1 /* hash - hashing table processing.
3 Copyright (C) 1998-2004, 2006-2007, 2009-2010 Free Software Foundation, Inc.
5 Written by Jim Meyering, 1992.
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.
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.
17 You should have received a copy of the GNU General Public License
18 along with this program. If not, see <http://www.gnu.org/licenses/>. */
20 /* A generic hash table package. */
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! */
29 #include "bitrotate.h"
38 # ifndef obstack_chunk_alloc
39 # define obstack_chunk_alloc malloc
41 # ifndef obstack_chunk_free
42 # define obstack_chunk_free free
49 struct hash_entry *next;
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;
60 size_t n_buckets_used;
63 /* Tuning arguments, kept in a physically separate structure. */
64 const Hash_tuning *tuning;
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. */
72 Hash_comparator comparator;
73 Hash_data_freer data_freer;
75 /* A linked list of freed struct hash_entry structs. */
76 struct hash_entry *free_entry_list;
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;
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.
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.
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. */
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 0.8
117 #define DEFAULT_GROWTH_FACTOR 1.414
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
125 #define DEFAULT_SHRINK_THRESHOLD 0.0
126 #define DEFAULT_SHRINK_FACTOR 1.0
128 /* Use this to initialize or reset a TUNING structure to
129 some sensible values. */
130 static const Hash_tuning default_tuning =
132 DEFAULT_SHRINK_THRESHOLD,
133 DEFAULT_SHRINK_FACTOR,
134 DEFAULT_GROWTH_THRESHOLD,
135 DEFAULT_GROWTH_FACTOR,
139 /* Information and lookup. */
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. */
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. */
150 hash_get_n_buckets (const Hash_table *table)
152 return table->n_buckets;
155 /* Return the number of slots in use (non-empty buckets). */
158 hash_get_n_buckets_used (const Hash_table *table)
160 return table->n_buckets_used;
163 /* Return the number of active entries. */
166 hash_get_n_entries (const Hash_table *table)
168 return table->n_entries;
171 /* Return the length of the longest chain (bucket). */
174 hash_get_max_bucket_length (const Hash_table *table)
176 struct hash_entry const *bucket;
177 size_t max_bucket_length = 0;
179 for (bucket = table->bucket; bucket < table->bucket_limit; bucket++)
183 struct hash_entry const *cursor = bucket;
184 size_t bucket_length = 1;
186 while (cursor = cursor->next, cursor)
189 if (bucket_length > max_bucket_length)
190 max_bucket_length = bucket_length;
194 return max_bucket_length;
197 /* Do a mild validation of a hash table, by traversing it and checking two
201 hash_table_ok (const Hash_table *table)
203 struct hash_entry const *bucket;
204 size_t n_buckets_used = 0;
205 size_t n_entries = 0;
207 for (bucket = table->bucket; bucket < table->bucket_limit; bucket++)
211 struct hash_entry const *cursor = bucket;
213 /* Count bucket head. */
217 /* Count bucket overflow. */
218 while (cursor = cursor->next, cursor)
223 if (n_buckets_used == table->n_buckets_used && n_entries == table->n_entries)
230 hash_print_statistics (const Hash_table *table, FILE *stream)
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);
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);
246 /* If ENTRY matches an entry already in the hash table, return the
247 entry from the table. Otherwise, return NULL. */
250 hash_lookup (const Hash_table *table, const void *entry)
252 struct hash_entry const *bucket
253 = table->bucket + table->hasher (entry, table->n_buckets);
254 struct hash_entry const *cursor;
256 if (! (bucket < table->bucket_limit))
259 if (bucket->data == NULL)
262 for (cursor = bucket; cursor; cursor = cursor->next)
263 if (entry == cursor->data || table->comparator (entry, cursor->data))
271 /* The functions in this page traverse the hash table and process the
272 contained entries. For the traversal to work properly, the hash table
273 should not be resized nor modified while any particular entry is being
274 processed. In particular, entries should not be added, and an entry
275 may be removed only if there is no shrink threshold and the entry being
276 removed has already been passed to hash_get_next. */
278 /* Return the first data in the table, or NULL if the table is empty. */
281 hash_get_first (const Hash_table *table)
283 struct hash_entry const *bucket;
285 if (table->n_entries == 0)
288 for (bucket = table->bucket; ; bucket++)
289 if (! (bucket < table->bucket_limit))
291 else if (bucket->data)
295 /* Return the user data for the entry following ENTRY, where ENTRY has been
296 returned by a previous call to either `hash_get_first' or `hash_get_next'.
