root
/
xz
mirror of https://git.tukaani.org/xz.git
1
0
Fork 0
Code Issues Releases Wiki Activity

liblzma: x86 CLMUL CRC: Rewrite 

It's faster with both tiny and large buffers and doesn't require
disabling any sanitizers. With large buffers the extra speed is
from folding four 16-byte chunks in parallel.

The 32-bit x86 with MSVC reportedly still needs a workaround.
Now the simpler "__asm mov ebx, ebx" trick is enough but it
needs to be in lzma_crc64() instead of crc64_arch_optimized().
Thanks to Iouri Kharon for testing and the fix.

Thanks to Ilya Kurdyukov for testing the speed with aligned and
unaligned buffers on a few x86 processors and on E2K v6.

Thanks to Sam James for general feedback.

Fixes: https://github.com/tukaani-project/xz/issues/112
Fixes: https://github.com/tukaani-project/xz/issues/122
This commit is contained in:
Lasse Collin 2024-06-16 13:21:34 +03:00
parent c0e7eaae8d
commit 54eaea5ea4
2 changed files with 253 additions and 224 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
src/liblzma/check/crc64_fast.c
View File

@ -133,6 +133,14 @@ 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(_MSC_VER) && !defined(__INTEL_COMPILER) && !defined(__clang__) \
&& defined(_M_IX86) && defined(CRC64_ARCH_OPTIMIZED)
// VS2015-2022 might corrupt the ebx register on 32-bit x86 when
// the CLMUL code is enabled. This hack forces MSVC to store and
// restore ebx. This is only needed here, not in lzma_crc32().
__asm mov ebx, ebx
#endif
#if defined(CRC64_GENERIC) && defined(CRC64_ARCH_OPTIMIZED)
return crc64_func(buf, size, crc);

