1 //===-- AtomicExpandPass.cpp - Expand atomic instructions -------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file contains a pass (at IR level) to replace atomic instructions with
11 // target specific instruction which implement the same semantics in a way
12 // which better fits the target backend. This can include the use of either
13 // (intrinsic-based) load-linked/store-conditional loops, AtomicCmpXchg, or
16 //===----------------------------------------------------------------------===//
18 #include "llvm/ADT/SetOperations.h"
19 #include "llvm/ADT/SetVector.h"
20 #include "llvm/ADT/SmallPtrSet.h"
21 #include "llvm/ADT/SmallVector.h"
22 #include "llvm/Analysis/MemoryLocation.h"
23 #include "llvm/CodeGen/AtomicExpandUtils.h"
24 #include "llvm/CodeGen/Passes.h"
25 #include "llvm/IR/Function.h"
26 #include "llvm/IR/IRBuilder.h"
27 #include "llvm/IR/InstIterator.h"
28 #include "llvm/IR/Instructions.h"
29 #include "llvm/IR/Intrinsics.h"
30 #include "llvm/IR/Module.h"
31 #include "llvm/IR/NoFolder.h"
32 #include "llvm/Support/Debug.h"
33 #include "llvm/Support/raw_ostream.h"
34 #include "llvm/Target/TargetLowering.h"
35 #include "llvm/Target/TargetMachine.h"
36 #include "llvm/Target/TargetSubtargetInfo.h"
37 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
41 #define DEBUG_TYPE "atomic-expand"
44 class AtomicExpand: public FunctionPass {
45 const TargetMachine *TM;
46 const TargetLowering *TLI;
48 static char ID; // Pass identification, replacement for typeid
49 explicit AtomicExpand(const TargetMachine *TM = nullptr)
50 : FunctionPass(ID), TM(TM), TLI(nullptr) {
51 initializeAtomicExpandPass(*PassRegistry::getPassRegistry());
54 bool runOnFunction(Function &F) override;
57 bool bracketInstWithFences(Instruction *I, AtomicOrdering Order,
58 bool IsStore, bool IsLoad);
59 IntegerType *getCorrespondingIntegerType(Type *T, const DataLayout &DL);
60 LoadInst *convertAtomicLoadToIntegerType(LoadInst *LI);
61 bool tryExpandAtomicLoad(LoadInst *LI);
62 bool expandAtomicLoadToLL(LoadInst *LI);
63 bool expandAtomicLoadToCmpXchg(LoadInst *LI);
64 StoreInst *convertAtomicStoreToIntegerType(StoreInst *SI);
65 bool expandAtomicStore(StoreInst *SI);
66 bool tryExpandAtomicRMW(AtomicRMWInst *AI);
67 bool expandAtomicOpToLLSC(
68 Instruction *I, Value *Addr, AtomicOrdering MemOpOrder,
69 std::function<Value *(IRBuilder<> &, Value *)> PerformOp);
70 bool expandAtomicCmpXchg(AtomicCmpXchgInst *CI);
71 bool isIdempotentRMW(AtomicRMWInst *AI);
72 bool simplifyIdempotentRMW(AtomicRMWInst *AI);
76 // If 'LI' is a relaxed load, and it is immediately followed by a
77 // atomic read-modify-write that has acq_rel parameter, we don't have to do
78 // anything since the rmw serves as a natural barrier.
79 void MarkRelaxedLoadBeforeAcqrelRMW(LoadInst* LI) {
80 auto* BB = LI->getParent();
81 auto BBI = LI->getIterator();
82 for (BBI++; BBI != BB->end(); BBI++) {
83 Instruction* CurInst = &*BBI;
87 if (!CurInst->isAtomic()) {
90 auto* RMW = dyn_cast<AtomicRMWInst>(CurInst);
94 if (RMW->getOrdering() == AcquireRelease ||
95 RMW->getOrdering() == SequentiallyConsistent) {
96 LI->setHasSubsequentAcqlRMW(true);
103 char AtomicExpand::ID = 0;
104 char &llvm::AtomicExpandID = AtomicExpand::ID;
105 INITIALIZE_TM_PASS(AtomicExpand, "atomic-expand",
106 "Expand Atomic calls in terms of either load-linked & store-conditional or cmpxchg",
109 FunctionPass *llvm::createAtomicExpandPass(const TargetMachine *TM) {
110 return new AtomicExpand(TM);
113 bool AtomicExpand::runOnFunction(Function &F) {
114 if (!TM || !TM->getSubtargetImpl(F)->enableAtomicExpand())
116 TLI = TM->getSubtargetImpl(F)->getTargetLowering();
118 SmallVector<Instruction *, 1> AtomicInsts;
119 SmallVector<LoadInst*, 1> MonotonicLoadInsts;
121 // Changing control-flow while iterating through it is a bad idea, so gather a
122 // list of all atomic instructions before we start.
