diff --git a/CMakeLists.txt b/CMakeLists.txt index e84f5f5..b2d80ee 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -17,8 +17,7 @@ add_executable(studproject ./src/includes/tests.h ./src/includes/aggregate_tests.h ./src/includes/3thparty/emilib/loguru.cpp - main.cpp - + ./src/main.cpp ) target_link_libraries(studproject diff --git a/results.csv b/results.csv index 72e08bd..1b65d12 100644 --- a/results.csv +++ b/results.csv @@ -302,4 +302,52 @@ int_delete, 'absl::node_hash_map, std::__cxx11: int_insert, 'absl::node_hash_map, std::__cxx11::basic_string >', 332, 351, 180, 308, 210, 445, 214, 245, 336, 490, 226, 245, 263, 578 int_succ_lookup, 'absl::node_hash_map, std::__cxx11::basic_string >', 161, 212, 192, 248, 238, 430, 252, 209, 304, 331, 264, 263, 268, 402 int_nosucc_lookup, 'absl::node_hash_map, std::__cxx11::basic_string >', 1630, 385, 182, 393, 208, 425, 209, 210, 369, 464, 234, 251, 284, 606 -int_delete, 'absl::node_hash_map, std::__cxx11::basic_string >', 362, 347, 213, 410, 267, 336, 253, 243, 464, 491, 287, 299, 299, 642 \ No newline at end of file +int_delete, 'absl::node_hash_map, std::__cxx11::basic_string >', 362, 347, 213, 410, 267, 336, 253, 243, 464, 491, 287, 299, 299, 642 +int_insert, 'google::sparse_hash_map', 177, 175, 155, 211, 162, 196, 215, 254, 237, 262, 300, 352, 281, 293 +int_succ_lookup, 'google::sparse_hash_map', 44, 44, 40, 56, 48, 59, 67, 76, 79, 95, 99, 114, 107, 112 +int_nosucc_lookup, 'google::sparse_hash_map', 551, 792, 168, 1570, 178, 195, 232, 279, 239, 360, 311, 378, 281, 294 +int_delete, 'google::sparse_hash_map', 81, 46, 53, 118, 62, 66, 80, 96, 88, 156, 121, 134, 119, 117 +int_insert, 'google::sparse_hash_map', 172, 214, 152, 187, 176, 191, 218, 283, 253, 258, 296, 344, 270, 294 +int_succ_lookup, 'google::sparse_hash_map', 40, 47, 40, 51, 53, 59, 74, 81, 81, 87, 102, 115, 99, 109 +int_nosucc_lookup, 'google::sparse_hash_map', 457, 916, 160, 1487, 185, 200, 244, 279, 236, 272, 298, 387, 279, 291 +int_delete, 'google::sparse_hash_map', 43, 53, 52, 55, 64, 69, 84, 93, 85, 105, 106, 197, 118, 118 +int_insert, 'google::sparse_hash_map, std::__cxx11::basic_string >', 1031, 1306, 1189, 1503, 1499, 1294, 1435, 1587, 1193, 1350, 1492, 1692, 1336, 1384 +int_succ_lookup, 'google::sparse_hash_map, std::__cxx11::basic_string >', 136, 157, 164, 187, 246, 177, 194, 204, 193, 196, 204, 223, 211, 215 +int_nosucc_lookup, 'google::sparse_hash_map, std::__cxx11::basic_string >', 2306, 3944, 1232, 7215, 1555, 1408, 1465, 1600, 1203, 1344, 1504, 1678, 1183, 1266 +int_delete, 'google::sparse_hash_map, std::__cxx11::basic_string >', 110, 137, 156, 158, 176, 180, 197, 218, 190, 199, 215, 249, 212, 219 +int_insert, 'google::sparse_hash_map, std::__cxx11::basic_string >', 1027, 1232, 1102, 1408, 1187, 1344, 1382, 1543, 1146, 1747, 1561, 1652, 1143, 1251 +int_succ_lookup, 'google::sparse_hash_map, std::__cxx11::basic_string >', 121, 161, 165, 188, 168, 175, 188, 204, 187, 301, 242, 234, 201, 218 +int_nosucc_lookup, 'google::sparse_hash_map, std::__cxx11::basic_string >', 2275, 3930, 1178, 7380, 1109, 1241, 1389, 1572, 1162, 1744, 1565, 1705, 1132, 1223 +int_delete, 'google::sparse_hash_map, std::__cxx11::basic_string >', 109, 142, 158, 291, 169, 182, 193, 253, 199, 218, 231, 245, 203, 224 +int_insert, 'google::dense_hash_map', 31, 54, 42, 41, 40, 38, 41, 44, 51, 39, 41, 44, 52, 56 +int_succ_lookup, 'google::dense_hash_map', 24, 19, 20, 20, 19, 21, 20, 22, 23, 22, 23, 23, 26, 26 +int_nosucc_lookup, 'google::dense_hash_map', 47, 42, 41, 46, 44, 39, 42, 47, 53, 41, 45, 52, 58, 68 +int_delete, 'google::dense_hash_map', 22, 38, 22, 17, 16, 13, 14, 15, 19, 14, 15, 17, 26, 26 +int_insert, 'google::dense_hash_map', 27, 79, 37, 35, 39, 36, 37, 42, 49, 40, 44, 44, 49, 64 +int_succ_lookup, 'google::dense_hash_map', 11, 32, 17, 17, 20, 22, 20, 21, 23, 22, 26, 26, 25, 27 +int_nosucc_lookup, 'google::dense_hash_map', 34, 87, 34, 39, 44, 40, 48, 45, 53, 42, 47, 49, 53, 61 +int_delete, 'google::dense_hash_map', 15, 46, 11, 12, 17, 15, 15, 15, 19, 14, 17, 18, 19, 24 +int_insert, 'google::dense_hash_map, std::__cxx11::basic_string >', 152, 154, 145, 158, 171, 156, 163, 169, 189, 175, 178, 191, 192, 197 +int_succ_lookup, 'google::dense_hash_map, std::__cxx11::basic_string >', 91, 98, 93, 108, 113, 105, 107, 110, 116, 116, 118, 121, 123, 125 +int_nosucc_lookup, 'google::dense_hash_map, std::__cxx11::basic_string >', 155, 163, 157, 170, 170, 157, 165, 178, 183, 167, 174, 181, 194, 190 +int_delete, 'google::dense_hash_map, std::__cxx11::basic_string >', 95, 90, 87, 99, 103, 92, 97, 110, 115, 100, 108, 110, 118, 123 +int_insert, 'google::dense_hash_map, std::__cxx11::basic_string >', 141, 144, 146, 172, 175, 158, 164, 176, 196, 171, 175, 182, 190, 195 +int_succ_lookup, 'google::dense_hash_map, std::__cxx11::basic_string >', 79, 95, 104, 106, 111, 102, 110, 111, 118, 111, 114, 123, 123, 125 +int_nosucc_lookup, 'google::dense_hash_map, std::__cxx11::basic_string >', 181, 164, 158, 165, 171, 166, 182, 186, 186, 169, 178, 189, 189, 193 +int_delete, 'google::dense_hash_map, std::__cxx11::basic_string >', 85, 78, 94, 94, 110, 97, 99, 110, 123, 113, 113, 109, 138, 124 +int_insert, 'std::unordered_map', 138, 228, 207, 219, 241, 250, 257, 265, 268, 286, 281, 274, 286, 291 +int_succ_lookup, 'std::unordered_map', 68, 162, 91, 105, 111, 112, 116, 121, 117, 126, 125, 139, 126, 127 +int_nosucc_lookup, 'std::unordered_map', 625, 836, 1392, 1522, 2005, 260, 2765, 281, 271, 4995, 281, 291, 7592, 293 +int_delete, 'std::unordered_map', 162, 183, 237, 265, 278, 320, 305, 311, 302, 329, 354, 336, 393, 335 +int_insert, 'std::unordered_map', 198, 262, 246, 227, 239, 241, 263, 257, 275, 272, 406, 471, 291, 285 +int_succ_lookup, 'std::unordered_map', 88, 94, 112, 104, 107, 111, 118, 120, 127, 121, 340, 249, 129, 130 +int_nosucc_lookup, 'std::unordered_map', 407, 1019, 1505, 1527, 2201, 259, 2727, 268, 258, 4865, 367, 442, 7269, 290 +int_delete, 'std::unordered_map', 223, 246, 402, 250, 347, 281, 300, 306, 305, 325, 384, 441, 342, 341 +int_insert, 'std::unordered_map, std::__cxx11::basic_string >', 248, 372, 335, 352, 351, 355, 372, 377, 414, 422, 449, 429, 440, 435 +int_succ_lookup, 'std::unordered_map, std::__cxx11::basic_string >', 171, 298, 276, 286, 297, 292, 304, 330, 302, 310, 308, 318, 319, 322 +int_nosucc_lookup, 'std::unordered_map, std::__cxx11::basic_string >', 886, 1381, 1780, 2165, 2671, 391, 3570, 434, 429, 6488, 414, 435, 9678, 435 +int_delete, 'std::unordered_map, std::__cxx11::basic_string >', 370, 348, 342, 366, 372, 373, 384, 417, 387, 393, 396, 402, 396, 401 +int_insert, 'std::unordered_map, std::__cxx11::basic_string >', 367, 298, 336, 354, 361, 376, 371, 370, 366, 382, 402, 386, 399, 397 +int_succ_lookup, 'std::unordered_map, std::__cxx11::basic_string >', 299, 242, 270, 282, 299, 292, 303, 417, 328, 316, 317, 316, 331, 362 +int_nosucc_lookup, 'std::unordered_map, std::__cxx11::basic_string >', 757, 1204, 1897, 2121, 2788, 382, 3596, 390, 405, 6269, 409, 412, 9452, 405 +int_delete, 'std::unordered_map, std::__cxx11::basic_string >', 402, 320, 330, 359, 366, 378, 384, 378, 373, 407, 389, 397, 410, 452 \ No newline at end of file diff --git a/src/includes/3thparty/sparsehash/dense_hash_map b/src/includes/3thparty/sparsehash/dense_hash_map new file mode 100644 index 0000000..05fd580 --- /dev/null +++ b/src/includes/3thparty/sparsehash/dense_hash_map @@ -0,0 +1,369 @@ +// Copyright (c) 2005, Google Inc. +// All rights reserved. +// +// Redistribution and use in source and binary forms, with or without +// modification, are permitted provided that the following conditions are +// met: +// +// * Redistributions of source code must retain the above copyright +// notice, this list of conditions and the following disclaimer. +// * Redistributions in binary form must reproduce the above +// copyright notice, this list of conditions and the following disclaimer +// in the documentation and/or other materials provided with the +// distribution. +// * Neither the name of Google Inc. nor the names of its +// contributors may be used to endorse or promote products derived from +// this software without specific prior written permission. +// +// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS +// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT +// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR +// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT +// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, +// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT +// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, +// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY +// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT +// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE +// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + +// ---- +// +// This is just a very thin wrapper over densehashtable.h, just +// like sgi stl's stl_hash_map is a very thin wrapper over +// stl_hashtable. The major thing we define is operator[], because +// we have a concept of a data_type which stl_hashtable doesn't +// (it only has a key and a value). +// +// NOTE: this is exactly like sparse_hash_map.h, with the word +// "sparse" replaced by "dense", except for the addition of +// set_empty_key(). +// +// YOU MUST CALL SET_EMPTY_KEY() IMMEDIATELY AFTER CONSTRUCTION. +// +// Otherwise your program will die in mysterious ways. (Note if you +// use the constructor that takes an InputIterator range, you pass in +// the empty key in the constructor, rather than after. As a result, +// this constructor differs from the standard STL version.) +// +// In other respects, we adhere mostly to the STL semantics for +// hash-map. One important exception is that insert() may invalidate +// iterators entirely -- STL semantics are that insert() may reorder +// iterators, but they all still refer to something valid in the +// hashtable. Not so for us. Likewise, insert() may invalidate +// pointers into the hashtable. (Whether insert invalidates iterators +// and pointers depends on whether it results in a hashtable resize). +// On the plus side, delete() doesn't invalidate iterators or pointers +// at all, or even change the ordering of elements. +// +// Here are a few "power user" tips: +// +// 1) set_deleted_key(): +// If you want to use erase() you *must* call set_deleted_key(), +// in addition to set_empty_key(), after construction. +// The deleted and empty keys must differ. +// +// 2) resize(0): +// When an item is deleted, its memory isn't freed right +// away. This allows you to iterate over a hashtable, +// and call erase(), without invalidating the iterator. +// To force the memory to be freed, call resize(0). +// For tr1 compatibility, this can also be called as rehash(0). +// +// 3) min_load_factor(0.0) +// Setting the minimum load factor to 0.0 guarantees that +// the hash table will never shrink. +// +// Roughly speaking: +// (1) dense_hash_map: fastest, uses the most memory unless entries are small +// (2) sparse_hash_map: slowest, uses the least memory +// (3) hash_map / unordered_map (STL): in the middle +// +// Typically I use sparse_hash_map when I care about space and/or when +// I need to save the hashtable on disk. I use hash_map otherwise. I +// don't personally use dense_hash_set ever; some people use it for +// small sets with lots of lookups. +// +// - dense_hash_map has, typically, about 78% memory overhead (if your +// data takes up X bytes, the hash_map uses .78X more bytes in overhead). +// - sparse_hash_map has about 4 bits overhead per entry. +// - sparse_hash_map can be 3-7 times slower than the others for lookup and, +// especially, inserts. See time_hash_map.cc for details. +// +// See /usr/(local/)?doc/sparsehash-*/dense_hash_map.html +// for information about how to use this class. + +#ifndef _DENSE_HASH_MAP_H_ +#define _DENSE_HASH_MAP_H_ + +#include +#include // needed by stl_alloc +#include // for equal_to<>, select1st<>, etc +#include // for alloc +#include // for pair<> +#include // IWYU pragma: export +#include +#include HASH_FUN_H // for hash<> +_START_GOOGLE_NAMESPACE_ + +template , // defined in sparseconfig.h + class EqualKey = std::equal_to, + class Alloc = libc_allocator_with_realloc > > +class dense_hash_map { + private: + // Apparently select1st is not stl-standard, so we define our own + struct SelectKey { + typedef const Key& result_type; + const Key& operator()(const std::pair& p) const { + return p.first; + } + }; + struct SetKey { + void operator()(std::pair* value, const Key& new_key) const { + *const_cast(&value->first) = new_key; + // It would be nice to clear the rest of value here as well, in + // case it's taking up a lot of memory. We do this by clearing + // the value. This assumes T has a zero-arg constructor! + value->second = T(); + } + }; + // For operator[]. + struct DefaultValue { + std::pair operator()(const Key& key) { + return std::make_pair(key, T()); + } + }; + + // The actual data + typedef dense_hashtable, Key, HashFcn, SelectKey, + SetKey, EqualKey, Alloc> ht; + ht rep; + + public: + typedef typename ht::key_type key_type; + typedef T data_type; + typedef T mapped_type; + typedef typename ht::value_type value_type; + typedef typename ht::hasher hasher; + typedef typename ht::key_equal key_equal; + typedef Alloc allocator_type; + + typedef typename ht::size_type size_type; + typedef typename ht::difference_type difference_type; + typedef typename ht::pointer pointer; + typedef typename ht::const_pointer const_pointer; + typedef typename ht::reference reference; + typedef typename ht::const_reference const_reference; + + typedef typename ht::iterator iterator; + typedef typename ht::const_iterator const_iterator; + typedef typename ht::local_iterator local_iterator; + typedef typename ht::const_local_iterator const_local_iterator; + + // Iterator functions + iterator begin() { return rep.begin(); } + iterator end() { return rep.end(); } + const_iterator begin() const { return rep.begin(); } + const_iterator end() const { return rep.end(); } + + + // These come from tr1's unordered_map. For us, a bucket has 0 or 1 elements. + local_iterator begin(size_type i) { return rep.begin(i); } + local_iterator end(size_type i) { return rep.end(i); } + const_local_iterator begin(size_type i) const { return rep.begin(i); } + const_local_iterator end(size_type i) const { return rep.end(i); } + + // Accessor functions + allocator_type get_allocator() const { return rep.get_allocator(); } + hasher hash_funct() const { return rep.hash_funct(); } + hasher hash_function() const { return hash_funct(); } + key_equal key_eq() const { return rep.key_eq(); } + + + // Constructors + explicit dense_hash_map(size_type expected_max_items_in_table = 0, + const hasher& hf = hasher(), + const key_equal& eql = key_equal(), + const allocator_type& alloc = allocator_type()) + : rep(expected_max_items_in_table, hf, eql, SelectKey(), SetKey(), alloc) { + } + + template + dense_hash_map(InputIterator f, InputIterator l, + const key_type& empty_key_val, + size_type expected_max_items_in_table = 0, + const hasher& hf = hasher(), + const key_equal& eql = key_equal(), + const allocator_type& alloc = allocator_type()) + : rep(expected_max_items_in_table, hf, eql, SelectKey(), SetKey(), alloc) { + set_empty_key(empty_key_val); + rep.insert(f, l); + } + // We use the default copy constructor + // We use the default operator=() + // We use the default destructor + + void clear() { rep.clear(); } + // This clears the hash map without resizing it down to the minimum + // bucket count, but rather keeps the number of buckets constant + void clear_no_resize() { rep.clear_no_resize(); } + void swap(dense_hash_map& hs) { rep.swap(hs.rep); } + + + // Functions concerning size + size_type size() const { return rep.size(); } + size_type max_size() const { return rep.max_size(); } + bool empty() const { return rep.empty(); } + size_type bucket_count() const { return rep.bucket_count(); } + size_type max_bucket_count() const { return rep.max_bucket_count(); } + + // These are tr1 methods. bucket() is the bucket the key is or would be in. + size_type bucket_size(size_type i) const { return rep.bucket_size(i); } + size_type bucket(const key_type& key) const { return rep.bucket(key); } + float load_factor() const { + return size() * 1.0f / bucket_count(); + } + float max_load_factor() const { + float shrink, grow; + rep.get_resizing_parameters(&shrink, &grow); + return grow; + } + void max_load_factor(float new_grow) { + float shrink, grow; + rep.get_resizing_parameters(&shrink, &grow); + rep.set_resizing_parameters(shrink, new_grow); + } + // These aren't tr1 methods but perhaps ought to be. + float min_load_factor() const { + float shrink, grow; + rep.get_resizing_parameters(&shrink, &grow); + return shrink; + } + void min_load_factor(float new_shrink) { + float shrink, grow; + rep.get_resizing_parameters(&shrink, &grow); + rep.set_resizing_parameters(new_shrink, grow); + } + // Deprecated; use min_load_factor() or max_load_factor() instead. + void set_resizing_parameters(float shrink, float grow) { + rep.set_resizing_parameters(shrink, grow); + } + + void resize(size_type hint) { rep.resize(hint); } + void rehash(size_type hint) { resize(hint); } // the tr1 name + + // Lookup routines + iterator find(const key_type& key) { return rep.find(key); } + const_iterator find(const key_type& key) const { return rep.find(key); } + + data_type& operator[](const key_type& key) { // This is our value-add! + // If key is in the hashtable, returns find(key)->second, + // otherwise returns insert(value_type(key, T()).first->second. + // Note it does not create an empty T unless the find fails. + return rep.template find_or_insert(key).second; + } + + size_type count(const key_type& key) const { return rep.count(key); } + + std::pair equal_range(const key_type& key) { + return rep.equal_range(key); + } + std::pair equal_range(const key_type& key) + const { + return rep.equal_range(key); + } + + + // Insertion routines + std::pair insert(const value_type& obj) { + return rep.insert(obj); + } + template void insert(InputIterator f, InputIterator l) { + rep.insert(f, l); + } + void insert(const_iterator f, const_iterator l) { + rep.insert(f, l); + } + // Required for std::insert_iterator; the passed-in iterator is ignored. + iterator insert(iterator, const value_type& obj) { + return insert(obj).first; + } + + // Deletion and empty routines + // THESE ARE NON-STANDARD! I make you specify an "impossible" key + // value to identify deleted and empty buckets. You can change the + // deleted key as time goes on, or get rid of it entirely to be insert-only. + void set_empty_key(const key_type& key) { // YOU MUST CALL THIS! + rep.set_empty_key(value_type(key, data_type())); // rep wants a value + } + key_type empty_key() const { + return rep.empty_key().first; // rep returns a value + } + + void set_deleted_key(const key_type& key) { rep.set_deleted_key(key); } + void clear_deleted_key() { rep.clear_deleted_key(); } + key_type deleted_key() const { return rep.deleted_key(); } + + // These are standard + size_type erase(const key_type& key) { return rep.erase(key); } + void erase(iterator it) { rep.erase(it); } + void erase(iterator f, iterator l) { rep.erase(f, l); } + + + // Comparison + bool operator==(const dense_hash_map& hs) const { return rep == hs.rep; } + bool operator!=(const dense_hash_map& hs) const { return rep != hs.rep; } + + + // I/O -- this is an add-on for writing hash map to disk + // + // For maximum flexibility, this does not assume a particular + // file type (though it will probably be a FILE *). We just pass + // the fp through to rep. + + // If your keys and values are simple enough, you can pass this + // serializer to serialize()/unserialize(). "Simple enough" means + // value_type is a POD type that contains no pointers. Note, + // however, we don't try to normalize endianness. + typedef typename ht::NopointerSerializer NopointerSerializer; + + // serializer: a class providing operator()(OUTPUT*, const value_type&) + // (writing value_type to OUTPUT). You can specify a + // NopointerSerializer object if appropriate (see above). + // fp: either a FILE*, OR an ostream*/subclass_of_ostream*, OR a + // pointer to a class providing size_t Write(const void*, size_t), + // which writes a buffer into a stream (which fp presumably + // owns) and returns the number of bytes successfully written. + // Note basic_ostream is not currently supported. + template + bool serialize(ValueSerializer serializer, OUTPUT* fp) { + return rep.serialize(serializer, fp); + } + + // serializer: a functor providing operator()(INPUT*, value_type*) + // (reading from INPUT and into value_type). You can specify a + // NopointerSerializer object if appropriate (see above). + // fp: either a FILE*, OR an istream*/subclass_of_istream*, OR a + // pointer to a class providing size_t Read(void*, size_t), + // which reads into a buffer from a stream (which fp presumably + // owns) and returns the number of bytes successfully read. + // Note basic_istream is not currently supported. + // NOTE: Since value_type is std::pair, ValueSerializer + // may need to do a const cast in order to fill in the key. + template + bool unserialize(ValueSerializer serializer, INPUT* fp) { + return rep.unserialize(serializer, fp); + } +}; + +// We need a global swap as well +template +inline void swap(dense_hash_map& hm1, + dense_hash_map& hm2) { + hm1.swap(hm2); +} + +_END_GOOGLE_NAMESPACE_ + +#endif /* _DENSE_HASH_MAP_H_ */ diff --git a/src/includes/3thparty/sparsehash/dense_hash_set b/src/includes/3thparty/sparsehash/dense_hash_set new file mode 100644 index 0000000..050b15d --- /dev/null +++ b/src/includes/3thparty/sparsehash/dense_hash_set @@ -0,0 +1,338 @@ +// Copyright (c) 2005, Google Inc. +// All rights reserved. +// +// Redistribution and use in source and binary forms, with or without +// modification, are permitted provided that the following conditions are +// met: +// +// * Redistributions of source code must retain the above copyright +// notice, this list of conditions and the following disclaimer. +// * Redistributions in binary form must reproduce the above +// copyright notice, this list of conditions and the following disclaimer +// in the documentation and/or other materials provided with the +// distribution. +// * Neither the name of Google Inc. nor the names of its +// contributors may be used to endorse or promote products derived from +// this software without specific prior written permission. +// +// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS +// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT +// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR +// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT +// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, +// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT +// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, +// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY +// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT +// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE +// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + +// --- +// +// This is just a very thin wrapper over densehashtable.h, just +// like sgi stl's stl_hash_set is a very thin wrapper over +// stl_hashtable. The major thing we define is operator[], because +// we have a concept of a data_type which stl_hashtable doesn't +// (it only has a key and a value). +// +// This is more different from dense_hash_map than you might think, +// because all iterators for sets are const (you obviously can't +// change the key, and for sets there is no value). +// +// NOTE: this is exactly like sparse_hash_set.h, with the word +// "sparse" replaced by "dense", except for the addition of +// set_empty_key(). +// +// YOU MUST CALL SET_EMPTY_KEY() IMMEDIATELY AFTER CONSTRUCTION. +// +// Otherwise your program will die in mysterious ways. (Note if you +// use the constructor that takes an InputIterator range, you pass in +// the empty key in the constructor, rather than after. As a result, +// this constructor differs from the standard STL version.) +// +// In other respects, we adhere mostly to the STL semantics for +// hash-map. One important exception is that insert() may invalidate +// iterators entirely -- STL semantics are that insert() may reorder +// iterators, but they all still refer to something valid in the +// hashtable. Not so for us. Likewise, insert() may invalidate +// pointers into the hashtable. (Whether insert invalidates iterators +// and pointers depends on whether it results in a hashtable resize). +// On the plus side, delete() doesn't invalidate iterators or pointers +// at all, or even change the ordering of elements. +// +// Here are a few "power user" tips: +// +// 1) set_deleted_key(): +// If you want to use erase() you must call set_deleted_key(), +// in addition to set_empty_key(), after construction. +// The deleted and empty keys must differ. +// +// 2) resize(0): +// When an item is deleted, its memory isn't freed right +// away. This allows you to iterate over a hashtable, +// and call erase(), without invalidating the iterator. +// To force the memory to be freed, call resize(0). +// For tr1 compatibility, this can also be called as rehash(0). +// +// 3) min_load_factor(0.0) +// Setting the minimum load factor to 0.0 guarantees that +// the hash table will never shrink. +// +// Roughly speaking: +// (1) dense_hash_set: fastest, uses the most memory unless entries are small +// (2) sparse_hash_set: slowest, uses the least memory +// (3) hash_set / unordered_set (STL): in the middle +// +// Typically I use sparse_hash_set when I care about space and/or when +// I need to save the hashtable on disk. I use hash_set otherwise. I +// don't personally use dense_hash_set ever; some people use it for +// small sets with lots of lookups. +// +// - dense_hash_set has, typically, about 78% memory overhead (if your +// data takes up X bytes, the hash_set uses .78X more bytes in overhead). +// - sparse_hash_set has about 4 bits overhead per entry. +// - sparse_hash_set can be 3-7 times slower than the others for lookup and, +// especially, inserts. See time_hash_map.cc for details. +// +// See /usr/(local/)?doc/sparsehash-*/dense_hash_set.html +// for information about how to use this class. + +#ifndef _DENSE_HASH_SET_H_ +#define _DENSE_HASH_SET_H_ + +#include +#include // needed by stl_alloc +#include // for equal_to<>, select1st<>, etc +#include // for alloc +#include // for pair<> +#include // IWYU pragma: export +#include +#include HASH_FUN_H // for hash<> +_START_GOOGLE_NAMESPACE_ + +template , // defined in sparseconfig.h + class EqualKey = std::equal_to, + class Alloc = libc_allocator_with_realloc > +class dense_hash_set { + private: + // Apparently identity is not stl-standard, so we define our own + struct Identity { + typedef const Value& result_type; + const Value& operator()(const Value& v) const { return v; } + }; + struct SetKey { + void operator()(Value* value, const Value& new_key) const { + *value = new_key; + } + }; + + // The actual data + typedef dense_hashtable ht; + ht rep; + + public: + typedef typename ht::key_type key_type; + typedef typename ht::value_type value_type; + typedef typename ht::hasher hasher; + typedef typename ht::key_equal key_equal; + typedef Alloc allocator_type; + + typedef typename ht::size_type size_type; + typedef typename ht::difference_type difference_type; + typedef typename ht::const_pointer pointer; + typedef typename ht::const_pointer const_pointer; + typedef typename ht::const_reference reference; + typedef typename ht::const_reference const_reference; + + typedef typename ht::const_iterator iterator; + typedef typename ht::const_iterator const_iterator; + typedef typename ht::const_local_iterator local_iterator; + typedef typename ht::const_local_iterator const_local_iterator; + + + // Iterator functions -- recall all iterators are const + iterator begin() const { return rep.begin(); } + iterator end() const { return rep.end(); } + + // These come from tr1's unordered_set. For us, a bucket has 0 or 1 elements. + local_iterator begin(size_type i) const { return rep.begin(i); } + local_iterator end(size_type i) const { return rep.end(i); } + + + // Accessor functions + allocator_type get_allocator() const { return rep.get_allocator(); } + hasher hash_funct() const { return rep.hash_funct(); } + hasher hash_function() const { return hash_funct(); } // tr1 name + key_equal key_eq() const { return rep.key_eq(); } + + + // Constructors + explicit dense_hash_set(size_type expected_max_items_in_table = 0, + const hasher& hf = hasher(), + const key_equal& eql = key_equal(), + const allocator_type& alloc = allocator_type()) + : rep(expected_max_items_in_table, hf, eql, Identity(), SetKey(), alloc) { + } + + template + dense_hash_set(InputIterator f, InputIterator l, + const key_type& empty_key_val, + size_type expected_max_items_in_table = 0, + const hasher& hf = hasher(), + const key_equal& eql = key_equal(), + const allocator_type& alloc = allocator_type()) + : rep(expected_max_items_in_table, hf, eql, Identity(), SetKey(), alloc) { + set_empty_key(empty_key_val); + rep.insert(f, l); + } + // We use the default copy constructor + // We use the default operator=() + // We use the default destructor + + void clear() { rep.clear(); } + // This clears the hash set without resizing it down to the minimum + // bucket count, but rather keeps the number of buckets constant + void clear_no_resize() { rep.clear_no_resize(); } + void swap(dense_hash_set& hs) { rep.swap(hs.rep); } + + + // Functions concerning size + size_type size() const { return rep.size(); } + size_type max_size() const { return rep.max_size(); } + bool empty() const { return rep.empty(); } + size_type bucket_count() const { return rep.bucket_count(); } + size_type max_bucket_count() const { return rep.max_bucket_count(); } + + // These are tr1 methods. bucket() is the bucket the key is or would be in. + size_type bucket_size(size_type i) const { return rep.bucket_size(i); } + size_type bucket(const key_type& key) const { return rep.bucket(key); } + float load_factor() const { + return size() * 1.0f / bucket_count(); + } + float max_load_factor() const { + float shrink, grow; + rep.get_resizing_parameters(&shrink, &grow); + return grow; + } + void max_load_factor(float new_grow) { + float shrink, grow; + rep.get_resizing_parameters(&shrink, &grow); + rep.set_resizing_parameters(shrink, new_grow); + } + // These aren't tr1 methods but perhaps ought to be. + float min_load_factor() const { + float shrink, grow; + rep.get_resizing_parameters(&shrink, &grow); + return shrink; + } + void min_load_factor(float new_shrink) { + float shrink, grow; + rep.get_resizing_parameters(&shrink, &grow); + rep.set_resizing_parameters(new_shrink, grow); + } + // Deprecated; use min_load_factor() or max_load_factor() instead. + void set_resizing_parameters(float shrink, float grow) { + rep.set_resizing_parameters(shrink, grow); + } + + void resize(size_type hint) { rep.resize(hint); } + void rehash(size_type hint) { resize(hint); } // the tr1 name + + // Lookup routines + iterator find(const key_type& key) const { return rep.find(key); } + + size_type count(const key_type& key) const { return rep.count(key); } + + std::pair equal_range(const key_type& key) const { + return rep.equal_range(key); + } + + + // Insertion routines + std::pair insert(const value_type& obj) { + std::pair p = rep.insert(obj); + return std::pair(p.first, p.second); // const to non-const + } + template void insert(InputIterator f, InputIterator l) { + rep.insert(f, l); + } + void insert(const_iterator f, const_iterator l) { + rep.insert(f, l); + } + // Required for std::insert_iterator; the passed-in iterator is ignored. + iterator insert(iterator, const value_type& obj) { + return insert(obj).first; + } + + // Deletion and empty routines + // THESE ARE NON-STANDARD! I make you specify an "impossible" key + // value to identify deleted and empty buckets. You can change the + // deleted key as time goes on, or get rid of it entirely to be insert-only. + void set_empty_key(const key_type& key) { rep.set_empty_key(key); } + key_type empty_key() const { return rep.empty_key(); } + + void set_deleted_key(const key_type& key) { rep.set_deleted_key(key); } + void clear_deleted_key() { rep.clear_deleted_key(); } + key_type deleted_key() const { return rep.deleted_key(); } + + // These are standard + size_type erase(const key_type& key) { return rep.erase(key); } + void erase(iterator it) { rep.erase(it); } + void erase(iterator f, iterator l) { rep.erase(f, l); } + + + // Comparison + bool operator==(const dense_hash_set& hs) const { return rep == hs.rep; } + bool operator!