1 | //===- AtomicExpandPass.cpp - Expand atomic instructions ------------------===// |
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 | // |
9 | // This file contains a pass (at IR level) to replace atomic instructions with |
10 | // __atomic_* library calls, or target specific instruction which implement the |
11 | // same semantics in a way which better fits the target backend. This can |
12 | // include the use of (intrinsic-based) load-linked/store-conditional loops, |
13 | // AtomicCmpXchg, or type coercions. |
14 | // |
15 | //===----------------------------------------------------------------------===// |
16 | |
17 | #include "llvm/ADT/ArrayRef.h" |
18 | #include "llvm/ADT/STLFunctionalExtras.h" |
19 | #include "llvm/ADT/SmallVector.h" |
20 | #include "llvm/Analysis/InstSimplifyFolder.h" |
21 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" |
22 | #include "llvm/CodeGen/AtomicExpand.h" |
23 | #include "llvm/CodeGen/AtomicExpandUtils.h" |
24 | #include "llvm/CodeGen/RuntimeLibcallUtil.h" |
25 | #include "llvm/CodeGen/TargetLowering.h" |
26 | #include "llvm/CodeGen/TargetPassConfig.h" |
27 | #include "llvm/CodeGen/TargetSubtargetInfo.h" |
28 | #include "llvm/CodeGen/ValueTypes.h" |
29 | #include "llvm/IR/Attributes.h" |
30 | #include "llvm/IR/BasicBlock.h" |
31 | #include "llvm/IR/Constant.h" |
32 | #include "llvm/IR/Constants.h" |
33 | #include "llvm/IR/DataLayout.h" |
34 | #include "llvm/IR/DerivedTypes.h" |
35 | #include "llvm/IR/Function.h" |
36 | #include "llvm/IR/IRBuilder.h" |
37 | #include "llvm/IR/InstIterator.h" |
38 | #include "llvm/IR/Instruction.h" |
39 | #include "llvm/IR/Instructions.h" |
40 | #include "llvm/IR/MDBuilder.h" |
41 | #include "llvm/IR/MemoryModelRelaxationAnnotations.h" |
42 | #include "llvm/IR/Module.h" |
43 | #include "llvm/IR/Type.h" |
44 | #include "llvm/IR/User.h" |
45 | #include "llvm/IR/Value.h" |
46 | #include "llvm/InitializePasses.h" |
47 | #include "llvm/Pass.h" |
48 | #include "llvm/Support/AtomicOrdering.h" |
49 | #include "llvm/Support/Casting.h" |
50 | #include "llvm/Support/Debug.h" |
51 | #include "llvm/Support/ErrorHandling.h" |
52 | #include "llvm/Support/raw_ostream.h" |
53 | #include "llvm/Target/TargetMachine.h" |
54 | #include "llvm/Transforms/Utils/LowerAtomic.h" |
55 | #include <cassert> |
56 | #include <cstdint> |
57 | #include <iterator> |
58 | |
59 | using namespace llvm; |
60 | |
61 | #define DEBUG_TYPE "atomic-expand" |
62 | |
63 | namespace { |
64 | |
65 | class AtomicExpandImpl { |
66 | const TargetLowering *TLI = nullptr; |
67 | const DataLayout *DL = nullptr; |
68 | |
69 | private: |
70 | bool bracketInstWithFences(Instruction *I, AtomicOrdering Order); |
71 | IntegerType *getCorrespondingIntegerType(Type *T, const DataLayout &DL); |
72 | LoadInst *convertAtomicLoadToIntegerType(LoadInst *LI); |
73 | bool tryExpandAtomicLoad(LoadInst *LI); |
74 | bool expandAtomicLoadToLL(LoadInst *LI); |
75 | bool expandAtomicLoadToCmpXchg(LoadInst *LI); |
76 | StoreInst *convertAtomicStoreToIntegerType(StoreInst *SI); |
77 | bool tryExpandAtomicStore(StoreInst *SI); |
78 | void expandAtomicStore(StoreInst *SI); |
79 | bool tryExpandAtomicRMW(AtomicRMWInst *AI); |
80 | AtomicRMWInst *convertAtomicXchgToIntegerType(AtomicRMWInst *RMWI); |
81 | Value * |
82 | insertRMWLLSCLoop(IRBuilderBase &Builder, Type *ResultTy, Value *Addr, |
83 | Align AddrAlign, AtomicOrdering MemOpOrder, |
84 | function_ref<Value *(IRBuilderBase &, Value *)> PerformOp); |
85 | void expandAtomicOpToLLSC( |
86 | Instruction *I, Type *ResultTy, Value *Addr, Align AddrAlign, |
87 | AtomicOrdering MemOpOrder, |
88 | function_ref<Value *(IRBuilderBase &, Value *)> PerformOp); |
89 | void expandPartwordAtomicRMW( |
90 | AtomicRMWInst *I, TargetLoweringBase::AtomicExpansionKind ExpansionKind); |
91 | AtomicRMWInst *widenPartwordAtomicRMW(AtomicRMWInst *AI); |
92 | bool expandPartwordCmpXchg(AtomicCmpXchgInst *I); |
93 | void expandAtomicRMWToMaskedIntrinsic(AtomicRMWInst *AI); |
94 | void expandAtomicCmpXchgToMaskedIntrinsic(AtomicCmpXchgInst *CI); |
95 | |
96 | AtomicCmpXchgInst *convertCmpXchgToIntegerType(AtomicCmpXchgInst *CI); |
97 | static Value *insertRMWCmpXchgLoop( |
98 | IRBuilderBase &Builder, Type *ResultType, Value *Addr, Align AddrAlign, |
99 | AtomicOrdering MemOpOrder, SyncScope::ID SSID, |
100 | function_ref<Value *(IRBuilderBase &, Value *)> PerformOp, |
101 | CreateCmpXchgInstFun CreateCmpXchg); |
102 | bool tryExpandAtomicCmpXchg(AtomicCmpXchgInst *CI); |
103 | |
104 | bool expandAtomicCmpXchg(AtomicCmpXchgInst *CI); |
105 | bool isIdempotentRMW(AtomicRMWInst *RMWI); |
106 | bool simplifyIdempotentRMW(AtomicRMWInst *RMWI); |
107 | |
108 | bool expandAtomicOpToLibcall(Instruction *I, unsigned Size, Align Alignment, |
109 | Value *PointerOperand, Value *ValueOperand, |
110 | Value *CASExpected, AtomicOrdering Ordering, |
111 | AtomicOrdering Ordering2, |
112 | ArrayRef<RTLIB::Libcall> Libcalls); |
113 | void expandAtomicLoadToLibcall(LoadInst *LI); |
114 | void expandAtomicStoreToLibcall(StoreInst *LI); |
115 | void expandAtomicRMWToLibcall(AtomicRMWInst *I); |
116 | void expandAtomicCASToLibcall(AtomicCmpXchgInst *I); |
117 | |
118 | friend bool |
119 | llvm::expandAtomicRMWToCmpXchg(AtomicRMWInst *AI, |
120 | CreateCmpXchgInstFun CreateCmpXchg); |
121 | |
122 | public: |
123 | bool run(Function &F, const TargetMachine *TM); |
124 | }; |
125 | |
126 | class AtomicExpandLegacy : public FunctionPass { |
127 | public: |
128 | static char ID; // Pass identification, replacement for typeid |
129 | |
130 | AtomicExpandLegacy() : FunctionPass(ID) { |
131 | initializeAtomicExpandLegacyPass(*PassRegistry::getPassRegistry()); |
132 | } |
133 | |
134 | bool runOnFunction(Function &F) override; |
135 | }; |
136 | |
137 | // IRBuilder to be used for replacement atomic instructions. |
138 | struct ReplacementIRBuilder |
139 | : IRBuilder<InstSimplifyFolder, IRBuilderCallbackInserter> { |
140 | MDNode *MMRAMD = nullptr; |
141 | |
142 | // Preserves the DebugLoc from I, and preserves still valid metadata. |
143 | // Enable StrictFP builder mode when appropriate. |
144 | explicit ReplacementIRBuilder(Instruction *I, const DataLayout &DL) |
145 | : IRBuilder(I->getContext(), DL, |
146 | IRBuilderCallbackInserter( |
147 | [this](Instruction *I) { addMMRAMD(I); })) { |
148 | SetInsertPoint(I); |
149 | this->CollectMetadataToCopy(Src: I, MetadataKinds: {LLVMContext::MD_pcsections}); |
150 | if (BB->getParent()->getAttributes().hasFnAttr(Kind: Attribute::StrictFP)) |
151 | this->setIsFPConstrained(true); |
152 | |
153 | MMRAMD = I->getMetadata(KindID: LLVMContext::MD_mmra); |
154 | } |
155 | |
156 | void addMMRAMD(Instruction *I) { |
157 | if (canInstructionHaveMMRAs(I: *I)) |
158 | I->setMetadata(KindID: LLVMContext::MD_mmra, Node: MMRAMD); |
159 | } |
160 | }; |
161 | |
162 | } // end anonymous namespace |
163 | |
164 | char AtomicExpandLegacy::ID = 0; |
165 | |
166 | char &llvm::AtomicExpandID = AtomicExpandLegacy::ID; |
167 | |
168 | INITIALIZE_PASS_BEGIN(AtomicExpandLegacy, DEBUG_TYPE, |
169 | "Expand Atomic instructions" , false, false) |
170 | INITIALIZE_PASS_DEPENDENCY(TargetPassConfig) |
171 | INITIALIZE_PASS_END(AtomicExpandLegacy, DEBUG_TYPE, |
172 | "Expand Atomic instructions" , false, false) |
173 | |
174 | // Helper functions to retrieve the size of atomic instructions. |
175 | static unsigned getAtomicOpSize(LoadInst *LI) { |
176 | const DataLayout &DL = LI->getDataLayout(); |
177 | return DL.getTypeStoreSize(Ty: LI->getType()); |
178 | } |
179 | |
180 | static unsigned getAtomicOpSize(StoreInst *SI) { |
181 | const DataLayout &DL = SI->getDataLayout(); |
182 | return DL.getTypeStoreSize(Ty: SI->getValueOperand()->getType()); |
183 | } |
184 | |
185 | static unsigned getAtomicOpSize(AtomicRMWInst *RMWI) { |
186 | const DataLayout &DL = RMWI->getDataLayout(); |
187 | return DL.getTypeStoreSize(Ty: RMWI->getValOperand()->getType()); |
188 | } |
189 | |
190 | static unsigned getAtomicOpSize(AtomicCmpXchgInst *CASI) { |
191 | const DataLayout &DL = CASI->getDataLayout(); |
192 | return DL.getTypeStoreSize(Ty: CASI->getCompareOperand()->getType()); |
193 | } |
194 | |
195 | // Determine if a particular atomic operation has a supported size, |
196 | // and is of appropriate alignment, to be passed through for target |
197 | // lowering. (Versus turning into a __atomic libcall) |
198 | template <typename Inst> |
199 | static bool atomicSizeSupported(const TargetLowering *TLI, Inst *I) { |
200 | unsigned Size = getAtomicOpSize(I); |
201 | Align Alignment = I->getAlign(); |
202 | return Alignment >= Size && |
203 | Size <= TLI->getMaxAtomicSizeInBitsSupported() / 8; |
204 | } |
205 | |
206 | bool AtomicExpandImpl::run(Function &F, const TargetMachine *TM) { |
207 | const auto *Subtarget = TM->getSubtargetImpl(F); |
208 | if (!Subtarget->enableAtomicExpand()) |
209 | return false; |
210 | TLI = Subtarget->getTargetLowering(); |
211 | DL = &F.getDataLayout(); |
212 | |
213 | SmallVector<Instruction *, 1> AtomicInsts; |
214 | |
215 | // Changing control-flow while iterating through it is a bad idea, so gather a |
216 | // list of all atomic instructions before we start. |
217 | for (Instruction &I : instructions(F)) |
218 | if (I.isAtomic() && !isa<FenceInst>(Val: &I)) |
219 | AtomicInsts.push_back(Elt: &I); |
220 | |
221 | bool MadeChange = false; |
222 | for (auto *I : AtomicInsts) { |
223 | auto LI = dyn_cast<LoadInst>(Val: I); |
224 | auto SI = dyn_cast<StoreInst>(Val: I); |
225 | auto RMWI = dyn_cast<AtomicRMWInst>(Val: I); |
226 | auto CASI = dyn_cast<AtomicCmpXchgInst>(Val: I); |
227 | assert((LI || SI || RMWI || CASI) && "Unknown atomic instruction" ); |
228 | |
229 | // If the Size/Alignment is not supported, replace with a libcall. |
230 | if (LI) { |
231 | if (!atomicSizeSupported(TLI, I: LI)) { |
232 | expandAtomicLoadToLibcall(LI); |
233 | MadeChange = true; |
234 | continue; |
235 | } |
236 | } else if (SI) { |
237 | if (!atomicSizeSupported(TLI, I: SI)) { |
238 | expandAtomicStoreToLibcall(LI: SI); |
239 | MadeChange = true; |
240 | continue; |
241 | } |
242 | } else if (RMWI) { |
243 | if (!atomicSizeSupported(TLI, I: RMWI)) { |
244 | expandAtomicRMWToLibcall(I: RMWI); |
245 | MadeChange = true; |
246 | continue; |
247 | } |
248 | } else if (CASI) { |
249 | if (!atomicSizeSupported(TLI, I: CASI)) { |
250 | expandAtomicCASToLibcall(I: CASI); |
251 | MadeChange = true; |
252 | continue; |
253 | } |
254 | } |
255 | |
256 | if (LI && TLI->shouldCastAtomicLoadInIR(LI) == |
257 | TargetLoweringBase::AtomicExpansionKind::CastToInteger) { |
258 | I = LI = convertAtomicLoadToIntegerType(LI); |
259 | MadeChange = true; |
260 | } else if (SI && |
261 | TLI->shouldCastAtomicStoreInIR(SI) == |
262 | TargetLoweringBase::AtomicExpansionKind::CastToInteger) { |
263 | I = SI = convertAtomicStoreToIntegerType(SI); |
264 | MadeChange = true; |
265 | } else if (RMWI && |
266 | TLI->shouldCastAtomicRMWIInIR(RMWI) == |
267 | TargetLoweringBase::AtomicExpansionKind::CastToInteger) { |
268 | I = RMWI = convertAtomicXchgToIntegerType(RMWI); |
269 | MadeChange = true; |
270 | } else if (CASI) { |
271 | // TODO: when we're ready to make the change at the IR level, we can |
272 | // extend convertCmpXchgToInteger for floating point too. |
273 | if (CASI->getCompareOperand()->getType()->isPointerTy()) { |
274 | // TODO: add a TLI hook to control this so that each target can |
275 | // convert to lowering the original type one at a time. |
276 | I = CASI = convertCmpXchgToIntegerType(CI: CASI); |
277 | MadeChange = true; |
278 | } |
279 | } |
280 | |
281 | if (TLI->shouldInsertFencesForAtomic(I)) { |
282 | auto FenceOrdering = AtomicOrdering::Monotonic; |
283 | if (LI && isAcquireOrStronger(AO: LI->getOrdering())) { |
