git.m6w6.name Git - awesomized/libmemcached/blob - libhashkit/jenkins.cc

   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 <libhashkit/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 #define JENKINS_INITVAL 13
  42
  43 /*
  44 jenkins_hash() -- hash a variable-length key into a 32-bit value
  45   k       : the key (the unaligned variable-length array of bytes)
  46   length  : the length of the key, counting by bytes
  47   initval : can be any 4-byte value
  48 Returns a 32-bit value.  Every bit of the key affects every bit of
  49 the return value.  Two keys differing by one or two bits will have
  50 totally different hash values.
  51
  52 The best hash table sizes are powers of 2.  There is no need to do
  53 mod a prime (mod is sooo slow!).  If you need less than 32 bits,
  54 use a bitmask.  For example, if you need only 10 bits, do
  55   h = (h & hashmask(10));
  56 In which case, the hash table should have hashsize(10) elements.
  57 */
  58
  59 uint32_t hashkit_jenkins(const char *key, size_t length, void *)
  60 {
  61   uint32_t a,b,c;                                          /* internal state */
  62   union { const void *ptr; size_t i; } u;     /* needed for Mac Powerbook G4 */
  63
  64   /* Set up the internal state */
  65   a = b = c = 0xdeadbeef + ((uint32_t)length) + JENKINS_INITVAL;
  66
  67   u.ptr = key;
  68 #ifndef WORDS_BIGENDIAN
  69   if ((u.i & 0x3) == 0)
  70   {
  71     const uint32_t *k = (const uint32_t *)key;         /* read 32-bit chunks */
  72
  73     /*------ all but last block: aligned reads and affect 32 bits of (a,b,c) */
  74     while (length > 12)
  75     {
  76       a += k[0];
  77       b += k[1];
  78       c += k[2];
  79       mix(a,b,c);
  80       length -= 12;
  81       k += 3;
  82     }
  83
  84     /*----------------------------- handle the last (probably partial) block */
  85     /*
  86      * "k[2]&0xffffff" actually reads beyond the end of the string, but
  87      * then masks off the part it's not allowed to read.  Because the
  88      * string is aligned, the masked-off tail is in the same word as the
  89      * rest of the string.  Every machine with memory protection I've seen
  90      * does it on word boundaries, so is OK with this.  But VALGRIND will
  91      * still catch it and complain.  The masking trick does make the hash
  92      * noticably faster for short strings (like English words).
  93      */
  94     switch(length)
  95     {
  96     case 12: c+=k[2]; b+=k[1]; a+=k[0]; break;
  97     case 11: c+=k[2]&0xffffff; b+=k[1]; a+=k[0]; break;
  98     case 10: c+=k[2]&0xffff; b+=k[1]; a+=k[0]; break;
  99     case 9 : c+=k[2]&0xff; b+=k[1]; a+=k[0]; break;
 100     case 8 : b+=k[1]; a+=k[0]; break;
 101     case 7 : b+=k[1]&0xffffff; a+=k[0]; break;
 102     case 6 : b+=k[1]&0xffff; a+=k[0]; break;
 103     case 5 : b+=k[1]&0xff; a+=k[0]; break;
 104     case 4 : a+=k[0]; break;
 105     case 3 : a+=k[0]&0xffffff; break;
 106     case 2 : a+=k[0]&0xffff; break;
 107     case 1 : a+=k[0]&0xff; break;
 108     case 0 : return c;              /* zero length strings require no mixing */
 109     default: return c;
 110     }
 111
 112   }
 113   else if ((u.i & 0x1) == 0)
 114   {
 115     const uint16_t *k = (const uint16_t *)key;         /* read 16-bit chunks */
 116     const uint8_t  *k8;
