1 // Protocol Buffers - Google's data interchange format 2 // Copyright 2008 Google Inc. All rights reserved. 3 // https://developers.google.com/protocol-buffers/ 4 // 5 // Redistribution and use in source and binary forms, with or without 6 // modification, are permitted provided that the following conditions are 7 // met: 8 // 9 // * Redistributions of source code must retain the above copyright 10 // notice, this list of conditions and the following disclaimer. 11 // * Redistributions in binary form must reproduce the above 12 // copyright notice, this list of conditions and the following disclaimer 13 // in the documentation and/or other materials provided with the 14 // distribution. 15 // * Neither the name of Google Inc. nor the names of its 16 // contributors may be used to endorse or promote products derived from 17 // this software without specific prior written permission. 18 // 19 // THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 20 // "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 21 // LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 22 // A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 23 // OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 24 // SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 25 // LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 26 // DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 27 // THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 28 // (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 29 // OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 30 31 // This file defines the map container and its helpers to support protobuf maps. 32 // 33 // The Map and MapIterator types are provided by this header file. 34 // Please avoid using other types defined here, unless they are public 35 // types within Map or MapIterator, such as Map::value_type. 36 37 #ifndef GOOGLE_PROTOBUF_MAP_H__ 38 #define GOOGLE_PROTOBUF_MAP_H__ 39 40 #include <functional> 41 #include <initializer_list> 42 #include <iterator> 43 #include <limits> // To support Visual Studio 2008 44 #include <map> 45 #include <string> 46 #include <type_traits> 47 #include <utility> 48 49 #if defined(__cpp_lib_string_view) 50 #include <string_view> 51 #endif // defined(__cpp_lib_string_view) 52 53 #include <google/protobuf/stubs/common.h> 54 #include <google/protobuf/arena.h> 55 #include <google/protobuf/generated_enum_util.h> 56 #include <google/protobuf/map_type_handler.h> 57 #include <google/protobuf/stubs/hash.h> 58 59 #ifdef SWIG 60 #error "You cannot SWIG proto headers" 61 #endif 62 63 #include <google/protobuf/port_def.inc> 64 65 namespace google { 66 namespace protobuf { 67 68 template <typename Key, typename T> 69 class Map; 70 71 class MapIterator; 72 73 template <typename Enum> 74 struct is_proto_enum; 75 76 namespace internal { 77 template <typename Derived, typename Key, typename T, 78 WireFormatLite::FieldType key_wire_type, 79 WireFormatLite::FieldType value_wire_type, int default_enum_value> 80 class MapFieldLite; 81 82 template <typename Derived, typename Key, typename T, 83 WireFormatLite::FieldType key_wire_type, 84 WireFormatLite::FieldType value_wire_type, int default_enum_value> 85 class MapField; 86 87 template <typename Key, typename T> 88 class TypeDefinedMapFieldBase; 89 90 class DynamicMapField; 91 92 class GeneratedMessageReflection; 93 94 // re-implement std::allocator to use arena allocator for memory allocation. 95 // Used for Map implementation. Users should not use this class 96 // directly. 97 template <typename U> 98 class MapAllocator { 99 public: 100 using value_type = U; 101 using pointer = value_type*; 102 using const_pointer = const value_type*; 103 using reference = value_type&; 104 using const_reference = const value_type&; 105 using size_type = size_t; 106 using difference_type = ptrdiff_t; 107 MapAllocator()108 MapAllocator() : arena_(nullptr) {} MapAllocator(Arena * arena)109 explicit MapAllocator(Arena* arena) : arena_(arena) {} 110 template <typename X> MapAllocator(const MapAllocator<X> & allocator)111 MapAllocator(const MapAllocator<X>& allocator) // NOLINT(runtime/explicit) 112 : arena_(allocator.arena()) {} 113 114 pointer allocate(size_type n, const void* /* hint */ = nullptr) { 115 // If arena is not given, malloc needs to be called which doesn't 116 // construct element object. 117 if (arena_ == nullptr) { 118 return static_cast<pointer>(::operator new(n * sizeof(value_type))); 119 } else { 120 return reinterpret_cast<pointer>( 121 Arena::CreateArray<uint8>(arena_, n * sizeof(value_type))); 122 } 123 } 124 deallocate(pointer p,size_type n)125 void deallocate(pointer p, size_type n) { 126 if (arena_ == nullptr) { 127 #if defined(__GXX_DELETE_WITH_SIZE__) || defined(__cpp_sized_deallocation) 128 ::operator delete(p, n * sizeof(value_type)); 129 #else 130 (void)n; 131 ::operator delete(p); 132 #endif 133 } 134 } 135 136 #if __cplusplus >= 201103L && !defined(GOOGLE_PROTOBUF_OS_APPLE) && \ 137 !defined(GOOGLE_PROTOBUF_OS_NACL) && \ 138 !defined(GOOGLE_PROTOBUF_OS_EMSCRIPTEN) 139 template <class NodeType, class... Args> construct(NodeType * p,Args &&...args)140 void construct(NodeType* p, Args&&... args) { 141 // Clang 3.6 doesn't compile static casting to void* directly. (Issue 142 // #1266) According C++ standard 5.2.9/1: "The static_cast operator shall 143 // not cast away constness". So first the maybe const pointer is casted to 144 // const void* and after the const void* is const casted. 