Merging upstream version 5.6.2.

Signed-off-by: Daniel Baumann <daniel.baumann@progress-linux.org>

22 files changed, 954 insertions, 535 deletions

diff --git a/src/liblzma/check/Makefile.inc b/src/liblzma/check/Makefile.inc
index dc011a3..04cf1ff 100644
--- a/src/liblzma/check/Makefile.inc
+++ b/src/liblzma/check/Makefile.inc
@@ -1,9 +1,8 @@
-##
+## SPDX-License-Identifier: 0BSD
 ## Author: Lasse Collin
-##
-## This file has been put into the public domain.
-## You can do whatever you want with this file.
-##
+
+## Note: There is no check for COND_CHECK_CRC32 because
+## currently crc32 is always enabled.
 
 EXTRA_DIST += \
 	check/crc32_tablegen.c \
@@ -12,9 +11,10 @@ EXTRA_DIST += \
 liblzma_la_SOURCES += \
 	check/check.c \
 	check/check.h \
-	check/crc_macros.h
+	check/crc_common.h \
+	check/crc_x86_clmul.h \
+	check/crc32_arm64.h
 
-if COND_CHECK_CRC32
 if COND_SMALL
 liblzma_la_SOURCES += check/crc32_small.c
 else
@@ -28,7 +28,6 @@ else
 liblzma_la_SOURCES += check/crc32_fast.c
 endif
 endif
-endif
 
 if COND_CHECK_CRC64
 if COND_SMALL
diff --git a/src/liblzma/check/check.c b/src/liblzma/check/check.c
index 428ddae..7734ace 100644
--- a/src/liblzma/check/check.c
+++ b/src/liblzma/check/check.c
@@ -1,3 +1,5 @@
+// SPDX-License-Identifier: 0BSD
+
 ///////////////////////////////////////////////////////////////////////////////
 //
 /// \file       check.c
@@ -5,9 +7,6 @@
 //
 //  Author:     Lasse Collin
 //
-//  This file has been put into the public domain.
-//  You can do whatever you want with this file.
-//
 ///////////////////////////////////////////////////////////////////////////////
 
 #include "check.h"
diff --git a/src/liblzma/check/check.h b/src/liblzma/check/check.h
index 8ae95d5..f0eb117 100644
--- a/src/liblzma/check/check.h
+++ b/src/liblzma/check/check.h
@@ -1,3 +1,5 @@
+// SPDX-License-Identifier: 0BSD
+
 ///////////////////////////////////////////////////////////////////////////////
 //
 /// \file       check.h
@@ -5,9 +7,6 @@
 //
 //  Author:     Lasse Collin
 //
-//  This file has been put into the public domain.
-//  You can do whatever you want with this file.
-//
 ///////////////////////////////////////////////////////////////////////////////
 
 #ifndef LZMA_CHECK_H
diff --git a/src/liblzma/check/crc32_arm64.h b/src/liblzma/check/crc32_arm64.h
new file mode 100644
index 0000000..39c1c63
--- /dev/null
+++ b/src/liblzma/check/crc32_arm64.h
@@ -0,0 +1,122 @@
+// SPDX-License-Identifier: 0BSD
+
+///////////////////////////////////////////////////////////////////////////////
+//
+/// \file       crc32_arm64.h
+/// \brief      CRC32 calculation with ARM64 optimization
+//
+//  Authors:    Chenxi Mao
+//              Jia Tan
+//              Hans Jansen
+//
+///////////////////////////////////////////////////////////////////////////////
+
+#ifndef LZMA_CRC32_ARM64_H
+#define LZMA_CRC32_ARM64_H
+
+// MSVC always has the CRC intrinsics available when building for ARM64
+// there is no need to include any header files.
+#ifndef _MSC_VER
+#	include <arm_acle.h>
+#endif
+
+// If both versions are going to be built, we need runtime detection
+// to check if the instructions are supported.
+#if defined(CRC32_GENERIC) && defined(CRC32_ARCH_OPTIMIZED)
+#	if defined(HAVE_GETAUXVAL) || defined(HAVE_ELF_AUX_INFO)
+#		include <sys/auxv.h>
+#	elif defined(_WIN32)
+#		include <processthreadsapi.h>
+#	elif defined(__APPLE__) && defined(HAVE_SYSCTLBYNAME)
+#		include <sys/sysctl.h>
+#	endif
+#endif
+
+// Some EDG-based compilers support ARM64 and define __GNUC__
+// (such as Nvidia's nvcc), but do not support function attributes.
+//
+// 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
+
+
+crc_attr_target
+static uint32_t
+crc32_arch_optimized(const uint8_t *buf, size_t size, uint32_t crc)
+{
+	crc = ~crc;
+
+	// Align the input buffer because this was shown to be
+	// significantly faster than unaligned accesses.
+	const size_t align_amount = my_min(size, (0U - (uintptr_t)buf) & 7);
+
+	for (const uint8_t *limit = buf + align_amount; buf < limit; ++buf)
+		crc = __crc32b(crc, *buf);
+
+	size -= align_amount;
+
+	// Process 8 bytes at a time. The end point is determined by
+	// ignoring the least significant three bits of size to ensure
+	// we do not process past the bounds of the buffer. This guarantees
+	// that limit is a multiple of 8 and is strictly less than size.
+	for (const uint8_t *limit = buf + (size & ~(size_t)7);
+			buf < limit; buf += 8)
+		crc = __crc32d(crc, aligned_read64le(buf));
+
+	// Process the remaining bytes that are not 8 byte aligned.
+	for (const uint8_t *limit = buf + (size & 7); buf < limit; ++buf)
+		crc = __crc32b(crc, *buf);
+
+	return ~crc;
+}
+
+
+#if defined(CRC32_GENERIC) && defined(CRC32_ARCH_OPTIMIZED)
+static inline bool
+is_arch_extension_supported(void)
+{
+#if defined(HAVE_GETAUXVAL)
+	return (getauxval(AT_HWCAP) & HWCAP_CRC32) != 0;
+
+#elif defined(HAVE_ELF_AUX_INFO)
+	unsigned long feature_flags;
+
+	if (elf_aux_info(AT_HWCAP, &feature_flags, sizeof(feature_flags)) != 0)
+		return false;
+
+	return (feature_flags & HWCAP_CRC32) != 0;
+
+#elif defined(_WIN32)
+	return IsProcessorFeaturePresent(
+			PF_ARM_V8_CRC32_INSTRUCTIONS_AVAILABLE);
+
+#elif defined(__APPLE__) && defined(HAVE_SYSCTLBYNAME)
+	int has_crc32 = 0;
+	size_t size = sizeof(has_crc32);
+
+	// The sysctlbyname() function requires a string identifier for the
+	// CPU feature it tests. The Apple documentation lists the string
+	// "hw.optional.armv8_crc32", which can be found here:
+	// https://developer.apple.com/documentation/kernel/1387446-sysctlbyname/determining_instruction_set_characteristics#3915619
+	if (sysctlbyname("hw.optional.armv8_crc32", &has_crc32,
+			&size, NULL, 0) != 0)
+		return false;
+
+	return has_crc32;
+
+#else
+	// If a runtime detection method cannot be found, then this must
+	// be a compile time error. The checks in crc_common.h should ensure
+	// a runtime detection method is always found if this function is
+	// built. It would be possible to just return false here, but this
+	// is inefficient for binary size and runtime since only the generic
+	// method could ever be used.
+#	error Runtime detection method unavailable.
+#endif
+}
+#endif
+
+#endif // LZMA_CRC32_ARM64_H
diff --git a/src/liblzma/check/crc32_fast.c b/src/liblzma/check/crc32_fast.c
index eed7350..16dbb74 100644
--- a/src/liblzma/check/crc32_fast.c
+++ b/src/liblzma/check/crc32_fast.c
@@ -1,35 +1,40 @@
+// SPDX-License-Identifier: 0BSD
+
 ///////////////////////////////////////////////////////////////////////////////
 //
 /// \file       crc32.c
 /// \brief      CRC32 calculation
-///
-/// Calculate the CRC32 using the slice-by-eight algorithm.
-/// It is explained in this document:
-/// http://www.intel.com/technology/comms/perfnet/download/CRC_generators.pdf
-/// The code in this file is not the same as in Intel's paper, but
-/// the basic principle is identical.
-//
-//  Author:     Lasse Collin
 //
-//  This file has been put into the public domain.
-//  You can do whatever you want with this file.
+//  Authors:    Lasse Collin
+//              Ilya Kurdyukov
+//              Hans Jansen
 //
 ///////////////////////////////////////////////////////////////////////////////
 
 #include "check.h"
-#include "crc_macros.h"
+#include "crc_common.h"
+
+#if defined(CRC_X86_CLMUL)
+#	define BUILDING_CRC32_CLMUL
+#	include "crc_x86_clmul.h"
+#elif defined(CRC32_ARM64)
+#	include "crc32_arm64.h"
+#endif
 
 
-// If you make any changes, do some benchmarking! Seemingly unrelated
-// changes can very easily ruin the performance (and very probably is
-// very compiler dependent).
-extern LZMA_API(uint32_t)
-lzma_crc32(const uint8_t *buf, size_t size, uint32_t crc)
+#ifdef CRC32_GENERIC
+
+///////////////////
+// Generic CRC32 //
+///////////////////
+
+static uint32_t
+crc32_generic(const uint8_t *buf, size_t size, uint32_t crc)
 {
 	crc = ~crc;
 
