1 | // Copyright (c) 2011 Google, Inc. |
2 | // |
3 | // Permission is hereby granted, free of charge, to any person obtaining a copy |
4 | // of this software and associated documentation files (the "Software"), to deal |
5 | // in the Software without restriction, including without limitation the rights |
6 | // to use, copy, modify, merge, publish, distribute, sublicense, and/or sell |
7 | // copies of the Software, and to permit persons to whom the Software is |
8 | // furnished to do so, subject to the following conditions: |
9 | // |
10 | // The above copyright notice and this permission notice shall be included in |
11 | // all copies or substantial portions of the Software. |
12 | // |
13 | // THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
14 | // IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
15 | // FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE |
16 | // AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
17 | // LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
18 | // OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN |
19 | // THE SOFTWARE. |
20 | // |
21 | // CityHash, by Geoff Pike and Jyrki Alakuijala |
22 | // |
23 | // This file provides CityHash64() and related functions. |
24 | // |
25 | // It's probably possible to create even faster hash functions by |
26 | // writing a program that systematically explores some of the space of |
27 | // possible hash functions, by using SIMD instructions, or by |
28 | // compromising on hash quality. |
29 | |
30 | #include "config.h" |
31 | #include <city.h> |
32 | |
33 | #include <algorithm> |
34 | #include <string.h> // for memcpy and memset |
35 | |
36 | using namespace std; |
37 | |
38 | static uint64 UNALIGNED_LOAD64(const char *p) { |
39 | uint64 result; |
40 | memcpy(&result, p, sizeof(result)); |
41 | return result; |
42 | } |
43 | |
44 | static uint32 UNALIGNED_LOAD32(const char *p) { |
45 | uint32 result; |
46 | memcpy(&result, p, sizeof(result)); |
47 | return result; |
48 | } |
49 | |
50 | #ifdef _MSC_VER |
51 | |
52 | #include <stdlib.h> |
53 | #define bswap_32(x) _byteswap_ulong(x) |
54 | #define bswap_64(x) _byteswap_uint64(x) |
55 | |
56 | #elif defined(__APPLE__) |
57 | |
58 | // Mac OS X / Darwin features |
59 | #include <libkern/OSByteOrder.h> |
60 | #define bswap_32(x) OSSwapInt32(x) |
61 | #define bswap_64(x) OSSwapInt64(x) |
62 | |
63 | #elif defined(__sun) || defined(sun) |
64 | |
65 | #include <sys/byteorder.h> |
66 | #define bswap_32(x) BSWAP_32(x) |
67 | #define bswap_64(x) BSWAP_64(x) |
68 | |
69 | #elif defined(__FreeBSD__) |
70 | |
71 | #include <sys/endian.h> |
72 | #define bswap_32(x) bswap32(x) |
73 | #define bswap_64(x) bswap64(x) |
74 | |
75 | #elif defined(__OpenBSD__) |
76 | |
77 | #include <sys/types.h> |
78 | #define bswap_32(x) swap32(x) |
79 | #define bswap_64(x) swap64(x) |
80 | |
81 | #elif defined(__NetBSD__) |
82 | |
83 | #include <sys/types.h> |
84 | #include <machine/bswap.h> |
85 | #if defined(__BSWAP_RENAME) && !defined(__bswap_32) |
86 | #define bswap_32(x) bswap32(x) |
87 | #define bswap_64(x) bswap64(x) |
88 | #endif |
89 | |
90 | #else |
91 | |
92 | #include <byteswap.h> |