123 for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I) {
124 // XXX-update: For relaxed loads, change them to acquire. This includes
125 // relaxed loads, relaxed atomic RMW & relaxed atomic compare exchange.
127 switch (I->getOpcode()) {
128 case Instruction::AtomicCmpXchg: {
129 // XXX-comment: AtomicCmpXchg in AArch64 will be translated to a
130 // conditional branch that contains the value of the load anyway, so
131 // we don't need to do anything.
133 auto* CmpXchg = dyn_cast<AtomicCmpXchgInst>(&*I);
134 auto SuccOrdering = CmpXchg->getSuccessOrdering();
135 if (SuccOrdering == Monotonic) {
136 CmpXchg->setSuccessOrdering(Acquire);
137 } else if (SuccOrdering == Release) {
138 CmpXchg->setSuccessOrdering(AcquireRelease);
143 case Instruction::AtomicRMW: {
144 // XXX-comment: Similar to AtomicCmpXchg. These instructions in
145 // AArch64 will be translated to a loop whose condition depends on the
146 // store status, which further depends on the load value.
148 auto* RMW = dyn_cast<AtomicRMWInst>(&*I);
149 if (RMW->getOrdering() == Monotonic) {
150 RMW->setOrdering(Acquire);
155 case Instruction::Load: {
156 auto* LI = dyn_cast<LoadInst>(&*I);
157 if (LI->getOrdering() == Monotonic) {
159 DEBUG(dbgs() << "Transforming relaxed loads to acquire loads: "
161 LI->setOrdering(Acquire);
163 // MonotonicLoadInsts.push_back(LI);
164 MarkRelaxedLoadBeforeAcqrelRMW(LI);
172 AtomicInsts.push_back(&*I);
176 bool MadeChange = false;
177 for (auto I : AtomicInsts) {
178 auto LI = dyn_cast<LoadInst>(I);
179 auto SI = dyn_cast<StoreInst>(I);
180 auto RMWI = dyn_cast<AtomicRMWInst>(I);
181 auto CASI = dyn_cast<AtomicCmpXchgInst>(I);
182 assert((LI || SI || RMWI || CASI || isa<FenceInst>(I)) &&
183 "Unknown atomic instruction");
185 auto FenceOrdering = Monotonic;
186 bool IsStore, IsLoad;
187 if (TLI->getInsertFencesForAtomic()) {
188 if (LI && isAtLeastAcquire(LI->getOrdering())) {
189 FenceOrdering = LI->getOrdering();
190 // AddFakeConditionalBranch(
193 } else if (SI && isAtLeastRelease(SI->getOrdering())) {
194 FenceOrdering = SI->getOrdering();
195 SI->setOrdering(Monotonic);
198 } else if (RMWI && (isAtLeastRelease(RMWI->getOrdering()) ||
199 isAtLeastAcquire(RMWI->getOrdering()))) {
200 FenceOrdering = RMWI->getOrdering();
201 RMWI->setOrdering(Monotonic);
202 IsStore = IsLoad = true;
203 } else if (CASI && !TLI->shouldExpandAtomicCmpXchgInIR(CASI) &&
204 (isAtLeastRelease(CASI->getSuccessOrdering()) ||
205 isAtLeastAcquire(CASI->getSuccessOrdering()))) {
206 // If a compare and swap is lowered to LL/SC, we can do smarter fence
207 // insertion, with a stronger one on the success path than on the
208 // failure path. As a result, fence insertion is directly done by
209 // expandAtomicCmpXchg in that case.
