Move adler32 C fallbacks to arch/generic

CMakeLists.txt

@@ -977,6 +977,7 @@ set(ZLIB_PUBLIC_HDRS
     ${CMAKE_CURRENT_BINARY_DIR}/zlib${SUFFIX}.h
 )
 set(ZLIB_PRIVATE_HDRS
+    arch/generic/adler32_fold_c.h
     adler32_p.h
     chunkset_tpl.h
     compare256_rle.h
@@ -1003,8 +1004,9 @@ set(ZLIB_PRIVATE_HDRS
     zutil.h
 )
 set(ZLIB_SRCS
+    arch/generic/adler32_c.c
+    arch/generic/adler32_fold_c.c
     adler32.c
-    adler32_fold.c
     chunkset.c
     compare256.c
     compress.c

Makefile.in

@@ -74,8 +74,9 @@ man3dir = ${mandir}/man3
 pkgconfigdir = ${libdir}/pkgconfig
 
 OBJZ = \
+	arch/generic/adler32_c.o \
+	arch/generic/adler32_fold_c.o \
 	adler32.o \
-	adler32_fold.o \
 	chunkset.o \
 	compare256.o \
 	compress.o \
@@ -112,8 +113,9 @@ TESTOBJG =
 OBJC = $(OBJZ) $(OBJG)
 
 PIC_OBJZ = \
+	arch/generic/adler32_c.lo \
+	arch/generic/adler32_fold_c.lo \
 	adler32.lo \
-	adler32_fold.lo \
 	chunkset.lo \
 	compare256.lo \
 	compress.lo \
@@ -169,6 +171,12 @@ $(ARCHDIR)/%.o: $(SRCDIR)/$(ARCHDIR)/%.c
 $(ARCHDIR)/%.lo: $(SRCDIR)/$(ARCHDIR)/%.c
 	$(MAKE) -C $(ARCHDIR) $(notdir $@)
 
+arch/generic/%.o: $(SRCDIR)/arch/generic/%.c
+	$(MAKE) -C arch/generic $(notdir $@)
+
+arch/generic/%.lo: $(SRCDIR)/arch/generic/%.c
+	$(MAKE) -C arch/generic $(notdir $@)
+
 %.o: $(ARCHDIR)/%.o
 	-cp $< $@
 

adler32.c

@@ -7,52 +7,6 @@
 #include "functable.h"
 #include "adler32_p.h"
 
-/* ========================================================================= */
-Z_INTERNAL uint32_t adler32_c(uint32_t adler, const uint8_t *buf, size_t len) {
-    uint32_t sum2;
-    unsigned n;
-
-    /* split Adler-32 into component sums */
-    sum2 = (adler >> 16) & 0xffff;
-    adler &= 0xffff;
-
-    /* in case user likes doing a byte at a time, keep it fast */
-    if (UNLIKELY(len == 1))
-        return adler32_len_1(adler, buf, sum2);
-
-    /* initial Adler-32 value (deferred check for len == 1 speed) */
-    if (UNLIKELY(buf == NULL))
-        return 1L;
-
-    /* in case short lengths are provided, keep it somewhat fast */
-    if (UNLIKELY(len < 16))
-        return adler32_len_16(adler, buf, len, sum2);
-
-    /* do length NMAX blocks -- requires just one modulo operation */
-    while (len >= NMAX) {
-        len -= NMAX;
-#ifdef UNROLL_MORE
-        n = NMAX / 16;          /* NMAX is divisible by 16 */
-#else
-        n = NMAX / 8;           /* NMAX is divisible by 8 */
-#endif
-        do {
-#ifdef UNROLL_MORE
-            DO16(adler, sum2, buf);          /* 16 sums unrolled */
-            buf += 16;
-#else
-            DO8(adler, sum2, buf, 0);         /* 8 sums unrolled */
-            buf += 8;
-#endif
-        } while (--n);
-        adler %= BASE;
-        sum2 %= BASE;
-    }
-
-    /* do remaining bytes (less than NMAX, still just one modulo) */
-    return adler32_len_64(adler, buf, len, sum2);
-}
-
 #ifdef ZLIB_COMPAT
 unsigned long Z_EXPORT PREFIX(adler32_z)(unsigned long adler, const unsigned char *buf, size_t len) {
     return (unsigned long)functable.adler32((uint32_t)adler, buf, len);

arch/generic/Makefile.in

@@ -1,5 +1,6 @@
-# Makefile for zlib
+# Makefile for zlib-ng
 # Copyright (C) 1995-2013 Jean-loup Gailly, Mark Adler
+# Copyright (C) 2024 Hans Kristian Rosbach
 # For conditions of distribution and use, see copyright notice in zlib.h
 
