2 * Copyright 2013 Facebook, Inc.
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
8 * http://www.apache.org/licenses/LICENSE-2.0
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
17 #ifndef FOLLY_BASE_HASH_H_
18 #define FOLLY_BASE_HASH_H_
25 #include "folly/SpookyHashV1.h"
26 #include "folly/SpookyHashV2.h"
29 * Various hashing functions.
32 namespace folly { namespace hash {
34 // This is a general-purpose way to create a single hash from multiple
35 // hashable objects. hash_combine_generic takes a class Hasher implementing
36 // hash<T>; hash_combine uses a default hasher StdHasher that uses std::hash.
37 // hash_combine_generic hashes each argument and combines those hashes in
38 // an order-dependent way to yield a new hash.
41 // This is the Hash128to64 function from Google's cityhash (available
42 // under the MIT License). We use it to reduce multiple 64 bit hashes
43 // into a single hash.
44 inline size_t hash_128_to_64(const size_t upper, const size_t lower) {
45 // Murmur-inspired hashing.
46 const size_t kMul = 0x9ddfea08eb382d69ULL;
47 size_t a = (lower ^ upper) * kMul;
49 size_t b = (upper ^ a) * kMul;
55 // Never used, but gcc demands it.
56 template <class Hasher>
57 inline size_t hash_combine_generic() {
61 template <class Hasher, typename T, typename... Ts>
62 size_t hash_combine_generic(const T& t, const Ts&... ts) {
63 size_t seed = Hasher::hash(t);
64 if (sizeof...(ts) == 0) {
67 size_t remainder = hash_combine_generic<Hasher>(ts...);
68 return hash_128_to_64(seed, remainder);
71 // Simply uses std::hash to hash. Note that std::hash is not guaranteed
72 // to be a very good hash function; provided std::hash doesn't collide on
73 // the individual inputs, you are fine, but that won't be true for, say,
78 static size_t hash(const T& t) {
79 return std::hash<T>()(t);
83 template <typename T, typename... Ts>
84 size_t hash_combine(const T& t, const Ts&... ts) {
85 return hash_combine_generic<StdHasher>(t, ts...);
88 //////////////////////////////////////////////////////////////////////
91 * Thomas Wang 64 bit mix hash function
94 inline uint64_t twang_mix64(uint64_t key) {
95 key = (~key) + (key << 21); // key *= (1 << 21) - 1; key -= 1;
96 key = key ^ (key >> 24);
97 key = key + (key << 3) + (key << 8); // key *= 1 + (1 << 3) + (1 << 8)
98 key = key ^ (key >> 14);
99 key = key + (key << 2) + (key << 4); // key *= 1 + (1 << 2) + (1 << 4)
100 key = key ^ (key >> 28);
101 key = key + (key << 31); // key *= 1 + (1 << 31)
106 * Inverse of twang_mix64
108 * Note that twang_unmix64 is significantly slower than twang_mix64.
111 inline uint64_t twang_unmix64(uint64_t key) {
112 // See the comments in jenkins_rev_unmix32 for an explanation as to how this
114 key *= 4611686016279904257U;
115 key ^= (key >> 28) ^ (key >> 56);
116 key *= 14933078535860113213U;
117 key ^= (key >> 14) ^ (key >> 28) ^ (key >> 42) ^ (key >> 56);
118 key *= 15244667743933553977U;
119 key ^= (key >> 24) ^ (key >> 48);
120 key = (key + 1) * 9223367638806167551U;
125 * Thomas Wang downscaling hash function
128 inline uint32_t twang_32from64(uint64_t key) {
129 key = (~key) + (key << 18);
130 key = key ^ (key >> 31);
132 key = key ^ (key >> 11);
133 key = key + (key << 6);
134 key = key ^ (key >> 22);
135 return (uint32_t) key;
139 * Robert Jenkins' reversible 32 bit mix hash function
142 inline uint32_t jenkins_rev_mix32(uint32_t key) {
143 key += (key << 12); // key *= (1 + (1 << 12))
145 key += (key << 4); // key *= (1 + (1 << 4))
147 key += (key << 10); // key *= (1 + (1 << 10))
149 // key *= (1 + (1 << 7)) * (1 + (1 << 12))
156 * Inverse of jenkins_rev_mix32
158 * Note that jenkinks_rev_unmix32 is significantly slower than
162 inline uint32_t jenkins_rev_unmix32(uint32_t key) {