93 | |
94 | #endif |
95 | |
96 | #ifdef WORDS_BIGENDIAN |
97 | #define uint32_in_expected_order(x) (bswap_32(x)) |
98 | #define uint64_in_expected_order(x) (bswap_64(x)) |
99 | #else |
100 | #define uint32_in_expected_order(x) (x) |
101 | #define uint64_in_expected_order(x) (x) |
102 | #endif |
103 | |
104 | #if !defined(LIKELY) |
105 | #if HAVE_BUILTIN_EXPECT |
106 | #define LIKELY(x) (__builtin_expect(!!(x), 1)) |
107 | #else |
108 | #define LIKELY(x) (x) |
109 | #endif |
110 | #endif |
111 | |
112 | static uint64 Fetch64(const char *p) { |
113 | return uint64_in_expected_order(UNALIGNED_LOAD64(p)); |
114 | } |
115 | |
116 | static uint32 Fetch32(const char *p) { |
117 | return uint32_in_expected_order(UNALIGNED_LOAD32(p)); |
118 | } |
119 | |
120 | // Some primes between 2^63 and 2^64 for various uses. |
121 | static const uint64 k0 = 0xc3a5c85c97cb3127ULL; |
122 | static const uint64 k1 = 0xb492b66fbe98f273ULL; |
123 | static const uint64 k2 = 0x9ae16a3b2f90404fULL; |
124 | |
125 | // Magic numbers for 32-bit hashing. Copied from Murmur3. |
126 | static const uint32 c1 = 0xcc9e2d51; |
127 | static const uint32 c2 = 0x1b873593; |
128 | |
129 | // A 32-bit to 32-bit integer hash copied from Murmur3. |
130 | static uint32 fmix(uint32 h) |
131 | { |
132 | h ^= h >> 16; |
133 | h *= 0x85ebca6b; |
134 | h ^= h >> 13; |
135 | h *= 0xc2b2ae35; |
136 | h ^= h >> 16; |
137 | return h; |
138 | } |
139 | |
140 | static uint32 Rotate32(uint32 val, int shift) { |
141 | // Avoid shifting by 32: doing so yields an undefined result. |
142 | return shift == 0 ? val : ((val >> shift) | (val << (32 - shift))); |
143 | } |
144 | |
145 | #undef PERMUTE3 |
146 | #define PERMUTE3(a, b, c) do { std::swap(a, b); std::swap(a, c); } while (0) |
147 | |
148 | static uint32 Mur(uint32 a, uint32 h) { |
149 | // Helper from Murmur3 for combining two 32-bit values. |
150 | a *= c1; |
151 | a = Rotate32(a, 17); |
152 | a *= c2; |
153 | h ^= a; |
154 | h = Rotate32(h, 19); |
155 | return h * 5 + 0xe6546b64; |
156 | } |
157 | |
158 | static uint32 Hash32Len13to24(const char *s, size_t len) { |
159 | uint32 a = Fetch32(s - 4 + (len >> 1)); |
160 | uint32 b = Fetch32(s + 4); |
161 | uint32 c = Fetch32(s + len - 8); |
162 | uint32 d = Fetch32(s + (len >> 1)); |
163 | uint32 e = Fetch32(s); |
164 | uint32 f = Fetch32(s + len - 4); |
165 | uint32 h = len; |
166 | |
167 | return fmix(Mur(f, Mur(e, Mur(d, Mur(c, Mur(b, Mur(a, h))))))); |
168 | } |
169 | |
170 | static uint32 Hash32Len0to4(const char *s, size_t len) { |
171 | uint32 b = 0; |
172 | uint32 c = 9; |
173 | for (size_t i = 0; i < len; i++) { |
174 | signed char v = s[i]; |
175 | b = b * c1 + v; |
176 | c ^= b; |
177 | } |
178 | return fmix(Mur(b, Mur(len, c))); |
179 | } |
180 | |
181 | static uint32 Hash32Len5to12(const char *s, size_t len) { |
182 | uint32 a = len, b = len * 5, c = 9, d = b; |
183 | a += Fetch32(s); |
184 | b += Fetch32(s + len - 4); |
185 | c += Fetch32(s + ((len >> 1) & 4)); |
186 | return fmix(Mur(c, Mur(b, Mur(a, d)))); |
187 | } |
188 | |
189 | uint32 CityHash32(const char *s, size_t len) { |
190 | if (len <= 24) { |
191 | return len <= 12 ? |
