city.cc
Go to the documentation of this file.
00001 // Copyright 2018 The Abseil Authors.
00002 //
00003 // Licensed under the Apache License, Version 2.0 (the "License");
00004 // you may not use this file except in compliance with the License.
00005 // You may obtain a copy of the License at
00006 //
00007 //      https://www.apache.org/licenses/LICENSE-2.0
00008 //
00009 // Unless required by applicable law or agreed to in writing, software
00010 // distributed under the License is distributed on an "AS IS" BASIS,
00011 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
00012 // See the License for the specific language governing permissions and
00013 // limitations under the License.
00014 //
00015 // This file provides CityHash64() and related functions.
00016 //
00017 // It's probably possible to create even faster hash functions by
00018 // writing a program that systematically explores some of the space of
00019 // possible hash functions, by using SIMD instructions, or by
00020 // compromising on hash quality.
00021 
00022 #include "absl/hash/internal/city.h"
00023 
00024 #include <string.h>  // for memcpy and memset
00025 #include <algorithm>
00026 
00027 #include "absl/base/config.h"
00028 #include "absl/base/internal/endian.h"
00029 #include "absl/base/internal/unaligned_access.h"
00030 #include "absl/base/optimization.h"
00031 
00032 namespace absl {
00033 namespace hash_internal {
00034 
00035 #ifdef ABSL_IS_BIG_ENDIAN
00036 #define uint32_in_expected_order(x) (absl::gbswap_32(x))
00037 #define uint64_in_expected_order(x) (absl::gbswap_64(x))
00038 #else
00039 #define uint32_in_expected_order(x) (x)
00040 #define uint64_in_expected_order(x) (x)
00041 #endif
00042 
00043 static uint64_t Fetch64(const char *p) {
00044   return uint64_in_expected_order(ABSL_INTERNAL_UNALIGNED_LOAD64(p));
00045 }
00046 
00047 static uint32_t Fetch32(const char *p) {
00048   return uint32_in_expected_order(ABSL_INTERNAL_UNALIGNED_LOAD32(p));
00049 }
00050 
00051 // Some primes between 2^63 and 2^64 for various uses.
00052 static const uint64_t k0 = 0xc3a5c85c97cb3127ULL;
00053 static const uint64_t k1 = 0xb492b66fbe98f273ULL;
00054 static const uint64_t k2 = 0x9ae16a3b2f90404fULL;
00055 
00056 // Magic numbers for 32-bit hashing.  Copied from Murmur3.
00057 static const uint32_t c1 = 0xcc9e2d51;
00058 static const uint32_t c2 = 0x1b873593;
00059 
00060 // A 32-bit to 32-bit integer hash copied from Murmur3.
00061 static uint32_t fmix(uint32_t h) {
00062   h ^= h >> 16;
00063   h *= 0x85ebca6b;
00064   h ^= h >> 13;
00065   h *= 0xc2b2ae35;
00066   h ^= h >> 16;
00067   return h;
00068 }
00069 
00070 static uint32_t Rotate32(uint32_t val, int shift) {
00071   // Avoid shifting by 32: doing so yields an undefined result.
00072   return shift == 0 ? val : ((val >> shift) | (val << (32 - shift)));
00073 }
00074 
00075 #undef PERMUTE3
00076 #define PERMUTE3(a, b, c) \
00077   do {                    \
00078     std::swap(a, b);      \
00079     std::swap(a, c);      \
00080   } while (0)
00081 
00082 static uint32_t Mur(uint32_t a, uint32_t h) {
00083   // Helper from Murmur3 for combining two 32-bit values.
