/src/php-src/ext/standard/crc32_x86.c

Source
/*
  +----------------------------------------------------------------------+
  | Copyright (c) The PHP Group                                          |
  +----------------------------------------------------------------------+
  | This source file is subject to version 3.01 of the PHP license,      |
  | that is bundled with this package in the file LICENSE, and is        |
  | available through the world-wide-web at the following url:           |
  | https://www.php.net/license/3_01.txt                                 |
  | If you did not receive a copy of the PHP license and are unable to   |
  | obtain it through the world-wide-web, please send a note to          |
  | license@php.net so we can mail you a copy immediately.               |
  +----------------------------------------------------------------------+
  | Author: Frank Du <frank.du@intel.com>                                |
  +----------------------------------------------------------------------+
  | Compute the crc32 of the buffer. Based on:                           |
  | "Fast CRC Computation for Generic Polynomials Using PCLMULQDQ"       |
  |  V. Gopal, E. Ozturk, et al., 2009, http://intel.ly/2ySEwL0          |
*/

#include "crc32_x86.h"

#if defined(ZEND_INTRIN_SSE4_2_PCLMUL_NATIVE) || defined(ZEND_INTRIN_SSE4_2_PCLMUL_RESOLVER)
# include <nmmintrin.h>
# include <wmmintrin.h>
#endif

#ifdef ZEND_INTRIN_SSE4_2_PCLMUL_RESOLVER
# include "Zend/zend_cpuinfo.h"
#endif

#if defined(ZEND_INTRIN_SSE4_2_PCLMUL_NATIVE) || defined(ZEND_INTRIN_SSE4_2_PCLMUL_RESOLVER)

typedef struct _crc32_pclmul_bit_consts {
  uint64_t k1k2[2];
  uint64_t k3k4[2];
  uint64_t k5k6[2];
  uint64_t uPx[2];
} crc32_pclmul_consts;

static const crc32_pclmul_consts crc32_pclmul_consts_maps[X86_CRC32_MAX] = {
  { /* X86_CRC32, polynomial: 0x04C11DB7 */
    {0x00e6228b11, 0x008833794c}, /* endianness swap */
    {0x00e8a45605, 0x00c5b9cd4c}, /* endianness swap */
    {0x00490d678d, 0x00f200aa66}, /* endianness swap */
    {0x0104d101df, 0x0104c11db7}
  },
  { /* X86_CRC32B, polynomial: 0x04C11DB7 with reversed ordering */
    {0x0154442bd4, 0x01c6e41596},
    {0x01751997d0, 0x00ccaa009e},
    {0x0163cd6124, 0x01db710640},
    {0x01f7011641, 0x01db710641},
  },
  { /* X86_CRC32C, polynomial: 0x1EDC6F41 with reversed ordering */
    {0x00740eef02, 0x009e4addf8},
    {0x00f20c0dfe, 0x014cd00bd6},
    {0x00dd45aab8, 0x0000000000},
    {0x00dea713f1, 0x0105ec76f0}
  }
};

static uint8_t pclmul_shuf_mask_table[16] = {
  0x0f, 0x0e, 0x0d, 0x0c, 0x0b, 0x0a, 0x09, 0x08,
  0x07, 0x06, 0x05, 0x04, 0x03, 0x02, 0x01, 0x00,
};

/* Folding of 128-bit data chunks */
#define CRC32_FOLDING_BLOCK_SIZE (16)

/* PCLMUL version of non-reflected crc32 */
ZEND_INTRIN_SSE4_2_PCLMUL_FUNC_DECL(size_t crc32_pclmul_batch(uint32_t *crc, const unsigned char *p, size_t nr, const crc32_pclmul_consts *consts));
size_t crc32_pclmul_batch(uint32_t *crc, const unsigned char *p, size_t nr, const crc32_pclmul_consts *consts)
{
  size_t nr_in = nr;
  __m128i x0, x1, x2, k, shuf_mask;

  if (nr < CRC32_FOLDING_BLOCK_SIZE) {
    return 0;
  }

  shuf_mask = _mm_loadu_si128((__m128i *)(pclmul_shuf_mask_table));
  x0 = _mm_cvtsi32_si128(*crc);
  x1 = _mm_loadu_si128((__m128i *)(p + 0x00));
  x0 = _mm_slli_si128(x0, 12);
  x1 = _mm_shuffle_epi8(x1, shuf_mask); /* endianness swap */
  x0 = _mm_xor_si128(x1, x0);
  p += CRC32_FOLDING_BLOCK_SIZE;
  nr -= CRC32_FOLDING_BLOCK_SIZE;

  if (nr >= (CRC32_FOLDING_BLOCK_SIZE * 3)) {
    __m128i x3, x4;

    x1 = _mm_loadu_si128((__m128i *)(p + 0x00));
    x1 = _mm_shuffle_epi8(x1, shuf_mask); /* endianness swap */
    x2 = _mm_loadu_si128((__m128i *)(p + 0x10));
    x2 = _mm_shuffle_epi8(x2, shuf_mask); /* endianness swap */
    x3 = _mm_loadu_si128((__m128i *)(p + 0x20));
    x3 = _mm_shuffle_epi8(x3, shuf_mask); /* endianness swap */
    p += CRC32_FOLDING_BLOCK_SIZE * 3;
    nr -= CRC32_FOLDING_BLOCK_SIZE * 3;