297 Return NULL if there are no more entries. */
300 hash_get_next (const Hash_table *table, const void *entry)
302 struct hash_entry const *bucket
303 = table->bucket + table->hasher (entry, table->n_buckets);
304 struct hash_entry const *cursor;
306 if (! (bucket < table->bucket_limit))
309 /* Find next entry in the same bucket. */
313 if (cursor->data == entry && cursor->next)
314 return cursor->next->data;
315 cursor = cursor->next;
317 while (cursor != NULL);
319 /* Find first entry in any subsequent bucket. */
320 while (++bucket < table->bucket_limit)
328 /* Fill BUFFER with pointers to active user entries in the hash table, then
329 return the number of pointers copied. Do not copy more than BUFFER_SIZE
333 hash_get_entries (const Hash_table *table, void **buffer,
337 struct hash_entry const *bucket;
338 struct hash_entry const *cursor;
340 for (bucket = table->bucket; bucket < table->bucket_limit; bucket++)
344 for (cursor = bucket; cursor; cursor = cursor->next)
346 if (counter >= buffer_size)
348 buffer[counter++] = cursor->data;
356 /* Call a PROCESSOR function for each entry of a hash table, and return the
357 number of entries for which the processor function returned success. A
358 pointer to some PROCESSOR_DATA which will be made available to each call to
359 the processor function. The PROCESSOR accepts two arguments: the first is
360 the user entry being walked into, the second is the value of PROCESSOR_DATA
361 as received. The walking continue for as long as the PROCESSOR function
362 returns nonzero. When it returns zero, the walking is interrupted. */
365 hash_do_for_each (const Hash_table *table, Hash_processor processor,
366 void *processor_data)
369 struct hash_entry const *bucket;
370 struct hash_entry const *cursor;
372 for (bucket = table->bucket; bucket < table->bucket_limit; bucket++)
376 for (cursor = bucket; cursor; cursor = cursor->next)
378 if (! processor (cursor->data, processor_data))
388 /* Allocation and clean-up. */
390 /* Return a hash index for a NUL-terminated STRING between 0 and N_BUCKETS-1.
391 This is a convenience routine for constructing other hashing functions. */
395 /* About hashings, Paul Eggert writes to me (FP), on 1994-01-01: "Please see
396 B. J. McKenzie, R. Harries & T. Bell, Selecting a hashing algorithm,
397 Software--practice & experience 20, 2 (Feb 1990), 209-224. Good hash
398 algorithms tend to be domain-specific, so what's good for [diffutils'] io.c
399 may not be good for your application." */
402 hash_string (const char *string, size_t n_buckets)
404 # define HASH_ONE_CHAR(Value, Byte) \
405 ((Byte) + rotl_sz (Value, 7))
410 for (; (ch = *string); string++)
411 value = HASH_ONE_CHAR (value, ch);
412 return value % n_buckets;
414 # undef HASH_ONE_CHAR
417 #else /* not USE_DIFF_HASH */
419 /* This one comes from `recode', and performs a bit better than the above as
420 per a few experiments. It is inspired from a hashing routine found in the
421 very old Cyber `snoop', itself written in typical Greg Mansfield style.
