| 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 "CGDebugInfo.h" |
| 15 | #include "CGOpenMPRuntime.h" |
| 16 | #include "CodeGenFunction.h" |
| 17 | #include "CodeGenModule.h" |
| 18 | #include "CodeGenPGO.h" |
| 19 | #include "TargetInfo.h" |
| 20 | #include "clang/AST/ASTContext.h" |
| 21 | #include "clang/AST/Attr.h" |
| 22 | #include "clang/AST/DeclOpenMP.h" |
| 23 | #include "clang/AST/OpenMPClause.h" |
| 24 | #include "clang/AST/Stmt.h" |
| 25 | #include "clang/AST/StmtOpenMP.h" |
| 26 | #include "clang/AST/StmtVisitor.h" |
| 27 | #include "clang/Basic/DiagnosticFrontend.h" |
| 28 | #include "clang/Basic/OpenMPKinds.h" |
| 29 | #include "clang/Basic/PrettyStackTrace.h" |
| 30 | #include "clang/Basic/SourceManager.h" |
| 31 | #include "llvm/ADT/SmallSet.h" |
| 32 | #include "llvm/BinaryFormat/Dwarf.h" |
| 33 | #include "llvm/Frontend/OpenMP/OMPConstants.h" |
| 34 | #include "llvm/Frontend/OpenMP/OMPIRBuilder.h" |
| 35 | #include "llvm/IR/Constants.h" |
| 36 | #include "llvm/IR/DebugInfoMetadata.h" |
| 37 | #include "llvm/IR/Instructions.h" |
| 38 | #include "llvm/IR/IntrinsicInst.h" |
| 39 | #include "llvm/IR/Metadata.h" |
| 40 | #include "llvm/Support/AtomicOrdering.h" |
| 41 | #include "llvm/Support/Debug.h" |
| 42 | #include <optional> |
| 43 | using namespace clang; |
| 44 | using namespace CodeGen; |
| 45 | using namespace llvm::omp; |
| 46 | |
| 47 | #define TTL_CODEGEN_TYPE "target-teams-loop-codegen" |
| 48 | |
| 49 | static const VarDecl *getBaseDecl(const Expr *Ref); |
| 50 | static OpenMPDirectiveKind |
| 51 | getEffectiveDirectiveKind(const OMPExecutableDirective &S); |
| 52 | |
| 53 | /// Whether a combined `distribute parallel for` may use the fused |
| 54 | /// distr_static_chunk + static_chunkone schedule (enum 93): one |
| 55 | /// for_static_init, no surrounding distribute_static_init. |
| 56 | static bool canEmitGPUFusedDistSchedule(const CodeGenModule &CGM, |
| 57 | const OMPLoopDirective &S, |
| 58 | OpenMPDirectiveKind DKind) { |
| 59 | // Reduction-only for now. Non-reduction cases might follow in the future, but |
| 60 | // need more analysis for maximum profit. |
| 61 | return CGM.getLangOpts().OpenMPIsTargetDevice && CGM.getTriple().isGPU() && |
| 62 | isOpenMPLoopBoundSharingDirective(Kind: DKind) && |
| 63 | S.hasClausesOfKind<OMPReductionClause>() && |
| 64 | !S.getSingleClause<OMPDistScheduleClause>() && |
| 65 | !S.getSingleClause<OMPScheduleClause>() && |
| 66 | !S.getSingleClause<OMPOrderedClause>(); |
| 67 | } |
| 68 | |
| 69 | namespace { |
| 70 | /// Lexical scope for OpenMP executable constructs, that handles correct codegen |
| 71 | /// for captured expressions. |
| 72 | class OMPLexicalScope : public CodeGenFunction::LexicalScope { |
| 73 | void emitPreInitStmt(CodeGenFunction &CGF, const OMPExecutableDirective &S) { |
| 74 | for (const auto *C : S.clauses()) { |
| 75 | if (const auto *CPI = OMPClauseWithPreInit::get(C)) { |
| 76 | if (const auto *PreInit = |
| 77 | cast_or_null<DeclStmt>(Val: CPI->getPreInitStmt())) { |
| 78 | for (const auto *I : PreInit->decls()) { |
| 79 | if (!I->hasAttr<OMPCaptureNoInitAttr>()) { |
| 80 | CGF.EmitVarDecl(D: cast<VarDecl>(Val: *I)); |
| 81 | } else { |
| 82 | CodeGenFunction::AutoVarEmission Emission = |
| 83 | CGF.EmitAutoVarAlloca(var: cast<VarDecl>(Val: *I)); |
| 84 | CGF.EmitAutoVarCleanups(emission: Emission); |
| 85 | } |
| 86 | } |
| 87 | } |
| 88 | } |
| 89 | } |
| 90 | } |
| 91 | CodeGenFunction::OMPPrivateScope InlinedShareds; |
| 92 | |
| 93 | static bool isCapturedVar(CodeGenFunction &CGF, const VarDecl *VD) { |
| 94 | return CGF.LambdaCaptureFields.lookup(Val: VD) || |
| 95 | (CGF.CapturedStmtInfo && CGF.CapturedStmtInfo->lookup(VD)) || |
| 96 | (isa_and_nonnull<BlockDecl>(Val: CGF.CurCodeDecl) && |
| 97 | cast<BlockDecl>(Val: CGF.CurCodeDecl)->capturesVariable(var: VD)); |
| 98 | } |
| 99 | |
| 100 | public: |
| 101 | OMPLexicalScope( |
| 102 | CodeGenFunction &CGF, const OMPExecutableDirective &S, |
| 103 | const std::optional<OpenMPDirectiveKind> CapturedRegion = std::nullopt, |
| 104 | const bool EmitPreInitStmt = true) |
| 105 | : CodeGenFunction::LexicalScope(CGF, S.getSourceRange()), |
| 106 | InlinedShareds(CGF) { |
| 107 | if (EmitPreInitStmt) |
| 108 | emitPreInitStmt(CGF, S); |
| 109 | if (!CapturedRegion) |
| 110 | return; |
| 111 | assert(S.hasAssociatedStmt() && |
| 112 | "Expected associated statement for inlined directive."); |
| 113 | const CapturedStmt *CS = S.getCapturedStmt(RegionKind: *CapturedRegion); |
| 114 | for (const auto &C : CS->captures()) { |
| 115 | if (C.capturesVariable() || C.capturesVariableByCopy()) { |
| 116 | auto *VD = C.getCapturedVar(); |
| 117 | assert(VD == VD->getCanonicalDecl() && |
| 118 | "Canonical decl must be captured."); |
| 119 | DeclRefExpr DRE( |
| 120 | CGF.getContext(), const_cast<VarDecl *>(VD), |
| 121 | isCapturedVar(CGF, VD) || (CGF.CapturedStmtInfo && |
| 122 | InlinedShareds.isGlobalVarCaptured(VD)), |
| 123 | VD->getType().getNonReferenceType(), VK_LValue, C.getLocation()); |
| 124 | InlinedShareds.addPrivate(LocalVD: VD, Addr: CGF.EmitLValue(E: &DRE).getAddress()); |
| 125 | } |
| 126 | } |
| 127 | (void)InlinedShareds.Privatize(); |
| 128 | } |
| 129 | }; |
| 130 | |
| 131 | /// Lexical scope for OpenMP parallel construct, that handles correct codegen |
| 132 | /// for captured expressions. |
| 133 | class OMPParallelScope final : public OMPLexicalScope { |
| 134 | bool EmitPreInitStmt(const OMPExecutableDirective &S) { |
| 135 | OpenMPDirectiveKind EKind = getEffectiveDirectiveKind(S); |
| 136 | return !(isOpenMPTargetExecutionDirective(DKind: EKind) || |
| 137 | isOpenMPLoopBoundSharingDirective(Kind: EKind)) && |
| 138 | isOpenMPParallelDirective(DKind: EKind); |
| 139 | } |
| 140 | |
| 141 | public: |
| 142 | OMPParallelScope(CodeGenFunction &CGF, const OMPExecutableDirective &S) |
| 143 | : OMPLexicalScope(CGF, S, /*CapturedRegion=*/std::nullopt, |
| 144 | EmitPreInitStmt(S)) {} |
| 145 | }; |
| 146 | |
| 147 | /// Lexical scope for OpenMP teams construct, that handles correct codegen |
| 148 | /// for captured expressions. |
| 149 | class OMPTeamsScope final : public OMPLexicalScope { |
| 150 | bool EmitPreInitStmt(const OMPExecutableDirective &S) { |
| 151 | OpenMPDirectiveKind EKind = getEffectiveDirectiveKind(S); |
| 152 | return !isOpenMPTargetExecutionDirective(DKind: EKind) && |
| 153 | isOpenMPTeamsDirective(DKind: EKind); |
| 154 | } |
| 155 | |
| 156 | public: |
| 157 | OMPTeamsScope(CodeGenFunction &CGF, const OMPExecutableDirective &S) |
| 158 | : OMPLexicalScope(CGF, S, /*CapturedRegion=*/std::nullopt, |
| 159 | EmitPreInitStmt(S)) {} |
| 160 | }; |
| 161 | |
| 162 | /// Private scope for OpenMP loop-based directives, that supports capturing |
| 163 | /// of used expression from loop statement. |
| 164 | class OMPLoopScope : public CodeGenFunction::RunCleanupsScope { |
| 165 | void emitPreInitStmt(CodeGenFunction &CGF, const OMPLoopBasedDirective &S) { |
| 166 | const Stmt *PreInits; |
| 167 | CodeGenFunction::OMPMapVars PreCondVars; |
| 168 | if (auto *LD = dyn_cast<OMPLoopDirective>(Val: &S)) { |
| 169 | // Emit init, __range, __begin and __end variables for C++ range loops. |
| 170 | (void)OMPLoopBasedDirective::doForAllLoops( |
| 171 | CurStmt: LD->getInnermostCapturedStmt()->getCapturedStmt(), |
| 172 | /*TryImperfectlyNestedLoops=*/true, NumLoops: LD->getLoopsNumber(), |
| 173 | Callback: [&CGF](unsigned Cnt, const Stmt *CurStmt) { |
| 174 | if (const auto *CXXFor = dyn_cast<CXXForRangeStmt>(Val: CurStmt)) { |
| 175 | if (const Stmt *Init = CXXFor->getInit()) |
| 176 | CGF.EmitStmt(S: Init); |
| 177 | CGF.EmitStmt(S: CXXFor->getRangeStmt()); |
| 178 | CGF.EmitStmt(S: CXXFor->getBeginStmt()); |
| 179 | CGF.EmitStmt(S: CXXFor->getEndStmt()); |
| 180 | } |
| 181 | return false; |
| 182 | }); |
| 183 | llvm::DenseSet<const VarDecl *> EmittedAsPrivate; |
| 184 | for (const auto *E : LD->counters()) { |
| 185 | const auto *VD = cast<VarDecl>(Val: cast<DeclRefExpr>(Val: E)->getDecl()); |
| 186 | EmittedAsPrivate.insert(V: VD->getCanonicalDecl()); |
| 187 | (void)PreCondVars.setVarAddr( |
| 188 | CGF, LocalVD: VD, TempAddr: CGF.CreateMemTemp(T: VD->getType().getNonReferenceType())); |
| 189 | } |
| 190 | // Mark private vars as undefs. |
| 191 | for (const auto *C : LD->getClausesOfKind<OMPPrivateClause>()) { |
| 192 | for (const Expr *IRef : C->varlist()) { |
| 193 | const auto *OrigVD = |
| 194 | cast<VarDecl>(Val: cast<DeclRefExpr>(Val: IRef)->getDecl()); |
| 195 | if (EmittedAsPrivate.insert(V: OrigVD->getCanonicalDecl()).second) { |
| 196 | QualType OrigVDTy = OrigVD->getType().getNonReferenceType(); |
| 197 | (void)PreCondVars.setVarAddr( |
| 198 | CGF, LocalVD: OrigVD, |
| 199 | TempAddr: Address(llvm::UndefValue::get(T: CGF.ConvertTypeForMem( |
| 200 | T: CGF.getContext().getPointerType(T: OrigVDTy))), |
| 201 | CGF.ConvertTypeForMem(T: OrigVDTy), |
| 202 | CGF.getContext().getDeclAlign(D: OrigVD))); |
| 203 | } |
| 204 | } |
| 205 | } |
| 206 | (void)PreCondVars.apply(CGF); |
| 207 | PreInits = LD->getPreInits(); |
| 208 | } else if (const auto *Tile = dyn_cast<OMPTileDirective>(Val: &S)) { |
| 209 | PreInits = Tile->getPreInits(); |
| 210 | } else if (const auto *Stripe = dyn_cast<OMPStripeDirective>(Val: &S)) { |
| 211 | PreInits = Stripe->getPreInits(); |
| 212 | } else if (const auto *Unroll = dyn_cast<OMPUnrollDirective>(Val: &S)) { |
| 213 | PreInits = Unroll->getPreInits(); |
| 214 | } else if (const auto *Reverse = dyn_cast<OMPReverseDirective>(Val: &S)) { |
| 215 | PreInits = Reverse->getPreInits(); |
| 216 | } else if (const auto *Split = dyn_cast<OMPSplitDirective>(Val: &S)) { |
| 217 | PreInits = Split->getPreInits(); |
| 218 | } else if (const auto *Interchange = |
| 219 | dyn_cast<OMPInterchangeDirective>(Val: &S)) { |
| 220 | PreInits = Interchange->getPreInits(); |
| 221 | } else { |
| 222 | llvm_unreachable("Unknown loop-based directive kind."); |
| 223 | } |
| 224 | doEmitPreinits(PreInits); |
| 225 | PreCondVars.restore(CGF); |
| 226 | } |
| 227 | |
| 228 | void |
| 229 | emitPreInitStmt(CodeGenFunction &CGF, |
| 230 | const OMPCanonicalLoopSequenceTransformationDirective &S) { |
| 231 | const Stmt *PreInits; |
| 232 | if (const auto *Fuse = dyn_cast<OMPFuseDirective>(Val: &S)) { |
| 233 | PreInits = Fuse->getPreInits(); |
| 234 | } else { |
| 235 | llvm_unreachable( |
| 236 | "Unknown canonical loop sequence transform directive kind."); |
| 237 | } |
| 238 | doEmitPreinits(PreInits); |
| 239 | } |
| 240 | |
| 241 | void doEmitPreinits(const Stmt *PreInits) { |
| 242 | if (PreInits) { |
| 243 | // CompoundStmts and DeclStmts are used as lists of PreInit statements and |
| 244 | // declarations. Since declarations must be visible in the the following |
| 245 | // that they initialize, unpack the CompoundStmt they are nested in. |
| 246 | SmallVector<const Stmt *> PreInitStmts; |
| 247 | if (auto *PreInitCompound = dyn_cast<CompoundStmt>(Val: PreInits)) |
| 248 | llvm::append_range(C&: PreInitStmts, R: PreInitCompound->body()); |
| 249 | else |
| 250 | PreInitStmts.push_back(Elt: PreInits); |
| 251 | |
| 252 | for (const Stmt *S : PreInitStmts) { |
| 253 | // EmitStmt skips any OMPCapturedExprDecls, but needs to be emitted |
| 254 | // here. |
| 255 | if (auto *PreInitDecl = dyn_cast<DeclStmt>(Val: S)) { |
| 256 | for (Decl *I : PreInitDecl->decls()) |
| 257 | CGF.EmitVarDecl(D: cast<VarDecl>(Val&: *I)); |
| 258 | continue; |
| 259 | } |
| 260 | CGF.EmitStmt(S); |
| 261 | } |
| 262 | } |
| 263 | } |
| 264 | |
| 265 | public: |
| 266 | OMPLoopScope(CodeGenFunction &CGF, const OMPLoopBasedDirective &S) |
| 267 | : CodeGenFunction::RunCleanupsScope(CGF) { |
| 268 | emitPreInitStmt(CGF, S); |
| 269 | } |
| 270 | OMPLoopScope(CodeGenFunction &CGF, |
| 271 | const OMPCanonicalLoopSequenceTransformationDirective &S) |
| 272 | : CodeGenFunction::RunCleanupsScope(CGF) { |
| 273 | emitPreInitStmt(CGF, S); |
| 274 | } |
| 275 | }; |
| 276 | |
| 277 | class OMPSimdLexicalScope : public CodeGenFunction::LexicalScope { |
| 278 | CodeGenFunction::OMPPrivateScope InlinedShareds; |
| 279 | |
| 280 | static bool isCapturedVar(CodeGenFunction &CGF, const VarDecl *VD) { |
| 281 | return CGF.LambdaCaptureFields.lookup(Val: VD) || |
| 282 | (CGF.CapturedStmtInfo && CGF.CapturedStmtInfo->lookup(VD)) || |
| 283 | (isa_and_nonnull<BlockDecl>(Val: CGF.CurCodeDecl) && |
| 284 | cast<BlockDecl>(Val: CGF.CurCodeDecl)->capturesVariable(var: VD)); |
| 285 | } |
| 286 | |
| 287 | public: |
| 288 | OMPSimdLexicalScope(CodeGenFunction &CGF, const OMPExecutableDirective &S) |
| 289 | : CodeGenFunction::LexicalScope(CGF, S.getSourceRange()), |
| 290 | InlinedShareds(CGF) { |
| 291 | for (const auto *C : S.clauses()) { |
| 292 | if (const auto *CPI = OMPClauseWithPreInit::get(C)) { |
| 293 | if (const auto *PreInit = |
| 294 | cast_or_null<DeclStmt>(Val: CPI->getPreInitStmt())) { |
| 295 | for (const auto *I : PreInit->decls()) { |
| 296 | if (!I->hasAttr<OMPCaptureNoInitAttr>()) { |
| 297 | CGF.EmitVarDecl(D: cast<VarDecl>(Val: *I)); |
| 298 | } else { |
| 299 | CodeGenFunction::AutoVarEmission Emission = |
| 300 | CGF.EmitAutoVarAlloca(var: cast<VarDecl>(Val: *I)); |
| 301 | CGF.EmitAutoVarCleanups(emission: Emission); |
| 302 | } |
| 303 | } |
| 304 | } |
| 305 | } else if (const auto *UDP = dyn_cast<OMPUseDevicePtrClause>(Val: C)) { |
| 306 | for (const Expr *E : UDP->varlist()) { |
| 307 | const Decl *D = cast<DeclRefExpr>(Val: E)->getDecl(); |
| 308 | if (const auto *OED = dyn_cast<OMPCapturedExprDecl>(Val: D)) |
| 309 | CGF.EmitVarDecl(D: *OED); |
| 310 | } |
| 311 | } else if (const auto *UDP = dyn_cast<OMPUseDeviceAddrClause>(Val: C)) { |
| 312 | for (const Expr *E : UDP->varlist()) { |
| 313 | const Decl *D = getBaseDecl(Ref: E); |
| 314 | if (const auto *OED = dyn_cast<OMPCapturedExprDecl>(Val: D)) |
| 315 | CGF.EmitVarDecl(D: *OED); |
| 316 | } |
| 317 | } |
| 318 | } |
| 319 | if (!isOpenMPSimdDirective(DKind: getEffectiveDirectiveKind(S))) |
| 320 | CGF.EmitOMPPrivateClause(D: S, PrivateScope&: InlinedShareds); |
| 321 | if (const auto *TG = dyn_cast<OMPTaskgroupDirective>(Val: &S)) { |
| 322 | if (const Expr *E = TG->getReductionRef()) |
| 323 | CGF.EmitVarDecl(D: *cast<VarDecl>(Val: cast<DeclRefExpr>(Val: E)->getDecl())); |
| 324 | } |
| 325 | // Temp copy arrays for inscan reductions should not be emitted as they are |
| 326 | // not used in simd only mode. |
| 327 | llvm::DenseSet<CanonicalDeclPtr<const Decl>> CopyArrayTemps; |
| 328 | for (const auto *C : S.getClausesOfKind<OMPReductionClause>()) { |
| 329 | if (C->getModifier() != OMPC_REDUCTION_inscan) |
| 330 | continue; |
| 331 | for (const Expr *E : C->copy_array_temps()) |
| 332 | CopyArrayTemps.insert(V: cast<DeclRefExpr>(Val: E)->getDecl()); |
| 333 | } |
| 334 | const auto *CS = cast_or_null<CapturedStmt>(Val: S.getAssociatedStmt()); |
| 335 | while (CS) { |
| 336 | for (auto &C : CS->captures()) { |
| 337 | if (C.capturesVariable() || C.capturesVariableByCopy()) { |
| 338 | auto *VD = C.getCapturedVar(); |
| 339 | if (CopyArrayTemps.contains(V: VD)) |
| 340 | continue; |
| 341 | assert(VD == VD->getCanonicalDecl() && |
| 342 | "Canonical decl must be captured."); |
| 343 | DeclRefExpr DRE(CGF.getContext(), const_cast<VarDecl *>(VD), |
| 344 | isCapturedVar(CGF, VD) || |
| 345 | (CGF.CapturedStmtInfo && |
| 346 | InlinedShareds.isGlobalVarCaptured(VD)), |
| 347 | VD->getType().getNonReferenceType(), VK_LValue, |
| 348 | C.getLocation()); |
| 349 | InlinedShareds.addPrivate(LocalVD: VD, Addr: CGF.EmitLValue(E: &DRE).getAddress()); |
| 350 | } |
| 351 | } |
| 352 | CS = dyn_cast<CapturedStmt>(Val: CS->getCapturedStmt()); |
| 353 | } |
| 354 | (void)InlinedShareds.Privatize(); |
| 355 | } |
| 356 | }; |
| 357 | |
| 358 | } // namespace |
| 359 | |
| 360 | // The loop directive with a bind clause will be mapped to a different |
| 361 | // directive with corresponding semantics. |
| 362 | static OpenMPDirectiveKind |
| 363 | getEffectiveDirectiveKind(const OMPExecutableDirective &S) { |
| 364 | OpenMPDirectiveKind Kind = S.getDirectiveKind(); |
| 365 | if (Kind != OMPD_loop) |
| 366 | return Kind; |
| 367 | |
| 368 | OpenMPBindClauseKind BindKind = OMPC_BIND_unknown; |
| 369 | if (const auto *C = S.getSingleClause<OMPBindClause>()) |
| 370 | BindKind = C->getBindKind(); |
| 371 | |
| 372 | switch (BindKind) { |
| 373 | case OMPC_BIND_parallel: |
| 374 | return OMPD_for; |
| 375 | case OMPC_BIND_teams: |
| 376 | return OMPD_distribute; |
| 377 | case OMPC_BIND_thread: |
| 378 | return OMPD_simd; |
| 379 | default: |
| 380 | return OMPD_loop; |
| 381 | } |
| 382 | } |
| 383 | |
| 384 | static void emitCommonOMPTargetDirective(CodeGenFunction &CGF, |
| 385 | const OMPExecutableDirective &S, |
| 386 | const RegionCodeGenTy &CodeGen); |
| 387 | |
| 388 | LValue CodeGenFunction::EmitOMPSharedLValue(const Expr *E) { |
| 389 | if (const auto *OrigDRE = dyn_cast<DeclRefExpr>(Val: E)) { |
| 390 | if (const auto *OrigVD = dyn_cast<VarDecl>(Val: OrigDRE->getDecl())) { |
| 391 | OrigVD = OrigVD->getCanonicalDecl(); |
| 392 | bool IsCaptured = |
| 393 | LambdaCaptureFields.lookup(Val: OrigVD) || |
| 394 | (CapturedStmtInfo && CapturedStmtInfo->lookup(VD: OrigVD)) || |
| 395 | (isa_and_nonnull<BlockDecl>(Val: CurCodeDecl)); |
| 396 | DeclRefExpr DRE(getContext(), const_cast<VarDecl *>(OrigVD), IsCaptured, |
| 397 | OrigDRE->getType(), VK_LValue, OrigDRE->getExprLoc()); |
| 398 | return EmitLValue(E: &DRE); |
| 399 | } |
| 400 | } |
| 401 | return EmitLValue(E); |
| 402 | } |
| 403 | |
| 404 | llvm::Value *CodeGenFunction::getTypeSize(QualType Ty) { |
| 405 | ASTContext &C = getContext(); |
| 406 | llvm::Value *Size = nullptr; |
| 407 | auto SizeInChars = C.getTypeSizeInChars(T: Ty); |
| 408 | if (SizeInChars.isZero()) { |
| 409 | // getTypeSizeInChars() returns 0 for a VLA. |
| 410 | while (const VariableArrayType *VAT = C.getAsVariableArrayType(T: Ty)) { |
| 411 | VlaSizePair VlaSize = getVLASize(vla: VAT); |
| 412 | Ty = VlaSize.Type; |
| 413 | Size = |
| 414 | Size ? Builder.CreateNUWMul(LHS: Size, RHS: VlaSize.NumElts) : VlaSize.NumElts; |
| 415 | } |
| 416 | SizeInChars = C.getTypeSizeInChars(T: Ty); |
| 417 | if (SizeInChars.isZero()) |
| 418 | return llvm::ConstantInt::get(Ty: SizeTy, /*V=*/0); |
| 419 | return Builder.CreateNUWMul(LHS: Size, RHS: CGM.getSize(numChars: SizeInChars)); |
| 420 | } |
| 421 | return CGM.getSize(numChars: SizeInChars); |
| 422 | } |
| 423 | |
| 424 | void CodeGenFunction::GenerateOpenMPCapturedVars( |
| 425 | const CapturedStmt &S, SmallVectorImpl<llvm::Value *> &CapturedVars) { |
| 426 | const RecordDecl *RD = S.getCapturedRecordDecl(); |
| 427 | auto CurField = RD->field_begin(); |
| 428 | auto CurCap = S.captures().begin(); |
| 429 | for (CapturedStmt::const_capture_init_iterator I = S.capture_init_begin(), |
| 430 | E = S.capture_init_end(); |
| 431 | I != E; ++I, ++CurField, ++CurCap) { |
| 432 | if (CurField->hasCapturedVLAType()) { |
| 433 | const VariableArrayType *VAT = CurField->getCapturedVLAType(); |
| 434 | llvm::Value *Val = VLASizeMap[VAT->getSizeExpr()]; |
| 435 | CapturedVars.push_back(Elt: Val); |
| 436 | } else if (CurCap->capturesThis()) { |
| 437 | CapturedVars.push_back(Elt: CXXThisValue); |
| 438 | } else if (CurCap->capturesVariableByCopy()) { |
| 439 | llvm::Value *CV = EmitLoadOfScalar(lvalue: EmitLValue(E: *I), Loc: CurCap->getLocation()); |
| 440 | |
| 441 | // If the field is not a pointer, we need to save the actual value |
| 442 | // and load it as a void pointer. |
| 443 | if (!CurField->getType()->isAnyPointerType()) { |
| 444 | ASTContext &Ctx = getContext(); |
| 445 | Address DstAddr = CreateMemTempWithoutCast( |
| 446 | T: Ctx.getUIntPtrType(), |
| 447 | Name: Twine(CurCap->getCapturedVar()->getName(), ".casted")); |
| 448 | LValue DstLV = MakeAddrLValue(Addr: DstAddr, T: Ctx.getUIntPtrType()); |
| 449 | |
| 450 | llvm::Value *SrcAddrVal = EmitScalarConversion( |
| 451 | Src: DstAddr.emitRawPointer(CGF&: *this), |
| 452 | SrcTy: Ctx.getPointerType(T: Ctx.getUIntPtrType()), |
| 453 | DstTy: Ctx.getPointerType(T: CurField->getType()), Loc: CurCap->getLocation()); |
| 454 | LValue SrcLV = |
| 455 | MakeNaturalAlignAddrLValue(V: SrcAddrVal, T: CurField->getType()); |
| 456 | |
| 457 | // Store the value using the source type pointer. |
| 458 | EmitStoreThroughLValue(Src: RValue::get(V: CV), Dst: SrcLV); |
| 459 | |
| 460 | // Load the value using the destination type pointer. |
| 461 | CV = EmitLoadOfScalar(lvalue: DstLV, Loc: CurCap->getLocation()); |
| 462 | } |
| 463 | CapturedVars.push_back(Elt: CV); |
| 464 | } else { |
| 465 | assert(CurCap->capturesVariable() && "Expected capture by reference."); |
| 466 | CapturedVars.push_back(Elt: EmitLValue(E: *I).getAddress().emitRawPointer(CGF&: *this)); |
| 467 | } |
| 468 | } |
| 469 | } |
| 470 | |
| 471 | static Address castValueFromUintptr(CodeGenFunction &CGF, SourceLocation Loc, |
| 472 | QualType DstType, StringRef Name, |
| 473 | LValue AddrLV) { |
| 474 | ASTContext &Ctx = CGF.getContext(); |
| 475 | |
| 476 | llvm::Value *CastedPtr = CGF.EmitScalarConversion( |
| 477 | Src: AddrLV.getAddress().emitRawPointer(CGF), SrcTy: Ctx.getUIntPtrType(), |
| 478 | DstTy: Ctx.getPointerType(T: DstType), Loc); |
| 479 | // FIXME: should the pointee type (DstType) be passed? |
| 480 | Address TmpAddr = |
| 481 | CGF.MakeNaturalAlignAddrLValue(V: CastedPtr, T: DstType).getAddress(); |
| 482 | return TmpAddr; |
| 483 | } |
| 484 | |
| 485 | static QualType getCanonicalParamType(ASTContext &C, QualType T) { |
| 486 | if (T->isLValueReferenceType()) |
| 487 | return C.getLValueReferenceType( |
| 488 | T: getCanonicalParamType(C, T: T.getNonReferenceType()), |
| 489 | /*SpelledAsLValue=*/false); |
| 490 | if (T->isPointerType()) |
| 491 | return C.getPointerType(T: getCanonicalParamType(C, T: T->getPointeeType())); |
| 492 | if (const ArrayType *A = T->getAsArrayTypeUnsafe()) { |
| 493 | if (const auto *VLA = dyn_cast<VariableArrayType>(Val: A)) |
| 494 | return getCanonicalParamType(C, T: VLA->getElementType()); |
| 495 | if (!A->isVariablyModifiedType()) |
| 496 | return C.getCanonicalType(T); |
| 497 | } |
| 498 | return C.getCanonicalParamType(T); |
| 499 | } |
| 500 | |
| 501 | namespace { |
| 502 | /// Contains required data for proper outlined function codegen. |
| 503 | struct FunctionOptions { |
| 504 | /// Captured statement for which the function is generated. |
| 505 | const CapturedStmt *S = nullptr; |
| 506 | /// true if cast to/from UIntPtr is required for variables captured by |
| 507 | /// value. |
| 508 | const bool UIntPtrCastRequired = true; |
| 509 | /// true if only casted arguments must be registered as local args or VLA |
| 510 | /// sizes. |
| 511 | const bool RegisterCastedArgsOnly = false; |
| 512 | /// Name of the generated function. |
| 513 | const StringRef FunctionName; |
| 514 | /// Location of the non-debug version of the outlined function. |
| 515 | SourceLocation Loc; |
| 516 | const bool IsDeviceKernel = false; |
| 517 | explicit FunctionOptions(const CapturedStmt *S, bool UIntPtrCastRequired, |
| 518 | bool RegisterCastedArgsOnly, StringRef FunctionName, |
| 519 | SourceLocation Loc, bool IsDeviceKernel) |
| 520 | : S(S), UIntPtrCastRequired(UIntPtrCastRequired), |
| 521 | RegisterCastedArgsOnly(UIntPtrCastRequired && RegisterCastedArgsOnly), |
| 522 | FunctionName(FunctionName), Loc(Loc), IsDeviceKernel(IsDeviceKernel) {} |
| 523 | }; |
| 524 | } // namespace |
| 525 | |
| 526 | static llvm::Function *emitOutlinedFunctionPrologue( |
| 527 | CodeGenFunction &CGF, FunctionArgList &Args, |
| 528 | llvm::MapVector<const Decl *, std::pair<const VarDecl *, Address>> |
| 529 | &LocalAddrs, |
| 530 | llvm::DenseMap<const Decl *, std::pair<const Expr *, llvm::Value *>> |
| 531 | &VLASizes, |
| 532 | llvm::Value *&CXXThisValue, const FunctionOptions &FO) { |
| 533 | const CapturedDecl *CD = FO.S->getCapturedDecl(); |
| 534 | const RecordDecl *RD = FO.S->getCapturedRecordDecl(); |
| 535 | assert(CD->hasBody() && "missing CapturedDecl body"); |
| 536 | |
| 537 | CXXThisValue = nullptr; |
| 538 | // Build the argument list. |
| 539 | CodeGenModule &CGM = CGF.CGM; |
| 540 | ASTContext &Ctx = CGM.getContext(); |
| 541 | FunctionArgList TargetArgs; |
| 542 | Args.append(in_start: CD->param_begin(), |
| 543 | in_end: std::next(x: CD->param_begin(), n: CD->getContextParamPosition())); |
| 544 | TargetArgs.append( |
| 545 | in_start: CD->param_begin(), |
| 546 | in_end: std::next(x: CD->param_begin(), n: CD->getContextParamPosition())); |
| 547 | auto I = FO.S->captures().begin(); |
| 548 | FunctionDecl *DebugFunctionDecl = nullptr; |
| 549 | if (!FO.UIntPtrCastRequired) { |
| 550 | FunctionProtoType::ExtProtoInfo EPI; |
| 551 | QualType FunctionTy = Ctx.getFunctionType(ResultTy: Ctx.VoidTy, Args: {}, EPI); |
| 552 | DebugFunctionDecl = FunctionDecl::Create( |
| 553 | C&: Ctx, DC: Ctx.getTranslationUnitDecl(), StartLoc: FO.S->getBeginLoc(), |
| 554 | NLoc: SourceLocation(), N: DeclarationName(), T: FunctionTy, |
| 555 | TInfo: Ctx.getTrivialTypeSourceInfo(T: FunctionTy), SC: SC_Static, |
| 556 | /*UsesFPIntrin=*/false, /*isInlineSpecified=*/false, |
| 557 | /*hasWrittenPrototype=*/false); |
| 558 | } |
| 559 | for (const FieldDecl *FD : RD->fields()) { |
| 560 | QualType ArgType = FD->getType(); |
| 561 | IdentifierInfo *II = nullptr; |
| 562 | VarDecl *CapVar = nullptr; |
| 563 | |
| 564 | // If this is a capture by copy and the type is not a pointer, the outlined |
| 565 | // function argument type should be uintptr and the value properly casted to |
| 566 | // uintptr. This is necessary given that the runtime library is only able to |
| 567 | // deal with pointers. We can pass in the same way the VLA type sizes to the |
| 568 | // outlined function. |
| 569 | if (FO.UIntPtrCastRequired && |
| 570 | ((I->capturesVariableByCopy() && !ArgType->isAnyPointerType()) || |
| 571 | I->capturesVariableArrayType())) |
| 572 | ArgType = Ctx.getUIntPtrType(); |
| 573 | |
| 574 | if (I->capturesVariable() || I->capturesVariableByCopy()) { |
| 575 | CapVar = I->getCapturedVar(); |
| 576 | II = CapVar->getIdentifier(); |
| 577 | } else if (I->capturesThis()) { |
| 578 | II = &Ctx.Idents.get(Name: "this"); |
| 579 | } else { |
| 580 | assert(I->capturesVariableArrayType()); |
| 581 | II = &Ctx.Idents.get(Name: "vla"); |
| 582 | } |
| 583 | if (ArgType->isVariablyModifiedType()) |
| 584 | ArgType = getCanonicalParamType(C&: Ctx, T: ArgType); |
| 585 | VarDecl *Arg; |
| 586 | if (CapVar && (CapVar->getTLSKind() != clang::VarDecl::TLS_None)) { |
| 587 | Arg = ImplicitParamDecl::Create(C&: Ctx, /*DC=*/nullptr, IdLoc: FD->getLocation(), |
| 588 | Id: II, T: ArgType, |
| 589 | ParamKind: ImplicitParamKind::ThreadPrivateVar); |
| 590 | } else if (DebugFunctionDecl && (CapVar || I->capturesThis())) { |
| 591 | Arg = ParmVarDecl::Create( |
| 592 | C&: Ctx, DC: DebugFunctionDecl, |
| 593 | StartLoc: CapVar ? CapVar->getBeginLoc() : FD->getBeginLoc(), |
| 594 | IdLoc: CapVar ? CapVar->getLocation() : FD->getLocation(), Id: II, T: ArgType, |
| 595 | /*TInfo=*/nullptr, S: SC_None, /*DefArg=*/nullptr); |
| 596 | } else { |
| 597 | Arg = ImplicitParamDecl::Create(C&: Ctx, /*DC=*/nullptr, IdLoc: FD->getLocation(), |
| 598 | Id: II, T: ArgType, ParamKind: ImplicitParamKind::Other); |
| 599 | } |
| 600 | Args.emplace_back(Args&: Arg); |
| 601 | // Do not cast arguments if we emit function with non-original types. |
| 602 | TargetArgs.emplace_back( |
| 603 | Args: FO.UIntPtrCastRequired |
| 604 | ? Arg |
| 605 | : CGM.getOpenMPRuntime().translateParameter(FD, NativeParam: Arg)); |
| 606 | ++I; |
| 607 | } |
| 608 | Args.append(in_start: std::next(x: CD->param_begin(), n: CD->getContextParamPosition() + 1), |
| 609 | in_end: CD->param_end()); |
| 610 | TargetArgs.append( |
| 611 | in_start: std::next(x: CD->param_begin(), n: CD->getContextParamPosition() + 1), |
| 612 | in_end: CD->param_end()); |
| 613 | |
| 614 | // Create the function declaration. |
| 615 | const CGFunctionInfo &FuncInfo = |
| 616 | FO.IsDeviceKernel |
| 617 | ? CGM.getTypes().arrangeDeviceKernelCallerDeclaration(resultType: Ctx.VoidTy, |
| 618 | args: TargetArgs) |
| 619 | : CGM.getTypes().arrangeBuiltinFunctionDeclaration(resultType: Ctx.VoidTy, |
| 620 | args: TargetArgs); |
| 621 | llvm::FunctionType *FuncLLVMTy = CGM.getTypes().GetFunctionType(Info: FuncInfo); |
| 622 | |
| 623 | auto *F = |
| 624 | llvm::Function::Create(Ty: FuncLLVMTy, Linkage: llvm::GlobalValue::InternalLinkage, |
| 625 | N: FO.FunctionName, M: &CGM.getModule()); |
| 626 | CGM.SetInternalFunctionAttributes(GD: CD, F, FI: FuncInfo); |
| 627 | |
| 628 | // Adjust the calling convention for SPIR-V targets to avoid mismatches |
| 629 | // between callee and caller. |
| 630 | if (CGM.getTriple().isSPIRV() && !FO.IsDeviceKernel) |
| 631 | F->setCallingConv(llvm::CallingConv::SPIR_FUNC); |
| 632 | |
| 633 | if (CD->isNothrow()) |
| 634 | F->setDoesNotThrow(); |
| 635 | F->setDoesNotRecurse(); |
| 636 | |
| 637 | // Always inline the outlined function if optimizations are enabled. |
| 638 | if (CGM.getCodeGenOpts().OptimizationLevel != 0) { |
| 639 | F->removeFnAttr(Kind: llvm::Attribute::NoInline); |
| 640 | F->addFnAttr(Kind: llvm::Attribute::AlwaysInline); |
| 641 | } |
| 642 | if (!CGM.getCodeGenOpts().SampleProfileFile.empty()) |
| 643 | F->addFnAttr(Kind: "sample-profile-suffix-elision-policy", Val: "selected"); |
| 644 | |
| 645 | // Generate the function. |
| 646 | CGF.StartFunction(GD: CD, RetTy: Ctx.VoidTy, Fn: F, FnInfo: FuncInfo, Args: TargetArgs, |
| 647 | Loc: FO.UIntPtrCastRequired ? FO.Loc : FO.S->getBeginLoc(), |
| 648 | StartLoc: FO.UIntPtrCastRequired ? FO.Loc |
| 649 | : CD->getBody()->getBeginLoc()); |
| 650 | unsigned Cnt = CD->getContextParamPosition(); |
| 651 | I = FO.S->captures().begin(); |
| 652 | for (const FieldDecl *FD : RD->fields()) { |
| 653 | // Do not map arguments if we emit function with non-original types. |
| 654 | Address LocalAddr(Address::invalid()); |
| 655 | if (!FO.UIntPtrCastRequired && Args[Cnt] != TargetArgs[Cnt]) { |
| 656 | LocalAddr = CGM.getOpenMPRuntime().getParameterAddress(CGF, NativeParam: Args[Cnt], |
| 657 | TargetParam: TargetArgs[Cnt]); |
| 658 | } else { |
| 659 | LocalAddr = CGF.GetAddrOfLocalVar(VD: Args[Cnt]); |
| 660 | } |
| 661 | // If we are capturing a pointer by copy we don't need to do anything, just |
| 662 | // use the value that we get from the arguments. |
| 663 | if (I->capturesVariableByCopy() && FD->getType()->isAnyPointerType()) { |
| 664 | const VarDecl *CurVD = I->getCapturedVar(); |
| 665 | if (!FO.RegisterCastedArgsOnly) |
| 666 | LocalAddrs.insert(KV: {Args[Cnt], {CurVD, LocalAddr}}); |
| 667 | ++Cnt; |
| 668 | ++I; |
| 669 | continue; |
| 670 | } |
| 671 | |
| 672 | LValue ArgLVal = CGF.MakeAddrLValue(Addr: LocalAddr, T: Args[Cnt]->getType(), |
| 673 | Source: AlignmentSource::Decl); |
| 674 | if (FD->hasCapturedVLAType()) { |
| 675 | if (FO.UIntPtrCastRequired) { |
| 676 | ArgLVal = CGF.MakeAddrLValue( |
| 677 | Addr: castValueFromUintptr(CGF, Loc: I->getLocation(), DstType: FD->getType(), |
| 678 | Name: Args[Cnt]->getName(), AddrLV: ArgLVal), |
| 679 | T: FD->getType(), Source: AlignmentSource::Decl); |
| 680 | } |
| 681 | llvm::Value *ExprArg = CGF.EmitLoadOfScalar(lvalue: ArgLVal, Loc: I->getLocation()); |
| 682 | const VariableArrayType *VAT = FD->getCapturedVLAType(); |
| 683 | VLASizes.try_emplace(Key: Args[Cnt], Args: VAT->getSizeExpr(), Args&: ExprArg); |
| 684 | } else if (I->capturesVariable()) { |
| 685 | const VarDecl *Var = I->getCapturedVar(); |
| 686 | QualType VarTy = Var->getType(); |
| 687 | Address ArgAddr = ArgLVal.getAddress(); |
| 688 | if (ArgLVal.getType()->isLValueReferenceType()) { |
| 689 | ArgAddr = CGF.EmitLoadOfReference(RefLVal: ArgLVal); |
| 690 | } else if (!VarTy->isVariablyModifiedType() || !VarTy->isPointerType()) { |
| 691 | assert(ArgLVal.getType()->isPointerType()); |
| 692 | ArgAddr = CGF.EmitLoadOfPointer( |
| 693 | Ptr: ArgAddr, PtrTy: ArgLVal.getType()->castAs<PointerType>()); |
| 694 | } |
| 695 | if (!FO.RegisterCastedArgsOnly) { |
| 696 | LocalAddrs.insert( |
| 697 | KV: {Args[Cnt], {Var, ArgAddr.withAlignment(NewAlignment: Ctx.getDeclAlign(D: Var))}}); |
| 698 | } |
| 699 | } else if (I->capturesVariableByCopy()) { |
| 700 | assert(!FD->getType()->isAnyPointerType() && |
| 701 | "Not expecting a captured pointer."); |
| 702 | const VarDecl *Var = I->getCapturedVar(); |
| 703 | LocalAddrs.insert(KV: {Args[Cnt], |
| 704 | {Var, FO.UIntPtrCastRequired |
| 705 | ? castValueFromUintptr( |
| 706 | CGF, Loc: I->getLocation(), DstType: FD->getType(), |
| 707 | Name: Args[Cnt]->getName(), AddrLV: ArgLVal) |
| 708 | : ArgLVal.getAddress()}}); |
| 709 | } else { |
| 710 | // If 'this' is captured, load it into CXXThisValue. |
| 711 | assert(I->capturesThis()); |
| 712 | CXXThisValue = CGF.EmitLoadOfScalar(lvalue: ArgLVal, Loc: I->getLocation()); |
| 713 | LocalAddrs.insert(KV: {Args[Cnt], {nullptr, ArgLVal.getAddress()}}); |
| 714 | } |
| 715 | ++Cnt; |
| 716 | ++I; |
| 717 | } |
| 718 | |
| 719 | return F; |
| 720 | } |
| 721 | |
| 722 | static llvm::Function *emitOutlinedFunctionPrologueAggregate( |
| 723 | CodeGenFunction &CGF, FunctionArgList &Args, |
| 724 | llvm::MapVector<const Decl *, std::pair<const VarDecl *, Address>> |
| 725 | &LocalAddrs, |
| 726 | llvm::DenseMap<const Decl *, std::pair<const Expr *, llvm::Value *>> |
| 727 | &VLASizes, |
| 728 | llvm::Value *&CXXThisValue, llvm::Value *&ContextV, const CapturedStmt &CS, |
| 729 | SourceLocation Loc, StringRef FunctionName) { |
| 730 | const CapturedDecl *CD = CS.getCapturedDecl(); |
| 731 | const RecordDecl *RD = CS.getCapturedRecordDecl(); |
| 732 | |
| 733 | CXXThisValue = nullptr; |
| 734 | CodeGenModule &CGM = CGF.CGM; |
| 735 | ASTContext &Ctx = CGM.getContext(); |
| 736 | Args.push_back(Elt: CD->getContextParam()); |
| 737 | |
| 738 | const CGFunctionInfo &FuncInfo = |
| 739 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(resultType: Ctx.VoidTy, args: Args); |
| 740 | llvm::FunctionType *FuncLLVMTy = CGM.getTypes().GetFunctionType(Info: FuncInfo); |
| 741 | |
| 742 | auto *F = |
| 743 | llvm::Function::Create(Ty: FuncLLVMTy, Linkage: llvm::GlobalValue::InternalLinkage, |
| 744 | N: FunctionName, M: &CGM.getModule()); |
| 745 | CGM.SetInternalFunctionAttributes(GD: CD, F, FI: FuncInfo); |
| 746 | if (CD->isNothrow()) |
| 747 | F->setDoesNotThrow(); |
| 748 | F->setDoesNotRecurse(); |
| 749 | |
| 750 | CGF.StartFunction(GD: CD, RetTy: Ctx.VoidTy, Fn: F, FnInfo: FuncInfo, Args, Loc, StartLoc: Loc); |
| 751 | Address ContextAddr = CGF.GetAddrOfLocalVar(VD: CD->getContextParam()); |
| 752 | ContextV = CGF.Builder.CreateLoad(Addr: ContextAddr); |
| 753 | |
| 754 | // The runtime passes arguments as an array of pointers. |
| 755 | llvm::Type *PtrTy = CGF.Builder.getPtrTy(); |
| 756 | llvm::Align PtrAlign = CGM.getDataLayout().getPointerABIAlignment(AS: 0); |
| 757 | CharUnits SlotAlign = CharUnits::fromQuantity(Quantity: PtrAlign.value()); |
| 758 | |
| 759 | for (auto [FD, C, FieldIdx] : |
| 760 | llvm::zip(t: RD->fields(), u: CS.captures(), |
| 761 | args: llvm::seq<unsigned>(Size: RD->getNumFields()))) { |
| 762 | llvm::Value *SlotPtr = |
| 763 | CGF.Builder.CreateConstInBoundsGEP1_32(Ty: PtrTy, Ptr: ContextV, Idx0: FieldIdx); |
| 764 | llvm::Value *Slot = CGF.Builder.CreateAlignedLoad(Ty: PtrTy, Ptr: SlotPtr, Align: PtrAlign); |
| 765 | |
| 766 | // Generate the appropriate load from the per-argument storage. This |
| 767 | // includes all of the user arguments as well as the implicit kernel |
| 768 | // argument pointer. |
| 769 | if (C.capturesVariableByCopy() && FD->getType()->isAnyPointerType()) { |
| 770 | const VarDecl *CurVD = C.getCapturedVar(); |
| 771 | Slot->setName(CurVD->getName()); |
| 772 | Address SlotAddr(Slot, PtrTy, SlotAlign); |
| 773 | LocalAddrs.insert(KV: {FD, {CurVD, SlotAddr}}); |
| 774 | } else if (FD->hasCapturedVLAType()) { |
| 775 | // VLA size is stored as intptr_t directly in the slot. |
| 776 | Address SlotAddr(Slot, CGF.ConvertTypeForMem(T: FD->getType()), SlotAlign); |
| 777 | LValue ArgLVal = |
| 778 | CGF.MakeAddrLValue(Addr: SlotAddr, T: FD->getType(), Source: AlignmentSource::Decl); |
| 779 | llvm::Value *ExprArg = CGF.EmitLoadOfScalar(lvalue: ArgLVal, Loc: C.getLocation()); |
| 780 | const VariableArrayType *VAT = FD->getCapturedVLAType(); |
| 781 | VLASizes.try_emplace(Key: FD, Args: VAT->getSizeExpr(), Args&: ExprArg); |
| 782 | } else if (C.capturesVariable()) { |
| 783 | const VarDecl *Var = C.getCapturedVar(); |
| 784 | QualType VarTy = Var->getType(); |
| 785 | |
| 786 | if (VarTy->isVariablyModifiedType() && VarTy->isPointerType()) { |
| 787 | Slot->setName(Var->getName() + ".addr"); |
| 788 | Address SlotAddr(Slot, PtrTy, SlotAlign); |
| 789 | LocalAddrs.insert(KV: {FD, {Var, SlotAddr}}); |
| 790 | } else { |
| 791 | llvm::Value *VarAddr = CGF.Builder.CreateAlignedLoad( |
| 792 | Ty: PtrTy, Ptr: Slot, Align: PtrAlign, Name: Var->getName()); |
| 793 | LocalAddrs.insert(KV: {FD, |
| 794 | {Var, Address(VarAddr, CGF.ConvertTypeForMem(T: VarTy), |
| 795 | Ctx.getDeclAlign(D: Var))}}); |
| 796 | } |
| 797 | } else if (C.capturesVariableByCopy()) { |
| 798 | assert(!FD->getType()->isAnyPointerType() && |
| 799 | "Not expecting a captured pointer."); |
| 800 | const VarDecl *Var = C.getCapturedVar(); |
| 801 | QualType FieldTy = FD->getType(); |
| 802 | |
| 803 | // Scalar values are promoted and stored directly in the slot. |
| 804 | Address SlotAddr(Slot, CGF.ConvertTypeForMem(T: FieldTy), SlotAlign); |
| 805 | Address CopyAddr = |
| 806 | CGF.CreateMemTemp(T: FieldTy, Align: Ctx.getDeclAlign(D: FD), Name: Var->getName()); |
| 807 | LValue SrcLVal = |
| 808 | CGF.MakeAddrLValue(Addr: SlotAddr, T: FieldTy, Source: AlignmentSource::Decl); |
| 809 | LValue CopyLVal = |
| 810 | CGF.MakeAddrLValue(Addr: CopyAddr, T: FieldTy, Source: AlignmentSource::Decl); |
| 811 | |
| 812 | RValue ArgRVal = CGF.EmitLoadOfLValue(V: SrcLVal, Loc: C.getLocation()); |
| 813 | CGF.EmitStoreThroughLValue(Src: ArgRVal, Dst: CopyLVal); |
| 814 | |
| 815 | LocalAddrs.insert(KV: {FD, {Var, CopyAddr}}); |
| 816 | } else { |
| 817 | assert(C.capturesThis() && "Default case expected to be CXX 'this'"); |
| 818 | CXXThisValue = |
| 819 | CGF.Builder.CreateAlignedLoad(Ty: PtrTy, Ptr: Slot, Align: PtrAlign, Name: "this"); |
| 820 | Address SlotAddr(Slot, PtrTy, SlotAlign); |
| 821 | LocalAddrs.insert(KV: {FD, {nullptr, SlotAddr}}); |
| 822 | } |
| 823 | } |
| 824 | |
| 825 | return F; |
| 826 | } |
| 827 | |
| 828 | llvm::Function *CodeGenFunction::GenerateOpenMPCapturedStmtFunction( |
| 829 | const CapturedStmt &S, const OMPExecutableDirective &D) { |
| 830 | SourceLocation Loc = D.getBeginLoc(); |
| 831 | assert( |
| 832 | CapturedStmtInfo && |
| 833 | "CapturedStmtInfo should be set when generating the captured function"); |
| 834 | const CapturedDecl *CD = S.getCapturedDecl(); |
| 835 | // Build the argument list. |
| 836 | bool NeedWrapperFunction = |
| 837 | getDebugInfo() && CGM.getCodeGenOpts().hasReducedDebugInfo(); |
| 838 | FunctionArgList Args, WrapperArgs; |
| 839 | llvm::MapVector<const Decl *, std::pair<const VarDecl *, Address>> LocalAddrs, |
| 840 | WrapperLocalAddrs; |
| 841 | llvm::DenseMap<const Decl *, std::pair<const Expr *, llvm::Value *>> VLASizes, |
| 842 | WrapperVLASizes; |
| 843 | SmallString<256> Buffer; |
| 844 | llvm::raw_svector_ostream Out(Buffer); |
| 845 | Out << CapturedStmtInfo->getHelperName(); |
| 846 | OpenMPDirectiveKind EKind = getEffectiveDirectiveKind(S: D); |
| 847 | bool IsDeviceKernel = CGM.getOpenMPRuntime().isGPU() && |
| 848 | isOpenMPTargetExecutionDirective(DKind: EKind) && |
| 849 | D.getCapturedStmt(RegionKind: OMPD_target) == &S; |
| 850 | CodeGenFunction WrapperCGF(CGM, /*suppressNewContext=*/true); |
| 851 | llvm::Function *WrapperF = nullptr; |
| 852 | if (NeedWrapperFunction) { |
| 853 | // Emit the final kernel early to allow attributes to be added by the |
| 854 | // OpenMPI-IR-Builder. |
| 855 | FunctionOptions WrapperFO(&S, /*UIntPtrCastRequired=*/true, |
| 856 | /*RegisterCastedArgsOnly=*/true, |
| 857 | CapturedStmtInfo->getHelperName(), Loc, |
| 858 | IsDeviceKernel); |
| 859 | WrapperCGF.CapturedStmtInfo = CapturedStmtInfo; |
| 860 | WrapperF = |
| 861 | emitOutlinedFunctionPrologue(CGF&: WrapperCGF, Args, LocalAddrs, VLASizes, |
| 862 | CXXThisValue&: WrapperCGF.CXXThisValue, FO: WrapperFO); |
| 863 | Out << "_debug__"; |
| 864 | } |
| 865 | FunctionOptions FO(&S, !NeedWrapperFunction, /*RegisterCastedArgsOnly=*/false, |
| 866 | Out.str(), Loc, !NeedWrapperFunction && IsDeviceKernel); |
| 867 | llvm::Function *F = emitOutlinedFunctionPrologue( |
| 868 | CGF&: *this, Args&: WrapperArgs, LocalAddrs&: WrapperLocalAddrs, VLASizes&: WrapperVLASizes, CXXThisValue, FO); |
| 869 | CodeGenFunction::OMPPrivateScope LocalScope(*this); |
| 870 | for (const auto &LocalAddrPair : WrapperLocalAddrs) { |
| 871 | if (LocalAddrPair.second.first) { |
| 872 | LocalScope.addPrivate(LocalVD: LocalAddrPair.second.first, |
| 873 | Addr: LocalAddrPair.second.second); |
| 874 | } |
| 875 | } |
| 876 | (void)LocalScope.Privatize(); |
| 877 | for (const auto &VLASizePair : WrapperVLASizes) |
| 878 | VLASizeMap[VLASizePair.second.first] = VLASizePair.second.second; |
| 879 | PGO->assignRegionCounters(GD: GlobalDecl(CD), Fn: F); |
| 880 | CapturedStmtInfo->EmitBody(CGF&: *this, S: CD->getBody()); |
| 881 | LocalScope.ForceCleanup(); |
| 882 | FinishFunction(EndLoc: CD->getBodyRBrace()); |
| 883 | if (!NeedWrapperFunction) |
| 884 | return F; |
| 885 | |
| 886 | // Reverse the order. |
| 887 | WrapperF->removeFromParent(); |
| 888 | F->getParent()->getFunctionList().insertAfter(where: F->getIterator(), New: WrapperF); |
| 889 | |
| 890 | llvm::SmallVector<llvm::Value *, 4> CallArgs; |
| 891 | auto *PI = F->arg_begin(); |
| 892 | for (const auto *Arg : Args) { |
| 893 | llvm::Value *CallArg; |
| 894 | auto I = LocalAddrs.find(Key: Arg); |
| 895 | if (I != LocalAddrs.end()) { |
| 896 | LValue LV = WrapperCGF.MakeAddrLValue( |
| 897 | Addr: I->second.second, |
| 898 | T: I->second.first ? I->second.first->getType() : Arg->getType(), |
| 899 | Source: AlignmentSource::Decl); |
| 900 | if (LV.getType()->isAnyComplexType()) |
| 901 | LV.setAddress(LV.getAddress().withElementType(ElemTy: PI->getType())); |
| 902 | CallArg = WrapperCGF.EmitLoadOfScalar(lvalue: LV, Loc: S.getBeginLoc()); |
| 903 | } else { |
| 904 | auto EI = VLASizes.find(Val: Arg); |
| 905 | if (EI != VLASizes.end()) { |
| 906 | CallArg = EI->second.second; |
| 907 | } else { |
| 908 | LValue LV = |
| 909 | WrapperCGF.MakeAddrLValue(Addr: WrapperCGF.GetAddrOfLocalVar(VD: Arg), |
| 910 | T: Arg->getType(), Source: AlignmentSource::Decl); |
| 911 | CallArg = WrapperCGF.EmitLoadOfScalar(lvalue: LV, Loc: S.getBeginLoc()); |
| 912 | } |
| 913 | } |
| 914 | CallArgs.emplace_back(Args: WrapperCGF.EmitFromMemory(Value: CallArg, Ty: Arg->getType())); |
| 915 | ++PI; |
| 916 | } |
| 917 | CGM.getOpenMPRuntime().emitOutlinedFunctionCall(CGF&: WrapperCGF, Loc, OutlinedFn: F, Args: CallArgs); |
| 918 | WrapperCGF.FinishFunction(); |
| 919 | return WrapperF; |
| 920 | } |
| 921 | |
| 922 | llvm::Function *CodeGenFunction::GenerateOpenMPCapturedStmtFunctionAggregate( |
| 923 | const CapturedStmt &S, const OMPExecutableDirective &D) { |
| 924 | SourceLocation Loc = D.getBeginLoc(); |
| 925 | assert( |
| 926 | CapturedStmtInfo && |
| 927 | "CapturedStmtInfo should be set when generating the captured function"); |
| 928 | const CapturedDecl *CD = S.getCapturedDecl(); |
| 929 | const RecordDecl *RD = S.getCapturedRecordDecl(); |
| 930 | StringRef FunctionName = CapturedStmtInfo->getHelperName(); |
| 931 | bool NeedWrapperFunction = |
| 932 | getDebugInfo() && CGM.getCodeGenOpts().hasReducedDebugInfo(); |
| 933 | |
| 934 | CodeGenFunction WrapperCGF(CGM, /*suppressNewContext=*/true); |
| 935 | llvm::Function *WrapperF = nullptr; |
| 936 | llvm::Value *WrapperContextV = nullptr; |
| 937 | if (NeedWrapperFunction) { |
| 938 | WrapperCGF.CapturedStmtInfo = CapturedStmtInfo; |
| 939 | FunctionArgList WrapperArgs; |
| 940 | llvm::MapVector<const Decl *, std::pair<const VarDecl *, Address>> |
| 941 | WrapperLocalAddrs; |
| 942 | llvm::DenseMap<const Decl *, std::pair<const Expr *, llvm::Value *>> |
| 943 | WrapperVLASizes; |
| 944 | WrapperF = emitOutlinedFunctionPrologueAggregate( |
| 945 | CGF&: WrapperCGF, Args&: WrapperArgs, LocalAddrs&: WrapperLocalAddrs, VLASizes&: WrapperVLASizes, |
| 946 | CXXThisValue&: WrapperCGF.CXXThisValue, ContextV&: WrapperContextV, CS: S, Loc, FunctionName); |
| 947 | } |
| 948 | |
| 949 | FunctionArgList Args; |
| 950 | llvm::MapVector<const Decl *, std::pair<const VarDecl *, Address>> LocalAddrs; |
| 951 | llvm::DenseMap<const Decl *, std::pair<const Expr *, llvm::Value *>> VLASizes; |
| 952 | llvm::Function *F; |
| 953 | |
| 954 | if (NeedWrapperFunction) { |
| 955 | SmallString<256> Buffer; |
| 956 | llvm::raw_svector_ostream Out(Buffer); |
| 957 | Out << FunctionName << "_debug__"; |
| 958 | |
| 959 | FunctionOptions FO(&S, /*UIntPtrCastRequired=*/false, |
| 960 | /*RegisterCastedArgsOnly=*/false, Out.str(), Loc, |
| 961 | /*IsDeviceKernel=*/false); |
| 962 | F = emitOutlinedFunctionPrologue(CGF&: *this, Args, LocalAddrs, VLASizes, |
| 963 | CXXThisValue, FO); |
| 964 | } else { |
| 965 | llvm::Value *ContextV = nullptr; |
| 966 | F = emitOutlinedFunctionPrologueAggregate(CGF&: *this, Args, LocalAddrs, VLASizes, |
| 967 | CXXThisValue, ContextV, CS: S, Loc, |
| 968 | FunctionName); |
| 969 | |
| 970 | const RecordDecl *RD = S.getCapturedRecordDecl(); |
| 971 | unsigned FieldIdx = RD->getNumFields(); |
| 972 | for (unsigned I = 0; I < CD->getNumParams(); ++I) { |
| 973 | const ImplicitParamDecl *Param = CD->getParam(i: I); |
| 974 | if (Param == CD->getContextParam()) |
| 975 | continue; |
| 976 | llvm::Align PtrAlign = CGM.getDataLayout().getPointerABIAlignment(AS: 0); |
| 977 | llvm::Value *SlotPtr = Builder.CreateConstInBoundsGEP1_32( |
| 978 | Ty: Builder.getPtrTy(), Ptr: ContextV, Idx0: FieldIdx, |
| 979 | Name: Twine(Param->getName()) + ".addr"); |
| 980 | llvm::Value *ParamAddr = |
| 981 | Builder.CreateAlignedLoad(Ty: Builder.getPtrTy(), Ptr: SlotPtr, Align: PtrAlign); |
| 982 | llvm::Value *ParamVal = Builder.CreateAlignedLoad( |
| 983 | Ty: Builder.getPtrTy(), Ptr: ParamAddr, Align: PtrAlign, Name: Param->getName()); |
| 984 | Address ParamLocalAddr = |
| 985 | CreateMemTemp(T: Param->getType(), Name: Param->getName()); |
| 986 | Builder.CreateStore(Val: ParamVal, Addr: ParamLocalAddr); |
| 987 | LocalAddrs.insert(KV: {Param, {Param, ParamLocalAddr}}); |
| 988 | ++FieldIdx; |
| 989 | } |
| 990 | } |
| 991 | |
| 992 | CodeGenFunction::OMPPrivateScope LocalScope(*this); |
| 993 | for (const auto &LocalAddrPair : LocalAddrs) { |
| 994 | if (LocalAddrPair.second.first) |
| 995 | LocalScope.addPrivate(LocalVD: LocalAddrPair.second.first, |
| 996 | Addr: LocalAddrPair.second.second); |
| 997 | } |
| 998 | (void)LocalScope.Privatize(); |
| 999 | for (const auto &VLASizePair : VLASizes) |
| 1000 | VLASizeMap[VLASizePair.second.first] = VLASizePair.second.second; |
| 1001 | PGO->assignRegionCounters(GD: GlobalDecl(CD), Fn: F); |
| 1002 | CapturedStmtInfo->EmitBody(CGF&: *this, S: CD->getBody()); |
| 1003 | (void)LocalScope.ForceCleanup(); |
| 1004 | FinishFunction(EndLoc: CD->getBodyRBrace()); |
| 1005 | |
| 1006 | if (!NeedWrapperFunction) |
| 1007 | return F; |
| 1008 | |
| 1009 | // Reverse the order. |
| 1010 | WrapperF->removeFromParent(); |
| 1011 | F->getParent()->getFunctionList().insertAfter(where: F->getIterator(), New: WrapperF); |
| 1012 | |
| 1013 | llvm::Align PtrAlign = CGM.getDataLayout().getPointerABIAlignment(AS: 0); |
| 1014 | llvm::SmallVector<llvm::Value *, 16> CallArgs; |
| 1015 | assert(CD->getContextParamPosition() == 0 && |
| 1016 | "Expected context param at position 0 for target regions"); |
| 1017 | assert(RD->getNumFields() + 1 == F->getNumOperands() && |
| 1018 | "Argument count mismatch"); |
| 1019 | |
| 1020 | for (auto [FD, InnerParam, SlotIdx] : llvm::zip( |
| 1021 | t: RD->fields(), u: F->args(), args: llvm::seq<unsigned>(Size: RD->getNumFields()))) { |
| 1022 | llvm::Value *SlotPtr = WrapperCGF.Builder.CreateConstInBoundsGEP1_32( |
| 1023 | Ty: WrapperCGF.Builder.getPtrTy(), Ptr: WrapperContextV, Idx0: SlotIdx); |
| 1024 | llvm::Value *Slot = WrapperCGF.Builder.CreateAlignedLoad( |
| 1025 | Ty: WrapperCGF.Builder.getPtrTy(), Ptr: SlotPtr, Align: PtrAlign); |
| 1026 | llvm::Value *Val = WrapperCGF.Builder.CreateAlignedLoad( |
| 1027 | Ty: InnerParam.getType(), Ptr: Slot, Align: PtrAlign, Name: InnerParam.getName()); |
| 1028 | CallArgs.push_back(Elt: Val); |
| 1029 | } |
| 1030 | |
| 1031 | // Handle the load from the implicit dyn_ptr at the end of the __context. |
| 1032 | unsigned SlotIdx = RD->getNumFields(); |
| 1033 | auto InnerParam = F->arg_begin() + SlotIdx; |
| 1034 | llvm::Value *SlotPtr = WrapperCGF.Builder.CreateConstInBoundsGEP1_32( |
| 1035 | Ty: WrapperCGF.Builder.getPtrTy(), Ptr: WrapperContextV, Idx0: SlotIdx); |
| 1036 | llvm::Value *Slot = WrapperCGF.Builder.CreateAlignedLoad( |
| 1037 | Ty: WrapperCGF.Builder.getPtrTy(), Ptr: SlotPtr, Align: PtrAlign); |
| 1038 | llvm::Value *Val = WrapperCGF.Builder.CreateAlignedLoad( |
| 1039 | Ty: InnerParam->getType(), Ptr: Slot, Align: PtrAlign, Name: InnerParam->getName()); |
| 1040 | CallArgs.push_back(Elt: Val); |
| 1041 | |
| 1042 | CGM.getOpenMPRuntime().emitOutlinedFunctionCall(CGF&: WrapperCGF, Loc, OutlinedFn: F, Args: CallArgs); |
| 1043 | WrapperCGF.FinishFunction(); |
| 1044 | return WrapperF; |
| 1045 | } |
| 1046 | |
| 1047 | //===----------------------------------------------------------------------===// |
| 1048 | // OpenMP Directive Emission |
| 1049 | //===----------------------------------------------------------------------===// |
| 1050 | void CodeGenFunction::EmitOMPAggregateAssign( |
| 1051 | Address DestAddr, Address SrcAddr, QualType OriginalType, |
| 1052 | const llvm::function_ref<void(Address, Address)> CopyGen) { |
| 1053 | // Perform element-by-element initialization. |
| 1054 | QualType ElementTy; |
| 1055 | |
| 1056 | // Drill down to the base element type on both arrays. |
| 1057 | const ArrayType *ArrayTy = OriginalType->getAsArrayTypeUnsafe(); |
| 1058 | llvm::Value *NumElements = emitArrayLength(arrayType: ArrayTy, baseType&: ElementTy, addr&: DestAddr); |
| 1059 | SrcAddr = SrcAddr.withElementType(ElemTy: DestAddr.getElementType()); |
| 1060 | |
| 1061 | llvm::Value *SrcBegin = SrcAddr.emitRawPointer(CGF&: *this); |
| 1062 | llvm::Value *DestBegin = DestAddr.emitRawPointer(CGF&: *this); |
| 1063 | // Cast from pointer to array type to pointer to single element. |
| 1064 | llvm::Value *DestEnd = Builder.CreateInBoundsGEP(Ty: DestAddr.getElementType(), |
| 1065 | Ptr: DestBegin, IdxList: NumElements); |
| 1066 | |
| 1067 | // The basic structure here is a while-do loop. |
| 1068 | llvm::BasicBlock *BodyBB = createBasicBlock(name: "omp.arraycpy.body"); |
| 1069 | llvm::BasicBlock *DoneBB = createBasicBlock(name: "omp.arraycpy.done"); |
| 1070 | llvm::Value *IsEmpty = |
| 1071 | Builder.CreateICmpEQ(LHS: DestBegin, RHS: DestEnd, Name: "omp.arraycpy.isempty"); |
| 1072 | Builder.CreateCondBr(Cond: IsEmpty, True: DoneBB, False: BodyBB); |
| 1073 | |
| 1074 | // Enter the loop body, making that address the current address. |
| 1075 | llvm::BasicBlock *EntryBB = Builder.GetInsertBlock(); |
| 1076 | EmitBlock(BB: BodyBB); |
| 1077 | |
| 1078 | CharUnits ElementSize = getContext().getTypeSizeInChars(T: ElementTy); |
| 1079 | |
| 1080 | llvm::PHINode *SrcElementPHI = |
| 1081 | Builder.CreatePHI(Ty: SrcBegin->getType(), NumReservedValues: 2, Name: "omp.arraycpy.srcElementPast"); |
| 1082 | SrcElementPHI->addIncoming(V: SrcBegin, BB: EntryBB); |
| 1083 | Address SrcElementCurrent = |
| 1084 | Address(SrcElementPHI, SrcAddr.getElementType(), |
| 1085 | SrcAddr.getAlignment().alignmentOfArrayElement(elementSize: ElementSize)); |
| 1086 | |
| 1087 | llvm::PHINode *DestElementPHI = Builder.CreatePHI( |
| 1088 | Ty: DestBegin->getType(), NumReservedValues: 2, Name: "omp.arraycpy.destElementPast"); |
| 1089 | DestElementPHI->addIncoming(V: DestBegin, BB: EntryBB); |
| 1090 | Address DestElementCurrent = |
| 1091 | Address(DestElementPHI, DestAddr.getElementType(), |
| 1092 | DestAddr.getAlignment().alignmentOfArrayElement(elementSize: ElementSize)); |
| 1093 | |
| 1094 | // Emit copy. |
| 1095 | CopyGen(DestElementCurrent, SrcElementCurrent); |
| 1096 | |
| 1097 | // Shift the address forward by one element. |
| 1098 | llvm::Value *DestElementNext = |
| 1099 | Builder.CreateConstGEP1_32(Ty: DestAddr.getElementType(), Ptr: DestElementPHI, |
| 1100 | /*Idx0=*/1, Name: "omp.arraycpy.dest.element"); |
| 1101 | llvm::Value *SrcElementNext = |
| 1102 | Builder.CreateConstGEP1_32(Ty: SrcAddr.getElementType(), Ptr: SrcElementPHI, |
| 1103 | /*Idx0=*/1, Name: "omp.arraycpy.src.element"); |
| 1104 | // Check whether we've reached the end. |
| 1105 | llvm::Value *Done = |
| 1106 | Builder.CreateICmpEQ(LHS: DestElementNext, RHS: DestEnd, Name: "omp.arraycpy.done"); |
| 1107 | Builder.CreateCondBr(Cond: Done, True: DoneBB, False: BodyBB); |
| 1108 | DestElementPHI->addIncoming(V: DestElementNext, BB: Builder.GetInsertBlock()); |
| 1109 | SrcElementPHI->addIncoming(V: SrcElementNext, BB: Builder.GetInsertBlock()); |
| 1110 | |
| 1111 | // Done. |
| 1112 | EmitBlock(BB: DoneBB, /*IsFinished=*/true); |
| 1113 | } |
| 1114 | |
| 1115 | void CodeGenFunction::EmitOMPCopy(QualType OriginalType, Address DestAddr, |
| 1116 | Address SrcAddr, const VarDecl *DestVD, |
| 1117 | const VarDecl *SrcVD, const Expr *Copy) { |
| 1118 | if (OriginalType->isArrayType()) { |
| 1119 | const auto *BO = dyn_cast<BinaryOperator>(Val: Copy); |
| 1120 | if (BO && BO->getOpcode() == BO_Assign) { |
| 1121 | // Perform simple memcpy for simple copying. |
| 1122 | LValue Dest = MakeAddrLValue(Addr: DestAddr, T: OriginalType); |
| 1123 | LValue Src = MakeAddrLValue(Addr: SrcAddr, T: OriginalType); |
| 1124 | EmitAggregateAssign(Dest, Src, EltTy: OriginalType); |
| 1125 | } else { |
| 1126 | // For arrays with complex element types perform element by element |
| 1127 | // copying. |
| 1128 | EmitOMPAggregateAssign( |
| 1129 | DestAddr, SrcAddr, OriginalType, |
| 1130 | CopyGen: [this, Copy, SrcVD, DestVD](Address DestElement, Address SrcElement) { |
| 1131 | // Working with the single array element, so have to remap |
| 1132 | // destination and source variables to corresponding array |
| 1133 | // elements. |
| 1134 | CodeGenFunction::OMPPrivateScope Remap(*this); |
| 1135 | Remap.addPrivate(LocalVD: DestVD, Addr: DestElement); |
| 1136 | Remap.addPrivate(LocalVD: SrcVD, Addr: SrcElement); |
| 1137 | (void)Remap.Privatize(); |
| 1138 | EmitIgnoredExpr(E: Copy); |
| 1139 | }); |
| 1140 | } |
| 1141 | } else { |
| 1142 | // Remap pseudo source variable to private copy. |
| 1143 | CodeGenFunction::OMPPrivateScope Remap(*this); |
| 1144 | Remap.addPrivate(LocalVD: SrcVD, Addr: SrcAddr); |
| 1145 | Remap.addPrivate(LocalVD: DestVD, Addr: DestAddr); |
| 1146 | (void)Remap.Privatize(); |
| 1147 | // Emit copying of the whole variable. |
| 1148 | EmitIgnoredExpr(E: Copy); |
| 1149 | } |
| 1150 | } |
| 1151 | |
| 1152 | bool CodeGenFunction::EmitOMPFirstprivateClause(const OMPExecutableDirective &D, |
| 1153 | OMPPrivateScope &PrivateScope) { |
| 1154 | if (!HaveInsertPoint()) |
| 1155 | return false; |
