2 -------------------------------------------------------------------------------
3 lookup3.c, by Bob Jenkins, May 2006, Public Domain.
5 These are functions for producing 32-bit hashes for hash table lookup.
6 hash_word(), hashlittle(), hashlittle2(), hashbig(), mix(), and final()
7 are externally useful functions. Routines to test the hash are included
8 if SELF_TEST is defined. You can use this free for any purpose. It's in
9 the public domain. It has no warranty.
11 You probably want to use hashlittle(). hashlittle() and hashbig()
12 hash byte arrays. hashlittle() is is faster than hashbig() on
13 little-endian machines. Intel and AMD are little-endian machines.
14 On second thought, you probably want hashlittle2(), which is identical to
15 hashlittle() except it returns two 32-bit hashes for the price of one.
16 You could implement hashbig2() if you wanted but I haven't bothered here.
18 If you want to find a hash of, say, exactly 7 integers, do
19 a = i1; b = i2; c = i3;
21 a += i4; b += i5; c += i6;
25 then use c as the hash value. If you have a variable length array of
26 4-byte integers to hash, use hash_word(). If you have a byte array (like
27 a character string), use hashlittle(). If you have several byte arrays, or
28 a mix of things, see the comments above hashlittle().
30 Why is this so big? I read 12 bytes at a time into 3 4-byte integers,
31 then mix those integers. This is fast (you can do a lot more thorough
32 mixing with 12*3 instructions on 3 integers than you can with 3 instructions
33 on 1 byte), but shoehorning those bytes into integers efficiently is messy.
34 -------------------------------------------------------------------------------
39 #include <stdio.h> /* defines printf for tests */
40 #include <time.h> /* defines time_t for timings in the test */
41 #include <stdint.h> /* defines uint32_t etc */
42 #include <sys/param.h> /* attempt to define endianness */
47 # include <endian.h> /* attempt to define endianness */
51 * My best guess at if you are big-endian or little-endian. This may
54 #if (defined(__BYTE_ORDER) && defined(__LITTLE_ENDIAN) && \
55 __BYTE_ORDER == __LITTLE_ENDIAN) || \
56 (defined(i386) || defined(__i386__) || defined(__i486__) || \
57 defined(__i586__) || defined(__i686__) || defined(vax) || defined(MIPSEL))
58 # define HASH_LITTLE_ENDIAN 1
59 # define HASH_BIG_ENDIAN 0
60 #elif (defined(__BYTE_ORDER) && defined(__BIG_ENDIAN) && \
61 __BYTE_ORDER == __BIG_ENDIAN) || \
62 (defined(sparc) || defined(POWERPC) || defined(mc68000) || defined(sel))
63 # define HASH_LITTLE_ENDIAN 0
64 # define HASH_BIG_ENDIAN 1
66 # error Unknown endian
69 #define hashsize(n) ((uint32_t)1<<(n))
70 #define hashmask(n) (hashsize(n)-1)
71 #define rot(x,k) (((x)<<(k)) | ((x)>>(32-(k))))
74 -------------------------------------------------------------------------------
75 mix -- mix 3 32-bit values reversibly.
77 This is reversible, so any information in (a,b,c) before mix() is
78 still in (a,b,c) after mix().
80 If four pairs of (a,b,c) inputs are run through mix(), or through
81 mix() in reverse, there are at least 32 bits of the output that
82 are sometimes the same for one pair and different for another pair.
84 * pairs that differed by one bit, by two bits, in any combination
85 of top bits of (a,b,c), or in any combination of bottom bits of
87 * "differ" is defined as +, -, ^, or ~^. For + and -, I transformed
88 the output delta to a Gray code (a^(a>>1)) so a string of 1's (as
89 is commonly produced by subtraction) look like a single 1-bit
91 * the base values were pseudorandom, all zero but one bit set, or
92 all zero plus a counter that starts at zero.
94 Some k values for my "a-=c; a^=rot(c,k); c+=b;" arrangement that
99 Well, "9 15 3 18 27 15" didn't quite get 32 bits diffing
100 for "differ" defined as + with a one-bit base and a two-bit delta. I
101 used http://burtleburtle.net/bob/hash/avalanche.html to choose
102 the operations, constants, and arrangements of the variables.
