/[pcsx2_0.9.7]/trunk/3rdparty/google/sparsehash/sparsehashtable.h
ViewVC logotype

Contents of /trunk/3rdparty/google/sparsehash/sparsehashtable.h

Parent Directory Parent Directory | Revision Log Revision Log


Revision 10 - (show annotations) (download)
Mon Sep 6 11:40:06 2010 UTC (9 years, 5 months ago) by william
File MIME type: text/plain
File size: 41129 byte(s)
exported r3113 from ./upstream/trunk
1 // Copyright (c) 2005, Google Inc.
2 // All rights reserved.
3 //
4 // Redistribution and use in source and binary forms, with or without
5 // modification, are permitted provided that the following conditions are
6 // met:
7 //
8 // * Redistributions of source code must retain the above copyright
9 // notice, this list of conditions and the following disclaimer.
10 // * Redistributions in binary form must reproduce the above
11 // copyright notice, this list of conditions and the following disclaimer
12 // in the documentation and/or other materials provided with the
13 // distribution.
14 // * Neither the name of Google Inc. nor the names of its
15 // contributors may be used to endorse or promote products derived from
16 // this software without specific prior written permission.
17 //
18 // THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
19 // "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
20 // LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
21 // A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
22 // OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
23 // SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
24 // LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
25 // DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
26 // THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
27 // (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
28 // OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
29
30 // ---
31 // Author: Craig Silverstein
32 //
33 // A sparse hashtable is a particular implementation of
34 // a hashtable: one that is meant to minimize memory use.
35 // It does this by using a *sparse table* (cf sparsetable.h),
36 // which uses between 1 and 2 bits to store empty buckets
37 // (we may need another bit for hashtables that support deletion).
38 //
39 // When empty buckets are so cheap, an appealing hashtable
40 // implementation is internal probing, in which the hashtable
41 // is a single table, and collisions are resolved by trying
42 // to insert again in another bucket. The most cache-efficient
43 // internal probing schemes are linear probing (which suffers,
44 // alas, from clumping) and quadratic probing, which is what
45 // we implement by default.
46 //
47 // Deleted buckets are a bit of a pain. We have to somehow mark
48 // deleted buckets (the probing must distinguish them from empty
49 // buckets). The most principled way is to have another bitmap,
50 // but that's annoying and takes up space. Instead we let the
51 // user specify an "impossible" key. We set deleted buckets
52 // to have the impossible key.
53 //
54 // Note it is possible to change the value of the delete key
55 // on the fly; you can even remove it, though after that point
56 // the hashtable is insert_only until you set it again.
57 //
58 // You probably shouldn't use this code directly. Use
59 // <google/sparse_hash_table> or <google/sparse_hash_set> instead.
60 //
61 // You can modify the following, below:
62 // HT_OCCUPANCY_FLT -- how full before we double size
63 // HT_EMPTY_FLT -- how empty before we halve size
64 // HT_MIN_BUCKETS -- smallest bucket size
65 // HT_DEFAULT_STARTING_BUCKETS -- default bucket size at construct-time
66 //
67 // You can also change enlarge_resize_percent (which defaults to
68 // HT_OCCUPANCY_FLT), and shrink_resize_percent (which defaults to
69 // HT_EMPTY_FLT) with set_resizing_parameters().
70 //
71 // How to decide what values to use?
72 // shrink_resize_percent's default of .4 * OCCUPANCY_FLT, is probably good.
73 // HT_MIN_BUCKETS is probably unnecessary since you can specify
74 // (indirectly) the starting number of buckets at construct-time.
75 // For enlarge_resize_percent, you can use this chart to try to trade-off
76 // expected lookup time to the space taken up. By default, this
77 // code uses quadratic probing, though you can change it to linear
78 // via _JUMP below if you really want to.
79 //
80 // From http://www.augustana.ca/~mohrj/courses/1999.fall/csc210/lecture_notes/hashing.html
81 // NUMBER OF PROBES / LOOKUP Successful Unsuccessful
82 // Quadratic collision resolution 1 - ln(1-L) - L/2 1/(1-L) - L - ln(1-L)
83 // Linear collision resolution [1+1/(1-L)]/2 [1+1/(1-L)2]/2
84 //
85 // -- enlarge_resize_percent -- 0.10 0.50 0.60 0.75 0.80 0.90 0.99
86 // QUADRATIC COLLISION RES.
87 // probes/successful lookup 1.05 1.44 1.62 2.01 2.21 2.85 5.11
88 // probes/unsuccessful lookup 1.11 2.19 2.82 4.64 5.81 11.4 103.6
89 // LINEAR COLLISION RES.
90 // probes/successful lookup 1.06 1.5 1.75 2.5 3.0 5.5 50.5
91 // probes/unsuccessful lookup 1.12 2.5 3.6 8.5 13.0 50.0 5000.0
92 //
93 // The value type is required to be copy constructible and default
94 // constructible, but it need not be (and commonly isn't) assignable.
95
96 #ifndef _SPARSEHASHTABLE_H_
97 #define _SPARSEHASHTABLE_H_
98
99 #ifndef SPARSEHASH_STAT_UPDATE
100 #define SPARSEHASH_STAT_UPDATE(x) ((void) 0)
101 #endif
102
103 // The probing method
104 // Linear probing
105 // #define JUMP_(key, num_probes) ( 1 )
106 // Quadratic-ish probing
107 #define JUMP_(key, num_probes) ( num_probes )
108
109
110 // Hashtable class, used to implement the hashed associative containers
111 // hash_set and hash_map.