145 new (const_cast<void*>(static_cast<const void*>(p))) 146 NodeType(std::forward<Args>(args)...); 147 } 148 149 template <class NodeType> destroy(NodeType * p)150 void destroy(NodeType* p) { 151 p->~NodeType(); 152 } 153 #else construct(pointer p,const_reference t)154 void construct(pointer p, const_reference t) { new (p) value_type(t); } 155 destroy(pointer p)156 void destroy(pointer p) { p->~value_type(); } 157 #endif 158 159 template <typename X> 160 struct rebind { 161 using other = MapAllocator<X>; 162 }; 163 164 template <typename X> 165 bool operator==(const MapAllocator<X>& other) const { 166 return arena_ == other.arena_; 167 } 168 169 template <typename X> 170 bool operator!=(const MapAllocator<X>& other) const { 171 return arena_ != other.arena_; 172 } 173 174 // To support Visual Studio 2008 max_size()175 size_type max_size() const { 176 // parentheses around (std::...:max) prevents macro warning of max() 177 return (std::numeric_limits<size_type>::max)(); 178 } 179 180 // To support gcc-4.4, which does not properly 181 // support templated friend classes arena()182 Arena* arena() const { return arena_; } 183 184 private: 185 using DestructorSkippable_ = void; 186 Arena* const arena_; 187 }; 188 189 template <typename T> 190 using KeyForTree = 191 typename std::conditional<std::is_scalar<T>::value, T, 192 std::reference_wrapper<const T>>::type; 193 194 // Default case: Not transparent. 195 // We use std::hash<key_type>/std::less<key_type> and all the lookup functions 196 // only accept `key_type`. 197 template <typename key_type> 198 struct TransparentSupport { 199 using hash = std::hash<key_type>; 200 using less = std::less<key_type>; 201 EqualsTransparentSupport202 static bool Equals(const key_type& a, const key_type& b) { return a == b; } 203 204 template <typename K> 205 using key_arg = key_type; 206 }; 207 208 #if defined(__cpp_lib_string_view) 209 // If std::string_view is available, we add transparent support for std::string 210 // keys. We use std::hash<std::string_view> as it supports the input types we 211 // care about. The lookup functions accept arbitrary `K`. This will include any 212 // key type that is convertible to std::string_view. 213 template <> 214 struct TransparentSupport<std::string> { 215 static std::string_view ImplicitConvert(std::string_view str) { return str; } 216 // If the element is not convertible to std::string_view, try to convert to 217 // std::string first. 218 // The template makes this overload lose resolution when both have the same 219 // rank otherwise. 220 template <typename = void> 221 static std::string_view ImplicitConvert(const std::string& str) { 222 return str; 223 } 224 225 struct hash : private std::hash<std::string_view> { 226 using is_transparent = void; 227 228 template <typename T> 229 size_t operator()(const T& str) const { 230 return base()(ImplicitConvert(str)); 231 } 232 233 private: 234 const std::hash<std::string_view>& base() const { return *this; } 235 }; 236 struct less { 237 using is_transparent = void; 238 239 template <typename T, typename U> 240 bool operator()(const T& t, const U& u) const { 241 return ImplicitConvert(t) < ImplicitConvert(u); 242 } 243 }; 244 245 template <typename T, typename U> 246 static bool Equals(const T& t, const U& u) { 247 return ImplicitConvert(t) == ImplicitConvert(u); 248 } 249 250 template <typename K> 251 using key_arg = K; 252 }; 253 #endif // defined(__cpp_lib_string_view) 254 255 } // namespace internal 256 257 // This is the class for Map's internal value_type. Instead of using 258 // std::pair as value_type, we use this class which provides us more control of 259 // its process of construction and destruction. 260 template <typename Key, typename T> 261 struct MapPair { 262 using first_type = const Key; 263 using second_type = T; 264 265 MapPair(const Key& other_first, const T& other_second) 266 : first(other_first), second(other_second) {} 267 explicit MapPair(const Key& other_first) : first(other_first), second() {} 268 MapPair(const MapPair& other) : first(other.first), second(other.second) {} 269 270 ~MapPair() {} 271 272 // Implicitly convertible to std::pair of compatible types. 273 template <typename T1, typename T2> 274 operator std::pair<T1, T2>() const { // NOLINT(runtime/explicit) 275 return std::pair<T1, T2>(first, second); 276 } 277 278 const Key first; 279 T second; 280 281 private: 282 friend class Arena; 283 friend class Map<Key, T>; 284 }; 285 286 // Map is an associative container type used to store protobuf map 287 // fields. Each Map instance may or may not use a different hash function, a 288 // different iteration order, and so on. E.g., please don't examine 289 // implementation details to decide if the following would work: 290 // Map<int, int> m0, m1; 291 // m0[0] = m1[0] = m0[1] = m1[1] = 0; 292 // assert(m0.begin()->first == m1.begin()->first); // Bug! 293 // 294 // Map's interface is similar to std::unordered_map, except that Map is not 295 // designed to play well with exceptions. 