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. */
30 /* With -DSTANDALONE, this file can be compiled outside Wget source
31 tree. To test, also use -DTEST. */
44 /* If running without Autoconf, go ahead and assume presence of
45 standard C89 headers. */
55 /* Get Wget's utility headers. */
59 /* Make do without them. */
60 # define xnew(x) xmalloc (sizeof (x))
61 # define xnew_array(type, x) xmalloc (sizeof (type) * (x))
62 # define xmalloc malloc /* or something that exits
63 if not enough memory */
65 # define countof(x) (sizeof (x) / sizeof ((x)[0]))
66 # define TOLOWER(x) ('A' <= (x) && (x) <= 'Z' ? (x) - 32 : (x))
74 Hash tables are a technique used to implement mapping between
75 objects with near-constant-time access and storage. The table
76 associates keys to values, and a value can be very quickly
77 retrieved by providing the key. Fast lookup tables are typically
78 implemented as hash tables.
81 hash_table_new -- creates the table.
82 hash_table_destroy -- destroys the table.
83 hash_table_put -- establishes or updates key->value mapping.
84 hash_table_get -- retrieves value of key.
85 hash_table_get_pair -- get key/value pair for key.
86 hash_table_contains -- test whether the table contains key.
87 hash_table_remove -- remove the key->value mapping for key.
88 hash_table_map -- iterate through table mappings.
89 hash_table_clear -- clear hash table contents.
90 hash_table_count -- return the number of entries in the table.
92 The hash table grows internally as new entries are added and is not
93 limited in size, except by available memory. The table doubles
94 with each resize, which ensures that the amortized time per
95 operation remains constant.
97 By default, tables created by hash_table_new consider the keys to
98 be equal if their pointer values are the same. You can use
99 make_string_hash_table to create tables whose keys are considered
100 equal if their string contents are the same. In the general case,
101 the criterion of equality used to compare keys is specified at
102 table creation time with two callback functions, "hash" and "test".
103 The hash function transforms the key into an arbitrary number that
104 must be the same for two equal keys. The test function accepts two
105 keys and returns non-zero if they are to be considered equal.
107 Note that neither keys nor values are copied when inserted into the
108 hash table, so they must exist for the lifetime of the table. This
109 means that e.g. the use of static strings is OK, but objects with a
110 shorter life-time need to be copied (with strdup() or the like in
111 the case of strings) before being inserted. */
115 The hash table is implemented as an open-addressed table with
116 linear probing collision resolution.
118 The above means that all the hash entries (pairs of pointers, key
119 and value) are stored in a contiguous array. The position of each
120 mapping is determined by the hash value of its key and the size of
121 the table: location := hash(key) % size. If two different keys end
122 up on the same position (collide), the one that came second is
123 placed at the next empty position following the occupied place.
124 This collision resolution technique is called "linear probing".
126 There are more advanced collision resolution methods (quadratic
127 probing, double hashing), but we don't use them because they incur
128 more non-sequential access to the array, which results in worse CPU
129 cache behavior. Linear probing works well as long as the
130 count/size ratio (fullness) is kept below 75%. We make sure to
131 grow and rehash the table whenever this threshold is exceeded.
133 Collisions make deletion tricky because clearing a position
134 followed by a colliding entry would make the position seem empty
135 and the colliding entry not found. One solution is to leave a
136 "tombstone" instead of clearing the entry, and another is to
137 carefully rehash the entries immediately following the deleted one.
