1 //===--- CGAtomic.cpp - Emit LLVM IR for atomic operations ----------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file contains the code for emitting atomic operations.
11 //
12 //===----------------------------------------------------------------------===//
13
14 #include "CodeGenFunction.h"
15 #include "CGCall.h"
16 #include "CGRecordLayout.h"
17 #include "CodeGenModule.h"
18 #include "clang/AST/ASTContext.h"
19 #include "clang/CodeGen/CGFunctionInfo.h"
20 #include "llvm/ADT/StringExtras.h"
21 #include "llvm/IR/DataLayout.h"
22 #include "llvm/IR/Intrinsics.h"
23 #include "llvm/IR/Operator.h"
24
25 using namespace clang;
26 using namespace CodeGen;
27
28 namespace {
29 class AtomicInfo {
30 CodeGenFunction &CGF;
31 QualType AtomicTy;
32 QualType ValueTy;
33 uint64_t AtomicSizeInBits;
34 uint64_t ValueSizeInBits;
35 CharUnits AtomicAlign;
36 CharUnits ValueAlign;
37 CharUnits LValueAlign;
38 TypeEvaluationKind EvaluationKind;
39 bool UseLibcall;
40 LValue LVal;
41 CGBitFieldInfo BFI;
42 public:
AtomicInfo(CodeGenFunction & CGF,LValue & lvalue)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 .RoundUpToAlignment(lvalue.getAlignment()));
83 auto VoidPtrAddr = CGF.EmitCastToVoidPtr(lvalue.getBitFieldAddr());
84 auto OffsetInChars =
85 (C.toCharUnitsFromBits(OrigBFI.Offset) / lvalue.getAlignment()) *
86 lvalue.getAlignment();
87 VoidPtrAddr = CGF.Builder.CreateConstGEP1_64(
88 VoidPtrAddr, OffsetInChars.getQuantity());
89 auto Addr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
90 VoidPtrAddr,
91 CGF.Builder.getIntNTy(AtomicSizeInBits)->getPointerTo(),
92 "atomic_bitfield_base");
93 BFI = OrigBFI;
94 BFI.Offset = Offset;
95 BFI.StorageSize = AtomicSizeInBits;
96 LVal = LValue::MakeBitfield(Addr, BFI, lvalue.getType(),
97 lvalue.getAlignment());
98 LVal.setTBAAInfo(lvalue.getTBAAInfo());
99 AtomicTy = C.getIntTypeForBitwidth(AtomicSizeInBits, OrigBFI.IsSigned);
100 if (AtomicTy.isNull()) {
101 llvm::APInt Size(
102 /*numBits=*/32,
103 C.toCharUnitsFromBits(AtomicSizeInBits).getQuantity());
104 AtomicTy = C.getConstantArrayType(C.CharTy, Size, ArrayType::Normal,
105 /*IndexTypeQuals=*/0);
106 }
107 AtomicAlign = ValueAlign = lvalue.getAlignment();
108 } else if (lvalue.isVectorElt()) {
109 ValueTy = lvalue.getType()->getAs<VectorType>()->getElementType();
110 ValueSizeInBits = C.getTypeSize(ValueTy);
111 AtomicTy = lvalue.getType();
112 AtomicSizeInBits = C.getTypeSize(AtomicTy);
113 AtomicAlign = ValueAlign = lvalue.getAlignment();
114 LVal = lvalue;
115 } else {
116 assert(lvalue.isExtVectorElt());
117 ValueTy = lvalue.getType();
118 ValueSizeInBits = C.getTypeSize(ValueTy);
119 AtomicTy = ValueTy = CGF.getContext().getExtVectorType(
120 lvalue.getType(), lvalue.getExtVectorAddr()
121 ->getType()
122 ->getPointerElementType()
123 ->getVectorNumElements());
124 AtomicSizeInBits = C.getTypeSize(AtomicTy);
125 AtomicAlign = ValueAlign = lvalue.getAlignment();
126 LVal = lvalue;
127 }
128 UseLibcall = !C.getTargetInfo().hasBuiltinAtomic(
129 AtomicSizeInBits, C.toBits(lvalue.getAlignment()));
130 }
131
getAtomicType() const132 QualType getAtomicType() const { return AtomicTy; }
getValueType() const133 QualType getValueType() const { return ValueTy; }
getAtomicAlignment() const134 CharUnits getAtomicAlignment() const { return AtomicAlign; }
getValueAlignment() const135 CharUnits getValueAlignment() const { return ValueAlign; }
getAtomicSizeInBits() const136 uint64_t getAtomicSizeInBits() const { return AtomicSizeInBits; }
getValueSizeInBits() const137 uint64_t getValueSizeInBits() const { return ValueSizeInBits; }
getEvaluationKind() const138 TypeEvaluationKind getEvaluationKind() const { return EvaluationKind; }
shouldUseLibcall() const139 bool shouldUseLibcall() const { return UseLibcall; }
getAtomicLValue() const140 const LValue &getAtomicLValue() const { return LVal; }
getAtomicAddress() const141 llvm::Value *getAtomicAddress() const {
142 if (LVal.isSimple())
143 return LVal.getAddress();
144 else if (LVal.isBitField())
145 return LVal.getBitFieldAddr();
146 else if (LVal.isVectorElt())
147 return LVal.getVectorAddr();
148 assert(LVal.isExtVectorElt());
149 return LVal.getExtVectorAddr();
150 }
151
152 /// Is the atomic size larger than the underlying value type?
153 ///
154 /// Note that the absence of padding does not mean that atomic
155 /// objects are completely interchangeable with non-atomic
156 /// objects: we might have promoted the alignment of a type
157 /// without making it bigger.
hasPadding() const158 bool hasPadding() const {
159 return (ValueSizeInBits != AtomicSizeInBits);
160 }
161
162 bool emitMemSetZeroIfNecessary() const;
163
getAtomicSizeValue() const164 llvm::Value *getAtomicSizeValue() const {
165 CharUnits size = CGF.getContext().toCharUnitsFromBits(AtomicSizeInBits);
166 return CGF.CGM.getSize(size);
167 }
168
169 /// Cast the given pointer to an integer pointer suitable for
170 /// atomic operations.
171 llvm::Value *emitCastToAtomicIntPointer(llvm::Value *addr) const;
172
173 /// Turn an atomic-layout object into an r-value.
174 RValue convertTempToRValue(llvm::Value *addr, AggValueSlot resultSlot,
175 SourceLocation loc, bool AsValue) const;
176
177 /// \brief Converts a rvalue to integer value.
178 llvm::Value *convertRValueToInt(RValue RVal) const;
179
180 RValue ConvertIntToValueOrAtomic(llvm::Value *IntVal,
181 AggValueSlot ResultSlot,
182 SourceLocation Loc, bool AsValue) const;
183
184 /// Copy an atomic r-value into atomic-layout memory.
185 void emitCopyIntoMemory(RValue rvalue) const;
186
187 /// Project an l-value down to the value field.
projectValue() const188 LValue projectValue() const {
189 assert(LVal.isSimple());
190 llvm::Value *addr = getAtomicAddress();
191 if (hasPadding())
192 addr = CGF.Builder.CreateStructGEP(nullptr, addr, 0);
193
194 return LValue::MakeAddr(addr, getValueType(), LVal.getAlignment(),
195 CGF.getContext(), LVal.getTBAAInfo());
196 }
197
198 /// \brief Emits atomic load.
199 /// \returns Loaded value.
200 RValue EmitAtomicLoad(AggValueSlot ResultSlot, SourceLocation Loc,
201 bool AsValue, llvm::AtomicOrdering AO,
202 bool IsVolatile);
203
204 /// \brief Emits atomic compare-and-exchange sequence.
205 /// \param Expected Expected value.
206 /// \param Desired Desired value.
207 /// \param Success Atomic ordering for success operation.
208 /// \param Failure Atomic ordering for failed operation.
209 /// \param IsWeak true if atomic operation is weak, false otherwise.
210 /// \returns Pair of values: previous value from storage (value type) and
211 /// boolean flag (i1 type) with true if success and false otherwise.
212 std::pair<RValue, llvm::Value *> EmitAtomicCompareExchange(
213 RValue Expected, RValue Desired,
214 llvm::AtomicOrdering Success = llvm::SequentiallyConsistent,
215 llvm::AtomicOrdering Failure = llvm::SequentiallyConsistent,
216 bool IsWeak = false);
217
218 /// Materialize an atomic r-value in atomic-layout memory.
219 llvm::Value *materializeRValue(RValue rvalue) const;
220
221 /// \brief Translates LLVM atomic ordering to GNU atomic ordering for
222 /// libcalls.
223 static AtomicExpr::AtomicOrderingKind
224 translateAtomicOrdering(const llvm::AtomicOrdering AO);
225
226 private:
227 bool requiresMemSetZero(llvm::Type *type) const;
228
229 /// \brief Creates temp alloca for intermediate operations on atomic value.
230 llvm::Value *CreateTempAlloca() const;
231
232 /// \brief Emits atomic load as a libcall.
233 void EmitAtomicLoadLibcall(llvm::Value *AddForLoaded,
234 llvm::AtomicOrdering AO, bool IsVolatile);
235 /// \brief Emits atomic load as LLVM instruction.
236 llvm::Value *EmitAtomicLoadOp(llvm::AtomicOrdering AO, bool IsVolatile);
237 /// \brief Emits atomic compare-and-exchange op as a libcall.
238 std::pair<RValue, llvm::Value *> EmitAtomicCompareExchangeLibcall(
239 RValue Expected, RValue DesiredAddr,
240 llvm::AtomicOrdering Success = llvm::SequentiallyConsistent,
241 llvm::AtomicOrdering Failure = llvm::SequentiallyConsistent);
242 /// \brief Emits atomic compare-and-exchange op as LLVM instruction.
243 std::pair<RValue, llvm::Value *> EmitAtomicCompareExchangeOp(
244 RValue Expected, RValue Desired,
245 llvm::AtomicOrdering Success = llvm::SequentiallyConsistent,
246 llvm::AtomicOrdering Failure = llvm::SequentiallyConsistent,
247 bool IsWeak = false);
248 };
249 }
250
251 AtomicExpr::AtomicOrderingKind
translateAtomicOrdering(const llvm::AtomicOrdering AO)252 AtomicInfo::translateAtomicOrdering(const llvm::AtomicOrdering AO) {
253 switch (AO) {
254 case llvm::Unordered:
255 case llvm::NotAtomic:
256 case llvm::Monotonic:
257 return AtomicExpr::AO_ABI_memory_order_relaxed;
258 case llvm::Acquire:
259 return AtomicExpr::AO_ABI_memory_order_acquire;
260 case llvm::Release:
261 return AtomicExpr::AO_ABI_memory_order_release;
262 case llvm::AcquireRelease:
263 return AtomicExpr::AO_ABI_memory_order_acq_rel;
264 case llvm::SequentiallyConsistent:
265 return AtomicExpr::AO_ABI_memory_order_seq_cst;
266 }
267 llvm_unreachable("Unhandled AtomicOrdering");
268 }
269
CreateTempAlloca() const270 llvm::Value *AtomicInfo::CreateTempAlloca() const {
271 auto *TempAlloca = CGF.CreateMemTemp(
272 (LVal.isBitField() && ValueSizeInBits > AtomicSizeInBits) ? ValueTy
273 : AtomicTy,
274 "atomic-temp");
275 TempAlloca->setAlignment(getAtomicAlignment().getQuantity());
276 // Cast to pointer to value type for bitfields.