00084   a *= c1;
00085   a = Rotate32(a, 17);
00086   a *= c2;
00087   h ^= a;
00088   h = Rotate32(h, 19);
00089   return h * 5 + 0xe6546b64;
00090 }
00091 
00092 static uint32_t Hash32Len13to24(const char *s, size_t len) {
00093   uint32_t a = Fetch32(s - 4 + (len >> 1));
00094   uint32_t b = Fetch32(s + 4);
00095   uint32_t c = Fetch32(s + len - 8);
00096   uint32_t d = Fetch32(s + (len >> 1));
00097   uint32_t e = Fetch32(s);
00098   uint32_t f = Fetch32(s + len - 4);
00099   uint32_t h = len;
00100 
00101   return fmix(Mur(f, Mur(e, Mur(d, Mur(c, Mur(b, Mur(a, h)))))));
00102 }
00103 
00104 static uint32_t Hash32Len0to4(const char *s, size_t len) {
00105   uint32_t b = 0;
00106   uint32_t c = 9;
00107   for (size_t i = 0; i < len; i++) {
00108     signed char v = s[i];
00109     b = b * c1 + v;
00110     c ^= b;
00111   }
00112   return fmix(Mur(b, Mur(len, c)));
00113 }
00114 
00115 static uint32_t Hash32Len5to12(const char *s, size_t len) {
00116   uint32_t a = len, b = len * 5, c = 9, d = b;
00117   a += Fetch32(s);
00118   b += Fetch32(s + len - 4);
00119   c += Fetch32(s + ((len >> 1) & 4));
00120   return fmix(Mur(c, Mur(b, Mur(a, d))));
00121 }
00122 
00123 uint32_t CityHash32(const char *s, size_t len) {
00124   if (len <= 24) {
00125     return len <= 12
00126                ? (len <= 4 ? Hash32Len0to4(s, len) : Hash32Len5to12(s, len))
00127                : Hash32Len13to24(s, len);
00128   }
00129 
00130   // len > 24
00131   uint32_t h = len, g = c1 * len, f = g;
00132 
00133   uint32_t a0 = Rotate32(Fetch32(s + len - 4) * c1, 17) * c2;
00134   uint32_t a1 = Rotate32(Fetch32(s + len - 8) * c1, 17) * c2;
00135   uint32_t a2 = Rotate32(Fetch32(s + len - 16) * c1, 17) * c2;
00136   uint32_t a3 = Rotate32(Fetch32(s + len - 12) * c1, 17) * c2;
00137   uint32_t a4 = Rotate32(Fetch32(s + len - 20) * c1, 17) * c2;
00138   h ^= a0;
00139   h = Rotate32(h, 19);
00140   h = h * 5 + 0xe6546b64;
00141   h ^= a2;
00142   h = Rotate32(h, 19);
00143   h = h * 5 + 0xe6546b64;
00144   g ^= a1;
00145   g = Rotate32(g, 19);
00146   g = g * 5 + 0xe6546b64;
00147   g ^= a3;
00148   g = Rotate32(g, 19);
00149   g = g * 5 + 0xe6546b64;
00150   f += a4;
00151   f = Rotate32(f, 19);
00152   f = f * 5 + 0xe6546b64;
00153   size_t iters = (len - 1) / 20;
00154   do {
00155     uint32_t b0 = Rotate32(Fetch32(s) * c1, 17) * c2;
00156     uint32_t b1 = Fetch32(s + 4);
00157     uint32_t b2 = Rotate32(Fetch32(s + 8) * c1, 17) * c2;
00158     uint32_t b3 = Rotate32(Fetch32(s + 12) * c1, 17) * c2;
00159     uint32_t b4 = Fetch32(s + 16);
00160     h ^= b0;
00161     h = Rotate32(h, 18);
00162     h = h * 5 + 0xe6546b64;
00163     f += b1;
00164     f = Rotate32(f, 19);
00165     f = f * c1;
00166     g += b2;
00167     g = Rotate32(g, 18);
00168     g = g * 5 + 0xe6546b64;
00169     h ^= b3 + b1;
00170     h = Rotate32(h, 19);
00171     h = h * 5 + 0xe6546b64;
00172     g ^= b4;
00173     g = absl::gbswap_32(g) * 5;
00174     h += b4 * 5;
00175     h = absl::gbswap_32(h);
00176     f += b0;
00177     PERMUTE3(f, h, g);
00178     s += 20;
00179   } while (--iters != 0);
00180   g = Rotate32(g, 11) * c1;
00181   g = Rotate32(g, 17) * c1;
00182   f = Rotate32(f, 11) * c1;
00183   f = Rotate32(f, 17) * c1;
00184   h = Rotate32(h + g, 19);
00185   h = h * 5 + 0xe6546b64;
00186   h = Rotate32(h, 17) * c1;
00187   h = Rotate32(h + f, 19);
00188   h = h * 5 + 0xe6546b64;
00189   h = Rotate32(h, 17) * c1;
00190   return h;
00191 }
00192 
00193 // Bitwise right rotate.  Normally this will compile to a single
00194 // instruction, especially if the shift is a manifest constant.
