1 /*
2 * Copyright (C) 2018 The Android Open Source Project
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 #include "utils/hash/cityhash.h"
18
19 #include <cstdint>
20
21 #include "absl/base/internal/endian.h"
22 #include "absl/numeric/int128.h"
23
24 namespace tc3farmhash {
25 namespace {
26 // Some primes between 2^63 and 2^64 for various uses.
27 static const uint64_t k0 = 0xa5b85c5e198ed849ULL;
28 static const uint64_t k1 = 0x8d58ac26afe12e47ULL;
29 static const uint64_t k2 = 0xc47b6e9e3a970ed3ULL;
30 static const uint64_t k3 = 0xc70f6907e782aa0bULL;
31
32 // Hash 128 input bits down to 64 bits of output.
33 // This is intended to be a reasonably good hash function.
34 // It may change from time to time.
Hash128to64(const absl::uint128 x)35 inline uint64_t Hash128to64(const absl::uint128 x) {
36 // Murmur-inspired hashing.
37 const uint64_t kMul = 0xc6a4a7935bd1e995ULL;
38 uint64_t a = (absl::Uint128Low64(x) ^ absl::Uint128High64(x)) * kMul;
39 a ^= (a >> 47);
40 uint64_t b = (absl::Uint128High64(x) ^ a) * kMul;
41 b ^= (b >> 47);
42 b *= kMul;
43 return b;
44 }
45
HashLen16(uint64_t u,uint64_t v)46 uint64_t HashLen16(uint64_t u, uint64_t v) {
47 return Hash128to64(absl::MakeUint128(u, v));
48 }
49
Rotate(uint64_t val,size_t shift)50 static uint64_t Rotate(uint64_t val, size_t shift) {
51 assert(shift <= 63);
52 return (val >> shift) | (val << (-shift & 63));
53 }
54
ShiftMix(uint64_t val)55 static uint64_t ShiftMix(uint64_t val) { return val ^ (val >> 47); }
56
HashLen0to16(const char * s,size_t len)57 uint64_t HashLen0to16(const char *s, size_t len) {
58 assert(len <= 16);
59 if (len > 8) {
60 uint64_t a = absl::little_endian::Load64(s);
61 uint64_t b = absl::little_endian::Load64(s + len - 8);
62 return HashLen16(a, Rotate(b + len, len)) ^ b;
63 }
64 if (len >= 4) {
65 uint64_t a = absl::little_endian::Load32(s);
66 return HashLen16(len + (a << 3), absl::little_endian::Load32(s + len - 4));
67 }
68 if (len > 0) {
69 uint8_t a = s[0];
70 uint8_t b = s[len >> 1];
71 uint8_t c = s[len - 1];
72 uint32_t y = static_cast<uint32_t>(a) + (static_cast<uint32_t>(b) << 8);
73 uint32_t z = len + (static_cast<uint32_t>(c) << 2);
74 return ShiftMix(y * k2 ^ z * k3) * k2;
75 }
76 return k2;
77 }
78
79 // Return a 16-byte hash for 48 bytes. Quick and dirty.
80 // Callers do best to use "random-looking" values for a and b.
81 // (For more, see the code review discussion of CL 18799087.)
WeakHashLen32WithSeeds(uint64_t w,uint64_t x,uint64_t y,uint64_t z,uint64_t a,uint64_t b)82 std::pair<uint64_t, uint64_t> WeakHashLen32WithSeeds(uint64_t w, uint64_t x,
83 uint64_t y, uint64_t z,
84 uint64_t a, uint64_t b) {
85 a += w;
86 b = Rotate(b + a + z, 51);
87 uint64_t c = a;
88 a += x;
89 a += y;
90 b += Rotate(a, 23);
91 return std::make_pair(a + z, b + c);
92 }
93
94 // Return a 16-byte hash for s[0] ... s[31], a, and b. Quick and dirty.
WeakHashLen32WithSeeds(const char * s,uint64_t a,uint64_t b)95 std::pair<uint64_t, uint64_t> WeakHashLen32WithSeeds(const char *s, uint64_t a,
96 uint64_t b) {
97 return WeakHashLen32WithSeeds(absl::little_endian::Load64(s),
98 absl::little_endian::Load64(s + 8),
99 absl::little_endian::Load64(s + 16),
100 absl::little_endian::Load64(s + 24), a, b);
101 }
102
103 } // namespace
104
105 // This probably works well for 16-byte strings as well, but it may be overkill
106 // in that case.