104 This does not achieve avalanche. There are input bits of (a,b,c)
105 that fail to affect some output bits of (a,b,c), especially of a. The
106 most thoroughly mixed value is c, but it doesn't really even achieve
109 This allows some parallelism. Read-after-writes are good at doubling
110 the number of bits affected, so the goal of mixing pulls in the opposite
111 direction as the goal of parallelism. I did what I could. Rotates
112 seem to cost as much as shifts on every machine I could lay my hands
113 on, and rotates are much kinder to the top and bottom bits, so I used
115 -------------------------------------------------------------------------------
119 a -= c; a ^= rot(c, 4); c += b; \
120 b -= a; b ^= rot(a, 6); a += c; \
121 c -= b; c ^= rot(b, 8); b += a; \
122 a -= c; a ^= rot(c,16); c += b; \
123 b -= a; b ^= rot(a,19); a += c; \
124 c -= b; c ^= rot(b, 4); b += a; \
128 -------------------------------------------------------------------------------
129 final -- final mixing of 3 32-bit values (a,b,c) into c
131 Pairs of (a,b,c) values differing in only a few bits will usually
132 produce values of c that look totally different. This was tested for
133 * pairs that differed by one bit, by two bits, in any combination
134 of top bits of (a,b,c), or in any combination of bottom bits of
136 * "differ" is defined as +, -, ^, or ~^. For + and -, I transformed
137 the output delta to a Gray code (a^(a>>1)) so a string of 1's (as
138 is commonly produced by subtraction) look like a single 1-bit
140 * the base values were pseudorandom, all zero but one bit set, or
141 all zero plus a counter that starts at zero.
143 These constants passed:
146 and these came close:
150 -------------------------------------------------------------------------------
152 #define final(a,b,c) \
154 c ^= b; c -= rot(b,14); \
155 a ^= c; a -= rot(c,11); \
156 b ^= a; b -= rot(a,25); \
157 c ^= b; c -= rot(b,16); \
158 a ^= c; a -= rot(c,4); \
159 b ^= a; b -= rot(a,14); \
160 c ^= b; c -= rot(b,24); \
164 --------------------------------------------------------------------
165 This works on all machines. To be useful, it requires
166 -- that the key be an array of uint32_t's, and
167 -- that the length be the number of uint32_t's in the key
169 The function hash_word() is identical to hashlittle() on little-endian
170 machines, and identical to hashbig() on big-endian machines,
171 except that the length has to be measured in uint32_ts rather than in
172 bytes. hashlittle() is more complicated than hash_word() only because
173 hashlittle() has to dance around fitting the key bytes into registers.
174 --------------------------------------------------------------------
177 const uint32_t *k, /* the key, an array of uint32_t values */
178 size_t length, /* the length of the key, in uint32_ts */
179 uint32_t initval) /* the previous hash, or an arbitrary value */
183 /* Set up the internal state */
184 a = b = c = 0xdeadbeef + (((uint32_t)length)<<2) + initval;
186 /*------------------------------------------------- handle most of the key */
197 /*------------------------------------------- handle the last 3 uint32_t's */
198 switch(length) /* all the case statements fall through */
204 case 0: /* case 0: nothing left to add */
207 /*------------------------------------------------------ report the result */
214 --------------------------------------------------------------------
215 hash_word2() -- same as hash_word(), but take two seeds and return two
216 32-bit values. pc and pb must both be nonnull, and *pc and *pb must
217 both be initialized with seeds. If you pass in (*pb)==0, the output
218 (*pc) will be the same as the return value from hash_word().
219 --------------------------------------------------------------------
222 const uint32_t *k, /* the key, an array of uint32_t values */
223 size_t length, /* the length of the key, in uint32_ts */
224 uint32_t *pc, /* IN: seed OUT: primary hash value */
225 uint32_t *pb) /* IN: more seed OUT: secondary hash value */
229 /* Set up the internal state */
230 a = b = c = 0xdeadbeef + ((uint32_t)(length<<2)) + *pc;
233 /*------------------------------------------------- handle most of the key */
244 /*------------------------------------------- handle the last 3 uint32_t's */
245 switch(length) /* all the case statements fall through */
251 case 0: /* case 0: nothing left to add */
254 /*------------------------------------------------------ report the result */
260 -------------------------------------------------------------------------------
261 hashlittle() -- hash a variable-length key into a 32-bit value
262 k : the key (the unaligned variable-length array of bytes)
263 length : the length of the key, counting by bytes
264 initval : can be any 4-byte value
265 Returns a 32-bit value. Every bit of the key affects every bit of
266 the return value. Two keys differing by one or two bits will have
267 totally different hash values.
269 The best hash table sizes are powers of 2. There is no need to do
270 mod a prime (mod is sooo slow!). If you need less than 32 bits,
271 use a bitmask. For example, if you need only 10 bits, do
272 h = (h & hashmask(10));
273 In which case, the hash table should have hashsize(10) elements.
275 If you are hashing n strings (uint8_t **)k, do it like this:
276 for (i=0, h=0; i<n; ++i) h = hashlittle( k[i], len[i], h);
278 By Bob Jenkins, 2006. bob_jenkins@burtleburtle.net. You may use this
279 code any way you wish, private, educational, or commercial. It's free.
