/src/rocksdb/util/crc32c.cc
Line | Count | Source (jump to first uncovered line) |
1 | | // Copyright (c) 2011-present, Facebook, Inc. All rights reserved. |
2 | | // This source code is licensed under both the GPLv2 (found in the |
3 | | // COPYING file in the root directory) and Apache 2.0 License |
4 | | // (found in the LICENSE.Apache file in the root directory). |
5 | | // |
6 | | // Copyright (c) 2011 The LevelDB Authors. All rights reserved. |
7 | | // Use of this source code is governed by a BSD-style license that can be |
8 | | // found in the LICENSE file. See the AUTHORS file for names of contributors. |
9 | | // |
10 | | // A portable implementation of crc32c, optimized to handle |
11 | | // four bytes at a time. |
12 | | #include "util/crc32c.h" |
13 | | |
14 | | #include <array> |
15 | | #include <cstdint> |
16 | | #include <utility> |
17 | | |
18 | | #include "port/lang.h" |
19 | | #include "util/coding.h" |
20 | | #include "util/crc32c_arm64.h" |
21 | | #include "util/math.h" |
22 | | |
23 | | #ifdef __powerpc64__ |
24 | | #include "util/crc32c_ppc.h" |
25 | | #include "util/crc32c_ppc_constants.h" |
26 | | |
27 | | #if __linux__ |
28 | | #ifdef ROCKSDB_AUXV_GETAUXVAL_PRESENT |
29 | | #include <sys/auxv.h> |
30 | | #endif |
31 | | |
32 | | #ifndef PPC_FEATURE2_VEC_CRYPTO |
33 | | #define PPC_FEATURE2_VEC_CRYPTO 0x02000000 |
34 | | #endif |
35 | | |
36 | | #ifndef AT_HWCAP2 |
37 | | #define AT_HWCAP2 26 |
38 | | #endif |
39 | | |
40 | | #elif __FreeBSD__ |
41 | | #include <machine/cpu.h> |
42 | | #include <sys/auxv.h> |
43 | | #include <sys/elf_common.h> |
44 | | #endif /* __linux__ */ |
45 | | |
46 | | #endif |
47 | | |
48 | | ASSERT_FEATURE_COMPAT_HEADER(); |
49 | | |
50 | | #ifdef __SSE4_2__ |
51 | | #include <nmmintrin.h> |
52 | | #include <wmmintrin.h> |
53 | | #endif |
54 | | |
55 | | #if defined(HAVE_ARM64_CRC) |
56 | | bool pmull_runtime_flag = false; |
57 | | #endif |
58 | | |
59 | | namespace ROCKSDB_NAMESPACE::crc32c { |
60 | | |
61 | | #if defined(HAVE_POWER8) && defined(HAS_ALTIVEC) |
62 | | #ifdef __powerpc64__ |
63 | | static int arch_ppc_crc32 = 0; |
64 | | #endif /* __powerpc64__ */ |
65 | | #endif |
66 | | |
67 | | static const uint32_t table0_[256] = { |
68 | | 0x00000000, 0xf26b8303, 0xe13b70f7, 0x1350f3f4, 0xc79a971f, 0x35f1141c, |
69 | | 0x26a1e7e8, 0xd4ca64eb, 0x8ad958cf, 0x78b2dbcc, 0x6be22838, 0x9989ab3b, |
70 | | 0x4d43cfd0, 0xbf284cd3, 0xac78bf27, 0x5e133c24, 0x105ec76f, 0xe235446c, |
71 | | 0xf165b798, 0x030e349b, 0xd7c45070, 0x25afd373, 0x36ff2087, 0xc494a384, |
72 | | 0x9a879fa0, 0x68ec1ca3, 0x7bbcef57, 0x89d76c54, 0x5d1d08bf, 0xaf768bbc, |
73 | | 0xbc267848, 0x4e4dfb4b, 0x20bd8ede, 0xd2d60ddd, 0xc186fe29, 0x33ed7d2a, |
74 | | 0xe72719c1, 0x154c9ac2, 0x061c6936, 0xf477ea35, 0xaa64d611, 0x580f5512, |
75 | | 0x4b5fa6e6, 0xb93425e5, 0x6dfe410e, 0x9f95c20d, 0x8cc531f9, 0x7eaeb2fa, |
76 | | 0x30e349b1, 0xc288cab2, 0xd1d83946, 0x23b3ba45, 0xf779deae, 0x05125dad, |
77 | | 0x1642ae59, 0xe4292d5a, 0xba3a117e, 0x4851927d, 0x5b016189, 0xa96ae28a, |
78 | | 0x7da08661, 0x8fcb0562, 0x9c9bf696, 0x6ef07595, 0x417b1dbc, 0xb3109ebf, |
79 | | 0xa0406d4b, 0x522bee48, 0x86e18aa3, 0x748a09a0, 0x67dafa54, 0x95b17957, |
80 | | 0xcba24573, 0x39c9c670, 0x2a993584, 0xd8f2b687, 0x0c38d26c, 0xfe53516f, |
81 | | 0xed03a29b, 0x1f682198, 0x5125dad3, 0xa34e59d0, 0xb01eaa24, 0x42752927, |
82 | | 0x96bf4dcc, 0x64d4cecf, 0x77843d3b, 0x85efbe38, 0xdbfc821c, 0x2997011f, |
83 | | 0x3ac7f2eb, 0xc8ac71e8, 0x1c661503, 0xee0d9600, 0xfd5d65f4, 0x0f36e6f7, |
84 | | 0x61c69362, 0x93ad1061, 0x80fde395, 0x72966096, 0xa65c047d, 0x5437877e, |
85 | | 0x4767748a, 0xb50cf789, 0xeb1fcbad, 0x197448ae, 0x0a24bb5a, 0xf84f3859, |
86 | | 0x2c855cb2, 0xdeeedfb1, 0xcdbe2c45, 0x3fd5af46, 0x7198540d, 0x83f3d70e, |
87 | | 0x90a324fa, 0x62c8a7f9, 0xb602c312, 0x44694011, 0x5739b3e5, 0xa55230e6, |
88 | | 0xfb410cc2, 0x092a8fc1, 0x1a7a7c35, 0xe811ff36, 0x3cdb9bdd, 0xceb018de, |
89 | | 0xdde0eb2a, 0x2f8b6829, 0x82f63b78, 0x709db87b, 0x63cd4b8f, 0x91a6c88c, |
90 | | 0x456cac67, 0xb7072f64, 0xa457dc90, 0x563c5f93, 0x082f63b7, 0xfa44e0b4, |
91 | | 0xe9141340, 0x1b7f9043, 0xcfb5f4a8, 0x3dde77ab, 0x2e8e845f, 0xdce5075c, |
92 | | 0x92a8fc17, 0x60c37f14, 0x73938ce0, 0x81f80fe3, 0x55326b08, 0xa759e80b, |
93 | | 0xb4091bff, 0x466298fc, 0x1871a4d8, 0xea1a27db, 0xf94ad42f, 0x0b21572c, |
94 | | 0xdfeb33c7, 0x2d80b0c4, 0x3ed04330, 0xccbbc033, 0xa24bb5a6, 0x502036a5, |
95 | | 0x4370c551, 0xb11b4652, 0x65d122b9, 0x97baa1ba, 0x84ea524e, 0x7681d14d, |
96 | | 0x2892ed69, 0xdaf96e6a, 0xc9a99d9e, 0x3bc21e9d, 0xef087a76, 0x1d63f975, |
97 | | 0x0e330a81, 0xfc588982, 0xb21572c9, 0x407ef1ca, 0x532e023e, 0xa145813d, |
98 | | 0x758fe5d6, 0x87e466d5, 0x94b49521, 0x66df1622, 0x38cc2a06, 0xcaa7a905, |
99 | | 0xd9f75af1, 0x2b9cd9f2, 0xff56bd19, 0x0d3d3e1a, 0x1e6dcdee, 0xec064eed, |
100 | | 0xc38d26c4, 0x31e6a5c7, 0x22b65633, 0xd0ddd530, 0x0417b1db, 0xf67c32d8, |
101 | | 0xe52cc12c, 0x1747422f, 0x49547e0b, 0xbb3ffd08, 0xa86f0efc, 0x5a048dff, |
102 | | 0x8ecee914, 0x7ca56a17, 0x6ff599e3, 0x9d9e1ae0, 0xd3d3e1ab, 0x21b862a8, |
103 | | 0x32e8915c, 0xc083125f, 0x144976b4, 0xe622f5b7, 0xf5720643, 0x07198540, |
104 | | 0x590ab964, 0xab613a67, 0xb831c993, 0x4a5a4a90, 0x9e902e7b, 0x6cfbad78, |
105 | | 0x7fab5e8c, 0x8dc0dd8f, 0xe330a81a, 0x115b2b19, 0x020bd8ed, 0xf0605bee, |
106 | | 0x24aa3f05, 0xd6c1bc06, 0xc5914ff2, 0x37faccf1, 0x69e9f0d5, 0x9b8273d6, |
107 | | 0x88d28022, 0x7ab90321, 0xae7367ca, 0x5c18e4c9, 0x4f48173d, 0xbd23943e, |
108 | | 0xf36e6f75, 0x0105ec76, 0x12551f82, 0xe03e9c81, 0x34f4f86a, 0xc69f7b69, |
109 | | 0xd5cf889d, 0x27a40b9e, 0x79b737ba, 0x8bdcb4b9, 0x988c474d, 0x6ae7c44e, |
110 | | 0xbe2da0a5, 0x4c4623a6, 0x5f16d052, 0xad7d5351}; |
111 | | #ifndef __SSE4_2__ |
112 | | static const uint32_t table1_[256] = { |
113 | | 0x00000000, 0x13a29877, 0x274530ee, 0x34e7a899, 0x4e8a61dc, 0x5d28f9ab, |
114 | | 0x69cf5132, 0x7a6dc945, 0x9d14c3b8, 0x8eb65bcf, 0xba51f356, 0xa9f36b21, |
115 | | 0xd39ea264, 0xc03c3a13, 0xf4db928a, 0xe7790afd, 0x3fc5f181, 0x2c6769f6, |
116 | | 0x1880c16f, 0x0b225918, 0x714f905d, 0x62ed082a, 0x560aa0b3, 0x45a838c4, |
117 | | 0xa2d13239, 0xb173aa4e, 0x859402d7, 0x96369aa0, 0xec5b53e5, 0xfff9cb92, |
118 | | 0xcb1e630b, 0xd8bcfb7c, 0x7f8be302, 0x6c297b75, 0x58ced3ec, 0x4b6c4b9b, |
119 | | 0x310182de, 0x22a31aa9, 0x1644b230, 0x05e62a47, 0xe29f20ba, 0xf13db8cd, |
120 | | 0xc5da1054, 0xd6788823, 0xac154166, 0xbfb7d911, 0x8b507188, 0x98f2e9ff, |
121 | | 0x404e1283, 0x53ec8af4, 0x670b226d, 0x74a9ba1a, 0x0ec4735f, 0x1d66eb28, |
122 | | 0x298143b1, 0x3a23dbc6, 0xdd5ad13b, 0xcef8494c, 0xfa1fe1d5, 0xe9bd79a2, |
123 | | 0x93d0b0e7, 0x80722890, 0xb4958009, 0xa737187e, 0xff17c604, 0xecb55e73, |
124 | | 0xd852f6ea, 0xcbf06e9d, 0xb19da7d8, 0xa23f3faf, 0x96d89736, 0x857a0f41, |
125 | | 0x620305bc, 0x71a19dcb, 0x45463552, 0x56e4ad25, 0x2c896460, 0x3f2bfc17, |
126 | | 0x0bcc548e, 0x186eccf9, 0xc0d23785, 0xd370aff2, 0xe797076b, 0xf4359f1c, |
127 | | 0x8e585659, 0x9dface2e, 0xa91d66b7, 0xbabffec0, 0x5dc6f43d, 0x4e646c4a, |
128 | | 0x7a83c4d3, 0x69215ca4, 0x134c95e1, 0x00ee0d96, 0x3409a50f, 0x27ab3d78, |
129 | | 0x809c2506, 0x933ebd71, 0xa7d915e8, 0xb47b8d9f, 0xce1644da, 0xddb4dcad, |
130 | | 0xe9537434, 0xfaf1ec43, 0x1d88e6be, 0x0e2a7ec9, 0x3acdd650, 0x296f4e27, |
131 | | 0x53028762, 0x40a01f15, 0x7447b78c, 0x67e52ffb, 0xbf59d487, 0xacfb4cf0, |
