1 /* hash - hashing table processing.
2
3 Copyright (C) 1998-2004, 2006-2007, 2009-2020 Free Software Foundation, Inc.
4
5 Written by Jim Meyering, 1992.
6
7 This program is free software: you can redistribute it and/or modify
8 it under the terms of the GNU General Public License as published by
9 the Free Software Foundation; either version 3 of the License, or
10 (at your option) any later version.
11
12 This program is distributed in the hope that it will be useful,
13 but WITHOUT ANY WARRANTY; without even the implied warranty of
14 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 GNU General Public License for more details.
16
17 You should have received a copy of the GNU General Public License
18 along with this program. If not, see <https://www.gnu.org/licenses/>. */
19
20 /* A generic hash table package. */
21
22 /* Define USE_OBSTACK to 1 if you want the allocator to use obstacks instead
23 of malloc. If you change USE_OBSTACK, you have to recompile! */
24
25 #include <config.h>
26
27 #include "hash.h"
28
29 #include "bitrotate.h"
30 #include "xalloc-oversized.h"
31
32 #include <stdint.h>
33 #include <stdio.h>
34 #include <stdlib.h>
35
36 #if USE_OBSTACK
37 # include "obstack.h"
38 # ifndef obstack_chunk_alloc
39 # define obstack_chunk_alloc malloc
40 # endif
41 # ifndef obstack_chunk_free
42 # define obstack_chunk_free free
43 # endif
44 #endif
45
46 struct hash_entry
47 {
48 void *data;
49 struct hash_entry *next;
50 };
51
52 struct hash_table
53 {
54 /* The array of buckets starts at BUCKET and extends to BUCKET_LIMIT-1,
55 for a possibility of N_BUCKETS. Among those, N_BUCKETS_USED buckets
56 are not empty, there are N_ENTRIES active entries in the table. */
57 struct hash_entry *bucket;
58 struct hash_entry const *bucket_limit;
59 size_t n_buckets;
60 size_t n_buckets_used;
61 size_t n_entries;
62
63 /* Tuning arguments, kept in a physically separate structure. */
64 const Hash_tuning *tuning;
65
66 /* Three functions are given to 'hash_initialize', see the documentation
67 block for this function. In a word, HASHER randomizes a user entry
68 into a number up from 0 up to some maximum minus 1; COMPARATOR returns
69 true if two user entries compare equally; and DATA_FREER is the cleanup
70 function for a user entry. */
71 Hash_hasher hasher;
72 Hash_comparator comparator;
73 Hash_data_freer data_freer;
74
75 /* A linked list of freed struct hash_entry structs. */
76 struct hash_entry *free_entry_list;
77
78 #if USE_OBSTACK
79 /* Whenever obstacks are used, it is possible to allocate all overflowed
80 entries into a single stack, so they all can be freed in a single
81 operation. It is not clear if the speedup is worth the trouble. */
82 struct obstack entry_stack;
83 #endif
84 };
85
86 /* A hash table contains many internal entries, each holding a pointer to
87 some user-provided data (also called a user entry). An entry indistinctly
88 refers to both the internal entry and its associated user entry. A user
89 entry contents may be hashed by a randomization function (the hashing
90 function, or just "hasher" for short) into a number (or "slot") between 0
91 and the current table size. At each slot position in the hash table,
92 starts a linked chain of entries for which the user data all hash to this
93 slot. A bucket is the collection of all entries hashing to the same slot.
94
95 A good "hasher" function will distribute entries rather evenly in buckets.
96 In the ideal case, the length of each bucket is roughly the number of
97 entries divided by the table size. Finding the slot for a data is usually
98 done in constant time by the "hasher", and the later finding of a precise
99 entry is linear in time with the size of the bucket. Consequently, a
100 larger hash table size (that is, a larger number of buckets) is prone to
101 yielding shorter chains, *given* the "hasher" function behaves properly.
