| File: | build/llvm-toolchain-snapshot-15~++20220420111733+e13d2efed663/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp |
| Warning: | line 1448, column 5 Value stored to 'Size' is never read |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
| 1 | //===---- CGOpenMPRuntimeGPU.cpp - Interface to OpenMP GPU Runtimes ----===// |
| 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 provides a generalized class for OpenMP runtime code generation |
| 10 | // specialized by GPU targets NVPTX and AMDGCN. |
| 11 | // |
| 12 | //===----------------------------------------------------------------------===// |
| 13 | |
| 14 | #include "CGOpenMPRuntimeGPU.h" |
| 15 | #include "CodeGenFunction.h" |
| 16 | #include "clang/AST/Attr.h" |
| 17 | #include "clang/AST/DeclOpenMP.h" |
| 18 | #include "clang/AST/StmtOpenMP.h" |
| 19 | #include "clang/AST/StmtVisitor.h" |
| 20 | #include "clang/Basic/Cuda.h" |
| 21 | #include "llvm/ADT/SmallPtrSet.h" |
| 22 | #include "llvm/Frontend/OpenMP/OMPGridValues.h" |
| 23 | #include "llvm/Support/MathExtras.h" |
| 24 | |
| 25 | using namespace clang; |
| 26 | using namespace CodeGen; |
| 27 | using namespace llvm::omp; |
| 28 | |
| 29 | namespace { |
| 30 | /// Pre(post)-action for different OpenMP constructs specialized for NVPTX. |
| 31 | class NVPTXActionTy final : public PrePostActionTy { |
| 32 | llvm::FunctionCallee EnterCallee = nullptr; |
| 33 | ArrayRef<llvm::Value *> EnterArgs; |
| 34 | llvm::FunctionCallee ExitCallee = nullptr; |
| 35 | ArrayRef<llvm::Value *> ExitArgs; |
| 36 | bool Conditional = false; |
| 37 | llvm::BasicBlock *ContBlock = nullptr; |
| 38 | |
| 39 | public: |
| 40 | NVPTXActionTy(llvm::FunctionCallee EnterCallee, |
| 41 | ArrayRef<llvm::Value *> EnterArgs, |
| 42 | llvm::FunctionCallee ExitCallee, |
| 43 | ArrayRef<llvm::Value *> ExitArgs, bool Conditional = false) |
| 44 | : EnterCallee(EnterCallee), EnterArgs(EnterArgs), ExitCallee(ExitCallee), |
| 45 | ExitArgs(ExitArgs), Conditional(Conditional) {} |
| 46 | void Enter(CodeGenFunction &CGF) override { |
| 47 | llvm::Value *EnterRes = CGF.EmitRuntimeCall(EnterCallee, EnterArgs); |
| 48 | if (Conditional) { |
| 49 | llvm::Value *CallBool = CGF.Builder.CreateIsNotNull(EnterRes); |
| 50 | auto *ThenBlock = CGF.createBasicBlock("omp_if.then"); |
| 51 | ContBlock = CGF.createBasicBlock("omp_if.end"); |
| 52 | // Generate the branch (If-stmt) |
| 53 | CGF.Builder.CreateCondBr(CallBool, ThenBlock, ContBlock); |
| 54 | CGF.EmitBlock(ThenBlock); |
| 55 | } |
| 56 | } |
| 57 | void Done(CodeGenFunction &CGF) { |
| 58 | // Emit the rest of blocks/branches |
| 59 | CGF.EmitBranch(ContBlock); |
| 60 | CGF.EmitBlock(ContBlock, true); |
| 61 | } |
| 62 | void Exit(CodeGenFunction &CGF) override { |
| 63 | CGF.EmitRuntimeCall(ExitCallee, ExitArgs); |
| 64 | } |
| 65 | }; |
| 66 | |
| 67 | /// A class to track the execution mode when codegening directives within |
| 68 | /// a target region. The appropriate mode (SPMD|NON-SPMD) is set on entry |
| 69 | /// to the target region and used by containing directives such as 'parallel' |
| 70 | /// to emit optimized code. |
| 71 | class ExecutionRuntimeModesRAII { |
| 72 | private: |
| 73 | CGOpenMPRuntimeGPU::ExecutionMode SavedExecMode = |
| 74 | CGOpenMPRuntimeGPU::EM_Unknown; |
| 75 | CGOpenMPRuntimeGPU::ExecutionMode &ExecMode; |
| 76 | bool SavedRuntimeMode = false; |
| 77 | bool *RuntimeMode = nullptr; |
| 78 | |
| 79 | public: |
| 80 | /// Constructor for Non-SPMD mode. |
| 81 | ExecutionRuntimeModesRAII(CGOpenMPRuntimeGPU::ExecutionMode &ExecMode) |
| 82 | : ExecMode(ExecMode) { |
| 83 | SavedExecMode = ExecMode; |
| 84 | ExecMode = CGOpenMPRuntimeGPU::EM_NonSPMD; |
| 85 | } |
| 86 | /// Constructor for SPMD mode. |
| 87 | ExecutionRuntimeModesRAII(CGOpenMPRuntimeGPU::ExecutionMode &ExecMode, |
| 88 | bool &RuntimeMode, bool FullRuntimeMode) |
| 89 | : ExecMode(ExecMode), RuntimeMode(&RuntimeMode) { |
| 90 | SavedExecMode = ExecMode; |
| 91 | SavedRuntimeMode = RuntimeMode; |
| 92 | ExecMode = CGOpenMPRuntimeGPU::EM_SPMD; |
| 93 | RuntimeMode = FullRuntimeMode; |
| 94 | } |
| 95 | ~ExecutionRuntimeModesRAII() { |
| 96 | ExecMode = SavedExecMode; |
| 97 | if (RuntimeMode) |
| 98 | *RuntimeMode = SavedRuntimeMode; |
| 99 | } |
| 100 | }; |
| 101 | |
| 102 | /// GPU Configuration: This information can be derived from cuda registers, |
| 103 | /// however, providing compile time constants helps generate more efficient |
| 104 | /// code. For all practical purposes this is fine because the configuration |
| 105 | /// is the same for all known NVPTX architectures. |
| 106 | enum MachineConfiguration : unsigned { |
| 107 | /// See "llvm/Frontend/OpenMP/OMPGridValues.h" for various related target |
| 108 | /// specific Grid Values like GV_Warp_Size, GV_Slot_Size |
| 109 | |
| 110 | /// Global memory alignment for performance. |
| 111 | GlobalMemoryAlignment = 128, |
| 112 | |
| 113 | /// Maximal size of the shared memory buffer. |
| 114 | SharedMemorySize = 128, |
| 115 | }; |
| 116 | |
| 117 | static const ValueDecl *getPrivateItem(const Expr *RefExpr) { |
| 118 | RefExpr = RefExpr->IgnoreParens(); |
| 119 | if (const auto *ASE = dyn_cast<ArraySubscriptExpr>(RefExpr)) { |
| 120 | const Expr *Base = ASE->getBase()->IgnoreParenImpCasts(); |
| 121 | while (const auto *TempASE = dyn_cast<ArraySubscriptExpr>(Base)) |
| 122 | Base = TempASE->getBase()->IgnoreParenImpCasts(); |
| 123 | RefExpr = Base; |
| 124 | } else if (auto *OASE = dyn_cast<OMPArraySectionExpr>(RefExpr)) { |
| 125 | const Expr *Base = OASE->getBase()->IgnoreParenImpCasts(); |
| 126 | while (const auto *TempOASE = dyn_cast<OMPArraySectionExpr>(Base)) |
| 127 | Base = TempOASE->getBase()->IgnoreParenImpCasts(); |
| 128 | while (const auto *TempASE = dyn_cast<ArraySubscriptExpr>(Base)) |
| 129 | Base = TempASE->getBase()->IgnoreParenImpCasts(); |
| 130 | RefExpr = Base; |
| 131 | } |
| 132 | RefExpr = RefExpr->IgnoreParenImpCasts(); |
| 133 | if (const auto *DE = dyn_cast<DeclRefExpr>(RefExpr)) |
| 134 | return cast<ValueDecl>(DE->getDecl()->getCanonicalDecl()); |
| 135 | const auto *ME = cast<MemberExpr>(RefExpr); |
| 136 | return cast<ValueDecl>(ME->getMemberDecl()->getCanonicalDecl()); |
| 137 | } |
| 138 | |
| 139 | |
| 140 | static RecordDecl *buildRecordForGlobalizedVars( |
| 141 | ASTContext &C, ArrayRef<const ValueDecl *> EscapedDecls, |
| 142 | ArrayRef<const ValueDecl *> EscapedDeclsForTeams, |
| 143 | llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> |
| 144 | &MappedDeclsFields, int BufSize) { |
| 145 | using VarsDataTy = std::pair<CharUnits /*Align*/, const ValueDecl *>; |
| 146 | if (EscapedDecls.empty() && EscapedDeclsForTeams.empty()) |
| 147 | return nullptr; |
| 148 | SmallVector<VarsDataTy, 4> GlobalizedVars; |
| 149 | for (const ValueDecl *D : EscapedDecls) |
| 150 | GlobalizedVars.emplace_back( |
| 151 | CharUnits::fromQuantity(std::max( |
| 152 | C.getDeclAlign(D).getQuantity(), |
| 153 | static_cast<CharUnits::QuantityType>(GlobalMemoryAlignment))), |
| 154 | D); |
| 155 | for (const ValueDecl *D : EscapedDeclsForTeams) |
| 156 | GlobalizedVars.emplace_back(C.getDeclAlign(D), D); |
| 157 | llvm::stable_sort(GlobalizedVars, [](VarsDataTy L, VarsDataTy R) { |
| 158 | return L.first > R.first; |
| 159 | }); |
| 160 | |
| 161 | // Build struct _globalized_locals_ty { |
| 162 | // /* globalized vars */[WarSize] align (max(decl_align, |
| 163 | // GlobalMemoryAlignment)) |
| 164 | // /* globalized vars */ for EscapedDeclsForTeams |
| 165 | // }; |
| 166 | RecordDecl *GlobalizedRD = C.buildImplicitRecord("_globalized_locals_ty"); |
| 167 | GlobalizedRD->startDefinition(); |
| 168 | llvm::SmallPtrSet<const ValueDecl *, 16> SingleEscaped( |
| 169 | EscapedDeclsForTeams.begin(), EscapedDeclsForTeams.end()); |
| 170 | for (const auto &Pair : GlobalizedVars) { |
| 171 | const ValueDecl *VD = Pair.second; |
| 172 | QualType Type = VD->getType(); |
| 173 | if (Type->isLValueReferenceType()) |
| 174 | Type = C.getPointerType(Type.getNonReferenceType()); |
| 175 | else |
| 176 | Type = Type.getNonReferenceType(); |
| 177 | SourceLocation Loc = VD->getLocation(); |
| 178 | FieldDecl *Field; |
| 179 | if (SingleEscaped.count(VD)) { |
| 180 | Field = FieldDecl::Create( |
| 181 | C, GlobalizedRD, Loc, Loc, VD->getIdentifier(), Type, |
| 182 | C.getTrivialTypeSourceInfo(Type, SourceLocation()), |
| 183 | /*BW=*/nullptr, /*Mutable=*/false, |
| 184 | /*InitStyle=*/ICIS_NoInit); |
| 185 | Field->setAccess(AS_public); |
| 186 | if (VD->hasAttrs()) { |
| 187 | for (specific_attr_iterator<AlignedAttr> I(VD->getAttrs().begin()), |
| 188 | E(VD->getAttrs().end()); |
| 189 | I != E; ++I) |
| 190 | Field->addAttr(*I); |
| 191 | } |
| 192 | } else { |
| 193 | llvm::APInt ArraySize(32, BufSize); |
| 194 | Type = C.getConstantArrayType(Type, ArraySize, nullptr, ArrayType::Normal, |
| 195 | 0); |
| 196 | Field = FieldDecl::Create( |
| 197 | C, GlobalizedRD, Loc, Loc, VD->getIdentifier(), Type, |
| 198 | C.getTrivialTypeSourceInfo(Type, SourceLocation()), |
| 199 | /*BW=*/nullptr, /*Mutable=*/false, |
| 200 | /*InitStyle=*/ICIS_NoInit); |
| 201 | Field->setAccess(AS_public); |
| 202 | llvm::APInt Align(32, std::max(C.getDeclAlign(VD).getQuantity(), |
| 203 | static_cast<CharUnits::QuantityType>( |
| 204 | GlobalMemoryAlignment))); |
| 205 | Field->addAttr(AlignedAttr::CreateImplicit( |
| 206 | C, /*IsAlignmentExpr=*/true, |
| 207 | IntegerLiteral::Create(C, Align, |
| 208 | C.getIntTypeForBitwidth(32, /*Signed=*/0), |
| 209 | SourceLocation()), |
| 210 | {}, AttributeCommonInfo::AS_GNU, AlignedAttr::GNU_aligned)); |
| 211 | } |
| 212 | GlobalizedRD->addDecl(Field); |
| 213 | MappedDeclsFields.try_emplace(VD, Field); |
| 214 | } |
| 215 | GlobalizedRD->completeDefinition(); |
| 216 | return GlobalizedRD; |
| 217 | } |
| 218 | |
| 219 | /// Get the list of variables that can escape their declaration context. |
| 220 | class CheckVarsEscapingDeclContext final |
| 221 | : public ConstStmtVisitor<CheckVarsEscapingDeclContext> { |
| 222 | CodeGenFunction &CGF; |
| 223 | llvm::SetVector<const ValueDecl *> EscapedDecls; |
| 224 | llvm::SetVector<const ValueDecl *> EscapedVariableLengthDecls; |
| 225 | llvm::SmallPtrSet<const Decl *, 4> EscapedParameters; |
| 226 | RecordDecl *GlobalizedRD = nullptr; |
| 227 | llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> MappedDeclsFields; |
| 228 | bool AllEscaped = false; |
| 229 | bool IsForCombinedParallelRegion = false; |
| 230 | |
| 231 | void markAsEscaped(const ValueDecl *VD) { |
| 232 | // Do not globalize declare target variables. |
| 233 | if (!isa<VarDecl>(VD) || |
| 234 | OMPDeclareTargetDeclAttr::isDeclareTargetDeclaration(VD)) |
| 235 | return; |
| 236 | VD = cast<ValueDecl>(VD->getCanonicalDecl()); |
| 237 | // Use user-specified allocation. |
| 238 | if (VD->hasAttrs() && VD->hasAttr<OMPAllocateDeclAttr>()) |
| 239 | return; |
| 240 | // Variables captured by value must be globalized. |
| 241 | if (auto *CSI = CGF.CapturedStmtInfo) { |
| 242 | if (const FieldDecl *FD = CSI->lookup(cast<VarDecl>(VD))) { |
| 243 | // Check if need to capture the variable that was already captured by |
| 244 | // value in the outer region. |
| 245 | if (!IsForCombinedParallelRegion) { |
| 246 | if (!FD->hasAttrs()) |
| 247 | return; |
| 248 | const auto *Attr = FD->getAttr<OMPCaptureKindAttr>(); |
| 249 | if (!Attr) |
| 250 | return; |
| 251 | if (((Attr->getCaptureKind() != OMPC_map) && |
| 252 | !isOpenMPPrivate(Attr->getCaptureKind())) || |
| 253 | ((Attr->getCaptureKind() == OMPC_map) && |
| 254 | !FD->getType()->isAnyPointerType())) |
| 255 | return; |
| 256 | } |
| 257 | if (!FD->getType()->isReferenceType()) { |
| 258 | assert(!VD->getType()->isVariablyModifiedType() &&(static_cast <bool> (!VD->getType()->isVariablyModifiedType () && "Parameter captured by value with variably modified type" ) ? void (0) : __assert_fail ("!VD->getType()->isVariablyModifiedType() && \"Parameter captured by value with variably modified type\"" , "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 259, __extension__ __PRETTY_FUNCTION__)) |
| 259 | "Parameter captured by value with variably modified type")(static_cast <bool> (!VD->getType()->isVariablyModifiedType () && "Parameter captured by value with variably modified type" ) ? void (0) : __assert_fail ("!VD->getType()->isVariablyModifiedType() && \"Parameter captured by value with variably modified type\"" , "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 259, __extension__ __PRETTY_FUNCTION__)); |
| 260 | EscapedParameters.insert(VD); |
| 261 | } else if (!IsForCombinedParallelRegion) { |
| 262 | return; |
| 263 | } |
| 264 | } |
| 265 | } |
| 266 | if ((!CGF.CapturedStmtInfo || |
| 267 | (IsForCombinedParallelRegion && CGF.CapturedStmtInfo)) && |
| 268 | VD->getType()->isReferenceType()) |
| 269 | // Do not globalize variables with reference type. |
| 270 | return; |
| 271 | if (VD->getType()->isVariablyModifiedType()) |
| 272 | EscapedVariableLengthDecls.insert(VD); |
| 273 | else |
| 274 | EscapedDecls.insert(VD); |
| 275 | } |
| 276 | |
| 277 | void VisitValueDecl(const ValueDecl *VD) { |
| 278 | if (VD->getType()->isLValueReferenceType()) |
| 279 | markAsEscaped(VD); |
| 280 | if (const auto *VarD = dyn_cast<VarDecl>(VD)) { |
| 281 | if (!isa<ParmVarDecl>(VarD) && VarD->hasInit()) { |
| 282 | const bool SavedAllEscaped = AllEscaped; |
| 283 | AllEscaped = VD->getType()->isLValueReferenceType(); |
| 284 | Visit(VarD->getInit()); |
| 285 | AllEscaped = SavedAllEscaped; |
| 286 | } |
| 287 | } |
| 288 | } |
| 289 | void VisitOpenMPCapturedStmt(const CapturedStmt *S, |
| 290 | ArrayRef<OMPClause *> Clauses, |
| 291 | bool IsCombinedParallelRegion) { |
| 292 | if (!S) |
| 293 | return; |
| 294 | for (const CapturedStmt::Capture &C : S->captures()) { |
| 295 | if (C.capturesVariable() && !C.capturesVariableByCopy()) { |
| 296 | const ValueDecl *VD = C.getCapturedVar(); |
| 297 | bool SavedIsForCombinedParallelRegion = IsForCombinedParallelRegion; |
| 298 | if (IsCombinedParallelRegion) { |
| 299 | // Check if the variable is privatized in the combined construct and |
| 300 | // those private copies must be shared in the inner parallel |
| 301 | // directive. |
| 302 | IsForCombinedParallelRegion = false; |
| 303 | for (const OMPClause *C : Clauses) { |
| 304 | if (!isOpenMPPrivate(C->getClauseKind()) || |
| 305 | C->getClauseKind() == OMPC_reduction || |
| 306 | C->getClauseKind() == OMPC_linear || |
| 307 | C->getClauseKind() == OMPC_private) |
| 308 | continue; |
| 309 | ArrayRef<const Expr *> Vars; |
| 310 | if (const auto *PC = dyn_cast<OMPFirstprivateClause>(C)) |
| 311 | Vars = PC->getVarRefs(); |
| 312 | else if (const auto *PC = dyn_cast<OMPLastprivateClause>(C)) |
| 313 | Vars = PC->getVarRefs(); |
| 314 | else |
| 315 | llvm_unreachable("Unexpected clause.")::llvm::llvm_unreachable_internal("Unexpected clause.", "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp" , 315); |
| 316 | for (const auto *E : Vars) { |
| 317 | const Decl *D = |
| 318 | cast<DeclRefExpr>(E)->getDecl()->getCanonicalDecl(); |
| 319 | if (D == VD->getCanonicalDecl()) { |
| 320 | IsForCombinedParallelRegion = true; |
| 321 | break; |
| 322 | } |
| 323 | } |
| 324 | if (IsForCombinedParallelRegion) |
| 325 | break; |
| 326 | } |
| 327 | } |
| 328 | markAsEscaped(VD); |
| 329 | if (isa<OMPCapturedExprDecl>(VD)) |
| 330 | VisitValueDecl(VD); |
| 331 | IsForCombinedParallelRegion = SavedIsForCombinedParallelRegion; |
| 332 | } |
| 333 | } |
| 334 | } |
| 335 | |
| 336 | void buildRecordForGlobalizedVars(bool IsInTTDRegion) { |
| 337 | assert(!GlobalizedRD &&(static_cast <bool> (!GlobalizedRD && "Record for globalized variables is built already." ) ? void (0) : __assert_fail ("!GlobalizedRD && \"Record for globalized variables is built already.\"" , "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 338, __extension__ __PRETTY_FUNCTION__)) |
| 338 | "Record for globalized variables is built already.")(static_cast <bool> (!GlobalizedRD && "Record for globalized variables is built already." ) ? void (0) : __assert_fail ("!GlobalizedRD && \"Record for globalized variables is built already.\"" , "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 338, __extension__ __PRETTY_FUNCTION__)); |
| 339 | ArrayRef<const ValueDecl *> EscapedDeclsForParallel, EscapedDeclsForTeams; |
| 340 | unsigned WarpSize = CGF.getTarget().getGridValue().GV_Warp_Size; |
| 341 | if (IsInTTDRegion) |
| 342 | EscapedDeclsForTeams = EscapedDecls.getArrayRef(); |
| 343 | else |
| 344 | EscapedDeclsForParallel = EscapedDecls.getArrayRef(); |
| 345 | GlobalizedRD = ::buildRecordForGlobalizedVars( |
| 346 | CGF.getContext(), EscapedDeclsForParallel, EscapedDeclsForTeams, |
| 347 | MappedDeclsFields, WarpSize); |
| 348 | } |
| 349 | |
| 350 | public: |
| 351 | CheckVarsEscapingDeclContext(CodeGenFunction &CGF, |
| 352 | ArrayRef<const ValueDecl *> TeamsReductions) |
| 353 | : CGF(CGF), EscapedDecls(TeamsReductions.begin(), TeamsReductions.end()) { |
| 354 | } |
| 355 | virtual ~CheckVarsEscapingDeclContext() = default; |
| 356 | void VisitDeclStmt(const DeclStmt *S) { |
| 357 | if (!S) |
| 358 | return; |
| 359 | for (const Decl *D : S->decls()) |
| 360 | if (const auto *VD = dyn_cast_or_null<ValueDecl>(D)) |
| 361 | VisitValueDecl(VD); |
| 362 | } |
| 363 | void VisitOMPExecutableDirective(const OMPExecutableDirective *D) { |
| 364 | if (!D) |
| 365 | return; |
| 366 | if (!D->hasAssociatedStmt()) |
| 367 | return; |
| 368 | if (const auto *S = |
| 369 | dyn_cast_or_null<CapturedStmt>(D->getAssociatedStmt())) { |
| 370 | // Do not analyze directives that do not actually require capturing, |
| 371 | // like `omp for` or `omp simd` directives. |
| 372 | llvm::SmallVector<OpenMPDirectiveKind, 4> CaptureRegions; |
| 373 | getOpenMPCaptureRegions(CaptureRegions, D->getDirectiveKind()); |
| 374 | if (CaptureRegions.size() == 1 && CaptureRegions.back() == OMPD_unknown) { |
| 375 | VisitStmt(S->getCapturedStmt()); |
| 376 | return; |
| 377 | } |
| 378 | VisitOpenMPCapturedStmt( |
| 379 | S, D->clauses(), |
| 380 | CaptureRegions.back() == OMPD_parallel && |
| 381 | isOpenMPDistributeDirective(D->getDirectiveKind())); |
| 382 | } |
| 383 | } |
| 384 | void VisitCapturedStmt(const CapturedStmt *S) { |
| 385 | if (!S) |
| 386 | return; |
| 387 | for (const CapturedStmt::Capture &C : S->captures()) { |
| 388 | if (C.capturesVariable() && !C.capturesVariableByCopy()) { |
| 389 | const ValueDecl *VD = C.getCapturedVar(); |
| 390 | markAsEscaped(VD); |
| 391 | if (isa<OMPCapturedExprDecl>(VD)) |
| 392 | VisitValueDecl(VD); |
| 393 | } |
| 394 | } |
| 395 | } |
| 396 | void VisitLambdaExpr(const LambdaExpr *E) { |
| 397 | if (!E) |
| 398 | return; |
| 399 | for (const LambdaCapture &C : E->captures()) { |
| 400 | if (C.capturesVariable()) { |
| 401 | if (C.getCaptureKind() == LCK_ByRef) { |
| 402 | const ValueDecl *VD = C.getCapturedVar(); |
| 403 | markAsEscaped(VD); |
| 404 | if (E->isInitCapture(&C) || isa<OMPCapturedExprDecl>(VD)) |
| 405 | VisitValueDecl(VD); |
| 406 | } |
| 407 | } |
| 408 | } |
| 409 | } |
| 410 | void VisitBlockExpr(const BlockExpr *E) { |
| 411 | if (!E) |
| 412 | return; |
| 413 | for (const BlockDecl::Capture &C : E->getBlockDecl()->captures()) { |
| 414 | if (C.isByRef()) { |
| 415 | const VarDecl *VD = C.getVariable(); |
| 416 | markAsEscaped(VD); |
| 417 | if (isa<OMPCapturedExprDecl>(VD) || VD->isInitCapture()) |
| 418 | VisitValueDecl(VD); |
| 419 | } |
| 420 | } |
| 421 | } |
| 422 | void VisitCallExpr(const CallExpr *E) { |
| 423 | if (!E) |
| 424 | return; |
| 425 | for (const Expr *Arg : E->arguments()) { |
| 426 | if (!Arg) |
| 427 | continue; |
| 428 | if (Arg->isLValue()) { |
| 429 | const bool SavedAllEscaped = AllEscaped; |
| 430 | AllEscaped = true; |
| 431 | Visit(Arg); |
| 432 | AllEscaped = SavedAllEscaped; |
| 433 | } else { |
| 434 | Visit(Arg); |
| 435 | } |
| 436 | } |
| 437 | Visit(E->getCallee()); |
| 438 | } |
| 439 | void VisitDeclRefExpr(const DeclRefExpr *E) { |
| 440 | if (!E) |
| 441 | return; |
| 442 | const ValueDecl *VD = E->getDecl(); |
| 443 | if (AllEscaped) |
| 444 | markAsEscaped(VD); |
| 445 | if (isa<OMPCapturedExprDecl>(VD)) |
| 446 | VisitValueDecl(VD); |
| 447 | else if (const auto *VarD = dyn_cast<VarDecl>(VD)) |
| 448 | if (VarD->isInitCapture()) |
| 449 | VisitValueDecl(VD); |
| 450 | } |
| 451 | void VisitUnaryOperator(const UnaryOperator *E) { |
| 452 | if (!E) |
| 453 | return; |
| 454 | if (E->getOpcode() == UO_AddrOf) { |
| 455 | const bool SavedAllEscaped = AllEscaped; |
| 456 | AllEscaped = true; |
| 457 | Visit(E->getSubExpr()); |
| 458 | AllEscaped = SavedAllEscaped; |
| 459 | } else { |
| 460 | Visit(E->getSubExpr()); |
| 461 | } |
| 462 | } |
| 463 | void VisitImplicitCastExpr(const ImplicitCastExpr *E) { |
| 464 | if (!E) |
| 465 | return; |
| 466 | if (E->getCastKind() == CK_ArrayToPointerDecay) { |
| 467 | const bool SavedAllEscaped = AllEscaped; |
| 468 | AllEscaped = true; |
| 469 | Visit(E->getSubExpr()); |
| 470 | AllEscaped = SavedAllEscaped; |
| 471 | } else { |
