xref: /freebsd/contrib/llvm-project/clang/lib/CodeGen/CGAtomic.cpp (revision 59c8e88e72633afbc47a4ace0d2170d00d51f7dc)
1 //===--- CGAtomic.cpp - Emit LLVM IR for atomic operations ----------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file contains the code for emitting atomic operations.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "CGCall.h"
14 #include "CGRecordLayout.h"
15 #include "CodeGenFunction.h"
16 #include "CodeGenModule.h"
17 #include "TargetInfo.h"
18 #include "clang/AST/ASTContext.h"
19 #include "clang/CodeGen/CGFunctionInfo.h"
20 #include "clang/Frontend/FrontendDiagnostic.h"
21 #include "llvm/ADT/DenseMap.h"
22 #include "llvm/IR/DataLayout.h"
23 #include "llvm/IR/Intrinsics.h"
24 #include "llvm/IR/Operator.h"
25 
26 using namespace clang;
27 using namespace CodeGen;
28 
29 namespace {
30   class AtomicInfo {
31     CodeGenFunction &CGF;
32     QualType AtomicTy;
33     QualType ValueTy;
34     uint64_t AtomicSizeInBits;
35     uint64_t ValueSizeInBits;
36     CharUnits AtomicAlign;
37     CharUnits ValueAlign;
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         llvm::Value *BitFieldPtr = lvalue.getBitFieldPointer();
84         auto OffsetInChars =
85             (C.toCharUnitsFromBits(OrigBFI.Offset) / lvalue.getAlignment()) *
86             lvalue.getAlignment();
87         llvm::Value *StoragePtr = CGF.Builder.CreateConstGEP1_64(
88             CGF.Int8Ty, BitFieldPtr, OffsetInChars.getQuantity());
89         StoragePtr = CGF.Builder.CreateAddrSpaceCast(
90             StoragePtr, llvm::PointerType::getUnqual(CGF.getLLVMContext()),
91             "atomic_bitfield_base");
92         BFI = OrigBFI;
93         BFI.Offset = Offset;
94         BFI.StorageSize = AtomicSizeInBits;
95         BFI.StorageOffset += OffsetInChars;
96         llvm::Type *StorageTy = CGF.Builder.getIntNTy(AtomicSizeInBits);
97         LVal = LValue::MakeBitfield(
98             Address(StoragePtr, StorageTy, lvalue.getAlignment()), BFI,
99             lvalue.getType(), lvalue.getBaseInfo(), lvalue.getTBAAInfo());
100         AtomicTy = C.getIntTypeForBitwidth(AtomicSizeInBits, OrigBFI.IsSigned);
101         if (AtomicTy.isNull()) {
102           llvm::APInt Size(
103               /*numBits=*/32,
104               C.toCharUnitsFromBits(AtomicSizeInBits).getQuantity());
105           AtomicTy =
106               C.getConstantArrayType(C.CharTy, Size, nullptr, ArrayType::Normal,
107                                      /*IndexTypeQuals=*/0);
108         }
109         AtomicAlign = ValueAlign = lvalue.getAlignment();
110       } else if (lvalue.isVectorElt()) {
111         ValueTy = lvalue.getType()->castAs<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(), cast<llvm::FixedVectorType>(
123                                   lvalue.getExtVectorAddress().getElementType())
124                                   ->getNumElements());
125         AtomicSizeInBits = C.getTypeSize(AtomicTy);
126         AtomicAlign = ValueAlign = lvalue.getAlignment();
127         LVal = lvalue;
128       }
129       UseLibcall = !C.getTargetInfo().hasBuiltinAtomic(
130           AtomicSizeInBits, C.toBits(lvalue.getAlignment()));
131     }
132 
133     QualType getAtomicType() const { return AtomicTy; }
134     QualType getValueType() const { return ValueTy; }
135     CharUnits getAtomicAlignment() const { return AtomicAlign; }
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(CGF);
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       llvm::Type *ElTy;
153       if (LVal.isSimple())
154         ElTy = LVal.getAddress(CGF).getElementType();
155       else if (LVal.isBitField())
156         ElTy = LVal.getBitFieldAddress().getElementType();
157       else if (LVal.isVectorElt())
158         ElTy = LVal.getVectorAddress().getElementType();
159       else
160         ElTy = LVal.getExtVectorAddress().getElementType();
161       return Address(getAtomicPointer(), ElTy, getAtomicAlignment());
162     }
163 
164     Address getAtomicAddressAsAtomicIntPointer() const {
165       return castToAtomicIntPointer(getAtomicAddress());
166     }
167 
168     /// Is the atomic size larger than the underlying value type?
169     ///
170     /// Note that the absence of padding does not mean that atomic
171     /// objects are completely interchangeable with non-atomic
172     /// objects: we might have promoted the alignment of a type
173     /// without making it bigger.
174     bool hasPadding() const {
175       return (ValueSizeInBits != AtomicSizeInBits);
176     }
177 
178     bool emitMemSetZeroIfNecessary() const;
179 
180     llvm::Value *getAtomicSizeValue() const {
181       CharUnits size = CGF.getContext().toCharUnitsFromBits(AtomicSizeInBits);
182       return CGF.CGM.getSize(size);
183     }
184 
185     /// Cast the given pointer to an integer pointer suitable for atomic
186     /// operations if the source.
187     Address castToAtomicIntPointer(Address Addr) const;
188 
189     /// If Addr is compatible with the iN that will be used for an atomic
190     /// operation, bitcast it. Otherwise, create a temporary that is suitable
191     /// and copy the value across.
192     Address convertToAtomicIntPointer(Address Addr) const;
193 
194     /// Turn an atomic-layout object into an r-value.
195     RValue convertAtomicTempToRValue(Address addr, AggValueSlot resultSlot,
196                                      SourceLocation loc, bool AsValue) const;
197 
198     /// Converts a rvalue to integer value.
199     llvm::Value *convertRValueToInt(RValue RVal) const;
200 
201     RValue ConvertIntToValueOrAtomic(llvm::Value *IntVal,
202                                      AggValueSlot ResultSlot,
203                                      SourceLocation Loc, bool AsValue) const;
204 
205     /// Copy an atomic r-value into atomic-layout memory.
206     void emitCopyIntoMemory(RValue rvalue) const;
207 
208     /// Project an l-value down to the value field.
209     LValue projectValue() const {
210       assert(LVal.isSimple());
211       Address addr = getAtomicAddress();
212       if (hasPadding())
213         addr = CGF.Builder.CreateStructGEP(addr, 0);
214 
215       return LValue::MakeAddr(addr, getValueType(), CGF.getContext(),
216                               LVal.getBaseInfo(), LVal.getTBAAInfo());
217     }
218 
219     /// Emits atomic load.
220     /// \returns Loaded value.
221     RValue EmitAtomicLoad(AggValueSlot ResultSlot, SourceLocation Loc,
222                           bool AsValue, llvm::AtomicOrdering AO,
223                           bool IsVolatile);
224 
225     /// Emits atomic compare-and-exchange sequence.
226     /// \param Expected Expected value.
227     /// \param Desired Desired value.
228     /// \param Success Atomic ordering for success operation.
229     /// \param Failure Atomic ordering for failed operation.
230     /// \param IsWeak true if atomic operation is weak, false otherwise.
231     /// \returns Pair of values: previous value from storage (value type) and
232     /// boolean flag (i1 type) with true if success and false otherwise.
233     std::pair<RValue, llvm::Value *>
234     EmitAtomicCompareExchange(RValue Expected, RValue Desired,
235                               llvm::AtomicOrdering Success =
236                                   llvm::AtomicOrdering::SequentiallyConsistent,
237                               llvm::AtomicOrdering Failure =
238                                   llvm::AtomicOrdering::SequentiallyConsistent,
239                               bool IsWeak = false);
240 
241     /// Emits atomic update.
242     /// \param AO Atomic ordering.
243     /// \param UpdateOp Update operation for the current lvalue.
244     void EmitAtomicUpdate(llvm::AtomicOrdering AO,
245                           const llvm::function_ref<RValue(RValue)> &UpdateOp,
246                           bool IsVolatile);
247     /// Emits atomic update.
248     /// \param AO Atomic ordering.
249     void EmitAtomicUpdate(llvm::AtomicOrdering AO, RValue UpdateRVal,
250                           bool IsVolatile);
251 
252     /// Materialize an atomic r-value in atomic-layout memory.
253     Address materializeRValue(RValue rvalue) const;
254 
255     /// Creates temp alloca for intermediate operations on atomic value.
256     Address CreateTempAlloca() const;
257   private:
258     bool requiresMemSetZero(llvm::Type *type) const;
259 
260 
261     /// Emits atomic load as a libcall.
262     void EmitAtomicLoadLibcall(llvm::Value *AddForLoaded,
263                                llvm::AtomicOrdering AO, bool IsVolatile);
264     /// Emits atomic load as LLVM instruction.
265     llvm::Value *EmitAtomicLoadOp(llvm::AtomicOrdering AO, bool IsVolatile);
266     /// Emits atomic compare-and-exchange op as a libcall.
267     llvm::Value *EmitAtomicCompareExchangeLibcall(
268         llvm::Value *ExpectedAddr, llvm::Value *DesiredAddr,
269         llvm::AtomicOrdering Success =
270             llvm::AtomicOrdering::SequentiallyConsistent,
271         llvm::AtomicOrdering Failure =
272             llvm::AtomicOrdering::SequentiallyConsistent);
273     /// Emits atomic compare-and-exchange op as LLVM instruction.
274     std::pair<llvm::Value *, llvm::Value *> EmitAtomicCompareExchangeOp(
275         llvm::Value *ExpectedVal, llvm::Value *DesiredVal,
276         llvm::AtomicOrdering Success =
277             llvm::AtomicOrdering::SequentiallyConsistent,
278         llvm::AtomicOrdering Failure =
279             llvm::AtomicOrdering::SequentiallyConsistent,
280         bool IsWeak = false);
281     /// Emit atomic update as libcalls.
282     void
283     EmitAtomicUpdateLibcall(llvm::AtomicOrdering AO,
284                             const llvm::function_ref<RValue(RValue)> &UpdateOp,
285                             bool IsVolatile);
286     /// Emit atomic update as LLVM instructions.
287     void EmitAtomicUpdateOp(llvm::AtomicOrdering AO,
288                             const llvm::function_ref<RValue(RValue)> &UpdateOp,
289                             bool IsVolatile);
290     /// Emit atomic update as libcalls.
291     void EmitAtomicUpdateLibcall(llvm::AtomicOrdering AO, RValue UpdateRVal,
292                                  bool IsVolatile);
293     /// Emit atomic update as LLVM instructions.
294     void EmitAtomicUpdateOp(llvm::AtomicOrdering AO, RValue UpdateRal,
295                             bool IsVolatile);
296   };
297 }
298 
299 Address AtomicInfo::CreateTempAlloca() const {
300   Address TempAlloca = CGF.CreateMemTemp(
301       (LVal.isBitField() && ValueSizeInBits > AtomicSizeInBits) ? ValueTy
302                                                                 : AtomicTy,
303       getAtomicAlignment(),
304       "atomic-temp");
305   // Cast to pointer to value type for bitfields.
306   if (LVal.isBitField())
307     return CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
308         TempAlloca, getAtomicAddress().getType(),
309         getAtomicAddress().getElementType());
310   return TempAlloca;
311 }
312 
313 static RValue emitAtomicLibcall(CodeGenFunction &CGF,
314                                 StringRef fnName,
315                                 QualType resultType,
316                                 CallArgList &args) {
317   const CGFunctionInfo &fnInfo =
318     CGF.CGM.getTypes().arrangeBuiltinFunctionCall(resultType, args);
319   llvm::FunctionType *fnTy = CGF.CGM.getTypes().GetFunctionType(fnInfo);
320   llvm::AttrBuilder fnAttrB(CGF.getLLVMContext());
321   fnAttrB.addAttribute(llvm::Attribute::NoUnwind);
322   fnAttrB.addAttribute(llvm::Attribute::WillReturn);
323   llvm::AttributeList fnAttrs = llvm::AttributeList::get(
324       CGF.getLLVMContext(), llvm::AttributeList::FunctionIndex, fnAttrB);
325 
326   llvm::FunctionCallee fn =
327       CGF.CGM.CreateRuntimeFunction(fnTy, fnName, fnAttrs);
328   auto callee = CGCallee::forDirect(fn);
329   return CGF.EmitCall(fnInfo, callee, ReturnValueSlot(), args);
330 }
331 
332 /// Does a store of the given IR type modify the full expected width?
333 static bool isFullSizeType(CodeGenModule &CGM, llvm::Type *type,
334                            uint64_t expectedSize) {
335   return (CGM.getDataLayout().getTypeStoreSize(type) * 8 == expectedSize);
336 }
337 
338 /// Does the atomic type require memsetting to zero before initialization?
339 ///
340 /// The IR type is provided as a way of making certain queries faster.
341 bool AtomicInfo::requiresMemSetZero(llvm::Type *type) const {
342   // If the atomic type has size padding, we definitely need a memset.
343   if (hasPadding()) return true;
344 
345   // Otherwise, do some simple heuristics to try to avoid it:
346   switch (getEvaluationKind()) {
347   // For scalars and complexes, check whether the store size of the
348   // type uses the full size.
349   case TEK_Scalar:
350     return !isFullSizeType(CGF.CGM, type, AtomicSizeInBits);
351   case TEK_Complex:
352     return !isFullSizeType(CGF.CGM, type->getStructElementType(0),
353                            AtomicSizeInBits / 2);
354 
355   // Padding in structs has an undefined bit pattern.  User beware.