 #ifdef WORDS_BIGENDIAN
-	crc = bswap32(crc);
+	crc = byteswap32(crc);
 #endif
 
 	if (size > 8) {
@@ -75,8 +80,125 @@ lzma_crc32(const uint8_t *buf, size_t size, uint32_t crc)
 		crc = lzma_crc32_table[0][*buf++ ^ A(crc)] ^ S8(crc);
 
 #ifdef WORDS_BIGENDIAN
-	crc = bswap32(crc);
+	crc = byteswap32(crc);
 #endif
 
 	return ~crc;
 }
+#endif
+
+
+#if defined(CRC32_GENERIC) && defined(CRC32_ARCH_OPTIMIZED)
+
+//////////////////////////
+// Function dispatching //
+//////////////////////////
+
+// If both the generic and arch-optimized implementations are built, then
+// the function to use is selected at runtime because the system running
+// the binary might not have the arch-specific instruction set extension(s)
+// available. The dispatch methods in order of priority:
+//
+// 1. Constructor. This method uses __attribute__((__constructor__)) to
+//    set crc32_func at load time. This avoids extra computation (and any
+//    unlikely threading bugs) on the first call to lzma_crc32() to decide
+//    which implementation should be used.
+//
+// 2. First Call Resolution. On the very first call to lzma_crc32(), the
+//    call will be directed to crc32_dispatch() instead. This will set the
+//    appropriate implementation function and will not be called again.
+//    This method does not use any kind of locking but is safe because if
+//    multiple threads run the dispatcher simultaneously then they will all
+//    set crc32_func to the same value.
+
+typedef uint32_t (*crc32_func_type)(
+		const uint8_t *buf, size_t size, uint32_t crc);
+
+// This resolver is shared between all dispatch methods.
+static crc32_func_type
+crc32_resolve(void)
+{
+	return is_arch_extension_supported()
+			? &crc32_arch_optimized : &crc32_generic;
+}
+
+
+#ifdef HAVE_FUNC_ATTRIBUTE_CONSTRUCTOR
+// Constructor method.
+#	define CRC32_SET_FUNC_ATTR __attribute__((__constructor__))
+static crc32_func_type crc32_func;
+#else
+// First Call Resolution method.
+#	define CRC32_SET_FUNC_ATTR
+static uint32_t crc32_dispatch(const uint8_t *buf, size_t size, uint32_t crc);
+static crc32_func_type crc32_func = &crc32_dispatch;
+#endif
+
+CRC32_SET_FUNC_ATTR
+static void
+crc32_set_func(void)
+{
+	crc32_func = crc32_resolve();
+	return;
+}
+
+#ifndef HAVE_FUNC_ATTRIBUTE_CONSTRUCTOR
+static uint32_t
+crc32_dispatch(const uint8_t *buf, size_t size, uint32_t crc)
+{
+	// When __attribute__((__constructor__)) isn't supported, set the
+	// function pointer without any locking. If multiple threads run
+	// the detection code in parallel, they will all end up setting
+	// the pointer to the same value. This avoids the use of
+	// mythread_once() on every call to lzma_crc32() but this likely
+	// isn't strictly standards compliant. Let's change it if it breaks.
+	crc32_set_func();
+	return crc32_func(buf, size, crc);
+}
+
+#endif
+#endif
+
+
+extern LZMA_API(uint32_t)
+lzma_crc32(const uint8_t *buf, size_t size, uint32_t crc)
+{
+#if defined(CRC32_GENERIC) && defined(CRC32_ARCH_OPTIMIZED)
+	// On x86-64, if CLMUL is available, it is the best for non-tiny
+	// inputs, being over twice as fast as the generic slice-by-four
+	// version. However, for size <= 16 it's different. In the extreme
+	// case of size == 1 the generic version can be five times faster.
+	// At size >= 8 the CLMUL starts to become reasonable. It
+	// varies depending on the alignment of buf too.
+	//
+	// The above doesn't include the overhead of mythread_once().
+	// At least on x86-64 GNU/Linux, pthread_once() is very fast but
+	// it still makes lzma_crc32(buf, 1, crc) 50-100 % slower. When
+	// size reaches 12-16 bytes the overhead becomes negligible.
+	//
+	// So using the generic version for size <= 16 may give better
+	// performance with tiny inputs but if such inputs happen rarely
+	// it's not so obvious because then the lookup table of the
+	// generic version may not be in the processor cache.
+#ifdef CRC_USE_GENERIC_FOR_SMALL_INPUTS
+	if (size <= 16)
+		return crc32_generic(buf, size, crc);
+#endif
+
+/*
+#ifndef HAVE_FUNC_ATTRIBUTE_CONSTRUCTOR
+	// See crc32_dispatch(). This would be the alternative which uses
+	// locking and doesn't use crc32_dispatch(). Note that on Windows
+	// this method needs Vista threads.
+	mythread_once(crc64_set_func);
+#endif
+*/
+	return crc32_func(buf, size, crc);
+
+#elif defined(CRC32_ARCH_OPTIMIZED)
+	return crc32_arch_optimized(buf, size, crc);
+
+#else
+	return crc32_generic(buf, size, crc);
+#endif
+}
diff --git a/src/liblzma/check/crc32_small.c b/src/liblzma/check/crc32_small.c
index 186966e..6a1bd66 100644
--- a/src/liblzma/check/crc32_small.c
+++ b/src/liblzma/check/crc32_small.c
@@ -1,3 +1,5 @@
+// SPDX-License-Identifier: 0BSD
+
 ///////////////////////////////////////////////////////////////////////////////
 //
 /// \file       crc32_small.c
@@ -5,9 +7,6 @@
 //
 //  Author:     Lasse Collin
 //
-//  This file has been put into the public domain.
-//  You can do whatever you want with this file.
-//
 ///////////////////////////////////////////////////////////////////////////////
 
 #include "check.h"
diff --git a/src/liblzma/check/crc32_table.c b/src/liblzma/check/crc32_table.c
index b11762a..c141cef 100644
--- a/src/liblzma/check/crc32_table.c
+++ b/src/liblzma/check/crc32_table.c
@@ -1,3 +1,5 @@
+// SPDX-License-Identifier: 0BSD
+
 ///////////////////////////////////////////////////////////////////////////////
 //
 /// \file       crc32_table.c
@@ -5,18 +7,36 @@
 //
 //  Author:     Lasse Collin
 //
-//  This file has been put into the public domain.
-//  You can do whatever you want with this file.
-//
 ///////////////////////////////////////////////////////////////////////////////
 
 #include "common.h"
 
+
+// FIXME: Compared to crc_common.h this has to check for __x86_64__ too
+// so that in 32-bit builds crc32_x86.S won't break due to a missing table.
+#if defined(HAVE_USABLE_CLMUL) && ((defined(__x86_64__) && defined(__SSSE3__) \
+			&& defined(__SSE4_1__) && defined(__PCLMUL__)) \
+		|| (defined(__e2k__) && __iset__ >= 6))
+#	define NO_CRC32_TABLE
+
+#elif defined(HAVE_ARM64_CRC32) \
+		&& !defined(WORDS_BIGENDIAN) \
+		&& defined(__ARM_FEATURE_CRC32)
+#	define NO_CRC32_TABLE
+#endif
+
+
+#if !defined(HAVE_ENCODERS) && defined(NO_CRC32_TABLE)
+// No table needed. Use a typedef to avoid an empty translation unit.
+typedef void lzma_crc32_dummy;
+
+#else
 // Having the declaration here silences clang -Wmissing-variable-declarations.
 extern const uint32_t lzma_crc32_table[8][256];
 
-#ifdef WORDS_BIGENDIAN
-#	include "crc32_table_be.h"
-#else
-#	include "crc32_table_le.h"
+#       ifdef WORDS_BIGENDIAN
+#       	include "crc32_table_be.h"
+#       else
+#       	include "crc32_table_le.h"
+#       endif
 #endif
diff --git a/src/liblzma/check/crc32_table_be.h b/src/liblzma/check/crc32_table_be.h
index c483cb6..505c230 100644
--- a/src/liblzma/check/crc32_table_be.h
+++ b/src/liblzma/check/crc32_table_be.h
@@ -1,4 +1,6 @@
-/* This file has been automatically generated by crc32_tablegen.c. */
+// SPDX-License-Identifier: 0BSD
+
+// This file has been generated by crc32_tablegen.c.
 
 const uint32_t lzma_crc32_table[8][256] = {
 	{
diff --git a/src/liblzma/check/crc32_table_le.h b/src/liblzma/check/crc32_table_le.h
index 25f4fc4..e89c21a 100644
--- a/src/liblzma/check/crc32_table_le.h
+++ b/src/liblzma/check/crc32_table_le.h
@@ -1,4 +1,6 @@
-/* This file has been automatically generated by crc32_tablegen.c. */
+// SPDX-License-Identifier: 0BSD
+
+// This file has been generated by crc32_tablegen.c.
 
 const uint32_t lzma_crc32_table[8][256] = {
 	{
diff --git a/src/liblzma/check/crc32_tablegen.c b/src/liblzma/check/crc32_tablegen.c
index 31a4d27..b8cf459 100644
--- a/src/liblzma/check/crc32_tablegen.c
+++ b/src/liblzma/check/crc32_tablegen.c
@@ -1,3 +1,5 @@
+// SPDX-License-Identifier: 0BSD
+
 ///////////////////////////////////////////////////////////////////////////////
 //
 /// \file       crc32_tablegen.c
@@ -9,9 +11,6 @@
 //
 //  Author:     Lasse Collin
 //
-//  This file has been put into the public domain.
-//  You can do whatever you want with this file.
-//
 ///////////////////////////////////////////////////////////////////////////////
 
