| 1 | /* |
| 2 | * Copyright 2015 WebAssembly Community Group participants |
| 3 | * |
| 4 | * Licensed under the Apache License, Version 2.0 (the "License"); |
| 5 | * you may not use this file except in compliance with the License. |
| 6 | * You may obtain a copy of the License at |
| 7 | * |
| 8 | * http://www.apache.org/licenses/LICENSE-2.0 |
| 9 | * |
| 10 | * Unless required by applicable law or agreed to in writing, software |
| 11 | * distributed under the License is distributed on an "AS IS" BASIS, |
| 12 | * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 13 | * See the License for the specific language governing permissions and |
| 14 | * limitations under the License. |
| 15 | */ |
| 16 | |
| 17 | #ifndef wasm_mixed_arena_h |
| 18 | #define wasm_mixed_arena_h |
| 19 | |
| 20 | #include <atomic> |
| 21 | #include <cassert> |
| 22 | #include <memory> |
| 23 | #include <mutex> |
| 24 | #include <thread> |
| 25 | #include <type_traits> |
| 26 | #include <vector> |
| 27 | |
| 28 | #include <support/alloc.h> |
| 29 | |
| 30 | // |
| 31 | // Arena allocation for mixed-type data. |
| 32 | // |
| 33 | // Arena-style bump allocation is important for two reasons: First, so that |
| 34 | // allocation is quick, and second, so that allocated items are close together, |
| 35 | // which is cache-friendy. Arena allocation is also useful for a minor third |
| 36 | // reason which is to make freeing all the items in an arena very quick. |
| 37 | // |
| 38 | // Each WebAssembly Module has an arena allocator, which should be used |
| 39 | // for all of its AST nodes and so forth. When the Module is destroyed, the |
| 40 | // entire arena is cleaned up. |
| 41 | // |
| 42 | // When allocating an object in an arena, the object's proper constructor |
| 43 | // is called. Note that destructors are not called, because to make the |
| 44 | // arena simple and fast we do not track internal allocations inside it |
| 45 | // (and we can also avoid the need for virtual destructors). |
| 46 | // |
| 47 | // In general, optimization passes avoid allocation as much as possible. |
| 48 | // Many passes only remove or modify nodes anyhow, others can often |
| 49 | // reuse nodes that are being optimized out. This keeps things |
| 50 | // cache-friendly, and also makes the operations trivially thread-safe. |
| 51 | // In the rare case that a pass does need to allocate, and it is a |
| 52 | // parallel pass (so multiple threads might access the allocator), |
| 53 | // the MixedArena instance will notice if it is on a different thread |
| 54 | // than that arena's original thread, and will perform the allocation |
| 55 | // in a side arena for that other thread. This is done in a transparent |
| 56 | // way to the outside; as a result, it is always safe to allocate using |
| 57 | // a MixedArena, no matter which thread you are on. Allocations will |
| 58 | // of course be fastest on the original thread for the arena. |
| 59 | // |
| 60 | |
| 61 | struct MixedArena { |
| 62 | // fast bump allocation |
| 63 | |
| 64 | static const size_t CHUNK_SIZE = 32768; |
| 65 | static const size_t MAX_ALIGN = 16; // allow 128bit SIMD |
| 66 | |
| 67 | // Each pointer in chunks is to a multiple of CHUNK_SIZE - typically 1, |
| 68 | // but possibly more. |
| 69 | std::vector<void*> chunks; |
| 70 | |
| 71 | size_t index = 0; // in last chunk |
| 72 | |
| 73 | std::thread::id threadId; |
| 74 | |
| 75 | // multithreaded allocation - each arena is valid on a specific thread. |
| 76 | // if we are on the wrong thread, we atomically look in the linked |
| 77 | // list of next, adding an allocator if necessary |
| 78 | std::atomic<MixedArena*> next; |