284 | FenceOrdering = LI->getOrdering(); |
285 | LI->setOrdering(AtomicOrdering::Monotonic); |
286 | } else if (SI && isReleaseOrStronger(AO: SI->getOrdering())) { |
287 | FenceOrdering = SI->getOrdering(); |
288 | SI->setOrdering(AtomicOrdering::Monotonic); |
289 | } else if (RMWI && (isReleaseOrStronger(AO: RMWI->getOrdering()) || |
290 | isAcquireOrStronger(AO: RMWI->getOrdering()))) { |
291 | FenceOrdering = RMWI->getOrdering(); |
292 | RMWI->setOrdering(AtomicOrdering::Monotonic); |
293 | } else if (CASI && |
294 | TLI->shouldExpandAtomicCmpXchgInIR(AI: CASI) == |
295 | TargetLoweringBase::AtomicExpansionKind::None && |
296 | (isReleaseOrStronger(AO: CASI->getSuccessOrdering()) || |
297 | isAcquireOrStronger(AO: CASI->getSuccessOrdering()) || |
298 | isAcquireOrStronger(AO: CASI->getFailureOrdering()))) { |
299 | // If a compare and swap is lowered to LL/SC, we can do smarter fence |
300 | // insertion, with a stronger one on the success path than on the |
301 | // failure path. As a result, fence insertion is directly done by |
302 | // expandAtomicCmpXchg in that case. |
303 | FenceOrdering = CASI->getMergedOrdering(); |
304 | CASI->setSuccessOrdering(AtomicOrdering::Monotonic); |
305 | CASI->setFailureOrdering(AtomicOrdering::Monotonic); |
306 | } |
307 | |
308 | if (FenceOrdering != AtomicOrdering::Monotonic) { |
309 | MadeChange |= bracketInstWithFences(I, Order: FenceOrdering); |
310 | } |
311 | } else if (I->hasAtomicStore() && |
312 | TLI->shouldInsertTrailingFenceForAtomicStore(I)) { |
313 | auto FenceOrdering = AtomicOrdering::Monotonic; |
314 | if (SI) |
315 | FenceOrdering = SI->getOrdering(); |
316 | else if (RMWI) |
317 | FenceOrdering = RMWI->getOrdering(); |
318 | else if (CASI && TLI->shouldExpandAtomicCmpXchgInIR(AI: CASI) != |
319 | TargetLoweringBase::AtomicExpansionKind::LLSC) |
320 | // LLSC is handled in expandAtomicCmpXchg(). |
321 | FenceOrdering = CASI->getSuccessOrdering(); |
322 | |
323 | IRBuilder Builder(I); |
324 | if (auto TrailingFence = |
325 | TLI->emitTrailingFence(Builder, Inst: I, Ord: FenceOrdering)) { |
326 | TrailingFence->moveAfter(MovePos: I); |
327 | MadeChange = true; |
328 | } |
329 | } |
330 | |
331 | if (LI) |
332 | MadeChange |= tryExpandAtomicLoad(LI); |
333 | else if (SI) |
334 | MadeChange |= tryExpandAtomicStore(SI); |
335 | else if (RMWI) { |
336 | // There are two different ways of expanding RMW instructions: |
337 | // - into a load if it is idempotent |
338 | // - into a Cmpxchg/LL-SC loop otherwise |
339 | // we try them in that order. |
340 | |
341 | if (isIdempotentRMW(RMWI) && simplifyIdempotentRMW(RMWI)) { |
342 | MadeChange = true; |
343 | } else { |
344 | MadeChange |= tryExpandAtomicRMW(AI: RMWI); |
345 | } |
346 | } else if (CASI) |
347 | MadeChange |= tryExpandAtomicCmpXchg(CI: CASI); |
348 | } |
349 | return MadeChange; |
350 | } |
351 | |
352 | bool AtomicExpandLegacy::runOnFunction(Function &F) { |
353 | |
354 | auto *TPC = getAnalysisIfAvailable<TargetPassConfig>(); |
355 | if (!TPC) |
356 | return false; |
357 | auto *TM = &TPC->getTM<TargetMachine>(); |
358 | AtomicExpandImpl AE; |
359 | return AE.run(F, TM); |
360 | } |
361 | |
362 | FunctionPass *llvm::createAtomicExpandLegacyPass() { |
363 | return new AtomicExpandLegacy(); |
364 | } |
365 | |
366 | PreservedAnalyses AtomicExpandPass::run(Function &F, |
367 | FunctionAnalysisManager &AM) { |
368 | AtomicExpandImpl AE; |
369 | |
370 | bool Changed = AE.run(F, TM); |
371 | if (!Changed) |
372 | return PreservedAnalyses::all(); |
373 | |
374 | return PreservedAnalyses::none(); |
375 | } |
376 | |
377 | bool AtomicExpandImpl::bracketInstWithFences(Instruction *I, |
378 | AtomicOrdering Order) { |
379 | ReplacementIRBuilder Builder(I, *DL); |
380 | |
381 | auto LeadingFence = TLI->emitLeadingFence(Builder, Inst: I, Ord: Order); |
382 | |
383 | auto TrailingFence = TLI->emitTrailingFence(Builder, Inst: I, Ord: Order); |
384 | // We have a guard here because not every atomic operation generates a |
385 | // trailing fence. |
386 | if (TrailingFence) |
387 | TrailingFence->moveAfter(MovePos: I); |
388 | |
389 | return (LeadingFence || TrailingFence); |
390 | } |
391 | |
392 | /// Get the iX type with the same bitwidth as T. |
393 | IntegerType * |
394 | AtomicExpandImpl::getCorrespondingIntegerType(Type *T, const DataLayout &DL) { |
395 | EVT VT = TLI->getMemValueType(DL, Ty: T); |
396 | unsigned BitWidth = VT.getStoreSizeInBits(); |
397 | assert(BitWidth == VT.getSizeInBits() && "must be a power of two" ); |
398 | return IntegerType::get(C&: T->getContext(), NumBits: BitWidth); |
399 | } |
400 | |
401 | /// Convert an atomic load of a non-integral type to an integer load of the |
402 | /// equivalent bitwidth. See the function comment on |
403 | /// convertAtomicStoreToIntegerType for background. |
404 | LoadInst *AtomicExpandImpl::convertAtomicLoadToIntegerType(LoadInst *LI) { |
405 | auto *M = LI->getModule(); |
406 | Type *NewTy = getCorrespondingIntegerType(T: LI->getType(), DL: M->getDataLayout()); |
407 | |
408 | ReplacementIRBuilder Builder(LI, *DL); |
409 | |
410 | Value *Addr = LI->getPointerOperand(); |
411 | |
412 | auto *NewLI = Builder.CreateLoad(Ty: NewTy, Ptr: Addr); |
413 | NewLI->setAlignment(LI->getAlign()); |
414 | NewLI->setVolatile(LI->isVolatile()); |
415 | NewLI->setAtomic(Ordering: LI->getOrdering(), SSID: LI->getSyncScopeID()); |
416 | LLVM_DEBUG(dbgs() << "Replaced " << *LI << " with " << *NewLI << "\n" ); |
417 | |
418 | Value *NewVal = Builder.CreateBitCast(V: NewLI, DestTy: LI->getType()); |
419 | LI->replaceAllUsesWith(V: NewVal); |
420 | LI->eraseFromParent(); |
421 | return NewLI; |
422 | } |
423 | |
424 | AtomicRMWInst * |
425 | AtomicExpandImpl::convertAtomicXchgToIntegerType(AtomicRMWInst *RMWI) { |
426 | auto *M = RMWI->getModule(); |
427 | Type *NewTy = |
428 | getCorrespondingIntegerType(T: RMWI->getType(), DL: M->getDataLayout()); |
429 | |
430 | ReplacementIRBuilder Builder(RMWI, *DL); |
431 | |
432 | Value *Addr = RMWI->getPointerOperand(); |
433 | Value *Val = RMWI->getValOperand(); |
434 | Value *NewVal = Val->getType()->isPointerTy() |
435 | ? Builder.CreatePtrToInt(V: Val, DestTy: NewTy) |
436 | : Builder.CreateBitCast(V: Val, DestTy: NewTy); |
437 | |
438 | auto *NewRMWI = Builder.CreateAtomicRMW(Op: AtomicRMWInst::Xchg, Ptr: Addr, Val: NewVal, |
439 | Align: RMWI->getAlign(), Ordering: RMWI->getOrdering(), |
440 | SSID: RMWI->getSyncScopeID()); |
441 | NewRMWI->setVolatile(RMWI->isVolatile()); |
442 | LLVM_DEBUG(dbgs() << "Replaced " << *RMWI << " with " << *NewRMWI << "\n" ); |
443 | |
444 | Value *NewRVal = RMWI->getType()->isPointerTy() |
445 | ? Builder.CreateIntToPtr(V: NewRMWI, DestTy: RMWI->getType()) |
446 | : Builder.CreateBitCast(V: NewRMWI, DestTy: RMWI->getType()); |
447 | RMWI->replaceAllUsesWith(V: NewRVal); |
448 | RMWI->eraseFromParent(); |
449 | return NewRMWI; |
450 | } |
451 | |
452 | bool AtomicExpandImpl::tryExpandAtomicLoad(LoadInst *LI) { |
453 | switch (TLI->shouldExpandAtomicLoadInIR(LI)) { |
454 | case TargetLoweringBase::AtomicExpansionKind::None: |
455 | return false; |
456 | case TargetLoweringBase::AtomicExpansionKind::LLSC: |
457 | expandAtomicOpToLLSC( |
458 | I: LI, ResultTy: LI->getType(), Addr: LI->getPointerOperand(), AddrAlign: LI->getAlign(), |
459 | MemOpOrder: LI->getOrdering(), |
460 | PerformOp: [](IRBuilderBase &Builder, Value *Loaded) { return Loaded; }); |
461 | return true; |
462 | case TargetLoweringBase::AtomicExpansionKind::LLOnly: |
463 | return expandAtomicLoadToLL(LI); |
464 | case TargetLoweringBase::AtomicExpansionKind::CmpXChg: |
465 | return expandAtomicLoadToCmpXchg(LI); |
466 | case TargetLoweringBase::AtomicExpansionKind::NotAtomic: |
467 | LI->setAtomic(Ordering: AtomicOrdering::NotAtomic); |
468 | return true; |
469 | default: |
470 | llvm_unreachable("Unhandled case in tryExpandAtomicLoad" ); |
471 | } |
472 | } |
473 | |
474 | bool AtomicExpandImpl::tryExpandAtomicStore(StoreInst *SI) { |
475 | switch (TLI->shouldExpandAtomicStoreInIR(SI)) { |
476 | case TargetLoweringBase::AtomicExpansionKind::None: |
477 | return false; |
478 | case TargetLoweringBase::AtomicExpansionKind::Expand: |
479 | expandAtomicStore(SI); |
480 | return true; |
481 | case TargetLoweringBase::AtomicExpansionKind::NotAtomic: |
482 | SI->setAtomic(Ordering: AtomicOrdering::NotAtomic); |
483 | return true; |
484 | default: |
485 | llvm_unreachable("Unhandled case in tryExpandAtomicStore" ); |
486 | } |
487 | } |
488 | |
489 | bool AtomicExpandImpl::expandAtomicLoadToLL(LoadInst *LI) { |
490 | ReplacementIRBuilder Builder(LI, *DL); |
491 | |
492 | // On some architectures, load-linked instructions are atomic for larger |
493 | // sizes than normal loads. For example, the only 64-bit load guaranteed |
494 | // to be single-copy atomic by ARM is an ldrexd (A3.5.3). |
495 | Value *Val = TLI->emitLoadLinked(Builder, ValueTy: LI->getType(), |
496 | Addr: LI->getPointerOperand(), Ord: LI->getOrdering()); |
497 | TLI->emitAtomicCmpXchgNoStoreLLBalance(Builder); |
498 | |
499 | LI->replaceAllUsesWith(V: Val); |
500 | LI->eraseFromParent(); |
501 | |
502 | return true; |
503 | } |
504 | |
505 | bool AtomicExpandImpl::expandAtomicLoadToCmpXchg(LoadInst *LI) { |
506 | ReplacementIRBuilder Builder(LI, *DL); |
507 | AtomicOrdering Order = LI->getOrdering(); |
508 | if (Order == AtomicOrdering::Unordered) |
509 | Order = AtomicOrdering::Monotonic; |
510 | |
511 | Value *Addr = LI->getPointerOperand(); |
512 | Type *Ty = LI->getType(); |
513 | Constant *DummyVal = Constant::getNullValue(Ty); |
514 | |
515 | Value *Pair = Builder.CreateAtomicCmpXchg( |
516 | Ptr: Addr, Cmp: DummyVal, New: DummyVal, Align: LI->getAlign(), SuccessOrdering: Order, |
517 | FailureOrdering: AtomicCmpXchgInst::getStrongestFailureOrdering(SuccessOrdering: Order)); |
518 | Value *Loaded = Builder.CreateExtractValue(Agg: Pair, Idxs: 0, Name: "loaded" ); |
519 | |
520 | LI->replaceAllUsesWith(V: Loaded); |
521 | LI->eraseFromParent(); |
522 | |
523 | return true; |
524 | } |
525 | |
526 | /// Convert an atomic store of a non-integral type to an integer store of the |
527 | /// equivalent bitwidth. We used to not support floating point or vector |
528 | /// atomics in the IR at all. The backends learned to deal with the bitcast |
529 | /// idiom because that was the only way of expressing the notion of a atomic |
530 | /// float or vector store. The long term plan is to teach each backend to |
531 | /// instruction select from the original atomic store, but as a migration |
532 | /// mechanism, we convert back to the old format which the backends understand. |
533 | /// Each backend will need individual work to recognize the new format. |
534 | StoreInst *AtomicExpandImpl::convertAtomicStoreToIntegerType(StoreInst *SI) { |
535 | ReplacementIRBuilder Builder(SI, *DL); |
536 | auto *M = SI->getModule(); |
537 | Type *NewTy = getCorrespondingIntegerType(T: SI->getValueOperand()->getType(), |
538 | DL: M->getDataLayout()); |
539 | Value *NewVal = Builder.CreateBitCast(V: SI->getValueOperand(), DestTy: NewTy); |
540 | |
541 | Value *Addr = SI->getPointerOperand(); |
542 | |
543 | StoreInst *NewSI = Builder.CreateStore(Val: NewVal, Ptr: Addr); |
544 | NewSI->setAlignment(SI->getAlign()); |
545 | NewSI->setVolatile(SI->isVolatile()); |
546 | NewSI->setAtomic(Ordering: SI->getOrdering(), SSID: SI->getSyncScopeID()); |
547 | LLVM_DEBUG(dbgs() << "Replaced " << *SI << " with " << *NewSI << "\n" ); |
548 | SI->eraseFromParent(); |
549 | return NewSI; |
550 | } |
551 | |
552 | void AtomicExpandImpl::expandAtomicStore(StoreInst *SI) { |
553 | // This function is only called on atomic stores that are too large to be |
554 | // atomic if implemented as a native store. So we replace them by an |
