1 //===-- AtomicExpandPass.cpp - Expand atomic instructions -------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
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
14 // type coercions.
15 //
16 //===----------------------------------------------------------------------===//
17 
18 #include "llvm/CodeGen/AtomicExpandUtils.h"
19 #include "llvm/CodeGen/Passes.h"
20 #include "llvm/IR/Function.h"
21 #include "llvm/IR/IRBuilder.h"
22 #include "llvm/IR/InstIterator.h"
23 #include "llvm/IR/Instructions.h"
24 #include "llvm/IR/Intrinsics.h"
25 #include "llvm/IR/Module.h"
26 #include "llvm/Support/Debug.h"
27 #include "llvm/Support/raw_ostream.h"
28 #include "llvm/Target/TargetLowering.h"
29 #include "llvm/Target/TargetMachine.h"
30 #include "llvm/Target/TargetSubtargetInfo.h"
31 
32 using namespace llvm;
33 
34 #define DEBUG_TYPE "atomic-expand"
35 
36 namespace {
37   class AtomicExpand: public FunctionPass {
38     const TargetMachine *TM;
39     const TargetLowering *TLI;
40   public:
41     static char ID; // Pass identification, replacement for typeid
AtomicExpand(const TargetMachine * TM=nullptr)42     explicit AtomicExpand(const TargetMachine *TM = nullptr)
43       : FunctionPass(ID), TM(TM), TLI(nullptr) {
44       initializeAtomicExpandPass(*PassRegistry::getPassRegistry());
45     }
46 
47     bool runOnFunction(Function &F) override;
48 
49   private:
50     bool bracketInstWithFences(Instruction *I, AtomicOrdering Order,
51                                bool IsStore, bool IsLoad);
52     IntegerType *getCorrespondingIntegerType(Type *T, const DataLayout &DL);
53     LoadInst *convertAtomicLoadToIntegerType(LoadInst *LI);
54     bool tryExpandAtomicLoad(LoadInst *LI);
55     bool expandAtomicLoadToLL(LoadInst *LI);
56     bool expandAtomicLoadToCmpXchg(LoadInst *LI);
57     StoreInst *convertAtomicStoreToIntegerType(StoreInst *SI);
58     bool expandAtomicStore(StoreInst *SI);
59     bool tryExpandAtomicRMW(AtomicRMWInst *AI);
60     bool expandAtomicOpToLLSC(
61         Instruction *I, Value *Addr, AtomicOrdering MemOpOrder,
62         std::function<Value *(IRBuilder<> &, Value *)> PerformOp);
63     bool expandAtomicCmpXchg(AtomicCmpXchgInst *CI);
64     bool isIdempotentRMW(AtomicRMWInst *AI);
65     bool simplifyIdempotentRMW(AtomicRMWInst *AI);
66   };
67 }
68 
69 char AtomicExpand::ID = 0;
70 char &llvm::AtomicExpandID = AtomicExpand::ID;
71 INITIALIZE_TM_PASS(AtomicExpand, "atomic-expand",
72     "Expand Atomic calls in terms of either load-linked & store-conditional or cmpxchg",
73     false, false)
74 
createAtomicExpandPass(const TargetMachine * TM)75 FunctionPass *llvm::createAtomicExpandPass(const TargetMachine *TM) {
76   return new AtomicExpand(TM);
77 }
78 
runOnFunction(Function & F)79 bool AtomicExpand::runOnFunction(Function &F) {
80   if (!TM || !TM->getSubtargetImpl(F)->enableAtomicExpand())
81     return false;
82   TLI = TM->getSubtargetImpl(F)->getTargetLowering();
83 
84   SmallVector<Instruction *, 1> AtomicInsts;
85 
86   // Changing control-flow while iterating through it is a bad idea, so gather a
87   // list of all atomic instructions before we start.