| 1156 | OpenMPDirectiveKind EKind = getEffectiveDirectiveKind(S: D); |
| 1157 | bool DeviceConstTarget = getLangOpts().OpenMPIsTargetDevice && |
| 1158 | isOpenMPTargetExecutionDirective(DKind: EKind); |
| 1159 | bool FirstprivateIsLastprivate = false; |
| 1160 | llvm::DenseMap<const VarDecl *, OpenMPLastprivateModifier> Lastprivates; |
| 1161 | for (const auto *C : D.getClausesOfKind<OMPLastprivateClause>()) { |
| 1162 | for (const auto *D : C->varlist()) |
| 1163 | Lastprivates.try_emplace( |
| 1164 | Key: cast<VarDecl>(Val: cast<DeclRefExpr>(Val: D)->getDecl())->getCanonicalDecl(), |
| 1165 | Args: C->getKind()); |
| 1166 | } |
| 1167 | llvm::DenseSet<const VarDecl *> EmittedAsFirstprivate; |
| 1168 | llvm::SmallVector<OpenMPDirectiveKind, 4> CaptureRegions; |
| 1169 | getOpenMPCaptureRegions(CaptureRegions, DKind: EKind); |
| 1170 | // Force emission of the firstprivate copy if the directive does not emit |
| 1171 | // outlined function, like omp for, omp simd, omp distribute etc. |
| 1172 | bool MustEmitFirstprivateCopy = |
| 1173 | CaptureRegions.size() == 1 && CaptureRegions.back() == OMPD_unknown; |
| 1174 | for (const auto *C : D.getClausesOfKind<OMPFirstprivateClause>()) { |
| 1175 | const auto *IRef = C->varlist_begin(); |
| 1176 | const auto *InitsRef = C->inits().begin(); |
| 1177 | for (const Expr *IInit : C->private_copies()) { |
| 1178 | const auto *OrigVD = cast<VarDecl>(Val: cast<DeclRefExpr>(Val: *IRef)->getDecl()); |
| 1179 | bool ThisFirstprivateIsLastprivate = |
| 1180 | Lastprivates.count(Val: OrigVD->getCanonicalDecl()) > 0; |
| 1181 | const FieldDecl *FD = CapturedStmtInfo->lookup(VD: OrigVD); |
| 1182 | const auto *VD = cast<VarDecl>(Val: cast<DeclRefExpr>(Val: IInit)->getDecl()); |
| 1183 | if (!MustEmitFirstprivateCopy && !ThisFirstprivateIsLastprivate && FD && |
| 1184 | !FD->getType()->isReferenceType() && |
| 1185 | (!VD || !VD->hasAttr<OMPAllocateDeclAttr>())) { |
| 1186 | EmittedAsFirstprivate.insert(V: OrigVD->getCanonicalDecl()); |
| 1187 | ++IRef; |
| 1188 | ++InitsRef; |
| 1189 | continue; |
| 1190 | } |
| 1191 | // Do not emit copy for firstprivate constant variables in target regions, |
| 1192 | // captured by reference. |
| 1193 | if (DeviceConstTarget && OrigVD->getType().isConstant(Ctx: getContext()) && |
| 1194 | FD && FD->getType()->isReferenceType() && |
| 1195 | (!VD || !VD->hasAttr<OMPAllocateDeclAttr>())) { |
| 1196 | EmittedAsFirstprivate.insert(V: OrigVD->getCanonicalDecl()); |
| 1197 | ++IRef; |
| 1198 | ++InitsRef; |
| 1199 | continue; |
| 1200 | } |
| 1201 | FirstprivateIsLastprivate = |
| 1202 | FirstprivateIsLastprivate || ThisFirstprivateIsLastprivate; |
| 1203 | if (EmittedAsFirstprivate.insert(V: OrigVD->getCanonicalDecl()).second) { |
| 1204 | const auto *VDInit = |
| 1205 | cast<VarDecl>(Val: cast<DeclRefExpr>(Val: *InitsRef)->getDecl()); |
| 1206 | bool IsRegistered; |
| 1207 | DeclRefExpr DRE(getContext(), const_cast<VarDecl *>(OrigVD), |
| 1208 | /*RefersToEnclosingVariableOrCapture=*/FD != nullptr, |
| 1209 | (*IRef)->getType(), VK_LValue, (*IRef)->getExprLoc()); |
| 1210 | LValue OriginalLVal; |
| 1211 | if (!FD) { |
| 1212 | // Check if the firstprivate variable is just a constant value. |
| 1213 | ConstantEmission CE = tryEmitAsConstant(RefExpr: &DRE); |
| 1214 | if (CE && !CE.isReference()) { |
| 1215 | // Constant value, no need to create a copy. |
| 1216 | ++IRef; |
| 1217 | ++InitsRef; |
| 1218 | continue; |
| 1219 | } |
| 1220 | if (CE && CE.isReference()) { |
| 1221 | OriginalLVal = CE.getReferenceLValue(CGF&: *this, RefExpr: &DRE); |
| 1222 | } else { |
| 1223 | assert(!CE && "Expected non-constant firstprivate."); |
| 1224 | OriginalLVal = EmitLValue(E: &DRE); |
| 1225 | } |
| 1226 | } else { |
| 1227 | OriginalLVal = EmitLValue(E: &DRE); |
| 1228 | } |
| 1229 | QualType Type = VD->getType(); |
| 1230 | if (Type->isArrayType()) { |
| 1231 | // Emit VarDecl with copy init for arrays. |
| 1232 | // Get the address of the original variable captured in current |
| 1233 | // captured region. |
| 1234 | AutoVarEmission Emission = EmitAutoVarAlloca(var: *VD); |
| 1235 | const Expr *Init = VD->getInit(); |
| 1236 | if (!isa<CXXConstructExpr>(Val: Init) || isTrivialInitializer(Init)) { |
| 1237 | // Perform simple memcpy. |
| 1238 | LValue Dest = MakeAddrLValue(Addr: Emission.getAllocatedAddress(), T: Type); |
| 1239 | EmitAggregateAssign(Dest, Src: OriginalLVal, EltTy: Type); |
| 1240 | } else { |
| 1241 | EmitOMPAggregateAssign( |
| 1242 | DestAddr: Emission.getAllocatedAddress(), SrcAddr: OriginalLVal.getAddress(), OriginalType: Type, |
| 1243 | CopyGen: [this, VDInit, Init](Address DestElement, Address SrcElement) { |
| 1244 | // Clean up any temporaries needed by the |
| 1245 | // initialization. |
| 1246 | RunCleanupsScope InitScope(*this); |
| 1247 | // Emit initialization for single element. |
| 1248 | setAddrOfLocalVar(VD: VDInit, Addr: SrcElement); |
| 1249 | EmitAnyExprToMem(E: Init, Location: DestElement, |
| 1250 | Quals: Init->getType().getQualifiers(), |
| 1251 | /*IsInitializer*/ false); |
| 1252 | LocalDeclMap.erase(Val: VDInit); |
| 1253 | }); |
| 1254 | } |
| 1255 | EmitAutoVarCleanups(emission: Emission); |
| 1256 | IsRegistered = |
| 1257 | PrivateScope.addPrivate(LocalVD: OrigVD, Addr: Emission.getAllocatedAddress()); |
| 1258 | } else { |
| 1259 | Address OriginalAddr = OriginalLVal.getAddress(); |
| 1260 | // Emit private VarDecl with copy init. |
| 1261 | // Remap temp VDInit variable to the address of the original |
| 1262 | // variable (for proper handling of captured global variables). |
| 1263 | setAddrOfLocalVar(VD: VDInit, Addr: OriginalAddr); |
| 1264 | EmitDecl(D: *VD); |
| 1265 | LocalDeclMap.erase(Val: VDInit); |
| 1266 | Address VDAddr = GetAddrOfLocalVar(VD); |
| 1267 | if (ThisFirstprivateIsLastprivate && |
| 1268 | Lastprivates[OrigVD->getCanonicalDecl()] == |
| 1269 | OMPC_LASTPRIVATE_conditional) { |
| 1270 | // Create/init special variable for lastprivate conditionals. |
| 1271 | llvm::Value *V = |
| 1272 | EmitLoadOfScalar(lvalue: MakeAddrLValue(Addr: VDAddr, T: (*IRef)->getType(), |
| 1273 | Source: AlignmentSource::Decl), |
| 1274 | Loc: (*IRef)->getExprLoc()); |
| 1275 | VDAddr = CGM.getOpenMPRuntime().emitLastprivateConditionalInit( |
| 1276 | CGF&: *this, VD: OrigVD); |
| 1277 | EmitStoreOfScalar(value: V, lvalue: MakeAddrLValue(Addr: VDAddr, T: (*IRef)->getType(), |
| 1278 | Source: AlignmentSource::Decl)); |
| 1279 | LocalDeclMap.erase(Val: VD); |
| 1280 | setAddrOfLocalVar(VD, Addr: VDAddr); |
| 1281 | } |
| 1282 | IsRegistered = PrivateScope.addPrivate(LocalVD: OrigVD, Addr: VDAddr); |
| 1283 | } |
| 1284 | assert(IsRegistered && |
| 1285 | "firstprivate var already registered as private"); |
| 1286 | // Silence the warning about unused variable. |
| 1287 | (void)IsRegistered; |
| 1288 | } |
| 1289 | ++IRef; |
| 1290 | ++InitsRef; |
| 1291 | } |
| 1292 | } |
| 1293 | return FirstprivateIsLastprivate && !EmittedAsFirstprivate.empty(); |
| 1294 | } |
| 1295 | |
| 1296 | void CodeGenFunction::EmitOMPPrivateClause( |
| 1297 | const OMPExecutableDirective &D, |
| 1298 | CodeGenFunction::OMPPrivateScope &PrivateScope) { |
| 1299 | if (!HaveInsertPoint()) |
| 1300 | return; |
| 1301 | llvm::DenseSet<const VarDecl *> EmittedAsPrivate; |
| 1302 | for (const auto *C : D.getClausesOfKind<OMPPrivateClause>()) { |
| 1303 | auto IRef = C->varlist_begin(); |
| 1304 | for (const Expr *IInit : C->private_copies()) { |
| 1305 | const auto *OrigVD = cast<VarDecl>(Val: cast<DeclRefExpr>(Val: *IRef)->getDecl()); |
| 1306 | if (EmittedAsPrivate.insert(V: OrigVD->getCanonicalDecl()).second) { |
| 1307 | const auto *VD = cast<VarDecl>(Val: cast<DeclRefExpr>(Val: IInit)->getDecl()); |
| 1308 | EmitDecl(D: *VD); |
| 1309 | // Emit private VarDecl with copy init. |
| 1310 | bool IsRegistered = |
| 1311 | PrivateScope.addPrivate(LocalVD: OrigVD, Addr: GetAddrOfLocalVar(VD)); |
| 1312 | assert(IsRegistered && "private var already registered as private"); |
| 1313 | // Silence the warning about unused variable. |
| 1314 | (void)IsRegistered; |
| 1315 | } |
| 1316 | ++IRef; |
| 1317 | } |
| 1318 | } |
| 1319 | } |
| 1320 | |
| 1321 | bool CodeGenFunction::EmitOMPCopyinClause(const OMPExecutableDirective &D) { |
| 1322 | if (!HaveInsertPoint()) |
| 1323 | return false; |
| 1324 | // threadprivate_var1 = master_threadprivate_var1; |
| 1325 | // operator=(threadprivate_var2, master_threadprivate_var2); |
| 1326 | // ... |
| 1327 | // __kmpc_barrier(&loc, global_tid); |
| 1328 | llvm::DenseSet<const VarDecl *> CopiedVars; |
| 1329 | llvm::BasicBlock *CopyBegin = nullptr, *CopyEnd = nullptr; |
| 1330 | for (const auto *C : D.getClausesOfKind<OMPCopyinClause>()) { |
| 1331 | auto IRef = C->varlist_begin(); |
| 1332 | auto ISrcRef = C->source_exprs().begin(); |
| 1333 | auto IDestRef = C->destination_exprs().begin(); |
| 1334 | for (const Expr *AssignOp : C->assignment_ops()) { |
| 1335 | const auto *VD = cast<VarDecl>(Val: cast<DeclRefExpr>(Val: *IRef)->getDecl()); |
| 1336 | QualType Type = VD->getType(); |
| 1337 | if (CopiedVars.insert(V: VD->getCanonicalDecl()).second) { |
| 1338 | // Get the address of the master variable. If we are emitting code with |
| 1339 | // TLS support, the address is passed from the master as field in the |
| 1340 | // captured declaration. |
| 1341 | Address MasterAddr = Address::invalid(); |
| 1342 | if (getLangOpts().OpenMPUseTLS && |
| 1343 | getContext().getTargetInfo().isTLSSupported()) { |
| 1344 | assert(CapturedStmtInfo->lookup(VD) && |
| 1345 | "Copyin threadprivates should have been captured!"); |
| 1346 | DeclRefExpr DRE(getContext(), const_cast<VarDecl *>(VD), true, |
| 1347 | (*IRef)->getType(), VK_LValue, (*IRef)->getExprLoc()); |
| 1348 | MasterAddr = EmitLValue(E: &DRE).getAddress(); |
| 1349 | LocalDeclMap.erase(Val: VD); |
| 1350 | } else { |
| 1351 | MasterAddr = |
| 1352 | Address(VD->isStaticLocal() ? CGM.getStaticLocalDeclAddress(D: VD) |
| 1353 | : CGM.GetAddrOfGlobal(GD: VD), |
| 1354 | CGM.getTypes().ConvertTypeForMem(T: VD->getType()), |
| 1355 | getContext().getDeclAlign(D: VD)); |
| 1356 | } |
| 1357 | // Get the address of the threadprivate variable. |
| 1358 | Address PrivateAddr = EmitLValue(E: *IRef).getAddress(); |
| 1359 | if (CopiedVars.size() == 1) { |
| 1360 | // At first check if current thread is a master thread. If it is, no |
| 1361 | // need to copy data. |
| 1362 | CopyBegin = createBasicBlock(name: "copyin.not.master"); |
| 1363 | CopyEnd = createBasicBlock(name: "copyin.not.master.end"); |
| 1364 | // TODO: Avoid ptrtoint conversion. |
| 1365 | auto *MasterAddrInt = Builder.CreatePtrToInt( |
| 1366 | V: MasterAddr.emitRawPointer(CGF&: *this), DestTy: CGM.IntPtrTy); |
| 1367 | auto *PrivateAddrInt = Builder.CreatePtrToInt( |
| 1368 | V: PrivateAddr.emitRawPointer(CGF&: *this), DestTy: CGM.IntPtrTy); |
| 1369 | Builder.CreateCondBr( |
| 1370 | Cond: Builder.CreateICmpNE(LHS: MasterAddrInt, RHS: PrivateAddrInt), True: CopyBegin, |
| 1371 | False: CopyEnd); |
| 1372 | EmitBlock(BB: CopyBegin); |
| 1373 | } |
| 1374 | const auto *SrcVD = |
| 1375 | cast<VarDecl>(Val: cast<DeclRefExpr>(Val: *ISrcRef)->getDecl()); |
| 1376 | const auto *DestVD = |
| 1377 | cast<VarDecl>(Val: cast<DeclRefExpr>(Val: *IDestRef)->getDecl()); |
| 1378 | EmitOMPCopy(OriginalType: Type, DestAddr: PrivateAddr, SrcAddr: MasterAddr, DestVD, SrcVD, Copy: AssignOp); |
| 1379 | } |
| 1380 | ++IRef; |
| 1381 | ++ISrcRef; |
| 1382 | ++IDestRef; |
| 1383 | } |
| 1384 | } |
| 1385 | if (CopyEnd) { |
| 1386 | // Exit out of copying procedure for non-master thread. |
| 1387 | EmitBlock(BB: CopyEnd, /*IsFinished=*/true); |
| 1388 | return true; |
| 1389 | } |
| 1390 | return false; |
| 1391 | } |
| 1392 | |
| 1393 | bool CodeGenFunction::EmitOMPLastprivateClauseInit( |
| 1394 | const OMPExecutableDirective &D, OMPPrivateScope &PrivateScope) { |
| 1395 | if (!HaveInsertPoint()) |
| 1396 | return false; |
| 1397 | bool HasAtLeastOneLastprivate = false; |
| 1398 | OpenMPDirectiveKind EKind = getEffectiveDirectiveKind(S: D); |
| 1399 | llvm::DenseSet<const VarDecl *> SIMDLCVs; |
| 1400 | if (isOpenMPSimdDirective(DKind: EKind)) { |
| 1401 | const auto *LoopDirective = cast<OMPLoopDirective>(Val: &D); |
| 1402 | for (const Expr *C : LoopDirective->counters()) { |
| 1403 | SIMDLCVs.insert( |
| 1404 | V: cast<VarDecl>(Val: cast<DeclRefExpr>(Val: C)->getDecl())->getCanonicalDecl()); |
| 1405 | } |
| 1406 | } |
| 1407 | llvm::DenseSet<const VarDecl *> AlreadyEmittedVars; |
| 1408 | for (const auto *C : D.getClausesOfKind<OMPLastprivateClause>()) { |
| 1409 | HasAtLeastOneLastprivate = true; |
| 1410 | if (isOpenMPTaskLoopDirective(DKind: EKind) && !getLangOpts().OpenMPSimd) |
| 1411 | break; |
| 1412 | const auto *IRef = C->varlist_begin(); |
| 1413 | const auto *IDestRef = C->destination_exprs().begin(); |
| 1414 | for (const Expr *IInit : C->private_copies()) { |
| 1415 | // Keep the address of the original variable for future update at the end |
| 1416 | // of the loop. |
| 1417 | const auto *OrigVD = cast<VarDecl>(Val: cast<DeclRefExpr>(Val: *IRef)->getDecl()); |
| 1418 | // Taskloops do not require additional initialization, it is done in |
| 1419 | // runtime support library. |
| 1420 | if (AlreadyEmittedVars.insert(V: OrigVD->getCanonicalDecl()).second) { |
| 1421 | const auto *DestVD = |
| 1422 | cast<VarDecl>(Val: cast<DeclRefExpr>(Val: *IDestRef)->getDecl()); |
| 1423 | DeclRefExpr DRE(getContext(), const_cast<VarDecl *>(OrigVD), |
| 1424 | /*RefersToEnclosingVariableOrCapture=*/ |
| 1425 | CapturedStmtInfo->lookup(VD: OrigVD) != nullptr, |
| 1426 | (*IRef)->getType(), VK_LValue, (*IRef)->getExprLoc()); |
| 1427 | PrivateScope.addPrivate(LocalVD: DestVD, Addr: EmitLValue(E: &DRE).getAddress()); |
| 1428 | // Check if the variable is also a firstprivate: in this case IInit is |
| 1429 | // not generated. Initialization of this variable will happen in codegen |
| 1430 | // for 'firstprivate' clause. |
| 1431 | if (IInit && !SIMDLCVs.count(V: OrigVD->getCanonicalDecl())) { |
| 1432 | const auto *VD = cast<VarDecl>(Val: cast<DeclRefExpr>(Val: IInit)->getDecl()); |
| 1433 | Address VDAddr = Address::invalid(); |
| 1434 | if (C->getKind() == OMPC_LASTPRIVATE_conditional) { |
| 1435 | VDAddr = CGM.getOpenMPRuntime().emitLastprivateConditionalInit( |
| 1436 | CGF&: *this, VD: OrigVD); |
| 1437 | setAddrOfLocalVar(VD, Addr: VDAddr); |
| 1438 | } else { |
| 1439 | // Emit private VarDecl with copy init. |
| 1440 | EmitDecl(D: *VD); |
| 1441 | VDAddr = GetAddrOfLocalVar(VD); |
| 1442 | } |
| 1443 | bool IsRegistered = PrivateScope.addPrivate(LocalVD: OrigVD, Addr: VDAddr); |
| 1444 | assert(IsRegistered && |
| 1445 | "lastprivate var already registered as private"); |
| 1446 | (void)IsRegistered; |
| 1447 | } |
| 1448 | } |
| 1449 | ++IRef; |
| 1450 | ++IDestRef; |
| 1451 | } |
| 1452 | } |
| 1453 | return HasAtLeastOneLastprivate; |
| 1454 | } |
| 1455 | |
| 1456 | void CodeGenFunction::EmitOMPLastprivateClauseFinal( |
| 1457 | const OMPExecutableDirective &D, bool NoFinals, |
| 1458 | llvm::Value *IsLastIterCond) { |
| 1459 | if (!HaveInsertPoint()) |
| 1460 | return; |
| 1461 | // Emit following code: |
| 1462 | // if (<IsLastIterCond>) { |
| 1463 | // orig_var1 = private_orig_var1; |
| 1464 | // ... |
| 1465 | // orig_varn = private_orig_varn; |
| 1466 | // } |
| 1467 | llvm::BasicBlock *ThenBB = nullptr; |
| 1468 | llvm::BasicBlock *DoneBB = nullptr; |
| 1469 | if (IsLastIterCond) { |
| 1470 | // Emit implicit barrier if at least one lastprivate conditional is found |
| 1471 | // and this is not a simd mode. |
| 1472 | if (!getLangOpts().OpenMPSimd && |
| 1473 | llvm::any_of(Range: D.getClausesOfKind<OMPLastprivateClause>(), |
| 1474 | P: [](const OMPLastprivateClause *C) { |
| 1475 | return C->getKind() == OMPC_LASTPRIVATE_conditional; |
| 1476 | })) { |
| 1477 | CGM.getOpenMPRuntime().emitBarrierCall(CGF&: *this, Loc: D.getBeginLoc(), |
| 1478 | Kind: OMPD_unknown, |
| 1479 | /*EmitChecks=*/false, |
| 1480 | /*ForceSimpleCall=*/true); |
| 1481 | } |
| 1482 | ThenBB = createBasicBlock(name: ".omp.lastprivate.then"); |
| 1483 | DoneBB = createBasicBlock(name: ".omp.lastprivate.done"); |
| 1484 | Builder.CreateCondBr(Cond: IsLastIterCond, True: ThenBB, False: DoneBB); |
| 1485 | EmitBlock(BB: ThenBB); |
| 1486 | } |
| 1487 | llvm::DenseSet<const VarDecl *> AlreadyEmittedVars; |
| 1488 | llvm::DenseMap<const VarDecl *, const Expr *> LoopCountersAndUpdates; |
| 1489 | if (const auto *LoopDirective = dyn_cast<OMPLoopDirective>(Val: &D)) { |
| 1490 | auto IC = LoopDirective->counters().begin(); |
| 1491 | for (const Expr *F : LoopDirective->finals()) { |
| 1492 | const auto *D = |
| 1493 | cast<VarDecl>(Val: cast<DeclRefExpr>(Val: *IC)->getDecl())->getCanonicalDecl(); |
| 1494 | if (NoFinals) |
| 1495 | AlreadyEmittedVars.insert(V: D); |
| 1496 | else |
| 1497 | LoopCountersAndUpdates[D] = F; |
| 1498 | ++IC; |
| 1499 | } |
| 1500 | } |
| 1501 | for (const auto *C : D.getClausesOfKind<OMPLastprivateClause>()) { |
| 1502 | auto IRef = C->varlist_begin(); |
| 1503 | auto ISrcRef = C->source_exprs().begin(); |
| 1504 | auto IDestRef = C->destination_exprs().begin(); |
| 1505 | for (const Expr *AssignOp : C->assignment_ops()) { |
| 1506 | const auto *PrivateVD = |
| 1507 | cast<VarDecl>(Val: cast<DeclRefExpr>(Val: *IRef)->getDecl()); |
| 1508 | QualType Type = PrivateVD->getType(); |
| 1509 | const auto *CanonicalVD = PrivateVD->getCanonicalDecl(); |
| 1510 | if (AlreadyEmittedVars.insert(V: CanonicalVD).second) { |
| 1511 | // If lastprivate variable is a loop control variable for loop-based |
| 1512 | // directive, update its value before copyin back to original |
| 1513 | // variable. |
| 1514 | if (const Expr *FinalExpr = LoopCountersAndUpdates.lookup(Val: CanonicalVD)) |
| 1515 | EmitIgnoredExpr(E: FinalExpr); |
| 1516 | const auto *SrcVD = |
| 1517 | cast<VarDecl>(Val: cast<DeclRefExpr>(Val: *ISrcRef)->getDecl()); |
| 1518 | const auto *DestVD = |
| 1519 | cast<VarDecl>(Val: cast<DeclRefExpr>(Val: *IDestRef)->getDecl()); |
| 1520 | // Get the address of the private variable. |
| 1521 | Address PrivateAddr = GetAddrOfLocalVar(VD: PrivateVD); |
| 1522 | if (const auto *RefTy = PrivateVD->getType()->getAs<ReferenceType>()) |
| 1523 | PrivateAddr = Address( |
| 1524 | Builder.CreateLoad(Addr: PrivateAddr), |
| 1525 | CGM.getTypes().ConvertTypeForMem(T: RefTy->getPointeeType()), |
| 1526 | CGM.getNaturalTypeAlignment(T: RefTy->getPointeeType())); |
| 1527 | // Store the last value to the private copy in the last iteration. |
| 1528 | if (C->getKind() == OMPC_LASTPRIVATE_conditional) |
| 1529 | CGM.getOpenMPRuntime().emitLastprivateConditionalFinalUpdate( |
| 1530 | CGF&: *this, PrivLVal: MakeAddrLValue(Addr: PrivateAddr, T: (*IRef)->getType()), VD: PrivateVD, |
| 1531 | Loc: (*IRef)->getExprLoc()); |
| 1532 | // Get the address of the original variable. |
| 1533 | Address OriginalAddr = GetAddrOfLocalVar(VD: DestVD); |
| 1534 | EmitOMPCopy(OriginalType: Type, DestAddr: OriginalAddr, SrcAddr: PrivateAddr, DestVD, SrcVD, Copy: AssignOp); |
| 1535 | } |
| 1536 | ++IRef; |
| 1537 | ++ISrcRef; |
| 1538 | ++IDestRef; |
| 1539 | } |
| 1540 | if (const Expr *PostUpdate = C->getPostUpdateExpr()) |
| 1541 | EmitIgnoredExpr(E: PostUpdate); |
| 1542 | } |
| 1543 | if (IsLastIterCond) |
| 1544 | EmitBlock(BB: DoneBB, /*IsFinished=*/true); |
| 1545 | } |
| 1546 | |
| 1547 | void CodeGenFunction::EmitOMPReductionClauseInit( |
| 1548 | const OMPExecutableDirective &D, |
| 1549 | CodeGenFunction::OMPPrivateScope &PrivateScope, bool ForInscan) { |
| 1550 | if (!HaveInsertPoint()) |
| 1551 | return; |
| 1552 | SmallVector<const Expr *, 4> Shareds; |
| 1553 | SmallVector<const Expr *, 4> Privates; |
| 1554 | SmallVector<const Expr *, 4> ReductionOps; |
| 1555 | SmallVector<const Expr *, 4> LHSs; |
| 1556 | SmallVector<const Expr *, 4> RHSs; |
| 1557 | OMPTaskDataTy Data; |
| 1558 | SmallVector<const Expr *, 4> TaskLHSs; |
| 1559 | SmallVector<const Expr *, 4> TaskRHSs; |
| 1560 | for (const auto *C : D.getClausesOfKind<OMPReductionClause>()) { |
| 1561 | if (ForInscan != (C->getModifier() == OMPC_REDUCTION_inscan)) |
| 1562 | continue; |
| 1563 | Shareds.append(in_start: C->varlist_begin(), in_end: C->varlist_end()); |
| 1564 | Privates.append(in_start: C->privates().begin(), in_end: C->privates().end()); |
| 1565 | ReductionOps.append(in_start: C->reduction_ops().begin(), in_end: C->reduction_ops().end()); |
| 1566 | LHSs.append(in_start: C->lhs_exprs().begin(), in_end: C->lhs_exprs().end()); |
| 1567 | RHSs.append(in_start: C->rhs_exprs().begin(), in_end: C->rhs_exprs().end()); |
| 1568 | if (C->getModifier() == OMPC_REDUCTION_task) { |
| 1569 | Data.ReductionVars.append(in_start: C->privates().begin(), in_end: C->privates().end()); |
| 1570 | Data.ReductionOrigs.append(in_start: C->varlist_begin(), in_end: C->varlist_end()); |
| 1571 | Data.ReductionCopies.append(in_start: C->privates().begin(), in_end: C->privates().end()); |
| 1572 | Data.ReductionOps.append(in_start: C->reduction_ops().begin(), |
| 1573 | in_end: C->reduction_ops().end()); |
| 1574 | TaskLHSs.append(in_start: C->lhs_exprs().begin(), in_end: C->lhs_exprs().end()); |
| 1575 | TaskRHSs.append(in_start: C->rhs_exprs().begin(), in_end: C->rhs_exprs().end()); |
| 1576 | } |
| 1577 | } |
| 1578 | ReductionCodeGen RedCG(Shareds, Shareds, Privates, ReductionOps); |
| 1579 | unsigned Count = 0; |
| 1580 | auto *ILHS = LHSs.begin(); |
| 1581 | auto *IRHS = RHSs.begin(); |
| 1582 | auto *IPriv = Privates.begin(); |
| 1583 | for (const Expr *IRef : Shareds) { |
| 1584 | const auto *PrivateVD = cast<VarDecl>(Val: cast<DeclRefExpr>(Val: *IPriv)->getDecl()); |
| 1585 | // Emit private VarDecl with reduction init. |
| 1586 | RedCG.emitSharedOrigLValue(CGF&: *this, N: Count); |
| 1587 | RedCG.emitAggregateType(CGF&: *this, N: Count); |
| 1588 | AutoVarEmission Emission = EmitAutoVarAlloca(var: *PrivateVD); |
| 1589 | RedCG.emitInitialization(CGF&: *this, N: Count, PrivateAddr: Emission.getAllocatedAddress(), |
| 1590 | SharedAddr: RedCG.getSharedLValue(N: Count).getAddress(), |
| 1591 | DefaultInit: [&Emission](CodeGenFunction &CGF) { |
| 1592 | CGF.EmitAutoVarInit(emission: Emission); |
| 1593 | return true; |
| 1594 | }); |
| 1595 | EmitAutoVarCleanups(emission: Emission); |
| 1596 | Address BaseAddr = RedCG.adjustPrivateAddress( |
| 1597 | CGF&: *this, N: Count, PrivateAddr: Emission.getAllocatedAddress()); |
| 1598 | bool IsRegistered = |
| 1599 | PrivateScope.addPrivate(LocalVD: RedCG.getBaseDecl(N: Count), Addr: BaseAddr); |
| 1600 | assert(IsRegistered && "private var already registered as private"); |
| 1601 | // Silence the warning about unused variable. |
| 1602 | (void)IsRegistered; |
| 1603 | |
| 1604 | const auto *LHSVD = cast<VarDecl>(Val: cast<DeclRefExpr>(Val: *ILHS)->getDecl()); |
| 1605 | const auto *RHSVD = cast<VarDecl>(Val: cast<DeclRefExpr>(Val: *IRHS)->getDecl()); |
| 1606 | QualType Type = PrivateVD->getType(); |
| 1607 | bool isaOMPArraySectionExpr = isa<ArraySectionExpr>(Val: IRef); |
| 1608 | if (isaOMPArraySectionExpr && Type->isVariablyModifiedType()) { |
| 1609 | // Store the address of the original variable associated with the LHS |
| 1610 | // implicit variable. |
| 1611 | PrivateScope.addPrivate(LocalVD: LHSVD, Addr: RedCG.getSharedLValue(N: Count).getAddress()); |
| 1612 | PrivateScope.addPrivate(LocalVD: RHSVD, Addr: GetAddrOfLocalVar(VD: PrivateVD)); |
| 1613 | } else if ((isaOMPArraySectionExpr && Type->isScalarType()) || |
| 1614 | isa<ArraySubscriptExpr>(Val: IRef)) { |
| 1615 | // Store the address of the original variable associated with the LHS |
| 1616 | // implicit variable. |
| 1617 | PrivateScope.addPrivate(LocalVD: LHSVD, Addr: RedCG.getSharedLValue(N: Count).getAddress()); |
| 1618 | PrivateScope.addPrivate(LocalVD: RHSVD, |
| 1619 | Addr: GetAddrOfLocalVar(VD: PrivateVD).withElementType( |
| 1620 | ElemTy: ConvertTypeForMem(T: RHSVD->getType()))); |
| 1621 | } else { |
| 1622 | QualType Type = PrivateVD->getType(); |
| 1623 | bool IsArray = getContext().getAsArrayType(T: Type) != nullptr; |
| 1624 | Address OriginalAddr = RedCG.getSharedLValue(N: Count).getAddress(); |
| 1625 | // Store the address of the original variable associated with the LHS |
| 1626 | // implicit variable. |
| 1627 | if (IsArray) { |
| 1628 | OriginalAddr = |
| 1629 | OriginalAddr.withElementType(ElemTy: ConvertTypeForMem(T: LHSVD->getType())); |
| 1630 | } |
| 1631 | PrivateScope.addPrivate(LocalVD: LHSVD, Addr: OriginalAddr); |
| 1632 | PrivateScope.addPrivate( |
| 1633 | LocalVD: RHSVD, Addr: IsArray ? GetAddrOfLocalVar(VD: PrivateVD).withElementType( |
| 1634 | ElemTy: ConvertTypeForMem(T: RHSVD->getType())) |
| 1635 | : GetAddrOfLocalVar(VD: PrivateVD)); |
| 1636 | } |
| 1637 | ++ILHS; |
| 1638 | ++IRHS; |
| 1639 | ++IPriv; |
| 1640 | ++Count; |
| 1641 | } |
| 1642 | if (!Data.ReductionVars.empty()) { |
| 1643 | OpenMPDirectiveKind EKind = getEffectiveDirectiveKind(S: D); |
| 1644 | Data.IsReductionWithTaskMod = true; |
| 1645 | Data.IsWorksharingReduction = isOpenMPWorksharingDirective(DKind: EKind); |
| 1646 | llvm::Value *ReductionDesc = CGM.getOpenMPRuntime().emitTaskReductionInit( |
| 1647 | CGF&: *this, Loc: D.getBeginLoc(), LHSExprs: TaskLHSs, RHSExprs: TaskRHSs, Data); |
| 1648 | const Expr *TaskRedRef = nullptr; |
| 1649 | switch (EKind) { |
| 1650 | case OMPD_parallel: |
| 1651 | TaskRedRef = cast<OMPParallelDirective>(Val: D).getTaskReductionRefExpr(); |
| 1652 | break; |
| 1653 | case OMPD_for: |
| 1654 | TaskRedRef = cast<OMPForDirective>(Val: D).getTaskReductionRefExpr(); |
| 1655 | break; |
| 1656 | case OMPD_sections: |
| 1657 | TaskRedRef = cast<OMPSectionsDirective>(Val: D).getTaskReductionRefExpr(); |
| 1658 | break; |
| 1659 | case OMPD_parallel_for: |
| 1660 | TaskRedRef = cast<OMPParallelForDirective>(Val: D).getTaskReductionRefExpr(); |
| 1661 | break; |
| 1662 | case OMPD_parallel_master: |
| 1663 | TaskRedRef = |
| 1664 | cast<OMPParallelMasterDirective>(Val: D).getTaskReductionRefExpr(); |
| 1665 | break; |
| 1666 | case OMPD_parallel_sections: |
| 1667 | TaskRedRef = |
| 1668 | cast<OMPParallelSectionsDirective>(Val: D).getTaskReductionRefExpr(); |
| 1669 | break; |
| 1670 | case OMPD_target_parallel: |
| 1671 | TaskRedRef = |
| 1672 | cast<OMPTargetParallelDirective>(Val: D).getTaskReductionRefExpr(); |
| 1673 | break; |
| 1674 | case OMPD_target_parallel_for: |
| 1675 | TaskRedRef = |
| 1676 | cast<OMPTargetParallelForDirective>(Val: D).getTaskReductionRefExpr(); |
| 1677 | break; |
| 1678 | case OMPD_distribute_parallel_for: |
| 1679 | TaskRedRef = |
| 1680 | cast<OMPDistributeParallelForDirective>(Val: D).getTaskReductionRefExpr(); |
| 1681 | break; |
| 1682 | case OMPD_teams_distribute_parallel_for: |
| 1683 | TaskRedRef = cast<OMPTeamsDistributeParallelForDirective>(Val: D) |
| 1684 | .getTaskReductionRefExpr(); |
| 1685 | break; |
| 1686 | case OMPD_target_teams_distribute_parallel_for: |
| 1687 | TaskRedRef = cast<OMPTargetTeamsDistributeParallelForDirective>(Val: D) |
| 1688 | .getTaskReductionRefExpr(); |
| 1689 | break; |
| 1690 | case OMPD_simd: |
| 1691 | case OMPD_for_simd: |
| 1692 | case OMPD_section: |
| 1693 | case OMPD_single: |
| 1694 | case OMPD_master: |
| 1695 | case OMPD_critical: |
| 1696 | case OMPD_parallel_for_simd: |
| 1697 | case OMPD_task: |
| 1698 | case OMPD_taskyield: |
| 1699 | case OMPD_error: |
| 1700 | case OMPD_barrier: |
| 1701 | case OMPD_taskwait: |
| 1702 | case OMPD_taskgroup: |
| 1703 | case OMPD_flush: |
| 1704 | case OMPD_depobj: |
| 1705 | case OMPD_scan: |
| 1706 | case OMPD_ordered_standalone: |
| 1707 | case OMPD_ordered_blockassoc: |
| 1708 | case OMPD_atomic: |
| 1709 | case OMPD_teams: |
| 1710 | case OMPD_target: |
| 1711 | case OMPD_cancellation_point: |
| 1712 | case OMPD_cancel: |
| 1713 | case OMPD_target_data: |
| 1714 | case OMPD_target_enter_data: |
| 1715 | case OMPD_target_exit_data: |
| 1716 | case OMPD_taskloop: |
| 1717 | case OMPD_taskloop_simd: |
| 1718 | case OMPD_master_taskloop: |
| 1719 | case OMPD_master_taskloop_simd: |
| 1720 | case OMPD_parallel_master_taskloop: |
| 1721 | case OMPD_parallel_master_taskloop_simd: |
| 1722 | case OMPD_distribute: |
| 1723 | case OMPD_target_update: |
| 1724 | case OMPD_distribute_parallel_for_simd: |
| 1725 | case OMPD_distribute_simd: |
| 1726 | case OMPD_target_parallel_for_simd: |
| 1727 | case OMPD_target_simd: |
| 1728 | case OMPD_teams_distribute: |
| 1729 | case OMPD_teams_distribute_simd: |
| 1730 | case OMPD_teams_distribute_parallel_for_simd: |
| 1731 | case OMPD_target_teams: |
| 1732 | case OMPD_target_teams_distribute: |
| 1733 | case OMPD_target_teams_distribute_parallel_for_simd: |
| 1734 | case OMPD_target_teams_distribute_simd: |
| 1735 | case OMPD_declare_target: |
| 1736 | case OMPD_end_declare_target: |
| 1737 | case OMPD_threadprivate: |
| 1738 | case OMPD_allocate: |
| 1739 | case OMPD_declare_reduction: |
| 1740 | case OMPD_declare_mapper: |
| 1741 | case OMPD_declare_simd: |
| 1742 | case OMPD_requires: |
| 1743 | case OMPD_declare_variant: |
| 1744 | case OMPD_begin_declare_variant: |
| 1745 | case OMPD_end_declare_variant: |
| 1746 | case OMPD_unknown: |
| 1747 | default: |
| 1748 | llvm_unreachable("Unexpected directive with task reductions."); |
| 1749 | } |
| 1750 | |
| 1751 | const auto *VD = cast<VarDecl>(Val: cast<DeclRefExpr>(Val: TaskRedRef)->getDecl()); |
| 1752 | EmitVarDecl(D: *VD); |
| 1753 | EmitStoreOfScalar(Value: ReductionDesc, Addr: GetAddrOfLocalVar(VD), |
| 1754 | /*Volatile=*/false, Ty: TaskRedRef->getType()); |
| 1755 | } |
| 1756 | } |
| 1757 | |
| 1758 | void CodeGenFunction::EmitOMPReductionClauseFinal( |
| 1759 | const OMPExecutableDirective &D, const OpenMPDirectiveKind ReductionKind) { |
| 1760 | if (!HaveInsertPoint()) |
| 1761 | return; |
| 1762 | llvm::SmallVector<const Expr *, 8> Privates; |
| 1763 | llvm::SmallVector<const Expr *, 8> LHSExprs; |
| 1764 | llvm::SmallVector<const Expr *, 8> RHSExprs; |
| 1765 | llvm::SmallVector<const Expr *, 8> ReductionOps; |
| 1766 | llvm::SmallVector<bool, 8> IsPrivateVarReduction; |
| 1767 | bool HasAtLeastOneReduction = false; |
| 1768 | bool IsReductionWithTaskMod = false; |
| 1769 | for (const auto *C : D.getClausesOfKind<OMPReductionClause>()) { |
| 1770 | // Do not emit for inscan reductions. |
| 1771 | if (C->getModifier() == OMPC_REDUCTION_inscan) |
| 1772 | continue; |
| 1773 | HasAtLeastOneReduction = true; |
| 1774 | Privates.append(in_start: C->privates().begin(), in_end: C->privates().end()); |
| 1775 | LHSExprs.append(in_start: C->lhs_exprs().begin(), in_end: C->lhs_exprs().end()); |
| 1776 | RHSExprs.append(in_start: C->rhs_exprs().begin(), in_end: C->rhs_exprs().end()); |
| 1777 | IsPrivateVarReduction.append(in_start: C->private_var_reduction_flags().begin(), |
| 1778 | in_end: C->private_var_reduction_flags().end()); |
| 1779 | ReductionOps.append(in_start: C->reduction_ops().begin(), in_end: C->reduction_ops().end()); |
| 1780 | IsReductionWithTaskMod = |
| 1781 | IsReductionWithTaskMod || C->getModifier() == OMPC_REDUCTION_task; |
| 1782 | } |
| 1783 | if (HasAtLeastOneReduction) { |
| 1784 | OpenMPDirectiveKind EKind = getEffectiveDirectiveKind(S: D); |
| 1785 | if (IsReductionWithTaskMod) { |
| 1786 | CGM.getOpenMPRuntime().emitTaskReductionFini( |
| 1787 | CGF&: *this, Loc: D.getBeginLoc(), IsWorksharingReduction: isOpenMPWorksharingDirective(DKind: EKind)); |
| 1788 | } |
| 1789 | bool TeamsLoopCanBeParallel = false; |
| 1790 | if (auto *TTLD = dyn_cast<OMPTargetTeamsGenericLoopDirective>(Val: &D)) |
| 1791 | TeamsLoopCanBeParallel = TTLD->canBeParallelFor(); |
| 1792 | bool WithNowait = D.getSingleClause<OMPNowaitClause>() || |
| 1793 | isOpenMPParallelDirective(DKind: EKind) || |
| 1794 | TeamsLoopCanBeParallel || ReductionKind == OMPD_simd; |
| 1795 | bool SimpleReduction = ReductionKind == OMPD_simd; |
| 1796 | // Emit nowait reduction if nowait clause is present or directive is a |
| 1797 | // parallel directive (it always has implicit barrier). |
| 1798 | CGM.getOpenMPRuntime().emitReduction( |
| 1799 | CGF&: *this, Loc: D.getEndLoc(), Privates, LHSExprs, RHSExprs, ReductionOps, |
| 1800 | Options: {.WithNowait: WithNowait, .SimpleReduction: SimpleReduction, .IsPrivateVarReduction: IsPrivateVarReduction, .ReductionKind: ReductionKind}); |
| 1801 | } |
| 1802 | } |
| 1803 | |
| 1804 | static void emitPostUpdateForReductionClause( |
| 1805 | CodeGenFunction &CGF, const OMPExecutableDirective &D, |
| 1806 | const llvm::function_ref<llvm::Value *(CodeGenFunction &)> CondGen) { |
| 1807 | if (!CGF.HaveInsertPoint()) |
| 1808 | return; |
| 1809 | llvm::BasicBlock *DoneBB = nullptr; |
| 1810 | for (const auto *C : D.getClausesOfKind<OMPReductionClause>()) { |
| 1811 | if (const Expr *PostUpdate = C->getPostUpdateExpr()) { |
| 1812 | if (!DoneBB) { |
| 1813 | if (llvm::Value *Cond = CondGen(CGF)) { |
| 1814 | // If the first post-update expression is found, emit conditional |
| 1815 | // block if it was requested. |
| 1816 | llvm::BasicBlock *ThenBB = CGF.createBasicBlock(name: ".omp.reduction.pu"); |
| 1817 | DoneBB = CGF.createBasicBlock(name: ".omp.reduction.pu.done"); |
| 1818 | CGF.Builder.CreateCondBr(Cond, True: ThenBB, False: DoneBB); |
| 1819 | CGF.EmitBlock(BB: ThenBB); |
| 1820 | } |
| 1821 | } |
| 1822 | CGF.EmitIgnoredExpr(E: PostUpdate); |
| 1823 | } |
| 1824 | } |
| 1825 | if (DoneBB) |
| 1826 | CGF.EmitBlock(BB: DoneBB, /*IsFinished=*/true); |
| 1827 | } |
| 1828 | |
| 1829 | namespace { |
| 1830 | /// Codegen lambda for appending distribute lower and upper bounds to outlined |
| 1831 | /// parallel function. This is necessary for combined constructs such as |
| 1832 | /// 'distribute parallel for' |
| 1833 | typedef llvm::function_ref<void(CodeGenFunction &, |
| 1834 | const OMPExecutableDirective &, |
| 1835 | llvm::SmallVectorImpl<llvm::Value *> &)> |
| 1836 | CodeGenBoundParametersTy; |
| 1837 | } // anonymous namespace |
| 1838 | |
| 1839 | static void |
| 1840 | checkForLastprivateConditionalUpdate(CodeGenFunction &CGF, |
| 1841 | const OMPExecutableDirective &S) { |
| 1842 | if (CGF.getLangOpts().OpenMP < 50) |
| 1843 | return; |
| 1844 | llvm::DenseSet<CanonicalDeclPtr<const VarDecl>> PrivateDecls; |
| 1845 | for (const auto *C : S.getClausesOfKind<OMPReductionClause>()) { |
| 1846 | for (const Expr *Ref : C->varlist()) { |
| 1847 | if (!Ref->getType()->isScalarType()) |
| 1848 | continue; |
| 1849 | const auto *DRE = dyn_cast<DeclRefExpr>(Val: Ref->IgnoreParenImpCasts()); |
| 1850 | if (!DRE) |
| 1851 | continue; |
| 1852 | PrivateDecls.insert(V: cast<VarDecl>(Val: DRE->getDecl())); |
| 1853 | CGF.CGM.getOpenMPRuntime().checkAndEmitLastprivateConditional(CGF, LHS: Ref); |
| 1854 | } |
| 1855 | } |
| 1856 | for (const auto *C : S.getClausesOfKind<OMPLastprivateClause>()) { |
| 1857 | for (const Expr *Ref : C->varlist()) { |
| 1858 | if (!Ref->getType()->isScalarType()) |
| 1859 | continue; |
| 1860 | const auto *DRE = dyn_cast<DeclRefExpr>(Val: Ref->IgnoreParenImpCasts()); |
| 1861 | if (!DRE) |
| 1862 | continue; |
| 1863 | PrivateDecls.insert(V: cast<VarDecl>(Val: DRE->getDecl())); |
| 1864 | CGF.CGM.getOpenMPRuntime().checkAndEmitLastprivateConditional(CGF, LHS: Ref); |
| 1865 | } |
| 1866 | } |
| 1867 | for (const auto *C : S.getClausesOfKind<OMPLinearClause>()) { |
| 1868 | for (const Expr *Ref : C->varlist()) { |
| 1869 | if (!Ref->getType()->isScalarType()) |
| 1870 | continue; |
| 1871 | const auto *DRE = dyn_cast<DeclRefExpr>(Val: Ref->IgnoreParenImpCasts()); |
| 1872 | if (!DRE) |
| 1873 | continue; |
| 1874 | PrivateDecls.insert(V: cast<VarDecl>(Val: DRE->getDecl())); |
| 1875 | CGF.CGM.getOpenMPRuntime().checkAndEmitLastprivateConditional(CGF, LHS: Ref); |
| 1876 | } |
| 1877 | } |
| 1878 | // Privates should ne analyzed since they are not captured at all. |
| 1879 | // Task reductions may be skipped - tasks are ignored. |
| 1880 | // Firstprivates do not return value but may be passed by reference - no need |
| 1881 | // to check for updated lastprivate conditional. |
| 1882 | for (const auto *C : S.getClausesOfKind<OMPFirstprivateClause>()) { |
| 1883 | for (const Expr *Ref : C->varlist()) { |
| 1884 | if (!Ref->getType()->isScalarType()) |
| 1885 | continue; |
| 1886 | const auto *DRE = dyn_cast<DeclRefExpr>(Val: Ref->IgnoreParenImpCasts()); |
| 1887 | if (!DRE) |
| 1888 | continue; |
| 1889 | PrivateDecls.insert(V: cast<VarDecl>(Val: DRE->getDecl())); |
| 1890 | } |
| 1891 | } |
| 1892 | CGF.CGM.getOpenMPRuntime().checkAndEmitSharedLastprivateConditional( |
| 1893 | CGF, D: S, IgnoredDecls: PrivateDecls); |
| 1894 | } |
| 1895 | |
| 1896 | static void emitCommonOMPParallelDirective( |
| 1897 | CodeGenFunction &CGF, const OMPExecutableDirective &S, |
| 1898 | OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen, |
| 1899 | const CodeGenBoundParametersTy &CodeGenBoundParameters) { |
| 1900 | const CapturedStmt *CS = S.getCapturedStmt(RegionKind: OMPD_parallel); |
| 1901 | llvm::Value *NumThreads = nullptr; |
| 1902 | OpenMPNumThreadsClauseModifier Modifier = OMPC_NUMTHREADS_unknown; |
| 1903 | // OpenMP 6.0, 10.4: "If no severity clause is specified then the effect is as |
| 1904 | // if sev-level is fatal." |
| 1905 | OpenMPSeverityClauseKind Severity = OMPC_SEVERITY_fatal; |
| 1906 | clang::Expr *Message = nullptr; |
| 1907 | SourceLocation SeverityLoc = SourceLocation(); |
| 1908 | SourceLocation MessageLoc = SourceLocation(); |
| 1909 | |
| 1910 | llvm::Function *OutlinedFn = |
| 1911 | CGF.CGM.getOpenMPRuntime().emitParallelOutlinedFunction( |
| 1912 | CGF, D: S, ThreadIDVar: *CS->getCapturedDecl()->param_begin(), InnermostKind, |
| 1913 | CodeGen); |
| 1914 | |
| 1915 | if (const auto *NumThreadsClause = S.getSingleClause<OMPNumThreadsClause>()) { |
| 1916 | CodeGenFunction::RunCleanupsScope NumThreadsScope(CGF); |
| 1917 | NumThreads = CGF.EmitScalarExpr(E: NumThreadsClause->getNumThreads().front(), |
| 1918 | /*IgnoreResultAssign=*/true); |
| 1919 | Modifier = NumThreadsClause->getPrescriptivenessModifier(); |
| 1920 | if (const auto *MessageClause = S.getSingleClause<OMPMessageClause>()) { |
| 1921 | Message = MessageClause->getMessageString(); |
| 1922 | MessageLoc = MessageClause->getBeginLoc(); |
| 1923 | } |
| 1924 | if (const auto *SeverityClause = S.getSingleClause<OMPSeverityClause>()) { |
| 1925 | Severity = SeverityClause->getSeverityKind(); |
| 1926 | SeverityLoc = SeverityClause->getBeginLoc(); |
| 1927 | } |
| 1928 | CGF.CGM.getOpenMPRuntime().emitNumThreadsClause( |
| 1929 | CGF, NumThreads, Loc: NumThreadsClause->getBeginLoc(), Modifier, Severity, |
| 1930 | SeverityLoc, Message, MessageLoc); |
| 1931 | } |
| 1932 | if (const auto *ProcBindClause = S.getSingleClause<OMPProcBindClause>()) { |
| 1933 | CodeGenFunction::RunCleanupsScope ProcBindScope(CGF); |
| 1934 | CGF.CGM.getOpenMPRuntime().emitProcBindClause( |
| 1935 | CGF, ProcBind: ProcBindClause->getProcBindKind(), Loc: ProcBindClause->getBeginLoc()); |
| 1936 | } |
| 1937 | const Expr *IfCond = nullptr; |
| 1938 | for (const auto *C : S.getClausesOfKind<OMPIfClause>()) { |
| 1939 | if (C->getNameModifier() == OMPD_unknown || |
| 1940 | C->getNameModifier() == OMPD_parallel) { |
| 1941 | IfCond = C->getCondition(); |
| 1942 | break; |
| 1943 | } |
| 1944 | } |
| 1945 | |
| 1946 | OMPParallelScope Scope(CGF, S); |
| 1947 | llvm::SmallVector<llvm::Value *, 16> CapturedVars; |
| 1948 | // Combining 'distribute' with 'for' requires sharing each 'distribute' chunk |
| 1949 | // lower and upper bounds with the pragma 'for' chunking mechanism. |
| 1950 | // The following lambda takes care of appending the lower and upper bound |
| 1951 | // parameters when necessary |
| 1952 | CodeGenBoundParameters(CGF, S, CapturedVars); |
| 1953 | CGF.GenerateOpenMPCapturedVars(S: *CS, CapturedVars); |
| 1954 | CGF.CGM.getOpenMPRuntime().emitParallelCall(CGF, Loc: S.getBeginLoc(), OutlinedFn, |
| 1955 | CapturedVars, IfCond, NumThreads, |
| 1956 | NumThreadsModifier: Modifier, Severity, Message); |
| 1957 | } |
| 1958 | |
| 1959 | static bool isAllocatableDecl(const VarDecl *VD) { |
| 1960 | const VarDecl *CVD = VD->getCanonicalDecl(); |
| 1961 | if (!CVD->hasAttr<OMPAllocateDeclAttr>()) |
| 1962 | return false; |
| 1963 | const auto *AA = CVD->getAttr<OMPAllocateDeclAttr>(); |
| 1964 | // Use the default allocation. |
| 1965 | return !((AA->getAllocatorType() == OMPAllocateDeclAttr::OMPDefaultMemAlloc || |
| 1966 | AA->getAllocatorType() == OMPAllocateDeclAttr::OMPNullMemAlloc) && |
| 1967 | !AA->getAllocator()); |
| 1968 | } |
| 1969 | |
| 1970 | static void emitEmptyBoundParameters(CodeGenFunction &, |
| 1971 | const OMPExecutableDirective &, |
| 1972 | llvm::SmallVectorImpl<llvm::Value *> &) {} |
| 1973 | |
| 1974 | static void emitOMPCopyinClause(CodeGenFunction &CGF, |
| 1975 | const OMPExecutableDirective &S) { |
| 1976 | bool Copyins = CGF.EmitOMPCopyinClause(D: S); |
| 1977 | if (Copyins) { |
| 1978 | // Emit implicit barrier to synchronize threads and avoid data races on |
| 1979 | // propagation master's thread values of threadprivate variables to local |
| 1980 | // instances of that variables of all other implicit threads. |
| 1981 | CGF.CGM.getOpenMPRuntime().emitBarrierCall( |
| 1982 | CGF, Loc: S.getBeginLoc(), Kind: OMPD_unknown, /*EmitChecks=*/false, |
| 1983 | /*ForceSimpleCall=*/true); |
| 1984 | } |
| 1985 | } |
| 1986 | |
| 1987 | Address CodeGenFunction::OMPBuilderCBHelpers::getAddressOfLocalVariable( |
| 1988 | CodeGenFunction &CGF, const VarDecl *VD) { |
| 1989 | CodeGenModule &CGM = CGF.CGM; |
| 1990 | auto &OMPBuilder = CGM.getOpenMPRuntime().getOMPBuilder(); |
| 1991 | |
| 1992 | if (!VD) |
| 1993 | return Address::invalid(); |
| 1994 | const VarDecl *CVD = VD->getCanonicalDecl(); |
| 1995 | if (!isAllocatableDecl(VD: CVD)) |
| 1996 | return Address::invalid(); |
| 1997 | llvm::Value *Size; |
| 1998 | CharUnits Align = CGM.getContext().getDeclAlign(D: CVD); |
| 1999 | if (CVD->getType()->isVariablyModifiedType()) { |
| 2000 | Size = CGF.getTypeSize(Ty: CVD->getType()); |
| 2001 | // Align the size: ((size + align - 1) / align) * align |
| 2002 | Size = CGF.Builder.CreateNUWAdd( |
| 2003 | LHS: Size, RHS: CGM.getSize(numChars: Align - CharUnits::fromQuantity(Quantity: 1))); |
| 2004 | Size = CGF.Builder.CreateUDiv(LHS: Size, RHS: CGM.getSize(numChars: Align)); |
| 2005 | Size = CGF.Builder.CreateNUWMul(LHS: Size, RHS: CGM.getSize(numChars: Align)); |
| 2006 | } else { |
| 2007 | CharUnits Sz = CGM.getContext().getTypeSizeInChars(T: CVD->getType()); |
| 2008 | Size = CGM.getSize(numChars: Sz.alignTo(Align)); |
| 2009 | } |
| 2010 | |
| 2011 | const auto *AA = CVD->getAttr<OMPAllocateDeclAttr>(); |
| 2012 | assert(AA->getAllocator() && |
| 2013 | "Expected allocator expression for non-default allocator."); |
| 2014 | llvm::Value *Allocator = CGF.EmitScalarExpr(E: AA->getAllocator()); |
| 2015 | // According to the standard, the original allocator type is a enum (integer). |
| 2016 | // Convert to pointer type, if required. |
| 2017 | if (Allocator->getType()->isIntegerTy()) |
| 2018 | Allocator = CGF.Builder.CreateIntToPtr(V: Allocator, DestTy: CGM.VoidPtrTy); |
| 2019 | else if (Allocator->getType()->isPointerTy()) |
| 2020 | Allocator = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(V: Allocator, |
| 2021 | DestTy: CGM.VoidPtrTy); |
| 2022 | |
| 2023 | llvm::Value *Addr = OMPBuilder.createOMPAlloc( |
| 2024 | Loc: CGF.Builder, Size, Allocator, |
| 2025 | Name: getNameWithSeparators(Parts: {CVD->getName(), ".void.addr"}, FirstSeparator: ".", Separator: ".")); |
| 2026 | llvm::CallInst *FreeCI = |
| 2027 | OMPBuilder.createOMPFree(Loc: CGF.Builder, Addr, Allocator); |
| 2028 | |
| 2029 | CGF.EHStack.pushCleanup<OMPAllocateCleanupTy>(Kind: NormalAndEHCleanup, A: FreeCI); |
| 2030 | Addr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( |
| 2031 | V: Addr, |
| 2032 | DestTy: CGF.ConvertTypeForMem(T: CGM.getContext().getPointerType(T: CVD->getType())), |
| 2033 | Name: getNameWithSeparators(Parts: {CVD->getName(), ".addr"}, FirstSeparator: ".", Separator: ".")); |
| 2034 | return Address(Addr, CGF.ConvertTypeForMem(T: CVD->getType()), Align); |
| 2035 | } |
| 2036 | |
| 2037 | Address CodeGenFunction::OMPBuilderCBHelpers::getAddrOfThreadPrivate( |
| 2038 | CodeGenFunction &CGF, const VarDecl *VD, Address VDAddr, |
| 2039 | SourceLocation Loc) { |
| 2040 | CodeGenModule &CGM = CGF.CGM; |
| 2041 | if (CGM.getLangOpts().OpenMPUseTLS && |
| 2042 | CGM.getContext().getTargetInfo().isTLSSupported()) |
| 2043 | return VDAddr; |
| 2044 | |
| 2045 | llvm::OpenMPIRBuilder &OMPBuilder = CGM.getOpenMPRuntime().getOMPBuilder(); |
| 2046 | |
| 2047 | llvm::Type *VarTy = VDAddr.getElementType(); |
| 2048 | llvm::Value *Data = |
| 2049 | CGF.Builder.CreatePointerCast(V: VDAddr.emitRawPointer(CGF), DestTy: CGM.Int8PtrTy); |
| 2050 | llvm::ConstantInt *Size = CGM.getSize(numChars: CGM.GetTargetTypeStoreSize(Ty: VarTy)); |
| 2051 | std::string Suffix = getNameWithSeparators(Parts: {"cache", ""}); |
| 2052 | llvm::Twine CacheName = Twine(CGM.getMangledName(GD: VD)).concat(Suffix); |
| 2053 | |
| 2054 | llvm::CallInst *ThreadPrivateCacheCall = |
| 2055 | OMPBuilder.createCachedThreadPrivate(Loc: CGF.Builder, Pointer: Data, Size, Name: CacheName); |
| 2056 | |
| 2057 | return Address(ThreadPrivateCacheCall, CGM.Int8Ty, VDAddr.getAlignment()); |
| 2058 | } |
| 2059 | |
| 2060 | std::string CodeGenFunction::OMPBuilderCBHelpers::getNameWithSeparators( |
| 2061 | ArrayRef<StringRef> Parts, StringRef FirstSeparator, StringRef Separator) { |
| 2062 | SmallString<128> Buffer; |
| 2063 | llvm::raw_svector_ostream OS(Buffer); |
| 2064 | StringRef Sep = FirstSeparator; |
| 2065 | for (StringRef Part : Parts) { |
| 2066 | OS << Sep << Part; |
| 2067 | Sep = Separator; |
| 2068 | } |
| 2069 | return OS.str().str(); |
| 2070 | } |
| 2071 | |
| 2072 | void CodeGenFunction::OMPBuilderCBHelpers::EmitOMPInlinedRegionBody( |
| 2073 | CodeGenFunction &CGF, const Stmt *RegionBodyStmt, InsertPointTy AllocaIP, |
| 2074 | InsertPointTy CodeGenIP, Twine RegionName) { |
| 2075 | CGBuilderTy &Builder = CGF.Builder; |
| 2076 | Builder.restoreIP(IP: CodeGenIP); |
| 2077 | llvm::BasicBlock *FiniBB = splitBBWithSuffix(Builder, /*CreateBranch=*/false, |
| 2078 | Suffix: "."+ RegionName + ".after"); |
| 2079 | |
| 2080 | { |
| 2081 | OMPBuilderCBHelpers::InlinedRegionBodyRAII IRB(CGF, AllocaIP, *FiniBB); |
| 2082 | CGF.EmitStmt(S: RegionBodyStmt); |
| 2083 | } |
| 2084 | |
| 2085 | if (Builder.saveIP().isSet()) |
| 2086 | Builder.CreateBr(Dest: FiniBB); |
| 2087 | } |
| 2088 | |
| 2089 | void CodeGenFunction::OMPBuilderCBHelpers::EmitOMPOutlinedRegionBody( |
| 2090 | CodeGenFunction &CGF, const Stmt *RegionBodyStmt, InsertPointTy AllocaIP, |
| 2091 | InsertPointTy CodeGenIP, Twine RegionName) { |
| 2092 | CGBuilderTy &Builder = CGF.Builder; |
| 2093 | Builder.restoreIP(IP: CodeGenIP); |
| 2094 | llvm::BasicBlock *FiniBB = splitBBWithSuffix(Builder, /*CreateBranch=*/false, |
| 2095 | Suffix: "."+ RegionName + ".after"); |
| 2096 | |
| 2097 | { |
| 2098 | OMPBuilderCBHelpers::OutlinedRegionBodyRAII IRB(CGF, AllocaIP, *FiniBB); |
| 2099 | CGF.EmitStmt(S: RegionBodyStmt); |
| 2100 | } |
| 2101 | |
| 2102 | if (Builder.saveIP().isSet()) |
| 2103 | Builder.CreateBr(Dest: FiniBB); |
| 2104 | } |
| 2105 | |
| 2106 | void CodeGenFunction::EmitOMPParallelDirective(const OMPParallelDirective &S) { |
| 2107 | if (CGM.getLangOpts().OpenMPIRBuilder) { |
| 2108 | llvm::OpenMPIRBuilder &OMPBuilder = CGM.getOpenMPRuntime().getOMPBuilder(); |
| 2109 | // Check if we have any if clause associated with the directive. |
| 2110 | llvm::Value *IfCond = nullptr; |
| 2111 | if (const auto *C = S.getSingleClause<OMPIfClause>()) |
| 2112 | IfCond = EmitScalarExpr(E: C->getCondition(), |
| 2113 | /*IgnoreResultAssign=*/true); |
| 2114 | |
| 2115 | llvm::Value *NumThreads = nullptr; |
| 2116 | if (const auto *NumThreadsClause = S.getSingleClause<OMPNumThreadsClause>()) |
| 2117 | NumThreads = EmitScalarExpr(E: NumThreadsClause->getNumThreads().front(), |
| 2118 | /*IgnoreResultAssign=*/true); |
| 2119 | |
| 2120 | ProcBindKind ProcBind = OMP_PROC_BIND_default; |
| 2121 | if (const auto *ProcBindClause = S.getSingleClause<OMPProcBindClause>()) |
| 2122 | ProcBind = ProcBindClause->getProcBindKind(); |
| 2123 | |
| 2124 | using InsertPointTy = llvm::OpenMPIRBuilder::InsertPointTy; |
| 2125 | |
| 2126 | // The cleanup callback that finalizes all variables at the given location, |
| 2127 | // thus calls destructors etc. |
| 2128 | auto FiniCB = [this](InsertPointTy IP) { |
| 2129 | OMPBuilderCBHelpers::FinalizeOMPRegion(CGF&: *this, IP); |
| 2130 | return llvm::Error::success(); |
| 2131 | }; |
| 2132 | |
| 2133 | // Privatization callback that performs appropriate action for |
| 2134 | // shared/private/firstprivate/lastprivate/copyin/... variables. |
| 2135 | // |
| 2136 | // TODO: This defaults to shared right now. |
| 2137 | auto PrivCB = [](InsertPointTy AllocaIP, InsertPointTy CodeGenIP, |
| 2138 | llvm::Value &, llvm::Value &Val, llvm::Value *&ReplVal) { |
| 2139 | // The next line is appropriate only for variables (Val) with the |
| 2140 | // data-sharing attribute "shared". |
| 2141 | ReplVal = &Val; |
| 2142 | |
| 2143 | return CodeGenIP; |
| 2144 | }; |
| 2145 | |
| 2146 | const CapturedStmt *CS = S.getCapturedStmt(RegionKind: OMPD_parallel); |
| 2147 | const Stmt *ParallelRegionBodyStmt = CS->getCapturedStmt(); |
| 2148 | |
| 2149 | auto BodyGenCB = [&, this](InsertPointTy AllocIP, InsertPointTy CodeGenIP, |
| 2150 | ArrayRef<llvm::BasicBlock *> DeallocBlocks) { |
| 2151 | OMPBuilderCBHelpers::EmitOMPOutlinedRegionBody( |
| 2152 | CGF&: *this, RegionBodyStmt: ParallelRegionBodyStmt, AllocaIP: AllocIP, CodeGenIP, RegionName: "parallel"); |
| 2153 | return llvm::Error::success(); |
| 2154 | }; |
| 2155 | |
| 2156 | CGCapturedStmtInfo CGSI(*CS, CR_OpenMP); |
| 2157 | CodeGenFunction::CGCapturedStmtRAII CapInfoRAII(*this, &CGSI); |
| 2158 | llvm::OpenMPIRBuilder::InsertPointTy AllocaIP( |
| 2159 | AllocaInsertPt->getParent(), AllocaInsertPt->getIterator()); |
| 2160 | llvm::OpenMPIRBuilder::InsertPointTy AfterIP = |
| 2161 | cantFail(ValOrErr: OMPBuilder.createParallel( |
| 2162 | Loc: Builder, AllocaIP, /*DeallocBlocks=*/{}, BodyGenCB, PrivCB, FiniCB, |
| 2163 | IfCondition: IfCond, NumThreads, ProcBind, IsCancellable: S.hasCancel())); |
| 2164 | Builder.restoreIP(IP: AfterIP); |
| 2165 | return; |
| 2166 | } |
| 2167 | |
| 2168 | // Emit parallel region as a standalone region. |
| 2169 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 2170 | Action.Enter(CGF); |
| 2171 | OMPPrivateScope PrivateScope(CGF); |
| 2172 | emitOMPCopyinClause(CGF, S); |
| 2173 | (void)CGF.EmitOMPFirstprivateClause(D: S, PrivateScope); |
| 2174 | CGF.EmitOMPPrivateClause(D: S, PrivateScope); |
| 2175 | CGF.EmitOMPReductionClauseInit(D: S, PrivateScope); |
| 2176 | (void)PrivateScope.Privatize(); |
| 2177 | CGF.EmitStmt(S: S.getCapturedStmt(RegionKind: OMPD_parallel)->getCapturedStmt()); |
| 2178 | CGF.EmitOMPReductionClauseFinal(D: S, /*ReductionKind=*/OMPD_parallel); |
| 2179 | }; |
| 2180 | { |
| 2181 | auto LPCRegion = |
| 2182 | CGOpenMPRuntime::LastprivateConditionalRAII::disable(CGF&: *this, S); |
| 2183 | emitCommonOMPParallelDirective(CGF&: *this, S, InnermostKind: OMPD_parallel, CodeGen, |
| 2184 | CodeGenBoundParameters: emitEmptyBoundParameters); |
| 2185 | emitPostUpdateForReductionClause(CGF&: *this, D: S, |
| 2186 | CondGen: [](CodeGenFunction &) { return nullptr; }); |
| 2187 | } |
| 2188 | // Check for outer lastprivate conditional update. |
| 2189 | checkForLastprivateConditionalUpdate(CGF&: *this, S); |
| 2190 | } |
| 2191 | |
| 2192 | void CodeGenFunction::EmitOMPMetaDirective(const OMPMetaDirective &S) { |
| 2193 | EmitStmt(S: S.getIfStmt()); |
| 2194 | } |
| 2195 | |
| 2196 | namespace { |
| 2197 | /// RAII to handle scopes for loop transformation directives. |
| 2198 | class OMPTransformDirectiveScopeRAII { |
| 2199 | OMPLoopScope *Scope = nullptr; |
| 2200 | CodeGenFunction::CGCapturedStmtInfo *CGSI = nullptr; |
| 2201 | CodeGenFunction::CGCapturedStmtRAII *CapInfoRAII = nullptr; |
| 2202 | |
| 2203 | OMPTransformDirectiveScopeRAII(const OMPTransformDirectiveScopeRAII &) = |
| 2204 | delete; |
| 2205 | OMPTransformDirectiveScopeRAII & |
| 2206 | operator=(const OMPTransformDirectiveScopeRAII &) = delete; |
| 2207 | |
| 2208 | public: |
| 2209 | OMPTransformDirectiveScopeRAII(CodeGenFunction &CGF, const Stmt *S) { |
| 2210 | if (const auto *Dir = dyn_cast<OMPLoopBasedDirective>(Val: S)) { |
| 2211 | Scope = new OMPLoopScope(CGF, *Dir); |
| 2212 | CGSI = new CodeGenFunction::CGCapturedStmtInfo(CR_OpenMP); |
| 2213 | CapInfoRAII = new CodeGenFunction::CGCapturedStmtRAII(CGF, CGSI); |
| 2214 | } else if (const auto *Dir = |
| 2215 | dyn_cast<OMPCanonicalLoopSequenceTransformationDirective>( |
| 2216 | Val: S)) { |
| 2217 | // For simplicity we reuse the loop scope similarly to what we do with |
| 2218 | // OMPCanonicalLoopNestTransformationDirective do by being a subclass |
| 2219 | // of OMPLoopBasedDirective. |
| 2220 | Scope = new OMPLoopScope(CGF, *Dir); |
| 2221 | CGSI = new CodeGenFunction::CGCapturedStmtInfo(CR_OpenMP); |
| 2222 | CapInfoRAII = new CodeGenFunction::CGCapturedStmtRAII(CGF, CGSI); |
| 2223 | } |
| 2224 | } |
| 2225 | ~OMPTransformDirectiveScopeRAII() { |
| 2226 | if (!Scope) |
| 2227 | return; |
| 2228 | delete CapInfoRAII; |
| 2229 | delete CGSI; |
| 2230 | delete Scope; |
| 2231 | } |
| 2232 | }; |
| 2233 | } // namespace |
| 2234 | |
| 2235 | static void emitBody(CodeGenFunction &CGF, const Stmt *S, const Stmt *NextLoop, |
| 2236 | int MaxLevel, int Level = 0) { |
| 2237 | assert(Level < MaxLevel && "Too deep lookup during loop body codegen."); |
| 2238 | const Stmt *SimplifiedS = S->IgnoreContainers(); |
| 2239 | if (const auto *CS = dyn_cast<CompoundStmt>(Val: SimplifiedS)) { |
| 2240 | PrettyStackTraceLoc CrashInfo( |
| 2241 | CGF.getContext().getSourceManager(), CS->getLBracLoc(), |
| 2242 | "LLVM IR generation of compound statement ('{}')"); |
| 2243 | |
| 2244 | // Keep track of the current cleanup stack depth, including debug scopes. |
| 2245 | CodeGenFunction::LexicalScope Scope(CGF, S->getSourceRange()); |
| 2246 | for (const Stmt *CurStmt : CS->body()) |
| 2247 | emitBody(CGF, S: CurStmt, NextLoop, MaxLevel, Level); |
| 2248 | return; |
| 2249 | } |
| 2250 | if (SimplifiedS == NextLoop) { |
| 2251 | if (auto *Dir = dyn_cast<OMPLoopTransformationDirective>(Val: SimplifiedS)) |
| 2252 | SimplifiedS = Dir->getTransformedStmt(); |
| 2253 | if (const auto *CanonLoop = dyn_cast<OMPCanonicalLoop>(Val: SimplifiedS)) |
| 2254 | SimplifiedS = CanonLoop->getLoopStmt(); |
| 2255 | if (const auto *For = dyn_cast<ForStmt>(Val: SimplifiedS)) { |
| 2256 | S = For->getBody(); |
| 2257 | } else { |
| 2258 | assert(isa<CXXForRangeStmt>(SimplifiedS) && |
| 2259 | "Expected canonical for loop or range-based for loop."); |
| 2260 | const auto *CXXFor = cast<CXXForRangeStmt>(Val: SimplifiedS); |
| 2261 | CGF.EmitStmt(S: CXXFor->getLoopVarStmt()); |
| 2262 | S = CXXFor->getBody(); |
| 2263 | } |
| 2264 | if (Level + 1 < MaxLevel) { |
| 2265 | NextLoop = OMPLoopDirective::tryToFindNextInnerLoop( |
| 2266 | CurStmt: S, /*TryImperfectlyNestedLoops=*/true); |
| 2267 | emitBody(CGF, S, NextLoop, MaxLevel, Level: Level + 1); |
| 2268 | return; |
| 2269 | } |
| 2270 | } |
| 2271 | CGF.EmitStmt(S); |
| 2272 | } |
| 2273 | |
| 2274 | void CodeGenFunction::EmitOMPLoopBody(const OMPLoopDirective &D, |
