2 * Copyright 2016 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.
23 #include <type_traits>
26 #include <folly/ApplyTuple.h>
27 #include <folly/Bits.h>
28 #include <folly/SpookyHashV1.h>
29 #include <folly/SpookyHashV2.h>
32 * Various hashing functions.
35 namespace folly { namespace hash {
37 // This is a general-purpose way to create a single hash from multiple
38 // hashable objects. hash_combine_generic takes a class Hasher implementing
39 // hash<T>; hash_combine uses a default hasher StdHasher that uses std::hash.
40 // hash_combine_generic hashes each argument and combines those hashes in
41 // an order-dependent way to yield a new hash.
44 // This is the Hash128to64 function from Google's cityhash (available
45 // under the MIT License). We use it to reduce multiple 64 bit hashes
46 // into a single hash.
47 inline uint64_t hash_128_to_64(const uint64_t upper, const uint64_t lower) {
48 // Murmur-inspired hashing.
49 const uint64_t kMul = 0x9ddfea08eb382d69ULL;
50 uint64_t a = (lower ^ upper) * kMul;
52 uint64_t b = (upper ^ a) * kMul;
58 // Never used, but gcc demands it.
59 template <class Hasher>
60 inline size_t hash_combine_generic() {
66 class Hash = std::hash<typename std::iterator_traits<Iter>::value_type>>
67 uint64_t hash_range(Iter begin,
70 Hash hasher = Hash()) {
71 for (; begin != end; ++begin) {
72 hash = hash_128_to_64(hash, hasher(*begin));
77 inline uint32_t twang_32from64(uint64_t key);
79 template <class Hasher, typename T, typename... Ts>
80 size_t hash_combine_generic(const T& t, const Ts&... ts) {
81 size_t seed = Hasher::hash(t);
82 if (sizeof...(ts) == 0) {
85 size_t remainder = hash_combine_generic<Hasher>(ts...);
86 /* static */ if (sizeof(size_t) == sizeof(uint32_t)) {
87 return twang_32from64((uint64_t(seed) << 32) | remainder);
89 return static_cast<size_t>(hash_128_to_64(seed, remainder));
93 // Simply uses std::hash to hash. Note that std::hash is not guaranteed
94 // to be a very good hash function; provided std::hash doesn't collide on
95 // the individual inputs, you are fine, but that won't be true for, say,
100 static size_t hash(const T& t) {
101 return std::hash<T>()(t);
105 template <typename T, typename... Ts>
106 size_t hash_combine(const T& t, const Ts&... ts) {
107 return hash_combine_generic<StdHasher>(t, ts...);
110 //////////////////////////////////////////////////////////////////////
113 * Thomas Wang 64 bit mix hash function
116 inline uint64_t twang_mix64(uint64_t key) {
117 key = (~key) + (key << 21); // key *= (1 << 21) - 1; key -= 1;
118 key = key ^ (key >> 24);
119 key = key + (key << 3) + (key << 8); // key *= 1 + (1 << 3) + (1 << 8)
120 key = key ^ (key >> 14);
121 key = key + (key << 2) + (key << 4); // key *= 1 + (1 << 2) + (1 << 4)
122 key = key ^ (key >> 28);
123 key = key + (key << 31); // key *= 1 + (1 << 31)
128 * Inverse of twang_mix64
130 * Note that twang_unmix64 is significantly slower than twang_mix64.