356   case TEK_Aggregate:
357     return false;
358   }
359   llvm_unreachable("bad evaluation kind");
360 }
361 
362 bool AtomicInfo::emitMemSetZeroIfNecessary() const {
363   assert(LVal.isSimple());
364   Address addr = LVal.getAddress(CGF);
365   if (!requiresMemSetZero(addr.getElementType()))
366     return false;
367 
368   CGF.Builder.CreateMemSet(
369       addr.getPointer(), llvm::ConstantInt::get(CGF.Int8Ty, 0),
370       CGF.getContext().toCharUnitsFromBits(AtomicSizeInBits).getQuantity(),
371       LVal.getAlignment().getAsAlign());
372   return true;
373 }
374 
375 static void emitAtomicCmpXchg(CodeGenFunction &CGF, AtomicExpr *E, bool IsWeak,
376                               Address Dest, Address Ptr,
377                               Address Val1, Address Val2,
378                               uint64_t Size,
379                               llvm::AtomicOrdering SuccessOrder,
380                               llvm::AtomicOrdering FailureOrder,
381                               llvm::SyncScope::ID Scope) {
382   // Note that cmpxchg doesn't support weak cmpxchg, at least at the moment.
383   llvm::Value *Expected = CGF.Builder.CreateLoad(Val1);
384   llvm::Value *Desired = CGF.Builder.CreateLoad(Val2);
385 
386   llvm::AtomicCmpXchgInst *Pair = CGF.Builder.CreateAtomicCmpXchg(
387       Ptr.getPointer(), Expected, Desired, SuccessOrder, FailureOrder,
388       Scope);
389   Pair->setVolatile(E->isVolatile());
390   Pair->setWeak(IsWeak);
391 
392   // Cmp holds the result of the compare-exchange operation: true on success,
393   // false on failure.
394   llvm::Value *Old = CGF.Builder.CreateExtractValue(Pair, 0);
395   llvm::Value *Cmp = CGF.Builder.CreateExtractValue(Pair, 1);
396 
397   // This basic block is used to hold the store instruction if the operation
398   // failed.
399   llvm::BasicBlock *StoreExpectedBB =
400       CGF.createBasicBlock("cmpxchg.store_expected", CGF.CurFn);
401 
402   // This basic block is the exit point of the operation, we should end up
403   // here regardless of whether or not the operation succeeded.
404   llvm::BasicBlock *ContinueBB =
405       CGF.createBasicBlock("cmpxchg.continue", CGF.CurFn);
406 
407   // Update Expected if Expected isn't equal to Old, otherwise branch to the
408   // exit point.
409   CGF.Builder.CreateCondBr(Cmp, ContinueBB, StoreExpectedBB);
410 
411   CGF.Builder.SetInsertPoint(StoreExpectedBB);
412   // Update the memory at Expected with Old's value.
413   CGF.Builder.CreateStore(Old, Val1);
414   // Finally, branch to the exit point.
415   CGF.Builder.CreateBr(ContinueBB);
416 
417   CGF.Builder.SetInsertPoint(ContinueBB);
418   // Update the memory at Dest with Cmp's value.
419   CGF.EmitStoreOfScalar(Cmp, CGF.MakeAddrLValue(Dest, E->getType()));
420 }
421 
422 /// Given an ordering required on success, emit all possible cmpxchg
423 /// instructions to cope with the provided (but possibly only dynamically known)
424 /// FailureOrder.
425 static void emitAtomicCmpXchgFailureSet(CodeGenFunction &CGF, AtomicExpr *E,
426                                         bool IsWeak, Address Dest, Address Ptr,
427                                         Address Val1, Address Val2,
428                                         llvm::Value *FailureOrderVal,
429                                         uint64_t Size,
430                                         llvm::AtomicOrdering SuccessOrder,
431                                         llvm::SyncScope::ID Scope) {
432   llvm::AtomicOrdering FailureOrder;
433   if (llvm::ConstantInt *FO = dyn_cast<llvm::ConstantInt>(FailureOrderVal)) {
434     auto FOS = FO->getSExtValue();
435     if (!llvm::isValidAtomicOrderingCABI(FOS))
436       FailureOrder = llvm::AtomicOrdering::Monotonic;
437     else
438       switch ((llvm::AtomicOrderingCABI)FOS) {
439       case llvm::AtomicOrderingCABI::relaxed:
440       // 31.7.2.18: "The failure argument shall not be memory_order_release
441       // nor memory_order_acq_rel". Fallback to monotonic.
442       case llvm::AtomicOrderingCABI::release:
443       case llvm::AtomicOrderingCABI::acq_rel:
444         FailureOrder = llvm::AtomicOrdering::Monotonic;
445         break;
446       case llvm::AtomicOrderingCABI::consume:
447       case llvm::AtomicOrderingCABI::acquire:
448         FailureOrder = llvm::AtomicOrdering::Acquire;
449         break;
450       case llvm::AtomicOrderingCABI::seq_cst:
451         FailureOrder = llvm::AtomicOrdering::SequentiallyConsistent;
452         break;
453       }
454     // Prior to c++17, "the failure argument shall be no stronger than the
455     // success argument". This condition has been lifted and the only
456     // precondition is 31.7.2.18. Effectively treat this as a DR and skip
457     // language version checks.
458     emitAtomicCmpXchg(CGF, E, IsWeak, Dest, Ptr, Val1, Val2, Size, SuccessOrder,
459                       FailureOrder, Scope);
460     return;
461   }
462 
463   // Create all the relevant BB's
464   auto *MonotonicBB = CGF.createBasicBlock("monotonic_fail", CGF.CurFn);
465   auto *AcquireBB = CGF.createBasicBlock("acquire_fail", CGF.CurFn);
466   auto *SeqCstBB = CGF.createBasicBlock("seqcst_fail", CGF.CurFn);
467   auto *ContBB = CGF.createBasicBlock("atomic.continue", CGF.CurFn);
468 
469   // MonotonicBB is arbitrarily chosen as the default case; in practice, this
470   // doesn't matter unless someone is crazy enough to use something that
471   // doesn't fold to a constant for the ordering.
472   llvm::SwitchInst *SI = CGF.Builder.CreateSwitch(FailureOrderVal, MonotonicBB);
473   // Implemented as acquire, since it's the closest in LLVM.
474   SI->addCase(CGF.Builder.getInt32((int)llvm::AtomicOrderingCABI::consume),
475               AcquireBB);
476   SI->addCase(CGF.Builder.getInt32((int)llvm::AtomicOrderingCABI::acquire),
477               AcquireBB);
478   SI->addCase(CGF.Builder.getInt32((int)llvm::AtomicOrderingCABI::seq_cst),
479               SeqCstBB);
480 
481   // Emit all the different atomics
482   CGF.Builder.SetInsertPoint(MonotonicBB);
483   emitAtomicCmpXchg(CGF, E, IsWeak, Dest, Ptr, Val1, Val2,
484                     Size, SuccessOrder, llvm::AtomicOrdering::Monotonic, Scope);
485   CGF.Builder.CreateBr(ContBB);
486 
487   CGF.Builder.SetInsertPoint(AcquireBB);
488   emitAtomicCmpXchg(CGF, E, IsWeak, Dest, Ptr, Val1, Val2, Size, SuccessOrder,
489                     llvm::AtomicOrdering::Acquire, Scope);
490   CGF.Builder.CreateBr(ContBB);
491 
492   CGF.Builder.SetInsertPoint(SeqCstBB);
493   emitAtomicCmpXchg(CGF, E, IsWeak, Dest, Ptr, Val1, Val2, Size, SuccessOrder,
494                     llvm::AtomicOrdering::SequentiallyConsistent, Scope);
495   CGF.Builder.CreateBr(ContBB);
496 
497   CGF.Builder.SetInsertPoint(ContBB);
498 }
499 
500 /// Duplicate the atomic min/max operation in conventional IR for the builtin
501 /// variants that return the new rather than the original value.
502 static llvm::Value *EmitPostAtomicMinMax(CGBuilderTy &Builder,
503                                          AtomicExpr::AtomicOp Op,
504                                          bool IsSigned,
505                                          llvm::Value *OldVal,
506                                          llvm::Value *RHS) {
507   llvm::CmpInst::Predicate Pred;
508   switch (Op) {
509   default:
510     llvm_unreachable("Unexpected min/max operation");
511   case AtomicExpr::AO__atomic_max_fetch:
512     Pred = IsSigned ? llvm::CmpInst::ICMP_SGT : llvm::CmpInst::ICMP_UGT;
513     break;
514   case AtomicExpr::AO__atomic_min_fetch:
515     Pred = IsSigned ? llvm::CmpInst::ICMP_SLT : llvm::CmpInst::ICMP_ULT;
516     break;
517   }
518   llvm::Value *Cmp = Builder.CreateICmp(Pred, OldVal, RHS, "tst");
519   return Builder.CreateSelect(Cmp, OldVal, RHS, "newval");
520 }
521 
522 static void EmitAtomicOp(CodeGenFunction &CGF, AtomicExpr *E, Address Dest,
523                          Address Ptr, Address Val1, Address Val2,
524                          llvm::Value *IsWeak, llvm::Value *FailureOrder,
525                          uint64_t Size, llvm::AtomicOrdering Order,
526                          llvm::SyncScope::ID Scope) {
527   llvm::AtomicRMWInst::BinOp Op = llvm::AtomicRMWInst::Add;
528   bool PostOpMinMax = false;
529   unsigned PostOp = 0;
530 
531   switch (E->getOp()) {
532   case AtomicExpr::AO__c11_atomic_init:
533   case AtomicExpr::AO__opencl_atomic_init:
534     llvm_unreachable("Already handled!");
535 
536   case AtomicExpr::AO__c11_atomic_compare_exchange_strong:
537   case AtomicExpr::AO__hip_atomic_compare_exchange_strong:
538   case AtomicExpr::AO__opencl_atomic_compare_exchange_strong:
539     emitAtomicCmpXchgFailureSet(CGF, E, false, Dest, Ptr, Val1, Val2,
540                                 FailureOrder, Size, Order, Scope);
541     return;
542   case AtomicExpr::AO__c11_atomic_compare_exchange_weak:
543   case AtomicExpr::AO__opencl_atomic_compare_exchange_weak:
544   case AtomicExpr::AO__hip_atomic_compare_exchange_weak:
545     emitAtomicCmpXchgFailureSet(CGF, E, true, Dest, Ptr, Val1, Val2,
546                                 FailureOrder, Size, Order, Scope);
547     return;
548   case AtomicExpr::AO__atomic_compare_exchange:
549   case AtomicExpr::AO__atomic_compare_exchange_n: {
550     if (llvm::ConstantInt *IsWeakC = dyn_cast<llvm::ConstantInt>(IsWeak)) {
551       emitAtomicCmpXchgFailureSet(CGF, E, IsWeakC->getZExtValue(), Dest, Ptr,
552                                   Val1, Val2, FailureOrder, Size, Order, Scope);
553     } else {
554       // Create all the relevant BB's
555       llvm::BasicBlock *StrongBB =
556           CGF.createBasicBlock("cmpxchg.strong", CGF.CurFn);
557       llvm::BasicBlock *WeakBB = CGF.createBasicBlock("cmxchg.weak", CGF.CurFn);
558       llvm::BasicBlock *ContBB =
559           CGF.createBasicBlock("cmpxchg.continue", CGF.CurFn);
560 
561       llvm::SwitchInst *SI = CGF.Builder.CreateSwitch(IsWeak, WeakBB);
562       SI->addCase(CGF.Builder.getInt1(false), StrongBB);
563 
564       CGF.Builder.SetInsertPoint(StrongBB);
565       emitAtomicCmpXchgFailureSet(CGF, E, false, Dest, Ptr, Val1, Val2,
566                                   FailureOrder, Size, Order, Scope);
567       CGF.Builder.CreateBr(ContBB);
568 
569       CGF.Builder.SetInsertPoint(WeakBB);
570       emitAtomicCmpXchgFailureSet(CGF, E, true, Dest, Ptr, Val1, Val2,
571                                   FailureOrder, Size, Order, Scope);
572       CGF.Builder.CreateBr(ContBB);
573 
574       CGF.Builder.SetInsertPoint(ContBB);
575     }
576     return;
577   }
578   case AtomicExpr::AO__c11_atomic_load:
579   case AtomicExpr::AO__opencl_atomic_load:
580   case AtomicExpr::AO__hip_atomic_load:
581   case AtomicExpr::AO__atomic_load_n:
582   case AtomicExpr::AO__atomic_load: {
583     llvm::LoadInst *Load = CGF.Builder.CreateLoad(Ptr);
584     Load->setAtomic(Order, Scope);
585     Load->setVolatile(E->isVolatile());
586     CGF.Builder.CreateStore(Load, Dest);
587     return;
588   }
589 
590   case AtomicExpr::AO__c11_atomic_store:
591   case AtomicExpr::AO__opencl_atomic_store:
592   case AtomicExpr::AO__hip_atomic_store:
593   case AtomicExpr::AO__atomic_store:
594   case AtomicExpr::AO__atomic_store_n: {
595     llvm::Value *LoadVal1 = CGF.Builder.CreateLoad(Val1);
596     llvm::StoreInst *Store = CGF.Builder.CreateStore(LoadVal1, Ptr);
597     Store->setAtomic(Order, Scope);
598     Store->setVolatile(E->isVolatile());
599     return;
600   }
601 
602   case AtomicExpr::AO__c11_atomic_exchange:
603   case AtomicExpr::AO__hip_atomic_exchange:
604   case AtomicExpr::AO__opencl_atomic_exchange:
605   case AtomicExpr::AO__atomic_exchange_n:
606   case AtomicExpr::AO__atomic_exchange:
607     Op = llvm::AtomicRMWInst::Xchg;
608     break;
609 
610   case AtomicExpr::AO__atomic_add_fetch:
611     PostOp = E->getValueType()->isFloatingType() ? llvm::Instruction::FAdd
612                                                  : llvm::Instruction::Add;
613     [[fallthrough]];
614   case AtomicExpr::AO__c11_atomic_fetch_add:
615   case AtomicExpr::AO__hip_atomic_fetch_add:
616   case AtomicExpr::AO__opencl_atomic_fetch_add:
617   case AtomicExpr::AO__atomic_fetch_add:
618     Op = E->getValueType()->isFloatingType() ? llvm::AtomicRMWInst::FAdd
619                                              : llvm::AtomicRMWInst::Add;
620     break;
621 
622   case AtomicExpr::AO__atomic_sub_fetch:
623     PostOp = E->getValueType()->isFloatingType() ? llvm::Instruction::FSub
624                                                  : llvm::Instruction::Sub;
625     [[fallthrough]];
626   case AtomicExpr::AO__c11_atomic_fetch_sub:
627   case AtomicExpr::AO__hip_atomic_fetch_sub:
628   case AtomicExpr::AO__opencl_atomic_fetch_sub:
629   case AtomicExpr::AO__atomic_fetch_sub:
630     Op = E->getValueType()->isFloatingType() ? llvm::AtomicRMWInst::FSub
631                                              : llvm::AtomicRMWInst::Sub;
632     break;
633 
634   case AtomicExpr::AO__atomic_min_fetch:
635     PostOpMinMax = true;
636     [[fallthrough]];
637   case AtomicExpr::AO__c11_atomic_fetch_min:
638   case AtomicExpr::AO__hip_atomic_fetch_min:
639   case AtomicExpr::AO__opencl_atomic_fetch_min:
640   case AtomicExpr::AO__atomic_fetch_min:
641     Op = E->getValueType()->isFloatingType()
642              ? llvm::AtomicRMWInst::FMin
643              : (E->getValueType()->isSignedIntegerType()
644                     ? llvm::AtomicRMWInst::Min
645                     : llvm::AtomicRMWInst::UMin);
646     break;
647 
648   case AtomicExpr::AO__atomic_max_fetch:
649     PostOpMinMax = true;
650     [[fallthrough]];
651   case AtomicExpr::AO__c11_atomic_fetch_max:
652   case AtomicExpr::AO__hip_atomic_fetch_max:
653   case AtomicExpr::AO__opencl_atomic_fetch_max:
654   case AtomicExpr::AO__atomic_fetch_max:
655     Op = E->getValueType()->isFloatingType()
656              ? llvm::AtomicRMWInst::FMax
657              : (E->getValueType()->isSignedIntegerType()
658                     ? llvm::AtomicRMWInst::Max
659                     : llvm::AtomicRMWInst::UMax);
660     break;
661 
662   case AtomicExpr::AO__atomic_and_fetch:
663     PostOp = llvm::Instruction::And;
664     [[fallthrough]];
665   case AtomicExpr::AO__c11_atomic_fetch_and:
666   case AtomicExpr::AO__hip_atomic_fetch_and:
667   case AtomicExpr::AO__opencl_atomic_fetch_and:
668   case AtomicExpr::AO__atomic_fetch_and:
669     Op = llvm::AtomicRMWInst::And;
670     break;
671 
672   case AtomicExpr::AO__atomic_or_fetch:
673     PostOp = llvm::Instruction::Or;
674     [[fallthrough]];
675   case AtomicExpr::AO__c11_atomic_fetch_or:
676   case AtomicExpr::AO__hip_atomic_fetch_or:
677   case AtomicExpr::AO__opencl_atomic_fetch_or:
678   case AtomicExpr::AO__atomic_fetch_or:
679     Op = llvm::AtomicRMWInst::Or;
680     break;
681 
682   case AtomicExpr::AO__atomic_xor_fetch:
683     PostOp = llvm::Instruction::Xor;
684     [[fallthrough]];
685   case AtomicExpr::AO__c11_atomic_fetch_xor:
686   case AtomicExpr::AO__hip_atomic_fetch_xor:
687   case AtomicExpr::AO__opencl_atomic_fetch_xor:
688   case AtomicExpr::AO__atomic_fetch_xor:
689     Op = llvm::AtomicRMWInst::Xor;
690     break;
691 
692   case AtomicExpr::AO__atomic_nand_fetch:
693     PostOp = llvm::Instruction::And; // the NOT is special cased below
694     [[fallthrough]];
695   case AtomicExpr::AO__c11_atomic_fetch_nand:
696   case AtomicExpr::AO__atomic_fetch_nand:
697     Op = llvm::AtomicRMWInst::Nand;
698     break;
699   }
700 
701   llvm::Value *LoadVal1 = CGF.Builder.CreateLoad(Val1);
702   llvm::AtomicRMWInst *RMWI =
703       CGF.Builder.CreateAtomicRMW(Op, Ptr.getPointer(), LoadVal1, Order, Scope);
704   RMWI->setVolatile(E->isVolatile());
705 
706   // For __atomic_*_fetch operations, perform the operation again to
707   // determine the value which was written.