    k = _mm_loadu_si128((__m128i *)consts->k1k2);
    /* parallel folding by 4 */
    while (nr >= (CRC32_FOLDING_BLOCK_SIZE * 4)) {
      __m128i x5, x6, x7, x8, x9, x10, x11;
      x4 = _mm_clmulepi64_si128(x0, k, 0x00);
      x5 = _mm_clmulepi64_si128(x1, k, 0x00);
      x6 = _mm_clmulepi64_si128(x2, k, 0x00);
      x7 = _mm_clmulepi64_si128(x3, k, 0x00);
      x0 = _mm_clmulepi64_si128(x0, k, 0x11);
      x1 = _mm_clmulepi64_si128(x1, k, 0x11);
      x2 = _mm_clmulepi64_si128(x2, k, 0x11);
      x3 = _mm_clmulepi64_si128(x3, k, 0x11);
      x8 = _mm_loadu_si128((__m128i *)(p + 0x00));
      x8 = _mm_shuffle_epi8(x8, shuf_mask); /* endianness swap */
      x9 = _mm_loadu_si128((__m128i *)(p + 0x10));
      x9 = _mm_shuffle_epi8(x9, shuf_mask); /* endianness swap */
      x10 = _mm_loadu_si128((__m128i *)(p + 0x20));
      x10 = _mm_shuffle_epi8(x10, shuf_mask); /* endianness swap */
      x11 = _mm_loadu_si128((__m128i *)(p + 0x30));
      x11 = _mm_shuffle_epi8(x11, shuf_mask); /* endianness swap */
      x0 = _mm_xor_si128(x0, x4);
      x1 = _mm_xor_si128(x1, x5);
      x2 = _mm_xor_si128(x2, x6);
      x3 = _mm_xor_si128(x3, x7);
      x0 = _mm_xor_si128(x0, x8);
      x1 = _mm_xor_si128(x1, x9);
      x2 = _mm_xor_si128(x2, x10);
      x3 = _mm_xor_si128(x3, x11);

      p += CRC32_FOLDING_BLOCK_SIZE * 4;
      nr -= CRC32_FOLDING_BLOCK_SIZE * 4;
    }

    k = _mm_loadu_si128((__m128i *)consts->k3k4);
    /* fold 4 to 1, [x1, x2, x3] -> x0 */
    x4 = _mm_clmulepi64_si128(x0, k, 0x00);
    x0 = _mm_clmulepi64_si128(x0, k, 0x11);
    x0 = _mm_xor_si128(x0, x1);
    x0 = _mm_xor_si128(x0, x4);
    x4 = _mm_clmulepi64_si128(x0, k, 0x00);
    x0 = _mm_clmulepi64_si128(x0, k, 0x11);
    x0 = _mm_xor_si128(x0, x2);
    x0 = _mm_xor_si128(x0, x4);
    x4 = _mm_clmulepi64_si128(x0, k, 0x00);
    x0 = _mm_clmulepi64_si128(x0, k, 0x11);
    x0 = _mm_xor_si128(x0, x3);
    x0 = _mm_xor_si128(x0, x4);
  }

  k = _mm_loadu_si128((__m128i *)consts->k3k4);
  /* folding by 1 */
  while (nr >= CRC32_FOLDING_BLOCK_SIZE) {
    /* load next to x2, fold to x0, x1 */
    x2 = _mm_loadu_si128((__m128i *)(p + 0x00));
    x2 = _mm_shuffle_epi8(x2, shuf_mask); /* endianness swap */
    x1 = _mm_clmulepi64_si128(x0, k, 0x00);
    x0 = _mm_clmulepi64_si128(x0, k, 0x11);
    x0 = _mm_xor_si128(x0, x2);
    x0 = _mm_xor_si128(x0, x1);
    p += CRC32_FOLDING_BLOCK_SIZE;
    nr -= CRC32_FOLDING_BLOCK_SIZE;
  }

  /* reduce 128-bits(final fold) to 96-bits */
  k = _mm_loadu_si128((__m128i*)consts->k5k6);
  x1 = _mm_clmulepi64_si128(x0, k, 0x11);
  x0 = _mm_slli_si128(x0, 8);
  x0 = _mm_srli_si128(x0, 4);
  x0 = _mm_xor_si128(x0, x1);
  /* reduce 96-bits to 64-bits */
  x1 = _mm_clmulepi64_si128(x0, k, 0x01);
  x0 = _mm_xor_si128(x0, x1);

  /* barrett reduction */
  k = _mm_loadu_si128((__m128i*)consts->uPx);
  x1 = _mm_move_epi64(x0);
  x1 = _mm_srli_si128(x1, 4);
  x1 = _mm_clmulepi64_si128(x1, k, 0x00);
  x1 = _mm_srli_si128(x1, 4);
  x1 = _mm_clmulepi64_si128(x1, k, 0x10);
  x0 = _mm_xor_si128(x1, x0);
  *crc =  _mm_extract_epi32(x0, 0);
  return (nr_in - nr); /* the nr processed */
}

