1 /* PSPP - a program for statistical analysis.
2 Copyright (C) 2008, 2009, 2010 Free Software Foundation, Inc.
4 This program is free software: you can redistribute it and/or modify
5 it under the terms of the GNU General Public License as published by
6 the Free Software Foundation, either version 3 of the License, or
7 (at your option) any later version.
9 This program is distributed in the hope that it will be useful,
10 but WITHOUT ANY WARRANTY; without even the implied warranty of
11 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 GNU General Public License for more details.
14 You should have received a copy of the GNU General Public License
15 along with this program. If not, see <http://www.gnu.org/licenses/>. */
17 /* Hash table with separate chaining.
19 This header (hmap.h) supplies an "embedded" implementation of
20 a hash table that uses linked lists to resolve collisions
21 ("separate chaining"). Its companion header (hmapx.h)
22 supplies a "external" implementation that is otherwise
23 similar. The two variants are described briefly here. The
24 embedded variant, for which this is the header, is described
25 in slightly more detail below. Each function also has a
26 detailed usage comment at its point of definition. (Many of
27 those definitions are inline in this file, because they are so
28 simple. Others are in hmap.c.)
30 The "hmap" embedded hash table implementation puts the hash
31 table node (which includes the linked list used for resolving
32 collisions) within the data structure that the hash table
33 contains. This makes allocation efficient, in space and time,
34 because no additional call into an allocator is needed to
35 obtain memory for the hash table node. It also makes it easy
36 to find the hash table node associated with a given object.
37 However, it's difficult to include a given object in an
38 arbitrary number of hash tables.
40 The "hmapx" external hash table implementation allocates hash
41 table nodes separately from the objects in the hash table.
42 Inserting and removing hash table elements requires dynamic
43 allocation, so it is normally slower and takes more memory
44 than the embedded implementation. It also requires searching
45 the table to find the node associated with a given object.
46 However, it's easy to include a given object in an arbitrary
47 number of hash tables. It's also possible to create an
48 external hash table without adding a member to the data
49 structure that the hash table contains. */
51 #ifndef LIBPSPP_HMAP_H
52 #define LIBPSPP_HMAP_H 1
54 /* Embedded hash table with separate chaining.
56 To create an embedded hash table, declare an instance of
57 struct hmap, then initialize it with hmap_init():
61 struct hmap map = HMAP_INITIALIZER (map);
63 Each node in the hash table, presumably a structure type, must
64 include a struct hmap_node member. Here's an example:
67 struct hmap_node node; // hmap_node member.
68 const char *string; // Another member.
70 The hash table functions work with pointers to struct
71 hmap_node. To obtain a pointer to your structure type given a
72 pointer to struct hmap_node, use the HMAP_DATA macro.
74 Inserting and deleting elements is straightforward. Use
75 hmap_insert() to insert an element and hmap_delete() to delete
78 my_foo.string = "My string";
79 hmap_insert (&map, &my_foo.node, hsh_hash_string (my_foo.string));
81 hmap_delete (&map, &my_foo.node);
82 You must pass the element's hash value as one of
83 hmap_insert()'s arguments. The hash table saves this hash
84 value for use later to speed searches and to rehash as the
87 hmap_insert() does not check whether the newly inserted
88 element duplicates an element already in the hash table. The
89 client is responsible for doing so, if this is desirable.
91 The hash table does not provide a direct way to search for an
92 existing element. Instead, it provides the means to iterate
93 over all the elements in the hash table with a given hash
94 value. It is easy to compose a search function from such a
95 building block. For example:
97 find_foo (const struct hmap *map, const char *name)
99 const struct foo *foo;
102 hash = hsh_hash_string (name);
103 HMAP_FOR_EACH_WITH_HASH (foo, struct foo, node, hash, map)
104 if (!strcmp (foo->name, name))
109 Here is how to iterate through the elements currently in the
112 HMAP_FOR_EACH (foo, struct foo, node, &map)
114 ...do something with foo...
