cityhash.cpp 10 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321
  1. // SPDX-FileCopyrightText: 2011 Google, Inc.
  2. // SPDX-FileContributor: Geoff Pike
  3. // SPDX-FileContributor: Jyrki Alakuijala
  4. // SPDX-License-Identifier: MIT
  5. // CityHash, by Geoff Pike and Jyrki Alakuijala
  6. //
  7. // This file provides CityHash64() and related functions.
  8. //
  9. // It's probably possible to create even faster hash functions by
  10. // writing a program that systematically explores some of the space of
  11. // possible hash functions, by using SIMD instructions, or by
  12. // compromising on hash quality.
  13. #include <algorithm>
  14. #include <cstring>
  15. #include <utility>
  16. #include "common/cityhash.h"
  17. #include "common/swap.h"
  18. // #include "config.h"
  19. #ifdef __GNUC__
  20. #define HAVE_BUILTIN_EXPECT 1
  21. #endif
  22. #ifdef COMMON_BIG_ENDIAN
  23. #define WORDS_BIGENDIAN 1
  24. #endif
  25. using namespace std;
  26. namespace Common {
  27. static u64 unaligned_load64(const char* p) {
  28. u64 result;
  29. std::memcpy(&result, p, sizeof(result));
  30. return result;
  31. }
  32. static u32 unaligned_load32(const char* p) {
  33. u32 result;
  34. std::memcpy(&result, p, sizeof(result));
  35. return result;
  36. }
  37. #ifdef WORDS_BIGENDIAN
  38. #define uint32_in_expected_order(x) (swap32(x))
  39. #define uint64_in_expected_order(x) (swap64(x))
  40. #else
  41. #define uint32_in_expected_order(x) (x)
  42. #define uint64_in_expected_order(x) (x)
  43. #endif
  44. #if !defined(LIKELY)
  45. #if HAVE_BUILTIN_EXPECT
  46. #define LIKELY(x) (__builtin_expect(!!(x), 1))
  47. #else
  48. #define LIKELY(x) (x)
  49. #endif
  50. #endif
  51. static u64 Fetch64(const char* p) {
  52. return uint64_in_expected_order(unaligned_load64(p));
  53. }
  54. static u32 Fetch32(const char* p) {
  55. return uint32_in_expected_order(unaligned_load32(p));
  56. }
  57. // Some primes between 2^63 and 2^64 for various uses.
  58. static constexpr u64 k0 = 0xc3a5c85c97cb3127ULL;
  59. static constexpr u64 k1 = 0xb492b66fbe98f273ULL;
  60. static constexpr u64 k2 = 0x9ae16a3b2f90404fULL;
  61. // Bitwise right rotate. Normally this will compile to a single
  62. // instruction, especially if the shift is a manifest constant.
  63. static u64 Rotate(u64 val, int shift) {
  64. // Avoid shifting by 64: doing so yields an undefined result.
  65. return shift == 0 ? val : ((val >> shift) | (val << (64 - shift)));
  66. }
  67. static u64 ShiftMix(u64 val) {
  68. return val ^ (val >> 47);
  69. }
  70. static u64 HashLen16(u64 u, u64 v) {
  71. return Hash128to64(u128{u, v});
  72. }
  73. static u64 HashLen16(u64 u, u64 v, u64 mul) {
  74. // Murmur-inspired hashing.
  75. u64 a = (u ^ v) * mul;
  76. a ^= (a >> 47);
  77. u64 b = (v ^ a) * mul;
  78. b ^= (b >> 47);
  79. b *= mul;
  80. return b;
  81. }
  82. static u64 HashLen0to16(const char* s, size_t len) {
  83. if (len >= 8) {
  84. u64 mul = k2 + len * 2;
  85. u64 a = Fetch64(s) + k2;
  86. u64 b = Fetch64(s + len - 8);
  87. u64 c = Rotate(b, 37) * mul + a;
  88. u64 d = (Rotate(a, 25) + b) * mul;
  89. return HashLen16(c, d, mul);
  90. }
  91. if (len >= 4) {
  92. u64 mul = k2 + len * 2;
  93. u64 a = Fetch32(s);
  94. return HashLen16(len + (a << 3), Fetch32(s + len - 4), mul);
  95. }
  96. if (len > 0) {
  97. u8 a = s[0];
  98. u8 b = s[len >> 1];
  99. u8 c = s[len - 1];
  100. u32 y = static_cast<u32>(a) + (static_cast<u32>(b) << 8);
  101. u32 z = static_cast<u32>(len) + (static_cast<u32>(c) << 2);
  102. return ShiftMix(y * k2 ^ z * k0) * k2;
  103. }
  104. return k2;
  105. }
  106. // This probably works well for 16-byte strings as well, but it may be overkill
  107. // in that case.