138 We use the latter method because it results in less bookkeeping and
139 faster retrieval at the (slight) expense of deletion. */
141 /* Maximum allowed fullness: when hash table's fullness exceeds this
142 value, the table is resized. */
143 #define HASH_MAX_FULLNESS 0.75
145 /* The hash table size is multiplied by this factor (and then rounded
146 to the next prime) with each resize. This guarantees infrequent
148 #define HASH_RESIZE_FACTOR 2
155 typedef unsigned long (*hashfun_t) PARAMS ((const void *));
156 typedef int (*testfun_t) PARAMS ((const void *, const void *));
159 hashfun_t hash_function;
160 testfun_t test_function;
162 struct mapping *mappings; /* pointer to the table entries. */
163 int size; /* size of the array. */
165 int count; /* number of non-empty entries. */
166 int resize_threshold; /* after size exceeds this number of
167 entries, resize the table. */
168 int prime_offset; /* the offset of the current prime in
172 /* We use the all-bits-set constant (INVALID_PTR) marker to mean that
173 a mapping is empty. It is unaligned and therefore illegal as a
174 pointer. INVALID_PTR_BYTE (0xff) is the one-byte value used to
175 initialize the mappings array as empty.
177 The all-bits-set value is a better choice than NULL because it
178 allows the use of NULL/0 keys. Since the keys are either integers
179 or pointers, the only key that cannot be used is the integer value
180 -1. This is acceptable because it still allows the use of
181 nonnegative integer keys. */
183 #define INVALID_PTR ((void *) ~(unsigned long)0)
185 # define UCHAR_MAX 0xff
187 #define INVALID_PTR_BYTE UCHAR_MAX
189 #define NON_EMPTY(mp) ((mp)->key != INVALID_PTR)
190 #define MARK_AS_EMPTY(mp) ((mp)->key = INVALID_PTR)
192 /* "Next" mapping is the mapping after MP, but wrapping back to
193 MAPPINGS when MP would reach MAPPINGS+SIZE. */
194 #define NEXT_MAPPING(mp, mappings, size) (mp != mappings + (size - 1) \
197 /* Loop over non-empty mappings starting at MP. */
198 #define LOOP_NON_EMPTY(mp, mappings, size) \
199 for (; NON_EMPTY (mp); mp = NEXT_MAPPING (mp, mappings, size))
201 /* Return the position of KEY in hash table SIZE large, hash function
203 #define HASH_POSITION(key, hashfun, size) ((hashfun) (key) % size)
205 /* Find a prime near, but greather than or equal to SIZE. The primes
206 are looked up from a table with a selection of primes convenient
209 PRIME_OFFSET is a minor optimization: it specifies start position
210 for the search for the large enough prime. The final offset is
211 stored in the same variable. That way the list of primes does not
212 have to be scanned from the beginning each time around. */
215 prime_size (int size, int *prime_offset)
217 static const int primes[] = {
218 13, 19, 29, 41, 59, 79, 107, 149, 197, 263, 347, 457, 599, 787, 1031,
219 1361, 1777, 2333, 3037, 3967, 5167, 6719, 8737, 11369, 14783,
220 19219, 24989, 32491, 42257, 54941, 71429, 92861, 120721, 156941,
221 204047, 265271, 344857, 448321, 582821, 757693, 985003, 1280519,
222 1664681, 2164111, 2813353, 3657361, 4754591, 6180989, 8035301,
223 10445899, 13579681, 17653589, 22949669, 29834603, 38784989,
224 50420551, 65546729, 85210757, 110774011, 144006217, 187208107,
225 243370577, 316381771, 411296309, 534685237, 695090819, 903618083,
226 1174703521, 1527114613, 1837299131, 2147483647
230 for (i = *prime_offset; i < countof (primes); i++)
231 if (primes[i] >= size)
233 /* Set the offset to the next prime. That is safe because,
234 next time we are called, it will be with a larger SIZE,
235 which means we could never return the same prime anyway.
236 (If that is not the case, the caller can simply reset
238 *prime_offset = i + 1;
245 static int cmp_pointer PARAMS ((const void *, const void *));
247 /* Create a hash table with hash function HASH_FUNCTION and test
248 function TEST_FUNCTION. The table is empty (its count is 0), but
249 pre-allocated to store at least ITEMS items.
251 ITEMS is the number of items that the table can accept without
252 needing to resize. It is useful when creating a table that is to
253 be immediately filled with a known number of items. In that case,
254 the regrows are a waste of time, and specifying ITEMS correctly
255 will avoid them altogether.
