2 * Copyright 2014 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_
26 #include <folly/SpookyHashV1.h>
27 #include <folly/SpookyHashV2.h>
30 * Various hashing functions.
33 namespace folly { namespace hash {
35 // This is a general-purpose way to create a single hash from multiple
36 // hashable objects. hash_combine_generic takes a class Hasher implementing
37 // hash<T>; hash_combine uses a default hasher StdHasher that uses std::hash.
38 // hash_combine_generic hashes each argument and combines those hashes in
39 // an order-dependent way to yield a new hash.
42 // This is the Hash128to64 function from Google's cityhash (available
43 // under the MIT License). We use it to reduce multiple 64 bit hashes
44 // into a single hash.
45 inline size_t hash_128_to_64(const size_t upper, const size_t lower) {
46 // Murmur-inspired hashing.
47 const size_t kMul = 0x9ddfea08eb382d69ULL;
48 size_t a = (lower ^ upper) * kMul;
50 size_t b = (upper ^ a) * kMul;
56 // Never used, but gcc demands it.
57 template <class Hasher>
58 inline size_t hash_combine_generic() {
64 class Hash = std::hash<typename std::iterator_traits<Iter>::value_type>>
65 uint64_t hash_range(Iter begin,
68 Hash hasher = Hash()) {
69 for (; begin != end; ++begin) {
70 hash = hash_128_to_64(hash, hasher(*begin));
75 template <class Hasher, typename T, typename... Ts>
76 size_t hash_combine_generic(const T& t, const Ts&... ts) {
77 size_t seed = Hasher::hash(t);
78 if (sizeof...(ts) == 0) {
81 size_t remainder = hash_combine_generic<Hasher>(ts...);
82 return hash_128_to_64(seed, remainder);
85 // Simply uses std::hash to hash. Note that std::hash is not guaranteed
86 // to be a very good hash function; provided std::hash doesn't collide on
87 // the individual inputs, you are fine, but that won't be true for, say,
92 static size_t hash(const T& t) {
93 return std::hash<T>()(t);
97 template <typename T, typename... Ts>
98 size_t hash_combine(const T& t, const Ts&... ts) {
99 return hash_combine_generic<StdHasher>(t, ts...);
102 //////////////////////////////////////////////////////////////////////
105 * Thomas Wang 64 bit mix hash function
108 inline uint64_t twang_mix64(uint64_t key) {
109 key = (~key) + (key << 21); // key *= (1 << 21) - 1; key -= 1;
110 key = key ^ (key >> 24);
111 key = key + (key << 3) + (key << 8); // key *= 1 + (1 << 3) + (1 << 8)
112 key = key ^ (key >> 14);
113 key = key + (key << 2) + (key << 4); // key *= 1 + (1 << 2) + (1 << 4)
114 key = key ^ (key >> 28);
115 key = key + (key << 31); // key *= 1 + (1 << 31)
120 * Inverse of twang_mix64
122 * Note that twang_unmix64 is significantly slower than twang_mix64.
125 inline uint64_t twang_unmix64(uint64_t key) {
126 // See the comments in jenkins_rev_unmix32 for an explanation as to how this
128 key *= 4611686016279904257U;
129 key ^= (key >> 28) ^ (key >> 56);
130 key *= 14933078535860113213U;
131 key ^= (key >> 14) ^ (key >> 28) ^ (key >> 42) ^ (key >> 56);
132 key *= 15244667743933553977U;
133 key ^= (key >> 24) ^ (key >> 48);
134 key = (key + 1) * 9223367638806167551U;
139 * Thomas Wang downscaling hash function
142 inline uint32_t twang_32from64(uint64_t key) {
143 key = (~key) + (key << 18);
144 key = key ^ (key >> 31);
146 key = key ^ (key >> 11);
147 key = key + (key << 6);
148 key = key ^ (key >> 22);
149 return (uint32_t) key;
153 * Robert Jenkins' reversible 32 bit mix hash function
156 inline uint32_t jenkins_rev_mix32(uint32_t key) {
157 key += (key << 12); // key *= (1 + (1 << 12))
159 key += (key << 4); // key *= (1 + (1 << 4))
161 key += (key << 10); // key *= (1 + (1 << 10))
163 // key *= (1 + (1 << 7)) * (1 + (1 << 12))
170 * Inverse of jenkins_rev_mix32
172 * Note that jenkinks_rev_unmix32 is significantly slower than
176 inline uint32_t jenkins_rev_unmix32(uint32_t key) {
177 // These are the modular multiplicative inverses (in Z_2^32) of the
178 // multiplication factors in jenkins_rev_mix32, in reverse order. They were
179 // computed using the Extended Euclidean algorithm, see
180 // http://en.wikipedia.org/wiki/Modular_multiplicative_inverse
183 // The inverse of a ^= (a >> n) is
185 // for (int i = n; i < 32; i += n) {
