2 * Copyright 2012 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 // This is version 1 of SpookyHash, incompatible with version 2.
19 // SpookyHash: a 128-bit noncryptographic hash function
20 // By Bob Jenkins, public domain
21 // Oct 31 2010: alpha, framework + SpookyHash::Mix appears right
22 // Oct 31 2011: alpha again, Mix only good to 2^^69 but rest appears right
23 // Dec 31 2011: beta, improved Mix, tested it for 2-bit deltas
24 // Feb 2 2012: production, same bits as beta
25 // Feb 5 2012: adjusted definitions of uint* to be more portable
26 // Mar 30 2012: 3 bytes/cycle, not 4. Alpha was 4 but wasn't thorough enough.
28 // Up to 3 bytes/cycle for long messages. Reasonably fast for short messages.
29 // All 1 or 2 bit deltas achieve avalanche within 1% bias per output bit.
31 // This was developed for and tested on 64-bit x86-compatible processors.
32 // It assumes the processor is little-endian. There is a macro
33 // controlling whether unaligned reads are allowed (by default they are).
34 // This should be an equally good hash on big-endian machines, but it will
35 // compute different results on them than on little-endian machines.
37 // Google's CityHash has similar specs to SpookyHash, and CityHash is faster
38 // on some platforms. MD4 and MD5 also have similar specs, but they are orders
39 // of magnitude slower. CRCs are two or more times slower, but unlike
40 // SpookyHash, they have nice math for combining the CRCs of pieces to form
41 // the CRCs of wholes. There are also cryptographic hashes, but those are even
45 #ifndef FOLLY_SPOOKYHASHV1_H_
46 #define FOLLY_SPOOKYHASHV1_H_
58 // SpookyHash: hash a single message in one call, produce 128-bit output
61 const void *message, // message to hash
62 size_t length, // length of message in bytes
63 uint64_t *hash1, // in/out: in seed 1, out hash value 1
64 uint64_t *hash2); // in/out: in seed 2, out hash value 2
67 // Hash64: hash a single message in one call, return 64-bit output
69 static uint64_t Hash64(
70 const void *message, // message to hash
71 size_t length, // length of message in bytes
72 uint64_t seed) // seed
74 uint64_t hash1 = seed;
75 Hash128(message, length, &hash1, &seed);
80 // Hash32: hash a single message in one call, produce 32-bit output
82 static uint32_t Hash32(
83 const void *message, // message to hash
84 size_t length, // length of message in bytes
85 uint32_t seed) // seed
87 uint64_t hash1 = seed, hash2 = seed;
88 Hash128(message, length, &hash1, &hash2);
89 return (uint32_t)hash1;
93 // Init: initialize the context of a SpookyHash
96 uint64_t seed1, // any 64-bit value will do, including 0
97 uint64_t seed2); // different seeds produce independent hashes
100 // Update: add a piece of a message to a SpookyHash state
103 const void *message, // message fragment
104 size_t length); // length of message fragment in bytes
108 // Final: compute the hash for the current SpookyHash state
110 // This does not modify the state; you can keep updating it afterward
112 // The result is the same as if SpookyHash() had been called with
113 // all the pieces concatenated into one message.
116 uint64_t *hash1, // out only: first 64 bits of hash value.
117 uint64_t *hash2); // out only: second 64 bits of hash value.
120 // left rotate a 64-bit value by k bytes
122 static inline uint64_t Rot64(uint64_t x, int k)
124 return (x << k) | (x >> (64 - k));
128 // This is used if the input is 96 bytes long or longer.
130 // The internal state is fully overwritten every 96 bytes.
131 // Every input bit appears to cause at least 128 bits of entropy
132 // before 96 other bytes are combined, when run forward or backward
133 // For every input bit,
134 // Two inputs differing in just that input bit
135 // Where "differ" means xor or subtraction
136 // And the base value is random
137 // When run forward or backwards one Mix
138 // I tried 3 pairs of each; they all differed by at least 212 bits.
