| 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 | |
| 41 | #include <functional> |
| 42 | #include <initializer_list> |
| 43 | #include <iterator> |
| 44 | #include <limits> // To support Visual Studio 2008 |
| 45 | #include <map> |
| 46 | #include <string> |
| 47 | #include <type_traits> |
| 48 | #include <utility> |
| 49 | |
| 50 | #if defined(__cpp_lib_string_view) |
| 51 | #include <string_view> |
| 52 | #endif // defined(__cpp_lib_string_view) |
| 53 | |
| 54 | #if !defined(GOOGLE_PROTOBUF_NO_RDTSC) && defined(__APPLE__) |
| 55 | #include <mach/mach_time.h> |
| 56 | #endif |
| 57 | |
| 58 | #include <google/protobuf/stubs/common.h> |
| 59 | #include <google/protobuf/arena.h> |
| 60 | #include <google/protobuf/generated_enum_util.h> |
| 61 | #include <google/protobuf/map_type_handler.h> |
| 62 | #include <google/protobuf/port.h> |
| 63 | #include <google/protobuf/stubs/hash.h> |
| 64 | |
| 65 | #ifdef SWIG |
| 66 | #error "You cannot SWIG proto headers" |
| 67 | #endif |
| 68 | |
| 69 | // Must be included last. |
| 70 | #include <google/protobuf/port_def.inc> |
| 71 | |
| 72 | namespace google { |
| 73 | namespace protobuf { |
| 74 | |
| 75 | template <typename Key, typename T> |
| 76 | class Map; |
| 77 | |
| 78 | class MapIterator; |
| 79 | |
| 80 | template <typename Enum> |
| 81 | struct is_proto_enum; |
| 82 | |
| 83 | namespace internal { |
| 84 | template <typename Derived, typename Key, typename T, |
| 85 | WireFormatLite::FieldType key_wire_type, |
| 86 | WireFormatLite::FieldType value_wire_type> |
| 87 | class MapFieldLite; |
| 88 | |
| 89 | template <typename Derived, typename Key, typename T, |
| 90 | WireFormatLite::FieldType key_wire_type, |
| 91 | WireFormatLite::FieldType value_wire_type> |
| 92 | class MapField; |
| 93 | |
| 94 | template <typename Key, typename T> |
| 95 | class TypeDefinedMapFieldBase; |
| 96 | |
| 97 | class DynamicMapField; |
| 98 | |
| 99 | class GeneratedMessageReflection; |
| 100 | |
| 101 | // re-implement std::allocator to use arena allocator for memory allocation. |
| 102 | // Used for Map implementation. Users should not use this class |
| 103 | // directly. |
| 104 | template <typename U> |
| 105 | class MapAllocator { |
| 106 | public: |
| 107 | using value_type = U; |
| 108 | using pointer = value_type*; |
| 109 | using const_pointer = const value_type*; |
| 110 | using reference = value_type&; |
| 111 | using const_reference = const value_type&; |
| 112 | using size_type = size_t; |
| 113 | using difference_type = ptrdiff_t; |
| 114 | |
| 115 | constexpr MapAllocator() : arena_(nullptr) {} |
| 116 | explicit constexpr MapAllocator(Arena* arena) : arena_(arena) {} |
| 117 | template <typename X> |
| 118 | MapAllocator(const MapAllocator<X>& allocator) // NOLINT(runtime/explicit) |
| 119 | : arena_(allocator.arena()) {} |
| 120 | |
| 121 | // MapAllocator does not support alignments beyond 8. Technically we should |
| 122 | // support up to std::max_align_t, but this fails with ubsan and tcmalloc |
| 123 | // debug allocation logic which assume 8 as default alignment. |
| 124 | static_assert(alignof(value_type) <= 8, "" ); |
| 125 | |
| 126 | pointer allocate(size_type n, const void* /* hint */ = nullptr) { |
| 127 | // If arena is not given, malloc needs to be called which doesn't |
| 128 | // construct element object. |
| 129 | if (arena_ == nullptr) { |
| 130 | return static_cast<pointer>(::operator new(n * sizeof(value_type))); |
| 131 | } else { |
| 132 | return reinterpret_cast<pointer>( |
| 133 | Arena::CreateArray<uint8_t>(arena: arena_, num_elements: n * sizeof(value_type))); |
| 134 | } |
| 135 | } |
| 136 | |
| 137 | void deallocate(pointer p, size_type n) { |
| 138 | if (arena_ == nullptr) { |
| 139 | internal::SizedDelete(p, size: n * sizeof(value_type)); |
| 140 | } |
| 141 | } |
| 142 | |
| 143 | #if !defined(GOOGLE_PROTOBUF_OS_APPLE) && !defined(GOOGLE_PROTOBUF_OS_NACL) && \ |
| 144 | !defined(GOOGLE_PROTOBUF_OS_EMSCRIPTEN) |
| 145 | template <class NodeType, class... Args> |
| 146 | void construct(NodeType* p, Args&&... args) { |
| 147 | // Clang 3.6 doesn't compile static casting to void* directly. (Issue |
| 148 | // #1266) According C++ standard 5.2.9/1: "The static_cast operator shall |
| 149 | // not cast away constness". So first the maybe const pointer is casted to |
| 150 | // const void* and after the const void* is const casted. |
| 151 | new (const_cast<void*>(static_cast<const void*>(p))) |
| 152 | NodeType(std::forward<Args>(args)...); |
| 153 | } |
| 154 | |
| 155 | template <class NodeType> |
| 156 | void destroy(NodeType* p) { |
| 157 | p->~NodeType(); |
| 158 | } |
| 159 | #else |
| 160 | void construct(pointer p, const_reference t) { new (p) value_type(t); } |
| 161 | |
| 162 | void destroy(pointer p) { p->~value_type(); } |
| 163 | #endif |
| 164 | |
| 165 | template <typename X> |
| 166 | struct rebind { |
| 167 | using other = MapAllocator<X>; |
| 168 | }; |
| 169 | |
| 170 | template <typename X> |
| 171 | bool operator==(const MapAllocator<X>& other) const { |
| 172 | return arena_ == other.arena_; |
| 173 | } |
| 174 | |
| 175 | template <typename X> |
| 176 | bool operator!=(const MapAllocator<X>& other) const { |
| 177 | return arena_ != other.arena_; |
| 178 | } |
| 179 | |
| 180 | // To support Visual Studio 2008 |
| 181 | size_type max_size() const { |
| 182 | // parentheses around (std::...:max) prevents macro warning of max() |
| 183 | return (std::numeric_limits<size_type>::max)(); |
| 184 | } |
| 185 | |
| 186 | // To support gcc-4.4, which does not properly |
| 187 | // support templated friend classes |
| 188 | Arena* arena() const { return arena_; } |
| 189 | |
| 190 | private: |
| 191 | using DestructorSkippable_ = void; |
| 192 | Arena* arena_; |
| 193 | }; |
| 194 | |
| 195 | template <typename T> |
| 196 | using KeyForTree = |
| 197 | typename std::conditional<std::is_scalar<T>::value, T, |
| 198 | std::reference_wrapper<const T>>::type; |
| 199 | |
| 200 | // Default case: Not transparent. |
| 201 | // We use std::hash<key_type>/std::less<key_type> and all the lookup functions |
| 202 | // only accept `key_type`. |
| 203 | template <typename key_type> |
| 204 | struct TransparentSupport { |
| 205 | using hash = std::hash<key_type>; |
| 206 | using less = std::less<key_type>; |
| 207 | |
| 208 | static bool Equals(const key_type& a, const key_type& b) { return a == b; } |
| 209 | |
| 210 | template <typename K> |
| 211 | using key_arg = key_type; |
| 212 | }; |
| 213 | |
| 214 | #if defined(__cpp_lib_string_view) |
| 215 | // If std::string_view is available, we add transparent support for std::string |
| 216 | // keys. We use std::hash<std::string_view> as it supports the input types we |
| 217 | // care about. The lookup functions accept arbitrary `K`. This will include any |
| 218 | // key type that is convertible to std::string_view. |
| 219 | template <> |
| 220 | struct TransparentSupport<std::string> { |
| 221 | static std::string_view ImplicitConvert(std::string_view str) { return str; } |
| 222 | // If the element is not convertible to std::string_view, try to convert to |
| 223 | // std::string first. |
| 224 | // The template makes this overload lose resolution when both have the same |
| 225 | // rank otherwise. |
| 226 | template <typename = void> |
