1 //===--- CGAtomic.cpp - Emit LLVM IR for atomic operations ----------------===//
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 the code for emitting atomic operations.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "CGCall.h"
15 #include "CGRecordLayout.h"
16 #include "CodeGenFunction.h"
17 #include "CodeGenModule.h"
18 #include "TargetInfo.h"
19 #include "clang/AST/ASTContext.h"
20 #include "clang/CodeGen/CGFunctionInfo.h"
21 #include "llvm/IR/DataLayout.h"
22 #include "llvm/IR/Intrinsics.h"
23 #include "llvm/IR/Operator.h"
24 
25 using namespace clang;
26 using namespace CodeGen;
27 
28 namespace {
29   class AtomicInfo {
30     CodeGenFunction &CGF;
31     QualType AtomicTy;
32     QualType ValueTy;
33     uint64_t AtomicSizeInBits;
34     uint64_t ValueSizeInBits;
35     CharUnits AtomicAlign;
36     CharUnits ValueAlign;
37     CharUnits LValueAlign;
38     TypeEvaluationKind EvaluationKind;
39     bool UseLibcall;
40     LValue LVal;
41     CGBitFieldInfo BFI;
42   public:
43     AtomicInfo(CodeGenFunction &CGF, LValue &lvalue)
44         : CGF(CGF), AtomicSizeInBits(0), ValueSizeInBits(0),
45           EvaluationKind(TEK_Scalar), UseLibcall(true) {
46       assert(!lvalue.isGlobalReg());
47       ASTContext &C = CGF.getContext();
48       if (lvalue.isSimple()) {
49         AtomicTy = lvalue.getType();
50         if (auto *ATy = AtomicTy->getAs<AtomicType>())
51           ValueTy = ATy->getValueType();
52         else
53           ValueTy = AtomicTy;
54         EvaluationKind = CGF.getEvaluationKind(ValueTy);
55 
56         uint64_t ValueAlignInBits;
57         uint64_t AtomicAlignInBits;
58         TypeInfo ValueTI = C.getTypeInfo(ValueTy);
59         ValueSizeInBits = ValueTI.Width;
60         ValueAlignInBits = ValueTI.Align;
61 
62         TypeInfo AtomicTI = C.getTypeInfo(AtomicTy);
63         AtomicSizeInBits = AtomicTI.Width;
64         AtomicAlignInBits = AtomicTI.Align;
65 
66         assert(ValueSizeInBits <= AtomicSizeInBits);
67         assert(ValueAlignInBits <= AtomicAlignInBits);
68 
69         AtomicAlign = C.toCharUnitsFromBits(AtomicAlignInBits);
70         ValueAlign = C.toCharUnitsFromBits(ValueAlignInBits);
71         if (lvalue.getAlignment().isZero())
72           lvalue.setAlignment(AtomicAlign);
73 
74         LVal = lvalue;
75       } else if (lvalue.isBitField()) {
76         ValueTy = lvalue.getType();
77         ValueSizeInBits = C.getTypeSize(ValueTy);
78         auto &OrigBFI = lvalue.getBitFieldInfo();
79         auto Offset = OrigBFI.Offset % C.toBits(lvalue.getAlignment());
80         AtomicSizeInBits = C.toBits(
81             C.toCharUnitsFromBits(Offset + OrigBFI.Size + C.getCharWidth() - 1)
82                 .alignTo(lvalue.getAlignment()));
83         auto VoidPtrAddr = CGF.EmitCastToVoidPtr(lvalue.getBitFieldPointer());
84         auto OffsetInChars =
85             (C.toCharUnitsFromBits(OrigBFI.Offset) / lvalue.getAlignment()) *
86             lvalue.getAlignment();
87         VoidPtrAddr = CGF.Builder.CreateConstGEP1_64(
88             VoidPtrAddr, OffsetInChars.getQuantity());
89         auto Addr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
90             VoidPtrAddr,
91             CGF.Builder.getIntNTy(AtomicSizeInBits)->getPointerTo(),
92             "atomic_bitfield_base");
93         BFI = OrigBFI;
94         BFI.Offset = Offset;
95         BFI.StorageSize = AtomicSizeInBits;
96         BFI.StorageOffset += OffsetInChars;
97         LVal = LValue::MakeBitfield(Address(Addr, lvalue.getAlignment()),
98                                     BFI, lvalue.getType(),
99                                     lvalue.getBaseInfo());
100         LVal.setTBAAInfo(lvalue.getTBAAInfo());
101         AtomicTy = C.getIntTypeForBitwidth(AtomicSizeInBits, OrigBFI.IsSigned);
102         if (AtomicTy.isNull()) {
103           llvm::APInt Size(
104               /*numBits=*/32,
105               C.toCharUnitsFromBits(AtomicSizeInBits).getQuantity());
106           AtomicTy = C.getConstantArrayType(C.CharTy, Size, ArrayType::Normal,
107                                             /*IndexTypeQuals=*/0);
108         }
109         AtomicAlign = ValueAlign = lvalue.getAlignment();
110       } else if (lvalue.isVectorElt()) {
111         ValueTy = lvalue.getType()->getAs<VectorType>()->getElementType();
112         ValueSizeInBits = C.getTypeSize(ValueTy);
113         AtomicTy = lvalue.getType();
114         AtomicSizeInBits = C.getTypeSize(AtomicTy);
115         AtomicAlign = ValueAlign = lvalue.getAlignment();
116         LVal = lvalue;
117       } else {
118         assert(lvalue.isExtVectorElt());
119         ValueTy = lvalue.getType();
120         ValueSizeInBits = C.getTypeSize(ValueTy);
121         AtomicTy = ValueTy = CGF.getContext().getExtVectorType(
122             lvalue.getType(), lvalue.getExtVectorAddress()
123                                   .getElementType()->getVectorNumElements());
124         AtomicSizeInBits = C.getTypeSize(AtomicTy);
125         AtomicAlign = ValueAlign = lvalue.getAlignment();
126         LVal = lvalue;
127       }
128       UseLibcall = !C.getTargetInfo().hasBuiltinAtomic(
129           AtomicSizeInBits, C.toBits(lvalue.getAlignment()));
130     }
131 
132     QualType getAtomicType() const { return AtomicTy; }
133     QualType getValueType() const { return ValueTy; }
134     CharUnits getAtomicAlignment() const { return AtomicAlign; }
135     CharUnits getValueAlignment() const { return ValueAlign; }
136     uint64_t getAtomicSizeInBits() const { return AtomicSizeInBits; }
137     uint64_t getValueSizeInBits() const { return ValueSizeInBits; }
138     TypeEvaluationKind getEvaluationKind() const { return EvaluationKind; }
139     bool shouldUseLibcall() const { return UseLibcall; }
140     const LValue &getAtomicLValue() const { return LVal; }
141     llvm::Value *getAtomicPointer() const {
142       if (LVal.isSimple())
143         return LVal.getPointer();
144       else if (LVal.isBitField())
145         return LVal.getBitFieldPointer();
146       else if (LVal.isVectorElt())
147         return LVal.getVectorPointer();
148       assert(LVal.isExtVectorElt());
149       return LVal.getExtVectorPointer();
150     }
151     Address getAtomicAddress() const {
152       return Address(getAtomicPointer(), getAtomicAlignment());
153     }
154 
155     Address getAtomicAddressAsAtomicIntPointer() const {
156       return emitCastToAtomicIntPointer(getAtomicAddress());
157     }
158 
159     /// Is the atomic size larger than the underlying value type?
160     ///
161     /// Note that the absence of padding does not mean that atomic
162     /// objects are completely interchangeable with non-atomic
163     /// objects: we might have promoted the alignment of a type
164     /// without making it bigger.
165     bool hasPadding() const {
166       return (ValueSizeInBits != AtomicSizeInBits);
167     }
168 
169     bool emitMemSetZeroIfNecessary() const;
170 
171     llvm::Value *getAtomicSizeValue() const {
172       CharUnits size = CGF.getContext().toCharUnitsFromBits(AtomicSizeInBits);
173       return CGF.CGM.getSize(size);
174     }
175 
176     /// Cast the given pointer to an integer pointer suitable for atomic
177     /// operations if the source.
178     Address emitCastToAtomicIntPointer(Address Addr) const;
179 
180     /// If Addr is compatible with the iN that will be used for an atomic
181     /// operation, bitcast it. Otherwise, create a temporary that is suitable
182     /// and copy the value across.
183     Address convertToAtomicIntPointer(Address Addr) const;
184 
185     /// Turn an atomic-layout object into an r-value.
186     RValue convertAtomicTempToRValue(Address addr, AggValueSlot resultSlot,
187                                      SourceLocation loc, bool AsValue) const;
188 
189     /// \brief Converts a rvalue to integer value.
190     llvm::Value *convertRValueToInt(RValue RVal) const;
191 
192     RValue ConvertIntToValueOrAtomic(llvm::Value *IntVal,
193                                      AggValueSlot ResultSlot,
194                                      SourceLocation Loc, bool AsValue) const;
195 
196     /// Copy an atomic r-value into atomic-layout memory.
197     void emitCopyIntoMemory(RValue rvalue) const;
198 
199     /// Project an l-value down to the value field.
200     LValue projectValue() const {
201       assert(LVal.isSimple());
202       Address addr = getAtomicAddress();
203       if (hasPadding())
204         addr = CGF.Builder.CreateStructGEP(addr, 0, CharUnits());
205 
206       return LValue::MakeAddr(addr, getValueType(), CGF.getContext(),
207                               LVal.getBaseInfo(), LVal.getTBAAInfo());
208     }
209 
210     /// \brief Emits atomic load.
211     /// \returns Loaded value.
212     RValue EmitAtomicLoad(AggValueSlot ResultSlot, SourceLocation Loc,
213                           bool AsValue, llvm::AtomicOrdering AO,
214                           bool IsVolatile);
215 
216     /// \brief Emits atomic compare-and-exchange sequence.
217     /// \param Expected Expected value.
218     /// \param Desired Desired value.
219     /// \param Success Atomic ordering for success operation.
220     /// \param Failure Atomic ordering for failed operation.
221     /// \param IsWeak true if atomic operation is weak, false otherwise.
222     /// \returns Pair of values: previous value from storage (value type) and
223     /// boolean flag (i1 type) with true if success and false otherwise.
224     std::pair<RValue, llvm::Value *>
225     EmitAtomicCompareExchange(RValue Expected, RValue Desired,
226                               llvm::AtomicOrdering Success =
227                                   llvm::AtomicOrdering::SequentiallyConsistent,
228                               llvm::AtomicOrdering Failure =
229                                   llvm::AtomicOrdering::SequentiallyConsistent,
230                               bool IsWeak = false);
231 
232     /// \brief Emits atomic update.
233     /// \param AO Atomic ordering.
234     /// \param UpdateOp Update operation for the current lvalue.
235     void EmitAtomicUpdate(llvm::AtomicOrdering AO,
236                           const llvm::function_ref<RValue(RValue)> &UpdateOp,
237                           bool IsVolatile);
238     /// \brief Emits atomic update.
239     /// \param AO Atomic ordering.
240     void EmitAtomicUpdate(llvm::AtomicOrdering AO, RValue UpdateRVal,
241                           bool IsVolatile);
242 
243     /// Materialize an atomic r-value in atomic-layout memory.
244     Address materializeRValue(RValue rvalue) const;
245 
246     /// \brief Creates temp alloca for intermediate operations on atomic value.
247     Address CreateTempAlloca() const;
248   private:
249     bool requiresMemSetZero(llvm::Type *type) const;
250 
251 
252     /// \brief Emits atomic load as a libcall.
253     void EmitAtomicLoadLibcall(llvm::Value *AddForLoaded,
254                                llvm::AtomicOrdering AO, bool IsVolatile);
255     /// \brief Emits atomic load as LLVM instruction.
256     llvm::Value *EmitAtomicLoadOp(llvm::AtomicOrdering AO, bool IsVolatile);
257     /// \brief Emits atomic compare-and-exchange op as a libcall.
258     llvm::Value *EmitAtomicCompareExchangeLibcall(
259         llvm::Value *ExpectedAddr, llvm::Value *DesiredAddr,
260         llvm::AtomicOrdering Success =
261             llvm::AtomicOrdering::SequentiallyConsistent,
262         llvm::AtomicOrdering Failure =
263             llvm::AtomicOrdering::SequentiallyConsistent);
264     /// \brief Emits atomic compare-and-exchange op as LLVM instruction.
265     std::pair<llvm::Value *, llvm::Value *> EmitAtomicCompareExchangeOp(
266         llvm::Value *ExpectedVal, llvm::Value *DesiredVal,
267         llvm::AtomicOrdering Success =
268             llvm::AtomicOrdering::SequentiallyConsistent,
269         llvm::AtomicOrdering Failure =
270             llvm::AtomicOrdering::SequentiallyConsistent,
271         bool IsWeak = false);
272     /// \brief Emit atomic update as libcalls.
