1/*
2 * Copyright 2017-present Facebook, Inc.
3 *
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
7 *
8 * http://www.apache.org/licenses/LICENSE-2.0
9 *
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.
15 */
16
17// This is version 1 of SpookyHash, incompatible with version 2.
18//
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.
27//
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.
30//
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.
36//
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
42// slower than MD5.
43//
44
45#pragma once
46
47#include <cstddef>
48#include <cstdint>
49
50namespace folly {
51namespace hash {
52
53// clang-format off
54
55class SpookyHashV1
56{
57public:
58 //
59 // SpookyHash: hash a single message in one call, produce 128-bit output
60 //
61 static void Hash128(
62 const void *message, // message to hash
63 size_t length, // length of message in bytes
64 uint64_t *hash1, // in/out: in seed 1, out hash value 1
65 uint64_t *hash2); // in/out: in seed 2, out hash value 2
66
67 //
68 // Hash64: hash a single message in one call, return 64-bit output
69 //
70 static uint64_t Hash64(
71 const void *message, // message to hash
72 size_t length, // length of message in bytes
73 uint64_t seed) // seed
74 {
75 uint64_t hash1 = seed;
76 Hash128(message, length, &hash1, &seed);
77 return hash1;
78 }
79
80 //
81 // Hash32: hash a single message in one call, produce 32-bit output
82 //
83 static uint32_t Hash32(
84 const void *message, // message to hash
85 size_t length, // length of message in bytes
86 uint32_t seed) // seed
87 {
88 uint64_t hash1 = seed, hash2 = seed;
89 Hash128(message, length, &hash1, &hash2);
90 return (uint32_t)hash1;
91 }
92
93 //
94 // Init: initialize the context of a SpookyHash
95 //
96 void Init(
97 uint64_t seed1, // any 64-bit value will do, including 0
98 uint64_t seed2); // different seeds produce independent hashes
99
100 //
101 // Update: add a piece of a message to a SpookyHash state
102 //
103 void Update(
104 const void *message, // message fragment
105 size_t length); // length of message fragment in bytes
106
107
108 //
109 // Final: compute the hash for the current SpookyHash state
110 //
111 // This does not modify the state; you can keep updating it afterward
112 //
113 // The result is the same as if SpookyHash() had been called with
114 // all the pieces concatenated into one message.
115 //
116 void Final(
117 uint64_t *hash1, // out only: first 64 bits of hash value.
118 uint64_t *hash2); // out only: second 64 bits of hash value.
119
120 //
121 // left rotate a 64-bit value by k bytes
122 //
123 static inline uint64_t Rot64(uint64_t x, int k)
124 {
125 return (x << k) | (x >> (64 - k));
126 }
127
128 //
129 // This is used if the input is 96 bytes long or longer.
130 //
131 // The internal state is fully overwritten every 96 bytes.
132 // Every input bit appears to cause at least 128 bits of entropy
133 // before 96 other bytes are combined, when run forward or backward
134 // For every input bit,
135 // Two inputs differing in just that input bit
136 // Where "differ" means xor or subtraction
137 // And the base value is random
138 // When run forward or backwards one Mix
139 // I tried 3 pairs of each; they all differed by at least 212 bits.