| 2275 | JumpDest LoopExit) { |
| 2276 | RunCleanupsScope BodyScope(*this); |
| 2277 | // Update counters values on current iteration. |
| 2278 | for (const Expr *UE : D.updates()) |
| 2279 | EmitIgnoredExpr(E: UE); |
| 2280 | // Update the linear variables. |
| 2281 | // In distribute directives only loop counters may be marked as linear, no |
| 2282 | // need to generate the code for them. |
| 2283 | OpenMPDirectiveKind EKind = getEffectiveDirectiveKind(S: D); |
| 2284 | if (!isOpenMPDistributeDirective(DKind: EKind)) { |
| 2285 | for (const auto *C : D.getClausesOfKind<OMPLinearClause>()) { |
| 2286 | for (const Expr *UE : C->updates()) |
| 2287 | EmitIgnoredExpr(E: UE); |
| 2288 | } |
| 2289 | } |
| 2290 | |
| 2291 | // On a continue in the body, jump to the end. |
| 2292 | JumpDest Continue = getJumpDestInCurrentScope(Name: "omp.body.continue"); |
| 2293 | BreakContinueStack.push_back(Elt: BreakContinue(D, LoopExit, Continue)); |
| 2294 | for (const Expr *E : D.finals_conditions()) { |
| 2295 | if (!E) |
| 2296 | continue; |
| 2297 | // Check that loop counter in non-rectangular nest fits into the iteration |
| 2298 | // space. |
| 2299 | llvm::BasicBlock *NextBB = createBasicBlock(name: "omp.body.next"); |
| 2300 | EmitBranchOnBoolExpr(Cond: E, TrueBlock: NextBB, FalseBlock: Continue.getBlock(), |
| 2301 | TrueCount: getProfileCount(S: D.getBody())); |
| 2302 | EmitBlock(BB: NextBB); |
| 2303 | } |
| 2304 | |
| 2305 | OMPPrivateScope InscanScope(*this); |
| 2306 | EmitOMPReductionClauseInit(D, PrivateScope&: InscanScope, /*ForInscan=*/true); |
| 2307 | bool IsInscanRegion = InscanScope.Privatize(); |
| 2308 | if (IsInscanRegion) { |
| 2309 | // Need to remember the block before and after scan directive |
| 2310 | // to dispatch them correctly depending on the clause used in |
| 2311 | // this directive, inclusive or exclusive. For inclusive scan the natural |
| 2312 | // order of the blocks is used, for exclusive clause the blocks must be |
| 2313 | // executed in reverse order. |
| 2314 | OMPBeforeScanBlock = createBasicBlock(name: "omp.before.scan.bb"); |
| 2315 | OMPAfterScanBlock = createBasicBlock(name: "omp.after.scan.bb"); |
| 2316 | // No need to allocate inscan exit block, in simd mode it is selected in the |
| 2317 | // codegen for the scan directive. |
| 2318 | if (EKind != OMPD_simd && !getLangOpts().OpenMPSimd) |
| 2319 | OMPScanExitBlock = createBasicBlock(name: "omp.exit.inscan.bb"); |
| 2320 | OMPScanDispatch = createBasicBlock(name: "omp.inscan.dispatch"); |
| 2321 | EmitBranch(Block: OMPScanDispatch); |
| 2322 | EmitBlock(BB: OMPBeforeScanBlock); |
| 2323 | } |
| 2324 | |
| 2325 | // Emit loop variables for C++ range loops. |
| 2326 | const Stmt *Body = |
| 2327 | D.getInnermostCapturedStmt()->getCapturedStmt()->IgnoreContainers(); |
| 2328 | // Emit loop body. |
| 2329 | emitBody(CGF&: *this, S: Body, |
| 2330 | NextLoop: OMPLoopBasedDirective::tryToFindNextInnerLoop( |
| 2331 | CurStmt: Body, /*TryImperfectlyNestedLoops=*/true), |
| 2332 | MaxLevel: D.getLoopsNumber()); |
| 2333 | |
| 2334 | // Jump to the dispatcher at the end of the loop body. |
| 2335 | if (IsInscanRegion) |
| 2336 | EmitBranch(Block: OMPScanExitBlock); |
| 2337 | |
| 2338 | // The end (updates/cleanups). |
| 2339 | EmitBlock(BB: Continue.getBlock()); |
| 2340 | BreakContinueStack.pop_back(); |
| 2341 | } |
| 2342 | |
| 2343 | using EmittedClosureTy = std::pair<llvm::Function *, llvm::Value *>; |
| 2344 | |
| 2345 | /// Emit a captured statement and return the function as well as its captured |
| 2346 | /// closure context. |
| 2347 | static EmittedClosureTy emitCapturedStmtFunc(CodeGenFunction &ParentCGF, |
| 2348 | const CapturedStmt *S) { |
| 2349 | LValue CapStruct = ParentCGF.InitCapturedStruct(S: *S); |
| 2350 | CodeGenFunction CGF(ParentCGF.CGM, /*suppressNewContext=*/true); |
| 2351 | std::unique_ptr<CodeGenFunction::CGCapturedStmtInfo> CSI = |
| 2352 | std::make_unique<CodeGenFunction::CGCapturedStmtInfo>(args: *S); |
| 2353 | CodeGenFunction::CGCapturedStmtRAII CapInfoRAII(CGF, CSI.get()); |
| 2354 | llvm::Function *F = CGF.GenerateCapturedStmtFunction(S: *S); |
| 2355 | |
| 2356 | return {F, CapStruct.getPointer(CGF&: ParentCGF)}; |
| 2357 | } |
| 2358 | |
| 2359 | /// Emit a call to a previously captured closure. |
| 2360 | static llvm::CallInst * |
| 2361 | emitCapturedStmtCall(CodeGenFunction &ParentCGF, EmittedClosureTy Cap, |
| 2362 | llvm::ArrayRef<llvm::Value *> Args) { |
| 2363 | // Append the closure context to the argument. |
| 2364 | SmallVector<llvm::Value *> EffectiveArgs; |
| 2365 | EffectiveArgs.reserve(N: Args.size() + 1); |
| 2366 | llvm::append_range(C&: EffectiveArgs, R&: Args); |
| 2367 | EffectiveArgs.push_back(Elt: Cap.second); |
| 2368 | |
| 2369 | return ParentCGF.Builder.CreateCall(Callee: Cap.first, Args: EffectiveArgs); |
| 2370 | } |
| 2371 | |
| 2372 | llvm::CanonicalLoopInfo * |
| 2373 | CodeGenFunction::EmitOMPCollapsedCanonicalLoopNest(const Stmt *S, int Depth) { |
| 2374 | assert(Depth == 1 && "Nested loops with OpenMPIRBuilder not yet implemented"); |
| 2375 | |
| 2376 | // The caller is processing the loop-associated directive processing the \p |
| 2377 | // Depth loops nested in \p S. Put the previous pending loop-associated |
| 2378 | // directive to the stack. If the current loop-associated directive is a loop |
| 2379 | // transformation directive, it will push its generated loops onto the stack |
| 2380 | // such that together with the loops left here they form the combined loop |
| 2381 | // nest for the parent loop-associated directive. |
| 2382 | int ParentExpectedOMPLoopDepth = ExpectedOMPLoopDepth; |
| 2383 | ExpectedOMPLoopDepth = Depth; |
| 2384 | |
| 2385 | EmitStmt(S); |
| 2386 | assert(OMPLoopNestStack.size() >= (size_t)Depth && "Found too few loops"); |
| 2387 | |
| 2388 | // The last added loop is the outermost one. |
| 2389 | llvm::CanonicalLoopInfo *Result = OMPLoopNestStack.back(); |
| 2390 | |
| 2391 | // Pop the \p Depth loops requested by the call from that stack and restore |
| 2392 | // the previous context. |
| 2393 | OMPLoopNestStack.pop_back_n(NumItems: Depth); |
| 2394 | ExpectedOMPLoopDepth = ParentExpectedOMPLoopDepth; |
| 2395 | |
| 2396 | return Result; |
| 2397 | } |
| 2398 | |
| 2399 | void CodeGenFunction::EmitOMPCanonicalLoop(const OMPCanonicalLoop *S) { |
| 2400 | const Stmt *SyntacticalLoop = S->getLoopStmt(); |
| 2401 | if (!getLangOpts().OpenMPIRBuilder) { |
| 2402 | // Ignore if OpenMPIRBuilder is not enabled. |
| 2403 | EmitStmt(S: SyntacticalLoop); |
| 2404 | return; |
| 2405 | } |
| 2406 | |
| 2407 | LexicalScope ForScope(*this, S->getSourceRange()); |
| 2408 | |
| 2409 | // Emit init statements. The Distance/LoopVar funcs may reference variable |
| 2410 | // declarations they contain. |
| 2411 | const Stmt *BodyStmt; |
| 2412 | if (const auto *For = dyn_cast<ForStmt>(Val: SyntacticalLoop)) { |
| 2413 | if (const Stmt *InitStmt = For->getInit()) |
| 2414 | EmitStmt(S: InitStmt); |
| 2415 | BodyStmt = For->getBody(); |
| 2416 | } else if (const auto *RangeFor = |
| 2417 | dyn_cast<CXXForRangeStmt>(Val: SyntacticalLoop)) { |
| 2418 | if (const DeclStmt *RangeStmt = RangeFor->getRangeStmt()) |
| 2419 | EmitStmt(S: RangeStmt); |
| 2420 | if (const DeclStmt *BeginStmt = RangeFor->getBeginStmt()) |
| 2421 | EmitStmt(S: BeginStmt); |
| 2422 | if (const DeclStmt *EndStmt = RangeFor->getEndStmt()) |
| 2423 | EmitStmt(S: EndStmt); |
| 2424 | if (const DeclStmt *LoopVarStmt = RangeFor->getLoopVarStmt()) |
| 2425 | EmitStmt(S: LoopVarStmt); |
| 2426 | BodyStmt = RangeFor->getBody(); |
| 2427 | } else |
| 2428 | llvm_unreachable("Expected for-stmt or range-based for-stmt"); |
| 2429 | |
| 2430 | // Emit closure for later use. By-value captures will be captured here. |
| 2431 | const CapturedStmt *DistanceFunc = S->getDistanceFunc(); |
| 2432 | EmittedClosureTy DistanceClosure = emitCapturedStmtFunc(ParentCGF&: *this, S: DistanceFunc); |
| 2433 | const CapturedStmt *LoopVarFunc = S->getLoopVarFunc(); |
| 2434 | EmittedClosureTy LoopVarClosure = emitCapturedStmtFunc(ParentCGF&: *this, S: LoopVarFunc); |
| 2435 | |
| 2436 | // Call the distance function to get the number of iterations of the loop to |
| 2437 | // come. |
| 2438 | QualType LogicalTy = DistanceFunc->getCapturedDecl() |
| 2439 | ->getParam(i: 0) |
| 2440 | ->getType() |
| 2441 | .getNonReferenceType(); |
| 2442 | RawAddress CountAddr = CreateMemTemp(T: LogicalTy, Name: ".count.addr"); |
| 2443 | emitCapturedStmtCall(ParentCGF&: *this, Cap: DistanceClosure, Args: {CountAddr.getPointer()}); |
| 2444 | llvm::Value *DistVal = Builder.CreateLoad(Addr: CountAddr, Name: ".count"); |
| 2445 | |
| 2446 | // Emit the loop structure. |
| 2447 | llvm::OpenMPIRBuilder &OMPBuilder = CGM.getOpenMPRuntime().getOMPBuilder(); |
| 2448 | auto BodyGen = [&, this](llvm::OpenMPIRBuilder::InsertPointTy CodeGenIP, |
| 2449 | llvm::Value *IndVar) { |
| 2450 | Builder.restoreIP(IP: CodeGenIP); |
| 2451 | |
| 2452 | // Emit the loop body: Convert the logical iteration number to the loop |
| 2453 | // variable and emit the body. |
| 2454 | const DeclRefExpr *LoopVarRef = S->getLoopVarRef(); |
| 2455 | LValue LCVal = EmitLValue(E: LoopVarRef); |
| 2456 | Address LoopVarAddress = LCVal.getAddress(); |
| 2457 | emitCapturedStmtCall(ParentCGF&: *this, Cap: LoopVarClosure, |
| 2458 | Args: {LoopVarAddress.emitRawPointer(CGF&: *this), IndVar}); |
| 2459 | |
| 2460 | RunCleanupsScope BodyScope(*this); |
| 2461 | EmitStmt(S: BodyStmt); |
| 2462 | return llvm::Error::success(); |
| 2463 | }; |
| 2464 | |
| 2465 | llvm::CanonicalLoopInfo *CL = |
| 2466 | cantFail(ValOrErr: OMPBuilder.createCanonicalLoop(Loc: Builder, BodyGenCB: BodyGen, TripCount: DistVal)); |
| 2467 | |
| 2468 | // Finish up the loop. |
| 2469 | Builder.restoreIP(IP: CL->getAfterIP()); |
| 2470 | ForScope.ForceCleanup(); |
| 2471 | |
| 2472 | // Remember the CanonicalLoopInfo for parent AST nodes consuming it. |
| 2473 | OMPLoopNestStack.push_back(Elt: CL); |
| 2474 | } |
| 2475 | |
| 2476 | void CodeGenFunction::EmitOMPInnerLoop( |
| 2477 | const OMPExecutableDirective &S, bool RequiresCleanup, const Expr *LoopCond, |
| 2478 | const Expr *IncExpr, |
| 2479 | const llvm::function_ref<void(CodeGenFunction &)> BodyGen, |
| 2480 | const llvm::function_ref<void(CodeGenFunction &)> PostIncGen) { |
| 2481 | auto LoopExit = getJumpDestInCurrentScope(Name: "omp.inner.for.end"); |
| 2482 | |
| 2483 | // Start the loop with a block that tests the condition. |
| 2484 | auto CondBlock = createBasicBlock(name: "omp.inner.for.cond"); |
| 2485 | EmitBlock(BB: CondBlock); |
| 2486 | const SourceRange R = S.getSourceRange(); |
| 2487 | |
| 2488 | // If attributes are attached, push to the basic block with them. |
| 2489 | const auto &OMPED = cast<OMPExecutableDirective>(Val: S); |
| 2490 | const CapturedStmt *ICS = OMPED.getInnermostCapturedStmt(); |
| 2491 | const Stmt *SS = ICS->getCapturedStmt(); |
| 2492 | const AttributedStmt *AS = dyn_cast_or_null<AttributedStmt>(Val: SS); |
| 2493 | OMPLoopNestStack.clear(); |
| 2494 | if (AS) |
| 2495 | LoopStack.push(Header: CondBlock, Ctx&: CGM.getContext(), CGOpts: CGM.getCodeGenOpts(), |
| 2496 | Attrs: AS->getAttrs(), StartLoc: SourceLocToDebugLoc(Location: R.getBegin()), |
| 2497 | EndLoc: SourceLocToDebugLoc(Location: R.getEnd())); |
| 2498 | else |
| 2499 | LoopStack.push(Header: CondBlock, StartLoc: SourceLocToDebugLoc(Location: R.getBegin()), |
| 2500 | EndLoc: SourceLocToDebugLoc(Location: R.getEnd())); |
| 2501 | |
| 2502 | // If there are any cleanups between here and the loop-exit scope, |
| 2503 | // create a block to stage a loop exit along. |
| 2504 | llvm::BasicBlock *ExitBlock = LoopExit.getBlock(); |
| 2505 | if (RequiresCleanup) |
| 2506 | ExitBlock = createBasicBlock(name: "omp.inner.for.cond.cleanup"); |
| 2507 | |
| 2508 | llvm::BasicBlock *LoopBody = createBasicBlock(name: "omp.inner.for.body"); |
| 2509 | |
| 2510 | // Emit condition. |
| 2511 | EmitBranchOnBoolExpr(Cond: LoopCond, TrueBlock: LoopBody, FalseBlock: ExitBlock, TrueCount: getProfileCount(S: &S)); |
| 2512 | if (ExitBlock != LoopExit.getBlock()) { |
| 2513 | EmitBlock(BB: ExitBlock); |
| 2514 | EmitBranchThroughCleanup(Dest: LoopExit); |
| 2515 | } |
| 2516 | |
| 2517 | EmitBlock(BB: LoopBody); |
| 2518 | incrementProfileCounter(S: &S); |
| 2519 | |
| 2520 | // Create a block for the increment. |
| 2521 | JumpDest Continue = getJumpDestInCurrentScope(Name: "omp.inner.for.inc"); |
| 2522 | BreakContinueStack.push_back(Elt: BreakContinue(S, LoopExit, Continue)); |
| 2523 | |
| 2524 | BodyGen(*this); |
| 2525 | |
| 2526 | // Emit "IV = IV + 1" and a back-edge to the condition block. |
| 2527 | EmitBlock(BB: Continue.getBlock()); |
| 2528 | EmitIgnoredExpr(E: IncExpr); |
| 2529 | PostIncGen(*this); |
| 2530 | BreakContinueStack.pop_back(); |
| 2531 | EmitBranch(Block: CondBlock); |
| 2532 | LoopStack.pop(); |
| 2533 | // Emit the fall-through block. |
| 2534 | EmitBlock(BB: LoopExit.getBlock()); |
| 2535 | } |
| 2536 | |
| 2537 | bool CodeGenFunction::EmitOMPLinearClauseInit(const OMPLoopDirective &D) { |
| 2538 | if (!HaveInsertPoint()) |
| 2539 | return false; |
| 2540 | // Emit inits for the linear variables. |
| 2541 | bool HasLinears = false; |
| 2542 | for (const auto *C : D.getClausesOfKind<OMPLinearClause>()) { |
| 2543 | for (const Expr *Init : C->inits()) { |
| 2544 | HasLinears = true; |
| 2545 | const auto *VD = cast<VarDecl>(Val: cast<DeclRefExpr>(Val: Init)->getDecl()); |
| 2546 | if (const auto *Ref = |
| 2547 | dyn_cast<DeclRefExpr>(Val: VD->getInit()->IgnoreImpCasts())) { |
| 2548 | AutoVarEmission Emission = EmitAutoVarAlloca(var: *VD); |
| 2549 | const auto *OrigVD = cast<VarDecl>(Val: Ref->getDecl()); |
| 2550 | DeclRefExpr DRE(getContext(), const_cast<VarDecl *>(OrigVD), |
| 2551 | CapturedStmtInfo->lookup(VD: OrigVD) != nullptr, |
| 2552 | VD->getInit()->getType(), VK_LValue, |
| 2553 | VD->getInit()->getExprLoc()); |
| 2554 | EmitExprAsInit( |
| 2555 | init: &DRE, D: VD, |
| 2556 | lvalue: MakeAddrLValue(Addr: Emission.getAllocatedAddress(), T: VD->getType()), |
| 2557 | /*capturedByInit=*/false); |
| 2558 | EmitAutoVarCleanups(emission: Emission); |
| 2559 | } else { |
| 2560 | EmitVarDecl(D: *VD); |
| 2561 | } |
| 2562 | } |
| 2563 | // Emit the linear steps for the linear clauses. |
| 2564 | // If a step is not constant, it is pre-calculated before the loop. |
| 2565 | if (const auto *CS = cast_or_null<BinaryOperator>(Val: C->getCalcStep())) |
| 2566 | if (const auto *SaveRef = cast<DeclRefExpr>(Val: CS->getLHS())) { |
| 2567 | EmitVarDecl(D: *cast<VarDecl>(Val: SaveRef->getDecl())); |
| 2568 | // Emit calculation of the linear step. |
| 2569 | EmitIgnoredExpr(E: CS); |
| 2570 | } |
| 2571 | } |
| 2572 | return HasLinears; |
| 2573 | } |
| 2574 | |
| 2575 | void CodeGenFunction::EmitOMPLinearClauseFinal( |
| 2576 | const OMPLoopDirective &D, |
| 2577 | const llvm::function_ref<llvm::Value *(CodeGenFunction &)> CondGen) { |
| 2578 | if (!HaveInsertPoint()) |
| 2579 | return; |
| 2580 | llvm::BasicBlock *DoneBB = nullptr; |
| 2581 | // Emit the final values of the linear variables. |
| 2582 | for (const auto *C : D.getClausesOfKind<OMPLinearClause>()) { |
| 2583 | auto IC = C->varlist_begin(); |
| 2584 | for (const Expr *F : C->finals()) { |
| 2585 | if (!DoneBB) { |
| 2586 | if (llvm::Value *Cond = CondGen(*this)) { |
| 2587 | // If the first post-update expression is found, emit conditional |
| 2588 | // block if it was requested. |
| 2589 | llvm::BasicBlock *ThenBB = createBasicBlock(name: ".omp.linear.pu"); |
| 2590 | DoneBB = createBasicBlock(name: ".omp.linear.pu.done"); |
| 2591 | Builder.CreateCondBr(Cond, True: ThenBB, False: DoneBB); |
| 2592 | EmitBlock(BB: ThenBB); |
| 2593 | } |
| 2594 | } |
| 2595 | const auto *OrigVD = cast<VarDecl>(Val: cast<DeclRefExpr>(Val: *IC)->getDecl()); |
| 2596 | DeclRefExpr DRE(getContext(), const_cast<VarDecl *>(OrigVD), |
| 2597 | CapturedStmtInfo->lookup(VD: OrigVD) != nullptr, |
| 2598 | (*IC)->getType(), VK_LValue, (*IC)->getExprLoc()); |
| 2599 | Address OrigAddr = EmitLValue(E: &DRE).getAddress(); |
| 2600 | CodeGenFunction::OMPPrivateScope VarScope(*this); |
| 2601 | VarScope.addPrivate(LocalVD: OrigVD, Addr: OrigAddr); |
| 2602 | (void)VarScope.Privatize(); |
| 2603 | EmitIgnoredExpr(E: F); |
| 2604 | ++IC; |
| 2605 | } |
| 2606 | if (const Expr *PostUpdate = C->getPostUpdateExpr()) |
| 2607 | EmitIgnoredExpr(E: PostUpdate); |
| 2608 | } |
| 2609 | if (DoneBB) |
| 2610 | EmitBlock(BB: DoneBB, /*IsFinished=*/true); |
| 2611 | } |
| 2612 | |
| 2613 | static void emitAlignedClause(CodeGenFunction &CGF, |
| 2614 | const OMPExecutableDirective &D) { |
| 2615 | if (!CGF.HaveInsertPoint()) |
| 2616 | return; |
| 2617 | for (const auto *Clause : D.getClausesOfKind<OMPAlignedClause>()) { |
| 2618 | llvm::APInt ClauseAlignment(64, 0); |
| 2619 | if (const Expr *AlignmentExpr = Clause->getAlignment()) { |
| 2620 | auto *AlignmentCI = |
| 2621 | cast<llvm::ConstantInt>(Val: CGF.EmitScalarExpr(E: AlignmentExpr)); |
| 2622 | ClauseAlignment = AlignmentCI->getValue(); |
| 2623 | } |
| 2624 | for (const Expr *E : Clause->varlist()) { |
| 2625 | llvm::APInt Alignment(ClauseAlignment); |
| 2626 | if (Alignment == 0) { |
| 2627 | // OpenMP [2.8.1, Description] |
| 2628 | // If no optional parameter is specified, implementation-defined default |
| 2629 | // alignments for SIMD instructions on the target platforms are assumed. |
| 2630 | Alignment = |
| 2631 | CGF.getContext() |
| 2632 | .toCharUnitsFromBits(BitSize: CGF.getContext().getOpenMPDefaultSimdAlign( |
| 2633 | T: E->getType()->getPointeeType())) |
| 2634 | .getQuantity(); |
| 2635 | } |
| 2636 | assert((Alignment == 0 || Alignment.isPowerOf2()) && |
| 2637 | "alignment is not power of 2"); |
| 2638 | if (Alignment != 0) { |
| 2639 | llvm::Value *PtrValue = CGF.EmitScalarExpr(E); |
| 2640 | CGF.emitAlignmentAssumption( |
| 2641 | PtrValue, E, /*No second loc needed*/ AssumptionLoc: SourceLocation(), |
| 2642 | Alignment: llvm::ConstantInt::get(Context&: CGF.getLLVMContext(), V: Alignment)); |
| 2643 | } |
| 2644 | } |
| 2645 | } |
| 2646 | } |
| 2647 | |
| 2648 | void CodeGenFunction::EmitOMPPrivateLoopCounters( |
| 2649 | const OMPLoopDirective &S, CodeGenFunction::OMPPrivateScope &LoopScope) { |
| 2650 | if (!HaveInsertPoint()) |
| 2651 | return; |
| 2652 | auto I = S.private_counters().begin(); |
| 2653 | for (const Expr *E : S.counters()) { |
| 2654 | const auto *VD = cast<VarDecl>(Val: cast<DeclRefExpr>(Val: E)->getDecl()); |
| 2655 | const auto *PrivateVD = cast<VarDecl>(Val: cast<DeclRefExpr>(Val: *I)->getDecl()); |
| 2656 | // Emit var without initialization. |
| 2657 | AutoVarEmission VarEmission = EmitAutoVarAlloca(var: *PrivateVD); |
| 2658 | EmitAutoVarCleanups(emission: VarEmission); |
| 2659 | LocalDeclMap.erase(Val: PrivateVD); |
| 2660 | (void)LoopScope.addPrivate(LocalVD: VD, Addr: VarEmission.getAllocatedAddress()); |
| 2661 | if (LocalDeclMap.count(Val: VD) || CapturedStmtInfo->lookup(VD) || |
| 2662 | VD->hasGlobalStorage()) { |
| 2663 | DeclRefExpr DRE(getContext(), const_cast<VarDecl *>(VD), |
| 2664 | LocalDeclMap.count(Val: VD) || CapturedStmtInfo->lookup(VD), |
| 2665 | E->getType(), VK_LValue, E->getExprLoc()); |
| 2666 | (void)LoopScope.addPrivate(LocalVD: PrivateVD, Addr: EmitLValue(E: &DRE).getAddress()); |
| 2667 | } else { |
| 2668 | (void)LoopScope.addPrivate(LocalVD: PrivateVD, Addr: VarEmission.getAllocatedAddress()); |
| 2669 | } |
| 2670 | ++I; |
| 2671 | } |
| 2672 | // Privatize extra loop counters used in loops for ordered(n) clauses. |
| 2673 | for (const auto *C : S.getClausesOfKind<OMPOrderedClause>()) { |
| 2674 | if (!C->getNumForLoops()) |
| 2675 | continue; |
| 2676 | for (unsigned I = S.getLoopsNumber(), E = C->getLoopNumIterations().size(); |
| 2677 | I < E; ++I) { |
| 2678 | const auto *DRE = cast<DeclRefExpr>(Val: C->getLoopCounter(NumLoop: I)); |
| 2679 | const auto *VD = cast<VarDecl>(Val: DRE->getDecl()); |
| 2680 | // Override only those variables that can be captured to avoid re-emission |
| 2681 | // of the variables declared within the loops. |
| 2682 | if (DRE->refersToEnclosingVariableOrCapture()) { |
| 2683 | (void)LoopScope.addPrivate( |
| 2684 | LocalVD: VD, Addr: CreateMemTemp(T: DRE->getType(), Name: VD->getName())); |
| 2685 | } |
| 2686 | } |
| 2687 | } |
| 2688 | } |
| 2689 | |
| 2690 | static void emitPreCond(CodeGenFunction &CGF, const OMPLoopDirective &S, |
| 2691 | const Expr *Cond, llvm::BasicBlock *TrueBlock, |
| 2692 | llvm::BasicBlock *FalseBlock, uint64_t TrueCount) { |
| 2693 | if (!CGF.HaveInsertPoint()) |
| 2694 | return; |
| 2695 | { |
| 2696 | CodeGenFunction::OMPPrivateScope PreCondScope(CGF); |
| 2697 | CGF.EmitOMPPrivateLoopCounters(S, LoopScope&: PreCondScope); |
| 2698 | (void)PreCondScope.Privatize(); |
| 2699 | // Get initial values of real counters. |
| 2700 | for (const Expr *I : S.inits()) { |
| 2701 | CGF.EmitIgnoredExpr(E: I); |
| 2702 | } |
| 2703 | } |
| 2704 | // Create temp loop control variables with their init values to support |
| 2705 | // non-rectangular loops. |
| 2706 | CodeGenFunction::OMPMapVars PreCondVars; |
| 2707 | for (const Expr *E : S.dependent_counters()) { |
| 2708 | if (!E) |
| 2709 | continue; |
| 2710 | assert(!E->getType().getNonReferenceType()->isRecordType() && |
| 2711 | "dependent counter must not be an iterator."); |
| 2712 | const auto *VD = cast<VarDecl>(Val: cast<DeclRefExpr>(Val: E)->getDecl()); |
| 2713 | Address CounterAddr = |
| 2714 | CGF.CreateMemTemp(T: VD->getType().getNonReferenceType()); |
| 2715 | (void)PreCondVars.setVarAddr(CGF, LocalVD: VD, TempAddr: CounterAddr); |
| 2716 | } |
| 2717 | (void)PreCondVars.apply(CGF); |
| 2718 | for (const Expr *E : S.dependent_inits()) { |
| 2719 | if (!E) |
| 2720 | continue; |
| 2721 | CGF.EmitIgnoredExpr(E); |
| 2722 | } |
| 2723 | // Check that loop is executed at least one time. |
| 2724 | CGF.EmitBranchOnBoolExpr(Cond, TrueBlock, FalseBlock, TrueCount); |
| 2725 | PreCondVars.restore(CGF); |
| 2726 | } |
| 2727 | |
| 2728 | void CodeGenFunction::EmitOMPLinearClause( |
| 2729 | const OMPLoopDirective &D, CodeGenFunction::OMPPrivateScope &PrivateScope) { |
| 2730 | if (!HaveInsertPoint()) |
| 2731 | return; |
| 2732 | llvm::DenseSet<const VarDecl *> SIMDLCVs; |
| 2733 | OpenMPDirectiveKind EKind = getEffectiveDirectiveKind(S: D); |
| 2734 | if (isOpenMPSimdDirective(DKind: EKind)) { |
| 2735 | const auto *LoopDirective = cast<OMPLoopDirective>(Val: &D); |
| 2736 | for (const Expr *C : LoopDirective->counters()) { |
| 2737 | SIMDLCVs.insert( |
| 2738 | V: cast<VarDecl>(Val: cast<DeclRefExpr>(Val: C)->getDecl())->getCanonicalDecl()); |
| 2739 | } |
| 2740 | } |
| 2741 | for (const auto *C : D.getClausesOfKind<OMPLinearClause>()) { |
| 2742 | auto CurPrivate = C->privates().begin(); |
| 2743 | for (const Expr *E : C->varlist()) { |
| 2744 | const auto *VD = cast<VarDecl>(Val: cast<DeclRefExpr>(Val: E)->getDecl()); |
| 2745 | const auto *PrivateVD = |
| 2746 | cast<VarDecl>(Val: cast<DeclRefExpr>(Val: *CurPrivate)->getDecl()); |
| 2747 | if (!SIMDLCVs.count(V: VD->getCanonicalDecl())) { |
| 2748 | // Emit private VarDecl with copy init. |
| 2749 | EmitVarDecl(D: *PrivateVD); |
| 2750 | bool IsRegistered = |
| 2751 | PrivateScope.addPrivate(LocalVD: VD, Addr: GetAddrOfLocalVar(VD: PrivateVD)); |
| 2752 | assert(IsRegistered && "linear var already registered as private"); |
| 2753 | // Silence the warning about unused variable. |
| 2754 | (void)IsRegistered; |
| 2755 | } else { |
| 2756 | EmitVarDecl(D: *PrivateVD); |
| 2757 | } |
| 2758 | ++CurPrivate; |
| 2759 | } |
| 2760 | } |
| 2761 | } |
| 2762 | |
| 2763 | static void emitSimdlenSafelenClause(CodeGenFunction &CGF, |
| 2764 | const OMPExecutableDirective &D) { |
| 2765 | if (!CGF.HaveInsertPoint()) |
| 2766 | return; |
| 2767 | if (const auto *C = D.getSingleClause<OMPSimdlenClause>()) { |
| 2768 | RValue Len = CGF.EmitAnyExpr(E: C->getSimdlen(), aggSlot: AggValueSlot::ignored(), |
| 2769 | /*ignoreResult=*/true); |
| 2770 | auto *Val = cast<llvm::ConstantInt>(Val: Len.getScalarVal()); |
| 2771 | CGF.LoopStack.setVectorizeWidth(Val->getZExtValue()); |
| 2772 | // In presence of finite 'safelen', it may be unsafe to mark all |
| 2773 | // the memory instructions parallel, because loop-carried |
| 2774 | // dependences of 'safelen' iterations are possible. |
| 2775 | CGF.LoopStack.setParallel(!D.getSingleClause<OMPSafelenClause>()); |
| 2776 | } else if (const auto *C = D.getSingleClause<OMPSafelenClause>()) { |
| 2777 | RValue Len = CGF.EmitAnyExpr(E: C->getSafelen(), aggSlot: AggValueSlot::ignored(), |
| 2778 | /*ignoreResult=*/true); |
| 2779 | auto *Val = cast<llvm::ConstantInt>(Val: Len.getScalarVal()); |
| 2780 | CGF.LoopStack.setVectorizeWidth(Val->getZExtValue()); |
| 2781 | // In presence of finite 'safelen', it may be unsafe to mark all |
| 2782 | // the memory instructions parallel, because loop-carried |
| 2783 | // dependences of 'safelen' iterations are possible. |
| 2784 | CGF.LoopStack.setParallel(/*Enable=*/false); |
| 2785 | } |
| 2786 | } |
| 2787 | |
| 2788 | // Check for the presence of an `OMPOrderedBlockAssocDirective`, |
| 2789 | // i.e., `ordered` in `#pragma omp ordered simd`. |
| 2790 | // |
| 2791 | // Consider the following source code: |
| 2792 | // ``` |
| 2793 | // __attribute__((noinline)) void omp_simd_loop(float X[ARRAY_SIZE][ARRAY_SIZE]) |
| 2794 | // { |
| 2795 | // for (int r = 1; r < ARRAY_SIZE; ++r) { |
| 2796 | // for (int c = 1; c < ARRAY_SIZE; ++c) { |
| 2797 | // #pragma omp simd |
| 2798 | // for (int k = 2; k < ARRAY_SIZE; ++k) { |
| 2799 | // #pragma omp ordered simd |
| 2800 | // X[r][k] = X[r][k - 2] + sinf((float)(r / c)); |
| 2801 | // } |
| 2802 | // } |
| 2803 | // } |
| 2804 | // } |
| 2805 | // ``` |
| 2806 | // |
| 2807 | // Suppose we are in `CodeGenFunction::EmitOMPSimdInit(const OMPLoopDirective |
| 2808 | // &D)`. By examining `D.dump()` we have the following AST containing |
| 2809 | // `OMPOrderedBlockAssocDirective`: |
| 2810 | // |
| 2811 | // ``` |
| 2812 | // OMPSimdDirective 0x1c32950 |
| 2813 | // `-CapturedStmt 0x1c32028 |
| 2814 | // |-CapturedDecl 0x1c310e8 |
| 2815 | // | |-ForStmt 0x1c31e30 |
| 2816 | // | | |-DeclStmt 0x1c31298 |
| 2817 | // | | | `-VarDecl 0x1c31208 used k 'int' cinit |
| 2818 | // | | | `-IntegerLiteral 0x1c31278 'int' 2 |
| 2819 | // | | |-<<<NULL>>> |
| 2820 | // | | |-BinaryOperator 0x1c31308 'int' '<' |
| 2821 | // | | | |-ImplicitCastExpr 0x1c312f0 'int' <LValueToRValue> |
| 2822 | // | | | | `-DeclRefExpr 0x1c312b0 'int' lvalue Var 0x1c31208 'k' 'int' |
| 2823 | // | | | `-IntegerLiteral 0x1c312d0 'int' 256 |
| 2824 | // | | |-UnaryOperator 0x1c31348 'int' prefix '++' |
| 2825 | // | | | `-DeclRefExpr 0x1c31328 'int' lvalue Var 0x1c31208 'k' 'int' |
| 2826 | // | | `-CompoundStmt 0x1c31e18 |
| 2827 | // | | `-OMPOrderedBlockAssocDirective 0x1c31dd8 |
| 2828 | // | | |-OMPSimdClause 0x1c31380 |
| 2829 | // | | `-CapturedStmt 0x1c31cd0 |
| 2830 | // ``` |
| 2831 | // |
| 2832 | // Note the presence of `OMPOrderedBlockAssocDirective` above: |
| 2833 | // It's (transitively) nested in a `CapturedStmt` representing the pragma |
| 2834 | // annotated compound statement. Thus, we need to consider this nesting and |
| 2835 | // include checking the `getCapturedStmt` in this case. |
| 2836 | static bool hasOrderedBlockAssocDirective(const Stmt *S) { |
| 2837 | if (isa<OMPOrderedBlockAssocDirective>(Val: S)) |
| 2838 | return true; |
| 2839 | |
| 2840 | if (const auto *CS = dyn_cast<CapturedStmt>(Val: S)) |
| 2841 | return hasOrderedBlockAssocDirective(S: CS->getCapturedStmt()); |
| 2842 | |
| 2843 | for (const Stmt *Child : S->children()) { |
| 2844 | if (Child && hasOrderedBlockAssocDirective(S: Child)) |
| 2845 | return true; |
| 2846 | } |
| 2847 | |
| 2848 | return false; |
| 2849 | } |
| 2850 | |
| 2851 | static void applyConservativeSimdOrderedDirective(const Stmt &AssociatedStmt, |
| 2852 | LoopInfoStack &LoopStack) { |
| 2853 | // Check for the presence of an `OMPOrderedBlockAssocDirective` |
| 2854 | // i.e., `ordered` in `#pragma omp ordered simd` |
| 2855 | bool HasOrderedDirective = hasOrderedBlockAssocDirective(S: &AssociatedStmt); |
| 2856 | // If present then conservatively disable loop vectorization |
| 2857 | // analogously to how `emitSimdlenSafelenClause` does. |
| 2858 | if (HasOrderedDirective) |
| 2859 | LoopStack.setParallel(/*Enable=*/false); |
| 2860 | } |
| 2861 | |
| 2862 | void CodeGenFunction::EmitOMPSimdInit(const OMPLoopDirective &D) { |
| 2863 | // Walk clauses and process safelen/lastprivate. |
| 2864 | LoopStack.setParallel(/*Enable=*/true); |
| 2865 | LoopStack.setVectorizeEnable(); |
| 2866 | const Stmt *AssociatedStmt = D.getAssociatedStmt(); |
| 2867 | applyConservativeSimdOrderedDirective(AssociatedStmt: *AssociatedStmt, LoopStack); |
| 2868 | emitSimdlenSafelenClause(CGF&: *this, D); |
| 2869 | if (const auto *C = D.getSingleClause<OMPOrderClause>()) |
| 2870 | if (C->getKind() == OMPC_ORDER_concurrent) |
| 2871 | LoopStack.setParallel(/*Enable=*/true); |
| 2872 | OpenMPDirectiveKind EKind = getEffectiveDirectiveKind(S: D); |
| 2873 | if ((EKind == OMPD_simd || |
| 2874 | (getLangOpts().OpenMPSimd && isOpenMPSimdDirective(DKind: EKind))) && |
| 2875 | llvm::any_of(Range: D.getClausesOfKind<OMPReductionClause>(), |
| 2876 | P: [](const OMPReductionClause *C) { |
| 2877 | return C->getModifier() == OMPC_REDUCTION_inscan; |
| 2878 | })) |
| 2879 | // Disable parallel access in case of prefix sum. |
| 2880 | LoopStack.setParallel(/*Enable=*/false); |
| 2881 | } |
| 2882 | |
| 2883 | void CodeGenFunction::EmitOMPSimdFinal( |
| 2884 | const OMPLoopDirective &D, |
| 2885 | const llvm::function_ref<llvm::Value *(CodeGenFunction &)> CondGen) { |
| 2886 | if (!HaveInsertPoint()) |
| 2887 | return; |
| 2888 | llvm::BasicBlock *DoneBB = nullptr; |
| 2889 | auto IC = D.counters().begin(); |
| 2890 | auto IPC = D.private_counters().begin(); |
| 2891 | for (const Expr *F : D.finals()) { |
| 2892 | const auto *OrigVD = cast<VarDecl>(Val: cast<DeclRefExpr>(Val: (*IC))->getDecl()); |
| 2893 | const auto *PrivateVD = cast<VarDecl>(Val: cast<DeclRefExpr>(Val: (*IPC))->getDecl()); |
| 2894 | const auto *CED = dyn_cast<OMPCapturedExprDecl>(Val: OrigVD); |
| 2895 | if (LocalDeclMap.count(Val: OrigVD) || CapturedStmtInfo->lookup(VD: OrigVD) || |
| 2896 | OrigVD->hasGlobalStorage() || CED) { |
| 2897 | if (!DoneBB) { |
| 2898 | if (llvm::Value *Cond = CondGen(*this)) { |
| 2899 | // If the first post-update expression is found, emit conditional |
| 2900 | // block if it was requested. |
| 2901 | llvm::BasicBlock *ThenBB = createBasicBlock(name: ".omp.final.then"); |
| 2902 | DoneBB = createBasicBlock(name: ".omp.final.done"); |
| 2903 | Builder.CreateCondBr(Cond, True: ThenBB, False: DoneBB); |
| 2904 | EmitBlock(BB: ThenBB); |
| 2905 | } |
| 2906 | } |
| 2907 | Address OrigAddr = Address::invalid(); |
| 2908 | if (CED) { |
| 2909 | OrigAddr = EmitLValue(E: CED->getInit()->IgnoreImpCasts()).getAddress(); |
| 2910 | } else { |
| 2911 | DeclRefExpr DRE(getContext(), const_cast<VarDecl *>(PrivateVD), |
| 2912 | /*RefersToEnclosingVariableOrCapture=*/false, |
| 2913 | (*IPC)->getType(), VK_LValue, (*IPC)->getExprLoc()); |
| 2914 | OrigAddr = EmitLValue(E: &DRE).getAddress(); |
| 2915 | } |
| 2916 | OMPPrivateScope VarScope(*this); |
| 2917 | VarScope.addPrivate(LocalVD: OrigVD, Addr: OrigAddr); |
| 2918 | (void)VarScope.Privatize(); |
| 2919 | EmitIgnoredExpr(E: F); |
| 2920 | } |
| 2921 | ++IC; |
| 2922 | ++IPC; |
| 2923 | } |
| 2924 | if (DoneBB) |
| 2925 | EmitBlock(BB: DoneBB, /*IsFinished=*/true); |
| 2926 | } |
| 2927 | |
| 2928 | static void emitOMPLoopBodyWithStopPoint(CodeGenFunction &CGF, |
| 2929 | const OMPLoopDirective &S, |
| 2930 | CodeGenFunction::JumpDest LoopExit) { |
| 2931 | CGF.EmitOMPLoopBody(D: S, LoopExit); |
| 2932 | CGF.EmitStopPoint(S: &S); |
| 2933 | } |
| 2934 | |
| 2935 | /// Emit a helper variable and return corresponding lvalue. |
| 2936 | static LValue EmitOMPHelperVar(CodeGenFunction &CGF, |
| 2937 | const DeclRefExpr *Helper) { |
| 2938 | auto VDecl = cast<VarDecl>(Val: Helper->getDecl()); |
| 2939 | CGF.EmitVarDecl(D: *VDecl); |
| 2940 | return CGF.EmitLValue(E: Helper); |
| 2941 | } |
| 2942 | |
| 2943 | static void emitCommonSimdLoop(CodeGenFunction &CGF, const OMPLoopDirective &S, |
| 2944 | const RegionCodeGenTy &SimdInitGen, |
| 2945 | const RegionCodeGenTy &BodyCodeGen) { |
| 2946 | auto &&ThenGen = [&S, &SimdInitGen, &BodyCodeGen](CodeGenFunction &CGF, |
| 2947 | PrePostActionTy &) { |
| 2948 | CGOpenMPRuntime::NontemporalDeclsRAII NontemporalsRegion(CGF.CGM, S); |
| 2949 | CodeGenFunction::OMPLocalDeclMapRAII Scope(CGF); |
| 2950 | SimdInitGen(CGF); |
| 2951 | |
| 2952 | BodyCodeGen(CGF); |
| 2953 | }; |
| 2954 | auto &&ElseGen = [&BodyCodeGen](CodeGenFunction &CGF, PrePostActionTy &) { |
| 2955 | CodeGenFunction::OMPLocalDeclMapRAII Scope(CGF); |
| 2956 | CGF.LoopStack.setVectorizeEnable(/*Enable=*/false); |
| 2957 | |
| 2958 | BodyCodeGen(CGF); |
| 2959 | }; |
| 2960 | const Expr *IfCond = nullptr; |
| 2961 | OpenMPDirectiveKind EKind = getEffectiveDirectiveKind(S); |
| 2962 | if (isOpenMPSimdDirective(DKind: EKind)) { |
| 2963 | for (const auto *C : S.getClausesOfKind<OMPIfClause>()) { |
| 2964 | if (CGF.getLangOpts().OpenMP >= 50 && |
| 2965 | (C->getNameModifier() == OMPD_unknown || |
| 2966 | C->getNameModifier() == OMPD_simd)) { |
| 2967 | IfCond = C->getCondition(); |
| 2968 | break; |
| 2969 | } |
| 2970 | } |
| 2971 | } |
| 2972 | if (IfCond) { |
| 2973 | CGF.CGM.getOpenMPRuntime().emitIfClause(CGF, Cond: IfCond, ThenGen, ElseGen); |
| 2974 | } else { |
| 2975 | RegionCodeGenTy ThenRCG(ThenGen); |
| 2976 | ThenRCG(CGF); |
| 2977 | } |
| 2978 | } |
| 2979 | |
| 2980 | static void emitOMPSimdRegion(CodeGenFunction &CGF, const OMPLoopDirective &S, |
| 2981 | PrePostActionTy &Action) { |
| 2982 | Action.Enter(CGF); |
| 2983 | OMPLoopScope PreInitScope(CGF, S); |
| 2984 | // if (PreCond) { |
| 2985 | // for (IV in 0..LastIteration) BODY; |
| 2986 | // <Final counter/linear vars updates>; |
| 2987 | // } |
| 2988 | |
| 2989 | // The presence of lower/upper bound variable depends on the actual directive |
| 2990 | // kind in the AST node. The variables must be emitted because some of the |
| 2991 | // expressions associated with the loop will use them. |
| 2992 | OpenMPDirectiveKind DKind = S.getDirectiveKind(); |
| 2993 | if (isOpenMPDistributeDirective(DKind) || |
| 2994 | isOpenMPWorksharingDirective(DKind) || isOpenMPTaskLoopDirective(DKind) || |
| 2995 | isOpenMPGenericLoopDirective(DKind)) { |
| 2996 | (void)EmitOMPHelperVar(CGF, Helper: cast<DeclRefExpr>(Val: S.getLowerBoundVariable())); |
| 2997 | (void)EmitOMPHelperVar(CGF, Helper: cast<DeclRefExpr>(Val: S.getUpperBoundVariable())); |
| 2998 | } |
| 2999 | |
| 3000 | OpenMPDirectiveKind EKind = getEffectiveDirectiveKind(S); |
| 3001 | // Emit: if (PreCond) - begin. |
| 3002 | // If the condition constant folds and can be elided, avoid emitting the |
| 3003 | // whole loop. |
| 3004 | bool CondConstant; |
| 3005 | llvm::BasicBlock *ContBlock = nullptr; |
| 3006 | if (CGF.ConstantFoldsToSimpleInteger(Cond: S.getPreCond(), Result&: CondConstant)) { |
| 3007 | if (!CondConstant) |
| 3008 | return; |
| 3009 | } else { |
| 3010 | llvm::BasicBlock *ThenBlock = CGF.createBasicBlock(name: "simd.if.then"); |
| 3011 | ContBlock = CGF.createBasicBlock(name: "simd.if.end"); |
| 3012 | emitPreCond(CGF, S, Cond: S.getPreCond(), TrueBlock: ThenBlock, FalseBlock: ContBlock, |
| 3013 | TrueCount: CGF.getProfileCount(S: &S)); |
| 3014 | CGF.EmitBlock(BB: ThenBlock); |
| 3015 | CGF.incrementProfileCounter(S: &S); |
| 3016 | } |
| 3017 | |
| 3018 | // Emit the loop iteration variable. |
| 3019 | const Expr *IVExpr = S.getIterationVariable(); |
| 3020 | const auto *IVDecl = cast<VarDecl>(Val: cast<DeclRefExpr>(Val: IVExpr)->getDecl()); |
| 3021 | CGF.EmitVarDecl(D: *IVDecl); |
| 3022 | CGF.EmitIgnoredExpr(E: S.getInit()); |
| 3023 | |
| 3024 | // Emit the iterations count variable. |
| 3025 | // If it is not a variable, Sema decided to calculate iterations count on |
| 3026 | // each iteration (e.g., it is foldable into a constant). |
| 3027 | if (const auto *LIExpr = dyn_cast<DeclRefExpr>(Val: S.getLastIteration())) { |
| 3028 | CGF.EmitVarDecl(D: *cast<VarDecl>(Val: LIExpr->getDecl())); |
| 3029 | // Emit calculation of the iterations count. |
| 3030 | CGF.EmitIgnoredExpr(E: S.getCalcLastIteration()); |
| 3031 | } |
| 3032 | |
| 3033 | emitAlignedClause(CGF, D: S); |
| 3034 | (void)CGF.EmitOMPLinearClauseInit(D: S); |
| 3035 | { |
| 3036 | CodeGenFunction::OMPPrivateScope LoopScope(CGF); |
| 3037 | CGF.EmitOMPPrivateClause(D: S, PrivateScope&: LoopScope); |
| 3038 | CGF.EmitOMPPrivateLoopCounters(S, LoopScope); |
| 3039 | CGF.EmitOMPLinearClause(D: S, PrivateScope&: LoopScope); |
| 3040 | CGF.EmitOMPReductionClauseInit(D: S, PrivateScope&: LoopScope); |
| 3041 | CGOpenMPRuntime::LastprivateConditionalRAII LPCRegion( |
| 3042 | CGF, S, CGF.EmitLValue(E: S.getIterationVariable())); |
| 3043 | bool HasLastprivateClause = CGF.EmitOMPLastprivateClauseInit(D: S, PrivateScope&: LoopScope); |
| 3044 | (void)LoopScope.Privatize(); |
| 3045 | if (isOpenMPTargetExecutionDirective(DKind: EKind)) |
| 3046 | CGF.CGM.getOpenMPRuntime().adjustTargetSpecificDataForLambdas(CGF, D: S); |
| 3047 | |
| 3048 | emitCommonSimdLoop( |
| 3049 | CGF, S, |
| 3050 | SimdInitGen: [&S](CodeGenFunction &CGF, PrePostActionTy &) { |
| 3051 | CGF.EmitOMPSimdInit(D: S); |
| 3052 | }, |
| 3053 | BodyCodeGen: [&S, &LoopScope](CodeGenFunction &CGF, PrePostActionTy &) { |
| 3054 | CGF.EmitOMPInnerLoop( |
| 3055 | S, RequiresCleanup: LoopScope.requiresCleanups(), LoopCond: S.getCond(), IncExpr: S.getInc(), |
| 3056 | BodyGen: [&S](CodeGenFunction &CGF) { |
| 3057 | emitOMPLoopBodyWithStopPoint(CGF, S, |
| 3058 | LoopExit: CodeGenFunction::JumpDest()); |
| 3059 | }, |
| 3060 | PostIncGen: [](CodeGenFunction &) {}); |
| 3061 | }); |
| 3062 | CGF.EmitOMPSimdFinal(D: S, CondGen: [](CodeGenFunction &) { return nullptr; }); |
| 3063 | // Emit final copy of the lastprivate variables at the end of loops. |
| 3064 | if (HasLastprivateClause) |
| 3065 | CGF.EmitOMPLastprivateClauseFinal(D: S, /*NoFinals=*/true); |
| 3066 | CGF.EmitOMPReductionClauseFinal(D: S, /*ReductionKind=*/OMPD_simd); |
| 3067 | emitPostUpdateForReductionClause(CGF, D: S, |
| 3068 | CondGen: [](CodeGenFunction &) { return nullptr; }); |
| 3069 | LoopScope.restoreMap(); |
| 3070 | CGF.EmitOMPLinearClauseFinal(D: S, CondGen: [](CodeGenFunction &) { return nullptr; }); |
| 3071 | } |
| 3072 | // Emit: if (PreCond) - end. |
| 3073 | if (ContBlock) { |
| 3074 | CGF.EmitBranch(Block: ContBlock); |
| 3075 | CGF.EmitBlock(BB: ContBlock, IsFinished: true); |
| 3076 | } |
| 3077 | } |
| 3078 | |
| 3079 | // Pass OMPLoopDirective (instead of OMPSimdDirective) to make this function |
| 3080 | // available for "loop bind(thread)", which maps to "simd". |
| 3081 | static bool isSimdSupportedByOpenMPIRBuilder(const OMPLoopDirective &S) { |
| 3082 | // Check for unsupported clauses |
| 3083 | for (OMPClause *C : S.clauses()) { |
| 3084 | // Currently only order, simdlen and safelen clauses are supported |
| 3085 | if (!(isa<OMPSimdlenClause>(Val: C) || isa<OMPSafelenClause>(Val: C) || |
| 3086 | isa<OMPOrderClause>(Val: C) || isa<OMPAlignedClause>(Val: C))) |
| 3087 | return false; |
| 3088 | } |
| 3089 | |
| 3090 | // Check if we have a statement with the ordered-blockassoc directive. |
| 3091 | // Visit the statement hierarchy to find a compound statement |
| 3092 | // with a ordered-blockassoc directive in it. |
| 3093 | if (const auto *CanonLoop = dyn_cast<OMPCanonicalLoop>(Val: S.getRawStmt())) { |
| 3094 | if (const Stmt *SyntacticalLoop = CanonLoop->getLoopStmt()) { |
| 3095 | for (const Stmt *SubStmt : SyntacticalLoop->children()) { |
| 3096 | if (!SubStmt) |
| 3097 | continue; |
| 3098 | if (const CompoundStmt *CS = dyn_cast<CompoundStmt>(Val: SubStmt)) { |
| 3099 | for (const Stmt *CSSubStmt : CS->children()) { |
| 3100 | if (!CSSubStmt) |
| 3101 | continue; |
| 3102 | if (isa<OMPOrderedBlockAssocDirective>(Val: CSSubStmt)) { |
| 3103 | return false; |
| 3104 | } |
| 3105 | } |
| 3106 | } |
| 3107 | } |
| 3108 | } |
| 3109 | } |
| 3110 | return true; |
| 3111 | } |
| 3112 | |
| 3113 | static llvm::MapVector<llvm::Value *, llvm::Value *> |
| 3114 | GetAlignedMapping(const OMPLoopDirective &S, CodeGenFunction &CGF) { |
| 3115 | llvm::MapVector<llvm::Value *, llvm::Value *> AlignedVars; |
| 3116 | for (const auto *Clause : S.getClausesOfKind<OMPAlignedClause>()) { |
| 3117 | llvm::APInt ClauseAlignment(64, 0); |
| 3118 | if (const Expr *AlignmentExpr = Clause->getAlignment()) { |
| 3119 | auto *AlignmentCI = |
| 3120 | cast<llvm::ConstantInt>(Val: CGF.EmitScalarExpr(E: AlignmentExpr)); |
| 3121 | ClauseAlignment = AlignmentCI->getValue(); |
| 3122 | } |
| 3123 | for (const Expr *E : Clause->varlist()) { |
| 3124 | llvm::APInt Alignment(ClauseAlignment); |
| 3125 | if (Alignment == 0) { |
| 3126 | // OpenMP [2.8.1, Description] |
| 3127 | // If no optional parameter is specified, implementation-defined default |
| 3128 | // alignments for SIMD instructions on the target platforms are assumed. |
| 3129 | Alignment = |
| 3130 | CGF.getContext() |
| 3131 | .toCharUnitsFromBits(BitSize: CGF.getContext().getOpenMPDefaultSimdAlign( |
| 3132 | T: E->getType()->getPointeeType())) |
| 3133 | .getQuantity(); |
| 3134 | } |
| 3135 | assert((Alignment == 0 || Alignment.isPowerOf2()) && |
| 3136 | "alignment is not power of 2"); |
| 3137 | llvm::Value *PtrValue = CGF.EmitScalarExpr(E); |
| 3138 | AlignedVars[PtrValue] = CGF.Builder.getInt64(C: Alignment.getSExtValue()); |
| 3139 | } |
| 3140 | } |
| 3141 | return AlignedVars; |
| 3142 | } |
| 3143 | |
| 3144 | // Pass OMPLoopDirective (instead of OMPSimdDirective) to make this function |
| 3145 | // available for "loop bind(thread)", which maps to "simd". |
| 3146 | static void emitOMPSimdDirective(const OMPLoopDirective &S, |
| 3147 | CodeGenFunction &CGF, CodeGenModule &CGM) { |
| 3148 | bool UseOMPIRBuilder = |
| 3149 | CGM.getLangOpts().OpenMPIRBuilder && isSimdSupportedByOpenMPIRBuilder(S); |
| 3150 | if (UseOMPIRBuilder) { |
| 3151 | auto &&CodeGenIRBuilder = [&S, &CGM, UseOMPIRBuilder](CodeGenFunction &CGF, |
| 3152 | PrePostActionTy &) { |
| 3153 | // Use the OpenMPIRBuilder if enabled. |
| 3154 | if (UseOMPIRBuilder) { |
| 3155 | llvm::MapVector<llvm::Value *, llvm::Value *> AlignedVars = |
| 3156 | GetAlignedMapping(S, CGF); |
| 3157 | // Emit the associated statement and get its loop representation. |
| 3158 | const Stmt *Inner = S.getRawStmt(); |
| 3159 | llvm::CanonicalLoopInfo *CLI = |
| 3160 | CGF.EmitOMPCollapsedCanonicalLoopNest(S: Inner, Depth: 1); |
| 3161 | |
| 3162 | llvm::OpenMPIRBuilder &OMPBuilder = |
| 3163 | CGM.getOpenMPRuntime().getOMPBuilder(); |
| 3164 | // Add SIMD specific metadata |
| 3165 | llvm::ConstantInt *Simdlen = nullptr; |
| 3166 | if (const auto *C = S.getSingleClause<OMPSimdlenClause>()) { |
| 3167 | RValue Len = CGF.EmitAnyExpr(E: C->getSimdlen(), aggSlot: AggValueSlot::ignored(), |
| 3168 | /*ignoreResult=*/true); |
| 3169 | auto *Val = cast<llvm::ConstantInt>(Val: Len.getScalarVal()); |
| 3170 | Simdlen = Val; |
| 3171 | } |
| 3172 | llvm::ConstantInt *Safelen = nullptr; |
| 3173 | if (const auto *C = S.getSingleClause<OMPSafelenClause>()) { |
| 3174 | RValue Len = CGF.EmitAnyExpr(E: C->getSafelen(), aggSlot: AggValueSlot::ignored(), |
| 3175 | /*ignoreResult=*/true); |
| 3176 | auto *Val = cast<llvm::ConstantInt>(Val: Len.getScalarVal()); |
| 3177 | Safelen = Val; |
| 3178 | } |
| 3179 | llvm::omp::OrderKind Order = llvm::omp::OrderKind::OMP_ORDER_unknown; |
| 3180 | if (const auto *C = S.getSingleClause<OMPOrderClause>()) { |
| 3181 | if (C->getKind() == OpenMPOrderClauseKind::OMPC_ORDER_concurrent) { |
| 3182 | Order = llvm::omp::OrderKind::OMP_ORDER_concurrent; |
| 3183 | } |
| 3184 | } |
| 3185 | // Add simd metadata to the collapsed loop. Do not generate |
| 3186 | // another loop for if clause. Support for if clause is done earlier. |
| 3187 | OMPBuilder.applySimd(Loop: CLI, AlignedVars, |
| 3188 | /*IfCond*/ nullptr, Order, Simdlen, Safelen); |
| 3189 | return; |
| 3190 | } |
| 3191 | }; |
| 3192 | { |
| 3193 | auto LPCRegion = |
| 3194 | CGOpenMPRuntime::LastprivateConditionalRAII::disable(CGF, S); |
| 3195 | OMPLexicalScope Scope(CGF, S, OMPD_unknown); |
| 3196 | CGM.getOpenMPRuntime().emitInlinedDirective(CGF, InnermostKind: OMPD_simd, |
| 3197 | CodeGen: CodeGenIRBuilder); |
| 3198 | } |
| 3199 | return; |
| 3200 | } |
| 3201 | |
| 3202 | CodeGenFunction::ParentLoopDirectiveForScanRegion ScanRegion(CGF, S); |
| 3203 | CGF.OMPFirstScanLoop = true; |
| 3204 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 3205 | emitOMPSimdRegion(CGF, S, Action); |
| 3206 | }; |
| 3207 | { |
| 3208 | auto LPCRegion = |
| 3209 | CGOpenMPRuntime::LastprivateConditionalRAII::disable(CGF, S); |
| 3210 | OMPLexicalScope Scope(CGF, S, OMPD_unknown); |
| 3211 | CGM.getOpenMPRuntime().emitInlinedDirective(CGF, InnermostKind: OMPD_simd, CodeGen); |
| 3212 | } |
| 3213 | // Check for outer lastprivate conditional update. |
| 3214 | checkForLastprivateConditionalUpdate(CGF, S); |
| 3215 | } |
| 3216 | |
| 3217 | void CodeGenFunction::EmitOMPSimdDirective(const OMPSimdDirective &S) { |
| 3218 | emitOMPSimdDirective(S, CGF&: *this, CGM); |
| 3219 | } |
| 3220 | |
| 3221 | void CodeGenFunction::EmitOMPTileDirective(const OMPTileDirective &S) { |
| 3222 | // Emit the de-sugared statement. |
| 3223 | OMPTransformDirectiveScopeRAII TileScope(*this, &S); |
| 3224 | EmitStmt(S: S.getTransformedStmt()); |
| 3225 | } |
| 3226 | |
| 3227 | void CodeGenFunction::EmitOMPStripeDirective(const OMPStripeDirective &S) { |
| 3228 | // Emit the de-sugared statement. |
| 3229 | OMPTransformDirectiveScopeRAII StripeScope(*this, &S); |
| 3230 | EmitStmt(S: S.getTransformedStmt()); |
| 3231 | } |
| 3232 | |
| 3233 | void CodeGenFunction::EmitOMPReverseDirective(const OMPReverseDirective &S) { |
| 3234 | // Emit the de-sugared statement. |
| 3235 | OMPTransformDirectiveScopeRAII ReverseScope(*this, &S); |
| 3236 | EmitStmt(S: S.getTransformedStmt()); |
| 3237 | } |
| 3238 | |
| 3239 | void CodeGenFunction::EmitOMPSplitDirective(const OMPSplitDirective &S) { |
| 3240 | // Emit the de-sugared statement (the split loops). |
| 3241 | OMPTransformDirectiveScopeRAII SplitScope(*this, &S); |
| 3242 | EmitStmt(S: S.getTransformedStmt()); |
| 3243 | } |
| 3244 | |
| 3245 | void CodeGenFunction::EmitOMPInterchangeDirective( |
| 3246 | const OMPInterchangeDirective &S) { |
| 3247 | // Emit the de-sugared statement. |
| 3248 | OMPTransformDirectiveScopeRAII InterchangeScope(*this, &S); |
| 3249 | EmitStmt(S: S.getTransformedStmt()); |
| 3250 | } |
| 3251 | |
| 3252 | void CodeGenFunction::EmitOMPFuseDirective(const OMPFuseDirective &S) { |
| 3253 | // Emit the de-sugared statement |
| 3254 | OMPTransformDirectiveScopeRAII FuseScope(*this, &S); |
| 3255 | EmitStmt(S: S.getTransformedStmt()); |
| 3256 | } |
| 3257 | |
| 3258 | void CodeGenFunction::EmitOMPUnrollDirective(const OMPUnrollDirective &S) { |
| 3259 | bool UseOMPIRBuilder = CGM.getLangOpts().OpenMPIRBuilder; |
| 3260 | |
| 3261 | if (UseOMPIRBuilder) { |
| 3262 | auto DL = SourceLocToDebugLoc(Location: S.getBeginLoc()); |
| 3263 | const Stmt *Inner = S.getRawStmt(); |
| 3264 | |
| 3265 | // Consume nested loop. Clear the entire remaining loop stack because a |
| 3266 | // fully unrolled loop is non-transformable. For partial unrolling the |
| 3267 | // generated outer loop is pushed back to the stack. |
| 3268 | llvm::CanonicalLoopInfo *CLI = EmitOMPCollapsedCanonicalLoopNest(S: Inner, Depth: 1); |
| 3269 | OMPLoopNestStack.clear(); |
| 3270 | |
| 3271 | llvm::OpenMPIRBuilder &OMPBuilder = CGM.getOpenMPRuntime().getOMPBuilder(); |
| 3272 | |
| 3273 | bool NeedsUnrolledCLI = ExpectedOMPLoopDepth >= 1; |
| 3274 | llvm::CanonicalLoopInfo *UnrolledCLI = nullptr; |
| 3275 | |
| 3276 | if (S.hasClausesOfKind<OMPFullClause>()) { |
| 3277 | assert(ExpectedOMPLoopDepth == 0); |
| 3278 | OMPBuilder.unrollLoopFull(DL, Loop: CLI); |
| 3279 | } else if (auto *PartialClause = S.getSingleClause<OMPPartialClause>()) { |
| 3280 | uint64_t Factor = 0; |
| 3281 | if (Expr *FactorExpr = PartialClause->getFactor()) { |
| 3282 | Factor = FactorExpr->EvaluateKnownConstInt(Ctx: getContext()).getZExtValue(); |
| 3283 | assert(Factor >= 1 && "Only positive factors are valid"); |
| 3284 | } |
| 3285 | OMPBuilder.unrollLoopPartial(DL, Loop: CLI, Factor, |
| 3286 | UnrolledCLI: NeedsUnrolledCLI ? &UnrolledCLI : nullptr); |
| 3287 | } else { |
| 3288 | OMPBuilder.unrollLoopHeuristic(DL, Loop: CLI); |
| 3289 | } |
| 3290 | |
| 3291 | assert((!NeedsUnrolledCLI || UnrolledCLI) && |
| 3292 | "NeedsUnrolledCLI implies UnrolledCLI to be set"); |
| 3293 | if (UnrolledCLI) |
| 3294 | OMPLoopNestStack.push_back(Elt: UnrolledCLI); |
| 3295 | |
| 3296 | return; |
| 3297 | } |
| 3298 | |
| 3299 | // This function is only called if the unrolled loop is not consumed by any |
| 3300 | // other loop-associated construct. Such a loop-associated construct will have |
| 3301 | // used the transformed AST. |
| 3302 | |
| 3303 | // Set the unroll metadata for the next emitted loop. |
| 3304 | LoopStack.setUnrollState(LoopAttributes::Enable); |
| 3305 | |
| 3306 | if (S.hasClausesOfKind<OMPFullClause>()) { |
| 3307 | LoopStack.setUnrollState(LoopAttributes::Full); |
| 3308 | } else if (auto *PartialClause = S.getSingleClause<OMPPartialClause>()) { |
| 3309 | if (Expr *FactorExpr = PartialClause->getFactor()) { |
| 3310 | uint64_t Factor = |
| 3311 | FactorExpr->EvaluateKnownConstInt(Ctx: getContext()).getZExtValue(); |
| 3312 | assert(Factor >= 1 && "Only positive factors are valid"); |
| 3313 | LoopStack.setUnrollCount(Factor); |
| 3314 | } |
| 3315 | } |
| 3316 | |
| 3317 | EmitStmt(S: S.getAssociatedStmt()); |
| 3318 | } |
| 3319 | |
| 3320 | void CodeGenFunction::EmitOMPOuterLoop( |
| 3321 | bool DynamicOrOrdered, bool IsMonotonic, const OMPLoopDirective &S, |
| 3322 | CodeGenFunction::OMPPrivateScope &LoopScope, |
| 3323 | const CodeGenFunction::OMPLoopArguments &LoopArgs, |
| 3324 | const CodeGenFunction::CodeGenLoopTy &CodeGenLoop, |
| 3325 | const CodeGenFunction::CodeGenOrderedTy &CodeGenOrdered) { |
| 3326 | CGOpenMPRuntime &RT = CGM.getOpenMPRuntime(); |
| 3327 | |
| 3328 | const Expr *IVExpr = S.getIterationVariable(); |
| 3329 | const unsigned IVSize = getContext().getTypeSize(T: IVExpr->getType()); |
| 3330 | const bool IVSigned = IVExpr->getType()->hasSignedIntegerRepresentation(); |
| 3331 | |
| 3332 | JumpDest LoopExit = getJumpDestInCurrentScope(Name: "omp.dispatch.end"); |
| 3333 | |
| 3334 | // Start the loop with a block that tests the condition. |
| 3335 | llvm::BasicBlock *CondBlock = createBasicBlock(name: "omp.dispatch.cond"); |
| 3336 | EmitBlock(BB: CondBlock); |
| 3337 | const SourceRange R = S.getSourceRange(); |
| 3338 | OMPLoopNestStack.clear(); |
| 3339 | LoopStack.push(Header: CondBlock, StartLoc: SourceLocToDebugLoc(Location: R.getBegin()), |
| 3340 | EndLoc: SourceLocToDebugLoc(Location: R.getEnd())); |
| 3341 | |
| 3342 | llvm::Value *BoolCondVal = nullptr; |
| 3343 | if (!DynamicOrOrdered) { |
| 3344 | // UB = min(UB, GlobalUB) or |
| 3345 | // UB = min(UB, PrevUB) for combined loop sharing constructs (e.g. |
| 3346 | // 'distribute parallel for') |
| 3347 | EmitIgnoredExpr(E: LoopArgs.EUB); |
| 3348 | // IV = LB |
| 3349 | EmitIgnoredExpr(E: LoopArgs.Init); |
| 3350 | // IV < UB |
| 3351 | BoolCondVal = EvaluateExprAsBool(E: LoopArgs.Cond); |
| 3352 | } else { |
| 3353 | BoolCondVal = |
| 3354 | RT.emitForNext(CGF&: *this, Loc: S.getBeginLoc(), IVSize, IVSigned, IL: LoopArgs.IL, |
| 3355 | LB: LoopArgs.LB, UB: LoopArgs.UB, ST: LoopArgs.ST); |
| 3356 | } |
| 3357 | |
| 3358 | // If there are any cleanups between here and the loop-exit scope, |
| 3359 | // create a block to stage a loop exit along. |
| 3360 | llvm::BasicBlock *ExitBlock = LoopExit.getBlock(); |
| 3361 | if (LoopScope.requiresCleanups()) |
| 3362 | ExitBlock = createBasicBlock(name: "omp.dispatch.cleanup"); |
| 3363 | |
| 3364 | llvm::BasicBlock *LoopBody = createBasicBlock(name: "omp.dispatch.body"); |
| 3365 | Builder.CreateCondBr(Cond: BoolCondVal, True: LoopBody, False: ExitBlock); |
| 3366 | if (ExitBlock != LoopExit.getBlock()) { |
| 3367 | EmitBlock(BB: ExitBlock); |
| 3368 | EmitBranchThroughCleanup(Dest: LoopExit); |
| 3369 | } |
| 3370 | EmitBlock(BB: LoopBody); |
| 3371 | |
| 3372 | // Emit "IV = LB" (in case of static schedule, we have already calculated new |
| 3373 | // LB for loop condition and emitted it above). |
| 3374 | if (DynamicOrOrdered) |
| 3375 | EmitIgnoredExpr(E: LoopArgs.Init); |
| 3376 | |
| 3377 | // Create a block for the increment. |
| 3378 | JumpDest Continue = getJumpDestInCurrentScope(Name: "omp.dispatch.inc"); |
| 3379 | BreakContinueStack.push_back(Elt: BreakContinue(S, LoopExit, Continue)); |
| 3380 | |
| 3381 | OpenMPDirectiveKind EKind = getEffectiveDirectiveKind(S); |
| 3382 | emitCommonSimdLoop( |
| 3383 | CGF&: *this, S, |
| 3384 | SimdInitGen: [&S, IsMonotonic, EKind](CodeGenFunction &CGF, PrePostActionTy &) { |
| 3385 | // Generate !llvm.loop.parallel metadata for loads and stores for loops |
| 3386 | // with dynamic/guided scheduling and without ordered clause. |
| 3387 | if (!isOpenMPSimdDirective(DKind: EKind)) { |
| 3388 | CGF.LoopStack.setParallel(!IsMonotonic); |
| 3389 | if (const auto *C = S.getSingleClause<OMPOrderClause>()) |
| 3390 | if (C->getKind() == OMPC_ORDER_concurrent) |
| 3391 | CGF.LoopStack.setParallel(/*Enable=*/true); |
| 3392 | } else { |
| 3393 | CGF.EmitOMPSimdInit(D: S); |
| 3394 | } |
| 3395 | }, |
| 3396 | BodyCodeGen: [&S, &LoopArgs, LoopExit, &CodeGenLoop, IVSize, IVSigned, &CodeGenOrdered, |
| 3397 | &LoopScope](CodeGenFunction &CGF, PrePostActionTy &) { |
| 3398 | SourceLocation Loc = S.getBeginLoc(); |
| 3399 | // when 'distribute' is not combined with a 'for': |
| 3400 | // while (idx <= UB) { BODY; ++idx; } |
| 3401 | // when 'distribute' is combined with a 'for' |
| 3402 | // (e.g. 'distribute parallel for') |
| 3403 | // while (idx <= UB) { <CodeGen rest of pragma>; idx += ST; } |
| 3404 | CGF.EmitOMPInnerLoop( |
| 3405 | S, RequiresCleanup: LoopScope.requiresCleanups(), LoopCond: LoopArgs.Cond, IncExpr: LoopArgs.IncExpr, |
| 3406 | BodyGen: [&S, LoopExit, &CodeGenLoop](CodeGenFunction &CGF) { |
| 3407 | CodeGenLoop(CGF, S, LoopExit); |
| 3408 | }, |
| 3409 | PostIncGen: [IVSize, IVSigned, Loc, &CodeGenOrdered](CodeGenFunction &CGF) { |
| 3410 | CodeGenOrdered(CGF, Loc, IVSize, IVSigned); |
| 3411 | }); |
| 3412 | }); |
| 3413 | |
| 3414 | EmitBlock(BB: Continue.getBlock()); |
| 3415 | BreakContinueStack.pop_back(); |
| 3416 | if (!DynamicOrOrdered) { |
| 3417 | // Emit "LB = LB + Stride", "UB = UB + Stride". |
| 3418 | EmitIgnoredExpr(E: LoopArgs.NextLB); |
| 3419 | EmitIgnoredExpr(E: LoopArgs.NextUB); |
| 3420 | } |
| 3421 | |
| 3422 | EmitBranch(Block: CondBlock); |
| 3423 | OMPLoopNestStack.clear(); |
| 3424 | LoopStack.pop(); |
| 3425 | // Emit the fall-through block. |
| 3426 | EmitBlock(BB: LoopExit.getBlock()); |
| 3427 | |
| 3428 | // Tell the runtime we are done. |
| 3429 | auto &&CodeGen = [DynamicOrOrdered, &S, &LoopArgs](CodeGenFunction &CGF) { |
| 3430 | if (!DynamicOrOrdered) |
