]>
Commit | Line | Data |
---|---|---|
7c673cae | 1 | // Copyright (c) 2011-present, Facebook, Inc. All rights reserved. |
11fdf7f2 TL |
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). | |
7c673cae FG |
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. | |
7c673cae | 12 | #include "util/crc32c.h" |
1e59de90 | 13 | |
7c673cae | 14 | #include <stdint.h> |
1e59de90 TL |
15 | |
16 | #include <array> | |
17 | #include <utility> | |
11fdf7f2 | 18 | #ifdef HAVE_SSE42 |
7c673cae | 19 | #include <nmmintrin.h> |
11fdf7f2 | 20 | #include <wmmintrin.h> |
7c673cae | 21 | #endif |
1e59de90 | 22 | |
20effc67 | 23 | #include "port/lang.h" |
11fdf7f2 | 24 | #include "util/coding.h" |
f67539c2 | 25 | #include "util/crc32c_arm64.h" |
1e59de90 | 26 | #include "util/math.h" |
f67539c2 | 27 | |
11fdf7f2 TL |
28 | #ifdef __powerpc64__ |
29 | #include "util/crc32c_ppc.h" | |
30 | #include "util/crc32c_ppc_constants.h" | |
31 | ||
32 | #if __linux__ | |
b3b6e05e | 33 | #ifdef ROCKSDB_AUXV_GETAUXVAL_PRESENT |
11fdf7f2 | 34 | #include <sys/auxv.h> |
b3b6e05e | 35 | #endif |
11fdf7f2 TL |
36 | |
37 | #ifndef PPC_FEATURE2_VEC_CRYPTO | |
38 | #define PPC_FEATURE2_VEC_CRYPTO 0x02000000 | |
7c673cae | 39 | #endif |
11fdf7f2 TL |
40 | |
41 | #ifndef AT_HWCAP2 | |
42 | #define AT_HWCAP2 26 | |
43 | #endif | |
44 | ||
1e59de90 TL |
45 | #elif __FreeBSD__ |
46 | #include <machine/cpu.h> | |
47 | #include <sys/auxv.h> | |
48 | #include <sys/elf_common.h> | |
11fdf7f2 TL |
49 | #endif /* __linux__ */ |
50 | ||
7c673cae | 51 | #endif |
7c673cae | 52 | |
1e59de90 | 53 | #if defined(HAVE_ARM64_CRC) |
b3b6e05e TL |
54 | bool pmull_runtime_flag = false; |
55 | #endif | |
56 | ||
f67539c2 | 57 | namespace ROCKSDB_NAMESPACE { |
7c673cae FG |
58 | namespace crc32c { |
59 | ||
11fdf7f2 TL |
60 | #if defined(HAVE_POWER8) && defined(HAS_ALTIVEC) |
61 | #ifdef __powerpc64__ | |
62 | static int arch_ppc_crc32 = 0; | |
63 | #endif /* __powerpc64__ */ | |
64 | #endif | |
65 | ||
7c673cae | 66 | static const uint32_t table0_[256] = { |
1e59de90 TL |
67 | 0x00000000, 0xf26b8303, 0xe13b70f7, 0x1350f3f4, 0xc79a971f, 0x35f1141c, |
68 | 0x26a1e7e8, 0xd4ca64eb, 0x8ad958cf, 0x78b2dbcc, 0x6be22838, 0x9989ab3b, | |
69 | 0x4d43cfd0, 0xbf284cd3, 0xac78bf27, 0x5e133c24, 0x105ec76f, 0xe235446c, | |
70 | 0xf165b798, 0x030e349b, 0xd7c45070, 0x25afd373, 0x36ff2087, 0xc494a384, | |
71 | 0x9a879fa0, 0x68ec1ca3, 0x7bbcef57, 0x89d76c54, 0x5d1d08bf, 0xaf768bbc, | |
72 | 0xbc267848, 0x4e4dfb4b, 0x20bd8ede, 0xd2d60ddd, 0xc186fe29, 0x33ed7d2a, | |
73 | 0xe72719c1, 0x154c9ac2, 0x061c6936, 0xf477ea35, 0xaa64d611, 0x580f5512, | |
74 | 0x4b5fa6e6, 0xb93425e5, 0x6dfe410e, 0x9f95c20d, 0x8cc531f9, 0x7eaeb2fa, | |
75 | 0x30e349b1, 0xc288cab2, 0xd1d83946, 0x23b3ba45, 0xf779deae, 0x05125dad, | |
76 | 0x1642ae59, 0xe4292d5a, 0xba3a117e, 0x4851927d, 0x5b016189, 0xa96ae28a, | |
77 | 0x7da08661, 0x8fcb0562, 0x9c9bf696, 0x6ef07595, 0x417b1dbc, 0xb3109ebf, | |
78 | 0xa0406d4b, 0x522bee48, 0x86e18aa3, 0x748a09a0, 0x67dafa54, 0x95b17957, | |
79 | 0xcba24573, 0x39c9c670, 0x2a993584, 0xd8f2b687, 0x0c38d26c, 0xfe53516f, | |
80 | 0xed03a29b, 0x1f682198, 0x5125dad3, 0xa34e59d0, 0xb01eaa24, 0x42752927, | |
81 | 0x96bf4dcc, 0x64d4cecf, 0x77843d3b, 0x85efbe38, 0xdbfc821c, 0x2997011f, | |
82 | 0x3ac7f2eb, 0xc8ac71e8, 0x1c661503, 0xee0d9600, 0xfd5d65f4, 0x0f36e6f7, | |
83 | 0x61c69362, 0x93ad1061, 0x80fde395, 0x72966096, 0xa65c047d, 0x5437877e, | |
84 | 0x4767748a, 0xb50cf789, 0xeb1fcbad, 0x197448ae, 0x0a24bb5a, 0xf84f3859, | |
85 | 0x2c855cb2, 0xdeeedfb1, 0xcdbe2c45, 0x3fd5af46, 0x7198540d, 0x83f3d70e, | |
86 | 0x90a324fa, 0x62c8a7f9, 0xb602c312, 0x44694011, 0x5739b3e5, 0xa55230e6, | |
87 | 0xfb410cc2, 0x092a8fc1, 0x1a7a7c35, 0xe811ff36, 0x3cdb9bdd, 0xceb018de, | |
88 | 0xdde0eb2a, 0x2f8b6829, 0x82f63b78, 0x709db87b, 0x63cd4b8f, 0x91a6c88c, | |
89 | 0x456cac67, 0xb7072f64, 0xa457dc90, 0x563c5f93, 0x082f63b7, 0xfa44e0b4, | |
90 | 0xe9141340, 0x1b7f9043, 0xcfb5f4a8, 0x3dde77ab, 0x2e8e845f, 0xdce5075c, | |
91 | 0x92a8fc17, 0x60c37f14, 0x73938ce0, 0x81f80fe3, 0x55326b08, 0xa759e80b, | |
92 | 0xb4091bff, 0x466298fc, 0x1871a4d8, 0xea1a27db, 0xf94ad42f, 0x0b21572c, | |
93 | 0xdfeb33c7, 0x2d80b0c4, 0x3ed04330, 0xccbbc033, 0xa24bb5a6, 0x502036a5, | |
94 | 0x4370c551, 0xb11b4652, 0x65d122b9, 0x97baa1ba, 0x84ea524e, 0x7681d14d, | |
95 | 0x2892ed69, 0xdaf96e6a, 0xc9a99d9e, 0x3bc21e9d, 0xef087a76, 0x1d63f975, | |
96 | 0x0e330a81, 0xfc588982, 0xb21572c9, 0x407ef1ca, 0x532e023e, 0xa145813d, | |
97 | 0x758fe5d6, 0x87e466d5, 0x94b49521, 0x66df1622, 0x38cc2a06, 0xcaa7a905, | |
98 | 0xd9f75af1, 0x2b9cd9f2, 0xff56bd19, 0x0d3d3e1a, 0x1e6dcdee, 0xec064eed, | |
99 | 0xc38d26c4, 0x31e6a5c7, 0x22b65633, 0xd0ddd530, 0x0417b1db, 0xf67c32d8, | |
100 | 0xe52cc12c, 0x1747422f, 0x49547e0b, 0xbb3ffd08, 0xa86f0efc, 0x5a048dff, | |
101 | 0x8ecee914, 0x7ca56a17, 0x6ff599e3, 0x9d9e1ae0, 0xd3d3e1ab, 0x21b862a8, | |
102 | 0x32e8915c, 0xc083125f, 0x144976b4, 0xe622f5b7, 0xf5720643, 0x07198540, | |
103 | 0x590ab964, 0xab613a67, 0xb831c993, 0x4a5a4a90, 0x9e902e7b, 0x6cfbad78, | |
104 | 0x7fab5e8c, 0x8dc0dd8f, 0xe330a81a, 0x115b2b19, 0x020bd8ed, 0xf0605bee, | |
105 | 0x24aa3f05, 0xd6c1bc06, 0xc5914ff2, 0x37faccf1, 0x69e9f0d5, 0x9b8273d6, | |
106 | 0x88d28022, 0x7ab90321, 0xae7367ca, 0x5c18e4c9, 0x4f48173d, 0xbd23943e, | |
107 | 0xf36e6f75, 0x0105ec76, 0x12551f82, 0xe03e9c81, 0x34f4f86a, 0xc69f7b69, | |
108 | 0xd5cf889d, 0x27a40b9e, 0x79b737ba, 0x8bdcb4b9, 0x988c474d, 0x6ae7c44e, | |
109 | 0xbe2da0a5, 0x4c4623a6, 0x5f16d052, 0xad7d5351}; | |
7c673cae | 110 | static const uint32_t table1_[256] = { |
1e59de90 TL |
111 | 0x00000000, 0x13a29877, 0x274530ee, 0x34e7a899, 0x4e8a61dc, 0x5d28f9ab, |
112 | 0x69cf5132, 0x7a6dc945, 0x9d14c3b8, 0x8eb65bcf, 0xba51f356, 0xa9f36b21, | |
113 | 0xd39ea264, 0xc03c3a13, 0xf4db928a, 0xe7790afd, 0x3fc5f181, 0x2c6769f6, | |
114 | 0x1880c16f, 0x0b225918, 0x714f905d, 0x62ed082a, 0x560aa0b3, 0x45a838c4, | |
115 | 0xa2d13239, 0xb173aa4e, 0x859402d7, 0x96369aa0, 0xec5b53e5, 0xfff9cb92, | |
116 | 0xcb1e630b, 0xd8bcfb7c, 0x7f8be302, 0x6c297b75, 0x58ced3ec, 0x4b6c4b9b, | |
117 | 0x310182de, 0x22a31aa9, 0x1644b230, 0x05e62a47, 0xe29f20ba, 0xf13db8cd, | |
118 | 0xc5da1054, 0xd6788823, 0xac154166, 0xbfb7d911, 0x8b507188, 0x98f2e9ff, | |
119 | 0x404e1283, 0x53ec8af4, 0x670b226d, 0x74a9ba1a, 0x0ec4735f, 0x1d66eb28, | |
120 | 0x298143b1, 0x3a23dbc6, 0xdd5ad13b, 0xcef8494c, 0xfa1fe1d5, 0xe9bd79a2, | |
121 | 0x93d0b0e7, 0x80722890, 0xb4958009, 0xa737187e, 0xff17c604, 0xecb55e73, | |
122 | 0xd852f6ea, 0xcbf06e9d, 0xb19da7d8, 0xa23f3faf, 0x96d89736, 0x857a0f41, | |
123 | 0x620305bc, 0x71a19dcb, 0x45463552, 0x56e4ad25, 0x2c896460, 0x3f2bfc17, | |
124 | 0x0bcc548e, 0x186eccf9, 0xc0d23785, 0xd370aff2, 0xe797076b, 0xf4359f1c, | |
125 | 0x8e585659, 0x9dface2e, 0xa91d66b7, 0xbabffec0, 0x5dc6f43d, 0x4e646c4a, | |
126 | 0x7a83c4d3, 0x69215ca4, 0x134c95e1, 0x00ee0d96, 0x3409a50f, 0x27ab3d78, | |
127 | 0x809c2506, 0x933ebd71, 0xa7d915e8, 0xb47b8d9f, 0xce1644da, 0xddb4dcad, | |
128 | 0xe9537434, 0xfaf1ec43, 0x1d88e6be, 0x0e2a7ec9, 0x3acdd650, 0x296f4e27, | |
129 | 0x53028762, 0x40a01f15, 0x7447b78c, 0x67e52ffb, 0xbf59d487, 0xacfb4cf0, | |
130 | 0x981ce469, 0x8bbe7c1e, 0xf1d3b55b, 0xe2712d2c, 0xd69685b5, 0xc5341dc2, | |
