michael@0: // Copyright (c) 2011 Google, Inc.
michael@0: //
michael@0: // Permission is hereby granted, free of charge, to any person obtaining a copy
michael@0: // of this software and associated documentation files (the "Software"), to deal
michael@0: // in the Software without restriction, including without limitation the rights
michael@0: // to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
michael@0: // copies of the Software, and to permit persons to whom the Software is
michael@0: // furnished to do so, subject to the following conditions:
michael@0: //
michael@0: // The above copyright notice and this permission notice shall be included in
michael@0: // all copies or substantial portions of the Software.
michael@0: //
michael@0: // THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
michael@0: // IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
michael@0: // FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
michael@0: // AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
michael@0: // LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
michael@0: // OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
michael@0: // THE SOFTWARE.
michael@0: //
michael@0: // CityHash Version 1, by Geoff Pike and Jyrki Alakuijala
michael@0: //
michael@0: // This file provides CityHash64() and related functions.
michael@0: //
michael@0: // It's probably possible to create even faster hash functions by
michael@0: // writing a program that systematically explores some of the space of
michael@0: // possible hash functions, by using SIMD instructions, or by
michael@0: // compromising on hash quality.
michael@0: 
michael@0: #include "city.h"
michael@0: 
michael@0: #include <algorithm>
michael@0: 
michael@0: using namespace std;
michael@0: 
michael@0: #define UNALIGNED_LOAD64(p) (*(const uint64*)(p))
michael@0: #define UNALIGNED_LOAD32(p) (*(const uint32*)(p))
michael@0: 
michael@0: #if !defined(LIKELY)
michael@0: #if defined(__GNUC__)
michael@0: #define LIKELY(x) (__builtin_expect(!!(x), 1))
michael@0: #else
michael@0: #define LIKELY(x) (x)
michael@0: #endif
michael@0: #endif
michael@0: 
michael@0: // Some primes between 2^63 and 2^64 for various uses.
michael@0: static const uint64 k0 = 0xc3a5c85c97cb3127;
michael@0: static const uint64 k1 = 0xb492b66fbe98f273;
michael@0: static const uint64 k2 = 0x9ae16a3b2f90404f;
michael@0: static const uint64 k3 = 0xc949d7c7509e6557;
michael@0: 
michael@0: // Bitwise right rotate.  Normally this will compile to a single
michael@0: // instruction, especially if the shift is a manifest constant.
michael@0: static uint64 Rotate(uint64 val, int shift) {
michael@0:   // Avoid shifting by 64: doing so yields an undefined result.
michael@0:   return shift == 0 ? val : ((val >> shift) | (val << (64 - shift)));
michael@0: }
michael@0: 
michael@0: // Equivalent to Rotate(), but requires the second arg to be non-zero.
michael@0: // On x86-64, and probably others, it's possible for this to compile
michael@0: // to a single instruction if both args are already in registers.
michael@0: static uint64 RotateByAtLeast1(uint64 val, int shift) {
michael@0:   return (val >> shift) | (val << (64 - shift));
michael@0: }
michael@0: 
michael@0: static uint64 ShiftMix(uint64 val) {
michael@0:   return val ^ (val >> 47);
michael@0: }
michael@0: 
michael@0: static uint64 HashLen16(uint64 u, uint64 v) {
michael@0:   return Hash128to64(uint128(u, v));
michael@0: }
michael@0: 
michael@0: static uint64 HashLen0to16(const char *s, size_t len) {
michael@0:   if (len > 8) {
michael@0:     uint64 a = UNALIGNED_LOAD64(s);
michael@0:     uint64 b = UNALIGNED_LOAD64(s + len - 8);
michael@0:     return HashLen16(a, RotateByAtLeast1(b + len, len)) ^ b;
michael@0:   }
michael@0:   if (len >= 4) {
michael@0:     uint64 a = UNALIGNED_LOAD32(s);
michael@0:     return HashLen16(len + (a << 3), UNALIGNED_LOAD32(s + len - 4));
michael@0:   }
michael@0:   if (len > 0) {
michael@0:     uint8 a = s[0];
michael@0:     uint8 b = s[len >> 1];
michael@0:     uint8 c = s[len - 1];
michael@0:     uint32 y = static_cast<uint32>(a) + (static_cast<uint32>(b) << 8);
michael@0:     uint32 z = len + (static_cast<uint32>(c) << 2);
michael@0:     return ShiftMix(y * k2 ^ z * k3) * k2;
michael@0:   }
michael@0:   return k2;
michael@0: }
michael@0: 
michael@0: // This probably works well for 16-byte strings as well, but it may be overkill
michael@0: // in that case.