132 | | 0x981ce469, 0x8bbe7c1e, 0xf1d3b55b, 0xe2712d2c, 0xd69685b5, 0xc5341dc2, |
133 | | 0x224d173f, 0x31ef8f48, 0x050827d1, 0x16aabfa6, 0x6cc776e3, 0x7f65ee94, |
134 | | 0x4b82460d, 0x5820de7a, 0xfbc3faf9, 0xe861628e, 0xdc86ca17, 0xcf245260, |
135 | | 0xb5499b25, 0xa6eb0352, 0x920cabcb, 0x81ae33bc, 0x66d73941, 0x7575a136, |
136 | | 0x419209af, 0x523091d8, 0x285d589d, 0x3bffc0ea, 0x0f186873, 0x1cbaf004, |
137 | | 0xc4060b78, 0xd7a4930f, 0xe3433b96, 0xf0e1a3e1, 0x8a8c6aa4, 0x992ef2d3, |
138 | | 0xadc95a4a, 0xbe6bc23d, 0x5912c8c0, 0x4ab050b7, 0x7e57f82e, 0x6df56059, |
139 | | 0x1798a91c, 0x043a316b, 0x30dd99f2, 0x237f0185, 0x844819fb, 0x97ea818c, |
140 | | 0xa30d2915, 0xb0afb162, 0xcac27827, 0xd960e050, 0xed8748c9, 0xfe25d0be, |
141 | | 0x195cda43, 0x0afe4234, 0x3e19eaad, 0x2dbb72da, 0x57d6bb9f, 0x447423e8, |
142 | | 0x70938b71, 0x63311306, 0xbb8de87a, 0xa82f700d, 0x9cc8d894, 0x8f6a40e3, |
143 | | 0xf50789a6, 0xe6a511d1, 0xd242b948, 0xc1e0213f, 0x26992bc2, 0x353bb3b5, |
144 | | 0x01dc1b2c, 0x127e835b, 0x68134a1e, 0x7bb1d269, 0x4f567af0, 0x5cf4e287, |
145 | | 0x04d43cfd, 0x1776a48a, 0x23910c13, 0x30339464, 0x4a5e5d21, 0x59fcc556, |
146 | | 0x6d1b6dcf, 0x7eb9f5b8, 0x99c0ff45, 0x8a626732, 0xbe85cfab, 0xad2757dc, |
147 | | 0xd74a9e99, 0xc4e806ee, 0xf00fae77, 0xe3ad3600, 0x3b11cd7c, 0x28b3550b, |
148 | | 0x1c54fd92, 0x0ff665e5, 0x759baca0, 0x663934d7, 0x52de9c4e, 0x417c0439, |
149 | | 0xa6050ec4, 0xb5a796b3, 0x81403e2a, 0x92e2a65d, 0xe88f6f18, 0xfb2df76f, |
150 | | 0xcfca5ff6, 0xdc68c781, 0x7b5fdfff, 0x68fd4788, 0x5c1aef11, 0x4fb87766, |
151 | | 0x35d5be23, 0x26772654, 0x12908ecd, 0x013216ba, 0xe64b1c47, 0xf5e98430, |
152 | | 0xc10e2ca9, 0xd2acb4de, 0xa8c17d9b, 0xbb63e5ec, 0x8f844d75, 0x9c26d502, |
153 | | 0x449a2e7e, 0x5738b609, 0x63df1e90, 0x707d86e7, 0x0a104fa2, 0x19b2d7d5, |
154 | | 0x2d557f4c, 0x3ef7e73b, 0xd98eedc6, 0xca2c75b1, 0xfecbdd28, 0xed69455f, |
155 | | 0x97048c1a, 0x84a6146d, 0xb041bcf4, 0xa3e32483}; |
156 | | static const uint32_t table2_[256] = { |
157 | | 0x00000000, 0xa541927e, 0x4f6f520d, 0xea2ec073, 0x9edea41a, 0x3b9f3664, |
158 | | 0xd1b1f617, 0x74f06469, 0x38513ec5, 0x9d10acbb, 0x773e6cc8, 0xd27ffeb6, |
159 | | 0xa68f9adf, 0x03ce08a1, 0xe9e0c8d2, 0x4ca15aac, 0x70a27d8a, 0xd5e3eff4, |
160 | | 0x3fcd2f87, 0x9a8cbdf9, 0xee7cd990, 0x4b3d4bee, 0xa1138b9d, 0x045219e3, |
161 | | 0x48f3434f, 0xedb2d131, 0x079c1142, 0xa2dd833c, 0xd62de755, 0x736c752b, |
162 | | 0x9942b558, 0x3c032726, 0xe144fb14, 0x4405696a, 0xae2ba919, 0x0b6a3b67, |
163 | | 0x7f9a5f0e, 0xdadbcd70, 0x30f50d03, 0x95b49f7d, 0xd915c5d1, 0x7c5457af, |
164 | | 0x967a97dc, 0x333b05a2, 0x47cb61cb, 0xe28af3b5, 0x08a433c6, 0xade5a1b8, |
165 | | 0x91e6869e, 0x34a714e0, 0xde89d493, 0x7bc846ed, 0x0f382284, 0xaa79b0fa, |
166 | | 0x40577089, 0xe516e2f7, 0xa9b7b85b, 0x0cf62a25, 0xe6d8ea56, 0x43997828, |
167 | | 0x37691c41, 0x92288e3f, 0x78064e4c, 0xdd47dc32, 0xc76580d9, 0x622412a7, |
168 | | 0x880ad2d4, 0x2d4b40aa, 0x59bb24c3, 0xfcfab6bd, 0x16d476ce, 0xb395e4b0, |
169 | | 0xff34be1c, 0x5a752c62, 0xb05bec11, 0x151a7e6f, 0x61ea1a06, 0xc4ab8878, |
170 | | 0x2e85480b, 0x8bc4da75, 0xb7c7fd53, 0x12866f2d, 0xf8a8af5e, 0x5de93d20, |
171 | | 0x29195949, 0x8c58cb37, 0x66760b44, 0xc337993a, 0x8f96c396, 0x2ad751e8, |
172 | | 0xc0f9919b, 0x65b803e5, 0x1148678c, 0xb409f5f2, 0x5e273581, 0xfb66a7ff, |
173 | | 0x26217bcd, 0x8360e9b3, 0x694e29c0, 0xcc0fbbbe, 0xb8ffdfd7, 0x1dbe4da9, |
174 | | 0xf7908dda, 0x52d11fa4, 0x1e704508, 0xbb31d776, 0x511f1705, 0xf45e857b, |
175 | | 0x80aee112, 0x25ef736c, 0xcfc1b31f, 0x6a802161, 0x56830647, 0xf3c29439, |
176 | | 0x19ec544a, 0xbcadc634, 0xc85da25d, 0x6d1c3023, 0x8732f050, 0x2273622e, |
177 | | 0x6ed23882, 0xcb93aafc, 0x21bd6a8f, 0x84fcf8f1, 0xf00c9c98, 0x554d0ee6, |
178 | | 0xbf63ce95, 0x1a225ceb, 0x8b277743, 0x2e66e53d, 0xc448254e, 0x6109b730, |
179 | | 0x15f9d359, 0xb0b84127, 0x5a968154, 0xffd7132a, 0xb3764986, 0x1637dbf8, |
180 | | 0xfc191b8b, 0x595889f5, 0x2da8ed9c, 0x88e97fe2, 0x62c7bf91, 0xc7862def, |
181 | | 0xfb850ac9, 0x5ec498b7, 0xb4ea58c4, 0x11abcaba, 0x655baed3, 0xc01a3cad, |
182 | | 0x2a34fcde, 0x8f756ea0, 0xc3d4340c, 0x6695a672, 0x8cbb6601, 0x29faf47f, |
183 | | 0x5d0a9016, 0xf84b0268, 0x1265c21b, 0xb7245065, 0x6a638c57, 0xcf221e29, |
184 | | 0x250cde5a, 0x804d4c24, 0xf4bd284d, 0x51fcba33, 0xbbd27a40, 0x1e93e83e, |
185 | | 0x5232b292, 0xf77320ec, 0x1d5de09f, 0xb81c72e1, 0xccec1688, 0x69ad84f6, |
186 | | 0x83834485, 0x26c2d6fb, 0x1ac1f1dd, 0xbf8063a3, 0x55aea3d0, 0xf0ef31ae, |
187 | | 0x841f55c7, 0x215ec7b9, 0xcb7007ca, 0x6e3195b4, 0x2290cf18, 0x87d15d66, |
188 | | 0x6dff9d15, 0xc8be0f6b, 0xbc4e6b02, 0x190ff97c, 0xf321390f, 0x5660ab71, |
189 | | 0x4c42f79a, 0xe90365e4, 0x032da597, 0xa66c37e9, 0xd29c5380, 0x77ddc1fe, |
190 | | 0x9df3018d, 0x38b293f3, 0x7413c95f, 0xd1525b21, 0x3b7c9b52, 0x9e3d092c, |
191 | | 0xeacd6d45, 0x4f8cff3b, 0xa5a23f48, 0x00e3ad36, 0x3ce08a10, 0x99a1186e, |
192 | | 0x738fd81d, 0xd6ce4a63, 0xa23e2e0a, 0x077fbc74, 0xed517c07, 0x4810ee79, |
193 | | 0x04b1b4d5, 0xa1f026ab, 0x4bdee6d8, 0xee9f74a6, 0x9a6f10cf, 0x3f2e82b1, |
194 | | 0xd50042c2, 0x7041d0bc, 0xad060c8e, 0x08479ef0, 0xe2695e83, 0x4728ccfd, |
195 | | 0x33d8a894, 0x96993aea, 0x7cb7fa99, 0xd9f668e7, 0x9557324b, 0x3016a035, |
196 | | 0xda386046, 0x7f79f238, 0x0b899651, 0xaec8042f, 0x44e6c45c, 0xe1a75622, |
197 | | 0xdda47104, 0x78e5e37a, 0x92cb2309, 0x378ab177, 0x437ad51e, 0xe63b4760, |
198 | | 0x0c158713, 0xa954156d, 0xe5f54fc1, 0x40b4ddbf, 0xaa9a1dcc, 0x0fdb8fb2, |
199 | | 0x7b2bebdb, 0xde6a79a5, 0x3444b9d6, 0x91052ba8}; |
200 | | static const uint32_t table3_[256] = { |
201 | | 0x00000000, 0xdd45aab8, 0xbf672381, 0x62228939, 0x7b2231f3, 0xa6679b4b, |
202 | | 0xc4451272, 0x1900b8ca, 0xf64463e6, 0x2b01c95e, 0x49234067, 0x9466eadf, |
203 | | 0x8d665215, 0x5023f8ad, 0x32017194, 0xef44db2c, 0xe964b13d, 0x34211b85, |
204 | | 0x560392bc, 0x8b463804, 0x924680ce, 0x4f032a76, 0x2d21a34f, 0xf06409f7, |
205 | | 0x1f20d2db, 0xc2657863, 0xa047f15a, 0x7d025be2, 0x6402e328, 0xb9474990, |
206 | | 0xdb65c0a9, 0x06206a11, 0xd725148b, 0x0a60be33, 0x6842370a, 0xb5079db2, |
207 | | 0xac072578, 0x71428fc0, 0x136006f9, 0xce25ac41, 0x2161776d, 0xfc24ddd5, |
208 | | 0x9e0654ec, 0x4343fe54, 0x5a43469e, 0x8706ec26, 0xe524651f, 0x3861cfa7, |
209 | | 0x3e41a5b6, 0xe3040f0e, 0x81268637, 0x5c632c8f, 0x45639445, 0x98263efd, |
210 | | 0xfa04b7c4, 0x27411d7c, 0xc805c650, 0x15406ce8, 0x7762e5d1, 0xaa274f69, |
211 | | 0xb327f7a3, 0x6e625d1b, 0x0c40d422, 0xd1057e9a, 0xaba65fe7, 0x76e3f55f, |
212 | | 0x14c17c66, 0xc984d6de, 0xd0846e14, 0x0dc1c4ac, 0x6fe34d95, 0xb2a6e72d, |
213 | | 0x5de23c01, 0x80a796b9, 0xe2851f80, 0x3fc0b538, 0x26c00df2, 0xfb85a74a, |
214 | | 0x99a72e73, 0x44e284cb, 0x42c2eeda, 0x9f874462, 0xfda5cd5b, 0x20e067e3, |
215 | | 0x39e0df29, 0xe4a57591, 0x8687fca8, 0x5bc25610, 0xb4868d3c, 0x69c32784, |
216 | | 0x0be1aebd, 0xd6a40405, 0xcfa4bccf, 0x12e11677, 0x70c39f4e, 0xad8635f6, |
217 | | 0x7c834b6c, 0xa1c6e1d4, 0xc3e468ed, 0x1ea1c255, 0x07a17a9f, 0xdae4d027, |
218 | | 0xb8c6591e, 0x6583f3a6, 0x8ac7288a, 0x57828232, 0x35a00b0b, 0xe8e5a1b3, |
219 | | 0xf1e51979, 0x2ca0b3c1, 0x4e823af8, 0x93c79040, 0x95e7fa51, 0x48a250e9, |
220 | | 0x2a80d9d0, 0xf7c57368, 0xeec5cba2, 0x3380611a, 0x51a2e823, 0x8ce7429b, |
221 | | 0x63a399b7, 0xbee6330f, 0xdcc4ba36, 0x0181108e, 0x1881a844, 0xc5c402fc, |