=(const dense_hash_set& hs) const { return rep != hs.rep; } + + + // I/O -- this is an add-on for writing metainformation to disk + // + // For maximum flexibility, this does not assume a particular + // file type (though it will probably be a FILE *). We just pass + // the fp through to rep. + + // If your keys and values are simple enough, you can pass this + // serializer to serialize()/unserialize(). "Simple enough" means + // value_type is a POD type that contains no pointers. Note, + // however, we don't try to normalize endianness. + typedef typename ht::NopointerSerializer NopointerSerializer; + + // serializer: a class providing operator()(OUTPUT*, const value_type&) + // (writing value_type to OUTPUT). You can specify a + // NopointerSerializer object if appropriate (see above). + // fp: either a FILE*, OR an ostream*/subclass_of_ostream*, OR a + // pointer to a class providing size_t Write(const void*, size_t), + // which writes a buffer into a stream (which fp presumably + // owns) and returns the number of bytes successfully written. + // Note basic_ostream is not currently supported. + template + bool serialize(ValueSerializer serializer, OUTPUT* fp) { + return rep.serialize(serializer, fp); + } + + // serializer: a functor providing operator()(INPUT*, value_type*) + // (reading from INPUT and into value_type). You can specify a + // NopointerSerializer object if appropriate (see above). + // fp: either a FILE*, OR an istream*/subclass_of_istream*, OR a + // pointer to a class providing size_t Read(void*, size_t), + // which reads into a buffer from a stream (which fp presumably + // owns) and returns the number of bytes successfully read. + // Note basic_istream is not currently supported. + template + bool unserialize(ValueSerializer serializer, INPUT* fp) { + return rep.unserialize(serializer, fp); + } +}; + +template +inline void swap(dense_hash_set& hs1, + dense_hash_set& hs2) { + hs1.swap(hs2); +} + +_END_GOOGLE_NAMESPACE_ + +#endif /* _DENSE_HASH_SET_H_ */ diff --git a/src/includes/3thparty/sparsehash/internal/densehashtable.h b/src/includes/3thparty/sparsehash/internal/densehashtable.h new file mode 100644 index 0000000..cdf4ff6 --- /dev/null +++ b/src/includes/3thparty/sparsehash/internal/densehashtable.h @@ -0,0 +1,1327 @@ +// Copyright (c) 2005, Google Inc. +// All rights reserved. +// +// Redistribution and use in source and binary forms, with or without +// modification, are permitted provided that the following conditions are +// met: +// +// * Redistributions of source code must retain the above copyright +// notice, this list of conditions and the following disclaimer. +// * Redistributions in binary form must reproduce the above +// copyright notice, this list of conditions and the following disclaimer +// in the documentation and/or other materials provided with the +// distribution. +// * Neither the name of Google Inc. nor the names of its +// contributors may be used to endorse or promote products derived from +// this software without specific prior written permission. +// +// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS +// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT +// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR +// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT +// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, +// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT +// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, +// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY +// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT +// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE +// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + +// --- +// +// A dense hashtable is a particular implementation of +// a hashtable: one that is meant to minimize memory allocation. +// It does this by using an array to store all the data. We +// steal a value from the key space to indicate "empty" array +// elements (ie indices where no item lives) and another to indicate +// "deleted" elements. +// +// (Note it is possible to change the value of the delete key +// on the fly; you can even remove it, though after that point +// the hashtable is insert_only until you set it again. The empty +// value however can't be changed.) +// +// To minimize allocation and pointer overhead, we use internal +// probing, in which the hashtable is a single table, and collisions +// are resolved by trying to insert again in another bucket. The +// most cache-efficient internal probing schemes are linear probing +// (which suffers, alas, from clumping) and quadratic probing, which +// is what we implement by default. +// +// Type requirements: value_type is required to be Copy Constructible +// and Default Constructible. It is not required to be (and commonly +// isn't) Assignable. +// +// You probably shouldn't use this code directly. Use dense_hash_map<> +// or dense_hash_set<> instead. + +// You can change the following below: +// HT_OCCUPANCY_PCT -- how full before we double size +// HT_EMPTY_PCT -- how empty before we halve size +// HT_MIN_BUCKETS -- default smallest bucket size +// +// You can also change enlarge_factor (which defaults to +// HT_OCCUPANCY_PCT), and shrink_factor (which defaults to +// HT_EMPTY_PCT) with set_resizing_parameters(). +// +// How to decide what values to use? +// shrink_factor's default of .4 * OCCUPANCY_PCT, is probably good. +// HT_MIN_BUCKETS is probably unnecessary since you can specify +// (indirectly) the starting number of buckets at construct-time. +// For enlarge_factor, you can use this chart to try to trade-off +// expected lookup time to the space taken up. By default, this +// code uses quadratic probing, though you can change it to linear +// via JUMP_ below if you really want to. +// +// From http://www.augustana.ca/~mohrj/courses/1999.fall/csc210/lecture_notes/hashing.html +// NUMBER OF PROBES / LOOKUP Successful Unsuccessful +// Quadratic collision resolution 1 - ln(1-L) - L/2 1/(1-L) - L - ln(1-L) +// Linear collision resolution [1+1/(1-L)]/2 [1+1/(1-L)2]/2 +// +// -- enlarge_factor -- 0.10 0.50 0.60 0.75 0.80 0.90 0.99 +// QUADRATIC COLLISION RES. +// probes/successful lookup 1.05 1.44 1.62 2.01 2.21 2.85 5.11 +// probes/unsuccessful lookup 1.11 2.19 2.82 4.64 5.81 11.4 103.6 +// LINEAR COLLISION RES. +// probes/successful lookup 1.06 1.5 1.75 2.5 3.0 5.5 50.5 +// probes/unsuccessful lookup 1.12 2.5 3.6 8.5 13.0 50.0 5000.0 + +#ifndef _DENSEHASHTABLE_H_ +#define _DENSEHASHTABLE_H_ + +#include +#include +#include // for FILE, fwrite, fread +#include // For swap(), eg +#include // For iterator tags +#include // for numeric_limits +#include // For uninitialized_fill +#include // for pair +#include +#include +#include +#include // For length_error + +_START_GOOGLE_NAMESPACE_ + +namespace base { // just to make google->opensource transition easier +using GOOGLE_NAMESPACE::true_type; +using GOOGLE_NAMESPACE::false_type; +using GOOGLE_NAMESPACE::integral_constant; +using GOOGLE_NAMESPACE::is_same; +using GOOGLE_NAMESPACE::remove_const; +} + +// The probing method +// Linear probing +// #define JUMP_(key, num_probes) ( 1 ) +// Quadratic probing +#define JUMP_(key, num_probes) ( num_probes ) + +// Hashtable class, used to implement the hashed associative containers +// hash_set and hash_map. + +// Value: what is stored in the table (each bucket is a Value). +// Key: something in a 1-to-1 correspondence to a Value, that can be used +// to search for a Value in the table (find() takes a Key). +// HashFcn: Takes a Key and returns an integer, the more unique the better. +// ExtractKey: given a Value, returns the unique Key associated with it. +// Must inherit from unary_function, or at least have a +// result_type enum indicating the return type of operator(). +// SetKey: given a Value* and a Key, modifies the value such that +// ExtractKey(value) == key. We guarantee this is only called +// with key == deleted_key or key == empty_key. +// EqualKey: Given two Keys, says whether they are the same (that is, +// if they are both associated with the same Value). +// Alloc: STL allocator to use to allocate memory. + +template +class dense_hashtable; + +template +struct dense_hashtable_iterator; + +template +struct dense_hashtable_const_iterator; + +// We're just an array, but we need to skip over empty and deleted elements +template +struct dense_hashtable_iterator { + private: + typedef typename A::template rebind::other value_alloc_type; + + public: + typedef dense_hashtable_iterator iterator; + typedef dense_hashtable_const_iterator const_iterator; + + typedef std::forward_iterator_tag iterator_category; // very little defined! + typedef V value_type; + typedef typename value_alloc_type::difference_type difference_type; + typedef typename value_alloc_type::size_type size_type; + typedef typename value_alloc_type::reference reference; + typedef typename value_alloc_type::pointer pointer; + + // "Real" constructor and default constructor + dense_hashtable_iterator(const dense_hashtable *h, + pointer it, pointer it_end, bool advance) + : ht(h), pos(it), end(it_end) { + if (advance) advance_past_empty_and_deleted(); + } + dense_hashtable_iterator() { } + // The default destructor is fine; we don't define one + // The default operator= is fine; we don't define one + + // Happy dereferencer + reference operator*() const { return *pos; } + pointer operator->() const { return &(operator*()); } + + // Arithmetic. The only hard part is making sure that + // we're not on an empty or marked-deleted array element + void advance_past_empty_and_deleted() { + while ( pos != end && (ht->test_empty(*this) || ht->test_deleted(*this)) ) + ++pos; + } + iterator& operator++() { + assert(pos != end); ++pos; advance_past_empty_and_deleted(); return *this; + } + iterator operator++(int) { iterator tmp(*this); ++*this; return tmp; } + + // Comparison. + bool operator==(const iterator& it) const { return pos == it.pos; } + bool operator!=(const iterator& it) const { return pos != it.pos; } + + + // The actual data + const dense_hashtable *ht; + pointer pos, end; +}; + + +// Now do it all again, but with const-ness! +template +struct dense_hashtable_const_iterator { + private: + typedef typename A::template rebind::other value_alloc_type; + + public: + typedef dense_hashtable_iterator iterator; + typedef dense_hashtable_const_iterator const_iterator; + + typedef std::forward_iterator_tag iterator_category; // very little defined! + typedef V value_type; + typedef typename value_alloc_type::difference_type difference_type; + typedef typename value_alloc_type::size_type size_type; + typedef typename value_alloc_type::const_reference reference; + typedef typename value_alloc_type::const_pointer pointer; + + // "Real" constructor and default constructor + dense_hashtable_const_iterator( + const dense_hashtable *h, + pointer it, pointer it_end, bool advance) + : ht(h), pos(it), end(it_end) { + if (advance) advance_past_empty_and_deleted(); + } + dense_hashtable_const_iterator() + : ht(NULL), pos(pointer()), end(pointer()) { } + // This lets us convert regular iterators to const iterators + dense_hashtable_const_iterator(const iterator &it) + : ht(it.ht), pos(it.pos), end(it.end) { } + // The default destructor is fine; we don't define one + // The default operator= is fine; we don't define one + + // Happy dereferencer + reference operator*() const { return *pos; } + pointer operator->() const { return &(operator*()); } + + // Arithmetic. The only hard part is making sure that + // we're not on an empty or marked-deleted array element + void advance_past_empty_and_deleted() { + while ( pos != end && (ht->test_empty(*this) || ht->test_deleted(*this)) ) + ++pos; + } + const_iterator& operator++() { + assert(pos != end); ++pos; advance_past_empty_and_deleted(); return *this; + } + const_iterator operator++(int) { const_iterator tmp(*this); ++*this; return tmp; } + + // Comparison. + bool operator==(const const_iterator& it) const { return pos == it.pos; } + bool operator!=(const const_iterator& it) const { return pos != it.pos; } + + + // The actual data + const dense_hashtable *ht; + pointer pos, end; +}; + +template +class dense_hashtable { + private: + typedef typename Alloc::template rebind::other value_alloc_type; + + public: + typedef Key key_type; + typedef Value value_type; + typedef HashFcn hasher; + typedef EqualKey key_equal; + typedef Alloc allocator_type; + + typedef typename value_alloc_type::size_type size_type; + typedef typename value_alloc_type::difference_type difference_type; + typedef typename value_alloc_type::reference reference; + typedef typename value_alloc_type::const_reference const_reference; + typedef typename value_alloc_type::pointer pointer; + typedef typename value_alloc_type::const_pointer const_pointer; + typedef dense_hashtable_iterator + iterator; + + typedef dense_hashtable_const_iterator + const_iterator; + + // These come from tr1. For us they're the same as regular iterators. + typedef iterator local_iterator; + typedef const_iterator const_local_iterator; + + // How full we let the table get before we resize, by default. + // Knuth says .8 is good -- higher causes us to probe too much, + // though it saves memory. + static const int HT_OCCUPANCY_PCT; // defined at the bottom of this file + + // How empty we let the table get before we resize lower, by default. + // (0.0 means never resize lower.) + // It should be less than OCCUPANCY_PCT / 2 or we thrash resizing + static const int HT_EMPTY_PCT; // defined at the bottom of this file + + // Minimum size we're willing to let hashtables be. + // Must be a power of two, and at least 4. + // Note, however, that for a given hashtable, the initial size is a + // function of the first constructor arg, and may be >HT_MIN_BUCKETS. + static const size_type HT_MIN_BUCKETS = 4; + + // By default, if you don't specify a hashtable size at + // construction-time, we use this size. Must be a power of two, and + // at least HT_MIN_BUCKETS. + static const size_type HT_DEFAULT_STARTING_BUCKETS = 32; + + // ITERATOR FUNCTIONS + iterator begin() { return iterator(this, table, + table + num_buckets, true); } + iterator end() { return iterator(this, table + num_buckets, + table + num_buckets, true); } + const_iterator begin() const { return const_iterator(this, table, + table+num_buckets,true);} + const_iterator end() const { return const_iterator(this, table + num_buckets, + table+num_buckets,true);} + + // These come from tr1 unordered_map. They iterate over 'bucket' n. + // We'll just consider bucket n to be the n-th element of the table. + local_iterator begin(size_type i) { + return local_iterator(this, table + i, table + i+1, false); + } + local_iterator end(size_type i) { + local_iterator it = begin(i); + if (!test_empty(i) && !test_deleted(i)) + ++it; + return it; + } + const_local_iterator begin(size_type i) const { + return const_local_iterator(this, table + i, table + i+1, false); + } + const_local_iterator end(size_type i) const { + const_local_iterator it = begin(i); + if (!test_empty(i) && !test_deleted(i)) + ++it; + return it; + } + + // ACCESSOR FUNCTIONS for the things we templatize on, basically + hasher hash_funct() const { return settings; } + key_equal key_eq() const { return key_info; } + allocator_type get_allocator() const { + return allocator_type(val_info); + } + + // Accessor function for statistics gathering. + int num_table_copies() const { return settings.num_ht_copies(); } + + private: + // Annoyingly, we can't copy values around, because they might have + // const components (they're probably pair). We use + // explicit destructor invocation and placement new to get around + // this. Arg. + void set_value(pointer dst, const_reference src) { + dst->~value_type(); // delete the old value, if any + new(dst) value_type(src); + } + + void destroy_buckets(size_type first, size_type last) { + for ( ; first != last; ++first) + table[first].~value_type(); + } + + // DELETE HELPER FUNCTIONS + // This lets the user describe a key that will indicate deleted + // table entries. This key should be an "impossible" entry -- + // if you try to insert it for real, you won't be able to retrieve it! + // (NB: while you pass in an entire value, only the key part is looked + // at. This is just because I don't know how to assign just a key.) + private: + void squash_deleted() { // gets rid of any deleted entries we have + if ( num_deleted ) { // get rid of deleted before writing + dense_hashtable tmp(*this); // copying will get rid of deleted + swap(tmp); // now we are tmp + } + assert(num_deleted == 0); + } + + // Test if the given key is the deleted indicator. Requires + // num_deleted > 0, for correctness of read(), and because that + // guarantees that key_info.delkey is valid. + bool test_deleted_key(const key_type& key) const { + assert(num_deleted > 0); + return equals(key_info.delkey, key); + } + + public: + void set_deleted_key(const key_type &key) { + // the empty indicator (if specified) and the deleted indicator + // must be different + assert((!settings.use_empty() || !equals(key, get_key(val_info.emptyval))) + && "Passed the empty-key to set_deleted_key"); + // It's only safe to change what "deleted" means if we purge deleted guys + squash_deleted(); + settings.set_use_deleted(true); + key_info.delkey = key; + } + void clear_deleted_key() { + squash_deleted(); + settings.set_use_deleted(false); + } + key_type deleted_key() const { + assert(settings.use_deleted() + && "Must set deleted key before calling deleted_key"); + return key_info.delkey; + } + + // These are public so the iterators can use them + // True if the item at position bucknum is "deleted" marker + bool test_deleted(size_type bucknum) const { + // Invariant: !use_deleted() implies num_deleted is 0. + assert(settings.use_deleted() || num_deleted == 0); + return num_deleted > 0 && test_deleted_key(get_key(table[bucknum])); + } + bool test_deleted(const iterator &it) const { + // Invariant: !use_deleted() implies num_deleted is 0. + assert(settings.use_deleted() || num_deleted == 0); + return num_deleted > 0 && test_deleted_key(get_key(*it)); + } + bool test_deleted(const const_iterator &it) const { + // Invariant: !use_deleted() implies num_deleted is 0. + assert(settings.use_deleted() || num_deleted == 0); + return num_deleted > 0 && test_deleted_key(get_key(*it)); + } + + private: + void check_use_deleted(const char* caller) { + (void)caller; // could log it if the assert failed + assert(settings.use_deleted()); + } + + // Set it so test_deleted is true. true if object didn't used to be deleted. + bool set_deleted(iterator &it) { + check_use_deleted("set_deleted()"); + bool retval = !test_deleted(it); + // &* converts from iterator to value-type. + set_key(&(*it), key_info.delkey); + return retval; + } + // Set it so test_deleted is false. true if object used to be deleted. + bool clear_deleted(iterator &it) { + check_use_deleted("clear_deleted()"); + // Happens automatically when we assign something else in its place. + return test_deleted(it); + } + + // We also allow to set/clear the deleted bit on a const iterator. + // We allow a const_iterator for the same reason you can delete a + // const pointer: it's convenient, and semantically you can't use + // 'it' after it's been deleted anyway, so its const-ness doesn't + // really matter. + bool set_deleted(const_iterator &it) { + check_use_deleted("set_deleted()"); + bool retval = !test_deleted(it); + set_key(const_cast(&(*it)), key_info.delkey); + return retval; + } + // Set it so test_deleted is false. true if object used to be deleted. + bool clear_deleted(const_iterator &it) { + check_use_deleted("clear_deleted()"); + return test_deleted(it); + } + + // EMPTY HELPER FUNCTIONS + // This lets the user describe a key that will indicate empty (unused) + // table entries. This key should be an "impossible" entry -- + // if you try to insert it for real, you won't be able to retrieve it! + // (NB: while you pass in an entire value, only the key part is looked + // at. This is just because I don't know how to assign just a key.) + public: + // These are public so the iterators can use them + // True if the item at position bucknum is "empty" marker + bool test_empty(size_type bucknum) const { + assert(settings.use_empty()); // we always need to know what's empty! + return equals(get_key(val_info.emptyval), get_key(table[bucknum])); + } + bool test_empty(const iterator &it) const { + assert(settings.use_empty()); // we always need to know what's empty! + return equals(get_key(val_info.emptyval), get_key(*it)); + } + bool test_empty(const const_iterator &it) const { + assert(settings.use_empty()); // we always need to know what's empty! + return equals(get_key(val_info.emptyval), get_key(*it)); + } + + private: + void fill_range_with_empty(pointer table_start, pointer table_end) { + std::uninitialized_fill(table_start, table_end, val_info.emptyval); + } + + public: + // TODO(csilvers): change all callers of this to pass in a key instead, + // and take a const key_type instead of const value_type. + void set_empty_key(const_reference val) { + // Once you set the empty key, you can't change it + assert(!settings.use_empty() && "Calling set_empty_key multiple times"); + // The deleted indicator (if specified) and the empty indicator + // must be different. + assert((!settings.use_deleted() || !equals(get_key(val), key_info.delkey)) + && "Setting the empty key the same as the deleted key"); + settings.set_use_empty(true); + set_value(&val_info.emptyval, val); + + assert(!table); // must set before first use + // num_buckets was set in constructor even though table was NULL + table = val_info.allocate(num_buckets); + assert(table); + fill_range_with_empty(table, table + num_buckets); + } + // TODO(user): return a key_type rather than a value_type + value_type empty_key() const { + assert(settings.use_empty()); + return val_info.emptyval; + } + + // FUNCTIONS CONCERNING SIZE + public: + size_type size() const { return num_elements - num_deleted; } + size_type max_size() const { return val_info.max_size(); } + bool empty() const { return size() == 0; } + size_type bucket_count() const { return num_buckets; } + size_type max_bucket_count() const { return max_size(); } + size_type nonempty_bucket_count() const { return num_elements; } + // These are tr1 methods. Their idea of 'bucket' doesn't map well to + // what we do. We just say every bucket has 0 or 1 items in it. + size_type bucket_size(size_type i) const { + return begin(i) == end(i) ? 0 : 1; + } + + private: + // Because of the above, size_type(-1) is never legal; use it for errors + static const size_type ILLEGAL_BUCKET = size_type(-1); + + // Used after a string of deletes. Returns true if we actually shrunk. + // TODO(csilvers): take a delta so we can take into account inserts + // done after shrinking. Maybe make part of the Settings class? + bool maybe_shrink() { + assert(num_elements >= num_deleted); + assert((bucket_count() & (bucket_count()-1)) == 0); // is a power of two + assert(bucket_count() >= HT_MIN_BUCKETS); + bool retval = false; + + // If you construct a hashtable with < HT_DEFAULT_STARTING_BUCKETS, + // we'll never shrink until you get relatively big, and we'll never + // shrink below HT_DEFAULT_STARTING_BUCKETS. Otherwise, something + // like "dense_hash_set x; x.insert(4); x.erase(4);" will + // shrink us down to HT_MIN_BUCKETS buckets, which is too small. + const size_type num_remain = num_elements - num_deleted; + const size_type shrink_threshold = settings.shrink_threshold(); + if (shrink_threshold > 0 && num_remain < shrink_threshold && + bucket_count() > HT_DEFAULT_STARTING_BUCKETS) { + const float shrink_factor = settings.shrink_factor(); + size_type sz = bucket_count() / 2; // find how much we should shrink + while (sz > HT_DEFAULT_STARTING_BUCKETS && + num_remain < sz * shrink_factor) { + sz /= 2; // stay a power of 2 + } + dense_hashtable tmp(*this, sz); // Do the actual resizing + swap(tmp); // now we are tmp + retval = true; + } + settings.set_consider_shrink(false); // because we just considered it + return retval; + } + + // We'll let you resize a hashtable -- though this makes us copy all! + // When you resize, you say, "make it big enough for this many more elements" + // Returns true if we actually resized, false if size was already ok. + bool resize_delta(size_type delta) { + bool did_resize = false; + if ( settings.consider_shrink() ) { // see if lots of deletes happened + if ( maybe_shrink() ) + did_resize = true; + } + if (num_elements >= + (std::numeric_limits::max)() - delta) { + throw std::length_error("resize overflow"); + } + if ( bucket_count() >= HT_MIN_BUCKETS && + (num_elements + delta) <= settings.enlarge_threshold() ) + return did_resize; // we're ok as we are + + // Sometimes, we need to resize just to get rid of all the + // "deleted" buckets that are clogging up the hashtable. So when + // deciding whether to resize, count the deleted buckets (which + // are currently taking up room). But later, when we decide what + // size to resize to, *don't* count deleted buckets, since they + // get discarded during the resize. + size_type needed_size = settings.min_buckets(num_elements + delta, 0); + if ( needed_size <= bucket_count() ) // we have enough buckets + return did_resize; + + size_type resize_to = + settings.min_buckets(num_elements - num_deleted + delta, bucket_count()); + + // When num_deleted is large, we may still grow but we do not want to + // over expand. So we reduce needed_size by a portion of num_deleted + // (the exact portion does not matter). This is especially helpful + // when min_load_factor is zero (no shrink at all) to avoid doubling + // the bucket count to infinity. See also test ResizeWithoutShrink. + needed_size = settings.min_buckets(num_elements - num_deleted / 4 + delta, 0); + if (resize_to < needed_size && // may double resize_to + resize_to < (std::numeric_limits::max)() / 2) { + // This situation means that we have enough deleted elements, + // that once we purge them, we won't actually have needed to + // grow. But we may want to grow anyway: if we just purge one + // element, say, we'll have to grow anyway next time we + // insert. Might as well grow now, since we're already going + // through the trouble of copying (in order to purge the + // deleted elements). + const size_type target = + static_cast(settings.shrink_size(resize_to*2)); + if (num_elements - num_deleted + delta >= target) { + // Good, we won't be below the shrink threshhold even if we double. + resize_to *= 2; + } + } + dense_hashtable tmp(*this, resize_to); + swap(tmp); // now we are tmp + return true; + } + + // We require table be not-NULL and empty before calling this. + void resize_table(size_type /*old_size*/, size_type new_size, + base::true_type) { + table = val_info.realloc_or_die(table, new_size); + } + + void resize_table(size_type old_size, size_type new_size, base::false_type) { + val_info.deallocate(table, old_size); + table = val_info.allocate(new_size); + } + + // Used to actually do the rehashing when we grow/shrink a hashtable + void copy_from(const dense_hashtable &ht, size_type min_buckets_wanted) { + clear_to_size(settings.min_buckets(ht.size(), min_buckets_wanted)); + + // We use a normal iterator to get non-deleted bcks from ht + // We could use insert() here, but since we know there are + // no duplicates and no deleted items, we can be more efficient + assert((bucket_count() & (bucket_count()-1)) == 0); // a power of two + for ( const_iterator it = ht.begin(); it != ht.end(); ++it ) { + size_type num_probes = 0; // how many times we've probed + size_type bucknum; + const size_type bucket_count_minus_one = bucket_count() - 1; + for (bucknum = hash(get_key(*it)) & bucket_count_minus_one; + !test_empty(bucknum); // not empty + bucknum = (bucknum + JUMP_(key, num_probes)) & bucket_count_minus_one) { + ++num_probes; + assert(num_probes < bucket_count() + && "Hashtable is full: an error in key_equal<> or hash<>"); + } + set_value(&table[bucknum], *it); // copies the value to here + num_elements++; + } + settings.inc_num_ht_copies(); + } + + // Required by the spec for hashed associative container + public: + // Though the docs say this should be num_buckets, I think it's much + // more useful as num_elements. As a special feature, calling with + // req_elements==0 will cause us to shrink if we can, saving space. + void resize(size_type req_elements) { // resize to this or larger + if ( settings.consider_shrink() || req_elements == 0 ) + maybe_shrink(); + if ( req_elements > num_elements ) + resize_delta(req_elements - num_elements); + } + + // Get and change the value of shrink_factor and enlarge_factor. The + // description at the beginning of this file explains how to choose + // the values. Setting the shrink parameter to 0.0 ensures that the + // table never shrinks. + void get_resizing_parameters(float* shrink, float* grow) const { + *shrink = settings.shrink_factor(); + *grow = settings.enlarge_factor(); + } + void set_resizing_parameters(float shrink, float grow) { + settings.set_resizing_parameters(shrink, grow); + settings.reset_thresholds(bucket_count()); + } + + // CONSTRUCTORS -- as required by the specs, we take a size, + // but also let you specify a hashfunction, key comparator, + // and key extractor. We also define a copy constructor and =. + // DESTRUCTOR -- needs to free the table + explicit dense_hashtable(size_type expected_max_items_in_table = 0, + const HashFcn& hf = HashFcn(), + const EqualKey& eql = EqualKey(), + const ExtractKey& ext = ExtractKey(), + const SetKey& set = SetKey(), + const Alloc& alloc = Alloc()) + : settings(hf), + key_info(ext, set, eql), + num_deleted(0), + num_elements(0), + num_buckets(expected_max_items_in_table == 0 + ? HT_DEFAULT_STARTING_BUCKETS + : settings.min_buckets(expected_max_items_in_table, 0)), + val_info(alloc_impl(alloc)), + table(NULL) { + // table is NULL until emptyval is set. However, we set num_buckets + // here so we know how much space to allocate once emptyval is set + settings.reset_thresholds(bucket_count()); + } + + // As a convenience for resize(), we allow an optional second argument + // which lets you make this new hashtable a different size than ht + dense_hashtable(const dense_hashtable& ht, + size_type min_buckets_wanted = HT_DEFAULT_STARTING_BUCKETS) + : settings(ht.settings), + key_info(ht.key_info), + num_deleted(0), + num_elements(0), + num_buckets(0), + val_info(ht.val_info), + table(NULL) { + if (!ht.settings.use_empty()) { + // If use_empty isn't set, copy_from will crash, so we do our own copying. + assert(ht.empty()); + num_buckets = settings.min_buckets(ht.size(), min_buckets_wanted); + settings.reset_thresholds(bucket_count()); + return; + } + settings.reset_thresholds(bucket_count()); + copy_from(ht, min_buckets_wanted); // copy_from() ignores deleted entries + } + + dense_hashtable& operator= (const dense_hashtable& ht) { + if (&ht == this) return *this; // don't copy onto ourselves + if (!ht.settings.use_empty()) { + assert(ht.empty()); + dense_hashtable empty_table(ht); // empty table with ht's thresholds + this->swap(empty_table); + return *this; + } + settings = ht.settings; + key_info = ht.key_info; + set_value(&val_info.emptyval, ht.val_info.emptyval); + // copy_from() calls clear and sets num_deleted to 0 too + copy_from(ht, HT_MIN_BUCKETS); + // we purposefully don't copy the allocator, which may not be copyable + return *this; + } + + ~dense_hashtable() { + if (table) { + destroy_buckets(0, num_buckets); + val_info.deallocate(table, num_buckets); + } + } + + // Many STL algorithms use swap instead of copy constructors + void swap(dense_hashtable& ht) { + std::swap(settings, ht.settings); + std::swap(key_info, ht.key_info); + std::swap(num_deleted, ht.num_deleted); + std::swap(num_elements, ht.num_elements); + std::swap(num_buckets, ht.num_buckets); + { value_type tmp; // for annoying reasons, swap() doesn't work + set_value(&tmp, val_info.emptyval); + set_value(&val_info.emptyval, ht.val_info.emptyval); + set_value(&ht.val_info.emptyval, tmp); + } + std::swap(table, ht.table); + settings.reset_thresholds(bucket_count()); // also resets consider_shrink + ht.settings.reset_thresholds(ht.bucket_count()); + // we purposefully don't swap the allocator, which may not be swap-able + } + + private: + void clear_to_size(size_type new_num_buckets) { + if (!table) { + table = val_info.allocate(new_num_buckets); + } else { + destroy_buckets(0, num_buckets); + if (new_num_buckets != num_buckets) { // resize, if necessary + typedef base::integral_constant >::value> + realloc_ok; + resize_table(num_buckets, new_num_buckets, realloc_ok()); + } + } + assert(table); + fill_range_with_empty(table, table + new_num_buckets); + num_elements = 0; + num_deleted = 0; + num_buckets = new_num_buckets; // our new size + settings.reset_thresholds(bucket_count()); + } + + public: + // It's always nice to be able to clear a table without deallocating it + void clear() { + // If the table is already empty, and the number of buckets is + // already as we desire, there's nothing to do. + const size_type new_num_buckets = settings.min_buckets(0, 0); + if (num_elements == 0 && new_num_buckets == num_buckets) { + return; + } + clear_to_size(new_num_buckets); + } + + // Clear the table without resizing it. + // Mimicks the stl_hashtable's behaviour when clear()-ing in that it + // does not modify the bucket count + void clear_no_resize() { + if (num_elements > 0) { + assert(table); + destroy_buckets(0, num_buckets); + fill_range_with_empty(table, table + num_buckets); + } + // don't consider to shrink before another erase() + settings.reset_thresholds(bucket_count()); + num_elements = 0; + num_deleted = 0; + } + + // LOOKUP ROUTINES + private: + // Returns a pair of positions: 1st where the object is, 2nd where + // it would go if you wanted to insert it. 1st is ILLEGAL_BUCKET + // if object is not found; 2nd is ILLEGAL_BUCKET if it is. + // Note: because of deletions where-to-insert is not trivial: it's the + // first deleted bucket we see, as long as we don't find the key later + std::pair find_position(const key_type &key) const { + size_type num_probes = 0; // how many times we've probed + const size_type bucket_count_minus_one = bucket_count() - 1; + size_type bucknum = hash(key) & bucket_count_minus_one; + size_type insert_pos = ILLEGAL_BUCKET; // where we would insert + while ( 1 ) { // probe until something happens + if ( test_empty(bucknum) ) { // bucket is empty + if ( insert_pos == ILLEGAL_BUCKET ) // found no prior place to insert + return std::pair(ILLEGAL_BUCKET, bucknum); + else + return std::pair(ILLEGAL_BUCKET, insert_pos); + + } else if ( test_deleted(bucknum) ) {// keep searching, but mark to insert + if ( insert_pos == ILLEGAL_BUCKET ) + insert_pos = bucknum; + + } else if ( equals(key, get_key(table[bucknum])) ) { + return std::pair(bucknum, ILLEGAL_BUCKET); + } + ++num_probes; // we're doing another probe + bucknum = (bucknum + JUMP_(key, num_probes)) & bucket_count_minus_one; + assert(num_probes < bucket_count() + && "Hashtable is full: an error in key_equal<> or hash<>"); + } + } + + public: + + iterator find(const key_type& key) { + if ( size() == 0 ) return end(); + std::pair pos = find_position(key); + if ( pos.first == ILLEGAL_BUCKET ) // alas, not there + return end(); + else + return iterator(this, table + pos.first, table + num_buckets, false); + } + + const_iterator find(const key_type& key) const { + if ( size() == 0 ) return end(); + std::pair pos = find_position(key); + if ( pos.first == ILLEGAL_BUCKET ) // alas, not there + return end(); + else + return const_iterator(this, table + pos.first, table+num_buckets, false); + } + + // This is a tr1 method: the bucket a given key is in, or what bucket + // it would be put in, if it were to be inserted. Shrug. + size_type bucket(const key_type& key) const { + std::pair pos = find_position(key); + return pos.first == ILLEGAL_BUCKET ? pos.second : pos.first; + } + + // Counts how many elements have key key. For maps, it's either 0 or 1. + size_type count(const key_type &key) const { + std::pair pos = find_position(key); + return pos.first == ILLEGAL_BUCKET ? 