273     void
274     EmitAtomicUpdateLibcall(llvm::AtomicOrdering AO,
275                             const llvm::function_ref<RValue(RValue)> &UpdateOp,
276                             bool IsVolatile);
277     /// \brief Emit atomic update as LLVM instructions.
278     void EmitAtomicUpdateOp(llvm::AtomicOrdering AO,
279                             const llvm::function_ref<RValue(RValue)> &UpdateOp,
280                             bool IsVolatile);
281     /// \brief Emit atomic update as libcalls.
282     void EmitAtomicUpdateLibcall(llvm::AtomicOrdering AO, RValue UpdateRVal,
283                                  bool IsVolatile);
284     /// \brief Emit atomic update as LLVM instructions.
285     void EmitAtomicUpdateOp(llvm::AtomicOrdering AO, RValue UpdateRal,
286                             bool IsVolatile);
287   };
288 }
289 
290 Address AtomicInfo::CreateTempAlloca() const {
291   Address TempAlloca = CGF.CreateMemTemp(
292       (LVal.isBitField() && ValueSizeInBits > AtomicSizeInBits) ? ValueTy
293                                                                 : AtomicTy,
294       getAtomicAlignment(),
295       "atomic-temp");
296   // Cast to pointer to value type for bitfields.
297   if (LVal.isBitField())
298     return CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
299         TempAlloca, getAtomicAddress().getType());
300   return TempAlloca;
301 }
302 
303 static RValue emitAtomicLibcall(CodeGenFunction &CGF,
304                                 StringRef fnName,
305                                 QualType resultType,
306                                 CallArgList &args) {
307   const CGFunctionInfo &fnInfo =
308     CGF.CGM.getTypes().arrangeBuiltinFunctionCall(resultType, args);
309   llvm::FunctionType *fnTy = CGF.CGM.getTypes().GetFunctionType(fnInfo);
310   llvm::Constant *fn = CGF.CGM.CreateRuntimeFunction(fnTy, fnName);
311   auto callee = CGCallee::forDirect(fn);
312   return CGF.EmitCall(fnInfo, callee, ReturnValueSlot(), args);
313 }
314 
315 /// Does a store of the given IR type modify the full expected width?
316 static bool isFullSizeType(CodeGenModule &CGM, llvm::Type *type,
317                            uint64_t expectedSize) {
318   return (CGM.getDataLayout().getTypeStoreSize(type) * 8 == expectedSize);
319 }
320 
321 /// Does the atomic type require memsetting to zero before initialization?
322 ///
323 /// The IR type is provided as a way of making certain queries faster.
324 bool AtomicInfo::requiresMemSetZero(llvm::Type *type) const {
325   // If the atomic type has size padding, we definitely need a memset.
326   if (hasPadding()) return true;
327 
328   // Otherwise, do some simple heuristics to try to avoid it:
329   switch (getEvaluationKind()) {
330   // For scalars and complexes, check whether the store size of the
331   // type uses the full size.
332   case TEK_Scalar:
333     return !isFullSizeType(CGF.CGM, type, AtomicSizeInBits);
334   case TEK_Complex:
335     return !isFullSizeType(CGF.CGM, type->getStructElementType(0),
336                            AtomicSizeInBits / 2);
337 
338   // Padding in structs has an undefined bit pattern.  User beware.
339   case TEK_Aggregate:
340     return false;
341   }
342   llvm_unreachable("bad evaluation kind");
343 }
344 
345 bool AtomicInfo::emitMemSetZeroIfNecessary() const {
346   assert(LVal.isSimple());
347   llvm::Value *addr = LVal.getPointer();
348   if (!requiresMemSetZero(addr->getType()->getPointerElementType()))
349     return false;
350 
351   CGF.Builder.CreateMemSet(
352       addr, llvm::ConstantInt::get(CGF.Int8Ty, 0),
353       CGF.getContext().toCharUnitsFromBits(AtomicSizeInBits).getQuantity(),
354       LVal.getAlignment().getQuantity());
355   return true;
356 }
357 
358 static void emitAtomicCmpXchg(CodeGenFunction &CGF, AtomicExpr *E, bool IsWeak,
359                               Address Dest, Address Ptr,
360                               Address Val1, Address Val2,
361                               uint64_t Size,
362                               llvm::AtomicOrdering SuccessOrder,
363                               llvm::AtomicOrdering FailureOrder,
364                               llvm::SyncScope::ID Scope) {
365   // Note that cmpxchg doesn't support weak cmpxchg, at least at the moment.
366   llvm::Value *Expected = CGF.Builder.CreateLoad(Val1);
367   llvm::Value *Desired = CGF.Builder.CreateLoad(Val2);
368 
369   llvm::AtomicCmpXchgInst *Pair = CGF.Builder.CreateAtomicCmpXchg(
370       Ptr.getPointer(), Expected, Desired, SuccessOrder, FailureOrder,
371       Scope);
372   Pair->setVolatile(E->isVolatile());
373   Pair->setWeak(IsWeak);
374 
375   // Cmp holds the result of the compare-exchange operation: true on success,
376   // false on failure.
377   llvm::Value *Old = CGF.Builder.CreateExtractValue(Pair, 0);
378   llvm::Value *Cmp = CGF.Builder.CreateExtractValue(Pair, 1);
379 
380   // This basic block is used to hold the store instruction if the operation
381   // failed.
382   llvm::BasicBlock *StoreExpectedBB =
383       CGF.createBasicBlock("cmpxchg.store_expected", CGF.CurFn);
384 
385   // This basic block is the exit point of the operation, we should end up
386   // here regardless of whether or not the operation succeeded.
387   llvm::BasicBlock *ContinueBB =
388       CGF.createBasicBlock("cmpxchg.continue", CGF.CurFn);
389 
390   // Update Expected if Expected isn't equal to Old, otherwise branch to the
391   // exit point.
392   CGF.Builder.CreateCondBr(Cmp, ContinueBB, StoreExpectedBB);
393 
394   CGF.Builder.SetInsertPoint(StoreExpectedBB);
395   // Update the memory at Expected with Old's value.
396   CGF.Builder.CreateStore(Old, Val1);
397   // Finally, branch to the exit point.
398   CGF.Builder.CreateBr(ContinueBB);
399 
400   CGF.Builder.SetInsertPoint(ContinueBB);
401   // Update the memory at Dest with Cmp's value.
402   CGF.EmitStoreOfScalar(Cmp, CGF.MakeAddrLValue(Dest, E->getType()));
403 }
404 
405 /// Given an ordering required on success, emit all possible cmpxchg
406 /// instructions to cope with the provided (but possibly only dynamically known)
407 /// FailureOrder.
408 static void emitAtomicCmpXchgFailureSet(CodeGenFunction &CGF, AtomicExpr *E,
409                                         bool IsWeak, Address Dest, Address Ptr,
410                                         Address Val1, Address Val2,
411                                         llvm::Value *FailureOrderVal,
412                                         uint64_t Size,
413                                         llvm::AtomicOrdering SuccessOrder,
414                                         llvm::SyncScope::ID Scope) {
415   llvm::AtomicOrdering FailureOrder;
416   if (llvm::ConstantInt *FO = dyn_cast<llvm::ConstantInt>(FailureOrderVal)) {
417     auto FOS = FO->getSExtValue();
418     if (!llvm::isValidAtomicOrderingCABI(FOS))
419       FailureOrder = llvm::AtomicOrdering::Monotonic;
420     else
421       switch ((llvm::AtomicOrderingCABI)FOS) {
422       case llvm::AtomicOrderingCABI::relaxed:
423       case llvm::AtomicOrderingCABI::release:
424       case llvm::AtomicOrderingCABI::acq_rel:
425         FailureOrder = llvm::AtomicOrdering::Monotonic;
426         break;
427       case llvm::AtomicOrderingCABI::consume:
428       case llvm::AtomicOrderingCABI::acquire:
429         FailureOrder = llvm::AtomicOrdering::Acquire;
430         break;
431       case llvm::AtomicOrderingCABI::seq_cst:
432         FailureOrder = llvm::AtomicOrdering::SequentiallyConsistent;
433         break;
434       }
435     if (isStrongerThan(FailureOrder, SuccessOrder)) {
436       // Don't assert on undefined behavior "failure argument shall be no
437       // stronger than the success argument".
438       FailureOrder =
439           llvm::AtomicCmpXchgInst::getStrongestFailureOrdering(SuccessOrder);
440     }
441     emitAtomicCmpXchg(CGF, E, IsWeak, Dest, Ptr, Val1, Val2, Size, SuccessOrder,
442                       FailureOrder, Scope);
443     return;
444   }
445 
446   // Create all the relevant BB's
447   llvm::BasicBlock *MonotonicBB = nullptr, *AcquireBB = nullptr,
448                    *SeqCstBB = nullptr;
449   MonotonicBB = CGF.createBasicBlock("monotonic_fail", CGF.CurFn);
450   if (SuccessOrder != llvm::AtomicOrdering::Monotonic &&
451       SuccessOrder != llvm::AtomicOrdering::Release)
452     AcquireBB = CGF.createBasicBlock("acquire_fail", CGF.CurFn);
453   if (SuccessOrder == llvm::AtomicOrdering::SequentiallyConsistent)
454     SeqCstBB = CGF.createBasicBlock("seqcst_fail", CGF.CurFn);
455 
456   llvm::BasicBlock *ContBB = CGF.createBasicBlock("atomic.continue", CGF.CurFn);
457 
458   llvm::SwitchInst *SI = CGF.Builder.CreateSwitch(FailureOrderVal, MonotonicBB);
459 
460   // Emit all the different atomics
461 
462   // MonotonicBB is arbitrarily chosen as the default case; in practice, this
463   // doesn't matter unless someone is crazy enough to use something that
464   // doesn't fold to a constant for the ordering.
465   CGF.Builder.SetInsertPoint(MonotonicBB);
466   emitAtomicCmpXchg(CGF, E, IsWeak, Dest, Ptr, Val1, Val2,
467                     Size, SuccessOrder, llvm::AtomicOrdering::Monotonic, Scope);
468   CGF.Builder.CreateBr(ContBB);
469 
470   if (AcquireBB) {
471     CGF.Builder.SetInsertPoint(AcquireBB);
472     emitAtomicCmpXchg(CGF, E, IsWeak, Dest, Ptr, Val1, Val2,
473                       Size, SuccessOrder, llvm::AtomicOrdering::Acquire, Scope);
474     CGF.Builder.CreateBr(ContBB);
475     SI->addCase(CGF.Builder.getInt32((int)llvm::AtomicOrderingCABI::consume),
476                 AcquireBB);
477     SI->addCase(CGF.Builder.getInt32((int)llvm::AtomicOrderingCABI::acquire),
478                 AcquireBB);
479   }
480   if (SeqCstBB) {
481     CGF.Builder.SetInsertPoint(SeqCstBB);
482     emitAtomicCmpXchg(CGF, E, IsWeak, Dest, Ptr, Val1, Val2, Size, SuccessOrder,
483                       llvm::AtomicOrdering::SequentiallyConsistent, Scope);
484     CGF.Builder.CreateBr(ContBB);
485     SI->addCase(CGF.Builder.getInt32((int)llvm::AtomicOrderingCABI::seq_cst),
486                 SeqCstBB);
487   }
488 
489   CGF.Builder.SetInsertPoint(ContBB);
490 }
491 
492 static void EmitAtomicOp(CodeGenFunction &CGF, AtomicExpr *E, Address Dest,
493                          Address Ptr, Address Val1, Address Val2,
494                          llvm::Value *IsWeak, llvm::Value *FailureOrder,
495                          uint64_t Size, llvm::AtomicOrdering Order,
496                          llvm::SyncScope::ID Scope) {
497   llvm::AtomicRMWInst::BinOp Op = llvm::AtomicRMWInst::Add;
498   llvm::Instruction::BinaryOps PostOp = (llvm::Instruction::BinaryOps)0;
499 
500   switch (E->getOp()) {
501   case AtomicExpr::AO__c11_atomic_init:
502   case AtomicExpr::AO__opencl_atomic_init:
503     llvm_unreachable("Already handled!");
504 
505   case AtomicExpr::AO__c11_atomic_compare_exchange_strong:
506   case AtomicExpr::AO__opencl_atomic_compare_exchange_strong:
507     emitAtomicCmpXchgFailureSet(CGF, E, false, Dest, Ptr, Val1, Val2,
508                                 FailureOrder, Size, Order, Scope);
509     return;
510   case AtomicExpr::AO__c11_atomic_compare_exchange_weak:
511   case AtomicExpr::AO__opencl_atomic_compare_exchange_weak:
512     emitAtomicCmpXchgFailureSet(CGF, E, true, Dest, Ptr, Val1, Val2,
513                                 FailureOrder, Size, Order, Scope);
514     return;
515   case AtomicExpr::AO__atomic_compare_exchange:
516   case AtomicExpr::AO__atomic_compare_exchange_n: {
517     if (llvm::ConstantInt *IsWeakC = dyn_cast<llvm::ConstantInt>(IsWeak)) {
518       emitAtomicCmpXchgFailureSet(CGF, E, IsWeakC->getZExtValue(), Dest, Ptr,
519                                   Val1, Val2, FailureOrder, Size, Order, Scope);
520     } else {
521       // Create all the relevant BB's
522       llvm::BasicBlock *StrongBB =
523           CGF.createBasicBlock("cmpxchg.strong", CGF.CurFn);
524       llvm::BasicBlock *WeakBB = CGF.createBasicBlock("cmxchg.weak", CGF.CurFn);
525       llvm::BasicBlock *ContBB =
526           CGF.createBasicBlock("cmpxchg.continue", CGF.CurFn);
527 
528       llvm::SwitchInst *SI = CGF.Builder.CreateSwitch(IsWeak, WeakBB);
529       SI->addCase(CGF.Builder.getInt1(false), StrongBB);
530 
531       CGF.Builder.SetInsertPoint(StrongBB);
532       emitAtomicCmpXchgFailureSet(CGF, E, false, Dest, Ptr, Val1, Val2,
533                                   FailureOrder, Size, Order, Scope);
534       CGF.Builder.CreateBr(ContBB);
535 
536       CGF.Builder.SetInsertPoint(WeakBB);
537       emitAtomicCmpXchgFailureSet(CGF, E, true, Dest, Ptr, Val1, Val2,
538                                   FailureOrder, Size, Order, Scope);
539       CGF.Builder.CreateBr(ContBB);
540 
541       CGF.Builder.SetInsertPoint(ContBB);
542     }
543     return;
544   }
545   case AtomicExpr::AO__c11_atomic_load:
546   case AtomicExpr::AO__opencl_atomic_load:
547   case AtomicExpr::AO__atomic_load_n:
548   case AtomicExpr::AO__atomic_load: {
549     llvm::LoadInst *Load = CGF.Builder.CreateLoad(Ptr);
550     Load->setAtomic(Order, Scope);
551     Load->setVolatile(E->isVolatile());
552     CGF.Builder.CreateStore(Load, Dest);
553     return;
554   }
555 
556   case AtomicExpr::AO__c11_atomic_store:
557   case AtomicExpr::AO__opencl_atomic_store:
558   case AtomicExpr::AO__atomic_store:
559   case AtomicExpr::AO__atomic_store_n: {
560     llvm::Value *LoadVal1 = CGF.Builder.CreateLoad(Val1);
561     llvm::StoreInst *Store = CGF.Builder.CreateStore(LoadVal1, Ptr);
562     Store->setAtomic(Order, Scope);
563     Store->setVolatile(E->isVolatile());
564     return;
565   }
566 
567   case AtomicExpr::AO__c11_atomic_exchange:
568   case AtomicExpr::AO__opencl_atomic_exchange:
569   case AtomicExpr::AO__atomic_exchange_n:
570   case AtomicExpr::AO__atomic_exchange:
571     Op = llvm::AtomicRMWInst::Xchg;
572     break;
573 
574   case AtomicExpr::AO__atomic_add_fetch:
575     PostOp = llvm::Instruction::Add;
576     // Fall through.
