Bug Summary

File:clang/lib/CodeGen/CGStmtOpenMP.cpp
Warning:line 5475, column 16
Called C++ object pointer is null

Annotated Source Code

Press '?' to see keyboard shortcuts

clang -cc1 -triple x86_64-pc-linux-gnu -analyze -disable-free -disable-llvm-verifier -discard-value-names -main-file-name CGStmtOpenMP.cpp -analyzer-store=region -analyzer-opt-analyze-nested-blocks -analyzer-checker=core -analyzer-checker=apiModeling -analyzer-checker=unix -analyzer-checker=deadcode -analyzer-checker=cplusplus -analyzer-checker=security.insecureAPI.UncheckedReturn -analyzer-checker=security.insecureAPI.getpw -analyzer-checker=security.insecureAPI.gets -analyzer-checker=security.insecureAPI.mktemp -analyzer-checker=security.insecureAPI.mkstemp -analyzer-checker=security.insecureAPI.vfork -analyzer-checker=nullability.NullPassedToNonnull -analyzer-checker=nullability.NullReturnedFromNonnull -analyzer-output plist -w -setup-static-analyzer -analyzer-config-compatibility-mode=true -mrelocation-model pic -pic-level 2 -mthread-model posix -mframe-pointer=none -relaxed-aliasing -fmath-errno -fno-rounding-math -masm-verbose -mconstructor-aliases -munwind-tables -target-cpu x86-64 -dwarf-column-info -fno-split-dwarf-inlining -debugger-tuning=gdb -ffunction-sections -fdata-sections -resource-dir /usr/lib/llvm-10/lib/clang/10.0.0 -D CLANG_VENDOR="Debian " -D _DEBUG -D _GNU_SOURCE -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -I /build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/build-llvm/tools/clang/lib/CodeGen -I /build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen -I /build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/include -I /build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/build-llvm/tools/clang/include -I /build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/build-llvm/include -I /build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/llvm/include -U NDEBUG -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/x86_64-linux-gnu/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/x86_64-linux-gnu/c++/6.3.0 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/6.3.0/../../../../include/c++/6.3.0/backward -internal-isystem /usr/local/include -internal-isystem /usr/lib/llvm-10/lib/clang/10.0.0/include -internal-externc-isystem /usr/include/x86_64-linux-gnu -internal-externc-isystem /include -internal-externc-isystem /usr/include -O2 -Wno-unused-parameter -Wwrite-strings -Wno-missing-field-initializers -Wno-long-long -Wno-maybe-uninitialized -Wno-comment -std=c++14 -fdeprecated-macro -fdebug-compilation-dir /build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/build-llvm/tools/clang/lib/CodeGen -fdebug-prefix-map=/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814=. -ferror-limit 19 -fmessage-length 0 -fvisibility-inlines-hidden -stack-protector 2 -fgnuc-version=4.2.1 -fobjc-runtime=gcc -fno-common -fdiagnostics-show-option -vectorize-loops -vectorize-slp -analyzer-output=html -analyzer-config stable-report-filename=true -faddrsig -o /tmp/scan-build-2020-01-11-115256-23437-1 -x c++ /build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp
1//===--- CGStmtOpenMP.cpp - Emit LLVM Code from Statements ----------------===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This contains code to emit OpenMP nodes as LLVM code.
10//
11//===----------------------------------------------------------------------===//
12
13#include "CGCleanup.h"
14#include "CGOpenMPRuntime.h"
15#include "CodeGenFunction.h"
16#include "CodeGenModule.h"
17#include "TargetInfo.h"
18#include "clang/AST/ASTContext.h"
19#include "clang/AST/Attr.h"
20#include "clang/AST/DeclOpenMP.h"
21#include "clang/AST/Stmt.h"
22#include "clang/AST/StmtOpenMP.h"
23#include "clang/Basic/PrettyStackTrace.h"
24#include "llvm/Frontend/OpenMP/OMPIRBuilder.h"
25using namespace clang;
26using namespace CodeGen;
27using namespace llvm::omp;
28
29namespace {
30/// Lexical scope for OpenMP executable constructs, that handles correct codegen
31/// for captured expressions.
32class OMPLexicalScope : public CodeGenFunction::LexicalScope {
33 void emitPreInitStmt(CodeGenFunction &CGF, const OMPExecutableDirective &S) {
34 for (const auto *C : S.clauses()) {
35 if (const auto *CPI = OMPClauseWithPreInit::get(C)) {
36 if (const auto *PreInit =
37 cast_or_null<DeclStmt>(CPI->getPreInitStmt())) {
38 for (const auto *I : PreInit->decls()) {
39 if (!I->hasAttr<OMPCaptureNoInitAttr>()) {
40 CGF.EmitVarDecl(cast<VarDecl>(*I));
41 } else {
42 CodeGenFunction::AutoVarEmission Emission =
43 CGF.EmitAutoVarAlloca(cast<VarDecl>(*I));
44 CGF.EmitAutoVarCleanups(Emission);
45 }
46 }
47 }
48 }
49 }
50 }
51 CodeGenFunction::OMPPrivateScope InlinedShareds;
52
53 static bool isCapturedVar(CodeGenFunction &CGF, const VarDecl *VD) {
54 return CGF.LambdaCaptureFields.lookup(VD) ||
55 (CGF.CapturedStmtInfo && CGF.CapturedStmtInfo->lookup(VD)) ||
56 (CGF.CurCodeDecl && isa<BlockDecl>(CGF.CurCodeDecl));
57 }
58
59public:
60 OMPLexicalScope(
61 CodeGenFunction &CGF, const OMPExecutableDirective &S,
62 const llvm::Optional<OpenMPDirectiveKind> CapturedRegion = llvm::None,
63 const bool EmitPreInitStmt = true)
64 : CodeGenFunction::LexicalScope(CGF, S.getSourceRange()),
65 InlinedShareds(CGF) {
66 if (EmitPreInitStmt)
67 emitPreInitStmt(CGF, S);
68 if (!CapturedRegion.hasValue())
69 return;
70 assert(S.hasAssociatedStmt() &&((S.hasAssociatedStmt() && "Expected associated statement for inlined directive."
) ? static_cast<void> (0) : __assert_fail ("S.hasAssociatedStmt() && \"Expected associated statement for inlined directive.\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 71, __PRETTY_FUNCTION__))
71 "Expected associated statement for inlined directive.")((S.hasAssociatedStmt() && "Expected associated statement for inlined directive."
) ? static_cast<void> (0) : __assert_fail ("S.hasAssociatedStmt() && \"Expected associated statement for inlined directive.\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 71, __PRETTY_FUNCTION__))
;
72 const CapturedStmt *CS = S.getCapturedStmt(*CapturedRegion);
73 for (const auto &C : CS->captures()) {
74 if (C.capturesVariable() || C.capturesVariableByCopy()) {
75 auto *VD = C.getCapturedVar();
76 assert(VD == VD->getCanonicalDecl() &&((VD == VD->getCanonicalDecl() && "Canonical decl must be captured."
) ? static_cast<void> (0) : __assert_fail ("VD == VD->getCanonicalDecl() && \"Canonical decl must be captured.\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 77, __PRETTY_FUNCTION__))
77 "Canonical decl must be captured.")((VD == VD->getCanonicalDecl() && "Canonical decl must be captured."
) ? static_cast<void> (0) : __assert_fail ("VD == VD->getCanonicalDecl() && \"Canonical decl must be captured.\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 77, __PRETTY_FUNCTION__))
;
78 DeclRefExpr DRE(
79 CGF.getContext(), const_cast<VarDecl *>(VD),
80 isCapturedVar(CGF, VD) || (CGF.CapturedStmtInfo &&
81 InlinedShareds.isGlobalVarCaptured(VD)),
82 VD->getType().getNonReferenceType(), VK_LValue, C.getLocation());
83 InlinedShareds.addPrivate(VD, [&CGF, &DRE]() -> Address {
84 return CGF.EmitLValue(&DRE).getAddress(CGF);
85 });
86 }
87 }
88 (void)InlinedShareds.Privatize();
89 }
90};
91
92/// Lexical scope for OpenMP parallel construct, that handles correct codegen
93/// for captured expressions.
94class OMPParallelScope final : public OMPLexicalScope {
95 bool EmitPreInitStmt(const OMPExecutableDirective &S) {
96 OpenMPDirectiveKind Kind = S.getDirectiveKind();
97 return !(isOpenMPTargetExecutionDirective(Kind) ||
98 isOpenMPLoopBoundSharingDirective(Kind)) &&
99 isOpenMPParallelDirective(Kind);
100 }
101
102public:
103 OMPParallelScope(CodeGenFunction &CGF, const OMPExecutableDirective &S)
104 : OMPLexicalScope(CGF, S, /*CapturedRegion=*/llvm::None,
105 EmitPreInitStmt(S)) {}
106};
107
108/// Lexical scope for OpenMP teams construct, that handles correct codegen
109/// for captured expressions.
110class OMPTeamsScope final : public OMPLexicalScope {
111 bool EmitPreInitStmt(const OMPExecutableDirective &S) {
112 OpenMPDirectiveKind Kind = S.getDirectiveKind();
113 return !isOpenMPTargetExecutionDirective(Kind) &&
114 isOpenMPTeamsDirective(Kind);
115 }
116
117public:
118 OMPTeamsScope(CodeGenFunction &CGF, const OMPExecutableDirective &S)
119 : OMPLexicalScope(CGF, S, /*CapturedRegion=*/llvm::None,
120 EmitPreInitStmt(S)) {}
121};
122
123/// Private scope for OpenMP loop-based directives, that supports capturing
124/// of used expression from loop statement.
125class OMPLoopScope : public CodeGenFunction::RunCleanupsScope {
126 void emitPreInitStmt(CodeGenFunction &CGF, const OMPLoopDirective &S) {
127 CodeGenFunction::OMPMapVars PreCondVars;
128 llvm::DenseSet<const VarDecl *> EmittedAsPrivate;
129 for (const auto *E : S.counters()) {
130 const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl());
131 EmittedAsPrivate.insert(VD->getCanonicalDecl());
132 (void)PreCondVars.setVarAddr(
133 CGF, VD, CGF.CreateMemTemp(VD->getType().getNonReferenceType()));
134 }
135 // Mark private vars as undefs.
136 for (const auto *C : S.getClausesOfKind<OMPPrivateClause>()) {
137 for (const Expr *IRef : C->varlists()) {
138 const auto *OrigVD = cast<VarDecl>(cast<DeclRefExpr>(IRef)->getDecl());
139 if (EmittedAsPrivate.insert(OrigVD->getCanonicalDecl()).second) {
140 (void)PreCondVars.setVarAddr(
141 CGF, OrigVD,
142 Address(llvm::UndefValue::get(
143 CGF.ConvertTypeForMem(CGF.getContext().getPointerType(
144 OrigVD->getType().getNonReferenceType()))),
145 CGF.getContext().getDeclAlign(OrigVD)));
146 }
147 }
148 }
149 (void)PreCondVars.apply(CGF);
150 // Emit init, __range and __end variables for C++ range loops.
151 const Stmt *Body =
152 S.getInnermostCapturedStmt()->getCapturedStmt()->IgnoreContainers();
153 for (unsigned Cnt = 0; Cnt < S.getCollapsedNumber(); ++Cnt) {
154 Body = OMPLoopDirective::tryToFindNextInnerLoop(
155 Body, /*TryImperfectlyNestedLoops=*/true);
156 if (auto *For = dyn_cast<ForStmt>(Body)) {
157 Body = For->getBody();
158 } else {
159 assert(isa<CXXForRangeStmt>(Body) &&((isa<CXXForRangeStmt>(Body) && "Expected canonical for loop or range-based for loop."
) ? static_cast<void> (0) : __assert_fail ("isa<CXXForRangeStmt>(Body) && \"Expected canonical for loop or range-based for loop.\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 160, __PRETTY_FUNCTION__))
160 "Expected canonical for loop or range-based for loop.")((isa<CXXForRangeStmt>(Body) && "Expected canonical for loop or range-based for loop."
) ? static_cast<void> (0) : __assert_fail ("isa<CXXForRangeStmt>(Body) && \"Expected canonical for loop or range-based for loop.\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 160, __PRETTY_FUNCTION__))
;
161 auto *CXXFor = cast<CXXForRangeStmt>(Body);
162 if (const Stmt *Init = CXXFor->getInit())
163 CGF.EmitStmt(Init);
164 CGF.EmitStmt(CXXFor->getRangeStmt());
165 CGF.EmitStmt(CXXFor->getEndStmt());
166 Body = CXXFor->getBody();
167 }
168 }
169 if (const auto *PreInits = cast_or_null<DeclStmt>(S.getPreInits())) {
170 for (const auto *I : PreInits->decls())
171 CGF.EmitVarDecl(cast<VarDecl>(*I));
172 }
173 PreCondVars.restore(CGF);
174 }
175
176public:
177 OMPLoopScope(CodeGenFunction &CGF, const OMPLoopDirective &S)
178 : CodeGenFunction::RunCleanupsScope(CGF) {
179 emitPreInitStmt(CGF, S);
180 }
181};
182
183class OMPSimdLexicalScope : public CodeGenFunction::LexicalScope {
184 CodeGenFunction::OMPPrivateScope InlinedShareds;
185
186 static bool isCapturedVar(CodeGenFunction &CGF, const VarDecl *VD) {
187 return CGF.LambdaCaptureFields.lookup(VD) ||
188 (CGF.CapturedStmtInfo && CGF.CapturedStmtInfo->lookup(VD)) ||
189 (CGF.CurCodeDecl && isa<BlockDecl>(CGF.CurCodeDecl) &&
190 cast<BlockDecl>(CGF.CurCodeDecl)->capturesVariable(VD));
191 }
192
193public:
194 OMPSimdLexicalScope(CodeGenFunction &CGF, const OMPExecutableDirective &S)
195 : CodeGenFunction::LexicalScope(CGF, S.getSourceRange()),
196 InlinedShareds(CGF) {
197 for (const auto *C : S.clauses()) {
198 if (const auto *CPI = OMPClauseWithPreInit::get(C)) {
199 if (const auto *PreInit =
200 cast_or_null<DeclStmt>(CPI->getPreInitStmt())) {
201 for (const auto *I : PreInit->decls()) {
202 if (!I->hasAttr<OMPCaptureNoInitAttr>()) {
203 CGF.EmitVarDecl(cast<VarDecl>(*I));
204 } else {
205 CodeGenFunction::AutoVarEmission Emission =
206 CGF.EmitAutoVarAlloca(cast<VarDecl>(*I));
207 CGF.EmitAutoVarCleanups(Emission);
208 }
209 }
210 }
211 } else if (const auto *UDP = dyn_cast<OMPUseDevicePtrClause>(C)) {
212 for (const Expr *E : UDP->varlists()) {
213 const Decl *D = cast<DeclRefExpr>(E)->getDecl();
214 if (const auto *OED = dyn_cast<OMPCapturedExprDecl>(D))
215 CGF.EmitVarDecl(*OED);
216 }
217 }
218 }
219 if (!isOpenMPSimdDirective(S.getDirectiveKind()))
220 CGF.EmitOMPPrivateClause(S, InlinedShareds);
221 if (const auto *TG = dyn_cast<OMPTaskgroupDirective>(&S)) {
222 if (const Expr *E = TG->getReductionRef())
223 CGF.EmitVarDecl(*cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl()));
224 }
225 const auto *CS = cast_or_null<CapturedStmt>(S.getAssociatedStmt());
226 while (CS) {
227 for (auto &C : CS->captures()) {
228 if (C.capturesVariable() || C.capturesVariableByCopy()) {
229 auto *VD = C.getCapturedVar();
230 assert(VD == VD->getCanonicalDecl() &&((VD == VD->getCanonicalDecl() && "Canonical decl must be captured."
) ? static_cast<void> (0) : __assert_fail ("VD == VD->getCanonicalDecl() && \"Canonical decl must be captured.\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 231, __PRETTY_FUNCTION__))
231 "Canonical decl must be captured.")((VD == VD->getCanonicalDecl() && "Canonical decl must be captured."
) ? static_cast<void> (0) : __assert_fail ("VD == VD->getCanonicalDecl() && \"Canonical decl must be captured.\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 231, __PRETTY_FUNCTION__))
;
232 DeclRefExpr DRE(CGF.getContext(), const_cast<VarDecl *>(VD),
233 isCapturedVar(CGF, VD) ||
234 (CGF.CapturedStmtInfo &&
235 InlinedShareds.isGlobalVarCaptured(VD)),
236 VD->getType().getNonReferenceType(), VK_LValue,
237 C.getLocation());
238 InlinedShareds.addPrivate(VD, [&CGF, &DRE]() -> Address {
239 return CGF.EmitLValue(&DRE).getAddress(CGF);
240 });
241 }
242 }
243 CS = dyn_cast<CapturedStmt>(CS->getCapturedStmt());
244 }
245 (void)InlinedShareds.Privatize();
246 }
247};
248
249} // namespace
250
251static void emitCommonOMPTargetDirective(CodeGenFunction &CGF,
252 const OMPExecutableDirective &S,
253 const RegionCodeGenTy &CodeGen);
254
255LValue CodeGenFunction::EmitOMPSharedLValue(const Expr *E) {
256 if (const auto *OrigDRE = dyn_cast<DeclRefExpr>(E)) {
257 if (const auto *OrigVD = dyn_cast<VarDecl>(OrigDRE->getDecl())) {
258 OrigVD = OrigVD->getCanonicalDecl();
259 bool IsCaptured =
260 LambdaCaptureFields.lookup(OrigVD) ||
261 (CapturedStmtInfo && CapturedStmtInfo->lookup(OrigVD)) ||
262 (CurCodeDecl && isa<BlockDecl>(CurCodeDecl));
263 DeclRefExpr DRE(getContext(), const_cast<VarDecl *>(OrigVD), IsCaptured,
264 OrigDRE->getType(), VK_LValue, OrigDRE->getExprLoc());
265 return EmitLValue(&DRE);
266 }
267 }
268 return EmitLValue(E);
269}
270
271llvm::Value *CodeGenFunction::getTypeSize(QualType Ty) {
272 ASTContext &C = getContext();
273 llvm::Value *Size = nullptr;
274 auto SizeInChars = C.getTypeSizeInChars(Ty);
275 if (SizeInChars.isZero()) {
276 // getTypeSizeInChars() returns 0 for a VLA.
277 while (const VariableArrayType *VAT = C.getAsVariableArrayType(Ty)) {
278 VlaSizePair VlaSize = getVLASize(VAT);
279 Ty = VlaSize.Type;
280 Size = Size ? Builder.CreateNUWMul(Size, VlaSize.NumElts)
281 : VlaSize.NumElts;
282 }
283 SizeInChars = C.getTypeSizeInChars(Ty);
284 if (SizeInChars.isZero())
285 return llvm::ConstantInt::get(SizeTy, /*V=*/0);
286 return Builder.CreateNUWMul(Size, CGM.getSize(SizeInChars));
287 }
288 return CGM.getSize(SizeInChars);
289}
290
291void CodeGenFunction::GenerateOpenMPCapturedVars(
292 const CapturedStmt &S, SmallVectorImpl<llvm::Value *> &CapturedVars) {
293 const RecordDecl *RD = S.getCapturedRecordDecl();
294 auto CurField = RD->field_begin();
295 auto CurCap = S.captures().begin();
296 for (CapturedStmt::const_capture_init_iterator I = S.capture_init_begin(),
297 E = S.capture_init_end();
298 I != E; ++I, ++CurField, ++CurCap) {
299 if (CurField->hasCapturedVLAType()) {
300 const VariableArrayType *VAT = CurField->getCapturedVLAType();
301 llvm::Value *Val = VLASizeMap[VAT->getSizeExpr()];
302 CapturedVars.push_back(Val);
303 } else if (CurCap->capturesThis()) {
304 CapturedVars.push_back(CXXThisValue);
305 } else if (CurCap->capturesVariableByCopy()) {
306 llvm::Value *CV = EmitLoadOfScalar(EmitLValue(*I), CurCap->getLocation());
307
308 // If the field is not a pointer, we need to save the actual value
309 // and load it as a void pointer.
310 if (!CurField->getType()->isAnyPointerType()) {
311 ASTContext &Ctx = getContext();
312 Address DstAddr = CreateMemTemp(
313 Ctx.getUIntPtrType(),
314 Twine(CurCap->getCapturedVar()->getName(), ".casted"));
315 LValue DstLV = MakeAddrLValue(DstAddr, Ctx.getUIntPtrType());
316
317 llvm::Value *SrcAddrVal = EmitScalarConversion(
318 DstAddr.getPointer(), Ctx.getPointerType(Ctx.getUIntPtrType()),
319 Ctx.getPointerType(CurField->getType()), CurCap->getLocation());
320 LValue SrcLV =
321 MakeNaturalAlignAddrLValue(SrcAddrVal, CurField->getType());
322
323 // Store the value using the source type pointer.
324 EmitStoreThroughLValue(RValue::get(CV), SrcLV);
325
326 // Load the value using the destination type pointer.
327 CV = EmitLoadOfScalar(DstLV, CurCap->getLocation());
328 }
329 CapturedVars.push_back(CV);
330 } else {
331 assert(CurCap->capturesVariable() && "Expected capture by reference.")((CurCap->capturesVariable() && "Expected capture by reference."
) ? static_cast<void> (0) : __assert_fail ("CurCap->capturesVariable() && \"Expected capture by reference.\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 331, __PRETTY_FUNCTION__))
;
332 CapturedVars.push_back(EmitLValue(*I).getAddress(*this).getPointer());
333 }
334 }
335}
336
337static Address castValueFromUintptr(CodeGenFunction &CGF, SourceLocation Loc,
338 QualType DstType, StringRef Name,
339 LValue AddrLV) {
340 ASTContext &Ctx = CGF.getContext();
341
342 llvm::Value *CastedPtr = CGF.EmitScalarConversion(
343 AddrLV.getAddress(CGF).getPointer(), Ctx.getUIntPtrType(),
344 Ctx.getPointerType(DstType), Loc);
345 Address TmpAddr =
346 CGF.MakeNaturalAlignAddrLValue(CastedPtr, Ctx.getPointerType(DstType))
347 .getAddress(CGF);
348 return TmpAddr;
349}
350
351static QualType getCanonicalParamType(ASTContext &C, QualType T) {
352 if (T->isLValueReferenceType())
353 return C.getLValueReferenceType(
354 getCanonicalParamType(C, T.getNonReferenceType()),
355 /*SpelledAsLValue=*/false);
356 if (T->isPointerType())
357 return C.getPointerType(getCanonicalParamType(C, T->getPointeeType()));
358 if (const ArrayType *A = T->getAsArrayTypeUnsafe()) {
359 if (const auto *VLA = dyn_cast<VariableArrayType>(A))
360 return getCanonicalParamType(C, VLA->getElementType());
361 if (!A->isVariablyModifiedType())
362 return C.getCanonicalType(T);
363 }
364 return C.getCanonicalParamType(T);
365}
366
367namespace {
368 /// Contains required data for proper outlined function codegen.
369 struct FunctionOptions {
370 /// Captured statement for which the function is generated.
371 const CapturedStmt *S = nullptr;
372 /// true if cast to/from UIntPtr is required for variables captured by
373 /// value.
374 const bool UIntPtrCastRequired = true;
375 /// true if only casted arguments must be registered as local args or VLA
376 /// sizes.
377 const bool RegisterCastedArgsOnly = false;
378 /// Name of the generated function.
379 const StringRef FunctionName;
380 explicit FunctionOptions(const CapturedStmt *S, bool UIntPtrCastRequired,
381 bool RegisterCastedArgsOnly,
382 StringRef FunctionName)
383 : S(S), UIntPtrCastRequired(UIntPtrCastRequired),
384 RegisterCastedArgsOnly(UIntPtrCastRequired && RegisterCastedArgsOnly),
385 FunctionName(FunctionName) {}
386 };
387}
388
389static llvm::Function *emitOutlinedFunctionPrologue(
390 CodeGenFunction &CGF, FunctionArgList &Args,
391 llvm::MapVector<const Decl *, std::pair<const VarDecl *, Address>>
392 &LocalAddrs,
393 llvm::DenseMap<const Decl *, std::pair<const Expr *, llvm::Value *>>
394 &VLASizes,
395 llvm::Value *&CXXThisValue, const FunctionOptions &FO) {
396 const CapturedDecl *CD = FO.S->getCapturedDecl();
397 const RecordDecl *RD = FO.S->getCapturedRecordDecl();
398 assert(CD->hasBody() && "missing CapturedDecl body")((CD->hasBody() && "missing CapturedDecl body") ? static_cast
<void> (0) : __assert_fail ("CD->hasBody() && \"missing CapturedDecl body\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 398, __PRETTY_FUNCTION__))
;
399
400 CXXThisValue = nullptr;
401 // Build the argument list.
402 CodeGenModule &CGM = CGF.CGM;
403 ASTContext &Ctx = CGM.getContext();
404 FunctionArgList TargetArgs;
405 Args.append(CD->param_begin(),
406 std::next(CD->param_begin(), CD->getContextParamPosition()));
407 TargetArgs.append(
408 CD->param_begin(),
409 std::next(CD->param_begin(), CD->getContextParamPosition()));
410 auto I = FO.S->captures().begin();
411 FunctionDecl *DebugFunctionDecl = nullptr;
412 if (!FO.UIntPtrCastRequired) {
413 FunctionProtoType::ExtProtoInfo EPI;
414 QualType FunctionTy = Ctx.getFunctionType(Ctx.VoidTy, llvm::None, EPI);
415 DebugFunctionDecl = FunctionDecl::Create(
416 Ctx, Ctx.getTranslationUnitDecl(), FO.S->getBeginLoc(),
417 SourceLocation(), DeclarationName(), FunctionTy,
418 Ctx.getTrivialTypeSourceInfo(FunctionTy), SC_Static,
419 /*isInlineSpecified=*/false, /*hasWrittenPrototype=*/false);
420 }
421 for (const FieldDecl *FD : RD->fields()) {
422 QualType ArgType = FD->getType();
423 IdentifierInfo *II = nullptr;
424 VarDecl *CapVar = nullptr;
425
426 // If this is a capture by copy and the type is not a pointer, the outlined
427 // function argument type should be uintptr and the value properly casted to
428 // uintptr. This is necessary given that the runtime library is only able to
429 // deal with pointers. We can pass in the same way the VLA type sizes to the
430 // outlined function.
431 if (FO.UIntPtrCastRequired &&
432 ((I->capturesVariableByCopy() && !ArgType->isAnyPointerType()) ||
433 I->capturesVariableArrayType()))
434 ArgType = Ctx.getUIntPtrType();
435
436 if (I->capturesVariable() || I->capturesVariableByCopy()) {
437 CapVar = I->getCapturedVar();
438 II = CapVar->getIdentifier();
439 } else if (I->capturesThis()) {
440 II = &Ctx.Idents.get("this");
441 } else {
442 assert(I->capturesVariableArrayType())((I->capturesVariableArrayType()) ? static_cast<void>
(0) : __assert_fail ("I->capturesVariableArrayType()", "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 442, __PRETTY_FUNCTION__))
;
443 II = &Ctx.Idents.get("vla");
444 }
445 if (ArgType->isVariablyModifiedType())
446 ArgType = getCanonicalParamType(Ctx, ArgType);
447 VarDecl *Arg;
448 if (DebugFunctionDecl && (CapVar || I->capturesThis())) {
449 Arg = ParmVarDecl::Create(
450 Ctx, DebugFunctionDecl,
451 CapVar ? CapVar->getBeginLoc() : FD->getBeginLoc(),
452 CapVar ? CapVar->getLocation() : FD->getLocation(), II, ArgType,
453 /*TInfo=*/nullptr, SC_None, /*DefArg=*/nullptr);
454 } else {
455 Arg = ImplicitParamDecl::Create(Ctx, /*DC=*/nullptr, FD->getLocation(),
456 II, ArgType, ImplicitParamDecl::Other);
457 }
458 Args.emplace_back(Arg);
459 // Do not cast arguments if we emit function with non-original types.
460 TargetArgs.emplace_back(
461 FO.UIntPtrCastRequired
462 ? Arg
463 : CGM.getOpenMPRuntime().translateParameter(FD, Arg));
464 ++I;
465 }
466 Args.append(
467 std::next(CD->param_begin(), CD->getContextParamPosition() + 1),
468 CD->param_end());
469 TargetArgs.append(
470 std::next(CD->param_begin(), CD->getContextParamPosition() + 1),
471 CD->param_end());
472
473 // Create the function declaration.
474 const CGFunctionInfo &FuncInfo =
475 CGM.getTypes().arrangeBuiltinFunctionDeclaration(Ctx.VoidTy, TargetArgs);
476 llvm::FunctionType *FuncLLVMTy = CGM.getTypes().GetFunctionType(FuncInfo);
477
478 auto *F =
479 llvm::Function::Create(FuncLLVMTy, llvm::GlobalValue::InternalLinkage,
480 FO.FunctionName, &CGM.getModule());
481 CGM.SetInternalFunctionAttributes(CD, F, FuncInfo);
482 if (CD->isNothrow())
483 F->setDoesNotThrow();
484 F->setDoesNotRecurse();
485
486 // Generate the function.
487 CGF.StartFunction(CD, Ctx.VoidTy, F, FuncInfo, TargetArgs,
488 FO.S->getBeginLoc(), CD->getBody()->getBeginLoc());
489 unsigned Cnt = CD->getContextParamPosition();
490 I = FO.S->captures().begin();
491 for (const FieldDecl *FD : RD->fields()) {
492 // Do not map arguments if we emit function with non-original types.
493 Address LocalAddr(Address::invalid());
494 if (!FO.UIntPtrCastRequired && Args[Cnt] != TargetArgs[Cnt]) {
495 LocalAddr = CGM.getOpenMPRuntime().getParameterAddress(CGF, Args[Cnt],
496 TargetArgs[Cnt]);
497 } else {
498 LocalAddr = CGF.GetAddrOfLocalVar(Args[Cnt]);
499 }
500 // If we are capturing a pointer by copy we don't need to do anything, just
501 // use the value that we get from the arguments.
502 if (I->capturesVariableByCopy() && FD->getType()->isAnyPointerType()) {
503 const VarDecl *CurVD = I->getCapturedVar();
504 if (!FO.RegisterCastedArgsOnly)
505 LocalAddrs.insert({Args[Cnt], {CurVD, LocalAddr}});
506 ++Cnt;
507 ++I;
508 continue;
509 }
510
511 LValue ArgLVal = CGF.MakeAddrLValue(LocalAddr, Args[Cnt]->getType(),
512 AlignmentSource::Decl);
513 if (FD->hasCapturedVLAType()) {
514 if (FO.UIntPtrCastRequired) {
515 ArgLVal = CGF.MakeAddrLValue(
516 castValueFromUintptr(CGF, I->getLocation(), FD->getType(),
517 Args[Cnt]->getName(), ArgLVal),
518 FD->getType(), AlignmentSource::Decl);
519 }
520 llvm::Value *ExprArg = CGF.EmitLoadOfScalar(ArgLVal, I->getLocation());
521 const VariableArrayType *VAT = FD->getCapturedVLAType();
522 VLASizes.try_emplace(Args[Cnt], VAT->getSizeExpr(), ExprArg);
523 } else if (I->capturesVariable()) {
524 const VarDecl *Var = I->getCapturedVar();
525 QualType VarTy = Var->getType();
526 Address ArgAddr = ArgLVal.getAddress(CGF);
527 if (ArgLVal.getType()->isLValueReferenceType()) {
528 ArgAddr = CGF.EmitLoadOfReference(ArgLVal);
529 } else if (!VarTy->isVariablyModifiedType() || !VarTy->isPointerType()) {
530 assert(ArgLVal.getType()->isPointerType())((ArgLVal.getType()->isPointerType()) ? static_cast<void
> (0) : __assert_fail ("ArgLVal.getType()->isPointerType()"
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 530, __PRETTY_FUNCTION__))
;
531 ArgAddr = CGF.EmitLoadOfPointer(
532 ArgAddr, ArgLVal.getType()->castAs<PointerType>());
533 }
534 if (!FO.RegisterCastedArgsOnly) {
535 LocalAddrs.insert(
536 {Args[Cnt],
537 {Var, Address(ArgAddr.getPointer(), Ctx.getDeclAlign(Var))}});
538 }
539 } else if (I->capturesVariableByCopy()) {
540 assert(!FD->getType()->isAnyPointerType() &&((!FD->getType()->isAnyPointerType() && "Not expecting a captured pointer."
) ? static_cast<void> (0) : __assert_fail ("!FD->getType()->isAnyPointerType() && \"Not expecting a captured pointer.\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 541, __PRETTY_FUNCTION__))
541 "Not expecting a captured pointer.")((!FD->getType()->isAnyPointerType() && "Not expecting a captured pointer."
) ? static_cast<void> (0) : __assert_fail ("!FD->getType()->isAnyPointerType() && \"Not expecting a captured pointer.\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 541, __PRETTY_FUNCTION__))
;
542 const VarDecl *Var = I->getCapturedVar();
543 LocalAddrs.insert({Args[Cnt],
544 {Var, FO.UIntPtrCastRequired
545 ? castValueFromUintptr(
546 CGF, I->getLocation(), FD->getType(),
547 Args[Cnt]->getName(), ArgLVal)
548 : ArgLVal.getAddress(CGF)}});
549 } else {
550 // If 'this' is captured, load it into CXXThisValue.
551 assert(I->capturesThis())((I->capturesThis()) ? static_cast<void> (0) : __assert_fail
("I->capturesThis()", "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 551, __PRETTY_FUNCTION__))
;
552 CXXThisValue = CGF.EmitLoadOfScalar(ArgLVal, I->getLocation());
553 LocalAddrs.insert({Args[Cnt], {nullptr, ArgLVal.getAddress(CGF)}});
554 }
555 ++Cnt;
556 ++I;
557 }
558
559 return F;
560}
561
562llvm::Function *
563CodeGenFunction::GenerateOpenMPCapturedStmtFunction(const CapturedStmt &S) {
564 assert(((CapturedStmtInfo && "CapturedStmtInfo should be set when generating the captured function"
) ? static_cast<void> (0) : __assert_fail ("CapturedStmtInfo && \"CapturedStmtInfo should be set when generating the captured function\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 566, __PRETTY_FUNCTION__))
565 CapturedStmtInfo &&((CapturedStmtInfo && "CapturedStmtInfo should be set when generating the captured function"
) ? static_cast<void> (0) : __assert_fail ("CapturedStmtInfo && \"CapturedStmtInfo should be set when generating the captured function\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 566, __PRETTY_FUNCTION__))
566 "CapturedStmtInfo should be set when generating the captured function")((CapturedStmtInfo && "CapturedStmtInfo should be set when generating the captured function"
) ? static_cast<void> (0) : __assert_fail ("CapturedStmtInfo && \"CapturedStmtInfo should be set when generating the captured function\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 566, __PRETTY_FUNCTION__))
;
567 const CapturedDecl *CD = S.getCapturedDecl();
568 // Build the argument list.
569 bool NeedWrapperFunction =
570 getDebugInfo() &&
571 CGM.getCodeGenOpts().getDebugInfo() >= codegenoptions::LimitedDebugInfo;
572 FunctionArgList Args;
573 llvm::MapVector<const Decl *, std::pair<const VarDecl *, Address>> LocalAddrs;
574 llvm::DenseMap<const Decl *, std::pair<const Expr *, llvm::Value *>> VLASizes;
575 SmallString<256> Buffer;
576 llvm::raw_svector_ostream Out(Buffer);
577 Out << CapturedStmtInfo->getHelperName();
578 if (NeedWrapperFunction)
579 Out << "_debug__";
580 FunctionOptions FO(&S, !NeedWrapperFunction, /*RegisterCastedArgsOnly=*/false,
581 Out.str());
582 llvm::Function *F = emitOutlinedFunctionPrologue(*this, Args, LocalAddrs,
583 VLASizes, CXXThisValue, FO);
584 CodeGenFunction::OMPPrivateScope LocalScope(*this);
585 for (const auto &LocalAddrPair : LocalAddrs) {
586 if (LocalAddrPair.second.first) {
587 LocalScope.addPrivate(LocalAddrPair.second.first, [&LocalAddrPair]() {
588 return LocalAddrPair.second.second;
589 });
590 }
591 }
592 (void)LocalScope.Privatize();
593 for (const auto &VLASizePair : VLASizes)
594 VLASizeMap[VLASizePair.second.first] = VLASizePair.second.second;
595 PGO.assignRegionCounters(GlobalDecl(CD), F);
596 CapturedStmtInfo->EmitBody(*this, CD->getBody());
597 (void)LocalScope.ForceCleanup();
598 FinishFunction(CD->getBodyRBrace());
599 if (!NeedWrapperFunction)
600 return F;
601
602 FunctionOptions WrapperFO(&S, /*UIntPtrCastRequired=*/true,
603 /*RegisterCastedArgsOnly=*/true,
604 CapturedStmtInfo->getHelperName());
605 CodeGenFunction WrapperCGF(CGM, /*suppressNewContext=*/true);
606 WrapperCGF.CapturedStmtInfo = CapturedStmtInfo;
607 Args.clear();
608 LocalAddrs.clear();
609 VLASizes.clear();
610 llvm::Function *WrapperF =
611 emitOutlinedFunctionPrologue(WrapperCGF, Args, LocalAddrs, VLASizes,
612 WrapperCGF.CXXThisValue, WrapperFO);
613 llvm::SmallVector<llvm::Value *, 4> CallArgs;
614 for (const auto *Arg : Args) {
615 llvm::Value *CallArg;
616 auto I = LocalAddrs.find(Arg);
617 if (I != LocalAddrs.end()) {
618 LValue LV = WrapperCGF.MakeAddrLValue(
619 I->second.second,
620 I->second.first ? I->second.first->getType() : Arg->getType(),
621 AlignmentSource::Decl);
622 CallArg = WrapperCGF.EmitLoadOfScalar(LV, S.getBeginLoc());
623 } else {
624 auto EI = VLASizes.find(Arg);
625 if (EI != VLASizes.end()) {
626 CallArg = EI->second.second;
627 } else {
628 LValue LV = WrapperCGF.MakeAddrLValue(WrapperCGF.GetAddrOfLocalVar(Arg),
629 Arg->getType(),
630 AlignmentSource::Decl);
631 CallArg = WrapperCGF.EmitLoadOfScalar(LV, S.getBeginLoc());
632 }
633 }
634 CallArgs.emplace_back(WrapperCGF.EmitFromMemory(CallArg, Arg->getType()));
635 }
636 CGM.getOpenMPRuntime().emitOutlinedFunctionCall(WrapperCGF, S.getBeginLoc(),
637 F, CallArgs);
638 WrapperCGF.FinishFunction();
639 return WrapperF;
640}
641
642//===----------------------------------------------------------------------===//
643// OpenMP Directive Emission
644//===----------------------------------------------------------------------===//
645void CodeGenFunction::EmitOMPAggregateAssign(
646 Address DestAddr, Address SrcAddr, QualType OriginalType,
647 const llvm::function_ref<void(Address, Address)> CopyGen) {
648 // Perform element-by-element initialization.
649 QualType ElementTy;
650
651 // Drill down to the base element type on both arrays.
652 const ArrayType *ArrayTy = OriginalType->getAsArrayTypeUnsafe();
653 llvm::Value *NumElements = emitArrayLength(ArrayTy, ElementTy, DestAddr);
654 SrcAddr = Builder.CreateElementBitCast(SrcAddr, DestAddr.getElementType());
655
656 llvm::Value *SrcBegin = SrcAddr.getPointer();
657 llvm::Value *DestBegin = DestAddr.getPointer();
658 // Cast from pointer to array type to pointer to single element.
659 llvm::Value *DestEnd = Builder.CreateGEP(DestBegin, NumElements);
660 // The basic structure here is a while-do loop.