281 Use for hash table lookup, or anything where one collision in 2^^32 is
282 acceptable. Do NOT use for cryptographic purposes.
283 -------------------------------------------------------------------------------
286 static uint32_t hashlittle( const void *key, size_t length, uint32_t initval)
288 uint32_t a,b,c; /* internal state */
289 union { const void *ptr; size_t i; } u; /* needed for Mac Powerbook G4 */
291 /* Set up the internal state */
292 a = b = c = 0xdeadbeef + ((uint32_t)length) + initval;
295 if (HASH_LITTLE_ENDIAN && ((u.i & 0x3) == 0)) {
296 const uint32_t *k = (const uint32_t *)key; /* read 32-bit chunks */
301 /*------ all but last block: aligned reads and affect 32 bits of (a,b,c) */
312 /*----------------------------- handle the last (probably partial) block */
314 * "k[2]&0xffffff" actually reads beyond the end of the string, but
315 * then masks off the part it's not allowed to read. Because the
316 * string is aligned, the masked-off tail is in the same word as the
317 * rest of the string. Every machine with memory protection I've seen
318 * does it on word boundaries, so is OK with this. But VALGRIND will
319 * still catch it and complain. The masking trick does make the hash
320 * noticably faster for short strings (like English words).
326 case 12: c+=k[2]; b+=k[1]; a+=k[0]; break;
327 case 11: c+=k[2]&0xffffff; b+=k[1]; a+=k[0]; break;
328 case 10: c+=k[2]&0xffff; b+=k[1]; a+=k[0]; break;
329 case 9 : c+=k[2]&0xff; b+=k[1]; a+=k[0]; break;
330 case 8 : b+=k[1]; a+=k[0]; break;
331 case 7 : b+=k[1]&0xffffff; a+=k[0]; break;
332 case 6 : b+=k[1]&0xffff; a+=k[0]; break;
333 case 5 : b+=k[1]&0xff; a+=k[0]; break;
334 case 4 : a+=k[0]; break;
335 case 3 : a+=k[0]&0xffffff; break;
336 case 2 : a+=k[0]&0xffff; break;
337 case 1 : a+=k[0]&0xff; break;
338 case 0 : return c; /* zero length strings require no mixing */
341 #else /* make valgrind happy */
343 k8 = (const uint8_t *)k;
346 case 12: c+=k[2]; b+=k[1]; a+=k[0]; break;
347 case 11: c+=((uint32_t)k8[10])<<16; /* fall through */
348 case 10: c+=((uint32_t)k8[9])<<8; /* fall through */
349 case 9 : c+=k8[8]; /* fall through */
350 case 8 : b+=k[1]; a+=k[0]; break;
351 case 7 : b+=((uint32_t)k8[6])<<16; /* fall through */
352 case 6 : b+=((uint32_t)k8[5])<<8; /* fall through */
353 case 5 : b+=k8[4]; /* fall through */
354 case 4 : a+=k[0]; break;
355 case 3 : a+=((uint32_t)k8[2])<<16; /* fall through */
356 case 2 : a+=((uint32_t)k8[1])<<8; /* fall through */
357 case 1 : a+=k8[0]; break;
361 #endif /* !valgrind */
363 } else if (HASH_LITTLE_ENDIAN && ((u.i & 0x1) == 0)) {
364 const uint16_t *k = (const uint16_t *)key; /* read 16-bit chunks */
367 /*--------------- all but last block: aligned reads and different mixing */
370 a += k[0] + (((uint32_t)k[1])<<16);
371 b += k[2] + (((uint32_t)k[3])<<16);
372 c += k[4] + (((uint32_t)k[5])<<16);
378 /*----------------------------- handle the last (probably partial) block */
379 k8 = (const uint8_t *)k;
382 case 12: c+=k[4]+(((uint32_t)k[5])<<16);
383 b+=k[2]+(((uint32_t)k[3])<<16);
384 a+=k[0]+(((uint32_t)k[1])<<16);
386 case 11: c+=((uint32_t)k8[10])<<16; /* fall through */
388 b+=k[2]+(((uint32_t)k[3])<<16);
389 a+=k[0]+(((uint32_t)k[1])<<16);
391 case 9 : c+=k8[8]; /* fall through */
392 case 8 : b+=k[2]+(((uint32_t)k[3])<<16);
393 a+=k[0]+(((uint32_t)k[1])<<16);
395 case 7 : b+=((uint32_t)k8[6])<<16; /* fall through */
397 a+=k[0]+(((uint32_t)k[1])<<16);
399 case 5 : b+=k8[4]; /* fall through */
400 case 4 : a+=k[0]+(((uint32_t)k[1])<<16);
402 case 3 : a+=((uint32_t)k8[2])<<16; /* fall through */
407 case 0 : return c; /* zero length requires no mixing */