112
113 #include <google/sparsehash/sparseconfig.h>
114 #include <assert.h>
115 #include <algorithm> // For swap(), eg
116 #include <iterator> // for facts about iterator tags
117 #include <utility> // for pair<>
118 #include <google/sparsetable> // Since that's basically what we are
119
120 _START_GOOGLE_NAMESPACE_
121
122 using STL_NAMESPACE::pair;
123
124 // Alloc is completely ignored. It is present as a template parameter only
125 // for the sake of being compatible with the old SGI hashtable interface.
126 // TODO(csilvers): is that the right thing to do?
127
128 template <class Value, class Key, class HashFcn,
129 class ExtractKey, class EqualKey, class Alloc>
130 class sparse_hashtable;
131
132 template <class V, class K, class HF, class ExK, class EqK, class A>
133 struct sparse_hashtable_iterator;
134
135 template <class V, class K, class HF, class ExK, class EqK, class A>
136 struct sparse_hashtable_const_iterator;
137
138 // As far as iterating, we're basically just a sparsetable
139 // that skips over deleted elements.
140 template <class V, class K, class HF, class ExK, class EqK, class A>
141 struct sparse_hashtable_iterator {
142 public:
143 typedef sparse_hashtable_iterator<V,K,HF,ExK,EqK,A> iterator;
144 typedef sparse_hashtable_const_iterator<V,K,HF,ExK,EqK,A> const_iterator;
145 typedef typename sparsetable<V>::nonempty_iterator st_iterator;
146
147 typedef STL_NAMESPACE::forward_iterator_tag iterator_category;
148 typedef V value_type;
149 typedef ptrdiff_t difference_type;
150 typedef size_t size_type;
151 typedef V& reference; // Value
152 typedef V* pointer;
153
154 // "Real" constructor and default constructor
155 sparse_hashtable_iterator(const sparse_hashtable<V,K,HF,ExK,EqK,A> *h,
156 st_iterator it, st_iterator it_end)
157 : ht(h), pos(it), end(it_end) { advance_past_deleted(); }
158 sparse_hashtable_iterator() { } // not ever used internally
159 // The default destructor is fine; we don't define one
160 // The default operator= is fine; we don't define one
161
162 // Happy dereferencer
163 reference operator*() const { return *pos; }
164 pointer operator->() const { return &(operator*()); }
165
166 // Arithmetic. The only hard part is making sure that
167 // we're not on a marked-deleted array element
168 void advance_past_deleted() {
169 while ( pos != end && ht->test_deleted(*this) )
170 ++pos;
171 }
172 iterator& operator++() {
173 assert(pos != end); ++pos; advance_past_deleted(); return *this;
174 }
175 iterator operator++(int) { iterator tmp(*this); ++*this; return tmp; }
176
177 // Comparison.
178 bool operator==(const iterator& it) const { return pos == it.pos; }
179 bool operator!=(const iterator& it) const { return pos != it.pos; }
180
181
182 // The actual data
183 const sparse_hashtable<V,K,HF,ExK,EqK,A> *ht;
184 st_iterator pos, end;
185 };
186
187 // Now do it all again, but with const-ness!
188 template <class V, class K, class HF, class ExK, class EqK, class A>
189 struct sparse_hashtable_const_iterator {
190 public:
191 typedef sparse_hashtable_iterator<V,K,HF,ExK,EqK,A> iterator;
192 typedef sparse_hashtable_const_iterator<V,K,HF,ExK,EqK,A> const_iterator;
193 typedef typename sparsetable<V>::const_nonempty_iterator st_iterator;
194
195 typedef STL_NAMESPACE::forward_iterator_tag iterator_category;
196 typedef V value_type;
197 typedef ptrdiff_t difference_type;
198 typedef size_t size_type;
199 typedef const V& reference; // Value
200 typedef const V* pointer;
201
202 // "Real" constructor and default constructor
203 sparse_hashtable_const_iterator(const sparse_hashtable<V,K,HF,ExK,EqK,A> *h,
204 st_iterator it, st_iterator it_end)
205 : ht(h), pos(it), end(it_end) { advance_past_deleted(); }
206 // This lets us convert regular iterators to const iterators
207 sparse_hashtable_const_iterator() { } // never used internally
208 sparse_hashtable_const_iterator(const iterator &it)
209 : ht(it.ht), pos(it.pos), end(it.end) { }
210 // The default destructor is fine; we don't define one
211 // The default operator= is fine; we don't define one
212
213 // Happy dereferencer
214 reference operator*() const { return *pos; }
215 pointer operator->() const { return &(operator*()); }
216
217 // Arithmetic. The only hard part is making sure that
218 // we're not on a marked-deleted array element
219 void advance_past_deleted() {
220 while ( pos != end && ht->test_deleted(*this) )
221 ++pos;
222 }
223 const_iterator& operator++() {
224 assert(pos != end); ++pos; advance_past_deleted(); return *this;
225 }
226 const_iterator operator++(int) { const_iterator tmp(*this); ++*this; return tmp; }
227
228 // Comparison.