708   llvm::Value *Result = RMWI;
709   if (PostOpMinMax)
710     Result = EmitPostAtomicMinMax(CGF.Builder, E->getOp(),
711                                   E->getValueType()->isSignedIntegerType(),
712                                   RMWI, LoadVal1);
713   else if (PostOp)
714     Result = CGF.Builder.CreateBinOp((llvm::Instruction::BinaryOps)PostOp, RMWI,
715                                      LoadVal1);
716   if (E->getOp() == AtomicExpr::AO__atomic_nand_fetch)
717     Result = CGF.Builder.CreateNot(Result);
718   CGF.Builder.CreateStore(Result, Dest);
719 }
720 
721 // This function emits any expression (scalar, complex, or aggregate)
722 // into a temporary alloca.
723 static Address
724 EmitValToTemp(CodeGenFunction &CGF, Expr *E) {
725   Address DeclPtr = CGF.CreateMemTemp(E->getType(), ".atomictmp");
726   CGF.EmitAnyExprToMem(E, DeclPtr, E->getType().getQualifiers(),
727                        /*Init*/ true);
728   return DeclPtr;
729 }
730 
731 static void EmitAtomicOp(CodeGenFunction &CGF, AtomicExpr *Expr, Address Dest,
732                          Address Ptr, Address Val1, Address Val2,
733                          llvm::Value *IsWeak, llvm::Value *FailureOrder,
734                          uint64_t Size, llvm::AtomicOrdering Order,
735                          llvm::Value *Scope) {
736   auto ScopeModel = Expr->getScopeModel();
737 
738   // LLVM atomic instructions always have synch scope. If clang atomic
739   // expression has no scope operand, use default LLVM synch scope.
740   if (!ScopeModel) {
741     EmitAtomicOp(CGF, Expr, Dest, Ptr, Val1, Val2, IsWeak, FailureOrder, Size,
742                  Order, CGF.CGM.getLLVMContext().getOrInsertSyncScopeID(""));
743     return;
744   }
745 
746   // Handle constant scope.
747   if (auto SC = dyn_cast<llvm::ConstantInt>(Scope)) {
748     auto SCID = CGF.getTargetHooks().getLLVMSyncScopeID(
749         CGF.CGM.getLangOpts(), ScopeModel->map(SC->getZExtValue()),
750         Order, CGF.CGM.getLLVMContext());
751     EmitAtomicOp(CGF, Expr, Dest, Ptr, Val1, Val2, IsWeak, FailureOrder, Size,
752                  Order, SCID);
753     return;
754   }
755 
756   // Handle non-constant scope.
757   auto &Builder = CGF.Builder;
758   auto Scopes = ScopeModel->getRuntimeValues();
759   llvm::DenseMap<unsigned, llvm::BasicBlock *> BB;
760   for (auto S : Scopes)
761     BB[S] = CGF.createBasicBlock(getAsString(ScopeModel->map(S)), CGF.CurFn);
762 
763   llvm::BasicBlock *ContBB =
764       CGF.createBasicBlock("atomic.scope.continue", CGF.CurFn);
765 
766   auto *SC = Builder.CreateIntCast(Scope, Builder.getInt32Ty(), false);
767   // If unsupported synch scope is encountered at run time, assume a fallback
768   // synch scope value.
769   auto FallBack = ScopeModel->getFallBackValue();
770   llvm::SwitchInst *SI = Builder.CreateSwitch(SC, BB[FallBack]);
771   for (auto S : Scopes) {
772     auto *B = BB[S];
773     if (S != FallBack)
774       SI->addCase(Builder.getInt32(S), B);
775 
776     Builder.SetInsertPoint(B);
777     EmitAtomicOp(CGF, Expr, Dest, Ptr, Val1, Val2, IsWeak, FailureOrder, Size,
778                  Order,
779                  CGF.getTargetHooks().getLLVMSyncScopeID(CGF.CGM.getLangOpts(),
780                                                          ScopeModel->map(S),
781                                                          Order,
782                                                          CGF.getLLVMContext()));
783     Builder.CreateBr(ContBB);
784   }
785 
786   Builder.SetInsertPoint(ContBB);
787 }
788 
789 static void
790 AddDirectArgument(CodeGenFunction &CGF, CallArgList &Args,
791                   bool UseOptimizedLibcall, llvm::Value *Val, QualType ValTy,
792                   SourceLocation Loc, CharUnits SizeInChars) {
793   if (UseOptimizedLibcall) {
794     // Load value and pass it to the function directly.
795     CharUnits Align = CGF.getContext().getTypeAlignInChars(ValTy);
796     int64_t SizeInBits = CGF.getContext().toBits(SizeInChars);
797     ValTy =
798         CGF.getContext().getIntTypeForBitwidth(SizeInBits, /*Signed=*/false);
799     llvm::Type *ITy = llvm::IntegerType::get(CGF.getLLVMContext(), SizeInBits);
800     Address Ptr = Address(Val, ITy, Align);
801     Val = CGF.EmitLoadOfScalar(Ptr, false,
802                                CGF.getContext().getPointerType(ValTy),
803                                Loc);
804     // Coerce the value into an appropriately sized integer type.
805     Args.add(RValue::get(Val), ValTy);
806   } else {
807     // Non-optimized functions always take a reference.
808     Args.add(RValue::get(Val), CGF.getContext().VoidPtrTy);
809   }
810 }
811 
812 RValue CodeGenFunction::EmitAtomicExpr(AtomicExpr *E) {
813   QualType AtomicTy = E->getPtr()->getType()->getPointeeType();
814   QualType MemTy = AtomicTy;
815   if (const AtomicType *AT = AtomicTy->getAs<AtomicType>())
816     MemTy = AT->getValueType();
817   llvm::Value *IsWeak = nullptr, *OrderFail = nullptr;
818 
819   Address Val1 = Address::invalid();
820   Address Val2 = Address::invalid();
821   Address Dest = Address::invalid();
822   Address Ptr = EmitPointerWithAlignment(E->getPtr());
823 
824   if (E->getOp() == AtomicExpr::AO__c11_atomic_init ||
825       E->getOp() == AtomicExpr::AO__opencl_atomic_init) {
826     LValue lvalue = MakeAddrLValue(Ptr, AtomicTy);
827     EmitAtomicInit(E->getVal1(), lvalue);
828     return RValue::get(nullptr);
829   }
830 
831   auto TInfo = getContext().getTypeInfoInChars(AtomicTy);
832   uint64_t Size = TInfo.Width.getQuantity();
833   unsigned MaxInlineWidthInBits = getTarget().getMaxAtomicInlineWidth();
834 
835   bool Oversized = getContext().toBits(TInfo.Width) > MaxInlineWidthInBits;
836   bool Misaligned = (Ptr.getAlignment() % TInfo.Width) != 0;
837   bool UseLibcall = Misaligned | Oversized;
838   bool ShouldCastToIntPtrTy = true;
839 
840   CharUnits MaxInlineWidth =
841       getContext().toCharUnitsFromBits(MaxInlineWidthInBits);
842 
843   DiagnosticsEngine &Diags = CGM.getDiags();
844 
845   if (Misaligned) {
846     Diags.Report(E->getBeginLoc(), diag::warn_atomic_op_misaligned)
847         << (int)TInfo.Width.getQuantity()
848         << (int)Ptr.getAlignment().getQuantity();
849   }
850 
851   if (Oversized) {
852     Diags.Report(E->getBeginLoc(), diag::warn_atomic_op_oversized)
853         << (int)TInfo.Width.getQuantity() << (int)MaxInlineWidth.getQuantity();
854   }
855 
856   llvm::Value *Order = EmitScalarExpr(E->getOrder());
857   llvm::Value *Scope =
858       E->getScopeModel() ? EmitScalarExpr(E->getScope()) : nullptr;
859 
860   switch (E->getOp()) {
861   case AtomicExpr::AO__c11_atomic_init:
862   case AtomicExpr::AO__opencl_atomic_init:
863     llvm_unreachable("Already handled above with EmitAtomicInit!");
864 
865   case AtomicExpr::AO__c11_atomic_load:
866   case AtomicExpr::AO__opencl_atomic_load:
867   case AtomicExpr::AO__hip_atomic_load:
868   case AtomicExpr::AO__atomic_load_n:
869     break;
870 
871   case AtomicExpr::AO__atomic_load:
872     Dest = EmitPointerWithAlignment(E->getVal1());
873     break;
874 
875   case AtomicExpr::AO__atomic_store:
876     Val1 = EmitPointerWithAlignment(E->getVal1());
877     break;
878 
879   case AtomicExpr::AO__atomic_exchange:
880     Val1 = EmitPointerWithAlignment(E->getVal1());
881     Dest = EmitPointerWithAlignment(E->getVal2());
882     break;
883 
884   case AtomicExpr::AO__c11_atomic_compare_exchange_strong:
885   case AtomicExpr::AO__c11_atomic_compare_exchange_weak:
886   case AtomicExpr::AO__opencl_atomic_compare_exchange_strong:
887   case AtomicExpr::AO__hip_atomic_compare_exchange_strong:
888   case AtomicExpr::AO__opencl_atomic_compare_exchange_weak:
889   case AtomicExpr::AO__hip_atomic_compare_exchange_weak:
890   case AtomicExpr::AO__atomic_compare_exchange_n:
891   case AtomicExpr::AO__atomic_compare_exchange:
892     Val1 = EmitPointerWithAlignment(E->getVal1());
893     if (E->getOp() == AtomicExpr::AO__atomic_compare_exchange)
894       Val2 = EmitPointerWithAlignment(E->getVal2());
895     else
896       Val2 = EmitValToTemp(*this, E->getVal2());
897     OrderFail = EmitScalarExpr(E->getOrderFail());
898     if (E->getOp() == AtomicExpr::AO__atomic_compare_exchange_n ||
899         E->getOp() == AtomicExpr::AO__atomic_compare_exchange)
900       IsWeak = EmitScalarExpr(E->getWeak());
901     break;
902 
903   case AtomicExpr::AO__c11_atomic_fetch_add:
904   case AtomicExpr::AO__c11_atomic_fetch_sub:
905   case AtomicExpr::AO__hip_atomic_fetch_add:
906   case AtomicExpr::AO__hip_atomic_fetch_sub:
907   case AtomicExpr::AO__opencl_atomic_fetch_add:
908   case AtomicExpr::AO__opencl_atomic_fetch_sub:
909     if (MemTy->isPointerType()) {
910       // For pointer arithmetic, we're required to do a bit of math:
911       // adding 1 to an int* is not the same as adding 1 to a uintptr_t.