210 FenceOrdering = CASI->getSuccessOrdering();
211 CASI->setSuccessOrdering(Monotonic);
212 CASI->setFailureOrdering(Monotonic);
213 IsStore = IsLoad = true;
216 if (FenceOrdering != Monotonic) {
217 MadeChange |= bracketInstWithFences(I, FenceOrdering, IsStore, IsLoad);
222 if (LI->getType()->isFloatingPointTy()) {
223 // TODO: add a TLI hook to control this so that each target can
224 // convert to lowering the original type one at a time.
225 LI = convertAtomicLoadToIntegerType(LI);
226 assert(LI->getType()->isIntegerTy() && "invariant broken");
230 MadeChange |= tryExpandAtomicLoad(LI);
232 if (SI->getValueOperand()->getType()->isFloatingPointTy()) {
233 // TODO: add a TLI hook to control this so that each target can
234 // convert to lowering the original type one at a time.
235 SI = convertAtomicStoreToIntegerType(SI);
236 assert(SI->getValueOperand()->getType()->isIntegerTy() &&
241 if (TLI->shouldExpandAtomicStoreInIR(SI))
242 MadeChange |= expandAtomicStore(SI);
244 // There are two different ways of expanding RMW instructions:
245 // - into a load if it is idempotent
246 // - into a Cmpxchg/LL-SC loop otherwise
247 // we try them in that order.
249 if (isIdempotentRMW(RMWI) && simplifyIdempotentRMW(RMWI)) {
252 MadeChange |= tryExpandAtomicRMW(RMWI);
254 } else if (CASI && TLI->shouldExpandAtomicCmpXchgInIR(CASI)) {
255 MadeChange |= expandAtomicCmpXchg(CASI);
262 bool AtomicExpand::bracketInstWithFences(Instruction *I, AtomicOrdering Order,
263 bool IsStore, bool IsLoad) {
264 IRBuilder<> Builder(I);
266 auto LeadingFence = TLI->emitLeadingFence(Builder, Order, IsStore, IsLoad);
268 auto TrailingFence = TLI->emitTrailingFence(Builder, Order, IsStore, IsLoad);
269 // The trailing fence is emitted before the instruction instead of after
270 // because there is no easy way of setting Builder insertion point after
271 // an instruction. So we must erase it from the BB, and insert it back
272 // in the right place.
273 // We have a guard here because not every atomic operation generates a
276 TrailingFence->removeFromParent();
277 TrailingFence->insertAfter(I);
280 return (LeadingFence || TrailingFence);
283 /// Get the iX type with the same bitwidth as T.
284 IntegerType *AtomicExpand::getCorrespondingIntegerType(Type *T,
285 const DataLayout &DL) {
286 EVT VT = TLI->getValueType(DL, T);
287 unsigned BitWidth = VT.getStoreSizeInBits();
288 assert(BitWidth == VT.getSizeInBits() && "must be a power of two");
289 return IntegerType::get(T->getContext(), BitWidth);
292 /// Convert an atomic load of a non-integral type to an integer load of the
293 /// equivelent bitwidth. See the function comment on
294 /// convertAtomicStoreToIntegerType for background.