michael@0: static uint64 HashLen17to32(const char *s, size_t len) {
michael@0:   uint64 a = UNALIGNED_LOAD64(s) * k1;
michael@0:   uint64 b = UNALIGNED_LOAD64(s + 8);
michael@0:   uint64 c = UNALIGNED_LOAD64(s + len - 8) * k2;
michael@0:   uint64 d = UNALIGNED_LOAD64(s + len - 16) * k0;
michael@0:   return HashLen16(Rotate(a - b, 43) + Rotate(c, 30) + d,
michael@0:                    a + Rotate(b ^ k3, 20) - c + len);
michael@0: }
michael@0: 
michael@0: // Return a 16-byte hash for 48 bytes.  Quick and dirty.
michael@0: // Callers do best to use "random-looking" values for a and b.
michael@0: static pair<uint64, uint64> WeakHashLen32WithSeeds(
michael@0:     uint64 w, uint64 x, uint64 y, uint64 z, uint64 a, uint64 b) {
michael@0:   a += w;
michael@0:   b = Rotate(b + a + z, 21);
michael@0:   uint64 c = a;
michael@0:   a += x;
michael@0:   a += y;
michael@0:   b += Rotate(a, 44);
michael@0:   return make_pair(a + z, b + c);
michael@0: }
michael@0: 
michael@0: // Return a 16-byte hash for s[0] ... s[31], a, and b.  Quick and dirty.
michael@0: static pair<uint64, uint64> WeakHashLen32WithSeeds(
michael@0:     const char* s, uint64 a, uint64 b) {
michael@0:   return WeakHashLen32WithSeeds(UNALIGNED_LOAD64(s),
michael@0:                                 UNALIGNED_LOAD64(s + 8),
michael@0:                                 UNALIGNED_LOAD64(s + 16),
michael@0:                                 UNALIGNED_LOAD64(s + 24),
michael@0:                                 a,
michael@0:                                 b);
michael@0: }
michael@0: 
michael@0: // Return an 8-byte hash for 33 to 64 bytes.
michael@0: static uint64 HashLen33to64(const char *s, size_t len) {
michael@0:   uint64 z = UNALIGNED_LOAD64(s + 24);
michael@0:   uint64 a = UNALIGNED_LOAD64(s) + (len + UNALIGNED_LOAD64(s + len - 16)) * k0;
michael@0:   uint64 b = Rotate(a + z, 52);
michael@0:   uint64 c = Rotate(a, 37);
michael@0:   a += UNALIGNED_LOAD64(s + 8);
michael@0:   c += Rotate(a, 7);
michael@0:   a += UNALIGNED_LOAD64(s + 16);
michael@0:   uint64 vf = a + z;
michael@0:   uint64 vs = b + Rotate(a, 31) + c;
michael@0:   a = UNALIGNED_LOAD64(s + 16) + UNALIGNED_LOAD64(s + len - 32);
michael@0:   z = UNALIGNED_LOAD64(s + len - 8);
michael@0:   b = Rotate(a + z, 52);
michael@0:   c = Rotate(a, 37);
michael@0:   a += UNALIGNED_LOAD64(s + len - 24);
michael@0:   c += Rotate(a, 7);
michael@0:   a += UNALIGNED_LOAD64(s + len - 16);
michael@0:   uint64 wf = a + z;
michael@0:   uint64 ws = b + Rotate(a, 31) + c;
michael@0:   uint64 r = ShiftMix((vf + ws) * k2 + (wf + vs) * k0);
michael@0:   return ShiftMix(r * k0 + vs) * k2;
michael@0: }
michael@0: 
michael@0: uint64 CityHash64(const char *s, size_t len) {
michael@0:   if (len <= 32) {
michael@0:     if (len <= 16) {
michael@0:       return HashLen0to16(s, len);
michael@0:     } else {
michael@0:       return HashLen17to32(s, len);
michael@0:     }
michael@0:   } else if (len <= 64) {
michael@0:     return HashLen33to64(s, len);
michael@0:   }
michael@0: 
michael@0:   // For strings over 64 bytes we hash the end first, and then as we
michael@0:   // loop we keep 56 bytes of state: v, w, x, y, and z.