296 template <typename Key, typename T> 297 class Map { 298 public: 299 using key_type = Key; 300 using mapped_type = T; 301 using value_type = MapPair<Key, T>; 302 303 using pointer = value_type*; 304 using const_pointer = const value_type*; 305 using reference = value_type&; 306 using const_reference = const value_type&; 307 308 using size_type = size_t; 309 using hasher = typename internal::TransparentSupport<Key>::hash; 310 311 Map() : arena_(nullptr), default_enum_value_(0) { Init(); } 312 explicit Map(Arena* arena) : arena_(arena), default_enum_value_(0) { Init(); } 313 314 Map(const Map& other) 315 : arena_(nullptr), default_enum_value_(other.default_enum_value_) { 316 Init(); 317 insert(other.begin(), other.end()); 318 } 319 320 Map(Map&& other) noexcept : Map() { 321 if (other.arena_) { 322 *this = other; 323 } else { 324 swap(other); 325 } 326 } 327 Map& operator=(Map&& other) noexcept { 328 if (this != &other) { 329 if (arena_ != other.arena_) { 330 *this = other; 331 } else { 332 swap(other); 333 } 334 } 335 return *this; 336 } 337 338 template <class InputIt> 339 Map(const InputIt& first, const InputIt& last) 340 : arena_(nullptr), default_enum_value_(0) { 341 Init(); 342 insert(first, last); 343 } 344 345 ~Map() { 346 if (arena_ == nullptr) { 347 clear(); 348 delete elements_; 349 } 350 } 351 352 private: 353 void Init() { elements_ = Arena::CreateMessage<InnerMap>(arena_, 0); } 354 355 using Allocator = internal::MapAllocator<void*>; 356 357 // InnerMap is a generic hash-based map. It doesn't contain any 358 // protocol-buffer-specific logic. It is a chaining hash map with the 359 // additional feature that some buckets can be converted to use an ordered 360 // container. This ensures O(lg n) bounds on find, insert, and erase, while 361 // avoiding the overheads of ordered containers most of the time. 362 // 363 // The implementation doesn't need the full generality of unordered_map, 364 // and it doesn't have it. More bells and whistles can be added as needed. 365 // Some implementation details: 366 // 1. The hash function has type hasher and the equality function 367 // equal_to<Key>. We inherit from hasher to save space 368 // (empty-base-class optimization). 369 // 2. The number of buckets is a power of two. 370 // 3. Buckets are converted to trees in pairs: if we convert bucket b then 371 // buckets b and b^1 will share a tree. Invariant: buckets b and b^1 have 372 // the same non-null value iff they are sharing a tree. (An alternative 373 // implementation strategy would be to have a tag bit per bucket.) 374 // 4. As is typical for hash_map and such, the Keys and Values are always 375 // stored in linked list nodes. Pointers to elements are never invalidated 376 // until the element is deleted. 377 // 5. The trees' payload type is pointer to linked-list node. Tree-converting 378 // a bucket doesn't copy Key-Value pairs. 379 // 6. Once we've tree-converted a bucket, it is never converted back. However, 380 // the items a tree contains may wind up assigned to trees or lists upon a 381 // rehash. 382 // 7. The code requires no C++ features from C++14 or later. 383 // 8. Mutations to a map do not invalidate the map's iterators, pointers to 384 // elements, or references to elements. 385 // 9. Except for erase(iterator), any non-const method can reorder iterators. 386 // 10. InnerMap uses KeyForTree<Key> when using the Tree representation, which 387 // is either `Key`, if Key is a scalar, or `reference_wrapper<const Key>` 388 // otherwise. This avoids unncessary copies of string keys, for example. 389 class InnerMap : private hasher { 390 public: 391 explicit InnerMap(size_type n) : InnerMap(nullptr, n) {} 392 InnerMap(Arena* arena, size_type n) 393 : hasher(), 394 num_elements_(0), 395 seed_(Seed()), 396 table_(nullptr), 397 alloc_(arena) { 398 n = TableSize(n); 399 table_ = CreateEmptyTable(n); 400 num_buckets_ = index_of_first_non_null_ = n; 401 } 402 403 ~InnerMap() { 404 if (table_ != nullptr) { 405 clear(); 406 Dealloc<void*>(table_, num_buckets_); 407 } 408 } 409 410 private: 411 enum { kMinTableSize = 8 }; 412 413 // Linked-list nodes, as one would expect for a chaining hash table. 414 struct Node { 415 value_type kv; 416 Node* next; 417 }; 418 419 // Trees. The payload type is a copy of Key, so that we can query the tree 420 // with Keys that are not in any particular data structure. 421 // The value is a void* pointing to Node. We use void* instead of Node* to 422 // avoid code bloat. That way there is only one instantiation of the tree 423 // class per key type. 424 using TreeAllocator = typename Allocator::template rebind< 425 std::pair<const internal::KeyForTree<Key>, void*>>::other; 426 using Tree = std::map<internal::KeyForTree<Key>, void*, 427 typename internal::TransparentSupport<Key>::less, 428 TreeAllocator>; 429 using TreeIterator = typename Tree::iterator; 430 431 static Node* NodeFromTreeIterator(TreeIterator it) { 432 return static_cast<Node*>(it->second); 433 } 434 435 // iterator and const_iterator are instantiations of iterator_base. 436 template <typename KeyValueType> 437 class iterator_base { 438 public: 439 using reference = KeyValueType&; 440 using pointer = KeyValueType*; 441 442 // Invariants: 443 // node_ is always correct. This is handy because the most common 444 // operations are operator* and operator-> and they only use node_. 445 // When node_ is set to a non-null value, all the other non-const fields 446 // are updated to be correct also, but those fields can become stale 447 // if the underlying map is modified. When those fields are needed they 448 // are rechecked, and updated if necessary. 449 iterator_base() : node_(nullptr), m_(nullptr), bucket_index_(0) {} 450 451 explicit iterator_base(const InnerMap* m) : m_(m) { 452 SearchFrom(m->index_of_first_non_null_); 453 } 454 455 // Any iterator_base can convert to any other. This is overkill, and we 456 // rely on the enclosing class to use it wisely. The standard "iterator 457 // can convert to const_iterator" is OK but the reverse direction is not. 458 template <typename U> 459 explicit iterator_base(const iterator_base<U>& it) 460 : node_(it.node_), m_(it.m_), bucket_index_(it.bucket_index_) {} 461 462 iterator_base(Node* n, const InnerMap* m, size_type index) 463 : node_(n), m_(m), bucket_index_(index) {} 464 465 iterator_base(TreeIterator tree_it, const InnerMap* m, size_type index) 466 : node_(NodeFromTreeIterator(tree_it)), m_(m), bucket_index_(index) { 467 // Invariant: iterators that use buckets with trees have an even 468 // bucket_index_. 