410 } else { /* need to read the key one byte at a time */
411 const uint8_t *k = (const uint8_t *)key;
413 /*--------------- all but the last block: affect some 32 bits of (a,b,c) */
417 a += ((uint32_t)k[1])<<8;
418 a += ((uint32_t)k[2])<<16;
419 a += ((uint32_t)k[3])<<24;
421 b += ((uint32_t)k[5])<<8;
422 b += ((uint32_t)k[6])<<16;
423 b += ((uint32_t)k[7])<<24;
425 c += ((uint32_t)k[9])<<8;
426 c += ((uint32_t)k[10])<<16;
427 c += ((uint32_t)k[11])<<24;
433 /*-------------------------------- last block: affect all 32 bits of (c) */
434 switch(length) /* all the case statements fall through */
436 case 12: c+=((uint32_t)k[11])<<24;
437 case 11: c+=((uint32_t)k[10])<<16;
438 case 10: c+=((uint32_t)k[9])<<8;
440 case 8 : b+=((uint32_t)k[7])<<24;
441 case 7 : b+=((uint32_t)k[6])<<16;
442 case 6 : b+=((uint32_t)k[5])<<8;
444 case 4 : a+=((uint32_t)k[3])<<24;
445 case 3 : a+=((uint32_t)k[2])<<16;
446 case 2 : a+=((uint32_t)k[1])<<8;
459 * hashlittle2: return 2 32-bit hash values
461 * This is identical to hashlittle(), except it returns two 32-bit hash
462 * values instead of just one. This is good enough for hash table
463 * lookup with 2^^64 buckets, or if you want a second hash if you're not
464 * happy with the first, or if you want a probably-unique 64-bit ID for
465 * the key. *pc is better mixed than *pb, so use *pc first. If you want
466 * a 64-bit value do something like "*pc + (((uint64_t)*pb)<<32)".
469 const void *key, /* the key to hash */
470 size_t length, /* length of the key */
471 uint32_t *pc, /* IN: primary initval, OUT: primary hash */
472 uint32_t *pb) /* IN: secondary initval, OUT: secondary hash */
474 uint32_t a,b,c; /* internal state */
475 union { const void *ptr; size_t i; } u; /* needed for Mac Powerbook G4 */
477 /* Set up the internal state */
478 a = b = c = 0xdeadbeef + ((uint32_t)length) + *pc;
482 if (HASH_LITTLE_ENDIAN && ((u.i & 0x3) == 0)) {
483 const uint32_t *k = (const uint32_t *)key; /* read 32-bit chunks */
486 /*------ all but last block: aligned reads and affect 32 bits of (a,b,c) */
497 /*----------------------------- handle the last (probably partial) block */
499 * "k[2]&0xffffff" actually reads beyond the end of the string, but
500 * then masks off the part it's not allowed to read. Because the
501 * string is aligned, the masked-off tail is in the same word as the
502 * rest of the string. Every machine with memory protection I've seen
503 * does it on word boundaries, so is OK with this. But VALGRIND will
504 * still catch it and complain. The masking trick does make the hash
505 * noticably faster for short strings (like English words).
511 case 12: c+=k[2]; b+=k[1]; a+=k[0]; break;
512 case 11: c+=k[2]&0xffffff; b+=k[1]; a+=k[0]; break;
513 case 10: c+=k[2]&0xffff; b+=k[1]; a+=k[0]; break;
514 case 9 : c+=k[2]&0xff; b+=k[1]; a+=k[0]; break;
515 case 8 : b+=k[1]; a+=k[0]; break;
516 case 7 : b+=k[1]&0xffffff; a+=k[0]; break;
517 case 6 : b+=k[1]&0xffff; a+=k[0]; break;
518 case 5 : b+=k[1]&0xff; a+=k[0]; break;
519 case 4 : a+=k[0]; break;
520 case 3 : a+=k[0]&0xffffff; break;
521 case 2 : a+=k[0]&0xffff; break;
522 case 1 : a+=k[0]&0xff; break;
523 case 0 : *pc=c; *pb=b; return; /* zero length strings require no mixing */
526 #else /* make valgrind happy */
528 k8 = (const uint8_t *)k;
531 case 12: c+=k[2]; b+=k[1]; a+=k[0]; break;
532 case 11: c+=((uint32_t)k8[10])<<16; /* fall through */
533 case 10: c+=((uint32_t)k8[9])<<8; /* fall through */
534 case 9 : c+=k8[8]; /* fall through */
535 case 8 : b+=k[1]; a+=k[0]; break;
536 case 7 : b+=((uint32_t)k8[6])<<16; /* fall through */
537 case 6 : b+=((uint32_t)k8[5])<<8; /* fall through */