140 static inline void Mix(
141 const uint64_t *data,
142 uint64_t &s0, uint64_t &s1, uint64_t &s2, uint64_t &s3,
143 uint64_t &s4, uint64_t &s5, uint64_t &s6, uint64_t &s7,
144 uint64_t &s8, uint64_t &s9, uint64_t &s10,uint64_t &s11)
146 s0 += data[0]; s2 ^= s10; s11 ^= s0; s0 = Rot64(s0,11); s11 += s1;
147 s1 += data[1]; s3 ^= s11; s0 ^= s1; s1 = Rot64(s1,32); s0 += s2;
148 s2 += data[2]; s4 ^= s0; s1 ^= s2; s2 = Rot64(s2,43); s1 += s3;
149 s3 += data[3]; s5 ^= s1; s2 ^= s3; s3 = Rot64(s3,31); s2 += s4;
150 s4 += data[4]; s6 ^= s2; s3 ^= s4; s4 = Rot64(s4,17); s3 += s5;
151 s5 += data[5]; s7 ^= s3; s4 ^= s5; s5 = Rot64(s5,28); s4 += s6;
152 s6 += data[6]; s8 ^= s4; s5 ^= s6; s6 = Rot64(s6,39); s5 += s7;
153 s7 += data[7]; s9 ^= s5; s6 ^= s7; s7 = Rot64(s7,57); s6 += s8;
154 s8 += data[8]; s10 ^= s6; s7 ^= s8; s8 = Rot64(s8,55); s7 += s9;
155 s9 += data[9]; s11 ^= s7; s8 ^= s9; s9 = Rot64(s9,54); s8 += s10;
156 s10 += data[10]; s0 ^= s8; s9 ^= s10; s10 = Rot64(s10,22); s9 += s11;
157 s11 += data[11]; s1 ^= s9; s10 ^= s11; s11 = Rot64(s11,46); s10 += s0;
161 // Mix all 12 inputs together so that h0, h1 are a hash of them all.
163 // For two inputs differing in just the input bits
164 // Where "differ" means xor or subtraction
165 // And the base value is random, or a counting value starting at that bit
166 // The final result will have each bit of h0, h1 flip
167 // For every input bit,
168 // with probability 50 +- .3%
169 // For every pair of input bits,
170 // with probability 50 +- 3%
172 // This does not rely on the last Mix() call having already mixed some.
173 // Two iterations was almost good enough for a 64-bit result, but a
174 // 128-bit result is reported, so End() does three iterations.
176 static inline void EndPartial(
177 uint64_t &h0, uint64_t &h1, uint64_t &h2, uint64_t &h3,
178 uint64_t &h4, uint64_t &h5, uint64_t &h6, uint64_t &h7,
179 uint64_t &h8, uint64_t &h9, uint64_t &h10,uint64_t &h11)
181 h11+= h1; h2 ^= h11; h1 = Rot64(h1,44);
182 h0 += h2; h3 ^= h0; h2 = Rot64(h2,15);
183 h1 += h3; h4 ^= h1; h3 = Rot64(h3,34);
184 h2 += h4; h5 ^= h2; h4 = Rot64(h4,21);
185 h3 += h5; h6 ^= h3; h5 = Rot64(h5,38);
186 h4 += h6; h7 ^= h4; h6 = Rot64(h6,33);
187 h5 += h7; h8 ^= h5; h7 = Rot64(h7,10);
188 h6 += h8; h9 ^= h6; h8 = Rot64(h8,13);
189 h7 += h9; h10^= h7; h9 = Rot64(h9,38);
190 h8 += h10; h11^= h8; h10= Rot64(h10,53);
191 h9 += h11; h0 ^= h9; h11= Rot64(h11,42);
192 h10+= h0; h1 ^= h10; h0 = Rot64(h0,54);
195 static inline void End(
196 uint64_t &h0, uint64_t &h1, uint64_t &h2, uint64_t &h3,
197 uint64_t &h4, uint64_t &h5, uint64_t &h6, uint64_t &h7,
198 uint64_t &h8, uint64_t &h9, uint64_t &h10,uint64_t &h11)
200 EndPartial(h0,h1,h2,h3,h4,h5,h6,h7,h8,h9,h10,h11);
201 EndPartial(h0,h1,h2,h3,h4,h5,h6,h7,h8,h9,h10,h11);
202 EndPartial(h0,h1,h2,h3,h4,h5,h6,h7,h8,h9,h10,h11);
206 // The goal is for each bit of the input to expand into 128 bits of
207 // apparent entropy before it is fully overwritten.