120 #include <libpspp/cast.h>
122 /* Returns the data structure corresponding to the given NODE,
123 assuming that NODE is embedded as the given MEMBER name in
124 data type STRUCT. NODE must not be a null pointer. */
125 #define HMAP_DATA(NODE, STRUCT, MEMBER) \
126 (CHECK_POINTER_HAS_TYPE (NODE, struct hmap_node *), \
127 UP_CAST (NODE, STRUCT, MEMBER))
129 /* Like HMAP_DATA, except that a null NODE yields a null pointer
131 #define HMAP_NULLABLE_DATA(NODE, STRUCT, MEMBER) \
132 hmap_nullable_data__ (NODE, offsetof (STRUCT, MEMBER))
134 /* Hash table node. */
137 struct hmap_node *next; /* Next in chain. */
138 size_t hash; /* Hash value. */
141 static inline size_t hmap_node_hash (const struct hmap_node *);
146 size_t count; /* Number of inserted nodes. */
147 size_t mask; /* Number of buckets (power of 2), minus 1. */
148 struct hmap_node **buckets; /* Array of buckets. */
149 struct hmap_node *one; /* One bucket, to eliminate corner cases. */
152 /* Suitable for use as the initializer for a struct hmap named
154 struct hmap map = HMAP_INITIALIZER (map);
155 HMAP_INITIALIZER() is an alternative to hmap_init(). */
156 #define HMAP_INITIALIZER(MAP) { 0, 0, &(MAP).one, NULL }
158 /* Creation and destruction. */
159 void hmap_init (struct hmap *);
160 void hmap_swap (struct hmap *, struct hmap *);
161 void hmap_clear (struct hmap *);
162 void hmap_destroy (struct hmap *);
164 /* Storage management. */
165 void hmap_reserve (struct hmap *, size_t capacity);
166 void hmap_shrink (struct hmap *);
168 /* Search. Refer to the large comment near the top of this file
170 static inline struct hmap_node *hmap_first_with_hash (const struct hmap *,
172 static inline struct hmap_node *hmap_next_with_hash (const struct hmap_node *);
174 /* Insertion and deletion. */
175 static inline void hmap_insert (struct hmap *, struct hmap_node *,
177 static inline void hmap_insert_fast (struct hmap *, struct hmap_node *,
179 static inline void hmap_delete (struct hmap *, struct hmap_node *);
182 static inline struct hmap_node *hmap_first (const struct hmap *);
183 static inline struct hmap_node *hmap_next (const struct hmap *,
184 const struct hmap_node *);
187 static bool hmap_is_empty (const struct hmap *);
188 static inline size_t hmap_count (const struct hmap *);
189 static inline size_t hmap_capacity (const struct hmap *);
191 /* Updating data elements. */
192 void hmap_changed (struct hmap *, struct hmap_node *, size_t new_hash);
193 void hmap_moved (struct hmap *,
194 struct hmap_node *, const struct hmap_node *old);
196 /* Convenience macros for search.
198 These macros automatically use HMAP_DATA to obtain the data
199 elements that encapsulate hmap nodes, which often saves typing
200 and can make code easier to read. Refer to the large comment
201 near the top of this file for an example.
203 These macros evaluate HASH only once. They evaluate their
204 other arguments many times. */
205 #define HMAP_FIRST_WITH_HASH(STRUCT, MEMBER, HMAP, HASH) \
206 HMAP_NULLABLE_DATA (hmap_first_with_hash (HMAP, HASH), STRUCT, MEMBER)
207 #define HMAP_NEXT_WITH_HASH(DATA, STRUCT, MEMBER) \
208 HMAP_NULLABLE_DATA (hmap_next_with_hash (&(DATA)->MEMBER), STRUCT, MEMBER)
209 #define HMAP_FOR_EACH_WITH_HASH(DATA, STRUCT, MEMBER, HASH, HMAP) \
210 for ((DATA) = HMAP_FIRST_WITH_HASH (STRUCT, MEMBER, HMAP, HASH); \
212 (DATA) = HMAP_NEXT_WITH_HASH (DATA, STRUCT, MEMBER))
213 #define HMAP_FOR_EACH_WITH_HASH_SAFE(DATA, NEXT, STRUCT, MEMBER, HASH, HMAP) \
214 for ((DATA) = HMAP_FIRST_WITH_HASH (STRUCT, MEMBER, HMAP, HASH); \
216 ? ((NEXT) = HMAP_NEXT_WITH_HASH (DATA, STRUCT, MEMBER), 1) \
220 /* Convenience macros for iteration.
