1
0
Fork 0
mirror of https://git.tukaani.org/xz.git synced 2024-04-04 12:36:23 +02:00

liblzma: Create crc_clmul.c.

Both crc32_clmul() and crc64_clmul() are now exported from
crc32_clmul.c as lzma_crc32_clmul() and lzma_crc64_clmul(). This
ensures that is_clmul_supported() (now lzma_is_clmul_supported()) is
not duplicated between crc32_fast.c and crc64_fast.c.

Also, it encapsulates the complexity of the CLMUL implementations into a
single file and reduces the complexity of crc32_fast.c and crc64_fast.c.
Before, CLMUL code was present in crc32_fast.c, crc64_fast.c, and
crc_common.h.

During the conversion, various cleanups were applied to code (thanks to
Lasse Collin) including:

- Require using semicolons with MASK_/L/H/LH macros.
- Variable typing and const handling improvements.
- Improvements to comments.
- Fixes to the pragmas used.
- Removed unneeded variables.
- Whitespace improvements.
- Fixed CRC_USE_GENERIC_FOR_SMALL_INPUTS handling.
- Silenced warnings and removed the need for some #pragmas
This commit is contained in:
Jia Tan 2023-10-14 12:17:57 +08:00
parent a3ebc2c516
commit 8c0f9376f5
7 changed files with 444 additions and 423 deletions

View file

@ -845,7 +845,11 @@ if(HAVE_IMMINTRIN_H)
int main(void) { return 0; } int main(void) { return 0; }
" "
HAVE_USABLE_CLMUL) HAVE_USABLE_CLMUL)
tuklib_add_definition_if(liblzma HAVE_USABLE_CLMUL)
if(HAVE_USABLE_CLMUL)
target_sources(liblzma PRIVATE src/liblzma/check/crc_clmul.c)
target_compile_definitions(liblzma PRIVATE HAVE_USABLE_CLMUL)
endif()
endif() endif()
# Support -fvisiblity=hidden when building shared liblzma. # Support -fvisiblity=hidden when building shared liblzma.

View file

@ -1035,11 +1035,13 @@ __m128i my_clmul(__m128i a)
[Define to 1 if _mm_set_epi64x and [Define to 1 if _mm_set_epi64x and
_mm_clmulepi64_si128 are usable. _mm_clmulepi64_si128 are usable.
See configure.ac for details.]) See configure.ac for details.])
AC_MSG_RESULT([yes]) enable_clmul_crc=yes
], [ ], [
AC_MSG_RESULT([no]) enable_clmul_crc=no
]) ])
AC_MSG_RESULT([$enable_clmul_crc])
]) ])
AM_CONDITIONAL([COND_CRC_CLMUL], [test "x$enable_clmul_crc" = xyes])
# Check for sandbox support. If one is found, set enable_sandbox=found. # Check for sandbox support. If one is found, set enable_sandbox=found.
AS_CASE([$enable_sandbox], AS_CASE([$enable_sandbox],

View file

@ -26,6 +26,9 @@ if COND_ASM_X86
liblzma_la_SOURCES += check/crc32_x86.S liblzma_la_SOURCES += check/crc32_x86.S
else else
liblzma_la_SOURCES += check/crc32_fast.c liblzma_la_SOURCES += check/crc32_fast.c
if COND_CRC_CLMUL
liblzma_la_SOURCES += check/crc_clmul.c
endif
endif endif
endif endif
endif endif

View file

@ -34,11 +34,11 @@
#include "check.h" #include "check.h"
#include "crc_common.h" #include "crc_common.h"
#ifdef CRC_GENERIC
/////////////////// ///////////////////
// Generic CRC32 // // Generic CRC32 //
/////////////////// ///////////////////
#ifdef CRC_GENERIC
static uint32_t static uint32_t
crc32_generic(const uint8_t *buf, size_t size, uint32_t crc) crc32_generic(const uint8_t *buf, size_t size, uint32_t crc)
@ -99,118 +99,6 @@ crc32_generic(const uint8_t *buf, size_t size, uint32_t crc)
} }
#endif #endif
/////////////////////
// x86 CLMUL CRC32 //
/////////////////////
#ifdef CRC_CLMUL
#include <immintrin.h>
/*
// These functions were used to generate the constants
// at the top of crc32_clmul().
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;
}
*/
// 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
static uint32_t
crc32_clmul(const uint8_t *buf, size_t size, uint32_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
// uint32_t poly = 0xedb88320;
uint64_t p = 0x1db710640; // p << 1
uint64_t mu = 0x1f7011641; // calc_lo(p, p, 32) << 1 | 1
uint64_t k5 = 0x163cd6124; // calc_hi(p, p, 32) << 1
uint64_t k4 = 0x0ccaa009e; // calc_hi(p, p, 64) << 1
uint64_t k3 = 0x1751997d0; // calc_hi(p, p, 128) << 1
__m128i vfold4 = _mm_set_epi64x(mu, p);
__m128i vfold8 = _mm_set_epi64x(0, k5);
__m128i vfold16 = _mm_set_epi64x(k4, k3);
__m128i v0, v1, v2;
crc_simd_body(buf, size, &v0, &v1, vfold16, _mm_cvtsi32_si128(~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 ~_mm_extract_epi32(v0, 2);
#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
#if defined(CRC_GENERIC) && defined(CRC_CLMUL) #if defined(CRC_GENERIC) && defined(CRC_CLMUL)
typedef uint32_t (*crc32_func_type)( typedef uint32_t (*crc32_func_type)(
const uint8_t *buf, size_t size, uint32_t crc); const uint8_t *buf, size_t size, uint32_t crc);
@ -226,7 +114,7 @@ typedef uint32_t (*crc32_func_type)(
static crc32_func_type static crc32_func_type
crc32_resolve(void) crc32_resolve(void)
{ {
return is_clmul_supported() ? &crc32_clmul : &crc32_generic; return lzma_is_clmul_supported() ? &lzma_crc32_clmul : &crc32_generic;
} }
#if defined(HAVE_FUNC_ATTRIBUTE_IFUNC) && defined(__clang__) #if defined(HAVE_FUNC_ATTRIBUTE_IFUNC) && defined(__clang__)
@ -305,7 +193,7 @@ lzma_crc32(const uint8_t *buf, size_t size, uint32_t crc)
return crc32_func(buf, size, crc); return crc32_func(buf, size, crc);
#elif defined(CRC_CLMUL) #elif defined(CRC_CLMUL)
return crc32_clmul(buf, size, crc); return lzma_crc32_clmul(buf, size, crc);
#else #else
return crc32_generic(buf, size, crc); return crc32_generic(buf, size, crc);

