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. */
28 #endif /* HAVE_STRING_H */
42 # define xmalloc malloc
43 # define xrealloc realloc
47 # define TOLOWER(x) ('A' <= (x) && (x) <= 'Z' ? (x) - 32 : (x))
52 Hash tables are an implementation technique used to implement
53 mapping between objects. Provided a good hashing function is used,
54 they guarantee constant-time access and storing of information.
55 Duplicate keys are not allowed.
57 The basics are all covered. hash_table_new creates a hash table,
58 and hash_table_destroy deletes it. hash_table_put establishes a
59 mapping between a key and a value. hash_table_get retrieves the
60 value that corresponds to a key. hash_table_contains queries
61 whether a key is stored in a table at all. hash_table_remove
62 removes a mapping that corresponds to a key. hash_table_map allows
63 you to map through all the entries in a hash table.
64 hash_table_clear clears all the entries from the hash table.
66 The number of mappings in a table is not limited, except by the
67 amount of memory. As you add new elements to a table, it regrows
68 as necessary. If you have an idea about how many elements you will
69 store, you can provide a hint to hash_table_new().
71 The hashing and equality functions are normally provided by the
72 user. For the special (and frequent) case of hashing strings, you
73 can use the pre-canned make_string_hash_table(), which provides an
74 efficient string hashing function, and a string equality wrapper
77 When specifying your own hash and test functions, make sure the
80 - The test function returns non-zero for keys that are considered
81 "equal", zero otherwise.
83 - The hash function returns a number that represents the
84 "distinctness" of the object. In more precise terms, it means
85 that for any two objects that test "equal" under the test
86 function, the hash function MUST produce the same result.
88 This does not mean that each distinct object must produce a
89 distinct value, only that non-distinct objects must produce the
90 same values! For instance, a hash function that returns 0 for
91 any given object is a perfectly valid (albeit extremely bad) hash
92 function. A hash function that hashes a string by adding up all
93 its characters is another example of a valid (but quite bad) hash
96 The above stated rule is quite easy to enforce. For example, if
97 your testing function compares strings case-insensitively, all
98 your function needs to do is lower-case the string characters
99 before calculating a hash. That way you have easily guaranteed
100 that case differences will not result in a different hash.
102 - (optional) Choose the hash function to get as good "spreading" as
103 possible. A good hash function will react to even a small change
104 in its input with a completely different resulting hash.
105 Finally, don't make your hash function extremely slow, because
106 you're then defeating the purpose of hashing.
108 Note that neither keys nor values are copied when inserted into the
109 hash table, so they must exist for the lifetime of the table. This
110 means that e.g. the use of static strings is OK, but objects with a
111 shorter life-time need to be copied (with strdup() or the like in
112 the case of strings) before being inserted. */
116 All the hash mappings (key-value pairs of pointers) are stored in a
117 contiguous array. The position of each mapping is determined by
118 applying the hash function to the key: location = hash(key) % size.
119 If two different keys end up on the same position, the collision is
120 resolved by placing the second mapping at the next empty place in
121 the array following the occupied place. This method of collision
122 resolution is called "linear probing".
124 There are more advanced collision resolution mechanisms (quadratic
125 probing, double hashing), but we don't use them because they
126 involve more non-sequential access to the array, and therefore
127 worse cache behavior. Linear probing works well as long as the
128 fullness/size ratio is kept below 75%. We make sure to regrow or
129 rehash the hash table whenever this threshold is exceeded.