/* PCLMUL version of reflected crc32 */
ZEND_INTRIN_SSE4_2_PCLMUL_FUNC_DECL(size_t crc32_pclmul_reflected_batch(uint32_t *crc, const unsigned char *p, size_t nr, const crc32_pclmul_consts *consts));
size_t crc32_pclmul_reflected_batch(uint32_t *crc, const unsigned char *p, size_t nr, const crc32_pclmul_consts *consts)
{
  size_t nr_in = nr;
  __m128i x0, x1, x2, k;

  if (nr < CRC32_FOLDING_BLOCK_SIZE) {
    return 0;
  }

  x0 = _mm_loadu_si128((__m128i *)(p + 0x00));
  x0 = _mm_xor_si128(x0, _mm_cvtsi32_si128(*crc));
  p += CRC32_FOLDING_BLOCK_SIZE;
  nr -= CRC32_FOLDING_BLOCK_SIZE;
  if (nr >= (CRC32_FOLDING_BLOCK_SIZE * 3)) {
    __m128i x3, x4;

    x1 = _mm_loadu_si128((__m128i *)(p + 0x00));
    x2 = _mm_loadu_si128((__m128i *)(p + 0x10));
    x3 = _mm_loadu_si128((__m128i *)(p + 0x20));
    p += CRC32_FOLDING_BLOCK_SIZE * 3;
    nr -= CRC32_FOLDING_BLOCK_SIZE * 3;

    k = _mm_loadu_si128((__m128i *)consts->k1k2);
    /* parallel folding by 4 */
    while (nr >= (CRC32_FOLDING_BLOCK_SIZE * 4)) {
      __m128i x5, x6, x7, x8, x9, x10, x11;
      x4 = _mm_clmulepi64_si128(x0, k, 0x00);
      x5 = _mm_clmulepi64_si128(x1, k, 0x00);
      x6 = _mm_clmulepi64_si128(x2, k, 0x00);
      x7 = _mm_clmulepi64_si128(x3, k, 0x00);
      x0 = _mm_clmulepi64_si128(x0, k, 0x11);
      x1 = _mm_clmulepi64_si128(x1, k, 0x11);
      x2 = _mm_clmulepi64_si128(x2, k, 0x11);
      x3 = _mm_clmulepi64_si128(x3, k, 0x11);
      x8 = _mm_loadu_si128((__m128i *)(p + 0x00));
      x9 = _mm_loadu_si128((__m128i *)(p + 0x10));
      x10 = _mm_loadu_si128((__m128i *)(p + 0x20));
      x11 = _mm_loadu_si128((__m128i *)(p + 0x30));
      x0 = _mm_xor_si128(x0, x4);
      x1 = _mm_xor_si128(x1, x5);
      x2 = _mm_xor_si128(x2, x6);
      x3 = _mm_xor_si128(x3, x7);
      x0 = _mm_xor_si128(x0, x8);
      x1 = _mm_xor_si128(x1, x9);
      x2 = _mm_xor_si128(x2, x10);
      x3 = _mm_xor_si128(x3, x11);

      p += CRC32_FOLDING_BLOCK_SIZE * 4;
      nr -= CRC32_FOLDING_BLOCK_SIZE * 4;
    }

    k = _mm_loadu_si128((__m128i *)consts->k3k4);
    /* fold 4 to 1, [x1, x2, x3] -> x0 */
    x4 = _mm_clmulepi64_si128(x0, k, 0x00);
    x0 = _mm_clmulepi64_si128(x0, k, 0x11);
    x0 = _mm_xor_si128(x0, x1);
    x0 = _mm_xor_si128(x0, x4);
    x4 = _mm_clmulepi64_si128(x0, k, 0x00);
    x0 = _mm_clmulepi64_si128(x0, k, 0x11);
    x0 = _mm_xor_si128(x0, x2);
    x0 = _mm_xor_si128(x0, x4);
    x4 = _mm_clmulepi64_si128(x0, k, 0x00);
    x0 = _mm_clmulepi64_si128(x0, k, 0x11);
    x0 = _mm_xor_si128(x0, x3);
    x0 = _mm_xor_si128(x0, x4);
  }

  k = _mm_loadu_si128((__m128i *)consts->k3k4);
  /* folding by 1 */
  while (nr >= CRC32_FOLDING_BLOCK_SIZE) {
    /* load next to x2, fold to x0, x1 */
    x2 = _mm_loadu_si128((__m128i *)(p + 0x00));
    x1 = _mm_clmulepi64_si128(x0, k, 0x00);
    x0 = _mm_clmulepi64_si128(x0, k, 0x11);
    x0 = _mm_xor_si128(x0, x2);
    x0 = _mm_xor_si128(x0, x1);
    p += CRC32_FOLDING_BLOCK_SIZE;
    nr -= CRC32_FOLDING_BLOCK_SIZE;
  }