| 3431 | CGF.CGM.getOpenMPRuntime().emitForStaticFinish(CGF, Loc: S.getEndLoc(), |
| 3432 | DKind: LoopArgs.DKind); |
| 3433 | }; |
| 3434 | OMPCancelStack.emitExit(CGF&: *this, Kind: EKind, CodeGen); |
| 3435 | } |
| 3436 | |
| 3437 | void CodeGenFunction::EmitOMPForOuterLoop( |
| 3438 | const OpenMPScheduleTy &ScheduleKind, bool IsMonotonic, |
| 3439 | const OMPLoopDirective &S, OMPPrivateScope &LoopScope, bool Ordered, |
| 3440 | const OMPLoopArguments &LoopArgs, |
| 3441 | const CodeGenDispatchBoundsTy &CGDispatchBounds) { |
| 3442 | CGOpenMPRuntime &RT = CGM.getOpenMPRuntime(); |
| 3443 | |
| 3444 | // Dynamic scheduling of the outer loop (dynamic, guided, auto, runtime). |
| 3445 | const bool DynamicOrOrdered = Ordered || RT.isDynamic(ScheduleKind: ScheduleKind.Schedule); |
| 3446 | |
| 3447 | assert((Ordered || !RT.isStaticNonchunked(ScheduleKind.Schedule, |
| 3448 | LoopArgs.Chunk != nullptr)) && |
| 3449 | "static non-chunked schedule does not need outer loop"); |
| 3450 | |
| 3451 | // Emit outer loop. |
| 3452 | // |
| 3453 | // OpenMP [2.7.1, Loop Construct, Description, table 2-1] |
| 3454 | // When schedule(dynamic,chunk_size) is specified, the iterations are |
| 3455 | // distributed to threads in the team in chunks as the threads request them. |
| 3456 | // Each thread executes a chunk of iterations, then requests another chunk, |
| 3457 | // until no chunks remain to be distributed. Each chunk contains chunk_size |
| 3458 | // iterations, except for the last chunk to be distributed, which may have |
| 3459 | // fewer iterations. When no chunk_size is specified, it defaults to 1. |
| 3460 | // |
| 3461 | // When schedule(guided,chunk_size) is specified, the iterations are assigned |
| 3462 | // to threads in the team in chunks as the executing threads request them. |
| 3463 | // Each thread executes a chunk of iterations, then requests another chunk, |
| 3464 | // until no chunks remain to be assigned. For a chunk_size of 1, the size of |
| 3465 | // each chunk is proportional to the number of unassigned iterations divided |
| 3466 | // by the number of threads in the team, decreasing to 1. For a chunk_size |
| 3467 | // with value k (greater than 1), the size of each chunk is determined in the |
| 3468 | // same way, with the restriction that the chunks do not contain fewer than k |
| 3469 | // iterations (except for the last chunk to be assigned, which may have fewer |
| 3470 | // than k iterations). |
| 3471 | // |
| 3472 | // When schedule(auto) is specified, the decision regarding scheduling is |
| 3473 | // delegated to the compiler and/or runtime system. The programmer gives the |
| 3474 | // implementation the freedom to choose any possible mapping of iterations to |
| 3475 | // threads in the team. |
| 3476 | // |
| 3477 | // When schedule(runtime) is specified, the decision regarding scheduling is |
| 3478 | // deferred until run time, and the schedule and chunk size are taken from the |
| 3479 | // run-sched-var ICV. If the ICV is set to auto, the schedule is |
| 3480 | // implementation defined |
| 3481 | // |
| 3482 | // __kmpc_dispatch_init(); |
| 3483 | // while(__kmpc_dispatch_next(&LB, &UB)) { |
| 3484 | // idx = LB; |
| 3485 | // while (idx <= UB) { BODY; ++idx; |
| 3486 | // __kmpc_dispatch_fini_(4|8)[u](); // For ordered loops only. |
| 3487 | // } // inner loop |
| 3488 | // } |
| 3489 | // __kmpc_dispatch_deinit(); |
| 3490 | // |
| 3491 | // OpenMP [2.7.1, Loop Construct, Description, table 2-1] |
| 3492 | // When schedule(static, chunk_size) is specified, iterations are divided into |
| 3493 | // chunks of size chunk_size, and the chunks are assigned to the threads in |
| 3494 | // the team in a round-robin fashion in the order of the thread number. |
| 3495 | // |
| 3496 | // while(UB = min(UB, GlobalUB), idx = LB, idx < UB) { |
| 3497 | // while (idx <= UB) { BODY; ++idx; } // inner loop |
| 3498 | // LB = LB + ST; |
| 3499 | // UB = UB + ST; |
| 3500 | // } |
| 3501 | // |
| 3502 | |
| 3503 | const Expr *IVExpr = S.getIterationVariable(); |
| 3504 | const unsigned IVSize = getContext().getTypeSize(T: IVExpr->getType()); |
| 3505 | const bool IVSigned = IVExpr->getType()->hasSignedIntegerRepresentation(); |
| 3506 | |
| 3507 | if (DynamicOrOrdered) { |
| 3508 | const std::pair<llvm::Value *, llvm::Value *> DispatchBounds = |
| 3509 | CGDispatchBounds(*this, S, LoopArgs.LB, LoopArgs.UB); |
| 3510 | llvm::Value *LBVal = DispatchBounds.first; |
| 3511 | llvm::Value *UBVal = DispatchBounds.second; |
| 3512 | CGOpenMPRuntime::DispatchRTInput DipatchRTInputValues = {LBVal, UBVal, |
| 3513 | LoopArgs.Chunk}; |
| 3514 | RT.emitForDispatchInit(CGF&: *this, Loc: S.getBeginLoc(), ScheduleKind, IVSize, |
| 3515 | IVSigned, Ordered, DispatchValues: DipatchRTInputValues); |
| 3516 | } else { |
| 3517 | CGOpenMPRuntime::StaticRTInput StaticInit( |
| 3518 | IVSize, IVSigned, Ordered, LoopArgs.IL, LoopArgs.LB, LoopArgs.UB, |
| 3519 | LoopArgs.ST, LoopArgs.Chunk); |
| 3520 | OpenMPDirectiveKind EKind = getEffectiveDirectiveKind(S); |
| 3521 | RT.emitForStaticInit(CGF&: *this, Loc: S.getBeginLoc(), DKind: EKind, ScheduleKind, |
| 3522 | Values: StaticInit); |
| 3523 | } |
| 3524 | |
| 3525 | auto &&CodeGenOrdered = [Ordered](CodeGenFunction &CGF, SourceLocation Loc, |
| 3526 | const unsigned IVSize, |
| 3527 | const bool IVSigned) { |
| 3528 | if (Ordered) { |
| 3529 | CGF.CGM.getOpenMPRuntime().emitForOrderedIterationEnd(CGF, Loc, IVSize, |
| 3530 | IVSigned); |
| 3531 | } |
| 3532 | }; |
| 3533 | |
| 3534 | OMPLoopArguments OuterLoopArgs(LoopArgs.LB, LoopArgs.UB, LoopArgs.ST, |
| 3535 | LoopArgs.IL, LoopArgs.Chunk, LoopArgs.EUB); |
| 3536 | OuterLoopArgs.IncExpr = S.getInc(); |
| 3537 | OuterLoopArgs.Init = S.getInit(); |
| 3538 | OuterLoopArgs.Cond = S.getCond(); |
| 3539 | OuterLoopArgs.NextLB = S.getNextLowerBound(); |
| 3540 | OuterLoopArgs.NextUB = S.getNextUpperBound(); |
| 3541 | OuterLoopArgs.DKind = LoopArgs.DKind; |
| 3542 | EmitOMPOuterLoop(DynamicOrOrdered, IsMonotonic, S, LoopScope, LoopArgs: OuterLoopArgs, |
| 3543 | CodeGenLoop: emitOMPLoopBodyWithStopPoint, CodeGenOrdered); |
| 3544 | if (DynamicOrOrdered) { |
| 3545 | RT.emitForDispatchDeinit(CGF&: *this, Loc: S.getBeginLoc()); |
| 3546 | } |
| 3547 | } |
| 3548 | |
| 3549 | static void emitEmptyOrdered(CodeGenFunction &, SourceLocation Loc, |
| 3550 | const unsigned IVSize, const bool IVSigned) {} |
| 3551 | |
| 3552 | void CodeGenFunction::EmitOMPDistributeOuterLoop( |
| 3553 | OpenMPDistScheduleClauseKind ScheduleKind, const OMPLoopDirective &S, |
| 3554 | OMPPrivateScope &LoopScope, const OMPLoopArguments &LoopArgs, |
| 3555 | const CodeGenLoopTy &CodeGenLoopContent) { |
| 3556 | |
| 3557 | CGOpenMPRuntime &RT = CGM.getOpenMPRuntime(); |
| 3558 | |
| 3559 | // Emit outer loop. |
| 3560 | // Same behavior as a OMPForOuterLoop, except that schedule cannot be |
| 3561 | // dynamic |
| 3562 | // |
| 3563 | |
| 3564 | const Expr *IVExpr = S.getIterationVariable(); |
| 3565 | const unsigned IVSize = getContext().getTypeSize(T: IVExpr->getType()); |
| 3566 | const bool IVSigned = IVExpr->getType()->hasSignedIntegerRepresentation(); |
| 3567 | OpenMPDirectiveKind EKind = getEffectiveDirectiveKind(S); |
| 3568 | |
| 3569 | CGOpenMPRuntime::StaticRTInput StaticInit( |
| 3570 | IVSize, IVSigned, /* Ordered = */ false, LoopArgs.IL, LoopArgs.LB, |
| 3571 | LoopArgs.UB, LoopArgs.ST, LoopArgs.Chunk); |
| 3572 | RT.emitDistributeStaticInit(CGF&: *this, Loc: S.getBeginLoc(), SchedKind: ScheduleKind, Values: StaticInit); |
| 3573 | |
| 3574 | // for combined 'distribute' and 'for' the increment expression of distribute |
| 3575 | // is stored in DistInc. For 'distribute' alone, it is in Inc. |
| 3576 | Expr *IncExpr; |
| 3577 | if (isOpenMPLoopBoundSharingDirective(Kind: EKind)) |
| 3578 | IncExpr = S.getDistInc(); |
| 3579 | else |
| 3580 | IncExpr = S.getInc(); |
| 3581 | |
| 3582 | // this routine is shared by 'omp distribute parallel for' and |
| 3583 | // 'omp distribute': select the right EUB expression depending on the |
| 3584 | // directive |
| 3585 | OMPLoopArguments OuterLoopArgs; |
| 3586 | OuterLoopArgs.LB = LoopArgs.LB; |
| 3587 | OuterLoopArgs.UB = LoopArgs.UB; |
| 3588 | OuterLoopArgs.ST = LoopArgs.ST; |
| 3589 | OuterLoopArgs.IL = LoopArgs.IL; |
| 3590 | OuterLoopArgs.Chunk = LoopArgs.Chunk; |
| 3591 | OuterLoopArgs.EUB = isOpenMPLoopBoundSharingDirective(Kind: EKind) |
| 3592 | ? S.getCombinedEnsureUpperBound() |
| 3593 | : S.getEnsureUpperBound(); |
| 3594 | OuterLoopArgs.IncExpr = IncExpr; |
| 3595 | OuterLoopArgs.Init = isOpenMPLoopBoundSharingDirective(Kind: EKind) |
| 3596 | ? S.getCombinedInit() |
| 3597 | : S.getInit(); |
| 3598 | OuterLoopArgs.Cond = isOpenMPLoopBoundSharingDirective(Kind: EKind) |
| 3599 | ? S.getCombinedCond() |
| 3600 | : S.getCond(); |
| 3601 | OuterLoopArgs.NextLB = isOpenMPLoopBoundSharingDirective(Kind: EKind) |
| 3602 | ? S.getCombinedNextLowerBound() |
| 3603 | : S.getNextLowerBound(); |
| 3604 | OuterLoopArgs.NextUB = isOpenMPLoopBoundSharingDirective(Kind: EKind) |
| 3605 | ? S.getCombinedNextUpperBound() |
| 3606 | : S.getNextUpperBound(); |
| 3607 | OuterLoopArgs.DKind = OMPD_distribute; |
| 3608 | |
| 3609 | EmitOMPOuterLoop(/* DynamicOrOrdered = */ false, /* IsMonotonic = */ false, S, |
| 3610 | LoopScope, LoopArgs: OuterLoopArgs, CodeGenLoop: CodeGenLoopContent, |
| 3611 | CodeGenOrdered: emitEmptyOrdered); |
| 3612 | } |
| 3613 | |
| 3614 | static std::pair<LValue, LValue> |
| 3615 | emitDistributeParallelForInnerBounds(CodeGenFunction &CGF, |
| 3616 | const OMPExecutableDirective &S) { |
| 3617 | const OMPLoopDirective &LS = cast<OMPLoopDirective>(Val: S); |
| 3618 | LValue LB = |
| 3619 | EmitOMPHelperVar(CGF, Helper: cast<DeclRefExpr>(Val: LS.getLowerBoundVariable())); |
| 3620 | LValue UB = |
| 3621 | EmitOMPHelperVar(CGF, Helper: cast<DeclRefExpr>(Val: LS.getUpperBoundVariable())); |
| 3622 | |
| 3623 | // When composing 'distribute' with 'for' (e.g. as in 'distribute |
| 3624 | // parallel for') we need to use the 'distribute' |
| 3625 | // chunk lower and upper bounds rather than the whole loop iteration |
| 3626 | // space. These are parameters to the outlined function for 'parallel' |
| 3627 | // and we copy the bounds of the previous schedule into the |
| 3628 | // the current ones. |
| 3629 | LValue PrevLB = CGF.EmitLValue(E: LS.getPrevLowerBoundVariable()); |
| 3630 | LValue PrevUB = CGF.EmitLValue(E: LS.getPrevUpperBoundVariable()); |
| 3631 | llvm::Value *PrevLBVal = CGF.EmitLoadOfScalar( |
| 3632 | lvalue: PrevLB, Loc: LS.getPrevLowerBoundVariable()->getExprLoc()); |
| 3633 | PrevLBVal = CGF.EmitScalarConversion( |
| 3634 | Src: PrevLBVal, SrcTy: LS.getPrevLowerBoundVariable()->getType(), |
| 3635 | DstTy: LS.getIterationVariable()->getType(), |
| 3636 | Loc: LS.getPrevLowerBoundVariable()->getExprLoc()); |
| 3637 | llvm::Value *PrevUBVal = CGF.EmitLoadOfScalar( |
| 3638 | lvalue: PrevUB, Loc: LS.getPrevUpperBoundVariable()->getExprLoc()); |
| 3639 | PrevUBVal = CGF.EmitScalarConversion( |
| 3640 | Src: PrevUBVal, SrcTy: LS.getPrevUpperBoundVariable()->getType(), |
| 3641 | DstTy: LS.getIterationVariable()->getType(), |
| 3642 | Loc: LS.getPrevUpperBoundVariable()->getExprLoc()); |
| 3643 | |
| 3644 | CGF.EmitStoreOfScalar(value: PrevLBVal, lvalue: LB); |
| 3645 | CGF.EmitStoreOfScalar(value: PrevUBVal, lvalue: UB); |
| 3646 | |
| 3647 | return {LB, UB}; |
| 3648 | } |
| 3649 | |
| 3650 | /// if the 'for' loop has a dispatch schedule (e.g. dynamic, guided) then |
| 3651 | /// we need to use the LB and UB expressions generated by the worksharing |
| 3652 | /// code generation support, whereas in non combined situations we would |
| 3653 | /// just emit 0 and the LastIteration expression |
| 3654 | /// This function is necessary due to the difference of the LB and UB |
| 3655 | /// types for the RT emission routines for 'for_static_init' and |
| 3656 | /// 'for_dispatch_init' |
| 3657 | static std::pair<llvm::Value *, llvm::Value *> |
| 3658 | emitDistributeParallelForDispatchBounds(CodeGenFunction &CGF, |
| 3659 | const OMPExecutableDirective &S, |
| 3660 | Address LB, Address UB) { |
| 3661 | const OMPLoopDirective &LS = cast<OMPLoopDirective>(Val: S); |
| 3662 | const Expr *IVExpr = LS.getIterationVariable(); |
| 3663 | // when implementing a dynamic schedule for a 'for' combined with a |
| 3664 | // 'distribute' (e.g. 'distribute parallel for'), the 'for' loop |
| 3665 | // is not normalized as each team only executes its own assigned |
| 3666 | // distribute chunk |
| 3667 | QualType IteratorTy = IVExpr->getType(); |
| 3668 | llvm::Value *LBVal = |
| 3669 | CGF.EmitLoadOfScalar(Addr: LB, /*Volatile=*/false, Ty: IteratorTy, Loc: S.getBeginLoc()); |
| 3670 | llvm::Value *UBVal = |
| 3671 | CGF.EmitLoadOfScalar(Addr: UB, /*Volatile=*/false, Ty: IteratorTy, Loc: S.getBeginLoc()); |
| 3672 | return {LBVal, UBVal}; |
| 3673 | } |
| 3674 | |
| 3675 | static void emitDistributeParallelForDistributeInnerBoundParams( |
| 3676 | CodeGenFunction &CGF, const OMPExecutableDirective &S, |
| 3677 | llvm::SmallVectorImpl<llvm::Value *> &CapturedVars) { |
| 3678 | const auto &Dir = cast<OMPLoopDirective>(Val: S); |
| 3679 | LValue LB = |
| 3680 | CGF.EmitLValue(E: cast<DeclRefExpr>(Val: Dir.getCombinedLowerBoundVariable())); |
| 3681 | llvm::Value *LBCast = CGF.Builder.CreateIntCast( |
| 3682 | V: CGF.Builder.CreateLoad(Addr: LB.getAddress()), DestTy: CGF.SizeTy, /*isSigned=*/false); |
| 3683 | CapturedVars.push_back(Elt: LBCast); |
| 3684 | LValue UB = |
| 3685 | CGF.EmitLValue(E: cast<DeclRefExpr>(Val: Dir.getCombinedUpperBoundVariable())); |
| 3686 | |
| 3687 | llvm::Value *UBCast = CGF.Builder.CreateIntCast( |
| 3688 | V: CGF.Builder.CreateLoad(Addr: UB.getAddress()), DestTy: CGF.SizeTy, /*isSigned=*/false); |
| 3689 | CapturedVars.push_back(Elt: UBCast); |
| 3690 | } |
| 3691 | |
| 3692 | static void |
| 3693 | emitInnerParallelForWhenCombined(CodeGenFunction &CGF, |
| 3694 | const OMPLoopDirective &S, |
| 3695 | CodeGenFunction::JumpDest LoopExit) { |
| 3696 | OpenMPDirectiveKind EKind = getEffectiveDirectiveKind(S); |
| 3697 | auto &&CGInlinedWorksharingLoop = [&S, EKind](CodeGenFunction &CGF, |
| 3698 | PrePostActionTy &Action) { |
| 3699 | Action.Enter(CGF); |
| 3700 | bool HasCancel = false; |
| 3701 | if (!isOpenMPSimdDirective(DKind: EKind)) { |
| 3702 | if (const auto *D = dyn_cast<OMPTeamsDistributeParallelForDirective>(Val: &S)) |
| 3703 | HasCancel = D->hasCancel(); |
| 3704 | else if (const auto *D = dyn_cast<OMPDistributeParallelForDirective>(Val: &S)) |
| 3705 | HasCancel = D->hasCancel(); |
| 3706 | else if (const auto *D = |
| 3707 | dyn_cast<OMPTargetTeamsDistributeParallelForDirective>(Val: &S)) |
| 3708 | HasCancel = D->hasCancel(); |
| 3709 | } |
| 3710 | CodeGenFunction::OMPCancelStackRAII CancelRegion(CGF, EKind, HasCancel); |
| 3711 | CGF.EmitOMPWorksharingLoop(S, EUB: S.getPrevEnsureUpperBound(), |
| 3712 | CodeGenLoopBounds: emitDistributeParallelForInnerBounds, |
| 3713 | CGDispatchBounds: emitDistributeParallelForDispatchBounds); |
| 3714 | }; |
| 3715 | |
| 3716 | emitCommonOMPParallelDirective( |
| 3717 | CGF, S, InnermostKind: isOpenMPSimdDirective(DKind: EKind) ? OMPD_for_simd : OMPD_for, |
| 3718 | CodeGen: CGInlinedWorksharingLoop, |
| 3719 | CodeGenBoundParameters: emitDistributeParallelForDistributeInnerBoundParams); |
| 3720 | } |
| 3721 | |
| 3722 | void CodeGenFunction::EmitOMPDistributeParallelForDirective( |
| 3723 | const OMPDistributeParallelForDirective &S) { |
| 3724 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &) { |
| 3725 | CGF.EmitOMPDistributeLoop(S, CodeGenLoop: emitInnerParallelForWhenCombined, |
| 3726 | IncExpr: S.getDistInc()); |
| 3727 | }; |
| 3728 | OMPLexicalScope Scope(*this, S, OMPD_parallel); |
| 3729 | CGM.getOpenMPRuntime().emitInlinedDirective(CGF&: *this, InnermostKind: OMPD_distribute, CodeGen); |
| 3730 | } |
| 3731 | |
| 3732 | void CodeGenFunction::EmitOMPDistributeParallelForSimdDirective( |
| 3733 | const OMPDistributeParallelForSimdDirective &S) { |
| 3734 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &) { |
| 3735 | CGF.EmitOMPDistributeLoop(S, CodeGenLoop: emitInnerParallelForWhenCombined, |
| 3736 | IncExpr: S.getDistInc()); |
| 3737 | }; |
| 3738 | OMPLexicalScope Scope(*this, S, OMPD_parallel); |
| 3739 | CGM.getOpenMPRuntime().emitInlinedDirective(CGF&: *this, InnermostKind: OMPD_distribute, CodeGen); |
| 3740 | } |
| 3741 | |
| 3742 | void CodeGenFunction::EmitOMPDistributeSimdDirective( |
| 3743 | const OMPDistributeSimdDirective &S) { |
| 3744 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &) { |
| 3745 | CGF.EmitOMPDistributeLoop(S, CodeGenLoop: emitOMPLoopBodyWithStopPoint, IncExpr: S.getInc()); |
| 3746 | }; |
| 3747 | OMPLexicalScope Scope(*this, S, OMPD_unknown); |
| 3748 | CGM.getOpenMPRuntime().emitInlinedDirective(CGF&: *this, InnermostKind: OMPD_simd, CodeGen); |
| 3749 | } |
| 3750 | |
| 3751 | void CodeGenFunction::EmitOMPTargetSimdDeviceFunction( |
| 3752 | CodeGenModule &CGM, StringRef ParentName, const OMPTargetSimdDirective &S) { |
| 3753 | // Emit SPMD target parallel for region as a standalone region. |
| 3754 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 3755 | emitOMPSimdRegion(CGF, S, Action); |
| 3756 | }; |
| 3757 | llvm::Function *Fn; |
| 3758 | llvm::Constant *Addr; |
| 3759 | // Emit target region as a standalone region. |
| 3760 | CGM.getOpenMPRuntime().emitTargetOutlinedFunction( |
| 3761 | D: S, ParentName, OutlinedFn&: Fn, OutlinedFnID&: Addr, /*IsOffloadEntry=*/true, CodeGen); |
| 3762 | assert(Fn && Addr && "Target device function emission failed."); |
| 3763 | } |
| 3764 | |
| 3765 | void CodeGenFunction::EmitOMPTargetSimdDirective( |
| 3766 | const OMPTargetSimdDirective &S) { |
| 3767 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 3768 | emitOMPSimdRegion(CGF, S, Action); |
| 3769 | }; |
| 3770 | emitCommonOMPTargetDirective(CGF&: *this, S, CodeGen); |
| 3771 | } |
| 3772 | |
| 3773 | namespace { |
| 3774 | struct ScheduleKindModifiersTy { |
| 3775 | OpenMPScheduleClauseKind Kind; |
| 3776 | OpenMPScheduleClauseModifier M1; |
| 3777 | OpenMPScheduleClauseModifier M2; |
| 3778 | ScheduleKindModifiersTy(OpenMPScheduleClauseKind Kind, |
| 3779 | OpenMPScheduleClauseModifier M1, |
| 3780 | OpenMPScheduleClauseModifier M2) |
| 3781 | : Kind(Kind), M1(M1), M2(M2) {} |
| 3782 | }; |
| 3783 | } // namespace |
| 3784 | |
| 3785 | bool CodeGenFunction::EmitOMPWorksharingLoop( |
| 3786 | const OMPLoopDirective &S, Expr *EUB, |
| 3787 | const CodeGenLoopBoundsTy &CodeGenLoopBounds, |
| 3788 | const CodeGenDispatchBoundsTy &CGDispatchBounds) { |
| 3789 | // Emit the loop iteration variable. |
| 3790 | const auto *IVExpr = cast<DeclRefExpr>(Val: S.getIterationVariable()); |
| 3791 | const auto *IVDecl = cast<VarDecl>(Val: IVExpr->getDecl()); |
| 3792 | EmitVarDecl(D: *IVDecl); |
| 3793 | |
| 3794 | // Emit the iterations count variable. |
| 3795 | // If it is not a variable, Sema decided to calculate iterations count on each |
| 3796 | // iteration (e.g., it is foldable into a constant). |
| 3797 | if (const auto *LIExpr = dyn_cast<DeclRefExpr>(Val: S.getLastIteration())) { |
| 3798 | EmitVarDecl(D: *cast<VarDecl>(Val: LIExpr->getDecl())); |
| 3799 | // Emit calculation of the iterations count. |
| 3800 | EmitIgnoredExpr(E: S.getCalcLastIteration()); |
| 3801 | } |
| 3802 | |
| 3803 | CGOpenMPRuntime &RT = CGM.getOpenMPRuntime(); |
| 3804 | |
| 3805 | bool HasLastprivateClause; |
| 3806 | // Check pre-condition. |
| 3807 | { |
| 3808 | OMPLoopScope PreInitScope(*this, S); |
| 3809 | // Skip the entire loop if we don't meet the precondition. |
| 3810 | // If the condition constant folds and can be elided, avoid emitting the |
| 3811 | // whole loop. |
| 3812 | bool CondConstant; |
| 3813 | llvm::BasicBlock *ContBlock = nullptr; |
| 3814 | if (ConstantFoldsToSimpleInteger(Cond: S.getPreCond(), Result&: CondConstant)) { |
| 3815 | if (!CondConstant) |
| 3816 | return false; |
| 3817 | } else { |
| 3818 | llvm::BasicBlock *ThenBlock = createBasicBlock(name: "omp.precond.then"); |
| 3819 | ContBlock = createBasicBlock(name: "omp.precond.end"); |
| 3820 | emitPreCond(CGF&: *this, S, Cond: S.getPreCond(), TrueBlock: ThenBlock, FalseBlock: ContBlock, |
| 3821 | TrueCount: getProfileCount(S: &S)); |
| 3822 | EmitBlock(BB: ThenBlock); |
| 3823 | incrementProfileCounter(S: &S); |
| 3824 | } |
| 3825 | |
| 3826 | RunCleanupsScope DoacrossCleanupScope(*this); |
| 3827 | bool Ordered = false; |
| 3828 | if (const auto *OrderedClause = S.getSingleClause<OMPOrderedClause>()) { |
| 3829 | if (OrderedClause->getNumForLoops()) |
| 3830 | RT.emitDoacrossInit(CGF&: *this, D: S, NumIterations: OrderedClause->getLoopNumIterations()); |
| 3831 | else |
| 3832 | Ordered = true; |
| 3833 | } |
| 3834 | |
| 3835 | emitAlignedClause(CGF&: *this, D: S); |
| 3836 | bool HasLinears = EmitOMPLinearClauseInit(D: S); |
| 3837 | // Emit helper vars inits. |
| 3838 | |
| 3839 | std::pair<LValue, LValue> Bounds = CodeGenLoopBounds(*this, S); |
| 3840 | LValue LB = Bounds.first; |
| 3841 | LValue UB = Bounds.second; |
| 3842 | LValue ST = |
| 3843 | EmitOMPHelperVar(CGF&: *this, Helper: cast<DeclRefExpr>(Val: S.getStrideVariable())); |
| 3844 | LValue IL = |
| 3845 | EmitOMPHelperVar(CGF&: *this, Helper: cast<DeclRefExpr>(Val: S.getIsLastIterVariable())); |
| 3846 | |
| 3847 | // Emit 'then' code. |
| 3848 | { |
| 3849 | OpenMPDirectiveKind EKind = getEffectiveDirectiveKind(S); |
| 3850 | OMPPrivateScope LoopScope(*this); |
| 3851 | if (EmitOMPFirstprivateClause(D: S, PrivateScope&: LoopScope) || HasLinears) { |
| 3852 | // Emit implicit barrier to synchronize threads and avoid data races on |
| 3853 | // initialization of firstprivate variables and post-update of |
| 3854 | // lastprivate variables. |
| 3855 | CGM.getOpenMPRuntime().emitBarrierCall( |
| 3856 | CGF&: *this, Loc: S.getBeginLoc(), Kind: OMPD_unknown, /*EmitChecks=*/false, |
| 3857 | /*ForceSimpleCall=*/true); |
| 3858 | } |
| 3859 | EmitOMPPrivateClause(D: S, PrivateScope&: LoopScope); |
| 3860 | CGOpenMPRuntime::LastprivateConditionalRAII LPCRegion( |
| 3861 | *this, S, EmitLValue(E: S.getIterationVariable())); |
| 3862 | HasLastprivateClause = EmitOMPLastprivateClauseInit(D: S, PrivateScope&: LoopScope); |
| 3863 | EmitOMPReductionClauseInit(D: S, PrivateScope&: LoopScope); |
| 3864 | EmitOMPPrivateLoopCounters(S, LoopScope); |
| 3865 | EmitOMPLinearClause(D: S, PrivateScope&: LoopScope); |
| 3866 | (void)LoopScope.Privatize(); |
| 3867 | if (isOpenMPTargetExecutionDirective(DKind: EKind)) |
| 3868 | CGM.getOpenMPRuntime().adjustTargetSpecificDataForLambdas(CGF&: *this, D: S); |
| 3869 | |
| 3870 | // Detect the loop schedule kind and chunk. |
| 3871 | const Expr *ChunkExpr = nullptr; |
| 3872 | OpenMPScheduleTy ScheduleKind; |
| 3873 | if (const auto *C = S.getSingleClause<OMPScheduleClause>()) { |
| 3874 | ScheduleKind.Schedule = C->getScheduleKind(); |
| 3875 | ScheduleKind.M1 = C->getFirstScheduleModifier(); |
| 3876 | ScheduleKind.M2 = C->getSecondScheduleModifier(); |
| 3877 | ChunkExpr = C->getChunkSize(); |
| 3878 | } else { |
| 3879 | // Default behaviour for schedule clause. |
| 3880 | CGM.getOpenMPRuntime().getDefaultScheduleAndChunk( |
| 3881 | CGF&: *this, S, ScheduleKind&: ScheduleKind.Schedule, ChunkExpr); |
| 3882 | } |
| 3883 | bool HasChunkSizeOne = false; |
| 3884 | llvm::Value *Chunk = nullptr; |
| 3885 | if (ChunkExpr) { |
| 3886 | Chunk = EmitScalarExpr(E: ChunkExpr); |
| 3887 | Chunk = EmitScalarConversion(Src: Chunk, SrcTy: ChunkExpr->getType(), |
| 3888 | DstTy: S.getIterationVariable()->getType(), |
| 3889 | Loc: S.getBeginLoc()); |
| 3890 | Expr::EvalResult Result; |
| 3891 | if (ChunkExpr->EvaluateAsInt(Result, Ctx: getContext())) { |
| 3892 | llvm::APSInt EvaluatedChunk = Result.Val.getInt(); |
| 3893 | HasChunkSizeOne = (EvaluatedChunk.getLimitedValue() == 1); |
| 3894 | } |
| 3895 | } |
| 3896 | const unsigned IVSize = getContext().getTypeSize(T: IVExpr->getType()); |
| 3897 | const bool IVSigned = IVExpr->getType()->hasSignedIntegerRepresentation(); |
| 3898 | // OpenMP 4.5, 2.7.1 Loop Construct, Description. |
| 3899 | // If the static schedule kind is specified or if the ordered clause is |
| 3900 | // specified, and if no monotonic modifier is specified, the effect will |
| 3901 | // be as if the monotonic modifier was specified. |
| 3902 | bool StaticChunkedOne = |
| 3903 | RT.isStaticChunked(ScheduleKind: ScheduleKind.Schedule, |
| 3904 | /* Chunked */ Chunk != nullptr) && |
| 3905 | HasChunkSizeOne && isOpenMPLoopBoundSharingDirective(Kind: EKind); |
| 3906 | // GPU combined `distribute parallel for`: emit a single |
| 3907 | // for_static_init with the fused distr_static_chunk + static_chunkone |
| 3908 | // schedule (enum 93). The surrounding EmitOMPDistributeLoop must skip |
| 3909 | // its distribute_static_init under the same conditions. Both sites are |
| 3910 | // guarded by canEmitGPUFusedDistSchedule() alone so they cannot |
| 3911 | // disagree; the assert guards the invariant that makes this safe today, |
| 3912 | // aka that the implicit GPU default schedule is always static chunk-one. |
| 3913 | ScheduleKind.UseFusedDistChunkSchedule = |
| 3914 | canEmitGPUFusedDistSchedule(CGM, S, DKind: EKind); |
| 3915 | assert((!ScheduleKind.UseFusedDistChunkSchedule || StaticChunkedOne) && |
| 3916 | "fused distribute schedule requires a static chunk-one schedule"); |
| 3917 | bool IsMonotonic = |
| 3918 | Ordered || |
| 3919 | (ScheduleKind.Schedule == OMPC_SCHEDULE_static && |
| 3920 | !(ScheduleKind.M1 == OMPC_SCHEDULE_MODIFIER_nonmonotonic || |
| 3921 | ScheduleKind.M2 == OMPC_SCHEDULE_MODIFIER_nonmonotonic)) || |
| 3922 | ScheduleKind.M1 == OMPC_SCHEDULE_MODIFIER_monotonic || |
| 3923 | ScheduleKind.M2 == OMPC_SCHEDULE_MODIFIER_monotonic; |
| 3924 | if ((RT.isStaticNonchunked(ScheduleKind: ScheduleKind.Schedule, |
| 3925 | /* Chunked */ Chunk != nullptr) || |
| 3926 | StaticChunkedOne) && |
| 3927 | !Ordered) { |
| 3928 | JumpDest LoopExit = |
| 3929 | getJumpDestInCurrentScope(Target: createBasicBlock(name: "omp.loop.exit")); |
| 3930 | emitCommonSimdLoop( |
| 3931 | CGF&: *this, S, |
| 3932 | SimdInitGen: [&S, EKind](CodeGenFunction &CGF, PrePostActionTy &) { |
| 3933 | if (isOpenMPSimdDirective(DKind: EKind)) { |
| 3934 | CGF.EmitOMPSimdInit(D: S); |
| 3935 | } else if (const auto *C = S.getSingleClause<OMPOrderClause>()) { |
| 3936 | if (C->getKind() == OMPC_ORDER_concurrent) |
| 3937 | CGF.LoopStack.setParallel(/*Enable=*/true); |
| 3938 | } |
| 3939 | }, |
| 3940 | BodyCodeGen: [IVSize, IVSigned, Ordered, IL, LB, UB, ST, StaticChunkedOne, Chunk, |
| 3941 | &S, ScheduleKind, LoopExit, EKind, |
| 3942 | &LoopScope](CodeGenFunction &CGF, PrePostActionTy &) { |
| 3943 | // OpenMP [2.7.1, Loop Construct, Description, table 2-1] |
| 3944 | // When no chunk_size is specified, the iteration space is divided |
| 3945 | // into chunks that are approximately equal in size, and at most |
| 3946 | // one chunk is distributed to each thread. Note that the size of |
| 3947 | // the chunks is unspecified in this case. |
| 3948 | CGOpenMPRuntime::StaticRTInput StaticInit( |
| 3949 | IVSize, IVSigned, Ordered, IL.getAddress(), LB.getAddress(), |
| 3950 | UB.getAddress(), ST.getAddress(), |
| 3951 | StaticChunkedOne ? Chunk : nullptr); |
| 3952 | CGF.CGM.getOpenMPRuntime().emitForStaticInit( |
| 3953 | CGF, Loc: S.getBeginLoc(), DKind: EKind, ScheduleKind, Values: StaticInit); |
| 3954 | // UB = min(UB, GlobalUB); |
| 3955 | if (!StaticChunkedOne) |
| 3956 | CGF.EmitIgnoredExpr(E: S.getEnsureUpperBound()); |
| 3957 | // IV = LB; |
| 3958 | CGF.EmitIgnoredExpr(E: S.getInit()); |
| 3959 | // For unchunked static schedule generate: |
| 3960 | // |
| 3961 | // while (idx <= UB) { |
| 3962 | // BODY; |
| 3963 | // ++idx; |
| 3964 | // } |
| 3965 | // |
| 3966 | // For static schedule with chunk one: |
| 3967 | // |
| 3968 | // while (IV <= PrevUB) { |
| 3969 | // BODY; |
| 3970 | // IV += ST; |
| 3971 | // } |
| 3972 | CGF.EmitOMPInnerLoop( |
| 3973 | S, RequiresCleanup: LoopScope.requiresCleanups(), |
| 3974 | LoopCond: StaticChunkedOne ? S.getCombinedParForInDistCond() |
| 3975 | : S.getCond(), |
| 3976 | IncExpr: StaticChunkedOne ? S.getDistInc() : S.getInc(), |
| 3977 | BodyGen: [&S, LoopExit](CodeGenFunction &CGF) { |
| 3978 | emitOMPLoopBodyWithStopPoint(CGF, S, LoopExit); |
| 3979 | }, |
| 3980 | PostIncGen: [](CodeGenFunction &) {}); |
| 3981 | }); |
| 3982 | EmitBlock(BB: LoopExit.getBlock()); |
| 3983 | // Tell the runtime we are done. |
| 3984 | auto &&CodeGen = [&S](CodeGenFunction &CGF) { |
| 3985 | CGF.CGM.getOpenMPRuntime().emitForStaticFinish(CGF, Loc: S.getEndLoc(), |
| 3986 | DKind: OMPD_for); |
| 3987 | }; |
| 3988 | OMPCancelStack.emitExit(CGF&: *this, Kind: EKind, CodeGen); |
| 3989 | } else { |
| 3990 | // Emit the outer loop, which requests its work chunk [LB..UB] from |
| 3991 | // runtime and runs the inner loop to process it. |
| 3992 | OMPLoopArguments LoopArguments(LB.getAddress(), UB.getAddress(), |
| 3993 | ST.getAddress(), IL.getAddress(), Chunk, |
| 3994 | EUB); |
| 3995 | LoopArguments.DKind = OMPD_for; |
| 3996 | EmitOMPForOuterLoop(ScheduleKind, IsMonotonic, S, LoopScope, Ordered, |
| 3997 | LoopArgs: LoopArguments, CGDispatchBounds); |
| 3998 | } |
| 3999 | if (isOpenMPSimdDirective(DKind: EKind)) { |
| 4000 | EmitOMPSimdFinal(D: S, CondGen: [IL, &S](CodeGenFunction &CGF) { |
| 4001 | return CGF.Builder.CreateIsNotNull( |
| 4002 | Arg: CGF.EmitLoadOfScalar(lvalue: IL, Loc: S.getBeginLoc())); |
| 4003 | }); |
| 4004 | } |
| 4005 | EmitOMPReductionClauseFinal( |
| 4006 | D: S, /*ReductionKind=*/isOpenMPSimdDirective(DKind: EKind) |
| 4007 | ? /*Parallel and Simd*/ OMPD_parallel_for_simd |
| 4008 | : /*Parallel only*/ OMPD_parallel); |
| 4009 | // Emit post-update of the reduction variables if IsLastIter != 0. |
| 4010 | emitPostUpdateForReductionClause( |
| 4011 | CGF&: *this, D: S, CondGen: [IL, &S](CodeGenFunction &CGF) { |
| 4012 | return CGF.Builder.CreateIsNotNull( |
| 4013 | Arg: CGF.EmitLoadOfScalar(lvalue: IL, Loc: S.getBeginLoc())); |
| 4014 | }); |
| 4015 | // Emit final copy of the lastprivate variables if IsLastIter != 0. |
| 4016 | if (HasLastprivateClause) |
| 4017 | EmitOMPLastprivateClauseFinal( |
| 4018 | D: S, NoFinals: isOpenMPSimdDirective(DKind: EKind), |
| 4019 | IsLastIterCond: Builder.CreateIsNotNull(Arg: EmitLoadOfScalar(lvalue: IL, Loc: S.getBeginLoc()))); |
| 4020 | LoopScope.restoreMap(); |
| 4021 | EmitOMPLinearClauseFinal(D: S, CondGen: [IL, &S](CodeGenFunction &CGF) { |
| 4022 | return CGF.Builder.CreateIsNotNull( |
| 4023 | Arg: CGF.EmitLoadOfScalar(lvalue: IL, Loc: S.getBeginLoc())); |
| 4024 | }); |
| 4025 | } |
| 4026 | DoacrossCleanupScope.ForceCleanup(); |
| 4027 | // We're now done with the loop, so jump to the continuation block. |
| 4028 | if (ContBlock) { |
| 4029 | EmitBranch(Block: ContBlock); |
| 4030 | EmitBlock(BB: ContBlock, /*IsFinished=*/true); |
| 4031 | } |
| 4032 | } |
| 4033 | return HasLastprivateClause; |
| 4034 | } |
| 4035 | |
| 4036 | /// The following two functions generate expressions for the loop lower |
| 4037 | /// and upper bounds in case of static and dynamic (dispatch) schedule |
| 4038 | /// of the associated 'for' or 'distribute' loop. |
| 4039 | static std::pair<LValue, LValue> |
| 4040 | emitForLoopBounds(CodeGenFunction &CGF, const OMPExecutableDirective &S) { |
| 4041 | const auto &LS = cast<OMPLoopDirective>(Val: S); |
| 4042 | LValue LB = |
| 4043 | EmitOMPHelperVar(CGF, Helper: cast<DeclRefExpr>(Val: LS.getLowerBoundVariable())); |
| 4044 | LValue UB = |
| 4045 | EmitOMPHelperVar(CGF, Helper: cast<DeclRefExpr>(Val: LS.getUpperBoundVariable())); |
| 4046 | return {LB, UB}; |
| 4047 | } |
| 4048 | |
| 4049 | /// When dealing with dispatch schedules (e.g. dynamic, guided) we do not |
| 4050 | /// consider the lower and upper bound expressions generated by the |
| 4051 | /// worksharing loop support, but we use 0 and the iteration space size as |
| 4052 | /// constants |
| 4053 | static std::pair<llvm::Value *, llvm::Value *> |
| 4054 | emitDispatchForLoopBounds(CodeGenFunction &CGF, const OMPExecutableDirective &S, |
| 4055 | Address LB, Address UB) { |
| 4056 | const auto &LS = cast<OMPLoopDirective>(Val: S); |
| 4057 | const Expr *IVExpr = LS.getIterationVariable(); |
| 4058 | const unsigned IVSize = CGF.getContext().getTypeSize(T: IVExpr->getType()); |
| 4059 | llvm::Value *LBVal = CGF.Builder.getIntN(N: IVSize, C: 0); |
| 4060 | llvm::Value *UBVal = CGF.EmitScalarExpr(E: LS.getLastIteration()); |
| 4061 | return {LBVal, UBVal}; |
| 4062 | } |
| 4063 | |
| 4064 | /// Emits internal temp array declarations for the directive with inscan |
| 4065 | /// reductions. |
| 4066 | /// The code is the following: |
| 4067 | /// \code |
| 4068 | /// size num_iters = <num_iters>; |
| 4069 | /// <type> buffer[num_iters]; |
| 4070 | /// \endcode |
| 4071 | static void emitScanBasedDirectiveDecls( |
| 4072 | CodeGenFunction &CGF, const OMPLoopDirective &S, |
| 4073 | llvm::function_ref<llvm::Value *(CodeGenFunction &)> NumIteratorsGen) { |
| 4074 | llvm::Value *OMPScanNumIterations = CGF.Builder.CreateIntCast( |
| 4075 | V: NumIteratorsGen(CGF), DestTy: CGF.SizeTy, /*isSigned=*/false); |
| 4076 | SmallVector<const Expr *, 4> Shareds; |
| 4077 | SmallVector<const Expr *, 4> Privates; |
| 4078 | SmallVector<const Expr *, 4> ReductionOps; |
| 4079 | SmallVector<const Expr *, 4> CopyArrayTemps; |
| 4080 | for (const auto *C : S.getClausesOfKind<OMPReductionClause>()) { |
| 4081 | assert(C->getModifier() == OMPC_REDUCTION_inscan && |
| 4082 | "Only inscan reductions are expected."); |
| 4083 | Shareds.append(in_start: C->varlist_begin(), in_end: C->varlist_end()); |
| 4084 | Privates.append(in_start: C->privates().begin(), in_end: C->privates().end()); |
| 4085 | ReductionOps.append(in_start: C->reduction_ops().begin(), in_end: C->reduction_ops().end()); |
| 4086 | CopyArrayTemps.append(in_start: C->copy_array_temps().begin(), |
| 4087 | in_end: C->copy_array_temps().end()); |
| 4088 | } |
| 4089 | { |
| 4090 | // Emit buffers for each reduction variables. |
| 4091 | // ReductionCodeGen is required to emit correctly the code for array |
| 4092 | // reductions. |
| 4093 | ReductionCodeGen RedCG(Shareds, Shareds, Privates, ReductionOps); |
| 4094 | unsigned Count = 0; |
| 4095 | auto *ITA = CopyArrayTemps.begin(); |
| 4096 | for (const Expr *IRef : Privates) { |
| 4097 | const auto *PrivateVD = cast<VarDecl>(Val: cast<DeclRefExpr>(Val: IRef)->getDecl()); |
| 4098 | // Emit variably modified arrays, used for arrays/array sections |
| 4099 | // reductions. |
| 4100 | if (PrivateVD->getType()->isVariablyModifiedType()) { |
| 4101 | RedCG.emitSharedOrigLValue(CGF, N: Count); |
| 4102 | RedCG.emitAggregateType(CGF, N: Count); |
| 4103 | } |
| 4104 | CodeGenFunction::OpaqueValueMapping DimMapping( |
| 4105 | CGF, |
| 4106 | cast<OpaqueValueExpr>( |
| 4107 | Val: cast<VariableArrayType>(Val: (*ITA)->getType()->getAsArrayTypeUnsafe()) |
| 4108 | ->getSizeExpr()), |
| 4109 | RValue::get(V: OMPScanNumIterations)); |
| 4110 | // Emit temp buffer. |
| 4111 | CGF.EmitVarDecl(D: *cast<VarDecl>(Val: cast<DeclRefExpr>(Val: *ITA)->getDecl())); |
| 4112 | ++ITA; |
| 4113 | ++Count; |
| 4114 | } |
| 4115 | } |
| 4116 | } |
| 4117 | |
| 4118 | /// Copies final inscan reductions values to the original variables. |
| 4119 | /// The code is the following: |
| 4120 | /// \code |
| 4121 | /// <orig_var> = buffer[num_iters-1]; |
| 4122 | /// \endcode |
| 4123 | static void emitScanBasedDirectiveFinals( |
| 4124 | CodeGenFunction &CGF, const OMPLoopDirective &S, |
| 4125 | llvm::function_ref<llvm::Value *(CodeGenFunction &)> NumIteratorsGen) { |
| 4126 | llvm::Value *OMPScanNumIterations = CGF.Builder.CreateIntCast( |
| 4127 | V: NumIteratorsGen(CGF), DestTy: CGF.SizeTy, /*isSigned=*/false); |
| 4128 | SmallVector<const Expr *, 4> Shareds; |
| 4129 | SmallVector<const Expr *, 4> LHSs; |
| 4130 | SmallVector<const Expr *, 4> RHSs; |
| 4131 | SmallVector<const Expr *, 4> Privates; |
| 4132 | SmallVector<const Expr *, 4> CopyOps; |
| 4133 | SmallVector<const Expr *, 4> CopyArrayElems; |
| 4134 | for (const auto *C : S.getClausesOfKind<OMPReductionClause>()) { |
| 4135 | assert(C->getModifier() == OMPC_REDUCTION_inscan && |
| 4136 | "Only inscan reductions are expected."); |
| 4137 | Shareds.append(in_start: C->varlist_begin(), in_end: C->varlist_end()); |
| 4138 | LHSs.append(in_start: C->lhs_exprs().begin(), in_end: C->lhs_exprs().end()); |
| 4139 | RHSs.append(in_start: C->rhs_exprs().begin(), in_end: C->rhs_exprs().end()); |
| 4140 | Privates.append(in_start: C->privates().begin(), in_end: C->privates().end()); |
| 4141 | CopyOps.append(in_start: C->copy_ops().begin(), in_end: C->copy_ops().end()); |
| 4142 | CopyArrayElems.append(in_start: C->copy_array_elems().begin(), |
| 4143 | in_end: C->copy_array_elems().end()); |
| 4144 | } |
| 4145 | // Create temp var and copy LHS value to this temp value. |
| 4146 | // LHS = TMP[LastIter]; |
| 4147 | llvm::Value *OMPLast = CGF.Builder.CreateNSWSub( |
| 4148 | LHS: OMPScanNumIterations, |
| 4149 | RHS: llvm::ConstantInt::get(Ty: CGF.SizeTy, V: 1, /*isSigned=*/IsSigned: false)); |
| 4150 | for (unsigned I = 0, E = CopyArrayElems.size(); I < E; ++I) { |
| 4151 | const Expr *PrivateExpr = Privates[I]; |
| 4152 | const Expr *OrigExpr = Shareds[I]; |
| 4153 | const Expr *CopyArrayElem = CopyArrayElems[I]; |
| 4154 | CodeGenFunction::OpaqueValueMapping IdxMapping( |
| 4155 | CGF, |
| 4156 | cast<OpaqueValueExpr>( |
| 4157 | Val: cast<ArraySubscriptExpr>(Val: CopyArrayElem)->getIdx()), |
| 4158 | RValue::get(V: OMPLast)); |
| 4159 | LValue DestLVal = CGF.EmitLValue(E: OrigExpr); |
| 4160 | LValue SrcLVal = CGF.EmitLValue(E: CopyArrayElem); |
| 4161 | CGF.EmitOMPCopy( |
| 4162 | OriginalType: PrivateExpr->getType(), DestAddr: DestLVal.getAddress(), SrcAddr: SrcLVal.getAddress(), |
| 4163 | DestVD: cast<VarDecl>(Val: cast<DeclRefExpr>(Val: LHSs[I])->getDecl()), |
| 4164 | SrcVD: cast<VarDecl>(Val: cast<DeclRefExpr>(Val: RHSs[I])->getDecl()), Copy: CopyOps[I]); |
| 4165 | } |
| 4166 | } |
| 4167 | |
| 4168 | /// Emits the code for the directive with inscan reductions. |
| 4169 | /// The code is the following: |
| 4170 | /// \code |
| 4171 | /// #pragma omp ... |
| 4172 | /// for (i: 0..<num_iters>) { |
| 4173 | /// <input phase>; |
| 4174 | /// buffer[i] = red; |
| 4175 | /// } |
| 4176 | /// #pragma omp master // in parallel region |
| 4177 | /// for (int k = 0; k != ceil(log2(num_iters)); ++k) |
| 4178 | /// for (size cnt = last_iter; cnt >= pow(2, k); --k) |
| 4179 | /// buffer[i] op= buffer[i-pow(2,k)]; |
| 4180 | /// #pragma omp barrier // in parallel region |
| 4181 | /// #pragma omp ... |
| 4182 | /// for (0..<num_iters>) { |
| 4183 | /// red = InclusiveScan ? buffer[i] : buffer[i-1]; |
| 4184 | /// <scan phase>; |
| 4185 | /// } |
| 4186 | /// \endcode |
| 4187 | static void emitScanBasedDirective( |
| 4188 | CodeGenFunction &CGF, const OMPLoopDirective &S, |
| 4189 | llvm::function_ref<llvm::Value *(CodeGenFunction &)> NumIteratorsGen, |
| 4190 | llvm::function_ref<void(CodeGenFunction &)> FirstGen, |
| 4191 | llvm::function_ref<void(CodeGenFunction &)> SecondGen) { |
| 4192 | llvm::Value *OMPScanNumIterations = CGF.Builder.CreateIntCast( |
| 4193 | V: NumIteratorsGen(CGF), DestTy: CGF.SizeTy, /*isSigned=*/false); |
| 4194 | SmallVector<const Expr *, 4> Privates; |
| 4195 | SmallVector<const Expr *, 4> ReductionOps; |
| 4196 | SmallVector<const Expr *, 4> LHSs; |
| 4197 | SmallVector<const Expr *, 4> RHSs; |
| 4198 | SmallVector<const Expr *, 4> CopyArrayElems; |
| 4199 | for (const auto *C : S.getClausesOfKind<OMPReductionClause>()) { |
| 4200 | assert(C->getModifier() == OMPC_REDUCTION_inscan && |
| 4201 | "Only inscan reductions are expected."); |
| 4202 | Privates.append(in_start: C->privates().begin(), in_end: C->privates().end()); |
| 4203 | ReductionOps.append(in_start: C->reduction_ops().begin(), in_end: C->reduction_ops().end()); |
| 4204 | LHSs.append(in_start: C->lhs_exprs().begin(), in_end: C->lhs_exprs().end()); |
| 4205 | RHSs.append(in_start: C->rhs_exprs().begin(), in_end: C->rhs_exprs().end()); |
| 4206 | CopyArrayElems.append(in_start: C->copy_array_elems().begin(), |
| 4207 | in_end: C->copy_array_elems().end()); |
| 4208 | } |
| 4209 | CodeGenFunction::ParentLoopDirectiveForScanRegion ScanRegion(CGF, S); |
| 4210 | { |
| 4211 | // Emit loop with input phase: |
| 4212 | // #pragma omp ... |
| 4213 | // for (i: 0..<num_iters>) { |
| 4214 | // <input phase>; |
| 4215 | // buffer[i] = red; |
| 4216 | // } |
| 4217 | CGF.OMPFirstScanLoop = true; |
| 4218 | CodeGenFunction::OMPLocalDeclMapRAII Scope(CGF); |
| 4219 | FirstGen(CGF); |
| 4220 | } |
| 4221 | // #pragma omp barrier // in parallel region |
| 4222 | auto &&CodeGen = [&S, OMPScanNumIterations, &LHSs, &RHSs, &CopyArrayElems, |
| 4223 | &ReductionOps, |
| 4224 | &Privates](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 4225 | Action.Enter(CGF); |
| 4226 | // Emit prefix reduction: |
| 4227 | // #pragma omp master // in parallel region |
| 4228 | // for (int k = 0; k <= ceil(log2(n)); ++k) |
| 4229 | llvm::BasicBlock *InputBB = CGF.Builder.GetInsertBlock(); |
| 4230 | llvm::BasicBlock *LoopBB = CGF.createBasicBlock(name: "omp.outer.log.scan.body"); |
| 4231 | llvm::BasicBlock *ExitBB = CGF.createBasicBlock(name: "omp.outer.log.scan.exit"); |
| 4232 | llvm::Function *F = |
| 4233 | CGF.CGM.getIntrinsic(IID: llvm::Intrinsic::log2, Tys: CGF.DoubleTy); |
| 4234 | llvm::Value *Arg = |
| 4235 | CGF.Builder.CreateUIToFP(V: OMPScanNumIterations, DestTy: CGF.DoubleTy); |
| 4236 | llvm::Value *LogVal = CGF.EmitNounwindRuntimeCall(callee: F, args: Arg); |
| 4237 | F = CGF.CGM.getIntrinsic(IID: llvm::Intrinsic::ceil, Tys: CGF.DoubleTy); |
| 4238 | LogVal = CGF.EmitNounwindRuntimeCall(callee: F, args: LogVal); |
| 4239 | LogVal = CGF.Builder.CreateFPToUI(V: LogVal, DestTy: CGF.IntTy); |
| 4240 | llvm::Value *NMin1 = CGF.Builder.CreateNUWSub( |
| 4241 | LHS: OMPScanNumIterations, RHS: llvm::ConstantInt::get(Ty: CGF.SizeTy, V: 1)); |
| 4242 | auto DL = ApplyDebugLocation::CreateDefaultArtificial(CGF, TemporaryLocation: S.getBeginLoc()); |
| 4243 | CGF.EmitBlock(BB: LoopBB); |
| 4244 | auto *Counter = CGF.Builder.CreatePHI(Ty: CGF.IntTy, NumReservedValues: 2); |
| 4245 | // size pow2k = 1; |
| 4246 | auto *Pow2K = CGF.Builder.CreatePHI(Ty: CGF.SizeTy, NumReservedValues: 2); |
| 4247 | Counter->addIncoming(V: llvm::ConstantInt::get(Ty: CGF.IntTy, V: 0), BB: InputBB); |
| 4248 | Pow2K->addIncoming(V: llvm::ConstantInt::get(Ty: CGF.SizeTy, V: 1), BB: InputBB); |
| 4249 | // for (size i = n - 1; i >= 2 ^ k; --i) |
| 4250 | // tmp[i] op= tmp[i-pow2k]; |
| 4251 | llvm::BasicBlock *InnerLoopBB = |
| 4252 | CGF.createBasicBlock(name: "omp.inner.log.scan.body"); |
| 4253 | llvm::BasicBlock *InnerExitBB = |
| 4254 | CGF.createBasicBlock(name: "omp.inner.log.scan.exit"); |
| 4255 | llvm::Value *CmpI = CGF.Builder.CreateICmpUGE(LHS: NMin1, RHS: Pow2K); |
| 4256 | CGF.Builder.CreateCondBr(Cond: CmpI, True: InnerLoopBB, False: InnerExitBB); |
| 4257 | CGF.EmitBlock(BB: InnerLoopBB); |
| 4258 | auto *IVal = CGF.Builder.CreatePHI(Ty: CGF.SizeTy, NumReservedValues: 2); |
| 4259 | IVal->addIncoming(V: NMin1, BB: LoopBB); |
| 4260 | { |
| 4261 | CodeGenFunction::OMPPrivateScope PrivScope(CGF); |
| 4262 | auto *ILHS = LHSs.begin(); |
| 4263 | auto *IRHS = RHSs.begin(); |
| 4264 | for (const Expr *CopyArrayElem : CopyArrayElems) { |
| 4265 | const auto *LHSVD = cast<VarDecl>(Val: cast<DeclRefExpr>(Val: *ILHS)->getDecl()); |
| 4266 | const auto *RHSVD = cast<VarDecl>(Val: cast<DeclRefExpr>(Val: *IRHS)->getDecl()); |
| 4267 | Address LHSAddr = Address::invalid(); |
| 4268 | { |
| 4269 | CodeGenFunction::OpaqueValueMapping IdxMapping( |
| 4270 | CGF, |
| 4271 | cast<OpaqueValueExpr>( |
| 4272 | Val: cast<ArraySubscriptExpr>(Val: CopyArrayElem)->getIdx()), |
| 4273 | RValue::get(V: IVal)); |
| 4274 | LHSAddr = CGF.EmitLValue(E: CopyArrayElem).getAddress(); |
| 4275 | } |
| 4276 | PrivScope.addPrivate(LocalVD: LHSVD, Addr: LHSAddr); |
| 4277 | Address RHSAddr = Address::invalid(); |
| 4278 | { |
| 4279 | llvm::Value *OffsetIVal = CGF.Builder.CreateNUWSub(LHS: IVal, RHS: Pow2K); |
| 4280 | CodeGenFunction::OpaqueValueMapping IdxMapping( |
| 4281 | CGF, |
| 4282 | cast<OpaqueValueExpr>( |
| 4283 | Val: cast<ArraySubscriptExpr>(Val: CopyArrayElem)->getIdx()), |
| 4284 | RValue::get(V: OffsetIVal)); |
| 4285 | RHSAddr = CGF.EmitLValue(E: CopyArrayElem).getAddress(); |
| 4286 | } |
| 4287 | PrivScope.addPrivate(LocalVD: RHSVD, Addr: RHSAddr); |
| 4288 | ++ILHS; |
| 4289 | ++IRHS; |
| 4290 | } |
| 4291 | PrivScope.Privatize(); |
| 4292 | CGF.CGM.getOpenMPRuntime().emitReduction( |
| 4293 | CGF, Loc: S.getEndLoc(), Privates, LHSExprs: LHSs, RHSExprs: RHSs, ReductionOps, |
| 4294 | Options: {/*WithNowait=*/true, /*SimpleReduction=*/true, |
| 4295 | /*IsPrivateVarReduction*/ {}, .ReductionKind: OMPD_unknown}); |
| 4296 | } |
| 4297 | llvm::Value *NextIVal = |
| 4298 | CGF.Builder.CreateNUWSub(LHS: IVal, RHS: llvm::ConstantInt::get(Ty: CGF.SizeTy, V: 1)); |
| 4299 | IVal->addIncoming(V: NextIVal, BB: CGF.Builder.GetInsertBlock()); |
| 4300 | CmpI = CGF.Builder.CreateICmpUGE(LHS: NextIVal, RHS: Pow2K); |
| 4301 | CGF.Builder.CreateCondBr(Cond: CmpI, True: InnerLoopBB, False: InnerExitBB); |
| 4302 | CGF.EmitBlock(BB: InnerExitBB); |
| 4303 | llvm::Value *Next = |
| 4304 | CGF.Builder.CreateNUWAdd(LHS: Counter, RHS: llvm::ConstantInt::get(Ty: CGF.IntTy, V: 1)); |
| 4305 | Counter->addIncoming(V: Next, BB: CGF.Builder.GetInsertBlock()); |
| 4306 | // pow2k <<= 1; |
| 4307 | llvm::Value *NextPow2K = |
| 4308 | CGF.Builder.CreateShl(LHS: Pow2K, RHS: 1, Name: "", /*HasNUW=*/true); |
| 4309 | Pow2K->addIncoming(V: NextPow2K, BB: CGF.Builder.GetInsertBlock()); |
| 4310 | llvm::Value *Cmp = CGF.Builder.CreateICmpNE(LHS: Next, RHS: LogVal); |
| 4311 | CGF.Builder.CreateCondBr(Cond: Cmp, True: LoopBB, False: ExitBB); |
| 4312 | auto DL1 = ApplyDebugLocation::CreateDefaultArtificial(CGF, TemporaryLocation: S.getEndLoc()); |
| 4313 | CGF.EmitBlock(BB: ExitBB); |
| 4314 | }; |
| 4315 | OpenMPDirectiveKind EKind = getEffectiveDirectiveKind(S); |
| 4316 | if (isOpenMPParallelDirective(DKind: EKind)) { |
| 4317 | CGF.CGM.getOpenMPRuntime().emitMasterRegion(CGF, MasterOpGen: CodeGen, Loc: S.getBeginLoc()); |
| 4318 | CGF.CGM.getOpenMPRuntime().emitBarrierCall( |
| 4319 | CGF, Loc: S.getBeginLoc(), Kind: OMPD_unknown, /*EmitChecks=*/false, |
| 4320 | /*ForceSimpleCall=*/true); |
| 4321 | } else { |
| 4322 | RegionCodeGenTy RCG(CodeGen); |
| 4323 | RCG(CGF); |
| 4324 | } |
| 4325 | |
| 4326 | CGF.OMPFirstScanLoop = false; |
| 4327 | SecondGen(CGF); |
| 4328 | } |
| 4329 | |
| 4330 | static bool emitWorksharingDirective(CodeGenFunction &CGF, |
| 4331 | const OMPLoopDirective &S, |
| 4332 | bool HasCancel) { |
| 4333 | bool HasLastprivates; |
| 4334 | OpenMPDirectiveKind EKind = getEffectiveDirectiveKind(S); |
| 4335 | if (llvm::any_of(Range: S.getClausesOfKind<OMPReductionClause>(), |
| 4336 | P: [](const OMPReductionClause *C) { |
| 4337 | return C->getModifier() == OMPC_REDUCTION_inscan; |
| 4338 | })) { |
| 4339 | const auto &&NumIteratorsGen = [&S](CodeGenFunction &CGF) { |
| 4340 | CodeGenFunction::OMPLocalDeclMapRAII Scope(CGF); |
| 4341 | OMPLoopScope LoopScope(CGF, S); |
| 4342 | return CGF.EmitScalarExpr(E: S.getNumIterations()); |
| 4343 | }; |
| 4344 | const auto &&FirstGen = [&S, HasCancel, EKind](CodeGenFunction &CGF) { |
| 4345 | CodeGenFunction::OMPCancelStackRAII CancelRegion(CGF, EKind, HasCancel); |
| 4346 | (void)CGF.EmitOMPWorksharingLoop(S, EUB: S.getEnsureUpperBound(), |
| 4347 | CodeGenLoopBounds: emitForLoopBounds, |
| 4348 | CGDispatchBounds: emitDispatchForLoopBounds); |
| 4349 | // Emit an implicit barrier at the end. |
| 4350 | CGF.CGM.getOpenMPRuntime().emitBarrierCall(CGF, Loc: S.getBeginLoc(), |
| 4351 | Kind: OMPD_for); |
| 4352 | }; |
| 4353 | const auto &&SecondGen = [&S, HasCancel, EKind, |
| 4354 | &HasLastprivates](CodeGenFunction &CGF) { |
| 4355 | CodeGenFunction::OMPCancelStackRAII CancelRegion(CGF, EKind, HasCancel); |
| 4356 | HasLastprivates = CGF.EmitOMPWorksharingLoop(S, EUB: S.getEnsureUpperBound(), |
| 4357 | CodeGenLoopBounds: emitForLoopBounds, |
| 4358 | CGDispatchBounds: emitDispatchForLoopBounds); |
| 4359 | }; |
| 4360 | if (!isOpenMPParallelDirective(DKind: EKind)) |
| 4361 | emitScanBasedDirectiveDecls(CGF, S, NumIteratorsGen); |
| 4362 | emitScanBasedDirective(CGF, S, NumIteratorsGen, FirstGen, SecondGen); |
| 4363 | if (!isOpenMPParallelDirective(DKind: EKind)) |
| 4364 | emitScanBasedDirectiveFinals(CGF, S, NumIteratorsGen); |
| 4365 | } else { |
| 4366 | CodeGenFunction::OMPCancelStackRAII CancelRegion(CGF, EKind, HasCancel); |
| 4367 | HasLastprivates = CGF.EmitOMPWorksharingLoop(S, EUB: S.getEnsureUpperBound(), |
| 4368 | CodeGenLoopBounds: emitForLoopBounds, |
| 4369 | CGDispatchBounds: emitDispatchForLoopBounds); |
| 4370 | } |
| 4371 | return HasLastprivates; |
| 4372 | } |
| 4373 | |
| 4374 | // Pass OMPLoopDirective (instead of OMPForDirective) to make this check |
| 4375 | // available for "loop bind(parallel)", which maps to "for". |
| 4376 | static bool isForSupportedByOpenMPIRBuilder(const OMPLoopDirective &S, |
| 4377 | bool HasCancel) { |
| 4378 | if (HasCancel) |
| 4379 | return false; |
| 4380 | for (OMPClause *C : S.clauses()) { |
| 4381 | if (isa<OMPNowaitClause, OMPBindClause>(Val: C)) |
| 4382 | continue; |
| 4383 | |
| 4384 | if (auto *SC = dyn_cast<OMPScheduleClause>(Val: C)) { |
| 4385 | if (SC->getFirstScheduleModifier() != OMPC_SCHEDULE_MODIFIER_unknown) |
| 4386 | return false; |
| 4387 | if (SC->getSecondScheduleModifier() != OMPC_SCHEDULE_MODIFIER_unknown) |
| 4388 | return false; |
| 4389 | switch (SC->getScheduleKind()) { |
| 4390 | case OMPC_SCHEDULE_auto: |
| 4391 | case OMPC_SCHEDULE_dynamic: |
| 4392 | case OMPC_SCHEDULE_runtime: |
| 4393 | case OMPC_SCHEDULE_guided: |
| 4394 | case OMPC_SCHEDULE_static: |
| 4395 | continue; |
| 4396 | case OMPC_SCHEDULE_unknown: |
| 4397 | return false; |
| 4398 | } |
| 4399 | } |
| 4400 | |
| 4401 | return false; |
| 4402 | } |
| 4403 | |
| 4404 | return true; |
| 4405 | } |
| 4406 | |
| 4407 | static llvm::omp::ScheduleKind |
| 4408 | convertClauseKindToSchedKind(OpenMPScheduleClauseKind ScheduleClauseKind) { |
| 4409 | switch (ScheduleClauseKind) { |
| 4410 | case OMPC_SCHEDULE_unknown: |
| 4411 | return llvm::omp::OMP_SCHEDULE_Default; |
| 4412 | case OMPC_SCHEDULE_auto: |
| 4413 | return llvm::omp::OMP_SCHEDULE_Auto; |
| 4414 | case OMPC_SCHEDULE_dynamic: |
| 4415 | return llvm::omp::OMP_SCHEDULE_Dynamic; |
| 4416 | case OMPC_SCHEDULE_guided: |
| 4417 | return llvm::omp::OMP_SCHEDULE_Guided; |
| 4418 | case OMPC_SCHEDULE_runtime: |
| 4419 | return llvm::omp::OMP_SCHEDULE_Runtime; |
| 4420 | case OMPC_SCHEDULE_static: |
| 4421 | return llvm::omp::OMP_SCHEDULE_Static; |
| 4422 | } |
| 4423 | llvm_unreachable("Unhandled schedule kind"); |
| 4424 | } |
| 4425 | |
| 4426 | // Pass OMPLoopDirective (instead of OMPForDirective) to make this function |
| 4427 | // available for "loop bind(parallel)", which maps to "for". |
| 4428 | static void emitOMPForDirective(const OMPLoopDirective &S, CodeGenFunction &CGF, |
| 4429 | CodeGenModule &CGM, bool HasCancel) { |
| 4430 | bool HasLastprivates = false; |
| 4431 | bool UseOMPIRBuilder = CGM.getLangOpts().OpenMPIRBuilder && |
| 4432 | isForSupportedByOpenMPIRBuilder(S, HasCancel); |
| 4433 | auto &&CodeGen = [&S, &CGM, HasCancel, &HasLastprivates, |
| 4434 | UseOMPIRBuilder](CodeGenFunction &CGF, PrePostActionTy &) { |
| 4435 | // Use the OpenMPIRBuilder if enabled. |
| 4436 | if (UseOMPIRBuilder) { |
| 4437 | bool NeedsBarrier = !S.getSingleClause<OMPNowaitClause>(); |
| 4438 | |
| 4439 | llvm::omp::ScheduleKind SchedKind = llvm::omp::OMP_SCHEDULE_Default; |
| 4440 | llvm::Value *ChunkSize = nullptr; |
| 4441 | if (auto *SchedClause = S.getSingleClause<OMPScheduleClause>()) { |
| 4442 | SchedKind = |
| 4443 | convertClauseKindToSchedKind(ScheduleClauseKind: SchedClause->getScheduleKind()); |
| 4444 | if (const Expr *ChunkSizeExpr = SchedClause->getChunkSize()) |
| 4445 | ChunkSize = CGF.EmitScalarExpr(E: ChunkSizeExpr); |
| 4446 | } |
| 4447 | |
| 4448 | // Emit the associated statement and get its loop representation. |
| 4449 | const Stmt *Inner = S.getRawStmt(); |
| 4450 | llvm::CanonicalLoopInfo *CLI = |
| 4451 | CGF.EmitOMPCollapsedCanonicalLoopNest(S: Inner, Depth: 1); |
| 4452 | |
| 4453 | llvm::OpenMPIRBuilder &OMPBuilder = |
| 4454 | CGM.getOpenMPRuntime().getOMPBuilder(); |
| 4455 | llvm::OpenMPIRBuilder::InsertPointTy AllocaIP( |
| 4456 | CGF.AllocaInsertPt->getParent(), CGF.AllocaInsertPt->getIterator()); |
| 4457 | cantFail(ValOrErr: OMPBuilder.applyWorkshareLoop( |
| 4458 | DL: CGF.Builder.getCurrentDebugLocation(), CLI, AllocaIP, NeedsBarrier, |
| 4459 | SchedKind, ChunkSize, /*HasSimdModifier=*/false, |
| 4460 | /*HasMonotonicModifier=*/false, /*HasNonmonotonicModifier=*/false, |
| 4461 | /*HasOrderedClause=*/false)); |
| 4462 | return; |
| 4463 | } |
| 4464 | |
| 4465 | HasLastprivates = emitWorksharingDirective(CGF, S, HasCancel); |
| 4466 | }; |
| 4467 | { |
| 4468 | auto LPCRegion = |
| 4469 | CGOpenMPRuntime::LastprivateConditionalRAII::disable(CGF, S); |
| 4470 | OMPLexicalScope Scope(CGF, S, OMPD_unknown); |
| 4471 | CGM.getOpenMPRuntime().emitInlinedDirective(CGF, InnermostKind: OMPD_for, CodeGen, |
| 4472 | HasCancel); |
| 4473 | } |
| 4474 | |
| 4475 | if (!UseOMPIRBuilder) { |
| 4476 | // Emit an implicit barrier at the end. |
| 4477 | if (!S.getSingleClause<OMPNowaitClause>() || HasLastprivates) |
| 4478 | CGM.getOpenMPRuntime().emitBarrierCall(CGF, Loc: S.getBeginLoc(), Kind: OMPD_for); |
| 4479 | } |
| 4480 | // Check for outer lastprivate conditional update. |
| 4481 | checkForLastprivateConditionalUpdate(CGF, S); |
| 4482 | } |
| 4483 | |
| 4484 | void CodeGenFunction::EmitOMPForDirective(const OMPForDirective &S) { |
| 4485 | return emitOMPForDirective(S, CGF&: *this, CGM, HasCancel: S.hasCancel()); |
| 4486 | } |
| 4487 | |
| 4488 | void CodeGenFunction::EmitOMPForSimdDirective(const OMPForSimdDirective &S) { |
| 4489 | bool HasLastprivates = false; |
| 4490 | auto &&CodeGen = [&S, &HasLastprivates](CodeGenFunction &CGF, |
| 4491 | PrePostActionTy &) { |
| 4492 | HasLastprivates = emitWorksharingDirective(CGF, S, /*HasCancel=*/false); |
| 4493 | }; |
| 4494 | { |
| 4495 | auto LPCRegion = |
| 4496 | CGOpenMPRuntime::LastprivateConditionalRAII::disable(CGF&: *this, S); |
| 4497 | OMPLexicalScope Scope(*this, S, OMPD_unknown); |
| 4498 | CGM.getOpenMPRuntime().emitInlinedDirective(CGF&: *this, InnermostKind: OMPD_simd, CodeGen); |
| 4499 | } |
| 4500 | |
| 4501 | // Emit an implicit barrier at the end. |
| 4502 | if (!S.getSingleClause<OMPNowaitClause>() || HasLastprivates) |
| 4503 | CGM.getOpenMPRuntime().emitBarrierCall(CGF&: *this, Loc: S.getBeginLoc(), Kind: OMPD_for); |
| 4504 | // Check for outer lastprivate conditional update. |
| 4505 | checkForLastprivateConditionalUpdate(CGF&: *this, S); |
| 4506 | } |
| 4507 | |
| 4508 | static LValue createSectionLVal(CodeGenFunction &CGF, QualType Ty, |
| 4509 | const Twine &Name, |
| 4510 | llvm::Value *Init = nullptr) { |
| 4511 | LValue LVal = CGF.MakeAddrLValue(Addr: CGF.CreateMemTemp(T: Ty, Name), T: Ty); |
| 4512 | if (Init) |