102
103 Long buckets slow down the lookup algorithm. One might use big hash table
104 sizes in hope to reduce the average length of buckets, but this might
105 become inordinate, as unused slots in the hash table take some space. The
106 best bet is to make sure you are using a good "hasher" function (beware
107 that those are not that easy to write! :-), and to use a table size
108 larger than the actual number of entries. */
109
110 /* If an insertion makes the ratio of nonempty buckets to table size larger
111 than the growth threshold (a number between 0.0 and 1.0), then increase
112 the table size by multiplying by the growth factor (a number greater than
113 1.0). The growth threshold defaults to 0.8, and the growth factor
114 defaults to 1.414, meaning that the table will have doubled its size
115 every second time 80% of the buckets get used. */
116 #define DEFAULT_GROWTH_THRESHOLD 0.8f
117 #define DEFAULT_GROWTH_FACTOR 1.414f
118
119 /* If a deletion empties a bucket and causes the ratio of used buckets to
120 table size to become smaller than the shrink threshold (a number between
121 0.0 and 1.0), then shrink the table by multiplying by the shrink factor (a
122 number greater than the shrink threshold but smaller than 1.0). The shrink
123 threshold and factor default to 0.0 and 1.0, meaning that the table never
124 shrinks. */
125 #define DEFAULT_SHRINK_THRESHOLD 0.0f
126 #define DEFAULT_SHRINK_FACTOR 1.0f
127
128 /* Use this to initialize or reset a TUNING structure to
129 some sensible values. */
130 static const Hash_tuning default_tuning =
131 {
132 DEFAULT_SHRINK_THRESHOLD,
133 DEFAULT_SHRINK_FACTOR,
134 DEFAULT_GROWTH_THRESHOLD,
135 DEFAULT_GROWTH_FACTOR,
136 false
137 };
138
139 /* Information and lookup. */
140
141 /* The following few functions provide information about the overall hash
142 table organization: the number of entries, number of buckets and maximum
143 length of buckets. */
144
145 /* Return the number of buckets in the hash table. The table size, the total
146 number of buckets (used plus unused), or the maximum number of slots, are
147 the same quantity. */
148
149 size_t
hash_get_n_buckets(const Hash_table * table)150 hash_get_n_buckets (const Hash_table *table)
151 {
152 return table->n_buckets;
153 }
154
155 /* Return the number of slots in use (non-empty buckets). */
156
157 size_t
hash_get_n_buckets_used(const Hash_table * table)158 hash_get_n_buckets_used (const Hash_table *table)
159 {
160 return table->n_buckets_used;
161 }
162
163 /* Return the number of active entries. */
164
165 size_t
hash_get_n_entries(const Hash_table * table)166 hash_get_n_entries (const Hash_table *table)
167 {
168 return table->n_entries;
169 }
170
171 /* Return the length of the longest chain (bucket). */
172
173 size_t
hash_get_max_bucket_length(const Hash_table * table)174 hash_get_max_bucket_length (const Hash_table *table)
175 {
176 struct hash_entry const *bucket;
177 size_t max_bucket_length = 0;
178
179 for (bucket = table->bucket; bucket < table->bucket_limit; bucket++)
180 {
181 if (bucket->data)
182 {
183 struct hash_entry const *cursor = bucket;
184 size_t bucket_length = 1;
185
186 while (cursor = cursor->next, cursor)
187 bucket_length++;
188
189 if (bucket_length > max_bucket_length)
190 max_bucket_length = bucket_length;
191 }
192 }
193
194 return max_bucket_length;
195 }
196
197 /* Do a mild validation of a hash table, by traversing it and checking two
198 statistics. */
199
200 bool
hash_table_ok(const Hash_table * table)201 hash_table_ok (const Hash_table *table)
202 {
203 struct hash_entry const *bucket;
204 size_t n_buckets_used = 0;
205 size_t n_entries = 0;
206
207 for (bucket = table->bucket; bucket < table->bucket_limit; bucket++)
208 {
209 if (bucket->data)
210 {
211 struct hash_entry const *cursor = bucket;
212
213 /* Count bucket head. */
214 n_buckets_used++;
215 n_entries++;
216
217 /* Count bucket overflow. */
218 while (cursor = cursor->next, cursor)
219 n_entries++;
220 }
221 }
222
223 if (n_buckets_used == table->n_buckets_used && n_entries == table->n_entries)
224 return true;
225
226 return false;
227 }
228
229 void
hash_print_statistics(const Hash_table * table,FILE * stream)230 hash_print_statistics (const Hash_table *table, FILE *stream)
231 {
232 size_t n_entries = hash_get_n_entries (table);
233 size_t n_buckets = hash_get_n_buckets (table);
234 size_t n_buckets_used = hash_get_n_buckets_used (table);
235 size_t max_bucket_length = hash_get_max_bucket_length (table);
236
237 fprintf (stream, "# entries: %lu\n", (unsigned long int) n_entries);
238 fprintf (stream, "# buckets: %lu\n", (unsigned long int) n_buckets);
239 fprintf (stream, "# buckets used: %lu (%.2f%%)\n",
240 (unsigned long int) n_buckets_used,
241 (100.0 * n_buckets_used) / n_buckets);
242 fprintf (stream, "max bucket length: %lu\n",
243 (unsigned long int) max_bucket_length);
244 }
245
246 /* Hash KEY and return a pointer to the selected bucket.
247 If TABLE->hasher misbehaves, abort. */
248 static struct hash_entry *
safe_hasher(const Hash_table * table,const void * key)249 safe_hasher (const Hash_table *table, const void *key)
250 {
251 size_t n = table->hasher (key, table->n_buckets);
252 if (! (n < table->n_buckets))
253 abort ();
254 return table->bucket + n;
255 }
256
257 /* If ENTRY matches an entry already in the hash table, return the
258 entry from the table. Otherwise, return NULL. */
259
260 void *
hash_lookup(const Hash_table * table,const void * entry)261 hash_lookup (const Hash_table *table, const void *entry)
262 {
263 struct hash_entry const *bucket = safe_hasher (table, entry);
264 struct hash_entry const *cursor;
265
266 if (bucket->data == NULL)
267 return NULL;
268
269 for (cursor = bucket; cursor; cursor = cursor->next)
270 if (entry == cursor->data || table->comparator (entry, cursor->data))
271 return cursor->data;
272
273 return NULL;
274 }
275
276 /* Walking. */
277
278 /* The functions in this page traverse the hash table and process the
279 contained entries. For the traversal to work properly, the hash table
280 should not be resized nor modified while any particular entry is being
281 processed. In particular, entries should not be added, and an entry
282 may be removed only if there is no shrink threshold and the entry being
283 removed has already been passed to hash_get_next. */
284
285 /* Return the first data in the table, or NULL if the table is empty. */
286
287 void *
hash_get_first(const Hash_table * table)288 hash_get_first (const Hash_table *table)
289 {
290 struct hash_entry const *bucket;
291
292 if (table->n_entries == 0)
293 return NULL;
294
295 for (bucket = table->bucket; ; bucket++)
296 if (! (bucket < table->bucket_limit))
297 abort ();
298 else if (bucket->data)
299 return bucket->data;
300 }
301
302 /* Return the user data for the entry following ENTRY, where ENTRY has been
303 returned by a previous call to either 'hash_get_first' or 'hash_get_next'.
