1 | //===- Allocator.h - Simple memory allocation abstraction -------*- C++ -*-===// |
2 | // |
3 | // The LLVM Compiler Infrastructure |
4 | // |
5 | // This file is distributed under the University of Illinois Open Source |
6 | // License. See LICENSE.TXT for details. |
7 | // |
8 | //===----------------------------------------------------------------------===// |
9 | /// \file |
10 | /// |
11 | /// This file defines the MallocAllocator and BumpPtrAllocator interfaces. Both |
12 | /// of these conform to an LLVM "Allocator" concept which consists of an |
13 | /// Allocate method accepting a size and alignment, and a Deallocate accepting |
14 | /// a pointer and size. Further, the LLVM "Allocator" concept has overloads of |
15 | /// Allocate and Deallocate for setting size and alignment based on the final |
16 | /// type. These overloads are typically provided by a base class template \c |
17 | /// AllocatorBase. |
18 | /// |
19 | //===----------------------------------------------------------------------===// |
20 | |
21 | #ifndef LLVM_SUPPORT_ALLOCATOR_H |
22 | #define LLVM_SUPPORT_ALLOCATOR_H |
23 | |
24 | #include "llvm/ADT/Optional.h" |
25 | #include "llvm/ADT/SmallVector.h" |
26 | #include "llvm/Support/Compiler.h" |
27 | #include "llvm/Support/ErrorHandling.h" |
28 | #include "llvm/Support/MathExtras.h" |
29 | #include "llvm/Support/MemAlloc.h" |
30 | #include <algorithm> |
31 | #include <cassert> |
32 | #include <cstddef> |
33 | #include <cstdint> |
34 | #include <cstdlib> |
35 | #include <iterator> |
36 | #include <type_traits> |
37 | #include <utility> |
38 | |
39 | namespace llvm { |
40 | |
41 | /// CRTP base class providing obvious overloads for the core \c |
42 | /// Allocate() methods of LLVM-style allocators. |
43 | /// |
44 | /// This base class both documents the full public interface exposed by all |
45 | /// LLVM-style allocators, and redirects all of the overloads to a single core |
46 | /// set of methods which the derived class must define. |
47 | template <typename DerivedT> class AllocatorBase { |
48 | public: |
49 | /// Allocate \a Size bytes of \a Alignment aligned memory. This method |
50 | /// must be implemented by \c DerivedT. |
51 | void *Allocate(size_t Size, size_t Alignment) { |
52 | #ifdef __clang__ |
53 | static_assert(static_cast<void *(AllocatorBase::*)(size_t, size_t)>( |
54 | &AllocatorBase::Allocate) != |
55 | static_cast<void *(DerivedT::*)(size_t, size_t)>( |
56 | &DerivedT::Allocate), |
57 | "Class derives from AllocatorBase without implementing the " |
58 | "core Allocate(size_t, size_t) overload!" ); |
59 | #endif |
60 | return static_cast<DerivedT *>(this)->Allocate(Size, Alignment); |
61 | } |
62 | |
63 | /// Deallocate \a Ptr to \a Size bytes of memory allocated by this |
64 | /// allocator. |
65 | void Deallocate(const void *Ptr, size_t Size) { |
66 | #ifdef __clang__ |
67 | static_assert(static_cast<void (AllocatorBase::*)(const void *, size_t)>( |
68 | &AllocatorBase::Deallocate) != |
69 | static_cast<void (DerivedT::*)(const void *, size_t)>( |
70 | &DerivedT::Deallocate), |
71 | "Class derives from AllocatorBase without implementing the " |
72 | "core Deallocate(void *) overload!" ); |
73 | #endif |
74 | return static_cast<DerivedT *>(this)->Deallocate(Ptr, Size); |
75 | } |
76 | |
77 | // The rest of these methods are helpers that redirect to one of the above |
78 | // core methods. |
79 | |
80 | /// Allocate space for a sequence of objects without constructing them. |
81 | template <typename T> T *Allocate(size_t Num = 1) { |