555 | // atomic swap, that can be implemented for example as a ldrex/strex on ARM |
556 | // or lock cmpxchg8/16b on X86, as these are atomic for larger sizes. |
557 | // It is the responsibility of the target to only signal expansion via |
558 | // shouldExpandAtomicRMW in cases where this is required and possible. |
559 | ReplacementIRBuilder Builder(SI, *DL); |
560 | AtomicOrdering Ordering = SI->getOrdering(); |
561 | assert(Ordering != AtomicOrdering::NotAtomic); |
562 | AtomicOrdering RMWOrdering = Ordering == AtomicOrdering::Unordered |
563 | ? AtomicOrdering::Monotonic |
564 | : Ordering; |
565 | AtomicRMWInst *AI = Builder.CreateAtomicRMW( |
566 | Op: AtomicRMWInst::Xchg, Ptr: SI->getPointerOperand(), Val: SI->getValueOperand(), |
567 | Align: SI->getAlign(), Ordering: RMWOrdering); |
568 | SI->eraseFromParent(); |
569 | |
570 | // Now we have an appropriate swap instruction, lower it as usual. |
571 | tryExpandAtomicRMW(AI); |
572 | } |
573 | |
574 | static void createCmpXchgInstFun(IRBuilderBase &Builder, Value *Addr, |
575 | Value *Loaded, Value *NewVal, Align AddrAlign, |
576 | AtomicOrdering MemOpOrder, SyncScope::ID SSID, |
577 | Value *&Success, Value *&NewLoaded) { |
578 | Type *OrigTy = NewVal->getType(); |
579 | |
580 | // This code can go away when cmpxchg supports FP and vector types. |
581 | assert(!OrigTy->isPointerTy()); |
582 | bool NeedBitcast = OrigTy->isFloatingPointTy() || OrigTy->isVectorTy(); |
583 | if (NeedBitcast) { |
584 | IntegerType *IntTy = Builder.getIntNTy(N: OrigTy->getPrimitiveSizeInBits()); |
585 | NewVal = Builder.CreateBitCast(V: NewVal, DestTy: IntTy); |
586 | Loaded = Builder.CreateBitCast(V: Loaded, DestTy: IntTy); |
587 | } |
588 | |
589 | Value *Pair = Builder.CreateAtomicCmpXchg( |
590 | Ptr: Addr, Cmp: Loaded, New: NewVal, Align: AddrAlign, SuccessOrdering: MemOpOrder, |
591 | FailureOrdering: AtomicCmpXchgInst::getStrongestFailureOrdering(SuccessOrdering: MemOpOrder), SSID); |
592 | Success = Builder.CreateExtractValue(Agg: Pair, Idxs: 1, Name: "success" ); |
593 | NewLoaded = Builder.CreateExtractValue(Agg: Pair, Idxs: 0, Name: "newloaded" ); |
594 | |
595 | if (NeedBitcast) |
596 | NewLoaded = Builder.CreateBitCast(V: NewLoaded, DestTy: OrigTy); |
597 | } |
598 | |
599 | bool AtomicExpandImpl::tryExpandAtomicRMW(AtomicRMWInst *AI) { |
600 | LLVMContext &Ctx = AI->getModule()->getContext(); |
601 | TargetLowering::AtomicExpansionKind Kind = TLI->shouldExpandAtomicRMWInIR(RMW: AI); |
602 | switch (Kind) { |
603 | case TargetLoweringBase::AtomicExpansionKind::None: |
604 | return false; |
605 | case TargetLoweringBase::AtomicExpansionKind::LLSC: { |
606 | unsigned MinCASSize = TLI->getMinCmpXchgSizeInBits() / 8; |
607 | unsigned ValueSize = getAtomicOpSize(RMWI: AI); |
608 | if (ValueSize < MinCASSize) { |
609 | expandPartwordAtomicRMW(I: AI, |
610 | ExpansionKind: TargetLoweringBase::AtomicExpansionKind::LLSC); |
611 | } else { |
612 | auto PerformOp = [&](IRBuilderBase &Builder, Value *Loaded) { |
613 | return buildAtomicRMWValue(Op: AI->getOperation(), Builder, Loaded, |
614 | Val: AI->getValOperand()); |
615 | }; |
616 | expandAtomicOpToLLSC(I: AI, ResultTy: AI->getType(), Addr: AI->getPointerOperand(), |
617 | AddrAlign: AI->getAlign(), MemOpOrder: AI->getOrdering(), PerformOp); |
618 | } |
619 | return true; |
620 | } |
621 | case TargetLoweringBase::AtomicExpansionKind::CmpXChg: { |
622 | unsigned MinCASSize = TLI->getMinCmpXchgSizeInBits() / 8; |
623 | unsigned ValueSize = getAtomicOpSize(RMWI: AI); |
624 | if (ValueSize < MinCASSize) { |
625 | expandPartwordAtomicRMW(I: AI, |
626 | ExpansionKind: TargetLoweringBase::AtomicExpansionKind::CmpXChg); |
627 | } else { |
628 | SmallVector<StringRef> SSNs; |
629 | Ctx.getSyncScopeNames(SSNs); |
630 | auto MemScope = SSNs[AI->getSyncScopeID()].empty() |
631 | ? "system" |
632 | : SSNs[AI->getSyncScopeID()]; |
633 | OptimizationRemarkEmitter ORE(AI->getFunction()); |
634 | ORE.emit(RemarkBuilder: [&]() { |
635 | return OptimizationRemark(DEBUG_TYPE, "Passed" , AI) |
636 | << "A compare and swap loop was generated for an atomic " |
637 | << AI->getOperationName(Op: AI->getOperation()) << " operation at " |
638 | << MemScope << " memory scope" ; |
639 | }); |
640 | expandAtomicRMWToCmpXchg(AI, CreateCmpXchg: createCmpXchgInstFun); |
641 | } |
642 | return true; |
643 | } |
644 | case TargetLoweringBase::AtomicExpansionKind::MaskedIntrinsic: { |
645 | unsigned MinCASSize = TLI->getMinCmpXchgSizeInBits() / 8; |
646 | unsigned ValueSize = getAtomicOpSize(RMWI: AI); |
647 | if (ValueSize < MinCASSize) { |
648 | AtomicRMWInst::BinOp Op = AI->getOperation(); |
649 | // Widen And/Or/Xor and give the target another chance at expanding it. |
650 | if (Op == AtomicRMWInst::Or || Op == AtomicRMWInst::Xor || |
651 | Op == AtomicRMWInst::And) { |
652 | tryExpandAtomicRMW(AI: widenPartwordAtomicRMW(AI)); |
653 | return true; |
654 | } |
655 | } |
656 | expandAtomicRMWToMaskedIntrinsic(AI); |
657 | return true; |
658 | } |
659 | case TargetLoweringBase::AtomicExpansionKind::BitTestIntrinsic: { |
660 | TLI->emitBitTestAtomicRMWIntrinsic(AI); |
661 | return true; |
662 | } |
663 | case TargetLoweringBase::AtomicExpansionKind::CmpArithIntrinsic: { |
664 | TLI->emitCmpArithAtomicRMWIntrinsic(AI); |
665 | return true; |
666 | } |
667 | case TargetLoweringBase::AtomicExpansionKind::NotAtomic: |
668 | return lowerAtomicRMWInst(RMWI: AI); |
669 | case TargetLoweringBase::AtomicExpansionKind::Expand: |
670 | TLI->emitExpandAtomicRMW(AI); |
671 | return true; |
672 | default: |
673 | llvm_unreachable("Unhandled case in tryExpandAtomicRMW" ); |
674 | } |
675 | } |
676 | |
677 | namespace { |
678 | |
679 | struct PartwordMaskValues { |
680 | // These three fields are guaranteed to be set by createMaskInstrs. |
681 | Type *WordType = nullptr; |
682 | Type *ValueType = nullptr; |
683 | Type *IntValueType = nullptr; |
684 | Value *AlignedAddr = nullptr; |
685 | Align AlignedAddrAlignment; |
686 | // The remaining fields can be null. |
687 | Value *ShiftAmt = nullptr; |
688 | Value *Mask = nullptr; |
689 | Value *Inv_Mask = nullptr; |
690 | }; |
691 | |
692 | LLVM_ATTRIBUTE_UNUSED |
693 | raw_ostream &operator<<(raw_ostream &O, const PartwordMaskValues &PMV) { |
694 | auto PrintObj = [&O](auto *V) { |
695 | if (V) |
696 | O << *V; |
697 | else |
698 | O << "nullptr" ; |
699 | O << '\n'; |
700 | }; |
701 | O << "PartwordMaskValues {\n" ; |
702 | O << " WordType: " ; |
703 | PrintObj(PMV.WordType); |
704 | O << " ValueType: " ; |
705 | PrintObj(PMV.ValueType); |
706 | O << " AlignedAddr: " ; |
707 | PrintObj(PMV.AlignedAddr); |
708 | O << " AlignedAddrAlignment: " << PMV.AlignedAddrAlignment.value() << '\n'; |
709 | O << " ShiftAmt: " ; |
710 | PrintObj(PMV.ShiftAmt); |
711 | O << " Mask: " ; |
712 | PrintObj(PMV.Mask); |
713 | O << " Inv_Mask: " ; |
714 | PrintObj(PMV.Inv_Mask); |
715 | O << "}\n" ; |
716 | return O; |
717 | } |
718 | |
719 | } // end anonymous namespace |
720 | |
721 | /// This is a helper function which builds instructions to provide |
722 | /// values necessary for partword atomic operations. It takes an |
723 | /// incoming address, Addr, and ValueType, and constructs the address, |
724 | /// shift-amounts and masks needed to work with a larger value of size |
725 | /// WordSize. |
726 | /// |
727 | /// AlignedAddr: Addr rounded down to a multiple of WordSize |
728 | /// |
729 | /// ShiftAmt: Number of bits to right-shift a WordSize value loaded |
730 | /// from AlignAddr for it to have the same value as if |
731 | /// ValueType was loaded from Addr. |
732 | /// |
733 | /// Mask: Value to mask with the value loaded from AlignAddr to |
734 | /// include only the part that would've been loaded from Addr. |
735 | /// |
736 | /// Inv_Mask: The inverse of Mask. |
737 | static PartwordMaskValues createMaskInstrs(IRBuilderBase &Builder, |
738 | Instruction *I, Type *ValueType, |
739 | Value *Addr, Align AddrAlign, |
740 | unsigned MinWordSize) { |
741 | PartwordMaskValues PMV; |
742 | |
743 | Module *M = I->getModule(); |
744 | LLVMContext &Ctx = M->getContext(); |
745 | const DataLayout &DL = M->getDataLayout(); |
746 | unsigned ValueSize = DL.getTypeStoreSize(Ty: ValueType); |
747 | |
748 | PMV.ValueType = PMV.IntValueType = ValueType; |
749 | if (PMV.ValueType->isFloatingPointTy() || PMV.ValueType->isVectorTy()) |
750 | PMV.IntValueType = |
751 | Type::getIntNTy(C&: Ctx, N: ValueType->getPrimitiveSizeInBits()); |
752 | |
753 | PMV.WordType = MinWordSize > ValueSize ? Type::getIntNTy(C&: Ctx, N: MinWordSize * 8) |
754 | : ValueType; |
755 | if (PMV.ValueType == PMV.WordType) { |
756 | PMV.AlignedAddr = Addr; |
757 | PMV.AlignedAddrAlignment = AddrAlign; |
758 | PMV.ShiftAmt = ConstantInt::get(Ty: PMV.ValueType, V: 0); |
759 | PMV.Mask = ConstantInt::get(Ty: PMV.ValueType, V: ~0, /*isSigned*/ IsSigned: true); |
760 | return PMV; |
761 | } |
762 | |
763 | PMV.AlignedAddrAlignment = Align(MinWordSize); |
764 | |
765 | assert(ValueSize < MinWordSize); |
766 | |
767 | PointerType *PtrTy = cast<PointerType>(Val: Addr->getType()); |
768 | IntegerType *IntTy = DL.getIndexType(C&: Ctx, AddressSpace: PtrTy->getAddressSpace()); |
769 | Value *PtrLSB; |
770 | |
771 | if (AddrAlign < MinWordSize) { |
772 | PMV.AlignedAddr = Builder.CreateIntrinsic( |
773 | ID: Intrinsic::ptrmask, Types: {PtrTy, IntTy}, |
774 | Args: {Addr, ConstantInt::get(Ty: IntTy, V: ~(uint64_t)(MinWordSize - 1))}, FMFSource: nullptr, |
775 | Name: "AlignedAddr" ); |
776 | |
777 | Value *AddrInt = Builder.CreatePtrToInt(V: Addr, DestTy: IntTy); |
778 | PtrLSB = Builder.CreateAnd(LHS: AddrInt, RHS: MinWordSize - 1, Name: "PtrLSB" ); |
779 | } else { |
780 | // If the alignment is high enough, the LSB are known 0. |
781 | PMV.AlignedAddr = Addr; |
782 | PtrLSB = ConstantInt::getNullValue(Ty: IntTy); |
783 | } |
784 | |
785 | if (DL.isLittleEndian()) { |
786 | // turn bytes into bits |
787 | PMV.ShiftAmt = Builder.CreateShl(LHS: PtrLSB, RHS: 3); |
788 | } else { |
789 | // turn bytes into bits, and count from the other side. |
790 | PMV.ShiftAmt = Builder.CreateShl( |
791 | LHS: Builder.CreateXor(LHS: PtrLSB, RHS: MinWordSize - ValueSize), RHS: 3); |
792 | } |
793 | |
794 | PMV.ShiftAmt = Builder.CreateTrunc(V: PMV.ShiftAmt, DestTy: PMV.WordType, Name: "ShiftAmt" ); |
795 | PMV.Mask = Builder.CreateShl( |
796 | LHS: ConstantInt::get(Ty: PMV.WordType, V: (1 << (ValueSize * 8)) - 1), RHS: PMV.ShiftAmt, |
797 | Name: "Mask" ); |
798 | |
799 | PMV.Inv_Mask = Builder.CreateNot(V: PMV.Mask, Name: "Inv_Mask" ); |
800 | |
801 | return PMV; |
802 | } |
803 | |
804 | static Value *(IRBuilderBase &Builder, Value *WideWord, |
805 | const PartwordMaskValues &PMV) { |
806 | assert(WideWord->getType() == PMV.WordType && "Widened type mismatch" ); |
807 | if (PMV.WordType == PMV.ValueType) |
808 | return WideWord; |
809 | |
810 | Value *Shift = Builder.CreateLShr(LHS: WideWord, RHS: PMV.ShiftAmt, Name: "shifted" ); |
811 | Value *Trunc = Builder.CreateTrunc(V: Shift, DestTy: PMV.IntValueType, Name: "extracted" ); |
812 | return Builder.CreateBitCast(V: Trunc, DestTy: PMV.ValueType); |
813 | } |
814 | |
815 | static Value *insertMaskedValue(IRBuilderBase &Builder, Value *WideWord, |
816 | Value *Updated, const PartwordMaskValues &PMV) { |
817 | assert(WideWord->getType() == PMV.WordType && "Widened type mismatch" ); |
818 | assert(Updated->getType() == PMV.ValueType && "Value type mismatch" ); |
819 | if (PMV.WordType == PMV.ValueType) |
820 | return Updated; |
821 | |
822 | Updated = Builder.CreateBitCast(V: Updated, DestTy: PMV.IntValueType); |
823 | |
824 | Value *ZExt = Builder.CreateZExt(V: Updated, DestTy: PMV.WordType, Name: "extended" ); |
825 | Value *Shift = |
