1 // Copyright 2018 The Abseil Authors.
2 //
3 // Licensed under the Apache License, Version 2.0 (the "License");
4 // you may not use this file except in compliance with the License.
5 // You may obtain a copy of the License at
6 //
7 //      https://www.apache.org/licenses/LICENSE-2.0
8 //
9 // Unless required by applicable law or agreed to in writing, software
10 // distributed under the License is distributed on an "AS IS" BASIS,
11 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12 // See the License for the specific language governing permissions and
13 // limitations under the License.
14 //
15 // This file provides CityHash64() and related functions.
16 //
17 // It's probably possible to create even faster hash functions by
18 // writing a program that systematically explores some of the space of
19 // possible hash functions, by using SIMD instructions, or by
20 // compromising on hash quality.
21 
22 #include "absl/hash/internal/city.h"
23 
24 #include <string.h>  // for memcpy and memset
25 #include <algorithm>
26 
27 #include "absl/base/config.h"
28 #include "absl/base/internal/endian.h"
29 #include "absl/base/internal/unaligned_access.h"
30 #include "absl/base/optimization.h"
31 
32 namespace absl {
33 ABSL_NAMESPACE_BEGIN
34 namespace hash_internal {
35 
36 #ifdef ABSL_IS_BIG_ENDIAN
37 #define uint32_in_expected_order(x) (absl::gbswap_32(x))
38 #define uint64_in_expected_order(x) (absl::gbswap_64(x))
39 #else
40 #define uint32_in_expected_order(x) (x)
41 #define uint64_in_expected_order(x) (x)
42 #endif
43 
Fetch64(const char * p)44 static uint64_t Fetch64(const char *p) {
45   return uint64_in_expected_order(ABSL_INTERNAL_UNALIGNED_LOAD64(p));
46 }
47 
Fetch32(const char * p)48 static uint32_t Fetch32(const char *p) {
49   return uint32_in_expected_order(ABSL_INTERNAL_UNALIGNED_LOAD32(p));
50 }
51 
52 // Some primes between 2^63 and 2^64 for various uses.
53 static const uint64_t k0 = 0xc3a5c85c97cb3127ULL;
54 static const uint64_t k1 = 0xb492b66fbe98f273ULL;
55 static const uint64_t k2 = 0x9ae16a3b2f90404fULL;
56 
57 // Magic numbers for 32-bit hashing.  Copied from Murmur3.
58 static const uint32_t c1 = 0xcc9e2d51;
59 static const uint32_t c2 = 0x1b873593;
60 
61 // A 32-bit to 32-bit integer hash copied from Murmur3.
fmix(uint32_t h)62 static uint32_t fmix(uint32_t h) {
63   h ^= h >> 16;
64   h *= 0x85ebca6b;
65   h ^= h >> 13;
66   h *= 0xc2b2ae35;
67   h ^= h >> 16;
68   return h;
69 }
70 
Rotate32(uint32_t val,int shift)71 static uint32_t Rotate32(uint32_t val, int shift) {
72   // Avoid shifting by 32: doing so yields an undefined result.
73   return shift == 0 ? val : ((val >> shift) | (val << (32 - shift)));
74 }
75 
76 #undef PERMUTE3
77 #define PERMUTE3(a, b, c) \
78   do {                    \
79     std::swap(a, b);      \
80     std::swap(a, c);      \
81   } while (0)
82 
Mur(uint32_t a,uint32_t h)83 static uint32_t Mur(uint32_t a, uint32_t h) {
84   // Helper from Murmur3 for combining two 32-bit values.
