1 //===--- CGStmtOpenMP.cpp - Emit LLVM Code from Statements ----------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This contains code to emit OpenMP nodes as LLVM code.
12 //===----------------------------------------------------------------------===//
14 #include "CGCleanup.h"
15 #include "CGOpenMPRuntime.h"
16 #include "CodeGenFunction.h"
17 #include "CodeGenModule.h"
18 #include "TargetInfo.h"
19 #include "clang/AST/Stmt.h"
20 #include "clang/AST/StmtOpenMP.h"
21 #include "clang/AST/DeclOpenMP.h"
22 #include "llvm/IR/CallSite.h"
23 using namespace clang;
24 using namespace CodeGen;
27 /// Lexical scope for OpenMP executable constructs, that handles correct codegen
28 /// for captured expressions.
29 class OMPLexicalScope : public CodeGenFunction::LexicalScope {
30 void emitPreInitStmt(CodeGenFunction &CGF, const OMPExecutableDirective &S) {
31 for (const auto *C : S.clauses()) {
32 if (const auto *CPI = OMPClauseWithPreInit::get(C)) {
33 if (const auto *PreInit =
34 cast_or_null<DeclStmt>(CPI->getPreInitStmt())) {
35 for (const auto *I : PreInit->decls()) {
36 if (!I->hasAttr<OMPCaptureNoInitAttr>()) {
37 CGF.EmitVarDecl(cast<VarDecl>(*I));
39 CodeGenFunction::AutoVarEmission Emission =
40 CGF.EmitAutoVarAlloca(cast<VarDecl>(*I));
41 CGF.EmitAutoVarCleanups(Emission);
48 CodeGenFunction::OMPPrivateScope InlinedShareds;
50 static bool isCapturedVar(CodeGenFunction &CGF, const VarDecl *VD) {
51 return CGF.LambdaCaptureFields.lookup(VD) ||
52 (CGF.CapturedStmtInfo && CGF.CapturedStmtInfo->lookup(VD)) ||
53 (CGF.CurCodeDecl && isa<BlockDecl>(CGF.CurCodeDecl));
58 CodeGenFunction &CGF, const OMPExecutableDirective &S,
59 const llvm::Optional<OpenMPDirectiveKind> CapturedRegion = llvm::None,
60 const bool EmitPreInitStmt = true)
61 : CodeGenFunction::LexicalScope(CGF, S.getSourceRange()),
64 emitPreInitStmt(CGF, S);
65 if (!CapturedRegion.hasValue())
67 assert(S.hasAssociatedStmt() &&
68 "Expected associated statement for inlined directive.");
69 const CapturedStmt *CS = S.getCapturedStmt(*CapturedRegion);
70 for (const auto &C : CS->captures()) {
71 if (C.capturesVariable() || C.capturesVariableByCopy()) {
72 auto *VD = C.getCapturedVar();
73 assert(VD == VD->getCanonicalDecl() &&
74 "Canonical decl must be captured.");
76 CGF.getContext(), const_cast<VarDecl *>(VD),
77 isCapturedVar(CGF, VD) || (CGF.CapturedStmtInfo &&
78 InlinedShareds.isGlobalVarCaptured(VD)),
79 VD->getType().getNonReferenceType(), VK_LValue, C.getLocation());
80 InlinedShareds.addPrivate(VD, [&CGF, &DRE]() -> Address {
81 return CGF.EmitLValue(&DRE).getAddress();
85 (void)InlinedShareds.Privatize();
89 /// Lexical scope for OpenMP parallel construct, that handles correct codegen
90 /// for captured expressions.
91 class OMPParallelScope final : public OMPLexicalScope {
92 bool EmitPreInitStmt(const OMPExecutableDirective &S) {
93 OpenMPDirectiveKind Kind = S.getDirectiveKind();
94 return !(isOpenMPTargetExecutionDirective(Kind) ||
95 isOpenMPLoopBoundSharingDirective(Kind)) &&
96 isOpenMPParallelDirective(Kind);
100 OMPParallelScope(CodeGenFunction &CGF, const OMPExecutableDirective &S)
101 : OMPLexicalScope(CGF, S, /*CapturedRegion=*/llvm::None,
102 EmitPreInitStmt(S)) {}
105 /// Lexical scope for OpenMP teams construct, that handles correct codegen
106 /// for captured expressions.
107 class OMPTeamsScope final : public OMPLexicalScope {
108 bool EmitPreInitStmt(const OMPExecutableDirective &S) {
109 OpenMPDirectiveKind Kind = S.getDirectiveKind();
110 return !isOpenMPTargetExecutionDirective(Kind) &&
111 isOpenMPTeamsDirective(Kind);
115 OMPTeamsScope(CodeGenFunction &CGF, const OMPExecutableDirective &S)
116 : OMPLexicalScope(CGF, S, /*CapturedRegion=*/llvm::None,
117 EmitPreInitStmt(S)) {}
120 /// Private scope for OpenMP loop-based directives, that supports capturing
121 /// of used expression from loop statement.
122 class OMPLoopScope : public CodeGenFunction::RunCleanupsScope {
123 void emitPreInitStmt(CodeGenFunction &CGF, const OMPLoopDirective &S) {
124 CodeGenFunction::OMPMapVars PreCondVars;
125 for (const auto *E : S.counters()) {
126 const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl());
127 (void)PreCondVars.setVarAddr(
128 CGF, VD, CGF.CreateMemTemp(VD->getType().getNonReferenceType()));
130 (void)PreCondVars.apply(CGF);
131 if (const auto *PreInits = cast_or_null<DeclStmt>(S.getPreInits())) {
132 for (const auto *I : PreInits->decls())
133 CGF.EmitVarDecl(cast<VarDecl>(*I));
135 PreCondVars.restore(CGF);
139 OMPLoopScope(CodeGenFunction &CGF, const OMPLoopDirective &S)
140 : CodeGenFunction::RunCleanupsScope(CGF) {
141 emitPreInitStmt(CGF, S);
145 class OMPSimdLexicalScope : public CodeGenFunction::LexicalScope {
146 CodeGenFunction::OMPPrivateScope InlinedShareds;
148 static bool isCapturedVar(CodeGenFunction &CGF, const VarDecl *VD) {
149 return CGF.LambdaCaptureFields.lookup(VD) ||
150 (CGF.CapturedStmtInfo && CGF.CapturedStmtInfo->lookup(VD)) ||
151 (CGF.CurCodeDecl && isa<BlockDecl>(CGF.CurCodeDecl) &&
152 cast<BlockDecl>(CGF.CurCodeDecl)->capturesVariable(VD));
156 OMPSimdLexicalScope(CodeGenFunction &CGF, const OMPExecutableDirective &S)
157 : CodeGenFunction::LexicalScope(CGF, S.getSourceRange()),
158 InlinedShareds(CGF) {
159 for (const auto *C : S.clauses()) {
160 if (const auto *CPI = OMPClauseWithPreInit::get(C)) {
161 if (const auto *PreInit =
162 cast_or_null<DeclStmt>(CPI->getPreInitStmt())) {
163 for (const auto *I : PreInit->decls()) {
164 if (!I->hasAttr<OMPCaptureNoInitAttr>()) {
165 CGF.EmitVarDecl(cast<VarDecl>(*I));
167 CodeGenFunction::AutoVarEmission Emission =
168 CGF.EmitAutoVarAlloca(cast<VarDecl>(*I));
169 CGF.EmitAutoVarCleanups(Emission);
173 } else if (const auto *UDP = dyn_cast<OMPUseDevicePtrClause>(C)) {
174 for (const Expr *E : UDP->varlists()) {
175 const Decl *D = cast<DeclRefExpr>(E)->getDecl();
176 if (const auto *OED = dyn_cast<OMPCapturedExprDecl>(D))
177 CGF.EmitVarDecl(*OED);
181 if (!isOpenMPSimdDirective(S.getDirectiveKind()))
182 CGF.EmitOMPPrivateClause(S, InlinedShareds);
183 if (const auto *TG = dyn_cast<OMPTaskgroupDirective>(&S)) {
184 if (const Expr *E = TG->getReductionRef())
185 CGF.EmitVarDecl(*cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl()));
187 const auto *CS = cast_or_null<CapturedStmt>(S.getAssociatedStmt());
189 for (auto &C : CS->captures()) {
190 if (C.capturesVariable() || C.capturesVariableByCopy()) {
191 auto *VD = C.getCapturedVar();
192 assert(VD == VD->getCanonicalDecl() &&
193 "Canonical decl must be captured.");
194 DeclRefExpr DRE(CGF.getContext(), const_cast<VarDecl *>(VD),
195 isCapturedVar(CGF, VD) ||
196 (CGF.CapturedStmtInfo &&
197 InlinedShareds.isGlobalVarCaptured(VD)),
198 VD->getType().getNonReferenceType(), VK_LValue,
200 InlinedShareds.addPrivate(VD, [&CGF, &DRE]() -> Address {
201 return CGF.EmitLValue(&DRE).getAddress();
205 CS = dyn_cast<CapturedStmt>(CS->getCapturedStmt());
207 (void)InlinedShareds.Privatize();
213 static void emitCommonOMPTargetDirective(CodeGenFunction &CGF,
214 const OMPExecutableDirective &S,
215 const RegionCodeGenTy &CodeGen);
217 LValue CodeGenFunction::EmitOMPSharedLValue(const Expr *E) {
218 if (const auto *OrigDRE = dyn_cast<DeclRefExpr>(E)) {
219 if (const auto *OrigVD = dyn_cast<VarDecl>(OrigDRE->getDecl())) {
220 OrigVD = OrigVD->getCanonicalDecl();
222 LambdaCaptureFields.lookup(OrigVD) ||
223 (CapturedStmtInfo && CapturedStmtInfo->lookup(OrigVD)) ||
224 (CurCodeDecl && isa<BlockDecl>(CurCodeDecl));
225 DeclRefExpr DRE(getContext(), const_cast<VarDecl *>(OrigVD), IsCaptured,
226 OrigDRE->getType(), VK_LValue, OrigDRE->getExprLoc());
227 return EmitLValue(&DRE);
230 return EmitLValue(E);
233 llvm::Value *CodeGenFunction::getTypeSize(QualType Ty) {
234 ASTContext &C = getContext();
235 llvm::Value *Size = nullptr;
236 auto SizeInChars = C.getTypeSizeInChars(Ty);
237 if (SizeInChars.isZero()) {
238 // getTypeSizeInChars() returns 0 for a VLA.
239 while (const VariableArrayType *VAT = C.getAsVariableArrayType(Ty)) {
240 VlaSizePair VlaSize = getVLASize(VAT);
242 Size = Size ? Builder.CreateNUWMul(Size, VlaSize.NumElts)
245 SizeInChars = C.getTypeSizeInChars(Ty);
246 if (SizeInChars.isZero())
247 return llvm::ConstantInt::get(SizeTy, /*V=*/0);
248 return Builder.CreateNUWMul(Size, CGM.getSize(SizeInChars));
250 return CGM.getSize(SizeInChars);
253 void CodeGenFunction::GenerateOpenMPCapturedVars(
254 const CapturedStmt &S, SmallVectorImpl<llvm::Value *> &CapturedVars) {
255 const RecordDecl *RD = S.getCapturedRecordDecl();
256 auto CurField = RD->field_begin();
257 auto CurCap = S.captures().begin();
258 for (CapturedStmt::const_capture_init_iterator I = S.capture_init_begin(),
259 E = S.capture_init_end();
260 I != E; ++I, ++CurField, ++CurCap) {
261 if (CurField->hasCapturedVLAType()) {
262 const VariableArrayType *VAT = CurField->getCapturedVLAType();
263 llvm::Value *Val = VLASizeMap[VAT->getSizeExpr()];
264 CapturedVars.push_back(Val);
265 } else if (CurCap->capturesThis()) {
266 CapturedVars.push_back(CXXThisValue);
267 } else if (CurCap->capturesVariableByCopy()) {
268 llvm::Value *CV = EmitLoadOfScalar(EmitLValue(*I), CurCap->getLocation());
270 // If the field is not a pointer, we need to save the actual value
271 // and load it as a void pointer.
272 if (!CurField->getType()->isAnyPointerType()) {
273 ASTContext &Ctx = getContext();
274 Address DstAddr = CreateMemTemp(
275 Ctx.getUIntPtrType(),
276 Twine(CurCap->getCapturedVar()->getName(), ".casted"));
277 LValue DstLV = MakeAddrLValue(DstAddr, Ctx.getUIntPtrType());
279 llvm::Value *SrcAddrVal = EmitScalarConversion(
280 DstAddr.getPointer(), Ctx.getPointerType(Ctx.getUIntPtrType()),
281 Ctx.getPointerType(CurField->getType()), CurCap->getLocation());
283 MakeNaturalAlignAddrLValue(SrcAddrVal, CurField->getType());
285 // Store the value using the source type pointer.
286 EmitStoreThroughLValue(RValue::get(CV), SrcLV);
288 // Load the value using the destination type pointer.
289 CV = EmitLoadOfScalar(DstLV, CurCap->getLocation());
291 CapturedVars.push_back(CV);
293 assert(CurCap->capturesVariable() && "Expected capture by reference.");
294 CapturedVars.push_back(EmitLValue(*I).getAddress().getPointer());
299 static Address castValueFromUintptr(CodeGenFunction &CGF, SourceLocation Loc,
300 QualType DstType, StringRef Name,
302 bool isReferenceType = false) {
303 ASTContext &Ctx = CGF.getContext();
305 llvm::Value *CastedPtr = CGF.EmitScalarConversion(
306 AddrLV.getAddress().getPointer(), Ctx.getUIntPtrType(),
307 Ctx.getPointerType(DstType), Loc);
309 CGF.MakeNaturalAlignAddrLValue(CastedPtr, Ctx.getPointerType(DstType))
312 // If we are dealing with references we need to return the address of the
313 // reference instead of the reference of the value.
314 if (isReferenceType) {
315 QualType RefType = Ctx.getLValueReferenceType(DstType);
316 llvm::Value *RefVal = TmpAddr.getPointer();
317 TmpAddr = CGF.CreateMemTemp(RefType, Twine(Name, ".ref"));
318 LValue TmpLVal = CGF.MakeAddrLValue(TmpAddr, RefType);
319 CGF.EmitStoreThroughLValue(RValue::get(RefVal), TmpLVal, /*isInit=*/true);
325 static QualType getCanonicalParamType(ASTContext &C, QualType T) {
326 if (T->isLValueReferenceType())
327 return C.getLValueReferenceType(
328 getCanonicalParamType(C, T.getNonReferenceType()),
329 /*SpelledAsLValue=*/false);
330 if (T->isPointerType())
331 return C.getPointerType(getCanonicalParamType(C, T->getPointeeType()));
332 if (const ArrayType *A = T->getAsArrayTypeUnsafe()) {
333 if (const auto *VLA = dyn_cast<VariableArrayType>(A))
334 return getCanonicalParamType(C, VLA->getElementType());
335 if (!A->isVariablyModifiedType())
336 return C.getCanonicalType(T);
338 return C.getCanonicalParamType(T);
342 /// Contains required data for proper outlined function codegen.
343 struct FunctionOptions {
344 /// Captured statement for which the function is generated.
345 const CapturedStmt *S = nullptr;
346 /// true if cast to/from UIntPtr is required for variables captured by
348 const bool UIntPtrCastRequired = true;
349 /// true if only casted arguments must be registered as local args or VLA
351 const bool RegisterCastedArgsOnly = false;
352 /// Name of the generated function.
353 const StringRef FunctionName;
354 explicit FunctionOptions(const CapturedStmt *S, bool UIntPtrCastRequired,
355 bool RegisterCastedArgsOnly,
356 StringRef FunctionName)
357 : S(S), UIntPtrCastRequired(UIntPtrCastRequired),
358 RegisterCastedArgsOnly(UIntPtrCastRequired && RegisterCastedArgsOnly),
359 FunctionName(FunctionName) {}
363 static llvm::Function *emitOutlinedFunctionPrologue(
364 CodeGenFunction &CGF, FunctionArgList &Args,
365 llvm::MapVector<const Decl *, std::pair<const VarDecl *, Address>>
367 llvm::DenseMap<const Decl *, std::pair<const Expr *, llvm::Value *>>
369 llvm::Value *&CXXThisValue, const FunctionOptions &FO) {
370 const CapturedDecl *CD = FO.S->getCapturedDecl();
371 const RecordDecl *RD = FO.S->getCapturedRecordDecl();
372 assert(CD->hasBody() && "missing CapturedDecl body");
374 CXXThisValue = nullptr;
375 // Build the argument list.
376 CodeGenModule &CGM = CGF.CGM;
377 ASTContext &Ctx = CGM.getContext();
378 FunctionArgList TargetArgs;
379 Args.append(CD->param_begin(),
380 std::next(CD->param_begin(), CD->getContextParamPosition()));
383 std::next(CD->param_begin(), CD->getContextParamPosition()));
384 auto I = FO.S->captures().begin();
385 FunctionDecl *DebugFunctionDecl = nullptr;
386 if (!FO.UIntPtrCastRequired) {
387 FunctionProtoType::ExtProtoInfo EPI;
388 QualType FunctionTy = Ctx.getFunctionType(Ctx.VoidTy, llvm::None, EPI);
389 DebugFunctionDecl = FunctionDecl::Create(
390 Ctx, Ctx.getTranslationUnitDecl(), FO.S->getBeginLoc(),
391 SourceLocation(), DeclarationName(), FunctionTy,
392 Ctx.getTrivialTypeSourceInfo(FunctionTy), SC_Static,
393 /*isInlineSpecified=*/false, /*hasWrittenPrototype=*/false);
395 for (const FieldDecl *FD : RD->fields()) {
396 QualType ArgType = FD->getType();
397 IdentifierInfo *II = nullptr;
398 VarDecl *CapVar = nullptr;
400 // If this is a capture by copy and the type is not a pointer, the outlined
401 // function argument type should be uintptr and the value properly casted to
402 // uintptr. This is necessary given that the runtime library is only able to
403 // deal with pointers. We can pass in the same way the VLA type sizes to the
404 // outlined function.
405 if (FO.UIntPtrCastRequired &&
406 ((I->capturesVariableByCopy() && !ArgType->isAnyPointerType()) ||
407 I->capturesVariableArrayType()))
408 ArgType = Ctx.getUIntPtrType();
410 if (I->capturesVariable() || I->capturesVariableByCopy()) {
411 CapVar = I->getCapturedVar();
412 II = CapVar->getIdentifier();
413 } else if (I->capturesThis()) {
414 II = &Ctx.Idents.get("this");
416 assert(I->capturesVariableArrayType());
417 II = &Ctx.Idents.get("vla");
419 if (ArgType->isVariablyModifiedType())
420 ArgType = getCanonicalParamType(Ctx, ArgType);
422 if (DebugFunctionDecl && (CapVar || I->capturesThis())) {
423 Arg = ParmVarDecl::Create(
424 Ctx, DebugFunctionDecl,
425 CapVar ? CapVar->getBeginLoc() : FD->getBeginLoc(),
426 CapVar ? CapVar->getLocation() : FD->getLocation(), II, ArgType,
427 /*TInfo=*/nullptr, SC_None, /*DefArg=*/nullptr);
429 Arg = ImplicitParamDecl::Create(Ctx, /*DC=*/nullptr, FD->getLocation(),
430 II, ArgType, ImplicitParamDecl::Other);
432 Args.emplace_back(Arg);
433 // Do not cast arguments if we emit function with non-original types.
434 TargetArgs.emplace_back(
435 FO.UIntPtrCastRequired
437 : CGM.getOpenMPRuntime().translateParameter(FD, Arg));
441 std::next(CD->param_begin(), CD->getContextParamPosition() + 1),
444 std::next(CD->param_begin(), CD->getContextParamPosition() + 1),
447 // Create the function declaration.
448 const CGFunctionInfo &FuncInfo =
449 CGM.getTypes().arrangeBuiltinFunctionDeclaration(Ctx.VoidTy, TargetArgs);
450 llvm::FunctionType *FuncLLVMTy = CGM.getTypes().GetFunctionType(FuncInfo);
453 llvm::Function::Create(FuncLLVMTy, llvm::GlobalValue::InternalLinkage,
454 FO.FunctionName, &CGM.getModule());
455 CGM.SetInternalFunctionAttributes(CD, F, FuncInfo);
457 F->setDoesNotThrow();
458 F->setDoesNotRecurse();
460 // Generate the function.
461 CGF.StartFunction(CD, Ctx.VoidTy, F, FuncInfo, TargetArgs,
462 FO.S->getBeginLoc(), CD->getBody()->getBeginLoc());
463 unsigned Cnt = CD->getContextParamPosition();
464 I = FO.S->captures().begin();
465 for (const FieldDecl *FD : RD->fields()) {
466 // Do not map arguments if we emit function with non-original types.
467 Address LocalAddr(Address::invalid());
468 if (!FO.UIntPtrCastRequired && Args[Cnt] != TargetArgs[Cnt]) {
469 LocalAddr = CGM.getOpenMPRuntime().getParameterAddress(CGF, Args[Cnt],
472 LocalAddr = CGF.GetAddrOfLocalVar(Args[Cnt]);
474 // If we are capturing a pointer by copy we don't need to do anything, just
475 // use the value that we get from the arguments.
476 if (I->capturesVariableByCopy() && FD->getType()->isAnyPointerType()) {
477 const VarDecl *CurVD = I->getCapturedVar();
478 // If the variable is a reference we need to materialize it here.
479 if (CurVD->getType()->isReferenceType()) {
480 Address RefAddr = CGF.CreateMemTemp(
481 CurVD->getType(), CGM.getPointerAlign(), ".materialized_ref");
482 CGF.EmitStoreOfScalar(LocalAddr.getPointer(), RefAddr,
483 /*Volatile=*/false, CurVD->getType());
486 if (!FO.RegisterCastedArgsOnly)
487 LocalAddrs.insert({Args[Cnt], {CurVD, LocalAddr}});
493 LValue ArgLVal = CGF.MakeAddrLValue(LocalAddr, Args[Cnt]->getType(),
494 AlignmentSource::Decl);
495 if (FD->hasCapturedVLAType()) {
496 if (FO.UIntPtrCastRequired) {
497 ArgLVal = CGF.MakeAddrLValue(
498 castValueFromUintptr(CGF, I->getLocation(), FD->getType(),
499 Args[Cnt]->getName(), ArgLVal),
500 FD->getType(), AlignmentSource::Decl);
502 llvm::Value *ExprArg = CGF.EmitLoadOfScalar(ArgLVal, I->getLocation());
503 const VariableArrayType *VAT = FD->getCapturedVLAType();
504 VLASizes.try_emplace(Args[Cnt], VAT->getSizeExpr(), ExprArg);
505 } else if (I->capturesVariable()) {
506 const VarDecl *Var = I->getCapturedVar();
507 QualType VarTy = Var->getType();
508 Address ArgAddr = ArgLVal.getAddress();
509 if (!VarTy->isReferenceType()) {
510 if (ArgLVal.getType()->isLValueReferenceType()) {
511 ArgAddr = CGF.EmitLoadOfReference(ArgLVal);
512 } else if (!VarTy->isVariablyModifiedType() ||
513 !VarTy->isPointerType()) {
514 assert(ArgLVal.getType()->isPointerType());
515 ArgAddr = CGF.EmitLoadOfPointer(
516 ArgAddr, ArgLVal.getType()->castAs<PointerType>());
519 if (!FO.RegisterCastedArgsOnly) {
522 {Var, Address(ArgAddr.getPointer(), Ctx.getDeclAlign(Var))}});
524 } else if (I->capturesVariableByCopy()) {
525 assert(!FD->getType()->isAnyPointerType() &&
526 "Not expecting a captured pointer.");
527 const VarDecl *Var = I->getCapturedVar();
528 QualType VarTy = Var->getType();
531 {Var, FO.UIntPtrCastRequired
532 ? castValueFromUintptr(CGF, I->getLocation(),
533 FD->getType(), Args[Cnt]->getName(),
534 ArgLVal, VarTy->isReferenceType())
535 : ArgLVal.getAddress()}});
537 // If 'this' is captured, load it into CXXThisValue.
538 assert(I->capturesThis());
539 CXXThisValue = CGF.EmitLoadOfScalar(ArgLVal, I->getLocation());
540 LocalAddrs.insert({Args[Cnt], {nullptr, ArgLVal.getAddress()}});
550 CodeGenFunction::GenerateOpenMPCapturedStmtFunction(const CapturedStmt &S) {
553 "CapturedStmtInfo should be set when generating the captured function");
554 const CapturedDecl *CD = S.getCapturedDecl();
555 // Build the argument list.
556 bool NeedWrapperFunction =
558 CGM.getCodeGenOpts().getDebugInfo() >= codegenoptions::LimitedDebugInfo;
559 FunctionArgList Args;
560 llvm::MapVector<const Decl *, std::pair<const VarDecl *, Address>> LocalAddrs;
561 llvm::DenseMap<const Decl *, std::pair<const Expr *, llvm::Value *>> VLASizes;
562 SmallString<256> Buffer;
563 llvm::raw_svector_ostream Out(Buffer);
564 Out << CapturedStmtInfo->getHelperName();
565 if (NeedWrapperFunction)
567 FunctionOptions FO(&S, !NeedWrapperFunction, /*RegisterCastedArgsOnly=*/false,
569 llvm::Function *F = emitOutlinedFunctionPrologue(*this, Args, LocalAddrs,
570 VLASizes, CXXThisValue, FO);
571 for (const auto &LocalAddrPair : LocalAddrs) {
572 if (LocalAddrPair.second.first) {
573 setAddrOfLocalVar(LocalAddrPair.second.first,
574 LocalAddrPair.second.second);
577 for (const auto &VLASizePair : VLASizes)
578 VLASizeMap[VLASizePair.second.first] = VLASizePair.second.second;
579 PGO.assignRegionCounters(GlobalDecl(CD), F);
580 CapturedStmtInfo->EmitBody(*this, CD->getBody());
581 FinishFunction(CD->getBodyRBrace());
582 if (!NeedWrapperFunction)
585 FunctionOptions WrapperFO(&S, /*UIntPtrCastRequired=*/true,
586 /*RegisterCastedArgsOnly=*/true,
587 CapturedStmtInfo->getHelperName());
588 CodeGenFunction WrapperCGF(CGM, /*suppressNewContext=*/true);
589 WrapperCGF.CapturedStmtInfo = CapturedStmtInfo;
593 llvm::Function *WrapperF =
594 emitOutlinedFunctionPrologue(WrapperCGF, Args, LocalAddrs, VLASizes,
595 WrapperCGF.CXXThisValue, WrapperFO);
596 llvm::SmallVector<llvm::Value *, 4> CallArgs;
597 for (const auto *Arg : Args) {
598 llvm::Value *CallArg;
599 auto I = LocalAddrs.find(Arg);
600 if (I != LocalAddrs.end()) {
601 LValue LV = WrapperCGF.MakeAddrLValue(
603 I->second.first ? I->second.first->getType() : Arg->getType(),
604 AlignmentSource::Decl);
605 CallArg = WrapperCGF.EmitLoadOfScalar(LV, S.getBeginLoc());
607 auto EI = VLASizes.find(Arg);
608 if (EI != VLASizes.end()) {
609 CallArg = EI->second.second;
611 LValue LV = WrapperCGF.MakeAddrLValue(WrapperCGF.GetAddrOfLocalVar(Arg),
613 AlignmentSource::Decl);
614 CallArg = WrapperCGF.EmitLoadOfScalar(LV, S.getBeginLoc());
617 CallArgs.emplace_back(WrapperCGF.EmitFromMemory(CallArg, Arg->getType()));
619 CGM.getOpenMPRuntime().emitOutlinedFunctionCall(WrapperCGF, S.getBeginLoc(),
621 WrapperCGF.FinishFunction();
625 //===----------------------------------------------------------------------===//
626 // OpenMP Directive Emission
627 //===----------------------------------------------------------------------===//
628 void CodeGenFunction::EmitOMPAggregateAssign(
629 Address DestAddr, Address SrcAddr, QualType OriginalType,
630 const llvm::function_ref<void(Address, Address)> CopyGen) {
631 // Perform element-by-element initialization.
634 // Drill down to the base element type on both arrays.
635 const ArrayType *ArrayTy = OriginalType->getAsArrayTypeUnsafe();
636 llvm::Value *NumElements = emitArrayLength(ArrayTy, ElementTy, DestAddr);
637 SrcAddr = Builder.CreateElementBitCast(SrcAddr, DestAddr.getElementType());
639 llvm::Value *SrcBegin = SrcAddr.getPointer();
640 llvm::Value *DestBegin = DestAddr.getPointer();
641 // Cast from pointer to array type to pointer to single element.
642 llvm::Value *DestEnd = Builder.CreateGEP(DestBegin, NumElements);
643 // The basic structure here is a while-do loop.
644 llvm::BasicBlock *BodyBB = createBasicBlock("omp.arraycpy.body");
645 llvm::BasicBlock *DoneBB = createBasicBlock("omp.arraycpy.done");
646 llvm::Value *IsEmpty =
647 Builder.CreateICmpEQ(DestBegin, DestEnd, "omp.arraycpy.isempty");
648 Builder.CreateCondBr(IsEmpty, DoneBB, BodyBB);
650 // Enter the loop body, making that address the current address.
651 llvm::BasicBlock *EntryBB = Builder.GetInsertBlock();
654 CharUnits ElementSize = getContext().getTypeSizeInChars(ElementTy);
656 llvm::PHINode *SrcElementPHI =
657 Builder.CreatePHI(SrcBegin->getType(), 2, "omp.arraycpy.srcElementPast");
658 SrcElementPHI->addIncoming(SrcBegin, EntryBB);
659 Address SrcElementCurrent =
660 Address(SrcElementPHI,
661 SrcAddr.getAlignment().alignmentOfArrayElement(ElementSize));
663 llvm::PHINode *DestElementPHI =
664 Builder.CreatePHI(DestBegin->getType(), 2, "omp.arraycpy.destElementPast");
665 DestElementPHI->addIncoming(DestBegin, EntryBB);
666 Address DestElementCurrent =
667 Address(DestElementPHI,
668 DestAddr.getAlignment().alignmentOfArrayElement(ElementSize));
671 CopyGen(DestElementCurrent, SrcElementCurrent);
673 // Shift the address forward by one element.