82 | return static_cast<T *>(Allocate(Num * sizeof(T), alignof(T))); |
83 | } |
84 | |
85 | /// Deallocate space for a sequence of objects without constructing them. |
86 | template <typename T> |
87 | typename std::enable_if< |
88 | !std::is_same<typename std::remove_cv<T>::type, void>::value, void>::type |
89 | Deallocate(T *Ptr, size_t Num = 1) { |
90 | Deallocate(static_cast<const void *>(Ptr), Num * sizeof(T)); |
91 | } |
92 | }; |
93 | |
94 | class MallocAllocator : public AllocatorBase<MallocAllocator> { |
95 | public: |
96 | void Reset() {} |
97 | |
98 | LLVM_ATTRIBUTE_RETURNS_NONNULL void *Allocate(size_t Size, |
99 | size_t /*Alignment*/) { |
100 | return safe_malloc(Size); |
101 | } |
102 | |
103 | // Pull in base class overloads. |
104 | using AllocatorBase<MallocAllocator>::Allocate; |
105 | |
106 | void Deallocate(const void *Ptr, size_t /*Size*/) { |
107 | free(const_cast<void *>(Ptr)); |
108 | } |
109 | |
110 | // Pull in base class overloads. |
111 | using AllocatorBase<MallocAllocator>::Deallocate; |
112 | |
113 | void PrintStats() const {} |
114 | }; |
115 | |
116 | namespace detail { |
117 | |
118 | // We call out to an external function to actually print the message as the |
119 | // printing code uses Allocator.h in its implementation. |
120 | void printBumpPtrAllocatorStats(unsigned NumSlabs, size_t BytesAllocated, |
121 | size_t TotalMemory); |
122 | |
123 | } // end namespace detail |
124 | |
125 | /// Allocate memory in an ever growing pool, as if by bump-pointer. |
126 | /// |
127 | /// This isn't strictly a bump-pointer allocator as it uses backing slabs of |
128 | /// memory rather than relying on a boundless contiguous heap. However, it has |
129 | /// bump-pointer semantics in that it is a monotonically growing pool of memory |
130 | /// where every allocation is found by merely allocating the next N bytes in |
131 | /// the slab, or the next N bytes in the next slab. |
132 | /// |
133 | /// Note that this also has a threshold for forcing allocations above a certain |
134 | /// size into their own slab. |
135 | /// |
136 | /// The BumpPtrAllocatorImpl template defaults to using a MallocAllocator |
137 | /// object, which wraps malloc, to allocate memory, but it can be changed to |
138 | /// use a custom allocator. |
139 | template <typename AllocatorT = MallocAllocator, size_t SlabSize = 4096, |
140 | size_t SizeThreshold = SlabSize> |
141 | class BumpPtrAllocatorImpl |
142 | : public AllocatorBase< |
143 | BumpPtrAllocatorImpl<AllocatorT, SlabSize, SizeThreshold>> { |
144 | public: |
145 | static_assert(SizeThreshold <= SlabSize, |
146 | "The SizeThreshold must be at most the SlabSize to ensure " |
147 | "that objects larger than a slab go into their own memory " |
148 | "allocation." ); |
149 | |
150 | BumpPtrAllocatorImpl() = default; |
151 | |
152 | template <typename T> |
153 | BumpPtrAllocatorImpl(T &&Allocator) |
154 | : Allocator(std::forward<T &&>(Allocator)) {} |
155 | |
156 | // Manually implement a move constructor as we must clear the old allocator's |
157 | // slabs as a matter of correctness. |
158 | BumpPtrAllocatorImpl(BumpPtrAllocatorImpl &&Old) |
159 | : CurPtr(Old.CurPtr), End(Old.End), Slabs(std::move(Old.Slabs)), |
160 | CustomSizedSlabs(std::move(Old.CustomSizedSlabs)), |
161 | BytesAllocated(Old.BytesAllocated), RedZoneSize(Old.RedZoneSize), |
162 | Allocator(std::move(Old.Allocator)) { |
163 | Old.CurPtr = Old.End = nullptr; |
164 | Old.BytesAllocated = 0; |
165 | Old.Slabs.clear(); |
166 | Old.CustomSizedSlabs.clear(); |
167 | } |
168 | |