00195 static uint64_t Rotate(uint64_t val, int shift) {
00196   // Avoid shifting by 64: doing so yields an undefined result.
00197   return shift == 0 ? val : ((val >> shift) | (val << (64 - shift)));
00198 }
00199 
00200 static uint64_t ShiftMix(uint64_t val) { return val ^ (val >> 47); }
00201 
00202 static uint64_t HashLen16(uint64_t u, uint64_t v) {
00203   return Hash128to64(uint128(u, v));
00204 }
00205 
00206 static uint64_t HashLen16(uint64_t u, uint64_t v, uint64_t mul) {
00207   // Murmur-inspired hashing.
00208   uint64_t a = (u ^ v) * mul;
00209   a ^= (a >> 47);
00210   uint64_t b = (v ^ a) * mul;
00211   b ^= (b >> 47);
00212   b *= mul;
00213   return b;
00214 }
00215 
00216 static uint64_t HashLen0to16(const char *s, size_t len) {
00217   if (len >= 8) {
00218     uint64_t mul = k2 + len * 2;
00219     uint64_t a = Fetch64(s) + k2;
00220     uint64_t b = Fetch64(s + len - 8);
00221     uint64_t c = Rotate(b, 37) * mul + a;
00222     uint64_t d = (Rotate(a, 25) + b) * mul;
00223     return HashLen16(c, d, mul);
00224   }
00225   if (len >= 4) {
00226     uint64_t mul = k2 + len * 2;
00227     uint64_t a = Fetch32(s);
00228     return HashLen16(len + (a << 3), Fetch32(s + len - 4), mul);
00229   }
00230   if (len > 0) {
00231     uint8_t a = s[0];
00232     uint8_t b = s[len >> 1];
00233     uint8_t c = s[len - 1];
00234     uint32_t y = static_cast<uint32_t>(a) + (static_cast<uint32_t>(b) << 8);
00235     uint32_t z = len + (static_cast<uint32_t>(c) << 2);
00236     return ShiftMix(y * k2 ^ z * k0) * k2;
00237   }
00238   return k2;
00239 }
00240 
00241 // This probably works well for 16-byte strings as well, but it may be overkill
00242 // in that case.
00243 static uint64_t HashLen17to32(const char *s, size_t len) {
00244   uint64_t mul = k2 + len * 2;
00245   uint64_t a = Fetch64(s) * k1;
00246   uint64_t b = Fetch64(s + 8);
00247   uint64_t c = Fetch64(s + len - 8) * mul;
00248   uint64_t d = Fetch64(s + len - 16) * k2;
00249   return HashLen16(Rotate(a + b, 43) + Rotate(c, 30) + d,
00250                    a + Rotate(b + k2, 18) + c, mul);
00251 }
00252 
00253 // Return a 16-byte hash for 48 bytes.  Quick and dirty.
00254 // Callers do best to use "random-looking" values for a and b.
00255 static std::pair<uint64_t, uint64_t> WeakHashLen32WithSeeds(uint64_t w, uint64_t x,
00256                                                         uint64_t y, uint64_t z,
00257                                                         uint64_t a, uint64_t b) {
00258   a += w;
00259   b = Rotate(b + a + z, 21);
00260   uint64_t c = a;
00261   a += x;
00262   a += y;
00263   b += Rotate(a, 44);
00264   return std::make_pair(a + z, b + c);
00265 }
00266 
00267 // Return a 16-byte hash for s[0] ... s[31], a, and b.  Quick and dirty.
00268 static std::pair<uint64_t, uint64_t> WeakHashLen32WithSeeds(const char *s, uint64_t a,
00269                                                         uint64_t b) {
00270   return WeakHashLen32WithSeeds(Fetch64(s), Fetch64(s + 8), Fetch64(s + 16),
00271                                 Fetch64(s + 24), a, b);
00272 }
00273 
00274 // Return an 8-byte hash for 33 to 64 bytes.