661 llvm::BasicBlock *BodyBB = createBasicBlock("omp.arraycpy.body");
662 llvm::BasicBlock *DoneBB = createBasicBlock("omp.arraycpy.done");
663 llvm::Value *IsEmpty =
664 Builder.CreateICmpEQ(DestBegin, DestEnd, "omp.arraycpy.isempty");
665 Builder.CreateCondBr(IsEmpty, DoneBB, BodyBB);
666
667 // Enter the loop body, making that address the current address.
668 llvm::BasicBlock *EntryBB = Builder.GetInsertBlock();
669 EmitBlock(BodyBB);
670
671 CharUnits ElementSize = getContext().getTypeSizeInChars(ElementTy);
672
673 llvm::PHINode *SrcElementPHI =
674 Builder.CreatePHI(SrcBegin->getType(), 2, "omp.arraycpy.srcElementPast");
675 SrcElementPHI->addIncoming(SrcBegin, EntryBB);
676 Address SrcElementCurrent =
677 Address(SrcElementPHI,
678 SrcAddr.getAlignment().alignmentOfArrayElement(ElementSize));
679
680 llvm::PHINode *DestElementPHI =
681 Builder.CreatePHI(DestBegin->getType(), 2, "omp.arraycpy.destElementPast");
682 DestElementPHI->addIncoming(DestBegin, EntryBB);
683 Address DestElementCurrent =
684 Address(DestElementPHI,
685 DestAddr.getAlignment().alignmentOfArrayElement(ElementSize));
686
687 // Emit copy.
688 CopyGen(DestElementCurrent, SrcElementCurrent);
689
690 // Shift the address forward by one element.
691 llvm::Value *DestElementNext = Builder.CreateConstGEP1_32(
692 DestElementPHI, /*Idx0=*/1, "omp.arraycpy.dest.element");
693 llvm::Value *SrcElementNext = Builder.CreateConstGEP1_32(
694 SrcElementPHI, /*Idx0=*/1, "omp.arraycpy.src.element");
695 // Check whether we've reached the end.
696 llvm::Value *Done =
697 Builder.CreateICmpEQ(DestElementNext, DestEnd, "omp.arraycpy.done");
698 Builder.CreateCondBr(Done, DoneBB, BodyBB);
699 DestElementPHI->addIncoming(DestElementNext, Builder.GetInsertBlock());
700 SrcElementPHI->addIncoming(SrcElementNext, Builder.GetInsertBlock());
701
702 // Done.
703 EmitBlock(DoneBB, /*IsFinished=*/true);
704}
705
706void CodeGenFunction::EmitOMPCopy(QualType OriginalType, Address DestAddr,
707 Address SrcAddr, const VarDecl *DestVD,
708 const VarDecl *SrcVD, const Expr *Copy) {
709 if (OriginalType->isArrayType()) {
710 const auto *BO = dyn_cast<BinaryOperator>(Copy);
711 if (BO && BO->getOpcode() == BO_Assign) {
712 // Perform simple memcpy for simple copying.
713 LValue Dest = MakeAddrLValue(DestAddr, OriginalType);
714 LValue Src = MakeAddrLValue(SrcAddr, OriginalType);
715 EmitAggregateAssign(Dest, Src, OriginalType);
716 } else {
717 // For arrays with complex element types perform element by element
718 // copying.
719 EmitOMPAggregateAssign(
720 DestAddr, SrcAddr, OriginalType,
721 [this, Copy, SrcVD, DestVD](Address DestElement, Address SrcElement) {
722 // Working with the single array element, so have to remap
723 // destination and source variables to corresponding array
724 // elements.
725 CodeGenFunction::OMPPrivateScope Remap(*this);
726 Remap.addPrivate(DestVD, [DestElement]() { return DestElement; });
727 Remap.addPrivate(SrcVD, [SrcElement]() { return SrcElement; });
728 (void)Remap.Privatize();
729 EmitIgnoredExpr(Copy);
730 });
731 }
732 } else {
733 // Remap pseudo source variable to private copy.
734 CodeGenFunction::OMPPrivateScope Remap(*this);
735 Remap.addPrivate(SrcVD, [SrcAddr]() { return SrcAddr; });
736 Remap.addPrivate(DestVD, [DestAddr]() { return DestAddr; });
737 (void)Remap.Privatize();
738 // Emit copying of the whole variable.
739 EmitIgnoredExpr(Copy);
740 }
741}
742
743bool CodeGenFunction::EmitOMPFirstprivateClause(const OMPExecutableDirective &D,
744 OMPPrivateScope &PrivateScope) {
745 if (!HaveInsertPoint())
746 return false;
747 bool DeviceConstTarget =
748 getLangOpts().OpenMPIsDevice &&
749 isOpenMPTargetExecutionDirective(D.getDirectiveKind());
750 bool FirstprivateIsLastprivate = false;
751 llvm::DenseSet<const VarDecl *> Lastprivates;
752 for (const auto *C : D.getClausesOfKind<OMPLastprivateClause>()) {
753 for (const auto *D : C->varlists())
754 Lastprivates.insert(
755 cast<VarDecl>(cast<DeclRefExpr>(D)->getDecl())->getCanonicalDecl());
756 }
757 llvm::DenseSet<const VarDecl *> EmittedAsFirstprivate;
758 llvm::SmallVector<OpenMPDirectiveKind, 4> CaptureRegions;
759 getOpenMPCaptureRegions(CaptureRegions, D.getDirectiveKind());
760 // Force emission of the firstprivate copy if the directive does not emit
761 // outlined function, like omp for, omp simd, omp distribute etc.
762 bool MustEmitFirstprivateCopy =
763 CaptureRegions.size() == 1 && CaptureRegions.back() == OMPD_unknown;
764 for (const auto *C : D.getClausesOfKind<OMPFirstprivateClause>()) {
765 auto IRef = C->varlist_begin();
766 auto InitsRef = C->inits().begin();
767 for (const Expr *IInit : C->private_copies()) {
768 const auto *OrigVD = cast<VarDecl>(cast<DeclRefExpr>(*IRef)->getDecl());
769 bool ThisFirstprivateIsLastprivate =
770 Lastprivates.count(OrigVD->getCanonicalDecl()) > 0;
771 const FieldDecl *FD = CapturedStmtInfo->lookup(OrigVD);
772 const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(IInit)->getDecl());
773 if (!MustEmitFirstprivateCopy && !ThisFirstprivateIsLastprivate && FD &&
774 !FD->getType()->isReferenceType() &&
775 (!VD || !VD->hasAttr<OMPAllocateDeclAttr>())) {
776 EmittedAsFirstprivate.insert(OrigVD->getCanonicalDecl());
777 ++IRef;
778 ++InitsRef;
779 continue;
780 }
781 // Do not emit copy for firstprivate constant variables in target regions,
782 // captured by reference.
783 if (DeviceConstTarget && OrigVD->getType().isConstant(getContext()) &&
784 FD && FD->getType()->isReferenceType() &&
785 (!VD || !VD->hasAttr<OMPAllocateDeclAttr>())) {
786 (void)CGM.getOpenMPRuntime().registerTargetFirstprivateCopy(*this,
787 OrigVD);
788 ++IRef;
789 ++InitsRef;
790 continue;
791 }
792 FirstprivateIsLastprivate =
793 FirstprivateIsLastprivate || ThisFirstprivateIsLastprivate;
794 if (EmittedAsFirstprivate.insert(OrigVD->getCanonicalDecl()).second) {
795 const auto *VDInit =
796 cast<VarDecl>(cast<DeclRefExpr>(*InitsRef)->getDecl());
797 bool IsRegistered;
798 DeclRefExpr DRE(getContext(), const_cast<VarDecl *>(OrigVD),
799 /*RefersToEnclosingVariableOrCapture=*/FD != nullptr,
800 (*IRef)->getType(), VK_LValue, (*IRef)->getExprLoc());
801 LValue OriginalLVal;
802 if (!FD) {
803 // Check if the firstprivate variable is just a constant value.
804 ConstantEmission CE = tryEmitAsConstant(&DRE);
805 if (CE && !CE.isReference()) {
806 // Constant value, no need to create a copy.
807 ++IRef;
808 ++InitsRef;
809 continue;
810 }
811 if (CE && CE.isReference()) {
812 OriginalLVal = CE.getReferenceLValue(*this, &DRE);
813 } else {
814 assert(!CE && "Expected non-constant firstprivate.")((!CE && "Expected non-constant firstprivate.") ? static_cast
<void> (0) : __assert_fail ("!CE && \"Expected non-constant firstprivate.\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 814, __PRETTY_FUNCTION__))
;
815 OriginalLVal = EmitLValue(&DRE);
816 }
817 } else {
818 OriginalLVal = EmitLValue(&DRE);
819 }
820 QualType Type = VD->getType();
821 if (Type->isArrayType()) {
822 // Emit VarDecl with copy init for arrays.
823 // Get the address of the original variable captured in current
824 // captured region.
825 IsRegistered = PrivateScope.addPrivate(
826 OrigVD, [this, VD, Type, OriginalLVal, VDInit]() {
827 AutoVarEmission Emission = EmitAutoVarAlloca(*VD);
828 const Expr *Init = VD->getInit();
829 if (!isa<CXXConstructExpr>(Init) ||
830 isTrivialInitializer(Init)) {
831 // Perform simple memcpy.
832 LValue Dest =
833 MakeAddrLValue(Emission.getAllocatedAddress(), Type);
834 EmitAggregateAssign(Dest, OriginalLVal, Type);
835 } else {
836 EmitOMPAggregateAssign(
837 Emission.getAllocatedAddress(),
838 OriginalLVal.getAddress(*this), Type,
839 [this, VDInit, Init](Address DestElement,
840 Address SrcElement) {
841 // Clean up any temporaries needed by the
842 // initialization.
843 RunCleanupsScope InitScope(*this);
844 // Emit initialization for single element.
845 setAddrOfLocalVar(VDInit, SrcElement);
846 EmitAnyExprToMem(Init, DestElement,
847 Init->getType().getQualifiers(),
848 /*IsInitializer*/ false);
849 LocalDeclMap.erase(VDInit);
850 });
851 }
852 EmitAutoVarCleanups(Emission);
853 return Emission.getAllocatedAddress();
854 });
855 } else {
856 Address OriginalAddr = OriginalLVal.getAddress(*this);
857 IsRegistered = PrivateScope.addPrivate(
858 OrigVD, [this, VDInit, OriginalAddr, VD]() {
859 // Emit private VarDecl with copy init.
860 // Remap temp VDInit variable to the address of the original
861 // variable (for proper handling of captured global variables).
862 setAddrOfLocalVar(VDInit, OriginalAddr);
863 EmitDecl(*VD);
864 LocalDeclMap.erase(VDInit);
865 return GetAddrOfLocalVar(VD);
866 });
867 }
868 assert(IsRegistered &&((IsRegistered && "firstprivate var already registered as private"
) ? static_cast<void> (0) : __assert_fail ("IsRegistered && \"firstprivate var already registered as private\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 869, __PRETTY_FUNCTION__))
869 "firstprivate var already registered as private")((IsRegistered && "firstprivate var already registered as private"
) ? static_cast<void> (0) : __assert_fail ("IsRegistered && \"firstprivate var already registered as private\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 869, __PRETTY_FUNCTION__))
;
870 // Silence the warning about unused variable.
871 (void)IsRegistered;
872 }
873 ++IRef;
874 ++InitsRef;
875 }
876 }
877 return FirstprivateIsLastprivate && !EmittedAsFirstprivate.empty();
878}
879
880void CodeGenFunction::EmitOMPPrivateClause(
881 const OMPExecutableDirective &D,
882 CodeGenFunction::OMPPrivateScope &PrivateScope) {
883 if (!HaveInsertPoint())
884 return;
885 llvm::DenseSet<const VarDecl *> EmittedAsPrivate;
886 for (const auto *C : D.getClausesOfKind<OMPPrivateClause>()) {
887 auto IRef = C->varlist_begin();
888 for (const Expr *IInit : C->private_copies()) {
889 const auto *OrigVD = cast<VarDecl>(cast<DeclRefExpr>(*IRef)->getDecl());
890 if (EmittedAsPrivate.insert(OrigVD->getCanonicalDecl()).second) {
891 const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(IInit)->getDecl());
892 bool IsRegistered = PrivateScope.addPrivate(OrigVD, [this, VD]() {
893 // Emit private VarDecl with copy init.
894 EmitDecl(*VD);
895 return GetAddrOfLocalVar(VD);
896 });
897 assert(IsRegistered && "private var already registered as private")((IsRegistered && "private var already registered as private"
) ? static_cast<void> (0) : __assert_fail ("IsRegistered && \"private var already registered as private\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 897, __PRETTY_FUNCTION__))
;
898 // Silence the warning about unused variable.
899 (void)IsRegistered;
900 }
901 ++IRef;
902 }
903 }
904}
905
906bool CodeGenFunction::EmitOMPCopyinClause(const OMPExecutableDirective &D) {
907 if (!HaveInsertPoint())
908 return false;
909 // threadprivate_var1 = master_threadprivate_var1;
910 // operator=(threadprivate_var2, master_threadprivate_var2);
911 // ...
912 // __kmpc_barrier(&loc, global_tid);
913 llvm::DenseSet<const VarDecl *> CopiedVars;
914 llvm::BasicBlock *CopyBegin = nullptr, *CopyEnd = nullptr;
915 for (const auto *C : D.getClausesOfKind<OMPCopyinClause>()) {
916 auto IRef = C->varlist_begin();
917 auto ISrcRef = C->source_exprs().begin();
918 auto IDestRef = C->destination_exprs().begin();
919 for (const Expr *AssignOp : C->assignment_ops()) {
920 const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(*IRef)->getDecl());
921 QualType Type = VD->getType();
922 if (CopiedVars.insert(VD->getCanonicalDecl()).second) {
923 // Get the address of the master variable. If we are emitting code with
924 // TLS support, the address is passed from the master as field in the
925 // captured declaration.
926 Address MasterAddr = Address::invalid();
927 if (getLangOpts().OpenMPUseTLS &&
928 getContext().getTargetInfo().isTLSSupported()) {
929 assert(CapturedStmtInfo->lookup(VD) &&((CapturedStmtInfo->lookup(VD) && "Copyin threadprivates should have been captured!"
) ? static_cast<void> (0) : __assert_fail ("CapturedStmtInfo->lookup(VD) && \"Copyin threadprivates should have been captured!\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 930, __PRETTY_FUNCTION__))
930 "Copyin threadprivates should have been captured!")((CapturedStmtInfo->lookup(VD) && "Copyin threadprivates should have been captured!"
) ? static_cast<void> (0) : __assert_fail ("CapturedStmtInfo->lookup(VD) && \"Copyin threadprivates should have been captured!\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 930, __PRETTY_FUNCTION__))
;
931 DeclRefExpr DRE(getContext(), const_cast<VarDecl *>(VD), true,
932 (*IRef)->getType(), VK_LValue, (*IRef)->getExprLoc());
933 MasterAddr = EmitLValue(&DRE).getAddress(*this);
934 LocalDeclMap.erase(VD);
935 } else {
936 MasterAddr =
937 Address(VD->isStaticLocal() ? CGM.getStaticLocalDeclAddress(VD)
938 : CGM.GetAddrOfGlobal(VD),
939 getContext().getDeclAlign(VD));
940 }
941 // Get the address of the threadprivate variable.
942 Address PrivateAddr = EmitLValue(*IRef).getAddress(*this);
943 if (CopiedVars.size() == 1) {
944 // At first check if current thread is a master thread. If it is, no
945 // need to copy data.
946 CopyBegin = createBasicBlock("copyin.not.master");
947 CopyEnd = createBasicBlock("copyin.not.master.end");
948 Builder.CreateCondBr(
949 Builder.CreateICmpNE(
950 Builder.CreatePtrToInt(MasterAddr.getPointer(), CGM.IntPtrTy),
951 Builder.CreatePtrToInt(PrivateAddr.getPointer(),
952 CGM.IntPtrTy)),
953 CopyBegin, CopyEnd);
954 EmitBlock(CopyBegin);
955 }
956 const auto *SrcVD =
957 cast<VarDecl>(cast<DeclRefExpr>(*ISrcRef)->getDecl());
958 const auto *DestVD =
959 cast<VarDecl>(cast<DeclRefExpr>(*IDestRef)->getDecl());
960 EmitOMPCopy(Type, PrivateAddr, MasterAddr, DestVD, SrcVD, AssignOp);
961 }
962 ++IRef;
963 ++ISrcRef;
964 ++IDestRef;
965 }
966 }
967 if (CopyEnd) {
968 // Exit out of copying procedure for non-master thread.
969 EmitBlock(CopyEnd, /*IsFinished=*/true);
970 return true;
971 }
972 return false;
973}
974
975bool CodeGenFunction::EmitOMPLastprivateClauseInit(
976 const OMPExecutableDirective &D, OMPPrivateScope &PrivateScope) {
977 if (!HaveInsertPoint())
978 return false;
979 bool HasAtLeastOneLastprivate = false;
980 llvm::DenseSet<const VarDecl *> SIMDLCVs;
981 if (isOpenMPSimdDirective(D.getDirectiveKind())) {
982 const auto *LoopDirective = cast<OMPLoopDirective>(&D);
983 for (const Expr *C : LoopDirective->counters()) {
984 SIMDLCVs.insert(
985 cast<VarDecl>(cast<DeclRefExpr>(C)->getDecl())->getCanonicalDecl());
986 }
987 }
988 llvm::DenseSet<const VarDecl *> AlreadyEmittedVars;
989 for (const auto *C : D.getClausesOfKind<OMPLastprivateClause>()) {
990 HasAtLeastOneLastprivate = true;
991 if (isOpenMPTaskLoopDirective(D.getDirectiveKind()) &&
992 !getLangOpts().OpenMPSimd)
993 break;
994 auto IRef = C->varlist_begin();
995 auto IDestRef = C->destination_exprs().begin();
996 for (const Expr *IInit : C->private_copies()) {
997 // Keep the address of the original variable for future update at the end
998 // of the loop.
999 const auto *OrigVD = cast<VarDecl>(cast<DeclRefExpr>(*IRef)->getDecl());
1000 // Taskloops do not require additional initialization, it is done in
1001 // runtime support library.
1002 if (AlreadyEmittedVars.insert(OrigVD->getCanonicalDecl()).second) {
1003 const auto *DestVD =
1004 cast<VarDecl>(cast<DeclRefExpr>(*IDestRef)->getDecl());
1005 PrivateScope.addPrivate(DestVD, [this, OrigVD, IRef]() {
1006 DeclRefExpr DRE(getContext(), const_cast<VarDecl *>(OrigVD),
1007 /*RefersToEnclosingVariableOrCapture=*/
1008 CapturedStmtInfo->lookup(OrigVD) != nullptr,
1009 (*IRef)->getType(), VK_LValue, (*IRef)->getExprLoc());
1010 return EmitLValue(&DRE).getAddress(*this);
1011 });
1012 // Check if the variable is also a firstprivate: in this case IInit is
1013 // not generated. Initialization of this variable will happen in codegen
1014 // for 'firstprivate' clause.
1015 if (IInit && !SIMDLCVs.count(OrigVD->getCanonicalDecl())) {
1016 const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(IInit)->getDecl());
1017 bool IsRegistered = PrivateScope.addPrivate(OrigVD, [this, VD]() {
1018 // Emit private VarDecl with copy init.
1019 EmitDecl(*VD);
1020 return GetAddrOfLocalVar(VD);
1021 });
1022 assert(IsRegistered &&((IsRegistered && "lastprivate var already registered as private"
) ? static_cast<void> (0) : __assert_fail ("IsRegistered && \"lastprivate var already registered as private\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 1023, __PRETTY_FUNCTION__))
1023 "lastprivate var already registered as private")((IsRegistered && "lastprivate var already registered as private"
) ? static_cast<void> (0) : __assert_fail ("IsRegistered && \"lastprivate var already registered as private\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 1023, __PRETTY_FUNCTION__))
;
1024 (void)IsRegistered;
1025 }
1026 }
1027 ++IRef;
1028 ++IDestRef;
1029 }
1030 }
1031 return HasAtLeastOneLastprivate;
1032}
1033
1034void CodeGenFunction::EmitOMPLastprivateClauseFinal(
1035 const OMPExecutableDirective &D, bool NoFinals,
1036 llvm::Value *IsLastIterCond) {
1037 if (!HaveInsertPoint())
1038 return;
1039 // Emit following code:
1040 // if (<IsLastIterCond>) {
1041 // orig_var1 = private_orig_var1;
1042 // ...
1043 // orig_varn = private_orig_varn;
1044 // }
1045 llvm::BasicBlock *ThenBB = nullptr;
1046 llvm::BasicBlock *DoneBB = nullptr;
1047 if (IsLastIterCond) {
1048 // Emit implicit barrier if at least one lastprivate conditional is found
1049 // and this is not a simd mode.
1050 if (!getLangOpts().OpenMPSimd &&
1051 llvm::any_of(D.getClausesOfKind<OMPLastprivateClause>(),
1052 [](const OMPLastprivateClause *C) {
1053 return C->getKind() == OMPC_LASTPRIVATE_conditional;
1054 })) {
1055 CGM.getOpenMPRuntime().emitBarrierCall(*this, D.getBeginLoc(),
1056 OMPD_unknown,
1057 /*EmitChecks=*/false,
1058 /*ForceSimpleCall=*/true);
1059 }
1060 ThenBB = createBasicBlock(".omp.lastprivate.then");
1061 DoneBB = createBasicBlock(".omp.lastprivate.done");
1062 Builder.CreateCondBr(IsLastIterCond, ThenBB, DoneBB);
1063 EmitBlock(ThenBB);
1064 }
1065 llvm::DenseSet<const VarDecl *> AlreadyEmittedVars;
1066 llvm::DenseMap<const VarDecl *, const Expr *> LoopCountersAndUpdates;
1067 if (const auto *LoopDirective = dyn_cast<OMPLoopDirective>(&D)) {
1068 auto IC = LoopDirective->counters().begin();
1069 for (const Expr *F : LoopDirective->finals()) {
1070 const auto *D =
1071 cast<VarDecl>(cast<DeclRefExpr>(*IC)->getDecl())->getCanonicalDecl();
1072 if (NoFinals)
1073 AlreadyEmittedVars.insert(D);
1074 else
1075 LoopCountersAndUpdates[D] = F;
1076 ++IC;
1077 }
1078 }
1079 for (const auto *C : D.getClausesOfKind<OMPLastprivateClause>()) {
1080 auto IRef = C->varlist_begin();
1081 auto ISrcRef = C->source_exprs().begin();
1082 auto IDestRef = C->destination_exprs().begin();
1083 for (const Expr *AssignOp : C->assignment_ops()) {
1084 const auto *PrivateVD =
1085 cast<VarDecl>(cast<DeclRefExpr>(*IRef)->getDecl());
1086 QualType Type = PrivateVD->getType();
1087 const auto *CanonicalVD = PrivateVD->getCanonicalDecl();
1088 if (AlreadyEmittedVars.insert(CanonicalVD).second) {
1089 // If lastprivate variable is a loop control variable for loop-based
1090 // directive, update its value before copyin back to original
1091 // variable.
1092 if (const Expr *FinalExpr = LoopCountersAndUpdates.lookup(CanonicalVD))
1093 EmitIgnoredExpr(FinalExpr);
1094 const auto *SrcVD =
1095 cast<VarDecl>(cast<DeclRefExpr>(*ISrcRef)->getDecl());
1096 const auto *DestVD =
1097 cast<VarDecl>(cast<DeclRefExpr>(*IDestRef)->getDecl());
1098 // Get the address of the private variable.
1099 Address PrivateAddr = GetAddrOfLocalVar(PrivateVD);
1100 if (const auto *RefTy = PrivateVD->getType()->getAs<ReferenceType>())
1101 PrivateAddr =
1102 Address(Builder.CreateLoad(PrivateAddr),
1103 getNaturalTypeAlignment(RefTy->getPointeeType()));
1104 // Store the last value to the private copy in the last iteration.
1105 if (C->getKind() == OMPC_LASTPRIVATE_conditional)
1106 CGM.getOpenMPRuntime().emitLastprivateConditionalFinalUpdate(
1107 *this, MakeAddrLValue(PrivateAddr, (*IRef)->getType()), PrivateVD,
1108 (*IRef)->getExprLoc());
1109 // Get the address of the original variable.
1110 Address OriginalAddr = GetAddrOfLocalVar(DestVD);
1111 EmitOMPCopy(Type, OriginalAddr, PrivateAddr, DestVD, SrcVD, AssignOp);
1112 }
1113 ++IRef;
1114 ++ISrcRef;
1115 ++IDestRef;
1116 }
1117 if (const Expr *PostUpdate = C->getPostUpdateExpr())
1118 EmitIgnoredExpr(PostUpdate);
1119 }
1120 if (IsLastIterCond)
1121 EmitBlock(DoneBB, /*IsFinished=*/true);
1122}
1123
1124void CodeGenFunction::EmitOMPReductionClauseInit(
1125 const OMPExecutableDirective &D,
1126 CodeGenFunction::OMPPrivateScope &PrivateScope) {
1127 if (!HaveInsertPoint())
1128 return;
1129 SmallVector<const Expr *, 4> Shareds;
1130 SmallVector<const Expr *, 4> Privates;
1131 SmallVector<const Expr *, 4> ReductionOps;
1132 SmallVector<const Expr *, 4> LHSs;
1133 SmallVector<const Expr *, 4> RHSs;
1134 for (const auto *C : D.getClausesOfKind<OMPReductionClause>()) {
1135 auto IPriv = C->privates().begin();
1136 auto IRed = C->reduction_ops().begin();
1137 auto ILHS = C->lhs_exprs().begin();
1138 auto IRHS = C->rhs_exprs().begin();
1139 for (const Expr *Ref : C->varlists()) {
1140 Shareds.emplace_back(Ref);
1141 Privates.emplace_back(*IPriv);
1142 ReductionOps.emplace_back(*IRed);
1143 LHSs.emplace_back(*ILHS);
1144 RHSs.emplace_back(*IRHS);
1145 std::advance(IPriv, 1);
1146 std::advance(IRed, 1);
1147 std::advance(ILHS, 1);
1148 std::advance(IRHS, 1);
1149 }
1150 }
1151 ReductionCodeGen RedCG(Shareds, Privates, ReductionOps);
1152 unsigned Count = 0;
1153 auto ILHS = LHSs.begin();
1154 auto IRHS = RHSs.begin();
1155 auto IPriv = Privates.begin();
1156 for (const Expr *IRef : Shareds) {
1157 const auto *PrivateVD = cast<VarDecl>(cast<DeclRefExpr>(*IPriv)->getDecl());
1158 // Emit private VarDecl with reduction init.
1159 RedCG.emitSharedLValue(*this, Count);
1160 RedCG.emitAggregateType(*this, Count);
1161 AutoVarEmission Emission = EmitAutoVarAlloca(*PrivateVD);
1162 RedCG.emitInitialization(*this, Count, Emission.getAllocatedAddress(),
1163 RedCG.getSharedLValue(Count),
1164 [&Emission](CodeGenFunction &CGF) {
1165 CGF.EmitAutoVarInit(Emission);
1166 return true;
1167 });
1168 EmitAutoVarCleanups(Emission);
1169 Address BaseAddr = RedCG.adjustPrivateAddress(
1170 *this, Count, Emission.getAllocatedAddress());
1171 bool IsRegistered = PrivateScope.addPrivate(
1172 RedCG.getBaseDecl(Count), [BaseAddr]() { return BaseAddr; });
1173 assert(IsRegistered && "private var already registered as private")((IsRegistered && "private var already registered as private"
) ? static_cast<void> (0) : __assert_fail ("IsRegistered && \"private var already registered as private\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 1173, __PRETTY_FUNCTION__))
;
1174 // Silence the warning about unused variable.
1175 (void)IsRegistered;
1176
1177 const auto *LHSVD = cast<VarDecl>(cast<DeclRefExpr>(*ILHS)->getDecl());
1178 const auto *RHSVD = cast<VarDecl>(cast<DeclRefExpr>(*IRHS)->getDecl());
1179 QualType Type = PrivateVD->getType();
1180 bool isaOMPArraySectionExpr = isa<OMPArraySectionExpr>(IRef);
1181 if (isaOMPArraySectionExpr && Type->isVariablyModifiedType()) {
1182 // Store the address of the original variable associated with the LHS
1183 // implicit variable.
1184 PrivateScope.addPrivate(LHSVD, [&RedCG, Count, this]() {
1185 return RedCG.getSharedLValue(Count).getAddress(*this);
1186 });
1187 PrivateScope.addPrivate(
1188 RHSVD, [this, PrivateVD]() { return GetAddrOfLocalVar(PrivateVD); });
1189 } else if ((isaOMPArraySectionExpr && Type->isScalarType()) ||
1190 isa<ArraySubscriptExpr>(IRef)) {
1191 // Store the address of the original variable associated with the LHS
1192 // implicit variable.
1193 PrivateScope.addPrivate(LHSVD, [&RedCG, Count, this]() {
1194 return RedCG.getSharedLValue(Count).getAddress(*this);
1195 });
1196 PrivateScope.addPrivate(RHSVD, [this, PrivateVD, RHSVD]() {
1197 return Builder.CreateElementBitCast(GetAddrOfLocalVar(PrivateVD),
1198 ConvertTypeForMem(RHSVD->getType()),
1199 "rhs.begin");
1200 });
1201 } else {
1202 QualType Type = PrivateVD->getType();
1203 bool IsArray = getContext().getAsArrayType(Type) != nullptr;
1204 Address OriginalAddr = RedCG.getSharedLValue(Count).getAddress(*this);
1205 // Store the address of the original variable associated with the LHS
1206 // implicit variable.
1207 if (IsArray) {
1208 OriginalAddr = Builder.CreateElementBitCast(
1209 OriginalAddr, ConvertTypeForMem(LHSVD->getType()), "lhs.begin");
1210 }
1211 PrivateScope.addPrivate(LHSVD, [OriginalAddr]() { return OriginalAddr; });
1212 PrivateScope.addPrivate(
1213 RHSVD, [this, PrivateVD, RHSVD, IsArray]() {
1214 return IsArray
1215 ? Builder.CreateElementBitCast(
1216 GetAddrOfLocalVar(PrivateVD),
1217 ConvertTypeForMem(RHSVD->getType()), "rhs.begin")
1218 : GetAddrOfLocalVar(PrivateVD);
1219 });
1220 }
1221 ++ILHS;
1222 ++IRHS;
1223 ++IPriv;
1224 ++Count;
1225 }
1226}
1227
1228void CodeGenFunction::EmitOMPReductionClauseFinal(
1229 const OMPExecutableDirective &D, const OpenMPDirectiveKind ReductionKind) {
1230 if (!HaveInsertPoint())
1231 return;
1232 llvm::SmallVector<const Expr *, 8> Privates;
1233 llvm::SmallVector<const Expr *, 8> LHSExprs;
1234 llvm::SmallVector<const Expr *, 8> RHSExprs;
1235 llvm::SmallVector<const Expr *, 8> ReductionOps;
1236 bool HasAtLeastOneReduction = false;
1237 for (const auto *C : D.getClausesOfKind<OMPReductionClause>()) {
1238 HasAtLeastOneReduction = true;
1239 Privates.append(C->privates().begin(), C->privates().end());
1240 LHSExprs.append(C->lhs_exprs().begin(), C->lhs_exprs().end());
1241 RHSExprs.append(C->rhs_exprs().begin(), C->rhs_exprs().end());
1242 ReductionOps.append(C->reduction_ops().begin(), C->reduction_ops().end());
1243 }
1244 if (HasAtLeastOneReduction) {
1245 bool WithNowait = D.getSingleClause<OMPNowaitClause>() ||
1246 isOpenMPParallelDirective(D.getDirectiveKind()) ||
1247 ReductionKind == OMPD_simd;
1248 bool SimpleReduction = ReductionKind == OMPD_simd;
1249 // Emit nowait reduction if nowait clause is present or directive is a
1250 // parallel directive (it always has implicit barrier).
1251 CGM.getOpenMPRuntime().emitReduction(
1252 *this, D.getEndLoc(), Privates, LHSExprs, RHSExprs, ReductionOps,
1253 {WithNowait, SimpleReduction, ReductionKind});
1254 }
1255}
1256
1257static void emitPostUpdateForReductionClause(
1258 CodeGenFunction &CGF, const OMPExecutableDirective &D,
1259 const llvm::function_ref<llvm::Value *(CodeGenFunction &)> CondGen) {
1260 if (!CGF.HaveInsertPoint())
1261 return;
1262 llvm::BasicBlock *DoneBB = nullptr;
1263 for (const auto *C : D.getClausesOfKind<OMPReductionClause>()) {
1264 if (const Expr *PostUpdate = C->getPostUpdateExpr()) {
1265 if (!DoneBB) {
1266 if (llvm::Value *Cond = CondGen(CGF)) {
1267 // If the first post-update expression is found, emit conditional
1268 // block if it was requested.
1269 llvm::BasicBlock *ThenBB = CGF.createBasicBlock(".omp.reduction.pu");
1270 DoneBB = CGF.createBasicBlock(".omp.reduction.pu.done");
1271 CGF.Builder.CreateCondBr(Cond, ThenBB, DoneBB);
1272 CGF.EmitBlock(ThenBB);
1273 }
1274 }
1275 CGF.EmitIgnoredExpr(PostUpdate);
1276 }
1277 }
1278 if (DoneBB)
1279 CGF.EmitBlock(DoneBB, /*IsFinished=*/true);
1280}
1281
1282namespace {
1283/// Codegen lambda for appending distribute lower and upper bounds to outlined
1284/// parallel function. This is necessary for combined constructs such as
1285/// 'distribute parallel for'
1286typedef llvm::function_ref<void(CodeGenFunction &,
1287 const OMPExecutableDirective &,
1288 llvm::SmallVectorImpl<llvm::Value *> &)>
1289 CodeGenBoundParametersTy;
1290} // anonymous namespace
1291
1292static void emitCommonOMPParallelDirective(
1293 CodeGenFunction &CGF, const OMPExecutableDirective &S,
1294 OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen,
1295 const CodeGenBoundParametersTy &CodeGenBoundParameters) {
1296 const CapturedStmt *CS = S.getCapturedStmt(OMPD_parallel);
1297 llvm::Function *OutlinedFn =
1298 CGF.CGM.getOpenMPRuntime().emitParallelOutlinedFunction(
1299 S, *CS->getCapturedDecl()->param_begin(), InnermostKind, CodeGen);
1300 if (const auto *NumThreadsClause = S.getSingleClause<OMPNumThreadsClause>()) {
1301 CodeGenFunction::RunCleanupsScope NumThreadsScope(CGF);
1302 llvm::Value *NumThreads =
1303 CGF.EmitScalarExpr(NumThreadsClause->getNumThreads(),
1304 /*IgnoreResultAssign=*/true);
1305 CGF.CGM.getOpenMPRuntime().emitNumThreadsClause(
1306 CGF, NumThreads, NumThreadsClause->getBeginLoc());
1307 }
1308 if (const auto *ProcBindClause = S.getSingleClause<OMPProcBindClause>()) {
1309 CodeGenFunction::RunCleanupsScope ProcBindScope(CGF);
1310 CGF.CGM.getOpenMPRuntime().emitProcBindClause(
1311 CGF, ProcBindClause->getProcBindKind(), ProcBindClause->getBeginLoc());
1312 }
1313 const Expr *IfCond = nullptr;
1314 for (const auto *C : S.getClausesOfKind<OMPIfClause>()) {
1315 if (C->getNameModifier() == OMPD_unknown ||
1316 C->getNameModifier() == OMPD_parallel) {
1317 IfCond = C->getCondition();
1318 break;
1319 }
1320 }
1321
1322 OMPParallelScope Scope(CGF, S);
1323 llvm::SmallVector<llvm::Value *, 16> CapturedVars;
1324 // Combining 'distribute' with 'for' requires sharing each 'distribute' chunk
1325 // lower and upper bounds with the pragma 'for' chunking mechanism.
1326 // The following lambda takes care of appending the lower and upper bound
1327 // parameters when necessary
1328 CodeGenBoundParameters(CGF, S, CapturedVars);
1329 CGF.GenerateOpenMPCapturedVars(*CS, CapturedVars);
1330 CGF.CGM.getOpenMPRuntime().emitParallelCall(CGF, S.getBeginLoc(), OutlinedFn,
1331 CapturedVars, IfCond);
1332}
1333
1334static void emitEmptyBoundParameters(CodeGenFunction &,
1335 const OMPExecutableDirective &,
1336 llvm::SmallVectorImpl<llvm::Value *> &) {}
1337
1338void CodeGenFunction::EmitOMPParallelDirective(const OMPParallelDirective &S) {
1339
1340 if (llvm::OpenMPIRBuilder *OMPBuilder = CGM.getOpenMPIRBuilder()) {
1341 // Check if we have any if clause associated with the directive.
1342 llvm::Value *IfCond = nullptr;
1343 if (const auto *C = S.getSingleClause<OMPIfClause>())
1344 IfCond = EmitScalarExpr(C->getCondition(),
1345 /*IgnoreResultAssign=*/true);
1346
1347 llvm::Value *NumThreads = nullptr;
1348 if (const auto *NumThreadsClause = S.getSingleClause<OMPNumThreadsClause>())
1349 NumThreads = EmitScalarExpr(NumThreadsClause->getNumThreads(),
1350 /*IgnoreResultAssign=*/true);
1351
1352 ProcBindKind ProcBind = OMP_PROC_BIND_default;
1353 if (const auto *ProcBindClause = S.getSingleClause<OMPProcBindClause>())
1354 ProcBind = ProcBindClause->getProcBindKind();
1355
1356 using InsertPointTy = llvm::OpenMPIRBuilder::InsertPointTy;
1357
1358 // The cleanup callback that finalizes all variabels at the given location,
1359 // thus calls destructors etc.
1360 auto FiniCB = [this](InsertPointTy IP) {
1361 CGBuilderTy::InsertPointGuard IPG(Builder);
1362 assert(IP.getBlock()->end() != IP.getPoint() &&((IP.getBlock()->end() != IP.getPoint() && "OpenMP IR Builder should cause terminated block!"
) ? static_cast<void> (0) : __assert_fail ("IP.getBlock()->end() != IP.getPoint() && \"OpenMP IR Builder should cause terminated block!\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 1363, __PRETTY_FUNCTION__))
1363 "OpenMP IR Builder should cause terminated block!")((IP.getBlock()->end() != IP.getPoint() && "OpenMP IR Builder should cause terminated block!"
) ? static_cast<void> (0) : __assert_fail ("IP.getBlock()->end() != IP.getPoint() && \"OpenMP IR Builder should cause terminated block!\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 1363, __PRETTY_FUNCTION__))
;
1364 llvm::BasicBlock *IPBB = IP.getBlock();
1365 llvm::BasicBlock *DestBB = IPBB->splitBasicBlock(IP.getPoint());
1366 IPBB->getTerminator()->eraseFromParent();
1367 Builder.SetInsertPoint(IPBB);
1368 CodeGenFunction::JumpDest Dest = getJumpDestInCurrentScope(DestBB);
1369 EmitBranchThroughCleanup(Dest);
1370 };
1371
1372 // Privatization callback that performs appropriate action for
1373 // shared/private/firstprivate/lastprivate/copyin/... variables.
1374 //
1375 // TODO: This defaults to shared right now.
1376 auto PrivCB = [](InsertPointTy AllocaIP, InsertPointTy CodeGenIP,
1377 llvm::Value &Val, llvm::Value *&ReplVal) {
1378 // The next line is appropriate only for variables (Val) with the
1379 // data-sharing attribute "shared".