131 | 0x224d173f, 0x31ef8f48, 0x050827d1, 0x16aabfa6, 0x6cc776e3, 0x7f65ee94, | |
132 | 0x4b82460d, 0x5820de7a, 0xfbc3faf9, 0xe861628e, 0xdc86ca17, 0xcf245260, | |
133 | 0xb5499b25, 0xa6eb0352, 0x920cabcb, 0x81ae33bc, 0x66d73941, 0x7575a136, | |
134 | 0x419209af, 0x523091d8, 0x285d589d, 0x3bffc0ea, 0x0f186873, 0x1cbaf004, | |
135 | 0xc4060b78, 0xd7a4930f, 0xe3433b96, 0xf0e1a3e1, 0x8a8c6aa4, 0x992ef2d3, | |
136 | 0xadc95a4a, 0xbe6bc23d, 0x5912c8c0, 0x4ab050b7, 0x7e57f82e, 0x6df56059, | |
137 | 0x1798a91c, 0x043a316b, 0x30dd99f2, 0x237f0185, 0x844819fb, 0x97ea818c, | |
138 | 0xa30d2915, 0xb0afb162, 0xcac27827, 0xd960e050, 0xed8748c9, 0xfe25d0be, | |
139 | 0x195cda43, 0x0afe4234, 0x3e19eaad, 0x2dbb72da, 0x57d6bb9f, 0x447423e8, | |
140 | 0x70938b71, 0x63311306, 0xbb8de87a, 0xa82f700d, 0x9cc8d894, 0x8f6a40e3, | |
141 | 0xf50789a6, 0xe6a511d1, 0xd242b948, 0xc1e0213f, 0x26992bc2, 0x353bb3b5, | |
142 | 0x01dc1b2c, 0x127e835b, 0x68134a1e, 0x7bb1d269, 0x4f567af0, 0x5cf4e287, | |
143 | 0x04d43cfd, 0x1776a48a, 0x23910c13, 0x30339464, 0x4a5e5d21, 0x59fcc556, | |
144 | 0x6d1b6dcf, 0x7eb9f5b8, 0x99c0ff45, 0x8a626732, 0xbe85cfab, 0xad2757dc, | |
145 | 0xd74a9e99, 0xc4e806ee, 0xf00fae77, 0xe3ad3600, 0x3b11cd7c, 0x28b3550b, | |
146 | 0x1c54fd92, 0x0ff665e5, 0x759baca0, 0x663934d7, 0x52de9c4e, 0x417c0439, | |
147 | 0xa6050ec4, 0xb5a796b3, 0x81403e2a, 0x92e2a65d, 0xe88f6f18, 0xfb2df76f, | |
148 | 0xcfca5ff6, 0xdc68c781, 0x7b5fdfff, 0x68fd4788, 0x5c1aef11, 0x4fb87766, | |
149 | 0x35d5be23, 0x26772654, 0x12908ecd, 0x013216ba, 0xe64b1c47, 0xf5e98430, | |
150 | 0xc10e2ca9, 0xd2acb4de, 0xa8c17d9b, 0xbb63e5ec, 0x8f844d75, 0x9c26d502, | |
151 | 0x449a2e7e, 0x5738b609, 0x63df1e90, 0x707d86e7, 0x0a104fa2, 0x19b2d7d5, | |
152 | 0x2d557f4c, 0x3ef7e73b, 0xd98eedc6, 0xca2c75b1, 0xfecbdd28, 0xed69455f, | |
153 | 0x97048c1a, 0x84a6146d, 0xb041bcf4, 0xa3e32483}; | |
7c673cae | 154 | static const uint32_t table2_[256] = { |
1e59de90 TL |
155 | 0x00000000, 0xa541927e, 0x4f6f520d, 0xea2ec073, 0x9edea41a, 0x3b9f3664, |
156 | 0xd1b1f617, 0x74f06469, 0x38513ec5, 0x9d10acbb, 0x773e6cc8, 0xd27ffeb6, | |
157 | 0xa68f9adf, 0x03ce08a1, 0xe9e0c8d2, 0x4ca15aac, 0x70a27d8a, 0xd5e3eff4, | |
158 | 0x3fcd2f87, 0x9a8cbdf9, 0xee7cd990, 0x4b3d4bee, 0xa1138b9d, 0x045219e3, | |
159 | 0x48f3434f, 0xedb2d131, 0x079c1142, 0xa2dd833c, 0xd62de755, 0x736c752b, | |
160 | 0x9942b558, 0x3c032726, 0xe144fb14, 0x4405696a, 0xae2ba919, 0x0b6a3b67, | |
161 | 0x7f9a5f0e, 0xdadbcd70, 0x30f50d03, 0x95b49f7d, 0xd915c5d1, 0x7c5457af, | |
162 | 0x967a97dc, 0x333b05a2, 0x47cb61cb, 0xe28af3b5, 0x08a433c6, 0xade5a1b8, | |
163 | 0x91e6869e, 0x34a714e0, 0xde89d493, 0x7bc846ed, 0x0f382284, 0xaa79b0fa, | |
164 | 0x40577089, 0xe516e2f7, 0xa9b7b85b, 0x0cf62a25, 0xe6d8ea56, 0x43997828, | |
165 | 0x37691c41, 0x92288e3f, 0x78064e4c, 0xdd47dc32, 0xc76580d9, 0x622412a7, | |
166 | 0x880ad2d4, 0x2d4b40aa, 0x59bb24c3, 0xfcfab6bd, 0x16d476ce, 0xb395e4b0, | |
167 | 0xff34be1c, 0x5a752c62, 0xb05bec11, 0x151a7e6f, 0x61ea1a06, 0xc4ab8878, | |
168 | 0x2e85480b, 0x8bc4da75, 0xb7c7fd53, 0x12866f2d, 0xf8a8af5e, 0x5de93d20, | |
169 | 0x29195949, 0x8c58cb37, 0x66760b44, 0xc337993a, 0x8f96c396, 0x2ad751e8, | |
170 | 0xc0f9919b, 0x65b803e5, 0x1148678c, 0xb409f5f2, 0x5e273581, 0xfb66a7ff, | |
171 | 0x26217bcd, 0x8360e9b3, 0x694e29c0, 0xcc0fbbbe, 0xb8ffdfd7, 0x1dbe4da9, | |
172 | 0xf7908dda, 0x52d11fa4, 0x1e704508, 0xbb31d776, 0x511f1705, 0xf45e857b, | |
173 | 0x80aee112, 0x25ef736c, 0xcfc1b31f, 0x6a802161, 0x56830647, 0xf3c29439, | |
174 | 0x19ec544a, 0xbcadc634, 0xc85da25d, 0x6d1c3023, 0x8732f050, 0x2273622e, | |
175 | 0x6ed23882, 0xcb93aafc, 0x21bd6a8f, 0x84fcf8f1, 0xf00c9c98, 0x554d0ee6, | |
176 | 0xbf63ce95, 0x1a225ceb, 0x8b277743, 0x2e66e53d, 0xc448254e, 0x6109b730, | |
177 | 0x15f9d359, 0xb0b84127, 0x5a968154, 0xffd7132a, 0xb3764986, 0x1637dbf8, | |
178 | 0xfc191b8b, 0x595889f5, 0x2da8ed9c, 0x88e97fe2, 0x62c7bf91, 0xc7862def, | |
179 | 0xfb850ac9, 0x5ec498b7, 0xb4ea58c4, 0x11abcaba, 0x655baed3, 0xc01a3cad, | |
180 | 0x2a34fcde, 0x8f756ea0, 0xc3d4340c, 0x6695a672, 0x8cbb6601, 0x29faf47f, | |
181 | 0x5d0a9016, 0xf84b0268, 0x1265c21b, 0xb7245065, 0x6a638c57, 0xcf221e29, | |
182 | 0x250cde5a, 0x804d4c24, 0xf4bd284d, 0x51fcba33, 0xbbd27a40, 0x1e93e83e, | |
183 | 0x5232b292, 0xf77320ec, 0x1d5de09f, 0xb81c72e1, 0xccec1688, 0x69ad84f6, | |
184 | 0x83834485, 0x26c2d6fb, 0x1ac1f1dd, 0xbf8063a3, 0x55aea3d0, 0xf0ef31ae, | |
185 | 0x841f55c7, 0x215ec7b9, 0xcb7007ca, 0x6e3195b4, 0x2290cf18, 0x87d15d66, | |
186 | 0x6dff9d15, 0xc8be0f6b, 0xbc4e6b02, 0x190ff97c, 0xf321390f, 0x5660ab71, | |
187 | 0x4c42f79a, 0xe90365e4, 0x032da597, 0xa66c37e9, 0xd29c5380, 0x77ddc1fe, | |
188 | 0x9df3018d, 0x38b293f3, 0x7413c95f, 0xd1525b21, 0x3b7c9b52, 0x9e3d092c, | |
189 | 0xeacd6d45, 0x4f8cff3b, 0xa5a23f48, 0x00e3ad36, 0x3ce08a10, 0x99a1186e, | |
190 | 0x738fd81d, 0xd6ce4a63, 0xa23e2e0a, 0x077fbc74, 0xed517c07, 0x4810ee79, | |
191 | 0x04b1b4d5, 0xa1f026ab, 0x4bdee6d8, 0xee9f74a6, 0x9a6f10cf, 0x3f2e82b1, | |
192 | 0xd50042c2, 0x7041d0bc, 0xad060c8e, 0x08479ef0, 0xe2695e83, 0x4728ccfd, | |
193 | 0x33d8a894, 0x96993aea, 0x7cb7fa99, 0xd9f668e7, 0x9557324b, 0x3016a035, | |
194 | 0xda386046, 0x7f79f238, 0x0b899651, 0xaec8042f, 0x44e6c45c, 0xe1a75622, | |
195 | 0xdda47104, 0x78e5e37a, 0x92cb2309, 0x378ab177, 0x437ad51e, 0xe63b4760, | |
196 | 0x0c158713, 0xa954156d, 0xe5f54fc1, 0x40b4ddbf, 0xaa9a1dcc, 0x0fdb8fb2, | |
197 | 0x7b2bebdb, 0xde6a79a5, 0x3444b9d6, 0x91052ba8}; | |
7c673cae | 198 | static const uint32_t table3_[256] = { |
1e59de90 TL |
199 | 0x00000000, 0xdd45aab8, 0xbf672381, 0x62228939, 0x7b2231f3, 0xa6679b4b, |
200 | 0xc4451272, 0x1900b8ca, 0xf64463e6, 0x2b01c95e, 0x49234067, 0x9466eadf, | |
201 | 0x8d665215, 0x5023f8ad, 0x32017194, 0xef44db2c, 0xe964b13d, 0x34211b85, | |
202 | 0x560392bc, 0x8b463804, 0x924680ce, 0x4f032a76, 0x2d21a34f, 0xf06409f7, | |
203 | 0x1f20d2db, 0xc2657863, 0xa047f15a, 0x7d025be2, 0x6402e328, 0xb9474990, | |
204 | 0xdb65c0a9, 0x06206a11, 0xd725148b, 0x0a60be33, 0x6842370a, 0xb5079db2, | |
205 | 0xac072578, 0x71428fc0, 0x136006f9, 0xce25ac41, 0x2161776d, 0xfc24ddd5, | |
206 | 0x9e0654ec, 0x4343fe54, 0x5a43469e, 0x8706ec26, 0xe524651f, 0x3861cfa7, | |
207 | 0x3e41a5b6, 0xe3040f0e, 0x81268637, 0x5c632c8f, 0x45639445, 0x98263efd, | |
208 | 0xfa04b7c4, 0x27411d7c, 0xc805c650, 0x15406ce8, 0x7762e5d1, 0xaa274f69, | |
209 | 0xb327f7a3, 0x6e625d1b, 0x0c40d422, 0xd1057e9a, 0xaba65fe7, 0x76e3f55f, | |
210 | 0x14c17c66, 0xc984d6de, 0xd0846e14, 0x0dc1c4ac, 0x6fe34d95, 0xb2a6e72d, | |
211 | 0x5de23c01, 0x80a796b9, 0xe2851f80, 0x3fc0b538, 0x26c00df2, 0xfb85a74a, | |
212 | 0x99a72e73, 0x44e284cb, 0x42c2eeda, 0x9f874462, 0xfda5cd5b, 0x20e067e3, | |
213 | 0x39e0df29, 0xe4a57591, 0x8687fca8, 0x5bc25610, 0xb4868d3c, 0x69c32784, | |
214 | 0x0be1aebd, 0xd6a40405, 0xcfa4bccf, 0x12e11677, 0x70c39f4e, 0xad8635f6, | |
215 | 0x7c834b6c, 0xa1c6e1d4, 0xc3e468ed, 0x1ea1c255, 0x07a17a9f, 0xdae4d027, | |
216 | 0xb8c6591e, 0x6583f3a6, 0x8ac7288a, 0x57828232, 0x35a00b0b, 0xe8e5a1b3, | |
217 | 0xf1e51979, 0x2ca0b3c1, 0x4e823af8, 0x93c79040, 0x95e7fa51, 0x48a250e9, | |
218 | 0x2a80d9d0, 0xf7c57368, 0xeec5cba2, 0x3380611a, 0x51a2e823, 0x8ce7429b, | |
219 | 0x63a399b7, 0xbee6330f, 0xdcc4ba36, 0x0181108e, 0x1881a844, 0xc5c402fc, | |