 #include <stdio.h>
@@ -44,7 +43,7 @@ init_crc32_table(void)
 #ifdef WORDS_BIGENDIAN
 	for (size_t s = 0; s < 8; ++s)
 		for (size_t b = 0; b < 256; ++b)
-			crc32_table[s][b] = bswap32(crc32_table[s][b]);
+			crc32_table[s][b] = byteswap32(crc32_table[s][b]);
 #endif
 
 	return;
@@ -54,9 +53,11 @@ init_crc32_table(void)
 static void
 print_crc32_table(void)
 {
-	printf("/* This file has been automatically generated by "
-			"crc32_tablegen.c. */\n\n"
-			"const uint32_t lzma_crc32_table[8][256] = {\n\t{");
+	// Split the SPDX string so that it won't accidentally match
+	// when tools search for the string.
+	printf("// SPDX" "-License-Identifier" ": 0BSD\n\n"
+		"// This file has been generated by crc32_tablegen.c.\n\n"
+		"const uint32_t lzma_crc32_table[8][256] = {\n\t{");
 
 	for (size_t s = 0; s < 8; ++s) {
 		for (size_t b = 0; b < 256; ++b) {
@@ -82,9 +83,11 @@ print_crc32_table(void)
 static void
 print_lz_table(void)
 {
-	printf("/* This file has been automatically generated by "
-			"crc32_tablegen.c. */\n\n"
-			"const uint32_t lzma_lz_hash_table[256] = {");
+	// Split the SPDX string so that it won't accidentally match
+	// when tools search for the string.
+	printf("// SPDX" "-License-Identifier" ": 0BSD\n\n"
+		"// This file has been generated by crc32_tablegen.c.\n\n"
+		"const uint32_t lzma_lz_hash_table[256] = {");
 
 	for (size_t b = 0; b < 256; ++b) {
 		if ((b % 4) == 0)
diff --git a/src/liblzma/check/crc32_x86.S b/src/liblzma/check/crc32_x86.S
index 4f395df..ddc3cee 100644
--- a/src/liblzma/check/crc32_x86.S
+++ b/src/liblzma/check/crc32_x86.S
@@ -1,3 +1,5 @@
+/* SPDX-License-Identifier: 0BSD */
+
 /*
  * Speed-optimized CRC32 using slicing-by-eight algorithm
  *
@@ -11,9 +13,6 @@
  * Authors: Igor Pavlov (original version)
  *          Lasse Collin (AT&T syntax, PIC support, better portability)
  *
- * This file has been put into the public domain.
- * You can do whatever you want with this file.
- *
  * This code needs lzma_crc32_table, which can be created using the
  * following C code:
 
diff --git a/src/liblzma/check/crc64_fast.c b/src/liblzma/check/crc64_fast.c
index 0c8622a..0ce83fe 100644
--- a/src/liblzma/check/crc64_fast.c
+++ b/src/liblzma/check/crc64_fast.c
@@ -1,85 +1,30 @@
+// SPDX-License-Identifier: 0BSD
+
 ///////////////////////////////////////////////////////////////////////////////
 //
 /// \file       crc64.c
 /// \brief      CRC64 calculation
-///
-/// There are two methods in this file. crc64_generic uses the
-/// the slice-by-four algorithm. This is the same idea that is
-/// used in crc32_fast.c, but for CRC64 we use only four tables
-/// instead of eight to avoid increasing CPU cache usage.
-///
-/// crc64_clmul uses 32/64-bit x86 SSSE3, SSE4.1, and CLMUL instructions.
-/// It was derived from
-/// https://www.intel.com/content/dam/www/public/us/en/documents/white-papers/fast-crc-computation-generic-polynomials-pclmulqdq-paper.pdf
-/// and the public domain code from https://github.com/rawrunprotected/crc
-/// (URLs were checked on 2022-11-07).
-///
-/// FIXME: Builds for 32-bit x86 use crc64_x86.S by default instead
-/// of this file and thus CLMUL version isn't available on 32-bit x86
-/// unless configured with --disable-assembler. Even then the lookup table
-/// isn't omitted in crc64_table.c since it doesn't know that assembly
-/// code has been disabled.
 //
 //  Authors:    Lasse Collin
 //              Ilya Kurdyukov
 //
-//  This file has been put into the public domain.
-//  You can do whatever you want with this file.
-//
 ///////////////////////////////////////////////////////////////////////////////
 
 #include "check.h"
+#include "crc_common.h"
 
-#undef CRC_GENERIC
-#undef CRC_CLMUL
-#undef CRC_USE_GENERIC_FOR_SMALL_INPUTS
-
-// If CLMUL cannot be used then only the generic slice-by-four 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 crc64_table.c.
-#elif (defined(__SSSE3__) && defined(__SSE4_1__) && defined(__PCLMUL__)) \
-		|| (defined(__e2k__) && __iset__ >= 6)
-#	define CRC_CLMUL 1
-
-// Otherwise build both and detect at runtime which version to use.
-#else
-#	define CRC_GENERIC 1
-#	define CRC_CLMUL 1
-
-/*
-	// The generic code is much faster with 1-8-byte inputs and has
-	// similar performance up to 16 bytes  at least in microbenchmarks
-	// (it depends on input buffer alignment too). If both versions are
-	// built, this #define will use the generic version for inputs up to
-	// 16 bytes and CLMUL for bigger inputs. It saves a little in code
-	// size since the special cases for 0-16-byte inputs will be omitted
-	// from the CLMUL code.
-#	define CRC_USE_GENERIC_FOR_SMALL_INPUTS 1
-*/
-
-#	if defined(_MSC_VER)
-#		include <intrin.h>
-#	elif defined(HAVE_CPUID_H)
-#		include <cpuid.h>
-#	endif
+#if defined(CRC_X86_CLMUL)
+#	define BUILDING_CRC64_CLMUL
+#	include "crc_x86_clmul.h"
 #endif
 
 
+#ifdef CRC64_GENERIC
+
 /////////////////////////////////
 // Generic slice-by-four CRC64 //
 /////////////////////////////////
 
-#ifdef CRC_GENERIC
-
-#include "crc_macros.h"
-
-
 #ifdef WORDS_BIGENDIAN
 #	define A1(x) ((x) >> 56)
 #else
@@ -94,7 +39,7 @@ crc64_generic(const uint8_t *buf, size_t size, uint64_t crc)
 	crc = ~crc;
 
 #ifdef WORDS_BIGENDIAN
-	crc = bswap64(crc);
+	crc = byteswap64(crc);
 #endif
 
 	if (size > 4) {
@@ -128,7 +73,7 @@ crc64_generic(const uint8_t *buf, size_t size, uint64_t crc)
 		crc = lzma_crc64_table[0][*buf++ ^ A1(crc)] ^ S8(crc);
 
 #ifdef WORDS_BIGENDIAN
-	crc = bswap64(crc);
+	crc = byteswap64(crc);
 #endif
 
 	return ~crc;
@@ -136,336 +81,40 @@ crc64_generic(const uint8_t *buf, size_t size, uint64_t crc)
 #endif
 