277 if (LVal.isBitField())
278 return CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
279 TempAlloca, getAtomicAddress()->getType());
280 return TempAlloca;
281 }
282
emitAtomicLibcall(CodeGenFunction & CGF,StringRef fnName,QualType resultType,CallArgList & args)283 static RValue emitAtomicLibcall(CodeGenFunction &CGF,
284 StringRef fnName,
285 QualType resultType,
286 CallArgList &args) {
287 const CGFunctionInfo &fnInfo =
288 CGF.CGM.getTypes().arrangeFreeFunctionCall(resultType, args,
289 FunctionType::ExtInfo(), RequiredArgs::All);
290 llvm::FunctionType *fnTy = CGF.CGM.getTypes().GetFunctionType(fnInfo);
291 llvm::Constant *fn = CGF.CGM.CreateRuntimeFunction(fnTy, fnName);
292 return CGF.EmitCall(fnInfo, fn, ReturnValueSlot(), args);
293 }
294
295 /// Does a store of the given IR type modify the full expected width?
isFullSizeType(CodeGenModule & CGM,llvm::Type * type,uint64_t expectedSize)296 static bool isFullSizeType(CodeGenModule &CGM, llvm::Type *type,
297 uint64_t expectedSize) {
298 return (CGM.getDataLayout().getTypeStoreSize(type) * 8 == expectedSize);
299 }
300
301 /// Does the atomic type require memsetting to zero before initialization?
302 ///
303 /// The IR type is provided as a way of making certain queries faster.
requiresMemSetZero(llvm::Type * type) const304 bool AtomicInfo::requiresMemSetZero(llvm::Type *type) const {
305 // If the atomic type has size padding, we definitely need a memset.
306 if (hasPadding()) return true;
307
308 // Otherwise, do some simple heuristics to try to avoid it:
309 switch (getEvaluationKind()) {
310 // For scalars and complexes, check whether the store size of the
311 // type uses the full size.
312 case TEK_Scalar:
313 return !isFullSizeType(CGF.CGM, type, AtomicSizeInBits);
314 case TEK_Complex:
315 return !isFullSizeType(CGF.CGM, type->getStructElementType(0),
316 AtomicSizeInBits / 2);
317
318 // Padding in structs has an undefined bit pattern. User beware.
319 case TEK_Aggregate:
320 return false;
321 }
322 llvm_unreachable("bad evaluation kind");
323 }
324
emitMemSetZeroIfNecessary() const325 bool AtomicInfo::emitMemSetZeroIfNecessary() const {
326 assert(LVal.isSimple());
327 llvm::Value *addr = LVal.getAddress();
328 if (!requiresMemSetZero(addr->getType()->getPointerElementType()))
329 return false;
330
331 CGF.Builder.CreateMemSet(
332 addr, llvm::ConstantInt::get(CGF.Int8Ty, 0),
333 CGF.getContext().toCharUnitsFromBits(AtomicSizeInBits).getQuantity(),
334 LVal.getAlignment().getQuantity());
335 return true;
336 }
337
emitAtomicCmpXchg(CodeGenFunction & CGF,AtomicExpr * E,bool IsWeak,llvm::Value * Dest,llvm::Value * Ptr,llvm::Value * Val1,llvm::Value * Val2,uint64_t Size,unsigned Align,llvm::AtomicOrdering SuccessOrder,llvm::AtomicOrdering FailureOrder)338 static void emitAtomicCmpXchg(CodeGenFunction &CGF, AtomicExpr *E, bool IsWeak,
339 llvm::Value *Dest, llvm::Value *Ptr,
340 llvm::Value *Val1, llvm::Value *Val2,
341 uint64_t Size, unsigned Align,
342 llvm::AtomicOrdering SuccessOrder,
343 llvm::AtomicOrdering FailureOrder) {
344 // Note that cmpxchg doesn't support weak cmpxchg, at least at the moment.
345 llvm::LoadInst *Expected = CGF.Builder.CreateLoad(Val1);
346 Expected->setAlignment(Align);
347 llvm::LoadInst *Desired = CGF.Builder.CreateLoad(Val2);
348 Desired->setAlignment(Align);
349
350 llvm::AtomicCmpXchgInst *Pair = CGF.Builder.CreateAtomicCmpXchg(
351 Ptr, Expected, Desired, SuccessOrder, FailureOrder);
352 Pair->setVolatile(E->isVolatile());
353 Pair->setWeak(IsWeak);
354
355 // Cmp holds the result of the compare-exchange operation: true on success,
356 // false on failure.
357 llvm::Value *Old = CGF.Builder.CreateExtractValue(Pair, 0);
358 llvm::Value *Cmp = CGF.Builder.CreateExtractValue(Pair, 1);
359
360 // This basic block is used to hold the store instruction if the operation
361 // failed.
362 llvm::BasicBlock *StoreExpectedBB =
363 CGF.createBasicBlock("cmpxchg.store_expected", CGF.CurFn);
364
365 // This basic block is the exit point of the operation, we should end up
366 // here regardless of whether or not the operation succeeded.
367 llvm::BasicBlock *ContinueBB =
368 CGF.createBasicBlock("cmpxchg.continue", CGF.CurFn);
369
370 // Update Expected if Expected isn't equal to Old, otherwise branch to the
371 // exit point.
372 CGF.Builder.CreateCondBr(Cmp, ContinueBB, StoreExpectedBB);
373
374 CGF.Builder.SetInsertPoint(StoreExpectedBB);
375 // Update the memory at Expected with Old's value.
376 llvm::StoreInst *StoreExpected = CGF.Builder.CreateStore(Old, Val1);
377 StoreExpected->setAlignment(Align);
378 // Finally, branch to the exit point.
379 CGF.Builder.CreateBr(ContinueBB);
380
381 CGF.Builder.SetInsertPoint(ContinueBB);
382 // Update the memory at Dest with Cmp's value.
383 CGF.EmitStoreOfScalar(Cmp, CGF.MakeAddrLValue(Dest, E->getType()));
384 return;
385 }
386
387 /// Given an ordering required on success, emit all possible cmpxchg
388 /// instructions to cope with the provided (but possibly only dynamically known)
389 /// FailureOrder.
emitAtomicCmpXchgFailureSet(CodeGenFunction & CGF,AtomicExpr * E,bool IsWeak,llvm::Value * Dest,llvm::Value * Ptr,llvm::Value * Val1,llvm::Value * Val2,llvm::Value * FailureOrderVal,uint64_t Size,unsigned Align,llvm::AtomicOrdering SuccessOrder)390 static void emitAtomicCmpXchgFailureSet(CodeGenFunction &CGF, AtomicExpr *E,
391 bool IsWeak, llvm::Value *Dest,
392 llvm::Value *Ptr, llvm::Value *Val1,
393 llvm::Value *Val2,
394 llvm::Value *FailureOrderVal,
395 uint64_t Size, unsigned Align,
396 llvm::AtomicOrdering SuccessOrder) {
397 llvm::AtomicOrdering FailureOrder;
398 if (llvm::ConstantInt *FO = dyn_cast<llvm::ConstantInt>(FailureOrderVal)) {
399 switch (FO->getSExtValue()) {
400 default:
401 FailureOrder = llvm::Monotonic;
402 break;
403 case AtomicExpr::AO_ABI_memory_order_consume:
404 case AtomicExpr::AO_ABI_memory_order_acquire:
405 FailureOrder = llvm::Acquire;
406 break;
407 case AtomicExpr::AO_ABI_memory_order_seq_cst:
408 FailureOrder = llvm::SequentiallyConsistent;
409 break;
410 }
411 if (FailureOrder >= SuccessOrder) {
412 // Don't assert on undefined behaviour.
413 FailureOrder =
414 llvm::AtomicCmpXchgInst::getStrongestFailureOrdering(SuccessOrder);
415 }
416 emitAtomicCmpXchg(CGF, E, IsWeak, Dest, Ptr, Val1, Val2, Size, Align,
417 SuccessOrder, FailureOrder);
418 return;
419 }
420
421 // Create all the relevant BB's
422 llvm::BasicBlock *MonotonicBB = nullptr, *AcquireBB = nullptr,
423 *SeqCstBB = nullptr;
424 MonotonicBB = CGF.createBasicBlock("monotonic_fail", CGF.CurFn);
425 if (SuccessOrder != llvm::Monotonic && SuccessOrder != llvm::Release)
426 AcquireBB = CGF.createBasicBlock("acquire_fail", CGF.CurFn);
427 if (SuccessOrder == llvm::SequentiallyConsistent)
428 SeqCstBB = CGF.createBasicBlock("seqcst_fail", CGF.CurFn);
429
430 llvm::BasicBlock *ContBB = CGF.createBasicBlock("atomic.continue", CGF.CurFn);
431
432 llvm::SwitchInst *SI = CGF.Builder.CreateSwitch(FailureOrderVal, MonotonicBB);
433
434 // Emit all the different atomics
435
436 // MonotonicBB is arbitrarily chosen as the default case; in practice, this
437 // doesn't matter unless someone is crazy enough to use something that
438 // doesn't fold to a constant for the ordering.