826 | Builder.CreateShl(LHS: ZExt, RHS: PMV.ShiftAmt, Name: "shifted" , /*HasNUW*/ true); |
827 | Value *And = Builder.CreateAnd(LHS: WideWord, RHS: PMV.Inv_Mask, Name: "unmasked" ); |
828 | Value *Or = Builder.CreateOr(LHS: And, RHS: Shift, Name: "inserted" ); |
829 | return Or; |
830 | } |
831 | |
832 | /// Emit IR to implement a masked version of a given atomicrmw |
833 | /// operation. (That is, only the bits under the Mask should be |
834 | /// affected by the operation) |
835 | static Value *performMaskedAtomicOp(AtomicRMWInst::BinOp Op, |
836 | IRBuilderBase &Builder, Value *Loaded, |
837 | Value *Shifted_Inc, Value *Inc, |
838 | const PartwordMaskValues &PMV) { |
839 | // TODO: update to use |
840 | // https://graphics.stanford.edu/~seander/bithacks.html#MaskedMerge in order |
841 | // to merge bits from two values without requiring PMV.Inv_Mask. |
842 | switch (Op) { |
843 | case AtomicRMWInst::Xchg: { |
844 | Value *Loaded_MaskOut = Builder.CreateAnd(LHS: Loaded, RHS: PMV.Inv_Mask); |
845 | Value *FinalVal = Builder.CreateOr(LHS: Loaded_MaskOut, RHS: Shifted_Inc); |
846 | return FinalVal; |
847 | } |
848 | case AtomicRMWInst::Or: |
849 | case AtomicRMWInst::Xor: |
850 | case AtomicRMWInst::And: |
851 | llvm_unreachable("Or/Xor/And handled by widenPartwordAtomicRMW" ); |
852 | case AtomicRMWInst::Add: |
853 | case AtomicRMWInst::Sub: |
854 | case AtomicRMWInst::Nand: { |
855 | // The other arithmetic ops need to be masked into place. |
856 | Value *NewVal = buildAtomicRMWValue(Op, Builder, Loaded, Val: Shifted_Inc); |
857 | Value *NewVal_Masked = Builder.CreateAnd(LHS: NewVal, RHS: PMV.Mask); |
858 | Value *Loaded_MaskOut = Builder.CreateAnd(LHS: Loaded, RHS: PMV.Inv_Mask); |
859 | Value *FinalVal = Builder.CreateOr(LHS: Loaded_MaskOut, RHS: NewVal_Masked); |
860 | return FinalVal; |
861 | } |
862 | case AtomicRMWInst::Max: |
863 | case AtomicRMWInst::Min: |
864 | case AtomicRMWInst::UMax: |
865 | case AtomicRMWInst::UMin: |
866 | case AtomicRMWInst::FAdd: |
867 | case AtomicRMWInst::FSub: |
868 | case AtomicRMWInst::FMin: |
869 | case AtomicRMWInst::FMax: |
870 | case AtomicRMWInst::UIncWrap: |
871 | case AtomicRMWInst::UDecWrap: { |
872 | // Finally, other ops will operate on the full value, so truncate down to |
873 | // the original size, and expand out again after doing the |
874 | // operation. Bitcasts will be inserted for FP values. |
875 | Value * = extractMaskedValue(Builder, WideWord: Loaded, PMV); |
876 | Value *NewVal = buildAtomicRMWValue(Op, Builder, Loaded: Loaded_Extract, Val: Inc); |
877 | Value *FinalVal = insertMaskedValue(Builder, WideWord: Loaded, Updated: NewVal, PMV); |
878 | return FinalVal; |
879 | } |
880 | default: |
881 | llvm_unreachable("Unknown atomic op" ); |
882 | } |
883 | } |
884 | |
885 | /// Expand a sub-word atomicrmw operation into an appropriate |
886 | /// word-sized operation. |
887 | /// |
888 | /// It will create an LL/SC or cmpxchg loop, as appropriate, the same |
889 | /// way as a typical atomicrmw expansion. The only difference here is |
890 | /// that the operation inside of the loop may operate upon only a |
891 | /// part of the value. |
892 | void AtomicExpandImpl::expandPartwordAtomicRMW( |
893 | AtomicRMWInst *AI, TargetLoweringBase::AtomicExpansionKind ExpansionKind) { |
894 | // Widen And/Or/Xor and give the target another chance at expanding it. |
895 | AtomicRMWInst::BinOp Op = AI->getOperation(); |
896 | if (Op == AtomicRMWInst::Or || Op == AtomicRMWInst::Xor || |
897 | Op == AtomicRMWInst::And) { |
898 | tryExpandAtomicRMW(AI: widenPartwordAtomicRMW(AI)); |
899 | return; |
900 | } |
901 | AtomicOrdering MemOpOrder = AI->getOrdering(); |
902 | SyncScope::ID SSID = AI->getSyncScopeID(); |
903 | |
904 | ReplacementIRBuilder Builder(AI, *DL); |
905 | |
906 | PartwordMaskValues PMV = |
907 | createMaskInstrs(Builder, I: AI, ValueType: AI->getType(), Addr: AI->getPointerOperand(), |
908 | AddrAlign: AI->getAlign(), MinWordSize: TLI->getMinCmpXchgSizeInBits() / 8); |
909 | |
910 | Value *ValOperand_Shifted = nullptr; |
911 | if (Op == AtomicRMWInst::Xchg || Op == AtomicRMWInst::Add || |
912 | Op == AtomicRMWInst::Sub || Op == AtomicRMWInst::Nand) { |
913 | Value *ValOp = Builder.CreateBitCast(V: AI->getValOperand(), DestTy: PMV.IntValueType); |
914 | ValOperand_Shifted = |
915 | Builder.CreateShl(LHS: Builder.CreateZExt(V: ValOp, DestTy: PMV.WordType), RHS: PMV.ShiftAmt, |
916 | Name: "ValOperand_Shifted" ); |
917 | } |
918 | |
919 | auto PerformPartwordOp = [&](IRBuilderBase &Builder, Value *Loaded) { |
920 | return performMaskedAtomicOp(Op, Builder, Loaded, Shifted_Inc: ValOperand_Shifted, |
921 | Inc: AI->getValOperand(), PMV); |
922 | }; |
923 | |
924 | Value *OldResult; |
925 | if (ExpansionKind == TargetLoweringBase::AtomicExpansionKind::CmpXChg) { |
926 | OldResult = insertRMWCmpXchgLoop(Builder, ResultType: PMV.WordType, Addr: PMV.AlignedAddr, |
927 | AddrAlign: PMV.AlignedAddrAlignment, MemOpOrder, SSID, |
928 | PerformOp: PerformPartwordOp, CreateCmpXchg: createCmpXchgInstFun); |
929 | } else { |
930 | assert(ExpansionKind == TargetLoweringBase::AtomicExpansionKind::LLSC); |
931 | OldResult = insertRMWLLSCLoop(Builder, ResultTy: PMV.WordType, Addr: PMV.AlignedAddr, |
932 | AddrAlign: PMV.AlignedAddrAlignment, MemOpOrder, |
933 | PerformOp: PerformPartwordOp); |
934 | } |
935 | |
936 | Value *FinalOldResult = extractMaskedValue(Builder, WideWord: OldResult, PMV); |
937 | AI->replaceAllUsesWith(V: FinalOldResult); |
938 | AI->eraseFromParent(); |
939 | } |
940 | |
941 | /// Copy metadata that's safe to preserve when widening atomics. |
942 | static void copyMetadataForAtomic(Instruction &Dest, |
943 | const Instruction &Source) { |
944 | SmallVector<std::pair<unsigned, MDNode *>, 8> MD; |
945 | Source.getAllMetadata(MDs&: MD); |
946 | LLVMContext &Ctx = Dest.getContext(); |
947 | MDBuilder MDB(Ctx); |
948 | |
949 | for (auto [ID, N] : MD) { |
950 | switch (ID) { |
951 | case LLVMContext::MD_dbg: |
952 | case LLVMContext::MD_tbaa: |
953 | case LLVMContext::MD_tbaa_struct: |
954 | case LLVMContext::MD_alias_scope: |
955 | case LLVMContext::MD_noalias: |
956 | case LLVMContext::MD_access_group: |
957 | case LLVMContext::MD_mmra: |
958 | Dest.setMetadata(KindID: ID, Node: N); |
959 | break; |
960 | default: |
961 | if (ID == Ctx.getMDKindID(Name: "amdgpu.no.remote.memory" )) |
962 | Dest.setMetadata(KindID: ID, Node: N); |
963 | else if (ID == Ctx.getMDKindID(Name: "amdgpu.no.fine.grained.memory" )) |
964 | Dest.setMetadata(KindID: ID, Node: N); |
965 | |
966 | break; |
967 | } |
968 | } |
969 | } |
970 | |
971 | // Widen the bitwise atomicrmw (or/xor/and) to the minimum supported width. |
972 | AtomicRMWInst *AtomicExpandImpl::widenPartwordAtomicRMW(AtomicRMWInst *AI) { |
973 | ReplacementIRBuilder Builder(AI, *DL); |
974 | AtomicRMWInst::BinOp Op = AI->getOperation(); |
975 | |
976 | assert((Op == AtomicRMWInst::Or || Op == AtomicRMWInst::Xor || |
977 | Op == AtomicRMWInst::And) && |
978 | "Unable to widen operation" ); |
979 | |
980 | PartwordMaskValues PMV = |
981 | createMaskInstrs(Builder, I: AI, ValueType: AI->getType(), Addr: AI->getPointerOperand(), |
982 | AddrAlign: AI->getAlign(), MinWordSize: TLI->getMinCmpXchgSizeInBits() / 8); |
983 | |
984 | Value *ValOperand_Shifted = |
985 | Builder.CreateShl(LHS: Builder.CreateZExt(V: AI->getValOperand(), DestTy: PMV.WordType), |
986 | RHS: PMV.ShiftAmt, Name: "ValOperand_Shifted" ); |
987 | |
988 | Value *NewOperand; |
989 | |
990 | if (Op == AtomicRMWInst::And) |
991 | NewOperand = |
992 | Builder.CreateOr(LHS: ValOperand_Shifted, RHS: PMV.Inv_Mask, Name: "AndOperand" ); |
993 | else |
994 | NewOperand = ValOperand_Shifted; |
995 | |
996 | AtomicRMWInst *NewAI = Builder.CreateAtomicRMW( |
997 | Op, Ptr: PMV.AlignedAddr, Val: NewOperand, Align: PMV.AlignedAddrAlignment, |
998 | Ordering: AI->getOrdering(), SSID: AI->getSyncScopeID()); |
999 | |
1000 | copyMetadataForAtomic(Dest&: *NewAI, Source: *AI); |
1001 | |
1002 | Value *FinalOldResult = extractMaskedValue(Builder, WideWord: NewAI, PMV); |
1003 | AI->replaceAllUsesWith(V: FinalOldResult); |
1004 | AI->eraseFromParent(); |
1005 | return NewAI; |
1006 | } |
1007 | |
1008 | bool AtomicExpandImpl::expandPartwordCmpXchg(AtomicCmpXchgInst *CI) { |
1009 | // The basic idea here is that we're expanding a cmpxchg of a |
1010 | // smaller memory size up to a word-sized cmpxchg. To do this, we |
1011 | // need to add a retry-loop for strong cmpxchg, so that |
1012 | // modifications to other parts of the word don't cause a spurious |
1013 | // failure. |
1014 | |
1015 | // This generates code like the following: |
1016 | // [[Setup mask values PMV.*]] |
1017 | // %NewVal_Shifted = shl i32 %NewVal, %PMV.ShiftAmt |
1018 | // %Cmp_Shifted = shl i32 %Cmp, %PMV.ShiftAmt |
1019 | // %InitLoaded = load i32* %addr |
1020 | // %InitLoaded_MaskOut = and i32 %InitLoaded, %PMV.Inv_Mask |
1021 | // br partword.cmpxchg.loop |
1022 | // partword.cmpxchg.loop: |
1023 | // %Loaded_MaskOut = phi i32 [ %InitLoaded_MaskOut, %entry ], |
1024 | // [ %OldVal_MaskOut, %partword.cmpxchg.failure ] |
1025 | // %FullWord_NewVal = or i32 %Loaded_MaskOut, %NewVal_Shifted |
1026 | // %FullWord_Cmp = or i32 %Loaded_MaskOut, %Cmp_Shifted |
1027 | // %NewCI = cmpxchg i32* %PMV.AlignedAddr, i32 %FullWord_Cmp, |
1028 | // i32 %FullWord_NewVal success_ordering failure_ordering |
1029 | // %OldVal = extractvalue { i32, i1 } %NewCI, 0 |
1030 | // %Success = extractvalue { i32, i1 } %NewCI, 1 |
1031 | // br i1 %Success, label %partword.cmpxchg.end, |
1032 | // label %partword.cmpxchg.failure |
1033 | // partword.cmpxchg.failure: |
1034 | // %OldVal_MaskOut = and i32 %OldVal, %PMV.Inv_Mask |
1035 | // %ShouldContinue = icmp ne i32 %Loaded_MaskOut, %OldVal_MaskOut |
1036 | // br i1 %ShouldContinue, label %partword.cmpxchg.loop, |
1037 | // label %partword.cmpxchg.end |
1038 | // partword.cmpxchg.end: |
1039 | // %tmp1 = lshr i32 %OldVal, %PMV.ShiftAmt |
1040 | // %FinalOldVal = trunc i32 %tmp1 to i8 |
1041 | // %tmp2 = insertvalue { i8, i1 } undef, i8 %FinalOldVal, 0 |
1042 | // %Res = insertvalue { i8, i1 } %25, i1 %Success, 1 |
1043 | |
1044 | Value *Addr = CI->getPointerOperand(); |
1045 | Value *Cmp = CI->getCompareOperand(); |
1046 | Value *NewVal = CI->getNewValOperand(); |
1047 | |
1048 | BasicBlock *BB = CI->getParent(); |
1049 | Function *F = BB->getParent(); |
1050 | ReplacementIRBuilder Builder(CI, *DL); |
1051 | LLVMContext &Ctx = Builder.getContext(); |
1052 | |
1053 | BasicBlock *EndBB = |
1054 | BB->splitBasicBlock(I: CI->getIterator(), BBName: "partword.cmpxchg.end" ); |
1055 | auto FailureBB = |
1056 | BasicBlock::Create(Context&: Ctx, Name: "partword.cmpxchg.failure" , Parent: F, InsertBefore: EndBB); |
1057 | auto LoopBB = BasicBlock::Create(Context&: Ctx, Name: "partword.cmpxchg.loop" , Parent: F, InsertBefore: FailureBB); |
1058 | |
1059 | // The split call above "helpfully" added a branch at the end of BB |
1060 | // (to the wrong place). |
1061 | std::prev(x: BB->end())->eraseFromParent(); |
1062 | Builder.SetInsertPoint(BB); |
1063 | |
1064 | PartwordMaskValues PMV = |
1065 | createMaskInstrs(Builder, I: CI, ValueType: CI->getCompareOperand()->getType(), Addr, |
1066 | AddrAlign: CI->getAlign(), MinWordSize: TLI->getMinCmpXchgSizeInBits() / 8); |
1067 | |
1068 | // Shift the incoming values over, into the right location in the word. |
1069 | Value *NewVal_Shifted = |
1070 | Builder.CreateShl(LHS: Builder.CreateZExt(V: NewVal, DestTy: PMV.WordType), RHS: PMV.ShiftAmt); |
1071 | Value *Cmp_Shifted = |
1072 | Builder.CreateShl(LHS: Builder.CreateZExt(V: Cmp, DestTy: PMV.WordType), RHS: PMV.ShiftAmt); |
1073 | |
1074 | // Load the entire current word, and mask into place the expected and new |