0 : 1; + } + + // Likewise, equal_range doesn't really make sense for us. Oh well. + std::pair equal_range(const key_type& key) { + iterator pos = find(key); // either an iterator or end + if (pos == end()) { + return std::pair(pos, pos); + } else { + const iterator startpos = pos++; + return std::pair(startpos, pos); + } + } + std::pair equal_range(const key_type& key) + const { + const_iterator pos = find(key); // either an iterator or end + if (pos == end()) { + return std::pair(pos, pos); + } else { + const const_iterator startpos = pos++; + return std::pair(startpos, pos); + } + } + + + // INSERTION ROUTINES + private: + // Private method used by insert_noresize and find_or_insert. + iterator insert_at(const_reference obj, size_type pos) { + if (size() >= max_size()) { + throw std::length_error("insert overflow"); + } + if ( test_deleted(pos) ) { // just replace if it's been del. + // shrug: shouldn't need to be const. + const_iterator delpos(this, table + pos, table + num_buckets, false); + clear_deleted(delpos); + assert( num_deleted > 0); + --num_deleted; // used to be, now it isn't + } else { + ++num_elements; // replacing an empty bucket + } + set_value(&table[pos], obj); + return iterator(this, table + pos, table + num_buckets, false); + } + + // If you know *this is big enough to hold obj, use this routine + std::pair insert_noresize(const_reference obj) { + // First, double-check we're not inserting delkey or emptyval + assert((!settings.use_empty() || !equals(get_key(obj), + get_key(val_info.emptyval))) + && "Inserting the empty key"); + assert((!settings.use_deleted() || !equals(get_key(obj), key_info.delkey)) + && "Inserting the deleted key"); + const std::pair pos = find_position(get_key(obj)); + if ( pos.first != ILLEGAL_BUCKET) { // object was already there + return std::pair(iterator(this, table + pos.first, + table + num_buckets, false), + false); // false: we didn't insert + } else { // pos.second says where to put it + return std::pair(insert_at(obj, pos.second), true); + } + } + + // Specializations of insert(it, it) depending on the power of the iterator: + // (1) Iterator supports operator-, resize before inserting + template + void insert(ForwardIterator f, ForwardIterator l, std::forward_iterator_tag) { + size_t dist = std::distance(f, l); + if (dist >= (std::numeric_limits::max)()) { + throw std::length_error("insert-range overflow"); + } + resize_delta(static_cast(dist)); + for ( ; dist > 0; --dist, ++f) { + insert_noresize(*f); + } + } + + // (2) Arbitrary iterator, can't tell how much to resize + template + void insert(InputIterator f, InputIterator l, std::input_iterator_tag) { + for ( ; f != l; ++f) + insert(*f); + } + + public: + // This is the normal insert routine, used by the outside world + std::pair insert(const_reference obj) { + resize_delta(1); // adding an object, grow if need be + return insert_noresize(obj); + } + + // When inserting a lot at a time, we specialize on the type of iterator + template + void insert(InputIterator f, InputIterator l) { + // specializes on iterator type + insert(f, l, + typename std::iterator_traits::iterator_category()); + } + + // DefaultValue is a functor that takes a key and returns a value_type + // representing the default value to be inserted if none is found. + template + value_type& find_or_insert(const key_type& key) { + // First, double-check we're not inserting emptykey or delkey + assert((!settings.use_empty() || !equals(key, get_key(val_info.emptyval))) + && "Inserting the empty key"); + assert((!settings.use_deleted() || !equals(key, key_info.delkey)) + && "Inserting the deleted key"); + const std::pair pos = find_position(key); + DefaultValue default_value; + if ( pos.first != ILLEGAL_BUCKET) { // object was already there + return table[pos.first]; + } else if (resize_delta(1)) { // needed to rehash to make room + // Since we resized, we can't use pos, so recalculate where to insert. + return *insert_noresize(default_value(key)).first; + } else { // no need to rehash, insert right here + return *insert_at(default_value(key), pos.second); + } + } + + + // DELETION ROUTINES + size_type erase(const key_type& key) { + // First, double-check we're not trying to erase delkey or emptyval. + assert((!settings.use_empty() || !equals(key, get_key(val_info.emptyval))) + && "Erasing the empty key"); + assert((!settings.use_deleted() || !equals(key, key_info.delkey)) + && "Erasing the deleted key"); + const_iterator pos = find(key); // shrug: shouldn't need to be const + if ( pos != end() ) { + assert(!test_deleted(pos)); // or find() shouldn't have returned it + set_deleted(pos); + ++num_deleted; + settings.set_consider_shrink(true); // will think about shrink after next insert + return 1; // because we deleted one thing + } else { + return 0; // because we deleted nothing + } + } + + // We return the iterator past the deleted item. + void erase(iterator pos) { + if ( pos == end() ) return; // sanity check + if ( set_deleted(pos) ) { // true if object has been newly deleted + ++num_deleted; + settings.set_consider_shrink(true); // will think about shrink after next insert + } + } + + void erase(iterator f, iterator l) { + for ( ; f != l; ++f) { + if ( set_deleted(f) ) // should always be true + ++num_deleted; + } + settings.set_consider_shrink(true); // will think about shrink after next insert + } + + // We allow you to erase a const_iterator just like we allow you to + // erase an iterator. This is in parallel to 'delete': you can delete + // a const pointer just like a non-const pointer. The logic is that + // you can't use the object after it's erased anyway, so it doesn't matter + // if it's const or not. + void erase(const_iterator pos) { + if ( pos == end() ) return; // sanity check + if ( set_deleted(pos) ) { // true if object has been newly deleted + ++num_deleted; + settings.set_consider_shrink(true); // will think about shrink after next insert + } + } + void erase(const_iterator f, const_iterator l) { + for ( ; f != l; ++f) { + if ( set_deleted(f) ) // should always be true + ++num_deleted; + } + settings.set_consider_shrink(true); // will think about shrink after next insert + } + + + // COMPARISON + bool operator==(const dense_hashtable& ht) const { + if (size() != ht.size()) { + return false; + } else if (this == &ht) { + return true; + } else { + // Iterate through the elements in "this" and see if the + // corresponding element is in ht + for ( const_iterator it = begin(); it != end(); ++it ) { + const_iterator it2 = ht.find(get_key(*it)); + if ((it2 == ht.end()) || (*it != *it2)) { + return false; + } + } + return true; + } + } + bool operator!=(const dense_hashtable& ht) const { + return !(*this == ht); + } + + + // I/O + // We support reading and writing hashtables to disk. Alas, since + // I don't know how to write a hasher or key_equal, you have to make + // sure everything but the table is the same. We compact before writing. + private: + // Every time the disk format changes, this should probably change too + typedef unsigned long MagicNumberType; + static const MagicNumberType MAGIC_NUMBER = 0x13578642; + + public: + // I/O -- this is an add-on for writing hash table to disk + // + // INPUT and OUTPUT must be either a FILE, *or* a C++ stream + // (istream, ostream, etc) *or* a class providing + // Read(void*, size_t) and Write(const void*, size_t) + // (respectively), which writes a buffer into a stream + // (which the INPUT/OUTPUT instance presumably owns). + + typedef sparsehash_internal::pod_serializer NopointerSerializer; + + // ValueSerializer: a functor. operator()(OUTPUT*, const value_type&) + template + bool serialize(ValueSerializer serializer, OUTPUT *fp) { + squash_deleted(); // so we don't have to worry about delkey + if ( !sparsehash_internal::write_bigendian_number(fp, MAGIC_NUMBER, 4) ) + return false; + if ( !sparsehash_internal::write_bigendian_number(fp, num_buckets, 8) ) + return false; + if ( !sparsehash_internal::write_bigendian_number(fp, num_elements, 8) ) + return false; + // Now write a bitmap of non-empty buckets. + for ( size_type i = 0; i < num_buckets; i += 8 ) { + unsigned char bits = 0; + for ( int bit = 0; bit < 8; ++bit ) { + if ( i + bit < num_buckets && !test_empty(i + bit) ) + bits |= (1 << bit); + } + if ( !sparsehash_internal::write_data(fp, &bits, sizeof(bits)) ) + return false; + for ( int bit = 0; bit < 8; ++bit ) { + if ( bits & (1 << bit) ) { + if ( !serializer(fp, table[i + bit]) ) return false; + } + } + } + return true; + } + + // INPUT: anything we've written an overload of read_data() for. + // ValueSerializer: a functor. operator()(INPUT*, value_type*) + template + bool unserialize(ValueSerializer serializer, INPUT *fp) { + assert(settings.use_empty() && "empty_key not set for read"); + + clear(); // just to be consistent + MagicNumberType magic_read; + if ( !sparsehash_internal::read_bigendian_number(fp, &magic_read, 4) ) + return false; + if ( magic_read != MAGIC_NUMBER ) { + return false; + } + size_type new_num_buckets; + if ( !sparsehash_internal::read_bigendian_number(fp, &new_num_buckets, 8) ) + return false; + clear_to_size(new_num_buckets); + if ( !sparsehash_internal::read_bigendian_number(fp, &num_elements, 8) ) + return false; + + // Read the bitmap of non-empty buckets. + for (size_type i = 0; i < num_buckets; i += 8) { + unsigned char bits; + if ( !sparsehash_internal::read_data(fp, &bits, sizeof(bits)) ) + return false; + for ( int bit = 0; bit < 8; ++bit ) { + if ( i + bit < num_buckets && (bits & (1 << bit)) ) { // not empty + if ( !serializer(fp, &table[i + bit]) ) return false; + } + } + } + return true; + } + + private: + template + class alloc_impl : public A { + public: + typedef typename A::pointer pointer; + typedef typename A::size_type size_type; + + // Convert a normal allocator to one that has realloc_or_die() + alloc_impl(const A& a) : A(a) { } + + // realloc_or_die should only be used when using the default + // allocator (libc_allocator_with_realloc). + pointer realloc_or_die(pointer /*ptr*/, size_type /*n*/) { + fprintf(stderr, "realloc_or_die is only supported for " + "libc_allocator_with_realloc\n"); + exit(1); + return NULL; + } + }; + + // A template specialization of alloc_impl for + // libc_allocator_with_realloc that can handle realloc_or_die. + template + class alloc_impl > + : public libc_allocator_with_realloc { + public: + typedef typename libc_allocator_with_realloc::pointer pointer; + typedef typename libc_allocator_with_realloc::size_type size_type; + + alloc_impl(const libc_allocator_with_realloc& a) + : libc_allocator_with_realloc(a) { } + + pointer realloc_or_die(pointer ptr, size_type n) { + pointer retval = this->reallocate(ptr, n); + if (retval == NULL) { + fprintf(stderr, + "sparsehash: FATAL ERROR: failed to reallocate " + "%lu elements for ptr %p", + static_cast(n), static_cast(ptr)); + exit(1); + } + return retval; + } + }; + + // Package allocator with emptyval to eliminate memory needed for + // the zero-size allocator. + // If new fields are added to this class, we should add them to + // operator= and swap. + class ValInfo : public alloc_impl { + public: + typedef typename alloc_impl::value_type value_type; + + ValInfo(const alloc_impl& a) + : alloc_impl(a), emptyval() { } + ValInfo(const ValInfo& v) + : alloc_impl(v), emptyval(v.emptyval) { } + + value_type emptyval; // which key marks unused entries + }; + + + // Package functors with another class to eliminate memory needed for + // zero-size functors. Since ExtractKey and hasher's operator() might + // have the same function signature, they must be packaged in + // different classes. + struct Settings : + sparsehash_internal::sh_hashtable_settings { + explicit Settings(const hasher& hf) + : sparsehash_internal::sh_hashtable_settings( + hf, HT_OCCUPANCY_PCT / 100.0f, HT_EMPTY_PCT / 100.0f) {} + }; + + // Packages ExtractKey and SetKey functors. + class KeyInfo : public ExtractKey, public SetKey, public EqualKey { + public: + KeyInfo(const ExtractKey& ek, const SetKey& sk, const EqualKey& eq) + : ExtractKey(ek), + SetKey(sk), + EqualKey(eq) { + } + + // We want to return the exact same type as ExtractKey: Key or const Key& + typename ExtractKey::result_type get_key(const_reference v) const { + return ExtractKey::operator()(v); + } + void set_key(pointer v, const key_type& k) const { + SetKey::operator()(v, k); + } + bool equals(const key_type& a, const key_type& b) const { + return EqualKey::operator()(a, b); + } + + // Which key marks deleted entries. + // TODO(csilvers): make a pointer, and get rid of use_deleted (benchmark!) + typename base::remove_const::type delkey; + }; + + // Utility functions to access the templated operators + size_type hash(const key_type& v) const { + return settings.hash(v); + } + bool equals(const key_type& a, const key_type& b) const { + return key_info.equals(a, b); + } + typename ExtractKey::result_type get_key(const_reference v) const { + return key_info.get_key(v); + } + void set_key(pointer v, const key_type& k) const { + key_info.set_key(v, k); + } + + private: + // Actual data + Settings settings; + KeyInfo key_info; + + size_type num_deleted; // how many occupied buckets are marked deleted + size_type num_elements; + size_type num_buckets; + ValInfo val_info; // holds emptyval, and also the allocator + pointer table; +}; + + +// We need a global swap as well +template +inline void swap(dense_hashtable &x, + dense_hashtable &y) { + x.swap(y); +} + +#undef JUMP_ + +template +const typename dense_hashtable::size_type + dense_hashtable::ILLEGAL_BUCKET; + +// How full we let the table get before we resize. Knuth says .8 is +// good -- higher causes us to probe too much, though saves memory. +// However, we go with .5, getting better performance at the cost of +// more space (a trade-off densehashtable explicitly chooses to make). +// Feel free to play around with different values, though, via +// max_load_factor() and/or set_resizing_parameters(). +template +const int dense_hashtable::HT_OCCUPANCY_PCT = 50; + +// How empty we let the table get before we resize lower. +// It should be less than OCCUPANCY_PCT / 2 or we thrash resizing. +template +const int dense_hashtable::HT_EMPTY_PCT + = static_cast(0.4 * + dense_hashtable::HT_OCCUPANCY_PCT); + +_END_GOOGLE_NAMESPACE_ + +#endif /* _DENSEHASHTABLE_H_ */ diff --git a/src/includes/3thparty/sparsehash/internal/hashtable-common.h b/src/includes/3thparty/sparsehash/internal/hashtable-common.h new file mode 100644 index 0000000..bac2b88 --- /dev/null +++ b/src/includes/3thparty/sparsehash/internal/hashtable-common.h @@ -0,0 +1,381 @@ +// Copyright (c) 2010, Google Inc. +// All rights reserved. +// +// Redistribution and use in source and binary forms, with or without +// modification, are permitted provided that the following conditions are +// met: +// +// * Redistributions of source code must retain the above copyright +// notice, this list of conditions and the following disclaimer. +// * Redistributions in binary form must reproduce the above +// copyright notice, this list of conditions and the following disclaimer +// in the documentation and/or other materials provided with the +// distribution. +// * Neither the name of Google Inc. nor the names of its +// contributors may be used to endorse or promote products derived from +// this software without specific prior written permission. +// +// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS +// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT +// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR +// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT +// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, +// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT +// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, +// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY +// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT +// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE +// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + +// --- +// +// Provides classes shared by both sparse and dense hashtable. +// +// sh_hashtable_settings has parameters for growing and shrinking +// a hashtable. It also packages zero-size functor (ie. hasher). +// +// Other functions and classes provide common code for serializing +// and deserializing hashtables to a stream (such as a FILE*). + +#ifndef UTIL_GTL_HASHTABLE_COMMON_H_ +#define UTIL_GTL_HASHTABLE_COMMON_H_ + +#include +#include +#include +#include // for size_t +#include +#include // For length_error + +_START_GOOGLE_NAMESPACE_ + +template struct SparsehashCompileAssert { }; +#define SPARSEHASH_COMPILE_ASSERT(expr, msg) \ + __attribute__((unused)) typedef SparsehashCompileAssert<(bool(expr))> msg[bool(expr) ? 1 : -1] + +namespace sparsehash_internal { + +// Adaptor methods for reading/writing data from an INPUT or OUPTUT +// variable passed to serialize() or unserialize(). For now we +// have implemented INPUT/OUTPUT for FILE*, istream*/ostream* (note +// they are pointers, unlike typical use), or else a pointer to +// something that supports a Read()/Write() method. +// +// For technical reasons, we implement read_data/write_data in two +// stages. The actual work is done in *_data_internal, which takes +// the stream argument twice: once as a template type, and once with +// normal type information. (We only use the second version.) We do +// this because of how C++ picks what function overload to use. If we +// implemented this the naive way: +// bool read_data(istream* is, const void* data, size_t length); +// template read_data(T* fp, const void* data, size_t length); +// C++ would prefer the second version for every stream type except +// istream. However, we want C++ to prefer the first version for +// streams that are *subclasses* of istream, such as istringstream. +// This is not possible given the way template types are resolved. So +// we split the stream argument in two, one of which is templated and +// one of which is not. The specialized functions (like the istream +// version above) ignore the template arg and use the second, 'type' +// arg, getting subclass matching as normal. The 'catch-all' +// functions (the second version above) use the template arg to deduce +// the type, and use a second, void* arg to achieve the desired +// 'catch-all' semantics. + +// ----- low-level I/O for FILE* ---- + +template +inline bool read_data_internal(Ignored*, FILE* fp, + void* data, size_t length) { + return fread(data, length, 1, fp) == 1; +} + +template +inline bool write_data_internal(Ignored*, FILE* fp, + const void* data, size_t length) { + return fwrite(data, length, 1, fp) == 1; +} + +// ----- low-level I/O for iostream ---- + +// We want the caller to be responsible for #including , not +// us, because iostream is a big header! According to the standard, +// it's only legal to delay the instantiation the way we want to if +// the istream/ostream is a template type. So we jump through hoops. +template +inline bool read_data_internal_for_istream(ISTREAM* fp, + void* data, size_t length) { + return fp->read(reinterpret_cast(data), length).good(); +} +template +inline bool read_data_internal(Ignored*, std::istream* fp, + void* data, size_t length) { + return read_data_internal_for_istream(fp, data, length); +} + +template +inline bool write_data_internal_for_ostream(OSTREAM* fp, + const void* data, size_t length) { + return fp->write(reinterpret_cast(data), length).good(); +} +template +inline bool write_data_internal(Ignored*, std::ostream* fp, + const void* data, size_t length) { + return write_data_internal_for_ostream(fp, data, length); +} + +// ----- low-level I/O for custom streams ---- + +// The INPUT type needs to support a Read() method that takes a +// buffer and a length and returns the number of bytes read. +template +inline bool read_data_internal(INPUT* fp, void*, + void* data, size_t length) { + return static_cast(fp->Read(data, length)) == length; +} + +// The OUTPUT type needs to support a Write() operation that takes +// a buffer and a length and returns the number of bytes written. +template +inline bool write_data_internal(OUTPUT* fp, void*, + const void* data, size_t length) { + return static_cast(fp->Write(data, length)) == length; +} + +// ----- low-level I/O: the public API ---- + +template +inline bool read_data(INPUT* fp, void* data, size_t length) { + return read_data_internal(fp, fp, data, length); +} + +template +inline bool write_data(OUTPUT* fp, const void* data, size_t length) { + return write_data_internal(fp, fp, data, length); +} + +// Uses read_data() and write_data() to read/write an integer. +// length is the number of bytes to read/write (which may differ +// from sizeof(IntType), allowing us to save on a 32-bit system +// and load on a 64-bit system). Excess bytes are taken to be 0. +// INPUT and OUTPUT must match legal inputs to read/write_data (above). +template +bool read_bigendian_number(INPUT* fp, IntType* value, size_t length) { + *value = 0; + unsigned char byte; + // We require IntType to be unsigned or else the shifting gets all screwy. + SPARSEHASH_COMPILE_ASSERT(static_cast(-1) > static_cast(0), + serializing_int_requires_an_unsigned_type); + for (size_t i = 0; i < length; ++i) { + if (!read_data(fp, &byte, sizeof(byte))) return false; + *value |= static_cast(byte) << ((length - 1 - i) * 8); + } + return true; +} + +template +bool write_bigendian_number(OUTPUT* fp, IntType value, size_t length) { + unsigned char byte; + // We require IntType to be unsigned or else the shifting gets all screwy. + SPARSEHASH_COMPILE_ASSERT(static_cast(-1) > static_cast(0), + serializing_int_requires_an_unsigned_type); + for (size_t i = 0; i < length; ++i) { + byte = (sizeof(value) <= length-1 - i) + ? 0 : static_cast((value >> ((length-1 - i) * 8)) & 255); + if (!write_data(fp, &byte, sizeof(byte))) return false; + } + return true; +} + +// If your keys and values are simple enough, you can pass this +// serializer to serialize()/unserialize(). "Simple enough" means +// value_type is a POD type that contains no pointers. Note, +// however, we don't try to normalize endianness. +// This is the type used for NopointerSerializer. +template struct pod_serializer { + template + bool operator()(INPUT* fp, value_type* value) const { + return read_data(fp, value, sizeof(*value)); + } + + template + bool operator()(OUTPUT* fp, const value_type& value) const { + return write_data(fp, &value, sizeof(value)); + } +}; + + +// Settings contains parameters for growing and shrinking the table. +// It also packages zero-size functor (ie. hasher). +// +// It does some munging of the hash value in cases where we think +// (fear) the original hash function might not be very good. In +// particular, the default hash of pointers is the identity hash, +// so probably all the low bits are 0. We identify when we think +// we're hashing a pointer, and chop off the low bits. Note this +// isn't perfect: even when the key is a pointer, we can't tell +// for sure that the hash is the identity hash. If it's not, this +// is needless work (and possibly, though not likely, harmful). + +template +class sh_hashtable_settings : public HashFunc { + public: + typedef Key key_type; + typedef HashFunc hasher; + typedef SizeType size_type; + + public: + sh_hashtable_settings(const hasher& hf, + const float ht_occupancy_flt, + const float ht_empty_flt) + : hasher(hf), + enlarge_threshold_(0), + shrink_threshold_(0), + consider_shrink_(false), + use_empty_(false), + use_deleted_(false), + num_ht_copies_(0) { + set_enlarge_factor(ht_occupancy_flt); + set_shrink_factor(ht_empty_flt); + } + + size_type hash(const key_type& v) const { + // We munge the hash value when we don't trust hasher::operator(). + return hash_munger::MungedHash(hasher::operator()(v)); + } + + float enlarge_factor() const { + return enlarge_factor_; + } + void set_enlarge_factor(float f) { + enlarge_factor_ = f; + } + float shrink_factor() const { + return shrink_factor_; + } + void set_shrink_factor(float f) { + shrink_factor_ = f; + } + + size_type enlarge_threshold() const { + return enlarge_threshold_; + } + void set_enlarge_threshold(size_type t) { + enlarge_threshold_ = t; + } + size_type shrink_threshold() const { + return shrink_threshold_; + } + void set_shrink_threshold(size_type t) { + shrink_threshold_ = t; + } + + size_type enlarge_size(size_type x) const { + return static_cast(x * enlarge_factor_); + } + size_type shrink_size(size_type x) const { + return static_cast(x * shrink_factor_); + } + + bool consider_shrink() const { + return consider_shrink_; + } + void set_consider_shrink(bool t) { + consider_shrink_ = t; + } + + bool use_empty() const { + return use_empty_; + } + void set_use_empty(bool t) { + use_empty_ = t; + } + + bool use_deleted() const { + return use_deleted_; + } + void set_use_deleted(bool t) { + use_deleted_ = t; + } + + size_type num_ht_copies() const { + return static_cast(num_ht_copies_); + } + void inc_num_ht_copies() { + ++num_ht_copies_; + } + + // Reset the enlarge and shrink thresholds + void reset_thresholds(size_type num_buckets) { + set_enlarge_threshold(enlarge_size(num_buckets)); + set_shrink_threshold(shrink_size(num_buckets)); + // whatever caused us to reset already considered + set_consider_shrink(false); + } + + // Caller is resposible for calling reset_threshold right after + // set_resizing_parameters. + void set_resizing_parameters(float shrink, float grow) { + assert(shrink >= 0.0); + assert(grow <= 1.0); + if (shrink > grow/2.0f) + shrink = grow / 2.0f; // otherwise we thrash hashtable size + set_shrink_factor(shrink); + set_enlarge_factor(grow); + } + + // This is the smallest size a hashtable can be without being too crowded + // If you like, you can give a min #buckets as well as a min #elts + size_type min_buckets(size_type num_elts, size_type min_buckets_wanted) { + float enlarge = enlarge_factor(); + size_type sz = HT_MIN_BUCKETS; // min buckets allowed + while ( sz < min_buckets_wanted || + num_elts >= static_cast(sz * enlarge) ) { + // This just prevents overflowing size_type, since sz can exceed + // max_size() here. + if (static_cast(sz * 2) < sz) { + throw std::length_error("resize overflow"); // protect against overflow + } + sz *= 2; + } + return sz; + } + + private: + template class hash_munger { + public: + static size_t MungedHash(size_t hash) { + return hash; + } + }; + // This matches when the hashtable key is a pointer. + template class hash_munger { + public: + static size_t MungedHash(size_t hash) { + // TODO(csilvers): consider rotating instead: + // static const int shift = (sizeof(void *) == 4) ? 2 : 3; + // return (hash << (sizeof(hash) * 8) - shift)) | (hash >> shift); + // This matters if we ever change sparse/dense_hash_* to compare + // hashes before comparing actual values. It's speedy on x86. + return hash / sizeof(void*); // get rid of known-0 bits + } + }; + + size_type enlarge_threshold_; // table.size() * enlarge_factor + size_type shrink_threshold_; // table.size() * shrink_factor + float enlarge_factor_; // how full before resize + float shrink_factor_; // how empty before resize + // consider_shrink=true if we should try to shrink before next insert + bool consider_shrink_; + bool use_empty_; // used only by densehashtable, not sparsehashtable + bool use_deleted_; // false until delkey has been set + // num_ht_copies is a counter incremented every Copy/Move + unsigned int num_ht_copies_; +}; + +} // namespace sparsehash_internal + +#undef SPARSEHASH_COMPILE_ASSERT +_END_GOOGLE_NAMESPACE_ + +#endif // UTIL_GTL_HASHTABLE_COMMON_H_ diff --git a/src/includes/3thparty/sparsehash/internal/libc_allocator_with_realloc.h b/src/includes/3thparty/sparsehash/internal/libc_allocator_with_realloc.h new file mode 100644 index 0000000..769329f --- /dev/null +++ b/src/includes/3thparty/sparsehash/internal/libc_allocator_with_realloc.h @@ -0,0 +1,122 @@ +// Copyright (c) 2010, Google Inc. +// All rights reserved. +// +// Redistribution and use in source and binary forms, with or without +// modification, are permitted provided that the following conditions are +// met: +// +// * Redistributions of source code must retain the above copyright +// notice, this list of conditions and the following disclaimer. +// * Redistributions in binary form must reproduce the above +// copyright notice, this list of conditions and the following disclaimer +// in the documentation and/or other materials provided with the +// distribution. +// * Neither the name of Google Inc. nor the names of its +// contributors may be used to endorse or promote products derived from +// this software without specific prior written permission. +// +// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS +// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT +// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR +// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT +// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, +// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT +// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, +// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY +// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT +// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE +// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + +// --- + +#ifndef UTIL_GTL_LIBC_ALLOCATOR_WITH_REALLOC_H_ +#define UTIL_GTL_LIBC_ALLOCATOR_WITH_REALLOC_H_ + +#include +#include // for malloc/realloc/free +#include // for ptrdiff_t +#include // for placement new + +_START_GOOGLE_NAMESPACE_ + +template +class libc_allocator_with_realloc { + public: + typedef T value_type; + typedef size_t size_type; + typedef ptrdiff_t difference_type; + + typedef T* pointer; + typedef const T* const_pointer; + typedef T& reference; + typedef const T& const_reference; + + libc_allocator_with_realloc() {} + libc_allocator_with_realloc(const libc_allocator_with_realloc&) {} + ~libc_allocator_with_realloc() {} + + pointer address(reference r) const { return &r; } + const_pointer address(const_reference r) const { return &r; } + + pointer allocate(size_type n, const_pointer = 0) { + return static_cast(malloc(n * sizeof(value_type))); + } + void deallocate(pointer p, size_type) { + free(p); + } + pointer reallocate(pointer p, size_type n) { + // p points to a storage array whose objects have already been destroyed + // cast to void* to prevent compiler warnings about calling realloc() on + // an object which cannot be relocated in memory + return static_cast(realloc(static_cast(p), n * sizeof(value_type))); + } + + size_type max_size() const { + return static_cast(-1) / sizeof(value_type); + } + + void construct(pointer p, const value_type& val) { + new(p) value_type(val); + } + void destroy(pointer p) { p->~value_type(); } + + template + libc_allocator_with_realloc(const libc_allocator_with_realloc&) {} + + template + struct rebind { + typedef libc_allocator_with_realloc other; + }; +}; + +// libc_allocator_with_realloc specialization. +template<> +class libc_allocator_with_realloc { + public: + typedef void value_type; + typedef size_t size_type; + typedef ptrdiff_t difference_type; + typedef void* pointer; + typedef const void* const_pointer; + + template + struct rebind { + typedef libc_allocator_with_realloc other; + }; +}; + +template +inline bool operator==(const libc_allocator_with_realloc&, + const libc_allocator_with_realloc&) { + return true; +} + +template +inline bool operator!