422 (By the way, what happened to this excellent man? Is he still alive?) */
425 hash_string (const char *string, size_t n_buckets)
430 for (; (ch = *string); string++)
431 value = (value * 31 + ch) % n_buckets;
435 #endif /* not USE_DIFF_HASH */
437 /* Return true if CANDIDATE is a prime number. CANDIDATE should be an odd
438 number at least equal to 11. */
441 is_prime (size_t candidate)
444 size_t square = divisor * divisor;
446 while (square < candidate && (candidate % divisor))
449 square += 4 * divisor;
453 return (candidate % divisor ? true : false);
456 /* Round a given CANDIDATE number up to the nearest prime, and return that
457 prime. Primes lower than 10 are merely skipped. */
460 next_prime (size_t candidate)
462 /* Skip small primes. */
466 /* Make it definitely odd. */
469 while (SIZE_MAX != candidate && !is_prime (candidate))
476 hash_reset_tuning (Hash_tuning *tuning)
478 *tuning = default_tuning;
481 /* If the user passes a NULL hasher, we hash the raw pointer. */
483 raw_hasher (const void *data, size_t n)
485 /* When hashing unique pointers, it is often the case that they were
486 generated by malloc and thus have the property that the low-order
487 bits are 0. As this tends to give poorer performance with small
488 tables, we rotate the pointer value before performing division,
489 in an attempt to improve hash quality. */
490 size_t val = rotr_sz ((size_t) data, 3);
494 /* If the user passes a NULL comparator, we use pointer comparison. */
496 raw_comparator (const void *a, const void *b)
502 /* For the given hash TABLE, check the user supplied tuning structure for
503 reasonable values, and return true if there is no gross error with it.
504 Otherwise, definitively reset the TUNING field to some acceptable default
505 in the hash table (that is, the user loses the right of further modifying
506 tuning arguments), and return false. */
509 check_tuning (Hash_table *table)
511 const Hash_tuning *tuning = table->tuning;
513 if (tuning == &default_tuning)
516 /* Be a bit stricter than mathematics would require, so that
517 rounding errors in size calculations do not cause allocations to
518 fail to grow or shrink as they should. The smallest allocation
519 is 11 (due to next_prime's algorithm), so an epsilon of 0.1
520 should be good enough. */
523 if (epsilon < tuning->growth_threshold
524 && tuning->growth_threshold < 1 - epsilon
525 && 1 + epsilon < tuning->growth_factor
526 && 0 <= tuning->shrink_threshold
527 && tuning->shrink_threshold + epsilon < tuning->shrink_factor
528 && tuning->shrink_factor <= 1
529 && tuning->shrink_threshold + epsilon < tuning->growth_threshold)
532 table->tuning = &default_tuning;
536 /* Compute the size of the bucket array for the given CANDIDATE and
537 TUNING, or return 0 if there is no possible way to allocate that
541 compute_bucket_size (size_t candidate, const Hash_tuning *tuning)
543 if (!tuning->is_n_buckets)
545 float new_candidate = candidate / tuning->growth_threshold;
546 if (SIZE_MAX <= new_candidate)
548 candidate = new_candidate;
550 candidate = next_prime (candidate);
551 if (xalloc_oversized (candidate, sizeof (struct hash_entry *)))
556 /* Allocate and return a new hash table, or NULL upon failure. The initial
557 number of buckets is automatically selected so as to _guarantee_ that you
558 may insert at least CANDIDATE different user entries before any growth of
559 the hash table size occurs. So, if have a reasonably tight a-priori upper
560 bound on the number of entries you intend to insert in the hash table, you
561 may save some table memory and insertion time, by specifying it here. If
562 the IS_N_BUCKETS field of the TUNING structure is true, the CANDIDATE
563 argument has its meaning changed to the wanted number of buckets.