169 | ~BumpPtrAllocatorImpl() { |
170 | DeallocateSlabs(Slabs.begin(), Slabs.end()); |
171 | DeallocateCustomSizedSlabs(); |
172 | } |
173 | |
174 | BumpPtrAllocatorImpl &operator=(BumpPtrAllocatorImpl &&RHS) { |
175 | DeallocateSlabs(Slabs.begin(), Slabs.end()); |
176 | DeallocateCustomSizedSlabs(); |
177 | |
178 | CurPtr = RHS.CurPtr; |
179 | End = RHS.End; |
180 | BytesAllocated = RHS.BytesAllocated; |
181 | RedZoneSize = RHS.RedZoneSize; |
182 | Slabs = std::move(RHS.Slabs); |
183 | CustomSizedSlabs = std::move(RHS.CustomSizedSlabs); |
184 | Allocator = std::move(RHS.Allocator); |
185 | |
186 | RHS.CurPtr = RHS.End = nullptr; |
187 | RHS.BytesAllocated = 0; |
188 | RHS.Slabs.clear(); |
189 | RHS.CustomSizedSlabs.clear(); |
190 | return *this; |
191 | } |
192 | |
193 | /// Deallocate all but the current slab and reset the current pointer |
194 | /// to the beginning of it, freeing all memory allocated so far. |
195 | void Reset() { |
196 | // Deallocate all but the first slab, and deallocate all custom-sized slabs. |
197 | DeallocateCustomSizedSlabs(); |
198 | CustomSizedSlabs.clear(); |
199 | |
200 | if (Slabs.empty()) |
201 | return; |
202 | |
203 | // Reset the state. |
204 | BytesAllocated = 0; |
205 | CurPtr = (char *)Slabs.front(); |
206 | End = CurPtr + SlabSize; |
207 | |
208 | __asan_poison_memory_region(*Slabs.begin(), computeSlabSize(0)); |
209 | DeallocateSlabs(std::next(Slabs.begin()), Slabs.end()); |
210 | Slabs.erase(std::next(Slabs.begin()), Slabs.end()); |
211 | } |
212 | |
213 | /// Allocate space at the specified alignment. |
214 | LLVM_ATTRIBUTE_RETURNS_NONNULL LLVM_ATTRIBUTE_RETURNS_NOALIAS void * |
215 | Allocate(size_t Size, size_t Alignment) { |
216 | assert(Alignment > 0 && "0-byte alignnment is not allowed. Use 1 instead." ); |
217 | |
218 | // Keep track of how many bytes we've allocated. |
219 | BytesAllocated += Size; |
220 | |
221 | size_t Adjustment = alignmentAdjustment(CurPtr, Alignment); |
222 | assert(Adjustment + Size >= Size && "Adjustment + Size must not overflow" ); |
223 | |
224 | size_t SizeToAllocate = Size; |
225 | #if LLVM_ADDRESS_SANITIZER_BUILD |
226 | // Add trailing bytes as a "red zone" under ASan. |
227 | SizeToAllocate += RedZoneSize; |
228 | #endif |
229 | |
230 | // Check if we have enough space. |
231 | if (Adjustment + SizeToAllocate <= size_t(End - CurPtr)) { |
232 | char *AlignedPtr = CurPtr + Adjustment; |
233 | CurPtr = AlignedPtr + SizeToAllocate; |
234 | // Update the allocation point of this memory block in MemorySanitizer. |
235 | // Without this, MemorySanitizer messages for values originated from here |
236 | // will point to the allocation of the entire slab. |
237 | __msan_allocated_memory(AlignedPtr, Size); |
238 | // Similarly, tell ASan about this space. |
239 | __asan_unpoison_memory_region(AlignedPtr, Size); |
240 | return AlignedPtr; |
241 | } |
242 | |
243 | // If Size is really big, allocate a separate slab for it. |
244 | size_t PaddedSize = SizeToAllocate + Alignment - 1; |
245 | if (PaddedSize > SizeThreshold) { |
246 | void *NewSlab = Allocator.Allocate(PaddedSize, 0); |
247 | // We own the new slab and don't want anyone reading anyting other than |
248 | // pieces returned from this method. So poison the whole slab. |
249 | __asan_poison_memory_region(NewSlab, PaddedSize); |
250 | CustomSizedSlabs.push_back(std::make_pair(NewSlab, PaddedSize)); |
251 | |
252 | uintptr_t AlignedAddr = alignAddr(NewSlab, Alignment); |
253 | assert(AlignedAddr + Size <= (uintptr_t)NewSlab + PaddedSize); |
254 | char *AlignedPtr = (char*)AlignedAddr; |
