1 | //===- Allocator.h - Simple memory allocation abstraction -------*- C++ -*-===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | /// \file |
9 | /// |
10 | /// This file defines the BumpPtrAllocator interface. BumpPtrAllocator conforms |
11 | /// to the LLVM "Allocator" concept and is similar to MallocAllocator, but |
12 | /// objects cannot be deallocated. Their lifetime is tied to the lifetime of the |
13 | /// allocator. |
14 | /// |
15 | //===----------------------------------------------------------------------===// |
16 | |
17 | #ifndef LLVM_SUPPORT_ALLOCATOR_H |
18 | #define LLVM_SUPPORT_ALLOCATOR_H |
19 | |
20 | #include "llvm/ADT/SmallVector.h" |
21 | #include "llvm/Support/Alignment.h" |
22 | #include "llvm/Support/AllocatorBase.h" |
23 | #include "llvm/Support/Compiler.h" |
24 | #include "llvm/Support/MathExtras.h" |
25 | #include <algorithm> |
26 | #include <cassert> |
27 | #include <cstddef> |
28 | #include <cstdint> |
29 | #include <iterator> |
30 | #include <optional> |
31 | #include <utility> |
32 | |
33 | namespace llvm { |
34 | |
35 | namespace detail { |
36 | |
37 | // We call out to an external function to actually print the message as the |
38 | // printing code uses Allocator.h in its implementation. |
39 | void printBumpPtrAllocatorStats(unsigned NumSlabs, size_t BytesAllocated, |
40 | size_t TotalMemory); |
41 | |
42 | } // end namespace detail |
43 | |
44 | /// Allocate memory in an ever growing pool, as if by bump-pointer. |
45 | /// |
46 | /// This isn't strictly a bump-pointer allocator as it uses backing slabs of |
47 | /// memory rather than relying on a boundless contiguous heap. However, it has |
48 | /// bump-pointer semantics in that it is a monotonically growing pool of memory |
49 | /// where every allocation is found by merely allocating the next N bytes in |
50 | /// the slab, or the next N bytes in the next slab. |
51 | /// |
52 | /// Note that this also has a threshold for forcing allocations above a certain |
53 | /// size into their own slab. |
54 | /// |
55 | /// The BumpPtrAllocatorImpl template defaults to using a MallocAllocator |
56 | /// object, which wraps malloc, to allocate memory, but it can be changed to |
57 | /// use a custom allocator. |
58 | /// |
59 | /// The GrowthDelay specifies after how many allocated slabs the allocator |
60 | /// increases the size of the slabs. |
61 | template <typename AllocatorT = MallocAllocator, size_t SlabSize = 4096, |
62 | size_t SizeThreshold = SlabSize, size_t GrowthDelay = 128> |
63 | class BumpPtrAllocatorImpl |
64 | : public AllocatorBase<BumpPtrAllocatorImpl<AllocatorT, SlabSize, |
65 | SizeThreshold, GrowthDelay>>, |
66 | private detail::AllocatorHolder<AllocatorT> { |
67 | using AllocTy = detail::AllocatorHolder<AllocatorT>; |
68 | |
69 | public: |
70 | static_assert(SizeThreshold <= SlabSize, |
71 | "The SizeThreshold must be at most the SlabSize to ensure " |
72 | "that objects larger than a slab go into their own memory " |
73 | "allocation." ); |
74 | static_assert(GrowthDelay > 0, |
75 | "GrowthDelay must be at least 1 which already increases the" |
76 | "slab size after each allocated slab." ); |
77 | |