222 These macros automatically use HMAP_DATA to obtain the data
223 elements that encapsulate hmap nodes, which often saves typing
224 and can make code easier to read. Refer to the large comment
225 near the top of this file for an example.
227 These macros evaluate their arguments many times. */
228 #define HMAP_FIRST(STRUCT, MEMBER, HMAP) \
229 HMAP_NULLABLE_DATA (hmap_first (HMAP), STRUCT, MEMBER)
230 #define HMAP_NEXT(DATA, STRUCT, MEMBER, HMAP) \
231 HMAP_NULLABLE_DATA (hmap_next (HMAP, &(DATA)->MEMBER), STRUCT, MEMBER)
232 #define HMAP_FOR_EACH(DATA, STRUCT, MEMBER, HMAP) \
233 for ((DATA) = HMAP_FIRST (STRUCT, MEMBER, HMAP); \
235 (DATA) = HMAP_NEXT (DATA, STRUCT, MEMBER, HMAP))
236 #define HMAP_FOR_EACH_SAFE(DATA, NEXT, STRUCT, MEMBER, HMAP) \
237 for ((DATA) = HMAP_FIRST (STRUCT, MEMBER, HMAP); \
239 ? ((NEXT) = HMAP_NEXT (DATA, STRUCT, MEMBER, HMAP), 1) \
243 /* Inline definitions. */
245 static inline struct hmap_node *hmap_find_hash__ (struct hmap_node *, size_t);
246 static inline struct hmap_node *hmap_first_nonempty_bucket__ (
247 const struct hmap *, size_t start);
248 static inline size_t hmap_mask_to_capacity__ (size_t mask);
250 /* Returns the hash value associated with NODE. */
252 hmap_node_hash (const struct hmap_node *node)
257 /* Returns the first node in MAP that has hash value HASH, or a
258 null pointer if MAP does not contain any node with that hash
261 Assuming uniform hashing and no duplicate data items in MAP,
262 this function runs in constant time. (Amortized over an
263 iteration over all data items with a given HASH, its runtime
264 is proportional to the length of the hash chain for HASH, so
265 given a pathological hash function, e.g. one that returns a
266 constant value, its runtime degenerates to linear in the
267 length of NODE's hash chain.)
269 Nodes are returned in arbitrary order that may change whenever
270 the hash table's current capacity changes, as reported by
271 hmap_capacity(). Calls to hmap_insert(), hmap_reserve(), and
272 hmap_shrink() can change the capacity of a hash map.
273 Inserting a node with hmap_insert_fast() or deleting one with
274 hmap_delete() will not change the relative ordering of nodes.
276 The HMAP_FOR_EACH_WITH_HASH and HMAP_FOR_EACH_WITH_HASH_SAFE
277 macros provide convenient ways to iterate over all the nodes
278 with a given hash. The HMAP_FIRST_WITH_HASH macro is an
279 interface to this particular function that is often more
281 static inline struct hmap_node *
282 hmap_first_with_hash (const struct hmap *map, size_t hash)
284 return hmap_find_hash__ (map->buckets[hash & map->mask], hash);
287 /* Returns the next node in MAP after NODE that has the same hash
288 value as NODE, or a null pointer if MAP does not contain any
289 more nodes with that hash value.
291 Assuming uniform hashing and no duplicate data items in MAP,
292 this function runs in constant time. (Amortized over an
293 iteration over all data items with a given HASH, its runtime
294 is proportional to the length of the hash chain for HASH, so
295 given a pathological hash function, e.g. one that returns a
296 constant value, its runtime degenerates to linear in the
297 length of NODE's hash chain.)