88   for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I) {
89     if (I->isAtomic())
90       AtomicInsts.push_back(&*I);
91   }
92 
93   bool MadeChange = false;
94   for (auto I : AtomicInsts) {
95     auto LI = dyn_cast<LoadInst>(I);
96     auto SI = dyn_cast<StoreInst>(I);
97     auto RMWI = dyn_cast<AtomicRMWInst>(I);
98     auto CASI = dyn_cast<AtomicCmpXchgInst>(I);
99     assert((LI || SI || RMWI || CASI || isa<FenceInst>(I)) &&
100            "Unknown atomic instruction");
101 
102     auto FenceOrdering = Monotonic;
103     bool IsStore, IsLoad;
104     if (TLI->getInsertFencesForAtomic()) {
105       if (LI && isAtLeastAcquire(LI->getOrdering())) {
106         FenceOrdering = LI->getOrdering();
107         LI->setOrdering(Monotonic);
108         IsStore = false;
109         IsLoad = true;
110       } else if (SI && isAtLeastRelease(SI->getOrdering())) {
111         FenceOrdering = SI->getOrdering();
112         SI->setOrdering(Monotonic);
113         IsStore = true;
114         IsLoad = false;
115       } else if (RMWI && (isAtLeastRelease(RMWI->getOrdering()) ||
116                           isAtLeastAcquire(RMWI->getOrdering()))) {
117         FenceOrdering = RMWI->getOrdering();
118         RMWI->setOrdering(Monotonic);
119         IsStore = IsLoad = true;
120       } else if (CASI && !TLI->shouldExpandAtomicCmpXchgInIR(CASI) &&
121                  (isAtLeastRelease(CASI->getSuccessOrdering()) ||
122                   isAtLeastAcquire(CASI->getSuccessOrdering()))) {
123         // If a compare and swap is lowered to LL/SC, we can do smarter fence
124         // insertion, with a stronger one on the success path than on the
125         // failure path. As a result, fence insertion is directly done by
126         // expandAtomicCmpXchg in that case.
127         FenceOrdering = CASI->getSuccessOrdering();
128         CASI->setSuccessOrdering(Monotonic);
129         CASI->setFailureOrdering(Monotonic);
130         IsStore = IsLoad = true;
131       }
132 
133       if (FenceOrdering != Monotonic) {
134         MadeChange |= bracketInstWithFences(I, FenceOrdering, IsStore, IsLoad);
135       }
136     }
137 
138     if (LI) {
139       if (LI->getType()->isFloatingPointTy()) {
140         // TODO: add a TLI hook to control this so that each target can
141         // convert to lowering the original type one at a time.
142         LI = convertAtomicLoadToIntegerType(LI);
143         assert(LI->getType()->isIntegerTy() && "invariant broken");
144         MadeChange = true;
145       }
146 
147       MadeChange |= tryExpandAtomicLoad(LI);
148     } else if (SI) {
149       if (SI->getValueOperand()->getType()->isFloatingPointTy()) {
150         // TODO: add a TLI hook to control this so that each target can
151         // convert to lowering the original type one at a time.
152         SI = convertAtomicStoreToIntegerType(SI);
153         assert(SI->getValueOperand()->getType()->isIntegerTy() &&
154                "invariant broken");
155         MadeChange = true;
156       }
157 
158       if (TLI->shouldExpandAtomicStoreInIR(SI))
159         MadeChange |= expandAtomicStore(SI);
160     } else if (RMWI) {
161       // There are two different ways of expanding RMW instructions:
162       // - into a load if it is idempotent
163       // - into a Cmpxchg/LL-SC loop otherwise
164       // we try them in that order.
165 
166       if (isIdempotentRMW(RMWI) && simplifyIdempotentRMW(RMWI)) {
167         MadeChange = true;
168       } else {
169         MadeChange |= tryExpandAtomicRMW(RMWI);
170       }
171     } else if (CASI && TLI->shouldExpandAtomicCmpXchgInIR(CASI)) {
172       MadeChange |= expandAtomicCmpXchg(CASI);
173     }
174   }
175   return MadeChange;
176 }
177 
bracketInstWithFences(Instruction * I,AtomicOrdering Order,bool IsStore,bool IsLoad)178 bool AtomicExpand::bracketInstWithFences(Instruction *I, AtomicOrdering Order,
179                                          bool IsStore, bool IsLoad) {
180   IRBuilder<> Builder(I);
181 
182   auto LeadingFence = TLI->emitLeadingFence(Builder, Order, IsStore, IsLoad);
183 
184   auto TrailingFence = TLI->emitTrailingFence(Builder, Order, IsStore, IsLoad);
185   // The trailing fence is emitted before the instruction instead of after
186   // because there is no easy way of setting Builder insertion point after
187   // an instruction. So we must erase it from the BB, and insert it back
188   // in the right place.
189   // We have a guard here because not every atomic operation generates a
190   // trailing fence.