HashLen17to32(const char * s,size_t len)107 static uint64_t HashLen17to32(const char *s, size_t len) {
108 assert(len >= 17);
109 assert(len <= 32);
110 uint64_t a = absl::little_endian::Load64(s) * k1;
111 uint64_t b = absl::little_endian::Load64(s + 8);
112 uint64_t c = absl::little_endian::Load64(s + len - 8) * k2;
113 uint64_t d = absl::little_endian::Load64(s + len - 16) * k0;
114 return HashLen16(Rotate(a - b, 43) + Rotate(c, 30) + d,
115 a + Rotate(b ^ k3, 20) - c + len);
116 }
117
118 // Return an 8-byte hash for 33 to 64 bytes.
HashLen33to64(const char * s,size_t len)119 static uint64_t HashLen33to64(const char *s, size_t len) {
120 uint64_t z = absl::little_endian::Load64(s + 24);
121 uint64_t a = absl::little_endian::Load64(s) +
122 (len + absl::little_endian::Load64(s + len - 16)) * k0;
123 uint64_t b = Rotate(a + z, 52);
124 uint64_t c = Rotate(a, 37);
125 a += absl::little_endian::Load64(s + 8);
126 c += Rotate(a, 7);
127 a += absl::little_endian::Load64(s + 16);
128 uint64_t vf = a + z;
129 uint64_t vs = b + Rotate(a, 31) + c;
130 a = absl::little_endian::Load64(s + 16) +
131 absl::little_endian::Load64(s + len - 32);
132 z += absl::little_endian::Load64(s + len - 8);
133 b = Rotate(a + z, 52);
134 c = Rotate(a, 37);
135 a += absl::little_endian::Load64(s + len - 24);
136 c += Rotate(a, 7);
137 a += absl::little_endian::Load64(s + len - 16);
138 uint64_t wf = a + z;
139 uint64_t ws = b + Rotate(a, 31) + c;
140 uint64_t r = ShiftMix((vf + ws) * k2 + (wf + vs) * k0);
141 return ShiftMix(r * k0 + vs) * k2;
142 }
143
CityHash64(const char * s,size_t len)144 uint64_t CityHash64(const char *s, size_t len) {
145 if (len <= 32) {
146 if (len <= 16) {
147 return HashLen0to16(s, len);
148 } else {
149 return HashLen17to32(s, len);
150 }
151 } else if (len <= 64) {
152 return HashLen33to64(s, len);
153 }
154
155 // For strings over 64 bytes we hash the end first, and then as we
156 // loop we keep 56 bytes of state: v, w, x, y, and z.
157 uint64_t x = absl::little_endian::Load64(s + len - 40);
158 uint64_t y = absl::little_endian::Load64(s + len - 16) +
159 absl::little_endian::Load64(s + len - 56);
160 uint64_t z = HashLen16(absl::little_endian::Load64(s + len - 48) + len,
161 absl::little_endian::Load64(s + len - 24));
162 std::pair<uint64_t, uint64_t> v =
163 WeakHashLen32WithSeeds(s + len - 64, len, z);
164 std::pair<uint64_t, uint64_t> w =
165 WeakHashLen32WithSeeds(s + len - 32, y + k1, x);
166 x = x * k1 + absl::little_endian::Load64(s);
167
168 // Decrease len to the nearest multiple of 64, and operate on 64-byte chunks.
169 len = (len - 1) & ~static_cast<size_t>(63);
170 assert(len > 0);
171 assert(len == len / 64 * 64);
172 do {
173 x = Rotate(x + y + v.first + absl::little_endian::Load64(s + 8), 37) * k1;
174 y = Rotate(y + v.second + absl::little_endian::Load64(s + 48), 42) * k1;
175 x ^= w.second;
176 y += v.first + absl::little_endian::Load64(s + 40);
177 z = Rotate(z + w.first, 33) * k1;
178 v = WeakHashLen32WithSeeds(s, v.second * k1, x + w.first);
179 w = WeakHashLen32WithSeeds(s + 32, z + w.second,
180 y + absl::little_endian::Load64(s + 16));
181 std::swap(z, x);
182 s += 64;
183 len -= 64;
184 } while (len != 0);
185 return HashLen16(HashLen16(v.first, w.first) + ShiftMix(y) * k1 + z,
186 HashLen16(v.second, w.second) + x);
187 }
188 } // namespace tc3farmhash
189