304 Return NULL if there are no more entries. */
305
306 void *
hash_get_next(const Hash_table * table,const void * entry)307 hash_get_next (const Hash_table *table, const void *entry)
308 {
309 struct hash_entry const *bucket = safe_hasher (table, entry);
310 struct hash_entry const *cursor;
311
312 /* Find next entry in the same bucket. */
313 cursor = bucket;
314 do
315 {
316 if (cursor->data == entry && cursor->next)
317 return cursor->next->data;
318 cursor = cursor->next;
319 }
320 while (cursor != NULL);
321
322 /* Find first entry in any subsequent bucket. */
323 while (++bucket < table->bucket_limit)
324 if (bucket->data)
325 return bucket->data;
326
327 /* None found. */
328 return NULL;
329 }
330
331 /* Fill BUFFER with pointers to active user entries in the hash table, then
332 return the number of pointers copied. Do not copy more than BUFFER_SIZE
333 pointers. */
334
335 size_t
hash_get_entries(const Hash_table * table,void ** buffer,size_t buffer_size)336 hash_get_entries (const Hash_table *table, void **buffer,
337 size_t buffer_size)
338 {
339 size_t counter = 0;
340 struct hash_entry const *bucket;
341 struct hash_entry const *cursor;
342
343 for (bucket = table->bucket; bucket < table->bucket_limit; bucket++)
344 {
345 if (bucket->data)
346 {
347 for (cursor = bucket; cursor; cursor = cursor->next)
348 {
349 if (counter >= buffer_size)
350 return counter;
351 buffer[counter++] = cursor->data;
352 }
353 }
354 }
355
356 return counter;
357 }
358
359 /* Call a PROCESSOR function for each entry of a hash table, and return the
360 number of entries for which the processor function returned success. A
361 pointer to some PROCESSOR_DATA which will be made available to each call to
362 the processor function. The PROCESSOR accepts two arguments: the first is
363 the user entry being walked into, the second is the value of PROCESSOR_DATA
364 as received. The walking continue for as long as the PROCESSOR function
365 returns nonzero. When it returns zero, the walking is interrupted. */
366
367 size_t
hash_do_for_each(const Hash_table * table,Hash_processor processor,void * processor_data)368 hash_do_for_each (const Hash_table *table, Hash_processor processor,
369 void *processor_data)
370 {
371 size_t counter = 0;
372 struct hash_entry const *bucket;
373 struct hash_entry const *cursor;
374
375 for (bucket = table->bucket; bucket < table->bucket_limit; bucket++)
376 {
377 if (bucket->data)
378 {
379 for (cursor = bucket; cursor; cursor = cursor->next)
380 {
381 if (! processor (cursor->data, processor_data))
382 return counter;
383 counter++;
384 }
385 }
386 }
387
388 return counter;
389 }
390
391 /* Allocation and clean-up. */
392
393 #if 0
394
395 /* Return a hash index for a NUL-terminated STRING between 0 and N_BUCKETS-1.
396 This is a convenience routine for constructing other hashing functions. */
397
398 #if USE_DIFF_HASH
399
400 /* About hashings, Paul Eggert writes to me (FP), on 1994-01-01: "Please see
401 B. J. McKenzie, R. Harries & T. Bell, Selecting a hashing algorithm,
402 Software--practice & experience 20, 2 (Feb 1990), 209-224. Good hash
403 algorithms tend to be domain-specific, so what's good for [diffutils'] io.c
404 may not be good for your application." */
405
406 size_t
407 hash_string (const char *string, size_t n_buckets)
408 {
409 # define HASH_ONE_CHAR(Value, Byte) \
410 ((Byte) + rotl_sz (Value, 7))
411
412 size_t value = 0;
413 unsigned char ch;
414
415 for (; (ch = *string); string++)
416 value = HASH_ONE_CHAR (value, ch);
417 return value % n_buckets;
418
419 # undef HASH_ONE_CHAR
420 }
421
422 #else /* not USE_DIFF_HASH */
423
424 /* This one comes from 'recode', and performs a bit better than the above as
425 per a few experiments. It is inspired from a hashing routine found in the
426 very old Cyber 'snoop', itself written in typical Greg Mansfield style.