538 case 5 : b+=k8[4]; /* fall through */
539 case 4 : a+=k[0]; break;
540 case 3 : a+=((uint32_t)k8[2])<<16; /* fall through */
541 case 2 : a+=((uint32_t)k8[1])<<8; /* fall through */
542 case 1 : a+=k8[0]; break;
543 case 0 : *pc=c; *pb=b; return; /* zero length strings require no mixing */
546 #endif /* !valgrind */
548 } else if (HASH_LITTLE_ENDIAN && ((u.i & 0x1) == 0)) {
549 const uint16_t *k = (const uint16_t *)key; /* read 16-bit chunks */
552 /*--------------- all but last block: aligned reads and different mixing */
555 a += k[0] + (((uint32_t)k[1])<<16);
556 b += k[2] + (((uint32_t)k[3])<<16);
557 c += k[4] + (((uint32_t)k[5])<<16);
563 /*----------------------------- handle the last (probably partial) block */
564 k8 = (const uint8_t *)k;
567 case 12: c+=k[4]+(((uint32_t)k[5])<<16);
568 b+=k[2]+(((uint32_t)k[3])<<16);
569 a+=k[0]+(((uint32_t)k[1])<<16);
571 case 11: c+=((uint32_t)k8[10])<<16; /* fall through */
573 b+=k[2]+(((uint32_t)k[3])<<16);
574 a+=k[0]+(((uint32_t)k[1])<<16);
576 case 9 : c+=k8[8]; /* fall through */
577 case 8 : b+=k[2]+(((uint32_t)k[3])<<16);
578 a+=k[0]+(((uint32_t)k[1])<<16);
580 case 7 : b+=((uint32_t)k8[6])<<16; /* fall through */
582 a+=k[0]+(((uint32_t)k[1])<<16);
584 case 5 : b+=k8[4]; /* fall through */
585 case 4 : a+=k[0]+(((uint32_t)k[1])<<16);
587 case 3 : a+=((uint32_t)k8[2])<<16; /* fall through */
592 case 0 : *pc=c; *pb=b; return; /* zero length strings require no mixing */
595 } else { /* need to read the key one byte at a time */
596 const uint8_t *k = (const uint8_t *)key;
598 /*--------------- all but the last block: affect some 32 bits of (a,b,c) */
602 a += ((uint32_t)k[1])<<8;
603 a += ((uint32_t)k[2])<<16;
604 a += ((uint32_t)k[3])<<24;
606 b += ((uint32_t)k[5])<<8;
607 b += ((uint32_t)k[6])<<16;
608 b += ((uint32_t)k[7])<<24;
610 c += ((uint32_t)k[9])<<8;
611 c += ((uint32_t)k[10])<<16;
612 c += ((uint32_t)k[11])<<24;
618 /*-------------------------------- last block: affect all 32 bits of (c) */
619 switch(length) /* all the case statements fall through */
621 case 12: c+=((uint32_t)k[11])<<24;
622 case 11: c+=((uint32_t)k[10])<<16;
623 case 10: c+=((uint32_t)k[9])<<8;
625 case 8 : b+=((uint32_t)k[7])<<24;
626 case 7 : b+=((uint32_t)k[6])<<16;
627 case 6 : b+=((uint32_t)k[5])<<8;
629 case 4 : a+=((uint32_t)k[3])<<24;
630 case 3 : a+=((uint32_t)k[2])<<16;
631 case 2 : a+=((uint32_t)k[1])<<8;
634 case 0 : *pc=c; *pb=b; return; /* zero length strings require no mixing */
646 * This is the same as hash_word() on big-endian machines. It is different
647 * from hashlittle() on all machines. hashbig() takes advantage of
648 * big-endian byte ordering.
650 static uint32_t hashbig( const void *key, size_t length, uint32_t initval)
653 union { const void *ptr; size_t i; } u; /* to cast key to (size_t) happily */
655 /* Set up the internal state */
656 a = b = c = 0xdeadbeef + ((uint32_t)length) + initval;
659 if (HASH_BIG_ENDIAN && ((u.i & 0x3) == 0)) {
660 const uint32_t *k = (const uint32_t *)key; /* read 32-bit chunks */
665 /*------ all but last block: aligned reads and affect 32 bits of (a,b,c) */
676 /*----------------------------- handle the last (probably partial) block */
678 * "k[2]<<8" actually reads beyond the end of the string, but
679 * then shifts out the part it's not allowed to read. Because the
680 * string is aligned, the illegal read is in the same word as the
681 * rest of the string. Every machine with memory protection I've seen
682 * does it on word boundaries, so is OK with this. But VALGRIND will
683 * still catch it and complain. The masking trick does make the hash
684 * noticably faster for short strings (like English words).
