blob: 6ece69d51daf0b87ac4af2c8fae4aa8bba78cd32 [file] [log] [blame]
//===--- CGStmtOpenMP.cpp - Emit LLVM Code from Statements ----------------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
//
// This contains code to emit OpenMP nodes as LLVM code.
//
//===----------------------------------------------------------------------===//
#include "CGCleanup.h"
#include "CGOpenMPRuntime.h"
#include "CodeGenFunction.h"
#include "CodeGenModule.h"
#include "TargetInfo.h"
#include "clang/AST/Stmt.h"
#include "clang/AST/StmtOpenMP.h"
#include "clang/AST/DeclOpenMP.h"
using namespace clang;
using namespace CodeGen;
namespace {
/// Lexical scope for OpenMP executable constructs, that handles correct codegen
/// for captured expressions.
class OMPLexicalScope : public CodeGenFunction::LexicalScope {
void emitPreInitStmt(CodeGenFunction &CGF, const OMPExecutableDirective &S) {
for (const auto *C : S.clauses()) {
if (const auto *CPI = OMPClauseWithPreInit::get(C)) {
if (const auto *PreInit =
cast_or_null<DeclStmt>(CPI->getPreInitStmt())) {
for (const auto *I : PreInit->decls()) {
if (!I->hasAttr<OMPCaptureNoInitAttr>()) {
CGF.EmitVarDecl(cast<VarDecl>(*I));
} else {
CodeGenFunction::AutoVarEmission Emission =
CGF.EmitAutoVarAlloca(cast<VarDecl>(*I));
CGF.EmitAutoVarCleanups(Emission);
}
}
}
}
}
}
CodeGenFunction::OMPPrivateScope InlinedShareds;
static bool isCapturedVar(CodeGenFunction &CGF, const VarDecl *VD) {
return CGF.LambdaCaptureFields.lookup(VD) ||
(CGF.CapturedStmtInfo && CGF.CapturedStmtInfo->lookup(VD)) ||
(CGF.CurCodeDecl && isa<BlockDecl>(CGF.CurCodeDecl));
}
public:
OMPLexicalScope(
CodeGenFunction &CGF, const OMPExecutableDirective &S,
const llvm::Optional<OpenMPDirectiveKind> CapturedRegion = llvm::None,
const bool EmitPreInitStmt = true)
: CodeGenFunction::LexicalScope(CGF, S.getSourceRange()),
InlinedShareds(CGF) {
if (EmitPreInitStmt)
emitPreInitStmt(CGF, S);
if (!CapturedRegion.hasValue())
return;
assert(S.hasAssociatedStmt() &&
"Expected associated statement for inlined directive.");
const CapturedStmt *CS = S.getCapturedStmt(*CapturedRegion);
for (const auto &C : CS->captures()) {
if (C.capturesVariable() || C.capturesVariableByCopy()) {
auto *VD = C.getCapturedVar();
assert(VD == VD->getCanonicalDecl() &&
"Canonical decl must be captured.");
DeclRefExpr DRE(
CGF.getContext(), const_cast<VarDecl *>(VD),
isCapturedVar(CGF, VD) || (CGF.CapturedStmtInfo &&
InlinedShareds.isGlobalVarCaptured(VD)),
VD->getType().getNonReferenceType(), VK_LValue, C.getLocation());
InlinedShareds.addPrivate(VD, [&CGF, &DRE]() -> Address {
return CGF.EmitLValue(&DRE).getAddress();
});
}
}
(void)InlinedShareds.Privatize();
}
};
/// Lexical scope for OpenMP parallel construct, that handles correct codegen
/// for captured expressions.
class OMPParallelScope final : public OMPLexicalScope {
bool EmitPreInitStmt(const OMPExecutableDirective &S) {
OpenMPDirectiveKind Kind = S.getDirectiveKind();
return !(isOpenMPTargetExecutionDirective(Kind) ||
isOpenMPLoopBoundSharingDirective(Kind)) &&
isOpenMPParallelDirective(Kind);
}
public:
OMPParallelScope(CodeGenFunction &CGF, const OMPExecutableDirective &S)
: OMPLexicalScope(CGF, S, /*CapturedRegion=*/llvm::None,
EmitPreInitStmt(S)) {}
};
/// Lexical scope for OpenMP teams construct, that handles correct codegen
/// for captured expressions.
class OMPTeamsScope final : public OMPLexicalScope {
bool EmitPreInitStmt(const OMPExecutableDirective &S) {
OpenMPDirectiveKind Kind = S.getDirectiveKind();
return !isOpenMPTargetExecutionDirective(Kind) &&
isOpenMPTeamsDirective(Kind);
}
public:
OMPTeamsScope(CodeGenFunction &CGF, const OMPExecutableDirective &S)
: OMPLexicalScope(CGF, S, /*CapturedRegion=*/llvm::None,
EmitPreInitStmt(S)) {}
};
/// Private scope for OpenMP loop-based directives, that supports capturing
/// of used expression from loop statement.
class OMPLoopScope : public CodeGenFunction::RunCleanupsScope {
void emitPreInitStmt(CodeGenFunction &CGF, const OMPLoopDirective &S) {
CodeGenFunction::OMPMapVars PreCondVars;
llvm::DenseSet<const VarDecl *> EmittedAsPrivate;
for (const auto *E : S.counters()) {
const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl());
EmittedAsPrivate.insert(VD->getCanonicalDecl());
(void)PreCondVars.setVarAddr(
CGF, VD, CGF.CreateMemTemp(VD->getType().getNonReferenceType()));
}
// Mark private vars as undefs.
for (const auto *C : S.getClausesOfKind<OMPPrivateClause>()) {
for (const Expr *IRef : C->varlists()) {
const auto *OrigVD = cast<VarDecl>(cast<DeclRefExpr>(IRef)->getDecl());
if (EmittedAsPrivate.insert(OrigVD->getCanonicalDecl()).second) {
(void)PreCondVars.setVarAddr(
CGF, OrigVD,
Address(llvm::UndefValue::get(
CGF.ConvertTypeForMem(CGF.getContext().getPointerType(
OrigVD->getType().getNonReferenceType()))),
CGF.getContext().getDeclAlign(OrigVD)));
}
}
}
(void)PreCondVars.apply(CGF);
// Emit init, __range and __end variables for C++ range loops.
const Stmt *Body =
S.getInnermostCapturedStmt()->getCapturedStmt()->IgnoreContainers();
for (unsigned Cnt = 0; Cnt < S.getCollapsedNumber(); ++Cnt) {
Body = Body->IgnoreContainers();
if (auto *For = dyn_cast<ForStmt>(Body)) {
Body = For->getBody();
} else {
assert(isa<CXXForRangeStmt>(Body) &&
"Expected canonical for loop or range-based for loop.");
auto *CXXFor = cast<CXXForRangeStmt>(Body);
if (const Stmt *Init = CXXFor->getInit())
CGF.EmitStmt(Init);
CGF.EmitStmt(CXXFor->getRangeStmt());
CGF.EmitStmt(CXXFor->getEndStmt());
Body = CXXFor->getBody();
}
}
if (const auto *PreInits = cast_or_null<DeclStmt>(S.getPreInits())) {
for (const auto *I : PreInits->decls())
CGF.EmitVarDecl(cast<VarDecl>(*I));
}
PreCondVars.restore(CGF);
}
public:
OMPLoopScope(CodeGenFunction &CGF, const OMPLoopDirective &S)
: CodeGenFunction::RunCleanupsScope(CGF) {
emitPreInitStmt(CGF, S);
}
};
class OMPSimdLexicalScope : public CodeGenFunction::LexicalScope {
CodeGenFunction::OMPPrivateScope InlinedShareds;
static bool isCapturedVar(CodeGenFunction &CGF, const VarDecl *VD) {
return CGF.LambdaCaptureFields.lookup(VD) ||
(CGF.CapturedStmtInfo && CGF.CapturedStmtInfo->lookup(VD)) ||
(CGF.CurCodeDecl && isa<BlockDecl>(CGF.CurCodeDecl) &&
cast<BlockDecl>(CGF.CurCodeDecl)->capturesVariable(VD));
}
public:
OMPSimdLexicalScope(CodeGenFunction &CGF, const OMPExecutableDirective &S)
: CodeGenFunction::LexicalScope(CGF, S.getSourceRange()),
InlinedShareds(CGF) {
for (const auto *C : S.clauses()) {
if (const auto *CPI = OMPClauseWithPreInit::get(C)) {
if (const auto *PreInit =
cast_or_null<DeclStmt>(CPI->getPreInitStmt())) {
for (const auto *I : PreInit->decls()) {
if (!I->hasAttr<OMPCaptureNoInitAttr>()) {
CGF.EmitVarDecl(cast<VarDecl>(*I));
} else {
CodeGenFunction::AutoVarEmission Emission =
CGF.EmitAutoVarAlloca(cast<VarDecl>(*I));
CGF.EmitAutoVarCleanups(Emission);
}
}
}
} else if (const auto *UDP = dyn_cast<OMPUseDevicePtrClause>(C)) {
for (const Expr *E : UDP->varlists()) {
const Decl *D = cast<DeclRefExpr>(E)->getDecl();
if (const auto *OED = dyn_cast<OMPCapturedExprDecl>(D))
CGF.EmitVarDecl(*OED);
}
}
}
if (!isOpenMPSimdDirective(S.getDirectiveKind()))
CGF.EmitOMPPrivateClause(S, InlinedShareds);
if (const auto *TG = dyn_cast<OMPTaskgroupDirective>(&S)) {
if (const Expr *E = TG->getReductionRef())
CGF.EmitVarDecl(*cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl()));
}
const auto *CS = cast_or_null<CapturedStmt>(S.getAssociatedStmt());
while (CS) {
for (auto &C : CS->captures()) {
if (C.capturesVariable() || C.capturesVariableByCopy()) {
auto *VD = C.getCapturedVar();
assert(VD == VD->getCanonicalDecl() &&
"Canonical decl must be captured.");
DeclRefExpr DRE(CGF.getContext(), const_cast<VarDecl *>(VD),
isCapturedVar(CGF, VD) ||
(CGF.CapturedStmtInfo &&
InlinedShareds.isGlobalVarCaptured(VD)),
VD->getType().getNonReferenceType(), VK_LValue,
C.getLocation());
InlinedShareds.addPrivate(VD, [&CGF, &DRE]() -> Address {
return CGF.EmitLValue(&DRE).getAddress();
});
}
}
CS = dyn_cast<CapturedStmt>(CS->getCapturedStmt());
}
(void)InlinedShareds.Privatize();
}
};
} // namespace
static void emitCommonOMPTargetDirective(CodeGenFunction &CGF,
const OMPExecutableDirective &S,
const RegionCodeGenTy &CodeGen);
LValue CodeGenFunction::EmitOMPSharedLValue(const Expr *E) {
if (const auto *OrigDRE = dyn_cast<DeclRefExpr>(E)) {
if (const auto *OrigVD = dyn_cast<VarDecl>(OrigDRE->getDecl())) {
OrigVD = OrigVD->getCanonicalDecl();
bool IsCaptured =
LambdaCaptureFields.lookup(OrigVD) ||
(CapturedStmtInfo && CapturedStmtInfo->lookup(OrigVD)) ||
(CurCodeDecl && isa<BlockDecl>(CurCodeDecl));
DeclRefExpr DRE(getContext(), const_cast<VarDecl *>(OrigVD), IsCaptured,
OrigDRE->getType(), VK_LValue, OrigDRE->getExprLoc());
return EmitLValue(&DRE);
}
}
return EmitLValue(E);
}
llvm::Value *CodeGenFunction::getTypeSize(QualType Ty) {
ASTContext &C = getContext();
llvm::Value *Size = nullptr;
auto SizeInChars = C.getTypeSizeInChars(Ty);
if (SizeInChars.isZero()) {
// getTypeSizeInChars() returns 0 for a VLA.
while (const VariableArrayType *VAT = C.getAsVariableArrayType(Ty)) {
VlaSizePair VlaSize = getVLASize(VAT);
Ty = VlaSize.Type;
Size = Size ? Builder.CreateNUWMul(Size, VlaSize.NumElts)
: VlaSize.NumElts;
}
SizeInChars = C.getTypeSizeInChars(Ty);
if (SizeInChars.isZero())
return llvm::ConstantInt::get(SizeTy, /*V=*/0);
return Builder.CreateNUWMul(Size, CGM.getSize(SizeInChars));
}
return CGM.getSize(SizeInChars);
}
void CodeGenFunction::GenerateOpenMPCapturedVars(
const CapturedStmt &S, SmallVectorImpl<llvm::Value *> &CapturedVars) {
const RecordDecl *RD = S.getCapturedRecordDecl();
auto CurField = RD->field_begin();
auto CurCap = S.captures().begin();
for (CapturedStmt::const_capture_init_iterator I = S.capture_init_begin(),
E = S.capture_init_end();
I != E; ++I, ++CurField, ++CurCap) {
if (CurField->hasCapturedVLAType()) {
const VariableArrayType *VAT = CurField->getCapturedVLAType();
llvm::Value *Val = VLASizeMap[VAT->getSizeExpr()];
CapturedVars.push_back(Val);
} else if (CurCap->capturesThis()) {
CapturedVars.push_back(CXXThisValue);
} else if (CurCap->capturesVariableByCopy()) {
llvm::Value *CV = EmitLoadOfScalar(EmitLValue(*I), CurCap->getLocation());
// If the field is not a pointer, we need to save the actual value
// and load it as a void pointer.
if (!CurField->getType()->isAnyPointerType()) {
ASTContext &Ctx = getContext();
Address DstAddr = CreateMemTemp(
Ctx.getUIntPtrType(),
Twine(CurCap->getCapturedVar()->getName(), ".casted"));
LValue DstLV = MakeAddrLValue(DstAddr, Ctx.getUIntPtrType());
llvm::Value *SrcAddrVal = EmitScalarConversion(
DstAddr.getPointer(), Ctx.getPointerType(Ctx.getUIntPtrType()),
Ctx.getPointerType(CurField->getType()), CurCap->getLocation());
LValue SrcLV =
MakeNaturalAlignAddrLValue(SrcAddrVal, CurField->getType());
// Store the value using the source type pointer.