View file

@ -31,13 +31,12 @@
#include "check.h" #include "check.h"
#include "crc_common.h" #include "crc_common.h"
#ifdef CRC_GENERIC
///////////////////////////////// /////////////////////////////////
// Generic slice-by-four CRC64 // // Generic slice-by-four CRC64 //
///////////////////////////////// /////////////////////////////////
#ifdef CRC_GENERIC
#ifdef WORDS_BIGENDIAN #ifdef WORDS_BIGENDIAN
# define A1(x) ((x) >> 56) # define A1(x) ((x) >> 56)
#else #else
@ -93,125 +92,6 @@ crc64_generic(const uint8_t *buf, size_t size, uint64_t crc)
} }
#endif #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;
}
*/
// 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
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 vfold8 = _mm_set_epi64x(p, mu);
const __m128i vfold16 = _mm_set_epi64x(k2, k1);
__m128i v0, v1, v2;
#if defined(__i386__) || defined(_M_IX86)
crc_simd_body(buf, size, &v0, &v1, vfold16, _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.
crc_simd_body(buf, size, &v0, &v1, vfold16, _mm_cvtsi64_si128(~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 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
#if defined(CRC_GENERIC) && defined(CRC_CLMUL) #if defined(CRC_GENERIC) && defined(CRC_CLMUL)
typedef uint64_t (*crc64_func_type)( typedef uint64_t (*crc64_func_type)(
const uint8_t *buf, size_t size, uint64_t crc); const uint8_t *buf, size_t size, uint64_t crc);
@ -227,7 +107,7 @@ typedef uint64_t (*crc64_func_type)(
static crc64_func_type static crc64_func_type
crc64_resolve(void) crc64_resolve(void)
{ {
return is_clmul_supported() ? &crc64_clmul : &crc64_generic; return lzma_is_clmul_supported() ? &lzma_crc64_clmul : &crc64_generic;
} }
#if defined(HAVE_FUNC_ATTRIBUTE_IFUNC) && defined(__clang__) #if defined(HAVE_FUNC_ATTRIBUTE_IFUNC) && defined(__clang__)
@ -322,7 +202,7 @@ lzma_crc64(const uint8_t *buf, size_t size, uint64_t crc)
// //
// FIXME: Lookup table isn't currently omitted on 32-bit x86, // FIXME: Lookup table isn't currently omitted on 32-bit x86,
// see crc64_table.c. // see crc64_table.c.
return crc64_clmul(buf, size, crc); return lzma_crc64_clmul(buf, size, crc);
#else #else
return crc64_generic(buf, size, crc); return crc64_generic(buf, size, crc);