192 | (len <= 4 ? Hash32Len0to4(s, len) : Hash32Len5to12(s, len)) : |
193 | Hash32Len13to24(s, len); |
194 | } |
195 | |
196 | // len > 24 |
197 | uint32 h = len, g = c1 * len, f = g; |
198 | uint32 a0 = Rotate32(Fetch32(s + len - 4) * c1, 17) * c2; |
199 | uint32 a1 = Rotate32(Fetch32(s + len - 8) * c1, 17) * c2; |
200 | uint32 a2 = Rotate32(Fetch32(s + len - 16) * c1, 17) * c2; |
201 | uint32 a3 = Rotate32(Fetch32(s + len - 12) * c1, 17) * c2; |
202 | uint32 a4 = Rotate32(Fetch32(s + len - 20) * c1, 17) * c2; |
203 | h ^= a0; |
204 | h = Rotate32(h, 19); |
205 | h = h * 5 + 0xe6546b64; |
206 | h ^= a2; |
207 | h = Rotate32(h, 19); |
208 | h = h * 5 + 0xe6546b64; |
209 | g ^= a1; |
210 | g = Rotate32(g, 19); |
211 | g = g * 5 + 0xe6546b64; |
212 | g ^= a3; |
213 | g = Rotate32(g, 19); |
214 | g = g * 5 + 0xe6546b64; |
215 | f += a4; |
216 | f = Rotate32(f, 19); |
217 | f = f * 5 + 0xe6546b64; |
218 | size_t iters = (len - 1) / 20; |
219 | do { |
220 | uint32 a0 = Rotate32(Fetch32(s) * c1, 17) * c2; |
221 | uint32 a1 = Fetch32(s + 4); |
222 | uint32 a2 = Rotate32(Fetch32(s + 8) * c1, 17) * c2; |
223 | uint32 a3 = Rotate32(Fetch32(s + 12) * c1, 17) * c2; |
224 | uint32 a4 = Fetch32(s + 16); |
225 | h ^= a0; |
226 | h = Rotate32(h, 18); |
227 | h = h * 5 + 0xe6546b64; |
228 | f += a1; |
229 | f = Rotate32(f, 19); |
230 | f = f * c1; |
231 | g += a2; |
232 | g = Rotate32(g, 18); |
233 | g = g * 5 + 0xe6546b64; |
234 | h ^= a3 + a1; |
235 | h = Rotate32(h, 19); |
236 | h = h * 5 + 0xe6546b64; |
237 | g ^= a4; |
238 | g = bswap_32(g) * 5; |
239 | h += a4 * 5; |
240 | h = bswap_32(h); |
241 | f += a0; |
242 | PERMUTE3(f, h, g); |
243 | s += 20; |
244 | } while (--iters != 0); |
245 | g = Rotate32(g, 11) * c1; |
246 | g = Rotate32(g, 17) * c1; |
247 | f = Rotate32(f, 11) * c1; |
248 | f = Rotate32(f, 17) * c1; |
249 | h = Rotate32(h + g, 19); |
250 | h = h * 5 + 0xe6546b64; |
251 | h = Rotate32(h, 17) * c1; |
252 | h = Rotate32(h + f, 19); |
253 | h = h * 5 + 0xe6546b64; |
254 | h = Rotate32(h, 17) * c1; |
255 | return h; |
256 | } |
257 | |
258 | // Bitwise right rotate. Normally this will compile to a single |
259 | // instruction, especially if the shift is a manifest constant. |
260 | static uint64 Rotate(uint64 val, int shift) { |
261 | // Avoid shifting by 64: doing so yields an undefined result. |
262 | return shift == 0 ? val : ((val >> shift) | (val << (64 - shift))); |
263 | } |
264 | |
265 | static uint64 ShiftMix(uint64 val) { |
266 | return val ^ (val >> 47); |
267 | } |
268 | |
269 | static uint64 HashLen16(uint64 u, uint64 v) { |
270 | return Hash128to64(uint128(u, v)); |
271 | } |
272 | |
273 | static uint64 HashLen16(uint64 u, uint64 v, uint64 mul) { |
274 | // Murmur-inspired hashing. |
275 | uint64 a = (u ^ v) * mul; |
276 | a ^= (a >> 47); |
277 | uint64 b = (v ^ a) * mul; |
278 | b ^= (b >> 47); |
279 | b *= mul; |
280 | return b; |
281 | } |
282 | |
283 | static uint64 HashLen0to16(const char *s, size_t len) { |
284 | if (len >= 8) { |
285 | uint64 mul = k2 + len * 2; |
286 | uint64 a = Fetch64(s) + k2; |
287 | uint64 b = Fetch64(s + len - 8); |
288 | uint64 c = Rotate(b, 37) * mul + a; |
