| 1 | /* |
| 2 | * Copyright 2020 Google LLC |
| 3 | * |
| 4 | * Use of this source code is governed by a BSD-style license that can be |
| 5 | * found in the LICENSE file. |
| 6 | */ |
| 7 | |
| 8 | #ifndef GrBlockAllocator_DEFINED |
| 9 | #define GrBlockAllocator_DEFINED |
| 10 | |
| 11 | #include "include/private/GrTypesPriv.h" |
| 12 | #include "include/private/SkNoncopyable.h" |
| 13 | |
| 14 | #include <memory> // std::unique_ptr |
| 15 | #include <cstddef> // max_align_t |
| 16 | |
| 17 | /** |
| 18 | * GrBlockAllocator provides low-level support for a block allocated arena with a dynamic tail that |
| 19 | * tracks space reservations within each block. Its APIs provide the ability to reserve space, |
| 20 | * resize reservations, and release reservations. It will automatically create new blocks if needed |
| 21 | * and destroy all remaining blocks when it is destructed. It assumes that anything allocated within |
| 22 | * its blocks has its destructors called externally. It is recommended that GrBlockAllocator is |
| 23 | * wrapped by a higher-level allocator that uses the low-level APIs to implement a simpler, |
| 24 | * purpose-focused API w/o having to worry as much about byte-level concerns. |
| 25 | * |
| 26 | * GrBlockAllocator has no limit to its total size, but each allocation is limited to 512MB (which |
| 27 | * should be sufficient for Ganesh's use cases). This upper allocation limit allows all internal |
| 28 | * operations to be performed using 'int' and avoid many overflow checks. Static asserts are used |
| 29 | * to ensure that those operations would not overflow when using the largest possible values. |
| 30 | * |
| 31 | * Possible use modes: |
| 32 | * 1. No upfront allocation, either on the stack or as a field |
| 33 | * GrBlockAllocator allocator(policy, heapAllocSize); |
| 34 | * |
| 35 | * 2. In-place new'd |
| 36 | * void* mem = operator new(totalSize); |
| 37 | * GrBlockAllocator* allocator = new (mem) GrBlockAllocator(policy, heapAllocSize, |
| 38 | * totalSize- sizeof(GrBlockAllocator)); |
| 39 | * delete allocator; |
| 40 | * |
| 41 | * 3. Use GrSBlockAllocator to increase the preallocation size |
| 42 | * GrSBlockAllocator<1024> allocator(policy, heapAllocSize); |
| 43 | * sizeof(allocator) == 1024; |
| 44 | */ |
| 45 | class GrBlockAllocator final : SkNoncopyable { |
| 46 | public: |
| 47 | // Largest size that can be requested from allocate(), chosen because it's the largest pow-2 |
| 48 | // that is less than int32_t::max()/2. |
| 49 | static constexpr int kMaxAllocationSize = 1 << 29; |
| 50 | |
| 51 | enum class GrowthPolicy : int { |
| 52 | kFixed, // Next block size = N |
| 53 | kLinear, // = #blocks * N |
| 54 | kFibonacci, // = fibonacci(#blocks) * N |
| 55 | kExponential, // = 2^#blocks * N |
| 56 | kLast = kExponential |
| 57 | }; |
| 58 | static constexpr int kGrowthPolicyCount = static_cast<int>(GrowthPolicy::kLast) + 1; |
| 59 | |
| 60 | class Block; |
| 61 | |
| 62 | // Tuple representing a range of bytes, marking the unaligned start, the first aligned point |
| 63 | // after any padding, and the upper limit depending on requested size. |
| 64 | struct ByteRange { |
| 65 | Block* fBlock; // Owning block |
| 66 | int fStart; // Inclusive byte lower limit of byte range |
| 67 | int fAlignedOffset; // >= start, matching alignment requirement (i.e. first real byte) |
| 68 | int fEnd; // Exclusive upper limit of byte range |
| 69 | }; |
| 70 | |
| 71 | class Block final { |
| 72 | public: |
| 73 | ~Block(); |
| 74 | void operator delete(void* p) { ::operator delete(p); } |
| 75 | |
| 76 | // Return the maximum allocation size with the given alignment that can fit in this block. |
| 77 | template <size_t Align = 1, size_t Padding = 0> |
| 78 | int avail() const { return std::max(0, fSize - this->cursor<Align, Padding>()); } |
| 79 | |
| 80 | // Return the aligned offset of the first allocation, assuming it was made with the |
| 81 | // specified Align, and Padding. The returned offset does not mean a valid allocation |
| 82 | // starts at that offset, this is a utility function for classes built on top to manage |
| 83 | // indexing into a block effectively. |
| 84 | template <size_t Align = 1, size_t Padding = 0> |
| 85 | int firstAlignedOffset() const { return this->alignedOffset<Align, Padding>(kDataStart); } |
| 86 | |
| 87 | // Convert an offset into this block's storage into a usable pointer. |
| 88 | void* ptr(int offset) { |
| 89 | SkASSERT(offset >= kDataStart && offset < fSize); |
| 90 | return reinterpret_cast<char*>(this) + offset; |
| 91 | } |
