| 1 | /* insert_string.h |
| 2 | * |
| 3 | * Copyright 2019 The Chromium Authors. All rights reserved. |
| 4 | * Use of this source code is governed by a BSD-style license that can be |
| 5 | * found in the Chromium source repository LICENSE file. |
| 6 | */ |
| 7 | |
| 8 | #if defined(_MSC_VER) |
| 9 | #define INLINE __inline |
| 10 | #else |
| 11 | #define INLINE inline |
| 12 | #endif |
| 13 | |
| 14 | #include "cpu_features.h" |
| 15 | |
| 16 | // clang-format off |
| 17 | #if defined(CRC32_SIMD_SSE42_PCLMUL) |
| 18 | #include <smmintrin.h> /* Required to make MSVC bot build pass. */ |
| 19 | |
| 20 | #if defined(__clang__) || defined(__GNUC__) |
| 21 | #define TARGET_CPU_WITH_CRC __attribute__((target("sse4.2"))) |
| 22 | #else |
| 23 | #define TARGET_CPU_WITH_CRC |
| 24 | #endif |
| 25 | |
| 26 | #define _cpu_crc32_u32 _mm_crc32_u32 |
| 27 | |
| 28 | #elif defined(CRC32_ARMV8_CRC32) |
| 29 | #if defined(__clang__) |
| 30 | #define __crc32cw __builtin_arm_crc32cw |
| 31 | #endif |
| 32 | |
| 33 | #if defined(__aarch64__) |
| 34 | #define TARGET_CPU_WITH_CRC __attribute__((target("crc"))) |
| 35 | #else // !defined(__aarch64__) |
| 36 | #define TARGET_CPU_WITH_CRC __attribute__((target("armv8-a,crc"))) |
| 37 | #endif // defined(__aarch64__) |
| 38 | |
| 39 | #define _cpu_crc32_u32 __crc32cw |
| 40 | |
| 41 | #endif |
| 42 | // clang-format on |
| 43 | |
| 44 | #if defined(TARGET_CPU_WITH_CRC) |
| 45 | |
| 46 | TARGET_CPU_WITH_CRC |
| 47 | local INLINE Pos insert_string_simd(deflate_state* const s, const Pos str) { |
| 48 | Pos ret; |
| 49 | unsigned *ip, val, h = 0; |
| 50 | |
| 51 | ip = (unsigned*)&s->window[str]; |
| 52 | val = *ip; |
| 53 | |
| 54 | if (s->level >= 6) |
| 55 | val &= 0xFFFFFF; |
| 56 | |
| 57 | /* Unlike the case of data integrity checks for GZIP format where the |
| 58 | * polynomial used is defined (https://tools.ietf.org/html/rfc1952#page-11), |
| 59 | * here it is just a hash function for the hash table used while |
| 60 | * performing compression. |
| 61 | */ |
| 62 | h = _cpu_crc32_u32(h, val); |
| 63 | |
| 64 | ret = s->head[h & s->hash_mask]; |
| 65 | s->head[h & s->hash_mask] = str; |
| 66 | s->prev[str & s->w_mask] = ret; |
| 67 | return ret; |
| 68 | } |
| 69 | |
| 70 | #endif // TARGET_CPU_WITH_CRC |
| 71 | |
| 72 | /* =========================================================================== |
| 73 | * Update a hash value with the given input byte |
| 74 | * IN assertion: all calls to UPDATE_HASH are made with consecutive input |
| 75 | * characters, so that a running hash key can be computed from the previous |
| 76 | * key instead of complete recalculation each time. |
| 77 | */ |
| 78 | #define UPDATE_HASH(s, h, c) (h = (((h) << s->hash_shift) ^ (c)) & s->hash_mask) |
| 79 | |
| 80 | /* =========================================================================== |
| 81 | * Insert string str in the dictionary and set match_head to the previous head |
| 82 | * of the hash chain (the most recent string with same hash key). Return |
| 83 | * the previous length of the hash chain. |
| 84 | * If this file is compiled with -DFASTEST, the compression level is forced |
| 85 | * to 1, and no hash chains are maintained. |
| 86 | * IN assertion: all calls to INSERT_STRING are made with consecutive input |
| 87 | * characters and the first MIN_MATCH bytes of str are valid (except for |
| 88 | * the last MIN_MATCH-1 bytes of the input file). |
| 89 | */ |
| 90 | local INLINE Pos insert_string_c(deflate_state* const s, const Pos str) { |
| 91 | Pos ret; |
| 92 | |
| 93 | UPDATE_HASH(s, s->ins_h, s->window[(str) + (MIN_MATCH - 1)]); |
| 94 | #ifdef FASTEST |
| 95 | ret = s->head[s->ins_h]; |
| 96 | #else |
| 97 | ret = s->prev[str & s->w_mask] = s->head[s->ins_h]; |
| 98 | #endif |
| 99 | s->head[s->ins_h] = str; |
| 100 | |
| 101 | return ret; |
| 102 | } |
| 103 | |
| 104 | local INLINE Pos insert_string(deflate_state* const s, const Pos str) { |
| 105 | /* insert_string_simd string dictionary insertion: this SIMD symbol hashing |
| 106 | * significantly improves data compression speed. |
| 107 | * |
| 108 | * Note: the generated compressed output is a valid DEFLATE stream but will |
| 109 | * differ from vanilla zlib output ... |
| 110 | */ |
| 111 | #if defined(CHROMIUM_ZLIB_NO_CASTAGNOLI) |
| 112 | /* ... so this build-time option can used to disable the SIMD symbol hasher |
| 113 | * if matching vanilla zlib DEFLATE output is required. |
| 114 | */ (;) /* FALLTHOUGH */ |
| 115 | #elif defined(TARGET_CPU_WITH_CRC) && defined(CRC32_SIMD_SSE42_PCLMUL) |
| 116 | if (x86_cpu_enable_simd) |
| 117 | return insert_string_simd(s, str); |
| 118 | #elif defined(TARGET_CPU_WITH_CRC) && defined(CRC32_ARMV8_CRC32) |
| 119 | if (arm_cpu_enable_crc32) |
| 120 | return insert_string_simd(s, str); |
| 121 | #endif |
| 122 | return insert_string_c(s, str); |
| 123 | } |
| 124 | |