427 (By the way, what happened to this excellent man? Is he still alive?) */
428
429 size_t
430 hash_string (const char *string, size_t n_buckets)
431 {
432 size_t value = 0;
433 unsigned char ch;
434
435 for (; (ch = *string); string++)
436 value = (value * 31 + ch) % n_buckets;
437 return value;
438 }
439
440 #endif /* not USE_DIFF_HASH */
441
442 #endif
443
444 /* Return true if CANDIDATE is a prime number. CANDIDATE should be an odd
445 number at least equal to 11. */
446
447 static bool _GL_ATTRIBUTE_CONST
is_prime(size_t candidate)448 is_prime (size_t candidate)
449 {
450 size_t divisor = 3;
451 size_t square = divisor * divisor;
452
453 while (square < candidate && (candidate % divisor))
454 {
455 divisor++;
456 square += 4 * divisor;
457 divisor++;
458 }
459
460 return (candidate % divisor ? true : false);
461 }
462
463 /* Round a given CANDIDATE number up to the nearest prime, and return that
464 prime. Primes lower than 10 are merely skipped. */
465
466 static size_t _GL_ATTRIBUTE_CONST
next_prime(size_t candidate)467 next_prime (size_t candidate)
468 {
469 /* Skip small primes. */
470 if (candidate < 10)
471 candidate = 10;
472
473 /* Make it definitely odd. */
474 candidate |= 1;
475
476 while (SIZE_MAX != candidate && !is_prime (candidate))
477 candidate += 2;
478
479 return candidate;
480 }
481
482 void
hash_reset_tuning(Hash_tuning * tuning)483 hash_reset_tuning (Hash_tuning *tuning)
484 {
485 *tuning = default_tuning;
486 }
487
488 /* If the user passes a NULL hasher, we hash the raw pointer. */
489 static size_t
raw_hasher(const void * data,size_t n)490 raw_hasher (const void *data, size_t n)
491 {
492 /* When hashing unique pointers, it is often the case that they were
493 generated by malloc and thus have the property that the low-order
494 bits are 0. As this tends to give poorer performance with small
495 tables, we rotate the pointer value before performing division,
496 in an attempt to improve hash quality. */
497 size_t val = rotr_sz ((size_t) data, 3);
498 return val % n;
499 }
500
501 /* If the user passes a NULL comparator, we use pointer comparison. */
502 static bool
raw_comparator(const void * a,const void * b)503 raw_comparator (const void *a, const void *b)
504 {
505 return a == b;
506 }
507
508
509 /* For the given hash TABLE, check the user supplied tuning structure for
510 reasonable values, and return true if there is no gross error with it.
511 Otherwise, definitively reset the TUNING field to some acceptable default
512 in the hash table (that is, the user loses the right of further modifying
513 tuning arguments), and return false. */
514
515 static bool
check_tuning(Hash_table * table)516 check_tuning (Hash_table *table)
517 {
518 const Hash_tuning *tuning = table->tuning;
519 float epsilon;
520 if (tuning == &default_tuning)
521 return true;
522
523 /* Be a bit stricter than mathematics would require, so that
524 rounding errors in size calculations do not cause allocations to
525 fail to grow or shrink as they should. The smallest allocation
526 is 11 (due to next_prime's algorithm), so an epsilon of 0.1
527 should be good enough. */
528 epsilon = 0.1f;
529
530 if (epsilon < tuning->growth_threshold
531 && tuning->growth_threshold < 1 - epsilon
532 && 1 + epsilon < tuning->growth_factor
533 && 0 <= tuning->shrink_threshold
534 && tuning->shrink_threshold + epsilon < tuning->shrink_factor
535 && tuning->shrink_factor <= 1
536 && tuning->shrink_threshold + epsilon < tuning->growth_threshold)
537 return true;
538
539 table->tuning = &default_tuning;
540 return false;
541 }
542
543 /* Compute the size of the bucket array for the given CANDIDATE and
544 TUNING, or return 0 if there is no possible way to allocate that
545 many entries. */
546
547 static size_t _GL_ATTRIBUTE_PURE
compute_bucket_size(size_t candidate,const Hash_tuning * tuning)548 compute_bucket_size (size_t candidate, const Hash_tuning *tuning)
549 {
550 if (!tuning->is_n_buckets)
551 {
552 float new_candidate = candidate / tuning->growth_threshold;
553 if (SIZE_MAX <= new_candidate)
554 return 0;
555 candidate = new_candidate;
556 }
557 candidate = next_prime (candidate);
558 if (xalloc_oversized (candidate, sizeof (struct hash_entry *)))
559 return 0;
560 return candidate;
561 }
562
563 /* Allocate and return a new hash table, or NULL upon failure. The initial
564 number of buckets is automatically selected so as to _guarantee_ that you
565 may insert at least CANDIDATE different user entries before any growth of
566 the hash table size occurs. So, if have a reasonably tight a-priori upper
567 bound on the number of entries you intend to insert in the hash table, you
568 may save some table memory and insertion time, by specifying it here. If
569 the IS_N_BUCKETS field of the TUNING structure is true, the CANDIDATE
570 argument has its meaning changed to the wanted number of buckets.
571
572 TUNING points to a structure of user-supplied values, in case some fine
573 tuning is wanted over the default behavior of the hasher. If TUNING is
574 NULL, the default tuning parameters are used instead. If TUNING is
575 provided but the values requested are out of bounds or might cause
576 rounding errors, return NULL.
577
578 The user-supplied HASHER function, when not NULL, accepts two
579 arguments ENTRY and TABLE_SIZE. It computes, by hashing ENTRY contents, a
580 slot number for that entry which should be in the range 0..TABLE_SIZE-1.
581 This slot number is then returned.
582
583 The user-supplied COMPARATOR function, when not NULL, accepts two
584 arguments pointing to user data, it then returns true for a pair of entries
585 that compare equal, or false otherwise. This function is internally called
586 on entries which are already known to hash to the same bucket index,
587 but which are distinct pointers.
588
589 The user-supplied DATA_FREER function, when not NULL, may be later called
590 with the user data as an argument, just before the entry containing the
591 data gets freed. This happens from within 'hash_free' or 'hash_clear'.