191   if (TrailingFence) {
192     TrailingFence->removeFromParent();
193     TrailingFence->insertAfter(I);
194   }
195 
196   return (LeadingFence || TrailingFence);
197 }
198 
199 /// Get the iX type with the same bitwidth as T.
getCorrespondingIntegerType(Type * T,const DataLayout & DL)200 IntegerType *AtomicExpand::getCorrespondingIntegerType(Type *T,
201                                                        const DataLayout &DL) {
202   EVT VT = TLI->getValueType(DL, T);
203   unsigned BitWidth = VT.getStoreSizeInBits();
204   assert(BitWidth == VT.getSizeInBits() && "must be a power of two");
205   return IntegerType::get(T->getContext(), BitWidth);
206 }
207 
208 /// Convert an atomic load of a non-integral type to an integer load of the
209 /// equivelent bitwidth.  See the function comment on
210 /// convertAtomicStoreToIntegerType for background.
convertAtomicLoadToIntegerType(LoadInst * LI)211 LoadInst *AtomicExpand::convertAtomicLoadToIntegerType(LoadInst *LI) {
212   auto *M = LI->getModule();
213   Type *NewTy = getCorrespondingIntegerType(LI->getType(),
214                                             M->getDataLayout());
215 
216   IRBuilder<> Builder(LI);
217 
218   Value *Addr = LI->getPointerOperand();
219   Type *PT = PointerType::get(NewTy,
220                               Addr->getType()->getPointerAddressSpace());
221   Value *NewAddr = Builder.CreateBitCast(Addr, PT);
222 
223   auto *NewLI = Builder.CreateLoad(NewAddr);
224   NewLI->setAlignment(LI->getAlignment());
225   NewLI->setVolatile(LI->isVolatile());
226   NewLI->setAtomic(LI->getOrdering(), LI->getSynchScope());
227   DEBUG(dbgs() << "Replaced " << *LI << " with " << *NewLI << "\n");
228 
229   Value *NewVal = Builder.CreateBitCast(NewLI, LI->getType());
230   LI->replaceAllUsesWith(NewVal);
231   LI->eraseFromParent();
232   return NewLI;
233 }
234 
tryExpandAtomicLoad(LoadInst * LI)235 bool AtomicExpand::tryExpandAtomicLoad(LoadInst *LI) {
236   switch (TLI->shouldExpandAtomicLoadInIR(LI)) {
237   case TargetLoweringBase::AtomicExpansionKind::None:
238     return false;
239   case TargetLoweringBase::AtomicExpansionKind::LLSC:
240     return expandAtomicOpToLLSC(
241         LI, LI->getPointerOperand(), LI->getOrdering(),
242         [](IRBuilder<> &Builder, Value *Loaded) { return Loaded; });
243   case TargetLoweringBase::AtomicExpansionKind::LLOnly:
244     return expandAtomicLoadToLL(LI);
245   case TargetLoweringBase::AtomicExpansionKind::CmpXChg:
246     return expandAtomicLoadToCmpXchg(LI);
247   }
248   llvm_unreachable("Unhandled case in tryExpandAtomicLoad");
249 }
250 
expandAtomicLoadToLL(LoadInst * LI)251 bool AtomicExpand::expandAtomicLoadToLL(LoadInst *LI) {
252   IRBuilder<> Builder(LI);
253 
254   // On some architectures, load-linked instructions are atomic for larger
255   // sizes than normal loads. For example, the only 64-bit load guaranteed
256   // to be single-copy atomic by ARM is an ldrexd (A3.5.3).
257   Value *Val =
258       TLI->emitLoadLinked(Builder, LI->getPointerOperand(), LI->getOrdering());
259   TLI->emitAtomicCmpXchgNoStoreLLBalance(Builder);
260 
261   LI->replaceAllUsesWith(Val);
262   LI->eraseFromParent();
263 
264   return true;
265 }
266 
expandAtomicLoadToCmpXchg(LoadInst * LI)267 bool AtomicExpand::expandAtomicLoadToCmpXchg(LoadInst *LI) {
268   IRBuilder<> Builder(LI);
269   AtomicOrdering Order = LI->getOrdering();
270   Value *Addr = LI->getPointerOperand();
271   Type *Ty = cast<PointerType>(Addr->getType())->getElementType();
272   Constant *DummyVal = Constant::getNullValue(Ty);
273 
274   Value *Pair = Builder.CreateAtomicCmpXchg(
275       Addr, DummyVal, DummyVal, Order,
276       AtomicCmpXchgInst::getStrongestFailureOrdering(Order));
277   Value *Loaded = Builder.CreateExtractValue(Pair, 0, "loaded");
278 
279   LI->replaceAllUsesWith(Loaded);
280   LI->eraseFromParent();
281 
282   return true;
283 }
284 
285 /// Convert an atomic store of a non-integral type to an integer store of the
286 /// equivelent bitwidth.  We used to not support floating point or vector
287 /// atomics in the IR at all.  The backends learned to deal with the bitcast
288 /// idiom because that was the only way of expressing the notion of a atomic
289 /// float or vector store.  The long term plan is to teach each backend to
290 /// instruction select from the original atomic store, but as a migration
291 /// mechanism, we convert back to the old format which the backends understand.