229 bool operator==(const const_iterator& it) const { return pos == it.pos; }
230 bool operator!=(const const_iterator& it) const { return pos != it.pos; }
231
232
233 // The actual data
234 const sparse_hashtable<V,K,HF,ExK,EqK,A> *ht;
235 st_iterator pos, end;
236 };
237
238 // And once again, but this time freeing up memory as we iterate
239 template <class V, class K, class HF, class ExK, class EqK, class A>
240 struct sparse_hashtable_destructive_iterator {
241 public:
242 typedef sparse_hashtable_destructive_iterator<V,K,HF,ExK,EqK,A> iterator;
243 typedef typename sparsetable<V>::destructive_iterator st_iterator;
244
245 typedef STL_NAMESPACE::forward_iterator_tag iterator_category;
246 typedef V value_type;
247 typedef ptrdiff_t difference_type;
248 typedef size_t size_type;
249 typedef V& reference; // Value
250 typedef V* pointer;
251
252 // "Real" constructor and default constructor
253 sparse_hashtable_destructive_iterator(const
254 sparse_hashtable<V,K,HF,ExK,EqK,A> *h,
255 st_iterator it, st_iterator it_end)
256 : ht(h), pos(it), end(it_end) { advance_past_deleted(); }
257 sparse_hashtable_destructive_iterator() { } // never used internally
258 // The default destructor is fine; we don't define one
259 // The default operator= is fine; we don't define one
260
261 // Happy dereferencer
262 reference operator*() const { return *pos; }
263 pointer operator->() const { return &(operator*()); }
264
265 // Arithmetic. The only hard part is making sure that
266 // we're not on a marked-deleted array element
267 void advance_past_deleted() {
268 while ( pos != end && ht->test_deleted(*this) )
269 ++pos;
270 }
271 iterator& operator++() {
272 assert(pos != end); ++pos; advance_past_deleted(); return *this;
273 }
274 iterator operator++(int) { iterator tmp(*this); ++*this; return tmp; }
275
276 // Comparison.
277 bool operator==(const iterator& it) const { return pos == it.pos; }
278 bool operator!=(const iterator& it) const { return pos != it.pos; }
279
280
281 // The actual data
282 const sparse_hashtable<V,K,HF,ExK,EqK,A> *ht;
283 st_iterator pos, end;
284 };
285
286
287 template <class Value, class Key, class HashFcn,
288 class ExtractKey, class EqualKey, class Alloc>
289 class sparse_hashtable {
290 public:
291 typedef Key key_type;
292 typedef Value value_type;
293 typedef HashFcn hasher;
294 typedef EqualKey key_equal;
295
296 typedef size_t size_type;
297 typedef ptrdiff_t difference_type;
298 typedef value_type* pointer;
299 typedef const value_type* const_pointer;
300 typedef value_type& reference;
301 typedef const value_type& const_reference;
302 typedef sparse_hashtable_iterator<Value, Key, HashFcn,
303 ExtractKey, EqualKey, Alloc>
304 iterator;
305
306 typedef sparse_hashtable_const_iterator<Value, Key, HashFcn,
307 ExtractKey, EqualKey, Alloc>
308 const_iterator;
309
310 typedef sparse_hashtable_destructive_iterator<Value, Key, HashFcn,
311 ExtractKey, EqualKey, Alloc>
312 destructive_iterator;
313
314
315 // How full we let the table get before we resize. Knuth says .8 is
316 // good -- higher causes us to probe too much, though saves memory
317 static const float HT_OCCUPANCY_FLT; // = 0.8f;
318
319 // How empty we let the table get before we resize lower.
320 // It should be less than OCCUPANCY_FLT / 2 or we thrash resizing
321 static const float HT_EMPTY_FLT; // = 0.4 * HT_OCCUPANCY_FLT;
322
323 // Minimum size we're willing to let hashtables be.
324 // Must be a power of two, and at least 4.
325 // Note, however, that for a given hashtable, the minimum size is
326 // determined by the first constructor arg, and may be >HT_MIN_BUCKETS.
327 static const size_t HT_MIN_BUCKETS = 4;
328
329 // By default, if you don't specify a hashtable size at
330 // construction-time, we use this size. Must be a power of two, and
331 // at least HT_MIN_BUCKETS.
332 static const size_t HT_DEFAULT_STARTING_BUCKETS = 32;
333
334 // ITERATOR FUNCTIONS
335 iterator begin() { return iterator(this, table.nonempty_begin(),
336 table.nonempty_end()); }
337 iterator end() { return iterator(this, table.nonempty_end(),
338 table.nonempty_end()); }
339 const_iterator begin() const { return const_iterator(this,
340 table.nonempty_begin(),
341 table.nonempty_end()); }
342 const_iterator end() const { return const_iterator(this,
343 table.nonempty_end(),
344 table.nonempty_end()); }
345
346 // This is used when resizing
347 destructive_iterator destructive_begin() {
348 return destructive_iterator(this, table.destructive_begin(),
349 table.destructive_end());
350 }
351 destructive_iterator destructive_end() {
352 return destructive_iterator(this, table.destructive_end(),
353 table.destructive_end());
354 }
355
356
357 // ACCESSOR FUNCTIONS for the things we templatize on, basically
358 hasher hash_funct() const { return hash; }
359 key_equal key_eq() const { return equals; }
360
361 // We need to copy values when we set the special marker for deleted
362 // elements, but, annoyingly, we can't just use the copy assignment
363 // operator because value_type might not be assignable (it's often
364 // pair<const X, Y>). We use explicit destructor invocation and
365 // placement new to get around this. Arg.