| 227 | static std::string_view ImplicitConvert(const std::string& str) { |
| 228 | return str; |
| 229 | } |
| 230 | |
| 231 | struct hash : private std::hash<std::string_view> { |
| 232 | using is_transparent = void; |
| 233 | |
| 234 | template <typename T> |
| 235 | size_t operator()(const T& str) const { |
| 236 | return base()(ImplicitConvert(str)); |
| 237 | } |
| 238 | |
| 239 | private: |
| 240 | const std::hash<std::string_view>& base() const { return *this; } |
| 241 | }; |
| 242 | struct less { |
| 243 | using is_transparent = void; |
| 244 | |
| 245 | template <typename T, typename U> |
| 246 | bool operator()(const T& t, const U& u) const { |
| 247 | return ImplicitConvert(t) < ImplicitConvert(u); |
| 248 | } |
| 249 | }; |
| 250 | |
| 251 | template <typename T, typename U> |
| 252 | static bool Equals(const T& t, const U& u) { |
| 253 | return ImplicitConvert(t) == ImplicitConvert(u); |
| 254 | } |
| 255 | |
| 256 | template <typename K> |
| 257 | using key_arg = K; |
| 258 | }; |
| 259 | #endif // defined(__cpp_lib_string_view) |
| 260 | |
| 261 | template <typename Key> |
| 262 | using TreeForMap = |
| 263 | std::map<KeyForTree<Key>, void*, typename TransparentSupport<Key>::less, |
| 264 | MapAllocator<std::pair<const KeyForTree<Key>, void*>>>; |
| 265 | |
| 266 | inline bool TableEntryIsEmpty(void* const* table, size_t b) { |
| 267 | return table[b] == nullptr; |
| 268 | } |
| 269 | inline bool TableEntryIsNonEmptyList(void* const* table, size_t b) { |
| 270 | return table[b] != nullptr && table[b] != table[b ^ 1]; |
| 271 | } |
| 272 | inline bool TableEntryIsTree(void* const* table, size_t b) { |
| 273 | return !TableEntryIsEmpty(table, b) && !TableEntryIsNonEmptyList(table, b); |
| 274 | } |
| 275 | inline bool TableEntryIsList(void* const* table, size_t b) { |
| 276 | return !TableEntryIsTree(table, b); |
| 277 | } |
| 278 | |
| 279 | // This captures all numeric types. |
| 280 | inline size_t MapValueSpaceUsedExcludingSelfLong(bool) { return 0; } |
| 281 | inline size_t MapValueSpaceUsedExcludingSelfLong(const std::string& str) { |
| 282 | return StringSpaceUsedExcludingSelfLong(str); |
| 283 | } |
| 284 | template <typename T, |
| 285 | typename = decltype(std::declval<const T&>().SpaceUsedLong())> |
| 286 | size_t MapValueSpaceUsedExcludingSelfLong(const T& message) { |
| 287 | return message.SpaceUsedLong() - sizeof(T); |
| 288 | } |
| 289 | |
| 290 | constexpr size_t kGlobalEmptyTableSize = 1; |
| 291 | PROTOBUF_EXPORT extern void* const kGlobalEmptyTable[kGlobalEmptyTableSize]; |
| 292 | |
| 293 | // Space used for the table, trees, and nodes. |
| 294 | // Does not include the indirect space used. Eg the data of a std::string. |
| 295 | template <typename Key> |
| 296 | PROTOBUF_NOINLINE size_t SpaceUsedInTable(void** table, size_t num_buckets, |
| 297 | size_t num_elements, |
| 298 | size_t sizeof_node) { |
| 299 | size_t size = 0; |
| 300 | // The size of the table. |
| 301 | size += sizeof(void*) * num_buckets; |
| 302 | // All the nodes. |
| 303 | size += sizeof_node * num_elements; |
| 304 | // For each tree, count the overhead of the those nodes. |
| 305 | // Two buckets at a time because we only care about trees. |
| 306 | for (size_t b = 0; b < num_buckets; b += 2) { |
| 307 | if (internal::TableEntryIsTree(table, b)) { |
| 308 | using Tree = TreeForMap<Key>; |
| 309 | Tree* tree = static_cast<Tree*>(table[b]); |
| 310 | // Estimated cost of the red-black tree nodes, 3 pointers plus a |
| 311 | // bool (plus alignment, so 4 pointers). |
| 312 | size += tree->size() * |
| 313 | (sizeof(typename Tree::value_type) + sizeof(void*) * 4); |
| 314 | } |
| 315 | } |
| 316 | return size; |
| 317 | } |
| 318 | |
| 319 | template <typename Map, |
| 320 | typename = typename std::enable_if< |
| 321 | !std::is_scalar<typename Map::key_type>::value || |
| 322 | !std::is_scalar<typename Map::mapped_type>::value>::type> |
| 323 | size_t SpaceUsedInValues(const Map* map) { |
| 324 | size_t size = 0; |
| 325 | for (const auto& v : *map) { |
| 326 | size += internal::MapValueSpaceUsedExcludingSelfLong(v.first) + |
| 327 | internal::MapValueSpaceUsedExcludingSelfLong(v.second); |
| 328 | } |
| 329 | return size; |
| 330 | } |
| 331 | |
| 332 | inline size_t SpaceUsedInValues(const void*) { return 0; } |
| 333 | |
| 334 | } // namespace internal |
| 335 | |
| 336 | // This is the class for Map's internal value_type. Instead of using |
| 337 | // std::pair as value_type, we use this class which provides us more control of |
| 338 | // its process of construction and destruction. |
| 339 | template <typename Key, typename T> |
| 340 | struct PROTOBUF_ATTRIBUTE_STANDALONE_DEBUG MapPair { |
| 341 | using first_type = const Key; |
| 342 | using second_type = T; |
| 343 | |
| 344 | MapPair(const Key& other_first, const T& other_second) |
| 345 | : first(other_first), second(other_second) {} |
| 346 | explicit MapPair(const Key& other_first) : first(other_first), second() {} |
| 347 | explicit MapPair(Key&& other_first) |
| 348 | : first(std::move(other_first)), second() {} |
| 349 | MapPair(const MapPair& other) : first(other.first), second(other.second) {} |
| 350 | |
| 351 | ~MapPair() {} |
| 352 | |
| 353 | // Implicitly convertible to std::pair of compatible types. |
| 354 | template <typename T1, typename T2> |
| 355 | operator std::pair<T1, T2>() const { // NOLINT(runtime/explicit) |
| 356 | return std::pair<T1, T2>(first, second); |
| 357 | } |
| 358 | |
| 359 | const Key first; |
| 360 | T second; |
| 361 | |
| 362 | private: |
| 363 | friend class Arena; |
| 364 | friend class Map<Key, T>; |
| 365 | }; |
| 366 | |
| 367 | // Map is an associative container type used to store protobuf map |
| 368 | // fields. Each Map instance may or may not use a different hash function, a |
| 369 | // different iteration order, and so on. E.g., please don't examine |
| 370 | // implementation details to decide if the following would work: |
| 371 | // Map<int, int> m0, m1; |
| 372 | // m0[0] = m1[0] = m0[1] = m1[1] = 0; |
| 373 | // assert(m0.begin()->first == m1.begin()->first); // Bug! |
| 374 | // |
| 375 | // Map's interface is similar to std::unordered_map, except that Map is not |
| 376 | // designed to play well with exceptions. |
| 377 | template <typename Key, typename T> |
| 378 | class Map { |
| 379 | public: |
| 380 | using key_type = Key; |
| 381 | using mapped_type = T; |
| 382 | using value_type = MapPair<Key, T>; |
| 383 | |
| 384 | using pointer = value_type*; |
| 385 | using const_pointer = const value_type*; |
| 386 | using reference = value_type&; |
| 387 | using const_reference = const value_type&; |
| 388 | |
| 389 | using size_type = size_t; |
| 390 | using hasher = typename internal::TransparentSupport<Key>::hash; |
| 391 | |
| 392 | constexpr Map() : elements_(nullptr) {} |
| 393 | explicit Map(Arena* arena) : elements_(arena) {} |
| 394 | |
| 395 | Map(const Map& other) : Map() { insert(other.begin(), other.end()); } |
| 396 | |
| 397 | Map(Map&& other) noexcept : Map() { |
| 398 | if (other.arena() != nullptr) { |
| 399 | *this = other; |
| 400 | } else { |
| 401 | swap(other); |
| 402 | } |
| 403 | } |
| 404 | |
| 405 | Map& operator=(Map&& other) noexcept { |
| 406 | if (this != &other) { |
| 407 | if (arena() != other.arena()) { |
| 408 | *this = other; |
| 409 | } else { |