 
-/////////////////////
-// x86 CLMUL CRC64 //
-/////////////////////
-
-#ifdef CRC_CLMUL
-
-#include <immintrin.h>
-
-
-/*
-// These functions were used to generate the constants
-// at the top of crc64_clmul().
-static uint64_t
-calc_lo(uint64_t poly)
-{
-	uint64_t a = poly;
-	uint64_t b = 0;
-
-	for (unsigned i = 0; i < 64; ++i) {
-		b = (b >> 1) | (a << 63);
-		a = (a >> 1) ^ (a & 1 ? poly : 0);
-	}
-
-	return b;
-}
-
-static uint64_t
-calc_hi(uint64_t poly, uint64_t a)
-{
-	for (unsigned i = 0; i < 64; ++i)
-		a = (a >> 1) ^ (a & 1 ? poly : 0);
-
-	return a;
-}
-*/
-
-
-#define MASK_L(in, mask, r) \
-	r = _mm_shuffle_epi8(in, mask)
-
-#define MASK_H(in, mask, r) \
-	r = _mm_shuffle_epi8(in, _mm_xor_si128(mask, vsign))
-
-#define MASK_LH(in, mask, low, high) \
-	MASK_L(in, mask, low); \
-	MASK_H(in, mask, high)
-
-
-// MSVC (VS2015 - VS2022) produces bad 32-bit x86 code from the CLMUL CRC
-// code when optimizations are enabled (release build). According to the bug
-// report, the ebx register is corrupted and the calculated result is wrong.
-// Trying to workaround the problem with "__asm mov ebx, ebx" didn't help.
-// The following pragma works and performance is still good. x86-64 builds
-// aren't affected by this problem.
-//
-// NOTE: Another pragma after the function restores the optimizations.
-// If the #if condition here is updated, the other one must be updated too.
-#if defined(_MSC_VER) && !defined(__INTEL_COMPILER) && !defined(__clang__) \
-		&& defined(_M_IX86)
-#	pragma optimize("g", off)
-#endif
-
-// 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__)
-__attribute__((__target__("ssse3,sse4.1,pclmul")))
-#endif
-// The intrinsics use 16-byte-aligned reads from buf, thus they may read
-// up to 15 bytes before or after the buffer (depending on the alignment
-// of the buf argument). The values of the extra bytes are ignored.
-// This unavoidably trips -fsanitize=address so address sanitizier has
-// to be disabled for this function.
-#if lzma_has_attribute(__no_sanitize_address__)
-__attribute__((__no_sanitize_address__))
-#endif
-static uint64_t
-crc64_clmul(const uint8_t *buf, size_t size, uint64_t crc)
-{
-	// The prototypes of the intrinsics use signed types while most of
-	// the values are treated as unsigned here. These warnings in this
-	// function have been checked and found to be harmless so silence them.
-#if TUKLIB_GNUC_REQ(4, 6) || defined(__clang__)
-#	pragma GCC diagnostic push
-#	pragma GCC diagnostic ignored "-Wsign-conversion"
-#	pragma GCC diagnostic ignored "-Wconversion"
-#endif
-
-#ifndef CRC_USE_GENERIC_FOR_SMALL_INPUTS
-	// The code assumes that there is at least one byte of input.
-	if (size == 0)
-		return crc;
-#endif
-
-	// const uint64_t poly = 0xc96c5795d7870f42; // CRC polynomial
-	const uint64_t p  = 0x92d8af2baf0e1e85; // (poly << 1) | 1
-	const uint64_t mu = 0x9c3e466c172963d5; // (calc_lo(poly) << 1) | 1
-	const uint64_t k2 = 0xdabe95afc7875f40; // calc_hi(poly, 1)
-	const uint64_t k1 = 0xe05dd497ca393ae4; // calc_hi(poly, k2)
-	const __m128i vfold0 = _mm_set_epi64x(p, mu);
-	const __m128i vfold1 = _mm_set_epi64x(k2, k1);
-
-	// Create a vector with 8-bit values 0 to 15. This is used to
-	// construct control masks for _mm_blendv_epi8 and _mm_shuffle_epi8.
-	const __m128i vramp = _mm_setr_epi32(
-			0x03020100, 0x07060504, 0x0b0a0908, 0x0f0e0d0c);
-
-	// This is used to inverse the control mask of _mm_shuffle_epi8
-	// so that bytes that wouldn't be picked with the original mask
-	// will be picked and vice versa.
-	const __m128i vsign = _mm_set1_epi8(0x80);
-
-	// Memory addresses A to D and the distances between them:
-	//
-	//     A           B     C         D
-	//     [skip_start][size][skip_end]
-	//     [     size2      ]
-	//
-	// A and D are 16-byte aligned. B and C are 1-byte aligned.
-	// skip_start and skip_end are 0-15 bytes. size is at least 1 byte.
-	//
-	// A = aligned_buf will initially point to this address.
-	// B = The address pointed by the caller-supplied buf.
-	// C = buf + size == aligned_buf + size2
-	// D = buf + size + skip_end == aligned_buf + size2 + skip_end
-	const size_t skip_start = (size_t)((uintptr_t)buf & 15);
-	const size_t skip_end = (size_t)((0U - (uintptr_t)(buf + size)) & 15);
-	const __m128i *aligned_buf = (const __m128i *)(
-			(uintptr_t)buf & ~(uintptr_t)15);
-
-	// If size2 <= 16 then the whole input fits into a single 16-byte
-	// vector. If size2 > 16 then at least two 16-byte vectors must
-	// be processed. If size2 > 16 && size <= 16 then there is only
-	// one 16-byte vector's worth of input but it is unaligned in memory.
-	//
-	// NOTE: There is no integer overflow here if the arguments are valid.
-	// If this overflowed, buf + size would too.
-	size_t size2 = skip_start + size;
-
-	// Masks to be used with _mm_blendv_epi8 and _mm_shuffle_epi8:
-	// The first skip_start or skip_end bytes in the vectors will have
-	// the high bit (0x80) set. _mm_blendv_epi8 and _mm_shuffle_epi8
-	// will produce zeros for these positions. (Bitwise-xor of these
-	// masks with vsign will produce the opposite behavior.)
-	const __m128i mask_start
-			= _mm_sub_epi8(vramp, _mm_set1_epi8(skip_start));
-	const __m128i mask_end = _mm_sub_epi8(vramp, _mm_set1_epi8(skip_end));
-
-	// Get the first 1-16 bytes into data0. If loading less than 16 bytes,
-	// the bytes are loaded to the high bits of the vector and the least
-	// significant positions are filled with zeros.
-	const __m128i data0 = _mm_blendv_epi8(_mm_load_si128(aligned_buf),
-			_mm_setzero_si128(), mask_start);
-	++aligned_buf;
-
-#if defined(__i386__) || defined(_M_IX86)
-	const __m128i initial_crc = _mm_set_epi64x(0, ~crc);
-#else
-	// GCC and Clang would produce good code with _mm_set_epi64x
-	// but MSVC needs _mm_cvtsi64_si128 on x86-64.
-	const __m128i initial_crc = _mm_cvtsi64_si128(~crc);
-#endif
+#if defined(CRC64_GENERIC) && defined(CRC64_ARCH_OPTIMIZED)
 
-	__m128i v0, v1, v2, v3;
-
-#ifndef CRC_USE_GENERIC_FOR_SMALL_INPUTS
-	if (size <= 16) {
-		// Right-shift initial_crc by 1-16 bytes based on "size"
-		// and store the result in v1 (high bytes) and v0 (low bytes).
-		//
-		// NOTE: The highest 8 bytes of initial_crc are zeros so
-		// v1 will be filled with zeros if size >= 8. The highest 8
-		// bytes of v1 will always become zeros.
-		//
-		// [      v1      ][      v0      ]
-		//  [ initial_crc  ]                  size == 1
-		//   [ initial_crc  ]                 size == 2
-		//                [ initial_crc  ]    size == 15
-		//                 [ initial_crc  ]   size == 16 (all in v0)
-		const __m128i mask_low = _mm_add_epi8(
-				vramp, _mm_set1_epi8(size - 16));
-		MASK_LH(initial_crc, mask_low, v0, v1);
-
-		if (size2 <= 16) {
-			// There are 1-16 bytes of input and it is all
-			// in data0. Copy the input bytes to v3. If there
-			// are fewer than 16 bytes, the low bytes in v3
-			// will be filled with zeros. That is, the input
-			// bytes are stored to the same position as
-			// (part of) initial_crc is in v0.
-			MASK_L(data0, mask_end, v3);
-		} else {
-			// There are 2-16 bytes of input but not all bytes
-			// are in data0.
-			const __m128i data1 = _mm_load_si128(aligned_buf);
-
-			// Collect the 2-16 input bytes from data0 and data1
-			// to v2 and v3, and bitwise-xor them with the
-			// low bits of initial_crc in v0. Note that the
-			// the second xor is below this else-block as it
-			// is shared with the other branch.
-			MASK_H(data0, mask_end, v2);
-			MASK_L(data1, mask_end, v3);
-			v0 = _mm_xor_si128(v0, v2);
-		}
+//////////////////////////
+// Function dispatching //
+//////////////////////////
 
-		v0 = _mm_xor_si128(v0, v3);
-		v1 = _mm_alignr_epi8(v1, v0, 8);
-	} else
-#endif
-	{
-		const __m128i data1 = _mm_load_si128(aligned_buf);
-		MASK_LH(initial_crc, mask_start, v0, v1);
-		v0 = _mm_xor_si128(v0, data0);
-		v1 = _mm_xor_si128(v1, data1);
-
-#define FOLD \
-	v1 = _mm_xor_si128(v1, _mm_clmulepi64_si128(v0, vfold1, 0x00)); \
-	v0 = _mm_xor_si128(v1, _mm_clmulepi64_si128(v0, vfold1, 0x11));
-
-		while (size2 > 32) {
-			++aligned_buf;
-			size2 -= 16;
-			FOLD
-			v1 = _mm_load_si128(aligned_buf);
-		}
-
-		if (size2 < 32) {
-			MASK_H(v0, mask_end, v2);
-			MASK_L(v0, mask_end, v0);
-			MASK_L(v1, mask_end, v3);
-			v1 = _mm_or_si128(v2, v3);
-		}
-
-		FOLD
-		v1 = _mm_srli_si128(v0, 8);
-#undef FOLD
-	}
+// If both the generic and arch-optimized implementations are usable, then
+// the function that is used is selected at runtime. See crc32_fast.c.
 
-	v1 = _mm_xor_si128(_mm_clmulepi64_si128(v0, vfold1, 0x10), v1);
-	v0 = _mm_clmulepi64_si128(v1, vfold0, 0x00);
-	v2 = _mm_clmulepi64_si128(v0, vfold0, 0x10);
-	v0 = _mm_xor_si128(_mm_xor_si128(v2, _mm_slli_si128(v0, 8)), v1);
+typedef uint64_t (*crc64_func_type)(
+		const uint8_t *buf, size_t size, uint64_t crc);
 