 CC=
@@ -11,12 +12,27 @@ SRCDIR=.
 SRCTOP=../..
 TOPDIR=$(SRCTOP)
 
-all:
+all: \
+ adler32_c.o adler32_c.lo \
+ adler32_fold_c.o adler32_fold_c.lo
+
+
+adler32_c.o: $(SRCDIR)/adler32_c.c  $(SRCTOP)/zbuild.h $(SRCTOP)/adler32_p.h
+	$(CC) $(CFLAGS) $(INCLUDES) -c -o $@ $(SRCDIR)/adler32_c.c
+
+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
+	$(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
+	$(CC) $(SFLAGS) $(INCLUDES) -c -o $@ $(SRCDIR)/adler32_fold_c.c
 
 
 mostlyclean: clean
 clean:
-	rm -f *.o *.lo *~ \
+	rm -f *.o *.lo *~
 	rm -rf objs
 	rm -f *.gcda *.gcno *.gcov
 

arch/generic/adler32_c.c

@@ -0,0 +1,54 @@
+/* adler32.c -- compute the Adler-32 checksum of a data stream
+ * Copyright (C) 1995-2011, 2016 Mark Adler
+ * For conditions of distribution and use, see copyright notice in zlib.h
+ */
+
+#include "zbuild.h"
+#include "functable.h"
+#include "adler32_p.h"
+
+/* ========================================================================= */
+Z_INTERNAL uint32_t adler32_c(uint32_t adler, const uint8_t *buf, size_t len) {
+    uint32_t sum2;
+    unsigned n;
+
+    /* split Adler-32 into component sums */
+    sum2 = (adler >> 16) & 0xffff;
+    adler &= 0xffff;
+
+    /* in case user likes doing a byte at a time, keep it fast */
+    if (UNLIKELY(len == 1))
+        return adler32_len_1(adler, buf, sum2);
+
+    /* initial Adler-32 value (deferred check for len == 1 speed) */
+    if (UNLIKELY(buf == NULL))
+        return 1L;
+
+    /* in case short lengths are provided, keep it somewhat fast */
+    if (UNLIKELY(len < 16))
+        return adler32_len_16(adler, buf, len, sum2);
+
+    /* do length NMAX blocks -- requires just one modulo operation */
+    while (len >= NMAX) {
+        len -= NMAX;
+#ifdef UNROLL_MORE
+        n = NMAX / 16;          /* NMAX is divisible by 16 */
+#else
+        n = NMAX / 8;           /* NMAX is divisible by 8 */
+#endif
+        do {
+#ifdef UNROLL_MORE
+            DO16(adler, sum2, buf);          /* 16 sums unrolled */
+            buf += 16;
+#else
+            DO8(adler, sum2, buf, 0);         /* 8 sums unrolled */
+            buf += 8;
+#endif
+        } while (--n);
+        adler %= BASE;
+        sum2 %= BASE;
+    }
+
+    /* do remaining bytes (less than NMAX, still just one modulo) */
+    return adler32_len_64(adler, buf, len, sum2);
+}

arch/generic/adler32_fold_c.c

@@ -5,7 +5,7 @@
 
 #include "zbuild.h"
 #include "functable.h"
-#include "adler32_fold.h"
+#include "adler32_fold_c.h"
 
 #include <limits.h>
 

arch/generic/adler32_fold_c.h

@@ -3,8 +3,8 @@
  * For conditions of distribution and use, see copyright notice in zlib.h
  */
 
-#ifndef ADLER32_FOLD_H_
+#ifndef ADLER32_FOLD_C_H_
-#define ADLER32_FOLD_H_
+#define ADLER32_FOLD_C_H_
 