| 410 | swap(other); |
| 411 | } |
| 412 | } |
| 413 | return *this; |
| 414 | } |
| 415 | |
| 416 | template <class InputIt> |
| 417 | Map(const InputIt& first, const InputIt& last) : Map() { |
| 418 | insert(first, last); |
| 419 | } |
| 420 | |
| 421 | ~Map() {} |
| 422 | |
| 423 | private: |
| 424 | using Allocator = internal::MapAllocator<void*>; |
| 425 | |
| 426 | // InnerMap is a generic hash-based map. It doesn't contain any |
| 427 | // protocol-buffer-specific logic. It is a chaining hash map with the |
| 428 | // additional feature that some buckets can be converted to use an ordered |
| 429 | // container. This ensures O(lg n) bounds on find, insert, and erase, while |
| 430 | // avoiding the overheads of ordered containers most of the time. |
| 431 | // |
| 432 | // The implementation doesn't need the full generality of unordered_map, |
| 433 | // and it doesn't have it. More bells and whistles can be added as needed. |
| 434 | // Some implementation details: |
| 435 | // 1. The hash function has type hasher and the equality function |
| 436 | // equal_to<Key>. We inherit from hasher to save space |
| 437 | // (empty-base-class optimization). |
| 438 | // 2. The number of buckets is a power of two. |
| 439 | // 3. Buckets are converted to trees in pairs: if we convert bucket b then |
| 440 | // buckets b and b^1 will share a tree. Invariant: buckets b and b^1 have |
| 441 | // the same non-null value iff they are sharing a tree. (An alternative |
| 442 | // implementation strategy would be to have a tag bit per bucket.) |
| 443 | // 4. As is typical for hash_map and such, the Keys and Values are always |
| 444 | // stored in linked list nodes. Pointers to elements are never invalidated |
| 445 | // until the element is deleted. |
| 446 | // 5. The trees' payload type is pointer to linked-list node. Tree-converting |
| 447 | // a bucket doesn't copy Key-Value pairs. |
| 448 | // 6. Once we've tree-converted a bucket, it is never converted back. However, |
| 449 | // the items a tree contains may wind up assigned to trees or lists upon a |
| 450 | // rehash. |
| 451 | // 7. The code requires no C++ features from C++14 or later. |
| 452 | // 8. Mutations to a map do not invalidate the map's iterators, pointers to |
| 453 | // elements, or references to elements. |
| 454 | // 9. Except for erase(iterator), any non-const method can reorder iterators. |
| 455 | // 10. InnerMap uses KeyForTree<Key> when using the Tree representation, which |
| 456 | // is either `Key`, if Key is a scalar, or `reference_wrapper<const Key>` |
| 457 | // otherwise. This avoids unnecessary copies of string keys, for example. |
| 458 | class InnerMap : private hasher { |
| 459 | public: |
| 460 | explicit constexpr InnerMap(Arena* arena) |
| 461 | : hasher(), |
| 462 | num_elements_(0), |
| 463 | num_buckets_(internal::kGlobalEmptyTableSize), |
| 464 | seed_(0), |
| 465 | index_of_first_non_null_(internal::kGlobalEmptyTableSize), |
| 466 | table_(const_cast<void**>(internal::kGlobalEmptyTable)), |
| 467 | alloc_(arena) {} |
| 468 | |
| 469 | ~InnerMap() { |
| 470 | if (alloc_.arena() == nullptr && |
| 471 | num_buckets_ != internal::kGlobalEmptyTableSize) { |
| 472 | clear(); |
| 473 | Dealloc<void*>(table_, num_buckets_); |
| 474 | } |
| 475 | } |
| 476 | |
| 477 | private: |
| 478 | enum { kMinTableSize = 8 }; |
| 479 | |
| 480 | // Linked-list nodes, as one would expect for a chaining hash table. |
| 481 | struct Node { |
| 482 | value_type kv; |
| 483 | Node* next; |
| 484 | }; |
| 485 | |
| 486 | // Trees. The payload type is a copy of Key, so that we can query the tree |
| 487 | // with Keys that are not in any particular data structure. |
| 488 | // The value is a void* pointing to Node. We use void* instead of Node* to |
| 489 | // avoid code bloat. That way there is only one instantiation of the tree |
| 490 | // class per key type. |
| 491 | using Tree = internal::TreeForMap<Key>; |
| 492 | using TreeIterator = typename Tree::iterator; |
| 493 | |
| 494 | static Node* NodeFromTreeIterator(TreeIterator it) { |
| 495 | return static_cast<Node*>(it->second); |
| 496 | } |
| 497 | |
| 498 | // iterator and const_iterator are instantiations of iterator_base. |
| 499 | template <typename KeyValueType> |
| 500 | class iterator_base { |
| 501 | public: |
| 502 | using reference = KeyValueType&; |
| 503 | using pointer = KeyValueType*; |
| 504 | |
| 505 | // Invariants: |
| 506 | // node_ is always correct. This is handy because the most common |
| 507 | // operations are operator* and operator-> and they only use node_. |
| 508 | // When node_ is set to a non-null value, all the other non-const fields |
| 509 | // are updated to be correct also, but those fields can become stale |
| 510 | // if the underlying map is modified. When those fields are needed they |
| 511 | // are rechecked, and updated if necessary. |
| 512 | iterator_base() : node_(nullptr), m_(nullptr), bucket_index_(0) {} |
| 513 | |
| 514 | explicit iterator_base(const InnerMap* m) : m_(m) { |
| 515 | SearchFrom(start_bucket: m->index_of_first_non_null_); |
| 516 | } |
| 517 | |
| 518 | // Any iterator_base can convert to any other. This is overkill, and we |
| 519 | // rely on the enclosing class to use it wisely. The standard "iterator |
| 520 | // can convert to const_iterator" is OK but the reverse direction is not. |
| 521 | template <typename U> |
| 522 | explicit iterator_base(const iterator_base<U>& it) |
| 523 | : node_(it.node_), m_(it.m_), bucket_index_(it.bucket_index_) {} |
| 524 | |
| 525 | iterator_base(Node* n, const InnerMap* m, size_type index) |
| 526 | : node_(n), m_(m), bucket_index_(index) {} |
| 527 | |
| 528 | iterator_base(TreeIterator tree_it, const InnerMap* m, size_type index) |
| 529 | : node_(NodeFromTreeIterator(it: tree_it)), m_(m), bucket_index_(index) { |
| 530 | // Invariant: iterators that use buckets with trees have an even |
| 531 | // bucket_index_. |
| 532 | GOOGLE_DCHECK_EQ(bucket_index_ % 2, 0u); |
| 533 | } |
| 534 | |
| 535 | // Advance through buckets, looking for the first that isn't empty. |
| 536 | // If nothing non-empty is found then leave node_ == nullptr. |
| 537 | void SearchFrom(size_type start_bucket) { |
| 538 | GOOGLE_DCHECK(m_->index_of_first_non_null_ == m_->num_buckets_ || |
| 539 | m_->table_[m_->index_of_first_non_null_] != nullptr); |
| 540 | node_ = nullptr; |
| 541 | for (bucket_index_ = start_bucket; bucket_index_ < m_->num_buckets_; |
| 542 | bucket_index_++) { |
| 543 | if (m_->TableEntryIsNonEmptyList(bucket_index_)) { |
| 544 | node_ = static_cast<Node*>(m_->table_[bucket_index_]); |
| 545 | break; |
| 546 | } else if (m_->TableEntryIsTree(bucket_index_)) { |
| 547 | Tree* tree = static_cast<Tree*>(m_->table_[bucket_index_]); |
| 548 | GOOGLE_DCHECK(!tree->empty()); |
| 549 | node_ = NodeFromTreeIterator(it: tree->begin()); |
| 550 | break; |
| 551 | } |
| 552 | } |
| 553 | } |
| 554 | |
| 555 | reference operator*() const { return node_->kv; } |
| 556 | pointer operator->() const { return &(operator*()); } |
| 557 | |
| 558 | friend bool operator==(const iterator_base& a, const iterator_base& b) { |
| 559 | return a.node_ == b.node_; |
| 560 | } |
| 561 | friend bool operator!=(const iterator_base& a, const iterator_base& b) { |
| 562 | return a.node_ != b.node_; |
