hash.c revision 219820
1114402Sru/*
2151497Sru * Copyright (c) 2005 Voltaire Inc.  All rights reserved.
3114402Sru *
4114402Sru * This software is available to you under a choice of one of two
5114402Sru * licenses.  You may choose to be licensed under the terms of the GNU
6114402Sru * General Public License (GPL) Version 2, available from the file
7114402Sru * COPYING in the main directory of this source tree, or the
8114402Sru * OpenIB.org BSD license below:
9114402Sru *
10114402Sru *     Redistribution and use in source and binary forms, with or
11114402Sru *     without modification, are permitted provided that the following
12114402Sru *     conditions are met:
13114402Sru *
14114402Sru *      - Redistributions of source code must retain the above
15114402Sru *        copyright notice, this list of conditions and the following
16114402Sru *        disclaimer.
17114402Sru *
18114402Sru *      - Redistributions in binary form must reproduce the above
19151497Sru *        copyright notice, this list of conditions and the following
20114402Sru *        disclaimer in the documentation and/or other materials
21114402Sru *        provided with the distribution.
22114402Sru *
23114402Sru * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24114402Sru * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25114402Sru * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26114402Sru * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27114402Sru * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28114402Sru * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29114402Sru * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30151497Sru * SOFTWARE.
31114402Sru *
32151497Sru */
33151497Sru
34114402Sru/*
35114402Sru * By Bob Jenkins, 1996.  bob_jenkins@burtleburtle.net.  You may use this
36114402Sru * code any way you wish, private, educational, or commercial.  It's free.
37114402Sru *
38114402Sru * See http://burtleburtle.net/bob/hash/evahash.html
39114402Sru * Use for hash table lookup, or anything where one collision in 2^^32 is
40114402Sru * acceptable.  Do NOT use for cryptographic purposes.
41114402Sru */
42151497Sru
43114402Sru#include <common.h>
44114402Sru
45151497Sru#define hashsize(n) ((uint32)1<<(n))
46114402Sru#define hashmask(n) (hashsize(n)-1)
47114402Sru
48151497Sru
49151497Sru/*
50151497Sru--------------------------------------------------------------------
51114402Srumix -- mix 3 32-bit values reversibly.
52151497SruFor every delta with one or two bits set, and the deltas of all three
53151497Sru  high bits or all three low bits, whether the original value of a,b,c
54114402Sru  is almost all zero or is uniformly distributed,
55151497Sru* If mix() is run forward or backward, at least 32 bits in a,b,c
56151497Sru  have at least 1/4 probability of changing.
57151497Sru* If mix() is run forward, every bit of c will change between 1/3 and
58151497Sru  2/3 of the time.  (Well, 22/100 and 78/100 for some 2-bit deltas.)
59151497Srumix() was built out of 36 single-cycle latency instructions in a
60151497Sru  structure that could supported 2x parallelism, like so:
61151497Sru      a -= b;
62151497Sru      a -= c; x = (c>>13);
63114402Sru      b -= c; a ^= x;
64114402Sru      b -= a; x = (a<<8);
65114402Sru      c -= a; b ^= x;
66151497Sru      c -= b; x = (b>>13);
67114402Sru      ...
68114402Sru  Unfortunately, superscalar Pentiums and Sparcs can't take advantage
69114402Sru  of that parallelism.  They've also turned some of those single-cycle
70114402Sru  latency instructions into multi-cycle latency instructions.  Still,
71114402Sru  this is the fastest good hash I could find.  There were about 2^^68
72114402Sru  to choose from.  I only looked at a billion or so.