220 | 0xa7e68bc5, 0x7aa3217d, 0x52a0c93f, 0x8fe56387, 0xedc7eabe, 0x30824006, | |
221 | 0x2982f8cc, 0xf4c75274, 0x96e5db4d, 0x4ba071f5, 0xa4e4aad9, 0x79a10061, | |
222 | 0x1b838958, 0xc6c623e0, 0xdfc69b2a, 0x02833192, 0x60a1b8ab, 0xbde41213, | |
223 | 0xbbc47802, 0x6681d2ba, 0x04a35b83, 0xd9e6f13b, 0xc0e649f1, 0x1da3e349, | |
224 | 0x7f816a70, 0xa2c4c0c8, 0x4d801be4, 0x90c5b15c, 0xf2e73865, 0x2fa292dd, | |
225 | 0x36a22a17, 0xebe780af, 0x89c50996, 0x5480a32e, 0x8585ddb4, 0x58c0770c, | |
226 | 0x3ae2fe35, 0xe7a7548d, 0xfea7ec47, 0x23e246ff, 0x41c0cfc6, 0x9c85657e, | |
227 | 0x73c1be52, 0xae8414ea, 0xcca69dd3, 0x11e3376b, 0x08e38fa1, 0xd5a62519, | |
228 | 0xb784ac20, 0x6ac10698, 0x6ce16c89, 0xb1a4c631, 0xd3864f08, 0x0ec3e5b0, | |
229 | 0x17c35d7a, 0xca86f7c2, 0xa8a47efb, 0x75e1d443, 0x9aa50f6f, 0x47e0a5d7, | |
230 | 0x25c22cee, 0xf8878656, 0xe1873e9c, 0x3cc29424, 0x5ee01d1d, 0x83a5b7a5, | |
231 | 0xf90696d8, 0x24433c60, 0x4661b559, 0x9b241fe1, 0x8224a72b, 0x5f610d93, | |
232 | 0x3d4384aa, 0xe0062e12, 0x0f42f53e, 0xd2075f86, 0xb025d6bf, 0x6d607c07, | |
233 | 0x7460c4cd, 0xa9256e75, 0xcb07e74c, 0x16424df4, 0x106227e5, 0xcd278d5d, | |
234 | 0xaf050464, 0x7240aedc, 0x6b401616, 0xb605bcae, 0xd4273597, 0x09629f2f, | |
235 | 0xe6264403, 0x3b63eebb, 0x59416782, 0x8404cd3a, 0x9d0475f0, 0x4041df48, | |
236 | 0x22635671, 0xff26fcc9, 0x2e238253, 0xf36628eb, 0x9144a1d2, 0x4c010b6a, | |
237 | 0x5501b3a0, 0x88441918, 0xea669021, 0x37233a99, 0xd867e1b5, 0x05224b0d, | |
238 | 0x6700c234, 0xba45688c, 0xa345d046, 0x7e007afe, 0x1c22f3c7, 0xc167597f, | |
239 | 0xc747336e, 0x1a0299d6, 0x782010ef, 0xa565ba57, 0xbc65029d, 0x6120a825, | |
240 | 0x0302211c, 0xde478ba4, 0x31035088, 0xec46fa30, 0x8e647309, 0x5321d9b1, | |
241 | 0x4a21617b, 0x9764cbc3, 0xf54642fa, 0x2803e842}; | |
7c673cae FG |
242 | |
243 | // Used to fetch a naturally-aligned 32-bit word in little endian byte-order | |
1e59de90 | 244 | static inline uint32_t LE_LOAD32(const uint8_t* p) { |
7c673cae FG |
245 | return DecodeFixed32(reinterpret_cast<const char*>(p)); |
246 | } | |
247 | ||
11fdf7f2 | 248 | #if defined(HAVE_SSE42) && (defined(__LP64__) || defined(_WIN64)) |
1e59de90 | 249 | static inline uint64_t LE_LOAD64(const uint8_t* p) { |
7c673cae FG |
250 | return DecodeFixed64(reinterpret_cast<const char*>(p)); |
251 | } | |
252 | #endif | |
7c673cae | 253 | |
1e59de90 | 254 | static inline void Slow_CRC32(uint64_t* l, uint8_t const** p) { |
7c673cae FG |
255 | uint32_t c = static_cast<uint32_t>(*l ^ LE_LOAD32(*p)); |
256 | *p += 4; | |
1e59de90 TL |
257 | *l = table3_[c & 0xff] ^ table2_[(c >> 8) & 0xff] ^ |
258 | table1_[(c >> 16) & 0xff] ^ table0_[c >> 24]; | |
7c673cae FG |
259 | // DO it twice. |
260 | c = static_cast<uint32_t>(*l ^ LE_LOAD32(*p)); | |
261 | *p += 4; | |
1e59de90 TL |
262 | *l = table3_[c & 0xff] ^ table2_[(c >> 8) & 0xff] ^ |
263 | table1_[(c >> 16) & 0xff] ^ table0_[c >> 24]; | |
7c673cae FG |
264 | } |
265 | ||
1e59de90 TL |
266 | #if (!(defined(HAVE_POWER8) && defined(HAS_ALTIVEC))) && \ |
267 | (!defined(HAVE_ARM64_CRC)) || \ | |
268 | defined(NO_THREEWAY_CRC32C) | |
269 | static inline void Fast_CRC32(uint64_t* l, uint8_t const** p) { | |
11fdf7f2 TL |
270 | #ifndef HAVE_SSE42 |
271 | Slow_CRC32(l, p); | |
272 | #elif defined(__LP64__) || defined(_WIN64) | |
7c673cae FG |
273 | *l = _mm_crc32_u64(*l, LE_LOAD64(*p)); |
274 | *p += 8; | |
275 | #else | |
276 | *l = _mm_crc32_u32(static_cast<unsigned int>(*l), LE_LOAD32(*p)); | |
277 | *p += 4; | |
278 | *l = _mm_crc32_u32(static_cast<unsigned int>(*l), LE_LOAD32(*p)); | |
279 | *p += 4; | |
280 | #endif | |
7c673cae | 281 | } |
1e59de90 | 282 | #endif |
7c673cae | 283 | |
1e59de90 | 284 | template <void (*CRC32)(uint64_t*, uint8_t const**)> |
7c673cae | 285 | uint32_t ExtendImpl(uint32_t crc, const char* buf, size_t size) { |
1e59de90 TL |
286 | const uint8_t* p = reinterpret_cast<const uint8_t*>(buf); |
287 | const uint8_t* e = p + size; | |
7c673cae FG |
288 | uint64_t l = crc ^ 0xffffffffu; |
289 | ||
290 | // Align n to (1 << m) byte boundary | |
1e59de90 | 291 | #define ALIGN(n, m) ((n + ((1 << m) - 1)) & ~((1 << m) - 1)) |
7c673cae | 292 | |
1e59de90 TL |
293 | #define STEP1 \ |
294 | do { \ | |
295 | int c = (l & 0xff) ^ *p++; \ | |
296 | l = table0_[c] ^ (l >> 8); \ | |
297 | } while (0) | |
7c673cae FG |
298 | |
299 | // Point x at first 16-byte aligned byte in string. This might be | |
300 | // just past the end of the string. | |
301 | const uintptr_t pval = reinterpret_cast<uintptr_t>(p); | |
302 | const uint8_t* x = reinterpret_cast<const uint8_t*>(ALIGN(pval, 4)); | |
303 | if (x <= e) { | |
304 | // Process bytes until finished or p is 16-byte aligned | |
305 | while (p != x) { | |
306 | STEP1; | |
307 | } | |
308 | } | |
309 | // Process bytes 16 at a time | |
1e59de90 | 310 | while ((e - p) >= 16) { |
7c673cae FG |
311 | CRC32(&l, &p); |
312 | CRC32(&l, &p); | |
313 | } | |
314 | // Process bytes 8 at a time | |
1e59de90 | 315 | while ((e - p) >= 8) { |
7c673cae FG |
316 | CRC32(&l, &p); |
317 | } | |
318 | // Process the last few bytes | |
319 | while (p != e) { | |
320 | STEP1; | |
321 | } | |
322 | #undef STEP1 | |
323 | #undef ALIGN | |
324 | return static_cast<uint32_t>(l ^ 0xffffffffu); | |
325 | } | |
326 | ||
f67539c2 TL |
327 | // Detect if ARM64 CRC or not. |
328 | #ifndef HAVE_ARM64_CRC | |
7c673cae | 329 | // Detect if SS42 or not. |
11fdf7f2 TL |
330 | #ifndef HAVE_POWER8 |
331 | ||
7c673cae | 332 | static bool isSSE42() { |
11fdf7f2 TL |
333 | #ifndef HAVE_SSE42 |
334 | return false; | |
335 | #elif defined(__GNUC__) && defined(__x86_64__) && !defined(IOS_CROSS_COMPILE) | |
7c673cae | 336 | uint32_t c_; |
181888fb | 337 | __asm__("cpuid" : "=c"(c_) : "a"(1) : "ebx", "edx"); |
11fdf7f2 | 338 | return c_ & (1U << 20); // copied from CpuId.h in Folly. Test SSE42 |
7c673cae FG |
339 | #elif defined(_WIN64) |
340 | int info[4]; | |
341 | __cpuidex(info, 0x00000001, 0); | |
342 | return (info[2] & ((int)1 << 20)) != 0; | |
343 | #else | |
344 | return false; | |
345 | #endif | |
346 | } | |
347 | ||
11fdf7f2 TL |
348 | static bool isPCLMULQDQ() { |
349 | #ifndef HAVE_SSE42 | |
1e59de90 TL |
350 | // in build_detect_platform we set this macro when both SSE42 and PCLMULQDQ |
351 | // are supported by compiler | |
11fdf7f2 TL |
352 | return false; |
353 | #elif defined(__GNUC__) && defined(__x86_64__) && !defined(IOS_CROSS_COMPILE) | |
354 | uint32_t c_; | |
355 | __asm__("cpuid" : "=c"(c_) : "a"(1) : "ebx", "edx"); | |
356 | return c_ & (1U << 1); // PCLMULQDQ is in bit 1 (not bit 0) | |
357 | #elif defined(_WIN64) | |
358 | int info[4]; | |
359 | __cpuidex(info, 0x00000001, 0); | |
360 | return (info[2] & ((int)1 << 1)) != 0; | |
361 | #else | |
362 | return false; | |
363 | #endif | |
364 | } | |
365 | ||
366 | #endif // HAVE_POWER8 | |
f67539c2 | 367 | #endif // HAVE_ARM64_CRC |
11fdf7f2 | 368 | |
1e59de90 | 369 | using Function = uint32_t (*)(uint32_t, const char*, size_t); |
7c673cae | 370 | |
11fdf7f2 | 371 | #if defined(HAVE_POWER8) && defined(HAS_ALTIVEC) |
1e59de90 TL |
372 | uint32_t ExtendPPCImpl(uint32_t crc, const char* buf, size_t size) { |
373 | return crc32c_ppc(crc, (const unsigned char*)buf, size); | |
7c673cae FG |
374 | } |
375 | ||
11fdf7f2 TL |
376 | #if __linux__ |