EmitStoreThroughLValue(RValue::get(CV), SrcLV);
// Load the value using the destination type pointer.
CV = EmitLoadOfScalar(DstLV, CurCap->getLocation());
}
CapturedVars.push_back(CV);
} else {
assert(CurCap->capturesVariable() && "Expected capture by reference.");
CapturedVars.push_back(EmitLValue(*I).getAddress().getPointer());
}
}
}
static Address castValueFromUintptr(CodeGenFunction &CGF, SourceLocation Loc,
QualType DstType, StringRef Name,
LValue AddrLV) {
ASTContext &Ctx = CGF.getContext();
llvm::Value *CastedPtr = CGF.EmitScalarConversion(
AddrLV.getAddress().getPointer(), Ctx.getUIntPtrType(),
Ctx.getPointerType(DstType), Loc);
Address TmpAddr =
CGF.MakeNaturalAlignAddrLValue(CastedPtr, Ctx.getPointerType(DstType))
.getAddress();
return TmpAddr;
}
static QualType getCanonicalParamType(ASTContext &C, QualType T) {
if (T->isLValueReferenceType())
return C.getLValueReferenceType(
getCanonicalParamType(C, T.getNonReferenceType()),
/*SpelledAsLValue=*/false);
if (T->isPointerType())
return C.getPointerType(getCanonicalParamType(C, T->getPointeeType()));
if (const ArrayType *A = T->getAsArrayTypeUnsafe()) {
if (const auto *VLA = dyn_cast<VariableArrayType>(A))
return getCanonicalParamType(C, VLA->getElementType());
if (!A->isVariablyModifiedType())
return C.getCanonicalType(T);
}
return C.getCanonicalParamType(T);
}
namespace {
/// Contains required data for proper outlined function codegen.
struct FunctionOptions {
/// Captured statement for which the function is generated.
const CapturedStmt *S = nullptr;
/// true if cast to/from UIntPtr is required for variables captured by
/// value.
const bool UIntPtrCastRequired = true;
/// true if only casted arguments must be registered as local args or VLA
/// sizes.
const bool RegisterCastedArgsOnly = false;
/// Name of the generated function.
const StringRef FunctionName;
explicit FunctionOptions(const CapturedStmt *S, bool UIntPtrCastRequired,
bool RegisterCastedArgsOnly,
StringRef FunctionName)
: S(S), UIntPtrCastRequired(UIntPtrCastRequired),
RegisterCastedArgsOnly(UIntPtrCastRequired && RegisterCastedArgsOnly),
FunctionName(FunctionName) {}
};
}
static llvm::Function *emitOutlinedFunctionPrologue(
CodeGenFunction &CGF, FunctionArgList &Args,
llvm::MapVector<const Decl *, std::pair<const VarDecl *, Address>>
&LocalAddrs,
llvm::DenseMap<const Decl *, std::pair<const Expr *, llvm::Value *>>
&VLASizes,
llvm::Value *&CXXThisValue, const FunctionOptions &FO) {
const CapturedDecl *CD = FO.S->getCapturedDecl();
const RecordDecl *RD = FO.S->getCapturedRecordDecl();
assert(CD->hasBody() && "missing CapturedDecl body");
CXXThisValue = nullptr;
// Build the argument list.
CodeGenModule &CGM = CGF.CGM;
ASTContext &Ctx = CGM.getContext();
FunctionArgList TargetArgs;
Args.append(CD->param_begin(),
std::next(CD->param_begin(), CD->getContextParamPosition()));
TargetArgs.append(
CD->param_begin(),
std::next(CD->param_begin(), CD->getContextParamPosition()));
auto I = FO.S->captures().begin();
FunctionDecl *DebugFunctionDecl = nullptr;
if (!FO.UIntPtrCastRequired) {
FunctionProtoType::ExtProtoInfo EPI;
QualType FunctionTy = Ctx.getFunctionType(Ctx.VoidTy, llvm::None, EPI);
DebugFunctionDecl = FunctionDecl::Create(
Ctx, Ctx.getTranslationUnitDecl(), FO.S->getBeginLoc(),
SourceLocation(), DeclarationName(), FunctionTy,
Ctx.getTrivialTypeSourceInfo(FunctionTy), SC_Static,
/*isInlineSpecified=*/false, /*hasWrittenPrototype=*/false);
}
for (const FieldDecl *FD : RD->fields()) {
QualType ArgType = FD->getType();
IdentifierInfo *II = nullptr;
VarDecl *CapVar = nullptr;
// If this is a capture by copy and the type is not a pointer, the outlined
// function argument type should be uintptr and the value properly casted to
// uintptr. This is necessary given that the runtime library is only able to
// deal with pointers. We can pass in the same way the VLA type sizes to the
// outlined function.
if (FO.UIntPtrCastRequired &&
((I->capturesVariableByCopy() && !ArgType->isAnyPointerType()) ||
I->capturesVariableArrayType()))
ArgType = Ctx.getUIntPtrType();
if (I->capturesVariable() || I->capturesVariableByCopy()) {
CapVar = I->getCapturedVar();
II = CapVar->getIdentifier();
} else if (I->capturesThis()) {
II = &Ctx.Idents.get("this");
} else {
assert(I->capturesVariableArrayType());
II = &Ctx.Idents.get("vla");
}
if (ArgType->isVariablyModifiedType())
ArgType = getCanonicalParamType(Ctx, ArgType);
VarDecl *Arg;
if (DebugFunctionDecl && (CapVar || I->capturesThis())) {
Arg = ParmVarDecl::Create(
Ctx, DebugFunctionDecl,
CapVar ? CapVar->getBeginLoc() : FD->getBeginLoc(),
CapVar ? CapVar->getLocation() : FD->getLocation(), II, ArgType,
/*TInfo=*/nullptr, SC_None, /*DefArg=*/nullptr);
} else {
Arg = ImplicitParamDecl::Create(Ctx, /*DC=*/nullptr, FD->getLocation(),
II, ArgType, ImplicitParamDecl::Other);
}
Args.emplace_back(Arg);
// Do not cast arguments if we emit function with non-original types.
TargetArgs.emplace_back(
FO.UIntPtrCastRequired
? Arg
: CGM.getOpenMPRuntime().translateParameter(FD, Arg));
++I;
}
Args.append(
std::next(CD->param_begin(), CD->getContextParamPosition() + 1),
CD->param_end());
TargetArgs.append(
std::next(CD->param_begin(), CD->getContextParamPosition() + 1),
CD->param_end());
// Create the function declaration.
const CGFunctionInfo &FuncInfo =
CGM.getTypes().arrangeBuiltinFunctionDeclaration(Ctx.VoidTy, TargetArgs);
llvm::FunctionType *FuncLLVMTy = CGM.getTypes().GetFunctionType(FuncInfo);
auto *F =
llvm::Function::Create(FuncLLVMTy, llvm::GlobalValue::InternalLinkage,
FO.FunctionName, &CGM.getModule());
CGM.SetInternalFunctionAttributes(CD, F, FuncInfo);
if (CD->isNothrow())
F->setDoesNotThrow();
F->setDoesNotRecurse();
// Generate the function.
CGF.StartFunction(CD, Ctx.VoidTy, F, FuncInfo, TargetArgs,
FO.S->getBeginLoc(), CD->getBody()->getBeginLoc());
unsigned Cnt = CD->getContextParamPosition();
I = FO.S->captures().begin();
for (const FieldDecl *FD : RD->fields()) {
// Do not map arguments if we emit function with non-original types.
Address LocalAddr(Address::invalid());
if (!FO.UIntPtrCastRequired && Args[Cnt] != TargetArgs[Cnt]) {
LocalAddr = CGM.getOpenMPRuntime().getParameterAddress(CGF, Args[Cnt],
TargetArgs[Cnt]);
} else {
LocalAddr = CGF.GetAddrOfLocalVar(Args[Cnt]);
}
// If we are capturing a pointer by copy we don't need to do anything, just
// use the value that we get from the arguments.
if (I->capturesVariableByCopy() && FD->getType()->isAnyPointerType()) {
const VarDecl *CurVD = I->getCapturedVar();
if (!FO.RegisterCastedArgsOnly)
LocalAddrs.insert({Args[Cnt], {CurVD, LocalAddr}});
++Cnt;
++I;
continue;
}
LValue ArgLVal = CGF.MakeAddrLValue(LocalAddr, Args[Cnt]->getType(),
AlignmentSource::Decl);
if (FD->hasCapturedVLAType()) {
if (FO.UIntPtrCastRequired) {
ArgLVal = CGF.MakeAddrLValue(
castValueFromUintptr(CGF, I->getLocation(), FD->getType(),
Args[Cnt]->getName(), ArgLVal),
FD->getType(), AlignmentSource::Decl);
}
llvm::Value *ExprArg = CGF.EmitLoadOfScalar(ArgLVal, I->getLocation());
const VariableArrayType *VAT = FD->getCapturedVLAType();
VLASizes.try_emplace(Args[Cnt], VAT->getSizeExpr(), ExprArg);
} else if (I->capturesVariable()) {
const VarDecl *Var = I->getCapturedVar();
QualType VarTy = Var->getType();
Address ArgAddr = ArgLVal.getAddress();
if (ArgLVal.getType()->isLValueReferenceType()) {
ArgAddr = CGF.EmitLoadOfReference(ArgLVal);
} else if (!VarTy->isVariablyModifiedType() || !VarTy->isPointerType()) {
assert(ArgLVal.getType()->isPointerType());
ArgAddr = CGF.EmitLoadOfPointer(
ArgAddr, ArgLVal.getType()->castAs<PointerType>());
}
if (!FO.RegisterCastedArgsOnly) {
LocalAddrs.insert(
{Args[Cnt],
{Var, Address(ArgAddr.getPointer(), Ctx.getDeclAlign(Var))}});
}
} else if (I->capturesVariableByCopy()) {
assert(!FD->getType()->isAnyPointerType() &&
"Not expecting a captured pointer.");
const VarDecl *Var = I->getCapturedVar();
LocalAddrs.insert({Args[Cnt],
{Var, FO.UIntPtrCastRequired
? castValueFromUintptr(
CGF, I->getLocation(), FD->getType(),
Args[Cnt]->getName(), ArgLVal)
: ArgLVal.getAddress()}});
} else {
// If 'this' is captured, load it into CXXThisValue.
assert(I->capturesThis());
CXXThisValue = CGF.EmitLoadOfScalar(ArgLVal, I->getLocation());
LocalAddrs.insert({Args[Cnt], {nullptr, ArgLVal.getAddress()}});
}
++Cnt;
++I;
}
return F;
}
llvm::Function *
CodeGenFunction::GenerateOpenMPCapturedStmtFunction(const CapturedStmt &S) {
assert(
CapturedStmtInfo &&
"CapturedStmtInfo should be set when generating the captured function");
const CapturedDecl *CD = S.getCapturedDecl();
// Build the argument list.
bool NeedWrapperFunction =
getDebugInfo() &&
CGM.getCodeGenOpts().getDebugInfo() >= codegenoptions::LimitedDebugInfo;
FunctionArgList Args;
llvm::MapVector<const Decl *, std::pair<const VarDecl *, Address>> LocalAddrs;
llvm::DenseMap<const Decl *, std::pair<const Expr *, llvm::Value *>> VLASizes;
SmallString<256> Buffer;
llvm::raw_svector_ostream Out(Buffer);
Out << CapturedStmtInfo->getHelperName();
if (NeedWrapperFunction)
Out << "_debug__";
FunctionOptions FO(&S, !NeedWrapperFunction, /*RegisterCastedArgsOnly=*/false,
Out.str());
llvm::Function *F = emitOutlinedFunctionPrologue(*this, Args, LocalAddrs,
VLASizes, CXXThisValue, FO);
CodeGenFunction::OMPPrivateScope LocalScope(*this);
for (const auto &LocalAddrPair : LocalAddrs) {
if (LocalAddrPair.second.first) {
LocalScope.addPrivate(LocalAddrPair.second.first, [&LocalAddrPair]() {
return LocalAddrPair.second.second;
});
}
}
(void)LocalScope.Privatize();
for (const auto &VLASizePair : VLASizes)
VLASizeMap[VLASizePair.second.first] = VLASizePair.second.second;
PGO.assignRegionCounters(GlobalDecl(CD), F);
CapturedStmtInfo->EmitBody(*this, CD->getBody());
(void)LocalScope.ForceCleanup();
FinishFunction(CD->getBodyRBrace());
if (!NeedWrapperFunction)
return F;
FunctionOptions WrapperFO(&S, /*UIntPtrCastRequired=*/true,
/*RegisterCastedArgsOnly=*/true,
CapturedStmtInfo->getHelperName());
CodeGenFunction WrapperCGF(CGM, /*suppressNewContext=*/true);
WrapperCGF.CapturedStmtInfo = CapturedStmtInfo;
Args.clear();
LocalAddrs.clear();
VLASizes.clear();
llvm::Function *WrapperF =
emitOutlinedFunctionPrologue(WrapperCGF, Args, LocalAddrs, VLASizes,
WrapperCGF.CXXThisValue, WrapperFO);
llvm::SmallVector<llvm::Value *, 4> CallArgs;
for (const auto *Arg : Args) {
llvm::Value *CallArg;
auto I = LocalAddrs.find(Arg);
if (I != LocalAddrs.end()) {
LValue LV = WrapperCGF.MakeAddrLValue(
I->second.second,
I->second.first ? I->second.first->getType() : Arg->getType(),
AlignmentSource::Decl);
CallArg = WrapperCGF.EmitLoadOfScalar(LV, S.getBeginLoc());
} else {
auto EI = VLASizes.find(Arg);
if (EI != VLASizes.end()) {
CallArg = EI->second.second;
} else {
LValue LV = WrapperCGF.MakeAddrLValue(WrapperCGF.GetAddrOfLocalVar(Arg),
Arg->getType(),
AlignmentSource::Decl);
CallArg = WrapperCGF.EmitLoadOfScalar(LV, S.getBeginLoc());
}
}
CallArgs.emplace_back(WrapperCGF.EmitFromMemory(CallArg, Arg->getType()));
}
CGM.getOpenMPRuntime().emitOutlinedFunctionCall(WrapperCGF, S.getBeginLoc(),
F, CallArgs);
WrapperCGF.FinishFunction();
return WrapperF;
}
//===----------------------------------------------------------------------===//
// OpenMP Directive Emission
//===----------------------------------------------------------------------===//
void CodeGenFunction::EmitOMPAggregateAssign(
Address DestAddr, Address SrcAddr, QualType OriginalType,
const llvm::function_ref<void(Address, Address)> CopyGen) {
// Perform element-by-element initialization.