78 | BumpPtrAllocatorImpl() = default; |
79 | |
80 | template <typename T> |
81 | BumpPtrAllocatorImpl(T &&Allocator) |
82 | : AllocTy(std::forward<T &&>(Allocator)) {} |
83 | |
84 | // Manually implement a move constructor as we must clear the old allocator's |
85 | // slabs as a matter of correctness. |
86 | BumpPtrAllocatorImpl(BumpPtrAllocatorImpl &&Old) |
87 | : AllocTy(std::move(Old.getAllocator())), CurPtr(Old.CurPtr), |
88 | End(Old.End), Slabs(std::move(Old.Slabs)), |
89 | CustomSizedSlabs(std::move(Old.CustomSizedSlabs)), |
90 | BytesAllocated(Old.BytesAllocated), RedZoneSize(Old.RedZoneSize) { |
91 | Old.CurPtr = Old.End = nullptr; |
92 | Old.BytesAllocated = 0; |
93 | Old.Slabs.clear(); |
94 | Old.CustomSizedSlabs.clear(); |
95 | } |
96 | |
97 | ~BumpPtrAllocatorImpl() { |
98 | DeallocateSlabs(I: Slabs.begin(), E: Slabs.end()); |
99 | DeallocateCustomSizedSlabs(); |
100 | } |
101 | |
102 | BumpPtrAllocatorImpl &operator=(BumpPtrAllocatorImpl &&RHS) { |
103 | DeallocateSlabs(I: Slabs.begin(), E: Slabs.end()); |
104 | DeallocateCustomSizedSlabs(); |
105 | |
106 | CurPtr = RHS.CurPtr; |
107 | End = RHS.End; |
108 | BytesAllocated = RHS.BytesAllocated; |
109 | RedZoneSize = RHS.RedZoneSize; |
110 | Slabs = std::move(RHS.Slabs); |
111 | CustomSizedSlabs = std::move(RHS.CustomSizedSlabs); |
112 | AllocTy::operator=(std::move(RHS.getAllocator())); |
113 | |
114 | RHS.CurPtr = RHS.End = nullptr; |
115 | RHS.BytesAllocated = 0; |
116 | RHS.Slabs.clear(); |
117 | RHS.CustomSizedSlabs.clear(); |
118 | return *this; |
119 | } |
120 | |
121 | /// Deallocate all but the current slab and reset the current pointer |
122 | /// to the beginning of it, freeing all memory allocated so far. |
123 | void Reset() { |
124 | // Deallocate all but the first slab, and deallocate all custom-sized slabs. |
125 | DeallocateCustomSizedSlabs(); |
126 | CustomSizedSlabs.clear(); |
127 | |
128 | if (Slabs.empty()) |
129 | return; |
130 | |
131 | // Reset the state. |
132 | BytesAllocated = 0; |
133 | CurPtr = (char *)Slabs.front(); |
134 | End = CurPtr + SlabSize; |
135 | |
136 | __asan_poison_memory_region(*Slabs.begin(), computeSlabSize(0)); |
137 | DeallocateSlabs(I: std::next(x: Slabs.begin()), E: Slabs.end()); |
138 | Slabs.erase(CS: std::next(x: Slabs.begin()), CE: Slabs.end()); |
139 | } |
140 | |
141 | /// Allocate space at the specified alignment. |
142 | // This method is *not* marked noalias, because |
143 | // SpecificBumpPtrAllocator::DestroyAll() loops over all allocations, and |
144 | // that loop is not based on the Allocate() return value. |
145 | // |
146 | // Allocate(0, N) is valid, it returns a non-null pointer (which should not |
147 | // be dereferenced). |
148 | LLVM_ATTRIBUTE_RETURNS_NONNULL void *Allocate(size_t Size, Align Alignment) { |
149 | // Keep track of how many bytes we've allocated. |
150 | BytesAllocated += Size; |
151 | |
152 | size_t Adjustment = offsetToAlignedAddr(Addr: CurPtr, Alignment); |
153 | assert(Adjustment + Size >= Size && "Adjustment + Size must not overflow" ); |
154 | |
155 | size_t SizeToAllocate = Size; |
156 | #if LLVM_ADDRESS_SANITIZER_BUILD |
157 | // Add trailing bytes as a "red zone" under ASan. |