| 4513 | CGF.EmitStoreThroughLValue(Src: RValue::get(V: Init), Dst: LVal, /*isInit*/ true); |
| 4514 | return LVal; |
| 4515 | } |
| 4516 | |
| 4517 | void CodeGenFunction::EmitSections(const OMPExecutableDirective &S) { |
| 4518 | const Stmt *CapturedStmt = S.getInnermostCapturedStmt()->getCapturedStmt(); |
| 4519 | const auto *CS = dyn_cast<CompoundStmt>(Val: CapturedStmt); |
| 4520 | bool HasLastprivates = false; |
| 4521 | OpenMPDirectiveKind EKind = getEffectiveDirectiveKind(S); |
| 4522 | auto &&CodeGen = [&S, CapturedStmt, CS, EKind, |
| 4523 | &HasLastprivates](CodeGenFunction &CGF, PrePostActionTy &) { |
| 4524 | const ASTContext &C = CGF.getContext(); |
| 4525 | QualType KmpInt32Ty = |
| 4526 | C.getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/1); |
| 4527 | // Emit helper vars inits. |
| 4528 | LValue LB = createSectionLVal(CGF, Ty: KmpInt32Ty, Name: ".omp.sections.lb.", |
| 4529 | Init: CGF.Builder.getInt32(C: 0)); |
| 4530 | llvm::ConstantInt *GlobalUBVal = CS != nullptr |
| 4531 | ? CGF.Builder.getInt32(C: CS->size() - 1) |
| 4532 | : CGF.Builder.getInt32(C: 0); |
| 4533 | LValue UB = |
| 4534 | createSectionLVal(CGF, Ty: KmpInt32Ty, Name: ".omp.sections.ub.", Init: GlobalUBVal); |
| 4535 | LValue ST = createSectionLVal(CGF, Ty: KmpInt32Ty, Name: ".omp.sections.st.", |
| 4536 | Init: CGF.Builder.getInt32(C: 1)); |
| 4537 | LValue IL = createSectionLVal(CGF, Ty: KmpInt32Ty, Name: ".omp.sections.il.", |
| 4538 | Init: CGF.Builder.getInt32(C: 0)); |
| 4539 | // Loop counter. |
| 4540 | LValue IV = createSectionLVal(CGF, Ty: KmpInt32Ty, Name: ".omp.sections.iv."); |
| 4541 | OpaqueValueExpr IVRefExpr(S.getBeginLoc(), KmpInt32Ty, VK_LValue); |
| 4542 | CodeGenFunction::OpaqueValueMapping OpaqueIV(CGF, &IVRefExpr, IV); |
| 4543 | OpaqueValueExpr UBRefExpr(S.getBeginLoc(), KmpInt32Ty, VK_LValue); |
| 4544 | CodeGenFunction::OpaqueValueMapping OpaqueUB(CGF, &UBRefExpr, UB); |
| 4545 | // Generate condition for loop. |
| 4546 | BinaryOperator *Cond = BinaryOperator::Create( |
| 4547 | C, lhs: &IVRefExpr, rhs: &UBRefExpr, opc: BO_LE, ResTy: C.BoolTy, VK: VK_PRValue, OK: OK_Ordinary, |
| 4548 | opLoc: S.getBeginLoc(), FPFeatures: FPOptionsOverride()); |
| 4549 | // Increment for loop counter. |
| 4550 | UnaryOperator *Inc = UnaryOperator::Create( |
| 4551 | C, input: &IVRefExpr, opc: UO_PreInc, type: KmpInt32Ty, VK: VK_PRValue, OK: OK_Ordinary, |
| 4552 | l: S.getBeginLoc(), CanOverflow: true, FPFeatures: FPOptionsOverride()); |
| 4553 | auto &&BodyGen = [CapturedStmt, CS, &S, &IV](CodeGenFunction &CGF) { |
| 4554 | // Iterate through all sections and emit a switch construct: |
| 4555 | // switch (IV) { |
| 4556 | // case 0: |
| 4557 | // <SectionStmt[0]>; |
| 4558 | // break; |
| 4559 | // ... |
| 4560 | // case <NumSection> - 1: |
| 4561 | // <SectionStmt[<NumSection> - 1]>; |
| 4562 | // break; |
| 4563 | // } |
| 4564 | // .omp.sections.exit: |
| 4565 | llvm::BasicBlock *ExitBB = CGF.createBasicBlock(name: ".omp.sections.exit"); |
| 4566 | llvm::SwitchInst *SwitchStmt = |
| 4567 | CGF.Builder.CreateSwitch(V: CGF.EmitLoadOfScalar(lvalue: IV, Loc: S.getBeginLoc()), |
| 4568 | Dest: ExitBB, NumCases: CS == nullptr ? 1 : CS->size()); |
| 4569 | if (CS) { |
| 4570 | unsigned CaseNumber = 0; |
| 4571 | for (const Stmt *SubStmt : CS->children()) { |
| 4572 | auto CaseBB = CGF.createBasicBlock(name: ".omp.sections.case"); |
| 4573 | CGF.EmitBlock(BB: CaseBB); |
| 4574 | SwitchStmt->addCase(OnVal: CGF.Builder.getInt32(C: CaseNumber), Dest: CaseBB); |
| 4575 | CGF.EmitStmt(S: SubStmt); |
| 4576 | CGF.EmitBranch(Block: ExitBB); |
| 4577 | ++CaseNumber; |
| 4578 | } |
| 4579 | } else { |
| 4580 | llvm::BasicBlock *CaseBB = CGF.createBasicBlock(name: ".omp.sections.case"); |
| 4581 | CGF.EmitBlock(BB: CaseBB); |
| 4582 | SwitchStmt->addCase(OnVal: CGF.Builder.getInt32(C: 0), Dest: CaseBB); |
| 4583 | CGF.EmitStmt(S: CapturedStmt); |
| 4584 | CGF.EmitBranch(Block: ExitBB); |
| 4585 | } |
| 4586 | CGF.EmitBlock(BB: ExitBB, /*IsFinished=*/true); |
| 4587 | }; |
| 4588 | |
| 4589 | CodeGenFunction::OMPPrivateScope LoopScope(CGF); |
| 4590 | if (CGF.EmitOMPFirstprivateClause(D: S, PrivateScope&: LoopScope)) { |
| 4591 | // Emit implicit barrier to synchronize threads and avoid data races on |
| 4592 | // initialization of firstprivate variables and post-update of lastprivate |
| 4593 | // variables. |
| 4594 | CGF.CGM.getOpenMPRuntime().emitBarrierCall( |
| 4595 | CGF, Loc: S.getBeginLoc(), Kind: OMPD_unknown, /*EmitChecks=*/false, |
| 4596 | /*ForceSimpleCall=*/true); |
| 4597 | } |
| 4598 | CGF.EmitOMPPrivateClause(D: S, PrivateScope&: LoopScope); |
| 4599 | CGOpenMPRuntime::LastprivateConditionalRAII LPCRegion(CGF, S, IV); |
| 4600 | HasLastprivates = CGF.EmitOMPLastprivateClauseInit(D: S, PrivateScope&: LoopScope); |
| 4601 | CGF.EmitOMPReductionClauseInit(D: S, PrivateScope&: LoopScope); |
| 4602 | (void)LoopScope.Privatize(); |
| 4603 | if (isOpenMPTargetExecutionDirective(DKind: EKind)) |
| 4604 | CGF.CGM.getOpenMPRuntime().adjustTargetSpecificDataForLambdas(CGF, D: S); |
| 4605 | |
| 4606 | // Emit static non-chunked loop. |
| 4607 | OpenMPScheduleTy ScheduleKind; |
| 4608 | ScheduleKind.Schedule = OMPC_SCHEDULE_static; |
| 4609 | CGOpenMPRuntime::StaticRTInput StaticInit( |
| 4610 | /*IVSize=*/32, /*IVSigned=*/true, /*Ordered=*/false, IL.getAddress(), |
| 4611 | LB.getAddress(), UB.getAddress(), ST.getAddress()); |
| 4612 | CGF.CGM.getOpenMPRuntime().emitForStaticInit(CGF, Loc: S.getBeginLoc(), DKind: EKind, |
| 4613 | ScheduleKind, Values: StaticInit); |
| 4614 | // UB = min(UB, GlobalUB); |
| 4615 | llvm::Value *UBVal = CGF.EmitLoadOfScalar(lvalue: UB, Loc: S.getBeginLoc()); |
| 4616 | llvm::Value *MinUBGlobalUB = CGF.Builder.CreateSelect( |
| 4617 | C: CGF.Builder.CreateICmpSLT(LHS: UBVal, RHS: GlobalUBVal), True: UBVal, False: GlobalUBVal); |
| 4618 | CGF.EmitStoreOfScalar(value: MinUBGlobalUB, lvalue: UB); |
| 4619 | // IV = LB; |
| 4620 | CGF.EmitStoreOfScalar(value: CGF.EmitLoadOfScalar(lvalue: LB, Loc: S.getBeginLoc()), lvalue: IV); |
| 4621 | // while (idx <= UB) { BODY; ++idx; } |
| 4622 | CGF.EmitOMPInnerLoop(S, /*RequiresCleanup=*/false, LoopCond: Cond, IncExpr: Inc, BodyGen, |
| 4623 | PostIncGen: [](CodeGenFunction &) {}); |
| 4624 | // Tell the runtime we are done. |
| 4625 | auto &&CodeGen = [&S](CodeGenFunction &CGF) { |
| 4626 | CGF.CGM.getOpenMPRuntime().emitForStaticFinish(CGF, Loc: S.getEndLoc(), |
| 4627 | DKind: OMPD_sections); |
| 4628 | }; |
| 4629 | CGF.OMPCancelStack.emitExit(CGF, Kind: EKind, CodeGen); |
| 4630 | CGF.EmitOMPReductionClauseFinal(D: S, /*ReductionKind=*/OMPD_parallel); |
| 4631 | // Emit post-update of the reduction variables if IsLastIter != 0. |
| 4632 | emitPostUpdateForReductionClause(CGF, D: S, CondGen: [IL, &S](CodeGenFunction &CGF) { |
| 4633 | return CGF.Builder.CreateIsNotNull( |
| 4634 | Arg: CGF.EmitLoadOfScalar(lvalue: IL, Loc: S.getBeginLoc())); |
| 4635 | }); |
| 4636 | |
| 4637 | // Emit final copy of the lastprivate variables if IsLastIter != 0. |
| 4638 | if (HasLastprivates) |
| 4639 | CGF.EmitOMPLastprivateClauseFinal( |
| 4640 | D: S, /*NoFinals=*/false, |
| 4641 | IsLastIterCond: CGF.Builder.CreateIsNotNull( |
| 4642 | Arg: CGF.EmitLoadOfScalar(lvalue: IL, Loc: S.getBeginLoc()))); |
| 4643 | }; |
| 4644 | |
| 4645 | bool HasCancel = false; |
| 4646 | if (auto *OSD = dyn_cast<OMPSectionsDirective>(Val: &S)) |
| 4647 | HasCancel = OSD->hasCancel(); |
| 4648 | else if (auto *OPSD = dyn_cast<OMPParallelSectionsDirective>(Val: &S)) |
| 4649 | HasCancel = OPSD->hasCancel(); |
| 4650 | OMPCancelStackRAII CancelRegion(*this, EKind, HasCancel); |
| 4651 | CGM.getOpenMPRuntime().emitInlinedDirective(CGF&: *this, InnermostKind: OMPD_sections, CodeGen, |
| 4652 | HasCancel); |
| 4653 | // Emit barrier for lastprivates only if 'sections' directive has 'nowait' |
| 4654 | // clause. Otherwise the barrier will be generated by the codegen for the |
| 4655 | // directive. |
| 4656 | if (HasLastprivates && S.getSingleClause<OMPNowaitClause>()) { |
| 4657 | // Emit implicit barrier to synchronize threads and avoid data races on |
| 4658 | // initialization of firstprivate variables. |
| 4659 | CGM.getOpenMPRuntime().emitBarrierCall(CGF&: *this, Loc: S.getBeginLoc(), |
| 4660 | Kind: OMPD_unknown); |
| 4661 | } |
| 4662 | } |
| 4663 | |
| 4664 | void CodeGenFunction::EmitOMPScopeDirective(const OMPScopeDirective &S) { |
| 4665 | { |
| 4666 | // Emit code for 'scope' region |
| 4667 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 4668 | Action.Enter(CGF); |
| 4669 | OMPPrivateScope PrivateScope(CGF); |
| 4670 | (void)CGF.EmitOMPFirstprivateClause(D: S, PrivateScope); |
| 4671 | CGF.EmitOMPPrivateClause(D: S, PrivateScope); |
| 4672 | CGF.EmitOMPReductionClauseInit(D: S, PrivateScope); |
| 4673 | (void)PrivateScope.Privatize(); |
| 4674 | CGF.EmitStmt(S: S.getInnermostCapturedStmt()->getCapturedStmt()); |
| 4675 | CGF.EmitOMPReductionClauseFinal(D: S, /*ReductionKind=*/OMPD_parallel); |
| 4676 | }; |
| 4677 | auto LPCRegion = |
| 4678 | CGOpenMPRuntime::LastprivateConditionalRAII::disable(CGF&: *this, S); |
| 4679 | OMPLexicalScope Scope(*this, S, OMPD_unknown); |
| 4680 | CGM.getOpenMPRuntime().emitInlinedDirective(CGF&: *this, InnermostKind: OMPD_scope, CodeGen); |
| 4681 | } |
| 4682 | // Emit an implicit barrier at the end. |
| 4683 | if (!S.getSingleClause<OMPNowaitClause>()) { |
| 4684 | CGM.getOpenMPRuntime().emitBarrierCall(CGF&: *this, Loc: S.getBeginLoc(), Kind: OMPD_scope); |
| 4685 | } |
| 4686 | // Check for outer lastprivate conditional update. |
| 4687 | checkForLastprivateConditionalUpdate(CGF&: *this, S); |
| 4688 | } |
| 4689 | |
| 4690 | void CodeGenFunction::EmitOMPSectionsDirective(const OMPSectionsDirective &S) { |
| 4691 | if (CGM.getLangOpts().OpenMPIRBuilder) { |
| 4692 | llvm::OpenMPIRBuilder &OMPBuilder = CGM.getOpenMPRuntime().getOMPBuilder(); |
| 4693 | using InsertPointTy = llvm::OpenMPIRBuilder::InsertPointTy; |
| 4694 | using BodyGenCallbackTy = llvm::OpenMPIRBuilder::StorableBodyGenCallbackTy; |
| 4695 | |
| 4696 | auto FiniCB = [](InsertPointTy IP) { |
| 4697 | // Don't FinalizeOMPRegion because this is done inside of OMPIRBuilder for |
| 4698 | // sections. |
| 4699 | return llvm::Error::success(); |
| 4700 | }; |
| 4701 | |
| 4702 | const CapturedStmt *ICS = S.getInnermostCapturedStmt(); |
| 4703 | const Stmt *CapturedStmt = S.getInnermostCapturedStmt()->getCapturedStmt(); |
| 4704 | const auto *CS = dyn_cast<CompoundStmt>(Val: CapturedStmt); |
| 4705 | llvm::SmallVector<BodyGenCallbackTy, 4> SectionCBVector; |
| 4706 | if (CS) { |
| 4707 | for (const Stmt *SubStmt : CS->children()) { |
| 4708 | auto SectionCB = [this, SubStmt]( |
| 4709 | InsertPointTy AllocIP, InsertPointTy CodeGenIP, |
| 4710 | ArrayRef<llvm::BasicBlock *> DeallocBlocks) { |
| 4711 | OMPBuilderCBHelpers::EmitOMPInlinedRegionBody(CGF&: *this, RegionBodyStmt: SubStmt, AllocaIP: AllocIP, |
| 4712 | CodeGenIP, RegionName: "section"); |
| 4713 | return llvm::Error::success(); |
| 4714 | }; |
| 4715 | SectionCBVector.push_back(Elt: SectionCB); |
| 4716 | } |
| 4717 | } else { |
| 4718 | auto SectionCB = |
| 4719 | [this, CapturedStmt](InsertPointTy AllocIP, InsertPointTy CodeGenIP, |
| 4720 | ArrayRef<llvm::BasicBlock *> DeallocBlocks) { |
| 4721 | OMPBuilderCBHelpers::EmitOMPInlinedRegionBody( |
| 4722 | CGF&: *this, RegionBodyStmt: CapturedStmt, AllocaIP: AllocIP, CodeGenIP, RegionName: "section"); |
| 4723 | return llvm::Error::success(); |
| 4724 | }; |
| 4725 | SectionCBVector.push_back(Elt: SectionCB); |
| 4726 | } |
| 4727 | |
| 4728 | // Privatization callback that performs appropriate action for |
| 4729 | // shared/private/firstprivate/lastprivate/copyin/... variables. |
| 4730 | // |
| 4731 | // TODO: This defaults to shared right now. |
| 4732 | auto PrivCB = [](InsertPointTy AllocaIP, InsertPointTy CodeGenIP, |
| 4733 | llvm::Value &, llvm::Value &Val, llvm::Value *&ReplVal) { |
| 4734 | // The next line is appropriate only for variables (Val) with the |
| 4735 | // data-sharing attribute "shared". |
| 4736 | ReplVal = &Val; |
| 4737 | |
| 4738 | return CodeGenIP; |
| 4739 | }; |
| 4740 | |
| 4741 | CGCapturedStmtInfo CGSI(*ICS, CR_OpenMP); |
| 4742 | CodeGenFunction::CGCapturedStmtRAII CapInfoRAII(*this, &CGSI); |
| 4743 | llvm::OpenMPIRBuilder::InsertPointTy AllocaIP( |
| 4744 | AllocaInsertPt->getParent(), AllocaInsertPt->getIterator()); |
| 4745 | llvm::OpenMPIRBuilder::InsertPointTy AfterIP = |
| 4746 | cantFail(ValOrErr: OMPBuilder.createSections( |
| 4747 | Loc: Builder, AllocaIP, SectionCBs: SectionCBVector, PrivCB, FiniCB, IsCancellable: S.hasCancel(), |
| 4748 | IsNowait: S.getSingleClause<OMPNowaitClause>())); |
| 4749 | Builder.restoreIP(IP: AfterIP); |
| 4750 | return; |
| 4751 | } |
| 4752 | { |
| 4753 | auto LPCRegion = |
| 4754 | CGOpenMPRuntime::LastprivateConditionalRAII::disable(CGF&: *this, S); |
| 4755 | OMPLexicalScope Scope(*this, S, OMPD_unknown); |
| 4756 | EmitSections(S); |
| 4757 | } |
| 4758 | // Emit an implicit barrier at the end. |
| 4759 | if (!S.getSingleClause<OMPNowaitClause>()) { |
| 4760 | CGM.getOpenMPRuntime().emitBarrierCall(CGF&: *this, Loc: S.getBeginLoc(), |
| 4761 | Kind: OMPD_sections); |
| 4762 | } |
| 4763 | // Check for outer lastprivate conditional update. |
| 4764 | checkForLastprivateConditionalUpdate(CGF&: *this, S); |
| 4765 | } |
| 4766 | |
| 4767 | void CodeGenFunction::EmitOMPSectionDirective(const OMPSectionDirective &S) { |
| 4768 | if (CGM.getLangOpts().OpenMPIRBuilder) { |
| 4769 | llvm::OpenMPIRBuilder &OMPBuilder = CGM.getOpenMPRuntime().getOMPBuilder(); |
| 4770 | using InsertPointTy = llvm::OpenMPIRBuilder::InsertPointTy; |
| 4771 | |
| 4772 | const Stmt *SectionRegionBodyStmt = S.getAssociatedStmt(); |
| 4773 | auto FiniCB = [this](InsertPointTy IP) { |
| 4774 | OMPBuilderCBHelpers::FinalizeOMPRegion(CGF&: *this, IP); |
| 4775 | return llvm::Error::success(); |
| 4776 | }; |
| 4777 | |
| 4778 | auto BodyGenCB = [SectionRegionBodyStmt, |
| 4779 | this](InsertPointTy AllocIP, InsertPointTy CodeGenIP, |
| 4780 | ArrayRef<llvm::BasicBlock *> DeallocBlocks) { |
| 4781 | OMPBuilderCBHelpers::EmitOMPInlinedRegionBody( |
| 4782 | CGF&: *this, RegionBodyStmt: SectionRegionBodyStmt, AllocaIP: AllocIP, CodeGenIP, RegionName: "section"); |
| 4783 | return llvm::Error::success(); |
| 4784 | }; |
| 4785 | |
| 4786 | LexicalScope Scope(*this, S.getSourceRange()); |
| 4787 | EmitStopPoint(S: &S); |
| 4788 | llvm::OpenMPIRBuilder::InsertPointTy AfterIP = |
| 4789 | cantFail(ValOrErr: OMPBuilder.createSection(Loc: Builder, BodyGenCB, FiniCB)); |
| 4790 | Builder.restoreIP(IP: AfterIP); |
| 4791 | |
| 4792 | return; |
| 4793 | } |
| 4794 | LexicalScope Scope(*this, S.getSourceRange()); |
| 4795 | EmitStopPoint(S: &S); |
| 4796 | EmitStmt(S: S.getAssociatedStmt()); |
| 4797 | } |
| 4798 | |
| 4799 | void CodeGenFunction::EmitOMPSingleDirective(const OMPSingleDirective &S) { |
| 4800 | llvm::SmallVector<const Expr *, 8> CopyprivateVars; |
| 4801 | llvm::SmallVector<const Expr *, 8> DestExprs; |
| 4802 | llvm::SmallVector<const Expr *, 8> SrcExprs; |
| 4803 | llvm::SmallVector<const Expr *, 8> AssignmentOps; |
| 4804 | // Check if there are any 'copyprivate' clauses associated with this |
| 4805 | // 'single' construct. |
| 4806 | // Build a list of copyprivate variables along with helper expressions |
| 4807 | // (<source>, <destination>, <destination>=<source> expressions) |
| 4808 | for (const auto *C : S.getClausesOfKind<OMPCopyprivateClause>()) { |
| 4809 | CopyprivateVars.append(in_start: C->varlist_begin(), in_end: C->varlist_end()); |
| 4810 | DestExprs.append(in_start: C->destination_exprs().begin(), |
| 4811 | in_end: C->destination_exprs().end()); |
| 4812 | SrcExprs.append(in_start: C->source_exprs().begin(), in_end: C->source_exprs().end()); |
| 4813 | AssignmentOps.append(in_start: C->assignment_ops().begin(), |
| 4814 | in_end: C->assignment_ops().end()); |
| 4815 | } |
| 4816 | // Emit code for 'single' region along with 'copyprivate' clauses |
| 4817 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 4818 | Action.Enter(CGF); |
| 4819 | OMPPrivateScope SingleScope(CGF); |
| 4820 | (void)CGF.EmitOMPFirstprivateClause(D: S, PrivateScope&: SingleScope); |
| 4821 | CGF.EmitOMPPrivateClause(D: S, PrivateScope&: SingleScope); |
| 4822 | (void)SingleScope.Privatize(); |
| 4823 | CGF.EmitStmt(S: S.getInnermostCapturedStmt()->getCapturedStmt()); |
| 4824 | }; |
| 4825 | { |
| 4826 | auto LPCRegion = |
| 4827 | CGOpenMPRuntime::LastprivateConditionalRAII::disable(CGF&: *this, S); |
| 4828 | OMPLexicalScope Scope(*this, S, OMPD_unknown); |
| 4829 | CGM.getOpenMPRuntime().emitSingleRegion(CGF&: *this, SingleOpGen: CodeGen, Loc: S.getBeginLoc(), |
| 4830 | CopyprivateVars, DestExprs, |
| 4831 | SrcExprs, AssignmentOps); |
| 4832 | } |
| 4833 | // Emit an implicit barrier at the end (to avoid data race on firstprivate |
| 4834 | // init or if no 'nowait' clause was specified and no 'copyprivate' clause). |
| 4835 | if (!S.getSingleClause<OMPNowaitClause>() && CopyprivateVars.empty()) { |
| 4836 | CGM.getOpenMPRuntime().emitBarrierCall( |
| 4837 | CGF&: *this, Loc: S.getBeginLoc(), |
| 4838 | Kind: S.getSingleClause<OMPNowaitClause>() ? OMPD_unknown : OMPD_single); |
| 4839 | } |
| 4840 | // Check for outer lastprivate conditional update. |
| 4841 | checkForLastprivateConditionalUpdate(CGF&: *this, S); |
| 4842 | } |
| 4843 | |
| 4844 | static void emitMaster(CodeGenFunction &CGF, const OMPExecutableDirective &S) { |
| 4845 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 4846 | Action.Enter(CGF); |
| 4847 | CGF.EmitStmt(S: S.getRawStmt()); |
| 4848 | }; |
| 4849 | CGF.CGM.getOpenMPRuntime().emitMasterRegion(CGF, MasterOpGen: CodeGen, Loc: S.getBeginLoc()); |
| 4850 | } |
| 4851 | |
| 4852 | void CodeGenFunction::EmitOMPMasterDirective(const OMPMasterDirective &S) { |
| 4853 | if (CGM.getLangOpts().OpenMPIRBuilder) { |
| 4854 | llvm::OpenMPIRBuilder &OMPBuilder = CGM.getOpenMPRuntime().getOMPBuilder(); |
| 4855 | using InsertPointTy = llvm::OpenMPIRBuilder::InsertPointTy; |
| 4856 | |
| 4857 | const Stmt *MasterRegionBodyStmt = S.getAssociatedStmt(); |
| 4858 | |
| 4859 | auto FiniCB = [this](InsertPointTy IP) { |
| 4860 | OMPBuilderCBHelpers::FinalizeOMPRegion(CGF&: *this, IP); |
| 4861 | return llvm::Error::success(); |
| 4862 | }; |
| 4863 | |
| 4864 | auto BodyGenCB = [MasterRegionBodyStmt, |
| 4865 | this](InsertPointTy AllocIP, InsertPointTy CodeGenIP, |
| 4866 | ArrayRef<llvm::BasicBlock *> DeallocBlocks) { |
| 4867 | OMPBuilderCBHelpers::EmitOMPInlinedRegionBody( |
| 4868 | CGF&: *this, RegionBodyStmt: MasterRegionBodyStmt, AllocaIP: AllocIP, CodeGenIP, RegionName: "master"); |
| 4869 | return llvm::Error::success(); |
| 4870 | }; |
| 4871 | |
| 4872 | LexicalScope Scope(*this, S.getSourceRange()); |
| 4873 | EmitStopPoint(S: &S); |
| 4874 | llvm::OpenMPIRBuilder::InsertPointTy AfterIP = |
| 4875 | cantFail(ValOrErr: OMPBuilder.createMaster(Loc: Builder, BodyGenCB, FiniCB)); |
| 4876 | Builder.restoreIP(IP: AfterIP); |
| 4877 | |
| 4878 | return; |
| 4879 | } |
| 4880 | LexicalScope Scope(*this, S.getSourceRange()); |
| 4881 | EmitStopPoint(S: &S); |
| 4882 | emitMaster(CGF&: *this, S); |
| 4883 | } |
| 4884 | |
| 4885 | static void emitMasked(CodeGenFunction &CGF, const OMPExecutableDirective &S) { |
| 4886 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 4887 | Action.Enter(CGF); |
| 4888 | CGF.EmitStmt(S: S.getRawStmt()); |
| 4889 | }; |
| 4890 | Expr *Filter = nullptr; |
| 4891 | if (const auto *FilterClause = S.getSingleClause<OMPFilterClause>()) |
| 4892 | Filter = FilterClause->getThreadID(); |
| 4893 | CGF.CGM.getOpenMPRuntime().emitMaskedRegion(CGF, MaskedOpGen: CodeGen, Loc: S.getBeginLoc(), |
| 4894 | Filter); |
| 4895 | } |
| 4896 | |
| 4897 | void CodeGenFunction::EmitOMPMaskedDirective(const OMPMaskedDirective &S) { |
| 4898 | if (CGM.getLangOpts().OpenMPIRBuilder) { |
| 4899 | llvm::OpenMPIRBuilder &OMPBuilder = CGM.getOpenMPRuntime().getOMPBuilder(); |
| 4900 | using InsertPointTy = llvm::OpenMPIRBuilder::InsertPointTy; |
| 4901 | |
| 4902 | const Stmt *MaskedRegionBodyStmt = S.getAssociatedStmt(); |
| 4903 | const Expr *Filter = nullptr; |
| 4904 | if (const auto *FilterClause = S.getSingleClause<OMPFilterClause>()) |
| 4905 | Filter = FilterClause->getThreadID(); |
| 4906 | llvm::Value *FilterVal = Filter |
| 4907 | ? EmitScalarExpr(E: Filter, IgnoreResultAssign: CGM.Int32Ty) |
| 4908 | : llvm::ConstantInt::get(Ty: CGM.Int32Ty, /*V=*/0); |
| 4909 | |
| 4910 | auto FiniCB = [this](InsertPointTy IP) { |
| 4911 | OMPBuilderCBHelpers::FinalizeOMPRegion(CGF&: *this, IP); |
| 4912 | return llvm::Error::success(); |
| 4913 | }; |
| 4914 | |
| 4915 | auto BodyGenCB = [MaskedRegionBodyStmt, |
| 4916 | this](InsertPointTy AllocIP, InsertPointTy CodeGenIP, |
| 4917 | ArrayRef<llvm::BasicBlock *> DeallocBlocks) { |
| 4918 | OMPBuilderCBHelpers::EmitOMPInlinedRegionBody( |
| 4919 | CGF&: *this, RegionBodyStmt: MaskedRegionBodyStmt, AllocaIP: AllocIP, CodeGenIP, RegionName: "masked"); |
| 4920 | return llvm::Error::success(); |
| 4921 | }; |
| 4922 | |
| 4923 | LexicalScope Scope(*this, S.getSourceRange()); |
| 4924 | EmitStopPoint(S: &S); |
| 4925 | llvm::OpenMPIRBuilder::InsertPointTy AfterIP = cantFail( |
| 4926 | ValOrErr: OMPBuilder.createMasked(Loc: Builder, BodyGenCB, FiniCB, Filter: FilterVal)); |
| 4927 | Builder.restoreIP(IP: AfterIP); |
| 4928 | |
| 4929 | return; |
| 4930 | } |
| 4931 | LexicalScope Scope(*this, S.getSourceRange()); |
| 4932 | EmitStopPoint(S: &S); |
| 4933 | emitMasked(CGF&: *this, S); |
| 4934 | } |
| 4935 | |
| 4936 | void CodeGenFunction::EmitOMPCriticalDirective(const OMPCriticalDirective &S) { |
| 4937 | if (CGM.getLangOpts().OpenMPIRBuilder) { |
| 4938 | llvm::OpenMPIRBuilder &OMPBuilder = CGM.getOpenMPRuntime().getOMPBuilder(); |
| 4939 | using InsertPointTy = llvm::OpenMPIRBuilder::InsertPointTy; |
| 4940 | |
| 4941 | const Stmt *CriticalRegionBodyStmt = S.getAssociatedStmt(); |
| 4942 | const Expr *Hint = nullptr; |
| 4943 | if (const auto *HintClause = S.getSingleClause<OMPHintClause>()) |
| 4944 | Hint = HintClause->getHint(); |
| 4945 | |
| 4946 | // TODO: This is slightly different from what's currently being done in |
| 4947 | // clang. Fix the Int32Ty to IntPtrTy (pointer width size) when everything |
| 4948 | // about typing is final. |
| 4949 | llvm::Value *HintInst = nullptr; |
| 4950 | if (Hint) |
| 4951 | HintInst = |
| 4952 | Builder.CreateIntCast(V: EmitScalarExpr(E: Hint), DestTy: CGM.Int32Ty, isSigned: false); |
| 4953 | |
| 4954 | auto FiniCB = [this](InsertPointTy IP) { |
| 4955 | OMPBuilderCBHelpers::FinalizeOMPRegion(CGF&: *this, IP); |
| 4956 | return llvm::Error::success(); |
| 4957 | }; |
| 4958 | |
| 4959 | auto BodyGenCB = [CriticalRegionBodyStmt, |
| 4960 | this](InsertPointTy AllocIP, InsertPointTy CodeGenIP, |
| 4961 | ArrayRef<llvm::BasicBlock *> DeallocBlocks) { |
| 4962 | OMPBuilderCBHelpers::EmitOMPInlinedRegionBody( |
| 4963 | CGF&: *this, RegionBodyStmt: CriticalRegionBodyStmt, AllocaIP: AllocIP, CodeGenIP, RegionName: "critical"); |
| 4964 | return llvm::Error::success(); |
| 4965 | }; |
| 4966 | |
| 4967 | LexicalScope Scope(*this, S.getSourceRange()); |
| 4968 | EmitStopPoint(S: &S); |
| 4969 | llvm::OpenMPIRBuilder::InsertPointTy AfterIP = |
| 4970 | cantFail(ValOrErr: OMPBuilder.createCritical(Loc: Builder, BodyGenCB, FiniCB, |
| 4971 | CriticalName: S.getDirectiveName().getAsString(), |
| 4972 | HintInst)); |
| 4973 | Builder.restoreIP(IP: AfterIP); |
| 4974 | |
| 4975 | return; |
| 4976 | } |
| 4977 | |
| 4978 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 4979 | Action.Enter(CGF); |
| 4980 | CGF.EmitStmt(S: S.getAssociatedStmt()); |
| 4981 | }; |
| 4982 | const Expr *Hint = nullptr; |
| 4983 | if (const auto *HintClause = S.getSingleClause<OMPHintClause>()) |
| 4984 | Hint = HintClause->getHint(); |
| 4985 | LexicalScope Scope(*this, S.getSourceRange()); |
| 4986 | EmitStopPoint(S: &S); |
| 4987 | CGM.getOpenMPRuntime().emitCriticalRegion(CGF&: *this, |
| 4988 | CriticalName: S.getDirectiveName().getAsString(), |
| 4989 | CriticalOpGen: CodeGen, Loc: S.getBeginLoc(), Hint); |
| 4990 | } |
| 4991 | |
| 4992 | void CodeGenFunction::EmitOMPParallelForDirective( |
| 4993 | const OMPParallelForDirective &S) { |
| 4994 | // Emit directive as a combined directive that consists of two implicit |
| 4995 | // directives: 'parallel' with 'for' directive. |
| 4996 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 4997 | Action.Enter(CGF); |
| 4998 | emitOMPCopyinClause(CGF, S); |
| 4999 | (void)emitWorksharingDirective(CGF, S, HasCancel: S.hasCancel()); |
| 5000 | }; |
| 5001 | { |
| 5002 | const auto &&NumIteratorsGen = [&S](CodeGenFunction &CGF) { |
| 5003 | CodeGenFunction::OMPLocalDeclMapRAII Scope(CGF); |
| 5004 | CGCapturedStmtInfo CGSI(CR_OpenMP); |
| 5005 | CodeGenFunction::CGCapturedStmtRAII CapInfoRAII(CGF, &CGSI); |
| 5006 | OMPLoopScope LoopScope(CGF, S); |
| 5007 | return CGF.EmitScalarExpr(E: S.getNumIterations()); |
| 5008 | }; |
| 5009 | bool IsInscan = llvm::any_of(Range: S.getClausesOfKind<OMPReductionClause>(), |
| 5010 | P: [](const OMPReductionClause *C) { |
| 5011 | return C->getModifier() == OMPC_REDUCTION_inscan; |
| 5012 | }); |
| 5013 | if (IsInscan) |
| 5014 | emitScanBasedDirectiveDecls(CGF&: *this, S, NumIteratorsGen); |
| 5015 | auto LPCRegion = |
| 5016 | CGOpenMPRuntime::LastprivateConditionalRAII::disable(CGF&: *this, S); |
| 5017 | emitCommonOMPParallelDirective(CGF&: *this, S, InnermostKind: OMPD_for, CodeGen, |
| 5018 | CodeGenBoundParameters: emitEmptyBoundParameters); |
| 5019 | if (IsInscan) |
| 5020 | emitScanBasedDirectiveFinals(CGF&: *this, S, NumIteratorsGen); |
| 5021 | } |
| 5022 | // Check for outer lastprivate conditional update. |
| 5023 | checkForLastprivateConditionalUpdate(CGF&: *this, S); |
| 5024 | } |
| 5025 | |
| 5026 | void CodeGenFunction::EmitOMPParallelForSimdDirective( |
| 5027 | const OMPParallelForSimdDirective &S) { |
| 5028 | // Emit directive as a combined directive that consists of two implicit |
| 5029 | // directives: 'parallel' with 'for' directive. |
| 5030 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 5031 | Action.Enter(CGF); |
| 5032 | emitOMPCopyinClause(CGF, S); |
| 5033 | (void)emitWorksharingDirective(CGF, S, /*HasCancel=*/false); |
| 5034 | }; |
| 5035 | { |
| 5036 | const auto &&NumIteratorsGen = [&S](CodeGenFunction &CGF) { |
| 5037 | CodeGenFunction::OMPLocalDeclMapRAII Scope(CGF); |
| 5038 | CGCapturedStmtInfo CGSI(CR_OpenMP); |
| 5039 | CodeGenFunction::CGCapturedStmtRAII CapInfoRAII(CGF, &CGSI); |
| 5040 | OMPLoopScope LoopScope(CGF, S); |
| 5041 | return CGF.EmitScalarExpr(E: S.getNumIterations()); |
| 5042 | }; |
| 5043 | bool IsInscan = llvm::any_of(Range: S.getClausesOfKind<OMPReductionClause>(), |
| 5044 | P: [](const OMPReductionClause *C) { |
| 5045 | return C->getModifier() == OMPC_REDUCTION_inscan; |
| 5046 | }); |
| 5047 | if (IsInscan) |
| 5048 | emitScanBasedDirectiveDecls(CGF&: *this, S, NumIteratorsGen); |
| 5049 | auto LPCRegion = |
| 5050 | CGOpenMPRuntime::LastprivateConditionalRAII::disable(CGF&: *this, S); |
| 5051 | emitCommonOMPParallelDirective(CGF&: *this, S, InnermostKind: OMPD_for_simd, CodeGen, |
| 5052 | CodeGenBoundParameters: emitEmptyBoundParameters); |
| 5053 | if (IsInscan) |
| 5054 | emitScanBasedDirectiveFinals(CGF&: *this, S, NumIteratorsGen); |
| 5055 | } |
| 5056 | // Check for outer lastprivate conditional update. |
| 5057 | checkForLastprivateConditionalUpdate(CGF&: *this, S); |
| 5058 | } |
| 5059 | |
| 5060 | void CodeGenFunction::EmitOMPParallelMasterDirective( |
| 5061 | const OMPParallelMasterDirective &S) { |
| 5062 | // Emit directive as a combined directive that consists of two implicit |
| 5063 | // directives: 'parallel' with 'master' directive. |
| 5064 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 5065 | Action.Enter(CGF); |
| 5066 | OMPPrivateScope PrivateScope(CGF); |
| 5067 | emitOMPCopyinClause(CGF, S); |
| 5068 | (void)CGF.EmitOMPFirstprivateClause(D: S, PrivateScope); |
| 5069 | CGF.EmitOMPPrivateClause(D: S, PrivateScope); |
| 5070 | CGF.EmitOMPReductionClauseInit(D: S, PrivateScope); |
| 5071 | (void)PrivateScope.Privatize(); |
| 5072 | emitMaster(CGF, S); |
| 5073 | CGF.EmitOMPReductionClauseFinal(D: S, /*ReductionKind=*/OMPD_parallel); |
| 5074 | }; |
| 5075 | { |
| 5076 | auto LPCRegion = |
| 5077 | CGOpenMPRuntime::LastprivateConditionalRAII::disable(CGF&: *this, S); |
| 5078 | emitCommonOMPParallelDirective(CGF&: *this, S, InnermostKind: OMPD_master, CodeGen, |
| 5079 | CodeGenBoundParameters: emitEmptyBoundParameters); |
| 5080 | emitPostUpdateForReductionClause(CGF&: *this, D: S, |
| 5081 | CondGen: [](CodeGenFunction &) { return nullptr; }); |
| 5082 | } |
| 5083 | // Check for outer lastprivate conditional update. |
| 5084 | checkForLastprivateConditionalUpdate(CGF&: *this, S); |
| 5085 | } |
| 5086 | |
| 5087 | void CodeGenFunction::EmitOMPParallelMaskedDirective( |
| 5088 | const OMPParallelMaskedDirective &S) { |
| 5089 | // Emit directive as a combined directive that consists of two implicit |
| 5090 | // directives: 'parallel' with 'masked' directive. |
| 5091 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 5092 | Action.Enter(CGF); |
| 5093 | OMPPrivateScope PrivateScope(CGF); |
| 5094 | emitOMPCopyinClause(CGF, S); |
| 5095 | (void)CGF.EmitOMPFirstprivateClause(D: S, PrivateScope); |
| 5096 | CGF.EmitOMPPrivateClause(D: S, PrivateScope); |
| 5097 | CGF.EmitOMPReductionClauseInit(D: S, PrivateScope); |
| 5098 | (void)PrivateScope.Privatize(); |
| 5099 | emitMasked(CGF, S); |
| 5100 | CGF.EmitOMPReductionClauseFinal(D: S, /*ReductionKind=*/OMPD_parallel); |
| 5101 | }; |
| 5102 | { |
| 5103 | auto LPCRegion = |
| 5104 | CGOpenMPRuntime::LastprivateConditionalRAII::disable(CGF&: *this, S); |
| 5105 | emitCommonOMPParallelDirective(CGF&: *this, S, InnermostKind: OMPD_masked, CodeGen, |
| 5106 | CodeGenBoundParameters: emitEmptyBoundParameters); |
| 5107 | emitPostUpdateForReductionClause(CGF&: *this, D: S, |
| 5108 | CondGen: [](CodeGenFunction &) { return nullptr; }); |
| 5109 | } |
| 5110 | // Check for outer lastprivate conditional update. |
| 5111 | checkForLastprivateConditionalUpdate(CGF&: *this, S); |
| 5112 | } |
| 5113 | |
| 5114 | void CodeGenFunction::EmitOMPParallelSectionsDirective( |
| 5115 | const OMPParallelSectionsDirective &S) { |
| 5116 | // Emit directive as a combined directive that consists of two implicit |
| 5117 | // directives: 'parallel' with 'sections' directive. |
| 5118 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 5119 | Action.Enter(CGF); |
| 5120 | emitOMPCopyinClause(CGF, S); |
| 5121 | CGF.EmitSections(S); |
| 5122 | }; |
| 5123 | { |
| 5124 | auto LPCRegion = |
| 5125 | CGOpenMPRuntime::LastprivateConditionalRAII::disable(CGF&: *this, S); |
| 5126 | emitCommonOMPParallelDirective(CGF&: *this, S, InnermostKind: OMPD_sections, CodeGen, |
| 5127 | CodeGenBoundParameters: emitEmptyBoundParameters); |
| 5128 | } |
| 5129 | // Check for outer lastprivate conditional update. |
| 5130 | checkForLastprivateConditionalUpdate(CGF&: *this, S); |
| 5131 | } |
| 5132 | |
| 5133 | namespace { |
| 5134 | /// Get the list of variables declared in the context of the untied tasks. |
| 5135 | class CheckVarsEscapingUntiedTaskDeclContext final |
| 5136 | : public ConstStmtVisitor<CheckVarsEscapingUntiedTaskDeclContext> { |
| 5137 | llvm::SmallVector<const VarDecl *, 4> PrivateDecls; |
| 5138 | |
| 5139 | public: |
| 5140 | explicit CheckVarsEscapingUntiedTaskDeclContext() = default; |
| 5141 | ~CheckVarsEscapingUntiedTaskDeclContext() = default; |
| 5142 | void VisitDeclStmt(const DeclStmt *S) { |
| 5143 | if (!S) |
| 5144 | return; |
| 5145 | // Need to privatize only local vars, static locals can be processed as is. |
| 5146 | for (const Decl *D : S->decls()) { |
| 5147 | if (const auto *VD = dyn_cast_or_null<VarDecl>(Val: D)) |
| 5148 | if (VD->hasLocalStorage()) |
| 5149 | PrivateDecls.push_back(Elt: VD); |
| 5150 | } |
| 5151 | } |
| 5152 | void VisitOMPExecutableDirective(const OMPExecutableDirective *) {} |
| 5153 | void VisitCapturedStmt(const CapturedStmt *) {} |
| 5154 | void VisitLambdaExpr(const LambdaExpr *) {} |
| 5155 | void VisitBlockExpr(const BlockExpr *) {} |
| 5156 | void VisitStmt(const Stmt *S) { |
| 5157 | if (!S) |
| 5158 | return; |
| 5159 | for (const Stmt *Child : S->children()) |
| 5160 | if (Child) |
| 5161 | Visit(S: Child); |
| 5162 | } |
| 5163 | |
| 5164 | /// Swaps list of vars with the provided one. |
| 5165 | ArrayRef<const VarDecl *> getPrivateDecls() const { return PrivateDecls; } |
| 5166 | }; |
| 5167 | } // anonymous namespace |
| 5168 | |
| 5169 | static void buildDependences(const OMPExecutableDirective &S, |
| 5170 | OMPTaskDataTy &Data) { |
| 5171 | |
| 5172 | // First look for 'omp_all_memory' and add this first. |
| 5173 | bool OmpAllMemory = false; |
| 5174 | if (llvm::any_of( |
| 5175 | Range: S.getClausesOfKind<OMPDependClause>(), P: [](const OMPDependClause *C) { |
| 5176 | return C->getDependencyKind() == OMPC_DEPEND_outallmemory || |
| 5177 | C->getDependencyKind() == OMPC_DEPEND_inoutallmemory; |
| 5178 | })) { |
| 5179 | OmpAllMemory = true; |
| 5180 | // Since both OMPC_DEPEND_outallmemory and OMPC_DEPEND_inoutallmemory are |
| 5181 | // equivalent to the runtime, always use OMPC_DEPEND_outallmemory to |
| 5182 | // simplify. |
| 5183 | OMPTaskDataTy::DependData &DD = |
| 5184 | Data.Dependences.emplace_back(Args: OMPC_DEPEND_outallmemory, |
| 5185 | /*IteratorExpr=*/Args: nullptr); |
| 5186 | // Add a nullptr Expr to simplify the codegen in emitDependData. |
| 5187 | DD.DepExprs.push_back(Elt: nullptr); |
| 5188 | } |
| 5189 | // Add remaining dependences skipping any 'out' or 'inout' if they are |
| 5190 | // overridden by 'omp_all_memory'. |
| 5191 | for (const auto *C : S.getClausesOfKind<OMPDependClause>()) { |
| 5192 | OpenMPDependClauseKind Kind = C->getDependencyKind(); |
| 5193 | if (Kind == OMPC_DEPEND_outallmemory || Kind == OMPC_DEPEND_inoutallmemory) |
| 5194 | continue; |
| 5195 | if (OmpAllMemory && (Kind == OMPC_DEPEND_out || Kind == OMPC_DEPEND_inout)) |
| 5196 | continue; |
| 5197 | OMPTaskDataTy::DependData &DD = |
| 5198 | Data.Dependences.emplace_back(Args: C->getDependencyKind(), Args: C->getModifier()); |
| 5199 | DD.DepExprs.append(in_start: C->varlist_begin(), in_end: C->varlist_end()); |
| 5200 | } |
| 5201 | } |
| 5202 | |
| 5203 | void CodeGenFunction::EmitOMPTaskBasedDirective( |
| 5204 | const OMPExecutableDirective &S, const OpenMPDirectiveKind CapturedRegion, |
| 5205 | const RegionCodeGenTy &BodyGen, const TaskGenTy &TaskGen, |
| 5206 | OMPTaskDataTy &Data) { |
| 5207 | // Emit outlined function for task construct. |
| 5208 | const CapturedStmt *CS = S.getCapturedStmt(RegionKind: CapturedRegion); |
| 5209 | auto I = CS->getCapturedDecl()->param_begin(); |
| 5210 | auto PartId = std::next(x: I); |
| 5211 | auto TaskT = std::next(x: I, n: 4); |
| 5212 | // Check if the task is final |
| 5213 | if (const auto *Clause = S.getSingleClause<OMPFinalClause>()) { |
| 5214 | // If the condition constant folds and can be elided, try to avoid emitting |
| 5215 | // the condition and the dead arm of the if/else. |
| 5216 | const Expr *Cond = Clause->getCondition(); |
| 5217 | bool CondConstant; |
| 5218 | if (ConstantFoldsToSimpleInteger(Cond, Result&: CondConstant)) |
| 5219 | Data.Final.setInt(CondConstant); |
| 5220 | else |
| 5221 | Data.Final.setPointer(EvaluateExprAsBool(E: Cond)); |
| 5222 | } else { |
| 5223 | // By default the task is not final. |
| 5224 | Data.Final.setInt(/*IntVal=*/false); |
| 5225 | } |
| 5226 | // Check if the task has 'priority' clause. |
| 5227 | if (const auto *Clause = S.getSingleClause<OMPPriorityClause>()) { |
| 5228 | const Expr *Prio = Clause->getPriority(); |
| 5229 | Data.Priority.setInt(/*IntVal=*/true); |
| 5230 | Data.Priority.setPointer(EmitScalarConversion( |
| 5231 | Src: EmitScalarExpr(E: Prio), SrcTy: Prio->getType(), |
| 5232 | DstTy: getContext().getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/1), |
| 5233 | Loc: Prio->getExprLoc())); |
| 5234 | } |
| 5235 | // The first function argument for tasks is a thread id, the second one is a |
| 5236 | // part id (0 for tied tasks, >=0 for untied task). |
| 5237 | llvm::DenseSet<const VarDecl *> EmittedAsPrivate; |
| 5238 | // Get list of private variables. |
| 5239 | for (const auto *C : S.getClausesOfKind<OMPPrivateClause>()) { |
| 5240 | auto IRef = C->varlist_begin(); |
| 5241 | for (const Expr *IInit : C->private_copies()) { |
| 5242 | const auto *OrigVD = cast<VarDecl>(Val: cast<DeclRefExpr>(Val: *IRef)->getDecl()); |
| 5243 | if (EmittedAsPrivate.insert(V: OrigVD->getCanonicalDecl()).second) { |
| 5244 | Data.PrivateVars.push_back(Elt: *IRef); |
| 5245 | Data.PrivateCopies.push_back(Elt: IInit); |
| 5246 | } |
| 5247 | ++IRef; |
| 5248 | } |
| 5249 | } |
| 5250 | EmittedAsPrivate.clear(); |
| 5251 | // Get list of firstprivate variables. |
| 5252 | for (const auto *C : S.getClausesOfKind<OMPFirstprivateClause>()) { |
| 5253 | auto IRef = C->varlist_begin(); |
| 5254 | auto IElemInitRef = C->inits().begin(); |
| 5255 | for (const Expr *IInit : C->private_copies()) { |
| 5256 | const auto *OrigVD = cast<VarDecl>(Val: cast<DeclRefExpr>(Val: *IRef)->getDecl()); |
| 5257 | if (EmittedAsPrivate.insert(V: OrigVD->getCanonicalDecl()).second) { |
| 5258 | Data.FirstprivateVars.push_back(Elt: *IRef); |
| 5259 | Data.FirstprivateCopies.push_back(Elt: IInit); |
| 5260 | Data.FirstprivateInits.push_back(Elt: *IElemInitRef); |
| 5261 | } |
| 5262 | ++IRef; |
| 5263 | ++IElemInitRef; |
| 5264 | } |
| 5265 | } |
| 5266 | // Get list of lastprivate variables (for taskloops). |
| 5267 | llvm::MapVector<const VarDecl *, const DeclRefExpr *> LastprivateDstsOrigs; |
| 5268 | for (const auto *C : S.getClausesOfKind<OMPLastprivateClause>()) { |
| 5269 | auto IRef = C->varlist_begin(); |
| 5270 | auto ID = C->destination_exprs().begin(); |
| 5271 | for (const Expr *IInit : C->private_copies()) { |
| 5272 | const auto *OrigVD = cast<VarDecl>(Val: cast<DeclRefExpr>(Val: *IRef)->getDecl()); |
| 5273 | if (EmittedAsPrivate.insert(V: OrigVD->getCanonicalDecl()).second) { |
| 5274 | Data.LastprivateVars.push_back(Elt: *IRef); |
| 5275 | Data.LastprivateCopies.push_back(Elt: IInit); |
| 5276 | } |
| 5277 | LastprivateDstsOrigs.insert( |
| 5278 | KV: std::make_pair(x: cast<VarDecl>(Val: cast<DeclRefExpr>(Val: *ID)->getDecl()), |
| 5279 | y: cast<DeclRefExpr>(Val: *IRef))); |
| 5280 | ++IRef; |
| 5281 | ++ID; |
| 5282 | } |
| 5283 | } |
| 5284 | SmallVector<const Expr *, 4> LHSs; |
| 5285 | SmallVector<const Expr *, 4> RHSs; |
| 5286 | for (const auto *C : S.getClausesOfKind<OMPReductionClause>()) { |
| 5287 | Data.ReductionVars.append(in_start: C->varlist_begin(), in_end: C->varlist_end()); |
| 5288 | Data.ReductionOrigs.append(in_start: C->varlist_begin(), in_end: C->varlist_end()); |
| 5289 | Data.ReductionCopies.append(in_start: C->privates().begin(), in_end: C->privates().end()); |
| 5290 | Data.ReductionOps.append(in_start: C->reduction_ops().begin(), |
| 5291 | in_end: C->reduction_ops().end()); |
| 5292 | LHSs.append(in_start: C->lhs_exprs().begin(), in_end: C->lhs_exprs().end()); |
| 5293 | RHSs.append(in_start: C->rhs_exprs().begin(), in_end: C->rhs_exprs().end()); |
| 5294 | } |
| 5295 | Data.Reductions = CGM.getOpenMPRuntime().emitTaskReductionInit( |
| 5296 | CGF&: *this, Loc: S.getBeginLoc(), LHSExprs: LHSs, RHSExprs: RHSs, Data); |
| 5297 | // Build list of dependences. |
| 5298 | buildDependences(S, Data); |
| 5299 | // Get list of local vars for untied tasks. |
| 5300 | if (!Data.Tied) { |
| 5301 | CheckVarsEscapingUntiedTaskDeclContext Checker; |
| 5302 | Checker.Visit(S: S.getInnermostCapturedStmt()->getCapturedStmt()); |
| 5303 | Data.PrivateLocals.append(in_start: Checker.getPrivateDecls().begin(), |
| 5304 | in_end: Checker.getPrivateDecls().end()); |
| 5305 | } |
| 5306 | auto &&CodeGen = [&Data, &S, CS, &BodyGen, &LastprivateDstsOrigs, |
| 5307 | CapturedRegion](CodeGenFunction &CGF, |
| 5308 | PrePostActionTy &Action) { |
| 5309 | llvm::MapVector<CanonicalDeclPtr<const VarDecl>, |
| 5310 | std::pair<Address, Address>> |
| 5311 | UntiedLocalVars; |
| 5312 | // Set proper addresses for generated private copies. |
| 5313 | OMPPrivateScope Scope(CGF); |
| 5314 | // Generate debug info for variables present in shared clause. |
| 5315 | if (auto *DI = CGF.getDebugInfo()) { |
| 5316 | llvm::SmallDenseMap<const VarDecl *, FieldDecl *> CaptureFields = |
| 5317 | CGF.CapturedStmtInfo->getCaptureFields(); |
| 5318 | llvm::Value *ContextValue = CGF.CapturedStmtInfo->getContextValue(); |
| 5319 | if (CaptureFields.size() && ContextValue) { |
| 5320 | unsigned CharWidth = CGF.getContext().getCharWidth(); |
| 5321 | // The shared variables are packed together as members of structure. |
| 5322 | // So the address of each shared variable can be computed by adding |
| 5323 | // offset of it (within record) to the base address of record. For each |
| 5324 | // shared variable, debug intrinsic llvm.dbg.declare is generated with |
| 5325 | // appropriate expressions (DIExpression). |
| 5326 | // Ex: |
| 5327 | // %12 = load %struct.anon*, %struct.anon** %__context.addr.i |
| 5328 | // call void @llvm.dbg.declare(metadata %struct.anon* %12, |
| 5329 | // metadata !svar1, |
| 5330 | // metadata !DIExpression(DW_OP_deref)) |
| 5331 | // call void @llvm.dbg.declare(metadata %struct.anon* %12, |
| 5332 | // metadata !svar2, |
| 5333 | // metadata !DIExpression(DW_OP_plus_uconst, 8, DW_OP_deref)) |
| 5334 | for (auto It = CaptureFields.begin(); It != CaptureFields.end(); ++It) { |
| 5335 | const VarDecl *SharedVar = It->first; |
| 5336 | RecordDecl *CaptureRecord = It->second->getParent(); |
| 5337 | const ASTRecordLayout &Layout = |
| 5338 | CGF.getContext().getASTRecordLayout(D: CaptureRecord); |
| 5339 | unsigned Offset = |
| 5340 | Layout.getFieldOffset(FieldNo: It->second->getFieldIndex()) / CharWidth; |
| 5341 | if (CGF.CGM.getCodeGenOpts().hasReducedDebugInfo()) |
| 5342 | (void)DI->EmitDeclareOfAutoVariable(Decl: SharedVar, AI: ContextValue, |
| 5343 | Builder&: CGF.Builder, UsePointerValue: false); |
| 5344 | // Get the call dbg.declare instruction we just created and update |
| 5345 | // its DIExpression to add offset to base address. |
| 5346 | auto UpdateExpr = [](llvm::LLVMContext &Ctx, auto *Declare, |
| 5347 | unsigned Offset) { |
| 5348 | SmallVector<uint64_t, 8> Ops; |
| 5349 | // Add offset to the base address if non zero. |
| 5350 | if (Offset) { |
| 5351 | Ops.push_back(Elt: llvm::dwarf::DW_OP_plus_uconst); |
| 5352 | Ops.push_back(Elt: Offset); |
| 5353 | } |
| 5354 | Ops.push_back(Elt: llvm::dwarf::DW_OP_deref); |
| 5355 | Declare->setExpression(llvm::DIExpression::get(Context&: Ctx, Elements: Ops)); |
| 5356 | }; |
| 5357 | llvm::Instruction &Last = CGF.Builder.GetInsertBlock()->back(); |
| 5358 | if (auto DDI = dyn_cast<llvm::DbgVariableIntrinsic>(Val: &Last)) |
| 5359 | UpdateExpr(DDI->getContext(), DDI, Offset); |
| 5360 | // If we're emitting using the new debug info format into a block |
| 5361 | // without a terminator, the record will be "trailing". |
| 5362 | assert(!Last.isTerminator() && "unexpected terminator"); |
| 5363 | if (auto *Marker = |
| 5364 | CGF.Builder.GetInsertBlock()->getTrailingDbgRecords()) { |
| 5365 | for (llvm::DbgVariableRecord &DVR : llvm::reverse( |
| 5366 | C: llvm::filterDbgVars(R: Marker->getDbgRecordRange()))) { |
| 5367 | UpdateExpr(Last.getContext(), &DVR, Offset); |
| 5368 | break; |
| 5369 | } |
| 5370 | } |
| 5371 | } |
| 5372 | } |
| 5373 | } |
| 5374 | llvm::SmallVector<std::pair<const VarDecl *, Address>, 16> FirstprivatePtrs; |
| 5375 | if (!Data.PrivateVars.empty() || !Data.FirstprivateVars.empty() || |
| 5376 | !Data.LastprivateVars.empty() || !Data.PrivateLocals.empty()) { |
| 5377 | enum { PrivatesParam = 2, CopyFnParam = 3 }; |
| 5378 | llvm::Value *CopyFn = CGF.Builder.CreateLoad( |
| 5379 | Addr: CGF.GetAddrOfLocalVar(VD: CS->getCapturedDecl()->getParam(i: CopyFnParam))); |
| 5380 | llvm::Value *PrivatesPtr = CGF.Builder.CreateLoad(Addr: CGF.GetAddrOfLocalVar( |
| 5381 | VD: CS->getCapturedDecl()->getParam(i: PrivatesParam))); |
| 5382 | // Map privates. |
| 5383 | llvm::SmallVector<std::pair<const VarDecl *, Address>, 16> PrivatePtrs; |
| 5384 | llvm::SmallVector<llvm::Value *, 16> CallArgs; |
| 5385 | llvm::SmallVector<llvm::Type *, 4> ParamTypes; |
| 5386 | CallArgs.push_back(Elt: PrivatesPtr); |
| 5387 | ParamTypes.push_back(Elt: PrivatesPtr->getType()); |
| 5388 | for (const Expr *E : Data.PrivateVars) { |
| 5389 | const auto *VD = cast<VarDecl>(Val: cast<DeclRefExpr>(Val: E)->getDecl()); |
| 5390 | RawAddress PrivatePtr = CGF.CreateMemTempWithoutCast( |
| 5391 | T: CGF.getContext().getPointerType(T: E->getType()), Name: ".priv.ptr.addr"); |
| 5392 | PrivatePtrs.emplace_back(Args&: VD, Args&: PrivatePtr); |
| 5393 | CallArgs.push_back(Elt: PrivatePtr.getPointer()); |
| 5394 | ParamTypes.push_back(Elt: PrivatePtr.getType()); |
| 5395 | } |
| 5396 | for (const Expr *E : Data.FirstprivateVars) { |
| 5397 | const auto *VD = cast<VarDecl>(Val: cast<DeclRefExpr>(Val: E)->getDecl()); |
| 5398 | RawAddress PrivatePtr = CGF.CreateMemTempWithoutCast( |
| 5399 | T: CGF.getContext().getPointerType(T: E->getType()), |
| 5400 | Name: ".firstpriv.ptr.addr"); |
| 5401 | PrivatePtrs.emplace_back(Args&: VD, Args&: PrivatePtr); |
| 5402 | FirstprivatePtrs.emplace_back(Args&: VD, Args&: PrivatePtr); |
| 5403 | CallArgs.push_back(Elt: PrivatePtr.getPointer()); |
| 5404 | ParamTypes.push_back(Elt: PrivatePtr.getType()); |
| 5405 | } |
| 5406 | for (const Expr *E : Data.LastprivateVars) { |
| 5407 | const auto *VD = cast<VarDecl>(Val: cast<DeclRefExpr>(Val: E)->getDecl()); |
| 5408 | RawAddress PrivatePtr = CGF.CreateMemTempWithoutCast( |
| 5409 | T: CGF.getContext().getPointerType(T: E->getType()), |
| 5410 | Name: ".lastpriv.ptr.addr"); |
| 5411 | PrivatePtrs.emplace_back(Args&: VD, Args&: PrivatePtr); |
| 5412 | CallArgs.push_back(Elt: PrivatePtr.getPointer()); |
| 5413 | ParamTypes.push_back(Elt: PrivatePtr.getType()); |
| 5414 | } |
| 5415 | for (const VarDecl *VD : Data.PrivateLocals) { |
| 5416 | QualType Ty = VD->getType().getNonReferenceType(); |
| 5417 | if (VD->getType()->isLValueReferenceType()) |
| 5418 | Ty = CGF.getContext().getPointerType(T: Ty); |
| 5419 | if (isAllocatableDecl(VD)) |
| 5420 | Ty = CGF.getContext().getPointerType(T: Ty); |
| 5421 | RawAddress PrivatePtr = CGF.CreateMemTempWithoutCast( |
| 5422 | T: CGF.getContext().getPointerType(T: Ty), Name: ".local.ptr.addr"); |
| 5423 | auto Result = UntiedLocalVars.insert( |
| 5424 | KV: std::make_pair(x&: VD, y: std::make_pair(x&: PrivatePtr, y: Address::invalid()))); |
| 5425 | // If key exists update in place. |
| 5426 | if (Result.second == false) |
| 5427 | *Result.first = std::make_pair( |
| 5428 | x&: VD, y: std::make_pair(x&: PrivatePtr, y: Address::invalid())); |
| 5429 | CallArgs.push_back(Elt: PrivatePtr.getPointer()); |
| 5430 | ParamTypes.push_back(Elt: PrivatePtr.getType()); |
| 5431 | } |
| 5432 | auto *CopyFnTy = llvm::FunctionType::get(Result: CGF.Builder.getVoidTy(), |
| 5433 | Params: ParamTypes, /*isVarArg=*/false); |
| 5434 | CGF.CGM.getOpenMPRuntime().emitOutlinedFunctionCall( |
| 5435 | CGF, Loc: S.getBeginLoc(), OutlinedFn: {CopyFnTy, CopyFn}, Args: CallArgs); |
| 5436 | for (const auto &Pair : LastprivateDstsOrigs) { |
| 5437 | const auto *OrigVD = cast<VarDecl>(Val: Pair.second->getDecl()); |
| 5438 | DeclRefExpr DRE(CGF.getContext(), const_cast<VarDecl *>(OrigVD), |
| 5439 | /*RefersToEnclosingVariableOrCapture=*/ |
| 5440 | CGF.CapturedStmtInfo->lookup(VD: OrigVD) != nullptr, |
| 5441 | Pair.second->getType(), VK_LValue, |
| 5442 | Pair.second->getExprLoc()); |
| 5443 | Scope.addPrivate(LocalVD: Pair.first, Addr: CGF.EmitLValue(E: &DRE).getAddress()); |
| 5444 | } |
| 5445 | for (const auto &Pair : PrivatePtrs) { |
| 5446 | Address Replacement = Address( |
| 5447 | CGF.Builder.CreateLoad(Addr: Pair.second), |
| 5448 | CGF.ConvertTypeForMem(T: Pair.first->getType().getNonReferenceType()), |
| 5449 | CGF.getContext().getDeclAlign(D: Pair.first)); |
| 5450 | Scope.addPrivate(LocalVD: Pair.first, Addr: Replacement); |
| 5451 | if (auto *DI = CGF.getDebugInfo()) |
| 5452 | if (CGF.CGM.getCodeGenOpts().hasReducedDebugInfo()) |
| 5453 | (void)DI->EmitDeclareOfAutoVariable( |
| 5454 | Decl: Pair.first, AI: Pair.second.getBasePointer(), Builder&: CGF.Builder, |
| 5455 | /*UsePointerValue*/ true); |
| 5456 | } |
| 5457 | // Adjust mapping for internal locals by mapping actual memory instead of |
| 5458 | // a pointer to this memory. |
| 5459 | for (auto &Pair : UntiedLocalVars) { |
| 5460 | QualType VDType = Pair.first->getType().getNonReferenceType(); |
| 5461 | if (Pair.first->getType()->isLValueReferenceType()) |
| 5462 | VDType = CGF.getContext().getPointerType(T: VDType); |
| 5463 | if (isAllocatableDecl(VD: Pair.first)) { |
| 5464 | llvm::Value *Ptr = CGF.Builder.CreateLoad(Addr: Pair.second.first); |
| 5465 | Address Replacement( |
| 5466 | Ptr, |
| 5467 | CGF.ConvertTypeForMem(T: CGF.getContext().getPointerType(T: VDType)), |
| 5468 | CGF.getPointerAlign()); |
| 5469 | Pair.second.first = Replacement; |
| 5470 | Ptr = CGF.Builder.CreateLoad(Addr: Replacement); |
| 5471 | Replacement = Address(Ptr, CGF.ConvertTypeForMem(T: VDType), |
| 5472 | CGF.getContext().getDeclAlign(D: Pair.first)); |
| 5473 | Pair.second.second = Replacement; |
| 5474 | } else { |
| 5475 | llvm::Value *Ptr = CGF.Builder.CreateLoad(Addr: Pair.second.first); |
| 5476 | Address Replacement(Ptr, CGF.ConvertTypeForMem(T: VDType), |
| 5477 | CGF.getContext().getDeclAlign(D: Pair.first)); |
| 5478 | Pair.second.first = Replacement; |
| 5479 | } |
| 5480 | } |
| 5481 | } |
| 5482 | if (Data.Reductions) { |
| 5483 | OMPPrivateScope FirstprivateScope(CGF); |
| 5484 | for (const auto &Pair : FirstprivatePtrs) { |
| 5485 | Address Replacement( |
| 5486 | CGF.Builder.CreateLoad(Addr: Pair.second), |
| 5487 | CGF.ConvertTypeForMem(T: Pair.first->getType().getNonReferenceType()), |
| 5488 | CGF.getContext().getDeclAlign(D: Pair.first)); |
| 5489 | FirstprivateScope.addPrivate(LocalVD: Pair.first, Addr: Replacement); |
| 5490 | } |
| 5491 | (void)FirstprivateScope.Privatize(); |
| 5492 | OMPLexicalScope LexScope(CGF, S, CapturedRegion); |
| 5493 | ReductionCodeGen RedCG(Data.ReductionVars, Data.ReductionVars, |
| 5494 | Data.ReductionCopies, Data.ReductionOps); |
| 5495 | llvm::Value *ReductionsPtr = CGF.Builder.CreateLoad( |
| 5496 | Addr: CGF.GetAddrOfLocalVar(VD: CS->getCapturedDecl()->getParam(i: 9))); |
| 5497 | for (unsigned Cnt = 0, E = Data.ReductionVars.size(); Cnt < E; ++Cnt) { |
| 5498 | RedCG.emitSharedOrigLValue(CGF, N: Cnt); |
| 5499 | RedCG.emitAggregateType(CGF, N: Cnt); |
| 5500 | // FIXME: This must removed once the runtime library is fixed. |
| 5501 | // Emit required threadprivate variables for |
| 5502 | // initializer/combiner/finalizer. |
| 5503 | CGF.CGM.getOpenMPRuntime().emitTaskReductionFixups(CGF, Loc: S.getBeginLoc(), |
| 5504 | RCG&: RedCG, N: Cnt); |
| 5505 | Address Replacement = CGF.CGM.getOpenMPRuntime().getTaskReductionItem( |
| 5506 | CGF, Loc: S.getBeginLoc(), ReductionsPtr, SharedLVal: RedCG.getSharedLValue(N: Cnt)); |
| 5507 | Replacement = Address( |
| 5508 | CGF.EmitScalarConversion(Src: Replacement.emitRawPointer(CGF), |
| 5509 | SrcTy: CGF.getContext().VoidPtrTy, |
| 5510 | DstTy: CGF.getContext().getPointerType( |
| 5511 | T: Data.ReductionCopies[Cnt]->getType()), |
| 5512 | Loc: Data.ReductionCopies[Cnt]->getExprLoc()), |
| 5513 | CGF.ConvertTypeForMem(T: Data.ReductionCopies[Cnt]->getType()), |
| 5514 | Replacement.getAlignment()); |
| 5515 | Replacement = RedCG.adjustPrivateAddress(CGF, N: Cnt, PrivateAddr: Replacement); |
| 5516 | Scope.addPrivate(LocalVD: RedCG.getBaseDecl(N: Cnt), Addr: Replacement); |
| 5517 | } |
| 5518 | } |
| 5519 | // Privatize all private variables except for in_reduction items. |
| 5520 | (void)Scope.Privatize(); |
| 5521 | SmallVector<const Expr *, 4> InRedVars; |
| 5522 | SmallVector<const Expr *, 4> InRedPrivs; |
| 5523 | SmallVector<const Expr *, 4> InRedOps; |
| 5524 | SmallVector<const Expr *, 4> TaskgroupDescriptors; |
| 5525 | for (const auto *C : S.getClausesOfKind<OMPInReductionClause>()) { |
| 5526 | auto IPriv = C->privates().begin(); |
| 5527 | auto IRed = C->reduction_ops().begin(); |
| 5528 | auto ITD = C->taskgroup_descriptors().begin(); |
| 5529 | for (const Expr *Ref : C->varlist()) { |
| 5530 | InRedVars.emplace_back(Args&: Ref); |
| 5531 | InRedPrivs.emplace_back(Args: *IPriv); |
| 5532 | InRedOps.emplace_back(Args: *IRed); |
| 5533 | TaskgroupDescriptors.emplace_back(Args: *ITD); |
| 5534 | std::advance(i&: IPriv, n: 1); |
| 5535 | std::advance(i&: IRed, n: 1); |
| 5536 | std::advance(i&: ITD, n: 1); |
| 5537 | } |
| 5538 | } |
| 5539 | // Privatize in_reduction items here, because taskgroup descriptors must be |
| 5540 | // privatized earlier. |
| 5541 | OMPPrivateScope InRedScope(CGF); |
| 5542 | if (!InRedVars.empty()) { |
| 5543 | ReductionCodeGen RedCG(InRedVars, InRedVars, InRedPrivs, InRedOps); |
| 5544 | for (unsigned Cnt = 0, E = InRedVars.size(); Cnt < E; ++Cnt) { |
| 5545 | RedCG.emitSharedOrigLValue(CGF, N: Cnt); |
| 5546 | RedCG.emitAggregateType(CGF, N: Cnt); |
| 5547 | // The taskgroup descriptor variable is always implicit firstprivate and |
| 5548 | // privatized already during processing of the firstprivates. |
| 5549 | // FIXME: This must removed once the runtime library is fixed. |
| 5550 | // Emit required threadprivate variables for |
| 5551 | // initializer/combiner/finalizer. |
| 5552 | CGF.CGM.getOpenMPRuntime().emitTaskReductionFixups(CGF, Loc: S.getBeginLoc(), |
| 5553 | RCG&: RedCG, N: Cnt); |
| 5554 | llvm::Value *ReductionsPtr; |
| 5555 | if (const Expr *TRExpr = TaskgroupDescriptors[Cnt]) { |
| 5556 | ReductionsPtr = CGF.EmitLoadOfScalar(lvalue: CGF.EmitLValue(E: TRExpr), |
| 5557 | Loc: TRExpr->getExprLoc()); |
| 5558 | } else { |
| 5559 | ReductionsPtr = llvm::ConstantPointerNull::get(T: CGF.VoidPtrTy); |
| 5560 | } |
| 5561 | Address Replacement = CGF.CGM.getOpenMPRuntime().getTaskReductionItem( |
| 5562 | CGF, Loc: S.getBeginLoc(), ReductionsPtr, SharedLVal: RedCG.getSharedLValue(N: Cnt)); |
| 5563 | Replacement = Address( |
| 5564 | CGF.EmitScalarConversion( |
| 5565 | Src: Replacement.emitRawPointer(CGF), SrcTy: CGF.getContext().VoidPtrTy, |
| 5566 | DstTy: CGF.getContext().getPointerType(T: InRedPrivs[Cnt]->getType()), |
| 5567 | Loc: InRedPrivs[Cnt]->getExprLoc()), |
| 5568 | CGF.ConvertTypeForMem(T: InRedPrivs[Cnt]->getType()), |
| 5569 | Replacement.getAlignment()); |
| 5570 | Replacement = RedCG.adjustPrivateAddress(CGF, N: Cnt, PrivateAddr: Replacement); |
| 5571 | InRedScope.addPrivate(LocalVD: RedCG.getBaseDecl(N: Cnt), Addr: Replacement); |
| 5572 | } |
| 5573 | } |
| 5574 | (void)InRedScope.Privatize(); |
| 5575 | |
| 5576 | CGOpenMPRuntime::UntiedTaskLocalDeclsRAII LocalVarsScope(CGF, |
| 5577 | UntiedLocalVars); |
| 5578 | Action.Enter(CGF); |
| 5579 | BodyGen(CGF); |
| 5580 | }; |
| 5581 | OpenMPDirectiveKind EKind = getEffectiveDirectiveKind(S); |
| 5582 | llvm::Function *OutlinedFn = CGM.getOpenMPRuntime().emitTaskOutlinedFunction( |
| 5583 | D: S, ThreadIDVar: *I, PartIDVar: *PartId, TaskTVar: *TaskT, InnermostKind: EKind, CodeGen, Tied: Data.Tied, NumberOfParts&: Data.NumberOfParts); |