912       // ... but only for the C11 builtins. The GNU builtins expect the
913       // user to multiply by sizeof(T).
914       QualType Val1Ty = E->getVal1()->getType();
915       llvm::Value *Val1Scalar = EmitScalarExpr(E->getVal1());
916       CharUnits PointeeIncAmt =
917           getContext().getTypeSizeInChars(MemTy->getPointeeType());
918       Val1Scalar = Builder.CreateMul(Val1Scalar, CGM.getSize(PointeeIncAmt));
919       auto Temp = CreateMemTemp(Val1Ty, ".atomictmp");
920       Val1 = Temp;
921       EmitStoreOfScalar(Val1Scalar, MakeAddrLValue(Temp, Val1Ty));
922       break;
923     }
924     [[fallthrough]];
925   case AtomicExpr::AO__atomic_fetch_add:
926   case AtomicExpr::AO__atomic_fetch_max:
927   case AtomicExpr::AO__atomic_fetch_min:
928   case AtomicExpr::AO__atomic_fetch_sub:
929   case AtomicExpr::AO__atomic_add_fetch:
930   case AtomicExpr::AO__atomic_max_fetch:
931   case AtomicExpr::AO__atomic_min_fetch:
932   case AtomicExpr::AO__atomic_sub_fetch:
933   case AtomicExpr::AO__c11_atomic_fetch_max:
934   case AtomicExpr::AO__c11_atomic_fetch_min:
935   case AtomicExpr::AO__opencl_atomic_fetch_max:
936   case AtomicExpr::AO__opencl_atomic_fetch_min:
937   case AtomicExpr::AO__hip_atomic_fetch_max:
938   case AtomicExpr::AO__hip_atomic_fetch_min:
939     ShouldCastToIntPtrTy = !MemTy->isFloatingType();
940     [[fallthrough]];
941 
942   case AtomicExpr::AO__c11_atomic_store:
943   case AtomicExpr::AO__c11_atomic_exchange:
944   case AtomicExpr::AO__opencl_atomic_store:
945   case AtomicExpr::AO__hip_atomic_store:
946   case AtomicExpr::AO__opencl_atomic_exchange:
947   case AtomicExpr::AO__hip_atomic_exchange:
948   case AtomicExpr::AO__atomic_store_n:
949   case AtomicExpr::AO__atomic_exchange_n:
950   case AtomicExpr::AO__c11_atomic_fetch_and:
951   case AtomicExpr::AO__c11_atomic_fetch_or:
952   case AtomicExpr::AO__c11_atomic_fetch_xor:
953   case AtomicExpr::AO__c11_atomic_fetch_nand:
954   case AtomicExpr::AO__opencl_atomic_fetch_and:
955   case AtomicExpr::AO__opencl_atomic_fetch_or:
956   case AtomicExpr::AO__opencl_atomic_fetch_xor:
957   case AtomicExpr::AO__atomic_fetch_and:
958   case AtomicExpr::AO__hip_atomic_fetch_and:
959   case AtomicExpr::AO__atomic_fetch_or:
960   case AtomicExpr::AO__hip_atomic_fetch_or:
961   case AtomicExpr::AO__atomic_fetch_xor:
962   case AtomicExpr::AO__hip_atomic_fetch_xor:
963   case AtomicExpr::AO__atomic_fetch_nand:
964   case AtomicExpr::AO__atomic_and_fetch:
965   case AtomicExpr::AO__atomic_or_fetch:
966   case AtomicExpr::AO__atomic_xor_fetch:
967   case AtomicExpr::AO__atomic_nand_fetch:
968     Val1 = EmitValToTemp(*this, E->getVal1());
969     break;
970   }
971 
972   QualType RValTy = E->getType().getUnqualifiedType();
973 
974   // The inlined atomics only function on iN types, where N is a power of 2. We
975   // need to make sure (via temporaries if necessary) that all incoming values
976   // are compatible.
977   LValue AtomicVal = MakeAddrLValue(Ptr, AtomicTy);
978   AtomicInfo Atomics(*this, AtomicVal);
979 
980   if (ShouldCastToIntPtrTy) {
981     Ptr = Atomics.castToAtomicIntPointer(Ptr);
982     if (Val1.isValid())
983       Val1 = Atomics.convertToAtomicIntPointer(Val1);
984     if (Val2.isValid())
985       Val2 = Atomics.convertToAtomicIntPointer(Val2);
986   }
987   if (Dest.isValid()) {
988     if (ShouldCastToIntPtrTy)
989       Dest = Atomics.castToAtomicIntPointer(Dest);
990   } else if (E->isCmpXChg())
991     Dest = CreateMemTemp(RValTy, "cmpxchg.bool");
992   else if (!RValTy->isVoidType()) {
993     Dest = Atomics.CreateTempAlloca();
994     if (ShouldCastToIntPtrTy)
995       Dest = Atomics.castToAtomicIntPointer(Dest);
996   }
997 
998   // Use a library call.  See: http://gcc.gnu.org/wiki/Atomic/GCCMM/LIbrary .
999   if (UseLibcall) {
1000     bool UseOptimizedLibcall = false;
1001     switch (E->getOp()) {
1002     case AtomicExpr::AO__c11_atomic_init:
1003     case AtomicExpr::AO__opencl_atomic_init:
1004       llvm_unreachable("Already handled above with EmitAtomicInit!");
1005 
1006     case AtomicExpr::AO__c11_atomic_fetch_add:
1007     case AtomicExpr::AO__opencl_atomic_fetch_add:
1008     case AtomicExpr::AO__atomic_fetch_add:
1009     case AtomicExpr::AO__hip_atomic_fetch_add:
1010     case AtomicExpr::AO__c11_atomic_fetch_and:
1011     case AtomicExpr::AO__opencl_atomic_fetch_and:
1012     case AtomicExpr::AO__hip_atomic_fetch_and:
1013     case AtomicExpr::AO__atomic_fetch_and:
1014     case AtomicExpr::AO__c11_atomic_fetch_or:
1015     case AtomicExpr::AO__opencl_atomic_fetch_or:
1016     case AtomicExpr::AO__hip_atomic_fetch_or:
1017     case AtomicExpr::AO__atomic_fetch_or:
1018     case AtomicExpr::AO__c11_atomic_fetch_nand:
1019     case AtomicExpr::AO__atomic_fetch_nand:
1020     case AtomicExpr::AO__c11_atomic_fetch_sub:
1021     case AtomicExpr::AO__opencl_atomic_fetch_sub:
1022     case AtomicExpr::AO__atomic_fetch_sub:
1023     case AtomicExpr::AO__hip_atomic_fetch_sub:
1024     case AtomicExpr::AO__c11_atomic_fetch_xor:
1025     case AtomicExpr::AO__opencl_atomic_fetch_xor:
1026     case AtomicExpr::AO__opencl_atomic_fetch_min:
1027     case AtomicExpr::AO__opencl_atomic_fetch_max:
1028     case AtomicExpr::AO__atomic_fetch_xor:
1029     case AtomicExpr::AO__hip_atomic_fetch_xor:
1030     case AtomicExpr::AO__c11_atomic_fetch_max:
1031     case AtomicExpr::AO__c11_atomic_fetch_min:
1032     case AtomicExpr::AO__atomic_add_fetch:
1033     case AtomicExpr::AO__atomic_and_fetch:
1034     case AtomicExpr::AO__atomic_nand_fetch:
1035     case AtomicExpr::AO__atomic_or_fetch:
1036     case AtomicExpr::AO__atomic_sub_fetch:
1037     case AtomicExpr::AO__atomic_xor_fetch:
1038     case AtomicExpr::AO__atomic_fetch_max:
1039     case AtomicExpr::AO__hip_atomic_fetch_max:
1040     case AtomicExpr::AO__atomic_fetch_min:
1041     case AtomicExpr::AO__hip_atomic_fetch_min:
1042     case AtomicExpr::AO__atomic_max_fetch:
1043     case AtomicExpr::AO__atomic_min_fetch:
1044       // For these, only library calls for certain sizes exist.
1045       UseOptimizedLibcall = true;
1046       break;
1047 
1048     case AtomicExpr::AO__atomic_load:
1049     case AtomicExpr::AO__atomic_store:
1050     case AtomicExpr::AO__atomic_exchange:
1051     case AtomicExpr::AO__atomic_compare_exchange:
1052       // Use the generic version if we don't know that the operand will be
1053       // suitably aligned for the optimized version.
1054       if (Misaligned)
1055         break;
1056       [[fallthrough]];
1057     case AtomicExpr::AO__c11_atomic_load:
1058     case AtomicExpr::AO__c11_atomic_store:
1059     case AtomicExpr::AO__c11_atomic_exchange:
1060     case AtomicExpr::AO__c11_atomic_compare_exchange_weak:
1061     case AtomicExpr::AO__c11_atomic_compare_exchange_strong:
1062     case AtomicExpr::AO__hip_atomic_compare_exchange_strong:
1063     case AtomicExpr::AO__opencl_atomic_load:
1064     case AtomicExpr::AO__hip_atomic_load:
1065     case AtomicExpr::AO__opencl_atomic_store:
1066     case AtomicExpr::AO__hip_atomic_store:
1067     case AtomicExpr::AO__opencl_atomic_exchange:
1068     case AtomicExpr::AO__hip_atomic_exchange:
1069     case AtomicExpr::AO__opencl_atomic_compare_exchange_weak:
1070     case AtomicExpr::AO__hip_atomic_compare_exchange_weak:
1071     case AtomicExpr::AO__opencl_atomic_compare_exchange_strong:
1072     case AtomicExpr::AO__atomic_load_n:
1073     case AtomicExpr::AO__atomic_store_n:
1074     case AtomicExpr::AO__atomic_exchange_n:
1075     case AtomicExpr::AO__atomic_compare_exchange_n:
1076       // Only use optimized library calls for sizes for which they exist.
1077       // FIXME: Size == 16 optimized library functions exist too.
1078       if (Size == 1 || Size == 2 || Size == 4 || Size == 8)
1079         UseOptimizedLibcall = true;
1080       break;
1081     }
1082 
1083     CallArgList Args;
1084     if (!UseOptimizedLibcall) {
1085       // For non-optimized library calls, the size is the first parameter
1086       Args.add(RValue::get(llvm::ConstantInt::get(SizeTy, Size)),
1087                getContext().getSizeType());
1088     }
1089     // Atomic address is the first or second parameter
1090     // The OpenCL atomic library functions only accept pointer arguments to
1091     // generic address space.
1092     auto CastToGenericAddrSpace = [&](llvm::Value *V, QualType PT) {
1093       if (!E->isOpenCL())
1094         return V;
1095       auto AS = PT->castAs<PointerType>()->getPointeeType().getAddressSpace();
1096       if (AS == LangAS::opencl_generic)
1097         return V;
1098       auto DestAS = getContext().getTargetAddressSpace(LangAS::opencl_generic);
1099       auto *DestType = llvm::PointerType::get(getLLVMContext(), DestAS);
1100 
1101       return getTargetHooks().performAddrSpaceCast(
1102           *this, V, AS, LangAS::opencl_generic, DestType, false);
1103     };
1104 
1105     Args.add(RValue::get(CastToGenericAddrSpace(Ptr.getPointer(),
1106                                                 E->getPtr()->getType())),
1107              getContext().VoidPtrTy);
1108 
1109     std::string LibCallName;
1110     QualType LoweredMemTy =
1111       MemTy->isPointerType() ? getContext().getIntPtrType() : MemTy;
1112     QualType RetTy;
1113     bool HaveRetTy = false;
1114     llvm::Instruction::BinaryOps PostOp = (llvm::Instruction::BinaryOps)0;
1115     bool PostOpMinMax = false;
1116     switch (E->getOp()) {
1117     case AtomicExpr::AO__c11_atomic_init:
1118     case AtomicExpr::AO__opencl_atomic_init:
1119       llvm_unreachable("Already handled!");
1120 
1121     // There is only one libcall for compare an exchange, because there is no
1122     // optimisation benefit possible from a libcall version of a weak compare
1123     // and exchange.