565 TUNING points to a structure of user-supplied values, in case some fine
566 tuning is wanted over the default behavior of the hasher. If TUNING is
567 NULL, the default tuning parameters are used instead. If TUNING is
568 provided but the values requested are out of bounds or might cause
569 rounding errors, return NULL.
571 The user-supplied HASHER function, when not NULL, accepts two
572 arguments ENTRY and TABLE_SIZE. It computes, by hashing ENTRY contents, a
573 slot number for that entry which should be in the range 0..TABLE_SIZE-1.
574 This slot number is then returned.
576 The user-supplied COMPARATOR function, when not NULL, accepts two
577 arguments pointing to user data, it then returns true for a pair of entries
578 that compare equal, or false otherwise. This function is internally called
579 on entries which are already known to hash to the same bucket index,
580 but which are distinct pointers.
582 The user-supplied DATA_FREER function, when not NULL, may be later called
583 with the user data as an argument, just before the entry containing the
584 data gets freed. This happens from within `hash_free' or `hash_clear'.
585 You should specify this function only if you want these functions to free
586 all of your `data' data. This is typically the case when your data is
587 simply an auxiliary struct that you have malloc'd to aggregate several
591 hash_initialize (size_t candidate, const Hash_tuning *tuning,
592 Hash_hasher hasher, Hash_comparator comparator,
593 Hash_data_freer data_freer)
599 if (comparator == NULL)
600 comparator = raw_comparator;
602 table = malloc (sizeof *table);
607 tuning = &default_tuning;
608 table->tuning = tuning;
609 if (!check_tuning (table))
611 /* Fail if the tuning options are invalid. This is the only occasion
612 when the user gets some feedback about it. Once the table is created,
613 if the user provides invalid tuning options, we silently revert to
614 using the defaults, and ignore further request to change the tuning
619 table->n_buckets = compute_bucket_size (candidate, tuning);
620 if (!table->n_buckets)
623 table->bucket = calloc (table->n_buckets, sizeof *table->bucket);
624 if (table->bucket == NULL)
626 table->bucket_limit = table->bucket + table->n_buckets;
627 table->n_buckets_used = 0;
628 table->n_entries = 0;
630 table->hasher = hasher;
631 table->comparator = comparator;
632 table->data_freer = data_freer;
634 table->free_entry_list = NULL;
636 obstack_init (&table->entry_stack);
645 /* Make all buckets empty, placing any chained entries on the free list.
646 Apply the user-specified function data_freer (if any) to the datas of any
650 hash_clear (Hash_table *table)
652 struct hash_entry *bucket;
654 for (bucket = table->bucket; bucket < table->bucket_limit; bucket++)
658 struct hash_entry *cursor;
659 struct hash_entry *next;
661 /* Free the bucket overflow. */
662 for (cursor = bucket->next; cursor; cursor = next)
664 if (table->data_freer)
665 table->data_freer (cursor->data);
669 /* Relinking is done one entry at a time, as it is to be expected
670 that overflows are either rare or short. */
671 cursor->next = table->free_entry_list;
672 table->free_entry_list = cursor;
675 /* Free the bucket head. */
676 if (table->data_freer)
677 table->data_freer (bucket->data);
683 table->n_buckets_used = 0;
684 table->n_entries = 0;
687 /* Reclaim all storage associated with a hash table. If a data_freer
688 function has been supplied by the user when the hash table was created,
689 this function applies it to the data of each entry before freeing that
693 hash_free (Hash_table *table)
695 struct hash_entry *bucket;
696 struct hash_entry *cursor;
697 struct hash_entry *next;
699 /* Call the user data_freer function. */
700 if (table->data_freer && table->n_entries)
702 for (bucket = table->bucket; bucket < table->bucket_limit; bucket++)
706 for (cursor = bucket; cursor; cursor = cursor->next)
707 table->data_freer (cursor->data);
714 obstack_free (&table->entry_stack, NULL);
718 /* Free all bucket overflowed entries. */