140 //
141 static inline void Mix(
142 const uint64_t *data,
143 uint64_t &s0, uint64_t &s1, uint64_t &s2, uint64_t &s3,
144 uint64_t &s4, uint64_t &s5, uint64_t &s6, uint64_t &s7,
145 uint64_t &s8, uint64_t &s9, uint64_t &s10,uint64_t &s11)
146 {
147 s0 += data[0]; s2 ^= s10; s11 ^= s0; s0 = Rot64(s0,11); s11 += s1;
148 s1 += data[1]; s3 ^= s11; s0 ^= s1; s1 = Rot64(s1,32); s0 += s2;
149 s2 += data[2]; s4 ^= s0; s1 ^= s2; s2 = Rot64(s2,43); s1 += s3;
150 s3 += data[3]; s5 ^= s1; s2 ^= s3; s3 = Rot64(s3,31); s2 += s4;
151 s4 += data[4]; s6 ^= s2; s3 ^= s4; s4 = Rot64(s4,17); s3 += s5;
152 s5 += data[5]; s7 ^= s3; s4 ^= s5; s5 = Rot64(s5,28); s4 += s6;
153 s6 += data[6]; s8 ^= s4; s5 ^= s6; s6 = Rot64(s6,39); s5 += s7;
154 s7 += data[7]; s9 ^= s5; s6 ^= s7; s7 = Rot64(s7,57); s6 += s8;
155 s8 += data[8]; s10 ^= s6; s7 ^= s8; s8 = Rot64(s8,55); s7 += s9;
156 s9 += data[9]; s11 ^= s7; s8 ^= s9; s9 = Rot64(s9,54); s8 += s10;
157 s10 += data[10]; s0 ^= s8; s9 ^= s10; s10 = Rot64(s10,22); s9 += s11;
158 s11 += data[11]; s1 ^= s9; s10 ^= s11; s11 = Rot64(s11,46); s10 += s0;
159 }
160
161 //
162 // Mix all 12 inputs together so that h0, h1 are a hash of them all.
163 //
164 // For two inputs differing in just the input bits
165 // Where "differ" means xor or subtraction
166 // And the base value is random, or a counting value starting at that bit
167 // The final result will have each bit of h0, h1 flip
168 // For every input bit,
169 // with probability 50 +- .3%
170 // For every pair of input bits,
171 // with probability 50 +- 3%
172 //
173 // This does not rely on the last Mix() call having already mixed some.
174 // Two iterations was almost good enough for a 64-bit result, but a
175 // 128-bit result is reported, so End() does three iterations.
176 //
177 static inline void EndPartial(
178 uint64_t &h0, uint64_t &h1, uint64_t &h2, uint64_t &h3,
179 uint64_t &h4, uint64_t &h5, uint64_t &h6, uint64_t &h7,
180 uint64_t &h8, uint64_t &h9, uint64_t &h10,uint64_t &h11)
181 {
182 h11+= h1; h2 ^= h11; h1 = Rot64(h1,44);
183 h0 += h2; h3 ^= h0; h2 = Rot64(h2,15);
184 h1 += h3; h4 ^= h1; h3 = Rot64(h3,34);
185 h2 += h4; h5 ^= h2; h4 = Rot64(h4,21);
186 h3 += h5; h6 ^= h3; h5 = Rot64(h5,38);
187 h4 += h6; h7 ^= h4; h6 = Rot64(h6,33);
188 h5 += h7; h8 ^= h5; h7 = Rot64(h7,10);
189 h6 += h8; h9 ^= h6; h8 = Rot64(h8,13);
190 h7 += h9; h10^= h7; h9 = Rot64(h9,38);
191 h8 += h10; h11^= h8; h10= Rot64(h10,53);
192 h9 += h11; h0 ^= h9; h11= Rot64(h11,42);
193 h10+= h0; h1 ^= h10; h0 = Rot64(h0,54);
194 }
195
196 static inline void End(
197 uint64_t &h0, uint64_t &h1, uint64_t &h2, uint64_t &h3,
198 uint64_t &h4, uint64_t &h5, uint64_t &h6, uint64_t &h7,
199 uint64_t &h8, uint64_t &h9, uint64_t &h10,uint64_t &h11)
200 {
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);
203 EndPartial(h0,h1,h2,h3,h4,h5,h6,h7,h8,h9,h10,h11);
204 }
205
206 //
207 // The goal is for each bit of the input to expand into 128 bits of
208 // apparent entropy before it is fully overwritten.