| 472 | Visit(E->getSubExpr()); |
| 473 | } |
| 474 | } |
| 475 | void VisitExpr(const Expr *E) { |
| 476 | if (!E) |
| 477 | return; |
| 478 | bool SavedAllEscaped = AllEscaped; |
| 479 | if (!E->isLValue()) |
| 480 | AllEscaped = false; |
| 481 | for (const Stmt *Child : E->children()) |
| 482 | if (Child) |
| 483 | Visit(Child); |
| 484 | AllEscaped = SavedAllEscaped; |
| 485 | } |
| 486 | void VisitStmt(const Stmt *S) { |
| 487 | if (!S) |
| 488 | return; |
| 489 | for (const Stmt *Child : S->children()) |
| 490 | if (Child) |
| 491 | Visit(Child); |
| 492 | } |
| 493 | |
| 494 | /// Returns the record that handles all the escaped local variables and used |
| 495 | /// instead of their original storage. |
| 496 | const RecordDecl *getGlobalizedRecord(bool IsInTTDRegion) { |
| 497 | if (!GlobalizedRD) |
| 498 | buildRecordForGlobalizedVars(IsInTTDRegion); |
| 499 | return GlobalizedRD; |
| 500 | } |
| 501 | |
| 502 | /// Returns the field in the globalized record for the escaped variable. |
| 503 | const FieldDecl *getFieldForGlobalizedVar(const ValueDecl *VD) const { |
| 504 | assert(GlobalizedRD &&(static_cast <bool> (GlobalizedRD && "Record for globalized variables must be generated already." ) ? void (0) : __assert_fail ("GlobalizedRD && \"Record for globalized variables must be generated already.\"" , "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 505, __extension__ __PRETTY_FUNCTION__)) |
| 505 | "Record for globalized variables must be generated already.")(static_cast <bool> (GlobalizedRD && "Record for globalized variables must be generated already." ) ? void (0) : __assert_fail ("GlobalizedRD && \"Record for globalized variables must be generated already.\"" , "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 505, __extension__ __PRETTY_FUNCTION__)); |
| 506 | auto I = MappedDeclsFields.find(VD); |
| 507 | if (I == MappedDeclsFields.end()) |
| 508 | return nullptr; |
| 509 | return I->getSecond(); |
| 510 | } |
| 511 | |
| 512 | /// Returns the list of the escaped local variables/parameters. |
| 513 | ArrayRef<const ValueDecl *> getEscapedDecls() const { |
| 514 | return EscapedDecls.getArrayRef(); |
| 515 | } |
| 516 | |
| 517 | /// Checks if the escaped local variable is actually a parameter passed by |
| 518 | /// value. |
| 519 | const llvm::SmallPtrSetImpl<const Decl *> &getEscapedParameters() const { |
| 520 | return EscapedParameters; |
| 521 | } |
| 522 | |
| 523 | /// Returns the list of the escaped variables with the variably modified |
| 524 | /// types. |
| 525 | ArrayRef<const ValueDecl *> getEscapedVariableLengthDecls() const { |
| 526 | return EscapedVariableLengthDecls.getArrayRef(); |
| 527 | } |
| 528 | }; |
| 529 | } // anonymous namespace |
| 530 | |
| 531 | /// Get the id of the warp in the block. |
| 532 | /// We assume that the warp size is 32, which is always the case |
| 533 | /// on the NVPTX device, to generate more efficient code. |
| 534 | static llvm::Value *getNVPTXWarpID(CodeGenFunction &CGF) { |
| 535 | CGBuilderTy &Bld = CGF.Builder; |
| 536 | unsigned LaneIDBits = |
| 537 | llvm::Log2_32(CGF.getTarget().getGridValue().GV_Warp_Size); |
| 538 | auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()); |
| 539 | return Bld.CreateAShr(RT.getGPUThreadID(CGF), LaneIDBits, "nvptx_warp_id"); |
| 540 | } |
| 541 | |
| 542 | /// Get the id of the current lane in the Warp. |
| 543 | /// We assume that the warp size is 32, which is always the case |
| 544 | /// on the NVPTX device, to generate more efficient code. |
| 545 | static llvm::Value *getNVPTXLaneID(CodeGenFunction &CGF) { |
| 546 | CGBuilderTy &Bld = CGF.Builder; |
| 547 | unsigned LaneIDBits = |
| 548 | llvm::Log2_32(CGF.getTarget().getGridValue().GV_Warp_Size); |
| 549 | unsigned LaneIDMask = ~0u >> (32u - LaneIDBits); |
| 550 | auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()); |
| 551 | return Bld.CreateAnd(RT.getGPUThreadID(CGF), Bld.getInt32(LaneIDMask), |
| 552 | "nvptx_lane_id"); |
| 553 | } |
| 554 | |
| 555 | CGOpenMPRuntimeGPU::ExecutionMode |
| 556 | CGOpenMPRuntimeGPU::getExecutionMode() const { |
| 557 | return CurrentExecutionMode; |
| 558 | } |
| 559 | |
| 560 | static CGOpenMPRuntimeGPU::DataSharingMode |
| 561 | getDataSharingMode(CodeGenModule &CGM) { |
| 562 | return CGM.getLangOpts().OpenMPCUDAMode ? CGOpenMPRuntimeGPU::CUDA |
| 563 | : CGOpenMPRuntimeGPU::Generic; |
| 564 | } |
| 565 | |
| 566 | /// Check for inner (nested) SPMD construct, if any |
| 567 | static bool hasNestedSPMDDirective(ASTContext &Ctx, |
| 568 | const OMPExecutableDirective &D) { |
| 569 | const auto *CS = D.getInnermostCapturedStmt(); |
| 570 | const auto *Body = |
| 571 | CS->getCapturedStmt()->IgnoreContainers(/*IgnoreCaptured=*/true); |
| 572 | const Stmt *ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body); |
| 573 | |
| 574 | if (const auto *NestedDir = |
| 575 | dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) { |
| 576 | OpenMPDirectiveKind DKind = NestedDir->getDirectiveKind(); |
| 577 | switch (D.getDirectiveKind()) { |
| 578 | case OMPD_target: |
| 579 | if (isOpenMPParallelDirective(DKind)) |
| 580 | return true; |
| 581 | if (DKind == OMPD_teams) { |
| 582 | Body = NestedDir->getInnermostCapturedStmt()->IgnoreContainers( |
| 583 | /*IgnoreCaptured=*/true); |
| 584 | if (!Body) |
| 585 | return false; |
| 586 | ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body); |
| 587 | if (const auto *NND = |
| 588 | dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) { |
| 589 | DKind = NND->getDirectiveKind(); |
| 590 | if (isOpenMPParallelDirective(DKind)) |
| 591 | return true; |
| 592 | } |
| 593 | } |
| 594 | return false; |
| 595 | case OMPD_target_teams: |
| 596 | return isOpenMPParallelDirective(DKind); |
| 597 | case OMPD_target_simd: |
| 598 | case OMPD_target_parallel: |
| 599 | case OMPD_target_parallel_for: |
| 600 | case OMPD_target_parallel_for_simd: |
| 601 | case OMPD_target_teams_distribute: |
| 602 | case OMPD_target_teams_distribute_simd: |
| 603 | case OMPD_target_teams_distribute_parallel_for: |
| 604 | case OMPD_target_teams_distribute_parallel_for_simd: |
| 605 | case OMPD_parallel: |
| 606 | case OMPD_for: |
| 607 | case OMPD_parallel_for: |
| 608 | case OMPD_parallel_master: |
| 609 | case OMPD_parallel_sections: |
| 610 | case OMPD_for_simd: |
| 611 | case OMPD_parallel_for_simd: |
| 612 | case OMPD_cancel: |
| 613 | case OMPD_cancellation_point: |
| 614 | case OMPD_ordered: |
| 615 | case OMPD_threadprivate: |
| 616 | case OMPD_allocate: |
| 617 | case OMPD_task: |
| 618 | case OMPD_simd: |
| 619 | case OMPD_sections: |
| 620 | case OMPD_section: |
| 621 | case OMPD_single: |
| 622 | case OMPD_master: |
| 623 | case OMPD_critical: |
| 624 | case OMPD_taskyield: |
| 625 | case OMPD_barrier: |
| 626 | case OMPD_taskwait: |
| 627 | case OMPD_taskgroup: |
| 628 | case OMPD_atomic: |
| 629 | case OMPD_flush: |
| 630 | case OMPD_depobj: |
| 631 | case OMPD_scan: |
| 632 | case OMPD_teams: |
| 633 | case OMPD_target_data: |
| 634 | case OMPD_target_exit_data: |
| 635 | case OMPD_target_enter_data: |
| 636 | case OMPD_distribute: |
| 637 | case OMPD_distribute_simd: |
| 638 | case OMPD_distribute_parallel_for: |
| 639 | case OMPD_distribute_parallel_for_simd: |
| 640 | case OMPD_teams_distribute: |
| 641 | case OMPD_teams_distribute_simd: |
| 642 | case OMPD_teams_distribute_parallel_for: |
| 643 | case OMPD_teams_distribute_parallel_for_simd: |
| 644 | case OMPD_target_update: |
| 645 | case OMPD_declare_simd: |
| 646 | case OMPD_declare_variant: |
| 647 | case OMPD_begin_declare_variant: |
| 648 | case OMPD_end_declare_variant: |
| 649 | case OMPD_declare_target: |
| 650 | case OMPD_end_declare_target: |
| 651 | case OMPD_declare_reduction: |
| 652 | case OMPD_declare_mapper: |
| 653 | case OMPD_taskloop: |
| 654 | case OMPD_taskloop_simd: |
| 655 | case OMPD_master_taskloop: |
| 656 | case OMPD_master_taskloop_simd: |
| 657 | case OMPD_parallel_master_taskloop: |
| 658 | case OMPD_parallel_master_taskloop_simd: |
| 659 | case OMPD_requires: |
| 660 | case OMPD_unknown: |
| 661 | default: |
| 662 | llvm_unreachable("Unexpected directive.")::llvm::llvm_unreachable_internal("Unexpected directive.", "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp" , 662); |
| 663 | } |
| 664 | } |
| 665 | |
| 666 | return false; |
| 667 | } |
| 668 | |
| 669 | static bool supportsSPMDExecutionMode(ASTContext &Ctx, |
| 670 | const OMPExecutableDirective &D) { |
| 671 | OpenMPDirectiveKind DirectiveKind = D.getDirectiveKind(); |
| 672 | switch (DirectiveKind) { |
| 673 | case OMPD_target: |
| 674 | case OMPD_target_teams: |
| 675 | return hasNestedSPMDDirective(Ctx, D); |
| 676 | case OMPD_target_parallel: |
| 677 | case OMPD_target_parallel_for: |
| 678 | case OMPD_target_parallel_for_simd: |
| 679 | case OMPD_target_teams_distribute_parallel_for: |
| 680 | case OMPD_target_teams_distribute_parallel_for_simd: |
| 681 | case OMPD_target_simd: |
| 682 | case OMPD_target_teams_distribute_simd: |
| 683 | return true; |
| 684 | case OMPD_target_teams_distribute: |
| 685 | return false; |
| 686 | case OMPD_parallel: |
| 687 | case OMPD_for: |
| 688 | case OMPD_parallel_for: |
| 689 | case OMPD_parallel_master: |
| 690 | case OMPD_parallel_sections: |
| 691 | case OMPD_for_simd: |
| 692 | case OMPD_parallel_for_simd: |
| 693 | case OMPD_cancel: |
| 694 | case OMPD_cancellation_point: |
| 695 | case OMPD_ordered: |
| 696 | case OMPD_threadprivate: |
| 697 | case OMPD_allocate: |
| 698 | case OMPD_task: |
| 699 | case OMPD_simd: |
| 700 | case OMPD_sections: |
| 701 | case OMPD_section: |
| 702 | case OMPD_single: |
| 703 | case OMPD_master: |
| 704 | case OMPD_critical: |
| 705 | case OMPD_taskyield: |
| 706 | case OMPD_barrier: |
| 707 | case OMPD_taskwait: |
| 708 | case OMPD_taskgroup: |
| 709 | case OMPD_atomic: |
| 710 | case OMPD_flush: |
| 711 | case OMPD_depobj: |
| 712 | case OMPD_scan: |
| 713 | case OMPD_teams: |
| 714 | case OMPD_target_data: |
| 715 | case OMPD_target_exit_data: |
| 716 | case OMPD_target_enter_data: |
| 717 | case OMPD_distribute: |
| 718 | case OMPD_distribute_simd: |
| 719 | case OMPD_distribute_parallel_for: |
| 720 | case OMPD_distribute_parallel_for_simd: |
| 721 | case OMPD_teams_distribute: |
| 722 | case OMPD_teams_distribute_simd: |
| 723 | case OMPD_teams_distribute_parallel_for: |
| 724 | case OMPD_teams_distribute_parallel_for_simd: |
| 725 | case OMPD_target_update: |
| 726 | case OMPD_declare_simd: |
| 727 | case OMPD_declare_variant: |
| 728 | case OMPD_begin_declare_variant: |
| 729 | case OMPD_end_declare_variant: |
| 730 | case OMPD_declare_target: |
| 731 | case OMPD_end_declare_target: |
| 732 | case OMPD_declare_reduction: |
| 733 | case OMPD_declare_mapper: |
| 734 | case OMPD_taskloop: |
| 735 | case OMPD_taskloop_simd: |
| 736 | case OMPD_master_taskloop: |
| 737 | case OMPD_master_taskloop_simd: |
| 738 | case OMPD_parallel_master_taskloop: |
| 739 | case OMPD_parallel_master_taskloop_simd: |
| 740 | case OMPD_requires: |
| 741 | case OMPD_unknown: |
| 742 | default: |
| 743 | break; |
| 744 | } |
| 745 | llvm_unreachable(::llvm::llvm_unreachable_internal("Unknown programming model for OpenMP directive on NVPTX target." , "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 746) |
| 746 | "Unknown programming model for OpenMP directive on NVPTX target.")::llvm::llvm_unreachable_internal("Unknown programming model for OpenMP directive on NVPTX target." , "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 746); |
| 747 | } |
| 748 | |
| 749 | /// Check if the directive is loops based and has schedule clause at all or has |
| 750 | /// static scheduling. |
| 751 | static bool hasStaticScheduling(const OMPExecutableDirective &D) { |
| 752 | assert(isOpenMPWorksharingDirective(D.getDirectiveKind()) &&(static_cast <bool> (isOpenMPWorksharingDirective(D.getDirectiveKind ()) && isOpenMPLoopDirective(D.getDirectiveKind()) && "Expected loop-based directive.") ? void (0) : __assert_fail ("isOpenMPWorksharingDirective(D.getDirectiveKind()) && isOpenMPLoopDirective(D.getDirectiveKind()) && \"Expected loop-based directive.\"" , "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 754, __extension__ __PRETTY_FUNCTION__)) |
| 753 | isOpenMPLoopDirective(D.getDirectiveKind()) &&(static_cast <bool> (isOpenMPWorksharingDirective(D.getDirectiveKind ()) && isOpenMPLoopDirective(D.getDirectiveKind()) && "Expected loop-based directive.") ? void (0) : __assert_fail ("isOpenMPWorksharingDirective(D.getDirectiveKind()) && isOpenMPLoopDirective(D.getDirectiveKind()) && \"Expected loop-based directive.\"" , "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 754, __extension__ __PRETTY_FUNCTION__)) |
| 754 | "Expected loop-based directive.")(static_cast <bool> (isOpenMPWorksharingDirective(D.getDirectiveKind ()) && isOpenMPLoopDirective(D.getDirectiveKind()) && "Expected loop-based directive.") ? void (0) : __assert_fail ("isOpenMPWorksharingDirective(D.getDirectiveKind()) && isOpenMPLoopDirective(D.getDirectiveKind()) && \"Expected loop-based directive.\"" , "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 754, __extension__ __PRETTY_FUNCTION__)); |
| 755 | return !D.hasClausesOfKind<OMPOrderedClause>() && |
| 756 | (!D.hasClausesOfKind<OMPScheduleClause>() || |
| 757 | llvm::any_of(D.getClausesOfKind<OMPScheduleClause>(), |
| 758 | [](const OMPScheduleClause *C) { |
| 759 | return C->getScheduleKind() == OMPC_SCHEDULE_static; |
| 760 | })); |
| 761 | } |
| 762 | |
| 763 | /// Check for inner (nested) lightweight runtime construct, if any |
| 764 | static bool hasNestedLightweightDirective(ASTContext &Ctx, |
| 765 | const OMPExecutableDirective &D) { |
| 766 | assert(supportsSPMDExecutionMode(Ctx, D) && "Expected SPMD mode directive.")(static_cast <bool> (supportsSPMDExecutionMode(Ctx, D) && "Expected SPMD mode directive.") ? void (0) : __assert_fail ( "supportsSPMDExecutionMode(Ctx, D) && \"Expected SPMD mode directive.\"" , "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 766, __extension__ __PRETTY_FUNCTION__)); |
| 767 | const auto *CS = D.getInnermostCapturedStmt(); |
| 768 | const auto *Body = |
| 769 | CS->getCapturedStmt()->IgnoreContainers(/*IgnoreCaptured=*/true); |
| 770 | const Stmt *ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body); |
| 771 | |
| 772 | if (const auto *NestedDir = |
| 773 | dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) { |
| 774 | OpenMPDirectiveKind DKind = NestedDir->getDirectiveKind(); |
| 775 | switch (D.getDirectiveKind()) { |
| 776 | case OMPD_target: |
| 777 | if (isOpenMPParallelDirective(DKind) && |
| 778 | isOpenMPWorksharingDirective(DKind) && isOpenMPLoopDirective(DKind) && |
| 779 | hasStaticScheduling(*NestedDir)) |
| 780 | return true; |
| 781 | if (DKind == OMPD_teams_distribute_simd || DKind == OMPD_simd) |
| 782 | return true; |
| 783 | if (DKind == OMPD_parallel) { |
| 784 | Body = NestedDir->getInnermostCapturedStmt()->IgnoreContainers( |
| 785 | /*IgnoreCaptured=*/true); |
| 786 | if (!Body) |
| 787 | return false; |
| 788 | ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body); |
| 789 | if (const auto *NND = |
| 790 | dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) { |
| 791 | DKind = NND->getDirectiveKind(); |
| 792 | if (isOpenMPWorksharingDirective(DKind) && |
| 793 | isOpenMPLoopDirective(DKind) && hasStaticScheduling(*NND)) |
| 794 | return true; |
| 795 | } |
| 796 | } else if (DKind == OMPD_teams) { |
| 797 | Body = NestedDir->getInnermostCapturedStmt()->IgnoreContainers( |
| 798 | /*IgnoreCaptured=*/true); |
| 799 | if (!Body) |
| 800 | return false; |
| 801 | ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body); |
| 802 | if (const auto *NND = |
| 803 | dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) { |
| 804 | DKind = NND->getDirectiveKind(); |
| 805 | if (isOpenMPParallelDirective(DKind) && |
| 806 | isOpenMPWorksharingDirective(DKind) && |
| 807 | isOpenMPLoopDirective(DKind) && hasStaticScheduling(*NND)) |
| 808 | return true; |
| 809 | if (DKind == OMPD_parallel) { |
| 810 | Body = NND->getInnermostCapturedStmt()->IgnoreContainers( |
| 811 | /*IgnoreCaptured=*/true); |
| 812 | if (!Body) |
| 813 | return false; |
| 814 | ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body); |
| 815 | if (const auto *NND = |
| 816 | dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) { |
| 817 | DKind = NND->getDirectiveKind(); |
| 818 | if (isOpenMPWorksharingDirective(DKind) && |
| 819 | isOpenMPLoopDirective(DKind) && hasStaticScheduling(*NND)) |
| 820 | return true; |
| 821 | } |
| 822 | } |
| 823 | } |
| 824 | } |
| 825 | return false; |
| 826 | case OMPD_target_teams: |
| 827 | if (isOpenMPParallelDirective(DKind) && |
| 828 | isOpenMPWorksharingDirective(DKind) && isOpenMPLoopDirective(DKind) && |
| 829 | hasStaticScheduling(*NestedDir)) |
| 830 | return true; |
| 831 | if (DKind == OMPD_distribute_simd || DKind == OMPD_simd) |
| 832 | return true; |
| 833 | if (DKind == OMPD_parallel) { |
| 834 | Body = NestedDir->getInnermostCapturedStmt()->IgnoreContainers( |
| 835 | /*IgnoreCaptured=*/true); |
| 836 | if (!Body) |
| 837 | return false; |
| 838 | ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body); |
| 839 | if (const auto *NND = |
| 840 | dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) { |
| 841 | DKind = NND->getDirectiveKind(); |
| 842 | if (isOpenMPWorksharingDirective(DKind) && |
| 843 | isOpenMPLoopDirective(DKind) && hasStaticScheduling(*NND)) |
| 844 | return true; |
| 845 | } |
| 846 | } |
| 847 | return false; |
| 848 | case OMPD_target_parallel: |
| 849 | if (DKind == OMPD_simd) |
| 850 | return true; |
| 851 | return isOpenMPWorksharingDirective(DKind) && |
| 852 | isOpenMPLoopDirective(DKind) && hasStaticScheduling(*NestedDir); |
| 853 | case OMPD_target_teams_distribute: |
| 854 | case OMPD_target_simd: |
| 855 | case OMPD_target_parallel_for: |
| 856 | case OMPD_target_parallel_for_simd: |
| 857 | case OMPD_target_teams_distribute_simd: |
| 858 | case OMPD_target_teams_distribute_parallel_for: |
| 859 | case OMPD_target_teams_distribute_parallel_for_simd: |
| 860 | case OMPD_parallel: |
| 861 | case OMPD_for: |
| 862 | case OMPD_parallel_for: |
| 863 | case OMPD_parallel_master: |
| 864 | case OMPD_parallel_sections: |
| 865 | case OMPD_for_simd: |
| 866 | case OMPD_parallel_for_simd: |
| 867 | case OMPD_cancel: |
| 868 | case OMPD_cancellation_point: |
| 869 | case OMPD_ordered: |
| 870 | case OMPD_threadprivate: |
| 871 | case OMPD_allocate: |
| 872 | case OMPD_task: |
| 873 | case OMPD_simd: |
| 874 | case OMPD_sections: |
| 875 | case OMPD_section: |
| 876 | case OMPD_single: |
| 877 | case OMPD_master: |
| 878 | case OMPD_critical: |
| 879 | case OMPD_taskyield: |
| 880 | case OMPD_barrier: |
| 881 | case OMPD_taskwait: |
| 882 | case OMPD_taskgroup: |
| 883 | case OMPD_atomic: |
| 884 | case OMPD_flush: |
| 885 | case OMPD_depobj: |
| 886 | case OMPD_scan: |
| 887 | case OMPD_teams: |
| 888 | case OMPD_target_data: |
| 889 | case OMPD_target_exit_data: |
| 890 | case OMPD_target_enter_data: |
| 891 | case OMPD_distribute: |
| 892 | case OMPD_distribute_simd: |
| 893 | case OMPD_distribute_parallel_for: |
| 894 | case OMPD_distribute_parallel_for_simd: |
| 895 | case OMPD_teams_distribute: |
| 896 | case OMPD_teams_distribute_simd: |
| 897 | case OMPD_teams_distribute_parallel_for: |
| 898 | case OMPD_teams_distribute_parallel_for_simd: |
| 899 | case OMPD_target_update: |
| 900 | case OMPD_declare_simd: |
| 901 | case OMPD_declare_variant: |
| 902 | case OMPD_begin_declare_variant: |
| 903 | case OMPD_end_declare_variant: |
| 904 | case OMPD_declare_target: |
| 905 | case OMPD_end_declare_target: |
| 906 | case OMPD_declare_reduction: |
| 907 | case OMPD_declare_mapper: |
| 908 | case OMPD_taskloop: |
| 909 | case OMPD_taskloop_simd: |
| 910 | case OMPD_master_taskloop: |
| 911 | case OMPD_master_taskloop_simd: |
| 912 | case OMPD_parallel_master_taskloop: |
| 913 | case OMPD_parallel_master_taskloop_simd: |
| 914 | case OMPD_requires: |
| 915 | case OMPD_unknown: |
| 916 | default: |
| 917 | llvm_unreachable("Unexpected directive.")::llvm::llvm_unreachable_internal("Unexpected directive.", "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp" , 917); |
| 918 | } |
| 919 | } |
| 920 | |
| 921 | return false; |
| 922 | } |
| 923 | |
| 924 | /// Checks if the construct supports lightweight runtime. It must be SPMD |
| 925 | /// construct + inner loop-based construct with static scheduling. |
| 926 | static bool supportsLightweightRuntime(ASTContext &Ctx, |
| 927 | const OMPExecutableDirective &D) { |
| 928 | if (!supportsSPMDExecutionMode(Ctx, D)) |
| 929 | return false; |
| 930 | OpenMPDirectiveKind DirectiveKind = D.getDirectiveKind(); |
| 931 | switch (DirectiveKind) { |
| 932 | case OMPD_target: |
| 933 | case OMPD_target_teams: |
| 934 | case OMPD_target_parallel: |
| 935 | return hasNestedLightweightDirective(Ctx, D); |
| 936 | case OMPD_target_parallel_for: |
| 937 | case OMPD_target_parallel_for_simd: |