  /* reduce 128-bits(final fold) to 96-bits */
  x1 = _mm_clmulepi64_si128(x0, k, 0x10);
  x0 = _mm_srli_si128(x0, 8);
  x0 = _mm_xor_si128(x0, x1);
  /* reduce 96-bits to 64-bits */
  x1 = _mm_shuffle_epi32(x0, 0xfc);
  x0 = _mm_shuffle_epi32(x0, 0xf9);
  k = _mm_loadu_si128((__m128i*)consts->k5k6);
  x1 = _mm_clmulepi64_si128(x1, k, 0x00);
  x0 = _mm_xor_si128(x0, x1);

  /* barrett reduction */
  x1 = _mm_shuffle_epi32(x0, 0xf3);
  x0 = _mm_slli_si128(x0, 4);
  k = _mm_loadu_si128((__m128i*)consts->uPx);
  x1 = _mm_clmulepi64_si128(x1, k, 0x00);
  x1 = _mm_clmulepi64_si128(x1, k, 0x10);
  x0 = _mm_xor_si128(x1, x0);
  *crc =  _mm_extract_epi32(x0, 2);
  return (nr_in - nr); /* the nr processed */
}

# if defined(ZEND_INTRIN_SSE4_2_PCLMUL_NATIVE)
size_t crc32_x86_simd_update(X86_CRC32_TYPE type, uint32_t *crc, const unsigned char *p, size_t nr)
# else /* ZEND_INTRIN_SSE4_2_PCLMUL_RESOLVER */
size_t crc32_sse42_pclmul_update(X86_CRC32_TYPE type, uint32_t *crc, const unsigned char *p, size_t nr)
# endif
{
  if (type > X86_CRC32_MAX) {
    return 0;
  }
  const crc32_pclmul_consts *consts = &crc32_pclmul_consts_maps[type];

  switch (type) {
  case X86_CRC32:
    return crc32_pclmul_batch(crc, p, nr, consts);
  case X86_CRC32B:
  case X86_CRC32C:
    return crc32_pclmul_reflected_batch(crc, p, nr, consts);
  default:
    return 0;
  }
}
#endif

#ifdef ZEND_INTRIN_SSE4_2_PCLMUL_RESOLVER
static size_t crc32_x86_simd_update_default(X86_CRC32_TYPE type, uint32_t *crc, const unsigned char *p, size_t nr)
{
  return 0;
}

# ifdef ZEND_INTRIN_SSE4_2_PCLMUL_FUNC_PROTO
size_t crc32_x86_simd_update(X86_CRC32_TYPE type, uint32_t *crc, const unsigned char *p, size_t nr) __attribute__((ifunc("resolve_crc32_x86_simd_update")));

typedef size_t (*crc32_x86_simd_func_t)(X86_CRC32_TYPE type, uint32_t *crc, const unsigned char *p, size_t nr);

ZEND_NO_SANITIZE_ADDRESS
ZEND_ATTRIBUTE_UNUSED /* clang mistakenly warns about this */
static crc32_x86_simd_func_t resolve_crc32_x86_simd_update(void) {
  if (zend_cpu_supports_sse42() && zend_cpu_supports_pclmul()) {
    return crc32_sse42_pclmul_update;
  }
  return crc32_x86_simd_update_default;
}
# else /* ZEND_INTRIN_SSE4_2_PCLMUL_FUNC_PTR */
static size_t (*crc32_x86_simd_ptr)(X86_CRC32_TYPE type, uint32_t *crc, const unsigned char *p, size_t nr) = crc32_x86_simd_update_default;

size_t crc32_x86_simd_update(X86_CRC32_TYPE type, uint32_t *crc, const unsigned char *p, size_t nr) {
  return crc32_x86_simd_ptr(type, crc, p, nr);
}

/* {{{ PHP_MINIT_FUNCTION */
PHP_MINIT_FUNCTION(crc32_x86_intrin)
{
  if (zend_cpu_supports_sse42() && zend_cpu_supports_pclmul()) {
    crc32_x86_simd_ptr = crc32_sse42_pclmul_update;
  }
  return SUCCESS;
}
/* }}} */
# endif
#endif