299 Nodes are returned in arbitrary order that may change whenever
300 the hash table's current capacity changes, as reported by
301 hmap_capacity(). Calls to hmap_insert(), hmap_reserve(), and
302 hmap_shrink() can change the capacity of a hash map.
303 Inserting a node with hmap_insert_fast() or deleting one with
304 hmap_delete() will not change the relative ordering of nodes.
306 The HMAP_FOR_EACH_WITH_HASH and HMAP_FOR_EACH_WITH_HASH_SAFE
307 macros provide convenient ways to iterate over all the nodes
308 with a given hash. The HMAP_NEXT_WITH_HASH macro is an
309 interface to this particular function that is often more
311 static inline struct hmap_node *
312 hmap_next_with_hash (const struct hmap_node *node)
314 return hmap_find_hash__ (node->next, node->hash);
317 /* Inserts NODE into MAP with hash value HASH. If the insertion
318 causes MAP's current capacity, as reported by hmap_capacity(),
319 to be exceeded, rehashes MAP with an increased number of hash
322 This function runs in constant time amortized over all the
325 This function does not verify that MAP does not already
326 contain a data item with the same value as NODE. If
327 duplicates should be disallowed (which is the usual case),
328 then the client must check for duplicates itself before
329 inserting the new node. */
331 hmap_insert (struct hmap *map, struct hmap_node *node, size_t hash)
333 hmap_insert_fast (map, node, hash);
334 if (map->count > hmap_capacity (map))
335 hmap_reserve (map, map->count);
338 /* Inserts NODE into MAP with hash value HASH. Does not check
339 whether this causes MAP's current capacity to be exceeded.
340 The caller must take responsibility for that (or use
341 hmap_insert() instead).
343 This function runs in constant time.
345 This function does not verify that MAP does not already
346 contain a data item with the same value as NODE. If
347 duplicates should be disallowed (which is the usual case),
348 then the client must check for duplicates itself before
349 inserting the new node. */
351 hmap_insert_fast (struct hmap *map, struct hmap_node *node, size_t hash)
353 struct hmap_node **bucket = &map->buckets[hash & map->mask];
355 node->next = *bucket;
360 /* Removes NODE from MAP. The client is responsible for freeing
361 any data associated with NODE, if necessary.
363 Assuming uniform hashing, this function runs in constant time.
364 (Its runtime is proportional to the position of NODE in its
365 hash chain, so given a pathological hash function, e.g. one
366 that returns a constant value, its runtime degenerates to
367 linear in the length of NODE's hash chain.)
369 This function never reduces the number of buckets in MAP.
370 When one deletes a large number of nodes from a hash table,
371 calling hmap_shrink() afterward may therefore save a small
372 amount of memory. It is also more expensive to iterate
373 through a very sparse hash table than a denser one, so
374 shrinking the hash table could also save some time. However,
375 rehashing has an immediate cost that must be weighed against
378 hmap_delete() does not change NODE's hash value reported by
381 hmap_delete (struct hmap *map, struct hmap_node *node)
383 struct hmap_node **bucket = &map->buckets[node->hash & map->mask];
384 while (*bucket != node)
385 bucket = &(*bucket)->next;
386 *bucket = (*bucket)->next;
390 /* Returns the first node in MAP, or a null pointer if MAP is
393 Amortized over iterating through every data element in MAP,
394 this function runs in constant time. However, this assumes
395 that MAP is not excessively sparse, that is, that
396 hmap_capacity(MAP) is at most a constant factor greater than
397 hmap_count(MAP). This will always be true unless many nodes
398 have been inserted into MAP and then most or all of them
399 deleted; in such a case, calling hmap_shrink() is advised.
401 Nodes are returned in arbitrary order that may change whenever
402 the hash table's current capacity changes, as reported by
403 hmap_capacity(). Calls to hmap_insert(), hmap_reserve(), and
404 hmap_shrink() can change the capacity of a hash map.