| 938 | case OMPD_target_teams_distribute_parallel_for: |
| 939 | case OMPD_target_teams_distribute_parallel_for_simd: |
| 940 | // (Last|First)-privates must be shared in parallel region. |
| 941 | return hasStaticScheduling(D); |
| 942 | case OMPD_target_simd: |
| 943 | case OMPD_target_teams_distribute_simd: |
| 944 | return true; |
| 945 | case OMPD_target_teams_distribute: |
| 946 | return false; |
| 947 | case OMPD_parallel: |
| 948 | case OMPD_for: |
| 949 | case OMPD_parallel_for: |
| 950 | case OMPD_parallel_master: |
| 951 | case OMPD_parallel_sections: |
| 952 | case OMPD_for_simd: |
| 953 | case OMPD_parallel_for_simd: |
| 954 | case OMPD_cancel: |
| 955 | case OMPD_cancellation_point: |
| 956 | case OMPD_ordered: |
| 957 | case OMPD_threadprivate: |
| 958 | case OMPD_allocate: |
| 959 | case OMPD_task: |
| 960 | case OMPD_simd: |
| 961 | case OMPD_sections: |
| 962 | case OMPD_section: |
| 963 | case OMPD_single: |
| 964 | case OMPD_master: |
| 965 | case OMPD_critical: |
| 966 | case OMPD_taskyield: |
| 967 | case OMPD_barrier: |
| 968 | case OMPD_taskwait: |
| 969 | case OMPD_taskgroup: |
| 970 | case OMPD_atomic: |
| 971 | case OMPD_flush: |
| 972 | case OMPD_depobj: |
| 973 | case OMPD_scan: |
| 974 | case OMPD_teams: |
| 975 | case OMPD_target_data: |
| 976 | case OMPD_target_exit_data: |
| 977 | case OMPD_target_enter_data: |
| 978 | case OMPD_distribute: |
| 979 | case OMPD_distribute_simd: |
| 980 | case OMPD_distribute_parallel_for: |
| 981 | case OMPD_distribute_parallel_for_simd: |
| 982 | case OMPD_teams_distribute: |
| 983 | case OMPD_teams_distribute_simd: |
| 984 | case OMPD_teams_distribute_parallel_for: |
| 985 | case OMPD_teams_distribute_parallel_for_simd: |
| 986 | case OMPD_target_update: |
| 987 | case OMPD_declare_simd: |
| 988 | case OMPD_declare_variant: |
| 989 | case OMPD_begin_declare_variant: |
| 990 | case OMPD_end_declare_variant: |
| 991 | case OMPD_declare_target: |
| 992 | case OMPD_end_declare_target: |
| 993 | case OMPD_declare_reduction: |
| 994 | case OMPD_declare_mapper: |
| 995 | case OMPD_taskloop: |
| 996 | case OMPD_taskloop_simd: |
| 997 | case OMPD_master_taskloop: |
| 998 | case OMPD_master_taskloop_simd: |
| 999 | case OMPD_parallel_master_taskloop: |
| 1000 | case OMPD_parallel_master_taskloop_simd: |
| 1001 | case OMPD_requires: |
| 1002 | case OMPD_unknown: |
| 1003 | default: |
| 1004 | break; |
| 1005 | } |
| 1006 | llvm_unreachable(::llvm::llvm_unreachable_internal("Unknown programming model for OpenMP directive on NVPTX target." , "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 1007) |
| 1007 | "Unknown programming model for OpenMP directive on NVPTX target.")::llvm::llvm_unreachable_internal("Unknown programming model for OpenMP directive on NVPTX target." , "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 1007); |
| 1008 | } |
| 1009 | |
| 1010 | void CGOpenMPRuntimeGPU::emitNonSPMDKernel(const OMPExecutableDirective &D, |
| 1011 | StringRef ParentName, |
| 1012 | llvm::Function *&OutlinedFn, |
| 1013 | llvm::Constant *&OutlinedFnID, |
| 1014 | bool IsOffloadEntry, |
| 1015 | const RegionCodeGenTy &CodeGen) { |
| 1016 | ExecutionRuntimeModesRAII ModeRAII(CurrentExecutionMode); |
| 1017 | EntryFunctionState EST; |
| 1018 | WrapperFunctionsMap.clear(); |
| 1019 | |
| 1020 | // Emit target region as a standalone region. |
| 1021 | class NVPTXPrePostActionTy : public PrePostActionTy { |
| 1022 | CGOpenMPRuntimeGPU::EntryFunctionState &EST; |
| 1023 | |
| 1024 | public: |
| 1025 | NVPTXPrePostActionTy(CGOpenMPRuntimeGPU::EntryFunctionState &EST) |
| 1026 | : EST(EST) {} |
| 1027 | void Enter(CodeGenFunction &CGF) override { |
| 1028 | auto &RT = |
| 1029 | static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()); |
| 1030 | RT.emitKernelInit(CGF, EST, /* IsSPMD */ false); |
| 1031 | // Skip target region initialization. |
| 1032 | RT.setLocThreadIdInsertPt(CGF, /*AtCurrentPoint=*/true); |
| 1033 | } |
| 1034 | void Exit(CodeGenFunction &CGF) override { |
| 1035 | auto &RT = |
| 1036 | static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()); |
| 1037 | RT.clearLocThreadIdInsertPt(CGF); |
| 1038 | RT.emitKernelDeinit(CGF, EST, /* IsSPMD */ false); |
| 1039 | } |
| 1040 | } Action(EST); |
| 1041 | CodeGen.setAction(Action); |
| 1042 | IsInTTDRegion = true; |
| 1043 | emitTargetOutlinedFunctionHelper(D, ParentName, OutlinedFn, OutlinedFnID, |
| 1044 | IsOffloadEntry, CodeGen); |
| 1045 | IsInTTDRegion = false; |
| 1046 | } |
| 1047 | |
| 1048 | void CGOpenMPRuntimeGPU::emitKernelInit(CodeGenFunction &CGF, |
| 1049 | EntryFunctionState &EST, bool IsSPMD) { |
| 1050 | CGBuilderTy &Bld = CGF.Builder; |
| 1051 | Bld.restoreIP(OMPBuilder.createTargetInit(Bld, IsSPMD, requiresFullRuntime())); |
| 1052 | IsInTargetMasterThreadRegion = IsSPMD; |
| 1053 | if (!IsSPMD) |
| 1054 | emitGenericVarsProlog(CGF, EST.Loc); |
| 1055 | } |
| 1056 | |
| 1057 | void CGOpenMPRuntimeGPU::emitKernelDeinit(CodeGenFunction &CGF, |
| 1058 | EntryFunctionState &EST, |
| 1059 | bool IsSPMD) { |
| 1060 | if (!IsSPMD) |
| 1061 | emitGenericVarsEpilog(CGF); |
| 1062 | |
| 1063 | CGBuilderTy &Bld = CGF.Builder; |
| 1064 | OMPBuilder.createTargetDeinit(Bld, IsSPMD, requiresFullRuntime()); |
| 1065 | } |
| 1066 | |
| 1067 | void CGOpenMPRuntimeGPU::emitSPMDKernel(const OMPExecutableDirective &D, |
| 1068 | StringRef ParentName, |
| 1069 | llvm::Function *&OutlinedFn, |
| 1070 | llvm::Constant *&OutlinedFnID, |
| 1071 | bool IsOffloadEntry, |
| 1072 | const RegionCodeGenTy &CodeGen) { |
| 1073 | ExecutionRuntimeModesRAII ModeRAII( |
| 1074 | CurrentExecutionMode, RequiresFullRuntime, |
| 1075 | CGM.getLangOpts().OpenMPCUDAForceFullRuntime || |
| 1076 | !supportsLightweightRuntime(CGM.getContext(), D)); |
| 1077 | EntryFunctionState EST; |
| 1078 | |
| 1079 | // Emit target region as a standalone region. |
| 1080 | class NVPTXPrePostActionTy : public PrePostActionTy { |
| 1081 | CGOpenMPRuntimeGPU &RT; |
| 1082 | CGOpenMPRuntimeGPU::EntryFunctionState &EST; |
| 1083 | |
| 1084 | public: |
| 1085 | NVPTXPrePostActionTy(CGOpenMPRuntimeGPU &RT, |
| 1086 | CGOpenMPRuntimeGPU::EntryFunctionState &EST) |
| 1087 | : RT(RT), EST(EST) {} |
| 1088 | void Enter(CodeGenFunction &CGF) override { |
| 1089 | RT.emitKernelInit(CGF, EST, /* IsSPMD */ true); |
| 1090 | // Skip target region initialization. |
| 1091 | RT.setLocThreadIdInsertPt(CGF, /*AtCurrentPoint=*/true); |
| 1092 | } |
| 1093 | void Exit(CodeGenFunction &CGF) override { |
| 1094 | RT.clearLocThreadIdInsertPt(CGF); |
| 1095 | RT.emitKernelDeinit(CGF, EST, /* IsSPMD */ true); |
| 1096 | } |
| 1097 | } Action(*this, EST); |
| 1098 | CodeGen.setAction(Action); |
| 1099 | IsInTTDRegion = true; |
| 1100 | emitTargetOutlinedFunctionHelper(D, ParentName, OutlinedFn, OutlinedFnID, |
| 1101 | IsOffloadEntry, CodeGen); |
| 1102 | IsInTTDRegion = false; |
| 1103 | } |
| 1104 | |
| 1105 | // Create a unique global variable to indicate the execution mode of this target |
| 1106 | // region. The execution mode is either 'generic', or 'spmd' depending on the |
| 1107 | // target directive. This variable is picked up by the offload library to setup |
| 1108 | // the device appropriately before kernel launch. If the execution mode is |
| 1109 | // 'generic', the runtime reserves one warp for the master, otherwise, all |
| 1110 | // warps participate in parallel work. |
| 1111 | static void setPropertyExecutionMode(CodeGenModule &CGM, StringRef Name, |
| 1112 | bool Mode) { |
| 1113 | auto *GVMode = new llvm::GlobalVariable( |
| 1114 | CGM.getModule(), CGM.Int8Ty, /*isConstant=*/true, |
| 1115 | llvm::GlobalValue::WeakAnyLinkage, |
| 1116 | llvm::ConstantInt::get(CGM.Int8Ty, Mode ? OMP_TGT_EXEC_MODE_SPMD |
| 1117 | : OMP_TGT_EXEC_MODE_GENERIC), |
| 1118 | Twine(Name, "_exec_mode")); |
| 1119 | CGM.addCompilerUsedGlobal(GVMode); |
| 1120 | } |
| 1121 | |
| 1122 | void CGOpenMPRuntimeGPU::createOffloadEntry(llvm::Constant *ID, |
| 1123 | llvm::Constant *Addr, |
| 1124 | uint64_t Size, int32_t, |
| 1125 | llvm::GlobalValue::LinkageTypes) { |
| 1126 | // TODO: Add support for global variables on the device after declare target |
| 1127 | // support. |
| 1128 | llvm::Function *Fn = dyn_cast<llvm::Function>(Addr); |
| 1129 | if (!Fn) |
| 1130 | return; |
| 1131 | |
| 1132 | llvm::Module &M = CGM.getModule(); |
| 1133 | llvm::LLVMContext &Ctx = CGM.getLLVMContext(); |
| 1134 | |
| 1135 | // Get "nvvm.annotations" metadata node. |
| 1136 | llvm::NamedMDNode *MD = M.getOrInsertNamedMetadata("nvvm.annotations"); |
| 1137 | |
| 1138 | llvm::Metadata *MDVals[] = { |
| 1139 | llvm::ConstantAsMetadata::get(Fn), llvm::MDString::get(Ctx, "kernel"), |
| 1140 | llvm::ConstantAsMetadata::get( |
| 1141 | llvm::ConstantInt::get(llvm::Type::getInt32Ty(Ctx), 1))}; |
| 1142 | // Append metadata to nvvm.annotations. |
| 1143 | MD->addOperand(llvm::MDNode::get(Ctx, MDVals)); |
| 1144 | |
| 1145 | // Add a function attribute for the kernel. |
| 1146 | Fn->addFnAttr(llvm::Attribute::get(Ctx, "kernel")); |
| 1147 | } |
| 1148 | |
| 1149 | void CGOpenMPRuntimeGPU::emitTargetOutlinedFunction( |
| 1150 | const OMPExecutableDirective &D, StringRef ParentName, |
| 1151 | llvm::Function *&OutlinedFn, llvm::Constant *&OutlinedFnID, |
| 1152 | bool IsOffloadEntry, const RegionCodeGenTy &CodeGen) { |
| 1153 | if (!IsOffloadEntry) // Nothing to do. |
| 1154 | return; |
| 1155 | |
| 1156 | assert(!ParentName.empty() && "Invalid target region parent name!")(static_cast <bool> (!ParentName.empty() && "Invalid target region parent name!" ) ? void (0) : __assert_fail ("!ParentName.empty() && \"Invalid target region parent name!\"" , "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 1156, __extension__ __PRETTY_FUNCTION__)); |
| 1157 | |
| 1158 | bool Mode = supportsSPMDExecutionMode(CGM.getContext(), D); |
| 1159 | if (Mode) |
| 1160 | emitSPMDKernel(D, ParentName, OutlinedFn, OutlinedFnID, IsOffloadEntry, |
| 1161 | CodeGen); |
| 1162 | else |
| 1163 | emitNonSPMDKernel(D, ParentName, OutlinedFn, OutlinedFnID, IsOffloadEntry, |
| 1164 | CodeGen); |
| 1165 | |
| 1166 | setPropertyExecutionMode(CGM, OutlinedFn->getName(), Mode); |
| 1167 | } |
| 1168 | |
| 1169 | namespace { |
| 1170 | LLVM_ENABLE_BITMASK_ENUMS_IN_NAMESPACE()using ::llvm::BitmaskEnumDetail::operator~; using ::llvm::BitmaskEnumDetail ::operator|; using ::llvm::BitmaskEnumDetail::operator&; using ::llvm::BitmaskEnumDetail::operator^; using ::llvm::BitmaskEnumDetail ::operator|=; using ::llvm::BitmaskEnumDetail::operator&= ; using ::llvm::BitmaskEnumDetail::operator^=; |
| 1171 | /// Enum for accesseing the reserved_2 field of the ident_t struct. |
| 1172 | enum ModeFlagsTy : unsigned { |
| 1173 | /// Bit set to 1 when in SPMD mode. |
| 1174 | KMP_IDENT_SPMD_MODE = 0x01, |
| 1175 | /// Bit set to 1 when a simplified runtime is used. |
| 1176 | KMP_IDENT_SIMPLE_RT_MODE = 0x02, |
| 1177 | LLVM_MARK_AS_BITMASK_ENUM(/*LargestValue=*/KMP_IDENT_SIMPLE_RT_MODE)LLVM_BITMASK_LARGEST_ENUMERATOR = KMP_IDENT_SIMPLE_RT_MODE |
| 1178 | }; |
| 1179 | |
| 1180 | /// Special mode Undefined. Is the combination of Non-SPMD mode + SimpleRuntime. |
| 1181 | static const ModeFlagsTy UndefinedMode = |
| 1182 | (~KMP_IDENT_SPMD_MODE) & KMP_IDENT_SIMPLE_RT_MODE; |
| 1183 | } // anonymous namespace |
| 1184 | |
| 1185 | unsigned CGOpenMPRuntimeGPU::getDefaultLocationReserved2Flags() const { |
| 1186 | switch (getExecutionMode()) { |
| 1187 | case EM_SPMD: |
| 1188 | if (requiresFullRuntime()) |
| 1189 | return KMP_IDENT_SPMD_MODE & (~KMP_IDENT_SIMPLE_RT_MODE); |
| 1190 | return KMP_IDENT_SPMD_MODE | KMP_IDENT_SIMPLE_RT_MODE; |
| 1191 | case EM_NonSPMD: |
| 1192 | assert(requiresFullRuntime() && "Expected full runtime.")(static_cast <bool> (requiresFullRuntime() && "Expected full runtime." ) ? void (0) : __assert_fail ("requiresFullRuntime() && \"Expected full runtime.\"" , "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 1192, __extension__ __PRETTY_FUNCTION__)); |
| 1193 | return (~KMP_IDENT_SPMD_MODE) & (~KMP_IDENT_SIMPLE_RT_MODE); |
| 1194 | case EM_Unknown: |
| 1195 | return UndefinedMode; |
| 1196 | } |
| 1197 | llvm_unreachable("Unknown flags are requested.")::llvm::llvm_unreachable_internal("Unknown flags are requested." , "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 1197); |
| 1198 | } |
| 1199 | |
| 1200 | CGOpenMPRuntimeGPU::CGOpenMPRuntimeGPU(CodeGenModule &CGM) |
| 1201 | : CGOpenMPRuntime(CGM, "_", "$") { |
| 1202 | if (!CGM.getLangOpts().OpenMPIsDevice) |
| 1203 | llvm_unreachable("OpenMP can only handle device code.")::llvm::llvm_unreachable_internal("OpenMP can only handle device code." , "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 1203); |
| 1204 | |
| 1205 | llvm::OpenMPIRBuilder &OMPBuilder = getOMPBuilder(); |
| 1206 | if (!CGM.getLangOpts().OMPHostIRFile.empty()) { |
| 1207 | OMPBuilder.createGlobalFlag(CGM.getLangOpts().OpenMPTargetDebug, |
| 1208 | "__omp_rtl_debug_kind"); |
| 1209 | OMPBuilder.createGlobalFlag(CGM.getLangOpts().OpenMPTeamSubscription, |
| 1210 | "__omp_rtl_assume_teams_oversubscription"); |
| 1211 | OMPBuilder.createGlobalFlag(CGM.getLangOpts().OpenMPThreadSubscription, |
| 1212 | "__omp_rtl_assume_threads_oversubscription"); |
| 1213 | OMPBuilder.createGlobalFlag(CGM.getLangOpts().OpenMPNoThreadState, |
| 1214 | "__omp_rtl_assume_no_thread_state"); |
| 1215 | } |
| 1216 | } |
| 1217 | |
| 1218 | void CGOpenMPRuntimeGPU::emitProcBindClause(CodeGenFunction &CGF, |
| 1219 | ProcBindKind ProcBind, |
| 1220 | SourceLocation Loc) { |
| 1221 | // Do nothing in case of SPMD mode and L0 parallel. |
| 1222 | if (getExecutionMode() == CGOpenMPRuntimeGPU::EM_SPMD) |
| 1223 | return; |
| 1224 | |
| 1225 | CGOpenMPRuntime::emitProcBindClause(CGF, ProcBind, Loc); |
| 1226 | } |
| 1227 | |
| 1228 | void CGOpenMPRuntimeGPU::emitNumThreadsClause(CodeGenFunction &CGF, |
| 1229 | llvm::Value *NumThreads, |
| 1230 | SourceLocation Loc) { |
| 1231 | // Nothing to do. |
| 1232 | } |
| 1233 | |
| 1234 | void CGOpenMPRuntimeGPU::emitNumTeamsClause(CodeGenFunction &CGF, |
| 1235 | const Expr *NumTeams, |
| 1236 | const Expr *ThreadLimit, |
| 1237 | SourceLocation Loc) {} |
| 1238 | |
| 1239 | llvm::Function *CGOpenMPRuntimeGPU::emitParallelOutlinedFunction( |
| 1240 | const OMPExecutableDirective &D, const VarDecl *ThreadIDVar, |
| 1241 | OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen) { |
| 1242 | // Emit target region as a standalone region. |
| 1243 | class NVPTXPrePostActionTy : public PrePostActionTy { |
| 1244 | bool &IsInParallelRegion; |
| 1245 | bool PrevIsInParallelRegion; |
| 1246 | |
| 1247 | public: |
| 1248 | NVPTXPrePostActionTy(bool &IsInParallelRegion) |
| 1249 | : IsInParallelRegion(IsInParallelRegion) {} |
| 1250 | void Enter(CodeGenFunction &CGF) override { |
| 1251 | PrevIsInParallelRegion = IsInParallelRegion; |
| 1252 | IsInParallelRegion = true; |
| 1253 | } |
| 1254 | void Exit(CodeGenFunction &CGF) override { |
| 1255 | IsInParallelRegion = PrevIsInParallelRegion; |
| 1256 | } |
| 1257 | } Action(IsInParallelRegion); |
| 1258 | CodeGen.setAction(Action); |
| 1259 | bool PrevIsInTTDRegion = IsInTTDRegion; |
| 1260 | IsInTTDRegion = false; |
| 1261 | bool PrevIsInTargetMasterThreadRegion = IsInTargetMasterThreadRegion; |
| 1262 | IsInTargetMasterThreadRegion = false; |
| 1263 | auto *OutlinedFun = |
| 1264 | cast<llvm::Function>(CGOpenMPRuntime::emitParallelOutlinedFunction( |
| 1265 | D, ThreadIDVar, InnermostKind, CodeGen)); |
| 1266 | IsInTargetMasterThreadRegion = PrevIsInTargetMasterThreadRegion; |
| 1267 | IsInTTDRegion = PrevIsInTTDRegion; |
| 1268 | if (getExecutionMode() != CGOpenMPRuntimeGPU::EM_SPMD && |
| 1269 | !IsInParallelRegion) { |
| 1270 | llvm::Function *WrapperFun = |
| 1271 | createParallelDataSharingWrapper(OutlinedFun, D); |
| 1272 | WrapperFunctionsMap[OutlinedFun] = WrapperFun; |
| 1273 | } |
| 1274 | |
| 1275 | return OutlinedFun; |
| 1276 | } |
| 1277 | |
| 1278 | /// Get list of lastprivate variables from the teams distribute ... or |
| 1279 | /// teams {distribute ...} directives. |
| 1280 | static void |
| 1281 | getDistributeLastprivateVars(ASTContext &Ctx, const OMPExecutableDirective &D, |
| 1282 | llvm::SmallVectorImpl<const ValueDecl *> &Vars) { |
| 1283 | assert(isOpenMPTeamsDirective(D.getDirectiveKind()) &&(static_cast <bool> (isOpenMPTeamsDirective(D.getDirectiveKind ()) && "expected teams directive.") ? void (0) : __assert_fail ("isOpenMPTeamsDirective(D.getDirectiveKind()) && \"expected teams directive.\"" , "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 1284, __extension__ __PRETTY_FUNCTION__)) |
| 1284 | "expected teams directive.")(static_cast <bool> (isOpenMPTeamsDirective(D.getDirectiveKind ()) && "expected teams directive.") ? void (0) : __assert_fail ("isOpenMPTeamsDirective(D.getDirectiveKind()) && \"expected teams directive.\"" , "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 1284, __extension__ __PRETTY_FUNCTION__)); |
| 1285 | const OMPExecutableDirective *Dir = &D; |
| 1286 | if (!isOpenMPDistributeDirective(D.getDirectiveKind())) { |
| 1287 | if (const Stmt *S = CGOpenMPRuntime::getSingleCompoundChild( |
| 1288 | Ctx, |
| 1289 | D.getInnermostCapturedStmt()->getCapturedStmt()->IgnoreContainers( |
| 1290 | /*IgnoreCaptured=*/true))) { |
| 1291 | Dir = dyn_cast_or_null<OMPExecutableDirective>(S); |
| 1292 | if (Dir && !isOpenMPDistributeDirective(Dir->getDirectiveKind())) |
| 1293 | Dir = nullptr; |
| 1294 | } |
| 1295 | } |
| 1296 | if (!Dir) |
| 1297 | return; |
| 1298 | for (const auto *C : Dir->getClausesOfKind<OMPLastprivateClause>()) { |
| 1299 | for (const Expr *E : C->getVarRefs()) |
| 1300 | Vars.push_back(getPrivateItem(E)); |
| 1301 | } |
| 1302 | } |
| 1303 | |
| 1304 | /// Get list of reduction variables from the teams ... directives. |
| 1305 | static void |
| 1306 | getTeamsReductionVars(ASTContext &Ctx, const OMPExecutableDirective &D, |
| 1307 | llvm::SmallVectorImpl<const ValueDecl *> &Vars) { |
| 1308 | assert(isOpenMPTeamsDirective(D.getDirectiveKind()) &&(static_cast <bool> (isOpenMPTeamsDirective(D.getDirectiveKind ()) && "expected teams directive.") ? void (0) : __assert_fail ("isOpenMPTeamsDirective(D.getDirectiveKind()) && \"expected teams directive.\"" , "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 1309, __extension__ __PRETTY_FUNCTION__)) |
| 1309 | "expected teams directive.")(static_cast <bool> (isOpenMPTeamsDirective(D.getDirectiveKind ()) && "expected teams directive.") ? void (0) : __assert_fail ("isOpenMPTeamsDirective(D.getDirectiveKind()) && \"expected teams directive.\"" , "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 1309, __extension__ __PRETTY_FUNCTION__)); |
| 1310 | for (const auto *C : D.getClausesOfKind<OMPReductionClause>()) { |
| 1311 | for (const Expr *E : C->privates()) |
| 1312 | Vars.push_back(getPrivateItem(E)); |
| 1313 | } |
| 1314 | } |
| 1315 | |
| 1316 | llvm::Function *CGOpenMPRuntimeGPU::emitTeamsOutlinedFunction( |
| 1317 | const OMPExecutableDirective &D, const VarDecl *ThreadIDVar, |
| 1318 | OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen) { |
| 1319 | SourceLocation Loc = D.getBeginLoc(); |
| 1320 | |
| 1321 | const RecordDecl *GlobalizedRD = nullptr; |
| 1322 | llvm::SmallVector<const ValueDecl *, 4> LastPrivatesReductions; |
| 1323 | llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> MappedDeclsFields; |
| 1324 | unsigned WarpSize = CGM.getTarget().getGridValue().GV_Warp_Size; |
| 1325 | // Globalize team reductions variable unconditionally in all modes. |
| 1326 | if (getExecutionMode() != CGOpenMPRuntimeGPU::EM_SPMD) |
| 1327 | getTeamsReductionVars(CGM.getContext(), D, LastPrivatesReductions); |
| 1328 | if (getExecutionMode() == CGOpenMPRuntimeGPU::EM_SPMD) { |
| 1329 | getDistributeLastprivateVars(CGM.getContext(), D, LastPrivatesReductions); |
| 1330 | if (!LastPrivatesReductions.empty()) { |
| 1331 | GlobalizedRD = ::buildRecordForGlobalizedVars( |
| 1332 | CGM.getContext(), llvm::None, LastPrivatesReductions, |
| 1333 | MappedDeclsFields, WarpSize); |
| 1334 | } |
| 1335 | } else if (!LastPrivatesReductions.empty()) { |
| 1336 | assert(!TeamAndReductions.first &&(static_cast <bool> (!TeamAndReductions.first && "Previous team declaration is not expected.") ? void (0) : __assert_fail ("!TeamAndReductions.first && \"Previous team declaration is not expected.\"" , "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 1337, __extension__ __PRETTY_FUNCTION__)) |
| 1337 | "Previous team declaration is not expected.")(static_cast <bool> (!TeamAndReductions.first && "Previous team declaration is not expected.") ? void (0) : __assert_fail ("!TeamAndReductions.first && \"Previous team declaration is not expected.\"" , "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 1337, __extension__ __PRETTY_FUNCTION__)); |
| 1338 | TeamAndReductions.first = D.getCapturedStmt(OMPD_teams)->getCapturedDecl(); |
| 1339 | std::swap(TeamAndReductions.second, LastPrivatesReductions); |
| 1340 | } |
| 1341 | |
| 1342 | // Emit target region as a standalone region. |