 Z_INTERNAL uint32_t adler32_fold_copy_c(uint32_t adler, uint8_t *dst, const uint8_t *src, size_t len);
 

arch/x86/adler32_avx2.c

@@ -11,7 +11,6 @@
 
 #include "zbuild.h"
 #include <immintrin.h>
-#include "adler32_fold.h"
 #include "adler32_p.h"
 #include "adler32_avx2_p.h"
 #include "x86_intrins.h"

arch/x86/adler32_avx512.c

@@ -10,7 +10,6 @@
 
 #include "zbuild.h"
 #include "adler32_p.h"
-#include "adler32_fold.h"
 #include "cpu_features.h"
 #include <immintrin.h>
 #include "x86_intrins.h"

arch/x86/adler32_avx512_vnni.c

@@ -13,7 +13,6 @@
 #include "adler32_p.h"
 #include "cpu_features.h"
 #include <immintrin.h>
-#include "adler32_fold.h"
 #include "x86_intrins.h"
 #include "adler32_avx512_p.h"
 #include "adler32_avx2_p.h"

arch/x86/adler32_sse42.c

@@ -8,7 +8,6 @@
 
 #include "zbuild.h"
 #include "adler32_p.h"
-#include "adler32_fold.h"
 #include "adler32_ssse3_p.h"
 #include <immintrin.h>
 

configure

@@ -2109,6 +2109,21 @@ for file in $SRCDIR/$ARCHDIR/*.c; do
     fi
 done
 
+# Generate Makefile in generic arch dir
+mkdir -p arch/generic
+
+sed < $SRCDIR/arch/generic/Makefile.in "
+/^CC *=/s#=.*#=$CC#
+/^CFLAGS *=/s#=.*#=$CFLAGS#
+/^SFLAGS *=/s#=.*#=$SFLAGS#
+/^INCLUDES *=/s#=.*#=$INCLUDES#
+/^SRCDIR *=/s#=.*#=$SRCDIR/arch/generic#
+/^SRCTOP *=/s#=.*#=$SRCDIR#
+/^BUILDDIR *=/s#=.*#=$BUILDDIR#
+" > arch/generic/Makefile
+
+## TODO: Process header dependencies
+
 # Emscripten does not support large amounts of data via stdin/out
 # https://github.com/emscripten-core/emscripten/issues/16755#issuecomment-1102732849
 if test "$CHOST" != "wasm32"; then

cpu_features.h

@@ -6,7 +6,6 @@
 #ifndef CPU_FEATURES_H_
 #define CPU_FEATURES_H_
 
-#include "adler32_fold.h"
 #include "crc32_fold.h"
 
 #if defined(X86_FEATURES)

deflate.h

@@ -12,7 +12,6 @@
 
 #include "zutil.h"
 #include "zendian.h"
-#include "adler32_fold.h"
 #include "crc32_fold.h"
 
 /* define NO_GZIP when compiling if you want to disable gzip header and

functable.h

@@ -8,7 +8,7 @@
 
 #include "deflate.h"
 #include "crc32_fold.h"
-#include "adler32_fold.h"
+#include "adler32_fold_c.h"
 
 #ifdef ZLIB_COMPAT
 typedef struct z_stream_s z_stream;

inflate.h

@@ -11,7 +11,6 @@
 #ifndef INFLATE_H_
 #define INFLATE_H_
 
-#include "adler32_fold.h"
 #include "crc32_fold.h"
 
 /* define NO_GZIP when compiling if you want to disable gzip header and trailer decoding by inflate().

win32/Makefile.a64

@@ -44,7 +44,8 @@ SUFFIX =
 
 OBJS = \
 	adler32.obj \
-	adler32_fold.obj \
+	adler32_c.obj \
+	adler32_fold_c.obj \
 	arm_features.obj \
 	chunkset.obj \
 	compare256.obj \
@@ -169,6 +170,9 @@ gzwrite2.obj: gzwrite.c
 {$(TOP)/arch/arm}.c.obj:
 	$(CC) -c -I$(TOP) $(WFLAGS) $(CFLAGS) $<
 
+{$(TOP)/arch/generic}.c.obj:
+	$(CC) -c -I$(TOP) $(WFLAGS) $(CFLAGS) $<
+
 {$(TOP)/test}.c.obj:
 	$(CC) -c -I$(TOP) $(WFLAGS) $(CFLAGS) -DWITH_GZFILEOP $<
 