257 Note that hash tables grow dynamically regardless of ITEMS. The
258 only use of ITEMS is to preallocate the table and avoid unnecessary
259 dynamic regrows. Don't bother making ITEMS prime because it's not
260 used as size unchanged. To start with a small table that grows as
261 needed, simply specify zero ITEMS.
263 If hash and test callbacks are not specified, identity mapping is
264 assumed, i.e. pointer values are used for key comparison. (Common
265 Lisp calls such tables EQ hash tables, and Java calls them
266 IdentityHashMaps.) If your keys require different comparison,
267 specify hash and test functions. For easy use of C strings as hash
268 keys, you can use the convenience functions make_string_hash_table
269 and make_nocase_string_hash_table. */
272 hash_table_new (int items,
273 unsigned long (*hash_function) (const void *),
274 int (*test_function) (const void *, const void *))
277 struct hash_table *ht = xnew (struct hash_table);
279 ht->hash_function = hash_function ? hash_function : hash_pointer;
280 ht->test_function = test_function ? test_function : cmp_pointer;
282 /* If the size of struct hash_table ever becomes a concern, this
283 field can go. (Wget doesn't create many hashes.) */
284 ht->prime_offset = 0;
286 /* Calculate the size that ensures that the table will store at
287 least ITEMS keys without the need to resize. */
288 size = 1 + items / HASH_MAX_FULLNESS;
289 size = prime_size (size, &ht->prime_offset);
291 ht->resize_threshold = size * HASH_MAX_FULLNESS;
292 /*assert (ht->resize_threshold >= items);*/
294 ht->mappings = xnew_array (struct mapping, ht->size);
296 /* Mark mappings as empty. We use 0xff rather than 0 to mark empty
297 keys because it allows us to use NULL/0 as keys. */
298 memset (ht->mappings, INVALID_PTR_BYTE, size * sizeof (struct mapping));
305 /* Free the data associated with hash table HT. */
308 hash_table_destroy (struct hash_table *ht)
310 xfree (ht->mappings);
314 /* The heart of most functions in this file -- find the mapping whose
315 KEY is equal to key, using linear probing. Returns the mapping
316 that matches KEY, or the first empty mapping if none matches. */
318 static inline struct mapping *
319 find_mapping (const struct hash_table *ht, const void *key)
321 struct mapping *mappings = ht->mappings;
323 struct mapping *mp = mappings + HASH_POSITION (key, ht->hash_function, size);
324 testfun_t equals = ht->test_function;
326 LOOP_NON_EMPTY (mp, mappings, size)
327 if (equals (key, mp->key))
332 /* Get the value that corresponds to the key KEY in the hash table HT.
333 If no value is found, return NULL. Note that NULL is a legal value
334 for value; if you are storing NULLs in your hash table, you can use
335 hash_table_contains to be sure that a (possibly NULL) value exists
336 in the table. Or, you can use hash_table_get_pair instead of this
340 hash_table_get (const struct hash_table *ht, const void *key)
342 struct mapping *mp = find_mapping (ht, key);
349 /* Like hash_table_get, but writes out the pointers to both key and
350 value. Returns non-zero on success. */
353 hash_table_get_pair (const struct hash_table *ht, const void *lookup_key,
354 void *orig_key, void *value)
356 struct mapping *mp = find_mapping (ht, lookup_key);
360 *(void **)orig_key = mp->key;
362 *(void **)value = mp->value;
369 /* Return 1 if HT contains KEY, 0 otherwise. */
372 hash_table_contains (const struct hash_table *ht, const void *key)
374 struct mapping *mp = find_mapping (ht, key);
375 return NON_EMPTY (mp);
378 /* Grow hash table HT as necessary, and rehash all the key-value
382 grow_hash_table (struct hash_table *ht)
384 hashfun_t hasher = ht->hash_function;
385 struct mapping *old_mappings = ht->mappings;
386 struct mapping *old_end = ht->mappings + ht->size;
387 struct mapping *mp, *mappings;
