2 Copyright (C) 2000, 2001 Free Software Foundation, Inc.
4 This file is part of GNU Wget.
6 GNU Wget is free software; you can redistribute it and/or modify
7 it under the terms of the GNU General Public License as published by
8 the Free Software Foundation; either version 2 of the License, or (at
9 your option) any later version.
11 GNU Wget is distributed in the hope that it will be useful,
12 but WITHOUT ANY WARRANTY; without even the implied warranty of
13 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 GNU General Public License for more details.
16 You should have received a copy of the GNU General Public License
17 along with Wget; if not, write to the Free Software
18 Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
20 In addition, as a special exception, the Free Software Foundation
21 gives permission to link the code of its release of Wget with the
22 OpenSSL project's "OpenSSL" library (or with modified versions of it
23 that use the same license as the "OpenSSL" library), and distribute
24 the linked executables. You must obey the GNU General Public License
25 in all respects for all of the code used other than "OpenSSL". If you
26 modify this file, you may extend this exception to your version of the
27 file, but you are not obligated to do so. If you do not wish to do
28 so, delete this exception statement from your version. */
38 #endif /* HAVE_STRING_H */
52 # define xmalloc malloc
53 # define xrealloc realloc
57 # define TOLOWER(x) ('A' <= (x) && (x) <= 'Z' ? (x) - 32 : (x))
62 Hash tables are a technique used to implement mapping between
63 objects with near-constant-time access and storage. The table
64 associates keys to values, and a value can be very quickly
65 retrieved by providing the key. Fast lookup tables are typically
66 implemented as hash tables.
69 hash_table_new -- creates the table.
70 hash_table_destroy -- destroys the table.
71 hash_table_put -- establishes or updates key->value mapping.
72 hash_table_get -- retrieves value of key.
73 hash_table_get_pair -- get key/value pair for key.
74 hash_table_contains -- test whether the table contains key.
75 hash_table_remove -- remove the key->value mapping for key.
76 hash_table_map -- iterate through table mappings.
77 hash_table_clear -- clear hash table contents.
78 hash_table_count -- return the number of entries in the table.
80 The hash table grows internally as new entries are added and is not
81 limited in size, except by available memory. The table doubles
82 with each resize, which ensures that the amortized time per
83 operation remains constant.
85 By default, tables created by hash_table_new consider the keys to
86 be equal if their pointer values are the same. You can use
87 make_string_hash_table to create tables whose keys are considered
88 equal if their string contents are the same. In the general case,
89 the criterion of equality used to compare keys is specified at
90 table creation time with two callback functions, "hash" and "test".
91 The hash function transforms the key into an arbitrary number that
92 must be the same for two equal keys. The test function accepts two
93 keys and returns non-zero if they are to be considered equal.
95 Note that neither keys nor values are copied when inserted into the
96 hash table, so they must exist for the lifetime of the table. This
97 means that e.g. the use of static strings is OK, but objects with a
98 shorter life-time need to be copied (with strdup() or the like in
99 the case of strings) before being inserted. */
103 The hash table is implemented as an open-addressed table with
104 linear probing collision resolution.
106 For those not up to CS parlance, it means that all the hash entries
107 (pairs of pointers key and value) are stored in a contiguous array.
108 The position of each mapping is determined by the hash value of its
109 key and the size of the table: location := hash(key) % size. If
110 two different keys end up on the same position (collide), the one
111 that came second is placed at the next empty position following the
112 occupied place. This collision resolution technique is called
115 There are more advanced collision resolution methods (quadratic
116 probing, double hashing), but we don't use them because they incur
117 more non-sequential access to the array, which results in worse CPU
118 cache behavior. Linear probing works well as long as the
119 count/size ratio (fullness) is kept below 75%. We make sure to
120 grow and rehash the table whenever this threshold is exceeded.
122 Collisions make deletion tricky because clearing a position
123 followed by a colliding entry would make the position seem empty
124 and the colliding entry not found. One solution is to leave a
125 "tombstone" instead of clearing the entry, and another is to
126 carefully rehash the entries immediately following the deleted one.