719 for (bucket = table->bucket; bucket < table->bucket_limit; bucket++)
721 for (cursor = bucket->next; cursor; cursor = next)
728 /* Also reclaim the internal list of previously freed entries. */
729 for (cursor = table->free_entry_list; cursor; cursor = next)
737 /* Free the remainder of the hash table structure. */
738 free (table->bucket);
742 /* Insertion and deletion. */
744 /* Get a new hash entry for a bucket overflow, possibly by recycling a
745 previously freed one. If this is not possible, allocate a new one. */
747 static struct hash_entry *
748 allocate_entry (Hash_table *table)
750 struct hash_entry *new;
752 if (table->free_entry_list)
754 new = table->free_entry_list;
755 table->free_entry_list = new->next;
760 new = obstack_alloc (&table->entry_stack, sizeof *new);
762 new = malloc (sizeof *new);
769 /* Free a hash entry which was part of some bucket overflow,
770 saving it for later recycling. */
773 free_entry (Hash_table *table, struct hash_entry *entry)
776 entry->next = table->free_entry_list;
777 table->free_entry_list = entry;
780 /* This private function is used to help with insertion and deletion. When
781 ENTRY matches an entry in the table, return a pointer to the corresponding
782 user data and set *BUCKET_HEAD to the head of the selected bucket.
783 Otherwise, return NULL. When DELETE is true and ENTRY matches an entry in
784 the table, unlink the matching entry. */
787 hash_find_entry (Hash_table *table, const void *entry,
788 struct hash_entry **bucket_head, bool delete)
790 struct hash_entry *bucket
791 = table->bucket + table->hasher (entry, table->n_buckets);
792 struct hash_entry *cursor;
794 if (! (bucket < table->bucket_limit))
797 *bucket_head = bucket;
799 /* Test for empty bucket. */
800 if (bucket->data == NULL)
803 /* See if the entry is the first in the bucket. */
804 if (entry == bucket->data || table->comparator (entry, bucket->data))
806 void *data = bucket->data;
812 struct hash_entry *next = bucket->next;
814 /* Bump the first overflow entry into the bucket head, then save
815 the previous first overflow entry for later recycling. */
817 free_entry (table, next);
828 /* Scan the bucket overflow. */
829 for (cursor = bucket; cursor->next; cursor = cursor->next)
831 if (entry == cursor->next->data
832 || table->comparator (entry, cursor->next->data))
834 void *data = cursor->next->data;
838 struct hash_entry *next = cursor->next;
840 /* Unlink the entry to delete, then save the freed entry for later
842 cursor->next = next->next;
843 free_entry (table, next);
850 /* No entry found. */
854 /* Internal helper, to move entries from SRC to DST. Both tables must
855 share the same free entry list. If SAFE, only move overflow
856 entries, saving bucket heads for later, so that no allocations will
857 occur. Return false if the free entry list is exhausted and an
861 transfer_entries (Hash_table *dst, Hash_table *src, bool safe)
863 struct hash_entry *bucket;
864 struct hash_entry *cursor;
865 struct hash_entry *next;
866 for (bucket = src->bucket; bucket < src->bucket_limit; bucket++)
870 struct hash_entry *new_bucket;
872 /* Within each bucket, transfer overflow entries first and
873 then the bucket head, to minimize memory pressure. After
874 all, the only time we might allocate is when moving the
875 bucket head, but moving overflow entries first may create
876 free entries that can be recycled by the time we finally
877 get to the bucket head. */
878 for (cursor = bucket->next; cursor; cursor = next)
881 new_bucket = (dst->bucket + dst->hasher (data, dst->n_buckets));
883 if (! (new_bucket < dst->bucket_limit))
888 if (new_bucket->data)
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;
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);
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
911 new_bucket = (dst->bucket + dst->hasher (data, dst->n_buckets));
913 if (! (new_bucket < dst->bucket_limit))
916 if (new_bucket->data)
918 /* Allocate or recycle an entry, when moving from a bucket
919 header into a bucket overflow. */