1124     // bool __atomic_compare_exchange(size_t size, void *mem, void *expected,
1125     //                                void *desired, int success, int failure)
1126     // bool __atomic_compare_exchange_N(T *mem, T *expected, T desired,
1127     //                                  int success, int failure)
1128     case AtomicExpr::AO__c11_atomic_compare_exchange_weak:
1129     case AtomicExpr::AO__c11_atomic_compare_exchange_strong:
1130     case AtomicExpr::AO__opencl_atomic_compare_exchange_weak:
1131     case AtomicExpr::AO__hip_atomic_compare_exchange_weak:
1132     case AtomicExpr::AO__opencl_atomic_compare_exchange_strong:
1133     case AtomicExpr::AO__hip_atomic_compare_exchange_strong:
1134     case AtomicExpr::AO__atomic_compare_exchange:
1135     case AtomicExpr::AO__atomic_compare_exchange_n:
1136       LibCallName = "__atomic_compare_exchange";
1137       RetTy = getContext().BoolTy;
1138       HaveRetTy = true;
1139       Args.add(RValue::get(CastToGenericAddrSpace(Val1.getPointer(),
1140                                                   E->getVal1()->getType())),
1141                getContext().VoidPtrTy);
1142       AddDirectArgument(*this, Args, UseOptimizedLibcall, Val2.getPointer(),
1143                         MemTy, E->getExprLoc(), TInfo.Width);
1144       Args.add(RValue::get(Order), getContext().IntTy);
1145       Order = OrderFail;
1146       break;
1147     // void __atomic_exchange(size_t size, void *mem, void *val, void *return,
1148     //                        int order)
1149     // T __atomic_exchange_N(T *mem, T val, int order)
1150     case AtomicExpr::AO__c11_atomic_exchange:
1151     case AtomicExpr::AO__opencl_atomic_exchange:
1152     case AtomicExpr::AO__atomic_exchange_n:
1153     case AtomicExpr::AO__atomic_exchange:
1154     case AtomicExpr::AO__hip_atomic_exchange:
1155       LibCallName = "__atomic_exchange";
1156       AddDirectArgument(*this, Args, UseOptimizedLibcall, Val1.getPointer(),
1157                         MemTy, E->getExprLoc(), TInfo.Width);
1158       break;
1159     // void __atomic_store(size_t size, void *mem, void *val, int order)
1160     // void __atomic_store_N(T *mem, T val, int order)
1161     case AtomicExpr::AO__c11_atomic_store:
1162     case AtomicExpr::AO__opencl_atomic_store:
1163     case AtomicExpr::AO__hip_atomic_store:
1164     case AtomicExpr::AO__atomic_store:
1165     case AtomicExpr::AO__atomic_store_n:
1166       LibCallName = "__atomic_store";
1167       RetTy = getContext().VoidTy;
1168       HaveRetTy = true;
1169       AddDirectArgument(*this, Args, UseOptimizedLibcall, Val1.getPointer(),
1170                         MemTy, E->getExprLoc(), TInfo.Width);
1171       break;
1172     // void __atomic_load(size_t size, void *mem, void *return, int order)
1173     // T __atomic_load_N(T *mem, int order)
1174     case AtomicExpr::AO__c11_atomic_load:
1175     case AtomicExpr::AO__opencl_atomic_load:
1176     case AtomicExpr::AO__hip_atomic_load:
1177     case AtomicExpr::AO__atomic_load:
1178     case AtomicExpr::AO__atomic_load_n:
1179       LibCallName = "__atomic_load";
1180       break;
1181     // T __atomic_add_fetch_N(T *mem, T val, int order)
1182     // T __atomic_fetch_add_N(T *mem, T val, int order)
1183     case AtomicExpr::AO__atomic_add_fetch:
1184       PostOp = llvm::Instruction::Add;
1185       [[fallthrough]];
1186     case AtomicExpr::AO__c11_atomic_fetch_add:
1187     case AtomicExpr::AO__opencl_atomic_fetch_add:
1188     case AtomicExpr::AO__atomic_fetch_add:
1189     case AtomicExpr::AO__hip_atomic_fetch_add:
1190       LibCallName = "__atomic_fetch_add";
1191       AddDirectArgument(*this, Args, UseOptimizedLibcall, Val1.getPointer(),
1192                         LoweredMemTy, E->getExprLoc(), TInfo.Width);
1193       break;
1194     // T __atomic_and_fetch_N(T *mem, T val, int order)
1195     // T __atomic_fetch_and_N(T *mem, T val, int order)
1196     case AtomicExpr::AO__atomic_and_fetch:
1197       PostOp = llvm::Instruction::And;
1198       [[fallthrough]];
1199     case AtomicExpr::AO__c11_atomic_fetch_and:
1200     case AtomicExpr::AO__opencl_atomic_fetch_and:
1201     case AtomicExpr::AO__hip_atomic_fetch_and:
1202     case AtomicExpr::AO__atomic_fetch_and:
1203       LibCallName = "__atomic_fetch_and";
1204       AddDirectArgument(*this, Args, UseOptimizedLibcall, Val1.getPointer(),
1205                         MemTy, E->getExprLoc(), TInfo.Width);
1206       break;
1207     // T __atomic_or_fetch_N(T *mem, T val, int order)
1208     // T __atomic_fetch_or_N(T *mem, T val, int order)
1209     case AtomicExpr::AO__atomic_or_fetch:
1210       PostOp = llvm::Instruction::Or;
1211       [[fallthrough]];
1212     case AtomicExpr::AO__c11_atomic_fetch_or:
1213     case AtomicExpr::AO__opencl_atomic_fetch_or:
1214     case AtomicExpr::AO__hip_atomic_fetch_or:
1215     case AtomicExpr::AO__atomic_fetch_or:
1216       LibCallName = "__atomic_fetch_or";
1217       AddDirectArgument(*this, Args, UseOptimizedLibcall, Val1.getPointer(),
1218                         MemTy, E->getExprLoc(), TInfo.Width);
1219       break;
1220     // T __atomic_sub_fetch_N(T *mem, T val, int order)
1221     // T __atomic_fetch_sub_N(T *mem, T val, int order)
1222     case AtomicExpr::AO__atomic_sub_fetch:
1223       PostOp = llvm::Instruction::Sub;
1224       [[fallthrough]];
1225     case AtomicExpr::AO__c11_atomic_fetch_sub:
1226     case AtomicExpr::AO__opencl_atomic_fetch_sub:
1227     case AtomicExpr::AO__hip_atomic_fetch_sub:
1228     case AtomicExpr::AO__atomic_fetch_sub:
1229       LibCallName = "__atomic_fetch_sub";
1230       AddDirectArgument(*this, Args, UseOptimizedLibcall, Val1.getPointer(),
1231                         LoweredMemTy, E->getExprLoc(), TInfo.Width);
1232       break;
1233     // T __atomic_xor_fetch_N(T *mem, T val, int order)
1234     // T __atomic_fetch_xor_N(T *mem, T val, int order)
1235     case AtomicExpr::AO__atomic_xor_fetch:
1236       PostOp = llvm::Instruction::Xor;
1237       [[fallthrough]];
1238     case AtomicExpr::AO__c11_atomic_fetch_xor:
1239     case AtomicExpr::AO__opencl_atomic_fetch_xor:
1240     case AtomicExpr::AO__hip_atomic_fetch_xor:
1241     case AtomicExpr::AO__atomic_fetch_xor:
1242       LibCallName = "__atomic_fetch_xor";
1243       AddDirectArgument(*this, Args, UseOptimizedLibcall, Val1.getPointer(),
1244                         MemTy, E->getExprLoc(), TInfo.Width);
1245       break;
1246     case AtomicExpr::AO__atomic_min_fetch:
1247       PostOpMinMax = true;
1248       [[fallthrough]];
1249     case AtomicExpr::AO__c11_atomic_fetch_min:
1250     case AtomicExpr::AO__atomic_fetch_min:
1251     case AtomicExpr::AO__hip_atomic_fetch_min:
1252     case AtomicExpr::AO__opencl_atomic_fetch_min:
1253       LibCallName = E->getValueType()->isSignedIntegerType()
1254                         ? "__atomic_fetch_min"
1255                         : "__atomic_fetch_umin";
1256       AddDirectArgument(*this, Args, UseOptimizedLibcall, Val1.getPointer(),
1257                         LoweredMemTy, E->getExprLoc(), TInfo.Width);
1258       break;
1259     case AtomicExpr::AO__atomic_max_fetch:
1260       PostOpMinMax = true;
1261       [[fallthrough]];
1262     case AtomicExpr::AO__c11_atomic_fetch_max:
1263     case AtomicExpr::AO__atomic_fetch_max:
1264     case AtomicExpr::AO__hip_atomic_fetch_max:
1265     case AtomicExpr::AO__opencl_atomic_fetch_max:
1266       LibCallName = E->getValueType()->isSignedIntegerType()
1267                         ? "__atomic_fetch_max"
1268                         : "__atomic_fetch_umax";
1269       AddDirectArgument(*this, Args, UseOptimizedLibcall, Val1.getPointer(),
1270                         LoweredMemTy, E->getExprLoc(), TInfo.Width);
1271       break;
1272     // T __atomic_nand_fetch_N(T *mem, T val, int order)
1273     // T __atomic_fetch_nand_N(T *mem, T val, int order)
1274     case AtomicExpr::AO__atomic_nand_fetch:
1275       PostOp = llvm::Instruction::And; // the NOT is special cased below
1276       [[fallthrough]];
1277     case AtomicExpr::AO__c11_atomic_fetch_nand:
1278     case AtomicExpr::AO__atomic_fetch_nand:
1279       LibCallName = "__atomic_fetch_nand";
1280       AddDirectArgument(*this, Args, UseOptimizedLibcall, Val1.getPointer(),
1281                         MemTy, E->getExprLoc(), TInfo.Width);
1282       break;
1283     }
1284 
1285     if (E->isOpenCL()) {
1286       LibCallName = std::string("__opencl") +
1287           StringRef(LibCallName).drop_front(1).str();
1288 
1289     }
1290     // Optimized functions have the size in their name.
1291     if (UseOptimizedLibcall)
1292       LibCallName += "_" + llvm::utostr(Size);
1293     // By default, assume we return a value of the atomic type.
1294     if (!HaveRetTy) {
1295       if (UseOptimizedLibcall) {
1296         // Value is returned directly.
1297         // The function returns an appropriately sized integer type.
1298         RetTy = getContext().getIntTypeForBitwidth(
1299             getContext().toBits(TInfo.Width), /*Signed=*/false);
1300       } else {
1301         // Value is returned through parameter before the order.
1302         RetTy = getContext().VoidTy;
1303         Args.add(RValue::get(Dest.getPointer()), getContext().VoidPtrTy);
1304       }
1305     }
1306     // order is always the last parameter
1307     Args.add(RValue::get(Order),
1308              getContext().IntTy);
1309     if (E->isOpenCL())
1310       Args.add(RValue::get(Scope), getContext().IntTy);
1311 
1312     // PostOp is only needed for the atomic_*_fetch operations, and
1313     // thus is only needed for and implemented in the
1314     // UseOptimizedLibcall codepath.
1315     assert(UseOptimizedLibcall || (!PostOp && !PostOpMinMax));
1316 
1317     RValue Res = emitAtomicLibcall(*this, LibCallName, RetTy, Args);
1318     // The value is returned directly from the libcall.
1319     if (E->isCmpXChg())
1320       return Res;
1321 
1322     // The value is returned directly for optimized libcalls but the expr
1323     // provided an out-param.
1324     if (UseOptimizedLibcall && Res.getScalarVal()) {
1325       llvm::Value *ResVal = Res.getScalarVal();
1326       if (PostOpMinMax) {
1327         llvm::Value *LoadVal1 = Args[1].getRValue(*this).getScalarVal();
1328         ResVal = EmitPostAtomicMinMax(Builder, E->getOp(),
1329                                       E->getValueType()->isSignedIntegerType(),
1330                                       ResVal, LoadVal1);
1331       } else if (PostOp) {
1332         llvm::Value *LoadVal1 = Args[1].getRValue(*this).getScalarVal();
1333         ResVal = Builder.CreateBinOp(PostOp, ResVal, LoadVal1);
1334       }
1335       if (E->getOp() == AtomicExpr::AO__atomic_nand_fetch)
1336         ResVal = Builder.CreateNot(ResVal);
1337 
1338       Builder.CreateStore(ResVal, Dest.withElementType(ResVal->getType()));
1339     }
1340 
1341     if (RValTy->isVoidType())
1342       return RValue::get(nullptr);
1343 
1344     return convertTempToRValue(Dest.withElementType(ConvertTypeForMem(RValTy)),
1345                                RValTy, E->getExprLoc());
1346   }
1347 
1348   bool IsStore = E->getOp() == AtomicExpr::AO__c11_atomic_store ||
1349                  E->getOp() == AtomicExpr::AO__opencl_atomic_store ||
1350                  E->getOp() == AtomicExpr::AO__hip_atomic_store ||
1351                  E->getOp() == AtomicExpr::AO__atomic_store ||
1352                  E->getOp() == AtomicExpr::AO__atomic_store_n;
1353   bool IsLoad = E->getOp() == AtomicExpr::AO__c11_atomic_load ||
1354                 E->getOp() == AtomicExpr::AO__opencl_atomic_load ||
1355                 E->getOp() == AtomicExpr::AO__hip_atomic_load ||
1356                 E->getOp() == AtomicExpr::AO__atomic_load ||
1357                 E->getOp() == AtomicExpr::AO__atomic_load_n;
1358 
1359   if (isa<llvm::ConstantInt>(Order)) {
1360     auto ord = cast<llvm::ConstantInt>(Order)->getZExtValue();
1361     // We should not ever get to a case where the ordering isn't a valid C ABI
1362     // value, but it's hard to enforce that in general.
1363     if (llvm::isValidAtomicOrderingCABI(ord))
1364       switch ((llvm::AtomicOrderingCABI)ord) {
1365       case llvm::AtomicOrderingCABI::relaxed:
1366         EmitAtomicOp(*this, E, Dest, Ptr, Val1, Val2, IsWeak, OrderFail, Size,
1367                      llvm::AtomicOrdering::Monotonic, Scope);
1368         break;
1369       case llvm::AtomicOrderingCABI::consume:
1370       case llvm::AtomicOrderingCABI::acquire:
1371         if (IsStore)
1372           break; // Avoid crashing on code with undefined behavior
1373         EmitAtomicOp(*this, E, Dest, Ptr, Val1, Val2, IsWeak, OrderFail, Size,
1374                      llvm::AtomicOrdering::Acquire, Scope);
1375         break;
1376       case llvm::AtomicOrderingCABI::release:
1377         if (IsLoad)
1378           break; // Avoid crashing on code with undefined behavior
1379         EmitAtomicOp(*this, E, Dest, Ptr, Val1, Val2, IsWeak, OrderFail, Size,
1380                      llvm::AtomicOrdering::Release, Scope);
1381         break;
1382       case llvm::AtomicOrderingCABI::acq_rel:
1383         if (IsLoad || IsStore)
1384           break; // Avoid crashing on code with undefined behavior
1385         EmitAtomicOp(*this, E, Dest, Ptr, Val1, Val2, IsWeak, OrderFail, Size,
1386                      llvm::AtomicOrdering::AcquireRelease, Scope);
1387         break;
1388       case llvm::AtomicOrderingCABI::seq_cst:
1389         EmitAtomicOp(*this, E, Dest, Ptr, Val1, Val2, IsWeak, OrderFail, Size,
1390                      llvm::AtomicOrdering::SequentiallyConsistent, Scope);
1391         break;
1392       }
1393     if (RValTy->isVoidType())
1394       return RValue::get(nullptr);
1395 
1396     return convertTempToRValue(Dest.withElementType(ConvertTypeForMem(RValTy)),
1397                                RValTy, E->getExprLoc());
1398   }
1399 
1400   // Long case, when Order isn't obviously constant.