674 llvm::Value *DestElementNext = Builder.CreateConstGEP1_32(
675 DestElementPHI, /*Idx0=*/1, "omp.arraycpy.dest.element");
676 llvm::Value *SrcElementNext = Builder.CreateConstGEP1_32(
677 SrcElementPHI, /*Idx0=*/1, "omp.arraycpy.src.element");
678 // Check whether we've reached the end.
680 Builder.CreateICmpEQ(DestElementNext, DestEnd, "omp.arraycpy.done");
681 Builder.CreateCondBr(Done, DoneBB, BodyBB);
682 DestElementPHI->addIncoming(DestElementNext, Builder.GetInsertBlock());
683 SrcElementPHI->addIncoming(SrcElementNext, Builder.GetInsertBlock());
686 EmitBlock(DoneBB, /*IsFinished=*/true);
689 void CodeGenFunction::EmitOMPCopy(QualType OriginalType, Address DestAddr,
690 Address SrcAddr, const VarDecl *DestVD,
691 const VarDecl *SrcVD, const Expr *Copy) {
692 if (OriginalType->isArrayType()) {
693 const auto *BO = dyn_cast<BinaryOperator>(Copy);
694 if (BO && BO->getOpcode() == BO_Assign) {
695 // Perform simple memcpy for simple copying.
696 LValue Dest = MakeAddrLValue(DestAddr, OriginalType);
697 LValue Src = MakeAddrLValue(SrcAddr, OriginalType);
698 EmitAggregateAssign(Dest, Src, OriginalType);
700 // For arrays with complex element types perform element by element
702 EmitOMPAggregateAssign(
703 DestAddr, SrcAddr, OriginalType,
704 [this, Copy, SrcVD, DestVD](Address DestElement, Address SrcElement) {
705 // Working with the single array element, so have to remap
706 // destination and source variables to corresponding array
708 CodeGenFunction::OMPPrivateScope Remap(*this);
709 Remap.addPrivate(DestVD, [DestElement]() { return DestElement; });
710 Remap.addPrivate(SrcVD, [SrcElement]() { return SrcElement; });
711 (void)Remap.Privatize();
712 EmitIgnoredExpr(Copy);
716 // Remap pseudo source variable to private copy.
717 CodeGenFunction::OMPPrivateScope Remap(*this);
718 Remap.addPrivate(SrcVD, [SrcAddr]() { return SrcAddr; });
719 Remap.addPrivate(DestVD, [DestAddr]() { return DestAddr; });
720 (void)Remap.Privatize();
721 // Emit copying of the whole variable.
722 EmitIgnoredExpr(Copy);
726 bool CodeGenFunction::EmitOMPFirstprivateClause(const OMPExecutableDirective &D,
727 OMPPrivateScope &PrivateScope) {
728 if (!HaveInsertPoint())
730 bool FirstprivateIsLastprivate = false;
731 llvm::DenseSet<const VarDecl *> Lastprivates;
732 for (const auto *C : D.getClausesOfKind<OMPLastprivateClause>()) {
733 for (const auto *D : C->varlists())
735 cast<VarDecl>(cast<DeclRefExpr>(D)->getDecl())->getCanonicalDecl());
737 llvm::DenseSet<const VarDecl *> EmittedAsFirstprivate;
738 llvm::SmallVector<OpenMPDirectiveKind, 4> CaptureRegions;
739 getOpenMPCaptureRegions(CaptureRegions, D.getDirectiveKind());
740 // Force emission of the firstprivate copy if the directive does not emit
741 // outlined function, like omp for, omp simd, omp distribute etc.
742 bool MustEmitFirstprivateCopy =
743 CaptureRegions.size() == 1 && CaptureRegions.back() == OMPD_unknown;
744 for (const auto *C : D.getClausesOfKind<OMPFirstprivateClause>()) {
745 auto IRef = C->varlist_begin();
746 auto InitsRef = C->inits().begin();
747 for (const Expr *IInit : C->private_copies()) {
748 const auto *OrigVD = cast<VarDecl>(cast<DeclRefExpr>(*IRef)->getDecl());
749 bool ThisFirstprivateIsLastprivate =
750 Lastprivates.count(OrigVD->getCanonicalDecl()) > 0;
751 const FieldDecl *FD = CapturedStmtInfo->lookup(OrigVD);
752 if (!MustEmitFirstprivateCopy && !ThisFirstprivateIsLastprivate && FD &&
753 !FD->getType()->isReferenceType()) {
754 EmittedAsFirstprivate.insert(OrigVD->getCanonicalDecl());
759 FirstprivateIsLastprivate =
760 FirstprivateIsLastprivate || ThisFirstprivateIsLastprivate;
761 if (EmittedAsFirstprivate.insert(OrigVD->getCanonicalDecl()).second) {
762 const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(IInit)->getDecl());
764 cast<VarDecl>(cast<DeclRefExpr>(*InitsRef)->getDecl());
766 DeclRefExpr DRE(getContext(), const_cast<VarDecl *>(OrigVD),
767 /*RefersToEnclosingVariableOrCapture=*/FD != nullptr,
768 (*IRef)->getType(), VK_LValue, (*IRef)->getExprLoc());
769 LValue OriginalLVal = EmitLValue(&DRE);
770 QualType Type = VD->getType();
771 if (Type->isArrayType()) {
772 // Emit VarDecl with copy init for arrays.
773 // Get the address of the original variable captured in current
775 IsRegistered = PrivateScope.addPrivate(
776 OrigVD, [this, VD, Type, OriginalLVal, VDInit]() {
777 AutoVarEmission Emission = EmitAutoVarAlloca(*VD);
778 const Expr *Init = VD->getInit();
779 if (!isa<CXXConstructExpr>(Init) ||
780 isTrivialInitializer(Init)) {
781 // Perform simple memcpy.
783 MakeAddrLValue(Emission.getAllocatedAddress(), Type);
784 EmitAggregateAssign(Dest, OriginalLVal, Type);
786 EmitOMPAggregateAssign(
787 Emission.getAllocatedAddress(), OriginalLVal.getAddress(),
789 [this, VDInit, Init](Address DestElement,
790 Address SrcElement) {
791 // Clean up any temporaries needed by the
793 RunCleanupsScope InitScope(*this);
794 // Emit initialization for single element.
795 setAddrOfLocalVar(VDInit, SrcElement);
796 EmitAnyExprToMem(Init, DestElement,
797 Init->getType().getQualifiers(),
798 /*IsInitializer*/ false);
799 LocalDeclMap.erase(VDInit);
802 EmitAutoVarCleanups(Emission);
803 return Emission.getAllocatedAddress();
806 Address OriginalAddr = OriginalLVal.getAddress();
807 IsRegistered = PrivateScope.addPrivate(
808 OrigVD, [this, VDInit, OriginalAddr, VD]() {
809 // Emit private VarDecl with copy init.
810 // Remap temp VDInit variable to the address of the original
811 // variable (for proper handling of captured global variables).
812 setAddrOfLocalVar(VDInit, OriginalAddr);
814 LocalDeclMap.erase(VDInit);
815 return GetAddrOfLocalVar(VD);
818 assert(IsRegistered &&
819 "firstprivate var already registered as private");
820 // Silence the warning about unused variable.
827 return FirstprivateIsLastprivate && !EmittedAsFirstprivate.empty();
830 void CodeGenFunction::EmitOMPPrivateClause(
831 const OMPExecutableDirective &D,
832 CodeGenFunction::OMPPrivateScope &PrivateScope) {
833 if (!HaveInsertPoint())
835 llvm::DenseSet<const VarDecl *> EmittedAsPrivate;
836 for (const auto *C : D.getClausesOfKind<OMPPrivateClause>()) {
837 auto IRef = C->varlist_begin();
838 for (const Expr *IInit : C->private_copies()) {
839 const auto *OrigVD = cast<VarDecl>(cast<DeclRefExpr>(*IRef)->getDecl());
840 if (EmittedAsPrivate.insert(OrigVD->getCanonicalDecl()).second) {
841 const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(IInit)->getDecl());
842 bool IsRegistered = PrivateScope.addPrivate(OrigVD, [this, VD]() {
843 // Emit private VarDecl with copy init.
845 return GetAddrOfLocalVar(VD);
847 assert(IsRegistered && "private var already registered as private");
848 // Silence the warning about unused variable.
856 bool CodeGenFunction::EmitOMPCopyinClause(const OMPExecutableDirective &D) {
857 if (!HaveInsertPoint())
859 // threadprivate_var1 = master_threadprivate_var1;
860 // operator=(threadprivate_var2, master_threadprivate_var2);
862 // __kmpc_barrier(&loc, global_tid);
863 llvm::DenseSet<const VarDecl *> CopiedVars;
864 llvm::BasicBlock *CopyBegin = nullptr, *CopyEnd = nullptr;
865 for (const auto *C : D.getClausesOfKind<OMPCopyinClause>()) {
866 auto IRef = C->varlist_begin();
867 auto ISrcRef = C->source_exprs().begin();
868 auto IDestRef = C->destination_exprs().begin();
869 for (const Expr *AssignOp : C->assignment_ops()) {
870 const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(*IRef)->getDecl());
871 QualType Type = VD->getType();
872 if (CopiedVars.insert(VD->getCanonicalDecl()).second) {
873 // Get the address of the master variable. If we are emitting code with
874 // TLS support, the address is passed from the master as field in the
875 // captured declaration.
876 Address MasterAddr = Address::invalid();
877 if (getLangOpts().OpenMPUseTLS &&
878 getContext().getTargetInfo().isTLSSupported()) {
879 assert(CapturedStmtInfo->lookup(VD) &&
880 "Copyin threadprivates should have been captured!");
881 DeclRefExpr DRE(getContext(), const_cast<VarDecl *>(VD), true,
882 (*IRef)->getType(), VK_LValue, (*IRef)->getExprLoc());
883 MasterAddr = EmitLValue(&DRE).getAddress();
884 LocalDeclMap.erase(VD);
887 Address(VD->isStaticLocal() ? CGM.getStaticLocalDeclAddress(VD)
888 : CGM.GetAddrOfGlobal(VD),
889 getContext().getDeclAlign(VD));
891 // Get the address of the threadprivate variable.
892 Address PrivateAddr = EmitLValue(*IRef).getAddress();
893 if (CopiedVars.size() == 1) {
894 // At first check if current thread is a master thread. If it is, no
895 // need to copy data.
896 CopyBegin = createBasicBlock("copyin.not.master");
897 CopyEnd = createBasicBlock("copyin.not.master.end");
898 Builder.CreateCondBr(
899 Builder.CreateICmpNE(
900 Builder.CreatePtrToInt(MasterAddr.getPointer(), CGM.IntPtrTy),
901 Builder.CreatePtrToInt(PrivateAddr.getPointer(),
904 EmitBlock(CopyBegin);
907 cast<VarDecl>(cast<DeclRefExpr>(*ISrcRef)->getDecl());
909 cast<VarDecl>(cast<DeclRefExpr>(*IDestRef)->getDecl());
910 EmitOMPCopy(Type, PrivateAddr, MasterAddr, DestVD, SrcVD, AssignOp);
918 // Exit out of copying procedure for non-master thread.
919 EmitBlock(CopyEnd, /*IsFinished=*/true);
925 bool CodeGenFunction::EmitOMPLastprivateClauseInit(
926 const OMPExecutableDirective &D, OMPPrivateScope &PrivateScope) {
927 if (!HaveInsertPoint())
929 bool HasAtLeastOneLastprivate = false;
930 llvm::DenseSet<const VarDecl *> SIMDLCVs;
931 if (isOpenMPSimdDirective(D.getDirectiveKind())) {
932 const auto *LoopDirective = cast<OMPLoopDirective>(&D);
933 for (const Expr *C : LoopDirective->counters()) {
935 cast<VarDecl>(cast<DeclRefExpr>(C)->getDecl())->getCanonicalDecl());
938 llvm::DenseSet<const VarDecl *> AlreadyEmittedVars;
939 for (const auto *C : D.getClausesOfKind<OMPLastprivateClause>()) {
940 HasAtLeastOneLastprivate = true;
941 if (isOpenMPTaskLoopDirective(D.getDirectiveKind()) &&
942 !getLangOpts().OpenMPSimd)
944 auto IRef = C->varlist_begin();
945 auto IDestRef = C->destination_exprs().begin();
946 for (const Expr *IInit : C->private_copies()) {
947 // Keep the address of the original variable for future update at the end
949 const auto *OrigVD = cast<VarDecl>(cast<DeclRefExpr>(*IRef)->getDecl());
950 // Taskloops do not require additional initialization, it is done in
951 // runtime support library.
952 if (AlreadyEmittedVars.insert(OrigVD->getCanonicalDecl()).second) {
954 cast<VarDecl>(cast<DeclRefExpr>(*IDestRef)->getDecl());
955 PrivateScope.addPrivate(DestVD, [this, OrigVD, IRef]() {
956 DeclRefExpr DRE(getContext(), const_cast<VarDecl *>(OrigVD),
957 /*RefersToEnclosingVariableOrCapture=*/
958 CapturedStmtInfo->lookup(OrigVD) != nullptr,
959 (*IRef)->getType(), VK_LValue, (*IRef)->getExprLoc());
960 return EmitLValue(&DRE).getAddress();
962 // Check if the variable is also a firstprivate: in this case IInit is
963 // not generated. Initialization of this variable will happen in codegen
964 // for 'firstprivate' clause.
965 if (IInit && !SIMDLCVs.count(OrigVD->getCanonicalDecl())) {
966 const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(IInit)->getDecl());
967 bool IsRegistered = PrivateScope.addPrivate(OrigVD, [this, VD]() {
968 // Emit private VarDecl with copy init.
970 return GetAddrOfLocalVar(VD);
972 assert(IsRegistered &&
973 "lastprivate var already registered as private");
981 return HasAtLeastOneLastprivate;
984 void CodeGenFunction::EmitOMPLastprivateClauseFinal(
985 const OMPExecutableDirective &D, bool NoFinals,
986 llvm::Value *IsLastIterCond) {
987 if (!HaveInsertPoint())
989 // Emit following code:
990 // if (<IsLastIterCond>) {
991 // orig_var1 = private_orig_var1;
993 // orig_varn = private_orig_varn;
995 llvm::BasicBlock *ThenBB = nullptr;
996 llvm::BasicBlock *DoneBB = nullptr;
997 if (IsLastIterCond) {
998 ThenBB = createBasicBlock(".omp.lastprivate.then");
999 DoneBB = createBasicBlock(".omp.lastprivate.done");
1000 Builder.CreateCondBr(IsLastIterCond, ThenBB, DoneBB);
1003 llvm::DenseSet<const VarDecl *> AlreadyEmittedVars;
1004 llvm::DenseMap<const VarDecl *, const Expr *> LoopCountersAndUpdates;
1005 if (const auto *LoopDirective = dyn_cast<OMPLoopDirective>(&D)) {
1006 auto IC = LoopDirective->counters().begin();
1007 for (const Expr *F : LoopDirective->finals()) {
1009 cast<VarDecl>(cast<DeclRefExpr>(*IC)->getDecl())->getCanonicalDecl();
1011 AlreadyEmittedVars.insert(D);
1013 LoopCountersAndUpdates[D] = F;
1017 for (const auto *C : D.getClausesOfKind<OMPLastprivateClause>()) {
1018 auto IRef = C->varlist_begin();
1019 auto ISrcRef = C->source_exprs().begin();
1020 auto IDestRef = C->destination_exprs().begin();
1021 for (const Expr *AssignOp : C->assignment_ops()) {
1022 const auto *PrivateVD =
1023 cast<VarDecl>(cast<DeclRefExpr>(*IRef)->getDecl());
1024 QualType Type = PrivateVD->getType();
1025 const auto *CanonicalVD = PrivateVD->getCanonicalDecl();
1026 if (AlreadyEmittedVars.insert(CanonicalVD).second) {
1027 // If lastprivate variable is a loop control variable for loop-based
1028 // directive, update its value before copyin back to original
1030 if (const Expr *FinalExpr = LoopCountersAndUpdates.lookup(CanonicalVD))
1031 EmitIgnoredExpr(FinalExpr);
1033 cast<VarDecl>(cast<DeclRefExpr>(*ISrcRef)->getDecl());
1034 const auto *DestVD =
1035 cast<VarDecl>(cast<DeclRefExpr>(*IDestRef)->getDecl());
1036 // Get the address of the original variable.
1037 Address OriginalAddr = GetAddrOfLocalVar(DestVD);
1038 // Get the address of the private variable.
1039 Address PrivateAddr = GetAddrOfLocalVar(PrivateVD);
1040 if (const auto *RefTy = PrivateVD->getType()->getAs<ReferenceType>())
1042 Address(Builder.CreateLoad(PrivateAddr),
1043 getNaturalTypeAlignment(RefTy->getPointeeType()));
1044 EmitOMPCopy(Type, OriginalAddr, PrivateAddr, DestVD, SrcVD, AssignOp);
1050 if (const Expr *PostUpdate = C->getPostUpdateExpr())
1051 EmitIgnoredExpr(PostUpdate);
1054 EmitBlock(DoneBB, /*IsFinished=*/true);
1057 void CodeGenFunction::EmitOMPReductionClauseInit(
1058 const OMPExecutableDirective &D,
1059 CodeGenFunction::OMPPrivateScope &PrivateScope) {
1060 if (!HaveInsertPoint())
1062 SmallVector<const Expr *, 4> Shareds;
1063 SmallVector<const Expr *, 4> Privates;
1064 SmallVector<const Expr *, 4> ReductionOps;
1065 SmallVector<const Expr *, 4> LHSs;
1066 SmallVector<const Expr *, 4> RHSs;
1067 for (const auto *C : D.getClausesOfKind<OMPReductionClause>()) {
1068 auto IPriv = C->privates().begin();
1069 auto IRed = C->reduction_ops().begin();
1070 auto ILHS = C->lhs_exprs().begin();
1071 auto IRHS = C->rhs_exprs().begin();
1072 for (const Expr *Ref : C->varlists()) {
1073 Shareds.emplace_back(Ref);
1074 Privates.emplace_back(*IPriv);
1075 ReductionOps.emplace_back(*IRed);
1076 LHSs.emplace_back(*ILHS);
1077 RHSs.emplace_back(*IRHS);
1078 std::advance(IPriv, 1);
1079 std::advance(IRed, 1);
1080 std::advance(ILHS, 1);
1081 std::advance(IRHS, 1);
1084 ReductionCodeGen RedCG(Shareds, Privates, ReductionOps);
1086 auto ILHS = LHSs.begin();
1087 auto IRHS = RHSs.begin();
1088 auto IPriv = Privates.begin();
1089 for (const Expr *IRef : Shareds) {
1090 const auto *PrivateVD = cast<VarDecl>(cast<DeclRefExpr>(*IPriv)->getDecl());
1091 // Emit private VarDecl with reduction init.
1092 RedCG.emitSharedLValue(*this, Count);
1093 RedCG.emitAggregateType(*this, Count);
1094 AutoVarEmission Emission = EmitAutoVarAlloca(*PrivateVD);
1095 RedCG.emitInitialization(*this, Count, Emission.getAllocatedAddress(),
1096 RedCG.getSharedLValue(Count),
1097 [&Emission](CodeGenFunction &CGF) {
1098 CGF.EmitAutoVarInit(Emission);
1101 EmitAutoVarCleanups(Emission);
1102 Address BaseAddr = RedCG.adjustPrivateAddress(
1103 *this, Count, Emission.getAllocatedAddress());
1104 bool IsRegistered = PrivateScope.addPrivate(
1105 RedCG.getBaseDecl(Count), [BaseAddr]() { return BaseAddr; });
1106 assert(IsRegistered && "private var already registered as private");
1107 // Silence the warning about unused variable.
1110 const auto *LHSVD = cast<VarDecl>(cast<DeclRefExpr>(*ILHS)->getDecl());
1111 const auto *RHSVD = cast<VarDecl>(cast<DeclRefExpr>(*IRHS)->getDecl());
1112 QualType Type = PrivateVD->getType();
1113 bool isaOMPArraySectionExpr = isa<OMPArraySectionExpr>(IRef);
1114 if (isaOMPArraySectionExpr && Type->isVariablyModifiedType()) {
1115 // Store the address of the original variable associated with the LHS
1116 // implicit variable.
1117 PrivateScope.addPrivate(LHSVD, [&RedCG, Count]() {
1118 return RedCG.getSharedLValue(Count).getAddress();
1120 PrivateScope.addPrivate(
1121 RHSVD, [this, PrivateVD]() { return GetAddrOfLocalVar(PrivateVD); });
1122 } else if ((isaOMPArraySectionExpr && Type->isScalarType()) ||
1123 isa<ArraySubscriptExpr>(IRef)) {
1124 // Store the address of the original variable associated with the LHS
1125 // implicit variable.
1126 PrivateScope.addPrivate(LHSVD, [&RedCG, Count]() {
1127 return RedCG.getSharedLValue(Count).getAddress();
1129 PrivateScope.addPrivate(RHSVD, [this, PrivateVD, RHSVD]() {
1130 return Builder.CreateElementBitCast(GetAddrOfLocalVar(PrivateVD),
1131 ConvertTypeForMem(RHSVD->getType()),
1135 QualType Type = PrivateVD->getType();
1136 bool IsArray = getContext().getAsArrayType(Type) != nullptr;
1137 Address OriginalAddr = RedCG.getSharedLValue(Count).getAddress();
1138 // Store the address of the original variable associated with the LHS
1139 // implicit variable.
1141 OriginalAddr = Builder.CreateElementBitCast(
1142 OriginalAddr, ConvertTypeForMem(LHSVD->getType()), "lhs.begin");
1144 PrivateScope.addPrivate(LHSVD, [OriginalAddr]() { return OriginalAddr; });
1145 PrivateScope.addPrivate(
1146 RHSVD, [this, PrivateVD, RHSVD, IsArray]() {
1148 ? Builder.CreateElementBitCast(
1149 GetAddrOfLocalVar(PrivateVD),
1150 ConvertTypeForMem(RHSVD->getType()), "rhs.begin")
1151 : GetAddrOfLocalVar(PrivateVD);
1161 void CodeGenFunction::EmitOMPReductionClauseFinal(
1162 const OMPExecutableDirective &D, const OpenMPDirectiveKind ReductionKind) {
1163 if (!HaveInsertPoint())
1165 llvm::SmallVector<const Expr *, 8> Privates;
1166 llvm::SmallVector<const Expr *, 8> LHSExprs;
1167 llvm::SmallVector<const Expr *, 8> RHSExprs;
1168 llvm::SmallVector<const Expr *, 8> ReductionOps;
1169 bool HasAtLeastOneReduction = false;
1170 for (const auto *C : D.getClausesOfKind<OMPReductionClause>()) {
1171 HasAtLeastOneReduction = true;
1172 Privates.append(C->privates().begin(), C->privates().end());
1173 LHSExprs.append(C->lhs_exprs().begin(), C->lhs_exprs().end());
1174 RHSExprs.append(C->rhs_exprs().begin(), C->rhs_exprs().end());
1175 ReductionOps.append(C->reduction_ops().begin(), C->reduction_ops().end());
1177 if (HasAtLeastOneReduction) {
1178 bool WithNowait = D.getSingleClause<OMPNowaitClause>() ||
1179 isOpenMPParallelDirective(D.getDirectiveKind()) ||
1180 ReductionKind == OMPD_simd;
1181 bool SimpleReduction = ReductionKind == OMPD_simd;
1182 // Emit nowait reduction if nowait clause is present or directive is a
1183 // parallel directive (it always has implicit barrier).
1184 CGM.getOpenMPRuntime().emitReduction(
1185 *this, D.getEndLoc(), Privates, LHSExprs, RHSExprs, ReductionOps,
1186 {WithNowait, SimpleReduction, ReductionKind});
1190 static void emitPostUpdateForReductionClause(
1191 CodeGenFunction &CGF, const OMPExecutableDirective &D,
1192 const llvm::function_ref<llvm::Value *(CodeGenFunction &)> CondGen) {
1193 if (!CGF.HaveInsertPoint())
1195 llvm::BasicBlock *DoneBB = nullptr;
1196 for (const auto *C : D.getClausesOfKind<OMPReductionClause>()) {
1197 if (const Expr *PostUpdate = C->getPostUpdateExpr()) {
1199 if (llvm::Value *Cond = CondGen(CGF)) {
1200 // If the first post-update expression is found, emit conditional
1201 // block if it was requested.
1202 llvm::BasicBlock *ThenBB = CGF.createBasicBlock(".omp.reduction.pu");
1203 DoneBB = CGF.createBasicBlock(".omp.reduction.pu.done");
1204 CGF.Builder.CreateCondBr(Cond, ThenBB, DoneBB);
1205 CGF.EmitBlock(ThenBB);
1208 CGF.EmitIgnoredExpr(PostUpdate);
1212 CGF.EmitBlock(DoneBB, /*IsFinished=*/true);
1216 /// Codegen lambda for appending distribute lower and upper bounds to outlined
1217 /// parallel function. This is necessary for combined constructs such as
1218 /// 'distribute parallel for'
1219 typedef llvm::function_ref<void(CodeGenFunction &,
1220 const OMPExecutableDirective &,
1221 llvm::SmallVectorImpl<llvm::Value *> &)>
1222 CodeGenBoundParametersTy;
1223 } // anonymous namespace
1225 static void emitCommonOMPParallelDirective(
1226 CodeGenFunction &CGF, const OMPExecutableDirective &S,
1227 OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen,
1228 const CodeGenBoundParametersTy &CodeGenBoundParameters) {
1229 const CapturedStmt *CS = S.getCapturedStmt(OMPD_parallel);
1230 llvm::Value *OutlinedFn =
1231 CGF.CGM.getOpenMPRuntime().emitParallelOutlinedFunction(
1232 S, *CS->getCapturedDecl()->param_begin(), InnermostKind, CodeGen);
1233 if (const auto *NumThreadsClause = S.getSingleClause<OMPNumThreadsClause>()) {
1234 CodeGenFunction::RunCleanupsScope NumThreadsScope(CGF);
1235 llvm::Value *NumThreads =
1236 CGF.EmitScalarExpr(NumThreadsClause->getNumThreads(),
1237 /*IgnoreResultAssign=*/true);
1238 CGF.CGM.getOpenMPRuntime().emitNumThreadsClause(
1239 CGF, NumThreads, NumThreadsClause->getBeginLoc());
1241 if (const auto *ProcBindClause = S.getSingleClause<OMPProcBindClause>()) {
1242 CodeGenFunction::RunCleanupsScope ProcBindScope(CGF);
1243 CGF.CGM.getOpenMPRuntime().emitProcBindClause(
1244 CGF, ProcBindClause->getProcBindKind(), ProcBindClause->getBeginLoc());
1246 const Expr *IfCond = nullptr;
1247 for (const auto *C : S.getClausesOfKind<OMPIfClause>()) {
1248 if (C->getNameModifier() == OMPD_unknown ||
1249 C->getNameModifier() == OMPD_parallel) {
1250 IfCond = C->getCondition();
1255 OMPParallelScope Scope(CGF, S);
1256 llvm::SmallVector<llvm::Value *, 16> CapturedVars;
1257 // Combining 'distribute' with 'for' requires sharing each 'distribute' chunk
1258 // lower and upper bounds with the pragma 'for' chunking mechanism.
1259 // The following lambda takes care of appending the lower and upper bound
1260 // parameters when necessary
1261 CodeGenBoundParameters(CGF, S, CapturedVars);
1262 CGF.GenerateOpenMPCapturedVars(*CS, CapturedVars);
1263 CGF.CGM.getOpenMPRuntime().emitParallelCall(CGF, S.getBeginLoc(), OutlinedFn,
1264 CapturedVars, IfCond);
1267 static void emitEmptyBoundParameters(CodeGenFunction &,
1268 const OMPExecutableDirective &,
1269 llvm::SmallVectorImpl<llvm::Value *> &) {}
1271 void CodeGenFunction::EmitOMPParallelDirective(const OMPParallelDirective &S) {
1272 // Emit parallel region as a standalone region.
1273 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
1275 OMPPrivateScope PrivateScope(CGF);
1276 bool Copyins = CGF.EmitOMPCopyinClause(S);
1277 (void)CGF.EmitOMPFirstprivateClause(S, PrivateScope);
1279 // Emit implicit barrier to synchronize threads and avoid data races on
1280 // propagation master's thread values of threadprivate variables to local
1281 // instances of that variables of all other implicit threads.
1282 CGF.CGM.getOpenMPRuntime().emitBarrierCall(
1283 CGF, S.getBeginLoc(), OMPD_unknown, /*EmitChecks=*/false,
1284 /*ForceSimpleCall=*/true);
1286 CGF.EmitOMPPrivateClause(S, PrivateScope);
1287 CGF.EmitOMPReductionClauseInit(S, PrivateScope);
1288 (void)PrivateScope.Privatize();
1289 CGF.EmitStmt(S.getCapturedStmt(OMPD_parallel)->getCapturedStmt());
1290 CGF.EmitOMPReductionClauseFinal(S, /*ReductionKind=*/OMPD_parallel);
1292 emitCommonOMPParallelDirective(*this, S, OMPD_parallel, CodeGen,
1293 emitEmptyBoundParameters);
1294 emitPostUpdateForReductionClause(*this, S,
1295 [](CodeGenFunction &) { return nullptr; });
1298 void CodeGenFunction::EmitOMPLoopBody(const OMPLoopDirective &D,
1299 JumpDest LoopExit) {
1300 RunCleanupsScope BodyScope(*this);
1301 // Update counters values on current iteration.
1302 for (const Expr *UE : D.updates())
1303 EmitIgnoredExpr(UE);
1304 // Update the linear variables.
1305 // In distribute directives only loop counters may be marked as linear, no
1306 // need to generate the code for them.
1307 if (!isOpenMPDistributeDirective(D.getDirectiveKind())) {
1308 for (const auto *C : D.getClausesOfKind<OMPLinearClause>()) {
1309 for (const Expr *UE : C->updates())
1310 EmitIgnoredExpr(UE);
1314 // On a continue in the body, jump to the end.
1315 JumpDest Continue = getJumpDestInCurrentScope("omp.body.continue");
1316 BreakContinueStack.push_back(BreakContinue(LoopExit, Continue));
1318 EmitStmt(D.getBody());
1319 // The end (updates/cleanups).