189 (key >> 2) ^ (key >> 4) ^ (key >> 6) ^ (key >> 8) ^
190 (key >> 10) ^ (key >> 12) ^ (key >> 14) ^ (key >> 16) ^
191 (key >> 18) ^ (key >> 20) ^ (key >> 22) ^ (key >> 24) ^
192 (key >> 26) ^ (key >> 28) ^ (key >> 30);
194 key ^= (key >> 9) ^ (key >> 18) ^ (key >> 27);
202 * Fowler / Noll / Vo (FNV) Hash
203 * http://www.isthe.com/chongo/tech/comp/fnv/
206 const uint32_t FNV_32_HASH_START = 2166136261UL;
207 const uint64_t FNV_64_HASH_START = 14695981039346656037ULL;
209 inline uint32_t fnv32(const char* s,
210 uint32_t hash = FNV_32_HASH_START) {
212 hash += (hash << 1) + (hash << 4) + (hash << 7) +
213 (hash << 8) + (hash << 24);
219 inline uint32_t fnv32_buf(const void* buf,
221 uint32_t hash = FNV_32_HASH_START) {
222 const char* char_buf = reinterpret_cast<const char*>(buf);
224 for (int i = 0; i < n; ++i) {
225 hash += (hash << 1) + (hash << 4) + (hash << 7) +
226 (hash << 8) + (hash << 24);
233 inline uint32_t fnv32(const std::string& str,
234 uint32_t hash = FNV_32_HASH_START) {
235 return fnv32_buf(str.data(), str.size(), hash);
238 inline uint64_t fnv64(const char* s,
239 uint64_t hash = FNV_64_HASH_START) {
241 hash += (hash << 1) + (hash << 4) + (hash << 5) + (hash << 7) +
242 (hash << 8) + (hash << 40);
248 inline uint64_t fnv64_buf(const void* buf,
250 uint64_t hash = FNV_64_HASH_START) {
251 const char* char_buf = reinterpret_cast<const char*>(buf);
253 for (int i = 0; i < n; ++i) {
254 hash += (hash << 1) + (hash << 4) + (hash << 5) + (hash << 7) +
255 (hash << 8) + (hash << 40);
261 inline uint64_t fnv64(const std::string& str,
262 uint64_t hash = FNV_64_HASH_START) {
263 return fnv64_buf(str.data(), str.size(), hash);
267 * Paul Hsieh: http://www.azillionmonkeys.com/qed/hash.html
270 #define get16bits(d) (*((const uint16_t*) (d)))
272 inline uint32_t hsieh_hash32_buf(const void* buf, int len) {
273 const char* s = reinterpret_cast<const char*>(buf);
278 if (len <= 0 || buf == 0) {
286 for (;len > 0; len--) {
287 hash += get16bits (s);
288 tmp = (get16bits (s+2) << 11) ^ hash;
289 hash = (hash << 16) ^ tmp;
290 s += 2*sizeof (uint16_t);
294 /* Handle end cases */
297 hash += get16bits(s);
299 hash ^= s[sizeof (uint16_t)] << 18;
303 hash += get16bits(s);
313 /* Force "avalanching" of final 127 bits */
326 inline uint32_t hsieh_hash32(const char* s) {
327 return hsieh_hash32_buf(s, std::strlen(s));
330 inline uint32_t hsieh_hash32_str(const std::string& str) {
331 return hsieh_hash32_buf(str.data(), str.size());
334 //////////////////////////////////////////////////////////////////////
341 template<> struct hasher<int32_t> {
342 size_t operator()(int32_t key) const {
343 return hash::jenkins_rev_mix32(uint32_t(key));
347 template<> struct hasher<uint32_t> {
348 size_t operator()(uint32_t key) const {
349 return hash::jenkins_rev_mix32(key);
353 template<> struct hasher<int64_t> {
354 size_t operator()(int64_t key) const {
355 return hash::twang_mix64(uint64_t(key));
359 template<> struct hasher<uint64_t> {
360 size_t operator()(uint64_t key) const {
361 return hash::twang_mix64(key);
366 template <size_t index, typename... Ts>
368 size_t operator()(std::tuple<Ts...> const& key) const {
369 return hash::hash_combine(
370 TupleHasher<index - 1, Ts...>()(key),
371 std::get<index>(key));
376 template <typename... Ts>
377 struct TupleHasher<0, Ts...> {
378 size_t operator()(std::tuple<Ts...> const& key) const {
379 // we could do std::hash here directly, but hash_combine hides all the
380 // ugly templating implicitly
381 return hash::hash_combine(std::get<0>(key));
387 // Custom hash functions.
389 // Hash function for pairs. Requires default hash functions for both
390 // items in the pair.
391 template <typename T1, typename T2>
392 class hash<std::pair<T1, T2> > {
394 size_t operator()(const std::pair<T1, T2>& x) const {
395 return folly::hash::hash_combine(x.first, x.second);
399 // Hash function for tuples. Requires default hash functions for all types.
400 template <typename... Ts>
401 struct hash<std::tuple<Ts...>> {
402 size_t operator()(std::tuple<Ts...> const& key) const {
404 std::tuple_size<std::tuple<Ts...>>::value - 1, // start index