1380 ReplVal = &Val;
1381
1382 return CodeGenIP;
1383 };
1384
1385 const CapturedStmt *CS = S.getCapturedStmt(OMPD_parallel);
1386 const Stmt *ParallelRegionBodyStmt = CS->getCapturedStmt();
1387
1388 auto BodyGenCB = [ParallelRegionBodyStmt,
1389 this](InsertPointTy AllocaIP, InsertPointTy CodeGenIP,
1390 llvm::BasicBlock &ContinuationBB) {
1391 auto OldAllocaIP = AllocaInsertPt;
1392 AllocaInsertPt = &*AllocaIP.getPoint();
1393
1394 auto OldReturnBlock = ReturnBlock;
1395 ReturnBlock = getJumpDestInCurrentScope(&ContinuationBB);
1396
1397 llvm::BasicBlock *CodeGenIPBB = CodeGenIP.getBlock();
1398 CodeGenIPBB->splitBasicBlock(CodeGenIP.getPoint());
1399 llvm::Instruction *CodeGenIPBBTI = CodeGenIPBB->getTerminator();
1400 CodeGenIPBBTI->removeFromParent();
1401
1402 Builder.SetInsertPoint(CodeGenIPBB);
1403
1404 EmitStmt(ParallelRegionBodyStmt);
1405
1406 Builder.Insert(CodeGenIPBBTI);
1407
1408 AllocaInsertPt = OldAllocaIP;
1409 ReturnBlock = OldReturnBlock;
1410 };
1411
1412 CGCapturedStmtInfo CGSI(*CS, CR_OpenMP);
1413 CodeGenFunction::CGCapturedStmtRAII CapInfoRAII(*this, &CGSI);
1414 Builder.restoreIP(OMPBuilder->CreateParallel(Builder, BodyGenCB, PrivCB,
1415 FiniCB, IfCond, NumThreads,
1416 ProcBind, S.hasCancel()));
1417 return;
1418 }
1419
1420 // Emit parallel region as a standalone region.
1421 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
1422 Action.Enter(CGF);
1423 OMPPrivateScope PrivateScope(CGF);
1424 bool Copyins = CGF.EmitOMPCopyinClause(S);
1425 (void)CGF.EmitOMPFirstprivateClause(S, PrivateScope);
1426 if (Copyins) {
1427 // Emit implicit barrier to synchronize threads and avoid data races on
1428 // propagation master's thread values of threadprivate variables to local
1429 // instances of that variables of all other implicit threads.
1430 CGF.CGM.getOpenMPRuntime().emitBarrierCall(
1431 CGF, S.getBeginLoc(), OMPD_unknown, /*EmitChecks=*/false,
1432 /*ForceSimpleCall=*/true);
1433 }
1434 CGF.EmitOMPPrivateClause(S, PrivateScope);
1435 CGF.EmitOMPReductionClauseInit(S, PrivateScope);
1436 (void)PrivateScope.Privatize();
1437 CGF.EmitStmt(S.getCapturedStmt(OMPD_parallel)->getCapturedStmt());
1438 CGF.EmitOMPReductionClauseFinal(S, /*ReductionKind=*/OMPD_parallel);
1439 };
1440 emitCommonOMPParallelDirective(*this, S, OMPD_parallel, CodeGen,
1441 emitEmptyBoundParameters);
1442 emitPostUpdateForReductionClause(*this, S,
1443 [](CodeGenFunction &) { return nullptr; });
1444}
1445
1446static void emitBody(CodeGenFunction &CGF, const Stmt *S, const Stmt *NextLoop,
1447 int MaxLevel, int Level = 0) {
1448 assert(Level < MaxLevel && "Too deep lookup during loop body codegen.")((Level < MaxLevel && "Too deep lookup during loop body codegen."
) ? static_cast<void> (0) : __assert_fail ("Level < MaxLevel && \"Too deep lookup during loop body codegen.\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 1448, __PRETTY_FUNCTION__))
;
1449 const Stmt *SimplifiedS = S->IgnoreContainers();
1450 if (const auto *CS = dyn_cast<CompoundStmt>(SimplifiedS)) {
1451 PrettyStackTraceLoc CrashInfo(
1452 CGF.getContext().getSourceManager(), CS->getLBracLoc(),
1453 "LLVM IR generation of compound statement ('{}')");
1454
1455 // Keep track of the current cleanup stack depth, including debug scopes.
1456 CodeGenFunction::LexicalScope Scope(CGF, S->getSourceRange());
1457 for (const Stmt *CurStmt : CS->body())
1458 emitBody(CGF, CurStmt, NextLoop, MaxLevel, Level);
1459 return;
1460 }
1461 if (SimplifiedS == NextLoop) {
1462 if (const auto *For = dyn_cast<ForStmt>(SimplifiedS)) {
1463 S = For->getBody();
1464 } else {
1465 assert(isa<CXXForRangeStmt>(SimplifiedS) &&((isa<CXXForRangeStmt>(SimplifiedS) && "Expected canonical for loop or range-based for loop."
) ? static_cast<void> (0) : __assert_fail ("isa<CXXForRangeStmt>(SimplifiedS) && \"Expected canonical for loop or range-based for loop.\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 1466, __PRETTY_FUNCTION__))
1466 "Expected canonical for loop or range-based for loop.")((isa<CXXForRangeStmt>(SimplifiedS) && "Expected canonical for loop or range-based for loop."
) ? static_cast<void> (0) : __assert_fail ("isa<CXXForRangeStmt>(SimplifiedS) && \"Expected canonical for loop or range-based for loop.\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 1466, __PRETTY_FUNCTION__))
;
1467 const auto *CXXFor = cast<CXXForRangeStmt>(SimplifiedS);
1468 CGF.EmitStmt(CXXFor->getLoopVarStmt());
1469 S = CXXFor->getBody();
1470 }
1471 if (Level + 1 < MaxLevel) {
1472 NextLoop = OMPLoopDirective::tryToFindNextInnerLoop(
1473 S, /*TryImperfectlyNestedLoops=*/true);
1474 emitBody(CGF, S, NextLoop, MaxLevel, Level + 1);
1475 return;
1476 }
1477 }
1478 CGF.EmitStmt(S);
1479}
1480
1481void CodeGenFunction::EmitOMPLoopBody(const OMPLoopDirective &D,
1482 JumpDest LoopExit) {
1483 RunCleanupsScope BodyScope(*this);
1484 // Update counters values on current iteration.
1485 for (const Expr *UE : D.updates())
1486 EmitIgnoredExpr(UE);
1487 // Update the linear variables.
1488 // In distribute directives only loop counters may be marked as linear, no
1489 // need to generate the code for them.
1490 if (!isOpenMPDistributeDirective(D.getDirectiveKind())) {
1491 for (const auto *C : D.getClausesOfKind<OMPLinearClause>()) {
1492 for (const Expr *UE : C->updates())
1493 EmitIgnoredExpr(UE);
1494 }
1495 }
1496
1497 // On a continue in the body, jump to the end.
1498 JumpDest Continue = getJumpDestInCurrentScope("omp.body.continue");
1499 BreakContinueStack.push_back(BreakContinue(LoopExit, Continue));
1500 for (const Expr *E : D.finals_conditions()) {
1501 if (!E)
1502 continue;
1503 // Check that loop counter in non-rectangular nest fits into the iteration
1504 // space.
1505 llvm::BasicBlock *NextBB = createBasicBlock("omp.body.next");
1506 EmitBranchOnBoolExpr(E, NextBB, Continue.getBlock(),
1507 getProfileCount(D.getBody()));
1508 EmitBlock(NextBB);
1509 }
1510 // Emit loop variables for C++ range loops.
1511 const Stmt *Body =
1512 D.getInnermostCapturedStmt()->getCapturedStmt()->IgnoreContainers();
1513 // Emit loop body.
1514 emitBody(*this, Body,
1515 OMPLoopDirective::tryToFindNextInnerLoop(
1516 Body, /*TryImperfectlyNestedLoops=*/true),
1517 D.getCollapsedNumber());
1518
1519 // The end (updates/cleanups).
1520 EmitBlock(Continue.getBlock());
1521 BreakContinueStack.pop_back();
1522}
1523
1524void CodeGenFunction::EmitOMPInnerLoop(
1525 const Stmt &S, bool RequiresCleanup, const Expr *LoopCond,
1526 const Expr *IncExpr,
1527 const llvm::function_ref<void(CodeGenFunction &)> BodyGen,
1528 const llvm::function_ref<void(CodeGenFunction &)> PostIncGen) {
1529 auto LoopExit = getJumpDestInCurrentScope("omp.inner.for.end");
1530
1531 // Start the loop with a block that tests the condition.
1532 auto CondBlock = createBasicBlock("omp.inner.for.cond");
1533 EmitBlock(CondBlock);
1534 const SourceRange R = S.getSourceRange();
1535 LoopStack.push(CondBlock, SourceLocToDebugLoc(R.getBegin()),
1536 SourceLocToDebugLoc(R.getEnd()));
1537
1538 // If there are any cleanups between here and the loop-exit scope,
1539 // create a block to stage a loop exit along.
1540 llvm::BasicBlock *ExitBlock = LoopExit.getBlock();
1541 if (RequiresCleanup)
1542 ExitBlock = createBasicBlock("omp.inner.for.cond.cleanup");
1543
1544 llvm::BasicBlock *LoopBody = createBasicBlock("omp.inner.for.body");
1545
1546 // Emit condition.
1547 EmitBranchOnBoolExpr(LoopCond, LoopBody, ExitBlock, getProfileCount(&S));
1548 if (ExitBlock != LoopExit.getBlock()) {
1549 EmitBlock(ExitBlock);
1550 EmitBranchThroughCleanup(LoopExit);
1551 }
1552
1553 EmitBlock(LoopBody);
1554 incrementProfileCounter(&S);
1555
1556 // Create a block for the increment.
1557 JumpDest Continue = getJumpDestInCurrentScope("omp.inner.for.inc");
1558 BreakContinueStack.push_back(BreakContinue(LoopExit, Continue));
1559
1560 BodyGen(*this);
1561
1562 // Emit "IV = IV + 1" and a back-edge to the condition block.
1563 EmitBlock(Continue.getBlock());
1564 EmitIgnoredExpr(IncExpr);
1565 PostIncGen(*this);
1566 BreakContinueStack.pop_back();
1567 EmitBranch(CondBlock);
1568 LoopStack.pop();
1569 // Emit the fall-through block.
1570 EmitBlock(LoopExit.getBlock());
1571}
1572
1573bool CodeGenFunction::EmitOMPLinearClauseInit(const OMPLoopDirective &D) {
1574 if (!HaveInsertPoint())
1575 return false;
1576 // Emit inits for the linear variables.
1577 bool HasLinears = false;
1578 for (const auto *C : D.getClausesOfKind<OMPLinearClause>()) {
1579 for (const Expr *Init : C->inits()) {
1580 HasLinears = true;
1581 const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(Init)->getDecl());
1582 if (const auto *Ref =
1583 dyn_cast<DeclRefExpr>(VD->getInit()->IgnoreImpCasts())) {
1584 AutoVarEmission Emission = EmitAutoVarAlloca(*VD);
1585 const auto *OrigVD = cast<VarDecl>(Ref->getDecl());
1586 DeclRefExpr DRE(getContext(), const_cast<VarDecl *>(OrigVD),
1587 CapturedStmtInfo->lookup(OrigVD) != nullptr,
1588 VD->getInit()->getType(), VK_LValue,
1589 VD->getInit()->getExprLoc());
1590 EmitExprAsInit(&DRE, VD, MakeAddrLValue(Emission.getAllocatedAddress(),
1591 VD->getType()),
1592 /*capturedByInit=*/false);
1593 EmitAutoVarCleanups(Emission);
1594 } else {
1595 EmitVarDecl(*VD);
1596 }
1597 }
1598 // Emit the linear steps for the linear clauses.
1599 // If a step is not constant, it is pre-calculated before the loop.
1600 if (const auto *CS = cast_or_null<BinaryOperator>(C->getCalcStep()))
1601 if (const auto *SaveRef = cast<DeclRefExpr>(CS->getLHS())) {
1602 EmitVarDecl(*cast<VarDecl>(SaveRef->getDecl()));
1603 // Emit calculation of the linear step.
1604 EmitIgnoredExpr(CS);
1605 }
1606 }
1607 return HasLinears;
1608}
1609
1610void CodeGenFunction::EmitOMPLinearClauseFinal(
1611 const OMPLoopDirective &D,
1612 const llvm::function_ref<llvm::Value *(CodeGenFunction &)> CondGen) {
1613 if (!HaveInsertPoint())
1614 return;
1615 llvm::BasicBlock *DoneBB = nullptr;
1616 // Emit the final values of the linear variables.
1617 for (const auto *C : D.getClausesOfKind<OMPLinearClause>()) {
1618 auto IC = C->varlist_begin();
1619 for (const Expr *F : C->finals()) {
1620 if (!DoneBB) {
1621 if (llvm::Value *Cond = CondGen(*this)) {
1622 // If the first post-update expression is found, emit conditional
1623 // block if it was requested.
1624 llvm::BasicBlock *ThenBB = createBasicBlock(".omp.linear.pu");
1625 DoneBB = createBasicBlock(".omp.linear.pu.done");
1626 Builder.CreateCondBr(Cond, ThenBB, DoneBB);
1627 EmitBlock(ThenBB);
1628 }
1629 }
1630 const auto *OrigVD = cast<VarDecl>(cast<DeclRefExpr>(*IC)->getDecl());
1631 DeclRefExpr DRE(getContext(), const_cast<VarDecl *>(OrigVD),
1632 CapturedStmtInfo->lookup(OrigVD) != nullptr,
1633 (*IC)->getType(), VK_LValue, (*IC)->getExprLoc());
1634 Address OrigAddr = EmitLValue(&DRE).getAddress(*this);
1635 CodeGenFunction::OMPPrivateScope VarScope(*this);
1636 VarScope.addPrivate(OrigVD, [OrigAddr]() { return OrigAddr; });
1637 (void)VarScope.Privatize();
1638 EmitIgnoredExpr(F);
1639 ++IC;
1640 }
1641 if (const Expr *PostUpdate = C->getPostUpdateExpr())
1642 EmitIgnoredExpr(PostUpdate);
1643 }
1644 if (DoneBB)
1645 EmitBlock(DoneBB, /*IsFinished=*/true);
1646}
1647
1648static void emitAlignedClause(CodeGenFunction &CGF,
1649 const OMPExecutableDirective &D) {
1650 if (!CGF.HaveInsertPoint())
1651 return;
1652 for (const auto *Clause : D.getClausesOfKind<OMPAlignedClause>()) {
1653 llvm::APInt ClauseAlignment(64, 0);
1654 if (const Expr *AlignmentExpr = Clause->getAlignment()) {
1655 auto *AlignmentCI =
1656 cast<llvm::ConstantInt>(CGF.EmitScalarExpr(AlignmentExpr));
1657 ClauseAlignment = AlignmentCI->getValue();
1658 }
1659 for (const Expr *E : Clause->varlists()) {
1660 llvm::APInt Alignment(ClauseAlignment);
1661 if (Alignment == 0) {
1662 // OpenMP [2.8.1, Description]
1663 // If no optional parameter is specified, implementation-defined default
1664 // alignments for SIMD instructions on the target platforms are assumed.
1665 Alignment =
1666 CGF.getContext()
1667 .toCharUnitsFromBits(CGF.getContext().getOpenMPDefaultSimdAlign(
1668 E->getType()->getPointeeType()))
1669 .getQuantity();
1670 }
1671 assert((Alignment == 0 || Alignment.isPowerOf2()) &&(((Alignment == 0 || Alignment.isPowerOf2()) && "alignment is not power of 2"
) ? static_cast<void> (0) : __assert_fail ("(Alignment == 0 || Alignment.isPowerOf2()) && \"alignment is not power of 2\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 1672, __PRETTY_FUNCTION__))
1672 "alignment is not power of 2")(((Alignment == 0 || Alignment.isPowerOf2()) && "alignment is not power of 2"
) ? static_cast<void> (0) : __assert_fail ("(Alignment == 0 || Alignment.isPowerOf2()) && \"alignment is not power of 2\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 1672, __PRETTY_FUNCTION__))
;
1673 if (Alignment != 0) {
1674 llvm::Value *PtrValue = CGF.EmitScalarExpr(E);
1675 CGF.EmitAlignmentAssumption(
1676 PtrValue, E, /*No second loc needed*/ SourceLocation(),
1677 llvm::ConstantInt::get(CGF.getLLVMContext(), Alignment));
1678 }
1679 }
1680 }
1681}
1682
1683void CodeGenFunction::EmitOMPPrivateLoopCounters(
1684 const OMPLoopDirective &S, CodeGenFunction::OMPPrivateScope &LoopScope) {
1685 if (!HaveInsertPoint())
1686 return;
1687 auto I = S.private_counters().begin();
1688 for (const Expr *E : S.counters()) {
1689 const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl());
1690 const auto *PrivateVD = cast<VarDecl>(cast<DeclRefExpr>(*I)->getDecl());
1691 // Emit var without initialization.
1692 AutoVarEmission VarEmission = EmitAutoVarAlloca(*PrivateVD);
1693 EmitAutoVarCleanups(VarEmission);
1694 LocalDeclMap.erase(PrivateVD);
1695 (void)LoopScope.addPrivate(VD, [&VarEmission]() {
1696 return VarEmission.getAllocatedAddress();
1697 });
1698 if (LocalDeclMap.count(VD) || CapturedStmtInfo->lookup(VD) ||
1699 VD->hasGlobalStorage()) {
1700 (void)LoopScope.addPrivate(PrivateVD, [this, VD, E]() {
1701 DeclRefExpr DRE(getContext(), const_cast<VarDecl *>(VD),
1702 LocalDeclMap.count(VD) || CapturedStmtInfo->lookup(VD),
1703 E->getType(), VK_LValue, E->getExprLoc());
1704 return EmitLValue(&DRE).getAddress(*this);
1705 });
1706 } else {
1707 (void)LoopScope.addPrivate(PrivateVD, [&VarEmission]() {
1708 return VarEmission.getAllocatedAddress();
1709 });
1710 }
1711 ++I;
1712 }
1713 // Privatize extra loop counters used in loops for ordered(n) clauses.
1714 for (const auto *C : S.getClausesOfKind<OMPOrderedClause>()) {
1715 if (!C->getNumForLoops())
1716 continue;
1717 for (unsigned I = S.getCollapsedNumber(),
1718 E = C->getLoopNumIterations().size();
1719 I < E; ++I) {
1720 const auto *DRE = cast<DeclRefExpr>(C->getLoopCounter(I));
1721 const auto *VD = cast<VarDecl>(DRE->getDecl());
1722 // Override only those variables that can be captured to avoid re-emission
1723 // of the variables declared within the loops.
1724 if (DRE->refersToEnclosingVariableOrCapture()) {
1725 (void)LoopScope.addPrivate(VD, [this, DRE, VD]() {
1726 return CreateMemTemp(DRE->getType(), VD->getName());
1727 });
1728 }
1729 }
1730 }
1731}
1732
1733static void emitPreCond(CodeGenFunction &CGF, const OMPLoopDirective &S,
1734 const Expr *Cond, llvm::BasicBlock *TrueBlock,
1735 llvm::BasicBlock *FalseBlock, uint64_t TrueCount) {
1736 if (!CGF.HaveInsertPoint())
1737 return;
1738 {
1739 CodeGenFunction::OMPPrivateScope PreCondScope(CGF);
1740 CGF.EmitOMPPrivateLoopCounters(S, PreCondScope);
1741 (void)PreCondScope.Privatize();
1742 // Get initial values of real counters.
1743 for (const Expr *I : S.inits()) {
1744 CGF.EmitIgnoredExpr(I);
1745 }
1746 }
1747 // Create temp loop control variables with their init values to support
1748 // non-rectangular loops.
1749 CodeGenFunction::OMPMapVars PreCondVars;
1750 for (const Expr * E: S.dependent_counters()) {
1751 if (!E)
1752 continue;
1753 assert(!E->getType().getNonReferenceType()->isRecordType() &&((!E->getType().getNonReferenceType()->isRecordType() &&
"dependent counter must not be an iterator.") ? static_cast<
void> (0) : __assert_fail ("!E->getType().getNonReferenceType()->isRecordType() && \"dependent counter must not be an iterator.\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 1754, __PRETTY_FUNCTION__))
1754 "dependent counter must not be an iterator.")((!E->getType().getNonReferenceType()->isRecordType() &&
"dependent counter must not be an iterator.") ? static_cast<
void> (0) : __assert_fail ("!E->getType().getNonReferenceType()->isRecordType() && \"dependent counter must not be an iterator.\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 1754, __PRETTY_FUNCTION__))
;
1755 const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl());
1756 Address CounterAddr =
1757 CGF.CreateMemTemp(VD->getType().getNonReferenceType());
1758 (void)PreCondVars.setVarAddr(CGF, VD, CounterAddr);
1759 }
1760 (void)PreCondVars.apply(CGF);
1761 for (const Expr *E : S.dependent_inits()) {
1762 if (!E)
1763 continue;
1764 CGF.EmitIgnoredExpr(E);
1765 }
1766 // Check that loop is executed at least one time.
1767 CGF.EmitBranchOnBoolExpr(Cond, TrueBlock, FalseBlock, TrueCount);
1768 PreCondVars.restore(CGF);
1769}
1770
1771void CodeGenFunction::EmitOMPLinearClause(
1772 const OMPLoopDirective &D, CodeGenFunction::OMPPrivateScope &PrivateScope) {
1773 if (!HaveInsertPoint())
1774 return;
1775 llvm::DenseSet<const VarDecl *> SIMDLCVs;
1776 if (isOpenMPSimdDirective(D.getDirectiveKind())) {
1777 const auto *LoopDirective = cast<OMPLoopDirective>(&D);
1778 for (const Expr *C : LoopDirective->counters()) {
1779 SIMDLCVs.insert(
1780 cast<VarDecl>(cast<DeclRefExpr>(C)->getDecl())->getCanonicalDecl());
1781 }
1782 }
1783 for (const auto *C : D.getClausesOfKind<OMPLinearClause>()) {
1784 auto CurPrivate = C->privates().begin();
1785 for (const Expr *E : C->varlists()) {
1786 const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl());
1787 const auto *PrivateVD =
1788 cast<VarDecl>(cast<DeclRefExpr>(*CurPrivate)->getDecl());
1789 if (!SIMDLCVs.count(VD->getCanonicalDecl())) {
1790 bool IsRegistered = PrivateScope.addPrivate(VD, [this, PrivateVD]() {
1791 // Emit private VarDecl with copy init.
1792 EmitVarDecl(*PrivateVD);
1793 return GetAddrOfLocalVar(PrivateVD);
1794 });
1795 assert(IsRegistered && "linear var already registered as private")((IsRegistered && "linear var already registered as private"
) ? static_cast<void> (0) : __assert_fail ("IsRegistered && \"linear var already registered as private\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 1795, __PRETTY_FUNCTION__))
;
1796 // Silence the warning about unused variable.
1797 (void)IsRegistered;
1798 } else {
1799 EmitVarDecl(*PrivateVD);
1800 }
1801 ++CurPrivate;
1802 }
1803 }
1804}
1805
1806static void emitSimdlenSafelenClause(CodeGenFunction &CGF,
1807 const OMPExecutableDirective &D,
1808 bool IsMonotonic) {
1809 if (!CGF.HaveInsertPoint())
1810 return;
1811 if (const auto *C = D.getSingleClause<OMPSimdlenClause>()) {
1812 RValue Len = CGF.EmitAnyExpr(C->getSimdlen(), AggValueSlot::ignored(),
1813 /*ignoreResult=*/true);
1814 auto *Val = cast<llvm::ConstantInt>(Len.getScalarVal());
1815 CGF.LoopStack.setVectorizeWidth(Val->getZExtValue());
1816 // In presence of finite 'safelen', it may be unsafe to mark all
1817 // the memory instructions parallel, because loop-carried
1818 // dependences of 'safelen' iterations are possible.
1819 if (!IsMonotonic)
1820 CGF.LoopStack.setParallel(!D.getSingleClause<OMPSafelenClause>());
1821 } else if (const auto *C = D.getSingleClause<OMPSafelenClause>()) {
1822 RValue Len = CGF.EmitAnyExpr(C->getSafelen(), AggValueSlot::ignored(),
1823 /*ignoreResult=*/true);
1824 auto *Val = cast<llvm::ConstantInt>(Len.getScalarVal());
1825 CGF.LoopStack.setVectorizeWidth(Val->getZExtValue());
1826 // In presence of finite 'safelen', it may be unsafe to mark all
1827 // the memory instructions parallel, because loop-carried
1828 // dependences of 'safelen' iterations are possible.
1829 CGF.LoopStack.setParallel(/*Enable=*/false);
1830 }
1831}
1832
1833void CodeGenFunction::EmitOMPSimdInit(const OMPLoopDirective &D,
1834 bool IsMonotonic) {
1835 // Walk clauses and process safelen/lastprivate.
1836 LoopStack.setParallel(!IsMonotonic);
1837 LoopStack.setVectorizeEnable();
1838 emitSimdlenSafelenClause(*this, D, IsMonotonic);
1839}
1840
1841void CodeGenFunction::EmitOMPSimdFinal(
1842 const OMPLoopDirective &D,
1843 const llvm::function_ref<llvm::Value *(CodeGenFunction &)> CondGen) {
1844 if (!HaveInsertPoint())
1845 return;
1846 llvm::BasicBlock *DoneBB = nullptr;
1847 auto IC = D.counters().begin();
1848 auto IPC = D.private_counters().begin();
1849 for (const Expr *F : D.finals()) {
1850 const auto *OrigVD = cast<VarDecl>(cast<DeclRefExpr>((*IC))->getDecl());
1851 const auto *PrivateVD = cast<VarDecl>(cast<DeclRefExpr>((*IPC))->getDecl());
1852 const auto *CED = dyn_cast<OMPCapturedExprDecl>(OrigVD);
1853 if (LocalDeclMap.count(OrigVD) || CapturedStmtInfo->lookup(OrigVD) ||
1854 OrigVD->hasGlobalStorage() || CED) {
1855 if (!DoneBB) {
1856 if (llvm::Value *Cond = CondGen(*this)) {
1857 // If the first post-update expression is found, emit conditional
1858 // block if it was requested.
1859 llvm::BasicBlock *ThenBB = createBasicBlock(".omp.final.then");
1860 DoneBB = createBasicBlock(".omp.final.done");
1861 Builder.CreateCondBr(Cond, ThenBB, DoneBB);
1862 EmitBlock(ThenBB);
1863 }
1864 }
1865 Address OrigAddr = Address::invalid();
1866 if (CED) {
1867 OrigAddr =
1868 EmitLValue(CED->getInit()->IgnoreImpCasts()).getAddress(*this);
1869 } else {
1870 DeclRefExpr DRE(getContext(), const_cast<VarDecl *>(PrivateVD),
1871 /*RefersToEnclosingVariableOrCapture=*/false,
1872 (*IPC)->getType(), VK_LValue, (*IPC)->getExprLoc());
1873 OrigAddr = EmitLValue(&DRE).getAddress(*this);
1874 }
1875 OMPPrivateScope VarScope(*this);
1876 VarScope.addPrivate(OrigVD, [OrigAddr]() { return OrigAddr; });
1877 (void)VarScope.Privatize();
1878 EmitIgnoredExpr(F);
1879 }
1880 ++IC;
1881 ++IPC;
1882 }
1883 if (DoneBB)
1884 EmitBlock(DoneBB, /*IsFinished=*/true);
1885}
1886
1887static void emitOMPLoopBodyWithStopPoint(CodeGenFunction &CGF,
1888 const OMPLoopDirective &S,
1889 CodeGenFunction::JumpDest LoopExit) {
1890 CGF.CGM.getOpenMPRuntime().initLastprivateConditionalCounter(CGF, S);
1891 CGF.EmitOMPLoopBody(S, LoopExit);
1892 CGF.EmitStopPoint(&S);
1893}
1894
1895/// Emit a helper variable and return corresponding lvalue.
1896static LValue EmitOMPHelperVar(CodeGenFunction &CGF,
1897 const DeclRefExpr *Helper) {
1898 auto VDecl = cast<VarDecl>(Helper->getDecl());
1899 CGF.EmitVarDecl(*VDecl);
1900 return CGF.EmitLValue(Helper);
1901}
1902
1903static void emitCommonSimdLoop(CodeGenFunction &CGF, const OMPLoopDirective &S,
1904 const RegionCodeGenTy &SimdInitGen,
1905 const RegionCodeGenTy &BodyCodeGen) {
1906 auto &&ThenGen = [&S, &SimdInitGen, &BodyCodeGen](CodeGenFunction &CGF,
1907 PrePostActionTy &) {
1908 CGOpenMPRuntime::NontemporalDeclsRAII NontemporalsRegion(CGF.CGM, S);
1909 CodeGenFunction::OMPLocalDeclMapRAII Scope(CGF);
1910 SimdInitGen(CGF);
1911
1912 BodyCodeGen(CGF);
1913 };
1914 auto &&ElseGen = [&BodyCodeGen](CodeGenFunction &CGF, PrePostActionTy &) {
1915 CodeGenFunction::OMPLocalDeclMapRAII Scope(CGF);
1916 CGF.LoopStack.setVectorizeEnable(/*Enable=*/false);
1917
1918 BodyCodeGen(CGF);
1919 };
1920 const Expr *IfCond = nullptr;
1921 for (const auto *C : S.getClausesOfKind<OMPIfClause>()) {
1922 if (CGF.getLangOpts().OpenMP >= 50 &&
1923 (C->getNameModifier() == OMPD_unknown ||
1924 C->getNameModifier() == OMPD_simd)) {
1925 IfCond = C->getCondition();
1926 break;
1927 }
1928 }
1929 if (IfCond) {
1930 CGF.CGM.getOpenMPRuntime().emitIfClause(CGF, IfCond, ThenGen, ElseGen);
1931 } else {
1932 RegionCodeGenTy ThenRCG(ThenGen);
1933 ThenRCG(CGF);
1934 }
1935}
1936
1937static void emitOMPSimdRegion(CodeGenFunction &CGF, const OMPLoopDirective &S,
1938 PrePostActionTy &Action) {
1939 Action.Enter(CGF);
1940 assert(isOpenMPSimdDirective(S.getDirectiveKind()) &&((isOpenMPSimdDirective(S.getDirectiveKind()) && "Expected simd directive"
) ? static_cast<void> (0) : __assert_fail ("isOpenMPSimdDirective(S.getDirectiveKind()) && \"Expected simd directive\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 1941, __PRETTY_FUNCTION__))
1941 "Expected simd directive")((isOpenMPSimdDirective(S.getDirectiveKind()) && "Expected simd directive"
) ? static_cast<void> (0) : __assert_fail ("isOpenMPSimdDirective(S.getDirectiveKind()) && \"Expected simd directive\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 1941, __PRETTY_FUNCTION__))
;
1942 OMPLoopScope PreInitScope(CGF, S);
1943 // if (PreCond) {
1944 // for (IV in 0..LastIteration) BODY;
1945 // <Final counter/linear vars updates>;
1946 // }
1947 //
1948 if (isOpenMPDistributeDirective(S.getDirectiveKind()) ||
1949 isOpenMPWorksharingDirective(S.getDirectiveKind()) ||
1950 isOpenMPTaskLoopDirective(S.getDirectiveKind())) {
1951 (void)EmitOMPHelperVar(CGF, cast<DeclRefExpr>(S.getLowerBoundVariable()));
1952 (void)EmitOMPHelperVar(CGF, cast<DeclRefExpr>(S.getUpperBoundVariable()));
1953 }
1954
1955 // Emit: if (PreCond) - begin.
1956 // If the condition constant folds and can be elided, avoid emitting the
1957 // whole loop.
1958 bool CondConstant;
1959 llvm::BasicBlock *ContBlock = nullptr;
1960 if (CGF.ConstantFoldsToSimpleInteger(S.getPreCond(), CondConstant)) {
1961 if (!CondConstant)
1962 return;
1963 } else {
1964 llvm::BasicBlock *ThenBlock = CGF.createBasicBlock("simd.if.then");
1965 ContBlock = CGF.createBasicBlock("simd.if.end");
1966 emitPreCond(CGF, S, S.getPreCond(), ThenBlock, ContBlock,
1967 CGF.getProfileCount(&S));
1968 CGF.EmitBlock(ThenBlock);
1969 CGF.incrementProfileCounter(&S);
1970 }
1971
1972 // Emit the loop iteration variable.
1973 const Expr *IVExpr = S.getIterationVariable();
1974 const auto *IVDecl = cast<VarDecl>(cast<DeclRefExpr>(IVExpr)->getDecl());
1975 CGF.EmitVarDecl(*IVDecl);
1976 CGF.EmitIgnoredExpr(S.getInit());
1977
1978 // Emit the iterations count variable.
1979 // If it is not a variable, Sema decided to calculate iterations count on
1980 // each iteration (e.g., it is foldable into a constant).
1981 if (const auto *LIExpr = dyn_cast<DeclRefExpr>(S.getLastIteration())) {
1982 CGF.EmitVarDecl(*cast<VarDecl>(LIExpr->getDecl()));
1983 // Emit calculation of the iterations count.
1984 CGF.EmitIgnoredExpr(S.getCalcLastIteration());
1985 }
1986
1987 emitAlignedClause(CGF, S);
1988 (void)CGF.EmitOMPLinearClauseInit(S);
1989 {
1990 CodeGenFunction::OMPPrivateScope LoopScope(CGF);
1991 CGF.EmitOMPPrivateLoopCounters(S, LoopScope);
1992 CGF.EmitOMPLinearClause(S, LoopScope);
1993 CGF.EmitOMPPrivateClause(S, LoopScope);
1994 CGF.EmitOMPReductionClauseInit(S, LoopScope);
1995 CGOpenMPRuntime::LastprivateConditionalRAII LPCRegion(
1996 CGF, S, CGF.EmitLValue(S.getIterationVariable()));
1997 bool HasLastprivateClause = CGF.EmitOMPLastprivateClauseInit(S, LoopScope);
1998 (void)LoopScope.Privatize();
1999 if (isOpenMPTargetExecutionDirective(S.getDirectiveKind()))
2000 CGF.CGM.getOpenMPRuntime().adjustTargetSpecificDataForLambdas(CGF, S);
2001
2002 emitCommonSimdLoop(
2003 CGF, S,
2004 [&S](CodeGenFunction &CGF, PrePostActionTy &) {
2005 CGF.EmitOMPSimdInit(S);
2006 },
2007 [&S, &LoopScope](CodeGenFunction &CGF, PrePostActionTy &) {
2008 CGF.EmitOMPInnerLoop(
2009 S, LoopScope.requiresCleanups(), S.getCond(), S.getInc(),
2010 [&S](CodeGenFunction &CGF) {
2011 CGF.CGM.getOpenMPRuntime().initLastprivateConditionalCounter(
2012 CGF, S);
2013 CGF.EmitOMPLoopBody(S, CodeGenFunction::JumpDest());
2014 CGF.EmitStopPoint(&S);
2015 },
2016 [](CodeGenFunction &) {});
2017 });
2018 CGF.EmitOMPSimdFinal(S, [](CodeGenFunction &) { return nullptr; });
2019 // Emit final copy of the lastprivate variables at the end of loops.
2020 if (HasLastprivateClause)
2021 CGF.EmitOMPLastprivateClauseFinal(S, /*NoFinals=*/true);
2022 CGF.EmitOMPReductionClauseFinal(S, /*ReductionKind=*/OMPD_simd);
2023 emitPostUpdateForReductionClause(CGF, S,
2024 [](CodeGenFunction &) { return nullptr; });
2025 }
2026 CGF.EmitOMPLinearClauseFinal(S, [](CodeGenFunction &) { return nullptr; });
2027 // Emit: if (PreCond) - end.
2028 if (ContBlock) {
2029 CGF.EmitBranch(ContBlock);
2030 CGF.EmitBlock(ContBlock, true);
2031 }
2032}
2033
2034void CodeGenFunction::EmitOMPSimdDirective(const OMPSimdDirective &S) {
2035 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
2036 emitOMPSimdRegion(CGF, S, Action);
2037 };
2038 OMPLexicalScope Scope(*this, S, OMPD_unknown);
2039 CGM.getOpenMPRuntime().emitInlinedDirective(*this, OMPD_simd, CodeGen);
2040}
2041
2042void CodeGenFunction::EmitOMPOuterLoop(
2043 bool DynamicOrOrdered, bool IsMonotonic, const OMPLoopDirective &S,
2044 CodeGenFunction::OMPPrivateScope &LoopScope,
2045 const CodeGenFunction::OMPLoopArguments &LoopArgs,
2046 const CodeGenFunction::CodeGenLoopTy &CodeGenLoop,
2047 const CodeGenFunction::CodeGenOrderedTy &CodeGenOrdered) {
2048 CGOpenMPRuntime &RT = CGM.getOpenMPRuntime();
2049
2050 const Expr *IVExpr = S.getIterationVariable();
2051 const unsigned IVSize = getContext().getTypeSize(IVExpr->getType());
2052 const bool IVSigned = IVExpr->getType()->hasSignedIntegerRepresentation();
2053
2054 JumpDest LoopExit = getJumpDestInCurrentScope("omp.dispatch.end");
2055
2056 // Start the loop with a block that tests the condition.
2057 llvm::BasicBlock *CondBlock = createBasicBlock("omp.dispatch.cond");
2058 EmitBlock(CondBlock);
2059 const SourceRange R = S.getSourceRange();
2060 LoopStack.push(CondBlock, SourceLocToDebugLoc(R.getBegin()),
2061 SourceLocToDebugLoc(R.getEnd()));
2062
2063 llvm::Value *BoolCondVal = nullptr;
2064 if (!DynamicOrOrdered) {
2065 // UB = min(UB, GlobalUB) or
2066 // UB = min(UB, PrevUB) for combined loop sharing constructs (e.g.
2067 // 'distribute parallel for')
2068 EmitIgnoredExpr(LoopArgs.EUB);
2069 // IV = LB
2070 EmitIgnoredExpr(LoopArgs.Init);
2071 // IV < UB
2072 BoolCondVal = EvaluateExprAsBool(LoopArgs.Cond);
2073 } else {
2074 BoolCondVal =
2075 RT.emitForNext(*this, S.getBeginLoc(), IVSize, IVSigned, LoopArgs.IL,
2076 LoopArgs.LB, LoopArgs.UB, LoopArgs.ST);
2077 }
2078
2079 // If there are any cleanups between here and the loop-exit scope,
2080 // create a block to stage a loop exit along.
2081 llvm::BasicBlock *ExitBlock = LoopExit.getBlock();
2082 if (LoopScope.requiresCleanups())
2083 ExitBlock = createBasicBlock("omp.dispatch.cleanup");
2084
2085 llvm::BasicBlock *LoopBody = createBasicBlock("omp.dispatch.body");
2086 Builder.CreateCondBr(BoolCondVal, LoopBody, ExitBlock);
2087 if (ExitBlock != LoopExit.getBlock()) {
2088 EmitBlock(ExitBlock);
2089 EmitBranchThroughCleanup(LoopExit);
2090 }
2091 EmitBlock(LoopBody);
2092
2093 // Emit "IV = LB" (in case of static schedule, we have already calculated new
2094 // LB for loop condition and emitted it above).
2095 if (DynamicOrOrdered)
2096 EmitIgnoredExpr(LoopArgs.Init);
2097
2098 // Create a block for the increment.
2099 JumpDest Continue = getJumpDestInCurrentScope("omp.dispatch.inc");
2100 BreakContinueStack.push_back(BreakContinue(LoopExit, Continue));
2101
2102 emitCommonSimdLoop(
2103 *this, S,
2104 [&S, IsMonotonic](CodeGenFunction &CGF, PrePostActionTy &) {
2105 // Generate !llvm.loop.parallel metadata for loads and stores for loops
2106 // with dynamic/guided scheduling and without ordered clause.