-#if defined(__i386__) || defined(_M_IX86)
-	return ~(((uint64_t)(uint32_t)_mm_extract_epi32(v0, 3) << 32) |
-			(uint64_t)(uint32_t)_mm_extract_epi32(v0, 2));
-#else
-	return ~(uint64_t)_mm_extract_epi64(v0, 1);
-#endif
-
-#if TUKLIB_GNUC_REQ(4, 6) || defined(__clang__)
-#	pragma GCC diagnostic pop
-#endif
-}
-#if defined(_MSC_VER) && !defined(__INTEL_COMPILER) && !defined(__clang__) \
-		&& defined(_M_IX86)
-#	pragma optimize("", on)
-#endif
-#endif
-
-
-////////////////////////
-// Detect CPU support //
-////////////////////////
-
-#if defined(CRC_GENERIC) && defined(CRC_CLMUL)
-static inline bool
-is_clmul_supported(void)
+static crc64_func_type
+crc64_resolve(void)
 {
-	int success = 1;
-	uint32_t r[4]; // eax, ebx, ecx, edx
-
-#if defined(_MSC_VER)
-	// This needs <intrin.h> with MSVC. ICC has it as a built-in
-	// on all platforms.
-	__cpuid(r, 1);
-#elif defined(HAVE_CPUID_H)
-	// Compared to just using __asm__ to run CPUID, this also checks
-	// that CPUID is supported and saves and restores ebx as that is
-	// needed with GCC < 5 with position-independent code (PIC).
-	success = __get_cpuid(1, &r[0], &r[1], &r[2], &r[3]);
-#else
-	// Just a fallback that shouldn't be needed.
-	__asm__("cpuid\n\t"
-			: "=a"(r[0]), "=b"(r[1]), "=c"(r[2]), "=d"(r[3])
-			: "a"(1), "c"(0));
-#endif
-
-	// Returns true if these are supported:
-	// CLMUL (bit 1 in ecx)
-	// SSSE3 (bit 9 in ecx)
-	// SSE4.1 (bit 19 in ecx)
-	const uint32_t ecx_mask = (1 << 1) | (1 << 9) | (1 << 19);
-	return success && (r[2] & ecx_mask) == ecx_mask;
-
-	// Alternative methods that weren't used:
-	//   - ICC's _may_i_use_cpu_feature: the other methods should work too.
-	//   - GCC >= 6 / Clang / ICX __builtin_cpu_supports("pclmul")
-	//
-	// CPUID decding is needed with MSVC anyway and older GCC. This keeps
-	// the feature checks in the build system simpler too. The nice thing
-	// about __builtin_cpu_supports would be that it generates very short
-	// code as is it only reads a variable set at startup but a few bytes
-	// doesn't matter here.
+	return is_arch_extension_supported()
+			? &crc64_arch_optimized : &crc64_generic;
 }
 
-
 #ifdef HAVE_FUNC_ATTRIBUTE_CONSTRUCTOR
-#	define CRC64_FUNC_INIT
 #	define CRC64_SET_FUNC_ATTR __attribute__((__constructor__))
+static crc64_func_type crc64_func;
 #else
-#	define CRC64_FUNC_INIT = &crc64_dispatch
 #	define CRC64_SET_FUNC_ATTR
 static uint64_t crc64_dispatch(const uint8_t *buf, size_t size, uint64_t crc);
+static crc64_func_type crc64_func = &crc64_dispatch;
 #endif
 
 
-// Pointer to the the selected CRC64 method.
-static uint64_t (*crc64_func)(const uint8_t *buf, size_t size, uint64_t crc)
-		CRC64_FUNC_INIT;
-
-
 CRC64_SET_FUNC_ATTR
 static void
 crc64_set_func(void)
 {
-	crc64_func = is_clmul_supported() ? &crc64_clmul : &crc64_generic;
+	crc64_func = crc64_resolve();
 	return;
 }
 
@@ -474,12 +123,6 @@ crc64_set_func(void)
 static uint64_t
 crc64_dispatch(const uint8_t *buf, size_t size, uint64_t crc)
 {
-	// When __attribute__((__constructor__)) isn't supported, set the
-	// function pointer without any locking. If multiple threads run
-	// the detection code in parallel, they will all end up setting
-	// the pointer to the same value. This avoids the use of
-	// mythread_once() on every call to lzma_crc64() but this likely
-	// isn't strictly standards compliant. Let's change it if it breaks.
 	crc64_set_func();
 	return crc64_func(buf, size, crc);
 }
@@ -490,47 +133,22 @@ crc64_dispatch(const uint8_t *buf, size_t size, uint64_t crc)
 extern LZMA_API(uint64_t)
 lzma_crc64(const uint8_t *buf, size_t size, uint64_t crc)
 {
-#if defined(CRC_GENERIC) && defined(CRC_CLMUL)
-	// If CLMUL is available, it is the best for non-tiny inputs,
-	// being over twice as fast as the generic slice-by-four version.
-	// However, for size <= 16 it's different. In the extreme case
-	// of size == 1 the generic version can be five times faster.
-	// At size >= 8 the CLMUL starts to become reasonable. It
-	// varies depending on the alignment of buf too.
-	//
-	// The above doesn't include the overhead of mythread_once().
-	// At least on x86-64 GNU/Linux, pthread_once() is very fast but
-	// it still makes lzma_crc64(buf, 1, crc) 50-100 % slower. When
-	// size reaches 12-16 bytes the overhead becomes negligible.
-	//
-	// So using the generic version for size <= 16 may give better
-	// performance with tiny inputs but if such inputs happen rarely
-	// it's not so obvious because then the lookup table of the
-	// generic version may not be in the processor cache.
+#if defined(CRC64_GENERIC) && defined(CRC64_ARCH_OPTIMIZED)
+
 #ifdef CRC_USE_GENERIC_FOR_SMALL_INPUTS
 	if (size <= 16)
 		return crc64_generic(buf, size, crc);
 #endif
-
-/*
-#ifndef HAVE_FUNC_ATTRIBUTE_CONSTRUCTOR
-	// See crc64_dispatch(). This would be the alternative which uses
-	// locking and doesn't use crc64_dispatch(). Note that on Windows
-	// this method needs Vista threads.
-	mythread_once(crc64_set_func);
-#endif
-*/
-
 	return crc64_func(buf, size, crc);
 
-#elif defined(CRC_CLMUL)
-	// If CLMUL is used unconditionally without runtime CPU detection
-	// then omitting the generic version and its 8 KiB lookup table
-	// makes the library smaller.
+#elif defined(CRC64_ARCH_OPTIMIZED)
+	// If arch-optimized version is used unconditionally without runtime
+	// CPU detection then omitting the generic version and its 8 KiB
+	// lookup table makes the library smaller.
 	//
 	// FIXME: Lookup table isn't currently omitted on 32-bit x86,
 	// see crc64_table.c.
-	return crc64_clmul(buf, size, crc);
+	return crc64_arch_optimized(buf, size, crc);
 
 #else
 	return crc64_generic(buf, size, crc);
diff --git a/src/liblzma/check/crc64_small.c b/src/liblzma/check/crc64_small.c
index 420f7cf..ee4ea26 100644
--- a/src/liblzma/check/crc64_small.c
+++ b/src/liblzma/check/crc64_small.c
@@ -1,3 +1,5 @@
+// SPDX-License-Identifier: 0BSD
+
 ///////////////////////////////////////////////////////////////////////////////
 //
 /// \file       crc64_small.c
@@ -5,9 +7,6 @@
 //
 //  Author:     Lasse Collin
 //
-//  This file has been put into the public domain.
-//  You can do whatever you want with this file.
-//
 ///////////////////////////////////////////////////////////////////////////////
 
 #include "check.h"
diff --git a/src/liblzma/check/crc64_table.c b/src/liblzma/check/crc64_table.c
index 688e527..78e4275 100644
--- a/src/liblzma/check/crc64_table.c
+++ b/src/liblzma/check/crc64_table.c
@@ -1,3 +1,5 @@
+// SPDX-License-Identifier: 0BSD
+
 ///////////////////////////////////////////////////////////////////////////////
 //
 /// \file       crc64_table.c
@@ -5,19 +7,21 @@
 //
 //  Author:     Lasse Collin
 //
-//  This file has been put into the public domain.
-//  You can do whatever you want with this file.
-//
 ///////////////////////////////////////////////////////////////////////////////
 
 #include "common.h"
 
 
-// FIXME: Compared to crc64_fast.c this has to check for __x86_64__ too
+// FIXME: Compared to crc_common.h this has to check for __x86_64__ too
 // so that in 32-bit builds crc64_x86.S won't break due to a missing table.
-#if (defined(__x86_64__) && defined(__SSSE3__) \
+#if defined(HAVE_USABLE_CLMUL) && ((defined(__x86_64__) && defined(__SSSE3__) \
 			&& defined(__SSE4_1__) && defined(__PCLMUL__)) \
-		|| (defined(__e2k__) && __iset__ >= 6)
+		|| (defined(__e2k__) && __iset__ >= 6))
+#	define NO_CRC64_TABLE
+#endif
+
+
+#ifdef NO_CRC64_TABLE
 // No table needed. Use a typedef to avoid an empty translation unit.
 typedef void lzma_crc64_dummy;
 
diff --git a/src/liblzma/check/crc64_table_be.h b/src/liblzma/check/crc64_table_be.h
index ea074f3..db76cc7 100644
--- a/src/liblzma/check/crc64_table_be.h
+++ b/src/liblzma/check/crc64_table_be.h
@@ -1,4 +1,6 @@
-/* This file has been automatically generated by crc64_tablegen.c. */
+// SPDX-License-Identifier: 0BSD
+
+// This file has been generated by crc64_tablegen.c.
 
 const uint64_t lzma_crc64_table[4][256] = {
 	{
diff --git a/src/liblzma/check/crc64_table_le.h b/src/liblzma/check/crc64_table_le.h
index 1196b31..e40a8c8 100644
--- a/src/liblzma/check/crc64_table_le.h
+++ b/src/liblzma/check/crc64_table_le.h
@@ -1,4 +1,6 @@
-/* This file has been automatically generated by crc64_tablegen.c. */
+// SPDX-License-Identifier: 0BSD
+
+// This file has been generated by crc64_tablegen.c.
 
 const uint64_t lzma_crc64_table[4][256] = {
 	{
diff --git a/src/liblzma/check/crc64_tablegen.c b/src/liblzma/check/crc64_tablegen.c
index fddaa7e..2035127 100644
--- a/src/liblzma/check/crc64_tablegen.c
+++ b/src/liblzma/check/crc64_tablegen.c
@@ -1,3 +1,5 @@
+// SPDX-License-Identifier: 0BSD
+
 ///////////////////////////////////////////////////////////////////////////////
 //
 /// \file       crc64_tablegen.c
@@ -8,9 +10,6 @@
 //
 //  Author:     Lasse Collin
 //
-//  This file has been put into the public domain.
-//  You can do whatever you want with this file.
-//
 ///////////////////////////////////////////////////////////////////////////////
 