439 CGF.Builder.SetInsertPoint(MonotonicBB);
440 emitAtomicCmpXchg(CGF, E, IsWeak, Dest, Ptr, Val1, Val2,
441 Size, Align, SuccessOrder, llvm::Monotonic);
442 CGF.Builder.CreateBr(ContBB);
443
444 if (AcquireBB) {
445 CGF.Builder.SetInsertPoint(AcquireBB);
446 emitAtomicCmpXchg(CGF, E, IsWeak, Dest, Ptr, Val1, Val2,
447 Size, Align, SuccessOrder, llvm::Acquire);
448 CGF.Builder.CreateBr(ContBB);
449 SI->addCase(CGF.Builder.getInt32(AtomicExpr::AO_ABI_memory_order_consume),
450 AcquireBB);
451 SI->addCase(CGF.Builder.getInt32(AtomicExpr::AO_ABI_memory_order_acquire),
452 AcquireBB);
453 }
454 if (SeqCstBB) {
455 CGF.Builder.SetInsertPoint(SeqCstBB);
456 emitAtomicCmpXchg(CGF, E, IsWeak, Dest, Ptr, Val1, Val2,
457 Size, Align, SuccessOrder, llvm::SequentiallyConsistent);
458 CGF.Builder.CreateBr(ContBB);
459 SI->addCase(CGF.Builder.getInt32(AtomicExpr::AO_ABI_memory_order_seq_cst),
460 SeqCstBB);
461 }
462
463 CGF.Builder.SetInsertPoint(ContBB);
464 }
465
EmitAtomicOp(CodeGenFunction & CGF,AtomicExpr * E,llvm::Value * Dest,llvm::Value * Ptr,llvm::Value * Val1,llvm::Value * Val2,llvm::Value * IsWeak,llvm::Value * FailureOrder,uint64_t Size,unsigned Align,llvm::AtomicOrdering Order)466 static void EmitAtomicOp(CodeGenFunction &CGF, AtomicExpr *E, llvm::Value *Dest,
467 llvm::Value *Ptr, llvm::Value *Val1, llvm::Value *Val2,
468 llvm::Value *IsWeak, llvm::Value *FailureOrder,
469 uint64_t Size, unsigned Align,
470 llvm::AtomicOrdering Order) {
471 llvm::AtomicRMWInst::BinOp Op = llvm::AtomicRMWInst::Add;
472 llvm::Instruction::BinaryOps PostOp = (llvm::Instruction::BinaryOps)0;
473
474 switch (E->getOp()) {
475 case AtomicExpr::AO__c11_atomic_init:
476 llvm_unreachable("Already handled!");
477
478 case AtomicExpr::AO__c11_atomic_compare_exchange_strong:
479 emitAtomicCmpXchgFailureSet(CGF, E, false, Dest, Ptr, Val1, Val2,
480 FailureOrder, Size, Align, Order);
481 return;
482 case AtomicExpr::AO__c11_atomic_compare_exchange_weak:
483 emitAtomicCmpXchgFailureSet(CGF, E, true, Dest, Ptr, Val1, Val2,
484 FailureOrder, Size, Align, Order);
485 return;
486 case AtomicExpr::AO__atomic_compare_exchange:
487 case AtomicExpr::AO__atomic_compare_exchange_n: {
488 if (llvm::ConstantInt *IsWeakC = dyn_cast<llvm::ConstantInt>(IsWeak)) {
489 emitAtomicCmpXchgFailureSet(CGF, E, IsWeakC->getZExtValue(), Dest, Ptr,
490 Val1, Val2, FailureOrder, Size, Align, Order);
491 } else {
492 // Create all the relevant BB's
493 llvm::BasicBlock *StrongBB =
494 CGF.createBasicBlock("cmpxchg.strong", CGF.CurFn);
495 llvm::BasicBlock *WeakBB = CGF.createBasicBlock("cmxchg.weak", CGF.CurFn);
496 llvm::BasicBlock *ContBB =
497 CGF.createBasicBlock("cmpxchg.continue", CGF.CurFn);
498
499 llvm::SwitchInst *SI = CGF.Builder.CreateSwitch(IsWeak, WeakBB);
500 SI->addCase(CGF.Builder.getInt1(false), StrongBB);
501
502 CGF.Builder.SetInsertPoint(StrongBB);
503 emitAtomicCmpXchgFailureSet(CGF, E, false, Dest, Ptr, Val1, Val2,
504 FailureOrder, Size, Align, Order);
505 CGF.Builder.CreateBr(ContBB);
506
507 CGF.Builder.SetInsertPoint(WeakBB);
508 emitAtomicCmpXchgFailureSet(CGF, E, true, Dest, Ptr, Val1, Val2,
509 FailureOrder, Size, Align, Order);
510 CGF.Builder.CreateBr(ContBB);
511
512 CGF.Builder.SetInsertPoint(ContBB);
513 }
514 return;
515 }
516 case AtomicExpr::AO__c11_atomic_load:
517 case AtomicExpr::AO__atomic_load_n:
518 case AtomicExpr::AO__atomic_load: {
519 llvm::LoadInst *Load = CGF.Builder.CreateLoad(Ptr);
520 Load->setAtomic(Order);
521 Load->setAlignment(Size);
522 Load->setVolatile(E->isVolatile());
523 llvm::StoreInst *StoreDest = CGF.Builder.CreateStore(Load, Dest);
524 StoreDest->setAlignment(Align);
525 return;
526 }
527
528 case AtomicExpr::AO__c11_atomic_store:
529 case AtomicExpr::AO__atomic_store:
530 case AtomicExpr::AO__atomic_store_n: {
531 assert(!Dest && "Store does not return a value");
532 llvm::LoadInst *LoadVal1 = CGF.Builder.CreateLoad(Val1);
533 LoadVal1->setAlignment(Align);
534 llvm::StoreInst *Store = CGF.Builder.CreateStore(LoadVal1, Ptr);
535 Store->setAtomic(Order);
536 Store->setAlignment(Size);
537 Store->setVolatile(E->isVolatile());
538 return;
539 }
540
541 case AtomicExpr::AO__c11_atomic_exchange:
542 case AtomicExpr::AO__atomic_exchange_n:
543 case AtomicExpr::AO__atomic_exchange:
544 Op = llvm::AtomicRMWInst::Xchg;
545 break;
546
547 case AtomicExpr::AO__atomic_add_fetch:
548 PostOp = llvm::Instruction::Add;
549 // Fall through.
550 case AtomicExpr::AO__c11_atomic_fetch_add:
551 case AtomicExpr::AO__atomic_fetch_add:
552 Op = llvm::AtomicRMWInst::Add;
553 break;
554
555 case AtomicExpr::AO__atomic_sub_fetch:
556 PostOp = llvm::Instruction::Sub;
557 // Fall through.
558 case AtomicExpr::AO__c11_atomic_fetch_sub:
559 case AtomicExpr::AO__atomic_fetch_sub:
560 Op = llvm::AtomicRMWInst::Sub;
561 break;
562
563 case AtomicExpr::AO__atomic_and_fetch:
564 PostOp = llvm::Instruction::And;
565 // Fall through.
566 case AtomicExpr::AO__c11_atomic_fetch_and:
567 case AtomicExpr::AO__atomic_fetch_and:
568 Op = llvm::AtomicRMWInst::And;
569 break;
570
571 case AtomicExpr::AO__atomic_or_fetch:
572 PostOp = llvm::Instruction::Or;
573 // Fall through.
574 case AtomicExpr::AO__c11_atomic_fetch_or:
575 case AtomicExpr::AO__atomic_fetch_or:
576 Op = llvm::AtomicRMWInst::Or;
577 break;
578
579 case AtomicExpr::AO__atomic_xor_fetch:
580 PostOp = llvm::Instruction::Xor;
581 // Fall through.
582 case AtomicExpr::AO__c11_atomic_fetch_xor:
583 case AtomicExpr::AO__atomic_fetch_xor:
584 Op = llvm::AtomicRMWInst::Xor;
585 break;
586
587 case AtomicExpr::AO__atomic_nand_fetch:
588 PostOp = llvm::Instruction::And;
589 // Fall through.
590 case AtomicExpr::AO__atomic_fetch_nand:
591 Op = llvm::AtomicRMWInst::Nand;
592 break;
593 }
594
595 llvm::LoadInst *LoadVal1 = CGF.Builder.CreateLoad(Val1);
596 LoadVal1->setAlignment(Align);
597 llvm::AtomicRMWInst *RMWI =
598 CGF.Builder.CreateAtomicRMW(Op, Ptr, LoadVal1, Order);
599 RMWI->setVolatile(E->isVolatile());
600
601 // For __atomic_*_fetch operations, perform the operation again to
602 // determine the value which was written.
603 llvm::Value *Result = RMWI;
604 if (PostOp)
605 Result = CGF.Builder.CreateBinOp(PostOp, RMWI, LoadVal1);
606 if (E->getOp() == AtomicExpr::AO__atomic_nand_fetch)
607 Result = CGF.Builder.CreateNot(Result);
608 llvm::StoreInst *StoreDest = CGF.Builder.CreateStore(Result, Dest);
609 StoreDest->setAlignment(Align);
610 }
611
612 // This function emits any expression (scalar, complex, or aggregate)
613 // into a temporary alloca.
614 static llvm::Value *
EmitValToTemp(CodeGenFunction & CGF,Expr * E)615 EmitValToTemp(CodeGenFunction &CGF, Expr *E) {
616 llvm::Value *DeclPtr = CGF.CreateMemTemp(E->getType(), ".atomictmp");
617 CGF.EmitAnyExprToMem(E, DeclPtr, E->getType().getQualifiers(),
618 /*Init*/ true);
619 return DeclPtr;
620 }
621
622 static void
AddDirectArgument(CodeGenFunction & CGF,CallArgList & Args,bool UseOptimizedLibcall,llvm::Value * Val,QualType ValTy,SourceLocation Loc,CharUnits SizeInChars)623 AddDirectArgument(CodeGenFunction &CGF, CallArgList &Args,
624 bool UseOptimizedLibcall, llvm::Value *Val, QualType ValTy,
625 SourceLocation Loc, CharUnits SizeInChars) {
626 if (UseOptimizedLibcall) {
627 // Load value and pass it to the function directly.
628 unsigned Align = CGF.getContext().getTypeAlignInChars(ValTy).getQuantity();
629 int64_t SizeInBits = CGF.getContext().toBits(SizeInChars);
630 ValTy =
631 CGF.getContext().getIntTypeForBitwidth(SizeInBits, /*Signed=*/false);
632 llvm::Type *IPtrTy = llvm::IntegerType::get(CGF.getLLVMContext(),
633 SizeInBits)->getPointerTo();
634 Val = CGF.EmitLoadOfScalar(CGF.Builder.CreateBitCast(Val, IPtrTy), false,
635 Align, CGF.getContext().getPointerType(ValTy),
636 Loc);
637 // Coerce the value into an appropriately sized integer type.
638 Args.add(RValue::get(Val), ValTy);
639 } else {
640 // Non-optimized functions always take a reference.