73114402Sru--------------------------------------------------------------------
74151497Sru*/
75151497Sru#define mix(a,b,c) \
76151497Sru{ \
77114402Sru	a -= b; a -= c; a ^= (c>>13);	\
78151497Sru	b -= c; b -= a; b ^= (a<<8);	\
79114402Sru	c -= a; c -= b; c ^= (b>>13);	\
80151497Sru	a -= b; a -= c; a ^= (c>>12);	\
81114402Sru	b -= c; b -= a; b ^= (a<<16);	\
82114402Sru	c -= a; c -= b; c ^= (b>>5);	\
83151497Sru	a -= b; a -= c; a ^= (c>>3);	\
84114402Sru	b -= c; b -= a; b ^= (a<<10);	\
85114402Sru	c -= a; c -= b; c ^= (b>>15);	\
86114402Sru}
87114402Sru
88151497Sru/*
89151497Sru--------------------------------------------------------------------
90151497Srufhash() -- hash a variable-length key into a 32-bit value
91151497Sru  k       : the key (the unaligned variable-length array of bytes)
92151497Sru  len     : the length of the key, counting by bytes
93114402Sru  initval : can be any 4-byte value
94151497SruReturns a 32-bit value.  Every bit of the key affects every bit of
95151497Sruthe return value.  Every 1-bit and 2-bit delta achieves avalanche.
96114402SruAbout 6*len+35 instructions.
97151497Sru
98151497SruThe best hash table sizes are powers of 2.  There is no need to do
99114402Srumod a prime (mod is sooo slow!).  If you need less than 32 bits,
100114402Sruuse a bitmask.  For example, if you need only 10 bits, do
101151497Sru  h = (h & hashmask(10));
102151497SruIn which case, the hash table should have hashsize(10) elements.
103151497Sru
104151497SruIf you are hashing n strings (uint8 **)k, do it like this:
105114402Sru  for (i=0, h=0; i<n; ++i) h = hash( k[i], len[i], h);
106151497Sru
107114402Sru--------------------------------------------------------------------
108114402Sru*/
109114402Sru
110114402Sruuint32_t
111114402Srufhash(uint8_t *k, int length, uint32_t initval)
112151497Sru{
113151497Sru	uint32_t a, b, c, len;
114114402Sru
115151497Sru	/* Set up the internal state */
116114402Sru	len = length;
117151497Sru	a = b = 0x9e3779b9;		/* the golden ratio; an arbitrary value */
118114402Sru	c = initval;			/* the previous hash value */
119151497Sru
120151497Sru	/* handle most of the key */
121151497Sru	while (len >= 12) {
122151497Sru		a += (k[0] + ((uint32_t)k[1]<<8) +
123151497Sru		      ((uint32_t)k[2]<<16) + ((uint32_t)k[3]<<24));
124114402Sru		b += (k[4] + ((uint32_t)k[5]<<8) + ((uint32_t)k[6]<<16) +
125114402Sru		      ((uint32_t)k[7]<<24));
126151497Sru		c += (k[8] + ((uint32_t)k[9]<<8) + ((uint32_t)k[10]<<16) +
127151497Sru		      ((uint32_t)k[11]<<24));
128114402Sru		mix(a, b, c);
129151497Sru		k += 12; len -= 12;
130151497Sru	}
131151497Sru
132151497Sru	/* handle the last 11 bytes */
133151497Sru	c += length;
134151497Sru	switch (len) {		/* all the case statements fall through */
135114402Sru	case 11: c += ((uint32_t)k[10]<<24);
136114402Sru	case 10: c += ((uint32_t)k[9]<<16);
137114402Sru	case 9 : c += ((uint32_t)k[8]<<8);
138114402Sru		/* the first byte of c is reserved for the length */
139114402Sru	case 8 : b += ((uint32_t)k[7]<<24);
140114402Sru	case 7 : b += ((uint32_t)k[6]<<16);
141114402Sru	case 6 : b += ((uint32_t)k[5]<<8);
142114402Sru	case 5 : b += k[4];
143114402Sru	case 4 : a += ((uint32_t)k[3]<<24);
144114402Sru	case 3 : a += ((uint32_t)k[2]<<16);
145114402Sru	case 2 : a += ((uint32_t)k[1]<<8);
146114402Sru	case 1 : a += k[0];
147114402Sru	  /* case 0: nothing left to add */
148151497Sru	}
149114402Sru
150114402Sru	mix(a, b, c);
151114402Sru
152114402Sru	return c;
153114402Sru}
154114402Sru