2 Copyright (c) 2003-2010, Troy D. Hanson http://uthash.sourceforge.net
5 Redistribution and use in source and binary forms, with or without
6 modification, are permitted provided that the following conditions are met:
8 * Redistributions of source code must retain the above copyright
9 notice, this list of conditions and the following disclaimer.
11 THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS
12 IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
13 TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A
14 PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
15 OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
16 EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
17 PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
18 PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
19 LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
20 NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
21 SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
27 #include <string.h> /* memcmp,strlen */
28 #include <stddef.h> /* ptrdiff_t */
30 /* These macros use decltype or the earlier __typeof GNU extension.
31 As decltype is only available in newer compilers (VS2010 or gcc 4.3+
32 when compiling c++ source) this code uses whatever method is needed
33 or, for VS2008 where neither is available, uses casting workarounds. */
34 #ifdef _MSC_VER /* MS compiler */
35 #if _MSC_VER >= 1600 && defined(__cplusplus) /* VS2010 or newer in C++ mode */
36 #define DECLTYPE(x) (decltype(x))
37 #else /* VS2008 or older (or VS2010 in C mode) */
41 #else /* GNU, Sun and other compilers */
42 #define DECLTYPE(x) (__typeof(x))
46 #define DECLTYPE_ASSIGN(dst,src) \
48 char **_da_dst = (char**)(&(dst)); \
49 *_da_dst = (char*)(src); \
52 #define DECLTYPE_ASSIGN(dst,src) \
54 (dst) = DECLTYPE(dst)(src); \
58 /* a number of the hash function use uint32_t which isn't defined on win32 */
60 typedef unsigned int uint32_t;
62 #include <inttypes.h> /* uint32_t */
65 #define UTHASH_VERSION 1.9.3
68 #define uthash_fatal(msg) exit(-1) /* fatal error (out of memory,etc) */
70 #define uthash_malloc(sz) malloc(sz) /* malloc fcn */
71 #define uthash_free(ptr,sz) free(ptr) /* free fcn */
73 #define uthash_noexpand_fyi(tbl) /* can be defined to log noexpand */
74 #define uthash_expand_fyi(tbl) /* can be defined to log expands */
76 /* initial number of buckets */
77 #define HASH_INITIAL_NUM_BUCKETS 32 /* initial number of buckets */
78 #define HASH_INITIAL_NUM_BUCKETS_LOG2 5 /* lg2 of initial number of buckets */
79 #define HASH_BKT_CAPACITY_THRESH 10 /* expand when bucket count reaches */
81 /* calculate the element whose hash handle address is hhe */
82 #define ELMT_FROM_HH(tbl,hhp) ((void*)(((char*)(hhp)) - ((tbl)->hho)))
84 #define HASH_FIND(hh,head,keyptr,keylen,out) \
86 unsigned _hf_bkt,_hf_hashv; \
89 HASH_FCN(keyptr,keylen, (head)->hh.tbl->num_buckets, _hf_hashv, _hf_bkt); \
90 if (HASH_BLOOM_TEST((head)->hh.tbl, _hf_hashv)) { \
91 HASH_FIND_IN_BKT((head)->hh.tbl, hh, (head)->hh.tbl->buckets[ _hf_bkt ], \
98 #define HASH_BLOOM_BITLEN (1ULL << HASH_BLOOM)
99 #define HASH_BLOOM_BYTELEN (HASH_BLOOM_BITLEN/8) + ((HASH_BLOOM_BITLEN%8) ? 1:0)
100 #define HASH_BLOOM_MAKE(tbl) \
102 (tbl)->bloom_nbits = HASH_BLOOM; \
103 (tbl)->bloom_bv = (uint8_t*)uthash_malloc(HASH_BLOOM_BYTELEN); \
104 if (!((tbl)->bloom_bv)) { uthash_fatal( "out of memory"); } \
105 memset((tbl)->bloom_bv, 0, HASH_BLOOM_BYTELEN); \
106 (tbl)->bloom_sig = HASH_BLOOM_SIGNATURE; \
109 #define HASH_BLOOM_FREE(tbl) \
111 uthash_free((tbl)->bloom_bv, HASH_BLOOM_BYTELEN); \
114 #define HASH_BLOOM_BITSET(bv,idx) (bv[(idx)/8] |= (1U << ((idx)%8)))
115 #define HASH_BLOOM_BITTEST(bv,idx) (bv[(idx)/8] & (1U << ((idx)%8)))
117 #define HASH_BLOOM_ADD(tbl,hashv) \
118 HASH_BLOOM_BITSET((tbl)->bloom_bv, (hashv & (uint32_t)((1ULL << (tbl)->bloom_nbits) - 1)))
120 #define HASH_BLOOM_TEST(tbl,hashv) \
121 HASH_BLOOM_BITTEST((tbl)->bloom_bv, (hashv & (uint32_t)((1ULL << (tbl)->bloom_nbits) - 1)))
124 #define HASH_BLOOM_MAKE(tbl)
125 #define HASH_BLOOM_FREE(tbl)
126 #define HASH_BLOOM_ADD(tbl,hashv)
127 #define HASH_BLOOM_TEST(tbl,hashv) (1)
130 #define HASH_MAKE_TABLE(hh,head) \
132 (head)->hh.tbl = (UT_hash_table*)uthash_malloc( \
133 sizeof(UT_hash_table)); \