469 GOOGLE_DCHECK_EQ(bucket_index_ % 2, 0u); 470 } 471 472 // Advance through buckets, looking for the first that isn't empty. 473 // If nothing non-empty is found then leave node_ == nullptr. 474 void SearchFrom(size_type start_bucket) { 475 GOOGLE_DCHECK(m_->index_of_first_non_null_ == m_->num_buckets_ || 476 m_->table_[m_->index_of_first_non_null_] != nullptr); 477 node_ = nullptr; 478 for (bucket_index_ = start_bucket; bucket_index_ < m_->num_buckets_; 479 bucket_index_++) { 480 if (m_->TableEntryIsNonEmptyList(bucket_index_)) { 481 node_ = static_cast<Node*>(m_->table_[bucket_index_]); 482 break; 483 } else if (m_->TableEntryIsTree(bucket_index_)) { 484 Tree* tree = static_cast<Tree*>(m_->table_[bucket_index_]); 485 GOOGLE_DCHECK(!tree->empty()); 486 node_ = NodeFromTreeIterator(tree->begin()); 487 break; 488 } 489 } 490 } 491 492 reference operator*() const { return node_->kv; } 493 pointer operator->() const { return &(operator*()); } 494 495 friend bool operator==(const iterator_base& a, const iterator_base& b) { 496 return a.node_ == b.node_; 497 } 498 friend bool operator!=(const iterator_base& a, const iterator_base& b) { 499 return a.node_ != b.node_; 500 } 501 502 iterator_base& operator++() { 503 if (node_->next == nullptr) { 504 TreeIterator tree_it; 505 const bool is_list = revalidate_if_necessary(&tree_it); 506 if (is_list) { 507 SearchFrom(bucket_index_ + 1); 508 } else { 509 GOOGLE_DCHECK_EQ(bucket_index_ & 1, 0u); 510 Tree* tree = static_cast<Tree*>(m_->table_[bucket_index_]); 511 if (++tree_it == tree->end()) { 512 SearchFrom(bucket_index_ + 2); 513 } else { 514 node_ = NodeFromTreeIterator(tree_it); 515 } 516 } 517 } else { 518 node_ = node_->next; 519 } 520 return *this; 521 } 522 523 iterator_base operator++(int /* unused */) { 524 iterator_base tmp = *this; 525 ++*this; 526 return tmp; 527 } 528 529 // Assumes node_ and m_ are correct and non-null, but other fields may be 530 // stale. Fix them as needed. Then return true iff node_ points to a 531 // Node in a list. If false is returned then *it is modified to be 532 // a valid iterator for node_. 533 bool revalidate_if_necessary(TreeIterator* it) { 534 GOOGLE_DCHECK(node_ != nullptr && m_ != nullptr); 535 // Force bucket_index_ to be in range. 536 bucket_index_ &= (m_->num_buckets_ - 1); 537 // Common case: the bucket we think is relevant points to node_. 538 if (m_->table_[bucket_index_] == static_cast<void*>(node_)) return true; 539 // Less common: the bucket is a linked list with node_ somewhere in it, 540 // but not at the head. 541 if (m_->TableEntryIsNonEmptyList(bucket_index_)) { 542 Node* l = static_cast<Node*>(m_->table_[bucket_index_]); 543 while ((l = l->next) != nullptr) { 544 if (l == node_) { 545 return true; 546 } 547 } 548 } 549 // Well, bucket_index_ still might be correct, but probably 550 // not. Revalidate just to be sure. This case is rare enough that we 551 // don't worry about potential optimizations, such as having a custom 552 // find-like method that compares Node* instead of the key. 553 iterator_base i(m_->find(node_->kv.first, it)); 554 bucket_index_ = i.bucket_index_; 555 return m_->TableEntryIsList(bucket_index_); 556 } 557 558 Node* node_; 559 const InnerMap* m_; 560 size_type bucket_index_; 561 }; 562 563 public: 564 using iterator = iterator_base<value_type>; 565 using const_iterator = iterator_base<const value_type>; 566 567 iterator begin() { return iterator(this); } 568 iterator end() { return iterator(); } 569 const_iterator begin() const { return const_iterator(this); } 570 const_iterator end() const { return const_iterator(); } 571 572 void clear() { 573 for (size_type b = 0; b < num_buckets_; b++) { 574 if (TableEntryIsNonEmptyList(b)) { 575 Node* node = static_cast<Node*>(table_[b]); 576 table_[b] = nullptr; 577 do { 578 Node* next = node->next; 579 DestroyNode(node); 580 node = next; 581 } while (node != nullptr); 582 } else if (TableEntryIsTree(b)) { 583 Tree* tree = static_cast<Tree*>(table_[b]); 584 GOOGLE_DCHECK(table_[b] == table_[b + 1] && (b & 1) == 0); 585 table_[b] = table_[b + 1] = nullptr; 586 typename Tree::iterator tree_it = tree->begin(); 587 do { 588 Node* node = NodeFromTreeIterator(tree_it); 589 typename Tree::iterator next = tree_it; 590 ++next; 591 tree->erase(tree_it); 592 DestroyNode(node); 593 tree_it = next; 594 } while (tree_it != tree->end()); 595 DestroyTree(tree); 596 b++; 597 } 598 } 599 num_elements_ = 0; 600 index_of_first_non_null_ = num_buckets_; 601 } 602 603 const hasher& hash_function() const { return *this; } 604 605 static size_type max_size() { 606 return static_cast<size_type>(1) << (sizeof(void**) >= 8 ? 60 : 28); 607 } 608 size_type size() const { return num_elements_; } 609 bool empty() const { return size() == 0; } 610 611 template <typename K> 612 iterator find(const K& k) { 613 return iterator(FindHelper(k).first); 614 } 615 616 // Insert the key into the map, if not present. In that case, the value will 617 // be value initialized. 618 std::pair<iterator, bool> insert(const Key& k) { 619 std::pair<const_iterator, size_type> p = FindHelper(k); 620 // Case 1: key was already present. 