289 | uint64 d = (Rotate(a, 25) + b) * mul; |
290 | return HashLen16(c, d, mul); |
291 | } |
292 | if (len >= 4) { |
293 | uint64 mul = k2 + len * 2; |
294 | uint64 a = Fetch32(s); |
295 | return HashLen16(len + (a << 3), Fetch32(s + len - 4), mul); |
296 | } |
297 | if (len > 0) { |
298 | uint8 a = s[0]; |
299 | uint8 b = s[len >> 1]; |
300 | uint8 c = s[len - 1]; |
301 | uint32 y = static_cast<uint32>(a) + (static_cast<uint32>(b) << 8); |
302 | uint32 z = len + (static_cast<uint32>(c) << 2); |
303 | return ShiftMix(y * k2 ^ z * k0) * k2; |
304 | } |
305 | return k2; |
306 | } |
307 | |
308 | // This probably works well for 16-byte strings as well, but it may be overkill |
309 | // in that case. |
310 | static uint64 HashLen17to32(const char *s, size_t len) { |
311 | uint64 mul = k2 + len * 2; |
312 | uint64 a = Fetch64(s) * k1; |
313 | uint64 b = Fetch64(s + 8); |
314 | uint64 c = Fetch64(s + len - 8) * mul; |
315 | uint64 d = Fetch64(s + len - 16) * k2; |
316 | return HashLen16(Rotate(a + b, 43) + Rotate(c, 30) + d, |
317 | a + Rotate(b + k2, 18) + c, mul); |
318 | } |
319 | |
320 | // Return a 16-byte hash for 48 bytes. Quick and dirty. |
321 | // Callers do best to use "random-looking" values for a and b. |
322 | static pair<uint64, uint64> WeakHashLen32WithSeeds( |
323 | uint64 w, uint64 x, uint64 y, uint64 z, uint64 a, uint64 b) { |
324 | a += w; |
325 | b = Rotate(b + a + z, 21); |
326 | uint64 c = a; |
327 | a += x; |
328 | a += y; |
329 | b += Rotate(a, 44); |
330 | return make_pair(a + z, b + c); |
331 | } |
332 | |
333 | // Return a 16-byte hash for s[0] ... s[31], a, and b. Quick and dirty. |
334 | static pair<uint64, uint64> WeakHashLen32WithSeeds( |
335 | const char* s, uint64 a, uint64 b) { |
336 | return WeakHashLen32WithSeeds(Fetch64(s), |
337 | Fetch64(s + 8), |
338 | Fetch64(s + 16), |
339 | Fetch64(s + 24), |
340 | a, |
341 | b); |
342 | } |
343 | |
344 | // Return an 8-byte hash for 33 to 64 bytes. |
345 | static uint64 HashLen33to64(const char *s, size_t len) { |
346 | uint64 mul = k2 + len * 2; |
347 | uint64 a = Fetch64(s) * k2; |
348 | uint64 b = Fetch64(s + 8); |
349 | uint64 c = Fetch64(s + len - 24); |
350 | uint64 d = Fetch64(s + len - 32); |
351 | uint64 e = Fetch64(s + 16) * k2; |
352 | uint64 f = Fetch64(s + 24) * 9; |
353 | uint64 g = Fetch64(s + len - 8); |
354 | uint64 h = Fetch64(s + len - 16) * mul; |
355 | uint64 u = Rotate(a + g, 43) + (Rotate(b, 30) + c) * 9; |
356 | uint64 v = ((a + g) ^ d) + f + 1; |
357 | uint64 w = bswap_64((u + v) * mul) + h; |
358 | uint64 x = Rotate(e + f, 42) + c; |
359 | uint64 y = (bswap_64((v + w) * mul) + g) * mul; |
360 | uint64 z = e + f + c; |
361 | a = bswap_64((x + z) * mul + y) + b; |
362 | b = ShiftMix((z + a) * mul + d + h) * mul; |
363 | return b + x; |
364 | } |
365 | |
366 | uint64 CityHash64(const char *s, size_t len) { |
367 | if (len <= 32) { |
368 | if (len <= 16) { |
369 | return HashLen0to16(s, len); |
370 | } else { |
371 | return HashLen17to32(s, len); |
372 | } |
373 | } else if (len <= 64) { |
374 | return HashLen33to64(s, len); |
375 | } |
376 | |
377 | // For strings over 64 bytes we hash the end first, and then as we |