390 newsize = prime_size (ht->size * HASH_RESIZE_FACTOR, &ht->prime_offset);
392 printf ("growing from %d to %d; fullness %.2f%% to %.2f%%\n",
394 100.0 * ht->count / ht->size,
395 100.0 * ht->count / newsize);
399 ht->resize_threshold = newsize * HASH_MAX_FULLNESS;
401 mappings = xnew_array (struct mapping, newsize);
402 memset (mappings, INVALID_PTR_BYTE, newsize * sizeof (struct mapping));
403 ht->mappings = mappings;
405 for (mp = old_mappings; mp < old_end; mp++)
408 struct mapping *new_mp;
409 /* We don't need to test for uniqueness of keys because they
410 come from the hash table and are therefore known to be
412 new_mp = mappings + HASH_POSITION (mp->key, hasher, newsize);
413 LOOP_NON_EMPTY (new_mp, mappings, newsize)
418 xfree (old_mappings);
421 /* Put VALUE in the hash table HT under the key KEY. This regrows the
422 table if necessary. */
425 hash_table_put (struct hash_table *ht, const void *key, void *value)
427 struct mapping *mp = find_mapping (ht, key);
430 /* update existing item */
431 mp->key = (void *)key; /* const? */
436 /* If adding the item would make the table exceed max. fullness,
437 grow the table first. */
438 if (ht->count >= ht->resize_threshold)
440 grow_hash_table (ht);
441 mp = find_mapping (ht, key);
446 mp->key = (void *)key; /* const? */
450 /* Remove a mapping that matches KEY from HT. Return 0 if there was
451 no such entry; return 1 if an entry was removed. */
454 hash_table_remove (struct hash_table *ht, const void *key)
456 struct mapping *mp = find_mapping (ht, key);
462 struct mapping *mappings = ht->mappings;
463 hashfun_t hasher = ht->hash_function;
468 /* Rehash all the entries following MP. The alternative
469 approach is to mark the entry as deleted, i.e. create a
470 "tombstone". That speeds up removal, but leaves a lot of
471 garbage and slows down hash_table_get and hash_table_put. */
473 mp = NEXT_MAPPING (mp, mappings, size);
474 LOOP_NON_EMPTY (mp, mappings, size)
476 const void *key2 = mp->key;
477 struct mapping *mp_new;
479 /* Find the new location for the key. */
480 mp_new = mappings + HASH_POSITION (key2, hasher, size);
481 LOOP_NON_EMPTY (mp_new, mappings, size)
482 if (key2 == mp_new->key)
483 /* The mapping MP (key2) is already where we want it (in
484 MP_NEW's "chain" of keys.) */
497 /* Clear HT of all entries. After calling this function, the count
498 and the fullness of the hash table will be zero. The size will
502 hash_table_clear (struct hash_table *ht)
504 memset (ht->mappings, INVALID_PTR_BYTE, ht->size * sizeof (struct mapping));
508 /* Map MAPFUN over all the mappings in hash table HT. MAPFUN is
509 called with three arguments: the key, the value, and MAPARG.
511 It is undefined what happens if you add or remove entries in the
512 hash table while hash_table_map is running. The exception is the
513 entry you're currently mapping over; you may remove or change that
517 hash_table_map (struct hash_table *ht,
518 int (*mapfun) (void *, void *, void *),
521 struct mapping *mp = ht->mappings;
522 struct mapping *end = ht->mappings + ht->size;
524 for (; mp < end; mp++)
530 if (mapfun (key, mp->value, maparg))
532 /* hash_table_remove might have moved the adjacent
534 if (mp->key != key && NON_EMPTY (mp))
539 /* Return the number of elements in the hash table. This is not the
540 same as the physical size of the hash table, which is always
541 greater than the number of elements. */
544 hash_table_count (const struct hash_table *ht)
549 /* Functions from this point onward are meant for convenience and
550 don't strictly belong to this file. However, this is as good a
551 place for them as any. */
553 /* Guidelines for creating custom hash and test functions:
555 - The test function returns non-zero for keys that are considered
556 "equal", zero otherwise.