2107 if (!isOpenMPSimdDirective(S.getDirectiveKind()))
2108 CGF.LoopStack.setParallel(!IsMonotonic);
2109 else
2110 CGF.EmitOMPSimdInit(S, IsMonotonic);
2111 },
2112 [&S, &LoopArgs, LoopExit, &CodeGenLoop, IVSize, IVSigned, &CodeGenOrdered,
2113 &LoopScope](CodeGenFunction &CGF, PrePostActionTy &) {
2114 SourceLocation Loc = S.getBeginLoc();
2115 // when 'distribute' is not combined with a 'for':
2116 // while (idx <= UB) { BODY; ++idx; }
2117 // when 'distribute' is combined with a 'for'
2118 // (e.g. 'distribute parallel for')
2119 // while (idx <= UB) { <CodeGen rest of pragma>; idx += ST; }
2120 CGF.EmitOMPInnerLoop(
2121 S, LoopScope.requiresCleanups(), LoopArgs.Cond, LoopArgs.IncExpr,
2122 [&S, LoopExit, &CodeGenLoop](CodeGenFunction &CGF) {
2123 CodeGenLoop(CGF, S, LoopExit);
2124 },
2125 [IVSize, IVSigned, Loc, &CodeGenOrdered](CodeGenFunction &CGF) {
2126 CodeGenOrdered(CGF, Loc, IVSize, IVSigned);
2127 });
2128 });
2129
2130 EmitBlock(Continue.getBlock());
2131 BreakContinueStack.pop_back();
2132 if (!DynamicOrOrdered) {
2133 // Emit "LB = LB + Stride", "UB = UB + Stride".
2134 EmitIgnoredExpr(LoopArgs.NextLB);
2135 EmitIgnoredExpr(LoopArgs.NextUB);
2136 }
2137
2138 EmitBranch(CondBlock);
2139 LoopStack.pop();
2140 // Emit the fall-through block.
2141 EmitBlock(LoopExit.getBlock());
2142
2143 // Tell the runtime we are done.
2144 auto &&CodeGen = [DynamicOrOrdered, &S](CodeGenFunction &CGF) {
2145 if (!DynamicOrOrdered)
2146 CGF.CGM.getOpenMPRuntime().emitForStaticFinish(CGF, S.getEndLoc(),
2147 S.getDirectiveKind());
2148 };
2149 OMPCancelStack.emitExit(*this, S.getDirectiveKind(), CodeGen);
2150}
2151
2152void CodeGenFunction::EmitOMPForOuterLoop(
2153 const OpenMPScheduleTy &ScheduleKind, bool IsMonotonic,
2154 const OMPLoopDirective &S, OMPPrivateScope &LoopScope, bool Ordered,
2155 const OMPLoopArguments &LoopArgs,
2156 const CodeGenDispatchBoundsTy &CGDispatchBounds) {
2157 CGOpenMPRuntime &RT = CGM.getOpenMPRuntime();
2158
2159 // Dynamic scheduling of the outer loop (dynamic, guided, auto, runtime).
2160 const bool DynamicOrOrdered =
2161 Ordered || RT.isDynamic(ScheduleKind.Schedule);
2162
2163 assert((Ordered ||(((Ordered || !RT.isStaticNonchunked(ScheduleKind.Schedule, LoopArgs
.Chunk != nullptr)) && "static non-chunked schedule does not need outer loop"
) ? static_cast<void> (0) : __assert_fail ("(Ordered || !RT.isStaticNonchunked(ScheduleKind.Schedule, LoopArgs.Chunk != nullptr)) && \"static non-chunked schedule does not need outer loop\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 2166, __PRETTY_FUNCTION__))
2164 !RT.isStaticNonchunked(ScheduleKind.Schedule,(((Ordered || !RT.isStaticNonchunked(ScheduleKind.Schedule, LoopArgs
.Chunk != nullptr)) && "static non-chunked schedule does not need outer loop"
) ? static_cast<void> (0) : __assert_fail ("(Ordered || !RT.isStaticNonchunked(ScheduleKind.Schedule, LoopArgs.Chunk != nullptr)) && \"static non-chunked schedule does not need outer loop\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 2166, __PRETTY_FUNCTION__))
2165 LoopArgs.Chunk != nullptr)) &&(((Ordered || !RT.isStaticNonchunked(ScheduleKind.Schedule, LoopArgs
.Chunk != nullptr)) && "static non-chunked schedule does not need outer loop"
) ? static_cast<void> (0) : __assert_fail ("(Ordered || !RT.isStaticNonchunked(ScheduleKind.Schedule, LoopArgs.Chunk != nullptr)) && \"static non-chunked schedule does not need outer loop\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 2166, __PRETTY_FUNCTION__))
2166 "static non-chunked schedule does not need outer loop")(((Ordered || !RT.isStaticNonchunked(ScheduleKind.Schedule, LoopArgs
.Chunk != nullptr)) && "static non-chunked schedule does not need outer loop"
) ? static_cast<void> (0) : __assert_fail ("(Ordered || !RT.isStaticNonchunked(ScheduleKind.Schedule, LoopArgs.Chunk != nullptr)) && \"static non-chunked schedule does not need outer loop\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 2166, __PRETTY_FUNCTION__))
;
2167
2168 // Emit outer loop.
2169 //
2170 // OpenMP [2.7.1, Loop Construct, Description, table 2-1]
2171 // When schedule(dynamic,chunk_size) is specified, the iterations are
2172 // distributed to threads in the team in chunks as the threads request them.
2173 // Each thread executes a chunk of iterations, then requests another chunk,
2174 // until no chunks remain to be distributed. Each chunk contains chunk_size
2175 // iterations, except for the last chunk to be distributed, which may have
2176 // fewer iterations. When no chunk_size is specified, it defaults to 1.
2177 //
2178 // When schedule(guided,chunk_size) is specified, the iterations are assigned
2179 // to threads in the team in chunks as the executing threads request them.
2180 // Each thread executes a chunk of iterations, then requests another chunk,
2181 // until no chunks remain to be assigned. For a chunk_size of 1, the size of
2182 // each chunk is proportional to the number of unassigned iterations divided
2183 // by the number of threads in the team, decreasing to 1. For a chunk_size
2184 // with value k (greater than 1), the size of each chunk is determined in the
2185 // same way, with the restriction that the chunks do not contain fewer than k
2186 // iterations (except for the last chunk to be assigned, which may have fewer
2187 // than k iterations).
2188 //
2189 // When schedule(auto) is specified, the decision regarding scheduling is
2190 // delegated to the compiler and/or runtime system. The programmer gives the
2191 // implementation the freedom to choose any possible mapping of iterations to
2192 // threads in the team.
2193 //
2194 // When schedule(runtime) is specified, the decision regarding scheduling is
2195 // deferred until run time, and the schedule and chunk size are taken from the
2196 // run-sched-var ICV. If the ICV is set to auto, the schedule is
2197 // implementation defined
2198 //
2199 // while(__kmpc_dispatch_next(&LB, &UB)) {
2200 // idx = LB;
2201 // while (idx <= UB) { BODY; ++idx;
2202 // __kmpc_dispatch_fini_(4|8)[u](); // For ordered loops only.
2203 // } // inner loop
2204 // }
2205 //
2206 // OpenMP [2.7.1, Loop Construct, Description, table 2-1]
2207 // When schedule(static, chunk_size) is specified, iterations are divided into
2208 // chunks of size chunk_size, and the chunks are assigned to the threads in
2209 // the team in a round-robin fashion in the order of the thread number.
2210 //
2211 // while(UB = min(UB, GlobalUB), idx = LB, idx < UB) {
2212 // while (idx <= UB) { BODY; ++idx; } // inner loop
2213 // LB = LB + ST;
2214 // UB = UB + ST;
2215 // }
2216 //
2217
2218 const Expr *IVExpr = S.getIterationVariable();
2219 const unsigned IVSize = getContext().getTypeSize(IVExpr->getType());
2220 const bool IVSigned = IVExpr->getType()->hasSignedIntegerRepresentation();
2221
2222 if (DynamicOrOrdered) {
2223 const std::pair<llvm::Value *, llvm::Value *> DispatchBounds =
2224 CGDispatchBounds(*this, S, LoopArgs.LB, LoopArgs.UB);
2225 llvm::Value *LBVal = DispatchBounds.first;
2226 llvm::Value *UBVal = DispatchBounds.second;
2227 CGOpenMPRuntime::DispatchRTInput DipatchRTInputValues = {LBVal, UBVal,
2228 LoopArgs.Chunk};
2229 RT.emitForDispatchInit(*this, S.getBeginLoc(), ScheduleKind, IVSize,
2230 IVSigned, Ordered, DipatchRTInputValues);
2231 } else {
2232 CGOpenMPRuntime::StaticRTInput StaticInit(
2233 IVSize, IVSigned, Ordered, LoopArgs.IL, LoopArgs.LB, LoopArgs.UB,
2234 LoopArgs.ST, LoopArgs.Chunk);
2235 RT.emitForStaticInit(*this, S.getBeginLoc(), S.getDirectiveKind(),
2236 ScheduleKind, StaticInit);
2237 }
2238
2239 auto &&CodeGenOrdered = [Ordered](CodeGenFunction &CGF, SourceLocation Loc,
2240 const unsigned IVSize,
2241 const bool IVSigned) {
2242 if (Ordered) {
2243 CGF.CGM.getOpenMPRuntime().emitForOrderedIterationEnd(CGF, Loc, IVSize,
2244 IVSigned);
2245 }
2246 };
2247
2248 OMPLoopArguments OuterLoopArgs(LoopArgs.LB, LoopArgs.UB, LoopArgs.ST,
2249 LoopArgs.IL, LoopArgs.Chunk, LoopArgs.EUB);
2250 OuterLoopArgs.IncExpr = S.getInc();
2251 OuterLoopArgs.Init = S.getInit();
2252 OuterLoopArgs.Cond = S.getCond();
2253 OuterLoopArgs.NextLB = S.getNextLowerBound();
2254 OuterLoopArgs.NextUB = S.getNextUpperBound();
2255 EmitOMPOuterLoop(DynamicOrOrdered, IsMonotonic, S, LoopScope, OuterLoopArgs,
2256 emitOMPLoopBodyWithStopPoint, CodeGenOrdered);
2257}
2258
2259static void emitEmptyOrdered(CodeGenFunction &, SourceLocation Loc,
2260 const unsigned IVSize, const bool IVSigned) {}
2261
2262void CodeGenFunction::EmitOMPDistributeOuterLoop(
2263 OpenMPDistScheduleClauseKind ScheduleKind, const OMPLoopDirective &S,
2264 OMPPrivateScope &LoopScope, const OMPLoopArguments &LoopArgs,
2265 const CodeGenLoopTy &CodeGenLoopContent) {
2266
2267 CGOpenMPRuntime &RT = CGM.getOpenMPRuntime();
2268
2269 // Emit outer loop.
2270 // Same behavior as a OMPForOuterLoop, except that schedule cannot be
2271 // dynamic
2272 //
2273
2274 const Expr *IVExpr = S.getIterationVariable();
2275 const unsigned IVSize = getContext().getTypeSize(IVExpr->getType());
2276 const bool IVSigned = IVExpr->getType()->hasSignedIntegerRepresentation();
2277
2278 CGOpenMPRuntime::StaticRTInput StaticInit(
2279 IVSize, IVSigned, /* Ordered = */ false, LoopArgs.IL, LoopArgs.LB,
2280 LoopArgs.UB, LoopArgs.ST, LoopArgs.Chunk);
2281 RT.emitDistributeStaticInit(*this, S.getBeginLoc(), ScheduleKind, StaticInit);
2282
2283 // for combined 'distribute' and 'for' the increment expression of distribute
2284 // is stored in DistInc. For 'distribute' alone, it is in Inc.
2285 Expr *IncExpr;
2286 if (isOpenMPLoopBoundSharingDirective(S.getDirectiveKind()))
2287 IncExpr = S.getDistInc();
2288 else
2289 IncExpr = S.getInc();
2290
2291 // this routine is shared by 'omp distribute parallel for' and
2292 // 'omp distribute': select the right EUB expression depending on the
2293 // directive
2294 OMPLoopArguments OuterLoopArgs;
2295 OuterLoopArgs.LB = LoopArgs.LB;
2296 OuterLoopArgs.UB = LoopArgs.UB;
2297 OuterLoopArgs.ST = LoopArgs.ST;
2298 OuterLoopArgs.IL = LoopArgs.IL;
2299 OuterLoopArgs.Chunk = LoopArgs.Chunk;
2300 OuterLoopArgs.EUB = isOpenMPLoopBoundSharingDirective(S.getDirectiveKind())
2301 ? S.getCombinedEnsureUpperBound()
2302 : S.getEnsureUpperBound();
2303 OuterLoopArgs.IncExpr = IncExpr;
2304 OuterLoopArgs.Init = isOpenMPLoopBoundSharingDirective(S.getDirectiveKind())
2305 ? S.getCombinedInit()
2306 : S.getInit();
2307 OuterLoopArgs.Cond = isOpenMPLoopBoundSharingDirective(S.getDirectiveKind())
2308 ? S.getCombinedCond()
2309 : S.getCond();
2310 OuterLoopArgs.NextLB = isOpenMPLoopBoundSharingDirective(S.getDirectiveKind())
2311 ? S.getCombinedNextLowerBound()
2312 : S.getNextLowerBound();
2313 OuterLoopArgs.NextUB = isOpenMPLoopBoundSharingDirective(S.getDirectiveKind())
2314 ? S.getCombinedNextUpperBound()
2315 : S.getNextUpperBound();
2316
2317 EmitOMPOuterLoop(/* DynamicOrOrdered = */ false, /* IsMonotonic = */ false, S,
2318 LoopScope, OuterLoopArgs, CodeGenLoopContent,
2319 emitEmptyOrdered);
2320}
2321
2322static std::pair<LValue, LValue>
2323emitDistributeParallelForInnerBounds(CodeGenFunction &CGF,
2324 const OMPExecutableDirective &S) {
2325 const OMPLoopDirective &LS = cast<OMPLoopDirective>(S);
2326 LValue LB =
2327 EmitOMPHelperVar(CGF, cast<DeclRefExpr>(LS.getLowerBoundVariable()));
2328 LValue UB =
2329 EmitOMPHelperVar(CGF, cast<DeclRefExpr>(LS.getUpperBoundVariable()));
2330
2331 // When composing 'distribute' with 'for' (e.g. as in 'distribute
2332 // parallel for') we need to use the 'distribute'
2333 // chunk lower and upper bounds rather than the whole loop iteration
2334 // space. These are parameters to the outlined function for 'parallel'
2335 // and we copy the bounds of the previous schedule into the
2336 // the current ones.
2337 LValue PrevLB = CGF.EmitLValue(LS.getPrevLowerBoundVariable());
2338 LValue PrevUB = CGF.EmitLValue(LS.getPrevUpperBoundVariable());
2339 llvm::Value *PrevLBVal = CGF.EmitLoadOfScalar(
2340 PrevLB, LS.getPrevLowerBoundVariable()->getExprLoc());
2341 PrevLBVal = CGF.EmitScalarConversion(
2342 PrevLBVal, LS.getPrevLowerBoundVariable()->getType(),
2343 LS.getIterationVariable()->getType(),
2344 LS.getPrevLowerBoundVariable()->getExprLoc());
2345 llvm::Value *PrevUBVal = CGF.EmitLoadOfScalar(
2346 PrevUB, LS.getPrevUpperBoundVariable()->getExprLoc());
2347 PrevUBVal = CGF.EmitScalarConversion(
2348 PrevUBVal, LS.getPrevUpperBoundVariable()->getType(),
2349 LS.getIterationVariable()->getType(),
2350 LS.getPrevUpperBoundVariable()->getExprLoc());
2351
2352 CGF.EmitStoreOfScalar(PrevLBVal, LB);
2353 CGF.EmitStoreOfScalar(PrevUBVal, UB);
2354
2355 return {LB, UB};
2356}
2357
2358/// if the 'for' loop has a dispatch schedule (e.g. dynamic, guided) then
2359/// we need to use the LB and UB expressions generated by the worksharing
2360/// code generation support, whereas in non combined situations we would
2361/// just emit 0 and the LastIteration expression
2362/// This function is necessary due to the difference of the LB and UB
2363/// types for the RT emission routines for 'for_static_init' and
2364/// 'for_dispatch_init'
2365static std::pair<llvm::Value *, llvm::Value *>
2366emitDistributeParallelForDispatchBounds(CodeGenFunction &CGF,
2367 const OMPExecutableDirective &S,
2368 Address LB, Address UB) {
2369 const OMPLoopDirective &LS = cast<OMPLoopDirective>(S);
2370 const Expr *IVExpr = LS.getIterationVariable();
2371 // when implementing a dynamic schedule for a 'for' combined with a
2372 // 'distribute' (e.g. 'distribute parallel for'), the 'for' loop
2373 // is not normalized as each team only executes its own assigned
2374 // distribute chunk
2375 QualType IteratorTy = IVExpr->getType();
2376 llvm::Value *LBVal =
2377 CGF.EmitLoadOfScalar(LB, /*Volatile=*/false, IteratorTy, S.getBeginLoc());
2378 llvm::Value *UBVal =
2379 CGF.EmitLoadOfScalar(UB, /*Volatile=*/false, IteratorTy, S.getBeginLoc());
2380 return {LBVal, UBVal};
2381}
2382
2383static void emitDistributeParallelForDistributeInnerBoundParams(
2384 CodeGenFunction &CGF, const OMPExecutableDirective &S,
2385 llvm::SmallVectorImpl<llvm::Value *> &CapturedVars) {
2386 const auto &Dir = cast<OMPLoopDirective>(S);
2387 LValue LB =
2388 CGF.EmitLValue(cast<DeclRefExpr>(Dir.getCombinedLowerBoundVariable()));
2389 llvm::Value *LBCast =
2390 CGF.Builder.CreateIntCast(CGF.Builder.CreateLoad(LB.getAddress(CGF)),
2391 CGF.SizeTy, /*isSigned=*/false);
2392 CapturedVars.push_back(LBCast);
2393 LValue UB =
2394 CGF.EmitLValue(cast<DeclRefExpr>(Dir.getCombinedUpperBoundVariable()));
2395
2396 llvm::Value *UBCast =
2397 CGF.Builder.CreateIntCast(CGF.Builder.CreateLoad(UB.getAddress(CGF)),
2398 CGF.SizeTy, /*isSigned=*/false);
2399 CapturedVars.push_back(UBCast);
2400}
2401
2402static void
2403emitInnerParallelForWhenCombined(CodeGenFunction &CGF,
2404 const OMPLoopDirective &S,
2405 CodeGenFunction::JumpDest LoopExit) {
2406 auto &&CGInlinedWorksharingLoop = [&S](CodeGenFunction &CGF,
2407 PrePostActionTy &Action) {
2408 Action.Enter(CGF);
2409 bool HasCancel = false;
2410 if (!isOpenMPSimdDirective(S.getDirectiveKind())) {
2411 if (const auto *D = dyn_cast<OMPTeamsDistributeParallelForDirective>(&S))
2412 HasCancel = D->hasCancel();
2413 else if (const auto *D = dyn_cast<OMPDistributeParallelForDirective>(&S))
2414 HasCancel = D->hasCancel();
2415 else if (const auto *D =
2416 dyn_cast<OMPTargetTeamsDistributeParallelForDirective>(&S))
2417 HasCancel = D->hasCancel();
2418 }
2419 CodeGenFunction::OMPCancelStackRAII CancelRegion(CGF, S.getDirectiveKind(),
2420 HasCancel);
2421 CGF.EmitOMPWorksharingLoop(S, S.getPrevEnsureUpperBound(),
2422 emitDistributeParallelForInnerBounds,
2423 emitDistributeParallelForDispatchBounds);
2424 };
2425
2426 emitCommonOMPParallelDirective(
2427 CGF, S,
2428 isOpenMPSimdDirective(S.getDirectiveKind()) ? OMPD_for_simd : OMPD_for,
2429 CGInlinedWorksharingLoop,
2430 emitDistributeParallelForDistributeInnerBoundParams);
2431}
2432
2433void CodeGenFunction::EmitOMPDistributeParallelForDirective(
2434 const OMPDistributeParallelForDirective &S) {
2435 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &) {
2436 CGF.EmitOMPDistributeLoop(S, emitInnerParallelForWhenCombined,
2437 S.getDistInc());
2438 };
2439 OMPLexicalScope Scope(*this, S, OMPD_parallel);
2440 CGM.getOpenMPRuntime().emitInlinedDirective(*this, OMPD_distribute, CodeGen);
2441}
2442
2443void CodeGenFunction::EmitOMPDistributeParallelForSimdDirective(
2444 const OMPDistributeParallelForSimdDirective &S) {
2445 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &) {
2446 CGF.EmitOMPDistributeLoop(S, emitInnerParallelForWhenCombined,
2447 S.getDistInc());
2448 };
2449 OMPLexicalScope Scope(*this, S, OMPD_parallel);
2450 CGM.getOpenMPRuntime().emitInlinedDirective(*this, OMPD_distribute, CodeGen);
2451}
2452
2453void CodeGenFunction::EmitOMPDistributeSimdDirective(
2454 const OMPDistributeSimdDirective &S) {
2455 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &) {
2456 CGF.EmitOMPDistributeLoop(S, emitOMPLoopBodyWithStopPoint, S.getInc());
2457 };
2458 OMPLexicalScope Scope(*this, S, OMPD_unknown);
2459 CGM.getOpenMPRuntime().emitInlinedDirective(*this, OMPD_simd, CodeGen);
2460}
2461
2462void CodeGenFunction::EmitOMPTargetSimdDeviceFunction(
2463 CodeGenModule &CGM, StringRef ParentName, const OMPTargetSimdDirective &S) {
2464 // Emit SPMD target parallel for region as a standalone region.
2465 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
2466 emitOMPSimdRegion(CGF, S, Action);
2467 };
2468 llvm::Function *Fn;
2469 llvm::Constant *Addr;
2470 // Emit target region as a standalone region.
2471 CGM.getOpenMPRuntime().emitTargetOutlinedFunction(
2472 S, ParentName, Fn, Addr, /*IsOffloadEntry=*/true, CodeGen);
2473 assert(Fn && Addr && "Target device function emission failed.")((Fn && Addr && "Target device function emission failed."
) ? static_cast<void> (0) : __assert_fail ("Fn && Addr && \"Target device function emission failed.\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 2473, __PRETTY_FUNCTION__))
;
2474}
2475
2476void CodeGenFunction::EmitOMPTargetSimdDirective(
2477 const OMPTargetSimdDirective &S) {
2478 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
2479 emitOMPSimdRegion(CGF, S, Action);
2480 };
2481 emitCommonOMPTargetDirective(*this, S, CodeGen);
2482}
2483
2484namespace {
2485 struct ScheduleKindModifiersTy {
2486 OpenMPScheduleClauseKind Kind;
2487 OpenMPScheduleClauseModifier M1;
2488 OpenMPScheduleClauseModifier M2;
2489 ScheduleKindModifiersTy(OpenMPScheduleClauseKind Kind,
2490 OpenMPScheduleClauseModifier M1,
2491 OpenMPScheduleClauseModifier M2)
2492 : Kind(Kind), M1(M1), M2(M2) {}
2493 };
2494} // namespace
2495
2496bool CodeGenFunction::EmitOMPWorksharingLoop(
2497 const OMPLoopDirective &S, Expr *EUB,
2498 const CodeGenLoopBoundsTy &CodeGenLoopBounds,
2499 const CodeGenDispatchBoundsTy &CGDispatchBounds) {
2500 // Emit the loop iteration variable.
2501 const auto *IVExpr = cast<DeclRefExpr>(S.getIterationVariable());
2502 const auto *IVDecl = cast<VarDecl>(IVExpr->getDecl());
2503 EmitVarDecl(*IVDecl);
2504
2505 // Emit the iterations count variable.
2506 // If it is not a variable, Sema decided to calculate iterations count on each
2507 // iteration (e.g., it is foldable into a constant).
2508 if (const auto *LIExpr = dyn_cast<DeclRefExpr>(S.getLastIteration())) {
2509 EmitVarDecl(*cast<VarDecl>(LIExpr->getDecl()));
2510 // Emit calculation of the iterations count.
2511 EmitIgnoredExpr(S.getCalcLastIteration());
2512 }
2513
2514 CGOpenMPRuntime &RT = CGM.getOpenMPRuntime();
2515
2516 bool HasLastprivateClause;
2517 // Check pre-condition.
2518 {
2519 OMPLoopScope PreInitScope(*this, S);
2520 // Skip the entire loop if we don't meet the precondition.
2521 // If the condition constant folds and can be elided, avoid emitting the
2522 // whole loop.
2523 bool CondConstant;
2524 llvm::BasicBlock *ContBlock = nullptr;
2525 if (ConstantFoldsToSimpleInteger(S.getPreCond(), CondConstant)) {
2526 if (!CondConstant)
2527 return false;
2528 } else {
2529 llvm::BasicBlock *ThenBlock = createBasicBlock("omp.precond.then");
2530 ContBlock = createBasicBlock("omp.precond.end");
2531 emitPreCond(*this, S, S.getPreCond(), ThenBlock, ContBlock,
2532 getProfileCount(&S));
2533 EmitBlock(ThenBlock);
2534 incrementProfileCounter(&S);
2535 }
2536
2537 RunCleanupsScope DoacrossCleanupScope(*this);
2538 bool Ordered = false;
2539 if (const auto *OrderedClause = S.getSingleClause<OMPOrderedClause>()) {
2540 if (OrderedClause->getNumForLoops())
2541 RT.emitDoacrossInit(*this, S, OrderedClause->getLoopNumIterations());
2542 else
2543 Ordered = true;
2544 }
2545
2546 llvm::DenseSet<const Expr *> EmittedFinals;
2547 emitAlignedClause(*this, S);
2548 bool HasLinears = EmitOMPLinearClauseInit(S);
2549 // Emit helper vars inits.
2550
2551 std::pair<LValue, LValue> Bounds = CodeGenLoopBounds(*this, S);
2552 LValue LB = Bounds.first;
2553 LValue UB = Bounds.second;
2554 LValue ST =
2555 EmitOMPHelperVar(*this, cast<DeclRefExpr>(S.getStrideVariable()));
2556 LValue IL =
2557 EmitOMPHelperVar(*this, cast<DeclRefExpr>(S.getIsLastIterVariable()));
2558
2559 // Emit 'then' code.
2560 {
2561 OMPPrivateScope LoopScope(*this);
2562 if (EmitOMPFirstprivateClause(S, LoopScope) || HasLinears) {
2563 // Emit implicit barrier to synchronize threads and avoid data races on
2564 // initialization of firstprivate variables and post-update of
2565 // lastprivate variables.
2566 CGM.getOpenMPRuntime().emitBarrierCall(
2567 *this, S.getBeginLoc(), OMPD_unknown, /*EmitChecks=*/false,
2568 /*ForceSimpleCall=*/true);
2569 }
2570 EmitOMPPrivateClause(S, LoopScope);
2571 CGOpenMPRuntime::LastprivateConditionalRAII LPCRegion(
2572 *this, S, EmitLValue(S.getIterationVariable()));
2573 HasLastprivateClause = EmitOMPLastprivateClauseInit(S, LoopScope);
2574 EmitOMPReductionClauseInit(S, LoopScope);
2575 EmitOMPPrivateLoopCounters(S, LoopScope);
2576 EmitOMPLinearClause(S, LoopScope);
2577 (void)LoopScope.Privatize();
2578 if (isOpenMPTargetExecutionDirective(S.getDirectiveKind()))
2579 CGM.getOpenMPRuntime().adjustTargetSpecificDataForLambdas(*this, S);
2580
2581 // Detect the loop schedule kind and chunk.
2582 const Expr *ChunkExpr = nullptr;
2583 OpenMPScheduleTy ScheduleKind;
2584 if (const auto *C = S.getSingleClause<OMPScheduleClause>()) {
2585 ScheduleKind.Schedule = C->getScheduleKind();
2586 ScheduleKind.M1 = C->getFirstScheduleModifier();
2587 ScheduleKind.M2 = C->getSecondScheduleModifier();
2588 ChunkExpr = C->getChunkSize();
2589 } else {
2590 // Default behaviour for schedule clause.
2591 CGM.getOpenMPRuntime().getDefaultScheduleAndChunk(
2592 *this, S, ScheduleKind.Schedule, ChunkExpr);
2593 }
2594 bool HasChunkSizeOne = false;
2595 llvm::Value *Chunk = nullptr;
2596 if (ChunkExpr) {
2597 Chunk = EmitScalarExpr(ChunkExpr);
2598 Chunk = EmitScalarConversion(Chunk, ChunkExpr->getType(),
2599 S.getIterationVariable()->getType(),
2600 S.getBeginLoc());
2601 Expr::EvalResult Result;
2602 if (ChunkExpr->EvaluateAsInt(Result, getContext())) {
2603 llvm::APSInt EvaluatedChunk = Result.Val.getInt();
2604 HasChunkSizeOne = (EvaluatedChunk.getLimitedValue() == 1);
2605 }
2606 }
2607 const unsigned IVSize = getContext().getTypeSize(IVExpr->getType());
2608 const bool IVSigned = IVExpr->getType()->hasSignedIntegerRepresentation();
2609 // OpenMP 4.5, 2.7.1 Loop Construct, Description.
2610 // If the static schedule kind is specified or if the ordered clause is
2611 // specified, and if no monotonic modifier is specified, the effect will
2612 // be as if the monotonic modifier was specified.
2613 bool StaticChunkedOne = RT.isStaticChunked(ScheduleKind.Schedule,
2614 /* Chunked */ Chunk != nullptr) && HasChunkSizeOne &&
2615 isOpenMPLoopBoundSharingDirective(S.getDirectiveKind());
2616 if ((RT.isStaticNonchunked(ScheduleKind.Schedule,
2617 /* Chunked */ Chunk != nullptr) ||
2618 StaticChunkedOne) &&
2619 !Ordered) {
2620 JumpDest LoopExit =
2621 getJumpDestInCurrentScope(createBasicBlock("omp.loop.exit"));
2622 emitCommonSimdLoop(
2623 *this, S,
2624 [&S](CodeGenFunction &CGF, PrePostActionTy &) {
2625 if (isOpenMPSimdDirective(S.getDirectiveKind()))
2626 CGF.EmitOMPSimdInit(S, /*IsMonotonic=*/true);
2627 },
2628 [IVSize, IVSigned, Ordered, IL, LB, UB, ST, StaticChunkedOne, Chunk,
2629 &S, ScheduleKind, LoopExit,
2630 &LoopScope](CodeGenFunction &CGF, PrePostActionTy &) {
2631 // OpenMP [2.7.1, Loop Construct, Description, table 2-1]
2632 // When no chunk_size is specified, the iteration space is divided
2633 // into chunks that are approximately equal in size, and at most
2634 // one chunk is distributed to each thread. Note that the size of
2635 // the chunks is unspecified in this case.
2636 CGOpenMPRuntime::StaticRTInput StaticInit(
2637 IVSize, IVSigned, Ordered, IL.getAddress(CGF),
2638 LB.getAddress(CGF), UB.getAddress(CGF), ST.getAddress(CGF),
2639 StaticChunkedOne ? Chunk : nullptr);
2640 CGF.CGM.getOpenMPRuntime().emitForStaticInit(
2641 CGF, S.getBeginLoc(), S.getDirectiveKind(), ScheduleKind,
2642 StaticInit);
2643 // UB = min(UB, GlobalUB);
2644 if (!StaticChunkedOne)
2645 CGF.EmitIgnoredExpr(S.getEnsureUpperBound());
2646 // IV = LB;
2647 CGF.EmitIgnoredExpr(S.getInit());
2648 // For unchunked static schedule generate:
2649 //
2650 // while (idx <= UB) {
2651 // BODY;
2652 // ++idx;
2653 // }
2654 //
2655 // For static schedule with chunk one:
2656 //
2657 // while (IV <= PrevUB) {
2658 // BODY;
2659 // IV += ST;
2660 // }
2661 CGF.EmitOMPInnerLoop(
2662 S, LoopScope.requiresCleanups(),
2663 StaticChunkedOne ? S.getCombinedParForInDistCond()
2664 : S.getCond(),
2665 StaticChunkedOne ? S.getDistInc() : S.getInc(),
2666 [&S, LoopExit](CodeGenFunction &CGF) {
2667 CGF.CGM.getOpenMPRuntime()
2668 .initLastprivateConditionalCounter(CGF, S);
2669 CGF.EmitOMPLoopBody(S, LoopExit);
2670 CGF.EmitStopPoint(&S);
2671 },
2672 [](CodeGenFunction &) {});
2673 });
2674 EmitBlock(LoopExit.getBlock());
2675 // Tell the runtime we are done.
2676 auto &&CodeGen = [&S](CodeGenFunction &CGF) {
2677 CGF.CGM.getOpenMPRuntime().emitForStaticFinish(CGF, S.getEndLoc(),
2678 S.getDirectiveKind());
2679 };
2680 OMPCancelStack.emitExit(*this, S.getDirectiveKind(), CodeGen);
2681 } else {
2682 const bool IsMonotonic =
2683 Ordered || ScheduleKind.Schedule == OMPC_SCHEDULE_static ||
2684 ScheduleKind.Schedule == OMPC_SCHEDULE_unknown ||
2685 ScheduleKind.M1 == OMPC_SCHEDULE_MODIFIER_monotonic ||
2686 ScheduleKind.M2 == OMPC_SCHEDULE_MODIFIER_monotonic;
2687 // Emit the outer loop, which requests its work chunk [LB..UB] from
2688 // runtime and runs the inner loop to process it.
2689 const OMPLoopArguments LoopArguments(
2690 LB.getAddress(*this), UB.getAddress(*this), ST.getAddress(*this),
2691 IL.getAddress(*this), Chunk, EUB);
2692 EmitOMPForOuterLoop(ScheduleKind, IsMonotonic, S, LoopScope, Ordered,
2693 LoopArguments, CGDispatchBounds);
2694 }
2695 if (isOpenMPSimdDirective(S.getDirectiveKind())) {
2696 EmitOMPSimdFinal(S, [IL, &S](CodeGenFunction &CGF) {
2697 return CGF.Builder.CreateIsNotNull(
2698 CGF.EmitLoadOfScalar(IL, S.getBeginLoc()));
2699 });
2700 }
2701 EmitOMPReductionClauseFinal(
2702 S, /*ReductionKind=*/isOpenMPSimdDirective(S.getDirectiveKind())
2703 ? /*Parallel and Simd*/ OMPD_parallel_for_simd
2704 : /*Parallel only*/ OMPD_parallel);
2705 // Emit post-update of the reduction variables if IsLastIter != 0.
2706 emitPostUpdateForReductionClause(
2707 *this, S, [IL, &S](CodeGenFunction &CGF) {
2708 return CGF.Builder.CreateIsNotNull(
2709 CGF.EmitLoadOfScalar(IL, S.getBeginLoc()));
2710 });
2711 // Emit final copy of the lastprivate variables if IsLastIter != 0.
2712 if (HasLastprivateClause)
2713 EmitOMPLastprivateClauseFinal(
2714 S, isOpenMPSimdDirective(S.getDirectiveKind()),
2715 Builder.CreateIsNotNull(EmitLoadOfScalar(IL, S.getBeginLoc())));
2716 }
2717 EmitOMPLinearClauseFinal(S, [IL, &S](CodeGenFunction &CGF) {
2718 return CGF.Builder.CreateIsNotNull(
2719 CGF.EmitLoadOfScalar(IL, S.getBeginLoc()));
2720 });
2721 DoacrossCleanupScope.ForceCleanup();
2722 // We're now done with the loop, so jump to the continuation block.
2723 if (ContBlock) {
2724 EmitBranch(ContBlock);
2725 EmitBlock(ContBlock, /*IsFinished=*/true);
2726 }
2727 }
2728 return HasLastprivateClause;
2729}
2730
2731/// The following two functions generate expressions for the loop lower
2732/// and upper bounds in case of static and dynamic (dispatch) schedule
2733/// of the associated 'for' or 'distribute' loop.
2734static std::pair<LValue, LValue>
2735emitForLoopBounds(CodeGenFunction &CGF, const OMPExecutableDirective &S) {
2736 const auto &LS = cast<OMPLoopDirective>(S);
2737 LValue LB =
2738 EmitOMPHelperVar(CGF, cast<DeclRefExpr>(LS.getLowerBoundVariable()));
2739 LValue UB =
2740 EmitOMPHelperVar(CGF, cast<DeclRefExpr>(LS.getUpperBoundVariable()));
2741 return {LB, UB};
2742}
2743
2744/// When dealing with dispatch schedules (e.g. dynamic, guided) we do not
2745/// consider the lower and upper bound expressions generated by the
2746/// worksharing loop support, but we use 0 and the iteration space size as
2747/// constants
2748static std::pair<llvm::Value *, llvm::Value *>
2749emitDispatchForLoopBounds(CodeGenFunction &CGF, const OMPExecutableDirective &S,
2750 Address LB, Address UB) {
2751 const auto &LS = cast<OMPLoopDirective>(S);
2752 const Expr *IVExpr = LS.getIterationVariable();
2753 const unsigned IVSize = CGF.getContext().getTypeSize(IVExpr->getType());
2754 llvm::Value *LBVal = CGF.Builder.getIntN(IVSize, 0);
2755 llvm::Value *UBVal = CGF.EmitScalarExpr(LS.getLastIteration());
2756 return {LBVal, UBVal};
2757}
2758
2759void CodeGenFunction::EmitOMPForDirective(const OMPForDirective &S) {
2760 bool HasLastprivates = false;
2761 auto &&CodeGen = [&S, &HasLastprivates](CodeGenFunction &CGF,
2762 PrePostActionTy &) {
2763 OMPCancelStackRAII CancelRegion(CGF, OMPD_for, S.hasCancel());
2764 HasLastprivates = CGF.EmitOMPWorksharingLoop(S, S.getEnsureUpperBound(),
2765 emitForLoopBounds,
2766 emitDispatchForLoopBounds);
2767 };
2768 {
2769 OMPLexicalScope Scope(*this, S, OMPD_unknown);
2770 CGM.getOpenMPRuntime().emitInlinedDirective(*this, OMPD_for, CodeGen,
2771 S.hasCancel());
2772 }
2773
2774 // Emit an implicit barrier at the end.
2775 if (!S.getSingleClause<OMPNowaitClause>() || HasLastprivates)
2776 CGM.getOpenMPRuntime().emitBarrierCall(*this, S.getBeginLoc(), OMPD_for);
2777}
2778
2779void CodeGenFunction::EmitOMPForSimdDirective(const OMPForSimdDirective &S) {
2780 bool HasLastprivates = false;
2781 auto &&CodeGen = [&S, &HasLastprivates](CodeGenFunction &CGF,
2782 PrePostActionTy &) {
2783 HasLastprivates = CGF.EmitOMPWorksharingLoop(S, S.getEnsureUpperBound(),
2784 emitForLoopBounds,
2785 emitDispatchForLoopBounds);
2786 };
2787 {
2788 OMPLexicalScope Scope(*this, S, OMPD_unknown);
2789 CGM.getOpenMPRuntime().emitInlinedDirective(*this, OMPD_simd, CodeGen);
2790 }
2791
2792 // Emit an implicit barrier at the end.