292 /// Each backend will need individual work to recognize the new format.
convertAtomicStoreToIntegerType(StoreInst * SI)293 StoreInst *AtomicExpand::convertAtomicStoreToIntegerType(StoreInst *SI) {
294   IRBuilder<> Builder(SI);
295   auto *M = SI->getModule();
296   Type *NewTy = getCorrespondingIntegerType(SI->getValueOperand()->getType(),
297                                             M->getDataLayout());
298   Value *NewVal = Builder.CreateBitCast(SI->getValueOperand(), NewTy);
299 
300   Value *Addr = SI->getPointerOperand();
301   Type *PT = PointerType::get(NewTy,
302                               Addr->getType()->getPointerAddressSpace());
303   Value *NewAddr = Builder.CreateBitCast(Addr, PT);
304 
305   StoreInst *NewSI = Builder.CreateStore(NewVal, NewAddr);
306   NewSI->setAlignment(SI->getAlignment());
307   NewSI->setVolatile(SI->isVolatile());
308   NewSI->setAtomic(SI->getOrdering(), SI->getSynchScope());
309   DEBUG(dbgs() << "Replaced " << *SI << " with " << *NewSI << "\n");
310   SI->eraseFromParent();
311   return NewSI;
312 }
313 
expandAtomicStore(StoreInst * SI)314 bool AtomicExpand::expandAtomicStore(StoreInst *SI) {
315   // This function is only called on atomic stores that are too large to be
316   // atomic if implemented as a native store. So we replace them by an
317   // atomic swap, that can be implemented for example as a ldrex/strex on ARM
318   // or lock cmpxchg8/16b on X86, as these are atomic for larger sizes.
319   // It is the responsibility of the target to only signal expansion via
320   // shouldExpandAtomicRMW in cases where this is required and possible.
321   IRBuilder<> Builder(SI);
322   AtomicRMWInst *AI =
323       Builder.CreateAtomicRMW(AtomicRMWInst::Xchg, SI->getPointerOperand(),
324                               SI->getValueOperand(), SI->getOrdering());
325   SI->eraseFromParent();
326 
327   // Now we have an appropriate swap instruction, lower it as usual.
328   return tryExpandAtomicRMW(AI);
329 }
330 
createCmpXchgInstFun(IRBuilder<> & Builder,Value * Addr,Value * Loaded,Value * NewVal,AtomicOrdering MemOpOrder,Value * & Success,Value * & NewLoaded)331 static void createCmpXchgInstFun(IRBuilder<> &Builder, Value *Addr,
332                                  Value *Loaded, Value *NewVal,
333                                  AtomicOrdering MemOpOrder,
334                                  Value *&Success, Value *&NewLoaded) {
335   Value* Pair = Builder.CreateAtomicCmpXchg(
336       Addr, Loaded, NewVal, MemOpOrder,
337       AtomicCmpXchgInst::getStrongestFailureOrdering(MemOpOrder));
338   Success = Builder.CreateExtractValue(Pair, 1, "success");
339   NewLoaded = Builder.CreateExtractValue(Pair, 0, "newloaded");
340 }
341 
342 /// Emit IR to implement the given atomicrmw operation on values in registers,
343 /// returning the new value.