222 | | 0xa7e68bc5, 0x7aa3217d, 0x52a0c93f, 0x8fe56387, 0xedc7eabe, 0x30824006, |
223 | | 0x2982f8cc, 0xf4c75274, 0x96e5db4d, 0x4ba071f5, 0xa4e4aad9, 0x79a10061, |
224 | | 0x1b838958, 0xc6c623e0, 0xdfc69b2a, 0x02833192, 0x60a1b8ab, 0xbde41213, |
225 | | 0xbbc47802, 0x6681d2ba, 0x04a35b83, 0xd9e6f13b, 0xc0e649f1, 0x1da3e349, |
226 | | 0x7f816a70, 0xa2c4c0c8, 0x4d801be4, 0x90c5b15c, 0xf2e73865, 0x2fa292dd, |
227 | | 0x36a22a17, 0xebe780af, 0x89c50996, 0x5480a32e, 0x8585ddb4, 0x58c0770c, |
228 | | 0x3ae2fe35, 0xe7a7548d, 0xfea7ec47, 0x23e246ff, 0x41c0cfc6, 0x9c85657e, |
229 | | 0x73c1be52, 0xae8414ea, 0xcca69dd3, 0x11e3376b, 0x08e38fa1, 0xd5a62519, |
230 | | 0xb784ac20, 0x6ac10698, 0x6ce16c89, 0xb1a4c631, 0xd3864f08, 0x0ec3e5b0, |
231 | | 0x17c35d7a, 0xca86f7c2, 0xa8a47efb, 0x75e1d443, 0x9aa50f6f, 0x47e0a5d7, |
232 | | 0x25c22cee, 0xf8878656, 0xe1873e9c, 0x3cc29424, 0x5ee01d1d, 0x83a5b7a5, |
233 | | 0xf90696d8, 0x24433c60, 0x4661b559, 0x9b241fe1, 0x8224a72b, 0x5f610d93, |
234 | | 0x3d4384aa, 0xe0062e12, 0x0f42f53e, 0xd2075f86, 0xb025d6bf, 0x6d607c07, |
235 | | 0x7460c4cd, 0xa9256e75, 0xcb07e74c, 0x16424df4, 0x106227e5, 0xcd278d5d, |
236 | | 0xaf050464, 0x7240aedc, 0x6b401616, 0xb605bcae, 0xd4273597, 0x09629f2f, |
237 | | 0xe6264403, 0x3b63eebb, 0x59416782, 0x8404cd3a, 0x9d0475f0, 0x4041df48, |
238 | | 0x22635671, 0xff26fcc9, 0x2e238253, 0xf36628eb, 0x9144a1d2, 0x4c010b6a, |
239 | | 0x5501b3a0, 0x88441918, 0xea669021, 0x37233a99, 0xd867e1b5, 0x05224b0d, |
240 | | 0x6700c234, 0xba45688c, 0xa345d046, 0x7e007afe, 0x1c22f3c7, 0xc167597f, |
241 | | 0xc747336e, 0x1a0299d6, 0x782010ef, 0xa565ba57, 0xbc65029d, 0x6120a825, |
242 | | 0x0302211c, 0xde478ba4, 0x31035088, 0xec46fa30, 0x8e647309, 0x5321d9b1, |
243 | | 0x4a21617b, 0x9764cbc3, 0xf54642fa, 0x2803e842}; |
244 | | |
245 | | // Used to fetch a naturally-aligned 32-bit word in little endian byte-order |
246 | | static inline uint32_t LE_LOAD32(const uint8_t* p) { |
247 | | return DecodeFixed32(reinterpret_cast<const char*>(p)); |
248 | | } |
249 | | #endif // !__SSE4_2__ |
250 | | |
251 | 0 | static inline void DefaultCRC32(uint64_t* l, uint8_t const** p) { |
252 | | #ifndef __SSE4_2__ |
253 | | uint32_t c = static_cast<uint32_t>(*l ^ LE_LOAD32(*p)); |
254 | | *p += 4; |
255 | | *l = table3_[c & 0xff] ^ table2_[(c >> 8) & 0xff] ^ |
256 | | table1_[(c >> 16) & 0xff] ^ table0_[c >> 24]; |
257 | | // DO it twice. |
258 | | c = static_cast<uint32_t>(*l ^ LE_LOAD32(*p)); |
259 | | *p += 4; |
260 | | *l = table3_[c & 0xff] ^ table2_[(c >> 8) & 0xff] ^ |
261 | | table1_[(c >> 16) & 0xff] ^ table0_[c >> 24]; |
262 | | #elif defined(__LP64__) || defined(_WIN64) |
263 | | *l = _mm_crc32_u64(*l, DecodeFixed64(reinterpret_cast<const char*>(*p))); |
264 | 0 | *p += 8; |
265 | | #else |
266 | | *l = _mm_crc32_u32(static_cast<unsigned int>(*l), LE_LOAD32(*p)); |
267 | | *p += 4; |
268 | | *l = _mm_crc32_u32(static_cast<unsigned int>(*l), LE_LOAD32(*p)); |
269 | | *p += 4; |
270 | | #endif |
271 | 0 | } |
272 | | |
273 | | template <void (*CRC32)(uint64_t*, uint8_t const**)> |
274 | 0 | uint32_t ExtendImpl(uint32_t crc, const char* buf, size_t size) { |
275 | 0 | const uint8_t* p = reinterpret_cast<const uint8_t*>(buf); |
276 | 0 | const uint8_t* e = p + size; |
277 | 0 | uint64_t l = crc ^ 0xffffffffu; |
278 | | |
279 | | // Align n to (1 << m) byte boundary |
280 | 0 | #define ALIGN(n, m) ((n + ((1 << m) - 1)) & ~((1 << m) - 1)) |
281 | |
|
282 | 0 | #define STEP1 \ |
283 | 0 | do { \ |
284 | 0 | int c = (l & 0xff) ^ *p++; \ |
285 | 0 | l = table0_[c] ^ (l >> 8); \ |
286 | 0 | } while (0) |
287 | | |
288 | | // Point x at first 16-byte aligned byte in string. This might be |
289 | | // just past the end of the string. |
290 | 0 | const uintptr_t pval = reinterpret_cast<uintptr_t>(p); |
291 | 0 | const uint8_t* x = reinterpret_cast<const uint8_t*>(ALIGN(pval, 4)); |
292 | 0 | if (x <= e) { |
293 | | // Process bytes until finished or p is 16-byte aligned |
294 | 0 | while (p != x) { |
295 | 0 | STEP1; |
296 | 0 | } |
297 | 0 | } |
298 | | // Process bytes 16 at a time |
299 | 0 | while ((e - p) >= 16) { |
300 | 0 | CRC32(&l, &p); |
301 | 0 | CRC32(&l, &p); |
302 | 0 | } |
303 | | // Process bytes 8 at a time |
304 | 0 | while ((e - p) >= 8) { |
305 | 0 | CRC32(&l, &p); |
306 | 0 | } |
307 | | // Process the last few bytes |
308 | 0 | while (p != e) { |
309 | 0 | STEP1; |
310 | 0 | } |
311 | 0 | #undef STEP1 |
312 | 0 | #undef ALIGN |
313 | 0 | return static_cast<uint32_t>(l ^ 0xffffffffu); |
314 | 0 | } |
315 | | |
316 | | using Function = uint32_t (*)(uint32_t, const char*, size_t); |
317 | | |
318 | | #if defined(HAVE_POWER8) && defined(HAS_ALTIVEC) |
319 | | uint32_t ExtendPPCImpl(uint32_t crc, const char* buf, size_t size) { |
320 | | return crc32c_ppc(crc, (const unsigned char*)buf, size); |
321 | | } |
322 | | |
323 | | #if __linux__ |
324 | | static int arch_ppc_probe(void) { |
325 | | arch_ppc_crc32 = 0; |
326 | | |
327 | | #if defined(__powerpc64__) && defined(ROCKSDB_AUXV_GETAUXVAL_PRESENT) |
328 | | if (getauxval(AT_HWCAP2) & PPC_FEATURE2_VEC_CRYPTO) arch_ppc_crc32 = 1; |
329 | | #endif /* __powerpc64__ */ |
330 | | |
331 | | return arch_ppc_crc32; |
332 | | } |
333 | | #elif __FreeBSD__ |
334 | | static int arch_ppc_probe(void) { |
335 | | unsigned long cpufeatures; |
336 | | arch_ppc_crc32 = 0; |
337 | | |
338 | | #if defined(__powerpc64__) |
339 | | elf_aux_info(AT_HWCAP2, &cpufeatures, sizeof(cpufeatures)); |
340 | | if (cpufeatures & PPC_FEATURE2_HAS_VEC_CRYPTO) arch_ppc_crc32 = 1; |
341 | | #endif /* __powerpc64__ */ |
342 | | |
343 | | return arch_ppc_crc32; |
344 | | } |
345 | | #endif // __linux__ |
346 | | |
347 | | static bool isAltiVec() { |
348 | | if (arch_ppc_probe()) { |
349 | | return true; |
350 | | } else { |
351 | | return false; |
352 | | } |
353 | | } |
354 | | #endif |
355 | | |
356 | | #if defined(HAVE_ARM64_CRC) |
357 | | uint32_t ExtendARMImpl(uint32_t crc, const char* buf, size_t size) { |
358 | | return crc32c_arm64(crc, (const unsigned char*)buf, size); |
359 | | } |
360 | | #endif |
361 | | |
362 | 9.02k | std::string IsFastCrc32Supported() { |
363 | 9.02k | bool has_fast_crc = false; |
364 | 9.02k | std::string fast_zero_msg; |
365 | 9.02k | std::string arch; |
366 | | #ifdef HAVE_POWER8 |
367 | | #ifdef HAS_ALTIVEC |
368 | | if (arch_ppc_probe()) { |
369 | | has_fast_crc = true; |
370 | | arch = "PPC"; |
371 | | } |
372 | | #else |
373 | | has_fast_crc = false; |
374 | | arch = "PPC"; |
375 | | #endif |
376 | | #elif defined(HAVE_ARM64_CRC) |
377 | | if (crc32c_runtime_check()) { |
378 | | has_fast_crc = true; |
379 | | arch = "Arm64"; |
380 | | pmull_runtime_flag = crc32c_pmull_runtime_check(); |
381 | | } else { |
382 | | has_fast_crc = false; |
383 | | arch = "Arm64"; |
384 | | } |
385 | | #else |
386 | 9.02k | #ifdef __SSE4_2__ |
387 | 9.02k | has_fast_crc = true; |
388 | 9.02k | #endif // __SSE4_2__ |
389 | 9.02k | arch = "x86"; |
390 | 9.02k | #endif |
391 | 9.02k | if (has_fast_crc) { |
392 | 9.02k | fast_zero_msg.append("Supported on " + arch); |
393 | 9.02k | } else { |
394 | 0 | fast_zero_msg.append("Not supported on " + arch); |
395 | 0 | } |
396 | 9.02k | return fast_zero_msg; |
397 | 9.02k | } |
398 | | |
399 | | /* |
400 | | * Copyright 2016 Ferry Toth, Exalon Delft BV, The Netherlands |