2793 if (!S.getSingleClause<OMPNowaitClause>() || HasLastprivates)
2794 CGM.getOpenMPRuntime().emitBarrierCall(*this, S.getBeginLoc(), OMPD_for);
2795}
2796
2797static LValue createSectionLVal(CodeGenFunction &CGF, QualType Ty,
2798 const Twine &Name,
2799 llvm::Value *Init = nullptr) {
2800 LValue LVal = CGF.MakeAddrLValue(CGF.CreateMemTemp(Ty, Name), Ty);
2801 if (Init)
2802 CGF.EmitStoreThroughLValue(RValue::get(Init), LVal, /*isInit*/ true);
2803 return LVal;
2804}
2805
2806void CodeGenFunction::EmitSections(const OMPExecutableDirective &S) {
2807 const Stmt *CapturedStmt = S.getInnermostCapturedStmt()->getCapturedStmt();
2808 const auto *CS = dyn_cast<CompoundStmt>(CapturedStmt);
2809 bool HasLastprivates = false;
2810 auto &&CodeGen = [&S, CapturedStmt, CS,
2811 &HasLastprivates](CodeGenFunction &CGF, PrePostActionTy &) {
2812 ASTContext &C = CGF.getContext();
2813 QualType KmpInt32Ty =
2814 C.getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/1);
2815 // Emit helper vars inits.
2816 LValue LB = createSectionLVal(CGF, KmpInt32Ty, ".omp.sections.lb.",
2817 CGF.Builder.getInt32(0));
2818 llvm::ConstantInt *GlobalUBVal = CS != nullptr
2819 ? CGF.Builder.getInt32(CS->size() - 1)
2820 : CGF.Builder.getInt32(0);
2821 LValue UB =
2822 createSectionLVal(CGF, KmpInt32Ty, ".omp.sections.ub.", GlobalUBVal);
2823 LValue ST = createSectionLVal(CGF, KmpInt32Ty, ".omp.sections.st.",
2824 CGF.Builder.getInt32(1));
2825 LValue IL = createSectionLVal(CGF, KmpInt32Ty, ".omp.sections.il.",
2826 CGF.Builder.getInt32(0));
2827 // Loop counter.
2828 LValue IV = createSectionLVal(CGF, KmpInt32Ty, ".omp.sections.iv.");
2829 OpaqueValueExpr IVRefExpr(S.getBeginLoc(), KmpInt32Ty, VK_LValue);
2830 CodeGenFunction::OpaqueValueMapping OpaqueIV(CGF, &IVRefExpr, IV);
2831 OpaqueValueExpr UBRefExpr(S.getBeginLoc(), KmpInt32Ty, VK_LValue);
2832 CodeGenFunction::OpaqueValueMapping OpaqueUB(CGF, &UBRefExpr, UB);
2833 // Generate condition for loop.
2834 BinaryOperator Cond(&IVRefExpr, &UBRefExpr, BO_LE, C.BoolTy, VK_RValue,
2835 OK_Ordinary, S.getBeginLoc(), FPOptions());
2836 // Increment for loop counter.
2837 UnaryOperator Inc(&IVRefExpr, UO_PreInc, KmpInt32Ty, VK_RValue, OK_Ordinary,
2838 S.getBeginLoc(), true);
2839 auto &&BodyGen = [CapturedStmt, CS, &S, &IV](CodeGenFunction &CGF) {
2840 // Iterate through all sections and emit a switch construct:
2841 // switch (IV) {
2842 // case 0:
2843 // <SectionStmt[0]>;
2844 // break;
2845 // ...
2846 // case <NumSection> - 1:
2847 // <SectionStmt[<NumSection> - 1]>;
2848 // break;
2849 // }
2850 // .omp.sections.exit:
2851 CGF.CGM.getOpenMPRuntime().initLastprivateConditionalCounter(CGF, S);
2852 llvm::BasicBlock *ExitBB = CGF.createBasicBlock(".omp.sections.exit");
2853 llvm::SwitchInst *SwitchStmt =
2854 CGF.Builder.CreateSwitch(CGF.EmitLoadOfScalar(IV, S.getBeginLoc()),
2855 ExitBB, CS == nullptr ? 1 : CS->size());
2856 if (CS) {
2857 unsigned CaseNumber = 0;
2858 for (const Stmt *SubStmt : CS->children()) {
2859 auto CaseBB = CGF.createBasicBlock(".omp.sections.case");
2860 CGF.EmitBlock(CaseBB);
2861 SwitchStmt->addCase(CGF.Builder.getInt32(CaseNumber), CaseBB);
2862 CGF.EmitStmt(SubStmt);
2863 CGF.EmitBranch(ExitBB);
2864 ++CaseNumber;
2865 }
2866 } else {
2867 llvm::BasicBlock *CaseBB = CGF.createBasicBlock(".omp.sections.case");
2868 CGF.EmitBlock(CaseBB);
2869 SwitchStmt->addCase(CGF.Builder.getInt32(0), CaseBB);
2870 CGF.EmitStmt(CapturedStmt);
2871 CGF.EmitBranch(ExitBB);
2872 }
2873 CGF.EmitBlock(ExitBB, /*IsFinished=*/true);
2874 };
2875
2876 CodeGenFunction::OMPPrivateScope LoopScope(CGF);
2877 if (CGF.EmitOMPFirstprivateClause(S, LoopScope)) {
2878 // Emit implicit barrier to synchronize threads and avoid data races on
2879 // initialization of firstprivate variables and post-update of lastprivate
2880 // variables.
2881 CGF.CGM.getOpenMPRuntime().emitBarrierCall(
2882 CGF, S.getBeginLoc(), OMPD_unknown, /*EmitChecks=*/false,
2883 /*ForceSimpleCall=*/true);
2884 }
2885 CGF.EmitOMPPrivateClause(S, LoopScope);
2886 CGOpenMPRuntime::LastprivateConditionalRAII LPCRegion(CGF, S, IV);
2887 HasLastprivates = CGF.EmitOMPLastprivateClauseInit(S, LoopScope);
2888 CGF.EmitOMPReductionClauseInit(S, LoopScope);
2889 (void)LoopScope.Privatize();
2890 if (isOpenMPTargetExecutionDirective(S.getDirectiveKind()))
2891 CGF.CGM.getOpenMPRuntime().adjustTargetSpecificDataForLambdas(CGF, S);
2892
2893 // Emit static non-chunked loop.
2894 OpenMPScheduleTy ScheduleKind;
2895 ScheduleKind.Schedule = OMPC_SCHEDULE_static;
2896 CGOpenMPRuntime::StaticRTInput StaticInit(
2897 /*IVSize=*/32, /*IVSigned=*/true, /*Ordered=*/false, IL.getAddress(CGF),
2898 LB.getAddress(CGF), UB.getAddress(CGF), ST.getAddress(CGF));
2899 CGF.CGM.getOpenMPRuntime().emitForStaticInit(
2900 CGF, S.getBeginLoc(), S.getDirectiveKind(), ScheduleKind, StaticInit);
2901 // UB = min(UB, GlobalUB);
2902 llvm::Value *UBVal = CGF.EmitLoadOfScalar(UB, S.getBeginLoc());
2903 llvm::Value *MinUBGlobalUB = CGF.Builder.CreateSelect(
2904 CGF.Builder.CreateICmpSLT(UBVal, GlobalUBVal), UBVal, GlobalUBVal);
2905 CGF.EmitStoreOfScalar(MinUBGlobalUB, UB);
2906 // IV = LB;
2907 CGF.EmitStoreOfScalar(CGF.EmitLoadOfScalar(LB, S.getBeginLoc()), IV);
2908 // while (idx <= UB) { BODY; ++idx; }
2909 CGF.EmitOMPInnerLoop(S, /*RequiresCleanup=*/false, &Cond, &Inc, BodyGen,
2910 [](CodeGenFunction &) {});
2911 // Tell the runtime we are done.
2912 auto &&CodeGen = [&S](CodeGenFunction &CGF) {
2913 CGF.CGM.getOpenMPRuntime().emitForStaticFinish(CGF, S.getEndLoc(),
2914 S.getDirectiveKind());
2915 };
2916 CGF.OMPCancelStack.emitExit(CGF, S.getDirectiveKind(), CodeGen);
2917 CGF.EmitOMPReductionClauseFinal(S, /*ReductionKind=*/OMPD_parallel);
2918 // Emit post-update of the reduction variables if IsLastIter != 0.
2919 emitPostUpdateForReductionClause(CGF, S, [IL, &S](CodeGenFunction &CGF) {
2920 return CGF.Builder.CreateIsNotNull(
2921 CGF.EmitLoadOfScalar(IL, S.getBeginLoc()));
2922 });
2923
2924 // Emit final copy of the lastprivate variables if IsLastIter != 0.
2925 if (HasLastprivates)
2926 CGF.EmitOMPLastprivateClauseFinal(
2927 S, /*NoFinals=*/false,
2928 CGF.Builder.CreateIsNotNull(
2929 CGF.EmitLoadOfScalar(IL, S.getBeginLoc())));
2930 };
2931
2932 bool HasCancel = false;
2933 if (auto *OSD = dyn_cast<OMPSectionsDirective>(&S))
2934 HasCancel = OSD->hasCancel();
2935 else if (auto *OPSD = dyn_cast<OMPParallelSectionsDirective>(&S))
2936 HasCancel = OPSD->hasCancel();
2937 OMPCancelStackRAII CancelRegion(*this, S.getDirectiveKind(), HasCancel);
2938 CGM.getOpenMPRuntime().emitInlinedDirective(*this, OMPD_sections, CodeGen,
2939 HasCancel);
2940 // Emit barrier for lastprivates only if 'sections' directive has 'nowait'
2941 // clause. Otherwise the barrier will be generated by the codegen for the
2942 // directive.
2943 if (HasLastprivates && S.getSingleClause<OMPNowaitClause>()) {
2944 // Emit implicit barrier to synchronize threads and avoid data races on
2945 // initialization of firstprivate variables.
2946 CGM.getOpenMPRuntime().emitBarrierCall(*this, S.getBeginLoc(),
2947 OMPD_unknown);
2948 }
2949}
2950
2951void CodeGenFunction::EmitOMPSectionsDirective(const OMPSectionsDirective &S) {
2952 {
2953 OMPLexicalScope Scope(*this, S, OMPD_unknown);
2954 EmitSections(S);
2955 }
2956 // Emit an implicit barrier at the end.
2957 if (!S.getSingleClause<OMPNowaitClause>()) {
2958 CGM.getOpenMPRuntime().emitBarrierCall(*this, S.getBeginLoc(),
2959 OMPD_sections);
2960 }
2961}
2962
2963void CodeGenFunction::EmitOMPSectionDirective(const OMPSectionDirective &S) {
2964 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &) {
2965 CGF.EmitStmt(S.getInnermostCapturedStmt()->getCapturedStmt());
2966 };
2967 OMPLexicalScope Scope(*this, S, OMPD_unknown);
2968 CGM.getOpenMPRuntime().emitInlinedDirective(*this, OMPD_section, CodeGen,
2969 S.hasCancel());
2970}
2971
2972void CodeGenFunction::EmitOMPSingleDirective(const OMPSingleDirective &S) {
2973 llvm::SmallVector<const Expr *, 8> CopyprivateVars;
2974 llvm::SmallVector<const Expr *, 8> DestExprs;
2975 llvm::SmallVector<const Expr *, 8> SrcExprs;
2976 llvm::SmallVector<const Expr *, 8> AssignmentOps;
2977 // Check if there are any 'copyprivate' clauses associated with this
2978 // 'single' construct.
2979 // Build a list of copyprivate variables along with helper expressions
2980 // (<source>, <destination>, <destination>=<source> expressions)
2981 for (const auto *C : S.getClausesOfKind<OMPCopyprivateClause>()) {
2982 CopyprivateVars.append(C->varlists().begin(), C->varlists().end());
2983 DestExprs.append(C->destination_exprs().begin(),
2984 C->destination_exprs().end());
2985 SrcExprs.append(C->source_exprs().begin(), C->source_exprs().end());
2986 AssignmentOps.append(C->assignment_ops().begin(),
2987 C->assignment_ops().end());
2988 }
2989 // Emit code for 'single' region along with 'copyprivate' clauses
2990 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
2991 Action.Enter(CGF);
2992 OMPPrivateScope SingleScope(CGF);
2993 (void)CGF.EmitOMPFirstprivateClause(S, SingleScope);
2994 CGF.EmitOMPPrivateClause(S, SingleScope);
2995 (void)SingleScope.Privatize();
2996 CGF.EmitStmt(S.getInnermostCapturedStmt()->getCapturedStmt());
2997 };
2998 {
2999 OMPLexicalScope Scope(*this, S, OMPD_unknown);
3000 CGM.getOpenMPRuntime().emitSingleRegion(*this, CodeGen, S.getBeginLoc(),
3001 CopyprivateVars, DestExprs,
3002 SrcExprs, AssignmentOps);
3003 }
3004 // Emit an implicit barrier at the end (to avoid data race on firstprivate
3005 // init or if no 'nowait' clause was specified and no 'copyprivate' clause).
3006 if (!S.getSingleClause<OMPNowaitClause>() && CopyprivateVars.empty()) {
3007 CGM.getOpenMPRuntime().emitBarrierCall(
3008 *this, S.getBeginLoc(),
3009 S.getSingleClause<OMPNowaitClause>() ? OMPD_unknown : OMPD_single);
3010 }
3011}
3012
3013static void emitMaster(CodeGenFunction &CGF, const OMPExecutableDirective &S) {
3014 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
3015 Action.Enter(CGF);
3016 CGF.EmitStmt(S.getInnermostCapturedStmt()->getCapturedStmt());
3017 };
3018 CGF.CGM.getOpenMPRuntime().emitMasterRegion(CGF, CodeGen, S.getBeginLoc());
3019}
3020
3021void CodeGenFunction::EmitOMPMasterDirective(const OMPMasterDirective &S) {
3022 OMPLexicalScope Scope(*this, S, OMPD_unknown);
3023 emitMaster(*this, S);
3024}
3025
3026void CodeGenFunction::EmitOMPCriticalDirective(const OMPCriticalDirective &S) {
3027 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
3028 Action.Enter(CGF);
3029 CGF.EmitStmt(S.getInnermostCapturedStmt()->getCapturedStmt());
3030 };
3031 const Expr *Hint = nullptr;
3032 if (const auto *HintClause = S.getSingleClause<OMPHintClause>())
3033 Hint = HintClause->getHint();
3034 OMPLexicalScope Scope(*this, S, OMPD_unknown);
3035 CGM.getOpenMPRuntime().emitCriticalRegion(*this,
3036 S.getDirectiveName().getAsString(),
3037 CodeGen, S.getBeginLoc(), Hint);
3038}
3039
3040void CodeGenFunction::EmitOMPParallelForDirective(
3041 const OMPParallelForDirective &S) {
3042 // Emit directive as a combined directive that consists of two implicit
3043 // directives: 'parallel' with 'for' directive.
3044 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
3045 Action.Enter(CGF);
3046 OMPCancelStackRAII CancelRegion(CGF, OMPD_parallel_for, S.hasCancel());
3047 CGF.EmitOMPWorksharingLoop(S, S.getEnsureUpperBound(), emitForLoopBounds,
3048 emitDispatchForLoopBounds);
3049 };
3050 emitCommonOMPParallelDirective(*this, S, OMPD_for, CodeGen,
3051 emitEmptyBoundParameters);
3052}
3053
3054void CodeGenFunction::EmitOMPParallelForSimdDirective(
3055 const OMPParallelForSimdDirective &S) {
3056 // Emit directive as a combined directive that consists of two implicit
3057 // directives: 'parallel' with 'for' directive.
3058 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
3059 Action.Enter(CGF);
3060 CGF.EmitOMPWorksharingLoop(S, S.getEnsureUpperBound(), emitForLoopBounds,
3061 emitDispatchForLoopBounds);
3062 };
3063 emitCommonOMPParallelDirective(*this, S, OMPD_simd, CodeGen,
3064 emitEmptyBoundParameters);
3065}
3066
3067void CodeGenFunction::EmitOMPParallelMasterDirective(
3068 const OMPParallelMasterDirective &S) {
3069 // Emit directive as a combined directive that consists of two implicit
3070 // directives: 'parallel' with 'master' directive.
3071 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
3072 Action.Enter(CGF);
3073 OMPPrivateScope PrivateScope(CGF);
3074 bool Copyins = CGF.EmitOMPCopyinClause(S);
3075 (void)CGF.EmitOMPFirstprivateClause(S, PrivateScope);
3076 if (Copyins) {
3077 // Emit implicit barrier to synchronize threads and avoid data races on
3078 // propagation master's thread values of threadprivate variables to local
3079 // instances of that variables of all other implicit threads.
3080 CGF.CGM.getOpenMPRuntime().emitBarrierCall(
3081 CGF, S.getBeginLoc(), OMPD_unknown, /*EmitChecks=*/false,
3082 /*ForceSimpleCall=*/true);
3083 }
3084 CGF.EmitOMPPrivateClause(S, PrivateScope);
3085 CGF.EmitOMPReductionClauseInit(S, PrivateScope);
3086 (void)PrivateScope.Privatize();
3087 emitMaster(CGF, S);
3088 CGF.EmitOMPReductionClauseFinal(S, /*ReductionKind=*/OMPD_parallel);
3089 };
3090 emitCommonOMPParallelDirective(*this, S, OMPD_master, CodeGen,
3091 emitEmptyBoundParameters);
3092 emitPostUpdateForReductionClause(*this, S,
3093 [](CodeGenFunction &) { return nullptr; });
3094}
3095
3096void CodeGenFunction::EmitOMPParallelSectionsDirective(
3097 const OMPParallelSectionsDirective &S) {
3098 // Emit directive as a combined directive that consists of two implicit
3099 // directives: 'parallel' with 'sections' directive.
3100 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
3101 Action.Enter(CGF);
3102 CGF.EmitSections(S);
3103 };
3104 emitCommonOMPParallelDirective(*this, S, OMPD_sections, CodeGen,
3105 emitEmptyBoundParameters);
3106}
3107
3108void CodeGenFunction::EmitOMPTaskBasedDirective(
3109 const OMPExecutableDirective &S, const OpenMPDirectiveKind CapturedRegion,
3110 const RegionCodeGenTy &BodyGen, const TaskGenTy &TaskGen,
3111 OMPTaskDataTy &Data) {
3112 // Emit outlined function for task construct.
3113 const CapturedStmt *CS = S.getCapturedStmt(CapturedRegion);
3114 auto I = CS->getCapturedDecl()->param_begin();
3115 auto PartId = std::next(I);
3116 auto TaskT = std::next(I, 4);
3117 // Check if the task is final
3118 if (const auto *Clause = S.getSingleClause<OMPFinalClause>()) {
3119 // If the condition constant folds and can be elided, try to avoid emitting
3120 // the condition and the dead arm of the if/else.
3121 const Expr *Cond = Clause->getCondition();
3122 bool CondConstant;
3123 if (ConstantFoldsToSimpleInteger(Cond, CondConstant))
3124 Data.Final.setInt(CondConstant);
3125 else
3126 Data.Final.setPointer(EvaluateExprAsBool(Cond));
3127 } else {
3128 // By default the task is not final.
3129 Data.Final.setInt(/*IntVal=*/false);
3130 }
3131 // Check if the task has 'priority' clause.
3132 if (const auto *Clause = S.getSingleClause<OMPPriorityClause>()) {
3133 const Expr *Prio = Clause->getPriority();
3134 Data.Priority.setInt(/*IntVal=*/true);
3135 Data.Priority.setPointer(EmitScalarConversion(
3136 EmitScalarExpr(Prio), Prio->getType(),
3137 getContext().getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/1),
3138 Prio->getExprLoc()));
3139 }
3140 // The first function argument for tasks is a thread id, the second one is a
3141 // part id (0 for tied tasks, >=0 for untied task).
3142 llvm::DenseSet<const VarDecl *> EmittedAsPrivate;
3143 // Get list of private variables.
3144 for (const auto *C : S.getClausesOfKind<OMPPrivateClause>()) {
3145 auto IRef = C->varlist_begin();
3146 for (const Expr *IInit : C->private_copies()) {
3147 const auto *OrigVD = cast<VarDecl>(cast<DeclRefExpr>(*IRef)->getDecl());
3148 if (EmittedAsPrivate.insert(OrigVD->getCanonicalDecl()).second) {
3149 Data.PrivateVars.push_back(*IRef);
3150 Data.PrivateCopies.push_back(IInit);
3151 }
3152 ++IRef;
3153 }
3154 }
3155 EmittedAsPrivate.clear();
3156 // Get list of firstprivate variables.
3157 for (const auto *C : S.getClausesOfKind<OMPFirstprivateClause>()) {
3158 auto IRef = C->varlist_begin();
3159 auto IElemInitRef = C->inits().begin();
3160 for (const Expr *IInit : C->private_copies()) {
3161 const auto *OrigVD = cast<VarDecl>(cast<DeclRefExpr>(*IRef)->getDecl());
3162 if (EmittedAsPrivate.insert(OrigVD->getCanonicalDecl()).second) {
3163 Data.FirstprivateVars.push_back(*IRef);
3164 Data.FirstprivateCopies.push_back(IInit);
3165 Data.FirstprivateInits.push_back(*IElemInitRef);
3166 }
3167 ++IRef;
3168 ++IElemInitRef;
3169 }
3170 }
3171 // Get list of lastprivate variables (for taskloops).
3172 llvm::DenseMap<const VarDecl *, const DeclRefExpr *> LastprivateDstsOrigs;
3173 for (const auto *C : S.getClausesOfKind<OMPLastprivateClause>()) {
3174 auto IRef = C->varlist_begin();
3175 auto ID = C->destination_exprs().begin();
3176 for (const Expr *IInit : C->private_copies()) {
3177 const auto *OrigVD = cast<VarDecl>(cast<DeclRefExpr>(*IRef)->getDecl());
3178 if (EmittedAsPrivate.insert(OrigVD->getCanonicalDecl()).second) {
3179 Data.LastprivateVars.push_back(*IRef);
3180 Data.LastprivateCopies.push_back(IInit);
3181 }
3182 LastprivateDstsOrigs.insert(
3183 {cast<VarDecl>(cast<DeclRefExpr>(*ID)->getDecl()),
3184 cast<DeclRefExpr>(*IRef)});
3185 ++IRef;
3186 ++ID;
3187 }
3188 }
3189 SmallVector<const Expr *, 4> LHSs;
3190 SmallVector<const Expr *, 4> RHSs;
3191 for (const auto *C : S.getClausesOfKind<OMPReductionClause>()) {
3192 auto IPriv = C->privates().begin();
3193 auto IRed = C->reduction_ops().begin();
3194 auto ILHS = C->lhs_exprs().begin();
3195 auto IRHS = C->rhs_exprs().begin();
3196 for (const Expr *Ref : C->varlists()) {
3197 Data.ReductionVars.emplace_back(Ref);
3198 Data.ReductionCopies.emplace_back(*IPriv);
3199 Data.ReductionOps.emplace_back(*IRed);
3200 LHSs.emplace_back(*ILHS);
3201 RHSs.emplace_back(*IRHS);
3202 std::advance(IPriv, 1);
3203 std::advance(IRed, 1);
3204 std::advance(ILHS, 1);
3205 std::advance(IRHS, 1);
3206 }
3207 }
3208 Data.Reductions = CGM.getOpenMPRuntime().emitTaskReductionInit(
3209 *this, S.getBeginLoc(), LHSs, RHSs, Data);
3210 // Build list of dependences.
3211 for (const auto *C : S.getClausesOfKind<OMPDependClause>())
3212 for (const Expr *IRef : C->varlists())
3213 Data.Dependences.emplace_back(C->getDependencyKind(), IRef);
3214 auto &&CodeGen = [&Data, &S, CS, &BodyGen, &LastprivateDstsOrigs,
3215 CapturedRegion](CodeGenFunction &CGF,
3216 PrePostActionTy &Action) {
3217 // Set proper addresses for generated private copies.
3218 OMPPrivateScope Scope(CGF);
3219 if (!Data.PrivateVars.empty() || !Data.FirstprivateVars.empty() ||
3220 !Data.LastprivateVars.empty()) {
3221 llvm::FunctionType *CopyFnTy = llvm::FunctionType::get(
3222 CGF.Builder.getVoidTy(), {CGF.Builder.getInt8PtrTy()}, true);
3223 enum { PrivatesParam = 2, CopyFnParam = 3 };
3224 llvm::Value *CopyFn = CGF.Builder.CreateLoad(
3225 CGF.GetAddrOfLocalVar(CS->getCapturedDecl()->getParam(CopyFnParam)));
3226 llvm::Value *PrivatesPtr = CGF.Builder.CreateLoad(CGF.GetAddrOfLocalVar(
3227 CS->getCapturedDecl()->getParam(PrivatesParam)));
3228 // Map privates.
3229 llvm::SmallVector<std::pair<const VarDecl *, Address>, 16> PrivatePtrs;
3230 llvm::SmallVector<llvm::Value *, 16> CallArgs;
3231 CallArgs.push_back(PrivatesPtr);
3232 for (const Expr *E : Data.PrivateVars) {
3233 const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl());
3234 Address PrivatePtr = CGF.CreateMemTemp(
3235 CGF.getContext().getPointerType(E->getType()), ".priv.ptr.addr");
3236 PrivatePtrs.emplace_back(VD, PrivatePtr);
3237 CallArgs.push_back(PrivatePtr.getPointer());
3238 }
3239 for (const Expr *E : Data.FirstprivateVars) {
3240 const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl());
3241 Address PrivatePtr =
3242 CGF.CreateMemTemp(CGF.getContext().getPointerType(E->getType()),
3243 ".firstpriv.ptr.addr");
3244 PrivatePtrs.emplace_back(VD, PrivatePtr);
3245 CallArgs.push_back(PrivatePtr.getPointer());
3246 }
3247 for (const Expr *E : Data.LastprivateVars) {
3248 const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl());
3249 Address PrivatePtr =
3250 CGF.CreateMemTemp(CGF.getContext().getPointerType(E->getType()),
3251 ".lastpriv.ptr.addr");
3252 PrivatePtrs.emplace_back(VD, PrivatePtr);
3253 CallArgs.push_back(PrivatePtr.getPointer());
3254 }
3255 CGF.CGM.getOpenMPRuntime().emitOutlinedFunctionCall(
3256 CGF, S.getBeginLoc(), {CopyFnTy, CopyFn}, CallArgs);
3257 for (const auto &Pair : LastprivateDstsOrigs) {
3258 const auto *OrigVD = cast<VarDecl>(Pair.second->getDecl());
3259 DeclRefExpr DRE(CGF.getContext(), const_cast<VarDecl *>(OrigVD),
3260 /*RefersToEnclosingVariableOrCapture=*/
3261 CGF.CapturedStmtInfo->lookup(OrigVD) != nullptr,
3262 Pair.second->getType(), VK_LValue,
3263 Pair.second->getExprLoc());
3264 Scope.addPrivate(Pair.first, [&CGF, &DRE]() {
3265 return CGF.EmitLValue(&DRE).getAddress(CGF);
3266 });
3267 }
3268 for (const auto &Pair : PrivatePtrs) {
3269 Address Replacement(CGF.Builder.CreateLoad(Pair.second),
3270 CGF.getContext().getDeclAlign(Pair.first));
3271 Scope.addPrivate(Pair.first, [Replacement]() { return Replacement; });
3272 }
3273 }
3274 if (Data.Reductions) {
3275 OMPLexicalScope LexScope(CGF, S, CapturedRegion);
3276 ReductionCodeGen RedCG(Data.ReductionVars, Data.ReductionCopies,
3277 Data.ReductionOps);
3278 llvm::Value *ReductionsPtr = CGF.Builder.CreateLoad(
3279 CGF.GetAddrOfLocalVar(CS->getCapturedDecl()->getParam(9)));
3280 for (unsigned Cnt = 0, E = Data.ReductionVars.size(); Cnt < E; ++Cnt) {
3281 RedCG.emitSharedLValue(CGF, Cnt);
3282 RedCG.emitAggregateType(CGF, Cnt);
3283 // FIXME: This must removed once the runtime library is fixed.
3284 // Emit required threadprivate variables for
3285 // initializer/combiner/finalizer.
3286 CGF.CGM.getOpenMPRuntime().emitTaskReductionFixups(CGF, S.getBeginLoc(),
3287 RedCG, Cnt);
3288 Address Replacement = CGF.CGM.getOpenMPRuntime().getTaskReductionItem(
3289 CGF, S.getBeginLoc(), ReductionsPtr, RedCG.getSharedLValue(Cnt));
3290 Replacement =
3291 Address(CGF.EmitScalarConversion(
3292 Replacement.getPointer(), CGF.getContext().VoidPtrTy,
3293 CGF.getContext().getPointerType(
3294 Data.ReductionCopies[Cnt]->getType()),
3295 Data.ReductionCopies[Cnt]->getExprLoc()),
3296 Replacement.getAlignment());
3297 Replacement = RedCG.adjustPrivateAddress(CGF, Cnt, Replacement);
3298 Scope.addPrivate(RedCG.getBaseDecl(Cnt),
3299 [Replacement]() { return Replacement; });
3300 }
3301 }
3302 // Privatize all private variables except for in_reduction items.
3303 (void)Scope.Privatize();
3304 SmallVector<const Expr *, 4> InRedVars;
3305 SmallVector<const Expr *, 4> InRedPrivs;
3306 SmallVector<const Expr *, 4> InRedOps;
3307 SmallVector<const Expr *, 4> TaskgroupDescriptors;
3308 for (const auto *C : S.getClausesOfKind<OMPInReductionClause>()) {
3309 auto IPriv = C->privates().begin();
3310 auto IRed = C->reduction_ops().begin();
3311 auto ITD = C->taskgroup_descriptors().begin();
3312 for (const Expr *Ref : C->varlists()) {
3313 InRedVars.emplace_back(Ref);
3314 InRedPrivs.emplace_back(*IPriv);
3315 InRedOps.emplace_back(*IRed);
3316 TaskgroupDescriptors.emplace_back(*ITD);
3317 std::advance(IPriv, 1);
3318 std::advance(IRed, 1);
3319 std::advance(ITD, 1);
3320 }
3321 }
3322 // Privatize in_reduction items here, because taskgroup descriptors must be
3323 // privatized earlier.
3324 OMPPrivateScope InRedScope(CGF);
3325 if (!InRedVars.empty()) {
3326 ReductionCodeGen RedCG(InRedVars, InRedPrivs, InRedOps);
3327 for (unsigned Cnt = 0, E = InRedVars.size(); Cnt < E; ++Cnt) {
3328 RedCG.emitSharedLValue(CGF, Cnt);
3329 RedCG.emitAggregateType(CGF, Cnt);
3330 // The taskgroup descriptor variable is always implicit firstprivate and
3331 // privatized already during processing of the firstprivates.
3332 // FIXME: This must removed once the runtime library is fixed.
3333 // Emit required threadprivate variables for
3334 // initializer/combiner/finalizer.
3335 CGF.CGM.getOpenMPRuntime().emitTaskReductionFixups(CGF, S.getBeginLoc(),
3336 RedCG, Cnt);
3337 llvm::Value *ReductionsPtr =
3338 CGF.EmitLoadOfScalar(CGF.EmitLValue(TaskgroupDescriptors[Cnt]),
3339 TaskgroupDescriptors[Cnt]->getExprLoc());
3340 Address Replacement = CGF.CGM.getOpenMPRuntime().getTaskReductionItem(
3341 CGF, S.getBeginLoc(), ReductionsPtr, RedCG.getSharedLValue(Cnt));
3342 Replacement = Address(
3343 CGF.EmitScalarConversion(
3344 Replacement.getPointer(), CGF.getContext().VoidPtrTy,
3345 CGF.getContext().getPointerType(InRedPrivs[Cnt]->getType()),
3346 InRedPrivs[Cnt]->getExprLoc()),
3347 Replacement.getAlignment());
3348 Replacement = RedCG.adjustPrivateAddress(CGF, Cnt, Replacement);
3349 InRedScope.addPrivate(RedCG.getBaseDecl(Cnt),
3350 [Replacement]() { return Replacement; });
3351 }
3352 }
3353 (void)InRedScope.Privatize();
3354
3355 Action.Enter(CGF);
3356 BodyGen(CGF);
3357 };
3358 llvm::Function *OutlinedFn = CGM.getOpenMPRuntime().emitTaskOutlinedFunction(
3359 S, *I, *PartId, *TaskT, S.getDirectiveKind(), CodeGen, Data.Tied,
3360 Data.NumberOfParts);
3361 OMPLexicalScope Scope(*this, S, llvm::None,
3362 !isOpenMPParallelDirective(S.getDirectiveKind()) &&
3363 !isOpenMPSimdDirective(S.getDirectiveKind()));
3364 TaskGen(*this, OutlinedFn, Data);
3365}
3366
3367static ImplicitParamDecl *
3368createImplicitFirstprivateForType(ASTContext &C, OMPTaskDataTy &Data,
3369 QualType Ty, CapturedDecl *CD,
3370 SourceLocation Loc) {
3371 auto *OrigVD = ImplicitParamDecl::Create(C, CD, Loc, /*Id=*/nullptr, Ty,
3372 ImplicitParamDecl::Other);
3373 auto *OrigRef = DeclRefExpr::Create(
3374 C, NestedNameSpecifierLoc(), SourceLocation(), OrigVD,
3375 /*RefersToEnclosingVariableOrCapture=*/false, Loc, Ty, VK_LValue);
3376 auto *PrivateVD = ImplicitParamDecl::Create(C, CD, Loc, /*Id=*/nullptr, Ty,
3377 ImplicitParamDecl::Other);
3378 auto *PrivateRef = DeclRefExpr::Create(
3379 C, NestedNameSpecifierLoc(), SourceLocation(), PrivateVD,
3380 /*RefersToEnclosingVariableOrCapture=*/false, Loc, Ty, VK_LValue);
3381 QualType ElemType = C.getBaseElementType(Ty);
3382 auto *InitVD = ImplicitParamDecl::Create(C, CD, Loc, /*Id=*/nullptr, ElemType,
3383 ImplicitParamDecl::Other);
3384 auto *InitRef = DeclRefExpr::Create(
3385 C, NestedNameSpecifierLoc(), SourceLocation(), InitVD,
3386 /*RefersToEnclosingVariableOrCapture=*/false, Loc, ElemType, VK_LValue);
3387 PrivateVD->setInitStyle(VarDecl::CInit);
3388 PrivateVD->setInit(ImplicitCastExpr::Create(C, ElemType, CK_LValueToRValue,
3389 InitRef, /*BasePath=*/nullptr,
3390 VK_RValue));
3391 Data.FirstprivateVars.emplace_back(OrigRef);
3392 Data.FirstprivateCopies.emplace_back(PrivateRef);
3393 Data.FirstprivateInits.emplace_back(InitRef);
3394 return OrigVD;
3395}
3396
3397void CodeGenFunction::EmitOMPTargetTaskBasedDirective(
3398 const OMPExecutableDirective &S, const RegionCodeGenTy &BodyGen,
3399 OMPTargetDataInfo &InputInfo) {
3400 // Emit outlined function for task construct.
3401 const CapturedStmt *CS = S.getCapturedStmt(OMPD_task);
3402 Address CapturedStruct = GenerateCapturedStmtArgument(*CS);
3403 QualType SharedsTy = getContext().getRecordType(CS->getCapturedRecordDecl());
3404 auto I = CS->getCapturedDecl()->param_begin();
3405 auto PartId = std::next(I);
3406 auto TaskT = std::next(I, 4);
3407 OMPTaskDataTy Data;
3408 // The task is not final.
3409 Data.Final.setInt(/*IntVal=*/false);
3410 // Get list of firstprivate variables.
3411 for (const auto *C : S.getClausesOfKind<OMPFirstprivateClause>()) {
3412 auto IRef = C->varlist_begin();
3413 auto IElemInitRef = C->inits().begin();
3414 for (auto *IInit : C->private_copies()) {
3415 Data.FirstprivateVars.push_back(*IRef);
3416 Data.FirstprivateCopies.push_back(IInit);
3417 Data.FirstprivateInits.push_back(*IElemInitRef);
3418 ++IRef;
3419 ++IElemInitRef;
3420 }
3421 }
3422 OMPPrivateScope TargetScope(*this);
3423 VarDecl *BPVD = nullptr;
3424 VarDecl *PVD = nullptr;
3425 VarDecl *SVD = nullptr;
3426 if (InputInfo.NumberOfTargetItems > 0) {
3427 auto *CD = CapturedDecl::Create(
3428 getContext(), getContext().getTranslationUnitDecl(), /*NumParams=*/0);
3429 llvm::APInt ArrSize(/*numBits=*/32, InputInfo.NumberOfTargetItems);
3430 QualType BaseAndPointersType = getContext().getConstantArrayType(
3431 getContext().VoidPtrTy, ArrSize, nullptr, ArrayType::Normal,
3432 /*IndexTypeQuals=*/0);
3433 BPVD = createImplicitFirstprivateForType(
3434 getContext(), Data, BaseAndPointersType, CD, S.getBeginLoc());
3435 PVD = createImplicitFirstprivateForType(
3436 getContext(), Data, BaseAndPointersType, CD, S.getBeginLoc());
3437 QualType SizesType = getContext().getConstantArrayType(
3438 getContext().getIntTypeForBitwidth(/*DestWidth=*/64, /*Signed=*/1),
3439 ArrSize, nullptr, ArrayType::Normal,
3440 /*IndexTypeQuals=*/0);
3441 SVD = createImplicitFirstprivateForType(getContext(), Data, SizesType, CD,
3442 S.getBeginLoc());
3443 TargetScope.addPrivate(
3444 BPVD, [&InputInfo]() { return InputInfo.BasePointersArray; });
3445 TargetScope.addPrivate(PVD,
3446 [&InputInfo]() { return InputInfo.PointersArray; });
3447 TargetScope.addPrivate(SVD,
3448 [&InputInfo]() { return InputInfo.SizesArray; });
3449 }
3450 (void)TargetScope.Privatize();
3451 // Build list of dependences.
3452 for (const auto *C : S.getClausesOfKind<OMPDependClause>())
3453 for (const Expr *IRef : C->varlists())
3454 Data.Dependences.emplace_back(C->getDependencyKind(), IRef);
3455 auto &&CodeGen = [&Data, &S, CS, &BodyGen, BPVD, PVD, SVD,
3456 &InputInfo](CodeGenFunction &CGF, PrePostActionTy &Action) {
3457 // Set proper addresses for generated private copies.
3458 OMPPrivateScope Scope(CGF);
3459 if (!Data.FirstprivateVars.empty()) {
3460 llvm::FunctionType *CopyFnTy = llvm::FunctionType::get(
3461 CGF.Builder.getVoidTy(), {CGF.Builder.getInt8PtrTy()}, true);
3462 enum { PrivatesParam = 2, CopyFnParam = 3 };
3463 llvm::Value *CopyFn = CGF.Builder.CreateLoad(
3464 CGF.GetAddrOfLocalVar(CS->getCapturedDecl()->getParam(CopyFnParam)));
3465 llvm::Value *PrivatesPtr = CGF.Builder.CreateLoad(CGF.GetAddrOfLocalVar(
3466 CS->getCapturedDecl()->getParam(PrivatesParam)));
3467 // Map privates.
3468 llvm::SmallVector<std::pair<const VarDecl *, Address>, 16> PrivatePtrs;
3469 llvm::SmallVector<llvm::Value *, 16> CallArgs;
3470 CallArgs.push_back(PrivatesPtr);
3471 for (const Expr *E : Data.FirstprivateVars) {
3472 const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl());
3473 Address PrivatePtr =
3474 CGF.CreateMemTemp(CGF.getContext().getPointerType(E->getType()),
3475 ".firstpriv.ptr.addr");
3476 PrivatePtrs.emplace_back(VD, PrivatePtr);
3477 CallArgs.push_back(PrivatePtr.getPointer());
3478 }
3479 CGF.CGM.getOpenMPRuntime().emitOutlinedFunctionCall(
3480 CGF, S.getBeginLoc(), {CopyFnTy, CopyFn}, CallArgs);
3481 for (const auto &Pair : PrivatePtrs) {
3482 Address Replacement(CGF.Builder.CreateLoad(Pair.second),
3483 CGF.getContext().getDeclAlign(Pair.first));
3484 Scope.addPrivate(Pair.first, [Replacement]() { return Replacement; });
3485 }
3486 }
3487 // Privatize all private variables except for in_reduction items.