  108. static u64 HashLen17to32(const char* s, size_t len) {
  109. u64 mul = k2 + len * 2;
  110. u64 a = Fetch64(s) * k1;
  111. u64 b = Fetch64(s + 8);
  112. u64 c = Fetch64(s + len - 8) * mul;
  113. u64 d = Fetch64(s + len - 16) * k2;
  114. return HashLen16(Rotate(a + b, 43) + Rotate(c, 30) + d, a + Rotate(b + k2, 18) + c, mul);
  115. }
  116. // Return a 16-byte hash for 48 bytes. Quick and dirty.
  117. // Callers do best to use "random-looking" values for a and b.
  118. static pair<u64, u64> WeakHashLen32WithSeeds(u64 w, u64 x, u64 y, u64 z, u64 a, u64 b) {
  119. a += w;
  120. b = Rotate(b + a + z, 21);
  121. u64 c = a;
  122. a += x;
  123. a += y;
  124. b += Rotate(a, 44);
  125. return make_pair(a + z, b + c);
  126. }
  127. // Return a 16-byte hash for s[0] ... s[31], a, and b. Quick and dirty.
  128. static pair<u64, u64> WeakHashLen32WithSeeds(const char* s, u64 a, u64 b) {
  129. return WeakHashLen32WithSeeds(Fetch64(s), Fetch64(s + 8), Fetch64(s + 16), Fetch64(s + 24), a,
  130. b);
  131. }
  132. // Return an 8-byte hash for 33 to 64 bytes.
  133. static u64 HashLen33to64(const char* s, size_t len) {
  134. u64 mul = k2 + len * 2;
  135. u64 a = Fetch64(s) * k2;
  136. u64 b = Fetch64(s + 8);
  137. u64 c = Fetch64(s + len - 24);
  138. u64 d = Fetch64(s + len - 32);
  139. u64 e = Fetch64(s + 16) * k2;
  140. u64 f = Fetch64(s + 24) * 9;
  141. u64 g = Fetch64(s + len - 8);
  142. u64 h = Fetch64(s + len - 16) * mul;
  143. u64 u = Rotate(a + g, 43) + (Rotate(b, 30) + c) * 9;
  144. u64 v = ((a + g) ^ d) + f + 1;
  145. u64 w = swap64((u + v) * mul) + h;
  146. u64 x = Rotate(e + f, 42) + c;
  147. u64 y = (swap64((v + w) * mul) + g) * mul;
  148. u64 z = e + f + c;
  149. a = swap64((x + z) * mul + y) + b;
  150. b = ShiftMix((z + a) * mul + d + h) * mul;
  151. return b + x;
  152. }
  153. u64 CityHash64(const char* s, size_t len) {
  154. if (len <= 32) {
  155. if (len <= 16) {
  156. return HashLen0to16(s, len);
  157. } else {
  158. return HashLen17to32(s, len);
  159. }
  160. } else if (len <= 64) {
  161. return HashLen33to64(s, len);
  162. }
  163. // For strings over 64 bytes we hash the end first, and then as we
  164. // loop we keep 56 bytes of state: v, w, x, y, and z.
  165. u64 x = Fetch64(s + len - 40);
  166. u64 y = Fetch64(s + len - 16) + Fetch64(s + len - 56);
  167. u64 z = HashLen16(Fetch64(s + len - 48) + len, Fetch64(s + len - 24));
  168. pair<u64, u64> v = WeakHashLen32WithSeeds(s + len - 64, len, z);
  169. pair<u64, u64> w = WeakHashLen32WithSeeds(s + len - 32, y + k1, x);
  170. x = x * k1 + Fetch64(s);
  171. // Decrease len to the nearest multiple of 64, and operate on 64-byte chunks.
  172. len = (len - 1) & ~static_cast<size_t>(63);
  173. do {
  174. x = Rotate(x + y + v.first + Fetch64(s + 8), 37) * k1;
  175. y = Rotate(y + v.second + Fetch64(s + 48), 42) * k1;
  176. x ^= w.second;
  177. y += v.first + Fetch64(s + 40);
  178. z = Rotate(z + w.first, 33) * k1;
  179. v = WeakHashLen32WithSeeds(s, v.second * k1, x + w.first);
  180. w = WeakHashLen32WithSeeds(s + 32, z + w.second, y + Fetch64(s + 16));
  181. std::swap(z, x);
  182. s += 64;
  183. len -= 64;
  184. } while (len != 0);
  185. return HashLen16(HashLen16(v.first, w.first) + ShiftMix(y) * k1 + z,
  186. HashLen16(v.second, w.second) + x);
  187. }
  188. u64 CityHash64WithSeed(const char* s, size_t len, u64 seed) {
  189. return CityHash64WithSeeds(s, len, k2, seed);
  190. }
  191. u64 CityHash64WithSeeds(const char* s, size_t len, u64 seed0, u64 seed1) {
  192. return HashLen16(CityHash64(s, len) - seed0, seed1);
  193. }
  194. // A subroutine for CityHash128(). Returns a decent 128-bit hash for strings
  195. // of any length representable in signed long. Based on City and Murmur.