255 | __msan_allocated_memory(AlignedPtr, Size); |
256 | __asan_unpoison_memory_region(AlignedPtr, Size); |
257 | return AlignedPtr; |
258 | } |
259 | |
260 | // Otherwise, start a new slab and try again. |
261 | StartNewSlab(); |
262 | uintptr_t AlignedAddr = alignAddr(CurPtr, Alignment); |
263 | assert(AlignedAddr + SizeToAllocate <= (uintptr_t)End && |
264 | "Unable to allocate memory!" ); |
265 | char *AlignedPtr = (char*)AlignedAddr; |
266 | CurPtr = AlignedPtr + SizeToAllocate; |
267 | __msan_allocated_memory(AlignedPtr, Size); |
268 | __asan_unpoison_memory_region(AlignedPtr, Size); |
269 | return AlignedPtr; |
270 | } |
271 | |
272 | // Pull in base class overloads. |
273 | using AllocatorBase<BumpPtrAllocatorImpl>::Allocate; |
274 | |
275 | // Bump pointer allocators are expected to never free their storage; and |
276 | // clients expect pointers to remain valid for non-dereferencing uses even |
277 | // after deallocation. |
278 | void Deallocate(const void *Ptr, size_t Size) { |
279 | __asan_poison_memory_region(Ptr, Size); |
280 | } |
281 | |
282 | // Pull in base class overloads. |
283 | using AllocatorBase<BumpPtrAllocatorImpl>::Deallocate; |
284 | |
285 | size_t GetNumSlabs() const { return Slabs.size() + CustomSizedSlabs.size(); } |
286 | |
287 | /// \return An index uniquely and reproducibly identifying |
288 | /// an input pointer \p Ptr in the given allocator. |
289 | /// The returned value is negative iff the object is inside a custom-size |
290 | /// slab. |
291 | /// Returns an empty optional if the pointer is not found in the allocator. |
292 | llvm::Optional<int64_t> identifyObject(const void *Ptr) { |
293 | const char *P = static_cast<const char *>(Ptr); |
294 | int64_t InSlabIdx = 0; |
295 | for (size_t Idx = 0, E = Slabs.size(); Idx < E; Idx++) { |
296 | const char *S = static_cast<const char *>(Slabs[Idx]); |
297 | if (P >= S && P < S + computeSlabSize(Idx)) |
298 | return InSlabIdx + static_cast<int64_t>(P - S); |
299 | InSlabIdx += static_cast<int64_t>(computeSlabSize(Idx)); |
300 | } |
301 | |
302 | // Use negative index to denote custom sized slabs. |
303 | int64_t InCustomSizedSlabIdx = -1; |
304 | for (size_t Idx = 0, E = CustomSizedSlabs.size(); Idx < E; Idx++) { |
305 | const char *S = static_cast<const char *>(CustomSizedSlabs[Idx].first); |
306 | size_t Size = CustomSizedSlabs[Idx].second; |
307 | if (P >= S && P < S + Size) |
308 | return InCustomSizedSlabIdx - static_cast<int64_t>(P - S); |
309 | InCustomSizedSlabIdx -= static_cast<int64_t>(Size); |
310 | } |
311 | return None; |
312 | } |
313 | |
314 | /// A wrapper around identifyObject that additionally asserts that |
315 | /// the object is indeed within the allocator. |
316 | /// \return An index uniquely and reproducibly identifying |
317 | /// an input pointer \p Ptr in the given allocator. |
318 | int64_t identifyKnownObject(const void *Ptr) { |
319 | Optional<int64_t> Out = identifyObject(Ptr); |
320 | assert(Out && "Wrong allocator used" ); |
321 | return *Out; |
322 | } |
323 | |
324 | /// A wrapper around identifyKnownObject. Accepts type information |
325 | /// about the object and produces a smaller identifier by relying on |
326 | /// the alignment information. Note that sub-classes may have different |
327 | /// alignment, so the most base class should be passed as template parameter |
328 | /// in order to obtain correct results. For that reason automatic template |
329 | /// parameter deduction is disabled. |
330 | /// \return An index uniquely and reproducibly identifying |