3488 (void)Scope.Privatize();
3489 if (InputInfo.NumberOfTargetItems > 0) {
3490 InputInfo.BasePointersArray = CGF.Builder.CreateConstArrayGEP(
3491 CGF.GetAddrOfLocalVar(BPVD), /*Index=*/0);
3492 InputInfo.PointersArray = CGF.Builder.CreateConstArrayGEP(
3493 CGF.GetAddrOfLocalVar(PVD), /*Index=*/0);
3494 InputInfo.SizesArray = CGF.Builder.CreateConstArrayGEP(
3495 CGF.GetAddrOfLocalVar(SVD), /*Index=*/0);
3496 }
3497
3498 Action.Enter(CGF);
3499 OMPLexicalScope LexScope(CGF, S, OMPD_task, /*EmitPreInitStmt=*/false);
3500 BodyGen(CGF);
3501 };
3502 llvm::Function *OutlinedFn = CGM.getOpenMPRuntime().emitTaskOutlinedFunction(
3503 S, *I, *PartId, *TaskT, S.getDirectiveKind(), CodeGen, /*Tied=*/true,
3504 Data.NumberOfParts);
3505 llvm::APInt TrueOrFalse(32, S.hasClausesOfKind<OMPNowaitClause>() ? 1 : 0);
3506 IntegerLiteral IfCond(getContext(), TrueOrFalse,
3507 getContext().getIntTypeForBitwidth(32, /*Signed=*/0),
3508 SourceLocation());
3509
3510 CGM.getOpenMPRuntime().emitTaskCall(*this, S.getBeginLoc(), S, OutlinedFn,
3511 SharedsTy, CapturedStruct, &IfCond, Data);
3512}
3513
3514void CodeGenFunction::EmitOMPTaskDirective(const OMPTaskDirective &S) {
3515 // Emit outlined function for task construct.
3516 const CapturedStmt *CS = S.getCapturedStmt(OMPD_task);
3517 Address CapturedStruct = GenerateCapturedStmtArgument(*CS);
3518 QualType SharedsTy = getContext().getRecordType(CS->getCapturedRecordDecl());
3519 const Expr *IfCond = nullptr;
3520 for (const auto *C : S.getClausesOfKind<OMPIfClause>()) {
3521 if (C->getNameModifier() == OMPD_unknown ||
3522 C->getNameModifier() == OMPD_task) {
3523 IfCond = C->getCondition();
3524 break;
3525 }
3526 }
3527
3528 OMPTaskDataTy Data;
3529 // Check if we should emit tied or untied task.
3530 Data.Tied = !S.getSingleClause<OMPUntiedClause>();
3531 auto &&BodyGen = [CS](CodeGenFunction &CGF, PrePostActionTy &) {
3532 CGF.EmitStmt(CS->getCapturedStmt());
3533 };
3534 auto &&TaskGen = [&S, SharedsTy, CapturedStruct,
3535 IfCond](CodeGenFunction &CGF, llvm::Function *OutlinedFn,
3536 const OMPTaskDataTy &Data) {
3537 CGF.CGM.getOpenMPRuntime().emitTaskCall(CGF, S.getBeginLoc(), S, OutlinedFn,
3538 SharedsTy, CapturedStruct, IfCond,
3539 Data);
3540 };
3541 EmitOMPTaskBasedDirective(S, OMPD_task, BodyGen, TaskGen, Data);
3542}
3543
3544void CodeGenFunction::EmitOMPTaskyieldDirective(
3545 const OMPTaskyieldDirective &S) {
3546 CGM.getOpenMPRuntime().emitTaskyieldCall(*this, S.getBeginLoc());
3547}
3548
3549void CodeGenFunction::EmitOMPBarrierDirective(const OMPBarrierDirective &S) {
3550 CGM.getOpenMPRuntime().emitBarrierCall(*this, S.getBeginLoc(), OMPD_barrier);
3551}
3552
3553void CodeGenFunction::EmitOMPTaskwaitDirective(const OMPTaskwaitDirective &S) {
3554 CGM.getOpenMPRuntime().emitTaskwaitCall(*this, S.getBeginLoc());
3555}
3556
3557void CodeGenFunction::EmitOMPTaskgroupDirective(
3558 const OMPTaskgroupDirective &S) {
3559 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
3560 Action.Enter(CGF);
3561 if (const Expr *E = S.getReductionRef()) {
3562 SmallVector<const Expr *, 4> LHSs;
3563 SmallVector<const Expr *, 4> RHSs;
3564 OMPTaskDataTy Data;
3565 for (const auto *C : S.getClausesOfKind<OMPTaskReductionClause>()) {
3566 auto IPriv = C->privates().begin();
3567 auto IRed = C->reduction_ops().begin();
3568 auto ILHS = C->lhs_exprs().begin();
3569 auto IRHS = C->rhs_exprs().begin();
3570 for (const Expr *Ref : C->varlists()) {
3571 Data.ReductionVars.emplace_back(Ref);
3572 Data.ReductionCopies.emplace_back(*IPriv);
3573 Data.ReductionOps.emplace_back(*IRed);
3574 LHSs.emplace_back(*ILHS);
3575 RHSs.emplace_back(*IRHS);
3576 std::advance(IPriv, 1);
3577 std::advance(IRed, 1);
3578 std::advance(ILHS, 1);
3579 std::advance(IRHS, 1);
3580 }
3581 }
3582 llvm::Value *ReductionDesc =
3583 CGF.CGM.getOpenMPRuntime().emitTaskReductionInit(CGF, S.getBeginLoc(),
3584 LHSs, RHSs, Data);
3585 const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl());
3586 CGF.EmitVarDecl(*VD);
3587 CGF.EmitStoreOfScalar(ReductionDesc, CGF.GetAddrOfLocalVar(VD),
3588 /*Volatile=*/false, E->getType());
3589 }
3590 CGF.EmitStmt(S.getInnermostCapturedStmt()->getCapturedStmt());
3591 };
3592 OMPLexicalScope Scope(*this, S, OMPD_unknown);
3593 CGM.getOpenMPRuntime().emitTaskgroupRegion(*this, CodeGen, S.getBeginLoc());
3594}
3595
3596void CodeGenFunction::EmitOMPFlushDirective(const OMPFlushDirective &S) {
3597 CGM.getOpenMPRuntime().emitFlush(
3598 *this,
3599 [&S]() -> ArrayRef<const Expr *> {
3600 if (const auto *FlushClause = S.getSingleClause<OMPFlushClause>())
3601 return llvm::makeArrayRef(FlushClause->varlist_begin(),
3602 FlushClause->varlist_end());
3603 return llvm::None;
3604 }(),
3605 S.getBeginLoc());
3606}
3607
3608void CodeGenFunction::EmitOMPDistributeLoop(const OMPLoopDirective &S,
3609 const CodeGenLoopTy &CodeGenLoop,
3610 Expr *IncExpr) {
3611 // Emit the loop iteration variable.
3612 const auto *IVExpr = cast<DeclRefExpr>(S.getIterationVariable());
3613 const auto *IVDecl = cast<VarDecl>(IVExpr->getDecl());
3614 EmitVarDecl(*IVDecl);
3615
3616 // Emit the iterations count variable.
3617 // If it is not a variable, Sema decided to calculate iterations count on each
3618 // iteration (e.g., it is foldable into a constant).
3619 if (const auto *LIExpr = dyn_cast<DeclRefExpr>(S.getLastIteration())) {
3620 EmitVarDecl(*cast<VarDecl>(LIExpr->getDecl()));
3621 // Emit calculation of the iterations count.
3622 EmitIgnoredExpr(S.getCalcLastIteration());
3623 }
3624
3625 CGOpenMPRuntime &RT = CGM.getOpenMPRuntime();
3626
3627 bool HasLastprivateClause = false;
3628 // Check pre-condition.
3629 {
3630 OMPLoopScope PreInitScope(*this, S);
3631 // Skip the entire loop if we don't meet the precondition.
3632 // If the condition constant folds and can be elided, avoid emitting the
3633 // whole loop.
3634 bool CondConstant;
3635 llvm::BasicBlock *ContBlock = nullptr;
3636 if (ConstantFoldsToSimpleInteger(S.getPreCond(), CondConstant)) {
3637 if (!CondConstant)
3638 return;
3639 } else {
3640 llvm::BasicBlock *ThenBlock = createBasicBlock("omp.precond.then");
3641 ContBlock = createBasicBlock("omp.precond.end");
3642 emitPreCond(*this, S, S.getPreCond(), ThenBlock, ContBlock,
3643 getProfileCount(&S));
3644 EmitBlock(ThenBlock);
3645 incrementProfileCounter(&S);
3646 }
3647
3648 emitAlignedClause(*this, S);
3649 // Emit 'then' code.
3650 {
3651 // Emit helper vars inits.
3652
3653 LValue LB = EmitOMPHelperVar(
3654 *this, cast<DeclRefExpr>(
3655 (isOpenMPLoopBoundSharingDirective(S.getDirectiveKind())
3656 ? S.getCombinedLowerBoundVariable()
3657 : S.getLowerBoundVariable())));
3658 LValue UB = EmitOMPHelperVar(
3659 *this, cast<DeclRefExpr>(
3660 (isOpenMPLoopBoundSharingDirective(S.getDirectiveKind())
3661 ? S.getCombinedUpperBoundVariable()
3662 : S.getUpperBoundVariable())));
3663 LValue ST =
3664 EmitOMPHelperVar(*this, cast<DeclRefExpr>(S.getStrideVariable()));
3665 LValue IL =
3666 EmitOMPHelperVar(*this, cast<DeclRefExpr>(S.getIsLastIterVariable()));
3667
3668 OMPPrivateScope LoopScope(*this);
3669 if (EmitOMPFirstprivateClause(S, LoopScope)) {
3670 // Emit implicit barrier to synchronize threads and avoid data races
3671 // on initialization of firstprivate variables and post-update of
3672 // lastprivate variables.
3673 CGM.getOpenMPRuntime().emitBarrierCall(
3674 *this, S.getBeginLoc(), OMPD_unknown, /*EmitChecks=*/false,
3675 /*ForceSimpleCall=*/true);
3676 }
3677 EmitOMPPrivateClause(S, LoopScope);
3678 if (isOpenMPSimdDirective(S.getDirectiveKind()) &&
3679 !isOpenMPParallelDirective(S.getDirectiveKind()) &&
3680 !isOpenMPTeamsDirective(S.getDirectiveKind()))
3681 EmitOMPReductionClauseInit(S, LoopScope);
3682 HasLastprivateClause = EmitOMPLastprivateClauseInit(S, LoopScope);
3683 EmitOMPPrivateLoopCounters(S, LoopScope);
3684 (void)LoopScope.Privatize();
3685 if (isOpenMPTargetExecutionDirective(S.getDirectiveKind()))
3686 CGM.getOpenMPRuntime().adjustTargetSpecificDataForLambdas(*this, S);
3687
3688 // Detect the distribute schedule kind and chunk.
3689 llvm::Value *Chunk = nullptr;
3690 OpenMPDistScheduleClauseKind ScheduleKind = OMPC_DIST_SCHEDULE_unknown;
3691 if (const auto *C = S.getSingleClause<OMPDistScheduleClause>()) {
3692 ScheduleKind = C->getDistScheduleKind();
3693 if (const Expr *Ch = C->getChunkSize()) {
3694 Chunk = EmitScalarExpr(Ch);
3695 Chunk = EmitScalarConversion(Chunk, Ch->getType(),
3696 S.getIterationVariable()->getType(),
3697 S.getBeginLoc());
3698 }
3699 } else {
3700 // Default behaviour for dist_schedule clause.
3701 CGM.getOpenMPRuntime().getDefaultDistScheduleAndChunk(
3702 *this, S, ScheduleKind, Chunk);
3703 }
3704 const unsigned IVSize = getContext().getTypeSize(IVExpr->getType());
3705 const bool IVSigned = IVExpr->getType()->hasSignedIntegerRepresentation();
3706
3707 // OpenMP [2.10.8, distribute Construct, Description]
3708 // If dist_schedule is specified, kind must be static. If specified,
3709 // iterations are divided into chunks of size chunk_size, chunks are
3710 // assigned to the teams of the league in a round-robin fashion in the
3711 // order of the team number. When no chunk_size is specified, the
3712 // iteration space is divided into chunks that are approximately equal
3713 // in size, and at most one chunk is distributed to each team of the
3714 // league. The size of the chunks is unspecified in this case.
3715 bool StaticChunked = RT.isStaticChunked(
3716 ScheduleKind, /* Chunked */ Chunk != nullptr) &&
3717 isOpenMPLoopBoundSharingDirective(S.getDirectiveKind());
3718 if (RT.isStaticNonchunked(ScheduleKind,
3719 /* Chunked */ Chunk != nullptr) ||
3720 StaticChunked) {
3721 CGOpenMPRuntime::StaticRTInput StaticInit(
3722 IVSize, IVSigned, /* Ordered = */ false, IL.getAddress(*this),
3723 LB.getAddress(*this), UB.getAddress(*this), ST.getAddress(*this),
3724 StaticChunked ? Chunk : nullptr);
3725 RT.emitDistributeStaticInit(*this, S.getBeginLoc(), ScheduleKind,
3726 StaticInit);
3727 JumpDest LoopExit =
3728 getJumpDestInCurrentScope(createBasicBlock("omp.loop.exit"));
3729 // UB = min(UB, GlobalUB);
3730 EmitIgnoredExpr(isOpenMPLoopBoundSharingDirective(S.getDirectiveKind())
3731 ? S.getCombinedEnsureUpperBound()
3732 : S.getEnsureUpperBound());
3733 // IV = LB;
3734 EmitIgnoredExpr(isOpenMPLoopBoundSharingDirective(S.getDirectiveKind())
3735 ? S.getCombinedInit()
3736 : S.getInit());
3737
3738 const Expr *Cond =
3739 isOpenMPLoopBoundSharingDirective(S.getDirectiveKind())
3740 ? S.getCombinedCond()
3741 : S.getCond();
3742
3743 if (StaticChunked)
3744 Cond = S.getCombinedDistCond();
3745
3746 // For static unchunked schedules generate:
3747 //
3748 // 1. For distribute alone, codegen
3749 // while (idx <= UB) {
3750 // BODY;
3751 // ++idx;
3752 // }
3753 //
3754 // 2. When combined with 'for' (e.g. as in 'distribute parallel for')
3755 // while (idx <= UB) {
3756 // <CodeGen rest of pragma>(LB, UB);
3757 // idx += ST;
3758 // }
3759 //
3760 // For static chunk one schedule generate:
3761 //
3762 // while (IV <= GlobalUB) {
3763 // <CodeGen rest of pragma>(LB, UB);
3764 // LB += ST;
3765 // UB += ST;
3766 // UB = min(UB, GlobalUB);
3767 // IV = LB;
3768 // }
3769 //
3770 emitCommonSimdLoop(
3771 *this, S,
3772 [&S](CodeGenFunction &CGF, PrePostActionTy &) {
3773 if (isOpenMPSimdDirective(S.getDirectiveKind()))
3774 CGF.EmitOMPSimdInit(S, /*IsMonotonic=*/true);
3775 },
3776 [&S, &LoopScope, Cond, IncExpr, LoopExit, &CodeGenLoop,
3777 StaticChunked](CodeGenFunction &CGF, PrePostActionTy &) {
3778 CGF.EmitOMPInnerLoop(
3779 S, LoopScope.requiresCleanups(), Cond, IncExpr,
3780 [&S, LoopExit, &CodeGenLoop](CodeGenFunction &CGF) {
3781 CodeGenLoop(CGF, S, LoopExit);
3782 },
3783 [&S, StaticChunked](CodeGenFunction &CGF) {
3784 if (StaticChunked) {
3785 CGF.EmitIgnoredExpr(S.getCombinedNextLowerBound());
3786 CGF.EmitIgnoredExpr(S.getCombinedNextUpperBound());
3787 CGF.EmitIgnoredExpr(S.getCombinedEnsureUpperBound());
3788 CGF.EmitIgnoredExpr(S.getCombinedInit());
3789 }
3790 });
3791 });
3792 EmitBlock(LoopExit.getBlock());
3793 // Tell the runtime we are done.
3794 RT.emitForStaticFinish(*this, S.getBeginLoc(), S.getDirectiveKind());
3795 } else {
3796 // Emit the outer loop, which requests its work chunk [LB..UB] from
3797 // runtime and runs the inner loop to process it.
3798 const OMPLoopArguments LoopArguments = {
3799 LB.getAddress(*this), UB.getAddress(*this), ST.getAddress(*this),
3800 IL.getAddress(*this), Chunk};
3801 EmitOMPDistributeOuterLoop(ScheduleKind, S, LoopScope, LoopArguments,
3802 CodeGenLoop);
3803 }
3804 if (isOpenMPSimdDirective(S.getDirectiveKind())) {
3805 EmitOMPSimdFinal(S, [IL, &S](CodeGenFunction &CGF) {
3806 return CGF.Builder.CreateIsNotNull(
3807 CGF.EmitLoadOfScalar(IL, S.getBeginLoc()));
3808 });
3809 }
3810 if (isOpenMPSimdDirective(S.getDirectiveKind()) &&
3811 !isOpenMPParallelDirective(S.getDirectiveKind()) &&
3812 !isOpenMPTeamsDirective(S.getDirectiveKind())) {
3813 EmitOMPReductionClauseFinal(S, OMPD_simd);
3814 // Emit post-update of the reduction variables if IsLastIter != 0.
3815 emitPostUpdateForReductionClause(
3816 *this, S, [IL, &S](CodeGenFunction &CGF) {
3817 return CGF.Builder.CreateIsNotNull(
3818 CGF.EmitLoadOfScalar(IL, S.getBeginLoc()));
3819 });
3820 }
3821 // Emit final copy of the lastprivate variables if IsLastIter != 0.
3822 if (HasLastprivateClause) {
3823 EmitOMPLastprivateClauseFinal(
3824 S, /*NoFinals=*/false,
3825 Builder.CreateIsNotNull(EmitLoadOfScalar(IL, S.getBeginLoc())));
3826 }
3827 }
3828
3829 // We're now done with the loop, so jump to the continuation block.
3830 if (ContBlock) {
3831 EmitBranch(ContBlock);
3832 EmitBlock(ContBlock, true);
3833 }
3834 }
3835}
3836
3837void CodeGenFunction::EmitOMPDistributeDirective(
3838 const OMPDistributeDirective &S) {
3839 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &) {
3840 CGF.EmitOMPDistributeLoop(S, emitOMPLoopBodyWithStopPoint, S.getInc());
3841 };
3842 OMPLexicalScope Scope(*this, S, OMPD_unknown);
3843 CGM.getOpenMPRuntime().emitInlinedDirective(*this, OMPD_distribute, CodeGen);
3844}
3845
3846static llvm::Function *emitOutlinedOrderedFunction(CodeGenModule &CGM,
3847 const CapturedStmt *S) {
3848 CodeGenFunction CGF(CGM, /*suppressNewContext=*/true);
3849 CodeGenFunction::CGCapturedStmtInfo CapStmtInfo;
3850 CGF.CapturedStmtInfo = &CapStmtInfo;
3851 llvm::Function *Fn = CGF.GenerateOpenMPCapturedStmtFunction(*S);
3852 Fn->setDoesNotRecurse();
3853 return Fn;
3854}
3855
3856void CodeGenFunction::EmitOMPOrderedDirective(const OMPOrderedDirective &S) {
3857 if (S.hasClausesOfKind<OMPDependClause>()) {
3858 assert(!S.getAssociatedStmt() &&((!S.getAssociatedStmt() && "No associated statement must be in ordered depend construct."
) ? static_cast<void> (0) : __assert_fail ("!S.getAssociatedStmt() && \"No associated statement must be in ordered depend construct.\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 3859, __PRETTY_FUNCTION__))
3859 "No associated statement must be in ordered depend construct.")((!S.getAssociatedStmt() && "No associated statement must be in ordered depend construct."
) ? static_cast<void> (0) : __assert_fail ("!S.getAssociatedStmt() && \"No associated statement must be in ordered depend construct.\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 3859, __PRETTY_FUNCTION__))
;
3860 for (const auto *DC : S.getClausesOfKind<OMPDependClause>())
3861 CGM.getOpenMPRuntime().emitDoacrossOrdered(*this, DC);
3862 return;
3863 }
3864 const auto *C = S.getSingleClause<OMPSIMDClause>();
3865 auto &&CodeGen = [&S, C, this](CodeGenFunction &CGF,
3866 PrePostActionTy &Action) {
3867 const CapturedStmt *CS = S.getInnermostCapturedStmt();
3868 if (C) {
3869 llvm::SmallVector<llvm::Value *, 16> CapturedVars;
3870 CGF.GenerateOpenMPCapturedVars(*CS, CapturedVars);
3871 llvm::Function *OutlinedFn = emitOutlinedOrderedFunction(CGM, CS);
3872 CGM.getOpenMPRuntime().emitOutlinedFunctionCall(CGF, S.getBeginLoc(),
3873 OutlinedFn, CapturedVars);
3874 } else {
3875 Action.Enter(CGF);
3876 CGF.EmitStmt(CS->getCapturedStmt());
3877 }
3878 };
3879 OMPLexicalScope Scope(*this, S, OMPD_unknown);
3880 CGM.getOpenMPRuntime().emitOrderedRegion(*this, CodeGen, S.getBeginLoc(), !C);
3881}
3882
3883static llvm::Value *convertToScalarValue(CodeGenFunction &CGF, RValue Val,
3884 QualType SrcType, QualType DestType,
3885 SourceLocation Loc) {
3886 assert(CGF.hasScalarEvaluationKind(DestType) &&((CGF.hasScalarEvaluationKind(DestType) && "DestType must have scalar evaluation kind."
) ? static_cast<void> (0) : __assert_fail ("CGF.hasScalarEvaluationKind(DestType) && \"DestType must have scalar evaluation kind.\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 3887, __PRETTY_FUNCTION__))
3887 "DestType must have scalar evaluation kind.")((CGF.hasScalarEvaluationKind(DestType) && "DestType must have scalar evaluation kind."
) ? static_cast<void> (0) : __assert_fail ("CGF.hasScalarEvaluationKind(DestType) && \"DestType must have scalar evaluation kind.\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 3887, __PRETTY_FUNCTION__))
;
3888 assert(!Val.isAggregate() && "Must be a scalar or complex.")((!Val.isAggregate() && "Must be a scalar or complex."
) ? static_cast<void> (0) : __assert_fail ("!Val.isAggregate() && \"Must be a scalar or complex.\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 3888, __PRETTY_FUNCTION__))
;
3889 return Val.isScalar() ? CGF.EmitScalarConversion(Val.getScalarVal(), SrcType,
3890 DestType, Loc)
3891 : CGF.EmitComplexToScalarConversion(
3892 Val.getComplexVal(), SrcType, DestType, Loc);
3893}
3894
3895static CodeGenFunction::ComplexPairTy
3896convertToComplexValue(CodeGenFunction &CGF, RValue Val, QualType SrcType,
3897 QualType DestType, SourceLocation Loc) {
3898 assert(CGF.getEvaluationKind(DestType) == TEK_Complex &&((CGF.getEvaluationKind(DestType) == TEK_Complex && "DestType must have complex evaluation kind."
) ? static_cast<void> (0) : __assert_fail ("CGF.getEvaluationKind(DestType) == TEK_Complex && \"DestType must have complex evaluation kind.\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 3899, __PRETTY_FUNCTION__))
3899 "DestType must have complex evaluation kind.")((CGF.getEvaluationKind(DestType) == TEK_Complex && "DestType must have complex evaluation kind."
) ? static_cast<void> (0) : __assert_fail ("CGF.getEvaluationKind(DestType) == TEK_Complex && \"DestType must have complex evaluation kind.\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 3899, __PRETTY_FUNCTION__))
;
3900 CodeGenFunction::ComplexPairTy ComplexVal;
3901 if (Val.isScalar()) {
3902 // Convert the input element to the element type of the complex.
3903 QualType DestElementType =
3904 DestType->castAs<ComplexType>()->getElementType();
3905 llvm::Value *ScalarVal = CGF.EmitScalarConversion(
3906 Val.getScalarVal(), SrcType, DestElementType, Loc);
3907 ComplexVal = CodeGenFunction::ComplexPairTy(
3908 ScalarVal, llvm::Constant::getNullValue(ScalarVal->getType()));
3909 } else {
3910 assert(Val.isComplex() && "Must be a scalar or complex.")((Val.isComplex() && "Must be a scalar or complex.") ?
static_cast<void> (0) : __assert_fail ("Val.isComplex() && \"Must be a scalar or complex.\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 3910, __PRETTY_FUNCTION__))
;
3911 QualType SrcElementType = SrcType->castAs<ComplexType>()->getElementType();
3912 QualType DestElementType =
3913 DestType->castAs<ComplexType>()->getElementType();
3914 ComplexVal.first = CGF.EmitScalarConversion(
3915 Val.getComplexVal().first, SrcElementType, DestElementType, Loc);
3916 ComplexVal.second = CGF.EmitScalarConversion(
3917 Val.getComplexVal().second, SrcElementType, DestElementType, Loc);
3918 }
3919 return ComplexVal;
3920}
3921
3922static void emitSimpleAtomicStore(CodeGenFunction &CGF, bool IsSeqCst,
3923 LValue LVal, RValue RVal) {
3924 if (LVal.isGlobalReg()) {
3925 CGF.EmitStoreThroughGlobalRegLValue(RVal, LVal);
3926 } else {
3927 CGF.EmitAtomicStore(RVal, LVal,
3928 IsSeqCst ? llvm::AtomicOrdering::SequentiallyConsistent
3929 : llvm::AtomicOrdering::Monotonic,
3930 LVal.isVolatile(), /*isInit=*/false);
3931 }
3932}
3933
3934void CodeGenFunction::emitOMPSimpleStore(LValue LVal, RValue RVal,
3935 QualType RValTy, SourceLocation Loc) {
3936 switch (getEvaluationKind(LVal.getType())) {
3937 case TEK_Scalar:
3938 EmitStoreThroughLValue(RValue::get(convertToScalarValue(
3939 *this, RVal, RValTy, LVal.getType(), Loc)),
3940 LVal);
3941 break;
3942 case TEK_Complex:
3943 EmitStoreOfComplex(
3944 convertToComplexValue(*this, RVal, RValTy, LVal.getType(), Loc), LVal,
3945 /*isInit=*/false);
3946 break;
3947 case TEK_Aggregate:
3948 llvm_unreachable("Must be a scalar or complex.")::llvm::llvm_unreachable_internal("Must be a scalar or complex."
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 3948)
;
3949 }
3950}
3951
3952static void emitOMPAtomicReadExpr(CodeGenFunction &CGF, bool IsSeqCst,
3953 const Expr *X, const Expr *V,
3954 SourceLocation Loc) {
3955 // v = x;
3956 assert(V->isLValue() && "V of 'omp atomic read' is not lvalue")((V->isLValue() && "V of 'omp atomic read' is not lvalue"
) ? static_cast<void> (0) : __assert_fail ("V->isLValue() && \"V of 'omp atomic read' is not lvalue\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 3956, __PRETTY_FUNCTION__))
;
3957 assert(X->isLValue() && "X of 'omp atomic read' is not lvalue")((X->isLValue() && "X of 'omp atomic read' is not lvalue"
) ? static_cast<void> (0) : __assert_fail ("X->isLValue() && \"X of 'omp atomic read' is not lvalue\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 3957, __PRETTY_FUNCTION__))
;
3958 LValue XLValue = CGF.EmitLValue(X);
3959 LValue VLValue = CGF.EmitLValue(V);
3960 RValue Res = XLValue.isGlobalReg()
3961 ? CGF.EmitLoadOfLValue(XLValue, Loc)
3962 : CGF.EmitAtomicLoad(
3963 XLValue, Loc,
3964 IsSeqCst ? llvm::AtomicOrdering::SequentiallyConsistent
3965 : llvm::AtomicOrdering::Monotonic,
3966 XLValue.isVolatile());
3967 // OpenMP, 2.12.6, atomic Construct
3968 // Any atomic construct with a seq_cst clause forces the atomically
3969 // performed operation to include an implicit flush operation without a
3970 // list.
3971 if (IsSeqCst)
3972 CGF.CGM.getOpenMPRuntime().emitFlush(CGF, llvm::None, Loc);
3973 CGF.emitOMPSimpleStore(VLValue, Res, X->getType().getNonReferenceType(), Loc);
3974}
3975
3976static void emitOMPAtomicWriteExpr(CodeGenFunction &CGF, bool IsSeqCst,
3977 const Expr *X, const Expr *E,
3978 SourceLocation Loc) {
3979 // x = expr;
3980 assert(X->isLValue() && "X of 'omp atomic write' is not lvalue")((X->isLValue() && "X of 'omp atomic write' is not lvalue"
) ? static_cast<void> (0) : __assert_fail ("X->isLValue() && \"X of 'omp atomic write' is not lvalue\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 3980, __PRETTY_FUNCTION__))
;
3981 emitSimpleAtomicStore(CGF, IsSeqCst, CGF.EmitLValue(X), CGF.EmitAnyExpr(E));
3982 // OpenMP, 2.12.6, atomic Construct
3983 // Any atomic construct with a seq_cst clause forces the atomically
3984 // performed operation to include an implicit flush operation without a
3985 // list.
3986 if (IsSeqCst)
3987 CGF.CGM.getOpenMPRuntime().emitFlush(CGF, llvm::None, Loc);
3988}
3989
3990static std::pair<bool, RValue> emitOMPAtomicRMW(CodeGenFunction &CGF, LValue X,
3991 RValue Update,
3992 BinaryOperatorKind BO,
3993 llvm::AtomicOrdering AO,
3994 bool IsXLHSInRHSPart) {
3995 ASTContext &Context = CGF.getContext();
3996 // Allow atomicrmw only if 'x' and 'update' are integer values, lvalue for 'x'
3997 // expression is simple and atomic is allowed for the given type for the
3998 // target platform.
3999 if (BO == BO_Comma || !Update.isScalar() ||
4000 !Update.getScalarVal()->getType()->isIntegerTy() || !X.isSimple() ||
4001 (!isa<llvm::ConstantInt>(Update.getScalarVal()) &&
4002 (Update.getScalarVal()->getType() !=
4003 X.getAddress(CGF).getElementType())) ||
4004 !X.getAddress(CGF).getElementType()->isIntegerTy() ||
4005 !Context.getTargetInfo().hasBuiltinAtomic(
4006 Context.getTypeSize(X.getType()), Context.toBits(X.getAlignment())))
4007 return std::make_pair(false, RValue::get(nullptr));
4008
4009 llvm::AtomicRMWInst::BinOp RMWOp;
4010 switch (BO) {
4011 case BO_Add:
4012 RMWOp = llvm::AtomicRMWInst::Add;
4013 break;
4014 case BO_Sub:
4015 if (!IsXLHSInRHSPart)
4016 return std::make_pair(false, RValue::get(nullptr));
4017 RMWOp = llvm::AtomicRMWInst::Sub;
4018 break;
4019 case BO_And:
4020 RMWOp = llvm::AtomicRMWInst::And;
4021 break;
4022 case BO_Or:
4023 RMWOp = llvm::AtomicRMWInst::Or;
4024 break;
4025 case BO_Xor:
4026 RMWOp = llvm::AtomicRMWInst::Xor;
4027 break;
4028 case BO_LT:
4029 RMWOp = X.getType()->hasSignedIntegerRepresentation()
4030 ? (IsXLHSInRHSPart ? llvm::AtomicRMWInst::Min
4031 : llvm::AtomicRMWInst::Max)
4032 : (IsXLHSInRHSPart ? llvm::AtomicRMWInst::UMin
4033 : llvm::AtomicRMWInst::UMax);
4034 break;
4035 case BO_GT:
4036 RMWOp = X.getType()->hasSignedIntegerRepresentation()
4037 ? (IsXLHSInRHSPart ? llvm::AtomicRMWInst::Max
4038 : llvm::AtomicRMWInst::Min)
4039 : (IsXLHSInRHSPart ? llvm::AtomicRMWInst::UMax
4040 : llvm::AtomicRMWInst::UMin);
4041 break;
4042 case BO_Assign:
4043 RMWOp = llvm::AtomicRMWInst::Xchg;
4044 break;
4045 case BO_Mul:
4046 case BO_Div:
4047 case BO_Rem:
4048 case BO_Shl:
4049 case BO_Shr:
4050 case BO_LAnd:
4051 case BO_LOr:
4052 return std::make_pair(false, RValue::get(nullptr));
4053 case BO_PtrMemD:
4054 case BO_PtrMemI:
4055 case BO_LE:
4056 case BO_GE:
4057 case BO_EQ:
4058 case BO_NE:
4059 case BO_Cmp:
4060 case BO_AddAssign:
4061 case BO_SubAssign:
4062 case BO_AndAssign:
4063 case BO_OrAssign:
4064 case BO_XorAssign:
4065 case BO_MulAssign:
4066 case BO_DivAssign:
4067 case BO_RemAssign:
4068 case BO_ShlAssign:
4069 case BO_ShrAssign:
4070 case BO_Comma:
4071 llvm_unreachable("Unsupported atomic update operation")::llvm::llvm_unreachable_internal("Unsupported atomic update operation"
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 4071)
;
4072 }
4073 llvm::Value *UpdateVal = Update.getScalarVal();
4074 if (auto *IC = dyn_cast<llvm::ConstantInt>(UpdateVal)) {
4075 UpdateVal = CGF.Builder.CreateIntCast(
4076 IC, X.getAddress(CGF).getElementType(),
4077 X.getType()->hasSignedIntegerRepresentation());
4078 }
4079 llvm::Value *Res =
4080 CGF.Builder.CreateAtomicRMW(RMWOp, X.getPointer(CGF), UpdateVal, AO);
4081 return std::make_pair(true, RValue::get(Res));
4082}
4083
4084std::pair<bool, RValue> CodeGenFunction::EmitOMPAtomicSimpleUpdateExpr(
4085 LValue X, RValue E, BinaryOperatorKind BO, bool IsXLHSInRHSPart,
4086 llvm::AtomicOrdering AO, SourceLocation Loc,
4087 const llvm::function_ref<RValue(RValue)> CommonGen) {
4088 // Update expressions are allowed to have the following forms:
4089 // x binop= expr; -> xrval + expr;
4090 // x++, ++x -> xrval + 1;
4091 // x--, --x -> xrval - 1;
4092 // x = x binop expr; -> xrval binop expr
4093 // x = expr Op x; - > expr binop xrval;
4094 auto Res = emitOMPAtomicRMW(*this, X, E, BO, AO, IsXLHSInRHSPart);
4095 if (!Res.first) {
4096 if (X.isGlobalReg()) {
4097 // Emit an update expression: 'xrval' binop 'expr' or 'expr' binop
4098 // 'xrval'.
4099 EmitStoreThroughLValue(CommonGen(EmitLoadOfLValue(X, Loc)), X);
4100 } else {
4101 // Perform compare-and-swap procedure.
4102 EmitAtomicUpdate(X, AO, CommonGen, X.getType().isVolatileQualified());
4103 }
4104 }
4105 return Res;
4106}
4107
4108static void emitOMPAtomicUpdateExpr(CodeGenFunction &CGF, bool IsSeqCst,
4109 const Expr *X, const Expr *E,
4110 const Expr *UE, bool IsXLHSInRHSPart,
4111 SourceLocation Loc) {
4112 assert(isa<BinaryOperator>(UE->IgnoreImpCasts()) &&((isa<BinaryOperator>(UE->IgnoreImpCasts()) &&
"Update expr in 'atomic update' must be a binary operator.")
? static_cast<void> (0) : __assert_fail ("isa<BinaryOperator>(UE->IgnoreImpCasts()) && \"Update expr in 'atomic update' must be a binary operator.\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 4113, __PRETTY_FUNCTION__))
4113 "Update expr in 'atomic update' must be a binary operator.")((isa<BinaryOperator>(UE->IgnoreImpCasts()) &&
"Update expr in 'atomic update' must be a binary operator.")
? static_cast<void> (0) : __assert_fail ("isa<BinaryOperator>(UE->IgnoreImpCasts()) && \"Update expr in 'atomic update' must be a binary operator.\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 4113, __PRETTY_FUNCTION__))
;
4114 const auto *BOUE = cast<BinaryOperator>(UE->IgnoreImpCasts());
4115 // Update expressions are allowed to have the following forms:
4116 // x binop= expr; -> xrval + expr;
4117 // x++, ++x -> xrval + 1;
4118 // x--, --x -> xrval - 1;
4119 // x = x binop expr; -> xrval binop expr
4120 // x = expr Op x; - > expr binop xrval;
4121 assert(X->isLValue() && "X of 'omp atomic update' is not lvalue")((X->isLValue() && "X of 'omp atomic update' is not lvalue"
) ? static_cast<void> (0) : __assert_fail ("X->isLValue() && \"X of 'omp atomic update' is not lvalue\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 4121, __PRETTY_FUNCTION__))
;
4122 LValue XLValue = CGF.EmitLValue(X);
4123 RValue ExprRValue = CGF.EmitAnyExpr(E);
4124 llvm::AtomicOrdering AO = IsSeqCst
4125 ? llvm::AtomicOrdering::SequentiallyConsistent
4126 : llvm::AtomicOrdering::Monotonic;
4127 const auto *LHS = cast<OpaqueValueExpr>(BOUE->getLHS()->IgnoreImpCasts());
4128 const auto *RHS = cast<OpaqueValueExpr>(BOUE->getRHS()->IgnoreImpCasts());
4129 const OpaqueValueExpr *XRValExpr = IsXLHSInRHSPart ? LHS : RHS;
4130 const OpaqueValueExpr *ERValExpr = IsXLHSInRHSPart ? RHS : LHS;
4131 auto &&Gen = [&CGF, UE, ExprRValue, XRValExpr, ERValExpr](RValue XRValue) {
4132 CodeGenFunction::OpaqueValueMapping MapExpr(CGF, ERValExpr, ExprRValue);
4133 CodeGenFunction::OpaqueValueMapping MapX(CGF, XRValExpr, XRValue);
4134 return CGF.EmitAnyExpr(UE);
4135 };
4136 (void)CGF.EmitOMPAtomicSimpleUpdateExpr(
4137 XLValue, ExprRValue, BOUE->getOpcode(), IsXLHSInRHSPart, AO, Loc, Gen);
4138 // OpenMP, 2.12.6, atomic Construct
4139 // Any atomic construct with a seq_cst clause forces the atomically
4140 // performed operation to include an implicit flush operation without a
4141 // list.
4142 if (IsSeqCst)
4143 CGF.CGM.getOpenMPRuntime().emitFlush(CGF, llvm::None, Loc);
4144}
4145
4146static RValue convertToType(CodeGenFunction &CGF, RValue Value,
4147 QualType SourceType, QualType ResType,
4148 SourceLocation Loc) {
4149 switch (CGF.getEvaluationKind(ResType)) {
4150 case TEK_Scalar:
4151 return RValue::get(
4152 convertToScalarValue(CGF, Value, SourceType, ResType, Loc));
4153 case TEK_Complex: {
4154 auto Res = convertToComplexValue(CGF, Value, SourceType, ResType, Loc);
4155 return RValue::getComplex(Res.first, Res.second);
4156 }
4157 case TEK_Aggregate:
4158 break;
4159 }
4160 llvm_unreachable("Must be a scalar or complex.")::llvm::llvm_unreachable_internal("Must be a scalar or complex."