| 1343 | class NVPTXPrePostActionTy : public PrePostActionTy { |
| 1344 | SourceLocation &Loc; |
| 1345 | const RecordDecl *GlobalizedRD; |
| 1346 | llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> |
| 1347 | &MappedDeclsFields; |
| 1348 | |
| 1349 | public: |
| 1350 | NVPTXPrePostActionTy( |
| 1351 | SourceLocation &Loc, const RecordDecl *GlobalizedRD, |
| 1352 | llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> |
| 1353 | &MappedDeclsFields) |
| 1354 | : Loc(Loc), GlobalizedRD(GlobalizedRD), |
| 1355 | MappedDeclsFields(MappedDeclsFields) {} |
| 1356 | void Enter(CodeGenFunction &CGF) override { |
| 1357 | auto &Rt = |
| 1358 | static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()); |
| 1359 | if (GlobalizedRD) { |
| 1360 | auto I = Rt.FunctionGlobalizedDecls.try_emplace(CGF.CurFn).first; |
| 1361 | I->getSecond().MappedParams = |
| 1362 | std::make_unique<CodeGenFunction::OMPMapVars>(); |
| 1363 | DeclToAddrMapTy &Data = I->getSecond().LocalVarData; |
| 1364 | for (const auto &Pair : MappedDeclsFields) { |
| 1365 | assert(Pair.getFirst()->isCanonicalDecl() &&(static_cast <bool> (Pair.getFirst()->isCanonicalDecl () && "Expected canonical declaration") ? void (0) : __assert_fail ("Pair.getFirst()->isCanonicalDecl() && \"Expected canonical declaration\"" , "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 1366, __extension__ __PRETTY_FUNCTION__)) |
| 1366 | "Expected canonical declaration")(static_cast <bool> (Pair.getFirst()->isCanonicalDecl () && "Expected canonical declaration") ? void (0) : __assert_fail ("Pair.getFirst()->isCanonicalDecl() && \"Expected canonical declaration\"" , "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 1366, __extension__ __PRETTY_FUNCTION__)); |
| 1367 | Data.insert(std::make_pair(Pair.getFirst(), MappedVarData())); |
| 1368 | } |
| 1369 | } |
| 1370 | Rt.emitGenericVarsProlog(CGF, Loc); |
| 1371 | } |
| 1372 | void Exit(CodeGenFunction &CGF) override { |
| 1373 | static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()) |
| 1374 | .emitGenericVarsEpilog(CGF); |
| 1375 | } |
| 1376 | } Action(Loc, GlobalizedRD, MappedDeclsFields); |
| 1377 | CodeGen.setAction(Action); |
| 1378 | llvm::Function *OutlinedFun = CGOpenMPRuntime::emitTeamsOutlinedFunction( |
| 1379 | D, ThreadIDVar, InnermostKind, CodeGen); |
| 1380 | |
| 1381 | return OutlinedFun; |
| 1382 | } |
| 1383 | |
| 1384 | void CGOpenMPRuntimeGPU::emitGenericVarsProlog(CodeGenFunction &CGF, |
| 1385 | SourceLocation Loc, |
| 1386 | bool WithSPMDCheck) { |
| 1387 | if (getDataSharingMode(CGM) != CGOpenMPRuntimeGPU::Generic && |
| 1388 | getExecutionMode() != CGOpenMPRuntimeGPU::EM_SPMD) |
| 1389 | return; |
| 1390 | |
| 1391 | CGBuilderTy &Bld = CGF.Builder; |
| 1392 | |
| 1393 | const auto I = FunctionGlobalizedDecls.find(CGF.CurFn); |
| 1394 | if (I == FunctionGlobalizedDecls.end()) |
| 1395 | return; |
| 1396 | |
| 1397 | for (auto &Rec : I->getSecond().LocalVarData) { |
| 1398 | const auto *VD = cast<VarDecl>(Rec.first); |
| 1399 | bool EscapedParam = I->getSecond().EscapedParameters.count(Rec.first); |
| 1400 | QualType VarTy = VD->getType(); |
| 1401 | |
| 1402 | // Get the local allocation of a firstprivate variable before sharing |
| 1403 | llvm::Value *ParValue; |
| 1404 | if (EscapedParam) { |
| 1405 | LValue ParLVal = |
| 1406 | CGF.MakeAddrLValue(CGF.GetAddrOfLocalVar(VD), VD->getType()); |
| 1407 | ParValue = CGF.EmitLoadOfScalar(ParLVal, Loc); |
| 1408 | } |
| 1409 | |
| 1410 | // Allocate space for the variable to be globalized |
| 1411 | llvm::Value *AllocArgs[] = {CGF.getTypeSize(VD->getType())}; |
| 1412 | llvm::CallBase *VoidPtr = |
| 1413 | CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( |
| 1414 | CGM.getModule(), OMPRTL___kmpc_alloc_shared), |
| 1415 | AllocArgs, VD->getName()); |
| 1416 | // FIXME: We should use the variables actual alignment as an argument. |
| 1417 | VoidPtr->addRetAttr(llvm::Attribute::get( |
| 1418 | CGM.getLLVMContext(), llvm::Attribute::Alignment, |
| 1419 | CGM.getContext().getTargetInfo().getNewAlign() / 8)); |
| 1420 | |
| 1421 | // Cast the void pointer and get the address of the globalized variable. |
| 1422 | llvm::PointerType *VarPtrTy = CGF.ConvertTypeForMem(VarTy)->getPointerTo(); |
| 1423 | llvm::Value *CastedVoidPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( |
| 1424 | VoidPtr, VarPtrTy, VD->getName() + "_on_stack"); |
| 1425 | LValue VarAddr = CGF.MakeNaturalAlignAddrLValue(CastedVoidPtr, VarTy); |
| 1426 | Rec.second.PrivateAddr = VarAddr.getAddress(CGF); |
| 1427 | Rec.second.GlobalizedVal = VoidPtr; |
| 1428 | |
| 1429 | // Assign the local allocation to the newly globalized location. |
| 1430 | if (EscapedParam) { |
| 1431 | CGF.EmitStoreOfScalar(ParValue, VarAddr); |
| 1432 | I->getSecond().MappedParams->setVarAddr(CGF, VD, VarAddr.getAddress(CGF)); |
| 1433 | } |
| 1434 | if (auto *DI = CGF.getDebugInfo()) |
| 1435 | VoidPtr->setDebugLoc(DI->SourceLocToDebugLoc(VD->getLocation())); |
| 1436 | } |
| 1437 | for (const auto *VD : I->getSecond().EscapedVariableLengthDecls) { |
| 1438 | // Use actual memory size of the VLA object including the padding |
| 1439 | // for alignment purposes. |
| 1440 | llvm::Value *Size = CGF.getTypeSize(VD->getType()); |
| 1441 | CharUnits Align = CGM.getContext().getDeclAlign(VD); |
| 1442 | Size = Bld.CreateNUWAdd( |
| 1443 | Size, llvm::ConstantInt::get(CGF.SizeTy, Align.getQuantity() - 1)); |
| 1444 | llvm::Value *AlignVal = |
| 1445 | llvm::ConstantInt::get(CGF.SizeTy, Align.getQuantity()); |
| 1446 | |
| 1447 | Size = Bld.CreateUDiv(Size, AlignVal); |
| 1448 | Size = Bld.CreateNUWMul(Size, AlignVal); |
Value stored to 'Size' is never read | |
| 1449 | |
| 1450 | // Allocate space for this VLA object to be globalized. |
| 1451 | llvm::Value *AllocArgs[] = {CGF.getTypeSize(VD->getType())}; |
| 1452 | llvm::CallBase *VoidPtr = |
| 1453 | CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( |
| 1454 | CGM.getModule(), OMPRTL___kmpc_alloc_shared), |
| 1455 | AllocArgs, VD->getName()); |
| 1456 | VoidPtr->addRetAttr( |
| 1457 | llvm::Attribute::get(CGM.getLLVMContext(), llvm::Attribute::Alignment, |
| 1458 | CGM.getContext().getTargetInfo().getNewAlign())); |
| 1459 | |
| 1460 | I->getSecond().EscapedVariableLengthDeclsAddrs.emplace_back( |
| 1461 | std::pair<llvm::Value *, llvm::Value *>( |
| 1462 | {VoidPtr, CGF.getTypeSize(VD->getType())})); |
| 1463 | LValue Base = CGF.MakeAddrLValue(VoidPtr, VD->getType(), |
| 1464 | CGM.getContext().getDeclAlign(VD), |
| 1465 | AlignmentSource::Decl); |
| 1466 | I->getSecond().MappedParams->setVarAddr(CGF, cast<VarDecl>(VD), |
| 1467 | Base.getAddress(CGF)); |
| 1468 | } |
| 1469 | I->getSecond().MappedParams->apply(CGF); |
| 1470 | } |
| 1471 | |
| 1472 | void CGOpenMPRuntimeGPU::emitGenericVarsEpilog(CodeGenFunction &CGF, |
| 1473 | bool WithSPMDCheck) { |
| 1474 | if (getDataSharingMode(CGM) != CGOpenMPRuntimeGPU::Generic && |
| 1475 | getExecutionMode() != CGOpenMPRuntimeGPU::EM_SPMD) |
| 1476 | return; |
| 1477 | |
| 1478 | const auto I = FunctionGlobalizedDecls.find(CGF.CurFn); |
| 1479 | if (I != FunctionGlobalizedDecls.end()) { |
| 1480 | // Deallocate the memory for each globalized VLA object |
| 1481 | for (auto AddrSizePair : |
| 1482 | llvm::reverse(I->getSecond().EscapedVariableLengthDeclsAddrs)) { |
| 1483 | CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( |
| 1484 | CGM.getModule(), OMPRTL___kmpc_free_shared), |
| 1485 | {AddrSizePair.first, AddrSizePair.second}); |
| 1486 | } |
| 1487 | // Deallocate the memory for each globalized value |
| 1488 | for (auto &Rec : llvm::reverse(I->getSecond().LocalVarData)) { |
| 1489 | const auto *VD = cast<VarDecl>(Rec.first); |
| 1490 | I->getSecond().MappedParams->restore(CGF); |
| 1491 | |
| 1492 | llvm::Value *FreeArgs[] = {Rec.second.GlobalizedVal, |
| 1493 | CGF.getTypeSize(VD->getType())}; |
| 1494 | CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( |
| 1495 | CGM.getModule(), OMPRTL___kmpc_free_shared), |
| 1496 | FreeArgs); |
| 1497 | } |
| 1498 | } |
| 1499 | } |
| 1500 | |
| 1501 | void CGOpenMPRuntimeGPU::emitTeamsCall(CodeGenFunction &CGF, |
| 1502 | const OMPExecutableDirective &D, |
| 1503 | SourceLocation Loc, |
| 1504 | llvm::Function *OutlinedFn, |
| 1505 | ArrayRef<llvm::Value *> CapturedVars) { |
| 1506 | if (!CGF.HaveInsertPoint()) |
| 1507 | return; |
| 1508 | |
| 1509 | Address ZeroAddr = CGF.CreateDefaultAlignTempAlloca(CGF.Int32Ty, |
| 1510 | /*Name=*/".zero.addr"); |
| 1511 | CGF.Builder.CreateStore(CGF.Builder.getInt32(/*C*/ 0), ZeroAddr); |
| 1512 | llvm::SmallVector<llvm::Value *, 16> OutlinedFnArgs; |
| 1513 | OutlinedFnArgs.push_back(emitThreadIDAddress(CGF, Loc).getPointer()); |
| 1514 | OutlinedFnArgs.push_back(ZeroAddr.getPointer()); |
| 1515 | OutlinedFnArgs.append(CapturedVars.begin(), CapturedVars.end()); |
| 1516 | emitOutlinedFunctionCall(CGF, Loc, OutlinedFn, OutlinedFnArgs); |
| 1517 | } |
| 1518 | |
| 1519 | void CGOpenMPRuntimeGPU::emitParallelCall(CodeGenFunction &CGF, |
| 1520 | SourceLocation Loc, |
| 1521 | llvm::Function *OutlinedFn, |
| 1522 | ArrayRef<llvm::Value *> CapturedVars, |
| 1523 | const Expr *IfCond, |
| 1524 | llvm::Value *NumThreads) { |
| 1525 | if (!CGF.HaveInsertPoint()) |
| 1526 | return; |
| 1527 | |
| 1528 | auto &&ParallelGen = [this, Loc, OutlinedFn, CapturedVars, IfCond, |
| 1529 | NumThreads](CodeGenFunction &CGF, |
| 1530 | PrePostActionTy &Action) { |
| 1531 | CGBuilderTy &Bld = CGF.Builder; |
| 1532 | llvm::Value *NumThreadsVal = NumThreads; |
| 1533 | llvm::Function *WFn = WrapperFunctionsMap[OutlinedFn]; |
| 1534 | llvm::Value *ID = llvm::ConstantPointerNull::get(CGM.Int8PtrTy); |
| 1535 | if (WFn) |
| 1536 | ID = Bld.CreateBitOrPointerCast(WFn, CGM.Int8PtrTy); |
| 1537 | llvm::Value *FnPtr = Bld.CreateBitOrPointerCast(OutlinedFn, CGM.Int8PtrTy); |
| 1538 | |
| 1539 | // Create a private scope that will globalize the arguments |
| 1540 | // passed from the outside of the target region. |
| 1541 | // TODO: Is that needed? |
| 1542 | CodeGenFunction::OMPPrivateScope PrivateArgScope(CGF); |
| 1543 | |
| 1544 | Address CapturedVarsAddrs = CGF.CreateDefaultAlignTempAlloca( |
| 1545 | llvm::ArrayType::get(CGM.VoidPtrTy, CapturedVars.size()), |
| 1546 | "captured_vars_addrs"); |
| 1547 | // There's something to share. |
| 1548 | if (!CapturedVars.empty()) { |
| 1549 | // Prepare for parallel region. Indicate the outlined function. |
| 1550 | ASTContext &Ctx = CGF.getContext(); |
| 1551 | unsigned Idx = 0; |
| 1552 | for (llvm::Value *V : CapturedVars) { |
| 1553 | Address Dst = Bld.CreateConstArrayGEP(CapturedVarsAddrs, Idx); |
| 1554 | llvm::Value *PtrV; |
| 1555 | if (V->getType()->isIntegerTy()) |
| 1556 | PtrV = Bld.CreateIntToPtr(V, CGF.VoidPtrTy); |
| 1557 | else |
| 1558 | PtrV = Bld.CreatePointerBitCastOrAddrSpaceCast(V, CGF.VoidPtrTy); |
| 1559 | CGF.EmitStoreOfScalar(PtrV, Dst, /*Volatile=*/false, |
| 1560 | Ctx.getPointerType(Ctx.VoidPtrTy)); |
| 1561 | ++Idx; |
| 1562 | } |
| 1563 | } |
| 1564 | |
| 1565 | llvm::Value *IfCondVal = nullptr; |
| 1566 | if (IfCond) |
| 1567 | IfCondVal = Bld.CreateIntCast(CGF.EvaluateExprAsBool(IfCond), CGF.Int32Ty, |
| 1568 | /* isSigned */ false); |
| 1569 | else |
| 1570 | IfCondVal = llvm::ConstantInt::get(CGF.Int32Ty, 1); |
| 1571 | |
| 1572 | if (!NumThreadsVal) |
| 1573 | NumThreadsVal = llvm::ConstantInt::get(CGF.Int32Ty, -1); |
| 1574 | else |
| 1575 | NumThreadsVal = Bld.CreateZExtOrTrunc(NumThreadsVal, CGF.Int32Ty), |
| 1576 | |
| 1577 | assert(IfCondVal && "Expected a value")(static_cast <bool> (IfCondVal && "Expected a value" ) ? void (0) : __assert_fail ("IfCondVal && \"Expected a value\"" , "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 1577, __extension__ __PRETTY_FUNCTION__)); |
| 1578 | llvm::Value *RTLoc = emitUpdateLocation(CGF, Loc); |
| 1579 | llvm::Value *Args[] = { |
| 1580 | RTLoc, |
| 1581 | getThreadID(CGF, Loc), |
| 1582 | IfCondVal, |
| 1583 | NumThreadsVal, |
| 1584 | llvm::ConstantInt::get(CGF.Int32Ty, -1), |
| 1585 | FnPtr, |
| 1586 | ID, |
| 1587 | Bld.CreateBitOrPointerCast(CapturedVarsAddrs.getPointer(), |
| 1588 | CGF.VoidPtrPtrTy), |
| 1589 | llvm::ConstantInt::get(CGM.SizeTy, CapturedVars.size())}; |
| 1590 | CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( |
| 1591 | CGM.getModule(), OMPRTL___kmpc_parallel_51), |
| 1592 | Args); |
| 1593 | }; |
| 1594 | |
| 1595 | RegionCodeGenTy RCG(ParallelGen); |
| 1596 | RCG(CGF); |
| 1597 | } |
| 1598 | |
| 1599 | void CGOpenMPRuntimeGPU::syncCTAThreads(CodeGenFunction &CGF) { |
| 1600 | // Always emit simple barriers! |
| 1601 | if (!CGF.HaveInsertPoint()) |
| 1602 | return; |
| 1603 | // Build call __kmpc_barrier_simple_spmd(nullptr, 0); |
| 1604 | // This function does not use parameters, so we can emit just default values. |
| 1605 | llvm::Value *Args[] = { |
| 1606 | llvm::ConstantPointerNull::get( |
| 1607 | cast<llvm::PointerType>(getIdentTyPointerTy())), |
| 1608 | llvm::ConstantInt::get(CGF.Int32Ty, /*V=*/0, /*isSigned=*/true)}; |
| 1609 | CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( |
| 1610 | CGM.getModule(), OMPRTL___kmpc_barrier_simple_spmd), |
| 1611 | Args); |
| 1612 | } |
| 1613 | |
| 1614 | void CGOpenMPRuntimeGPU::emitBarrierCall(CodeGenFunction &CGF, |
| 1615 | SourceLocation Loc, |
| 1616 | OpenMPDirectiveKind Kind, bool, |
| 1617 | bool) { |
| 1618 | // Always emit simple barriers! |
| 1619 | if (!CGF.HaveInsertPoint()) |
| 1620 | return; |
| 1621 | // Build call __kmpc_cancel_barrier(loc, thread_id); |
| 1622 | unsigned Flags = getDefaultFlagsForBarriers(Kind); |
| 1623 | llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc, Flags), |
| 1624 | getThreadID(CGF, Loc)}; |
| 1625 | |
| 1626 | CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( |
| 1627 | CGM.getModule(), OMPRTL___kmpc_barrier), |
| 1628 | Args); |
| 1629 | } |
| 1630 | |
| 1631 | void CGOpenMPRuntimeGPU::emitCriticalRegion( |
| 1632 | CodeGenFunction &CGF, StringRef CriticalName, |
| 1633 | const RegionCodeGenTy &CriticalOpGen, SourceLocation Loc, |
| 1634 | const Expr *Hint) { |
| 1635 | llvm::BasicBlock *LoopBB = CGF.createBasicBlock("omp.critical.loop"); |
| 1636 | llvm::BasicBlock *TestBB = CGF.createBasicBlock("omp.critical.test"); |
| 1637 | llvm::BasicBlock *SyncBB = CGF.createBasicBlock("omp.critical.sync"); |
| 1638 | llvm::BasicBlock *BodyBB = CGF.createBasicBlock("omp.critical.body"); |
| 1639 | llvm::BasicBlock *ExitBB = CGF.createBasicBlock("omp.critical.exit"); |
| 1640 | |
| 1641 | auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()); |
| 1642 | |
| 1643 | // Get the mask of active threads in the warp. |
| 1644 | llvm::Value *Mask = CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( |
| 1645 | CGM.getModule(), OMPRTL___kmpc_warp_active_thread_mask)); |
| 1646 | // Fetch team-local id of the thread. |
| 1647 | llvm::Value *ThreadID = RT.getGPUThreadID(CGF); |
| 1648 | |
| 1649 | // Get the width of the team. |
| 1650 | llvm::Value *TeamWidth = RT.getGPUNumThreads(CGF); |
| 1651 | |
| 1652 | // Initialize the counter variable for the loop. |
| 1653 | QualType Int32Ty = |
| 1654 | CGF.getContext().getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/0); |
| 1655 | Address Counter = CGF.CreateMemTemp(Int32Ty, "critical_counter"); |
| 1656 | LValue CounterLVal = CGF.MakeAddrLValue(Counter, Int32Ty); |
| 1657 | CGF.EmitStoreOfScalar(llvm::Constant::getNullValue(CGM.Int32Ty), CounterLVal, |
| 1658 | /*isInit=*/true); |
| 1659 | |
| 1660 | // Block checks if loop counter exceeds upper bound. |
| 1661 | CGF.EmitBlock(LoopBB); |
| 1662 | llvm::Value *CounterVal = CGF.EmitLoadOfScalar(CounterLVal, Loc); |
| 1663 | llvm::Value *CmpLoopBound = CGF.Builder.CreateICmpSLT(CounterVal, TeamWidth); |
| 1664 | CGF.Builder.CreateCondBr(CmpLoopBound, TestBB, ExitBB); |
| 1665 | |
| 1666 | // Block tests which single thread should execute region, and which threads |
| 1667 | // should go straight to synchronisation point. |
| 1668 | CGF.EmitBlock(TestBB); |
| 1669 | CounterVal = CGF.EmitLoadOfScalar(CounterLVal, Loc); |
| 1670 | llvm::Value *CmpThreadToCounter = |
| 1671 | CGF.Builder.CreateICmpEQ(ThreadID, CounterVal); |
| 1672 | CGF.Builder.CreateCondBr(CmpThreadToCounter, BodyBB, SyncBB); |
| 1673 | |
| 1674 | // Block emits the body of the critical region. |
| 1675 | CGF.EmitBlock(BodyBB); |
| 1676 | |
| 1677 | // Output the critical statement. |
| 1678 | CGOpenMPRuntime::emitCriticalRegion(CGF, CriticalName, CriticalOpGen, Loc, |
| 1679 | Hint); |
| 1680 | |
| 1681 | // After the body surrounded by the critical region, the single executing |
| 1682 | // thread will jump to the synchronisation point. |
| 1683 | // Block waits for all threads in current team to finish then increments the |
| 1684 | // counter variable and returns to the loop. |
| 1685 | CGF.EmitBlock(SyncBB); |
| 1686 | // Reconverge active threads in the warp. |
| 1687 | (void)CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( |
| 1688 | CGM.getModule(), OMPRTL___kmpc_syncwarp), |
| 1689 | Mask); |
| 1690 | |
| 1691 | llvm::Value *IncCounterVal = |
| 1692 | CGF.Builder.CreateNSWAdd(CounterVal, CGF.Builder.getInt32(1)); |
| 1693 | CGF.EmitStoreOfScalar(IncCounterVal, CounterLVal); |
| 1694 | CGF.EmitBranch(LoopBB); |
| 1695 | |
| 1696 | // Block that is reached when all threads in the team complete the region. |
| 1697 | CGF.EmitBlock(ExitBB, /*IsFinished=*/true); |
| 1698 | } |
| 1699 | |
| 1700 | /// Cast value to the specified type. |
| 1701 | static llvm::Value *castValueToType(CodeGenFunction &CGF, llvm::Value *Val, |
| 1702 | QualType ValTy, QualType CastTy, |
| 1703 | SourceLocation Loc) { |
| 1704 | assert(!CGF.getContext().getTypeSizeInChars(CastTy).isZero() &&(static_cast <bool> (!CGF.getContext().getTypeSizeInChars (CastTy).isZero() && "Cast type must sized.") ? void ( 0) : __assert_fail ("!CGF.getContext().getTypeSizeInChars(CastTy).isZero() && \"Cast type must sized.\"" , "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 1705, __extension__ __PRETTY_FUNCTION__)) |
| 1705 | "Cast type must sized.")(static_cast <bool> (!CGF.getContext().getTypeSizeInChars (CastTy).isZero() && "Cast type must sized.") ? void ( 0) : __assert_fail ("!CGF.getContext().getTypeSizeInChars(CastTy).isZero() && \"Cast type must sized.\"" , "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 1705, __extension__ __PRETTY_FUNCTION__)); |
| 1706 | assert(!CGF.getContext().getTypeSizeInChars(ValTy).isZero() &&(static_cast <bool> (!CGF.getContext().getTypeSizeInChars (ValTy).isZero() && "Val type must sized.") ? void (0 ) : __assert_fail ("!CGF.getContext().getTypeSizeInChars(ValTy).isZero() && \"Val type must sized.\"" , "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 1707, __extension__ __PRETTY_FUNCTION__)) |
| 1707 | "Val type must sized.")(static_cast <bool> (!CGF.getContext().getTypeSizeInChars (ValTy).isZero() && "Val type must sized.") ? void (0 ) : __assert_fail ("!CGF.getContext().getTypeSizeInChars(ValTy).isZero() && \"Val type must sized.\"" , "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 1707, __extension__ __PRETTY_FUNCTION__)); |
| 1708 | llvm::Type *LLVMCastTy = CGF.ConvertTypeForMem(CastTy); |
| 1709 | if (ValTy == CastTy) |
| 1710 | return Val; |
| 1711 | if (CGF.getContext().getTypeSizeInChars(ValTy) == |
| 1712 | CGF.getContext().getTypeSizeInChars(CastTy)) |
| 1713 | return CGF.Builder.CreateBitCast(Val, LLVMCastTy); |
| 1714 | if (CastTy->isIntegerType() && ValTy->isIntegerType()) |
| 1715 | return CGF.Builder.CreateIntCast(Val, LLVMCastTy, |
| 1716 | CastTy->hasSignedIntegerRepresentation()); |
| 1717 | Address CastItem = CGF.CreateMemTemp(CastTy); |
| 1718 | Address ValCastItem = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( |
| 1719 | CastItem, Val->getType()->getPointerTo(CastItem.getAddressSpace()), |
| 1720 | Val->getType()); |
| 1721 | CGF.EmitStoreOfScalar(Val, ValCastItem, /*Volatile=*/false, ValTy, |
| 1722 | LValueBaseInfo(AlignmentSource::Type), |
| 1723 | TBAAAccessInfo()); |
| 1724 | return CGF.EmitLoadOfScalar(CastItem, /*Volatile=*/false, CastTy, Loc, |
| 1725 | LValueBaseInfo(AlignmentSource::Type), |
| 1726 | TBAAAccessInfo()); |
| 1727 | } |
| 1728 | |
| 1729 | /// This function creates calls to one of two shuffle functions to copy |
| 1730 | /// variables between lanes in a warp. |
| 1731 | static llvm::Value *createRuntimeShuffleFunction(CodeGenFunction &CGF, |
| 1732 | llvm::Value *Elem, |
| 1733 | QualType ElemType, |
| 1734 | llvm::Value *Offset, |
| 1735 | SourceLocation Loc) { |
| 1736 | CodeGenModule &CGM = CGF.CGM; |
| 1737 | CGBuilderTy &Bld = CGF.Builder; |
| 1738 | CGOpenMPRuntimeGPU &RT = |
| 1739 | *(static_cast<CGOpenMPRuntimeGPU *>(&CGM.getOpenMPRuntime())); |
| 1740 | llvm::OpenMPIRBuilder &OMPBuilder = RT.getOMPBuilder(); |
| 1741 | |
| 1742 | CharUnits Size = CGF.getContext().getTypeSizeInChars(ElemType); |
| 1743 | assert(Size.getQuantity() <= 8 &&(static_cast <bool> (Size.getQuantity() <= 8 && "Unsupported bitwidth in shuffle instruction.") ? void (0) : __assert_fail ("Size.getQuantity() <= 8 && \"Unsupported bitwidth in shuffle instruction.\"" , "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 1744, __extension__ __PRETTY_FUNCTION__)) |