performAtomicOp(AtomicRMWInst::BinOp Op,IRBuilder<> & Builder,Value * Loaded,Value * Inc)344 static Value *performAtomicOp(AtomicRMWInst::BinOp Op, IRBuilder<> &Builder,
345                               Value *Loaded, Value *Inc) {
346   Value *NewVal;
347   switch (Op) {
348   case AtomicRMWInst::Xchg:
349     return Inc;
350   case AtomicRMWInst::Add:
351     return Builder.CreateAdd(Loaded, Inc, "new");
352   case AtomicRMWInst::Sub:
353     return Builder.CreateSub(Loaded, Inc, "new");
354   case AtomicRMWInst::And:
355     return Builder.CreateAnd(Loaded, Inc, "new");
356   case AtomicRMWInst::Nand:
357     return Builder.CreateNot(Builder.CreateAnd(Loaded, Inc), "new");
358   case AtomicRMWInst::Or:
359     return Builder.CreateOr(Loaded, Inc, "new");
360   case AtomicRMWInst::Xor:
361     return Builder.CreateXor(Loaded, Inc, "new");
362   case AtomicRMWInst::Max:
363     NewVal = Builder.CreateICmpSGT(Loaded, Inc);
364     return Builder.CreateSelect(NewVal, Loaded, Inc, "new");
365   case AtomicRMWInst::Min:
366     NewVal = Builder.CreateICmpSLE(Loaded, Inc);
367     return Builder.CreateSelect(NewVal, Loaded, Inc, "new");
368   case AtomicRMWInst::UMax:
369     NewVal = Builder.CreateICmpUGT(Loaded, Inc);
370     return Builder.CreateSelect(NewVal, Loaded, Inc, "new");
371   case AtomicRMWInst::UMin:
372     NewVal = Builder.CreateICmpULE(Loaded, Inc);
373     return Builder.CreateSelect(NewVal, Loaded, Inc, "new");
374   default:
375     llvm_unreachable("Unknown atomic op");
376   }
377 }
378 
tryExpandAtomicRMW(AtomicRMWInst * AI)379 bool AtomicExpand::tryExpandAtomicRMW(AtomicRMWInst *AI) {
380   switch (TLI->shouldExpandAtomicRMWInIR(AI)) {
381   case TargetLoweringBase::AtomicExpansionKind::None:
382     return false;
383   case TargetLoweringBase::AtomicExpansionKind::LLSC:
384     return expandAtomicOpToLLSC(AI, AI->getPointerOperand(), AI->getOrdering(),
385                                 [&](IRBuilder<> &Builder, Value *Loaded) {
386                                   return performAtomicOp(AI->getOperation(),
387                                                          Builder, Loaded,
388                                                          AI->getValOperand());
389                                 });
390   case TargetLoweringBase::AtomicExpansionKind::CmpXChg:
391     return expandAtomicRMWToCmpXchg(AI, createCmpXchgInstFun);
392   default:
393     llvm_unreachable("Unhandled case in tryExpandAtomicRMW");
394   }
395 }
396 
expandAtomicOpToLLSC(Instruction * I,Value * Addr,AtomicOrdering MemOpOrder,std::function<Value * (IRBuilder<> &,Value *)> PerformOp)397 bool AtomicExpand::expandAtomicOpToLLSC(
398     Instruction *I, Value *Addr, AtomicOrdering MemOpOrder,
399     std::function<Value *(IRBuilder<> &, Value *)> PerformOp) {
400   BasicBlock *BB = I->getParent();
401   Function *F = BB->getParent();
402   LLVMContext &Ctx = F->getContext();
403 
404   // Given: atomicrmw some_op iN* %addr, iN %incr ordering
405   //
406   // The standard expansion we produce is:
407   //     [...]
408   //     fence?
409   // atomicrmw.start:
410   //     %loaded = @load.linked(%addr)
411   //     %new = some_op iN %loaded, %incr
412   //     %stored = @store_conditional(%new, %addr)
413   //     %try_again = icmp i32 ne %stored, 0
414   //     br i1 %try_again, label %loop, label %atomicrmw.end
415   // atomicrmw.end:
416   //     fence?
417   //     [...]
418   BasicBlock *ExitBB = BB->splitBasicBlock(I->getIterator(), "atomicrmw.end");
419   BasicBlock *LoopBB =  BasicBlock::Create(Ctx, "atomicrmw.start", F, ExitBB);
420 
421   // This grabs the DebugLoc from I.
422   IRBuilder<> Builder(I);
423 
424   // The split call above "helpfully" added a branch at the end of BB (to the
425   // wrong place), but we might want a fence too. It's easiest to just remove
426   // the branch entirely.
427   std::prev(BB->end())->eraseFromParent();
428   Builder.SetInsertPoint(BB);
429   Builder.CreateBr(LoopBB);
430 
431   // Start the main loop block now that we've taken care of the preliminaries.