158 | SizeToAllocate += RedZoneSize; |
159 | #endif |
160 | |
161 | // Check if we have enough space. |
162 | if (Adjustment + SizeToAllocate <= size_t(End - CurPtr) |
163 | // We can't return nullptr even for a zero-sized allocation! |
164 | && CurPtr != nullptr) { |
165 | char *AlignedPtr = CurPtr + Adjustment; |
166 | CurPtr = AlignedPtr + SizeToAllocate; |
167 | // Update the allocation point of this memory block in MemorySanitizer. |
168 | // Without this, MemorySanitizer messages for values originated from here |
169 | // will point to the allocation of the entire slab. |
170 | __msan_allocated_memory(AlignedPtr, Size); |
171 | // Similarly, tell ASan about this space. |
172 | __asan_unpoison_memory_region(AlignedPtr, Size); |
173 | return AlignedPtr; |
174 | } |
175 | |
176 | // If Size is really big, allocate a separate slab for it. |
177 | size_t PaddedSize = SizeToAllocate + Alignment.value() - 1; |
178 | if (PaddedSize > SizeThreshold) { |
179 | void *NewSlab = |
180 | this->getAllocator().Allocate(PaddedSize, alignof(std::max_align_t)); |
181 | // We own the new slab and don't want anyone reading anyting other than |
182 | // pieces returned from this method. So poison the whole slab. |
183 | __asan_poison_memory_region(NewSlab, PaddedSize); |
184 | CustomSizedSlabs.push_back(Elt: std::make_pair(x&: NewSlab, y&: PaddedSize)); |
185 | |
186 | uintptr_t AlignedAddr = alignAddr(Addr: NewSlab, Alignment); |
187 | assert(AlignedAddr + Size <= (uintptr_t)NewSlab + PaddedSize); |
188 | char *AlignedPtr = (char*)AlignedAddr; |
189 | __msan_allocated_memory(AlignedPtr, Size); |
190 | __asan_unpoison_memory_region(AlignedPtr, Size); |
191 | return AlignedPtr; |
192 | } |
193 | |
194 | // Otherwise, start a new slab and try again. |
195 | StartNewSlab(); |
196 | uintptr_t AlignedAddr = alignAddr(Addr: CurPtr, Alignment); |
197 | assert(AlignedAddr + SizeToAllocate <= (uintptr_t)End && |
198 | "Unable to allocate memory!" ); |
199 | char *AlignedPtr = (char*)AlignedAddr; |
200 | CurPtr = AlignedPtr + SizeToAllocate; |
201 | __msan_allocated_memory(AlignedPtr, Size); |
202 | __asan_unpoison_memory_region(AlignedPtr, Size); |
203 | return AlignedPtr; |
204 | } |
205 | |
206 | inline LLVM_ATTRIBUTE_RETURNS_NONNULL void * |
207 | Allocate(size_t Size, size_t Alignment) { |
208 | assert(Alignment > 0 && "0-byte alignment is not allowed. Use 1 instead." ); |
209 | return Allocate(Size, Align(Alignment)); |
210 | } |
211 | |
212 | // Pull in base class overloads. |
213 | using AllocatorBase<BumpPtrAllocatorImpl>::Allocate; |
214 | |
215 | // Bump pointer allocators are expected to never free their storage; and |
216 | // clients expect pointers to remain valid for non-dereferencing uses even |
217 | // after deallocation. |
218 | void Deallocate(const void *Ptr, size_t Size, size_t /*Alignment*/) { |
219 | __asan_poison_memory_region(Ptr, Size); |
220 | } |
221 | |
222 | // Pull in base class overloads. |
223 | using AllocatorBase<BumpPtrAllocatorImpl>::Deallocate; |
224 | |
225 | size_t GetNumSlabs() const { return Slabs.size() + CustomSizedSlabs.size(); } |
226 | |
227 | /// \return An index uniquely and reproducibly identifying |