134 if (!((head)->hh.tbl)) { uthash_fatal( "out of memory"); } \
135 memset((head)->hh.tbl, 0, sizeof(UT_hash_table)); \
136 (head)->hh.tbl->tail = &((head)->hh); \
137 (head)->hh.tbl->num_buckets = HASH_INITIAL_NUM_BUCKETS; \
138 (head)->hh.tbl->log2_num_buckets = HASH_INITIAL_NUM_BUCKETS_LOG2; \
139 (head)->hh.tbl->hho = (char*)(&(head)->hh) - (char*)(head); \
140 (head)->hh.tbl->buckets = (UT_hash_bucket*)uthash_malloc( \
141 HASH_INITIAL_NUM_BUCKETS*sizeof(struct UT_hash_bucket)); \
142 if (! (head)->hh.tbl->buckets) { uthash_fatal( "out of memory"); } \
143 memset((head)->hh.tbl->buckets, 0, \
144 HASH_INITIAL_NUM_BUCKETS*sizeof(struct UT_hash_bucket)); \
145 HASH_BLOOM_MAKE((head)->hh.tbl); \
146 (head)->hh.tbl->signature = HASH_SIGNATURE; \
149 #define HASH_ADD(hh,head,fieldname,keylen_in,add) \
150 HASH_ADD_KEYPTR(hh,head,&add->fieldname,keylen_in,add)
152 #define HASH_ADD_KEYPTR(hh,head,keyptr,keylen_in,add) \
155 (add)->hh.next = NULL; \
156 (add)->hh.key = (char*)keyptr; \
157 (add)->hh.keylen = keylen_in; \
160 (head)->hh.prev = NULL; \
161 HASH_MAKE_TABLE(hh,head); \
163 (head)->hh.tbl->tail->next = (add); \
164 (add)->hh.prev = ELMT_FROM_HH((head)->hh.tbl, (head)->hh.tbl->tail); \
165 (head)->hh.tbl->tail = &((add)->hh); \
167 (head)->hh.tbl->num_items++; \
168 (add)->hh.tbl = (head)->hh.tbl; \
169 HASH_FCN(keyptr,keylen_in, (head)->hh.tbl->num_buckets, \
170 (add)->hh.hashv, _ha_bkt); \
171 HASH_ADD_TO_BKT((head)->hh.tbl->buckets[_ha_bkt],&(add)->hh); \
172 HASH_BLOOM_ADD((head)->hh.tbl,(add)->hh.hashv); \
173 HASH_EMIT_KEY(hh,head,keyptr,keylen_in); \
174 HASH_FSCK(hh,head); \
177 #define HASH_TO_BKT( hashv, num_bkts, bkt ) \
179 bkt = ((hashv) & ((num_bkts) - 1)); \
182 /* delete "delptr" from the hash table.
183 * "the usual" patch-up process for the app-order doubly-linked-list.
184 * The use of _hd_hh_del below deserves special explanation.
185 * These used to be expressed using (delptr) but that led to a bug
186 * if someone used the same symbol for the head and deletee, like
187 * HASH_DELETE(hh,users,users);
188 * We want that to work, but by changing the head (users) below
189 * we were forfeiting our ability to further refer to the deletee (users)
190 * in the patch-up process. Solution: use scratch space to
191 * copy the deletee pointer, then the latter references are via that
192 * scratch pointer rather than through the repointed (users) symbol.
194 #define HASH_DELETE(hh,head,delptr) \
197 struct UT_hash_handle *_hd_hh_del; \
198 if ( ((delptr)->hh.prev == NULL) && ((delptr)->hh.next == NULL) ) { \
199 uthash_free((head)->hh.tbl->buckets, \
200 (head)->hh.tbl->num_buckets*sizeof(struct UT_hash_bucket) ); \
201 HASH_BLOOM_FREE((head)->hh.tbl); \
202 uthash_free((head)->hh.tbl, sizeof(UT_hash_table)); \
205 _hd_hh_del = &((delptr)->hh); \
206 if ((delptr) == ELMT_FROM_HH((head)->hh.tbl,(head)->hh.tbl->tail)) { \
207 (head)->hh.tbl->tail = \
208 (UT_hash_handle*)((char*)((delptr)->hh.prev) + \
209 (head)->hh.tbl->hho); \
211 if ((delptr)->hh.prev) { \
212 ((UT_hash_handle*)((char*)((delptr)->hh.prev) + \
213 (head)->hh.tbl->hho))->next = (delptr)->hh.next; \
215 DECLTYPE_ASSIGN(head,(delptr)->hh.next); \
217 if (_hd_hh_del->next) { \
218 ((UT_hash_handle*)((char*)_hd_hh_del->next + \
219 (head)->hh.tbl->hho))->prev = \
222 HASH_TO_BKT( _hd_hh_del->hashv, (head)->hh.tbl->num_buckets, _hd_bkt); \
223 HASH_DEL_IN_BKT(hh,(head)->hh.tbl->buckets[_hd_bkt], _hd_hh_del); \
224 (head)->hh.tbl->num_items--; \
226 HASH_FSCK(hh,head); \
229 /* convenience forms of HASH_FIND/HASH_ADD/HASH_DEL */
230 #define HASH_FIND_STR(head,findstr,out) \
231 HASH_FIND(hh,head,findstr,strlen(findstr),out)
232 #define HASH_ADD_STR(head,strfield,add) \
233 HASH_ADD(hh,head,strfield,strlen(add->strfield),add)
234 #define HASH_FIND_INT(head,findint,out) \
235 HASH_FIND(hh,head,findint,sizeof(int),out)
236 #define HASH_ADD_INT(head,intfield,add) \
237 HASH_ADD(hh,head,intfield,sizeof(int),add)
238 #define HASH_FIND_PTR(head,findptr,out) \
239 HASH_FIND(hh,head,findptr,sizeof(void *),out)
240 #define HASH_ADD_PTR(head,ptrfield,add) \
241 HASH_ADD(hh,head,ptrfield,sizeof(void *),add)
242 #define HASH_DEL(head,delptr) \
243 HASH_DELETE(hh,head,delptr)
245 /* HASH_FSCK checks hash integrity on every add/delete when HASH_DEBUG is defined.
