| 1 | #ifndef _LINUX_HASH_H |
| 2 | #define _LINUX_HASH_H |
| 3 | |
| 4 | #include <inttypes.h> |
| 5 | #include "arch/arch.h" |
| 6 | #include "compiler/compiler.h" |
| 7 | |
| 8 | /* Fast hashing routine for a long. |
| 9 | (C) 2002 William Lee Irwin III, IBM */ |
| 10 | |
| 11 | /* |
| 12 | * Knuth recommends primes in approximately golden ratio to the maximum |
| 13 | * integer representable by a machine word for multiplicative hashing. |
| 14 | * Chuck Lever verified the effectiveness of this technique: |
| 15 | * http://www.citi.umich.edu/techreports/reports/citi-tr-00-1.pdf |
| 16 | * |
| 17 | * These primes are chosen to be bit-sparse, that is operations on |
| 18 | * them can use shifts and additions instead of multiplications for |
| 19 | * machines where multiplications are slow. |
| 20 | */ |
| 21 | |
| 22 | #if BITS_PER_LONG == 32 |
| 23 | /* 2^31 + 2^29 - 2^25 + 2^22 - 2^19 - 2^16 + 1 */ |
| 24 | #define GOLDEN_RATIO_PRIME 0x9e370001UL |
| 25 | #elif BITS_PER_LONG == 64 |
| 26 | /* 2^63 + 2^61 - 2^57 + 2^54 - 2^51 - 2^18 + 1 */ |
| 27 | #define GOLDEN_RATIO_PRIME 0x9e37fffffffc0001UL |
| 28 | #else |
| 29 | #error Define GOLDEN_RATIO_PRIME for your wordsize. |
| 30 | #endif |
| 31 | |
| 32 | /* |
| 33 | * The above primes are actively bad for hashing, since they are |
| 34 | * too sparse. The 32-bit one is mostly ok, the 64-bit one causes |
| 35 | * real problems. Besides, the "prime" part is pointless for the |
| 36 | * multiplicative hash. |
| 37 | * |
| 38 | * Although a random odd number will do, it turns out that the golden |
| 39 | * ratio phi = (sqrt(5)-1)/2, or its negative, has particularly nice |
| 40 | * properties. |
| 41 | * |
| 42 | * These are the negative, (1 - phi) = (phi^2) = (3 - sqrt(5))/2. |
| 43 | * (See Knuth vol 3, section 6.4, exercise 9.) |
| 44 | */ |
| 45 | #define GOLDEN_RATIO_32 0x61C88647 |
| 46 | #define GOLDEN_RATIO_64 0x61C8864680B583EBull |
| 47 | |
| 48 | static inline unsigned long __hash_long(uint64_t val) |
| 49 | { |
| 50 | uint64_t hash = val; |
| 51 | |
| 52 | #if BITS_PER_LONG == 64 |
| 53 | hash *= GOLDEN_RATIO_64; |
| 54 | #else |
| 55 | /* Sigh, gcc can't optimise this alone like it does for 32 bits. */ |
| 56 | uint64_t n = hash; |
| 57 | n <<= 18; |
| 58 | hash -= n; |
| 59 | n <<= 33; |
| 60 | hash -= n; |
| 61 | n <<= 3; |
| 62 | hash += n; |
| 63 | n <<= 3; |
| 64 | hash -= n; |
| 65 | n <<= 4; |
| 66 | hash += n; |
| 67 | n <<= 2; |
| 68 | hash += n; |
| 69 | #endif |
| 70 | |
| 71 | return hash; |
| 72 | } |
| 73 | |
| 74 | static inline unsigned long hash_long(unsigned long val, unsigned int bits) |
| 75 | { |
| 76 | /* High bits are more random, so use them. */ |
| 77 | return __hash_long(val) >> (BITS_PER_LONG - bits); |
| 78 | } |
| 79 | |
| 80 | static inline uint64_t __hash_u64(uint64_t val) |
| 81 | { |
| 82 | return val * GOLDEN_RATIO_64; |