469
src/liblzma/check/crc_x86_clmul.h
View File

@ -8,10 +8,11 @@
/// 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
/// See the Intel white paper "Fast CRC Computation for Generic Polynomials
/// Using PCLMULQDQ Instruction" from 2009. The original file seems to be
/// gone from Intel's website but a version is available here:
/// 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).
/// (The link was checked on 2024-06-11.)
///
/// While this file has both CRC32 and CRC64 implementations, only one
/// can be built at a time. The version to build is selected by defining
@ -21,11 +22,11 @@
/// 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.
///
/// NOTE: The x86 CLMUL CRC implementation was rewritten for XZ Utils 5.8.0.
//
// Authors: Ilya Kurdyukov
// Hans Jansen
// Lasse Collin
// Jia Tan
// Authors: Lasse Collin
// Ilya Kurdyukov
//
///////////////////////////////////////////////////////////////////////////////
@ -61,257 +62,277 @@
#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)
// GCC and Clang would produce good code with _mm_set_epi64x
// but MSVC needs _mm_cvtsi64_si128 on x86-64.
#if defined(__i386__) || defined(_M_IX86)
# define my_set_low64(a) _mm_set_epi64x(0, (a))
#else
# define my_set_low64(a) _mm_cvtsi64_si128(a)
#endif
// Align it so that the whole array is within the same cache line.
// More than one unaligned load can be done from this during the
// same CRC function call.
//
// The bytes [0] to [31] are used with AND to clear the low bytes. (With ANDN
// those could be used to clear the high bytes too but it's not needed here.)
//
// The bytes [16] to [47] are for left shifts.
// The bytes [32] to [63] are for right shifts.
alignas(64)
static uint8_t vmasks[64] = {
0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07,
0x08, 0x09, 0x0A, 0x0B, 0x0C, 0x0D, 0x0E, 0x0F,
0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
};
// *Unaligned* 128-bit load
crc_attr_target
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)
static inline __m128i
my_load128(const uint8_t *p)
{
// 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;
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 {
// 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);
}
return _mm_loadu_si128((const __m128i *)p);
}
/////////////////////
// x86 CLMUL CRC32 //
/////////////////////
// Keep the highest "count" bytes as is and clear the remaining low bytes.
crc_attr_target
static inline __m128i
keep_high_bytes(__m128i v, size_t count)
{
return _mm_and_si128(my_load128((vmasks + count)), v);
}
// Shift the 128-bit value left by "amount" bytes (not bits).
crc_attr_target
static inline __m128i
shift_left(__m128i v, size_t amount)
{
return _mm_shuffle_epi8(v, my_load128((vmasks + 32 - amount)));
}
// Shift the 128-bit value right by "amount" bytes (not bits).
crc_attr_target
static inline __m128i
shift_right(__m128i v, size_t amount)
{
return _mm_shuffle_epi8(v, my_load128((vmasks + 32 + amount)));
}
crc_attr_target
static inline __m128i
fold(__m128i v, __m128i k)
{
__m128i a = _mm_clmulepi64_si128(v, k, 0x00);
__m128i b = _mm_clmulepi64_si128(v, k, 0x11);
return _mm_xor_si128(a, b);
}
crc_attr_target
static inline __m128i
fold_xor(__m128i v, __m128i k, const uint8_t *buf)
{
return _mm_xor_si128(my_load128(buf), fold(v, k));
}
#if BUILDING_CRC_CLMUL == 32
crc_attr_target
static uint32_t
crc32_arch_optimized(const uint8_t *buf, size_t size, uint32_t crc)
{
// The code assumes that there is at least one byte of input.
if (size == 0)
return crc;
// See crc_clmul_consts_gen.c.
const __m128i vfold16 = _mm_set_epi64x(0xccaa009e, 0xae689191);
const __m128i mu_p = _mm_set_epi64x(
(int64_t)0xb4e5b025f7011641, 0x1db710640);
__m128i v0, v1;
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
v0 = _mm_clmulepi64_si128(v1, mu_p, 0x10); // v1 * mu
v0 = _mm_clmulepi64_si128(v0, mu_p, 0x00); // v0 * p
v0 = _mm_xor_si128(v0, v1);
return ~(uint32_t)_mm_extract_epi32(v0, 2);
}
#endif // BUILDING_CRC_CLMUL == 32
/////////////////////
// x86 CLMUL CRC64 //
/////////////////////
#if BUILDING_CRC_CLMUL == 64
// 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
#else
crc_attr_target
static uint64_t
crc64_arch_optimized(const uint8_t *buf, size_t size, uint64_t crc)
#endif
{
// The code assumes that there is at least one byte of input.
// We will assume that there is at least one byte of input.
if (size == 0)
return crc;
// See crc_clmul_consts_gen.c.
const __m128i vfold16 = _mm_set_epi64x(
#if BUILDING_CRC_CLMUL == 32
const __m128i fold512 = _mm_set_epi64x(0x1d9513d7, 0x8f352d95);
const __m128i fold128 = _mm_set_epi64x(0xccaa009e, 0xae689191);
const __m128i mu_p = _mm_set_epi64x(
(int64_t)0xb4e5b025f7011641, 0x1db710640);
#else
const __m128i fold512 = _mm_set_epi64x(