| 5584 | OMPLexicalScope Scope(*this, S, std::nullopt, |
| 5585 | !isOpenMPParallelDirective(DKind: EKind) && |
| 5586 | !isOpenMPSimdDirective(DKind: EKind)); |
| 5587 | TaskGen(*this, OutlinedFn, Data); |
| 5588 | } |
| 5589 | |
| 5590 | static ImplicitParamDecl * |
| 5591 | createImplicitFirstprivateForType(ASTContext &C, OMPTaskDataTy &Data, |
| 5592 | QualType Ty, CapturedDecl *CD, |
| 5593 | SourceLocation Loc) { |
| 5594 | auto *OrigVD = ImplicitParamDecl::Create(C, DC: CD, IdLoc: Loc, /*Id=*/nullptr, T: Ty, |
| 5595 | ParamKind: ImplicitParamKind::Other); |
| 5596 | auto *OrigRef = DeclRefExpr::Create( |
| 5597 | Context: C, QualifierLoc: NestedNameSpecifierLoc(), TemplateKWLoc: SourceLocation(), D: OrigVD, |
| 5598 | /*RefersToEnclosingVariableOrCapture=*/false, NameLoc: Loc, T: Ty, VK: VK_LValue); |
| 5599 | auto *PrivateVD = ImplicitParamDecl::Create(C, DC: CD, IdLoc: Loc, /*Id=*/nullptr, T: Ty, |
| 5600 | ParamKind: ImplicitParamKind::Other); |
| 5601 | auto *PrivateRef = DeclRefExpr::Create( |
| 5602 | Context: C, QualifierLoc: NestedNameSpecifierLoc(), TemplateKWLoc: SourceLocation(), D: PrivateVD, |
| 5603 | /*RefersToEnclosingVariableOrCapture=*/false, NameLoc: Loc, T: Ty, VK: VK_LValue); |
| 5604 | QualType ElemType = C.getBaseElementType(QT: Ty); |
| 5605 | auto *InitVD = ImplicitParamDecl::Create(C, DC: CD, IdLoc: Loc, /*Id=*/nullptr, T: ElemType, |
| 5606 | ParamKind: ImplicitParamKind::Other); |
| 5607 | auto *InitRef = DeclRefExpr::Create( |
| 5608 | Context: C, QualifierLoc: NestedNameSpecifierLoc(), TemplateKWLoc: SourceLocation(), D: InitVD, |
| 5609 | /*RefersToEnclosingVariableOrCapture=*/false, NameLoc: Loc, T: ElemType, VK: VK_LValue); |
| 5610 | PrivateVD->setInitStyle(VarDecl::CInit); |
| 5611 | PrivateVD->setInit(ImplicitCastExpr::Create(Context: C, T: ElemType, Kind: CK_LValueToRValue, |
| 5612 | Operand: InitRef, /*BasePath=*/nullptr, |
| 5613 | Cat: VK_PRValue, FPO: FPOptionsOverride())); |
| 5614 | Data.FirstprivateVars.emplace_back(Args&: OrigRef); |
| 5615 | Data.FirstprivateCopies.emplace_back(Args&: PrivateRef); |
| 5616 | Data.FirstprivateInits.emplace_back(Args&: InitRef); |
| 5617 | return OrigVD; |
| 5618 | } |
| 5619 | |
| 5620 | void CodeGenFunction::EmitOMPTargetTaskBasedDirective( |
| 5621 | const OMPExecutableDirective &S, const RegionCodeGenTy &BodyGen, |
| 5622 | OMPTargetDataInfo &InputInfo) { |
| 5623 | // Emit outlined function for task construct. |
| 5624 | const CapturedStmt *CS = S.getCapturedStmt(RegionKind: OMPD_task); |
| 5625 | Address CapturedStruct = GenerateCapturedStmtArgument(S: *CS); |
| 5626 | CanQualType SharedsTy = |
| 5627 | getContext().getCanonicalTagType(TD: CS->getCapturedRecordDecl()); |
| 5628 | auto I = CS->getCapturedDecl()->param_begin(); |
| 5629 | auto PartId = std::next(x: I); |
| 5630 | auto TaskT = std::next(x: I, n: 4); |
| 5631 | OMPTaskDataTy Data; |
| 5632 | // The task is not final. |
| 5633 | Data.Final.setInt(/*IntVal=*/false); |
| 5634 | // Get list of firstprivate variables. |
| 5635 | for (const auto *C : S.getClausesOfKind<OMPFirstprivateClause>()) { |
| 5636 | auto IRef = C->varlist_begin(); |
| 5637 | auto IElemInitRef = C->inits().begin(); |
| 5638 | for (auto *IInit : C->private_copies()) { |
| 5639 | Data.FirstprivateVars.push_back(Elt: *IRef); |
| 5640 | Data.FirstprivateCopies.push_back(Elt: IInit); |
| 5641 | Data.FirstprivateInits.push_back(Elt: *IElemInitRef); |
| 5642 | ++IRef; |
| 5643 | ++IElemInitRef; |
| 5644 | } |
| 5645 | } |
| 5646 | SmallVector<const Expr *, 4> LHSs; |
| 5647 | SmallVector<const Expr *, 4> RHSs; |
| 5648 | for (const auto *C : S.getClausesOfKind<OMPInReductionClause>()) { |
| 5649 | Data.ReductionVars.append(in_start: C->varlist_begin(), in_end: C->varlist_end()); |
| 5650 | Data.ReductionOrigs.append(in_start: C->varlist_begin(), in_end: C->varlist_end()); |
| 5651 | Data.ReductionCopies.append(in_start: C->privates().begin(), in_end: C->privates().end()); |
| 5652 | Data.ReductionOps.append(in_start: C->reduction_ops().begin(), |
| 5653 | in_end: C->reduction_ops().end()); |
| 5654 | LHSs.append(in_start: C->lhs_exprs().begin(), in_end: C->lhs_exprs().end()); |
| 5655 | RHSs.append(in_start: C->rhs_exprs().begin(), in_end: C->rhs_exprs().end()); |
| 5656 | } |
| 5657 | OMPPrivateScope TargetScope(*this); |
| 5658 | VarDecl *BPVD = nullptr; |
| 5659 | VarDecl *PVD = nullptr; |
| 5660 | VarDecl *SVD = nullptr; |
| 5661 | VarDecl *MVD = nullptr; |
| 5662 | if (InputInfo.NumberOfTargetItems > 0) { |
| 5663 | auto *CD = CapturedDecl::Create( |
| 5664 | C&: getContext(), DC: getContext().getTranslationUnitDecl(), /*NumParams=*/0); |
| 5665 | llvm::APInt ArrSize(/*numBits=*/32, InputInfo.NumberOfTargetItems); |
| 5666 | QualType BaseAndPointerAndMapperType = getContext().getConstantArrayType( |
| 5667 | EltTy: getContext().VoidPtrTy, ArySize: ArrSize, SizeExpr: nullptr, ASM: ArraySizeModifier::Normal, |
| 5668 | /*IndexTypeQuals=*/0); |
| 5669 | BPVD = createImplicitFirstprivateForType( |
| 5670 | C&: getContext(), Data, Ty: BaseAndPointerAndMapperType, CD, Loc: S.getBeginLoc()); |
| 5671 | PVD = createImplicitFirstprivateForType( |
| 5672 | C&: getContext(), Data, Ty: BaseAndPointerAndMapperType, CD, Loc: S.getBeginLoc()); |
| 5673 | QualType SizesType = getContext().getConstantArrayType( |
| 5674 | EltTy: getContext().getIntTypeForBitwidth(/*DestWidth=*/64, /*Signed=*/1), |
| 5675 | ArySize: ArrSize, SizeExpr: nullptr, ASM: ArraySizeModifier::Normal, |
| 5676 | /*IndexTypeQuals=*/0); |
| 5677 | SVD = createImplicitFirstprivateForType(C&: getContext(), Data, Ty: SizesType, CD, |
| 5678 | Loc: S.getBeginLoc()); |
| 5679 | TargetScope.addPrivate(LocalVD: BPVD, Addr: InputInfo.BasePointersArray); |
| 5680 | TargetScope.addPrivate(LocalVD: PVD, Addr: InputInfo.PointersArray); |
| 5681 | TargetScope.addPrivate(LocalVD: SVD, Addr: InputInfo.SizesArray); |
| 5682 | // If there is no user-defined mapper, the mapper array will be nullptr. In |
| 5683 | // this case, we don't need to privatize it. |
| 5684 | if (!isa_and_nonnull<llvm::ConstantPointerNull>( |
| 5685 | Val: InputInfo.MappersArray.emitRawPointer(CGF&: *this))) { |
| 5686 | MVD = createImplicitFirstprivateForType( |
| 5687 | C&: getContext(), Data, Ty: BaseAndPointerAndMapperType, CD, Loc: S.getBeginLoc()); |
| 5688 | TargetScope.addPrivate(LocalVD: MVD, Addr: InputInfo.MappersArray); |
| 5689 | } |
| 5690 | } |
| 5691 | (void)TargetScope.Privatize(); |
| 5692 | buildDependences(S, Data); |
| 5693 | OpenMPDirectiveKind EKind = getEffectiveDirectiveKind(S); |
| 5694 | auto &&CodeGen = [&Data, &S, CS, &BodyGen, BPVD, PVD, SVD, MVD, EKind, |
| 5695 | &InputInfo](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 5696 | // Set proper addresses for generated private copies. |
| 5697 | OMPPrivateScope Scope(CGF); |
| 5698 | if (!Data.FirstprivateVars.empty()) { |
| 5699 | enum { PrivatesParam = 2, CopyFnParam = 3 }; |
| 5700 | llvm::Value *CopyFn = CGF.Builder.CreateLoad( |
| 5701 | Addr: CGF.GetAddrOfLocalVar(VD: CS->getCapturedDecl()->getParam(i: CopyFnParam))); |
| 5702 | llvm::Value *PrivatesPtr = CGF.Builder.CreateLoad(Addr: CGF.GetAddrOfLocalVar( |
| 5703 | VD: CS->getCapturedDecl()->getParam(i: PrivatesParam))); |
| 5704 | // Map privates. |
| 5705 | llvm::SmallVector<std::pair<const VarDecl *, Address>, 16> PrivatePtrs; |
| 5706 | llvm::SmallVector<llvm::Value *, 16> CallArgs; |
| 5707 | llvm::SmallVector<llvm::Type *, 4> ParamTypes; |
| 5708 | CallArgs.push_back(Elt: PrivatesPtr); |
| 5709 | ParamTypes.push_back(Elt: PrivatesPtr->getType()); |
| 5710 | for (const Expr *E : Data.FirstprivateVars) { |
| 5711 | const auto *VD = cast<VarDecl>(Val: cast<DeclRefExpr>(Val: E)->getDecl()); |
| 5712 | RawAddress PrivatePtr = CGF.CreateMemTempWithoutCast( |
| 5713 | T: CGF.getContext().getPointerType(T: E->getType()), |
| 5714 | Name: ".firstpriv.ptr.addr"); |
| 5715 | PrivatePtrs.emplace_back(Args&: VD, Args&: PrivatePtr); |
| 5716 | CallArgs.push_back(Elt: PrivatePtr.getPointer()); |
| 5717 | ParamTypes.push_back(Elt: PrivatePtr.getType()); |
| 5718 | } |
| 5719 | auto *CopyFnTy = llvm::FunctionType::get(Result: CGF.Builder.getVoidTy(), |
| 5720 | Params: ParamTypes, /*isVarArg=*/false); |
| 5721 | CGF.CGM.getOpenMPRuntime().emitOutlinedFunctionCall( |
| 5722 | CGF, Loc: S.getBeginLoc(), OutlinedFn: {CopyFnTy, CopyFn}, Args: CallArgs); |
| 5723 | for (const auto &Pair : PrivatePtrs) { |
| 5724 | Address Replacement( |
| 5725 | CGF.Builder.CreateLoad(Addr: Pair.second), |
| 5726 | CGF.ConvertTypeForMem(T: Pair.first->getType().getNonReferenceType()), |
| 5727 | CGF.getContext().getDeclAlign(D: Pair.first)); |
| 5728 | Scope.addPrivate(LocalVD: Pair.first, Addr: Replacement); |
| 5729 | } |
| 5730 | } |
| 5731 | CGF.processInReduction(S, Data, CGF, CS, Scope); |
| 5732 | if (InputInfo.NumberOfTargetItems > 0) { |
| 5733 | InputInfo.BasePointersArray = CGF.Builder.CreateConstArrayGEP( |
| 5734 | Addr: CGF.GetAddrOfLocalVar(VD: BPVD), /*Index=*/0); |
| 5735 | InputInfo.PointersArray = CGF.Builder.CreateConstArrayGEP( |
| 5736 | Addr: CGF.GetAddrOfLocalVar(VD: PVD), /*Index=*/0); |
| 5737 | InputInfo.SizesArray = CGF.Builder.CreateConstArrayGEP( |
| 5738 | Addr: CGF.GetAddrOfLocalVar(VD: SVD), /*Index=*/0); |
| 5739 | // If MVD is nullptr, the mapper array is not privatized |
| 5740 | if (MVD) |
| 5741 | InputInfo.MappersArray = CGF.Builder.CreateConstArrayGEP( |
| 5742 | Addr: CGF.GetAddrOfLocalVar(VD: MVD), /*Index=*/0); |
| 5743 | } |
| 5744 | |
| 5745 | Action.Enter(CGF); |
| 5746 | OMPLexicalScope LexScope(CGF, S, OMPD_task, /*EmitPreInitStmt=*/false); |
| 5747 | auto *TL = S.getSingleClause<OMPThreadLimitClause>(); |
| 5748 | if (CGF.CGM.getLangOpts().OpenMP >= 51 && |
| 5749 | needsTaskBasedThreadLimit(DKind: EKind) && TL) { |
| 5750 | // Emit __kmpc_set_thread_limit() to set the thread_limit for the task |
| 5751 | // enclosing this target region. This will indirectly set the thread_limit |
| 5752 | // for every applicable construct within target region. |
| 5753 | CGF.CGM.getOpenMPRuntime().emitThreadLimitClause( |
| 5754 | CGF, ThreadLimit: TL->getThreadLimit().front(), Loc: S.getBeginLoc()); |
| 5755 | } |
| 5756 | BodyGen(CGF); |
| 5757 | }; |
| 5758 | llvm::Function *OutlinedFn = CGM.getOpenMPRuntime().emitTaskOutlinedFunction( |
| 5759 | D: S, ThreadIDVar: *I, PartIDVar: *PartId, TaskTVar: *TaskT, InnermostKind: EKind, CodeGen, /*Tied=*/true, |
| 5760 | NumberOfParts&: Data.NumberOfParts); |
| 5761 | llvm::APInt TrueOrFalse(32, S.hasClausesOfKind<OMPNowaitClause>() ? 1 : 0); |
| 5762 | IntegerLiteral IfCond(getContext(), TrueOrFalse, |
| 5763 | getContext().getIntTypeForBitwidth(DestWidth: 32, /*Signed=*/0), |
| 5764 | SourceLocation()); |
| 5765 | CGM.getOpenMPRuntime().emitTaskCall(CGF&: *this, Loc: S.getBeginLoc(), D: S, TaskFunction: OutlinedFn, |
| 5766 | SharedsTy, Shareds: CapturedStruct, IfCond: &IfCond, Data); |
| 5767 | } |
| 5768 | |
| 5769 | void CodeGenFunction::processInReduction(const OMPExecutableDirective &S, |
| 5770 | OMPTaskDataTy &Data, |
| 5771 | CodeGenFunction &CGF, |
| 5772 | const CapturedStmt *CS, |
| 5773 | OMPPrivateScope &Scope) { |
| 5774 | OpenMPDirectiveKind EKind = getEffectiveDirectiveKind(S); |
| 5775 | if (Data.Reductions) { |
| 5776 | OpenMPDirectiveKind CapturedRegion = EKind; |
| 5777 | OMPLexicalScope LexScope(CGF, S, CapturedRegion); |
| 5778 | ReductionCodeGen RedCG(Data.ReductionVars, Data.ReductionVars, |
| 5779 | Data.ReductionCopies, Data.ReductionOps); |
| 5780 | llvm::Value *ReductionsPtr = CGF.Builder.CreateLoad( |
| 5781 | Addr: CGF.GetAddrOfLocalVar(VD: CS->getCapturedDecl()->getParam(i: 4))); |
| 5782 | for (unsigned Cnt = 0, E = Data.ReductionVars.size(); Cnt < E; ++Cnt) { |
| 5783 | RedCG.emitSharedOrigLValue(CGF, N: Cnt); |
| 5784 | RedCG.emitAggregateType(CGF, N: Cnt); |
| 5785 | // FIXME: This must removed once the runtime library is fixed. |
| 5786 | // Emit required threadprivate variables for |
| 5787 | // initializer/combiner/finalizer. |
| 5788 | CGF.CGM.getOpenMPRuntime().emitTaskReductionFixups(CGF, Loc: S.getBeginLoc(), |
| 5789 | RCG&: RedCG, N: Cnt); |
| 5790 | Address Replacement = CGF.CGM.getOpenMPRuntime().getTaskReductionItem( |
| 5791 | CGF, Loc: S.getBeginLoc(), ReductionsPtr, SharedLVal: RedCG.getSharedLValue(N: Cnt)); |
| 5792 | Replacement = Address( |
| 5793 | CGF.EmitScalarConversion(Src: Replacement.emitRawPointer(CGF), |
| 5794 | SrcTy: CGF.getContext().VoidPtrTy, |
| 5795 | DstTy: CGF.getContext().getPointerType( |
| 5796 | T: Data.ReductionCopies[Cnt]->getType()), |
| 5797 | Loc: Data.ReductionCopies[Cnt]->getExprLoc()), |
| 5798 | CGF.ConvertTypeForMem(T: Data.ReductionCopies[Cnt]->getType()), |
| 5799 | Replacement.getAlignment()); |
| 5800 | Replacement = RedCG.adjustPrivateAddress(CGF, N: Cnt, PrivateAddr: Replacement); |
| 5801 | Scope.addPrivate(LocalVD: RedCG.getBaseDecl(N: Cnt), Addr: Replacement); |
| 5802 | } |
| 5803 | } |
| 5804 | (void)Scope.Privatize(); |
| 5805 | SmallVector<const Expr *, 4> InRedVars; |
| 5806 | SmallVector<const Expr *, 4> InRedPrivs; |
| 5807 | SmallVector<const Expr *, 4> InRedOps; |
| 5808 | SmallVector<const Expr *, 4> TaskgroupDescriptors; |
| 5809 | for (const auto *C : S.getClausesOfKind<OMPInReductionClause>()) { |
| 5810 | auto IPriv = C->privates().begin(); |
| 5811 | auto IRed = C->reduction_ops().begin(); |
| 5812 | auto ITD = C->taskgroup_descriptors().begin(); |
| 5813 | for (const Expr *Ref : C->varlist()) { |
| 5814 | InRedVars.emplace_back(Args&: Ref); |
| 5815 | InRedPrivs.emplace_back(Args: *IPriv); |
| 5816 | InRedOps.emplace_back(Args: *IRed); |
| 5817 | TaskgroupDescriptors.emplace_back(Args: *ITD); |
| 5818 | std::advance(i&: IPriv, n: 1); |
| 5819 | std::advance(i&: IRed, n: 1); |
| 5820 | std::advance(i&: ITD, n: 1); |
| 5821 | } |
| 5822 | } |
| 5823 | OMPPrivateScope InRedScope(CGF); |
| 5824 | if (!InRedVars.empty()) { |
| 5825 | ReductionCodeGen RedCG(InRedVars, InRedVars, InRedPrivs, InRedOps); |
| 5826 | for (unsigned Cnt = 0, E = InRedVars.size(); Cnt < E; ++Cnt) { |
| 5827 | RedCG.emitSharedOrigLValue(CGF, N: Cnt); |
| 5828 | RedCG.emitAggregateType(CGF, N: Cnt); |
| 5829 | // FIXME: This must removed once the runtime library is fixed. |
| 5830 | // Emit required threadprivate variables for |
| 5831 | // initializer/combiner/finalizer. |
| 5832 | CGF.CGM.getOpenMPRuntime().emitTaskReductionFixups(CGF, Loc: S.getBeginLoc(), |
| 5833 | RCG&: RedCG, N: Cnt); |
| 5834 | llvm::Value *ReductionsPtr; |
| 5835 | if (const Expr *TRExpr = TaskgroupDescriptors[Cnt]) { |
| 5836 | ReductionsPtr = |
| 5837 | CGF.EmitLoadOfScalar(lvalue: CGF.EmitLValue(E: TRExpr), Loc: TRExpr->getExprLoc()); |
| 5838 | } else { |
| 5839 | ReductionsPtr = llvm::ConstantPointerNull::get(T: CGF.VoidPtrTy); |
| 5840 | } |
| 5841 | Address Replacement = CGF.CGM.getOpenMPRuntime().getTaskReductionItem( |
| 5842 | CGF, Loc: S.getBeginLoc(), ReductionsPtr, SharedLVal: RedCG.getSharedLValue(N: Cnt)); |
| 5843 | Replacement = Address( |
| 5844 | CGF.EmitScalarConversion( |
| 5845 | Src: Replacement.emitRawPointer(CGF), SrcTy: CGF.getContext().VoidPtrTy, |
| 5846 | DstTy: CGF.getContext().getPointerType(T: InRedPrivs[Cnt]->getType()), |
| 5847 | Loc: InRedPrivs[Cnt]->getExprLoc()), |
| 5848 | CGF.ConvertTypeForMem(T: InRedPrivs[Cnt]->getType()), |
| 5849 | Replacement.getAlignment()); |
| 5850 | Replacement = RedCG.adjustPrivateAddress(CGF, N: Cnt, PrivateAddr: Replacement); |
| 5851 | InRedScope.addPrivate(LocalVD: RedCG.getBaseDecl(N: Cnt), Addr: Replacement); |
| 5852 | } |
| 5853 | } |
| 5854 | (void)InRedScope.Privatize(); |
| 5855 | } |
| 5856 | |
| 5857 | void CodeGenFunction::EmitOMPTaskDirective(const OMPTaskDirective &S) { |
| 5858 | // Emit outlined function for task construct. |
| 5859 | const CapturedStmt *CS = S.getCapturedStmt(RegionKind: OMPD_task); |
| 5860 | Address CapturedStruct = GenerateCapturedStmtArgument(S: *CS); |
| 5861 | CanQualType SharedsTy = |
| 5862 | getContext().getCanonicalTagType(TD: CS->getCapturedRecordDecl()); |
| 5863 | const Expr *IfCond = nullptr; |
| 5864 | for (const auto *C : S.getClausesOfKind<OMPIfClause>()) { |
| 5865 | if (C->getNameModifier() == OMPD_unknown || |
| 5866 | C->getNameModifier() == OMPD_task) { |
| 5867 | IfCond = C->getCondition(); |
| 5868 | break; |
| 5869 | } |
| 5870 | } |
| 5871 | |
| 5872 | OMPTaskDataTy Data; |
| 5873 | // Check if we should emit tied or untied task. |
| 5874 | Data.Tied = !S.getSingleClause<OMPUntiedClause>(); |
| 5875 | auto &&BodyGen = [CS](CodeGenFunction &CGF, PrePostActionTy &) { |
| 5876 | CGF.EmitStmt(S: CS->getCapturedStmt()); |
| 5877 | }; |
| 5878 | auto &&TaskGen = [&S, SharedsTy, CapturedStruct, |
| 5879 | IfCond](CodeGenFunction &CGF, llvm::Function *OutlinedFn, |
| 5880 | const OMPTaskDataTy &Data) { |
| 5881 | CGF.CGM.getOpenMPRuntime().emitTaskCall(CGF, Loc: S.getBeginLoc(), D: S, TaskFunction: OutlinedFn, |
| 5882 | SharedsTy, Shareds: CapturedStruct, IfCond, |
| 5883 | Data); |
| 5884 | }; |
| 5885 | auto LPCRegion = |
| 5886 | CGOpenMPRuntime::LastprivateConditionalRAII::disable(CGF&: *this, S); |
| 5887 | EmitOMPTaskBasedDirective(S, CapturedRegion: OMPD_task, BodyGen, TaskGen, Data); |
| 5888 | } |
| 5889 | |
| 5890 | void CodeGenFunction::EmitOMPTaskyieldDirective( |
| 5891 | const OMPTaskyieldDirective &S) { |
| 5892 | CGM.getOpenMPRuntime().emitTaskyieldCall(CGF&: *this, Loc: S.getBeginLoc()); |
| 5893 | } |
| 5894 | |
| 5895 | void CodeGenFunction::EmitOMPErrorDirective(const OMPErrorDirective &S) { |
| 5896 | const OMPMessageClause *MC = S.getSingleClause<OMPMessageClause>(); |
| 5897 | Expr *ME = MC ? MC->getMessageString() : nullptr; |
| 5898 | const OMPSeverityClause *SC = S.getSingleClause<OMPSeverityClause>(); |
| 5899 | bool IsFatal = false; |
| 5900 | if (!SC || SC->getSeverityKind() == OMPC_SEVERITY_fatal) |
| 5901 | IsFatal = true; |
| 5902 | CGM.getOpenMPRuntime().emitErrorCall(CGF&: *this, Loc: S.getBeginLoc(), ME, IsFatal); |
| 5903 | } |
| 5904 | |
| 5905 | void CodeGenFunction::EmitOMPBarrierDirective(const OMPBarrierDirective &S) { |
| 5906 | CGM.getOpenMPRuntime().emitBarrierCall(CGF&: *this, Loc: S.getBeginLoc(), Kind: OMPD_barrier); |
| 5907 | } |
| 5908 | |
| 5909 | void CodeGenFunction::EmitOMPTaskwaitDirective(const OMPTaskwaitDirective &S) { |
| 5910 | OMPTaskDataTy Data; |
| 5911 | // Build list of dependences |
| 5912 | buildDependences(S, Data); |
| 5913 | Data.HasNowaitClause = S.hasClausesOfKind<OMPNowaitClause>(); |
| 5914 | CGM.getOpenMPRuntime().emitTaskwaitCall(CGF&: *this, Loc: S.getBeginLoc(), Data); |
| 5915 | } |
| 5916 | |
| 5917 | static bool isSupportedByOpenMPIRBuilder(const OMPTaskgroupDirective &T) { |
| 5918 | return T.clauses().empty(); |
| 5919 | } |
| 5920 | |
| 5921 | void CodeGenFunction::EmitOMPTaskgroupDirective( |
| 5922 | const OMPTaskgroupDirective &S) { |
| 5923 | OMPLexicalScope Scope(*this, S, OMPD_unknown); |
| 5924 | if (CGM.getLangOpts().OpenMPIRBuilder && isSupportedByOpenMPIRBuilder(T: S)) { |
| 5925 | llvm::OpenMPIRBuilder &OMPBuilder = CGM.getOpenMPRuntime().getOMPBuilder(); |
| 5926 | using InsertPointTy = llvm::OpenMPIRBuilder::InsertPointTy; |
| 5927 | InsertPointTy AllocaIP(AllocaInsertPt->getParent(), |
| 5928 | AllocaInsertPt->getIterator()); |
| 5929 | |
| 5930 | auto BodyGenCB = [&, this](InsertPointTy AllocIP, InsertPointTy CodeGenIP, |
| 5931 | ArrayRef<llvm::BasicBlock *> DeallocBlocks) { |
| 5932 | Builder.restoreIP(IP: CodeGenIP); |
| 5933 | EmitStmt(S: S.getInnermostCapturedStmt()->getCapturedStmt()); |
| 5934 | return llvm::Error::success(); |
| 5935 | }; |
| 5936 | CodeGenFunction::CGCapturedStmtInfo CapStmtInfo; |
| 5937 | if (!CapturedStmtInfo) |
| 5938 | CapturedStmtInfo = &CapStmtInfo; |
| 5939 | llvm::OpenMPIRBuilder::InsertPointTy AfterIP = |
| 5940 | cantFail(ValOrErr: OMPBuilder.createTaskgroup(Loc: Builder, AllocaIP, |
| 5941 | /*DeallocBlocks=*/{}, BodyGenCB)); |
| 5942 | Builder.restoreIP(IP: AfterIP); |
| 5943 | return; |
| 5944 | } |
| 5945 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 5946 | Action.Enter(CGF); |
| 5947 | if (const Expr *E = S.getReductionRef()) { |
| 5948 | SmallVector<const Expr *, 4> LHSs; |
| 5949 | SmallVector<const Expr *, 4> RHSs; |
| 5950 | OMPTaskDataTy Data; |
| 5951 | for (const auto *C : S.getClausesOfKind<OMPTaskReductionClause>()) { |
| 5952 | Data.ReductionVars.append(in_start: C->varlist_begin(), in_end: C->varlist_end()); |
| 5953 | Data.ReductionOrigs.append(in_start: C->varlist_begin(), in_end: C->varlist_end()); |
| 5954 | Data.ReductionCopies.append(in_start: C->privates().begin(), in_end: C->privates().end()); |
| 5955 | Data.ReductionOps.append(in_start: C->reduction_ops().begin(), |
| 5956 | in_end: C->reduction_ops().end()); |
| 5957 | LHSs.append(in_start: C->lhs_exprs().begin(), in_end: C->lhs_exprs().end()); |
| 5958 | RHSs.append(in_start: C->rhs_exprs().begin(), in_end: C->rhs_exprs().end()); |
| 5959 | } |
| 5960 | llvm::Value *ReductionDesc = |
| 5961 | CGF.CGM.getOpenMPRuntime().emitTaskReductionInit(CGF, Loc: S.getBeginLoc(), |
| 5962 | LHSExprs: LHSs, RHSExprs: RHSs, Data); |
| 5963 | const auto *VD = cast<VarDecl>(Val: cast<DeclRefExpr>(Val: E)->getDecl()); |
| 5964 | CGF.EmitVarDecl(D: *VD); |
| 5965 | CGF.EmitStoreOfScalar(Value: ReductionDesc, Addr: CGF.GetAddrOfLocalVar(VD), |
| 5966 | /*Volatile=*/false, Ty: E->getType()); |
| 5967 | } |
| 5968 | CGF.EmitStmt(S: S.getInnermostCapturedStmt()->getCapturedStmt()); |
| 5969 | }; |
| 5970 | CGM.getOpenMPRuntime().emitTaskgroupRegion(CGF&: *this, TaskgroupOpGen: CodeGen, Loc: S.getBeginLoc()); |
| 5971 | } |
| 5972 | |
| 5973 | void CodeGenFunction::EmitOMPFlushDirective(const OMPFlushDirective &S) { |
| 5974 | llvm::AtomicOrdering AO = S.getSingleClause<OMPFlushClause>() |
| 5975 | ? llvm::AtomicOrdering::NotAtomic |
| 5976 | : llvm::AtomicOrdering::AcquireRelease; |
| 5977 | CGM.getOpenMPRuntime().emitFlush( |
| 5978 | CGF&: *this, |
| 5979 | Vars: [&S]() -> ArrayRef<const Expr *> { |
| 5980 | if (const auto *FlushClause = S.getSingleClause<OMPFlushClause>()) |
| 5981 | return llvm::ArrayRef(FlushClause->varlist_begin(), |
| 5982 | FlushClause->varlist_end()); |
| 5983 | return {}; |
| 5984 | }(), |
| 5985 | Loc: S.getBeginLoc(), AO); |
| 5986 | } |
| 5987 | |
| 5988 | void CodeGenFunction::EmitOMPDepobjDirective(const OMPDepobjDirective &S) { |
| 5989 | const auto *DO = S.getSingleClause<OMPDepobjClause>(); |
| 5990 | LValue DOLVal = EmitLValue(E: DO->getDepobj()); |
| 5991 | if (const auto *DC = S.getSingleClause<OMPDependClause>()) { |
| 5992 | // Build list and emit dependences |
| 5993 | OMPTaskDataTy Data; |
| 5994 | buildDependences(S, Data); |
| 5995 | for (auto &Dep : Data.Dependences) { |
| 5996 | Address DepAddr = CGM.getOpenMPRuntime().emitDepobjDependClause( |
| 5997 | CGF&: *this, Dependencies: Dep, Loc: DC->getBeginLoc()); |
| 5998 | EmitStoreOfScalar(value: DepAddr.emitRawPointer(CGF&: *this), lvalue: DOLVal); |
| 5999 | } |
| 6000 | return; |
| 6001 | } |
| 6002 | if (const auto *DC = S.getSingleClause<OMPDestroyClause>()) { |
| 6003 | CGM.getOpenMPRuntime().emitDestroyClause(CGF&: *this, DepobjLVal: DOLVal, Loc: DC->getBeginLoc()); |
| 6004 | return; |
| 6005 | } |
| 6006 | if (const auto *UC = S.getSingleClause<OMPUpdateDependObjectsClause>()) { |
| 6007 | CGM.getOpenMPRuntime().emitUpdateDependObjectsClause( |
| 6008 | CGF&: *this, DepobjLVal: DOLVal, NewDepKind: UC->getDependencyKind(), Loc: UC->getBeginLoc()); |
| 6009 | return; |
| 6010 | } |
| 6011 | } |
| 6012 | |
| 6013 | void CodeGenFunction::EmitOMPScanDirective(const OMPScanDirective &S) { |
| 6014 | if (!OMPParentLoopDirectiveForScan) |
| 6015 | return; |
| 6016 | const OMPExecutableDirective &ParentDir = *OMPParentLoopDirectiveForScan; |
| 6017 | bool IsInclusive = S.hasClausesOfKind<OMPInclusiveClause>(); |
| 6018 | SmallVector<const Expr *, 4> Shareds; |
| 6019 | SmallVector<const Expr *, 4> Privates; |
| 6020 | SmallVector<const Expr *, 4> LHSs; |
| 6021 | SmallVector<const Expr *, 4> RHSs; |
| 6022 | SmallVector<const Expr *, 4> ReductionOps; |
| 6023 | SmallVector<const Expr *, 4> CopyOps; |
| 6024 | SmallVector<const Expr *, 4> CopyArrayTemps; |
| 6025 | SmallVector<const Expr *, 4> CopyArrayElems; |
| 6026 | for (const auto *C : ParentDir.getClausesOfKind<OMPReductionClause>()) { |
| 6027 | if (C->getModifier() != OMPC_REDUCTION_inscan) |
| 6028 | continue; |
| 6029 | Shareds.append(in_start: C->varlist_begin(), in_end: C->varlist_end()); |
| 6030 | Privates.append(in_start: C->privates().begin(), in_end: C->privates().end()); |
| 6031 | LHSs.append(in_start: C->lhs_exprs().begin(), in_end: C->lhs_exprs().end()); |
| 6032 | RHSs.append(in_start: C->rhs_exprs().begin(), in_end: C->rhs_exprs().end()); |
| 6033 | ReductionOps.append(in_start: C->reduction_ops().begin(), in_end: C->reduction_ops().end()); |
| 6034 | CopyOps.append(in_start: C->copy_ops().begin(), in_end: C->copy_ops().end()); |
| 6035 | CopyArrayTemps.append(in_start: C->copy_array_temps().begin(), |
| 6036 | in_end: C->copy_array_temps().end()); |
| 6037 | CopyArrayElems.append(in_start: C->copy_array_elems().begin(), |
| 6038 | in_end: C->copy_array_elems().end()); |
| 6039 | } |
| 6040 | if (ParentDir.getDirectiveKind() == OMPD_simd || |
| 6041 | (getLangOpts().OpenMPSimd && |
| 6042 | isOpenMPSimdDirective(DKind: ParentDir.getDirectiveKind()))) { |
| 6043 | // For simd directive and simd-based directives in simd only mode, use the |
| 6044 | // following codegen: |
| 6045 | // int x = 0; |
| 6046 | // #pragma omp simd reduction(inscan, +: x) |
| 6047 | // for (..) { |
| 6048 | // <first part> |
| 6049 | // #pragma omp scan inclusive(x) |
| 6050 | // <second part> |
| 6051 | // } |
| 6052 | // is transformed to: |
| 6053 | // int x = 0; |
| 6054 | // for (..) { |
| 6055 | // int x_priv = 0; |
| 6056 | // <first part> |
| 6057 | // x = x_priv + x; |
| 6058 | // x_priv = x; |
| 6059 | // <second part> |
| 6060 | // } |
| 6061 | // and |
| 6062 | // int x = 0; |
| 6063 | // #pragma omp simd reduction(inscan, +: x) |
| 6064 | // for (..) { |
| 6065 | // <first part> |
| 6066 | // #pragma omp scan exclusive(x) |
| 6067 | // <second part> |
| 6068 | // } |
| 6069 | // to |
| 6070 | // int x = 0; |
| 6071 | // for (..) { |
| 6072 | // int x_priv = 0; |
| 6073 | // <second part> |
| 6074 | // int temp = x; |
| 6075 | // x = x_priv + x; |
| 6076 | // x_priv = temp; |
| 6077 | // <first part> |
| 6078 | // } |
| 6079 | llvm::BasicBlock *OMPScanReduce = createBasicBlock(name: "omp.inscan.reduce"); |
| 6080 | EmitBranch(Block: IsInclusive |
| 6081 | ? OMPScanReduce |
| 6082 | : BreakContinueStack.back().ContinueBlock.getBlock()); |
| 6083 | EmitBlock(BB: OMPScanDispatch); |
| 6084 | { |
| 6085 | // New scope for correct construction/destruction of temp variables for |
| 6086 | // exclusive scan. |
| 6087 | LexicalScope Scope(*this, S.getSourceRange()); |
| 6088 | EmitBranch(Block: IsInclusive ? OMPBeforeScanBlock : OMPAfterScanBlock); |
| 6089 | EmitBlock(BB: OMPScanReduce); |
| 6090 | if (!IsInclusive) { |
| 6091 | // Create temp var and copy LHS value to this temp value. |
| 6092 | // TMP = LHS; |
| 6093 | for (unsigned I = 0, E = CopyArrayElems.size(); I < E; ++I) { |
| 6094 | const Expr *PrivateExpr = Privates[I]; |
| 6095 | const Expr *TempExpr = CopyArrayTemps[I]; |
| 6096 | EmitAutoVarDecl( |
| 6097 | D: *cast<VarDecl>(Val: cast<DeclRefExpr>(Val: TempExpr)->getDecl())); |
| 6098 | LValue DestLVal = EmitLValue(E: TempExpr); |
| 6099 | LValue SrcLVal = EmitLValue(E: LHSs[I]); |
| 6100 | EmitOMPCopy(OriginalType: PrivateExpr->getType(), DestAddr: DestLVal.getAddress(), |
| 6101 | SrcAddr: SrcLVal.getAddress(), |
| 6102 | DestVD: cast<VarDecl>(Val: cast<DeclRefExpr>(Val: LHSs[I])->getDecl()), |
| 6103 | SrcVD: cast<VarDecl>(Val: cast<DeclRefExpr>(Val: RHSs[I])->getDecl()), |
| 6104 | Copy: CopyOps[I]); |
| 6105 | } |
| 6106 | } |
| 6107 | CGM.getOpenMPRuntime().emitReduction( |
| 6108 | CGF&: *this, Loc: ParentDir.getEndLoc(), Privates, LHSExprs: LHSs, RHSExprs: RHSs, ReductionOps, |
| 6109 | Options: {/*WithNowait=*/true, /*SimpleReduction=*/true, |
| 6110 | /*IsPrivateVarReduction*/ {}, .ReductionKind: OMPD_simd}); |
| 6111 | for (unsigned I = 0, E = CopyArrayElems.size(); I < E; ++I) { |
| 6112 | const Expr *PrivateExpr = Privates[I]; |
| 6113 | LValue DestLVal; |
| 6114 | LValue SrcLVal; |
| 6115 | if (IsInclusive) { |
| 6116 | DestLVal = EmitLValue(E: RHSs[I]); |
| 6117 | SrcLVal = EmitLValue(E: LHSs[I]); |
| 6118 | } else { |
| 6119 | const Expr *TempExpr = CopyArrayTemps[I]; |
| 6120 | DestLVal = EmitLValue(E: RHSs[I]); |
| 6121 | SrcLVal = EmitLValue(E: TempExpr); |
| 6122 | } |
| 6123 | EmitOMPCopy( |
| 6124 | OriginalType: PrivateExpr->getType(), DestAddr: DestLVal.getAddress(), SrcAddr: SrcLVal.getAddress(), |
| 6125 | DestVD: cast<VarDecl>(Val: cast<DeclRefExpr>(Val: LHSs[I])->getDecl()), |
| 6126 | SrcVD: cast<VarDecl>(Val: cast<DeclRefExpr>(Val: RHSs[I])->getDecl()), Copy: CopyOps[I]); |
| 6127 | } |
| 6128 | } |
| 6129 | EmitBranch(Block: IsInclusive ? OMPAfterScanBlock : OMPBeforeScanBlock); |
| 6130 | OMPScanExitBlock = IsInclusive |
| 6131 | ? BreakContinueStack.back().ContinueBlock.getBlock() |
| 6132 | : OMPScanReduce; |
| 6133 | EmitBlock(BB: OMPAfterScanBlock); |
| 6134 | return; |
| 6135 | } |
| 6136 | if (!IsInclusive) { |
| 6137 | EmitBranch(Block: BreakContinueStack.back().ContinueBlock.getBlock()); |
| 6138 | EmitBlock(BB: OMPScanExitBlock); |
| 6139 | } |
| 6140 | if (OMPFirstScanLoop) { |
| 6141 | // Emit buffer[i] = red; at the end of the input phase. |
| 6142 | const auto *IVExpr = cast<OMPLoopDirective>(Val: ParentDir) |
| 6143 | .getIterationVariable() |
| 6144 | ->IgnoreParenImpCasts(); |
| 6145 | LValue IdxLVal = EmitLValue(E: IVExpr); |
| 6146 | llvm::Value *IdxVal = EmitLoadOfScalar(lvalue: IdxLVal, Loc: IVExpr->getExprLoc()); |
| 6147 | IdxVal = Builder.CreateIntCast(V: IdxVal, DestTy: SizeTy, /*isSigned=*/false); |
| 6148 | for (unsigned I = 0, E = CopyArrayElems.size(); I < E; ++I) { |
| 6149 | const Expr *PrivateExpr = Privates[I]; |
| 6150 | const Expr *OrigExpr = Shareds[I]; |
| 6151 | const Expr *CopyArrayElem = CopyArrayElems[I]; |
| 6152 | OpaqueValueMapping IdxMapping( |
| 6153 | *this, |
| 6154 | cast<OpaqueValueExpr>( |
| 6155 | Val: cast<ArraySubscriptExpr>(Val: CopyArrayElem)->getIdx()), |
| 6156 | RValue::get(V: IdxVal)); |
| 6157 | LValue DestLVal = EmitLValue(E: CopyArrayElem); |
| 6158 | LValue SrcLVal = EmitLValue(E: OrigExpr); |
| 6159 | EmitOMPCopy( |
| 6160 | OriginalType: PrivateExpr->getType(), DestAddr: DestLVal.getAddress(), SrcAddr: SrcLVal.getAddress(), |
| 6161 | DestVD: cast<VarDecl>(Val: cast<DeclRefExpr>(Val: LHSs[I])->getDecl()), |
| 6162 | SrcVD: cast<VarDecl>(Val: cast<DeclRefExpr>(Val: RHSs[I])->getDecl()), Copy: CopyOps[I]); |
| 6163 | } |
| 6164 | } |
| 6165 | EmitBranch(Block: BreakContinueStack.back().ContinueBlock.getBlock()); |
| 6166 | if (IsInclusive) { |
| 6167 | EmitBlock(BB: OMPScanExitBlock); |
| 6168 | EmitBranch(Block: BreakContinueStack.back().ContinueBlock.getBlock()); |
| 6169 | } |
| 6170 | EmitBlock(BB: OMPScanDispatch); |
| 6171 | if (!OMPFirstScanLoop) { |
| 6172 | // Emit red = buffer[i]; at the entrance to the scan phase. |
| 6173 | const auto *IVExpr = cast<OMPLoopDirective>(Val: ParentDir) |
| 6174 | .getIterationVariable() |
| 6175 | ->IgnoreParenImpCasts(); |
| 6176 | LValue IdxLVal = EmitLValue(E: IVExpr); |
| 6177 | llvm::Value *IdxVal = EmitLoadOfScalar(lvalue: IdxLVal, Loc: IVExpr->getExprLoc()); |
| 6178 | IdxVal = Builder.CreateIntCast(V: IdxVal, DestTy: SizeTy, /*isSigned=*/false); |
| 6179 | llvm::BasicBlock *ExclusiveExitBB = nullptr; |
| 6180 | if (!IsInclusive) { |
| 6181 | llvm::BasicBlock *ContBB = createBasicBlock(name: "omp.exclusive.dec"); |
| 6182 | ExclusiveExitBB = createBasicBlock(name: "omp.exclusive.copy.exit"); |
| 6183 | llvm::Value *Cmp = Builder.CreateIsNull(Arg: IdxVal); |
| 6184 | Builder.CreateCondBr(Cond: Cmp, True: ExclusiveExitBB, False: ContBB); |
| 6185 | EmitBlock(BB: ContBB); |
| 6186 | // Use idx - 1 iteration for exclusive scan. |
| 6187 | IdxVal = Builder.CreateNUWSub(LHS: IdxVal, RHS: llvm::ConstantInt::get(Ty: SizeTy, V: 1)); |
| 6188 | } |
| 6189 | for (unsigned I = 0, E = CopyArrayElems.size(); I < E; ++I) { |
| 6190 | const Expr *PrivateExpr = Privates[I]; |
| 6191 | const Expr *OrigExpr = Shareds[I]; |
| 6192 | const Expr *CopyArrayElem = CopyArrayElems[I]; |
| 6193 | OpaqueValueMapping IdxMapping( |
| 6194 | *this, |
| 6195 | cast<OpaqueValueExpr>( |
| 6196 | Val: cast<ArraySubscriptExpr>(Val: CopyArrayElem)->getIdx()), |
| 6197 | RValue::get(V: IdxVal)); |
| 6198 | LValue SrcLVal = EmitLValue(E: CopyArrayElem); |
| 6199 | LValue DestLVal = EmitLValue(E: OrigExpr); |
| 6200 | EmitOMPCopy( |
| 6201 | OriginalType: PrivateExpr->getType(), DestAddr: DestLVal.getAddress(), SrcAddr: SrcLVal.getAddress(), |
| 6202 | DestVD: cast<VarDecl>(Val: cast<DeclRefExpr>(Val: LHSs[I])->getDecl()), |
| 6203 | SrcVD: cast<VarDecl>(Val: cast<DeclRefExpr>(Val: RHSs[I])->getDecl()), Copy: CopyOps[I]); |
| 6204 | } |
| 6205 | if (!IsInclusive) { |
| 6206 | EmitBlock(BB: ExclusiveExitBB); |
| 6207 | } |
| 6208 | } |
| 6209 | EmitBranch(Block: (OMPFirstScanLoop == IsInclusive) ? OMPBeforeScanBlock |
| 6210 | : OMPAfterScanBlock); |
| 6211 | EmitBlock(BB: OMPAfterScanBlock); |
| 6212 | } |
| 6213 | |
| 6214 | void CodeGenFunction::EmitOMPDistributeLoop(const OMPLoopDirective &S, |
| 6215 | const CodeGenLoopTy &CodeGenLoop, |
| 6216 | Expr *IncExpr) { |
| 6217 | // Emit the loop iteration variable. |
| 6218 | const auto *IVExpr = cast<DeclRefExpr>(Val: S.getIterationVariable()); |
| 6219 | const auto *IVDecl = cast<VarDecl>(Val: IVExpr->getDecl()); |
| 6220 | EmitVarDecl(D: *IVDecl); |
| 6221 | |
| 6222 | // Emit the iterations count variable. |
| 6223 | // If it is not a variable, Sema decided to calculate iterations count on each |
| 6224 | // iteration (e.g., it is foldable into a constant). |
| 6225 | if (const auto *LIExpr = dyn_cast<DeclRefExpr>(Val: S.getLastIteration())) { |
| 6226 | EmitVarDecl(D: *cast<VarDecl>(Val: LIExpr->getDecl())); |
| 6227 | // Emit calculation of the iterations count. |
| 6228 | EmitIgnoredExpr(E: S.getCalcLastIteration()); |
| 6229 | } |
| 6230 | |
| 6231 | CGOpenMPRuntime &RT = CGM.getOpenMPRuntime(); |
| 6232 | |
| 6233 | bool HasLastprivateClause = false; |
| 6234 | // Check pre-condition. |
| 6235 | { |
| 6236 | OMPLoopScope PreInitScope(*this, S); |
| 6237 | // Skip the entire loop if we don't meet the precondition. |
| 6238 | // If the condition constant folds and can be elided, avoid emitting the |
| 6239 | // whole loop. |
| 6240 | bool CondConstant; |
| 6241 | llvm::BasicBlock *ContBlock = nullptr; |
| 6242 | if (ConstantFoldsToSimpleInteger(Cond: S.getPreCond(), Result&: CondConstant)) { |
| 6243 | if (!CondConstant) |
| 6244 | return; |
| 6245 | } else { |
| 6246 | llvm::BasicBlock *ThenBlock = createBasicBlock(name: "omp.precond.then"); |
| 6247 | ContBlock = createBasicBlock(name: "omp.precond.end"); |
| 6248 | emitPreCond(CGF&: *this, S, Cond: S.getPreCond(), TrueBlock: ThenBlock, FalseBlock: ContBlock, |
| 6249 | TrueCount: getProfileCount(S: &S)); |
| 6250 | EmitBlock(BB: ThenBlock); |
| 6251 | incrementProfileCounter(S: &S); |
| 6252 | } |
| 6253 | |
| 6254 | emitAlignedClause(CGF&: *this, D: S); |
| 6255 | // Emit 'then' code. |
| 6256 | { |
| 6257 | // Emit helper vars inits. |
| 6258 | |
| 6259 | LValue LB = EmitOMPHelperVar( |
| 6260 | CGF&: *this, Helper: cast<DeclRefExpr>( |
| 6261 | Val: (isOpenMPLoopBoundSharingDirective(Kind: S.getDirectiveKind()) |
| 6262 | ? S.getCombinedLowerBoundVariable() |
| 6263 | : S.getLowerBoundVariable()))); |
| 6264 | LValue UB = EmitOMPHelperVar( |
| 6265 | CGF&: *this, Helper: cast<DeclRefExpr>( |
| 6266 | Val: (isOpenMPLoopBoundSharingDirective(Kind: S.getDirectiveKind()) |
| 6267 | ? S.getCombinedUpperBoundVariable() |
| 6268 | : S.getUpperBoundVariable()))); |
| 6269 | LValue ST = |
| 6270 | EmitOMPHelperVar(CGF&: *this, Helper: cast<DeclRefExpr>(Val: S.getStrideVariable())); |
| 6271 | LValue IL = |
| 6272 | EmitOMPHelperVar(CGF&: *this, Helper: cast<DeclRefExpr>(Val: S.getIsLastIterVariable())); |
| 6273 | |
| 6274 | OMPPrivateScope LoopScope(*this); |
| 6275 | if (EmitOMPFirstprivateClause(D: S, PrivateScope&: LoopScope)) { |
| 6276 | // Emit implicit barrier to synchronize threads and avoid data races |
| 6277 | // on initialization of firstprivate variables and post-update of |
| 6278 | // lastprivate variables. |
| 6279 | CGM.getOpenMPRuntime().emitBarrierCall( |
| 6280 | CGF&: *this, Loc: S.getBeginLoc(), Kind: OMPD_unknown, /*EmitChecks=*/false, |
| 6281 | /*ForceSimpleCall=*/true); |
| 6282 | } |
| 6283 | EmitOMPPrivateClause(D: S, PrivateScope&: LoopScope); |
| 6284 | if (isOpenMPSimdDirective(DKind: S.getDirectiveKind()) && |
| 6285 | !isOpenMPParallelDirective(DKind: S.getDirectiveKind()) && |
| 6286 | !isOpenMPTeamsDirective(DKind: S.getDirectiveKind())) |
| 6287 | EmitOMPReductionClauseInit(D: S, PrivateScope&: LoopScope); |
| 6288 | HasLastprivateClause = EmitOMPLastprivateClauseInit(D: S, PrivateScope&: LoopScope); |
| 6289 | EmitOMPPrivateLoopCounters(S, LoopScope); |
| 6290 | (void)LoopScope.Privatize(); |
| 6291 | if (isOpenMPTargetExecutionDirective(DKind: S.getDirectiveKind())) |
| 6292 | CGM.getOpenMPRuntime().adjustTargetSpecificDataForLambdas(CGF&: *this, D: S); |
| 6293 | |
| 6294 | // Detect the distribute schedule kind and chunk. |
| 6295 | llvm::Value *Chunk = nullptr; |
| 6296 | OpenMPDistScheduleClauseKind ScheduleKind = OMPC_DIST_SCHEDULE_unknown; |
| 6297 | if (const auto *C = S.getSingleClause<OMPDistScheduleClause>()) { |
| 6298 | ScheduleKind = C->getDistScheduleKind(); |
| 6299 | if (const Expr *Ch = C->getChunkSize()) { |
| 6300 | Chunk = EmitScalarExpr(E: Ch); |
| 6301 | Chunk = EmitScalarConversion(Src: Chunk, SrcTy: Ch->getType(), |
| 6302 | DstTy: S.getIterationVariable()->getType(), |
| 6303 | Loc: S.getBeginLoc()); |
| 6304 | } |
| 6305 | } else { |
| 6306 | // Default behaviour for dist_schedule clause. |
| 6307 | CGM.getOpenMPRuntime().getDefaultDistScheduleAndChunk( |
| 6308 | CGF&: *this, S, ScheduleKind, Chunk); |
| 6309 | } |
| 6310 | const unsigned IVSize = getContext().getTypeSize(T: IVExpr->getType()); |
| 6311 | const bool IVSigned = IVExpr->getType()->hasSignedIntegerRepresentation(); |
| 6312 | |
| 6313 | // GPU fused schedule: omit the outer distribute loop and let the inner |
| 6314 | // worksharing loop schedule the flattened team/thread iteration space. |
| 6315 | if (canEmitGPUFusedDistSchedule(CGM, S, DKind: S.getDirectiveKind())) { |
| 6316 | JumpDest LoopExit = |
| 6317 | getJumpDestInCurrentScope(Target: createBasicBlock(name: "omp.loop.exit")); |
| 6318 | CodeGenLoop(*this, S, LoopExit); |
| 6319 | EmitBlock(BB: LoopExit.getBlock()); |
| 6320 | } else { |
| 6321 | // OpenMP [2.10.8, distribute Construct, Description] |
| 6322 | // If dist_schedule is specified, kind must be static. If specified, |
| 6323 | // iterations are divided into chunks of size chunk_size, chunks are |
| 6324 | // assigned to the teams of the league in a round-robin fashion in the |
| 6325 | // order of the team number. When no chunk_size is specified, the |
| 6326 | // iteration space is divided into chunks that are approximately equal |
| 6327 | // in size, and at most one chunk is distributed to each team of the |
| 6328 | // league. The size of the chunks is unspecified in this case. |
| 6329 | bool StaticChunked = |
| 6330 | RT.isStaticChunked(ScheduleKind, /* Chunked */ Chunk != nullptr) && |
| 6331 | isOpenMPLoopBoundSharingDirective(Kind: S.getDirectiveKind()); |
| 6332 | if (RT.isStaticNonchunked(ScheduleKind, |
| 6333 | /* Chunked */ Chunk != nullptr) || |
| 6334 | StaticChunked) { |
| 6335 | CGOpenMPRuntime::StaticRTInput StaticInit( |
| 6336 | IVSize, IVSigned, /* Ordered = */ false, IL.getAddress(), |
| 6337 | LB.getAddress(), UB.getAddress(), ST.getAddress(), |
| 6338 | StaticChunked ? Chunk : nullptr); |
| 6339 | RT.emitDistributeStaticInit(CGF&: *this, Loc: S.getBeginLoc(), SchedKind: ScheduleKind, |
| 6340 | Values: StaticInit); |
| 6341 | JumpDest LoopExit = |
| 6342 | getJumpDestInCurrentScope(Target: createBasicBlock(name: "omp.loop.exit")); |
| 6343 | // UB = min(UB, GlobalUB); |
| 6344 | EmitIgnoredExpr( |
| 6345 | E: isOpenMPLoopBoundSharingDirective(Kind: S.getDirectiveKind()) |
| 6346 | ? S.getCombinedEnsureUpperBound() |
| 6347 | : S.getEnsureUpperBound()); |
| 6348 | // IV = LB; |
| 6349 | EmitIgnoredExpr( |
| 6350 | E: isOpenMPLoopBoundSharingDirective(Kind: S.getDirectiveKind()) |
| 6351 | ? S.getCombinedInit() |
| 6352 | : S.getInit()); |
| 6353 | |
| 6354 | const Expr *Cond = |
| 6355 | isOpenMPLoopBoundSharingDirective(Kind: S.getDirectiveKind()) |
| 6356 | ? S.getCombinedCond() |
| 6357 | : S.getCond(); |
| 6358 | |
| 6359 | if (StaticChunked) |
| 6360 | Cond = S.getCombinedDistCond(); |
| 6361 | |
| 6362 | // For static unchunked schedules generate: |
| 6363 | // |
| 6364 | // 1. For distribute alone, codegen |
| 6365 | // while (idx <= UB) { |
| 6366 | // BODY; |
| 6367 | // ++idx; |
| 6368 | // } |
| 6369 | // |
| 6370 | // 2. When combined with 'for' (e.g. as in 'distribute parallel for') |
| 6371 | // while (idx <= UB) { |
| 6372 | // <CodeGen rest of pragma>(LB, UB); |
| 6373 | // idx += ST; |
| 6374 | // } |
| 6375 | // |
| 6376 | // For static chunk one schedule generate: |
| 6377 | // |
| 6378 | // while (IV <= GlobalUB) { |
| 6379 | // <CodeGen rest of pragma>(LB, UB); |
| 6380 | // LB += ST; |
| 6381 | // UB += ST; |
| 6382 | // UB = min(UB, GlobalUB); |
| 6383 | // IV = LB; |
| 6384 | // } |
| 6385 | // |
| 6386 | emitCommonSimdLoop( |
| 6387 | CGF&: *this, S, |
| 6388 | SimdInitGen: [&S](CodeGenFunction &CGF, PrePostActionTy &) { |
| 6389 | if (isOpenMPSimdDirective(DKind: S.getDirectiveKind())) |
| 6390 | CGF.EmitOMPSimdInit(D: S); |
| 6391 | }, |
| 6392 | BodyCodeGen: [&S, &LoopScope, Cond, IncExpr, LoopExit, &CodeGenLoop, |
| 6393 | StaticChunked](CodeGenFunction &CGF, PrePostActionTy &) { |
| 6394 | CGF.EmitOMPInnerLoop( |
| 6395 | S, RequiresCleanup: LoopScope.requiresCleanups(), LoopCond: Cond, IncExpr, |
| 6396 | BodyGen: [&S, LoopExit, &CodeGenLoop](CodeGenFunction &CGF) { |
| 6397 | CodeGenLoop(CGF, S, LoopExit); |
| 6398 | }, |
| 6399 | PostIncGen: [&S, StaticChunked](CodeGenFunction &CGF) { |
| 6400 | if (StaticChunked) { |
| 6401 | CGF.EmitIgnoredExpr(E: S.getCombinedNextLowerBound()); |
| 6402 | CGF.EmitIgnoredExpr(E: S.getCombinedNextUpperBound()); |
| 6403 | CGF.EmitIgnoredExpr(E: S.getCombinedEnsureUpperBound()); |
| 6404 | CGF.EmitIgnoredExpr(E: S.getCombinedInit()); |
| 6405 | } |
| 6406 | }); |
| 6407 | }); |
| 6408 | EmitBlock(BB: LoopExit.getBlock()); |
| 6409 | // Tell the runtime we are done. |
| 6410 | RT.emitForStaticFinish(CGF&: *this, Loc: S.getEndLoc(), DKind: OMPD_distribute); |
| 6411 | } else { |
| 6412 | // Emit the outer loop, which requests its work chunk [LB..UB] from |
| 6413 | // runtime and runs the inner loop to process it. |
| 6414 | const OMPLoopArguments LoopArguments = { |
| 6415 | LB.getAddress(), UB.getAddress(), ST.getAddress(), |
| 6416 | IL.getAddress(), Chunk}; |
| 6417 | EmitOMPDistributeOuterLoop(ScheduleKind, S, LoopScope, LoopArgs: LoopArguments, |
| 6418 | CodeGenLoopContent: CodeGenLoop); |
| 6419 | } |
| 6420 | } |
| 6421 | if (isOpenMPSimdDirective(DKind: S.getDirectiveKind())) { |
| 6422 | EmitOMPSimdFinal(D: S, CondGen: [IL, &S](CodeGenFunction &CGF) { |
| 6423 | return CGF.Builder.CreateIsNotNull( |
| 6424 | Arg: CGF.EmitLoadOfScalar(lvalue: IL, Loc: S.getBeginLoc())); |
| 6425 | }); |
| 6426 | } |
| 6427 | if (isOpenMPSimdDirective(DKind: S.getDirectiveKind()) && |
| 6428 | !isOpenMPParallelDirective(DKind: S.getDirectiveKind()) && |
| 6429 | !isOpenMPTeamsDirective(DKind: S.getDirectiveKind())) { |
| 6430 | EmitOMPReductionClauseFinal(D: S, ReductionKind: OMPD_simd); |
| 6431 | // Emit post-update of the reduction variables if IsLastIter != 0. |
| 6432 | emitPostUpdateForReductionClause( |
| 6433 | CGF&: *this, D: S, CondGen: [IL, &S](CodeGenFunction &CGF) { |
| 6434 | return CGF.Builder.CreateIsNotNull( |
| 6435 | Arg: CGF.EmitLoadOfScalar(lvalue: IL, Loc: S.getBeginLoc())); |
| 6436 | }); |
| 6437 | } |
| 6438 | // Emit final copy of the lastprivate variables if IsLastIter != 0. |
| 6439 | if (HasLastprivateClause) { |
| 6440 | EmitOMPLastprivateClauseFinal( |
| 6441 | D: S, /*NoFinals=*/false, |
| 6442 | IsLastIterCond: Builder.CreateIsNotNull(Arg: EmitLoadOfScalar(lvalue: IL, Loc: S.getBeginLoc()))); |
| 6443 | } |
| 6444 | } |
| 6445 | |
| 6446 | // We're now done with the loop, so jump to the continuation block. |
| 6447 | if (ContBlock) { |
| 6448 | EmitBranch(Block: ContBlock); |
| 6449 | EmitBlock(BB: ContBlock, IsFinished: true); |
| 6450 | } |
| 6451 | } |
| 6452 | } |
| 6453 | |
| 6454 | // Pass OMPLoopDirective (instead of OMPDistributeDirective) to make this |
| 6455 | // function available for "loop bind(teams)", which maps to "distribute". |
| 6456 | static void emitOMPDistributeDirective(const OMPLoopDirective &S, |
| 6457 | CodeGenFunction &CGF, |
| 6458 | CodeGenModule &CGM) { |
| 6459 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &) { |
| 6460 | CGF.EmitOMPDistributeLoop(S, CodeGenLoop: emitOMPLoopBodyWithStopPoint, IncExpr: S.getInc()); |
| 6461 | }; |
| 6462 | OMPLexicalScope Scope(CGF, S, OMPD_unknown); |
| 6463 | CGM.getOpenMPRuntime().emitInlinedDirective(CGF, InnermostKind: OMPD_distribute, CodeGen); |
| 6464 | } |
| 6465 | |
| 6466 | void CodeGenFunction::EmitOMPDistributeDirective( |
| 6467 | const OMPDistributeDirective &S) { |
| 6468 | emitOMPDistributeDirective(S, CGF&: *this, CGM); |
| 6469 | } |
| 6470 | |
| 6471 | static llvm::Function * |
| 6472 | emitOutlinedOrderedFunction(CodeGenModule &CGM, const CapturedStmt *S, |
| 6473 | const OMPExecutableDirective &D) { |
| 6474 | CodeGenFunction CGF(CGM, /*suppressNewContext=*/true); |
| 6475 | CodeGenFunction::CGCapturedStmtInfo CapStmtInfo; |
| 6476 | CGF.CapturedStmtInfo = &CapStmtInfo; |
| 6477 | llvm::Function *Fn = CGF.GenerateOpenMPCapturedStmtFunction(S: *S, D); |
| 6478 | Fn->setDoesNotRecurse(); |
| 6479 | return Fn; |
| 6480 | } |
| 6481 | |
| 6482 | template <typename T> |
| 6483 | static void emitRestoreIP(CodeGenFunction &CGF, const T *C, |
| 6484 | llvm::OpenMPIRBuilder::InsertPointTy AllocaIP, |
| 6485 | llvm::OpenMPIRBuilder &OMPBuilder) { |
| 6486 | |
| 6487 | unsigned NumLoops = C->getNumLoops(); |
| 6488 | QualType Int64Ty = CGF.CGM.getContext().getIntTypeForBitwidth( |
| 6489 | /*DestWidth=*/64, /*Signed=*/1); |
| 6490 | llvm::SmallVector<llvm::Value *> StoreValues; |
| 6491 | for (unsigned I = 0; I < NumLoops; I++) { |
| 6492 | const Expr *CounterVal = C->getLoopData(I); |
| 6493 | assert(CounterVal); |
| 6494 | llvm::Value *StoreValue = CGF.EmitScalarConversion( |
| 6495 | Src: CGF.EmitScalarExpr(E: CounterVal), SrcTy: CounterVal->getType(), DstTy: Int64Ty, |
| 6496 | Loc: CounterVal->getExprLoc()); |
| 6497 | StoreValues.emplace_back(Args&: StoreValue); |
| 6498 | } |
| 6499 | OMPDoacrossKind<T> ODK; |
| 6500 | bool IsDependSource = ODK.isSource(C); |
| 6501 | CGF.Builder.restoreIP( |
| 6502 | IP: OMPBuilder.createOrderedDepend(Loc: CGF.Builder, AllocaIP, NumLoops, |
| 6503 | StoreValues, Name: ".cnt.addr", IsDependSource)); |
| 6504 | } |
| 6505 | |
| 6506 | void CodeGenFunction::EmitOMPOrderedStandaloneDirective( |
| 6507 | const OMPOrderedStandaloneDirective &S) { |
| 6508 | assert((S.hasClausesOfKind<OMPDependClause>() || |
| 6509 | S.hasClausesOfKind<OMPDoacrossClause>()) && |
| 6510 | "Standalone ordered directive should have either depend or doacross " |
| 6511 | "clause"); |
| 6512 | // The ordered-standalone directive. |
| 6513 | assert(!S.hasAssociatedStmt() && "No associated statement must be in " |
| 6514 | "ordered depend|doacross construct."); |
| 6515 | |
| 6516 | if (CGM.getLangOpts().OpenMPIRBuilder) { |
| 6517 | llvm::OpenMPIRBuilder &OMPBuilder = CGM.getOpenMPRuntime().getOMPBuilder(); |
| 6518 | using InsertPointTy = llvm::OpenMPIRBuilder::InsertPointTy; |
| 6519 | |
| 6520 | InsertPointTy AllocaIP(AllocaInsertPt->getParent(), |
| 6521 | AllocaInsertPt->getIterator()); |
| 6522 | for (const auto *DC : S.getClausesOfKind<OMPDependClause>()) |
| 6523 | emitRestoreIP(CGF&: *this, C: DC, AllocaIP, OMPBuilder); |
| 6524 | for (const auto *DC : S.getClausesOfKind<OMPDoacrossClause>()) |
| 6525 | emitRestoreIP(CGF&: *this, C: DC, AllocaIP, OMPBuilder); |
| 6526 | return; |
| 6527 | } |
| 6528 | |
| 6529 | if (S.hasClausesOfKind<OMPDependClause>()) { |
| 6530 | for (const auto *DC : S.getClausesOfKind<OMPDependClause>()) |
| 6531 | CGM.getOpenMPRuntime().emitDoacrossOrdered(CGF&: *this, C: DC); |
| 6532 | } else if (S.hasClausesOfKind<OMPDoacrossClause>()) { |
| 6533 | for (const auto *DC : S.getClausesOfKind<OMPDoacrossClause>()) |
| 6534 | CGM.getOpenMPRuntime().emitDoacrossOrdered(CGF&: *this, C: DC); |
| 6535 | } |
| 6536 | } |
| 6537 | |
| 6538 | void CodeGenFunction::EmitOMPOrderedBlockAssocDirective( |
| 6539 | const OMPOrderedBlockAssocDirective &S) { |
| 6540 | if (CGM.getLangOpts().OpenMPIRBuilder) { |
| 6541 | llvm::OpenMPIRBuilder &OMPBuilder = CGM.getOpenMPRuntime().getOMPBuilder(); |
| 6542 | using InsertPointTy = llvm::OpenMPIRBuilder::InsertPointTy; |
| 6543 | |
| 6544 | // The ordered directive with threads or simd clause, or without clause. |
| 6545 | // Without clause, it behaves as if the threads clause is specified. |
| 6546 | const auto *C = S.getSingleClause<OMPSIMDClause>(); |
| 6547 | |
| 6548 | auto FiniCB = [this](InsertPointTy IP) { |
| 6549 | OMPBuilderCBHelpers::FinalizeOMPRegion(CGF&: *this, IP); |
| 6550 | return llvm::Error::success(); |
| 6551 | }; |
| 6552 | |
| 6553 | auto BodyGenCB = [&S, C, this](InsertPointTy AllocIP, |
| 6554 | InsertPointTy CodeGenIP, |
| 6555 | ArrayRef<llvm::BasicBlock *> DeallocBlocks) { |
| 6556 | Builder.restoreIP(IP: CodeGenIP); |
| 6557 | |
| 6558 | const CapturedStmt *CS = S.getInnermostCapturedStmt(); |
| 6559 | if (C) { |
| 6560 | llvm::BasicBlock *FiniBB = splitBBWithSuffix( |
| 6561 | Builder, /*CreateBranch=*/false, Suffix: ".ordered.after"); |
| 6562 | llvm::SmallVector<llvm::Value *, 16> CapturedVars; |
| 6563 | GenerateOpenMPCapturedVars(S: *CS, CapturedVars); |
| 6564 | llvm::Function *OutlinedFn = emitOutlinedOrderedFunction(CGM, S: CS, D: S); |
| 6565 | assert(S.getBeginLoc().isValid() && |
| 6566 | "Outlined function call location must be valid."); |
| 6567 | ApplyDebugLocation::CreateDefaultArtificial(CGF&: *this, TemporaryLocation: S.getBeginLoc()); |
| 6568 | OMPBuilderCBHelpers::EmitCaptureStmt(CGF&: *this, CodeGenIP, FiniBB&: *FiniBB, |
| 6569 | Fn: OutlinedFn, Args: CapturedVars); |
| 6570 | } else { |
| 6571 | OMPBuilderCBHelpers::EmitOMPInlinedRegionBody( |
| 6572 | CGF&: *this, RegionBodyStmt: CS->getCapturedStmt(), AllocaIP: AllocIP, CodeGenIP, RegionName: "ordered"); |
| 6573 | } |
| 6574 | return llvm::Error::success(); |
| 6575 | }; |
| 6576 | |
| 6577 | OMPLexicalScope Scope(*this, S, OMPD_unknown); |
| 6578 | llvm::OpenMPIRBuilder::InsertPointTy AfterIP = cantFail( |
| 6579 | ValOrErr: OMPBuilder.createOrderedThreadsSimd(Loc: Builder, BodyGenCB, FiniCB, IsThreads: !C)); |
| 6580 | Builder.restoreIP(IP: AfterIP); |
| 6581 | return; |
| 6582 | } |
| 6583 | |
| 6584 | const auto *C = S.getSingleClause<OMPSIMDClause>(); |
| 6585 | auto &&CodeGen = [&S, C, this](CodeGenFunction &CGF, |
| 6586 | PrePostActionTy &Action) { |
| 6587 | const CapturedStmt *CS = S.getInnermostCapturedStmt(); |
| 6588 | if (C) { |
| 6589 | llvm::SmallVector<llvm::Value *, 16> CapturedVars; |
| 6590 | CGF.GenerateOpenMPCapturedVars(S: *CS, CapturedVars); |
| 6591 | llvm::Function *OutlinedFn = emitOutlinedOrderedFunction(CGM, S: CS, D: S); |
| 6592 | CGM.getOpenMPRuntime().emitOutlinedFunctionCall(CGF, Loc: S.getBeginLoc(), |
| 6593 | OutlinedFn, Args: CapturedVars); |
| 6594 | } else { |
| 6595 | Action.Enter(CGF); |
| 6596 | CGF.EmitStmt(S: CS->getCapturedStmt()); |
| 6597 | } |
| 6598 | }; |
| 6599 | OMPLexicalScope Scope(*this, S, OMPD_unknown); |
| 6600 | CGM.getOpenMPRuntime().emitOrderedRegion(CGF&: *this, OrderedOpGen: CodeGen, Loc: S.getBeginLoc(), IsThreads: !C); |
| 6601 | } |
| 6602 | |
| 6603 | static llvm::Value *convertToScalarValue(CodeGenFunction &CGF, RValue Val, |
| 6604 | QualType SrcType, QualType DestType, |
| 6605 | SourceLocation Loc) { |
| 6606 | assert(CGF.hasScalarEvaluationKind(DestType) && |
| 6607 | "DestType must have scalar evaluation kind."); |
| 6608 | assert(!Val.isAggregate() && "Must be a scalar or complex."); |
| 6609 | return Val.isScalar() ? CGF.EmitScalarConversion(Src: Val.getScalarVal(), SrcTy: SrcType, |
| 6610 | DstTy: DestType, Loc) |
| 6611 | : CGF.EmitComplexToScalarConversion( |
| 6612 | Src: Val.getComplexVal(), SrcTy: SrcType, DstTy: DestType, Loc); |
| 6613 | } |
| 6614 | |
| 6615 | static CodeGenFunction::ComplexPairTy |
| 6616 | convertToComplexValue(CodeGenFunction &CGF, RValue Val, QualType SrcType, |
| 6617 | QualType DestType, SourceLocation Loc) { |
| 6618 | assert(CGF.getEvaluationKind(DestType) == TEK_Complex && |
| 6619 | "DestType must have complex evaluation kind."); |
| 6620 | CodeGenFunction::ComplexPairTy ComplexVal; |
| 6621 | if (Val.isScalar()) { |
| 6622 | // Convert the input element to the element type of the complex. |
| 6623 | QualType DestElementType = |
| 6624 | DestType->castAs<ComplexType>()->getElementType(); |
| 6625 | llvm::Value *ScalarVal = CGF.EmitScalarConversion( |
| 6626 | Src: Val.getScalarVal(), SrcTy: SrcType, DstTy: DestElementType, Loc); |
| 6627 | ComplexVal = CodeGenFunction::ComplexPairTy( |
| 6628 | ScalarVal, llvm::Constant::getNullValue(Ty: ScalarVal->getType())); |
| 6629 | } else { |
| 6630 | assert(Val.isComplex() && "Must be a scalar or complex."); |
| 6631 | QualType SrcElementType = SrcType->castAs<ComplexType>()->getElementType(); |
| 6632 | QualType DestElementType = |
| 6633 | DestType->castAs<ComplexType>()->getElementType(); |
| 6634 | ComplexVal.first = CGF.EmitScalarConversion( |
| 6635 | Src: Val.getComplexVal().first, SrcTy: SrcElementType, DstTy: DestElementType, Loc); |
| 6636 | ComplexVal.second = CGF.EmitScalarConversion( |
| 6637 | Src: Val.getComplexVal().second, SrcTy: SrcElementType, DstTy: DestElementType, Loc); |
| 6638 | } |
| 6639 | return ComplexVal; |
| 6640 | } |
| 6641 | |
| 6642 | static void emitSimpleAtomicStore(CodeGenFunction &CGF, llvm::AtomicOrdering AO, |
| 6643 | LValue LVal, RValue RVal) { |
| 6644 | if (LVal.isGlobalReg()) |
| 6645 | CGF.EmitStoreThroughGlobalRegLValue(Src: RVal, Dst: LVal); |
| 6646 | else |
| 6647 | CGF.EmitAtomicStore(rvalue: RVal, lvalue: LVal, AO, IsVolatile: LVal.isVolatile(), /*isInit=*/false); |
| 6648 | } |
| 6649 | |
| 6650 | static RValue emitSimpleAtomicLoad(CodeGenFunction &CGF, |
| 6651 | llvm::AtomicOrdering AO, LValue LVal, |
| 6652 | SourceLocation Loc) { |
| 6653 | if (LVal.isGlobalReg()) |
| 6654 | return CGF.EmitLoadOfLValue(V: LVal, Loc); |
| 6655 | return CGF.EmitAtomicLoad( |
| 6656 | lvalue: LVal, loc: Loc, AO: llvm::AtomicCmpXchgInst::getStrongestFailureOrdering(SuccessOrdering: AO), |
| 6657 | IsVolatile: LVal.isVolatile()); |
| 6658 | } |
| 6659 | |
| 6660 | void CodeGenFunction::emitOMPSimpleStore(LValue LVal, RValue RVal, |
| 6661 | QualType RValTy, SourceLocation Loc) { |
| 6662 | switch (getEvaluationKind(T: LVal.getType())) { |
| 6663 | case TEK_Scalar: |