 #include <stdio.h>
@@ -43,7 +42,7 @@ init_crc64_table(void)
 #ifdef WORDS_BIGENDIAN
 	for (size_t s = 0; s < 4; ++s)
 		for (size_t b = 0; b < 256; ++b)
-			crc64_table[s][b] = bswap64(crc64_table[s][b]);
+			crc64_table[s][b] = byteswap64(crc64_table[s][b]);
 #endif
 
 	return;
@@ -53,9 +52,11 @@ init_crc64_table(void)
 static void
 print_crc64_table(void)
 {
-	printf("/* This file has been automatically generated by "
-			"crc64_tablegen.c. */\n\n"
-			"const uint64_t lzma_crc64_table[4][256] = {\n\t{");
+	// Split the SPDX string so that it won't accidentally match
+	// when tools search for the string.
+	printf("// SPDX" "-License-Identifier" ": 0BSD\n\n"
+		"// This file has been generated by crc64_tablegen.c.\n\n"
+		"const uint64_t lzma_crc64_table[4][256] = {\n\t{");
 
 	for (size_t s = 0; s < 4; ++s) {
 		for (size_t b = 0; b < 256; ++b) {
diff --git a/src/liblzma/check/crc64_x86.S b/src/liblzma/check/crc64_x86.S
index 9aecf58..47f6081 100644
--- a/src/liblzma/check/crc64_x86.S
+++ b/src/liblzma/check/crc64_x86.S
@@ -1,3 +1,5 @@
+/* SPDX-License-Identifier: 0BSD */
+
 /*
  * Speed-optimized CRC64 using slicing-by-four algorithm
  *
@@ -7,9 +9,6 @@
  * Authors: Igor Pavlov (original CRC32 assembly code)
  *          Lasse Collin (CRC64 adaptation of the modified CRC32 code)
  *
- * This file has been put into the public domain.
- * You can do whatever you want with this file.
- *
  * This code needs lzma_crc64_table, which can be created using the
  * following C code:
 