| 1744 | "Unsupported bitwidth in shuffle instruction.")(static_cast <bool> (Size.getQuantity() <= 8 && "Unsupported bitwidth in shuffle instruction.") ? void (0) : __assert_fail ("Size.getQuantity() <= 8 && \"Unsupported bitwidth in shuffle instruction.\"" , "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 1744, __extension__ __PRETTY_FUNCTION__)); |
| 1745 | |
| 1746 | RuntimeFunction ShuffleFn = Size.getQuantity() <= 4 |
| 1747 | ? OMPRTL___kmpc_shuffle_int32 |
| 1748 | : OMPRTL___kmpc_shuffle_int64; |
| 1749 | |
| 1750 | // Cast all types to 32- or 64-bit values before calling shuffle routines. |
| 1751 | QualType CastTy = CGF.getContext().getIntTypeForBitwidth( |
| 1752 | Size.getQuantity() <= 4 ? 32 : 64, /*Signed=*/1); |
| 1753 | llvm::Value *ElemCast = castValueToType(CGF, Elem, ElemType, CastTy, Loc); |
| 1754 | llvm::Value *WarpSize = |
| 1755 | Bld.CreateIntCast(RT.getGPUWarpSize(CGF), CGM.Int16Ty, /*isSigned=*/true); |
| 1756 | |
| 1757 | llvm::Value *ShuffledVal = CGF.EmitRuntimeCall( |
| 1758 | OMPBuilder.getOrCreateRuntimeFunction(CGM.getModule(), ShuffleFn), |
| 1759 | {ElemCast, Offset, WarpSize}); |
| 1760 | |
| 1761 | return castValueToType(CGF, ShuffledVal, CastTy, ElemType, Loc); |
| 1762 | } |
| 1763 | |
| 1764 | static void shuffleAndStore(CodeGenFunction &CGF, Address SrcAddr, |
| 1765 | Address DestAddr, QualType ElemType, |
| 1766 | llvm::Value *Offset, SourceLocation Loc) { |
| 1767 | CGBuilderTy &Bld = CGF.Builder; |
| 1768 | |
| 1769 | CharUnits Size = CGF.getContext().getTypeSizeInChars(ElemType); |
| 1770 | // Create the loop over the big sized data. |
| 1771 | // ptr = (void*)Elem; |
| 1772 | // ptrEnd = (void*) Elem + 1; |
| 1773 | // Step = 8; |
| 1774 | // while (ptr + Step < ptrEnd) |
| 1775 | // shuffle((int64_t)*ptr); |
| 1776 | // Step = 4; |
| 1777 | // while (ptr + Step < ptrEnd) |
| 1778 | // shuffle((int32_t)*ptr); |
| 1779 | // ... |
| 1780 | Address ElemPtr = DestAddr; |
| 1781 | Address Ptr = SrcAddr; |
| 1782 | Address PtrEnd = Bld.CreatePointerBitCastOrAddrSpaceCast( |
| 1783 | Bld.CreateConstGEP(SrcAddr, 1), CGF.VoidPtrTy, CGF.Int8Ty); |
| 1784 | for (int IntSize = 8; IntSize >= 1; IntSize /= 2) { |
| 1785 | if (Size < CharUnits::fromQuantity(IntSize)) |
| 1786 | continue; |
| 1787 | QualType IntType = CGF.getContext().getIntTypeForBitwidth( |
| 1788 | CGF.getContext().toBits(CharUnits::fromQuantity(IntSize)), |
| 1789 | /*Signed=*/1); |
| 1790 | llvm::Type *IntTy = CGF.ConvertTypeForMem(IntType); |
| 1791 | Ptr = Bld.CreatePointerBitCastOrAddrSpaceCast(Ptr, IntTy->getPointerTo(), |
| 1792 | IntTy); |
| 1793 | ElemPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( |
| 1794 | ElemPtr, IntTy->getPointerTo(), IntTy); |
| 1795 | if (Size.getQuantity() / IntSize > 1) { |
| 1796 | llvm::BasicBlock *PreCondBB = CGF.createBasicBlock(".shuffle.pre_cond"); |
| 1797 | llvm::BasicBlock *ThenBB = CGF.createBasicBlock(".shuffle.then"); |
| 1798 | llvm::BasicBlock *ExitBB = CGF.createBasicBlock(".shuffle.exit"); |
| 1799 | llvm::BasicBlock *CurrentBB = Bld.GetInsertBlock(); |
| 1800 | CGF.EmitBlock(PreCondBB); |
| 1801 | llvm::PHINode *PhiSrc = |
| 1802 | Bld.CreatePHI(Ptr.getType(), /*NumReservedValues=*/2); |
| 1803 | PhiSrc->addIncoming(Ptr.getPointer(), CurrentBB); |
| 1804 | llvm::PHINode *PhiDest = |
| 1805 | Bld.CreatePHI(ElemPtr.getType(), /*NumReservedValues=*/2); |
| 1806 | PhiDest->addIncoming(ElemPtr.getPointer(), CurrentBB); |
| 1807 | Ptr = Address(PhiSrc, Ptr.getElementType(), Ptr.getAlignment()); |
| 1808 | ElemPtr = |
| 1809 | Address(PhiDest, ElemPtr.getElementType(), ElemPtr.getAlignment()); |
| 1810 | llvm::Value *PtrDiff = Bld.CreatePtrDiff( |
| 1811 | CGF.Int8Ty, PtrEnd.getPointer(), |
| 1812 | Bld.CreatePointerBitCastOrAddrSpaceCast(Ptr.getPointer(), |
| 1813 | CGF.VoidPtrTy)); |
| 1814 | Bld.CreateCondBr(Bld.CreateICmpSGT(PtrDiff, Bld.getInt64(IntSize - 1)), |
| 1815 | ThenBB, ExitBB); |
| 1816 | CGF.EmitBlock(ThenBB); |
| 1817 | llvm::Value *Res = createRuntimeShuffleFunction( |
| 1818 | CGF, |
| 1819 | CGF.EmitLoadOfScalar(Ptr, /*Volatile=*/false, IntType, Loc, |
| 1820 | LValueBaseInfo(AlignmentSource::Type), |
| 1821 | TBAAAccessInfo()), |
| 1822 | IntType, Offset, Loc); |
| 1823 | CGF.EmitStoreOfScalar(Res, ElemPtr, /*Volatile=*/false, IntType, |
| 1824 | LValueBaseInfo(AlignmentSource::Type), |
| 1825 | TBAAAccessInfo()); |
| 1826 | Address LocalPtr = Bld.CreateConstGEP(Ptr, 1); |
| 1827 | Address LocalElemPtr = Bld.CreateConstGEP(ElemPtr, 1); |
| 1828 | PhiSrc->addIncoming(LocalPtr.getPointer(), ThenBB); |
| 1829 | PhiDest->addIncoming(LocalElemPtr.getPointer(), ThenBB); |
| 1830 | CGF.EmitBranch(PreCondBB); |
| 1831 | CGF.EmitBlock(ExitBB); |
| 1832 | } else { |
| 1833 | llvm::Value *Res = createRuntimeShuffleFunction( |
| 1834 | CGF, |
| 1835 | CGF.EmitLoadOfScalar(Ptr, /*Volatile=*/false, IntType, Loc, |
| 1836 | LValueBaseInfo(AlignmentSource::Type), |
| 1837 | TBAAAccessInfo()), |
| 1838 | IntType, Offset, Loc); |
| 1839 | CGF.EmitStoreOfScalar(Res, ElemPtr, /*Volatile=*/false, IntType, |
| 1840 | LValueBaseInfo(AlignmentSource::Type), |
| 1841 | TBAAAccessInfo()); |
| 1842 | Ptr = Bld.CreateConstGEP(Ptr, 1); |
| 1843 | ElemPtr = Bld.CreateConstGEP(ElemPtr, 1); |
| 1844 | } |
| 1845 | Size = Size % IntSize; |
| 1846 | } |
| 1847 | } |
| 1848 | |
| 1849 | namespace { |
| 1850 | enum CopyAction : unsigned { |
| 1851 | // RemoteLaneToThread: Copy over a Reduce list from a remote lane in |
| 1852 | // the warp using shuffle instructions. |
| 1853 | RemoteLaneToThread, |
| 1854 | // ThreadCopy: Make a copy of a Reduce list on the thread's stack. |
| 1855 | ThreadCopy, |
| 1856 | // ThreadToScratchpad: Copy a team-reduced array to the scratchpad. |
| 1857 | ThreadToScratchpad, |
| 1858 | // ScratchpadToThread: Copy from a scratchpad array in global memory |
| 1859 | // containing team-reduced data to a thread's stack. |
| 1860 | ScratchpadToThread, |
| 1861 | }; |
| 1862 | } // namespace |
| 1863 | |
| 1864 | struct CopyOptionsTy { |
| 1865 | llvm::Value *RemoteLaneOffset; |
| 1866 | llvm::Value *ScratchpadIndex; |
| 1867 | llvm::Value *ScratchpadWidth; |
| 1868 | }; |
| 1869 | |
| 1870 | /// Emit instructions to copy a Reduce list, which contains partially |
| 1871 | /// aggregated values, in the specified direction. |
| 1872 | static void emitReductionListCopy( |
| 1873 | CopyAction Action, CodeGenFunction &CGF, QualType ReductionArrayTy, |
| 1874 | ArrayRef<const Expr *> Privates, Address SrcBase, Address DestBase, |
| 1875 | CopyOptionsTy CopyOptions = {nullptr, nullptr, nullptr}) { |
| 1876 | |
| 1877 | CodeGenModule &CGM = CGF.CGM; |
| 1878 | ASTContext &C = CGM.getContext(); |
| 1879 | CGBuilderTy &Bld = CGF.Builder; |
| 1880 | |
| 1881 | llvm::Value *RemoteLaneOffset = CopyOptions.RemoteLaneOffset; |
| 1882 | llvm::Value *ScratchpadIndex = CopyOptions.ScratchpadIndex; |
| 1883 | llvm::Value *ScratchpadWidth = CopyOptions.ScratchpadWidth; |
| 1884 | |
| 1885 | // Iterates, element-by-element, through the source Reduce list and |
| 1886 | // make a copy. |
| 1887 | unsigned Idx = 0; |
| 1888 | unsigned Size = Privates.size(); |
| 1889 | for (const Expr *Private : Privates) { |
| 1890 | Address SrcElementAddr = Address::invalid(); |
| 1891 | Address DestElementAddr = Address::invalid(); |
| 1892 | Address DestElementPtrAddr = Address::invalid(); |
| 1893 | // Should we shuffle in an element from a remote lane? |
| 1894 | bool ShuffleInElement = false; |
| 1895 | // Set to true to update the pointer in the dest Reduce list to a |
| 1896 | // newly created element. |
| 1897 | bool UpdateDestListPtr = false; |
| 1898 | // Increment the src or dest pointer to the scratchpad, for each |
| 1899 | // new element. |
| 1900 | bool IncrScratchpadSrc = false; |
| 1901 | bool IncrScratchpadDest = false; |
| 1902 | QualType PrivatePtrType = C.getPointerType(Private->getType()); |
| 1903 | llvm::Type *PrivateLlvmPtrType = CGF.ConvertType(PrivatePtrType); |
| 1904 | |
| 1905 | switch (Action) { |
| 1906 | case RemoteLaneToThread: { |
| 1907 | // Step 1.1: Get the address for the src element in the Reduce list. |
| 1908 | Address SrcElementPtrAddr = Bld.CreateConstArrayGEP(SrcBase, Idx); |
| 1909 | SrcElementAddr = |
| 1910 | CGF.EmitLoadOfPointer(CGF.Builder.CreateElementBitCast( |
| 1911 | SrcElementPtrAddr, PrivateLlvmPtrType), |
| 1912 | PrivatePtrType->castAs<PointerType>()); |
| 1913 | |
| 1914 | // Step 1.2: Create a temporary to store the element in the destination |
| 1915 | // Reduce list. |
| 1916 | DestElementPtrAddr = Bld.CreateConstArrayGEP(DestBase, Idx); |
| 1917 | DestElementAddr = |
| 1918 | CGF.CreateMemTemp(Private->getType(), ".omp.reduction.element"); |
| 1919 | ShuffleInElement = true; |
| 1920 | UpdateDestListPtr = true; |
| 1921 | break; |
| 1922 | } |
| 1923 | case ThreadCopy: { |
| 1924 | // Step 1.1: Get the address for the src element in the Reduce list. |
| 1925 | Address SrcElementPtrAddr = Bld.CreateConstArrayGEP(SrcBase, Idx); |
| 1926 | SrcElementAddr = |
| 1927 | CGF.EmitLoadOfPointer(CGF.Builder.CreateElementBitCast( |
| 1928 | SrcElementPtrAddr, PrivateLlvmPtrType), |
| 1929 | PrivatePtrType->castAs<PointerType>()); |
| 1930 | |
| 1931 | // Step 1.2: Get the address for dest element. The destination |
| 1932 | // element has already been created on the thread's stack. |
| 1933 | DestElementPtrAddr = Bld.CreateConstArrayGEP(DestBase, Idx); |
| 1934 | DestElementAddr = |
| 1935 | CGF.EmitLoadOfPointer(CGF.Builder.CreateElementBitCast( |
| 1936 | DestElementPtrAddr, PrivateLlvmPtrType), |
| 1937 | PrivatePtrType->castAs<PointerType>()); |
| 1938 | break; |
| 1939 | } |
| 1940 | case ThreadToScratchpad: { |
| 1941 | // Step 1.1: Get the address for the src element in the Reduce list. |
| 1942 | Address SrcElementPtrAddr = Bld.CreateConstArrayGEP(SrcBase, Idx); |
| 1943 | SrcElementAddr = |
| 1944 | CGF.EmitLoadOfPointer(CGF.Builder.CreateElementBitCast( |
| 1945 | SrcElementPtrAddr, PrivateLlvmPtrType), |
| 1946 | PrivatePtrType->castAs<PointerType>()); |
| 1947 | |
| 1948 | // Step 1.2: Get the address for dest element: |
| 1949 | // address = base + index * ElementSizeInChars. |
| 1950 | llvm::Value *ElementSizeInChars = CGF.getTypeSize(Private->getType()); |
| 1951 | llvm::Value *CurrentOffset = |
| 1952 | Bld.CreateNUWMul(ElementSizeInChars, ScratchpadIndex); |
| 1953 | llvm::Value *ScratchPadElemAbsolutePtrVal = |
| 1954 | Bld.CreateNUWAdd(DestBase.getPointer(), CurrentOffset); |
| 1955 | ScratchPadElemAbsolutePtrVal = |
| 1956 | Bld.CreateIntToPtr(ScratchPadElemAbsolutePtrVal, CGF.VoidPtrTy); |
| 1957 | DestElementAddr = Address(ScratchPadElemAbsolutePtrVal, CGF.Int8Ty, |
| 1958 | C.getTypeAlignInChars(Private->getType())); |
| 1959 | IncrScratchpadDest = true; |
| 1960 | break; |
| 1961 | } |
| 1962 | case ScratchpadToThread: { |
| 1963 | // Step 1.1: Get the address for the src element in the scratchpad. |
| 1964 | // address = base + index * ElementSizeInChars. |
| 1965 | llvm::Value *ElementSizeInChars = CGF.getTypeSize(Private->getType()); |
| 1966 | llvm::Value *CurrentOffset = |
| 1967 | Bld.CreateNUWMul(ElementSizeInChars, ScratchpadIndex); |
| 1968 | llvm::Value *ScratchPadElemAbsolutePtrVal = |
| 1969 | Bld.CreateNUWAdd(SrcBase.getPointer(), CurrentOffset); |
| 1970 | ScratchPadElemAbsolutePtrVal = |
| 1971 | Bld.CreateIntToPtr(ScratchPadElemAbsolutePtrVal, CGF.VoidPtrTy); |
| 1972 | SrcElementAddr = Address(ScratchPadElemAbsolutePtrVal, CGF.Int8Ty, |
| 1973 | C.getTypeAlignInChars(Private->getType())); |
| 1974 | IncrScratchpadSrc = true; |
| 1975 | |
| 1976 | // Step 1.2: Create a temporary to store the element in the destination |
| 1977 | // Reduce list. |
| 1978 | DestElementPtrAddr = Bld.CreateConstArrayGEP(DestBase, Idx); |
| 1979 | DestElementAddr = |
| 1980 | CGF.CreateMemTemp(Private->getType(), ".omp.reduction.element"); |
| 1981 | UpdateDestListPtr = true; |
| 1982 | break; |
| 1983 | } |
| 1984 | } |
| 1985 | |
| 1986 | // Regardless of src and dest of copy, we emit the load of src |
| 1987 | // element as this is required in all directions |
| 1988 | SrcElementAddr = Bld.CreateElementBitCast( |
| 1989 | SrcElementAddr, CGF.ConvertTypeForMem(Private->getType())); |
| 1990 | DestElementAddr = Bld.CreateElementBitCast(DestElementAddr, |
| 1991 | SrcElementAddr.getElementType()); |
| 1992 | |
| 1993 | // Now that all active lanes have read the element in the |
| 1994 | // Reduce list, shuffle over the value from the remote lane. |
| 1995 | if (ShuffleInElement) { |
| 1996 | shuffleAndStore(CGF, SrcElementAddr, DestElementAddr, Private->getType(), |
| 1997 | RemoteLaneOffset, Private->getExprLoc()); |
| 1998 | } else { |
| 1999 | switch (CGF.getEvaluationKind(Private->getType())) { |
| 2000 | case TEK_Scalar: { |
| 2001 | llvm::Value *Elem = CGF.EmitLoadOfScalar( |
| 2002 | SrcElementAddr, /*Volatile=*/false, Private->getType(), |
| 2003 | Private->getExprLoc(), LValueBaseInfo(AlignmentSource::Type), |
| 2004 | TBAAAccessInfo()); |
| 2005 | // Store the source element value to the dest element address. |
| 2006 | CGF.EmitStoreOfScalar( |
| 2007 | Elem, DestElementAddr, /*Volatile=*/false, Private->getType(), |
| 2008 | LValueBaseInfo(AlignmentSource::Type), TBAAAccessInfo()); |
| 2009 | break; |
| 2010 | } |
| 2011 | case TEK_Complex: { |
| 2012 | CodeGenFunction::ComplexPairTy Elem = CGF.EmitLoadOfComplex( |
| 2013 | CGF.MakeAddrLValue(SrcElementAddr, Private->getType()), |
| 2014 | Private->getExprLoc()); |
| 2015 | CGF.EmitStoreOfComplex( |
| 2016 | Elem, CGF.MakeAddrLValue(DestElementAddr, Private->getType()), |
| 2017 | /*isInit=*/false); |
| 2018 | break; |
| 2019 | } |
| 2020 | case TEK_Aggregate: |
| 2021 | CGF.EmitAggregateCopy( |
| 2022 | CGF.MakeAddrLValue(DestElementAddr, Private->getType()), |
| 2023 | CGF.MakeAddrLValue(SrcElementAddr, Private->getType()), |
| 2024 | Private->getType(), AggValueSlot::DoesNotOverlap); |
| 2025 | break; |
| 2026 | } |
| 2027 | } |
| 2028 | |
| 2029 | // Step 3.1: Modify reference in dest Reduce list as needed. |
| 2030 | // Modifying the reference in Reduce list to point to the newly |
| 2031 | // created element. The element is live in the current function |
| 2032 | // scope and that of functions it invokes (i.e., reduce_function). |
| 2033 | // RemoteReduceData[i] = (void*)&RemoteElem |
| 2034 | if (UpdateDestListPtr) { |
| 2035 | CGF.EmitStoreOfScalar(Bld.CreatePointerBitCastOrAddrSpaceCast( |
| 2036 | DestElementAddr.getPointer(), CGF.VoidPtrTy), |
| 2037 | DestElementPtrAddr, /*Volatile=*/false, |
| 2038 | C.VoidPtrTy); |
| 2039 | } |
| 2040 | |
| 2041 | // Step 4.1: Increment SrcBase/DestBase so that it points to the starting |
| 2042 | // address of the next element in scratchpad memory, unless we're currently |
| 2043 | // processing the last one. Memory alignment is also taken care of here. |
| 2044 | if ((IncrScratchpadDest || IncrScratchpadSrc) && (Idx + 1 < Size)) { |
| 2045 | // FIXME: This code doesn't make any sense, it's trying to perform |
| 2046 | // integer arithmetic on pointers. |
| 2047 | llvm::Value *ScratchpadBasePtr = |
| 2048 | IncrScratchpadDest ? DestBase.getPointer() : SrcBase.getPointer(); |
| 2049 | llvm::Value *ElementSizeInChars = CGF.getTypeSize(Private->getType()); |
| 2050 | ScratchpadBasePtr = Bld.CreateNUWAdd( |
| 2051 | ScratchpadBasePtr, |
| 2052 | Bld.CreateNUWMul(ScratchpadWidth, ElementSizeInChars)); |
| 2053 | |
| 2054 | // Take care of global memory alignment for performance |
| 2055 | ScratchpadBasePtr = Bld.CreateNUWSub( |
| 2056 | ScratchpadBasePtr, llvm::ConstantInt::get(CGM.SizeTy, 1)); |
| 2057 | ScratchpadBasePtr = Bld.CreateUDiv( |
| 2058 | ScratchpadBasePtr, |
| 2059 | llvm::ConstantInt::get(CGM.SizeTy, GlobalMemoryAlignment)); |
| 2060 | ScratchpadBasePtr = Bld.CreateNUWAdd( |
| 2061 | ScratchpadBasePtr, llvm::ConstantInt::get(CGM.SizeTy, 1)); |
| 2062 | ScratchpadBasePtr = Bld.CreateNUWMul( |
| 2063 | ScratchpadBasePtr, |
| 2064 | llvm::ConstantInt::get(CGM.SizeTy, GlobalMemoryAlignment)); |
| 2065 | |
| 2066 | if (IncrScratchpadDest) |
| 2067 | DestBase = |
| 2068 | Address(ScratchpadBasePtr, CGF.VoidPtrTy, CGF.getPointerAlign()); |
| 2069 | else /* IncrScratchpadSrc = true */ |
| 2070 | SrcBase = |
| 2071 | Address(ScratchpadBasePtr, CGF.VoidPtrTy, CGF.getPointerAlign()); |
| 2072 | } |
| 2073 | |
| 2074 | ++Idx; |
| 2075 | } |
| 2076 | } |
| 2077 | |
| 2078 | /// This function emits a helper that gathers Reduce lists from the first |
| 2079 | /// lane of every active warp to lanes in the first warp. |
| 2080 | /// |
| 2081 | /// void inter_warp_copy_func(void* reduce_data, num_warps) |
| 2082 | /// shared smem[warp_size]; |
| 2083 | /// For all data entries D in reduce_data: |
| 2084 | /// sync |
| 2085 | /// If (I am the first lane in each warp) |
| 2086 | /// Copy my local D to smem[warp_id] |
| 2087 | /// sync |
| 2088 | /// if (I am the first warp) |
| 2089 | /// Copy smem[thread_id] to my local D |
| 2090 | static llvm::Value *emitInterWarpCopyFunction(CodeGenModule &CGM, |
| 2091 | ArrayRef<const Expr *> Privates, |
| 2092 | QualType ReductionArrayTy, |
| 2093 | SourceLocation Loc) { |
| 2094 | ASTContext &C = CGM.getContext(); |
| 2095 | llvm::Module &M = CGM.getModule(); |
| 2096 | |
| 2097 | // ReduceList: thread local Reduce list. |
| 2098 | // At the stage of the computation when this function is called, partially |
| 2099 | // aggregated values reside in the first lane of every active warp. |
| 2100 | ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, |
| 2101 | C.VoidPtrTy, ImplicitParamDecl::Other); |
| 2102 | // NumWarps: number of warps active in the parallel region. This could |
| 2103 | // be smaller than 32 (max warps in a CTA) for partial block reduction. |
| 2104 | ImplicitParamDecl NumWarpsArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, |
| 2105 | C.getIntTypeForBitwidth(32, /* Signed */ true), |
| 2106 | ImplicitParamDecl::Other); |
| 2107 | FunctionArgList Args; |
| 2108 | Args.push_back(&ReduceListArg); |
| 2109 | Args.push_back(&NumWarpsArg); |
| 2110 | |
| 2111 | const CGFunctionInfo &CGFI = |
| 2112 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); |
| 2113 | auto *Fn = llvm::Function::Create(CGM.getTypes().GetFunctionType(CGFI), |
| 2114 | llvm::GlobalValue::InternalLinkage, |
| 2115 | "_omp_reduction_inter_warp_copy_func", &M); |
| 2116 | CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI); |
| 2117 | Fn->setDoesNotRecurse(); |
| 2118 | CodeGenFunction CGF(CGM); |
| 2119 | CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc); |
| 2120 | |
| 2121 | CGBuilderTy &Bld = CGF.Builder; |
| 2122 | |
| 2123 | // This array is used as a medium to transfer, one reduce element at a time, |
| 2124 | // the data from the first lane of every warp to lanes in the first warp |
| 2125 | // in order to perform the final step of a reduction in a parallel region |
| 2126 | // (reduction across warps). The array is placed in NVPTX __shared__ memory |
| 2127 | // for reduced latency, as well as to have a distinct copy for concurrently |
| 2128 | // executing target regions. The array is declared with common linkage so |
| 2129 | // as to be shared across compilation units. |
| 2130 | StringRef TransferMediumName = |
| 2131 | "__openmp_nvptx_data_transfer_temporary_storage"; |
| 2132 | llvm::GlobalVariable *TransferMedium = |
| 2133 | M.getGlobalVariable(TransferMediumName); |
| 2134 | unsigned WarpSize = CGF.getTarget().getGridValue().GV_Warp_Size; |
| 2135 | if (!TransferMedium) { |
| 2136 | auto *Ty = llvm::ArrayType::get(CGM.Int32Ty, WarpSize); |
| 2137 | unsigned SharedAddressSpace = C.getTargetAddressSpace(LangAS::cuda_shared); |
| 2138 | TransferMedium = new llvm::GlobalVariable( |
| 2139 | M, Ty, /*isConstant=*/false, llvm::GlobalVariable::WeakAnyLinkage, |
| 2140 | llvm::UndefValue::get(Ty), TransferMediumName, |
| 2141 | /*InsertBefore=*/nullptr, llvm::GlobalVariable::NotThreadLocal, |
| 2142 | SharedAddressSpace); |
| 2143 | CGM.addCompilerUsedGlobal(TransferMedium); |
| 2144 | } |
| 2145 | |
| 2146 | auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()); |
| 2147 | // Get the CUDA thread id of the current OpenMP thread on the GPU. |
| 2148 | llvm::Value *ThreadID = RT.getGPUThreadID(CGF); |
| 2149 | // nvptx_lane_id = nvptx_id % warpsize |
| 2150 | llvm::Value *LaneID = getNVPTXLaneID(CGF); |
| 2151 | // nvptx_warp_id = nvptx_id / warpsize |
| 2152 | llvm::Value *WarpID = getNVPTXWarpID(CGF); |
| 2153 | |
| 2154 | Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg); |
| 2155 | llvm::Type *ElemTy = CGF.ConvertTypeForMem(ReductionArrayTy); |
| 2156 | Address LocalReduceList( |
| 2157 | Bld.CreatePointerBitCastOrAddrSpaceCast( |
| 2158 | CGF.EmitLoadOfScalar( |
| 2159 | AddrReduceListArg, /*Volatile=*/false, C.VoidPtrTy, Loc, |
| 2160 | LValueBaseInfo(AlignmentSource::Type), TBAAAccessInfo()), |
| 2161 | ElemTy->getPointerTo()), |
| 2162 | ElemTy, CGF.getPointerAlign()); |