(int64_t)0x081f6054a7842df4, (int64_t)0x6ae3efbb9dd441f3);
const __m128i fold128 = _mm_set_epi64x(
(int64_t)0xdabe95afc7875f40, (int64_t)0xe05dd497ca393ae4);
const __m128i mu_p = _mm_set_epi64x(
(int64_t)0x9c3e466c172963d5, (int64_t)0x92d8af2baf0e1e84);
__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, mu_p, 0x10);
v2 = _mm_clmulepi64_si128(v0, mu_p, 0x00);
v0 = _mm_xor_si128(_mm_xor_si128(v1, _mm_slli_si128(v0, 8)), v2);
__m128i v0, v1, v2, v3;
crc = ~crc;
if (size < 8) {
uint64_t x = crc;
size_t i = 0;
// Checking the bit instead of comparing the size means
// that we don't need to update the size between the steps.
if (size & 4) {
x ^= read32le(buf);
buf += 4;
i = 32;
}
if (size & 2) {
x ^= (uint64_t)read16le(buf) << i;
buf += 2;
i += 16;
}
if (size & 1)
x ^= (uint64_t)*buf << i;
v0 = my_set_low64((int64_t)x);
v0 = shift_left(v0, 8 - size);
} else if (size < 16) {
v0 = my_set_low64((int64_t)(crc ^ read64le(buf)));
// NOTE: buf is intentionally left 8 bytes behind so that
// we can read the last 1-7 bytes with read64le(buf + size).
size -= 8;
// Handling 8-byte input specially is a speed optimization
// as the clmul can be skipped. A branch is also needed to
// avoid a too high shift amount.
if (size > 0) {
const size_t padding = 8 - size;
uint64_t high = read64le(buf + size) >> (padding * 8);
#if defined(__i386__) || defined(_M_IX86)
// Simple but likely not the best code for 32-bit x86.
v0 = _mm_insert_epi32(v0, (int32_t)high, 2);
v0 = _mm_insert_epi32(v0, (int32_t)(high >> 32), 3);
#else
v0 = _mm_insert_epi64(v0, (int64_t)high, 1);
#endif
v0 = shift_left(v0, padding);
v1 = _mm_srli_si128(v0, 8);
v0 = _mm_clmulepi64_si128(v0, fold128, 0x10);
v0 = _mm_xor_si128(v0, v1);
}
} else {
v0 = my_set_low64((int64_t)crc);
// To align or not to align the buf pointer? If the end of
// the buffer isn't aligned, aligning the pointer here would
// make us do an extra folding step with the associated byte
// shuffling overhead. The cost of that would need to be
// lower than the benefit of aligned reads. Testing on an old
// Intel Ivy Bridge processor suggested that aligning isn't
// worth the cost but it likely depends on the processor and
// buffer size. Unaligned loads (MOVDQU) should be fast on
// x86 processors that support PCLMULQDQ, so we don't align
// the buf pointer here.
// Read the first (and possibly the only) full 16 bytes.
v0 = _mm_xor_si128(v0, my_load128(buf));
buf += 16;
size -= 16;
if (size >= 48) {
v1 = my_load128(buf);
v2 = my_load128(buf + 16);
v3 = my_load128(buf + 32);
buf += 48;
size -= 48;
while (size >= 64) {
v0 = fold_xor(v0, fold512, buf);
v1 = fold_xor(v1, fold512, buf + 16);
v2 = fold_xor(v2, fold512, buf + 32);
v3 = fold_xor(v3, fold512, buf + 48);
buf += 64;
size -= 64;
}
v0 = _mm_xor_si128(v1, fold(v0, fold128));
v0 = _mm_xor_si128(v2, fold(v0, fold128));
v0 = _mm_xor_si128(v3, fold(v0, fold128));
}
while (size >= 16) {
v0 = fold_xor(v0, fold128, buf);
buf += 16;
size -= 16;
}
if (size > 0) {
// We want the last "size" number of input bytes to
// be at the high bits of v1. First do a full 16-byte
// load and then mask the low bytes to zeros.
v1 = my_load128(buf + size - 16);
v1 = keep_high_bytes(v1, size);
// Shift high bytes from v0 to the low bytes of v1.
//
// Alternatively we could replace the combination
// keep_high_bytes + shift_right + _mm_or_si128 with
// _mm_shuffle_epi8 + _mm_blendv_epi8 but that would
// require larger tables for the masks. Now there are
// three loads (instead of two) from the mask tables
// but they all are from the same cache line.
v1 = _mm_or_si128(v1, shift_right(v0, size));
// Shift high bytes of v0 away, padding the
// low bytes with zeros.
v0 = shift_left(v0, 16 - size);
v0 = _mm_xor_si128(v1, fold(v0, fold128));
}
v1 = _mm_srli_si128(v0, 8);
v0 = _mm_clmulepi64_si128(v0, fold128, 0x10);
v0 = _mm_xor_si128(v0, v1);
}
// Barrett reduction
#if BUILDING_CRC_CLMUL == 32
v1 = _mm_clmulepi64_si128(v0, mu_p, 0x10); // v0 * mu
v1 = _mm_clmulepi64_si128(v1, mu_p, 0x00); // v1 * p
v0 = _mm_xor_si128(v0, v1);
return ~(uint32_t)_mm_extract_epi32(v0, 2);
#else
// Because p is 65 bits but one bit doesn't fit into the 64-bit
// half of __m128i, finish the second clmul by shifting v1 left
// by 64 bits and xorring it to the final result.
v1 = _mm_clmulepi64_si128(v0, mu_p, 0x10); // v0 * mu
v2 = _mm_slli_si128(v1, 8);
v1 = _mm_clmulepi64_si128(v1, mu_p, 0x00); // v1 * p
v0 = _mm_xor_si128(v0, v2);
v0 = _mm_xor_si128(v0, 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 defined(_MSC_VER) && !defined(__INTEL_COMPILER) && !defined(__clang__) \
&& defined(_M_IX86)
# pragma optimize("", on)
#endif
#endif // BUILDING_CRC_CLMUL == 64
}
// Inlining this function duplicates the function body in crc32_resolve() and