git.m6w6.name Git - awesomized/libmemcached/blob - libmemcached/jenkins_hash.c

   1 /*
   2 *
   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
   8 *
   9 * Modified by Brian Pontz for libmemcached
  10 * TODO:
  11 * Add big endian support
  12 */
  13
  14 #include "common.h"
  15
  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))))
  19
  20 #define mix(a,b,c) \
  21 { \
  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; \
  28 }
  29
  30 #define final(a,b,c) \
  31 { \
  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); \
  39 }
  40
  41 /*
  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.
  49
  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.
  55 */
  56
  57 uint32_t jenkins_hash(const void *key, size_t length, uint32_t initval)
  58 {
  59   uint32_t a,b,c;                                          /* internal state */
  60   union { const void *ptr; size_t i; } u;     /* needed for Mac Powerbook G4 */
  61
  62   /* Set up the internal state */
  63   a = b = c = 0xdeadbeef + ((uint32_t)length) + initval;
  64
  65   u.ptr = key;
  66   if ((u.i & 0x3) == 0) {
  67     const uint32_t *k = (const uint32_t *)key;         /* read 32-bit chunks */
  68     const uint8_t  *k8;
  69
  70     /*------ all but last block: aligned reads and affect 32 bits of (a,b,c) */
  71     while (length > 12)
  72     {
  73       a += k[0];
  74       b += k[1];
  75       c += k[2];
  76       mix(a,b,c);
  77       length -= 12;
  78       k += 3;
  79     }
  80
  81     /*----------------------------- handle the last (probably partial) block */
  82     /*
  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).
  90      */
  91     switch(length)
  92     {
  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 */
 106     }
 107
 108   } else if ((u.i & 0x1) == 0) {
 109     const uint16_t *k = (const uint16_t *)key;         /* read 16-bit chunks */
 110     const uint8_t  *k8;
 111
 112     /*--------------- all but last block: aligned reads and different mixing */
 113     while (length > 12)
 114     {
 115       a += k[0] + (((uint32_t)k[1])<<16);
 116       b += k[2] + (((uint32_t)k[3])<<16);
 117       c += k[4] + (((uint32_t)k[5])<<16);
 118       mix(a,b,c);
 119       length -= 12;
 120       k += 6;
 121     }
 122
 123     /*----------------------------- handle the last (probably partial) block */
 124     k8 = (const uint8_t *)k;
 125     switch(length)
 126     {
 127     case 12: c+=k[4]+(((uint32_t)k[5])<<16);
 128              b+=k[2]+(((uint32_t)k[3])<<16);
 129              a+=k[0]+(((uint32_t)k[1])<<16);
 130              break;
 131     case 11: c+=((uint32_t)k8[10])<<16;     /* fall through */
 132     case 10: c+=k[4];
 133              b+=k[2]+(((uint32_t)k[3])<<16);
 134              a+=k[0]+(((uint32_t)k[1])<<16);
 135              break;
 136     case 9 : c+=k8[8];                      /* fall through */
 137     case 8 : b+=k[2]+(((uint32_t)k[3])<<16);
 138              a+=k[0]+(((uint32_t)k[1])<<16);
 139              break;
 140     case 7 : b+=((uint32_t)k8[6])<<16;      /* fall through */
 141     case 6 : b+=k[2];
 142              a+=k[0]+(((uint32_t)k[1])<<16);
 143              break;
 144     case 5 : b+=k8[4];                      /* fall through */
 145     case 4 : a+=k[0]+(((uint32_t)k[1])<<16);
 146              break;
 147     case 3 : a+=((uint32_t)k8[2])<<16;      /* fall through */
 148     case 2 : a+=k[0];
 149              break;
 150     case 1 : a+=k8[0];
 151              break;
 152     case 0 : return c;                     /* zero length requires no mixing */
 153     }
 154
 155   } else {                        /* need to read the key one byte at a time */
 156     const uint8_t *k = (const uint8_t *)key;
 157
 158     /*--------------- all but the last block: affect some 32 bits of (a,b,c) */
 159     while (length > 12)
 160     {
 161       a += k[0];
 162       a += ((uint32_t)k[1])<<8;
 163       a += ((uint32_t)k[2])<<16;
 164       a += ((uint32_t)k[3])<<24;
 165       b += k[4];
 166       b += ((uint32_t)k[5])<<8;
 167       b += ((uint32_t)k[6])<<16;
 168       b += ((uint32_t)k[7])<<24;
 169       c += k[8];
 170       c += ((uint32_t)k[9])<<8;
 171       c += ((uint32_t)k[10])<<16;
 172       c += ((uint32_t)k[11])<<24;
 173       mix(a,b,c);
 174       length -= 12;
 175       k += 12;
 176     }
 177
 178     /*-------------------------------- last block: affect all 32 bits of (c) */
 179     switch(length)                   /* all the case statements fall through */
 180     {
 181     case 12: c+=((uint32_t)k[11])<<24;
 182     case 11: c+=((uint32_t)k[10])<<16;
 183     case 10: c+=((uint32_t)k[9])<<8;
 184     case 9 : c+=k[8];
 185     case 8 : b+=((uint32_t)k[7])<<24;
 186     case 7 : b+=((uint32_t)k[6])<<16;
 187     case 6 : b+=((uint32_t)k[5])<<8;
 188     case 5 : b+=k[4];
 189     case 4 : a+=((uint32_t)k[3])<<24;
 190     case 3 : a+=((uint32_t)k[2])<<16;
 191     case 2 : a+=((uint32_t)k[1])<<8;
 192     case 1 : a+=k[0];
 193              break;
 194     case 0 : return c;
 195     }
 196   }
 197
 198   final(a,b,c);
 199   return c;
 200 }
 201
 202