liblzma: Add fast CRC64 for 32/64-bit x86 using SSSE3 + SSE4.1 + CLMUL.

It also works on E2K as it supports these intrinsics.

On x86-64 runtime detection is used so the code keeps working on
older processors too. A CLMUL-only build can be done by using
-msse4.1 -mpclmul in CFLAGS and this will reduce the library
size since the generic implementation and its 8 KiB lookup table
will be omitted.

On 32-bit x86 this isn't used by default for now because by default
on 32-bit x86 the separate assembly file crc64_x86.S is used.
If --disable-assembler is used then this new CLMUL code is used
the same way as on 64-bit x86. However, a CLMUL-only build
(-msse4.1 -mpclmul) won't omit the 8 KiB lookup table on
32-bit x86 due to a currently-missing check for disabled
assembler usage.

The configure.ac check should be such that the code won't be
built if something in the toolchain doesn't support it but
--disable-clmul-crc option can be used to unconditionally
disable this feature.

CLMUL speeds up decompression of files that have compressed very
well (assuming CRC64 is used as a check type). It is know that
the CLMUL code is significantly slower than the generic code for
tiny inputs (especially 1-8 bytes but up to 16 bytes). If that
is a real-world problem then there is already a commented-out
variant that uses the generic version for small inputs.

Thanks to Ilya Kurdyukov for the original patch which was
derived from a white paper from Intel [1] (published in 2009)
and public domain code from [2] (released in 2016).

[1] https://www.intel.com/content/dam/www/public/us/en/documents/white-papers/fast-crc-computation-generic-polynomials-pclmulqdq-paper.pdf
[2] https://github.com/rawrunprotected/crc

1 files changed, 444 insertions, 5 deletions

diff --git a/src/liblzma/check/crc64_fast.c b/src/liblzma/check/crc64_fast.c
index cf18ebc5..e3cbf1b1 100644
--- a/src/liblzma/check/crc64_fast.c
+++ b/src/liblzma/check/crc64_fast.c
@@ -3,11 +3,25 @@
 /// \file       crc64.c
 /// \brief      CRC64 calculation
 ///
-/// Calculate the CRC64 using 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
+/// 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.
 //
-//  Author:     Lasse Collin
+//  Authors:    Lasse Collin
+//              Ilya Kurdyukov
 //
 //  This file has been put into the public domain.
 //  You can do whatever you want with this file.
@@ -15,6 +29,54 @@
 ///////////////////////////////////////////////////////////////////////////////
 
 #include "check.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
+#endif
+
+
+/////////////////////////////////
+// Generic slice-by-four CRC64 //
+/////////////////////////////////
+
+#ifdef CRC_GENERIC
+
 #include "crc_macros.h"
 
 
@@ -26,8 +88,8 @@
 
 
 // See the comments in crc32_fast.c. They aren't duplicated here.
-extern LZMA_API(uint64_t)
-lzma_crc64(const uint8_t *buf, size_t size, uint64_t crc)
+static uint64_t
+crc64_generic(const uint8_t *buf, size_t size, uint64_t crc)
 {
 	crc = ~crc;
 
@@ -71,3 +133,380 @@ lzma_crc64(const uint8_t *buf, size_t size, uint64_t crc)
 
 	return ~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)
+
+
+// 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
+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)(-(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
+
+	__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);
+		}
+
+		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
+	}
+
+	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);
+
+#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
+}
+#endif
+
+
+////////////////////////
+// Detect CPU support //
+////////////////////////
+
+#if defined(CRC_GENERIC) && defined(CRC_CLMUL)
+static inline bool
+is_clmul_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.
+}
+
+
+#ifdef HAVE_FUNC_ATTRIBUTE_CONSTRUCTOR
+#	define CRC64_FUNC_INIT
+#	define CRC64_SET_FUNC_ATTR __attribute__((__constructor__))
+#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);
+#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;
+	return;
+}
+
+
+#ifndef HAVE_FUNC_ATTRIBUTE_CONSTRUCTOR
+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);
+}
+#endif
+#endif
+
+
+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.
+#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.
+	//
+	// FIXME: Lookup table isn't currently omitted on 32-bit x86,
+	// see crc64_table.c.
+	return crc64_clmul(buf, size, crc);
+
+#else
+	return crc64_generic(buf, size, crc);
+#endif
+}