377 | static int arch_ppc_probe(void) { | |
378 | arch_ppc_crc32 = 0; | |
379 | ||
b3b6e05e | 380 | #if defined(__powerpc64__) && defined(ROCKSDB_AUXV_GETAUXVAL_PRESENT) |
11fdf7f2 TL |
381 | if (getauxval(AT_HWCAP2) & PPC_FEATURE2_VEC_CRYPTO) arch_ppc_crc32 = 1; |
382 | #endif /* __powerpc64__ */ | |
383 | ||
384 | return arch_ppc_crc32; | |
385 | } | |
1e59de90 TL |
386 | #elif __FreeBSD__ |
387 | static int arch_ppc_probe(void) { | |
388 | unsigned long cpufeatures; | |
389 | arch_ppc_crc32 = 0; | |
390 | ||
391 | #if defined(__powerpc64__) | |
392 | elf_aux_info(AT_HWCAP2, &cpufeatures, sizeof(cpufeatures)); | |
393 | if (cpufeatures & PPC_FEATURE2_HAS_VEC_CRYPTO) arch_ppc_crc32 = 1; | |
394 | #endif /* __powerpc64__ */ | |
395 | ||
396 | return arch_ppc_crc32; | |
397 | } | |
11fdf7f2 TL |
398 | #endif // __linux__ |
399 | ||
400 | static bool isAltiVec() { | |
401 | if (arch_ppc_probe()) { | |
402 | return true; | |
403 | } else { | |
404 | return false; | |
405 | } | |
406 | } | |
407 | #endif | |
408 | ||
1e59de90 TL |
409 | #if defined(HAVE_ARM64_CRC) |
410 | uint32_t ExtendARMImpl(uint32_t crc, const char* buf, size_t size) { | |
411 | return crc32c_arm64(crc, (const unsigned char*)buf, size); | |
f67539c2 TL |
412 | } |
413 | #endif | |
11fdf7f2 TL |
414 | |
415 | std::string IsFastCrc32Supported() { | |
416 | bool has_fast_crc = false; | |
417 | std::string fast_zero_msg; | |
418 | std::string arch; | |
419 | #ifdef HAVE_POWER8 | |
420 | #ifdef HAS_ALTIVEC | |
421 | if (arch_ppc_probe()) { | |
422 | has_fast_crc = true; | |
423 | arch = "PPC"; | |
424 | } | |
7c673cae | 425 | #else |
11fdf7f2 TL |
426 | has_fast_crc = false; |
427 | arch = "PPC"; | |
428 | #endif | |
1e59de90 | 429 | #elif defined(HAVE_ARM64_CRC) |
f67539c2 TL |
430 | if (crc32c_runtime_check()) { |
431 | has_fast_crc = true; | |
432 | arch = "Arm64"; | |
b3b6e05e | 433 | pmull_runtime_flag = crc32c_pmull_runtime_check(); |
f67539c2 TL |
434 | } else { |
435 | has_fast_crc = false; | |
436 | arch = "Arm64"; | |
437 | } | |
11fdf7f2 TL |
438 | #else |
439 | has_fast_crc = isSSE42(); | |
440 | arch = "x86"; | |
7c673cae | 441 | #endif |
11fdf7f2 TL |
442 | if (has_fast_crc) { |
443 | fast_zero_msg.append("Supported on " + arch); | |
1e59de90 | 444 | } else { |
11fdf7f2 TL |
445 | fast_zero_msg.append("Not supported on " + arch); |
446 | } | |
447 | return fast_zero_msg; | |
7c673cae FG |
448 | } |
449 | ||
11fdf7f2 TL |
450 | /* |
451 | * Copyright 2016 Ferry Toth, Exalon Delft BV, The Netherlands | |
452 | * This software is provided 'as-is', without any express or implied | |
453 | * warranty. In no event will the author be held liable for any damages | |
454 | * arising from the use of this software. | |
455 | * Permission is granted to anyone to use this software for any purpose, | |
456 | * including commercial applications, and to alter it and redistribute it | |
457 | * freely, subject to the following restrictions: | |
458 | * 1. The origin of this software must not be misrepresented; you must not | |
459 | * claim that you wrote the original software. If you use this software | |
460 | * in a product, an acknowledgment in the product documentation would be | |
461 | * appreciated but is not required. | |
462 | * 2. Altered source versions must be plainly marked as such, and must not be | |
463 | * misrepresented as being the original software. | |
464 | * 3. This notice may not be removed or altered from any source distribution. | |
465 | * Ferry Toth | |
466 | * ftoth@exalondelft.nl | |
467 | * | |
468 | * https://github.com/htot/crc32c | |
469 | * | |
470 | * Modified by Facebook | |
471 | * | |
472 | * Original intel whitepaper: | |
473 | * "Fast CRC Computation for iSCSI Polynomial Using CRC32 Instruction" | |
474 | * https://www.intel.com/content/dam/www/public/us/en/documents/white-papers/crc-iscsi-polynomial-crc32-instruction-paper.pdf | |
475 | * | |
1e59de90 TL |
476 | * This version is from the folly library, created by Dave Watson |
477 | * <davejwatson@fb.com> | |
11fdf7f2 | 478 | * |
1e59de90 | 479 | */ |
11fdf7f2 TL |
480 | #if defined HAVE_SSE42 && defined HAVE_PCLMUL |
481 | ||
482 | #define CRCtriplet(crc, buf, offset) \ | |
483 | crc##0 = _mm_crc32_u64(crc##0, *(buf##0 + offset)); \ | |
484 | crc##1 = _mm_crc32_u64(crc##1, *(buf##1 + offset)); \ | |
485 | crc##2 = _mm_crc32_u64(crc##2, *(buf##2 + offset)); | |
486 | ||
487 | #define CRCduplet(crc, buf, offset) \ | |
488 | crc##0 = _mm_crc32_u64(crc##0, *(buf##0 + offset)); \ | |
489 | crc##1 = _mm_crc32_u64(crc##1, *(buf##1 + offset)); | |
490 | ||
1e59de90 | 491 | #define CRCsinglet(crc, buf, offset) \ |
11fdf7f2 TL |
492 | crc = _mm_crc32_u64(crc, *(uint64_t*)(buf + offset)); |
493 | ||
11fdf7f2 TL |
494 | // Numbers taken directly from intel whitepaper. |
495 | // clang-format off | |
496 | const uint64_t clmul_constants[] = { | |
497 | 0x14cd00bd6, 0x105ec76f0, 0x0ba4fc28e, 0x14cd00bd6, | |
498 | 0x1d82c63da, 0x0f20c0dfe, 0x09e4addf8, 0x0ba4fc28e, | |
499 | 0x039d3b296, 0x1384aa63a, 0x102f9b8a2, 0x1d82c63da, | |
500 | 0x14237f5e6, 0x01c291d04, 0x00d3b6092, 0x09e4addf8, | |
501 | 0x0c96cfdc0, 0x0740eef02, 0x18266e456, 0x039d3b296, | |
502 | 0x0daece73e, 0x0083a6eec, 0x0ab7aff2a, 0x102f9b8a2, | |
503 | 0x1248ea574, 0x1c1733996, 0x083348832, 0x14237f5e6, | |
504 | 0x12c743124, 0x02ad91c30, 0x0b9e02b86, 0x00d3b6092, | |
505 | 0x018b33a4e, 0x06992cea2, 0x1b331e26a, 0x0c96cfdc0, | |
506 | 0x17d35ba46, 0x07e908048, 0x1bf2e8b8a, 0x18266e456, | |
507 | 0x1a3e0968a, 0x11ed1f9d8, 0x0ce7f39f4, 0x0daece73e, | |
508 | 0x061d82e56, 0x0f1d0f55e, 0x0d270f1a2, 0x0ab7aff2a, | |
509 | 0x1c3f5f66c, 0x0a87ab8a8, 0x12ed0daac, 0x1248ea574, | |
510 | 0x065863b64, 0x08462d800, 0x11eef4f8e, 0x083348832, | |
511 | 0x1ee54f54c, 0x071d111a8, 0x0b3e32c28, 0x12c743124, | |
512 | 0x0064f7f26, 0x0ffd852c6, 0x0dd7e3b0c, 0x0b9e02b86, | |
513 | 0x0f285651c, 0x0dcb17aa4, 0x010746f3c, 0x018b33a4e, | |
514 | 0x1c24afea4, 0x0f37c5aee, 0x0271d9844, 0x1b331e26a, | |
515 | 0x08e766a0c, 0x06051d5a2, 0x093a5f730, 0x17d35ba46, | |
516 | 0x06cb08e5c, 0x11d5ca20e, 0x06b749fb2, 0x1bf2e8b8a, | |
517 | 0x1167f94f2, 0x021f3d99c, 0x0cec3662e, 0x1a3e0968a, | |
518 | 0x19329634a, 0x08f158014, 0x0e6fc4e6a, 0x0ce7f39f4, | |
519 | 0x08227bb8a, 0x1a5e82106, 0x0b0cd4768, 0x061d82e56, | |
520 | 0x13c2b89c4, 0x188815ab2, 0x0d7a4825c, 0x0d270f1a2, | |
521 | 0x10f5ff2ba, 0x105405f3e, 0x00167d312, 0x1c3f5f66c, | |
522 | 0x0f6076544, 0x0e9adf796, 0x026f6a60a, 0x12ed0daac, | |
523 | 0x1a2adb74e, 0x096638b34, 0x19d34af3a, 0x065863b64, | |
524 | 0x049c3cc9c, 0x1e50585a0, 0x068bce87a, 0x11eef4f8e, | |
525 | 0x1524fa6c6, 0x19f1c69dc, 0x16cba8aca, 0x1ee54f54c, | |
526 | 0x042d98888, 0x12913343e, 0x1329d9f7e, 0x0b3e32c28, | |
527 | 0x1b1c69528, 0x088f25a3a, 0x02178513a, 0x0064f7f26, | |
528 | 0x0e0ac139e, 0x04e36f0b0, 0x0170076fa, 0x0dd7e3b0c, | |
529 | 0x141a1a2e2, 0x0bd6f81f8, 0x16ad828b4, 0x0f285651c, | |
530 | 0x041d17b64, 0x19425cbba, 0x1fae1cc66, 0x010746f3c, | |
531 | 0x1a75b4b00, 0x18db37e8a, 0x0f872e54c, 0x1c24afea4, | |
532 | 0x01e41e9fc, 0x04c144932, 0x086d8e4d2, 0x0271d9844, | |
533 | 0x160f7af7a, 0x052148f02, 0x05bb8f1bc, 0x08e766a0c, | |
534 | 0x0a90fd27a, 0x0a3c6f37a, 0x0b3af077a, 0x093a5f730, | |