228 | /// an input pointer \p Ptr in the given allocator. |
229 | /// The returned value is negative iff the object is inside a custom-size |
230 | /// slab. |
231 | /// Returns an empty optional if the pointer is not found in the allocator. |
232 | std::optional<int64_t> identifyObject(const void *Ptr) { |
233 | const char *P = static_cast<const char *>(Ptr); |
234 | int64_t InSlabIdx = 0; |
235 | for (size_t Idx = 0, E = Slabs.size(); Idx < E; Idx++) { |
236 | const char *S = static_cast<const char *>(Slabs[Idx]); |
237 | if (P >= S && P < S + computeSlabSize(SlabIdx: Idx)) |
238 | return InSlabIdx + static_cast<int64_t>(P - S); |
239 | InSlabIdx += static_cast<int64_t>(computeSlabSize(SlabIdx: Idx)); |
240 | } |
241 | |
242 | // Use negative index to denote custom sized slabs. |
243 | int64_t InCustomSizedSlabIdx = -1; |
244 | for (size_t Idx = 0, E = CustomSizedSlabs.size(); Idx < E; Idx++) { |
245 | const char *S = static_cast<const char *>(CustomSizedSlabs[Idx].first); |
246 | size_t Size = CustomSizedSlabs[Idx].second; |
247 | if (P >= S && P < S + Size) |
248 | return InCustomSizedSlabIdx - static_cast<int64_t>(P - S); |
249 | InCustomSizedSlabIdx -= static_cast<int64_t>(Size); |
250 | } |
251 | return std::nullopt; |
252 | } |
253 | |
254 | /// A wrapper around identifyObject that additionally asserts that |
255 | /// the object is indeed within the allocator. |
256 | /// \return An index uniquely and reproducibly identifying |
257 | /// an input pointer \p Ptr in the given allocator. |
258 | int64_t identifyKnownObject(const void *Ptr) { |
259 | std::optional<int64_t> Out = identifyObject(Ptr); |
260 | assert(Out && "Wrong allocator used" ); |
261 | return *Out; |
262 | } |
263 | |
264 | /// A wrapper around identifyKnownObject. Accepts type information |
265 | /// about the object and produces a smaller identifier by relying on |
266 | /// the alignment information. Note that sub-classes may have different |
267 | /// alignment, so the most base class should be passed as template parameter |
268 | /// in order to obtain correct results. For that reason automatic template |
269 | /// parameter deduction is disabled. |
270 | /// \return An index uniquely and reproducibly identifying |
271 | /// an input pointer \p Ptr in the given allocator. This identifier is |
272 | /// different from the ones produced by identifyObject and |
273 | /// identifyAlignedObject. |
274 | template <typename T> |
275 | int64_t identifyKnownAlignedObject(const void *Ptr) { |
276 | int64_t Out = identifyKnownObject(Ptr); |
277 | assert(Out % alignof(T) == 0 && "Wrong alignment information" ); |
278 | return Out / alignof(T); |
279 | } |
280 | |
281 | size_t getTotalMemory() const { |
282 | size_t TotalMemory = 0; |
283 | for (auto I = Slabs.begin(), E = Slabs.end(); I != E; ++I) |
284 | TotalMemory += computeSlabSize(SlabIdx: std::distance(first: Slabs.begin(), last: I)); |
285 | for (const auto &PtrAndSize : CustomSizedSlabs) |
286 | TotalMemory += PtrAndSize.second; |
287 | return TotalMemory; |
288 | } |
289 | |
290 | size_t getBytesAllocated() const { return BytesAllocated; } |
291 | |
292 | void setRedZoneSize(size_t NewSize) { |