621 if (p.first.node_ != nullptr) 622 return std::make_pair(iterator(p.first), false); 623 // Case 2: insert. 624 if (ResizeIfLoadIsOutOfRange(num_elements_ + 1)) { 625 p = FindHelper(k); 626 } 627 const size_type b = p.second; // bucket number 628 Node* node; 629 if (alloc_.arena() == nullptr) { 630 node = new Node{value_type(k), nullptr}; 631 } else { 632 node = Alloc<Node>(1); 633 Arena::CreateInArenaStorage(const_cast<Key*>(&node->kv.first), 634 alloc_.arena(), k); 635 Arena::CreateInArenaStorage(&node->kv.second, alloc_.arena()); 636 } 637 638 iterator result = InsertUnique(b, node); 639 ++num_elements_; 640 return std::make_pair(result, true); 641 } 642 643 value_type& operator[](const Key& k) { return *insert(k).first; } 644 645 void erase(iterator it) { 646 GOOGLE_DCHECK_EQ(it.m_, this); 647 typename Tree::iterator tree_it; 648 const bool is_list = it.revalidate_if_necessary(&tree_it); 649 size_type b = it.bucket_index_; 650 Node* const item = it.node_; 651 if (is_list) { 652 GOOGLE_DCHECK(TableEntryIsNonEmptyList(b)); 653 Node* head = static_cast<Node*>(table_[b]); 654 head = EraseFromLinkedList(item, head); 655 table_[b] = static_cast<void*>(head); 656 } else { 657 GOOGLE_DCHECK(TableEntryIsTree(b)); 658 Tree* tree = static_cast<Tree*>(table_[b]); 659 tree->erase(tree_it); 660 if (tree->empty()) { 661 // Force b to be the minimum of b and b ^ 1. This is important 662 // only because we want index_of_first_non_null_ to be correct. 663 b &= ~static_cast<size_type>(1); 664 DestroyTree(tree); 665 table_[b] = table_[b + 1] = nullptr; 666 } 667 } 668 DestroyNode(item); 669 --num_elements_; 670 if (PROTOBUF_PREDICT_FALSE(b == index_of_first_non_null_)) { 671 while (index_of_first_non_null_ < num_buckets_ && 672 table_[index_of_first_non_null_] == nullptr) { 673 ++index_of_first_non_null_; 674 } 675 } 676 } 677 678 private: 679 const_iterator find(const Key& k, TreeIterator* it) const { 680 return FindHelper(k, it).first; 681 } 682 template <typename K> 683 std::pair<const_iterator, size_type> FindHelper(const K& k) const { 684 return FindHelper(k, nullptr); 685 } 686 template <typename K> 687 std::pair<const_iterator, size_type> FindHelper(const K& k, 688 TreeIterator* it) const { 689 size_type b = BucketNumber(k); 690 if (TableEntryIsNonEmptyList(b)) { 691 Node* node = static_cast<Node*>(table_[b]); 692 do { 693 if (internal::TransparentSupport<Key>::Equals(node->kv.first, k)) { 694 return std::make_pair(const_iterator(node, this, b), b); 695 } else { 696 node = node->next; 697 } 698 } while (node != nullptr); 699 } else if (TableEntryIsTree(b)) { 700 GOOGLE_DCHECK_EQ(table_[b], table_[b ^ 1]); 701 b &= ~static_cast<size_t>(1); 702 Tree* tree = static_cast<Tree*>(table_[b]); 703 auto tree_it = tree->find(k); 704 if (tree_it != tree->end()) { 705 if (it != nullptr) *it = tree_it; 706 return std::make_pair(const_iterator(tree_it, this, b), b); 707 } 708 } 709 return std::make_pair(end(), b); 710 } 711 712 // Insert the given Node in bucket b. If that would make bucket b too big, 713 // and bucket b is not a tree, create a tree for buckets b and b^1 to share. 714 // Requires count(*KeyPtrFromNodePtr(node)) == 0 and that b is the correct 715 // bucket. num_elements_ is not modified. 716 iterator InsertUnique(size_type b, Node* node) { 717 GOOGLE_DCHECK(index_of_first_non_null_ == num_buckets_ || 718 table_[index_of_first_non_null_] != nullptr); 719 // In practice, the code that led to this point may have already 720 // determined whether we are inserting into an empty list, a short list, 721 // or whatever. But it's probably cheap enough to recompute that here; 722 // it's likely that we're inserting into an empty or short list. 723 iterator result; 724 GOOGLE_DCHECK(find(node->kv.first) == end()); 725 if (TableEntryIsEmpty(b)) { 726 result = InsertUniqueInList(b, node); 727 } else if (TableEntryIsNonEmptyList(b)) { 728 if (PROTOBUF_PREDICT_FALSE(TableEntryIsTooLong(b))) { 729 TreeConvert(b); 730 result = InsertUniqueInTree(b, node); 731 GOOGLE_DCHECK_EQ(result.bucket_index_, b & ~static_cast<size_type>(1)); 732 } else { 733 // Insert into a pre-existing list. This case cannot modify 734 // index_of_first_non_null_, so we skip the code to update it. 735 return InsertUniqueInList(b, node); 736 } 737 } else { 738 // Insert into a pre-existing tree. This case cannot modify 739 // index_of_first_non_null_, so we skip the code to update it. 740 return InsertUniqueInTree(b, node); 741 } 742 // parentheses around (std::min) prevents macro expansion of min(...) 743 index_of_first_non_null_ = 744 (std::min)(index_of_first_non_null_, result.bucket_index_); 745 return result; 746 } 747 748 // Returns whether we should insert after the head of the list. For 749 // non-optimized builds, we randomly decide whether to insert right at the 750 // head of the list or just after the head. This helps add a little bit of 751 // non-determinism to the map ordering. 752 bool ShouldInsertAfterHead(void* node) { 753 #ifdef NDEBUG 754 return false; 755 #else 756 // Doing modulo with a prime mixes the bits more. 757 return (reinterpret_cast<uintptr_t>(node) ^ seed_) % 13 > 6; 758 #endif 759 } 760 761 // Helper for InsertUnique. Handles the case where bucket b is a 762 // not-too-long linked list. 763 iterator InsertUniqueInList(size_type b, Node* node) { 764 if (table_[b] != nullptr && ShouldInsertAfterHead(node)) { 765 Node* first = static_cast<Node*>(table_[b]); 766 node->next = first->next; 767 first->next = node; 768 return iterator(node, this, b); 769 } 770 771 node->next = static_cast<Node*>(table_[b]); 772 table_[b] = static_cast<void*>(node); 773 return iterator(node, this, b); 774 } 775 776 // Helper for InsertUnique. Handles the case where bucket b points to a 777 // Tree. 778 iterator InsertUniqueInTree(size_type b, Node* node) { 779 GOOGLE_DCHECK_EQ(table_[b], table_[b ^ 1]); 780 // Maintain the invariant that node->next is null for all Nodes in Trees. 