1401 
1402   // Create all the relevant BB's
1403   llvm::BasicBlock *MonotonicBB = nullptr, *AcquireBB = nullptr,
1404                    *ReleaseBB = nullptr, *AcqRelBB = nullptr,
1405                    *SeqCstBB = nullptr;
1406   MonotonicBB = createBasicBlock("monotonic", CurFn);
1407   if (!IsStore)
1408     AcquireBB = createBasicBlock("acquire", CurFn);
1409   if (!IsLoad)
1410     ReleaseBB = createBasicBlock("release", CurFn);
1411   if (!IsLoad && !IsStore)
1412     AcqRelBB = createBasicBlock("acqrel", CurFn);
1413   SeqCstBB = createBasicBlock("seqcst", CurFn);
1414   llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn);
1415 
1416   // Create the switch for the split
1417   // MonotonicBB is arbitrarily chosen as the default case; in practice, this
1418   // doesn't matter unless someone is crazy enough to use something that
1419   // doesn't fold to a constant for the ordering.
1420   Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false);
1421   llvm::SwitchInst *SI = Builder.CreateSwitch(Order, MonotonicBB);
1422 
1423   // Emit all the different atomics
1424   Builder.SetInsertPoint(MonotonicBB);
1425   EmitAtomicOp(*this, E, Dest, Ptr, Val1, Val2, IsWeak, OrderFail, Size,
1426                llvm::AtomicOrdering::Monotonic, Scope);
1427   Builder.CreateBr(ContBB);
1428   if (!IsStore) {
1429     Builder.SetInsertPoint(AcquireBB);
1430     EmitAtomicOp(*this, E, Dest, Ptr, Val1, Val2, IsWeak, OrderFail, Size,
1431                  llvm::AtomicOrdering::Acquire, Scope);
1432     Builder.CreateBr(ContBB);
1433     SI->addCase(Builder.getInt32((int)llvm::AtomicOrderingCABI::consume),
1434                 AcquireBB);
1435     SI->addCase(Builder.getInt32((int)llvm::AtomicOrderingCABI::acquire),
1436                 AcquireBB);
1437   }
1438   if (!IsLoad) {
1439     Builder.SetInsertPoint(ReleaseBB);
1440     EmitAtomicOp(*this, E, Dest, Ptr, Val1, Val2, IsWeak, OrderFail, Size,
1441                  llvm::AtomicOrdering::Release, Scope);
1442     Builder.CreateBr(ContBB);
1443     SI->addCase(Builder.getInt32((int)llvm::AtomicOrderingCABI::release),
1444                 ReleaseBB);
1445   }
1446   if (!IsLoad && !IsStore) {
1447     Builder.SetInsertPoint(AcqRelBB);
1448     EmitAtomicOp(*this, E, Dest, Ptr, Val1, Val2, IsWeak, OrderFail, Size,
1449                  llvm::AtomicOrdering::AcquireRelease, Scope);
1450     Builder.CreateBr(ContBB);
1451     SI->addCase(Builder.getInt32((int)llvm::AtomicOrderingCABI::acq_rel),
1452                 AcqRelBB);
1453   }
1454   Builder.SetInsertPoint(SeqCstBB);
1455   EmitAtomicOp(*this, E, Dest, Ptr, Val1, Val2, IsWeak, OrderFail, Size,
1456                llvm::AtomicOrdering::SequentiallyConsistent, Scope);
1457   Builder.CreateBr(ContBB);
1458   SI->addCase(Builder.getInt32((int)llvm::AtomicOrderingCABI::seq_cst),
1459               SeqCstBB);
1460 
1461   // Cleanup and return
1462   Builder.SetInsertPoint(ContBB);
1463   if (RValTy->isVoidType())
1464     return RValue::get(nullptr);
1465 
1466   assert(Atomics.getValueSizeInBits() <= Atomics.getAtomicSizeInBits());
1467   return convertTempToRValue(Dest.withElementType(ConvertTypeForMem(RValTy)),
1468                              RValTy, E->getExprLoc());
1469 }
1470 
1471 Address AtomicInfo::castToAtomicIntPointer(Address addr) const {
1472   llvm::IntegerType *ty =
1473     llvm::IntegerType::get(CGF.getLLVMContext(), AtomicSizeInBits);
1474   return addr.withElementType(ty);
1475 }
1476 
1477 Address AtomicInfo::convertToAtomicIntPointer(Address Addr) const {
1478   llvm::Type *Ty = Addr.getElementType();
1479   uint64_t SourceSizeInBits = CGF.CGM.getDataLayout().getTypeSizeInBits(Ty);
1480   if (SourceSizeInBits != AtomicSizeInBits) {
1481     Address Tmp = CreateTempAlloca();
1482     CGF.Builder.CreateMemCpy(Tmp, Addr,
1483                              std::min(AtomicSizeInBits, SourceSizeInBits) / 8);
1484     Addr = Tmp;
1485   }
1486 
1487   return castToAtomicIntPointer(Addr);
1488 }
1489 
1490 RValue AtomicInfo::convertAtomicTempToRValue(Address addr,
1491                                              AggValueSlot resultSlot,
1492                                              SourceLocation loc,
1493                                              bool asValue) const {
1494   if (LVal.isSimple()) {
1495     if (EvaluationKind == TEK_Aggregate)
1496       return resultSlot.asRValue();
1497 
1498     // Drill into the padding structure if we have one.
1499     if (hasPadding())
1500       addr = CGF.Builder.CreateStructGEP(addr, 0);
1501 
1502     // Otherwise, just convert the temporary to an r-value using the
1503     // normal conversion routine.
1504     return CGF.convertTempToRValue(addr, getValueType(), loc);
1505   }
1506   if (!asValue)
1507     // Get RValue from temp memory as atomic for non-simple lvalues
1508     return RValue::get(CGF.Builder.CreateLoad(addr));
1509   if (LVal.isBitField())
1510     return CGF.EmitLoadOfBitfieldLValue(
1511         LValue::MakeBitfield(addr, LVal.getBitFieldInfo(), LVal.getType(),
1512                              LVal.getBaseInfo(), TBAAAccessInfo()), loc);
1513   if (LVal.isVectorElt())
1514     return CGF.EmitLoadOfLValue(
1515         LValue::MakeVectorElt(addr, LVal.getVectorIdx(), LVal.getType(),
1516                               LVal.getBaseInfo(), TBAAAccessInfo()), loc);
1517   assert(LVal.isExtVectorElt());
1518   return CGF.EmitLoadOfExtVectorElementLValue(LValue::MakeExtVectorElt(
1519       addr, LVal.getExtVectorElts(), LVal.getType(),
1520       LVal.getBaseInfo(), TBAAAccessInfo()));
1521 }
1522 
1523 RValue AtomicInfo::ConvertIntToValueOrAtomic(llvm::Value *IntVal,
1524                                              AggValueSlot ResultSlot,
1525                                              SourceLocation Loc,
1526                                              bool AsValue) const {
1527   // Try not to in some easy cases.
1528   assert(IntVal->getType()->isIntegerTy() && "Expected integer value");
1529   if (getEvaluationKind() == TEK_Scalar &&
1530       (((!LVal.isBitField() ||
1531          LVal.getBitFieldInfo().Size == ValueSizeInBits) &&
1532         !hasPadding()) ||
1533        !AsValue)) {
1534     auto *ValTy = AsValue
1535                       ? CGF.ConvertTypeForMem(ValueTy)
1536                       : getAtomicAddress().getElementType();
1537     if (ValTy->isIntegerTy()) {
1538       assert(IntVal->getType() == ValTy && "Different integer types.");
1539       return RValue::get(CGF.EmitFromMemory(IntVal, ValueTy));
1540     } else if (ValTy->isPointerTy())
1541       return RValue::get(CGF.Builder.CreateIntToPtr(IntVal, ValTy));
1542     else if (llvm::CastInst::isBitCastable(IntVal->getType(), ValTy))
1543       return RValue::get(CGF.Builder.CreateBitCast(IntVal, ValTy));
1544   }
1545 
1546   // Create a temporary.  This needs to be big enough to hold the
1547   // atomic integer.
1548   Address Temp = Address::invalid();
1549   bool TempIsVolatile = false;
1550   if (AsValue && getEvaluationKind() == TEK_Aggregate) {
1551     assert(!ResultSlot.isIgnored());
1552     Temp = ResultSlot.getAddress();
1553     TempIsVolatile = ResultSlot.isVolatile();
1554   } else {
1555     Temp = CreateTempAlloca();
1556   }
1557 
1558   // Slam the integer into the temporary.
1559   Address CastTemp = castToAtomicIntPointer(Temp);
1560   CGF.Builder.CreateStore(IntVal, CastTemp)
1561       ->setVolatile(TempIsVolatile);
1562 
1563   return convertAtomicTempToRValue(Temp, ResultSlot, Loc, AsValue);
1564 }
1565 
1566 void AtomicInfo::EmitAtomicLoadLibcall(llvm::Value *AddForLoaded,
1567                                        llvm::AtomicOrdering AO, bool) {
1568   // void __atomic_load(size_t size, void *mem, void *return, int order);
1569   CallArgList Args;
1570   Args.add(RValue::get(getAtomicSizeValue()), CGF.getContext().getSizeType());
1571   Args.add(RValue::get(getAtomicPointer()), CGF.getContext().VoidPtrTy);
1572   Args.add(RValue::get(AddForLoaded), CGF.getContext().VoidPtrTy);
1573   Args.add(
1574       RValue::get(llvm::ConstantInt::get(CGF.IntTy, (int)llvm::toCABI(AO))),
1575       CGF.getContext().IntTy);
1576   emitAtomicLibcall(CGF, "__atomic_load", CGF.getContext().VoidTy, Args);
1577 }
1578 
1579 llvm::Value *AtomicInfo::EmitAtomicLoadOp(llvm::AtomicOrdering AO,
1580                                           bool IsVolatile) {
1581   // Okay, we're doing this natively.
1582   Address Addr = getAtomicAddressAsAtomicIntPointer();
1583   llvm::LoadInst *Load = CGF.Builder.CreateLoad(Addr, "atomic-load");
1584   Load->setAtomic(AO);
1585 
1586   // Other decoration.
1587   if (IsVolatile)
1588     Load->setVolatile(true);
1589   CGF.CGM.DecorateInstructionWithTBAA(Load, LVal.getTBAAInfo());
1590   return Load;
1591 }
1592 
1593 /// An LValue is a candidate for having its loads and stores be made atomic if
1594 /// we are operating under /volatile:ms *and* the LValue itself is volatile and
1595 /// performing such an operation can be performed without a libcall.
1596 bool CodeGenFunction::LValueIsSuitableForInlineAtomic(LValue LV) {
1597   if (!CGM.getLangOpts().MSVolatile) return false;
1598   AtomicInfo AI(*this, LV);
1599   bool IsVolatile = LV.isVolatile() || hasVolatileMember(LV.getType());
1600   // An atomic is inline if we don't need to use a libcall.
1601   bool AtomicIsInline = !AI.shouldUseLibcall();
1602   // MSVC doesn't seem to do this for types wider than a pointer.
1603   if (getContext().getTypeSize(LV.getType()) >
1604       getContext().getTypeSize(getContext().getIntPtrType()))
1605     return false;
1606   return IsVolatile && AtomicIsInline;
1607 }
1608 
1609 RValue CodeGenFunction::EmitAtomicLoad(LValue LV, SourceLocation SL,
1610                                        AggValueSlot Slot) {
1611   llvm::AtomicOrdering AO;
1612   bool IsVolatile = LV.isVolatileQualified();
1613   if (LV.getType()->isAtomicType()) {
1614     AO = llvm::AtomicOrdering::SequentiallyConsistent;
1615   } else {
1616     AO = llvm::AtomicOrdering::Acquire;
1617     IsVolatile = true;
1618   }
1619   return EmitAtomicLoad(LV, SL, AO, IsVolatile, Slot);
1620 }
1621 
1622 RValue AtomicInfo::EmitAtomicLoad(AggValueSlot ResultSlot, SourceLocation Loc,
1623                                   bool AsValue, llvm::AtomicOrdering AO,
1624                                   bool IsVolatile) {
1625   // Check whether we should use a library call.
1626   if (shouldUseLibcall()) {
1627     Address TempAddr = Address::invalid();
1628     if (LVal.isSimple() && !ResultSlot.isIgnored()) {
1629       assert(getEvaluationKind() == TEK_Aggregate);
1630       TempAddr = ResultSlot.getAddress();
1631     } else
1632       TempAddr = CreateTempAlloca();
1633 
1634     EmitAtomicLoadLibcall(TempAddr.getPointer(), AO, IsVolatile);
1635 
1636     // Okay, turn that back into the original value or whole atomic (for
1637     // non-simple lvalues) type.
1638     return convertAtomicTempToRValue(TempAddr, ResultSlot, Loc, AsValue);
1639   }
1640 
1641   // Okay, we're doing this natively.
1642   auto *Load = EmitAtomicLoadOp(AO, IsVolatile);
1643 
1644   // If we're ignoring an aggregate return, don't do anything.
1645   if (getEvaluationKind() == TEK_Aggregate && ResultSlot.isIgnored())
1646     return RValue::getAggregate(Address::invalid(), false);
1647 
1648   // Okay, turn that back into the original value or atomic (for non-simple
1649   // lvalues) type.
1650   return ConvertIntToValueOrAtomic(Load, ResultSlot, Loc, AsValue);
1651 }
1652 
1653 /// Emit a load from an l-value of atomic type.  Note that the r-value
1654 /// we produce is an r-value of the atomic *value* type.