133 inline uint64_t twang_unmix64(uint64_t key) {
134 // See the comments in jenkins_rev_unmix32 for an explanation as to how this
136 key *= 4611686016279904257U;
137 key ^= (key >> 28) ^ (key >> 56);
138 key *= 14933078535860113213U;
139 key ^= (key >> 14) ^ (key >> 28) ^ (key >> 42) ^ (key >> 56);
140 key *= 15244667743933553977U;
141 key ^= (key >> 24) ^ (key >> 48);
142 key = (key + 1) * 9223367638806167551U;
147 * Thomas Wang downscaling hash function
150 inline uint32_t twang_32from64(uint64_t key) {
151 key = (~key) + (key << 18);
152 key = key ^ (key >> 31);
154 key = key ^ (key >> 11);
155 key = key + (key << 6);
156 key = key ^ (key >> 22);
157 return (uint32_t) key;
161 * Robert Jenkins' reversible 32 bit mix hash function
164 inline uint32_t jenkins_rev_mix32(uint32_t key) {
165 key += (key << 12); // key *= (1 + (1 << 12))
167 key += (key << 4); // key *= (1 + (1 << 4))
169 key += (key << 10); // key *= (1 + (1 << 10))
171 // key *= (1 + (1 << 7)) * (1 + (1 << 12))
178 * Inverse of jenkins_rev_mix32
180 * Note that jenkinks_rev_unmix32 is significantly slower than
184 inline uint32_t jenkins_rev_unmix32(uint32_t key) {
185 // These are the modular multiplicative inverses (in Z_2^32) of the
186 // multiplication factors in jenkins_rev_mix32, in reverse order. They were
187 // computed using the Extended Euclidean algorithm, see
188 // http://en.wikipedia.org/wiki/Modular_multiplicative_inverse
191 // The inverse of a ^= (a >> n) is
193 // for (int i = n; i < 32; i += n) {
197 (key >> 2) ^ (key >> 4) ^ (key >> 6) ^ (key >> 8) ^
198 (key >> 10) ^ (key >> 12) ^ (key >> 14) ^ (key >> 16) ^
199 (key >> 18) ^ (key >> 20) ^ (key >> 22) ^ (key >> 24) ^
200 (key >> 26) ^ (key >> 28) ^ (key >> 30);
202 key ^= (key >> 9) ^ (key >> 18) ^ (key >> 27);
210 * Fowler / Noll / Vo (FNV) Hash
211 * http://www.isthe.com/chongo/tech/comp/fnv/
214 const uint32_t FNV_32_HASH_START = 2166136261UL;
215 const uint64_t FNV_64_HASH_START = 14695981039346656037ULL;
217 inline uint32_t fnv32(const char* s,
218 uint32_t hash = FNV_32_HASH_START) {
220 hash += (hash << 1) + (hash << 4) + (hash << 7) +
221 (hash << 8) + (hash << 24);
227 inline uint32_t fnv32_buf(const void* buf,
229 uint32_t hash = FNV_32_HASH_START) {
230 // forcing signed char, since other platforms can use unsigned
231 const signed char* char_buf = reinterpret_cast<const signed char*>(buf);
233 for (size_t i = 0; i < n; ++i) {
234 hash += (hash << 1) + (hash << 4) + (hash << 7) +
235 (hash << 8) + (hash << 24);
242 inline uint32_t fnv32(const std::string& str,
243 uint32_t hash = FNV_32_HASH_START) {
244 return fnv32_buf(str.data(), str.size(), hash);
247 inline uint64_t fnv64(const char* s,
248 uint64_t hash = FNV_64_HASH_START) {
250 hash += (hash << 1) + (hash << 4) + (hash << 5) + (hash << 7) +
251 (hash << 8) + (hash << 40);
257 inline uint64_t fnv64_buf(const void* buf,
259 uint64_t hash = FNV_64_HASH_START) {
260 // forcing signed char, since other platforms can use unsigned
261 const signed char* char_buf = reinterpret_cast<const signed char*>(buf);
263 for (size_t i = 0; i < n; ++i) {
264 hash += (hash << 1) + (hash << 4) + (hash << 5) + (hash << 7) +
265 (hash << 8) + (hash << 40);
271 inline uint64_t fnv64(const std::string& str,
272 uint64_t hash = FNV_64_HASH_START) {
273 return fnv64_buf(str.data(), str.size(), hash);
277 * Paul Hsieh: http://www.azillionmonkeys.com/qed/hash.html
280 #define get16bits(d) folly::loadUnaligned<uint16_t>(d)