View file

@ -0,0 +1,414 @@
///////////////////////////////////////////////////////////////////////////////
//
/// \file crc_clmul.c
/// \brief CRC32 and CRC64 implementations using CLMUL instructions.
///
/// lzma_crc32_clmul() and lzma_crc64_clmul() 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).
///
/// 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 has been put into the public domain.
// You can do whatever you want with this file.
//
///////////////////////////////////////////////////////////////////////////////
#include "common.h"
#include "crc_common.h"
#include <immintrin.h>
#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 lzma_crc64_clmul() 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
#if (defined(__GNUC__) || defined(__clang__)) && !defined(__EDG__)
__attribute__((__target__("ssse3,sse4.1,pclmul")))
#endif
#if lzma_has_attribute(__no_sanitize_address__)
__attribute__((__no_sanitize_address__))
#endif
static 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 lzma_crc32_clmul().
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 HAVE_CHECK_CRC32
// 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
extern uint32_t
lzma_crc32_clmul(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 // HAVE_CHECK_CRC32
/////////////////////
// x86 CLMUL CRC64 //
/////////////////////
/*
// These functions were used to generate the constants
// at the top of lzma_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;
}
*/
#ifdef HAVE_CHECK_CRC64
#if (defined(__GNUC__) || defined(__clang__)) && !defined(__EDG__)
__attribute__((__target__("ssse3,sse4.1,pclmul")))
#endif
extern uint64_t
lzma_crc64_clmul(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
}
#endif // HAVE_CHECK_CRC64
#if defined(_MSC_VER) && !defined(__INTEL_COMPILER) && !defined(__clang__) \
&& defined(_M_IX86)
# pragma optimize("", on)
#endif
////////////////////////
// Detect CPU support //
////////////////////////
extern bool
lzma_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.
}

View file

@ -6,6 +6,7 @@
// Authors: Lasse Collin // Authors: Lasse Collin
// Ilya Kurdyukov // Ilya Kurdyukov
// Hans Jansen // Hans Jansen
// Jia Tan
// //
// This file has been put into the public domain. // This file has been put into the public domain.
// You can do whatever you want with this file. // You can do whatever you want with this file.
@ -77,185 +78,14 @@
# endif # endif
#endif #endif
//////////////////////// /// Detect at runtime if the CPU supports the x86 CLMUL instruction when
// Detect CPU support // /// both the generic and CLMUL implementations are built.
//////////////////////// extern bool lzma_is_clmul_supported(void);
#if defined(CRC_GENERIC) && defined(CRC_CLMUL) /// CRC32 implemented with the x86 CLMUL instruction.
static inline bool extern uint32_t lzma_crc32_clmul(const uint8_t *buf, size_t size,
is_clmul_supported(void) uint32_t crc);
{
int success = 1;
uint32_t r[4]; // eax, ebx, ecx, edx
#if defined(_MSC_VER) /// CRC64 implemented with the x86 CLMUL instruction.
// This needs <intrin.h> with MSVC. ICC has it as a built-in extern uint64_t lzma_crc64_clmul(const uint8_t *buf, size_t size,
// on all platforms. uint64_t crc);
__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.
}
#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)
#ifdef CRC_CLMUL
#include <immintrin.h>
#if (defined(__GNUC__) || defined(__clang__)) && !defined(__EDG__)
__attribute__((__target__("ssse3,sse4.1,pclmul")))
#endif
#if lzma_has_attribute(__no_sanitize_address__)
__attribute__((__no_sanitize_address__))
#endif
static inline void
crc_simd_body(const uint8_t *buf, size_t size, __m128i *v0, __m128i *v1,
__m128i vfold16, __m128i crc2vec)
{
#if TUKLIB_GNUC_REQ(4, 6) || defined(__clang__)
# pragma GCC diagnostic push
# pragma GCC diagnostic ignored "-Wsign-conversion"
#endif
// 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
uintptr_t skip_start = (uintptr_t)buf & 15;
uintptr_t skip_end = -(uintptr_t)(buf + size) & 15;
// 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.
__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.
__m128i vsign = _mm_set1_epi8(-0x80);
// Masks to be used with _mm_blendv_epi8 and _mm_shuffle_epi8
// The first skip_start or skip_end bytes in the vectors will hav
// the high bit (0x80) set. _mm_blendv_epi8 and _mm_shuffle_epi
// will produce zeros for these positions. (Bitwise-xor of thes
// masks with vsign will produce the opposite behavior.)
__m128i mask_start = _mm_sub_epi8(vramp, _mm_set1_epi8(skip_start));
__m128i mask_end = _mm_sub_epi8(vramp, _mm_set1_epi8(skip_end));
// 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.
uintptr_t size2 = skip_start + size;
const __m128i *aligned_buf = (const __m128i*)((uintptr_t)buf & -16);
__m128i v2, v3, vcrc, data0;
vcrc = crc2vec;
// 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.
data0 = _mm_load_si128(aligned_buf);
data0 = _mm_blendv_epi8(data0, _mm_setzero_si128(), mask_start);
aligned_buf++;
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.
__m128i mask_low = _mm_add_epi8(
vramp, _mm_set1_epi8(size - 16));
MASK_LH(vcrc, mask_low, *v0, *v1)
MASK_L(data0, mask_end, v3)
*v0 = _mm_xor_si128(*v0, v3);
*v1 = _mm_alignr_epi8(*v1, *v0, 8);
} else {
__m128i data1 = _mm_load_si128(aligned_buf);
if (size <= 16) {
// 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.
__m128i mask_low = _mm_add_epi8(
vramp, _mm_set1_epi8(size - 16));
MASK_LH(vcrc, mask_low, *v0, *v1);
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 {
const __m128i *end = (const __m128i*)(
(char*)aligned_buf++ - 16 + size2);
MASK_LH(vcrc, 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);
}
}
}
#endif