781 node->next = nullptr; 782 return iterator( 783 static_cast<Tree*>(table_[b])->insert({node->kv.first, node}).first, 784 this, b & ~static_cast<size_t>(1)); 785 } 786 787 // Returns whether it did resize. Currently this is only used when 788 // num_elements_ increases, though it could be used in other situations. 789 // It checks for load too low as well as load too high: because any number 790 // of erases can occur between inserts, the load could be as low as 0 here. 791 // Resizing to a lower size is not always helpful, but failing to do so can 792 // destroy the expected big-O bounds for some operations. By having the 793 // policy that sometimes we resize down as well as up, clients can easily 794 // keep O(size()) = O(number of buckets) if they want that. 795 bool ResizeIfLoadIsOutOfRange(size_type new_size) { 796 const size_type kMaxMapLoadTimes16 = 12; // controls RAM vs CPU tradeoff 797 const size_type hi_cutoff = num_buckets_ * kMaxMapLoadTimes16 / 16; 798 const size_type lo_cutoff = hi_cutoff / 4; 799 // We don't care how many elements are in trees. If a lot are, 800 // we may resize even though there are many empty buckets. In 801 // practice, this seems fine. 802 if (PROTOBUF_PREDICT_FALSE(new_size >= hi_cutoff)) { 803 if (num_buckets_ <= max_size() / 2) { 804 Resize(num_buckets_ * 2); 805 return true; 806 } 807 } else if (PROTOBUF_PREDICT_FALSE(new_size <= lo_cutoff && 808 num_buckets_ > kMinTableSize)) { 809 size_type lg2_of_size_reduction_factor = 1; 810 // It's possible we want to shrink a lot here... size() could even be 0. 811 // So, estimate how much to shrink by making sure we don't shrink so 812 // much that we would need to grow the table after a few inserts. 813 const size_type hypothetical_size = new_size * 5 / 4 + 1; 814 while ((hypothetical_size << lg2_of_size_reduction_factor) < 815 hi_cutoff) { 816 ++lg2_of_size_reduction_factor; 817 } 818 size_type new_num_buckets = std::max<size_type>( 819 kMinTableSize, num_buckets_ >> lg2_of_size_reduction_factor); 820 if (new_num_buckets != num_buckets_) { 821 Resize(new_num_buckets); 822 return true; 823 } 824 } 825 return false; 826 } 827 828 // Resize to the given number of buckets. 829 void Resize(size_t new_num_buckets) { 830 GOOGLE_DCHECK_GE(new_num_buckets, kMinTableSize); 831 void** const old_table = table_; 832 const size_type old_table_size = num_buckets_; 833 num_buckets_ = new_num_buckets; 834 table_ = CreateEmptyTable(num_buckets_); 835 const size_type start = index_of_first_non_null_; 836 index_of_first_non_null_ = num_buckets_; 837 for (size_type i = start; i < old_table_size; i++) { 838 if (TableEntryIsNonEmptyList(old_table, i)) { 839 TransferList(old_table, i); 840 } else if (TableEntryIsTree(old_table, i)) { 841 TransferTree(old_table, i++); 842 } 843 } 844 Dealloc<void*>(old_table, old_table_size); 845 } 846 847 void TransferList(void* const* table, size_type index) { 848 Node* node = static_cast<Node*>(table[index]); 849 do { 850 Node* next = node->next; 851 InsertUnique(BucketNumber(node->kv.first), node); 852 node = next; 853 } while (node != nullptr); 854 } 855 856 void TransferTree(void* const* table, size_type index) { 857 Tree* tree = static_cast<Tree*>(table[index]); 858 typename Tree::iterator tree_it = tree->begin(); 859 do { 860 InsertUnique(BucketNumber(std::cref(tree_it->first).get()), 861 NodeFromTreeIterator(tree_it)); 862 } while (++tree_it != tree->end()); 863 DestroyTree(tree); 864 } 865 866 Node* EraseFromLinkedList(Node* item, Node* head) { 867 if (head == item) { 868 return head->next; 869 } else { 870 head->next = EraseFromLinkedList(item, head->next); 871 return head; 872 } 873 } 874 875 bool TableEntryIsEmpty(size_type b) const { 876 return TableEntryIsEmpty(table_, b); 877 } 878 bool TableEntryIsNonEmptyList(size_type b) const { 879 return TableEntryIsNonEmptyList(table_, b); 880 } 881 bool TableEntryIsTree(size_type b) const { 882 return TableEntryIsTree(table_, b); 883 } 884 bool TableEntryIsList(size_type b) const { 885 return TableEntryIsList(table_, b); 886 } 887 static bool TableEntryIsEmpty(void* const* table, size_type b) { 888 return table[b] == nullptr; 889 } 890 static bool TableEntryIsNonEmptyList(void* const* table, size_type b) { 891 return table[b] != nullptr && table[b] != table[b ^ 1]; 892 } 893 static bool TableEntryIsTree(void* const* table, size_type b) { 894 return !TableEntryIsEmpty(table, b) && 895 !TableEntryIsNonEmptyList(table, b); 896 } 897 static bool TableEntryIsList(void* const* table, size_type b) { 898 return !TableEntryIsTree(table, b); 899 } 900 901 void TreeConvert(size_type b) { 902 GOOGLE_DCHECK(!TableEntryIsTree(b) && !TableEntryIsTree(b ^ 1)); 903 Tree* tree = 904 Arena::Create<Tree>(alloc_.arena(), typename Tree::key_compare(), 905 typename Tree::allocator_type(alloc_)); 906 size_type count = CopyListToTree(b, tree) + CopyListToTree(b ^ 1, tree); 907 GOOGLE_DCHECK_EQ(count, tree->size()); 908 table_[b] = table_[b ^ 1] = static_cast<void*>(tree); 909 } 910 911 // Copy a linked list in the given bucket to a tree. 912 // Returns the number of things it copied. 