401 | | * This software is provided 'as-is', without any express or implied |
402 | | * warranty. In no event will the author be held liable for any damages |
403 | | * arising from the use of this software. |
404 | | * Permission is granted to anyone to use this software for any purpose, |
405 | | * including commercial applications, and to alter it and redistribute it |
406 | | * freely, subject to the following restrictions: |
407 | | * 1. The origin of this software must not be misrepresented; you must not |
408 | | * claim that you wrote the original software. If you use this software |
409 | | * in a product, an acknowledgment in the product documentation would be |
410 | | * appreciated but is not required. |
411 | | * 2. Altered source versions must be plainly marked as such, and must not be |
412 | | * misrepresented as being the original software. |
413 | | * 3. This notice may not be removed or altered from any source distribution. |
414 | | * Ferry Toth |
415 | | * ftoth@exalondelft.nl |
416 | | * |
417 | | * https://github.com/htot/crc32c |
418 | | * |
419 | | * Modified by Facebook |
420 | | * |
421 | | * Original intel whitepaper: |
422 | | * "Fast CRC Computation for iSCSI Polynomial Using CRC32 Instruction" |
423 | | * https://www.intel.com/content/dam/www/public/us/en/documents/white-papers/crc-iscsi-polynomial-crc32-instruction-paper.pdf |
424 | | * |
425 | | * This version is from the folly library, created by Dave Watson |
426 | | * <davejwatson@fb.com> |
427 | | * |
428 | | */ |
429 | | #if defined(__SSE4_2__) && defined(__PCLMUL__) |
430 | | |
431 | | #define CRCtriplet(crc, buf, offset) \ |
432 | 27.4M | crc##0 = _mm_crc32_u64(crc##0, *(buf##0 + offset)); \ |
433 | 27.4M | crc##1 = _mm_crc32_u64(crc##1, *(buf##1 + offset)); \ |
434 | 27.4M | crc##2 = _mm_crc32_u64(crc##2, *(buf##2 + offset)); |
435 | | |
436 | | #define CRCduplet(crc, buf, offset) \ |
437 | 412k | crc##0 = _mm_crc32_u64(crc##0, *(buf##0 + offset)); \ |
438 | 412k | crc##1 = _mm_crc32_u64(crc##1, *(buf##1 + offset)); |
439 | | |
440 | | #define CRCsinglet(crc, buf, offset) \ |
441 | 10.3M | crc = _mm_crc32_u64(crc, *(uint64_t*)(buf + offset)); |
442 | | |
443 | | // Numbers taken directly from intel whitepaper. |
444 | | // clang-format off |
445 | | const uint64_t clmul_constants[] = { |
446 | | 0x14cd00bd6, 0x105ec76f0, 0x0ba4fc28e, 0x14cd00bd6, |
447 | | 0x1d82c63da, 0x0f20c0dfe, 0x09e4addf8, 0x0ba4fc28e, |
448 | | 0x039d3b296, 0x1384aa63a, 0x102f9b8a2, 0x1d82c63da, |
449 | | 0x14237f5e6, 0x01c291d04, 0x00d3b6092, 0x09e4addf8, |
450 | | 0x0c96cfdc0, 0x0740eef02, 0x18266e456, 0x039d3b296, |
451 | | 0x0daece73e, 0x0083a6eec, 0x0ab7aff2a, 0x102f9b8a2, |
452 | | 0x1248ea574, 0x1c1733996, 0x083348832, 0x14237f5e6, |
453 | | 0x12c743124, 0x02ad91c30, 0x0b9e02b86, 0x00d3b6092, |
454 | | 0x018b33a4e, 0x06992cea2, 0x1b331e26a, 0x0c96cfdc0, |
455 | | 0x17d35ba46, 0x07e908048, 0x1bf2e8b8a, 0x18266e456, |
456 | | 0x1a3e0968a, 0x11ed1f9d8, 0x0ce7f39f4, 0x0daece73e, |
457 | | 0x061d82e56, 0x0f1d0f55e, 0x0d270f1a2, 0x0ab7aff2a, |
458 | | 0x1c3f5f66c, 0x0a87ab8a8, 0x12ed0daac, 0x1248ea574, |
459 | | 0x065863b64, 0x08462d800, 0x11eef4f8e, 0x083348832, |
460 | | 0x1ee54f54c, 0x071d111a8, 0x0b3e32c28, 0x12c743124, |
461 | | 0x0064f7f26, 0x0ffd852c6, 0x0dd7e3b0c, 0x0b9e02b86, |
462 | | 0x0f285651c, 0x0dcb17aa4, 0x010746f3c, 0x018b33a4e, |
463 | | 0x1c24afea4, 0x0f37c5aee, 0x0271d9844, 0x1b331e26a, |
464 | | 0x08e766a0c, 0x06051d5a2, 0x093a5f730, 0x17d35ba46, |
465 | | 0x06cb08e5c, 0x11d5ca20e, 0x06b749fb2, 0x1bf2e8b8a, |
466 | | 0x1167f94f2, 0x021f3d99c, 0x0cec3662e, 0x1a3e0968a, |
467 | | 0x19329634a, 0x08f158014, 0x0e6fc4e6a, 0x0ce7f39f4, |
468 | | 0x08227bb8a, 0x1a5e82106, 0x0b0cd4768, 0x061d82e56, |
469 | | 0x13c2b89c4, 0x188815ab2, 0x0d7a4825c, 0x0d270f1a2, |
470 | | 0x10f5ff2ba, 0x105405f3e, 0x00167d312, 0x1c3f5f66c, |
471 | | 0x0f6076544, 0x0e9adf796, 0x026f6a60a, 0x12ed0daac, |
472 | | 0x1a2adb74e, 0x096638b34, 0x19d34af3a, 0x065863b64, |
473 | | 0x049c3cc9c, 0x1e50585a0, 0x068bce87a, 0x11eef4f8e, |
474 | | 0x1524fa6c6, 0x19f1c69dc, 0x16cba8aca, 0x1ee54f54c, |
475 | | 0x042d98888, 0x12913343e, 0x1329d9f7e, 0x0b3e32c28, |
476 | | 0x1b1c69528, 0x088f25a3a, 0x02178513a, 0x0064f7f26, |
477 | | 0x0e0ac139e, 0x04e36f0b0, 0x0170076fa, 0x0dd7e3b0c, |
478 | | 0x141a1a2e2, 0x0bd6f81f8, 0x16ad828b4, 0x0f285651c, |
479 | | 0x041d17b64, 0x19425cbba, 0x1fae1cc66, 0x010746f3c, |
480 | | 0x1a75b4b00, 0x18db37e8a, 0x0f872e54c, 0x1c24afea4, |
481 | | 0x01e41e9fc, 0x04c144932, 0x086d8e4d2, 0x0271d9844, |
482 | | 0x160f7af7a, 0x052148f02, 0x05bb8f1bc, 0x08e766a0c, |
483 | | 0x0a90fd27a, 0x0a3c6f37a, 0x0b3af077a, 0x093a5f730, |
484 | | 0x04984d782, 0x1d22c238e, 0x0ca6ef3ac, 0x06cb08e5c, |
485 | | 0x0234e0b26, 0x063ded06a, 0x1d88abd4a, 0x06b749fb2, |
486 | | 0x04597456a, 0x04d56973c, 0x0e9e28eb4, 0x1167f94f2, |
487 | | 0x07b3ff57a, 0x19385bf2e, 0x0c9c8b782, 0x0cec3662e, |
488 | | 0x13a9cba9e, 0x0e417f38a, 0x093e106a4, 0x19329634a, |
489 | | 0x167001a9c, 0x14e727980, 0x1ddffc5d4, 0x0e6fc4e6a, |
490 | | 0x00df04680, 0x0d104b8fc, 0x02342001e, 0x08227bb8a, |
491 | | 0x00a2a8d7e, 0x05b397730, 0x168763fa6, 0x0b0cd4768, |
492 | | 0x1ed5a407a, 0x0e78eb416, 0x0d2c3ed1a, 0x13c2b89c4, |
493 | | 0x0995a5724, 0x1641378f0, 0x19b1afbc4, 0x0d7a4825c, |
494 | | 0x109ffedc0, 0x08d96551c, 0x0f2271e60, 0x10f5ff2ba, |
495 | | 0x00b0bf8ca, 0x00bf80dd2, 0x123888b7a, 0x00167d312, |
496 | | 0x1e888f7dc, 0x18dcddd1c, 0x002ee03b2, 0x0f6076544, |
497 | | 0x183e8d8fe, 0x06a45d2b2, 0x133d7a042, 0x026f6a60a, |
498 | | 0x116b0f50c, 0x1dd3e10e8, 0x05fabe670, 0x1a2adb74e, |
499 | | 0x130004488, 0x0de87806c, 0x000bcf5f6, 0x19d34af3a, |
500 | | 0x18f0c7078, 0x014338754, 0x017f27698, 0x049c3cc9c, |
501 | | 0x058ca5f00, 0x15e3e77ee, 0x1af900c24, 0x068bce87a, |
502 | | 0x0b5cfca28, 0x0dd07448e, 0x0ded288f8, 0x1524fa6c6, |
503 | | 0x059f229bc, 0x1d8048348, 0x06d390dec, 0x16cba8aca, |
504 | | 0x037170390, 0x0a3e3e02c, 0x06353c1cc, 0x042d98888, |
505 | | 0x0c4584f5c, 0x0d73c7bea, 0x1f16a3418, 0x1329d9f7e, |
506 | | 0x0531377e2, 0x185137662, 0x1d8d9ca7c, 0x1b1c69528, |
507 | | 0x0b25b29f2, 0x18a08b5bc, 0x19fb2a8b0, 0x02178513a, |
508 | | 0x1a08fe6ac, 0x1da758ae0, 0x045cddf4e, 0x0e0ac139e, |
509 | | 0x1a91647f2, 0x169cf9eb0, 0x1a0f717c4, 0x0170076fa, |
510 | | }; |
511 | | |
512 | | // Compute the crc32c value for buffer smaller than 8 |
513 | | #ifdef ROCKSDB_UBSAN_RUN |
514 | | #if defined(__clang__) |
515 | | __attribute__((__no_sanitize__("alignment"))) |
516 | | #elif defined(__GNUC__) |
517 | | __attribute__((__no_sanitize_undefined__)) |
518 | | #endif |
519 | | #endif |
520 | | inline void align_to_8( |
521 | | size_t len, |
522 | | uint64_t& crc0, // crc so far, updated on return |
523 | 4.42M | const unsigned char*& next) { // next data pointer, updated on return |
524 | 4.42M | uint32_t crc32bit = static_cast<uint32_t>(crc0); |
525 | 4.42M | if (len & 0x04) { |
526 | 1.39M | crc32bit = _mm_crc32_u32(crc32bit, *(uint32_t*)next); |
527 | 1.39M | next += sizeof(uint32_t); |
528 | 1.39M | } |
529 | 4.42M | if (len & 0x02) { |
530 | 2.31M | crc32bit = _mm_crc32_u16(crc32bit, *(uint16_t*)next); |
531 | 2.31M | next += sizeof(uint16_t); |
532 | 2.31M | } |
533 | 4.42M | if (len & 0x01) { |
534 | 2.54M | crc32bit = _mm_crc32_u8(crc32bit, *(next)); |