535 | 0x04984d782, 0x1d22c238e, 0x0ca6ef3ac, 0x06cb08e5c, | |
536 | 0x0234e0b26, 0x063ded06a, 0x1d88abd4a, 0x06b749fb2, | |
537 | 0x04597456a, 0x04d56973c, 0x0e9e28eb4, 0x1167f94f2, | |
538 | 0x07b3ff57a, 0x19385bf2e, 0x0c9c8b782, 0x0cec3662e, | |
539 | 0x13a9cba9e, 0x0e417f38a, 0x093e106a4, 0x19329634a, | |
540 | 0x167001a9c, 0x14e727980, 0x1ddffc5d4, 0x0e6fc4e6a, | |
541 | 0x00df04680, 0x0d104b8fc, 0x02342001e, 0x08227bb8a, | |
542 | 0x00a2a8d7e, 0x05b397730, 0x168763fa6, 0x0b0cd4768, | |
543 | 0x1ed5a407a, 0x0e78eb416, 0x0d2c3ed1a, 0x13c2b89c4, | |
544 | 0x0995a5724, 0x1641378f0, 0x19b1afbc4, 0x0d7a4825c, | |
545 | 0x109ffedc0, 0x08d96551c, 0x0f2271e60, 0x10f5ff2ba, | |
546 | 0x00b0bf8ca, 0x00bf80dd2, 0x123888b7a, 0x00167d312, | |
547 | 0x1e888f7dc, 0x18dcddd1c, 0x002ee03b2, 0x0f6076544, | |
548 | 0x183e8d8fe, 0x06a45d2b2, 0x133d7a042, 0x026f6a60a, | |
549 | 0x116b0f50c, 0x1dd3e10e8, 0x05fabe670, 0x1a2adb74e, | |
550 | 0x130004488, 0x0de87806c, 0x000bcf5f6, 0x19d34af3a, | |
551 | 0x18f0c7078, 0x014338754, 0x017f27698, 0x049c3cc9c, | |
552 | 0x058ca5f00, 0x15e3e77ee, 0x1af900c24, 0x068bce87a, | |
553 | 0x0b5cfca28, 0x0dd07448e, 0x0ded288f8, 0x1524fa6c6, | |
554 | 0x059f229bc, 0x1d8048348, 0x06d390dec, 0x16cba8aca, | |
555 | 0x037170390, 0x0a3e3e02c, 0x06353c1cc, 0x042d98888, | |
556 | 0x0c4584f5c, 0x0d73c7bea, 0x1f16a3418, 0x1329d9f7e, | |
557 | 0x0531377e2, 0x185137662, 0x1d8d9ca7c, 0x1b1c69528, | |
558 | 0x0b25b29f2, 0x18a08b5bc, 0x19fb2a8b0, 0x02178513a, | |
559 | 0x1a08fe6ac, 0x1da758ae0, 0x045cddf4e, 0x0e0ac139e, | |
560 | 0x1a91647f2, 0x169cf9eb0, 0x1a0f717c4, 0x0170076fa, | |
561 | }; | |
562 | ||
563 | // Compute the crc32c value for buffer smaller than 8 | |
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 | inline void align_to_8( | |
572 | size_t len, | |
573 | uint64_t& crc0, // crc so far, updated on return | |
574 | const unsigned char*& next) { // next data pointer, updated on return | |
575 | uint32_t crc32bit = static_cast<uint32_t>(crc0); | |
576 | if (len & 0x04) { | |
577 | crc32bit = _mm_crc32_u32(crc32bit, *(uint32_t*)next); | |
578 | next += sizeof(uint32_t); | |
579 | } | |
580 | if (len & 0x02) { | |
581 | crc32bit = _mm_crc32_u16(crc32bit, *(uint16_t*)next); | |
582 | next += sizeof(uint16_t); | |
583 | } | |
584 | if (len & 0x01) { | |
585 | crc32bit = _mm_crc32_u8(crc32bit, *(next)); | |
586 | next++; | |
587 | } | |
588 | crc0 = crc32bit; | |
589 | } | |
590 | ||
591 | // | |
592 | // CombineCRC performs pclmulqdq multiplication of 2 partial CRC's and a well | |
593 | // chosen constant and xor's these with the remaining CRC. | |
594 | // | |
595 | inline uint64_t CombineCRC( | |
596 | size_t block_size, | |
597 | uint64_t crc0, | |
598 | uint64_t crc1, | |
599 | uint64_t crc2, | |
600 | const uint64_t* next2) { | |
601 | const auto multiplier = | |
602 | *(reinterpret_cast<const __m128i*>(clmul_constants) + block_size - 1); | |
603 | const auto crc0_xmm = _mm_set_epi64x(0, crc0); | |
604 | const auto res0 = _mm_clmulepi64_si128(crc0_xmm, multiplier, 0x00); | |
605 | const auto crc1_xmm = _mm_set_epi64x(0, crc1); | |
606 | const auto res1 = _mm_clmulepi64_si128(crc1_xmm, multiplier, 0x10); | |
607 | const auto res = _mm_xor_si128(res0, res1); | |
608 | crc0 = _mm_cvtsi128_si64(res); | |
609 | crc0 = crc0 ^ *((uint64_t*)next2 - 1); | |
610 | crc2 = _mm_crc32_u64(crc2, crc0); | |
611 | return crc2; | |
612 | } | |
613 | ||
614 | // Compute CRC-32C using the Intel hardware instruction. | |
615 | #ifdef ROCKSDB_UBSAN_RUN | |
616 | #if defined(__clang__) | |
617 | __attribute__((__no_sanitize__("alignment"))) | |
618 | #elif defined(__GNUC__) | |
619 | __attribute__((__no_sanitize_undefined__)) | |
620 | #endif | |
621 | #endif | |
622 | uint32_t crc32c_3way(uint32_t crc, const char* buf, size_t len) { | |
623 | const unsigned char* next = (const unsigned char*)buf; | |
624 | uint64_t count; | |
625 | uint64_t crc0, crc1, crc2; | |
626 | crc0 = crc ^ 0xffffffffu; | |
627 | ||
628 | ||
629 | if (len >= 8) { | |
630 | // if len > 216 then align and use triplets | |
631 | if (len > 216) { | |
632 | { | |
633 | // Work on the bytes (< 8) before the first 8-byte alignment addr starts | |
634 | uint64_t align_bytes = (8 - (uintptr_t)next) & 7; | |
635 | len -= align_bytes; | |
636 | align_to_8(align_bytes, crc0, next); | |
637 | } | |
638 | ||
639 | // Now work on the remaining blocks | |
640 | count = len / 24; // number of triplets | |
641 | len %= 24; // bytes remaining | |
642 | uint64_t n = count >> 7; // #blocks = first block + full blocks | |
643 | uint64_t block_size = count & 127; | |
644 | if (block_size == 0) { | |
645 | block_size = 128; | |
646 | } else { | |
647 | n++; | |
648 | } | |
649 | // points to the first byte of the next block | |
650 | const uint64_t* next0 = (uint64_t*)next + block_size; | |
651 | const uint64_t* next1 = next0 + block_size; | |
652 | const uint64_t* next2 = next1 + block_size; | |
653 | ||
654 | crc1 = crc2 = 0; | |
655 | // Use Duff's device, a for() loop inside a switch() | |
656 | // statement. This needs to execute at least once, round len | |
657 | // down to nearest triplet multiple | |
658 | switch (block_size) { | |
659 | case 128: | |
660 | do { | |
661 | // jumps here for a full block of len 128 | |
662 | CRCtriplet(crc, next, -128); | |
20effc67 | 663 | FALLTHROUGH_INTENDED; |
11fdf7f2 TL |
664 | case 127: |
665 | // jumps here or below for the first block smaller | |
666 | CRCtriplet(crc, next, -127); | |
20effc67 | 667 | FALLTHROUGH_INTENDED; |
11fdf7f2 TL |
668 | case 126: |
669 | CRCtriplet(crc, next, -126); // than 128 | |
20effc67 | 670 | FALLTHROUGH_INTENDED; |
11fdf7f2 TL |
671 | case 125: |
672 | CRCtriplet(crc, next, -125); | |
20effc67 | 673 | FALLTHROUGH_INTENDED; |
11fdf7f2 TL |
674 | case 124: |
675 | CRCtriplet(crc, next, -124); | |
20effc67 | 676 | FALLTHROUGH_INTENDED; |
11fdf7f2 TL |
677 | case 123: |
678 | CRCtriplet(crc, next, -123); | |
20effc67 | 679 | FALLTHROUGH_INTENDED; |
11fdf7f2 TL |
680 | case 122: |
681 | CRCtriplet(crc, next, -122); | |
20effc67 | 682 | FALLTHROUGH_INTENDED; |
11fdf7f2 TL |
683 | case 121: |
684 | CRCtriplet(crc, next, -121); | |
20effc67 | 685 | FALLTHROUGH_INTENDED; |
11fdf7f2 TL |
686 | case 120: |
687 | CRCtriplet(crc, next, -120); | |
688 | FALLTHROUGH_INTENDED; | |
689 | case 119: | |
690 | CRCtriplet(crc, next, -119); | |
691 | FALLTHROUGH_INTENDED; | |
692 | case 118: | |
693 | CRCtriplet(crc, next, -118); | |
694 | FALLTHROUGH_INTENDED; | |
695 | case 117: | |
696 | CRCtriplet(crc, next, -117); | |
697 | FALLTHROUGH_INTENDED; | |
698 | case 116: | |
699 | CRCtriplet(crc, next, -116); | |
700 | FALLTHROUGH_INTENDED; | |
701 | case 115: | |
702 | CRCtriplet(crc, next, -115); | |
703 | FALLTHROUGH_INTENDED; | |
704 | case 114: | |
705 | CRCtriplet(crc, next, -114); | |
706 | FALLTHROUGH_INTENDED; | |
707 | case 113: | |
708 | CRCtriplet(crc, next, -113); | |
709 | FALLTHROUGH_INTENDED; | |
710 | case 112: | |
711 | CRCtriplet(crc, next, -112); | |
712 | FALLTHROUGH_INTENDED; | |
713 | case 111: | |
714 | CRCtriplet(crc, next, -111); | |
715 | FALLTHROUGH_INTENDED; | |
716 | case 110: | |
717 | CRCtriplet(crc, next, -110); | |
718 | FALLTHROUGH_INTENDED; | |
719 | case 109: | |