577   case AtomicExpr::AO__c11_atomic_fetch_add:
578   case AtomicExpr::AO__opencl_atomic_fetch_add:
579   case AtomicExpr::AO__atomic_fetch_add:
580     Op = llvm::AtomicRMWInst::Add;
581     break;
582 
583   case AtomicExpr::AO__atomic_sub_fetch:
584     PostOp = llvm::Instruction::Sub;
585     // Fall through.
586   case AtomicExpr::AO__c11_atomic_fetch_sub:
587   case AtomicExpr::AO__opencl_atomic_fetch_sub:
588   case AtomicExpr::AO__atomic_fetch_sub:
589     Op = llvm::AtomicRMWInst::Sub;
590     break;
591 
592   case AtomicExpr::AO__opencl_atomic_fetch_min:
593     Op = E->getValueType()->isSignedIntegerType() ? llvm::AtomicRMWInst::Min
594                                                   : llvm::AtomicRMWInst::UMin;
595     break;
596 
597   case AtomicExpr::AO__opencl_atomic_fetch_max:
598     Op = E->getValueType()->isSignedIntegerType() ? llvm::AtomicRMWInst::Max
599                                                   : llvm::AtomicRMWInst::UMax;
600     break;
601 
602   case AtomicExpr::AO__atomic_and_fetch:
603     PostOp = llvm::Instruction::And;
604     // Fall through.
605   case AtomicExpr::AO__c11_atomic_fetch_and:
606   case AtomicExpr::AO__opencl_atomic_fetch_and:
607   case AtomicExpr::AO__atomic_fetch_and:
608     Op = llvm::AtomicRMWInst::And;
609     break;
610 
611   case AtomicExpr::AO__atomic_or_fetch:
612     PostOp = llvm::Instruction::Or;
613     // Fall through.
614   case AtomicExpr::AO__c11_atomic_fetch_or:
615   case AtomicExpr::AO__opencl_atomic_fetch_or:
616   case AtomicExpr::AO__atomic_fetch_or:
617     Op = llvm::AtomicRMWInst::Or;
618     break;
619 
620   case AtomicExpr::AO__atomic_xor_fetch:
621     PostOp = llvm::Instruction::Xor;
622     // Fall through.
623   case AtomicExpr::AO__c11_atomic_fetch_xor:
624   case AtomicExpr::AO__opencl_atomic_fetch_xor:
625   case AtomicExpr::AO__atomic_fetch_xor:
626     Op = llvm::AtomicRMWInst::Xor;
627     break;
628 
629   case AtomicExpr::AO__atomic_nand_fetch:
630     PostOp = llvm::Instruction::And; // the NOT is special cased below
631   // Fall through.
632   case AtomicExpr::AO__atomic_fetch_nand:
633     Op = llvm::AtomicRMWInst::Nand;
634     break;
635   }
636 
637   llvm::Value *LoadVal1 = CGF.Builder.CreateLoad(Val1);
638   llvm::AtomicRMWInst *RMWI =
639       CGF.Builder.CreateAtomicRMW(Op, Ptr.getPointer(), LoadVal1, Order, Scope);
640   RMWI->setVolatile(E->isVolatile());
641 
642   // For __atomic_*_fetch operations, perform the operation again to
643   // determine the value which was written.
644   llvm::Value *Result = RMWI;
645   if (PostOp)
646     Result = CGF.Builder.CreateBinOp(PostOp, RMWI, LoadVal1);
647   if (E->getOp() == AtomicExpr::AO__atomic_nand_fetch)
648     Result = CGF.Builder.CreateNot(Result);
649   CGF.Builder.CreateStore(Result, Dest);
650 }
651 
652 // This function emits any expression (scalar, complex, or aggregate)
653 // into a temporary alloca.
654 static Address
655 EmitValToTemp(CodeGenFunction &CGF, Expr *E) {
656   Address DeclPtr = CGF.CreateMemTemp(E->getType(), ".atomictmp");
657   CGF.EmitAnyExprToMem(E, DeclPtr, E->getType().getQualifiers(),
658                        /*Init*/ true);
659   return DeclPtr;
660 }
661 
662 static void
663 AddDirectArgument(CodeGenFunction &CGF, CallArgList &Args,
664                   bool UseOptimizedLibcall, llvm::Value *Val, QualType ValTy,
665                   SourceLocation Loc, CharUnits SizeInChars) {
666   if (UseOptimizedLibcall) {
667     // Load value and pass it to the function directly.
668     CharUnits Align = CGF.getContext().getTypeAlignInChars(ValTy);
669     int64_t SizeInBits = CGF.getContext().toBits(SizeInChars);
670     ValTy =
671         CGF.getContext().getIntTypeForBitwidth(SizeInBits, /*Signed=*/false);
672     llvm::Type *IPtrTy = llvm::IntegerType::get(CGF.getLLVMContext(),
673                                                 SizeInBits)->getPointerTo();
674     Address Ptr = Address(CGF.Builder.CreateBitCast(Val, IPtrTy), Align);
675     Val = CGF.EmitLoadOfScalar(Ptr, false,
676                                CGF.getContext().getPointerType(ValTy),
677                                Loc);
678     // Coerce the value into an appropriately sized integer type.
679     Args.add(RValue::get(Val), ValTy);
680   } else {
681     // Non-optimized functions always take a reference.
682     Args.add(RValue::get(CGF.EmitCastToVoidPtr(Val)),
683                          CGF.getContext().VoidPtrTy);
684   }
685 }
686 
687 RValue CodeGenFunction::EmitAtomicExpr(AtomicExpr *E) {
688   QualType AtomicTy = E->getPtr()->getType()->getPointeeType();
689   QualType MemTy = AtomicTy;
690   if (const AtomicType *AT = AtomicTy->getAs<AtomicType>())
691     MemTy = AT->getValueType();
692   CharUnits sizeChars, alignChars;
693   std::tie(sizeChars, alignChars) = getContext().getTypeInfoInChars(AtomicTy);
694   uint64_t Size = sizeChars.getQuantity();
695   unsigned MaxInlineWidthInBits = getTarget().getMaxAtomicInlineWidth();
696   bool UseLibcall = (sizeChars != alignChars ||
697                      getContext().toBits(sizeChars) > MaxInlineWidthInBits);
698 
699   llvm::Value *IsWeak = nullptr, *OrderFail = nullptr;
700 
701   Address Val1 = Address::invalid();
702   Address Val2 = Address::invalid();
703   Address Dest = Address::invalid();
704   Address Ptr(EmitScalarExpr(E->getPtr()), alignChars);
705 
706   if (E->getOp() == AtomicExpr::AO__c11_atomic_init ||
707       E->getOp() == AtomicExpr::AO__opencl_atomic_init) {
708     LValue lvalue = MakeAddrLValue(Ptr, AtomicTy);
709     EmitAtomicInit(E->getVal1(), lvalue);
710     return RValue::get(nullptr);
711   }
712 
713   llvm::Value *Order = EmitScalarExpr(E->getOrder());
714   llvm::Value *Scope = EmitScalarExpr(E->getScope());
715 
716   switch (E->getOp()) {
717   case AtomicExpr::AO__c11_atomic_init:
718   case AtomicExpr::AO__opencl_atomic_init:
719     llvm_unreachable("Already handled above with EmitAtomicInit!");
720 
721   case AtomicExpr::AO__c11_atomic_load:
722   case AtomicExpr::AO__opencl_atomic_load:
723   case AtomicExpr::AO__atomic_load_n:
724     break;
725 
726   case AtomicExpr::AO__atomic_load:
727     Dest = EmitPointerWithAlignment(E->getVal1());
728     break;
729 
730   case AtomicExpr::AO__atomic_store:
731     Val1 = EmitPointerWithAlignment(E->getVal1());
732     break;
733 
734   case AtomicExpr::AO__atomic_exchange:
735     Val1 = EmitPointerWithAlignment(E->getVal1());
736     Dest = EmitPointerWithAlignment(E->getVal2());
737     break;
738 
739   case AtomicExpr::AO__c11_atomic_compare_exchange_strong:
740   case AtomicExpr::AO__c11_atomic_compare_exchange_weak:
741   case AtomicExpr::AO__opencl_atomic_compare_exchange_strong:
742   case AtomicExpr::AO__opencl_atomic_compare_exchange_weak:
743   case AtomicExpr::AO__atomic_compare_exchange_n:
744   case AtomicExpr::AO__atomic_compare_exchange:
745     Val1 = EmitPointerWithAlignment(E->getVal1());
746     if (E->getOp() == AtomicExpr::AO__atomic_compare_exchange)
747       Val2 = EmitPointerWithAlignment(E->getVal2());
748     else
749       Val2 = EmitValToTemp(*this, E->getVal2());
750     OrderFail = EmitScalarExpr(E->getOrderFail());
751     if (E->getOp() == AtomicExpr::AO__atomic_compare_exchange_n ||
752         E->getOp() == AtomicExpr::AO__atomic_compare_exchange)
753       IsWeak = EmitScalarExpr(E->getWeak());
754     break;
755 
756   case AtomicExpr::AO__c11_atomic_fetch_add:
757   case AtomicExpr::AO__c11_atomic_fetch_sub:
758   case AtomicExpr::AO__opencl_atomic_fetch_add:
759   case AtomicExpr::AO__opencl_atomic_fetch_sub:
760     if (MemTy->isPointerType()) {
761       // For pointer arithmetic, we're required to do a bit of math:
762       // adding 1 to an int* is not the same as adding 1 to a uintptr_t.
763       // ... but only for the C11 builtins. The GNU builtins expect the
764       // user to multiply by sizeof(T).
765       QualType Val1Ty = E->getVal1()->getType();
766       llvm::Value *Val1Scalar = EmitScalarExpr(E->getVal1());
767       CharUnits PointeeIncAmt =
768           getContext().getTypeSizeInChars(MemTy->getPointeeType());
769       Val1Scalar = Builder.CreateMul(Val1Scalar, CGM.getSize(PointeeIncAmt));
770       auto Temp = CreateMemTemp(Val1Ty, ".atomictmp");
771       Val1 = Temp;
772       EmitStoreOfScalar(Val1Scalar, MakeAddrLValue(Temp, Val1Ty));
773       break;
774     }
775     // Fall through.