592 You should specify this function only if you want these functions to free
593 all of your 'data' data. This is typically the case when your data is
594 simply an auxiliary struct that you have malloc'd to aggregate several
595 values. */
596
597 Hash_table *
hash_initialize(size_t candidate,const Hash_tuning * tuning,Hash_hasher hasher,Hash_comparator comparator,Hash_data_freer data_freer)598 hash_initialize (size_t candidate, const Hash_tuning *tuning,
599 Hash_hasher hasher, Hash_comparator comparator,
600 Hash_data_freer data_freer)
601 {
602 Hash_table *table;
603
604 if (hasher == NULL)
605 hasher = raw_hasher;
606 if (comparator == NULL)
607 comparator = raw_comparator;
608
609 table = malloc (sizeof *table);
610 if (table == NULL)
611 return NULL;
612
613 if (!tuning)
614 tuning = &default_tuning;
615 table->tuning = tuning;
616 if (!check_tuning (table))
617 {
618 /* Fail if the tuning options are invalid. This is the only occasion
619 when the user gets some feedback about it. Once the table is created,
620 if the user provides invalid tuning options, we silently revert to
621 using the defaults, and ignore further request to change the tuning
622 options. */
623 goto fail;
624 }
625
626 table->n_buckets = compute_bucket_size (candidate, tuning);
627 if (!table->n_buckets)
628 goto fail;
629
630 table->bucket = calloc (table->n_buckets, sizeof *table->bucket);
631 if (table->bucket == NULL)
632 goto fail;
633 table->bucket_limit = table->bucket + table->n_buckets;
634 table->n_buckets_used = 0;
635 table->n_entries = 0;
636
637 table->hasher = hasher;
638 table->comparator = comparator;
639 table->data_freer = data_freer;
640
641 table->free_entry_list = NULL;
642 #if USE_OBSTACK
643 obstack_init (&table->entry_stack);
644 #endif
645 return table;
646
647 fail:
648 free (table);
649 return NULL;
650 }
651
652 /* Make all buckets empty, placing any chained entries on the free list.
653 Apply the user-specified function data_freer (if any) to the datas of any
654 affected entries. */
655
656 void
hash_clear(Hash_table * table)657 hash_clear (Hash_table *table)
658 {
659 struct hash_entry *bucket;
660
661 for (bucket = table->bucket; bucket < table->bucket_limit; bucket++)
662 {
663 if (bucket->data)
664 {
665 struct hash_entry *cursor;
666 struct hash_entry *next;
667
668 /* Free the bucket overflow. */
669 for (cursor = bucket->next; cursor; cursor = next)
670 {
671 if (table->data_freer)
672 table->data_freer (cursor->data);
673 cursor->data = NULL;
674
675 next = cursor->next;
676 /* Relinking is done one entry at a time, as it is to be expected
677 that overflows are either rare or short. */
678 cursor->next = table->free_entry_list;
679 table->free_entry_list = cursor;
680 }
681
682 /* Free the bucket head. */
683 if (table->data_freer)
684 table->data_freer (bucket->data);
685 bucket->data = NULL;
686 bucket->next = NULL;
687 }
688 }
689
690 table->n_buckets_used = 0;
691 table->n_entries = 0;
692 }
693
694 /* Reclaim all storage associated with a hash table. If a data_freer
695 function has been supplied by the user when the hash table was created,
696 this function applies it to the data of each entry before freeing that
697 entry. */
698
699 void
hash_free(Hash_table * table)700 hash_free (Hash_table *table)
701 {
702 struct hash_entry *bucket;
703 struct hash_entry *cursor;
704 struct hash_entry *next;
705
706 /* Call the user data_freer function. */
707 if (table->data_freer && table->n_entries)
708 {
709 for (bucket = table->bucket; bucket < table->bucket_limit; bucket++)
710 {
711 if (bucket->data)
712 {
713 for (cursor = bucket; cursor; cursor = cursor->next)
714 table->data_freer (cursor->data);
715 }
716 }
717 }
718
719 #if USE_OBSTACK
720
721 obstack_free (&table->entry_stack, NULL);
722
723 #else
724
725 /* Free all bucket overflowed entries. */
726 for (bucket = table->bucket; bucket < table->bucket_limit; bucket++)
727 {
728 for (cursor = bucket->next; cursor; cursor = next)
729 {
730 next = cursor->next;
731 free (cursor);
732 }
733 }
734
735 /* Also reclaim the internal list of previously freed entries. */
736 for (cursor = table->free_entry_list; cursor; cursor = next)
737 {
738 next = cursor->next;
739 free (cursor);
740 }
741
742 #endif
743
744 /* Free the remainder of the hash table structure. */
745 free (table->bucket);
746 free (table);
747 }
748
749 /* Insertion and deletion. */
750
751 /* Get a new hash entry for a bucket overflow, possibly by recycling a
752 previously freed one. If this is not possible, allocate a new one. */
753
754 static struct hash_entry *
allocate_entry(Hash_table * table)755 allocate_entry (Hash_table *table)
756 {
757 struct hash_entry *new;
758
759 if (table->free_entry_list)
760 {
761 new = table->free_entry_list;
762 table->free_entry_list = new->next;
763 }
764 else
765 {
766 #if USE_OBSTACK
767 new = obstack_alloc (&table->entry_stack, sizeof *new);
768 #else
769 new = malloc (sizeof *new);
770 #endif
771 }
772
773 return new;
774 }
775
776 /* Free a hash entry which was part of some bucket overflow,
777 saving it for later recycling. */
778
779 static void
free_entry(Hash_table * table,struct hash_entry * entry)780 free_entry (Hash_table *table, struct hash_entry *entry)
781 {
782 entry->data = NULL;
783 entry->next = table->free_entry_list;
784 table->free_entry_list = entry;
785 }
786
787 /* This private function is used to help with insertion and deletion. When
788 ENTRY matches an entry in the table, return a pointer to the corresponding
789 user data and set *BUCKET_HEAD to the head of the selected bucket.