295 LoadInst *AtomicExpand::convertAtomicLoadToIntegerType(LoadInst *LI) {
296 auto *M = LI->getModule();
297 Type *NewTy = getCorrespondingIntegerType(LI->getType(),
300 IRBuilder<> Builder(LI);
302 Value *Addr = LI->getPointerOperand();
303 Type *PT = PointerType::get(NewTy,
304 Addr->getType()->getPointerAddressSpace());
305 Value *NewAddr = Builder.CreateBitCast(Addr, PT);
307 auto *NewLI = Builder.CreateLoad(NewAddr);
308 NewLI->setAlignment(LI->getAlignment());
309 NewLI->setVolatile(LI->isVolatile());
310 NewLI->setAtomic(LI->getOrdering(), LI->getSynchScope());
311 DEBUG(dbgs() << "Replaced " << *LI << " with " << *NewLI << "\n");
313 Value *NewVal = Builder.CreateBitCast(NewLI, LI->getType());
314 LI->replaceAllUsesWith(NewVal);
315 LI->eraseFromParent();
319 bool AtomicExpand::tryExpandAtomicLoad(LoadInst *LI) {
320 switch (TLI->shouldExpandAtomicLoadInIR(LI)) {
321 case TargetLoweringBase::AtomicExpansionKind::None:
323 case TargetLoweringBase::AtomicExpansionKind::LLSC:
324 return expandAtomicOpToLLSC(
325 LI, LI->getPointerOperand(), LI->getOrdering(),
326 [](IRBuilder<> &Builder, Value *Loaded) { return Loaded; });
327 case TargetLoweringBase::AtomicExpansionKind::LLOnly:
328 return expandAtomicLoadToLL(LI);
329 case TargetLoweringBase::AtomicExpansionKind::CmpXChg:
330 return expandAtomicLoadToCmpXchg(LI);
332 llvm_unreachable("Unhandled case in tryExpandAtomicLoad");
335 bool AtomicExpand::expandAtomicLoadToLL(LoadInst *LI) {
336 IRBuilder<> Builder(LI);
338 // On some architectures, load-linked instructions are atomic for larger
339 // sizes than normal loads. For example, the only 64-bit load guaranteed
340 // to be single-copy atomic by ARM is an ldrexd (A3.5.3).
342 TLI->emitLoadLinked(Builder, LI->getPointerOperand(), LI->getOrdering());
343 TLI->emitAtomicCmpXchgNoStoreLLBalance(Builder);
345 LI->replaceAllUsesWith(Val);
346 LI->eraseFromParent();
351 bool AtomicExpand::expandAtomicLoadToCmpXchg(LoadInst *LI) {
352 IRBuilder<> Builder(LI);
353 AtomicOrdering Order = LI->getOrdering();
354 Value *Addr = LI->getPointerOperand();
355 Type *Ty = cast<PointerType>(Addr->getType())->getElementType();
356 Constant *DummyVal = Constant::getNullValue(Ty);
358 Value *Pair = Builder.CreateAtomicCmpXchg(
359 Addr, DummyVal, DummyVal, Order,
360 AtomicCmpXchgInst::getStrongestFailureOrdering(Order));
361 Value *Loaded = Builder.CreateExtractValue(Pair, 0, "loaded");
363 LI->replaceAllUsesWith(Loaded);
364 LI->eraseFromParent();
369 /// Convert an atomic store of a non-integral type to an integer store of the
370 /// equivelent bitwidth. We used to not support floating point or vector
371 /// atomics in the IR at all. The backends learned to deal with the bitcast
372 /// idiom because that was the only way of expressing the notion of a atomic
373 /// float or vector store. The long term plan is to teach each backend to
374 /// instruction select from the original atomic store, but as a migration
375 /// mechanism, we convert back to the old format which the backends understand.
376 /// Each backend will need individual work to recognize the new format.
377 StoreInst *AtomicExpand::convertAtomicStoreToIntegerType(StoreInst *SI) {
378 IRBuilder<> Builder(SI);
379 auto *M = SI->getModule();
380 Type *NewTy = getCorrespondingIntegerType(SI->getValueOperand()->getType(),
382 Value *NewVal = Builder.CreateBitCast(SI->getValueOperand(), NewTy);
384 Value *Addr = SI->getPointerOperand();
385 Type *PT = PointerType::get(NewTy,
386 Addr->getType()->getPointerAddressSpace());
387 Value *NewAddr = Builder.CreateBitCast(Addr, PT);
389 StoreInst *NewSI = Builder.CreateStore(NewVal, NewAddr);
390 NewSI->setAlignment(SI->getAlignment());
391 NewSI->setVolatile(SI->isVolatile());
392 NewSI->setAtomic(SI->getOrdering(), SI->getSynchScope());
393 DEBUG(dbgs() << "Replaced " << *SI << " with " << *NewSI << "\n");
394 SI->eraseFromParent();
398 bool AtomicExpand::expandAtomicStore(StoreInst *SI) {
399 // This function is only called on atomic stores that are too large to be
400 // atomic if implemented as a native store. So we replace them by an
401 // atomic swap, that can be implemented for example as a ldrex/strex on ARM
402 // or lock cmpxchg8/16b on X86, as these are atomic for larger sizes.