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 4160)
;
4161}
4162
4163static void emitOMPAtomicCaptureExpr(CodeGenFunction &CGF, bool IsSeqCst,
4164 bool IsPostfixUpdate, const Expr *V,
4165 const Expr *X, const Expr *E,
4166 const Expr *UE, bool IsXLHSInRHSPart,
4167 SourceLocation Loc) {
4168 assert(X->isLValue() && "X of 'omp atomic capture' is not lvalue")((X->isLValue() && "X of 'omp atomic capture' is not lvalue"
) ? static_cast<void> (0) : __assert_fail ("X->isLValue() && \"X of 'omp atomic capture' is not lvalue\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 4168, __PRETTY_FUNCTION__))
;
4169 assert(V->isLValue() && "V of 'omp atomic capture' is not lvalue")((V->isLValue() && "V of 'omp atomic capture' is not lvalue"
) ? static_cast<void> (0) : __assert_fail ("V->isLValue() && \"V of 'omp atomic capture' is not lvalue\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 4169, __PRETTY_FUNCTION__))
;
4170 RValue NewVVal;
4171 LValue VLValue = CGF.EmitLValue(V);
4172 LValue XLValue = CGF.EmitLValue(X);
4173 RValue ExprRValue = CGF.EmitAnyExpr(E);
4174 llvm::AtomicOrdering AO = IsSeqCst
4175 ? llvm::AtomicOrdering::SequentiallyConsistent
4176 : llvm::AtomicOrdering::Monotonic;
4177 QualType NewVValType;
4178 if (UE) {
4179 // 'x' is updated with some additional value.
4180 assert(isa<BinaryOperator>(UE->IgnoreImpCasts()) &&((isa<BinaryOperator>(UE->IgnoreImpCasts()) &&
"Update expr in 'atomic capture' must be a binary operator."
) ? static_cast<void> (0) : __assert_fail ("isa<BinaryOperator>(UE->IgnoreImpCasts()) && \"Update expr in 'atomic capture' must be a binary operator.\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 4181, __PRETTY_FUNCTION__))
4181 "Update expr in 'atomic capture' must be a binary operator.")((isa<BinaryOperator>(UE->IgnoreImpCasts()) &&
"Update expr in 'atomic capture' must be a binary operator."
) ? static_cast<void> (0) : __assert_fail ("isa<BinaryOperator>(UE->IgnoreImpCasts()) && \"Update expr in 'atomic capture' must be a binary operator.\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 4181, __PRETTY_FUNCTION__))
;
4182 const auto *BOUE = cast<BinaryOperator>(UE->IgnoreImpCasts());
4183 // Update expressions are allowed to have the following forms:
4184 // x binop= expr; -> xrval + expr;
4185 // x++, ++x -> xrval + 1;
4186 // x--, --x -> xrval - 1;
4187 // x = x binop expr; -> xrval binop expr
4188 // x = expr Op x; - > expr binop xrval;
4189 const auto *LHS = cast<OpaqueValueExpr>(BOUE->getLHS()->IgnoreImpCasts());
4190 const auto *RHS = cast<OpaqueValueExpr>(BOUE->getRHS()->IgnoreImpCasts());
4191 const OpaqueValueExpr *XRValExpr = IsXLHSInRHSPart ? LHS : RHS;
4192 NewVValType = XRValExpr->getType();
4193 const OpaqueValueExpr *ERValExpr = IsXLHSInRHSPart ? RHS : LHS;
4194 auto &&Gen = [&CGF, &NewVVal, UE, ExprRValue, XRValExpr, ERValExpr,
4195 IsPostfixUpdate](RValue XRValue) {
4196 CodeGenFunction::OpaqueValueMapping MapExpr(CGF, ERValExpr, ExprRValue);
4197 CodeGenFunction::OpaqueValueMapping MapX(CGF, XRValExpr, XRValue);
4198 RValue Res = CGF.EmitAnyExpr(UE);
4199 NewVVal = IsPostfixUpdate ? XRValue : Res;
4200 return Res;
4201 };
4202 auto Res = CGF.EmitOMPAtomicSimpleUpdateExpr(
4203 XLValue, ExprRValue, BOUE->getOpcode(), IsXLHSInRHSPart, AO, Loc, Gen);
4204 if (Res.first) {
4205 // 'atomicrmw' instruction was generated.
4206 if (IsPostfixUpdate) {
4207 // Use old value from 'atomicrmw'.
4208 NewVVal = Res.second;
4209 } else {
4210 // 'atomicrmw' does not provide new value, so evaluate it using old
4211 // value of 'x'.
4212 CodeGenFunction::OpaqueValueMapping MapExpr(CGF, ERValExpr, ExprRValue);
4213 CodeGenFunction::OpaqueValueMapping MapX(CGF, XRValExpr, Res.second);
4214 NewVVal = CGF.EmitAnyExpr(UE);
4215 }
4216 }
4217 } else {
4218 // 'x' is simply rewritten with some 'expr'.
4219 NewVValType = X->getType().getNonReferenceType();
4220 ExprRValue = convertToType(CGF, ExprRValue, E->getType(),
4221 X->getType().getNonReferenceType(), Loc);
4222 auto &&Gen = [&NewVVal, ExprRValue](RValue XRValue) {
4223 NewVVal = XRValue;
4224 return ExprRValue;
4225 };
4226 // Try to perform atomicrmw xchg, otherwise simple exchange.
4227 auto Res = CGF.EmitOMPAtomicSimpleUpdateExpr(
4228 XLValue, ExprRValue, /*BO=*/BO_Assign, /*IsXLHSInRHSPart=*/false, AO,
4229 Loc, Gen);
4230 if (Res.first) {
4231 // 'atomicrmw' instruction was generated.
4232 NewVVal = IsPostfixUpdate ? Res.second : ExprRValue;
4233 }
4234 }
4235 // Emit post-update store to 'v' of old/new 'x' value.
4236 CGF.emitOMPSimpleStore(VLValue, NewVVal, NewVValType, Loc);
4237 // OpenMP, 2.12.6, atomic Construct
4238 // Any atomic construct with a seq_cst clause forces the atomically
4239 // performed operation to include an implicit flush operation without a
4240 // list.
4241 if (IsSeqCst)
4242 CGF.CGM.getOpenMPRuntime().emitFlush(CGF, llvm::None, Loc);
4243}
4244
4245static void emitOMPAtomicExpr(CodeGenFunction &CGF, OpenMPClauseKind Kind,
4246 bool IsSeqCst, bool IsPostfixUpdate,
4247 const Expr *X, const Expr *V, const Expr *E,
4248 const Expr *UE, bool IsXLHSInRHSPart,
4249 SourceLocation Loc) {
4250 switch (Kind) {
4251 case OMPC_read:
4252 emitOMPAtomicReadExpr(CGF, IsSeqCst, X, V, Loc);
4253 break;
4254 case OMPC_write:
4255 emitOMPAtomicWriteExpr(CGF, IsSeqCst, X, E, Loc);
4256 break;
4257 case OMPC_unknown:
4258 case OMPC_update:
4259 emitOMPAtomicUpdateExpr(CGF, IsSeqCst, X, E, UE, IsXLHSInRHSPart, Loc);
4260 break;
4261 case OMPC_capture:
4262 emitOMPAtomicCaptureExpr(CGF, IsSeqCst, IsPostfixUpdate, V, X, E, UE,
4263 IsXLHSInRHSPart, Loc);
4264 break;
4265 case OMPC_if:
4266 case OMPC_final:
4267 case OMPC_num_threads:
4268 case OMPC_private:
4269 case OMPC_firstprivate:
4270 case OMPC_lastprivate:
4271 case OMPC_reduction:
4272 case OMPC_task_reduction:
4273 case OMPC_in_reduction:
4274 case OMPC_safelen:
4275 case OMPC_simdlen:
4276 case OMPC_allocator:
4277 case OMPC_allocate:
4278 case OMPC_collapse:
4279 case OMPC_default:
4280 case OMPC_seq_cst:
4281 case OMPC_shared:
4282 case OMPC_linear:
4283 case OMPC_aligned:
4284 case OMPC_copyin:
4285 case OMPC_copyprivate:
4286 case OMPC_flush:
4287 case OMPC_proc_bind:
4288 case OMPC_schedule:
4289 case OMPC_ordered:
4290 case OMPC_nowait:
4291 case OMPC_untied:
4292 case OMPC_threadprivate:
4293 case OMPC_depend:
4294 case OMPC_mergeable:
4295 case OMPC_device:
4296 case OMPC_threads:
4297 case OMPC_simd:
4298 case OMPC_map:
4299 case OMPC_num_teams:
4300 case OMPC_thread_limit:
4301 case OMPC_priority:
4302 case OMPC_grainsize:
4303 case OMPC_nogroup:
4304 case OMPC_num_tasks:
4305 case OMPC_hint:
4306 case OMPC_dist_schedule:
4307 case OMPC_defaultmap:
4308 case OMPC_uniform:
4309 case OMPC_to:
4310 case OMPC_from:
4311 case OMPC_use_device_ptr:
4312 case OMPC_is_device_ptr:
4313 case OMPC_unified_address:
4314 case OMPC_unified_shared_memory:
4315 case OMPC_reverse_offload:
4316 case OMPC_dynamic_allocators:
4317 case OMPC_atomic_default_mem_order:
4318 case OMPC_device_type:
4319 case OMPC_match:
4320 case OMPC_nontemporal:
4321 llvm_unreachable("Clause is not allowed in 'omp atomic'.")::llvm::llvm_unreachable_internal("Clause is not allowed in 'omp atomic'."
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 4321)
;
4322 }
4323}
4324
4325void CodeGenFunction::EmitOMPAtomicDirective(const OMPAtomicDirective &S) {
4326 bool IsSeqCst = S.getSingleClause<OMPSeqCstClause>();
4327 OpenMPClauseKind Kind = OMPC_unknown;
4328 for (const OMPClause *C : S.clauses()) {
4329 // Find first clause (skip seq_cst clause, if it is first).
4330 if (C->getClauseKind() != OMPC_seq_cst) {
4331 Kind = C->getClauseKind();
4332 break;
4333 }
4334 }
4335
4336 const Stmt *CS = S.getInnermostCapturedStmt()->IgnoreContainers();
4337 if (const auto *FE = dyn_cast<FullExpr>(CS))
4338 enterFullExpression(FE);
4339 // Processing for statements under 'atomic capture'.
4340 if (const auto *Compound = dyn_cast<CompoundStmt>(CS)) {
4341 for (const Stmt *C : Compound->body()) {
4342 if (const auto *FE = dyn_cast<FullExpr>(C))
4343 enterFullExpression(FE);
4344 }
4345 }
4346
4347 auto &&CodeGen = [&S, Kind, IsSeqCst, CS](CodeGenFunction &CGF,
4348 PrePostActionTy &) {
4349 CGF.EmitStopPoint(CS);
4350 emitOMPAtomicExpr(CGF, Kind, IsSeqCst, S.isPostfixUpdate(), S.getX(),
4351 S.getV(), S.getExpr(), S.getUpdateExpr(),
4352 S.isXLHSInRHSPart(), S.getBeginLoc());
4353 };
4354 OMPLexicalScope Scope(*this, S, OMPD_unknown);
4355 CGM.getOpenMPRuntime().emitInlinedDirective(*this, OMPD_atomic, CodeGen);
4356}
4357
4358static void emitCommonOMPTargetDirective(CodeGenFunction &CGF,
4359 const OMPExecutableDirective &S,
4360 const RegionCodeGenTy &CodeGen) {
4361 assert(isOpenMPTargetExecutionDirective(S.getDirectiveKind()))((isOpenMPTargetExecutionDirective(S.getDirectiveKind())) ? static_cast
<void> (0) : __assert_fail ("isOpenMPTargetExecutionDirective(S.getDirectiveKind())"
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 4361, __PRETTY_FUNCTION__))
;
4362 CodeGenModule &CGM = CGF.CGM;
4363
4364 // On device emit this construct as inlined code.
4365 if (CGM.getLangOpts().OpenMPIsDevice) {
4366 OMPLexicalScope Scope(CGF, S, OMPD_target);
4367 CGM.getOpenMPRuntime().emitInlinedDirective(
4368 CGF, OMPD_target, [&S](CodeGenFunction &CGF, PrePostActionTy &) {
4369 CGF.EmitStmt(S.getInnermostCapturedStmt()->getCapturedStmt());
4370 });
4371 return;
4372 }
4373
4374 llvm::Function *Fn = nullptr;
4375 llvm::Constant *FnID = nullptr;
4376
4377 const Expr *IfCond = nullptr;
4378 // Check for the at most one if clause associated with the target region.
4379 for (const auto *C : S.getClausesOfKind<OMPIfClause>()) {
4380 if (C->getNameModifier() == OMPD_unknown ||
4381 C->getNameModifier() == OMPD_target) {
4382 IfCond = C->getCondition();
4383 break;
4384 }
4385 }
4386
4387 // Check if we have any device clause associated with the directive.
4388 const Expr *Device = nullptr;
4389 if (auto *C = S.getSingleClause<OMPDeviceClause>())
4390 Device = C->getDevice();
4391
4392 // Check if we have an if clause whose conditional always evaluates to false
4393 // or if we do not have any targets specified. If so the target region is not
4394 // an offload entry point.
4395 bool IsOffloadEntry = true;
4396 if (IfCond) {
4397 bool Val;
4398 if (CGF.ConstantFoldsToSimpleInteger(IfCond, Val) && !Val)
4399 IsOffloadEntry = false;
4400 }
4401 if (CGM.getLangOpts().OMPTargetTriples.empty())
4402 IsOffloadEntry = false;
4403
4404 assert(CGF.CurFuncDecl && "No parent declaration for target region!")((CGF.CurFuncDecl && "No parent declaration for target region!"
) ? static_cast<void> (0) : __assert_fail ("CGF.CurFuncDecl && \"No parent declaration for target region!\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 4404, __PRETTY_FUNCTION__))
;
4405 StringRef ParentName;
4406 // In case we have Ctors/Dtors we use the complete type variant to produce
4407 // the mangling of the device outlined kernel.
4408 if (const auto *D = dyn_cast<CXXConstructorDecl>(CGF.CurFuncDecl))
4409 ParentName = CGM.getMangledName(GlobalDecl(D, Ctor_Complete));
4410 else if (const auto *D = dyn_cast<CXXDestructorDecl>(CGF.CurFuncDecl))
4411 ParentName = CGM.getMangledName(GlobalDecl(D, Dtor_Complete));
4412 else
4413 ParentName =
4414 CGM.getMangledName(GlobalDecl(cast<FunctionDecl>(CGF.CurFuncDecl)));
4415
4416 // Emit target region as a standalone region.
4417 CGM.getOpenMPRuntime().emitTargetOutlinedFunction(S, ParentName, Fn, FnID,
4418 IsOffloadEntry, CodeGen);
4419 OMPLexicalScope Scope(CGF, S, OMPD_task);
4420 auto &&SizeEmitter =
4421 [IsOffloadEntry](CodeGenFunction &CGF,
4422 const OMPLoopDirective &D) -> llvm::Value * {
4423 if (IsOffloadEntry) {
4424 OMPLoopScope(CGF, D);
4425 // Emit calculation of the iterations count.
4426 llvm::Value *NumIterations = CGF.EmitScalarExpr(D.getNumIterations());
4427 NumIterations = CGF.Builder.CreateIntCast(NumIterations, CGF.Int64Ty,
4428 /*isSigned=*/false);
4429 return NumIterations;
4430 }
4431 return nullptr;
4432 };
4433 CGM.getOpenMPRuntime().emitTargetCall(CGF, S, Fn, FnID, IfCond, Device,
4434 SizeEmitter);
4435}
4436
4437static void emitTargetRegion(CodeGenFunction &CGF, const OMPTargetDirective &S,
4438 PrePostActionTy &Action) {
4439 Action.Enter(CGF);
4440 CodeGenFunction::OMPPrivateScope PrivateScope(CGF);
4441 (void)CGF.EmitOMPFirstprivateClause(S, PrivateScope);
4442 CGF.EmitOMPPrivateClause(S, PrivateScope);
4443 (void)PrivateScope.Privatize();
4444 if (isOpenMPTargetExecutionDirective(S.getDirectiveKind()))
4445 CGF.CGM.getOpenMPRuntime().adjustTargetSpecificDataForLambdas(CGF, S);
4446
4447 CGF.EmitStmt(S.getCapturedStmt(OMPD_target)->getCapturedStmt());
4448}
4449
4450void CodeGenFunction::EmitOMPTargetDeviceFunction(CodeGenModule &CGM,
4451 StringRef ParentName,
4452 const OMPTargetDirective &S) {
4453 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
4454 emitTargetRegion(CGF, S, Action);
4455 };
4456 llvm::Function *Fn;
4457 llvm::Constant *Addr;
4458 // Emit target region as a standalone region.
4459 CGM.getOpenMPRuntime().emitTargetOutlinedFunction(
4460 S, ParentName, Fn, Addr, /*IsOffloadEntry=*/true, CodeGen);
4461 assert(Fn && Addr && "Target device function emission failed.")((Fn && Addr && "Target device function emission failed."
) ? static_cast<void> (0) : __assert_fail ("Fn && Addr && \"Target device function emission failed.\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 4461, __PRETTY_FUNCTION__))
;
4462}
4463
4464void CodeGenFunction::EmitOMPTargetDirective(const OMPTargetDirective &S) {
4465 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
4466 emitTargetRegion(CGF, S, Action);
4467 };
4468 emitCommonOMPTargetDirective(*this, S, CodeGen);
4469}
4470
4471static void emitCommonOMPTeamsDirective(CodeGenFunction &CGF,
4472 const OMPExecutableDirective &S,
4473 OpenMPDirectiveKind InnermostKind,
4474 const RegionCodeGenTy &CodeGen) {
4475 const CapturedStmt *CS = S.getCapturedStmt(OMPD_teams);
4476 llvm::Function *OutlinedFn =
4477 CGF.CGM.getOpenMPRuntime().emitTeamsOutlinedFunction(
4478 S, *CS->getCapturedDecl()->param_begin(), InnermostKind, CodeGen);
4479
4480 const auto *NT = S.getSingleClause<OMPNumTeamsClause>();
4481 const auto *TL = S.getSingleClause<OMPThreadLimitClause>();
4482 if (NT || TL) {
4483 const Expr *NumTeams = NT ? NT->getNumTeams() : nullptr;
4484 const Expr *ThreadLimit = TL ? TL->getThreadLimit() : nullptr;
4485
4486 CGF.CGM.getOpenMPRuntime().emitNumTeamsClause(CGF, NumTeams, ThreadLimit,
4487 S.getBeginLoc());
4488 }
4489
4490 OMPTeamsScope Scope(CGF, S);
4491 llvm::SmallVector<llvm::Value *, 16> CapturedVars;
4492 CGF.GenerateOpenMPCapturedVars(*CS, CapturedVars);
4493 CGF.CGM.getOpenMPRuntime().emitTeamsCall(CGF, S, S.getBeginLoc(), OutlinedFn,
4494 CapturedVars);
4495}
4496
4497void CodeGenFunction::EmitOMPTeamsDirective(const OMPTeamsDirective &S) {
4498 // Emit teams region as a standalone region.
4499 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
4500 Action.Enter(CGF);
4501 OMPPrivateScope PrivateScope(CGF);
4502 (void)CGF.EmitOMPFirstprivateClause(S, PrivateScope);
4503 CGF.EmitOMPPrivateClause(S, PrivateScope);
4504 CGF.EmitOMPReductionClauseInit(S, PrivateScope);
4505 (void)PrivateScope.Privatize();
4506 CGF.EmitStmt(S.getCapturedStmt(OMPD_teams)->getCapturedStmt());
4507 CGF.EmitOMPReductionClauseFinal(S, /*ReductionKind=*/OMPD_teams);
4508 };
4509 emitCommonOMPTeamsDirective(*this, S, OMPD_distribute, CodeGen);
4510 emitPostUpdateForReductionClause(*this, S,
4511 [](CodeGenFunction &) { return nullptr; });
4512}
4513
4514static void emitTargetTeamsRegion(CodeGenFunction &CGF, PrePostActionTy &Action,
4515 const OMPTargetTeamsDirective &S) {
4516 auto *CS = S.getCapturedStmt(OMPD_teams);
4517 Action.Enter(CGF);
4518 // Emit teams region as a standalone region.
4519 auto &&CodeGen = [&S, CS](CodeGenFunction &CGF, PrePostActionTy &Action) {
4520 Action.Enter(CGF);
4521 CodeGenFunction::OMPPrivateScope PrivateScope(CGF);
4522 (void)CGF.EmitOMPFirstprivateClause(S, PrivateScope);
4523 CGF.EmitOMPPrivateClause(S, PrivateScope);
4524 CGF.EmitOMPReductionClauseInit(S, PrivateScope);
4525 (void)PrivateScope.Privatize();
4526 if (isOpenMPTargetExecutionDirective(S.getDirectiveKind()))
4527 CGF.CGM.getOpenMPRuntime().adjustTargetSpecificDataForLambdas(CGF, S);
4528 CGF.EmitStmt(CS->getCapturedStmt());
4529 CGF.EmitOMPReductionClauseFinal(S, /*ReductionKind=*/OMPD_teams);
4530 };
4531 emitCommonOMPTeamsDirective(CGF, S, OMPD_teams, CodeGen);
4532 emitPostUpdateForReductionClause(CGF, S,
4533 [](CodeGenFunction &) { return nullptr; });
4534}
4535
4536void CodeGenFunction::EmitOMPTargetTeamsDeviceFunction(
4537 CodeGenModule &CGM, StringRef ParentName,
4538 const OMPTargetTeamsDirective &S) {
4539 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
4540 emitTargetTeamsRegion(CGF, Action, S);
4541 };
4542 llvm::Function *Fn;
4543 llvm::Constant *Addr;
4544 // Emit target region as a standalone region.
4545 CGM.getOpenMPRuntime().emitTargetOutlinedFunction(
4546 S, ParentName, Fn, Addr, /*IsOffloadEntry=*/true, CodeGen);
4547 assert(Fn && Addr && "Target device function emission failed.")((Fn && Addr && "Target device function emission failed."
) ? static_cast<void> (0) : __assert_fail ("Fn && Addr && \"Target device function emission failed.\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 4547, __PRETTY_FUNCTION__))
;
4548}
4549
4550void CodeGenFunction::EmitOMPTargetTeamsDirective(
4551 const OMPTargetTeamsDirective &S) {
4552 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
4553 emitTargetTeamsRegion(CGF, Action, S);
4554 };
4555 emitCommonOMPTargetDirective(*this, S, CodeGen);
4556}
4557
4558static void
4559emitTargetTeamsDistributeRegion(CodeGenFunction &CGF, PrePostActionTy &Action,
4560 const OMPTargetTeamsDistributeDirective &S) {
4561 Action.Enter(CGF);
4562 auto &&CodeGenDistribute = [&S](CodeGenFunction &CGF, PrePostActionTy &) {
4563 CGF.EmitOMPDistributeLoop(S, emitOMPLoopBodyWithStopPoint, S.getInc());
4564 };
4565
4566 // Emit teams region as a standalone region.
4567 auto &&CodeGen = [&S, &CodeGenDistribute](CodeGenFunction &CGF,
4568 PrePostActionTy &Action) {
4569 Action.Enter(CGF);
4570 CodeGenFunction::OMPPrivateScope PrivateScope(CGF);
4571 CGF.EmitOMPReductionClauseInit(S, PrivateScope);
4572 (void)PrivateScope.Privatize();
4573 CGF.CGM.getOpenMPRuntime().emitInlinedDirective(CGF, OMPD_distribute,
4574 CodeGenDistribute);
4575 CGF.EmitOMPReductionClauseFinal(S, /*ReductionKind=*/OMPD_teams);
4576 };
4577 emitCommonOMPTeamsDirective(CGF, S, OMPD_distribute, CodeGen);
4578 emitPostUpdateForReductionClause(CGF, S,
4579 [](CodeGenFunction &) { return nullptr; });
4580}
4581
4582void CodeGenFunction::EmitOMPTargetTeamsDistributeDeviceFunction(
4583 CodeGenModule &CGM, StringRef ParentName,
4584 const OMPTargetTeamsDistributeDirective &S) {
4585 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
4586 emitTargetTeamsDistributeRegion(CGF, Action, S);
4587 };
4588 llvm::Function *Fn;
4589 llvm::Constant *Addr;
4590 // Emit target region as a standalone region.
4591 CGM.getOpenMPRuntime().emitTargetOutlinedFunction(
4592 S, ParentName, Fn, Addr, /*IsOffloadEntry=*/true, CodeGen);
4593 assert(Fn && Addr && "Target device function emission failed.")((Fn && Addr && "Target device function emission failed."
) ? static_cast<void> (0) : __assert_fail ("Fn && Addr && \"Target device function emission failed.\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 4593, __PRETTY_FUNCTION__))
;
4594}
4595
4596void CodeGenFunction::EmitOMPTargetTeamsDistributeDirective(
4597 const OMPTargetTeamsDistributeDirective &S) {
4598 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
4599 emitTargetTeamsDistributeRegion(CGF, Action, S);
4600 };
4601 emitCommonOMPTargetDirective(*this, S, CodeGen);
4602}
4603
4604static void emitTargetTeamsDistributeSimdRegion(
4605 CodeGenFunction &CGF, PrePostActionTy &Action,
4606 const OMPTargetTeamsDistributeSimdDirective &S) {
4607 Action.Enter(CGF);
4608 auto &&CodeGenDistribute = [&S](CodeGenFunction &CGF, PrePostActionTy &) {
4609 CGF.EmitOMPDistributeLoop(S, emitOMPLoopBodyWithStopPoint, S.getInc());
4610 };
4611
4612 // Emit teams region as a standalone region.
4613 auto &&CodeGen = [&S, &CodeGenDistribute](CodeGenFunction &CGF,
4614 PrePostActionTy &Action) {
4615 Action.Enter(CGF);
4616 CodeGenFunction::OMPPrivateScope PrivateScope(CGF);
4617 CGF.EmitOMPReductionClauseInit(S, PrivateScope);
4618 (void)PrivateScope.Privatize();
4619 CGF.CGM.getOpenMPRuntime().emitInlinedDirective(CGF, OMPD_distribute,
4620 CodeGenDistribute);
4621 CGF.EmitOMPReductionClauseFinal(S, /*ReductionKind=*/OMPD_teams);
4622 };
4623 emitCommonOMPTeamsDirective(CGF, S, OMPD_distribute_simd, CodeGen);
4624 emitPostUpdateForReductionClause(CGF, S,
4625 [](CodeGenFunction &) { return nullptr; });
4626}
4627
4628void CodeGenFunction::EmitOMPTargetTeamsDistributeSimdDeviceFunction(
4629 CodeGenModule &CGM, StringRef ParentName,
4630 const OMPTargetTeamsDistributeSimdDirective &S) {
4631 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
4632 emitTargetTeamsDistributeSimdRegion(CGF, Action, S);
4633 };
4634 llvm::Function *Fn;
4635 llvm::Constant *Addr;
4636 // Emit target region as a standalone region.
4637 CGM.getOpenMPRuntime().emitTargetOutlinedFunction(
4638 S, ParentName, Fn, Addr, /*IsOffloadEntry=*/true, CodeGen);
4639 assert(Fn && Addr && "Target device function emission failed.")((Fn && Addr && "Target device function emission failed."
) ? static_cast<void> (0) : __assert_fail ("Fn && Addr && \"Target device function emission failed.\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 4639, __PRETTY_FUNCTION__))
;
4640}
4641
4642void CodeGenFunction::EmitOMPTargetTeamsDistributeSimdDirective(
4643 const OMPTargetTeamsDistributeSimdDirective &S) {
4644 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
4645 emitTargetTeamsDistributeSimdRegion(CGF, Action, S);
4646 };
4647 emitCommonOMPTargetDirective(*this, S, CodeGen);
4648}
4649
4650void CodeGenFunction::EmitOMPTeamsDistributeDirective(
4651 const OMPTeamsDistributeDirective &S) {
4652
4653 auto &&CodeGenDistribute = [&S](CodeGenFunction &CGF, PrePostActionTy &) {
4654 CGF.EmitOMPDistributeLoop(S, emitOMPLoopBodyWithStopPoint, S.getInc());
4655 };
4656
4657 // Emit teams region as a standalone region.
4658 auto &&CodeGen = [&S, &CodeGenDistribute](CodeGenFunction &CGF,
4659 PrePostActionTy &Action) {
4660 Action.Enter(CGF);
4661 OMPPrivateScope PrivateScope(CGF);
4662 CGF.EmitOMPReductionClauseInit(S, PrivateScope);
4663 (void)PrivateScope.Privatize();
4664 CGF.CGM.getOpenMPRuntime().emitInlinedDirective(CGF, OMPD_distribute,
4665 CodeGenDistribute);
4666 CGF.EmitOMPReductionClauseFinal(S, /*ReductionKind=*/OMPD_teams);
4667 };
4668 emitCommonOMPTeamsDirective(*this, S, OMPD_distribute, CodeGen);
4669 emitPostUpdateForReductionClause(*this, S,
4670 [](CodeGenFunction &) { return nullptr; });
4671}
4672
4673void CodeGenFunction::EmitOMPTeamsDistributeSimdDirective(
4674 const OMPTeamsDistributeSimdDirective &S) {
4675 auto &&CodeGenDistribute = [&S](CodeGenFunction &CGF, PrePostActionTy &) {
4676 CGF.EmitOMPDistributeLoop(S, emitOMPLoopBodyWithStopPoint, S.getInc());
4677 };
4678
4679 // Emit teams region as a standalone region.
4680 auto &&CodeGen = [&S, &CodeGenDistribute](CodeGenFunction &CGF,
4681 PrePostActionTy &Action) {
4682 Action.Enter(CGF);
4683 OMPPrivateScope PrivateScope(CGF);
4684 CGF.EmitOMPReductionClauseInit(S, PrivateScope);
4685 (void)PrivateScope.Privatize();
4686 CGF.CGM.getOpenMPRuntime().emitInlinedDirective(CGF, OMPD_simd,
4687 CodeGenDistribute);
4688 CGF.EmitOMPReductionClauseFinal(S, /*ReductionKind=*/OMPD_teams);
4689 };
4690 emitCommonOMPTeamsDirective(*this, S, OMPD_distribute_simd, CodeGen);
4691 emitPostUpdateForReductionClause(*this, S,
4692 [](CodeGenFunction &) { return nullptr; });
4693}
4694
4695void CodeGenFunction::EmitOMPTeamsDistributeParallelForDirective(
4696 const OMPTeamsDistributeParallelForDirective &S) {
4697 auto &&CodeGenDistribute = [&S](CodeGenFunction &CGF, PrePostActionTy &) {
4698 CGF.EmitOMPDistributeLoop(S, emitInnerParallelForWhenCombined,
4699 S.getDistInc());
4700 };
4701
4702 // Emit teams region as a standalone region.
4703 auto &&CodeGen = [&S, &CodeGenDistribute](CodeGenFunction &CGF,
4704 PrePostActionTy &Action) {
4705 Action.Enter(CGF);
4706 OMPPrivateScope PrivateScope(CGF);
4707 CGF.EmitOMPReductionClauseInit(S, PrivateScope);
4708 (void)PrivateScope.Privatize();
4709 CGF.CGM.getOpenMPRuntime().emitInlinedDirective(CGF, OMPD_distribute,
4710 CodeGenDistribute);
4711 CGF.EmitOMPReductionClauseFinal(S, /*ReductionKind=*/OMPD_teams);
4712 };
4713 emitCommonOMPTeamsDirective(*this, S, OMPD_distribute_parallel_for, CodeGen);
4714 emitPostUpdateForReductionClause(*this, S,
4715 [](CodeGenFunction &) { return nullptr; });
4716}
4717
4718void CodeGenFunction::EmitOMPTeamsDistributeParallelForSimdDirective(
4719 const OMPTeamsDistributeParallelForSimdDirective &S) {
4720 auto &&CodeGenDistribute = [&S](CodeGenFunction &CGF, PrePostActionTy &) {
4721 CGF.EmitOMPDistributeLoop(S, emitInnerParallelForWhenCombined,
4722 S.getDistInc());
4723 };
4724
4725 // Emit teams region as a standalone region.
4726 auto &&CodeGen = [&S, &CodeGenDistribute](CodeGenFunction &CGF,
4727 PrePostActionTy &Action) {
4728 Action.Enter(CGF);
4729 OMPPrivateScope PrivateScope(CGF);
4730 CGF.EmitOMPReductionClauseInit(S, PrivateScope);
4731 (void)PrivateScope.Privatize();
4732 CGF.CGM.getOpenMPRuntime().emitInlinedDirective(
4733 CGF, OMPD_distribute, CodeGenDistribute, /*HasCancel=*/false);
4734 CGF.EmitOMPReductionClauseFinal(S, /*ReductionKind=*/OMPD_teams);
4735 };
4736 emitCommonOMPTeamsDirective(*this, S, OMPD_distribute_parallel_for_simd,
4737 CodeGen);
4738 emitPostUpdateForReductionClause(*this, S,
4739 [](CodeGenFunction &) { return nullptr; });
4740}
4741
4742static void emitTargetTeamsDistributeParallelForRegion(
4743 CodeGenFunction &CGF, const OMPTargetTeamsDistributeParallelForDirective &S,
4744 PrePostActionTy &Action) {
4745 Action.Enter(CGF);
4746 auto &&CodeGenDistribute = [&S](CodeGenFunction &CGF, PrePostActionTy &) {
4747 CGF.EmitOMPDistributeLoop(S, emitInnerParallelForWhenCombined,
4748 S.getDistInc());
4749 };
4750
4751 // Emit teams region as a standalone region.
4752 auto &&CodeGenTeams = [&S, &CodeGenDistribute](CodeGenFunction &CGF,
4753 PrePostActionTy &Action) {
4754 Action.Enter(CGF);
4755 CodeGenFunction::OMPPrivateScope PrivateScope(CGF);
4756 CGF.EmitOMPReductionClauseInit(S, PrivateScope);
4757 (void)PrivateScope.Privatize();
4758 CGF.CGM.getOpenMPRuntime().emitInlinedDirective(
4759 CGF, OMPD_distribute, CodeGenDistribute, /*HasCancel=*/false);
4760 CGF.EmitOMPReductionClauseFinal(S, /*ReductionKind=*/OMPD_teams);
4761 };
4762
4763 emitCommonOMPTeamsDirective(CGF, S, OMPD_distribute_parallel_for,
4764 CodeGenTeams);
4765 emitPostUpdateForReductionClause(CGF, S,
4766 [](CodeGenFunction &) { return nullptr; });
4767}
4768
4769void CodeGenFunction::EmitOMPTargetTeamsDistributeParallelForDeviceFunction(
4770 CodeGenModule &CGM, StringRef ParentName,
4771 const OMPTargetTeamsDistributeParallelForDirective &S) {
4772 // Emit SPMD target teams distribute parallel for region as a standalone
4773 // region.
4774 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
4775 emitTargetTeamsDistributeParallelForRegion(CGF, S, Action);
4776 };
4777 llvm::Function *Fn;
4778 llvm::Constant *Addr;
4779 // Emit target region as a standalone region.
4780 CGM.getOpenMPRuntime().emitTargetOutlinedFunction(
4781 S, ParentName, Fn, Addr, /*IsOffloadEntry=*/true, CodeGen);
4782 assert(Fn && Addr && "Target device function emission failed.")((Fn && Addr && "Target device function emission failed."
) ? static_cast<void> (0) : __assert_fail ("Fn && Addr && \"Target device function emission failed.\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 4782, __PRETTY_FUNCTION__))
;
4783}
4784
4785void CodeGenFunction::EmitOMPTargetTeamsDistributeParallelForDirective(
4786 const OMPTargetTeamsDistributeParallelForDirective &S) {
4787 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
4788 emitTargetTeamsDistributeParallelForRegion(CGF, S, Action);
4789 };
4790 emitCommonOMPTargetDirective(*this, S, CodeGen);
4791}
4792
4793static void emitTargetTeamsDistributeParallelForSimdRegion(
4794 CodeGenFunction &CGF,
4795 const OMPTargetTeamsDistributeParallelForSimdDirective &S,
4796 PrePostActionTy &Action) {
4797 Action.Enter(CGF);
4798 auto &&CodeGenDistribute = [&S](CodeGenFunction &CGF, PrePostActionTy &) {
4799 CGF.EmitOMPDistributeLoop(S, emitInnerParallelForWhenCombined,
4800 S.getDistInc());
4801 };
4802
4803 // Emit teams region as a standalone region.
4804 auto &&CodeGenTeams = [&S, &CodeGenDistribute](CodeGenFunction &CGF,
4805 PrePostActionTy &Action) {
4806 Action.Enter(CGF);
4807 CodeGenFunction::OMPPrivateScope PrivateScope(CGF);
4808 CGF.EmitOMPReductionClauseInit(S, PrivateScope);
4809 (void)PrivateScope.Privatize();
4810 CGF.CGM.getOpenMPRuntime().emitInlinedDirective(
4811 CGF, OMPD_distribute, CodeGenDistribute, /*HasCancel=*/false);
4812 CGF.EmitOMPReductionClauseFinal(S, /*ReductionKind=*/OMPD_teams);
4813 };
4814
4815 emitCommonOMPTeamsDirective(CGF, S, OMPD_distribute_parallel_for_simd,
4816 CodeGenTeams);
4817 emitPostUpdateForReductionClause(CGF, S,
4818 [](CodeGenFunction &) { return nullptr; });
4819}
4820
4821void CodeGenFunction::EmitOMPTargetTeamsDistributeParallelForSimdDeviceFunction(
4822 CodeGenModule &CGM, StringRef ParentName,
4823 const OMPTargetTeamsDistributeParallelForSimdDirective &S) {
4824 // Emit SPMD target teams distribute parallel for simd region as a standalone
4825 // region.
4826 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
4827 emitTargetTeamsDistributeParallelForSimdRegion(CGF, S, Action);
4828 };
4829 llvm::Function *Fn;
4830 llvm::Constant *Addr;
4831 // Emit target region as a standalone region.
4832 CGM.getOpenMPRuntime().emitTargetOutlinedFunction(
4833 S, ParentName, Fn, Addr, /*IsOffloadEntry=*/true, CodeGen);
4834 assert(Fn && Addr && "Target device function emission failed.")((Fn && Addr && "Target device function emission failed."
) ? static_cast<void> (0) : __assert_fail ("Fn && Addr && \"Target device function emission failed.\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 4834, __PRETTY_FUNCTION__))
;
4835}
4836
4837void CodeGenFunction::EmitOMPTargetTeamsDistributeParallelForSimdDirective(
4838 const OMPTargetTeamsDistributeParallelForSimdDirective &S) {
4839 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
4840 emitTargetTeamsDistributeParallelForSimdRegion(CGF, S, Action);
4841 };
4842 emitCommonOMPTargetDirective(*this, S, CodeGen);
4843}
4844
4845void CodeGenFunction::EmitOMPCancellationPointDirective(
4846 const OMPCancellationPointDirective &S) {
4847 CGM.getOpenMPRuntime().emitCancellationPointCall(*this, S.getBeginLoc(),
4848 S.getCancelRegion());
4849}
4850
4851void CodeGenFunction::EmitOMPCancelDirective(const OMPCancelDirective &S) {
4852 const Expr *IfCond = nullptr;
4853 for (const auto *C : S.getClausesOfKind<OMPIfClause>()) {
4854 if (C->getNameModifier() == OMPD_unknown ||
4855 C->getNameModifier() == OMPD_cancel) {
4856 IfCond = C->getCondition();
4857 break;
4858 }
4859 }
4860 if (llvm::OpenMPIRBuilder *OMPBuilder = CGM.getOpenMPIRBuilder()) {
4861 // TODO: This check is necessary as we only generate `omp parallel` through
4862 // the OpenMPIRBuilder for now.