720 | CRCtriplet(crc, next, -109); | |
721 | FALLTHROUGH_INTENDED; | |
722 | case 108: | |
723 | CRCtriplet(crc, next, -108); | |
724 | FALLTHROUGH_INTENDED; | |
725 | case 107: | |
726 | CRCtriplet(crc, next, -107); | |
727 | FALLTHROUGH_INTENDED; | |
728 | case 106: | |
729 | CRCtriplet(crc, next, -106); | |
730 | FALLTHROUGH_INTENDED; | |
731 | case 105: | |
732 | CRCtriplet(crc, next, -105); | |
733 | FALLTHROUGH_INTENDED; | |
734 | case 104: | |
735 | CRCtriplet(crc, next, -104); | |
736 | FALLTHROUGH_INTENDED; | |
737 | case 103: | |
738 | CRCtriplet(crc, next, -103); | |
739 | FALLTHROUGH_INTENDED; | |
740 | case 102: | |
741 | CRCtriplet(crc, next, -102); | |
742 | FALLTHROUGH_INTENDED; | |
743 | case 101: | |
744 | CRCtriplet(crc, next, -101); | |
745 | FALLTHROUGH_INTENDED; | |
746 | case 100: | |
747 | CRCtriplet(crc, next, -100); | |
748 | FALLTHROUGH_INTENDED; | |
749 | case 99: | |
750 | CRCtriplet(crc, next, -99); | |
751 | FALLTHROUGH_INTENDED; | |
752 | case 98: | |
753 | CRCtriplet(crc, next, -98); | |
754 | FALLTHROUGH_INTENDED; | |
755 | case 97: | |
756 | CRCtriplet(crc, next, -97); | |
757 | FALLTHROUGH_INTENDED; | |
758 | case 96: | |
759 | CRCtriplet(crc, next, -96); | |
760 | FALLTHROUGH_INTENDED; | |
761 | case 95: | |
762 | CRCtriplet(crc, next, -95); | |
763 | FALLTHROUGH_INTENDED; | |
764 | case 94: | |
765 | CRCtriplet(crc, next, -94); | |
766 | FALLTHROUGH_INTENDED; | |
767 | case 93: | |
768 | CRCtriplet(crc, next, -93); | |
769 | FALLTHROUGH_INTENDED; | |
770 | case 92: | |
771 | CRCtriplet(crc, next, -92); | |
772 | FALLTHROUGH_INTENDED; | |
773 | case 91: | |
774 | CRCtriplet(crc, next, -91); | |
775 | FALLTHROUGH_INTENDED; | |
776 | case 90: | |
777 | CRCtriplet(crc, next, -90); | |
778 | FALLTHROUGH_INTENDED; | |
779 | case 89: | |
780 | CRCtriplet(crc, next, -89); | |
781 | FALLTHROUGH_INTENDED; | |
782 | case 88: | |
783 | CRCtriplet(crc, next, -88); | |
784 | FALLTHROUGH_INTENDED; | |
785 | case 87: | |
786 | CRCtriplet(crc, next, -87); | |
787 | FALLTHROUGH_INTENDED; | |
788 | case 86: | |
789 | CRCtriplet(crc, next, -86); | |
790 | FALLTHROUGH_INTENDED; | |
791 | case 85: | |
792 | CRCtriplet(crc, next, -85); | |
793 | FALLTHROUGH_INTENDED; | |
794 | case 84: | |
795 | CRCtriplet(crc, next, -84); | |
796 | FALLTHROUGH_INTENDED; | |
797 | case 83: | |
798 | CRCtriplet(crc, next, -83); | |
799 | FALLTHROUGH_INTENDED; | |
800 | case 82: | |
801 | CRCtriplet(crc, next, -82); | |
802 | FALLTHROUGH_INTENDED; | |
803 | case 81: | |
804 | CRCtriplet(crc, next, -81); | |
805 | FALLTHROUGH_INTENDED; | |
806 | case 80: | |
807 | CRCtriplet(crc, next, -80); | |
808 | FALLTHROUGH_INTENDED; | |
809 | case 79: | |
810 | CRCtriplet(crc, next, -79); | |
811 | FALLTHROUGH_INTENDED; | |
812 | case 78: | |
813 | CRCtriplet(crc, next, -78); | |
814 | FALLTHROUGH_INTENDED; | |
815 | case 77: | |
816 | CRCtriplet(crc, next, -77); | |
817 | FALLTHROUGH_INTENDED; | |
818 | case 76: | |
819 | CRCtriplet(crc, next, -76); | |
820 | FALLTHROUGH_INTENDED; | |
821 | case 75: | |
822 | CRCtriplet(crc, next, -75); | |
823 | FALLTHROUGH_INTENDED; | |
824 | case 74: | |
825 | CRCtriplet(crc, next, -74); | |
826 | FALLTHROUGH_INTENDED; | |
827 | case 73: | |
828 | CRCtriplet(crc, next, -73); | |
829 | FALLTHROUGH_INTENDED; | |
830 | case 72: | |
831 | CRCtriplet(crc, next, -72); | |
832 | FALLTHROUGH_INTENDED; | |
833 | case 71: | |
834 | CRCtriplet(crc, next, -71); | |
835 | FALLTHROUGH_INTENDED; | |
836 | case 70: | |
837 | CRCtriplet(crc, next, -70); | |
838 | FALLTHROUGH_INTENDED; | |
839 | case 69: | |
840 | CRCtriplet(crc, next, -69); | |
841 | FALLTHROUGH_INTENDED; | |
842 | case 68: | |
843 | CRCtriplet(crc, next, -68); | |
844 | FALLTHROUGH_INTENDED; | |
845 | case 67: | |
846 | CRCtriplet(crc, next, -67); | |
847 | FALLTHROUGH_INTENDED; | |
848 | case 66: | |
849 | CRCtriplet(crc, next, -66); | |
850 | FALLTHROUGH_INTENDED; | |
851 | case 65: | |
852 | CRCtriplet(crc, next, -65); | |
853 | FALLTHROUGH_INTENDED; | |
854 | case 64: | |
855 | CRCtriplet(crc, next, -64); | |
856 | FALLTHROUGH_INTENDED; | |
857 | case 63: | |
858 | CRCtriplet(crc, next, -63); | |
859 | FALLTHROUGH_INTENDED; | |
860 | case 62: | |
861 | CRCtriplet(crc, next, -62); | |
862 | FALLTHROUGH_INTENDED; | |
863 | case 61: | |
864 | CRCtriplet(crc, next, -61); | |
865 | FALLTHROUGH_INTENDED; | |
866 | case 60: | |
867 | CRCtriplet(crc, next, -60); | |
868 | FALLTHROUGH_INTENDED; | |
869 | case 59: | |
870 | CRCtriplet(crc, next, -59); | |
871 | FALLTHROUGH_INTENDED; | |
872 | case 58: | |
873 | CRCtriplet(crc, next, -58); | |
874 | FALLTHROUGH_INTENDED; | |
875 | case 57: | |
876 | CRCtriplet(crc, next, -57); | |
877 | FALLTHROUGH_INTENDED; | |
878 | case 56: | |
879 | CRCtriplet(crc, next, -56); | |
880 | FALLTHROUGH_INTENDED; | |
881 | case 55: | |
882 | CRCtriplet(crc, next, -55); | |
883 | FALLTHROUGH_INTENDED; | |
884 | case 54: | |
885 | CRCtriplet(crc, next, -54); | |
886 | FALLTHROUGH_INTENDED; | |
887 | case 53: | |
888 | CRCtriplet(crc, next, -53); | |
889 | FALLTHROUGH_INTENDED; | |
890 | case 52: | |
891 | CRCtriplet(crc, next, -52); | |
892 | FALLTHROUGH_INTENDED; | |
893 | case 51: | |
894 | CRCtriplet(crc, next, -51); | |
895 | FALLTHROUGH_INTENDED; | |
896 | case 50: | |
897 | CRCtriplet(crc, next, -50); | |
898 | FALLTHROUGH_INTENDED; | |
899 | case 49: | |
900 | CRCtriplet(crc, next, -49); | |
901 | FALLTHROUGH_INTENDED; | |
902 | case 48: | |
903 | CRCtriplet(crc, next, -48); | |
904 | FALLTHROUGH_INTENDED; | |
905 | case 47: | |
906 | CRCtriplet(crc, next, -47); | |
907 | FALLTHROUGH_INTENDED; | |
908 | case 46: | |
909 | CRCtriplet(crc, next, -46); | |
910 | FALLTHROUGH_INTENDED; | |
911 | case 45: | |
912 | CRCtriplet(crc, next, -45); | |
913 | FALLTHROUGH_INTENDED; | |
914 | case 44: | |
915 | CRCtriplet(crc, next, -44); | |
916 | FALLTHROUGH_INTENDED; | |
917 | case 43: | |
918 | CRCtriplet(crc, next, -43); | |
919 | FALLTHROUGH_INTENDED; | |
920 | case 42: | |
921 | CRCtriplet(crc, next, -42); | |
922 | FALLTHROUGH_INTENDED; | |
923 | case 41: | |
924 | CRCtriplet(crc, next, -41); | |
925 | FALLTHROUGH_INTENDED; | |
926 | case 40: | |
927 | CRCtriplet(crc, next, -40); | |
928 | FALLTHROUGH_INTENDED; | |
929 | case 39: | |
930 | CRCtriplet(crc, next, -39); | |
931 | FALLTHROUGH_INTENDED; | |
932 | case 38: | |
933 | CRCtriplet(crc, next, -38); | |
934 | FALLTHROUGH_INTENDED; | |
935 | case 37: | |
936 | CRCtriplet(crc, next, -37); | |
937 | FALLTHROUGH_INTENDED; | |
938 | case 36: | |
939 | CRCtriplet(crc, next, -36); | |
940 | FALLTHROUGH_INTENDED; | |
941 | case 35: | |
942 | CRCtriplet(crc, next, -35); | |
943 | FALLTHROUGH_INTENDED; | |
944 | case 34: | |
945 | CRCtriplet(crc, next, -34); | |
946 | FALLTHROUGH_INTENDED; | |
947 | case 33: | |
948 | CRCtriplet(crc, next, -33); | |
949 | FALLTHROUGH_INTENDED; | |
950 | case 32: | |
951 | CRCtriplet(crc, next, -32); | |
952 | FALLTHROUGH_INTENDED; | |
953 | case 31: | |
954 | CRCtriplet(crc, next, -31); | |
955 | FALLTHROUGH_INTENDED; | |
956 | case 30: | |
957 | CRCtriplet(crc, next, -30); | |
958 | FALLTHROUGH_INTENDED; | |
959 | case 29: | |
960 | CRCtriplet(crc, next, -29); | |
961 | FALLTHROUGH_INTENDED; | |
962 | case 28: | |