1655 RValue CodeGenFunction::EmitAtomicLoad(LValue src, SourceLocation loc,
1656                                        llvm::AtomicOrdering AO, bool IsVolatile,
1657                                        AggValueSlot resultSlot) {
1658   AtomicInfo Atomics(*this, src);
1659   return Atomics.EmitAtomicLoad(resultSlot, loc, /*AsValue=*/true, AO,
1660                                 IsVolatile);
1661 }
1662 
1663 /// Copy an r-value into memory as part of storing to an atomic type.
1664 /// This needs to create a bit-pattern suitable for atomic operations.
1665 void AtomicInfo::emitCopyIntoMemory(RValue rvalue) const {
1666   assert(LVal.isSimple());
1667   // If we have an r-value, the rvalue should be of the atomic type,
1668   // which means that the caller is responsible for having zeroed
1669   // any padding.  Just do an aggregate copy of that type.
1670   if (rvalue.isAggregate()) {
1671     LValue Dest = CGF.MakeAddrLValue(getAtomicAddress(), getAtomicType());
1672     LValue Src = CGF.MakeAddrLValue(rvalue.getAggregateAddress(),
1673                                     getAtomicType());
1674     bool IsVolatile = rvalue.isVolatileQualified() ||
1675                       LVal.isVolatileQualified();
1676     CGF.EmitAggregateCopy(Dest, Src, getAtomicType(),
1677                           AggValueSlot::DoesNotOverlap, IsVolatile);
1678     return;
1679   }
1680 
1681   // Okay, otherwise we're copying stuff.
1682 
1683   // Zero out the buffer if necessary.
1684   emitMemSetZeroIfNecessary();
1685 
1686   // Drill past the padding if present.
1687   LValue TempLVal = projectValue();
1688 
1689   // Okay, store the rvalue in.
1690   if (rvalue.isScalar()) {
1691     CGF.EmitStoreOfScalar(rvalue.getScalarVal(), TempLVal, /*init*/ true);
1692   } else {
1693     CGF.EmitStoreOfComplex(rvalue.getComplexVal(), TempLVal, /*init*/ true);
1694   }
1695 }
1696 
1697 
1698 /// Materialize an r-value into memory for the purposes of storing it
1699 /// to an atomic type.
1700 Address AtomicInfo::materializeRValue(RValue rvalue) const {
1701   // Aggregate r-values are already in memory, and EmitAtomicStore
1702   // requires them to be values of the atomic type.
1703   if (rvalue.isAggregate())
1704     return rvalue.getAggregateAddress();
1705 
1706   // Otherwise, make a temporary and materialize into it.
1707   LValue TempLV = CGF.MakeAddrLValue(CreateTempAlloca(), getAtomicType());
1708   AtomicInfo Atomics(CGF, TempLV);
1709   Atomics.emitCopyIntoMemory(rvalue);
1710   return TempLV.getAddress(CGF);
1711 }
1712 
1713 llvm::Value *AtomicInfo::convertRValueToInt(RValue RVal) const {
1714   // If we've got a scalar value of the right size, try to avoid going
1715   // through memory.
1716   if (RVal.isScalar() && (!hasPadding() || !LVal.isSimple())) {
1717     llvm::Value *Value = RVal.getScalarVal();
1718     if (isa<llvm::IntegerType>(Value->getType()))
1719       return CGF.EmitToMemory(Value, ValueTy);
1720     else {
1721       llvm::IntegerType *InputIntTy = llvm::IntegerType::get(
1722           CGF.getLLVMContext(),
1723           LVal.isSimple() ? getValueSizeInBits() : getAtomicSizeInBits());
1724       if (isa<llvm::PointerType>(Value->getType()))
1725         return CGF.Builder.CreatePtrToInt(Value, InputIntTy);
1726       else if (llvm::BitCastInst::isBitCastable(Value->getType(), InputIntTy))
1727         return CGF.Builder.CreateBitCast(Value, InputIntTy);
1728     }
1729   }
1730   // Otherwise, we need to go through memory.
1731   // Put the r-value in memory.
1732   Address Addr = materializeRValue(RVal);
1733 
1734   // Cast the temporary to the atomic int type and pull a value out.
1735   Addr = castToAtomicIntPointer(Addr);
1736   return CGF.Builder.CreateLoad(Addr);
1737 }
1738 
1739 std::pair<llvm::Value *, llvm::Value *> AtomicInfo::EmitAtomicCompareExchangeOp(
1740     llvm::Value *ExpectedVal, llvm::Value *DesiredVal,
1741     llvm::AtomicOrdering Success, llvm::AtomicOrdering Failure, bool IsWeak) {
1742   // Do the atomic store.
1743   Address Addr = getAtomicAddressAsAtomicIntPointer();
1744   auto *Inst = CGF.Builder.CreateAtomicCmpXchg(Addr.getPointer(),
1745                                                ExpectedVal, DesiredVal,
1746                                                Success, Failure);
1747   // Other decoration.
1748   Inst->setVolatile(LVal.isVolatileQualified());
1749   Inst->setWeak(IsWeak);
1750 
1751   // Okay, turn that back into the original value type.
1752   auto *PreviousVal = CGF.Builder.CreateExtractValue(Inst, /*Idxs=*/0);
1753   auto *SuccessFailureVal = CGF.Builder.CreateExtractValue(Inst, /*Idxs=*/1);
1754   return std::make_pair(PreviousVal, SuccessFailureVal);
1755 }
1756 
1757 llvm::Value *
1758 AtomicInfo::EmitAtomicCompareExchangeLibcall(llvm::Value *ExpectedAddr,
1759                                              llvm::Value *DesiredAddr,
1760                                              llvm::AtomicOrdering Success,
1761                                              llvm::AtomicOrdering Failure) {
1762   // bool __atomic_compare_exchange(size_t size, void *obj, void *expected,
1763   // void *desired, int success, int failure);
1764   CallArgList Args;
1765   Args.add(RValue::get(getAtomicSizeValue()), CGF.getContext().getSizeType());
1766   Args.add(RValue::get(getAtomicPointer()), CGF.getContext().VoidPtrTy);
1767   Args.add(RValue::get(ExpectedAddr), CGF.getContext().VoidPtrTy);
1768   Args.add(RValue::get(DesiredAddr), CGF.getContext().VoidPtrTy);
1769   Args.add(RValue::get(
1770                llvm::ConstantInt::get(CGF.IntTy, (int)llvm::toCABI(Success))),
1771            CGF.getContext().IntTy);
1772   Args.add(RValue::get(
1773                llvm::ConstantInt::get(CGF.IntTy, (int)llvm::toCABI(Failure))),
1774            CGF.getContext().IntTy);
1775   auto SuccessFailureRVal = emitAtomicLibcall(CGF, "__atomic_compare_exchange",
1776                                               CGF.getContext().BoolTy, Args);
1777 
1778   return SuccessFailureRVal.getScalarVal();
1779 }
1780 
1781 std::pair<RValue, llvm::Value *> AtomicInfo::EmitAtomicCompareExchange(
1782     RValue Expected, RValue Desired, llvm::AtomicOrdering Success,
1783     llvm::AtomicOrdering Failure, bool IsWeak) {
1784   // Check whether we should use a library call.
1785   if (shouldUseLibcall()) {
1786     // Produce a source address.
1787     Address ExpectedAddr = materializeRValue(Expected);
1788     Address DesiredAddr = materializeRValue(Desired);
1789     auto *Res = EmitAtomicCompareExchangeLibcall(ExpectedAddr.getPointer(),
1790                                                  DesiredAddr.getPointer(),
1791                                                  Success, Failure);
1792     return std::make_pair(
1793         convertAtomicTempToRValue(ExpectedAddr, AggValueSlot::ignored(),
1794                                   SourceLocation(), /*AsValue=*/false),
1795         Res);
1796   }
1797 
1798   // If we've got a scalar value of the right size, try to avoid going
1799   // through memory.
1800   auto *ExpectedVal = convertRValueToInt(Expected);
1801   auto *DesiredVal = convertRValueToInt(Desired);
1802   auto Res = EmitAtomicCompareExchangeOp(ExpectedVal, DesiredVal, Success,
1803                                          Failure, IsWeak);
1804   return std::make_pair(
1805       ConvertIntToValueOrAtomic(Res.first, AggValueSlot::ignored(),
1806                                 SourceLocation(), /*AsValue=*/false),
1807       Res.second);
1808 }
1809 
1810 static void
1811 EmitAtomicUpdateValue(CodeGenFunction &CGF, AtomicInfo &Atomics, RValue OldRVal,
1812                       const llvm::function_ref<RValue(RValue)> &UpdateOp,
1813                       Address DesiredAddr) {
1814   RValue UpRVal;
1815   LValue AtomicLVal = Atomics.getAtomicLValue();
1816   LValue DesiredLVal;
1817   if (AtomicLVal.isSimple()) {
1818     UpRVal = OldRVal;
1819     DesiredLVal = CGF.MakeAddrLValue(DesiredAddr, AtomicLVal.getType());
1820   } else {
1821     // Build new lvalue for temp address.
1822     Address Ptr = Atomics.materializeRValue(OldRVal);
1823     LValue UpdateLVal;
1824     if (AtomicLVal.isBitField()) {
1825       UpdateLVal =
1826           LValue::MakeBitfield(Ptr, AtomicLVal.getBitFieldInfo(),
1827                                AtomicLVal.getType(),
1828                                AtomicLVal.getBaseInfo(),
1829                                AtomicLVal.getTBAAInfo());
1830       DesiredLVal =
1831           LValue::MakeBitfield(DesiredAddr, AtomicLVal.getBitFieldInfo(),
1832                                AtomicLVal.getType(), AtomicLVal.getBaseInfo(),
1833                                AtomicLVal.getTBAAInfo());
1834     } else if (AtomicLVal.isVectorElt()) {
1835       UpdateLVal = LValue::MakeVectorElt(Ptr, AtomicLVal.getVectorIdx(),
1836                                          AtomicLVal.getType(),
1837                                          AtomicLVal.getBaseInfo(),
1838                                          AtomicLVal.getTBAAInfo());
1839       DesiredLVal = LValue::MakeVectorElt(
1840           DesiredAddr, AtomicLVal.getVectorIdx(), AtomicLVal.getType(),
1841           AtomicLVal.getBaseInfo(), AtomicLVal.getTBAAInfo());
1842     } else {
1843       assert(AtomicLVal.isExtVectorElt());
1844       UpdateLVal = LValue::MakeExtVectorElt(Ptr, AtomicLVal.getExtVectorElts(),
1845                                             AtomicLVal.getType(),
1846                                             AtomicLVal.getBaseInfo(),
1847                                             AtomicLVal.getTBAAInfo());
1848       DesiredLVal = LValue::MakeExtVectorElt(
1849           DesiredAddr, AtomicLVal.getExtVectorElts(), AtomicLVal.getType(),
1850           AtomicLVal.getBaseInfo(), AtomicLVal.getTBAAInfo());
1851     }
1852     UpRVal = CGF.EmitLoadOfLValue(UpdateLVal, SourceLocation());
1853   }
1854   // Store new value in the corresponding memory area.
1855   RValue NewRVal = UpdateOp(UpRVal);
1856   if (NewRVal.isScalar()) {
1857     CGF.EmitStoreThroughLValue(NewRVal, DesiredLVal);
1858   } else {
1859     assert(NewRVal.isComplex());
1860     CGF.EmitStoreOfComplex(NewRVal.getComplexVal(), DesiredLVal,
1861                            /*isInit=*/false);
1862   }
1863 }
1864 
1865 void AtomicInfo::EmitAtomicUpdateLibcall(
1866     llvm::AtomicOrdering AO, const llvm::function_ref<RValue(RValue)> &UpdateOp,
1867     bool IsVolatile) {
1868   auto Failure = llvm::AtomicCmpXchgInst::getStrongestFailureOrdering(AO);
1869 
1870   Address ExpectedAddr = CreateTempAlloca();
1871 
1872   EmitAtomicLoadLibcall(ExpectedAddr.getPointer(), AO, IsVolatile);
1873   auto *ContBB = CGF.createBasicBlock("atomic_cont");
1874   auto *ExitBB = CGF.createBasicBlock("atomic_exit");
1875   CGF.EmitBlock(ContBB);
1876   Address DesiredAddr = CreateTempAlloca();
1877   if ((LVal.isBitField() && BFI.Size != ValueSizeInBits) ||
1878       requiresMemSetZero(getAtomicAddress().getElementType())) {
1879     auto *OldVal = CGF.Builder.CreateLoad(ExpectedAddr);
1880     CGF.Builder.CreateStore(OldVal, DesiredAddr);
1881   }
1882   auto OldRVal = convertAtomicTempToRValue(ExpectedAddr,
1883                                            AggValueSlot::ignored(),
1884                                            SourceLocation(), /*AsValue=*/false);
1885   EmitAtomicUpdateValue(CGF, *this, OldRVal, UpdateOp, DesiredAddr);
1886   auto *Res =
1887       EmitAtomicCompareExchangeLibcall(ExpectedAddr.getPointer(),
1888                                        DesiredAddr.getPointer(),
1889                                        AO, Failure);
1890   CGF.Builder.CreateCondBr(Res, ExitBB, ContBB);
1891   CGF.EmitBlock(ExitBB, /*IsFinished=*/true);
1892 }
1893 
1894 void AtomicInfo::EmitAtomicUpdateOp(
1895     llvm::AtomicOrdering AO, const llvm::function_ref<RValue(RValue)> &UpdateOp,
1896     bool IsVolatile) {
1897   auto Failure = llvm::AtomicCmpXchgInst::getStrongestFailureOrdering(AO);
1898 
1899   // Do the atomic load.
1900   auto *OldVal = EmitAtomicLoadOp(Failure, IsVolatile);
1901   // For non-simple lvalues perform compare-and-swap procedure.