127 We use the latter method because it results in less bookkeeping and
128 faster retrieval at the (slight) expense of deletion. */
130 /* Maximum allowed fullness: when hash table's fullness exceeds this
131 value, the table is resized. */
132 #define HASH_MAX_FULLNESS 0.75
134 /* The hash table size is multiplied by this factor (and then rounded
135 to the next prime) with each resize. This guarantees infrequent
137 #define HASH_RESIZE_FACTOR 2
144 typedef unsigned long (*hashfun_t) PARAMS ((const void *));
145 typedef int (*testfun_t) PARAMS ((const void *, const void *));
148 hashfun_t hash_function;
149 testfun_t test_function;
151 struct mapping *mappings; /* pointer to the table entries. */
152 int size; /* size of the array. */
154 int count; /* number of non-empty entries. */
155 int resize_threshold; /* after size exceeds this number of
156 entries, resize the table. */
157 int prime_offset; /* the offset of the current prime in
161 /* We use all-bit-set marker to mean that a mapping is empty. It is
162 (hopefully) illegal as a pointer, and it allows the users to use
163 NULL (as well as any non-negative integer) as key. */
165 #define NON_EMPTY(mp) (mp->key != (void *)~(unsigned long)0)
166 #define MARK_AS_EMPTY(mp) (mp->key = (void *)~(unsigned long)0)
168 /* "Next" mapping is the mapping after MP, but wrapping back to
169 MAPPINGS when MP would reach MAPPINGS+SIZE. */
170 #define NEXT_MAPPING(mp, mappings, size) (mp != mappings + (size - 1) \
173 /* Loop over non-empty mappings starting at MP. */
174 #define LOOP_NON_EMPTY(mp, mappings, size) \
175 for (; NON_EMPTY (mp); mp = NEXT_MAPPING (mp, mappings, size))
177 /* Return the position of KEY in hash table SIZE large, hash function
178 being HASHFUN. #### Some implementations multiply HASHFUN's output
179 with the table's "golden ratio" to get better spreading of keys.
180 I'm not sure if that is necessary with our hash functions. */
181 #define HASH_POSITION(key, hashfun, size) ((hashfun) (key) % size)
183 /* Find a prime near, but greather than or equal to SIZE. Of course,
184 the primes are not calculated, but looked up from a table. The
185 table does not contain all primes in range, just a selection useful
188 PRIME_OFFSET is a minor optimization: it specifies start position
189 for the search for the large enough prime. The final offset is
190 stored in the same variable. That way the list of primes does not
191 have to be scanned from the beginning each time around. */
194 prime_size (int size, int *prime_offset)
196 static const unsigned long primes [] = {
197 13, 19, 29, 41, 59, 79, 107, 149, 197, 263, 347, 457, 599, 787, 1031,
198 1361, 1777, 2333, 3037, 3967, 5167, 6719, 8737, 11369, 14783,
199 19219, 24989, 32491, 42257, 54941, 71429, 92861, 120721, 156941,
200 204047, 265271, 344857, 448321, 582821, 757693, 985003, 1280519,
201 1664681, 2164111, 2813353, 3657361, 4754591, 6180989, 8035301,
202 10445899, 13579681, 17653589, 22949669, 29834603, 38784989,
203 50420551, 65546729, 85210757, 110774011, 144006217, 187208107,
204 243370577, 316381771, 411296309, 534685237, 695090819, 903618083,
205 1174703521, 1527114613, 1985248999,
206 (unsigned long)0x99d43ea5, (unsigned long)0xc7fa5177
210 for (i = *prime_offset; i < countof (primes); i++)
211 if (primes[i] >= size)
213 /* Set the offset to the next prime. That is safe because,
214 next time we are called, it will be with a larger SIZE,
215 which means we could never return the same prime anyway.