1075 | // values |
1076 | LoadInst *InitLoaded = Builder.CreateLoad(Ty: PMV.WordType, Ptr: PMV.AlignedAddr); |
1077 | InitLoaded->setVolatile(CI->isVolatile()); |
1078 | Value *InitLoaded_MaskOut = Builder.CreateAnd(LHS: InitLoaded, RHS: PMV.Inv_Mask); |
1079 | Builder.CreateBr(Dest: LoopBB); |
1080 | |
1081 | // partword.cmpxchg.loop: |
1082 | Builder.SetInsertPoint(LoopBB); |
1083 | PHINode *Loaded_MaskOut = Builder.CreatePHI(Ty: PMV.WordType, NumReservedValues: 2); |
1084 | Loaded_MaskOut->addIncoming(V: InitLoaded_MaskOut, BB); |
1085 | |
1086 | // Mask/Or the expected and new values into place in the loaded word. |
1087 | Value *FullWord_NewVal = Builder.CreateOr(LHS: Loaded_MaskOut, RHS: NewVal_Shifted); |
1088 | Value *FullWord_Cmp = Builder.CreateOr(LHS: Loaded_MaskOut, RHS: Cmp_Shifted); |
1089 | AtomicCmpXchgInst *NewCI = Builder.CreateAtomicCmpXchg( |
1090 | Ptr: PMV.AlignedAddr, Cmp: FullWord_Cmp, New: FullWord_NewVal, Align: PMV.AlignedAddrAlignment, |
1091 | SuccessOrdering: CI->getSuccessOrdering(), FailureOrdering: CI->getFailureOrdering(), SSID: CI->getSyncScopeID()); |
1092 | NewCI->setVolatile(CI->isVolatile()); |
1093 | // When we're building a strong cmpxchg, we need a loop, so you |
1094 | // might think we could use a weak cmpxchg inside. But, using strong |
1095 | // allows the below comparison for ShouldContinue, and we're |
1096 | // expecting the underlying cmpxchg to be a machine instruction, |
1097 | // which is strong anyways. |
1098 | NewCI->setWeak(CI->isWeak()); |
1099 | |
1100 | Value *OldVal = Builder.CreateExtractValue(Agg: NewCI, Idxs: 0); |
1101 | Value *Success = Builder.CreateExtractValue(Agg: NewCI, Idxs: 1); |
1102 | |
1103 | if (CI->isWeak()) |
1104 | Builder.CreateBr(Dest: EndBB); |
1105 | else |
1106 | Builder.CreateCondBr(Cond: Success, True: EndBB, False: FailureBB); |
1107 | |
1108 | // partword.cmpxchg.failure: |
1109 | Builder.SetInsertPoint(FailureBB); |
1110 | // Upon failure, verify that the masked-out part of the loaded value |
1111 | // has been modified. If it didn't, abort the cmpxchg, since the |
1112 | // masked-in part must've. |
1113 | Value *OldVal_MaskOut = Builder.CreateAnd(LHS: OldVal, RHS: PMV.Inv_Mask); |
1114 | Value *ShouldContinue = Builder.CreateICmpNE(LHS: Loaded_MaskOut, RHS: OldVal_MaskOut); |
1115 | Builder.CreateCondBr(Cond: ShouldContinue, True: LoopBB, False: EndBB); |
1116 | |
1117 | // Add the second value to the phi from above |
1118 | Loaded_MaskOut->addIncoming(V: OldVal_MaskOut, BB: FailureBB); |
1119 | |
1120 | // partword.cmpxchg.end: |
1121 | Builder.SetInsertPoint(CI); |
1122 | |
1123 | Value *FinalOldVal = extractMaskedValue(Builder, WideWord: OldVal, PMV); |
1124 | Value *Res = PoisonValue::get(T: CI->getType()); |
1125 | Res = Builder.CreateInsertValue(Agg: Res, Val: FinalOldVal, Idxs: 0); |
1126 | Res = Builder.CreateInsertValue(Agg: Res, Val: Success, Idxs: 1); |
1127 | |
1128 | CI->replaceAllUsesWith(V: Res); |
1129 | CI->eraseFromParent(); |
1130 | return true; |
1131 | } |
1132 | |
1133 | void AtomicExpandImpl::expandAtomicOpToLLSC( |
1134 | Instruction *I, Type *ResultType, Value *Addr, Align AddrAlign, |
1135 | AtomicOrdering MemOpOrder, |
1136 | function_ref<Value *(IRBuilderBase &, Value *)> PerformOp) { |
1137 | ReplacementIRBuilder Builder(I, *DL); |
1138 | Value *Loaded = insertRMWLLSCLoop(Builder, ResultTy: ResultType, Addr, AddrAlign, |
1139 | MemOpOrder, PerformOp); |
1140 | |
1141 | I->replaceAllUsesWith(V: Loaded); |
1142 | I->eraseFromParent(); |
1143 | } |
1144 | |
1145 | void AtomicExpandImpl::expandAtomicRMWToMaskedIntrinsic(AtomicRMWInst *AI) { |
1146 | ReplacementIRBuilder Builder(AI, *DL); |
1147 | |
1148 | PartwordMaskValues PMV = |
1149 | createMaskInstrs(Builder, I: AI, ValueType: AI->getType(), Addr: AI->getPointerOperand(), |
1150 | AddrAlign: AI->getAlign(), MinWordSize: TLI->getMinCmpXchgSizeInBits() / 8); |
1151 | |
1152 | // The value operand must be sign-extended for signed min/max so that the |
1153 | // target's signed comparison instructions can be used. Otherwise, just |
1154 | // zero-ext. |
1155 | Instruction::CastOps CastOp = Instruction::ZExt; |
1156 | AtomicRMWInst::BinOp RMWOp = AI->getOperation(); |
1157 | if (RMWOp == AtomicRMWInst::Max || RMWOp == AtomicRMWInst::Min) |
1158 | CastOp = Instruction::SExt; |
1159 | |
1160 | Value *ValOperand_Shifted = Builder.CreateShl( |
1161 | LHS: Builder.CreateCast(Op: CastOp, V: AI->getValOperand(), DestTy: PMV.WordType), |
1162 | RHS: PMV.ShiftAmt, Name: "ValOperand_Shifted" ); |
1163 | Value *OldResult = TLI->emitMaskedAtomicRMWIntrinsic( |
1164 | Builder, AI, AlignedAddr: PMV.AlignedAddr, Incr: ValOperand_Shifted, Mask: PMV.Mask, ShiftAmt: PMV.ShiftAmt, |
1165 | Ord: AI->getOrdering()); |
1166 | Value *FinalOldResult = extractMaskedValue(Builder, WideWord: OldResult, PMV); |
1167 | AI->replaceAllUsesWith(V: FinalOldResult); |
1168 | AI->eraseFromParent(); |
1169 | } |
1170 | |
1171 | void AtomicExpandImpl::expandAtomicCmpXchgToMaskedIntrinsic( |
1172 | AtomicCmpXchgInst *CI) { |
1173 | ReplacementIRBuilder Builder(CI, *DL); |
1174 | |
1175 | PartwordMaskValues PMV = createMaskInstrs( |
1176 | Builder, I: CI, ValueType: CI->getCompareOperand()->getType(), Addr: CI->getPointerOperand(), |
1177 | AddrAlign: CI->getAlign(), MinWordSize: TLI->getMinCmpXchgSizeInBits() / 8); |
1178 | |
1179 | Value *CmpVal_Shifted = Builder.CreateShl( |
1180 | LHS: Builder.CreateZExt(V: CI->getCompareOperand(), DestTy: PMV.WordType), RHS: PMV.ShiftAmt, |
1181 | Name: "CmpVal_Shifted" ); |
1182 | Value *NewVal_Shifted = Builder.CreateShl( |
1183 | LHS: Builder.CreateZExt(V: CI->getNewValOperand(), DestTy: PMV.WordType), RHS: PMV.ShiftAmt, |
1184 | Name: "NewVal_Shifted" ); |
1185 | Value *OldVal = TLI->emitMaskedAtomicCmpXchgIntrinsic( |
1186 | Builder, CI, AlignedAddr: PMV.AlignedAddr, CmpVal: CmpVal_Shifted, NewVal: NewVal_Shifted, Mask: PMV.Mask, |
1187 | Ord: CI->getMergedOrdering()); |
1188 | Value *FinalOldVal = extractMaskedValue(Builder, WideWord: OldVal, PMV); |
1189 | Value *Res = PoisonValue::get(T: CI->getType()); |
1190 | Res = Builder.CreateInsertValue(Agg: Res, Val: FinalOldVal, Idxs: 0); |
1191 | Value *Success = Builder.CreateICmpEQ( |
1192 | LHS: CmpVal_Shifted, RHS: Builder.CreateAnd(LHS: OldVal, RHS: PMV.Mask), Name: "Success" ); |
1193 | Res = Builder.CreateInsertValue(Agg: Res, Val: Success, Idxs: 1); |
1194 | |
1195 | CI->replaceAllUsesWith(V: Res); |
1196 | CI->eraseFromParent(); |
1197 | } |
1198 | |
1199 | Value *AtomicExpandImpl::insertRMWLLSCLoop( |
1200 | IRBuilderBase &Builder, Type *ResultTy, Value *Addr, Align AddrAlign, |
1201 | AtomicOrdering MemOpOrder, |
1202 | function_ref<Value *(IRBuilderBase &, Value *)> PerformOp) { |
1203 | LLVMContext &Ctx = Builder.getContext(); |
1204 | BasicBlock *BB = Builder.GetInsertBlock(); |
1205 | Function *F = BB->getParent(); |
1206 | |
1207 | assert(AddrAlign >= |
1208 | F->getDataLayout().getTypeStoreSize(ResultTy) && |
1209 | "Expected at least natural alignment at this point." ); |
1210 | |
1211 | // Given: atomicrmw some_op iN* %addr, iN %incr ordering |
1212 | // |
1213 | // The standard expansion we produce is: |
1214 | // [...] |
1215 | // atomicrmw.start: |
1216 | // %loaded = @load.linked(%addr) |
1217 | // %new = some_op iN %loaded, %incr |
1218 | // %stored = @store_conditional(%new, %addr) |
1219 | // %try_again = icmp i32 ne %stored, 0 |
1220 | // br i1 %try_again, label %loop, label %atomicrmw.end |
1221 | // atomicrmw.end: |
1222 | // [...] |
1223 | BasicBlock *ExitBB = |
1224 | BB->splitBasicBlock(I: Builder.GetInsertPoint(), BBName: "atomicrmw.end" ); |
1225 | BasicBlock *LoopBB = BasicBlock::Create(Context&: Ctx, Name: "atomicrmw.start" , Parent: F, InsertBefore: ExitBB); |
1226 | |
1227 | // The split call above "helpfully" added a branch at the end of BB (to the |
1228 | // wrong place). |
1229 | std::prev(x: BB->end())->eraseFromParent(); |
1230 | Builder.SetInsertPoint(BB); |
1231 | Builder.CreateBr(Dest: LoopBB); |
1232 | |
1233 | // Start the main loop block now that we've taken care of the preliminaries. |
1234 | Builder.SetInsertPoint(LoopBB); |
1235 | Value *Loaded = TLI->emitLoadLinked(Builder, ValueTy: ResultTy, Addr, Ord: MemOpOrder); |
1236 | |
1237 | Value *NewVal = PerformOp(Builder, Loaded); |
1238 | |
1239 | Value *StoreSuccess = |
1240 | TLI->emitStoreConditional(Builder, Val: NewVal, Addr, Ord: MemOpOrder); |
1241 | Value *TryAgain = Builder.CreateICmpNE( |
1242 | LHS: StoreSuccess, RHS: ConstantInt::get(Ty: IntegerType::get(C&: Ctx, NumBits: 32), V: 0), Name: "tryagain" ); |
1243 | Builder.CreateCondBr(Cond: TryAgain, True: LoopBB, False: ExitBB); |
1244 | |
1245 | Builder.SetInsertPoint(TheBB: ExitBB, IP: ExitBB->begin()); |
1246 | return Loaded; |
1247 | } |
1248 | |
1249 | /// Convert an atomic cmpxchg of a non-integral type to an integer cmpxchg of |
1250 | /// the equivalent bitwidth. We used to not support pointer cmpxchg in the |
1251 | /// IR. As a migration step, we convert back to what use to be the standard |
1252 | /// way to represent a pointer cmpxchg so that we can update backends one by |
1253 | /// one. |
1254 | AtomicCmpXchgInst * |
1255 | AtomicExpandImpl::convertCmpXchgToIntegerType(AtomicCmpXchgInst *CI) { |
1256 | auto *M = CI->getModule(); |
1257 | Type *NewTy = getCorrespondingIntegerType(T: CI->getCompareOperand()->getType(), |
1258 | DL: M->getDataLayout()); |
1259 | |
1260 | ReplacementIRBuilder Builder(CI, *DL); |
1261 | |
1262 | Value *Addr = CI->getPointerOperand(); |
1263 | |
1264 | Value *NewCmp = Builder.CreatePtrToInt(V: CI->getCompareOperand(), DestTy: NewTy); |
1265 | Value *NewNewVal = Builder.CreatePtrToInt(V: CI->getNewValOperand(), DestTy: NewTy); |
1266 | |
1267 | auto *NewCI = Builder.CreateAtomicCmpXchg( |
1268 | Ptr: Addr, Cmp: NewCmp, New: NewNewVal, Align: CI->getAlign(), SuccessOrdering: CI->getSuccessOrdering(), |
1269 | FailureOrdering: CI->getFailureOrdering(), SSID: CI->getSyncScopeID()); |
1270 | NewCI->setVolatile(CI->isVolatile()); |
1271 | NewCI->setWeak(CI->isWeak()); |
1272 | LLVM_DEBUG(dbgs() << "Replaced " << *CI << " with " << *NewCI << "\n" ); |
1273 | |
1274 | Value *OldVal = Builder.CreateExtractValue(Agg: NewCI, Idxs: 0); |
1275 | Value *Succ = Builder.CreateExtractValue(Agg: NewCI, Idxs: 1); |
1276 | |
1277 | OldVal = Builder.CreateIntToPtr(V: OldVal, DestTy: CI->getCompareOperand()->getType()); |
1278 | |
1279 | Value *Res = PoisonValue::get(T: CI->getType()); |
1280 | Res = Builder.CreateInsertValue(Agg: Res, Val: OldVal, Idxs: 0); |
1281 | Res = Builder.CreateInsertValue(Agg: Res, Val: Succ, Idxs: 1); |
1282 | |
1283 | CI->replaceAllUsesWith(V: Res); |
1284 | CI->eraseFromParent(); |
1285 | return NewCI; |
1286 | } |
1287 | |
1288 | bool AtomicExpandImpl::expandAtomicCmpXchg(AtomicCmpXchgInst *CI) { |
1289 | AtomicOrdering SuccessOrder = CI->getSuccessOrdering(); |
1290 | AtomicOrdering FailureOrder = CI->getFailureOrdering(); |
1291 | Value *Addr = CI->getPointerOperand(); |
1292 | BasicBlock *BB = CI->getParent(); |
1293 | Function *F = BB->getParent(); |
1294 | LLVMContext &Ctx = F->getContext(); |
1295 | // If shouldInsertFencesForAtomic() returns true, then the target does not |
1296 | // want to deal with memory orders, and emitLeading/TrailingFence should take |
1297 | // care of everything. Otherwise, emitLeading/TrailingFence are no-op and we |
1298 | // should preserve the ordering. |
1299 | bool ShouldInsertFencesForAtomic = TLI->shouldInsertFencesForAtomic(I: CI); |
1300 | AtomicOrdering MemOpOrder = ShouldInsertFencesForAtomic |
1301 | ? AtomicOrdering::Monotonic |
1302 | : CI->getMergedOrdering(); |
1303 | |
1304 | // In implementations which use a barrier to achieve release semantics, we can |