963 | CRCtriplet(crc, next, -28); | |
964 | FALLTHROUGH_INTENDED; | |
965 | case 27: | |
966 | CRCtriplet(crc, next, -27); | |
967 | FALLTHROUGH_INTENDED; | |
968 | case 26: | |
969 | CRCtriplet(crc, next, -26); | |
970 | FALLTHROUGH_INTENDED; | |
971 | case 25: | |
972 | CRCtriplet(crc, next, -25); | |
973 | FALLTHROUGH_INTENDED; | |
974 | case 24: | |
975 | CRCtriplet(crc, next, -24); | |
976 | FALLTHROUGH_INTENDED; | |
977 | case 23: | |
978 | CRCtriplet(crc, next, -23); | |
979 | FALLTHROUGH_INTENDED; | |
980 | case 22: | |
981 | CRCtriplet(crc, next, -22); | |
982 | FALLTHROUGH_INTENDED; | |
983 | case 21: | |
984 | CRCtriplet(crc, next, -21); | |
985 | FALLTHROUGH_INTENDED; | |
986 | case 20: | |
987 | CRCtriplet(crc, next, -20); | |
988 | FALLTHROUGH_INTENDED; | |
989 | case 19: | |
990 | CRCtriplet(crc, next, -19); | |
991 | FALLTHROUGH_INTENDED; | |
992 | case 18: | |
993 | CRCtriplet(crc, next, -18); | |
994 | FALLTHROUGH_INTENDED; | |
995 | case 17: | |
996 | CRCtriplet(crc, next, -17); | |
997 | FALLTHROUGH_INTENDED; | |
998 | case 16: | |
999 | CRCtriplet(crc, next, -16); | |
1000 | FALLTHROUGH_INTENDED; | |
1001 | case 15: | |
1002 | CRCtriplet(crc, next, -15); | |
1003 | FALLTHROUGH_INTENDED; | |
1004 | case 14: | |
1005 | CRCtriplet(crc, next, -14); | |
1006 | FALLTHROUGH_INTENDED; | |
1007 | case 13: | |
1008 | CRCtriplet(crc, next, -13); | |
1009 | FALLTHROUGH_INTENDED; | |
1010 | case 12: | |
1011 | CRCtriplet(crc, next, -12); | |
1012 | FALLTHROUGH_INTENDED; | |
1013 | case 11: | |
1014 | CRCtriplet(crc, next, -11); | |
1015 | FALLTHROUGH_INTENDED; | |
1016 | case 10: | |
1017 | CRCtriplet(crc, next, -10); | |
1018 | FALLTHROUGH_INTENDED; | |
1019 | case 9: | |
1020 | CRCtriplet(crc, next, -9); | |
1021 | FALLTHROUGH_INTENDED; | |
1022 | case 8: | |
1023 | CRCtriplet(crc, next, -8); | |
1024 | FALLTHROUGH_INTENDED; | |
1025 | case 7: | |
1026 | CRCtriplet(crc, next, -7); | |
1027 | FALLTHROUGH_INTENDED; | |
1028 | case 6: | |
1029 | CRCtriplet(crc, next, -6); | |
1030 | FALLTHROUGH_INTENDED; | |
1031 | case 5: | |
1032 | CRCtriplet(crc, next, -5); | |
1033 | FALLTHROUGH_INTENDED; | |
1034 | case 4: | |
1035 | CRCtriplet(crc, next, -4); | |
1036 | FALLTHROUGH_INTENDED; | |
1037 | case 3: | |
1038 | CRCtriplet(crc, next, -3); | |
1039 | FALLTHROUGH_INTENDED; | |
1040 | case 2: | |
1041 | CRCtriplet(crc, next, -2); | |
1042 | FALLTHROUGH_INTENDED; | |
1043 | case 1: | |
1044 | CRCduplet(crc, next, -1); // the final triplet is actually only 2 | |
1045 | //{ CombineCRC(); } | |
1046 | crc0 = CombineCRC(block_size, crc0, crc1, crc2, next2); | |
1047 | if (--n > 0) { | |
1048 | crc1 = crc2 = 0; | |
1049 | block_size = 128; | |
1050 | // points to the first byte of the next block | |
1051 | next0 = next2 + 128; | |
1052 | next1 = next0 + 128; // from here on all blocks are 128 long | |
1053 | next2 = next1 + 128; | |
1054 | } | |
1055 | FALLTHROUGH_INTENDED; | |
1056 | case 0:; | |
1057 | } while (n > 0); | |
1058 | } | |
1059 | next = (const unsigned char*)next2; | |
1060 | } | |
1061 | uint64_t count2 = len >> 3; // 216 of less bytes is 27 or less singlets | |
1062 | len = len & 7; | |
1063 | next += (count2 * 8); | |
1064 | switch (count2) { | |
1065 | case 27: | |
1066 | CRCsinglet(crc0, next, -27 * 8); | |
1067 | FALLTHROUGH_INTENDED; | |
1068 | case 26: | |
1069 | CRCsinglet(crc0, next, -26 * 8); | |
1070 | FALLTHROUGH_INTENDED; | |
1071 | case 25: | |
1072 | CRCsinglet(crc0, next, -25 * 8); | |
1073 | FALLTHROUGH_INTENDED; | |
1074 | case 24: | |
1075 | CRCsinglet(crc0, next, -24 * 8); | |
1076 | FALLTHROUGH_INTENDED; | |
1077 | case 23: | |
1078 | CRCsinglet(crc0, next, -23 * 8); | |
1079 | FALLTHROUGH_INTENDED; | |
1080 | case 22: | |
1081 | CRCsinglet(crc0, next, -22 * 8); | |
1082 | FALLTHROUGH_INTENDED; | |
1083 | case 21: | |
1084 | CRCsinglet(crc0, next, -21 * 8); | |
1085 | FALLTHROUGH_INTENDED; | |
1086 | case 20: | |
1087 | CRCsinglet(crc0, next, -20 * 8); | |
1088 | FALLTHROUGH_INTENDED; | |
1089 | case 19: | |
1090 | CRCsinglet(crc0, next, -19 * 8); | |
1091 | FALLTHROUGH_INTENDED; | |
1092 | case 18: | |
1093 | CRCsinglet(crc0, next, -18 * 8); | |
1094 | FALLTHROUGH_INTENDED; | |
1095 | case 17: | |
1096 | CRCsinglet(crc0, next, -17 * 8); | |
1097 | FALLTHROUGH_INTENDED; | |
1098 | case 16: | |
1099 | CRCsinglet(crc0, next, -16 * 8); | |
1100 | FALLTHROUGH_INTENDED; | |
1101 | case 15: | |
1102 | CRCsinglet(crc0, next, -15 * 8); | |
1103 | FALLTHROUGH_INTENDED; | |
1104 | case 14: | |
1105 | CRCsinglet(crc0, next, -14 * 8); | |
1106 | FALLTHROUGH_INTENDED; | |
1107 | case 13: | |
1108 | CRCsinglet(crc0, next, -13 * 8); | |
1109 | FALLTHROUGH_INTENDED; | |
1110 | case 12: | |
1111 | CRCsinglet(crc0, next, -12 * 8); | |
1112 | FALLTHROUGH_INTENDED; | |
1113 | case 11: | |
1114 | CRCsinglet(crc0, next, -11 * 8); | |
1115 | FALLTHROUGH_INTENDED; | |
1116 | case 10: | |
1117 | CRCsinglet(crc0, next, -10 * 8); | |
1118 | FALLTHROUGH_INTENDED; | |
1119 | case 9: | |
1120 | CRCsinglet(crc0, next, -9 * 8); | |
1121 | FALLTHROUGH_INTENDED; | |
1122 | case 8: | |
1123 | CRCsinglet(crc0, next, -8 * 8); | |
1124 | FALLTHROUGH_INTENDED; | |
1125 | case 7: | |
1126 | CRCsinglet(crc0, next, -7 * 8); | |
1127 | FALLTHROUGH_INTENDED; | |
1128 | case 6: | |
1129 | CRCsinglet(crc0, next, -6 * 8); | |
1130 | FALLTHROUGH_INTENDED; | |
1131 | case 5: | |
1132 | CRCsinglet(crc0, next, -5 * 8); | |
1133 | FALLTHROUGH_INTENDED; | |
1134 | case 4: | |
1135 | CRCsinglet(crc0, next, -4 * 8); | |
1136 | FALLTHROUGH_INTENDED; | |
1137 | case 3: | |
1138 | CRCsinglet(crc0, next, -3 * 8); | |
1139 | FALLTHROUGH_INTENDED; | |
1140 | case 2: | |
1141 | CRCsinglet(crc0, next, -2 * 8); | |
1142 | FALLTHROUGH_INTENDED; | |
1143 | case 1: | |
1144 | CRCsinglet(crc0, next, -1 * 8); | |
1145 | FALLTHROUGH_INTENDED; | |
1146 | case 0:; | |
1147 | } | |
1148 | } | |
1149 | { | |
1150 | align_to_8(len, crc0, next); | |
1151 | return (uint32_t)crc0 ^ 0xffffffffu; | |
1152 | } | |
1153 | } | |
1154 | ||
1155 | #endif //HAVE_SSE42 && HAVE_PCLMUL | |
1156 | ||
1157 | static inline Function Choose_Extend() { | |
f67539c2 TL |
1158 | #ifdef HAVE_POWER8 |
1159 | return isAltiVec() ? ExtendPPCImpl : ExtendImpl<Slow_CRC32>; | |
1e59de90 | 1160 | #elif defined(HAVE_ARM64_CRC) |
f67539c2 | 1161 | if(crc32c_runtime_check()) { |
b3b6e05e | 1162 | pmull_runtime_flag = crc32c_pmull_runtime_check(); |
f67539c2 TL |
1163 | return ExtendARMImpl; |
1164 | } else { | |
1165 | return ExtendImpl<Slow_CRC32>; | |
1166 | } | |
1167 | #else | |
11fdf7f2 TL |
1168 | if (isSSE42()) { |
1169 | if (isPCLMULQDQ()) { | |
1e59de90 | 1170 | #if (defined HAVE_SSE42 && defined HAVE_PCLMUL) && !defined NO_THREEWAY_CRC32C |
11fdf7f2 TL |
1171 | return crc32c_3way; |
1172 | #else | |
1173 | return ExtendImpl<Fast_CRC32>; // Fast_CRC32 will check HAVE_SSE42 itself | |
1174 | #endif | |
1175 | } | |
1176 | else { // no runtime PCLMULQDQ support but has SSE42 support | |
1177 | return ExtendImpl<Fast_CRC32>; | |
1178 | } | |
1179 | } // end of isSSE42() | |
1180 | else { | |
1181 | return ExtendImpl<Slow_CRC32>; | |
1182 | } | |
11fdf7f2 TL |
1183 | #endif |
1184 | } | |
1185 | ||
1186 | static Function ChosenExtend = Choose_Extend(); | |
7c673cae FG |