246 * This is for uthash developer only; it compiles away if HASH_DEBUG isn't defined.
249 #define HASH_OOPS(...) do { fprintf(stderr,__VA_ARGS__); exit(-1); } while (0)
250 #define HASH_FSCK(hh,head) \
253 unsigned _count, _bkt_count; \
255 struct UT_hash_handle *_thh; \
258 for( _bkt_i = 0; _bkt_i < (head)->hh.tbl->num_buckets; _bkt_i++) { \
260 _thh = (head)->hh.tbl->buckets[_bkt_i].hh_head; \
263 if (_prev != (char*)(_thh->hh_prev)) { \
264 HASH_OOPS("invalid hh_prev %p, actual %p\n", \
265 _thh->hh_prev, _prev ); \
268 _prev = (char*)(_thh); \
269 _thh = _thh->hh_next; \
271 _count += _bkt_count; \
272 if ((head)->hh.tbl->buckets[_bkt_i].count != _bkt_count) { \
273 HASH_OOPS("invalid bucket count %d, actual %d\n", \
274 (head)->hh.tbl->buckets[_bkt_i].count, _bkt_count); \
277 if (_count != (head)->hh.tbl->num_items) { \
278 HASH_OOPS("invalid hh item count %d, actual %d\n", \
279 (head)->hh.tbl->num_items, _count ); \
281 /* traverse hh in app order; check next/prev integrity, count */ \
284 _thh = &(head)->hh; \
287 if (_prev !=(char*)(_thh->prev)) { \
288 HASH_OOPS("invalid prev %p, actual %p\n", \
289 _thh->prev, _prev ); \
291 _prev = (char*)ELMT_FROM_HH((head)->hh.tbl, _thh); \
292 _thh = ( _thh->next ? (UT_hash_handle*)((char*)(_thh->next) + \
293 (head)->hh.tbl->hho) : NULL ); \
295 if (_count != (head)->hh.tbl->num_items) { \
296 HASH_OOPS("invalid app item count %d, actual %d\n", \
297 (head)->hh.tbl->num_items, _count ); \
302 #define HASH_FSCK(hh,head)
305 /* When compiled with -DHASH_EMIT_KEYS, length-prefixed keys are emitted to
306 * the descriptor to which this macro is defined for tuning the hash function.
307 * The app can #include <unistd.h> to get the prototype for write(2). */
308 #ifdef HASH_EMIT_KEYS
309 #define HASH_EMIT_KEY(hh,head,keyptr,fieldlen) \
311 unsigned _klen = fieldlen; \
312 write(HASH_EMIT_KEYS, &_klen, sizeof(_klen)); \
313 write(HASH_EMIT_KEYS, keyptr, fieldlen); \
316 #define HASH_EMIT_KEY(hh,head,keyptr,fieldlen)
319 /* default to Jenkin's hash unless overridden e.g. DHASH_FUNCTION=HASH_SAX */
321 #define HASH_FCN HASH_FUNCTION
323 #define HASH_FCN HASH_JEN
326 /* The Bernstein hash function, used in Perl prior to v5.6 */
327 #define HASH_BER(key,keylen,num_bkts,hashv,bkt) \
329 unsigned _hb_keylen=keylen; \
330 char *_hb_key=(char*)(key); \
332 while (_hb_keylen--) { (hashv) = ((hashv) * 33) + *_hb_key++; } \
333 bkt = (hashv) & (num_bkts-1); \
336 /* SAX/FNV/OAT/JEN hash functions are macro variants of those listed at
337 * http://eternallyconfuzzled.com/tuts/algorithms/jsw_tut_hashing.aspx */
338 #define HASH_SAX(key,keylen,num_bkts,hashv,bkt) \
341 char *_hs_key=(char*)(key); \
343 for(_sx_i=0; _sx_i < keylen; _sx_i++) \
344 hashv ^= (hashv << 5) + (hashv >> 2) + _hs_key[_sx_i]; \
345 bkt = hashv & (num_bkts-1); \
348 #define HASH_FNV(key,keylen,num_bkts,hashv,bkt) \
351 char *_hf_key=(char*)(key); \
352 hashv = 2166136261UL; \
353 for(_fn_i=0; _fn_i < keylen; _fn_i++) \
354 hashv = (hashv * 16777619) ^ _hf_key[_fn_i]; \
355 bkt = hashv & (num_bkts-1); \
358 #define HASH_OAT(key,keylen,num_bkts,hashv,bkt) \
361 char *_ho_key=(char*)(key); \
363 for(_ho_i=0; _ho_i < keylen; _ho_i++) { \
364 hashv += _ho_key[_ho_i]; \
365 hashv += (hashv << 10); \
366 hashv ^= (hashv >> 6); \
368 hashv += (hashv << 3); \
369 hashv ^= (hashv >> 11); \