1305 | // delay emitting this barrier until we know a store is actually going to be |
1306 | // attempted. The cost of this delay is that we need 2 copies of the block |
1307 | // emitting the load-linked, affecting code size. |
1308 | // |
1309 | // Ideally, this logic would be unconditional except for the minsize check |
1310 | // since in other cases the extra blocks naturally collapse down to the |
1311 | // minimal loop. Unfortunately, this puts too much stress on later |
1312 | // optimisations so we avoid emitting the extra logic in those cases too. |
1313 | bool HasReleasedLoadBB = !CI->isWeak() && ShouldInsertFencesForAtomic && |
1314 | SuccessOrder != AtomicOrdering::Monotonic && |
1315 | SuccessOrder != AtomicOrdering::Acquire && |
1316 | !F->hasMinSize(); |
1317 | |
1318 | // There's no overhead for sinking the release barrier in a weak cmpxchg, so |
1319 | // do it even on minsize. |
1320 | bool UseUnconditionalReleaseBarrier = F->hasMinSize() && !CI->isWeak(); |
1321 | |
1322 | // Given: cmpxchg some_op iN* %addr, iN %desired, iN %new success_ord fail_ord |
1323 | // |
1324 | // The full expansion we produce is: |
1325 | // [...] |
1326 | // %aligned.addr = ... |
1327 | // cmpxchg.start: |
1328 | // %unreleasedload = @load.linked(%aligned.addr) |
1329 | // %unreleasedload.extract = extract value from %unreleasedload |
1330 | // %should_store = icmp eq %unreleasedload.extract, %desired |
1331 | // br i1 %should_store, label %cmpxchg.releasingstore, |
1332 | // label %cmpxchg.nostore |
1333 | // cmpxchg.releasingstore: |
1334 | // fence? |
1335 | // br label cmpxchg.trystore |
1336 | // cmpxchg.trystore: |
1337 | // %loaded.trystore = phi [%unreleasedload, %cmpxchg.releasingstore], |
1338 | // [%releasedload, %cmpxchg.releasedload] |
1339 | // %updated.new = insert %new into %loaded.trystore |
1340 | // %stored = @store_conditional(%updated.new, %aligned.addr) |
1341 | // %success = icmp eq i32 %stored, 0 |
1342 | // br i1 %success, label %cmpxchg.success, |
1343 | // label %cmpxchg.releasedload/%cmpxchg.failure |
1344 | // cmpxchg.releasedload: |
1345 | // %releasedload = @load.linked(%aligned.addr) |
1346 | // %releasedload.extract = extract value from %releasedload |
1347 | // %should_store = icmp eq %releasedload.extract, %desired |
1348 | // br i1 %should_store, label %cmpxchg.trystore, |
1349 | // label %cmpxchg.failure |
1350 | // cmpxchg.success: |
1351 | // fence? |
1352 | // br label %cmpxchg.end |
1353 | // cmpxchg.nostore: |
1354 | // %loaded.nostore = phi [%unreleasedload, %cmpxchg.start], |
1355 | // [%releasedload, |
1356 | // %cmpxchg.releasedload/%cmpxchg.trystore] |
1357 | // @load_linked_fail_balance()? |
1358 | // br label %cmpxchg.failure |
1359 | // cmpxchg.failure: |
1360 | // fence? |
1361 | // br label %cmpxchg.end |
1362 | // cmpxchg.end: |
1363 | // %loaded.exit = phi [%loaded.nostore, %cmpxchg.failure], |
1364 | // [%loaded.trystore, %cmpxchg.trystore] |
1365 | // %success = phi i1 [true, %cmpxchg.success], [false, %cmpxchg.failure] |
1366 | // %loaded = extract value from %loaded.exit |
1367 | // %restmp = insertvalue { iN, i1 } undef, iN %loaded, 0 |
1368 | // %res = insertvalue { iN, i1 } %restmp, i1 %success, 1 |
1369 | // [...] |
1370 | BasicBlock *ExitBB = BB->splitBasicBlock(I: CI->getIterator(), BBName: "cmpxchg.end" ); |
1371 | auto FailureBB = BasicBlock::Create(Context&: Ctx, Name: "cmpxchg.failure" , Parent: F, InsertBefore: ExitBB); |
1372 | auto NoStoreBB = BasicBlock::Create(Context&: Ctx, Name: "cmpxchg.nostore" , Parent: F, InsertBefore: FailureBB); |
1373 | auto SuccessBB = BasicBlock::Create(Context&: Ctx, Name: "cmpxchg.success" , Parent: F, InsertBefore: NoStoreBB); |
1374 | auto ReleasedLoadBB = |
1375 | BasicBlock::Create(Context&: Ctx, Name: "cmpxchg.releasedload" , Parent: F, InsertBefore: SuccessBB); |
1376 | auto TryStoreBB = |
1377 | BasicBlock::Create(Context&: Ctx, Name: "cmpxchg.trystore" , Parent: F, InsertBefore: ReleasedLoadBB); |
1378 | auto ReleasingStoreBB = |
1379 | BasicBlock::Create(Context&: Ctx, Name: "cmpxchg.fencedstore" , Parent: F, InsertBefore: TryStoreBB); |
1380 | auto StartBB = BasicBlock::Create(Context&: Ctx, Name: "cmpxchg.start" , Parent: F, InsertBefore: ReleasingStoreBB); |
1381 | |
1382 | ReplacementIRBuilder Builder(CI, *DL); |
1383 | |
1384 | // The split call above "helpfully" added a branch at the end of BB (to the |
1385 | // wrong place), but we might want a fence too. It's easiest to just remove |
1386 | // the branch entirely. |
1387 | std::prev(x: BB->end())->eraseFromParent(); |
1388 | Builder.SetInsertPoint(BB); |
1389 | if (ShouldInsertFencesForAtomic && UseUnconditionalReleaseBarrier) |
1390 | TLI->emitLeadingFence(Builder, Inst: CI, Ord: SuccessOrder); |
1391 | |
1392 | PartwordMaskValues PMV = |
1393 | createMaskInstrs(Builder, I: CI, ValueType: CI->getCompareOperand()->getType(), Addr, |
1394 | AddrAlign: CI->getAlign(), MinWordSize: TLI->getMinCmpXchgSizeInBits() / 8); |
1395 | Builder.CreateBr(Dest: StartBB); |
1396 | |
1397 | // Start the main loop block now that we've taken care of the preliminaries. |
1398 | Builder.SetInsertPoint(StartBB); |
1399 | Value *UnreleasedLoad = |
1400 | TLI->emitLoadLinked(Builder, ValueTy: PMV.WordType, Addr: PMV.AlignedAddr, Ord: MemOpOrder); |
1401 | Value * = |
1402 | extractMaskedValue(Builder, WideWord: UnreleasedLoad, PMV); |
1403 | Value *ShouldStore = Builder.CreateICmpEQ( |
1404 | LHS: UnreleasedLoadExtract, RHS: CI->getCompareOperand(), Name: "should_store" ); |
1405 | |
1406 | // If the cmpxchg doesn't actually need any ordering when it fails, we can |
1407 | // jump straight past that fence instruction (if it exists). |
1408 | Builder.CreateCondBr(Cond: ShouldStore, True: ReleasingStoreBB, False: NoStoreBB); |
1409 | |
1410 | Builder.SetInsertPoint(ReleasingStoreBB); |
1411 | if (ShouldInsertFencesForAtomic && !UseUnconditionalReleaseBarrier) |
1412 | TLI->emitLeadingFence(Builder, Inst: CI, Ord: SuccessOrder); |
1413 | Builder.CreateBr(Dest: TryStoreBB); |
1414 | |
1415 | Builder.SetInsertPoint(TryStoreBB); |
1416 | PHINode *LoadedTryStore = |
1417 | Builder.CreatePHI(Ty: PMV.WordType, NumReservedValues: 2, Name: "loaded.trystore" ); |
1418 | LoadedTryStore->addIncoming(V: UnreleasedLoad, BB: ReleasingStoreBB); |
1419 | Value *NewValueInsert = |
1420 | insertMaskedValue(Builder, WideWord: LoadedTryStore, Updated: CI->getNewValOperand(), PMV); |
1421 | Value *StoreSuccess = TLI->emitStoreConditional(Builder, Val: NewValueInsert, |
1422 | Addr: PMV.AlignedAddr, Ord: MemOpOrder); |
1423 | StoreSuccess = Builder.CreateICmpEQ( |
1424 | LHS: StoreSuccess, RHS: ConstantInt::get(Ty: Type::getInt32Ty(C&: Ctx), V: 0), Name: "success" ); |
1425 | BasicBlock *RetryBB = HasReleasedLoadBB ? ReleasedLoadBB : StartBB; |
1426 | Builder.CreateCondBr(Cond: StoreSuccess, True: SuccessBB, |
1427 | False: CI->isWeak() ? FailureBB : RetryBB); |
1428 | |
1429 | Builder.SetInsertPoint(ReleasedLoadBB); |
1430 | Value *SecondLoad; |
1431 | if (HasReleasedLoadBB) { |
1432 | SecondLoad = |
1433 | TLI->emitLoadLinked(Builder, ValueTy: PMV.WordType, Addr: PMV.AlignedAddr, Ord: MemOpOrder); |
1434 | Value * = extractMaskedValue(Builder, WideWord: SecondLoad, PMV); |
1435 | ShouldStore = Builder.CreateICmpEQ(LHS: SecondLoadExtract, |
1436 | RHS: CI->getCompareOperand(), Name: "should_store" ); |
1437 | |
1438 | // If the cmpxchg doesn't actually need any ordering when it fails, we can |
1439 | // jump straight past that fence instruction (if it exists). |
1440 | Builder.CreateCondBr(Cond: ShouldStore, True: TryStoreBB, False: NoStoreBB); |
1441 | // Update PHI node in TryStoreBB. |
1442 | LoadedTryStore->addIncoming(V: SecondLoad, BB: ReleasedLoadBB); |
1443 | } else |
1444 | Builder.CreateUnreachable(); |
1445 | |
1446 | // Make sure later instructions don't get reordered with a fence if |
1447 | // necessary. |
1448 | Builder.SetInsertPoint(SuccessBB); |
1449 | if (ShouldInsertFencesForAtomic || |
1450 | TLI->shouldInsertTrailingFenceForAtomicStore(I: CI)) |
1451 | TLI->emitTrailingFence(Builder, Inst: CI, Ord: SuccessOrder); |
1452 | Builder.CreateBr(Dest: ExitBB); |
1453 | |
1454 | Builder.SetInsertPoint(NoStoreBB); |
1455 | PHINode *LoadedNoStore = |
1456 | Builder.CreatePHI(Ty: UnreleasedLoad->getType(), NumReservedValues: 2, Name: "loaded.nostore" ); |
1457 | LoadedNoStore->addIncoming(V: UnreleasedLoad, BB: StartBB); |
1458 | if (HasReleasedLoadBB) |
1459 | LoadedNoStore->addIncoming(V: SecondLoad, BB: ReleasedLoadBB); |
1460 | |
1461 | // In the failing case, where we don't execute the store-conditional, the |
1462 | // target might want to balance out the load-linked with a dedicated |
1463 | // instruction (e.g., on ARM, clearing the exclusive monitor). |
1464 | TLI->emitAtomicCmpXchgNoStoreLLBalance(Builder); |
1465 | Builder.CreateBr(Dest: FailureBB); |
1466 | |
1467 | Builder.SetInsertPoint(FailureBB); |
1468 | PHINode *LoadedFailure = |
1469 | Builder.CreatePHI(Ty: UnreleasedLoad->getType(), NumReservedValues: 2, Name: "loaded.failure" ); |
1470 | LoadedFailure->addIncoming(V: LoadedNoStore, BB: NoStoreBB); |
1471 | if (CI->isWeak()) |
1472 | LoadedFailure->addIncoming(V: LoadedTryStore, BB: TryStoreBB); |
1473 | if (ShouldInsertFencesForAtomic) |
1474 | TLI->emitTrailingFence(Builder, Inst: CI, Ord: FailureOrder); |
1475 | Builder.CreateBr(Dest: ExitBB); |
1476 | |
1477 | // Finally, we have control-flow based knowledge of whether the cmpxchg |
1478 | // succeeded or not. We expose this to later passes by converting any |
1479 | // subsequent "icmp eq/ne %loaded, %oldval" into a use of an appropriate |
1480 | // PHI. |
1481 | Builder.SetInsertPoint(TheBB: ExitBB, IP: ExitBB->begin()); |
1482 | PHINode *LoadedExit = |
1483 | Builder.CreatePHI(Ty: UnreleasedLoad->getType(), NumReservedValues: 2, Name: "loaded.exit" ); |
1484 | LoadedExit->addIncoming(V: LoadedTryStore, BB: SuccessBB); |
1485 | LoadedExit->addIncoming(V: LoadedFailure, BB: FailureBB); |
1486 | PHINode *Success = Builder.CreatePHI(Ty: Type::getInt1Ty(C&: Ctx), NumReservedValues: 2, Name: "success" ); |
1487 | Success->addIncoming(V: ConstantInt::getTrue(Context&: Ctx), BB: SuccessBB); |
1488 | Success->addIncoming(V: ConstantInt::getFalse(Context&: Ctx), BB: FailureBB); |
1489 | |
1490 | // This is the "exit value" from the cmpxchg expansion. It may be of |
1491 | // a type wider than the one in the cmpxchg instruction. |
1492 | Value *LoadedFull = LoadedExit; |
1493 | |
1494 | Builder.SetInsertPoint(TheBB: ExitBB, IP: std::next(x: Success->getIterator())); |
1495 | Value *Loaded = extractMaskedValue(Builder, WideWord: LoadedFull, PMV); |
1496 | |
1497 | // Look for any users of the cmpxchg that are just comparing the loaded value |
1498 | // against the desired one, and replace them with the CFG-derived version. |
1499 | SmallVector<ExtractValueInst *, 2> PrunedInsts; |
1500 | for (auto *User : CI->users()) { |
1501 | ExtractValueInst *EV = dyn_cast<ExtractValueInst>(Val: User); |
1502 | if (!EV) |
1503 | continue; |
1504 | |
1505 | assert(EV->getNumIndices() == 1 && EV->getIndices()[0] <= 1 && |
1506 | "weird extraction from { iN, i1 }" ); |
1507 | |
1508 | if (EV->getIndices()[0] == 0) |
1509 | EV->replaceAllUsesWith(V: Loaded); |
1510 | else |
1511 | EV->replaceAllUsesWith(V: Success); |
1512 | |
1513 | PrunedInsts.push_back(Elt: EV); |
1514 | } |
1515 | |
1516 | // We can remove the instructions now we're no longer iterating through them. |
1517 | for (auto *EV : PrunedInsts) |
1518 | EV->eraseFromParent(); |
1519 | |
1520 | if (!CI->use_empty()) { |
1521 | // Some use of the full struct return that we don't understand has happened, |
1522 | // so we've got to reconstruct it properly. |
1523 | Value *Res; |