331 | /// an input pointer \p Ptr in the given allocator. This identifier is |
332 | /// different from the ones produced by identifyObject and |
333 | /// identifyAlignedObject. |
334 | template <typename T> |
335 | int64_t identifyKnownAlignedObject(const void *Ptr) { |
336 | int64_t Out = identifyKnownObject(Ptr); |
337 | assert(Out % alignof(T) == 0 && "Wrong alignment information" ); |
338 | return Out / alignof(T); |
339 | } |
340 | |
341 | size_t getTotalMemory() const { |
342 | size_t TotalMemory = 0; |
343 | for (auto I = Slabs.begin(), E = Slabs.end(); I != E; ++I) |
344 | TotalMemory += computeSlabSize(std::distance(Slabs.begin(), I)); |
345 | for (auto &PtrAndSize : CustomSizedSlabs) |
346 | TotalMemory += PtrAndSize.second; |
347 | return TotalMemory; |
348 | } |
349 | |
350 | size_t getBytesAllocated() const { return BytesAllocated; } |
351 | |
352 | void setRedZoneSize(size_t NewSize) { |
353 | RedZoneSize = NewSize; |
354 | } |
355 | |
356 | void PrintStats() const { |
357 | detail::printBumpPtrAllocatorStats(Slabs.size(), BytesAllocated, |
358 | getTotalMemory()); |
359 | } |
360 | |
361 | private: |
362 | /// The current pointer into the current slab. |
363 | /// |
364 | /// This points to the next free byte in the slab. |
365 | char *CurPtr = nullptr; |
366 | |
367 | /// The end of the current slab. |
368 | char *End = nullptr; |
369 | |
370 | /// The slabs allocated so far. |
371 | SmallVector<void *, 4> Slabs; |
372 | |
373 | /// Custom-sized slabs allocated for too-large allocation requests. |
374 | SmallVector<std::pair<void *, size_t>, 0> CustomSizedSlabs; |
375 | |
376 | /// How many bytes we've allocated. |
377 | /// |
378 | /// Used so that we can compute how much space was wasted. |
379 | size_t BytesAllocated = 0; |
380 | |
381 | /// The number of bytes to put between allocations when running under |
382 | /// a sanitizer. |
383 | size_t RedZoneSize = 1; |
384 | |
385 | /// The allocator instance we use to get slabs of memory. |
386 | AllocatorT Allocator; |
387 | |
388 | static size_t computeSlabSize(unsigned SlabIdx) { |
389 | // Scale the actual allocated slab size based on the number of slabs |
390 | // allocated. Every 128 slabs allocated, we double the allocated size to |
391 | // reduce allocation frequency, but saturate at multiplying the slab size by |
392 | // 2^30. |
393 | return SlabSize * ((size_t)1 << std::min<size_t>(30, SlabIdx / 128)); |
394 | } |
395 | |
396 | /// Allocate a new slab and move the bump pointers over into the new |
397 | /// slab, modifying CurPtr and End. |
398 | void StartNewSlab() { |
399 | size_t AllocatedSlabSize = computeSlabSize(Slabs.size()); |
400 | |
401 | void *NewSlab = Allocator.Allocate(AllocatedSlabSize, 0); |
402 | // We own the new slab and don't want anyone reading anything other than |
403 | // pieces returned from this method. So poison the whole slab. |
404 | __asan_poison_memory_region(NewSlab, AllocatedSlabSize); |
405 | |
406 | Slabs.push_back(NewSlab); |
407 | CurPtr = (char *)(NewSlab); |
408 | End = ((char *)NewSlab) + AllocatedSlabSize; |
409 | } |
410 | |
411 | /// Deallocate a sequence of slabs. |
412 | void DeallocateSlabs(SmallVectorImpl<void *>::iterator I, |
413 | SmallVectorImpl<void *>::iterator E) { |
414 | for (; I != E; ++I) { |
415 | size_t AllocatedSlabSize = |
416 | computeSlabSize(std::distance(Slabs.begin(), I)); |
417 | Allocator.Deallocate(*I, AllocatedSlabSize); |
418 | } |
419 | } |
420 | |
421 | /// Deallocate all memory for custom sized slabs. |
422 | void DeallocateCustomSizedSlabs() { |