=(const libc_allocator_with_realloc&, + const libc_allocator_with_realloc&) { + return false; +} + +_END_GOOGLE_NAMESPACE_ + +#endif // UTIL_GTL_LIBC_ALLOCATOR_WITH_REALLOC_H_ diff --git a/src/includes/3thparty/sparsehash/internal/sparseconfig.h b/src/includes/3thparty/sparsehash/internal/sparseconfig.h new file mode 100644 index 0000000..18f4b2d --- /dev/null +++ b/src/includes/3thparty/sparsehash/internal/sparseconfig.h @@ -0,0 +1,46 @@ +/* + * NOTE: This file is for internal use only. + * Do not use these #defines in your own program! + */ + +/* Namespace for Google classes */ +#define GOOGLE_NAMESPACE ::google + +/* the location of the header defining hash functions */ +#define HASH_FUN_H + +/* the namespace of the hash<> function */ +#define HASH_NAMESPACE std + +/* Define to 1 if you have the header file. */ +#define HAVE_INTTYPES_H 1 + +/* Define to 1 if the system has the type `long long'. */ +#define HAVE_LONG_LONG 1 + +/* Define to 1 if you have the `memcpy' function. */ +#define HAVE_MEMCPY 1 + +/* Define to 1 if you have the header file. */ +#define HAVE_STDINT_H 1 + +/* Define to 1 if you have the header file. */ +#define HAVE_SYS_TYPES_H 1 + +/* Define to 1 if the system has the type `uint16_t'. */ +#define HAVE_UINT16_T 1 + +/* Define to 1 if the system has the type `u_int16_t'. */ +#define HAVE_U_INT16_T 1 + +/* Define to 1 if the system has the type `__uint16'. */ +/* #undef HAVE___UINT16 */ + +/* The system-provided hash function including the namespace. */ +#define SPARSEHASH_HASH HASH_NAMESPACE::hash + +/* Stops putting the code inside the Google namespace */ +#define _END_GOOGLE_NAMESPACE_ } + +/* Puts following code inside the Google namespace */ +#define _START_GOOGLE_NAMESPACE_ namespace google { diff --git a/src/includes/3thparty/sparsehash/internal/sparsehashtable.h b/src/includes/3thparty/sparsehash/internal/sparsehashtable.h new file mode 100644 index 0000000..f54ea51 --- /dev/null +++ b/src/includes/3thparty/sparsehash/internal/sparsehashtable.h @@ -0,0 +1,1247 @@ +// Copyright (c) 2005, Google Inc. +// All rights reserved. +// +// Redistribution and use in source and binary forms, with or without +// modification, are permitted provided that the following conditions are +// met: +// +// * Redistributions of source code must retain the above copyright +// notice, this list of conditions and the following disclaimer. +// * Redistributions in binary form must reproduce the above +// copyright notice, this list of conditions and the following disclaimer +// in the documentation and/or other materials provided with the +// distribution. +// * Neither the name of Google Inc. nor the names of its +// contributors may be used to endorse or promote products derived from +// this software without specific prior written permission. +// +// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS +// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT +// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR +// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT +// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, +// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT +// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, +// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY +// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT +// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE +// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + +// --- +// +// A sparse hashtable is a particular implementation of +// a hashtable: one that is meant to minimize memory use. +// It does this by using a *sparse table* (cf sparsetable.h), +// which uses between 1 and 2 bits to store empty buckets +// (we may need another bit for hashtables that support deletion). +// +// When empty buckets are so cheap, an appealing hashtable +// implementation is internal probing, in which the hashtable +// is a single table, and collisions are resolved by trying +// to insert again in another bucket. The most cache-efficient +// internal probing schemes are linear probing (which suffers, +// alas, from clumping) and quadratic probing, which is what +// we implement by default. +// +// Deleted buckets are a bit of a pain. We have to somehow mark +// deleted buckets (the probing must distinguish them from empty +// buckets). The most principled way is to have another bitmap, +// but that's annoying and takes up space. Instead we let the +// user specify an "impossible" key. We set deleted buckets +// to have the impossible key. +// +// Note it is possible to change the value of the delete key +// on the fly; you can even remove it, though after that point +// the hashtable is insert_only until you set it again. +// +// You probably shouldn't use this code directly. Use +// sparse_hash_map<> or sparse_hash_set<> instead. +// +// You can modify the following, below: +// HT_OCCUPANCY_PCT -- how full before we double size +// HT_EMPTY_PCT -- how empty before we halve size +// HT_MIN_BUCKETS -- smallest bucket size +// HT_DEFAULT_STARTING_BUCKETS -- default bucket size at construct-time +// +// You can also change enlarge_factor (which defaults to +// HT_OCCUPANCY_PCT), and shrink_factor (which defaults to +// HT_EMPTY_PCT) with set_resizing_parameters(). +// +// How to decide what values to use? +// shrink_factor's default of .4 * OCCUPANCY_PCT, is probably good. +// HT_MIN_BUCKETS is probably unnecessary since you can specify +// (indirectly) the starting number of buckets at construct-time. +// For enlarge_factor, you can use this chart to try to trade-off +// expected lookup time to the space taken up. By default, this +// code uses quadratic probing, though you can change it to linear +// via _JUMP below if you really want to. +// +// From http://www.augustana.ca/~mohrj/courses/1999.fall/csc210/lecture_notes/hashing.html +// NUMBER OF PROBES / LOOKUP Successful Unsuccessful +// Quadratic collision resolution 1 - ln(1-L) - L/2 1/(1-L) - L - ln(1-L) +// Linear collision resolution [1+1/(1-L)]/2 [1+1/(1-L)2]/2 +// +// -- enlarge_factor -- 0.10 0.50 0.60 0.75 0.80 0.90 0.99 +// QUADRATIC COLLISION RES. +// probes/successful lookup 1.05 1.44 1.62 2.01 2.21 2.85 5.11 +// probes/unsuccessful lookup 1.11 2.19 2.82 4.64 5.81 11.4 103.6 +// LINEAR COLLISION RES. +// probes/successful lookup 1.06 1.5 1.75 2.5 3.0 5.5 50.5 +// probes/unsuccessful lookup 1.12 2.5 3.6 8.5 13.0 50.0 5000.0 +// +// The value type is required to be copy constructible and default +// constructible, but it need not be (and commonly isn't) assignable. + +#ifndef _SPARSEHASHTABLE_H_ +#define _SPARSEHASHTABLE_H_ + +#include +#include +#include // For swap(), eg +#include // for iterator tags +#include // for numeric_limits +#include // for pair +#include // for remove_const +#include +#include // IWYU pragma: export +#include // For length_error + +_START_GOOGLE_NAMESPACE_ + +namespace base { // just to make google->opensource transition easier +using GOOGLE_NAMESPACE::remove_const; +} + +#ifndef SPARSEHASH_STAT_UPDATE +#define SPARSEHASH_STAT_UPDATE(x) ((void) 0) +#endif + +// The probing method +// Linear probing +// #define JUMP_(key, num_probes) ( 1 ) +// Quadratic probing +#define JUMP_(key, num_probes) ( num_probes ) + +// The smaller this is, the faster lookup is (because the group bitmap is +// smaller) and the faster insert is, because there's less to move. +// On the other hand, there are more groups. Since group::size_type is +// a short, this number should be of the form 32*x + 16 to avoid waste. +static const u_int16_t DEFAULT_GROUP_SIZE = 48; // fits in 1.5 words + +// Hashtable class, used to implement the hashed associative containers +// hash_set and hash_map. +// +// Value: what is stored in the table (each bucket is a Value). +// Key: something in a 1-to-1 correspondence to a Value, that can be used +// to search for a Value in the table (find() takes a Key). +// HashFcn: Takes a Key and returns an integer, the more unique the better. +// ExtractKey: given a Value, returns the unique Key associated with it. +// Must inherit from unary_function, or at least have a +// result_type enum indicating the return type of operator(). +// SetKey: given a Value* and a Key, modifies the value such that +// ExtractKey(value) == key. We guarantee this is only called +// with key == deleted_key. +// EqualKey: Given two Keys, says whether they are the same (that is, +// if they are both associated with the same Value). +// Alloc: STL allocator to use to allocate memory. + +template +class sparse_hashtable; + +template +struct sparse_hashtable_iterator; + +template +struct sparse_hashtable_const_iterator; + +// As far as iterating, we're basically just a sparsetable +// that skips over deleted elements. +template +struct sparse_hashtable_iterator { + private: + typedef typename A::template rebind::other value_alloc_type; + + public: + typedef sparse_hashtable_iterator iterator; + typedef sparse_hashtable_const_iterator const_iterator; + typedef typename sparsetable::nonempty_iterator + st_iterator; + + typedef std::forward_iterator_tag iterator_category; // very little defined! + typedef V value_type; + typedef typename value_alloc_type::difference_type difference_type; + typedef typename value_alloc_type::size_type size_type; + typedef typename value_alloc_type::reference reference; + typedef typename value_alloc_type::pointer pointer; + + // "Real" constructor and default constructor + sparse_hashtable_iterator(const sparse_hashtable *h, + st_iterator it, st_iterator it_end) + : ht(h), pos(it), end(it_end) { advance_past_deleted(); } + sparse_hashtable_iterator() { } // not ever used internally + // The default destructor is fine; we don't define one + // The default operator= is fine; we don't define one + + // Happy dereferencer + reference operator*() const { return *pos; } + pointer operator->() const { return &(operator*()); } + + // Arithmetic. The only hard part is making sure that + // we're not on a marked-deleted array element + void advance_past_deleted() { + while ( pos != end && ht->test_deleted(*this) ) + ++pos; + } + iterator& operator++() { + assert(pos != end); ++pos; advance_past_deleted(); return *this; + } + iterator operator++(int) { iterator tmp(*this); ++*this; return tmp; } + + // Comparison. + bool operator==(const iterator& it) const { return pos == it.pos; } + bool operator!=(const iterator& it) const { return pos != it.pos; } + + + // The actual data + const sparse_hashtable *ht; + st_iterator pos, end; +}; + +// Now do it all again, but with const-ness! +template +struct sparse_hashtable_const_iterator { + private: + typedef typename A::template rebind::other value_alloc_type; + + public: + typedef sparse_hashtable_iterator iterator; + typedef sparse_hashtable_const_iterator const_iterator; + typedef typename sparsetable::const_nonempty_iterator + st_iterator; + + typedef std::forward_iterator_tag iterator_category; // very little defined! + typedef V value_type; + typedef typename value_alloc_type::difference_type difference_type; + typedef typename value_alloc_type::size_type size_type; + typedef typename value_alloc_type::const_reference reference; + typedef typename value_alloc_type::const_pointer pointer; + + // "Real" constructor and default constructor + sparse_hashtable_const_iterator(const sparse_hashtable *h, + st_iterator it, st_iterator it_end) + : ht(h), pos(it), end(it_end) { advance_past_deleted(); } + // This lets us convert regular iterators to const iterators + sparse_hashtable_const_iterator() { } // never used internally + sparse_hashtable_const_iterator(const iterator &it) + : ht(it.ht), pos(it.pos), end(it.end) { } + // The default destructor is fine; we don't define one + // The default operator= is fine; we don't define one + + // Happy dereferencer + reference operator*() const { return *pos; } + pointer operator->() const { return &(operator*()); } + + // Arithmetic. The only hard part is making sure that + // we're not on a marked-deleted array element + void advance_past_deleted() { + while ( pos != end && ht->test_deleted(*this) ) + ++pos; + } + const_iterator& operator++() { + assert(pos != end); ++pos; advance_past_deleted(); return *this; + } + const_iterator operator++(int) { const_iterator tmp(*this); ++*this; return tmp; } + + // Comparison. + bool operator==(const const_iterator& it) const { return pos == it.pos; } + bool operator!=(const const_iterator& it) const { return pos != it.pos; } + + + // The actual data + const sparse_hashtable *ht; + st_iterator pos, end; +}; + +// And once again, but this time freeing up memory as we iterate +template +struct sparse_hashtable_destructive_iterator { + private: + typedef typename A::template rebind::other value_alloc_type; + + public: + typedef sparse_hashtable_destructive_iterator iterator; + typedef typename sparsetable::destructive_iterator + st_iterator; + + typedef std::forward_iterator_tag iterator_category; // very little defined! + typedef V value_type; + typedef typename value_alloc_type::difference_type difference_type; + typedef typename value_alloc_type::size_type size_type; + typedef typename value_alloc_type::reference reference; + typedef typename value_alloc_type::pointer pointer; + + // "Real" constructor and default constructor + sparse_hashtable_destructive_iterator(const + sparse_hashtable *h, + st_iterator it, st_iterator it_end) + : ht(h), pos(it), end(it_end) { advance_past_deleted(); } + sparse_hashtable_destructive_iterator() { } // never used internally + // The default destructor is fine; we don't define one + // The default operator= is fine; we don't define one + + // Happy dereferencer + reference operator*() const { return *pos; } + pointer operator->() const { return &(operator*()); } + + // Arithmetic. The only hard part is making sure that + // we're not on a marked-deleted array element + void advance_past_deleted() { + while ( pos != end && ht->test_deleted(*this) ) + ++pos; + } + iterator& operator++() { + assert(pos != end); ++pos; advance_past_deleted(); return *this; + } + iterator operator++(int) { iterator tmp(*this); ++*this; return tmp; } + + // Comparison. + bool operator==(const iterator& it) const { return pos == it.pos; } + bool operator!=(const iterator& it) const { return pos != it.pos; } + + + // The actual data + const sparse_hashtable *ht; + st_iterator pos, end; +}; + + +template +class sparse_hashtable { + private: + typedef typename Alloc::template rebind::other value_alloc_type; + + public: + typedef Key key_type; + typedef Value value_type; + typedef HashFcn hasher; + typedef EqualKey key_equal; + typedef Alloc allocator_type; + + typedef typename value_alloc_type::size_type size_type; + typedef typename value_alloc_type::difference_type difference_type; + typedef typename value_alloc_type::reference reference; + typedef typename value_alloc_type::const_reference const_reference; + typedef typename value_alloc_type::pointer pointer; + typedef typename value_alloc_type::const_pointer const_pointer; + typedef sparse_hashtable_iterator + iterator; + + typedef sparse_hashtable_const_iterator + const_iterator; + + typedef sparse_hashtable_destructive_iterator + destructive_iterator; + + // These come from tr1. For us they're the same as regular iterators. + typedef iterator local_iterator; + typedef const_iterator const_local_iterator; + + // How full we let the table get before we resize, by default. + // Knuth says .8 is good -- higher causes us to probe too much, + // though it saves memory. + static const int HT_OCCUPANCY_PCT; // = 80 (out of 100); + + // How empty we let the table get before we resize lower, by default. + // (0.0 means never resize lower.) + // It should be less than OCCUPANCY_PCT / 2 or we thrash resizing + static const int HT_EMPTY_PCT; // = 0.4 * HT_OCCUPANCY_PCT; + + // Minimum size we're willing to let hashtables be. + // Must be a power of two, and at least 4. + // Note, however, that for a given hashtable, the initial size is a + // function of the first constructor arg, and may be >HT_MIN_BUCKETS. + static const size_type HT_MIN_BUCKETS = 4; + + // By default, if you don't specify a hashtable size at + // construction-time, we use this size. Must be a power of two, and + // at least HT_MIN_BUCKETS. + static const size_type HT_DEFAULT_STARTING_BUCKETS = 32; + + // ITERATOR FUNCTIONS + iterator begin() { return iterator(this, table.nonempty_begin(), + table.nonempty_end()); } + iterator end() { return iterator(this, table.nonempty_end(), + table.nonempty_end()); } + const_iterator begin() const { return const_iterator(this, + table.nonempty_begin(), + table.nonempty_end()); } + const_iterator end() const { return const_iterator(this, + table.nonempty_end(), + table.nonempty_end()); } + + // These come from tr1 unordered_map. They iterate over 'bucket' n. + // For sparsehashtable, we could consider each 'group' to be a bucket, + // I guess, but I don't really see the point. We'll just consider + // bucket n to be the n-th element of the sparsetable, if it's occupied, + // or some empty element, otherwise. + local_iterator begin(size_type i) { + if (table.test(i)) + return local_iterator(this, table.get_iter(i), table.nonempty_end()); + else + return local_iterator(this, table.nonempty_end(), table.nonempty_end()); + } + local_iterator end(size_type i) { + local_iterator it = begin(i); + if (table.test(i) && !test_deleted(i)) + ++it; + return it; + } + const_local_iterator begin(size_type i) const { + if (table.test(i)) + return const_local_iterator(this, table.get_iter(i), + table.nonempty_end()); + else + return const_local_iterator(this, table.nonempty_end(), + table.nonempty_end()); + } + const_local_iterator end(size_type i) const { + const_local_iterator it = begin(i); + if (table.test(i) && !test_deleted(i)) + ++it; + return it; + } + + // This is used when resizing + destructive_iterator destructive_begin() { + return destructive_iterator(this, table.destructive_begin(), + table.destructive_end()); + } + destructive_iterator destructive_end() { + return destructive_iterator(this, table.destructive_end(), + table.destructive_end()); + } + + + // ACCESSOR FUNCTIONS for the things we templatize on, basically + hasher hash_funct() const { return settings; } + key_equal key_eq() const { return key_info; } + allocator_type get_allocator() const { return table.get_allocator(); } + + // Accessor function for statistics gathering. + int num_table_copies() const { return settings.num_ht_copies(); } + + private: + // We need to copy values when we set the special marker for deleted + // elements, but, annoyingly, we can't just use the copy assignment + // operator because value_type might not be assignable (it's often + // pair). We use explicit destructor invocation and + // placement new to get around this. Arg. + void set_value(pointer dst, const_reference src) { + dst->~value_type(); // delete the old value, if any + new(dst) value_type(src); + } + + // This is used as a tag for the copy constructor, saying to destroy its + // arg We have two ways of destructively copying: with potentially growing + // the hashtable as we copy, and without. To make sure the outside world + // can't do a destructive copy, we make the typename private. + enum MoveDontCopyT {MoveDontCopy, MoveDontGrow}; + + // DELETE HELPER FUNCTIONS + // This lets the user describe a key that will indicate deleted + // table entries. This key should be an "impossible" entry -- + // if you try to insert it for real, you won't be able to retrieve it! + // (NB: while you pass in an entire value, only the key part is looked + // at. This is just because I don't know how to assign just a key.) + private: + void squash_deleted() { // gets rid of any deleted entries we have + if ( num_deleted ) { // get rid of deleted before writing + sparse_hashtable tmp(MoveDontGrow, *this); + swap(tmp); // now we are tmp + } + assert(num_deleted == 0); + } + + // Test if the given key is the deleted indicator. Requires + // num_deleted > 0, for correctness of read(), and because that + // guarantees that key_info.delkey is valid. + bool test_deleted_key(const key_type& key) const { + assert(num_deleted > 0); + return equals(key_info.delkey, key); + } + + public: + void set_deleted_key(const key_type &key) { + // It's only safe to change what "deleted" means if we purge deleted guys + squash_deleted(); + settings.set_use_deleted(true); + key_info.delkey = key; + } + void clear_deleted_key() { + squash_deleted(); + settings.set_use_deleted(false); + } + key_type deleted_key() const { + assert(settings.use_deleted() + && "Must set deleted key before calling deleted_key"); + return key_info.delkey; + } + + // These are public so the iterators can use them + // True if the item at position bucknum is "deleted" marker + bool test_deleted(size_type bucknum) const { + // Invariant: !use_deleted() implies num_deleted is 0. + assert(settings.use_deleted() || num_deleted == 0); + return num_deleted > 0 && table.test(bucknum) && + test_deleted_key(get_key(table.unsafe_get(bucknum))); + } + bool test_deleted(const iterator &it) const { + // Invariant: !use_deleted() implies num_deleted is 0. + assert(settings.use_deleted() || num_deleted == 0); + return num_deleted > 0 && test_deleted_key(get_key(*it)); + } + bool test_deleted(const const_iterator &it) const { + // Invariant: !use_deleted() implies num_deleted is 0. + assert(settings.use_deleted() || num_deleted == 0); + return num_deleted > 0 && test_deleted_key(get_key(*it)); + } + bool test_deleted(const destructive_iterator &it) const { + // Invariant: !use_deleted() implies num_deleted is 0. + assert(settings.use_deleted() || num_deleted == 0); + return num_deleted > 0 && test_deleted_key(get_key(*it)); + } + + private: + void check_use_deleted(const char* caller) { + (void)caller; // could log it if the assert failed + assert(settings.use_deleted()); + } + + // Set it so test_deleted is true. true if object didn't used to be deleted. + // TODO(csilvers): make these private (also in densehashtable.h) + bool set_deleted(iterator &it) { + check_use_deleted("set_deleted()"); + bool retval = !test_deleted(it); + // &* converts from iterator to value-type. + set_key(&(*it), key_info.delkey); + return retval; + } + // Set it so test_deleted is false. true if object used to be deleted. + bool clear_deleted(iterator &it) { + check_use_deleted("clear_deleted()"); + // Happens automatically when we assign something else in its place. + return test_deleted(it); + } + + // We also allow to set/clear the deleted bit on a const iterator. + // We allow a const_iterator for the same reason you can delete a + // const pointer: it's convenient, and semantically you can't use + // 'it' after it's been deleted anyway, so its const-ness doesn't + // really matter. + bool set_deleted(const_iterator &it) { + check_use_deleted("set_deleted()"); + bool retval = !test_deleted(it); + set_key(const_cast(&(*it)), key_info.delkey); + return retval; + } + // Set it so test_deleted is false. true if object used to be deleted. + bool clear_deleted(const_iterator &it) { + check_use_deleted("clear_deleted()"); + return test_deleted(it); + } + + // FUNCTIONS CONCERNING SIZE + public: + size_type size() const { return table.num_nonempty() - num_deleted; } + size_type max_size() const { return table.max_size(); } + bool empty() const { return size() == 0; } + size_type bucket_count() const { return table.size(); } + size_type max_bucket_count() const { return max_size(); } + // These are tr1 methods. Their idea of 'bucket' doesn't map well to + // what we do. We just say every bucket has 0 or 1 items in it. + size_type bucket_size(size_type i) const { + return begin(i) == end(i) ? 0 : 1; + } + + private: + // Because of the above, size_type(-1) is never legal; use it for errors + static const size_type ILLEGAL_BUCKET = size_type(-1); + + // Used after a string of deletes. Returns true if we actually shrunk. + // TODO(csilvers): take a delta so we can take into account inserts + // done after shrinking. Maybe make part of the Settings class? + bool maybe_shrink() { + assert(table.num_nonempty() >= num_deleted); + assert((bucket_count() & (bucket_count()-1)) == 0); // is a power of two + assert(bucket_count() >= HT_MIN_BUCKETS); + bool retval = false; + + // If you construct a hashtable with < HT_DEFAULT_STARTING_BUCKETS, + // we'll never shrink until you get relatively big, and we'll never + // shrink below HT_DEFAULT_STARTING_BUCKETS. Otherwise, something + // like "dense_hash_set x; x.insert(4); x.erase(4);" will + // shrink us down to HT_MIN_BUCKETS buckets, which is too small. + const size_type num_remain = table.num_nonempty() - num_deleted; + const size_type shrink_threshold = settings.shrink_threshold(); + if (shrink_threshold > 0 && num_remain < shrink_threshold && + bucket_count() > HT_DEFAULT_STARTING_BUCKETS) { + const float shrink_factor = settings.shrink_factor(); + size_type sz = bucket_count() / 2; // find how much we should shrink + while (sz > HT_DEFAULT_STARTING_BUCKETS && + num_remain < static_cast(sz * shrink_factor)) { + sz /= 2; // stay a power of 2 + } + sparse_hashtable tmp(MoveDontCopy, *this, sz); + swap(tmp); // now we are tmp + retval = true; + } + settings.set_consider_shrink(false); // because we just considered it + return retval; + } + + // We'll let you resize a hashtable -- though this makes us copy all! + // When you resize, you say, "make it big enough for this many more elements" + // Returns true if we actually resized, false if size was already ok. + bool resize_delta(size_type delta) { + bool did_resize = false; + if ( settings.consider_shrink() ) { // see if lots of deletes happened + if ( maybe_shrink() ) + did_resize = true; + } + if (table.num_nonempty() >= + (std::numeric_limits::max)() - delta) { + throw std::length_error("resize overflow"); + } + if ( bucket_count() >= HT_MIN_BUCKETS && + (table.num_nonempty() + delta) <= settings.enlarge_threshold() ) + return did_resize; // we're ok as we are + + // Sometimes, we need to resize just to get rid of all the + // "deleted" buckets that are clogging up the hashtable. So when + // deciding whether to resize, count the deleted buckets (which + // are currently taking up room). But later, when we decide what + // size to resize to, *don't* count deleted buckets, since they + // get discarded during the resize. + const size_type needed_size = + settings.min_buckets(table.num_nonempty() + delta, 0); + if ( needed_size <= bucket_count() ) // we have enough buckets + return did_resize; + + size_type resize_to = + settings.min_buckets(table.num_nonempty() - num_deleted + delta, + bucket_count()); + if (resize_to < needed_size && // may double resize_to + resize_to < (std::numeric_limits::max)() / 2) { + // This situation means that we have enough deleted elements, + // that once we purge them, we won't actually have needed to + // grow. But we may want to grow anyway: if we just purge one + // element, say, we'll have to grow anyway next time we + // insert. Might as well grow now, since we're already going + // through the trouble of copying (in order to purge the + // deleted elements). + const size_type target = + static_cast(settings.shrink_size(resize_to*2)); + if (table.num_nonempty() - num_deleted + delta >= target) { + // Good, we won't be below the shrink threshhold even if we double. + resize_to *= 2; + } + } + + sparse_hashtable tmp(MoveDontCopy, *this, resize_to); + swap(tmp); // now we are tmp + return true; + } + + // Used to actually do the rehashing when we grow/shrink a hashtable + void copy_from(const sparse_hashtable &ht, size_type min_buckets_wanted) { + clear(); // clear table, set num_deleted to 0 + + // If we need to change the size of our table, do it now + const size_type resize_to = + settings.min_buckets(ht.size(), min_buckets_wanted); + if ( resize_to > bucket_count() ) { // we don't have enough buckets + table.resize(resize_to); // sets the number of buckets + settings.reset_thresholds(bucket_count()); + } + + // We use a normal iterator to get non-deleted bcks from ht + // We could use insert() here, but since we know there are + // no duplicates and no deleted items, we can be more efficient + assert((bucket_count() & (bucket_count()-1)) == 0); // a power of two + for ( const_iterator it = ht.begin(); it != ht.end(); ++it ) { + size_type num_probes = 0; // how many times we've probed + size_type bucknum; + const size_type bucket_count_minus_one = bucket_count() - 1; + for (bucknum = hash(get_key(*it)) & bucket_count_minus_one; + table.test(bucknum); // not empty + bucknum = (bucknum + JUMP_(key, num_probes)) & bucket_count_minus_one) { + ++num_probes; + assert(num_probes < bucket_count() + && "Hashtable is full: an error in key_equal<> or hash<>"); + } + table.set(bucknum, *it); // copies the value to here + } + settings.inc_num_ht_copies(); + } + + // Implementation is like copy_from, but it destroys the table of the + // "from" guy by freeing sparsetable memory as we iterate. This is + // useful in resizing, since we're throwing away the "from" guy anyway. + void move_from(MoveDontCopyT mover, sparse_hashtable &ht, + size_type min_buckets_wanted) { + clear(); // clear table, set num_deleted to 0 + + // If we need to change the size of our table, do it now + size_type resize_to; + if ( mover == MoveDontGrow ) + resize_to = ht.bucket_count(); // keep same size as old ht + else // MoveDontCopy + resize_to = settings.min_buckets(ht.size(), min_buckets_wanted); + if ( resize_to > bucket_count() ) { // we don't have enough buckets + table.resize(resize_to); // sets the number of buckets + settings.reset_thresholds(bucket_count()); + } + + // We use a normal iterator to get non-deleted bcks from ht + // We could use insert() here, but since we know there are + // no duplicates and no deleted items, we can be more efficient + assert( (bucket_count() & (bucket_count()-1)) == 0); // a power of two + // THIS IS THE MAJOR LINE THAT DIFFERS FROM COPY_FROM(): + for ( destructive_iterator it = ht.destructive_begin(); + it != ht.destructive_end(); ++it ) { + size_type num_probes = 0; // how many times we've probed + size_type bucknum; + for ( bucknum = hash(get_key(*it)) & (bucket_count()-1); // h % buck_cnt + table.test(bucknum); // not empty + bucknum = (bucknum + JUMP_(key, num_probes)) & (bucket_count()-1) ) { + ++num_probes; + assert(num_probes < bucket_count() + && "Hashtable is full: an error in key_equal<> or hash<>"); + } + table.set(bucknum, *it); // copies the value to here + } + settings.inc_num_ht_copies(); + } + + + // Required by the spec for hashed associative container + public: + // Though the docs say this should be num_buckets, I think it's much + // more useful as num_elements. As a special feature, calling with + // req_elements==0 will cause us to shrink if we can, saving space. + void resize(size_type req_elements) { // resize to this or larger + if ( settings.consider_shrink() || req_elements == 0 ) + maybe_shrink(); + if ( req_elements > table.num_nonempty() ) // we only grow + resize_delta(req_elements - table.num_nonempty()); + } + + // Get and change the value of shrink_factor and enlarge_factor. The + // description at the beginning of this file explains how to choose + // the values. Setting the shrink parameter to 0.0 ensures that the + // table never shrinks. + void get_resizing_parameters(float* shrink, float* grow) const { + *shrink = settings.shrink_factor(); + *grow = settings.enlarge_factor(); + } + void set_resizing_parameters(float shrink, float grow) { + settings.set_resizing_parameters(shrink, grow); + settings.reset_thresholds(bucket_count()); + } + + // CONSTRUCTORS -- as required by the specs, we take a size, + // but also let you specify a hashfunction, key comparator, + // and key extractor. We also define a copy constructor and =. + // DESTRUCTOR -- the default is fine, surprisingly. + explicit sparse_hashtable(size_type expected_max_items_in_table = 0, + const HashFcn& hf = HashFcn(), + const EqualKey& eql = EqualKey(), + const ExtractKey& ext = ExtractKey(), + const SetKey& set = SetKey(), + const Alloc& alloc = Alloc()) + : settings(hf), + key_info(ext, set, eql), + num_deleted(0), + table((expected_max_items_in_table == 0 + ? HT_DEFAULT_STARTING_BUCKETS + : settings.min_buckets(expected_max_items_in_table, 0)), + alloc) { + settings.reset_thresholds(bucket_count()); + } + + // As a convenience for resize(), we allow an optional second argument + // which lets you make this new hashtable a different size than ht. + // We also provide a mechanism of saying you want to "move" the ht argument + // into us instead of copying. + sparse_hashtable(const sparse_hashtable& ht, + size_type min_buckets_wanted = HT_DEFAULT_STARTING_BUCKETS) + : settings(ht.settings), + key_info(ht.key_info), + num_deleted(0), + table(0, ht.get_allocator()) { + settings.reset_thresholds(bucket_count()); + copy_from(ht, min_buckets_wanted); // copy_from() ignores deleted entries + } + sparse_hashtable(MoveDontCopyT mover, sparse_hashtable& ht, + size_type min_buckets_wanted = HT_DEFAULT_STARTING_BUCKETS) + : settings(ht.settings), + key_info(ht.key_info), + num_deleted(0), + table(0, ht.get_allocator()) { + settings.reset_thresholds(bucket_count()); + move_from(mover, ht, min_buckets_wanted); // ignores deleted entries + } + + sparse_hashtable& operator= (const sparse_hashtable& ht) { + if (&ht == this) return *this; // don't copy onto ourselves + settings = ht.settings; + key_info = ht.key_info; + num_deleted = ht.num_deleted; + // copy_from() calls clear and sets num_deleted to 0 too + copy_from(ht, HT_MIN_BUCKETS); + // we purposefully don't copy the allocator, which may not be copyable + return *this; + } + + // Many STL algorithms use swap instead of copy constructors + void swap(sparse_hashtable& ht) { + std::swap(settings, ht.settings); + std::swap(key_info, ht.key_info); + std::swap(num_deleted, ht.num_deleted); + table.swap(ht.table); + settings.reset_thresholds(bucket_count()); // also resets consider_shrink + ht.settings.reset_thresholds(ht.bucket_count()); + // we purposefully don't swap the allocator, which may not be swap-able + } + + // It's always nice to be able to clear a table without deallocating it + void clear() { + if (!empty() || (num_deleted != 0)) { + table.clear(); + } + settings.reset_thresholds(bucket_count()); + num_deleted = 0; + } + + // LOOKUP ROUTINES + private: + // Returns a pair of positions: 1st where the object is, 2nd where + // it would go if you wanted to insert it. 1st is ILLEGAL_BUCKET + // if object is not found; 2nd is ILLEGAL_BUCKET if it is. + // Note: because of deletions where-to-insert is not trivial: it's the + // first deleted bucket we see, as long as we don't find the key later + std::pair find_position(const key_type &key) const { + size_type num_probes = 0; // how many times we've probed + const size_type bucket_count_minus_one = bucket_count() - 1; + size_type bucknum = hash(key) & bucket_count_minus_one; + size_type insert_pos = ILLEGAL_BUCKET; // where we would insert + SPARSEHASH_STAT_UPDATE(total_lookups += 1); + while ( 1 ) { // probe until something happens + if ( !table.test(bucknum) ) { // bucket is empty + SPARSEHASH_STAT_UPDATE(total_probes += num_probes); + if ( insert_pos == ILLEGAL_BUCKET ) // found no prior place to insert + return std::pair(ILLEGAL_BUCKET, bucknum); + else + return std::pair(ILLEGAL_BUCKET, insert_pos); + + } else if ( test_deleted(bucknum) ) {// keep searching, but mark to insert + if ( insert_pos == ILLEGAL_BUCKET ) + insert_pos = bucknum; + + } else if ( equals(key, get_key(table.unsafe_get(bucknum))) ) { + SPARSEHASH_STAT_UPDATE(total_probes += num_probes); + return std::pair(bucknum, ILLEGAL_BUCKET); + } + ++num_probes; // we're doing another probe + bucknum = (bucknum + JUMP_(key, num_probes)) & bucket_count_minus_one; + assert(num_probes < bucket_count() + && "Hashtable is full: an error in key_equal<> or hash<>"); + } + } + + public: + + iterator find(const key_type& key) { + if ( size() == 0 ) return end(); + std::pair pos = find_position(key); + if ( pos.first == ILLEGAL_BUCKET ) // alas, not there + return end(); + else + return iterator(this, table.get_iter(pos.first), table.nonempty_end()); + } + + const_iterator find(const key_type& key) const { + if ( size() == 0 ) return end(); + std::pair pos = find_position(key); + if ( pos.first == ILLEGAL_BUCKET ) // alas, not there + return end(); + else + return const_iterator(this, + table.get_iter(pos.first), table.nonempty_end()); + } + + // This is a tr1 method: the bucket a given key is in, or what bucket + // it would be put in, if it were to be inserted. Shrug. + size_type bucket(const key_type& key) const { + std::pair pos = find_position(key); + return pos.first == ILLEGAL_BUCKET ? pos.second : pos.first; + } + + // Counts how many elements have key key. For maps, it's either 0 or 1. + size_type count(const key_type &key) const { + std::pair pos = find_position(key); + return pos.first == ILLEGAL_BUCKET ? 