| 79 | |
| 80 | MixedArena() { |
| 81 | threadId = std::this_thread::get_id(); |
| 82 | next.store(nullptr); |
| 83 | } |
| 84 | |
| 85 | // Allocate an amount of space with a guaranteed alignment |
| 86 | void* allocSpace(size_t size, size_t align) { |
| 87 | // the bump allocator data should not be modified by multiple threads at |
| 88 | // once. |
| 89 | auto myId = std::this_thread::get_id(); |
| 90 | if (myId != threadId) { |
| 91 | MixedArena* curr = this; |
| 92 | MixedArena* allocated = nullptr; |
| 93 | while (myId != curr->threadId) { |
| 94 | auto seen = curr->next.load(); |
| 95 | if (seen) { |
| 96 | curr = seen; |
| 97 | continue; |
| 98 | } |
| 99 | // there is a nullptr for next, so we may be able to place a new |
| 100 | // allocator for us there. but carefully, as others may do so as |
| 101 | // well. we may waste a few allocations here, but it doesn't matter |
| 102 | // as this can only happen as the chain is built up, i.e., |
| 103 | // O(# of cores) per allocator, and our allocatrs are long-lived. |
| 104 | if (!allocated) { |
| 105 | allocated = new MixedArena(); // has our thread id |
| 106 | } |
| 107 | if (curr->next.compare_exchange_strong(seen, allocated)) { |
| 108 | // we replaced it, so we are the next in the chain |
| 109 | // we can forget about allocated, it is owned by the chain now |
| 110 | allocated = nullptr; |
| 111 | break; |
| 112 | } |
| 113 | // otherwise, the cmpxchg updated seen, and we continue to loop |
| 114 | curr = seen; |
| 115 | } |
| 116 | if (allocated) { |
| 117 | delete allocated; |
| 118 | } |
| 119 | return curr->allocSpace(size, align); |
| 120 | } |
| 121 | // First, move the current index in the last chunk to an aligned position. |
| 122 | index = (index + align - 1) & (-align); |
| 123 | if (index + size > CHUNK_SIZE || chunks.size() == 0) { |
| 124 | // Allocate a new chunk. |
| 125 | auto numChunks = (size + CHUNK_SIZE - 1) / CHUNK_SIZE; |
| 126 | assert(size <= numChunks * CHUNK_SIZE); |
| 127 | auto* allocation = |
| 128 | wasm::aligned_malloc(MAX_ALIGN, numChunks * CHUNK_SIZE); |
| 129 | if (!allocation) { |
| 130 | abort(); |
| 131 | } |
| 132 | chunks.push_back(allocation); |
| 133 | index = 0; |
| 134 | } |
| 135 | uint8_t* ret = static_cast<uint8_t*>(chunks.back()); |
| 136 | ret += index; |
| 137 | index += size; // TODO: if we allocated more than 1 chunk, reuse the |
| 138 | // remainder, right now we allocate another next time |
| 139 | return static_cast<void*>(ret); |
| 140 | } |
| 141 | |
| 142 | template<class T> T* alloc() { |
| 143 | static_assert(alignof(T) <= MAX_ALIGN, |
| 144 | "maximum alignment not large enough" ); |
| 145 | auto* ret = static_cast<T*>(allocSpace(sizeof(T), alignof(T))); |
| 146 | new (ret) T(*this); // allocated objects receive the allocator, so they can |
| 147 | // allocate more later if necessary |
| 148 | return ret; |
| 149 | } |
| 150 | |
| 151 | void clear() { |
| 152 | for (auto* chunk : chunks) { |
| 153 | wasm::aligned_free(chunk); |
| 154 | } |
| 155 | chunks.clear(); |
| 156 | } |
| 157 | |
| 158 | ~MixedArena() { |
| 159 | clear(); |
| 160 | if (next.load()) { |
| 161 | delete next.load(); |
| 162 | } |
| 163 | } |
| 164 | }; |
| 165 | |
| 166 | // |
| 167 | // A vector that allocates in an arena. |
| 168 | // |
| 169 | // TODO: specialize on the initial size of the array |
| 170 | |
| 171 | template<typename SubType, typename T> class ArenaVectorBase { |
| 172 | protected: |
| 173 | T* data = nullptr; |
| 174 | size_t usedElements = 0, allocatedElements = 0; |
| 175 | |
| 176 | void reallocate(size_t size) { |
| 177 | T* old = data; |
| 178 | static_cast<SubType*>(this)->allocate(size); |
