c
 
-adler32_fold_c.o: $(SRCDIR)/adler32_fold_c.c  $(SRCTOP)/zbuild.h $(SRCTOP)/functable.h $(SRCDIR)/adler32_fold_c.h
+adler32_fold_c.o: $(SRCDIR)/adler32_fold_c.c  $(SRCTOP)/zbuild.h $(SRCTOP)/functable.h
 	$(CC) $(CFLAGS) $(INCLUDES) -c -o $@ $(SRCDIR)/adler32_fold_c.c
 
-adler32_fold_c.lo: $(SRCDIR)/adler32_fold_c.c  $(SRCTOP)/zbuild.h $(SRCTOP)/functable.h $(SRCDIR)/adler32_fold_c.h
+adler32_fold_c.lo: $(SRCDIR)/adler32_fold_c.c  $(SRCTOP)/zbuild.h $(SRCTOP)/functable.h
 	$(CC) $(SFLAGS) $(INCLUDES) -c -o $@ $(SRCDIR)/adler32_fold_c.c
 
 chunkset_c.o: $(SRCDIR)/chunkset_c.c  $(SRCTOP)/zbuild.h $(SRCTOP)/chunkset_tpl.h $(SRCTOP)/inffast_tpl.h
@@ -53,10 +53,10 @@ crc32_braid_c.o: $(SRCDIR)/crc32_braid_c.c  $(SRCTOP)/zbuild.h $(SRCTOP)/crc32_b
 crc32_braid_c.lo: $(SRCDIR)/crc32_braid_c.c  $(SRCTOP)/zbuild.h $(SRCTOP)/crc32_braid_p.h $(SRCTOP)/crc32_braid_tbl.h
 	$(CC) $(SFLAGS) $(INCLUDES) -c -o $@ $(SRCDIR)/crc32_braid_c.c
 
-crc32_fold_c.o: $(SRCDIR)/crc32_fold_c.c  $(SRCTOP)/zbuild.h $(SRCTOP)/functable.h $(SRCDIR)/crc32_fold_c.h
+crc32_fold_c.o: $(SRCDIR)/crc32_fold_c.c  $(SRCTOP)/zbuild.h $(SRCTOP)/functable.h
 	$(CC) $(CFLAGS) $(INCLUDES) -c -o $@ $(SRCDIR)/crc32_fold_c.c
 
-crc32_fold_c.lo: $(SRCDIR)/crc32_fold_c.c  $(SRCTOP)/zbuild.h $(SRCTOP)/functable.h $(SRCDIR)/crc32_fold_c.h
+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

arch/generic/adler32_fold_c.c

@@ -5,7 +5,6 @@
 
 #include "zbuild.h"
 #include "functable.h"
-#include "adler32_fold_c.h"
 
 #include <limits.h>
 

arch/generic/crc32_fold_c.c

@@ -3,11 +3,10 @@
  * For conditions of distribution and use, see copyright notice in zlib.h
  */
 #include "zbuild.h"
+#include "zutil.h"
 #include "functable.h"
 #include "crc32.h"
 
-#include "crc32_fold_c.h"
-
 Z_INTERNAL uint32_t crc32_fold_reset_c(crc32_fold *crc) {
     crc->value = CRC32_INITIAL_VALUE;
     return crc->value;

arch/generic/generic_features.h → arch/generic/generic_functions.h

@@ -1,20 +1,29 @@
-/* generic_features.h -- generic C implementations for arch-specific features
+/* generic_functions.h -- generic C implementations for arch-specific functions.
  * For conditions of distribution and use, see copyright notice in zlib.h
  */
 
-#ifndef GENERIC_FEATURES_H_
-#define GENERIC_FEATURES_H_
+#ifndef GENERIC_FUNCTIONS_H_
+#define GENERIC_FUNCTIONS_H_
+
+#include "zendian.h"
+
+Z_INTERNAL uint32_t crc32_fold_reset_c(crc32_fold *crc);
+Z_INTERNAL void     crc32_fold_copy_c(crc32_fold *crc, uint8_t *dst, const uint8_t *src, size_t len);
+Z_INTERNAL void     crc32_fold_c(crc32_fold *crc, const uint8_t *src, size_t len, uint32_t init_crc);
+Z_INTERNAL uint32_t crc32_fold_final_c(crc32_fold *crc);
+
+Z_INTERNAL uint32_t adler32_fold_copy_c(uint32_t adler, uint8_t *dst, const uint8_t *src, size_t len);
+
 
 typedef uint32_t (*adler32_func)(uint32_t adler, const uint8_t *buf, size_t len);
 typedef uint32_t (*compare256_func)(const uint8_t *src0, const uint8_t *src1);
 typedef uint32_t (*crc32_func)(uint32_t crc32, const uint8_t *buf, size_t len);
 
 uint32_t adler32_c(uint32_t adler, const uint8_t *buf, size_t len);
+
 uint32_t chunksize_c(void);
 uint8_t* chunkmemset_safe_c(uint8_t *out, unsigned dist, unsigned len, unsigned left);
-#ifdef INFLATE_H_
 void     inflate_fast_c(PREFIX3(stream) *strm, uint32_t start);
-#endif
 
 uint32_t PREFIX(crc32_braid)(uint32_t crc, const uint8_t *buf, size_t len);
 