790 Otherwise, return NULL. When DELETE is true and ENTRY matches an entry in
791 the table, unlink the matching entry. */
792
793 static void *
hash_find_entry(Hash_table * table,const void * entry,struct hash_entry ** bucket_head,bool delete)794 hash_find_entry (Hash_table *table, const void *entry,
795 struct hash_entry **bucket_head, bool delete)
796 {
797 struct hash_entry *bucket = safe_hasher (table, entry);
798 struct hash_entry *cursor;
799
800 *bucket_head = bucket;
801
802 /* Test for empty bucket. */
803 if (bucket->data == NULL)
804 return NULL;
805
806 /* See if the entry is the first in the bucket. */
807 if (entry == bucket->data || table->comparator (entry, bucket->data))
808 {
809 void *data = bucket->data;
810
811 if (delete)
812 {
813 if (bucket->next)
814 {
815 struct hash_entry *next = bucket->next;
816
817 /* Bump the first overflow entry into the bucket head, then save
818 the previous first overflow entry for later recycling. */
819 *bucket = *next;
820 free_entry (table, next);
821 }
822 else
823 {
824 bucket->data = NULL;
825 }
826 }
827
828 return data;
829 }
830
831 /* Scan the bucket overflow. */
832 for (cursor = bucket; cursor->next; cursor = cursor->next)
833 {
834 if (entry == cursor->next->data
835 || table->comparator (entry, cursor->next->data))
836 {
837 void *data = cursor->next->data;
838
839 if (delete)
840 {
841 struct hash_entry *next = cursor->next;
842
843 /* Unlink the entry to delete, then save the freed entry for later
844 recycling. */
845 cursor->next = next->next;
846 free_entry (table, next);
847 }
848
849 return data;
850 }
851 }
852
853 /* No entry found. */
854 return NULL;
855 }
856
857 /* Internal helper, to move entries from SRC to DST. Both tables must
858 share the same free entry list. If SAFE, only move overflow
859 entries, saving bucket heads for later, so that no allocations will
860 occur. Return false if the free entry list is exhausted and an
861 allocation fails. */
862
863 static bool
transfer_entries(Hash_table * dst,Hash_table * src,bool safe)864 transfer_entries (Hash_table *dst, Hash_table *src, bool safe)
865 {
866 struct hash_entry *bucket;
867 struct hash_entry *cursor;
868 struct hash_entry *next;
869 for (bucket = src->bucket; bucket < src->bucket_limit; bucket++)
870 if (bucket->data)
871 {
872 void *data;
873 struct hash_entry *new_bucket;
874
875 /* Within each bucket, transfer overflow entries first and
876 then the bucket head, to minimize memory pressure. After
877 all, the only time we might allocate is when moving the
878 bucket head, but moving overflow entries first may create
879 free entries that can be recycled by the time we finally
880 get to the bucket head. */
881 for (cursor = bucket->next; cursor; cursor = next)
882 {
883 data = cursor->data;
884 new_bucket = safe_hasher (dst, data);
885
886 next = cursor->next;
887
888 if (new_bucket->data)
889 {
890 /* Merely relink an existing entry, when moving from a
891 bucket overflow into a bucket overflow. */
892 cursor->next = new_bucket->next;
893 new_bucket->next = cursor;
894 }
895 else
896 {
897 /* Free an existing entry, when moving from a bucket
898 overflow into a bucket header. */
899 new_bucket->data = data;
900 dst->n_buckets_used++;
901 free_entry (dst, cursor);
902 }
903 }
904 /* Now move the bucket head. Be sure that if we fail due to
905 allocation failure that the src table is in a consistent
906 state. */
907 data = bucket->data;
908 bucket->next = NULL;
909 if (safe)
910 continue;
911 new_bucket = safe_hasher (dst, data);
912
913 if (new_bucket->data)
914 {
915 /* Allocate or recycle an entry, when moving from a bucket
916 header into a bucket overflow. */
917 struct hash_entry *new_entry = allocate_entry (dst);
918
919 if (new_entry == NULL)
920 return false;
921
922 new_entry->data = data;
923 new_entry->next = new_bucket->next;
924 new_bucket->next = new_entry;
925 }
926 else
927 {
928 /* Move from one bucket header to another. */
929 new_bucket->data = data;
930 dst->n_buckets_used++;
931 }
932 bucket->data = NULL;
933 src->n_buckets_used--;
934 }
935 return true;
936 }
937
938 /* For an already existing hash table, change the number of buckets through
939 specifying CANDIDATE. The contents of the hash table are preserved. The
940 new number of buckets is automatically selected so as to _guarantee_ that
941 the table may receive at least CANDIDATE different user entries, including
942 those already in the table, before any other growth of the hash table size
943 occurs. If TUNING->IS_N_BUCKETS is true, then CANDIDATE specifies the
944 exact number of buckets desired. Return true iff the rehash succeeded. */
945
946 bool
hash_rehash(Hash_table * table,size_t candidate)947 hash_rehash (Hash_table *table, size_t candidate)
948 {
949 Hash_table storage;
950 Hash_table *new_table;
951 size_t new_size = compute_bucket_size (candidate, table->tuning);
952
953 if (!new_size)
954 return false;
955 if (new_size == table->n_buckets)
956 return true;
957 new_table = &storage;
958 new_table->bucket = calloc (new_size, sizeof *new_table->bucket);
959 if (new_table->bucket == NULL)
960 return false;
961 new_table->n_buckets = new_size;
962 new_table->bucket_limit = new_table->bucket + new_size;
963 new_table->n_buckets_used = 0;
964 new_table->n_entries = 0;
965 new_table->tuning = table->tuning;
966 new_table->hasher = table->hasher;
967 new_table->comparator = table->comparator;
968 new_table->data_freer = table->data_freer;
969
970 /* In order for the transfer to successfully complete, we need
971 additional overflow entries when distinct buckets in the old
972 table collide into a common bucket in the new table. The worst
973 case possible is a hasher that gives a good spread with the old
974 size, but returns a constant with the new size; if we were to
975 guarantee table->n_buckets_used-1 free entries in advance, then
976 the transfer would be guaranteed to not allocate memory.