209 // n trials both set and cleared at least m bits of h0 h1 h2 h3
210 // n: 2 m: 29
211 // n: 3 m: 46
212 // n: 4 m: 57
213 // n: 5 m: 107
214 // n: 6 m: 146
215 // n: 7 m: 152
216 // when run forwards or backwards
217 // for all 1-bit and 2-bit diffs
218 // with diffs defined by either xor or subtraction
219 // with a base of all zeros plus a counter, or plus another bit, or random
220 //
221 static inline void ShortMix(uint64_t &h0, uint64_t &h1,
222 uint64_t &h2, uint64_t &h3)
223 {
224 h2 = Rot64(h2,50); h2 += h3; h0 ^= h2;
225 h3 = Rot64(h3,52); h3 += h0; h1 ^= h3;
226 h0 = Rot64(h0,30); h0 += h1; h2 ^= h0;
227 h1 = Rot64(h1,41); h1 += h2; h3 ^= h1;
228 h2 = Rot64(h2,54); h2 += h3; h0 ^= h2;
229 h3 = Rot64(h3,48); h3 += h0; h1 ^= h3;
230 h0 = Rot64(h0,38); h0 += h1; h2 ^= h0;
231 h1 = Rot64(h1,37); h1 += h2; h3 ^= h1;
232 h2 = Rot64(h2,62); h2 += h3; h0 ^= h2;
233 h3 = Rot64(h3,34); h3 += h0; h1 ^= h3;
234 h0 = Rot64(h0,5); h0 += h1; h2 ^= h0;
235 h1 = Rot64(h1,36); h1 += h2; h3 ^= h1;
236 }
237
238 //
239 // Mix all 4 inputs together so that h0, h1 are a hash of them all.
240 //
241 // For two inputs differing in just the input bits
242 // Where "differ" means xor or subtraction
243 // And the base value is random, or a counting value starting at that bit
244 // The final result will have each bit of h0, h1 flip
245 // For every input bit,
246 // with probability 50 +- .3% (it is probably better than that)
247 // For every pair of input bits,
248 // with probability 50 +- .75% (the worst case is approximately that)
249 //
250 static inline void ShortEnd(uint64_t &h0, uint64_t &h1,
251 uint64_t &h2, uint64_t &h3)
252 {
253 h3 ^= h2; h2 = Rot64(h2,15); h3 += h2;
254 h0 ^= h3; h3 = Rot64(h3,52); h0 += h3;
255 h1 ^= h0; h0 = Rot64(h0,26); h1 += h0;
256 h2 ^= h1; h1 = Rot64(h1,51); h2 += h1;
257 h3 ^= h2; h2 = Rot64(h2,28); h3 += h2;
258 h0 ^= h3; h3 = Rot64(h3,9); h0 += h3;
259 h1 ^= h0; h0 = Rot64(h0,47); h1 += h0;
260 h2 ^= h1; h1 = Rot64(h1,54); h2 += h1;
261 h3 ^= h2; h2 = Rot64(h2,32); h3 += h2;
262 h0 ^= h3; h3 = Rot64(h3,25); h0 += h3;
263 h1 ^= h0; h0 = Rot64(h0,63); h1 += h0;
264 }
265
266private:
267
268 //
269 // Short is used for messages under 192 bytes in length
270 // Short has a low startup cost, the normal mode is good for long
271 // keys, the cost crossover is at about 192 bytes. The two modes were
272 // held to the same quality bar.
273 //
274 static void Short(
275 const void *message, // message (byte array, not necessarily aligned)
276 size_t length, // length of message (in bytes)
277 uint64_t *hash1, // in/out: in the seed, out the hash value
278 uint64_t *hash2); // in/out: in the seed, out the hash value
279
280 // number of uint64_t's in internal state
281 static const size_t sc_numVars = 12;
282
283 // size of the internal state
284 static const size_t sc_blockSize = sc_numVars*8;
285
286 // size of buffer of unhashed data, in bytes
287 static const size_t sc_bufSize = 2*sc_blockSize;
288
289 //
290 // sc_const: a constant which:
291 // * is not zero
292 // * is odd
293 // * is a not-very-regular mix of 1's and 0's
294 // * does not need any other special mathematical properties
295 //
296 static const uint64_t sc_const = 0xdeadbeefdeadbeefULL;
297
298 uint64_t m_data[2*sc_numVars]; // unhashed data, for partial messages
299 uint64_t m_state[sc_numVars]; // internal state of the hash
300 size_t m_length; // total length of the input so far
301 uint8_t m_remainder; // length of unhashed data stashed in m_data
302};
303
304// clang-format on
305
306} // namespace hash
307} // namespace folly
308