85   a *= c1;
86   a = Rotate32(a, 17);
87   a *= c2;
88   h ^= a;
89   h = Rotate32(h, 19);
90   return h * 5 + 0xe6546b64;
91 }
92 
Hash32Len13to24(const char * s,size_t len)93 static uint32_t Hash32Len13to24(const char *s, size_t len) {
94   uint32_t a = Fetch32(s - 4 + (len >> 1));
95   uint32_t b = Fetch32(s + 4);
96   uint32_t c = Fetch32(s + len - 8);
97   uint32_t d = Fetch32(s + (len >> 1));
98   uint32_t e = Fetch32(s);
99   uint32_t f = Fetch32(s + len - 4);
100   uint32_t h = len;
101 
102   return fmix(Mur(f, Mur(e, Mur(d, Mur(c, Mur(b, Mur(a, h)))))));
103 }
104 
Hash32Len0to4(const char * s,size_t len)105 static uint32_t Hash32Len0to4(const char *s, size_t len) {
106   uint32_t b = 0;
107   uint32_t c = 9;
108   for (size_t i = 0; i < len; i++) {
109     signed char v = s[i];
110     b = b * c1 + v;
111     c ^= b;
112   }
113   return fmix(Mur(b, Mur(len, c)));
114 }
115 
Hash32Len5to12(const char * s,size_t len)116 static uint32_t Hash32Len5to12(const char *s, size_t len) {
117   uint32_t a = len, b = len * 5, c = 9, d = b;
118   a += Fetch32(s);
119   b += Fetch32(s + len - 4);
120   c += Fetch32(s + ((len >> 1) & 4));
121   return fmix(Mur(c, Mur(b, Mur(a, d))));
122 }
123 
CityHash32(const char * s,size_t len)124 uint32_t CityHash32(const char *s, size_t len) {
125   if (len <= 24) {
126     return len <= 12
127                ? (len <= 4 ? Hash32Len0to4(s, len) : Hash32Len5to12(s, len))
128                : Hash32Len13to24(s, len);
129   }
130 
131   // len > 24
132   uint32_t h = len, g = c1 * len, f = g;
133 
134   uint32_t a0 = Rotate32(Fetch32(s + len - 4) * c1, 17) * c2;
135   uint32_t a1 = Rotate32(Fetch32(s + len - 8) * c1, 17) * c2;
136   uint32_t a2 = Rotate32(Fetch32(s + len - 16) * c1, 17) * c2;
137   uint32_t a3 = Rotate32(Fetch32(s + len - 12) * c1, 17) * c2;
138   uint32_t a4 = Rotate32(Fetch32(s + len - 20) * c1, 17) * c2;
139   h ^= a0;
140   h = Rotate32(h, 19);
141   h = h * 5 + 0xe6546b64;
142   h ^= a2;
143   h = Rotate32(h, 19);
144   h = h * 5 + 0xe6546b64;
145   g ^= a1;
146   g = Rotate32(g, 19);
147   g = g * 5 + 0xe6546b64;
148   g ^= a3;
149   g = Rotate32(g, 19);
150   g = g * 5 + 0xe6546b64;
151   f += a4;
152   f = Rotate32(f, 19);
153   f = f * 5 + 0xe6546b64;
154   size_t iters = (len - 1) / 20;
155   do {
156     uint32_t b0 = Rotate32(Fetch32(s) * c1, 17) * c2;
157     uint32_t b1 = Fetch32(s + 4);
158     uint32_t b2 = Rotate32(Fetch32(s + 8) * c1, 17) * c2;
159     uint32_t b3 = Rotate32(Fetch32(s + 12) * c1, 17) * c2;
160     uint32_t b4 = Fetch32(s + 16);
161     h ^= b0;
162     h = Rotate32(h, 18);
163     h = h * 5 + 0xe6546b64;
164     f += b1;
165     f = Rotate32(f, 19);
166     f = f * c1;
167     g += b2;
168     g = Rotate32(g, 18);
169     g = g * 5 + 0xe6546b64;
170     h ^= b3 + b1;
171     h = Rotate32(h, 19);
172     h = h * 5 + 0xe6546b64;
173     g ^= b4;
174     g = absl::gbswap_32(g) * 5;
175     h += b4 * 5;
176     h = absl::gbswap_32(h);
177     f += b0;
178     PERMUTE3(f, h, g);
179     s += 20;
180   } while (--iters != 0);
181   g = Rotate32(g, 11) * c1;
182   g = Rotate32(g, 17) * c1;
183   f = Rotate32(f, 11) * c1;
184   f = Rotate32(f, 17) * c1;
185   h = Rotate32(h + g, 19);
186   h = h * 5 + 0xe6546b64;
187   h = Rotate32(h, 17) * c1;
188   h = Rotate32(h + f, 19);
189   h = h * 5 + 0xe6546b64;
190   h = Rotate32(h, 17) * c1;
191   return h;
192 }
193 
194 // Bitwise right rotate.  Normally this will compile to a single
195 // instruction, especially if the shift is a manifest constant.