| 2163 | |
| 2164 | unsigned Idx = 0; |
| 2165 | for (const Expr *Private : Privates) { |
| 2166 | // |
| 2167 | // Warp master copies reduce element to transfer medium in __shared__ |
| 2168 | // memory. |
| 2169 | // |
| 2170 | unsigned RealTySize = |
| 2171 | C.getTypeSizeInChars(Private->getType()) |
| 2172 | .alignTo(C.getTypeAlignInChars(Private->getType())) |
| 2173 | .getQuantity(); |
| 2174 | for (unsigned TySize = 4; TySize > 0 && RealTySize > 0; TySize /=2) { |
| 2175 | unsigned NumIters = RealTySize / TySize; |
| 2176 | if (NumIters == 0) |
| 2177 | continue; |
| 2178 | QualType CType = C.getIntTypeForBitwidth( |
| 2179 | C.toBits(CharUnits::fromQuantity(TySize)), /*Signed=*/1); |
| 2180 | llvm::Type *CopyType = CGF.ConvertTypeForMem(CType); |
| 2181 | CharUnits Align = CharUnits::fromQuantity(TySize); |
| 2182 | llvm::Value *Cnt = nullptr; |
| 2183 | Address CntAddr = Address::invalid(); |
| 2184 | llvm::BasicBlock *PrecondBB = nullptr; |
| 2185 | llvm::BasicBlock *ExitBB = nullptr; |
| 2186 | if (NumIters > 1) { |
| 2187 | CntAddr = CGF.CreateMemTemp(C.IntTy, ".cnt.addr"); |
| 2188 | CGF.EmitStoreOfScalar(llvm::Constant::getNullValue(CGM.IntTy), CntAddr, |
| 2189 | /*Volatile=*/false, C.IntTy); |
| 2190 | PrecondBB = CGF.createBasicBlock("precond"); |
| 2191 | ExitBB = CGF.createBasicBlock("exit"); |
| 2192 | llvm::BasicBlock *BodyBB = CGF.createBasicBlock("body"); |
| 2193 | // There is no need to emit line number for unconditional branch. |
| 2194 | (void)ApplyDebugLocation::CreateEmpty(CGF); |
| 2195 | CGF.EmitBlock(PrecondBB); |
| 2196 | Cnt = CGF.EmitLoadOfScalar(CntAddr, /*Volatile=*/false, C.IntTy, Loc); |
| 2197 | llvm::Value *Cmp = |
| 2198 | Bld.CreateICmpULT(Cnt, llvm::ConstantInt::get(CGM.IntTy, NumIters)); |
| 2199 | Bld.CreateCondBr(Cmp, BodyBB, ExitBB); |
| 2200 | CGF.EmitBlock(BodyBB); |
| 2201 | } |
| 2202 | // kmpc_barrier. |
| 2203 | CGM.getOpenMPRuntime().emitBarrierCall(CGF, Loc, OMPD_unknown, |
| 2204 | /*EmitChecks=*/false, |
| 2205 | /*ForceSimpleCall=*/true); |
| 2206 | llvm::BasicBlock *ThenBB = CGF.createBasicBlock("then"); |
| 2207 | llvm::BasicBlock *ElseBB = CGF.createBasicBlock("else"); |
| 2208 | llvm::BasicBlock *MergeBB = CGF.createBasicBlock("ifcont"); |
| 2209 | |
| 2210 | // if (lane_id == 0) |
| 2211 | llvm::Value *IsWarpMaster = Bld.CreateIsNull(LaneID, "warp_master"); |
| 2212 | Bld.CreateCondBr(IsWarpMaster, ThenBB, ElseBB); |
| 2213 | CGF.EmitBlock(ThenBB); |
| 2214 | |
| 2215 | // Reduce element = LocalReduceList[i] |
| 2216 | Address ElemPtrPtrAddr = Bld.CreateConstArrayGEP(LocalReduceList, Idx); |
| 2217 | llvm::Value *ElemPtrPtr = CGF.EmitLoadOfScalar( |
| 2218 | ElemPtrPtrAddr, /*Volatile=*/false, C.VoidPtrTy, SourceLocation()); |
| 2219 | // elemptr = ((CopyType*)(elemptrptr)) + I |
| 2220 | Address ElemPtr(ElemPtrPtr, CGF.Int8Ty, Align); |
| 2221 | ElemPtr = Bld.CreateElementBitCast(ElemPtr, CopyType); |
| 2222 | if (NumIters > 1) |
| 2223 | ElemPtr = Bld.CreateGEP(ElemPtr, Cnt); |
| 2224 | |
| 2225 | // Get pointer to location in transfer medium. |
| 2226 | // MediumPtr = &medium[warp_id] |
| 2227 | llvm::Value *MediumPtrVal = Bld.CreateInBoundsGEP( |
| 2228 | TransferMedium->getValueType(), TransferMedium, |
| 2229 | {llvm::Constant::getNullValue(CGM.Int64Ty), WarpID}); |
| 2230 | // Casting to actual data type. |
| 2231 | // MediumPtr = (CopyType*)MediumPtrAddr; |
| 2232 | Address MediumPtr( |
| 2233 | Bld.CreateBitCast( |
| 2234 | MediumPtrVal, |
| 2235 | CopyType->getPointerTo( |
| 2236 | MediumPtrVal->getType()->getPointerAddressSpace())), |
| 2237 | CopyType, Align); |
| 2238 | |
| 2239 | // elem = *elemptr |
| 2240 | //*MediumPtr = elem |
| 2241 | llvm::Value *Elem = CGF.EmitLoadOfScalar( |
| 2242 | ElemPtr, /*Volatile=*/false, CType, Loc, |
| 2243 | LValueBaseInfo(AlignmentSource::Type), TBAAAccessInfo()); |
| 2244 | // Store the source element value to the dest element address. |
| 2245 | CGF.EmitStoreOfScalar(Elem, MediumPtr, /*Volatile=*/true, CType, |
| 2246 | LValueBaseInfo(AlignmentSource::Type), |
| 2247 | TBAAAccessInfo()); |
| 2248 | |
| 2249 | Bld.CreateBr(MergeBB); |
| 2250 | |
| 2251 | CGF.EmitBlock(ElseBB); |
| 2252 | Bld.CreateBr(MergeBB); |
| 2253 | |
| 2254 | CGF.EmitBlock(MergeBB); |
| 2255 | |
| 2256 | // kmpc_barrier. |
| 2257 | CGM.getOpenMPRuntime().emitBarrierCall(CGF, Loc, OMPD_unknown, |
| 2258 | /*EmitChecks=*/false, |
| 2259 | /*ForceSimpleCall=*/true); |
| 2260 | |
| 2261 | // |
| 2262 | // Warp 0 copies reduce element from transfer medium. |
| 2263 | // |
| 2264 | llvm::BasicBlock *W0ThenBB = CGF.createBasicBlock("then"); |
| 2265 | llvm::BasicBlock *W0ElseBB = CGF.createBasicBlock("else"); |
| 2266 | llvm::BasicBlock *W0MergeBB = CGF.createBasicBlock("ifcont"); |
| 2267 | |
| 2268 | Address AddrNumWarpsArg = CGF.GetAddrOfLocalVar(&NumWarpsArg); |
| 2269 | llvm::Value *NumWarpsVal = CGF.EmitLoadOfScalar( |
| 2270 | AddrNumWarpsArg, /*Volatile=*/false, C.IntTy, Loc); |
| 2271 | |
| 2272 | // Up to 32 threads in warp 0 are active. |
| 2273 | llvm::Value *IsActiveThread = |
| 2274 | Bld.CreateICmpULT(ThreadID, NumWarpsVal, "is_active_thread"); |
| 2275 | Bld.CreateCondBr(IsActiveThread, W0ThenBB, W0ElseBB); |
| 2276 | |
| 2277 | CGF.EmitBlock(W0ThenBB); |
| 2278 | |
| 2279 | // SrcMediumPtr = &medium[tid] |
| 2280 | llvm::Value *SrcMediumPtrVal = Bld.CreateInBoundsGEP( |
| 2281 | TransferMedium->getValueType(), TransferMedium, |
| 2282 | {llvm::Constant::getNullValue(CGM.Int64Ty), ThreadID}); |
| 2283 | // SrcMediumVal = *SrcMediumPtr; |
| 2284 | Address SrcMediumPtr( |
| 2285 | Bld.CreateBitCast( |
| 2286 | SrcMediumPtrVal, |
| 2287 | CopyType->getPointerTo( |
| 2288 | SrcMediumPtrVal->getType()->getPointerAddressSpace())), |
| 2289 | CopyType, Align); |
| 2290 | |
| 2291 | // TargetElemPtr = (CopyType*)(SrcDataAddr[i]) + I |
| 2292 | Address TargetElemPtrPtr = Bld.CreateConstArrayGEP(LocalReduceList, Idx); |
| 2293 | llvm::Value *TargetElemPtrVal = CGF.EmitLoadOfScalar( |
| 2294 | TargetElemPtrPtr, /*Volatile=*/false, C.VoidPtrTy, Loc); |
| 2295 | Address TargetElemPtr(TargetElemPtrVal, CGF.Int8Ty, Align); |
| 2296 | TargetElemPtr = Bld.CreateElementBitCast(TargetElemPtr, CopyType); |
| 2297 | if (NumIters > 1) |
| 2298 | TargetElemPtr = Bld.CreateGEP(TargetElemPtr, Cnt); |
| 2299 | |
| 2300 | // *TargetElemPtr = SrcMediumVal; |
| 2301 | llvm::Value *SrcMediumValue = |
| 2302 | CGF.EmitLoadOfScalar(SrcMediumPtr, /*Volatile=*/true, CType, Loc); |
| 2303 | CGF.EmitStoreOfScalar(SrcMediumValue, TargetElemPtr, /*Volatile=*/false, |
| 2304 | CType); |
| 2305 | Bld.CreateBr(W0MergeBB); |
| 2306 | |
| 2307 | CGF.EmitBlock(W0ElseBB); |
| 2308 | Bld.CreateBr(W0MergeBB); |
| 2309 | |
| 2310 | CGF.EmitBlock(W0MergeBB); |
| 2311 | |
| 2312 | if (NumIters > 1) { |
| 2313 | Cnt = Bld.CreateNSWAdd(Cnt, llvm::ConstantInt::get(CGM.IntTy, /*V=*/1)); |
| 2314 | CGF.EmitStoreOfScalar(Cnt, CntAddr, /*Volatile=*/false, C.IntTy); |
| 2315 | CGF.EmitBranch(PrecondBB); |
| 2316 | (void)ApplyDebugLocation::CreateEmpty(CGF); |
| 2317 | CGF.EmitBlock(ExitBB); |
| 2318 | } |
| 2319 | RealTySize %= TySize; |
| 2320 | } |
| 2321 | ++Idx; |
| 2322 | } |
| 2323 | |
| 2324 | CGF.FinishFunction(); |
| 2325 | return Fn; |
| 2326 | } |
| 2327 | |
| 2328 | /// Emit a helper that reduces data across two OpenMP threads (lanes) |
| 2329 | /// in the same warp. It uses shuffle instructions to copy over data from |
| 2330 | /// a remote lane's stack. The reduction algorithm performed is specified |
| 2331 | /// by the fourth parameter. |
| 2332 | /// |
| 2333 | /// Algorithm Versions. |
| 2334 | /// Full Warp Reduce (argument value 0): |
| 2335 | /// This algorithm assumes that all 32 lanes are active and gathers |
| 2336 | /// data from these 32 lanes, producing a single resultant value. |
| 2337 | /// Contiguous Partial Warp Reduce (argument value 1): |
| 2338 | /// This algorithm assumes that only a *contiguous* subset of lanes |
| 2339 | /// are active. This happens for the last warp in a parallel region |
| 2340 | /// when the user specified num_threads is not an integer multiple of |
| 2341 | /// 32. This contiguous subset always starts with the zeroth lane. |
| 2342 | /// Partial Warp Reduce (argument value 2): |
| 2343 | /// This algorithm gathers data from any number of lanes at any position. |
| 2344 | /// All reduced values are stored in the lowest possible lane. The set |
| 2345 | /// of problems every algorithm addresses is a super set of those |
| 2346 | /// addressable by algorithms with a lower version number. Overhead |
| 2347 | /// increases as algorithm version increases. |
| 2348 | /// |
| 2349 | /// Terminology |
| 2350 | /// Reduce element: |
| 2351 | /// Reduce element refers to the individual data field with primitive |
| 2352 | /// data types to be combined and reduced across threads. |
| 2353 | /// Reduce list: |
| 2354 | /// Reduce list refers to a collection of local, thread-private |
| 2355 | /// reduce elements. |
| 2356 | /// Remote Reduce list: |
| 2357 | /// Remote Reduce list refers to a collection of remote (relative to |
| 2358 | /// the current thread) reduce elements. |
| 2359 | /// |
| 2360 | /// We distinguish between three states of threads that are important to |
| 2361 | /// the implementation of this function. |
| 2362 | /// Alive threads: |
| 2363 | /// Threads in a warp executing the SIMT instruction, as distinguished from |
| 2364 | /// threads that are inactive due to divergent control flow. |
| 2365 | /// Active threads: |
| 2366 | /// The minimal set of threads that has to be alive upon entry to this |
| 2367 | /// function. The computation is correct iff active threads are alive. |
| 2368 | /// Some threads are alive but they are not active because they do not |
| 2369 | /// contribute to the computation in any useful manner. Turning them off |
| 2370 | /// may introduce control flow overheads without any tangible benefits. |
| 2371 | /// Effective threads: |
| 2372 | /// In order to comply with the argument requirements of the shuffle |
| 2373 | /// function, we must keep all lanes holding data alive. But at most |
| 2374 | /// half of them perform value aggregation; we refer to this half of |
| 2375 | /// threads as effective. The other half is simply handing off their |
| 2376 | /// data. |
| 2377 | /// |
| 2378 | /// Procedure |
| 2379 | /// Value shuffle: |
| 2380 | /// In this step active threads transfer data from higher lane positions |
| 2381 | /// in the warp to lower lane positions, creating Remote Reduce list. |
| 2382 | /// Value aggregation: |
| 2383 | /// In this step, effective threads combine their thread local Reduce list |
| 2384 | /// with Remote Reduce list and store the result in the thread local |
| 2385 | /// Reduce list. |
| 2386 | /// Value copy: |
| 2387 | /// In this step, we deal with the assumption made by algorithm 2 |
| 2388 | /// (i.e. contiguity assumption). When we have an odd number of lanes |
| 2389 | /// active, say 2k+1, only k threads will be effective and therefore k |
| 2390 | /// new values will be produced. However, the Reduce list owned by the |
| 2391 | /// (2k+1)th thread is ignored in the value aggregation. Therefore |
| 2392 | /// we copy the Reduce list from the (2k+1)th lane to (k+1)th lane so |
| 2393 | /// that the contiguity assumption still holds. |
| 2394 | static llvm::Function *emitShuffleAndReduceFunction( |
| 2395 | CodeGenModule &CGM, ArrayRef<const Expr *> Privates, |
| 2396 | QualType ReductionArrayTy, llvm::Function *ReduceFn, SourceLocation Loc) { |
| 2397 | ASTContext &C = CGM.getContext(); |
| 2398 | |
| 2399 | // Thread local Reduce list used to host the values of data to be reduced. |
| 2400 | ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, |
| 2401 | C.VoidPtrTy, ImplicitParamDecl::Other); |
| 2402 | // Current lane id; could be logical. |
| 2403 | ImplicitParamDecl LaneIDArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.ShortTy, |
| 2404 | ImplicitParamDecl::Other); |
| 2405 | // Offset of the remote source lane relative to the current lane. |
| 2406 | ImplicitParamDecl RemoteLaneOffsetArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, |
| 2407 | C.ShortTy, ImplicitParamDecl::Other); |
| 2408 | // Algorithm version. This is expected to be known at compile time. |
| 2409 | ImplicitParamDecl AlgoVerArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, |
| 2410 | C.ShortTy, ImplicitParamDecl::Other); |
| 2411 | FunctionArgList Args; |
| 2412 | Args.push_back(&ReduceListArg); |
| 2413 | Args.push_back(&LaneIDArg); |
| 2414 | Args.push_back(&RemoteLaneOffsetArg); |
| 2415 | Args.push_back(&AlgoVerArg); |
| 2416 | |
| 2417 | const CGFunctionInfo &CGFI = |
| 2418 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); |
| 2419 | auto *Fn = llvm::Function::Create( |
| 2420 | CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage, |
| 2421 | "_omp_reduction_shuffle_and_reduce_func", &CGM.getModule()); |
| 2422 | CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI); |
| 2423 | Fn->setDoesNotRecurse(); |
| 2424 | |
| 2425 | CodeGenFunction CGF(CGM); |
| 2426 | CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc); |
| 2427 | |
| 2428 | CGBuilderTy &Bld = CGF.Builder; |
| 2429 | |
| 2430 | Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg); |
| 2431 | llvm::Type *ElemTy = CGF.ConvertTypeForMem(ReductionArrayTy); |
| 2432 | Address LocalReduceList( |
| 2433 | Bld.CreatePointerBitCastOrAddrSpaceCast( |
| 2434 | CGF.EmitLoadOfScalar(AddrReduceListArg, /*Volatile=*/false, |
| 2435 | C.VoidPtrTy, SourceLocation()), |
| 2436 | ElemTy->getPointerTo()), |
| 2437 | ElemTy, CGF.getPointerAlign()); |
| 2438 | |
| 2439 | Address AddrLaneIDArg = CGF.GetAddrOfLocalVar(&LaneIDArg); |
| 2440 | llvm::Value *LaneIDArgVal = CGF.EmitLoadOfScalar( |
| 2441 | AddrLaneIDArg, /*Volatile=*/false, C.ShortTy, SourceLocation()); |
| 2442 | |
| 2443 | Address AddrRemoteLaneOffsetArg = CGF.GetAddrOfLocalVar(&RemoteLaneOffsetArg); |
| 2444 | llvm::Value *RemoteLaneOffsetArgVal = CGF.EmitLoadOfScalar( |
| 2445 | AddrRemoteLaneOffsetArg, /*Volatile=*/false, C.ShortTy, SourceLocation()); |
| 2446 | |
| 2447 | Address AddrAlgoVerArg = CGF.GetAddrOfLocalVar(&AlgoVerArg); |
| 2448 | llvm::Value *AlgoVerArgVal = CGF.EmitLoadOfScalar( |
| 2449 | AddrAlgoVerArg, /*Volatile=*/false, C.ShortTy, SourceLocation()); |
| 2450 | |
| 2451 | // Create a local thread-private variable to host the Reduce list |
| 2452 | // from a remote lane. |
| 2453 | Address RemoteReduceList = |
| 2454 | CGF.CreateMemTemp(ReductionArrayTy, ".omp.reduction.remote_reduce_list"); |
| 2455 | |
| 2456 | // This loop iterates through the list of reduce elements and copies, |
| 2457 | // element by element, from a remote lane in the warp to RemoteReduceList, |
| 2458 | // hosted on the thread's stack. |
| 2459 | emitReductionListCopy(RemoteLaneToThread, CGF, ReductionArrayTy, Privates, |
| 2460 | LocalReduceList, RemoteReduceList, |
| 2461 | {/*RemoteLaneOffset=*/RemoteLaneOffsetArgVal, |
| 2462 | /*ScratchpadIndex=*/nullptr, |
| 2463 | /*ScratchpadWidth=*/nullptr}); |
| 2464 | |
| 2465 | // The actions to be performed on the Remote Reduce list is dependent |
| 2466 | // on the algorithm version. |
| 2467 | // |
| 2468 | // if (AlgoVer==0) || (AlgoVer==1 && (LaneId < Offset)) || (AlgoVer==2 && |
| 2469 | // LaneId % 2 == 0 && Offset > 0): |
| 2470 | // do the reduction value aggregation |
| 2471 | // |
| 2472 | // The thread local variable Reduce list is mutated in place to host the |
| 2473 | // reduced data, which is the aggregated value produced from local and |
| 2474 | // remote lanes. |
| 2475 | // |
| 2476 | // Note that AlgoVer is expected to be a constant integer known at compile |
| 2477 | // time. |
| 2478 | // When AlgoVer==0, the first conjunction evaluates to true, making |
| 2479 | // the entire predicate true during compile time. |
| 2480 | // When AlgoVer==1, the second conjunction has only the second part to be |
| 2481 | // evaluated during runtime. Other conjunctions evaluates to false |
| 2482 | // during compile time. |
| 2483 | // When AlgoVer==2, the third conjunction has only the second part to be |
| 2484 | // evaluated during runtime. Other conjunctions evaluates to false |
| 2485 | // during compile time. |
| 2486 | llvm::Value *CondAlgo0 = Bld.CreateIsNull(AlgoVerArgVal); |
| 2487 | |
| 2488 | llvm::Value *Algo1 = Bld.CreateICmpEQ(AlgoVerArgVal, Bld.getInt16(1)); |
| 2489 | llvm::Value *CondAlgo1 = Bld.CreateAnd( |
| 2490 | Algo1, Bld.CreateICmpULT(LaneIDArgVal, RemoteLaneOffsetArgVal)); |
| 2491 | |
| 2492 | llvm::Value *Algo2 = Bld.CreateICmpEQ(AlgoVerArgVal, Bld.getInt16(2)); |
| 2493 | llvm::Value *CondAlgo2 = Bld.CreateAnd( |
| 2494 | Algo2, Bld.CreateIsNull(Bld.CreateAnd(LaneIDArgVal, Bld.getInt16(1)))); |
| 2495 | CondAlgo2 = Bld.CreateAnd( |
| 2496 | CondAlgo2, Bld.CreateICmpSGT(RemoteLaneOffsetArgVal, Bld.getInt16(0))); |
| 2497 | |
| 2498 | llvm::Value *CondReduce = Bld.CreateOr(CondAlgo0, CondAlgo1); |
| 2499 | CondReduce = Bld.CreateOr(CondReduce, CondAlgo2); |
| 2500 | |
| 2501 | llvm::BasicBlock *ThenBB = CGF.createBasicBlock("then"); |
| 2502 | llvm::BasicBlock *ElseBB = CGF.createBasicBlock("else"); |
| 2503 | llvm::BasicBlock *MergeBB = CGF.createBasicBlock("ifcont"); |
| 2504 | Bld.CreateCondBr(CondReduce, ThenBB, ElseBB); |
| 2505 | |
| 2506 | CGF.EmitBlock(ThenBB); |
| 2507 | // reduce_function(LocalReduceList, RemoteReduceList) |
| 2508 | llvm::Value *LocalReduceListPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( |
| 2509 | LocalReduceList.getPointer(), CGF.VoidPtrTy); |
| 2510 | llvm::Value *RemoteReduceListPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( |
| 2511 | RemoteReduceList.getPointer(), CGF.VoidPtrTy); |
| 2512 | CGM.getOpenMPRuntime().emitOutlinedFunctionCall( |
| 2513 | CGF, Loc, ReduceFn, {LocalReduceListPtr, RemoteReduceListPtr}); |
| 2514 | Bld.CreateBr(MergeBB); |
| 2515 | |
| 2516 | CGF.EmitBlock(ElseBB); |
| 2517 | Bld.CreateBr(MergeBB); |
| 2518 | |
| 2519 | CGF.EmitBlock(MergeBB); |
| 2520 | |
| 2521 | // if (AlgoVer==1 && (LaneId >= Offset)) copy Remote Reduce list to local |
| 2522 | // Reduce list. |
| 2523 | Algo1 = Bld.CreateICmpEQ(AlgoVerArgVal, Bld.getInt16(1)); |
| 2524 | llvm::Value *CondCopy = Bld.CreateAnd( |
| 2525 | Algo1, Bld.CreateICmpUGE(LaneIDArgVal, RemoteLaneOffsetArgVal)); |
| 2526 | |
| 2527 | llvm::BasicBlock *CpyThenBB = CGF.createBasicBlock("then"); |
| 2528 | llvm::BasicBlock *CpyElseBB = CGF.createBasicBlock("else"); |
| 2529 | llvm::BasicBlock *CpyMergeBB = CGF.createBasicBlock("ifcont"); |
| 2530 | Bld.CreateCondBr(CondCopy, CpyThenBB, CpyElseBB); |
| 2531 | |
| 2532 | CGF.EmitBlock(CpyThenBB); |
| 2533 | emitReductionListCopy(ThreadCopy, CGF, ReductionArrayTy, Privates, |
| 2534 | RemoteReduceList, LocalReduceList); |
| 2535 | Bld.CreateBr(CpyMergeBB); |
| 2536 | |
| 2537 | CGF.EmitBlock(CpyElseBB); |
| 2538 | Bld.CreateBr(CpyMergeBB); |
| 2539 | |
| 2540 | CGF.EmitBlock(CpyMergeBB); |
| 2541 | |
| 2542 | CGF.FinishFunction(); |
| 2543 | return Fn; |
| 2544 | } |
| 2545 | |
| 2546 | /// This function emits a helper that copies all the reduction variables from |
| 2547 | /// the team into the provided global buffer for the reduction variables. |
| 2548 | /// |
| 2549 | /// void list_to_global_copy_func(void *buffer, int Idx, void *reduce_data) |
| 2550 | /// For all data entries D in reduce_data: |
| 2551 | /// Copy local D to buffer.D[Idx] |
| 2552 | static llvm::Value *emitListToGlobalCopyFunction( |
| 2553 | CodeGenModule &CGM, ArrayRef<const Expr *> Privates, |
| 2554 | QualType ReductionArrayTy, SourceLocation Loc, |
| 2555 | const RecordDecl *TeamReductionRec, |
| 2556 | const llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> |
| 2557 | &VarFieldMap) { |
| 2558 | ASTContext &C = CGM.getContext(); |
| 2559 | |
| 2560 | // Buffer: global reduction buffer. |
| 2561 | ImplicitParamDecl BufferArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, |
| 2562 | C.VoidPtrTy, ImplicitParamDecl::Other); |
| 2563 | // Idx: index of the buffer. |
| 2564 | ImplicitParamDecl IdxArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.IntTy, |
| 2565 | ImplicitParamDecl::Other); |
| 2566 | // ReduceList: thread local Reduce list. |
| 2567 | ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, |
| 2568 | C.VoidPtrTy, ImplicitParamDecl::Other); |
| 2569 | FunctionArgList Args; |
| 2570 | Args.push_back(&BufferArg); |
| 2571 | Args.push_back(&IdxArg); |
| 2572 | Args.push_back(&ReduceListArg); |
| 2573 | |
| 2574 | const CGFunctionInfo &CGFI = |
| 2575 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); |
| 2576 | auto *Fn = llvm::Function::Create( |
| 2577 | CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage, |
| 2578 | "_omp_reduction_list_to_global_copy_func", &CGM.getModule()); |
| 2579 | CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI); |
| 2580 | Fn->setDoesNotRecurse(); |
| 2581 | CodeGenFunction CGF(CGM); |
| 2582 | CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc); |
| 2583 | |