0 : 1; + } + + // Likewise, equal_range doesn't really make sense for us. Oh well. + std::pair equal_range(const key_type& key) { + iterator pos = find(key); // either an iterator or end + if (pos == end()) { + return std::pair(pos, pos); + } else { + const iterator startpos = pos++; + return std::pair(startpos, pos); + } + } + std::pair equal_range(const key_type& key) + const { + const_iterator pos = find(key); // either an iterator or end + if (pos == end()) { + return std::pair(pos, pos); + } else { + const const_iterator startpos = pos++; + return std::pair(startpos, pos); + } + } + + + // INSERTION ROUTINES + private: + // Private method used by insert_noresize and find_or_insert. + iterator insert_at(const_reference obj, size_type pos) { + if (size() >= max_size()) { + throw std::length_error("insert overflow"); + } + if ( test_deleted(pos) ) { // just replace if it's been deleted + // The set() below will undelete this object. We just worry about stats + assert(num_deleted > 0); + --num_deleted; // used to be, now it isn't + } + table.set(pos, obj); + return iterator(this, table.get_iter(pos), table.nonempty_end()); + } + + // If you know *this is big enough to hold obj, use this routine + std::pair insert_noresize(const_reference obj) { + // First, double-check we're not inserting delkey + assert((!settings.use_deleted() || !equals(get_key(obj), key_info.delkey)) + && "Inserting the deleted key"); + const std::pair pos = find_position(get_key(obj)); + if ( pos.first != ILLEGAL_BUCKET) { // object was already there + return std::pair(iterator(this, table.get_iter(pos.first), + table.nonempty_end()), + false); // false: we didn't insert + } else { // pos.second says where to put it + return std::pair(insert_at(obj, pos.second), true); + } + } + + // Specializations of insert(it, it) depending on the power of the iterator: + // (1) Iterator supports operator-, resize before inserting + template + void insert(ForwardIterator f, ForwardIterator l, std::forward_iterator_tag) { + size_t dist = std::distance(f, l); + if (dist >= (std::numeric_limits::max)()) { + throw std::length_error("insert-range overflow"); + } + resize_delta(static_cast(dist)); + for ( ; dist > 0; --dist, ++f) { + insert_noresize(*f); + } + } + + // (2) Arbitrary iterator, can't tell how much to resize + template + void insert(InputIterator f, InputIterator l, std::input_iterator_tag) { + for ( ; f != l; ++f) + insert(*f); + } + + public: + // This is the normal insert routine, used by the outside world + std::pair insert(const_reference obj) { + resize_delta(1); // adding an object, grow if need be + return insert_noresize(obj); + } + + // When inserting a lot at a time, we specialize on the type of iterator + template + void insert(InputIterator f, InputIterator l) { + // specializes on iterator type + insert(f, l, + typename std::iterator_traits::iterator_category()); + } + + // DefaultValue is a functor that takes a key and returns a value_type + // representing the default value to be inserted if none is found. + template + value_type& find_or_insert(const key_type& key) { + // First, double-check we're not inserting delkey + assert((!settings.use_deleted() || !equals(key, key_info.delkey)) + && "Inserting the deleted key"); + const std::pair pos = find_position(key); + DefaultValue default_value; + if ( pos.first != ILLEGAL_BUCKET) { // object was already there + return *table.get_iter(pos.first); + } else if (resize_delta(1)) { // needed to rehash to make room + // Since we resized, we can't use pos, so recalculate where to insert. + return *insert_noresize(default_value(key)).first; + } else { // no need to rehash, insert right here + return *insert_at(default_value(key), pos.second); + } + } + + // DELETION ROUTINES + size_type erase(const key_type& key) { + // First, double-check we're not erasing delkey. + assert((!settings.use_deleted() || !equals(key, key_info.delkey)) + && "Erasing the deleted key"); + assert(!settings.use_deleted() || !equals(key, key_info.delkey)); + const_iterator pos = find(key); // shrug: shouldn't need to be const + if ( pos != end() ) { + assert(!test_deleted(pos)); // or find() shouldn't have returned it + set_deleted(pos); + ++num_deleted; + // will think about shrink after next insert + settings.set_consider_shrink(true); + return 1; // because we deleted one thing + } else { + return 0; // because we deleted nothing + } + } + + // We return the iterator past the deleted item. + void erase(iterator pos) { + if ( pos == end() ) return; // sanity check + if ( set_deleted(pos) ) { // true if object has been newly deleted + ++num_deleted; + // will think about shrink after next insert + settings.set_consider_shrink(true); + } + } + + void erase(iterator f, iterator l) { + for ( ; f != l; ++f) { + if ( set_deleted(f) ) // should always be true + ++num_deleted; + } + // will think about shrink after next insert + settings.set_consider_shrink(true); + } + + // We allow you to erase a const_iterator just like we allow you to + // erase an iterator. This is in parallel to 'delete': you can delete + // a const pointer just like a non-const pointer. The logic is that + // you can't use the object after it's erased anyway, so it doesn't matter + // if it's const or not. + void erase(const_iterator pos) { + if ( pos == end() ) return; // sanity check + if ( set_deleted(pos) ) { // true if object has been newly deleted + ++num_deleted; + // will think about shrink after next insert + settings.set_consider_shrink(true); + } + } + void erase(const_iterator f, const_iterator l) { + for ( ; f != l; ++f) { + if ( set_deleted(f) ) // should always be true + ++num_deleted; + } + // will think about shrink after next insert + settings.set_consider_shrink(true); + } + + + // COMPARISON + bool operator==(const sparse_hashtable& ht) const { + if (size() != ht.size()) { + return false; + } else if (this == &ht) { + return true; + } else { + // Iterate through the elements in "this" and see if the + // corresponding element is in ht + for ( const_iterator it = begin(); it != end(); ++it ) { + const_iterator it2 = ht.find(get_key(*it)); + if ((it2 == ht.end()) || (*it != *it2)) { + return false; + } + } + return true; + } + } + bool operator!=(const sparse_hashtable& ht) const { + return !(*this == ht); + } + + + // I/O + // We support reading and writing hashtables to disk. NOTE that + // this only stores the hashtable metadata, not the stuff you've + // actually put in the hashtable! Alas, since I don't know how to + // write a hasher or key_equal, you have to make sure everything + // but the table is the same. We compact before writing. + // + // The OUTPUT type needs to support a Write() operation. File and + // OutputBuffer are appropriate types to pass in. + // + // The INPUT type needs to support a Read() operation. File and + // InputBuffer are appropriate types to pass in. + template + bool write_metadata(OUTPUT *fp) { + squash_deleted(); // so we don't have to worry about delkey + return table.write_metadata(fp); + } + + template + bool read_metadata(INPUT *fp) { + num_deleted = 0; // since we got rid before writing + const bool result = table.read_metadata(fp); + settings.reset_thresholds(bucket_count()); + return result; + } + + // Only meaningful if value_type is a POD. + template + bool write_nopointer_data(OUTPUT *fp) { + return table.write_nopointer_data(fp); + } + + // Only meaningful if value_type is a POD. + template + bool read_nopointer_data(INPUT *fp) { + return table.read_nopointer_data(fp); + } + + // INPUT and OUTPUT must be either a FILE, *or* a C++ stream + // (istream, ostream, etc) *or* a class providing + // Read(void*, size_t) and Write(const void*, size_t) + // (respectively), which writes a buffer into a stream + // (which the INPUT/OUTPUT instance presumably owns). + + typedef sparsehash_internal::pod_serializer NopointerSerializer; + + // ValueSerializer: a functor. operator()(OUTPUT*, const value_type&) + template + bool serialize(ValueSerializer serializer, OUTPUT *fp) { + squash_deleted(); // so we don't have to worry about delkey + return table.serialize(serializer, fp); + } + + // ValueSerializer: a functor. operator()(INPUT*, value_type*) + template + bool unserialize(ValueSerializer serializer, INPUT *fp) { + num_deleted = 0; // since we got rid before writing + const bool result = table.unserialize(serializer, fp); + settings.reset_thresholds(bucket_count()); + return result; + } + + private: + // Table is the main storage class. + typedef sparsetable Table; + + // Package templated functors with the other types to eliminate memory + // needed for storing these zero-size operators. Since ExtractKey and + // hasher's operator() might have the same function signature, they + // must be packaged in different classes. + struct Settings : + sparsehash_internal::sh_hashtable_settings { + explicit Settings(const hasher& hf) + : sparsehash_internal::sh_hashtable_settings( + hf, HT_OCCUPANCY_PCT / 100.0f, HT_EMPTY_PCT / 100.0f) {} + }; + + // KeyInfo stores delete key and packages zero-size functors: + // ExtractKey and SetKey. + class KeyInfo : public ExtractKey, public SetKey, public EqualKey { + public: + KeyInfo(const ExtractKey& ek, const SetKey& sk, const EqualKey& eq) + : ExtractKey(ek), + SetKey(sk), + EqualKey(eq) { + } + // We want to return the exact same type as ExtractKey: Key or const Key& + typename ExtractKey::result_type get_key(const_reference v) const { + return ExtractKey::operator()(v); + } + void set_key(pointer v, const key_type& k) const { + SetKey::operator()(v, k); + } + bool equals(const key_type& a, const key_type& b) const { + return EqualKey::operator()(a, b); + } + + // Which key marks deleted entries. + // TODO(csilvers): make a pointer, and get rid of use_deleted (benchmark!) + typename base::remove_const::type delkey; + }; + + // Utility functions to access the templated operators + size_type hash(const key_type& v) const { + return settings.hash(v); + } + bool equals(const key_type& a, const key_type& b) const { + return key_info.equals(a, b); + } + typename ExtractKey::result_type get_key(const_reference v) const { + return key_info.get_key(v); + } + void set_key(pointer v, const key_type& k) const { + key_info.set_key(v, k); + } + + private: + // Actual data + Settings settings; + KeyInfo key_info; + size_type num_deleted; // how many occupied buckets are marked deleted + Table table; // holds num_buckets and num_elements too +}; + + +// We need a global swap as well +template +inline void swap(sparse_hashtable &x, + sparse_hashtable &y) { + x.swap(y); +} + +#undef JUMP_ + +template +const typename sparse_hashtable::size_type + sparse_hashtable::ILLEGAL_BUCKET; + +// How full we let the table get before we resize. Knuth says .8 is +// good -- higher causes us to probe too much, though saves memory +template +const int sparse_hashtable::HT_OCCUPANCY_PCT = 80; + +// How empty we let the table get before we resize lower. +// It should be less than OCCUPANCY_PCT / 2 or we thrash resizing +template +const int sparse_hashtable::HT_EMPTY_PCT + = static_cast(0.4 * + sparse_hashtable::HT_OCCUPANCY_PCT); + +_END_GOOGLE_NAMESPACE_ + +#endif /* _SPARSEHASHTABLE_H_ */ diff --git a/src/includes/3thparty/sparsehash/sparse_hash_map b/src/includes/3thparty/sparsehash/sparse_hash_map new file mode 100644 index 0000000..1687a8b --- /dev/null +++ b/src/includes/3thparty/sparsehash/sparse_hash_map @@ -0,0 +1,363 @@ +// Copyright (c) 2005, Google Inc. +// All rights reserved. +// +// Redistribution and use in source and binary forms, with or without +// modification, are permitted provided that the following conditions are +// met: +// +// * Redistributions of source code must retain the above copyright +// notice, this list of conditions and the following disclaimer. +// * Redistributions in binary form must reproduce the above +// copyright notice, this list of conditions and the following disclaimer +// in the documentation and/or other materials provided with the +// distribution. +// * Neither the name of Google Inc. nor the names of its +// contributors may be used to endorse or promote products derived from +// this software without specific prior written permission. +// +// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS +// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT +// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR +// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT +// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, +// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT +// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, +// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY +// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT +// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE +// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + +// --- +// +// This is just a very thin wrapper over sparsehashtable.h, just +// like sgi stl's stl_hash_map is a very thin wrapper over +// stl_hashtable. The major thing we define is operator[], because +// we have a concept of a data_type which stl_hashtable doesn't +// (it only has a key and a value). +// +// We adhere mostly to the STL semantics for hash-map. One important +// exception is that insert() may invalidate iterators entirely -- STL +// semantics are that insert() may reorder iterators, but they all +// still refer to something valid in the hashtable. Not so for us. +// Likewise, insert() may invalidate pointers into the hashtable. +// (Whether insert invalidates iterators and pointers depends on +// whether it results in a hashtable resize). On the plus side, +// delete() doesn't invalidate iterators or pointers at all, or even +// change the ordering of elements. +// +// Here are a few "power user" tips: +// +// 1) set_deleted_key(): +// Unlike STL's hash_map, if you want to use erase() you +// *must* call set_deleted_key() after construction. +// +// 2) resize(0): +// When an item is deleted, its memory isn't freed right +// away. This is what allows you to iterate over a hashtable +// and call erase() without invalidating the iterator. +// To force the memory to be freed, call resize(0). +// For tr1 compatibility, this can also be called as rehash(0). +// +// 3) min_load_factor(0.0) +// Setting the minimum load factor to 0.0 guarantees that +// the hash table will never shrink. +// +// Roughly speaking: +// (1) dense_hash_map: fastest, uses the most memory unless entries are small +// (2) sparse_hash_map: slowest, uses the least memory +// (3) hash_map / unordered_map (STL): in the middle +// +// Typically I use sparse_hash_map when I care about space and/or when +// I need to save the hashtable on disk. I use hash_map otherwise. I +// don't personally use dense_hash_map ever; some people use it for +// small maps with lots of lookups. +// +// - dense_hash_map has, typically, about 78% memory overhead (if your +// data takes up X bytes, the hash_map uses .78X more bytes in overhead). +// - sparse_hash_map has about 4 bits overhead per entry. +// - sparse_hash_map can be 3-7 times slower than the others for lookup and, +// especially, inserts. See time_hash_map.cc for details. +// +// See /usr/(local/)?doc/sparsehash-*/sparse_hash_map.html +// for information about how to use this class. + +#ifndef _SPARSE_HASH_MAP_H_ +#define _SPARSE_HASH_MAP_H_ + +#include +#include // needed by stl_alloc +#include // for equal_to<>, select1st<>, etc +#include // for alloc +#include // for pair<> +#include +#include // IWYU pragma: export +#include HASH_FUN_H // for hash<> +_START_GOOGLE_NAMESPACE_ + +template , // defined in sparseconfig.h + class EqualKey = std::equal_to, + class Alloc = libc_allocator_with_realloc > > +class sparse_hash_map { + private: + // Apparently select1st is not stl-standard, so we define our own + struct SelectKey { + typedef const Key& result_type; + const Key& operator()(const std::pair& p) const { + return p.first; + } + }; + struct SetKey { + void operator()(std::pair* value, const Key& new_key) const { + *const_cast(&value->first) = new_key; + // It would be nice to clear the rest of value here as well, in + // case it's taking up a lot of memory. We do this by clearing + // the value. This assumes T has a zero-arg constructor! + value->second = T(); + } + }; + // For operator[]. + struct DefaultValue { + std::pair operator()(const Key& key) { + return std::make_pair(key, T()); + } + }; + + // The actual data + typedef sparse_hashtable, Key, HashFcn, SelectKey, + SetKey, EqualKey, Alloc> ht; + ht rep; + + public: + typedef typename ht::key_type key_type; + typedef T data_type; + typedef T mapped_type; + typedef typename ht::value_type value_type; + typedef typename ht::hasher hasher; + typedef typename ht::key_equal key_equal; + typedef Alloc allocator_type; + + typedef typename ht::size_type size_type; + typedef typename ht::difference_type difference_type; + typedef typename ht::pointer pointer; + typedef typename ht::const_pointer const_pointer; + typedef typename ht::reference reference; + typedef typename ht::const_reference const_reference; + + typedef typename ht::iterator iterator; + typedef typename ht::const_iterator const_iterator; + typedef typename ht::local_iterator local_iterator; + typedef typename ht::const_local_iterator const_local_iterator; + + // Iterator functions + iterator begin() { return rep.begin(); } + iterator end() { return rep.end(); } + const_iterator begin() const { return rep.begin(); } + const_iterator end() const { return rep.end(); } + + // These come from tr1's unordered_map. For us, a bucket has 0 or 1 elements. + local_iterator begin(size_type i) { return rep.begin(i); } + local_iterator end(size_type i) { return rep.end(i); } + const_local_iterator begin(size_type i) const { return rep.begin(i); } + const_local_iterator end(size_type i) const { return rep.end(i); } + + // Accessor functions + allocator_type get_allocator() const { return rep.get_allocator(); } + hasher hash_funct() const { return rep.hash_funct(); } + hasher hash_function() const { return hash_funct(); } + key_equal key_eq() const { return rep.key_eq(); } + + + // Constructors + explicit sparse_hash_map(size_type expected_max_items_in_table = 0, + const hasher& hf = hasher(), + const key_equal& eql = key_equal(), + const allocator_type& alloc = allocator_type()) + : rep(expected_max_items_in_table, hf, eql, SelectKey(), SetKey(), alloc) { + } + + template + sparse_hash_map(InputIterator f, InputIterator l, + size_type expected_max_items_in_table = 0, + const hasher& hf = hasher(), + const key_equal& eql = key_equal(), + const allocator_type& alloc = allocator_type()) + : rep(expected_max_items_in_table, hf, eql, SelectKey(), SetKey(), alloc) { + rep.insert(f, l); + } + // We use the default copy constructor + // We use the default operator=() + // We use the default destructor + + void clear() { rep.clear(); } + void swap(sparse_hash_map& hs) { rep.swap(hs.rep); } + + + // Functions concerning size + size_type size() const { return rep.size(); } + size_type max_size() const { return rep.max_size(); } + bool empty() const { return rep.empty(); } + size_type bucket_count() const { return rep.bucket_count(); } + size_type max_bucket_count() const { return rep.max_bucket_count(); } + + // These are tr1 methods. bucket() is the bucket the key is or would be in. + size_type bucket_size(size_type i) const { return rep.bucket_size(i); } + size_type bucket(const key_type& key) const { return rep.bucket(key); } + float load_factor() const { + return size() * 1.0f / bucket_count(); + } + float max_load_factor() const { + float shrink, grow; + rep.get_resizing_parameters(&shrink, &grow); + return grow; + } + void max_load_factor(float new_grow) { + float shrink, grow; + rep.get_resizing_parameters(&shrink, &grow); + rep.set_resizing_parameters(shrink, new_grow); + } + // These aren't tr1 methods but perhaps ought to be. + float min_load_factor() const { + float shrink, grow; + rep.get_resizing_parameters(&shrink, &grow); + return shrink; + } + void min_load_factor(float new_shrink) { + float shrink, grow; + rep.get_resizing_parameters(&shrink, &grow); + rep.set_resizing_parameters(new_shrink, grow); + } + // Deprecated; use min_load_factor() or max_load_factor() instead. + void set_resizing_parameters(float shrink, float grow) { + rep.set_resizing_parameters(shrink, grow); + } + + void resize(size_type hint) { rep.resize(hint); } + void rehash(size_type hint) { resize(hint); } // the tr1 name + + // Lookup routines + iterator find(const key_type& key) { return rep.find(key); } + const_iterator find(const key_type& key) const { return rep.find(key); } + + data_type& operator[](const key_type& key) { // This is our value-add! + // If key is in the hashtable, returns find(key)->second, + // otherwise returns insert(value_type(key, T()).first->second. + // Note it does not create an empty T unless the find fails. + return rep.template find_or_insert(key).second; + } + + size_type count(const key_type& key) const { return rep.count(key); } + + std::pair equal_range(const key_type& key) { + return rep.equal_range(key); + } + std::pair equal_range(const key_type& key) + const { + return rep.equal_range(key); + } + + // Insertion routines + std::pair insert(const value_type& obj) { + return rep.insert(obj); + } + template void insert(InputIterator f, InputIterator l) { + rep.insert(f, l); + } + void insert(const_iterator f, const_iterator l) { + rep.insert(f, l); + } + // Required for std::insert_iterator; the passed-in iterator is ignored. + iterator insert(iterator, const value_type& obj) { + return insert(obj).first; + } + + // Deletion routines + // THESE ARE NON-STANDARD! I make you specify an "impossible" key + // value to identify deleted buckets. You can change the key as + // time goes on, or get rid of it entirely to be insert-only. + void set_deleted_key(const key_type& key) { + rep.set_deleted_key(key); + } + void clear_deleted_key() { rep.clear_deleted_key(); } + key_type deleted_key() const { return rep.deleted_key(); } + + // These are standard + size_type erase(const key_type& key) { return rep.erase(key); } + void erase(iterator it) { rep.erase(it); } + void erase(iterator f, iterator l) { rep.erase(f, l); } + + + // Comparison + bool operator==(const sparse_hash_map& hs) const { return rep == hs.rep; } + bool operator!=(const sparse_hash_map& hs) const { return rep != hs.rep; } + + + // I/O -- this is an add-on for writing metainformation to disk + // + // For maximum flexibility, this does not assume a particular + // file type (though it will probably be a FILE *). We just pass + // the fp through to rep. + + // If your keys and values are simple enough, you can pass this + // serializer to serialize()/unserialize(). "Simple enough" means + // value_type is a POD type that contains no pointers. Note, + // however, we don't try to normalize endianness. + typedef typename ht::NopointerSerializer NopointerSerializer; + + // serializer: a class providing operator()(OUTPUT*, const value_type&) + // (writing value_type to OUTPUT). You can specify a + // NopointerSerializer object if appropriate (see above). + // fp: either a FILE*, OR an ostream*/subclass_of_ostream*, OR a + // pointer to a class providing size_t Write(const void*, size_t), + // which writes a buffer into a stream (which fp presumably + // owns) and returns the number of bytes successfully written. + // Note basic_ostream is not currently supported. + template + bool serialize(ValueSerializer serializer, OUTPUT* fp) { + return rep.serialize(serializer, fp); + } + + // serializer: a functor providing operator()(INPUT*, value_type*) + // (reading from INPUT and into value_type). You can specify a + // NopointerSerializer object if appropriate (see above). + // fp: either a FILE*, OR an istream*/subclass_of_istream*, OR a + // pointer to a class providing size_t Read(void*, size_t), + // which reads into a buffer from a stream (which fp presumably + // owns) and returns the number of bytes successfully read. + // Note basic_istream is not currently supported. + // NOTE: Since value_type is std::pair, ValueSerializer + // may need to do a const cast in order to fill in the key. + // NOTE: if Key or T are not POD types, the serializer MUST use + // placement-new to initialize their values, rather than a normal + // equals-assignment or similar. (The value_type* passed into the + // serializer points to garbage memory.) + template + bool unserialize(ValueSerializer serializer, INPUT* fp) { + return rep.unserialize(serializer, fp); + } + + // The four methods below are DEPRECATED. + // Use serialize() and unserialize() for new code. + template + bool write_metadata(OUTPUT *fp) { return rep.write_metadata(fp); } + + template + bool read_metadata(INPUT *fp) { return rep.read_metadata(fp); } + + template + bool write_nopointer_data(OUTPUT *fp) { return rep.write_nopointer_data(fp); } + + template + bool read_nopointer_data(INPUT *fp) { return rep.read_nopointer_data(fp); } +}; + +// We need a global swap as well +template +inline void swap(sparse_hash_map& hm1, + sparse_hash_map& hm2) { + hm1.swap(hm2); +} + +_END_GOOGLE_NAMESPACE_ + +#endif /* _SPARSE_HASH_MAP_H_ */ diff --git a/src/includes/3thparty/sparsehash/sparse_hash_set b/src/includes/3thparty/sparsehash/sparse_hash_set new file mode 100644 index 0000000..ae4a97a --- /dev/null +++ b/src/includes/3thparty/sparsehash/sparse_hash_set @@ -0,0 +1,338 @@ +// Copyright (c) 2005, Google Inc. +// All rights reserved. +// +// Redistribution and use in source and binary forms, with or without +// modification, are permitted provided that the following conditions are +// met: +// +// * Redistributions of source code must retain the above copyright +// notice, this list of conditions and the following disclaimer. +// * Redistributions in binary form must reproduce the above +// copyright notice, this list of conditions and the following disclaimer +// in the documentation and/or other materials provided with the +// distribution. +// * Neither the name of Google Inc. nor the names of its +// contributors may be used to endorse or promote products derived from +// this software without specific prior written permission. +// +// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS +// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT +// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR +// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT +// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, +// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT +// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, +// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY +// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT +// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE +// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + +// --- +// +// This is just a very thin wrapper over sparsehashtable.h, just +// like sgi stl's stl_hash_set is a very thin wrapper over +// stl_hashtable. The major thing we define is operator[], because +// we have a concept of a data_type which stl_hashtable doesn't +// (it only has a key and a value). +// +// This is more different from sparse_hash_map than you might think, +// because all iterators for sets are const (you obviously can't +// change the key, and for sets there is no value). +// +// We adhere mostly to the STL semantics for hash-map. One important +// exception is that insert() may invalidate iterators entirely -- STL +// semantics are that insert() may reorder iterators, but they all +// still refer to something valid in the hashtable. Not so for us. +// Likewise, insert() may invalidate pointers into the hashtable. +// (Whether insert invalidates iterators and pointers depends on +// whether it results in a hashtable resize). On the plus side, +// delete() doesn't invalidate iterators or pointers at all, or even +// change the ordering of elements. +// +// Here are a few "power user" tips: +// +// 1) set_deleted_key(): +// Unlike STL's hash_map, if you want to use erase() you +// *must* call set_deleted_key() after construction. +// +// 2) resize(0): +// When an item is deleted, its memory isn't freed right +// away. This allows you to iterate over a hashtable, +// and call erase(), without invalidating the iterator. +// To force the memory to be freed, call resize(0). +// For tr1 compatibility, this can also be called as rehash(0). +// +// 3) min_load_factor(0.0) +// Setting the minimum load factor to 0.0 guarantees that +// the hash table will never shrink. +// +// Roughly speaking: +// (1) dense_hash_set: fastest, uses the most memory unless entries are small +// (2) sparse_hash_set: slowest, uses the least memory +// (3) hash_set / unordered_set (STL): in the middle +// +// Typically I use sparse_hash_set when I care about space and/or when +// I need to save the hashtable on disk. I use hash_set otherwise. I +// don't personally use dense_hash_set ever; some people use it for +// small sets with lots of lookups. +// +// - dense_hash_set has, typically, about 78% memory overhead (if your +// data takes up X bytes, the hash_set uses .78X more bytes in overhead). +// - sparse_hash_set has about 4 bits overhead per entry. +// - sparse_hash_set can be 3-7 times slower than the others for lookup and, +// especially, inserts. See time_hash_map.cc for details. +// +// See /usr/(local/)?doc/sparsehash-*/sparse_hash_set.html +// for information about how to use this class. + +#ifndef _SPARSE_HASH_SET_H_ +#define _SPARSE_HASH_SET_H_ + +#include +#include // needed by stl_alloc +#include // for equal_to<> +#include // for alloc (which we don't use) +#include // for pair<> +#include +#include // IWYU pragma: export +#include HASH_FUN_H // for hash<> + +_START_GOOGLE_NAMESPACE_ + +template , // defined in sparseconfig.h + class EqualKey = std::equal_to, + class Alloc = libc_allocator_with_realloc > +class sparse_hash_set { + private: + // Apparently identity is not stl-standard, so we define our own + struct Identity { + typedef const Value& result_type; + const Value& operator()(const Value& v) const { return v; } + }; + struct SetKey { + void operator()(Value* value, const Value& new_key) const { + *value = new_key; + } + }; + + typedef sparse_hashtable ht; + ht rep; + + public: + typedef typename ht::key_type key_type; + typedef typename ht::value_type value_type; + typedef typename ht::hasher hasher; + typedef typename ht::key_equal key_equal; + typedef Alloc allocator_type; + + typedef typename ht::size_type size_type; + typedef typename ht::difference_type difference_type; + typedef typename ht::const_pointer pointer; + typedef typename ht::const_pointer const_pointer; + typedef typename ht::const_reference reference; + typedef typename ht::const_reference const_reference; + + typedef typename ht::const_iterator iterator; + typedef typename ht::const_iterator const_iterator; + typedef typename ht::const_local_iterator local_iterator; + typedef typename ht::const_local_iterator const_local_iterator; + + + // Iterator functions -- recall all iterators are const + iterator begin() const { return rep.begin(); } + iterator end() const { return rep.end(); } + + // These come from tr1's unordered_set. For us, a bucket has 0 or 1 elements. + local_iterator begin(size_type i) const { return rep.begin(i); } + local_iterator end(size_type i) const { return rep.end(i); } + + + // Accessor functions + allocator_type get_allocator() const { return rep.get_allocator(); } + hasher hash_funct() const { return rep.hash_funct(); } + hasher hash_function() const { return hash_funct(); } // tr1 name + key_equal key_eq() const { return rep.key_eq(); } + + + // Constructors + explicit sparse_hash_set(size_type expected_max_items_in_table = 0, + const hasher& hf = hasher(), + const key_equal& eql = key_equal(), + const allocator_type& alloc = allocator_type()) + : rep(expected_max_items_in_table, hf, eql, Identity(), SetKey(), alloc) { + } + + template + sparse_hash_set(InputIterator f, InputIterator l, + size_type expected_max_items_in_table = 0, + const hasher& hf = hasher(), + const key_equal& eql = key_equal(), + const allocator_type& alloc = allocator_type()) + : rep(expected_max_items_in_table, hf, eql, Identity(), SetKey(), alloc) { + rep.insert(f, l); + } + // We use the default copy constructor + // We use the default operator=() + // We use the default destructor + + void clear() { rep.clear(); } + void swap(sparse_hash_set& hs) { rep.swap(hs.rep); } + + + // Functions concerning size + size_type size() const { return rep.size(); } + size_type max_size() const { return rep.max_size(); } + bool empty() const { return rep.empty(); } + size_type bucket_count() const { return rep.bucket_count(); } + size_type max_bucket_count() const { return rep.max_bucket_count(); } + + // These are tr1 methods. bucket() is the bucket the key is or would be in. + size_type bucket_size(size_type i) const { return rep.bucket_size(i); } + size_type bucket(const key_type& key) const { return rep.bucket(key); } + float load_factor() const { + return size() * 1.0f / bucket_count(); + } + float max_load_factor() const { + float shrink, grow; + rep.get_resizing_parameters(&shrink, &grow); + return grow; + } + void max_load_factor(float new_grow) { + float shrink, grow; + rep.get_resizing_parameters(&shrink, &grow); + rep.set_resizing_parameters(shrink, new_grow); + } + // These aren't tr1 methods but perhaps ought to be. + float min_load_factor() const { + float shrink, grow; + rep.get_resizing_parameters(&shrink, &grow); + return shrink; + } + void min_load_factor(float new_shrink) { + float shrink, grow; + rep.get_resizing_parameters(&shrink, &grow); + rep.set_resizing_parameters(new_shrink, grow); + } + // Deprecated; use min_load_factor() or max_load_factor() instead. + void set_resizing_parameters(float shrink, float grow) { + rep.set_resizing_parameters(shrink, grow); + } + + void resize(size_type hint) { rep.resize(hint); } + void rehash(size_type hint) { resize(hint); } // the tr1 name + + // Lookup routines + iterator find(const key_type& key) const { return rep.find(key); } + + size_type count(const key_type& key) const { return rep.count(key); } + + std::pair equal_range(const key_type& key) const { + return rep.equal_range(key); + } + + + // Insertion routines + std::pair insert(const value_type& obj) { + std::pair p = rep.insert(obj); + return std::pair(p.first, p.second); // const to non-const + } + template void insert(InputIterator f, InputIterator l) { + rep.insert(f, l); + } + void insert(const_iterator f, const_iterator l) { + rep.insert(f, l); + } + // Required for std::insert_iterator; the passed-in iterator is ignored. + iterator insert(iterator, const value_type& obj) { + return insert(obj).first; + } + + // Deletion routines + // THESE ARE NON-STANDARD! I make you specify an "impossible" key + // value to identify deleted buckets. You can change the key as + // time goes on, or get rid of it entirely to be insert-only. + void set_deleted_key(const key_type& key) { rep.set_deleted_key(key); } + void clear_deleted_key() { rep.clear_deleted_key(); } + key_type deleted_key() const { return rep.deleted_key(); } + + // These are standard + size_type erase(const key_type& key) { return rep.erase(key); } + void erase(iterator it) { rep.erase(it); } + void erase(iterator f, iterator l) { rep.erase(f, l); } + + + // Comparison + bool operator==(const sparse_hash_set& hs) const { return rep == hs.rep; } + bool operator!