432   Builder.SetInsertPoint(LoopBB);
433   Value *Loaded = TLI->emitLoadLinked(Builder, Addr, MemOpOrder);
434 
435   Value *NewVal = PerformOp(Builder, Loaded);
436 
437   Value *StoreSuccess =
438       TLI->emitStoreConditional(Builder, NewVal, Addr, MemOpOrder);
439   Value *TryAgain = Builder.CreateICmpNE(
440       StoreSuccess, ConstantInt::get(IntegerType::get(Ctx, 32), 0), "tryagain");
441   Builder.CreateCondBr(TryAgain, LoopBB, ExitBB);
442 
443   Builder.SetInsertPoint(ExitBB, ExitBB->begin());
444 
445   I->replaceAllUsesWith(Loaded);
446   I->eraseFromParent();
447 
448   return true;
449 }
450 
expandAtomicCmpXchg(AtomicCmpXchgInst * CI)451 bool AtomicExpand::expandAtomicCmpXchg(AtomicCmpXchgInst *CI) {
452   AtomicOrdering SuccessOrder = CI->getSuccessOrdering();
453   AtomicOrdering FailureOrder = CI->getFailureOrdering();
454   Value *Addr = CI->getPointerOperand();
455   BasicBlock *BB = CI->getParent();
456   Function *F = BB->getParent();
457   LLVMContext &Ctx = F->getContext();
458   // If getInsertFencesForAtomic() returns true, then the target does not want
459   // to deal with memory orders, and emitLeading/TrailingFence should take care
460   // of everything. Otherwise, emitLeading/TrailingFence are no-op and we
461   // should preserve the ordering.
462   AtomicOrdering MemOpOrder =
463       TLI->getInsertFencesForAtomic() ? Monotonic : SuccessOrder;
464 
465   // Given: cmpxchg some_op iN* %addr, iN %desired, iN %new success_ord fail_ord
466   //
467   // The full expansion we produce is:
468   //     [...]
469   //     fence?
470   // cmpxchg.start:
471   //     %loaded = @load.linked(%addr)
472   //     %should_store = icmp eq %loaded, %desired
473   //     br i1 %should_store, label %cmpxchg.trystore,
474   //                          label %cmpxchg.nostore
475   // cmpxchg.trystore:
476   //     %stored = @store_conditional(%new, %addr)
477   //     %success = icmp eq i32 %stored, 0
478   //     br i1 %success, label %cmpxchg.success, label %loop/%cmpxchg.failure
479   // cmpxchg.success:
480   //     fence?
481   //     br label %cmpxchg.end
482   // cmpxchg.nostore:
483   //     @load_linked_fail_balance()?
484   //     br label %cmpxchg.failure
485   // cmpxchg.failure:
486   //     fence?
487   //     br label %cmpxchg.end
488   // cmpxchg.end:
489   //     %success = phi i1 [true, %cmpxchg.success], [false, %cmpxchg.failure]
490   //     %restmp = insertvalue { iN, i1 } undef, iN %loaded, 0
491   //     %res = insertvalue { iN, i1 } %restmp, i1 %success, 1
492   //     [...]
493   BasicBlock *ExitBB = BB->splitBasicBlock(CI->getIterator(), "cmpxchg.end");
494   auto FailureBB = BasicBlock::Create(Ctx, "cmpxchg.failure", F, ExitBB);
495   auto NoStoreBB = BasicBlock::Create(Ctx, "cmpxchg.nostore", F, FailureBB);
496   auto SuccessBB = BasicBlock::Create(Ctx, "cmpxchg.success", F, NoStoreBB);
497   auto TryStoreBB = BasicBlock::Create(Ctx, "cmpxchg.trystore", F, SuccessBB);
498   auto LoopBB = BasicBlock::Create(Ctx, "cmpxchg.start", F, TryStoreBB);
499 
500   // This grabs the DebugLoc from CI
501   IRBuilder<> Builder(CI);
502 
503   // The split call above "helpfully" added a branch at the end of BB (to the
504   // wrong place), but we might want a fence too. It's easiest to just remove
505   // the branch entirely.
506   std::prev(BB->end())->eraseFromParent();
507   Builder.SetInsertPoint(BB);
508   TLI->emitLeadingFence(Builder, SuccessOrder, /*IsStore=*/true,
509                         /*IsLoad=*/true);
510   Builder.CreateBr(LoopBB);
511 
512   // Start the main loop block now that we've taken care of the preliminaries.