@@ -176,8 +180,9 @@ $(TOP)/zconf$(SUFFIX).h: zconf
 
 SRCDIR = $(TOP)
 # Keep the dependences in sync with top-level Makefile.in
-adler32.obj: $(SRCDIR)/adler32.c $(SRCDIR)/zbuild.h $(SRCDIR)/zutil.h $(SRCDIR)/functable.h $(SRCDIR)/adler32_p.h
+adler32.obj: $(SRCDIR)/adler32.c $(SRCDIR)/zbuild.h $(SRCDIR)/functable.h $(SRCDIR)/adler32_p.h
-adler32_fold.obj: $(SRCDIR)/adler32_fold.c $(SRCDIR)/zbuild.h $(SRCDIR)/adler32_fold.h $(SRCDIR)/functable.h
+adler32_c.obj: $(SRCDIR)/arch/generic/adler32_c.c $(SRCDIR)/zbuild.h $(SRCDIR)/functable.h $(SRCDIR)/adler32_p.h
+adler32_fold_c.obj: $(SRCDIR)/arch/generic/adler32_fold_c.c $(SRCDIR)/zbuild.h $(SRCDIR)/functable.h $(SRCDIR)/arch/generic/adler32_fold_c.h
 chunkset.obj: $(SRCDIR)/zbuild.h $(SRCDIR)/zutil.h
 functable.obj: $(SRCDIR)/functable.c $(SRCDIR)/zbuild.h $(SRCDIR)/functable.h $(SRCDIR)/deflate.h $(SRCDIR)/deflate_p.h $(SRCDIR)/zendian.h $(SRCDIR)/arch/x86/x86_features.h
 gzlib.obj: $(SRCDIR)/gzlib.c $(SRCDIR)/zbuild.h $(SRCDIR)/gzguts.h $(SRCDIR)/zutil_p.h

win32/Makefile.arm

@@ -49,7 +49,8 @@ SUFFIX =
 
 OBJS = \
 	adler32.obj \
-	adler32_fold.obj \
+	adler32_c.obj \
+	adler32_fold_c.obj \
 	arm_features.obj \
 	chunkset.obj \
 	compare256.obj \
@@ -190,6 +191,9 @@ gzwrite2.obj: gzwrite.c
 {$(TOP)/arch/arm}.c.obj:
 	$(CC) -c -I$(TOP) $(WFLAGS) $(CFLAGS) $<
 
+{$(TOP)/arch/generic}.c.obj:
+	$(CC) -c -I$(TOP) $(WFLAGS) $(CFLAGS) $<
+
 {$(TOP)/test}.c.obj:
 	$(CC) -c -I$(TOP) $(WFLAGS) $(CFLAGS) -DWITH_GZFILEOP $<
 
@@ -197,8 +201,9 @@ $(TOP)/zconf$(SUFFIX).h: zconf
 
 SRCDIR = $(TOP)
 # Keep the dependences in sync with top-level Makefile.in
-adler32.obj: $(SRCDIR)/adler32.c $(SRCDIR)/zbuild.h $(SRCDIR)/zutil.h $(SRCDIR)/functable.h $(SRCDIR)/adler32_p.h
+adler32.obj: $(SRCDIR)/adler32.c $(SRCDIR)/zbuild.h $(SRCDIR)/functable.h $(SRCDIR)/adler32_p.h
-adler32_fold.obj: $(SRCDIR)/adler32_fold.c $(SRCDIR)/zbuild.h $(SRCDIR)/adler32_fold.h $(SRCDIR)/functable.h
+adler32_c.obj: $(SRCDIR)/arch/generic/adler32_c.c $(SRCDIR)/zbuild.h $(SRCDIR)/functable.h $(SRCDIR)/adler32_p.h
+adler32_fold_c.obj: $(SRCDIR)/arch/generic/adler32_fold_c.c $(SRCDIR)/zbuild.h $(SRCDIR)/functable.h $(SRCDIR)/arch/generic/adler32_fold_c.h
 functable.obj: $(SRCDIR)/functable.c $(SRCDIR)/zbuild.h $(SRCDIR)/functable.h $(SRCDIR)/deflate.h $(SRCDIR)/deflate_p.h $(SRCDIR)/zendian.h $(SRCDIR)/arch/x86/x86_features.h
 gzlib.obj: $(SRCDIR)/gzlib.c $(SRCDIR)/zbuild.h $(SRCDIR)/gzguts.h $(SRCDIR)/zutil_p.h
 gzread.obj: $(SRCDIR)/gzread.c $(SRCDIR)/zbuild.h $(SRCDIR)/gzguts.h $(SRCDIR)/zutil_p.h

win32/Makefile.msc

@@ -47,12 +47,13 @@ SUFFIX =
 
 OBJS = \
 	adler32.obj \
+	adler32_c.obj \
 	adler32_avx2.obj \
 	adler32_avx512.obj \
 	adler32_avx512_vnni.obj \
 	adler32_sse42.obj \
 	adler32_ssse3.obj \
-	adler32_fold.obj \
+	adler32_fold_c.obj \
 	chunkset.obj \
 	chunkset_avx2.obj \
 	chunkset_sse2.obj \
@@ -178,6 +179,9 @@ gzwrite2.obj: gzwrite.c
 {$(TOP)/arch/x86}.c.obj:
 	$(CC) -c -I$(TOP) $(WFLAGS) $(CFLAGS) $<
 