Coverage Report

Created: 2025-12-31 07:28

Line	Count	Source
1		/*
2		+----------------------------------------------------------------------+
3		\| Copyright (c) The PHP Group \|
4		+----------------------------------------------------------------------+
5		\| This source file is subject to version 3.01 of the PHP license, \|
6		\| that is bundled with this package in the file LICENSE, and is \|
7		\| available through the world-wide-web at the following url: \|
8		\| https://www.php.net/license/3_01.txt \|
9		\| If you did not receive a copy of the PHP license and are unable to \|
10		\| obtain it through the world-wide-web, please send a note to \|
11		\| license@php.net so we can mail you a copy immediately. \|
12		+----------------------------------------------------------------------+
13		\| Author: Frank Du <frank.du@intel.com> \|
14		+----------------------------------------------------------------------+
15		\| Compute the crc32 of the buffer. Based on: \|
16		\| "Fast CRC Computation for Generic Polynomials Using PCLMULQDQ" \|
17		\| V. Gopal, E. Ozturk, et al., 2009, http://intel.ly/2ySEwL0 \|
18		*/
19
20		#include "crc32_x86.h"
21
22		#if defined(ZEND_INTRIN_SSE4_2_PCLMUL_NATIVE) \|\| defined(ZEND_INTRIN_SSE4_2_PCLMUL_RESOLVER)
23		# include <nmmintrin.h>
24		# include <wmmintrin.h>
25		#endif
26
27		#ifdef ZEND_INTRIN_SSE4_2_PCLMUL_RESOLVER
28		# include "Zend/zend_cpuinfo.h"
29		#endif
30
31		#if defined(ZEND_INTRIN_SSE4_2_PCLMUL_NATIVE) \|\| defined(ZEND_INTRIN_SSE4_2_PCLMUL_RESOLVER)
32
33		typedef struct _crc32_pclmul_bit_consts {
34		uint64_t k1k2[2];
35		uint64_t k3k4[2];
36		uint64_t k5k6[2];
37		uint64_t uPx[2];
38		} crc32_pclmul_consts;
39
40		static const crc32_pclmul_consts crc32_pclmul_consts_maps[X86_CRC32_MAX] = {
41		{ /* X86_CRC32, polynomial: 0x04C11DB7 */
42		{0x00e6228b11, 0x008833794c}, /* endianness swap */
43		{0x00e8a45605, 0x00c5b9cd4c}, /* endianness swap */
44		{0x00490d678d, 0x00f200aa66}, /* endianness swap */
45		{0x0104d101df, 0x0104c11db7}
46		},
47		{ /* X86_CRC32B, polynomial: 0x04C11DB7 with reversed ordering */
48		{0x0154442bd4, 0x01c6e41596},
49		{0x01751997d0, 0x00ccaa009e},
50		{0x0163cd6124, 0x01db710640},
51		{0x01f7011641, 0x01db710641},
52		},
53		{ /* X86_CRC32C, polynomial: 0x1EDC6F41 with reversed ordering */
54		{0x00740eef02, 0x009e4addf8},
55		{0x00f20c0dfe, 0x014cd00bd6},
56		{0x00dd45aab8, 0x0000000000},
57		{0x00dea713f1, 0x0105ec76f0}
58		}
59		};
60
61		static uint8_t pclmul_shuf_mask_table[16] = {
62		0x0f, 0x0e, 0x0d, 0x0c, 0x0b, 0x0a, 0x09, 0x08,
63		0x07, 0x06, 0x05, 0x04, 0x03, 0x02, 0x01, 0x00,
64		};
65
66		/* Folding of 128-bit data chunks */
67	24.8k	#define CRC32_FOLDING_BLOCK_SIZE (16)
68
69		/* PCLMUL version of non-reflected crc32 */
70		ZEND_INTRIN_SSE4_2_PCLMUL_FUNC_DECL(size_t crc32_pclmul_batch(uint32_t crc, const unsigned char p, size_t nr, const crc32_pclmul_consts *consts));
71		size_t crc32_pclmul_batch(uint32_t crc, const unsigned char p, size_t nr, const crc32_pclmul_consts *consts)
72	35	{
73	35	size_t nr_in = nr;
74	35	__m128i x0, x1, x2, k, shuf_mask;
75
76	35	if (nr < CRC32_FOLDING_BLOCK_SIZE) {
77	6	return 0;
78	6	}
79
80	29	shuf_mask = _mm_loadu_si128((__m128i *)(pclmul_shuf_mask_table));
81	29	x0 = _mm_cvtsi32_si128(*crc);
82	29	x1 = _mm_loadu_si128((__m128i *)(p + 0x00));
83	29	x0 = _mm_slli_si128(x0, 12);
84	29	x1 = _mm_shuffle_epi8(x1, shuf_mask); /* endianness swap */
85	29	x0 = _mm_xor_si128(x1, x0);