| 92 | const void* ptr(int offset) const { return const_cast<Block*>(this)->ptr(offset); } |
| 93 | |
| 94 | // Every block has an extra 'int' for clients to use however they want. It will start |
| 95 | // at 0 when a new block is made, or when the head block is reset. |
| 96 | int metadata() const { return fMetadata; } |
| 97 | void setMetadata(int value) { fMetadata = value; } |
| 98 | |
| 99 | /** |
| 100 | * Release the byte range between offset 'start' (inclusive) and 'end' (exclusive). This |
| 101 | * will return true if those bytes were successfully reclaimed, i.e. a subsequent allocation |
| 102 | * request could occupy the space. Regardless of return value, the provided byte range that |
| 103 | * [start, end) represents should not be used until it's re-allocated with allocate<...>(). |
| 104 | */ |
| 105 | inline bool release(int start, int end); |
| 106 | |
| 107 | /** |
| 108 | * Resize a previously reserved byte range of offset 'start' (inclusive) to 'end' |
| 109 | * (exclusive). 'deltaBytes' is the SIGNED change to length of the reservation. |
| 110 | * |
| 111 | * When negative this means the reservation is shrunk and the new length is (end - start - |
| 112 | * |deltaBytes|). If this new length would be 0, the byte range can no longer be used (as if |
| 113 | * it were released instead). Asserts that it would not shrink the reservation below 0. |
| 114 | * |
| 115 | * If 'deltaBytes' is positive, the allocator attempts to increase the length of the |
| 116 | * reservation. If 'deltaBytes' is less than or equal to avail() and it was the last |
| 117 | * allocation in the block, it can be resized. If there is not enough available bytes to |
| 118 | * accommodate the increase in size, or another allocation is blocking the increase in size, |
| 119 | * then false will be returned and the reserved byte range is unmodified. |
| 120 | */ |
| 121 | inline bool resize(int start, int end, int deltaBytes); |
| 122 | |
| 123 | private: |
| 124 | friend class GrBlockAllocator; |
| 125 | |
| 126 | Block(Block* prev, int allocationSize); |
| 127 | |
| 128 | // Get fCursor, but aligned such that ptr(rval) satisfies Align. |
| 129 | template <size_t Align, size_t Padding> |
| 130 | int cursor() const { return this->alignedOffset<Align, Padding>(fCursor); } |
| 131 | |
| 132 | template <size_t Align, size_t Padding> |
| 133 | int alignedOffset(int offset) const; |
| 134 | |
| 135 | bool isScratch() const { return fCursor < 0; } |
| 136 | void markAsScratch() { fCursor = -1; } |
| 137 | |
| 138 | SkDEBUGCODE(int fSentinel;) // known value to check for bad back pointers to blocks |
| 139 | |
| 140 | Block* fNext; // doubly-linked list of blocks |
| 141 | Block* fPrev; |
| 142 | |
| 143 | // Each block tracks its own cursor because as later blocks are released, an older block |
| 144 | // may become the active tail again. |
| 145 | int fSize; // includes the size of the BlockHeader and requested metadata |
| 146 | int fCursor; // (this + fCursor) points to next available allocation |
| 147 | int fMetadata; |
| 148 | |
| 149 | // On release builds, a Block's other 2 pointers and 3 int fields leaves 4 bytes of padding |
| 150 | // for 8 and 16 aligned systems. Currently this is only manipulated in the head block for |
| 151 | // an allocator-level metadata and is explicitly not reset when the head block is "released" |
| 152 | // Down the road we could instead choose to offer multiple metadata slots per block. |
| 153 | int fAllocatorMetadata; |
| 154 | }; |
| 155 | |
| 156 | // The size of the head block is determined by 'additionalPreallocBytes'. Subsequent heap blocks |
| 157 | // are determined by 'policy' and 'blockIncrementBytes', although 'blockIncrementBytes' will be |
| 158 | // aligned to std::max_align_t. |
| 159 | // |
| 160 | // When 'additionalPreallocBytes' > 0, the allocator assumes that many extra bytes immediately |
| 161 | // after the allocator can be used by its inline head block. This is useful when the allocator |
| 162 | // is in-place new'ed into a larger block of memory, but it should remain set to 0 if stack |
| 163 | // allocated or if the class layout does not guarantee that space is present. |
| 164 | GrBlockAllocator(GrowthPolicy policy, size_t blockIncrementBytes, |
| 165 | size_t additionalPreallocBytes = 0); |
| 166 | |
| 167 | ~GrBlockAllocator() { this->reset(); } |
| 168 | void operator delete(void* p) { ::operator delete(p); } |
| 169 | |
| 170 | /** |
| 171 | * Helper to calculate the minimum number of bytes needed for heap block size, under the |
