3 * By Bob Jenkins, 2006. bob_jenkins@burtleburtle.net. You may use this
4 * code any way you wish, private, educational, or commercial. It's free.
5 * Use for hash table lookup, or anything where one collision in 2^^32 is
6 * acceptable. Do NOT use for cryptographic purposes.
7 * http://burtleburtle.net/bob/hash/index.html
9 * Modified by Brian Pontz for libmemcached
11 * Add big endian support
16 #define hashsize(n) ((uint32_t)1<<(n))
17 #define hashmask(n) (hashsize(n)-1)
18 #define rot(x,k) (((x)<<(k)) | ((x)>>(32-(k))))
22 a -= c; a ^= rot(c, 4); c += b; \
23 b -= a; b ^= rot(a, 6); a += c; \
24 c -= b; c ^= rot(b, 8); b += a; \
25 a -= c; a ^= rot(c,16); c += b; \
26 b -= a; b ^= rot(a,19); a += c; \
27 c -= b; c ^= rot(b, 4); b += a; \
30 #define final(a,b,c) \
32 c ^= b; c -= rot(b,14); \
33 a ^= c; a -= rot(c,11); \
34 b ^= a; b -= rot(a,25); \
35 c ^= b; c -= rot(b,16); \
36 a ^= c; a -= rot(c,4); \
37 b ^= a; b -= rot(a,14); \
38 c ^= b; c -= rot(b,24); \
42 jenkins_hash() -- hash a variable-length key into a 32-bit value
43 k : the key (the unaligned variable-length array of bytes)
44 length : the length of the key, counting by bytes
45 initval : can be any 4-byte value
46 Returns a 32-bit value. Every bit of the key affects every bit of
47 the return value. Two keys differing by one or two bits will have
48 totally different hash values.
50 The best hash table sizes are powers of 2. There is no need to do
51 mod a prime (mod is sooo slow!). If you need less than 32 bits,
52 use a bitmask. For example, if you need only 10 bits, do
53 h = (h & hashmask(10));
54 In which case, the hash table should have hashsize(10) elements.
57 uint32_t jenkins_hash(const void *key
, size_t length
, uint32_t initval
)
59 uint32_t a
,b
,c
; /* internal state */
60 union { const void *ptr
; size_t i
; } u
; /* needed for Mac Powerbook G4 */
62 /* Set up the internal state */
63 a
= b
= c
= 0xdeadbeef + ((uint32_t)length
) + initval
;
68 const uint32_t *k
= (const uint32_t *)key
; /* read 32-bit chunks */
70 /*------ all but last block: aligned reads and affect 32 bits of (a,b,c) */
81 /*----------------------------- handle the last (probably partial) block */
83 * "k[2]&0xffffff" actually reads beyond the end of the string, but
84 * then masks off the part it's not allowed to read. Because the
85 * string is aligned, the masked-off tail is in the same word as the
86 * rest of the string. Every machine with memory protection I've seen
87 * does it on word boundaries, so is OK with this. But VALGRIND will
88 * still catch it and complain. The masking trick does make the hash
89 * noticably faster for short strings (like English words).
93 case 12: c
+=k
[2]; b
+=k
[1]; a
+=k
[0]; break;
94 case 11: c
+=k
[2]&0xffffff; b
+=k
[1]; a
+=k
[0]; break;
95 case 10: c
+=k
[2]&0xffff; b
+=k
[1]; a
+=k
[0]; break;
96 case 9 : c
+=k
[2]&0xff; b
+=k
[1]; a
+=k
[0]; break;
97 case 8 : b
+=k
[1]; a
+=k
[0]; break;
98 case 7 : b
+=k
[1]&0xffffff; a
+=k
[0]; break;
99 case 6 : b
+=k
[1]&0xffff; a
+=k
[0]; break;
100 case 5 : b
+=k
[1]&0xff; a
+=k
[0]; break;
101 case 4 : a
+=k
[0]; break;
102 case 3 : a
+=k
[0]&0xffffff; break;
103 case 2 : a
+=k
[0]&0xffff; break;
104 case 1 : a
+=k
[0]&0xff; break;
105 case 0 : return c
; /* zero length strings require no mixing */
109 else if ((u
.i
& 0x1) == 0)
111 const uint16_t *k
= (const uint16_t *)key
; /* read 16-bit chunks */
114 /*--------------- all but last block: aligned reads and different mixing */
117 a
+= k
[0] + (((uint32_t)k
[1])<<16);
118 b
+= k
[2] + (((uint32_t)k
[3])<<16);
119 c
+= k
[4] + (((uint32_t)k
[5])<<16);
125 /*----------------------------- handle the last (probably partial) block */
126 k8
= (const uint8_t *)k
;
129 case 12: c
+=k
[4]+(((uint32_t)k
[5])<<16);
130 b
+=k
[2]+(((uint32_t)k
[3])<<16);
131 a
+=k
[0]+(((uint32_t)k
[1])<<16);
133 case 11: c
+=((uint32_t)k8
[10])<<16; /* fall through */
135 b
+=k
[2]+(((uint32_t)k
[3])<<16);
136 a
+=k
[0]+(((uint32_t)k
[1])<<16);
138 case 9 : c
+=k8
[8]; /* fall through */
139 case 8 : b
+=k
[2]+(((uint32_t)k
[3])<<16);
140 a
+=k
[0]+(((uint32_t)k
[1])<<16);
142 case 7 : b
+=((uint32_t)k8
[6])<<16; /* fall through */
144 a
+=k
[0]+(((uint32_t)k
[1])<<16);
146 case 5 : b
+=k8
[4]; /* fall through */
147 case 4 : a
+=k
[0]+(((uint32_t)k
[1])<<16);
149 case 3 : a
+=((uint32_t)k8
[2])<<16; /* fall through */
154 case 0 : return c
; /* zero length requires no mixing */
159 { /* need to read the key one byte at a time */
160 const uint8_t *k
= (const uint8_t *)key
;
162 /*--------------- all but the last block: affect some 32 bits of (a,b,c) */
166 a
+= ((uint32_t)k
[1])<<8;
167 a
+= ((uint32_t)k
[2])<<16;
168 a
+= ((uint32_t)k
[3])<<24;
170 b
+= ((uint32_t)k
[5])<<8;
171 b
+= ((uint32_t)k
[6])<<16;
172 b
+= ((uint32_t)k
[7])<<24;
174 c
+= ((uint32_t)k
[9])<<8;
175 c
+= ((uint32_t)k
[10])<<16;
176 c
+= ((uint32_t)k
[11])<<24;
182 /*-------------------------------- last block: affect all 32 bits of (c) */
183 switch(length
) /* all the case statements fall through */
185 case 12: c
+=((uint32_t)k
[11])<<24;
186 case 11: c
+=((uint32_t)k
[10])<<16;
187 case 10: c
+=((uint32_t)k
[9])<<8;
189 case 8 : b
+=((uint32_t)k
[7])<<24;
190 case 7 : b
+=((uint32_t)k
[6])<<16;
191 case 6 : b
+=((uint32_t)k
[5])<<8;
193 case 4 : a
+=((uint32_t)k
[3])<<24;
194 case 3 : a
+=((uint32_t)k
[2])<<16;
195 case 2 : a
+=((uint32_t)k
[1])<<8;