4863 if (S.getCancelRegion() == OMPD_parallel) {
4864 llvm::Value *IfCondition = nullptr;
4865 if (IfCond)
4866 IfCondition = EmitScalarExpr(IfCond,
4867 /*IgnoreResultAssign=*/true);
4868 return Builder.restoreIP(
4869 OMPBuilder->CreateCancel(Builder, IfCondition, S.getCancelRegion()));
4870 }
4871 }
4872
4873 CGM.getOpenMPRuntime().emitCancelCall(*this, S.getBeginLoc(), IfCond,
4874 S.getCancelRegion());
4875}
4876
4877CodeGenFunction::JumpDest
4878CodeGenFunction::getOMPCancelDestination(OpenMPDirectiveKind Kind) {
4879 if (Kind == OMPD_parallel || Kind == OMPD_task ||
4880 Kind == OMPD_target_parallel)
4881 return ReturnBlock;
4882 assert(Kind == OMPD_for || Kind == OMPD_section || Kind == OMPD_sections ||((Kind == OMPD_for || Kind == OMPD_section || Kind == OMPD_sections
|| Kind == OMPD_parallel_sections || Kind == OMPD_parallel_for
|| Kind == OMPD_distribute_parallel_for || Kind == OMPD_target_parallel_for
|| Kind == OMPD_teams_distribute_parallel_for || Kind == OMPD_target_teams_distribute_parallel_for
) ? static_cast<void> (0) : __assert_fail ("Kind == OMPD_for || Kind == OMPD_section || Kind == OMPD_sections || Kind == OMPD_parallel_sections || Kind == OMPD_parallel_for || Kind == OMPD_distribute_parallel_for || Kind == OMPD_target_parallel_for || Kind == OMPD_teams_distribute_parallel_for || Kind == OMPD_target_teams_distribute_parallel_for"
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 4887, __PRETTY_FUNCTION__))
4883 Kind == OMPD_parallel_sections || Kind == OMPD_parallel_for ||((Kind == OMPD_for || Kind == OMPD_section || Kind == OMPD_sections
|| Kind == OMPD_parallel_sections || Kind == OMPD_parallel_for
|| Kind == OMPD_distribute_parallel_for || Kind == OMPD_target_parallel_for
|| Kind == OMPD_teams_distribute_parallel_for || Kind == OMPD_target_teams_distribute_parallel_for
) ? static_cast<void> (0) : __assert_fail ("Kind == OMPD_for || Kind == OMPD_section || Kind == OMPD_sections || Kind == OMPD_parallel_sections || Kind == OMPD_parallel_for || Kind == OMPD_distribute_parallel_for || Kind == OMPD_target_parallel_for || Kind == OMPD_teams_distribute_parallel_for || Kind == OMPD_target_teams_distribute_parallel_for"
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 4887, __PRETTY_FUNCTION__))
4884 Kind == OMPD_distribute_parallel_for ||((Kind == OMPD_for || Kind == OMPD_section || Kind == OMPD_sections
|| Kind == OMPD_parallel_sections || Kind == OMPD_parallel_for
|| Kind == OMPD_distribute_parallel_for || Kind == OMPD_target_parallel_for
|| Kind == OMPD_teams_distribute_parallel_for || Kind == OMPD_target_teams_distribute_parallel_for
) ? static_cast<void> (0) : __assert_fail ("Kind == OMPD_for || Kind == OMPD_section || Kind == OMPD_sections || Kind == OMPD_parallel_sections || Kind == OMPD_parallel_for || Kind == OMPD_distribute_parallel_for || Kind == OMPD_target_parallel_for || Kind == OMPD_teams_distribute_parallel_for || Kind == OMPD_target_teams_distribute_parallel_for"
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 4887, __PRETTY_FUNCTION__))
4885 Kind == OMPD_target_parallel_for ||((Kind == OMPD_for || Kind == OMPD_section || Kind == OMPD_sections
|| Kind == OMPD_parallel_sections || Kind == OMPD_parallel_for
|| Kind == OMPD_distribute_parallel_for || Kind == OMPD_target_parallel_for
|| Kind == OMPD_teams_distribute_parallel_for || Kind == OMPD_target_teams_distribute_parallel_for
) ? static_cast<void> (0) : __assert_fail ("Kind == OMPD_for || Kind == OMPD_section || Kind == OMPD_sections || Kind == OMPD_parallel_sections || Kind == OMPD_parallel_for || Kind == OMPD_distribute_parallel_for || Kind == OMPD_target_parallel_for || Kind == OMPD_teams_distribute_parallel_for || Kind == OMPD_target_teams_distribute_parallel_for"
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 4887, __PRETTY_FUNCTION__))
4886 Kind == OMPD_teams_distribute_parallel_for ||((Kind == OMPD_for || Kind == OMPD_section || Kind == OMPD_sections
|| Kind == OMPD_parallel_sections || Kind == OMPD_parallel_for
|| Kind == OMPD_distribute_parallel_for || Kind == OMPD_target_parallel_for
|| Kind == OMPD_teams_distribute_parallel_for || Kind == OMPD_target_teams_distribute_parallel_for
) ? static_cast<void> (0) : __assert_fail ("Kind == OMPD_for || Kind == OMPD_section || Kind == OMPD_sections || Kind == OMPD_parallel_sections || Kind == OMPD_parallel_for || Kind == OMPD_distribute_parallel_for || Kind == OMPD_target_parallel_for || Kind == OMPD_teams_distribute_parallel_for || Kind == OMPD_target_teams_distribute_parallel_for"
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 4887, __PRETTY_FUNCTION__))
4887 Kind == OMPD_target_teams_distribute_parallel_for)((Kind == OMPD_for || Kind == OMPD_section || Kind == OMPD_sections
|| Kind == OMPD_parallel_sections || Kind == OMPD_parallel_for
|| Kind == OMPD_distribute_parallel_for || Kind == OMPD_target_parallel_for
|| Kind == OMPD_teams_distribute_parallel_for || Kind == OMPD_target_teams_distribute_parallel_for
) ? static_cast<void> (0) : __assert_fail ("Kind == OMPD_for || Kind == OMPD_section || Kind == OMPD_sections || Kind == OMPD_parallel_sections || Kind == OMPD_parallel_for || Kind == OMPD_distribute_parallel_for || Kind == OMPD_target_parallel_for || Kind == OMPD_teams_distribute_parallel_for || Kind == OMPD_target_teams_distribute_parallel_for"
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 4887, __PRETTY_FUNCTION__))
;
4888 return OMPCancelStack.getExitBlock();
4889}
4890
4891void CodeGenFunction::EmitOMPUseDevicePtrClause(
4892 const OMPClause &NC, OMPPrivateScope &PrivateScope,
4893 const llvm::DenseMap<const ValueDecl *, Address> &CaptureDeviceAddrMap) {
4894 const auto &C = cast<OMPUseDevicePtrClause>(NC);
4895 auto OrigVarIt = C.varlist_begin();
4896 auto InitIt = C.inits().begin();
4897 for (const Expr *PvtVarIt : C.private_copies()) {
4898 const auto *OrigVD = cast<VarDecl>(cast<DeclRefExpr>(*OrigVarIt)->getDecl());
4899 const auto *InitVD = cast<VarDecl>(cast<DeclRefExpr>(*InitIt)->getDecl());
4900 const auto *PvtVD = cast<VarDecl>(cast<DeclRefExpr>(PvtVarIt)->getDecl());
4901
4902 // In order to identify the right initializer we need to match the
4903 // declaration used by the mapping logic. In some cases we may get
4904 // OMPCapturedExprDecl that refers to the original declaration.
4905 const ValueDecl *MatchingVD = OrigVD;
4906 if (const auto *OED = dyn_cast<OMPCapturedExprDecl>(MatchingVD)) {
4907 // OMPCapturedExprDecl are used to privative fields of the current
4908 // structure.
4909 const auto *ME = cast<MemberExpr>(OED->getInit());
4910 assert(isa<CXXThisExpr>(ME->getBase()) &&((isa<CXXThisExpr>(ME->getBase()) && "Base should be the current struct!"
) ? static_cast<void> (0) : __assert_fail ("isa<CXXThisExpr>(ME->getBase()) && \"Base should be the current struct!\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 4911, __PRETTY_FUNCTION__))
4911 "Base should be the current struct!")((isa<CXXThisExpr>(ME->getBase()) && "Base should be the current struct!"
) ? static_cast<void> (0) : __assert_fail ("isa<CXXThisExpr>(ME->getBase()) && \"Base should be the current struct!\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 4911, __PRETTY_FUNCTION__))
;
4912 MatchingVD = ME->getMemberDecl();
4913 }
4914
4915 // If we don't have information about the current list item, move on to
4916 // the next one.
4917 auto InitAddrIt = CaptureDeviceAddrMap.find(MatchingVD);
4918 if (InitAddrIt == CaptureDeviceAddrMap.end())
4919 continue;
4920
4921 bool IsRegistered = PrivateScope.addPrivate(OrigVD, [this, OrigVD,
4922 InitAddrIt, InitVD,
4923 PvtVD]() {
4924 // Initialize the temporary initialization variable with the address we
4925 // get from the runtime library. We have to cast the source address
4926 // because it is always a void *. References are materialized in the
4927 // privatization scope, so the initialization here disregards the fact
4928 // the original variable is a reference.
4929 QualType AddrQTy =
4930 getContext().getPointerType(OrigVD->getType().getNonReferenceType());
4931 llvm::Type *AddrTy = ConvertTypeForMem(AddrQTy);
4932 Address InitAddr = Builder.CreateBitCast(InitAddrIt->second, AddrTy);
4933 setAddrOfLocalVar(InitVD, InitAddr);
4934
4935 // Emit private declaration, it will be initialized by the value we
4936 // declaration we just added to the local declarations map.
4937 EmitDecl(*PvtVD);
4938
4939 // The initialization variables reached its purpose in the emission
4940 // of the previous declaration, so we don't need it anymore.
4941 LocalDeclMap.erase(InitVD);
4942
4943 // Return the address of the private variable.
4944 return GetAddrOfLocalVar(PvtVD);
4945 });
4946 assert(IsRegistered && "firstprivate var already registered as private")((IsRegistered && "firstprivate var already registered as private"
) ? static_cast<void> (0) : __assert_fail ("IsRegistered && \"firstprivate var already registered as private\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 4946, __PRETTY_FUNCTION__))
;
4947 // Silence the warning about unused variable.
4948 (void)IsRegistered;
4949
4950 ++OrigVarIt;
4951 ++InitIt;
4952 }
4953}
4954
4955// Generate the instructions for '#pragma omp target data' directive.
4956void CodeGenFunction::EmitOMPTargetDataDirective(
4957 const OMPTargetDataDirective &S) {
4958 CGOpenMPRuntime::TargetDataInfo Info(/*RequiresDevicePointerInfo=*/true);
4959
4960 // Create a pre/post action to signal the privatization of the device pointer.
4961 // This action can be replaced by the OpenMP runtime code generation to
4962 // deactivate privatization.
4963 bool PrivatizeDevicePointers = false;
4964 class DevicePointerPrivActionTy : public PrePostActionTy {
4965 bool &PrivatizeDevicePointers;
4966
4967 public:
4968 explicit DevicePointerPrivActionTy(bool &PrivatizeDevicePointers)
4969 : PrePostActionTy(), PrivatizeDevicePointers(PrivatizeDevicePointers) {}
4970 void Enter(CodeGenFunction &CGF) override {
4971 PrivatizeDevicePointers = true;
4972 }
4973 };
4974 DevicePointerPrivActionTy PrivAction(PrivatizeDevicePointers);
4975
4976 auto &&CodeGen = [&S, &Info, &PrivatizeDevicePointers](
4977 CodeGenFunction &CGF, PrePostActionTy &Action) {
4978 auto &&InnermostCodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &) {
4979 CGF.EmitStmt(S.getInnermostCapturedStmt()->getCapturedStmt());
4980 };
4981
4982 // Codegen that selects whether to generate the privatization code or not.
4983 auto &&PrivCodeGen = [&S, &Info, &PrivatizeDevicePointers,
4984 &InnermostCodeGen](CodeGenFunction &CGF,
4985 PrePostActionTy &Action) {
4986 RegionCodeGenTy RCG(InnermostCodeGen);
4987 PrivatizeDevicePointers = false;
4988
4989 // Call the pre-action to change the status of PrivatizeDevicePointers if
4990 // needed.
4991 Action.Enter(CGF);
4992
4993 if (PrivatizeDevicePointers) {
4994 OMPPrivateScope PrivateScope(CGF);
4995 // Emit all instances of the use_device_ptr clause.
4996 for (const auto *C : S.getClausesOfKind<OMPUseDevicePtrClause>())
4997 CGF.EmitOMPUseDevicePtrClause(*C, PrivateScope,
4998 Info.CaptureDeviceAddrMap);
4999 (void)PrivateScope.Privatize();
5000 RCG(CGF);
5001 } else {
5002 RCG(CGF);
5003 }
5004 };
5005
5006 // Forward the provided action to the privatization codegen.
5007 RegionCodeGenTy PrivRCG(PrivCodeGen);
5008 PrivRCG.setAction(Action);
5009
5010 // Notwithstanding the body of the region is emitted as inlined directive,
5011 // we don't use an inline scope as changes in the references inside the
5012 // region are expected to be visible outside, so we do not privative them.
5013 OMPLexicalScope Scope(CGF, S);
5014 CGF.CGM.getOpenMPRuntime().emitInlinedDirective(CGF, OMPD_target_data,
5015 PrivRCG);
5016 };
5017
5018 RegionCodeGenTy RCG(CodeGen);
5019
5020 // If we don't have target devices, don't bother emitting the data mapping
5021 // code.
5022 if (CGM.getLangOpts().OMPTargetTriples.empty()) {
5023 RCG(*this);
5024 return;
5025 }
5026
5027 // Check if we have any if clause associated with the directive.
5028 const Expr *IfCond = nullptr;
5029 if (const auto *C = S.getSingleClause<OMPIfClause>())
5030 IfCond = C->getCondition();
5031
5032 // Check if we have any device clause associated with the directive.
5033 const Expr *Device = nullptr;
5034 if (const auto *C = S.getSingleClause<OMPDeviceClause>())
5035 Device = C->getDevice();
5036
5037 // Set the action to signal privatization of device pointers.
5038 RCG.setAction(PrivAction);
5039
5040 // Emit region code.
5041 CGM.getOpenMPRuntime().emitTargetDataCalls(*this, S, IfCond, Device, RCG,
5042 Info);
5043}
5044
5045void CodeGenFunction::EmitOMPTargetEnterDataDirective(
5046 const OMPTargetEnterDataDirective &S) {
5047 // If we don't have target devices, don't bother emitting the data mapping
5048 // code.
5049 if (CGM.getLangOpts().OMPTargetTriples.empty())
5050 return;
5051
5052 // Check if we have any if clause associated with the directive.
5053 const Expr *IfCond = nullptr;
5054 if (const auto *C = S.getSingleClause<OMPIfClause>())
5055 IfCond = C->getCondition();
5056
5057 // Check if we have any device clause associated with the directive.
5058 const Expr *Device = nullptr;
5059 if (const auto *C = S.getSingleClause<OMPDeviceClause>())
5060 Device = C->getDevice();
5061
5062 OMPLexicalScope Scope(*this, S, OMPD_task);
5063 CGM.getOpenMPRuntime().emitTargetDataStandAloneCall(*this, S, IfCond, Device);
5064}
5065
5066void CodeGenFunction::EmitOMPTargetExitDataDirective(
5067 const OMPTargetExitDataDirective &S) {
5068 // If we don't have target devices, don't bother emitting the data mapping
5069 // code.
5070 if (CGM.getLangOpts().OMPTargetTriples.empty())
5071 return;
5072
5073 // Check if we have any if clause associated with the directive.
5074 const Expr *IfCond = nullptr;
5075 if (const auto *C = S.getSingleClause<OMPIfClause>())
5076 IfCond = C->getCondition();
5077
5078 // Check if we have any device clause associated with the directive.
5079 const Expr *Device = nullptr;
5080 if (const auto *C = S.getSingleClause<OMPDeviceClause>())
5081 Device = C->getDevice();
5082
5083 OMPLexicalScope Scope(*this, S, OMPD_task);
5084 CGM.getOpenMPRuntime().emitTargetDataStandAloneCall(*this, S, IfCond, Device);
5085}
5086
5087static void emitTargetParallelRegion(CodeGenFunction &CGF,
5088 const OMPTargetParallelDirective &S,
5089 PrePostActionTy &Action) {
5090 // Get the captured statement associated with the 'parallel' region.
5091 const CapturedStmt *CS = S.getCapturedStmt(OMPD_parallel);
5092 Action.Enter(CGF);
5093 auto &&CodeGen = [&S, CS](CodeGenFunction &CGF, PrePostActionTy &Action) {
5094 Action.Enter(CGF);
5095 CodeGenFunction::OMPPrivateScope PrivateScope(CGF);
5096 (void)CGF.EmitOMPFirstprivateClause(S, PrivateScope);
5097 CGF.EmitOMPPrivateClause(S, PrivateScope);
5098 CGF.EmitOMPReductionClauseInit(S, PrivateScope);
5099 (void)PrivateScope.Privatize();
5100 if (isOpenMPTargetExecutionDirective(S.getDirectiveKind()))
5101 CGF.CGM.getOpenMPRuntime().adjustTargetSpecificDataForLambdas(CGF, S);
5102 // TODO: Add support for clauses.
5103 CGF.EmitStmt(CS->getCapturedStmt());
5104 CGF.EmitOMPReductionClauseFinal(S, /*ReductionKind=*/OMPD_parallel);
5105 };
5106 emitCommonOMPParallelDirective(CGF, S, OMPD_parallel, CodeGen,
5107 emitEmptyBoundParameters);
5108 emitPostUpdateForReductionClause(CGF, S,
5109 [](CodeGenFunction &) { return nullptr; });
5110}
5111
5112void CodeGenFunction::EmitOMPTargetParallelDeviceFunction(
5113 CodeGenModule &CGM, StringRef ParentName,
5114 const OMPTargetParallelDirective &S) {
5115 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
5116 emitTargetParallelRegion(CGF, S, Action);
5117 };
5118 llvm::Function *Fn;
5119 llvm::Constant *Addr;
5120 // Emit target region as a standalone region.
5121 CGM.getOpenMPRuntime().emitTargetOutlinedFunction(
5122 S, ParentName, Fn, Addr, /*IsOffloadEntry=*/true, CodeGen);
5123 assert(Fn && Addr && "Target device function emission failed.")((Fn && Addr && "Target device function emission failed."
) ? static_cast<void> (0) : __assert_fail ("Fn && Addr && \"Target device function emission failed.\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 5123, __PRETTY_FUNCTION__))
;
5124}
5125
5126void CodeGenFunction::EmitOMPTargetParallelDirective(
5127 const OMPTargetParallelDirective &S) {
5128 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
5129 emitTargetParallelRegion(CGF, S, Action);
5130 };
5131 emitCommonOMPTargetDirective(*this, S, CodeGen);
5132}
5133
5134static void emitTargetParallelForRegion(CodeGenFunction &CGF,
5135 const OMPTargetParallelForDirective &S,
5136 PrePostActionTy &Action) {
5137 Action.Enter(CGF);
5138 // Emit directive as a combined directive that consists of two implicit
5139 // directives: 'parallel' with 'for' directive.
5140 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
5141 Action.Enter(CGF);
5142 CodeGenFunction::OMPCancelStackRAII CancelRegion(
5143 CGF, OMPD_target_parallel_for, S.hasCancel());
5144 CGF.EmitOMPWorksharingLoop(S, S.getEnsureUpperBound(), emitForLoopBounds,
5145 emitDispatchForLoopBounds);
5146 };
5147 emitCommonOMPParallelDirective(CGF, S, OMPD_for, CodeGen,
5148 emitEmptyBoundParameters);
5149}
5150
5151void CodeGenFunction::EmitOMPTargetParallelForDeviceFunction(
5152 CodeGenModule &CGM, StringRef ParentName,
5153 const OMPTargetParallelForDirective &S) {
5154 // Emit SPMD target parallel for region as a standalone region.
5155 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
5156 emitTargetParallelForRegion(CGF, S, Action);
5157 };
5158 llvm::Function *Fn;
5159 llvm::Constant *Addr;
5160 // Emit target region as a standalone region.
5161 CGM.getOpenMPRuntime().emitTargetOutlinedFunction(
5162 S, ParentName, Fn, Addr, /*IsOffloadEntry=*/true, CodeGen);
5163 assert(Fn && Addr && "Target device function emission failed.")((Fn && Addr && "Target device function emission failed."
) ? static_cast<void> (0) : __assert_fail ("Fn && Addr && \"Target device function emission failed.\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 5163, __PRETTY_FUNCTION__))
;
5164}
5165
5166void CodeGenFunction::EmitOMPTargetParallelForDirective(
5167 const OMPTargetParallelForDirective &S) {
5168 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
5169 emitTargetParallelForRegion(CGF, S, Action);
5170 };
5171 emitCommonOMPTargetDirective(*this, S, CodeGen);
5172}
5173
5174static void
5175emitTargetParallelForSimdRegion(CodeGenFunction &CGF,
5176 const OMPTargetParallelForSimdDirective &S,
5177 PrePostActionTy &Action) {
5178 Action.Enter(CGF);
5179 // Emit directive as a combined directive that consists of two implicit
5180 // directives: 'parallel' with 'for' directive.
5181 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
5182 Action.Enter(CGF);
5183 CGF.EmitOMPWorksharingLoop(S, S.getEnsureUpperBound(), emitForLoopBounds,
5184 emitDispatchForLoopBounds);
5185 };
5186 emitCommonOMPParallelDirective(CGF, S, OMPD_simd, CodeGen,
5187 emitEmptyBoundParameters);
5188}
5189
5190void CodeGenFunction::EmitOMPTargetParallelForSimdDeviceFunction(
5191 CodeGenModule &CGM, StringRef ParentName,
5192 const OMPTargetParallelForSimdDirective &S) {
5193 // Emit SPMD target parallel for region as a standalone region.
5194 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
5195 emitTargetParallelForSimdRegion(CGF, S, Action);
5196 };
5197 llvm::Function *Fn;
5198 llvm::Constant *Addr;
5199 // Emit target region as a standalone region.
5200 CGM.getOpenMPRuntime().emitTargetOutlinedFunction(
5201 S, ParentName, Fn, Addr, /*IsOffloadEntry=*/true, CodeGen);
5202 assert(Fn && Addr && "Target device function emission failed.")((Fn && Addr && "Target device function emission failed."
) ? static_cast<void> (0) : __assert_fail ("Fn && Addr && \"Target device function emission failed.\""
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 5202, __PRETTY_FUNCTION__))
;
5203}
5204
5205void CodeGenFunction::EmitOMPTargetParallelForSimdDirective(
5206 const OMPTargetParallelForSimdDirective &S) {
5207 auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) {
5208 emitTargetParallelForSimdRegion(CGF, S, Action);
5209 };
5210 emitCommonOMPTargetDirective(*this, S, CodeGen);
5211}
5212
5213/// Emit a helper variable and return corresponding lvalue.
5214static void mapParam(CodeGenFunction &CGF, const DeclRefExpr *Helper,
5215 const ImplicitParamDecl *PVD,
5216 CodeGenFunction::OMPPrivateScope &Privates) {
5217 const auto *VDecl = cast<VarDecl>(Helper->getDecl());
5218 Privates.addPrivate(VDecl,
5219 [&CGF, PVD]() { return CGF.GetAddrOfLocalVar(PVD); });
5220}
5221
5222void CodeGenFunction::EmitOMPTaskLoopBasedDirective(const OMPLoopDirective &S) {
5223 assert(isOpenMPTaskLoopDirective(S.getDirectiveKind()))((isOpenMPTaskLoopDirective(S.getDirectiveKind())) ? static_cast
<void> (0) : __assert_fail ("isOpenMPTaskLoopDirective(S.getDirectiveKind())"
, "/build/llvm-toolchain-snapshot-10~++20200110111110+a1cc19b5814/clang/lib/CodeGen/CGStmtOpenMP.cpp"
, 5223, __PRETTY_FUNCTION__))
;
5224 // Emit outlined function for task construct.
5225 const CapturedStmt *CS = S.getCapturedStmt(OMPD_taskloop);
5226 Address CapturedStruct = GenerateCapturedStmtArgument(*CS);
5227 QualType SharedsTy = getContext().getRecordType(CS->getCapturedRecordDecl());
5228 const Expr *IfCond = nullptr;
5229 for (const auto *C : S.getClausesOfKind<OMPIfClause>()) {
5230 if (C->getNameModifier() == OMPD_unknown ||
5231 C->getNameModifier() == OMPD_taskloop) {
5232 IfCond = C->getCondition();
5233 break;
5234 }
5235 }
5236
5237 OMPTaskDataTy Data;
5238 // Check if taskloop must be emitted without taskgroup.
5239 Data.Nogroup = S.getSingleClause<OMPNogroupClause>();
5240 // TODO: Check if we should emit tied or untied task.
5241 Data.Tied = true;
5242 // Set scheduling for taskloop
5243 if (const auto* Clause = S.getSingleClause<OMPGrainsizeClause>()) {
5244 // grainsize clause
5245 Data.Schedule.setInt(/*IntVal=*/false);
5246 Data.Schedule.setPointer(EmitScalarExpr(Clause->getGrainsize()));
5247 } else if (const auto* Clause = S.getSingleClause<OMPNumTasksClause>()) {
5248 // num_tasks clause
5249 Data.Schedule.setInt(/*IntVal=*/true);
5250 Data.Schedule.setPointer(EmitScalarExpr(Clause->getNumTasks()));
5251 }
5252
5253 auto &&BodyGen = [CS, &S](CodeGenFunction &CGF, PrePostActionTy &) {
5254 // if (PreCond) {
5255 // for (IV in 0..LastIteration) BODY;
5256 // <Final counter/linear vars updates>;
5257 // }
5258 //
5259
5260 // Emit: if (PreCond) - begin.
5261 // If the condition constant folds and can be elided, avoid emitting the
5262 // whole loop.
5263 bool CondConstant;
5264 llvm::BasicBlock *ContBlock = nullptr;
5265 OMPLoopScope PreInitScope(CGF, S);
5266 if (CGF.ConstantFoldsToSimpleInteger(S.getPreCond(), CondConstant)) {
5267 if (!CondConstant)
5268 return;
5269 } else {
5270 llvm::BasicBlock *ThenBlock = CGF.createBasicBlock("taskloop.if.then");
5271 ContBlock = CGF.createBasicBlock("taskloop.if.end");
5272 emitPreCond(CGF, S, S.getPreCond(), ThenBlock, ContBlock,
5273 CGF.getProfileCount(&S));
5274 CGF.EmitBlock(ThenBlock);
5275 CGF.incrementProfileCounter(&S);
5276 }
5277
5278 (void)CGF.EmitOMPLinearClauseInit(S);
5279
5280 OMPPrivateScope LoopScope(CGF);
5281 // Emit helper vars inits.
5282 enum { LowerBound = 5, UpperBound, Stride, LastIter };
5283 auto *I = CS->getCapturedDecl()->param_begin();
5284 auto *LBP = std::next(I, LowerBound);
5285 auto *UBP = std::next(I, UpperBound);
5286 auto *STP = std::next(I, Stride);
5287 auto *LIP = std::next(I, LastIter);
5288 mapParam(CGF, cast<DeclRefExpr>(S.getLowerBoundVariable()), *LBP,
5289 LoopScope);
5290 mapParam(CGF, cast<DeclRefExpr>(S.getUpperBoundVariable()), *UBP,
5291 LoopScope);
5292 mapParam(CGF, cast<DeclRefExpr>(S.getStrideVariable()), *STP, LoopScope);
5293 mapParam(CGF, cast<DeclRefExpr>(S.getIsLastIterVariable()), *LIP,
5294 LoopScope);
5295 CGF.EmitOMPPrivateLoopCounters(S, LoopScope);
5296 CGF.EmitOMPLinearClause(S, LoopScope);
5297 bool HasLastprivateClause = CGF.EmitOMPLastprivateClauseInit(S, LoopScope);
5298 (void)LoopScope.Privatize();
5299 // Emit the loop iteration variable.
5300 const Expr *IVExpr = S.getIterationVariable();
5301 const auto *IVDecl = cast<VarDecl>(cast<DeclRefExpr>(IVExpr)->getDecl());
5302 CGF.EmitVarDecl(*IVDecl);
5303 CGF.EmitIgnoredExpr(S.getInit());
5304
5305 // Emit the iterations count variable.
5306 // If it is not a variable, Sema decided to calculate iterations count on
5307 // each iteration (e.g., it is foldable into a constant).
5308 if (const auto *LIExpr = dyn_cast<DeclRefExpr>(S.getLastIteration())) {
5309 CGF.EmitVarDecl(*cast<VarDecl>(LIExpr->getDecl()));
5310 // Emit calculation of the iterations count.
5311 CGF.EmitIgnoredExpr(S.getCalcLastIteration());
5312 }
5313
5314 {
5315 OMPLexicalScope Scope(CGF, S, OMPD_taskloop, /*EmitPreInitStmt=*/false);
5316 emitCommonSimdLoop(
5317 CGF, S,
5318 [&S](CodeGenFunction &CGF, PrePostActionTy &) {
5319 if (isOpenMPSimdDirective(S.getDirectiveKind()))
5320 CGF.EmitOMPSimdInit(S);
5321 },
5322 [&S, &LoopScope](CodeGenFunction &CGF, PrePostActionTy &) {
5323 CGF.EmitOMPInnerLoop(
5324 S, LoopScope.requiresCleanups(), S.getCond(), S.getInc(),
5325 [&S](CodeGenFunction &CGF) {
5326 CGF.EmitOMPLoopBody(S, CodeGenFunction::JumpDest());
5327 CGF.EmitStopPoint(&S);
5328 },
5329 [](CodeGenFunction &) {});
5330 });
5331 }
5332 // Emit: if (PreCond) - end.
5333 if (ContBlock) {
5334 CGF.EmitBranch(ContBlock);
5335 CGF.EmitBlock(ContBlock, true);
5336 }
5337 // Emit final copy of the lastprivate variables if IsLastIter != 0.
5338 if (HasLastprivateClause) {
5339 CGF.EmitOMPLastprivateClauseFinal(
5340 S, isOpenMPSimdDirective(S.getDirectiveKind()),
5341 CGF.Builder.CreateIsNotNull(CGF.EmitLoadOfScalar(
5342 CGF.GetAddrOfLocalVar(*LIP), /*Volatile=*/false,
5343 (*LIP)->getType(), S.getBeginLoc())));
5344 }
5345 CGF.EmitOMPLinearClauseFinal(S, [LIP, &S](CodeGenFunction &CGF) {
5346 return CGF.Builder.CreateIsNotNull(
5347 CGF.EmitLoadOfScalar(CGF.GetAddrOfLocalVar(*LIP), /*Volatile=*/false,
5348 (*LIP)->getType(), S.getBeginLoc()));
5349 });
5350 };
5351 auto &&TaskGen = [&S, SharedsTy, CapturedStruct,
5352 IfCond](CodeGenFunction &CGF, llvm::Function *OutlinedFn,
5353 const OMPTaskDataTy &Data) {
5354 auto &&CodeGen = [&S, OutlinedFn, SharedsTy, CapturedStruct, IfCond,
5355 &Data](CodeGenFunction &CGF, PrePostActionTy &) {
5356 OMPLoopScope PreInitScope(CGF, S);
5357 CGF.CGM.getOpenMPRuntime().emitTaskLoopCall(CGF, S.getBeginLoc(), S,
5358 OutlinedFn, SharedsTy,
5359 CapturedStruct, IfCond, Data);
5360 };
5361 CGF.CGM.getOpenMPRuntime().emitInlinedDirective(CGF, OMPD_taskloop,
5362 CodeGen);
5363 };
5364 if (Data.Nogroup) {
5365 EmitOMPTaskBasedDirective(S, OMPD_taskloop, BodyGen, TaskGen, Data);
5366 } else {
5367 CGM.getOpenMPRuntime().emitTaskgroupRegion(
5368 *this,
5369 [&S, &BodyGen, &TaskGen, &Data](CodeGenFunction &CGF,
5370 PrePostActionTy &Action) {
5371 Action.Enter(CGF);
5372 CGF.EmitOMPTaskBasedDirective(S, OMPD_taskloop, BodyGen, TaskGen,
5373 Data);
5374 },
5375 S.getBeginLoc());
5376 }
5377}
5378
5379void CodeGenFunction::EmitOMPTaskLoopDirective(const OMPTaskLoopDirective &S) {
5380 EmitOMPTaskLoopBasedDirective(S);
5381}
5382
5383void CodeGenFunction::EmitOMPTaskLoopSimdDirective(
5384 const OMPTaskLoopSimdDirective &S) {
5385 OMPLexicalScope Scope(*this, S);
5386 EmitOMPTaskLoopBasedDirective(S);
5387}
5388
5389void CodeGenFunction::EmitOMPMasterTaskLoopDirective(
5390 const OMPMasterTaskLoopDirective &S) {
5391 auto &&CodeGen = [this, &S](CodeGenFunction &CGF, PrePostActionTy &Action) {
5392 Action.Enter(CGF);
5393 EmitOMPTaskLoopBasedDirective(S);
5394 };
5395 OMPLexicalScope Scope(*this, S, llvm::None, /*EmitPreInitStmt=*/false);
5396 CGM.getOpenMPRuntime().emitMasterRegion(*this, CodeGen, S.getBeginLoc());
5397}
5398
5399void CodeGenFunction::EmitOMPMasterTaskLoopSimdDirective(
5400 const OMPMasterTaskLoopSimdDirective &S) {
5401 auto &&CodeGen = [this, &S](CodeGenFunction &CGF, PrePostActionTy &Action) {
5402 Action.Enter(CGF);
5403 EmitOMPTaskLoopBasedDirective(S);
5404 };
5405 OMPLexicalScope Scope(*this, S);
5406 CGM.getOpenMPRuntime().emitMasterRegion(*this, CodeGen, S.getBeginLoc());
5407}
5408
5409void CodeGenFunction::EmitOMPParallelMasterTaskLoopDirective(
5410 const OMPParallelMasterTaskLoopDirective &S) {
5411 auto &&CodeGen = [this, &S](CodeGenFunction &CGF, PrePostActionTy &Action) {
5412 auto &&TaskLoopCodeGen = [&S](CodeGenFunction &CGF,
5413 PrePostActionTy &Action) {
5414 Action.Enter(CGF);
5415 CGF.EmitOMPTaskLoopBasedDirective(S);
5416 };
5417 OMPLexicalScope Scope(CGF, S, llvm::None, /*EmitPreInitStmt=*/false);
5418 CGM.getOpenMPRuntime().emitMasterRegion(CGF, TaskLoopCodeGen,
5419 S.getBeginLoc());
5420 };
5421 emitCommonOMPParallelDirective(*this, S, OMPD_master_taskloop, CodeGen,
5422 emitEmptyBoundParameters);
5423}
5424
5425void CodeGenFunction::EmitOMPParallelMasterTaskLoopSimdDirective(
5426 const OMPParallelMasterTaskLoopSimdDirective &S) {
5427 auto &&CodeGen = [this, &S](CodeGenFunction &CGF, PrePostActionTy &Action) {
5428 auto &&TaskLoopCodeGen = [&S](CodeGenFunction &CGF,
5429 PrePostActionTy &Action) {
5430 Action.Enter(CGF);
5431 CGF.EmitOMPTaskLoopBasedDirective(S);
5432 };
5433 OMPLexicalScope Scope(CGF, S, OMPD_parallel, /*EmitPreInitStmt=*/false);
5434 CGM.getOpenMPRuntime().emitMasterRegion(CGF, TaskLoopCodeGen,
5435 S.getBeginLoc());
5436 };
5437 emitCommonOMPParallelDirective(*this, S, OMPD_master_taskloop_simd, CodeGen,
5438 emitEmptyBoundParameters);
5439}
5440
5441// Generate the instructions for '#pragma omp target update' directive.
5442void CodeGenFunction::EmitOMPTargetUpdateDirective(
5443 const OMPTargetUpdateDirective &S) {
5444 // If we don't have target devices, don't bother emitting the data mapping
5445 // code.
5446 if (CGM.getLangOpts().OMPTargetTriples.empty())
5447 return;
5448
5449 // Check if we have any if clause associated with the directive.
5450 const Expr *IfCond = nullptr;
5451 if (const auto *C = S.getSingleClause<OMPIfClause>())
5452 IfCond = C->getCondition();
5453
5454 // Check if we have any device clause associated with the directive.
5455 const Expr *Device = nullptr;
5456 if (const auto *C = S.getSingleClause<OMPDeviceClause>())
5457 Device = C->getDevice();
5458
5459 OMPLexicalScope Scope(*this, S, OMPD_task);
5460 CGM.getOpenMPRuntime().emitTargetDataStandAloneCall(*this, S, IfCond, Device);
5461}
5462
5463void CodeGenFunction::EmitSimpleOMPExecutableDirective(
5464 const OMPExecutableDirective &D) {
5465 if (!D.hasAssociatedStmt() || !D.getAssociatedStmt())
5466 return;
5467 auto &&CodeGen = [&D](CodeGenFunction &CGF, PrePostActionTy &Action) {
5468 if (isOpenMPSimdDirective(D.getDirectiveKind())) {
1
Assuming the condition is false
2
Taking false branch
5469 emitOMPSimdRegion(CGF, cast<OMPLoopDirective>(D), Action);
5470 } else {
5471 OMPPrivateScope LoopGlobals(CGF);
5472 if (const auto *LD
3.1
'LD' is non-null
= dyn_cast<OMPLoopDirective>(&D)) {
3
Assuming the object is a 'OMPLoopDirective'
4
Taking true branch
5473 for (const Expr *E : LD->counters()) {
5
Assuming '__begin3' is not equal to '__end3'
5474 const auto *VD = dyn_cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl());
6
'E' is a 'DeclRefExpr'
7
Assuming the object is not a 'VarDecl'
8
'VD' initialized to a null pointer value
5475 if (!VD->hasLocalStorage() && !CGF.LocalDeclMap.count(VD)) {
9
Called C++ object pointer is null
5476 LValue GlobLVal = CGF.EmitLValue(E);
5477 LoopGlobals.addPrivate(
5478 VD, [&GlobLVal, &CGF]() { return GlobLVal.getAddress(CGF); });
5479 }
5480 if (isa<OMPCapturedExprDecl>(VD)) {
5481 // Emit only those that were not explicitly referenced in clauses.
5482 if (!CGF.LocalDeclMap.count(VD))
5483 CGF.EmitVarDecl(*VD);
5484 }
5485 }
5486 for (const auto *C : D.getClausesOfKind<OMPOrderedClause>()) {
5487 if (!C->getNumForLoops())
5488 continue;
5489 for (unsigned I = LD->getCollapsedNumber(),
5490 E = C->getLoopNumIterations().size();
5491 I < E; ++I) {
5492 if (const auto *VD = dyn_cast<OMPCapturedExprDecl>(
5493 cast<DeclRefExpr>(C->getLoopCounter(I))->getDecl())) {
5494 // Emit only those that were not explicitly referenced in clauses.
5495 if (!CGF.LocalDeclMap.count(VD))
5496 CGF.EmitVarDecl(*VD);
5497 }
5498 }
5499 }
5500 }
5501 LoopGlobals.Privatize();
5502 CGF.EmitStmt(D.getInnermostCapturedStmt()->getCapturedStmt());
5503 }
5504 };
5505 OMPSimdLexicalScope Scope(*this, D);
5506 CGM.getOpenMPRuntime().emitInlinedDirective(
5507 *this,
5508 isOpenMPSimdDirective(D.getDirectiveKind()) ? OMPD_simd
5509 : D.getDirectiveKind(),
5510 CodeGen);
5511}