| 172 | * assumption that Align will be the requested alignment of the first call to allocate(). |
| 173 | * Ex. To store N instances of T in a heap block, the 'blockIncrementBytes' should be set to |
| 174 | * BlockOverhead<alignof(T)>() + N * sizeof(T) when making the GrBlockAllocator. |
| 175 | */ |
| 176 | template<size_t Align = 1, size_t Padding = 0> |
| 177 | static constexpr size_t BlockOverhead(); |
| 178 | |
| 179 | /** |
| 180 | * Helper to calculate the minimum number of bytes needed for a preallocation, under the |
| 181 | * assumption that Align will be the requested alignment of the first call to allocate(). |
| 182 | * Ex. To preallocate a GrSBlockAllocator to hold N instances of T, its arge should be |
| 183 | * Overhead<alignof(T)>() + N * sizeof(T) |
| 184 | */ |
| 185 | template<size_t Align = 1, size_t Padding = 0> |
| 186 | static constexpr size_t Overhead(); |
| 187 | |
| 188 | /** |
| 189 | * Return the total number of bytes of the allocator, including its instance overhead, per-block |
| 190 | * overhead and space used for allocations. |
| 191 | */ |
| 192 | size_t totalSize() const; |
| 193 | /** |
| 194 | * Return the total number of bytes usable for allocations. This includes bytes that have |
| 195 | * been reserved already by a call to allocate() and bytes that are still available. It is |
| 196 | * totalSize() minus all allocator and block-level overhead. |
| 197 | */ |
| 198 | size_t totalUsableSpace() const; |
| 199 | /** |
| 200 | * Return the total number of usable bytes that have been reserved by allocations. This will |
| 201 | * be less than or equal to totalUsableSpace(). |
| 202 | */ |
| 203 | size_t totalSpaceInUse() const; |
| 204 | |
| 205 | /** |
| 206 | * Return the total number of bytes that were pre-allocated for the GrBlockAllocator. This will |
| 207 | * include 'additionalPreallocBytes' passed to the constructor, and represents what the total |
| 208 | * size would become after a call to reset(). |
| 209 | */ |
| 210 | size_t preallocSize() const { |
| 211 | // Don't double count fHead's Block overhead in both sizeof(GrBlockAllocator) and fSize. |
| 212 | return sizeof(GrBlockAllocator) + fHead.fSize - BaseHeadBlockSize(); |
| 213 | } |
| 214 | /** |
| 215 | * Return the usable size of the inline head block; this will be equal to |
| 216 | * 'additionalPreallocBytes' plus any alignment padding that the system had to add to Block. |
| 217 | * The returned value represents what could be allocated before a heap block is be created. |
| 218 | */ |
| 219 | size_t preallocUsableSpace() const { |
| 220 | return fHead.fSize - kDataStart; |
| 221 | } |
| 222 | |
| 223 | /** |
| 224 | * Get the current value of the allocator-level metadata (a user-oriented slot). This is |
| 225 | * separate from any block-level metadata, but can serve a similar purpose to compactly support |
| 226 | * data collections on top of GrBlockAllocator. |
| 227 | */ |
| 228 | int metadata() const { return fHead.fAllocatorMetadata; } |
| 229 | |
| 230 | /** |
| 231 | * Set the current value of the allocator-level metadata. |
| 232 | */ |
| 233 | void setMetadata(int value) { fHead.fAllocatorMetadata = value; } |
| 234 | |
| 235 | /** |
| 236 | * Reserve space that will hold 'size' bytes. This will automatically allocate a new block if |
| 237 | * there is not enough available space in the current block to provide 'size' bytes. The |
| 238 | * returned ByteRange tuple specifies the Block owning the reserved memory, the full byte range, |
| 239 | * and the aligned offset within that range to use for the user-facing pointer. The following |
| 240 | * invariants hold: |
| 241 | * |
| 242 | * 1. block->ptr(alignedOffset) is aligned to Align |
| 243 | * 2. end - alignedOffset == size |
| 244 | * 3. Padding <= alignedOffset - start <= Padding + Align - 1 |
| 245 | * |
| 246 | * Invariant #3, when Padding > 0, allows intermediate allocators to embed metadata along with |
| 247 | * the allocations. If the Padding bytes are used for some 'struct Meta', then |
| 248 | * ptr(alignedOffset - sizeof(Meta)) can be safely used as a Meta* if Meta's alignment |
| 249 | * requirements are less than or equal to the alignment specified in allocate<>. This can be |
| 250 | * easily guaranteed by using the pattern: |
| 251 | * |
| 252 | * allocate<max(UserAlign, alignof(Meta)), sizeof(Meta)>(userSize); |
| 253 | * |
| 254 | * This ensures that ptr(alignedOffset) will always satisfy UserAlign and |
| 255 | * ptr(alignedOffset - sizeof(Meta)) will always satisfy alignof(Meta). Alternatively, memcpy |
| 256 | * can be used to read and write values between start and alignedOffset without worrying about |