1320 EmitBlock(Continue.getBlock());
1321 BreakContinueStack.pop_back();
1324 void CodeGenFunction::EmitOMPInnerLoop(
1325 const Stmt &S, bool RequiresCleanup, const Expr *LoopCond,
1326 const Expr *IncExpr,
1327 const llvm::function_ref<void(CodeGenFunction &)> BodyGen,
1328 const llvm::function_ref<void(CodeGenFunction &)> PostIncGen) {
1329 auto LoopExit = getJumpDestInCurrentScope("omp.inner.for.end");
1331 // Start the loop with a block that tests the condition.
1332 auto CondBlock = createBasicBlock("omp.inner.for.cond");
1333 EmitBlock(CondBlock);
1334 const SourceRange R = S.getSourceRange();
1335 LoopStack.push(CondBlock, SourceLocToDebugLoc(R.getBegin()),
1336 SourceLocToDebugLoc(R.getEnd()));
1338 // If there are any cleanups between here and the loop-exit scope,
1339 // create a block to stage a loop exit along.
1340 llvm::BasicBlock *ExitBlock = LoopExit.getBlock();
1341 if (RequiresCleanup)
1342 ExitBlock = createBasicBlock("omp.inner.for.cond.cleanup");
1344 llvm::BasicBlock *LoopBody = createBasicBlock("omp.inner.for.body");
1347 EmitBranchOnBoolExpr(LoopCond, LoopBody, ExitBlock, getProfileCount(&S));
1348 if (ExitBlock != LoopExit.getBlock()) {
1349 EmitBlock(ExitBlock);
1350 EmitBranchThroughCleanup(LoopExit);
1353 EmitBlock(LoopBody);
1354 incrementProfileCounter(&S);
1356 // Create a block for the increment.
1357 JumpDest Continue = getJumpDestInCurrentScope("omp.inner.for.inc");
1358 BreakContinueStack.push_back(BreakContinue(LoopExit, Continue));
1362 // Emit "IV = IV + 1" and a back-edge to the condition block.
1363 EmitBlock(Continue.getBlock());
1364 EmitIgnoredExpr(IncExpr);
1366 BreakContinueStack.pop_back();
1367 EmitBranch(CondBlock);
1369 // Emit the fall-through block.
1370 EmitBlock(LoopExit.getBlock());
1373 bool CodeGenFunction::EmitOMPLinearClauseInit(const OMPLoopDirective &D) {
1374 if (!HaveInsertPoint())
1376 // Emit inits for the linear variables.
1377 bool HasLinears = false;
1378 for (const auto *C : D.getClausesOfKind<OMPLinearClause>()) {
1379 for (const Expr *Init : C->inits()) {
1381 const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(Init)->getDecl());
1382 if (const auto *Ref =
1383 dyn_cast<DeclRefExpr>(VD->getInit()->IgnoreImpCasts())) {
1384 AutoVarEmission Emission = EmitAutoVarAlloca(*VD);
1385 const auto *OrigVD = cast<VarDecl>(Ref->getDecl());
1386 DeclRefExpr DRE(getContext(), const_cast<VarDecl *>(OrigVD),
1387 CapturedStmtInfo->lookup(OrigVD) != nullptr,
1388 VD->getInit()->getType(), VK_LValue,
1389 VD->getInit()->getExprLoc());
1390 EmitExprAsInit(&DRE, VD, MakeAddrLValue(Emission.getAllocatedAddress(),
1392 /*capturedByInit=*/false);
1393 EmitAutoVarCleanups(Emission);
1398 // Emit the linear steps for the linear clauses.
1399 // If a step is not constant, it is pre-calculated before the loop.
1400 if (const auto *CS = cast_or_null<BinaryOperator>(C->getCalcStep()))
1401 if (const auto *SaveRef = cast<DeclRefExpr>(CS->getLHS())) {
1402 EmitVarDecl(*cast<VarDecl>(SaveRef->getDecl()));
1403 // Emit calculation of the linear step.
1404 EmitIgnoredExpr(CS);
1410 void CodeGenFunction::EmitOMPLinearClauseFinal(
1411 const OMPLoopDirective &D,
1412 const llvm::function_ref<llvm::Value *(CodeGenFunction &)> CondGen) {
1413 if (!HaveInsertPoint())
1415 llvm::BasicBlock *DoneBB = nullptr;
1416 // Emit the final values of the linear variables.
1417 for (const auto *C : D.getClausesOfKind<OMPLinearClause>()) {
1418 auto IC = C->varlist_begin();
1419 for (const Expr *F : C->finals()) {
1421 if (llvm::Value *Cond = CondGen(*this)) {
1422 // If the first post-update expression is found, emit conditional
1423 // block if it was requested.
1424 llvm::BasicBlock *ThenBB = createBasicBlock(".omp.linear.pu");
1425 DoneBB = createBasicBlock(".omp.linear.pu.done");
1426 Builder.CreateCondBr(Cond, ThenBB, DoneBB);
1430 const auto *OrigVD = cast<VarDecl>(cast<DeclRefExpr>(*IC)->getDecl());
1431 DeclRefExpr DRE(getContext(), const_cast<VarDecl *>(OrigVD),
1432 CapturedStmtInfo->lookup(OrigVD) != nullptr,
1433 (*IC)->getType(), VK_LValue, (*IC)->getExprLoc());
1434 Address OrigAddr = EmitLValue(&DRE).getAddress();
1435 CodeGenFunction::OMPPrivateScope VarScope(*this);
1436 VarScope.addPrivate(OrigVD, [OrigAddr]() { return OrigAddr; });
1437 (void)VarScope.Privatize();
1441 if (const Expr *PostUpdate = C->getPostUpdateExpr())
1442 EmitIgnoredExpr(PostUpdate);
1445 EmitBlock(DoneBB, /*IsFinished=*/true);
1448 static void emitAlignedClause(CodeGenFunction &CGF,
1449 const OMPExecutableDirective &D) {
1450 if (!CGF.HaveInsertPoint())
1452 for (const auto *Clause : D.getClausesOfKind<OMPAlignedClause>()) {
1453 unsigned ClauseAlignment = 0;
1454 if (const Expr *AlignmentExpr = Clause->getAlignment()) {
1456 cast<llvm::ConstantInt>(CGF.EmitScalarExpr(AlignmentExpr));
1457 ClauseAlignment = static_cast<unsigned>(AlignmentCI->getZExtValue());
1459 for (const Expr *E : Clause->varlists()) {
1460 unsigned Alignment = ClauseAlignment;
1461 if (Alignment == 0) {
1462 // OpenMP [2.8.1, Description]
1463 // If no optional parameter is specified, implementation-defined default
1464 // alignments for SIMD instructions on the target platforms are assumed.
1467 .toCharUnitsFromBits(CGF.getContext().getOpenMPDefaultSimdAlign(
1468 E->getType()->getPointeeType()))
1471 assert((Alignment == 0 || llvm::isPowerOf2_32(Alignment)) &&
1472 "alignment is not power of 2");
1473 if (Alignment != 0) {
1474 llvm::Value *PtrValue = CGF.EmitScalarExpr(E);
1475 CGF.EmitAlignmentAssumption(
1476 PtrValue, E, /*No second loc needed*/ SourceLocation(), Alignment);
1482 void CodeGenFunction::EmitOMPPrivateLoopCounters(
1483 const OMPLoopDirective &S, CodeGenFunction::OMPPrivateScope &LoopScope) {
1484 if (!HaveInsertPoint())
1486 auto I = S.private_counters().begin();
1487 for (const Expr *E : S.counters()) {
1488 const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl());
1489 const auto *PrivateVD = cast<VarDecl>(cast<DeclRefExpr>(*I)->getDecl());
1490 // Emit var without initialization.
1491 AutoVarEmission VarEmission = EmitAutoVarAlloca(*PrivateVD);
1492 EmitAutoVarCleanups(VarEmission);
1493 LocalDeclMap.erase(PrivateVD);
1494 (void)LoopScope.addPrivate(VD, [&VarEmission]() {
1495 return VarEmission.getAllocatedAddress();
1497 if (LocalDeclMap.count(VD) || CapturedStmtInfo->lookup(VD) ||
1498 VD->hasGlobalStorage()) {
1499 (void)LoopScope.addPrivate(PrivateVD, [this, VD, E]() {
1500 DeclRefExpr DRE(getContext(), const_cast<VarDecl *>(VD),
1501 LocalDeclMap.count(VD) || CapturedStmtInfo->lookup(VD),
1502 E->getType(), VK_LValue, E->getExprLoc());
1503 return EmitLValue(&DRE).getAddress();
1506 (void)LoopScope.addPrivate(PrivateVD, [&VarEmission]() {
1507 return VarEmission.getAllocatedAddress();
1512 // Privatize extra loop counters used in loops for ordered(n) clauses.
1513 for (const auto *C : S.getClausesOfKind<OMPOrderedClause>()) {
1514 if (!C->getNumForLoops())
1516 for (unsigned I = S.getCollapsedNumber(),
1517 E = C->getLoopNumIterations().size();
1519 const auto *DRE = cast<DeclRefExpr>(C->getLoopCounter(I));
1520 const auto *VD = cast<VarDecl>(DRE->getDecl());
1521 // Override only those variables that are really emitted already.
1522 if (LocalDeclMap.count(VD)) {
1523 (void)LoopScope.addPrivate(VD, [this, DRE, VD]() {
1524 return CreateMemTemp(DRE->getType(), VD->getName());
1531 static void emitPreCond(CodeGenFunction &CGF, const OMPLoopDirective &S,
1532 const Expr *Cond, llvm::BasicBlock *TrueBlock,
1533 llvm::BasicBlock *FalseBlock, uint64_t TrueCount) {
1534 if (!CGF.HaveInsertPoint())
1537 CodeGenFunction::OMPPrivateScope PreCondScope(CGF);
1538 CGF.EmitOMPPrivateLoopCounters(S, PreCondScope);
1539 (void)PreCondScope.Privatize();
1540 // Get initial values of real counters.
1541 for (const Expr *I : S.inits()) {
1542 CGF.EmitIgnoredExpr(I);
1545 // Check that loop is executed at least one time.
1546 CGF.EmitBranchOnBoolExpr(Cond, TrueBlock, FalseBlock, TrueCount);
1549 void CodeGenFunction::EmitOMPLinearClause(
1550 const OMPLoopDirective &D, CodeGenFunction::OMPPrivateScope &PrivateScope) {
1551 if (!HaveInsertPoint())
1553 llvm::DenseSet<const VarDecl *> SIMDLCVs;
1554 if (isOpenMPSimdDirective(D.getDirectiveKind())) {
1555 const auto *LoopDirective = cast<OMPLoopDirective>(&D);
1556 for (const Expr *C : LoopDirective->counters()) {
1558 cast<VarDecl>(cast<DeclRefExpr>(C)->getDecl())->getCanonicalDecl());
1561 for (const auto *C : D.getClausesOfKind<OMPLinearClause>()) {
1562 auto CurPrivate = C->privates().begin();
1563 for (const Expr *E : C->varlists()) {
1564 const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl());
1565 const auto *PrivateVD =
1566 cast<VarDecl>(cast<DeclRefExpr>(*CurPrivate)->getDecl());
1567 if (!SIMDLCVs.count(VD->getCanonicalDecl())) {
1568 bool IsRegistered = PrivateScope.addPrivate(VD, [this, PrivateVD]() {
1569 // Emit private VarDecl with copy init.
1570 EmitVarDecl(*PrivateVD);
1571 return GetAddrOfLocalVar(PrivateVD);
1573 assert(IsRegistered && "linear var already registered as private");
1574 // Silence the warning about unused variable.
1577 EmitVarDecl(*PrivateVD);
1584 static void emitSimdlenSafelenClause(CodeGenFunction &CGF,
1585 const OMPExecutableDirective &D,
1587 if (!CGF.HaveInsertPoint())
1589 if (const auto *C = D.getSingleClause<OMPSimdlenClause>()) {
1590 RValue Len = CGF.EmitAnyExpr(C->getSimdlen(), AggValueSlot::ignored(),
1591 /*ignoreResult=*/true);
1592 auto *Val = cast<llvm::ConstantInt>(Len.getScalarVal());
1593 CGF.LoopStack.setVectorizeWidth(Val->getZExtValue());
1594 // In presence of finite 'safelen', it may be unsafe to mark all
1595 // the memory instructions parallel, because loop-carried
1596 // dependences of 'safelen' iterations are possible.
1598 CGF.LoopStack.setParallel(!D.getSingleClause<OMPSafelenClause>());
1599 } else if (const auto *C = D.getSingleClause<OMPSafelenClause>()) {
1600 RValue Len = CGF.EmitAnyExpr(C->getSafelen(), AggValueSlot::ignored(),
1601 /*ignoreResult=*/true);
1602 auto *Val = cast<llvm::ConstantInt>(Len.getScalarVal());
1603 CGF.LoopStack.setVectorizeWidth(Val->getZExtValue());
1604 // In presence of finite 'safelen', it may be unsafe to mark all
1605 // the memory instructions parallel, because loop-carried
1606 // dependences of 'safelen' iterations are possible.
1607 CGF.LoopStack.setParallel(/*Enable=*/false);
1611 void CodeGenFunction::EmitOMPSimdInit(const OMPLoopDirective &D,
1613 // Walk clauses and process safelen/lastprivate.
1614 LoopStack.setParallel(!IsMonotonic);
1615 LoopStack.setVectorizeEnable();
1616 emitSimdlenSafelenClause(*this, D, IsMonotonic);
1619 void CodeGenFunction::EmitOMPSimdFinal(
1620 const OMPLoopDirective &D,
1621 const llvm::function_ref<llvm::Value *(CodeGenFunction &)> CondGen) {
1622 if (!HaveInsertPoint())
1624 llvm::BasicBlock *DoneBB = nullptr;
1625 auto IC = D.counters().begin();
1626 auto IPC = D.private_counters().begin();
1627 for (const Expr *F : D.finals()) {
1628 const auto *OrigVD = cast<VarDecl>(cast<DeclRefExpr>((*IC))->getDecl());
1629 const auto *PrivateVD = cast<VarDecl>(cast<DeclRefExpr>((*IPC))->getDecl());
1630 const auto *CED = dyn_cast<OMPCapturedExprDecl>(OrigVD);
1631 if (LocalDeclMap.count(OrigVD) || CapturedStmtInfo->lookup(OrigVD) ||
1632 OrigVD->hasGlobalStorage() || CED) {
1634 if (llvm::Value *Cond = CondGen(*this)) {
1635 // If the first post-update expression is found, emit conditional
1636 // block if it was requested.
1637 llvm::BasicBlock *ThenBB = createBasicBlock(".omp.final.then");
1638 DoneBB = createBasicBlock(".omp.final.done");
1639 Builder.CreateCondBr(Cond, ThenBB, DoneBB);
1643 Address OrigAddr = Address::invalid();
1645 OrigAddr = EmitLValue(CED->getInit()->IgnoreImpCasts()).getAddress();
1647 DeclRefExpr DRE(getContext(), const_cast<VarDecl *>(PrivateVD),
1648 /*RefersToEnclosingVariableOrCapture=*/false,
1649 (*IPC)->getType(), VK_LValue, (*IPC)->getExprLoc());
1650 OrigAddr = EmitLValue(&DRE).getAddress();
1652 OMPPrivateScope VarScope(*this);
1653 VarScope.addPrivate(OrigVD, [OrigAddr]() { return OrigAddr; });
1654 (void)VarScope.Privatize();
1661 EmitBlock(DoneBB, /*IsFinished=*/true);
1664 static void emitOMPLoopBodyWithStopPoint(CodeGenFunction &CGF,
1665 const OMPLoopDirective &S,
1666 CodeGenFunction::JumpDest LoopExit) {
1667 CGF.EmitOMPLoopBody(S, LoopExit);
1668 CGF.EmitStopPoint(&S);
1671 /// Emit a helper variable and return corresponding lvalue.
1672 static LValue EmitOMPHelperVar(CodeGenFunction &CGF,
1673 const DeclRefExpr *Helper) {
1674 auto VDecl = cast<VarDecl>(Helper->getDecl());
1675 CGF.EmitVarDecl(*VDecl);
1676 return CGF.EmitLValue(Helper);
1679 static void emitOMPSimdRegion(CodeGenFunction &CGF, const OMPLoopDirective &S,
1680 PrePostActionTy &Action) {
1682 assert(isOpenMPSimdDirective(S.getDirectiveKind()) &&
1683 "Expected simd directive");
1684 OMPLoopScope PreInitScope(CGF, S);
1686 // for (IV in 0..LastIteration) BODY;
1687 // <Final counter/linear vars updates>;
1690 if (isOpenMPDistributeDirective(S.getDirectiveKind()) ||
1691 isOpenMPWorksharingDirective(S.getDirectiveKind()) ||
1692 isOpenMPTaskLoopDirective(S.getDirectiveKind())) {
1693 (void)EmitOMPHelperVar(CGF, cast<DeclRefExpr>(S.getLowerBoundVariable()));
1694 (void)EmitOMPHelperVar(CGF, cast<DeclRefExpr>(S.getUpperBoundVariable()));
1697 // Emit: if (PreCond) - begin.
1698 // If the condition constant folds and can be elided, avoid emitting the
1701 llvm::BasicBlock *ContBlock = nullptr;
1702 if (CGF.ConstantFoldsToSimpleInteger(S.getPreCond(), CondConstant)) {
1706 llvm::BasicBlock *ThenBlock = CGF.createBasicBlock("simd.if.then");
1707 ContBlock = CGF.createBasicBlock("simd.if.end");
1708 emitPreCond(CGF, S, S.getPreCond(), ThenBlock, ContBlock,
1709 CGF.getProfileCount(&S));
1710 CGF.EmitBlock(ThenBlock);
1711 CGF.incrementProfileCounter(&S);
1714 // Emit the loop iteration variable.
1715 const Expr *IVExpr = S.getIterationVariable();
1716 const auto *IVDecl = cast<VarDecl>(cast<DeclRefExpr>(IVExpr)->getDecl());
1717 CGF.EmitVarDecl(*IVDecl);
1718 CGF.EmitIgnoredExpr(S.getInit());
1720 // Emit the iterations count variable.
1721 // If it is not a variable, Sema decided to calculate iterations count on
1722 // each iteration (e.g., it is foldable into a constant).
1723 if (const auto *LIExpr = dyn_cast<DeclRefExpr>(S.getLastIteration())) {
1724 CGF.EmitVarDecl(*cast<VarDecl>(LIExpr->getDecl()));
1725 // Emit calculation of the iterations count.
1726 CGF.EmitIgnoredExpr(S.getCalcLastIteration());
1729 CGF.EmitOMPSimdInit(S);
1731 emitAlignedClause(CGF, S);
1732 (void)CGF.EmitOMPLinearClauseInit(S);
1734 CodeGenFunction::OMPPrivateScope LoopScope(CGF);
1735 CGF.EmitOMPPrivateLoopCounters(S, LoopScope);
1736 CGF.EmitOMPLinearClause(S, LoopScope);
1737 CGF.EmitOMPPrivateClause(S, LoopScope);
1738 CGF.EmitOMPReductionClauseInit(S, LoopScope);
1739 bool HasLastprivateClause = CGF.EmitOMPLastprivateClauseInit(S, LoopScope);
1740 (void)LoopScope.Privatize();
1741 if (isOpenMPTargetExecutionDirective(S.getDirectiveKind()))
1742 CGF.CGM.getOpenMPRuntime().adjustTargetSpecificDataForLambdas(CGF, S);
1743 CGF.EmitOMPInnerLoop(S, LoopScope.requiresCleanups(), S.getCond(),
1745 [&S](CodeGenFunction &CGF) {
1746 CGF.EmitOMPLoopBody(S, CodeGenFunction::JumpDest());
1747 CGF.EmitStopPoint(&S);
1749 [](CodeGenFunction &) {});
1750 CGF.EmitOMPSimdFinal(S, [](CodeGenFunction &) { return nullptr; });
1751 // Emit final copy of the lastprivate variables at the end of loops.
1752 if (HasLastprivateClause)
1753 CGF.EmitOMPLastprivateClauseFinal(S, /*NoFinals=*/true);
1754 CGF.EmitOMPReductionClauseFinal(S, /*ReductionKind=*/OMPD_simd);
1755 emitPostUpdateForReductionClause(CGF, S,
1756 [](CodeGenFunction &) { return nullptr; });
1758 CGF.EmitOMPLinearClauseFinal(S, [](CodeGenFunction &) { return nullptr; });
1759 // Emit: if (PreCond) - end.
1761 CGF.EmitBranch(ContBlock);
1762 CGF.EmitBlock(ContBlock, true);
1766 void CodeGenFunction::EmitOMPSimdDirective(const OMPSimdDirective &S) {
1767 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
1768 emitOMPSimdRegion(CGF, S, Action);
1770 OMPLexicalScope Scope(*this, S, OMPD_unknown);
1771 CGM.getOpenMPRuntime().emitInlinedDirective(*this, OMPD_simd, CodeGen);
1774 void CodeGenFunction::EmitOMPOuterLoop(
1775 bool DynamicOrOrdered, bool IsMonotonic, const OMPLoopDirective &S,
1776 CodeGenFunction::OMPPrivateScope &LoopScope,
1777 const CodeGenFunction::OMPLoopArguments &LoopArgs,
1778 const CodeGenFunction::CodeGenLoopTy &CodeGenLoop,
1779 const CodeGenFunction::CodeGenOrderedTy &CodeGenOrdered) {
1780 CGOpenMPRuntime &RT = CGM.getOpenMPRuntime();
1782 const Expr *IVExpr = S.getIterationVariable();
1783 const unsigned IVSize = getContext().getTypeSize(IVExpr->getType());
1784 const bool IVSigned = IVExpr->getType()->hasSignedIntegerRepresentation();
1786 JumpDest LoopExit = getJumpDestInCurrentScope("omp.dispatch.end");
1788 // Start the loop with a block that tests the condition.
1789 llvm::BasicBlock *CondBlock = createBasicBlock("omp.dispatch.cond");
1790 EmitBlock(CondBlock);
1791 const SourceRange R = S.getSourceRange();
1792 LoopStack.push(CondBlock, SourceLocToDebugLoc(R.getBegin()),
1793 SourceLocToDebugLoc(R.getEnd()));
1795 llvm::Value *BoolCondVal = nullptr;
1796 if (!DynamicOrOrdered) {
1797 // UB = min(UB, GlobalUB) or
1798 // UB = min(UB, PrevUB) for combined loop sharing constructs (e.g.
1799 // 'distribute parallel for')
1800 EmitIgnoredExpr(LoopArgs.EUB);
1802 EmitIgnoredExpr(LoopArgs.Init);
1804 BoolCondVal = EvaluateExprAsBool(LoopArgs.Cond);
1807 RT.emitForNext(*this, S.getBeginLoc(), IVSize, IVSigned, LoopArgs.IL,
1808 LoopArgs.LB, LoopArgs.UB, LoopArgs.ST);
1811 // If there are any cleanups between here and the loop-exit scope,
1812 // create a block to stage a loop exit along.
1813 llvm::BasicBlock *ExitBlock = LoopExit.getBlock();
1814 if (LoopScope.requiresCleanups())
1815 ExitBlock = createBasicBlock("omp.dispatch.cleanup");
1817 llvm::BasicBlock *LoopBody = createBasicBlock("omp.dispatch.body");
1818 Builder.CreateCondBr(BoolCondVal, LoopBody, ExitBlock);
1819 if (ExitBlock != LoopExit.getBlock()) {
1820 EmitBlock(ExitBlock);
1821 EmitBranchThroughCleanup(LoopExit);
1823 EmitBlock(LoopBody);
1825 // Emit "IV = LB" (in case of static schedule, we have already calculated new
1826 // LB for loop condition and emitted it above).
1827 if (DynamicOrOrdered)
1828 EmitIgnoredExpr(LoopArgs.Init);
1830 // Create a block for the increment.
1831 JumpDest Continue = getJumpDestInCurrentScope("omp.dispatch.inc");
1832 BreakContinueStack.push_back(BreakContinue(LoopExit, Continue));
1834 // Generate !llvm.loop.parallel metadata for loads and stores for loops
1835 // with dynamic/guided scheduling and without ordered clause.
1836 if (!isOpenMPSimdDirective(S.getDirectiveKind()))
1837 LoopStack.setParallel(!IsMonotonic);
1839 EmitOMPSimdInit(S, IsMonotonic);
1841 SourceLocation Loc = S.getBeginLoc();
1843 // when 'distribute' is not combined with a 'for':
1844 // while (idx <= UB) { BODY; ++idx; }
1845 // when 'distribute' is combined with a 'for'
1846 // (e.g. 'distribute parallel for')
1847 // while (idx <= UB) { <CodeGen rest of pragma>; idx += ST; }
1849 S, LoopScope.requiresCleanups(), LoopArgs.Cond, LoopArgs.IncExpr,
1850 [&S, LoopExit, &CodeGenLoop](CodeGenFunction &CGF) {
1851 CodeGenLoop(CGF, S, LoopExit);
1853 [IVSize, IVSigned, Loc, &CodeGenOrdered](CodeGenFunction &CGF) {
1854 CodeGenOrdered(CGF, Loc, IVSize, IVSigned);
1857 EmitBlock(Continue.getBlock());
1858 BreakContinueStack.pop_back();
1859 if (!DynamicOrOrdered) {
1860 // Emit "LB = LB + Stride", "UB = UB + Stride".
1861 EmitIgnoredExpr(LoopArgs.NextLB);
1862 EmitIgnoredExpr(LoopArgs.NextUB);
1865 EmitBranch(CondBlock);
1867 // Emit the fall-through block.
1868 EmitBlock(LoopExit.getBlock());
1870 // Tell the runtime we are done.
1871 auto &&CodeGen = [DynamicOrOrdered, &S](CodeGenFunction &CGF) {
1872 if (!DynamicOrOrdered)
1873 CGF.CGM.getOpenMPRuntime().emitForStaticFinish(CGF, S.getEndLoc(),
1874 S.getDirectiveKind());
1876 OMPCancelStack.emitExit(*this, S.getDirectiveKind(), CodeGen);
1879 void CodeGenFunction::EmitOMPForOuterLoop(
1880 const OpenMPScheduleTy &ScheduleKind, bool IsMonotonic,
1881 const OMPLoopDirective &S, OMPPrivateScope &LoopScope, bool Ordered,
1882 const OMPLoopArguments &LoopArgs,
1883 const CodeGenDispatchBoundsTy &CGDispatchBounds) {
1884 CGOpenMPRuntime &RT = CGM.getOpenMPRuntime();
1886 // Dynamic scheduling of the outer loop (dynamic, guided, auto, runtime).
1887 const bool DynamicOrOrdered =
1888 Ordered || RT.isDynamic(ScheduleKind.Schedule);
1891 !RT.isStaticNonchunked(ScheduleKind.Schedule,
1892 LoopArgs.Chunk != nullptr)) &&
1893 "static non-chunked schedule does not need outer loop");
1897 // OpenMP [2.7.1, Loop Construct, Description, table 2-1]
1898 // When schedule(dynamic,chunk_size) is specified, the iterations are
1899 // distributed to threads in the team in chunks as the threads request them.
1900 // Each thread executes a chunk of iterations, then requests another chunk,
1901 // until no chunks remain to be distributed. Each chunk contains chunk_size
1902 // iterations, except for the last chunk to be distributed, which may have
1903 // fewer iterations. When no chunk_size is specified, it defaults to 1.
1905 // When schedule(guided,chunk_size) is specified, the iterations are assigned
1906 // to threads in the team in chunks as the executing threads request them.
1907 // Each thread executes a chunk of iterations, then requests another chunk,
1908 // until no chunks remain to be assigned. For a chunk_size of 1, the size of
1909 // each chunk is proportional to the number of unassigned iterations divided
1910 // by the number of threads in the team, decreasing to 1. For a chunk_size
1911 // with value k (greater than 1), the size of each chunk is determined in the
1912 // same way, with the restriction that the chunks do not contain fewer than k
1913 // iterations (except for the last chunk to be assigned, which may have fewer
1914 // than k iterations).
1916 // When schedule(auto) is specified, the decision regarding scheduling is
1917 // delegated to the compiler and/or runtime system. The programmer gives the
1918 // implementation the freedom to choose any possible mapping of iterations to
1919 // threads in the team.
1921 // When schedule(runtime) is specified, the decision regarding scheduling is
1922 // deferred until run time, and the schedule and chunk size are taken from the
1923 // run-sched-var ICV. If the ICV is set to auto, the schedule is
1924 // implementation defined
1926 // while(__kmpc_dispatch_next(&LB, &UB)) {
1928 // while (idx <= UB) { BODY; ++idx;
1929 // __kmpc_dispatch_fini_(4|8)[u](); // For ordered loops only.
1933 // OpenMP [2.7.1, Loop Construct, Description, table 2-1]
1934 // When schedule(static, chunk_size) is specified, iterations are divided into
1935 // chunks of size chunk_size, and the chunks are assigned to the threads in
1936 // the team in a round-robin fashion in the order of the thread number.