513   Builder.SetInsertPoint(LoopBB);
514   Value *Loaded = TLI->emitLoadLinked(Builder, Addr, MemOpOrder);
515   Value *ShouldStore =
516       Builder.CreateICmpEQ(Loaded, CI->getCompareOperand(), "should_store");
517 
518   // If the cmpxchg doesn't actually need any ordering when it fails, we can
519   // jump straight past that fence instruction (if it exists).
520   Builder.CreateCondBr(ShouldStore, TryStoreBB, NoStoreBB);
521 
522   Builder.SetInsertPoint(TryStoreBB);
523   Value *StoreSuccess = TLI->emitStoreConditional(
524       Builder, CI->getNewValOperand(), Addr, MemOpOrder);
525   StoreSuccess = Builder.CreateICmpEQ(
526       StoreSuccess, ConstantInt::get(Type::getInt32Ty(Ctx), 0), "success");
527   Builder.CreateCondBr(StoreSuccess, SuccessBB,
528                        CI->isWeak() ? FailureBB : LoopBB);
529 
530   // Make sure later instructions don't get reordered with a fence if necessary.
531   Builder.SetInsertPoint(SuccessBB);
532   TLI->emitTrailingFence(Builder, SuccessOrder, /*IsStore=*/true,
533                          /*IsLoad=*/true);
534   Builder.CreateBr(ExitBB);
535 
536   Builder.SetInsertPoint(NoStoreBB);
537   // In the failing case, where we don't execute the store-conditional, the
538   // target might want to balance out the load-linked with a dedicated
539   // instruction (e.g., on ARM, clearing the exclusive monitor).
540   TLI->emitAtomicCmpXchgNoStoreLLBalance(Builder);
541   Builder.CreateBr(FailureBB);
542 
543   Builder.SetInsertPoint(FailureBB);
544   TLI->emitTrailingFence(Builder, FailureOrder, /*IsStore=*/true,
545                          /*IsLoad=*/true);
546   Builder.CreateBr(ExitBB);
547 
548   // Finally, we have control-flow based knowledge of whether the cmpxchg
549   // succeeded or not. We expose this to later passes by converting any
550   // subsequent "icmp eq/ne %loaded, %oldval" into a use of an appropriate PHI.
551 
552   // Setup the builder so we can create any PHIs we need.
553   Builder.SetInsertPoint(ExitBB, ExitBB->begin());
554   PHINode *Success = Builder.CreatePHI(Type::getInt1Ty(Ctx), 2);
555   Success->addIncoming(ConstantInt::getTrue(Ctx), SuccessBB);
556   Success->addIncoming(ConstantInt::getFalse(Ctx), FailureBB);
557 
558   // Look for any users of the cmpxchg that are just comparing the loaded value
559   // against the desired one, and replace them with the CFG-derived version.
560   SmallVector<ExtractValueInst *, 2> PrunedInsts;
561   for (auto User : CI->users()) {
562     ExtractValueInst *EV = dyn_cast<ExtractValueInst>(User);
563     if (!EV)
564       continue;
565 
566     assert(EV->getNumIndices() == 1 && EV->getIndices()[0] <= 1 &&
567            "weird extraction from { iN, i1 }");
568 
569     if (EV->getIndices()[0] == 0)
570       EV->replaceAllUsesWith(Loaded);
571     else
572       EV->replaceAllUsesWith(Success);
573 
574     PrunedInsts.push_back(EV);
575   }
576 
577   // We can remove the instructions now we're no longer iterating through them.
578   for (auto EV : PrunedInsts)
579     EV->eraseFromParent();
580 
581   if (!CI->use_empty()) {
582     // Some use of the full struct return that we don't understand has happened,
583     // so we've got to reconstruct it properly.