423 | for (auto &PtrAndSize : CustomSizedSlabs) { |
424 | void *Ptr = PtrAndSize.first; |
425 | size_t Size = PtrAndSize.second; |
426 | Allocator.Deallocate(Ptr, Size); |
427 | } |
428 | } |
429 | |
430 | template <typename T> friend class SpecificBumpPtrAllocator; |
431 | }; |
432 | |
433 | /// The standard BumpPtrAllocator which just uses the default template |
434 | /// parameters. |
435 | typedef BumpPtrAllocatorImpl<> BumpPtrAllocator; |
436 | |
437 | /// A BumpPtrAllocator that allows only elements of a specific type to be |
438 | /// allocated. |
439 | /// |
440 | /// This allows calling the destructor in DestroyAll() and when the allocator is |
441 | /// destroyed. |
442 | template <typename T> class SpecificBumpPtrAllocator { |
443 | BumpPtrAllocator Allocator; |
444 | |
445 | public: |
446 | SpecificBumpPtrAllocator() { |
447 | // Because SpecificBumpPtrAllocator walks the memory to call destructors, |
448 | // it can't have red zones between allocations. |
449 | Allocator.setRedZoneSize(0); |
450 | } |
451 | SpecificBumpPtrAllocator(SpecificBumpPtrAllocator &&Old) |
452 | : Allocator(std::move(Old.Allocator)) {} |
453 | ~SpecificBumpPtrAllocator() { DestroyAll(); } |
454 | |
455 | SpecificBumpPtrAllocator &operator=(SpecificBumpPtrAllocator &&RHS) { |
456 | Allocator = std::move(RHS.Allocator); |
457 | return *this; |
458 | } |
459 | |
460 | /// Call the destructor of each allocated object and deallocate all but the |
461 | /// current slab and reset the current pointer to the beginning of it, freeing |
462 | /// all memory allocated so far. |
463 | void DestroyAll() { |
464 | auto DestroyElements = [](char *Begin, char *End) { |
465 | assert(Begin == (char *)alignAddr(Begin, alignof(T))); |
466 | for (char *Ptr = Begin; Ptr + sizeof(T) <= End; Ptr += sizeof(T)) |
467 | reinterpret_cast<T *>(Ptr)->~T(); |
468 | }; |
469 | |
470 | for (auto I = Allocator.Slabs.begin(), E = Allocator.Slabs.end(); I != E; |
471 | ++I) { |
472 | size_t AllocatedSlabSize = BumpPtrAllocator::computeSlabSize( |
473 | std::distance(Allocator.Slabs.begin(), I)); |
474 | char *Begin = (char *)alignAddr(*I, alignof(T)); |
475 | char *End = *I == Allocator.Slabs.back() ? Allocator.CurPtr |
476 | : (char *)*I + AllocatedSlabSize; |
477 | |
478 | DestroyElements(Begin, End); |
479 | } |
480 | |
481 | for (auto &PtrAndSize : Allocator.CustomSizedSlabs) { |
482 | void *Ptr = PtrAndSize.first; |
483 | size_t Size = PtrAndSize.second; |
484 | DestroyElements((char *)alignAddr(Ptr, alignof(T)), (char *)Ptr + Size); |
485 | } |
486 | |
487 | Allocator.Reset(); |
488 | } |
489 | |
490 | /// Allocate space for an array of objects without constructing them. |
491 | T *Allocate(size_t num = 1) { return Allocator.Allocate<T>(num); } |
492 | }; |
493 | |
494 | } // end namespace llvm |
495 | |
496 | template <typename AllocatorT, size_t SlabSize, size_t SizeThreshold> |
497 | void *operator new(size_t Size, |
498 | llvm::BumpPtrAllocatorImpl<AllocatorT, SlabSize, |
499 | SizeThreshold> &Allocator) { |
500 | struct S { |
501 | char c; |
502 | union { |
503 | double D; |
504 | long double LD; |
505 | long long L; |
506 | void *P; |
507 | } x; |
508 | }; |
509 | return Allocator.Allocate( |
510 | Size, std::min((size_t)llvm::NextPowerOf2(Size), offsetof(S, x))); |
511 | } |
512 | |
513 | template <typename AllocatorT, size_t SlabSize, size_t SizeThreshold> |
514 | void operator delete( |
515 | void *, llvm::BumpPtrAllocatorImpl<AllocatorT, SlabSize, SizeThreshold> &) { |
516 | } |
517 | |
518 | #endif // LLVM_SUPPORT_ALLOCATOR_H |
519 | |