913 size_type CopyListToTree(size_type b, Tree* tree) { 914 size_type count = 0; 915 Node* node = static_cast<Node*>(table_[b]); 916 while (node != nullptr) { 917 tree->insert({node->kv.first, node}); 918 ++count; 919 Node* next = node->next; 920 node->next = nullptr; 921 node = next; 922 } 923 return count; 924 } 925 926 // Return whether table_[b] is a linked list that seems awfully long. 927 // Requires table_[b] to point to a non-empty linked list. 928 bool TableEntryIsTooLong(size_type b) { 929 const size_type kMaxLength = 8; 930 size_type count = 0; 931 Node* node = static_cast<Node*>(table_[b]); 932 do { 933 ++count; 934 node = node->next; 935 } while (node != nullptr); 936 // Invariant: no linked list ever is more than kMaxLength in length. 937 GOOGLE_DCHECK_LE(count, kMaxLength); 938 return count >= kMaxLength; 939 } 940 941 template <typename K> 942 size_type BucketNumber(const K& k) const { 943 // We xor the hash value against the random seed so that we effectively 944 // have a random hash function. 945 uint64 h = hash_function()(k) ^ seed_; 946 947 // We use the multiplication method to determine the bucket number from 948 // the hash value. The constant kPhi (suggested by Knuth) is roughly 949 // (sqrt(5) - 1) / 2 * 2^64. 950 constexpr uint64 kPhi = uint64{0x9e3779b97f4a7c15}; 951 return ((kPhi * h) >> 32) & (num_buckets_ - 1); 952 } 953 954 // Return a power of two no less than max(kMinTableSize, n). 955 // Assumes either n < kMinTableSize or n is a power of two. 956 size_type TableSize(size_type n) { 957 return n < static_cast<size_type>(kMinTableSize) 958 ? static_cast<size_type>(kMinTableSize) 959 : n; 960 } 961 962 // Use alloc_ to allocate an array of n objects of type U. 963 template <typename U> 964 U* Alloc(size_type n) { 965 using alloc_type = typename Allocator::template rebind<U>::other; 966 return alloc_type(alloc_).allocate(n); 967 } 968 969 // Use alloc_ to deallocate an array of n objects of type U. 970 template <typename U> 971 void Dealloc(U* t, size_type n) { 972 using alloc_type = typename Allocator::template rebind<U>::other; 973 alloc_type(alloc_).deallocate(t, n); 974 } 975 976 void DestroyNode(Node* node) { 977 if (alloc_.arena() == nullptr) { 978 delete node; 979 } 980 } 981 982 void DestroyTree(Tree* tree) { 983 if (alloc_.arena() == nullptr) { 984 delete tree; 985 } 986 } 987 988 void** CreateEmptyTable(size_type n) { 989 GOOGLE_DCHECK(n >= kMinTableSize); 990 GOOGLE_DCHECK_EQ(n & (n - 1), 0); 991 void** result = Alloc<void*>(n); 992 memset(result, 0, n * sizeof(result[0])); 993 return result; 994 } 995 996 // Return a randomish value. 997 size_type Seed() const { 998 // We get a little bit of randomness from the address of the map. The 999 // lower bits are not very random, due to alignment, so we discard them 1000 // and shift the higher bits into their place. 1001 size_type s = reinterpret_cast<uintptr_t>(this) >> 12; 1002 #if defined(__x86_64__) && defined(__GNUC__) && \ 1003 !defined(GOOGLE_PROTOBUF_NO_RDTSC) 1004 uint32 hi, lo; 1005 asm("rdtsc" : "=a"(lo), "=d"(hi)); 1006 s += ((static_cast<uint64>(hi) << 32) | lo); 1007 #endif 1008 return s; 1009 } 1010 1011 friend class Arena; 1012 using InternalArenaConstructable_ = void; 1013 using DestructorSkippable_ = void; 1014 1015 size_type num_elements_; 1016 size_type num_buckets_; 1017 size_type seed_; 1018 size_type index_of_first_non_null_; 1019 void** table_; // an array with num_buckets_ entries 1020 Allocator alloc_; 1021 GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(InnerMap); 1022 }; // end of class InnerMap 1023 1024 template <typename LookupKey> 1025 using key_arg = typename internal::TransparentSupport< 1026 key_type>::template key_arg<LookupKey>; 1027 1028 public: 1029 // Iterators 1030 class const_iterator { 1031 using InnerIt = typename InnerMap::const_iterator; 1032 1033 public: 1034 using iterator_category = std::forward_iterator_tag; 1035 using value_type = typename Map::value_type; 1036 using difference_type = ptrdiff_t; 1037 using pointer = const value_type*; 1038 using reference = const value_type&; 1039 1040 const_iterator() {} 1041 explicit const_iterator(const InnerIt& it) : it_(it) {} 1042 1043 const_reference operator*() const { return *it_; } 1044 const_pointer operator->() const { return &(operator*()); } 1045 1046 const_iterator& operator++() { 1047 ++it_; 1048 return *this; 1049 } 1050 const_iterator operator++(int) { return const_iterator(it_++); } 1051 1052 friend bool operator==(const const_iterator& a, const const_iterator& b) { 1053 return a.it_ == b.it_; 1054 } 1055 friend bool operator!=(const const_iterator& a, const const_iterator& b) { 1056 return !(a == b); 1057 } 1058 1059 private: 1060 InnerIt it_; 1061 }; 1062 1063 class iterator { 1064 using InnerIt = typename InnerMap::iterator; 1065 1066 public: 1067 using iterator_category = std::forward_iterator_tag; 1068 using value_type = typename Map::value_type; 1069 using difference_type = ptrdiff_t; 1070 using pointer = value_type*; 1071 using reference = value_type&; 1072 1073 iterator() {} 1074 explicit iterator(const InnerIt& it) : it_(it) {} 1075 1076 reference operator*() const { return *it_; } 1077 pointer operator->() const { return &(operator*()); } 1078 1079 iterator& operator++() { 1080 ++it_; 1081 return *this; 1082 } 1083 iterator operator++(int) { return iterator(it_++); } 1084 1085 // Allow implicit conversion to const_iterator. 1086 operator const_iterator() const { // NOLINT(runtime/explicit) 1087 return const_iterator(typename InnerMap::const_iterator(it_)); 1088 } 1089 1090 friend bool operator==(const iterator& a, const iterator& b) { 1091 return a.it_ == b.it_; 1092 } 1093 friend bool operator!