Rotate(uint64_t val,int shift)196 static uint64_t Rotate(uint64_t val, int shift) {
197   // Avoid shifting by 64: doing so yields an undefined result.
198   return shift == 0 ? val : ((val >> shift) | (val << (64 - shift)));
199 }
200 
ShiftMix(uint64_t val)201 static uint64_t ShiftMix(uint64_t val) { return val ^ (val >> 47); }
202 
HashLen16(uint64_t u,uint64_t v)203 static uint64_t HashLen16(uint64_t u, uint64_t v) {
204   return Hash128to64(uint128(u, v));
205 }
206 
HashLen16(uint64_t u,uint64_t v,uint64_t mul)207 static uint64_t HashLen16(uint64_t u, uint64_t v, uint64_t mul) {
208   // Murmur-inspired hashing.
209   uint64_t a = (u ^ v) * mul;
210   a ^= (a >> 47);
211   uint64_t b = (v ^ a) * mul;
212   b ^= (b >> 47);
213   b *= mul;
214   return b;
215 }
216 
HashLen0to16(const char * s,size_t len)217 static uint64_t HashLen0to16(const char *s, size_t len) {
218   if (len >= 8) {
219     uint64_t mul = k2 + len * 2;
220     uint64_t a = Fetch64(s) + k2;
221     uint64_t b = Fetch64(s + len - 8);
222     uint64_t c = Rotate(b, 37) * mul + a;
223     uint64_t d = (Rotate(a, 25) + b) * mul;
224     return HashLen16(c, d, mul);
225   }
226   if (len >= 4) {
227     uint64_t mul = k2 + len * 2;
228     uint64_t a = Fetch32(s);
229     return HashLen16(len + (a << 3), Fetch32(s + len - 4), mul);
230   }
231   if (len > 0) {
232     uint8_t a = s[0];
233     uint8_t b = s[len >> 1];
234     uint8_t c = s[len - 1];
235     uint32_t y = static_cast<uint32_t>(a) + (static_cast<uint32_t>(b) << 8);
236     uint32_t z = len + (static_cast<uint32_t>(c) << 2);
237     return ShiftMix(y * k2 ^ z * k0) * k2;
238   }
239   return k2;
240 }
241 
242 // This probably works well for 16-byte strings as well, but it may be overkill
243 // in that case.
HashLen17to32(const char * s,size_t len)244 static uint64_t HashLen17to32(const char *s, size_t len) {
245   uint64_t mul = k2 + len * 2;
246   uint64_t a = Fetch64(s) * k1;
247   uint64_t b = Fetch64(s + 8);
248   uint64_t c = Fetch64(s + len - 8) * mul;
249   uint64_t d = Fetch64(s + len - 16) * k2;
250   return HashLen16(Rotate(a + b, 43) + Rotate(c, 30) + d,
251                    a + Rotate(b + k2, 18) + c, mul);
252 }
253 
254 // Return a 16-byte hash for 48 bytes.  Quick and dirty.
255 // Callers do best to use "random-looking" values for a and b.
WeakHashLen32WithSeeds(uint64_t w,uint64_t x,uint64_t y,uint64_t z,uint64_t a,uint64_t b)256 static std::pair<uint64_t, uint64_t> WeakHashLen32WithSeeds(uint64_t w, uint64_t x,
257                                                         uint64_t y, uint64_t z,
258                                                         uint64_t a, uint64_t b) {
259   a += w;
260   b = Rotate(b + a + z, 21);
261   uint64_t c = a;
262   a += x;
263   a += y;
264   b += Rotate(a, 44);
265   return std::make_pair(a + z, b + c);
266 }
267 
268 // Return a 16-byte hash for s[0] ... s[31], a, and b.  Quick and dirty.
WeakHashLen32WithSeeds(const char * s,uint64_t a,uint64_t b)269 static std::pair<uint64_t, uint64_t> WeakHashLen32WithSeeds(const char *s, uint64_t a,
270                                                         uint64_t b) {
271   return WeakHashLen32WithSeeds(Fetch64(s), Fetch64(s + 8), Fetch64(s + 16),
272                                 Fetch64(s + 24), a, b);
273 }
274 
275 // Return an 8-byte hash for 33 to 64 bytes.