86	29	p += CRC32_FOLDING_BLOCK_SIZE;
87	29	nr -= CRC32_FOLDING_BLOCK_SIZE;
88
89	29	if (nr >= (CRC32_FOLDING_BLOCK_SIZE * 3)) {
90	25	__m128i x3, x4;
91
92	25	x1 = _mm_loadu_si128((__m128i *)(p + 0x00));
93	25	x1 = _mm_shuffle_epi8(x1, shuf_mask); /* endianness swap */
94	25	x2 = _mm_loadu_si128((__m128i *)(p + 0x10));
95	25	x2 = _mm_shuffle_epi8(x2, shuf_mask); /* endianness swap */
96	25	x3 = _mm_loadu_si128((__m128i *)(p + 0x20));
97	25	x3 = _mm_shuffle_epi8(x3, shuf_mask); /* endianness swap */
98	25	p += CRC32_FOLDING_BLOCK_SIZE * 3;
99	25	nr -= CRC32_FOLDING_BLOCK_SIZE * 3;
100
101	25	k = _mm_loadu_si128((__m128i *)consts->k1k2);
102		/* parallel folding by 4 */
103	3.19k	while (nr >= (CRC32_FOLDING_BLOCK_SIZE * 4)) {
104	3.16k	__m128i x5, x6, x7, x8, x9, x10, x11;
105	3.16k	x4 = _mm_clmulepi64_si128(x0, k, 0x00);
106	3.16k	x5 = _mm_clmulepi64_si128(x1, k, 0x00);
107	3.16k	x6 = _mm_clmulepi64_si128(x2, k, 0x00);
108	3.16k	x7 = _mm_clmulepi64_si128(x3, k, 0x00);
109	3.16k	x0 = _mm_clmulepi64_si128(x0, k, 0x11);
110	3.16k	x1 = _mm_clmulepi64_si128(x1, k, 0x11);
111	3.16k	x2 = _mm_clmulepi64_si128(x2, k, 0x11);
112	3.16k	x3 = _mm_clmulepi64_si128(x3, k, 0x11);
113	3.16k	x8 = _mm_loadu_si128((__m128i *)(p + 0x00));
114	3.16k	x8 = _mm_shuffle_epi8(x8, shuf_mask); /* endianness swap */
115	3.16k	x9 = _mm_loadu_si128((__m128i *)(p + 0x10));
116	3.16k	x9 = _mm_shuffle_epi8(x9, shuf_mask); /* endianness swap */
117	3.16k	x10 = _mm_loadu_si128((__m128i *)(p + 0x20));
118	3.16k	x10 = _mm_shuffle_epi8(x10, shuf_mask); /* endianness swap */
119	3.16k	x11 = _mm_loadu_si128((__m128i *)(p + 0x30));
120	3.16k	x11 = _mm_shuffle_epi8(x11, shuf_mask); /* endianness swap */
121	3.16k	x0 = _mm_xor_si128(x0, x4);
122	3.16k	x1 = _mm_xor_si128(x1, x5);
123	3.16k	x2 = _mm_xor_si128(x2, x6);
124	3.16k	x3 = _mm_xor_si128(x3, x7);
125	3.16k	x0 = _mm_xor_si128(x0, x8);
126	3.16k	x1 = _mm_xor_si128(x1, x9);
127	3.16k	x2 = _mm_xor_si128(x2, x10);
128	3.16k	x3 = _mm_xor_si128(x3, x11);
129
130	3.16k	p += CRC32_FOLDING_BLOCK_SIZE * 4;
131	3.16k	nr -= CRC32_FOLDING_BLOCK_SIZE * 4;
132	3.16k	}
133
134	25	k = _mm_loadu_si128((__m128i *)consts->k3k4);
135		/* fold 4 to 1, [x1, x2, x3] -> x0 */
136	25	x4 = _mm_clmulepi64_si128(x0, k, 0x00);
137	25	x0 = _mm_clmulepi64_si128(x0, k, 0x11);
138	25	x0 = _mm_xor_si128(x0, x1);
139	25	x0 = _mm_xor_si128(x0, x4);
140	25	x4 = _mm_clmulepi64_si128(x0, k, 0x00);
141	25	x0 = _mm_clmulepi64_si128(x0, k, 0x11);
142	25	x0 = _mm_xor_si128(x0, x2);
143	25	x0 = _mm_xor_si128(x0, x4);
144	25	x4 = _mm_clmulepi64_si128(x0, k, 0x00);
145	25	x0 = _mm_clmulepi64_si128(x0, k, 0x11);
146	25	x0 = _mm_xor_si128(x0, x3);
147	25	x0 = _mm_xor_si128(x0, x4);
148	25	}
149
150	29	k = _mm_loadu_si128((__m128i *)consts->k3k4);
151		/* folding by 1 */
152	61	while (nr >= CRC32_FOLDING_BLOCK_SIZE) {
153		/* load next to x2, fold to x0, x1 */
154	32	x2 = _mm_loadu_si128((__m128i *)(p + 0x00));
155	32	x2 = _mm_shuffle_epi8(x2, shuf_mask); /* endianness swap */
156	32	x1 = _mm_clmulepi64_si128(x0, k, 0x00);
157	32	x0 = _mm_clmulepi64_si128(x0, k, 0x11);
158	32	x0 = _mm_xor_si128(x0, x2);
159	32	x0 = _mm_xor_si128(x0, x1);
160	32	p += CRC32_FOLDING_BLOCK_SIZE;
161	32	nr -= CRC32_FOLDING_BLOCK_SIZE;
162	32	}
163
164		/* reduce 128-bits(final fold) to 96-bits */
165	29	k = _mm_loadu_si128((__m128i*)consts->k5k6);
166	29	x1 = _mm_clmulepi64_si128(x0, k, 0x11);
167	29	x0 = _mm_slli_si128(x0, 8);
168	29	x0 = _mm_srli_si128(x0, 4);