| 257 | * alignment requirements of the metadata. |
| 258 | * |
| 259 | * For over-aligned allocations, the alignedOffset (as an int) may not be a multiple of Align, |
| 260 | * but the result of ptr(alignedOffset) will be a multiple of Align. |
| 261 | */ |
| 262 | template <size_t Align, size_t Padding = 0> |
| 263 | ByteRange allocate(size_t size); |
| 264 | |
| 265 | enum ReserveFlags : unsigned { |
| 266 | // If provided to reserve(), the input 'size' will be rounded up to the next size determined |
| 267 | // by the growth policy of the GrBlockAllocator. If not, 'size' will be aligned to max_align |
| 268 | kIgnoreGrowthPolicy_Flag = 0b01, |
| 269 | // If provided to reserve(), the number of available bytes of the current block will not |
| 270 | // be used to satisfy the reservation (assuming the contiguous range was long enough to |
| 271 | // begin with). |
| 272 | kIgnoreExistingBytes_Flag = 0b10, |
| 273 | |
| 274 | kNo_ReserveFlags = 0b00 |
| 275 | }; |
| 276 | |
| 277 | /** |
| 278 | * Ensure the block allocator has 'size' contiguous available bytes. After calling this |
| 279 | * function, currentBlock()->avail<Align, Padding>() may still report less than 'size' if the |
| 280 | * reserved space was added as a scratch block. This is done so that anything remaining in |
| 281 | * the current block can still be used if a smaller-than-size allocation is requested. If 'size' |
| 282 | * is requested by a subsequent allocation, the scratch block will automatically be activated |
| 283 | * and the request will not itself trigger any malloc. |
| 284 | * |
| 285 | * The optional 'flags' controls how the input size is allocated; by default it will attempt |
| 286 | * to use available contiguous bytes in the current block and will respect the growth policy |
| 287 | * of the allocator. |
| 288 | */ |
| 289 | template <size_t Align = 1, size_t Padding = 0> |
| 290 | void reserve(size_t size, ReserveFlags flags = kNo_ReserveFlags); |
| 291 | |
| 292 | /** |
| 293 | * Return a pointer to the start of the current block. This will never be null. |
| 294 | */ |
| 295 | const Block* currentBlock() const { return fTail; } |
| 296 | Block* currentBlock() { return fTail; } |
| 297 | |
| 298 | const Block* headBlock() const { return &fHead; } |
| 299 | Block* headBlock() { return &fHead; } |
| 300 | |
| 301 | /** |
| 302 | * Return the block that owns the allocated 'ptr'. Assuming that earlier, an allocation was |
| 303 | * returned as {b, start, alignedOffset, end}, and 'p = b->ptr(alignedOffset)', then a call |
| 304 | * to 'owningBlock<Align, Padding>(p, start) == b'. |
| 305 | * |
| 306 | * If calling code has already made a pointer to their metadata, i.e. 'm = p - Padding', then |
| 307 | * 'owningBlock<Align, 0>(m, start)' will also return b, allowing you to recover the block from |
| 308 | * the metadata pointer. |
| 309 | * |
| 310 | * If calling code has access to the original alignedOffset, this function should not be used |
| 311 | * since the owning block is just 'p - alignedOffset', regardless of original Align or Padding. |
| 312 | */ |
| 313 | template <size_t Align, size_t Padding = 0> |
| 314 | Block* owningBlock(const void* ptr, int start); |
| 315 | |
| 316 | template <size_t Align, size_t Padding = 0> |
| 317 | const Block* owningBlock(const void* ptr, int start) const { |
| 318 | return const_cast<GrBlockAllocator*>(this)->owningBlock<Align, Padding>(ptr, start); |
| 319 | } |
| 320 | |
| 321 | /** |
| 322 | * Find the owning block of the allocated pointer, 'p'. Without any additional information this |
| 323 | * is O(N) on the number of allocated blocks. |
| 324 | */ |
| 325 | Block* findOwningBlock(const void* ptr); |
| 326 | const Block* findOwningBlock(const void* ptr) const { |
| 327 | return const_cast<GrBlockAllocator*>(this)->findOwningBlock(ptr); |
| 328 | } |
| 329 | |
| 330 | /** |
| 331 | * Explicitly free an entire block, invalidating any remaining allocations from the block. |
| 332 | * GrBlockAllocator will release all alive blocks automatically when it is destroyed, but this |
| 333 | * function can be used to reclaim memory over the lifetime of the allocator. The provided |
| 334 | * 'block' pointer must have previously come from a call to currentBlock() or allocate(). |
| 335 | * |
| 336 | * If 'block' represents the inline-allocated head block, its cursor and metadata are instead |
| 337 | * reset to their defaults. |
| 338 | * |
| 339 | * If the block is not the head block, it may be kept as a scratch block to be reused for |