+{$(TOP)/arch/generic}.c.obj:
+	$(CC) -c -I$(TOP) $(WFLAGS) $(CFLAGS) $<
+
 {$(TOP)/test}.c.obj:
 	$(CC) -c -I$(TOP) $(WFLAGS) $(CFLAGS) -DWITH_GZFILEOP $<
 
@@ -185,15 +189,16 @@ $(TOP)/zconf$(SUFFIX).h: zconf
 
 SRCDIR = $(TOP)
 # Keep the dependences in sync with top-level Makefile.in
-adler32.obj: $(SRCDIR)/adler32.c $(SRCDIR)/zbuild.h $(SRCDIR)/zutil.h $(SRCDIR)/functable.h $(SRCDIR)/adler32_p.h
+adler32.obj: $(SRCDIR)/adler32.c $(SRCDIR)/zbuild.h $(SRCDIR)/functable.h $(SRCDIR)/adler32_p.h
+adler32_c.obj: $(SRCDIR)/arch/generic/adler32_c.c $(SRCDIR)/zbuild.h $(SRCDIR)/functable.h $(SRCDIR)/adler32_p.h
 adler32_avx2.obj: $(SRCDIR)/arch/x86/adler32_avx2.c $(SRCDIR)/zbuild.h $(SRCDIR)/cpu_features.h $(SRCDIR)/adler32_p.h $(SRCDIR)/fallback_builtins.h
 adler32_avx512.obj: $(SRCDIR)/arch/x86/adler32_avx512.c $(SRCDIR)/zbuild.h $(SRCDIR)/cpu_features.h $(SRCDIR)/adler32_p.h $(SRCDIR)/arch/x86/adler32_avx512_p.h
 adler32_avx512_vnni.obj: $(SRCDIR)/arch/x86/adler32_avx512_vnni.c $(SRCDIR)/zbuild.h $(SRCDIR)/cpu_features.h $(SRCDIR)/adler32_p.h $(SRCDIR)/arch/x86/adler32_avx512_p.h
-adler32_sse42.obj: $(SRCDIR)/arch/x86/adler32_sse42.c $(SRCDIR)/zbuild.h $(SRCDIR)/cpu_features.h $(SRCDIR)/adler32_p.h $(SRCDIR)/adler32_fold.h \
+adler32_sse42.obj: $(SRCDIR)/arch/x86/adler32_sse42.c $(SRCDIR)/zbuild.h $(SRCDIR)/cpu_features.h $(SRCDIR)/adler32_p.h \
                    $(SRCDIR)/arch/x86/adler32_ssse3_p.h
-adler32_ssse3.obj: $(SRCDIR)/arch/x86/adler32_ssse3.c $(SRCDIR)/zbuild.h $(SRCDIR)/cpu_features.h $(SRCDIR)/adler32_p.h $(SRCDIR)/adler32_fold.h \
+adler32_ssse3.obj: $(SRCDIR)/arch/x86/adler32_ssse3.c $(SRCDIR)/zbuild.h $(SRCDIR)/cpu_features.h $(SRCDIR)/adler32_p.h \
                    $(SRCDIR)/arch/x86/adler32_ssse3_p.h
-adler32_fold.obj: $(SRCDIR)/adler32_fold.c $(SRCDIR)/zbuild.h $(SRCDIR)/adler32_fold.h $(SRCDIR)/functable.h
+adler32_fold_c.obj: $(SRCDIR)/arch/generic/adler32_fold_c.c $(SRCDIR)/zbuild.h $(SRCDIR)/functable.h $(SRCDIR)/arch/generic/adler32_fold_c.h
 functable.obj: $(SRCDIR)/functable.c $(SRCDIR)/zbuild.h $(SRCDIR)/functable.h $(SRCDIR)/deflate.h $(SRCDIR)/deflate_p.h $(SRCDIR)/zendian.h $(SRCDIR)/arch/x86/x86_features.h
 gzlib.obj: $(SRCDIR)/gzlib.c $(SRCDIR)/zbuild.h $(SRCDIR)/gzguts.h $(SRCDIR)/zutil_p.h
 gzread.obj: $(SRCDIR)/gzread.c $(SRCDIR)/zbuild.h $(SRCDIR)/gzguts.h $(SRCDIR)/zutil_p.h