405 Inserting a node with hmap_insert_fast() or deleting one with
406 hmap_delete() will not change the relative ordering of nodes.
408 The HMAP_FOR_EACH and HMAP_FOR_EACH_SAFE macros provide
409 convenient ways to iterate over all the nodes in a hash map.
410 The HMAP_FIRST macro is an interface to this particular
411 function that is often more convenient. */
412 static inline struct hmap_node *
413 hmap_first (const struct hmap *map)
415 return hmap_first_nonempty_bucket__ (map, 0);
418 /* Returns the next node in MAP following NODE, or a null pointer
419 if NODE is the last node in MAP.
421 Amortized over iterating through every data element in MAP,
422 this function runs in constant time. However, this assumes
423 that MAP is not excessively sparse, that is, that
424 hmap_capacity(MAP) is at most a constant factor greater than
425 hmap_count(MAP). This will always be true unless many nodes
426 have been inserted into MAP and then most or all of them
427 deleted; in such a case, calling hmap_shrink() is advised.
429 Nodes are returned in arbitrary order that may change whenever
430 the hash table's current capacity changes, as reported by
431 hmap_capacity(). Calls to hmap_insert(), hmap_reserve(), and
432 hmap_shrink() can change the capacity of a hash map.
433 Inserting a node with hmap_insert_fast() or deleting one with
434 hmap_delete() will not change the relative ordering of nodes.
436 The HMAP_FOR_EACH and HMAP_FOR_EACH_SAFE macros provide
437 convenient ways to iterate over all the nodes in a hash map.
438 The HMAP_NEXT macro is an interface to this particular
439 function that is often more convenient. */
440 static inline struct hmap_node *
441 hmap_next (const struct hmap *map, const struct hmap_node *node)
443 return (node->next != NULL
445 : hmap_first_nonempty_bucket__ (map, (node->hash & map->mask) + 1));
448 /* Returns true if MAP currently contains no data items, false
451 hmap_is_empty (const struct hmap *map)
453 return map->count == 0;
456 /* Returns the number of data items currently in MAP. */
458 hmap_count (const struct hmap *map)
463 /* Returns the current capacity of MAP, that is, the maximum
464 number of data elements that MAP may hold before it becomes
467 The capacity is advisory only: it is possible to insert any
468 number of data elements into a hash map regardless of its
469 capacity. However, inserting many more elements than the
470 map's capacity will degrade search performance. */
472 hmap_capacity (const struct hmap *map)
474 return hmap_mask_to_capacity__ (map->mask);
477 /* Implementation details. */
479 /* Returns the first node at or after NODE in NODE's chain that
480 has hash value HASH. */
481 static inline struct hmap_node *
482 hmap_find_hash__ (struct hmap_node *node, size_t hash)
484 for (; node != NULL; node = node->next)
485 if (node->hash == hash)
490 /* Returns the first node in the lowest-numbered nonempty bucket
491 in MAP whose index is START or higher, or a null pointer if
492 all such buckets are empty. */
493 static inline struct hmap_node *
494 hmap_first_nonempty_bucket__ (const struct hmap *map, size_t start)
498 for (i = start; i <= map->mask; i++)
499 if (map->buckets[i] != NULL)
500 return map->buckets[i];
504 /* Returns the hash table capacity associated with a given MASK,
505 which should be a value for the "mask" member of struct hmap.
506 MASK must be a power of 2 minus 1 (including 0), that is, its
507 value in binary must be all 1-bits. */
509 hmap_mask_to_capacity__ (size_t mask)
511 return (mask + 1) * 2;
514 /* Helper for HMAP_NULLABLE_DATA (to avoid evaluating its NODE
515 argument more than once). */
517 hmap_nullable_data__ (struct hmap_node *node, size_t member_offset)
519 return node != NULL ? (char *) node - member_offset : NULL;
522 #endif /* libpspp/hmap.h */