370 hashv += (hashv << 15); \
371 bkt = hashv & (num_bkts-1); \
374 #define HASH_JEN_MIX(a,b,c) \
376 a -= b; a -= c; a ^= ( c >> 13 ); \
377 b -= c; b -= a; b ^= ( a << 8 ); \
378 c -= a; c -= b; c ^= ( b >> 13 ); \
379 a -= b; a -= c; a ^= ( c >> 12 ); \
380 b -= c; b -= a; b ^= ( a << 16 ); \
381 c -= a; c -= b; c ^= ( b >> 5 ); \
382 a -= b; a -= c; a ^= ( c >> 3 ); \
383 b -= c; b -= a; b ^= ( a << 10 ); \
384 c -= a; c -= b; c ^= ( b >> 15 ); \
387 #define HASH_JEN(key,keylen,num_bkts,hashv,bkt) \
389 unsigned _hj_i,_hj_j,_hj_k; \
390 char *_hj_key=(char*)(key); \
391 hashv = 0xfeedbeef; \
392 _hj_i = _hj_j = 0x9e3779b9; \
394 while (_hj_k >= 12) { \
395 _hj_i += (_hj_key[0] + ( (unsigned)_hj_key[1] << 8 ) \
396 + ( (unsigned)_hj_key[2] << 16 ) \
397 + ( (unsigned)_hj_key[3] << 24 ) ); \
398 _hj_j += (_hj_key[4] + ( (unsigned)_hj_key[5] << 8 ) \
399 + ( (unsigned)_hj_key[6] << 16 ) \
400 + ( (unsigned)_hj_key[7] << 24 ) ); \
401 hashv += (_hj_key[8] + ( (unsigned)_hj_key[9] << 8 ) \
402 + ( (unsigned)_hj_key[10] << 16 ) \
403 + ( (unsigned)_hj_key[11] << 24 ) ); \
405 HASH_JEN_MIX(_hj_i, _hj_j, hashv); \
412 case 11: hashv += ( (unsigned)_hj_key[10] << 24 ); \
413 case 10: hashv += ( (unsigned)_hj_key[9] << 16 ); \
414 case 9: hashv += ( (unsigned)_hj_key[8] << 8 ); \
415 case 8: _hj_j += ( (unsigned)_hj_key[7] << 24 ); \
416 case 7: _hj_j += ( (unsigned)_hj_key[6] << 16 ); \
417 case 6: _hj_j += ( (unsigned)_hj_key[5] << 8 ); \
418 case 5: _hj_j += _hj_key[4]; \
419 case 4: _hj_i += ( (unsigned)_hj_key[3] << 24 ); \
420 case 3: _hj_i += ( (unsigned)_hj_key[2] << 16 ); \
421 case 2: _hj_i += ( (unsigned)_hj_key[1] << 8 ); \
422 case 1: _hj_i += _hj_key[0]; \
424 HASH_JEN_MIX(_hj_i, _hj_j, hashv); \
425 bkt = hashv & (num_bkts-1); \
428 /* The Paul Hsieh hash function */
430 #if (defined(__GNUC__) && defined(__i386__)) || defined(__WATCOMC__) \
431 || defined(_MSC_VER) || defined (__BORLANDC__) || defined (__TURBOC__)
432 #define get16bits(d) (*((const uint16_t *) (d)))
435 #if !defined (get16bits)
436 #define get16bits(d) ((((uint32_t)(((const uint8_t *)(d))[1])) << 8) \
437 +(uint32_t)(((const uint8_t *)(d))[0]) )
439 #define HASH_SFH(key,keylen,num_bkts,hashv,bkt) \
441 char *_sfh_key=(char*)(key); \
442 uint32_t _sfh_tmp, _sfh_len = keylen; \
444 int _sfh_rem = _sfh_len & 3; \
446 hashv = 0xcafebabe; \
449 for (;_sfh_len > 0; _sfh_len--) { \
450 hashv += get16bits (_sfh_key); \
451 _sfh_tmp = (get16bits (_sfh_key+2) << 11) ^ hashv; \
452 hashv = (hashv << 16) ^ _sfh_tmp; \
453 _sfh_key += 2*sizeof (uint16_t); \
454 hashv += hashv >> 11; \
457 /* Handle end cases */ \
458 switch (_sfh_rem) { \
459 case 3: hashv += get16bits (_sfh_key); \
460 hashv ^= hashv << 16; \
461 hashv ^= _sfh_key[sizeof (uint16_t)] << 18; \
462 hashv += hashv >> 11; \
464 case 2: hashv += get16bits (_sfh_key); \
465 hashv ^= hashv << 11; \
466 hashv += hashv >> 17; \
468 case 1: hashv += *_sfh_key; \
469 hashv ^= hashv << 10; \
470 hashv += hashv >> 1; \
473 /* Force "avalanching" of final 127 bits */ \
474 hashv ^= hashv << 3; \
475 hashv += hashv >> 5; \
476 hashv ^= hashv << 4; \
477 hashv += hashv >> 17; \
478 hashv ^= hashv << 25; \
479 hashv += hashv >> 6; \
480 bkt = hashv & (num_bkts-1); \
483 #ifdef HASH_USING_NO_STRICT_ALIASING
484 /* The MurmurHash exploits some CPU's (e.g. x86) tolerance for unaligned reads.