QualType ElementTy;
// Drill down to the base element type on both arrays.
const ArrayType *ArrayTy = OriginalType->getAsArrayTypeUnsafe();
llvm::Value *NumElements = emitArrayLength(ArrayTy, ElementTy, DestAddr);
SrcAddr = Builder.CreateElementBitCast(SrcAddr, DestAddr.getElementType());
llvm::Value *SrcBegin = SrcAddr.getPointer();
llvm::Value *DestBegin = DestAddr.getPointer();
// Cast from pointer to array type to pointer to single element.
llvm::Value *DestEnd = Builder.CreateGEP(DestBegin, NumElements);
// The basic structure here is a while-do loop.
llvm::BasicBlock *BodyBB = createBasicBlock("omp.arraycpy.body");
llvm::BasicBlock *DoneBB = createBasicBlock("omp.arraycpy.done");
llvm::Value *IsEmpty =
Builder.CreateICmpEQ(DestBegin, DestEnd, "omp.arraycpy.isempty");
Builder.CreateCondBr(IsEmpty, DoneBB, BodyBB);
// Enter the loop body, making that address the current address.
llvm::BasicBlock *EntryBB = Builder.GetInsertBlock();
EmitBlock(BodyBB);
CharUnits ElementSize = getContext().getTypeSizeInChars(ElementTy);
llvm::PHINode *SrcElementPHI =
Builder.CreatePHI(SrcBegin->getType(), 2, "omp.arraycpy.srcElementPast");
SrcElementPHI->addIncoming(SrcBegin, EntryBB);
Address SrcElementCurrent =
Address(SrcElementPHI,
SrcAddr.getAlignment().alignmentOfArrayElement(ElementSize));
llvm::PHINode *DestElementPHI =
Builder.CreatePHI(DestBegin->getType(), 2, "omp.arraycpy.destElementPast");
DestElementPHI->addIncoming(DestBegin, EntryBB);
Address DestElementCurrent =
Address(DestElementPHI,
DestAddr.getAlignment().alignmentOfArrayElement(ElementSize));
// Emit copy.
CopyGen(DestElementCurrent, SrcElementCurrent);
// Shift the address forward by one element.
llvm::Value *DestElementNext = Builder.CreateConstGEP1_32(
DestElementPHI, /*Idx0=*/1, "omp.arraycpy.dest.element");
llvm::Value *SrcElementNext = Builder.CreateConstGEP1_32(
SrcElementPHI, /*Idx0=*/1, "omp.arraycpy.src.element");
// Check whether we've reached the end.
llvm::Value *Done =
Builder.CreateICmpEQ(DestElementNext, DestEnd, "omp.arraycpy.done");
Builder.CreateCondBr(Done, DoneBB, BodyBB);
DestElementPHI->addIncoming(DestElementNext, Builder.GetInsertBlock());
SrcElementPHI->addIncoming(SrcElementNext, Builder.GetInsertBlock());
// Done.
EmitBlock(DoneBB, /*IsFinished=*/true);
}
void CodeGenFunction::EmitOMPCopy(QualType OriginalType, Address DestAddr,
Address SrcAddr, const VarDecl *DestVD,
const VarDecl *SrcVD, const Expr *Copy) {
if (OriginalType->isArrayType()) {
const auto *BO = dyn_cast<BinaryOperator>(Copy);
if (BO && BO->getOpcode() == BO_Assign) {
// Perform simple memcpy for simple copying.
LValue Dest = MakeAddrLValue(DestAddr, OriginalType);
LValue Src = MakeAddrLValue(SrcAddr, OriginalType);
EmitAggregateAssign(Dest, Src, OriginalType);
} else {
// For arrays with complex element types perform element by element
// copying.
EmitOMPAggregateAssign(
DestAddr, SrcAddr, OriginalType,
[this, Copy, SrcVD, DestVD](Address DestElement, Address SrcElement) {
// Working with the single array element, so have to remap
// destination and source variables to corresponding array
// elements.
CodeGenFunction::OMPPrivateScope Remap(*this);
Remap.addPrivate(DestVD, [DestElement]() { return DestElement; });
Remap.addPrivate(SrcVD, [SrcElement]() { return SrcElement; });
(void)Remap.Privatize();
EmitIgnoredExpr(Copy);
});
}
} else {
// Remap pseudo source variable to private copy.
CodeGenFunction::OMPPrivateScope Remap(*this);
Remap.addPrivate(SrcVD, [SrcAddr]() { return SrcAddr; });
Remap.addPrivate(DestVD, [DestAddr]() { return DestAddr; });
(void)Remap.Privatize();
// Emit copying of the whole variable.
EmitIgnoredExpr(Copy);
}
}
bool CodeGenFunction::EmitOMPFirstprivateClause(const OMPExecutableDirective &D,
OMPPrivateScope &PrivateScope) {
if (!HaveInsertPoint())
return false;
bool DeviceConstTarget =
getLangOpts().OpenMPIsDevice &&
isOpenMPTargetExecutionDirective(D.getDirectiveKind());
bool FirstprivateIsLastprivate = false;
llvm::DenseSet<const VarDecl *> Lastprivates;
for (const auto *C : D.getClausesOfKind<OMPLastprivateClause>()) {
for (const auto *D : C->varlists())
Lastprivates.insert(
cast<VarDecl>(cast<DeclRefExpr>(D)->getDecl())->getCanonicalDecl());
}
llvm::DenseSet<const VarDecl *> EmittedAsFirstprivate;
llvm::SmallVector<OpenMPDirectiveKind, 4> CaptureRegions;
getOpenMPCaptureRegions(CaptureRegions, D.getDirectiveKind());
// Force emission of the firstprivate copy if the directive does not emit
// outlined function, like omp for, omp simd, omp distribute etc.
bool MustEmitFirstprivateCopy =
CaptureRegions.size() == 1 && CaptureRegions.back() == OMPD_unknown;
for (const auto *C : D.getClausesOfKind<OMPFirstprivateClause>()) {
auto IRef = C->varlist_begin();
auto InitsRef = C->inits().begin();
for (const Expr *IInit : C->private_copies()) {
const auto *OrigVD = cast<VarDecl>(cast<DeclRefExpr>(*IRef)->getDecl());
bool ThisFirstprivateIsLastprivate =
Lastprivates.count(OrigVD->getCanonicalDecl()) > 0;
const FieldDecl *FD = CapturedStmtInfo->lookup(OrigVD);
const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(IInit)->getDecl());
if (!MustEmitFirstprivateCopy && !ThisFirstprivateIsLastprivate && FD &&
!FD->getType()->isReferenceType() &&
(!VD || !VD->hasAttr<OMPAllocateDeclAttr>())) {
EmittedAsFirstprivate.insert(OrigVD->getCanonicalDecl());
++IRef;
++InitsRef;
continue;
}
// Do not emit copy for firstprivate constant variables in target regions,
// captured by reference.
if (DeviceConstTarget && OrigVD->getType().isConstant(getContext()) &&
FD && FD->getType()->isReferenceType() &&
(!VD || !VD->hasAttr<OMPAllocateDeclAttr>())) {
(void)CGM.getOpenMPRuntime().registerTargetFirstprivateCopy(*this,
OrigVD);
++IRef;
++InitsRef;
continue;
}
FirstprivateIsLastprivate =
FirstprivateIsLastprivate || ThisFirstprivateIsLastprivate;
if (EmittedAsFirstprivate.insert(OrigVD->getCanonicalDecl()).second) {
const auto *VDInit =
cast<VarDecl>(cast<DeclRefExpr>(*InitsRef)->getDecl());
bool IsRegistered;
DeclRefExpr DRE(getContext(), const_cast<VarDecl *>(OrigVD),
/*RefersToEnclosingVariableOrCapture=*/FD != nullptr,
(*IRef)->getType(), VK_LValue, (*IRef)->getExprLoc());
LValue OriginalLVal;
if (!FD) {
// Check if the firstprivate variable is just a constant value.
ConstantEmission CE = tryEmitAsConstant(&DRE);
if (CE && !CE.isReference()) {
// Constant value, no need to create a copy.
++IRef;
++InitsRef;
continue;
}
if (CE && CE.isReference()) {
OriginalLVal = CE.getReferenceLValue(*this, &DRE);
} else {
assert(!CE && "Expected non-constant firstprivate.");
OriginalLVal = EmitLValue(&DRE);
}
} else {
OriginalLVal = EmitLValue(&DRE);
}
QualType Type = VD->getType();
if (Type->isArrayType()) {
// Emit VarDecl with copy init for arrays.
// Get the address of the original variable captured in current
// captured region.
IsRegistered = PrivateScope.addPrivate(
OrigVD, [this, VD, Type, OriginalLVal, VDInit]() {
AutoVarEmission Emission = EmitAutoVarAlloca(*VD);
const Expr *Init = VD->getInit();
if (!isa<CXXConstructExpr>(Init) ||
isTrivialInitializer(Init)) {
// Perform simple memcpy.
LValue Dest =
MakeAddrLValue(Emission.getAllocatedAddress(), Type);
EmitAggregateAssign(Dest, OriginalLVal, Type);
} else {
EmitOMPAggregateAssign(
Emission.getAllocatedAddress(), OriginalLVal.getAddress(),
Type,
[this, VDInit, Init](Address DestElement,
Address SrcElement) {
// Clean up any temporaries needed by the
// initialization.
RunCleanupsScope InitScope(*this);
// Emit initialization for single element.
setAddrOfLocalVar(VDInit, SrcElement);
EmitAnyExprToMem(Init, DestElement,
Init->getType().getQualifiers(),
/*IsInitializer*/ false);
LocalDeclMap.erase(VDInit);
});
}
EmitAutoVarCleanups(Emission);
return Emission.getAllocatedAddress();
});
} else {
Address OriginalAddr = OriginalLVal.getAddress();
IsRegistered = PrivateScope.addPrivate(
OrigVD, [this, VDInit, OriginalAddr, VD]() {
// Emit private VarDecl with copy init.
// Remap temp VDInit variable to the address of the original
// variable (for proper handling of captured global variables).
setAddrOfLocalVar(VDInit, OriginalAddr);
EmitDecl(*VD);
LocalDeclMap.erase(VDInit);
return GetAddrOfLocalVar(VD);
});
}
assert(IsRegistered &&
"firstprivate var already registered as private");
// Silence the warning about unused variable.
(void)IsRegistered;
}
++IRef;
++InitsRef;
}
}
return FirstprivateIsLastprivate && !EmittedAsFirstprivate.empty();
}
void CodeGenFunction::EmitOMPPrivateClause(
const OMPExecutableDirective &D,
CodeGenFunction::OMPPrivateScope &PrivateScope) {
if (!HaveInsertPoint())
return;
llvm::DenseSet<const VarDecl *> EmittedAsPrivate;
for (const auto *C : D.getClausesOfKind<OMPPrivateClause>()) {
auto IRef = C->varlist_begin();
for (const Expr *IInit : C->private_copies()) {
const auto *OrigVD = cast<VarDecl>(cast<DeclRefExpr>(*IRef)->getDecl());
if (EmittedAsPrivate.insert(OrigVD->getCanonicalDecl()).second) {
const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(IInit)->getDecl());
bool IsRegistered = PrivateScope.addPrivate(OrigVD, [this, VD]() {
// Emit private VarDecl with copy init.
EmitDecl(*VD);
return GetAddrOfLocalVar(VD);
});
assert(IsRegistered && "private var already registered as private");
// Silence the warning about unused variable.
(void)IsRegistered;
}
++IRef;
}
}
}
bool CodeGenFunction::EmitOMPCopyinClause(const OMPExecutableDirective &D) {
if (!HaveInsertPoint())
return false;
// threadprivate_var1 = master_threadprivate_var1;
// operator=(threadprivate_var2, master_threadprivate_var2);
// ...
// __kmpc_barrier(&loc, global_tid);
llvm::DenseSet<const VarDecl *> CopiedVars;
llvm::BasicBlock *CopyBegin = nullptr, *CopyEnd = nullptr;
for (const auto *C : D.getClausesOfKind<OMPCopyinClause>()) {
auto IRef = C->varlist_begin();
auto ISrcRef = C->source_exprs().begin();
auto IDestRef = C->destination_exprs().begin();
for (const Expr *AssignOp : C->assignment_ops()) {
const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(*IRef)->getDecl());
QualType Type = VD->getType();
if (CopiedVars.insert(VD->getCanonicalDecl()).second) {
// Get the address of the master variable. If we are emitting code with
// TLS support, the address is passed from the master as field in the
// captured declaration.
Address MasterAddr = Address::invalid();
if (getLangOpts().OpenMPUseTLS &&
getContext().getTargetInfo().isTLSSupported()) {
assert(CapturedStmtInfo->lookup(VD) &&
"Copyin threadprivates should have been captured!");
DeclRefExpr DRE(getContext(), const_cast<VarDecl *>(VD), true,
(*IRef)->getType(), VK_LValue, (*IRef)->getExprLoc());
MasterAddr = EmitLValue(&DRE).getAddress();
LocalDeclMap.erase(VD);
} else {
MasterAddr =
Address(VD->isStaticLocal() ? CGM.getStaticLocalDeclAddress(VD)
: CGM.GetAddrOfGlobal(VD),
getContext().getDeclAlign(VD));
}
// Get the address of the threadprivate variable.
Address PrivateAddr = EmitLValue(*IRef).getAddress();
if (CopiedVars.size() == 1) {
// At first check if current thread is a master thread. If it is, no
// need to copy data.