403 // It is the responsibility of the target to only signal expansion via
404 // shouldExpandAtomicRMW in cases where this is required and possible.
405 IRBuilder<> Builder(SI);
407 Builder.CreateAtomicRMW(AtomicRMWInst::Xchg, SI->getPointerOperand(),
408 SI->getValueOperand(), SI->getOrdering());
409 SI->eraseFromParent();
411 // Now we have an appropriate swap instruction, lower it as usual.
412 return tryExpandAtomicRMW(AI);
415 static void createCmpXchgInstFun(IRBuilder<> &Builder, Value *Addr,
416 Value *Loaded, Value *NewVal,
417 AtomicOrdering MemOpOrder,
418 Value *&Success, Value *&NewLoaded) {
419 Value* Pair = Builder.CreateAtomicCmpXchg(
420 Addr, Loaded, NewVal, MemOpOrder,
421 AtomicCmpXchgInst::getStrongestFailureOrdering(MemOpOrder));
422 Success = Builder.CreateExtractValue(Pair, 1, "success");
423 NewLoaded = Builder.CreateExtractValue(Pair, 0, "newloaded");
426 /// Emit IR to implement the given atomicrmw operation on values in registers,
427 /// returning the new value.
428 static Value *performAtomicOp(AtomicRMWInst::BinOp Op, IRBuilder<> &Builder,
429 Value *Loaded, Value *Inc) {
432 case AtomicRMWInst::Xchg:
434 case AtomicRMWInst::Add:
435 return Builder.CreateAdd(Loaded, Inc, "new");
436 case AtomicRMWInst::Sub:
437 return Builder.CreateSub(Loaded, Inc, "new");
438 case AtomicRMWInst::And:
439 return Builder.CreateAnd(Loaded, Inc, "new");
440 case AtomicRMWInst::Nand:
441 return Builder.CreateNot(Builder.CreateAnd(Loaded, Inc), "new");
442 case AtomicRMWInst::Or:
443 return Builder.CreateOr(Loaded, Inc, "new");
444 case AtomicRMWInst::Xor:
445 return Builder.CreateXor(Loaded, Inc, "new");
446 case AtomicRMWInst::Max:
447 NewVal = Builder.CreateICmpSGT(Loaded, Inc);
448 return Builder.CreateSelect(NewVal, Loaded, Inc, "new");
449 case AtomicRMWInst::Min:
450 NewVal = Builder.CreateICmpSLE(Loaded, Inc);
451 return Builder.CreateSelect(NewVal, Loaded, Inc, "new");
452 case AtomicRMWInst::UMax:
453 NewVal = Builder.CreateICmpUGT(Loaded, Inc);
454 return Builder.CreateSelect(NewVal, Loaded, Inc, "new");
455 case AtomicRMWInst::UMin:
456 NewVal = Builder.CreateICmpULE(Loaded, Inc);
457 return Builder.CreateSelect(NewVal, Loaded, Inc, "new");
459 llvm_unreachable("Unknown atomic op");
463 bool AtomicExpand::tryExpandAtomicRMW(AtomicRMWInst *AI) {
464 switch (TLI->shouldExpandAtomicRMWInIR(AI)) {
465 case TargetLoweringBase::AtomicExpansionKind::None:
467 case TargetLoweringBase::AtomicExpansionKind::LLSC:
468 return expandAtomicOpToLLSC(AI, AI->getPointerOperand(), AI->getOrdering(),
469 [&](IRBuilder<> &Builder, Value *Loaded) {
470 return performAtomicOp(AI->getOperation(),
472 AI->getValOperand());
474 case TargetLoweringBase::AtomicExpansionKind::CmpXChg:
475 return expandAtomicRMWToCmpXchg(AI, createCmpXchgInstFun);
477 llvm_unreachable("Unhandled case in tryExpandAtomicRMW");
481 bool AtomicExpand::expandAtomicOpToLLSC(
482 Instruction *I, Value *Addr, AtomicOrdering MemOpOrder,
483 std::function<Value *(IRBuilder<> &, Value *)> PerformOp) {
484 BasicBlock *BB = I->getParent();
485 Function *F = BB->getParent();
486 LLVMContext &Ctx = F->getContext();
488 // Given: atomicrmw some_op iN* %addr, iN %incr ordering
490 // The standard expansion we produce is:
494 // %loaded = @load.linked(%addr)
495 // %new = some_op iN %loaded, %incr
496 // %stored = @store_conditional(%new, %addr)
497 // %try_again = icmp i32 ne %stored, 0
498 // br i1 %try_again, label %loop, label %atomicrmw.end
502 BasicBlock *ExitBB = BB->splitBasicBlock(I->getIterator(), "atomicrmw.end");
503 BasicBlock *LoopBB = BasicBlock::Create(Ctx, "atomicrmw.start", F, ExitBB);
505 // This grabs the DebugLoc from I.