=(const sparse_hash_set& hs) const { return rep != hs.rep; } + + + // I/O -- this is an add-on for writing metainformation to disk + // + // For maximum flexibility, this does not assume a particular + // file type (though it will probably be a FILE *). We just pass + // the fp through to rep. + + // If your keys and values are simple enough, you can pass this + // serializer to serialize()/unserialize(). "Simple enough" means + // value_type is a POD type that contains no pointers. Note, + // however, we don't try to normalize endianness. + typedef typename ht::NopointerSerializer NopointerSerializer; + + // serializer: a class providing operator()(OUTPUT*, const value_type&) + // (writing value_type to OUTPUT). You can specify a + // NopointerSerializer object if appropriate (see above). + // fp: either a FILE*, OR an ostream*/subclass_of_ostream*, OR a + // pointer to a class providing size_t Write(const void*, size_t), + // which writes a buffer into a stream (which fp presumably + // owns) and returns the number of bytes successfully written. + // Note basic_ostream is not currently supported. + template + bool serialize(ValueSerializer serializer, OUTPUT* fp) { + return rep.serialize(serializer, fp); + } + + // serializer: a functor providing operator()(INPUT*, value_type*) + // (reading from INPUT and into value_type). You can specify a + // NopointerSerializer object if appropriate (see above). + // fp: either a FILE*, OR an istream*/subclass_of_istream*, OR a + // pointer to a class providing size_t Read(void*, size_t), + // which reads into a buffer from a stream (which fp presumably + // owns) and returns the number of bytes successfully read. + // Note basic_istream is not currently supported. + // NOTE: Since value_type is const Key, ValueSerializer + // may need to do a const cast in order to fill in the key. + // NOTE: if Key is not a POD type, the serializer MUST use + // placement-new to initialize its value, rather than a normal + // equals-assignment or similar. (The value_type* passed into + // the serializer points to garbage memory.) + template + bool unserialize(ValueSerializer serializer, INPUT* fp) { + return rep.unserialize(serializer, fp); + } + + // The four methods below are DEPRECATED. + // Use serialize() and unserialize() for new code. + template + bool write_metadata(OUTPUT *fp) { return rep.write_metadata(fp); } + + template + bool read_metadata(INPUT *fp) { return rep.read_metadata(fp); } + + template + bool write_nopointer_data(OUTPUT *fp) { return rep.write_nopointer_data(fp); } + + template + bool read_nopointer_data(INPUT *fp) { return rep.read_nopointer_data(fp); } +}; + +template +inline void swap(sparse_hash_set& hs1, + sparse_hash_set& hs2) { + hs1.swap(hs2); +} + +_END_GOOGLE_NAMESPACE_ + +#endif /* _SPARSE_HASH_SET_H_ */ diff --git a/src/includes/3thparty/sparsehash/sparsetable b/src/includes/3thparty/sparsehash/sparsetable new file mode 100644 index 0000000..6259ebd --- /dev/null +++ b/src/includes/3thparty/sparsehash/sparsetable @@ -0,0 +1,1825 @@ +// Copyright (c) 2005, Google Inc. +// All rights reserved. +// +// Redistribution and use in source and binary forms, with or without +// modification, are permitted provided that the following conditions are +// met: +// +// * Redistributions of source code must retain the above copyright +// notice, this list of conditions and the following disclaimer. +// * Redistributions in binary form must reproduce the above +// copyright notice, this list of conditions and the following disclaimer +// in the documentation and/or other materials provided with the +// distribution. +// * Neither the name of Google Inc. nor the names of its +// contributors may be used to endorse or promote products derived from +// this software without specific prior written permission. +// +// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS +// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT +// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR +// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT +// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, +// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT +// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, +// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY +// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT +// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE +// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + +// --- +// +// +// A sparsetable is a random container that implements a sparse array, +// that is, an array that uses very little memory to store unassigned +// indices (in this case, between 1-2 bits per unassigned index). For +// instance, if you allocate an array of size 5 and assign a[2] = , then a[2] will take up a lot of memory but a[0], a[1], +// a[3], and a[4] will not. Array elements that have a value are +// called "assigned". Array elements that have no value yet, or have +// had their value cleared using erase() or clear(), are called +// "unassigned". +// +// Unassigned values seem to have the default value of T (see below). +// Nevertheless, there is a difference between an unassigned index and +// one explicitly assigned the value of T(). The latter is considered +// assigned. +// +// Access to an array element is constant time, as is insertion and +// deletion. Insertion and deletion may be fairly slow, however: +// because of this container's memory economy, each insert and delete +// causes a memory reallocation. +// +// NOTE: You should not test(), get(), or set() any index that is +// greater than sparsetable.size(). If you need to do that, call +// resize() first. +// +// --- Template parameters +// PARAMETER DESCRIPTION DEFAULT +// T The value of the array: the type of -- +// object that is stored in the array. +// +// GROUP_SIZE How large each "group" in the table 48 +// is (see below). Larger values use +// a little less memory but cause most +// operations to be a little slower +// +// Alloc: Allocator to use to allocate memory. libc_allocator_with_realloc +// +// --- Model of +// Random Access Container +// +// --- Type requirements +// T must be Copy Constructible. It need not be Assignable. +// +// --- Public base classes +// None. +// +// --- Members +// Type members +// +// MEMBER WHERE DEFINED DESCRIPTION +// value_type container The type of object, T, stored in the array +// allocator_type container Allocator to use +// pointer container Pointer to p +// const_pointer container Const pointer to p +// reference container Reference to t +// const_reference container Const reference to t +// size_type container An unsigned integral type +// difference_type container A signed integral type +// iterator [*] container Iterator used to iterate over a sparsetable +// const_iterator container Const iterator used to iterate over a table +// reverse_iterator reversible Iterator used to iterate backwards over +// container a sparsetable +// const_reverse_iterator reversible container Guess +// nonempty_iterator [+] sparsetable Iterates over assigned +// array elements only +// const_nonempty_iterator sparsetable Iterates over assigned +// array elements only +// reverse_nonempty_iterator sparsetable Iterates backwards over +// assigned array elements only +// const_reverse_nonempty_iterator sparsetable Iterates backwards over +// assigned array elements only +// +// [*] All iterators are const in a sparsetable (though nonempty_iterators +// may not be). Use get() and set() to assign values, not iterators. +// +// [+] iterators are random-access iterators. nonempty_iterators are +// bidirectional iterators. + +// Iterator members +// MEMBER WHERE DEFINED DESCRIPTION +// +// iterator begin() container An iterator to the beginning of the table +// iterator end() container An iterator to the end of the table +// const_iterator container A const_iterator pointing to the +// begin() const beginning of a sparsetable +// const_iterator container A const_iterator pointing to the +// end() const end of a sparsetable +// +// reverse_iterator reversable Points to beginning of a reversed +// rbegin() container sparsetable +// reverse_iterator reversable Points to end of a reversed table +// rend() container +// const_reverse_iterator reversable Points to beginning of a +// rbegin() const container reversed sparsetable +// const_reverse_iterator reversable Points to end of a reversed table +// rend() const container +// +// nonempty_iterator sparsetable Points to first assigned element +// begin() of a sparsetable +// nonempty_iterator sparsetable Points past last assigned element +// end() of a sparsetable +// const_nonempty_iterator sparsetable Points to first assigned element +// begin() const of a sparsetable +// const_nonempty_iterator sparsetable Points past last assigned element +// end() const of a sparsetable +// +// reverse_nonempty_iterator sparsetable Points to first assigned element +// begin() of a reversed sparsetable +// reverse_nonempty_iterator sparsetable Points past last assigned element +// end() of a reversed sparsetable +// const_reverse_nonempty_iterator sparsetable Points to first assigned +// begin() const elt of a reversed sparsetable +// const_reverse_nonempty_iterator sparsetable Points past last assigned +// end() const elt of a reversed sparsetable +// +// +// Other members +// MEMBER WHERE DEFINED DESCRIPTION +// sparsetable() sparsetable A table of size 0; must resize() +// before using. +// sparsetable(size_type size) sparsetable A table of size size. All +// indices are unassigned. +// sparsetable( +// const sparsetable &tbl) sparsetable Copy constructor +// ~sparsetable() sparsetable The destructor +// sparsetable &operator=( sparsetable The assignment operator +// const sparsetable &tbl) +// +// void resize(size_type size) sparsetable Grow or shrink a table to +// have size indices [*] +// +// void swap(sparsetable &x) sparsetable Swap two sparsetables +// void swap(sparsetable &x, sparsetable Swap two sparsetables +// sparsetable &y) (global, not member, function) +// +// size_type size() const sparsetable Number of "buckets" in the table +// size_type max_size() const sparsetable Max allowed size of a sparsetable +// bool empty() const sparsetable true if size() == 0 +// size_type num_nonempty() const sparsetable Number of assigned "buckets" +// +// const_reference get( sparsetable Value at index i, or default +// size_type i) const value if i is unassigned +// const_reference operator[]( sparsetable Identical to get(i) [+] +// difference_type i) const +// reference set(size_type i, sparsetable Set element at index i to +// const_reference val) be a copy of val +// bool test(size_type i) sparsetable True if element at index i +// const has been assigned to +// bool test(iterator pos) sparsetable True if element pointed to +// const by pos has been assigned to +// void erase(iterator pos) sparsetable Set element pointed to by +// pos to be unassigned [!] +// void erase(size_type i) sparsetable Set element i to be unassigned +// void erase(iterator start, sparsetable Erases all elements between +// iterator end) start and end +// void clear() sparsetable Erases all elements in the table +// +// I/O versions exist for both FILE* and for File* (Google2-style files): +// bool write_metadata(FILE *fp) sparsetable Writes a sparsetable to the +// bool write_metadata(File *fp) given file. true if write +// completes successfully +// bool read_metadata(FILE *fp) sparsetable Replaces sparsetable with +// bool read_metadata(File *fp) version read from fp. true +// if read completes sucessfully +// bool write_nopointer_data(FILE *fp) Read/write the data stored in +// bool read_nopointer_data(FILE*fp) the table, if it's simple +// +// bool operator==( forward Tests two tables for equality. +// const sparsetable &t1, container This is a global function, +// const sparsetable &t2) not a member function. +// bool operator<( forward Lexicographical comparison. +// const sparsetable &t1, container This is a global function, +// const sparsetable &t2) not a member function. +// +// [*] If you shrink a sparsetable using resize(), assigned elements +// past the end of the table are removed using erase(). If you grow +// a sparsetable, new unassigned indices are created. +// +// [+] Note that operator[] returns a const reference. You must use +// set() to change the value of a table element. +// +// [!] Unassignment also calls the destructor. +// +// Iterators are invalidated whenever an item is inserted or +// deleted (ie set() or erase() is used) or when the size of +// the table changes (ie resize() or clear() is used). +// +// See doc/sparsetable.html for more information about how to use this class. + +// Note: this uses STL style for naming, rather than Google naming. +// That's because this is an STL-y container + +#ifndef UTIL_GTL_SPARSETABLE_H_ +#define UTIL_GTL_SPARSETABLE_H_ + +#include +#include // for malloc/free +#include // to read/write tables +#include // for memcpy +#ifdef HAVE_STDINT_H +#include // the normal place uint16_t is defined +#endif +#ifdef HAVE_SYS_TYPES_H +#include // the normal place u_int16_t is defined +#endif +#ifdef HAVE_INTTYPES_H +#include // a third place for uint16_t or u_int16_t +#endif +#include // for bounds checking +#include // to define reverse_iterator for me +#include // equal, lexicographical_compare, swap,... +#include // uninitialized_copy, uninitialized_fill +#include // a sparsetable is a vector of groups +#include +#include +#include + +// A lot of work to get a type that's guaranteed to be 16 bits... +#ifndef HAVE_U_INT16_T +# if defined HAVE_UINT16_T + typedef uint16_t u_int16_t; // true on solaris, possibly other C99 libc's +# elif defined HAVE___UINT16 + typedef __int16 int16_t; // true on vc++7 + typedef unsigned __int16 u_int16_t; +# else + // Cannot find a 16-bit integer type. Hoping for the best with "short"... + typedef short int int16_t; + typedef unsigned short int u_int16_t; +# endif +#endif + +_START_GOOGLE_NAMESPACE_ + +namespace base { // just to make google->opensource transition easier +using GOOGLE_NAMESPACE::true_type; +using GOOGLE_NAMESPACE::false_type; +using GOOGLE_NAMESPACE::integral_constant; +using GOOGLE_NAMESPACE::has_trivial_copy; +using GOOGLE_NAMESPACE::has_trivial_destructor; +using GOOGLE_NAMESPACE::is_same; +} + + +// The smaller this is, the faster lookup is (because the group bitmap is +// smaller) and the faster insert is, because there's less to move. +// On the other hand, there are more groups. Since group::size_type is +// a short, this number should be of the form 32*x + 16 to avoid waste. +static const u_int16_t DEFAULT_SPARSEGROUP_SIZE = 48; // fits in 1.5 words + + +// Our iterator as simple as iterators can be: basically it's just +// the index into our table. Dereference, the only complicated +// thing, we punt to the table class. This just goes to show how +// much machinery STL requires to do even the most trivial tasks. +// +// A NOTE ON ASSIGNING: +// A sparse table does not actually allocate memory for entries +// that are not filled. Because of this, it becomes complicated +// to have a non-const iterator: we don't know, if the iterator points +// to a not-filled bucket, whether you plan to fill it with something +// or whether you plan to read its value (in which case you'll get +// the default bucket value). Therefore, while we can define const +// operations in a pretty 'normal' way, for non-const operations, we +// define something that returns a helper object with operator= and +// operator& that allocate a bucket lazily. We use this for table[] +// and also for regular table iterators. + +template +class table_element_adaptor { + public: + typedef typename tabletype::value_type value_type; + typedef typename tabletype::size_type size_type; + typedef typename tabletype::reference reference; + typedef typename tabletype::pointer pointer; + + table_element_adaptor(tabletype *tbl, size_type p) + : table(tbl), pos(p) { } + table_element_adaptor& operator= (const value_type &val) { + table->set(pos, val); + return *this; + } + operator value_type() { return table->get(pos); } // we look like a value + pointer operator& () { return &table->mutating_get(pos); } + + private: + tabletype* table; + size_type pos; +}; + +// Our iterator as simple as iterators can be: basically it's just +// the index into our table. Dereference, the only complicated +// thing, we punt to the table class. This just goes to show how +// much machinery STL requires to do even the most trivial tasks. +// +// By templatizing over tabletype, we have one iterator type which +// we can use for both sparsetables and sparsebins. In fact it +// works on any class that allows size() and operator[] (eg vector), +// as long as it does the standard STL typedefs too (eg value_type). + +template +class table_iterator { + public: + typedef table_iterator iterator; + + typedef std::random_access_iterator_tag iterator_category; + typedef typename tabletype::value_type value_type; + typedef typename tabletype::difference_type difference_type; + typedef typename tabletype::size_type size_type; + typedef table_element_adaptor reference; + typedef table_element_adaptor* pointer; + + // The "real" constructor + table_iterator(tabletype *tbl, size_type p) + : table(tbl), pos(p) { } + // The default constructor, used when I define vars of type table::iterator + table_iterator() : table(NULL), pos(0) { } + // The copy constructor, for when I say table::iterator foo = tbl.begin() + // The default destructor is fine; we don't define one + // The default operator= is fine; we don't define one + + // The main thing our iterator does is dereference. If the table entry + // we point to is empty, we return the default value type. + // This is the big different function from the const iterator. + reference operator*() { + return table_element_adaptor(table, pos); + } + pointer operator->() { return &(operator*()); } + + // Helper function to assert things are ok; eg pos is still in range + void check() const { + assert(table); + assert(pos <= table->size()); + } + + // Arithmetic: we just do arithmetic on pos. We don't even need to + // do bounds checking, since STL doesn't consider that its job. :-) + iterator& operator+=(size_type t) { pos += t; check(); return *this; } + iterator& operator-=(size_type t) { pos -= t; check(); return *this; } + iterator& operator++() { ++pos; check(); return *this; } + iterator& operator--() { --pos; check(); return *this; } + iterator operator++(int) { iterator tmp(*this); // for x++ + ++pos; check(); return tmp; } + iterator operator--(int) { iterator tmp(*this); // for x-- + --pos; check(); return tmp; } + iterator operator+(difference_type i) const { iterator tmp(*this); + tmp += i; return tmp; } + iterator operator-(difference_type i) const { iterator tmp(*this); + tmp -= i; return tmp; } + difference_type operator-(iterator it) const { // for "x = it2 - it" + assert(table == it.table); + return pos - it.pos; + } + reference operator[](difference_type n) const { + return *(*this + n); // simple though not totally efficient + } + + // Comparisons. + bool operator==(const iterator& it) const { + return table == it.table && pos == it.pos; + } + bool operator<(const iterator& it) const { + assert(table == it.table); // life is bad bad bad otherwise + return pos < it.pos; + } + bool operator!=(const iterator& it) const { return !(*this == it); } + bool operator<=(const iterator& it) const { return !(it < *this); } + bool operator>(const iterator& it) const { return it < *this; } + bool operator>=(const iterator& it) const { return !(*this < it); } + + // Here's the info we actually need to be an iterator + tabletype *table; // so we can dereference and bounds-check + size_type pos; // index into the table +}; + +// support for "3 + iterator" has to be defined outside the class, alas +template +table_iterator operator+(typename table_iterator::difference_type i, + table_iterator it) { + return it + i; // so people can say it2 = 3 + it +} + +template +class const_table_iterator { + public: + typedef table_iterator iterator; + typedef const_table_iterator const_iterator; + + typedef std::random_access_iterator_tag iterator_category; + typedef typename tabletype::value_type value_type; + typedef typename tabletype::difference_type difference_type; + typedef typename tabletype::size_type size_type; + typedef typename tabletype::const_reference reference; // we're const-only + typedef typename tabletype::const_pointer pointer; + + // The "real" constructor + const_table_iterator(const tabletype *tbl, size_type p) + : table(tbl), pos(p) { } + // The default constructor, used when I define vars of type table::iterator + const_table_iterator() : table(NULL), pos(0) { } + // The copy constructor, for when I say table::iterator foo = tbl.begin() + // Also converts normal iterators to const iterators + const_table_iterator(const iterator &from) + : table(from.table), pos(from.pos) { } + // The default destructor is fine; we don't define one + // The default operator= is fine; we don't define one + + // The main thing our iterator does is dereference. If the table entry + // we point to is empty, we return the default value type. + reference operator*() const { return (*table)[pos]; } + pointer operator->() const { return &(operator*()); } + + // Helper function to assert things are ok; eg pos is still in range + void check() const { + assert(table); + assert(pos <= table->size()); + } + + // Arithmetic: we just do arithmetic on pos. We don't even need to + // do bounds checking, since STL doesn't consider that its job. :-) + const_iterator& operator+=(size_type t) { pos += t; check(); return *this; } + const_iterator& operator-=(size_type t) { pos -= t; check(); return *this; } + const_iterator& operator++() { ++pos; check(); return *this; } + const_iterator& operator--() { --pos; check(); return *this; } + const_iterator operator++(int) { const_iterator tmp(*this); // for x++ + ++pos; check(); return tmp; } + const_iterator operator--(int) { const_iterator tmp(*this); // for x-- + --pos; check(); return tmp; } + const_iterator operator+(difference_type i) const { const_iterator tmp(*this); + tmp += i; return tmp; } + const_iterator operator-(difference_type i) const { const_iterator tmp(*this); + tmp -= i; return tmp; } + difference_type operator-(const_iterator it) const { // for "x = it2 - it" + assert(table == it.table); + return pos - it.pos; + } + reference operator[](difference_type n) const { + return *(*this + n); // simple though not totally efficient + } + + // Comparisons. + bool operator==(const const_iterator& it) const { + return table == it.table && pos == it.pos; + } + bool operator<(const const_iterator& it) const { + assert(table == it.table); // life is bad bad bad otherwise + return pos < it.pos; + } + bool operator!=(const const_iterator& it) const { return !(*this == it); } + bool operator<=(const const_iterator& it) const { return !(it < *this); } + bool operator>(const const_iterator& it) const { return it < *this; } + bool operator>=(const const_iterator& it) const { return !(*this < it); } + + // Here's the info we actually need to be an iterator + const tabletype *table; // so we can dereference and bounds-check + size_type pos; // index into the table +}; + +// support for "3 + iterator" has to be defined outside the class, alas +template +const_table_iterator operator+(typename + const_table_iterator::difference_type i, + const_table_iterator it) { + return it + i; // so people can say it2 = 3 + it +} + + +// --------------------------------------------------------------------------- + + +/* +// This is a 2-D iterator. You specify a begin and end over a list +// of *containers*. We iterate over each container by iterating over +// it. It's actually simple: +// VECTOR.begin() VECTOR[0].begin() --------> VECTOR[0].end() ---, +// | ________________________________________________/ +// | \_> VECTOR[1].begin() --------> VECTOR[1].end() -, +// | ___________________________________________________/ +// v \_> ...... +// VECTOR.end() +// +// It's impossible to do random access on one of these things in constant +// time, so it's just a bidirectional iterator. +// +// Unfortunately, because we need to use this for a non-empty iterator, +// we use nonempty_begin() and nonempty_end() instead of begin() and end() +// (though only going across, not down). +*/ + +#define TWOD_BEGIN_ nonempty_begin +#define TWOD_END_ nonempty_end +#define TWOD_ITER_ nonempty_iterator +#define TWOD_CONST_ITER_ const_nonempty_iterator + +template +class two_d_iterator { + public: + typedef two_d_iterator iterator; + + typedef std::bidirectional_iterator_tag iterator_category; + // apparently some versions of VC++ have trouble with two ::'s in a typename + typedef typename containertype::value_type _tmp_vt; + typedef typename _tmp_vt::value_type value_type; + typedef typename _tmp_vt::difference_type difference_type; + typedef typename _tmp_vt::reference reference; + typedef typename _tmp_vt::pointer pointer; + + // The "real" constructor. begin and end specify how many rows we have + // (in the diagram above); we always iterate over each row completely. + two_d_iterator(typename containertype::iterator begin, + typename containertype::iterator end, + typename containertype::iterator curr) + : row_begin(begin), row_end(end), row_current(curr), col_current() { + if ( row_current != row_end ) { + col_current = row_current->TWOD_BEGIN_(); + advance_past_end(); // in case cur->begin() == cur->end() + } + } + // If you want to start at an arbitrary place, you can, I guess + two_d_iterator(typename containertype::iterator begin, + typename containertype::iterator end, + typename containertype::iterator curr, + typename containertype::value_type::TWOD_ITER_ col) + : row_begin(begin), row_end(end), row_current(curr), col_current(col) { + advance_past_end(); // in case cur->begin() == cur->end() + } + // The default constructor, used when I define vars of type table::iterator + two_d_iterator() : row_begin(), row_end(), row_current(), col_current() { } + // The default destructor is fine; we don't define one + // The default operator= is fine; we don't define one + + // Happy dereferencer + reference operator*() const { return *col_current; } + pointer operator->() const { return &(operator*()); } + + // Arithmetic: we just do arithmetic on pos. We don't even need to + // do bounds checking, since STL doesn't consider that its job. :-) + // NOTE: this is not amortized constant time! What do we do about it? + void advance_past_end() { // used when col_current points to end() + while ( col_current == row_current->TWOD_END_() ) { // end of current row + ++row_current; // go to beginning of next + if ( row_current != row_end ) // col is irrelevant at end + col_current = row_current->TWOD_BEGIN_(); + else + break; // don't go past row_end + } + } + + iterator& operator++() { + assert(row_current != row_end); // how to ++ from there? + ++col_current; + advance_past_end(); // in case col_current is at end() + return *this; + } + iterator& operator--() { + while ( row_current == row_end || + col_current == row_current->TWOD_BEGIN_() ) { + assert(row_current != row_begin); + --row_current; + col_current = row_current->TWOD_END_(); // this is 1 too far + } + --col_current; + return *this; + } + iterator operator++(int) { iterator tmp(*this); ++*this; return tmp; } + iterator operator--(int) { iterator tmp(*this); --*this; return tmp; } + + + // Comparisons. + bool operator==(const iterator& it) const { + return ( row_begin == it.row_begin && + row_end == it.row_end && + row_current == it.row_current && + (row_current == row_end || col_current == it.col_current) ); + } + bool operator!=(const iterator& it) const { return !(*this == it); } + + + // Here's the info we actually need to be an iterator + // These need to be public so we convert from iterator to const_iterator + typename containertype::iterator row_begin, row_end, row_current; + typename containertype::value_type::TWOD_ITER_ col_current; +}; + +// The same thing again, but this time const. :-( +template +class const_two_d_iterator { + public: + typedef const_two_d_iterator iterator; + + typedef std::bidirectional_iterator_tag iterator_category; + // apparently some versions of VC++ have trouble with two ::'s in a typename + typedef typename containertype::value_type _tmp_vt; + typedef typename _tmp_vt::value_type value_type; + typedef typename _tmp_vt::difference_type difference_type; + typedef typename _tmp_vt::const_reference reference; + typedef typename _tmp_vt::const_pointer pointer; + + const_two_d_iterator(typename containertype::const_iterator begin, + typename containertype::const_iterator end, + typename containertype::const_iterator curr) + : row_begin(begin), row_end(end), row_current(curr), col_current() { + if ( curr != end ) { + col_current = curr->TWOD_BEGIN_(); + advance_past_end(); // in case cur->begin() == cur->end() + } + } + const_two_d_iterator(typename containertype::const_iterator begin, + typename containertype::const_iterator end, + typename containertype::const_iterator curr, + typename containertype::value_type::TWOD_CONST_ITER_ col) + : row_begin(begin), row_end(end), row_current(curr), col_current(col) { + advance_past_end(); // in case cur->begin() == cur->end() + } + const_two_d_iterator() + : row_begin(), row_end(), row_current(), col_current() { + } + // Need this explicitly so we can convert normal iterators to const iterators + const_two_d_iterator(const two_d_iterator& it) : + row_begin(it.row_begin), row_end(it.row_end), row_current(it.row_current), + col_current(it.col_current) { } + + typename containertype::const_iterator row_begin, row_end, row_current; + typename containertype::value_type::TWOD_CONST_ITER_ col_current; + + + // EVERYTHING FROM HERE DOWN IS THE SAME AS THE NON-CONST ITERATOR + reference operator*() const { return *col_current; } + pointer operator->() const { return &(operator*()); } + + void advance_past_end() { // used when col_current points to end() + while ( col_current == row_current->TWOD_END_() ) { // end of current row + ++row_current; // go to beginning of next + if ( row_current != row_end ) // col is irrelevant at end + col_current = row_current->TWOD_BEGIN_(); + else + break; // don't go past row_end + } + } + iterator& operator++() { + assert(row_current != row_end); // how to ++ from there? + ++col_current; + advance_past_end(); // in case col_current is at end() + return *this; + } + iterator& operator--() { + while ( row_current == row_end || + col_current == row_current->TWOD_BEGIN_() ) { + assert(row_current != row_begin); + --row_current; + col_current = row_current->TWOD_END_(); // this is 1 too far + } + --col_current; + return *this; + } + iterator operator++(int) { iterator tmp(*this); ++*this; return tmp; } + iterator operator--(int) { iterator tmp(*this); --*this; return tmp; } + + bool operator==(const iterator& it) const { + return ( row_begin == it.row_begin && + row_end == it.row_end && + row_current == it.row_current && + (row_current == row_end || col_current == it.col_current) ); + } + bool operator!=(const iterator& it) const { return !