CopyBegin = createBasicBlock("copyin.not.master");
CopyEnd = createBasicBlock("copyin.not.master.end");
Builder.CreateCondBr(
Builder.CreateICmpNE(
Builder.CreatePtrToInt(MasterAddr.getPointer(), CGM.IntPtrTy),
Builder.CreatePtrToInt(PrivateAddr.getPointer(),
CGM.IntPtrTy)),
CopyBegin, CopyEnd);
EmitBlock(CopyBegin);
}
const auto *SrcVD =
cast<VarDecl>(cast<DeclRefExpr>(*ISrcRef)->getDecl());
const auto *DestVD =
cast<VarDecl>(cast<DeclRefExpr>(*IDestRef)->getDecl());
EmitOMPCopy(Type, PrivateAddr, MasterAddr, DestVD, SrcVD, AssignOp);
}
++IRef;
++ISrcRef;
++IDestRef;
}
}
if (CopyEnd) {
// Exit out of copying procedure for non-master thread.
EmitBlock(CopyEnd, /*IsFinished=*/true);
return true;
}
return false;
}
bool CodeGenFunction::EmitOMPLastprivateClauseInit(
const OMPExecutableDirective &D, OMPPrivateScope &PrivateScope) {
if (!HaveInsertPoint())
return false;
bool HasAtLeastOneLastprivate = false;
llvm::DenseSet<const VarDecl *> SIMDLCVs;
if (isOpenMPSimdDirective(D.getDirectiveKind())) {
const auto *LoopDirective = cast<OMPLoopDirective>(&D);
for (const Expr *C : LoopDirective->counters()) {
SIMDLCVs.insert(
cast<VarDecl>(cast<DeclRefExpr>(C)->getDecl())->getCanonicalDecl());
}
}
llvm::DenseSet<const VarDecl *> AlreadyEmittedVars;
for (const auto *C : D.getClausesOfKind<OMPLastprivateClause>()) {
HasAtLeastOneLastprivate = true;
if (isOpenMPTaskLoopDirective(D.getDirectiveKind()) &&
!getLangOpts().OpenMPSimd)
break;
auto IRef = C->varlist_begin();
auto IDestRef = C->destination_exprs().begin();
for (const Expr *IInit : C->private_copies()) {
// Keep the address of the original variable for future update at the end
// of the loop.
const auto *OrigVD = cast<VarDecl>(cast<DeclRefExpr>(*IRef)->getDecl());
// Taskloops do not require additional initialization, it is done in
// runtime support library.
if (AlreadyEmittedVars.insert(OrigVD->getCanonicalDecl()).second) {
const auto *DestVD =
cast<VarDecl>(cast<DeclRefExpr>(*IDestRef)->getDecl());
PrivateScope.addPrivate(DestVD, [this, OrigVD, IRef]() {
DeclRefExpr DRE(getContext(), const_cast<VarDecl *>(OrigVD),
/*RefersToEnclosingVariableOrCapture=*/
CapturedStmtInfo->lookup(OrigVD) != nullptr,
(*IRef)->getType(), VK_LValue, (*IRef)->getExprLoc());
return EmitLValue(&DRE).getAddress();
});
// Check if the variable is also a firstprivate: in this case IInit is
// not generated. Initialization of this variable will happen in codegen
// for 'firstprivate' clause.
if (IInit && !SIMDLCVs.count(OrigVD->getCanonicalDecl())) {
const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(IInit)->getDecl());
bool IsRegistered = PrivateScope.addPrivate(OrigVD, [this, VD]() {
// Emit private VarDecl with copy init.
EmitDecl(*VD);
return GetAddrOfLocalVar(VD);
});
assert(IsRegistered &&
"lastprivate var already registered as private");
(void)IsRegistered;
}
}
++IRef;
++IDestRef;
}
}
return HasAtLeastOneLastprivate;
}
void CodeGenFunction::EmitOMPLastprivateClauseFinal(
const OMPExecutableDirective &D, bool NoFinals,
llvm::Value *IsLastIterCond) {
if (!HaveInsertPoint())
return;
// Emit following code:
// if (<IsLastIterCond>) {
// orig_var1 = private_orig_var1;
// ...
// orig_varn = private_orig_varn;
// }
llvm::BasicBlock *ThenBB = nullptr;
llvm::BasicBlock *DoneBB = nullptr;
if (IsLastIterCond) {
ThenBB = createBasicBlock(".omp.lastprivate.then");
DoneBB = createBasicBlock(".omp.lastprivate.done");
Builder.CreateCondBr(IsLastIterCond, ThenBB, DoneBB);
EmitBlock(ThenBB);
}
llvm::DenseSet<const VarDecl *> AlreadyEmittedVars;
llvm::DenseMap<const VarDecl *, const Expr *> LoopCountersAndUpdates;
if (const auto *LoopDirective = dyn_cast<OMPLoopDirective>(&D)) {
auto IC = LoopDirective->counters().begin();
for (const Expr *F : LoopDirective->finals()) {
const auto *D =
cast<VarDecl>(cast<DeclRefExpr>(*IC)->getDecl())->getCanonicalDecl();
if (NoFinals)
AlreadyEmittedVars.insert(D);
else
LoopCountersAndUpdates[D] = F;
++IC;
}
}
for (const auto *C : D.getClausesOfKind<OMPLastprivateClause>()) {
auto IRef = C->varlist_begin();
auto ISrcRef = C->source_exprs().begin();
auto IDestRef = C->destination_exprs().begin();
for (const Expr *AssignOp : C->assignment_ops()) {
const auto *PrivateVD =
cast<VarDecl>(cast<DeclRefExpr>(*IRef)->getDecl());
QualType Type = PrivateVD->getType();
const auto *CanonicalVD = PrivateVD->getCanonicalDecl();
if (AlreadyEmittedVars.insert(CanonicalVD).second) {
// If lastprivate variable is a loop control variable for loop-based
// directive, update its value before copyin back to original
// variable.
if (const Expr *FinalExpr = LoopCountersAndUpdates.lookup(CanonicalVD))
EmitIgnoredExpr(FinalExpr);
const auto *SrcVD =
cast<VarDecl>(cast<DeclRefExpr>(*ISrcRef)->getDecl());
const auto *DestVD =
cast<VarDecl>(cast<DeclRefExpr>(*IDestRef)->getDecl());
// Get the address of the original variable.
Address OriginalAddr = GetAddrOfLocalVar(DestVD);
// Get the address of the private variable.
Address PrivateAddr = GetAddrOfLocalVar(PrivateVD);
if (const auto *RefTy = PrivateVD->getType()->getAs<ReferenceType>())
PrivateAddr =
Address(Builder.CreateLoad(PrivateAddr),
getNaturalTypeAlignment(RefTy->getPointeeType()));
EmitOMPCopy(Type, OriginalAddr, PrivateAddr, DestVD, SrcVD, AssignOp);
}
++IRef;
++ISrcRef;
++IDestRef;
}
if (const Expr *PostUpdate = C->getPostUpdateExpr())
EmitIgnoredExpr(PostUpdate);
}
if (IsLastIterCond)
EmitBlock(DoneBB, /*IsFinished=*/true);
}
void CodeGenFunction::EmitOMPReductionClauseInit(
const OMPExecutableDirective &D,
CodeGenFunction::OMPPrivateScope &PrivateScope) {
if (!HaveInsertPoint())
return;
SmallVector<const Expr *, 4> Shareds;
SmallVector<const Expr *, 4> Privates;
SmallVector<const Expr *, 4> ReductionOps;
SmallVector<const Expr *, 4> LHSs;
SmallVector<const Expr *, 4> RHSs;
for (const auto *C : D.getClausesOfKind<OMPReductionClause>()) {
auto IPriv = C->privates().begin();
auto IRed = C->reduction_ops().begin();
auto ILHS = C->lhs_exprs().begin();
auto IRHS = C->rhs_exprs().begin();
for (const Expr *Ref : C->varlists()) {
Shareds.emplace_back(Ref);
Privates.emplace_back(*IPriv);
ReductionOps.emplace_back(*IRed);
LHSs.emplace_back(*ILHS);
RHSs.emplace_back(*IRHS);
std::advance(IPriv, 1);
std::advance(IRed, 1);
std::advance(ILHS, 1);
std::advance(IRHS, 1);
}
}
ReductionCodeGen RedCG(Shareds, Privates, ReductionOps);
unsigned Count = 0;
auto ILHS = LHSs.begin();
auto IRHS = RHSs.begin();
auto IPriv = Privates.begin();
for (const Expr *IRef : Shareds) {
const auto *PrivateVD = cast<VarDecl>(cast<DeclRefExpr>(*IPriv)->getDecl());
// Emit private VarDecl with reduction init.
RedCG.emitSharedLValue(*this, Count);
RedCG.emitAggregateType(*this, Count);
AutoVarEmission Emission = EmitAutoVarAlloca(*PrivateVD);
RedCG.emitInitialization(*this, Count, Emission.getAllocatedAddress(),
RedCG.getSharedLValue(Count),
[&Emission](CodeGenFunction &CGF) {
CGF.EmitAutoVarInit(Emission);
return true;
});
EmitAutoVarCleanups(Emission);
Address BaseAddr = RedCG.adjustPrivateAddress(
*this, Count, Emission.getAllocatedAddress());
bool IsRegistered = PrivateScope.addPrivate(
RedCG.getBaseDecl(Count), [BaseAddr]() { return BaseAddr; });
assert(IsRegistered && "private var already registered as private");
// Silence the warning about unused variable.
(void)IsRegistered;
const auto *LHSVD = cast<VarDecl>(cast<DeclRefExpr>(*ILHS)->getDecl());
const auto *RHSVD = cast<VarDecl>(cast<DeclRefExpr>(*IRHS)->getDecl());
QualType Type = PrivateVD->getType();
bool isaOMPArraySectionExpr = isa<OMPArraySectionExpr>(IRef);
if (isaOMPArraySectionExpr && Type->isVariablyModifiedType()) {
// Store the address of the original variable associated with the LHS
// implicit variable.
PrivateScope.addPrivate(LHSVD, [&RedCG, Count]() {
return RedCG.getSharedLValue(Count).getAddress();
});
PrivateScope.addPrivate(
RHSVD, [this, PrivateVD]() { return GetAddrOfLocalVar(PrivateVD); });
} else if ((isaOMPArraySectionExpr && Type->isScalarType()) ||
isa<ArraySubscriptExpr>(IRef)) {
// Store the address of the original variable associated with the LHS
// implicit variable.
PrivateScope.addPrivate(LHSVD, [&RedCG, Count]() {
return RedCG.getSharedLValue(Count).getAddress();
});
PrivateScope.addPrivate(RHSVD, [this, PrivateVD, RHSVD]() {
return Builder.CreateElementBitCast(GetAddrOfLocalVar(PrivateVD),
ConvertTypeForMem(RHSVD->getType()),
"rhs.begin");
});
} else {
QualType Type = PrivateVD->getType();
bool IsArray = getContext().getAsArrayType(Type) != nullptr;
Address OriginalAddr = RedCG.getSharedLValue(Count).getAddress();
// Store the address of the original variable associated with the LHS
// implicit variable.
if (IsArray) {
OriginalAddr = Builder.CreateElementBitCast(
OriginalAddr, ConvertTypeForMem(LHSVD->getType()), "lhs.begin");
}
PrivateScope.addPrivate(LHSVD, [OriginalAddr]() { return OriginalAddr; });
PrivateScope.addPrivate(
RHSVD, [this, PrivateVD, RHSVD, IsArray]() {
return IsArray
? Builder.CreateElementBitCast(
GetAddrOfLocalVar(PrivateVD),
ConvertTypeForMem(RHSVD->getType()), "rhs.begin")
: GetAddrOfLocalVar(PrivateVD);
});
}
++ILHS;
++IRHS;
++IPriv;
++Count;
}
}
void CodeGenFunction::EmitOMPReductionClauseFinal(
const OMPExecutableDirective &D, const OpenMPDirectiveKind ReductionKind) {
if (!HaveInsertPoint())
return;
llvm::SmallVector<const Expr *, 8> Privates;
llvm::SmallVector<const Expr *, 8> LHSExprs;
llvm::SmallVector<const Expr *, 8> RHSExprs;
llvm::SmallVector<const Expr *, 8> ReductionOps;
bool HasAtLeastOneReduction = false;
for (const auto *C : D.getClausesOfKind<OMPReductionClause>()) {
HasAtLeastOneReduction = true;
Privates.append(C->privates().begin(), C->privates().end());
LHSExprs.append(C->lhs_exprs().begin(), C->lhs_exprs().end());
RHSExprs.append(C->rhs_exprs().begin(), C->rhs_exprs().end());
ReductionOps.append(C->reduction_ops().begin(), C->reduction_ops().end());
}
if (HasAtLeastOneReduction) {
bool WithNowait = D.getSingleClause<OMPNowaitClause>() ||
isOpenMPParallelDirective(D.getDirectiveKind()) ||
ReductionKind == OMPD_simd;
bool SimpleReduction = ReductionKind == OMPD_simd;
// Emit nowait reduction if nowait clause is present or directive is a
// parallel directive (it always has implicit barrier).
