zipf/pareto: mix blocks with hashing
[fio.git] / hash.h
CommitLineData
bdc7211e
JA
1#ifndef _LINUX_HASH_H
2#define _LINUX_HASH_H
daaa166f 3
dadf66c5 4#include <inttypes.h>
daaa166f
JA
5#include "arch/arch.h"
6
bdc7211e
JA
7/* Fast hashing routine for a long.
8 (C) 2002 William Lee Irwin III, IBM */
9
10/*
11 * Knuth recommends primes in approximately golden ratio to the maximum
12 * integer representable by a machine word for multiplicative hashing.
13 * Chuck Lever verified the effectiveness of this technique:
14 * http://www.citi.umich.edu/techreports/reports/citi-tr-00-1.pdf
15 *
16 * These primes are chosen to be bit-sparse, that is operations on
17 * them can use shifts and additions instead of multiplications for
18 * machines where multiplications are slow.
19 */
5921e80c 20
bdc7211e
JA
21#if BITS_PER_LONG == 32
22/* 2^31 + 2^29 - 2^25 + 2^22 - 2^19 - 2^16 + 1 */
23#define GOLDEN_RATIO_PRIME 0x9e370001UL
24#elif BITS_PER_LONG == 64
25/* 2^63 + 2^61 - 2^57 + 2^54 - 2^51 - 2^18 + 1 */
26#define GOLDEN_RATIO_PRIME 0x9e37fffffffc0001UL
27#else
28#error Define GOLDEN_RATIO_PRIME for your wordsize.
29#endif
30
ed1860cd 31static inline unsigned long __hash_long(unsigned long val)
bdc7211e
JA
32{
33 unsigned long hash = val;
34
35#if BITS_PER_LONG == 64
36 /* Sigh, gcc can't optimise this alone like it does for 32 bits. */
37 unsigned long n = hash;
38 n <<= 18;
39 hash -= n;
40 n <<= 33;
41 hash -= n;
42 n <<= 3;
43 hash += n;
44 n <<= 3;
45 hash -= n;
46 n <<= 4;
47 hash += n;
48 n <<= 2;
49 hash += n;
50#else
51 /* On some cpus multiply is faster, on others gcc will do shifts */
52 hash *= GOLDEN_RATIO_PRIME;
53#endif
54
ed1860cd
JA
55 return hash;
56}
57
58static inline unsigned long hash_long(unsigned long val, unsigned int bits)
59{
bdc7211e 60 /* High bits are more random, so use them. */
ed1860cd 61 return __hash_long(val) >> (BITS_PER_LONG - bits);
bdc7211e
JA
62}
63
64static inline unsigned long hash_ptr(void *ptr, unsigned int bits)
65{
e43606c2 66 return hash_long((uintptr_t)ptr, bits);
bdc7211e 67}
dadf66c5
JA
68
69/*
70 * Bob Jenkins jhash
71 */
72
73#define JHASH_INITVAL GOLDEN_RATIO_PRIME
74
75static inline uint32_t rol32(uint32_t word, uint32_t shift)
76{
77 return (word << shift) | (word >> (32 - shift));
78}
79
80/* __jhash_mix -- mix 3 32-bit values reversibly. */
81#define __jhash_mix(a, b, c) \
82{ \
83 a -= c; a ^= rol32(c, 4); c += b; \
84 b -= a; b ^= rol32(a, 6); a += c; \
85 c -= b; c ^= rol32(b, 8); b += a; \
86 a -= c; a ^= rol32(c, 16); c += b; \
87 b -= a; b ^= rol32(a, 19); a += c; \
88 c -= b; c ^= rol32(b, 4); b += a; \
89}
90
91/* __jhash_final - final mixing of 3 32-bit values (a,b,c) into c */
92#define __jhash_final(a, b, c) \
93{ \
94 c ^= b; c -= rol32(b, 14); \
95 a ^= c; a -= rol32(c, 11); \
96 b ^= a; b -= rol32(a, 25); \
97 c ^= b; c -= rol32(b, 16); \
98 a ^= c; a -= rol32(c, 4); \
99 b ^= a; b -= rol32(a, 14); \
100 c ^= b; c -= rol32(b, 24); \
101}
102
103static inline uint32_t jhash(const void *key, uint32_t length, uint32_t initval)
104{
105 const uint8_t *k = key;
106 uint32_t a, b, c;
107
108 /* Set up the internal state */
109 a = b = c = JHASH_INITVAL + length + initval;
110
111 /* All but the last block: affect some 32 bits of (a,b,c) */
112 while (length > 12) {
113 a += *k;
114 b += *(k + 4);
115 c += *(k + 8);
116 __jhash_mix(a, b, c);
117 length -= 12;
118 k += 12;
119 }
120
121 /* Last block: affect all 32 bits of (c) */
122 /* All the case statements fall through */
123 switch (length) {
124 case 12: c += (uint32_t) k[11] << 24;
125 case 11: c += (uint32_t) k[10] << 16;
126 case 10: c += (uint32_t) k[9] << 8;
127 case 9: c += k[8];
128 case 8: b += (uint32_t) k[7] << 24;
129 case 7: b += (uint32_t) k[6] << 16;
130 case 6: b += (uint32_t) k[5] << 8;
131 case 5: b += k[4];
132 case 4: a += (uint32_t) k[3] << 24;
133 case 3: a += (uint32_t) k[2] << 16;
134 case 2: a += (uint32_t) k[1] << 8;
135 case 1: a += k[0];
136 __jhash_final(a, b, c);
137 case 0: /* Nothing left to add */
138 break;
139 }
140
141 return c;
142}
143
bdc7211e 144#endif /* _LINUX_HASH_H */