776   case AtomicExpr::AO__atomic_fetch_add:
777   case AtomicExpr::AO__atomic_fetch_sub:
778   case AtomicExpr::AO__atomic_add_fetch:
779   case AtomicExpr::AO__atomic_sub_fetch:
780   case AtomicExpr::AO__c11_atomic_store:
781   case AtomicExpr::AO__c11_atomic_exchange:
782   case AtomicExpr::AO__opencl_atomic_store:
783   case AtomicExpr::AO__opencl_atomic_exchange:
784   case AtomicExpr::AO__atomic_store_n:
785   case AtomicExpr::AO__atomic_exchange_n:
786   case AtomicExpr::AO__c11_atomic_fetch_and:
787   case AtomicExpr::AO__c11_atomic_fetch_or:
788   case AtomicExpr::AO__c11_atomic_fetch_xor:
789   case AtomicExpr::AO__opencl_atomic_fetch_and:
790   case AtomicExpr::AO__opencl_atomic_fetch_or:
791   case AtomicExpr::AO__opencl_atomic_fetch_xor:
792   case AtomicExpr::AO__opencl_atomic_fetch_min:
793   case AtomicExpr::AO__opencl_atomic_fetch_max:
794   case AtomicExpr::AO__atomic_fetch_and:
795   case AtomicExpr::AO__atomic_fetch_or:
796   case AtomicExpr::AO__atomic_fetch_xor:
797   case AtomicExpr::AO__atomic_fetch_nand:
798   case AtomicExpr::AO__atomic_and_fetch:
799   case AtomicExpr::AO__atomic_or_fetch:
800   case AtomicExpr::AO__atomic_xor_fetch:
801   case AtomicExpr::AO__atomic_nand_fetch:
802     Val1 = EmitValToTemp(*this, E->getVal1());
803     break;
804   }
805 
806   QualType RValTy = E->getType().getUnqualifiedType();
807 
808   // The inlined atomics only function on iN types, where N is a power of 2. We
809   // need to make sure (via temporaries if necessary) that all incoming values
810   // are compatible.
811   LValue AtomicVal = MakeAddrLValue(Ptr, AtomicTy);
812   AtomicInfo Atomics(*this, AtomicVal);
813 
814   Ptr = Atomics.emitCastToAtomicIntPointer(Ptr);
815   if (Val1.isValid()) Val1 = Atomics.convertToAtomicIntPointer(Val1);
816   if (Val2.isValid()) Val2 = Atomics.convertToAtomicIntPointer(Val2);
817   if (Dest.isValid())
818     Dest = Atomics.emitCastToAtomicIntPointer(Dest);
819   else if (E->isCmpXChg())
820     Dest = CreateMemTemp(RValTy, "cmpxchg.bool");
821   else if (!RValTy->isVoidType())
822     Dest = Atomics.emitCastToAtomicIntPointer(Atomics.CreateTempAlloca());
823 
824   // Use a library call.  See: http://gcc.gnu.org/wiki/Atomic/GCCMM/LIbrary .
825   if (UseLibcall) {
826     bool UseOptimizedLibcall = false;
827     switch (E->getOp()) {
828     case AtomicExpr::AO__c11_atomic_init:
829     case AtomicExpr::AO__opencl_atomic_init:
830       llvm_unreachable("Already handled above with EmitAtomicInit!");
831 
832     case AtomicExpr::AO__c11_atomic_fetch_add:
833     case AtomicExpr::AO__opencl_atomic_fetch_add:
834     case AtomicExpr::AO__atomic_fetch_add:
835     case AtomicExpr::AO__c11_atomic_fetch_and:
836     case AtomicExpr::AO__opencl_atomic_fetch_and:
837     case AtomicExpr::AO__atomic_fetch_and:
838     case AtomicExpr::AO__c11_atomic_fetch_or:
839     case AtomicExpr::AO__opencl_atomic_fetch_or:
840     case AtomicExpr::AO__atomic_fetch_or:
841     case AtomicExpr::AO__atomic_fetch_nand:
842     case AtomicExpr::AO__c11_atomic_fetch_sub:
843     case AtomicExpr::AO__opencl_atomic_fetch_sub:
844     case AtomicExpr::AO__atomic_fetch_sub:
845     case AtomicExpr::AO__c11_atomic_fetch_xor:
846     case AtomicExpr::AO__opencl_atomic_fetch_xor:
847     case AtomicExpr::AO__opencl_atomic_fetch_min:
848     case AtomicExpr::AO__opencl_atomic_fetch_max:
849     case AtomicExpr::AO__atomic_fetch_xor:
850     case AtomicExpr::AO__atomic_add_fetch:
851     case AtomicExpr::AO__atomic_and_fetch:
852     case AtomicExpr::AO__atomic_nand_fetch:
853     case AtomicExpr::AO__atomic_or_fetch:
854     case AtomicExpr::AO__atomic_sub_fetch:
855     case AtomicExpr::AO__atomic_xor_fetch:
856       // For these, only library calls for certain sizes exist.
857       UseOptimizedLibcall = true;
858       break;
859 
860     case AtomicExpr::AO__c11_atomic_load:
861     case AtomicExpr::AO__c11_atomic_store:
862     case AtomicExpr::AO__c11_atomic_exchange:
863     case AtomicExpr::AO__c11_atomic_compare_exchange_weak:
864     case AtomicExpr::AO__c11_atomic_compare_exchange_strong:
865     case AtomicExpr::AO__opencl_atomic_load:
866     case AtomicExpr::AO__opencl_atomic_store:
867     case AtomicExpr::AO__opencl_atomic_exchange:
868     case AtomicExpr::AO__opencl_atomic_compare_exchange_weak:
869     case AtomicExpr::AO__opencl_atomic_compare_exchange_strong:
870     case AtomicExpr::AO__atomic_load_n:
871     case AtomicExpr::AO__atomic_load:
872     case AtomicExpr::AO__atomic_store_n:
873     case AtomicExpr::AO__atomic_store:
874     case AtomicExpr::AO__atomic_exchange_n:
875     case AtomicExpr::AO__atomic_exchange:
876     case AtomicExpr::AO__atomic_compare_exchange_n:
877     case AtomicExpr::AO__atomic_compare_exchange:
878       // Only use optimized library calls for sizes for which they exist.
879       if (Size == 1 || Size == 2 || Size == 4 || Size == 8)
880         UseOptimizedLibcall = true;
881       break;
882     }
883 
884     CallArgList Args;
885     if (!UseOptimizedLibcall) {
886       // For non-optimized library calls, the size is the first parameter
887       Args.add(RValue::get(llvm::ConstantInt::get(SizeTy, Size)),
888                getContext().getSizeType());
889     }
890     // Atomic address is the first or second parameter
891     // The OpenCL atomic library functions only accept pointer arguments to
892     // generic address space.
893     auto CastToGenericAddrSpace = [&](llvm::Value *V, QualType PT) {
894       if (!E->isOpenCL())
895         return V;
896       auto AS = PT->getAs<PointerType>()->getPointeeType().getAddressSpace();
897       if (AS == LangAS::opencl_generic)
898         return V;
899       auto DestAS = getContext().getTargetAddressSpace(LangAS::opencl_generic);
900       auto T = V->getType();
901       auto *DestType = T->getPointerElementType()->getPointerTo(DestAS);
902 
903       return getTargetHooks().performAddrSpaceCast(
904           *this, V, AS, LangAS::opencl_generic, DestType, false);
905     };
906 
907     Args.add(RValue::get(CastToGenericAddrSpace(
908                  EmitCastToVoidPtr(Ptr.getPointer()), E->getPtr()->getType())),
909              getContext().VoidPtrTy);
910 
911     std::string LibCallName;
912     QualType LoweredMemTy =
913       MemTy->isPointerType() ? getContext().getIntPtrType() : MemTy;
914     QualType RetTy;
915     bool HaveRetTy = false;
916     llvm::Instruction::BinaryOps PostOp = (llvm::Instruction::BinaryOps)0;
917     switch (E->getOp()) {
918     case AtomicExpr::AO__c11_atomic_init:
919     case AtomicExpr::AO__opencl_atomic_init:
920       llvm_unreachable("Already handled!");
921 
922     // There is only one libcall for compare an exchange, because there is no
923     // optimisation benefit possible from a libcall version of a weak compare
924     // and exchange.
925     // bool __atomic_compare_exchange(size_t size, void *mem, void *expected,
926     //                                void *desired, int success, int failure)
927     // bool __atomic_compare_exchange_N(T *mem, T *expected, T desired,
928     //                                  int success, int failure)
929     case AtomicExpr::AO__c11_atomic_compare_exchange_weak:
930     case AtomicExpr::AO__c11_atomic_compare_exchange_strong:
931     case AtomicExpr::AO__opencl_atomic_compare_exchange_weak:
932     case AtomicExpr::AO__opencl_atomic_compare_exchange_strong:
933     case AtomicExpr::AO__atomic_compare_exchange:
934     case AtomicExpr::AO__atomic_compare_exchange_n:
935       LibCallName = "__atomic_compare_exchange";
936       RetTy = getContext().BoolTy;
937       HaveRetTy = true;
938       Args.add(
939           RValue::get(CastToGenericAddrSpace(
940               EmitCastToVoidPtr(Val1.getPointer()), E->getVal1()->getType())),
941           getContext().VoidPtrTy);
942       AddDirectArgument(*this, Args, UseOptimizedLibcall, Val2.getPointer(),
943                         MemTy, E->getExprLoc(), sizeChars);
944       Args.add(RValue::get(Order), getContext().IntTy);
945       Order = OrderFail;
946       break;
947     // void __atomic_exchange(size_t size, void *mem, void *val, void *return,
948     //                        int order)
949     // T __atomic_exchange_N(T *mem, T val, int order)
950     case AtomicExpr::AO__c11_atomic_exchange:
951     case AtomicExpr::AO__opencl_atomic_exchange:
952     case AtomicExpr::AO__atomic_exchange_n:
953     case AtomicExpr::AO__atomic_exchange:
954       LibCallName = "__atomic_exchange";
955       AddDirectArgument(*this, Args, UseOptimizedLibcall, Val1.getPointer(),
956                         MemTy, E->getExprLoc(), sizeChars);
957       break;
958     // void __atomic_store(size_t size, void *mem, void *val, int order)
959     // void __atomic_store_N(T *mem, T val, int order)
960     case AtomicExpr::AO__c11_atomic_store:
961     case AtomicExpr::AO__opencl_atomic_store:
962     case AtomicExpr::AO__atomic_store:
963     case AtomicExpr::AO__atomic_store_n:
964       LibCallName = "__atomic_store";
965       RetTy = getContext().VoidTy;
966       HaveRetTy = true;
967       AddDirectArgument(*this, Args, UseOptimizedLibcall, Val1.getPointer(),
968                         MemTy, E->getExprLoc(), sizeChars);
969       break;
970     // void __atomic_load(size_t size, void *mem, void *return, int order)
971     // T __atomic_load_N(T *mem, int order)
972     case AtomicExpr::AO__c11_atomic_load:
973     case AtomicExpr::AO__opencl_atomic_load:
974     case AtomicExpr::AO__atomic_load:
975     case AtomicExpr::AO__atomic_load_n:
976       LibCallName = "__atomic_load";
977       break;
978     // T __atomic_add_fetch_N(T *mem, T val, int order)
979     // T __atomic_fetch_add_N(T *mem, T val, int order)
980     case AtomicExpr::AO__atomic_add_fetch:
981       PostOp = llvm::Instruction::Add;
982     // Fall through.
983     case AtomicExpr::AO__c11_atomic_fetch_add:
984     case AtomicExpr::AO__opencl_atomic_fetch_add:
985     case AtomicExpr::AO__atomic_fetch_add:
986       LibCallName = "__atomic_fetch_add";
987       AddDirectArgument(*this, Args, UseOptimizedLibcall, Val1.getPointer(),
988                         LoweredMemTy, E->getExprLoc(), sizeChars);
989       break;
990     // T __atomic_and_fetch_N(T *mem, T val, int order)
991     // T __atomic_fetch_and_N(T *mem, T val, int order)
992     case AtomicExpr::AO__atomic_and_fetch:
993       PostOp = llvm::Instruction::And;
994     // Fall through.
995     case AtomicExpr::AO__c11_atomic_fetch_and:
996     case AtomicExpr::AO__opencl_atomic_fetch_and:
997     case AtomicExpr::AO__atomic_fetch_and:
998       LibCallName = "__atomic_fetch_and";
999       AddDirectArgument(*this, Args, UseOptimizedLibcall, Val1.getPointer(),
1000                         MemTy, E->getExprLoc(), sizeChars);
1001       break;
1002     // T __atomic_or_fetch_N(T *mem, T val, int order)
1003     // T __atomic_fetch_or_N(T *mem, T val, int order)
1004     case AtomicExpr::AO__atomic_or_fetch:
1005       PostOp = llvm::Instruction::Or;
1006     // Fall through.