CGM.getOpenMPRuntime().emitReduction(
*this, D.getEndLoc(), Privates, LHSExprs, RHSExprs, ReductionOps,
{WithNowait, SimpleReduction, ReductionKind});
}
}
static void emitPostUpdateForReductionClause(
CodeGenFunction &CGF, const OMPExecutableDirective &D,
const llvm::function_ref<llvm::Value *(CodeGenFunction &)> CondGen) {
if (!CGF.HaveInsertPoint())
return;
llvm::BasicBlock *DoneBB = nullptr;
for (const auto *C : D.getClausesOfKind<OMPReductionClause>()) {
if (const Expr *PostUpdate = C->getPostUpdateExpr()) {
if (!DoneBB) {
if (llvm::Value *Cond = CondGen(CGF)) {
// If the first post-update expression is found, emit conditional
// block if it was requested.
llvm::BasicBlock *ThenBB = CGF.createBasicBlock(".omp.reduction.pu");
DoneBB = CGF.createBasicBlock(".omp.reduction.pu.done");
CGF.Builder.CreateCondBr(Cond, ThenBB, DoneBB);
CGF.EmitBlock(ThenBB);
}
}
CGF.EmitIgnoredExpr(PostUpdate);
}
}
if (DoneBB)
CGF.EmitBlock(DoneBB, /*IsFinished=*/true);
}
namespace {
/// Codegen lambda for appending distribute lower and upper bounds to outlined
/// parallel function. This is necessary for combined constructs such as
/// 'distribute parallel for'
typedef llvm::function_ref<void(CodeGenFunction &,
const OMPExecutableDirective &,
llvm::SmallVectorImpl<llvm::Value *> &)>
CodeGenBoundParametersTy;
} // anonymous namespace
static void emitCommonOMPParallelDirective(
CodeGenFunction &CGF, const OMPExecutableDirective &S,
OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen,
const CodeGenBoundParametersTy &CodeGenBoundParameters) {
const CapturedStmt *CS = S.getCapturedStmt(OMPD_parallel);
llvm::Function *OutlinedFn =
CGF.CGM.getOpenMPRuntime().emitParallelOutlinedFunction(
S, *CS->getCapturedDecl()->param_begin(), InnermostKind, CodeGen);
if (const auto *NumThreadsClause = S.getSingleClause<OMPNumThreadsClause>()) {
CodeGenFunction::RunCleanupsScope NumThreadsScope(CGF);
llvm::Value *NumThreads =
CGF.EmitScalarExpr(NumThreadsClause->getNumThreads(),
/*IgnoreResultAssign=*/true);
CGF.CGM.getOpenMPRuntime().emitNumThreadsClause(
CGF, NumThreads, NumThreadsClause->getBeginLoc());
}
if (const auto *ProcBindClause = S.getSingleClause<OMPProcBindClause>()) {
CodeGenFunction::RunCleanupsScope ProcBindScope(CGF);
CGF.CGM.getOpenMPRuntime().emitProcBindClause(
CGF, ProcBindClause->getProcBindKind(), ProcBindClause->getBeginLoc());
}
const Expr *IfCond = nullptr;
for (const auto *C : S.getClausesOfKind<OMPIfClause>()) {
if (C->getNameModifier() == OMPD_unknown ||
C->getNameModifier() == OMPD_parallel) {
IfCond = C->getCondition();
break;
}
}
OMPParallelScope Scope(CGF, S);
llvm::SmallVector<llvm::Value *, 16> CapturedVars;
// Combining 'distribute' with 'for' requires sharing each 'distribute' chunk
// lower and upper bounds with the pragma 'for' chunking mechanism.
// The following lambda takes care of appending the lower and upper bound
// parameters when necessary
CodeGenBoundParameters(CGF, S, CapturedVars);
CGF.GenerateOpenMPCapturedVars(*CS, CapturedVars);
CGF.CGM.getOpenMPRuntime().emitParallelCall(CGF, S.getBeginLoc(), OutlinedFn,
CapturedVars, IfCond);
}
static void emitEmptyBoundParameters(CodeGenFunction &,
const OMPExecutableDirective &,
llvm::SmallVectorImpl<llvm::Value *> &) {}
void CodeGenFunction::EmitOMPParallelDirective(const OMPParallelDirective &S) {
// Emit parallel region as a standalone region.
auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
Action.Enter(CGF);
OMPPrivateScope PrivateScope(CGF);
bool Copyins = CGF.EmitOMPCopyinClause(S);
(void)CGF.EmitOMPFirstprivateClause(S, PrivateScope);
if (Copyins) {
// Emit implicit barrier to synchronize threads and avoid data races on
// propagation master's thread values of threadprivate variables to local
// instances of that variables of all other implicit threads.
CGF.CGM.getOpenMPRuntime().emitBarrierCall(
CGF, S.getBeginLoc(), OMPD_unknown, /*EmitChecks=*/false,
/*ForceSimpleCall=*/true);
}
CGF.EmitOMPPrivateClause(S, PrivateScope);
CGF.EmitOMPReductionClauseInit(S, PrivateScope);
(void)PrivateScope.Privatize();
CGF.EmitStmt(S.getCapturedStmt(OMPD_parallel)->getCapturedStmt());
CGF.EmitOMPReductionClauseFinal(S, /*ReductionKind=*/OMPD_parallel);
};
emitCommonOMPParallelDirective(*this, S, OMPD_parallel, CodeGen,
emitEmptyBoundParameters);
emitPostUpdateForReductionClause(*this, S,
[](CodeGenFunction &) { return nullptr; });
}
void CodeGenFunction::EmitOMPLoopBody(const OMPLoopDirective &D,
JumpDest LoopExit) {
RunCleanupsScope BodyScope(*this);
// Update counters values on current iteration.
for (const Expr *UE : D.updates())
EmitIgnoredExpr(UE);
// Update the linear variables.
// In distribute directives only loop counters may be marked as linear, no
// need to generate the code for them.
if (!isOpenMPDistributeDirective(D.getDirectiveKind())) {
for (const auto *C : D.getClausesOfKind<OMPLinearClause>()) {
for (const Expr *UE : C->updates())
EmitIgnoredExpr(UE);
}
}
// On a continue in the body, jump to the end.
JumpDest Continue = getJumpDestInCurrentScope("omp.body.continue");
BreakContinueStack.push_back(BreakContinue(LoopExit, Continue));
for (const Expr *E : D.finals_conditions()) {
if (!E)
continue;
// Check that loop counter in non-rectangular nest fits into the iteration
// space.
llvm::BasicBlock *NextBB = createBasicBlock("omp.body.next");
EmitBranchOnBoolExpr(E, NextBB, Continue.getBlock(),
getProfileCount(D.getBody()));
EmitBlock(NextBB);
}
// Emit loop variables for C++ range loops.
const Stmt *Body =
D.getInnermostCapturedStmt()->getCapturedStmt()->IgnoreContainers();
for (unsigned Cnt = 0; Cnt < D.getCollapsedNumber(); ++Cnt) {
Body = Body->IgnoreContainers();
if (auto *For = dyn_cast<ForStmt>(Body)) {
Body = For->getBody();
} else {
assert(isa<CXXForRangeStmt>(Body) &&
"Expected canonical for loop or range-based for loop.");
auto *CXXFor = cast<CXXForRangeStmt>(Body);
EmitStmt(CXXFor->getLoopVarStmt());
Body = CXXFor->getBody();
}
}
// Emit loop body.
EmitStmt(D.getBody());
// The end (updates/cleanups).
EmitBlock(Continue.getBlock());
BreakContinueStack.pop_back();
}
void CodeGenFunction::EmitOMPInnerLoop(
const Stmt &S, bool RequiresCleanup, const Expr *LoopCond,
const Expr *IncExpr,
const llvm::function_ref<void(CodeGenFunction &)> BodyGen,
const llvm::function_ref<void(CodeGenFunction &)> PostIncGen) {
auto LoopExit = getJumpDestInCurrentScope("omp.inner.for.end");
// Start the loop with a block that tests the condition.
auto CondBlock = createBasicBlock("omp.inner.for.cond");
EmitBlock(CondBlock);
const SourceRange R = S.getSourceRange();
LoopStack.push(CondBlock, SourceLocToDebugLoc(R.getBegin()),
SourceLocToDebugLoc(R.getEnd()));
// If there are any cleanups between here and the loop-exit scope,
// create a block to stage a loop exit along.
llvm::BasicBlock *ExitBlock = LoopExit.getBlock();
if (RequiresCleanup)
ExitBlock = createBasicBlock("omp.inner.for.cond.cleanup");
llvm::BasicBlock *LoopBody = createBasicBlock("omp.inner.for.body");
// Emit condition.
EmitBranchOnBoolExpr(LoopCond, LoopBody, ExitBlock, getProfileCount(&S));
if (ExitBlock != LoopExit.getBlock()) {
EmitBlock(ExitBlock);
EmitBranchThroughCleanup(LoopExit);
}
EmitBlock(LoopBody);
incrementProfileCounter(&S);
// Create a block for the increment.
JumpDest Continue = getJumpDestInCurrentScope("omp.inner.for.inc");
BreakContinueStack.push_back(BreakContinue(LoopExit, Continue));
BodyGen(*this);
// Emit "IV = IV + 1" and a back-edge to the condition block.
EmitBlock(Continue.getBlock());
EmitIgnoredExpr(IncExpr);
PostIncGen(*this);
BreakContinueStack.pop_back();
EmitBranch(CondBlock);
LoopStack.pop();
// Emit the fall-through block.
EmitBlock(LoopExit.getBlock());
}
bool CodeGenFunction::EmitOMPLinearClauseInit(const OMPLoopDirective &D) {
if (!HaveInsertPoint())
return false;
// Emit inits for the linear variables.
bool HasLinears = false;
for (const auto *C : D.getClausesOfKind<OMPLinearClause>()) {
for (const Expr *Init : C->inits()) {
HasLinears = true;
const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(Init)->getDecl());
if (const auto *Ref =
dyn_cast<DeclRefExpr>(VD->getInit()->IgnoreImpCasts())) {
AutoVarEmission Emission = EmitAutoVarAlloca(*VD);
const auto *OrigVD = cast<VarDecl>(Ref->getDecl());
DeclRefExpr DRE(getContext(), const_cast<VarDecl *>(OrigVD),
CapturedStmtInfo->lookup(OrigVD) != nullptr,
VD->getInit()->getType(), VK_LValue,
VD->getInit()->getExprLoc());
EmitExprAsInit(&DRE, VD, MakeAddrLValue(Emission.getAllocatedAddress(),
VD->getType()),
/*capturedByInit=*/false);
EmitAutoVarCleanups(Emission);
} else {
EmitVarDecl(*VD);
}
}
// Emit the linear steps for the linear clauses.
// If a step is not constant, it is pre-calculated before the loop.
if (const auto *CS = cast_or_null<BinaryOperator>(C->getCalcStep()))
if (const auto *SaveRef = cast<DeclRefExpr>(CS->getLHS())) {
EmitVarDecl(*cast<VarDecl>(SaveRef->getDecl()));
// Emit calculation of the linear step.
EmitIgnoredExpr(CS);
}
}
return HasLinears;
}
void CodeGenFunction::EmitOMPLinearClauseFinal(
const OMPLoopDirective &D,
const llvm::function_ref<llvm::Value *(CodeGenFunction &)> CondGen) {
if (!HaveInsertPoint())
return;
llvm::BasicBlock *DoneBB = nullptr;
// Emit the final values of the linear variables.
for (const auto *C : D.getClausesOfKind<OMPLinearClause>()) {
auto IC = C->varlist_begin();
for (const Expr *F : C->finals()) {
if (!DoneBB) {
if (llvm::Value *Cond = CondGen(*this)) {
// If the first post-update expression is found, emit conditional
// block if it was requested.
llvm::BasicBlock *ThenBB = createBasicBlock(".omp.linear.pu");
DoneBB = createBasicBlock(".omp.linear.pu.done");
Builder.CreateCondBr(Cond, ThenBB, DoneBB);
EmitBlock(ThenBB);
}
}
const auto *OrigVD = cast<VarDecl>(cast<DeclRefExpr>(*IC)->getDecl());
DeclRefExpr DRE(getContext(), const_cast<VarDecl *>(OrigVD),
CapturedStmtInfo->lookup(OrigVD) != nullptr,
(*IC)->getType(), VK_LValue, (*IC)->getExprLoc());
Address OrigAddr = EmitLValue(&DRE).getAddress();
CodeGenFunction::OMPPrivateScope VarScope(*this);
VarScope.addPrivate(OrigVD, [OrigAddr]() { return OrigAddr; });
(void)VarScope.Privatize();
EmitIgnoredExpr(F);
++IC;
}
if (const Expr *PostUpdate = C->getPostUpdateExpr())
EmitIgnoredExpr(PostUpdate);
}
if (DoneBB)
EmitBlock(DoneBB, /*IsFinished=*/true);
}
static void emitAlignedClause(CodeGenFunction &CGF,
const OMPExecutableDirective &D) {
if (!CGF.HaveInsertPoint())
return;
for (const auto *Clause : D.getClausesOfKind<OMPAlignedClause>()) {
llvm::APInt ClauseAlignment(64, 0);
if (const Expr *AlignmentExpr = Clause->getAlignment()) {
auto *AlignmentCI =
cast<llvm::ConstantInt>(CGF.EmitScalarExpr(AlignmentExpr));
ClauseAlignment = AlignmentCI->getValue();
}
for (const Expr *E : Clause->varlists()) {
llvm::APInt Alignment(ClauseAlignment);
if (Alignment == 0) {
// OpenMP [2.8.1, Description]
// If no optional parameter is specified, implementation-defined default
// alignments for SIMD instructions on the target platforms are assumed.
Alignment =
CGF.getContext()
.toCharUnitsFromBits(CGF.getContext().getOpenMPDefaultSimdAlign(
E->getType()->getPointeeType()))
.getQuantity();
}
assert((Alignment == 0 || Alignment.isPowerOf2()) &&
"alignment is not power of 2");
if (Alignment != 0) {
llvm::Value *PtrValue = CGF.EmitScalarExpr(E);
CGF.EmitAlignmentAssumption(
PtrValue, E, /*No second loc needed*/ SourceLocation(),
llvm::ConstantInt::get(CGF.getLLVMContext(), Alignment));
}
}
}
}
void CodeGenFunction::EmitOMPPrivateLoopCounters(
const OMPLoopDirective &S, CodeGenFunction::OMPPrivateScope &LoopScope) {
if (!HaveInsertPoint())
return;
auto I = S.private_counters().begin();
for (const Expr *E : S.counters()) {
const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl());
const auto *PrivateVD = cast<VarDecl>(cast<DeclRefExpr>(*I)->getDecl());
// Emit var without initialization.