169	29	x0 = _mm_xor_si128(x0, x1);
170		/* reduce 96-bits to 64-bits */
171	29	x1 = _mm_clmulepi64_si128(x0, k, 0x01);
172	29	x0 = _mm_xor_si128(x0, x1);
173
174		/* barrett reduction */
175	29	k = _mm_loadu_si128((__m128i*)consts->uPx);
176	29	x1 = _mm_move_epi64(x0);
177	29	x1 = _mm_srli_si128(x1, 4);
178	29	x1 = _mm_clmulepi64_si128(x1, k, 0x00);
179	29	x1 = _mm_srli_si128(x1, 4);
180	29	x1 = _mm_clmulepi64_si128(x1, k, 0x10);
181	29	x0 = _mm_xor_si128(x1, x0);
182	29	*crc = _mm_extract_epi32(x0, 0);
183	29	return (nr_in - nr); /* the nr processed */
184	35	}
185
186		/* PCLMUL version of reflected crc32 */
187		ZEND_INTRIN_SSE4_2_PCLMUL_FUNC_DECL(size_t crc32_pclmul_reflected_batch(uint32_t crc, const unsigned char p, size_t nr, const crc32_pclmul_consts *consts));
188		size_t crc32_pclmul_reflected_batch(uint32_t crc, const unsigned char p, size_t nr, const crc32_pclmul_consts *consts)
189	58	{
190	58	size_t nr_in = nr;
191	58	__m128i x0, x1, x2, k;
192
193	58	if (nr < CRC32_FOLDING_BLOCK_SIZE) {
194	8	return 0;
195	8	}
196
197	50	x0 = _mm_loadu_si128((__m128i *)(p + 0x00));
198	50	x0 = _mm_xor_si128(x0, _mm_cvtsi32_si128(*crc));
199	50	p += CRC32_FOLDING_BLOCK_SIZE;
200	50	nr -= CRC32_FOLDING_BLOCK_SIZE;
201	50	if (nr >= (CRC32_FOLDING_BLOCK_SIZE * 3)) {
202	22	__m128i x3, x4;
203
204	22	x1 = _mm_loadu_si128((__m128i *)(p + 0x00));
205	22	x2 = _mm_loadu_si128((__m128i *)(p + 0x10));
206	22	x3 = _mm_loadu_si128((__m128i *)(p + 0x20));
207	22	p += CRC32_FOLDING_BLOCK_SIZE * 3;
208	22	nr -= CRC32_FOLDING_BLOCK_SIZE * 3;
209
210	22	k = _mm_loadu_si128((__m128i *)consts->k1k2);
211		/* parallel folding by 4 */
212	4.88k	while (nr >= (CRC32_FOLDING_BLOCK_SIZE * 4)) {
213	4.85k	__m128i x5, x6, x7, x8, x9, x10, x11;
214	4.85k	x4 = _mm_clmulepi64_si128(x0, k, 0x00);
215	4.85k	x5 = _mm_clmulepi64_si128(x1, k, 0x00);
216	4.85k	x6 = _mm_clmulepi64_si128(x2, k, 0x00);
217	4.85k	x7 = _mm_clmulepi64_si128(x3, k, 0x00);
218	4.85k	x0 = _mm_clmulepi64_si128(x0, k, 0x11);
219	4.85k	x1 = _mm_clmulepi64_si128(x1, k, 0x11);
220	4.85k	x2 = _mm_clmulepi64_si128(x2, k, 0x11);
221	4.85k	x3 = _mm_clmulepi64_si128(x3, k, 0x11);
222	4.85k	x8 = _mm_loadu_si128((__m128i *)(p + 0x00));
223	4.85k	x9 = _mm_loadu_si128((__m128i *)(p + 0x10));
224	4.85k	x10 = _mm_loadu_si128((__m128i *)(p + 0x20));
225	4.85k	x11 = _mm_loadu_si128((__m128i *)(p + 0x30));
226	4.85k	x0 = _mm_xor_si128(x0, x4);
227	4.85k	x1 = _mm_xor_si128(x1, x5);
228	4.85k	x2 = _mm_xor_si128(x2, x6);
229	4.85k	x3 = _mm_xor_si128(x3, x7);
230	4.85k	x0 = _mm_xor_si128(x0, x8);
231	4.85k	x1 = _mm_xor_si128(x1, x9);
232	4.85k	x2 = _mm_xor_si128(x2, x10);
233	4.85k	x3 = _mm_xor_si128(x3, x11);
234
235	4.85k	p += CRC32_FOLDING_BLOCK_SIZE * 4;
236	4.85k	nr -= CRC32_FOLDING_BLOCK_SIZE * 4;
237	4.85k	}
238
239	22	k = _mm_loadu_si128((__m128i *)consts->k3k4);
240		/* fold 4 to 1, [x1, x2, x3] -> x0 */
241	22	x4 = _mm_clmulepi64_si128(x0, k, 0x00);
242	22	x0 = _mm_clmulepi64_si128(x0, k, 0x11);
243	22	x0 = _mm_xor_si128(x0, x1);
244	22	x0 = _mm_xor_si128(x0, x4);
245	22	x4 = _mm_clmulepi64_si128(x0, k, 0x00);
246	22	x0 = _mm_clmulepi64_si128(x0, k, 0x11);
247	22	x0 = _mm_xor_si128(x0, x2);
248	22	x0 = _mm_xor_si128(x0, x4);
249	22	x4 = _mm_clmulepi64_si128(x0, k, 0x00);
250	22	x0 = _mm_clmulepi64_si128(x0, k, 0x11);
251	22	x0 = _mm_xor_si128(x0, x3);
252	22	x0 = _mm_xor_si128(x0, x4);
253	22	}
254
255	50	k = _mm_loadu_si128((__m128i *)consts->k3k4);
256		/* folding by 1 */