293 | RedZoneSize = NewSize; |
294 | } |
295 | |
296 | void PrintStats() const { |
297 | detail::printBumpPtrAllocatorStats(NumSlabs: Slabs.size(), BytesAllocated, |
298 | TotalMemory: getTotalMemory()); |
299 | } |
300 | |
301 | private: |
302 | /// The current pointer into the current slab. |
303 | /// |
304 | /// This points to the next free byte in the slab. |
305 | char *CurPtr = nullptr; |
306 | |
307 | /// The end of the current slab. |
308 | char *End = nullptr; |
309 | |
310 | /// The slabs allocated so far. |
311 | SmallVector<void *, 4> Slabs; |
312 | |
313 | /// Custom-sized slabs allocated for too-large allocation requests. |
314 | SmallVector<std::pair<void *, size_t>, 0> CustomSizedSlabs; |
315 | |
316 | /// How many bytes we've allocated. |
317 | /// |
318 | /// Used so that we can compute how much space was wasted. |
319 | size_t BytesAllocated = 0; |
320 | |
321 | /// The number of bytes to put between allocations when running under |
322 | /// a sanitizer. |
323 | size_t RedZoneSize = 1; |
324 | |
325 | static size_t computeSlabSize(unsigned SlabIdx) { |
326 | // Scale the actual allocated slab size based on the number of slabs |
327 | // allocated. Every GrowthDelay slabs allocated, we double |
328 | // the allocated size to reduce allocation frequency, but saturate at |
329 | // multiplying the slab size by 2^30. |
330 | return SlabSize * |
331 | ((size_t)1 << std::min<size_t>(a: 30, b: SlabIdx / GrowthDelay)); |
332 | } |
333 | |
334 | /// Allocate a new slab and move the bump pointers over into the new |
335 | /// slab, modifying CurPtr and End. |
336 | void StartNewSlab() { |
337 | size_t AllocatedSlabSize = computeSlabSize(SlabIdx: Slabs.size()); |
338 | |
339 | void *NewSlab = this->getAllocator().Allocate(AllocatedSlabSize, |
340 | alignof(std::max_align_t)); |
341 | // We own the new slab and don't want anyone reading anything other than |
342 | // pieces returned from this method. So poison the whole slab. |
343 | __asan_poison_memory_region(NewSlab, AllocatedSlabSize); |
344 | |
345 | Slabs.push_back(Elt: NewSlab); |
346 | CurPtr = (char *)(NewSlab); |
347 | End = ((char *)NewSlab) + AllocatedSlabSize; |
348 | } |
349 | |
350 | /// Deallocate a sequence of slabs. |
351 | void DeallocateSlabs(SmallVectorImpl<void *>::iterator I, |
352 | SmallVectorImpl<void *>::iterator E) { |
353 | for (; I != E; ++I) { |
354 | size_t AllocatedSlabSize = |
355 | computeSlabSize(SlabIdx: std::distance(first: Slabs.begin(), last: I)); |
356 | this->getAllocator().Deallocate(*I, AllocatedSlabSize, |
357 | alignof(std::max_align_t)); |
358 | } |
359 | } |
360 | |
361 | /// Deallocate all memory for custom sized slabs. |
362 | void DeallocateCustomSizedSlabs() { |
363 | for (auto &PtrAndSize : CustomSizedSlabs) { |
364 | void *Ptr = PtrAndSize.first; |
365 | size_t Size = PtrAndSize.second; |
366 | this->getAllocator().Deallocate(Ptr, Size, alignof(std::max_align_t)); |
367 | } |
368 | } |
369 | |
370 | template <typename T> friend class SpecificBumpPtrAllocator; |
371 | }; |
372 | |
373 | /// The standard BumpPtrAllocator which just uses the default template |
374 | /// parameters. |