AutoVarEmission VarEmission = EmitAutoVarAlloca(*PrivateVD);
EmitAutoVarCleanups(VarEmission);
LocalDeclMap.erase(PrivateVD);
(void)LoopScope.addPrivate(VD, [&VarEmission]() {
return VarEmission.getAllocatedAddress();
});
if (LocalDeclMap.count(VD) || CapturedStmtInfo->lookup(VD) ||
VD->hasGlobalStorage()) {
(void)LoopScope.addPrivate(PrivateVD, [this, VD, E]() {
DeclRefExpr DRE(getContext(), const_cast<VarDecl *>(VD),
LocalDeclMap.count(VD) || CapturedStmtInfo->lookup(VD),
E->getType(), VK_LValue, E->getExprLoc());
return EmitLValue(&DRE).getAddress();
});
} else {
(void)LoopScope.addPrivate(PrivateVD, [&VarEmission]() {
return VarEmission.getAllocatedAddress();
});
}
++I;
}
// Privatize extra loop counters used in loops for ordered(n) clauses.
for (const auto *C : S.getClausesOfKind<OMPOrderedClause>()) {
if (!C->getNumForLoops())
continue;
for (unsigned I = S.getCollapsedNumber(),
E = C->getLoopNumIterations().size();
I < E; ++I) {
const auto *DRE = cast<DeclRefExpr>(C->getLoopCounter(I));
const auto *VD = cast<VarDecl>(DRE->getDecl());
// Override only those variables that can be captured to avoid re-emission
// of the variables declared within the loops.
if (DRE->refersToEnclosingVariableOrCapture()) {
(void)LoopScope.addPrivate(VD, [this, DRE, VD]() {
return CreateMemTemp(DRE->getType(), VD->getName());
});
}
}
}
}
static void emitPreCond(CodeGenFunction &CGF, const OMPLoopDirective &S,
const Expr *Cond, llvm::BasicBlock *TrueBlock,
llvm::BasicBlock *FalseBlock, uint64_t TrueCount) {
if (!CGF.HaveInsertPoint())
return;
{
CodeGenFunction::OMPPrivateScope PreCondScope(CGF);
CGF.EmitOMPPrivateLoopCounters(S, PreCondScope);
(void)PreCondScope.Privatize();
// Get initial values of real counters.
for (const Expr *I : S.inits()) {
CGF.EmitIgnoredExpr(I);
}
}
// Create temp loop control variables with their init values to support
// non-rectangular loops.
CodeGenFunction::OMPMapVars PreCondVars;
for (const Expr * E: S.dependent_counters()) {
if (!E)
continue;
assert(!E->getType().getNonReferenceType()->isRecordType() &&
"dependent counter must not be an iterator.");
const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl());
Address CounterAddr =
CGF.CreateMemTemp(VD->getType().getNonReferenceType());
(void)PreCondVars.setVarAddr(CGF, VD, CounterAddr);
}
(void)PreCondVars.apply(CGF);
for (const Expr *E : S.dependent_inits()) {
if (!E)
continue;
CGF.EmitIgnoredExpr(E);
}
// Check that loop is executed at least one time.
CGF.EmitBranchOnBoolExpr(Cond, TrueBlock, FalseBlock, TrueCount);
PreCondVars.restore(CGF);
}
void CodeGenFunction::EmitOMPLinearClause(
const OMPLoopDirective &D, CodeGenFunction::OMPPrivateScope &PrivateScope) {
if (!HaveInsertPoint())
return;
llvm::DenseSet<const VarDecl *> SIMDLCVs;
if (isOpenMPSimdDirective(D.getDirectiveKind())) {
const auto *LoopDirective = cast<OMPLoopDirective>(&D);
for (const Expr *C : LoopDirective->counters()) {
SIMDLCVs.insert(
cast<VarDecl>(cast<DeclRefExpr>(C)->getDecl())->getCanonicalDecl());
}
}
for (const auto *C : D.getClausesOfKind<OMPLinearClause>()) {
auto CurPrivate = C->privates().begin();
for (const Expr *E : C->varlists()) {
const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl());
const auto *PrivateVD =
cast<VarDecl>(cast<DeclRefExpr>(*CurPrivate)->getDecl());
if (!SIMDLCVs.count(VD->getCanonicalDecl())) {
bool IsRegistered = PrivateScope.addPrivate(VD, [this, PrivateVD]() {
// Emit private VarDecl with copy init.
EmitVarDecl(*PrivateVD);
return GetAddrOfLocalVar(PrivateVD);
});
assert(IsRegistered && "linear var already registered as private");
// Silence the warning about unused variable.
(void)IsRegistered;
} else {
EmitVarDecl(*PrivateVD);
}
++CurPrivate;
}
}
}
static void emitSimdlenSafelenClause(CodeGenFunction &CGF,
const OMPExecutableDirective &D,
bool IsMonotonic) {
if (!CGF.HaveInsertPoint())
return;
if (const auto *C = D.getSingleClause<OMPSimdlenClause>()) {
RValue Len = CGF.EmitAnyExpr(C->getSimdlen(), AggValueSlot::ignored(),
/*ignoreResult=*/true);
auto *Val = cast<llvm::ConstantInt>(Len.getScalarVal());
CGF.LoopStack.setVectorizeWidth(Val->getZExtValue());
// In presence of finite 'safelen', it may be unsafe to mark all
// the memory instructions parallel, because loop-carried
// dependences of 'safelen' iterations are possible.
if (!IsMonotonic)
CGF.LoopStack.setParallel(!D.getSingleClause<OMPSafelenClause>());
} else if (const auto *C = D.getSingleClause<OMPSafelenClause>()) {
RValue Len = CGF.EmitAnyExpr(C->getSafelen(), AggValueSlot::ignored(),
/*ignoreResult=*/true);
auto *Val = cast<llvm::ConstantInt>(Len.getScalarVal());
CGF.LoopStack.setVectorizeWidth(Val->getZExtValue());
// In presence of finite 'safelen', it may be unsafe to mark all
// the memory instructions parallel, because loop-carried
// dependences of 'safelen' iterations are possible.
CGF.LoopStack.setParallel(/*Enable=*/false);
}
}
void CodeGenFunction::EmitOMPSimdInit(const OMPLoopDirective &D,
bool IsMonotonic) {
// Walk clauses and process safelen/lastprivate.
LoopStack.setParallel(!IsMonotonic);
LoopStack.setVectorizeEnable();
emitSimdlenSafelenClause(*this, D, IsMonotonic);
}
void CodeGenFunction::EmitOMPSimdFinal(
const OMPLoopDirective &D,
const llvm::function_ref<llvm::Value *(CodeGenFunction &)> CondGen) {
if (!HaveInsertPoint())
return;
llvm::BasicBlock *DoneBB = nullptr;
auto IC = D.counters().begin();
auto IPC = D.private_counters().begin();
for (const Expr *F : D.finals()) {
const auto *OrigVD = cast<VarDecl>(cast<DeclRefExpr>((*IC))->getDecl());
const auto *PrivateVD = cast<VarDecl>(cast<DeclRefExpr>((*IPC))->getDecl());
const auto *CED = dyn_cast<OMPCapturedExprDecl>(OrigVD);
if (LocalDeclMap.count(OrigVD) || CapturedStmtInfo->lookup(OrigVD) ||
OrigVD->hasGlobalStorage() || CED) {
if (!DoneBB) {
if (llvm::Value *Cond = CondGen(*this)) {
// If the first post-update expression is found, emit conditional
// block if it was requested.
llvm::BasicBlock *ThenBB = createBasicBlock(".omp.final.then");
DoneBB = createBasicBlock(".omp.final.done");
Builder.CreateCondBr(Cond, ThenBB, DoneBB);
EmitBlock(ThenBB);
}
}
Address OrigAddr = Address::invalid();
if (CED) {
OrigAddr = EmitLValue(CED->getInit()->IgnoreImpCasts()).getAddress();
} else {
DeclRefExpr DRE(getContext(), const_cast<VarDecl *>(PrivateVD),
/*RefersToEnclosingVariableOrCapture=*/false,
(*IPC)->getType(), VK_LValue, (*IPC)->getExprLoc());
OrigAddr = EmitLValue(&DRE).getAddress();
}
OMPPrivateScope VarScope(*this);
VarScope.addPrivate(OrigVD, [OrigAddr]() { return OrigAddr; });
(void)VarScope.Privatize();
EmitIgnoredExpr(F);
}
++IC;
++IPC;
}
if (DoneBB)
EmitBlock(DoneBB, /*IsFinished=*/true);
}
static void emitOMPLoopBodyWithStopPoint(CodeGenFunction &CGF,
const OMPLoopDirective &S,
CodeGenFunction::JumpDest LoopExit) {
CGF.EmitOMPLoopBody(S, LoopExit);
CGF.EmitStopPoint(&S);
}
/// Emit a helper variable and return corresponding lvalue.
static LValue EmitOMPHelperVar(CodeGenFunction &CGF,
const DeclRefExpr *Helper) {
auto VDecl = cast<VarDecl>(Helper->getDecl());
CGF.EmitVarDecl(*VDecl);
return CGF.EmitLValue(Helper);
}
static void emitOMPSimdRegion(CodeGenFunction &CGF, const OMPLoopDirective &S,
PrePostActionTy &Action) {
Action.Enter(CGF);
assert(isOpenMPSimdDirective(S.getDirectiveKind()) &&
"Expected simd directive");
OMPLoopScope PreInitScope(CGF, S);
// if (PreCond) {
// for (IV in 0..LastIteration) BODY;
// <Final counter/linear vars updates>;
// }
//
if (isOpenMPDistributeDirective(S.getDirectiveKind()) ||
isOpenMPWorksharingDirective(S.getDirectiveKind()) ||
isOpenMPTaskLoopDirective(S.getDirectiveKind())) {
(void)EmitOMPHelperVar(CGF, cast<DeclRefExpr>(S.getLowerBoundVariable()));
(void)EmitOMPHelperVar(CGF, cast<DeclRefExpr>(S.getUpperBoundVariable()));
}
// Emit: if (PreCond) - begin.
// If the condition constant folds and can be elided, avoid emitting the
// whole loop.
bool CondConstant;
llvm::BasicBlock *ContBlock = nullptr;
if (CGF.ConstantFoldsToSimpleInteger(S.getPreCond(), CondConstant)) {
if (!CondConstant)
return;
} else {
llvm::BasicBlock *ThenBlock = CGF.createBasicBlock("simd.if.then");
ContBlock = CGF.createBasicBlock("simd.if.end");
emitPreCond(CGF, S, S.getPreCond(), ThenBlock, ContBlock,
CGF.getProfileCount(&S));
CGF.EmitBlock(ThenBlock);
CGF.incrementProfileCounter(&S);
}
// Emit the loop iteration variable.
const Expr *IVExpr = S.getIterationVariable();
const auto *IVDecl = cast<VarDecl>(cast<DeclRefExpr>(IVExpr)->getDecl());
CGF.EmitVarDecl(*IVDecl);
CGF.EmitIgnoredExpr(S.getInit());
// Emit the iterations count variable.
// If it is not a variable, Sema decided to calculate iterations count on
// each iteration (e.g., it is foldable into a constant).
if (const auto *LIExpr = dyn_cast<DeclRefExpr>(S.getLastIteration())) {
CGF.EmitVarDecl(*cast<VarDecl>(LIExpr->getDecl()));
// Emit calculation of the iterations count.
CGF.EmitIgnoredExpr(S.getCalcLastIteration());
}
CGF.EmitOMPSimdInit(S);
emitAlignedClause(CGF, S);
(void)CGF.EmitOMPLinearClauseInit(S);
{
CodeGenFunction::OMPPrivateScope LoopScope(CGF);
CGF.EmitOMPPrivateLoopCounters(S, LoopScope);
CGF.EmitOMPLinearClause(S, LoopScope);
CGF.EmitOMPPrivateClause(S, LoopScope);
CGF.EmitOMPReductionClauseInit(S, LoopScope);
bool HasLastprivateClause = CGF.EmitOMPLastprivateClauseInit(S, LoopScope);
(void)LoopScope.Privatize();
if (isOpenMPTargetExecutionDirective(S.getDirectiveKind()))
CGF.CGM.getOpenMPRuntime().adjustTargetSpecificDataForLambdas(CGF, S);
CGF.EmitOMPInnerLoop(S, LoopScope.requiresCleanups(), S.getCond(),
S.getInc(),
[&S](CodeGenFunction &CGF) {
CGF.EmitOMPLoopBody(S, CodeGenFunction::JumpDest());
CGF.EmitStopPoint(&S);
},
[](CodeGenFunction &) {});
CGF.EmitOMPSimdFinal(S, [](CodeGenFunction &) { return nullptr; });
// Emit final copy of the lastprivate variables at the end of loops.
if (HasLastprivateClause)
CGF.EmitOMPLastprivateClauseFinal(S, /*NoFinals=*/true);
CGF.EmitOMPReductionClauseFinal(S, /*ReductionKind=*/OMPD_simd);
emitPostUpdateForReductionClause(CGF, S,
[](CodeGenFunction &) { return nullptr; });
}
CGF.EmitOMPLinearClauseFinal(S, [](CodeGenFunction &) { return nullptr; });
// Emit: if (PreCond) - end.
if (ContBlock) {
CGF.EmitBranch(ContBlock);
CGF.EmitBlock(ContBlock, true);
}
}
void CodeGenFunction::EmitOMPSimdDirective(const OMPSimdDirective &S) {
auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
emitOMPSimdRegion(CGF, S, Action);
};
OMPLexicalScope Scope(*this, S, OMPD_unknown);
CGM.getOpenMPRuntime().emitInlinedDirective(*this, OMPD_simd, CodeGen);
}
void CodeGenFunction::EmitOMPOuterLoop(
bool DynamicOrOrdered, bool IsMonotonic, const OMPLoopDirective &S,
CodeGenFunction::OMPPrivateScope &LoopScope,
const CodeGenFunction::OMPLoopArguments &LoopArgs,
const CodeGenFunction::CodeGenLoopTy &CodeGenLoop,
const CodeGenFunction::CodeGenOrderedTy &CodeGenOrdered) {
CGOpenMPRuntime &RT = CGM.getOpenMPRuntime();
const Expr *IVExpr = S.getIterationVariable();
const unsigned IVSize = getContext().getTypeSize(IVExpr->getType());
const bool IVSigned = IVExpr->getType()->hasSignedIntegerRepresentation();
JumpDest LoopExit = getJumpDestInCurrentScope("omp.dispatch.end");
// Start the loop with a block that tests the condition.
llvm::BasicBlock *CondBlock = createBasicBlock("omp.dispatch.cond");
EmitBlock(CondBlock);
const SourceRange R = S.getSourceRange();
LoopStack.push(CondBlock, SourceLocToDebugLoc(R.getBegin()),
SourceLocToDebugLoc(R.getEnd()));
llvm::Value *BoolCondVal = nullptr;
if (!DynamicOrOrdered) {
// UB = min(UB, GlobalUB) or
// UB = min(UB, PrevUB) for combined loop sharing constructs (e.g.