977 However, for large tables, a guarantee of no further allocation
978 introduces a lot of extra memory pressure, all for an unlikely
979 corner case (most rehashes reduce, rather than increase, the
980 number of overflow entries needed). So, we instead ensure that
981 the transfer process can be reversed if we hit a memory
982 allocation failure mid-transfer. */
983
984 /* Merely reuse the extra old space into the new table. */
985 #if USE_OBSTACK
986 new_table->entry_stack = table->entry_stack;
987 #endif
988 new_table->free_entry_list = table->free_entry_list;
989
990 if (transfer_entries (new_table, table, false))
991 {
992 /* Entries transferred successfully; tie up the loose ends. */
993 free (table->bucket);
994 table->bucket = new_table->bucket;
995 table->bucket_limit = new_table->bucket_limit;
996 table->n_buckets = new_table->n_buckets;
997 table->n_buckets_used = new_table->n_buckets_used;
998 table->free_entry_list = new_table->free_entry_list;
999 /* table->n_entries and table->entry_stack already hold their value. */
1000 return true;
1001 }
1002
1003 /* We've allocated new_table->bucket (and possibly some entries),
1004 exhausted the free list, and moved some but not all entries into
1005 new_table. We must undo the partial move before returning
1006 failure. The only way to get into this situation is if new_table
1007 uses fewer buckets than the old table, so we will reclaim some
1008 free entries as overflows in the new table are put back into
1009 distinct buckets in the old table.
1010
1011 There are some pathological cases where a single pass through the
1012 table requires more intermediate overflow entries than using two
1013 passes. Two passes give worse cache performance and takes
1014 longer, but at this point, we're already out of memory, so slow
1015 and safe is better than failure. */
1016 table->free_entry_list = new_table->free_entry_list;
1017 if (! (transfer_entries (table, new_table, true)
1018 && transfer_entries (table, new_table, false)))
1019 abort ();
1020 /* table->n_entries already holds its value. */
1021 free (new_table->bucket);
1022 return false;
1023 }
1024
1025 /* Insert ENTRY into hash TABLE if there is not already a matching entry.
1026
1027 Return -1 upon memory allocation failure.
1028 Return 1 if insertion succeeded.
1029 Return 0 if there is already a matching entry in the table,
1030 and in that case, if MATCHED_ENT is non-NULL, set *MATCHED_ENT
1031 to that entry.
