git.m6w6.name Git - awesomized/libmemcached/blob - libhashkit/jenkins.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 #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 *context)
  60 {
  61   uint32_t a,b,c;                                          /* internal state */
  62   union { const void *ptr; size_t i; } u;     /* needed for Mac Powerbook G4 */
  63   (void)context;
  64
  65   /* Set up the internal state */
  66   a = b = c = 0xdeadbeef + ((uint32_t)length) + JENKINS_INITVAL;
  67
  68   u.ptr = key;
  69 #ifndef WORDS_BIGENDIAN
  70   if ((u.i & 0x3) == 0)
  71   {
  72     const uint32_t *k = (const uint32_t *)key;         /* read 32-bit chunks */
  73
  74     /*------ all but last block: aligned reads and affect 32 bits of (a,b,c) */
  75     while (length > 12)
  76     {
  77       a += k[0];
  78       b += k[1];
  79       c += k[2];
  80       mix(a,b,c);
  81       length -= 12;
  82       k += 3;
  83     }
  84
  85     /*----------------------------- handle the last (probably partial) block */
  86     /*
  87      * "k[2]&0xffffff" actually reads beyond the end of the string, but
  88      * then masks off the part it's not allowed to read.  Because the
  89      * string is aligned, the masked-off tail is in the same word as the
  90      * rest of the string.  Every machine with memory protection I've seen
  91      * does it on word boundaries, so is OK with this.  But VALGRIND will
  92      * still catch it and complain.  The masking trick does make the hash
  93      * noticably faster for short strings (like English words).
  94      */
  95     switch(length)
  96     {
  97     case 12: c+=k[2]; b+=k[1]; a+=k[0]; break;
  98     case 11: c+=k[2]&0xffffff; b+=k[1]; a+=k[0]; break;
  99     case 10: c+=k[2]&0xffff; b+=k[1]; a+=k[0]; break;
 100     case 9 : c+=k[2]&0xff; b+=k[1]; a+=k[0]; break;
 101     case 8 : b+=k[1]; a+=k[0]; break;
 102     case 7 : b+=k[1]&0xffffff; a+=k[0]; break;
 103     case 6 : b+=k[1]&0xffff; a+=k[0]; break;
 104     case 5 : b+=k[1]&0xff; a+=k[0]; break;
 105     case 4 : a+=k[0]; break;
 106     case 3 : a+=k[0]&0xffffff; break;
 107     case 2 : a+=k[0]&0xffff; break;
 108     case 1 : a+=k[0]&0xff; break;
 109     case 0 : return c;              /* zero length strings require no mixing */
 110     default: return c;
 111     }
 112
 113   }
 114   else if ((u.i & 0x1) == 0)
 115   {
 116     const uint16_t *k = (const uint16_t *)key;         /* read 16-bit chunks */
 117     const uint8_t  *k8;
 118
 119     /*--------------- all but last block: aligned reads and different mixing */
 120     while (length > 12)
 121     {
 122       a += k[0] + (((uint32_t)k[1])<<16);
 123       b += k[2] + (((uint32_t)k[3])<<16);
 124       c += k[4] + (((uint32_t)k[5])<<16);
 125       mix(a,b,c);
 126       length -= 12;
 127       k += 6;
 128     }
 129
 130     /*----------------------------- handle the last (probably partial) block */
 131     k8 = (const uint8_t *)k;
 132     switch(length)
 133     {
 134     case 12: c+=k[4]+(((uint32_t)k[5])<<16);
 135              b+=k[2]+(((uint32_t)k[3])<<16);
 136              a+=k[0]+(((uint32_t)k[1])<<16);
 137              break;
 138     case 11: c+=((uint32_t)k8[10])<<16;     /* fall through */
 139     case 10: c+=k[4];
 140              b+=k[2]+(((uint32_t)k[3])<<16);
 141              a+=k[0]+(((uint32_t)k[1])<<16);
 142              break;
 143     case 9 : c+=k8[8];                      /* fall through */
 144     case 8 : b+=k[2]+(((uint32_t)k[3])<<16);
 145              a+=k[0]+(((uint32_t)k[1])<<16);
 146              break;
 147     case 7 : b+=((uint32_t)k8[6])<<16;      /* fall through */
 148     case 6 : b+=k[2];
 149              a+=k[0]+(((uint32_t)k[1])<<16);
 150              break;
 151     case 5 : b+=k8[4];                      /* fall through */
 152     case 4 : a+=k[0]+(((uint32_t)k[1])<<16);
 153              break;
 154     case 3 : a+=((uint32_t)k8[2])<<16;      /* fall through */
 155     case 2 : a+=k[0];
 156              break;
 157     case 1 : a+=k8[0];
 158              break;
 159     case 0 : return c;                     /* zero length requires no mixing */
 160     default: return c;
 161     }
 162
 163   }
 164   else
 165   {                        /* need to read the key one byte at a time */
 166 #endif /* little endian */
 167     const uint8_t *k = (const uint8_t *)key;
 168
 169     /*--------------- all but the last block: affect some 32 bits of (a,b,c) */
 170     while (length > 12)
 171     {
 172       a += k[0];
 173       a += ((uint32_t)k[1])<<8;
 174       a += ((uint32_t)k[2])<<16;
 175       a += ((uint32_t)k[3])<<24;
 176       b += k[4];
 177       b += ((uint32_t)k[5])<<8;
 178       b += ((uint32_t)k[6])<<16;
 179       b += ((uint32_t)k[7])<<24;
 180       c += k[8];
 181       c += ((uint32_t)k[9])<<8;
 182       c += ((uint32_t)k[10])<<16;
 183       c += ((uint32_t)k[11])<<24;
 184       mix(a,b,c);
 185       length -= 12;
 186       k += 12;
 187     }
 188
 189     /*-------------------------------- last block: affect all 32 bits of (c) */
 190     switch(length)                   /* all the case statements fall through */
 191     {
 192     case 12: c+=((uint32_t)k[11])<<24;
 193     case 11: c+=((uint32_t)k[10])<<16;
 194     case 10: c+=((uint32_t)k[9])<<8;
 195     case 9 : c+=k[8];
 196     case 8 : b+=((uint32_t)k[7])<<24;
 197     case 7 : b+=((uint32_t)k[6])<<16;
 198     case 6 : b+=((uint32_t)k[5])<<8;
 199     case 5 : b+=k[4];
 200     case 4 : a+=((uint32_t)k[3])<<24;
 201     case 3 : a+=((uint32_t)k[2])<<16;
 202     case 2 : a+=((uint32_t)k[1])<<8;
 203     case 1 : a+=k[0];
 204              break;
 205     case 0 : return c;
 206     default : return c;
 207     }
 208 #ifndef WORDS_BIGENDIAN
 209   }
 210 #endif
 211
 212   final(a,b,c);
 213   return c;
 214 }