535 | 2.54M | next++; |
536 | 2.54M | } |
537 | 4.42M | crc0 = crc32bit; |
538 | 4.42M | } |
539 | | |
540 | | // |
541 | | // CombineCRC performs pclmulqdq multiplication of 2 partial CRC's and a well |
542 | | // chosen constant and xor's these with the remaining CRC. |
543 | | // |
544 | | inline uint64_t CombineCRC( |
545 | | size_t block_size, |
546 | | uint64_t crc0, |
547 | | uint64_t crc1, |
548 | | uint64_t crc2, |
549 | 412k | const uint64_t* next2) { |
550 | 412k | const auto multiplier = |
551 | 412k | *(reinterpret_cast<const __m128i*>(clmul_constants) + block_size - 1); |
552 | 412k | const auto crc0_xmm = _mm_set_epi64x(0, crc0); |
553 | 412k | const auto res0 = _mm_clmulepi64_si128(crc0_xmm, multiplier, 0x00); |
554 | 412k | const auto crc1_xmm = _mm_set_epi64x(0, crc1); |
555 | 412k | const auto res1 = _mm_clmulepi64_si128(crc1_xmm, multiplier, 0x10); |
556 | 412k | const auto res = _mm_xor_si128(res0, res1); |
557 | 412k | crc0 = _mm_cvtsi128_si64(res); |
558 | 412k | crc0 = crc0 ^ *((uint64_t*)next2 - 1); |
559 | 412k | crc2 = _mm_crc32_u64(crc2, crc0); |
560 | 412k | return crc2; |
561 | 412k | } |
562 | | |
563 | | // Compute CRC-32C using the Intel hardware instruction. |
564 | | #ifdef ROCKSDB_UBSAN_RUN |
565 | | #if defined(__clang__) |
566 | | __attribute__((__no_sanitize__("alignment"))) |
567 | | #elif defined(__GNUC__) |
568 | | __attribute__((__no_sanitize_undefined__)) |
569 | | #endif |
570 | | #endif |
571 | 4.16M | uint32_t crc32c_3way(uint32_t crc, const char* buf, size_t len) { |
572 | 4.16M | const unsigned char* next = (const unsigned char*)buf; |
573 | 4.16M | uint64_t count; |
574 | 4.16M | uint64_t crc0, crc1, crc2; |
575 | 4.16M | crc0 = crc ^ 0xffffffffu; |
576 | | |
577 | | |
578 | 4.16M | if (len >= 8) { |
579 | | // if len > 216 then align and use triplets |
580 | 2.76M | if (len > 216) { |
581 | 258k | { |
582 | | // Work on the bytes (< 8) before the first 8-byte alignment addr starts |
583 | 258k | uint64_t align_bytes = (8 - (uintptr_t)next) & 7; |
584 | 258k | len -= align_bytes; |
585 | 258k | align_to_8(align_bytes, crc0, next); |
586 | 258k | } |
587 | | |
588 | | // Now work on the remaining blocks |
589 | 258k | count = len / 24; // number of triplets |
590 | 258k | len %= 24; // bytes remaining |
591 | 258k | uint64_t n = count >> 7; // #blocks = first block + full blocks |
592 | 258k | uint64_t block_size = count & 127; |
593 | 258k | if (block_size == 0) { |
594 | 249 | block_size = 128; |
595 | 258k | } else { |
596 | 258k | n++; |
597 | 258k | } |
598 | | // points to the first byte of the next block |
599 | 258k | const uint64_t* next0 = (uint64_t*)next + block_size; |
600 | 258k | const uint64_t* next1 = next0 + block_size; |
601 | 258k | const uint64_t* next2 = next1 + block_size; |
602 | | |
603 | 258k | crc1 = crc2 = 0; |
604 | | // Use Duff's device, a for() loop inside a switch() |
605 | | // statement. This needs to execute at least once, round len |
606 | | // down to nearest triplet multiple |
607 | 258k | switch (block_size) { |
608 | 249 | case 128: |
609 | 153k | do { |
610 | | // jumps here for a full block of len 128 |
611 | 153k | CRCtriplet(crc, next, -128); |
612 | 153k | FALLTHROUGH_INTENDED; |
613 | 153k | case 127: |
614 | | // jumps here or below for the first block smaller |
615 | 153k | CRCtriplet(crc, next, -127); |
616 | 153k | FALLTHROUGH_INTENDED; |
617 | 154k | case 126: |
618 | 154k | CRCtriplet(crc, next, -126); // than 128 |
619 | 154k | FALLTHROUGH_INTENDED; |
620 | 154k | case 125: |
621 | 154k | CRCtriplet(crc, next, -125); |
622 | 154k | FALLTHROUGH_INTENDED; |
623 | 155k | case 124: |
624 | 155k | CRCtriplet(crc, next, -124); |
625 | 155k | FALLTHROUGH_INTENDED; |
626 | 155k | case 123: |
627 | 155k | CRCtriplet(crc, next, -123); |
628 | 155k | FALLTHROUGH_INTENDED; |
629 | 155k | case 122: |
630 | 155k | CRCtriplet(crc, next, -122); |
631 | 155k | FALLTHROUGH_INTENDED; |
632 | 156k | case 121: |
633 | 156k | CRCtriplet(crc, next, -121); |
634 | 156k | FALLTHROUGH_INTENDED; |
635 | 156k | case 120: |
636 | 156k | CRCtriplet(crc, next, -120); |
637 | 156k | FALLTHROUGH_INTENDED; |
638 | 156k | case 119: |
639 | 156k | CRCtriplet(crc, next, -119); |
640 | 156k | FALLTHROUGH_INTENDED; |
641 | 157k | case 118: |
642 | 157k | CRCtriplet(crc, next, -118); |
643 | 157k | FALLTHROUGH_INTENDED; |
644 | 157k | case 117: |
645 | 157k | CRCtriplet(crc, next, -117); |
646 | 157k | FALLTHROUGH_INTENDED; |
647 | 157k | case 116: |
648 | 157k | CRCtriplet(crc, next, -116); |
649 | 157k | FALLTHROUGH_INTENDED; |
650 | 158k | case 115: |
651 | 158k | CRCtriplet(crc, next, -115); |
652 | 158k | FALLTHROUGH_INTENDED; |
653 | 159k | case 114: |
654 | 159k | CRCtriplet(crc, next, -114); |
655 | 159k | FALLTHROUGH_INTENDED; |
656 | 160k | case 113: |
657 | 160k | CRCtriplet(crc, next, -113); |
658 | 160k | FALLTHROUGH_INTENDED; |
659 | 160k | case 112: |
660 | 160k | CRCtriplet(crc, next, -112); |
661 | 160k | FALLTHROUGH_INTENDED; |
662 | 161k | case 111: |
663 | 161k | CRCtriplet(crc, next, -111); |
664 | 161k | FALLTHROUGH_INTENDED; |
665 | 162k | case 110: |
666 | 162k | CRCtriplet(crc, next, -110); |
667 | 162k | FALLTHROUGH_INTENDED; |
668 | 162k | case 109: |
669 | 162k | CRCtriplet(crc, next, -109); |
670 | 162k | FALLTHROUGH_INTENDED; |
671 | 163k | case 108: |
672 | 163k | CRCtriplet(crc, next, -108); |
673 | 163k | FALLTHROUGH_INTENDED; |
674 | 163k | case 107: |
675 | 163k | CRCtriplet(crc, next, -107); |
676 | 163k | FALLTHROUGH_INTENDED; |
677 | 163k | case 106: |
678 | 163k | CRCtriplet(crc, next, -106); |
679 | 163k | FALLTHROUGH_INTENDED; |
680 | 163k | case 105: |
681 | 163k | CRCtriplet(crc, next, -105); |
682 | 163k | FALLTHROUGH_INTENDED; |
683 | 164k | case 104: |
684 | 164k | CRCtriplet(crc, next, -104); |
685 | 164k | FALLTHROUGH_INTENDED; |
686 | 164k | case 103: |
687 | 164k | CRCtriplet(crc, next, -103); |
688 | 164k | FALLTHROUGH_INTENDED; |
689 | 165k | case 102: |
690 | 165k | CRCtriplet(crc, next, -102); |
691 | 165k | FALLTHROUGH_INTENDED; |
692 | 165k | case 101: |
693 | 165k | CRCtriplet(crc, next, -101); |
694 | 165k | FALLTHROUGH_INTENDED; |
695 | 165k | case 100: |
696 | 165k | CRCtriplet(crc, next, -100); |
697 | 165k | FALLTHROUGH_INTENDED; |
698 | 166k | case 99: |
699 | 166k | CRCtriplet(crc, next, -99); |
700 | 166k | FALLTHROUGH_INTENDED; |
701 | 166k | case 98: |
702 | 166k | CRCtriplet(crc, next, -98); |
703 | 166k | FALLTHROUGH_INTENDED; |
704 | 168k | case 97: |
705 | 168k | CRCtriplet(crc, next, -97); |
706 | 168k | FALLTHROUGH_INTENDED; |
707 | 168k | case 96: |
708 | 168k | CRCtriplet(crc, next, -96); |
709 | 168k | FALLTHROUGH_INTENDED; |
710 | 168k | case 95: |
711 | 168k | CRCtriplet(crc, next, -95); |
712 | 168k | FALLTHROUGH_INTENDED; |
713 | 169k | case 94: |
714 | 169k | CRCtriplet(crc, next, -94); |
715 | 169k | FALLTHROUGH_INTENDED; |
716 | 171k | case 93: |
717 | 171k | CRCtriplet(crc, next, -93); |
718 | 171k | FALLTHROUGH_INTENDED; |
719 | 172k | case 92: |
720 | 172k | CRCtriplet(crc, next, -92); |
721 | 172k | FALLTHROUGH_INTENDED; |
722 | 173k | case 91: |
723 | 173k | CRCtriplet(crc, next, -91); |
724 | 173k | FALLTHROUGH_INTENDED; |
725 | 174k | case 90: |
726 | 174k | CRCtriplet(crc, next, -90); |