257	101	while (nr >= CRC32_FOLDING_BLOCK_SIZE) {
258		/* load next to x2, fold to x0, x1 */
259	51	x2 = _mm_loadu_si128((__m128i *)(p + 0x00));
260	51	x1 = _mm_clmulepi64_si128(x0, k, 0x00);
261	51	x0 = _mm_clmulepi64_si128(x0, k, 0x11);
262	51	x0 = _mm_xor_si128(x0, x2);
263	51	x0 = _mm_xor_si128(x0, x1);
264	51	p += CRC32_FOLDING_BLOCK_SIZE;
265	51	nr -= CRC32_FOLDING_BLOCK_SIZE;
266	51	}
267
268		/* reduce 128-bits(final fold) to 96-bits */
269	50	x1 = _mm_clmulepi64_si128(x0, k, 0x10);
270	50	x0 = _mm_srli_si128(x0, 8);
271	50	x0 = _mm_xor_si128(x0, x1);
272		/* reduce 96-bits to 64-bits */
273	50	x1 = _mm_shuffle_epi32(x0, 0xfc);
274	50	x0 = _mm_shuffle_epi32(x0, 0xf9);
275	50	k = _mm_loadu_si128((__m128i*)consts->k5k6);
276	50	x1 = _mm_clmulepi64_si128(x1, k, 0x00);
277	50	x0 = _mm_xor_si128(x0, x1);
278
279		/* barrett reduction */
280	50	x1 = _mm_shuffle_epi32(x0, 0xf3);
281	50	x0 = _mm_slli_si128(x0, 4);
282	50	k = _mm_loadu_si128((__m128i*)consts->uPx);
283	50	x1 = _mm_clmulepi64_si128(x1, k, 0x00);
284	50	x1 = _mm_clmulepi64_si128(x1, k, 0x10);
285	50	x0 = _mm_xor_si128(x1, x0);
286	50	*crc = _mm_extract_epi32(x0, 2);
287	50	return (nr_in - nr); /* the nr processed */
288	58	}
289
290		# if defined(ZEND_INTRIN_SSE4_2_PCLMUL_NATIVE)
291		size_t crc32_x86_simd_update(X86_CRC32_TYPE type, uint32_t crc, const unsigned char p, size_t nr)
292		# else /* ZEND_INTRIN_SSE4_2_PCLMUL_RESOLVER */
293		size_t crc32_sse42_pclmul_update(X86_CRC32_TYPE type, uint32_t crc, const unsigned char p, size_t nr)
294		# endif
295	93	{
296	93	if (type > X86_CRC32_MAX) {
297	0	return 0;
298	0	}
299	93	const crc32_pclmul_consts *consts = &crc32_pclmul_consts_maps[type];
300
301	93	switch (type) {
302	35	case X86_CRC32:
303	35	return crc32_pclmul_batch(crc, p, nr, consts);
304	35	case X86_CRC32B:
305	58	case X86_CRC32C:
306	58	return crc32_pclmul_reflected_batch(crc, p, nr, consts);
307	0	default:
308	0	return 0;
309	93	}
310	93	}
311		#endif
312
313		#ifdef ZEND_INTRIN_SSE4_2_PCLMUL_RESOLVER
314		static size_t crc32_x86_simd_update_default(X86_CRC32_TYPE type, uint32_t crc, const unsigned char p, size_t nr)
315	0	{
316	0	return 0;
317	0	}
318
319		# ifdef ZEND_INTRIN_SSE4_2_PCLMUL_FUNC_PROTO
320		size_t crc32_x86_simd_update(X86_CRC32_TYPE type, uint32_t crc, const unsigned char p, size_t nr) __attribute__((ifunc("resolve_crc32_x86_simd_update")));
321
322		typedef size_t (crc32_x86_simd_func_t)(X86_CRC32_TYPE type, uint32_t crc, const unsigned char *p, size_t nr);
323
324		ZEND_NO_SANITIZE_ADDRESS
325		ZEND_ATTRIBUTE_UNUSED /* clang mistakenly warns about this */
326		static crc32_x86_simd_func_t resolve_crc32_x86_simd_update(void) {
327		if (zend_cpu_supports_sse42() && zend_cpu_supports_pclmul()) {
328		return crc32_sse42_pclmul_update;
329		}
330		return crc32_x86_simd_update_default;
331		}
332		# else /* ZEND_INTRIN_SSE4_2_PCLMUL_FUNC_PTR */
333		static size_t (crc32_x86_simd_ptr)(X86_CRC32_TYPE type, uint32_t crc, const unsigned char *p, size_t nr) = crc32_x86_simd_update_default;
334
335	93	size_t crc32_x86_simd_update(X86_CRC32_TYPE type, uint32_t crc, const unsigned char p, size_t nr) {
336	93	return crc32_x86_simd_ptr(type, crc, p, nr);
337	93	}
338
339		/* {{{ PHP_MINIT_FUNCTION */
340		PHP_MINIT_FUNCTION(crc32_x86_intrin)
341	14	{
342	14	if (zend_cpu_supports_sse42() && zend_cpu_supports_pclmul()) {
343	14	crc32_x86_simd_ptr = crc32_sse42_pclmul_update;
344	14	}
345	14	return SUCCESS;
346	14	}
347		/* }}} */
348		# endif
349		#endif