1938 // while(UB = min(UB, GlobalUB), idx = LB, idx < UB) {
1939 // while (idx <= UB) { BODY; ++idx; } // inner loop
1945 const Expr *IVExpr = S.getIterationVariable();
1946 const unsigned IVSize = getContext().getTypeSize(IVExpr->getType());
1947 const bool IVSigned = IVExpr->getType()->hasSignedIntegerRepresentation();
1949 if (DynamicOrOrdered) {
1950 const std::pair<llvm::Value *, llvm::Value *> DispatchBounds =
1951 CGDispatchBounds(*this, S, LoopArgs.LB, LoopArgs.UB);
1952 llvm::Value *LBVal = DispatchBounds.first;
1953 llvm::Value *UBVal = DispatchBounds.second;
1954 CGOpenMPRuntime::DispatchRTInput DipatchRTInputValues = {LBVal, UBVal,
1956 RT.emitForDispatchInit(*this, S.getBeginLoc(), ScheduleKind, IVSize,
1957 IVSigned, Ordered, DipatchRTInputValues);
1959 CGOpenMPRuntime::StaticRTInput StaticInit(
1960 IVSize, IVSigned, Ordered, LoopArgs.IL, LoopArgs.LB, LoopArgs.UB,
1961 LoopArgs.ST, LoopArgs.Chunk);
1962 RT.emitForStaticInit(*this, S.getBeginLoc(), S.getDirectiveKind(),
1963 ScheduleKind, StaticInit);
1966 auto &&CodeGenOrdered = [Ordered](CodeGenFunction &CGF, SourceLocation Loc,
1967 const unsigned IVSize,
1968 const bool IVSigned) {
1970 CGF.CGM.getOpenMPRuntime().emitForOrderedIterationEnd(CGF, Loc, IVSize,
1975 OMPLoopArguments OuterLoopArgs(LoopArgs.LB, LoopArgs.UB, LoopArgs.ST,
1976 LoopArgs.IL, LoopArgs.Chunk, LoopArgs.EUB);
1977 OuterLoopArgs.IncExpr = S.getInc();
1978 OuterLoopArgs.Init = S.getInit();
1979 OuterLoopArgs.Cond = S.getCond();
1980 OuterLoopArgs.NextLB = S.getNextLowerBound();
1981 OuterLoopArgs.NextUB = S.getNextUpperBound();
1982 EmitOMPOuterLoop(DynamicOrOrdered, IsMonotonic, S, LoopScope, OuterLoopArgs,
1983 emitOMPLoopBodyWithStopPoint, CodeGenOrdered);
1986 static void emitEmptyOrdered(CodeGenFunction &, SourceLocation Loc,
1987 const unsigned IVSize, const bool IVSigned) {}
1989 void CodeGenFunction::EmitOMPDistributeOuterLoop(
1990 OpenMPDistScheduleClauseKind ScheduleKind, const OMPLoopDirective &S,
1991 OMPPrivateScope &LoopScope, const OMPLoopArguments &LoopArgs,
1992 const CodeGenLoopTy &CodeGenLoopContent) {
1994 CGOpenMPRuntime &RT = CGM.getOpenMPRuntime();
1997 // Same behavior as a OMPForOuterLoop, except that schedule cannot be
2001 const Expr *IVExpr = S.getIterationVariable();
2002 const unsigned IVSize = getContext().getTypeSize(IVExpr->getType());
2003 const bool IVSigned = IVExpr->getType()->hasSignedIntegerRepresentation();
2005 CGOpenMPRuntime::StaticRTInput StaticInit(
2006 IVSize, IVSigned, /* Ordered = */ false, LoopArgs.IL, LoopArgs.LB,
2007 LoopArgs.UB, LoopArgs.ST, LoopArgs.Chunk);
2008 RT.emitDistributeStaticInit(*this, S.getBeginLoc(), ScheduleKind, StaticInit);
2010 // for combined 'distribute' and 'for' the increment expression of distribute
2011 // is stored in DistInc. For 'distribute' alone, it is in Inc.
2013 if (isOpenMPLoopBoundSharingDirective(S.getDirectiveKind()))
2014 IncExpr = S.getDistInc();
2016 IncExpr = S.getInc();
2018 // this routine is shared by 'omp distribute parallel for' and
2019 // 'omp distribute': select the right EUB expression depending on the
2021 OMPLoopArguments OuterLoopArgs;
2022 OuterLoopArgs.LB = LoopArgs.LB;
2023 OuterLoopArgs.UB = LoopArgs.UB;
2024 OuterLoopArgs.ST = LoopArgs.ST;
2025 OuterLoopArgs.IL = LoopArgs.IL;
2026 OuterLoopArgs.Chunk = LoopArgs.Chunk;
2027 OuterLoopArgs.EUB = isOpenMPLoopBoundSharingDirective(S.getDirectiveKind())
2028 ? S.getCombinedEnsureUpperBound()
2029 : S.getEnsureUpperBound();
2030 OuterLoopArgs.IncExpr = IncExpr;
2031 OuterLoopArgs.Init = isOpenMPLoopBoundSharingDirective(S.getDirectiveKind())
2032 ? S.getCombinedInit()
2034 OuterLoopArgs.Cond = isOpenMPLoopBoundSharingDirective(S.getDirectiveKind())
2035 ? S.getCombinedCond()
2037 OuterLoopArgs.NextLB = isOpenMPLoopBoundSharingDirective(S.getDirectiveKind())
2038 ? S.getCombinedNextLowerBound()
2039 : S.getNextLowerBound();
2040 OuterLoopArgs.NextUB = isOpenMPLoopBoundSharingDirective(S.getDirectiveKind())
2041 ? S.getCombinedNextUpperBound()
2042 : S.getNextUpperBound();
2044 EmitOMPOuterLoop(/* DynamicOrOrdered = */ false, /* IsMonotonic = */ false, S,
2045 LoopScope, OuterLoopArgs, CodeGenLoopContent,
2049 static std::pair<LValue, LValue>
2050 emitDistributeParallelForInnerBounds(CodeGenFunction &CGF,
2051 const OMPExecutableDirective &S) {
2052 const OMPLoopDirective &LS = cast<OMPLoopDirective>(S);
2054 EmitOMPHelperVar(CGF, cast<DeclRefExpr>(LS.getLowerBoundVariable()));
2056 EmitOMPHelperVar(CGF, cast<DeclRefExpr>(LS.getUpperBoundVariable()));
2058 // When composing 'distribute' with 'for' (e.g. as in 'distribute
2059 // parallel for') we need to use the 'distribute'
2060 // chunk lower and upper bounds rather than the whole loop iteration
2061 // space. These are parameters to the outlined function for 'parallel'
2062 // and we copy the bounds of the previous schedule into the
2063 // the current ones.
2064 LValue PrevLB = CGF.EmitLValue(LS.getPrevLowerBoundVariable());
2065 LValue PrevUB = CGF.EmitLValue(LS.getPrevUpperBoundVariable());
2066 llvm::Value *PrevLBVal = CGF.EmitLoadOfScalar(
2067 PrevLB, LS.getPrevLowerBoundVariable()->getExprLoc());
2068 PrevLBVal = CGF.EmitScalarConversion(
2069 PrevLBVal, LS.getPrevLowerBoundVariable()->getType(),
2070 LS.getIterationVariable()->getType(),
2071 LS.getPrevLowerBoundVariable()->getExprLoc());
2072 llvm::Value *PrevUBVal = CGF.EmitLoadOfScalar(
2073 PrevUB, LS.getPrevUpperBoundVariable()->getExprLoc());
2074 PrevUBVal = CGF.EmitScalarConversion(
2075 PrevUBVal, LS.getPrevUpperBoundVariable()->getType(),
2076 LS.getIterationVariable()->getType(),
2077 LS.getPrevUpperBoundVariable()->getExprLoc());
2079 CGF.EmitStoreOfScalar(PrevLBVal, LB);
2080 CGF.EmitStoreOfScalar(PrevUBVal, UB);
2085 /// if the 'for' loop has a dispatch schedule (e.g. dynamic, guided) then
2086 /// we need to use the LB and UB expressions generated by the worksharing
2087 /// code generation support, whereas in non combined situations we would
2088 /// just emit 0 and the LastIteration expression
2089 /// This function is necessary due to the difference of the LB and UB
2090 /// types for the RT emission routines for 'for_static_init' and
2091 /// 'for_dispatch_init'
2092 static std::pair<llvm::Value *, llvm::Value *>
2093 emitDistributeParallelForDispatchBounds(CodeGenFunction &CGF,
2094 const OMPExecutableDirective &S,
2095 Address LB, Address UB) {
2096 const OMPLoopDirective &LS = cast<OMPLoopDirective>(S);
2097 const Expr *IVExpr = LS.getIterationVariable();
2098 // when implementing a dynamic schedule for a 'for' combined with a
2099 // 'distribute' (e.g. 'distribute parallel for'), the 'for' loop
2100 // is not normalized as each team only executes its own assigned
2102 QualType IteratorTy = IVExpr->getType();
2103 llvm::Value *LBVal =
2104 CGF.EmitLoadOfScalar(LB, /*Volatile=*/false, IteratorTy, S.getBeginLoc());
2105 llvm::Value *UBVal =
2106 CGF.EmitLoadOfScalar(UB, /*Volatile=*/false, IteratorTy, S.getBeginLoc());
2107 return {LBVal, UBVal};
2110 static void emitDistributeParallelForDistributeInnerBoundParams(
2111 CodeGenFunction &CGF, const OMPExecutableDirective &S,
2112 llvm::SmallVectorImpl<llvm::Value *> &CapturedVars) {
2113 const auto &Dir = cast<OMPLoopDirective>(S);
2115 CGF.EmitLValue(cast<DeclRefExpr>(Dir.getCombinedLowerBoundVariable()));
2116 llvm::Value *LBCast = CGF.Builder.CreateIntCast(
2117 CGF.Builder.CreateLoad(LB.getAddress()), CGF.SizeTy, /*isSigned=*/false);
2118 CapturedVars.push_back(LBCast);
2120 CGF.EmitLValue(cast<DeclRefExpr>(Dir.getCombinedUpperBoundVariable()));
2122 llvm::Value *UBCast = CGF.Builder.CreateIntCast(
2123 CGF.Builder.CreateLoad(UB.getAddress()), CGF.SizeTy, /*isSigned=*/false);
2124 CapturedVars.push_back(UBCast);
2128 emitInnerParallelForWhenCombined(CodeGenFunction &CGF,
2129 const OMPLoopDirective &S,
2130 CodeGenFunction::JumpDest LoopExit) {
2131 auto &&CGInlinedWorksharingLoop = [&S](CodeGenFunction &CGF,
2132 PrePostActionTy &Action) {
2134 bool HasCancel = false;
2135 if (!isOpenMPSimdDirective(S.getDirectiveKind())) {
2136 if (const auto *D = dyn_cast<OMPTeamsDistributeParallelForDirective>(&S))
2137 HasCancel = D->hasCancel();
2138 else if (const auto *D = dyn_cast<OMPDistributeParallelForDirective>(&S))
2139 HasCancel = D->hasCancel();
2140 else if (const auto *D =
2141 dyn_cast<OMPTargetTeamsDistributeParallelForDirective>(&S))
2142 HasCancel = D->hasCancel();
2144 CodeGenFunction::OMPCancelStackRAII CancelRegion(CGF, S.getDirectiveKind(),
2146 CGF.EmitOMPWorksharingLoop(S, S.getPrevEnsureUpperBound(),
2147 emitDistributeParallelForInnerBounds,
2148 emitDistributeParallelForDispatchBounds);
2151 emitCommonOMPParallelDirective(
2153 isOpenMPSimdDirective(S.getDirectiveKind()) ? OMPD_for_simd : OMPD_for,
2154 CGInlinedWorksharingLoop,
2155 emitDistributeParallelForDistributeInnerBoundParams);
2158 void CodeGenFunction::EmitOMPDistributeParallelForDirective(
2159 const OMPDistributeParallelForDirective &S) {
2160 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &) {
2161 CGF.EmitOMPDistributeLoop(S, emitInnerParallelForWhenCombined,
2164 OMPLexicalScope Scope(*this, S, OMPD_parallel);
2165 CGM.getOpenMPRuntime().emitInlinedDirective(*this, OMPD_distribute, CodeGen);
2168 void CodeGenFunction::EmitOMPDistributeParallelForSimdDirective(
2169 const OMPDistributeParallelForSimdDirective &S) {
2170 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &) {
2171 CGF.EmitOMPDistributeLoop(S, emitInnerParallelForWhenCombined,
2174 OMPLexicalScope Scope(*this, S, OMPD_parallel);
2175 CGM.getOpenMPRuntime().emitInlinedDirective(*this, OMPD_distribute, CodeGen);
2178 void CodeGenFunction::EmitOMPDistributeSimdDirective(
2179 const OMPDistributeSimdDirective &S) {
2180 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &) {
2181 CGF.EmitOMPDistributeLoop(S, emitOMPLoopBodyWithStopPoint, S.getInc());
2183 OMPLexicalScope Scope(*this, S, OMPD_unknown);
2184 CGM.getOpenMPRuntime().emitInlinedDirective(*this, OMPD_simd, CodeGen);
2187 void CodeGenFunction::EmitOMPTargetSimdDeviceFunction(
2188 CodeGenModule &CGM, StringRef ParentName, const OMPTargetSimdDirective &S) {
2189 // Emit SPMD target parallel for region as a standalone region.
2190 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
2191 emitOMPSimdRegion(CGF, S, Action);
2194 llvm::Constant *Addr;
2195 // Emit target region as a standalone region.
2196 CGM.getOpenMPRuntime().emitTargetOutlinedFunction(
2197 S, ParentName, Fn, Addr, /*IsOffloadEntry=*/true, CodeGen);
2198 assert(Fn && Addr && "Target device function emission failed.");
2201 void CodeGenFunction::EmitOMPTargetSimdDirective(
2202 const OMPTargetSimdDirective &S) {
2203 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
2204 emitOMPSimdRegion(CGF, S, Action);
2206 emitCommonOMPTargetDirective(*this, S, CodeGen);
2210 struct ScheduleKindModifiersTy {
2211 OpenMPScheduleClauseKind Kind;
2212 OpenMPScheduleClauseModifier M1;
2213 OpenMPScheduleClauseModifier M2;
2214 ScheduleKindModifiersTy(OpenMPScheduleClauseKind Kind,
2215 OpenMPScheduleClauseModifier M1,
2216 OpenMPScheduleClauseModifier M2)
2217 : Kind(Kind), M1(M1), M2(M2) {}
2221 bool CodeGenFunction::EmitOMPWorksharingLoop(
2222 const OMPLoopDirective &S, Expr *EUB,
2223 const CodeGenLoopBoundsTy &CodeGenLoopBounds,
2224 const CodeGenDispatchBoundsTy &CGDispatchBounds) {
2225 // Emit the loop iteration variable.
2226 const auto *IVExpr = cast<DeclRefExpr>(S.getIterationVariable());
2227 const auto *IVDecl = cast<VarDecl>(IVExpr->getDecl());
2228 EmitVarDecl(*IVDecl);
2230 // Emit the iterations count variable.
2231 // If it is not a variable, Sema decided to calculate iterations count on each
2232 // iteration (e.g., it is foldable into a constant).
2233 if (const auto *LIExpr = dyn_cast<DeclRefExpr>(S.getLastIteration())) {
2234 EmitVarDecl(*cast<VarDecl>(LIExpr->getDecl()));
2235 // Emit calculation of the iterations count.
2236 EmitIgnoredExpr(S.getCalcLastIteration());
2239 CGOpenMPRuntime &RT = CGM.getOpenMPRuntime();
2241 bool HasLastprivateClause;
2242 // Check pre-condition.
2244 OMPLoopScope PreInitScope(*this, S);
2245 // Skip the entire loop if we don't meet the precondition.
2246 // If the condition constant folds and can be elided, avoid emitting the
2249 llvm::BasicBlock *ContBlock = nullptr;
2250 if (ConstantFoldsToSimpleInteger(S.getPreCond(), CondConstant)) {
2254 llvm::BasicBlock *ThenBlock = createBasicBlock("omp.precond.then");
2255 ContBlock = createBasicBlock("omp.precond.end");
2256 emitPreCond(*this, S, S.getPreCond(), ThenBlock, ContBlock,
2257 getProfileCount(&S));
2258 EmitBlock(ThenBlock);
2259 incrementProfileCounter(&S);
2262 RunCleanupsScope DoacrossCleanupScope(*this);
2263 bool Ordered = false;
2264 if (const auto *OrderedClause = S.getSingleClause<OMPOrderedClause>()) {
2265 if (OrderedClause->getNumForLoops())
2266 RT.emitDoacrossInit(*this, S, OrderedClause->getLoopNumIterations());
2271 llvm::DenseSet<const Expr *> EmittedFinals;
2272 emitAlignedClause(*this, S);
2273 bool HasLinears = EmitOMPLinearClauseInit(S);
2274 // Emit helper vars inits.
2276 std::pair<LValue, LValue> Bounds = CodeGenLoopBounds(*this, S);
2277 LValue LB = Bounds.first;
2278 LValue UB = Bounds.second;
2280 EmitOMPHelperVar(*this, cast<DeclRefExpr>(S.getStrideVariable()));
2282 EmitOMPHelperVar(*this, cast<DeclRefExpr>(S.getIsLastIterVariable()));
2284 // Emit 'then' code.
2286 OMPPrivateScope LoopScope(*this);
2287 if (EmitOMPFirstprivateClause(S, LoopScope) || HasLinears) {
2288 // Emit implicit barrier to synchronize threads and avoid data races on
2289 // initialization of firstprivate variables and post-update of
2290 // lastprivate variables.
2291 CGM.getOpenMPRuntime().emitBarrierCall(
2292 *this, S.getBeginLoc(), OMPD_unknown, /*EmitChecks=*/false,
2293 /*ForceSimpleCall=*/true);
2295 EmitOMPPrivateClause(S, LoopScope);
2296 HasLastprivateClause = EmitOMPLastprivateClauseInit(S, LoopScope);
2297 EmitOMPReductionClauseInit(S, LoopScope);
2298 EmitOMPPrivateLoopCounters(S, LoopScope);
2299 EmitOMPLinearClause(S, LoopScope);
2300 (void)LoopScope.Privatize();
2301 if (isOpenMPTargetExecutionDirective(S.getDirectiveKind()))
2302 CGM.getOpenMPRuntime().adjustTargetSpecificDataForLambdas(*this, S);
2304 // Detect the loop schedule kind and chunk.
2305 const Expr *ChunkExpr = nullptr;
2306 OpenMPScheduleTy ScheduleKind;
2307 if (const auto *C = S.getSingleClause<OMPScheduleClause>()) {
2308 ScheduleKind.Schedule = C->getScheduleKind();
2309 ScheduleKind.M1 = C->getFirstScheduleModifier();
2310 ScheduleKind.M2 = C->getSecondScheduleModifier();
2311 ChunkExpr = C->getChunkSize();
2313 // Default behaviour for schedule clause.
2314 CGM.getOpenMPRuntime().getDefaultScheduleAndChunk(
2315 *this, S, ScheduleKind.Schedule, ChunkExpr);
2317 bool HasChunkSizeOne = false;
2318 llvm::Value *Chunk = nullptr;
2320 Chunk = EmitScalarExpr(ChunkExpr);
2321 Chunk = EmitScalarConversion(Chunk, ChunkExpr->getType(),
2322 S.getIterationVariable()->getType(),
2324 Expr::EvalResult Result;
2325 if (ChunkExpr->EvaluateAsInt(Result, getContext())) {
2326 llvm::APSInt EvaluatedChunk = Result.Val.getInt();
2327 HasChunkSizeOne = (EvaluatedChunk.getLimitedValue() == 1);
2330 const unsigned IVSize = getContext().getTypeSize(IVExpr->getType());
2331 const bool IVSigned = IVExpr->getType()->hasSignedIntegerRepresentation();
2332 // OpenMP 4.5, 2.7.1 Loop Construct, Description.
2333 // If the static schedule kind is specified or if the ordered clause is
2334 // specified, and if no monotonic modifier is specified, the effect will
2335 // be as if the monotonic modifier was specified.
2336 bool StaticChunkedOne = RT.isStaticChunked(ScheduleKind.Schedule,
2337 /* Chunked */ Chunk != nullptr) && HasChunkSizeOne &&
2338 isOpenMPLoopBoundSharingDirective(S.getDirectiveKind());
2339 if ((RT.isStaticNonchunked(ScheduleKind.Schedule,
2340 /* Chunked */ Chunk != nullptr) ||
2341 StaticChunkedOne) &&
2343 if (isOpenMPSimdDirective(S.getDirectiveKind()))
2344 EmitOMPSimdInit(S, /*IsMonotonic=*/true);
2345 // OpenMP [2.7.1, Loop Construct, Description, table 2-1]
2346 // When no chunk_size is specified, the iteration space is divided into
2347 // chunks that are approximately equal in size, and at most one chunk is
2348 // distributed to each thread. Note that the size of the chunks is
2349 // unspecified in this case.
2350 CGOpenMPRuntime::StaticRTInput StaticInit(
2351 IVSize, IVSigned, Ordered, IL.getAddress(), LB.getAddress(),
2352 UB.getAddress(), ST.getAddress(),
2353 StaticChunkedOne ? Chunk : nullptr);
2354 RT.emitForStaticInit(*this, S.getBeginLoc(), S.getDirectiveKind(),
2355 ScheduleKind, StaticInit);
2357 getJumpDestInCurrentScope(createBasicBlock("omp.loop.exit"));
2358 // UB = min(UB, GlobalUB);
2359 if (!StaticChunkedOne)
2360 EmitIgnoredExpr(S.getEnsureUpperBound());
2362 EmitIgnoredExpr(S.getInit());
2363 // For unchunked static schedule generate:
2365 // while (idx <= UB) {
2370 // For static schedule with chunk one:
2372 // while (IV <= PrevUB) {
2376 EmitOMPInnerLoop(S, LoopScope.requiresCleanups(),
2377 StaticChunkedOne ? S.getCombinedParForInDistCond() : S.getCond(),
2378 StaticChunkedOne ? S.getDistInc() : S.getInc(),
2379 [&S, LoopExit](CodeGenFunction &CGF) {
2380 CGF.EmitOMPLoopBody(S, LoopExit);
2381 CGF.EmitStopPoint(&S);
2383 [](CodeGenFunction &) {});
2384 EmitBlock(LoopExit.getBlock());
2385 // Tell the runtime we are done.
2386 auto &&CodeGen = [&S](CodeGenFunction &CGF) {
2387 CGF.CGM.getOpenMPRuntime().emitForStaticFinish(CGF, S.getEndLoc(),
2388 S.getDirectiveKind());
2390 OMPCancelStack.emitExit(*this, S.getDirectiveKind(), CodeGen);
2392 const bool IsMonotonic =
2393 Ordered || ScheduleKind.Schedule == OMPC_SCHEDULE_static ||
2394 ScheduleKind.Schedule == OMPC_SCHEDULE_unknown ||
2395 ScheduleKind.M1 == OMPC_SCHEDULE_MODIFIER_monotonic ||
2396 ScheduleKind.M2 == OMPC_SCHEDULE_MODIFIER_monotonic;
2397 // Emit the outer loop, which requests its work chunk [LB..UB] from
2398 // runtime and runs the inner loop to process it.
2399 const OMPLoopArguments LoopArguments(LB.getAddress(), UB.getAddress(),
2400 ST.getAddress(), IL.getAddress(),
2402 EmitOMPForOuterLoop(ScheduleKind, IsMonotonic, S, LoopScope, Ordered,
2403 LoopArguments, CGDispatchBounds);
2405 if (isOpenMPSimdDirective(S.getDirectiveKind())) {
2406 EmitOMPSimdFinal(S, [IL, &S](CodeGenFunction &CGF) {
2407 return CGF.Builder.CreateIsNotNull(
2408 CGF.EmitLoadOfScalar(IL, S.getBeginLoc()));
2411 EmitOMPReductionClauseFinal(
2412 S, /*ReductionKind=*/isOpenMPSimdDirective(S.getDirectiveKind())
2413 ? /*Parallel and Simd*/ OMPD_parallel_for_simd
2414 : /*Parallel only*/ OMPD_parallel);
2415 // Emit post-update of the reduction variables if IsLastIter != 0.
2416 emitPostUpdateForReductionClause(
2417 *this, S, [IL, &S](CodeGenFunction &CGF) {
2418 return CGF.Builder.CreateIsNotNull(
2419 CGF.EmitLoadOfScalar(IL, S.getBeginLoc()));
2421 // Emit final copy of the lastprivate variables if IsLastIter != 0.
2422 if (HasLastprivateClause)
2423 EmitOMPLastprivateClauseFinal(
2424 S, isOpenMPSimdDirective(S.getDirectiveKind()),
2425 Builder.CreateIsNotNull(EmitLoadOfScalar(IL, S.getBeginLoc())));
2427 EmitOMPLinearClauseFinal(S, [IL, &S](CodeGenFunction &CGF) {
2428 return CGF.Builder.CreateIsNotNull(
2429 CGF.EmitLoadOfScalar(IL, S.getBeginLoc()));
2431 DoacrossCleanupScope.ForceCleanup();
2432 // We're now done with the loop, so jump to the continuation block.
2434 EmitBranch(ContBlock);
2435 EmitBlock(ContBlock, /*IsFinished=*/true);
2438 return HasLastprivateClause;
2441 /// The following two functions generate expressions for the loop lower
2442 /// and upper bounds in case of static and dynamic (dispatch) schedule
2443 /// of the associated 'for' or 'distribute' loop.
2444 static std::pair<LValue, LValue>
2445 emitForLoopBounds(CodeGenFunction &CGF, const OMPExecutableDirective &S) {
2446 const auto &LS = cast<OMPLoopDirective>(S);
2448 EmitOMPHelperVar(CGF, cast<DeclRefExpr>(LS.getLowerBoundVariable()));
2450 EmitOMPHelperVar(CGF, cast<DeclRefExpr>(LS.getUpperBoundVariable()));
2454 /// When dealing with dispatch schedules (e.g. dynamic, guided) we do not
2455 /// consider the lower and upper bound expressions generated by the
2456 /// worksharing loop support, but we use 0 and the iteration space size as
2458 static std::pair<llvm::Value *, llvm::Value *>
2459 emitDispatchForLoopBounds(CodeGenFunction &CGF, const OMPExecutableDirective &S,
2460 Address LB, Address UB) {
2461 const auto &LS = cast<OMPLoopDirective>(S);
2462 const Expr *IVExpr = LS.getIterationVariable();
2463 const unsigned IVSize = CGF.getContext().getTypeSize(IVExpr->getType());
2464 llvm::Value *LBVal = CGF.Builder.getIntN(IVSize, 0);
2465 llvm::Value *UBVal = CGF.EmitScalarExpr(LS.getLastIteration());
2466 return {LBVal, UBVal};
2469 void CodeGenFunction::EmitOMPForDirective(const OMPForDirective &S) {
2470 bool HasLastprivates = false;
2471 auto &&CodeGen = [&S, &HasLastprivates](CodeGenFunction &CGF,
2472 PrePostActionTy &) {
2473 OMPCancelStackRAII CancelRegion(CGF, OMPD_for, S.hasCancel());
2474 HasLastprivates = CGF.EmitOMPWorksharingLoop(S, S.getEnsureUpperBound(),
2476 emitDispatchForLoopBounds);
2479 OMPLexicalScope Scope(*this, S, OMPD_unknown);
2480 CGM.getOpenMPRuntime().emitInlinedDirective(*this, OMPD_for, CodeGen,
2484 // Emit an implicit barrier at the end.
2485 if (!S.getSingleClause<OMPNowaitClause>() || HasLastprivates)
2486 CGM.getOpenMPRuntime().emitBarrierCall(*this, S.getBeginLoc(), OMPD_for);
2489 void CodeGenFunction::EmitOMPForSimdDirective(const OMPForSimdDirective &S) {
2490 bool HasLastprivates = false;
2491 auto &&CodeGen = [&S, &HasLastprivates](CodeGenFunction &CGF,
2492 PrePostActionTy &) {
2493 HasLastprivates = CGF.EmitOMPWorksharingLoop(S, S.getEnsureUpperBound(),
2495 emitDispatchForLoopBounds);
2498 OMPLexicalScope Scope(*this, S, OMPD_unknown);
2499 CGM.getOpenMPRuntime().emitInlinedDirective(*this, OMPD_simd, CodeGen);
2502 // Emit an implicit barrier at the end.
2503 if (!S.getSingleClause<OMPNowaitClause>() || HasLastprivates)
2504 CGM.getOpenMPRuntime().emitBarrierCall(*this, S.getBeginLoc(), OMPD_for);
2507 static LValue createSectionLVal(CodeGenFunction &CGF, QualType Ty,
2509 llvm::Value *Init = nullptr) {
2510 LValue LVal = CGF.MakeAddrLValue(CGF.CreateMemTemp(Ty, Name), Ty);
2512 CGF.EmitStoreThroughLValue(RValue::get(Init), LVal, /*isInit*/ true);
2516 void CodeGenFunction::EmitSections(const OMPExecutableDirective &S) {
2517 const Stmt *CapturedStmt = S.getInnermostCapturedStmt()->getCapturedStmt();
2518 const auto *CS = dyn_cast<CompoundStmt>(CapturedStmt);
2519 bool HasLastprivates = false;
2520 auto &&CodeGen = [&S, CapturedStmt, CS,
2521 &HasLastprivates](CodeGenFunction &CGF, PrePostActionTy &) {
2522 ASTContext &C = CGF.getContext();
2523 QualType KmpInt32Ty =
2524 C.getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/1);
2525 // Emit helper vars inits.
2526 LValue LB = createSectionLVal(CGF, KmpInt32Ty, ".omp.sections.lb.",
2527 CGF.Builder.getInt32(0));
2528 llvm::ConstantInt *GlobalUBVal = CS != nullptr
2529 ? CGF.Builder.getInt32(CS->size() - 1)
2530 : CGF.Builder.getInt32(0);
2532 createSectionLVal(CGF, KmpInt32Ty, ".omp.sections.ub.", GlobalUBVal);
2533 LValue ST = createSectionLVal(CGF, KmpInt32Ty, ".omp.sections.st.",
2534 CGF.Builder.getInt32(1));
2535 LValue IL = createSectionLVal(CGF, KmpInt32Ty, ".omp.sections.il.",
2536 CGF.Builder.getInt32(0));
2538 LValue IV = createSectionLVal(CGF, KmpInt32Ty, ".omp.sections.iv.");
2539 OpaqueValueExpr IVRefExpr(S.getBeginLoc(), KmpInt32Ty, VK_LValue);
2540 CodeGenFunction::OpaqueValueMapping OpaqueIV(CGF, &IVRefExpr, IV);
2541 OpaqueValueExpr UBRefExpr(S.getBeginLoc(), KmpInt32Ty, VK_LValue);
2542 CodeGenFunction::OpaqueValueMapping OpaqueUB(CGF, &UBRefExpr, UB);
2543 // Generate condition for loop.
2544 BinaryOperator Cond(&IVRefExpr, &UBRefExpr, BO_LE, C.BoolTy, VK_RValue,
2545 OK_Ordinary, S.getBeginLoc(), FPOptions());
2546 // Increment for loop counter.