1524 | Res = Builder.CreateInsertValue(Agg: PoisonValue::get(T: CI->getType()), Val: Loaded, Idxs: 0); |
1525 | Res = Builder.CreateInsertValue(Agg: Res, Val: Success, Idxs: 1); |
1526 | |
1527 | CI->replaceAllUsesWith(V: Res); |
1528 | } |
1529 | |
1530 | CI->eraseFromParent(); |
1531 | return true; |
1532 | } |
1533 | |
1534 | bool AtomicExpandImpl::isIdempotentRMW(AtomicRMWInst *RMWI) { |
1535 | auto C = dyn_cast<ConstantInt>(Val: RMWI->getValOperand()); |
1536 | if (!C) |
1537 | return false; |
1538 | |
1539 | AtomicRMWInst::BinOp Op = RMWI->getOperation(); |
1540 | switch (Op) { |
1541 | case AtomicRMWInst::Add: |
1542 | case AtomicRMWInst::Sub: |
1543 | case AtomicRMWInst::Or: |
1544 | case AtomicRMWInst::Xor: |
1545 | return C->isZero(); |
1546 | case AtomicRMWInst::And: |
1547 | return C->isMinusOne(); |
1548 | // FIXME: we could also treat Min/Max/UMin/UMax by the INT_MIN/INT_MAX/... |
1549 | default: |
1550 | return false; |
1551 | } |
1552 | } |
1553 | |
1554 | bool AtomicExpandImpl::simplifyIdempotentRMW(AtomicRMWInst *RMWI) { |
1555 | if (auto ResultingLoad = TLI->lowerIdempotentRMWIntoFencedLoad(RMWI)) { |
1556 | tryExpandAtomicLoad(LI: ResultingLoad); |
1557 | return true; |
1558 | } |
1559 | return false; |
1560 | } |
1561 | |
1562 | Value *AtomicExpandImpl::insertRMWCmpXchgLoop( |
1563 | IRBuilderBase &Builder, Type *ResultTy, Value *Addr, Align AddrAlign, |
1564 | AtomicOrdering MemOpOrder, SyncScope::ID SSID, |
1565 | function_ref<Value *(IRBuilderBase &, Value *)> PerformOp, |
1566 | CreateCmpXchgInstFun CreateCmpXchg) { |
1567 | LLVMContext &Ctx = Builder.getContext(); |
1568 | BasicBlock *BB = Builder.GetInsertBlock(); |
1569 | Function *F = BB->getParent(); |
1570 | |
1571 | // Given: atomicrmw some_op iN* %addr, iN %incr ordering |
1572 | // |
1573 | // The standard expansion we produce is: |
1574 | // [...] |
1575 | // %init_loaded = load atomic iN* %addr |
1576 | // br label %loop |
1577 | // loop: |
1578 | // %loaded = phi iN [ %init_loaded, %entry ], [ %new_loaded, %loop ] |
1579 | // %new = some_op iN %loaded, %incr |
1580 | // %pair = cmpxchg iN* %addr, iN %loaded, iN %new |
1581 | // %new_loaded = extractvalue { iN, i1 } %pair, 0 |
1582 | // %success = extractvalue { iN, i1 } %pair, 1 |
1583 | // br i1 %success, label %atomicrmw.end, label %loop |
1584 | // atomicrmw.end: |
1585 | // [...] |
1586 | BasicBlock *ExitBB = |
1587 | BB->splitBasicBlock(I: Builder.GetInsertPoint(), BBName: "atomicrmw.end" ); |
1588 | BasicBlock *LoopBB = BasicBlock::Create(Context&: Ctx, Name: "atomicrmw.start" , Parent: F, InsertBefore: ExitBB); |
1589 | |
1590 | // The split call above "helpfully" added a branch at the end of BB (to the |
1591 | // wrong place), but we want a load. It's easiest to just remove |
1592 | // the branch entirely. |
1593 | std::prev(x: BB->end())->eraseFromParent(); |
1594 | Builder.SetInsertPoint(BB); |
1595 | LoadInst *InitLoaded = Builder.CreateAlignedLoad(Ty: ResultTy, Ptr: Addr, Align: AddrAlign); |
1596 | Builder.CreateBr(Dest: LoopBB); |
1597 | |
1598 | // Start the main loop block now that we've taken care of the preliminaries. |
1599 | Builder.SetInsertPoint(LoopBB); |
1600 | PHINode *Loaded = Builder.CreatePHI(Ty: ResultTy, NumReservedValues: 2, Name: "loaded" ); |
1601 | Loaded->addIncoming(V: InitLoaded, BB); |
1602 | |
1603 | Value *NewVal = PerformOp(Builder, Loaded); |
1604 | |
1605 | Value *NewLoaded = nullptr; |
1606 | Value *Success = nullptr; |
1607 | |
1608 | CreateCmpXchg(Builder, Addr, Loaded, NewVal, AddrAlign, |
1609 | MemOpOrder == AtomicOrdering::Unordered |
1610 | ? AtomicOrdering::Monotonic |
1611 | : MemOpOrder, |
1612 | SSID, Success, NewLoaded); |
1613 | assert(Success && NewLoaded); |
1614 | |
1615 | Loaded->addIncoming(V: NewLoaded, BB: LoopBB); |
1616 | |
1617 | Builder.CreateCondBr(Cond: Success, True: ExitBB, False: LoopBB); |
1618 | |
1619 | Builder.SetInsertPoint(TheBB: ExitBB, IP: ExitBB->begin()); |
1620 | return NewLoaded; |
1621 | } |
1622 | |
1623 | bool AtomicExpandImpl::tryExpandAtomicCmpXchg(AtomicCmpXchgInst *CI) { |
1624 | unsigned MinCASSize = TLI->getMinCmpXchgSizeInBits() / 8; |
1625 | unsigned ValueSize = getAtomicOpSize(CASI: CI); |
1626 | |
1627 | switch (TLI->shouldExpandAtomicCmpXchgInIR(AI: CI)) { |
1628 | default: |
1629 | llvm_unreachable("Unhandled case in tryExpandAtomicCmpXchg" ); |
1630 | case TargetLoweringBase::AtomicExpansionKind::None: |
1631 | if (ValueSize < MinCASSize) |
1632 | return expandPartwordCmpXchg(CI); |
1633 | return false; |
1634 | case TargetLoweringBase::AtomicExpansionKind::LLSC: { |
1635 | return expandAtomicCmpXchg(CI); |
1636 | } |
1637 | case TargetLoweringBase::AtomicExpansionKind::MaskedIntrinsic: |
1638 | expandAtomicCmpXchgToMaskedIntrinsic(CI); |
1639 | return true; |
1640 | case TargetLoweringBase::AtomicExpansionKind::NotAtomic: |
1641 | return lowerAtomicCmpXchgInst(CXI: CI); |
1642 | } |
1643 | } |
1644 | |
1645 | // Note: This function is exposed externally by AtomicExpandUtils.h |
1646 | bool llvm::expandAtomicRMWToCmpXchg(AtomicRMWInst *AI, |
1647 | CreateCmpXchgInstFun CreateCmpXchg) { |
1648 | ReplacementIRBuilder Builder(AI, AI->getDataLayout()); |
1649 | Builder.setIsFPConstrained( |
1650 | AI->getFunction()->hasFnAttribute(Kind: Attribute::StrictFP)); |
1651 | |
1652 | // FIXME: If FP exceptions are observable, we should force them off for the |
1653 | // loop for the FP atomics. |
1654 | Value *Loaded = AtomicExpandImpl::insertRMWCmpXchgLoop( |
1655 | Builder, ResultTy: AI->getType(), Addr: AI->getPointerOperand(), AddrAlign: AI->getAlign(), |
1656 | MemOpOrder: AI->getOrdering(), SSID: AI->getSyncScopeID(), |
1657 | PerformOp: [&](IRBuilderBase &Builder, Value *Loaded) { |
1658 | return buildAtomicRMWValue(Op: AI->getOperation(), Builder, Loaded, |
1659 | Val: AI->getValOperand()); |
1660 | }, |
1661 | CreateCmpXchg); |
1662 | |
1663 | AI->replaceAllUsesWith(V: Loaded); |
1664 | AI->eraseFromParent(); |
1665 | return true; |
1666 | } |
1667 | |
1668 | // In order to use one of the sized library calls such as |
1669 | // __atomic_fetch_add_4, the alignment must be sufficient, the size |
1670 | // must be one of the potentially-specialized sizes, and the value |
1671 | // type must actually exist in C on the target (otherwise, the |
1672 | // function wouldn't actually be defined.) |
1673 | static bool canUseSizedAtomicCall(unsigned Size, Align Alignment, |
1674 | const DataLayout &DL) { |
1675 | // TODO: "LargestSize" is an approximation for "largest type that |
1676 | // you can express in C". It seems to be the case that int128 is |
1677 | // supported on all 64-bit platforms, otherwise only up to 64-bit |
1678 | // integers are supported. If we get this wrong, then we'll try to |
1679 | // call a sized libcall that doesn't actually exist. There should |
1680 | // really be some more reliable way in LLVM of determining integer |
1681 | // sizes which are valid in the target's C ABI... |
1682 | unsigned LargestSize = DL.getLargestLegalIntTypeSizeInBits() >= 64 ? 16 : 8; |
1683 | return Alignment >= Size && |
1684 | (Size == 1 || Size == 2 || Size == 4 || Size == 8 || Size == 16) && |
1685 | Size <= LargestSize; |
1686 | } |
1687 | |
1688 | void AtomicExpandImpl::expandAtomicLoadToLibcall(LoadInst *I) { |
1689 | static const RTLIB::Libcall Libcalls[6] = { |
1690 | RTLIB::ATOMIC_LOAD, RTLIB::ATOMIC_LOAD_1, RTLIB::ATOMIC_LOAD_2, |
1691 | RTLIB::ATOMIC_LOAD_4, RTLIB::ATOMIC_LOAD_8, RTLIB::ATOMIC_LOAD_16}; |
1692 | unsigned Size = getAtomicOpSize(LI: I); |
1693 | |
1694 | bool expanded = expandAtomicOpToLibcall( |
1695 | I, Size, Alignment: I->getAlign(), PointerOperand: I->getPointerOperand(), ValueOperand: nullptr, CASExpected: nullptr, |
1696 | Ordering: I->getOrdering(), Ordering2: AtomicOrdering::NotAtomic, Libcalls); |
1697 | if (!expanded) |
1698 | report_fatal_error(reason: "expandAtomicOpToLibcall shouldn't fail for Load" ); |
1699 | } |
1700 | |
1701 | void AtomicExpandImpl::expandAtomicStoreToLibcall(StoreInst *I) { |
1702 | static const RTLIB::Libcall Libcalls[6] = { |
1703 | RTLIB::ATOMIC_STORE, RTLIB::ATOMIC_STORE_1, RTLIB::ATOMIC_STORE_2, |
1704 | RTLIB::ATOMIC_STORE_4, RTLIB::ATOMIC_STORE_8, RTLIB::ATOMIC_STORE_16}; |
1705 | unsigned Size = getAtomicOpSize(SI: I); |
1706 | |
1707 | bool expanded = expandAtomicOpToLibcall( |
1708 | I, Size, Alignment: I->getAlign(), PointerOperand: I->getPointerOperand(), ValueOperand: I->getValueOperand(), |
1709 | CASExpected: nullptr, Ordering: I->getOrdering(), Ordering2: AtomicOrdering::NotAtomic, Libcalls); |
1710 | if (!expanded) |
1711 | report_fatal_error(reason: "expandAtomicOpToLibcall shouldn't fail for Store" ); |
1712 | } |
1713 | |
1714 | void AtomicExpandImpl::expandAtomicCASToLibcall(AtomicCmpXchgInst *I) { |
1715 | static const RTLIB::Libcall Libcalls[6] = { |
1716 | RTLIB::ATOMIC_COMPARE_EXCHANGE, RTLIB::ATOMIC_COMPARE_EXCHANGE_1, |
1717 | RTLIB::ATOMIC_COMPARE_EXCHANGE_2, RTLIB::ATOMIC_COMPARE_EXCHANGE_4, |
1718 | RTLIB::ATOMIC_COMPARE_EXCHANGE_8, RTLIB::ATOMIC_COMPARE_EXCHANGE_16}; |
1719 | unsigned Size = getAtomicOpSize(CASI: I); |
1720 | |
1721 | bool expanded = expandAtomicOpToLibcall( |
1722 | I, Size, Alignment: I->getAlign(), PointerOperand: I->getPointerOperand(), ValueOperand: I->getNewValOperand(), |
1723 | CASExpected: I->getCompareOperand(), Ordering: I->getSuccessOrdering(), Ordering2: I->getFailureOrdering(), |
1724 | Libcalls); |
1725 | if (!expanded) |
1726 | report_fatal_error(reason: "expandAtomicOpToLibcall shouldn't fail for CAS" ); |
1727 | } |
1728 | |
1729 | static ArrayRef<RTLIB::Libcall> GetRMWLibcall(AtomicRMWInst::BinOp Op) { |
1730 | static const RTLIB::Libcall LibcallsXchg[6] = { |
1731 | RTLIB::ATOMIC_EXCHANGE, RTLIB::ATOMIC_EXCHANGE_1, |
1732 | RTLIB::ATOMIC_EXCHANGE_2, RTLIB::ATOMIC_EXCHANGE_4, |
1733 | RTLIB::ATOMIC_EXCHANGE_8, RTLIB::ATOMIC_EXCHANGE_16}; |
1734 | static const RTLIB::Libcall LibcallsAdd[6] = { |
1735 | RTLIB::UNKNOWN_LIBCALL, RTLIB::ATOMIC_FETCH_ADD_1, |
1736 | RTLIB::ATOMIC_FETCH_ADD_2, RTLIB::ATOMIC_FETCH_ADD_4, |
1737 | RTLIB::ATOMIC_FETCH_ADD_8, RTLIB::ATOMIC_FETCH_ADD_16}; |
1738 | static const RTLIB::Libcall LibcallsSub[6] = { |
1739 | RTLIB::UNKNOWN_LIBCALL, RTLIB::ATOMIC_FETCH_SUB_1, |
1740 | RTLIB::ATOMIC_FETCH_SUB_2, RTLIB::ATOMIC_FETCH_SUB_4, |
1741 | RTLIB::ATOMIC_FETCH_SUB_8, RTLIB::ATOMIC_FETCH_SUB_16}; |
1742 | static const RTLIB::Libcall LibcallsAnd[6] = { |
1743 | RTLIB::UNKNOWN_LIBCALL, RTLIB::ATOMIC_FETCH_AND_1, |
1744 | RTLIB::ATOMIC_FETCH_AND_2, RTLIB::ATOMIC_FETCH_AND_4, |
1745 | RTLIB::ATOMIC_FETCH_AND_8, RTLIB::ATOMIC_FETCH_AND_16}; |
1746 | static const RTLIB::Libcall LibcallsOr[6] = { |
1747 | RTLIB::UNKNOWN_LIBCALL, RTLIB::ATOMIC_FETCH_OR_1, |
1748 | RTLIB::ATOMIC_FETCH_OR_2, RTLIB::ATOMIC_FETCH_OR_4, |
1749 | RTLIB::ATOMIC_FETCH_OR_8, RTLIB::ATOMIC_FETCH_OR_16}; |
1750 | static const RTLIB::Libcall LibcallsXor[6] = { |
1751 | RTLIB::UNKNOWN_LIBCALL, RTLIB::ATOMIC_FETCH_XOR_1, |
1752 | RTLIB::ATOMIC_FETCH_XOR_2, RTLIB::ATOMIC_FETCH_XOR_4, |
1753 | RTLIB::ATOMIC_FETCH_XOR_8, RTLIB::ATOMIC_FETCH_XOR_16}; |
1754 | static const RTLIB::Libcall LibcallsNand[6] = { |
1755 | RTLIB::UNKNOWN_LIBCALL, RTLIB::ATOMIC_FETCH_NAND_1, |
1756 | RTLIB::ATOMIC_FETCH_NAND_2, RTLIB::ATOMIC_FETCH_NAND_4, |
1757 | RTLIB::ATOMIC_FETCH_NAND_8, RTLIB::ATOMIC_FETCH_NAND_16}; |
1758 | |
1759 | switch (Op) { |
1760 | case AtomicRMWInst::BAD_BINOP: |
1761 | llvm_unreachable("Should not have BAD_BINOP." ); |
1762 | case AtomicRMWInst::Xchg: |
1763 | return ArrayRef(LibcallsXchg); |
1764 | case AtomicRMWInst::Add: |
1765 | return ArrayRef(LibcallsAdd); |
1766 | case AtomicRMWInst::Sub: |
1767 | return ArrayRef(LibcallsSub); |
1768 | case AtomicRMWInst::And: |