375 | typedef BumpPtrAllocatorImpl<> BumpPtrAllocator; |
376 | |
377 | /// A BumpPtrAllocator that allows only elements of a specific type to be |
378 | /// allocated. |
379 | /// |
380 | /// This allows calling the destructor in DestroyAll() and when the allocator is |
381 | /// destroyed. |
382 | template <typename T> class SpecificBumpPtrAllocator { |
383 | BumpPtrAllocator Allocator; |
384 | |
385 | public: |
386 | SpecificBumpPtrAllocator() { |
387 | // Because SpecificBumpPtrAllocator walks the memory to call destructors, |
388 | // it can't have red zones between allocations. |
389 | Allocator.setRedZoneSize(0); |
390 | } |
391 | SpecificBumpPtrAllocator(SpecificBumpPtrAllocator &&Old) |
392 | : Allocator(std::move(Old.Allocator)) {} |
393 | ~SpecificBumpPtrAllocator() { DestroyAll(); } |
394 | |
395 | SpecificBumpPtrAllocator &operator=(SpecificBumpPtrAllocator &&RHS) { |
396 | Allocator = std::move(RHS.Allocator); |
397 | return *this; |
398 | } |
399 | |
400 | /// Call the destructor of each allocated object and deallocate all but the |
401 | /// current slab and reset the current pointer to the beginning of it, freeing |
402 | /// all memory allocated so far. |
403 | void DestroyAll() { |
404 | auto DestroyElements = [](char *Begin, char *End) { |
405 | assert(Begin == (char *)alignAddr(Begin, Align::Of<T>())); |
406 | for (char *Ptr = Begin; Ptr + sizeof(T) <= End; Ptr += sizeof(T)) |
407 | reinterpret_cast<T *>(Ptr)->~T(); |
408 | }; |
409 | |
410 | for (auto I = Allocator.Slabs.begin(), E = Allocator.Slabs.end(); I != E; |
411 | ++I) { |
412 | size_t AllocatedSlabSize = BumpPtrAllocator::computeSlabSize( |
413 | SlabIdx: std::distance(first: Allocator.Slabs.begin(), last: I)); |
414 | char *Begin = (char *)alignAddr(*I, Align::Of<T>()); |
415 | char *End = *I == Allocator.Slabs.back() ? Allocator.CurPtr |
416 | : (char *)*I + AllocatedSlabSize; |
417 | |
418 | DestroyElements(Begin, End); |
419 | } |
420 | |
421 | for (auto &PtrAndSize : Allocator.CustomSizedSlabs) { |
422 | void *Ptr = PtrAndSize.first; |
423 | size_t Size = PtrAndSize.second; |
424 | DestroyElements((char *)alignAddr(Ptr, Align::Of<T>()), |
425 | (char *)Ptr + Size); |
426 | } |
427 | |
428 | Allocator.Reset(); |
429 | } |
430 | |
431 | /// Allocate space for an array of objects without constructing them. |
432 | T *Allocate(size_t num = 1) { return Allocator.Allocate<T>(num); } |
433 | }; |
434 | |
435 | } // end namespace llvm |
436 | |
437 | template <typename AllocatorT, size_t SlabSize, size_t SizeThreshold, |
438 | size_t GrowthDelay> |
439 | void * |
440 | operator new(size_t Size, |
441 | llvm::BumpPtrAllocatorImpl<AllocatorT, SlabSize, SizeThreshold, |
442 | GrowthDelay> &Allocator) { |
443 | return Allocator.Allocate(Size, std::min(a: (size_t)llvm::NextPowerOf2(A: Size), |
444 | b: alignof(std::max_align_t))); |
445 | } |
446 | |
447 | template <typename AllocatorT, size_t SlabSize, size_t SizeThreshold, |
448 | size_t GrowthDelay> |
449 | void operator delete(void *, |
450 | llvm::BumpPtrAllocatorImpl<AllocatorT, SlabSize, |
451 | SizeThreshold, GrowthDelay> &) { |
452 | } |
453 | |
454 | #endif // LLVM_SUPPORT_ALLOCATOR_H |
455 | |