2547 UnaryOperator Inc(&IVRefExpr, UO_PreInc, KmpInt32Ty, VK_RValue, OK_Ordinary,
2548 S.getBeginLoc(), true);
2549 auto &&BodyGen = [CapturedStmt, CS, &S, &IV](CodeGenFunction &CGF) {
2550 // Iterate through all sections and emit a switch construct:
2553 // <SectionStmt[0]>;
2556 // case <NumSection> - 1:
2557 // <SectionStmt[<NumSection> - 1]>;
2560 // .omp.sections.exit:
2561 llvm::BasicBlock *ExitBB = CGF.createBasicBlock(".omp.sections.exit");
2562 llvm::SwitchInst *SwitchStmt =
2563 CGF.Builder.CreateSwitch(CGF.EmitLoadOfScalar(IV, S.getBeginLoc()),
2564 ExitBB, CS == nullptr ? 1 : CS->size());
2566 unsigned CaseNumber = 0;
2567 for (const Stmt *SubStmt : CS->children()) {
2568 auto CaseBB = CGF.createBasicBlock(".omp.sections.case");
2569 CGF.EmitBlock(CaseBB);
2570 SwitchStmt->addCase(CGF.Builder.getInt32(CaseNumber), CaseBB);
2571 CGF.EmitStmt(SubStmt);
2572 CGF.EmitBranch(ExitBB);
2576 llvm::BasicBlock *CaseBB = CGF.createBasicBlock(".omp.sections.case");
2577 CGF.EmitBlock(CaseBB);
2578 SwitchStmt->addCase(CGF.Builder.getInt32(0), CaseBB);
2579 CGF.EmitStmt(CapturedStmt);
2580 CGF.EmitBranch(ExitBB);
2582 CGF.EmitBlock(ExitBB, /*IsFinished=*/true);
2585 CodeGenFunction::OMPPrivateScope LoopScope(CGF);
2586 if (CGF.EmitOMPFirstprivateClause(S, LoopScope)) {
2587 // Emit implicit barrier to synchronize threads and avoid data races on
2588 // initialization of firstprivate variables and post-update of lastprivate
2590 CGF.CGM.getOpenMPRuntime().emitBarrierCall(
2591 CGF, S.getBeginLoc(), OMPD_unknown, /*EmitChecks=*/false,
2592 /*ForceSimpleCall=*/true);
2594 CGF.EmitOMPPrivateClause(S, LoopScope);
2595 HasLastprivates = CGF.EmitOMPLastprivateClauseInit(S, LoopScope);
2596 CGF.EmitOMPReductionClauseInit(S, LoopScope);
2597 (void)LoopScope.Privatize();
2598 if (isOpenMPTargetExecutionDirective(S.getDirectiveKind()))
2599 CGF.CGM.getOpenMPRuntime().adjustTargetSpecificDataForLambdas(CGF, S);
2601 // Emit static non-chunked loop.
2602 OpenMPScheduleTy ScheduleKind;
2603 ScheduleKind.Schedule = OMPC_SCHEDULE_static;
2604 CGOpenMPRuntime::StaticRTInput StaticInit(
2605 /*IVSize=*/32, /*IVSigned=*/true, /*Ordered=*/false, IL.getAddress(),
2606 LB.getAddress(), UB.getAddress(), ST.getAddress());
2607 CGF.CGM.getOpenMPRuntime().emitForStaticInit(
2608 CGF, S.getBeginLoc(), S.getDirectiveKind(), ScheduleKind, StaticInit);
2609 // UB = min(UB, GlobalUB);
2610 llvm::Value *UBVal = CGF.EmitLoadOfScalar(UB, S.getBeginLoc());
2611 llvm::Value *MinUBGlobalUB = CGF.Builder.CreateSelect(
2612 CGF.Builder.CreateICmpSLT(UBVal, GlobalUBVal), UBVal, GlobalUBVal);
2613 CGF.EmitStoreOfScalar(MinUBGlobalUB, UB);
2615 CGF.EmitStoreOfScalar(CGF.EmitLoadOfScalar(LB, S.getBeginLoc()), IV);
2616 // while (idx <= UB) { BODY; ++idx; }
2617 CGF.EmitOMPInnerLoop(S, /*RequiresCleanup=*/false, &Cond, &Inc, BodyGen,
2618 [](CodeGenFunction &) {});
2619 // Tell the runtime we are done.
2620 auto &&CodeGen = [&S](CodeGenFunction &CGF) {
2621 CGF.CGM.getOpenMPRuntime().emitForStaticFinish(CGF, S.getEndLoc(),
2622 S.getDirectiveKind());
2624 CGF.OMPCancelStack.emitExit(CGF, S.getDirectiveKind(), CodeGen);
2625 CGF.EmitOMPReductionClauseFinal(S, /*ReductionKind=*/OMPD_parallel);
2626 // Emit post-update of the reduction variables if IsLastIter != 0.
2627 emitPostUpdateForReductionClause(CGF, S, [IL, &S](CodeGenFunction &CGF) {
2628 return CGF.Builder.CreateIsNotNull(
2629 CGF.EmitLoadOfScalar(IL, S.getBeginLoc()));
2632 // Emit final copy of the lastprivate variables if IsLastIter != 0.
2633 if (HasLastprivates)
2634 CGF.EmitOMPLastprivateClauseFinal(
2635 S, /*NoFinals=*/false,
2636 CGF.Builder.CreateIsNotNull(
2637 CGF.EmitLoadOfScalar(IL, S.getBeginLoc())));
2640 bool HasCancel = false;
2641 if (auto *OSD = dyn_cast<OMPSectionsDirective>(&S))
2642 HasCancel = OSD->hasCancel();
2643 else if (auto *OPSD = dyn_cast<OMPParallelSectionsDirective>(&S))
2644 HasCancel = OPSD->hasCancel();
2645 OMPCancelStackRAII CancelRegion(*this, S.getDirectiveKind(), HasCancel);
2646 CGM.getOpenMPRuntime().emitInlinedDirective(*this, OMPD_sections, CodeGen,
2648 // Emit barrier for lastprivates only if 'sections' directive has 'nowait'
2649 // clause. Otherwise the barrier will be generated by the codegen for the
2651 if (HasLastprivates && S.getSingleClause<OMPNowaitClause>()) {
2652 // Emit implicit barrier to synchronize threads and avoid data races on
2653 // initialization of firstprivate variables.
2654 CGM.getOpenMPRuntime().emitBarrierCall(*this, S.getBeginLoc(),
2659 void CodeGenFunction::EmitOMPSectionsDirective(const OMPSectionsDirective &S) {
2661 OMPLexicalScope Scope(*this, S, OMPD_unknown);
2664 // Emit an implicit barrier at the end.
2665 if (!S.getSingleClause<OMPNowaitClause>()) {
2666 CGM.getOpenMPRuntime().emitBarrierCall(*this, S.getBeginLoc(),
2671 void CodeGenFunction::EmitOMPSectionDirective(const OMPSectionDirective &S) {
2672 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &) {
2673 CGF.EmitStmt(S.getInnermostCapturedStmt()->getCapturedStmt());
2675 OMPLexicalScope Scope(*this, S, OMPD_unknown);
2676 CGM.getOpenMPRuntime().emitInlinedDirective(*this, OMPD_section, CodeGen,
2680 void CodeGenFunction::EmitOMPSingleDirective(const OMPSingleDirective &S) {
2681 llvm::SmallVector<const Expr *, 8> CopyprivateVars;
2682 llvm::SmallVector<const Expr *, 8> DestExprs;
2683 llvm::SmallVector<const Expr *, 8> SrcExprs;
2684 llvm::SmallVector<const Expr *, 8> AssignmentOps;
2685 // Check if there are any 'copyprivate' clauses associated with this
2686 // 'single' construct.
2687 // Build a list of copyprivate variables along with helper expressions
2688 // (<source>, <destination>, <destination>=<source> expressions)
2689 for (const auto *C : S.getClausesOfKind<OMPCopyprivateClause>()) {
2690 CopyprivateVars.append(C->varlists().begin(), C->varlists().end());
2691 DestExprs.append(C->destination_exprs().begin(),
2692 C->destination_exprs().end());
2693 SrcExprs.append(C->source_exprs().begin(), C->source_exprs().end());
2694 AssignmentOps.append(C->assignment_ops().begin(),
2695 C->assignment_ops().end());
2697 // Emit code for 'single' region along with 'copyprivate' clauses
2698 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
2700 OMPPrivateScope SingleScope(CGF);
2701 (void)CGF.EmitOMPFirstprivateClause(S, SingleScope);
2702 CGF.EmitOMPPrivateClause(S, SingleScope);
2703 (void)SingleScope.Privatize();
2704 CGF.EmitStmt(S.getInnermostCapturedStmt()->getCapturedStmt());
2707 OMPLexicalScope Scope(*this, S, OMPD_unknown);
2708 CGM.getOpenMPRuntime().emitSingleRegion(*this, CodeGen, S.getBeginLoc(),
2709 CopyprivateVars, DestExprs,
2710 SrcExprs, AssignmentOps);
2712 // Emit an implicit barrier at the end (to avoid data race on firstprivate
2713 // init or if no 'nowait' clause was specified and no 'copyprivate' clause).
2714 if (!S.getSingleClause<OMPNowaitClause>() && CopyprivateVars.empty()) {
2715 CGM.getOpenMPRuntime().emitBarrierCall(
2716 *this, S.getBeginLoc(),
2717 S.getSingleClause<OMPNowaitClause>() ? OMPD_unknown : OMPD_single);
2721 void CodeGenFunction::EmitOMPMasterDirective(const OMPMasterDirective &S) {
2722 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
2724 CGF.EmitStmt(S.getInnermostCapturedStmt()->getCapturedStmt());
2726 OMPLexicalScope Scope(*this, S, OMPD_unknown);
2727 CGM.getOpenMPRuntime().emitMasterRegion(*this, CodeGen, S.getBeginLoc());
2730 void CodeGenFunction::EmitOMPCriticalDirective(const OMPCriticalDirective &S) {
2731 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
2733 CGF.EmitStmt(S.getInnermostCapturedStmt()->getCapturedStmt());
2735 const Expr *Hint = nullptr;
2736 if (const auto *HintClause = S.getSingleClause<OMPHintClause>())
2737 Hint = HintClause->getHint();
2738 OMPLexicalScope Scope(*this, S, OMPD_unknown);
2739 CGM.getOpenMPRuntime().emitCriticalRegion(*this,
2740 S.getDirectiveName().getAsString(),
2741 CodeGen, S.getBeginLoc(), Hint);
2744 void CodeGenFunction::EmitOMPParallelForDirective(
2745 const OMPParallelForDirective &S) {
2746 // Emit directive as a combined directive that consists of two implicit
2747 // directives: 'parallel' with 'for' directive.
2748 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
2750 OMPCancelStackRAII CancelRegion(CGF, OMPD_parallel_for, S.hasCancel());
2751 CGF.EmitOMPWorksharingLoop(S, S.getEnsureUpperBound(), emitForLoopBounds,
2752 emitDispatchForLoopBounds);
2754 emitCommonOMPParallelDirective(*this, S, OMPD_for, CodeGen,
2755 emitEmptyBoundParameters);
2758 void CodeGenFunction::EmitOMPParallelForSimdDirective(
2759 const OMPParallelForSimdDirective &S) {
2760 // Emit directive as a combined directive that consists of two implicit
2761 // directives: 'parallel' with 'for' directive.
2762 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
2764 CGF.EmitOMPWorksharingLoop(S, S.getEnsureUpperBound(), emitForLoopBounds,
2765 emitDispatchForLoopBounds);
2767 emitCommonOMPParallelDirective(*this, S, OMPD_simd, CodeGen,
2768 emitEmptyBoundParameters);
2771 void CodeGenFunction::EmitOMPParallelSectionsDirective(
2772 const OMPParallelSectionsDirective &S) {
2773 // Emit directive as a combined directive that consists of two implicit
2774 // directives: 'parallel' with 'sections' directive.
2775 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
2777 CGF.EmitSections(S);
2779 emitCommonOMPParallelDirective(*this, S, OMPD_sections, CodeGen,
2780 emitEmptyBoundParameters);
2783 void CodeGenFunction::EmitOMPTaskBasedDirective(
2784 const OMPExecutableDirective &S, const OpenMPDirectiveKind CapturedRegion,
2785 const RegionCodeGenTy &BodyGen, const TaskGenTy &TaskGen,
2786 OMPTaskDataTy &Data) {
2787 // Emit outlined function for task construct.
2788 const CapturedStmt *CS = S.getCapturedStmt(CapturedRegion);
2789 auto I = CS->getCapturedDecl()->param_begin();
2790 auto PartId = std::next(I);
2791 auto TaskT = std::next(I, 4);
2792 // Check if the task is final
2793 if (const auto *Clause = S.getSingleClause<OMPFinalClause>()) {
2794 // If the condition constant folds and can be elided, try to avoid emitting
2795 // the condition and the dead arm of the if/else.
2796 const Expr *Cond = Clause->getCondition();
2798 if (ConstantFoldsToSimpleInteger(Cond, CondConstant))
2799 Data.Final.setInt(CondConstant);
2801 Data.Final.setPointer(EvaluateExprAsBool(Cond));
2803 // By default the task is not final.
2804 Data.Final.setInt(/*IntVal=*/false);
2806 // Check if the task has 'priority' clause.
2807 if (const auto *Clause = S.getSingleClause<OMPPriorityClause>()) {
2808 const Expr *Prio = Clause->getPriority();
2809 Data.Priority.setInt(/*IntVal=*/true);
2810 Data.Priority.setPointer(EmitScalarConversion(
2811 EmitScalarExpr(Prio), Prio->getType(),
2812 getContext().getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/1),
2813 Prio->getExprLoc()));
2815 // The first function argument for tasks is a thread id, the second one is a
2816 // part id (0 for tied tasks, >=0 for untied task).
2817 llvm::DenseSet<const VarDecl *> EmittedAsPrivate;
2818 // Get list of private variables.
2819 for (const auto *C : S.getClausesOfKind<OMPPrivateClause>()) {
2820 auto IRef = C->varlist_begin();
2821 for (const Expr *IInit : C->private_copies()) {
2822 const auto *OrigVD = cast<VarDecl>(cast<DeclRefExpr>(*IRef)->getDecl());
2823 if (EmittedAsPrivate.insert(OrigVD->getCanonicalDecl()).second) {
2824 Data.PrivateVars.push_back(*IRef);
2825 Data.PrivateCopies.push_back(IInit);
2830 EmittedAsPrivate.clear();
2831 // Get list of firstprivate variables.
2832 for (const auto *C : S.getClausesOfKind<OMPFirstprivateClause>()) {
2833 auto IRef = C->varlist_begin();
2834 auto IElemInitRef = C->inits().begin();
2835 for (const Expr *IInit : C->private_copies()) {
2836 const auto *OrigVD = cast<VarDecl>(cast<DeclRefExpr>(*IRef)->getDecl());
2837 if (EmittedAsPrivate.insert(OrigVD->getCanonicalDecl()).second) {
2838 Data.FirstprivateVars.push_back(*IRef);
2839 Data.FirstprivateCopies.push_back(IInit);
2840 Data.FirstprivateInits.push_back(*IElemInitRef);
2846 // Get list of lastprivate variables (for taskloops).
2847 llvm::DenseMap<const VarDecl *, const DeclRefExpr *> LastprivateDstsOrigs;
2848 for (const auto *C : S.getClausesOfKind<OMPLastprivateClause>()) {
2849 auto IRef = C->varlist_begin();
2850 auto ID = C->destination_exprs().begin();
2851 for (const Expr *IInit : C->private_copies()) {
2852 const auto *OrigVD = cast<VarDecl>(cast<DeclRefExpr>(*IRef)->getDecl());
2853 if (EmittedAsPrivate.insert(OrigVD->getCanonicalDecl()).second) {
2854 Data.LastprivateVars.push_back(*IRef);
2855 Data.LastprivateCopies.push_back(IInit);
2857 LastprivateDstsOrigs.insert(
2858 {cast<VarDecl>(cast<DeclRefExpr>(*ID)->getDecl()),
2859 cast<DeclRefExpr>(*IRef)});
2864 SmallVector<const Expr *, 4> LHSs;
2865 SmallVector<const Expr *, 4> RHSs;
2866 for (const auto *C : S.getClausesOfKind<OMPReductionClause>()) {
2867 auto IPriv = C->privates().begin();
2868 auto IRed = C->reduction_ops().begin();
2869 auto ILHS = C->lhs_exprs().begin();
2870 auto IRHS = C->rhs_exprs().begin();
2871 for (const Expr *Ref : C->varlists()) {
2872 Data.ReductionVars.emplace_back(Ref);
2873 Data.ReductionCopies.emplace_back(*IPriv);
2874 Data.ReductionOps.emplace_back(*IRed);
2875 LHSs.emplace_back(*ILHS);
2876 RHSs.emplace_back(*IRHS);
2877 std::advance(IPriv, 1);
2878 std::advance(IRed, 1);
2879 std::advance(ILHS, 1);
2880 std::advance(IRHS, 1);
2883 Data.Reductions = CGM.getOpenMPRuntime().emitTaskReductionInit(
2884 *this, S.getBeginLoc(), LHSs, RHSs, Data);
2885 // Build list of dependences.
2886 for (const auto *C : S.getClausesOfKind<OMPDependClause>())
2887 for (const Expr *IRef : C->varlists())
2888 Data.Dependences.emplace_back(C->getDependencyKind(), IRef);
2889 auto &&CodeGen = [&Data, &S, CS, &BodyGen, &LastprivateDstsOrigs,
2890 CapturedRegion](CodeGenFunction &CGF,
2891 PrePostActionTy &Action) {
2892 // Set proper addresses for generated private copies.
2893 OMPPrivateScope Scope(CGF);
2894 if (!Data.PrivateVars.empty() || !Data.FirstprivateVars.empty() ||
2895 !Data.LastprivateVars.empty()) {
2896 enum { PrivatesParam = 2, CopyFnParam = 3 };
2897 llvm::Value *CopyFn = CGF.Builder.CreateLoad(
2898 CGF.GetAddrOfLocalVar(CS->getCapturedDecl()->getParam(CopyFnParam)));
2899 llvm::Value *PrivatesPtr = CGF.Builder.CreateLoad(CGF.GetAddrOfLocalVar(
2900 CS->getCapturedDecl()->getParam(PrivatesParam)));
2902 llvm::SmallVector<std::pair<const VarDecl *, Address>, 16> PrivatePtrs;
2903 llvm::SmallVector<llvm::Value *, 16> CallArgs;
2904 CallArgs.push_back(PrivatesPtr);
2905 for (const Expr *E : Data.PrivateVars) {
2906 const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl());
2907 Address PrivatePtr = CGF.CreateMemTemp(
2908 CGF.getContext().getPointerType(E->getType()), ".priv.ptr.addr");
2909 PrivatePtrs.emplace_back(VD, PrivatePtr);
2910 CallArgs.push_back(PrivatePtr.getPointer());
2912 for (const Expr *E : Data.FirstprivateVars) {
2913 const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl());
2914 Address PrivatePtr =
2915 CGF.CreateMemTemp(CGF.getContext().getPointerType(E->getType()),
2916 ".firstpriv.ptr.addr");
2917 PrivatePtrs.emplace_back(VD, PrivatePtr);
2918 CallArgs.push_back(PrivatePtr.getPointer());
2920 for (const Expr *E : Data.LastprivateVars) {
2921 const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl());
2922 Address PrivatePtr =
2923 CGF.CreateMemTemp(CGF.getContext().getPointerType(E->getType()),
2924 ".lastpriv.ptr.addr");
2925 PrivatePtrs.emplace_back(VD, PrivatePtr);
2926 CallArgs.push_back(PrivatePtr.getPointer());
2928 CGF.CGM.getOpenMPRuntime().emitOutlinedFunctionCall(CGF, S.getBeginLoc(),
2930 for (const auto &Pair : LastprivateDstsOrigs) {
2931 const auto *OrigVD = cast<VarDecl>(Pair.second->getDecl());
2932 DeclRefExpr DRE(CGF.getContext(), const_cast<VarDecl *>(OrigVD),
2933 /*RefersToEnclosingVariableOrCapture=*/
2934 CGF.CapturedStmtInfo->lookup(OrigVD) != nullptr,
2935 Pair.second->getType(), VK_LValue,
2936 Pair.second->getExprLoc());
2937 Scope.addPrivate(Pair.first, [&CGF, &DRE]() {
2938 return CGF.EmitLValue(&DRE).getAddress();
2941 for (const auto &Pair : PrivatePtrs) {
2942 Address Replacement(CGF.Builder.CreateLoad(Pair.second),
2943 CGF.getContext().getDeclAlign(Pair.first));
2944 Scope.addPrivate(Pair.first, [Replacement]() { return Replacement; });
2947 if (Data.Reductions) {
2948 OMPLexicalScope LexScope(CGF, S, CapturedRegion);
2949 ReductionCodeGen RedCG(Data.ReductionVars, Data.ReductionCopies,
2951 llvm::Value *ReductionsPtr = CGF.Builder.CreateLoad(
2952 CGF.GetAddrOfLocalVar(CS->getCapturedDecl()->getParam(9)));
2953 for (unsigned Cnt = 0, E = Data.ReductionVars.size(); Cnt < E; ++Cnt) {
2954 RedCG.emitSharedLValue(CGF, Cnt);
2955 RedCG.emitAggregateType(CGF, Cnt);
2956 // FIXME: This must removed once the runtime library is fixed.
2957 // Emit required threadprivate variables for
2958 // initializer/combiner/finalizer.
2959 CGF.CGM.getOpenMPRuntime().emitTaskReductionFixups(CGF, S.getBeginLoc(),
2961 Address Replacement = CGF.CGM.getOpenMPRuntime().getTaskReductionItem(
2962 CGF, S.getBeginLoc(), ReductionsPtr, RedCG.getSharedLValue(Cnt));
2964 Address(CGF.EmitScalarConversion(
2965 Replacement.getPointer(), CGF.getContext().VoidPtrTy,
2966 CGF.getContext().getPointerType(
2967 Data.ReductionCopies[Cnt]->getType()),
2968 Data.ReductionCopies[Cnt]->getExprLoc()),
2969 Replacement.getAlignment());
2970 Replacement = RedCG.adjustPrivateAddress(CGF, Cnt, Replacement);
2971 Scope.addPrivate(RedCG.getBaseDecl(Cnt),
2972 [Replacement]() { return Replacement; });
2975 // Privatize all private variables except for in_reduction items.
2976 (void)Scope.Privatize();
2977 SmallVector<const Expr *, 4> InRedVars;
2978 SmallVector<const Expr *, 4> InRedPrivs;
2979 SmallVector<const Expr *, 4> InRedOps;
2980 SmallVector<const Expr *, 4> TaskgroupDescriptors;
2981 for (const auto *C : S.getClausesOfKind<OMPInReductionClause>()) {
2982 auto IPriv = C->privates().begin();
2983 auto IRed = C->reduction_ops().begin();
2984 auto ITD = C->taskgroup_descriptors().begin();
2985 for (const Expr *Ref : C->varlists()) {
2986 InRedVars.emplace_back(Ref);
2987 InRedPrivs.emplace_back(*IPriv);
2988 InRedOps.emplace_back(*IRed);
2989 TaskgroupDescriptors.emplace_back(*ITD);
2990 std::advance(IPriv, 1);
2991 std::advance(IRed, 1);
2992 std::advance(ITD, 1);
2995 // Privatize in_reduction items here, because taskgroup descriptors must be
2996 // privatized earlier.
2997 OMPPrivateScope InRedScope(CGF);
2998 if (!InRedVars.empty()) {
2999 ReductionCodeGen RedCG(InRedVars, InRedPrivs, InRedOps);
3000 for (unsigned Cnt = 0, E = InRedVars.size(); Cnt < E; ++Cnt) {
3001 RedCG.emitSharedLValue(CGF, Cnt);
3002 RedCG.emitAggregateType(CGF, Cnt);
3003 // The taskgroup descriptor variable is always implicit firstprivate and
3004 // privatized already during processing of the firstprivates.
3005 // FIXME: This must removed once the runtime library is fixed.
3006 // Emit required threadprivate variables for
3007 // initializer/combiner/finalizer.
3008 CGF.CGM.getOpenMPRuntime().emitTaskReductionFixups(CGF, S.getBeginLoc(),
3010 llvm::Value *ReductionsPtr =
3011 CGF.EmitLoadOfScalar(CGF.EmitLValue(TaskgroupDescriptors[Cnt]),
3012 TaskgroupDescriptors[Cnt]->getExprLoc());
3013 Address Replacement = CGF.CGM.getOpenMPRuntime().getTaskReductionItem(
3014 CGF, S.getBeginLoc(), ReductionsPtr, RedCG.getSharedLValue(Cnt));
3015 Replacement = Address(
3016 CGF.EmitScalarConversion(
3017 Replacement.getPointer(), CGF.getContext().VoidPtrTy,
3018 CGF.getContext().getPointerType(InRedPrivs[Cnt]->getType()),
3019 InRedPrivs[Cnt]->getExprLoc()),
3020 Replacement.getAlignment());
3021 Replacement = RedCG.adjustPrivateAddress(CGF, Cnt, Replacement);
3022 InRedScope.addPrivate(RedCG.getBaseDecl(Cnt),
3023 [Replacement]() { return Replacement; });
3026 (void)InRedScope.Privatize();
3031 llvm::Value *OutlinedFn = CGM.getOpenMPRuntime().emitTaskOutlinedFunction(
3032 S, *I, *PartId, *TaskT, S.getDirectiveKind(), CodeGen, Data.Tied,
3033 Data.NumberOfParts);
3034 OMPLexicalScope Scope(*this, S);
3035 TaskGen(*this, OutlinedFn, Data);
3038 static ImplicitParamDecl *
3039 createImplicitFirstprivateForType(ASTContext &C, OMPTaskDataTy &Data,
3040 QualType Ty, CapturedDecl *CD,
3041 SourceLocation Loc) {
3042 auto *OrigVD = ImplicitParamDecl::Create(C, CD, Loc, /*Id=*/nullptr, Ty,
3043 ImplicitParamDecl::Other);
3044 auto *OrigRef = DeclRefExpr::Create(
3045 C, NestedNameSpecifierLoc(), SourceLocation(), OrigVD,
3046 /*RefersToEnclosingVariableOrCapture=*/false, Loc, Ty, VK_LValue);
3047 auto *PrivateVD = ImplicitParamDecl::Create(C, CD, Loc, /*Id=*/nullptr, Ty,
3048 ImplicitParamDecl::Other);
3049 auto *PrivateRef = DeclRefExpr::Create(
3050 C, NestedNameSpecifierLoc(), SourceLocation(), PrivateVD,
3051 /*RefersToEnclosingVariableOrCapture=*/false, Loc, Ty, VK_LValue);
3052 QualType ElemType = C.getBaseElementType(Ty);
3053 auto *InitVD = ImplicitParamDecl::Create(C, CD, Loc, /*Id=*/nullptr, ElemType,
3054 ImplicitParamDecl::Other);
3055 auto *InitRef = DeclRefExpr::Create(
3056 C, NestedNameSpecifierLoc(), SourceLocation(), InitVD,
3057 /*RefersToEnclosingVariableOrCapture=*/false, Loc, ElemType, VK_LValue);
3058 PrivateVD->setInitStyle(VarDecl::CInit);
3059 PrivateVD->setInit(ImplicitCastExpr::Create(C, ElemType, CK_LValueToRValue,
3060 InitRef, /*BasePath=*/nullptr,
3062 Data.FirstprivateVars.emplace_back(OrigRef);
3063 Data.FirstprivateCopies.emplace_back(PrivateRef);
3064 Data.FirstprivateInits.emplace_back(InitRef);
3068 void CodeGenFunction::EmitOMPTargetTaskBasedDirective(
3069 const OMPExecutableDirective &S, const RegionCodeGenTy &BodyGen,
3070 OMPTargetDataInfo &InputInfo) {
3071 // Emit outlined function for task construct.
3072 const CapturedStmt *CS = S.getCapturedStmt(OMPD_task);
3073 Address CapturedStruct = GenerateCapturedStmtArgument(*CS);
3074 QualType SharedsTy = getContext().getRecordType(CS->getCapturedRecordDecl());
3075 auto I = CS->getCapturedDecl()->param_begin();
3076 auto PartId = std::next(I);
3077 auto TaskT = std::next(I, 4);
3079 // The task is not final.
3080 Data.Final.setInt(/*IntVal=*/false);
3081 // Get list of firstprivate variables.
3082 for (const auto *C : S.getClausesOfKind<OMPFirstprivateClause>()) {
3083 auto IRef = C->varlist_begin();
3084 auto IElemInitRef = C->inits().begin();
3085 for (auto *IInit : C->private_copies()) {
3086 Data.FirstprivateVars.push_back(*IRef);
3087 Data.FirstprivateCopies.push_back(IInit);
3088 Data.FirstprivateInits.push_back(*IElemInitRef);
3093 OMPPrivateScope TargetScope(*this);
3094 VarDecl *BPVD = nullptr;
3095 VarDecl *PVD = nullptr;
3096 VarDecl *SVD = nullptr;
3097 if (InputInfo.NumberOfTargetItems > 0) {
3098 auto *CD = CapturedDecl::Create(
3099 getContext(), getContext().getTranslationUnitDecl(), /*NumParams=*/0);
3100 llvm::APInt ArrSize(/*numBits=*/32, InputInfo.NumberOfTargetItems);
3101 QualType BaseAndPointersType = getContext().getConstantArrayType(
3102 getContext().VoidPtrTy, ArrSize, ArrayType::Normal,
3103 /*IndexTypeQuals=*/0);
3104 BPVD = createImplicitFirstprivateForType(
3105 getContext(), Data, BaseAndPointersType, CD, S.getBeginLoc());
3106 PVD = createImplicitFirstprivateForType(
3107 getContext(), Data, BaseAndPointersType, CD, S.getBeginLoc());
3108 QualType SizesType = getContext().getConstantArrayType(
3109 getContext().getSizeType(), ArrSize, ArrayType::Normal,
3110 /*IndexTypeQuals=*/0);
3111 SVD = createImplicitFirstprivateForType(getContext(), Data, SizesType, CD,
3113 TargetScope.addPrivate(
3114 BPVD, [&InputInfo]() { return InputInfo.BasePointersArray; });
3115 TargetScope.addPrivate(PVD,
3116 [&InputInfo]() { return InputInfo.PointersArray; });
3117 TargetScope.addPrivate(SVD,
3118 [&InputInfo]() { return InputInfo.SizesArray; });
3120 (void)TargetScope.Privatize();
3121 // Build list of dependences.