| 83 | } |
| 84 | |
| 85 | static inline unsigned long hash_ptr(void *ptr, unsigned int bits) |
| 86 | { |
| 87 | return hash_long((uintptr_t)ptr, bits); |
| 88 | } |
| 89 | |
| 90 | /* |
| 91 | * Bob Jenkins jhash |
| 92 | */ |
| 93 | |
| 94 | #define JHASH_INITVAL GOLDEN_RATIO_32 |
| 95 | |
| 96 | static inline uint32_t rol32(uint32_t word, uint32_t shift) |
| 97 | { |
| 98 | return (word << shift) | (word >> (32 - shift)); |
| 99 | } |
| 100 | |
| 101 | /* __jhash_mix -- mix 3 32-bit values reversibly. */ |
| 102 | #define __jhash_mix(a, b, c) \ |
| 103 | { \ |
| 104 | a -= c; a ^= rol32(c, 4); c += b; \ |
| 105 | b -= a; b ^= rol32(a, 6); a += c; \ |
| 106 | c -= b; c ^= rol32(b, 8); b += a; \ |
| 107 | a -= c; a ^= rol32(c, 16); c += b; \ |
| 108 | b -= a; b ^= rol32(a, 19); a += c; \ |
| 109 | c -= b; c ^= rol32(b, 4); b += a; \ |
| 110 | } |
| 111 | |
| 112 | /* __jhash_final - final mixing of 3 32-bit values (a,b,c) into c */ |
| 113 | #define __jhash_final(a, b, c) \ |
| 114 | { \ |
| 115 | c ^= b; c -= rol32(b, 14); \ |
| 116 | a ^= c; a -= rol32(c, 11); \ |
| 117 | b ^= a; b -= rol32(a, 25); \ |
| 118 | c ^= b; c -= rol32(b, 16); \ |
| 119 | a ^= c; a -= rol32(c, 4); \ |
| 120 | b ^= a; b -= rol32(a, 14); \ |
| 121 | c ^= b; c -= rol32(b, 24); \ |
| 122 | } |
| 123 | |
| 124 | static inline uint32_t jhash(const void *key, uint32_t length, uint32_t initval) |
| 125 | { |
| 126 | const uint8_t *k = key; |
| 127 | uint32_t a, b, c; |
| 128 | |
| 129 | /* Set up the internal state */ |
| 130 | a = b = c = JHASH_INITVAL + length + initval; |
| 131 | |
| 132 | /* All but the last block: affect some 32 bits of (a,b,c) */ |
| 133 | while (length > 12) { |
| 134 | a += *k; |
| 135 | b += *(k + 4); |
| 136 | c += *(k + 8); |
| 137 | __jhash_mix(a, b, c); |
| 138 | length -= 12; |
| 139 | k += 12; |
| 140 | } |
| 141 | |
| 142 | /* Last block: affect all 32 bits of (c) */ |
| 143 | /* All the case statements fall through */ |
| 144 | switch (length) { |
| 145 | case 12: c += (uint32_t) k[11] << 24; fallthrough; |
| 146 | case 11: c += (uint32_t) k[10] << 16; fallthrough; |
| 147 | case 10: c += (uint32_t) k[9] << 8; fallthrough; |
| 148 | case 9: c += k[8]; fallthrough; |
| 149 | case 8: b += (uint32_t) k[7] << 24; fallthrough; |
| 150 | case 7: b += (uint32_t) k[6] << 16; fallthrough; |
| 151 | case 6: b += (uint32_t) k[5] << 8; fallthrough; |
| 152 | case 5: b += k[4]; fallthrough; |
| 153 | case 4: a += (uint32_t) k[3] << 24; fallthrough; |
| 154 | case 3: a += (uint32_t) k[2] << 16; fallthrough; |
| 155 | case 2: a += (uint32_t) k[1] << 8; fallthrough; |
| 156 | case 1: a += k[0]; |
| 157 | __jhash_final(a, b, c); |
| 158 | fallthrough; |
| 159 | case 0: /* Nothing left to add */ |
| 160 | break; |
| 161 | } |
| 162 | |
| 163 | return c; |
| 164 | } |
| 165 | |
| 166 | #endif /* _LINUX_HASH_H */ |