920 struct hash_entry *new_entry = allocate_entry (dst);
922 if (new_entry == NULL)
925 new_entry->data = data;
926 new_entry->next = new_bucket->next;
927 new_bucket->next = new_entry;
931 /* Move from one bucket header to another. */
932 new_bucket->data = data;
933 dst->n_buckets_used++;
936 src->n_buckets_used--;
941 /* For an already existing hash table, change the number of buckets through
942 specifying CANDIDATE. The contents of the hash table are preserved. The
943 new number of buckets is automatically selected so as to _guarantee_ that
944 the table may receive at least CANDIDATE different user entries, including
945 those already in the table, before any other growth of the hash table size
946 occurs. If TUNING->IS_N_BUCKETS is true, then CANDIDATE specifies the
947 exact number of buckets desired. Return true iff the rehash succeeded. */
950 hash_rehash (Hash_table *table, size_t candidate)
953 Hash_table *new_table;
954 size_t new_size = compute_bucket_size (candidate, table->tuning);
958 if (new_size == table->n_buckets)
960 new_table = &storage;
961 new_table->bucket = calloc (new_size, sizeof *new_table->bucket);
962 if (new_table->bucket == NULL)
964 new_table->n_buckets = new_size;
965 new_table->bucket_limit = new_table->bucket + new_size;
966 new_table->n_buckets_used = 0;
967 new_table->n_entries = 0;
968 new_table->tuning = table->tuning;
969 new_table->hasher = table->hasher;
970 new_table->comparator = table->comparator;
971 new_table->data_freer = table->data_freer;
973 /* In order for the transfer to successfully complete, we need
974 additional overflow entries when distinct buckets in the old
975 table collide into a common bucket in the new table. The worst
976 case possible is a hasher that gives a good spread with the old
977 size, but returns a constant with the new size; if we were to
978 guarantee table->n_buckets_used-1 free entries in advance, then
979 the transfer would be guaranteed to not allocate memory.
980 However, for large tables, a guarantee of no further allocation
981 introduces a lot of extra memory pressure, all for an unlikely
982 corner case (most rehashes reduce, rather than increase, the
983 number of overflow entries needed). So, we instead ensure that
984 the transfer process can be reversed if we hit a memory
985 allocation failure mid-transfer. */
987 /* Merely reuse the extra old space into the new table. */
989 new_table->entry_stack = table->entry_stack;
991 new_table->free_entry_list = table->free_entry_list;
993 if (transfer_entries (new_table, table, false))
995 /* Entries transferred successfully; tie up the loose ends. */
996 free (table->bucket);
997 table->bucket = new_table->bucket;
998 table->bucket_limit = new_table->bucket_limit;
999 table->n_buckets = new_table->n_buckets;
1000 table->n_buckets_used = new_table->n_buckets_used;
1001 table->free_entry_list = new_table->free_entry_list;
1002 /* table->n_entries and table->entry_stack already hold their value. */
1006 /* We've allocated new_table->bucket (and possibly some entries),
1007 exhausted the free list, and moved some but not all entries into
1008 new_table. We must undo the partial move before returning
1009 failure. The only way to get into this situation is if new_table
1010 uses fewer buckets than the old table, so we will reclaim some
1011 free entries as overflows in the new table are put back into
1012 distinct buckets in the old table.
1014 There are some pathological cases where a single pass through the
1015 table requires more intermediate overflow entries than using two
1016 passes. Two passes give worse cache performance and takes
1017 longer, but at this point, we're already out of memory, so slow
1018 and safe is better than failure. */
1019 table->free_entry_list = new_table->free_entry_list;
1020 if (! (transfer_entries (table, new_table, true)
1021 && transfer_entries (table, new_table, false)))
1023 /* table->n_entries already holds its value. */
1024 free (new_table->bucket);
1028 /* Return -1 upon memory allocation failure.