131 Collisions make deletion tricky because finding collisions again
132 relies on new empty spots not being created. That's why
133 hash_table_remove is careful to rehash the mappings that follow the
142 unsigned long (*hash_function) PARAMS ((const void *));
143 int (*test_function) PARAMS ((const void *, const void *));
145 int size; /* size of the array */
146 int count; /* number of non-empty, non-deleted
149 int resize_threshold; /* after size exceeds this number of
150 entries, resize the table. */
151 int prime_offset; /* the offset of the current prime in
154 struct mapping *mappings; /* the array of mapping pairs. */
157 #define EMPTY_MAPPING_P(mp) ((mp)->key == NULL)
158 #define NEXT_MAPPING(mp, mappings, size) (mp == mappings + (size - 1) \
161 #define LOOP_NON_EMPTY(mp, mappings, size) \
162 for (; !EMPTY_MAPPING_P (mp); mp = NEXT_MAPPING (mp, mappings, size))
164 /* #### We might want to multiply with the "golden ratio" here to get
165 better randomness for keys that do not result from a good hash
166 function. This is currently not a problem in Wget because we only
167 use the string hash tables. */
169 #define HASH_POSITION(ht, key) (ht->hash_function (key) % ht->size)
171 /* Find a prime near, but greather than or equal to SIZE. Of course,
172 the primes are not calculated, but looked up from a table. The
173 table does not contain all primes in range, just a selection useful
176 PRIME_OFFSET is a micro-optimization: if specified, it starts the
177 search for the prime number beginning with the specific offset in
178 the prime number table. The final offset is stored in the same
182 prime_size (int size, int *prime_offset)
184 static const unsigned long primes [] = {
185 13, 19, 29, 41, 59, 79, 107, 149, 197, 263, 347, 457, 599, 787, 1031,
186 1361, 1777, 2333, 3037, 3967, 5167, 6719, 8737, 11369, 14783,
187 19219, 24989, 32491, 42257, 54941, 71429, 92861, 120721, 156941,
188 204047, 265271, 344857, 448321, 582821, 757693, 985003, 1280519,
189 1664681, 2164111, 2813353, 3657361, 4754591, 6180989, 8035301,
190 10445899, 13579681, 17653589, 22949669, 29834603, 38784989,
191 50420551, 65546729, 85210757, 110774011, 144006217, 187208107,
192 243370577, 316381771, 411296309, 534685237, 695090819, 903618083,
193 1174703521, 1527114613, 1985248999,
194 (unsigned long)0x99d43ea5, (unsigned long)0xc7fa5177
196 int i = *prime_offset;
198 for (; i < ARRAY_SIZE (primes); i++)
199 if (primes[i] >= size)
201 /* Set the offset to the next prime. That is safe because,
202 next time we are called, it will be with a larger SIZE,
203 which means we could never return the same prime anyway.
204 (If that is not the case, the caller can simply reset
206 *prime_offset = i + 1;
214 /* Create a hash table of INITIAL_SIZE with hash function
215 HASH_FUNCTION and test function TEST_FUNCTION. INITIAL_SIZE will
216 be rounded to the next prime, so you don't have to worry about it
217 being a prime number.
219 Consequently, if you wish to start out with a "small" table which
220 will be regrown as needed, specify INITIAL_SIZE 0. */
223 hash_table_new (int initial_size,
224 unsigned long (*hash_function) (const void *),
225 int (*test_function) (const void *, const void *))
227 struct hash_table *ht
228 = (struct hash_table *)xmalloc (sizeof (struct hash_table));
230 ht->hash_function = hash_function;
231 ht->test_function = test_function;
233 ht->prime_offset = 0;
234 ht->size = prime_size (initial_size, &ht->prime_offset);
235 ht->resize_threshold = ht->size * 3 / 4;
239 ht->mappings = xmalloc (ht->size * sizeof (struct mapping));
240 memset (ht->mappings, '\0', ht->size * sizeof (struct mapping));
245 /* Free the data associated with hash table HT. */
248 hash_table_destroy (struct hash_table *ht)
250 xfree (ht->mappings);
254 /* The heart of almost all functions in this file -- find the mapping
255 whose KEY is equal to key, using linear probing. Returns the
256 mapping that matches KEY, or NULL if none matches. */
258 static inline struct mapping *
259 find_mapping (struct hash_table *ht, const void *key)
261 struct mapping *mappings = ht->mappings;
263 struct mapping *mp = mappings + HASH_POSITION (ht, key);
264 int (*equals) PARAMS ((const void *, const void *)) = ht->test_function;
266 LOOP_NON_EMPTY (mp, mappings, size)
267 if (equals (key, mp->key))
272 /* Get the value that corresponds to the key KEY in the hash table HT.