378 | // loop we keep 56 bytes of state: v, w, x, y, and z. |
379 | uint64 x = Fetch64(s + len - 40); |
380 | uint64 y = Fetch64(s + len - 16) + Fetch64(s + len - 56); |
381 | uint64 z = HashLen16(Fetch64(s + len - 48) + len, Fetch64(s + len - 24)); |
382 | pair<uint64, uint64> v = WeakHashLen32WithSeeds(s + len - 64, len, z); |
383 | pair<uint64, uint64> w = WeakHashLen32WithSeeds(s + len - 32, y + k1, x); |
384 | x = x * k1 + Fetch64(s); |
385 | |
386 | // Decrease len to the nearest multiple of 64, and operate on 64-byte chunks. |
387 | len = (len - 1) & ~static_cast<size_t>(63); |
388 | do { |
389 | x = Rotate(x + y + v.first + Fetch64(s + 8), 37) * k1; |
390 | y = Rotate(y + v.second + Fetch64(s + 48), 42) * k1; |
391 | x ^= w.second; |
392 | y += v.first + Fetch64(s + 40); |
393 | z = Rotate(z + w.first, 33) * k1; |
394 | v = WeakHashLen32WithSeeds(s, v.second * k1, x + w.first); |
395 | w = WeakHashLen32WithSeeds(s + 32, z + w.second, y + Fetch64(s + 16)); |
396 | std::swap(z, x); |
397 | s += 64; |
398 | len -= 64; |
399 | } while (len != 0); |
400 | return HashLen16(HashLen16(v.first, w.first) + ShiftMix(y) * k1 + z, |
401 | HashLen16(v.second, w.second) + x); |
402 | } |
403 | |
404 | uint64 CityHash64WithSeed(const char *s, size_t len, uint64 seed) { |
405 | return CityHash64WithSeeds(s, len, k2, seed); |
406 | } |
407 | |
408 | uint64 CityHash64WithSeeds(const char *s, size_t len, |
409 | uint64 seed0, uint64 seed1) { |
410 | return HashLen16(CityHash64(s, len) - seed0, seed1); |
411 | } |
412 | |
413 | // A subroutine for CityHash128(). Returns a decent 128-bit hash for strings |
414 | // of any length representable in signed long. Based on City and Murmur. |
415 | static uint128 CityMurmur(const char *s, size_t len, uint128 seed) { |
416 | uint64 a = Uint128Low64(seed); |
417 | uint64 b = Uint128High64(seed); |
418 | uint64 c = 0; |
419 | uint64 d = 0; |
420 | signed long l = len - 16; |
421 | if (l <= 0) { // len <= 16 |
422 | a = ShiftMix(a * k1) * k1; |
423 | c = b * k1 + HashLen0to16(s, len); |
424 | d = ShiftMix(a + (len >= 8 ? Fetch64(s) : c)); |
425 | } else { // len > 16 |
426 | c = HashLen16(Fetch64(s + len - 8) + k1, a); |
427 | d = HashLen16(b + len, c + Fetch64(s + len - 16)); |
428 | a += d; |
429 | do { |
430 | a ^= ShiftMix(Fetch64(s) * k1) * k1; |
431 | a *= k1; |
432 | b ^= a; |
433 | c ^= ShiftMix(Fetch64(s + 8) * k1) * k1; |
434 | c *= k1; |
435 | d ^= c; |
436 | s += 16; |
437 | l -= 16; |
438 | } while (l > 0); |
439 | } |
440 | a = HashLen16(a, c); |
441 | b = HashLen16(d, b); |
442 | return uint128(a ^ b, HashLen16(b, a)); |
443 | } |
444 | |
445 | uint128 CityHash128WithSeed(const char *s, size_t len, uint128 seed) { |
446 | if (len < 128) { |
447 | return CityMurmur(s, len, seed); |
448 | } |
449 | |
450 | // We expect len >= 128 to be the common case. Keep 56 bytes of state: |
451 | // v, w, x, y, and z. |
452 | pair<uint64, uint64> v, w; |
453 | uint64 x = Uint128Low64(seed); |
454 | uint64 y = Uint128High64(seed); |
455 | uint64 z = len * k1; |
456 | v.first = Rotate(y ^ k1, 49) * k1 + Fetch64(s); |
457 | v.second = Rotate(v.first, 42) * k1 + Fetch64(s + 8); |