@@ -29,8 +38,7 @@ uint32_t compare256_unaligned_16(const uint8_t *src0, const uint8_t *src1);
 #  endif
 #endif
 
-#ifdef DEFLATE_H_
-typedef void    (*slide_hash_func)(deflate_state *s);
+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);
@@ -58,5 +66,3 @@ uint32_t longest_match_slow_c(deflate_state *const s, Pos cur_match);
 #  endif
 
 #endif
-
-#endif

arch/power/power_features.h

@@ -4,8 +4,8 @@
  * For conditions of distribution and use, see copyright notice in zlib.h
  */
 
-#ifndef POWER_H_
-#define POWER_H_
+#ifndef POWER_FEATURES_H_
+#define POWER_FEATURES_H_
 
 struct power_cpu_features {
     int has_altivec;
@@ -15,36 +15,4 @@ struct power_cpu_features {
 
 void Z_INTERNAL power_check_features(struct power_cpu_features *features);
 
-#ifdef CPU_FEATURES_H_
-
-#ifdef PPC_VMX
-uint32_t adler32_vmx(uint32_t adler, const uint8_t *buf, size_t len);
-#  ifdef DEFLATE_H_
-    void slide_hash_vmx(deflate_state *s);
-#  endif
-#endif
-
-#ifdef POWER8_VSX
-uint32_t adler32_power8(uint32_t adler, const uint8_t *buf, size_t len);
-uint32_t chunksize_power8(void);
-uint8_t* chunkmemset_safe_power8(uint8_t *out, unsigned dist, unsigned len, unsigned left);
-uint32_t crc32_power8(uint32_t crc, const uint8_t *buf, size_t len);
-#  ifdef DEFLATE_H_
-    void slide_hash_power8(deflate_state *s);
-#  endif
-#  ifdef INFLATE_H_
-    void inflate_fast_power8(PREFIX3(stream) *strm, uint32_t start);
-#  endif
-#endif
-
-#ifdef POWER9
-uint32_t compare256_power9(const uint8_t *src0, const uint8_t *src1);
-#  ifdef DEFLATE_H_
-    uint32_t longest_match_power9(deflate_state *const s, Pos cur_match);
-    uint32_t longest_match_slow_power9(deflate_state *const s, Pos cur_match);
-#  endif
-#endif
-
-#endif
-
-#endif /* POWER_H_ */
+#endif /* POWER_FEATURES_H_ */

arch/power/power_functions.h

@@ -0,0 +1,30 @@
+/* power_functions.h -- POWER implementations for arch-specific functions.
+ * Copyright (C) 2020 Matheus Castanho <[email protected]>, IBM
+ * Copyright (C) 2021 Mika T. Lindqvist <[email protected]>
+ * For conditions of distribution and use, see copyright notice in zlib.h
+ */
+
+#ifndef POWER_FUNCTIONS_H_
+#define POWER_FUNCTIONS_H_
+
+#ifdef PPC_VMX
+uint32_t adler32_vmx(uint32_t adler, const uint8_t *buf, size_t len);
+void slide_hash_vmx(deflate_state *s);
+#endif
+
+#ifdef POWER8_VSX
+uint32_t adler32_power8(uint32_t adler, const uint8_t *buf, size_t len);
+uint32_t chunksize_power8(void);
+uint8_t* chunkmemset_safe_power8(uint8_t *out, unsigned dist, unsigned len, unsigned left);
+uint32_t crc32_power8(uint32_t crc, const uint8_t *buf, size_t len);
+void slide_hash_power8(deflate_state *s);
+void inflate_fast_power8(PREFIX3(stream) *strm, uint32_t start);
+#endif
+
+#ifdef POWER9
+uint32_t compare256_power9(const uint8_t *src0, const uint8_t *src1);
+uint32_t longest_match_power9(deflate_state *const s, Pos cur_match);
+uint32_t longest_match_slow_power9(deflate_state *const s, Pos cur_match);
+#endif
+
+#endif /* POWER_FUNCTIONS_H_ */