1902   auto *ContBB = CGF.createBasicBlock("atomic_cont");
1903   auto *ExitBB = CGF.createBasicBlock("atomic_exit");
1904   auto *CurBB = CGF.Builder.GetInsertBlock();
1905   CGF.EmitBlock(ContBB);
1906   llvm::PHINode *PHI = CGF.Builder.CreatePHI(OldVal->getType(),
1907                                              /*NumReservedValues=*/2);
1908   PHI->addIncoming(OldVal, CurBB);
1909   Address NewAtomicAddr = CreateTempAlloca();
1910   Address NewAtomicIntAddr = castToAtomicIntPointer(NewAtomicAddr);
1911   if ((LVal.isBitField() && BFI.Size != ValueSizeInBits) ||
1912       requiresMemSetZero(getAtomicAddress().getElementType())) {
1913     CGF.Builder.CreateStore(PHI, NewAtomicIntAddr);
1914   }
1915   auto OldRVal = ConvertIntToValueOrAtomic(PHI, AggValueSlot::ignored(),
1916                                            SourceLocation(), /*AsValue=*/false);
1917   EmitAtomicUpdateValue(CGF, *this, OldRVal, UpdateOp, NewAtomicAddr);
1918   auto *DesiredVal = CGF.Builder.CreateLoad(NewAtomicIntAddr);
1919   // Try to write new value using cmpxchg operation.
1920   auto Res = EmitAtomicCompareExchangeOp(PHI, DesiredVal, AO, Failure);
1921   PHI->addIncoming(Res.first, CGF.Builder.GetInsertBlock());
1922   CGF.Builder.CreateCondBr(Res.second, ExitBB, ContBB);
1923   CGF.EmitBlock(ExitBB, /*IsFinished=*/true);
1924 }
1925 
1926 static void EmitAtomicUpdateValue(CodeGenFunction &CGF, AtomicInfo &Atomics,
1927                                   RValue UpdateRVal, Address DesiredAddr) {
1928   LValue AtomicLVal = Atomics.getAtomicLValue();
1929   LValue DesiredLVal;
1930   // Build new lvalue for temp address.
1931   if (AtomicLVal.isBitField()) {
1932     DesiredLVal =
1933         LValue::MakeBitfield(DesiredAddr, AtomicLVal.getBitFieldInfo(),
1934                              AtomicLVal.getType(), AtomicLVal.getBaseInfo(),
1935                              AtomicLVal.getTBAAInfo());
1936   } else if (AtomicLVal.isVectorElt()) {
1937     DesiredLVal =
1938         LValue::MakeVectorElt(DesiredAddr, AtomicLVal.getVectorIdx(),
1939                               AtomicLVal.getType(), AtomicLVal.getBaseInfo(),
1940                               AtomicLVal.getTBAAInfo());
1941   } else {
1942     assert(AtomicLVal.isExtVectorElt());
1943     DesiredLVal = LValue::MakeExtVectorElt(
1944         DesiredAddr, AtomicLVal.getExtVectorElts(), AtomicLVal.getType(),
1945         AtomicLVal.getBaseInfo(), AtomicLVal.getTBAAInfo());
1946   }
1947   // Store new value in the corresponding memory area.
1948   assert(UpdateRVal.isScalar());
1949   CGF.EmitStoreThroughLValue(UpdateRVal, DesiredLVal);
1950 }
1951 
1952 void AtomicInfo::EmitAtomicUpdateLibcall(llvm::AtomicOrdering AO,
1953                                          RValue UpdateRVal, bool IsVolatile) {
1954   auto Failure = llvm::AtomicCmpXchgInst::getStrongestFailureOrdering(AO);
1955 
1956   Address ExpectedAddr = CreateTempAlloca();
1957 
1958   EmitAtomicLoadLibcall(ExpectedAddr.getPointer(), AO, IsVolatile);
1959   auto *ContBB = CGF.createBasicBlock("atomic_cont");
1960   auto *ExitBB = CGF.createBasicBlock("atomic_exit");
1961   CGF.EmitBlock(ContBB);
1962   Address DesiredAddr = CreateTempAlloca();
1963   if ((LVal.isBitField() && BFI.Size != ValueSizeInBits) ||
1964       requiresMemSetZero(getAtomicAddress().getElementType())) {
1965     auto *OldVal = CGF.Builder.CreateLoad(ExpectedAddr);
1966     CGF.Builder.CreateStore(OldVal, DesiredAddr);
1967   }
1968   EmitAtomicUpdateValue(CGF, *this, UpdateRVal, DesiredAddr);
1969   auto *Res =
1970       EmitAtomicCompareExchangeLibcall(ExpectedAddr.getPointer(),
1971                                        DesiredAddr.getPointer(),
1972                                        AO, Failure);
1973   CGF.Builder.CreateCondBr(Res, ExitBB, ContBB);
1974   CGF.EmitBlock(ExitBB, /*IsFinished=*/true);
1975 }
1976 
1977 void AtomicInfo::EmitAtomicUpdateOp(llvm::AtomicOrdering AO, RValue UpdateRVal,
1978                                     bool IsVolatile) {
1979   auto Failure = llvm::AtomicCmpXchgInst::getStrongestFailureOrdering(AO);
1980 
1981   // Do the atomic load.
1982   auto *OldVal = EmitAtomicLoadOp(Failure, IsVolatile);
1983   // For non-simple lvalues perform compare-and-swap procedure.
1984   auto *ContBB = CGF.createBasicBlock("atomic_cont");
1985   auto *ExitBB = CGF.createBasicBlock("atomic_exit");
1986   auto *CurBB = CGF.Builder.GetInsertBlock();
1987   CGF.EmitBlock(ContBB);
1988   llvm::PHINode *PHI = CGF.Builder.CreatePHI(OldVal->getType(),
1989                                              /*NumReservedValues=*/2);
1990   PHI->addIncoming(OldVal, CurBB);
1991   Address NewAtomicAddr = CreateTempAlloca();
1992   Address NewAtomicIntAddr = castToAtomicIntPointer(NewAtomicAddr);
1993   if ((LVal.isBitField() && BFI.Size != ValueSizeInBits) ||
1994       requiresMemSetZero(getAtomicAddress().getElementType())) {
1995     CGF.Builder.CreateStore(PHI, NewAtomicIntAddr);
1996   }
1997   EmitAtomicUpdateValue(CGF, *this, UpdateRVal, NewAtomicAddr);
1998   auto *DesiredVal = CGF.Builder.CreateLoad(NewAtomicIntAddr);
1999   // Try to write new value using cmpxchg operation.
2000   auto Res = EmitAtomicCompareExchangeOp(PHI, DesiredVal, AO, Failure);
2001   PHI->addIncoming(Res.first, CGF.Builder.GetInsertBlock());
2002   CGF.Builder.CreateCondBr(Res.second, ExitBB, ContBB);
2003   CGF.EmitBlock(ExitBB, /*IsFinished=*/true);
2004 }
2005 
2006 void AtomicInfo::EmitAtomicUpdate(
2007     llvm::AtomicOrdering AO, const llvm::function_ref<RValue(RValue)> &UpdateOp,
2008     bool IsVolatile) {
2009   if (shouldUseLibcall()) {
2010     EmitAtomicUpdateLibcall(AO, UpdateOp, IsVolatile);
2011   } else {
2012     EmitAtomicUpdateOp(AO, UpdateOp, IsVolatile);
2013   }
2014 }
2015 
2016 void AtomicInfo::EmitAtomicUpdate(llvm::AtomicOrdering AO, RValue UpdateRVal,
2017                                   bool IsVolatile) {
2018   if (shouldUseLibcall()) {
2019     EmitAtomicUpdateLibcall(AO, UpdateRVal, IsVolatile);
2020   } else {
2021     EmitAtomicUpdateOp(AO, UpdateRVal, IsVolatile);
2022   }
2023 }
2024 
2025 void CodeGenFunction::EmitAtomicStore(RValue rvalue, LValue lvalue,
2026                                       bool isInit) {
2027   bool IsVolatile = lvalue.isVolatileQualified();
2028   llvm::AtomicOrdering AO;
2029   if (lvalue.getType()->isAtomicType()) {
2030     AO = llvm::AtomicOrdering::SequentiallyConsistent;
2031   } else {
2032     AO = llvm::AtomicOrdering::Release;
2033     IsVolatile = true;
2034   }
2035   return EmitAtomicStore(rvalue, lvalue, AO, IsVolatile, isInit);
2036 }
2037 
2038 /// Emit a store to an l-value of atomic type.
2039 ///
2040 /// Note that the r-value is expected to be an r-value *of the atomic
2041 /// type*; this means that for aggregate r-values, it should include
2042 /// storage for any padding that was necessary.
2043 void CodeGenFunction::EmitAtomicStore(RValue rvalue, LValue dest,
2044                                       llvm::AtomicOrdering AO, bool IsVolatile,
2045                                       bool isInit) {
2046   // If this is an aggregate r-value, it should agree in type except
2047   // maybe for address-space qualification.
2048   assert(!rvalue.isAggregate() ||
2049          rvalue.getAggregateAddress().getElementType() ==
2050              dest.getAddress(*this).getElementType());
2051 
2052   AtomicInfo atomics(*this, dest);
2053   LValue LVal = atomics.getAtomicLValue();
2054 
2055   // If this is an initialization, just put the value there normally.
2056   if (LVal.isSimple()) {
2057     if (isInit) {
2058       atomics.emitCopyIntoMemory(rvalue);
2059       return;
2060     }
2061 
2062     // Check whether we should use a library call.
2063     if (atomics.shouldUseLibcall()) {
2064       // Produce a source address.
2065       Address srcAddr = atomics.materializeRValue(rvalue);
2066 
2067       // void __atomic_store(size_t size, void *mem, void *val, int order)
2068       CallArgList args;
2069       args.add(RValue::get(atomics.getAtomicSizeValue()),
2070                getContext().getSizeType());
2071       args.add(RValue::get(atomics.getAtomicPointer()), getContext().VoidPtrTy);
2072       args.add(RValue::get(srcAddr.getPointer()), getContext().VoidPtrTy);
2073       args.add(
2074           RValue::get(llvm::ConstantInt::get(IntTy, (int)llvm::toCABI(AO))),
2075           getContext().IntTy);
2076       emitAtomicLibcall(*this, "__atomic_store", getContext().VoidTy, args);
2077       return;
2078     }
2079 
2080     // Okay, we're doing this natively.
2081     llvm::Value *intValue = atomics.convertRValueToInt(rvalue);
2082 
2083     // Do the atomic store.
2084     Address addr = atomics.castToAtomicIntPointer(atomics.getAtomicAddress());
2085     intValue = Builder.CreateIntCast(
2086         intValue, addr.getElementType(), /*isSigned=*/false);
2087     llvm::StoreInst *store = Builder.CreateStore(intValue, addr);
2088 
2089     if (AO == llvm::AtomicOrdering::Acquire)
2090       AO = llvm::AtomicOrdering::Monotonic;
2091     else if (AO == llvm::AtomicOrdering::AcquireRelease)
2092       AO = llvm::AtomicOrdering::Release;
2093     // Initializations don't need to be atomic.
2094     if (!isInit)
2095       store->setAtomic(AO);
2096 
2097     // Other decoration.
2098     if (IsVolatile)
2099       store->setVolatile(true);
2100     CGM.DecorateInstructionWithTBAA(store, dest.getTBAAInfo());
2101     return;
2102   }
2103 
2104   // Emit simple atomic update operation.
2105   atomics.EmitAtomicUpdate(AO, rvalue, IsVolatile);
2106 }
2107 
2108 /// Emit a compare-and-exchange op for atomic type.
2109 ///
2110 std::pair<RValue, llvm::Value *> CodeGenFunction::EmitAtomicCompareExchange(
2111     LValue Obj, RValue Expected, RValue Desired, SourceLocation Loc,
2112     llvm::AtomicOrdering Success, llvm::AtomicOrdering Failure, bool IsWeak,
2113     AggValueSlot Slot) {
2114   // If this is an aggregate r-value, it should agree in type except
2115   // maybe for address-space qualification.
2116   assert(!Expected.isAggregate() ||
2117          Expected.getAggregateAddress().getElementType() ==
2118              Obj.getAddress(*this).getElementType());
2119   assert(!Desired.isAggregate() ||
2120          Desired.getAggregateAddress().getElementType() ==
2121              Obj.getAddress(*this).getElementType());
2122   AtomicInfo Atomics(*this, Obj);
2123 
2124   return Atomics.EmitAtomicCompareExchange(Expected, Desired, Success, Failure,
2125                                            IsWeak);
2126 }
2127 
2128 void CodeGenFunction::EmitAtomicUpdate(
2129     LValue LVal, llvm::AtomicOrdering AO,
2130     const llvm::function_ref<RValue(RValue)> &UpdateOp, bool IsVolatile) {
2131   AtomicInfo Atomics(*this, LVal);
2132   Atomics.EmitAtomicUpdate(AO, UpdateOp, IsVolatile);
2133 }
2134 
2135 void CodeGenFunction::EmitAtomicInit(Expr *init, LValue dest) {
2136   AtomicInfo atomics(*this, dest);
2137 
2138   switch (atomics.getEvaluationKind()) {
2139   case TEK_Scalar: {
2140     llvm::Value *value = EmitScalarExpr(init);
2141     atomics.emitCopyIntoMemory(RValue::get(value));
2142     return;
2143   }
2144 
2145   case TEK_Complex: {
2146     ComplexPairTy value = EmitComplexExpr(init);
2147     atomics.emitCopyIntoMemory(RValue::getComplex(value));
2148     return;
2149   }
2150 
2151   case TEK_Aggregate: {
2152     // Fix up the destination if the initializer isn't an expression
2153     // of atomic type.
2154     bool Zeroed = false;
2155     if (!init->getType()->isAtomicType()) {
2156       Zeroed = atomics.emitMemSetZeroIfNecessary();
2157       dest = atomics.projectValue();
2158     }
2159 
2160     // Evaluate the expression directly into the destination.
2161     AggValueSlot slot = AggValueSlot::forLValue(
2162         dest, *this, AggValueSlot::IsNotDestructed,
2163         AggValueSlot::DoesNotNeedGCBarriers, AggValueSlot::IsNotAliased,
2164         AggValueSlot::DoesNotOverlap,
2165         Zeroed ? AggValueSlot::IsZeroed : AggValueSlot::IsNotZeroed);
2166 
2167     EmitAggExpr(init, slot);
2168     return;
2169   }
2170   }
2171   llvm_unreachable("bad evaluation kind");
2172 }
2173