Speed up CRC32 calculation on ARM64

The CRC32 instructions in ARM64 can calculate the CRC32 result
for 8 bytes in a single operation, making the use of ARM64
instructions much faster compared to the general CRC32 algorithm.

Optimized CRC32 will be enabled if ARM64 has CRC extension
running on Linux.

Signed-off-by: Chenxi Mao <chenxi.mao2013@gmail.com>

CMakeLists.txt

@@ -230,6 +230,7 @@ add_library(liblzma
     src/liblzma/check/check.h
     src/liblzma/check/crc_common.h
     src/liblzma/check/crc_x86_clmul.h
+    src/liblzma/check/crc32_aarch64.h
     src/liblzma/common/block_util.c
     src/liblzma/common/common.c
     src/liblzma/common/common.h

src/liblzma/check/Makefile.inc

@@ -15,7 +15,8 @@ liblzma_la_SOURCES += \
 	check/check.c \
 	check/check.h \
 	check/crc_common.h \
-	check/crc_x86_clmul.h
+	check/crc_x86_clmul.h \
+	check/crc32_aarch64.h
 
 if COND_SMALL
 liblzma_la_SOURCES += check/crc32_small.c

src/liblzma/check/crc32_aarch64.h

@@ -0,0 +1,109 @@
+///////////////////////////////////////////////////////////////////////////////
+//
+/// \file       crc32_aarch64.c
+/// \brief      CRC32 calculation with aarch64 optimization
+//
+//  Authors:    Chenxi Mao
+//
+//  This file has been put into the public domain.
+//  You can do whatever you want with this file.
+//
+///////////////////////////////////////////////////////////////////////////////
+#ifdef LZMA_CRC_CRC32_AARCH64_H
+#	error crc_arm64_clmul.h was included twice.
+#endif
+#define LZMA_CRC_CRC32_AARCH64_H
+#include <sys/auxv.h>
+// EDG-based compilers (Intel's classic compiler and compiler for E2K) can
+// define __GNUC__ but the attribute must not be used with them.
+// The new Clang-based ICX needs the attribute.
+//
+// NOTE: Build systems check for this too, keep them in sync with this.
+#if (defined(__GNUC__) || defined(__clang__)) && !defined(__EDG__)
+#	define crc_attr_target \
+        __attribute__((__target__("+crc")))
+#else
+#	define crc_attr_target
+#endif
+#ifdef BUILDING_CRC32_AARCH64
+crc_attr_target
+crc_attr_no_sanitize_address
+static uint32_t
+crc32_arch_optimized(const uint8_t *buf, size_t size, uint32_t crc)
+{
+	crc = ~crc;
+	while ((uintptr_t)(buf) & 7) {
+		crc = __builtin_aarch64_crc32b(crc, *buf);
+		buf++;
+		size--;
+	}
+	for (;size>=8;size-=8,buf+=8) {
+		crc = __builtin_aarch64_crc32x(crc, aligned_read64le(buf));
+	}
+	for (;size>0;size--,buf++)
+		crc = __builtin_aarch64_crc32b(crc, *buf);
+	return ~crc;
+}
+#endif
+#ifdef BUILDING_CRC64_AARCH64
+//FIXME: there is no crc64_arch_optimized implementation,
+// to make compiler happy, add crc64_generic here.
+#ifdef WORDS_BIGENDIAN
+#	define A1(x) ((x) >> 56)
+#else
+#	define A1 A
+#endif
+crc_attr_target
+crc_attr_no_sanitize_address
+static uint64_t
+crc64_arch_optimized(const uint8_t *buf, size_t size, uint64_t crc)
+{
+	crc = ~crc;
+
+#ifdef WORDS_BIGENDIAN
+	crc = bswap64(crc);
+#endif
+
+	if (size > 4) {
+		while ((uintptr_t)(buf) & 3) {
+			crc = lzma_crc64_table[0][*buf++ ^ A1(crc)] ^ S8(crc);
+			--size;
+		}
+
+		const uint8_t *const limit = buf + (size & ~(size_t)(3));
+		size &= (size_t)(3);
+
+		while (buf < limit) {
+#ifdef WORDS_BIGENDIAN
+			const uint32_t tmp = (uint32_t)(crc >> 32)
+					^ aligned_read32ne(buf);
+#else
+			const uint32_t tmp = (uint32_t)crc
+					^ aligned_read32ne(buf);
+#endif
+			buf += 4;
+
+			crc = lzma_crc64_table[3][A(tmp)]
+			    ^ lzma_crc64_table[2][B(tmp)]
+			    ^ S32(crc)
+			    ^ lzma_crc64_table[1][C(tmp)]
+			    ^ lzma_crc64_table[0][D(tmp)];
+		}
+	}
+
+	while (size-- != 0)
+		crc = lzma_crc64_table[0][*buf++ ^ A1(crc)] ^ S8(crc);
+
+#ifdef WORDS_BIGENDIAN
+	crc = bswap64(crc);
+#endif
+
+	return ~crc;
+}
+#endif
+static inline bool
+is_arch_extension_supported(void)
+{
+	return (getauxval(AT_HWCAP) & HWCAP_CRC32)!=0;
+}
+