| 563 | } |
| 564 | |
| 565 | iterator_base& operator++() { |
| 566 | if (node_->next == nullptr) { |
| 567 | TreeIterator tree_it; |
| 568 | const bool is_list = revalidate_if_necessary(it: &tree_it); |
| 569 | if (is_list) { |
| 570 | SearchFrom(start_bucket: bucket_index_ + 1); |
| 571 | } else { |
| 572 | GOOGLE_DCHECK_EQ(bucket_index_ & 1, 0u); |
| 573 | Tree* tree = static_cast<Tree*>(m_->table_[bucket_index_]); |
| 574 | if (++tree_it == tree->end()) { |
| 575 | SearchFrom(start_bucket: bucket_index_ + 2); |
| 576 | } else { |
| 577 | node_ = NodeFromTreeIterator(it: tree_it); |
| 578 | } |
| 579 | } |
| 580 | } else { |
| 581 | node_ = node_->next; |
| 582 | } |
| 583 | return *this; |
| 584 | } |
| 585 | |
| 586 | iterator_base operator++(int /* unused */) { |
| 587 | iterator_base tmp = *this; |
| 588 | ++*this; |
| 589 | return tmp; |
| 590 | } |
| 591 | |
| 592 | // Assumes node_ and m_ are correct and non-null, but other fields may be |
| 593 | // stale. Fix them as needed. Then return true iff node_ points to a |
| 594 | // Node in a list. If false is returned then *it is modified to be |
| 595 | // a valid iterator for node_. |
| 596 | bool revalidate_if_necessary(TreeIterator* it) { |
| 597 | GOOGLE_DCHECK(node_ != nullptr && m_ != nullptr); |
| 598 | // Force bucket_index_ to be in range. |
| 599 | bucket_index_ &= (m_->num_buckets_ - 1); |
| 600 | // Common case: the bucket we think is relevant points to node_. |
| 601 | if (m_->table_[bucket_index_] == static_cast<void*>(node_)) return true; |
| 602 | // Less common: the bucket is a linked list with node_ somewhere in it, |
| 603 | // but not at the head. |
| 604 | if (m_->TableEntryIsNonEmptyList(bucket_index_)) { |
| 605 | Node* l = static_cast<Node*>(m_->table_[bucket_index_]); |
| 606 | while ((l = l->next) != nullptr) { |
| 607 | if (l == node_) { |
| 608 | return true; |
| 609 | } |
| 610 | } |
| 611 | } |
| 612 | // Well, bucket_index_ still might be correct, but probably |
| 613 | // not. Revalidate just to be sure. This case is rare enough that we |
| 614 | // don't worry about potential optimizations, such as having a custom |
| 615 | // find-like method that compares Node* instead of the key. |
| 616 | iterator_base i(m_->find(node_->kv.first, it)); |
| 617 | bucket_index_ = i.bucket_index_; |
| 618 | return m_->TableEntryIsList(bucket_index_); |
| 619 | } |
| 620 | |
| 621 | Node* node_; |
| 622 | const InnerMap* m_; |
| 623 | size_type bucket_index_; |
| 624 | }; |
| 625 | |
| 626 | public: |
| 627 | using iterator = iterator_base<value_type>; |
| 628 | using const_iterator = iterator_base<const value_type>; |
| 629 | |
| 630 | Arena* arena() const { return alloc_.arena(); } |
| 631 | |
| 632 | void Swap(InnerMap* other) { |
| 633 | std::swap(num_elements_, other->num_elements_); |
| 634 | std::swap(num_buckets_, other->num_buckets_); |
| 635 | std::swap(seed_, other->seed_); |
| 636 | std::swap(index_of_first_non_null_, other->index_of_first_non_null_); |
| 637 | std::swap(table_, other->table_); |
| 638 | std::swap(alloc_, other->alloc_); |
| 639 | } |
| 640 | |
| 641 | iterator begin() { return iterator(this); } |
| 642 | iterator end() { return iterator(); } |
| 643 | const_iterator begin() const { return const_iterator(this); } |
| 644 | const_iterator end() const { return const_iterator(); } |
| 645 | |
| 646 | void clear() { |
| 647 | for (size_type b = 0; b < num_buckets_; b++) { |
| 648 | if (TableEntryIsNonEmptyList(b)) { |
| 649 | Node* node = static_cast<Node*>(table_[b]); |
| 650 | table_[b] = nullptr; |
| 651 | do { |
| 652 | Node* next = node->next; |
| 653 | DestroyNode(node); |
| 654 | node = next; |
| 655 | } while (node != nullptr); |
| 656 | } else if (TableEntryIsTree(b)) { |
| 657 | Tree* tree = static_cast<Tree*>(table_[b]); |
| 658 | GOOGLE_DCHECK(table_[b] == table_[b + 1] && (b & 1) == 0); |
| 659 | table_[b] = table_[b + 1] = nullptr; |
| 660 | typename Tree::iterator tree_it = tree->begin(); |
| 661 | do { |
| 662 | Node* node = NodeFromTreeIterator(it: tree_it); |
| 663 | typename Tree::iterator next = tree_it; |
| 664 | ++next; |
| 665 | tree->erase(tree_it); |
| 666 | DestroyNode(node); |
| 667 | tree_it = next; |
| 668 | } while (tree_it != tree->end()); |
| 669 | DestroyTree(tree); |
| 670 | b++; |
| 671 | } |
| 672 | } |
| 673 | num_elements_ = 0; |
| 674 | index_of_first_non_null_ = num_buckets_; |
| 675 | } |
| 676 | |
| 677 | const hasher& hash_function() const { return *this; } |
| 678 | |
| 679 | static size_type max_size() { |
| 680 | return static_cast<size_type>(1) << (sizeof(void**) >= 8 ? 60 : 28); |
| 681 | } |
| 682 | size_type size() const { return num_elements_; } |
| 683 | bool empty() const { return size() == 0; } |
| 684 | |
| 685 | template <typename K> |
| 686 | iterator find(const K& k) { |
| 687 | return iterator(FindHelper(k).first); |
| 688 | } |
| 689 | |
| 690 | template <typename K> |
| 691 | const_iterator find(const K& k) const { |
| 692 | return FindHelper(k).first; |
| 693 | } |
| 694 | |
| 695 | // Inserts a new element into the container if there is no element with the |
| 696 | // key in the container. |
| 697 | // The new element is: |
| 698 | // (1) Constructed in-place with the given args, if mapped_type is not |
| 699 | // arena constructible. |
| 700 | // (2) Constructed in-place with the arena and then assigned with a |
| 701 | // mapped_type temporary constructed with the given args, otherwise. |
| 702 | template <typename K, typename... Args> |
| 703 | std::pair<iterator, bool> try_emplace(K&& k, Args&&... args) { |
| 704 | return ArenaAwareTryEmplace(Arena::is_arena_constructable<mapped_type>(), |
| 705 | std::forward<K>(k), |
| 706 | std::forward<Args>(args)...); |
| 707 | } |
| 708 | |
| 709 | // Inserts the key into the map, if not present. In that case, the value |
| 710 | // will be value initialized. |
| 711 | template <typename K> |
| 712 | std::pair<iterator, bool> insert(K&& k) { |
| 713 | return try_emplace(std::forward<K>(k)); |
| 714 | } |
| 715 | |
| 716 | template <typename K> |
| 717 | value_type& operator[](K&& k) { |
| 718 | return *try_emplace(std::forward<K>(k)).first; |
| 719 | } |
| 720 | |
| 721 | void erase(iterator it) { |
| 722 | GOOGLE_DCHECK_EQ(it.m_, this); |
| 723 | typename Tree::iterator tree_it; |
| 724 | const bool is_list = it.revalidate_if_necessary(&tree_it); |
| 725 | size_type b = it.bucket_index_; |
| 726 | Node* const item = it.node_; |
| 727 | if (is_list) { |
| 728 | GOOGLE_DCHECK(TableEntryIsNonEmptyList(b)); |
| 729 | Node* head = static_cast<Node*>(table_[b]); |
| 730 | head = EraseFromLinkedList(item, head); |
| 731 | table_[b] = static_cast<void*>(head); |
| 732 | } else { |
| 733 | GOOGLE_DCHECK(TableEntryIsTree(b)); |
| 734 | Tree* tree = static_cast<Tree*>(table_[b]); |
| 735 | tree->erase(tree_it); |
| 736 | if (tree->empty()) { |
| 737 | // Force b to be the minimum of b and b ^ 1. This is important |
| 738 | // only because we want index_of_first_non_null_ to be correct. |
| 739 | b &= ~static_cast<size_type>(1); |
| 740 | DestroyTree(tree); |
| 741 | table_[b] = table_[b + 1] = nullptr; |
| 742 | } |
| 743 | } |
| 744 | DestroyNode(node: item); |
| 745 | --num_elements_; |
| 746 | if (PROTOBUF_PREDICT_FALSE(b == index_of_first_non_null_)) { |