00275 static uint64_t HashLen33to64(const char *s, size_t len) {
00276   uint64_t mul = k2 + len * 2;
00277   uint64_t a = Fetch64(s) * k2;
00278   uint64_t b = Fetch64(s + 8);
00279   uint64_t c = Fetch64(s + len - 24);
00280   uint64_t d = Fetch64(s + len - 32);
00281   uint64_t e = Fetch64(s + 16) * k2;
00282   uint64_t f = Fetch64(s + 24) * 9;
00283   uint64_t g = Fetch64(s + len - 8);
00284   uint64_t h = Fetch64(s + len - 16) * mul;
00285   uint64_t u = Rotate(a + g, 43) + (Rotate(b, 30) + c) * 9;
00286   uint64_t v = ((a + g) ^ d) + f + 1;
00287   uint64_t w = absl::gbswap_64((u + v) * mul) + h;
00288   uint64_t x = Rotate(e + f, 42) + c;
00289   uint64_t y = (absl::gbswap_64((v + w) * mul) + g) * mul;
00290   uint64_t z = e + f + c;
00291   a = absl::gbswap_64((x + z) * mul + y) + b;
00292   b = ShiftMix((z + a) * mul + d + h) * mul;
00293   return b + x;
00294 }
00295 
00296 uint64_t CityHash64(const char *s, size_t len) {
00297   if (len <= 32) {
00298     if (len <= 16) {
00299       return HashLen0to16(s, len);
00300     } else {
00301       return HashLen17to32(s, len);
00302     }
00303   } else if (len <= 64) {
00304     return HashLen33to64(s, len);
00305   }
00306 
00307   // For strings over 64 bytes we hash the end first, and then as we
00308   // loop we keep 56 bytes of state: v, w, x, y, and z.
00309   uint64_t x = Fetch64(s + len - 40);
00310   uint64_t y = Fetch64(s + len - 16) + Fetch64(s + len - 56);
00311   uint64_t z = HashLen16(Fetch64(s + len - 48) + len, Fetch64(s + len - 24));
00312   std::pair<uint64_t, uint64_t> v = WeakHashLen32WithSeeds(s + len - 64, len, z);
00313   std::pair<uint64_t, uint64_t> w = WeakHashLen32WithSeeds(s + len - 32, y + k1, x);
00314   x = x * k1 + Fetch64(s);
00315 
00316   // Decrease len to the nearest multiple of 64, and operate on 64-byte chunks.
00317   len = (len - 1) & ~static_cast<size_t>(63);
00318   do {
00319     x = Rotate(x + y + v.first + Fetch64(s + 8), 37) * k1;
00320     y = Rotate(y + v.second + Fetch64(s + 48), 42) * k1;
00321     x ^= w.second;
00322     y += v.first + Fetch64(s + 40);
00323     z = Rotate(z + w.first, 33) * k1;
00324     v = WeakHashLen32WithSeeds(s, v.second * k1, x + w.first);
00325     w = WeakHashLen32WithSeeds(s + 32, z + w.second, y + Fetch64(s + 16));
00326     std::swap(z, x);
00327     s += 64;
00328     len -= 64;
00329   } while (len != 0);
00330   return HashLen16(HashLen16(v.first, w.first) + ShiftMix(y) * k1 + z,
00331                    HashLen16(v.second, w.second) + x);
00332 }
00333 
00334 uint64_t CityHash64WithSeed(const char *s, size_t len, uint64_t seed) {
00335   return CityHash64WithSeeds(s, len, k2, seed);
00336 }
00337 
00338 uint64_t CityHash64WithSeeds(const char *s, size_t len, uint64_t seed0,
00339                            uint64_t seed1) {
00340   return HashLen16(CityHash64(s, len) - seed0, seed1);
00341 }
00342 
00343 }  // namespace hash_internal
00344 }  // namespace absl


abseil_cpp
Author(s):
autogenerated on Wed Jun 19 2019 19:42:14