src/liblzma/check/crc32_fast.c

@@ -15,9 +15,12 @@
 #include "check.h"
 #include "crc_common.h"
 
-#ifdef CRC_X86_CLMUL
+#if defined(CRC_X86_CLMUL)
 #	define BUILDING_CRC32_CLMUL
 #	include "crc_x86_clmul.h"
+#elif defined(CRC32_ARM64)
+#	define BUILDING_CRC32_AARCH64
+#	include "crc32_aarch64.h"
 #endif
 
 

src/liblzma/check/crc64_fast.c

@@ -14,9 +14,12 @@
 #include "check.h"
 #include "crc_common.h"
 
-#ifdef CRC_X86_CLMUL
+#if defined(CRC_X86_CLMUL)
 #	define BUILDING_CRC64_CLMUL
 #	include "crc_x86_clmul.h"
+#elif defined(CRC32_ARM64)
+#	define BUILDING_CRC64_AARCH64
+#	include "crc32_aarch64.h"
 #endif
 
 

src/liblzma/check/crc_common.h

@@ -52,29 +52,33 @@
 #undef CRC_GENERIC
 #undef CRC_ARCH_OPTIMIZED
 #undef CRC_X86_CLMUL
+#undef CRC32_ARM64
 #undef CRC_USE_IFUNC
 #undef CRC_USE_GENERIC_FOR_SMALL_INPUTS
 
-// If CLMUL cannot be used then only the generic slice-by-eight (CRC32)
-// or slice-by-four (CRC64) is built.
-#if !defined(HAVE_USABLE_CLMUL)
-#	define CRC_GENERIC 1
-
 // If CLMUL is allowed unconditionally in the compiler options then the
 // generic version can be omitted. Note that this doesn't work with MSVC
 // as I don't know how to detect the features here.
 //
 // NOTE: Keep this this in sync with crc32_table.c.
-#elif (defined(__SSSE3__) && defined(__SSE4_1__) && defined(__PCLMUL__)) \
+#if (defined(__SSSE3__) && defined(__SSE4_1__) && defined(__PCLMUL__)) \
 		|| (defined(__e2k__) && __iset__ >= 6)
 #	define CRC_ARCH_OPTIMIZED 1
 #	define CRC_X86_CLMUL 1
 
+#elif (defined(__aarch64__))
+#	define CRC_ARCH_OPTIMIZED 1
+#	define CRC32_ARM64 1
+// If CLMUL cannot be used then only the generic slice-by-eight (CRC32)
+// or slice-by-four (CRC64) is built.
+#elif !defined(HAVE_USABLE_CLMUL)
+#	define CRC_GENERIC 1
 // Otherwise build both and detect at runtime which version to use.
 #else
 #	define CRC_GENERIC 1
 #	define CRC_ARCH_OPTIMIZED 1
 #	define CRC_X86_CLMUL 1
+#	define CRC32_ARM64 1
 
 #	ifdef HAVE_FUNC_ATTRIBUTE_IFUNC
 #		define CRC_USE_IFUNC 1