1007     case AtomicExpr::AO__c11_atomic_fetch_or:
1008     case AtomicExpr::AO__opencl_atomic_fetch_or:
1009     case AtomicExpr::AO__atomic_fetch_or:
1010       LibCallName = "__atomic_fetch_or";
1011       AddDirectArgument(*this, Args, UseOptimizedLibcall, Val1.getPointer(),
1012                         MemTy, E->getExprLoc(), sizeChars);
1013       break;
1014     // T __atomic_sub_fetch_N(T *mem, T val, int order)
1015     // T __atomic_fetch_sub_N(T *mem, T val, int order)
1016     case AtomicExpr::AO__atomic_sub_fetch:
1017       PostOp = llvm::Instruction::Sub;
1018     // Fall through.
1019     case AtomicExpr::AO__c11_atomic_fetch_sub:
1020     case AtomicExpr::AO__opencl_atomic_fetch_sub:
1021     case AtomicExpr::AO__atomic_fetch_sub:
1022       LibCallName = "__atomic_fetch_sub";
1023       AddDirectArgument(*this, Args, UseOptimizedLibcall, Val1.getPointer(),
1024                         LoweredMemTy, E->getExprLoc(), sizeChars);
1025       break;
1026     // T __atomic_xor_fetch_N(T *mem, T val, int order)
1027     // T __atomic_fetch_xor_N(T *mem, T val, int order)
1028     case AtomicExpr::AO__atomic_xor_fetch:
1029       PostOp = llvm::Instruction::Xor;
1030     // Fall through.
1031     case AtomicExpr::AO__c11_atomic_fetch_xor:
1032     case AtomicExpr::AO__opencl_atomic_fetch_xor:
1033     case AtomicExpr::AO__atomic_fetch_xor:
1034       LibCallName = "__atomic_fetch_xor";
1035       AddDirectArgument(*this, Args, UseOptimizedLibcall, Val1.getPointer(),
1036                         MemTy, E->getExprLoc(), sizeChars);
1037       break;
1038     case AtomicExpr::AO__opencl_atomic_fetch_min:
1039       LibCallName = E->getValueType()->isSignedIntegerType()
1040                         ? "__atomic_fetch_min"
1041                         : "__atomic_fetch_umin";
1042       AddDirectArgument(*this, Args, UseOptimizedLibcall, Val1.getPointer(),
1043                         LoweredMemTy, E->getExprLoc(), sizeChars);
1044       break;
1045     case AtomicExpr::AO__opencl_atomic_fetch_max:
1046       LibCallName = E->getValueType()->isSignedIntegerType()
1047                         ? "__atomic_fetch_max"
1048                         : "__atomic_fetch_umax";
1049       AddDirectArgument(*this, Args, UseOptimizedLibcall, Val1.getPointer(),
1050                         LoweredMemTy, E->getExprLoc(), sizeChars);
1051       break;
1052     // T __atomic_nand_fetch_N(T *mem, T val, int order)
1053     // T __atomic_fetch_nand_N(T *mem, T val, int order)
1054     case AtomicExpr::AO__atomic_nand_fetch:
1055       PostOp = llvm::Instruction::And; // the NOT is special cased below
1056     // Fall through.
1057     case AtomicExpr::AO__atomic_fetch_nand:
1058       LibCallName = "__atomic_fetch_nand";
1059       AddDirectArgument(*this, Args, UseOptimizedLibcall, Val1.getPointer(),
1060                         MemTy, E->getExprLoc(), sizeChars);
1061       break;
1062     }
1063 
1064     if (E->isOpenCL()) {
1065       LibCallName = std::string("__opencl") +
1066           StringRef(LibCallName).drop_front(1).str();
1067 
1068     }
1069     // Optimized functions have the size in their name.
1070     if (UseOptimizedLibcall)
1071       LibCallName += "_" + llvm::utostr(Size);
1072     // By default, assume we return a value of the atomic type.
1073     if (!HaveRetTy) {
1074       if (UseOptimizedLibcall) {
1075         // Value is returned directly.
1076         // The function returns an appropriately sized integer type.
1077         RetTy = getContext().getIntTypeForBitwidth(
1078             getContext().toBits(sizeChars), /*Signed=*/false);
1079       } else {
1080         // Value is returned through parameter before the order.
1081         RetTy = getContext().VoidTy;
1082         Args.add(RValue::get(EmitCastToVoidPtr(Dest.getPointer())),
1083                  getContext().VoidPtrTy);
1084       }
1085     }
1086     // order is always the last parameter
1087     Args.add(RValue::get(Order),
1088              getContext().IntTy);
1089     if (E->isOpenCL())
1090       Args.add(RValue::get(Scope), getContext().IntTy);
1091 
1092     // PostOp is only needed for the atomic_*_fetch operations, and
1093     // thus is only needed for and implemented in the
1094     // UseOptimizedLibcall codepath.
1095     assert(UseOptimizedLibcall || !PostOp);
1096 
1097     RValue Res = emitAtomicLibcall(*this, LibCallName, RetTy, Args);
1098     // The value is returned directly from the libcall.
1099     if (E->isCmpXChg())
1100       return Res;
1101 
1102     // The value is returned directly for optimized libcalls but the expr
1103     // provided an out-param.
1104     if (UseOptimizedLibcall && Res.getScalarVal()) {
1105       llvm::Value *ResVal = Res.getScalarVal();
1106       if (PostOp) {
1107         llvm::Value *LoadVal1 = Args[1].RV.getScalarVal();
1108         ResVal = Builder.CreateBinOp(PostOp, ResVal, LoadVal1);
1109       }
1110       if (E->getOp() == AtomicExpr::AO__atomic_nand_fetch)
1111         ResVal = Builder.CreateNot(ResVal);
1112 
1113       Builder.CreateStore(
1114           ResVal,
1115           Builder.CreateBitCast(Dest, ResVal->getType()->getPointerTo()));
1116     }
1117 
1118     if (RValTy->isVoidType())
1119       return RValue::get(nullptr);
1120 
1121     return convertTempToRValue(
1122         Builder.CreateBitCast(Dest, ConvertTypeForMem(RValTy)->getPointerTo()),
1123         RValTy, E->getExprLoc());
1124   }
1125 
1126   bool IsStore = E->getOp() == AtomicExpr::AO__c11_atomic_store ||
1127                  E->getOp() == AtomicExpr::AO__opencl_atomic_store ||
1128                  E->getOp() == AtomicExpr::AO__atomic_store ||
1129                  E->getOp() == AtomicExpr::AO__atomic_store_n;
1130   bool IsLoad = E->getOp() == AtomicExpr::AO__c11_atomic_load ||
1131                 E->getOp() == AtomicExpr::AO__opencl_atomic_load ||
1132                 E->getOp() == AtomicExpr::AO__atomic_load ||
1133                 E->getOp() == AtomicExpr::AO__atomic_load_n;
1134 
1135   assert(isa<llvm::ConstantInt>(Scope) &&
1136       "Non-constant synchronization scope not supported");
1137   auto SCID = getTargetHooks().getLLVMSyncScopeID(
1138       static_cast<SyncScope>(cast<llvm::ConstantInt>(Scope)->getZExtValue()),
1139       getLLVMContext());
1140 
1141   if (isa<llvm::ConstantInt>(Order)) {
1142     auto ord = cast<llvm::ConstantInt>(Order)->getZExtValue();
1143     // We should not ever get to a case where the ordering isn't a valid C ABI
1144     // value, but it's hard to enforce that in general.
1145     if (llvm::isValidAtomicOrderingCABI(ord))
1146       switch ((llvm::AtomicOrderingCABI)ord) {
1147       case llvm::AtomicOrderingCABI::relaxed:
1148         EmitAtomicOp(*this, E, Dest, Ptr, Val1, Val2, IsWeak, OrderFail, Size,
1149                      llvm::AtomicOrdering::Monotonic, SCID);
1150         break;
1151       case llvm::AtomicOrderingCABI::consume:
1152       case llvm::AtomicOrderingCABI::acquire:
1153         if (IsStore)
1154           break; // Avoid crashing on code with undefined behavior
1155         EmitAtomicOp(*this, E, Dest, Ptr, Val1, Val2, IsWeak, OrderFail, Size,
1156                      llvm::AtomicOrdering::Acquire, SCID);
1157         break;
1158       case llvm::AtomicOrderingCABI::release:
1159         if (IsLoad)
1160           break; // Avoid crashing on code with undefined behavior
1161         EmitAtomicOp(*this, E, Dest, Ptr, Val1, Val2, IsWeak, OrderFail, Size,
1162                      llvm::AtomicOrdering::Release, SCID);
1163         break;
1164       case llvm::AtomicOrderingCABI::acq_rel:
1165         if (IsLoad || IsStore)
1166           break; // Avoid crashing on code with undefined behavior
1167         EmitAtomicOp(*this, E, Dest, Ptr, Val1, Val2, IsWeak, OrderFail, Size,
1168                      llvm::AtomicOrdering::AcquireRelease, SCID);
1169         break;
1170       case llvm::AtomicOrderingCABI::seq_cst:
1171         EmitAtomicOp(*this, E, Dest, Ptr, Val1, Val2, IsWeak, OrderFail, Size,
1172                      llvm::AtomicOrdering::SequentiallyConsistent, SCID);
1173         break;
1174       }
1175     if (RValTy->isVoidType())
1176       return RValue::get(nullptr);
1177 
1178     return convertTempToRValue(
1179         Builder.CreateBitCast(Dest, ConvertTypeForMem(RValTy)->getPointerTo()),
1180         RValTy, E->getExprLoc());
1181   }
1182 
1183   // Long case, when Order isn't obviously constant.
1184 
1185   // Create all the relevant BB's
1186   llvm::BasicBlock *MonotonicBB = nullptr, *AcquireBB = nullptr,
1187                    *ReleaseBB = nullptr, *AcqRelBB = nullptr,
1188                    *SeqCstBB = nullptr;
1189   MonotonicBB = createBasicBlock("monotonic", CurFn);
1190   if (!IsStore)
1191     AcquireBB = createBasicBlock("acquire", CurFn);
1192   if (!IsLoad)
1193     ReleaseBB = createBasicBlock("release", CurFn);
1194   if (!IsLoad && !IsStore)
1195     AcqRelBB = createBasicBlock("acqrel", CurFn);
1196   SeqCstBB = createBasicBlock("seqcst", CurFn);
1197   llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn);
1198 
1199   // Create the switch for the split
1200   // MonotonicBB is arbitrarily chosen as the default case; in practice, this
1201   // doesn't matter unless someone is crazy enough to use something that
1202   // doesn't fold to a constant for the ordering.
