2 * Copyright (c) 2008, 2009, 2010, 2012, 2013, 2014 Nicira, 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.
29 static inline uint32_t
30 hash_rot(uint32_t x, int k)
32 return (x << k) | (x >> (32 - k));
35 uint32_t hash_bytes(const void *, size_t n_bytes, uint32_t basis);
36 void hash_bytes128(const void *_, size_t n_bytes, uint32_t basis,
39 static inline uint32_t hash_int(uint32_t x, uint32_t basis);
40 static inline uint32_t hash_2words(uint32_t, uint32_t);
41 static inline uint32_t hash_uint64(const uint64_t);
42 static inline uint32_t hash_uint64_basis(const uint64_t x,
43 const uint32_t basis);
44 uint32_t hash_3words(uint32_t, uint32_t, uint32_t);
46 static inline uint32_t hash_boolean(bool x, uint32_t basis);
47 uint32_t hash_double(double, uint32_t basis);
49 static inline uint32_t hash_pointer(const void *, uint32_t basis);
50 static inline uint32_t hash_string(const char *, uint32_t basis);
52 /* Murmurhash by Austin Appleby,
53 * from http://code.google.com/p/smhasher/source/browse/trunk/MurmurHash3.cpp.
55 * The upstream license there says:
57 * // MurmurHash3 was written by Austin Appleby, and is placed in the public
58 * // domain. The author hereby disclaims copyright to this source code.
60 * See hash_words() for sample usage. */
62 static inline uint32_t mhash_add__(uint32_t hash, uint32_t data)
65 data = hash_rot(data, 15);
70 static inline uint32_t mhash_add(uint32_t hash, uint32_t data)
72 hash = mhash_add__(hash, data);
73 hash = hash_rot(hash, 13);
74 return hash * 5 + 0xe6546b64;
77 static inline uint32_t mhash_finish(uint32_t hash)
87 #if !(defined(__SSE4_2__) && defined(__x86_64__))
88 /* Mhash-based implementation. */
90 static inline uint32_t hash_add(uint32_t hash, uint32_t data)
92 return mhash_add(hash, data);
95 static inline uint32_t hash_finish(uint32_t hash, uint32_t final)
97 return mhash_finish(hash ^ final);
100 /* Returns the hash of the 'n' 32-bit words at 'p', starting from 'basis'.
101 * 'p' must be properly aligned.
103 * This is inlined for the compiler to have access to the 'n_words', which
104 * in many cases is a constant. */
105 static inline uint32_t
106 hash_words_inline(const uint32_t p[], size_t n_words, uint32_t basis)
112 for (i = 0; i < n_words; i++) {
113 hash = hash_add(hash, p[i]);
115 return hash_finish(hash, n_words * 4);
118 static inline uint32_t
119 hash_words64_inline(const uint64_t p[], size_t n_words, uint64_t basis)
121 return hash_words_inline((uint32_t *)p, n_words * 2,
122 (uint32_t)basis ^ basis >> 32);
125 static inline uint32_t hash_pointer(const void *p, uint32_t basis)
127 /* Often pointers are hashed simply by casting to integer type, but that
128 * has pitfalls since the lower bits of a pointer are often all 0 for
129 * alignment reasons. It's hard to guess where the entropy really is, so
130 * we give up here and just use a high-quality hash function.
132 * The double cast suppresses a warning on 64-bit systems about casting to
133 * an integer to different size. That's OK in this case, since most of the
134 * entropy in the pointer is almost certainly in the lower 32 bits. */
135 return hash_int((uint32_t) (uintptr_t) p, basis);
138 static inline uint32_t hash_2words(uint32_t x, uint32_t y)
140 return hash_finish(hash_add(hash_add(x, 0), y), 8);
143 static inline uint32_t hash_uint64(const uint64_t x)
145 return hash_2words((uint32_t)(x >> 32), (uint32_t)x);
148 static inline uint32_t hash_uint64_basis(const uint64_t x,
149 const uint32_t basis)
151 return hash_3words((uint32_t)(x >> 32), (uint32_t)x, basis);
154 #else /* __SSE4_2__ && __x86_64__ */
155 #include <smmintrin.h>
157 static inline uint32_t hash_add(uint32_t hash, uint32_t data)
159 return _mm_crc32_u32(hash, data);
162 static inline uint32_t hash_finish(uint64_t hash, uint64_t final)
164 /* The finishing multiplier 0x805204f3 has been experimentally
165 * derived to pass the testsuite hash tests. */
166 hash = _mm_crc32_u64(hash, final) * 0x805204f3;
167 return hash ^ (uint32_t)hash >> 16; /* Increase entropy in LSBs. */
170 /* Returns the hash of the 'n' 32-bit words at 'p_', starting from 'basis'.
171 * We access 'p_' as a uint64_t pointer, which is fine for __SSE_4_2__.