458 | w.first = Rotate(y + z, 35) * k1 + x; |
459 | w.second = Rotate(x + Fetch64(s + 88), 53) * k1; |
460 | |
461 | // This is the same inner loop as CityHash64(), manually unrolled. |
462 | do { |
463 | x = Rotate(x + y + v.first + Fetch64(s + 8), 37) * k1; |
464 | y = Rotate(y + v.second + Fetch64(s + 48), 42) * k1; |
465 | x ^= w.second; |
466 | y += v.first + Fetch64(s + 40); |
467 | z = Rotate(z + w.first, 33) * k1; |
468 | v = WeakHashLen32WithSeeds(s, v.second * k1, x + w.first); |
469 | w = WeakHashLen32WithSeeds(s + 32, z + w.second, y + Fetch64(s + 16)); |
470 | std::swap(z, x); |
471 | s += 64; |
472 | x = Rotate(x + y + v.first + Fetch64(s + 8), 37) * k1; |
473 | y = Rotate(y + v.second + Fetch64(s + 48), 42) * k1; |
474 | x ^= w.second; |
475 | y += v.first + Fetch64(s + 40); |
476 | z = Rotate(z + w.first, 33) * k1; |
477 | v = WeakHashLen32WithSeeds(s, v.second * k1, x + w.first); |
478 | w = WeakHashLen32WithSeeds(s + 32, z + w.second, y + Fetch64(s + 16)); |
479 | std::swap(z, x); |
480 | s += 64; |
481 | len -= 128; |
482 | } while (LIKELY(len >= 128)); |
483 | x += Rotate(v.first + z, 49) * k0; |
484 | y = y * k0 + Rotate(w.second, 37); |
485 | z = z * k0 + Rotate(w.first, 27); |
486 | w.first *= 9; |
487 | v.first *= k0; |
488 | // If 0 < len < 128, hash up to 4 chunks of 32 bytes each from the end of s. |
489 | for (size_t tail_done = 0; tail_done < len; ) { |
490 | tail_done += 32; |
491 | y = Rotate(x + y, 42) * k0 + v.second; |
492 | w.first += Fetch64(s + len - tail_done + 16); |
493 | x = x * k0 + w.first; |
494 | z += w.second + Fetch64(s + len - tail_done); |
495 | w.second += v.first; |
496 | v = WeakHashLen32WithSeeds(s + len - tail_done, v.first + z, v.second); |
497 | v.first *= k0; |
498 | } |
499 | // At this point our 56 bytes of state should contain more than |
500 | // enough information for a strong 128-bit hash. We use two |
501 | // different 56-byte-to-8-byte hashes to get a 16-byte final result. |
502 | x = HashLen16(x, v.first); |
503 | y = HashLen16(y + z, w.first); |
504 | return uint128(HashLen16(x + v.second, w.second) + y, |
505 | HashLen16(x + w.second, y + v.second)); |
506 | } |
507 | |
508 | uint128 CityHash128(const char *s, size_t len) { |
509 | return len >= 16 ? |
510 | CityHash128WithSeed(s + 16, len - 16, |
511 | uint128(Fetch64(s), Fetch64(s + 8) + k0)) : |
512 | CityHash128WithSeed(s, len, uint128(k0, k1)); |
513 | } |
514 | |
515 | #ifdef __SSE4_2__ |
516 | #include <citycrc.h> |
517 | #include <nmmintrin.h> |
518 | |
519 | // Requires len >= 240. |
520 | static void CityHashCrc256Long(const char *s, size_t len, |
521 | uint32 seed, uint64 *result) { |
522 | uint64 a = Fetch64(s + 56) + k0; |
523 | uint64 b = Fetch64(s + 96) + k0; |
524 | uint64 c = result[0] = HashLen16(b, len); |
525 | uint64 d = result[1] = Fetch64(s + 120) * k0 + len; |
526 | uint64 e = Fetch64(s + 184) + seed; |
527 | uint64 f = 0; |
528 | uint64 g = 0; |
529 | uint64 h = c + d; |
530 | uint64 x = seed; |
531 | uint64 y = 0; |
532 | uint64 z = 0; |
533 | |
534 | // 240 bytes of input per iter. |
535 | size_t iters = len / 240; |
536 | len -= iters * 240; |
537 | do { |
538 | #undef CHUNK |