1029 Return 1 if insertion succeeded.
1030 Return 0 if there is already a matching entry in the table,
1031 and in that case, if MATCHED_ENT is non-NULL, set *MATCHED_ENT
1034 This interface is easier to use than hash_insert when you must
1035 distinguish between the latter two cases. More importantly,
1036 hash_insert is unusable for some types of ENTRY values. When using
1037 hash_insert, the only way to distinguish those cases is to compare
1038 the return value and ENTRY. That works only when you can have two
1039 different ENTRY values that point to data that compares "equal". Thus,
1040 when the ENTRY value is a simple scalar, you must use hash_insert0.
1041 ENTRY must not be NULL. */
1043 hash_insert0 (Hash_table *table, void const *entry, void const **matched_ent)
1046 struct hash_entry *bucket;
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. */
1054 /* If there's a matching entry already in the table, return that. */
1055 if ((data = hash_find_entry (table, entry, &bucket, false)) != NULL)
1058 *matched_ent = data;
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. */
1067 if (table->n_buckets_used
1068 > table->tuning->growth_threshold * table->n_buckets)
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)
1076 const Hash_tuning *tuning = table->tuning;
1078 (tuning->is_n_buckets
1079 ? (table->n_buckets * tuning->growth_factor)
1080 : (table->n_buckets * tuning->growth_factor
1081 * tuning->growth_threshold));
1083 if (SIZE_MAX <= candidate)
1086 /* If the rehash fails, arrange to return NULL. */
1087 if (!hash_rehash (table, candidate))
1090 /* Update the bucket we are interested in. */
1091 if (hash_find_entry (table, entry, &bucket, false) != NULL)
1096 /* ENTRY is not matched, it should be inserted. */
1100 struct hash_entry *new_entry = allocate_entry (table);
1102 if (new_entry == NULL)
1105 /* Add ENTRY in the overflow of the bucket. */
1107 new_entry->data = (void *) entry;
1108 new_entry->next = bucket->next;
1109 bucket->next = new_entry;
1114 /* Add ENTRY right in the bucket head. */
1116 bucket->data = (void *) entry;
1118 table->n_buckets_used++;
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
1130 hash_insert (Hash_table *table, void const *entry)
1132 void const *matched_ent;
1133 int err = hash_insert0 (table, entry, &matched_ent);
1136 : (void *) (err == 0 ? matched_ent : entry));
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. */
1144 hash_delete (Hash_table *table, const void *entry)
1147 struct hash_entry *bucket;
1149 data = hash_find_entry (table, entry, &bucket, true);
1156 table->n_buckets_used--;
1158 /* If the shrink threshold of the buckets in use has been reached,
1159 rehash into a smaller table. */
1161 if (table->n_buckets_used
1162 < table->tuning->shrink_threshold * table->n_buckets)
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)
1170 const Hash_tuning *tuning = table->tuning;
1172 (tuning->is_n_buckets
1173 ? table->n_buckets * tuning->shrink_factor
1174 : (table->n_buckets * tuning->shrink_factor
1175 * tuning->growth_threshold));
1177 if (!hash_rehash (table, candidate))
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. */
1185 struct hash_entry *cursor = table->free_entry_list;
1186 struct hash_entry *next;
1189 next = cursor->next;
1193 table->free_entry_list = NULL;
1208 hash_print (const Hash_table *table)
1210 struct hash_entry *bucket = (struct hash_entry *) table->bucket;
1212 for ( ; bucket < table->bucket_limit; bucket++)
1214 struct hash_entry *cursor;
1217 printf ("%lu:\n", (unsigned long int) (bucket - table->bucket));
1219 for (cursor = bucket; cursor; cursor = cursor->next)
1221 char const *s = cursor->data;
1224 printf (" %s\n", s);
1229 #endif /* TESTING */