558 - The hash function returns a number that represents the
559 "distinctness" of the object. In more precise terms, it means
560 that for any two objects that test "equal" under the test
561 function, the hash function MUST produce the same result.
563 This does not mean that all different objects must produce
564 different values (that would be "perfect" hashing), only that
565 non-distinct objects must produce the same values! For instance,
566 a hash function that returns 0 for any given object is a
567 perfectly valid (albeit extremely bad) hash function. A hash
568 function that hashes a string by adding up all its characters is
569 another example of a valid (but quite bad) hash function.
571 It is not hard to make hash and test functions agree about
572 equality. For example, if the test function compares strings
573 case-insensitively, the hash function can lower-case the
574 characters when calculating the hash value. That ensures that
575 two strings differing only in case will hash the same.
577 - If you care about performance, choose a hash function with as
578 good "spreading" as possible. A good hash function will use all
579 the bits of the input when calculating the hash, and will react
580 to even small changes in input with a completely different
581 output. Finally, don't make the hash function itself overly
582 slow, because you'll be incurring a non-negligible overhead to
583 all hash table operations. */
586 * Support for hash tables whose keys are strings.
590 /* 31 bit hash function. Taken from Gnome's glib, modified to use
593 We used to use the popular hash function from the Dragon Book, but
594 this one seems to perform much better. */
597 hash_string (const void *key)
603 for (p += 1; *p != '\0'; p++)
604 h = (h << 5) - h + *p;
609 /* Frontend for strcmp usable for hash tables. */
612 cmp_string (const void *s1, const void *s2)
614 return !strcmp ((const char *)s1, (const char *)s2);
617 /* Return a hash table of preallocated to store at least ITEMS items
618 suitable to use strings as keys. */
621 make_string_hash_table (int items)
623 return hash_table_new (items, hash_string, cmp_string);
627 * Support for hash tables whose keys are strings, but which are
628 * compared case-insensitively.
632 /* Like hash_string, but produce the same hash regardless of the case. */
635 hash_string_nocase (const void *key)
638 unsigned int h = TOLOWER (*p);
641 for (p += 1; *p != '\0'; p++)
642 h = (h << 5) - h + TOLOWER (*p);
647 /* Like string_cmp, but doing case-insensitive compareison. */
650 string_cmp_nocase (const void *s1, const void *s2)
652 return !strcasecmp ((const char *)s1, (const char *)s2);
655 /* Like make_string_hash_table, but uses string_hash_nocase and
656 string_cmp_nocase. */
659 make_nocase_string_hash_table (int items)
661 return hash_table_new (items, hash_string_nocase, string_cmp_nocase);
664 /* Hashing of numeric values, such as pointers and integers.
666 This implementation is the Robert Jenkins' 32 bit Mix Function,
667 with a simple adaptation for 64-bit values. It offers excellent
668 spreading of values and doesn't need to know the hash table size to
669 work (unlike the very popular Knuth's multiplication hash). */
672 hash_pointer (const void *ptr)
674 unsigned long key = (unsigned long)ptr;
697 cmp_pointer (const void *ptr1, const void *ptr2)
708 print_hash_table_mapper (void *key, void *value, void *count)
711 printf ("%s: %s\n", (const char *)key, (char *)value);
716 print_hash (struct hash_table *sht)
719 hash_table_map (sht, print_hash_table_mapper, &debug_count);
720 assert (debug_count == sht->count);
726 struct hash_table *ht = make_string_hash_table (0);
728 while ((fgets (line, sizeof (line), stdin)))
730 int len = strlen (line);
734 if (!hash_table_contains (ht, line))
735 hash_table_put (ht, strdup (line), "here I am!");
740 if (hash_table_get_pair (ht, line, &line_copy, NULL))
742 hash_table_remove (ht, line);
752 printf ("%d %d\n", ht->count, ht->size);