208 // n trials both set and cleared at least m bits of h0 h1 h2 h3
215 // when run forwards or backwards
216 // for all 1-bit and 2-bit diffs
217 // with diffs defined by either xor or subtraction
218 // with a base of all zeros plus a counter, or plus another bit, or random
220 static inline void ShortMix(uint64_t &h0, uint64_t &h1, uint64_t &h2, uint64_t &h3)
222 h2 = Rot64(h2,50); h2 += h3; h0 ^= h2;
223 h3 = Rot64(h3,52); h3 += h0; h1 ^= h3;
224 h0 = Rot64(h0,30); h0 += h1; h2 ^= h0;
225 h1 = Rot64(h1,41); h1 += h2; h3 ^= h1;
226 h2 = Rot64(h2,54); h2 += h3; h0 ^= h2;
227 h3 = Rot64(h3,48); h3 += h0; h1 ^= h3;
228 h0 = Rot64(h0,38); h0 += h1; h2 ^= h0;
229 h1 = Rot64(h1,37); h1 += h2; h3 ^= h1;
230 h2 = Rot64(h2,62); h2 += h3; h0 ^= h2;
231 h3 = Rot64(h3,34); h3 += h0; h1 ^= h3;
232 h0 = Rot64(h0,5); h0 += h1; h2 ^= h0;
233 h1 = Rot64(h1,36); h1 += h2; h3 ^= h1;
237 // Mix all 4 inputs together so that h0, h1 are a hash of them all.
239 // For two inputs differing in just the input bits
240 // Where "differ" means xor or subtraction
241 // And the base value is random, or a counting value starting at that bit
242 // The final result will have each bit of h0, h1 flip
243 // For every input bit,
244 // with probability 50 +- .3% (it is probably better than that)
245 // For every pair of input bits,
246 // with probability 50 +- .75% (the worst case is approximately that)
248 static inline void ShortEnd(uint64_t &h0, uint64_t &h1, uint64_t &h2, uint64_t &h3)
250 h3 ^= h2; h2 = Rot64(h2,15); h3 += h2;
251 h0 ^= h3; h3 = Rot64(h3,52); h0 += h3;
252 h1 ^= h0; h0 = Rot64(h0,26); h1 += h0;
253 h2 ^= h1; h1 = Rot64(h1,51); h2 += h1;
254 h3 ^= h2; h2 = Rot64(h2,28); h3 += h2;
255 h0 ^= h3; h3 = Rot64(h3,9); h0 += h3;
256 h1 ^= h0; h0 = Rot64(h0,47); h1 += h0;
257 h2 ^= h1; h1 = Rot64(h1,54); h2 += h1;
258 h3 ^= h2; h2 = Rot64(h2,32); h3 += h2;
259 h0 ^= h3; h3 = Rot64(h3,25); h0 += h3;
260 h1 ^= h0; h0 = Rot64(h0,63); h1 += h0;
266 // Short is used for messages under 192 bytes in length
267 // Short has a low startup cost, the normal mode is good for long
268 // keys, the cost crossover is at about 192 bytes. The two modes were
269 // held to the same quality bar.
272 const void *message, // message (array of bytes, not necessarily aligned)
273 size_t length, // length of message (in bytes)
274 uint64_t *hash1, // in/out: in the seed, out the hash value
275 uint64_t *hash2); // in/out: in the seed, out the hash value
277 // number of uint64_t's in internal state
278 static const size_t sc_numVars = 12;
280 // size of the internal state
281 static const size_t sc_blockSize = sc_numVars*8;
283 // size of buffer of unhashed data, in bytes
284 static const size_t sc_bufSize = 2*sc_blockSize;
287 // sc_const: a constant which:
290 // * is a not-very-regular mix of 1's and 0's
291 // * does not need any other special mathematical properties
293 static const uint64_t sc_const = 0xdeadbeefdeadbeefLL;
295 uint64_t m_data[2*sc_numVars]; // unhashed data, for partial messages
296 uint64_t m_state[sc_numVars]; // internal state of the hash
297 size_t m_length; // total length of the input so far
298 uint8_t m_remainder; // length of unhashed data stashed in m_data