michael@0:   uint64 x = UNALIGNED_LOAD64(s);
michael@0:   uint64 y = UNALIGNED_LOAD64(s + len - 16) ^ k1;
michael@0:   uint64 z = UNALIGNED_LOAD64(s + len - 56) ^ k0;
michael@0:   pair<uint64, uint64> v = WeakHashLen32WithSeeds(s + len - 64, len, y);
michael@0:   pair<uint64, uint64> w = WeakHashLen32WithSeeds(s + len - 32, len * k1, k0);
michael@0:   z += ShiftMix(v.second) * k1;
michael@0:   x = Rotate(z + x, 39) * k1;
michael@0:   y = Rotate(y, 33) * k1;
michael@0: 
michael@0:   // Decrease len to the nearest multiple of 64, and operate on 64-byte chunks.
michael@0:   len = (len - 1) & ~static_cast<size_t>(63);
michael@0:   do {
michael@0:     x = Rotate(x + y + v.first + UNALIGNED_LOAD64(s + 16), 37) * k1;
michael@0:     y = Rotate(y + v.second + UNALIGNED_LOAD64(s + 48), 42) * k1;
michael@0:     x ^= w.second;
michael@0:     y ^= v.first;
michael@0:     z = Rotate(z ^ w.first, 33);
michael@0:     v = WeakHashLen32WithSeeds(s, v.second * k1, x + w.first);
michael@0:     w = WeakHashLen32WithSeeds(s + 32, z + w.second, y);
michael@0:     std::swap(z, x);
michael@0:     s += 64;
michael@0:     len -= 64;
michael@0:   } while (len != 0);
michael@0:   return HashLen16(HashLen16(v.first, w.first) + ShiftMix(y) * k1 + z,
michael@0:                    HashLen16(v.second, w.second) + x);
michael@0: }
michael@0: 
michael@0: uint64 CityHash64WithSeed(const char *s, size_t len, uint64 seed) {
michael@0:   return CityHash64WithSeeds(s, len, k2, seed);
michael@0: }
michael@0: 
michael@0: uint64 CityHash64WithSeeds(const char *s, size_t len,
michael@0:                            uint64 seed0, uint64 seed1) {
michael@0:   return HashLen16(CityHash64(s, len) - seed0, seed1);
michael@0: }
michael@0: 
michael@0: // A subroutine for CityHash128().  Returns a decent 128-bit hash for strings
michael@0: // of any length representable in ssize_t.  Based on City and Murmur.
michael@0: static uint128 CityMurmur(const char *s, size_t len, uint128 seed) {
michael@0:   uint64 a = Uint128Low64(seed);
michael@0:   uint64 b = Uint128High64(seed);
michael@0:   uint64 c = 0;
michael@0:   uint64 d = 0;
michael@0:   ssize_t l = len - 16;
michael@0:   if (l <= 0) {  // len <= 16
michael@0:     c = b * k1 + HashLen0to16(s, len);
michael@0:     d = Rotate(a + (len >= 8 ? UNALIGNED_LOAD64(s) : c), 32);
michael@0:   } else {  // len > 16
michael@0:     c = HashLen16(UNALIGNED_LOAD64(s + len - 8) + k1, a);
michael@0:     d = HashLen16(b + len, c + UNALIGNED_LOAD64(s + len - 16));
michael@0:     a += d;
michael@0:     do {
michael@0:       a ^= ShiftMix(UNALIGNED_LOAD64(s) * k1) * k1;
michael@0:       a *= k1;
michael@0:       b ^= a;
michael@0:       c ^= ShiftMix(UNALIGNED_LOAD64(s + 8) * k1) * k1;
michael@0:       c *= k1;
michael@0:       d ^= c;
michael@0:       s += 16;
michael@0:       l -= 16;
michael@0:     } while (l > 0);
michael@0:   }
michael@0:   a = HashLen16(a, c);
michael@0:   b = HashLen16(d, b);
michael@0:   return uint128(a ^ b, HashLen16(b, a));
michael@0: }
michael@0: 
michael@0: uint128 CityHash128WithSeed(const char *s, size_t len, uint128 seed) {
michael@0:   if (len < 128) {
michael@0:     return CityMurmur(s, len, seed);
michael@0:   }
michael@0: 
michael@0:   // We expect len >= 128 to be the common case.  Keep 56 bytes of state:
michael@0:   // v, w, x, y, and z.