3122 for (const auto *C : S.getClausesOfKind<OMPDependClause>())
3123 for (const Expr *IRef : C->varlists())
3124 Data.Dependences.emplace_back(C->getDependencyKind(), IRef);
3125 auto &&CodeGen = [&Data, &S, CS, &BodyGen, BPVD, PVD, SVD,
3126 &InputInfo](CodeGenFunction &CGF, PrePostActionTy &Action) {
3127 // Set proper addresses for generated private copies.
3128 OMPPrivateScope Scope(CGF);
3129 if (!Data.FirstprivateVars.empty()) {
3130 enum { PrivatesParam = 2, CopyFnParam = 3 };
3131 llvm::Value *CopyFn = CGF.Builder.CreateLoad(
3132 CGF.GetAddrOfLocalVar(CS->getCapturedDecl()->getParam(CopyFnParam)));
3133 llvm::Value *PrivatesPtr = CGF.Builder.CreateLoad(CGF.GetAddrOfLocalVar(
3134 CS->getCapturedDecl()->getParam(PrivatesParam)));
3136 llvm::SmallVector<std::pair<const VarDecl *, Address>, 16> PrivatePtrs;
3137 llvm::SmallVector<llvm::Value *, 16> CallArgs;
3138 CallArgs.push_back(PrivatesPtr);
3139 for (const Expr *E : Data.FirstprivateVars) {
3140 const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl());
3141 Address PrivatePtr =
3142 CGF.CreateMemTemp(CGF.getContext().getPointerType(E->getType()),
3143 ".firstpriv.ptr.addr");
3144 PrivatePtrs.emplace_back(VD, PrivatePtr);
3145 CallArgs.push_back(PrivatePtr.getPointer());
3147 CGF.CGM.getOpenMPRuntime().emitOutlinedFunctionCall(CGF, S.getBeginLoc(),
3149 for (const auto &Pair : PrivatePtrs) {
3150 Address Replacement(CGF.Builder.CreateLoad(Pair.second),
3151 CGF.getContext().getDeclAlign(Pair.first));
3152 Scope.addPrivate(Pair.first, [Replacement]() { return Replacement; });
3155 // Privatize all private variables except for in_reduction items.
3156 (void)Scope.Privatize();
3157 if (InputInfo.NumberOfTargetItems > 0) {
3158 InputInfo.BasePointersArray = CGF.Builder.CreateConstArrayGEP(
3159 CGF.GetAddrOfLocalVar(BPVD), /*Index=*/0, CGF.getPointerSize());
3160 InputInfo.PointersArray = CGF.Builder.CreateConstArrayGEP(
3161 CGF.GetAddrOfLocalVar(PVD), /*Index=*/0, CGF.getPointerSize());
3162 InputInfo.SizesArray = CGF.Builder.CreateConstArrayGEP(
3163 CGF.GetAddrOfLocalVar(SVD), /*Index=*/0, CGF.getSizeSize());
3167 OMPLexicalScope LexScope(CGF, S, OMPD_task, /*EmitPreInitStmt=*/false);
3170 llvm::Value *OutlinedFn = CGM.getOpenMPRuntime().emitTaskOutlinedFunction(
3171 S, *I, *PartId, *TaskT, S.getDirectiveKind(), CodeGen, /*Tied=*/true,
3172 Data.NumberOfParts);
3173 llvm::APInt TrueOrFalse(32, S.hasClausesOfKind<OMPNowaitClause>() ? 1 : 0);
3174 IntegerLiteral IfCond(getContext(), TrueOrFalse,
3175 getContext().getIntTypeForBitwidth(32, /*Signed=*/0),
3178 CGM.getOpenMPRuntime().emitTaskCall(*this, S.getBeginLoc(), S, OutlinedFn,
3179 SharedsTy, CapturedStruct, &IfCond, Data);
3182 void CodeGenFunction::EmitOMPTaskDirective(const OMPTaskDirective &S) {
3183 // Emit outlined function for task construct.
3184 const CapturedStmt *CS = S.getCapturedStmt(OMPD_task);
3185 Address CapturedStruct = GenerateCapturedStmtArgument(*CS);
3186 QualType SharedsTy = getContext().getRecordType(CS->getCapturedRecordDecl());
3187 const Expr *IfCond = nullptr;
3188 for (const auto *C : S.getClausesOfKind<OMPIfClause>()) {
3189 if (C->getNameModifier() == OMPD_unknown ||
3190 C->getNameModifier() == OMPD_task) {
3191 IfCond = C->getCondition();
3197 // Check if we should emit tied or untied task.
3198 Data.Tied = !S.getSingleClause<OMPUntiedClause>();
3199 auto &&BodyGen = [CS](CodeGenFunction &CGF, PrePostActionTy &) {
3200 CGF.EmitStmt(CS->getCapturedStmt());
3202 auto &&TaskGen = [&S, SharedsTy, CapturedStruct,
3203 IfCond](CodeGenFunction &CGF, llvm::Value *OutlinedFn,
3204 const OMPTaskDataTy &Data) {
3205 CGF.CGM.getOpenMPRuntime().emitTaskCall(CGF, S.getBeginLoc(), S, OutlinedFn,
3206 SharedsTy, CapturedStruct, IfCond,
3209 EmitOMPTaskBasedDirective(S, OMPD_task, BodyGen, TaskGen, Data);
3212 void CodeGenFunction::EmitOMPTaskyieldDirective(
3213 const OMPTaskyieldDirective &S) {
3214 CGM.getOpenMPRuntime().emitTaskyieldCall(*this, S.getBeginLoc());
3217 void CodeGenFunction::EmitOMPBarrierDirective(const OMPBarrierDirective &S) {
3218 CGM.getOpenMPRuntime().emitBarrierCall(*this, S.getBeginLoc(), OMPD_barrier);
3221 void CodeGenFunction::EmitOMPTaskwaitDirective(const OMPTaskwaitDirective &S) {
3222 CGM.getOpenMPRuntime().emitTaskwaitCall(*this, S.getBeginLoc());
3225 void CodeGenFunction::EmitOMPTaskgroupDirective(
3226 const OMPTaskgroupDirective &S) {
3227 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
3229 if (const Expr *E = S.getReductionRef()) {
3230 SmallVector<const Expr *, 4> LHSs;
3231 SmallVector<const Expr *, 4> RHSs;
3233 for (const auto *C : S.getClausesOfKind<OMPTaskReductionClause>()) {
3234 auto IPriv = C->privates().begin();
3235 auto IRed = C->reduction_ops().begin();
3236 auto ILHS = C->lhs_exprs().begin();
3237 auto IRHS = C->rhs_exprs().begin();
3238 for (const Expr *Ref : C->varlists()) {
3239 Data.ReductionVars.emplace_back(Ref);
3240 Data.ReductionCopies.emplace_back(*IPriv);
3241 Data.ReductionOps.emplace_back(*IRed);
3242 LHSs.emplace_back(*ILHS);
3243 RHSs.emplace_back(*IRHS);
3244 std::advance(IPriv, 1);
3245 std::advance(IRed, 1);
3246 std::advance(ILHS, 1);
3247 std::advance(IRHS, 1);
3250 llvm::Value *ReductionDesc =
3251 CGF.CGM.getOpenMPRuntime().emitTaskReductionInit(CGF, S.getBeginLoc(),
3253 const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl());
3254 CGF.EmitVarDecl(*VD);
3255 CGF.EmitStoreOfScalar(ReductionDesc, CGF.GetAddrOfLocalVar(VD),
3256 /*Volatile=*/false, E->getType());
3258 CGF.EmitStmt(S.getInnermostCapturedStmt()->getCapturedStmt());
3260 OMPLexicalScope Scope(*this, S, OMPD_unknown);
3261 CGM.getOpenMPRuntime().emitTaskgroupRegion(*this, CodeGen, S.getBeginLoc());
3264 void CodeGenFunction::EmitOMPFlushDirective(const OMPFlushDirective &S) {
3265 CGM.getOpenMPRuntime().emitFlush(
3267 [&S]() -> ArrayRef<const Expr *> {
3268 if (const auto *FlushClause = S.getSingleClause<OMPFlushClause>())
3269 return llvm::makeArrayRef(FlushClause->varlist_begin(),
3270 FlushClause->varlist_end());
3276 void CodeGenFunction::EmitOMPDistributeLoop(const OMPLoopDirective &S,
3277 const CodeGenLoopTy &CodeGenLoop,
3279 // Emit the loop iteration variable.
3280 const auto *IVExpr = cast<DeclRefExpr>(S.getIterationVariable());
3281 const auto *IVDecl = cast<VarDecl>(IVExpr->getDecl());
3282 EmitVarDecl(*IVDecl);
3284 // Emit the iterations count variable.
3285 // If it is not a variable, Sema decided to calculate iterations count on each
3286 // iteration (e.g., it is foldable into a constant).
3287 if (const auto *LIExpr = dyn_cast<DeclRefExpr>(S.getLastIteration())) {
3288 EmitVarDecl(*cast<VarDecl>(LIExpr->getDecl()));
3289 // Emit calculation of the iterations count.
3290 EmitIgnoredExpr(S.getCalcLastIteration());
3293 CGOpenMPRuntime &RT = CGM.getOpenMPRuntime();
3295 bool HasLastprivateClause = false;
3296 // Check pre-condition.
3298 OMPLoopScope PreInitScope(*this, S);
3299 // Skip the entire loop if we don't meet the precondition.
3300 // If the condition constant folds and can be elided, avoid emitting the
3303 llvm::BasicBlock *ContBlock = nullptr;
3304 if (ConstantFoldsToSimpleInteger(S.getPreCond(), CondConstant)) {
3308 llvm::BasicBlock *ThenBlock = createBasicBlock("omp.precond.then");
3309 ContBlock = createBasicBlock("omp.precond.end");
3310 emitPreCond(*this, S, S.getPreCond(), ThenBlock, ContBlock,
3311 getProfileCount(&S));
3312 EmitBlock(ThenBlock);
3313 incrementProfileCounter(&S);
3316 emitAlignedClause(*this, S);
3317 // Emit 'then' code.
3319 // Emit helper vars inits.
3321 LValue LB = EmitOMPHelperVar(
3322 *this, cast<DeclRefExpr>(
3323 (isOpenMPLoopBoundSharingDirective(S.getDirectiveKind())
3324 ? S.getCombinedLowerBoundVariable()
3325 : S.getLowerBoundVariable())));
3326 LValue UB = EmitOMPHelperVar(
3327 *this, cast<DeclRefExpr>(
3328 (isOpenMPLoopBoundSharingDirective(S.getDirectiveKind())
3329 ? S.getCombinedUpperBoundVariable()
3330 : S.getUpperBoundVariable())));
3332 EmitOMPHelperVar(*this, cast<DeclRefExpr>(S.getStrideVariable()));
3334 EmitOMPHelperVar(*this, cast<DeclRefExpr>(S.getIsLastIterVariable()));
3336 OMPPrivateScope LoopScope(*this);
3337 if (EmitOMPFirstprivateClause(S, LoopScope)) {
3338 // Emit implicit barrier to synchronize threads and avoid data races
3339 // on initialization of firstprivate variables and post-update of
3340 // lastprivate variables.
3341 CGM.getOpenMPRuntime().emitBarrierCall(
3342 *this, S.getBeginLoc(), OMPD_unknown, /*EmitChecks=*/false,
3343 /*ForceSimpleCall=*/true);
3345 EmitOMPPrivateClause(S, LoopScope);
3346 if (isOpenMPSimdDirective(S.getDirectiveKind()) &&
3347 !isOpenMPParallelDirective(S.getDirectiveKind()) &&
3348 !isOpenMPTeamsDirective(S.getDirectiveKind()))
3349 EmitOMPReductionClauseInit(S, LoopScope);
3350 HasLastprivateClause = EmitOMPLastprivateClauseInit(S, LoopScope);
3351 EmitOMPPrivateLoopCounters(S, LoopScope);
3352 (void)LoopScope.Privatize();
3353 if (isOpenMPTargetExecutionDirective(S.getDirectiveKind()))
3354 CGM.getOpenMPRuntime().adjustTargetSpecificDataForLambdas(*this, S);
3356 // Detect the distribute schedule kind and chunk.
3357 llvm::Value *Chunk = nullptr;
3358 OpenMPDistScheduleClauseKind ScheduleKind = OMPC_DIST_SCHEDULE_unknown;
3359 if (const auto *C = S.getSingleClause<OMPDistScheduleClause>()) {
3360 ScheduleKind = C->getDistScheduleKind();
3361 if (const Expr *Ch = C->getChunkSize()) {
3362 Chunk = EmitScalarExpr(Ch);
3363 Chunk = EmitScalarConversion(Chunk, Ch->getType(),
3364 S.getIterationVariable()->getType(),
3368 // Default behaviour for dist_schedule clause.
3369 CGM.getOpenMPRuntime().getDefaultDistScheduleAndChunk(
3370 *this, S, ScheduleKind, Chunk);
3372 const unsigned IVSize = getContext().getTypeSize(IVExpr->getType());
3373 const bool IVSigned = IVExpr->getType()->hasSignedIntegerRepresentation();
3375 // OpenMP [2.10.8, distribute Construct, Description]
3376 // If dist_schedule is specified, kind must be static. If specified,
3377 // iterations are divided into chunks of size chunk_size, chunks are
3378 // assigned to the teams of the league in a round-robin fashion in the
3379 // order of the team number. When no chunk_size is specified, the
3380 // iteration space is divided into chunks that are approximately equal
3381 // in size, and at most one chunk is distributed to each team of the
3382 // league. The size of the chunks is unspecified in this case.
3383 bool StaticChunked = RT.isStaticChunked(
3384 ScheduleKind, /* Chunked */ Chunk != nullptr) &&
3385 isOpenMPLoopBoundSharingDirective(S.getDirectiveKind());
3386 if (RT.isStaticNonchunked(ScheduleKind,
3387 /* Chunked */ Chunk != nullptr) ||
3389 if (isOpenMPSimdDirective(S.getDirectiveKind()))
3390 EmitOMPSimdInit(S, /*IsMonotonic=*/true);
3391 CGOpenMPRuntime::StaticRTInput StaticInit(
3392 IVSize, IVSigned, /* Ordered = */ false, IL.getAddress(),
3393 LB.getAddress(), UB.getAddress(), ST.getAddress(),
3394 StaticChunked ? Chunk : nullptr);
3395 RT.emitDistributeStaticInit(*this, S.getBeginLoc(), ScheduleKind,
3398 getJumpDestInCurrentScope(createBasicBlock("omp.loop.exit"));
3399 // UB = min(UB, GlobalUB);
3400 EmitIgnoredExpr(isOpenMPLoopBoundSharingDirective(S.getDirectiveKind())
3401 ? S.getCombinedEnsureUpperBound()
3402 : S.getEnsureUpperBound());
3404 EmitIgnoredExpr(isOpenMPLoopBoundSharingDirective(S.getDirectiveKind())
3405 ? S.getCombinedInit()
3409 isOpenMPLoopBoundSharingDirective(S.getDirectiveKind())
3410 ? S.getCombinedCond()
3414 Cond = S.getCombinedDistCond();
3416 // For static unchunked schedules generate:
3418 // 1. For distribute alone, codegen
3419 // while (idx <= UB) {
3424 // 2. When combined with 'for' (e.g. as in 'distribute parallel for')
3425 // while (idx <= UB) {
3426 // <CodeGen rest of pragma>(LB, UB);
3430 // For static chunk one schedule generate:
3432 // while (IV <= GlobalUB) {
3433 // <CodeGen rest of pragma>(LB, UB);
3436 // UB = min(UB, GlobalUB);
3440 EmitOMPInnerLoop(S, LoopScope.requiresCleanups(), Cond, IncExpr,
3441 [&S, LoopExit, &CodeGenLoop](CodeGenFunction &CGF) {
3442 CodeGenLoop(CGF, S, LoopExit);
3444 [&S, StaticChunked](CodeGenFunction &CGF) {
3445 if (StaticChunked) {
3446 CGF.EmitIgnoredExpr(S.getCombinedNextLowerBound());
3447 CGF.EmitIgnoredExpr(S.getCombinedNextUpperBound());
3448 CGF.EmitIgnoredExpr(S.getCombinedEnsureUpperBound());
3449 CGF.EmitIgnoredExpr(S.getCombinedInit());
3452 EmitBlock(LoopExit.getBlock());
3453 // Tell the runtime we are done.
3454 RT.emitForStaticFinish(*this, S.getBeginLoc(), S.getDirectiveKind());
3456 // Emit the outer loop, which requests its work chunk [LB..UB] from
3457 // runtime and runs the inner loop to process it.
3458 const OMPLoopArguments LoopArguments = {
3459 LB.getAddress(), UB.getAddress(), ST.getAddress(), IL.getAddress(),
3461 EmitOMPDistributeOuterLoop(ScheduleKind, S, LoopScope, LoopArguments,
3464 if (isOpenMPSimdDirective(S.getDirectiveKind())) {
3465 EmitOMPSimdFinal(S, [IL, &S](CodeGenFunction &CGF) {
3466 return CGF.Builder.CreateIsNotNull(
3467 CGF.EmitLoadOfScalar(IL, S.getBeginLoc()));
3470 if (isOpenMPSimdDirective(S.getDirectiveKind()) &&
3471 !isOpenMPParallelDirective(S.getDirectiveKind()) &&
3472 !isOpenMPTeamsDirective(S.getDirectiveKind())) {
3473 EmitOMPReductionClauseFinal(S, OMPD_simd);
3474 // Emit post-update of the reduction variables if IsLastIter != 0.
3475 emitPostUpdateForReductionClause(
3476 *this, S, [IL, &S](CodeGenFunction &CGF) {
3477 return CGF.Builder.CreateIsNotNull(
3478 CGF.EmitLoadOfScalar(IL, S.getBeginLoc()));
3481 // Emit final copy of the lastprivate variables if IsLastIter != 0.
3482 if (HasLastprivateClause) {
3483 EmitOMPLastprivateClauseFinal(
3484 S, /*NoFinals=*/false,
3485 Builder.CreateIsNotNull(EmitLoadOfScalar(IL, S.getBeginLoc())));
3489 // We're now done with the loop, so jump to the continuation block.
3491 EmitBranch(ContBlock);
3492 EmitBlock(ContBlock, true);
3497 void CodeGenFunction::EmitOMPDistributeDirective(
3498 const OMPDistributeDirective &S) {
3499 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &) {
3500 CGF.EmitOMPDistributeLoop(S, emitOMPLoopBodyWithStopPoint, S.getInc());
3502 OMPLexicalScope Scope(*this, S, OMPD_unknown);
3503 CGM.getOpenMPRuntime().emitInlinedDirective(*this, OMPD_distribute, CodeGen);
3506 static llvm::Function *emitOutlinedOrderedFunction(CodeGenModule &CGM,
3507 const CapturedStmt *S) {
3508 CodeGenFunction CGF(CGM, /*suppressNewContext=*/true);
3509 CodeGenFunction::CGCapturedStmtInfo CapStmtInfo;
3510 CGF.CapturedStmtInfo = &CapStmtInfo;
3511 llvm::Function *Fn = CGF.GenerateOpenMPCapturedStmtFunction(*S);
3512 Fn->setDoesNotRecurse();
3516 void CodeGenFunction::EmitOMPOrderedDirective(const OMPOrderedDirective &S) {
3517 if (S.hasClausesOfKind<OMPDependClause>()) {
3518 assert(!S.getAssociatedStmt() &&
3519 "No associated statement must be in ordered depend construct.");
3520 for (const auto *DC : S.getClausesOfKind<OMPDependClause>())
3521 CGM.getOpenMPRuntime().emitDoacrossOrdered(*this, DC);
3524 const auto *C = S.getSingleClause<OMPSIMDClause>();
3525 auto &&CodeGen = [&S, C, this](CodeGenFunction &CGF,
3526 PrePostActionTy &Action) {
3527 const CapturedStmt *CS = S.getInnermostCapturedStmt();
3529 llvm::SmallVector<llvm::Value *, 16> CapturedVars;
3530 CGF.GenerateOpenMPCapturedVars(*CS, CapturedVars);
3531 llvm::Function *OutlinedFn = emitOutlinedOrderedFunction(CGM, CS);
3532 CGM.getOpenMPRuntime().emitOutlinedFunctionCall(CGF, S.getBeginLoc(),
3533 OutlinedFn, CapturedVars);
3536 CGF.EmitStmt(CS->getCapturedStmt());
3539 OMPLexicalScope Scope(*this, S, OMPD_unknown);
3540 CGM.getOpenMPRuntime().emitOrderedRegion(*this, CodeGen, S.getBeginLoc(), !C);
3543 static llvm::Value *convertToScalarValue(CodeGenFunction &CGF, RValue Val,
3544 QualType SrcType, QualType DestType,
3545 SourceLocation Loc) {
3546 assert(CGF.hasScalarEvaluationKind(DestType) &&
3547 "DestType must have scalar evaluation kind.");
3548 assert(!Val.isAggregate() && "Must be a scalar or complex.");
3549 return Val.isScalar() ? CGF.EmitScalarConversion(Val.getScalarVal(), SrcType,
3551 : CGF.EmitComplexToScalarConversion(
3552 Val.getComplexVal(), SrcType, DestType, Loc);
3555 static CodeGenFunction::ComplexPairTy
3556 convertToComplexValue(CodeGenFunction &CGF, RValue Val, QualType SrcType,
3557 QualType DestType, SourceLocation Loc) {
3558 assert(CGF.getEvaluationKind(DestType) == TEK_Complex &&
3559 "DestType must have complex evaluation kind.");
3560 CodeGenFunction::ComplexPairTy ComplexVal;
3561 if (Val.isScalar()) {
3562 // Convert the input element to the element type of the complex.
3563 QualType DestElementType =
3564 DestType->castAs<ComplexType>()->getElementType();
3565 llvm::Value *ScalarVal = CGF.EmitScalarConversion(
3566 Val.getScalarVal(), SrcType, DestElementType, Loc);
3567 ComplexVal = CodeGenFunction::ComplexPairTy(
3568 ScalarVal, llvm::Constant::getNullValue(ScalarVal->getType()));
3570 assert(Val.isComplex() && "Must be a scalar or complex.");
3571 QualType SrcElementType = SrcType->castAs<ComplexType>()->getElementType();
3572 QualType DestElementType =
3573 DestType->castAs<ComplexType>()->getElementType();
3574 ComplexVal.first = CGF.EmitScalarConversion(
3575 Val.getComplexVal().first, SrcElementType, DestElementType, Loc);
3576 ComplexVal.second = CGF.EmitScalarConversion(
3577 Val.getComplexVal().second, SrcElementType, DestElementType, Loc);
3582 static void emitSimpleAtomicStore(CodeGenFunction &CGF, bool IsSeqCst,
3583 LValue LVal, RValue RVal) {
3584 if (LVal.isGlobalReg()) {
3585 CGF.EmitStoreThroughGlobalRegLValue(RVal, LVal);
3587 CGF.EmitAtomicStore(RVal, LVal,
3588 IsSeqCst ? llvm::AtomicOrdering::SequentiallyConsistent
3589 : llvm::AtomicOrdering::Monotonic,
3590 LVal.isVolatile(), /*IsInit=*/false);
3594 void CodeGenFunction::emitOMPSimpleStore(LValue LVal, RValue RVal,
3595 QualType RValTy, SourceLocation Loc) {
3596 switch (getEvaluationKind(LVal.getType())) {
3598 EmitStoreThroughLValue(RValue::get(convertToScalarValue(
3599 *this, RVal, RValTy, LVal.getType(), Loc)),
3604 convertToComplexValue(*this, RVal, RValTy, LVal.getType(), Loc), LVal,
3608 llvm_unreachable("Must be a scalar or complex.");
3612 static void emitOMPAtomicReadExpr(CodeGenFunction &CGF, bool IsSeqCst,
3613 const Expr *X, const Expr *V,
3614 SourceLocation Loc) {
3616 assert(V->isLValue() && "V of 'omp atomic read' is not lvalue");
3617 assert(X->isLValue() && "X of 'omp atomic read' is not lvalue");
3618 LValue XLValue = CGF.EmitLValue(X);
3619 LValue VLValue = CGF.EmitLValue(V);
3620 RValue Res = XLValue.isGlobalReg()
3621 ? CGF.EmitLoadOfLValue(XLValue, Loc)
3622 : CGF.EmitAtomicLoad(
3624 IsSeqCst ? llvm::AtomicOrdering::SequentiallyConsistent
3625 : llvm::AtomicOrdering::Monotonic,
3626 XLValue.isVolatile());
3627 // OpenMP, 2.12.6, atomic Construct
3628 // Any atomic construct with a seq_cst clause forces the atomically
3629 // performed operation to include an implicit flush operation without a
3632 CGF.CGM.getOpenMPRuntime().emitFlush(CGF, llvm::None, Loc);
3633 CGF.emitOMPSimpleStore(VLValue, Res, X->getType().getNonReferenceType(), Loc);
3636 static void emitOMPAtomicWriteExpr(CodeGenFunction &CGF, bool IsSeqCst,
3637 const Expr *X, const Expr *E,
3638 SourceLocation Loc) {
3640 assert(X->isLValue() && "X of 'omp atomic write' is not lvalue");
3641 emitSimpleAtomicStore(CGF, IsSeqCst, CGF.EmitLValue(X), CGF.EmitAnyExpr(E));
3642 // OpenMP, 2.12.6, atomic Construct
3643 // Any atomic construct with a seq_cst clause forces the atomically
3644 // performed operation to include an implicit flush operation without a
3647 CGF.CGM.getOpenMPRuntime().emitFlush(CGF, llvm::None, Loc);
3650 static std::pair<bool, RValue> emitOMPAtomicRMW(CodeGenFunction &CGF, LValue X,
3652 BinaryOperatorKind BO,
3653 llvm::AtomicOrdering AO,
3654 bool IsXLHSInRHSPart) {
3655 ASTContext &Context = CGF.getContext();
3656 // Allow atomicrmw only if 'x' and 'update' are integer values, lvalue for 'x'
3657 // expression is simple and atomic is allowed for the given type for the
3659 if (BO == BO_Comma || !Update.isScalar() ||
3660 !Update.getScalarVal()->getType()->isIntegerTy() ||
3661 !X.isSimple() || (!isa<llvm::ConstantInt>(Update.getScalarVal()) &&
3662 (Update.getScalarVal()->getType() !=
3663 X.getAddress().getElementType())) ||
3664 !X.getAddress().getElementType()->isIntegerTy() ||
3665 !Context.getTargetInfo().hasBuiltinAtomic(
3666 Context.getTypeSize(X.getType()), Context.toBits(X.getAlignment())))
3667 return std::make_pair(false, RValue::get(nullptr));
3669 llvm::AtomicRMWInst::BinOp RMWOp;
3672 RMWOp = llvm::AtomicRMWInst::Add;
3675 if (!IsXLHSInRHSPart)
3676 return std::make_pair(false, RValue::get(nullptr));
3677 RMWOp = llvm::AtomicRMWInst::Sub;
3680 RMWOp = llvm::AtomicRMWInst::And;
3683 RMWOp = llvm::AtomicRMWInst::Or;
3686 RMWOp = llvm::AtomicRMWInst::Xor;
3689 RMWOp = X.getType()->hasSignedIntegerRepresentation()
3690 ? (IsXLHSInRHSPart ? llvm::AtomicRMWInst::Min
3691 : llvm::AtomicRMWInst::Max)
3692 : (IsXLHSInRHSPart ? llvm::AtomicRMWInst::UMin
3693 : llvm::AtomicRMWInst::UMax);
3696 RMWOp = X.getType()->hasSignedIntegerRepresentation()
3697 ? (IsXLHSInRHSPart ? llvm::AtomicRMWInst::Max
3698 : llvm::AtomicRMWInst::Min)
3699 : (IsXLHSInRHSPart ? llvm::AtomicRMWInst::UMax
3700 : llvm::AtomicRMWInst::UMin);
3703 RMWOp = llvm::AtomicRMWInst::Xchg;
3712 return std::make_pair(false, RValue::get(nullptr));
3731 llvm_unreachable("Unsupported atomic update operation");
3733 llvm::Value *UpdateVal = Update.getScalarVal();
3734 if (auto *IC = dyn_cast<llvm::ConstantInt>(UpdateVal)) {
3735 UpdateVal = CGF.Builder.CreateIntCast(
3736 IC, X.getAddress().getElementType(),
3737 X.getType()->hasSignedIntegerRepresentation());
3740 CGF.Builder.CreateAtomicRMW(RMWOp, X.getPointer(), UpdateVal, AO);
3741 return std::make_pair(true, RValue::get(Res));
3744 std::pair<bool, RValue> CodeGenFunction::EmitOMPAtomicSimpleUpdateExpr(
3745 LValue X, RValue E, BinaryOperatorKind BO, bool IsXLHSInRHSPart,
3746 llvm::AtomicOrdering AO, SourceLocation Loc,
3747 const llvm::function_ref<RValue(RValue)> CommonGen) {
3748 // Update expressions are allowed to have the following forms:
3749 // x binop= expr; -> xrval + expr;
3750 // x++, ++x -> xrval + 1;
3751 // x--, --x -> xrval - 1;
3752 // x = x binop expr; -> xrval binop expr
3753 // x = expr Op x; - > expr binop xrval;
3754 auto Res = emitOMPAtomicRMW(*this, X, E, BO, AO, IsXLHSInRHSPart);
3756 if (X.isGlobalReg()) {
3757 // Emit an update expression: 'xrval' binop 'expr' or 'expr' binop
3759 EmitStoreThroughLValue(CommonGen(EmitLoadOfLValue(X, Loc)), X);
3761 // Perform compare-and-swap procedure.
