2 * Copyright 2015 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/ApplyTuple.h>
27 #include <folly/SpookyHashV1.h>
28 #include <folly/SpookyHashV2.h>
31 * Various hashing functions.
34 namespace folly { namespace hash {
36 // This is a general-purpose way to create a single hash from multiple
37 // hashable objects. hash_combine_generic takes a class Hasher implementing
38 // hash<T>; hash_combine uses a default hasher StdHasher that uses std::hash.
39 // hash_combine_generic hashes each argument and combines those hashes in
40 // an order-dependent way to yield a new hash.
43 // This is the Hash128to64 function from Google's cityhash (available
44 // under the MIT License). We use it to reduce multiple 64 bit hashes
45 // into a single hash.
46 inline uint64_t hash_128_to_64(const uint64_t upper, const uint64_t lower) {
47 // Murmur-inspired hashing.
48 const uint64_t kMul = 0x9ddfea08eb382d69ULL;
49 uint64_t a = (lower ^ upper) * kMul;
51 uint64_t b = (upper ^ a) * kMul;
57 // Never used, but gcc demands it.
58 template <class Hasher>
59 inline size_t hash_combine_generic() {
65 class Hash = std::hash<typename std::iterator_traits<Iter>::value_type>>
66 uint64_t hash_range(Iter begin,
69 Hash hasher = Hash()) {
70 for (; begin != end; ++begin) {
71 hash = hash_128_to_64(hash, hasher(*begin));
76 template <class Hasher, typename T, typename... Ts>
77 size_t hash_combine_generic(const T& t, const Ts&... ts) {
78 size_t seed = Hasher::hash(t);
79 if (sizeof...(ts) == 0) {
82 size_t remainder = hash_combine_generic<Hasher>(ts...);
83 return hash_128_to_64(seed, remainder);
86 // Simply uses std::hash to hash. Note that std::hash is not guaranteed
87 // to be a very good hash function; provided std::hash doesn't collide on
88 // the individual inputs, you are fine, but that won't be true for, say,
93 static size_t hash(const T& t) {
94 return std::hash<T>()(t);
98 template <typename T, typename... Ts>
99 size_t hash_combine(const T& t, const Ts&... ts) {
100 return hash_combine_generic<StdHasher>(t, ts...);
103 //////////////////////////////////////////////////////////////////////
106 * Thomas Wang 64 bit mix hash function
109 inline uint64_t twang_mix64(uint64_t key) {
110 key = (~key) + (key << 21); // key *= (1 << 21) - 1; key -= 1;
111 key = key ^ (key >> 24);
112 key = key + (key << 3) + (key << 8); // key *= 1 + (1 << 3) + (1 << 8)
113 key = key ^ (key >> 14);
114 key = key + (key << 2) + (key << 4); // key *= 1 + (1 << 2) + (1 << 4)
115 key = key ^ (key >> 28);
116 key = key + (key << 31); // key *= 1 + (1 << 31)
121 * Inverse of twang_mix64
123 * Note that twang_unmix64 is significantly slower than twang_mix64.
126 inline uint64_t twang_unmix64(uint64_t key) {
127 // See the comments in jenkins_rev_unmix32 for an explanation as to how this
129 key *= 4611686016279904257U;
130 key ^= (key >> 28) ^ (key >> 56);
131 key *= 14933078535860113213U;
132 key ^= (key >> 14) ^ (key >> 28) ^ (key >> 42) ^ (key >> 56);
133 key *= 15244667743933553977U;
134 key ^= (key >> 24) ^ (key >> 48);
135 key = (key + 1) * 9223367638806167551U;
140 * Thomas Wang downscaling hash function
143 inline uint32_t twang_32from64(uint64_t key) {
144 key = (~key) + (key << 18);
145 key = key ^ (key >> 31);
147 key = key ^ (key >> 11);
148 key = key + (key << 6);
149 key = key ^ (key >> 22);
150 return (uint32_t) key;
154 * Robert Jenkins' reversible 32 bit mix hash function
157 inline uint32_t jenkins_rev_mix32(uint32_t key) {
158 key += (key << 12); // key *= (1 + (1 << 12))
160 key += (key << 4); // key *= (1 + (1 << 4))
162 key += (key << 10); // key *= (1 + (1 << 10))
164 // key *= (1 + (1 << 7)) * (1 + (1 << 12))
171 * Inverse of jenkins_rev_mix32
173 * Note that jenkinks_rev_unmix32 is significantly slower than
177 inline uint32_t jenkins_rev_unmix32(uint32_t key) {
178 // These are the modular multiplicative inverses (in Z_2^32) of the
179 // multiplication factors in jenkins_rev_mix32, in reverse order. They were
180 // computed using the Extended Euclidean algorithm, see
181 // http://en.wikipedia.org/wiki/Modular_multiplicative_inverse
184 // The inverse of a ^= (a >> n) is
186 // for (int i = n; i < 32; i += n) {
190 (key >> 2) ^ (key >> 4) ^ (key >> 6) ^ (key >> 8) ^
191 (key >> 10) ^ (key >> 12) ^ (key >> 14) ^ (key >> 16) ^
192 (key >> 18) ^ (key >> 20) ^ (key >> 22) ^ (key >> 24) ^
193 (key >> 26) ^ (key >> 28) ^ (key >> 30);
195 key ^= (key >> 9) ^ (key >> 18) ^ (key >> 27);
203 * Fowler / Noll / Vo (FNV) Hash
204 * http://www.isthe.com/chongo/tech/comp/fnv/
207 const uint32_t FNV_32_HASH_START = 2166136261UL;