1203   Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false);
1204   llvm::SwitchInst *SI = Builder.CreateSwitch(Order, MonotonicBB);
1205 
1206   // Emit all the different atomics
1207   Builder.SetInsertPoint(MonotonicBB);
1208   EmitAtomicOp(*this, E, Dest, Ptr, Val1, Val2, IsWeak, OrderFail, Size,
1209                llvm::AtomicOrdering::Monotonic, SCID);
1210   Builder.CreateBr(ContBB);
1211   if (!IsStore) {
1212     Builder.SetInsertPoint(AcquireBB);
1213     EmitAtomicOp(*this, E, Dest, Ptr, Val1, Val2, IsWeak, OrderFail, Size,
1214                  llvm::AtomicOrdering::Acquire, SCID);
1215     Builder.CreateBr(ContBB);
1216     SI->addCase(Builder.getInt32((int)llvm::AtomicOrderingCABI::consume),
1217                 AcquireBB);
1218     SI->addCase(Builder.getInt32((int)llvm::AtomicOrderingCABI::acquire),
1219                 AcquireBB);
1220   }
1221   if (!IsLoad) {
1222     Builder.SetInsertPoint(ReleaseBB);
1223     EmitAtomicOp(*this, E, Dest, Ptr, Val1, Val2, IsWeak, OrderFail, Size,
1224                  llvm::AtomicOrdering::Release, SCID);
1225     Builder.CreateBr(ContBB);
1226     SI->addCase(Builder.getInt32((int)llvm::AtomicOrderingCABI::release),
1227                 ReleaseBB);
1228   }
1229   if (!IsLoad && !IsStore) {
1230     Builder.SetInsertPoint(AcqRelBB);
1231     EmitAtomicOp(*this, E, Dest, Ptr, Val1, Val2, IsWeak, OrderFail, Size,
1232                  llvm::AtomicOrdering::AcquireRelease, SCID);
1233     Builder.CreateBr(ContBB);
1234     SI->addCase(Builder.getInt32((int)llvm::AtomicOrderingCABI::acq_rel),
1235                 AcqRelBB);
1236   }
1237   Builder.SetInsertPoint(SeqCstBB);
1238   EmitAtomicOp(*this, E, Dest, Ptr, Val1, Val2, IsWeak, OrderFail, Size,
1239                llvm::AtomicOrdering::SequentiallyConsistent, SCID);
1240   Builder.CreateBr(ContBB);
1241   SI->addCase(Builder.getInt32((int)llvm::AtomicOrderingCABI::seq_cst),
1242               SeqCstBB);
1243 
1244   // Cleanup and return
1245   Builder.SetInsertPoint(ContBB);
1246   if (RValTy->isVoidType())
1247     return RValue::get(nullptr);
1248 
1249   assert(Atomics.getValueSizeInBits() <= Atomics.getAtomicSizeInBits());
1250   return convertTempToRValue(
1251       Builder.CreateBitCast(Dest, ConvertTypeForMem(RValTy)->getPointerTo()),
1252       RValTy, E->getExprLoc());
1253 }
1254 
1255 Address AtomicInfo::emitCastToAtomicIntPointer(Address addr) const {
1256   unsigned addrspace =
1257     cast<llvm::PointerType>(addr.getPointer()->getType())->getAddressSpace();
1258   llvm::IntegerType *ty =
1259     llvm::IntegerType::get(CGF.getLLVMContext(), AtomicSizeInBits);
1260   return CGF.Builder.CreateBitCast(addr, ty->getPointerTo(addrspace));
1261 }
1262 
1263 Address AtomicInfo::convertToAtomicIntPointer(Address Addr) const {
1264   llvm::Type *Ty = Addr.getElementType();
1265   uint64_t SourceSizeInBits = CGF.CGM.getDataLayout().getTypeSizeInBits(Ty);
1266   if (SourceSizeInBits != AtomicSizeInBits) {
1267     Address Tmp = CreateTempAlloca();
1268     CGF.Builder.CreateMemCpy(Tmp, Addr,
1269                              std::min(AtomicSizeInBits, SourceSizeInBits) / 8);
1270     Addr = Tmp;
1271   }
1272 
1273   return emitCastToAtomicIntPointer(Addr);
1274 }
1275 
1276 RValue AtomicInfo::convertAtomicTempToRValue(Address addr,
1277                                              AggValueSlot resultSlot,
1278                                              SourceLocation loc,
1279                                              bool asValue) const {
1280   if (LVal.isSimple()) {
1281     if (EvaluationKind == TEK_Aggregate)
1282       return resultSlot.asRValue();
1283 
1284     // Drill into the padding structure if we have one.
1285     if (hasPadding())
1286       addr = CGF.Builder.CreateStructGEP(addr, 0, CharUnits());
1287 
1288     // Otherwise, just convert the temporary to an r-value using the
1289     // normal conversion routine.
1290     return CGF.convertTempToRValue(addr, getValueType(), loc);
1291   }
1292   if (!asValue)
1293     // Get RValue from temp memory as atomic for non-simple lvalues
1294     return RValue::get(CGF.Builder.CreateLoad(addr));
1295   if (LVal.isBitField())
1296     return CGF.EmitLoadOfBitfieldLValue(
1297         LValue::MakeBitfield(addr, LVal.getBitFieldInfo(), LVal.getType(),
1298                              LVal.getBaseInfo()), loc);
1299   if (LVal.isVectorElt())
1300     return CGF.EmitLoadOfLValue(
1301         LValue::MakeVectorElt(addr, LVal.getVectorIdx(), LVal.getType(),
1302                               LVal.getBaseInfo()), loc);
1303   assert(LVal.isExtVectorElt());
1304   return CGF.EmitLoadOfExtVectorElementLValue(LValue::MakeExtVectorElt(
1305       addr, LVal.getExtVectorElts(), LVal.getType(),
1306       LVal.getBaseInfo()));
1307 }
1308 
1309 RValue AtomicInfo::ConvertIntToValueOrAtomic(llvm::Value *IntVal,
1310                                              AggValueSlot ResultSlot,
1311                                              SourceLocation Loc,
1312                                              bool AsValue) const {
1313   // Try not to in some easy cases.
1314   assert(IntVal->getType()->isIntegerTy() && "Expected integer value");
1315   if (getEvaluationKind() == TEK_Scalar &&
1316       (((!LVal.isBitField() ||
1317          LVal.getBitFieldInfo().Size == ValueSizeInBits) &&
1318         !hasPadding()) ||
1319        !AsValue)) {
1320     auto *ValTy = AsValue
1321                       ? CGF.ConvertTypeForMem(ValueTy)
1322                       : getAtomicAddress().getType()->getPointerElementType();
1323     if (ValTy->isIntegerTy()) {
1324       assert(IntVal->getType() == ValTy && "Different integer types.");
1325       return RValue::get(CGF.EmitFromMemory(IntVal, ValueTy));
1326     } else if (ValTy->isPointerTy())
1327       return RValue::get(CGF.Builder.CreateIntToPtr(IntVal, ValTy));
1328     else if (llvm::CastInst::isBitCastable(IntVal->getType(), ValTy))
1329       return RValue::get(CGF.Builder.CreateBitCast(IntVal, ValTy));
1330   }
1331 
1332   // Create a temporary.  This needs to be big enough to hold the
1333   // atomic integer.
1334   Address Temp = Address::invalid();
1335   bool TempIsVolatile = false;
1336   if (AsValue && getEvaluationKind() == TEK_Aggregate) {
1337     assert(!ResultSlot.isIgnored());
1338     Temp = ResultSlot.getAddress();
1339     TempIsVolatile = ResultSlot.isVolatile();
1340   } else {
1341     Temp = CreateTempAlloca();
1342   }
1343 
1344   // Slam the integer into the temporary.
1345   Address CastTemp = emitCastToAtomicIntPointer(Temp);
1346   CGF.Builder.CreateStore(IntVal, CastTemp)
1347       ->setVolatile(TempIsVolatile);
1348 
1349   return convertAtomicTempToRValue(Temp, ResultSlot, Loc, AsValue);
1350 }
1351 
1352 void AtomicInfo::EmitAtomicLoadLibcall(llvm::Value *AddForLoaded,
1353                                        llvm::AtomicOrdering AO, bool) {
1354   // void __atomic_load(size_t size, void *mem, void *return, int order);
1355   CallArgList Args;
1356   Args.add(RValue::get(getAtomicSizeValue()), CGF.getContext().getSizeType());
1357   Args.add(RValue::get(CGF.EmitCastToVoidPtr(getAtomicPointer())),
1358            CGF.getContext().VoidPtrTy);
1359   Args.add(RValue::get(CGF.EmitCastToVoidPtr(AddForLoaded)),
1360            CGF.getContext().VoidPtrTy);
1361   Args.add(
1362       RValue::get(llvm::ConstantInt::get(CGF.IntTy, (int)llvm::toCABI(AO))),
1363       CGF.getContext().IntTy);
1364   emitAtomicLibcall(CGF, "__atomic_load", CGF.getContext().VoidTy, Args);
1365 }
1366 
1367 llvm::Value *AtomicInfo::EmitAtomicLoadOp(llvm::AtomicOrdering AO,
1368                                           bool IsVolatile) {
1369   // Okay, we're doing this natively.
1370   Address Addr = getAtomicAddressAsAtomicIntPointer();
1371   llvm::LoadInst *Load = CGF.Builder.CreateLoad(Addr, "atomic-load");
1372   Load->setAtomic(AO);
1373 
1374   // Other decoration.
1375   if (IsVolatile)
1376     Load->setVolatile(true);
1377   if (LVal.getTBAAInfo())
1378     CGF.CGM.DecorateInstructionWithTBAA(Load, LVal.getTBAAInfo());
1379   return Load;
1380 }
1381 
1382 /// An LValue is a candidate for having its loads and stores be made atomic if
1383 /// we are operating under /volatile:ms *and* the LValue itself is volatile and
1384 /// performing such an operation can be performed without a libcall.
1385 bool CodeGenFunction::LValueIsSuitableForInlineAtomic(LValue LV) {
1386   if (!CGM.getCodeGenOpts().MSVolatile) return false;
1387   AtomicInfo AI(*this, LV);
1388   bool IsVolatile = LV.isVolatile() || hasVolatileMember(LV.getType());
1389   // An atomic is inline if we don't need to use a libcall.
1390   bool AtomicIsInline = !AI.shouldUseLibcall();
1391   // MSVC doesn't seem to do this for types wider than a pointer.
1392   if (getContext().getTypeSize(LV.getType()) >
1393       getContext().getTypeSize(getContext().getIntPtrType()))
1394     return false;
1395   return IsVolatile && AtomicIsInline;
1396 }
1397 
1398 RValue CodeGenFunction::EmitAtomicLoad(LValue LV, SourceLocation SL,
1399                                        AggValueSlot Slot) {
1400   llvm::AtomicOrdering AO;
1401   bool IsVolatile = LV.isVolatileQualified();
1402   if (LV.getType()->isAtomicType()) {
1403     AO = llvm::AtomicOrdering::SequentiallyConsistent;
1404   } else {
1405     AO = llvm::AtomicOrdering::Acquire;
1406     IsVolatile = true;
1407   }
1408   return EmitAtomicLoad(LV, SL, AO, IsVolatile, Slot);
1409 }
1410 
1411 RValue AtomicInfo::EmitAtomicLoad(AggValueSlot ResultSlot, SourceLocation Loc,
1412                                   bool AsValue, llvm::AtomicOrdering AO,
1413                                   bool IsVolatile) {
1414   // Check whether we should use a library call.
1415   if (shouldUseLibcall()) {
1416     Address TempAddr = Address::invalid();
1417     if (LVal.isSimple() && !ResultSlot.isIgnored()) {
1418       assert(getEvaluationKind() == TEK_Aggregate);
1419       TempAddr = ResultSlot.getAddress();
1420     } else
1421       TempAddr = CreateTempAlloca();
1422 
1423     EmitAtomicLoadLibcall(TempAddr.getPointer(), AO, IsVolatile);
1424 
1425     // Okay, turn that back into the original value or whole atomic (for
1426     // non-simple lvalues) type.
1427     return convertAtomicTempToRValue(TempAddr, ResultSlot, Loc, AsValue);
1428   }
1429 
1430   // Okay, we're doing this natively.
1431   auto *Load = EmitAtomicLoadOp(AO, IsVolatile);
1432 
1433   // If we're ignoring an aggregate return, don't do anything.
1434   if (getEvaluationKind() == TEK_Aggregate && ResultSlot.isIgnored())
1435     return RValue::getAggregate(Address::invalid(), false);
1436 
1437   // Okay, turn that back into the original value or atomic (for non-simple
1438   // lvalues) type.
1439   return ConvertIntToValueOrAtomic(Load, ResultSlot, Loc, AsValue);
1440 }
1441 
1442 /// Emit a load from an l-value of atomic type.  Note that the r-value
1443 /// we produce is an r-value of the atomic *value* type.
1444 RValue CodeGenFunction::EmitAtomicLoad(LValue src, SourceLocation loc,
1445                                        llvm::AtomicOrdering AO, bool IsVolatile,
1446                                        AggValueSlot resultSlot) {
1447   AtomicInfo Atomics(*this, src);
1448   return Atomics.EmitAtomicLoad(resultSlot, loc, /*AsValue=*/true, AO,
1449                                 IsVolatile);
1450 }
1451 
1452 /// Copy an r-value into memory as part of storing to an atomic type.
1453 /// This needs to create a bit-pattern suitable for atomic operations.
1454 void AtomicInfo::emitCopyIntoMemory(RValue rvalue) const {
1455   assert(LVal.isSimple());
1456   // If we have an r-value, the rvalue should be of the atomic type,
1457   // which means that the caller is responsible for having zeroed
1458   // any padding.  Just do an aggregate copy of that type.
1459   if (rvalue.isAggregate()) {
1460     CGF.EmitAggregateCopy(getAtomicAddress(),
1461                           rvalue.getAggregateAddress(),
1462                           getAtomicType(),
1463                           (rvalue.isVolatileQualified()
1464                            || LVal.isVolatileQualified()));
1465     return;
1466   }
1467 
1468   // Okay, otherwise we're copying stuff.
1469 
1470   // Zero out the buffer if necessary.
1471   emitMemSetZeroIfNecessary();
1472 
1473   // Drill past the padding if present.
1474   LValue TempLVal = projectValue();
1475 
1476   // Okay, store the rvalue in.
1477   if (rvalue.isScalar()) {
1478     CGF.EmitStoreOfScalar(rvalue.getScalarVal(), TempLVal, /*init*/ true);
1479   } else {
1480     CGF.EmitStoreOfComplex(rvalue.getComplexVal(), TempLVal, /*init*/ true);
1481   }
1482 }
1483 
1484 
1485 /// Materialize an r-value into memory for the purposes of storing it
1486 /// to an atomic type.
1487 Address AtomicInfo::materializeRValue(RValue rvalue) const {
1488   // Aggregate r-values are already in memory, and EmitAtomicStore
1489   // requires them to be values of the atomic type.
1490   if (rvalue.isAggregate())
1491     return rvalue.getAggregateAddress();
1492 
1493   // Otherwise, make a temporary and materialize into it.
