ARM CPU feature cleanups

.github/workflows/ci.yml

@@ -17,6 +17,7 @@ jobs:
       run: |
         sudo apt-get update
         sudo apt-get install -y clang llvm libz-dev valgrind
+    - run: sudo sysctl kernel.randomize_va_space=0 # https://bugs.launchpad.net/ubuntu/+source/llvm-toolchain-14/+bug/2048768
     - run: scripts/run_tests.sh
     - name: Direct compilation without official build system
       run: $CC -O2 -Wall -Werror lib/*{,/*}.c programs/{gzip,prog_util,tgetopt}.c -o libdeflate-gzip
@@ -287,5 +288,6 @@ jobs:
         sudo apt-get install -y clang llvm
     - name: Fuzz
       run: |
+        sudo sysctl kernel.randomize_va_space=0 # https://bugs.launchpad.net/ubuntu/+source/llvm-toolchain-14/+bug/2048768
         scripts/libFuzzer/fuzz.sh --time=120 ${{matrix.sanitizer}} \
             ${{matrix.target}}

lib/arm/adler32_impl.h

@@ -32,17 +32,20 @@
 
 /* Regular NEON implementation */
 #if HAVE_NEON_INTRIN && CPU_IS_LITTLE_ENDIAN()
-#  define adler32_arm_neon		adler32_arm_neon
+#  define adler32_arm_neon	adler32_arm_neon
 #  if HAVE_NEON_NATIVE
+     /*
+      * Use no attributes if none are needed, to support old versions of clang
+      * that don't accept the simd target attribute.
+      */
 #    define ATTRIBUTES
+#  elif defined(ARCH_ARM32)
+#    define ATTRIBUTES	_target_attribute("fpu=neon")
+#  elif defined(__clang__)
+#    define ATTRIBUTES	_target_attribute("simd")
 #  else
-#    ifdef ARCH_ARM32
-#      define ATTRIBUTES	_target_attribute("fpu=neon")
-#    else
-#      define ATTRIBUTES	_target_attribute("+simd")
-#    endif
+#    define ATTRIBUTES	_target_attribute("+simd")
 #  endif
-#  include <arm_neon.h>
 static ATTRIBUTES MAYBE_UNUSED u32
 adler32_arm_neon(u32 adler, const u8 *p, size_t len)
 {
@@ -208,23 +211,18 @@ adler32_arm_neon(u32 adler, const u8 *p, size_t len)
 /* NEON+dotprod implementation */
 #if HAVE_DOTPROD_INTRIN && CPU_IS_LITTLE_ENDIAN()
 #  define adler32_arm_neon_dotprod	adler32_arm_neon_dotprod
-#  if HAVE_DOTPROD_NATIVE
-#    define ATTRIBUTES
+#  ifdef __clang__
+#    define ATTRIBUTES	_target_attribute("dotprod")
+   /*
+    * With gcc, arch=armv8.2-a is needed for dotprod intrinsics, unless the
+    * default target is armv8.3-a or later in which case it must be omitted.
+    * armv8.3-a or later can be detected by checking for __ARM_FEATURE_JCVT.
+    */
+#  elif defined(__ARM_FEATURE_JCVT)
+#    define ATTRIBUTES	_target_attribute("+dotprod")
 #  else
-#    ifdef __clang__
-#      define ATTRIBUTES  _target_attribute("dotprod")
-     /*
-      * With gcc, arch=armv8.2-a is needed for dotprod intrinsics, unless the
-      * default target is armv8.3-a or later in which case it must be omitted.
-      * armv8.3-a or later can be detected by checking for __ARM_FEATURE_JCVT.
-      */
-#    elif defined(__ARM_FEATURE_JCVT)
-#      define ATTRIBUTES  _target_attribute("+dotprod")
-#    else
-#      define ATTRIBUTES  _target_attribute("arch=armv8.2-a+dotprod")
-#    endif
+#    define ATTRIBUTES	_target_attribute("arch=armv8.2-a+dotprod")
 #  endif
-#  include <arm_neon.h>
 static ATTRIBUTES u32
 adler32_arm_neon_dotprod(u32 adler, const u8 *p, size_t len)
 {
@@ -334,7 +332,7 @@ adler32_arm_neon_dotprod(u32 adler, const u8 *p, size_t len)
 #undef ATTRIBUTES
 #endif /* NEON+dotprod implementation */
 
-#if defined(adler32_arm_neon_dotprod) && HAVE_DOTPROD_NATIVE
+#if defined(adler32_arm_neon_dotprod) && defined(__ARM_FEATURE_DOTPROD)
 #define DEFAULT_IMPL	adler32_arm_neon_dotprod
 #else
 static inline adler32_func_t

lib/arm/cpu_features.c

@@ -113,10 +113,6 @@ static u32 query_arm_cpu_features(void)
 	STATIC_ASSERT(sizeof(long) == 4);
 	if (hwcap & (1 << 12))	/* HWCAP_NEON */
 		features |= ARM_CPU_FEATURE_NEON;
-	if (hwcap2 & (1 << 1))	/* HWCAP2_PMULL */
-		features |= ARM_CPU_FEATURE_PMULL;
-	if (hwcap2 & (1 << 4))	/* HWCAP2_CRC32 */
-		features |= ARM_CPU_FEATURE_CRC32;
 #else
 	STATIC_ASSERT(sizeof(long) == 8);
 	if (hwcap & (1 << 1))	/* HWCAP_ASIMD */
@@ -138,6 +134,7 @@ static u32 query_arm_cpu_features(void)
 
 #include <sys/types.h>
 #include <sys/sysctl.h>
+#include <TargetConditionals.h>
 
 static const struct {
 	const char *name;
@@ -192,6 +189,7 @@ static u32 query_arm_cpu_features(void)
 static const struct cpu_feature arm_cpu_feature_table[] = {
 	{ARM_CPU_FEATURE_NEON,		"neon"},
 	{ARM_CPU_FEATURE_PMULL,		"pmull"},
+	{ARM_CPU_FEATURE_PREFER_PMULL,  "prefer_pmull"},
 	{ARM_CPU_FEATURE_CRC32,		"crc32"},
 	{ARM_CPU_FEATURE_SHA3,		"sha3"},
 	{ARM_CPU_FEATURE_DOTPROD,	"dotprod"},
@@ -203,6 +201,19 @@ void libdeflate_init_arm_cpu_features(void)
 {
 	u32 features = query_arm_cpu_features();
 
+	/*
+	 * On the Apple M1 processor, crc32 instructions max out at about 25.5
+	 * GB/s in the best case of using a 3-way or greater interleaved chunked
+	 * implementation, whereas a pmull-based implementation achieves 68 GB/s
+	 * provided that the stride length is large enough (about 10+ vectors
+	 * with eor3, or 12+ without).
+	 *
+	 * Assume that crc32 instructions are preferable in other cases.
+	 */
+#if (defined(__APPLE__) && TARGET_OS_OSX) || defined(TEST_SUPPORT__DO_NOT_USE)
+	features |= ARM_CPU_FEATURE_PREFER_PMULL;
+#endif
+
 	disable_cpu_features_for_testing(&features, arm_cpu_feature_table,
 					 ARRAY_LEN(arm_cpu_feature_table));