208 const uint64_t FNV_64_HASH_START = 14695981039346656037ULL;
210 inline uint32_t fnv32(const char* s,
211 uint32_t hash = FNV_32_HASH_START) {
213 hash += (hash << 1) + (hash << 4) + (hash << 7) +
214 (hash << 8) + (hash << 24);
220 inline uint32_t fnv32_buf(const void* buf,
222 uint32_t hash = FNV_32_HASH_START) {
223 const char* char_buf = reinterpret_cast<const char*>(buf);
225 for (size_t i = 0; i < n; ++i) {
226 hash += (hash << 1) + (hash << 4) + (hash << 7) +
227 (hash << 8) + (hash << 24);
234 inline uint32_t fnv32(const std::string& str,
235 uint32_t hash = FNV_32_HASH_START) {
236 return fnv32_buf(str.data(), str.size(), hash);
239 inline uint64_t fnv64(const char* s,
240 uint64_t hash = FNV_64_HASH_START) {
242 hash += (hash << 1) + (hash << 4) + (hash << 5) + (hash << 7) +
243 (hash << 8) + (hash << 40);
249 inline uint64_t fnv64_buf(const void* buf,
251 uint64_t hash = FNV_64_HASH_START) {
252 const char* char_buf = reinterpret_cast<const char*>(buf);
254 for (size_t i = 0; i < n; ++i) {
255 hash += (hash << 1) + (hash << 4) + (hash << 5) + (hash << 7) +
256 (hash << 8) + (hash << 40);
262 inline uint64_t fnv64(const std::string& str,
263 uint64_t hash = FNV_64_HASH_START) {
264 return fnv64_buf(str.data(), str.size(), hash);
268 * Paul Hsieh: http://www.azillionmonkeys.com/qed/hash.html
271 #define get16bits(d) (*((const uint16_t*) (d)))
273 inline uint32_t hsieh_hash32_buf(const void* buf, size_t len) {
274 const char* s = reinterpret_cast<const char*>(buf);
275 uint32_t hash = static_cast<uint32_t>(len);
279 if (len <= 0 || buf == 0) {
287 for (;len > 0; len--) {
288 hash += get16bits (s);
289 tmp = (get16bits (s+2) << 11) ^ hash;
290 hash = (hash << 16) ^ tmp;
291 s += 2*sizeof (uint16_t);
295 /* Handle end cases */
298 hash += get16bits(s);
300 hash ^= s[sizeof (uint16_t)] << 18;
304 hash += get16bits(s);
314 /* Force "avalanching" of final 127 bits */
327 inline uint32_t hsieh_hash32(const char* s) {
328 return hsieh_hash32_buf(s, std::strlen(s));
331 inline uint32_t hsieh_hash32_str(const std::string& str) {
332 return hsieh_hash32_buf(str.data(), str.size());
335 //////////////////////////////////////////////////////////////////////
339 template<class Key, class Enable = void>
344 size_t operator()(const T& v) const {
345 return hasher<T>()(v);
348 template <class T, class... Ts>
349 size_t operator()(const T& t, const Ts&... ts) const {
350 return hash::hash_128_to_64((*this)(t), (*this)(ts...));
354 template<> struct hasher<int32_t> {
355 size_t operator()(int32_t key) const {
356 return hash::jenkins_rev_mix32(uint32_t(key));
360 template<> struct hasher<uint32_t> {
361 size_t operator()(uint32_t key) const {
362 return hash::jenkins_rev_mix32(key);
366 template<> struct hasher<int64_t> {
367 size_t operator()(int64_t key) const {
368 return hash::twang_mix64(uint64_t(key));
372 template<> struct hasher<uint64_t> {
373 size_t operator()(uint64_t key) const {
374 return hash::twang_mix64(key);
379 struct hasher<T, typename std::enable_if<std::is_enum<T>::value, void>::type> {
380 size_t operator()(T key) const {
381 return Hash()(static_cast<typename std::underlying_type<T>::type>(key));
385 template <class T1, class T2>
386 struct hasher<std::pair<T1, T2>> {
387 size_t operator()(const std::pair<T1, T2>& key) const {
388 return Hash()(key.first, key.second);
392 template <typename... Ts>
393 struct hasher<std::tuple<Ts...>> {
394 size_t operator() (const std::tuple<Ts...>& key) const {
395 return applyTuple(Hash(), key);
400 template <size_t index, typename... Ts>
402 size_t operator()(std::tuple<Ts...> const& key) const {
403 return hash::hash_combine(
404 TupleHasher<index - 1, Ts...>()(key),
405 std::get<index>(key));
410 template <typename... Ts>
411 struct TupleHasher<0, Ts...> {
412 size_t operator()(std::tuple<Ts...> const& key) const {
413 // we could do std::hash here directly, but hash_combine hides all the
414 // ugly templating implicitly
415 return hash::hash_combine(std::get<0>(key));
421 // Custom hash functions.
423 // Hash function for pairs. Requires default hash functions for both
424 // items in the pair.
425 template <typename T1, typename T2>
426 struct hash<std::pair<T1, T2> > {
428 size_t operator()(const std::pair<T1, T2>& x) const {
429 return folly::hash::hash_combine(x.first, x.second);
433 // Hash function for tuples. Requires default hash functions for all types.
434 template <typename... Ts>
435 struct hash<std::tuple<Ts...>> {
436 size_t operator()(std::tuple<Ts...> const& key) const {
438 std::tuple_size<std::tuple<Ts...>>::value - 1, // start index