506 IRBuilder<> Builder(I);
508 // The split call above "helpfully" added a branch at the end of BB (to the
509 // wrong place), but we might want a fence too. It's easiest to just remove
510 // the branch entirely.
511 std::prev(BB->end())->eraseFromParent();
512 Builder.SetInsertPoint(BB);
513 Builder.CreateBr(LoopBB);
515 // Start the main loop block now that we've taken care of the preliminaries.
516 Builder.SetInsertPoint(LoopBB);
517 Value *Loaded = TLI->emitLoadLinked(Builder, Addr, MemOpOrder);
519 Value *NewVal = PerformOp(Builder, Loaded);
521 Value *StoreSuccess =
522 TLI->emitStoreConditional(Builder, NewVal, Addr, MemOpOrder);
523 Value *TryAgain = Builder.CreateICmpNE(
524 StoreSuccess, ConstantInt::get(IntegerType::get(Ctx, 32), 0), "tryagain");
525 Builder.CreateCondBr(TryAgain, LoopBB, ExitBB);
527 Builder.SetInsertPoint(ExitBB, ExitBB->begin());
529 I->replaceAllUsesWith(Loaded);
530 I->eraseFromParent();
535 bool AtomicExpand::expandAtomicCmpXchg(AtomicCmpXchgInst *CI) {
536 AtomicOrdering SuccessOrder = CI->getSuccessOrdering();
537 AtomicOrdering FailureOrder = CI->getFailureOrdering();
538 Value *Addr = CI->getPointerOperand();
539 BasicBlock *BB = CI->getParent();
540 Function *F = BB->getParent();
541 LLVMContext &Ctx = F->getContext();
542 // If getInsertFencesForAtomic() returns true, then the target does not want
543 // to deal with memory orders, and emitLeading/TrailingFence should take care
544 // of everything. Otherwise, emitLeading/TrailingFence are no-op and we
545 // should preserve the ordering.
546 AtomicOrdering MemOpOrder =
547 TLI->getInsertFencesForAtomic() ? Monotonic : SuccessOrder;
549 // Given: cmpxchg some_op iN* %addr, iN %desired, iN %new success_ord fail_ord
551 // The full expansion we produce is:
555 // %loaded = @load.linked(%addr)
556 // %should_store = icmp eq %loaded, %desired
557 // br i1 %should_store, label %cmpxchg.trystore,
558 // label %cmpxchg.nostore
560 // %stored = @store_conditional(%new, %addr)
561 // %success = icmp eq i32 %stored, 0
562 // br i1 %success, label %cmpxchg.success, label %loop/%cmpxchg.failure
565 // br label %cmpxchg.end
567 // @load_linked_fail_balance()?
568 // br label %cmpxchg.failure
571 // br label %cmpxchg.end
573 // %success = phi i1 [true, %cmpxchg.success], [false, %cmpxchg.failure]
574 // %restmp = insertvalue { iN, i1 } undef, iN %loaded, 0
575 // %res = insertvalue { iN, i1 } %restmp, i1 %success, 1
577 BasicBlock *ExitBB = BB->splitBasicBlock(CI->getIterator(), "cmpxchg.end");
578 auto FailureBB = BasicBlock::Create(Ctx, "cmpxchg.failure", F, ExitBB);
579 auto NoStoreBB = BasicBlock::Create(Ctx, "cmpxchg.nostore", F, FailureBB);
580 auto SuccessBB = BasicBlock::Create(Ctx, "cmpxchg.success", F, NoStoreBB);
581 auto TryStoreBB = BasicBlock::Create(Ctx, "cmpxchg.trystore", F, SuccessBB);
582 auto LoopBB = BasicBlock::Create(Ctx, "cmpxchg.start", F, TryStoreBB);
584 // This grabs the DebugLoc from CI
585 IRBuilder<> Builder(CI);
587 // The split call above "helpfully" added a branch at the end of BB (to the
588 // wrong place), but we might want a fence too. It's easiest to just remove
589 // the branch entirely.