| 340 | * subsequent allocation requests, instead of making an entirely new block. A scratch block is |
| 341 | * not visible when iterating over blocks but is reported in the total size of the allocator. |
| 342 | */ |
| 343 | void releaseBlock(Block* block); |
| 344 | |
| 345 | /** |
| 346 | * Detach every heap-allocated block owned by 'other' and concatenate them to this allocator's |
| 347 | * list of blocks. This memory is now managed by this allocator. Since this only transfers |
| 348 | * ownership of a Block, and a Block itself does not move, any previous allocations remain |
| 349 | * valid and associated with their original Block instances. GrBlockAllocator-level functions |
| 350 | * that accept allocated pointers (e.g. findOwningBlock), must now use this allocator and not |
| 351 | * 'other' for these allocations. |
| 352 | * |
| 353 | * The head block of 'other' cannot be stolen, so higher-level allocators and memory structures |
| 354 | * must handle that data differently. |
| 355 | */ |
| 356 | void stealHeapBlocks(GrBlockAllocator* other); |
| 357 | |
| 358 | /** |
| 359 | * Explicitly free all blocks (invalidating all allocations), and resets the head block to its |
| 360 | * default state. The allocator-level metadata is reset to 0 as well. |
| 361 | */ |
| 362 | void reset(); |
| 363 | |
| 364 | /** |
| 365 | * Remove any reserved scratch space, either from calling reserve() or releaseBlock(). |
| 366 | */ |
| 367 | void resetScratchSpace(); |
| 368 | |
| 369 | template <bool Forward, bool Const> class BlockIter; |
| 370 | |
| 371 | /** |
| 372 | * Clients can iterate over all active Blocks in the GrBlockAllocator using for loops: |
| 373 | * |
| 374 | * Forward iteration from head to tail block (or non-const variant): |
| 375 | * for (const Block* b : this->blocks()) { } |
| 376 | * Reverse iteration from tail to head block: |
| 377 | * for (const Block* b : this->rblocks()) { } |
| 378 | * |
| 379 | * It is safe to call releaseBlock() on the active block while looping. |
| 380 | */ |
| 381 | inline BlockIter<true, false> blocks(); |
| 382 | inline BlockIter<true, true> blocks() const; |
| 383 | inline BlockIter<false, false> rblocks(); |
| 384 | inline BlockIter<false, true> rblocks() const; |
| 385 | |
| 386 | #ifdef SK_DEBUG |
| 387 | static constexpr int kAssignedMarker = 0xBEEFFACE; |
| 388 | static constexpr int kFreedMarker = 0xCAFEBABE; |
| 389 | |
| 390 | void validate() const; |
| 391 | #endif |
| 392 | |
| 393 | #if GR_TEST_UTILS |
| 394 | int testingOnly_scratchBlockSize() const { return this->scratchBlockSize(); } |
| 395 | #endif |
| 396 | |
| 397 | private: |
| 398 | static constexpr int kDataStart = sizeof(Block); |
| 399 | #ifdef SK_FORCE_8_BYTE_ALIGNMENT |
| 400 | // This is an issue for WASM builds using emscripten, which had std::max_align_t = 16, but |
| 401 | // was returning pointers only aligned to 8 bytes. |
| 402 | // https://github.com/emscripten-core/emscripten/issues/10072 |
| 403 | // |
| 404 | // Setting this to 8 will let GrBlockAllocator properly correct for the pointer address if |
| 405 | // a 16-byte aligned allocation is requested in wasm (unlikely since we don't use long |
| 406 | // doubles). |
| 407 | static constexpr size_t kAddressAlign = 8; |
| 408 | #else |
| 409 | // The alignment Block addresses will be at when created using operator new |
| 410 | // (spec-compliant is pointers are aligned to max_align_t). |
| 411 | static constexpr size_t kAddressAlign = alignof(std::max_align_t); |
| 412 | #endif |
| 413 | |
| 414 | // Calculates the size of a new Block required to store a kMaxAllocationSize request for the |
| 415 | // given alignment and padding bytes. Also represents maximum valid fCursor value in a Block. |
| 416 | template<size_t Align, size_t Padding> |
| 417 | static constexpr size_t MaxBlockSize(); |
| 418 | |
| 419 | static constexpr int BaseHeadBlockSize() { |
| 420 | return sizeof(GrBlockAllocator) - offsetof(GrBlockAllocator, fHead); |
| 421 | } |
| 422 | |
| 423 | // Append a new block to the end of the block linked list, updating fTail. 'minSize' must |
| 424 | // have enough room for sizeof(Block). 'maxSize' is the upper limit of fSize for the new block |
| 425 | // that will preserve the static guarantees GrBlockAllocator makes. |
| 426 | void addBlock(int minSize, int maxSize); |
| 427 | |
| 428 | int scratchBlockSize() const { return fHead.fPrev ? fHead.fPrev->fSize : 0; } |
| 429 | |
| 430 | Block* fTail; // All non-head blocks are heap allocated; tail will never be null. |
| 431 | |