366 private:
367 void set_value(value_type* dst, const value_type src) {
368 dst->~value_type(); // delete the old value, if any
369 new(dst) value_type(src);
370 }
371
372 // This is used as a tag for the copy constructor, saying to destroy its
373 // arg We have two ways of destructively copying: with potentially growing
374 // the hashtable as we copy, and without. To make sure the outside world
375 // can't do a destructive copy, we make the typename private.
376 enum MoveDontCopyT {MoveDontCopy, MoveDontGrow};
377
378
379 // DELETE HELPER FUNCTIONS
380 // This lets the user describe a key that will indicate deleted
381 // table entries. This key should be an "impossible" entry --
382 // if you try to insert it for real, you won't be able to retrieve it!
383 // (NB: while you pass in an entire value, only the key part is looked
384 // at. This is just because I don't know how to assign just a key.)
385 private:
386 void squash_deleted() { // gets rid of any deleted entries we have
387 if ( num_deleted ) { // get rid of deleted before writing
388 sparse_hashtable tmp(MoveDontGrow, *this);
389 swap(tmp); // now we are tmp
390 }
391 assert(num_deleted == 0);
392 }
393
394 public:
395 void set_deleted_key(const value_type &val) {
396 // It's only safe to change what "deleted" means if we purge deleted guys
397 squash_deleted();
398 use_deleted = true;
399 set_value(&delval, val); // save the key (and rest of val too)
400 }
401 void clear_deleted_key() {
402 squash_deleted();
403 use_deleted = false;
404 }
405
406 // These are public so the iterators can use them
407 // True if the item at position bucknum is "deleted" marker
408 bool test_deleted(size_type bucknum) const {
409 // The num_deleted test is crucial for read(): after read(), the ht values
410 // are garbage, and we don't want to think some of them are deleted.
411 return (use_deleted && num_deleted > 0 && table.test(bucknum) &&
412 equals(get_key(delval), get_key(table.get(bucknum))));
413 }
414 bool test_deleted(const iterator &it) const {
415 return (use_deleted && num_deleted > 0 &&
416 equals(get_key(delval), get_key(*it)));
417 }
418 bool test_deleted(const const_iterator &it) const {
419 return (use_deleted && num_deleted > 0 &&
420 equals(get_key(delval), get_key(*it)));
421 }
422 bool test_deleted(const destructive_iterator &it) const {
423 return (use_deleted && num_deleted > 0 &&
424 equals(get_key(delval), get_key(*it)));
425 }
426 // Set it so test_deleted is true. true if object didn't used to be deleted
427 // See below (at erase()) to explain why we allow const_iterators
428 bool set_deleted(const_iterator &it) {
429 assert(use_deleted); // bad if set_deleted_key() wasn't called
430 bool retval = !test_deleted(it);
431 // &* converts from iterator to value-type
432 set_value(const_cast<value_type*>(&(*it)), delval);
433 return retval;
434 }
435 // Set it so test_deleted is false. true if object used to be deleted
436 bool clear_deleted(const_iterator &it) {
437 assert(use_deleted); // bad if set_deleted_key() wasn't called
438 // happens automatically when we assign something else in its place
439 return test_deleted(it);
440 }
441
442
443 // FUNCTIONS CONCERNING SIZE
444 size_type size() const { return table.num_nonempty() - num_deleted; }
445 // Buckets are always a power of 2
446 size_type max_size() const { return (size_type(-1) >> 1U) + 1; }
447 bool empty() const { return size() == 0; }
448 size_type bucket_count() const { return table.size(); }
449 size_type max_bucket_count() const { return max_size(); }
450
451 private:
452 // Because of the above, size_type(-1) is never legal; use it for errors
453 static const size_type ILLEGAL_BUCKET = size_type(-1);
454
455 private:
456 // This is the smallest size a hashtable can be without being too crowded
457 // If you like, you can give a min #buckets as well as a min #elts
458 size_type min_size(size_type num_elts, size_type min_buckets_wanted) {
459 size_type sz = HT_MIN_BUCKETS;
460 while ( sz < min_buckets_wanted || num_elts >= sz * enlarge_resize_percent )
461 sz *= 2;
462 return sz;
463 }
464
465 // Used after a string of deletes
466 void maybe_shrink() {
467 assert(table.num_nonempty() >= num_deleted);
468 assert((bucket_count() & (bucket_count()-1)) == 0); // is a power of two
469 assert(bucket_count() >= HT_MIN_BUCKETS);
470
471 // If you construct a hashtable with < HT_DEFAULT_STARTING_BUCKETS,
472 // we'll never shrink until you get relatively big, and we'll never
473 // shrink below HT_DEFAULT_STARTING_BUCKETS. Otherwise, something
474 // like "dense_hash_set<int> x; x.insert(4); x.erase(4);" will
475 // shrink us down to HT_MIN_BUCKETS buckets, which is too small.