  196. static u128 CityMurmur(const char* s, size_t len, u128 seed) {
  197. u64 a = seed[0];
  198. u64 b = seed[1];
  199. u64 c = 0;
  200. u64 d = 0;
  201. signed long l = static_cast<long>(len) - 16;
  202. if (l <= 0) { // len <= 16
  203. a = ShiftMix(a * k1) * k1;
  204. c = b * k1 + HashLen0to16(s, len);
  205. d = ShiftMix(a + (len >= 8 ? Fetch64(s) : c));
  206. } else { // len > 16
  207. c = HashLen16(Fetch64(s + len - 8) + k1, a);
  208. d = HashLen16(b + len, c + Fetch64(s + len - 16));
  209. a += d;
  210. do {
  211. a ^= ShiftMix(Fetch64(s) * k1) * k1;
  212. a *= k1;
  213. b ^= a;
  214. c ^= ShiftMix(Fetch64(s + 8) * k1) * k1;
  215. c *= k1;
  216. d ^= c;
  217. s += 16;
  218. l -= 16;
  219. } while (l > 0);
  220. }
  221. a = HashLen16(a, c);
  222. b = HashLen16(d, b);
  223. return u128{a ^ b, HashLen16(b, a)};
  224. }
  225. u128 CityHash128WithSeed(const char* s, size_t len, u128 seed) {
  226. if (len < 128) {
  227. return CityMurmur(s, len, seed);
  228. }
  229. // We expect len >= 128 to be the common case. Keep 56 bytes of state:
  230. // v, w, x, y, and z.
  231. pair<u64, u64> v, w;
  232. u64 x = seed[0];
  233. u64 y = seed[1];
  234. u64 z = len * k1;
  235. v.first = Rotate(y ^ k1, 49) * k1 + Fetch64(s);
  236. v.second = Rotate(v.first, 42) * k1 + Fetch64(s + 8);
  237. w.first = Rotate(y + z, 35) * k1 + x;
  238. w.second = Rotate(x + Fetch64(s + 88), 53) * k1;
  239. // This is the same inner loop as CityHash64(), manually unrolled.
  240. do {
  241. x = Rotate(x + y + v.first + Fetch64(s + 8), 37) * k1;
  242. y = Rotate(y + v.second + Fetch64(s + 48), 42) * k1;
  243. x ^= w.second;
  244. y += v.first + Fetch64(s + 40);
  245. z = Rotate(z + w.first, 33) * k1;
  246. v = WeakHashLen32WithSeeds(s, v.second * k1, x + w.first);
  247. w = WeakHashLen32WithSeeds(s + 32, z + w.second, y + Fetch64(s + 16));
  248. std::swap(z, x);
  249. s += 64;
  250. x = Rotate(x + y + v.first + Fetch64(s + 8), 37) * k1;
  251. y = Rotate(y + v.second + Fetch64(s + 48), 42) * k1;
  252. x ^= w.second;
  253. y += v.first + Fetch64(s + 40);
  254. z = Rotate(z + w.first, 33) * k1;
  255. v = WeakHashLen32WithSeeds(s, v.second * k1, x + w.first);
  256. w = WeakHashLen32WithSeeds(s + 32, z + w.second, y + Fetch64(s + 16));
  257. std::swap(z, x);
  258. s += 64;
  259. len -= 128;
  260. } while (LIKELY(len >= 128));
  261. x += Rotate(v.first + z, 49) * k0;
  262. y = y * k0 + Rotate(w.second, 37);
  263. z = z * k0 + Rotate(w.first, 27);
  264. w.first *= 9;
  265. v.first *= k0;
  266. // If 0 < len < 128, hash up to 4 chunks of 32 bytes each from the end of s.
  267. for (size_t tail_done = 0; tail_done < len;) {
  268. tail_done += 32;
  269. y = Rotate(x + y, 42) * k0 + v.second;
  270. w.first += Fetch64(s + len - tail_done + 16);
  271. x = x * k0 + w.first;
  272. z += w.second + Fetch64(s + len - tail_done);
  273. w.second += v.first;
  274. v = WeakHashLen32WithSeeds(s + len - tail_done, v.first + z, v.second);
  275. v.first *= k0;
  276. }
  277. // At this point our 56 bytes of state should contain more than
  278. // enough information for a strong 128-bit hash. We use two
  279. // different 56-byte-to-8-byte hashes to get a 16-byte final result.
  280. x = HashLen16(x, v.first);
  281. y = HashLen16(y + z, w.first);
  282. return u128{HashLen16(x + v.second, w.second) + y, HashLen16(x + w.second, y + v.second)};
  283. }
  284. u128 CityHash128(const char* s, size_t len) {
  285. return len >= 16 ? CityHash128WithSeed(s + 16, len - 16, u128{Fetch64(s), Fetch64(s + 8) + k0})
  286. : CityHash128WithSeed(s, len, u128{k0, k1});
  287. }
  288. } // namespace Common