690 case 12: c+=k[2]; b+=k[1]; a+=k[0]; break;
691 case 11: c+=k[2]&0xffffff00; b+=k[1]; a+=k[0]; break;
692 case 10: c+=k[2]&0xffff0000; b+=k[1]; a+=k[0]; break;
693 case 9 : c+=k[2]&0xff000000; b+=k[1]; a+=k[0]; break;
694 case 8 : b+=k[1]; a+=k[0]; break;
695 case 7 : b+=k[1]&0xffffff00; a+=k[0]; break;
696 case 6 : b+=k[1]&0xffff0000; a+=k[0]; break;
697 case 5 : b+=k[1]&0xff000000; a+=k[0]; break;
698 case 4 : a+=k[0]; break;
699 case 3 : a+=k[0]&0xffffff00; break;
700 case 2 : a+=k[0]&0xffff0000; break;
701 case 1 : a+=k[0]&0xff000000; break;
702 case 0 : return c; /* zero length strings require no mixing */
705 #else /* make valgrind happy */
707 k8 = (const uint8_t *)k;
708 switch(length) /* all the case statements fall through */
710 case 12: c+=k[2]; b+=k[1]; a+=k[0]; break;
711 case 11: c+=((uint32_t)k8[10])<<8; /* fall through */
712 case 10: c+=((uint32_t)k8[9])<<16; /* fall through */
713 case 9 : c+=((uint32_t)k8[8])<<24; /* fall through */
714 case 8 : b+=k[1]; a+=k[0]; break;
715 case 7 : b+=((uint32_t)k8[6])<<8; /* fall through */
716 case 6 : b+=((uint32_t)k8[5])<<16; /* fall through */
717 case 5 : b+=((uint32_t)k8[4])<<24; /* fall through */
718 case 4 : a+=k[0]; break;
719 case 3 : a+=((uint32_t)k8[2])<<8; /* fall through */
720 case 2 : a+=((uint32_t)k8[1])<<16; /* fall through */
721 case 1 : a+=((uint32_t)k8[0])<<24; break;
725 #endif /* !VALGRIND */
727 } else { /* need to read the key one byte at a time */
728 const uint8_t *k = (const uint8_t *)key;
730 /*--------------- all but the last block: affect some 32 bits of (a,b,c) */
733 a += ((uint32_t)k[0])<<24;
734 a += ((uint32_t)k[1])<<16;
735 a += ((uint32_t)k[2])<<8;
736 a += ((uint32_t)k[3]);
737 b += ((uint32_t)k[4])<<24;
738 b += ((uint32_t)k[5])<<16;
739 b += ((uint32_t)k[6])<<8;
740 b += ((uint32_t)k[7]);
741 c += ((uint32_t)k[8])<<24;
742 c += ((uint32_t)k[9])<<16;
743 c += ((uint32_t)k[10])<<8;
744 c += ((uint32_t)k[11]);
750 /*-------------------------------- last block: affect all 32 bits of (c) */
751 switch(length) /* all the case statements fall through */
754 case 11: c+=((uint32_t)k[10])<<8;
755 case 10: c+=((uint32_t)k[9])<<16;
756 case 9 : c+=((uint32_t)k[8])<<24;
758 case 7 : b+=((uint32_t)k[6])<<8;
759 case 6 : b+=((uint32_t)k[5])<<16;
760 case 5 : b+=((uint32_t)k[4])<<24;
762 case 3 : a+=((uint32_t)k[2])<<8;
763 case 2 : a+=((uint32_t)k[1])<<16;
764 case 1 : a+=((uint32_t)k[0])<<24;
774 /* I basically use hashlittle here, but use native endian within each
775 * element. This delivers least-surprise: hash such as "int arr[] = {
776 * 1, 2 }; hash_stable(arr, 2, 0);" will be the same on big and little
777 * endian machines, even though a bytewise hash wouldn't be. */
778 uint32_t hash_stable_64(const void *key, size_t n, uint32_t base)
780 const uint64_t *k = key;
783 /* Set up the internal state */
784 a = b = c = 0xdeadbeef + ((uint32_t)n*8) + base;
788 b += (uint32_t)(k[0] >> 32);
791 a += (uint32_t)(k[1] >> 32);
793 c += (uint32_t)(k[2] >> 32);
801 b += (uint32_t)(k[0] >> 32);
804 a += (uint32_t)(k[1] >> 32);
808 b += (uint32_t)(k[0] >> 32);
817 uint32_t hash_stable_32(const void *key, size_t n, uint32_t base)
819 const uint32_t *k = key;
822 /* Set up the internal state */
823 a = b = c = 0xdeadbeef + ((uint32_t)n*4) + base;