476 if (shrink_threshold > 0
477 && (table.num_nonempty()-num_deleted) < shrink_threshold &&
478 bucket_count() > HT_DEFAULT_STARTING_BUCKETS ) {
479 size_type sz = bucket_count() / 2; // find how much we should shrink
480 while ( sz > HT_DEFAULT_STARTING_BUCKETS &&
481 (table.num_nonempty() - num_deleted) <= sz *
482 shrink_resize_percent )
483 sz /= 2; // stay a power of 2
484 sparse_hashtable tmp(MoveDontCopy, *this, sz);
485 swap(tmp); // now we are tmp
486 }
487 consider_shrink = false; // because we just considered it
488 }
489
490 // We'll let you resize a hashtable -- though this makes us copy all!
491 // When you resize, you say, "make it big enough for this many more elements"
492 void resize_delta(size_type delta) {
493 if ( consider_shrink ) // see if lots of deletes happened
494 maybe_shrink();
495 if ( bucket_count() >= HT_MIN_BUCKETS &&
496 (table.num_nonempty() + delta) <= enlarge_threshold )
497 return; // we're ok as we are
498
499 // Sometimes, we need to resize just to get rid of all the
500 // "deleted" buckets that are clogging up the hashtable. So when
501 // deciding whether to resize, count the deleted buckets (which
502 // are currently taking up room). But later, when we decide what
503 // size to resize to, *don't* count deleted buckets, since they
504 // get discarded during the resize.
505 const size_type needed_size = min_size(table.num_nonempty() + delta, 0);
506 if ( needed_size > bucket_count() ) { // we don't have enough buckets
507 const size_type resize_to = min_size(table.num_nonempty() - num_deleted
508 + delta, 0);
509 sparse_hashtable tmp(MoveDontCopy, *this, resize_to);
510 swap(tmp); // now we are tmp
511 }
512 }
513
514 // Used to actually do the rehashing when we grow/shrink a hashtable
515 void copy_from(const sparse_hashtable &ht, size_type min_buckets_wanted) {
516 clear(); // clear table, set num_deleted to 0
517
518 // If we need to change the size of our table, do it now
519 const size_type resize_to = min_size(ht.size(), min_buckets_wanted);
520 if ( resize_to > bucket_count() ) { // we don't have enough buckets
521 table.resize(resize_to); // sets the number of buckets
522 reset_thresholds();
523 }
524
525 // We use a normal iterator to get non-deleted bcks from ht
526 // We could use insert() here, but since we know there are
527 // no duplicates and no deleted items, we can be more efficient
528 assert( (bucket_count() & (bucket_count()-1)) == 0); // a power of two
529 for ( const_iterator it = ht.begin(); it != ht.end(); ++it ) {
530 size_type num_probes = 0; // how many times we've probed
531 size_type bucknum;
532 const size_type bucket_count_minus_one = bucket_count() - 1;
533 for (bucknum = hash(get_key(*it)) & bucket_count_minus_one;
534 table.test(bucknum); // not empty
535 bucknum = (bucknum + JUMP_(key, num_probes)) & bucket_count_minus_one) {
536 ++num_probes;
537 assert(num_probes < bucket_count()); // or else the hashtable is full
538 }
539 table.set(bucknum, *it); // copies the value to here
540 }
541 }
542
543 // Implementation is like copy_from, but it destroys the table of the
544 // "from" guy by freeing sparsetable memory as we iterate. This is
545 // useful in resizing, since we're throwing away the "from" guy anyway.
546 void move_from(MoveDontCopyT mover, sparse_hashtable &ht,
547 size_type min_buckets_wanted) {
548 clear(); // clear table, set num_deleted to 0
549
550 // If we need to change the size of our table, do it now
551 size_t resize_to;
552 if ( mover == MoveDontGrow )
553 resize_to = ht.bucket_count(); // keep same size as old ht
554 else // MoveDontCopy
555 resize_to = min_size(ht.size(), min_buckets_wanted);
556 if ( resize_to > bucket_count() ) { // we don't have enough buckets
557 table.resize(resize_to); // sets the number of buckets
558 reset_thresholds();
559 }
560
561 // We use a normal iterator to get non-deleted bcks from ht
562 // We could use insert() here, but since we know there are
563 // no duplicates and no deleted items, we can be more efficient
564 assert( (bucket_count() & (bucket_count()-1)) == 0); // a power of two
565 // THIS IS THE MAJOR LINE THAT DIFFERS FROM COPY_FROM():
566 for ( destructive_iterator it = ht.destructive_begin();
567 it != ht.destructive_end(); ++it ) {
568 size_type num_probes = 0; // how many times we've probed
569 size_type bucknum;
570 for ( bucknum = hash(get_key(*it)) & (bucket_count()-1); // h % buck_cnt
571 table.test(bucknum); // not empty
572 bucknum = (bucknum + JUMP_(key, num_probes)) & (bucket_count()-1) ) {
573 ++num_probes;
574 assert(num_probes < bucket_count()); // or else the hashtable is full
575 }
576 table.set(bucknum, *it); // copies the value to here
577 }
578 }
579
580
581 // Required by the spec for hashed associative container
582 public:
583 // Though the docs say this should be num_buckets, I think it's much
584 // more useful as num_elements. As a special feature, calling with
585 // req_elements==0 will cause us to shrink if we can, saving space.