485 * For other types of CPU's (e.g. Sparc) an unaligned read causes a bus error.
486 * So MurmurHash comes in two versions, the faster unaligned one and the slower
487 * aligned one. We only use the faster one on CPU's where we know it's safe.
489 * Note the preprocessor built-in defines can be emitted using:
491 * gcc -m64 -dM -E - < /dev/null (on gcc)
492 * cc -## a.c (where a.c is a simple test file) (Sun Studio)
494 #if (defined(__i386__) || defined(__x86_64__))
495 #define HASH_MUR HASH_MUR_UNALIGNED
497 #define HASH_MUR HASH_MUR_ALIGNED
500 /* Appleby's MurmurHash fast version for unaligned-tolerant archs like i386 */
501 #define HASH_MUR_UNALIGNED(key,keylen,num_bkts,hashv,bkt) \
503 const unsigned int _mur_m = 0x5bd1e995; \
504 const int _mur_r = 24; \
505 hashv = 0xcafebabe ^ keylen; \
506 char *_mur_key = (char *)(key); \
507 uint32_t _mur_tmp, _mur_len = keylen; \
509 for (;_mur_len >= 4; _mur_len-=4) { \
510 _mur_tmp = *(uint32_t *)_mur_key; \
511 _mur_tmp *= _mur_m; \
512 _mur_tmp ^= _mur_tmp >> _mur_r; \
513 _mur_tmp *= _mur_m; \
521 case 3: hashv ^= _mur_key[2] << 16; \
522 case 2: hashv ^= _mur_key[1] << 8; \
523 case 1: hashv ^= _mur_key[0]; \
527 hashv ^= hashv >> 13; \
529 hashv ^= hashv >> 15; \
531 bkt = hashv & (num_bkts-1); \
534 /* Appleby's MurmurHash version for alignment-sensitive archs like Sparc */
535 #define HASH_MUR_ALIGNED(key,keylen,num_bkts,hashv,bkt) \
537 const unsigned int _mur_m = 0x5bd1e995; \
538 const int _mur_r = 24; \
539 hashv = 0xcafebabe ^ (keylen); \
540 char *_mur_key = (char *)(key); \
541 uint32_t _mur_len = keylen; \
542 int _mur_align = (int)_mur_key & 3; \
544 if (_mur_align && (_mur_len >= 4)) { \
545 unsigned _mur_t = 0, _mur_d = 0; \
546 switch(_mur_align) { \
547 case 1: _mur_t |= _mur_key[2] << 16; \
548 case 2: _mur_t |= _mur_key[1] << 8; \
549 case 3: _mur_t |= _mur_key[0]; \
551 _mur_t <<= (8 * _mur_align); \
552 _mur_key += 4-_mur_align; \
553 _mur_len -= 4-_mur_align; \
554 int _mur_sl = 8 * (4-_mur_align); \
555 int _mur_sr = 8 * _mur_align; \
557 for (;_mur_len >= 4; _mur_len-=4) { \
558 _mur_d = *(unsigned *)_mur_key; \
559 _mur_t = (_mur_t >> _mur_sr) | (_mur_d << _mur_sl); \
560 unsigned _mur_k = _mur_t; \
562 _mur_k ^= _mur_k >> _mur_r; \
570 if(_mur_len >= _mur_align) { \
571 switch(_mur_align) { \
572 case 3: _mur_d |= _mur_key[2] << 16; \
573 case 2: _mur_d |= _mur_key[1] << 8; \
574 case 1: _mur_d |= _mur_key[0]; \
576 unsigned _mur_k = (_mur_t >> _mur_sr) | (_mur_d << _mur_sl); \
578 _mur_k ^= _mur_k >> _mur_r; \
582 _mur_k += _mur_align; \
583 _mur_len -= _mur_align; \
587 case 3: hashv ^= _mur_key[2] << 16; \
588 case 2: hashv ^= _mur_key[1] << 8; \
589 case 1: hashv ^= _mur_key[0]; \
595 case 3: _mur_d ^= _mur_key[2] << 16; \
596 case 2: _mur_d ^= _mur_key[1] << 8; \
597 case 1: _mur_d ^= _mur_key[0]; \
598 case 0: hashv ^= (_mur_t >> _mur_sr) | (_mur_d << _mur_sl); \
603 hashv ^= hashv >> 13; \
605 hashv ^= hashv >> 15; \
607 for (;_mur_len >= 4; _mur_len-=4) { \
608 unsigned _mur_k = *(unsigned*)_mur_key; \
610 _mur_k ^= _mur_k >> _mur_r; \
618 case 3: hashv ^= _mur_key[2] << 16; \
619 case 2: hashv ^= _mur_key[1] << 8; \
620 case 1: hashv ^= _mur_key[0]; \
624 hashv ^= hashv >> 13; \
626 hashv ^= hashv >> 15; \
628 bkt = hashv & (num_bkts-1); \
630 #endif /* HASH_USING_NO_STRICT_ALIASING */
632 /* key comparison function; return 0 if keys equal */
633 #define HASH_KEYCMP(a,b,len) memcmp(a,b,len)
635 /* iterate over items in a known bucket to find desired item */
636 #define HASH_FIND_IN_BKT(tbl,hh,head,keyptr,keylen_in,out) \
638 if (head.hh_head) DECLTYPE_ASSIGN(out,ELMT_FROM_HH(tbl,head.hh_head)); \
641 if (out->hh.keylen == keylen_in) { \
642 if ((HASH_KEYCMP(out->hh.key,keyptr,keylen_in)) == 0) break; \
644 if (out->hh.hh_next) DECLTYPE_ASSIGN(out,ELMT_FROM_HH(tbl,out->hh.hh_next)); \
649 /* add an item to a bucket */
650 #define HASH_ADD_TO_BKT(head,addhh) \