1032
1033 This interface is easier to use than hash_insert when you must
1034 distinguish between the latter two cases. More importantly,
1035 hash_insert is unusable for some types of ENTRY values. When using
1036 hash_insert, the only way to distinguish those cases is to compare
1037 the return value and ENTRY. That works only when you can have two
1038 different ENTRY values that point to data that compares "equal". Thus,
1039 when the ENTRY value is a simple scalar, you must use
1040 hash_insert_if_absent. ENTRY must not be NULL. */
1041 int
hash_insert_if_absent(Hash_table * table,void const * entry,void const ** matched_ent)1042 hash_insert_if_absent (Hash_table *table, void const *entry,
1043 void const **matched_ent)
1044 {
1045 void *data;
1046 struct hash_entry *bucket;
1047
1048 /* The caller cannot insert a NULL entry, since hash_lookup returns NULL
1049 to indicate "not found", and hash_find_entry uses "bucket->data == NULL"
1050 to indicate an empty bucket. */
1051 if (! entry)
1052 abort ();
1053
1054 /* If there's a matching entry already in the table, return that. */
1055 if ((data = hash_find_entry (table, entry, &bucket, false)) != NULL)
1056 {
1057 if (matched_ent)
1058 *matched_ent = data;
1059 return 0;
1060 }
1061
1062 /* If the growth threshold of the buckets in use has been reached, increase
1063 the table size and rehash. There's no point in checking the number of
1064 entries: if the hashing function is ill-conditioned, rehashing is not
1065 likely to improve it. */
1066
1067 if (table->n_buckets_used
1068 > table->tuning->growth_threshold * table->n_buckets)
1069 {
1070 /* Check more fully, before starting real work. If tuning arguments
1071 became invalid, the second check will rely on proper defaults. */
1072 check_tuning (table);
1073 if (table->n_buckets_used
1074 > table->tuning->growth_threshold * table->n_buckets)
1075 {
1076 const Hash_tuning *tuning = table->tuning;
1077 float candidate =
1078 (tuning->is_n_buckets
1079 ? (table->n_buckets * tuning->growth_factor)
1080 : (table->n_buckets * tuning->growth_factor
1081 * tuning->growth_threshold));
1082
1083 if (SIZE_MAX <= candidate)
1084 return -1;
1085
1086 /* If the rehash fails, arrange to return NULL. */
1087 if (!hash_rehash (table, candidate))
1088 return -1;
1089
1090 /* Update the bucket we are interested in. */
1091 if (hash_find_entry (table, entry, &bucket, false) != NULL)
1092 abort ();
1093 }
1094 }
1095
1096 /* ENTRY is not matched, it should be inserted. */
1097
1098 if (bucket->data)
1099 {
1100 struct hash_entry *new_entry = allocate_entry (table);
1101
1102 if (new_entry == NULL)
1103 return -1;
1104
1105 /* Add ENTRY in the overflow of the bucket. */
1106
1107 new_entry->data = (void *) entry;
1108 new_entry->next = bucket->next;
1109 bucket->next = new_entry;
1110 table->n_entries++;
1111 return 1;
1112 }
1113
1114 /* Add ENTRY right in the bucket head. */
1115
1116 bucket->data = (void *) entry;
1117 table->n_entries++;
1118 table->n_buckets_used++;
1119
1120 return 1;
1121 }
1122
1123 /* If ENTRY matches an entry already in the hash table, return the pointer
1124 to the entry from the table. Otherwise, insert ENTRY and return ENTRY.
1125 Return NULL if the storage required for insertion cannot be allocated.
1126 This implementation does not support duplicate entries or insertion of
1127 NULL. */
1128
1129 void *
hash_insert(Hash_table * table,void const * entry)1130 hash_insert (Hash_table *table, void const *entry)
1131 {
1132 void const *matched_ent;
1133 int err = hash_insert_if_absent (table, entry, &matched_ent);
1134 return (err == -1
1135 ? NULL
1136 : (void *) (err == 0 ? matched_ent : entry));
1137 }
1138
1139 /* If ENTRY is already in the table, remove it and return the just-deleted
1140 data (the user may want to deallocate its storage). If ENTRY is not in the
1141 table, don't modify the table and return NULL. */
1142
1143 void *
hash_delete(Hash_table * table,const void * entry)1144 hash_delete (Hash_table *table, const void *entry)
1145 {
1146 void *data;
1147 struct hash_entry *bucket;
1148
1149 data = hash_find_entry (table, entry, &bucket, true);
1150 if (!data)
1151 return NULL;
1152
1153 table->n_entries--;
1154 if (!bucket->data)
1155 {
1156 table->n_buckets_used--;
1157
1158 /* If the shrink threshold of the buckets in use has been reached,
1159 rehash into a smaller table. */
1160
1161 if (table->n_buckets_used
1162 < table->tuning->shrink_threshold * table->n_buckets)
1163 {
1164 /* Check more fully, before starting real work. If tuning arguments
1165 became invalid, the second check will rely on proper defaults. */
1166 check_tuning (table);
1167 if (table->n_buckets_used
1168 < table->tuning->shrink_threshold * table->n_buckets)
1169 {
1170 const Hash_tuning *tuning = table->tuning;
1171 size_t candidate =
1172 (tuning->is_n_buckets
1173 ? table->n_buckets * tuning->shrink_factor
1174 : (table->n_buckets * tuning->shrink_factor
1175 * tuning->growth_threshold));
1176
1177 if (!hash_rehash (table, candidate))
1178 {
1179 /* Failure to allocate memory in an attempt to
1180 shrink the table is not fatal. But since memory
1181 is low, we can at least be kind and free any
1182 spare entries, rather than keeping them tied up
1183 in the free entry list. */
1184 #if ! USE_OBSTACK
1185 struct hash_entry *cursor = table->free_entry_list;
1186 struct hash_entry *next;
1187 while (cursor)
1188 {
1189 next = cursor->next;
1190 free (cursor);
1191 cursor = next;
1192 }
1193 table->free_entry_list = NULL;
1194 #endif
1195 }
1196 }
1197 }
1198 }
1199
1200 return data;
1201 }
1202
1203 /* Testing. */
1204
1205 #if TESTING
1206
1207 void
hash_print(const Hash_table * table)1208 hash_print (const Hash_table *table)
1209 {
1210 struct hash_entry *bucket = (struct hash_entry *) table->bucket;
1211
1212 for ( ; bucket < table->bucket_limit; bucket++)
1213 {
1214 struct hash_entry *cursor;
1215
1216 if (bucket)
1217 printf ("%lu:\n", (unsigned long int) (bucket - table->bucket));
1218
1219 for (cursor = bucket; cursor; cursor = cursor->next)
1220 {
1221 char const *s = cursor->data;
1222 /* FIXME */
1223 if (s)
1224 printf (" %s\n", s);
1225 }
1226 }
1227 }
1228
1229 #endif /* TESTING */
1230