| 747 | while (index_of_first_non_null_ < num_buckets_ && |
| 748 | table_[index_of_first_non_null_] == nullptr) { |
| 749 | ++index_of_first_non_null_; |
| 750 | } |
| 751 | } |
| 752 | } |
| 753 | |
| 754 | size_t SpaceUsedInternal() const { |
| 755 | return internal::SpaceUsedInTable<Key>(table_, num_buckets_, |
| 756 | num_elements_, sizeof(Node)); |
| 757 | } |
| 758 | |
| 759 | private: |
| 760 | template <typename K, typename... Args> |
| 761 | std::pair<iterator, bool> TryEmplaceInternal(K&& k, Args&&... args) { |
| 762 | std::pair<const_iterator, size_type> p = FindHelper(k); |
| 763 | // Case 1: key was already present. |
| 764 | if (p.first.node_ != nullptr) |
| 765 | return std::make_pair(iterator(p.first), false); |
| 766 | // Case 2: insert. |
| 767 | if (ResizeIfLoadIsOutOfRange(new_size: num_elements_ + 1)) { |
| 768 | p = FindHelper(k); |
| 769 | } |
| 770 | const size_type b = p.second; // bucket number |
| 771 | // If K is not key_type, make the conversion to key_type explicit. |
| 772 | using TypeToInit = typename std::conditional< |
| 773 | std::is_same<typename std::decay<K>::type, key_type>::value, K&&, |
| 774 | key_type>::type; |
| 775 | Node* node = Alloc<Node>(1); |
| 776 | // Even when arena is nullptr, CreateInArenaStorage is still used to |
| 777 | // ensure the arena of submessage will be consistent. Otherwise, |
| 778 | // submessage may have its own arena when message-owned arena is enabled. |
| 779 | // Note: This only works if `Key` is not arena constructible. |
| 780 | Arena::CreateInArenaStorage(const_cast<Key*>(&node->kv.first), |
| 781 | alloc_.arena(), |
| 782 | static_cast<TypeToInit>(std::forward<K>(k))); |
| 783 | // Note: if `T` is arena constructible, `Args` needs to be empty. |
| 784 | Arena::CreateInArenaStorage(&node->kv.second, alloc_.arena(), |
| 785 | std::forward<Args>(args)...); |
| 786 | |
| 787 | iterator result = InsertUnique(b, node); |
| 788 | ++num_elements_; |
| 789 | return std::make_pair(result, true); |
| 790 | } |
| 791 | |
| 792 | // A helper function to perform an assignment of `mapped_type`. |
| 793 | // If the first argument is true, then it is a regular assignment. |
| 794 | // Otherwise, we first create a temporary and then perform an assignment. |
| 795 | template <typename V> |
| 796 | static void AssignMapped(std::true_type, mapped_type& mapped, V&& v) { |
| 797 | mapped = std::forward<V>(v); |
| 798 | } |
| 799 | template <typename... Args> |
| 800 | static void AssignMapped(std::false_type, mapped_type& mapped, |
| 801 | Args&&... args) { |
| 802 | mapped = mapped_type(std::forward<Args>(args)...); |
| 803 | } |
| 804 | |
| 805 | // Case 1: `mapped_type` is arena constructible. A temporary object is |
| 806 | // created and then (if `Args` are not empty) assigned to a mapped value |
| 807 | // that was created with the arena. |
| 808 | template <typename K> |
| 809 | std::pair<iterator, bool> ArenaAwareTryEmplace(std::true_type, K&& k) { |
| 810 | // case 1.1: "default" constructed (e.g. from arena only). |
| 811 | return TryEmplaceInternal(std::forward<K>(k)); |
| 812 | } |
| 813 | template <typename K, typename... Args> |
| 814 | std::pair<iterator, bool> ArenaAwareTryEmplace(std::true_type, K&& k, |
| 815 | Args&&... args) { |
| 816 | // case 1.2: "default" constructed + copy/move assignment |
| 817 | auto p = TryEmplaceInternal(std::forward<K>(k)); |
| 818 | if (p.second) { |
| 819 | AssignMapped(std::is_same<void(typename std::decay<Args>::type...), |
| 820 | void(mapped_type)>(), |
| 821 | p.first->second, std::forward<Args>(args)...); |
| 822 | } |
| 823 | return p; |
| 824 | } |
| 825 | // Case 2: `mapped_type` is not arena constructible. Using in-place |
| 826 | // construction. |
| 827 | template <typename... Args> |
| 828 | std::pair<iterator, bool> ArenaAwareTryEmplace(std::false_type, |
| 829 | Args&&... args) { |
| 830 | return TryEmplaceInternal(std::forward<Args>(args)...); |
| 831 | } |
| 832 | |
| 833 | const_iterator find(const Key& k, TreeIterator* it) const { |
| 834 | return FindHelper(k, it).first; |
| 835 | } |
| 836 | template <typename K> |
| 837 | std::pair<const_iterator, size_type> FindHelper(const K& k) const { |
| 838 | return FindHelper(k, nullptr); |
| 839 | } |
| 840 | template <typename K> |
| 841 | std::pair<const_iterator, size_type> FindHelper(const K& k, |
| 842 | TreeIterator* it) const { |
| 843 | size_type b = BucketNumber(k); |
| 844 | if (TableEntryIsNonEmptyList(b)) { |
| 845 | Node* node = static_cast<Node*>(table_[b]); |
| 846 | do { |
| 847 | if (internal::TransparentSupport<Key>::Equals(node->kv.first, k)) { |
| 848 | return std::make_pair(const_iterator(node, this, b), b); |
| 849 | } else { |
| 850 | node = node->next; |
| 851 | } |
| 852 | } while (node != nullptr); |
| 853 | } else if (TableEntryIsTree(b)) { |
| 854 | GOOGLE_DCHECK_EQ(table_[b], table_[b ^ 1]); |
| 855 | b &= ~static_cast<size_t>(1); |
| 856 | Tree* tree = static_cast<Tree*>(table_[b]); |
| 857 | auto tree_it = tree->find(k); |
| 858 | if (tree_it != tree->end()) { |
| 859 | if (it != nullptr) *it = tree_it; |
| 860 | return std::make_pair(const_iterator(tree_it, this, b), b); |
| 861 | } |
| 862 | } |
| 863 | return std::make_pair(end(), b); |
| 864 | } |
| 865 | |
| 866 | // Insert the given Node in bucket b. If that would make bucket b too big, |
| 867 | // and bucket b is not a tree, create a tree for buckets b and b^1 to share. |
| 868 | // Requires count(*KeyPtrFromNodePtr(node)) == 0 and that b is the correct |
| 869 | // bucket. num_elements_ is not modified. |
| 870 | iterator InsertUnique(size_type b, Node* node) { |
| 871 | GOOGLE_DCHECK(index_of_first_non_null_ == num_buckets_ || |
| 872 | table_[index_of_first_non_null_] != nullptr); |
| 873 | // In practice, the code that led to this point may have already |
| 874 | // determined whether we are inserting into an empty list, a short list, |
| 875 | // or whatever. But it's probably cheap enough to recompute that here; |
| 876 | // it's likely that we're inserting into an empty or short list. |
| 877 | iterator result; |
| 878 | GOOGLE_DCHECK(find(node->kv.first) == end()); |
| 879 | if (TableEntryIsEmpty(b)) { |
| 880 | result = InsertUniqueInList(b, node); |
| 881 | } else if (TableEntryIsNonEmptyList(b)) { |
| 882 | if (PROTOBUF_PREDICT_FALSE(TableEntryIsTooLong(b))) { |
| 883 | TreeConvert(b); |
| 884 | result = InsertUniqueInTree(b, node); |
| 885 | GOOGLE_DCHECK_EQ(result.bucket_index_, b & ~static_cast<size_type>(1)); |
| 886 | } else { |
| 887 | // Insert into a pre-existing list. This case cannot modify |
| 888 | // index_of_first_non_null_, so we skip the code to update it. |
| 889 | return InsertUniqueInList(b, node); |
| 890 | } |
| 891 | } else { |
| 892 | // Insert into a pre-existing tree. This case cannot modify |
| 893 | // index_of_first_non_null_, so we skip the code to update it. |
| 894 | return InsertUniqueInTree(b, node); |
| 895 | } |
| 896 | // parentheses around (std::min) prevents macro expansion of min(...) |
| 897 | index_of_first_non_null_ = |
| 898 | (std::min)(index_of_first_non_null_, result.bucket_index_); |
| 899 | return result; |
| 900 | } |
| 901 | |
| 902 | // Returns whether we should insert after the head of the list. For |
| 903 | // non-optimized builds, we randomly decide whether to insert right at the |