653 (addhh)->hh_next = head.hh_head; \
654 (addhh)->hh_prev = NULL; \
655 if (head.hh_head) { (head).hh_head->hh_prev = (addhh); } \
656 (head).hh_head=addhh; \
657 if (head.count >= ((head.expand_mult+1) * HASH_BKT_CAPACITY_THRESH) \
658 && (addhh)->tbl->noexpand != 1) { \
659 HASH_EXPAND_BUCKETS((addhh)->tbl); \
663 /* remove an item from a given bucket */
664 #define HASH_DEL_IN_BKT(hh,head,hh_del) \
666 if ((head).hh_head == hh_del) { \
667 (head).hh_head = hh_del->hh_next; \
669 if (hh_del->hh_prev) { \
670 hh_del->hh_prev->hh_next = hh_del->hh_next; \
672 if (hh_del->hh_next) { \
673 hh_del->hh_next->hh_prev = hh_del->hh_prev; \
676 /* Bucket expansion has the effect of doubling the number of buckets
677 * and redistributing the items into the new buckets. Ideally the
678 * items will distribute more or less evenly into the new buckets
679 * (the extent to which this is true is a measure of the quality of
680 * the hash function as it applies to the key domain).
682 * With the items distributed into more buckets, the chain length
683 * (item count) in each bucket is reduced. Thus by expanding buckets
684 * the hash keeps a bound on the chain length. This bounded chain
685 * length is the essence of how a hash provides constant time lookup.
687 * The calculation of tbl->ideal_chain_maxlen below deserves some
688 * explanation. First, keep in mind that we're calculating the ideal
689 * maximum chain length based on the *new* (doubled) bucket count.
690 * In fractions this is just n/b (n=number of items,b=new num buckets).
691 * Since the ideal chain length is an integer, we want to calculate
692 * ceil(n/b). We don't depend on floating point arithmetic in this
693 * hash, so to calculate ceil(n/b) with integers we could write
695 * ceil(n/b) = (n/b) + ((n%b)?1:0)
697 * and in fact a previous version of this hash did just that.
698 * But now we have improved things a bit by recognizing that b is
699 * always a power of two. We keep its base 2 log handy (call it lb),
700 * so now we can write this with a bit shift and logical AND:
702 * ceil(n/b) = (n>>lb) + ( (n & (b-1)) ? 1:0)
705 #define HASH_EXPAND_BUCKETS(tbl) \
708 unsigned _he_bkt_i; \
709 struct UT_hash_handle *_he_thh, *_he_hh_nxt; \
710 UT_hash_bucket *_he_new_buckets, *_he_newbkt; \
711 _he_new_buckets = (UT_hash_bucket*)uthash_malloc( \
712 2 * tbl->num_buckets * sizeof(struct UT_hash_bucket)); \
713 if (!_he_new_buckets) { uthash_fatal( "out of memory"); } \
714 memset(_he_new_buckets, 0, \
715 2 * tbl->num_buckets * sizeof(struct UT_hash_bucket)); \
716 tbl->ideal_chain_maxlen = \
717 (tbl->num_items >> (tbl->log2_num_buckets+1)) + \
718 ((tbl->num_items & ((tbl->num_buckets*2)-1)) ? 1 : 0); \
719 tbl->nonideal_items = 0; \
720 for(_he_bkt_i = 0; _he_bkt_i < tbl->num_buckets; _he_bkt_i++) \
722 _he_thh = tbl->buckets[ _he_bkt_i ].hh_head; \
724 _he_hh_nxt = _he_thh->hh_next; \
725 HASH_TO_BKT( _he_thh->hashv, tbl->num_buckets*2, _he_bkt); \
726 _he_newbkt = &(_he_new_buckets[ _he_bkt ]); \
727 if (++(_he_newbkt->count) > tbl->ideal_chain_maxlen) { \
728 tbl->nonideal_items++; \
729 _he_newbkt->expand_mult = _he_newbkt->count / \
730 tbl->ideal_chain_maxlen; \
732 _he_thh->hh_prev = NULL; \
733 _he_thh->hh_next = _he_newbkt->hh_head; \
734 if (_he_newbkt->hh_head) _he_newbkt->hh_head->hh_prev = \
736 _he_newbkt->hh_head = _he_thh; \
737 _he_thh = _he_hh_nxt; \
740 uthash_free( tbl->buckets, tbl->num_buckets*sizeof(struct UT_hash_bucket) ); \
741 tbl->num_buckets *= 2; \
742 tbl->log2_num_buckets++; \
743 tbl->buckets = _he_new_buckets; \
744 tbl->ineff_expands = (tbl->nonideal_items > (tbl->num_items >> 1)) ? \
745 (tbl->ineff_expands+1) : 0; \
746 if (tbl->ineff_expands > 1) { \
748 uthash_noexpand_fyi(tbl); \
750 uthash_expand_fyi(tbl); \
753 /* This is an adaptation of Simon Tatham's O(n log(n)) mergesort */
754 /* Note that HASH_SORT assumes the hash handle name to be hh.