diff --git a/src/liblzma/check/crc_common.h b/src/liblzma/check/crc_common.h
new file mode 100644
index 0000000..63a7b5c
--- /dev/null
+++ b/src/liblzma/check/crc_common.h
@@ -0,0 +1,137 @@
+// SPDX-License-Identifier: 0BSD
+
+///////////////////////////////////////////////////////////////////////////////
+//
+/// \file       crc_common.h
+/// \brief      Some functions and macros for CRC32 and CRC64
+//
+//  Authors:    Lasse Collin
+//              Ilya Kurdyukov
+//              Hans Jansen
+//              Jia Tan
+//
+///////////////////////////////////////////////////////////////////////////////
+
+#ifndef LZMA_CRC_COMMON_H
+#define LZMA_CRC_COMMON_H
+
+#include "common.h"
+
+
+#ifdef WORDS_BIGENDIAN
+#	define A(x) ((x) >> 24)
+#	define B(x) (((x) >> 16) & 0xFF)
+#	define C(x) (((x) >> 8) & 0xFF)
+#	define D(x) ((x) & 0xFF)
+
+#	define S8(x) ((x) << 8)
+#	define S32(x) ((x) << 32)
+
+#else
+#	define A(x) ((x) & 0xFF)
+#	define B(x) (((x) >> 8) & 0xFF)
+#	define C(x) (((x) >> 16) & 0xFF)
+#	define D(x) ((x) >> 24)
+
+#	define S8(x) ((x) >> 8)
+#	define S32(x) ((x) >> 32)
+#endif
+
+
+// CRC CLMUL code needs this because accessing input buffers that aren't
+// aligned to the vector size will inherently trip the address sanitizer.
+#if lzma_has_attribute(__no_sanitize_address__)
+#	define crc_attr_no_sanitize_address \
+			__attribute__((__no_sanitize_address__))
+#else
+#	define crc_attr_no_sanitize_address
+#endif
+
+// Keep this in sync with changes to crc32_arm64.h
+#if defined(_WIN32) || defined(HAVE_GETAUXVAL) \
+		|| defined(HAVE_ELF_AUX_INFO) \
+		|| (defined(__APPLE__) && defined(HAVE_SYSCTLBYNAME))
+#	define ARM64_RUNTIME_DETECTION 1
+#endif
+
+
+#undef CRC32_GENERIC
+#undef CRC64_GENERIC
+
+#undef CRC32_ARCH_OPTIMIZED
+#undef CRC64_ARCH_OPTIMIZED
+
+// The x86 CLMUL is used for both CRC32 and CRC64.
+#undef CRC_X86_CLMUL
+
+#undef CRC32_ARM64
+#undef CRC64_ARM64_CLMUL
+
+#undef CRC_USE_GENERIC_FOR_SMALL_INPUTS
+
+// ARM64 CRC32 instruction is only useful for CRC32. Currently, only
+// little endian is supported since we were unable to test on a big
+// endian machine.
+//
+// NOTE: Keep this and the next check in sync with the macro
+//       NO_CRC32_TABLE in crc32_table.c
+#if defined(HAVE_ARM64_CRC32) && !defined(WORDS_BIGENDIAN)
+// Allow ARM64 CRC32 instruction without a runtime check if
+// __ARM_FEATURE_CRC32 is defined. GCC and Clang only define this if the
+// proper compiler options are used.
+#	if defined(__ARM_FEATURE_CRC32)
+#		define CRC32_ARCH_OPTIMIZED 1
+#		define CRC32_ARM64 1
+#	elif defined(ARM64_RUNTIME_DETECTION)
+#		define CRC32_ARCH_OPTIMIZED 1
+#		define CRC32_ARM64 1
+#		define CRC32_GENERIC 1
+#	endif
+#endif
+
+#if defined(HAVE_USABLE_CLMUL)
+// 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 in sync with the NO_CRC32_TABLE macro in crc32_table.c
+// and NO_CRC64_TABLE in crc64_table.c.
+#	if (defined(__SSSE3__) && defined(__SSE4_1__) && defined(__PCLMUL__)) \
+		|| (defined(__e2k__) && __iset__ >= 6)
+#		define CRC32_ARCH_OPTIMIZED 1
+#		define CRC64_ARCH_OPTIMIZED 1
+#		define CRC_X86_CLMUL 1
+#	else
+#		define CRC32_GENERIC 1
+#		define CRC64_GENERIC 1
+#		define CRC32_ARCH_OPTIMIZED 1
+#		define CRC64_ARCH_OPTIMIZED 1
+#		define CRC_X86_CLMUL 1
+
+/*
+		// The generic code is much faster with 1-8-byte inputs and
+		// has similar performance up to 16 bytes  at least in
+		// microbenchmarks (it depends on input buffer alignment
+		// too). If both versions are built, this #define will use
+		// the generic version for inputs up to 16 bytes and CLMUL
+		// for bigger inputs. It saves a little in code size since
+		// the special cases for 0-16-byte inputs will be omitted
+		// from the CLMUL code.
+#		define CRC_USE_GENERIC_FOR_SMALL_INPUTS 1
+*/
+#	endif
+#endif
+
+// For CRC32 use the generic slice-by-eight implementation if no optimized
+// version is available.
+#if !defined(CRC32_ARCH_OPTIMIZED) && !defined(CRC32_GENERIC)
+#	define CRC32_GENERIC 1
+#endif
+
+// For CRC64 use the generic slice-by-four implementation if no optimized
+// version is available.
+#if !defined(CRC64_ARCH_OPTIMIZED) && !defined(CRC64_GENERIC)
+#	define CRC64_GENERIC 1
+#endif
+
+#endif
diff --git a/src/liblzma/check/crc_macros.h b/src/liblzma/check/crc_macros.h
deleted file mode 100644
index a7c21b7..0000000
--- a/src/liblzma/check/crc_macros.h
+++ /dev/null
@@ -1,30 +0,0 @@
-///////////////////////////////////////////////////////////////////////////////
-//
-/// \file       crc_macros.h
-/// \brief      Some endian-dependent macros for CRC32 and CRC64
-//
-//  Author:     Lasse Collin
-//
-//  This file has been put into the public domain.
-//  You can do whatever you want with this file.
-//
-///////////////////////////////////////////////////////////////////////////////
-
-#ifdef WORDS_BIGENDIAN
-#	define A(x) ((x) >> 24)
-#	define B(x) (((x) >> 16) & 0xFF)
-#	define C(x) (((x) >> 8) & 0xFF)
-#	define D(x) ((x) & 0xFF)
-
-#	define S8(x) ((x) << 8)
-#	define S32(x) ((x) << 32)
-
-#else
-#	define A(x) ((x) & 0xFF)
-#	define B(x) (((x) >> 8) & 0xFF)
-#	define C(x) (((x) >> 16) & 0xFF)
-#	define D(x) ((x) >> 24)
-
-#	define S8(x) ((x) >> 8)
-#	define S32(x) ((x) >> 32)
-#endif
diff --git a/src/liblzma/check/crc_x86_clmul.h b/src/liblzma/check/crc_x86_clmul.h
new file mode 100644
index 0000000..f1254ec
--- /dev/null
+++ b/src/liblzma/check/crc_x86_clmul.h
@@ -0,0 +1,428 @@
+// SPDX-License-Identifier: 0BSD
+
+///////////////////////////////////////////////////////////////////////////////
+//
+/// \file       crc_x86_clmul.h
+/// \brief      CRC32 and CRC64 implementations using CLMUL instructions.
+///
+/// The CRC32 and CRC64 implementations use 32/64-bit x86 SSSE3, SSE4.1, and
+/// CLMUL instructions. This is compatible with Elbrus 2000 (E2K) too.
+///
+/// They were derived from
+/// https://www.researchgate.net/publication/263424619_Fast_CRC_computation
+/// and the public domain code from https://github.com/rawrunprotected/crc
+/// (URLs were checked on 2023-10-14).
+///
+/// While this file has both CRC32 and CRC64 implementations, only one
+/// should be built at a time to ensure that crc_simd_body() is inlined
+/// even with compilers with which lzma_always_inline expands to plain inline.
+/// The version to build is selected by defining BUILDING_CRC32_CLMUL or
+/// BUILDING_CRC64_CLMUL before including this file.
+///
+/// FIXME: Builds for 32-bit x86 use the assembly .S files by default
+/// unless configured with --disable-assembler. Even then the lookup table
+/// isn't omitted in crc64_table.c since it doesn't know that assembly
+/// code has been disabled.
+//
+//  Authors:    Ilya Kurdyukov
+//              Hans Jansen
+//              Lasse Collin
+//              Jia Tan
+//
+///////////////////////////////////////////////////////////////////////////////
+
+// This file must not be included more than once.
+#ifdef LZMA_CRC_X86_CLMUL_H
+#	error crc_x86_clmul.h was included twice.
+#endif
+#define LZMA_CRC_X86_CLMUL_H
+
+#include <immintrin.h>
+
+#if defined(_MSC_VER)
+#	include <intrin.h>
+#elif defined(HAVE_CPUID_H)
+#	include <cpuid.h>
+#endif
+
+
+// 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__("ssse3,sse4.1,pclmul")))
+#else
+#	define crc_attr_target
+#endif
+
+
+#define MASK_L(in, mask, r) r = _mm_shuffle_epi8(in, mask)
+
+#define MASK_H(in, mask, r) \
+	r = _mm_shuffle_epi8(in, _mm_xor_si128(mask, vsign))
+
+#define MASK_LH(in, mask, low, high) \
+	MASK_L(in, mask, low); \
+	MASK_H(in, mask, high)
+
+
+crc_attr_target
+crc_attr_no_sanitize_address
+static lzma_always_inline void
+crc_simd_body(const uint8_t *buf, const size_t size, __m128i *v0, __m128i *v1,
+		const __m128i vfold16, const __m128i initial_crc)
+{
+	// Create a vector with 8-bit values 0 to 15. This is used to
+	// construct control masks for _mm_blendv_epi8 and _mm_shuffle_epi8.
+	const __m128i vramp = _mm_setr_epi32(
+			0x03020100, 0x07060504, 0x0b0a0908, 0x0f0e0d0c);
+
+	// This is used to inverse the control mask of _mm_shuffle_epi8
+	// so that bytes that wouldn't be picked with the original mask
+	// will be picked and vice versa.
+	const __m128i vsign = _mm_set1_epi8(-0x80);
+
+	// Memory addresses A to D and the distances between them:
+	//
+	//     A           B     C         D
+	//     [skip_start][size][skip_end]
+	//     [     size2      ]
+	//
+	// A and D are 16-byte aligned. B and C are 1-byte aligned.
+	// skip_start and skip_end are 0-15 bytes. size is at least 1 byte.
+	//
+	// A = aligned_buf will initially point to this address.
+	// B = The address pointed by the caller-supplied buf.
+	// C = buf + size == aligned_buf + size2
+	// D = buf + size + skip_end == aligned_buf + size2 + skip_end
+	const size_t skip_start = (size_t)((uintptr_t)buf & 15);
+	const size_t skip_end = (size_t)((0U - (uintptr_t)(buf + size)) & 15);
+	const __m128i *aligned_buf = (const __m128i *)(
+			(uintptr_t)buf & ~(uintptr_t)15);
+
+	// If size2 <= 16 then the whole input fits into a single 16-byte
+	// vector. If size2 > 16 then at least two 16-byte vectors must
+	// be processed. If size2 > 16 && size <= 16 then there is only
+	// one 16-byte vector's worth of input but it is unaligned in memory.
+	//
+	// NOTE: There is no integer overflow here if the arguments
+	// are valid. If this overflowed, buf + size would too.
+	const size_t size2 = skip_start + size;
+
+	// Masks to be used with _mm_blendv_epi8 and _mm_shuffle_epi8:
+	// The first skip_start or skip_end bytes in the vectors will have
+	// the high bit (0x80) set. _mm_blendv_epi8 and _mm_shuffle_epi8
+	// will produce zeros for these positions. (Bitwise-xor of these
+	// masks with vsign will produce the opposite behavior.)
+	const __m128i mask_start
+			= _mm_sub_epi8(vramp, _mm_set1_epi8((char)skip_start));
+	const __m128i mask_end
+			= _mm_sub_epi8(vramp, _mm_set1_epi8((char)skip_end));
+
+	// Get the first 1-16 bytes into data0. If loading less than 16
+	// bytes, the bytes are loaded to the high bits of the vector and
+	// the least significant positions are filled with zeros.
+	const __m128i data0 = _mm_blendv_epi8(_mm_load_si128(aligned_buf),
+			_mm_setzero_si128(), mask_start);
+	aligned_buf++;
+
+	__m128i v2, v3;
+
+#ifndef CRC_USE_GENERIC_FOR_SMALL_INPUTS
+	if (size <= 16) {
+		// Right-shift initial_crc by 1-16 bytes based on "size"
+		// and store the result in v1 (high bytes) and v0 (low bytes).
