git.m6w6.name Git - m6w6/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 #ifndef WORDS_BIGENDIAN
  67   if ((u.i & 0x3) == 0)
  68   {
  69     const uint32_t *k = (const uint32_t *)key;         /* read 32-bit chunks */
  70
  71     /*------ all but last block: aligned reads and affect 32 bits of (a,b,c) */
  72     while (length > 12)
  73     {
  74       a += k[0];
  75       b += k[1];
  76       c += k[2];
  77       mix(a,b,c);
  78       length -= 12;
  79       k += 3;
  80     }
  81
  82     /*----------------------------- handle the last (probably partial) block */
  83     /*
  84      * "k[2]&0xffffff" actually reads beyond the end of the string, but
  85      * then masks off the part it's not allowed to read.  Because the
  86      * string is aligned, the masked-off tail is in the same word as the
  87      * rest of the string.  Every machine with memory protection I've seen
  88      * does it on word boundaries, so is OK with this.  But VALGRIND will
  89      * still catch it and complain.  The masking trick does make the hash
  90      * noticably faster for short strings (like English words).
  91      */
  92     switch(length)
  93     {
  94     case 12: c+=k[2]; b+=k[1]; a+=k[0]; break;
  95     case 11: c+=k[2]&0xffffff; b+=k[1]; a+=k[0]; break;
  96     case 10: c+=k[2]&0xffff; b+=k[1]; a+=k[0]; break;
  97     case 9 : c+=k[2]&0xff; b+=k[1]; a+=k[0]; break;
  98     case 8 : b+=k[1]; a+=k[0]; break;
  99     case 7 : b+=k[1]&0xffffff; a+=k[0]; break;
 100     case 6 : b+=k[1]&0xffff; a+=k[0]; break;
 101     case 5 : b+=k[1]&0xff; a+=k[0]; break;
 102     case 4 : a+=k[0]; break;
 103     case 3 : a+=k[0]&0xffffff; break;
 104     case 2 : a+=k[0]&0xffff; break;
 105     case 1 : a+=k[0]&0xff; break;
 106     case 0 : return c;              /* zero length strings require no mixing */
 107     default: return c;
 108     }
 109
 110   }
 111   else if ((u.i & 0x1) == 0)
 112   {
 113     const uint16_t *k = (const uint16_t *)key;         /* read 16-bit chunks */
 114     const uint8_t  *k8;
 115
 116     /*--------------- all but last block: aligned reads and different mixing */
 117     while (length > 12)
 118     {
 119       a += k[0] + (((uint32_t)k[1])<<16);
 120       b += k[2] + (((uint32_t)k[3])<<16);
 121       c += k[4] + (((uint32_t)k[5])<<16);
 122       mix(a,b,c);
 123       length -= 12;
 124       k += 6;
 125     }
 126
 127     /*----------------------------- handle the last (probably partial) block */
 128     k8 = (const uint8_t *)k;
 129     switch(length)
 130     {
 131     case 12: c+=k[4]+(((uint32_t)k[5])<<16);
 132              b+=k[2]+(((uint32_t)k[3])<<16);
 133              a+=k[0]+(((uint32_t)k[1])<<16);
 134              break;
 135     case 11: c+=((uint32_t)k8[10])<<16;     /* fall through */
 136     case 10: c+=k[4];
 137              b+=k[2]+(((uint32_t)k[3])<<16);
 138              a+=k[0]+(((uint32_t)k[1])<<16);
 139              break;
 140     case 9 : c+=k8[8];                      /* fall through */
 141     case 8 : b+=k[2]+(((uint32_t)k[3])<<16);
 142              a+=k[0]+(((uint32_t)k[1])<<16);
 143              break;
 144     case 7 : b+=((uint32_t)k8[6])<<16;      /* fall through */
 145     case 6 : b+=k[2];
 146              a+=k[0]+(((uint32_t)k[1])<<16);
 147              break;
 148     case 5 : b+=k8[4];                      /* fall through */
 149     case 4 : a+=k[0]+(((uint32_t)k[1])<<16);
 150              break;
 151     case 3 : a+=((uint32_t)k8[2])<<16;      /* fall through */
 152     case 2 : a+=k[0];
 153              break;
 154     case 1 : a+=k8[0];
 155              break;
 156     case 0 : return c;                     /* zero length requires no mixing */
 157     default: return c;
 158     }
 159
 160   }
 161   else
 162   {                        /* need to read the key one byte at a time */
 163 #endif /* little endian */
 164     const uint8_t *k = (const uint8_t *)key;
 165
 166     /*--------------- all but the last block: affect some 32 bits of (a,b,c) */
 167     while (length > 12)
 168     {
 169       a += k[0];
 170       a += ((uint32_t)k[1])<<8;
 171       a += ((uint32_t)k[2])<<16;
 172       a += ((uint32_t)k[3])<<24;
 173       b += k[4];
 174       b += ((uint32_t)k[5])<<8;
 175       b += ((uint32_t)k[6])<<16;
 176       b += ((uint32_t)k[7])<<24;
 177       c += k[8];
 178       c += ((uint32_t)k[9])<<8;
 179       c += ((uint32_t)k[10])<<16;
 180       c += ((uint32_t)k[11])<<24;
 181       mix(a,b,c);
 182       length -= 12;
 183       k += 12;
 184     }
 185
 186     /*-------------------------------- last block: affect all 32 bits of (c) */
 187     switch(length)                   /* all the case statements fall through */
 188     {
 189     case 12: c+=((uint32_t)k[11])<<24;
 190     case 11: c+=((uint32_t)k[10])<<16;
 191     case 10: c+=((uint32_t)k[9])<<8;
 192     case 9 : c+=k[8];
 193     case 8 : b+=((uint32_t)k[7])<<24;
 194     case 7 : b+=((uint32_t)k[6])<<16;
 195     case 6 : b+=((uint32_t)k[5])<<8;
 196     case 5 : b+=k[4];
 197     case 4 : a+=((uint32_t)k[3])<<24;
 198     case 3 : a+=((uint32_t)k[2])<<16;
 199     case 2 : a+=((uint32_t)k[1])<<8;
 200     case 1 : a+=k[0];
 201              break;
 202     case 0 : return c;
 203     default : return c;
 204     }
 205 #ifndef WORDS_BIGENDIAN
 206   }
 207 #endif
 208
 209   final(a,b,c);
 210   return c;
 211 }
 212
 213