| 904 | // head of the list or just after the head. This helps add a little bit of |
| 905 | // non-determinism to the map ordering. |
| 906 | bool ShouldInsertAfterHead(void* node) { |
| 907 | #ifdef NDEBUG |
| 908 | (void)node; |
| 909 | return false; |
| 910 | #else |
| 911 | // Doing modulo with a prime mixes the bits more. |
| 912 | return (reinterpret_cast<uintptr_t>(node) ^ seed_) % 13 > 6; |
| 913 | #endif |
| 914 | } |
| 915 | |
| 916 | // Helper for InsertUnique. Handles the case where bucket b is a |
| 917 | // not-too-long linked list. |
| 918 | iterator InsertUniqueInList(size_type b, Node* node) { |
| 919 | if (table_[b] != nullptr && ShouldInsertAfterHead(node)) { |
| 920 | Node* first = static_cast<Node*>(table_[b]); |
| 921 | node->next = first->next; |
| 922 | first->next = node; |
| 923 | return iterator(node, this, b); |
| 924 | } |
| 925 | |
| 926 | node->next = static_cast<Node*>(table_[b]); |
| 927 | table_[b] = static_cast<void*>(node); |
| 928 | return iterator(node, this, b); |
| 929 | } |
| 930 | |
| 931 | // Helper for InsertUnique. Handles the case where bucket b points to a |
| 932 | // Tree. |
| 933 | iterator InsertUniqueInTree(size_type b, Node* node) { |
| 934 | GOOGLE_DCHECK_EQ(table_[b], table_[b ^ 1]); |
| 935 | // Maintain the invariant that node->next is null for all Nodes in Trees. |
| 936 | node->next = nullptr; |
| 937 | return iterator( |
| 938 | static_cast<Tree*>(table_[b])->insert({node->kv.first, node}).first, |
| 939 | this, b & ~static_cast<size_t>(1)); |
| 940 | } |
| 941 | |
| 942 | // Returns whether it did resize. Currently this is only used when |
| 943 | // num_elements_ increases, though it could be used in other situations. |
| 944 | // It checks for load too low as well as load too high: because any number |
| 945 | // of erases can occur between inserts, the load could be as low as 0 here. |
| 946 | // Resizing to a lower size is not always helpful, but failing to do so can |
| 947 | // destroy the expected big-O bounds for some operations. By having the |
| 948 | // policy that sometimes we resize down as well as up, clients can easily |
| 949 | // keep O(size()) = O(number of buckets) if they want that. |
| 950 | bool ResizeIfLoadIsOutOfRange(size_type new_size) { |
| 951 | const size_type kMaxMapLoadTimes16 = 12; // controls RAM vs CPU tradeoff |
| 952 | const size_type hi_cutoff = num_buckets_ * kMaxMapLoadTimes16 / 16; |
| 953 | const size_type lo_cutoff = hi_cutoff / 4; |
| 954 | // We don't care how many elements are in trees. If a lot are, |
| 955 | // we may resize even though there are many empty buckets. In |
| 956 | // practice, this seems fine. |
| 957 | if (PROTOBUF_PREDICT_FALSE(new_size >= hi_cutoff)) { |
| 958 | if (num_buckets_ <= max_size() / 2) { |
| 959 | Resize(new_num_buckets: num_buckets_ * 2); |
| 960 | return true; |
| 961 | } |
| 962 | } else if (PROTOBUF_PREDICT_FALSE(new_size <= lo_cutoff && |
| 963 | num_buckets_ > kMinTableSize)) { |
| 964 | size_type lg2_of_size_reduction_factor = 1; |
| 965 | // It's possible we want to shrink a lot here... size() could even be 0. |
| 966 | // So, estimate how much to shrink by making sure we don't shrink so |
| 967 | // much that we would need to grow the table after a few inserts. |
| 968 | const size_type hypothetical_size = new_size * 5 / 4 + 1; |
| 969 | while ((hypothetical_size << lg2_of_size_reduction_factor) < |
| 970 | hi_cutoff) { |
| 971 | ++lg2_of_size_reduction_factor; |
| 972 | } |
| 973 | size_type new_num_buckets = std::max<size_type>( |
| 974 | kMinTableSize, num_buckets_ >> lg2_of_size_reduction_factor); |
| 975 | if (new_num_buckets != num_buckets_) { |
| 976 | Resize(new_num_buckets); |
| 977 | return true; |
| 978 | } |
| 979 | } |
| 980 | return false; |
| 981 | } |
| 982 | |
| 983 | // Resize to the given number of buckets. |
| 984 | void Resize(size_t new_num_buckets) { |
| 985 | if (num_buckets_ == internal::kGlobalEmptyTableSize) { |
| 986 | // This is the global empty array. |
| 987 | // Just overwrite with a new one. No need to transfer or free anything. |
| 988 | num_buckets_ = index_of_first_non_null_ = kMinTableSize; |
| 989 | table_ = CreateEmptyTable(n: num_buckets_); |
| 990 | seed_ = Seed(); |
| 991 | return; |
| 992 | } |
| 993 | |
| 994 | GOOGLE_DCHECK_GE(new_num_buckets, kMinTableSize); |
| 995 | void** const old_table = table_; |
| 996 | const size_type old_table_size = num_buckets_; |
| 997 | num_buckets_ = new_num_buckets; |
| 998 | table_ = CreateEmptyTable(n: num_buckets_); |
| 999 | const size_type start = index_of_first_non_null_; |
| 1000 | index_of_first_non_null_ = num_buckets_; |
| 1001 | for (size_type i = start; i < old_table_size; i++) { |
| 1002 | if (internal::TableEntryIsNonEmptyList(table: old_table, b: i)) { |
| 1003 | TransferList(table: old_table, index: i); |
| 1004 | } else if (internal::TableEntryIsTree(table: old_table, b: i)) { |
| 1005 | TransferTree(table: old_table, index: i++); |
| 1006 | } |
| 1007 | } |
| 1008 | Dealloc<void*>(old_table, old_table_size); |
| 1009 | } |
| 1010 | |
| 1011 | void TransferList(void* const* table, size_type index) { |
| 1012 | Node* node = static_cast<Node*>(table[index]); |
| 1013 | do { |
| 1014 | Node* next = node->next; |
| 1015 | InsertUnique(b: BucketNumber(node->kv.first), node); |
| 1016 | node = next; |
| 1017 | } while (node != nullptr); |
| 1018 | } |
| 1019 | |
| 1020 | void TransferTree(void* const* table, size_type index) { |
| 1021 | Tree* tree = static_cast<Tree*>(table[index]); |
| 1022 | typename Tree::iterator tree_it = tree->begin(); |
| 1023 | do { |
| 1024 | InsertUnique(b: BucketNumber(std::cref(tree_it->first).get()), |
| 1025 | node: NodeFromTreeIterator(it: tree_it)); |
| 1026 | } while (++tree_it != tree->end()); |
| 1027 | DestroyTree(tree); |
| 1028 | } |
| 1029 | |
| 1030 | Node* EraseFromLinkedList(Node* item, Node* head) { |
| 1031 | if (head == item) { |
| 1032 | return head->next; |
| 1033 | } else { |
| 1034 | head->next = EraseFromLinkedList(item, head: head->next); |
| 1035 | return head; |
| 1036 | } |
| 1037 | } |
| 1038 | |
| 1039 | bool TableEntryIsEmpty(size_type b) const { |
| 1040 | return internal::TableEntryIsEmpty(table: table_, b); |
| 1041 | } |
| 1042 | bool TableEntryIsNonEmptyList(size_type b) const { |
| 1043 | return internal::TableEntryIsNonEmptyList(table: table_, b); |
| 1044 | } |
| 1045 | bool TableEntryIsTree(size_type b) const { |
| 1046 | return internal::TableEntryIsTree(table: table_, b); |
| 1047 | } |
| 1048 | bool TableEntryIsList(size_type b) const { |
| 1049 | return internal::TableEntryIsList(table: table_, b); |
| 1050 | } |
| 1051 | |
| 1052 | void TreeConvert(size_type b) { |
| 1053 | GOOGLE_DCHECK(!TableEntryIsTree(b) && !TableEntryIsTree(b ^ 1)); |
| 1054 | Tree* tree = |
| 1055 | Arena::Create<Tree>(alloc_.arena(), typename Tree::key_compare(), |
| 1056 | typename Tree::allocator_type(alloc_)); |
| 1057 | size_type count = CopyListToTree(b, tree) + CopyListToTree(b: b ^ 1, tree); |
| 1058 | GOOGLE_DCHECK_EQ(count, tree->size()); |
| 1059 | table_[b] = table_[b ^ 1] = static_cast<void*>(tree); |
| 1060 | } |
| 1061 | |
| 1062 | // Copy a linked list in the given bucket to a tree. |
| 1063 | // Returns the number of things it copied. |
| 1064 | size_type CopyListToTree(size_type b, Tree* tree) { |
