jenkins_hash.c revision 218965
1#ifndef __LIBKERN_JENKINS_H__ 2#define __LIBKERN_JENKINS_H__ 3/* 4 * Taken from http://burtleburtle.net/bob/c/lookup3.c 5 * $FreeBSD: head/sys/libkern/jenkins.h 218965 2011-02-23 09:22:33Z brucec $ 6 */ 7 8/* 9------------------------------------------------------------------------------- 10 lookup3.c, by Bob Jenkins, May 2006, Public Domain. 11 12 These are functions for producing 32-bit hashes for hash table lookup. 13 hashword(), hashlittle(), hashlittle2(), hashbig(), mix(), and final() 14 are externally useful functions. Routines to test the hash are included 15 if SELF_TEST is defined. You can use this free for any purpose. It's in 16 the public domain. It has no warranty. 17 18 You probably want to use hashlittle(). hashlittle() and hashbig() 19 hash byte arrays. hashlittle() is faster than hashbig() on 20 little-endian machines. Intel and AMD are little-endian machines. 21 On second thought, you probably want hashlittle2(), which is identical to 22 hashlittle() except it returns two 32-bit hashes for the price of one. 23 You could implement hashbig2() if you wanted but I haven't bothered here. 24 25 If you want to find a hash of, say, exactly 7 integers, do 26 a = i1; b = i2; c = i3; 27 mix(a,b,c); 28 a += i4; b += i5; c += i6; 29 mix(a,b,c); 30 a += i7; 31 final(a,b,c); 32 then use c as the hash value. If you have a variable length array of 33 4-byte integers to hash, use hashword(). If you have a byte array (like 34 a character string), use hashlittle(). If you have several byte arrays, or 35 a mix of things, see the comments above hashlittle(). 36 37 Why is this so big? I read 12 bytes at a time into 3 4-byte integers, 38 then mix those integers. This is fast (you can do a lot more thorough 39 mixing with 12*3 instructions on 3 integers than you can with 3 instructions 40 on 1 byte), but shoehorning those bytes into integers efficiently is messy. 41------------------------------------------------------------------------------- 42*/ 43 44#define rot(x,k) (((x)<<(k)) | ((x)>>(32-(k)))) 45 46/* 47------------------------------------------------------------------------------- 48mix -- mix 3 32-bit values reversibly. 49 50This is reversible, so any information in (a,b,c) before mix() is 51still in (a,b,c) after mix(). 52 53If four pairs of (a,b,c) inputs are run through mix(), or through 54mix() in reverse, there are at least 32 bits of the output that 55are sometimes the same for one pair and different for another pair. 56This was tested for: 57* pairs that differed by one bit, by two bits, in any combination 58 of top bits of (a,b,c), or in any combination of bottom bits of 59 (a,b,c). 60* "differ" is defined as +, -, ^, or ~^. For + and -, I transformed 61 the output delta to a Gray code (a^(a>>1)) so a string of 1's (as 62 is commonly produced by subtraction) look like a single 1-bit 63 difference. 64* the base values were pseudorandom, all zero but one bit set, or 65 all zero plus a counter that starts at zero. 66 67Some k values for my "a-=c; a^=rot(c,k); c+=b;" arrangement that 68satisfy this are 69 4 6 8 16 19 4 70 9 15 3 18 27 15 71 14 9 3 7 17 3 72Well, "9 15 3 18 27 15" didn't quite get 32 bits diffing 73for "differ" defined as + with a one-bit base and a two-bit delta. I 74used http://burtleburtle.net/bob/hash/avalanche.html to choose 75the operations, constants, and arrangements of the variables. 