+		//
+		// NOTE: The highest 8 bytes of initial_crc are zeros so
+		// v1 will be filled with zeros if size >= 8. The highest
+		// 8 bytes of v1 will always become zeros.
+		//
+		// [      v1      ][      v0      ]
+		//  [ initial_crc  ]                  size == 1
+		//   [ initial_crc  ]                 size == 2
+		//                [ initial_crc  ]    size == 15
+		//                 [ initial_crc  ]   size == 16 (all in v0)
+		const __m128i mask_low = _mm_add_epi8(
+				vramp, _mm_set1_epi8((char)(size - 16)));
+		MASK_LH(initial_crc, mask_low, *v0, *v1);
+
+		if (size2 <= 16) {
+			// There are 1-16 bytes of input and it is all
+			// in data0. Copy the input bytes to v3. If there
+			// are fewer than 16 bytes, the low bytes in v3
+			// will be filled with zeros. That is, the input
+			// bytes are stored to the same position as
+			// (part of) initial_crc is in v0.
+			MASK_L(data0, mask_end, v3);
+		} else {
+			// There are 2-16 bytes of input but not all bytes
+			// are in data0.
+			const __m128i data1 = _mm_load_si128(aligned_buf);
+
+			// Collect the 2-16 input bytes from data0 and data1
+			// to v2 and v3, and bitwise-xor them with the
+			// low bits of initial_crc in v0. Note that the
+			// the second xor is below this else-block as it
+			// is shared with the other branch.
+			MASK_H(data0, mask_end, v2);
+			MASK_L(data1, mask_end, v3);
+			*v0 = _mm_xor_si128(*v0, v2);
+		}
+
+		*v0 = _mm_xor_si128(*v0, v3);
+		*v1 = _mm_alignr_epi8(*v1, *v0, 8);
+	} else
+#endif
+	{
+		// There is more than 16 bytes of input.
+		const __m128i data1 = _mm_load_si128(aligned_buf);
+		const __m128i *end = (const __m128i*)(
+				(const char *)aligned_buf - 16 + size2);
+		aligned_buf++;
+
+		MASK_LH(initial_crc, mask_start, *v0, *v1);
+		*v0 = _mm_xor_si128(*v0, data0);
+		*v1 = _mm_xor_si128(*v1, data1);
+
+		while (aligned_buf < end) {
+			*v1 = _mm_xor_si128(*v1, _mm_clmulepi64_si128(
+					*v0, vfold16, 0x00));
+			*v0 = _mm_xor_si128(*v1, _mm_clmulepi64_si128(
+					*v0, vfold16, 0x11));
+			*v1 = _mm_load_si128(aligned_buf++);
+		}
+
+		if (aligned_buf != end) {
+			MASK_H(*v0, mask_end, v2);
+			MASK_L(*v0, mask_end, *v0);
+			MASK_L(*v1, mask_end, v3);
+			*v1 = _mm_or_si128(v2, v3);
+		}
+
+		*v1 = _mm_xor_si128(*v1, _mm_clmulepi64_si128(
+				*v0, vfold16, 0x00));
+		*v0 = _mm_xor_si128(*v1, _mm_clmulepi64_si128(
+				*v0, vfold16, 0x11));
+		*v1 = _mm_srli_si128(*v0, 8);
+	}
+}
+
+
+/////////////////////
+// x86 CLMUL CRC32 //
+/////////////////////
+
+/*
+// These functions were used to generate the constants
+// at the top of crc32_arch_optimized().
+static uint64_t
+calc_lo(uint64_t p, uint64_t a, int n)
+{
+	uint64_t b = 0; int i;
+	for (i = 0; i < n; i++) {
+		b = b >> 1 | (a & 1) << (n - 1);
+		a = (a >> 1) ^ ((0 - (a & 1)) & p);
+	}
+	return b;
+}
+
+// same as ~crc(&a, sizeof(a), ~0)
+static uint64_t
+calc_hi(uint64_t p, uint64_t a, int n)
+{
+	int i;
+	for (i = 0; i < n; i++)
+		a = (a >> 1) ^ ((0 - (a & 1)) & p);
+	return a;
+}
+*/
+
+#ifdef BUILDING_CRC32_CLMUL
+
+crc_attr_target
+crc_attr_no_sanitize_address
+static uint32_t
+crc32_arch_optimized(const uint8_t *buf, size_t size, uint32_t crc)
+{
+#ifndef CRC_USE_GENERIC_FOR_SMALL_INPUTS
+	// The code assumes that there is at least one byte of input.
+	if (size == 0)
+		return crc;
+#endif
+
+	// uint32_t poly = 0xedb88320;
+	const int64_t p = 0x1db710640; // p << 1
+	const int64_t mu = 0x1f7011641; // calc_lo(p, p, 32) << 1 | 1
+	const int64_t k5 = 0x163cd6124; // calc_hi(p, p, 32) << 1
+	const int64_t k4 = 0x0ccaa009e; // calc_hi(p, p, 64) << 1
+	const int64_t k3 = 0x1751997d0; // calc_hi(p, p, 128) << 1
+
+	const __m128i vfold4 = _mm_set_epi64x(mu, p);
+	const __m128i vfold8 = _mm_set_epi64x(0, k5);
+	const __m128i vfold16 = _mm_set_epi64x(k4, k3);
+
+	__m128i v0, v1, v2;
+
+	crc_simd_body(buf,  size, &v0, &v1, vfold16,
+			_mm_cvtsi32_si128((int32_t)~crc));
+
+	v1 = _mm_xor_si128(
+			_mm_clmulepi64_si128(v0, vfold16, 0x10), v1); // xxx0
+	v2 = _mm_shuffle_epi32(v1, 0xe7); // 0xx0
+	v0 = _mm_slli_epi64(v1, 32);  // [0]
+	v0 = _mm_clmulepi64_si128(v0, vfold8, 0x00);
+	v0 = _mm_xor_si128(v0, v2);   // [1] [2]
+	v2 = _mm_clmulepi64_si128(v0, vfold4, 0x10);
+	v2 = _mm_clmulepi64_si128(v2, vfold4, 0x00);
+	v0 = _mm_xor_si128(v0, v2);   // [2]
+	return ~(uint32_t)_mm_extract_epi32(v0, 2);
+}
+#endif // BUILDING_CRC32_CLMUL
+
+
+/////////////////////
+// x86 CLMUL CRC64 //
+/////////////////////
+
+/*
+// These functions were used to generate the constants
+// at the top of crc64_arch_optimized().
+static uint64_t
+calc_lo(uint64_t poly)
+{
+	uint64_t a = poly;
+	uint64_t b = 0;
+
+	for (unsigned i = 0; i < 64; ++i) {
+		b = (b >> 1) | (a << 63);
+		a = (a >> 1) ^ (a & 1 ? poly : 0);
+	}
+
+	return b;
+}
+
+static uint64_t
+calc_hi(uint64_t poly, uint64_t a)
+{
+	for (unsigned i = 0; i < 64; ++i)
+		a = (a >> 1) ^ (a & 1 ? poly : 0);
+
+	return a;
+}
+*/
+
+#ifdef BUILDING_CRC64_CLMUL
+
+// MSVC (VS2015 - VS2022) produces bad 32-bit x86 code from the CLMUL CRC
+// code when optimizations are enabled (release build). According to the bug
+// report, the ebx register is corrupted and the calculated result is wrong.
+// Trying to workaround the problem with "__asm mov ebx, ebx" didn't help.
+// The following pragma works and performance is still good. x86-64 builds
+// and CRC32 CLMUL aren't affected by this problem. The problem does not
+// happen in crc_simd_body() either (which is shared with CRC32 CLMUL anyway).
+//
+// NOTE: Another pragma after crc64_arch_optimized() restores
+// the optimizations. If the #if condition here is updated,
+// the other one must be updated too.
+#if defined(_MSC_VER) && !defined(__INTEL_COMPILER) && !defined(__clang__) \
+		&& defined(_M_IX86)
+#	pragma optimize("g", off)
+#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)
+{
+#ifndef CRC_USE_GENERIC_FOR_SMALL_INPUTS
+	// The code assumes that there is at least one byte of input.
+	if (size == 0)
+		return crc;
+#endif
+
+	// const uint64_t poly = 0xc96c5795d7870f42; // CRC polynomial
+	const uint64_t p  = 0x92d8af2baf0e1e85; // (poly << 1) | 1
+	const uint64_t mu = 0x9c3e466c172963d5; // (calc_lo(poly) << 1) | 1
+	const uint64_t k2 = 0xdabe95afc7875f40; // calc_hi(poly, 1)
+	const uint64_t k1 = 0xe05dd497ca393ae4; // calc_hi(poly, k2)
+
+	const __m128i vfold8 = _mm_set_epi64x((int64_t)p, (int64_t)mu);
+	const __m128i vfold16 = _mm_set_epi64x((int64_t)k2, (int64_t)k1);
+
+	__m128i v0, v1, v2;
+
+#if defined(__i386__) || defined(_M_IX86)
+	crc_simd_body(buf,  size, &v0, &v1, vfold16,
+			_mm_set_epi64x(0, (int64_t)~crc));
+#else
+	// GCC and Clang would produce good code with _mm_set_epi64x
+	// but MSVC needs _mm_cvtsi64_si128 on x86-64.
+	crc_simd_body(buf,  size, &v0, &v1, vfold16,
+			_mm_cvtsi64_si128((int64_t)~crc));
+#endif
+
+	v1 = _mm_xor_si128(_mm_clmulepi64_si128(v0, vfold16, 0x10), v1);
+	v0 = _mm_clmulepi64_si128(v1, vfold8, 0x00);
+	v2 = _mm_clmulepi64_si128(v0, vfold8, 0x10);
+	v0 = _mm_xor_si128(_mm_xor_si128(v1, _mm_slli_si128(v0, 8)), v2);
+
+#if defined(__i386__) || defined(_M_IX86)
+	return ~(((uint64_t)(uint32_t)_mm_extract_epi32(v0, 3) << 32) |
+			(uint64_t)(uint32_t)_mm_extract_epi32(v0, 2));
+#else
+	return ~(uint64_t)_mm_extract_epi64(v0, 1);
+#endif
+}
+
+#if defined(_MSC_VER) && !defined(__INTEL_COMPILER) && !defined(__clang__) \
+		&& defined(_M_IX86)
+#	pragma optimize("", on)
+#endif
+
+#endif // BUILDING_CRC64_CLMUL
+
+
+// Inlining this function duplicates the function body in crc32_resolve() and
+// crc64_resolve(), but this is acceptable because this is a tiny function.
+static inline bool
+is_arch_extension_supported(void)
+{
+	int success = 1;
+	uint32_t r[4]; // eax, ebx, ecx, edx
+
+#if defined(_MSC_VER)
+	// This needs <intrin.h> with MSVC. ICC has it as a built-in
+	// on all platforms.
+	__cpuid(r, 1);
+#elif defined(HAVE_CPUID_H)
+	// Compared to just using __asm__ to run CPUID, this also checks
+	// that CPUID is supported and saves and restores ebx as that is
+	// needed with GCC < 5 with position-independent code (PIC).
+	success = __get_cpuid(1, &r[0], &r[1], &r[2], &r[3]);
+#else
+	// Just a fallback that shouldn't be needed.
+	__asm__("cpuid\n\t"
+			: "=a"(r[0]), "=b"(r[1]), "=c"(r[2]), "=d"(r[3])
+			: "a"(1), "c"(0));
+#endif
+
+	// Returns true if these are supported:
+	// CLMUL (bit 1 in ecx)
+	// SSSE3 (bit 9 in ecx)
+	// SSE4.1 (bit 19 in ecx)
+	const uint32_t ecx_mask = (1 << 1) | (1 << 9) | (1 << 19);
+	return success && (r[2] & ecx_mask) == ecx_mask;
+
+	// Alternative methods that weren't used:
+	//   - ICC's _may_i_use_cpu_feature: the other methods should work too.
+	//   - GCC >= 6 / Clang / ICX __builtin_cpu_supports("pclmul")
+	//
+	// CPUID decding is needed with MSVC anyway and older GCC. This keeps
+	// the feature checks in the build system simpler too. The nice thing
+	// about __builtin_cpu_supports would be that it generates very short
+	// code as is it only reads a variable set at startup but a few bytes
+	// doesn't matter here.
+}
diff --git a/src/liblzma/check/sha256.c b/src/liblzma/check/sha256.c
index 6feb342..bd0d280 100644
--- a/src/liblzma/check/sha256.c
+++ b/src/liblzma/check/sha256.c
@@ -1,24 +1,17 @@
+// SPDX-License-Identifier: 0BSD
+
 ///////////////////////////////////////////////////////////////////////////////
 //
 /// \file       sha256.c
 /// \brief      SHA-256
-///
-/// \todo       Crypto++ has x86 ASM optimizations. They use SSE so if they
-///             are imported to liblzma, SSE instructions need to be used
-///             conditionally to keep the code working on older boxes.
 //
-//  This code is based on the code found from 7-Zip, which has a modified
-//  version of the SHA-256 found from Crypto++ <https://www.cryptopp.com/>.
-//  The code was modified a little to fit into liblzma.
+//  The C code is based on the public domain SHA-256 code found from
+//  Crypto++ Library 5.5.1 released in 2007: https://www.cryptopp.com/
+//  A few minor tweaks have been made in liblzma.
 //
-//  Authors:    Kevin Springle
-//              Wei Dai
-//              Igor Pavlov
+//  Authors:    Wei Dai
 //              Lasse Collin
 //
-//  This file has been put into the public domain.
-//  You can do whatever you want with this file.
-//
 ///////////////////////////////////////////////////////////////////////////////
 
 #include "check.h"