1494   LValue TempLV = CGF.MakeAddrLValue(CreateTempAlloca(), getAtomicType());
1495   AtomicInfo Atomics(CGF, TempLV);
1496   Atomics.emitCopyIntoMemory(rvalue);
1497   return TempLV.getAddress();
1498 }
1499 
1500 llvm::Value *AtomicInfo::convertRValueToInt(RValue RVal) const {
1501   // If we've got a scalar value of the right size, try to avoid going
1502   // through memory.
1503   if (RVal.isScalar() && (!hasPadding() || !LVal.isSimple())) {
1504     llvm::Value *Value = RVal.getScalarVal();
1505     if (isa<llvm::IntegerType>(Value->getType()))
1506       return CGF.EmitToMemory(Value, ValueTy);
1507     else {
1508       llvm::IntegerType *InputIntTy = llvm::IntegerType::get(
1509           CGF.getLLVMContext(),
1510           LVal.isSimple() ? getValueSizeInBits() : getAtomicSizeInBits());
1511       if (isa<llvm::PointerType>(Value->getType()))
1512         return CGF.Builder.CreatePtrToInt(Value, InputIntTy);
1513       else if (llvm::BitCastInst::isBitCastable(Value->getType(), InputIntTy))
1514         return CGF.Builder.CreateBitCast(Value, InputIntTy);
1515     }
1516   }
1517   // Otherwise, we need to go through memory.
1518   // Put the r-value in memory.
1519   Address Addr = materializeRValue(RVal);
1520 
1521   // Cast the temporary to the atomic int type and pull a value out.
1522   Addr = emitCastToAtomicIntPointer(Addr);
1523   return CGF.Builder.CreateLoad(Addr);
1524 }
1525 
1526 std::pair<llvm::Value *, llvm::Value *> AtomicInfo::EmitAtomicCompareExchangeOp(
1527     llvm::Value *ExpectedVal, llvm::Value *DesiredVal,
1528     llvm::AtomicOrdering Success, llvm::AtomicOrdering Failure, bool IsWeak) {
1529   // Do the atomic store.
1530   Address Addr = getAtomicAddressAsAtomicIntPointer();
1531   auto *Inst = CGF.Builder.CreateAtomicCmpXchg(Addr.getPointer(),
1532                                                ExpectedVal, DesiredVal,
1533                                                Success, Failure);
1534   // Other decoration.
1535   Inst->setVolatile(LVal.isVolatileQualified());
1536   Inst->setWeak(IsWeak);
1537 
1538   // Okay, turn that back into the original value type.
1539   auto *PreviousVal = CGF.Builder.CreateExtractValue(Inst, /*Idxs=*/0);
1540   auto *SuccessFailureVal = CGF.Builder.CreateExtractValue(Inst, /*Idxs=*/1);
1541   return std::make_pair(PreviousVal, SuccessFailureVal);
1542 }
1543 
1544 llvm::Value *
1545 AtomicInfo::EmitAtomicCompareExchangeLibcall(llvm::Value *ExpectedAddr,
1546                                              llvm::Value *DesiredAddr,
1547                                              llvm::AtomicOrdering Success,
1548                                              llvm::AtomicOrdering Failure) {
1549   // bool __atomic_compare_exchange(size_t size, void *obj, void *expected,
1550   // void *desired, int success, int failure);
1551   CallArgList Args;
1552   Args.add(RValue::get(getAtomicSizeValue()), CGF.getContext().getSizeType());
1553   Args.add(RValue::get(CGF.EmitCastToVoidPtr(getAtomicPointer())),
1554            CGF.getContext().VoidPtrTy);
1555   Args.add(RValue::get(CGF.EmitCastToVoidPtr(ExpectedAddr)),
1556            CGF.getContext().VoidPtrTy);
1557   Args.add(RValue::get(CGF.EmitCastToVoidPtr(DesiredAddr)),
1558            CGF.getContext().VoidPtrTy);
1559   Args.add(RValue::get(
1560                llvm::ConstantInt::get(CGF.IntTy, (int)llvm::toCABI(Success))),
1561            CGF.getContext().IntTy);
1562   Args.add(RValue::get(
1563                llvm::ConstantInt::get(CGF.IntTy, (int)llvm::toCABI(Failure))),
1564            CGF.getContext().IntTy);
1565   auto SuccessFailureRVal = emitAtomicLibcall(CGF, "__atomic_compare_exchange",
1566                                               CGF.getContext().BoolTy, Args);
1567 
1568   return SuccessFailureRVal.getScalarVal();
1569 }
1570 
1571 std::pair<RValue, llvm::Value *> AtomicInfo::EmitAtomicCompareExchange(
1572     RValue Expected, RValue Desired, llvm::AtomicOrdering Success,
1573     llvm::AtomicOrdering Failure, bool IsWeak) {
1574   if (isStrongerThan(Failure, Success))
1575     // Don't assert on undefined behavior "failure argument shall be no stronger
1576     // than the success argument".
1577     Failure = llvm::AtomicCmpXchgInst::getStrongestFailureOrdering(Success);
1578 
1579   // Check whether we should use a library call.
1580   if (shouldUseLibcall()) {
1581     // Produce a source address.
1582     Address ExpectedAddr = materializeRValue(Expected);
1583     Address DesiredAddr = materializeRValue(Desired);
1584     auto *Res = EmitAtomicCompareExchangeLibcall(ExpectedAddr.getPointer(),
1585                                                  DesiredAddr.getPointer(),
1586                                                  Success, Failure);
1587     return std::make_pair(
1588         convertAtomicTempToRValue(ExpectedAddr, AggValueSlot::ignored(),
1589                                   SourceLocation(), /*AsValue=*/false),
1590         Res);
1591   }
1592 
1593   // If we've got a scalar value of the right size, try to avoid going
1594   // through memory.
1595   auto *ExpectedVal = convertRValueToInt(Expected);
1596   auto *DesiredVal = convertRValueToInt(Desired);
1597   auto Res = EmitAtomicCompareExchangeOp(ExpectedVal, DesiredVal, Success,
1598                                          Failure, IsWeak);
1599   return std::make_pair(
1600       ConvertIntToValueOrAtomic(Res.first, AggValueSlot::ignored(),
1601                                 SourceLocation(), /*AsValue=*/false),
1602       Res.second);
1603 }
1604 
1605 static void
1606 EmitAtomicUpdateValue(CodeGenFunction &CGF, AtomicInfo &Atomics, RValue OldRVal,
1607                       const llvm::function_ref<RValue(RValue)> &UpdateOp,
1608                       Address DesiredAddr) {
1609   RValue UpRVal;
1610   LValue AtomicLVal = Atomics.getAtomicLValue();
1611   LValue DesiredLVal;
1612   if (AtomicLVal.isSimple()) {
1613     UpRVal = OldRVal;
1614     DesiredLVal = CGF.MakeAddrLValue(DesiredAddr, AtomicLVal.getType());
1615   } else {
1616     // Build new lvalue for temp address
1617     Address Ptr = Atomics.materializeRValue(OldRVal);
1618     LValue UpdateLVal;
1619     if (AtomicLVal.isBitField()) {
1620       UpdateLVal =
1621           LValue::MakeBitfield(Ptr, AtomicLVal.getBitFieldInfo(),
1622                                AtomicLVal.getType(),
1623                                AtomicLVal.getBaseInfo());
1624       DesiredLVal =
1625           LValue::MakeBitfield(DesiredAddr, AtomicLVal.getBitFieldInfo(),
1626                                AtomicLVal.getType(),
1627                                AtomicLVal.getBaseInfo());
1628     } else if (AtomicLVal.isVectorElt()) {
1629       UpdateLVal = LValue::MakeVectorElt(Ptr, AtomicLVal.getVectorIdx(),
1630                                          AtomicLVal.getType(),
1631                                          AtomicLVal.getBaseInfo());
1632       DesiredLVal = LValue::MakeVectorElt(
1633           DesiredAddr, AtomicLVal.getVectorIdx(), AtomicLVal.getType(),
1634           AtomicLVal.getBaseInfo());
1635     } else {
1636       assert(AtomicLVal.isExtVectorElt());
1637       UpdateLVal = LValue::MakeExtVectorElt(Ptr, AtomicLVal.getExtVectorElts(),
1638                                             AtomicLVal.getType(),
1639                                             AtomicLVal.getBaseInfo());
1640       DesiredLVal = LValue::MakeExtVectorElt(
1641           DesiredAddr, AtomicLVal.getExtVectorElts(), AtomicLVal.getType(),
1642           AtomicLVal.getBaseInfo());
1643     }
1644     UpdateLVal.setTBAAInfo(AtomicLVal.getTBAAInfo());
1645     DesiredLVal.setTBAAInfo(AtomicLVal.getTBAAInfo());
1646     UpRVal = CGF.EmitLoadOfLValue(UpdateLVal, SourceLocation());
1647   }
1648   // Store new value in the corresponding memory area
1649   RValue NewRVal = UpdateOp(UpRVal);
1650   if (NewRVal.isScalar()) {
1651     CGF.EmitStoreThroughLValue(NewRVal, DesiredLVal);
1652   } else {
1653     assert(NewRVal.isComplex());
1654     CGF.EmitStoreOfComplex(NewRVal.getComplexVal(), DesiredLVal,
1655                            /*isInit=*/false);
1656   }
1657 }
1658 
1659 void AtomicInfo::EmitAtomicUpdateLibcall(
1660     llvm::AtomicOrdering AO, const llvm::function_ref<RValue(RValue)> &UpdateOp,
1661     bool IsVolatile) {
1662   auto Failure = llvm::AtomicCmpXchgInst::getStrongestFailureOrdering(AO);
1663 
1664   Address ExpectedAddr = CreateTempAlloca();
1665 
1666   EmitAtomicLoadLibcall(ExpectedAddr.getPointer(), AO, IsVolatile);
1667   auto *ContBB = CGF.createBasicBlock("atomic_cont");
1668   auto *ExitBB = CGF.createBasicBlock("atomic_exit");
1669   CGF.EmitBlock(ContBB);
1670   Address DesiredAddr = CreateTempAlloca();
1671   if ((LVal.isBitField() && BFI.Size != ValueSizeInBits) ||
1672       requiresMemSetZero(getAtomicAddress().getElementType())) {
1673     auto *OldVal = CGF.Builder.CreateLoad(ExpectedAddr);
1674     CGF.Builder.CreateStore(OldVal, DesiredAddr);
1675   }
1676   auto OldRVal = convertAtomicTempToRValue(ExpectedAddr,
1677                                            AggValueSlot::ignored(),
1678                                            SourceLocation(), /*AsValue=*/false);
1679   EmitAtomicUpdateValue(CGF, *this, OldRVal, UpdateOp, DesiredAddr);
1680   auto *Res =
1681       EmitAtomicCompareExchangeLibcall(ExpectedAddr.getPointer(),
1682                                        DesiredAddr.getPointer(),
1683                                        AO, Failure);
1684   CGF.Builder.CreateCondBr(Res, ExitBB, ContBB);
1685   CGF.EmitBlock(ExitBB, /*IsFinished=*/true);
1686 }
1687 
1688 void AtomicInfo::EmitAtomicUpdateOp(
1689     llvm::AtomicOrdering AO, const llvm::function_ref<RValue(RValue)> &UpdateOp,
1690     bool IsVolatile) {
1691   auto Failure = llvm::AtomicCmpXchgInst::getStrongestFailureOrdering(AO);
1692 
1693   // Do the atomic load.
1694   auto *OldVal = EmitAtomicLoadOp(AO, IsVolatile);
1695   // For non-simple lvalues perform compare-and-swap procedure.