1187 | uint32_t Extend(uint32_t crc, const char* buf, size_t size) { |
1188 | return ChosenExtend(crc, buf, size); | |
1189 | } | |
1190 | ||
1e59de90 TL |
1191 | // The code for crc32c combine, copied with permission from folly |
1192 | ||
1193 | // Standard galois-field multiply. The only modification is that a, | |
1194 | // b, m, and p are all bit-reflected. | |
1195 | // | |
1196 | // https://en.wikipedia.org/wiki/Finite_field_arithmetic | |
1197 | static constexpr uint32_t gf_multiply_sw_1( | |
1198 | size_t i, uint32_t p, uint32_t a, uint32_t b, uint32_t m) { | |
1199 | // clang-format off | |
1200 | return i == 32 ? p : gf_multiply_sw_1( | |
1201 | /* i = */ i + 1, | |
1202 | /* p = */ p ^ ((0u-((b >> 31) & 1)) & a), | |
1203 | /* a = */ (a >> 1) ^ ((0u-(a & 1)) & m), | |
1204 | /* b = */ b << 1, | |
1205 | /* m = */ m); | |
1206 | // clang-format on | |
1207 | } | |
1208 | static constexpr uint32_t gf_multiply_sw(uint32_t a, uint32_t b, uint32_t m) { | |
1209 | return gf_multiply_sw_1(/* i = */ 0, /* p = */ 0, a, b, m); | |
1210 | } | |
1211 | ||
1212 | static constexpr uint32_t gf_square_sw(uint32_t a, uint32_t m) { | |
1213 | return gf_multiply_sw(a, a, m); | |
1214 | } | |
1215 | ||
1216 | template <size_t i, uint32_t m> | |
1217 | struct gf_powers_memo { | |
1218 | static constexpr uint32_t value = | |
1219 | gf_square_sw(gf_powers_memo<i - 1, m>::value, m); | |
1220 | }; | |
1221 | template <uint32_t m> | |
1222 | struct gf_powers_memo<0, m> { | |
1223 | static constexpr uint32_t value = m; | |
1224 | }; | |
1225 | ||
1226 | template <typename T, T... Ints> | |
1227 | struct integer_sequence { | |
1228 | using value_type = T; | |
1229 | static constexpr size_t size() { return sizeof...(Ints); } | |
1230 | }; | |
1231 | ||
1232 | template <typename T, std::size_t N, T... Is> | |
1233 | struct make_integer_sequence : make_integer_sequence<T, N - 1, N - 1, Is...> {}; | |
1234 | ||
1235 | template <typename T, T... Is> | |
1236 | struct make_integer_sequence<T, 0, Is...> : integer_sequence<T, Is...> {}; | |
1237 | ||
1238 | template <std::size_t N> | |
1239 | using make_index_sequence = make_integer_sequence<std::size_t, N>; | |
1240 | ||
1241 | template <uint32_t m> | |
1242 | struct gf_powers_make { | |
1243 | template <size_t... i> | |
1244 | using index_sequence = integer_sequence<size_t, i...>; | |
1245 | template <size_t... i> | |
1246 | constexpr std::array<uint32_t, sizeof...(i)> operator()( | |
1247 | index_sequence<i...>) const { | |
1248 | return std::array<uint32_t, sizeof...(i)>{{gf_powers_memo<i, m>::value...}}; | |
1249 | } | |
1250 | }; | |
1251 | ||
1252 | static constexpr uint32_t crc32c_m = 0x82f63b78; | |
1253 | ||
1254 | static constexpr std::array<uint32_t, 62> const crc32c_powers = | |
1255 | gf_powers_make<crc32c_m>{}(make_index_sequence<62>{}); | |
1256 | ||
1257 | // Expects a "pure" crc (see Crc32cCombine) | |
1258 | static uint32_t Crc32AppendZeroes( | |
1259 | uint32_t crc, size_t len_over_4, uint32_t polynomial, | |
1260 | std::array<uint32_t, 62> const& powers_array) { | |
1261 | auto powers = powers_array.data(); | |
1262 | // Append by multiplying by consecutive powers of two of the zeroes | |
1263 | // array | |
1264 | size_t len_bits = len_over_4; | |
1265 | ||
1266 | while (len_bits) { | |
1267 | // Advance directly to next bit set. | |
1268 | auto r = CountTrailingZeroBits(len_bits); | |
1269 | len_bits >>= r; | |
1270 | powers += r; | |
1271 | ||
1272 | crc = gf_multiply_sw(crc, *powers, polynomial); | |
1273 | ||
1274 | len_bits >>= 1; | |
1275 | powers++; | |
1276 | } | |
1277 | ||
1278 | return crc; | |
1279 | } | |
1280 | ||
1281 | static inline uint32_t InvertedToPure(uint32_t crc) { return ~crc; } | |
1282 | ||
1283 | static inline uint32_t PureToInverted(uint32_t crc) { return ~crc; } | |
1284 | ||
1285 | static inline uint32_t PureExtend(uint32_t crc, const char* buf, size_t size) { | |
1286 | return InvertedToPure(Extend(PureToInverted(crc), buf, size)); | |
1287 | } | |
1288 | ||
1289 | // Background: | |
1290 | // RocksDB uses two kinds of crc32c values: masked and unmasked. Neither is | |
1291 | // a "pure" CRC because a pure CRC satisfies (^ for xor) | |
1292 | // crc(a ^ b) = crc(a) ^ crc(b) | |
1293 | // The unmasked is closest, and this function takes unmasked crc32c values. | |
1294 | // The unmasked values are impure in two ways: | |
1295 | // * The initial setting at the start of CRC computation is all 1 bits | |
1296 | // (like -1) instead of zero. | |
1297 | // * The result has all bits invered. | |
1298 | // Note that together, these result in the empty string having a crc32c of | |
1299 | // zero. See | |
1300 | // https://en.wikipedia.org/wiki/Computation_of_cyclic_redundancy_checks#CRC_variants | |
1301 | // | |
1302 | // Simplified version of strategy, using xor through pure CRCs (+ for concat): | |
1303 | // | |
1304 | // pure_crc(str1 + str2) = pure_crc(str1 + zeros(len(str2))) ^ | |
1305 | // pure_crc(zeros(len(str1)) + str2) | |
1306 | // | |
1307 | // because the xor of these two zero-padded strings is str1 + str2. For pure | |
1308 | // CRC, leading zeros don't affect the result, so we only need | |
1309 | // | |
1310 | // pure_crc(str1 + str2) = pure_crc(str1 + zeros(len(str2))) ^ | |
1311 | // pure_crc(str2) | |
1312 | // | |
1313 | // Considering we aren't working with pure CRCs, what is actually in the input? | |
1314 | // | |
1315 | // crc1 = PureToInverted(PureExtendCrc32c(-1, zeros, crc1len) ^ | |
1316 | // PureCrc32c(str1, crc1len)) | |
1317 | // crc2 = PureToInverted(PureExtendCrc32c(-1, zeros, crc2len) ^ | |
1318 | // PureCrc32c(str2, crc2len)) | |
1319 | // | |
1320 | // The result we want to compute is | |
1321 | // combined = PureToInverted(PureExtendCrc32c(PureExtendCrc32c(-1, zeros, | |
1322 | // crc1len) ^ | |
1323 | // PureCrc32c(str1, crc1len), | |
1324 | // zeros, crc2len) ^ | |
1325 | // PureCrc32c(str2, crc2len)) | |
1326 | // | |
1327 | // Thus, in addition to extending crc1 over the length of str2 in (virtual) | |
1328 | // zeros, we need to cancel out the -1 initializer that was used in computing | |
1329 | // crc2. To cancel it out, we also need to extend it over crc2len in zeros. | |
1330 | // To simplify, since the end of str1 and that -1 initializer for crc2 are at | |
1331 | // the same logical position, we can combine them before we extend over the | |
1332 | // zeros. | |
1333 | uint32_t Crc32cCombine(uint32_t crc1, uint32_t crc2, size_t crc2len) { | |
1334 | uint32_t pure_crc1_with_init = InvertedToPure(crc1); | |
1335 | uint32_t pure_crc2_with_init = InvertedToPure(crc2); | |
1336 | uint32_t pure_crc2_init = static_cast<uint32_t>(-1); | |
1337 | ||
1338 | // Append up to 32 bits of zeroes in the normal way | |
1339 | char zeros[4] = {0, 0, 0, 0}; | |
1340 | auto len = crc2len & 3; | |
1341 | uint32_t tmp = pure_crc1_with_init ^ pure_crc2_init; | |
1342 | if (len) { | |
1343 | tmp = PureExtend(tmp, zeros, len); | |
1344 | } | |
1345 | return PureToInverted( | |
1346 | Crc32AppendZeroes(tmp, crc2len / 4, crc32c_m, crc32c_powers) ^ | |
1347 | pure_crc2_with_init); | |
1348 | } | |
11fdf7f2 | 1349 | |
7c673cae | 1350 | } // namespace crc32c |
f67539c2 | 1351 | } // namespace ROCKSDB_NAMESPACE |