| 1065 | size_type count = 0; |
| 1066 | Node* node = static_cast<Node*>(table_[b]); |
| 1067 | while (node != nullptr) { |
| 1068 | tree->insert({node->kv.first, node}); |
| 1069 | ++count; |
| 1070 | Node* next = node->next; |
| 1071 | node->next = nullptr; |
| 1072 | node = next; |
| 1073 | } |
| 1074 | return count; |
| 1075 | } |
| 1076 | |
| 1077 | // Return whether table_[b] is a linked list that seems awfully long. |
| 1078 | // Requires table_[b] to point to a non-empty linked list. |
| 1079 | bool TableEntryIsTooLong(size_type b) { |
| 1080 | const size_type kMaxLength = 8; |
| 1081 | size_type count = 0; |
| 1082 | Node* node = static_cast<Node*>(table_[b]); |
| 1083 | do { |
| 1084 | ++count; |
| 1085 | node = node->next; |
| 1086 | } while (node != nullptr); |
| 1087 | // Invariant: no linked list ever is more than kMaxLength in length. |
| 1088 | GOOGLE_DCHECK_LE(count, kMaxLength); |
| 1089 | return count >= kMaxLength; |
| 1090 | } |
| 1091 | |
| 1092 | template <typename K> |
| 1093 | size_type BucketNumber(const K& k) const { |
| 1094 | // We xor the hash value against the random seed so that we effectively |
| 1095 | // have a random hash function. |
| 1096 | uint64_t h = hash_function()(k) ^ seed_; |
| 1097 | |
| 1098 | // We use the multiplication method to determine the bucket number from |
| 1099 | // the hash value. The constant kPhi (suggested by Knuth) is roughly |
| 1100 | // (sqrt(5) - 1) / 2 * 2^64. |
| 1101 | constexpr uint64_t kPhi = uint64_t{0x9e3779b97f4a7c15}; |
| 1102 | return ((kPhi * h) >> 32) & (num_buckets_ - 1); |
| 1103 | } |
| 1104 | |
| 1105 | // Return a power of two no less than max(kMinTableSize, n). |
| 1106 | // Assumes either n < kMinTableSize or n is a power of two. |
| 1107 | size_type TableSize(size_type n) { |
| 1108 | return n < static_cast<size_type>(kMinTableSize) |
| 1109 | ? static_cast<size_type>(kMinTableSize) |
| 1110 | : n; |
| 1111 | } |
| 1112 | |
| 1113 | // Use alloc_ to allocate an array of n objects of type U. |
| 1114 | template <typename U> |
| 1115 | U* Alloc(size_type n) { |
| 1116 | using alloc_type = typename Allocator::template rebind<U>::other; |
| 1117 | return alloc_type(alloc_).allocate(n); |
| 1118 | } |
| 1119 | |
| 1120 | // Use alloc_ to deallocate an array of n objects of type U. |
| 1121 | template <typename U> |
| 1122 | void Dealloc(U* t, size_type n) { |
| 1123 | using alloc_type = typename Allocator::template rebind<U>::other; |
| 1124 | alloc_type(alloc_).deallocate(t, n); |
| 1125 | } |
| 1126 | |
| 1127 | void DestroyNode(Node* node) { |
| 1128 | if (alloc_.arena() == nullptr) { |
| 1129 | delete node; |
| 1130 | } |
| 1131 | } |
| 1132 | |
| 1133 | void DestroyTree(Tree* tree) { |
| 1134 | if (alloc_.arena() == nullptr) { |
| 1135 | delete tree; |
| 1136 | } |
| 1137 | } |
| 1138 | |
| 1139 | void** CreateEmptyTable(size_type n) { |
| 1140 | GOOGLE_DCHECK(n >= kMinTableSize); |
| 1141 | GOOGLE_DCHECK_EQ(n & (n - 1), 0u); |
| 1142 | void** result = Alloc<void*>(n); |
| 1143 | memset(s: result, c: 0, n: n * sizeof(result[0])); |
| 1144 | return result; |
| 1145 | } |
| 1146 | |
| 1147 | // Return a randomish value. |
| 1148 | size_type Seed() const { |
| 1149 | // We get a little bit of randomness from the address of the map. The |
| 1150 | // lower bits are not very random, due to alignment, so we discard them |
| 1151 | // and shift the higher bits into their place. |
| 1152 | size_type s = reinterpret_cast<uintptr_t>(this) >> 4; |
| 1153 | #if !defined(GOOGLE_PROTOBUF_NO_RDTSC) |
| 1154 | #if defined(__APPLE__) |
| 1155 | // Use a commpage-based fast time function on Apple environments (MacOS, |
| 1156 | // iOS, tvOS, watchOS, etc). |
| 1157 | s += mach_absolute_time(); |
| 1158 | #elif defined(__x86_64__) && defined(__GNUC__) |
| 1159 | uint32_t hi, lo; |
| 1160 | asm volatile("rdtsc" : "=a" (lo), "=d" (hi)); |
| 1161 | s += ((static_cast<uint64_t>(hi) << 32) | lo); |
| 1162 | #elif defined(__aarch64__) && defined(__GNUC__) |
| 1163 | // There is no rdtsc on ARMv8. CNTVCT_EL0 is the virtual counter of the |
| 1164 | // system timer. It runs at a different frequency than the CPU's, but is |
| 1165 | // the best source of time-based entropy we get. |
| 1166 | uint64_t virtual_timer_value; |
| 1167 | asm volatile("mrs %0, cntvct_el0" : "=r" (virtual_timer_value)); |
| 1168 | s += virtual_timer_value; |
| 1169 | #endif |
| 1170 | #endif // !defined(GOOGLE_PROTOBUF_NO_RDTSC) |
| 1171 | return s; |
| 1172 | } |
| 1173 | |
| 1174 | friend class Arena; |
| 1175 | using InternalArenaConstructable_ = void; |
| 1176 | using DestructorSkippable_ = void; |
| 1177 | |
| 1178 | size_type num_elements_; |
| 1179 | size_type num_buckets_; |
| 1180 | size_type seed_; |
| 1181 | size_type index_of_first_non_null_; |
| 1182 | void** table_; // an array with num_buckets_ entries |
| 1183 | Allocator alloc_; |
| 1184 | GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(InnerMap); |
| 1185 | }; // end of class InnerMap |
| 1186 | |
| 1187 | template <typename LookupKey> |
| 1188 | using key_arg = typename internal::TransparentSupport< |
| 1189 | key_type>::template key_arg<LookupKey>; |
| 1190 | |
| 1191 | public: |
| 1192 | // Iterators |
| 1193 | class const_iterator { |
| 1194 | using InnerIt = typename InnerMap::const_iterator; |
| 1195 | |
| 1196 | public: |
| 1197 | using iterator_category = std::forward_iterator_tag; |
| 1198 | using value_type = typename Map::value_type; |
| 1199 | using difference_type = ptrdiff_t; |
| 1200 | using pointer = const value_type*; |
| 1201 | using reference = const value_type&; |
| 1202 | |
| 1203 | const_iterator() {} |
| 1204 | explicit const_iterator(const InnerIt& it) : it_(it) {} |
| 1205 | |
| 1206 | const_reference operator*() const { return *it_; } |
| 1207 | const_pointer operator->() const { return &(operator*()); } |
| 1208 | |
| 1209 | const_iterator& operator++() { |
| 1210 | ++it_; |
| 1211 | return *this; |
| 1212 | } |
| 1213 | const_iterator operator++(int) { return const_iterator(it_++); } |
| 1214 | |
| 1215 | friend bool operator==(const const_iterator& a, const const_iterator& b) { |
| 1216 | return a.it_ == b.it_; |
| 1217 | } |
| 1218 | friend bool operator!=(const const_iterator& a, const const_iterator& b) { |
| 1219 | return !(a == b); |
| 1220 | } |
| 1221 | |
| 1222 | private: |
| 1223 | InnerIt it_; |
| 1224 | }; |
| 1225 | |
| 1226 | class iterator { |
| 1227 | using InnerIt = typename InnerMap::iterator; |
| 1228 | |
| 1229 | public: |
| 1230 | using iterator_category = std::forward_iterator_tag; |
| 1231 | using value_type = typename Map::value_type; |
| 1232 | using difference_type = ptrdiff_t; |
| 1233 | using pointer = value_type*; |
| 1234 | using reference = value_type&; |
| 1235 | |
| 1236 | iterator() {} |
| 1237 | explicit iterator(const InnerIt& it) : it_(it) {} |
| 1238 | |
| 1239 | reference operator*() const { return *it_; } |
| 1240 | pointer operator->() const { return &(operator*()); } |
| 1241 | |
| 1242 | iterator& operator++() { |
| 1243 | ++it_; |
| 1244 | return *this; |
| 1245 | } |
| 1246 | iterator operator++(int) { return iterator(it_++); } |
| 1247 | |
| 1248 | // Allow implicit conversion to const_iterator. |
| 1249 | operator const_iterator() const { // NOLINT(runtime/explicit) |
| 1250 | return const_iterator(typename InnerMap::const_iterator(it_)); |