// 'distribute parallel for')
EmitIgnoredExpr(LoopArgs.EUB);
// IV = LB
EmitIgnoredExpr(LoopArgs.Init);
// IV < UB
BoolCondVal = EvaluateExprAsBool(LoopArgs.Cond);
} else {
BoolCondVal =
RT.emitForNext(*this, S.getBeginLoc(), IVSize, IVSigned, LoopArgs.IL,
LoopArgs.LB, LoopArgs.UB, LoopArgs.ST);
}
// If there are any cleanups between here and the loop-exit scope,
// create a block to stage a loop exit along.
llvm::BasicBlock *ExitBlock = LoopExit.getBlock();
if (LoopScope.requiresCleanups())
ExitBlock = createBasicBlock("omp.dispatch.cleanup");
llvm::BasicBlock *LoopBody = createBasicBlock("omp.dispatch.body");
Builder.CreateCondBr(BoolCondVal, LoopBody, ExitBlock);
if (ExitBlock != LoopExit.getBlock()) {
EmitBlock(ExitBlock);
EmitBranchThroughCleanup(LoopExit);
}
EmitBlock(LoopBody);
// Emit "IV = LB" (in case of static schedule, we have already calculated new
// LB for loop condition and emitted it above).
if (DynamicOrOrdered)
EmitIgnoredExpr(LoopArgs.Init);
// Create a block for the increment.
JumpDest Continue = getJumpDestInCurrentScope("omp.dispatch.inc");
BreakContinueStack.push_back(BreakContinue(LoopExit, Continue));
// Generate !llvm.loop.parallel metadata for loads and stores for loops
// with dynamic/guided scheduling and without ordered clause.
if (!isOpenMPSimdDirective(S.getDirectiveKind()))
LoopStack.setParallel(!IsMonotonic);
else
EmitOMPSimdInit(S, IsMonotonic);
SourceLocation Loc = S.getBeginLoc();
// when 'distribute' is not combined with a 'for':
// while (idx <= UB) { BODY; ++idx; }
// when 'distribute' is combined with a 'for'
// (e.g. 'distribute parallel for')
// while (idx <= UB) { <CodeGen rest of pragma>; idx += ST; }
EmitOMPInnerLoop(
S, LoopScope.requiresCleanups(), LoopArgs.Cond, LoopArgs.IncExpr,
[&S, LoopExit, &CodeGenLoop](CodeGenFunction &CGF) {
CodeGenLoop(CGF, S, LoopExit);
},
[IVSize, IVSigned, Loc, &CodeGenOrdered](CodeGenFunction &CGF) {
CodeGenOrdered(CGF, Loc, IVSize, IVSigned);
});
EmitBlock(Continue.getBlock());
BreakContinueStack.pop_back();
if (!DynamicOrOrdered) {
// Emit "LB = LB + Stride", "UB = UB + Stride".
EmitIgnoredExpr(LoopArgs.NextLB);
EmitIgnoredExpr(LoopArgs.NextUB);
}
EmitBranch(CondBlock);
LoopStack.pop();
// Emit the fall-through block.
EmitBlock(LoopExit.getBlock());
// Tell the runtime we are done.
auto &&CodeGen = [DynamicOrOrdered, &S](CodeGenFunction &CGF) {
if (!DynamicOrOrdered)
CGF.CGM.getOpenMPRuntime().emitForStaticFinish(CGF, S.getEndLoc(),
S.getDirectiveKind());
};
OMPCancelStack.emitExit(*this, S.getDirectiveKind(), CodeGen);
}
void CodeGenFunction::EmitOMPForOuterLoop(
const OpenMPScheduleTy &ScheduleKind, bool IsMonotonic,
const OMPLoopDirective &S, OMPPrivateScope &LoopScope, bool Ordered,
const OMPLoopArguments &LoopArgs,
const CodeGenDispatchBoundsTy &CGDispatchBounds) {
CGOpenMPRuntime &RT = CGM.getOpenMPRuntime();
// Dynamic scheduling of the outer loop (dynamic, guided, auto, runtime).
const bool DynamicOrOrdered =
Ordered || RT.isDynamic(ScheduleKind.Schedule);
assert((Ordered ||
!RT.isStaticNonchunked(ScheduleKind.Schedule,
LoopArgs.Chunk != nullptr)) &&
"static non-chunked schedule does not need outer loop");
// Emit outer loop.
//
// OpenMP [2.7.1, Loop Construct, Description, table 2-1]
// When schedule(dynamic,chunk_size) is specified, the iterations are
// distributed to threads in the team in chunks as the threads request them.
// Each thread executes a chunk of iterations, then requests another chunk,
// until no chunks remain to be distributed. Each chunk contains chunk_size
// iterations, except for the last chunk to be distributed, which may have
// fewer iterations. When no chunk_size is specified, it defaults to 1.
//
// When schedule(guided,chunk_size) is specified, the iterations are assigned
// to threads in the team in chunks as the executing threads request them.
// Each thread executes a chunk of iterations, then requests another chunk,
// until no chunks remain to be assigned. For a chunk_size of 1, the size of
// each chunk is proportional to the number of unassigned iterations divided
// by the number of threads in the team, decreasing to 1. For a chunk_size
// with value k (greater than 1), the size of each chunk is determined in the
// same way, with the restriction that the chunks do not contain fewer than k
// iterations (except for the last chunk to be assigned, which may have fewer
// than k iterations).
//
// When schedule(auto) is specified, the decision regarding scheduling is
// delegated to the compiler and/or runtime system. The programmer gives the
// implementation the freedom to choose any possible mapping of iterations to
// threads in the team.
//
// When schedule(runtime) is specified, the decision regarding scheduling is
// deferred until run time, and the schedule and chunk size are taken from the
// run-sched-var ICV. If the ICV is set to auto, the schedule is
// implementation defined
//
// while(__kmpc_dispatch_next(&LB, &UB)) {
// idx = LB;
// while (idx <= UB) { BODY; ++idx;
// __kmpc_dispatch_fini_(4|8)[u](); // For ordered loops only.
// } // inner loop
// }
//
// OpenMP [2.7.1, Loop Construct, Description, table 2-1]
// When schedule(static, chunk_size) is specified, iterations are divided into
// chunks of size chunk_size, and the chunks are assigned to the threads in
// the team in a round-robin fashion in the order of the thread number.
//
// while(UB = min(UB, GlobalUB), idx = LB, idx < UB) {
// while (idx <= UB) { BODY; ++idx; } // inner loop
// LB = LB + ST;
// UB = UB + ST;
// }
//
const Expr *IVExpr = S.getIterationVariable();
const unsigned IVSize = getContext().getTypeSize(IVExpr->getType());
const bool IVSigned = IVExpr->getType()->hasSignedIntegerRepresentation();
if (DynamicOrOrdered) {
const std::pair<llvm::Value *, llvm::Value *> DispatchBounds =
CGDispatchBounds(*this, S, LoopArgs.LB, LoopArgs.UB);
llvm::Value *LBVal = DispatchBounds.first;
llvm::Value *UBVal = DispatchBounds.second;
CGOpenMPRuntime::DispatchRTInput DipatchRTInputValues = {LBVal, UBVal,
LoopArgs.Chunk};
RT.emitForDispatchInit(*this, S.getBeginLoc(), ScheduleKind, IVSize,
IVSigned, Ordered, DipatchRTInputValues);
} else {
CGOpenMPRuntime::StaticRTInput StaticInit(
IVSize, IVSigned, Ordered, LoopArgs.IL, LoopArgs.LB, LoopArgs.UB,
LoopArgs.ST, LoopArgs.Chunk);
RT.emitForStaticInit(*this, S.getBeginLoc(), S.getDirectiveKind(),
ScheduleKind, StaticInit);
}
auto &&CodeGenOrdered = [Ordered](CodeGenFunction &CGF, SourceLocation Loc,
const unsigned IVSize,
const bool IVSigned) {
if (Ordered) {
CGF.CGM.getOpenMPRuntime().emitForOrderedIterationEnd(CGF, Loc, IVSize,
IVSigned);
}
};
OMPLoopArguments OuterLoopArgs(LoopArgs.LB, LoopArgs.UB, LoopArgs.ST,
LoopArgs.IL, LoopArgs.Chunk, LoopArgs.EUB);
OuterLoopArgs.IncExpr = S.getInc();
OuterLoopArgs.Init = S.getInit();
OuterLoopArgs.Cond = S.getCond();
OuterLoopArgs.NextLB = S.getNextLowerBound();
OuterLoopArgs.NextUB = S.getNextUpperBound();
EmitOMPOuterLoop(DynamicOrOrdered, IsMonotonic, S, LoopScope, OuterLoopArgs,
emitOMPLoopBodyWithStopPoint, CodeGenOrdered);
}
static void emitEmptyOrdered(CodeGenFunction &, SourceLocation Loc,
const unsigned IVSize, const bool IVSigned) {}
void CodeGenFunction::EmitOMPDistributeOuterLoop(
OpenMPDistScheduleClauseKind ScheduleKind, const OMPLoopDirective &S,
OMPPrivateScope &LoopScope, const OMPLoopArguments &LoopArgs,
const CodeGenLoopTy &CodeGenLoopContent) {
CGOpenMPRuntime &RT = CGM.getOpenMPRuntime();
// Emit outer loop.
// Same behavior as a OMPForOuterLoop, except that schedule cannot be
// dynamic
//
const Expr *IVExpr = S.getIterationVariable();
const unsigned IVSize = getContext().getTypeSize(IVExpr->getType());
const bool IVSigned = IVExpr->getType()->hasSignedIntegerRepresentation();
CGOpenMPRuntime::StaticRTInput StaticInit(
IVSize, IVSigned, /* Ordered = */ false, LoopArgs.IL, LoopArgs.LB,
LoopArgs.UB, LoopArgs.ST, LoopArgs.Chunk);
RT.emitDistributeStaticInit(*this, S.getBeginLoc(), ScheduleKind, StaticInit);
// for combined 'distribute' and 'for' the increment expression of distribute
// is stored in DistInc. For 'distribute' alone, it is in Inc.
Expr *IncExpr;
if (isOpenMPLoopBoundSharingDirective(S.getDirectiveKind()))
IncExpr = S.getDistInc();
else
IncExpr = S.getInc();
// this routine is shared by 'omp distribute parallel for' and
// 'omp distribute': select the right EUB expression depending on the
// directive
OMPLoopArguments OuterLoopArgs;
OuterLoopArgs.LB = LoopArgs.LB;
OuterLoopArgs.UB = LoopArgs.UB;
OuterLoopArgs.ST = LoopArgs.ST;
OuterLoopArgs.IL = LoopArgs.IL;
OuterLoopArgs.Chunk = LoopArgs.Chunk;
OuterLoopArgs.EUB = isOpenMPLoopBoundSharingDirective(S.getDirectiveKind())
? S.getCombinedEnsureUpperBound()
: S.getEnsureUpperBound();
OuterLoopArgs.IncExpr = IncExpr;
OuterLoopArgs.Init = isOpenMPLoopBoundSharingDirective(S.getDirectiveKind())
? S.getCombinedInit()
: S.getInit();
OuterLoopArgs.Cond = isOpenMPLoopBoundSharingDirective(S.getDirectiveKind())
? S.getCombinedCond()
: S.getCond();
OuterLoopArgs.NextLB = isOpenMPLoopBoundSharingDirective(S.getDirectiveKind())
? S.getCombinedNextLowerBound()
: S.getNextLowerBound();
OuterLoopArgs.NextUB = isOpenMPLoopBoundSharingDirective(S.getDirectiveKind())
? S.getCombinedNextUpperBound()
: S.getNextUpperBound();
EmitOMPOuterLoop(/* DynamicOrOrdered = */ false, /* IsMonotonic = */ false, S,
LoopScope, OuterLoopArgs, CodeGenLoopContent,
emitEmptyOrdered);
}
static std::pair<LValue, LValue>
emitDistributeParallelForInnerBounds(CodeGenFunction &CGF,
const OMPExecutableDirective &S) {
const OMPLoopDirective &LS = cast<OMPLoopDirective>(S);
LValue LB =
EmitOMPHelperVar(CGF, cast<DeclRefExpr>(LS.getLowerBoundVariable()));
LValue UB =
EmitOMPHelperVar(CGF, cast<DeclRefExpr>(LS.getUpperBoundVariable()));
// When composing 'distribute' with 'for' (e.g. as in 'distribute
// parallel for') we need to use the 'distribute'
// chunk lower and upper bounds rather than the whole loop iteration
// space. These are parameters to the outlined function for 'parallel'
// and we copy the bounds of the previous schedule into the
// the current ones.