586 void resize(size_type req_elements) { // resize to this or larger
587 if ( consider_shrink || req_elements == 0 )
588 maybe_shrink();
589 if ( req_elements > table.num_nonempty() ) // we only grow
590 resize_delta(req_elements - table.num_nonempty());
591 }
592
593 // Change the value of shrink_resize_percent and
594 // enlarge_resize_percent. The description at the beginning of this
595 // file explains how to choose the values. Setting the shrink
596 // parameter to 0.0 ensures that the table never shrinks.
597 void set_resizing_parameters(float shrink, float grow) {
598 assert(shrink >= 0.0);
599 assert(grow <= 1.0);
600 assert(shrink <= grow/2.0);
601 shrink_resize_percent = shrink;
602 enlarge_resize_percent = grow;
603 reset_thresholds();
604 }
605
606 // CONSTRUCTORS -- as required by the specs, we take a size,
607 // but also let you specify a hashfunction, key comparator,
608 // and key extractor. We also define a copy constructor and =.
609 // DESTRUCTOR -- the default is fine, surprisingly.
610 explicit sparse_hashtable(size_type expected_max_items_in_table = 0,
611 const HashFcn& hf = HashFcn(),
612 const EqualKey& eql = EqualKey(),
613 const ExtractKey& ext = ExtractKey())
614 : hash(hf), equals(eql), get_key(ext), num_deleted(0), use_deleted(false),
615 delval(), enlarge_resize_percent(HT_OCCUPANCY_FLT),
616 shrink_resize_percent(HT_EMPTY_FLT),
617 table(expected_max_items_in_table == 0
618 ? HT_DEFAULT_STARTING_BUCKETS
619 : min_size(expected_max_items_in_table, 0)) {
620 reset_thresholds();
621 }
622
623 // As a convenience for resize(), we allow an optional second argument
624 // which lets you make this new hashtable a different size than ht.
625 // We also provide a mechanism of saying you want to "move" the ht argument
626 // into us instead of copying.
627 sparse_hashtable(const sparse_hashtable& ht,
628 size_type min_buckets_wanted = HT_DEFAULT_STARTING_BUCKETS)
629 : hash(ht.hash), equals(ht.equals), get_key(ht.get_key),
630 num_deleted(0), use_deleted(ht.use_deleted), delval(ht.delval),
631 enlarge_resize_percent(ht.enlarge_resize_percent),
632 shrink_resize_percent(ht.shrink_resize_percent),
633 table() {
634 reset_thresholds();
635 copy_from(ht, min_buckets_wanted); // copy_from() ignores deleted entries
636 }
637 sparse_hashtable(MoveDontCopyT mover, sparse_hashtable& ht,
638 size_type min_buckets_wanted = HT_DEFAULT_STARTING_BUCKETS)
639 : hash(ht.hash), equals(ht.equals), get_key(ht.get_key),
640 num_deleted(0), use_deleted(ht.use_deleted), delval(ht.delval),
641 enlarge_resize_percent(ht.enlarge_resize_percent),
642 shrink_resize_percent(ht.shrink_resize_percent),
643 table() {
644 reset_thresholds();
645 move_from(mover, ht, min_buckets_wanted); // ignores deleted entries
646 }
647
648 sparse_hashtable& operator= (const sparse_hashtable& ht) {
649 if (&ht == this) return *this; // don't copy onto ourselves
650 clear();
651 hash = ht.hash;
652 equals = ht.equals;
653 get_key = ht.get_key;
654 use_deleted = ht.use_deleted;
655 set_value(&delval, ht.delval);
656 copy_from(ht, HT_MIN_BUCKETS); // sets num_deleted to 0 too
657 return *this;
658 }
659
660 // Many STL algorithms use swap instead of copy constructors
661 void swap(sparse_hashtable& ht) {
662 STL_NAMESPACE::swap(hash, ht.hash);
663 STL_NAMESPACE::swap(equals, ht.equals);
664 STL_NAMESPACE::swap(get_key, ht.get_key);
665 STL_NAMESPACE::swap(num_deleted, ht.num_deleted);
666 STL_NAMESPACE::swap(use_deleted, ht.use_deleted);
667 STL_NAMESPACE::swap(enlarge_resize_percent, ht.enlarge_resize_percent);
668 STL_NAMESPACE::swap(shrink_resize_percent, ht.shrink_resize_percent);
669 { value_type tmp; // for annoying reasons, swap() doesn't work
670 set_value(&tmp, delval);
671 set_value(&delval, ht.delval);
672 set_value(&ht.delval, tmp);
673 }
674 table.swap(ht.table);
675 reset_thresholds();
676 ht.reset_thresholds();
677 }
678
679 // It's always nice to be able to clear a table without deallocating it
680 void clear() {
681 table.clear();
682 reset_thresholds();
683 num_deleted = 0;
684 }
685
686
687 // LOOKUP ROUTINES
688 private:
689 // Returns a pair of positions: 1st where the object is, 2nd where
690 // it would go if you wanted to insert it. 1st is ILLEGAL_BUCKET
691 // if object is not found; 2nd is ILLEGAL_BUCKET if it is.