727 | 174k | FALLTHROUGH_INTENDED; |
728 | 175k | case 89: |
729 | 175k | CRCtriplet(crc, next, -89); |
730 | 175k | FALLTHROUGH_INTENDED; |
731 | 175k | case 88: |
732 | 175k | CRCtriplet(crc, next, -88); |
733 | 175k | FALLTHROUGH_INTENDED; |
734 | 176k | case 87: |
735 | 176k | CRCtriplet(crc, next, -87); |
736 | 176k | FALLTHROUGH_INTENDED; |
737 | 176k | case 86: |
738 | 176k | CRCtriplet(crc, next, -86); |
739 | 176k | FALLTHROUGH_INTENDED; |
740 | 181k | case 85: |
741 | 181k | CRCtriplet(crc, next, -85); |
742 | 181k | FALLTHROUGH_INTENDED; |
743 | 185k | case 84: |
744 | 185k | CRCtriplet(crc, next, -84); |
745 | 185k | FALLTHROUGH_INTENDED; |
746 | 186k | case 83: |
747 | 186k | CRCtriplet(crc, next, -83); |
748 | 186k | FALLTHROUGH_INTENDED; |
749 | 187k | case 82: |
750 | 187k | CRCtriplet(crc, next, -82); |
751 | 187k | FALLTHROUGH_INTENDED; |
752 | 188k | case 81: |
753 | 188k | CRCtriplet(crc, next, -81); |
754 | 188k | FALLTHROUGH_INTENDED; |
755 | 189k | case 80: |
756 | 189k | CRCtriplet(crc, next, -80); |
757 | 189k | FALLTHROUGH_INTENDED; |
758 | 190k | case 79: |
759 | 190k | CRCtriplet(crc, next, -79); |
760 | 190k | FALLTHROUGH_INTENDED; |
761 | 191k | case 78: |
762 | 191k | CRCtriplet(crc, next, -78); |
763 | 191k | FALLTHROUGH_INTENDED; |
764 | 191k | case 77: |
765 | 191k | CRCtriplet(crc, next, -77); |
766 | 191k | FALLTHROUGH_INTENDED; |
767 | 192k | case 76: |
768 | 192k | CRCtriplet(crc, next, -76); |
769 | 192k | FALLTHROUGH_INTENDED; |
770 | 192k | case 75: |
771 | 192k | CRCtriplet(crc, next, -75); |
772 | 192k | FALLTHROUGH_INTENDED; |
773 | 192k | case 74: |
774 | 192k | CRCtriplet(crc, next, -74); |
775 | 192k | FALLTHROUGH_INTENDED; |
776 | 193k | case 73: |
777 | 193k | CRCtriplet(crc, next, -73); |
778 | 193k | FALLTHROUGH_INTENDED; |
779 | 193k | case 72: |
780 | 193k | CRCtriplet(crc, next, -72); |
781 | 193k | FALLTHROUGH_INTENDED; |
782 | 194k | case 71: |
783 | 194k | CRCtriplet(crc, next, -71); |
784 | 194k | FALLTHROUGH_INTENDED; |
785 | 195k | case 70: |
786 | 195k | CRCtriplet(crc, next, -70); |
787 | 195k | FALLTHROUGH_INTENDED; |
788 | 195k | case 69: |
789 | 195k | CRCtriplet(crc, next, -69); |
790 | 195k | FALLTHROUGH_INTENDED; |
791 | 196k | case 68: |
792 | 196k | CRCtriplet(crc, next, -68); |
793 | 196k | FALLTHROUGH_INTENDED; |
794 | 197k | case 67: |
795 | 197k | CRCtriplet(crc, next, -67); |
796 | 197k | FALLTHROUGH_INTENDED; |
797 | 197k | case 66: |
798 | 197k | CRCtriplet(crc, next, -66); |
799 | 197k | FALLTHROUGH_INTENDED; |
800 | 198k | case 65: |
801 | 198k | CRCtriplet(crc, next, -65); |
802 | 198k | FALLTHROUGH_INTENDED; |
803 | 199k | case 64: |
804 | 199k | CRCtriplet(crc, next, -64); |
805 | 199k | FALLTHROUGH_INTENDED; |
806 | 200k | case 63: |
807 | 200k | CRCtriplet(crc, next, -63); |
808 | 200k | FALLTHROUGH_INTENDED; |
809 | 201k | case 62: |
810 | 201k | CRCtriplet(crc, next, -62); |
811 | 201k | FALLTHROUGH_INTENDED; |
812 | 202k | case 61: |
813 | 202k | CRCtriplet(crc, next, -61); |
814 | 202k | FALLTHROUGH_INTENDED; |
815 | 203k | case 60: |
816 | 203k | CRCtriplet(crc, next, -60); |
817 | 203k | FALLTHROUGH_INTENDED; |
818 | 203k | case 59: |
819 | 203k | CRCtriplet(crc, next, -59); |
820 | 203k | FALLTHROUGH_INTENDED; |
821 | 204k | case 58: |
822 | 204k | CRCtriplet(crc, next, -58); |
823 | 204k | FALLTHROUGH_INTENDED; |
824 | 207k | case 57: |
825 | 207k | CRCtriplet(crc, next, -57); |
826 | 207k | FALLTHROUGH_INTENDED; |
827 | 208k | case 56: |
828 | 208k | CRCtriplet(crc, next, -56); |
829 | 208k | FALLTHROUGH_INTENDED; |
830 | 209k | case 55: |
831 | 209k | CRCtriplet(crc, next, -55); |
832 | 209k | FALLTHROUGH_INTENDED; |
833 | 209k | case 54: |
834 | 209k | CRCtriplet(crc, next, -54); |
835 | 209k | FALLTHROUGH_INTENDED; |
836 | 210k | case 53: |
837 | 210k | CRCtriplet(crc, next, -53); |
838 | 210k | FALLTHROUGH_INTENDED; |
839 | 211k | case 52: |
840 | 211k | CRCtriplet(crc, next, -52); |
841 | 211k | FALLTHROUGH_INTENDED; |
842 | 211k | case 51: |
843 | 211k | CRCtriplet(crc, next, -51); |
844 | 211k | FALLTHROUGH_INTENDED; |
845 | 215k | case 50: |
846 | 215k | CRCtriplet(crc, next, -50); |
847 | 215k | FALLTHROUGH_INTENDED; |
848 | 216k | case 49: |
849 | 216k | CRCtriplet(crc, next, -49); |
850 | 216k | FALLTHROUGH_INTENDED; |
851 | 218k | case 48: |
852 | 218k | CRCtriplet(crc, next, -48); |
853 | 218k | FALLTHROUGH_INTENDED; |
854 | 219k | case 47: |
855 | 219k | CRCtriplet(crc, next, -47); |
856 | 219k | FALLTHROUGH_INTENDED; |
857 | 220k | case 46: |
858 | 220k | CRCtriplet(crc, next, -46); |
859 | 220k | FALLTHROUGH_INTENDED; |
860 | 221k | case 45: |
861 | 221k | CRCtriplet(crc, next, -45); |
862 | 221k | FALLTHROUGH_INTENDED; |
863 | 224k | case 44: |
864 | 224k | CRCtriplet(crc, next, -44); |
865 | 224k | FALLTHROUGH_INTENDED; |
866 | 225k | case 43: |
867 | 225k | CRCtriplet(crc, next, -43); |
868 | 225k | FALLTHROUGH_INTENDED; |
869 | 226k | case 42: |
870 | 226k | CRCtriplet(crc, next, -42); |
871 | 226k | FALLTHROUGH_INTENDED; |
872 | 226k | case 41: |
873 | 226k | CRCtriplet(crc, next, -41); |
874 | 226k | FALLTHROUGH_INTENDED; |
875 | 228k | case 40: |
876 | 228k | CRCtriplet(crc, next, -40); |
877 | 228k | FALLTHROUGH_INTENDED; |
878 | 231k | case 39: |
879 | 231k | CRCtriplet(crc, next, -39); |
880 | 231k | FALLTHROUGH_INTENDED; |
881 | 232k | case 38: |
882 | 232k | CRCtriplet(crc, next, -38); |
883 | 232k | FALLTHROUGH_INTENDED; |
884 | 233k | case 37: |
885 | 233k | CRCtriplet(crc, next, -37); |
886 | 233k | FALLTHROUGH_INTENDED; |
887 | 235k | case 36: |
888 | 235k | CRCtriplet(crc, next, -36); |
889 | 235k | FALLTHROUGH_INTENDED; |
890 | 235k | case 35: |
891 | 235k | CRCtriplet(crc, next, -35); |
892 | 235k | FALLTHROUGH_INTENDED; |
893 | 237k | case 34: |
894 | 237k | CRCtriplet(crc, next, -34); |
895 | 237k | FALLTHROUGH_INTENDED; |
896 | 238k | case 33: |
897 | 238k | CRCtriplet(crc, next, -33); |
898 | 238k | FALLTHROUGH_INTENDED; |
899 | 239k | case 32: |
900 | 239k | CRCtriplet(crc, next, -32); |
901 | 239k | FALLTHROUGH_INTENDED; |
902 | 239k | case 31: |
903 | 239k | CRCtriplet(crc, next, -31); |
904 | 239k | FALLTHROUGH_INTENDED; |
905 | 241k | case 30: |
906 | 241k | CRCtriplet(crc, next, -30); |
907 | 241k | FALLTHROUGH_INTENDED; |
908 | 243k | case 29: |
909 | 243k | CRCtriplet(crc, next, -29); |
910 | 243k | FALLTHROUGH_INTENDED; |
911 | 244k | case 28: |
912 | 244k | CRCtriplet(crc, next, -28); |
913 | 244k | FALLTHROUGH_INTENDED; |
914 | 245k | case 27: |
915 | 245k | CRCtriplet(crc, next, -27); |
916 | 245k | FALLTHROUGH_INTENDED; |
917 | 246k | case 26: |
918 | 246k | CRCtriplet(crc, next, -26); |
919 | 246k | FALLTHROUGH_INTENDED; |
920 | 249k | case 25: |
921 | 249k | CRCtriplet(crc, next, -25); |
922 | 249k | FALLTHROUGH_INTENDED; |
923 | 250k | case 24: |
924 | 250k | CRCtriplet(crc, next, -24); |
925 | 250k | FALLTHROUGH_INTENDED; |
926 | 250k | case 23: |
927 | 250k | CRCtriplet(crc, next, -23); |
928 | 250k | FALLTHROUGH_INTENDED; |
929 | 252k | case 22: |
930 | 252k | CRCtriplet(crc, next, -22); |
931 | 252k | FALLTHROUGH_INTENDED; |
932 | 253k | case 21: |
933 | 253k | CRCtriplet(crc, next, -21); |
934 | 253k | FALLTHROUGH_INTENDED; |
935 | 256k | case 20: |
936 | 256k | CRCtriplet(crc, next, -20); |
