| 1 | /* crc32.c -- compute the CRC-32 of a data stream |
| 2 | * Copyright (C) 1995-2006, 2010, 2011, 2012, 2016, 2018 Mark Adler |
| 3 | * For conditions of distribution and use, see copyright notice in zlib.h |
| 4 | * |
| 5 | * Thanks to Rodney Brown <rbrown64@csc.com.au> for his contribution of faster |
| 6 | * CRC methods: exclusive-oring 32 bits of data at a time, and pre-computing |
| 7 | * tables for updating the shift register in one step with three exclusive-ors |
| 8 | * instead of four steps with four exclusive-ors. This results in about a |
| 9 | * factor of two increase in speed on a Power PC G4 (PPC7455) using gcc -O3. |
| 10 | */ |
| 11 | |
| 12 | #include "zbuild.h" |
| 13 | #include "zendian.h" |
| 14 | #include <inttypes.h> |
| 15 | #include "deflate.h" |
| 16 | #include "functable.h" |
| 17 | #include "crc32_tbl.h" |
| 18 | |
| 19 | /* ========================================================================= |
| 20 | * This function can be used by asm versions of crc32() |
| 21 | */ |
| 22 | const uint32_t * Z_EXPORT PREFIX(get_crc_table)(void) { |
| 23 | return (const uint32_t *)crc_table; |
| 24 | } |
| 25 | |
| 26 | #ifdef ZLIB_COMPAT |
| 27 | unsigned long Z_EXPORT PREFIX(crc32_z)(unsigned long crc, const unsigned char *buf, size_t len) { |
| 28 | if (buf == NULL) return 0; |
| 29 | |
| 30 | return (unsigned long)functable.crc32((uint32_t)crc, buf, len); |
| 31 | } |
| 32 | #else |
| 33 | uint32_t Z_EXPORT PREFIX(crc32_z)(uint32_t crc, const unsigned char *buf, size_t len) { |
| 34 | if (buf == NULL) return 0; |
| 35 | |
| 36 | return functable.crc32(crc, buf, len); |
| 37 | } |
| 38 | #endif |
| 39 | /* ========================================================================= */ |
| 40 | #define DO1 crc = crc_table[0][((int)crc ^ (*buf++)) & 0xff] ^ (crc >> 8) |
| 41 | #define DO8 DO1; DO1; DO1; DO1; DO1; DO1; DO1; DO1 |
| 42 | #define DO4 DO1; DO1; DO1; DO1 |
| 43 | |
| 44 | /* ========================================================================= */ |
| 45 | Z_INTERNAL uint32_t crc32_generic(uint32_t crc, const unsigned char *buf, uint64_t len) { |
| 46 | crc = crc ^ 0xffffffff; |
| 47 | |
| 48 | #ifdef UNROLL_MORE |
| 49 | while (len >= 8) { |
| 50 | DO8; |
| 51 | len -= 8; |
| 52 | } |
| 53 | #else |
| 54 | while (len >= 4) { |
| 55 | DO4; |
| 56 | len -= 4; |
| 57 | } |
| 58 | #endif |
| 59 | |
| 60 | if (len) do { |
| 61 | DO1; |
| 62 | } while (--len); |
| 63 | return crc ^ 0xffffffff; |
| 64 | } |
| 65 | |
| 66 | #ifdef ZLIB_COMPAT |
| 67 | unsigned long Z_EXPORT PREFIX(crc32)(unsigned long crc, const unsigned char *buf, unsigned int len) { |
| 68 | return (unsigned long)PREFIX(crc32_z)(crc: (uint32_t)crc, buf, len); |
| 69 | } |
| 70 | #else |
| 71 | uint32_t Z_EXPORT PREFIX(crc32)(uint32_t crc, const unsigned char *buf, uint32_t len) { |
| 72 | return PREFIX(crc32_z)(crc, buf, len); |
| 73 | } |
| 74 | #endif |
| 75 | |
| 76 | /* |
| 77 | This BYFOUR code accesses the passed unsigned char * buffer with a 32-bit |
| 78 | integer pointer type. This violates the strict aliasing rule, where a |
| 79 | compiler can assume, for optimization purposes, that two pointers to |
| 80 | fundamentally different types won't ever point to the same memory. This can |
| 81 | manifest as a problem only if one of the pointers is written to. This code |
| 82 | only reads from those pointers. So long as this code remains isolated in |
| 83 | this compilation unit, there won't be a problem. For this reason, this code |
| 84 | should not be copied and pasted into a compilation unit in which other code |
| 85 | writes to the buffer that is passed to these routines. |
| 86 | */ |
| 87 | |
| 88 | /* ========================================================================= */ |
| 89 | #if BYTE_ORDER == LITTLE_ENDIAN |