692 // Note: because of deletions where-to-insert is not trivial: it's the
693 // first deleted bucket we see, as long as we don't find the key later
694 pair<size_type, size_type> find_position(const key_type &key) const {
695 size_type num_probes = 0; // how many times we've probed
696 const size_type bucket_count_minus_one = bucket_count() - 1;
697 size_type bucknum = hash(key) & bucket_count_minus_one;
698 size_type insert_pos = ILLEGAL_BUCKET; // where we would insert
699 SPARSEHASH_STAT_UPDATE(total_lookups += 1);
700 while ( 1 ) { // probe until something happens
701 if ( !table.test(bucknum) ) { // bucket is empty
702 SPARSEHASH_STAT_UPDATE(total_probes += num_probes);
703 if ( insert_pos == ILLEGAL_BUCKET ) // found no prior place to insert
704 return pair<size_type,size_type>(ILLEGAL_BUCKET, bucknum);
705 else
706 return pair<size_type,size_type>(ILLEGAL_BUCKET, insert_pos);
707
708 } else if ( test_deleted(bucknum) ) {// keep searching, but mark to insert
709 if ( insert_pos == ILLEGAL_BUCKET )
710 insert_pos = bucknum;
711
712 } else if ( equals(key, get_key(table.get(bucknum))) ) {
713 SPARSEHASH_STAT_UPDATE(total_probes += num_probes);
714 return pair<size_type,size_type>(bucknum, ILLEGAL_BUCKET);
715 }
716 ++num_probes; // we're doing another probe
717 bucknum = (bucknum + JUMP_(key, num_probes)) & bucket_count_minus_one;
718 assert(num_probes < bucket_count()); // don't probe too many times!
719 }
720 }
721
722 public:
723 iterator find(const key_type& key) {
724 if ( size() == 0 ) return end();
725 pair<size_type, size_type> pos = find_position(key);
726 if ( pos.first == ILLEGAL_BUCKET ) // alas, not there
727 return end();
728 else
729 return iterator(this, table.get_iter(pos.first), table.nonempty_end());
730 }
731
732 const_iterator find(const key_type& key) const {
733 if ( size() == 0 ) return end();
734 pair<size_type, size_type> pos = find_position(key);
735 if ( pos.first == ILLEGAL_BUCKET ) // alas, not there
736 return end();
737 else
738 return const_iterator(this,
739 table.get_iter(pos.first), table.nonempty_end());
740 }
741
742 // Counts how many elements have key key. For maps, it's either 0 or 1.
743 size_type count(const key_type &key) const {
744 pair<size_type, size_type> pos = find_position(key);
745 return pos.first == ILLEGAL_BUCKET ? 0 : 1;
746 }
747
748 // Likewise, equal_range doesn't really make sense for us. Oh well.
749 pair<iterator,iterator> equal_range(const key_type& key) {
750 const iterator pos = find(key); // either an iterator or end
751 return pair<iterator,iterator>(pos, pos);
752 }
753 pair<const_iterator,const_iterator> equal_range(const key_type& key) const {
754 const const_iterator pos = find(key); // either an iterator or end
755 return pair<iterator,iterator>(pos, pos);
756 }
757
758
759 // INSERTION ROUTINES
760 private:
761 // If you know *this is big enough to hold obj, use this routine
762 pair<iterator, bool> insert_noresize(const value_type& obj) {
763 // First, double-check we're not inserting delval
764 assert(!use_deleted || !equals(get_key(obj), get_key(delval)));
765 const pair<size_type,size_type> pos = find_position(get_key(obj));
766 if ( pos.first != ILLEGAL_BUCKET) { // object was already there
767 return pair<iterator,bool>(iterator(this, table.get_iter(pos.first),
768 table.nonempty_end()),
769 false); // false: we didn't insert
770 } else { // pos.second says where to put it
771 if ( test_deleted(pos.second) ) { // just replace if it's been del.
772 // The set() below will undelete this object. We just worry about stats
773 assert(num_deleted > 0);
774 --num_deleted; // used to be, now it isn't
775 }
776 table.set(pos.second, obj);
777 return pair<iterator,bool>(iterator(this, table.get_iter(pos.second),
778 table.nonempty_end()),
779 true); // true: we did insert
780 }
781 }
782
783 public:
784 // This is the normal insert routine, used by the outside world
785 pair<iterator, bool> insert(const value_type& obj) {
786 resize_delta(1); // adding an object, grow if need be
787 return insert_noresize(obj);
788 }
789
790 // When inserting a lot at a time, we specialize on the type of iterator
791 template <class InputIterator>
792 void insert(InputIterator f, InputIterator l) {
793 // specializes on iterator type
794 insert(f, l, typename STL_NAMESPACE::iterator_traits<InputIterator>::iterator_category());
795 }
796
797 // Iterator supports operator-, resize before inserting
798 template <class ForwardIterator>
799 void insert(ForwardIterator f, ForwardIterator l,
800 STL_NAMESPACE::forward_iterator_tag) {
801 size_type n = STL_NAMESPACE::distance(f, l); // TODO(csilvers): standard?