641 Args.add(RValue::get(CGF.EmitCastToVoidPtr(Val)),
642 CGF.getContext().VoidPtrTy);
643 }
644 }
645
EmitAtomicExpr(AtomicExpr * E,llvm::Value * Dest)646 RValue CodeGenFunction::EmitAtomicExpr(AtomicExpr *E, llvm::Value *Dest) {
647 QualType AtomicTy = E->getPtr()->getType()->getPointeeType();
648 QualType MemTy = AtomicTy;
649 if (const AtomicType *AT = AtomicTy->getAs<AtomicType>())
650 MemTy = AT->getValueType();
651 CharUnits sizeChars = getContext().getTypeSizeInChars(AtomicTy);
652 uint64_t Size = sizeChars.getQuantity();
653 CharUnits alignChars = getContext().getTypeAlignInChars(AtomicTy);
654 unsigned Align = alignChars.getQuantity();
655 unsigned MaxInlineWidthInBits =
656 getTarget().getMaxAtomicInlineWidth();
657 bool UseLibcall = (Size != Align ||
658 getContext().toBits(sizeChars) > MaxInlineWidthInBits);
659
660 llvm::Value *IsWeak = nullptr, *OrderFail = nullptr, *Val1 = nullptr,
661 *Val2 = nullptr;
662 llvm::Value *Ptr = EmitScalarExpr(E->getPtr());
663
664 if (E->getOp() == AtomicExpr::AO__c11_atomic_init) {
665 assert(!Dest && "Init does not return a value");
666 LValue lvalue = LValue::MakeAddr(Ptr, AtomicTy, alignChars, getContext());
667 EmitAtomicInit(E->getVal1(), lvalue);
668 return RValue::get(nullptr);
669 }
670
671 llvm::Value *Order = EmitScalarExpr(E->getOrder());
672
673 switch (E->getOp()) {
674 case AtomicExpr::AO__c11_atomic_init:
675 llvm_unreachable("Already handled!");
676
677 case AtomicExpr::AO__c11_atomic_load:
678 case AtomicExpr::AO__atomic_load_n:
679 break;
680
681 case AtomicExpr::AO__atomic_load:
682 Dest = EmitScalarExpr(E->getVal1());
683 break;
684
685 case AtomicExpr::AO__atomic_store:
686 Val1 = EmitScalarExpr(E->getVal1());
687 break;
688
689 case AtomicExpr::AO__atomic_exchange:
690 Val1 = EmitScalarExpr(E->getVal1());
691 Dest = EmitScalarExpr(E->getVal2());
692 break;
693
694 case AtomicExpr::AO__c11_atomic_compare_exchange_strong:
695 case AtomicExpr::AO__c11_atomic_compare_exchange_weak:
696 case AtomicExpr::AO__atomic_compare_exchange_n:
697 case AtomicExpr::AO__atomic_compare_exchange:
698 Val1 = EmitScalarExpr(E->getVal1());
699 if (E->getOp() == AtomicExpr::AO__atomic_compare_exchange)
700 Val2 = EmitScalarExpr(E->getVal2());
701 else
702 Val2 = EmitValToTemp(*this, E->getVal2());
703 OrderFail = EmitScalarExpr(E->getOrderFail());
704 if (E->getNumSubExprs() == 6)
705 IsWeak = EmitScalarExpr(E->getWeak());
706 break;
707
708 case AtomicExpr::AO__c11_atomic_fetch_add:
709 case AtomicExpr::AO__c11_atomic_fetch_sub:
710 if (MemTy->isPointerType()) {
711 // For pointer arithmetic, we're required to do a bit of math:
712 // adding 1 to an int* is not the same as adding 1 to a uintptr_t.
713 // ... but only for the C11 builtins. The GNU builtins expect the
714 // user to multiply by sizeof(T).
715 QualType Val1Ty = E->getVal1()->getType();
716 llvm::Value *Val1Scalar = EmitScalarExpr(E->getVal1());
717 CharUnits PointeeIncAmt =
718 getContext().getTypeSizeInChars(MemTy->getPointeeType());
719 Val1Scalar = Builder.CreateMul(Val1Scalar, CGM.getSize(PointeeIncAmt));
720 Val1 = CreateMemTemp(Val1Ty, ".atomictmp");
721 EmitStoreOfScalar(Val1Scalar, MakeAddrLValue(Val1, Val1Ty));
722 break;
723 }
724 // Fall through.
725 case AtomicExpr::AO__atomic_fetch_add:
726 case AtomicExpr::AO__atomic_fetch_sub:
727 case AtomicExpr::AO__atomic_add_fetch:
728 case AtomicExpr::AO__atomic_sub_fetch:
729 case AtomicExpr::AO__c11_atomic_store:
730 case AtomicExpr::AO__c11_atomic_exchange:
731 case AtomicExpr::AO__atomic_store_n:
732 case AtomicExpr::AO__atomic_exchange_n:
733 case AtomicExpr::AO__c11_atomic_fetch_and:
734 case AtomicExpr::AO__c11_atomic_fetch_or:
735 case AtomicExpr::AO__c11_atomic_fetch_xor:
736 case AtomicExpr::AO__atomic_fetch_and:
737 case AtomicExpr::AO__atomic_fetch_or:
738 case AtomicExpr::AO__atomic_fetch_xor:
739 case AtomicExpr::AO__atomic_fetch_nand:
740 case AtomicExpr::AO__atomic_and_fetch:
741 case AtomicExpr::AO__atomic_or_fetch:
742 case AtomicExpr::AO__atomic_xor_fetch:
743 case AtomicExpr::AO__atomic_nand_fetch:
744 Val1 = EmitValToTemp(*this, E->getVal1());
745 break;
746 }
747
748 QualType RValTy = E->getType().getUnqualifiedType();
749
750 auto GetDest = [&] {
751 if (!RValTy->isVoidType() && !Dest) {
752 Dest = CreateMemTemp(RValTy, ".atomicdst");
753 }
754 return Dest;
755 };
756
757 // Use a library call. See: http://gcc.gnu.org/wiki/Atomic/GCCMM/LIbrary .
758 if (UseLibcall) {
759 bool UseOptimizedLibcall = false;
760 switch (E->getOp()) {
761 case AtomicExpr::AO__c11_atomic_fetch_add:
762 case AtomicExpr::AO__atomic_fetch_add:
763 case AtomicExpr::AO__c11_atomic_fetch_and:
764 case AtomicExpr::AO__atomic_fetch_and:
765 case AtomicExpr::AO__c11_atomic_fetch_or:
766 case AtomicExpr::AO__atomic_fetch_or:
767 case AtomicExpr::AO__c11_atomic_fetch_sub:
768 case AtomicExpr::AO__atomic_fetch_sub:
769 case AtomicExpr::AO__c11_atomic_fetch_xor:
770 case AtomicExpr::AO__atomic_fetch_xor:
771 // For these, only library calls for certain sizes exist.
772 UseOptimizedLibcall = true;
773 break;
774 default:
775 // Only use optimized library calls for sizes for which they exist.
776 if (Size == 1 || Size == 2 || Size == 4 || Size == 8)
777 UseOptimizedLibcall = true;
778 break;
779 }
780
781 CallArgList Args;
782 if (!UseOptimizedLibcall) {
783 // For non-optimized library calls, the size is the first parameter
784 Args.add(RValue::get(llvm::ConstantInt::get(SizeTy, Size)),
785 getContext().getSizeType());
786 }
787 // Atomic address is the first or second parameter
788 Args.add(RValue::get(EmitCastToVoidPtr(Ptr)), getContext().VoidPtrTy);
789
790 std::string LibCallName;
791 QualType LoweredMemTy =
792 MemTy->isPointerType() ? getContext().getIntPtrType() : MemTy;
793 QualType RetTy;
794 bool HaveRetTy = false;
795 switch (E->getOp()) {
796 // There is only one libcall for compare an exchange, because there is no
797 // optimisation benefit possible from a libcall version of a weak compare
798 // and exchange.
799 // bool __atomic_compare_exchange(size_t size, void *mem, void *expected,
800 // void *desired, int success, int failure)
801 // bool __atomic_compare_exchange_N(T *mem, T *expected, T desired,
802 // int success, int failure)
803 case AtomicExpr::AO__c11_atomic_compare_exchange_weak:
804 case AtomicExpr::AO__c11_atomic_compare_exchange_strong:
805 case AtomicExpr::AO__atomic_compare_exchange:
806 case AtomicExpr::AO__atomic_compare_exchange_n:
807 LibCallName = "__atomic_compare_exchange";
808 RetTy = getContext().BoolTy;
809 HaveRetTy = true;
810 Args.add(RValue::get(EmitCastToVoidPtr(Val1)), getContext().VoidPtrTy);
811 AddDirectArgument(*this, Args, UseOptimizedLibcall, Val2, MemTy,
812 E->getExprLoc(), sizeChars);
813 Args.add(RValue::get(Order), getContext().IntTy);
814 Order = OrderFail;
815 break;
816 // void __atomic_exchange(size_t size, void *mem, void *val, void *return,
817 // int order)
818 // T __atomic_exchange_N(T *mem, T val, int order)
819 case AtomicExpr::AO__c11_atomic_exchange:
820 case AtomicExpr::AO__atomic_exchange_n:
821 case AtomicExpr::AO__atomic_exchange:
822 LibCallName = "__atomic_exchange";
823 AddDirectArgument(*this, Args, UseOptimizedLibcall, Val1, MemTy,
824 E->getExprLoc(), sizeChars);
825 break;
826 // void __atomic_store(size_t size, void *mem, void *val, int order)
827 // void __atomic_store_N(T *mem, T val, int order)
828 case AtomicExpr::AO__c11_atomic_store:
829 case AtomicExpr::AO__atomic_store:
830 case AtomicExpr::AO__atomic_store_n:
831 LibCallName = "__atomic_store";
832 RetTy = getContext().VoidTy;
833 HaveRetTy = true;
834 AddDirectArgument(*this, Args, UseOptimizedLibcall, Val1, MemTy,
835 E->getExprLoc(), sizeChars);
836 break;
837 // void __atomic_load(size_t size, void *mem, void *return, int order)
838 // T __atomic_load_N(T *mem, int order)
839 case AtomicExpr::AO__c11_atomic_load:
840 case AtomicExpr::AO__atomic_load:
841 case AtomicExpr::AO__atomic_load_n:
842 LibCallName = "__atomic_load";
843 break;
844 // T __atomic_fetch_add_N(T *mem, T val, int order)
845 case AtomicExpr::AO__c11_atomic_fetch_add:
846 case AtomicExpr::AO__atomic_fetch_add:
847 LibCallName = "__atomic_fetch_add";
848 AddDirectArgument(*this, Args, UseOptimizedLibcall, Val1, LoweredMemTy,
849 E->getExprLoc(), sizeChars);
850 break;
851 // T __atomic_fetch_and_N(T *mem, T val, int order)
852 case AtomicExpr::AO__c11_atomic_fetch_and:
853 case AtomicExpr::AO__atomic_fetch_and:
854 LibCallName = "__atomic_fetch_and";
855 AddDirectArgument(*this, Args, UseOptimizedLibcall, Val1, MemTy,
856 E->getExprLoc(), sizeChars);
857 break;
858 // T __atomic_fetch_or_N(T *mem, T val, int order)
859 case AtomicExpr::AO__c11_atomic_fetch_or:
860 case AtomicExpr::AO__atomic_fetch_or:
861 LibCallName = "__atomic_fetch_or";
862 AddDirectArgument(*this, Args, UseOptimizedLibcall, Val1, MemTy,
863 E->getExprLoc(), sizeChars);
864 break;
865 // T __atomic_fetch_sub_N(T *mem, T val, int order)
866 case AtomicExpr::AO__c11_atomic_fetch_sub:
867 case AtomicExpr::AO__atomic_fetch_sub:
868 LibCallName = "__atomic_fetch_sub";
869 AddDirectArgument(*this, Args, UseOptimizedLibcall, Val1, LoweredMemTy,
870 E->getExprLoc(), sizeChars);
871 break;
872 // T __atomic_fetch_xor_N(T *mem, T val, int order)
873 case AtomicExpr::AO__c11_atomic_fetch_xor:
874 case AtomicExpr::AO__atomic_fetch_xor:
875 LibCallName = "__atomic_fetch_xor";
876 AddDirectArgument(*this, Args, UseOptimizedLibcall, Val1, MemTy,
877 E->getExprLoc(), sizeChars);
878 break;
879 default: return EmitUnsupportedRValue(E, "atomic library call");
880 }
881
882 // Optimized functions have the size in their name.