| 2584 | CGBuilderTy &Bld = CGF.Builder; |
| 2585 | |
| 2586 | Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg); |
| 2587 | Address AddrBufferArg = CGF.GetAddrOfLocalVar(&BufferArg); |
| 2588 | llvm::Type *ElemTy = CGF.ConvertTypeForMem(ReductionArrayTy); |
| 2589 | Address LocalReduceList( |
| 2590 | Bld.CreatePointerBitCastOrAddrSpaceCast( |
| 2591 | CGF.EmitLoadOfScalar(AddrReduceListArg, /*Volatile=*/false, |
| 2592 | C.VoidPtrTy, Loc), |
| 2593 | ElemTy->getPointerTo()), |
| 2594 | ElemTy, CGF.getPointerAlign()); |
| 2595 | QualType StaticTy = C.getRecordType(TeamReductionRec); |
| 2596 | llvm::Type *LLVMReductionsBufferTy = |
| 2597 | CGM.getTypes().ConvertTypeForMem(StaticTy); |
| 2598 | llvm::Value *BufferArrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( |
| 2599 | CGF.EmitLoadOfScalar(AddrBufferArg, /*Volatile=*/false, C.VoidPtrTy, Loc), |
| 2600 | LLVMReductionsBufferTy->getPointerTo()); |
| 2601 | llvm::Value *Idxs[] = {llvm::ConstantInt::getNullValue(CGF.Int32Ty), |
| 2602 | CGF.EmitLoadOfScalar(CGF.GetAddrOfLocalVar(&IdxArg), |
| 2603 | /*Volatile=*/false, C.IntTy, |
| 2604 | Loc)}; |
| 2605 | unsigned Idx = 0; |
| 2606 | for (const Expr *Private : Privates) { |
| 2607 | // Reduce element = LocalReduceList[i] |
| 2608 | Address ElemPtrPtrAddr = Bld.CreateConstArrayGEP(LocalReduceList, Idx); |
| 2609 | llvm::Value *ElemPtrPtr = CGF.EmitLoadOfScalar( |
| 2610 | ElemPtrPtrAddr, /*Volatile=*/false, C.VoidPtrTy, SourceLocation()); |
| 2611 | // elemptr = ((CopyType*)(elemptrptr)) + I |
| 2612 | ElemTy = CGF.ConvertTypeForMem(Private->getType()); |
| 2613 | ElemPtrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( |
| 2614 | ElemPtrPtr, ElemTy->getPointerTo()); |
| 2615 | Address ElemPtr = |
| 2616 | Address(ElemPtrPtr, ElemTy, C.getTypeAlignInChars(Private->getType())); |
| 2617 | const ValueDecl *VD = cast<DeclRefExpr>(Private)->getDecl(); |
| 2618 | // Global = Buffer.VD[Idx]; |
| 2619 | const FieldDecl *FD = VarFieldMap.lookup(VD); |
| 2620 | LValue GlobLVal = CGF.EmitLValueForField( |
| 2621 | CGF.MakeNaturalAlignAddrLValue(BufferArrPtr, StaticTy), FD); |
| 2622 | Address GlobAddr = GlobLVal.getAddress(CGF); |
| 2623 | llvm::Value *BufferPtr = Bld.CreateInBoundsGEP(GlobAddr.getElementType(), |
| 2624 | GlobAddr.getPointer(), Idxs); |
| 2625 | GlobLVal.setAddress(Address(BufferPtr, |
| 2626 | CGF.ConvertTypeForMem(Private->getType()), |
| 2627 | GlobAddr.getAlignment())); |
| 2628 | switch (CGF.getEvaluationKind(Private->getType())) { |
| 2629 | case TEK_Scalar: { |
| 2630 | llvm::Value *V = CGF.EmitLoadOfScalar( |
| 2631 | ElemPtr, /*Volatile=*/false, Private->getType(), Loc, |
| 2632 | LValueBaseInfo(AlignmentSource::Type), TBAAAccessInfo()); |
| 2633 | CGF.EmitStoreOfScalar(V, GlobLVal); |
| 2634 | break; |
| 2635 | } |
| 2636 | case TEK_Complex: { |
| 2637 | CodeGenFunction::ComplexPairTy V = CGF.EmitLoadOfComplex( |
| 2638 | CGF.MakeAddrLValue(ElemPtr, Private->getType()), Loc); |
| 2639 | CGF.EmitStoreOfComplex(V, GlobLVal, /*isInit=*/false); |
| 2640 | break; |
| 2641 | } |
| 2642 | case TEK_Aggregate: |
| 2643 | CGF.EmitAggregateCopy(GlobLVal, |
| 2644 | CGF.MakeAddrLValue(ElemPtr, Private->getType()), |
| 2645 | Private->getType(), AggValueSlot::DoesNotOverlap); |
| 2646 | break; |
| 2647 | } |
| 2648 | ++Idx; |
| 2649 | } |
| 2650 | |
| 2651 | CGF.FinishFunction(); |
| 2652 | return Fn; |
| 2653 | } |
| 2654 | |
| 2655 | /// This function emits a helper that reduces all the reduction variables from |
| 2656 | /// the team into the provided global buffer for the reduction variables. |
| 2657 | /// |
| 2658 | /// void list_to_global_reduce_func(void *buffer, int Idx, void *reduce_data) |
| 2659 | /// void *GlobPtrs[]; |
| 2660 | /// GlobPtrs[0] = (void*)&buffer.D0[Idx]; |
| 2661 | /// ... |
| 2662 | /// GlobPtrs[N] = (void*)&buffer.DN[Idx]; |
| 2663 | /// reduce_function(GlobPtrs, reduce_data); |
| 2664 | static llvm::Value *emitListToGlobalReduceFunction( |
| 2665 | CodeGenModule &CGM, ArrayRef<const Expr *> Privates, |
| 2666 | QualType ReductionArrayTy, SourceLocation Loc, |
| 2667 | const RecordDecl *TeamReductionRec, |
| 2668 | const llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> |
| 2669 | &VarFieldMap, |
| 2670 | llvm::Function *ReduceFn) { |
| 2671 | ASTContext &C = CGM.getContext(); |
| 2672 | |
| 2673 | // Buffer: global reduction buffer. |
| 2674 | ImplicitParamDecl BufferArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, |
| 2675 | C.VoidPtrTy, ImplicitParamDecl::Other); |
| 2676 | // Idx: index of the buffer. |
| 2677 | ImplicitParamDecl IdxArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.IntTy, |
| 2678 | ImplicitParamDecl::Other); |
| 2679 | // ReduceList: thread local Reduce list. |
| 2680 | ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, |
| 2681 | C.VoidPtrTy, ImplicitParamDecl::Other); |
| 2682 | FunctionArgList Args; |
| 2683 | Args.push_back(&BufferArg); |
| 2684 | Args.push_back(&IdxArg); |
| 2685 | Args.push_back(&ReduceListArg); |
| 2686 | |
| 2687 | const CGFunctionInfo &CGFI = |
| 2688 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); |
| 2689 | auto *Fn = llvm::Function::Create( |
| 2690 | CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage, |
| 2691 | "_omp_reduction_list_to_global_reduce_func", &CGM.getModule()); |
| 2692 | CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI); |
| 2693 | Fn->setDoesNotRecurse(); |
| 2694 | CodeGenFunction CGF(CGM); |
| 2695 | CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc); |
| 2696 | |
| 2697 | CGBuilderTy &Bld = CGF.Builder; |
| 2698 | |
| 2699 | Address AddrBufferArg = CGF.GetAddrOfLocalVar(&BufferArg); |
| 2700 | QualType StaticTy = C.getRecordType(TeamReductionRec); |
| 2701 | llvm::Type *LLVMReductionsBufferTy = |
| 2702 | CGM.getTypes().ConvertTypeForMem(StaticTy); |
| 2703 | llvm::Value *BufferArrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( |
| 2704 | CGF.EmitLoadOfScalar(AddrBufferArg, /*Volatile=*/false, C.VoidPtrTy, Loc), |
| 2705 | LLVMReductionsBufferTy->getPointerTo()); |
| 2706 | |
| 2707 | // 1. Build a list of reduction variables. |
| 2708 | // void *RedList[<n>] = {<ReductionVars>[0], ..., <ReductionVars>[<n>-1]}; |
| 2709 | Address ReductionList = |
| 2710 | CGF.CreateMemTemp(ReductionArrayTy, ".omp.reduction.red_list"); |
| 2711 | auto IPriv = Privates.begin(); |
| 2712 | llvm::Value *Idxs[] = {llvm::ConstantInt::getNullValue(CGF.Int32Ty), |
| 2713 | CGF.EmitLoadOfScalar(CGF.GetAddrOfLocalVar(&IdxArg), |
| 2714 | /*Volatile=*/false, C.IntTy, |
| 2715 | Loc)}; |
| 2716 | unsigned Idx = 0; |
| 2717 | for (unsigned I = 0, E = Privates.size(); I < E; ++I, ++IPriv, ++Idx) { |
| 2718 | Address Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx); |
| 2719 | // Global = Buffer.VD[Idx]; |
| 2720 | const ValueDecl *VD = cast<DeclRefExpr>(*IPriv)->getDecl(); |
| 2721 | const FieldDecl *FD = VarFieldMap.lookup(VD); |
| 2722 | LValue GlobLVal = CGF.EmitLValueForField( |
| 2723 | CGF.MakeNaturalAlignAddrLValue(BufferArrPtr, StaticTy), FD); |
| 2724 | Address GlobAddr = GlobLVal.getAddress(CGF); |
| 2725 | llvm::Value *BufferPtr = Bld.CreateInBoundsGEP( |
| 2726 | GlobAddr.getElementType(), GlobAddr.getPointer(), Idxs); |
| 2727 | llvm::Value *Ptr = CGF.EmitCastToVoidPtr(BufferPtr); |
| 2728 | CGF.EmitStoreOfScalar(Ptr, Elem, /*Volatile=*/false, C.VoidPtrTy); |
| 2729 | if ((*IPriv)->getType()->isVariablyModifiedType()) { |
| 2730 | // Store array size. |
| 2731 | ++Idx; |
| 2732 | Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx); |
| 2733 | llvm::Value *Size = CGF.Builder.CreateIntCast( |
| 2734 | CGF.getVLASize( |
| 2735 | CGF.getContext().getAsVariableArrayType((*IPriv)->getType())) |
| 2736 | .NumElts, |
| 2737 | CGF.SizeTy, /*isSigned=*/false); |
| 2738 | CGF.Builder.CreateStore(CGF.Builder.CreateIntToPtr(Size, CGF.VoidPtrTy), |
| 2739 | Elem); |
| 2740 | } |
| 2741 | } |
| 2742 | |
| 2743 | // Call reduce_function(GlobalReduceList, ReduceList) |
| 2744 | llvm::Value *GlobalReduceList = |
| 2745 | CGF.EmitCastToVoidPtr(ReductionList.getPointer()); |
| 2746 | Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg); |
| 2747 | llvm::Value *ReducedPtr = CGF.EmitLoadOfScalar( |
| 2748 | AddrReduceListArg, /*Volatile=*/false, C.VoidPtrTy, Loc); |
| 2749 | CGM.getOpenMPRuntime().emitOutlinedFunctionCall( |
| 2750 | CGF, Loc, ReduceFn, {GlobalReduceList, ReducedPtr}); |
| 2751 | CGF.FinishFunction(); |
| 2752 | return Fn; |
| 2753 | } |
| 2754 | |
| 2755 | /// This function emits a helper that copies all the reduction variables from |
| 2756 | /// the team into the provided global buffer for the reduction variables. |
| 2757 | /// |
| 2758 | /// void list_to_global_copy_func(void *buffer, int Idx, void *reduce_data) |
| 2759 | /// For all data entries D in reduce_data: |
| 2760 | /// Copy buffer.D[Idx] to local D; |
| 2761 | static llvm::Value *emitGlobalToListCopyFunction( |
| 2762 | CodeGenModule &CGM, ArrayRef<const Expr *> Privates, |
| 2763 | QualType ReductionArrayTy, SourceLocation Loc, |
| 2764 | const RecordDecl *TeamReductionRec, |
| 2765 | const llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> |
| 2766 | &VarFieldMap) { |
| 2767 | ASTContext &C = CGM.getContext(); |
| 2768 | |
| 2769 | // Buffer: global reduction buffer. |
| 2770 | ImplicitParamDecl BufferArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, |
| 2771 | C.VoidPtrTy, ImplicitParamDecl::Other); |
| 2772 | // Idx: index of the buffer. |
| 2773 | ImplicitParamDecl IdxArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.IntTy, |
| 2774 | ImplicitParamDecl::Other); |
| 2775 | // ReduceList: thread local Reduce list. |
| 2776 | ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, |
| 2777 | C.VoidPtrTy, ImplicitParamDecl::Other); |
| 2778 | FunctionArgList Args; |
| 2779 | Args.push_back(&BufferArg); |
| 2780 | Args.push_back(&IdxArg); |
| 2781 | Args.push_back(&ReduceListArg); |
| 2782 | |
| 2783 | const CGFunctionInfo &CGFI = |
| 2784 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); |
| 2785 | auto *Fn = llvm::Function::Create( |
| 2786 | CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage, |
| 2787 | "_omp_reduction_global_to_list_copy_func", &CGM.getModule()); |
| 2788 | CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI); |
| 2789 | Fn->setDoesNotRecurse(); |
| 2790 | CodeGenFunction CGF(CGM); |
| 2791 | CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc); |
| 2792 | |
| 2793 | CGBuilderTy &Bld = CGF.Builder; |
| 2794 | |
| 2795 | Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg); |
| 2796 | Address AddrBufferArg = CGF.GetAddrOfLocalVar(&BufferArg); |
| 2797 | llvm::Type *ElemTy = CGF.ConvertTypeForMem(ReductionArrayTy); |
| 2798 | Address LocalReduceList( |
| 2799 | Bld.CreatePointerBitCastOrAddrSpaceCast( |
| 2800 | CGF.EmitLoadOfScalar(AddrReduceListArg, /*Volatile=*/false, |
| 2801 | C.VoidPtrTy, Loc), |
| 2802 | ElemTy->getPointerTo()), |
| 2803 | ElemTy, CGF.getPointerAlign()); |
| 2804 | QualType StaticTy = C.getRecordType(TeamReductionRec); |
| 2805 | llvm::Type *LLVMReductionsBufferTy = |
| 2806 | CGM.getTypes().ConvertTypeForMem(StaticTy); |
| 2807 | llvm::Value *BufferArrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( |
| 2808 | CGF.EmitLoadOfScalar(AddrBufferArg, /*Volatile=*/false, C.VoidPtrTy, Loc), |
| 2809 | LLVMReductionsBufferTy->getPointerTo()); |
| 2810 | |
| 2811 | llvm::Value *Idxs[] = {llvm::ConstantInt::getNullValue(CGF.Int32Ty), |
| 2812 | CGF.EmitLoadOfScalar(CGF.GetAddrOfLocalVar(&IdxArg), |
| 2813 | /*Volatile=*/false, C.IntTy, |
| 2814 | Loc)}; |
| 2815 | unsigned Idx = 0; |
| 2816 | for (const Expr *Private : Privates) { |
| 2817 | // Reduce element = LocalReduceList[i] |
| 2818 | Address ElemPtrPtrAddr = Bld.CreateConstArrayGEP(LocalReduceList, Idx); |
| 2819 | llvm::Value *ElemPtrPtr = CGF.EmitLoadOfScalar( |
| 2820 | ElemPtrPtrAddr, /*Volatile=*/false, C.VoidPtrTy, SourceLocation()); |
| 2821 | // elemptr = ((CopyType*)(elemptrptr)) + I |
| 2822 | ElemTy = CGF.ConvertTypeForMem(Private->getType()); |
| 2823 | ElemPtrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( |
| 2824 | ElemPtrPtr, ElemTy->getPointerTo()); |
| 2825 | Address ElemPtr = |
| 2826 | Address(ElemPtrPtr, ElemTy, C.getTypeAlignInChars(Private->getType())); |
| 2827 | const ValueDecl *VD = cast<DeclRefExpr>(Private)->getDecl(); |
| 2828 | // Global = Buffer.VD[Idx]; |
| 2829 | const FieldDecl *FD = VarFieldMap.lookup(VD); |
| 2830 | LValue GlobLVal = CGF.EmitLValueForField( |
| 2831 | CGF.MakeNaturalAlignAddrLValue(BufferArrPtr, StaticTy), FD); |
| 2832 | Address GlobAddr = GlobLVal.getAddress(CGF); |
| 2833 | llvm::Value *BufferPtr = Bld.CreateInBoundsGEP(GlobAddr.getElementType(), |
| 2834 | GlobAddr.getPointer(), Idxs); |
| 2835 | GlobLVal.setAddress(Address(BufferPtr, |
| 2836 | CGF.ConvertTypeForMem(Private->getType()), |
| 2837 | GlobAddr.getAlignment())); |
| 2838 | switch (CGF.getEvaluationKind(Private->getType())) { |
| 2839 | case TEK_Scalar: { |
| 2840 | llvm::Value *V = CGF.EmitLoadOfScalar(GlobLVal, Loc); |
| 2841 | CGF.EmitStoreOfScalar(V, ElemPtr, /*Volatile=*/false, Private->getType(), |
| 2842 | LValueBaseInfo(AlignmentSource::Type), |
| 2843 | TBAAAccessInfo()); |
| 2844 | break; |
| 2845 | } |
| 2846 | case TEK_Complex: { |
| 2847 | CodeGenFunction::ComplexPairTy V = CGF.EmitLoadOfComplex(GlobLVal, Loc); |
| 2848 | CGF.EmitStoreOfComplex(V, CGF.MakeAddrLValue(ElemPtr, Private->getType()), |
| 2849 | /*isInit=*/false); |
| 2850 | break; |
| 2851 | } |
| 2852 | case TEK_Aggregate: |
| 2853 | CGF.EmitAggregateCopy(CGF.MakeAddrLValue(ElemPtr, Private->getType()), |
| 2854 | GlobLVal, Private->getType(), |
| 2855 | AggValueSlot::DoesNotOverlap); |
| 2856 | break; |
| 2857 | } |
| 2858 | ++Idx; |
| 2859 | } |
| 2860 | |
| 2861 | CGF.FinishFunction(); |
| 2862 | return Fn; |
| 2863 | } |
| 2864 | |
| 2865 | /// This function emits a helper that reduces all the reduction variables from |
| 2866 | /// the team into the provided global buffer for the reduction variables. |
| 2867 | /// |
| 2868 | /// void global_to_list_reduce_func(void *buffer, int Idx, void *reduce_data) |
| 2869 | /// void *GlobPtrs[]; |
| 2870 | /// GlobPtrs[0] = (void*)&buffer.D0[Idx]; |
| 2871 | /// ... |
| 2872 | /// GlobPtrs[N] = (void*)&buffer.DN[Idx]; |
| 2873 | /// reduce_function(reduce_data, GlobPtrs); |
| 2874 | static llvm::Value *emitGlobalToListReduceFunction( |
| 2875 | CodeGenModule &CGM, ArrayRef<const Expr *> Privates, |
| 2876 | QualType ReductionArrayTy, SourceLocation Loc, |
| 2877 | const RecordDecl *TeamReductionRec, |
| 2878 | const llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> |
| 2879 | &VarFieldMap, |
| 2880 | llvm::Function *ReduceFn) { |
| 2881 | ASTContext &C = CGM.getContext(); |
| 2882 | |
| 2883 | // Buffer: global reduction buffer. |
| 2884 | ImplicitParamDecl BufferArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, |
| 2885 | C.VoidPtrTy, ImplicitParamDecl::Other); |
| 2886 | // Idx: index of the buffer. |
| 2887 | ImplicitParamDecl IdxArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.IntTy, |
| 2888 | ImplicitParamDecl::Other); |
| 2889 | // ReduceList: thread local Reduce list. |
| 2890 | ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, |
| 2891 | C.VoidPtrTy, ImplicitParamDecl::Other); |
| 2892 | FunctionArgList Args; |
| 2893 | Args.push_back(&BufferArg); |
| 2894 | Args.push_back(&IdxArg); |
| 2895 | Args.push_back(&ReduceListArg); |
| 2896 | |
| 2897 | const CGFunctionInfo &CGFI = |
| 2898 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); |
| 2899 | auto *Fn = llvm::Function::Create( |
| 2900 | CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage, |
| 2901 | "_omp_reduction_global_to_list_reduce_func", &CGM.getModule()); |
| 2902 | CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI); |
| 2903 | Fn->setDoesNotRecurse(); |
| 2904 | CodeGenFunction CGF(CGM); |
| 2905 | CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc); |
| 2906 | |
| 2907 | CGBuilderTy &Bld = CGF.Builder; |
| 2908 | |
| 2909 | Address AddrBufferArg = CGF.GetAddrOfLocalVar(&BufferArg); |
| 2910 | QualType StaticTy = C.getRecordType(TeamReductionRec); |
| 2911 | llvm::Type *LLVMReductionsBufferTy = |
| 2912 | CGM.getTypes().ConvertTypeForMem(StaticTy); |
| 2913 | llvm::Value *BufferArrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( |
| 2914 | CGF.EmitLoadOfScalar(AddrBufferArg, /*Volatile=*/false, C.VoidPtrTy, Loc), |
| 2915 | LLVMReductionsBufferTy->getPointerTo()); |
| 2916 | |
| 2917 | // 1. Build a list of reduction variables. |
| 2918 | // void *RedList[<n>] = {<ReductionVars>[0], ..., <ReductionVars>[<n>-1]}; |
| 2919 | Address ReductionList = |
| 2920 | CGF.CreateMemTemp(ReductionArrayTy, ".omp.reduction.red_list"); |
| 2921 | auto IPriv = Privates.begin(); |
| 2922 | llvm::Value *Idxs[] = {llvm::ConstantInt::getNullValue(CGF.Int32Ty), |
| 2923 | CGF.EmitLoadOfScalar(CGF.GetAddrOfLocalVar(&IdxArg), |
| 2924 | /*Volatile=*/false, C.IntTy, |
| 2925 | Loc)}; |
| 2926 | unsigned Idx = 0; |
| 2927 | for (unsigned I = 0, E = Privates.size(); I < E; ++I, ++IPriv, ++Idx) { |
| 2928 | Address Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx); |
| 2929 | // Global = Buffer.VD[Idx]; |
| 2930 | const ValueDecl *VD = cast<DeclRefExpr>(*IPriv)->getDecl(); |
| 2931 | const FieldDecl *FD = VarFieldMap.lookup(VD); |
| 2932 | LValue GlobLVal = CGF.EmitLValueForField( |
| 2933 | CGF.MakeNaturalAlignAddrLValue(BufferArrPtr, StaticTy), FD); |
| 2934 | Address GlobAddr = GlobLVal.getAddress(CGF); |
| 2935 | llvm::Value *BufferPtr = Bld.CreateInBoundsGEP( |
| 2936 | GlobAddr.getElementType(), GlobAddr.getPointer(), Idxs); |
| 2937 | llvm::Value *Ptr = CGF.EmitCastToVoidPtr(BufferPtr); |
| 2938 | CGF.EmitStoreOfScalar(Ptr, Elem, /*Volatile=*/false, C.VoidPtrTy); |
| 2939 | if ((*IPriv)->getType()->isVariablyModifiedType()) { |
| 2940 | // Store array size. |
| 2941 | ++Idx; |
| 2942 | Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx); |
| 2943 | llvm::Value *Size = CGF.Builder.CreateIntCast( |
| 2944 | CGF.getVLASize( |
| 2945 | CGF.getContext().getAsVariableArrayType((*IPriv)->getType())) |
| 2946 | .NumElts, |
| 2947 | CGF.SizeTy, /*isSigned=*/false); |
| 2948 | CGF.Builder.CreateStore(CGF.Builder.CreateIntToPtr(Size, CGF.VoidPtrTy), |
| 2949 | Elem); |
| 2950 | } |
| 2951 | } |
| 2952 | |
| 2953 | // Call reduce_function(ReduceList, GlobalReduceList) |
| 2954 | llvm::Value *GlobalReduceList = |
| 2955 | CGF.EmitCastToVoidPtr(ReductionList.getPointer()); |
| 2956 | Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg); |
| 2957 | llvm::Value *ReducedPtr = CGF.EmitLoadOfScalar( |
| 2958 | AddrReduceListArg, /*Volatile=*/false, C.VoidPtrTy, Loc); |
| 2959 | CGM.getOpenMPRuntime().emitOutlinedFunctionCall( |
| 2960 | CGF, Loc, ReduceFn, {ReducedPtr, GlobalReduceList}); |
| 2961 | CGF.FinishFunction(); |
| 2962 | return Fn; |
| 2963 | } |
| 2964 | |
| 2965 | /// |
| 2966 | /// Design of OpenMP reductions on the GPU |
| 2967 | /// |
| 2968 | /// Consider a typical OpenMP program with one or more reduction |
| 2969 | /// clauses: |
| 2970 | /// |
| 2971 | /// float foo; |
| 2972 | /// double bar; |
| 2973 | /// #pragma omp target teams distribute parallel for \ |
| 2974 | /// reduction(+:foo) reduction(*:bar) |
| 2975 | /// for (int i = 0; i < N; i++) { |
| 2976 | /// foo += A[i]; bar *= B[i]; |
| 2977 | /// } |
| 2978 | /// |
| 2979 | /// where 'foo' and 'bar' are reduced across all OpenMP threads in |
| 2980 | /// all teams. In our OpenMP implementation on the NVPTX device an |
| 2981 | /// OpenMP team is mapped to a CUDA threadblock and OpenMP threads |
| 2982 | /// within a team are mapped to CUDA threads within a threadblock. |
| 2983 | /// Our goal is to efficiently aggregate values across all OpenMP |
| 2984 | /// threads such that: |
| 2985 | /// |
| 2986 | /// - the compiler and runtime are logically concise, and |
| 2987 | /// - the reduction is performed efficiently in a hierarchical |
| 2988 | /// manner as follows: within OpenMP threads in the same warp, |
| 2989 | /// across warps in a threadblock, and finally across teams on |
| 2990 | /// the NVPTX device. |
| 2991 | /// |
| 2992 | /// Introduction to Decoupling |
| 2993 | /// |
| 2994 | /// We would like to decouple the compiler and the runtime so that the |
| 2995 | /// latter is ignorant of the reduction variables (number, data types) |
| 2996 | /// and the reduction operators. This allows a simpler interface |
| 2997 | /// and implementation while still attaining good performance. |
| 2998 | /// |
| 2999 | /// Pseudocode for the aforementioned OpenMP program generated by the |
| 3000 | /// compiler is as follows: |
| 3001 | /// |
| 3002 | /// 1. Create private copies of reduction variables on each OpenMP |
| 3003 | /// thread: 'foo_private', 'bar_private' |
| 3004 | /// 2. Each OpenMP thread reduces the chunk of 'A' and 'B' assigned |
| 3005 | /// to it and writes the result in 'foo_private' and 'bar_private' |
| 3006 | /// respectively. |
| 3007 | /// 3. Call the OpenMP runtime on the GPU to reduce within a team |