590 std::prev(BB->end())->eraseFromParent();
591 Builder.SetInsertPoint(BB);
592 TLI->emitLeadingFence(Builder, SuccessOrder, /*IsStore=*/true,
594 Builder.CreateBr(LoopBB);
596 // Start the main loop block now that we've taken care of the preliminaries.
597 Builder.SetInsertPoint(LoopBB);
598 Value *Loaded = TLI->emitLoadLinked(Builder, Addr, MemOpOrder);
600 Builder.CreateICmpEQ(Loaded, CI->getCompareOperand(), "should_store");
602 // If the cmpxchg doesn't actually need any ordering when it fails, we can
603 // jump straight past that fence instruction (if it exists).
604 Builder.CreateCondBr(ShouldStore, TryStoreBB, NoStoreBB);
606 Builder.SetInsertPoint(TryStoreBB);
607 Value *StoreSuccess = TLI->emitStoreConditional(
608 Builder, CI->getNewValOperand(), Addr, MemOpOrder);
609 StoreSuccess = Builder.CreateICmpEQ(
610 StoreSuccess, ConstantInt::get(Type::getInt32Ty(Ctx), 0), "success");
611 Builder.CreateCondBr(StoreSuccess, SuccessBB,
612 CI->isWeak() ? FailureBB : LoopBB);
614 // Make sure later instructions don't get reordered with a fence if necessary.
615 Builder.SetInsertPoint(SuccessBB);
616 TLI->emitTrailingFence(Builder, SuccessOrder, /*IsStore=*/true,
618 Builder.CreateBr(ExitBB);
620 Builder.SetInsertPoint(NoStoreBB);
621 // In the failing case, where we don't execute the store-conditional, the
622 // target might want to balance out the load-linked with a dedicated
623 // instruction (e.g., on ARM, clearing the exclusive monitor).
624 TLI->emitAtomicCmpXchgNoStoreLLBalance(Builder);
625 Builder.CreateBr(FailureBB);
627 Builder.SetInsertPoint(FailureBB);
628 TLI->emitTrailingFence(Builder, FailureOrder, /*IsStore=*/true,
630 Builder.CreateBr(ExitBB);
632 // Finally, we have control-flow based knowledge of whether the cmpxchg
633 // succeeded or not. We expose this to later passes by converting any
634 // subsequent "icmp eq/ne %loaded, %oldval" into a use of an appropriate PHI.
636 // Setup the builder so we can create any PHIs we need.
637 Builder.SetInsertPoint(ExitBB, ExitBB->begin());
638 PHINode *Success = Builder.CreatePHI(Type::getInt1Ty(Ctx), 2);
639 Success->addIncoming(ConstantInt::getTrue(Ctx), SuccessBB);
640 Success->addIncoming(ConstantInt::getFalse(Ctx), FailureBB);
642 // Look for any users of the cmpxchg that are just comparing the loaded value
643 // against the desired one, and replace them with the CFG-derived version.
644 SmallVector<ExtractValueInst *, 2> PrunedInsts;
645 for (auto User : CI->users()) {
646 ExtractValueInst *EV = dyn_cast<ExtractValueInst>(User);
650 assert(EV->getNumIndices() == 1 && EV->getIndices()[0] <= 1 &&
651 "weird extraction from { iN, i1 }");
653 if (EV->getIndices()[0] == 0)
654 EV->replaceAllUsesWith(Loaded);
656 EV->replaceAllUsesWith(Success);
658 PrunedInsts.push_back(EV);
661 // We can remove the instructions now we're no longer iterating through them.