michael@0:   pair<uint64, uint64> v, w;
michael@0:   uint64 x = Uint128Low64(seed);
michael@0:   uint64 y = Uint128High64(seed);
michael@0:   uint64 z = len * k1;
michael@0:   v.first = Rotate(y ^ k1, 49) * k1 + UNALIGNED_LOAD64(s);
michael@0:   v.second = Rotate(v.first, 42) * k1 + UNALIGNED_LOAD64(s + 8);
michael@0:   w.first = Rotate(y + z, 35) * k1 + x;
michael@0:   w.second = Rotate(x + UNALIGNED_LOAD64(s + 88), 53) * k1;
michael@0: 
michael@0:   // This is the same inner loop as CityHash64(), manually unrolled.
michael@0:   do {
michael@0:     x = Rotate(x + y + v.first + UNALIGNED_LOAD64(s + 16), 37) * k1;
michael@0:     y = Rotate(y + v.second + UNALIGNED_LOAD64(s + 48), 42) * k1;
michael@0:     x ^= w.second;
michael@0:     y ^= v.first;
michael@0:     z = Rotate(z ^ w.first, 33);
michael@0:     v = WeakHashLen32WithSeeds(s, v.second * k1, x + w.first);
michael@0:     w = WeakHashLen32WithSeeds(s + 32, z + w.second, y);
michael@0:     std::swap(z, x);
michael@0:     s += 64;
michael@0:     x = Rotate(x + y + v.first + UNALIGNED_LOAD64(s + 16), 37) * k1;
michael@0:     y = Rotate(y + v.second + UNALIGNED_LOAD64(s + 48), 42) * k1;
michael@0:     x ^= w.second;
michael@0:     y ^= v.first;
michael@0:     z = Rotate(z ^ w.first, 33);
michael@0:     v = WeakHashLen32WithSeeds(s, v.second * k1, x + w.first);
michael@0:     w = WeakHashLen32WithSeeds(s + 32, z + w.second, y);
michael@0:     std::swap(z, x);
michael@0:     s += 64;
michael@0:     len -= 128;
michael@0:   } while (LIKELY(len >= 128));
michael@0:   y += Rotate(w.first, 37) * k0 + z;
michael@0:   x += Rotate(v.first + z, 49) * k0;
michael@0:   // If 0 < len < 128, hash up to 4 chunks of 32 bytes each from the end of s.
michael@0:   for (size_t tail_done = 0; tail_done < len; ) {
michael@0:     tail_done += 32;
michael@0:     y = Rotate(y - x, 42) * k0 + v.second;
michael@0:     w.first += UNALIGNED_LOAD64(s + len - tail_done + 16);
michael@0:     x = Rotate(x, 49) * k0 + w.first;
michael@0:     w.first += v.first;
michael@0:     v = WeakHashLen32WithSeeds(s + len - tail_done, v.first, v.second);
michael@0:   }
michael@0:   // At this point our 48 bytes of state should contain more than
michael@0:   // enough information for a strong 128-bit hash.  We use two
michael@0:   // different 48-byte-to-8-byte hashes to get a 16-byte final result.
michael@0:   x = HashLen16(x, v.first);
michael@0:   y = HashLen16(y, w.first);
michael@0:   return uint128(HashLen16(x + v.second, w.second) + y,
michael@0:                  HashLen16(x + w.second, y + v.second));
michael@0: }
michael@0: 
michael@0: uint128 CityHash128(const char *s, size_t len) {
michael@0:   if (len >= 16) {
michael@0:     return CityHash128WithSeed(s + 16,
michael@0:                                len - 16,
michael@0:                                uint128(UNALIGNED_LOAD64(s) ^ k3,
michael@0:                                        UNALIGNED_LOAD64(s + 8)));
michael@0:   } else if (len >= 8) {
michael@0:     return CityHash128WithSeed(NULL,
michael@0:                                0,
michael@0:                                uint128(UNALIGNED_LOAD64(s) ^ (len * k0),
michael@0:                                        UNALIGNED_LOAD64(s + len - 8) ^ k1));
michael@0:   } else {
michael@0:     return CityHash128WithSeed(s, len, uint128(k0, k1));
michael@0:   }
michael@0: }