| 1251 | } |
| 1252 | |
| 1253 | friend bool operator==(const iterator& a, const iterator& b) { |
| 1254 | return a.it_ == b.it_; |
| 1255 | } |
| 1256 | friend bool operator!=(const iterator& a, const iterator& b) { |
| 1257 | return !(a == b); |
| 1258 | } |
| 1259 | |
| 1260 | private: |
| 1261 | friend class Map; |
| 1262 | |
| 1263 | InnerIt it_; |
| 1264 | }; |
| 1265 | |
| 1266 | iterator begin() { return iterator(elements_.begin()); } |
| 1267 | iterator end() { return iterator(elements_.end()); } |
| 1268 | const_iterator begin() const { return const_iterator(elements_.begin()); } |
| 1269 | const_iterator end() const { return const_iterator(elements_.end()); } |
| 1270 | const_iterator cbegin() const { return begin(); } |
| 1271 | const_iterator cend() const { return end(); } |
| 1272 | |
| 1273 | // Capacity |
| 1274 | size_type size() const { return elements_.size(); } |
| 1275 | bool empty() const { return size() == 0; } |
| 1276 | |
| 1277 | // Element access |
| 1278 | template <typename K = key_type> |
| 1279 | T& operator[](const key_arg<K>& key) { |
| 1280 | return elements_[key].second; |
| 1281 | } |
| 1282 | template < |
| 1283 | typename K = key_type, |
| 1284 | // Disable for integral types to reduce code bloat. |
| 1285 | typename = typename std::enable_if<!std::is_integral<K>::value>::type> |
| 1286 | T& operator[](key_arg<K>&& key) { |
| 1287 | return elements_[std::forward<K>(key)].second; |
| 1288 | } |
| 1289 | |
| 1290 | template <typename K = key_type> |
| 1291 | const T& at(const key_arg<K>& key) const { |
| 1292 | const_iterator it = find(key); |
| 1293 | GOOGLE_CHECK(it != end()) << "key not found: " << static_cast<Key>(key); |
| 1294 | return it->second; |
| 1295 | } |
| 1296 | |
| 1297 | template <typename K = key_type> |
| 1298 | T& at(const key_arg<K>& key) { |
| 1299 | iterator it = find(key); |
| 1300 | GOOGLE_CHECK(it != end()) << "key not found: " << static_cast<Key>(key); |
| 1301 | return it->second; |
| 1302 | } |
| 1303 | |
| 1304 | // Lookup |
| 1305 | template <typename K = key_type> |
| 1306 | size_type count(const key_arg<K>& key) const { |
| 1307 | return find(key) == end() ? 0 : 1; |
| 1308 | } |
| 1309 | |
| 1310 | template <typename K = key_type> |
| 1311 | const_iterator find(const key_arg<K>& key) const { |
| 1312 | return const_iterator(elements_.find(key)); |
| 1313 | } |
| 1314 | template <typename K = key_type> |
| 1315 | iterator find(const key_arg<K>& key) { |
| 1316 | return iterator(elements_.find(key)); |
| 1317 | } |
| 1318 | |
| 1319 | template <typename K = key_type> |
| 1320 | bool contains(const key_arg<K>& key) const { |
| 1321 | return find(key) != end(); |
| 1322 | } |
| 1323 | |
| 1324 | template <typename K = key_type> |
| 1325 | std::pair<const_iterator, const_iterator> equal_range( |
| 1326 | const key_arg<K>& key) const { |
| 1327 | const_iterator it = find(key); |
| 1328 | if (it == end()) { |
| 1329 | return std::pair<const_iterator, const_iterator>(it, it); |
| 1330 | } else { |
| 1331 | const_iterator begin = it++; |
| 1332 | return std::pair<const_iterator, const_iterator>(begin, it); |
| 1333 | } |
| 1334 | } |
| 1335 | |
| 1336 | template <typename K = key_type> |
| 1337 | std::pair<iterator, iterator> equal_range(const key_arg<K>& key) { |
| 1338 | iterator it = find(key); |
| 1339 | if (it == end()) { |
| 1340 | return std::pair<iterator, iterator>(it, it); |
| 1341 | } else { |
| 1342 | iterator begin = it++; |
| 1343 | return std::pair<iterator, iterator>(begin, it); |
| 1344 | } |
| 1345 | } |
| 1346 | |
| 1347 | // insert |
| 1348 | template <typename K, typename... Args> |
| 1349 | std::pair<iterator, bool> try_emplace(K&& k, Args&&... args) { |
| 1350 | auto p = |
| 1351 | elements_.try_emplace(std::forward<K>(k), std::forward<Args>(args)...); |
| 1352 | return std::pair<iterator, bool>(iterator(p.first), p.second); |
| 1353 | } |
| 1354 | std::pair<iterator, bool> insert(const value_type& value) { |
| 1355 | return try_emplace(value.first, value.second); |
| 1356 | } |
| 1357 | std::pair<iterator, bool> insert(value_type&& value) { |
| 1358 | return try_emplace(value.first, std::move(value.second)); |
| 1359 | } |
| 1360 | template <typename... Args> |
| 1361 | std::pair<iterator, bool> emplace(Args&&... args) { |
| 1362 | return insert(value_type(std::forward<Args>(args)...)); |
| 1363 | } |
| 1364 | template <class InputIt> |
| 1365 | void insert(InputIt first, InputIt last) { |
| 1366 | for (; first != last; ++first) { |
| 1367 | try_emplace(first->first, first->second); |
| 1368 | } |
| 1369 | } |
| 1370 | void insert(std::initializer_list<value_type> values) { |
| 1371 | insert(values.begin(), values.end()); |
| 1372 | } |
| 1373 | |
| 1374 | // Erase and clear |
| 1375 | template <typename K = key_type> |
| 1376 | size_type erase(const key_arg<K>& key) { |
| 1377 | iterator it = find(key); |
| 1378 | if (it == end()) { |
| 1379 | return 0; |
| 1380 | } else { |
| 1381 | erase(it); |
| 1382 | return 1; |
| 1383 | } |
| 1384 | } |
| 1385 | iterator erase(iterator pos) { |
| 1386 | iterator i = pos++; |
| 1387 | elements_.erase(i.it_); |
| 1388 | return pos; |
| 1389 | } |
| 1390 | void erase(iterator first, iterator last) { |
| 1391 | while (first != last) { |
| 1392 | first = erase(first); |
| 1393 | } |
| 1394 | } |
| 1395 | void clear() { elements_.clear(); } |
| 1396 | |
| 1397 | // Assign |
| 1398 | Map& operator=(const Map& other) { |
| 1399 | if (this != &other) { |
| 1400 | clear(); |
| 1401 | insert(other.begin(), other.end()); |
| 1402 | } |
| 1403 | return *this; |
| 1404 | } |
| 1405 | |
| 1406 | void swap(Map& other) { |
| 1407 | if (arena() == other.arena()) { |
| 1408 | InternalSwap(other); |
| 1409 | } else { |
| 1410 | // TODO(zuguang): optimize this. The temporary copy can be allocated |
| 1411 | // in the same arena as the other message, and the "other = copy" can |
| 1412 | // be replaced with the fast-path swap above. |
| 1413 | Map copy = *this; |
| 1414 | *this = other; |
| 1415 | other = copy; |
| 1416 | } |
| 1417 | } |
| 1418 | |
| 1419 | void InternalSwap(Map& other) { elements_.Swap(&other.elements_); } |
| 1420 | |
| 1421 | // Access to hasher. Currently this returns a copy, but it may |
| 1422 | // be modified to return a const reference in the future. |
| 1423 | hasher hash_function() const { return elements_.hash_function(); } |
| 1424 | |
| 1425 | size_t SpaceUsedExcludingSelfLong() const { |
| 1426 | if (empty()) return 0; |
| 1427 | return elements_.SpaceUsedInternal() + internal::SpaceUsedInValues(this); |
| 1428 | } |
| 1429 | |
| 1430 | private: |
| 1431 | Arena* arena() const { return elements_.arena(); } |
| 1432 | InnerMap elements_; |
| 1433 | |
| 1434 | friend class Arena; |
| 1435 | using InternalArenaConstructable_ = void; |
| 1436 | using DestructorSkippable_ = void; |
| 1437 | template <typename Derived, typename K, typename V, |
| 1438 | internal::WireFormatLite::FieldType key_wire_type, |
| 1439 | internal::WireFormatLite::FieldType value_wire_type> |
| 1440 | friend class internal::MapFieldLite; |
| 1441 | }; |
| 1442 | |
| 1443 | } // namespace protobuf |
| 1444 | } // namespace google |
| 1445 | |
| 1446 | #include <google/protobuf/port_undef.inc> |
| 1447 | |
| 1448 | #endif // GOOGLE_PROTOBUF_MAP_H__ |
| 1449 | |