| 179 | for (size_t i = 0; i < usedElements; i++) { |
| 180 | data[i] = old[i]; |
| 181 | } |
| 182 | } |
| 183 | |
| 184 | public: |
| 185 | struct Iterator; |
| 186 | |
| 187 | T& operator[](size_t index) const { |
| 188 | assert(index < usedElements); |
| 189 | return data[index]; |
| 190 | } |
| 191 | |
| 192 | size_t size() const { return usedElements; } |
| 193 | |
| 194 | bool empty() const { return size() == 0; } |
| 195 | |
| 196 | void resize(size_t size) { |
| 197 | if (size > allocatedElements) { |
| 198 | reallocate(size); |
| 199 | } |
| 200 | // construct new elements |
| 201 | for (size_t i = usedElements; i < size; i++) { |
| 202 | new (data + i) T(); |
| 203 | } |
| 204 | usedElements = size; |
| 205 | } |
| 206 | |
| 207 | T& back() const { |
| 208 | assert(usedElements > 0); |
| 209 | return data[usedElements - 1]; |
| 210 | } |
| 211 | |
| 212 | T& pop_back() { |
| 213 | assert(usedElements > 0); |
| 214 | usedElements--; |
| 215 | return data[usedElements]; |
| 216 | } |
| 217 | |
| 218 | void push_back(T item) { |
| 219 | if (usedElements == allocatedElements) { |
| 220 | reallocate((allocatedElements + 1) * 2); // TODO: optimize |
| 221 | } |
| 222 | data[usedElements] = item; |
| 223 | usedElements++; |
| 224 | } |
| 225 | |
| 226 | T& front() const { |
| 227 | assert(usedElements > 0); |
| 228 | return data[0]; |
| 229 | } |
| 230 | |
| 231 | void erase(Iterator start_it, Iterator end_it) { |
| 232 | assert(start_it.parent == end_it.parent && start_it.parent == this); |
| 233 | assert(start_it.index <= end_it.index && end_it.index <= usedElements); |
| 234 | size_t size = end_it.index - start_it.index; |
| 235 | for (size_t cur = start_it.index; cur + size < usedElements; ++cur) { |
| 236 | data[cur] = data[cur + size]; |
| 237 | } |
| 238 | usedElements -= size; |
| 239 | } |
| 240 | |
| 241 | void erase(Iterator it) { erase(it, it + 1); } |
| 242 | |
| 243 | void clear() { usedElements = 0; } |
| 244 | |
| 245 | void reserve(size_t size) { |
| 246 | if (size > allocatedElements) { |
| 247 | reallocate(size); |
| 248 | } |
| 249 | } |
| 250 | |
| 251 | template<typename ListType> void set(const ListType& list) { |
| 252 | size_t size = list.size(); |
| 253 | if (allocatedElements < size) { |
| 254 | static_cast<SubType*>(this)->allocate(size); |
| 255 | } |
| 256 | size_t i = 0; |
| 257 | for (auto elem : list) { |
| 258 | data[i++] = elem; |
| 259 | } |
| 260 | usedElements = size; |
| 261 | } |
| 262 | |
| 263 | void operator=(SubType& other) { set(other); } |
| 264 | |
| 265 | void swap(SubType& other) { |
| 266 | data = other.data; |
| 267 | usedElements = other.usedElements; |
| 268 | allocatedElements = other.allocatedElements; |
| 269 | |
| 270 | other.data = nullptr; |
| 271 | other.usedElements = other.allocatedElements = 0; |
| 272 | } |
| 273 | |
| 274 | // iteration |
| 275 | |
| 276 | struct Iterator { |
| 277 | using iterator_category = std::random_access_iterator_tag; |
| 278 | using value_type = T; |
| 279 | using difference_type = std::ptrdiff_t; |
| 280 | using pointer = T*; |
| 281 | using reference = T&; |
| 282 | |
| 283 | const SubType* parent; |
| 284 | size_t index; |
| 285 | |
| 286 | Iterator() : parent(nullptr), index(0) {} |
| 287 | Iterator(const SubType* parent, size_t index) |
| 288 | : parent(parent), index(index) {} |
| 289 | |
| 290 | bool operator==(const Iterator& other) const { |
| 291 | return index == other.index && parent == other.parent; |
| 292 | } |
| 293 | |
| 294 | bool operator!=(const Iterator& other) const { return !(*this == other); } |
| 295 | |
| 296 | bool operator<(const Iterator& other) const { |
| 297 | assert(parent == other.parent); |