584     Value *Res;
585     Res = Builder.CreateInsertValue(UndefValue::get(CI->getType()), Loaded, 0);
586     Res = Builder.CreateInsertValue(Res, Success, 1);
587 
588     CI->replaceAllUsesWith(Res);
589   }
590 
591   CI->eraseFromParent();
592   return true;
593 }
594 
isIdempotentRMW(AtomicRMWInst * RMWI)595 bool AtomicExpand::isIdempotentRMW(AtomicRMWInst* RMWI) {
596   auto C = dyn_cast<ConstantInt>(RMWI->getValOperand());
597   if(!C)
598     return false;
599 
600   AtomicRMWInst::BinOp Op = RMWI->getOperation();
601   switch(Op) {
602     case AtomicRMWInst::Add:
603     case AtomicRMWInst::Sub:
604     case AtomicRMWInst::Or:
605     case AtomicRMWInst::Xor:
606       return C->isZero();
607     case AtomicRMWInst::And:
608       return C->isMinusOne();
609     // FIXME: we could also treat Min/Max/UMin/UMax by the INT_MIN/INT_MAX/...
610     default:
611       return false;
612   }
613 }
614 
simplifyIdempotentRMW(AtomicRMWInst * RMWI)615 bool AtomicExpand::simplifyIdempotentRMW(AtomicRMWInst* RMWI) {
616   if (auto ResultingLoad = TLI->lowerIdempotentRMWIntoFencedLoad(RMWI)) {
617     tryExpandAtomicLoad(ResultingLoad);
618     return true;
619   }
620   return false;
621 }
622 
expandAtomicRMWToCmpXchg(AtomicRMWInst * AI,CreateCmpXchgInstFun CreateCmpXchg)623 bool llvm::expandAtomicRMWToCmpXchg(AtomicRMWInst *AI,
624                                     CreateCmpXchgInstFun CreateCmpXchg) {
625   assert(AI);
626 
627   AtomicOrdering MemOpOrder =
628       AI->getOrdering() == Unordered ? Monotonic : AI->getOrdering();
629   Value *Addr = AI->getPointerOperand();
630   BasicBlock *BB = AI->getParent();
631   Function *F = BB->getParent();
632   LLVMContext &Ctx = F->getContext();
633 
634   // Given: atomicrmw some_op iN* %addr, iN %incr ordering
635   //
636   // The standard expansion we produce is:
637   //     [...]
638   //     %init_loaded = load atomic iN* %addr
639   //     br label %loop
640   // loop:
641   //     %loaded = phi iN [ %init_loaded, %entry ], [ %new_loaded, %loop ]
642   //     %new = some_op iN %loaded, %incr
643   //     %pair = cmpxchg iN* %addr, iN %loaded, iN %new
644   //     %new_loaded = extractvalue { iN, i1 } %pair, 0
645   //     %success = extractvalue { iN, i1 } %pair, 1
646   //     br i1 %success, label %atomicrmw.end, label %loop
647   // atomicrmw.end:
648   //     [...]
649   BasicBlock *ExitBB = BB->splitBasicBlock(AI->getIterator(), "atomicrmw.end");
650   BasicBlock *LoopBB = BasicBlock::Create(Ctx, "atomicrmw.start", F, ExitBB);
651 
652   // This grabs the DebugLoc from AI.
653   IRBuilder<> Builder(AI);
654 
655   // The split call above "helpfully" added a branch at the end of BB (to the
656   // wrong place), but we want a load. It's easiest to just remove
657   // the branch entirely.
658   std::prev(BB->end())->eraseFromParent();
659   Builder.SetInsertPoint(BB);
660   LoadInst *InitLoaded = Builder.CreateLoad(Addr);
661   // Atomics require at least natural alignment.
662   InitLoaded->setAlignment(AI->getType()->getPrimitiveSizeInBits() / 8);
663   Builder.CreateBr(LoopBB);
664 
665   // Start the main loop block now that we've taken care of the preliminaries.
666   Builder.SetInsertPoint(LoopBB);
667   PHINode *Loaded = Builder.CreatePHI(AI->getType(), 2, "loaded");
668   Loaded->addIncoming(InitLoaded, BB);
669 
670   Value *NewVal =
671       performAtomicOp(AI->getOperation(), Builder, Loaded, AI->getValOperand());
672 
673   Value *NewLoaded = nullptr;
674   Value *Success = nullptr;
675 
676   CreateCmpXchg(Builder, Addr, Loaded, NewVal, MemOpOrder,
677                 Success, NewLoaded);
678   assert(Success && NewLoaded);
679 
680   Loaded->addIncoming(NewLoaded, LoopBB);
681 
682   Builder.CreateCondBr(Success, ExitBB, LoopBB);
683 
684   Builder.SetInsertPoint(ExitBB, ExitBB->begin());
685 
686   AI->replaceAllUsesWith(NewLoaded);
687   AI->eraseFromParent();
688 
689   return true;
690 }
691