| 6664 | EmitStoreThroughLValue(Src: RValue::get(V: convertToScalarValue( |
| 6665 | CGF&: *this, Val: RVal, SrcType: RValTy, DestType: LVal.getType(), Loc)), |
| 6666 | Dst: LVal); |
| 6667 | break; |
| 6668 | case TEK_Complex: |
| 6669 | EmitStoreOfComplex( |
| 6670 | V: convertToComplexValue(CGF&: *this, Val: RVal, SrcType: RValTy, DestType: LVal.getType(), Loc), dest: LVal, |
| 6671 | /*isInit=*/false); |
| 6672 | break; |
| 6673 | case TEK_Aggregate: |
| 6674 | llvm_unreachable("Must be a scalar or complex."); |
| 6675 | } |
| 6676 | } |
| 6677 | |
| 6678 | static void emitOMPAtomicReadExpr(CodeGenFunction &CGF, llvm::AtomicOrdering AO, |
| 6679 | const Expr *X, const Expr *V, |
| 6680 | SourceLocation Loc) { |
| 6681 | // v = x; |
| 6682 | assert(V->isLValue() && "V of 'omp atomic read' is not lvalue"); |
| 6683 | assert(X->isLValue() && "X of 'omp atomic read' is not lvalue"); |
| 6684 | LValue XLValue = CGF.EmitLValue(E: X); |
| 6685 | LValue VLValue = CGF.EmitLValue(E: V); |
| 6686 | RValue Res = emitSimpleAtomicLoad(CGF, AO, LVal: XLValue, Loc); |
| 6687 | // OpenMP, 2.17.7, atomic Construct |
| 6688 | // If the read or capture clause is specified and the acquire, acq_rel, or |
| 6689 | // seq_cst clause is specified then the strong flush on exit from the atomic |
| 6690 | // operation is also an acquire flush. |
| 6691 | switch (AO) { |
| 6692 | case llvm::AtomicOrdering::Acquire: |
| 6693 | case llvm::AtomicOrdering::AcquireRelease: |
| 6694 | case llvm::AtomicOrdering::SequentiallyConsistent: |
| 6695 | CGF.CGM.getOpenMPRuntime().emitFlush(CGF, Vars: {}, Loc, |
| 6696 | AO: llvm::AtomicOrdering::Acquire); |
| 6697 | break; |
| 6698 | case llvm::AtomicOrdering::Monotonic: |
| 6699 | case llvm::AtomicOrdering::Release: |
| 6700 | break; |
| 6701 | case llvm::AtomicOrdering::NotAtomic: |
| 6702 | case llvm::AtomicOrdering::Unordered: |
| 6703 | llvm_unreachable("Unexpected ordering."); |
| 6704 | } |
| 6705 | CGF.emitOMPSimpleStore(LVal: VLValue, RVal: Res, RValTy: X->getType().getNonReferenceType(), Loc); |
| 6706 | CGF.CGM.getOpenMPRuntime().checkAndEmitLastprivateConditional(CGF, LHS: V); |
| 6707 | } |
| 6708 | |
| 6709 | static void emitOMPAtomicWriteExpr(CodeGenFunction &CGF, |
| 6710 | llvm::AtomicOrdering AO, const Expr *X, |
| 6711 | const Expr *E, SourceLocation Loc) { |
| 6712 | // x = expr; |
| 6713 | assert(X->isLValue() && "X of 'omp atomic write' is not lvalue"); |
| 6714 | emitSimpleAtomicStore(CGF, AO, LVal: CGF.EmitLValue(E: X), RVal: CGF.EmitAnyExpr(E)); |
| 6715 | CGF.CGM.getOpenMPRuntime().checkAndEmitLastprivateConditional(CGF, LHS: X); |
| 6716 | // OpenMP, 2.17.7, atomic Construct |
| 6717 | // If the write, update, or capture clause is specified and the release, |
| 6718 | // acq_rel, or seq_cst clause is specified then the strong flush on entry to |
| 6719 | // the atomic operation is also a release flush. |
| 6720 | switch (AO) { |
| 6721 | case llvm::AtomicOrdering::Release: |
| 6722 | case llvm::AtomicOrdering::AcquireRelease: |
| 6723 | case llvm::AtomicOrdering::SequentiallyConsistent: |
| 6724 | CGF.CGM.getOpenMPRuntime().emitFlush(CGF, Vars: {}, Loc, |
| 6725 | AO: llvm::AtomicOrdering::Release); |
| 6726 | break; |
| 6727 | case llvm::AtomicOrdering::Acquire: |
| 6728 | case llvm::AtomicOrdering::Monotonic: |
| 6729 | break; |
| 6730 | case llvm::AtomicOrdering::NotAtomic: |
| 6731 | case llvm::AtomicOrdering::Unordered: |
| 6732 | llvm_unreachable("Unexpected ordering."); |
| 6733 | } |
| 6734 | } |
| 6735 | |
| 6736 | static std::pair<bool, RValue> emitOMPAtomicRMW(CodeGenFunction &CGF, LValue X, |
| 6737 | RValue Update, |
| 6738 | BinaryOperatorKind BO, |
| 6739 | llvm::AtomicOrdering AO, |
| 6740 | bool IsXLHSInRHSPart) { |
| 6741 | ASTContext &Context = CGF.getContext(); |
| 6742 | // Allow atomicrmw only if 'x' and 'update' are integer values, lvalue for 'x' |
| 6743 | // expression is simple and atomic is allowed for the given type for the |
| 6744 | // target platform. |
| 6745 | if (BO == BO_Comma || !Update.isScalar() || !X.isSimple() || |
| 6746 | (!isa<llvm::ConstantInt>(Val: Update.getScalarVal()) && |
| 6747 | (Update.getScalarVal()->getType() != X.getAddress().getElementType())) || |
| 6748 | !Context.getTargetInfo().hasBuiltinAtomic( |
| 6749 | AtomicSizeInBits: Context.getTypeSize(T: X.getType()), AlignmentInBits: Context.toBits(CharSize: X.getAlignment()))) |
| 6750 | return std::make_pair(x: false, y: RValue::get(V: nullptr)); |
| 6751 | |
| 6752 | auto &&CheckAtomicSupport = [&CGF](llvm::Type *T, BinaryOperatorKind BO) { |
| 6753 | if (T->isIntegerTy()) |
| 6754 | return true; |
| 6755 | |
| 6756 | if (T->isFloatingPointTy() && (BO == BO_Add || BO == BO_Sub)) |
| 6757 | return llvm::isPowerOf2_64(Value: CGF.CGM.getDataLayout().getTypeStoreSize(Ty: T)); |
| 6758 | |
| 6759 | return false; |
| 6760 | }; |
| 6761 | |
| 6762 | if (!CheckAtomicSupport(Update.getScalarVal()->getType(), BO) || |
| 6763 | !CheckAtomicSupport(X.getAddress().getElementType(), BO)) |
| 6764 | return std::make_pair(x: false, y: RValue::get(V: nullptr)); |
| 6765 | |
| 6766 | bool IsInteger = X.getAddress().getElementType()->isIntegerTy(); |
| 6767 | llvm::AtomicRMWInst::BinOp RMWOp; |
| 6768 | switch (BO) { |
| 6769 | case BO_Add: |
| 6770 | RMWOp = IsInteger ? llvm::AtomicRMWInst::Add : llvm::AtomicRMWInst::FAdd; |
| 6771 | break; |
| 6772 | case BO_Sub: |
| 6773 | if (!IsXLHSInRHSPart) |
| 6774 | return std::make_pair(x: false, y: RValue::get(V: nullptr)); |
| 6775 | RMWOp = IsInteger ? llvm::AtomicRMWInst::Sub : llvm::AtomicRMWInst::FSub; |
| 6776 | break; |
| 6777 | case BO_And: |
| 6778 | RMWOp = llvm::AtomicRMWInst::And; |
| 6779 | break; |
| 6780 | case BO_Or: |
| 6781 | RMWOp = llvm::AtomicRMWInst::Or; |
| 6782 | break; |
| 6783 | case BO_Xor: |
| 6784 | RMWOp = llvm::AtomicRMWInst::Xor; |
| 6785 | break; |
| 6786 | case BO_LT: |
| 6787 | if (IsInteger) |
| 6788 | RMWOp = X.getType()->hasSignedIntegerRepresentation() |
| 6789 | ? (IsXLHSInRHSPart ? llvm::AtomicRMWInst::Min |
| 6790 | : llvm::AtomicRMWInst::Max) |
| 6791 | : (IsXLHSInRHSPart ? llvm::AtomicRMWInst::UMin |
| 6792 | : llvm::AtomicRMWInst::UMax); |
| 6793 | else |
| 6794 | RMWOp = IsXLHSInRHSPart ? llvm::AtomicRMWInst::FMin |
| 6795 | : llvm::AtomicRMWInst::FMax; |
| 6796 | break; |
| 6797 | case BO_GT: |
| 6798 | if (IsInteger) |
| 6799 | RMWOp = X.getType()->hasSignedIntegerRepresentation() |
| 6800 | ? (IsXLHSInRHSPart ? llvm::AtomicRMWInst::Max |
| 6801 | : llvm::AtomicRMWInst::Min) |
| 6802 | : (IsXLHSInRHSPart ? llvm::AtomicRMWInst::UMax |
| 6803 | : llvm::AtomicRMWInst::UMin); |
| 6804 | else |
| 6805 | RMWOp = IsXLHSInRHSPart ? llvm::AtomicRMWInst::FMax |
| 6806 | : llvm::AtomicRMWInst::FMin; |
| 6807 | break; |
| 6808 | case BO_Assign: |
| 6809 | RMWOp = llvm::AtomicRMWInst::Xchg; |
| 6810 | break; |
| 6811 | case BO_Mul: |
| 6812 | case BO_Div: |
| 6813 | case BO_Rem: |
| 6814 | case BO_Shl: |
| 6815 | case BO_Shr: |
| 6816 | case BO_LAnd: |
| 6817 | case BO_LOr: |
| 6818 | return std::make_pair(x: false, y: RValue::get(V: nullptr)); |
| 6819 | case BO_PtrMemD: |
| 6820 | case BO_PtrMemI: |
| 6821 | case BO_LE: |
| 6822 | case BO_GE: |
| 6823 | case BO_EQ: |
| 6824 | case BO_NE: |
| 6825 | case BO_Cmp: |
| 6826 | case BO_AddAssign: |
| 6827 | case BO_SubAssign: |
| 6828 | case BO_AndAssign: |
| 6829 | case BO_OrAssign: |
| 6830 | case BO_XorAssign: |
| 6831 | case BO_MulAssign: |
| 6832 | case BO_DivAssign: |
| 6833 | case BO_RemAssign: |
| 6834 | case BO_ShlAssign: |
| 6835 | case BO_ShrAssign: |
| 6836 | case BO_Comma: |
| 6837 | llvm_unreachable("Unsupported atomic update operation"); |
| 6838 | } |
| 6839 | llvm::Value *UpdateVal = Update.getScalarVal(); |
| 6840 | if (auto *IC = dyn_cast<llvm::ConstantInt>(Val: UpdateVal)) { |
| 6841 | if (IsInteger) |
| 6842 | UpdateVal = CGF.Builder.CreateIntCast( |
| 6843 | V: IC, DestTy: X.getAddress().getElementType(), |
| 6844 | isSigned: X.getType()->hasSignedIntegerRepresentation()); |
| 6845 | else |
| 6846 | UpdateVal = CGF.Builder.CreateCast(Op: llvm::Instruction::CastOps::UIToFP, V: IC, |
| 6847 | DestTy: X.getAddress().getElementType()); |
| 6848 | } |
| 6849 | llvm::AtomicRMWInst *Res = |
| 6850 | CGF.emitAtomicRMWInst(Op: RMWOp, Addr: X.getAddress(), Val: UpdateVal, Order: AO); |
| 6851 | return std::make_pair(x: true, y: RValue::get(V: Res)); |
| 6852 | } |
| 6853 | |
| 6854 | std::pair<bool, RValue> CodeGenFunction::EmitOMPAtomicSimpleUpdateExpr( |
| 6855 | LValue X, RValue E, BinaryOperatorKind BO, bool IsXLHSInRHSPart, |
| 6856 | llvm::AtomicOrdering AO, SourceLocation Loc, |
| 6857 | const llvm::function_ref<RValue(RValue)> CommonGen) { |
| 6858 | // Update expressions are allowed to have the following forms: |
| 6859 | // x binop= expr; -> xrval + expr; |
| 6860 | // x++, ++x -> xrval + 1; |
| 6861 | // x--, --x -> xrval - 1; |
| 6862 | // x = x binop expr; -> xrval binop expr |
| 6863 | // x = expr Op x; - > expr binop xrval; |
| 6864 | auto Res = emitOMPAtomicRMW(CGF&: *this, X, Update: E, BO, AO, IsXLHSInRHSPart); |
| 6865 | if (!Res.first) { |
| 6866 | if (X.isGlobalReg()) { |
| 6867 | // Emit an update expression: 'xrval' binop 'expr' or 'expr' binop |
| 6868 | // 'xrval'. |
| 6869 | EmitStoreThroughLValue(Src: CommonGen(EmitLoadOfLValue(V: X, Loc)), Dst: X); |
| 6870 | } else { |
| 6871 | // Perform compare-and-swap procedure. |
| 6872 | EmitAtomicUpdate(LVal: X, AO, UpdateOp: CommonGen, IsVolatile: X.getType().isVolatileQualified()); |
| 6873 | } |
| 6874 | } |
| 6875 | return Res; |
| 6876 | } |
| 6877 | |
| 6878 | static void emitOMPAtomicUpdateExpr(CodeGenFunction &CGF, |
| 6879 | llvm::AtomicOrdering AO, const Expr *X, |
| 6880 | const Expr *E, const Expr *UE, |
| 6881 | bool IsXLHSInRHSPart, SourceLocation Loc) { |
| 6882 | assert(isa<BinaryOperator>(UE->IgnoreImpCasts()) && |
| 6883 | "Update expr in 'atomic update' must be a binary operator."); |
| 6884 | const auto *BOUE = cast<BinaryOperator>(Val: UE->IgnoreImpCasts()); |
| 6885 | // Update expressions are allowed to have the following forms: |
| 6886 | // x binop= expr; -> xrval + expr; |
| 6887 | // x++, ++x -> xrval + 1; |
| 6888 | // x--, --x -> xrval - 1; |
| 6889 | // x = x binop expr; -> xrval binop expr |
| 6890 | // x = expr Op x; - > expr binop xrval; |
| 6891 | assert(X->isLValue() && "X of 'omp atomic update' is not lvalue"); |
| 6892 | LValue XLValue = CGF.EmitLValue(E: X); |
| 6893 | RValue ExprRValue = CGF.EmitAnyExpr(E); |
| 6894 | const auto *LHS = cast<OpaqueValueExpr>(Val: BOUE->getLHS()->IgnoreImpCasts()); |
| 6895 | const auto *RHS = cast<OpaqueValueExpr>(Val: BOUE->getRHS()->IgnoreImpCasts()); |
| 6896 | const OpaqueValueExpr *XRValExpr = IsXLHSInRHSPart ? LHS : RHS; |
| 6897 | const OpaqueValueExpr *ERValExpr = IsXLHSInRHSPart ? RHS : LHS; |
| 6898 | auto &&Gen = [&CGF, UE, ExprRValue, XRValExpr, ERValExpr](RValue XRValue) { |
| 6899 | CodeGenFunction::OpaqueValueMapping MapExpr(CGF, ERValExpr, ExprRValue); |
| 6900 | CodeGenFunction::OpaqueValueMapping MapX(CGF, XRValExpr, XRValue); |
| 6901 | return CGF.EmitAnyExpr(E: UE); |
| 6902 | }; |
| 6903 | (void)CGF.EmitOMPAtomicSimpleUpdateExpr( |
| 6904 | X: XLValue, E: ExprRValue, BO: BOUE->getOpcode(), IsXLHSInRHSPart, AO, Loc, CommonGen: Gen); |
| 6905 | CGF.CGM.getOpenMPRuntime().checkAndEmitLastprivateConditional(CGF, LHS: X); |
| 6906 | // OpenMP, 2.17.7, atomic Construct |
| 6907 | // If the write, update, or capture clause is specified and the release, |
| 6908 | // acq_rel, or seq_cst clause is specified then the strong flush on entry to |
| 6909 | // the atomic operation is also a release flush. |
| 6910 | switch (AO) { |
| 6911 | case llvm::AtomicOrdering::Release: |
| 6912 | case llvm::AtomicOrdering::AcquireRelease: |
| 6913 | case llvm::AtomicOrdering::SequentiallyConsistent: |
| 6914 | CGF.CGM.getOpenMPRuntime().emitFlush(CGF, Vars: {}, Loc, |
| 6915 | AO: llvm::AtomicOrdering::Release); |
| 6916 | break; |
| 6917 | case llvm::AtomicOrdering::Acquire: |
| 6918 | case llvm::AtomicOrdering::Monotonic: |
| 6919 | break; |
| 6920 | case llvm::AtomicOrdering::NotAtomic: |
| 6921 | case llvm::AtomicOrdering::Unordered: |
| 6922 | llvm_unreachable("Unexpected ordering."); |
| 6923 | } |
| 6924 | } |
| 6925 | |
| 6926 | static RValue convertToType(CodeGenFunction &CGF, RValue Value, |
| 6927 | QualType SourceType, QualType ResType, |
| 6928 | SourceLocation Loc) { |
| 6929 | switch (CGF.getEvaluationKind(T: ResType)) { |
| 6930 | case TEK_Scalar: |
| 6931 | return RValue::get( |
| 6932 | V: convertToScalarValue(CGF, Val: Value, SrcType: SourceType, DestType: ResType, Loc)); |
| 6933 | case TEK_Complex: { |
| 6934 | auto Res = convertToComplexValue(CGF, Val: Value, SrcType: SourceType, DestType: ResType, Loc); |
| 6935 | return RValue::getComplex(V1: Res.first, V2: Res.second); |
| 6936 | } |
| 6937 | case TEK_Aggregate: |
| 6938 | break; |
| 6939 | } |
| 6940 | llvm_unreachable("Must be a scalar or complex."); |
| 6941 | } |
| 6942 | |
| 6943 | static void emitOMPAtomicCaptureExpr(CodeGenFunction &CGF, |
| 6944 | llvm::AtomicOrdering AO, |
| 6945 | bool IsPostfixUpdate, const Expr *V, |
| 6946 | const Expr *X, const Expr *E, |
| 6947 | const Expr *UE, bool IsXLHSInRHSPart, |
| 6948 | SourceLocation Loc) { |
| 6949 | assert(X->isLValue() && "X of 'omp atomic capture' is not lvalue"); |
| 6950 | assert(V->isLValue() && "V of 'omp atomic capture' is not lvalue"); |
| 6951 | RValue NewVVal; |
| 6952 | LValue VLValue = CGF.EmitLValue(E: V); |
| 6953 | LValue XLValue = CGF.EmitLValue(E: X); |
| 6954 | RValue ExprRValue = CGF.EmitAnyExpr(E); |
| 6955 | QualType NewVValType; |
| 6956 | if (UE) { |
| 6957 | // 'x' is updated with some additional value. |
| 6958 | assert(isa<BinaryOperator>(UE->IgnoreImpCasts()) && |
| 6959 | "Update expr in 'atomic capture' must be a binary operator."); |
| 6960 | const auto *BOUE = cast<BinaryOperator>(Val: UE->IgnoreImpCasts()); |
| 6961 | // Update expressions are allowed to have the following forms: |
| 6962 | // x binop= expr; -> xrval + expr; |
| 6963 | // x++, ++x -> xrval + 1; |
| 6964 | // x--, --x -> xrval - 1; |
| 6965 | // x = x binop expr; -> xrval binop expr |
| 6966 | // x = expr Op x; - > expr binop xrval; |
| 6967 | const auto *LHS = cast<OpaqueValueExpr>(Val: BOUE->getLHS()->IgnoreImpCasts()); |
| 6968 | const auto *RHS = cast<OpaqueValueExpr>(Val: BOUE->getRHS()->IgnoreImpCasts()); |
| 6969 | const OpaqueValueExpr *XRValExpr = IsXLHSInRHSPart ? LHS : RHS; |
| 6970 | NewVValType = XRValExpr->getType(); |
| 6971 | const OpaqueValueExpr *ERValExpr = IsXLHSInRHSPart ? RHS : LHS; |
| 6972 | auto &&Gen = [&CGF, &NewVVal, UE, ExprRValue, XRValExpr, ERValExpr, |
| 6973 | IsPostfixUpdate](RValue XRValue) { |
| 6974 | CodeGenFunction::OpaqueValueMapping MapExpr(CGF, ERValExpr, ExprRValue); |
| 6975 | CodeGenFunction::OpaqueValueMapping MapX(CGF, XRValExpr, XRValue); |
| 6976 | RValue Res = CGF.EmitAnyExpr(E: UE); |
| 6977 | NewVVal = IsPostfixUpdate ? XRValue : Res; |
| 6978 | return Res; |
| 6979 | }; |
| 6980 | auto Res = CGF.EmitOMPAtomicSimpleUpdateExpr( |
| 6981 | X: XLValue, E: ExprRValue, BO: BOUE->getOpcode(), IsXLHSInRHSPart, AO, Loc, CommonGen: Gen); |
| 6982 | CGF.CGM.getOpenMPRuntime().checkAndEmitLastprivateConditional(CGF, LHS: X); |
| 6983 | if (Res.first) { |
| 6984 | // 'atomicrmw' instruction was generated. |
| 6985 | if (IsPostfixUpdate) { |
| 6986 | // Use old value from 'atomicrmw'. |
| 6987 | NewVVal = Res.second; |
| 6988 | } else { |
| 6989 | // 'atomicrmw' does not provide new value, so evaluate it using old |
| 6990 | // value of 'x'. |
| 6991 | CodeGenFunction::OpaqueValueMapping MapExpr(CGF, ERValExpr, ExprRValue); |
| 6992 | CodeGenFunction::OpaqueValueMapping MapX(CGF, XRValExpr, Res.second); |
| 6993 | NewVVal = CGF.EmitAnyExpr(E: UE); |
| 6994 | } |
| 6995 | } |
| 6996 | } else { |
| 6997 | // 'x' is simply rewritten with some 'expr'. |
| 6998 | NewVValType = X->getType().getNonReferenceType(); |
| 6999 | ExprRValue = convertToType(CGF, Value: ExprRValue, SourceType: E->getType(), |
| 7000 | ResType: X->getType().getNonReferenceType(), Loc); |
| 7001 | auto &&Gen = [&NewVVal, ExprRValue](RValue XRValue) { |
| 7002 | NewVVal = XRValue; |
| 7003 | return ExprRValue; |
| 7004 | }; |
| 7005 | // Try to perform atomicrmw xchg, otherwise simple exchange. |
| 7006 | auto Res = CGF.EmitOMPAtomicSimpleUpdateExpr( |
| 7007 | X: XLValue, E: ExprRValue, /*BO=*/BO_Assign, /*IsXLHSInRHSPart=*/false, AO, |
| 7008 | Loc, CommonGen: Gen); |
| 7009 | CGF.CGM.getOpenMPRuntime().checkAndEmitLastprivateConditional(CGF, LHS: X); |
| 7010 | if (Res.first) { |
| 7011 | // 'atomicrmw' instruction was generated. |
| 7012 | NewVVal = IsPostfixUpdate ? Res.second : ExprRValue; |
| 7013 | } |
| 7014 | } |
| 7015 | // Emit post-update store to 'v' of old/new 'x' value. |
| 7016 | CGF.emitOMPSimpleStore(LVal: VLValue, RVal: NewVVal, RValTy: NewVValType, Loc); |
| 7017 | CGF.CGM.getOpenMPRuntime().checkAndEmitLastprivateConditional(CGF, LHS: V); |
| 7018 | // OpenMP 5.1 removes the required flush for capture clause. |
| 7019 | if (CGF.CGM.getLangOpts().OpenMP < 51) { |
| 7020 | // OpenMP, 2.17.7, atomic Construct |
| 7021 | // If the write, update, or capture clause is specified and the release, |
| 7022 | // acq_rel, or seq_cst clause is specified then the strong flush on entry to |
| 7023 | // the atomic operation is also a release flush. |
| 7024 | // If the read or capture clause is specified and the acquire, acq_rel, or |
| 7025 | // seq_cst clause is specified then the strong flush on exit from the atomic |
| 7026 | // operation is also an acquire flush. |
| 7027 | switch (AO) { |
| 7028 | case llvm::AtomicOrdering::Release: |
| 7029 | CGF.CGM.getOpenMPRuntime().emitFlush(CGF, Vars: {}, Loc, |
| 7030 | AO: llvm::AtomicOrdering::Release); |
| 7031 | break; |
| 7032 | case llvm::AtomicOrdering::Acquire: |
| 7033 | CGF.CGM.getOpenMPRuntime().emitFlush(CGF, Vars: {}, Loc, |
| 7034 | AO: llvm::AtomicOrdering::Acquire); |
| 7035 | break; |
| 7036 | case llvm::AtomicOrdering::AcquireRelease: |
| 7037 | case llvm::AtomicOrdering::SequentiallyConsistent: |
| 7038 | CGF.CGM.getOpenMPRuntime().emitFlush( |
| 7039 | CGF, Vars: {}, Loc, AO: llvm::AtomicOrdering::AcquireRelease); |
| 7040 | break; |
| 7041 | case llvm::AtomicOrdering::Monotonic: |
| 7042 | break; |
| 7043 | case llvm::AtomicOrdering::NotAtomic: |
| 7044 | case llvm::AtomicOrdering::Unordered: |
| 7045 | llvm_unreachable("Unexpected ordering."); |
| 7046 | } |
| 7047 | } |
| 7048 | } |
| 7049 | |
| 7050 | static void emitOMPAtomicCompareExpr( |
| 7051 | CodeGenFunction &CGF, llvm::AtomicOrdering AO, llvm::AtomicOrdering FailAO, |
| 7052 | const Expr *X, const Expr *V, const Expr *R, const Expr *E, const Expr *D, |
| 7053 | const Expr *CE, bool IsXBinopExpr, bool IsPostfixUpdate, bool IsFailOnly, |
| 7054 | SourceLocation Loc) { |
| 7055 | llvm::OpenMPIRBuilder &OMPBuilder = |
| 7056 | CGF.CGM.getOpenMPRuntime().getOMPBuilder(); |
| 7057 | |
| 7058 | OMPAtomicCompareOp Op; |
| 7059 | assert(isa<BinaryOperator>(CE) && "CE is not a BinaryOperator"); |
| 7060 | switch (cast<BinaryOperator>(Val: CE)->getOpcode()) { |
| 7061 | case BO_EQ: |
| 7062 | Op = OMPAtomicCompareOp::EQ; |
| 7063 | break; |
| 7064 | case BO_LT: |
| 7065 | Op = OMPAtomicCompareOp::MIN; |
| 7066 | break; |
| 7067 | case BO_GT: |
| 7068 | Op = OMPAtomicCompareOp::MAX; |
| 7069 | break; |
| 7070 | default: |
| 7071 | llvm_unreachable("unsupported atomic compare binary operator"); |
| 7072 | } |
| 7073 | |
| 7074 | LValue XLVal = CGF.EmitLValue(E: X); |
| 7075 | Address XAddr = XLVal.getAddress(); |
| 7076 | |
| 7077 | auto EmitRValueWithCastIfNeeded = [&CGF, Loc](const Expr *X, const Expr *E) { |
| 7078 | if (X->getType() == E->getType()) |
| 7079 | return CGF.EmitScalarExpr(E); |
| 7080 | const Expr *NewE = E->IgnoreImplicitAsWritten(); |
| 7081 | llvm::Value *V = CGF.EmitScalarExpr(E: NewE); |
| 7082 | if (NewE->getType() == X->getType()) |
| 7083 | return V; |
| 7084 | return CGF.EmitScalarConversion(Src: V, SrcTy: NewE->getType(), DstTy: X->getType(), Loc); |
| 7085 | }; |
| 7086 | |
| 7087 | llvm::Value *EVal = EmitRValueWithCastIfNeeded(X, E); |
| 7088 | llvm::Value *DVal = D ? EmitRValueWithCastIfNeeded(X, D) : nullptr; |
| 7089 | if (auto *CI = dyn_cast<llvm::ConstantInt>(Val: EVal)) |
| 7090 | EVal = CGF.Builder.CreateIntCast( |
| 7091 | V: CI, DestTy: XLVal.getAddress().getElementType(), |
| 7092 | isSigned: E->getType()->hasSignedIntegerRepresentation()); |
| 7093 | if (DVal) |
| 7094 | if (auto *CI = dyn_cast<llvm::ConstantInt>(Val: DVal)) |
| 7095 | DVal = CGF.Builder.CreateIntCast( |
| 7096 | V: CI, DestTy: XLVal.getAddress().getElementType(), |
| 7097 | isSigned: D->getType()->hasSignedIntegerRepresentation()); |
| 7098 | |
| 7099 | llvm::OpenMPIRBuilder::AtomicOpValue XOpVal{ |
| 7100 | .Var: XAddr.emitRawPointer(CGF), .ElemTy: XAddr.getElementType(), |
| 7101 | .IsSigned: X->getType()->hasSignedIntegerRepresentation(), |
| 7102 | .IsVolatile: X->getType().isVolatileQualified()}; |
| 7103 | llvm::OpenMPIRBuilder::AtomicOpValue VOpVal, ROpVal; |
| 7104 | if (V) { |
| 7105 | LValue LV = CGF.EmitLValue(E: V); |
| 7106 | Address Addr = LV.getAddress(); |
| 7107 | VOpVal = {.Var: Addr.emitRawPointer(CGF), .ElemTy: Addr.getElementType(), |
| 7108 | .IsSigned: V->getType()->hasSignedIntegerRepresentation(), |
| 7109 | .IsVolatile: V->getType().isVolatileQualified()}; |
| 7110 | } |
| 7111 | if (R) { |
| 7112 | LValue LV = CGF.EmitLValue(E: R); |
| 7113 | Address Addr = LV.getAddress(); |
| 7114 | ROpVal = {.Var: Addr.emitRawPointer(CGF), .ElemTy: Addr.getElementType(), |
| 7115 | .IsSigned: R->getType()->hasSignedIntegerRepresentation(), |
| 7116 | .IsVolatile: R->getType().isVolatileQualified()}; |
| 7117 | } |
| 7118 | |
| 7119 | if (FailAO == llvm::AtomicOrdering::NotAtomic) { |
| 7120 | // fail clause was not mentioned on the |
| 7121 | // "#pragma omp atomic compare" construct. |
| 7122 | CGF.Builder.restoreIP(IP: OMPBuilder.createAtomicCompare( |
| 7123 | Loc: CGF.Builder, X&: XOpVal, V&: VOpVal, R&: ROpVal, E: EVal, D: DVal, AO, Op, IsXBinopExpr, |
| 7124 | IsPostfixUpdate, IsFailOnly)); |
| 7125 | } else |
| 7126 | CGF.Builder.restoreIP(IP: OMPBuilder.createAtomicCompare( |
| 7127 | Loc: CGF.Builder, X&: XOpVal, V&: VOpVal, R&: ROpVal, E: EVal, D: DVal, AO, Op, IsXBinopExpr, |
| 7128 | IsPostfixUpdate, IsFailOnly, Failure: FailAO)); |
| 7129 | } |
| 7130 | |
| 7131 | static void emitOMPAtomicExpr(CodeGenFunction &CGF, OpenMPClauseKind Kind, |
| 7132 | llvm::AtomicOrdering AO, |
| 7133 | llvm::AtomicOrdering FailAO, bool IsPostfixUpdate, |
| 7134 | const Expr *X, const Expr *V, const Expr *R, |
| 7135 | const Expr *E, const Expr *UE, const Expr *D, |
| 7136 | const Expr *CE, bool IsXLHSInRHSPart, |
| 7137 | bool IsFailOnly, SourceLocation Loc) { |
| 7138 | switch (Kind) { |
| 7139 | case OMPC_read: |
| 7140 | emitOMPAtomicReadExpr(CGF, AO, X, V, Loc); |
| 7141 | break; |
| 7142 | case OMPC_write: |
| 7143 | emitOMPAtomicWriteExpr(CGF, AO, X, E, Loc); |
| 7144 | break; |
| 7145 | case OMPC_unknown: |
| 7146 | case OMPC_update: |
| 7147 | emitOMPAtomicUpdateExpr(CGF, AO, X, E, UE, IsXLHSInRHSPart, Loc); |
| 7148 | break; |
| 7149 | case OMPC_capture: |
| 7150 | emitOMPAtomicCaptureExpr(CGF, AO, IsPostfixUpdate, V, X, E, UE, |
| 7151 | IsXLHSInRHSPart, Loc); |
| 7152 | break; |
| 7153 | case OMPC_compare: { |
| 7154 | emitOMPAtomicCompareExpr(CGF, AO, FailAO, X, V, R, E, D, CE, |
| 7155 | IsXBinopExpr: IsXLHSInRHSPart, IsPostfixUpdate, IsFailOnly, Loc); |
| 7156 | break; |
| 7157 | } |
| 7158 | default: |
| 7159 | llvm_unreachable("Clause is not allowed in 'omp atomic'."); |
| 7160 | } |
| 7161 | } |
| 7162 | |
| 7163 | void CodeGenFunction::EmitOMPAtomicDirective(const OMPAtomicDirective &S) { |
| 7164 | llvm::AtomicOrdering AO = CGM.getOpenMPRuntime().getDefaultMemoryOrdering(); |
| 7165 | // Fail Memory Clause Ordering. |
| 7166 | llvm::AtomicOrdering FailAO = llvm::AtomicOrdering::NotAtomic; |
| 7167 | bool MemOrderingSpecified = false; |
| 7168 | if (S.getSingleClause<OMPSeqCstClause>()) { |
| 7169 | AO = llvm::AtomicOrdering::SequentiallyConsistent; |
| 7170 | MemOrderingSpecified = true; |
| 7171 | } else if (S.getSingleClause<OMPAcqRelClause>()) { |
| 7172 | AO = llvm::AtomicOrdering::AcquireRelease; |
| 7173 | MemOrderingSpecified = true; |
| 7174 | } else if (S.getSingleClause<OMPAcquireClause>()) { |
| 7175 | AO = llvm::AtomicOrdering::Acquire; |
| 7176 | MemOrderingSpecified = true; |
| 7177 | } else if (S.getSingleClause<OMPReleaseClause>()) { |
| 7178 | AO = llvm::AtomicOrdering::Release; |
| 7179 | MemOrderingSpecified = true; |
| 7180 | } else if (S.getSingleClause<OMPRelaxedClause>()) { |
| 7181 | AO = llvm::AtomicOrdering::Monotonic; |
| 7182 | MemOrderingSpecified = true; |
| 7183 | } |
| 7184 | llvm::SmallSet<OpenMPClauseKind, 2> KindsEncountered; |
| 7185 | OpenMPClauseKind Kind = OMPC_unknown; |
| 7186 | for (const OMPClause *C : S.clauses()) { |
| 7187 | // Find first clause (skip seq_cst|acq_rel|aqcuire|release|relaxed clause, |
| 7188 | // if it is first). |
| 7189 | OpenMPClauseKind K = C->getClauseKind(); |
| 7190 | // TBD |
| 7191 | if (K == OMPC_weak) |
| 7192 | return; |
| 7193 | if (K == OMPC_seq_cst || K == OMPC_acq_rel || K == OMPC_acquire || |
| 7194 | K == OMPC_release || K == OMPC_relaxed || K == OMPC_hint) |
| 7195 | continue; |
| 7196 | Kind = K; |
| 7197 | KindsEncountered.insert(V: K); |
| 7198 | } |
| 7199 | // We just need to correct Kind here. No need to set a bool saying it is |
| 7200 | // actually compare capture because we can tell from whether V and R are |
| 7201 | // nullptr. |
| 7202 | if (KindsEncountered.contains(V: OMPC_compare) && |
| 7203 | KindsEncountered.contains(V: OMPC_capture)) |
| 7204 | Kind = OMPC_compare; |
| 7205 | if (!MemOrderingSpecified) { |
| 7206 | llvm::AtomicOrdering DefaultOrder = |
| 7207 | CGM.getOpenMPRuntime().getDefaultMemoryOrdering(); |
| 7208 | if (DefaultOrder == llvm::AtomicOrdering::Monotonic || |
| 7209 | DefaultOrder == llvm::AtomicOrdering::SequentiallyConsistent || |
| 7210 | (DefaultOrder == llvm::AtomicOrdering::AcquireRelease && |
| 7211 | Kind == OMPC_capture)) { |
| 7212 | AO = DefaultOrder; |
| 7213 | } else if (DefaultOrder == llvm::AtomicOrdering::AcquireRelease) { |
| 7214 | if (Kind == OMPC_unknown || Kind == OMPC_update || Kind == OMPC_write) { |
| 7215 | AO = llvm::AtomicOrdering::Release; |
| 7216 | } else if (Kind == OMPC_read) { |
| 7217 | assert(Kind == OMPC_read && "Unexpected atomic kind."); |
| 7218 | AO = llvm::AtomicOrdering::Acquire; |
| 7219 | } |
| 7220 | } |
| 7221 | } |
| 7222 | |
| 7223 | if (KindsEncountered.contains(V: OMPC_compare) && |
| 7224 | KindsEncountered.contains(V: OMPC_fail)) { |
| 7225 | Kind = OMPC_compare; |
| 7226 | const auto *FailClause = S.getSingleClause<OMPFailClause>(); |
| 7227 | if (FailClause) { |
| 7228 | OpenMPClauseKind FailParameter = FailClause->getFailParameter(); |
| 7229 | if (FailParameter == llvm::omp::OMPC_relaxed) |
| 7230 | FailAO = llvm::AtomicOrdering::Monotonic; |
| 7231 | else if (FailParameter == llvm::omp::OMPC_acquire) |
| 7232 | FailAO = llvm::AtomicOrdering::Acquire; |
| 7233 | else if (FailParameter == llvm::omp::OMPC_seq_cst) |
| 7234 | FailAO = llvm::AtomicOrdering::SequentiallyConsistent; |
| 7235 | } |
| 7236 | } |
| 7237 | |
| 7238 | LexicalScope Scope(*this, S.getSourceRange()); |
| 7239 | EmitStopPoint(S: S.getAssociatedStmt()); |
| 7240 | emitOMPAtomicExpr(CGF&: *this, Kind, AO, FailAO, IsPostfixUpdate: S.isPostfixUpdate(), X: S.getX(), |
| 7241 | V: S.getV(), R: S.getR(), E: S.getExpr(), UE: S.getUpdateExpr(), |
| 7242 | D: S.getD(), CE: S.getCondExpr(), IsXLHSInRHSPart: S.isXLHSInRHSPart(), |
| 7243 | IsFailOnly: S.isFailOnly(), Loc: S.getBeginLoc()); |
| 7244 | } |
| 7245 | |
| 7246 | static void emitCommonOMPTargetDirective(CodeGenFunction &CGF, |
| 7247 | const OMPExecutableDirective &S, |
| 7248 | const RegionCodeGenTy &CodeGen) { |
| 7249 | assert(isOpenMPTargetExecutionDirective(S.getDirectiveKind())); |
| 7250 | CodeGenModule &CGM = CGF.CGM; |
| 7251 | |
| 7252 | // On device emit this construct as inlined code. |
| 7253 | if (CGM.getLangOpts().OpenMPIsTargetDevice) { |
| 7254 | OMPLexicalScope Scope(CGF, S, OMPD_target); |
| 7255 | CGM.getOpenMPRuntime().emitInlinedDirective( |
| 7256 | CGF, InnermostKind: OMPD_target, CodeGen: [&S](CodeGenFunction &CGF, PrePostActionTy &) { |
| 7257 | CGF.EmitStmt(S: S.getInnermostCapturedStmt()->getCapturedStmt()); |
| 7258 | }); |
| 7259 | return; |
| 7260 | } |
| 7261 | |
| 7262 | auto LPCRegion = CGOpenMPRuntime::LastprivateConditionalRAII::disable(CGF, S); |
| 7263 | llvm::Function *Fn = nullptr; |
| 7264 | llvm::Constant *FnID = nullptr; |
| 7265 | |
| 7266 | const Expr *IfCond = nullptr; |
| 7267 | // Check for the at most one if clause associated with the target region. |
| 7268 | for (const auto *C : S.getClausesOfKind<OMPIfClause>()) { |
| 7269 | if (C->getNameModifier() == OMPD_unknown || |
| 7270 | C->getNameModifier() == OMPD_target) { |
| 7271 | IfCond = C->getCondition(); |
| 7272 | break; |
| 7273 | } |
| 7274 | } |
| 7275 | |
| 7276 | // Check if we have any device clause associated with the directive. |
| 7277 | llvm::PointerIntPair<const Expr *, 2, OpenMPDeviceClauseModifier> Device( |
| 7278 | nullptr, OMPC_DEVICE_unknown); |
| 7279 | if (auto *C = S.getSingleClause<OMPDeviceClause>()) |
| 7280 | Device.setPointerAndInt(PtrVal: C->getDevice(), IntVal: C->getModifier()); |
| 7281 | |
| 7282 | // Check if we have an if clause whose conditional always evaluates to false |
| 7283 | // or if we do not have any targets specified. If so the target region is not |
| 7284 | // an offload entry point. |
| 7285 | bool IsOffloadEntry = true; |
| 7286 | if (IfCond) { |
| 7287 | bool Val; |
| 7288 | if (CGF.ConstantFoldsToSimpleInteger(Cond: IfCond, Result&: Val) && !Val) |
| 7289 | IsOffloadEntry = false; |
| 7290 | } |
| 7291 | if (CGM.getLangOpts().OMPTargetTriples.empty()) |
| 7292 | IsOffloadEntry = false; |
| 7293 | |
| 7294 | if (CGM.getLangOpts().OpenMPOffloadMandatory && !IsOffloadEntry) { |
| 7295 | CGM.getDiags().Report(DiagID: diag::err_missing_mandatory_offloading); |
| 7296 | } |
| 7297 | |
| 7298 | assert(CGF.CurFuncDecl && "No parent declaration for target region!"); |
| 7299 | StringRef ParentName; |
| 7300 | // In case we have Ctors/Dtors we use the complete type variant to produce |
| 7301 | // the mangling of the device outlined kernel. |
| 7302 | if (const auto *D = dyn_cast<CXXConstructorDecl>(Val: CGF.CurFuncDecl)) |
| 7303 | ParentName = CGM.getMangledName(GD: GlobalDecl(D, Ctor_Complete)); |
| 7304 | else if (const auto *D = dyn_cast<CXXDestructorDecl>(Val: CGF.CurFuncDecl)) |
| 7305 | ParentName = CGM.getMangledName(GD: GlobalDecl(D, Dtor_Complete)); |
| 7306 | else |
| 7307 | ParentName = |
| 7308 | CGM.getMangledName(GD: GlobalDecl(cast<FunctionDecl>(Val: CGF.CurFuncDecl))); |
| 7309 | |
| 7310 | // Emit target region as a standalone region. |
| 7311 | CGM.getOpenMPRuntime().emitTargetOutlinedFunction(D: S, ParentName, OutlinedFn&: Fn, OutlinedFnID&: FnID, |
| 7312 | IsOffloadEntry, CodeGen); |
| 7313 | OMPLexicalScope Scope(CGF, S, OMPD_task); |
| 7314 | auto &&SizeEmitter = |
| 7315 | [IsOffloadEntry](CodeGenFunction &CGF, |
| 7316 | const OMPLoopDirective &D) -> llvm::Value * { |
| 7317 | if (IsOffloadEntry) { |
| 7318 | OMPLoopScope PreInitScope(CGF, D); |
| 7319 | // Emit calculation of the iterations count. |
| 7320 | llvm::Value *NumIterations = CGF.EmitScalarExpr(E: D.getNumIterations()); |
| 7321 | NumIterations = CGF.Builder.CreateIntCast(V: NumIterations, DestTy: CGF.Int64Ty, |
| 7322 | /*isSigned=*/false); |
| 7323 | return NumIterations; |
| 7324 | } |
| 7325 | return nullptr; |
| 7326 | }; |
| 7327 | CGM.getOpenMPRuntime().emitTargetCall(CGF, D: S, OutlinedFn: Fn, OutlinedFnID: FnID, IfCond, Device, |
| 7328 | SizeEmitter); |
| 7329 | } |
| 7330 | |
| 7331 | static void emitTargetRegion(CodeGenFunction &CGF, const OMPTargetDirective &S, |
| 7332 | PrePostActionTy &Action) { |
| 7333 | Action.Enter(CGF); |
| 7334 | CodeGenFunction::OMPPrivateScope PrivateScope(CGF); |
| 7335 | (void)CGF.EmitOMPFirstprivateClause(D: S, PrivateScope); |
| 7336 | CGF.EmitOMPPrivateClause(D: S, PrivateScope); |
| 7337 | (void)PrivateScope.Privatize(); |
| 7338 | if (isOpenMPTargetExecutionDirective(DKind: S.getDirectiveKind())) |
| 7339 | CGF.CGM.getOpenMPRuntime().adjustTargetSpecificDataForLambdas(CGF, D: S); |
| 7340 | |
| 7341 | CGF.EmitStmt(S: S.getCapturedStmt(RegionKind: OMPD_target)->getCapturedStmt()); |
| 7342 | CGF.EnsureInsertPoint(); |
| 7343 | } |
| 7344 | |
| 7345 | void CodeGenFunction::EmitOMPTargetDeviceFunction(CodeGenModule &CGM, |
| 7346 | StringRef ParentName, |
| 7347 | const OMPTargetDirective &S) { |
| 7348 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 7349 | emitTargetRegion(CGF, S, Action); |
| 7350 | }; |
| 7351 | llvm::Function *Fn; |
| 7352 | llvm::Constant *Addr; |
| 7353 | // Emit target region as a standalone region. |
| 7354 | CGM.getOpenMPRuntime().emitTargetOutlinedFunction( |
| 7355 | D: S, ParentName, OutlinedFn&: Fn, OutlinedFnID&: Addr, /*IsOffloadEntry=*/true, CodeGen); |
| 7356 | assert(Fn && Addr && "Target device function emission failed."); |
| 7357 | } |
| 7358 | |
| 7359 | void CodeGenFunction::EmitOMPTargetDirective(const OMPTargetDirective &S) { |
| 7360 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 7361 | emitTargetRegion(CGF, S, Action); |
| 7362 | }; |
| 7363 | emitCommonOMPTargetDirective(CGF&: *this, S, CodeGen); |
| 7364 | } |
| 7365 | |
| 7366 | static void emitCommonOMPTeamsDirective(CodeGenFunction &CGF, |
| 7367 | const OMPExecutableDirective &S, |
| 7368 | OpenMPDirectiveKind InnermostKind, |
| 7369 | const RegionCodeGenTy &CodeGen) { |
| 7370 | const CapturedStmt *CS = S.getCapturedStmt(RegionKind: OMPD_teams); |
| 7371 | llvm::Function *OutlinedFn = |
| 7372 | CGF.CGM.getOpenMPRuntime().emitTeamsOutlinedFunction( |
| 7373 | CGF, D: S, ThreadIDVar: *CS->getCapturedDecl()->param_begin(), InnermostKind, |
| 7374 | CodeGen); |
| 7375 | |
| 7376 | OMPTeamsScope Scope(CGF, S); |
| 7377 | auto ParallelLeague = [&S](CodeGenFunction &CGF, PrePostActionTy &) { |
| 7378 | const auto *NT = S.getSingleClause<OMPNumTeamsClause>(); |
| 7379 | const auto *TL = S.getSingleClause<OMPThreadLimitClause>(); |
| 7380 | if (NT || TL) { |
| 7381 | const Expr *NumTeams = NT ? NT->getNumTeams().front() : nullptr; |
| 7382 | const Expr *ThreadLimit = TL ? TL->getThreadLimit().front() : nullptr; |
| 7383 | |
| 7384 | CGF.CGM.getOpenMPRuntime().emitNumTeamsClause(CGF, NumTeams, ThreadLimit, |
| 7385 | Loc: S.getBeginLoc()); |
| 7386 | } |
| 7387 | }; |
| 7388 | |
| 7389 | const Expr *IfCond = nullptr; |
| 7390 | for (const auto *C : S.getClausesOfKind<OMPIfClause>()) { |
| 7391 | if (C->getNameModifier() == OMPD_unknown || |
| 7392 | C->getNameModifier() == OMPD_teams) { |
| 7393 | IfCond = C->getCondition(); |
| 7394 | break; |
| 7395 | } |
| 7396 | } |
| 7397 | if (IfCond && CGF.CGM.getLangOpts().OpenMP >= 52) { |
| 7398 | auto SerialLeague = [&S](CodeGenFunction &CGF, PrePostActionTy &) { |
| 7399 | // OpenMP 5.2, 10.2, teams Construct |
| 7400 | // When an if clause is present on a teams construct and the if clause |
| 7401 | // expression evaluates to false, the number of created teams is one. |
| 7402 | const llvm::APInt One(32, 1); |
| 7403 | IntegerLiteral NumTeams( |
| 7404 | CGF.getContext(), One, |
| 7405 | CGF.getContext().getIntTypeForBitwidth(DestWidth: 32, /*Signed=*/0), |
| 7406 | SourceLocation()); |
| 7407 | // The thread_limit clause is unaffected by the if clause. |
| 7408 | const auto *TL = S.getSingleClause<OMPThreadLimitClause>(); |
| 7409 | const Expr *ThreadLimit = TL ? TL->getThreadLimit().front() : nullptr; |
| 7410 | CGF.CGM.getOpenMPRuntime().emitNumTeamsClause(CGF, NumTeams: &NumTeams, ThreadLimit, |
| 7411 | Loc: S.getBeginLoc()); |
| 7412 | }; |
| 7413 | CGF.CGM.getOpenMPRuntime().emitIfClause(CGF, Cond: IfCond, ThenGen: ParallelLeague, |
| 7414 | ElseGen: SerialLeague); |
| 7415 | } else { |
| 7416 | const RegionCodeGenTy ThenRCG(ParallelLeague); |
| 7417 | ThenRCG(CGF); |
| 7418 | } |
| 7419 | |
| 7420 | llvm::SmallVector<llvm::Value *, 16> CapturedVars; |
| 7421 | CGF.GenerateOpenMPCapturedVars(S: *CS, CapturedVars); |
| 7422 | CGF.CGM.getOpenMPRuntime().emitTeamsCall(CGF, D: S, Loc: S.getBeginLoc(), OutlinedFn, |
| 7423 | CapturedVars); |
| 7424 | } |
| 7425 | |
| 7426 | void CodeGenFunction::EmitOMPTeamsDirective(const OMPTeamsDirective &S) { |
| 7427 | // Emit teams region as a standalone region. |
| 7428 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 7429 | Action.Enter(CGF); |
| 7430 | OMPPrivateScope PrivateScope(CGF); |
| 7431 | (void)CGF.EmitOMPFirstprivateClause(D: S, PrivateScope); |
| 7432 | CGF.EmitOMPPrivateClause(D: S, PrivateScope); |
| 7433 | CGF.EmitOMPReductionClauseInit(D: S, PrivateScope); |
| 7434 | (void)PrivateScope.Privatize(); |
| 7435 | CGF.EmitStmt(S: S.getCapturedStmt(RegionKind: OMPD_teams)->getCapturedStmt()); |
| 7436 | CGF.EmitOMPReductionClauseFinal(D: S, /*ReductionKind=*/OMPD_teams); |
| 7437 | }; |
| 7438 | emitCommonOMPTeamsDirective(CGF&: *this, S, InnermostKind: OMPD_distribute, CodeGen); |
| 7439 | emitPostUpdateForReductionClause(CGF&: *this, D: S, |
| 7440 | CondGen: [](CodeGenFunction &) { return nullptr; }); |
| 7441 | } |
| 7442 | |
| 7443 | static void emitTargetTeamsRegion(CodeGenFunction &CGF, PrePostActionTy &Action, |
| 7444 | const OMPTargetTeamsDirective &S) { |
| 7445 | auto *CS = S.getCapturedStmt(RegionKind: OMPD_teams); |
| 7446 | Action.Enter(CGF); |
| 7447 | // Emit teams region as a standalone region. |
| 7448 | auto &&CodeGen = [&S, CS](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 7449 | Action.Enter(CGF); |
| 7450 | CodeGenFunction::OMPPrivateScope PrivateScope(CGF); |
| 7451 | (void)CGF.EmitOMPFirstprivateClause(D: S, PrivateScope); |
| 7452 | CGF.EmitOMPPrivateClause(D: S, PrivateScope); |
| 7453 | CGF.EmitOMPReductionClauseInit(D: S, PrivateScope); |
| 7454 | (void)PrivateScope.Privatize(); |
| 7455 | if (isOpenMPTargetExecutionDirective(DKind: S.getDirectiveKind())) |
| 7456 | CGF.CGM.getOpenMPRuntime().adjustTargetSpecificDataForLambdas(CGF, D: S); |
| 7457 | CGF.EmitStmt(S: CS->getCapturedStmt()); |
| 7458 | CGF.EmitOMPReductionClauseFinal(D: S, /*ReductionKind=*/OMPD_teams); |
| 7459 | }; |
| 7460 | emitCommonOMPTeamsDirective(CGF, S, InnermostKind: OMPD_teams, CodeGen); |
| 7461 | emitPostUpdateForReductionClause(CGF, D: S, |
| 7462 | CondGen: [](CodeGenFunction &) { return nullptr; }); |
| 7463 | } |
| 7464 | |
| 7465 | void CodeGenFunction::EmitOMPTargetTeamsDeviceFunction( |
| 7466 | CodeGenModule &CGM, StringRef ParentName, |
| 7467 | const OMPTargetTeamsDirective &S) { |
| 7468 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 7469 | emitTargetTeamsRegion(CGF, Action, S); |
| 7470 | }; |
| 7471 | llvm::Function *Fn; |
| 7472 | llvm::Constant *Addr; |
| 7473 | // Emit target region as a standalone region. |
| 7474 | CGM.getOpenMPRuntime().emitTargetOutlinedFunction( |
| 7475 | D: S, ParentName, OutlinedFn&: Fn, OutlinedFnID&: Addr, /*IsOffloadEntry=*/true, CodeGen); |
| 7476 | assert(Fn && Addr && "Target device function emission failed."); |
| 7477 | } |
| 7478 | |
| 7479 | void CodeGenFunction::EmitOMPTargetTeamsDirective( |
| 7480 | const OMPTargetTeamsDirective &S) { |
| 7481 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 7482 | emitTargetTeamsRegion(CGF, Action, S); |
| 7483 | }; |
| 7484 | emitCommonOMPTargetDirective(CGF&: *this, S, CodeGen); |
| 7485 | } |
| 7486 | |
| 7487 | static void |
| 7488 | emitTargetTeamsDistributeRegion(CodeGenFunction &CGF, PrePostActionTy &Action, |
| 7489 | const OMPTargetTeamsDistributeDirective &S) { |
| 7490 | Action.Enter(CGF); |
| 7491 | auto &&CodeGenDistribute = [&S](CodeGenFunction &CGF, PrePostActionTy &) { |
| 7492 | CGF.EmitOMPDistributeLoop(S, CodeGenLoop: emitOMPLoopBodyWithStopPoint, IncExpr: S.getInc()); |
| 7493 | }; |
| 7494 | |
| 7495 | // Emit teams region as a standalone region. |
| 7496 | auto &&CodeGen = [&S, &CodeGenDistribute](CodeGenFunction &CGF, |
| 7497 | PrePostActionTy &Action) { |
| 7498 | Action.Enter(CGF); |
| 7499 | CodeGenFunction::OMPPrivateScope PrivateScope(CGF); |
| 7500 | CGF.EmitOMPReductionClauseInit(D: S, PrivateScope); |
| 7501 | (void)PrivateScope.Privatize(); |
| 7502 | CGF.CGM.getOpenMPRuntime().emitInlinedDirective(CGF, InnermostKind: OMPD_distribute, |
| 7503 | CodeGen: CodeGenDistribute); |
| 7504 | CGF.EmitOMPReductionClauseFinal(D: S, /*ReductionKind=*/OMPD_teams); |
| 7505 | }; |
| 7506 | emitCommonOMPTeamsDirective(CGF, S, InnermostKind: OMPD_distribute, CodeGen); |
| 7507 | emitPostUpdateForReductionClause(CGF, D: S, |
| 7508 | CondGen: [](CodeGenFunction &) { return nullptr; }); |
| 7509 | } |
| 7510 | |
| 7511 | void CodeGenFunction::EmitOMPTargetTeamsDistributeDeviceFunction( |
| 7512 | CodeGenModule &CGM, StringRef ParentName, |
| 7513 | const OMPTargetTeamsDistributeDirective &S) { |
| 7514 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 7515 | emitTargetTeamsDistributeRegion(CGF, Action, S); |
| 7516 | }; |
| 7517 | llvm::Function *Fn; |
| 7518 | llvm::Constant *Addr; |
| 7519 | // Emit target region as a standalone region. |
| 7520 | CGM.getOpenMPRuntime().emitTargetOutlinedFunction( |
| 7521 | D: S, ParentName, OutlinedFn&: Fn, OutlinedFnID&: Addr, /*IsOffloadEntry=*/true, CodeGen); |
| 7522 | assert(Fn && Addr && "Target device function emission failed."); |
| 7523 | } |
| 7524 | |
| 7525 | void CodeGenFunction::EmitOMPTargetTeamsDistributeDirective( |
| 7526 | const OMPTargetTeamsDistributeDirective &S) { |
| 7527 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 7528 | emitTargetTeamsDistributeRegion(CGF, Action, S); |
| 7529 | }; |
| 7530 | emitCommonOMPTargetDirective(CGF&: *this, S, CodeGen); |
| 7531 | } |
| 7532 | |
| 7533 | static void emitTargetTeamsDistributeSimdRegion( |
| 7534 | CodeGenFunction &CGF, PrePostActionTy &Action, |
| 7535 | const OMPTargetTeamsDistributeSimdDirective &S) { |
| 7536 | Action.Enter(CGF); |
| 7537 | auto &&CodeGenDistribute = [&S](CodeGenFunction &CGF, PrePostActionTy &) { |
| 7538 | CGF.EmitOMPDistributeLoop(S, CodeGenLoop: emitOMPLoopBodyWithStopPoint, IncExpr: S.getInc()); |
| 7539 | }; |
| 7540 | |
| 7541 | // Emit teams region as a standalone region. |
| 7542 | auto &&CodeGen = [&S, &CodeGenDistribute](CodeGenFunction &CGF, |
| 7543 | PrePostActionTy &Action) { |
| 7544 | Action.Enter(CGF); |
| 7545 | CodeGenFunction::OMPPrivateScope PrivateScope(CGF); |
| 7546 | CGF.EmitOMPReductionClauseInit(D: S, PrivateScope); |
| 7547 | (void)PrivateScope.Privatize(); |
| 7548 | CGF.CGM.getOpenMPRuntime().emitInlinedDirective(CGF, InnermostKind: OMPD_distribute, |
| 7549 | CodeGen: CodeGenDistribute); |
| 7550 | CGF.EmitOMPReductionClauseFinal(D: S, /*ReductionKind=*/OMPD_teams); |
| 7551 | }; |
| 7552 | emitCommonOMPTeamsDirective(CGF, S, InnermostKind: OMPD_distribute_simd, CodeGen); |
| 7553 | emitPostUpdateForReductionClause(CGF, D: S, |
| 7554 | CondGen: [](CodeGenFunction &) { return nullptr; }); |
| 7555 | } |
| 7556 | |
| 7557 | void CodeGenFunction::EmitOMPTargetTeamsDistributeSimdDeviceFunction( |
| 7558 | CodeGenModule &CGM, StringRef ParentName, |
| 7559 | const OMPTargetTeamsDistributeSimdDirective &S) { |
| 7560 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 7561 | emitTargetTeamsDistributeSimdRegion(CGF, Action, S); |
| 7562 | }; |
| 7563 | llvm::Function *Fn; |
| 7564 | llvm::Constant *Addr; |
| 7565 | // Emit target region as a standalone region. |
| 7566 | CGM.getOpenMPRuntime().emitTargetOutlinedFunction( |
| 7567 | D: S, ParentName, OutlinedFn&: Fn, OutlinedFnID&: Addr, /*IsOffloadEntry=*/true, CodeGen); |
| 7568 | assert(Fn && Addr && "Target device function emission failed."); |
| 7569 | } |
| 7570 | |
| 7571 | void CodeGenFunction::EmitOMPTargetTeamsDistributeSimdDirective( |
| 7572 | const OMPTargetTeamsDistributeSimdDirective &S) { |
| 7573 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 7574 | emitTargetTeamsDistributeSimdRegion(CGF, Action, S); |
| 7575 | }; |
| 7576 | emitCommonOMPTargetDirective(CGF&: *this, S, CodeGen); |
| 7577 | } |
| 7578 | |
| 7579 | void CodeGenFunction::EmitOMPTeamsDistributeDirective( |
| 7580 | const OMPTeamsDistributeDirective &S) { |
| 7581 | |
| 7582 | auto &&CodeGenDistribute = [&S](CodeGenFunction &CGF, PrePostActionTy &) { |
| 7583 | CGF.EmitOMPDistributeLoop(S, CodeGenLoop: emitOMPLoopBodyWithStopPoint, IncExpr: S.getInc()); |
| 7584 | }; |
| 7585 | |
| 7586 | // Emit teams region as a standalone region. |
| 7587 | auto &&CodeGen = [&S, &CodeGenDistribute](CodeGenFunction &CGF, |
| 7588 | PrePostActionTy &Action) { |
| 7589 | Action.Enter(CGF); |
| 7590 | OMPPrivateScope PrivateScope(CGF); |
| 7591 | CGF.EmitOMPReductionClauseInit(D: S, PrivateScope); |
| 7592 | (void)PrivateScope.Privatize(); |
| 7593 | CGF.CGM.getOpenMPRuntime().emitInlinedDirective(CGF, InnermostKind: OMPD_distribute, |
| 7594 | CodeGen: CodeGenDistribute); |
| 7595 | CGF.EmitOMPReductionClauseFinal(D: S, /*ReductionKind=*/OMPD_teams); |
| 7596 | }; |
| 7597 | emitCommonOMPTeamsDirective(CGF&: *this, S, InnermostKind: OMPD_distribute, CodeGen); |
| 7598 | emitPostUpdateForReductionClause(CGF&: *this, D: S, |
| 7599 | CondGen: [](CodeGenFunction &) { return nullptr; }); |
| 7600 | } |
| 7601 | |
| 7602 | void CodeGenFunction::EmitOMPTeamsDistributeSimdDirective( |
| 7603 | const OMPTeamsDistributeSimdDirective &S) { |
| 7604 | auto &&CodeGenDistribute = [&S](CodeGenFunction &CGF, PrePostActionTy &) { |
| 7605 | CGF.EmitOMPDistributeLoop(S, CodeGenLoop: emitOMPLoopBodyWithStopPoint, IncExpr: S.getInc()); |
| 7606 | }; |
| 7607 | |
| 7608 | // Emit teams region as a standalone region. |
| 7609 | auto &&CodeGen = [&S, &CodeGenDistribute](CodeGenFunction &CGF, |
| 7610 | PrePostActionTy &Action) { |
| 7611 | Action.Enter(CGF); |
| 7612 | OMPPrivateScope PrivateScope(CGF); |
| 7613 | CGF.EmitOMPReductionClauseInit(D: S, PrivateScope); |
| 7614 | (void)PrivateScope.Privatize(); |
| 7615 | CGF.CGM.getOpenMPRuntime().emitInlinedDirective(CGF, InnermostKind: OMPD_simd, |
| 7616 | CodeGen: CodeGenDistribute); |
| 7617 | CGF.EmitOMPReductionClauseFinal(D: S, /*ReductionKind=*/OMPD_teams); |
| 7618 | }; |
| 7619 | emitCommonOMPTeamsDirective(CGF&: *this, S, InnermostKind: OMPD_distribute_simd, CodeGen); |
| 7620 | emitPostUpdateForReductionClause(CGF&: *this, D: S, |
| 7621 | CondGen: [](CodeGenFunction &) { return nullptr; }); |
| 7622 | } |
| 7623 | |
| 7624 | void CodeGenFunction::EmitOMPTeamsDistributeParallelForDirective( |
| 7625 | const OMPTeamsDistributeParallelForDirective &S) { |
| 7626 | auto &&CodeGenDistribute = [&S](CodeGenFunction &CGF, PrePostActionTy &) { |
| 7627 | CGF.EmitOMPDistributeLoop(S, CodeGenLoop: emitInnerParallelForWhenCombined, |
| 7628 | IncExpr: S.getDistInc()); |
| 7629 | }; |
| 7630 | |
| 7631 | // Emit teams region as a standalone region. |
| 7632 | auto &&CodeGen = [&S, &CodeGenDistribute](CodeGenFunction &CGF, |
| 7633 | PrePostActionTy &Action) { |
| 7634 | Action.Enter(CGF); |
| 7635 | OMPPrivateScope PrivateScope(CGF); |
| 7636 | CGF.EmitOMPReductionClauseInit(D: S, PrivateScope); |
| 7637 | (void)PrivateScope.Privatize(); |
| 7638 | CGF.CGM.getOpenMPRuntime().emitInlinedDirective(CGF, InnermostKind: OMPD_distribute, |
| 7639 | CodeGen: CodeGenDistribute); |
| 7640 | CGF.EmitOMPReductionClauseFinal(D: S, /*ReductionKind=*/OMPD_teams); |
| 7641 | }; |
| 7642 | emitCommonOMPTeamsDirective(CGF&: *this, S, InnermostKind: OMPD_distribute_parallel_for, CodeGen); |
| 7643 | emitPostUpdateForReductionClause(CGF&: *this, D: S, |
| 7644 | CondGen: [](CodeGenFunction &) { return nullptr; }); |
| 7645 | } |
| 7646 | |
| 7647 | void CodeGenFunction::EmitOMPTeamsDistributeParallelForSimdDirective( |
| 7648 | const OMPTeamsDistributeParallelForSimdDirective &S) { |
| 7649 | auto &&CodeGenDistribute = [&S](CodeGenFunction &CGF, PrePostActionTy &) { |
| 7650 | CGF.EmitOMPDistributeLoop(S, CodeGenLoop: emitInnerParallelForWhenCombined, |
| 7651 | IncExpr: S.getDistInc()); |
| 7652 | }; |
| 7653 | |
| 7654 | // Emit teams region as a standalone region. |
| 7655 | auto &&CodeGen = [&S, &CodeGenDistribute](CodeGenFunction &CGF, |
| 7656 | PrePostActionTy &Action) { |
| 7657 | Action.Enter(CGF); |
| 7658 | OMPPrivateScope PrivateScope(CGF); |
| 7659 | CGF.EmitOMPReductionClauseInit(D: S, PrivateScope); |
| 7660 | (void)PrivateScope.Privatize(); |
| 7661 | CGF.CGM.getOpenMPRuntime().emitInlinedDirective( |
| 7662 | CGF, InnermostKind: OMPD_distribute, CodeGen: CodeGenDistribute, /*HasCancel=*/false); |
| 7663 | CGF.EmitOMPReductionClauseFinal(D: S, /*ReductionKind=*/OMPD_teams); |
| 7664 | }; |
| 7665 | emitCommonOMPTeamsDirective(CGF&: *this, S, InnermostKind: OMPD_distribute_parallel_for_simd, |
| 7666 | CodeGen); |
| 7667 | emitPostUpdateForReductionClause(CGF&: *this, D: S, |
| 7668 | CondGen: [](CodeGenFunction &) { return nullptr; }); |
| 7669 | } |
| 7670 | |
| 7671 | void CodeGenFunction::EmitOMPInteropDirective(const OMPInteropDirective &S) { |
| 7672 | llvm::OpenMPIRBuilder &OMPBuilder = CGM.getOpenMPRuntime().getOMPBuilder(); |
| 7673 | llvm::Value *Device = nullptr; |
| 7674 | llvm::Value *NumDependences = nullptr; |
| 7675 | llvm::Value *DependenceList = nullptr; |
| 7676 | |
| 7677 | if (const auto *C = S.getSingleClause<OMPDeviceClause>()) |
| 7678 | Device = EmitScalarExpr(E: C->getDevice()); |
| 7679 | |
| 7680 | // Build list and emit dependences |
| 7681 | OMPTaskDataTy Data; |
| 7682 | buildDependences(S, Data); |
| 7683 | if (!Data.Dependences.empty()) { |
| 7684 | Address DependenciesArray = Address::invalid(); |
| 7685 | std::tie(args&: NumDependences, args&: DependenciesArray) = |
| 7686 | CGM.getOpenMPRuntime().emitDependClause(CGF&: *this, Dependencies: Data.Dependences, |
| 7687 | Loc: S.getBeginLoc()); |
| 7688 | DependenceList = DependenciesArray.emitRawPointer(CGF&: *this); |
| 7689 | } |
| 7690 | Data.HasNowaitClause = S.hasClausesOfKind<OMPNowaitClause>(); |
| 7691 | |
| 7692 | assert(!(Data.HasNowaitClause && !(S.getSingleClause<OMPInitClause>() || |
| 7693 | S.getSingleClause<OMPDestroyClause>() || |
| 7694 | S.getSingleClause<OMPUseClause>())) && |
| 7695 | "OMPNowaitClause clause is used separately in OMPInteropDirective."); |
| 7696 | |
| 7697 | auto ItOMPInitClause = S.getClausesOfKind<OMPInitClause>(); |
| 7698 | if (!ItOMPInitClause.empty()) { |
| 7699 | // Look at the multiple init clauses |
| 7700 | for (const OMPInitClause *C : ItOMPInitClause) { |
| 7701 | llvm::Value *InteropvarPtr = |
| 7702 | EmitLValue(E: C->getInteropVar()).getPointer(CGF&: *this); |
| 7703 | llvm::omp::OMPInteropType InteropType = |
| 7704 | llvm::omp::OMPInteropType::Unknown; |
| 7705 | if (C->getIsTarget()) { |
| 7706 | InteropType = llvm::omp::OMPInteropType::Target; |
| 7707 | } else { |
| 7708 | assert(C->getIsTargetSync() && |
| 7709 | "Expected interop-type target/targetsync"); |
| 7710 | InteropType = llvm::omp::OMPInteropType::TargetSync; |
| 7711 | } |
| 7712 | OMPBuilder.createOMPInteropInit(Loc: Builder, InteropVar: InteropvarPtr, InteropType, |
| 7713 | Device, NumDependences, DependenceAddress: DependenceList, |
| 7714 | HaveNowaitClause: Data.HasNowaitClause); |
| 7715 | } |
| 7716 | } |
| 7717 | auto ItOMPDestroyClause = S.getClausesOfKind<OMPDestroyClause>(); |
| 7718 | if (!ItOMPDestroyClause.empty()) { |
| 7719 | // Look at the multiple destroy clauses |
| 7720 | for (const OMPDestroyClause *C : ItOMPDestroyClause) { |
| 7721 | llvm::Value *InteropvarPtr = |
| 7722 | EmitLValue(E: C->getInteropVar()).getPointer(CGF&: *this); |
| 7723 | OMPBuilder.createOMPInteropDestroy(Loc: Builder, InteropVar: InteropvarPtr, Device, |
| 7724 | NumDependences, DependenceAddress: DependenceList, |
| 7725 | HaveNowaitClause: Data.HasNowaitClause); |
| 7726 | } |
| 7727 | } |
| 7728 | auto ItOMPUseClause = S.getClausesOfKind<OMPUseClause>(); |
| 7729 | if (!ItOMPUseClause.empty()) { |
| 7730 | // Look at the multiple use clauses |
| 7731 | for (const OMPUseClause *C : ItOMPUseClause) { |
| 7732 | llvm::Value *InteropvarPtr = |
| 7733 | EmitLValue(E: C->getInteropVar()).getPointer(CGF&: *this); |
| 7734 | OMPBuilder.createOMPInteropUse(Loc: Builder, InteropVar: InteropvarPtr, Device, |
| 7735 | NumDependences, DependenceAddress: DependenceList, |
| 7736 | HaveNowaitClause: Data.HasNowaitClause); |
| 7737 | } |
| 7738 | } |
| 7739 | } |
| 7740 | |
| 7741 | static void emitTargetTeamsDistributeParallelForRegion( |
| 7742 | CodeGenFunction &CGF, const OMPTargetTeamsDistributeParallelForDirective &S, |
| 7743 | PrePostActionTy &Action) { |
| 7744 | Action.Enter(CGF); |
| 7745 | auto &&CodeGenDistribute = [&S](CodeGenFunction &CGF, PrePostActionTy &) { |
| 7746 | CGF.EmitOMPDistributeLoop(S, CodeGenLoop: emitInnerParallelForWhenCombined, |
| 7747 | IncExpr: S.getDistInc()); |
| 7748 | }; |
| 7749 | |
| 7750 | // Emit teams region as a standalone region. |
| 7751 | auto &&CodeGenTeams = [&S, &CodeGenDistribute](CodeGenFunction &CGF, |
| 7752 | PrePostActionTy &Action) { |
| 7753 | Action.Enter(CGF); |
| 7754 | CodeGenFunction::OMPPrivateScope PrivateScope(CGF); |
| 7755 | CGF.EmitOMPReductionClauseInit(D: S, PrivateScope); |
| 7756 | (void)PrivateScope.Privatize(); |
| 7757 | CGF.CGM.getOpenMPRuntime().emitInlinedDirective( |
| 7758 | CGF, InnermostKind: OMPD_distribute, CodeGen: CodeGenDistribute, /*HasCancel=*/false); |
| 7759 | CGF.EmitOMPReductionClauseFinal(D: S, /*ReductionKind=*/OMPD_teams); |
| 7760 | }; |
| 7761 | |
| 7762 | emitCommonOMPTeamsDirective(CGF, S, InnermostKind: OMPD_distribute_parallel_for, |
| 7763 | CodeGen: CodeGenTeams); |
| 7764 | emitPostUpdateForReductionClause(CGF, D: S, |
| 7765 | CondGen: [](CodeGenFunction &) { return nullptr; }); |
| 7766 | } |
| 7767 | |
| 7768 | void CodeGenFunction::EmitOMPTargetTeamsDistributeParallelForDeviceFunction( |
| 7769 | CodeGenModule &CGM, StringRef ParentName, |
| 7770 | const OMPTargetTeamsDistributeParallelForDirective &S) { |
| 7771 | // Emit SPMD target teams distribute parallel for region as a standalone |
| 7772 | // region. |
| 7773 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 7774 | emitTargetTeamsDistributeParallelForRegion(CGF, S, Action); |
| 7775 | }; |
| 7776 | llvm::Function *Fn; |
| 7777 | llvm::Constant *Addr; |
| 7778 | // Emit target region as a standalone region. |
| 7779 | CGM.getOpenMPRuntime().emitTargetOutlinedFunction( |
| 7780 | D: S, ParentName, OutlinedFn&: Fn, OutlinedFnID&: Addr, /*IsOffloadEntry=*/true, CodeGen); |
| 7781 | assert(Fn && Addr && "Target device function emission failed."); |
| 7782 | } |
| 7783 | |
| 7784 | void CodeGenFunction::EmitOMPTargetTeamsDistributeParallelForDirective( |
| 7785 | const OMPTargetTeamsDistributeParallelForDirective &S) { |
| 7786 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 7787 | emitTargetTeamsDistributeParallelForRegion(CGF, S, Action); |
| 7788 | }; |
| 7789 | emitCommonOMPTargetDirective(CGF&: *this, S, CodeGen); |
| 7790 | } |
| 7791 | |
| 7792 | static void emitTargetTeamsDistributeParallelForSimdRegion( |
| 7793 | CodeGenFunction &CGF, |
| 7794 | const OMPTargetTeamsDistributeParallelForSimdDirective &S, |
| 7795 | PrePostActionTy &Action) { |
| 7796 | Action.Enter(CGF); |
| 7797 | auto &&CodeGenDistribute = [&S](CodeGenFunction &CGF, PrePostActionTy &) { |