802 resize_delta(n);
803 for ( ; n > 0; --n, ++f)
804 insert_noresize(*f);
805 }
806
807 // Arbitrary iterator, can't tell how much to resize
808 template <class InputIterator>
809 void insert(InputIterator f, InputIterator l,
810 STL_NAMESPACE::input_iterator_tag) {
811 for ( ; f != l; ++f)
812 insert(*f);
813 }
814
815
816 // DELETION ROUTINES
817 size_type erase(const key_type& key) {
818 // First, double-check we're not erasing delval
819 assert(!use_deleted || !equals(key, get_key(delval)));
820 const_iterator pos = find(key); // shrug: shouldn't need to be const
821 if ( pos != end() ) {
822 assert(!test_deleted(pos)); // or find() shouldn't have returned it
823 set_deleted(pos);
824 ++num_deleted;
825 consider_shrink = true; // will think about shrink after next insert
826 return 1; // because we deleted one thing
827 } else {
828 return 0; // because we deleted nothing
829 }
830 }
831
832 // This is really evil: really it should be iterator, not const_iterator.
833 // But...the only reason keys are const is to allow lookup.
834 // Since that's a moot issue for deleted keys, we allow const_iterators
835 void erase(const_iterator pos) {
836 if ( pos == end() ) return; // sanity check
837 if ( set_deleted(pos) ) { // true if object has been newly deleted
838 ++num_deleted;
839 consider_shrink = true; // will think about shrink after next insert
840 }
841 }
842
843 void erase(const_iterator f, const_iterator l) {
844 for ( ; f != l; ++f) {
845 if ( set_deleted(f) ) // should always be true
846 ++num_deleted;
847 }
848 consider_shrink = true; // will think about shrink after next insert
849 }
850
851
852 // COMPARISON
853 bool operator==(const sparse_hashtable& ht) const {
854 // We really want to check that the hash functions are the same
855 // but alas there's no way to do this. We just hope.
856 return ( num_deleted == ht.num_deleted && table == ht.table );
857 }
858 bool operator!=(const sparse_hashtable& ht) const {
859 return !(*this == ht);
860 }
861
862
863 // I/O
864 // We support reading and writing hashtables to disk. NOTE that
865 // this only stores the hashtable metadata, not the stuff you've
866 // actually put in the hashtable! Alas, since I don't know how to
867 // write a hasher or key_equal, you have to make sure everything
868 // but the table is the same. We compact before writing.
869 bool write_metadata(FILE *fp) {
870 squash_deleted(); // so we don't have to worry about delkey
871 return table.write_metadata(fp);
872 }
873
874 bool read_metadata(FILE *fp) {
875 num_deleted = 0; // since we got rid before writing
876 bool result = table.read_metadata(fp);
877 reset_thresholds();
878 return result;
879 }
880
881 // Only meaningful if value_type is a POD.
882 bool write_nopointer_data(FILE *fp) {
883 return table.write_nopointer_data(fp);
884 }
885
886 // Only meaningful if value_type is a POD.
887 bool read_nopointer_data(FILE *fp) {
888 return table.read_nopointer_data(fp);
889 }
890
891 private:
892 // The actual data
893 hasher hash; // required by hashed_associative_container
894 key_equal equals;
895 ExtractKey get_key;
896 size_type num_deleted; // how many occupied buckets are marked deleted
897 bool use_deleted; // false until delval has been set
898 value_type delval; // which key marks deleted entries
899 float enlarge_resize_percent; // how full before resize
900 float shrink_resize_percent; // how empty before resize
901 size_type shrink_threshold; // table.size() * shrink_resize_percent
902 size_type enlarge_threshold; // table.size() * enlarge_resize_percent
903 sparsetable<value_type> table; // holds num_buckets and num_elements too
904 bool consider_shrink; // true if we should try to shrink before next insert
905
906 void reset_thresholds() {
907 enlarge_threshold = static_cast<size_type>(table.size()
908 * enlarge_resize_percent);
909 shrink_threshold = static_cast<size_type>(table.size()
910 * shrink_resize_percent);
911 consider_shrink = false; // whatever caused us to reset already considered
912 }
913 };
914
915 // We need a global swap as well
916 template <class V, class K, class HF, class ExK, class EqK, class A>
917 inline void swap(sparse_hashtable<V,K,HF,ExK,EqK,A> &x,
918 sparse_hashtable<V,K,HF,ExK,EqK,A> &y) {
919 x.swap(y);
920 }
921
922 #undef JUMP_
923
924 template <class V, class K, class HF, class ExK, class EqK, class A>
925 const typename sparse_hashtable<V,K,HF,ExK,EqK,A>::size_type
926 sparse_hashtable<V,K,HF,ExK,EqK,A>::ILLEGAL_BUCKET;
927
928 // How full we let the table get before we resize. Knuth says .8 is
929 // good -- higher causes us to probe too much, though saves memory
930 template <class V, class K, class HF, class ExK, class EqK, class A>
931 const float sparse_hashtable<V,K,HF,ExK,EqK,A>::HT_OCCUPANCY_FLT = 0.8f;
932
933 // How empty we let the table get before we resize lower.
934 // It should be less than OCCUPANCY_FLT / 2 or we thrash resizing
935 template <class V, class K, class HF, class ExK, class EqK, class A>
936 const float sparse_hashtable<V,K,HF,ExK,EqK,A>::HT_EMPTY_FLT = 0.4f *
937 sparse_hashtable<V,K,HF,ExK,EqK,A>::HT_OCCUPANCY_FLT;
938
939 _END_GOOGLE_NAMESPACE_
940
941 #endif /* _SPARSEHASHTABLE_H_ */

  ViewVC Help
Powered by ViewVC 1.1.22