163 // These are the modular multiplicative inverses (in Z_2^32) of the
164 // multiplication factors in jenkins_rev_mix32, in reverse order. They were
165 // computed using the Extended Euclidean algorithm, see
166 // http://en.wikipedia.org/wiki/Modular_multiplicative_inverse
169 // The inverse of a ^= (a >> n) is
171 // for (int i = n; i < 32; i += n) {
175 (key >> 2) ^ (key >> 4) ^ (key >> 6) ^ (key >> 8) ^
176 (key >> 10) ^ (key >> 12) ^ (key >> 14) ^ (key >> 16) ^
177 (key >> 18) ^ (key >> 20) ^ (key >> 22) ^ (key >> 24) ^
178 (key >> 26) ^ (key >> 28) ^ (key >> 30);
180 key ^= (key >> 9) ^ (key >> 18) ^ (key >> 27);
188 * Fowler / Noll / Vo (FNV) Hash
189 * http://www.isthe.com/chongo/tech/comp/fnv/
192 const uint32_t FNV_32_HASH_START = 216613626UL;
193 const uint64_t FNV_64_HASH_START = 14695981039346656037ULL;
195 inline uint32_t fnv32(const char* s,
196 uint32_t hash = FNV_32_HASH_START) {
198 hash += (hash << 1) + (hash << 4) + (hash << 7) +
199 (hash << 8) + (hash << 24);
205 inline uint32_t fnv32_buf(const void* buf,
207 uint32_t hash = FNV_32_HASH_START) {
208 const char* char_buf = reinterpret_cast<const char*>(buf);
210 for (int i = 0; i < n; ++i) {
211 hash += (hash << 1) + (hash << 4) + (hash << 7) +
212 (hash << 8) + (hash << 24);
219 inline uint32_t fnv32(const std::string& str,
220 uint64_t hash = FNV_32_HASH_START) {
221 return fnv32_buf(str.data(), str.size(), hash);
224 inline uint64_t fnv64(const char* s,
225 uint64_t hash = FNV_64_HASH_START) {
227 hash += (hash << 1) + (hash << 4) + (hash << 5) + (hash << 7) +
228 (hash << 8) + (hash << 40);
234 inline uint64_t fnv64_buf(const void* buf,
236 uint64_t hash = FNV_64_HASH_START) {
237 const char* char_buf = reinterpret_cast<const char*>(buf);
239 for (int i = 0; i < n; ++i) {
240 hash += (hash << 1) + (hash << 4) + (hash << 5) + (hash << 7) +
241 (hash << 8) + (hash << 40);
247 inline uint64_t fnv64(const std::string& str,
248 uint64_t hash = FNV_64_HASH_START) {
249 return fnv64_buf(str.data(), str.size(), hash);
253 * Paul Hsieh: http://www.azillionmonkeys.com/qed/hash.html
256 #define get16bits(d) (*((const uint16_t*) (d)))
258 inline uint32_t hsieh_hash32_buf(const void* buf, int len) {
259 const char* s = reinterpret_cast<const char*>(buf);
264 if (len <= 0 || buf == 0) {
272 for (;len > 0; len--) {
273 hash += get16bits (s);
274 tmp = (get16bits (s+2) << 11) ^ hash;
275 hash = (hash << 16) ^ tmp;
276 s += 2*sizeof (uint16_t);
280 /* Handle end cases */
283 hash += get16bits(s);
285 hash ^= s[sizeof (uint16_t)] << 18;
289 hash += get16bits(s);
299 /* Force "avalanching" of final 127 bits */
312 inline uint32_t hsieh_hash32(const char* s) {
313 return hsieh_hash32_buf(s, std::strlen(s));
316 inline uint32_t hsieh_hash32_str(const std::string& str) {
317 return hsieh_hash32_buf(str.data(), str.size());
320 //////////////////////////////////////////////////////////////////////
327 template<> struct hasher<int32_t> {
328 size_t operator()(int32_t key) const {
329 return hash::jenkins_rev_mix32(uint32_t(key));
333 template<> struct hasher<uint32_t> {
334 size_t operator()(uint32_t key) const {
335 return hash::jenkins_rev_mix32(key);
339 template<> struct hasher<int64_t> {
340 size_t operator()(int64_t key) const {
341 return hash::twang_mix64(uint64_t(key));
345 template<> struct hasher<uint64_t> {
346 size_t operator()(uint64_t key) const {
347 return hash::twang_mix64(key);
353 // Custom hash functions.
355 // Hash function for pairs. Requires default hash functions for both
356 // items in the pair.
357 template <typename T1, typename T2>
358 class hash<std::pair<T1, T2> > {
360 size_t operator()(const std::pair<T1, T2>& x) const {
361 return folly::hash::hash_combine(x.first, x.second);