| 7798 | CGF.EmitOMPDistributeLoop(S, CodeGenLoop: emitInnerParallelForWhenCombined, |
| 7799 | IncExpr: S.getDistInc()); |
| 7800 | }; |
| 7801 | |
| 7802 | // Emit teams region as a standalone region. |
| 7803 | auto &&CodeGenTeams = [&S, &CodeGenDistribute](CodeGenFunction &CGF, |
| 7804 | PrePostActionTy &Action) { |
| 7805 | Action.Enter(CGF); |
| 7806 | CodeGenFunction::OMPPrivateScope PrivateScope(CGF); |
| 7807 | CGF.EmitOMPReductionClauseInit(D: S, PrivateScope); |
| 7808 | (void)PrivateScope.Privatize(); |
| 7809 | CGF.CGM.getOpenMPRuntime().emitInlinedDirective( |
| 7810 | CGF, InnermostKind: OMPD_distribute, CodeGen: CodeGenDistribute, /*HasCancel=*/false); |
| 7811 | CGF.EmitOMPReductionClauseFinal(D: S, /*ReductionKind=*/OMPD_teams); |
| 7812 | }; |
| 7813 | |
| 7814 | emitCommonOMPTeamsDirective(CGF, S, InnermostKind: OMPD_distribute_parallel_for_simd, |
| 7815 | CodeGen: CodeGenTeams); |
| 7816 | emitPostUpdateForReductionClause(CGF, D: S, |
| 7817 | CondGen: [](CodeGenFunction &) { return nullptr; }); |
| 7818 | } |
| 7819 | |
| 7820 | void CodeGenFunction::EmitOMPTargetTeamsDistributeParallelForSimdDeviceFunction( |
| 7821 | CodeGenModule &CGM, StringRef ParentName, |
| 7822 | const OMPTargetTeamsDistributeParallelForSimdDirective &S) { |
| 7823 | // Emit SPMD target teams distribute parallel for simd region as a standalone |
| 7824 | // region. |
| 7825 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 7826 | emitTargetTeamsDistributeParallelForSimdRegion(CGF, S, Action); |
| 7827 | }; |
| 7828 | llvm::Function *Fn; |
| 7829 | llvm::Constant *Addr; |
| 7830 | // Emit target region as a standalone region. |
| 7831 | CGM.getOpenMPRuntime().emitTargetOutlinedFunction( |
| 7832 | D: S, ParentName, OutlinedFn&: Fn, OutlinedFnID&: Addr, /*IsOffloadEntry=*/true, CodeGen); |
| 7833 | assert(Fn && Addr && "Target device function emission failed."); |
| 7834 | } |
| 7835 | |
| 7836 | void CodeGenFunction::EmitOMPTargetTeamsDistributeParallelForSimdDirective( |
| 7837 | const OMPTargetTeamsDistributeParallelForSimdDirective &S) { |
| 7838 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 7839 | emitTargetTeamsDistributeParallelForSimdRegion(CGF, S, Action); |
| 7840 | }; |
| 7841 | emitCommonOMPTargetDirective(CGF&: *this, S, CodeGen); |
| 7842 | } |
| 7843 | |
| 7844 | void CodeGenFunction::EmitOMPCancellationPointDirective( |
| 7845 | const OMPCancellationPointDirective &S) { |
| 7846 | CGM.getOpenMPRuntime().emitCancellationPointCall(CGF&: *this, Loc: S.getBeginLoc(), |
| 7847 | CancelRegion: S.getCancelRegion()); |
| 7848 | } |
| 7849 | |
| 7850 | void CodeGenFunction::EmitOMPCancelDirective(const OMPCancelDirective &S) { |
| 7851 | const Expr *IfCond = nullptr; |
| 7852 | for (const auto *C : S.getClausesOfKind<OMPIfClause>()) { |
| 7853 | if (C->getNameModifier() == OMPD_unknown || |
| 7854 | C->getNameModifier() == OMPD_cancel) { |
| 7855 | IfCond = C->getCondition(); |
| 7856 | break; |
| 7857 | } |
| 7858 | } |
| 7859 | if (CGM.getLangOpts().OpenMPIRBuilder) { |
| 7860 | llvm::OpenMPIRBuilder &OMPBuilder = CGM.getOpenMPRuntime().getOMPBuilder(); |
| 7861 | // TODO: This check is necessary as we only generate `omp parallel` through |
| 7862 | // the OpenMPIRBuilder for now. |
| 7863 | if (S.getCancelRegion() == OMPD_parallel || |
| 7864 | S.getCancelRegion() == OMPD_sections || |
| 7865 | S.getCancelRegion() == OMPD_section) { |
| 7866 | llvm::Value *IfCondition = nullptr; |
| 7867 | if (IfCond) |
| 7868 | IfCondition = EmitScalarExpr(E: IfCond, |
| 7869 | /*IgnoreResultAssign=*/true); |
| 7870 | llvm::OpenMPIRBuilder::InsertPointTy AfterIP = cantFail( |
| 7871 | ValOrErr: OMPBuilder.createCancel(Loc: Builder, IfCondition, CanceledDirective: S.getCancelRegion())); |
| 7872 | return Builder.restoreIP(IP: AfterIP); |
| 7873 | } |
| 7874 | } |
| 7875 | |
| 7876 | CGM.getOpenMPRuntime().emitCancelCall(CGF&: *this, Loc: S.getBeginLoc(), IfCond, |
| 7877 | CancelRegion: S.getCancelRegion()); |
| 7878 | } |
| 7879 | |
| 7880 | CodeGenFunction::JumpDest |
| 7881 | CodeGenFunction::getOMPCancelDestination(OpenMPDirectiveKind Kind) { |
| 7882 | if (Kind == OMPD_parallel || Kind == OMPD_task || |
| 7883 | Kind == OMPD_target_parallel || Kind == OMPD_taskloop || |
| 7884 | Kind == OMPD_master_taskloop || Kind == OMPD_parallel_master_taskloop) |
| 7885 | return ReturnBlock; |
| 7886 | assert(Kind == OMPD_for || Kind == OMPD_section || Kind == OMPD_sections || |
| 7887 | Kind == OMPD_parallel_sections || Kind == OMPD_parallel_for || |
| 7888 | Kind == OMPD_distribute_parallel_for || |
| 7889 | Kind == OMPD_target_parallel_for || |
| 7890 | Kind == OMPD_teams_distribute_parallel_for || |
| 7891 | Kind == OMPD_target_teams_distribute_parallel_for); |
| 7892 | return OMPCancelStack.getExitBlock(); |
| 7893 | } |
| 7894 | |
| 7895 | void CodeGenFunction::EmitOMPUseDevicePtrClause( |
| 7896 | const OMPUseDevicePtrClause &C, OMPPrivateScope &PrivateScope, |
| 7897 | const llvm::DenseMap<const ValueDecl *, llvm::Value *> |
| 7898 | CaptureDeviceAddrMap) { |
| 7899 | llvm::SmallDenseSet<CanonicalDeclPtr<const Decl>, 4> Processed; |
| 7900 | for (const Expr *OrigVarIt : C.varlist()) { |
| 7901 | const auto *OrigVD = cast<VarDecl>(Val: cast<DeclRefExpr>(Val: OrigVarIt)->getDecl()); |
| 7902 | if (!Processed.insert(V: OrigVD).second) |
| 7903 | continue; |
| 7904 | |
| 7905 | // In order to identify the right initializer we need to match the |
| 7906 | // declaration used by the mapping logic. In some cases we may get |
| 7907 | // OMPCapturedExprDecl that refers to the original declaration. |
| 7908 | const ValueDecl *MatchingVD = OrigVD; |
| 7909 | if (const auto *OED = dyn_cast<OMPCapturedExprDecl>(Val: MatchingVD)) { |
| 7910 | // OMPCapturedExprDecl are used to privative fields of the current |
| 7911 | // structure. |
| 7912 | const auto *ME = cast<MemberExpr>(Val: OED->getInit()); |
| 7913 | assert(isa<CXXThisExpr>(ME->getBase()->IgnoreImpCasts()) && |
| 7914 | "Base should be the current struct!"); |
| 7915 | MatchingVD = ME->getMemberDecl(); |
| 7916 | } |
| 7917 | |
| 7918 | // If we don't have information about the current list item, move on to |
| 7919 | // the next one. |
| 7920 | auto InitAddrIt = CaptureDeviceAddrMap.find(Val: MatchingVD); |
| 7921 | if (InitAddrIt == CaptureDeviceAddrMap.end()) |
| 7922 | continue; |
| 7923 | |
| 7924 | llvm::Type *Ty = ConvertTypeForMem(T: OrigVD->getType().getNonReferenceType()); |
| 7925 | |
| 7926 | // Return the address of the private variable. |
| 7927 | bool IsRegistered = PrivateScope.addPrivate( |
| 7928 | LocalVD: OrigVD, |
| 7929 | Addr: Address(InitAddrIt->second, Ty, |
| 7930 | getContext().getTypeAlignInChars(T: getContext().VoidPtrTy))); |
| 7931 | assert(IsRegistered && "firstprivate var already registered as private"); |
| 7932 | // Silence the warning about unused variable. |
| 7933 | (void)IsRegistered; |
| 7934 | } |
| 7935 | } |
| 7936 | |
| 7937 | static const VarDecl *getBaseDecl(const Expr *Ref) { |
| 7938 | const Expr *Base = Ref->IgnoreParenImpCasts(); |
| 7939 | while (const auto *OASE = dyn_cast<ArraySectionExpr>(Val: Base)) |
| 7940 | Base = OASE->getBase()->IgnoreParenImpCasts(); |
| 7941 | while (const auto *ASE = dyn_cast<ArraySubscriptExpr>(Val: Base)) |
| 7942 | Base = ASE->getBase()->IgnoreParenImpCasts(); |
| 7943 | return cast<VarDecl>(Val: cast<DeclRefExpr>(Val: Base)->getDecl()); |
| 7944 | } |
| 7945 | |
| 7946 | void CodeGenFunction::EmitOMPUseDeviceAddrClause( |
| 7947 | const OMPUseDeviceAddrClause &C, OMPPrivateScope &PrivateScope, |
| 7948 | const llvm::DenseMap<const ValueDecl *, llvm::Value *> |
| 7949 | CaptureDeviceAddrMap) { |
| 7950 | llvm::SmallDenseSet<CanonicalDeclPtr<const Decl>, 4> Processed; |
| 7951 | for (const Expr *Ref : C.varlist()) { |
| 7952 | const VarDecl *OrigVD = getBaseDecl(Ref); |
| 7953 | if (!Processed.insert(V: OrigVD).second) |
| 7954 | continue; |
| 7955 | // In order to identify the right initializer we need to match the |
| 7956 | // declaration used by the mapping logic. In some cases we may get |
| 7957 | // OMPCapturedExprDecl that refers to the original declaration. |
| 7958 | const ValueDecl *MatchingVD = OrigVD; |
| 7959 | if (const auto *OED = dyn_cast<OMPCapturedExprDecl>(Val: MatchingVD)) { |
| 7960 | // OMPCapturedExprDecl are used to privative fields of the current |
| 7961 | // structure. |
| 7962 | const auto *ME = cast<MemberExpr>(Val: OED->getInit()); |
| 7963 | assert(isa<CXXThisExpr>(ME->getBase()) && |
| 7964 | "Base should be the current struct!"); |
| 7965 | MatchingVD = ME->getMemberDecl(); |
| 7966 | } |
| 7967 | |
| 7968 | // If we don't have information about the current list item, move on to |
| 7969 | // the next one. |
| 7970 | auto InitAddrIt = CaptureDeviceAddrMap.find(Val: MatchingVD); |
| 7971 | if (InitAddrIt == CaptureDeviceAddrMap.end()) |
| 7972 | continue; |
| 7973 | |
| 7974 | llvm::Type *Ty = ConvertTypeForMem(T: OrigVD->getType().getNonReferenceType()); |
| 7975 | |
| 7976 | Address PrivAddr = |
| 7977 | Address(InitAddrIt->second, Ty, |
| 7978 | getContext().getTypeAlignInChars(T: getContext().VoidPtrTy)); |
| 7979 | // For declrefs and variable length array need to load the pointer for |
| 7980 | // correct mapping, since the pointer to the data was passed to the runtime. |
| 7981 | if (isa<DeclRefExpr>(Val: Ref->IgnoreParenImpCasts()) || |
| 7982 | MatchingVD->getType()->isArrayType()) { |
| 7983 | QualType PtrTy = getContext().getPointerType( |
| 7984 | T: OrigVD->getType().getNonReferenceType()); |
| 7985 | PrivAddr = |
| 7986 | EmitLoadOfPointer(Ptr: PrivAddr.withElementType(ElemTy: ConvertTypeForMem(T: PtrTy)), |
| 7987 | PtrTy: PtrTy->castAs<PointerType>()); |
| 7988 | } |
| 7989 | |
| 7990 | (void)PrivateScope.addPrivate(LocalVD: OrigVD, Addr: PrivAddr); |
| 7991 | } |
| 7992 | } |
| 7993 | |
| 7994 | // Generate the instructions for '#pragma omp target data' directive. |
| 7995 | void CodeGenFunction::EmitOMPTargetDataDirective( |
| 7996 | const OMPTargetDataDirective &S) { |
| 7997 | // Emit vtable only from host for target data directive. |
| 7998 | if (!CGM.getLangOpts().OpenMPIsTargetDevice) |
| 7999 | CGM.getOpenMPRuntime().registerVTable(D: S); |
| 8000 | |
| 8001 | CGOpenMPRuntime::TargetDataInfo Info(/*RequiresDevicePointerInfo=*/true, |
| 8002 | /*SeparateBeginEndCalls=*/true); |
| 8003 | |
| 8004 | // Create a pre/post action to signal the privatization of the device pointer. |
| 8005 | // This action can be replaced by the OpenMP runtime code generation to |
| 8006 | // deactivate privatization. |
| 8007 | bool PrivatizeDevicePointers = false; |
| 8008 | class DevicePointerPrivActionTy : public PrePostActionTy { |
| 8009 | bool &PrivatizeDevicePointers; |
| 8010 | |
| 8011 | public: |
| 8012 | explicit DevicePointerPrivActionTy(bool &PrivatizeDevicePointers) |
| 8013 | : PrivatizeDevicePointers(PrivatizeDevicePointers) {} |
| 8014 | void Enter(CodeGenFunction &CGF) override { |
| 8015 | PrivatizeDevicePointers = true; |
| 8016 | } |
| 8017 | }; |
| 8018 | DevicePointerPrivActionTy PrivAction(PrivatizeDevicePointers); |
| 8019 | |
| 8020 | auto &&CodeGen = [&](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 8021 | auto &&InnermostCodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &) { |
| 8022 | CGF.EmitStmt(S: S.getInnermostCapturedStmt()->getCapturedStmt()); |
| 8023 | }; |
| 8024 | |
| 8025 | // Codegen that selects whether to generate the privatization code or not. |
| 8026 | auto &&PrivCodeGen = [&](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 8027 | RegionCodeGenTy RCG(InnermostCodeGen); |
| 8028 | PrivatizeDevicePointers = false; |
| 8029 | |
| 8030 | // Call the pre-action to change the status of PrivatizeDevicePointers if |
| 8031 | // needed. |
| 8032 | Action.Enter(CGF); |
| 8033 | |
| 8034 | if (PrivatizeDevicePointers) { |
| 8035 | OMPPrivateScope PrivateScope(CGF); |
| 8036 | // Emit all instances of the use_device_ptr clause. |
| 8037 | for (const auto *C : S.getClausesOfKind<OMPUseDevicePtrClause>()) |
| 8038 | CGF.EmitOMPUseDevicePtrClause(C: *C, PrivateScope, |
| 8039 | CaptureDeviceAddrMap: Info.CaptureDeviceAddrMap); |
| 8040 | for (const auto *C : S.getClausesOfKind<OMPUseDeviceAddrClause>()) |
| 8041 | CGF.EmitOMPUseDeviceAddrClause(C: *C, PrivateScope, |
| 8042 | CaptureDeviceAddrMap: Info.CaptureDeviceAddrMap); |
| 8043 | (void)PrivateScope.Privatize(); |
| 8044 | RCG(CGF); |
| 8045 | } else { |
| 8046 | // If we don't have target devices, don't bother emitting the data |
| 8047 | // mapping code. |
| 8048 | std::optional<OpenMPDirectiveKind> CaptureRegion; |
| 8049 | if (CGM.getLangOpts().OMPTargetTriples.empty()) { |
| 8050 | // Emit helper decls of the use_device_ptr/use_device_addr clauses. |
| 8051 | for (const auto *C : S.getClausesOfKind<OMPUseDevicePtrClause>()) |
| 8052 | for (const Expr *E : C->varlist()) { |
| 8053 | const Decl *D = cast<DeclRefExpr>(Val: E)->getDecl(); |
| 8054 | if (const auto *OED = dyn_cast<OMPCapturedExprDecl>(Val: D)) |
| 8055 | CGF.EmitVarDecl(D: *OED); |
| 8056 | } |
| 8057 | for (const auto *C : S.getClausesOfKind<OMPUseDeviceAddrClause>()) |
| 8058 | for (const Expr *E : C->varlist()) { |
| 8059 | const Decl *D = getBaseDecl(Ref: E); |
| 8060 | if (const auto *OED = dyn_cast<OMPCapturedExprDecl>(Val: D)) |
| 8061 | CGF.EmitVarDecl(D: *OED); |
| 8062 | } |
| 8063 | } else { |
| 8064 | CaptureRegion = OMPD_unknown; |
| 8065 | } |
| 8066 | |
| 8067 | OMPLexicalScope Scope(CGF, S, CaptureRegion); |
| 8068 | RCG(CGF); |
| 8069 | } |
| 8070 | }; |
| 8071 | |
| 8072 | // Forward the provided action to the privatization codegen. |
| 8073 | RegionCodeGenTy PrivRCG(PrivCodeGen); |
| 8074 | PrivRCG.setAction(Action); |
| 8075 | |
| 8076 | // Notwithstanding the body of the region is emitted as inlined directive, |
| 8077 | // we don't use an inline scope as changes in the references inside the |
| 8078 | // region are expected to be visible outside, so we do not privative them. |
| 8079 | OMPLexicalScope Scope(CGF, S); |
| 8080 | CGF.CGM.getOpenMPRuntime().emitInlinedDirective(CGF, InnermostKind: OMPD_target_data, |
| 8081 | CodeGen: PrivRCG); |
| 8082 | }; |
| 8083 | |
| 8084 | RegionCodeGenTy RCG(CodeGen); |
| 8085 | |
| 8086 | // If we don't have target devices, don't bother emitting the data mapping |
| 8087 | // code. |
| 8088 | if (CGM.getLangOpts().OMPTargetTriples.empty()) { |
| 8089 | RCG(*this); |
| 8090 | return; |
| 8091 | } |
| 8092 | |
| 8093 | // Check if we have any if clause associated with the directive. |
| 8094 | const Expr *IfCond = nullptr; |
| 8095 | if (const auto *C = S.getSingleClause<OMPIfClause>()) |
| 8096 | IfCond = C->getCondition(); |
| 8097 | |
| 8098 | // Check if we have any device clause associated with the directive. |
| 8099 | const Expr *Device = nullptr; |
| 8100 | if (const auto *C = S.getSingleClause<OMPDeviceClause>()) |
| 8101 | Device = C->getDevice(); |
| 8102 | |
| 8103 | // Set the action to signal privatization of device pointers. |
| 8104 | RCG.setAction(PrivAction); |
| 8105 | |
| 8106 | // Emit region code. |
| 8107 | CGM.getOpenMPRuntime().emitTargetDataCalls(CGF&: *this, D: S, IfCond, Device, CodeGen: RCG, |
| 8108 | Info); |
| 8109 | } |
| 8110 | |
| 8111 | void CodeGenFunction::EmitOMPTargetEnterDataDirective( |
| 8112 | const OMPTargetEnterDataDirective &S) { |
| 8113 | // If we don't have target devices, don't bother emitting the data mapping |
| 8114 | // code. |
| 8115 | if (CGM.getLangOpts().OMPTargetTriples.empty()) |
| 8116 | return; |
| 8117 | |
| 8118 | // Check if we have any if clause associated with the directive. |
| 8119 | const Expr *IfCond = nullptr; |
| 8120 | if (const auto *C = S.getSingleClause<OMPIfClause>()) |
| 8121 | IfCond = C->getCondition(); |
| 8122 | |
| 8123 | // Check if we have any device clause associated with the directive. |
| 8124 | const Expr *Device = nullptr; |
| 8125 | if (const auto *C = S.getSingleClause<OMPDeviceClause>()) |
| 8126 | Device = C->getDevice(); |
| 8127 | |
| 8128 | OMPLexicalScope Scope(*this, S, OMPD_task); |
| 8129 | CGM.getOpenMPRuntime().emitTargetDataStandAloneCall(CGF&: *this, D: S, IfCond, Device); |
| 8130 | } |
| 8131 | |
| 8132 | void CodeGenFunction::EmitOMPTargetExitDataDirective( |
| 8133 | const OMPTargetExitDataDirective &S) { |
| 8134 | // If we don't have target devices, don't bother emitting the data mapping |
| 8135 | // code. |
| 8136 | if (CGM.getLangOpts().OMPTargetTriples.empty()) |
| 8137 | return; |
| 8138 | |
| 8139 | // Check if we have any if clause associated with the directive. |
| 8140 | const Expr *IfCond = nullptr; |
| 8141 | if (const auto *C = S.getSingleClause<OMPIfClause>()) |
| 8142 | IfCond = C->getCondition(); |
| 8143 | |
| 8144 | // Check if we have any device clause associated with the directive. |
| 8145 | const Expr *Device = nullptr; |
| 8146 | if (const auto *C = S.getSingleClause<OMPDeviceClause>()) |
| 8147 | Device = C->getDevice(); |
| 8148 | |
| 8149 | OMPLexicalScope Scope(*this, S, OMPD_task); |
| 8150 | CGM.getOpenMPRuntime().emitTargetDataStandAloneCall(CGF&: *this, D: S, IfCond, Device); |
| 8151 | } |
| 8152 | |
| 8153 | static void emitTargetParallelRegion(CodeGenFunction &CGF, |
| 8154 | const OMPTargetParallelDirective &S, |
| 8155 | PrePostActionTy &Action) { |
| 8156 | // Get the captured statement associated with the 'parallel' region. |
| 8157 | const CapturedStmt *CS = S.getCapturedStmt(RegionKind: OMPD_parallel); |
| 8158 | Action.Enter(CGF); |
| 8159 | auto &&CodeGen = [&S, CS](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 8160 | Action.Enter(CGF); |
| 8161 | CodeGenFunction::OMPPrivateScope PrivateScope(CGF); |
| 8162 | (void)CGF.EmitOMPFirstprivateClause(D: S, PrivateScope); |
| 8163 | CGF.EmitOMPPrivateClause(D: S, PrivateScope); |
| 8164 | CGF.EmitOMPReductionClauseInit(D: S, PrivateScope); |
| 8165 | (void)PrivateScope.Privatize(); |
| 8166 | if (isOpenMPTargetExecutionDirective(DKind: S.getDirectiveKind())) |
| 8167 | CGF.CGM.getOpenMPRuntime().adjustTargetSpecificDataForLambdas(CGF, D: S); |
| 8168 | // TODO: Add support for clauses. |
| 8169 | CGF.EmitStmt(S: CS->getCapturedStmt()); |
| 8170 | CGF.EmitOMPReductionClauseFinal(D: S, /*ReductionKind=*/OMPD_parallel); |
| 8171 | }; |
| 8172 | emitCommonOMPParallelDirective(CGF, S, InnermostKind: OMPD_parallel, CodeGen, |
| 8173 | CodeGenBoundParameters: emitEmptyBoundParameters); |
| 8174 | emitPostUpdateForReductionClause(CGF, D: S, |
| 8175 | CondGen: [](CodeGenFunction &) { return nullptr; }); |
| 8176 | } |
| 8177 | |
| 8178 | void CodeGenFunction::EmitOMPTargetParallelDeviceFunction( |
| 8179 | CodeGenModule &CGM, StringRef ParentName, |
| 8180 | const OMPTargetParallelDirective &S) { |
| 8181 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 8182 | emitTargetParallelRegion(CGF, S, Action); |
| 8183 | }; |
| 8184 | llvm::Function *Fn; |
| 8185 | llvm::Constant *Addr; |
| 8186 | // Emit target region as a standalone region. |
| 8187 | CGM.getOpenMPRuntime().emitTargetOutlinedFunction( |
| 8188 | D: S, ParentName, OutlinedFn&: Fn, OutlinedFnID&: Addr, /*IsOffloadEntry=*/true, CodeGen); |
| 8189 | assert(Fn && Addr && "Target device function emission failed."); |
| 8190 | } |
| 8191 | |
| 8192 | void CodeGenFunction::EmitOMPTargetParallelDirective( |
| 8193 | const OMPTargetParallelDirective &S) { |
| 8194 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 8195 | emitTargetParallelRegion(CGF, S, Action); |
| 8196 | }; |
| 8197 | emitCommonOMPTargetDirective(CGF&: *this, S, CodeGen); |
| 8198 | } |
| 8199 | |
| 8200 | static void emitTargetParallelForRegion(CodeGenFunction &CGF, |
| 8201 | const OMPTargetParallelForDirective &S, |
| 8202 | PrePostActionTy &Action) { |
| 8203 | Action.Enter(CGF); |
| 8204 | // Emit directive as a combined directive that consists of two implicit |
| 8205 | // directives: 'parallel' with 'for' directive. |
| 8206 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 8207 | Action.Enter(CGF); |
| 8208 | CodeGenFunction::OMPCancelStackRAII CancelRegion( |
| 8209 | CGF, OMPD_target_parallel_for, S.hasCancel()); |
| 8210 | CGF.EmitOMPWorksharingLoop(S, EUB: S.getEnsureUpperBound(), CodeGenLoopBounds: emitForLoopBounds, |
| 8211 | CGDispatchBounds: emitDispatchForLoopBounds); |
| 8212 | }; |
| 8213 | emitCommonOMPParallelDirective(CGF, S, InnermostKind: OMPD_for, CodeGen, |
| 8214 | CodeGenBoundParameters: emitEmptyBoundParameters); |
| 8215 | } |
| 8216 | |
| 8217 | void CodeGenFunction::EmitOMPTargetParallelForDeviceFunction( |
| 8218 | CodeGenModule &CGM, StringRef ParentName, |
| 8219 | const OMPTargetParallelForDirective &S) { |
| 8220 | // Emit SPMD target parallel for region as a standalone region. |
| 8221 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 8222 | emitTargetParallelForRegion(CGF, S, Action); |
| 8223 | }; |
| 8224 | llvm::Function *Fn; |
| 8225 | llvm::Constant *Addr; |
| 8226 | // Emit target region as a standalone region. |
| 8227 | CGM.getOpenMPRuntime().emitTargetOutlinedFunction( |
| 8228 | D: S, ParentName, OutlinedFn&: Fn, OutlinedFnID&: Addr, /*IsOffloadEntry=*/true, CodeGen); |
| 8229 | assert(Fn && Addr && "Target device function emission failed."); |
| 8230 | } |
| 8231 | |
| 8232 | void CodeGenFunction::EmitOMPTargetParallelForDirective( |
| 8233 | const OMPTargetParallelForDirective &S) { |
| 8234 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 8235 | emitTargetParallelForRegion(CGF, S, Action); |
| 8236 | }; |
| 8237 | emitCommonOMPTargetDirective(CGF&: *this, S, CodeGen); |
| 8238 | } |
| 8239 | |
| 8240 | static void |
| 8241 | emitTargetParallelForSimdRegion(CodeGenFunction &CGF, |
| 8242 | const OMPTargetParallelForSimdDirective &S, |
| 8243 | PrePostActionTy &Action) { |
| 8244 | Action.Enter(CGF); |
| 8245 | // Emit directive as a combined directive that consists of two implicit |
| 8246 | // directives: 'parallel' with 'for' directive. |
| 8247 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 8248 | Action.Enter(CGF); |
| 8249 | CGF.EmitOMPWorksharingLoop(S, EUB: S.getEnsureUpperBound(), CodeGenLoopBounds: emitForLoopBounds, |
| 8250 | CGDispatchBounds: emitDispatchForLoopBounds); |
| 8251 | }; |
| 8252 | emitCommonOMPParallelDirective(CGF, S, InnermostKind: OMPD_simd, CodeGen, |
| 8253 | CodeGenBoundParameters: emitEmptyBoundParameters); |
| 8254 | } |
| 8255 | |
| 8256 | void CodeGenFunction::EmitOMPTargetParallelForSimdDeviceFunction( |
| 8257 | CodeGenModule &CGM, StringRef ParentName, |
| 8258 | const OMPTargetParallelForSimdDirective &S) { |
| 8259 | // Emit SPMD target parallel for region as a standalone region. |
| 8260 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 8261 | emitTargetParallelForSimdRegion(CGF, S, Action); |
| 8262 | }; |
| 8263 | llvm::Function *Fn; |
| 8264 | llvm::Constant *Addr; |
| 8265 | // Emit target region as a standalone region. |
| 8266 | CGM.getOpenMPRuntime().emitTargetOutlinedFunction( |
| 8267 | D: S, ParentName, OutlinedFn&: Fn, OutlinedFnID&: Addr, /*IsOffloadEntry=*/true, CodeGen); |
| 8268 | assert(Fn && Addr && "Target device function emission failed."); |
| 8269 | } |
| 8270 | |
| 8271 | void CodeGenFunction::EmitOMPTargetParallelForSimdDirective( |
| 8272 | const OMPTargetParallelForSimdDirective &S) { |
| 8273 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 8274 | emitTargetParallelForSimdRegion(CGF, S, Action); |
| 8275 | }; |
| 8276 | emitCommonOMPTargetDirective(CGF&: *this, S, CodeGen); |
| 8277 | } |
| 8278 | |
| 8279 | /// Emit a helper variable and return corresponding lvalue. |
| 8280 | static void mapParam(CodeGenFunction &CGF, const DeclRefExpr *Helper, |
| 8281 | const ImplicitParamDecl *PVD, |
| 8282 | CodeGenFunction::OMPPrivateScope &Privates) { |
| 8283 | const auto *VDecl = cast<VarDecl>(Val: Helper->getDecl()); |
| 8284 | Privates.addPrivate(LocalVD: VDecl, Addr: CGF.GetAddrOfLocalVar(VD: PVD)); |
| 8285 | } |
| 8286 | |
| 8287 | void CodeGenFunction::EmitOMPTaskLoopBasedDirective(const OMPLoopDirective &S) { |
| 8288 | assert(isOpenMPTaskLoopDirective(S.getDirectiveKind())); |
| 8289 | // Emit outlined function for task construct. |
| 8290 | const CapturedStmt *CS = S.getCapturedStmt(RegionKind: OMPD_taskloop); |
| 8291 | Address CapturedStruct = Address::invalid(); |
| 8292 | { |
| 8293 | OMPLexicalScope Scope(*this, S, OMPD_taskloop, /*EmitPreInitStmt=*/false); |
| 8294 | CapturedStruct = GenerateCapturedStmtArgument(S: *CS); |
| 8295 | } |
| 8296 | CanQualType SharedsTy = |
| 8297 | getContext().getCanonicalTagType(TD: CS->getCapturedRecordDecl()); |
| 8298 | const Expr *IfCond = nullptr; |
| 8299 | for (const auto *C : S.getClausesOfKind<OMPIfClause>()) { |
| 8300 | if (C->getNameModifier() == OMPD_unknown || |
| 8301 | C->getNameModifier() == OMPD_taskloop) { |
| 8302 | IfCond = C->getCondition(); |
| 8303 | break; |
| 8304 | } |
| 8305 | } |
| 8306 | |
| 8307 | OMPTaskDataTy Data; |
| 8308 | // Check if taskloop must be emitted without taskgroup. |
| 8309 | Data.Nogroup = S.getSingleClause<OMPNogroupClause>(); |
| 8310 | // TODO: Check if we should emit tied or untied task. |
| 8311 | Data.Tied = true; |
| 8312 | // Set scheduling for taskloop |
| 8313 | if (const auto *Clause = S.getSingleClause<OMPGrainsizeClause>()) { |
| 8314 | // grainsize clause |
| 8315 | Data.Schedule.setInt(/*IntVal=*/false); |
| 8316 | Data.Schedule.setPointer(EmitScalarExpr(E: Clause->getGrainsize())); |
| 8317 | Data.HasModifier = |
| 8318 | (Clause->getModifier() == OMPC_GRAINSIZE_strict) ? true : false; |
| 8319 | } else if (const auto *Clause = S.getSingleClause<OMPNumTasksClause>()) { |
| 8320 | // num_tasks clause |
| 8321 | Data.Schedule.setInt(/*IntVal=*/true); |
| 8322 | Data.Schedule.setPointer(EmitScalarExpr(E: Clause->getNumTasks())); |
| 8323 | Data.HasModifier = |
| 8324 | (Clause->getModifier() == OMPC_NUMTASKS_strict) ? true : false; |
| 8325 | } |
| 8326 | |
| 8327 | auto &&BodyGen = [CS, &S](CodeGenFunction &CGF, PrePostActionTy &) { |
| 8328 | // if (PreCond) { |
| 8329 | // for (IV in 0..LastIteration) BODY; |
| 8330 | // <Final counter/linear vars updates>; |
| 8331 | // } |
| 8332 | // |
| 8333 | |
| 8334 | // Emit: if (PreCond) - begin. |
| 8335 | // If the condition constant folds and can be elided, avoid emitting the |
| 8336 | // whole loop. |
| 8337 | bool CondConstant; |
| 8338 | llvm::BasicBlock *ContBlock = nullptr; |
| 8339 | OMPLoopScope PreInitScope(CGF, S); |
| 8340 | if (CGF.ConstantFoldsToSimpleInteger(Cond: S.getPreCond(), Result&: CondConstant)) { |
| 8341 | if (!CondConstant) |
| 8342 | return; |
| 8343 | } else { |
| 8344 | llvm::BasicBlock *ThenBlock = CGF.createBasicBlock(name: "taskloop.if.then"); |
| 8345 | ContBlock = CGF.createBasicBlock(name: "taskloop.if.end"); |
| 8346 | emitPreCond(CGF, S, Cond: S.getPreCond(), TrueBlock: ThenBlock, FalseBlock: ContBlock, |
| 8347 | TrueCount: CGF.getProfileCount(S: &S)); |
| 8348 | CGF.EmitBlock(BB: ThenBlock); |
| 8349 | CGF.incrementProfileCounter(S: &S); |
| 8350 | } |
| 8351 | |
| 8352 | (void)CGF.EmitOMPLinearClauseInit(D: S); |
| 8353 | |
| 8354 | OMPPrivateScope LoopScope(CGF); |
| 8355 | // Emit helper vars inits. |
| 8356 | enum { LowerBound = 5, UpperBound, Stride, LastIter }; |
| 8357 | auto *I = CS->getCapturedDecl()->param_begin(); |
| 8358 | auto *LBP = std::next(x: I, n: LowerBound); |
| 8359 | auto *UBP = std::next(x: I, n: UpperBound); |
| 8360 | auto *STP = std::next(x: I, n: Stride); |
| 8361 | auto *LIP = std::next(x: I, n: LastIter); |
| 8362 | mapParam(CGF, Helper: cast<DeclRefExpr>(Val: S.getLowerBoundVariable()), PVD: *LBP, |
| 8363 | Privates&: LoopScope); |
| 8364 | mapParam(CGF, Helper: cast<DeclRefExpr>(Val: S.getUpperBoundVariable()), PVD: *UBP, |
| 8365 | Privates&: LoopScope); |
| 8366 | mapParam(CGF, Helper: cast<DeclRefExpr>(Val: S.getStrideVariable()), PVD: *STP, Privates&: LoopScope); |
| 8367 | mapParam(CGF, Helper: cast<DeclRefExpr>(Val: S.getIsLastIterVariable()), PVD: *LIP, |
| 8368 | Privates&: LoopScope); |
| 8369 | CGF.EmitOMPPrivateLoopCounters(S, LoopScope); |
| 8370 | CGF.EmitOMPLinearClause(D: S, PrivateScope&: LoopScope); |
| 8371 | bool HasLastprivateClause = CGF.EmitOMPLastprivateClauseInit(D: S, PrivateScope&: LoopScope); |
| 8372 | (void)LoopScope.Privatize(); |
| 8373 | // Emit the loop iteration variable. |
| 8374 | const Expr *IVExpr = S.getIterationVariable(); |
| 8375 | const auto *IVDecl = cast<VarDecl>(Val: cast<DeclRefExpr>(Val: IVExpr)->getDecl()); |
| 8376 | CGF.EmitVarDecl(D: *IVDecl); |
| 8377 | CGF.EmitIgnoredExpr(E: S.getInit()); |
| 8378 | |
| 8379 | // Emit the iterations count variable. |
| 8380 | // If it is not a variable, Sema decided to calculate iterations count on |
| 8381 | // each iteration (e.g., it is foldable into a constant). |
| 8382 | if (const auto *LIExpr = dyn_cast<DeclRefExpr>(Val: S.getLastIteration())) { |
| 8383 | CGF.EmitVarDecl(D: *cast<VarDecl>(Val: LIExpr->getDecl())); |
| 8384 | // Emit calculation of the iterations count. |
| 8385 | CGF.EmitIgnoredExpr(E: S.getCalcLastIteration()); |
| 8386 | } |
| 8387 | |
| 8388 | { |
| 8389 | OMPLexicalScope Scope(CGF, S, OMPD_taskloop, /*EmitPreInitStmt=*/false); |
| 8390 | emitCommonSimdLoop( |
| 8391 | CGF, S, |
| 8392 | SimdInitGen: [&S](CodeGenFunction &CGF, PrePostActionTy &) { |
| 8393 | if (isOpenMPSimdDirective(DKind: S.getDirectiveKind())) |
| 8394 | CGF.EmitOMPSimdInit(D: S); |
| 8395 | }, |
| 8396 | BodyCodeGen: [&S, &LoopScope](CodeGenFunction &CGF, PrePostActionTy &) { |
| 8397 | CGF.EmitOMPInnerLoop( |
| 8398 | S, RequiresCleanup: LoopScope.requiresCleanups(), LoopCond: S.getCond(), IncExpr: S.getInc(), |
| 8399 | BodyGen: [&S](CodeGenFunction &CGF) { |
| 8400 | emitOMPLoopBodyWithStopPoint(CGF, S, |
| 8401 | LoopExit: CodeGenFunction::JumpDest()); |
| 8402 | }, |
| 8403 | PostIncGen: [](CodeGenFunction &) {}); |
| 8404 | }); |
| 8405 | } |
| 8406 | // Emit: if (PreCond) - end. |
| 8407 | if (ContBlock) { |
| 8408 | CGF.EmitBranch(Block: ContBlock); |
| 8409 | CGF.EmitBlock(BB: ContBlock, IsFinished: true); |
| 8410 | } |
| 8411 | // Emit final copy of the lastprivate variables if IsLastIter != 0. |
| 8412 | if (HasLastprivateClause) { |
| 8413 | CGF.EmitOMPLastprivateClauseFinal( |
| 8414 | D: S, NoFinals: isOpenMPSimdDirective(DKind: S.getDirectiveKind()), |
| 8415 | IsLastIterCond: CGF.Builder.CreateIsNotNull(Arg: CGF.EmitLoadOfScalar( |
| 8416 | Addr: CGF.GetAddrOfLocalVar(VD: *LIP), /*Volatile=*/false, |
| 8417 | Ty: (*LIP)->getType(), Loc: S.getBeginLoc()))); |
| 8418 | } |
| 8419 | LoopScope.restoreMap(); |
| 8420 | CGF.EmitOMPLinearClauseFinal(D: S, CondGen: [LIP, &S](CodeGenFunction &CGF) { |
| 8421 | return CGF.Builder.CreateIsNotNull( |
| 8422 | Arg: CGF.EmitLoadOfScalar(Addr: CGF.GetAddrOfLocalVar(VD: *LIP), /*Volatile=*/false, |
| 8423 | Ty: (*LIP)->getType(), Loc: S.getBeginLoc())); |
| 8424 | }); |
| 8425 | }; |
| 8426 | auto &&TaskGen = [&S, SharedsTy, CapturedStruct, |
| 8427 | IfCond](CodeGenFunction &CGF, llvm::Function *OutlinedFn, |
| 8428 | const OMPTaskDataTy &Data) { |
| 8429 | auto &&CodeGen = [&S, OutlinedFn, SharedsTy, CapturedStruct, IfCond, |
| 8430 | &Data](CodeGenFunction &CGF, PrePostActionTy &) { |
| 8431 | OMPLoopScope PreInitScope(CGF, S); |
| 8432 | CGF.CGM.getOpenMPRuntime().emitTaskLoopCall(CGF, Loc: S.getBeginLoc(), D: S, |
| 8433 | TaskFunction: OutlinedFn, SharedsTy, |
| 8434 | Shareds: CapturedStruct, IfCond, Data); |
| 8435 | }; |
| 8436 | CGF.CGM.getOpenMPRuntime().emitInlinedDirective(CGF, InnermostKind: OMPD_taskloop, |
| 8437 | CodeGen); |
| 8438 | }; |
| 8439 | if (Data.Nogroup) { |
| 8440 | EmitOMPTaskBasedDirective(S, CapturedRegion: OMPD_taskloop, BodyGen, TaskGen, Data); |
| 8441 | } else { |
| 8442 | CGM.getOpenMPRuntime().emitTaskgroupRegion( |
| 8443 | CGF&: *this, |
| 8444 | TaskgroupOpGen: [&S, &BodyGen, &TaskGen, &Data](CodeGenFunction &CGF, |
| 8445 | PrePostActionTy &Action) { |
| 8446 | Action.Enter(CGF); |
| 8447 | CGF.EmitOMPTaskBasedDirective(S, CapturedRegion: OMPD_taskloop, BodyGen, TaskGen, |
| 8448 | Data); |
| 8449 | }, |
| 8450 | Loc: S.getBeginLoc()); |
| 8451 | } |
| 8452 | } |
| 8453 | |
| 8454 | void CodeGenFunction::EmitOMPTaskLoopDirective(const OMPTaskLoopDirective &S) { |
| 8455 | auto LPCRegion = |
| 8456 | CGOpenMPRuntime::LastprivateConditionalRAII::disable(CGF&: *this, S); |
| 8457 | EmitOMPTaskLoopBasedDirective(S); |
| 8458 | } |
| 8459 | |
| 8460 | void CodeGenFunction::EmitOMPTaskLoopSimdDirective( |
| 8461 | const OMPTaskLoopSimdDirective &S) { |
| 8462 | auto LPCRegion = |
| 8463 | CGOpenMPRuntime::LastprivateConditionalRAII::disable(CGF&: *this, S); |
| 8464 | OMPLexicalScope Scope(*this, S); |
| 8465 | EmitOMPTaskLoopBasedDirective(S); |
| 8466 | } |
| 8467 | |
| 8468 | void CodeGenFunction::EmitOMPMasterTaskLoopDirective( |
| 8469 | const OMPMasterTaskLoopDirective &S) { |
| 8470 | auto &&CodeGen = [this, &S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 8471 | Action.Enter(CGF); |
| 8472 | EmitOMPTaskLoopBasedDirective(S); |
| 8473 | }; |
| 8474 | auto LPCRegion = |
| 8475 | CGOpenMPRuntime::LastprivateConditionalRAII::disable(CGF&: *this, S); |
| 8476 | OMPLexicalScope Scope(*this, S, std::nullopt, /*EmitPreInitStmt=*/false); |
| 8477 | CGM.getOpenMPRuntime().emitMasterRegion(CGF&: *this, MasterOpGen: CodeGen, Loc: S.getBeginLoc()); |
| 8478 | } |
| 8479 | |
| 8480 | void CodeGenFunction::EmitOMPMaskedTaskLoopDirective( |
| 8481 | const OMPMaskedTaskLoopDirective &S) { |
| 8482 | auto &&CodeGen = [this, &S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 8483 | Action.Enter(CGF); |
| 8484 | EmitOMPTaskLoopBasedDirective(S); |
| 8485 | }; |
| 8486 | auto LPCRegion = |
| 8487 | CGOpenMPRuntime::LastprivateConditionalRAII::disable(CGF&: *this, S); |
| 8488 | OMPLexicalScope Scope(*this, S, std::nullopt, /*EmitPreInitStmt=*/false); |
| 8489 | CGM.getOpenMPRuntime().emitMaskedRegion(CGF&: *this, MaskedOpGen: CodeGen, Loc: S.getBeginLoc()); |
| 8490 | } |
| 8491 | |
| 8492 | void CodeGenFunction::EmitOMPMasterTaskLoopSimdDirective( |
| 8493 | const OMPMasterTaskLoopSimdDirective &S) { |
| 8494 | auto &&CodeGen = [this, &S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 8495 | Action.Enter(CGF); |
| 8496 | EmitOMPTaskLoopBasedDirective(S); |
| 8497 | }; |
| 8498 | auto LPCRegion = |
| 8499 | CGOpenMPRuntime::LastprivateConditionalRAII::disable(CGF&: *this, S); |
| 8500 | OMPLexicalScope Scope(*this, S); |
| 8501 | CGM.getOpenMPRuntime().emitMasterRegion(CGF&: *this, MasterOpGen: CodeGen, Loc: S.getBeginLoc()); |
| 8502 | } |
| 8503 | |
| 8504 | void CodeGenFunction::EmitOMPMaskedTaskLoopSimdDirective( |
| 8505 | const OMPMaskedTaskLoopSimdDirective &S) { |
| 8506 | auto &&CodeGen = [this, &S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 8507 | Action.Enter(CGF); |
| 8508 | EmitOMPTaskLoopBasedDirective(S); |
| 8509 | }; |
| 8510 | auto LPCRegion = |
| 8511 | CGOpenMPRuntime::LastprivateConditionalRAII::disable(CGF&: *this, S); |
| 8512 | OMPLexicalScope Scope(*this, S); |
| 8513 | CGM.getOpenMPRuntime().emitMaskedRegion(CGF&: *this, MaskedOpGen: CodeGen, Loc: S.getBeginLoc()); |
| 8514 | } |
| 8515 | |
| 8516 | void CodeGenFunction::EmitOMPParallelMasterTaskLoopDirective( |
| 8517 | const OMPParallelMasterTaskLoopDirective &S) { |
| 8518 | auto &&CodeGen = [this, &S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 8519 | auto &&TaskLoopCodeGen = [&S](CodeGenFunction &CGF, |
| 8520 | PrePostActionTy &Action) { |
| 8521 | Action.Enter(CGF); |
| 8522 | CGF.EmitOMPTaskLoopBasedDirective(S); |
| 8523 | }; |
| 8524 | OMPLexicalScope Scope(CGF, S, OMPD_parallel, /*EmitPreInitStmt=*/false); |
| 8525 | CGM.getOpenMPRuntime().emitMasterRegion(CGF, MasterOpGen: TaskLoopCodeGen, |
| 8526 | Loc: S.getBeginLoc()); |
| 8527 | }; |
| 8528 | auto LPCRegion = |
| 8529 | CGOpenMPRuntime::LastprivateConditionalRAII::disable(CGF&: *this, S); |
| 8530 | emitCommonOMPParallelDirective(CGF&: *this, S, InnermostKind: OMPD_master_taskloop, CodeGen, |
| 8531 | CodeGenBoundParameters: emitEmptyBoundParameters); |
| 8532 | } |
| 8533 | |
| 8534 | void CodeGenFunction::EmitOMPParallelMaskedTaskLoopDirective( |
| 8535 | const OMPParallelMaskedTaskLoopDirective &S) { |
| 8536 | auto &&CodeGen = [this, &S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 8537 | auto &&TaskLoopCodeGen = [&S](CodeGenFunction &CGF, |
| 8538 | PrePostActionTy &Action) { |
| 8539 | Action.Enter(CGF); |
| 8540 | CGF.EmitOMPTaskLoopBasedDirective(S); |
| 8541 | }; |
| 8542 | OMPLexicalScope Scope(CGF, S, OMPD_parallel, /*EmitPreInitStmt=*/false); |
| 8543 | CGM.getOpenMPRuntime().emitMaskedRegion(CGF, MaskedOpGen: TaskLoopCodeGen, |
| 8544 | Loc: S.getBeginLoc()); |
| 8545 | }; |
| 8546 | auto LPCRegion = |
| 8547 | CGOpenMPRuntime::LastprivateConditionalRAII::disable(CGF&: *this, S); |
| 8548 | emitCommonOMPParallelDirective(CGF&: *this, S, InnermostKind: OMPD_masked_taskloop, CodeGen, |
| 8549 | CodeGenBoundParameters: emitEmptyBoundParameters); |
| 8550 | } |
| 8551 | |
| 8552 | void CodeGenFunction::EmitOMPParallelMasterTaskLoopSimdDirective( |
| 8553 | const OMPParallelMasterTaskLoopSimdDirective &S) { |
| 8554 | auto &&CodeGen = [this, &S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 8555 | auto &&TaskLoopCodeGen = [&S](CodeGenFunction &CGF, |
| 8556 | PrePostActionTy &Action) { |
| 8557 | Action.Enter(CGF); |
| 8558 | CGF.EmitOMPTaskLoopBasedDirective(S); |
| 8559 | }; |
| 8560 | OMPLexicalScope Scope(CGF, S, OMPD_parallel, /*EmitPreInitStmt=*/false); |
| 8561 | CGM.getOpenMPRuntime().emitMasterRegion(CGF, MasterOpGen: TaskLoopCodeGen, |
| 8562 | Loc: S.getBeginLoc()); |
| 8563 | }; |
| 8564 | auto LPCRegion = |
| 8565 | CGOpenMPRuntime::LastprivateConditionalRAII::disable(CGF&: *this, S); |
| 8566 | emitCommonOMPParallelDirective(CGF&: *this, S, InnermostKind: OMPD_master_taskloop_simd, CodeGen, |
| 8567 | CodeGenBoundParameters: emitEmptyBoundParameters); |
| 8568 | } |
| 8569 | |
| 8570 | void CodeGenFunction::EmitOMPParallelMaskedTaskLoopSimdDirective( |
| 8571 | const OMPParallelMaskedTaskLoopSimdDirective &S) { |
| 8572 | auto &&CodeGen = [this, &S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 8573 | auto &&TaskLoopCodeGen = [&S](CodeGenFunction &CGF, |
| 8574 | PrePostActionTy &Action) { |
| 8575 | Action.Enter(CGF); |
| 8576 | CGF.EmitOMPTaskLoopBasedDirective(S); |
| 8577 | }; |
| 8578 | OMPLexicalScope Scope(CGF, S, OMPD_parallel, /*EmitPreInitStmt=*/false); |
| 8579 | CGM.getOpenMPRuntime().emitMaskedRegion(CGF, MaskedOpGen: TaskLoopCodeGen, |
| 8580 | Loc: S.getBeginLoc()); |
| 8581 | }; |
| 8582 | auto LPCRegion = |
| 8583 | CGOpenMPRuntime::LastprivateConditionalRAII::disable(CGF&: *this, S); |
| 8584 | emitCommonOMPParallelDirective(CGF&: *this, S, InnermostKind: OMPD_masked_taskloop_simd, CodeGen, |
| 8585 | CodeGenBoundParameters: emitEmptyBoundParameters); |
| 8586 | } |
| 8587 | |
| 8588 | // Generate the instructions for '#pragma omp target update' directive. |
| 8589 | void CodeGenFunction::EmitOMPTargetUpdateDirective( |
| 8590 | const OMPTargetUpdateDirective &S) { |
| 8591 | // If we don't have target devices, don't bother emitting the data mapping |
| 8592 | // code. |
| 8593 | if (CGM.getLangOpts().OMPTargetTriples.empty()) |
| 8594 | return; |
| 8595 | |
| 8596 | // Check if we have any if clause associated with the directive. |
| 8597 | const Expr *IfCond = nullptr; |
| 8598 | if (const auto *C = S.getSingleClause<OMPIfClause>()) |
| 8599 | IfCond = C->getCondition(); |
| 8600 | |
| 8601 | // Check if we have any device clause associated with the directive. |
| 8602 | const Expr *Device = nullptr; |
| 8603 | if (const auto *C = S.getSingleClause<OMPDeviceClause>()) |
| 8604 | Device = C->getDevice(); |
| 8605 | |
| 8606 | OMPLexicalScope Scope(*this, S, OMPD_task); |
| 8607 | CGM.getOpenMPRuntime().emitTargetDataStandAloneCall(CGF&: *this, D: S, IfCond, Device); |
| 8608 | } |
| 8609 | |
| 8610 | void CodeGenFunction::EmitOMPGenericLoopDirective( |
| 8611 | const OMPGenericLoopDirective &S) { |
| 8612 | // Always expect a bind clause on the loop directive. It it wasn't |
| 8613 | // in the source, it should have been added in sema. |
| 8614 | |
| 8615 | OpenMPBindClauseKind BindKind = OMPC_BIND_unknown; |
| 8616 | if (const auto *C = S.getSingleClause<OMPBindClause>()) |
| 8617 | BindKind = C->getBindKind(); |
| 8618 | |
| 8619 | switch (BindKind) { |
| 8620 | case OMPC_BIND_parallel: // for |
| 8621 | return emitOMPForDirective(S, CGF&: *this, CGM, /*HasCancel=*/false); |
| 8622 | case OMPC_BIND_teams: // distribute |
| 8623 | return emitOMPDistributeDirective(S, CGF&: *this, CGM); |
| 8624 | case OMPC_BIND_thread: // simd |
| 8625 | return emitOMPSimdDirective(S, CGF&: *this, CGM); |
| 8626 | case OMPC_BIND_unknown: |
| 8627 | break; |
| 8628 | } |
| 8629 | |
| 8630 | // Unimplemented, just inline the underlying statement for now. |
| 8631 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 8632 | // Emit the loop iteration variable. |
| 8633 | const Stmt *CS = |
| 8634 | cast<CapturedStmt>(Val: S.getAssociatedStmt())->getCapturedStmt(); |
| 8635 | const auto *ForS = dyn_cast<ForStmt>(Val: CS); |
| 8636 | if (ForS && !isa<DeclStmt>(Val: ForS->getInit())) { |
| 8637 | OMPPrivateScope LoopScope(CGF); |
| 8638 | CGF.EmitOMPPrivateLoopCounters(S, LoopScope); |
| 8639 | (void)LoopScope.Privatize(); |
| 8640 | CGF.EmitStmt(S: CS); |
| 8641 | LoopScope.restoreMap(); |
| 8642 | } else { |
| 8643 | CGF.EmitStmt(S: CS); |
| 8644 | } |
| 8645 | }; |
| 8646 | OMPLexicalScope Scope(*this, S, OMPD_unknown); |
| 8647 | CGM.getOpenMPRuntime().emitInlinedDirective(CGF&: *this, InnermostKind: OMPD_loop, CodeGen); |
| 8648 | } |
| 8649 | |
| 8650 | void CodeGenFunction::EmitOMPParallelGenericLoopDirective( |
| 8651 | const OMPLoopDirective &S) { |
| 8652 | // Emit combined directive as if its constituent constructs are 'parallel' |
| 8653 | // and 'for'. |
| 8654 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 8655 | Action.Enter(CGF); |
| 8656 | emitOMPCopyinClause(CGF, S); |
| 8657 | (void)emitWorksharingDirective(CGF, S, /*HasCancel=*/false); |
| 8658 | }; |
| 8659 | { |
| 8660 | auto LPCRegion = |
| 8661 | CGOpenMPRuntime::LastprivateConditionalRAII::disable(CGF&: *this, S); |
| 8662 | emitCommonOMPParallelDirective(CGF&: *this, S, InnermostKind: OMPD_for, CodeGen, |
| 8663 | CodeGenBoundParameters: emitEmptyBoundParameters); |
| 8664 | } |
| 8665 | // Check for outer lastprivate conditional update. |
| 8666 | checkForLastprivateConditionalUpdate(CGF&: *this, S); |
| 8667 | } |
| 8668 | |
| 8669 | void CodeGenFunction::EmitOMPTeamsGenericLoopDirective( |
| 8670 | const OMPTeamsGenericLoopDirective &S) { |
| 8671 | // To be consistent with current behavior of 'target teams loop', emit |
| 8672 | // 'teams loop' as if its constituent constructs are 'teams' and 'distribute'. |
| 8673 | auto &&CodeGenDistribute = [&S](CodeGenFunction &CGF, PrePostActionTy &) { |
| 8674 | CGF.EmitOMPDistributeLoop(S, CodeGenLoop: emitOMPLoopBodyWithStopPoint, IncExpr: S.getInc()); |
| 8675 | }; |
| 8676 | |
| 8677 | // Emit teams region as a standalone region. |
| 8678 | auto &&CodeGen = [&S, &CodeGenDistribute](CodeGenFunction &CGF, |
| 8679 | PrePostActionTy &Action) { |
| 8680 | Action.Enter(CGF); |
| 8681 | OMPPrivateScope PrivateScope(CGF); |
| 8682 | CGF.EmitOMPReductionClauseInit(D: S, PrivateScope); |
| 8683 | (void)PrivateScope.Privatize(); |
| 8684 | CGF.CGM.getOpenMPRuntime().emitInlinedDirective(CGF, InnermostKind: OMPD_distribute, |
| 8685 | CodeGen: CodeGenDistribute); |
| 8686 | CGF.EmitOMPReductionClauseFinal(D: S, /*ReductionKind=*/OMPD_teams); |
| 8687 | }; |
| 8688 | emitCommonOMPTeamsDirective(CGF&: *this, S, InnermostKind: OMPD_distribute, CodeGen); |
| 8689 | emitPostUpdateForReductionClause(CGF&: *this, D: S, |
| 8690 | CondGen: [](CodeGenFunction &) { return nullptr; }); |
| 8691 | } |
| 8692 | |
| 8693 | #ifndef NDEBUG |
| 8694 | static void emitTargetTeamsLoopCodegenStatus(CodeGenFunction &CGF, |
| 8695 | std::string StatusMsg, |
| 8696 | const OMPExecutableDirective &D) { |
| 8697 | bool IsDevice = CGF.CGM.getLangOpts().OpenMPIsTargetDevice; |
| 8698 | if (IsDevice) |
| 8699 | StatusMsg += ": DEVICE"; |
| 8700 | else |
| 8701 | StatusMsg += ": HOST"; |
| 8702 | SourceLocation L = D.getBeginLoc(); |
| 8703 | auto &SM = CGF.getContext().getSourceManager(); |
| 8704 | PresumedLoc PLoc = SM.getPresumedLoc(L); |
| 8705 | const char *FileName = PLoc.isValid() ? PLoc.getFilename() : nullptr; |
| 8706 | unsigned LineNo = |
| 8707 | PLoc.isValid() ? PLoc.getLine() : SM.getExpansionLineNumber(L); |
| 8708 | llvm::dbgs() << StatusMsg << ": "<< FileName << ": "<< LineNo << "\n"; |
| 8709 | } |
| 8710 | #endif |
| 8711 | |
| 8712 | static void emitTargetTeamsGenericLoopRegionAsParallel( |
| 8713 | CodeGenFunction &CGF, PrePostActionTy &Action, |
| 8714 | const OMPTargetTeamsGenericLoopDirective &S) { |
| 8715 | Action.Enter(CGF); |
| 8716 | // Emit 'teams loop' as if its constituent constructs are 'distribute, |
| 8717 | // 'parallel, and 'for'. |
| 8718 | auto &&CodeGenDistribute = [&S](CodeGenFunction &CGF, PrePostActionTy &) { |
| 8719 | CGF.EmitOMPDistributeLoop(S, CodeGenLoop: emitInnerParallelForWhenCombined, |
| 8720 | IncExpr: S.getDistInc()); |
| 8721 | }; |
| 8722 | |
| 8723 | // Emit teams region as a standalone region. |
| 8724 | auto &&CodeGenTeams = [&S, &CodeGenDistribute](CodeGenFunction &CGF, |
| 8725 | PrePostActionTy &Action) { |
| 8726 | Action.Enter(CGF); |
| 8727 | CodeGenFunction::OMPPrivateScope PrivateScope(CGF); |
| 8728 | CGF.EmitOMPReductionClauseInit(D: S, PrivateScope); |
| 8729 | (void)PrivateScope.Privatize(); |
| 8730 | CGF.CGM.getOpenMPRuntime().emitInlinedDirective( |
| 8731 | CGF, InnermostKind: OMPD_distribute, CodeGen: CodeGenDistribute, /*HasCancel=*/false); |
| 8732 | CGF.EmitOMPReductionClauseFinal(D: S, /*ReductionKind=*/OMPD_teams); |
| 8733 | }; |
| 8734 | DEBUG_WITH_TYPE(TTL_CODEGEN_TYPE, |
| 8735 | emitTargetTeamsLoopCodegenStatus( |
| 8736 | CGF, TTL_CODEGEN_TYPE " as parallel for", S)); |
| 8737 | emitCommonOMPTeamsDirective(CGF, S, InnermostKind: OMPD_distribute_parallel_for, |
| 8738 | CodeGen: CodeGenTeams); |
| 8739 | emitPostUpdateForReductionClause(CGF, D: S, |
| 8740 | CondGen: [](CodeGenFunction &) { return nullptr; }); |
| 8741 | } |
| 8742 | |
| 8743 | static void emitTargetTeamsGenericLoopRegionAsDistribute( |
| 8744 | CodeGenFunction &CGF, PrePostActionTy &Action, |
| 8745 | const OMPTargetTeamsGenericLoopDirective &S) { |
| 8746 | Action.Enter(CGF); |
| 8747 | // Emit 'teams loop' as if its constituent construct is 'distribute'. |
| 8748 | auto &&CodeGenDistribute = [&S](CodeGenFunction &CGF, PrePostActionTy &) { |
| 8749 | CGF.EmitOMPDistributeLoop(S, CodeGenLoop: emitOMPLoopBodyWithStopPoint, IncExpr: S.getInc()); |
| 8750 | }; |
| 8751 | |
| 8752 | // Emit teams region as a standalone region. |
| 8753 | auto &&CodeGen = [&S, &CodeGenDistribute](CodeGenFunction &CGF, |
| 8754 | PrePostActionTy &Action) { |
| 8755 | Action.Enter(CGF); |
| 8756 | CodeGenFunction::OMPPrivateScope PrivateScope(CGF); |
| 8757 | CGF.EmitOMPReductionClauseInit(D: S, PrivateScope); |
| 8758 | (void)PrivateScope.Privatize(); |
| 8759 | CGF.CGM.getOpenMPRuntime().emitInlinedDirective( |
| 8760 | CGF, InnermostKind: OMPD_distribute, CodeGen: CodeGenDistribute, /*HasCancel=*/false); |
| 8761 | CGF.EmitOMPReductionClauseFinal(D: S, /*ReductionKind=*/OMPD_teams); |
| 8762 | }; |
| 8763 | DEBUG_WITH_TYPE(TTL_CODEGEN_TYPE, |
| 8764 | emitTargetTeamsLoopCodegenStatus( |
| 8765 | CGF, TTL_CODEGEN_TYPE " as distribute", S)); |
| 8766 | emitCommonOMPTeamsDirective(CGF, S, InnermostKind: OMPD_distribute, CodeGen); |
| 8767 | emitPostUpdateForReductionClause(CGF, D: S, |
| 8768 | CondGen: [](CodeGenFunction &) { return nullptr; }); |
| 8769 | } |
| 8770 | |
| 8771 | void CodeGenFunction::EmitOMPTargetTeamsGenericLoopDirective( |
| 8772 | const OMPTargetTeamsGenericLoopDirective &S) { |
| 8773 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 8774 | if (S.canBeParallelFor()) |
| 8775 | emitTargetTeamsGenericLoopRegionAsParallel(CGF, Action, S); |
| 8776 | else |
| 8777 | emitTargetTeamsGenericLoopRegionAsDistribute(CGF, Action, S); |
| 8778 | }; |
| 8779 | emitCommonOMPTargetDirective(CGF&: *this, S, CodeGen); |
| 8780 | } |
| 8781 | |
| 8782 | void CodeGenFunction::EmitOMPTargetTeamsGenericLoopDeviceFunction( |
| 8783 | CodeGenModule &CGM, StringRef ParentName, |
| 8784 | const OMPTargetTeamsGenericLoopDirective &S) { |
| 8785 | // Emit SPMD target parallel loop region as a standalone region. |
| 8786 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 8787 | if (S.canBeParallelFor()) |
| 8788 | emitTargetTeamsGenericLoopRegionAsParallel(CGF, Action, S); |
| 8789 | else |
| 8790 | emitTargetTeamsGenericLoopRegionAsDistribute(CGF, Action, S); |
| 8791 | }; |
| 8792 | llvm::Function *Fn; |
| 8793 | llvm::Constant *Addr; |
| 8794 | // Emit target region as a standalone region. |
| 8795 | CGM.getOpenMPRuntime().emitTargetOutlinedFunction( |
| 8796 | D: S, ParentName, OutlinedFn&: Fn, OutlinedFnID&: Addr, /*IsOffloadEntry=*/true, CodeGen); |
| 8797 | assert(Fn && Addr && |
| 8798 | "Target device function emission failed for 'target teams loop'."); |
| 8799 | } |
| 8800 | |
| 8801 | static void emitTargetParallelGenericLoopRegion( |
| 8802 | CodeGenFunction &CGF, const OMPTargetParallelGenericLoopDirective &S, |
| 8803 | PrePostActionTy &Action) { |
| 8804 | Action.Enter(CGF); |
| 8805 | // Emit as 'parallel for'. |
| 8806 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 8807 | Action.Enter(CGF); |
| 8808 | CodeGenFunction::OMPCancelStackRAII CancelRegion( |
| 8809 | CGF, OMPD_target_parallel_loop, /*hasCancel=*/false); |
| 8810 | CGF.EmitOMPWorksharingLoop(S, EUB: S.getEnsureUpperBound(), CodeGenLoopBounds: emitForLoopBounds, |
| 8811 | CGDispatchBounds: emitDispatchForLoopBounds); |
| 8812 | }; |
| 8813 | emitCommonOMPParallelDirective(CGF, S, InnermostKind: OMPD_for, CodeGen, |
| 8814 | CodeGenBoundParameters: emitEmptyBoundParameters); |
| 8815 | } |
| 8816 | |
| 8817 | void CodeGenFunction::EmitOMPTargetParallelGenericLoopDeviceFunction( |
| 8818 | CodeGenModule &CGM, StringRef ParentName, |
| 8819 | const OMPTargetParallelGenericLoopDirective &S) { |
| 8820 | // Emit target parallel loop region as a standalone region. |
| 8821 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 8822 | emitTargetParallelGenericLoopRegion(CGF, S, Action); |
| 8823 | }; |
| 8824 | llvm::Function *Fn; |
| 8825 | llvm::Constant *Addr; |
| 8826 | // Emit target region as a standalone region. |
| 8827 | CGM.getOpenMPRuntime().emitTargetOutlinedFunction( |
| 8828 | D: S, ParentName, OutlinedFn&: Fn, OutlinedFnID&: Addr, /*IsOffloadEntry=*/true, CodeGen); |
| 8829 | assert(Fn && Addr && "Target device function emission failed."); |
| 8830 | } |
| 8831 | |
| 8832 | /// Emit combined directive 'target parallel loop' as if its constituent |
| 8833 | /// constructs are 'target', 'parallel', and 'for'. |
| 8834 | void CodeGenFunction::EmitOMPTargetParallelGenericLoopDirective( |
| 8835 | const OMPTargetParallelGenericLoopDirective &S) { |
| 8836 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 8837 | emitTargetParallelGenericLoopRegion(CGF, S, Action); |
| 8838 | }; |
| 8839 | emitCommonOMPTargetDirective(CGF&: *this, S, CodeGen); |
| 8840 | } |
| 8841 | |
| 8842 | void CodeGenFunction::EmitSimpleOMPExecutableDirective( |
| 8843 | const OMPExecutableDirective &D) { |
| 8844 | if (const auto *SD = dyn_cast<OMPScanDirective>(Val: &D)) { |
| 8845 | EmitOMPScanDirective(S: *SD); |
| 8846 | return; |
| 8847 | } |
| 8848 | if (!D.hasAssociatedStmt() || !D.getAssociatedStmt()) |
| 8849 | return; |
| 8850 | auto &&CodeGen = [&D](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 8851 | OMPPrivateScope GlobalsScope(CGF); |
| 8852 | if (isOpenMPTaskingDirective(Kind: D.getDirectiveKind())) { |
| 8853 | // Capture global firstprivates to avoid crash. |
| 8854 | for (const auto *C : D.getClausesOfKind<OMPFirstprivateClause>()) { |
| 8855 | for (const Expr *Ref : C->varlist()) { |
| 8856 | const auto *DRE = cast<DeclRefExpr>(Val: Ref->IgnoreParenImpCasts()); |
| 8857 | if (!DRE) |
| 8858 | continue; |
| 8859 | const auto *VD = dyn_cast<VarDecl>(Val: DRE->getDecl()); |
| 8860 | if (!VD || VD->hasLocalStorage()) |
| 8861 | continue; |
| 8862 | if (!CGF.LocalDeclMap.count(Val: VD)) { |
| 8863 | LValue GlobLVal = CGF.EmitLValue(E: Ref); |
| 8864 | GlobalsScope.addPrivate(LocalVD: VD, Addr: GlobLVal.getAddress()); |
| 8865 | } |
| 8866 | } |
| 8867 | } |
| 8868 | } |
| 8869 | if (isOpenMPSimdDirective(DKind: D.getDirectiveKind())) { |
| 8870 | (void)GlobalsScope.Privatize(); |
| 8871 | ParentLoopDirectiveForScanRegion ScanRegion(CGF, D); |
| 8872 | emitOMPSimdRegion(CGF, S: cast<OMPLoopDirective>(Val: D), Action); |
| 8873 | } else { |
| 8874 | if (const auto *LD = dyn_cast<OMPLoopDirective>(Val: &D)) { |
| 8875 | for (const Expr *E : LD->counters()) { |
| 8876 | const auto *VD = cast<VarDecl>(Val: cast<DeclRefExpr>(Val: E)->getDecl()); |
| 8877 | if (!VD->hasLocalStorage() && !CGF.LocalDeclMap.count(Val: VD)) { |
| 8878 | LValue GlobLVal = CGF.EmitLValue(E); |
| 8879 | GlobalsScope.addPrivate(LocalVD: VD, Addr: GlobLVal.getAddress()); |
| 8880 | } |
| 8881 | if (isa<OMPCapturedExprDecl>(Val: VD)) { |
| 8882 | // Emit only those that were not explicitly referenced in clauses. |
| 8883 | if (!CGF.LocalDeclMap.count(Val: VD)) |
| 8884 | CGF.EmitVarDecl(D: *VD); |
| 8885 | } |
| 8886 | } |
| 8887 | for (const auto *C : D.getClausesOfKind<OMPOrderedClause>()) { |
| 8888 | if (!C->getNumForLoops()) |
| 8889 | continue; |
| 8890 | for (unsigned I = LD->getLoopsNumber(), |
| 8891 | E = C->getLoopNumIterations().size(); |
| 8892 | I < E; ++I) { |
| 8893 | if (const auto *VD = dyn_cast<OMPCapturedExprDecl>( |
| 8894 | Val: cast<DeclRefExpr>(Val: C->getLoopCounter(NumLoop: I))->getDecl())) { |
| 8895 | // Emit only those that were not explicitly referenced in clauses. |
| 8896 | if (!CGF.LocalDeclMap.count(Val: VD)) |
| 8897 | CGF.EmitVarDecl(D: *VD); |
| 8898 | } |
| 8899 | } |
| 8900 | } |
| 8901 | } |
| 8902 | (void)GlobalsScope.Privatize(); |
| 8903 | CGF.EmitStmt(S: D.getInnermostCapturedStmt()->getCapturedStmt()); |
| 8904 | } |
| 8905 | }; |
| 8906 | if (D.getDirectiveKind() == OMPD_atomic || |
| 8907 | D.getDirectiveKind() == OMPD_critical || |
| 8908 | D.getDirectiveKind() == OMPD_section || |
| 8909 | D.getDirectiveKind() == OMPD_master || |
| 8910 | D.getDirectiveKind() == OMPD_masked || |
| 8911 | D.getDirectiveKind() == OMPD_unroll || |
| 8912 | D.getDirectiveKind() == OMPD_assume) { |
| 8913 | EmitStmt(S: D.getAssociatedStmt()); |
| 8914 | } else { |
| 8915 | auto LPCRegion = |
| 8916 | CGOpenMPRuntime::LastprivateConditionalRAII::disable(CGF&: *this, S: D); |
| 8917 | OMPSimdLexicalScope Scope(*this, D); |
| 8918 | CGM.getOpenMPRuntime().emitInlinedDirective( |
| 8919 | CGF&: *this, |
| 8920 | InnermostKind: isOpenMPSimdDirective(DKind: D.getDirectiveKind()) ? OMPD_simd |
| 8921 | : D.getDirectiveKind(), |
| 8922 | CodeGen); |
| 8923 | } |
| 8924 | // Check for outer lastprivate conditional update. |
| 8925 | checkForLastprivateConditionalUpdate(CGF&: *this, S: D); |
| 8926 | } |
| 8927 | |
| 8928 | void CodeGenFunction::EmitOMPAssumeDirective(const OMPAssumeDirective &S) { |
| 8929 | EmitStmt(S: S.getAssociatedStmt()); |
| 8930 | } |
| 8931 |
Generated on 2026-Aug-30 from project llvm_projects revision 868246f06
Powered by 
Code Browser 2.1
Generator usage only permitted with license.