1769 | return ArrayRef(LibcallsAnd); |
1770 | case AtomicRMWInst::Or: |
1771 | return ArrayRef(LibcallsOr); |
1772 | case AtomicRMWInst::Xor: |
1773 | return ArrayRef(LibcallsXor); |
1774 | case AtomicRMWInst::Nand: |
1775 | return ArrayRef(LibcallsNand); |
1776 | case AtomicRMWInst::Max: |
1777 | case AtomicRMWInst::Min: |
1778 | case AtomicRMWInst::UMax: |
1779 | case AtomicRMWInst::UMin: |
1780 | case AtomicRMWInst::FMax: |
1781 | case AtomicRMWInst::FMin: |
1782 | case AtomicRMWInst::FAdd: |
1783 | case AtomicRMWInst::FSub: |
1784 | case AtomicRMWInst::UIncWrap: |
1785 | case AtomicRMWInst::UDecWrap: |
1786 | // No atomic libcalls are available for max/min/umax/umin. |
1787 | return {}; |
1788 | } |
1789 | llvm_unreachable("Unexpected AtomicRMW operation." ); |
1790 | } |
1791 | |
1792 | void AtomicExpandImpl::expandAtomicRMWToLibcall(AtomicRMWInst *I) { |
1793 | ArrayRef<RTLIB::Libcall> Libcalls = GetRMWLibcall(Op: I->getOperation()); |
1794 | |
1795 | unsigned Size = getAtomicOpSize(RMWI: I); |
1796 | |
1797 | bool Success = false; |
1798 | if (!Libcalls.empty()) |
1799 | Success = expandAtomicOpToLibcall( |
1800 | I, Size, Alignment: I->getAlign(), PointerOperand: I->getPointerOperand(), ValueOperand: I->getValOperand(), |
1801 | CASExpected: nullptr, Ordering: I->getOrdering(), Ordering2: AtomicOrdering::NotAtomic, Libcalls); |
1802 | |
1803 | // The expansion failed: either there were no libcalls at all for |
1804 | // the operation (min/max), or there were only size-specialized |
1805 | // libcalls (add/sub/etc) and we needed a generic. So, expand to a |
1806 | // CAS libcall, via a CAS loop, instead. |
1807 | if (!Success) { |
1808 | expandAtomicRMWToCmpXchg( |
1809 | AI: I, CreateCmpXchg: [this](IRBuilderBase &Builder, Value *Addr, Value *Loaded, |
1810 | Value *NewVal, Align Alignment, AtomicOrdering MemOpOrder, |
1811 | SyncScope::ID SSID, Value *&Success, Value *&NewLoaded) { |
1812 | // Create the CAS instruction normally... |
1813 | AtomicCmpXchgInst *Pair = Builder.CreateAtomicCmpXchg( |
1814 | Ptr: Addr, Cmp: Loaded, New: NewVal, Align: Alignment, SuccessOrdering: MemOpOrder, |
1815 | FailureOrdering: AtomicCmpXchgInst::getStrongestFailureOrdering(SuccessOrdering: MemOpOrder), SSID); |
1816 | Success = Builder.CreateExtractValue(Agg: Pair, Idxs: 1, Name: "success" ); |
1817 | NewLoaded = Builder.CreateExtractValue(Agg: Pair, Idxs: 0, Name: "newloaded" ); |
1818 | |
1819 | // ...and then expand the CAS into a libcall. |
1820 | expandAtomicCASToLibcall(I: Pair); |
1821 | }); |
1822 | } |
1823 | } |
1824 | |
1825 | // A helper routine for the above expandAtomic*ToLibcall functions. |
1826 | // |
1827 | // 'Libcalls' contains an array of enum values for the particular |
1828 | // ATOMIC libcalls to be emitted. All of the other arguments besides |
1829 | // 'I' are extracted from the Instruction subclass by the |
1830 | // caller. Depending on the particular call, some will be null. |
1831 | bool AtomicExpandImpl::expandAtomicOpToLibcall( |
1832 | Instruction *I, unsigned Size, Align Alignment, Value *PointerOperand, |
1833 | Value *ValueOperand, Value *CASExpected, AtomicOrdering Ordering, |
1834 | AtomicOrdering Ordering2, ArrayRef<RTLIB::Libcall> Libcalls) { |
1835 | assert(Libcalls.size() == 6); |
1836 | |
1837 | LLVMContext &Ctx = I->getContext(); |
1838 | Module *M = I->getModule(); |
1839 | const DataLayout &DL = M->getDataLayout(); |
1840 | IRBuilder<> Builder(I); |
1841 | IRBuilder<> AllocaBuilder(&I->getFunction()->getEntryBlock().front()); |
1842 | |
1843 | bool UseSizedLibcall = canUseSizedAtomicCall(Size, Alignment, DL); |
1844 | Type *SizedIntTy = Type::getIntNTy(C&: Ctx, N: Size * 8); |
1845 | |
1846 | const Align AllocaAlignment = DL.getPrefTypeAlign(Ty: SizedIntTy); |
1847 | |
1848 | // TODO: the "order" argument type is "int", not int32. So |
1849 | // getInt32Ty may be wrong if the arch uses e.g. 16-bit ints. |
1850 | ConstantInt *SizeVal64 = ConstantInt::get(Ty: Type::getInt64Ty(C&: Ctx), V: Size); |
1851 | assert(Ordering != AtomicOrdering::NotAtomic && "expect atomic MO" ); |
1852 | Constant *OrderingVal = |
1853 | ConstantInt::get(Ty: Type::getInt32Ty(C&: Ctx), V: (int)toCABI(AO: Ordering)); |
1854 | Constant *Ordering2Val = nullptr; |
1855 | if (CASExpected) { |
1856 | assert(Ordering2 != AtomicOrdering::NotAtomic && "expect atomic MO" ); |
1857 | Ordering2Val = |
1858 | ConstantInt::get(Ty: Type::getInt32Ty(C&: Ctx), V: (int)toCABI(AO: Ordering2)); |
1859 | } |
1860 | bool HasResult = I->getType() != Type::getVoidTy(C&: Ctx); |
1861 | |
1862 | RTLIB::Libcall RTLibType; |
1863 | if (UseSizedLibcall) { |
1864 | switch (Size) { |
1865 | case 1: |
1866 | RTLibType = Libcalls[1]; |
1867 | break; |
1868 | case 2: |
1869 | RTLibType = Libcalls[2]; |
1870 | break; |
1871 | case 4: |
1872 | RTLibType = Libcalls[3]; |
1873 | break; |
1874 | case 8: |
1875 | RTLibType = Libcalls[4]; |
1876 | break; |
1877 | case 16: |
1878 | RTLibType = Libcalls[5]; |
1879 | break; |
1880 | } |
1881 | } else if (Libcalls[0] != RTLIB::UNKNOWN_LIBCALL) { |
1882 | RTLibType = Libcalls[0]; |
1883 | } else { |
1884 | // Can't use sized function, and there's no generic for this |
1885 | // operation, so give up. |
1886 | return false; |
1887 | } |
1888 | |
1889 | if (!TLI->getLibcallName(Call: RTLibType)) { |
1890 | // This target does not implement the requested atomic libcall so give up. |
1891 | return false; |
1892 | } |
1893 | |
1894 | // Build up the function call. There's two kinds. First, the sized |
1895 | // variants. These calls are going to be one of the following (with |
1896 | // N=1,2,4,8,16): |
1897 | // iN __atomic_load_N(iN *ptr, int ordering) |
1898 | // void __atomic_store_N(iN *ptr, iN val, int ordering) |
1899 | // iN __atomic_{exchange|fetch_*}_N(iN *ptr, iN val, int ordering) |
1900 | // bool __atomic_compare_exchange_N(iN *ptr, iN *expected, iN desired, |
1901 | // int success_order, int failure_order) |
1902 | // |
1903 | // Note that these functions can be used for non-integer atomic |
1904 | // operations, the values just need to be bitcast to integers on the |
1905 | // way in and out. |
1906 | // |
1907 | // And, then, the generic variants. They look like the following: |
1908 | // void __atomic_load(size_t size, void *ptr, void *ret, int ordering) |
1909 | // void __atomic_store(size_t size, void *ptr, void *val, int ordering) |
1910 | // void __atomic_exchange(size_t size, void *ptr, void *val, void *ret, |
1911 | // int ordering) |
1912 | // bool __atomic_compare_exchange(size_t size, void *ptr, void *expected, |
1913 | // void *desired, int success_order, |
1914 | // int failure_order) |
1915 | // |
1916 | // The different signatures are built up depending on the |
1917 | // 'UseSizedLibcall', 'CASExpected', 'ValueOperand', and 'HasResult' |
1918 | // variables. |
1919 | |
1920 | AllocaInst *AllocaCASExpected = nullptr; |
1921 | AllocaInst *AllocaValue = nullptr; |
1922 | AllocaInst *AllocaResult = nullptr; |
1923 | |
1924 | Type *ResultTy; |
1925 | SmallVector<Value *, 6> Args; |
1926 | AttributeList Attr; |
1927 | |
1928 | // 'size' argument. |
1929 | if (!UseSizedLibcall) { |
1930 | // Note, getIntPtrType is assumed equivalent to size_t. |
1931 | Args.push_back(Elt: ConstantInt::get(Ty: DL.getIntPtrType(C&: Ctx), V: Size)); |
1932 | } |
1933 | |
1934 | // 'ptr' argument. |
1935 | // note: This assumes all address spaces share a common libfunc |
1936 | // implementation and that addresses are convertable. For systems without |
1937 | // that property, we'd need to extend this mechanism to support AS-specific |
1938 | // families of atomic intrinsics. |
1939 | Value *PtrVal = PointerOperand; |
1940 | PtrVal = Builder.CreateAddrSpaceCast(V: PtrVal, DestTy: PointerType::getUnqual(C&: Ctx)); |
1941 | Args.push_back(Elt: PtrVal); |
1942 | |
1943 | // 'expected' argument, if present. |
1944 | if (CASExpected) { |
1945 | AllocaCASExpected = AllocaBuilder.CreateAlloca(Ty: CASExpected->getType()); |
1946 | AllocaCASExpected->setAlignment(AllocaAlignment); |
1947 | Builder.CreateLifetimeStart(Ptr: AllocaCASExpected, Size: SizeVal64); |
1948 | Builder.CreateAlignedStore(Val: CASExpected, Ptr: AllocaCASExpected, Align: AllocaAlignment); |
1949 | Args.push_back(Elt: AllocaCASExpected); |
1950 | } |
1951 | |
1952 | // 'val' argument ('desired' for cas), if present. |
1953 | if (ValueOperand) { |
1954 | if (UseSizedLibcall) { |
1955 | Value *IntValue = |
1956 | Builder.CreateBitOrPointerCast(V: ValueOperand, DestTy: SizedIntTy); |
1957 | Args.push_back(Elt: IntValue); |
1958 | } else { |
1959 | AllocaValue = AllocaBuilder.CreateAlloca(Ty: ValueOperand->getType()); |
1960 | AllocaValue->setAlignment(AllocaAlignment); |
1961 | Builder.CreateLifetimeStart(Ptr: AllocaValue, Size: SizeVal64); |
1962 | Builder.CreateAlignedStore(Val: ValueOperand, Ptr: AllocaValue, Align: AllocaAlignment); |
1963 | Args.push_back(Elt: AllocaValue); |
1964 | } |
1965 | } |
1966 | |
1967 | // 'ret' argument. |
1968 | if (!CASExpected && HasResult && !UseSizedLibcall) { |
1969 | AllocaResult = AllocaBuilder.CreateAlloca(Ty: I->getType()); |
1970 | AllocaResult->setAlignment(AllocaAlignment); |
1971 | Builder.CreateLifetimeStart(Ptr: AllocaResult, Size: SizeVal64); |
1972 | Args.push_back(Elt: AllocaResult); |
1973 | } |
1974 | |
1975 | // 'ordering' ('success_order' for cas) argument. |
1976 | Args.push_back(Elt: OrderingVal); |
1977 | |
1978 | // 'failure_order' argument, if present. |
1979 | if (Ordering2Val) |
1980 | Args.push_back(Elt: Ordering2Val); |
1981 | |
1982 | // Now, the return type. |
1983 | if (CASExpected) { |
1984 | ResultTy = Type::getInt1Ty(C&: Ctx); |
1985 | Attr = Attr.addRetAttribute(C&: Ctx, Kind: Attribute::ZExt); |
1986 | } else if (HasResult && UseSizedLibcall) |
1987 | ResultTy = SizedIntTy; |
1988 | else |
1989 | ResultTy = Type::getVoidTy(C&: Ctx); |
1990 | |
1991 | // Done with setting up arguments and return types, create the call: |
1992 | SmallVector<Type *, 6> ArgTys; |
1993 | for (Value *Arg : Args) |
1994 | ArgTys.push_back(Elt: Arg->getType()); |
1995 | FunctionType *FnType = FunctionType::get(Result: ResultTy, Params: ArgTys, isVarArg: false); |
1996 | FunctionCallee LibcallFn = |
1997 | M->getOrInsertFunction(Name: TLI->getLibcallName(Call: RTLibType), T: FnType, AttributeList: Attr); |
1998 | CallInst *Call = Builder.CreateCall(Callee: LibcallFn, Args); |
1999 | Call->setAttributes(Attr); |
2000 | Value *Result = Call; |
2001 | |
2002 | // And then, extract the results... |
2003 | if (ValueOperand && !UseSizedLibcall) |
2004 | Builder.CreateLifetimeEnd(Ptr: AllocaValue, Size: SizeVal64); |
2005 | |
2006 | if (CASExpected) { |
2007 | // The final result from the CAS is {load of 'expected' alloca, bool result |
2008 | // from call} |
2009 | Type *FinalResultTy = I->getType(); |
2010 | Value *V = PoisonValue::get(T: FinalResultTy); |
2011 | Value *ExpectedOut = Builder.CreateAlignedLoad( |
2012 | Ty: CASExpected->getType(), Ptr: AllocaCASExpected, Align: AllocaAlignment); |
2013 | Builder.CreateLifetimeEnd(Ptr: AllocaCASExpected, Size: SizeVal64); |
2014 | V = Builder.CreateInsertValue(Agg: V, Val: ExpectedOut, Idxs: 0); |
2015 | V = Builder.CreateInsertValue(Agg: V, Val: Result, Idxs: 1); |
2016 | I->replaceAllUsesWith(V); |
2017 | } else if (HasResult) { |
2018 | Value *V; |
2019 | if (UseSizedLibcall) |
2020 | V = Builder.CreateBitOrPointerCast(V: Result, DestTy: I->getType()); |
2021 | else { |
2022 | V = Builder.CreateAlignedLoad(Ty: I->getType(), Ptr: AllocaResult, |
2023 | Align: AllocaAlignment); |
2024 | Builder.CreateLifetimeEnd(Ptr: AllocaResult, Size: SizeVal64); |
2025 | } |
2026 | I->replaceAllUsesWith(V); |
2027 | } |
2028 | I->eraseFromParent(); |
2029 | return true; |
2030 | } |
2031 | |