173 * This is inlined for the compiler to have access to the 'n_words', which
174 * in many cases is a constant. */
175 static inline uint32_t
176 hash_words_inline(const uint32_t p_[], size_t n_words, uint32_t basis)
178 const uint64_t *p = (const void *)p_;
179 uint64_t hash1 = basis;
181 uint64_t hash3 = n_words;
182 const uint32_t *endp = (const uint32_t *)p + n_words;
183 const uint64_t *limit = p + n_words / 2 - 3;
186 hash1 = _mm_crc32_u64(hash1, p[0]);
187 hash2 = _mm_crc32_u64(hash2, p[1]);
188 hash3 = _mm_crc32_u64(hash3, p[2]);
191 switch (endp - (const uint32_t *)p) {
193 hash1 = _mm_crc32_u32(hash1, *(const uint32_t *)&p[0]);
196 hash1 = _mm_crc32_u64(hash1, p[0]);
199 hash1 = _mm_crc32_u64(hash1, p[0]);
200 hash2 = _mm_crc32_u32(hash2, *(const uint32_t *)&p[1]);
203 hash1 = _mm_crc32_u64(hash1, p[0]);
204 hash2 = _mm_crc32_u64(hash2, p[1]);
207 hash1 = _mm_crc32_u64(hash1, p[0]);
208 hash2 = _mm_crc32_u64(hash2, p[1]);
209 hash3 = _mm_crc32_u32(hash3, *(const uint32_t *)&p[2]);
212 return hash_finish(hash1, hash2 << 32 | hash3);
215 /* A simpler version for 64-bit data.
216 * 'n_words' is the count of 64-bit words, basis is 64 bits. */
217 static inline uint32_t
218 hash_words64_inline(const uint64_t p[], size_t n_words, uint64_t basis)
220 uint64_t hash1 = (uint32_t)basis;
221 uint64_t hash2 = basis >> 32;
222 uint64_t hash3 = n_words;
223 const uint64_t *endp = p + n_words;
224 const uint64_t *limit = endp - 3;
227 hash1 = _mm_crc32_u64(hash1, p[0]);
228 hash2 = _mm_crc32_u64(hash2, p[1]);
229 hash3 = _mm_crc32_u64(hash3, p[2]);
234 hash1 = _mm_crc32_u64(hash1, p[0]);
237 hash1 = _mm_crc32_u64(hash1, p[0]);
238 hash2 = _mm_crc32_u64(hash2, p[1]);
241 return hash_finish(hash1, hash2 << 32 | hash3);
244 static inline uint32_t hash_uint64_basis(const uint64_t x,
245 const uint32_t basis)
247 /* '23' chosen to mix bits enough for the test-hash to pass. */
248 return hash_finish(_mm_crc32_u64(basis, x), 23);
251 static inline uint32_t hash_uint64(const uint64_t x)
253 return hash_uint64_basis(x, 0);
256 static inline uint32_t hash_2words(uint32_t x, uint32_t y)
258 return hash_uint64((uint64_t)y << 32 | x);
261 static inline uint32_t hash_pointer(const void *p, uint32_t basis)
263 return hash_uint64_basis((uint64_t) (uintptr_t) p, basis);
267 uint32_t hash_words__(const uint32_t p[], size_t n_words, uint32_t basis);
268 uint32_t hash_words64__(const uint64_t p[], size_t n_words, uint64_t basis);
270 /* Inline the larger hash functions only when 'n_words' is known to be
271 * compile-time constant. */
273 static inline uint32_t
274 hash_words(const uint32_t p[], size_t n_words, uint32_t basis)
276 if (__builtin_constant_p(n_words)) {
277 return hash_words_inline(p, n_words, basis);
279 return hash_words__(p, n_words, basis);
283 static inline uint32_t
284 hash_words64(const uint64_t p[], size_t n_words, uint64_t basis)
286 if (__builtin_constant_p(n_words)) {
287 return hash_words64_inline(p, n_words, basis);
289 return hash_words64__(p, n_words, basis);
295 static inline uint32_t
296 hash_words(const uint32_t p[], size_t n_words, uint32_t basis)
298 return hash_words__(p, n_words, basis);
301 static inline uint32_t
302 hash_words64(const uint64_t p[], size_t n_words, uint64_t basis)
304 return hash_words64__(p, n_words, basis);
308 static inline uint32_t hash_string(const char *s, uint32_t basis)
310 return hash_bytes(s, strlen(s), basis);
313 static inline uint32_t hash_int(uint32_t x, uint32_t basis)
315 return hash_2words(x, basis);
318 /* An attempt at a useful 1-bit hash function. Has not been analyzed for
320 static inline uint32_t hash_boolean(bool x, uint32_t basis)
322 const uint32_t P0 = 0xc2b73583; /* This is hash_int(1, 0). */
323 const uint32_t P1 = 0xe90f1258; /* This is hash_int(2, 0). */
324 return (x ? P0 : P1) ^ hash_rot(basis, 1);