HashLen33to64(const char * s,size_t len)276 static uint64_t HashLen33to64(const char *s, size_t len) {
277   uint64_t mul = k2 + len * 2;
278   uint64_t a = Fetch64(s) * k2;
279   uint64_t b = Fetch64(s + 8);
280   uint64_t c = Fetch64(s + len - 24);
281   uint64_t d = Fetch64(s + len - 32);
282   uint64_t e = Fetch64(s + 16) * k2;
283   uint64_t f = Fetch64(s + 24) * 9;
284   uint64_t g = Fetch64(s + len - 8);
285   uint64_t h = Fetch64(s + len - 16) * mul;
286   uint64_t u = Rotate(a + g, 43) + (Rotate(b, 30) + c) * 9;
287   uint64_t v = ((a + g) ^ d) + f + 1;
288   uint64_t w = absl::gbswap_64((u + v) * mul) + h;
289   uint64_t x = Rotate(e + f, 42) + c;
290   uint64_t y = (absl::gbswap_64((v + w) * mul) + g) * mul;
291   uint64_t z = e + f + c;
292   a = absl::gbswap_64((x + z) * mul + y) + b;
293   b = ShiftMix((z + a) * mul + d + h) * mul;
294   return b + x;
295 }
296 
CityHash64(const char * s,size_t len)297 uint64_t CityHash64(const char *s, size_t len) {
298   if (len <= 32) {
299     if (len <= 16) {
300       return HashLen0to16(s, len);
301     } else {
302       return HashLen17to32(s, len);
303     }
304   } else if (len <= 64) {
305     return HashLen33to64(s, len);
306   }
307 
308   // For strings over 64 bytes we hash the end first, and then as we
309   // loop we keep 56 bytes of state: v, w, x, y, and z.
310   uint64_t x = Fetch64(s + len - 40);
311   uint64_t y = Fetch64(s + len - 16) + Fetch64(s + len - 56);
312   uint64_t z = HashLen16(Fetch64(s + len - 48) + len, Fetch64(s + len - 24));
313   std::pair<uint64_t, uint64_t> v = WeakHashLen32WithSeeds(s + len - 64, len, z);
314   std::pair<uint64_t, uint64_t> w = WeakHashLen32WithSeeds(s + len - 32, y + k1, x);
315   x = x * k1 + Fetch64(s);
316 
317   // Decrease len to the nearest multiple of 64, and operate on 64-byte chunks.
318   len = (len - 1) & ~static_cast<size_t>(63);
319   do {
320     x = Rotate(x + y + v.first + Fetch64(s + 8), 37) * k1;
321     y = Rotate(y + v.second + Fetch64(s + 48), 42) * k1;
322     x ^= w.second;
323     y += v.first + Fetch64(s + 40);
324     z = Rotate(z + w.first, 33) * k1;
325     v = WeakHashLen32WithSeeds(s, v.second * k1, x + w.first);
326     w = WeakHashLen32WithSeeds(s + 32, z + w.second, y + Fetch64(s + 16));
327     std::swap(z, x);
328     s += 64;
329     len -= 64;
330   } while (len != 0);
331   return HashLen16(HashLen16(v.first, w.first) + ShiftMix(y) * k1 + z,
332                    HashLen16(v.second, w.second) + x);
333 }
334 
CityHash64WithSeed(const char * s,size_t len,uint64_t seed)335 uint64_t CityHash64WithSeed(const char *s, size_t len, uint64_t seed) {
336   return CityHash64WithSeeds(s, len, k2, seed);
337 }
338 
CityHash64WithSeeds(const char * s,size_t len,uint64_t seed0,uint64_t seed1)339 uint64_t CityHash64WithSeeds(const char *s, size_t len, uint64_t seed0,
340                            uint64_t seed1) {
341   return HashLen16(CityHash64(s, len) - seed0, seed1);
342 }
343 
344 }  // namespace hash_internal
345 ABSL_NAMESPACE_END
346 }  // namespace absl
347