| 432 | // All remaining state is packed into 64 bits to keep GrBlockAllocator at 16 bytes + head block |
| 433 | // (on a 64-bit system). |
| 434 | |
| 435 | // Growth of the block size is controlled by four factors: BlockIncrement, N0 and N1, and a |
| 436 | // policy defining how N0 is updated. When a new block is needed, we calculate N1' = N0 + N1. |
| 437 | // Depending on the policy, N0' = N0 (no growth or linear growth), or N0' = N1 (Fibonacci), or |
| 438 | // N0' = N1' (exponential). The size of the new block is N1' * BlockIncrement * MaxAlign, |
| 439 | // after which fN0 and fN1 store N0' and N1' clamped into 23 bits. With current bit allocations, |
| 440 | // N1' is limited to 2^24, and assuming MaxAlign=16, then BlockIncrement must be '2' in order to |
| 441 | // eventually reach the hard 2^29 size limit of GrBlockAllocator. |
| 442 | |
| 443 | // Next heap block size = (fBlockIncrement * alignof(std::max_align_t) * (fN0 + fN1)) |
| 444 | uint64_t fBlockIncrement : 16; |
| 445 | uint64_t fGrowthPolicy : 2; // GrowthPolicy |
| 446 | uint64_t fN0 : 23; // = 1 for linear/exp.; = 0 for fixed/fibonacci, initially |
| 447 | uint64_t fN1 : 23; // = 1 initially |
| 448 | |
| 449 | // Inline head block, must be at the end so that it can utilize any additional reserved space |
| 450 | // from the initial allocation. |
| 451 | // The head block's prev pointer may be non-null, which signifies a scratch block that may be |
| 452 | // reused instead of allocating an entirely new block (this helps when allocate+release calls |
| 453 | // bounce back and forth across the capacity of a block). |
| 454 | alignas(kAddressAlign) Block fHead; |
| 455 | |
| 456 | static_assert(kGrowthPolicyCount <= 4); |
| 457 | }; |
| 458 | |
| 459 | // A wrapper around GrBlockAllocator that includes preallocated storage for the head block. |
| 460 | // N will be the preallocSize() reported by the allocator. |
| 461 | template<size_t N> |
| 462 | class GrSBlockAllocator : SkNoncopyable { |
| 463 | public: |
| 464 | using GrowthPolicy = GrBlockAllocator::GrowthPolicy; |
| 465 | |
| 466 | GrSBlockAllocator() { |
| 467 | new (fStorage) GrBlockAllocator(GrowthPolicy::kFixed, N, N - sizeof(GrBlockAllocator)); |
| 468 | } |
| 469 | explicit GrSBlockAllocator(GrowthPolicy policy) { |
| 470 | new (fStorage) GrBlockAllocator(policy, N, N - sizeof(GrBlockAllocator)); |
| 471 | } |
| 472 | |
| 473 | GrSBlockAllocator(GrowthPolicy policy, size_t blockIncrementBytes) { |
| 474 | new (fStorage) GrBlockAllocator(policy, blockIncrementBytes, N - sizeof(GrBlockAllocator)); |
| 475 | } |
| 476 | |
| 477 | ~GrSBlockAllocator() { |
| 478 | this->allocator()->~GrBlockAllocator(); |
| 479 | } |
| 480 | |
| 481 | GrBlockAllocator* operator->() { return this->allocator(); } |
| 482 | const GrBlockAllocator* operator->() const { return this->allocator(); } |
| 483 | |
| 484 | GrBlockAllocator* allocator() { return reinterpret_cast<GrBlockAllocator*>(fStorage); } |
| 485 | const GrBlockAllocator* allocator() const { |
| 486 | return reinterpret_cast<const GrBlockAllocator*>(fStorage); |
| 487 | } |
| 488 | |
| 489 | private: |
| 490 | static_assert(N >= sizeof(GrBlockAllocator)); |
| 491 | |
| 492 | // Will be used to placement new the allocator |
| 493 | alignas(GrBlockAllocator) char fStorage[N]; |
| 494 | }; |
| 495 | |
| 496 | /////////////////////////////////////////////////////////////////////////////////////////////////// |
| 497 | // Template and inline implementations |
| 498 | |
| 499 | GR_MAKE_BITFIELD_OPS(GrBlockAllocator::ReserveFlags) |
| 500 | |
| 501 | template<size_t Align, size_t Padding> |
| 502 | constexpr size_t GrBlockAllocator::BlockOverhead() { |
| 503 | static_assert(GrAlignTo(kDataStart + Padding, Align) >= sizeof(Block)); |
| 504 | return GrAlignTo(kDataStart + Padding, Align); |
| 505 | } |
| 506 | |
| 507 | template<size_t Align, size_t Padding> |
| 508 | constexpr size_t GrBlockAllocator::Overhead() { |
| 509 | // NOTE: On most platforms, GrBlockAllocator is packed; this is not the case on debug builds |
| 510 | // due to extra fields, or on WASM due to 4byte pointers but 16byte max align. |
| 511 | return std::max(sizeof(GrBlockAllocator), |
| 512 | offsetof(GrBlockAllocator, fHead) + BlockOverhead<Align, Padding>()); |
| 513 | } |
| 514 | |
| 515 | template<size_t Align, size_t Padding> |
| 516 | constexpr size_t GrBlockAllocator::MaxBlockSize() { |
| 517 | // Without loss of generality, assumes 'align' will be the largest encountered alignment for the |
| 518 | // allocator (if it's not, the largest align will be encountered by the compiler and pass/fail |