273 If no value is found, return NULL. Note that NULL is a legal value
274 for value; if you are storing NULLs in your hash table, you can use
275 hash_table_contains to be sure that a (possibly NULL) value exists
276 in the table. Or, you can use hash_table_get_pair instead of this
280 hash_table_get (struct hash_table *ht, const void *key)
282 struct mapping *mp = find_mapping (ht, key);
289 /* Like hash_table_get, but writes out the pointers to both key and
290 value. Returns non-zero on success. */
293 hash_table_get_pair (struct hash_table *ht, const void *lookup_key,
294 void *orig_key, void *value)
296 struct mapping *mp = find_mapping (ht, lookup_key);
301 *(void **)orig_key = mp->key;
303 *(void **)value = mp->value;
310 /* Return 1 if HT contains KEY, 0 otherwise. */
313 hash_table_contains (struct hash_table *ht, const void *key)
315 return find_mapping (ht, key) != NULL;
318 /* Grow hash table HT as necessary, and rehash all the key-value
322 grow_hash_table (struct hash_table *ht)
324 struct mapping *old_mappings = ht->mappings;
325 struct mapping *old_end = ht->mappings + ht->size;
326 struct mapping *mp, *mappings;
329 newsize = prime_size (ht->size * 2, &ht->prime_offset);
331 printf ("growing from %d to %d; fullness %.2f%% to %.2f%%\n",
333 (double)100 * ht->count / ht->size,
334 (double)100 * ht->count / newsize);
338 ht->resize_threshold = newsize * 3 / 4;
340 mappings = xmalloc (ht->size * sizeof (struct mapping));
341 memset (mappings, '\0', ht->size * sizeof (struct mapping));
342 ht->mappings = mappings;
344 for (mp = old_mappings; mp < old_end; mp++)
345 if (!EMPTY_MAPPING_P (mp))
347 struct mapping *new_mp = mappings + HASH_POSITION (ht, mp->key);
348 /* We don't need to call test function and worry about
349 collisions because all the keys come from the hash table
350 and are therefore guaranteed to be unique. */
351 LOOP_NON_EMPTY (new_mp, mappings, newsize)
356 xfree (old_mappings);
359 /* Put VALUE in the hash table HT under the key KEY. This regrows the
360 table if necessary. */
363 hash_table_put (struct hash_table *ht, const void *key, void *value)
365 struct mapping *mappings = ht->mappings;
367 int (*equals) PARAMS ((const void *, const void *)) = ht->test_function;
369 struct mapping *mp = mappings + HASH_POSITION (ht, key);
371 LOOP_NON_EMPTY (mp, mappings, size)
372 if (equals (key, mp->key))
374 mp->key = (void *)key; /* const? */
380 mp->key = (void *)key; /* const? */
383 if (ht->count > ht->resize_threshold)
384 /* When table is 75% full, regrow it. */
385 grow_hash_table (ht);
388 /* Remove a mapping that matches KEY from HT. Return 0 if there was
389 no such entry; return 1 if an entry was removed. */
392 hash_table_remove (struct hash_table *ht, const void *key)
394 struct mapping *mp = find_mapping (ht, key);
400 struct mapping *mappings = ht->mappings;
405 /* Rehash all the entries following MP. The alternative
406 approach is to mark the entry as deleted, i.e. create a
407 "tombstone". That makes remove faster, but leaves a lot of
408 garbage and slows down hash_table_get and hash_table_put. */
410 mp = NEXT_MAPPING (mp, mappings, size);
411 LOOP_NON_EMPTY (mp, mappings, size)
413 const void *key2 = mp->key;
414 struct mapping *mp_new = mappings + HASH_POSITION (ht, key2);
416 /* Find the new location for the key. */
418 LOOP_NON_EMPTY (mp_new, mappings, size)
419 if (key2 == mp_new->key)
420 /* The mapping MP (key2) is already where we want it (in
421 MP_NEW's "chain" of keys.) */
434 /* Clear HT of all entries. After calling this function, the count
435 and the fullness of the hash table will be zero. The size will
439 hash_table_clear (struct hash_table *ht)
441 memset (ht->mappings, '\0', ht->size * sizeof (struct mapping));
445 /* Map MAPFUN over all the mappings in hash table HT. MAPFUN is
446 called with three arguments: the key, the value, and the CLOSURE.