3762 EmitAtomicUpdate(X, AO, CommonGen, X.getType().isVolatileQualified());
3768 static void emitOMPAtomicUpdateExpr(CodeGenFunction &CGF, bool IsSeqCst,
3769 const Expr *X, const Expr *E,
3770 const Expr *UE, bool IsXLHSInRHSPart,
3771 SourceLocation Loc) {
3772 assert(isa<BinaryOperator>(UE->IgnoreImpCasts()) &&
3773 "Update expr in 'atomic update' must be a binary operator.");
3774 const auto *BOUE = cast<BinaryOperator>(UE->IgnoreImpCasts());
3775 // Update expressions are allowed to have the following forms:
3776 // x binop= expr; -> xrval + expr;
3777 // x++, ++x -> xrval + 1;
3778 // x--, --x -> xrval - 1;
3779 // x = x binop expr; -> xrval binop expr
3780 // x = expr Op x; - > expr binop xrval;
3781 assert(X->isLValue() && "X of 'omp atomic update' is not lvalue");
3782 LValue XLValue = CGF.EmitLValue(X);
3783 RValue ExprRValue = CGF.EmitAnyExpr(E);
3784 llvm::AtomicOrdering AO = IsSeqCst
3785 ? llvm::AtomicOrdering::SequentiallyConsistent
3786 : llvm::AtomicOrdering::Monotonic;
3787 const auto *LHS = cast<OpaqueValueExpr>(BOUE->getLHS()->IgnoreImpCasts());
3788 const auto *RHS = cast<OpaqueValueExpr>(BOUE->getRHS()->IgnoreImpCasts());
3789 const OpaqueValueExpr *XRValExpr = IsXLHSInRHSPart ? LHS : RHS;
3790 const OpaqueValueExpr *ERValExpr = IsXLHSInRHSPart ? RHS : LHS;
3791 auto &&Gen = [&CGF, UE, ExprRValue, XRValExpr, ERValExpr](RValue XRValue) {
3792 CodeGenFunction::OpaqueValueMapping MapExpr(CGF, ERValExpr, ExprRValue);
3793 CodeGenFunction::OpaqueValueMapping MapX(CGF, XRValExpr, XRValue);
3794 return CGF.EmitAnyExpr(UE);
3796 (void)CGF.EmitOMPAtomicSimpleUpdateExpr(
3797 XLValue, ExprRValue, BOUE->getOpcode(), IsXLHSInRHSPart, AO, Loc, Gen);
3798 // OpenMP, 2.12.6, atomic Construct
3799 // Any atomic construct with a seq_cst clause forces the atomically
3800 // performed operation to include an implicit flush operation without a
3803 CGF.CGM.getOpenMPRuntime().emitFlush(CGF, llvm::None, Loc);
3806 static RValue convertToType(CodeGenFunction &CGF, RValue Value,
3807 QualType SourceType, QualType ResType,
3808 SourceLocation Loc) {
3809 switch (CGF.getEvaluationKind(ResType)) {
3812 convertToScalarValue(CGF, Value, SourceType, ResType, Loc));
3814 auto Res = convertToComplexValue(CGF, Value, SourceType, ResType, Loc);
3815 return RValue::getComplex(Res.first, Res.second);
3820 llvm_unreachable("Must be a scalar or complex.");
3823 static void emitOMPAtomicCaptureExpr(CodeGenFunction &CGF, bool IsSeqCst,
3824 bool IsPostfixUpdate, const Expr *V,
3825 const Expr *X, const Expr *E,
3826 const Expr *UE, bool IsXLHSInRHSPart,
3827 SourceLocation Loc) {
3828 assert(X->isLValue() && "X of 'omp atomic capture' is not lvalue");
3829 assert(V->isLValue() && "V of 'omp atomic capture' is not lvalue");
3831 LValue VLValue = CGF.EmitLValue(V);
3832 LValue XLValue = CGF.EmitLValue(X);
3833 RValue ExprRValue = CGF.EmitAnyExpr(E);
3834 llvm::AtomicOrdering AO = IsSeqCst
3835 ? llvm::AtomicOrdering::SequentiallyConsistent
3836 : llvm::AtomicOrdering::Monotonic;
3837 QualType NewVValType;
3839 // 'x' is updated with some additional value.
3840 assert(isa<BinaryOperator>(UE->IgnoreImpCasts()) &&
3841 "Update expr in 'atomic capture' must be a binary operator.");
3842 const auto *BOUE = cast<BinaryOperator>(UE->IgnoreImpCasts());
3843 // Update expressions are allowed to have the following forms:
3844 // x binop= expr; -> xrval + expr;
3845 // x++, ++x -> xrval + 1;
3846 // x--, --x -> xrval - 1;
3847 // x = x binop expr; -> xrval binop expr
3848 // x = expr Op x; - > expr binop xrval;
3849 const auto *LHS = cast<OpaqueValueExpr>(BOUE->getLHS()->IgnoreImpCasts());
3850 const auto *RHS = cast<OpaqueValueExpr>(BOUE->getRHS()->IgnoreImpCasts());
3851 const OpaqueValueExpr *XRValExpr = IsXLHSInRHSPart ? LHS : RHS;
3852 NewVValType = XRValExpr->getType();
3853 const OpaqueValueExpr *ERValExpr = IsXLHSInRHSPart ? RHS : LHS;
3854 auto &&Gen = [&CGF, &NewVVal, UE, ExprRValue, XRValExpr, ERValExpr,
3855 IsPostfixUpdate](RValue XRValue) {
3856 CodeGenFunction::OpaqueValueMapping MapExpr(CGF, ERValExpr, ExprRValue);
3857 CodeGenFunction::OpaqueValueMapping MapX(CGF, XRValExpr, XRValue);
3858 RValue Res = CGF.EmitAnyExpr(UE);
3859 NewVVal = IsPostfixUpdate ? XRValue : Res;
3862 auto Res = CGF.EmitOMPAtomicSimpleUpdateExpr(
3863 XLValue, ExprRValue, BOUE->getOpcode(), IsXLHSInRHSPart, AO, Loc, Gen);
3865 // 'atomicrmw' instruction was generated.
3866 if (IsPostfixUpdate) {
3867 // Use old value from 'atomicrmw'.
3868 NewVVal = Res.second;
3870 // 'atomicrmw' does not provide new value, so evaluate it using old
3872 CodeGenFunction::OpaqueValueMapping MapExpr(CGF, ERValExpr, ExprRValue);
3873 CodeGenFunction::OpaqueValueMapping MapX(CGF, XRValExpr, Res.second);
3874 NewVVal = CGF.EmitAnyExpr(UE);
3878 // 'x' is simply rewritten with some 'expr'.
3879 NewVValType = X->getType().getNonReferenceType();
3880 ExprRValue = convertToType(CGF, ExprRValue, E->getType(),
3881 X->getType().getNonReferenceType(), Loc);
3882 auto &&Gen = [&NewVVal, ExprRValue](RValue XRValue) {
3886 // Try to perform atomicrmw xchg, otherwise simple exchange.
3887 auto Res = CGF.EmitOMPAtomicSimpleUpdateExpr(
3888 XLValue, ExprRValue, /*BO=*/BO_Assign, /*IsXLHSInRHSPart=*/false, AO,
3891 // 'atomicrmw' instruction was generated.
3892 NewVVal = IsPostfixUpdate ? Res.second : ExprRValue;
3895 // Emit post-update store to 'v' of old/new 'x' value.
3896 CGF.emitOMPSimpleStore(VLValue, NewVVal, NewVValType, Loc);
3897 // OpenMP, 2.12.6, atomic Construct
3898 // Any atomic construct with a seq_cst clause forces the atomically
3899 // performed operation to include an implicit flush operation without a
3902 CGF.CGM.getOpenMPRuntime().emitFlush(CGF, llvm::None, Loc);
3905 static void emitOMPAtomicExpr(CodeGenFunction &CGF, OpenMPClauseKind Kind,
3906 bool IsSeqCst, bool IsPostfixUpdate,
3907 const Expr *X, const Expr *V, const Expr *E,
3908 const Expr *UE, bool IsXLHSInRHSPart,
3909 SourceLocation Loc) {
3912 emitOMPAtomicReadExpr(CGF, IsSeqCst, X, V, Loc);
3915 emitOMPAtomicWriteExpr(CGF, IsSeqCst, X, E, Loc);
3919 emitOMPAtomicUpdateExpr(CGF, IsSeqCst, X, E, UE, IsXLHSInRHSPart, Loc);
3922 emitOMPAtomicCaptureExpr(CGF, IsSeqCst, IsPostfixUpdate, V, X, E, UE,
3923 IsXLHSInRHSPart, Loc);
3927 case OMPC_num_threads:
3929 case OMPC_firstprivate:
3930 case OMPC_lastprivate:
3931 case OMPC_reduction:
3932 case OMPC_task_reduction:
3933 case OMPC_in_reduction:
3943 case OMPC_copyprivate:
3945 case OMPC_proc_bind:
3950 case OMPC_threadprivate:
3952 case OMPC_mergeable:
3957 case OMPC_num_teams:
3958 case OMPC_thread_limit:
3960 case OMPC_grainsize:
3962 case OMPC_num_tasks:
3964 case OMPC_dist_schedule:
3965 case OMPC_defaultmap:
3969 case OMPC_use_device_ptr:
3970 case OMPC_is_device_ptr:
3971 case OMPC_unified_address:
3972 case OMPC_unified_shared_memory:
3973 case OMPC_reverse_offload:
3974 case OMPC_dynamic_allocators:
3975 case OMPC_atomic_default_mem_order:
3976 llvm_unreachable("Clause is not allowed in 'omp atomic'.");
3980 void CodeGenFunction::EmitOMPAtomicDirective(const OMPAtomicDirective &S) {
3981 bool IsSeqCst = S.getSingleClause<OMPSeqCstClause>();
3982 OpenMPClauseKind Kind = OMPC_unknown;
3983 for (const OMPClause *C : S.clauses()) {
3984 // Find first clause (skip seq_cst clause, if it is first).
3985 if (C->getClauseKind() != OMPC_seq_cst) {
3986 Kind = C->getClauseKind();
3991 const Stmt *CS = S.getInnermostCapturedStmt()->IgnoreContainers();
3992 if (const auto *FE = dyn_cast<FullExpr>(CS))
3993 enterFullExpression(FE);
3994 // Processing for statements under 'atomic capture'.
3995 if (const auto *Compound = dyn_cast<CompoundStmt>(CS)) {
3996 for (const Stmt *C : Compound->body()) {
3997 if (const auto *FE = dyn_cast<FullExpr>(C))
3998 enterFullExpression(FE);
4002 auto &&CodeGen = [&S, Kind, IsSeqCst, CS](CodeGenFunction &CGF,
4003 PrePostActionTy &) {
4004 CGF.EmitStopPoint(CS);
4005 emitOMPAtomicExpr(CGF, Kind, IsSeqCst, S.isPostfixUpdate(), S.getX(),
4006 S.getV(), S.getExpr(), S.getUpdateExpr(),
4007 S.isXLHSInRHSPart(), S.getBeginLoc());
4009 OMPLexicalScope Scope(*this, S, OMPD_unknown);
4010 CGM.getOpenMPRuntime().emitInlinedDirective(*this, OMPD_atomic, CodeGen);
4013 static void emitCommonOMPTargetDirective(CodeGenFunction &CGF,
4014 const OMPExecutableDirective &S,
4015 const RegionCodeGenTy &CodeGen) {
4016 assert(isOpenMPTargetExecutionDirective(S.getDirectiveKind()));
4017 CodeGenModule &CGM = CGF.CGM;
4019 // On device emit this construct as inlined code.
4020 if (CGM.getLangOpts().OpenMPIsDevice) {
4021 OMPLexicalScope Scope(CGF, S, OMPD_target);
4022 CGM.getOpenMPRuntime().emitInlinedDirective(
4023 CGF, OMPD_target, [&S](CodeGenFunction &CGF, PrePostActionTy &) {
4024 CGF.EmitStmt(S.getInnermostCapturedStmt()->getCapturedStmt());
4029 llvm::Function *Fn = nullptr;
4030 llvm::Constant *FnID = nullptr;
4032 const Expr *IfCond = nullptr;
4033 // Check for the at most one if clause associated with the target region.
4034 for (const auto *C : S.getClausesOfKind<OMPIfClause>()) {
4035 if (C->getNameModifier() == OMPD_unknown ||
4036 C->getNameModifier() == OMPD_target) {
4037 IfCond = C->getCondition();
4042 // Check if we have any device clause associated with the directive.
4043 const Expr *Device = nullptr;
4044 if (auto *C = S.getSingleClause<OMPDeviceClause>())
4045 Device = C->getDevice();
4047 // Check if we have an if clause whose conditional always evaluates to false
4048 // or if we do not have any targets specified. If so the target region is not
4049 // an offload entry point.
4050 bool IsOffloadEntry = true;
4053 if (CGF.ConstantFoldsToSimpleInteger(IfCond, Val) && !Val)
4054 IsOffloadEntry = false;
4056 if (CGM.getLangOpts().OMPTargetTriples.empty())
4057 IsOffloadEntry = false;
4059 assert(CGF.CurFuncDecl && "No parent declaration for target region!");
4060 StringRef ParentName;
4061 // In case we have Ctors/Dtors we use the complete type variant to produce
4062 // the mangling of the device outlined kernel.
4063 if (const auto *D = dyn_cast<CXXConstructorDecl>(CGF.CurFuncDecl))
4064 ParentName = CGM.getMangledName(GlobalDecl(D, Ctor_Complete));
4065 else if (const auto *D = dyn_cast<CXXDestructorDecl>(CGF.CurFuncDecl))
4066 ParentName = CGM.getMangledName(GlobalDecl(D, Dtor_Complete));
4069 CGM.getMangledName(GlobalDecl(cast<FunctionDecl>(CGF.CurFuncDecl)));
4071 // Emit target region as a standalone region.
4072 CGM.getOpenMPRuntime().emitTargetOutlinedFunction(S, ParentName, Fn, FnID,
4073 IsOffloadEntry, CodeGen);
4074 OMPLexicalScope Scope(CGF, S, OMPD_task);
4075 auto &&SizeEmitter = [](CodeGenFunction &CGF, const OMPLoopDirective &D) {
4076 OMPLoopScope(CGF, D);
4077 // Emit calculation of the iterations count.
4078 llvm::Value *NumIterations = CGF.EmitScalarExpr(D.getNumIterations());
4079 NumIterations = CGF.Builder.CreateIntCast(NumIterations, CGF.Int64Ty,
4080 /*IsSigned=*/false);
4081 return NumIterations;
4083 CGM.getOpenMPRuntime().emitTargetNumIterationsCall(CGF, S, Device,
4085 CGM.getOpenMPRuntime().emitTargetCall(CGF, S, Fn, FnID, IfCond, Device);
4088 static void emitTargetRegion(CodeGenFunction &CGF, const OMPTargetDirective &S,
4089 PrePostActionTy &Action) {
4091 CodeGenFunction::OMPPrivateScope PrivateScope(CGF);
4092 (void)CGF.EmitOMPFirstprivateClause(S, PrivateScope);
4093 CGF.EmitOMPPrivateClause(S, PrivateScope);
4094 (void)PrivateScope.Privatize();
4095 if (isOpenMPTargetExecutionDirective(S.getDirectiveKind()))
4096 CGF.CGM.getOpenMPRuntime().adjustTargetSpecificDataForLambdas(CGF, S);
4098 CGF.EmitStmt(S.getCapturedStmt(OMPD_target)->getCapturedStmt());
4101 void CodeGenFunction::EmitOMPTargetDeviceFunction(CodeGenModule &CGM,
4102 StringRef ParentName,
4103 const OMPTargetDirective &S) {
4104 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
4105 emitTargetRegion(CGF, S, Action);
4108 llvm::Constant *Addr;
4109 // Emit target region as a standalone region.
4110 CGM.getOpenMPRuntime().emitTargetOutlinedFunction(
4111 S, ParentName, Fn, Addr, /*IsOffloadEntry=*/true, CodeGen);
4112 assert(Fn && Addr && "Target device function emission failed.");
4115 void CodeGenFunction::EmitOMPTargetDirective(const OMPTargetDirective &S) {
4116 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
4117 emitTargetRegion(CGF, S, Action);
4119 emitCommonOMPTargetDirective(*this, S, CodeGen);
4122 static void emitCommonOMPTeamsDirective(CodeGenFunction &CGF,
4123 const OMPExecutableDirective &S,
4124 OpenMPDirectiveKind InnermostKind,
4125 const RegionCodeGenTy &CodeGen) {
4126 const CapturedStmt *CS = S.getCapturedStmt(OMPD_teams);
4127 llvm::Value *OutlinedFn =
4128 CGF.CGM.getOpenMPRuntime().emitTeamsOutlinedFunction(
4129 S, *CS->getCapturedDecl()->param_begin(), InnermostKind, CodeGen);
4131 const auto *NT = S.getSingleClause<OMPNumTeamsClause>();
4132 const auto *TL = S.getSingleClause<OMPThreadLimitClause>();
4134 const Expr *NumTeams = NT ? NT->getNumTeams() : nullptr;
4135 const Expr *ThreadLimit = TL ? TL->getThreadLimit() : nullptr;
4137 CGF.CGM.getOpenMPRuntime().emitNumTeamsClause(CGF, NumTeams, ThreadLimit,
4141 OMPTeamsScope Scope(CGF, S);
4142 llvm::SmallVector<llvm::Value *, 16> CapturedVars;
4143 CGF.GenerateOpenMPCapturedVars(*CS, CapturedVars);
4144 CGF.CGM.getOpenMPRuntime().emitTeamsCall(CGF, S, S.getBeginLoc(), OutlinedFn,
4148 void CodeGenFunction::EmitOMPTeamsDirective(const OMPTeamsDirective &S) {
4149 // Emit teams region as a standalone region.
4150 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
4152 OMPPrivateScope PrivateScope(CGF);
4153 (void)CGF.EmitOMPFirstprivateClause(S, PrivateScope);
4154 CGF.EmitOMPPrivateClause(S, PrivateScope);
4155 CGF.EmitOMPReductionClauseInit(S, PrivateScope);
4156 (void)PrivateScope.Privatize();
4157 CGF.EmitStmt(S.getCapturedStmt(OMPD_teams)->getCapturedStmt());
4158 CGF.EmitOMPReductionClauseFinal(S, /*ReductionKind=*/OMPD_teams);
4160 emitCommonOMPTeamsDirective(*this, S, OMPD_distribute, CodeGen);
4161 emitPostUpdateForReductionClause(*this, S,
4162 [](CodeGenFunction &) { return nullptr; });
4165 static void emitTargetTeamsRegion(CodeGenFunction &CGF, PrePostActionTy &Action,
4166 const OMPTargetTeamsDirective &S) {
4167 auto *CS = S.getCapturedStmt(OMPD_teams);
4169 // Emit teams region as a standalone region.
4170 auto &&CodeGen = [&S, CS](CodeGenFunction &CGF, PrePostActionTy &Action) {
4172 CodeGenFunction::OMPPrivateScope PrivateScope(CGF);
4173 (void)CGF.EmitOMPFirstprivateClause(S, PrivateScope);
4174 CGF.EmitOMPPrivateClause(S, PrivateScope);
4175 CGF.EmitOMPReductionClauseInit(S, PrivateScope);
4176 (void)PrivateScope.Privatize();
4177 if (isOpenMPTargetExecutionDirective(S.getDirectiveKind()))
4178 CGF.CGM.getOpenMPRuntime().adjustTargetSpecificDataForLambdas(CGF, S);
4179 CGF.EmitStmt(CS->getCapturedStmt());
4180 CGF.EmitOMPReductionClauseFinal(S, /*ReductionKind=*/OMPD_teams);
4182 emitCommonOMPTeamsDirective(CGF, S, OMPD_teams, CodeGen);
4183 emitPostUpdateForReductionClause(CGF, S,
4184 [](CodeGenFunction &) { return nullptr; });
4187 void CodeGenFunction::EmitOMPTargetTeamsDeviceFunction(
4188 CodeGenModule &CGM, StringRef ParentName,
4189 const OMPTargetTeamsDirective &S) {
4190 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
4191 emitTargetTeamsRegion(CGF, Action, S);
4194 llvm::Constant *Addr;
4195 // Emit target region as a standalone region.
4196 CGM.getOpenMPRuntime().emitTargetOutlinedFunction(
4197 S, ParentName, Fn, Addr, /*IsOffloadEntry=*/true, CodeGen);
4198 assert(Fn && Addr && "Target device function emission failed.");
4201 void CodeGenFunction::EmitOMPTargetTeamsDirective(
4202 const OMPTargetTeamsDirective &S) {
4203 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
4204 emitTargetTeamsRegion(CGF, Action, S);
4206 emitCommonOMPTargetDirective(*this, S, CodeGen);
4210 emitTargetTeamsDistributeRegion(CodeGenFunction &CGF, PrePostActionTy &Action,
4211 const OMPTargetTeamsDistributeDirective &S) {
4213 auto &&CodeGenDistribute = [&S](CodeGenFunction &CGF, PrePostActionTy &) {
4214 CGF.EmitOMPDistributeLoop(S, emitOMPLoopBodyWithStopPoint, S.getInc());
4217 // Emit teams region as a standalone region.
4218 auto &&CodeGen = [&S, &CodeGenDistribute](CodeGenFunction &CGF,
4219 PrePostActionTy &Action) {
4221 CodeGenFunction::OMPPrivateScope PrivateScope(CGF);
4222 CGF.EmitOMPReductionClauseInit(S, PrivateScope);
4223 (void)PrivateScope.Privatize();
4224 CGF.CGM.getOpenMPRuntime().emitInlinedDirective(CGF, OMPD_distribute,
4226 CGF.EmitOMPReductionClauseFinal(S, /*ReductionKind=*/OMPD_teams);
4228 emitCommonOMPTeamsDirective(CGF, S, OMPD_distribute, CodeGen);
4229 emitPostUpdateForReductionClause(CGF, S,
4230 [](CodeGenFunction &) { return nullptr; });
4233 void CodeGenFunction::EmitOMPTargetTeamsDistributeDeviceFunction(
4234 CodeGenModule &CGM, StringRef ParentName,
4235 const OMPTargetTeamsDistributeDirective &S) {
4236 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
4237 emitTargetTeamsDistributeRegion(CGF, Action, S);
4240 llvm::Constant *Addr;
4241 // Emit target region as a standalone region.
4242 CGM.getOpenMPRuntime().emitTargetOutlinedFunction(
4243 S, ParentName, Fn, Addr, /*IsOffloadEntry=*/true, CodeGen);
4244 assert(Fn && Addr && "Target device function emission failed.");
4247 void CodeGenFunction::EmitOMPTargetTeamsDistributeDirective(
4248 const OMPTargetTeamsDistributeDirective &S) {
4249 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
4250 emitTargetTeamsDistributeRegion(CGF, Action, S);
4252 emitCommonOMPTargetDirective(*this, S, CodeGen);
4255 static void emitTargetTeamsDistributeSimdRegion(
4256 CodeGenFunction &CGF, PrePostActionTy &Action,
4257 const OMPTargetTeamsDistributeSimdDirective &S) {
4259 auto &&CodeGenDistribute = [&S](CodeGenFunction &CGF, PrePostActionTy &) {
4260 CGF.EmitOMPDistributeLoop(S, emitOMPLoopBodyWithStopPoint, S.getInc());
4263 // Emit teams region as a standalone region.
4264 auto &&CodeGen = [&S, &CodeGenDistribute](CodeGenFunction &CGF,
4265 PrePostActionTy &Action) {
4267 CodeGenFunction::OMPPrivateScope PrivateScope(CGF);
4268 CGF.EmitOMPReductionClauseInit(S, PrivateScope);
4269 (void)PrivateScope.Privatize();
4270 CGF.CGM.getOpenMPRuntime().emitInlinedDirective(CGF, OMPD_distribute,
4272 CGF.EmitOMPReductionClauseFinal(S, /*ReductionKind=*/OMPD_teams);
4274 emitCommonOMPTeamsDirective(CGF, S, OMPD_distribute_simd, CodeGen);
4275 emitPostUpdateForReductionClause(CGF, S,
4276 [](CodeGenFunction &) { return nullptr; });
4279 void CodeGenFunction::EmitOMPTargetTeamsDistributeSimdDeviceFunction(
4280 CodeGenModule &CGM, StringRef ParentName,
4281 const OMPTargetTeamsDistributeSimdDirective &S) {
4282 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
4283 emitTargetTeamsDistributeSimdRegion(CGF, Action, S);
4286 llvm::Constant *Addr;
4287 // Emit target region as a standalone region.
4288 CGM.getOpenMPRuntime().emitTargetOutlinedFunction(
4289 S, ParentName, Fn, Addr, /*IsOffloadEntry=*/true, CodeGen);
4290 assert(Fn && Addr && "Target device function emission failed.");
4293 void CodeGenFunction::EmitOMPTargetTeamsDistributeSimdDirective(
4294 const OMPTargetTeamsDistributeSimdDirective &S) {
4295 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
4296 emitTargetTeamsDistributeSimdRegion(CGF, Action, S);
4298 emitCommonOMPTargetDirective(*this, S, CodeGen);
4301 void CodeGenFunction::EmitOMPTeamsDistributeDirective(
4302 const OMPTeamsDistributeDirective &S) {
4304 auto &&CodeGenDistribute = [&S](CodeGenFunction &CGF, PrePostActionTy &) {
4305 CGF.EmitOMPDistributeLoop(S, emitOMPLoopBodyWithStopPoint, S.getInc());
4308 // Emit teams region as a standalone region.
4309 auto &&CodeGen = [&S, &CodeGenDistribute](CodeGenFunction &CGF,
4310 PrePostActionTy &Action) {
4312 OMPPrivateScope PrivateScope(CGF);
4313 CGF.EmitOMPReductionClauseInit(S, PrivateScope);
4314 (void)PrivateScope.Privatize();
4315 CGF.CGM.getOpenMPRuntime().emitInlinedDirective(CGF, OMPD_distribute,
4317 CGF.EmitOMPReductionClauseFinal(S, /*ReductionKind=*/OMPD_teams);
4319 emitCommonOMPTeamsDirective(*this, S, OMPD_distribute, CodeGen);
4320 emitPostUpdateForReductionClause(*this, S,
4321 [](CodeGenFunction &) { return nullptr; });
4324 void CodeGenFunction::EmitOMPTeamsDistributeSimdDirective(
4325 const OMPTeamsDistributeSimdDirective &S) {
4326 auto &&CodeGenDistribute = [&S](CodeGenFunction &CGF, PrePostActionTy &) {
4327 CGF.EmitOMPDistributeLoop(S, emitOMPLoopBodyWithStopPoint, S.getInc());
4330 // Emit teams region as a standalone region.
4331 auto &&CodeGen = [&S, &CodeGenDistribute](CodeGenFunction &CGF,
4332 PrePostActionTy &Action) {
4334 OMPPrivateScope PrivateScope(CGF);
4335 CGF.EmitOMPReductionClauseInit(S, PrivateScope);
4336 (void)PrivateScope.Privatize();
4337 CGF.CGM.getOpenMPRuntime().emitInlinedDirective(CGF, OMPD_simd,
4339 CGF.EmitOMPReductionClauseFinal(S, /*ReductionKind=*/OMPD_teams);
4341 emitCommonOMPTeamsDirective(*this, S, OMPD_distribute_simd, CodeGen);
4342 emitPostUpdateForReductionClause(*this, S,
4343 [](CodeGenFunction &) { return nullptr; });
4346 void CodeGenFunction::EmitOMPTeamsDistributeParallelForDirective(
4347 const OMPTeamsDistributeParallelForDirective &S) {
4348 auto &&CodeGenDistribute = [&S](CodeGenFunction &CGF, PrePostActionTy &) {
4349 CGF.EmitOMPDistributeLoop(S, emitInnerParallelForWhenCombined,
4353 // Emit teams region as a standalone region.
4354 auto &&CodeGen = [&S, &CodeGenDistribute](CodeGenFunction &CGF,
4355 PrePostActionTy &Action) {
4357 OMPPrivateScope PrivateScope(CGF);
4358 CGF.EmitOMPReductionClauseInit(S, PrivateScope);
4359 (void)PrivateScope.Privatize();
4360 CGF.CGM.getOpenMPRuntime().emitInlinedDirective(CGF, OMPD_distribute,
4362 CGF.EmitOMPReductionClauseFinal(S, /*ReductionKind=*/OMPD_teams);
4364 emitCommonOMPTeamsDirective(*this, S, OMPD_distribute_parallel_for, CodeGen);
4365 emitPostUpdateForReductionClause(*this, S,
4366 [](CodeGenFunction &) { return nullptr; });
4369 void CodeGenFunction::EmitOMPTeamsDistributeParallelForSimdDirective(
4370 const OMPTeamsDistributeParallelForSimdDirective &S) {
4371 auto &&CodeGenDistribute = [&S](CodeGenFunction &CGF, PrePostActionTy &) {
4372 CGF.EmitOMPDistributeLoop(S, emitInnerParallelForWhenCombined,
4376 // Emit teams region as a standalone region.
4377 auto &&CodeGen = [&S, &CodeGenDistribute](CodeGenFunction &CGF,
4378 PrePostActionTy &Action) {
4380 OMPPrivateScope PrivateScope(CGF);
4381 CGF.EmitOMPReductionClauseInit(S, PrivateScope);
4382 (void)PrivateScope.Privatize();
4383 CGF.CGM.getOpenMPRuntime().emitInlinedDirective(
4384 CGF, OMPD_distribute, CodeGenDistribute, /*HasCancel=*/false);
4385 CGF.EmitOMPReductionClauseFinal(S, /*ReductionKind=*/OMPD_teams);
4387 emitCommonOMPTeamsDirective(*this, S, OMPD_distribute_parallel_for, CodeGen);
4388 emitPostUpdateForReductionClause(*this, S,
4389 [](CodeGenFunction &) { return nullptr; });
4392 static void emitTargetTeamsDistributeParallelForRegion(
4393 CodeGenFunction &CGF, const OMPTargetTeamsDistributeParallelForDirective &S,
4394 PrePostActionTy &Action) {
4396 auto &&CodeGenDistribute = [&S](CodeGenFunction &CGF, PrePostActionTy &) {
4397 CGF.EmitOMPDistributeLoop(S, emitInnerParallelForWhenCombined,
4401 // Emit teams region as a standalone region.