847 uint32_t hash_stable_16(const void *key, size_t n, uint32_t base)
849 const uint16_t *k = key;
852 /* Set up the internal state */
853 a = b = c = 0xdeadbeef + ((uint32_t)n*2) + base;
856 a += (uint32_t)k[0] + ((uint32_t)k[1] << 16);
857 b += (uint32_t)k[2] + ((uint32_t)k[3] << 16);
858 c += (uint32_t)k[4] + ((uint32_t)k[5] << 16);
869 b += ((uint32_t)k[3] << 16);
873 a += ((uint32_t)k[1] << 16);
884 uint32_t hash_stable_8(const void *key, size_t n, uint32_t base)
886 return hashlittle(key, n, base);
889 uint32_t hash_any(const void *key, size_t length, uint32_t base)
892 return hashbig(key, length, base);
894 return hashlittle(key, length, base);
899 /* used for timings */
908 for (i=0; i<256; ++i) buf[i] = 'x';
911 h = hashlittle(&buf[0],1,h);
914 if (z-a > 0) printf("time %d %.8x\n", z-a, h);
917 /* check that every input bit changes every output bit half the time */
924 uint8_t qa[MAXLEN+1], qb[MAXLEN+2], *a = &qa[0], *b = &qb[1];
925 uint32_t c[HASHSTATE], d[HASHSTATE], i=0, j=0, k, l, m=0, z;
926 uint32_t e[HASHSTATE],f[HASHSTATE],g[HASHSTATE],h[HASHSTATE];
927 uint32_t x[HASHSTATE],y[HASHSTATE];
930 printf("No more than %d trials should ever be needed \n",MAXPAIR/2);
931 for (hlen=0; hlen < MAXLEN; ++hlen)
934 for (i=0; i<hlen; ++i) /*----------------------- for each input byte, */
936 for (j=0; j<8; ++j) /*------------------------ for each input bit, */
938 for (m=1; m<8; ++m) /*------------ for serveral possible initvals, */
940 for (l=0; l<HASHSTATE; ++l)
941 e[l]=f[l]=g[l]=h[l]=x[l]=y[l]=~((uint32_t)0);
943 /*---- check that every output bit is affected by that input bit */
944 for (k=0; k<MAXPAIR; k+=2)
947 /* keys have one bit different */
948 for (l=0; l<hlen+1; ++l) {a[l] = b[l] = (uint8_t)0;}
949 /* have a and b be two keys differing in only one bit */
952 c[0] = hashlittle(a, hlen, m);
954 b[i] ^= ((k+1)>>(8-j));
955 d[0] = hashlittle(b, hlen, m);
956 /* check every bit is 1, 0, set, and not set at least once */
957 for (l=0; l<HASHSTATE; ++l)
960 f[l] &= ~(c[l]^d[l]);
965 if (e[l]|f[l]|g[l]|h[l]|x[l]|y[l]) finished=0;
972 printf("Some bit didn't change: ");
973 printf("%.8x %.8x %.8x %.8x %.8x %.8x ",
974 e[0],f[0],g[0],h[0],x[0],y[0]);
975 printf("i %d j %d m %d len %d\n", i, j, m, hlen);
977 if (z==MAXPAIR) goto done;
984 printf("Mix success %2d bytes %2d initvals ",i,m);
985 printf("required %d trials\n", z/2);
991 /* Check for reading beyond the end of the buffer and alignment problems */
994 uint8_t buf[MAXLEN+20], *b;
996 uint8_t q[] = "This is the time for all good men to come to the aid of their country...";
998 uint8_t qq[] = "xThis is the time for all good men to come to the aid of their country...";
1000 uint8_t qqq[] = "xxThis is the time for all good men to come to the aid of their country...";
1002 uint8_t qqqq[] = "xxxThis is the time for all good men to come to the aid of their country...";
1006 printf("Endianness. These lines should all be the same (for values filled in):\n");
1007 printf("%.8x %.8x %.8x\n",
1008 hash_word((const uint32_t *)q, (sizeof(q)-1)/4, 13),
1009 hash_word((const uint32_t *)q, (sizeof(q)-5)/4, 13),
1010 hash_word((const uint32_t *)q, (sizeof(q)-9)/4, 13));
1012 printf("%.8x %.8x %.8x %.8x %.8x %.8x %.8x %.8x %.8x %.8x %.8x %.8x\n",
1013 hashlittle(p, sizeof(q)-1, 13), hashlittle(p, sizeof(q)-2, 13),