 117
 118     /*--------------- all but last block: aligned reads and different mixing */
 119     while (length > 12)
 120     {
 121       a += k[0] + (((uint32_t)k[1])<<16);
 122       b += k[2] + (((uint32_t)k[3])<<16);
 123       c += k[4] + (((uint32_t)k[5])<<16);
 124       mix(a,b,c);
 125       length -= 12;
 126       k += 6;
 127     }
 128
 129     /*----------------------------- handle the last (probably partial) block */
 130     k8 = (const uint8_t *)k;
 131     switch(length)
 132     {
 133     case 12: c+=k[4]+(((uint32_t)k[5])<<16);
 134              b+=k[2]+(((uint32_t)k[3])<<16);
 135              a+=k[0]+(((uint32_t)k[1])<<16);
 136              break;
 137     case 11: c+=((uint32_t)k8[10])<<16;     /* fall through */
 138     case 10: c+=k[4];
 139              b+=k[2]+(((uint32_t)k[3])<<16);
 140              a+=k[0]+(((uint32_t)k[1])<<16);
 141              break;
 142     case 9 : c+=k8[8];                      /* fall through */
 143     case 8 : b+=k[2]+(((uint32_t)k[3])<<16);
 144              a+=k[0]+(((uint32_t)k[1])<<16);
 145              break;
 146     case 7 : b+=((uint32_t)k8[6])<<16;      /* fall through */
 147     case 6 : b+=k[2];
 148              a+=k[0]+(((uint32_t)k[1])<<16);
 149              break;
 150     case 5 : b+=k8[4];                      /* fall through */
 151     case 4 : a+=k[0]+(((uint32_t)k[1])<<16);
 152              break;
 153     case 3 : a+=((uint32_t)k8[2])<<16;      /* fall through */
 154     case 2 : a+=k[0];
 155              break;
 156     case 1 : a+=k8[0];
 157              break;
 158     case 0 : return c;                     /* zero length requires no mixing */
 159     default: return c;
 160     }
 161
 162   }
 163   else
 164   {                        /* need to read the key one byte at a time */
 165 #endif /* little endian */
 166     const uint8_t *k = (const uint8_t *)key;
 167
 168     /*--------------- all but the last block: affect some 32 bits of (a,b,c) */
 169     while (length > 12)
 170     {
 171       a += k[0];
 172       a += ((uint32_t)k[1])<<8;
 173       a += ((uint32_t)k[2])<<16;
 174       a += ((uint32_t)k[3])<<24;
 175       b += k[4];
 176       b += ((uint32_t)k[5])<<8;
 177       b += ((uint32_t)k[6])<<16;
 178       b += ((uint32_t)k[7])<<24;
 179       c += k[8];
 180       c += ((uint32_t)k[9])<<8;
 181       c += ((uint32_t)k[10])<<16;
 182       c += ((uint32_t)k[11])<<24;
 183       mix(a,b,c);
 184       length -= 12;
 185       k += 12;
 186     }
 187
 188     /*-------------------------------- last block: affect all 32 bits of (c) */
 189     switch(length)                   /* all the case statements fall through */
 190     {
 191     case 12: c+=((uint32_t)k[11])<<24;
 192     case 11: c+=((uint32_t)k[10])<<16;
 193     case 10: c+=((uint32_t)k[9])<<8;
 194     case 9 : c+=k[8];
 195     case 8 : b+=((uint32_t)k[7])<<24;
 196     case 7 : b+=((uint32_t)k[6])<<16;
 197     case 6 : b+=((uint32_t)k[5])<<8;
 198     case 5 : b+=k[4];
 199     case 4 : a+=((uint32_t)k[3])<<24;
 200     case 3 : a+=((uint32_t)k[2])<<16;
 201     case 2 : a+=((uint32_t)k[1])<<8;
 202     case 1 : a+=k[0];
 203              break;
 204     case 0 : return c;
 205     default : return c;
 206     }
 207 #ifndef WORDS_BIGENDIAN
 208   }
 209 #endif
 210
 211   final(a,b,c);
 212   return c;
 213 }