539 | #define CHUNK(r) \ |
540 | PERMUTE3(x, z, y); \ |
541 | b += Fetch64(s); \ |
542 | c += Fetch64(s + 8); \ |
543 | d += Fetch64(s + 16); \ |
544 | e += Fetch64(s + 24); \ |
545 | f += Fetch64(s + 32); \ |
546 | a += b; \ |
547 | h += f; \ |
548 | b += c; \ |
549 | f += d; \ |
550 | g += e; \ |
551 | e += z; \ |
552 | g += x; \ |
553 | z = _mm_crc32_u64(z, b + g); \ |
554 | y = _mm_crc32_u64(y, e + h); \ |
555 | x = _mm_crc32_u64(x, f + a); \ |
556 | e = Rotate(e, r); \ |
557 | c += e; \ |
558 | s += 40 |
559 | |
560 | CHUNK(0); PERMUTE3(a, h, c); |
561 | CHUNK(33); PERMUTE3(a, h, f); |
562 | CHUNK(0); PERMUTE3(b, h, f); |
563 | CHUNK(42); PERMUTE3(b, h, d); |
564 | CHUNK(0); PERMUTE3(b, h, e); |
565 | CHUNK(33); PERMUTE3(a, h, e); |
566 | } while (--iters > 0); |
567 | |
568 | while (len >= 40) { |
569 | CHUNK(29); |
570 | e ^= Rotate(a, 20); |
571 | h += Rotate(b, 30); |
572 | g ^= Rotate(c, 40); |
573 | f += Rotate(d, 34); |
574 | PERMUTE3(c, h, g); |
575 | len -= 40; |
576 | } |
577 | if (len > 0) { |
578 | s = s + len - 40; |
579 | CHUNK(33); |
580 | e ^= Rotate(a, 43); |
581 | h += Rotate(b, 42); |
582 | g ^= Rotate(c, 41); |
583 | f += Rotate(d, 40); |
584 | } |
585 | result[0] ^= h; |
586 | result[1] ^= g; |
587 | g += h; |
588 | a = HashLen16(a, g + z); |
589 | x += y << 32; |
590 | b += x; |
591 | c = HashLen16(c, z) + h; |
592 | d = HashLen16(d, e + result[0]); |
593 | g += e; |
594 | h += HashLen16(x, f); |
595 | e = HashLen16(a, d) + g; |
596 | z = HashLen16(b, c) + a; |
597 | y = HashLen16(g, h) + c; |
598 | result[0] = e + z + y + x; |
599 | a = ShiftMix((a + y) * k0) * k0 + b; |
600 | result[1] += a + result[0]; |
601 | a = ShiftMix(a * k0) * k0 + c; |
602 | result[2] = a + result[1]; |
603 | a = ShiftMix((a + e) * k0) * k0; |
604 | result[3] = a + result[2]; |
605 | } |
606 | |
607 | // Requires len < 240. |
608 | static void CityHashCrc256Short(const char *s, size_t len, uint64 *result) { |
609 | char buf[240]; |
610 | memcpy(buf, s, len); |
611 | memset(buf + len, 0, 240 - len); |
612 | CityHashCrc256Long(buf, 240, ~static_cast<uint32>(len), result); |
613 | } |
614 | |
615 | void CityHashCrc256(const char *s, size_t len, uint64 *result) { |
616 | if (LIKELY(len >= 240)) { |
617 | CityHashCrc256Long(s, len, 0, result); |
618 | } else { |
619 | CityHashCrc256Short(s, len, result); |
620 | } |
621 | } |
622 | |
623 | uint128 CityHashCrc128WithSeed(const char *s, size_t len, uint128 seed) { |
624 | if (len <= 900) { |
625 | return CityHash128WithSeed(s, len, seed); |
626 | } else { |
627 | uint64 result[4]; |
628 | CityHashCrc256(s, len, result); |
629 | uint64 u = Uint128High64(seed) + result[0]; |
630 | uint64 v = Uint128Low64(seed) + result[1]; |
631 | return uint128(HashLen16(u, v + result[2]), |
632 | HashLen16(Rotate(v, 32), u * k0 + result[3])); |
633 | } |
634 | } |
635 | |
636 | uint128 CityHashCrc128(const char *s, size_t len) { |
637 | if (len <= 900) { |
638 | return CityHash128(s, len); |
639 | } else { |
640 | uint64 result[4]; |
641 | CityHashCrc256(s, len, result); |
642 | return uint128(result[2], result[3]); |
643 | } |
644 | } |
645 | |
646 | #endif |
647 | |