216 (If that is not the case, the caller can simply reset
218 *prime_offset = i + 1;
226 static unsigned long ptrhash PARAMS ((const void *));
227 static int ptrcmp PARAMS ((const void *, const void *));
229 /* Create a hash table with hash function HASH_FUNCTION and test
230 function TEST_FUNCTION. The table is empty (its count is 0), but
231 pre-allocated to store at least ITEMS items.
233 ITEMS is the number of items that the table can accept without
234 needing to resize. It is useful when creating a table that is to
235 be immediately filled with a known number of items. In that case,
236 the regrows are a waste of time, and specifying ITEMS correctly
237 will avoid them altogether.
239 Note that hash tables grow dynamically regardless of ITEMS. The
240 only use of ITEMS is to preallocate the table and avoid unnecessary
241 dynamic regrows. Don't bother making ITEMS prime because it's not
242 used as size unchanged. To start with a small table that grows as
243 needed, simply specify zero ITEMS.
245 If hash and test callbacks are not specified, identity mapping is
246 assumed, i.e. pointer values are used for key comparison. If,
247 instead of that, you want strings with equal contents to hash the
248 same, use make_string_hash_table. */
251 hash_table_new (int items,
252 unsigned long (*hash_function) (const void *),
253 int (*test_function) (const void *, const void *))
256 struct hash_table *ht = xnew (struct hash_table);
258 ht->hash_function = hash_function ? hash_function : ptrhash;
259 ht->test_function = test_function ? test_function : ptrcmp;
261 /* If the size of struct hash_table ever becomes a concern, this
262 field can go. (Wget doesn't create many hashes.) */
263 ht->prime_offset = 0;
265 /* Calculate the size that ensures that the table will store at
266 least ITEMS keys without the need to resize. */
267 size = 1 + items / HASH_MAX_FULLNESS;
268 size = prime_size (size, &ht->prime_offset);
270 ht->resize_threshold = size * HASH_MAX_FULLNESS;
271 /*assert (ht->resize_threshold >= items);*/
273 ht->mappings = xnew_array (struct mapping, ht->size);
274 /* Mark mappings as empty. We use 0xff rather than 0 to mark empty
275 keys because it allows us to store NULL keys to the table. */
276 memset (ht->mappings, 0xff, size * sizeof (struct mapping));
283 /* Free the data associated with hash table HT. */
286 hash_table_destroy (struct hash_table *ht)
288 xfree (ht->mappings);
292 /* The heart of most functions in this file -- find the mapping whose
293 KEY is equal to key, using linear probing. Returns the mapping
294 that matches KEY, or the first empty mapping if none matches. */
296 static inline struct mapping *
297 find_mapping (const struct hash_table *ht, const void *key)
299 struct mapping *mappings = ht->mappings;
301 struct mapping *mp = mappings + HASH_POSITION (key, ht->hash_function, size);
302 testfun_t equals = ht->test_function;
304 LOOP_NON_EMPTY (mp, mappings, size)
305 if (equals (key, mp->key))
310 /* Get the value that corresponds to the key KEY in the hash table HT.