883 if (UseOptimizedLibcall)
884 LibCallName += "_" + llvm::utostr(Size);
885 // By default, assume we return a value of the atomic type.
886 if (!HaveRetTy) {
887 if (UseOptimizedLibcall) {
888 // Value is returned directly.
889 // The function returns an appropriately sized integer type.
890 RetTy = getContext().getIntTypeForBitwidth(
891 getContext().toBits(sizeChars), /*Signed=*/false);
892 } else {
893 // Value is returned through parameter before the order.
894 RetTy = getContext().VoidTy;
895 Args.add(RValue::get(EmitCastToVoidPtr(Dest)), getContext().VoidPtrTy);
896 }
897 }
898 // order is always the last parameter
899 Args.add(RValue::get(Order),
900 getContext().IntTy);
901
902 RValue Res = emitAtomicLibcall(*this, LibCallName, RetTy, Args);
903 // The value is returned directly from the libcall.
904 if (HaveRetTy && !RetTy->isVoidType())
905 return Res;
906 // The value is returned via an explicit out param.
907 if (RetTy->isVoidType())
908 return RValue::get(nullptr);
909 // The value is returned directly for optimized libcalls but the caller is
910 // expected an out-param.
911 if (UseOptimizedLibcall) {
912 llvm::Value *ResVal = Res.getScalarVal();
913 llvm::StoreInst *StoreDest = Builder.CreateStore(
914 ResVal,
915 Builder.CreateBitCast(GetDest(), ResVal->getType()->getPointerTo()));
916 StoreDest->setAlignment(Align);
917 }
918 return convertTempToRValue(Dest, RValTy, E->getExprLoc());
919 }
920
921 bool IsStore = E->getOp() == AtomicExpr::AO__c11_atomic_store ||
922 E->getOp() == AtomicExpr::AO__atomic_store ||
923 E->getOp() == AtomicExpr::AO__atomic_store_n;
924 bool IsLoad = E->getOp() == AtomicExpr::AO__c11_atomic_load ||
925 E->getOp() == AtomicExpr::AO__atomic_load ||
926 E->getOp() == AtomicExpr::AO__atomic_load_n;
927
928 llvm::Type *ITy =
929 llvm::IntegerType::get(getLLVMContext(), Size * 8);
930 llvm::Value *OrigDest = GetDest();
931 Ptr = Builder.CreateBitCast(
932 Ptr, ITy->getPointerTo(Ptr->getType()->getPointerAddressSpace()));
933 if (Val1) Val1 = Builder.CreateBitCast(Val1, ITy->getPointerTo());
934 if (Val2) Val2 = Builder.CreateBitCast(Val2, ITy->getPointerTo());
935 if (Dest && !E->isCmpXChg())
936 Dest = Builder.CreateBitCast(Dest, ITy->getPointerTo());
937
938 if (isa<llvm::ConstantInt>(Order)) {
939 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue();
940 switch (ord) {
941 case AtomicExpr::AO_ABI_memory_order_relaxed:
942 EmitAtomicOp(*this, E, Dest, Ptr, Val1, Val2, IsWeak, OrderFail,
943 Size, Align, llvm::Monotonic);
944 break;
945 case AtomicExpr::AO_ABI_memory_order_consume:
946 case AtomicExpr::AO_ABI_memory_order_acquire:
947 if (IsStore)
948 break; // Avoid crashing on code with undefined behavior
949 EmitAtomicOp(*this, E, Dest, Ptr, Val1, Val2, IsWeak, OrderFail,
950 Size, Align, llvm::Acquire);
951 break;
952 case AtomicExpr::AO_ABI_memory_order_release:
953 if (IsLoad)
954 break; // Avoid crashing on code with undefined behavior
955 EmitAtomicOp(*this, E, Dest, Ptr, Val1, Val2, IsWeak, OrderFail,
956 Size, Align, llvm::Release);
957 break;
958 case AtomicExpr::AO_ABI_memory_order_acq_rel:
959 if (IsLoad || IsStore)
960 break; // Avoid crashing on code with undefined behavior
961 EmitAtomicOp(*this, E, Dest, Ptr, Val1, Val2, IsWeak, OrderFail,
962 Size, Align, llvm::AcquireRelease);
963 break;
964 case AtomicExpr::AO_ABI_memory_order_seq_cst:
965 EmitAtomicOp(*this, E, Dest, Ptr, Val1, Val2, IsWeak, OrderFail,
966 Size, Align, llvm::SequentiallyConsistent);
967 break;
968 default: // invalid order
969 // We should not ever get here normally, but it's hard to
970 // enforce that in general.
971 break;
972 }
973 if (RValTy->isVoidType())
974 return RValue::get(nullptr);
975 return convertTempToRValue(OrigDest, RValTy, E->getExprLoc());
976 }
977
978 // Long case, when Order isn't obviously constant.
979
980 // Create all the relevant BB's
981 llvm::BasicBlock *MonotonicBB = nullptr, *AcquireBB = nullptr,
982 *ReleaseBB = nullptr, *AcqRelBB = nullptr,
983 *SeqCstBB = nullptr;
984 MonotonicBB = createBasicBlock("monotonic", CurFn);
985 if (!IsStore)
986 AcquireBB = createBasicBlock("acquire", CurFn);
987 if (!IsLoad)
988 ReleaseBB = createBasicBlock("release", CurFn);
989 if (!IsLoad && !IsStore)
990 AcqRelBB = createBasicBlock("acqrel", CurFn);
991 SeqCstBB = createBasicBlock("seqcst", CurFn);
992 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn);
993
994 // Create the switch for the split
995 // MonotonicBB is arbitrarily chosen as the default case; in practice, this
996 // doesn't matter unless someone is crazy enough to use something that
997 // doesn't fold to a constant for the ordering.
998 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false);
999 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, MonotonicBB);
1000
1001 // Emit all the different atomics
1002 Builder.SetInsertPoint(MonotonicBB);
1003 EmitAtomicOp(*this, E, Dest, Ptr, Val1, Val2, IsWeak, OrderFail,
1004 Size, Align, llvm::Monotonic);
1005 Builder.CreateBr(ContBB);
1006 if (!IsStore) {
1007 Builder.SetInsertPoint(AcquireBB);
1008 EmitAtomicOp(*this, E, Dest, Ptr, Val1, Val2, IsWeak, OrderFail,
1009 Size, Align, llvm::Acquire);
1010 Builder.CreateBr(ContBB);
1011 SI->addCase(Builder.getInt32(AtomicExpr::AO_ABI_memory_order_consume),
1012 AcquireBB);
1013 SI->addCase(Builder.getInt32(AtomicExpr::AO_ABI_memory_order_acquire),
1014 AcquireBB);
1015 }
1016 if (!IsLoad) {
1017 Builder.SetInsertPoint(ReleaseBB);
1018 EmitAtomicOp(*this, E, Dest, Ptr, Val1, Val2, IsWeak, OrderFail,
1019 Size, Align, llvm::Release);
1020 Builder.CreateBr(ContBB);
1021 SI->addCase(Builder.getInt32(AtomicExpr::AO_ABI_memory_order_release),
1022 ReleaseBB);
1023 }
1024 if (!IsLoad && !IsStore) {
1025 Builder.SetInsertPoint(AcqRelBB);
1026 EmitAtomicOp(*this, E, Dest, Ptr, Val1, Val2, IsWeak, OrderFail,
1027 Size, Align, llvm::AcquireRelease);
1028 Builder.CreateBr(ContBB);
1029 SI->addCase(Builder.getInt32(AtomicExpr::AO_ABI_memory_order_acq_rel),
1030 AcqRelBB);
1031 }
1032 Builder.SetInsertPoint(SeqCstBB);
1033 EmitAtomicOp(*this, E, Dest, Ptr, Val1, Val2, IsWeak, OrderFail,
1034 Size, Align, llvm::SequentiallyConsistent);
1035 Builder.CreateBr(ContBB);
1036 SI->addCase(Builder.getInt32(AtomicExpr::AO_ABI_memory_order_seq_cst),
1037 SeqCstBB);
1038
1039 // Cleanup and return
1040 Builder.SetInsertPoint(ContBB);
1041 if (RValTy->isVoidType())
1042 return RValue::get(nullptr);
1043 return convertTempToRValue(OrigDest, RValTy, E->getExprLoc());
1044 }
1045
emitCastToAtomicIntPointer(llvm::Value * addr) const1046 llvm::Value *AtomicInfo::emitCastToAtomicIntPointer(llvm::Value *addr) const {
1047 unsigned addrspace =
1048 cast<llvm::PointerType>(addr->getType())->getAddressSpace();
1049 llvm::IntegerType *ty =
1050 llvm::IntegerType::get(CGF.getLLVMContext(), AtomicSizeInBits);
1051 return CGF.Builder.CreateBitCast(addr, ty->getPointerTo(addrspace));
1052 }
1053
convertTempToRValue(llvm::Value * addr,AggValueSlot resultSlot,SourceLocation loc,bool AsValue) const1054 RValue AtomicInfo::convertTempToRValue(llvm::Value *addr,
1055 AggValueSlot resultSlot,
1056 SourceLocation loc, bool AsValue) const {
1057 if (LVal.isSimple()) {
1058 if (EvaluationKind == TEK_Aggregate)
1059 return resultSlot.asRValue();
1060
1061 // Drill into the padding structure if we have one.