| 298 | return index < other.index; |
| 299 | } |
| 300 | |
| 301 | bool operator>(const Iterator& other) const { return other < *this; } |
| 302 | |
| 303 | bool operator<=(const Iterator& other) const { return !(other < *this); } |
| 304 | |
| 305 | bool operator>=(const Iterator& other) const { return !(*this < other); } |
| 306 | |
| 307 | Iterator& operator++() { |
| 308 | index++; |
| 309 | return *this; |
| 310 | } |
| 311 | |
| 312 | Iterator& operator--() { |
| 313 | index--; |
| 314 | return *this; |
| 315 | } |
| 316 | |
| 317 | Iterator operator++(int) { |
| 318 | Iterator it = *this; |
| 319 | ++*this; |
| 320 | return it; |
| 321 | } |
| 322 | |
| 323 | Iterator operator--(int) { |
| 324 | Iterator it = *this; |
| 325 | --*this; |
| 326 | return it; |
| 327 | } |
| 328 | |
| 329 | Iterator& operator+=(std::ptrdiff_t off) { |
| 330 | index += off; |
| 331 | return *this; |
| 332 | } |
| 333 | |
| 334 | Iterator& operator-=(std::ptrdiff_t off) { return *this += -off; } |
| 335 | |
| 336 | Iterator operator+(std::ptrdiff_t off) const { |
| 337 | return Iterator(*this) += off; |
| 338 | } |
| 339 | |
| 340 | Iterator operator-(std::ptrdiff_t off) const { return *this + -off; } |
| 341 | |
| 342 | std::ptrdiff_t operator-(const Iterator& other) const { |
| 343 | assert(parent == other.parent); |
| 344 | return index - other.index; |
| 345 | } |
| 346 | |
| 347 | friend Iterator operator+(std::ptrdiff_t off, const Iterator& it) { |
| 348 | return it + off; |
| 349 | } |
| 350 | |
| 351 | T& operator*() const { return (*parent)[index]; } |
| 352 | |
| 353 | T& operator[](std::ptrdiff_t off) const { return (*parent)[index + off]; } |
| 354 | |
| 355 | T* operator->() const { return &(*parent)[index]; } |
| 356 | }; |
| 357 | |
| 358 | Iterator begin() const { |
| 359 | return Iterator(static_cast<const SubType*>(this), 0); |
| 360 | } |
| 361 | Iterator end() const { |
| 362 | return Iterator(static_cast<const SubType*>(this), usedElements); |
| 363 | } |
| 364 | |
| 365 | void allocate(size_t size) { |
| 366 | abort(); // must be implemented in children |
| 367 | } |
| 368 | |
| 369 | // C-API |
| 370 | |
| 371 | void insertAt(size_t index, T item) { |
| 372 | assert(index <= size()); // appending is ok |
| 373 | resize(size() + 1); |
| 374 | for (auto i = size() - 1; i > index; --i) { |
| 375 | data[i] = data[i - 1]; |
| 376 | } |
| 377 | data[index] = item; |
| 378 | } |
| 379 | |
| 380 | T removeAt(size_t index) { |
| 381 | assert(index < size()); |
| 382 | auto item = data[index]; |
| 383 | for (auto i = index; i < size() - 1; ++i) { |
| 384 | data[i] = data[i + 1]; |
| 385 | } |
| 386 | resize(size() - 1); |
| 387 | return item; |
| 388 | } |
| 389 | }; |
| 390 | |
| 391 | // A vector that has an allocator for arena allocation |
| 392 | // |
| 393 | // TODO: consider not saving the allocator, but requiring it be |
| 394 | // passed in when needed, would make this (and thus Blocks etc. |
| 395 | // smaller) |
| 396 | |
| 397 | template<typename T> |
| 398 | class ArenaVector : public ArenaVectorBase<ArenaVector<T>, T> { |
| 399 | private: |
| 400 | MixedArena& allocator; |
| 401 | |
| 402 | public: |
| 403 | ArenaVector(MixedArena& allocator) : allocator(allocator) {} |
| 404 | |
| 405 | ArenaVector(ArenaVector<T>&& other) : allocator(other.allocator) { |
| 406 | *this = other; |
| 407 | } |
| 408 | |
| 409 | void allocate(size_t size) { |
| 410 | this->allocatedElements = size; |
| 411 | this->data = static_cast<T*>( |
| 412 | allocator.allocSpace(sizeof(T) * this->allocatedElements, alignof(T))); |
| 413 | } |
| 414 | }; |
| 415 | |
| 416 | #endif // wasm_mixed_arena_h |
| 417 | |