76 77This does not achieve avalanche. There are input bits of (a,b,c) 78that fail to affect some output bits of (a,b,c), especially of a. The 79most thoroughly mixed value is c, but it doesn't really even achieve 80avalanche in c. 81 82This allows some parallelism. Read-after-writes are good at doubling 83the number of bits affected, so the goal of mixing pulls in the opposite 84direction as the goal of parallelism. I did what I could. Rotates 85seem to cost as much as shifts on every machine I could lay my hands 86on, and rotates are much kinder to the top and bottom bits, so I used 87rotates. 88------------------------------------------------------------------------------- 89*/ 90#define mix(a,b,c) \ 91{ \ 92 a -= c; a ^= rot(c, 4); c += b; \ 93 b -= a; b ^= rot(a, 6); a += c; \ 94 c -= b; c ^= rot(b, 8); b += a; \ 95 a -= c; a ^= rot(c,16); c += b; \ 96 b -= a; b ^= rot(a,19); a += c; \ 97 c -= b; c ^= rot(b, 4); b += a; \ 98} 99 100/* 101------------------------------------------------------------------------------- 102final -- final mixing of 3 32-bit values (a,b,c) into c 103 104Pairs of (a,b,c) values differing in only a few bits will usually 105produce values of c that look totally different. This was tested for 106* pairs that differed by one bit, by two bits, in any combination 107 of top bits of (a,b,c), or in any combination of bottom bits of 108 (a,b,c). 109* "differ" is defined as +, -, ^, or ~^. For + and -, I transformed 110 the output delta to a Gray code (a^(a>>1)) so a string of 1's (as 111 is commonly produced by subtraction) look like a single 1-bit 112 difference. 113* the base values were pseudorandom, all zero but one bit set, or 114 all zero plus a counter that starts at zero. 115 116These constants passed: 117 14 11 25 16 4 14 24 118 12 14 25 16 4 14 24 119and these came close: 120 4 8 15 26 3 22 24 121 10 8 15 26 3 22 24 122 11 8 15 26 3 22 24 123------------------------------------------------------------------------------- 124*/ 125#define final(a,b,c) \ 126{ \ 127 c ^= b; c -= rot(b,14); \ 128 a ^= c; a -= rot(c,11); \ 129 b ^= a; b -= rot(a,25); \ 130 c ^= b; c -= rot(b,16); \ 131 a ^= c; a -= rot(c,4); \ 132 b ^= a; b -= rot(a,14); \ 133 c ^= b; c -= rot(b,24); \ 134} 135 136/* 137-------------------------------------------------------------------- 138 This works on all machines. To be useful, it requires 139 -- that the key be an array of uint32_t's, and 140 -- that the length be the number of uint32_t's in the key 141 142 The function hashword() is identical to hashlittle() on little-endian 143 machines, and identical to hashbig() on big-endian machines, 144 except that the length has to be measured in uint32_ts rather than in 145 bytes. hashlittle() is more complicated than hashword() only because 146 hashlittle() has to dance around fitting the key bytes into registers. 147-------------------------------------------------------------------- 148*/ 149static uint32_t 150jenkins_hashword( 151 const uint32_t *k, /* the key, an array of uint32_t values */ 152 size_t length, /* the length of the key, in uint32_ts */ 153 uint32_t initval /* the previous hash, or an arbitrary value */ 154) 155{ 156 uint32_t a,b,c; 157 158 /* Set up the internal state */ 159 a = b = c = 0xdeadbeef + (((uint32_t)length)<<2) + initval; 160 161 /*------------------------------------------------- handle most of the key */ 162 while (length > 3) 163 { 164 a += k[0]; 165 b += k[1]; 166 c += k[2]; 167 mix(a,b,c); 168 length -= 3; 169 k += 3; 170 } 171 172 /*------------------------------------------- handle the last 3 uint32_t's */ 173 switch(length) /* all the case statements fall through */ 174 { 175 case 3 : c+=k[2]; 176 case 2 : b+=k[1]; 177 case 1 : a+=k[0]; 178 final(a,b,c); 179 case 0: /* case 0: nothing left to add */ 180 break; 181 } 182 /*------------------------------------------------------ report the result */ 183 return c; 184} 185#endif 186