1062 if (hasPadding())
1063 addr = CGF.Builder.CreateStructGEP(nullptr, addr, 0);
1064
1065 // Otherwise, just convert the temporary to an r-value using the
1066 // normal conversion routine.
1067 return CGF.convertTempToRValue(addr, getValueType(), loc);
1068 }
1069 if (!AsValue)
1070 // Get RValue from temp memory as atomic for non-simple lvalues
1071 return RValue::get(
1072 CGF.Builder.CreateAlignedLoad(addr, AtomicAlign.getQuantity()));
1073 if (LVal.isBitField())
1074 return CGF.EmitLoadOfBitfieldLValue(LValue::MakeBitfield(
1075 addr, LVal.getBitFieldInfo(), LVal.getType(), LVal.getAlignment()));
1076 if (LVal.isVectorElt())
1077 return CGF.EmitLoadOfLValue(LValue::MakeVectorElt(addr, LVal.getVectorIdx(),
1078 LVal.getType(),
1079 LVal.getAlignment()),
1080 loc);
1081 assert(LVal.isExtVectorElt());
1082 return CGF.EmitLoadOfExtVectorElementLValue(LValue::MakeExtVectorElt(
1083 addr, LVal.getExtVectorElts(), LVal.getType(), LVal.getAlignment()));
1084 }
1085
ConvertIntToValueOrAtomic(llvm::Value * IntVal,AggValueSlot ResultSlot,SourceLocation Loc,bool AsValue) const1086 RValue AtomicInfo::ConvertIntToValueOrAtomic(llvm::Value *IntVal,
1087 AggValueSlot ResultSlot,
1088 SourceLocation Loc,
1089 bool AsValue) const {
1090 // Try not to in some easy cases.
1091 assert(IntVal->getType()->isIntegerTy() && "Expected integer value");
1092 if (getEvaluationKind() == TEK_Scalar &&
1093 (((!LVal.isBitField() ||
1094 LVal.getBitFieldInfo().Size == ValueSizeInBits) &&
1095 !hasPadding()) ||
1096 !AsValue)) {
1097 auto *ValTy = AsValue
1098 ? CGF.ConvertTypeForMem(ValueTy)
1099 : getAtomicAddress()->getType()->getPointerElementType();
1100 if (ValTy->isIntegerTy()) {
1101 assert(IntVal->getType() == ValTy && "Different integer types.");
1102 return RValue::get(CGF.EmitFromMemory(IntVal, ValueTy));
1103 } else if (ValTy->isPointerTy())
1104 return RValue::get(CGF.Builder.CreateIntToPtr(IntVal, ValTy));
1105 else if (llvm::CastInst::isBitCastable(IntVal->getType(), ValTy))
1106 return RValue::get(CGF.Builder.CreateBitCast(IntVal, ValTy));
1107 }
1108
1109 // Create a temporary. This needs to be big enough to hold the
1110 // atomic integer.
1111 llvm::Value *Temp;
1112 bool TempIsVolatile = false;
1113 CharUnits TempAlignment;
1114 if (AsValue && getEvaluationKind() == TEK_Aggregate) {
1115 assert(!ResultSlot.isIgnored());
1116 Temp = ResultSlot.getAddr();
1117 TempAlignment = getValueAlignment();
1118 TempIsVolatile = ResultSlot.isVolatile();
1119 } else {
1120 Temp = CreateTempAlloca();
1121 TempAlignment = getAtomicAlignment();
1122 }
1123
1124 // Slam the integer into the temporary.
1125 llvm::Value *CastTemp = emitCastToAtomicIntPointer(Temp);
1126 CGF.Builder.CreateAlignedStore(IntVal, CastTemp, TempAlignment.getQuantity())
1127 ->setVolatile(TempIsVolatile);
1128
1129 return convertTempToRValue(Temp, ResultSlot, Loc, AsValue);
1130 }
1131
EmitAtomicLoadLibcall(llvm::Value * AddForLoaded,llvm::AtomicOrdering AO,bool)1132 void AtomicInfo::EmitAtomicLoadLibcall(llvm::Value *AddForLoaded,
1133 llvm::AtomicOrdering AO, bool) {
1134 // void __atomic_load(size_t size, void *mem, void *return, int order);
1135 CallArgList Args;
1136 Args.add(RValue::get(getAtomicSizeValue()), CGF.getContext().getSizeType());
1137 Args.add(RValue::get(CGF.EmitCastToVoidPtr(getAtomicAddress())),
1138 CGF.getContext().VoidPtrTy);
1139 Args.add(RValue::get(CGF.EmitCastToVoidPtr(AddForLoaded)),
1140 CGF.getContext().VoidPtrTy);
1141 Args.add(RValue::get(
1142 llvm::ConstantInt::get(CGF.IntTy, translateAtomicOrdering(AO))),
1143 CGF.getContext().IntTy);
1144 emitAtomicLibcall(CGF, "__atomic_load", CGF.getContext().VoidTy, Args);
1145 }
1146
EmitAtomicLoadOp(llvm::AtomicOrdering AO,bool IsVolatile)1147 llvm::Value *AtomicInfo::EmitAtomicLoadOp(llvm::AtomicOrdering AO,
1148 bool IsVolatile) {
1149 // Okay, we're doing this natively.
1150 llvm::Value *Addr = emitCastToAtomicIntPointer(getAtomicAddress());
1151 llvm::LoadInst *Load = CGF.Builder.CreateLoad(Addr, "atomic-load");
1152 Load->setAtomic(AO);
1153
1154 // Other decoration.
1155 Load->setAlignment(getAtomicAlignment().getQuantity());
1156 if (IsVolatile)
1157 Load->setVolatile(true);
1158 if (LVal.getTBAAInfo())
1159 CGF.CGM.DecorateInstruction(Load, LVal.getTBAAInfo());
1160 return Load;
1161 }
1162
1163 /// An LValue is a candidate for having its loads and stores be made atomic if
1164 /// we are operating under /volatile:ms *and* the LValue itself is volatile and
1165 /// performing such an operation can be performed without a libcall.
LValueIsSuitableForInlineAtomic(LValue LV)1166 bool CodeGenFunction::LValueIsSuitableForInlineAtomic(LValue LV) {
1167 AtomicInfo AI(*this, LV);
1168 bool IsVolatile = LV.isVolatile() || hasVolatileMember(LV.getType());
1169 // An atomic is inline if we don't need to use a libcall.
1170 bool AtomicIsInline = !AI.shouldUseLibcall();
1171 return CGM.getCodeGenOpts().MSVolatile && IsVolatile && AtomicIsInline;
1172 }
1173
1174 /// An type is a candidate for having its loads and stores be made atomic if
1175 /// we are operating under /volatile:ms *and* we know the access is volatile and
1176 /// performing such an operation can be performed without a libcall.
typeIsSuitableForInlineAtomic(QualType Ty,bool IsVolatile) const1177 bool CodeGenFunction::typeIsSuitableForInlineAtomic(QualType Ty,
1178 bool IsVolatile) const {
1179 // An atomic is inline if we don't need to use a libcall (e.g. it is builtin).
1180 bool AtomicIsInline = getContext().getTargetInfo().hasBuiltinAtomic(
1181 getContext().getTypeSize(Ty), getContext().getTypeAlign(Ty));
1182 return CGM.getCodeGenOpts().MSVolatile && IsVolatile && AtomicIsInline;
1183 }
1184
EmitAtomicLoad(LValue LV,SourceLocation SL,AggValueSlot Slot)1185 RValue CodeGenFunction::EmitAtomicLoad(LValue LV, SourceLocation SL,
1186 AggValueSlot Slot) {
1187 llvm::AtomicOrdering AO;
1188 bool IsVolatile = LV.isVolatileQualified();
1189 if (LV.getType()->isAtomicType()) {
1190 AO = llvm::SequentiallyConsistent;
1191 } else {
1192 AO = llvm::Acquire;
1193 IsVolatile = true;
1194 }
1195 return EmitAtomicLoad(LV, SL, AO, IsVolatile, Slot);
1196 }
1197
EmitAtomicLoad(AggValueSlot ResultSlot,SourceLocation Loc,bool AsValue,llvm::AtomicOrdering AO,bool IsVolatile)1198 RValue AtomicInfo::EmitAtomicLoad(AggValueSlot ResultSlot, SourceLocation Loc,
1199 bool AsValue, llvm::AtomicOrdering AO,
1200 bool IsVolatile) {
1201 // Check whether we should use a library call.
1202 if (shouldUseLibcall()) {
1203 llvm::Value *TempAddr;
1204 if (LVal.isSimple() && !ResultSlot.isIgnored()) {
1205 assert(getEvaluationKind() == TEK_Aggregate);
1206 TempAddr = ResultSlot.getAddr();
1207 } else
1208 TempAddr = CreateTempAlloca();
1209
1210 EmitAtomicLoadLibcall(TempAddr, AO, IsVolatile);
1211
1212 // Okay, turn that back into the original value or whole atomic (for
1213 // non-simple lvalues) type.
1214 return convertTempToRValue(TempAddr, ResultSlot, Loc, AsValue);
1215 }
1216
1217 // Okay, we're doing this natively.
1218 auto *Load = EmitAtomicLoadOp(AO, IsVolatile);
1219
1220 // If we're ignoring an aggregate return, don't do anything.
1221 if (getEvaluationKind() == TEK_Aggregate && ResultSlot.isIgnored())
1222 return RValue::getAggregate(nullptr, false);
1223
1224 // Okay, turn that back into the original value or atomic (for non-simple
1225 // lvalues) type.
1226 return ConvertIntToValueOrAtomic(Load, ResultSlot, Loc, AsValue);
1227 }
1228
1229 /// Emit a load from an l-value of atomic type. Note that the r-value
1230 /// we produce is an r-value of the atomic *value* type.
EmitAtomicLoad(LValue src,SourceLocation loc,llvm::AtomicOrdering AO,bool IsVolatile,AggValueSlot resultSlot)1231 RValue CodeGenFunction::EmitAtomicLoad(LValue src, SourceLocation loc,
1232 llvm::AtomicOrdering AO, bool IsVolatile,
1233 AggValueSlot resultSlot) {
1234 AtomicInfo Atomics(*this, src);
1235 return Atomics.EmitAtomicLoad(resultSlot, loc, /*AsValue=*/true, AO,
1236 IsVolatile);
1237 }
1238
1239 /// Copy an r-value into memory as part of storing to an atomic type.
1240 /// This needs to create a bit-pattern suitable for atomic operations.
emitCopyIntoMemory(RValue rvalue) const1241 void AtomicInfo::emitCopyIntoMemory(RValue rvalue) const {
1242 assert(LVal.isSimple());
1243 // If we have an r-value, the rvalue should be of the atomic type,
1244 // which means that the caller is responsible for having zeroed
1245 // any padding. Just do an aggregate copy of that type.