282 inline uint32_t hsieh_hash32_buf(const void* buf, size_t len) {
283 // forcing signed char, since other platforms can use unsigned
284 const unsigned char* s = reinterpret_cast<const unsigned char*>(buf);
285 uint32_t hash = static_cast<uint32_t>(len);
289 if (len <= 0 || buf == 0) {
297 for (;len > 0; len--) {
298 hash += get16bits (s);
299 tmp = (get16bits (s+2) << 11) ^ hash;
300 hash = (hash << 16) ^ tmp;
301 s += 2*sizeof (uint16_t);
305 /* Handle end cases */
308 hash += get16bits(s);
310 hash ^= s[sizeof (uint16_t)] << 18;
314 hash += get16bits(s);
324 /* Force "avalanching" of final 127 bits */
337 inline uint32_t hsieh_hash32(const char* s) {
338 return hsieh_hash32_buf(s, std::strlen(s));
341 inline uint32_t hsieh_hash32_str(const std::string& str) {
342 return hsieh_hash32_buf(str.data(), str.size());
345 //////////////////////////////////////////////////////////////////////
349 template<class Key, class Enable = void>
354 size_t operator()(const T& v) const {
355 return hasher<T>()(v);
358 template <class T, class... Ts>
359 size_t operator()(const T& t, const Ts&... ts) const {
360 return hash::hash_128_to_64((*this)(t), (*this)(ts...));
364 template<> struct hasher<int32_t> {
365 size_t operator()(int32_t key) const {
366 return hash::jenkins_rev_mix32(uint32_t(key));
370 template<> struct hasher<uint32_t> {
371 size_t operator()(uint32_t key) const {
372 return hash::jenkins_rev_mix32(key);
376 template<> struct hasher<int64_t> {
377 size_t operator()(int64_t key) const {
378 return static_cast<size_t>(hash::twang_mix64(uint64_t(key)));
382 template<> struct hasher<uint64_t> {
383 size_t operator()(uint64_t key) const {
384 return static_cast<size_t>(hash::twang_mix64(key));
388 template<> struct hasher<std::string> {
389 size_t operator()(const std::string& key) const {
390 return static_cast<size_t>(
391 hash::SpookyHashV2::Hash64(key.data(), key.size(), 0));
396 struct hasher<T, typename std::enable_if<std::is_enum<T>::value, void>::type> {
397 size_t operator()(T key) const {
398 return Hash()(static_cast<typename std::underlying_type<T>::type>(key));
402 template <class T1, class T2>
403 struct hasher<std::pair<T1, T2>> {
404 size_t operator()(const std::pair<T1, T2>& key) const {
405 return Hash()(key.first, key.second);
409 template <typename... Ts>
410 struct hasher<std::tuple<Ts...>> {
411 size_t operator() (const std::tuple<Ts...>& key) const {
412 return applyTuple(Hash(), key);
417 template <size_t index, typename... Ts>
419 size_t operator()(std::tuple<Ts...> const& key) const {
420 return hash::hash_combine(
421 TupleHasher<index - 1, Ts...>()(key),
422 std::get<index>(key));
427 template <typename... Ts>
428 struct TupleHasher<0, Ts...> {
429 size_t operator()(std::tuple<Ts...> const& key) const {
430 // we could do std::hash here directly, but hash_combine hides all the
431 // ugly templating implicitly
432 return hash::hash_combine(std::get<0>(key));
438 // Custom hash functions.
440 // Hash function for pairs. Requires default hash functions for both
441 // items in the pair.
442 template <typename T1, typename T2>
443 struct hash<std::pair<T1, T2> > {
445 size_t operator()(const std::pair<T1, T2>& x) const {
446 return folly::hash::hash_combine(x.first, x.second);
450 // Hash function for tuples. Requires default hash functions for all types.
451 template <typename... Ts>
452 struct hash<std::tuple<Ts...>> {
453 size_t operator()(std::tuple<Ts...> const& key) const {
455 std::tuple_size<std::tuple<Ts...>>::value - 1, // start index