662 for (auto EV : PrunedInsts)
663 EV->eraseFromParent();
665 if (!CI->use_empty()) {
666 // Some use of the full struct return that we don't understand has happened,
667 // so we've got to reconstruct it properly.
669 Res = Builder.CreateInsertValue(UndefValue::get(CI->getType()), Loaded, 0);
670 Res = Builder.CreateInsertValue(Res, Success, 1);
672 CI->replaceAllUsesWith(Res);
675 CI->eraseFromParent();
679 bool AtomicExpand::isIdempotentRMW(AtomicRMWInst* RMWI) {
680 auto C = dyn_cast<ConstantInt>(RMWI->getValOperand());
684 AtomicRMWInst::BinOp Op = RMWI->getOperation();
686 case AtomicRMWInst::Add:
687 case AtomicRMWInst::Sub:
688 case AtomicRMWInst::Or:
689 case AtomicRMWInst::Xor:
691 case AtomicRMWInst::And:
692 return C->isMinusOne();
693 // FIXME: we could also treat Min/Max/UMin/UMax by the INT_MIN/INT_MAX/...
699 bool AtomicExpand::simplifyIdempotentRMW(AtomicRMWInst* RMWI) {
700 if (auto ResultingLoad = TLI->lowerIdempotentRMWIntoFencedLoad(RMWI)) {
701 tryExpandAtomicLoad(ResultingLoad);
707 bool llvm::expandAtomicRMWToCmpXchg(AtomicRMWInst *AI,
708 CreateCmpXchgInstFun CreateCmpXchg) {
711 AtomicOrdering MemOpOrder =
712 AI->getOrdering() == Unordered ? Monotonic : AI->getOrdering();
713 Value *Addr = AI->getPointerOperand();
714 BasicBlock *BB = AI->getParent();
715 Function *F = BB->getParent();
716 LLVMContext &Ctx = F->getContext();
718 // Given: atomicrmw some_op iN* %addr, iN %incr ordering
720 // The standard expansion we produce is:
722 // %init_loaded = load atomic iN* %addr
725 // %loaded = phi iN [ %init_loaded, %entry ], [ %new_loaded, %loop ]
726 // %new = some_op iN %loaded, %incr
727 // %pair = cmpxchg iN* %addr, iN %loaded, iN %new
728 // %new_loaded = extractvalue { iN, i1 } %pair, 0
729 // %success = extractvalue { iN, i1 } %pair, 1
730 // br i1 %success, label %atomicrmw.end, label %loop
733 BasicBlock *ExitBB = BB->splitBasicBlock(AI->getIterator(), "atomicrmw.end");
734 BasicBlock *LoopBB = BasicBlock::Create(Ctx, "atomicrmw.start", F, ExitBB);
736 // This grabs the DebugLoc from AI.
737 IRBuilder<> Builder(AI);
739 // The split call above "helpfully" added a branch at the end of BB (to the
740 // wrong place), but we want a load. It's easiest to just remove
741 // the branch entirely.
742 std::prev(BB->end())->eraseFromParent();
743 Builder.SetInsertPoint(BB);
744 LoadInst *InitLoaded = Builder.CreateLoad(Addr);
745 // Atomics require at least natural alignment.
746 InitLoaded->setAlignment(AI->getType()->getPrimitiveSizeInBits() / 8);
747 Builder.CreateBr(LoopBB);
749 // Start the main loop block now that we've taken care of the preliminaries.
750 Builder.SetInsertPoint(LoopBB);
751 PHINode *Loaded = Builder.CreatePHI(AI->getType(), 2, "loaded");
752 Loaded->addIncoming(InitLoaded, BB);
755 performAtomicOp(AI->getOperation(), Builder, Loaded, AI->getValOperand());
757 Value *NewLoaded = nullptr;
758 Value *Success = nullptr;
760 CreateCmpXchg(Builder, Addr, Loaded, NewVal, MemOpOrder,
762 assert(Success && NewLoaded);
764 Loaded->addIncoming(NewLoaded, LoopBB);
766 Builder.CreateCondBr(Success, ExitBB, LoopBB);
768 Builder.SetInsertPoint(ExitBB, ExitBB->begin());
770 AI->replaceAllUsesWith(NewLoaded);
771 AI->eraseFromParent();