1246 if (rvalue.isAggregate()) {
1247 CGF.EmitAggregateCopy(getAtomicAddress(),
1248 rvalue.getAggregateAddr(),
1249 getAtomicType(),
1250 (rvalue.isVolatileQualified()
1251 || LVal.isVolatileQualified()),
1252 LVal.getAlignment());
1253 return;
1254 }
1255
1256 // Okay, otherwise we're copying stuff.
1257
1258 // Zero out the buffer if necessary.
1259 emitMemSetZeroIfNecessary();
1260
1261 // Drill past the padding if present.
1262 LValue TempLVal = projectValue();
1263
1264 // Okay, store the rvalue in.
1265 if (rvalue.isScalar()) {
1266 CGF.EmitStoreOfScalar(rvalue.getScalarVal(), TempLVal, /*init*/ true);
1267 } else {
1268 CGF.EmitStoreOfComplex(rvalue.getComplexVal(), TempLVal, /*init*/ true);
1269 }
1270 }
1271
1272
1273 /// Materialize an r-value into memory for the purposes of storing it
1274 /// to an atomic type.
materializeRValue(RValue rvalue) const1275 llvm::Value *AtomicInfo::materializeRValue(RValue rvalue) const {
1276 // Aggregate r-values are already in memory, and EmitAtomicStore
1277 // requires them to be values of the atomic type.
1278 if (rvalue.isAggregate())
1279 return rvalue.getAggregateAddr();
1280
1281 // Otherwise, make a temporary and materialize into it.
1282 LValue TempLV = CGF.MakeAddrLValue(CreateTempAlloca(), getAtomicType(),
1283 getAtomicAlignment());
1284 AtomicInfo Atomics(CGF, TempLV);
1285 Atomics.emitCopyIntoMemory(rvalue);
1286 return TempLV.getAddress();
1287 }
1288
convertRValueToInt(RValue RVal) const1289 llvm::Value *AtomicInfo::convertRValueToInt(RValue RVal) const {
1290 // If we've got a scalar value of the right size, try to avoid going
1291 // through memory.
1292 if (RVal.isScalar() && (!hasPadding() || !LVal.isSimple())) {
1293 llvm::Value *Value = RVal.getScalarVal();
1294 if (isa<llvm::IntegerType>(Value->getType()))
1295 return CGF.EmitToMemory(Value, ValueTy);
1296 else {
1297 llvm::IntegerType *InputIntTy = llvm::IntegerType::get(
1298 CGF.getLLVMContext(),
1299 LVal.isSimple() ? getValueSizeInBits() : getAtomicSizeInBits());
1300 if (isa<llvm::PointerType>(Value->getType()))
1301 return CGF.Builder.CreatePtrToInt(Value, InputIntTy);
1302 else if (llvm::BitCastInst::isBitCastable(Value->getType(), InputIntTy))
1303 return CGF.Builder.CreateBitCast(Value, InputIntTy);
1304 }
1305 }
1306 // Otherwise, we need to go through memory.
1307 // Put the r-value in memory.
1308 llvm::Value *Addr = materializeRValue(RVal);
1309
1310 // Cast the temporary to the atomic int type and pull a value out.
1311 Addr = emitCastToAtomicIntPointer(Addr);
1312 return CGF.Builder.CreateAlignedLoad(Addr,
1313 getAtomicAlignment().getQuantity());
1314 }
1315
EmitAtomicCompareExchangeOp(RValue Expected,RValue Desired,llvm::AtomicOrdering Success,llvm::AtomicOrdering Failure,bool IsWeak)1316 std::pair<RValue, llvm::Value *> AtomicInfo::EmitAtomicCompareExchangeOp(
1317 RValue Expected, RValue Desired, llvm::AtomicOrdering Success,
1318 llvm::AtomicOrdering Failure, bool IsWeak) {
1319 // Do the atomic store.
1320 auto *ExpectedVal = convertRValueToInt(Expected);
1321 auto *DesiredVal = convertRValueToInt(Desired);
1322 auto *Addr = emitCastToAtomicIntPointer(getAtomicAddress());
1323 auto *Inst = CGF.Builder.CreateAtomicCmpXchg(Addr, ExpectedVal, DesiredVal,
1324 Success, Failure);
1325 // Other decoration.
1326 Inst->setVolatile(LVal.isVolatileQualified());
1327 Inst->setWeak(IsWeak);
1328
1329 // Okay, turn that back into the original value type.
1330 auto *PreviousVal = CGF.Builder.CreateExtractValue(Inst, /*Idxs=*/0);
1331 auto *SuccessFailureVal = CGF.Builder.CreateExtractValue(Inst, /*Idxs=*/1);
1332 return std::make_pair(
1333 ConvertIntToValueOrAtomic(PreviousVal, AggValueSlot::ignored(),
1334 SourceLocation(), /*AsValue=*/false),
1335 SuccessFailureVal);
1336 }
1337
1338 std::pair<RValue, llvm::Value *>
EmitAtomicCompareExchangeLibcall(RValue Expected,RValue Desired,llvm::AtomicOrdering Success,llvm::AtomicOrdering Failure)1339 AtomicInfo::EmitAtomicCompareExchangeLibcall(RValue Expected, RValue Desired,
1340 llvm::AtomicOrdering Success,
1341 llvm::AtomicOrdering Failure) {
1342 // bool __atomic_compare_exchange(size_t size, void *obj, void *expected,
1343 // void *desired, int success, int failure);
1344 auto *ExpectedAddr = materializeRValue(Expected);
1345 auto *DesiredAddr = materializeRValue(Desired);
1346 CallArgList Args;
1347 Args.add(RValue::get(getAtomicSizeValue()), CGF.getContext().getSizeType());
1348 Args.add(RValue::get(CGF.EmitCastToVoidPtr(getAtomicAddress())),
1349 CGF.getContext().VoidPtrTy);
1350 Args.add(RValue::get(CGF.EmitCastToVoidPtr(ExpectedAddr)),
1351 CGF.getContext().VoidPtrTy);
1352 Args.add(RValue::get(CGF.EmitCastToVoidPtr(DesiredAddr)),
1353 CGF.getContext().VoidPtrTy);
1354 Args.add(RValue::get(llvm::ConstantInt::get(
1355 CGF.IntTy, translateAtomicOrdering(Success))),
1356 CGF.getContext().IntTy);
1357 Args.add(RValue::get(llvm::ConstantInt::get(
1358 CGF.IntTy, translateAtomicOrdering(Failure))),
1359 CGF.getContext().IntTy);
1360 auto SuccessFailureRVal = emitAtomicLibcall(CGF, "__atomic_compare_exchange",
1361 CGF.getContext().BoolTy, Args);
1362
1363 return std::make_pair(
1364 convertTempToRValue(ExpectedAddr, AggValueSlot::ignored(),
1365 SourceLocation(), /*AsValue=*/false),
1366 SuccessFailureRVal.getScalarVal());
1367 }
1368
EmitAtomicCompareExchange(RValue Expected,RValue Desired,llvm::AtomicOrdering Success,llvm::AtomicOrdering Failure,bool IsWeak)1369 std::pair<RValue, llvm::Value *> AtomicInfo::EmitAtomicCompareExchange(
1370 RValue Expected, RValue Desired, llvm::AtomicOrdering Success,
1371 llvm::AtomicOrdering Failure, bool IsWeak) {
1372 if (Failure >= Success)
1373 // Don't assert on undefined behavior.
1374 Failure = llvm::AtomicCmpXchgInst::getStrongestFailureOrdering(Success);
1375
1376 // Check whether we should use a library call.
1377 if (shouldUseLibcall()) {
1378 // Produce a source address.
1379 return EmitAtomicCompareExchangeLibcall(Expected, Desired, Success,
1380 Failure);
1381 }
1382
1383 // If we've got a scalar value of the right size, try to avoid going
1384 // through memory.
1385 return EmitAtomicCompareExchangeOp(Expected, Desired, Success, Failure,
1386 IsWeak);
1387 }
1388
EmitAtomicStore(RValue rvalue,LValue lvalue,bool isInit)1389 void CodeGenFunction::EmitAtomicStore(RValue rvalue, LValue lvalue,
1390 bool isInit) {
1391 bool IsVolatile = lvalue.isVolatileQualified();
1392 llvm::AtomicOrdering AO;
1393 if (lvalue.getType()->isAtomicType()) {
1394 AO = llvm::SequentiallyConsistent;
1395 } else {
1396 AO = llvm::Release;
1397 IsVolatile = true;
1398 }
1399 return EmitAtomicStore(rvalue, lvalue, AO, IsVolatile, isInit);
1400 }
1401
1402 /// Emit a store to an l-value of atomic type.
1403 ///
1404 /// Note that the r-value is expected to be an r-value *of the atomic
1405 /// type*; this means that for aggregate r-values, it should include
1406 /// storage for any padding that was necessary.
EmitAtomicStore(RValue rvalue,LValue dest,llvm::AtomicOrdering AO,bool IsVolatile,bool isInit)1407 void CodeGenFunction::EmitAtomicStore(RValue rvalue, LValue dest,
1408 llvm::AtomicOrdering AO, bool IsVolatile,
1409 bool isInit) {
1410 // If this is an aggregate r-value, it should agree in type except
1411 // maybe for address-space qualification.
1412 assert(!rvalue.isAggregate() ||
1413 rvalue.getAggregateAddr()->getType()->getPointerElementType()
1414 == dest.getAddress()->getType()->getPointerElementType());
1415
1416 AtomicInfo atomics(*this, dest);
1417 LValue LVal = atomics.getAtomicLValue();
1418
1419 // If this is an initialization, just put the value there normally.
1420 if (LVal.isSimple()) {
1421 if (isInit) {
1422 atomics.emitCopyIntoMemory(rvalue);
1423 return;
1424 }
1425
1426 // Check whether we should use a library call.
1427 if (atomics.shouldUseLibcall()) {
1428 // Produce a source address.
1429 llvm::Value *srcAddr = atomics.materializeRValue(rvalue);
1430
1431 // void __atomic_store(size_t size, void *mem, void *val, int order)
1432 CallArgList args;
1433 args.add(RValue::get(atomics.getAtomicSizeValue()),
1434 getContext().getSizeType());
1435 args.add(RValue::get(EmitCastToVoidPtr(atomics.getAtomicAddress())),
1436 getContext().VoidPtrTy);
1437 args.add(RValue::get(EmitCastToVoidPtr(srcAddr)), getContext().VoidPtrTy);
1438 args.add(RValue::get(llvm::ConstantInt::get(
1439 IntTy, AtomicInfo::translateAtomicOrdering(AO))),
1440 getContext().IntTy);
1441 emitAtomicLibcall(*this, "__atomic_store", getContext().VoidTy, args);
1442 return;
1443 }
1444
1445 // Okay, we're doing this natively.