311 If no value is found, return NULL. Note that NULL is a legal value
312 for value; if you are storing NULLs in your hash table, you can use
313 hash_table_contains to be sure that a (possibly NULL) value exists
314 in the table. Or, you can use hash_table_get_pair instead of this
318 hash_table_get (const struct hash_table *ht, const void *key)
320 struct mapping *mp = find_mapping (ht, key);
327 /* Like hash_table_get, but writes out the pointers to both key and
328 value. Returns non-zero on success. */
331 hash_table_get_pair (const struct hash_table *ht, const void *lookup_key,
332 void *orig_key, void *value)
334 struct mapping *mp = find_mapping (ht, lookup_key);
338 *(void **)orig_key = mp->key;
340 *(void **)value = mp->value;
347 /* Return 1 if HT contains KEY, 0 otherwise. */
350 hash_table_contains (const struct hash_table *ht, const void *key)
352 struct mapping *mp = find_mapping (ht, key);
353 return NON_EMPTY (mp);
356 /* Grow hash table HT as necessary, and rehash all the key-value
360 grow_hash_table (struct hash_table *ht)
362 hashfun_t hasher = ht->hash_function;
363 struct mapping *old_mappings = ht->mappings;
364 struct mapping *old_end = ht->mappings + ht->size;
365 struct mapping *mp, *mappings;
368 newsize = prime_size (ht->size * HASH_RESIZE_FACTOR, &ht->prime_offset);
370 printf ("growing from %d to %d; fullness %.2f%% to %.2f%%\n",
372 100.0 * ht->count / ht->size,
373 100.0 * ht->count / newsize);
377 ht->resize_threshold = newsize * HASH_MAX_FULLNESS;
379 mappings = xnew_array (struct mapping, newsize);
380 memset (mappings, 0xff, newsize * sizeof (struct mapping));
381 ht->mappings = mappings;
383 for (mp = old_mappings; mp < old_end; mp++)
386 struct mapping *new_mp;
387 /* We don't need to test for uniqueness of keys because they
388 come from the hash table and are therefore known to be
390 new_mp = mappings + HASH_POSITION (mp->key, hasher, newsize);
391 LOOP_NON_EMPTY (new_mp, mappings, newsize)
396 xfree (old_mappings);
399 /* Put VALUE in the hash table HT under the key KEY. This regrows the
400 table if necessary. */
403 hash_table_put (struct hash_table *ht, const void *key, void *value)
405 struct mapping *mp = find_mapping (ht, key);
408 /* update existing item */
409 mp->key = (void *)key; /* const? */
414 /* If adding the item would make the table exceed max. fullness,
415 grow the table first. */
416 if (ht->count >= ht->resize_threshold)
418 grow_hash_table (ht);
419 mp = find_mapping (ht, key);
424 mp->key = (void *)key; /* const? */
428 /* Remove a mapping that matches KEY from HT. Return 0 if there was
429 no such entry; return 1 if an entry was removed. */
432 hash_table_remove (struct hash_table *ht, const void *key)
434 struct mapping *mp = find_mapping (ht, key);
440 struct mapping *mappings = ht->mappings;
441 hashfun_t hasher = ht->hash_function;
446 /* Rehash all the entries following MP. The alternative
447 approach is to mark the entry as deleted, i.e. create a
448 "tombstone". That speeds up removal, but leaves a lot of
449 garbage and slows down hash_table_get and hash_table_put. */
451 mp = NEXT_MAPPING (mp, mappings, size);
452 LOOP_NON_EMPTY (mp, mappings, size)
454 const void *key2 = mp->key;
455 struct mapping *mp_new;
457 /* Find the new location for the key. */
458 mp_new = mappings + HASH_POSITION (key2, hasher, size);
459 LOOP_NON_EMPTY (mp_new, mappings, size)
460 if (key2 == mp_new->key)
461 /* The mapping MP (key2) is already where we want it (in
462 MP_NEW's "chain" of keys.) */
475 /* Clear HT of all entries. After calling this function, the count
476 and the fullness of the hash table will be zero. The size will
480 hash_table_clear (struct hash_table *ht)
482 memset (ht->mappings, 0xff, ht->size * sizeof (struct mapping));
486 /* Map MAPFUN over all the mappings in hash table HT. MAPFUN is
487 called with three arguments: the key, the value, and MAPARG.
489 It is undefined what happens if you add or remove entries in the
490 hash table while hash_table_map is running. The exception is the
491 entry you're currently mapping over; you may remove or change that
495 hash_table_map (struct hash_table *ht,
496 int (*mapfun) (void *, void *, void *),
499 struct mapping *mp = ht->mappings;
500 struct mapping *end = ht->mappings + ht->size;
502 for (; mp < end; mp++)
508 if (mapfun (key, mp->value, maparg))
510 /* hash_table_remove might have moved the adjacent
512 if (mp->key != key && NON_EMPTY (mp))
517 /* Return the number of elements in the hash table. This is not the
518 same as the physical size of the hash table, which is always
519 greater than the number of elements. */
522 hash_table_count (const struct hash_table *ht)
527 /* Functions from this point onward are meant for convenience and
528 don't strictly belong to this file. However, this is as good a
529 place for them as any. */
531 /* Rules for creating custom hash and test functions:
533 - The test function returns non-zero for keys that are considered
534 "equal", zero otherwise.