LValue PrevLB = CGF.EmitLValue(LS.getPrevLowerBoundVariable());
LValue PrevUB = CGF.EmitLValue(LS.getPrevUpperBoundVariable());
llvm::Value *PrevLBVal = CGF.EmitLoadOfScalar(
PrevLB, LS.getPrevLowerBoundVariable()->getExprLoc());
PrevLBVal = CGF.EmitScalarConversion(
PrevLBVal, LS.getPrevLowerBoundVariable()->getType(),
LS.getIterationVariable()->getType(),
LS.getPrevLowerBoundVariable()->getExprLoc());
llvm::Value *PrevUBVal = CGF.EmitLoadOfScalar(
PrevUB, LS.getPrevUpperBoundVariable()->getExprLoc());
PrevUBVal = CGF.EmitScalarConversion(
PrevUBVal, LS.getPrevUpperBoundVariable()->getType(),
LS.getIterationVariable()->getType(),
LS.getPrevUpperBoundVariable()->getExprLoc());
CGF.EmitStoreOfScalar(PrevLBVal, LB);
CGF.EmitStoreOfScalar(PrevUBVal, UB);
return {LB, UB};
}
/// if the 'for' loop has a dispatch schedule (e.g. dynamic, guided) then
/// we need to use the LB and UB expressions generated by the worksharing
/// code generation support, whereas in non combined situations we would
/// just emit 0 and the LastIteration expression
/// This function is necessary due to the difference of the LB and UB
/// types for the RT emission routines for 'for_static_init' and
/// 'for_dispatch_init'
static std::pair<llvm::Value *, llvm::Value *>
emitDistributeParallelForDispatchBounds(CodeGenFunction &CGF,
const OMPExecutableDirective &S,
Address LB, Address UB) {
const OMPLoopDirective &LS = cast<OMPLoopDirective>(S);
const Expr *IVExpr = LS.getIterationVariable();
// when implementing a dynamic schedule for a 'for' combined with a
// 'distribute' (e.g. 'distribute parallel for'), the 'for' loop
// is not normalized as each team only executes its own assigned
// distribute chunk
QualType IteratorTy = IVExpr->getType();
llvm::Value *LBVal =
CGF.EmitLoadOfScalar(LB, /*Volatile=*/false, IteratorTy, S.getBeginLoc());
llvm::Value *UBVal =
CGF.EmitLoadOfScalar(UB, /*Volatile=*/false, IteratorTy, S.getBeginLoc());
return {LBVal, UBVal};
}
static void emitDistributeParallelForDistributeInnerBoundParams(
CodeGenFunction &CGF, const OMPExecutableDirective &S,
llvm::SmallVectorImpl<llvm::Value *> &CapturedVars) {
const auto &Dir = cast<OMPLoopDirective>(S);
LValue LB =
CGF.EmitLValue(cast<DeclRefExpr>(Dir.getCombinedLowerBoundVariable()));
llvm::Value *LBCast = CGF.Builder.CreateIntCast(
CGF.Builder.CreateLoad(LB.getAddress()), CGF.SizeTy, /*isSigned=*/false);
CapturedVars.push_back(LBCast);
LValue UB =
CGF.EmitLValue(cast<DeclRefExpr>(Dir.getCombinedUpperBoundVariable()));
llvm::Value *UBCast = CGF.Builder.CreateIntCast(
CGF.Builder.CreateLoad(UB.getAddress()), CGF.SizeTy, /*isSigned=*/false);
CapturedVars.push_back(UBCast);
}
static void
emitInnerParallelForWhenCombined(CodeGenFunction &CGF,
const OMPLoopDirective &S,
CodeGenFunction::JumpDest LoopExit) {
auto &&CGInlinedWorksharingLoop = [&S](CodeGenFunction &CGF,
PrePostActionTy &Action) {
Action.Enter(CGF);
bool HasCancel = false;
if (!isOpenMPSimdDirective(S.getDirectiveKind())) {
if (const auto *D = dyn_cast<OMPTeamsDistributeParallelForDirective>(&S))
HasCancel = D->hasCancel();
else if (const auto *D = dyn_cast<OMPDistributeParallelForDirective>(&S))
HasCancel = D->hasCancel();
else if (const auto *D =
dyn_cast<OMPTargetTeamsDistributeParallelForDirective>(&S))
HasCancel = D->hasCancel();
}
CodeGenFunction::OMPCancelStackRAII CancelRegion(CGF, S.getDirectiveKind(),
HasCancel);
CGF.EmitOMPWorksharingLoop(S, S.getPrevEnsureUpperBound(),
emitDistributeParallelForInnerBounds,
emitDistributeParallelForDispatchBounds);
};
emitCommonOMPParallelDirective(
CGF, S,
isOpenMPSimdDirective(S.getDirectiveKind()) ? OMPD_for_simd : OMPD_for,
CGInlinedWorksharingLoop,
emitDistributeParallelForDistributeInnerBoundParams);
}
void CodeGenFunction::EmitOMPDistributeParallelForDirective(
const OMPDistributeParallelForDirective &S) {
auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &) {
CGF.EmitOMPDistributeLoop(S, emitInnerParallelForWhenCombined,
S.getDistInc());
};
OMPLexicalScope Scope(*this, S, OMPD_parallel);
CGM.getOpenMPRuntime().emitInlinedDirective(*this, OMPD_distribute, CodeGen);
}
void CodeGenFunction::EmitOMPDistributeParallelForSimdDirective(
const OMPDistributeParallelForSimdDirective &S) {
auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &) {
CGF.EmitOMPDistributeLoop(S, emitInnerParallelForWhenCombined,
S.getDistInc());
};
OMPLexicalScope Scope(*this, S, OMPD_parallel);
CGM.getOpenMPRuntime().emitInlinedDirective(*this, OMPD_distribute, CodeGen);
}
void CodeGenFunction::EmitOMPDistributeSimdDirective(
const OMPDistributeSimdDirective &S) {
auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &) {
CGF.EmitOMPDistributeLoop(S, emitOMPLoopBodyWithStopPoint, S.getInc());
};
OMPLexicalScope Scope(*this, S, OMPD_unknown);
CGM.getOpenMPRuntime().emitInlinedDirective(*this, OMPD_simd, CodeGen);
}
void CodeGenFunction::EmitOMPTargetSimdDeviceFunction(
CodeGenModule &CGM, StringRef ParentName, const OMPTargetSimdDirective &S) {
// Emit SPMD target parallel for region as a standalone region.
auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
emitOMPSimdRegion(CGF, S, Action);
};
llvm::Function *Fn;
llvm::Constant *Addr;
// Emit target region as a standalone region.
CGM.getOpenMPRuntime().emitTargetOutlinedFunction(
S, ParentName, Fn, Addr, /*IsOffloadEntry=*/true, CodeGen);
assert(Fn && Addr && "Target device function emission failed.");
}
void CodeGenFunction::EmitOMPTargetSimdDirective(
const OMPTargetSimdDirective &S) {
auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
emitOMPSimdRegion(CGF, S, Action);
};
emitCommonOMPTargetDirective(*this, S, CodeGen);
}
namespace {
struct ScheduleKindModifiersTy {
OpenMPScheduleClauseKind Kind;
OpenMPScheduleClauseModifier M1;
OpenMPScheduleClauseModifier M2;
ScheduleKindModifiersTy(OpenMPScheduleClauseKind Kind,
OpenMPScheduleClauseModifier M1,
OpenMPScheduleClauseModifier M2)
: Kind(Kind), M1(M1), M2(M2) {}
};
} // namespace
bool CodeGenFunction::EmitOMPWorksharingLoop(
const OMPLoopDirective &S, Expr *EUB,
const CodeGenLoopBoundsTy &CodeGenLoopBounds,
const CodeGenDispatchBoundsTy &CGDispatchBounds) {
// Emit the loop iteration variable.
const auto *IVExpr = cast<DeclRefExpr>(S.getIterationVariable());
const auto *IVDecl = cast<VarDecl>(IVExpr->getDecl());
EmitVarDecl(*IVDecl);
// Emit the iterations count variable.
// If it is not a variable, Sema decided to calculate iterations count on each
// iteration (e.g., it is foldable into a constant).
if (const auto *LIExpr = dyn_cast<DeclRefExpr>(S.getLastIteration())) {
EmitVarDecl(*cast<VarDecl>(LIExpr->getDecl()));
// Emit calculation of the iterations count.
EmitIgnoredExpr(S.getCalcLastIteration());
}
CGOpenMPRuntime &RT = CGM.getOpenMPRuntime();
bool HasLastprivateClause;
// Check pre-condition.
{
OMPLoopScope PreInitScope(*this, S);
// Skip the entire loop if we don't meet the precondition.
// If the condition constant folds and can be elided, avoid emitting the
// whole loop.
bool CondConstant;
llvm::BasicBlock *ContBlock = nullptr;
if (ConstantFoldsToSimpleInteger(S.getPreCond(), CondConstant)) {
if (!CondConstant)
return false;
} else {
llvm::BasicBlock *ThenBlock = createBasicBlock("omp.precond.then");
ContBlock = createBasicBlock("omp.precond.end");
emitPreCond(*this, S, S.getPreCond(), ThenBlock, ContBlock,
getProfileCount(&S));
EmitBlock(ThenBlock);
incrementProfileCounter(&S);
}
RunCleanupsScope DoacrossCleanupScope(*this);
bool Ordered = false;
if (const auto *OrderedClause = S.getSingleClause<OMPOrderedClause>()) {
if (OrderedClause->getNumForLoops())
RT.emitDoacrossInit(*this, S, OrderedClause->getLoopNumIterations());
else
Ordered = true;
}
llvm::DenseSet<const Expr *> EmittedFinals;
emitAlignedClause(*this, S);
bool HasLinears = EmitOMPLinearClauseInit(S);
// Emit helper vars inits.
std::pair<LValue, LValue> Bounds = CodeGenLoopBounds(*this, S);
LValue LB = Bounds.first;
LValue UB = Bounds.second;
LValue ST =
EmitOMPHelperVar(*this, cast<DeclRefExpr>(S.getStrideVariable()));
LValue IL =
EmitOMPHelperVar(*this, cast<DeclRefExpr>(S.getIsLastIterVariable()));
// Emit 'then' code.
{
OMPPrivateScope LoopScope(*this);
if (EmitOMPFirstprivateClause(S, LoopScope) || HasLinears) {
// Emit implicit barrier to synchronize threads and avoid data races on
// initialization of firstprivate variables and post-update of
// lastprivate variables.
CGM.getOpenMPRuntime().emitBarrierCall(
*this, S.getBeginLoc(), OMPD_unknown, /*EmitChecks=*/false,
/*ForceSimpleCall=*/true);
}
EmitOMPPrivateClause(S, LoopScope);
HasLastprivateClause = EmitOMPLastprivateClauseInit(S, LoopScope);
EmitOMPReductionClauseInit(S, LoopScope);
EmitOMPPrivateLoopCounters(S, LoopScope);
EmitOMPLinearClause(S, LoopScope);
(void)LoopScope.Privatize();
if (isOpenMPTargetExecutionDirective(S.getDirectiveKind()))
CGM.getOpenMPRuntime().adjustTargetSpecificDataForLambdas(*this, S);
// Detect the loop schedule kind and chunk.
const Expr *ChunkExpr = nullptr;
OpenMPScheduleTy ScheduleKind;
if (const auto *C = S.getSingleClause<OMPScheduleClause>()) {
ScheduleKind.Schedule = C->getScheduleKind();
ScheduleKind.M1 = C->getFirstScheduleModifier();
ScheduleKind.M2 = C->getSecondScheduleModifier();
ChunkExpr = C->getChunkSize();
} else {
// Default behaviour for schedule clause.
CGM.getOpenMPRuntime().getDefaultScheduleAndChunk(
*this, S, ScheduleKind.Schedule, ChunkExpr);
}
bool HasChunkSizeOne = false;
llvm::Value *Chunk = nullptr;
if (ChunkExpr) {
Chunk = EmitScalarExpr(ChunkExpr);
Chunk = EmitScalarConversion(Chunk, ChunkExpr->getType(),
S.getIterationVariable()->getType(),
S.getBeginLoc());
Expr::EvalResult Result;
if (ChunkExpr->EvaluateAsInt(Result, getContext())) {
llvm::APSInt EvaluatedChunk = Result.Val.getInt();
HasChunkSizeOne = (EvaluatedChunk.getLimitedValue() == 1);
}
}
const unsigned IVSize = getContext().getTypeSize(IVExpr->getType());
const bool IVSigned = IVExpr->getType()->hasSignedIntegerRepresentation();
// OpenMP 4.5, 2.7.1 Loop Construct, Description.
// If the static schedule kind is specified or if the ordered clause is
// specified, and if no monotonic modifier is specified, the effect will
// be as if the monotonic modifier was specified.
bool StaticChunkedOne = RT.isStaticChunked(ScheduleKind.Schedule,
/* Chunked */ Chunk != nullptr) && HasChunkSizeOne &&
isOpenMPLoopBoundSharingDirective(S.getDirectiveKind());
if ((RT.isStaticNonchunked(ScheduleKind.Schedule,
/* Chunked */ Chunk != nullptr) ||
StaticChunkedOne) &&
!Ordered) {
if (isOpenMPSimdDirective(S.getDirectiveKind()))
EmitOMPSimdInit(S, /*IsMonotonic=*/true);
// OpenMP [2.7.1, Loop Construct, Description, table 2-1]
// When no chunk_size is specified, the iteration space is divided into
// chunks that are approximately equal in size, and at most one chunk is
// distributed to each thread. Note that the size of the chunks is
// unspecified in this case.
CGOpenMPRuntime::StaticRTInput StaticInit(
IVSize, IVSigned, Ordered, IL.getAddress(), LB.getAddress(),
UB.getAddress(), ST.getAddress(),
StaticChunkedOne ? Chunk : nullptr);
RT.emitForStaticInit(*this, S.getBeginLoc(), S.getDirectiveKind(),
ScheduleKind, StaticInit);
JumpDest LoopExit =
getJumpDestInCurrentScope(createBasicBlock("omp.loop.exit"));
// UB = min(UB, GlobalUB);
if (!StaticChunkedOne)
EmitIgnoredExpr(S.getEnsureUpperBound());
// IV = LB;
EmitIgnoredExpr(S.getInit());
// For unchunked static schedule generate:
//
// while (idx <= UB) {
// BODY;
// ++idx;
// }
//
// For static schedule with chunk one:
//
// while (IV <= PrevUB) {
// BODY;
// IV += ST;
// }
EmitOMPInnerLoop(S, LoopScope.requiresCleanups(),
StaticChunkedOne ? S.getCombinedParForInDistCond() : S.getCond(),
StaticChunkedOne ? S.getDistInc() : S.getInc(),
[&S, LoopExit](CodeGenFunction &CGF) {
CGF.EmitOMPLoopBody(S, LoopExit);
CGF.EmitStopPoint(&S);