755 * HASH_SRT was added to allow the hash handle name to be passed in. */
756 #define HASH_SORT(head,cmpfcn) HASH_SRT(hh,head,cmpfcn)
757 #define HASH_SRT(hh,head,cmpfcn) \
760 unsigned _hs_looping,_hs_nmerges,_hs_insize,_hs_psize,_hs_qsize; \
761 struct UT_hash_handle *_hs_p, *_hs_q, *_hs_e, *_hs_list, *_hs_tail; \
765 _hs_list = &((head)->hh); \
766 while (_hs_looping) { \
775 for ( _hs_i = 0; _hs_i < _hs_insize; _hs_i++ ) { \
777 _hs_q = (UT_hash_handle*)((_hs_q->next) ? \
778 ((void*)((char*)(_hs_q->next) + \
779 (head)->hh.tbl->hho)) : NULL); \
780 if (! (_hs_q) ) break; \
782 _hs_qsize = _hs_insize; \
783 while ((_hs_psize > 0) || ((_hs_qsize > 0) && _hs_q )) { \
784 if (_hs_psize == 0) { \
786 _hs_q = (UT_hash_handle*)((_hs_q->next) ? \
787 ((void*)((char*)(_hs_q->next) + \
788 (head)->hh.tbl->hho)) : NULL); \
790 } else if ( (_hs_qsize == 0) || !(_hs_q) ) { \
792 _hs_p = (UT_hash_handle*)((_hs_p->next) ? \
793 ((void*)((char*)(_hs_p->next) + \
794 (head)->hh.tbl->hho)) : NULL); \
797 cmpfcn(DECLTYPE(head)(ELMT_FROM_HH((head)->hh.tbl,_hs_p)), \
798 DECLTYPE(head)(ELMT_FROM_HH((head)->hh.tbl,_hs_q))) \
801 _hs_p = (UT_hash_handle*)((_hs_p->next) ? \
802 ((void*)((char*)(_hs_p->next) + \
803 (head)->hh.tbl->hho)) : NULL); \
807 _hs_q = (UT_hash_handle*)((_hs_q->next) ? \
808 ((void*)((char*)(_hs_q->next) + \
809 (head)->hh.tbl->hho)) : NULL); \
813 _hs_tail->next = ((_hs_e) ? \
814 ELMT_FROM_HH((head)->hh.tbl,_hs_e) : NULL); \
818 _hs_e->prev = ((_hs_tail) ? \
819 ELMT_FROM_HH((head)->hh.tbl,_hs_tail) : NULL); \
824 _hs_tail->next = NULL; \
825 if ( _hs_nmerges <= 1 ) { \
827 (head)->hh.tbl->tail = _hs_tail; \
828 DECLTYPE_ASSIGN(head,ELMT_FROM_HH((head)->hh.tbl, _hs_list)); \
832 HASH_FSCK(hh,head); \
836 /* This function selects items from one hash into another hash.