937 | 256k | FALLTHROUGH_INTENDED; |
938 | 257k | case 19: |
939 | 257k | CRCtriplet(crc, next, -19); |
940 | 257k | FALLTHROUGH_INTENDED; |
941 | 262k | case 18: |
942 | 262k | CRCtriplet(crc, next, -18); |
943 | 262k | FALLTHROUGH_INTENDED; |
944 | 265k | case 17: |
945 | 265k | CRCtriplet(crc, next, -17); |
946 | 265k | FALLTHROUGH_INTENDED; |
947 | 266k | case 16: |
948 | 266k | CRCtriplet(crc, next, -16); |
949 | 266k | FALLTHROUGH_INTENDED; |
950 | 276k | case 15: |
951 | 276k | CRCtriplet(crc, next, -15); |
952 | 276k | FALLTHROUGH_INTENDED; |
953 | 285k | case 14: |
954 | 285k | CRCtriplet(crc, next, -14); |
955 | 285k | FALLTHROUGH_INTENDED; |
956 | 296k | case 13: |
957 | 296k | CRCtriplet(crc, next, -13); |
958 | 296k | FALLTHROUGH_INTENDED; |
959 | 299k | case 12: |
960 | 299k | CRCtriplet(crc, next, -12); |
961 | 299k | FALLTHROUGH_INTENDED; |
962 | 324k | case 11: |
963 | 324k | CRCtriplet(crc, next, -11); |
964 | 324k | FALLTHROUGH_INTENDED; |
965 | 373k | case 10: |
966 | 373k | CRCtriplet(crc, next, -10); |
967 | 373k | FALLTHROUGH_INTENDED; |
968 | 406k | case 9: |
969 | 406k | CRCtriplet(crc, next, -9); |
970 | 406k | FALLTHROUGH_INTENDED; |
971 | 408k | case 8: |
972 | 408k | CRCtriplet(crc, next, -8); |
973 | 408k | FALLTHROUGH_INTENDED; |
974 | 409k | case 7: |
975 | 409k | CRCtriplet(crc, next, -7); |
976 | 409k | FALLTHROUGH_INTENDED; |
977 | 410k | case 6: |
978 | 410k | CRCtriplet(crc, next, -6); |
979 | 410k | FALLTHROUGH_INTENDED; |
980 | 410k | case 5: |
981 | 410k | CRCtriplet(crc, next, -5); |
982 | 410k | FALLTHROUGH_INTENDED; |
983 | 411k | case 4: |
984 | 411k | CRCtriplet(crc, next, -4); |
985 | 411k | FALLTHROUGH_INTENDED; |
986 | 411k | case 3: |
987 | 411k | CRCtriplet(crc, next, -3); |
988 | 411k | FALLTHROUGH_INTENDED; |
989 | 411k | case 2: |
990 | 411k | CRCtriplet(crc, next, -2); |
991 | 411k | FALLTHROUGH_INTENDED; |
992 | 412k | case 1: |
993 | 412k | CRCduplet(crc, next, -1); // the final triplet is actually only 2 |
994 | | //{ CombineCRC(); } |
995 | 412k | crc0 = CombineCRC(block_size, crc0, crc1, crc2, next2); |
996 | 412k | if (--n > 0) { |
997 | 153k | crc1 = crc2 = 0; |
998 | 153k | block_size = 128; |
999 | | // points to the first byte of the next block |
1000 | 153k | next0 = next2 + 128; |
1001 | 153k | next1 = next0 + 128; // from here on all blocks are 128 long |
1002 | 153k | next2 = next1 + 128; |
1003 | 153k | } |
1004 | 412k | FALLTHROUGH_INTENDED; |
1005 | 412k | case 0:; |
1006 | 412k | } while (n > 0); |
1007 | 258k | } |
1008 | 258k | next = (const unsigned char*)next2; |
1009 | 258k | } |
1010 | 2.76M | uint64_t count2 = len >> 3; // 216 of less bytes is 27 or less singlets |
1011 | 2.76M | len = len & 7; |
1012 | 2.76M | next += (count2 * 8); |
1013 | 2.76M | switch (count2) { |
1014 | 5.81k | case 27: |
1015 | 5.81k | CRCsinglet(crc0, next, -27 * 8); |
1016 | 5.81k | FALLTHROUGH_INTENDED; |
1017 | 31.4k | case 26: |
1018 | 31.4k | CRCsinglet(crc0, next, -26 * 8); |
1019 | 31.4k | FALLTHROUGH_INTENDED; |
1020 | 50.2k | case 25: |
1021 | 50.2k | CRCsinglet(crc0, next, -25 * 8); |
1022 | 50.2k | FALLTHROUGH_INTENDED; |
1023 | 59.2k | case 24: |
1024 | 59.2k | CRCsinglet(crc0, next, -24 * 8); |
1025 | 59.2k | FALLTHROUGH_INTENDED; |
1026 | 71.0k | case 23: |
1027 | 71.0k | CRCsinglet(crc0, next, -23 * 8); |
1028 | 71.0k | FALLTHROUGH_INTENDED; |
1029 | 89.5k | case 22: |
1030 | 89.5k | CRCsinglet(crc0, next, -22 * 8); |
1031 | 89.5k | FALLTHROUGH_INTENDED; |
1032 | 107k | case 21: |
1033 | 107k | CRCsinglet(crc0, next, -21 * 8); |
1034 | 107k | FALLTHROUGH_INTENDED; |
1035 | 112k | case 20: |
1036 | 112k | CRCsinglet(crc0, next, -20 * 8); |
1037 | 112k | FALLTHROUGH_INTENDED; |
1038 | 135k | case 19: |
1039 | 135k | CRCsinglet(crc0, next, -19 * 8); |
1040 | 135k | FALLTHROUGH_INTENDED; |
1041 | 144k | case 18: |
1042 | 144k | CRCsinglet(crc0, next, -18 * 8); |
1043 | 144k | FALLTHROUGH_INTENDED; |
1044 | 156k | case 17: |
1045 | 156k | CRCsinglet(crc0, next, -17 * 8); |
1046 | 156k | FALLTHROUGH_INTENDED; |
1047 | 170k | case 16: |
1048 | 170k | CRCsinglet(crc0, next, -16 * 8); |
1049 | 170k | FALLTHROUGH_INTENDED; |
1050 | 185k | case 15: |
1051 | 185k | CRCsinglet(crc0, next, -15 * 8); |
1052 | 185k | FALLTHROUGH_INTENDED; |
1053 | 202k | case 14: |
1054 | 202k | CRCsinglet(crc0, next, -14 * 8); |
1055 | 202k | FALLTHROUGH_INTENDED; |
1056 | 214k | case 13: |
1057 | 214k | CRCsinglet(crc0, next, -13 * 8); |
1058 | 214k | FALLTHROUGH_INTENDED; |
1059 | 228k | case 12: |
1060 | 228k | CRCsinglet(crc0, next, -12 * 8); |
1061 | 228k | FALLTHROUGH_INTENDED; |
1062 | 250k | case 11: |
1063 | 250k | CRCsinglet(crc0, next, -11 * 8); |
1064 | 250k | FALLTHROUGH_INTENDED; |
1065 | 278k | case 10: |
1066 | 278k | CRCsinglet(crc0, next, -10 * 8); |
1067 | 278k | FALLTHROUGH_INTENDED; |
1068 | 311k | case 9: |
1069 | 311k | CRCsinglet(crc0, next, -9 * 8); |
1070 | 311k | FALLTHROUGH_INTENDED; |
1071 | 328k | case 8: |
1072 | 328k | CRCsinglet(crc0, next, -8 * 8); |
1073 | 328k | FALLTHROUGH_INTENDED; |
1074 | 369k | case 7: |
1075 | 369k | CRCsinglet(crc0, next, -7 * 8); |
1076 | 369k | FALLTHROUGH_INTENDED; |
1077 | 408k | case 6: |
1078 | 408k | CRCsinglet(crc0, next, -6 * 8); |
1079 | 408k | FALLTHROUGH_INTENDED; |
1080 | 496k | case 5: |
1081 | 496k | CRCsinglet(crc0, next, -5 * 8); |
1082 | 496k | FALLTHROUGH_INTENDED; |
1083 | 595k | case 4: |
1084 | 595k | CRCsinglet(crc0, next, -4 * 8); |
1085 | 595k | FALLTHROUGH_INTENDED; |
1086 | 742k | case 3: |
1087 | 742k | CRCsinglet(crc0, next, -3 * 8); |
1088 | 742k | FALLTHROUGH_INTENDED; |
1089 | 1.91M | case 2: |
1090 | 1.91M | CRCsinglet(crc0, next, -2 * 8); |
1091 | 1.91M | FALLTHROUGH_INTENDED; |
1092 | 2.66M | case 1: |
1093 | 2.66M | CRCsinglet(crc0, next, -1 * 8); |
1094 | 2.66M | FALLTHROUGH_INTENDED; |
1095 | 2.76M | case 0:; |
1096 | 2.76M | } |
1097 | 2.76M | } |
1098 | 4.16M | { |
1099 | 4.16M | align_to_8(len, crc0, next); |
1100 | 4.16M | return (uint32_t)crc0 ^ 0xffffffffu; |
1101 | 4.16M | } |
1102 | 4.16M | } |
1103 | | |
1104 | | #endif //__SSE4_2__ && __PCLMUL__ |
1105 | | |
1106 | 2 | static inline Function Choose_Extend() { |
1107 | | #ifdef HAVE_POWER8 |
1108 | | return isAltiVec() ? ExtendPPCImpl : ExtendImpl<DefaultCRC32>; |
1109 | | #elif defined(HAVE_ARM64_CRC) |
1110 | | if(crc32c_runtime_check()) { |
1111 | | pmull_runtime_flag = crc32c_pmull_runtime_check(); |
1112 | | return ExtendARMImpl; |
1113 | | } else { |
1114 | | return ExtendImpl<DefaultCRC32>; |
1115 | | } |
1116 | | #elif defined(__SSE4_2__) && defined(__PCLMUL__) && !defined NO_THREEWAY_CRC32C |
1117 | | // NOTE: runtime detection no longer supported on x86 |
1118 | | #ifdef _MSC_VER |
1119 | | #pragma warning(disable: 4551) |
1120 | | #endif |
1121 | 2 | (void)ExtendImpl<DefaultCRC32>; // suppress unused warning |
1122 | | #ifdef _MSC_VER |
1123 | | #pragma warning(default: 4551) |
1124 | | #endif |
1125 | 2 | return crc32c_3way; |
1126 | | #else |
1127 | | return ExtendImpl<DefaultCRC32>; |
1128 | | #endif |
1129 | 2 | } |
1130 | | |
1131 | | static Function ChosenExtend = Choose_Extend(); |
1132 | 4.16M | uint32_t Extend(uint32_t crc, const char* buf, size_t size) { |
1133 | 4.16M | return ChosenExtend(crc, buf, size); |
1134 | 4.16M | } |
1135 | | |