448 It is undefined what happens if you add or remove entries in the
449 hash table while hash_table_map is running. The exception is the
450 entry you're currently mapping over; you may remove or change that
454 hash_table_map (struct hash_table *ht,
455 int (*mapfun) (void *, void *, void *),
458 struct mapping *mp = ht->mappings;
459 struct mapping *end = ht->mappings + ht->size;
461 for (; mp < end; mp++)
462 if (!EMPTY_MAPPING_P (mp))
467 if (mapfun (key, mp->value, closure))
469 /* hash_table_remove might have moved the adjacent
471 if (mp->key != key && !EMPTY_MAPPING_P (mp))
476 /* Return the number of elements in the hash table. This is not the
477 same as the physical size of the hash table, which is always
478 greater than the number of elements. */
481 hash_table_count (struct hash_table *ht)
486 /* Functions from this point onward are meant for convenience and
487 don't strictly belong to this file. However, this is as good a
488 place for them as any. */
491 Support for hash tables whose keys are strings.
494 /* 31 bit hash function. Taken from Gnome's glib, modified to use
497 We used to use the popular hash function from the Dragon Book, but
498 this one seems to perform much better. */
501 string_hash (const void *key)
507 for (p += 1; *p != '\0'; p++)
508 h = (h << 5) - h + *p;
513 /* Frontend for strcmp usable for hash tables. */
516 string_cmp (const void *s1, const void *s2)
518 return !strcmp ((const char *)s1, (const char *)s2);
521 /* Return a hash table of initial size INITIAL_SIZE suitable to use
525 make_string_hash_table (int initial_size)
527 return hash_table_new (initial_size, string_hash, string_cmp);
531 Support for hash tables whose keys are strings, but which are
532 compared case-insensitively.
535 /* Like string_hash, but produce the same hash regardless of the case. */
538 string_hash_nocase (const void *key)
541 unsigned int h = TOLOWER (*p);
544 for (p += 1; *p != '\0'; p++)
545 h = (h << 5) - h + TOLOWER (*p);
550 /* Like string_cmp, but doing case-insensitive compareison. */
553 string_cmp_nocase (const void *s1, const void *s2)
555 return !strcasecmp ((const char *)s1, (const char *)s2);
558 /* Like make_string_hash_table, but uses string_hash_nocase and
559 string_cmp_nocase. */
562 make_nocase_string_hash_table (int initial_size)
564 return hash_table_new (initial_size, string_hash_nocase, string_cmp_nocase);
568 /* If I ever need it: hashing of integers. */
571 inthash (unsigned int key)
591 print_hash_table_mapper (void *key, void *value, void *count)
594 printf ("%s: %s\n", (const char *)key, (char *)value);
599 print_hash (struct hash_table *sht)
602 hash_table_map (sht, print_hash_table_mapper, &debug_count);
603 assert (debug_count == sht->count);
609 struct hash_table *ht = make_string_hash_table (0);
611 while ((fgets (line, sizeof (line), stdin)))
613 int len = strlen (line);
617 if (!hash_table_contains (ht, line))
618 hash_table_put (ht, strdup (line), "here I am!");
623 if (hash_table_get_pair (ht, line, &line_copy, NULL))
625 hash_table_remove (ht, line);
635 printf ("%d %d\n", ht->count, ht->size);