| 519 | // the same set of static asserts). |
| 520 | return BlockOverhead<Align, Padding>() + kMaxAllocationSize; |
| 521 | } |
| 522 | |
| 523 | template<size_t Align, size_t Padding> |
| 524 | void GrBlockAllocator::reserve(size_t size, ReserveFlags flags) { |
| 525 | if (size > kMaxAllocationSize) { |
| 526 | SK_ABORT("Allocation too large (%zu bytes requested)" , size); |
| 527 | } |
| 528 | int iSize = (int) size; |
| 529 | if ((flags & kIgnoreExistingBytes_Flag) || |
| 530 | this->currentBlock()->avail<Align, Padding>() < iSize) { |
| 531 | |
| 532 | int blockSize = BlockOverhead<Align, Padding>() + iSize; |
| 533 | int maxSize = (flags & kIgnoreGrowthPolicy_Flag) ? blockSize |
| 534 | : MaxBlockSize<Align, Padding>(); |
| 535 | SkASSERT((size_t) maxSize <= (MaxBlockSize<Align, Padding>())); |
| 536 | |
| 537 | SkDEBUGCODE(auto oldTail = fTail;) |
| 538 | this->addBlock(blockSize, maxSize); |
| 539 | SkASSERT(fTail != oldTail); |
| 540 | // Releasing the just added block will move it into scratch space, allowing the original |
| 541 | // tail's bytes to be used first before the scratch block is activated. |
| 542 | this->releaseBlock(fTail); |
| 543 | } |
| 544 | } |
| 545 | |
| 546 | template <size_t Align, size_t Padding> |
| 547 | GrBlockAllocator::ByteRange GrBlockAllocator::allocate(size_t size) { |
| 548 | // Amount of extra space for a new block to make sure the allocation can succeed. |
| 549 | static constexpr int kBlockOverhead = (int) BlockOverhead<Align, Padding>(); |
| 550 | |
| 551 | // Ensures 'offset' and 'end' calculations will be valid |
| 552 | static_assert((kMaxAllocationSize + GrAlignTo(MaxBlockSize<Align, Padding>(), Align)) |
| 553 | <= (size_t) std::numeric_limits<int32_t>::max()); |
| 554 | // Ensures size + blockOverhead + addBlock's alignment operations will be valid |
| 555 | static_assert(kMaxAllocationSize + kBlockOverhead + ((1 << 12) - 1) // 4K align for large blocks |
| 556 | <= std::numeric_limits<int32_t>::max()); |
| 557 | |
| 558 | if (size > kMaxAllocationSize) { |
| 559 | SK_ABORT("Allocation too large (%zu bytes requested)" , size); |
| 560 | } |
| 561 | |
| 562 | int iSize = (int) size; |
| 563 | int offset = fTail->cursor<Align, Padding>(); |
| 564 | int end = offset + iSize; |
| 565 | if (end > fTail->fSize) { |
| 566 | this->addBlock(iSize + kBlockOverhead, MaxBlockSize<Align, Padding>()); |
| 567 | offset = fTail->cursor<Align, Padding>(); |
| 568 | end = offset + iSize; |
| 569 | } |
| 570 | |
| 571 | // Check invariants |
| 572 | SkASSERT(end <= fTail->fSize); |
| 573 | SkASSERT(end - offset == iSize); |
| 574 | SkASSERT(offset - fTail->fCursor >= (int) Padding && |
| 575 | offset - fTail->fCursor <= (int) (Padding + Align - 1)); |
| 576 | SkASSERT(reinterpret_cast<uintptr_t>(fTail->ptr(offset)) % Align == 0); |
| 577 | |
| 578 | int start = fTail->fCursor; |
| 579 | fTail->fCursor = end; |
| 580 | return {fTail, start, offset, end}; |
| 581 | } |
| 582 | |
| 583 | template <size_t Align, size_t Padding> |
| 584 | GrBlockAllocator::Block* GrBlockAllocator::owningBlock(const void* p, int start) { |
| 585 | // 'p' was originally formed by aligning 'block + start + Padding', producing the inequality: |
| 586 | // block + start + Padding <= p <= block + start + Padding + Align-1 |
| 587 | // Rearranging this yields: |
| 588 | // block <= p - start - Padding <= block + Align-1 |
| 589 | // Masking these terms by ~(Align-1) reconstructs 'block' if the alignment of the block is |
| 590 | // greater than or equal to Align (since block & ~(Align-1) == (block + Align-1) & ~(Align-1) |
| 591 | // in that case). Overalignment does not reduce to inequality unfortunately. |
| 592 | if /* constexpr */ (Align <= kAddressAlign) { |
| 593 | Block* block = reinterpret_cast<Block*>( |
| 594 | (reinterpret_cast<uintptr_t>(p) - start - Padding) & ~(Align - 1)); |
| 595 | SkASSERT(block->fSentinel == kAssignedMarker); |
| 596 | return block; |
| 597 | } else { |
| 598 | // There's not a constant-time expression available to reconstruct the block from 'p', |
| 599 | // but this is unlikely to happen frequently. |
| 600 | return this->findOwningBlock(p); |
| 601 | } |
| 602 | } |
| 603 | |
| 604 | template <size_t Align, size_t Padding> |
| 605 | int GrBlockAllocator::Block::alignedOffset(int offset) const { |
| 606 | static_assert(SkIsPow2(Align)); |
| 607 | // Aligning adds (Padding + Align - 1) as an intermediate step, so ensure that can't overflow |
| 608 | static_assert(MaxBlockSize<Align, Padding>() + Padding + Align - 1 |
| 609 | <= (size_t) std::numeric_limits<int32_t>::max()); |