=(const iterator& a, const iterator& b) { 1094 return !(a == b); 1095 } 1096 1097 private: 1098 friend class Map; 1099 1100 InnerIt it_; 1101 }; 1102 1103 iterator begin() { return iterator(elements_->begin()); } 1104 iterator end() { return iterator(elements_->end()); } 1105 const_iterator begin() const { 1106 return const_iterator(iterator(elements_->begin())); 1107 } 1108 const_iterator end() const { 1109 return const_iterator(iterator(elements_->end())); 1110 } 1111 const_iterator cbegin() const { return begin(); } 1112 const_iterator cend() const { return end(); } 1113 1114 // Capacity 1115 size_type size() const { return elements_->size(); } 1116 bool empty() const { return size() == 0; } 1117 1118 // Element access 1119 T& operator[](const key_type& key) { return (*elements_)[key].second; } 1120 1121 template <typename K = key_type> 1122 const T& at(const key_arg<K>& key) const { 1123 const_iterator it = find(key); 1124 GOOGLE_CHECK(it != end()) << "key not found: " << static_cast<Key>(key); 1125 return it->second; 1126 } 1127 1128 template <typename K = key_type> 1129 T& at(const key_arg<K>& key) { 1130 iterator it = find(key); 1131 GOOGLE_CHECK(it != end()) << "key not found: " << static_cast<Key>(key); 1132 return it->second; 1133 } 1134 1135 // Lookup 1136 template <typename K = key_type> 1137 size_type count(const key_arg<K>& key) const { 1138 return find(key) == end() ? 0 : 1; 1139 } 1140 1141 template <typename K = key_type> 1142 const_iterator find(const key_arg<K>& key) const { 1143 return const_iterator(iterator(elements_->find(key))); 1144 } 1145 template <typename K = key_type> 1146 iterator find(const key_arg<K>& key) { 1147 return iterator(elements_->find(key)); 1148 } 1149 1150 template <typename K = key_type> 1151 bool contains(const key_arg<K>& key) const { 1152 return find(key) != end(); 1153 } 1154 1155 template <typename K = key_type> 1156 std::pair<const_iterator, const_iterator> equal_range( 1157 const key_arg<K>& key) const { 1158 const_iterator it = find(key); 1159 if (it == end()) { 1160 return std::pair<const_iterator, const_iterator>(it, it); 1161 } else { 1162 const_iterator begin = it++; 1163 return std::pair<const_iterator, const_iterator>(begin, it); 1164 } 1165 } 1166 1167 template <typename K = key_type> 1168 std::pair<iterator, iterator> equal_range(const key_arg<K>& key) { 1169 iterator it = find(key); 1170 if (it == end()) { 1171 return std::pair<iterator, iterator>(it, it); 1172 } else { 1173 iterator begin = it++; 1174 return std::pair<iterator, iterator>(begin, it); 1175 } 1176 } 1177 1178 // insert 1179 std::pair<iterator, bool> insert(const value_type& value) { 1180 std::pair<typename InnerMap::iterator, bool> p = 1181 elements_->insert(value.first); 1182 if (p.second) { 1183 p.first->second = value.second; 1184 } 1185 return std::pair<iterator, bool>(iterator(p.first), p.second); 1186 } 1187 template <class InputIt> 1188 void insert(InputIt first, InputIt last) { 1189 for (InputIt it = first; it != last; ++it) { 1190 iterator exist_it = find(it->first); 1191 if (exist_it == end()) { 1192 operator[](it->first) = it->second; 1193 } 1194 } 1195 } 1196 void insert(std::initializer_list<value_type> values) { 1197 insert(values.begin(), values.end()); 1198 } 1199 1200 // Erase and clear 1201 template <typename K = key_type> 1202 size_type erase(const key_arg<K>& key) { 1203 iterator it = find(key); 1204 if (it == end()) { 1205 return 0; 1206 } else { 1207 erase(it); 1208 return 1; 1209 } 1210 } 1211 iterator erase(iterator pos) { 1212 iterator i = pos++; 1213 elements_->erase(i.it_); 1214 return pos; 1215 } 1216 void erase(iterator first, iterator last) { 1217 while (first != last) { 1218 first = erase(first); 1219 } 1220 } 1221 void clear() { elements_->clear(); } 1222 1223 // Assign 1224 Map& operator=(const Map& other) { 1225 if (this != &other) { 1226 clear(); 1227 insert(other.begin(), other.end()); 1228 } 1229 return *this; 1230 } 1231 1232 void swap(Map& other) { 1233 if (arena_ == other.arena_) { 1234 std::swap(default_enum_value_, other.default_enum_value_); 1235 std::swap(elements_, other.elements_); 1236 } else { 1237 // TODO(zuguang): optimize this. The temporary copy can be allocated 1238 // in the same arena as the other message, and the "other = copy" can 1239 // be replaced with the fast-path swap above. 1240 Map copy = *this; 1241 *this = other; 1242 other = copy; 1243 } 1244 } 1245 1246 // Access to hasher. Currently this returns a copy, but it may 1247 // be modified to return a const reference in the future. 1248 hasher hash_function() const { return elements_->hash_function(); } 1249 1250 private: 1251 // Set default enum value only for proto2 map field whose value is enum type. 1252 void SetDefaultEnumValue(int default_enum_value) { 1253 default_enum_value_ = default_enum_value; 1254 } 1255 1256 Arena* arena_; 1257 int default_enum_value_; 1258 InnerMap* elements_; 1259 1260 friend class Arena; 1261 using InternalArenaConstructable_ = void; 1262 using DestructorSkippable_ = void; 1263 template <typename Derived, typename K, typename V, 1264 internal::WireFormatLite::FieldType key_wire_type, 1265 internal::WireFormatLite::FieldType value_wire_type, 1266 int default_enum_value> 1267 friend class internal::MapFieldLite; 1268 }; 1269 1270 } // namespace protobuf 1271 } // namespace google 1272 1273 #include <google/protobuf/port_undef.inc> 1274 1275 #endif // GOOGLE_PROTOBUF_MAP_H__ 1276