1136 | | // The code for crc32c combine, copied with permission from folly |
1137 | | |
1138 | | // Standard galois-field multiply. The only modification is that a, |
1139 | | // b, m, and p are all bit-reflected. |
1140 | | // |
1141 | | // https://en.wikipedia.org/wiki/Finite_field_arithmetic |
1142 | | static constexpr uint32_t gf_multiply_sw_1( |
1143 | 102M | size_t i, uint32_t p, uint32_t a, uint32_t b, uint32_t m) { |
1144 | | // clang-format off |
1145 | 102M | return i == 32 ? p : gf_multiply_sw_1( |
1146 | 99.4M | /* i = */ i + 1, |
1147 | 99.4M | /* p = */ p ^ ((0u-((b >> 31) & 1)) & a), |
1148 | 99.4M | /* a = */ (a >> 1) ^ ((0u-(a & 1)) & m), |
1149 | 99.4M | /* b = */ b << 1, |
1150 | 99.4M | /* m = */ m); |
1151 | | // clang-format on |
1152 | 102M | } |
1153 | 3.10M | static constexpr uint32_t gf_multiply_sw(uint32_t a, uint32_t b, uint32_t m) { |
1154 | 3.10M | return gf_multiply_sw_1(/* i = */ 0, /* p = */ 0, a, b, m); |
1155 | 3.10M | } |
1156 | | |
1157 | 0 | static constexpr uint32_t gf_square_sw(uint32_t a, uint32_t m) { |
1158 | 0 | return gf_multiply_sw(a, a, m); |
1159 | 0 | } |
1160 | | |
1161 | | template <size_t i, uint32_t m> |
1162 | | struct gf_powers_memo { |
1163 | | static constexpr uint32_t value = |
1164 | | gf_square_sw(gf_powers_memo<i - 1, m>::value, m); |
1165 | | }; |
1166 | | template <uint32_t m> |
1167 | | struct gf_powers_memo<0, m> { |
1168 | | static constexpr uint32_t value = m; |
1169 | | }; |
1170 | | |
1171 | | template <typename T, T... Ints> |
1172 | | struct integer_sequence { |
1173 | | using value_type = T; |
1174 | | static constexpr size_t size() { return sizeof...(Ints); } |
1175 | | }; |
1176 | | |
1177 | | template <typename T, std::size_t N, T... Is> |
1178 | | struct make_integer_sequence : make_integer_sequence<T, N - 1, N - 1, Is...> {}; |
1179 | | |
1180 | | template <typename T, T... Is> |
1181 | | struct make_integer_sequence<T, 0, Is...> : integer_sequence<T, Is...> {}; |
1182 | | |
1183 | | template <std::size_t N> |
1184 | | using make_index_sequence = make_integer_sequence<std::size_t, N>; |
1185 | | |
1186 | | template <uint32_t m> |
1187 | | struct gf_powers_make { |
1188 | | template <size_t... i> |
1189 | | using index_sequence = integer_sequence<size_t, i...>; |
1190 | | template <size_t... i> |
1191 | | constexpr std::array<uint32_t, sizeof...(i)> operator()( |
1192 | 0 | index_sequence<i...>) const { |
1193 | 0 | return std::array<uint32_t, sizeof...(i)>{{gf_powers_memo<i, m>::value...}}; |
1194 | 0 | } |
1195 | | }; |
1196 | | |
1197 | | static constexpr uint32_t crc32c_m = 0x82f63b78; |
1198 | | |
1199 | | static constexpr std::array<uint32_t, 62> const crc32c_powers = |
1200 | | gf_powers_make<crc32c_m>{}(make_index_sequence<62>{}); |
1201 | | |
1202 | | // Expects a "pure" crc (see Crc32cCombine) |
1203 | | static uint32_t Crc32AppendZeroes( |
1204 | | uint32_t crc, size_t len_over_4, uint32_t polynomial, |
1205 | 1.42M | std::array<uint32_t, 62> const& powers_array) { |
1206 | 1.42M | auto powers = powers_array.data(); |
1207 | | // Append by multiplying by consecutive powers of two of the zeroes |
1208 | | // array |
1209 | 1.42M | size_t len_bits = len_over_4; |
1210 | | |
1211 | 4.52M | while (len_bits) { |
1212 | | // Advance directly to next bit set. |
1213 | 3.10M | auto r = CountTrailingZeroBits(len_bits); |
1214 | 3.10M | len_bits >>= r; |
1215 | 3.10M | powers += r; |
1216 | | |
1217 | 3.10M | crc = gf_multiply_sw(crc, *powers, polynomial); |
1218 | | |
1219 | 3.10M | len_bits >>= 1; |
1220 | 3.10M | powers++; |
1221 | 3.10M | } |
1222 | | |
1223 | 1.42M | return crc; |
1224 | 1.42M | } |
1225 | | |
1226 | 3.92M | static inline uint32_t InvertedToPure(uint32_t crc) { return ~crc; } |
1227 | | |
1228 | 2.50M | static inline uint32_t PureToInverted(uint32_t crc) { return ~crc; } |
1229 | | |
1230 | 1.07M | static inline uint32_t PureExtend(uint32_t crc, const char* buf, size_t size) { |
1231 | 1.07M | return InvertedToPure(Extend(PureToInverted(crc), buf, size)); |
1232 | 1.07M | } |
1233 | | |
1234 | | // Background: |
1235 | | // RocksDB uses two kinds of crc32c values: masked and unmasked. Neither is |
1236 | | // a "pure" CRC because a pure CRC satisfies (^ for xor) |
1237 | | // crc(a ^ b) = crc(a) ^ crc(b) |
1238 | | // The unmasked is closest, and this function takes unmasked crc32c values. |
1239 | | // The unmasked values are impure in two ways: |
1240 | | // * The initial setting at the start of CRC computation is all 1 bits |
1241 | | // (like -1) instead of zero. |
1242 | | // * The result has all bits invered. |
1243 | | // Note that together, these result in the empty string having a crc32c of |
1244 | | // zero. See |
1245 | | // https://en.wikipedia.org/wiki/Computation_of_cyclic_redundancy_checks#CRC_variants |
1246 | | // |
1247 | | // Simplified version of strategy, using xor through pure CRCs (+ for concat): |
1248 | | // |
1249 | | // pure_crc(str1 + str2) = pure_crc(str1 + zeros(len(str2))) ^ |
1250 | | // pure_crc(zeros(len(str1)) + str2) |
1251 | | // |
1252 | | // because the xor of these two zero-padded strings is str1 + str2. For pure |
1253 | | // CRC, leading zeros don't affect the result, so we only need |
1254 | | // |
1255 | | // pure_crc(str1 + str2) = pure_crc(str1 + zeros(len(str2))) ^ |
1256 | | // pure_crc(str2) |
1257 | | // |
1258 | | // Considering we aren't working with pure CRCs, what is actually in the input? |
1259 | | // |
1260 | | // crc1 = PureToInverted(PureExtendCrc32c(-1, zeros, crc1len) ^ |
1261 | | // PureCrc32c(str1, crc1len)) |
1262 | | // crc2 = PureToInverted(PureExtendCrc32c(-1, zeros, crc2len) ^ |
1263 | | // PureCrc32c(str2, crc2len)) |
1264 | | // |
1265 | | // The result we want to compute is |
1266 | | // combined = PureToInverted(PureExtendCrc32c(PureExtendCrc32c(-1, zeros, |
1267 | | // crc1len) ^ |
1268 | | // PureCrc32c(str1, crc1len), |
1269 | | // zeros, crc2len) ^ |
1270 | | // PureCrc32c(str2, crc2len)) |
1271 | | // |
1272 | | // Thus, in addition to extending crc1 over the length of str2 in (virtual) |
1273 | | // zeros, we need to cancel out the -1 initializer that was used in computing |
1274 | | // crc2. To cancel it out, we also need to extend it over crc2len in zeros. |
1275 | | // To simplify, since the end of str1 and that -1 initializer for crc2 are at |
1276 | | // the same logical position, we can combine them before we extend over the |
1277 | | // zeros. |
1278 | 1.42M | uint32_t Crc32cCombine(uint32_t crc1, uint32_t crc2, size_t crc2len) { |
1279 | 1.42M | uint32_t pure_crc1_with_init = InvertedToPure(crc1); |
1280 | 1.42M | uint32_t pure_crc2_with_init = InvertedToPure(crc2); |
1281 | 1.42M | uint32_t pure_crc2_init = static_cast<uint32_t>(-1); |
1282 | | |
1283 | | // Append up to 32 bits of zeroes in the normal way |
1284 | 1.42M | char zeros[4] = {0, 0, 0, 0}; |
1285 | 1.42M | auto len = crc2len & 3; |
1286 | 1.42M | uint32_t tmp = pure_crc1_with_init ^ pure_crc2_init; |
1287 | 1.42M | if (len) { |
1288 | 1.07M | tmp = PureExtend(tmp, zeros, len); |
1289 | 1.07M | } |
1290 | 1.42M | return PureToInverted( |
1291 | 1.42M | Crc32AppendZeroes(tmp, crc2len / 4, crc32c_m, crc32c_powers) ^ |
1292 | 1.42M | pure_crc2_with_init); |
1293 | 1.42M | } |
1294 | | |
1295 | | } // namespace ROCKSDB_NAMESPACE::crc32c |