1014 hashlittle(p, sizeof(q)-3, 13), hashlittle(p, sizeof(q)-4, 13),
1015 hashlittle(p, sizeof(q)-5, 13), hashlittle(p, sizeof(q)-6, 13),
1016 hashlittle(p, sizeof(q)-7, 13), hashlittle(p, sizeof(q)-8, 13),
1017 hashlittle(p, sizeof(q)-9, 13), hashlittle(p, sizeof(q)-10, 13),
1018 hashlittle(p, sizeof(q)-11, 13), hashlittle(p, sizeof(q)-12, 13));
1020 printf("%.8x %.8x %.8x %.8x %.8x %.8x %.8x %.8x %.8x %.8x %.8x %.8x\n",
1021 hashlittle(p, sizeof(q)-1, 13), hashlittle(p, sizeof(q)-2, 13),
1022 hashlittle(p, sizeof(q)-3, 13), hashlittle(p, sizeof(q)-4, 13),
1023 hashlittle(p, sizeof(q)-5, 13), hashlittle(p, sizeof(q)-6, 13),
1024 hashlittle(p, sizeof(q)-7, 13), hashlittle(p, sizeof(q)-8, 13),
1025 hashlittle(p, sizeof(q)-9, 13), hashlittle(p, sizeof(q)-10, 13),
1026 hashlittle(p, sizeof(q)-11, 13), hashlittle(p, sizeof(q)-12, 13));
1028 printf("%.8x %.8x %.8x %.8x %.8x %.8x %.8x %.8x %.8x %.8x %.8x %.8x\n",
1029 hashlittle(p, sizeof(q)-1, 13), hashlittle(p, sizeof(q)-2, 13),
1030 hashlittle(p, sizeof(q)-3, 13), hashlittle(p, sizeof(q)-4, 13),
1031 hashlittle(p, sizeof(q)-5, 13), hashlittle(p, sizeof(q)-6, 13),
1032 hashlittle(p, sizeof(q)-7, 13), hashlittle(p, sizeof(q)-8, 13),
1033 hashlittle(p, sizeof(q)-9, 13), hashlittle(p, sizeof(q)-10, 13),
1034 hashlittle(p, sizeof(q)-11, 13), hashlittle(p, sizeof(q)-12, 13));
1036 printf("%.8x %.8x %.8x %.8x %.8x %.8x %.8x %.8x %.8x %.8x %.8x %.8x\n",
1037 hashlittle(p, sizeof(q)-1, 13), hashlittle(p, sizeof(q)-2, 13),
1038 hashlittle(p, sizeof(q)-3, 13), hashlittle(p, sizeof(q)-4, 13),
1039 hashlittle(p, sizeof(q)-5, 13), hashlittle(p, sizeof(q)-6, 13),
1040 hashlittle(p, sizeof(q)-7, 13), hashlittle(p, sizeof(q)-8, 13),
1041 hashlittle(p, sizeof(q)-9, 13), hashlittle(p, sizeof(q)-10, 13),
1042 hashlittle(p, sizeof(q)-11, 13), hashlittle(p, sizeof(q)-12, 13));
1045 /* check that hashlittle2 and hashlittle produce the same results */
1047 hashlittle2(q, sizeof(q), &i, &j);
1048 if (hashlittle(q, sizeof(q), 47) != i)
1049 printf("hashlittle2 and hashlittle mismatch\n");
1051 /* check that hash_word2 and hash_word produce the same results */
1054 hash_word2(&len, 1, &i, &j);
1055 if (hash_word(&len, 1, 47) != i)
1056 printf("hash_word2 and hash_word mismatch %x %x\n",
1057 i, hash_word(&len, 1, 47));
1059 /* check hashlittle doesn't read before or after the ends of the string */
1060 for (h=0, b=buf+1; h<8; ++h, ++b)
1062 for (i=0; i<MAXLEN; ++i)
1065 for (j=0; j<i; ++j) *(b+j)=0;
1067 /* these should all be equal */
1068 ref = hashlittle(b, len, (uint32_t)1);
1071 x = hashlittle(b, len, (uint32_t)1);
1072 y = hashlittle(b, len, (uint32_t)1);
1073 if ((ref != x) || (ref != y))
1075 printf("alignment error: %.8x %.8x %.8x %d %d\n",ref,x,y,
1082 /* check for problems with nulls */
1086 uint32_t h,i,state[HASHSTATE];
1090 for (i=0; i<HASHSTATE; ++i) state[i] = 1;
1091 printf("These should all be different\n");
1092 for (i=0, h=0; i<8; ++i)
1094 h = hashlittle(buf, 0, h);
1095 printf("%2ld 0-byte strings, hash is %.8x\n", i, h);
1102 driver1(); /* test that the key is hashed: used for timings */
1103 driver2(); /* test that whole key is hashed thoroughly */
1104 driver3(); /* test that nothing but the key is hashed */
1105 driver4(); /* test hashing multiple buffers (all buffers are null) */
1109 #endif /* SELF_TEST */