1446 llvm::Value *intValue = atomics.convertRValueToInt(rvalue);
1447
1448 // Do the atomic store.
1449 llvm::Value *addr =
1450 atomics.emitCastToAtomicIntPointer(atomics.getAtomicAddress());
1451 intValue = Builder.CreateIntCast(
1452 intValue, addr->getType()->getPointerElementType(), /*isSigned=*/false);
1453 llvm::StoreInst *store = Builder.CreateStore(intValue, addr);
1454
1455 // Initializations don't need to be atomic.
1456 if (!isInit)
1457 store->setAtomic(AO);
1458
1459 // Other decoration.
1460 store->setAlignment(dest.getAlignment().getQuantity());
1461 if (IsVolatile)
1462 store->setVolatile(true);
1463 if (dest.getTBAAInfo())
1464 CGM.DecorateInstruction(store, dest.getTBAAInfo());
1465 return;
1466 }
1467
1468 // Atomic load of prev value.
1469 RValue OldRVal =
1470 atomics.EmitAtomicLoad(AggValueSlot::ignored(), SourceLocation(),
1471 /*AsValue=*/false, AO, IsVolatile);
1472 // For non-simple lvalues perform compare-and-swap procedure.
1473 auto *ContBB = createBasicBlock("atomic_cont");
1474 auto *ExitBB = createBasicBlock("atomic_exit");
1475 auto *CurBB = Builder.GetInsertBlock();
1476 EmitBlock(ContBB);
1477 llvm::PHINode *PHI = Builder.CreatePHI(OldRVal.getScalarVal()->getType(),
1478 /*NumReservedValues=*/2);
1479 PHI->addIncoming(OldRVal.getScalarVal(), CurBB);
1480 RValue OriginalRValue = RValue::get(PHI);
1481 // Build new lvalue for temp address
1482 auto *Ptr = atomics.materializeRValue(OriginalRValue);
1483 // Build new lvalue for temp address
1484 LValue UpdateLVal;
1485 if (LVal.isBitField())
1486 UpdateLVal = LValue::MakeBitfield(Ptr, LVal.getBitFieldInfo(),
1487 LVal.getType(), LVal.getAlignment());
1488 else if (LVal.isVectorElt())
1489 UpdateLVal = LValue::MakeVectorElt(Ptr, LVal.getVectorIdx(), LVal.getType(),
1490 LVal.getAlignment());
1491 else {
1492 assert(LVal.isExtVectorElt());
1493 UpdateLVal = LValue::MakeExtVectorElt(Ptr, LVal.getExtVectorElts(),
1494 LVal.getType(), LVal.getAlignment());
1495 }
1496 UpdateLVal.setTBAAInfo(LVal.getTBAAInfo());
1497 // Store new value in the corresponding memory area
1498 EmitStoreThroughLValue(rvalue, UpdateLVal);
1499 // Load new value
1500 RValue NewRValue = RValue::get(EmitLoadOfScalar(
1501 Ptr, LVal.isVolatile(), atomics.getAtomicAlignment().getQuantity(),
1502 atomics.getAtomicType(), SourceLocation()));
1503 // Try to write new value using cmpxchg operation
1504 auto Pair = atomics.EmitAtomicCompareExchange(OriginalRValue, NewRValue, AO);
1505 PHI->addIncoming(Pair.first.getScalarVal(), ContBB);
1506 Builder.CreateCondBr(Pair.second, ExitBB, ContBB);
1507 EmitBlock(ExitBB, /*IsFinished=*/true);
1508 }
1509
1510 /// Emit a compare-and-exchange op for atomic type.
1511 ///
EmitAtomicCompareExchange(LValue Obj,RValue Expected,RValue Desired,SourceLocation Loc,llvm::AtomicOrdering Success,llvm::AtomicOrdering Failure,bool IsWeak,AggValueSlot Slot)1512 std::pair<RValue, llvm::Value *> CodeGenFunction::EmitAtomicCompareExchange(
1513 LValue Obj, RValue Expected, RValue Desired, SourceLocation Loc,
1514 llvm::AtomicOrdering Success, llvm::AtomicOrdering Failure, bool IsWeak,
1515 AggValueSlot Slot) {
1516 // If this is an aggregate r-value, it should agree in type except
1517 // maybe for address-space qualification.
1518 assert(!Expected.isAggregate() ||
1519 Expected.getAggregateAddr()->getType()->getPointerElementType() ==
1520 Obj.getAddress()->getType()->getPointerElementType());
1521 assert(!Desired.isAggregate() ||
1522 Desired.getAggregateAddr()->getType()->getPointerElementType() ==
1523 Obj.getAddress()->getType()->getPointerElementType());
1524 AtomicInfo Atomics(*this, Obj);
1525
1526 return Atomics.EmitAtomicCompareExchange(Expected, Desired, Success, Failure,
1527 IsWeak);
1528 }
1529
EmitAtomicUpdate(LValue LVal,llvm::AtomicOrdering AO,const std::function<RValue (RValue)> & UpdateOp,bool IsVolatile)1530 void CodeGenFunction::EmitAtomicUpdate(
1531 LValue LVal, llvm::AtomicOrdering AO,
1532 const std::function<RValue(RValue)> &UpdateOp, bool IsVolatile) {
1533 AtomicInfo Atomics(*this, LVal);
1534 LValue AtomicLVal = Atomics.getAtomicLValue();
1535
1536 // Atomic load of prev value.
1537 RValue OldRVal =
1538 Atomics.EmitAtomicLoad(AggValueSlot::ignored(), SourceLocation(),
1539 /*AsValue=*/false, AO, IsVolatile);
1540 bool IsScalar = OldRVal.isScalar();
1541 auto *OldVal =
1542 IsScalar ? OldRVal.getScalarVal() : Atomics.convertRValueToInt(OldRVal);
1543 // For non-simple lvalues perform compare-and-swap procedure.
1544 auto *ContBB = createBasicBlock("atomic_cont");
1545 auto *ExitBB = createBasicBlock("atomic_exit");
1546 auto *CurBB = Builder.GetInsertBlock();
1547 EmitBlock(ContBB);
1548 llvm::PHINode *PHI = Builder.CreatePHI(OldVal->getType(),
1549 /*NumReservedValues=*/2);
1550 PHI->addIncoming(OldVal, CurBB);
1551 RValue OriginalRValue =
1552 IsScalar ? RValue::get(PHI) : Atomics.ConvertIntToValueOrAtomic(
1553 PHI, AggValueSlot::ignored(),
1554 SourceLocation(), /*AsValue=*/false);
1555 // Build new lvalue for temp address
1556 LValue UpdateLVal;
1557 llvm::Value *Ptr = nullptr;
1558 RValue UpRVal;
1559 if (AtomicLVal.isSimple()) {
1560 UpRVal = OriginalRValue;
1561 } else {
1562 // Build new lvalue for temp address
1563 Ptr = Atomics.materializeRValue(OriginalRValue);
1564 if (AtomicLVal.isBitField())
1565 UpdateLVal =
1566 LValue::MakeBitfield(Ptr, AtomicLVal.getBitFieldInfo(),
1567 AtomicLVal.getType(), AtomicLVal.getAlignment());
1568 else if (AtomicLVal.isVectorElt())
1569 UpdateLVal = LValue::MakeVectorElt(Ptr, AtomicLVal.getVectorIdx(),
1570 AtomicLVal.getType(),
1571 AtomicLVal.getAlignment());
1572 else {
1573 assert(AtomicLVal.isExtVectorElt());
1574 UpdateLVal = LValue::MakeExtVectorElt(Ptr, AtomicLVal.getExtVectorElts(),
1575 AtomicLVal.getType(),
1576 AtomicLVal.getAlignment());
1577 }
1578 UpdateLVal.setTBAAInfo(LVal.getTBAAInfo());
1579 UpRVal = EmitLoadOfLValue(UpdateLVal, SourceLocation());
1580 }
1581 // Store new value in the corresponding memory area
1582 RValue NewRVal = UpdateOp(UpRVal);
1583 if (!AtomicLVal.isSimple()) {
1584 EmitStoreThroughLValue(NewRVal, UpdateLVal);
1585 // Load new value
1586 NewRVal = RValue::get(
1587 EmitLoadOfScalar(Ptr, AtomicLVal.isVolatile(),
1588 Atomics.getAtomicAlignment().getQuantity(),
1589 Atomics.getAtomicType(), SourceLocation()));
1590 }
1591 // Try to write new value using cmpxchg operation
1592 auto Pair = Atomics.EmitAtomicCompareExchange(OriginalRValue, NewRVal, AO);
1593 OldVal = IsScalar ? Pair.first.getScalarVal()
1594 : Atomics.convertRValueToInt(Pair.first);
1595 PHI->addIncoming(OldVal, ContBB);
1596 Builder.CreateCondBr(Pair.second, ExitBB, ContBB);
1597 EmitBlock(ExitBB, /*IsFinished=*/true);
1598 }
1599
EmitAtomicInit(Expr * init,LValue dest)1600 void CodeGenFunction::EmitAtomicInit(Expr *init, LValue dest) {
1601 AtomicInfo atomics(*this, dest);
1602
1603 switch (atomics.getEvaluationKind()) {
1604 case TEK_Scalar: {
1605 llvm::Value *value = EmitScalarExpr(init);
1606 atomics.emitCopyIntoMemory(RValue::get(value));
1607 return;
1608 }
1609
1610 case TEK_Complex: {
1611 ComplexPairTy value = EmitComplexExpr(init);
1612 atomics.emitCopyIntoMemory(RValue::getComplex(value));
1613 return;
1614 }
1615
1616 case TEK_Aggregate: {
1617 // Fix up the destination if the initializer isn't an expression
1618 // of atomic type.
1619 bool Zeroed = false;
1620 if (!init->getType()->isAtomicType()) {
1621 Zeroed = atomics.emitMemSetZeroIfNecessary();
1622 dest = atomics.projectValue();
1623 }
1624
1625 // Evaluate the expression directly into the destination.
1626 AggValueSlot slot = AggValueSlot::forLValue(dest,
1627 AggValueSlot::IsNotDestructed,
1628 AggValueSlot::DoesNotNeedGCBarriers,
1629 AggValueSlot::IsNotAliased,
1630 Zeroed ? AggValueSlot::IsZeroed :
1631 AggValueSlot::IsNotZeroed);
1632
1633 EmitAggExpr(init, slot);
1634 return;
1635 }
1636 }
1637 llvm_unreachable("bad evaluation kind");
1638 }
1639