(*this == it); } +}; + +// We provide yet another version, to be as frugal with memory as +// possible. This one frees each block of memory as it finishes +// iterating over it. By the end, the entire table is freed. +// For understandable reasons, you can only iterate over it once, +// which is why it's an input iterator +template +class destructive_two_d_iterator { + public: + typedef destructive_two_d_iterator iterator; + + typedef std::input_iterator_tag iterator_category; + // apparently some versions of VC++ have trouble with two ::'s in a typename + typedef typename containertype::value_type _tmp_vt; + typedef typename _tmp_vt::value_type value_type; + typedef typename _tmp_vt::difference_type difference_type; + typedef typename _tmp_vt::reference reference; + typedef typename _tmp_vt::pointer pointer; + + destructive_two_d_iterator(typename containertype::iterator begin, + typename containertype::iterator end, + typename containertype::iterator curr) + : row_begin(begin), row_end(end), row_current(curr), col_current() { + if ( curr != end ) { + col_current = curr->TWOD_BEGIN_(); + advance_past_end(); // in case cur->begin() == cur->end() + } + } + destructive_two_d_iterator(typename containertype::iterator begin, + typename containertype::iterator end, + typename containertype::iterator curr, + typename containertype::value_type::TWOD_ITER_ col) + : row_begin(begin), row_end(end), row_current(curr), col_current(col) { + advance_past_end(); // in case cur->begin() == cur->end() + } + destructive_two_d_iterator() + : row_begin(), row_end(), row_current(), col_current() { + } + + typename containertype::iterator row_begin, row_end, row_current; + typename containertype::value_type::TWOD_ITER_ col_current; + + // This is the part that destroys + void advance_past_end() { // used when col_current points to end() + while ( col_current == row_current->TWOD_END_() ) { // end of current row + row_current->clear(); // the destructive part + // It would be nice if we could decrement sparsetable->num_buckets here + ++row_current; // go to beginning of next + if ( row_current != row_end ) // col is irrelevant at end + col_current = row_current->TWOD_BEGIN_(); + else + break; // don't go past row_end + } + } + + // EVERYTHING FROM HERE DOWN IS THE SAME AS THE REGULAR ITERATOR + reference operator*() const { return *col_current; } + pointer operator->() const { return &(operator*()); } + + iterator& operator++() { + assert(row_current != row_end); // how to ++ from there? + ++col_current; + advance_past_end(); // in case col_current is at end() + return *this; + } + iterator operator++(int) { iterator tmp(*this); ++*this; return tmp; } + + bool operator==(const iterator& it) const { + return ( row_begin == it.row_begin && + row_end == it.row_end && + row_current == it.row_current && + (row_current == row_end || col_current == it.col_current) ); + } + bool operator!=(const iterator& it) const { return !(*this == it); } +}; + +#undef TWOD_BEGIN_ +#undef TWOD_END_ +#undef TWOD_ITER_ +#undef TWOD_CONST_ITER_ + + + + +// SPARSE-TABLE +// ------------ +// The idea is that a table with (logically) t buckets is divided +// into t/M *groups* of M buckets each. (M is a constant set in +// GROUP_SIZE for efficiency.) Each group is stored sparsely. +// Thus, inserting into the table causes some array to grow, which is +// slow but still constant time. Lookup involves doing a +// logical-position-to-sparse-position lookup, which is also slow but +// constant time. The larger M is, the slower these operations are +// but the less overhead (slightly). +// +// To store the sparse array, we store a bitmap B, where B[i] = 1 iff +// bucket i is non-empty. Then to look up bucket i we really look up +// array[# of 1s before i in B]. This is constant time for fixed M. +// +// Terminology: the position of an item in the overall table (from +// 1 .. t) is called its "location." The logical position in a group +// (from 1 .. M ) is called its "position." The actual location in +// the array (from 1 .. # of non-empty buckets in the group) is +// called its "offset." + +template +class sparsegroup { + private: + typedef typename Alloc::template rebind::other value_alloc_type; + + public: + // Basic types + typedef T value_type; + typedef Alloc allocator_type; + typedef typename value_alloc_type::reference reference; + typedef typename value_alloc_type::const_reference const_reference; + typedef typename value_alloc_type::pointer pointer; + typedef typename value_alloc_type::const_pointer const_pointer; + + typedef table_iterator > iterator; + typedef const_table_iterator > + const_iterator; + typedef table_element_adaptor > + element_adaptor; + typedef u_int16_t size_type; // max # of buckets + typedef int16_t difference_type; + typedef std::reverse_iterator const_reverse_iterator; + typedef std::reverse_iterator reverse_iterator; // from iterator.h + + // These are our special iterators, that go over non-empty buckets in a + // group. These aren't const-only because you can change non-empty bcks. + typedef pointer nonempty_iterator; + typedef const_pointer const_nonempty_iterator; + typedef std::reverse_iterator reverse_nonempty_iterator; + typedef std::reverse_iterator const_reverse_nonempty_iterator; + + // Iterator functions + iterator begin() { return iterator(this, 0); } + const_iterator begin() const { return const_iterator(this, 0); } + iterator end() { return iterator(this, size()); } + const_iterator end() const { return const_iterator(this, size()); } + reverse_iterator rbegin() { return reverse_iterator(end()); } + const_reverse_iterator rbegin() const { return const_reverse_iterator(end()); } + reverse_iterator rend() { return reverse_iterator(begin()); } + const_reverse_iterator rend() const { return const_reverse_iterator(begin()); } + + // We'll have versions for our special non-empty iterator too + nonempty_iterator nonempty_begin() { return group; } + const_nonempty_iterator nonempty_begin() const { return group; } + nonempty_iterator nonempty_end() { + return group + settings.num_buckets; + } + const_nonempty_iterator nonempty_end() const { + return group + settings.num_buckets; + } + reverse_nonempty_iterator nonempty_rbegin() { + return reverse_nonempty_iterator(nonempty_end()); + } + const_reverse_nonempty_iterator nonempty_rbegin() const { + return const_reverse_nonempty_iterator(nonempty_end()); + } + reverse_nonempty_iterator nonempty_rend() { + return reverse_nonempty_iterator(nonempty_begin()); + } + const_reverse_nonempty_iterator nonempty_rend() const { + return const_reverse_nonempty_iterator(nonempty_begin()); + } + + + // This gives us the "default" value to return for an empty bucket. + // We just use the default constructor on T, the template type + const_reference default_value() const { + static value_type defaultval = value_type(); + return defaultval; + } + + + private: + // We need to do all this bit manipulation, of course. ick + static size_type charbit(size_type i) { return i >> 3; } + static size_type modbit(size_type i) { return 1 << (i&7); } + int bmtest(size_type i) const { return bitmap[charbit(i)] & modbit(i); } + void bmset(size_type i) { bitmap[charbit(i)] |= modbit(i); } + void bmclear(size_type i) { bitmap[charbit(i)] &= ~modbit(i); } + + pointer allocate_group(size_type n) { + pointer retval = settings.allocate(n); + if (retval == NULL) { + // We really should use PRIuS here, but I don't want to have to add + // a whole new configure option, with concomitant macro namespace + // pollution, just to print this (unlikely) error message. So I cast. + fprintf(stderr, "sparsehash FATAL ERROR: failed to allocate %lu groups\n", + static_cast(n)); + exit(1); + } + return retval; + } + + void free_group() { + if (!group) return; + pointer end_it = group + settings.num_buckets; + for (pointer p = group; p != end_it; ++p) + p->~value_type(); + settings.deallocate(group, settings.num_buckets); + group = NULL; + } + + static size_type bits_in_char(unsigned char c) { + // We could make these ints. The tradeoff is size (eg does it overwhelm + // the cache?) vs efficiency in referencing sub-word-sized array elements. + static const char bits_in[256] = { + 0, 1, 1, 2, 1, 2, 2, 3, 1, 2, 2, 3, 2, 3, 3, 4, + 1, 2, 2, 3, 2, 3, 3, 4, 2, 3, 3, 4, 3, 4, 4, 5, + 1, 2, 2, 3, 2, 3, 3, 4, 2, 3, 3, 4, 3, 4, 4, 5, + 2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6, + 1, 2, 2, 3, 2, 3, 3, 4, 2, 3, 3, 4, 3, 4, 4, 5, + 2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6, + 2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6, + 3, 4, 4, 5, 4, 5, 5, 6, 4, 5, 5, 6, 5, 6, 6, 7, + 1, 2, 2, 3, 2, 3, 3, 4, 2, 3, 3, 4, 3, 4, 4, 5, + 2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6, + 2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6, + 3, 4, 4, 5, 4, 5, 5, 6, 4, 5, 5, 6, 5, 6, 6, 7, + 2, 3, 3, 4, 3, 4, 4, 5, 3, 4, 4, 5, 4, 5, 5, 6, + 3, 4, 4, 5, 4, 5, 5, 6, 4, 5, 5, 6, 5, 6, 6, 7, + 3, 4, 4, 5, 4, 5, 5, 6, 4, 5, 5, 6, 5, 6, 6, 7, + 4, 5, 5, 6, 5, 6, 6, 7, 5, 6, 6, 7, 6, 7, 7, 8, + }; + return bits_in[c]; + } + + public: // get_iter() in sparsetable needs it + // We need a small function that tells us how many set bits there are + // in positions 0..i-1 of the bitmap. It uses a big table. + // We make it static so templates don't allocate lots of these tables. + // There are lots of ways to do this calculation (called 'popcount'). + // The 8-bit table lookup is one of the fastest, though this + // implementation suffers from not doing any loop unrolling. See, eg, + // http://www.dalkescientific.com/writings/diary/archive/2008/07/03/hakmem_and_other_popcounts.html + // http://gurmeetsingh.wordpress.com/2008/08/05/fast-bit-counting-routines/ + static size_type pos_to_offset(const unsigned char *bm, size_type pos) { + size_type retval = 0; + + // [Note: condition pos > 8 is an optimization; convince yourself we + // give exactly the same result as if we had pos >= 8 here instead.] + for ( ; pos > 8; pos -= 8 ) // bm[0..pos/8-1] + retval += bits_in_char(*bm++); // chars we want *all* bits in + return retval + bits_in_char(*bm & ((1 << pos)-1)); // char including pos + } + + size_type pos_to_offset(size_type pos) const { // not static but still const + return pos_to_offset(bitmap, pos); + } + + // Returns the (logical) position in the bm[] array, i, such that + // bm[i] is the offset-th set bit in the array. It is the inverse + // of pos_to_offset. get_pos() uses this function to find the index + // of an nonempty_iterator in the table. Bit-twiddling from + // http://hackersdelight.org/basics.pdf + static size_type offset_to_pos(const unsigned char *bm, size_type offset) { + size_type retval = 0; + // This is sizeof(this->bitmap). + const size_type group_size = (GROUP_SIZE-1) / 8 + 1; + for (size_type i = 0; i < group_size; i++) { // forward scan + const size_type pop_count = bits_in_char(*bm); + if (pop_count > offset) { + unsigned char last_bm = *bm; + for (; offset > 0; offset--) { + last_bm &= (last_bm-1); // remove right-most set bit + } + // Clear all bits to the left of the rightmost bit (the &), + // and then clear the rightmost bit but set all bits to the + // right of it (the -1). + last_bm = (last_bm & -last_bm) - 1; + retval += bits_in_char(last_bm); + return retval; + } + offset -= pop_count; + retval += 8; + bm++; + } + return retval; + } + + size_type offset_to_pos(size_type offset) const { + return offset_to_pos(bitmap, offset); + } + + + public: + // Constructors -- default and copy -- and destructor + explicit sparsegroup(allocator_type& a) : + group(0), settings(alloc_impl(a)) { + memset(bitmap, 0, sizeof(bitmap)); + } + sparsegroup(const sparsegroup& x) : group(0), settings(x.settings) { + if ( settings.num_buckets ) { + group = allocate_group(x.settings.num_buckets); + std::uninitialized_copy(x.group, x.group + x.settings.num_buckets, group); + } + memcpy(bitmap, x.bitmap, sizeof(bitmap)); + } + ~sparsegroup() { free_group(); } + + // Operator= is just like the copy constructor, I guess + // TODO(austern): Make this exception safe. Handle exceptions in value_type's + // copy constructor. + sparsegroup &operator=(const sparsegroup& x) { + if ( &x == this ) return *this; // x = x + if ( x.settings.num_buckets == 0 ) { + free_group(); + } else { + pointer p = allocate_group(x.settings.num_buckets); + std::uninitialized_copy(x.group, x.group + x.settings.num_buckets, p); + free_group(); + group = p; + } + memcpy(bitmap, x.bitmap, sizeof(bitmap)); + settings.num_buckets = x.settings.num_buckets; + return *this; + } + + // Many STL algorithms use swap instead of copy constructors + void swap(sparsegroup& x) { + std::swap(group, x.group); // defined in + for ( int i = 0; i < sizeof(bitmap) / sizeof(*bitmap); ++i ) + std::swap(bitmap[i], x.bitmap[i]); // swap not defined on arrays + std::swap(settings.num_buckets, x.settings.num_buckets); + // we purposefully don't swap the allocator, which may not be swap-able + } + + // It's always nice to be able to clear a table without deallocating it + void clear() { + free_group(); + memset(bitmap, 0, sizeof(bitmap)); + settings.num_buckets = 0; + } + + // Functions that tell you about size. Alas, these aren't so useful + // because our table is always fixed size. + size_type size() const { return GROUP_SIZE; } + size_type max_size() const { return GROUP_SIZE; } + bool empty() const { return false; } + // We also may want to know how many *used* buckets there are + size_type num_nonempty() const { return settings.num_buckets; } + + + // get()/set() are explicitly const/non-const. You can use [] if + // you want something that can be either (potentially more expensive). + const_reference get(size_type i) const { + if ( bmtest(i) ) // bucket i is occupied + return group[pos_to_offset(bitmap, i)]; + else + return default_value(); // return the default reference + } + + // TODO(csilvers): make protected + friend + // This is used by sparse_hashtable to get an element from the table + // when we know it exists. + const_reference unsafe_get(size_type i) const { + assert(bmtest(i)); + return group[pos_to_offset(bitmap, i)]; + } + + // TODO(csilvers): make protected + friend + reference mutating_get(size_type i) { // fills bucket i before getting + if ( !bmtest(i) ) + set(i, default_value()); + return group[pos_to_offset(bitmap, i)]; + } + + // Syntactic sugar. It's easy to return a const reference. To + // return a non-const reference, we need to use the assigner adaptor. + const_reference operator[](size_type i) const { + return get(i); + } + + element_adaptor operator[](size_type i) { + return element_adaptor(this, i); + } + + private: + // Create space at group[offset], assuming value_type has trivial + // copy constructor and destructor, and the allocator_type is + // the default libc_allocator_with_alloc. (Really, we want it to have + // "trivial move", because that's what realloc and memmove both do. + // But there's no way to capture that using type_traits, so we + // pretend that move(x, y) is equivalent to "x.~T(); new(x) T(y);" + // which is pretty much correct, if a bit conservative.) + void set_aux(size_type offset, base::true_type) { + group = settings.realloc_or_die(group, settings.num_buckets+1); + // This is equivalent to memmove(), but faster on my Intel P4, + // at least with gcc4.1 -O2 / glibc 2.3.6. + for (size_type i = settings.num_buckets; i > offset; --i) + // cast to void* to prevent compiler warnings about writing to an object + // with no trivial copy-assignment + memcpy(static_cast(group + i), group + i-1, sizeof(*group)); + } + + // Create space at group[offset], without special assumptions about value_type + // and allocator_type. + void set_aux(size_type offset, base::false_type) { + // This is valid because 0 <= offset <= num_buckets + pointer p = allocate_group(settings.num_buckets + 1); + std::uninitialized_copy(group, group + offset, p); + std::uninitialized_copy(group + offset, group + settings.num_buckets, + p + offset + 1); + free_group(); + group = p; + } + + public: + // This returns a reference to the inserted item (which is a copy of val). + // TODO(austern): Make this exception safe: handle exceptions from + // value_type's copy constructor. + reference set(size_type i, const_reference val) { + size_type offset = pos_to_offset(bitmap, i); // where we'll find (or insert) + if ( bmtest(i) ) { + // Delete the old value, which we're replacing with the new one + group[offset].~value_type(); + } else { + typedef base::integral_constant::value && + base::has_trivial_destructor::value && + base::is_same< + allocator_type, + libc_allocator_with_realloc >::value)> + realloc_and_memmove_ok; // we pretend mv(x,y) == "x.~T(); new(x) T(y)" + set_aux(offset, realloc_and_memmove_ok()); + ++settings.num_buckets; + bmset(i); + } + // This does the actual inserting. Since we made the array using + // malloc, we use "placement new" to just call the constructor. + new(&group[offset]) value_type(val); + return group[offset]; + } + + // We let you see if a bucket is non-empty without retrieving it + bool test(size_type i) const { + return bmtest(i) != 0; + } + bool test(iterator pos) const { + return bmtest(pos.pos) != 0; + } + + private: + // Shrink the array, assuming value_type has trivial copy + // constructor and destructor, and the allocator_type is the default + // libc_allocator_with_alloc. (Really, we want it to have "trivial + // move", because that's what realloc and memmove both do. But + // there's no way to capture that using type_traits, so we pretend + // that move(x, y) is equivalent to ""x.~T(); new(x) T(y);" + // which is pretty much correct, if a bit conservative.) + void erase_aux(size_type offset, base::true_type) { + // This isn't technically necessary, since we know we have a + // trivial destructor, but is a cheap way to get a bit more safety. + group[offset].~value_type(); + // This is equivalent to memmove(), but faster on my Intel P4, + // at lesat with gcc4.1 -O2 / glibc 2.3.6. + assert(settings.num_buckets > 0); + for (size_type i = offset; i < settings.num_buckets-1; ++i) + // cast to void* to prevent compiler warnings about writing to an object + // with no trivial copy-assignment + // hopefully inlined! + memcpy(static_cast(group + i), group + i+1, sizeof(*group)); + group = settings.realloc_or_die(group, settings.num_buckets-1); + } + + // Shrink the array, without any special assumptions about value_type and + // allocator_type. + void erase_aux(size_type offset, base::false_type) { + // This is valid because 0 <= offset < num_buckets. Note the inequality. + pointer p = allocate_group(settings.num_buckets - 1); + std::uninitialized_copy(group, group + offset, p); + std::uninitialized_copy(group + offset + 1, group + settings.num_buckets, + p + offset); + free_group(); + group = p; + } + + public: + // This takes the specified elements out of the group. This is + // "undefining", rather than "clearing". + // TODO(austern): Make this exception safe: handle exceptions from + // value_type's copy constructor. + void erase(size_type i) { + if ( bmtest(i) ) { // trivial to erase empty bucket + size_type offset = pos_to_offset(bitmap,i); // where we'll find (or insert) + if ( settings.num_buckets == 1 ) { + free_group(); + group = NULL; + } else { + typedef base::integral_constant::value && + base::has_trivial_destructor::value && + base::is_same< + allocator_type, + libc_allocator_with_realloc >::value)> + realloc_and_memmove_ok; // pretend mv(x,y) == "x.~T(); new(x) T(y)" + erase_aux(offset, realloc_and_memmove_ok()); + } + --settings.num_buckets; + bmclear(i); + } + } + + void erase(iterator pos) { + erase(pos.pos); + } + + void erase(iterator start_it, iterator end_it) { + // This could be more efficient, but to do so we'd need to make + // bmclear() clear a range of indices. Doesn't seem worth it. + for ( ; start_it != end_it; ++start_it ) + erase(start_it); + } + + + // I/O + // We support reading and writing groups to disk. We don't store + // the actual array contents (which we don't know how to store), + // just the bitmap and size. Meant to be used with table I/O. + + template bool write_metadata(OUTPUT *fp) const { + // we explicitly set to u_int16_t + assert(sizeof(settings.num_buckets) == 2); + if ( !sparsehash_internal::write_bigendian_number(fp, settings.num_buckets, + 2) ) + return false; + if ( !sparsehash_internal::write_data(fp, bitmap, sizeof(bitmap)) ) + return false; + return true; + } + + // Reading destroys the old group contents! Returns true if all was ok. + template bool read_metadata(INPUT *fp) { + clear(); + if ( !sparsehash_internal::read_bigendian_number(fp, &settings.num_buckets, + 2) ) + return false; + if ( !sparsehash_internal::read_data(fp, bitmap, sizeof(bitmap)) ) + return false; + // We'll allocate the space, but we won't fill it: it will be + // left as uninitialized raw memory. + group = allocate_group(settings.num_buckets); + return true; + } + + // Again, only meaningful if value_type is a POD. + template bool read_nopointer_data(INPUT *fp) { + for ( nonempty_iterator it = nonempty_begin(); + it != nonempty_end(); ++it ) { + if ( !sparsehash_internal::read_data(fp, &(*it), sizeof(*it)) ) + return false; + } + return true; + } + + // If your keys and values are simple enough, we can write them + // to disk for you. "simple enough" means POD and no pointers. + // However, we don't try to normalize endianness. + template bool write_nopointer_data(OUTPUT *fp) const { + for ( const_nonempty_iterator it = nonempty_begin(); + it != nonempty_end(); ++it ) { + if ( !sparsehash_internal::write_data(fp, &(*it), sizeof(*it)) ) + return false; + } + return true; + } + + + // Comparisons. We only need to define == and < -- we get + // != > <= >= via relops.h (which we happily included above). + // Note the comparisons are pretty arbitrary: we compare + // values of the first index that isn't equal (using default + // value for empty buckets). + bool operator==(const sparsegroup& x) const { + return ( settings.num_buckets == x.settings.num_buckets && + memcmp(bitmap, x.bitmap, sizeof(bitmap)) == 0 && + std::equal(begin(), end(), x.begin()) ); // from + } + + bool operator<(const sparsegroup& x) const { // also from + return std::lexicographical_compare(begin(), end(), x.begin(), x.end()); + } + bool operator!=(const sparsegroup& x) const { return !(*this == x); } + bool operator<=(const sparsegroup& x) const { return !(x < *this); } + bool operator>(const sparsegroup& x) const { return x < *this; } + bool operator>=(const sparsegroup& x) const { return !(*this < x); } + + private: + template + class alloc_impl : public A { + public: + typedef typename A::pointer pointer; + typedef typename A::size_type size_type; + + // Convert a normal allocator to one that has realloc_or_die() + alloc_impl(const A& a) : A(a) { } + + // realloc_or_die should only be used when using the default + // allocator (libc_allocator_with_realloc). + pointer realloc_or_die(pointer /*ptr*/, size_type /*n*/) { + fprintf(stderr, "realloc_or_die is only supported for " + "libc_allocator_with_realloc\n"); + exit(1); + return NULL; + } + }; + + // A template specialization of alloc_impl for + // libc_allocator_with_realloc that can handle realloc_or_die. + template + class alloc_impl > + : public libc_allocator_with_realloc { + public: + typedef typename libc_allocator_with_realloc::pointer pointer; + typedef typename libc_allocator_with_realloc::size_type size_type; + + alloc_impl(const libc_allocator_with_realloc& a) + : libc_allocator_with_realloc(a) { } + + pointer realloc_or_die(pointer ptr, size_type n) { + pointer retval = this->reallocate(ptr, n); + if (retval == NULL) { + fprintf(stderr, "sparsehash: FATAL ERROR: failed to reallocate " + "%lu elements for ptr %p", static_cast(n), ptr); + exit(1); + } + return retval; + } + }; + + // Package allocator with num_buckets to eliminate memory needed for the + // zero-size allocator. + // If new fields are added to this class, we should add them to + // operator= and swap. + class Settings : public alloc_impl { + public: + Settings(const alloc_impl& a, u_int16_t n = 0) + : alloc_impl(a), num_buckets(n) { } + Settings(const Settings& s) + : alloc_impl(s), num_buckets(s.num_buckets) { } + + u_int16_t num_buckets; // limits GROUP_SIZE to 64K + }; + + // The actual data + pointer group; // (small) array of T's + Settings settings; // allocator and num_buckets + unsigned char bitmap[(GROUP_SIZE-1)/8 + 1]; // fancy math is so we round up +}; + +// We need a global swap as well +template +inline void swap(sparsegroup &x, + sparsegroup &y) { + x.swap(y); +} + +// --------------------------------------------------------------------------- + + +template > +class sparsetable { + private: + typedef typename Alloc::template rebind::other value_alloc_type; + typedef typename Alloc::template rebind< + sparsegroup >::other vector_alloc; + + public: + // Basic types + typedef T value_type; // stolen from stl_vector.h + typedef Alloc allocator_type; + typedef typename value_alloc_type::size_type size_type; + typedef typename value_alloc_type::difference_type difference_type; + typedef typename value_alloc_type::reference reference; + typedef typename value_alloc_type::const_reference const_reference; + typedef typename value_alloc_type::pointer pointer; + typedef typename value_alloc_type::const_pointer const_pointer; + typedef table_iterator > iterator; + typedef const_table_iterator > + const_iterator; + typedef table_element_adaptor > + element_adaptor; + typedef std::reverse_iterator const_reverse_iterator; + typedef std::reverse_iterator reverse_iterator; // from iterator.h + + // These are our special iterators, that go over non-empty buckets in a + // table. These aren't const only because you can change non-empty bcks. + typedef two_d_iterator< std::vector< sparsegroup, + vector_alloc> > + nonempty_iterator; + typedef const_two_d_iterator< std::vector< sparsegroup, + vector_alloc> > + const_nonempty_iterator; + typedef std::reverse_iterator reverse_nonempty_iterator; + typedef std::reverse_iterator const_reverse_nonempty_iterator; + // Another special iterator: it frees memory as it iterates (used to resize) + typedef destructive_two_d_iterator< std::vector< sparsegroup, + vector_alloc> > + destructive_iterator; + + // Iterator functions + iterator begin() { return iterator(this, 0); } + const_iterator begin() const { return const_iterator(this, 0); } + iterator end() { return iterator(this, size()); } + const_iterator end() const { return const_iterator(this, size()); } + reverse_iterator rbegin() { return reverse_iterator(end()); } + const_reverse_iterator rbegin() const { return const_reverse_iterator(end()); } + reverse_iterator rend() { return reverse_iterator(begin()); } + const_reverse_iterator rend() const { return const_reverse_iterator(begin()); } + + // Versions for our special non-empty iterator + nonempty_iterator nonempty_begin() { + return nonempty_iterator(groups.begin(), groups.end(), groups.begin()); + } + const_nonempty_iterator nonempty_begin() const { + return const_nonempty_iterator(groups.begin(),groups.end(), groups.begin()); + } + nonempty_iterator nonempty_end() { + return nonempty_iterator(groups.begin(), groups.end(), groups.end()); + } + const_nonempty_iterator nonempty_end() const { + return const_nonempty_iterator(groups.begin(), groups.end(), groups.end()); + } + reverse_nonempty_iterator nonempty_rbegin() { + return reverse_nonempty_iterator(nonempty_end()); + } + const_reverse_nonempty_iterator nonempty_rbegin() const { + return const_reverse_nonempty_iterator(nonempty_end()); + } + reverse_nonempty_iterator nonempty_rend() { + return reverse_nonempty_iterator(nonempty_begin()); + } + const_reverse_nonempty_iterator nonempty_rend() const { + return const_reverse_nonempty_iterator(nonempty_begin()); + } + destructive_iterator destructive_begin() { + return destructive_iterator(groups.begin(), groups.end(), groups.begin()); + } + destructive_iterator destructive_end() { + return destructive_iterator(groups.begin(), groups.end(), groups.end()); + } + + typedef sparsegroup group_type; + typedef std::vector group_vector_type; + + typedef typename group_vector_type::reference GroupsReference; + typedef typename group_vector_type::const_reference GroupsConstReference; + typedef typename group_vector_type::iterator GroupsIterator; + typedef typename group_vector_type::const_iterator GroupsConstIterator; + + // How to deal with the proper group + static size_type num_groups(size_type num) { // how many to hold num buckets + return num == 0 ? 0 : ((num-1) / GROUP_SIZE) + 1; + } + + u_int16_t pos_in_group(size_type i) const { + return static_cast(i % GROUP_SIZE); + } + size_type group_num(size_type i) const { + return i / GROUP_SIZE; + } + GroupsReference which_group(size_type i) { + return groups[group_num(i)]; + } + GroupsConstReference which_group(size_type i) const { + return groups[group_num(i)]; + } + + public: + // Constructors -- default, normal (when you specify size), and copy + explicit sparsetable(size_type sz = 0, Alloc alloc = Alloc()) + : groups(vector_alloc(alloc)), settings(alloc, sz) { + groups.resize(num_groups(sz), group_type(settings)); + } + // We can get away with using the default copy constructor, + // and default destructor, and hence the default operator=. Huzzah! + + // Many STL algorithms use swap instead of copy constructors + void swap(sparsetable& x) { + std::swap(groups, x.groups); // defined in stl_algobase.h + std::swap(settings.table_size, x.settings.table_size); + std::swap(settings.num_buckets, x.settings.num_buckets); + } + + // It's always nice to be able to clear a table without deallocating it + void clear() { + GroupsIterator group; + for ( group = groups.begin(); group != groups.end(); ++group ) { + group->clear(); + } + settings.num_buckets = 0; + } + + // ACCESSOR FUNCTIONS for the things we templatize on, basically + allocator_type get_allocator() const { + return allocator_type(settings); + } + + + // Functions that tell you about size. + // NOTE: empty() is non-intuitive! It does not tell you the number + // of not-empty buckets (use num_nonempty() for that). Instead + // it says whether you've allocated any buckets or not. + size_type size() const { return settings.table_size; } + size_type max_size() const { return settings.max_size(); } + bool empty() const { return settings.table_size == 0; } + // We also may want to know how many *used* buckets there are + size_type num_nonempty() const { return settings.num_buckets; } + + // OK, we'll let you resize one of these puppies + void resize(size_type new_size) { + groups.resize(num_groups(new_size), group_type(settings)); + if ( new_size < settings.table_size) { + // lower num_buckets, clear last group + if ( pos_in_group(new_size) > 0 ) // need to clear inside last group + groups.back().erase(groups.back().begin() + pos_in_group(new_size), + groups.back().end()); + settings.num_buckets = 0; // refigure # of used buckets + GroupsConstIterator group; + for ( group = groups.begin(); group != groups.end(); ++group ) + settings.num_buckets += group->num_nonempty(); + } + settings.table_size = new_size; + } + + + // We let you see if a bucket is non-empty without retrieving it + bool test(size_type i) const { + assert(i < settings.table_size); + return which_group(i).test(pos_in_group(i)); + } + bool test(iterator pos) const { + return which_group(pos.pos).test(pos_in_group(pos.pos)); + } + bool test(const_iterator pos) const { + return which_group(pos.pos).test(pos_in_group(pos.pos)); + } + + // We only return const_references because it's really hard to + // return something settable for empty buckets. Use set() instead. + const_reference get(size_type i) const { + assert(i < settings.table_size); + return which_group(i).get(pos_in_group(i)); + } + + // TODO(csilvers): make protected + friend + // This is used by sparse_hashtable to get an element from the table + // when we know it exists (because the caller has called test(i)). + const_reference unsafe_get(size_type i) const { + assert(i < settings.table_size); + assert(test(i)); + return which_group(i).unsafe_get(pos_in_group(i)); + } + + // TODO(csilvers): make protected + friend element_adaptor + reference mutating_get(size_type i) { // fills bucket i before getting + assert(i < settings.table_size); + typename group_type::size_type old_numbuckets = which_group(i).num_nonempty(); + reference retval = which_group(i).mutating_get(pos_in_group(i)); + settings.num_buckets += which_group(i).num_nonempty() - old_numbuckets; + return retval; + } + + // Syntactic sugar. As in sparsegroup, the non-const version is harder + const_reference operator[](size_type i) const { + return get(i); + } + + element_adaptor operator[](size_type i) { + return element_adaptor(this, i); + } + + // Needed for hashtables, gets as a nonempty_iterator. Crashes for empty bcks + const_nonempty_iterator get_iter(size_type i) const { + assert(test(i)); // how can a nonempty_iterator point to an empty bucket? + return const_nonempty_iterator( + groups.begin(), groups.end(), + groups.begin() + group_num(i), + (groups[group_num(i)].nonempty_begin() + + groups[group_num(i)].pos_to_offset(pos_in_group(i)))); + } + // For nonempty we can return a non-const version + nonempty_iterator get_iter(size_type i) { + assert(test(i)); // how can a nonempty_iterator point to an empty bucket? + return nonempty_iterator( + groups.begin(), groups.end(), + groups.begin() + group_num(i), + (groups[group_num(i)].nonempty_begin() + + groups[group_num(i)].pos_to_offset(pos_in_group(i)))); + } + + // And the reverse transformation. + size_type get_pos(const const_nonempty_iterator& it) const { + difference_type current_row = it.row_current - it.row_begin; + difference_type current_col = (it.col_current - + groups[current_row].nonempty_begin()); + return ((current_row * GROUP_SIZE) + + groups[current_row].offset_to_pos(current_col)); + } + + + // This returns a reference to the inserted item (which is a copy of val) + // The trick is to figure out whether we're replacing or inserting anew + reference set(size_type i, const_reference val) { + assert(i < settings.table_size); + typename group_type::size_type old_numbuckets = which_group(i).num_nonempty(); + reference retval = which_group(i).set(pos_in_group(i), val); + settings.num_buckets += which_group(i).num_nonempty() - old_numbuckets; + return retval; + } + + // This takes the specified elements out of the table. This is + // "undefining", rather than "clearing". + void erase(size_type i) { + assert(i < settings.table_size); + typename group_type::size_type old_numbuckets = which_group(i).num_nonempty(); + which_group(i).erase(pos_in_group(i)); + settings.num_buckets += which_group(i).num_nonempty() - old_numbuckets; + } + + void erase(iterator pos) { + erase(pos.pos); + } + + void erase(iterator start_it, iterator end_it) { + // This could be more efficient, but then we'd need to figure + // out if we spanned groups or not. Doesn't seem worth it. + for ( ; start_it != end_it; ++start_it ) + erase(start_it); + } + + + // We support reading and writing tables to disk. We don't store + // the actual array contents (which we don't know how to store), + // just the groups and sizes. Returns true if all went ok. + + private: + // Every time the disk format changes, this should probably change too + typedef unsigned long MagicNumberType; + static const MagicNumberType MAGIC_NUMBER = 0x24687531; + + // Old versions of this code write all data in 32 bits. We need to + // support these files as well as having support for 64-bit systems. + // So we use the following encoding scheme: for values < 2^32-1, we + // store in 4 bytes in big-endian order. For values > 2^32, we + // store 0xFFFFFFF followed by 8 bytes in big-endian order. This + // causes us to mis-read old-version code that stores exactly + // 0xFFFFFFF, but I don't think that is likely to have happened for + // these particular values. + template + static bool write_32_or_64(OUTPUT* fp, IntType value) { + if ( value < 0xFFFFFFFFULL ) { // fits in 4 bytes + if ( !sparsehash_internal::write_bigendian_number(fp, value, 4) ) + return false; + } else { + if ( !sparsehash_internal::write_bigendian_number(fp, 0xFFFFFFFFUL, 4) ) + return false; + if ( !sparsehash_internal::write_bigendian_number(fp, value, 8) ) + return false; + } + return true; + } + + template + static bool read_32_or_64(INPUT* fp, IntType *value) { // reads into value + MagicNumberType first4 = 0; // a convenient 32-bit unsigned type + if ( !sparsehash_internal::read_bigendian_number(fp, &first4, 4) ) + return false; + if ( first4 < 0xFFFFFFFFULL ) { + *value = first4; + } else { + if ( !sparsehash_internal::read_bigendian_number(fp, value, 8) ) + return false; + } + return true; + } + + public: + // read/write_metadata() and read_write/nopointer_data() are DEPRECATED. + // Use serialize() and unserialize(), below, for new code. + + template bool write_metadata(OUTPUT *fp) const { + if ( !write_32_or_64(fp, MAGIC_NUMBER) ) return false; + if ( !write_32_or_64(fp, settings.table_size) ) return false; + if ( !write_32_or_64(fp, settings.num_buckets) ) return false; + + GroupsConstIterator group; + for ( group = groups.begin(); group != groups.end(); ++group ) + if ( group->write_metadata(fp) == false ) return false; + return true; + } + + // Reading destroys the old table contents! Returns true if read ok. + template bool read_metadata(INPUT *fp) { + size_type magic_read = 0; + if ( !read_32_or_64(fp, &magic_read) ) return false; + if ( magic_read != MAGIC_NUMBER ) { + clear(); // just to be consistent + return false; + } + + if ( !read_32_or_64(fp, &settings.table_size) ) return false; + if ( !read_32_or_64(fp, &settings.num_buckets) ) return false; + + resize(settings.table_size); // so the vector's sized ok + GroupsIterator group; + for ( group = groups.begin(); group != groups.end(); ++group ) + if ( group->read_metadata(fp) == false ) return false; + return true; + } + + // This code is identical to that for SparseGroup + // If your keys and values are simple enough, we can write them + // to disk for you. "simple enough" means no pointers. + // However, we don't try to normalize endianness + bool write_nopointer_data(FILE *fp) const { + for ( const_nonempty_iterator it = nonempty_begin(); + it != nonempty_end(); ++it ) { + if ( !fwrite(&*it, sizeof(*it), 1, fp) ) return false; + } + return true; + } + + // When reading, we have to override the potential const-ness of *it + bool read_nopointer_data(FILE *fp) { + for ( nonempty_iterator it = nonempty_begin(); + it != nonempty_end(); ++it ) { + if ( !fread(reinterpret_cast(&(*it)), sizeof(*it), 1, fp) ) + return false; + } + return true; + } + + // INPUT and OUTPUT must be either a FILE, *or* a C++ stream + // (istream, ostream, etc) *or* a class providing + // Read(void*, size_t) and Write(const void*, size_t) + // (respectively), which writes a buffer into a stream + // (which the INPUT/OUTPUT instance presumably owns). + + typedef sparsehash_internal::pod_serializer NopointerSerializer; + + // ValueSerializer: a functor. operator()(OUTPUT*, const value_type&) + template + bool serialize(ValueSerializer serializer, OUTPUT *fp) { + if ( !write_metadata(fp) ) + return false; + for ( const_nonempty_iterator it = nonempty_begin(); + it != nonempty_end(); ++it ) { + if ( !serializer(fp, *it) ) return false; + } + return true; + } + + // ValueSerializer: a functor. operator()(INPUT*, value_type*) + template + bool unserialize(ValueSerializer serializer, INPUT *fp) { + clear(); + if ( !read_metadata(fp) ) + return false; + for ( nonempty_iterator it = nonempty_begin(); + it != nonempty_end(); ++it ) { + if ( !serializer(fp, &*it) ) return false; + } + return true; + } + + // Comparisons. Note the comparisons are pretty arbitrary: we + // compare values of the first index that isn't equal (using default + // value for empty buckets). + bool operator==(const sparsetable& x) const { + return ( settings.table_size == x.settings.table_size && + settings.num_buckets == x.settings.num_buckets && + groups == x.groups ); + } + + bool operator<(const sparsetable& x) const { + return std::lexicographical_compare(begin(), end(), x.begin(), x.end()); + } + bool operator!=(const sparsetable& x) const { return !(*this == x); } + bool operator<=(const sparsetable& x) const { return !(x < *this); } + bool operator>(const sparsetable& x) const { return x < *this; } + bool operator>=(const sparsetable& x) const { return !(*this < x); } + + + private: + // Package allocator with table_size and num_buckets to eliminate memory + // needed for the zero-size allocator. + // If new fields are added to this class, we should add them to + // operator= and swap. + class Settings : public allocator_type { + public: + typedef typename allocator_type::size_type size_type; + + Settings(const allocator_type& a, size_type sz = 0, size_type n = 0) + : allocator_type(a), table_size(sz), num_buckets(n) { } + + Settings(const Settings& s) + : allocator_type(s), + table_size(s.table_size), num_buckets(s.num_buckets) { } + + size_type table_size; // how many buckets they want + size_type num_buckets; // number of non-empty buckets + }; + + // The actual data + group_vector_type groups; // our list of groups + Settings settings; // allocator, table size, buckets +}; + +// We need a global swap as well +template +inline void swap(sparsetable &x, + sparsetable &y) { + x.swap(y); +} + +_END_GOOGLE_NAMESPACE_ + +#endif // UTIL_GTL_SPARSETABLE_H_ diff --git a/src/includes/3thparty/sparsehash/template_util.h b/src/includes/3thparty/sparsehash/template_util.h new file mode 100644 index 0000000..6fec3d0 --- /dev/null +++ b/src/includes/3thparty/sparsehash/template_util.h @@ -0,0 +1,134 @@ +// Copyright 2005 Google Inc. +// All rights reserved. +// +// Redistribution and use in source and binary forms, with or without +// modification, are permitted provided that the following conditions are +// met: +// +// * Redistributions of source code must retain the above copyright +// notice, this list of conditions and the following disclaimer. +// * Redistributions in binary form must reproduce the above +// copyright notice, this list of conditions and the following disclaimer +// in the documentation and/or other materials provided with the +// distribution. +// * Neither the name of Google Inc. nor the names of its +// contributors may be used to endorse or promote products derived from +// this software without specific prior written permission. +// +// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS +// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT +// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR +// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT +// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, +// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT +// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, +// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY +// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT +// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE +// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + +// ---- +// +// Template metaprogramming utility functions. +// +// This code is compiled directly on many platforms, including client +// platforms like Windows, Mac, and embedded systems. Before making +// any changes here, make sure that you're not breaking any platforms. +// +// +// The names choosen here reflect those used in tr1 and the boost::mpl +// library, there are similar operations used in the Loki library as +// well. I prefer the boost names for 2 reasons: +// 1. I think that portions of the Boost libraries are more likely to +// be included in the c++ standard. +// 2. It is not impossible that some of the boost libraries will be +// included in our own build in the future. +// Both of these outcomes means that we may be able to directly replace +// some of these with boost equivalents. +// +#ifndef BASE_TEMPLATE_UTIL_H_ +#define BASE_TEMPLATE_UTIL_H_ + +#include +_START_GOOGLE_NAMESPACE_ + +// Types small_ and big_ are guaranteed such that sizeof(small_) < +// sizeof(big_) +typedef char small_; + +struct big_ { + char dummy[2]; +}; + +// Identity metafunction. +template +struct identity_ { + typedef T type; +}; + +// integral_constant, defined in tr1, is a wrapper for an integer +// value. We don't really need this generality; we could get away +// with hardcoding the integer type to bool. We use the fully +// general integer_constant for compatibility with tr1. + +template +struct integral_constant { + static const T value = v; + typedef T value_type; + typedef integral_constant type; +}; + +template const T integral_constant::value; + + +// Abbreviations: true_type and false_type are structs that represent boolean +// true and false values. Also define the boost::mpl versions of those names, +// true_ and false_. +typedef integral_constant true_type; +typedef integral_constant false_type; +typedef true_type true_; +typedef false_type false_; + +// if_ is a templatized conditional statement. +// if_ is a compile time evaluation of cond. +// if_<>::type contains A if cond is true, B otherwise. +template +struct if_{ + typedef A type; +}; + +template +struct if_ { + typedef B type; +}; + + +// type_equals_ is a template type comparator, similar to Loki IsSameType. +// type_equals_::value is true iff "A" is the same type as "B". +// +// New code should prefer base::is_same, defined in base/type_traits.h. +// It is functionally identical, but is_same is the standard spelling. +template +struct type_equals_ : public false_ { +}; + +template +struct type_equals_ : public true_ { +}; + +// and_ is a template && operator. +// and_::value evaluates "A::value && B::value". +template +struct and_ : public integral_constant { +}; + +// or_ is a template || operator. +// or_::value evaluates "A::value || B::value". +template +struct or_ : public integral_constant { +}; + + +_END_GOOGLE_NAMESPACE_ + +#endif // BASE_TEMPLATE_UTIL_H_ diff --git a/src/includes/3thparty/sparsehash/type_traits.h b/src/includes/3thparty/sparsehash/type_traits.h new file mode 100644 index 0000000..f909cf9 --- /dev/null +++ b/src/includes/3thparty/sparsehash/type_traits.h @@ -0,0 +1,342 @@ +// Copyright (c) 2006, Google Inc. +// All rights reserved. +// +// Redistribution and use in source and binary forms, with or without +// modification, are permitted provided that the following conditions are +// met: +// +// * Redistributions of source code must retain the above copyright +// notice, this list of conditions and the following disclaimer. +// * Redistributions in binary form must reproduce the above +// copyright notice, this list of conditions and the following disclaimer +// in the documentation and/or other materials provided with the +// distribution. +// * Neither the name of Google Inc. nor the names of its +// contributors may be used to endorse or promote products derived from +// this software without specific prior written permission. +// +// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS +// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT +// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR +// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT +// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, +// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT +// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, +// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY +// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT +// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE +// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + +// ---- +// +// This code is compiled directly on many platforms, including client +// platforms like Windows, Mac, and embedded systems. Before making +// any changes here, make sure that you're not breaking any platforms. +// +// Define a small subset of tr1 type traits. The traits we define are: +// is_integral +// is_floating_point +// is_pointer +// is_enum +// is_reference +// is_pod +// has_trivial_constructor +// has_trivial_copy +// has_trivial_assign +// has_trivial_destructor +// remove_const +// remove_volatile +// remove_cv +// remove_reference +// add_reference +// remove_pointer +// is_same +// is_convertible +// We can add more type traits as required. + +#ifndef BASE_TYPE_TRAITS_H_ +#define BASE_TYPE_TRAITS_H_ + +#include +#include // For pair + +#include // For true_type and false_type + +_START_GOOGLE_NAMESPACE_ + +template struct is_integral; +template struct is_floating_point; +template struct is_pointer; +// MSVC can't compile this correctly, and neither can gcc 3.3.5 (at least) +#if !defined(_MSC_VER) && !(defined(__GNUC__) && __GNUC__ <= 3) +// is_enum uses is_convertible, which is not available on MSVC. +template struct is_enum; +#endif +template struct is_reference; +template struct is_pod; +template struct has_trivial_constructor; +template struct has_trivial_copy; +template struct has_trivial_assign; +template struct has_trivial_destructor; +template struct remove_const; +template struct remove_volatile; +template struct remove_cv; +template struct remove_reference; +template struct add_reference; +template struct remove_pointer; +template struct is_same; +#if !defined(_MSC_VER) && !(defined(__GNUC__) && __GNUC__ <= 3) +template struct is_convertible; +#endif + +// is_integral is false except for the built-in integer types. A +// cv-qualified type is integral if and only if the underlying type is. +template struct is_integral : false_type { }; +template<> struct is_integral : true_type { }; +template<> struct is_integral : true_type { }; +template<> struct is_integral : true_type { }; +template<> struct is_integral : true_type { }; +#if defined(_MSC_VER) +// wchar_t is not by default a distinct type from unsigned short in +// Microsoft C. +// See http://msdn2.microsoft.com/en-us/library/dh8che7s(VS.80).aspx +template<> struct is_integral<__wchar_t> : true_type { }; +#else +template<> struct is_integral : true_type { }; +#endif +template<> struct is_integral : true_type { }; +template<> struct is_integral : true_type { }; +template<> struct is_integral : true_type { }; +template<> struct is_integral : true_type { }; +template<> struct is_integral : true_type { }; +template<> struct is_integral : true_type { }; +#ifdef HAVE_LONG_LONG +template<> struct is_integral : true_type { }; +template<> struct is_integral : true_type { }; +#endif +template struct is_integral : is_integral { }; +template struct is_integral : is_integral { }; +template struct is_integral : is_integral { }; + +// is_floating_point is false except for the built-in floating-point types. +// A cv-qualified type is integral if and only if the underlying type is. +template struct is_floating_point : false_type { }; +template<> struct is_floating_point : true_type { }; +template<> struct is_floating_point : true_type { }; +template<> struct is_floating_point : true_type { }; +template struct is_floating_point + : is_floating_point { }; +template struct is_floating_point + : is_floating_point { }; +template struct is_floating_point + : is_floating_point { }; + +// is_pointer is false except for pointer types. A cv-qualified type (e.g. +// "int* const", as opposed to "int const*") is cv-qualified if and only if +// the underlying type is. +template struct is_pointer : false_type { }; +template struct is_pointer : true_type { }; +template struct is_pointer : is_pointer { }; +template struct is_pointer : is_pointer { }; +template struct is_pointer : is_pointer { }; + +#if !defined(_MSC_VER) && !(defined(__GNUC__) && __GNUC__ <= 3) + +namespace internal { + +template struct is_class_or_union { + template static small_ tester(void (U::*)()); + template static big_ tester(...); + static const bool value = sizeof(tester(0)) == sizeof(small_); +}; + +// is_convertible chokes if the first argument is an array. That's why +// we use add_reference here. +template struct is_enum_impl + : is_convertible::type, int> { }; + +template struct is_enum_impl : false_type { }; + +} // namespace internal + +// Specified by TR1 [4.5.1] primary type categories. + +// Implementation note: +// +// Each type is either void, integral, floating point, array, pointer, +// reference, member object pointer, member function pointer, enum, +// union or class. Out of these, only integral, floating point, reference, +// class and enum types are potentially convertible to int. Therefore, +// if a type is not a reference, integral, floating point or class and +// is convertible to int, it's a enum. Adding cv-qualification to a type +// does not change whether it's an enum. +// +// Is-convertible-to-int check is done only if all other checks pass, +// because it can't be used with some types (e.g. void or classes with +// inaccessible conversion operators). +template struct is_enum + : internal::is_enum_impl< + is_same::value || + is_integral::value || + is_floating_point::value || + is_reference::value || + internal::is_class_or_union::value, + T> { }; + +template struct is_enum : is_enum { }; +template struct is_enum : is_enum { }; +template struct is_enum : is_enum { }; + +#endif + +// is_reference is false except for reference types. +template struct is_reference : false_type {}; +template struct is_reference : true_type {}; + + +// We can't get is_pod right without compiler help, so fail conservatively. +// We will assume it's false except for arithmetic types, enumerations, +// pointers and cv-qualified versions thereof. Note that std::pair +// is not a POD even if T and U are PODs. +template struct is_pod + : integral_constant::value || + is_floating_point::value || +#if !defined(_MSC_VER) && !(defined(__GNUC__) && __GNUC__ <= 3) + // is_enum is not available on MSVC. + is_enum::value || +#endif + is_pointer::value)> { }; +template struct is_pod : is_pod { }; +template struct is_pod : is_pod { }; +template struct is_pod : is_pod { }; + + +// We can't get has_trivial_constructor right without compiler help, so +// fail conservatively. We will assume it's false except for: (1) types +// for which is_pod is true. (2) std::pair of types with trivial +// constructors. (3) array of a type with a trivial constructor. +// (4) const versions thereof. +template struct has_trivial_constructor : is_pod { }; +template struct has_trivial_constructor > + : integral_constant::value && + has_trivial_constructor::value)> { }; +template struct has_trivial_constructor + : has_trivial_constructor { }; +template struct has_trivial_constructor + : has_trivial_constructor { }; + +// We can't get has_trivial_copy right without compiler help, so fail +// conservatively. We will assume it's false except for: (1) types +// for which is_pod is true. (2) std::pair of types with trivial copy +// constructors. (3) array of a type with a trivial copy constructor. +// (4) const versions thereof. +template struct has_trivial_copy : is_pod { }; +template struct has_trivial_copy > + : integral_constant::value && + has_trivial_copy::value)> { }; +template struct has_trivial_copy + : has_trivial_copy { }; +template struct has_trivial_copy : has_trivial_copy { }; + +// We can't get has_trivial_assign right without compiler help, so fail +// conservatively. We will assume it's false except for: (1) types +// for which is_pod is true. (2) std::pair of types with trivial copy +// constructors. (3) array of a type with a trivial assign constructor. +template struct has_trivial_assign : is_pod { }; +template struct has_trivial_assign > + : integral_constant::value && + has_trivial_assign::value)> { }; +template struct has_trivial_assign + : has_trivial_assign { }; + +// We can't get has_trivial_destructor right without compiler help, so +// fail conservatively. We will assume it's false except for: (1) types +// for which is_pod is true. (2) std::pair of types with trivial +// destructors. (3) array of a type with a trivial destructor. +// (4) const versions thereof. +template struct has_trivial_destructor : is_pod { }; +template struct has_trivial_destructor > + : integral_constant::value && + has_trivial_destructor::value)> { }; +template struct has_trivial_destructor + : has_trivial_destructor { }; +template struct has_trivial_destructor + : has_trivial_destructor { }; + +// Specified by TR1 [4.7.1] +template struct remove_const { typedef T type; }; +template struct remove_const { typedef T type; }; +template struct remove_volatile { typedef T type; }; +template struct remove_volatile { typedef T type; }; +template struct remove_cv { + typedef typename remove_const::type>::type type; +}; + + +// Specified by TR1 [4.7.2] Reference modifications. +template struct remove_reference { typedef T type; }; +template struct remove_reference { typedef T type; }; + +template struct add_reference { typedef T& type; }; +template struct add_reference { typedef T& type; }; + +// Specified by TR1 [4.7.4] Pointer modifications. +template struct remove_pointer { typedef T type; }; +template struct remove_pointer { typedef T type; }; +template struct remove_pointer { typedef T type; }; +template struct remove_pointer { typedef T type; }; +template struct remove_pointer { + typedef T type; }; + +// Specified by TR1 [4.6] Relationships between types +template struct is_same : public false_type { }; +template struct is_same : public true_type { }; + +// Specified by TR1 [4.6] Relationships between types +#if !defined(_MSC_VER) && !(defined(__GNUC__) && __GNUC__ <= 3) +namespace internal { + +// This class is an implementation detail for is_convertible, and you +// don't need to know how it works to use is_convertible. For those +// who care: we declare two different functions, one whose argument is +// of type To and one with a variadic argument list. We give them +// return types of different size, so we can use sizeof to trick the +// compiler into telling us which function it would have chosen if we +// had called it with an argument of type From. See Alexandrescu's +// _Modern C++ Design_ for more details on this sort of trick. + +template +struct ConvertHelper { + static small_ Test(To); + static big_ Test(...); + static From Create(); +}; +} // namespace internal + +// Inherits from true_type if From is convertible to To, false_type otherwise. +template +struct is_convertible + : integral_constant::Test( + internal::ConvertHelper::Create())) + == sizeof(small_)> { +}; +#endif + +_END_GOOGLE_NAMESPACE_ + +// Right now these macros are no-ops, and mostly just document the fact +// these types are PODs, for human use. They may be made more contentful +// later. The typedef is just to make it legal to put a semicolon after +// these macros. +#define DECLARE_POD(TypeName) typedef int Dummy_Type_For_DECLARE_POD +#define DECLARE_NESTED_POD(TypeName) DECLARE_POD(TypeName) +#define PROPAGATE_POD_FROM_TEMPLATE_ARGUMENT(TemplateName) \ + typedef int Dummy_Type_For_PROPAGATE_POD_FROM_TEMPLATE_ARGUMENT +#define ENFORCE_POD(TypeName) typedef int Dummy_Type_For_ENFORCE_POD + +#endif // BASE_TYPE_TRAITS_H_ diff --git a/src/includes/aggregate_tests.h b/src/includes/aggregate_tests.h index de7fdec..ea793d3 100644 --- a/src/includes/aggregate_tests.h +++ b/src/includes/aggregate_tests.h @@ -1,12 +1,8 @@ #ifndef ATESTS_H #define ATESTS_H -#include -#include "./tests.h" -#include -#include -#include -#include +#include +#include "./tests.h" vector sizes = { 50000, 100000, 150000, 200000, 250000, 300000, 350000, 400000, 500000, diff --git a/src/includes/prepare.h b/src/includes/prepare.h index bf2f860..dd80e99 100644 --- a/src/includes/prepare.h +++ b/src/includes/prepare.h @@ -7,8 +7,8 @@ // hashmaps and hash #include -#include -#include +#include "3thparty/sparsehash/sparse_hash_map" +#include "3thparty/sparsehash/dense_hash_map" #include "./3thparty/abseil-cpp/absl/container/node_hash_map.h" #include "./3thparty/abseil-cpp/absl/container/flat_hash_map.h" #include "./3thparty/abseil-cpp/absl/hash/hash.h" @@ -24,29 +24,22 @@ #include "./3thparty/parallel_hashmap/phmap.h" #include "./3thparty/emilib/hash_map.hpp" #include "3thparty/robinhood/robin_hood.h" +#include using std::string; using absl::Hash; - -// since my testing is based on this hashmap, this one doesn't need no prep -void prepare(std::unordered_map& map,int size){ - map.reserve(size); -} -void prepare(std::unordered_map& map,int size){ +template +void prepare(T& map, int size){ map.reserve(size); } - -// goooooogle void prepare(google::sparse_hash_map& map, int size){ - map.set_deleted_key(0); + map.set_deleted_key(-1); } void prepare(google::sparse_hash_map& map, int size){ - map.set_deleted_key(""); + map.set_deleted_key("a"); } - - void prepare(google::dense_hash_map& map, int size){ map.set_empty_key(0); map.set_deleted_key(-1); @@ -56,124 +49,4 @@ void prepare(google::dense_hash_map& map, int size){ map.set_empty_key(""); } -//absl -void prepare(absl::node_hash_map& map, int size){ - map.reserve(size); -} -void prepare(absl::node_hash_map& map, int size){ - map.reserve(size); -} - - -void prepare(absl::flat_hash_map& map, int size){ - map.reserve(size); -} -void prepare(absl::flat_hash_map& map, int size){ - map.reserve(size); -} - -// tessil -void prepare(tsl::sparse_map& map, int size){ - map.reserve(size); -} -void prepare(tsl::sparse_map& map, int size){ - map.reserve(size); -} - -void prepare(tsl::array_map& map, int size){ - // map.reserve(size); -} -void prepare(tsl::array_map& map, int size){ - // map.reserve(size); -} - -void prepare(tsl::ordered_map& map, int size){ - map.reserve(size); -} -void prepare(tsl::ordered_map& map, int size){ - map.reserve(size); -} - -void prepare(tsl::robin_map& map, int size){ - map.reserve(size); -} -void prepare(tsl::robin_map& map, int size){ - map.reserve(size); -} - -void prepare(tsl::hopscotch_map& map, int size){ - map.reserve(size); -} -void prepare(tsl::hopscotch_map& map, int size){ - map.reserve(size); -} - -// booooooost -void prepare(boost::unordered_map& map, int size){ - map.reserve(size); -} -void prepare(boost::unordered_map& map, int size){ - map.reserve(size); -} - -// skarupke's maps -void prepare(ska::bytell_hash_map& map, int size){ - map.reserve(size); -} -void prepare(ska::bytell_hash_map& map, int size){ - map.reserve(size); -} - -void prepare(ska::flat_hash_map& map, int size){ - map.reserve(size); -} -void prepare(ska::flat_hash_map& map, int size){ - map.reserve(size); -} - -void prepare(ska::unordered_map& map, int size){ - map.reserve(size); -} -void prepare(ska::unordered_map& map, int size){ - map.reserve(size); -} -// Gregory Popovitch' paralel hashmaps -void prepare(phmap::parallel_flat_hash_map& map, int size){ - map.reserve(size); -} -void prepare(phmap::parallel_flat_hash_map& map, int size){ - map.reserve(size); -} - -void prepare(phmap::parallel_node_hash_map& map, int size){ - map.reserve(size); -} -void prepare(phmap::parallel_node_hash_map& map, int size){ - map.reserve(size); -} - -// emilib (Emil Ernerfeld library) -void prepare(emilib::HashMap& map, int size){ - map.reserve(size); -} -void prepare(emilib::HashMap& map, int size){ - map.reserve(size); -} -// martin robinhood -void prepare(robin_hood::unordered_flat_map& map, int size){ - map.reserve(size); -} -void prepare(robin_hood::unordered_flat_map& map, int size){ - map.reserve(size); -} - -void prepare(robin_hood::unordered_node_map& map, int size){ - map.reserve(size); -} -void prepare(robin_hood::unordered_node_map& map, int size){ - map.reserve(size); -} - - - #endif /* TESTS_H */ \ No newline at end of file diff --git a/src/includes/tests.h b/src/includes/tests.h index d41cf2b..b2c3f9a 100644 --- a/src/includes/tests.h +++ b/src/includes/tests.h @@ -2,10 +2,12 @@ #ifndef TESTS_H #define TESTS_H +#include #include #include #include #include +#include // own #include "./generator.h" @@ -17,8 +19,6 @@ using std::vector; using std::cout; - - template vector int_test(T map, int size){ vector results; // insert, lookup, unsuccesful lookup, delete times diff --git a/main.cpp b/src/main.cpp similarity index 60% rename from main.cpp rename to src/main.cpp index 40bbd52..61aff2e 100644 --- a/main.cpp +++ b/src/main.cpp @@ -1,14 +1,11 @@ -#include -#include +#include "./includes/aggregate_tests.h" -#include "./src/includes/aggregate_tests.h" - -// // we can use to switch the map implementations to that +// // we can use typedefs and switch statements for the map implementations // // we can add some cli handling so we can specify which maps to tests (or all) -// typedef std::unordered_map intmap; -// typedef std::unordered_map stringmap; -// +typedef std::unordered_map intmap; +typedef std::unordered_map stringmap; + // // google sparse // typedef google::sparse_hash_map intmap; // typedef google::sparse_hash_map stringmap; @@ -18,8 +15,8 @@ // typedef google::dense_hash_map stringmap; // // // abseil nodehashmap -typedef absl::node_hash_map intmap; -typedef absl::node_hash_map stringmap; +// typedef absl::node_hash_map intmap; +// typedef absl::node_hash_map stringmap; // // // flat hashmap // typedef absl::flat_hash_map intmap; @@ -74,43 +71,44 @@ typedef absl::node_hash_map stringmap; // // martin flat map // typedef robin_hood::unordered_flat_map intmap; // typedef robin_hood::unordered_flat_map stringmap; -// // martin flat map +// martin flat map // typedef robin_hood::unordered_node_map intmap; // typedef robin_hood::unordered_node_map stringmap; +/* test takes 2hrs for 30 runs for unordered_map hashmap +// if the other maps have about the same operation times, it'll +// take around ~40hrs total ************ +// priorities are that the interface must be the same/similar to unordered_map +// and that we don't have to jump through hoops to get it to work +1. Google dense_hash_map [y] https://github.com/sparsehash/sparsehash +2. Google sparse_hash_map [y] +3. abseil node_hash_map [y] https://abseil.io/docs/cpp/tools/cmake-installs +4. abseil flat_hash_map [y] +5. Tessil/sparse-map/ [y] header only implementation for all tessil +6. Tessil/hopscotch-map[y] +7. tessil/robin-map[y] [y] +8. Boost unordered_map [y] just install boost with your package manager +9. skarupke/flat_hash_map [y] header only implementation +10. skarupke /bytell_hash_map [y] +11. skarupke/unordered_map [y] +12. greg7mdp/parallel-hashmap/paralel_flat [y] header only +13. greg7mdp/parallel-hashmap/paralel_node [y] +17. emilk/emilib emilib::hashmap [y] header only +18. martinus robin_hood::unordered_node_map [y] +19. martinus/robin_hood/ flatmap [y] + +3. folly F14ValueMap // lots of dependencies, even tho it really seems cool +4. folly F14NodeMap +5. Tessil/ordered-map [n] something is wrong with this one, verrrrry slow +6. Tessil/array-hash[n] (not with a small modification of the insert function) +*/ + int main() { time_point start_test = steady_clock::now(); - // string_test(stringmap{}, 1); // process gets killed for sizes >35000 int_test_aggregate(intmap{}, 2); string_test_aggregate(stringmap{}, 2); time_point end_test = steady_clock::now(); std::cout << "\n\n 30 runs for all tests for 1 map: " << duration_cast(end_test-start_test).count() << " seconds\n\n"; - - // test takes 2hrs for 30 runs for one hashmap - /* if the other maps have about the same operation times ************ - // possible maps to bench. priorities are that the interface must be the same/similar to unordered_map - // and that we don't have to jump through hoops to get it to work - 1. Google dense_hash_map [y] https://github.com/sparsehash/sparsehash - 2. Google sparse_hash_map [y] - 3. abseil node_hash_map [y] https://abseil.io/docs/cpp/tools/cmake-installs - 4. abseil flat_hash_map [y] - 5. Tessil/sparse-map/ [y] header only implementation for all tessil - 6. Tessil/hopscotch-map[y] - 7. tessil/robin-map[y] [y] - 8. Boost unordered_map [y] just install boost with your package manager - 9. skarupke/flat_hash_map [y] header only implementation - 10. skarupke /bytell_hash_map [y] - 11. skarupke/unordered_map [y] - 12. greg7mdp/parallel-hashmap/paralel_flat [y] header only - 13. greg7mdp/parallel-hashmap/paralel_node [y] - 17. emilk/emilib emilib::hashmap [y] header only - 18. martinus robin_hood::unordered_node_map [y] - 19. martinus/robin_hood/ flatmap [y] - 3. folly F14ValueMap - 4. folly F14NodeMap - 5. Tessil/ordered-map [n] something is wrong with this one, verrrrry slow - 6. Tessil/array-hash[n] (not with a small modification of the insert function) - */ - + }