4402 auto &&CodeGenTeams = [&S, &CodeGenDistribute](CodeGenFunction &CGF,
4403 PrePostActionTy &Action) {
4405 CodeGenFunction::OMPPrivateScope PrivateScope(CGF);
4406 CGF.EmitOMPReductionClauseInit(S, PrivateScope);
4407 (void)PrivateScope.Privatize();
4408 CGF.CGM.getOpenMPRuntime().emitInlinedDirective(
4409 CGF, OMPD_distribute, CodeGenDistribute, /*HasCancel=*/false);
4410 CGF.EmitOMPReductionClauseFinal(S, /*ReductionKind=*/OMPD_teams);
4413 emitCommonOMPTeamsDirective(CGF, S, OMPD_distribute_parallel_for,
4415 emitPostUpdateForReductionClause(CGF, S,
4416 [](CodeGenFunction &) { return nullptr; });
4419 void CodeGenFunction::EmitOMPTargetTeamsDistributeParallelForDeviceFunction(
4420 CodeGenModule &CGM, StringRef ParentName,
4421 const OMPTargetTeamsDistributeParallelForDirective &S) {
4422 // Emit SPMD target teams distribute parallel for region as a standalone
4424 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
4425 emitTargetTeamsDistributeParallelForRegion(CGF, S, Action);
4428 llvm::Constant *Addr;
4429 // Emit target region as a standalone region.
4430 CGM.getOpenMPRuntime().emitTargetOutlinedFunction(
4431 S, ParentName, Fn, Addr, /*IsOffloadEntry=*/true, CodeGen);
4432 assert(Fn && Addr && "Target device function emission failed.");
4435 void CodeGenFunction::EmitOMPTargetTeamsDistributeParallelForDirective(
4436 const OMPTargetTeamsDistributeParallelForDirective &S) {
4437 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
4438 emitTargetTeamsDistributeParallelForRegion(CGF, S, Action);
4440 emitCommonOMPTargetDirective(*this, S, CodeGen);
4443 static void emitTargetTeamsDistributeParallelForSimdRegion(
4444 CodeGenFunction &CGF,
4445 const OMPTargetTeamsDistributeParallelForSimdDirective &S,
4446 PrePostActionTy &Action) {
4448 auto &&CodeGenDistribute = [&S](CodeGenFunction &CGF, PrePostActionTy &) {
4449 CGF.EmitOMPDistributeLoop(S, emitInnerParallelForWhenCombined,
4453 // Emit teams region as a standalone region.
4454 auto &&CodeGenTeams = [&S, &CodeGenDistribute](CodeGenFunction &CGF,
4455 PrePostActionTy &Action) {
4457 CodeGenFunction::OMPPrivateScope PrivateScope(CGF);
4458 CGF.EmitOMPReductionClauseInit(S, PrivateScope);
4459 (void)PrivateScope.Privatize();
4460 CGF.CGM.getOpenMPRuntime().emitInlinedDirective(
4461 CGF, OMPD_distribute, CodeGenDistribute, /*HasCancel=*/false);
4462 CGF.EmitOMPReductionClauseFinal(S, /*ReductionKind=*/OMPD_teams);
4465 emitCommonOMPTeamsDirective(CGF, S, OMPD_distribute_parallel_for_simd,
4467 emitPostUpdateForReductionClause(CGF, S,
4468 [](CodeGenFunction &) { return nullptr; });
4471 void CodeGenFunction::EmitOMPTargetTeamsDistributeParallelForSimdDeviceFunction(
4472 CodeGenModule &CGM, StringRef ParentName,
4473 const OMPTargetTeamsDistributeParallelForSimdDirective &S) {
4474 // Emit SPMD target teams distribute parallel for simd region as a standalone
4476 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
4477 emitTargetTeamsDistributeParallelForSimdRegion(CGF, S, Action);
4480 llvm::Constant *Addr;
4481 // Emit target region as a standalone region.
4482 CGM.getOpenMPRuntime().emitTargetOutlinedFunction(
4483 S, ParentName, Fn, Addr, /*IsOffloadEntry=*/true, CodeGen);
4484 assert(Fn && Addr && "Target device function emission failed.");
4487 void CodeGenFunction::EmitOMPTargetTeamsDistributeParallelForSimdDirective(
4488 const OMPTargetTeamsDistributeParallelForSimdDirective &S) {
4489 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
4490 emitTargetTeamsDistributeParallelForSimdRegion(CGF, S, Action);
4492 emitCommonOMPTargetDirective(*this, S, CodeGen);
4495 void CodeGenFunction::EmitOMPCancellationPointDirective(
4496 const OMPCancellationPointDirective &S) {
4497 CGM.getOpenMPRuntime().emitCancellationPointCall(*this, S.getBeginLoc(),
4498 S.getCancelRegion());
4501 void CodeGenFunction::EmitOMPCancelDirective(const OMPCancelDirective &S) {
4502 const Expr *IfCond = nullptr;
4503 for (const auto *C : S.getClausesOfKind<OMPIfClause>()) {
4504 if (C->getNameModifier() == OMPD_unknown ||
4505 C->getNameModifier() == OMPD_cancel) {
4506 IfCond = C->getCondition();
4510 CGM.getOpenMPRuntime().emitCancelCall(*this, S.getBeginLoc(), IfCond,
4511 S.getCancelRegion());
4514 CodeGenFunction::JumpDest
4515 CodeGenFunction::getOMPCancelDestination(OpenMPDirectiveKind Kind) {
4516 if (Kind == OMPD_parallel || Kind == OMPD_task ||
4517 Kind == OMPD_target_parallel)
4519 assert(Kind == OMPD_for || Kind == OMPD_section || Kind == OMPD_sections ||
4520 Kind == OMPD_parallel_sections || Kind == OMPD_parallel_for ||
4521 Kind == OMPD_distribute_parallel_for ||
4522 Kind == OMPD_target_parallel_for ||
4523 Kind == OMPD_teams_distribute_parallel_for ||
4524 Kind == OMPD_target_teams_distribute_parallel_for);
4525 return OMPCancelStack.getExitBlock();
4528 void CodeGenFunction::EmitOMPUseDevicePtrClause(
4529 const OMPClause &NC, OMPPrivateScope &PrivateScope,
4530 const llvm::DenseMap<const ValueDecl *, Address> &CaptureDeviceAddrMap) {
4531 const auto &C = cast<OMPUseDevicePtrClause>(NC);
4532 auto OrigVarIt = C.varlist_begin();
4533 auto InitIt = C.inits().begin();
4534 for (const Expr *PvtVarIt : C.private_copies()) {
4535 const auto *OrigVD = cast<VarDecl>(cast<DeclRefExpr>(*OrigVarIt)->getDecl());
4536 const auto *InitVD = cast<VarDecl>(cast<DeclRefExpr>(*InitIt)->getDecl());
4537 const auto *PvtVD = cast<VarDecl>(cast<DeclRefExpr>(PvtVarIt)->getDecl());
4539 // In order to identify the right initializer we need to match the
4540 // declaration used by the mapping logic. In some cases we may get
4541 // OMPCapturedExprDecl that refers to the original declaration.
4542 const ValueDecl *MatchingVD = OrigVD;
4543 if (const auto *OED = dyn_cast<OMPCapturedExprDecl>(MatchingVD)) {
4544 // OMPCapturedExprDecl are used to privative fields of the current
4546 const auto *ME = cast<MemberExpr>(OED->getInit());
4547 assert(isa<CXXThisExpr>(ME->getBase()) &&
4548 "Base should be the current struct!");
4549 MatchingVD = ME->getMemberDecl();
4552 // If we don't have information about the current list item, move on to
4554 auto InitAddrIt = CaptureDeviceAddrMap.find(MatchingVD);
4555 if (InitAddrIt == CaptureDeviceAddrMap.end())
4558 bool IsRegistered = PrivateScope.addPrivate(OrigVD, [this, OrigVD,
4561 // Initialize the temporary initialization variable with the address we
4562 // get from the runtime library. We have to cast the source address
4563 // because it is always a void *. References are materialized in the
4564 // privatization scope, so the initialization here disregards the fact
4565 // the original variable is a reference.
4567 getContext().getPointerType(OrigVD->getType().getNonReferenceType());
4568 llvm::Type *AddrTy = ConvertTypeForMem(AddrQTy);
4569 Address InitAddr = Builder.CreateBitCast(InitAddrIt->second, AddrTy);
4570 setAddrOfLocalVar(InitVD, InitAddr);
4572 // Emit private declaration, it will be initialized by the value we
4573 // declaration we just added to the local declarations map.
4576 // The initialization variables reached its purpose in the emission
4577 // of the previous declaration, so we don't need it anymore.
4578 LocalDeclMap.erase(InitVD);
4580 // Return the address of the private variable.
4581 return GetAddrOfLocalVar(PvtVD);
4583 assert(IsRegistered && "firstprivate var already registered as private");
4584 // Silence the warning about unused variable.
4592 // Generate the instructions for '#pragma omp target data' directive.
4593 void CodeGenFunction::EmitOMPTargetDataDirective(
4594 const OMPTargetDataDirective &S) {
4595 CGOpenMPRuntime::TargetDataInfo Info(/*RequiresDevicePointerInfo=*/true);
4597 // Create a pre/post action to signal the privatization of the device pointer.
4598 // This action can be replaced by the OpenMP runtime code generation to
4599 // deactivate privatization.
4600 bool PrivatizeDevicePointers = false;
4601 class DevicePointerPrivActionTy : public PrePostActionTy {
4602 bool &PrivatizeDevicePointers;
4605 explicit DevicePointerPrivActionTy(bool &PrivatizeDevicePointers)
4606 : PrePostActionTy(), PrivatizeDevicePointers(PrivatizeDevicePointers) {}
4607 void Enter(CodeGenFunction &CGF) override {
4608 PrivatizeDevicePointers = true;
4611 DevicePointerPrivActionTy PrivAction(PrivatizeDevicePointers);
4613 auto &&CodeGen = [&S, &Info, &PrivatizeDevicePointers](
4614 CodeGenFunction &CGF, PrePostActionTy &Action) {
4615 auto &&InnermostCodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &) {
4616 CGF.EmitStmt(S.getInnermostCapturedStmt()->getCapturedStmt());
4619 // Codegen that selects whether to generate the privatization code or not.
4620 auto &&PrivCodeGen = [&S, &Info, &PrivatizeDevicePointers,
4621 &InnermostCodeGen](CodeGenFunction &CGF,
4622 PrePostActionTy &Action) {
4623 RegionCodeGenTy RCG(InnermostCodeGen);
4624 PrivatizeDevicePointers = false;
4626 // Call the pre-action to change the status of PrivatizeDevicePointers if
4630 if (PrivatizeDevicePointers) {
4631 OMPPrivateScope PrivateScope(CGF);
4632 // Emit all instances of the use_device_ptr clause.
4633 for (const auto *C : S.getClausesOfKind<OMPUseDevicePtrClause>())
4634 CGF.EmitOMPUseDevicePtrClause(*C, PrivateScope,
4635 Info.CaptureDeviceAddrMap);
4636 (void)PrivateScope.Privatize();
4643 // Forward the provided action to the privatization codegen.
4644 RegionCodeGenTy PrivRCG(PrivCodeGen);
4645 PrivRCG.setAction(Action);
4647 // Notwithstanding the body of the region is emitted as inlined directive,
4648 // we don't use an inline scope as changes in the references inside the
4649 // region are expected to be visible outside, so we do not privative them.
4650 OMPLexicalScope Scope(CGF, S);
4651 CGF.CGM.getOpenMPRuntime().emitInlinedDirective(CGF, OMPD_target_data,
4655 RegionCodeGenTy RCG(CodeGen);
4657 // If we don't have target devices, don't bother emitting the data mapping
4659 if (CGM.getLangOpts().OMPTargetTriples.empty()) {
4664 // Check if we have any if clause associated with the directive.
4665 const Expr *IfCond = nullptr;
4666 if (const auto *C = S.getSingleClause<OMPIfClause>())
4667 IfCond = C->getCondition();
4669 // Check if we have any device clause associated with the directive.
4670 const Expr *Device = nullptr;
4671 if (const auto *C = S.getSingleClause<OMPDeviceClause>())
4672 Device = C->getDevice();
4674 // Set the action to signal privatization of device pointers.
4675 RCG.setAction(PrivAction);
4677 // Emit region code.
4678 CGM.getOpenMPRuntime().emitTargetDataCalls(*this, S, IfCond, Device, RCG,
4682 void CodeGenFunction::EmitOMPTargetEnterDataDirective(
4683 const OMPTargetEnterDataDirective &S) {
4684 // If we don't have target devices, don't bother emitting the data mapping
4686 if (CGM.getLangOpts().OMPTargetTriples.empty())
4689 // Check if we have any if clause associated with the directive.
4690 const Expr *IfCond = nullptr;
4691 if (const auto *C = S.getSingleClause<OMPIfClause>())
4692 IfCond = C->getCondition();
4694 // Check if we have any device clause associated with the directive.
4695 const Expr *Device = nullptr;
4696 if (const auto *C = S.getSingleClause<OMPDeviceClause>())
4697 Device = C->getDevice();
4699 OMPLexicalScope Scope(*this, S, OMPD_task);
4700 CGM.getOpenMPRuntime().emitTargetDataStandAloneCall(*this, S, IfCond, Device);
4703 void CodeGenFunction::EmitOMPTargetExitDataDirective(
4704 const OMPTargetExitDataDirective &S) {
4705 // If we don't have target devices, don't bother emitting the data mapping
4707 if (CGM.getLangOpts().OMPTargetTriples.empty())
4710 // Check if we have any if clause associated with the directive.
4711 const Expr *IfCond = nullptr;
4712 if (const auto *C = S.getSingleClause<OMPIfClause>())
4713 IfCond = C->getCondition();
4715 // Check if we have any device clause associated with the directive.
4716 const Expr *Device = nullptr;
4717 if (const auto *C = S.getSingleClause<OMPDeviceClause>())
4718 Device = C->getDevice();
4720 OMPLexicalScope Scope(*this, S, OMPD_task);
4721 CGM.getOpenMPRuntime().emitTargetDataStandAloneCall(*this, S, IfCond, Device);
4724 static void emitTargetParallelRegion(CodeGenFunction &CGF,
4725 const OMPTargetParallelDirective &S,
4726 PrePostActionTy &Action) {
4727 // Get the captured statement associated with the 'parallel' region.
4728 const CapturedStmt *CS = S.getCapturedStmt(OMPD_parallel);
4730 auto &&CodeGen = [&S, CS](CodeGenFunction &CGF, PrePostActionTy &Action) {
4732 CodeGenFunction::OMPPrivateScope PrivateScope(CGF);
4733 (void)CGF.EmitOMPFirstprivateClause(S, PrivateScope);
4734 CGF.EmitOMPPrivateClause(S, PrivateScope);
4735 CGF.EmitOMPReductionClauseInit(S, PrivateScope);
4736 (void)PrivateScope.Privatize();
4737 if (isOpenMPTargetExecutionDirective(S.getDirectiveKind()))
4738 CGF.CGM.getOpenMPRuntime().adjustTargetSpecificDataForLambdas(CGF, S);
4739 // TODO: Add support for clauses.
4740 CGF.EmitStmt(CS->getCapturedStmt());
4741 CGF.EmitOMPReductionClauseFinal(S, /*ReductionKind=*/OMPD_parallel);
4743 emitCommonOMPParallelDirective(CGF, S, OMPD_parallel, CodeGen,
4744 emitEmptyBoundParameters);
4745 emitPostUpdateForReductionClause(CGF, S,
4746 [](CodeGenFunction &) { return nullptr; });
4749 void CodeGenFunction::EmitOMPTargetParallelDeviceFunction(
4750 CodeGenModule &CGM, StringRef ParentName,
4751 const OMPTargetParallelDirective &S) {
4752 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
4753 emitTargetParallelRegion(CGF, S, Action);
4756 llvm::Constant *Addr;
4757 // Emit target region as a standalone region.
4758 CGM.getOpenMPRuntime().emitTargetOutlinedFunction(
4759 S, ParentName, Fn, Addr, /*IsOffloadEntry=*/true, CodeGen);
4760 assert(Fn && Addr && "Target device function emission failed.");
4763 void CodeGenFunction::EmitOMPTargetParallelDirective(
4764 const OMPTargetParallelDirective &S) {
4765 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
4766 emitTargetParallelRegion(CGF, S, Action);
4768 emitCommonOMPTargetDirective(*this, S, CodeGen);
4771 static void emitTargetParallelForRegion(CodeGenFunction &CGF,
4772 const OMPTargetParallelForDirective &S,
4773 PrePostActionTy &Action) {
4775 // Emit directive as a combined directive that consists of two implicit
4776 // directives: 'parallel' with 'for' directive.
4777 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
4779 CodeGenFunction::OMPCancelStackRAII CancelRegion(
4780 CGF, OMPD_target_parallel_for, S.hasCancel());
4781 CGF.EmitOMPWorksharingLoop(S, S.getEnsureUpperBound(), emitForLoopBounds,
4782 emitDispatchForLoopBounds);
4784 emitCommonOMPParallelDirective(CGF, S, OMPD_for, CodeGen,
4785 emitEmptyBoundParameters);
4788 void CodeGenFunction::EmitOMPTargetParallelForDeviceFunction(
4789 CodeGenModule &CGM, StringRef ParentName,
4790 const OMPTargetParallelForDirective &S) {
4791 // Emit SPMD target parallel for region as a standalone region.
4792 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
4793 emitTargetParallelForRegion(CGF, S, Action);
4796 llvm::Constant *Addr;
4797 // Emit target region as a standalone region.
4798 CGM.getOpenMPRuntime().emitTargetOutlinedFunction(
4799 S, ParentName, Fn, Addr, /*IsOffloadEntry=*/true, CodeGen);
4800 assert(Fn && Addr && "Target device function emission failed.");
4803 void CodeGenFunction::EmitOMPTargetParallelForDirective(
4804 const OMPTargetParallelForDirective &S) {
4805 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
4806 emitTargetParallelForRegion(CGF, S, Action);
4808 emitCommonOMPTargetDirective(*this, S, CodeGen);
4812 emitTargetParallelForSimdRegion(CodeGenFunction &CGF,
4813 const OMPTargetParallelForSimdDirective &S,
4814 PrePostActionTy &Action) {
4816 // Emit directive as a combined directive that consists of two implicit
4817 // directives: 'parallel' with 'for' directive.
4818 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
4820 CGF.EmitOMPWorksharingLoop(S, S.getEnsureUpperBound(), emitForLoopBounds,
4821 emitDispatchForLoopBounds);
4823 emitCommonOMPParallelDirective(CGF, S, OMPD_simd, CodeGen,
4824 emitEmptyBoundParameters);
4827 void CodeGenFunction::EmitOMPTargetParallelForSimdDeviceFunction(
4828 CodeGenModule &CGM, StringRef ParentName,
4829 const OMPTargetParallelForSimdDirective &S) {
4830 // Emit SPMD target parallel for region as a standalone region.
4831 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
4832 emitTargetParallelForSimdRegion(CGF, S, Action);
4835 llvm::Constant *Addr;
4836 // Emit target region as a standalone region.
4837 CGM.getOpenMPRuntime().emitTargetOutlinedFunction(
4838 S, ParentName, Fn, Addr, /*IsOffloadEntry=*/true, CodeGen);
4839 assert(Fn && Addr && "Target device function emission failed.");
4842 void CodeGenFunction::EmitOMPTargetParallelForSimdDirective(
4843 const OMPTargetParallelForSimdDirective &S) {
4844 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
4845 emitTargetParallelForSimdRegion(CGF, S, Action);
4847 emitCommonOMPTargetDirective(*this, S, CodeGen);
4850 /// Emit a helper variable and return corresponding lvalue.
4851 static void mapParam(CodeGenFunction &CGF, const DeclRefExpr *Helper,
4852 const ImplicitParamDecl *PVD,
4853 CodeGenFunction::OMPPrivateScope &Privates) {
4854 const auto *VDecl = cast<VarDecl>(Helper->getDecl());
4855 Privates.addPrivate(VDecl,
4856 [&CGF, PVD]() { return CGF.GetAddrOfLocalVar(PVD); });
4859 void CodeGenFunction::EmitOMPTaskLoopBasedDirective(const OMPLoopDirective &S) {
4860 assert(isOpenMPTaskLoopDirective(S.getDirectiveKind()));
4861 // Emit outlined function for task construct.
4862 const CapturedStmt *CS = S.getCapturedStmt(OMPD_taskloop);
4863 Address CapturedStruct = GenerateCapturedStmtArgument(*CS);
4864 QualType SharedsTy = getContext().getRecordType(CS->getCapturedRecordDecl());
4865 const Expr *IfCond = nullptr;
4866 for (const auto *C : S.getClausesOfKind<OMPIfClause>()) {
4867 if (C->getNameModifier() == OMPD_unknown ||
4868 C->getNameModifier() == OMPD_taskloop) {
4869 IfCond = C->getCondition();
4875 // Check if taskloop must be emitted without taskgroup.
4876 Data.Nogroup = S.getSingleClause<OMPNogroupClause>();
4877 // TODO: Check if we should emit tied or untied task.
4879 // Set scheduling for taskloop
4880 if (const auto* Clause = S.getSingleClause<OMPGrainsizeClause>()) {
4882 Data.Schedule.setInt(/*IntVal=*/false);
4883 Data.Schedule.setPointer(EmitScalarExpr(Clause->getGrainsize()));
4884 } else if (const auto* Clause = S.getSingleClause<OMPNumTasksClause>()) {
4886 Data.Schedule.setInt(/*IntVal=*/true);
4887 Data.Schedule.setPointer(EmitScalarExpr(Clause->getNumTasks()));
4890 auto &&BodyGen = [CS, &S](CodeGenFunction &CGF, PrePostActionTy &) {
4892 // for (IV in 0..LastIteration) BODY;
4893 // <Final counter/linear vars updates>;
4897 // Emit: if (PreCond) - begin.
4898 // If the condition constant folds and can be elided, avoid emitting the
4901 llvm::BasicBlock *ContBlock = nullptr;
4902 OMPLoopScope PreInitScope(CGF, S);
4903 if (CGF.ConstantFoldsToSimpleInteger(S.getPreCond(), CondConstant)) {
4907 llvm::BasicBlock *ThenBlock = CGF.createBasicBlock("taskloop.if.then");
4908 ContBlock = CGF.createBasicBlock("taskloop.if.end");
4909 emitPreCond(CGF, S, S.getPreCond(), ThenBlock, ContBlock,
4910 CGF.getProfileCount(&S));
4911 CGF.EmitBlock(ThenBlock);
4912 CGF.incrementProfileCounter(&S);
4915 if (isOpenMPSimdDirective(S.getDirectiveKind()))
4916 CGF.EmitOMPSimdInit(S);
4918 OMPPrivateScope LoopScope(CGF);
4919 // Emit helper vars inits.
4920 enum { LowerBound = 5, UpperBound, Stride, LastIter };
4921 auto *I = CS->getCapturedDecl()->param_begin();
4922 auto *LBP = std::next(I, LowerBound);
4923 auto *UBP = std::next(I, UpperBound);
4924 auto *STP = std::next(I, Stride);
4925 auto *LIP = std::next(I, LastIter);
4926 mapParam(CGF, cast<DeclRefExpr>(S.getLowerBoundVariable()), *LBP,
4928 mapParam(CGF, cast<DeclRefExpr>(S.getUpperBoundVariable()), *UBP,
4930 mapParam(CGF, cast<DeclRefExpr>(S.getStrideVariable()), *STP, LoopScope);
4931 mapParam(CGF, cast<DeclRefExpr>(S.getIsLastIterVariable()), *LIP,
4933 CGF.EmitOMPPrivateLoopCounters(S, LoopScope);
4934 bool HasLastprivateClause = CGF.EmitOMPLastprivateClauseInit(S, LoopScope);
4935 (void)LoopScope.Privatize();
4936 // Emit the loop iteration variable.
4937 const Expr *IVExpr = S.getIterationVariable();
4938 const auto *IVDecl = cast<VarDecl>(cast<DeclRefExpr>(IVExpr)->getDecl());
4939 CGF.EmitVarDecl(*IVDecl);
4940 CGF.EmitIgnoredExpr(S.getInit());
4942 // Emit the iterations count variable.
4943 // If it is not a variable, Sema decided to calculate iterations count on
4944 // each iteration (e.g., it is foldable into a constant).
4945 if (const auto *LIExpr = dyn_cast<DeclRefExpr>(S.getLastIteration())) {
4946 CGF.EmitVarDecl(*cast<VarDecl>(LIExpr->getDecl()));
4947 // Emit calculation of the iterations count.
4948 CGF.EmitIgnoredExpr(S.getCalcLastIteration());
4951 CGF.EmitOMPInnerLoop(S, LoopScope.requiresCleanups(), S.getCond(),
4953 [&S](CodeGenFunction &CGF) {
4954 CGF.EmitOMPLoopBody(S, JumpDest());
4955 CGF.EmitStopPoint(&S);
4957 [](CodeGenFunction &) {});
4958 // Emit: if (PreCond) - end.
4960 CGF.EmitBranch(ContBlock);
4961 CGF.EmitBlock(ContBlock, true);
4963 // Emit final copy of the lastprivate variables if IsLastIter != 0.
4964 if (HasLastprivateClause) {
4965 CGF.EmitOMPLastprivateClauseFinal(
4966 S, isOpenMPSimdDirective(S.getDirectiveKind()),
4967 CGF.Builder.CreateIsNotNull(CGF.EmitLoadOfScalar(
4968 CGF.GetAddrOfLocalVar(*LIP), /*Volatile=*/false,
4969 (*LIP)->getType(), S.getBeginLoc())));
4972 auto &&TaskGen = [&S, SharedsTy, CapturedStruct,
4973 IfCond](CodeGenFunction &CGF, llvm::Value *OutlinedFn,
4974 const OMPTaskDataTy &Data) {
4975 auto &&CodeGen = [&S, OutlinedFn, SharedsTy, CapturedStruct, IfCond,
4976 &Data](CodeGenFunction &CGF, PrePostActionTy &) {
4977 OMPLoopScope PreInitScope(CGF, S);
4978 CGF.CGM.getOpenMPRuntime().emitTaskLoopCall(CGF, S.getBeginLoc(), S,
4979 OutlinedFn, SharedsTy,
4980 CapturedStruct, IfCond, Data);
4982 CGF.CGM.getOpenMPRuntime().emitInlinedDirective(CGF, OMPD_taskloop,
4986 EmitOMPTaskBasedDirective(S, OMPD_taskloop, BodyGen, TaskGen, Data);
4988 CGM.getOpenMPRuntime().emitTaskgroupRegion(
4990 [&S, &BodyGen, &TaskGen, &Data](CodeGenFunction &CGF,
4991 PrePostActionTy &Action) {
4993 CGF.EmitOMPTaskBasedDirective(S, OMPD_taskloop, BodyGen, TaskGen,
5000 void CodeGenFunction::EmitOMPTaskLoopDirective(const OMPTaskLoopDirective &S) {
5001 EmitOMPTaskLoopBasedDirective(S);
5004 void CodeGenFunction::EmitOMPTaskLoopSimdDirective(
5005 const OMPTaskLoopSimdDirective &S) {
5006 EmitOMPTaskLoopBasedDirective(S);
5009 // Generate the instructions for '#pragma omp target update' directive.
5010 void CodeGenFunction::EmitOMPTargetUpdateDirective(
5011 const OMPTargetUpdateDirective &S) {
5012 // If we don't have target devices, don't bother emitting the data mapping
5014 if (CGM.getLangOpts().OMPTargetTriples.empty())
5017 // Check if we have any if clause associated with the directive.
5018 const Expr *IfCond = nullptr;
5019 if (const auto *C = S.getSingleClause<OMPIfClause>())
5020 IfCond = C->getCondition();
5022 // Check if we have any device clause associated with the directive.
5023 const Expr *Device = nullptr;
5024 if (const auto *C = S.getSingleClause<OMPDeviceClause>())
5025 Device = C->getDevice();
5027 OMPLexicalScope Scope(*this, S, OMPD_task);
5028 CGM.getOpenMPRuntime().emitTargetDataStandAloneCall(*this, S, IfCond, Device);
5031 void CodeGenFunction::EmitSimpleOMPExecutableDirective(
5032 const OMPExecutableDirective &D) {
5033 if (!D.hasAssociatedStmt() || !D.getAssociatedStmt())
5035 auto &&CodeGen = [&D](CodeGenFunction &CGF, PrePostActionTy &Action) {
5036 if (isOpenMPSimdDirective(D.getDirectiveKind())) {
5037 emitOMPSimdRegion(CGF, cast<OMPLoopDirective>(D), Action);
5039 OMPPrivateScope LoopGlobals(CGF);
5040 if (const auto *LD = dyn_cast<OMPLoopDirective>(&D)) {
5041 for (const Expr *E : LD->counters()) {
5042 const auto *VD = dyn_cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl());
5043 if (!VD->hasLocalStorage() && !CGF.LocalDeclMap.count(VD)) {
5044 LValue GlobLVal = CGF.EmitLValue(E);
5045 LoopGlobals.addPrivate(
5046 VD, [&GlobLVal]() { return GlobLVal.getAddress(); });
5048 if (isa<OMPCapturedExprDecl>(VD)) {
5049 // Emit only those that were not explicitly referenced in clauses.
5050 if (!CGF.LocalDeclMap.count(VD))
5051 CGF.EmitVarDecl(*VD);
5054 for (const auto *C : D.getClausesOfKind<OMPOrderedClause>()) {
5055 if (!C->getNumForLoops())
5057 for (unsigned I = LD->getCollapsedNumber(),
5058 E = C->getLoopNumIterations().size();
5060 if (const auto *VD = dyn_cast<OMPCapturedExprDecl>(
5061 cast<DeclRefExpr>(C->getLoopCounter(I))->getDecl())) {
5062 // Emit only those that were not explicitly referenced in clauses.
5063 if (!CGF.LocalDeclMap.count(VD))
5064 CGF.EmitVarDecl(*VD);
5069 LoopGlobals.Privatize();
5070 CGF.EmitStmt(D.getInnermostCapturedStmt()->getCapturedStmt());
5073 OMPSimdLexicalScope Scope(*this, D);
5074 CGM.getOpenMPRuntime().emitInlinedDirective(
5076 isOpenMPSimdDirective(D.getDirectiveKind()) ? OMPD_simd
5077 : D.getDirectiveKind(),