837 * The end result is that the selected items have dual presence
838 * in both hashes. There is no copy of the items made; rather
839 * they are added into the new hash through a secondary hash
840 * hash handle that must be present in the structure. */
841 #define HASH_SELECT(hh_dst, dst, hh_src, src, cond) \
843 unsigned _src_bkt, _dst_bkt; \
844 void *_last_elt=NULL, *_elt; \
845 UT_hash_handle *_src_hh, *_dst_hh, *_last_elt_hh=NULL; \
846 ptrdiff_t _dst_hho = ((char*)(&(dst)->hh_dst) - (char*)(dst)); \
848 for(_src_bkt=0; _src_bkt < (src)->hh_src.tbl->num_buckets; _src_bkt++) { \
849 for(_src_hh = (src)->hh_src.tbl->buckets[_src_bkt].hh_head; \
851 _src_hh = _src_hh->hh_next) { \
852 _elt = ELMT_FROM_HH((src)->hh_src.tbl, _src_hh); \
854 _dst_hh = (UT_hash_handle*)(((char*)_elt) + _dst_hho); \
855 _dst_hh->key = _src_hh->key; \
856 _dst_hh->keylen = _src_hh->keylen; \
857 _dst_hh->hashv = _src_hh->hashv; \
858 _dst_hh->prev = _last_elt; \
859 _dst_hh->next = NULL; \
860 if (_last_elt_hh) { _last_elt_hh->next = _elt; } \
862 DECLTYPE_ASSIGN(dst,_elt); \
863 HASH_MAKE_TABLE(hh_dst,dst); \
865 _dst_hh->tbl = (dst)->hh_dst.tbl; \
867 HASH_TO_BKT(_dst_hh->hashv, _dst_hh->tbl->num_buckets, _dst_bkt); \
868 HASH_ADD_TO_BKT(_dst_hh->tbl->buckets[_dst_bkt],_dst_hh); \
869 (dst)->hh_dst.tbl->num_items++; \
871 _last_elt_hh = _dst_hh; \
876 HASH_FSCK(hh_dst,dst); \
879 #define HASH_CLEAR(hh,head) \
882 uthash_free((head)->hh.tbl->buckets, \
883 (head)->hh.tbl->num_buckets*sizeof(struct UT_hash_bucket)); \
884 uthash_free((head)->hh.tbl, sizeof(UT_hash_table)); \
890 #define HASH_ITER(hh,head,el,tmp) \
891 for((el)=(head), (*(char**)(&(tmp)))=(char*)((head)?(head)->hh.next:NULL); \
892 el; (el)=(tmp),(*(char**)(&(tmp)))=(char*)((tmp)?(tmp)->hh.next:NULL))
894 #define HASH_ITER(hh,head,el,tmp) \
895 for((el)=(head),(tmp)=DECLTYPE(el)((head)?(head)->hh.next:NULL); \
896 el; (el)=(tmp),(tmp)=DECLTYPE(el)((tmp)?(tmp)->hh.next:NULL))
899 /* obtain a count of items in the hash */
900 #define HASH_COUNT(head) HASH_CNT(hh,head)
901 #define HASH_CNT(hh,head) ((head)?((head)->hh.tbl->num_items):0)
903 typedef struct UT_hash_bucket
905 struct UT_hash_handle *hh_head;
908 /* expand_mult is normally set to 0. In this situation, the max chain length
909 * threshold is enforced at its default value, HASH_BKT_CAPACITY_THRESH. (If
910 * the bucket's chain exceeds this length, bucket expansion is triggered).
911 * However, setting expand_mult to a non-zero value delays bucket expansion
912 * (that would be triggered by additions to this particular bucket)
913 * until its chain length reaches a *multiple* of HASH_BKT_CAPACITY_THRESH.
914 * (The multiplier is simply expand_mult+1). The whole idea of this
915 * multiplier is to reduce bucket expansions, since they are expensive, in
916 * situations where we know that a particular bucket tends to be overused.
917 * It is better to let its chain length grow to a longer yet-still-bounded
918 * value, than to do an O(n) bucket expansion too often.
920 unsigned expand_mult;
924 /* random signature used only to find hash tables in external analysis */
925 #define HASH_SIGNATURE 0xa0111fe1
926 #define HASH_BLOOM_SIGNATURE 0xb12220f2
928 typedef struct UT_hash_table
930 UT_hash_bucket *buckets;
931 unsigned num_buckets, log2_num_buckets;
933 struct UT_hash_handle *tail; /* tail hh in app order, for fast append */
934 ptrdiff_t hho; /* hash handle offset (byte pos of hash handle in element */
936 /* in an ideal situation (all buckets used equally), no bucket would have
937 * more than ceil(#items/#buckets) items. that's the ideal chain length. */
938 unsigned ideal_chain_maxlen;
940 /* nonideal_items is the number of items in the hash whose chain position
941 * exceeds the ideal chain maxlen. these items pay the penalty for an uneven
942 * hash distribution; reaching them in a chain traversal takes >ideal steps */
943 unsigned nonideal_items;
945 /* ineffective expands occur when a bucket doubling was performed, but
946 * afterward, more than half the items in the hash had nonideal chain
947 * positions. If this happens on two consecutive expansions we inhibit any
948 * further expansion, as it's not helping; this happens when the hash
949 * function isn't a good fit for the key domain. When expansion is inhibited
950 * the hash will still work, albeit no longer in constant time. */
951 unsigned ineff_expands, noexpand;
953 uint32_t signature; /* used only to find hash tables in external analysis */
955 uint32_t bloom_sig; /* used only to test bloom exists in external analysis */
962 typedef struct UT_hash_handle
964 struct UT_hash_table *tbl;
965 void *prev; /* prev element in app order */
966 void *next; /* next element in app order */
967 struct UT_hash_handle *hh_prev; /* previous hh in bucket order */
968 struct UT_hash_handle *hh_next; /* next hh in bucket order */
969 void *key; /* ptr to enclosing struct's key */
970 unsigned keylen; /* enclosing struct's key len */
971 unsigned hashv; /* result of hash-fcn(key) */
974 #endif /* UTHASH_H */