| 610 | |
| 611 | if /* constexpr */ (Align <= kAddressAlign) { |
| 612 | // Same as GrAlignTo, but operates on ints instead of size_t |
| 613 | return (offset + Padding + Align - 1) & ~(Align - 1); |
| 614 | } else { |
| 615 | // Must take into account that 'this' may be starting at a pointer that doesn't satisfy the |
| 616 | // larger alignment request, so must align the entire pointer, not just offset |
| 617 | uintptr_t blockPtr = reinterpret_cast<uintptr_t>(this); |
| 618 | uintptr_t alignedPtr = (blockPtr + offset + Padding + Align - 1) & ~(Align - 1); |
| 619 | SkASSERT(alignedPtr - blockPtr <= (uintptr_t) std::numeric_limits<int32_t>::max()); |
| 620 | return (int) (alignedPtr - blockPtr); |
| 621 | } |
| 622 | } |
| 623 | |
| 624 | bool GrBlockAllocator::Block::resize(int start, int end, int deltaBytes) { |
| 625 | SkASSERT(fSentinel == kAssignedMarker); |
| 626 | SkASSERT(start >= kDataStart && end <= fSize && start < end); |
| 627 | |
| 628 | if (deltaBytes > kMaxAllocationSize || deltaBytes < -kMaxAllocationSize) { |
| 629 | // Cannot possibly satisfy the resize and could overflow subsequent math |
| 630 | return false; |
| 631 | } |
| 632 | if (fCursor == end) { |
| 633 | int nextCursor = end + deltaBytes; |
| 634 | SkASSERT(nextCursor >= start); |
| 635 | // We still check nextCursor >= start for release builds that wouldn't assert. |
| 636 | if (nextCursor <= fSize && nextCursor >= start) { |
| 637 | fCursor = nextCursor; |
| 638 | return true; |
| 639 | } |
| 640 | } |
| 641 | return false; |
| 642 | } |
| 643 | |
| 644 | // NOTE: release is equivalent to resize(start, end, start - end), and the compiler can optimize |
| 645 | // most of the operations away, but it wasn't able to remove the unnecessary branch comparing the |
| 646 | // new cursor to the block size or old start, so release() gets a specialization. |
| 647 | bool GrBlockAllocator::Block::release(int start, int end) { |
| 648 | SkASSERT(fSentinel == kAssignedMarker); |
| 649 | SkASSERT(start >= kDataStart && end <= fSize && start < end); |
| 650 | if (fCursor == end) { |
| 651 | fCursor = start; |
| 652 | return true; |
| 653 | } else { |
| 654 | return false; |
| 655 | } |
| 656 | } |
| 657 | |
| 658 | ///////// Block iteration |
| 659 | template <bool Forward, bool Const> |
| 660 | class GrBlockAllocator::BlockIter { |
| 661 | private: |
| 662 | using BlockT = typename std::conditional<Const, const Block, Block>::type; |
| 663 | using AllocatorT = |
| 664 | typename std::conditional<Const, const GrBlockAllocator, GrBlockAllocator>::type; |
| 665 | |
| 666 | public: |
| 667 | BlockIter(AllocatorT* allocator) : fAllocator(allocator) {} |
| 668 | |
| 669 | class Item { |
| 670 | public: |
| 671 | bool operator!=(const Item& other) const { return fBlock != other.fBlock; } |
| 672 | |
| 673 | BlockT* operator*() const { return fBlock; } |
| 674 | |
| 675 | Item& operator++() { |
| 676 | this->advance(fNext); |
| 677 | return *this; |
| 678 | } |
| 679 | |
| 680 | private: |
| 681 | friend BlockIter; |
| 682 | |
| 683 | Item(BlockT* block) { this->advance(block); } |
| 684 | |
| 685 | void advance(BlockT* block) { |
| 686 | fBlock = block; |
| 687 | fNext = block ? (Forward ? block->fNext : block->fPrev) : nullptr; |
| 688 | if (!Forward && fNext && fNext->isScratch()) { |
| 689 | // For reverse-iteration only, we need to stop at the head, not the scratch block |
| 690 | // possibly stashed in head->prev. |
| 691 | fNext = nullptr; |
| 692 | } |
| 693 | SkASSERT(!fNext || !fNext->isScratch()); |
| 694 | } |
| 695 | |
| 696 | BlockT* fBlock; |
| 697 | // Cache this before operator++ so that fBlock can be released during iteration |
| 698 | BlockT* fNext; |
| 699 | }; |
| 700 | |
| 701 | Item begin() const { return Item(Forward ? &fAllocator->fHead : fAllocator->fTail); } |
| 702 | Item end() const { return Item(nullptr); } |
| 703 | |
| 704 | private: |
| 705 | AllocatorT* fAllocator; |
| 706 | }; |
| 707 | |
| 708 | GrBlockAllocator::BlockIter<true, false> GrBlockAllocator::blocks() { |
| 709 | return BlockIter<true, false>(this); |
| 710 | } |
| 711 | GrBlockAllocator::BlockIter<true, true> GrBlockAllocator::blocks() const { |
| 712 | return BlockIter<true, true>(this); |
| 713 | } |
| 714 | GrBlockAllocator::BlockIter<false, false> GrBlockAllocator::rblocks() { |
| 715 | return BlockIter<false, false>(this); |
| 716 | } |
| 717 | GrBlockAllocator::BlockIter<false, true> GrBlockAllocator::rblocks() const { |
| 718 | return BlockIter<false, true>(this); |
| 719 | } |
| 720 | |
| 721 | #endif // GrBlockAllocator_DEFINED |
| 722 | |