| 90 | #define DOLIT4 c ^= *buf4++; \ |
| 91 | c = crc_table[3][c & 0xff] ^ crc_table[2][(c >> 8) & 0xff] ^ \ |
| 92 | crc_table[1][(c >> 16) & 0xff] ^ crc_table[0][c >> 24] |
| 93 | #define DOLIT32 DOLIT4; DOLIT4; DOLIT4; DOLIT4; DOLIT4; DOLIT4; DOLIT4; DOLIT4 |
| 94 | |
| 95 | /* ========================================================================= */ |
| 96 | Z_INTERNAL uint32_t crc32_little(uint32_t crc, const unsigned char *buf, uint64_t len) { |
| 97 | Z_REGISTER uint32_t c; |
| 98 | Z_REGISTER const uint32_t *buf4; |
| 99 | |
| 100 | c = crc; |
| 101 | c = ~c; |
| 102 | while (len && ((ptrdiff_t)buf & 3)) { |
| 103 | c = crc_table[0][(c ^ *buf++) & 0xff] ^ (c >> 8); |
| 104 | len--; |
| 105 | } |
| 106 | |
| 107 | buf4 = (const uint32_t *)(const void *)buf; |
| 108 | |
| 109 | #ifdef UNROLL_MORE |
| 110 | while (len >= 32) { |
| 111 | DOLIT32; |
| 112 | len -= 32; |
| 113 | } |
| 114 | #endif |
| 115 | |
| 116 | while (len >= 4) { |
| 117 | DOLIT4; |
| 118 | len -= 4; |
| 119 | } |
| 120 | buf = (const unsigned char *)buf4; |
| 121 | |
| 122 | if (len) do { |
| 123 | c = crc_table[0][(c ^ *buf++) & 0xff] ^ (c >> 8); |
| 124 | } while (--len); |
| 125 | c = ~c; |
| 126 | return c; |
| 127 | } |
| 128 | #endif /* BYTE_ORDER == LITTLE_ENDIAN */ |
| 129 | |
| 130 | /* ========================================================================= */ |
| 131 | #if BYTE_ORDER == BIG_ENDIAN |
| 132 | #define DOBIG4 c ^= *buf4++; \ |
| 133 | c = crc_table[4][c & 0xff] ^ crc_table[5][(c >> 8) & 0xff] ^ \ |
| 134 | crc_table[6][(c >> 16) & 0xff] ^ crc_table[7][c >> 24] |
| 135 | #define DOBIG32 DOBIG4; DOBIG4; DOBIG4; DOBIG4; DOBIG4; DOBIG4; DOBIG4; DOBIG4 |
| 136 | |
| 137 | /* ========================================================================= */ |
| 138 | Z_INTERNAL uint32_t crc32_big(uint32_t crc, const unsigned char *buf, uint64_t len) { |
| 139 | Z_REGISTER uint32_t c; |
| 140 | Z_REGISTER const uint32_t *buf4; |
| 141 | |
| 142 | c = ZSWAP32(crc); |
| 143 | c = ~c; |
| 144 | while (len && ((ptrdiff_t)buf & 3)) { |
| 145 | c = crc_table[4][(c >> 24) ^ *buf++] ^ (c << 8); |
| 146 | len--; |
| 147 | } |
| 148 | |
| 149 | buf4 = (const uint32_t *)(const void *)buf; |
| 150 | |
| 151 | #ifdef UNROLL_MORE |
| 152 | while (len >= 32) { |
| 153 | DOBIG32; |
| 154 | len -= 32; |
| 155 | } |
| 156 | #endif |
| 157 | |
| 158 | while (len >= 4) { |
| 159 | DOBIG4; |
| 160 | len -= 4; |
| 161 | } |
| 162 | buf = (const unsigned char *)buf4; |
| 163 | |
| 164 | if (len) do { |
| 165 | c = crc_table[4][(c >> 24) ^ *buf++] ^ (c << 8); |
| 166 | } while (--len); |
| 167 | c = ~c; |
| 168 | return ZSWAP32(c); |
| 169 | } |
| 170 | #endif /* BYTE_ORDER == BIG_ENDIAN */ |
| 171 | |
| 172 | #ifdef X86_PCLMULQDQ_CRC |
| 173 | #include "arch/x86/x86.h" |
| 174 | #include "arch/x86/crc_folding.h" |
| 175 | |
| 176 | Z_INTERNAL void crc_finalize(deflate_state *const s) { |
| 177 | if (x86_cpu_has_pclmulqdq) |
| 178 | s->strm->adler = crc_fold_512to32(s); |
| 179 | } |
| 180 | #endif |
| 181 | |
| 182 | Z_INTERNAL void crc_reset(deflate_state *const s) { |
| 183 | #ifdef X86_PCLMULQDQ_CRC |
| 184 | x86_check_features(); |
| 185 | if (x86_cpu_has_pclmulqdq) { |
| 186 | crc_fold_init(s); |
| 187 | return; |
| 188 | } |
| 189 | #endif |
| 190 | s->strm->adler = PREFIX(crc32)(crc: 0L, NULL, len: 0); |
| 191 | } |
| 192 | |
| 193 | Z_INTERNAL void copy_with_crc(PREFIX3(stream) *strm, unsigned char *dst, unsigned long size) { |
| 194 | #ifdef X86_PCLMULQDQ_CRC |
| 195 | if (x86_cpu_has_pclmulqdq) { |
| 196 | crc_fold_copy(strm->state, dst, strm->next_in, size); |
| 197 | return; |
| 198 | } |
| 199 | #endif |
| 200 | memcpy(dest: dst, src: strm->next_in, n: size); |
| 201 | strm->adler = PREFIX(crc32)(crc: strm->adler, buf: dst, len: size); |
| 202 | } |
| 203 | |