1696   auto *ContBB = CGF.createBasicBlock("atomic_cont");
1697   auto *ExitBB = CGF.createBasicBlock("atomic_exit");
1698   auto *CurBB = CGF.Builder.GetInsertBlock();
1699   CGF.EmitBlock(ContBB);
1700   llvm::PHINode *PHI = CGF.Builder.CreatePHI(OldVal->getType(),
1701                                              /*NumReservedValues=*/2);
1702   PHI->addIncoming(OldVal, CurBB);
1703   Address NewAtomicAddr = CreateTempAlloca();
1704   Address NewAtomicIntAddr = emitCastToAtomicIntPointer(NewAtomicAddr);
1705   if ((LVal.isBitField() && BFI.Size != ValueSizeInBits) ||
1706       requiresMemSetZero(getAtomicAddress().getElementType())) {
1707     CGF.Builder.CreateStore(PHI, NewAtomicIntAddr);
1708   }
1709   auto OldRVal = ConvertIntToValueOrAtomic(PHI, AggValueSlot::ignored(),
1710                                            SourceLocation(), /*AsValue=*/false);
1711   EmitAtomicUpdateValue(CGF, *this, OldRVal, UpdateOp, NewAtomicAddr);
1712   auto *DesiredVal = CGF.Builder.CreateLoad(NewAtomicIntAddr);
1713   // Try to write new value using cmpxchg operation
1714   auto Res = EmitAtomicCompareExchangeOp(PHI, DesiredVal, AO, Failure);
1715   PHI->addIncoming(Res.first, CGF.Builder.GetInsertBlock());
1716   CGF.Builder.CreateCondBr(Res.second, ExitBB, ContBB);
1717   CGF.EmitBlock(ExitBB, /*IsFinished=*/true);
1718 }
1719 
1720 static void EmitAtomicUpdateValue(CodeGenFunction &CGF, AtomicInfo &Atomics,
1721                                   RValue UpdateRVal, Address DesiredAddr) {
1722   LValue AtomicLVal = Atomics.getAtomicLValue();
1723   LValue DesiredLVal;
1724   // Build new lvalue for temp address
1725   if (AtomicLVal.isBitField()) {
1726     DesiredLVal =
1727         LValue::MakeBitfield(DesiredAddr, AtomicLVal.getBitFieldInfo(),
1728                              AtomicLVal.getType(),
1729                              AtomicLVal.getBaseInfo());
1730   } else if (AtomicLVal.isVectorElt()) {
1731     DesiredLVal =
1732         LValue::MakeVectorElt(DesiredAddr, AtomicLVal.getVectorIdx(),
1733                               AtomicLVal.getType(),
1734                               AtomicLVal.getBaseInfo());
1735   } else {
1736     assert(AtomicLVal.isExtVectorElt());
1737     DesiredLVal = LValue::MakeExtVectorElt(
1738         DesiredAddr, AtomicLVal.getExtVectorElts(), AtomicLVal.getType(),
1739         AtomicLVal.getBaseInfo());
1740   }
1741   DesiredLVal.setTBAAInfo(AtomicLVal.getTBAAInfo());
1742   // Store new value in the corresponding memory area
1743   assert(UpdateRVal.isScalar());
1744   CGF.EmitStoreThroughLValue(UpdateRVal, DesiredLVal);
1745 }
1746 
1747 void AtomicInfo::EmitAtomicUpdateLibcall(llvm::AtomicOrdering AO,
1748                                          RValue UpdateRVal, bool IsVolatile) {
1749   auto Failure = llvm::AtomicCmpXchgInst::getStrongestFailureOrdering(AO);
1750 
1751   Address ExpectedAddr = CreateTempAlloca();
1752 
1753   EmitAtomicLoadLibcall(ExpectedAddr.getPointer(), AO, IsVolatile);
1754   auto *ContBB = CGF.createBasicBlock("atomic_cont");
1755   auto *ExitBB = CGF.createBasicBlock("atomic_exit");
1756   CGF.EmitBlock(ContBB);
1757   Address DesiredAddr = CreateTempAlloca();
1758   if ((LVal.isBitField() && BFI.Size != ValueSizeInBits) ||
1759       requiresMemSetZero(getAtomicAddress().getElementType())) {
1760     auto *OldVal = CGF.Builder.CreateLoad(ExpectedAddr);
1761     CGF.Builder.CreateStore(OldVal, DesiredAddr);
1762   }
1763   EmitAtomicUpdateValue(CGF, *this, UpdateRVal, DesiredAddr);
1764   auto *Res =
1765       EmitAtomicCompareExchangeLibcall(ExpectedAddr.getPointer(),
1766                                        DesiredAddr.getPointer(),
1767                                        AO, Failure);
1768   CGF.Builder.CreateCondBr(Res, ExitBB, ContBB);
1769   CGF.EmitBlock(ExitBB, /*IsFinished=*/true);
1770 }
1771 
1772 void AtomicInfo::EmitAtomicUpdateOp(llvm::AtomicOrdering AO, RValue UpdateRVal,
1773                                     bool IsVolatile) {
1774   auto Failure = llvm::AtomicCmpXchgInst::getStrongestFailureOrdering(AO);
1775 
1776   // Do the atomic load.
1777   auto *OldVal = EmitAtomicLoadOp(AO, IsVolatile);
1778   // For non-simple lvalues perform compare-and-swap procedure.
1779   auto *ContBB = CGF.createBasicBlock("atomic_cont");
1780   auto *ExitBB = CGF.createBasicBlock("atomic_exit");
1781   auto *CurBB = CGF.Builder.GetInsertBlock();
1782   CGF.EmitBlock(ContBB);
1783   llvm::PHINode *PHI = CGF.Builder.CreatePHI(OldVal->getType(),
1784                                              /*NumReservedValues=*/2);
1785   PHI->addIncoming(OldVal, CurBB);
1786   Address NewAtomicAddr = CreateTempAlloca();
1787   Address NewAtomicIntAddr = emitCastToAtomicIntPointer(NewAtomicAddr);
1788   if ((LVal.isBitField() && BFI.Size != ValueSizeInBits) ||
1789       requiresMemSetZero(getAtomicAddress().getElementType())) {
1790     CGF.Builder.CreateStore(PHI, NewAtomicIntAddr);
1791   }
1792   EmitAtomicUpdateValue(CGF, *this, UpdateRVal, NewAtomicAddr);
1793   auto *DesiredVal = CGF.Builder.CreateLoad(NewAtomicIntAddr);
1794   // Try to write new value using cmpxchg operation
1795   auto Res = EmitAtomicCompareExchangeOp(PHI, DesiredVal, AO, Failure);
1796   PHI->addIncoming(Res.first, CGF.Builder.GetInsertBlock());
1797   CGF.Builder.CreateCondBr(Res.second, ExitBB, ContBB);
1798   CGF.EmitBlock(ExitBB, /*IsFinished=*/true);
1799 }
1800 
1801 void AtomicInfo::EmitAtomicUpdate(
1802     llvm::AtomicOrdering AO, const llvm::function_ref<RValue(RValue)> &UpdateOp,
1803     bool IsVolatile) {
1804   if (shouldUseLibcall()) {
1805     EmitAtomicUpdateLibcall(AO, UpdateOp, IsVolatile);
1806   } else {
1807     EmitAtomicUpdateOp(AO, UpdateOp, IsVolatile);
1808   }
1809 }
1810 
1811 void AtomicInfo::EmitAtomicUpdate(llvm::AtomicOrdering AO, RValue UpdateRVal,
1812                                   bool IsVolatile) {
1813   if (shouldUseLibcall()) {
1814     EmitAtomicUpdateLibcall(AO, UpdateRVal, IsVolatile);
1815   } else {
1816     EmitAtomicUpdateOp(AO, UpdateRVal, IsVolatile);
1817   }
1818 }
1819 
1820 void CodeGenFunction::EmitAtomicStore(RValue rvalue, LValue lvalue,
1821                                       bool isInit) {
1822   bool IsVolatile = lvalue.isVolatileQualified();
1823   llvm::AtomicOrdering AO;
1824   if (lvalue.getType()->isAtomicType()) {
1825     AO = llvm::AtomicOrdering::SequentiallyConsistent;
1826   } else {
1827     AO = llvm::AtomicOrdering::Release;
1828     IsVolatile = true;
1829   }
1830   return EmitAtomicStore(rvalue, lvalue, AO, IsVolatile, isInit);
1831 }
1832 
1833 /// Emit a store to an l-value of atomic type.
1834 ///
1835 /// Note that the r-value is expected to be an r-value *of the atomic
1836 /// type*; this means that for aggregate r-values, it should include
1837 /// storage for any padding that was necessary.
1838 void CodeGenFunction::EmitAtomicStore(RValue rvalue, LValue dest,
1839                                       llvm::AtomicOrdering AO, bool IsVolatile,
1840                                       bool isInit) {
1841   // If this is an aggregate r-value, it should agree in type except
1842   // maybe for address-space qualification.
1843   assert(!rvalue.isAggregate() ||
1844          rvalue.getAggregateAddress().getElementType()
1845            == dest.getAddress().getElementType());
1846 
1847   AtomicInfo atomics(*this, dest);
1848   LValue LVal = atomics.getAtomicLValue();
1849 
1850   // If this is an initialization, just put the value there normally.
1851   if (LVal.isSimple()) {
1852     if (isInit) {
1853       atomics.emitCopyIntoMemory(rvalue);
1854       return;
1855     }
1856 
1857     // Check whether we should use a library call.
1858     if (atomics.shouldUseLibcall()) {
1859       // Produce a source address.
1860       Address srcAddr = atomics.materializeRValue(rvalue);
1861 
1862       // void __atomic_store(size_t size, void *mem, void *val, int order)
1863       CallArgList args;
1864       args.add(RValue::get(atomics.getAtomicSizeValue()),
1865                getContext().getSizeType());
1866       args.add(RValue::get(EmitCastToVoidPtr(atomics.getAtomicPointer())),
1867                getContext().VoidPtrTy);
1868       args.add(RValue::get(EmitCastToVoidPtr(srcAddr.getPointer())),
1869                getContext().VoidPtrTy);
1870       args.add(
1871           RValue::get(llvm::ConstantInt::get(IntTy, (int)llvm::toCABI(AO))),
1872           getContext().IntTy);
1873       emitAtomicLibcall(*this, "__atomic_store", getContext().VoidTy, args);
1874       return;
1875     }
1876 
1877     // Okay, we're doing this natively.
1878     llvm::Value *intValue = atomics.convertRValueToInt(rvalue);
1879 
1880     // Do the atomic store.
1881     Address addr =
1882         atomics.emitCastToAtomicIntPointer(atomics.getAtomicAddress());
1883     intValue = Builder.CreateIntCast(
1884         intValue, addr.getElementType(), /*isSigned=*/false);
1885     llvm::StoreInst *store = Builder.CreateStore(intValue, addr);
1886 
1887     // Initializations don't need to be atomic.
1888     if (!isInit)
1889       store->setAtomic(AO);
1890 
1891     // Other decoration.
1892     if (IsVolatile)
1893       store->setVolatile(true);
1894     if (dest.getTBAAInfo())
1895       CGM.DecorateInstructionWithTBAA(store, dest.getTBAAInfo());
1896     return;
1897   }
1898 
1899   // Emit simple atomic update operation.
1900   atomics.EmitAtomicUpdate(AO, rvalue, IsVolatile);
1901 }
1902 
1903 /// Emit a compare-and-exchange op for atomic type.
1904 ///
1905 std::pair<RValue, llvm::Value *> CodeGenFunction::EmitAtomicCompareExchange(
1906     LValue Obj, RValue Expected, RValue Desired, SourceLocation Loc,
1907     llvm::AtomicOrdering Success, llvm::AtomicOrdering Failure, bool IsWeak,
1908     AggValueSlot Slot) {
1909   // If this is an aggregate r-value, it should agree in type except
1910   // maybe for address-space qualification.
1911   assert(!Expected.isAggregate() ||
1912          Expected.getAggregateAddress().getElementType() ==
1913              Obj.getAddress().getElementType());
1914   assert(!Desired.isAggregate() ||
1915          Desired.getAggregateAddress().getElementType() ==
1916              Obj.getAddress().getElementType());
1917   AtomicInfo Atomics(*this, Obj);
1918 
1919   return Atomics.EmitAtomicCompareExchange(Expected, Desired, Success, Failure,
1920                                            IsWeak);
1921 }
1922 
1923 void CodeGenFunction::EmitAtomicUpdate(
1924     LValue LVal, llvm::AtomicOrdering AO,
1925     const llvm::function_ref<RValue(RValue)> &UpdateOp, bool IsVolatile) {
1926   AtomicInfo Atomics(*this, LVal);
1927   Atomics.EmitAtomicUpdate(AO, UpdateOp, IsVolatile);
1928 }
1929 
1930 void CodeGenFunction::EmitAtomicInit(Expr *init, LValue dest) {
1931   AtomicInfo atomics(*this, dest);
1932 
1933   switch (atomics.getEvaluationKind()) {
1934   case TEK_Scalar: {
1935     llvm::Value *value = EmitScalarExpr(init);
1936     atomics.emitCopyIntoMemory(RValue::get(value));
1937     return;
1938   }
1939 
1940   case TEK_Complex: {
1941     ComplexPairTy value = EmitComplexExpr(init);
1942     atomics.emitCopyIntoMemory(RValue::getComplex(value));
1943     return;
1944   }
1945 
1946   case TEK_Aggregate: {
1947     // Fix up the destination if the initializer isn't an expression
1948     // of atomic type.
1949     bool Zeroed = false;
1950     if (!init->getType()->isAtomicType()) {
1951       Zeroed = atomics.emitMemSetZeroIfNecessary();
1952       dest = atomics.projectValue();
1953     }
1954 
1955     // Evaluate the expression directly into the destination.
1956     AggValueSlot slot = AggValueSlot::forLValue(dest,
1957                                         AggValueSlot::IsNotDestructed,
1958                                         AggValueSlot::DoesNotNeedGCBarriers,
1959                                         AggValueSlot::IsNotAliased,
1960                                         Zeroed ? AggValueSlot::IsZeroed :
1961                                                  AggValueSlot::IsNotZeroed);
1962 
1963     EmitAggExpr(init, slot);
1964     return;
1965   }
1966   }
1967   llvm_unreachable("bad evaluation kind");
1968 }
1969