536 - The hash function returns a number that represents the
537 "distinctness" of the object. In more precise terms, it means
538 that for any two objects that test "equal" under the test
539 function, the hash function MUST produce the same result.
541 This does not mean that all different objects must produce
542 different values (that would be "perfect" hashing), only that
543 non-distinct objects must produce the same values! For instance,
544 a hash function that returns 0 for any given object is a
545 perfectly valid (albeit extremely bad) hash function. A hash
546 function that hashes a string by adding up all its characters is
547 another example of a valid (but quite bad) hash function.
549 It is not hard to make hash and test functions agree about
550 equality. For example, if the test function compares strings
551 case-insensitively, the hash function can lower-case the
552 characters when calculating the hash value. That ensures that
553 two strings differing only in case will hash the same.
555 - If you care about performance, choose a hash function with as
556 good "spreading" as possible. A good hash function will use all
557 the bits of the input when calculating the hash, and will react
558 to even small changes in input with a completely different
559 output. Finally, don't make the hash function itself overly
560 slow, because you'll be incurring a non-negligible overhead to
561 all hash table operations. */
564 * Support for hash tables whose keys are strings.
568 /* 31 bit hash function. Taken from Gnome's glib, modified to use
571 We used to use the popular hash function from the Dragon Book, but
572 this one seems to perform much better. */
575 string_hash (const void *key)
581 for (p += 1; *p != '\0'; p++)
582 h = (h << 5) - h + *p;
587 /* Frontend for strcmp usable for hash tables. */
590 string_cmp (const void *s1, const void *s2)
592 return !strcmp ((const char *)s1, (const char *)s2);
595 /* Return a hash table of preallocated to store at least ITEMS items
596 suitable to use strings as keys. */
599 make_string_hash_table (int items)
601 return hash_table_new (items, string_hash, string_cmp);
605 * Support for hash tables whose keys are strings, but which are
606 * compared case-insensitively.
610 /* Like string_hash, but produce the same hash regardless of the case. */
613 string_hash_nocase (const void *key)
616 unsigned int h = TOLOWER (*p);
619 for (p += 1; *p != '\0'; p++)
620 h = (h << 5) - h + TOLOWER (*p);
625 /* Like string_cmp, but doing case-insensitive compareison. */
628 string_cmp_nocase (const void *s1, const void *s2)
630 return !strcasecmp ((const char *)s1, (const char *)s2);
633 /* Like make_string_hash_table, but uses string_hash_nocase and
634 string_cmp_nocase. */
637 make_nocase_string_hash_table (int items)
639 return hash_table_new (items, string_hash_nocase, string_cmp_nocase);
642 /* Hashing of pointers. Used for hash tables that are keyed by
643 pointer identity. (Common Lisp calls them EQ hash tables, and Java
644 calls them IdentityHashMaps.) */
647 ptrhash (const void *ptr)
649 unsigned long key = (unsigned long)ptr;
672 ptrcmp (const void *ptr1, const void *ptr2)
683 print_hash_table_mapper (void *key, void *value, void *count)
686 printf ("%s: %s\n", (const char *)key, (char *)value);
691 print_hash (struct hash_table *sht)
694 hash_table_map (sht, print_hash_table_mapper, &debug_count);
695 assert (debug_count == sht->count);
701 struct hash_table *ht = make_string_hash_table (0);
703 while ((fgets (line, sizeof (line), stdin)))
705 int len = strlen (line);
709 if (!hash_table_contains (ht, line))
710 hash_table_put (ht, strdup (line), "here I am!");
715 if (hash_table_get_pair (ht, line, &line_copy, NULL))
717 hash_table_remove (ht, line);
727 printf ("%d %d\n", ht->count, ht->size);