| 3008 | /// and store the result on the team master: |
| 3009 | /// |
| 3010 | /// __kmpc_nvptx_parallel_reduce_nowait_v2(..., |
| 3011 | /// reduceData, shuffleReduceFn, interWarpCpyFn) |
| 3012 | /// |
| 3013 | /// where: |
| 3014 | /// struct ReduceData { |
| 3015 | /// double *foo; |
| 3016 | /// double *bar; |
| 3017 | /// } reduceData |
| 3018 | /// reduceData.foo = &foo_private |
| 3019 | /// reduceData.bar = &bar_private |
| 3020 | /// |
| 3021 | /// 'shuffleReduceFn' and 'interWarpCpyFn' are pointers to two |
| 3022 | /// auxiliary functions generated by the compiler that operate on |
| 3023 | /// variables of type 'ReduceData'. They aid the runtime perform |
| 3024 | /// algorithmic steps in a data agnostic manner. |
| 3025 | /// |
| 3026 | /// 'shuffleReduceFn' is a pointer to a function that reduces data |
| 3027 | /// of type 'ReduceData' across two OpenMP threads (lanes) in the |
| 3028 | /// same warp. It takes the following arguments as input: |
| 3029 | /// |
| 3030 | /// a. variable of type 'ReduceData' on the calling lane, |
| 3031 | /// b. its lane_id, |
| 3032 | /// c. an offset relative to the current lane_id to generate a |
| 3033 | /// remote_lane_id. The remote lane contains the second |
| 3034 | /// variable of type 'ReduceData' that is to be reduced. |
| 3035 | /// d. an algorithm version parameter determining which reduction |
| 3036 | /// algorithm to use. |
| 3037 | /// |
| 3038 | /// 'shuffleReduceFn' retrieves data from the remote lane using |
| 3039 | /// efficient GPU shuffle intrinsics and reduces, using the |
| 3040 | /// algorithm specified by the 4th parameter, the two operands |
| 3041 | /// element-wise. The result is written to the first operand. |
| 3042 | /// |
| 3043 | /// Different reduction algorithms are implemented in different |
| 3044 | /// runtime functions, all calling 'shuffleReduceFn' to perform |
| 3045 | /// the essential reduction step. Therefore, based on the 4th |
| 3046 | /// parameter, this function behaves slightly differently to |
| 3047 | /// cooperate with the runtime to ensure correctness under |
| 3048 | /// different circumstances. |
| 3049 | /// |
| 3050 | /// 'InterWarpCpyFn' is a pointer to a function that transfers |
| 3051 | /// reduced variables across warps. It tunnels, through CUDA |
| 3052 | /// shared memory, the thread-private data of type 'ReduceData' |
| 3053 | /// from lane 0 of each warp to a lane in the first warp. |
| 3054 | /// 4. Call the OpenMP runtime on the GPU to reduce across teams. |
| 3055 | /// The last team writes the global reduced value to memory. |
| 3056 | /// |
| 3057 | /// ret = __kmpc_nvptx_teams_reduce_nowait(..., |
| 3058 | /// reduceData, shuffleReduceFn, interWarpCpyFn, |
| 3059 | /// scratchpadCopyFn, loadAndReduceFn) |
| 3060 | /// |
| 3061 | /// 'scratchpadCopyFn' is a helper that stores reduced |
| 3062 | /// data from the team master to a scratchpad array in |
| 3063 | /// global memory. |
| 3064 | /// |
| 3065 | /// 'loadAndReduceFn' is a helper that loads data from |
| 3066 | /// the scratchpad array and reduces it with the input |
| 3067 | /// operand. |
| 3068 | /// |
| 3069 | /// These compiler generated functions hide address |
| 3070 | /// calculation and alignment information from the runtime. |
| 3071 | /// 5. if ret == 1: |
| 3072 | /// The team master of the last team stores the reduced |
| 3073 | /// result to the globals in memory. |
| 3074 | /// foo += reduceData.foo; bar *= reduceData.bar |
| 3075 | /// |
| 3076 | /// |
| 3077 | /// Warp Reduction Algorithms |
| 3078 | /// |
| 3079 | /// On the warp level, we have three algorithms implemented in the |
| 3080 | /// OpenMP runtime depending on the number of active lanes: |
| 3081 | /// |
| 3082 | /// Full Warp Reduction |
| 3083 | /// |
| 3084 | /// The reduce algorithm within a warp where all lanes are active |
| 3085 | /// is implemented in the runtime as follows: |
| 3086 | /// |
| 3087 | /// full_warp_reduce(void *reduce_data, |
| 3088 | /// kmp_ShuffleReductFctPtr ShuffleReduceFn) { |
| 3089 | /// for (int offset = WARPSIZE/2; offset > 0; offset /= 2) |
| 3090 | /// ShuffleReduceFn(reduce_data, 0, offset, 0); |
| 3091 | /// } |
| 3092 | /// |
| 3093 | /// The algorithm completes in log(2, WARPSIZE) steps. |
| 3094 | /// |
| 3095 | /// 'ShuffleReduceFn' is used here with lane_id set to 0 because it is |
| 3096 | /// not used therefore we save instructions by not retrieving lane_id |
| 3097 | /// from the corresponding special registers. The 4th parameter, which |
| 3098 | /// represents the version of the algorithm being used, is set to 0 to |
| 3099 | /// signify full warp reduction. |
| 3100 | /// |
| 3101 | /// In this version, 'ShuffleReduceFn' behaves, per element, as follows: |
| 3102 | /// |
| 3103 | /// #reduce_elem refers to an element in the local lane's data structure |
| 3104 | /// #remote_elem is retrieved from a remote lane |
| 3105 | /// remote_elem = shuffle_down(reduce_elem, offset, WARPSIZE); |
| 3106 | /// reduce_elem = reduce_elem REDUCE_OP remote_elem; |
| 3107 | /// |
| 3108 | /// Contiguous Partial Warp Reduction |
| 3109 | /// |
| 3110 | /// This reduce algorithm is used within a warp where only the first |
| 3111 | /// 'n' (n <= WARPSIZE) lanes are active. It is typically used when the |
| 3112 | /// number of OpenMP threads in a parallel region is not a multiple of |
| 3113 | /// WARPSIZE. The algorithm is implemented in the runtime as follows: |
| 3114 | /// |
| 3115 | /// void |
| 3116 | /// contiguous_partial_reduce(void *reduce_data, |
| 3117 | /// kmp_ShuffleReductFctPtr ShuffleReduceFn, |
| 3118 | /// int size, int lane_id) { |
| 3119 | /// int curr_size; |
| 3120 | /// int offset; |
| 3121 | /// curr_size = size; |
| 3122 | /// mask = curr_size/2; |
| 3123 | /// while (offset>0) { |
| 3124 | /// ShuffleReduceFn(reduce_data, lane_id, offset, 1); |
| 3125 | /// curr_size = (curr_size+1)/2; |
| 3126 | /// offset = curr_size/2; |
| 3127 | /// } |
| 3128 | /// } |
| 3129 | /// |
| 3130 | /// In this version, 'ShuffleReduceFn' behaves, per element, as follows: |
| 3131 | /// |
| 3132 | /// remote_elem = shuffle_down(reduce_elem, offset, WARPSIZE); |
| 3133 | /// if (lane_id < offset) |
| 3134 | /// reduce_elem = reduce_elem REDUCE_OP remote_elem |
| 3135 | /// else |
| 3136 | /// reduce_elem = remote_elem |
| 3137 | /// |
| 3138 | /// This algorithm assumes that the data to be reduced are located in a |
| 3139 | /// contiguous subset of lanes starting from the first. When there is |
| 3140 | /// an odd number of active lanes, the data in the last lane is not |
| 3141 | /// aggregated with any other lane's dat but is instead copied over. |
| 3142 | /// |
| 3143 | /// Dispersed Partial Warp Reduction |
| 3144 | /// |
| 3145 | /// This algorithm is used within a warp when any discontiguous subset of |
| 3146 | /// lanes are active. It is used to implement the reduction operation |
| 3147 | /// across lanes in an OpenMP simd region or in a nested parallel region. |
| 3148 | /// |
| 3149 | /// void |
| 3150 | /// dispersed_partial_reduce(void *reduce_data, |
| 3151 | /// kmp_ShuffleReductFctPtr ShuffleReduceFn) { |
| 3152 | /// int size, remote_id; |
| 3153 | /// int logical_lane_id = number_of_active_lanes_before_me() * 2; |
| 3154 | /// do { |
| 3155 | /// remote_id = next_active_lane_id_right_after_me(); |
| 3156 | /// # the above function returns 0 of no active lane |
| 3157 | /// # is present right after the current lane. |
| 3158 | /// size = number_of_active_lanes_in_this_warp(); |
| 3159 | /// logical_lane_id /= 2; |
| 3160 | /// ShuffleReduceFn(reduce_data, logical_lane_id, |
| 3161 | /// remote_id-1-threadIdx.x, 2); |
| 3162 | /// } while (logical_lane_id % 2 == 0 && size > 1); |
| 3163 | /// } |
| 3164 | /// |
| 3165 | /// There is no assumption made about the initial state of the reduction. |
| 3166 | /// Any number of lanes (>=1) could be active at any position. The reduction |
| 3167 | /// result is returned in the first active lane. |
| 3168 | /// |
| 3169 | /// In this version, 'ShuffleReduceFn' behaves, per element, as follows: |
| 3170 | /// |
| 3171 | /// remote_elem = shuffle_down(reduce_elem, offset, WARPSIZE); |
| 3172 | /// if (lane_id % 2 == 0 && offset > 0) |
| 3173 | /// reduce_elem = reduce_elem REDUCE_OP remote_elem |
| 3174 | /// else |
| 3175 | /// reduce_elem = remote_elem |
| 3176 | /// |
| 3177 | /// |
| 3178 | /// Intra-Team Reduction |
| 3179 | /// |
| 3180 | /// This function, as implemented in the runtime call |
| 3181 | /// '__kmpc_nvptx_parallel_reduce_nowait_v2', aggregates data across OpenMP |
| 3182 | /// threads in a team. It first reduces within a warp using the |
| 3183 | /// aforementioned algorithms. We then proceed to gather all such |
| 3184 | /// reduced values at the first warp. |
| 3185 | /// |
| 3186 | /// The runtime makes use of the function 'InterWarpCpyFn', which copies |
| 3187 | /// data from each of the "warp master" (zeroth lane of each warp, where |
| 3188 | /// warp-reduced data is held) to the zeroth warp. This step reduces (in |
| 3189 | /// a mathematical sense) the problem of reduction across warp masters in |
| 3190 | /// a block to the problem of warp reduction. |
| 3191 | /// |
| 3192 | /// |
| 3193 | /// Inter-Team Reduction |
| 3194 | /// |
| 3195 | /// Once a team has reduced its data to a single value, it is stored in |
| 3196 | /// a global scratchpad array. Since each team has a distinct slot, this |
| 3197 | /// can be done without locking. |
| 3198 | /// |
| 3199 | /// The last team to write to the scratchpad array proceeds to reduce the |
| 3200 | /// scratchpad array. One or more workers in the last team use the helper |
| 3201 | /// 'loadAndReduceDataFn' to load and reduce values from the array, i.e., |
| 3202 | /// the k'th worker reduces every k'th element. |
| 3203 | /// |
| 3204 | /// Finally, a call is made to '__kmpc_nvptx_parallel_reduce_nowait_v2' to |
| 3205 | /// reduce across workers and compute a globally reduced value. |
| 3206 | /// |
| 3207 | void CGOpenMPRuntimeGPU::emitReduction( |
| 3208 | CodeGenFunction &CGF, SourceLocation Loc, ArrayRef<const Expr *> Privates, |
| 3209 | ArrayRef<const Expr *> LHSExprs, ArrayRef<const Expr *> RHSExprs, |
| 3210 | ArrayRef<const Expr *> ReductionOps, ReductionOptionsTy Options) { |
| 3211 | if (!CGF.HaveInsertPoint()) |
| 3212 | return; |
| 3213 | |
| 3214 | bool ParallelReduction = isOpenMPParallelDirective(Options.ReductionKind); |
| 3215 | #ifndef NDEBUG |
| 3216 | bool TeamsReduction = isOpenMPTeamsDirective(Options.ReductionKind); |
| 3217 | #endif |
| 3218 | |
| 3219 | if (Options.SimpleReduction) { |
| 3220 | assert(!TeamsReduction && !ParallelReduction &&(static_cast <bool> (!TeamsReduction && !ParallelReduction && "Invalid reduction selection in emitReduction.") ? void (0) : __assert_fail ("!TeamsReduction && !ParallelReduction && \"Invalid reduction selection in emitReduction.\"" , "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 3221, __extension__ __PRETTY_FUNCTION__)) |
| 3221 | "Invalid reduction selection in emitReduction.")(static_cast <bool> (!TeamsReduction && !ParallelReduction && "Invalid reduction selection in emitReduction.") ? void (0) : __assert_fail ("!TeamsReduction && !ParallelReduction && \"Invalid reduction selection in emitReduction.\"" , "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 3221, __extension__ __PRETTY_FUNCTION__)); |
| 3222 | CGOpenMPRuntime::emitReduction(CGF, Loc, Privates, LHSExprs, RHSExprs, |
| 3223 | ReductionOps, Options); |
| 3224 | return; |
| 3225 | } |
| 3226 | |
| 3227 | assert((TeamsReduction || ParallelReduction) &&(static_cast <bool> ((TeamsReduction || ParallelReduction ) && "Invalid reduction selection in emitReduction.") ? void (0) : __assert_fail ("(TeamsReduction || ParallelReduction) && \"Invalid reduction selection in emitReduction.\"" , "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 3228, __extension__ __PRETTY_FUNCTION__)) |
| 3228 | "Invalid reduction selection in emitReduction.")(static_cast <bool> ((TeamsReduction || ParallelReduction ) && "Invalid reduction selection in emitReduction.") ? void (0) : __assert_fail ("(TeamsReduction || ParallelReduction) && \"Invalid reduction selection in emitReduction.\"" , "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 3228, __extension__ __PRETTY_FUNCTION__)); |
| 3229 | |
| 3230 | // Build res = __kmpc_reduce{_nowait}(<gtid>, <n>, sizeof(RedList), |
| 3231 | // RedList, shuffle_reduce_func, interwarp_copy_func); |
| 3232 | // or |
| 3233 | // Build res = __kmpc_reduce_teams_nowait_simple(<loc>, <gtid>, <lck>); |
| 3234 | llvm::Value *RTLoc = emitUpdateLocation(CGF, Loc); |
| 3235 | llvm::Value *ThreadId = getThreadID(CGF, Loc); |
| 3236 | |
| 3237 | llvm::Value *Res; |
| 3238 | ASTContext &C = CGM.getContext(); |
| 3239 | // 1. Build a list of reduction variables. |
| 3240 | // void *RedList[<n>] = {<ReductionVars>[0], ..., <ReductionVars>[<n>-1]}; |
| 3241 | auto Size = RHSExprs.size(); |
| 3242 | for (const Expr *E : Privates) { |
| 3243 | if (E->getType()->isVariablyModifiedType()) |
| 3244 | // Reserve place for array size. |
| 3245 | ++Size; |
| 3246 | } |
| 3247 | llvm::APInt ArraySize(/*unsigned int numBits=*/32, Size); |
| 3248 | QualType ReductionArrayTy = |
| 3249 | C.getConstantArrayType(C.VoidPtrTy, ArraySize, nullptr, ArrayType::Normal, |
| 3250 | /*IndexTypeQuals=*/0); |
| 3251 | Address ReductionList = |
| 3252 | CGF.CreateMemTemp(ReductionArrayTy, ".omp.reduction.red_list"); |
| 3253 | auto IPriv = Privates.begin(); |
| 3254 | unsigned Idx = 0; |
| 3255 | for (unsigned I = 0, E = RHSExprs.size(); I < E; ++I, ++IPriv, ++Idx) { |
| 3256 | Address Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx); |
| 3257 | CGF.Builder.CreateStore( |
| 3258 | CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( |
| 3259 | CGF.EmitLValue(RHSExprs[I]).getPointer(CGF), CGF.VoidPtrTy), |
| 3260 | Elem); |
| 3261 | if ((*IPriv)->getType()->isVariablyModifiedType()) { |
| 3262 | // Store array size. |
| 3263 | ++Idx; |
| 3264 | Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx); |
| 3265 | llvm::Value *Size = CGF.Builder.CreateIntCast( |
| 3266 | CGF.getVLASize( |
| 3267 | CGF.getContext().getAsVariableArrayType((*IPriv)->getType())) |
| 3268 | .NumElts, |
| 3269 | CGF.SizeTy, /*isSigned=*/false); |
| 3270 | CGF.Builder.CreateStore(CGF.Builder.CreateIntToPtr(Size, CGF.VoidPtrTy), |
| 3271 | Elem); |
| 3272 | } |
| 3273 | } |
| 3274 | |
| 3275 | llvm::Value *RL = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( |
| 3276 | ReductionList.getPointer(), CGF.VoidPtrTy); |
| 3277 | llvm::Function *ReductionFn = |
| 3278 | emitReductionFunction(Loc, CGF.ConvertTypeForMem(ReductionArrayTy), |
| 3279 | Privates, LHSExprs, RHSExprs, ReductionOps); |
| 3280 | llvm::Value *ReductionArrayTySize = CGF.getTypeSize(ReductionArrayTy); |
| 3281 | llvm::Function *ShuffleAndReduceFn = emitShuffleAndReduceFunction( |
| 3282 | CGM, Privates, ReductionArrayTy, ReductionFn, Loc); |
| 3283 | llvm::Value *InterWarpCopyFn = |
| 3284 | emitInterWarpCopyFunction(CGM, Privates, ReductionArrayTy, Loc); |
| 3285 | |
| 3286 | if (ParallelReduction) { |
| 3287 | llvm::Value *Args[] = {RTLoc, |
| 3288 | ThreadId, |
| 3289 | CGF.Builder.getInt32(RHSExprs.size()), |
| 3290 | ReductionArrayTySize, |
| 3291 | RL, |
| 3292 | ShuffleAndReduceFn, |
| 3293 | InterWarpCopyFn}; |
| 3294 | |
| 3295 | Res = CGF.EmitRuntimeCall( |
| 3296 | OMPBuilder.getOrCreateRuntimeFunction( |
| 3297 | CGM.getModule(), OMPRTL___kmpc_nvptx_parallel_reduce_nowait_v2), |
| 3298 | Args); |
| 3299 | } else { |
| 3300 | assert(TeamsReduction && "expected teams reduction.")(static_cast <bool> (TeamsReduction && "expected teams reduction." ) ? void (0) : __assert_fail ("TeamsReduction && \"expected teams reduction.\"" , "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 3300, __extension__ __PRETTY_FUNCTION__)); |
| 3301 | llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> VarFieldMap; |
| 3302 | llvm::SmallVector<const ValueDecl *, 4> PrivatesReductions(Privates.size()); |
| 3303 | int Cnt = 0; |
| 3304 | for (const Expr *DRE : Privates) { |
| 3305 | PrivatesReductions[Cnt] = cast<DeclRefExpr>(DRE)->getDecl(); |
| 3306 | ++Cnt; |
| 3307 | } |
| 3308 | const RecordDecl *TeamReductionRec = ::buildRecordForGlobalizedVars( |
| 3309 | CGM.getContext(), PrivatesReductions, llvm::None, VarFieldMap, |
| 3310 | C.getLangOpts().OpenMPCUDAReductionBufNum); |
| 3311 | TeamsReductions.push_back(TeamReductionRec); |
| 3312 | if (!KernelTeamsReductionPtr) { |
| 3313 | KernelTeamsReductionPtr = new llvm::GlobalVariable( |
| 3314 | CGM.getModule(), CGM.VoidPtrTy, /*isConstant=*/true, |
| 3315 | llvm::GlobalValue::InternalLinkage, nullptr, |
| 3316 | "_openmp_teams_reductions_buffer_$_$ptr"); |
| 3317 | } |
| 3318 | llvm::Value *GlobalBufferPtr = CGF.EmitLoadOfScalar( |
| 3319 | Address(KernelTeamsReductionPtr, CGF.VoidPtrTy, CGM.getPointerAlign()), |
| 3320 | /*Volatile=*/false, C.getPointerType(C.VoidPtrTy), Loc); |
| 3321 | llvm::Value *GlobalToBufferCpyFn = ::emitListToGlobalCopyFunction( |
| 3322 | CGM, Privates, ReductionArrayTy, Loc, TeamReductionRec, VarFieldMap); |
| 3323 | llvm::Value *GlobalToBufferRedFn = ::emitListToGlobalReduceFunction( |
| 3324 | CGM, Privates, ReductionArrayTy, Loc, TeamReductionRec, VarFieldMap, |
| 3325 | ReductionFn); |
| 3326 | llvm::Value *BufferToGlobalCpyFn = ::emitGlobalToListCopyFunction( |
| 3327 | CGM, Privates, ReductionArrayTy, Loc, TeamReductionRec, VarFieldMap); |
| 3328 | llvm::Value *BufferToGlobalRedFn = ::emitGlobalToListReduceFunction( |
| 3329 | CGM, Privates, ReductionArrayTy, Loc, TeamReductionRec, VarFieldMap, |
| 3330 | ReductionFn); |
| 3331 | |
| 3332 | llvm::Value *Args[] = { |
| 3333 | RTLoc, |
| 3334 | ThreadId, |
| 3335 | GlobalBufferPtr, |
| 3336 | CGF.Builder.getInt32(C.getLangOpts().OpenMPCUDAReductionBufNum), |
| 3337 | RL, |
| 3338 | ShuffleAndReduceFn, |
| 3339 | InterWarpCopyFn, |
| 3340 | GlobalToBufferCpyFn, |
| 3341 | GlobalToBufferRedFn, |
| 3342 | BufferToGlobalCpyFn, |
| 3343 | BufferToGlobalRedFn}; |
| 3344 | |
| 3345 | Res = CGF.EmitRuntimeCall( |
| 3346 | OMPBuilder.getOrCreateRuntimeFunction( |
| 3347 | CGM.getModule(), OMPRTL___kmpc_nvptx_teams_reduce_nowait_v2), |
| 3348 | Args); |
| 3349 | } |
| 3350 | |
| 3351 | // 5. Build if (res == 1) |
| 3352 | llvm::BasicBlock *ExitBB = CGF.createBasicBlock(".omp.reduction.done"); |
| 3353 | llvm::BasicBlock *ThenBB = CGF.createBasicBlock(".omp.reduction.then"); |
| 3354 | llvm::Value *Cond = CGF.Builder.CreateICmpEQ( |
| 3355 | Res, llvm::ConstantInt::get(CGM.Int32Ty, /*V=*/1)); |
| 3356 | CGF.Builder.CreateCondBr(Cond, ThenBB, ExitBB); |
| 3357 | |
| 3358 | // 6. Build then branch: where we have reduced values in the master |
| 3359 | // thread in each team. |
| 3360 | // __kmpc_end_reduce{_nowait}(<gtid>); |
| 3361 | // break; |
| 3362 | CGF.EmitBlock(ThenBB); |
| 3363 | |
| 3364 | // Add emission of __kmpc_end_reduce{_nowait}(<gtid>); |
| 3365 | auto &&CodeGen = [Privates, LHSExprs, RHSExprs, ReductionOps, |
| 3366 | this](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 3367 | auto IPriv = Privates.begin(); |
| 3368 | auto ILHS = LHSExprs.begin(); |
| 3369 | auto IRHS = RHSExprs.begin(); |
| 3370 | for (const Expr *E : ReductionOps) { |
| 3371 | emitSingleReductionCombiner(CGF, E, *IPriv, cast<DeclRefExpr>(*ILHS), |
| 3372 | cast<DeclRefExpr>(*IRHS)); |
| 3373 | ++IPriv; |
| 3374 | ++ILHS; |
| 3375 | ++IRHS; |
| 3376 | } |
| 3377 | }; |
| 3378 | llvm::Value *EndArgs[] = {ThreadId}; |
| 3379 | RegionCodeGenTy RCG(CodeGen); |
| 3380 | NVPTXActionTy Action( |
| 3381 | nullptr, llvm::None, |
| 3382 | OMPBuilder.getOrCreateRuntimeFunction( |
| 3383 | CGM.getModule(), OMPRTL___kmpc_nvptx_end_reduce_nowait), |
| 3384 | EndArgs); |
| 3385 | RCG.setAction(Action); |
| 3386 | RCG(CGF); |
| 3387 | // There is no need to emit line number for unconditional branch. |
| 3388 | (void)ApplyDebugLocation::CreateEmpty(CGF); |
| 3389 | CGF.EmitBlock(ExitBB, /*IsFinished=*/true); |
| 3390 | } |
| 3391 | |
| 3392 | const VarDecl * |
| 3393 | CGOpenMPRuntimeGPU::translateParameter(const FieldDecl *FD, |
| 3394 | const VarDecl *NativeParam) const { |
| 3395 | if (!NativeParam->getType()->isReferenceType()) |
| 3396 | return NativeParam; |
| 3397 | QualType ArgType = NativeParam->getType(); |
| 3398 | QualifierCollector QC; |
| 3399 | const Type *NonQualTy = QC.strip(ArgType); |
| 3400 | QualType PointeeTy = cast<ReferenceType>(NonQualTy)->getPointeeType(); |
| 3401 | if (const auto *Attr = FD->getAttr<OMPCaptureKindAttr>()) { |
| 3402 | if (Attr->getCaptureKind() == OMPC_map) { |
| 3403 | PointeeTy = CGM.getContext().getAddrSpaceQualType(PointeeTy, |
| 3404 | LangAS::opencl_global); |
| 3405 | } |
| 3406 | } |
| 3407 | ArgType = CGM.getContext().getPointerType(PointeeTy); |
| 3408 | QC.addRestrict(); |
| 3409 | enum { NVPTX_local_addr = 5 }; |
| 3410 | QC.addAddressSpace(getLangASFromTargetAS(NVPTX_local_addr)); |
| 3411 | ArgType = QC.apply(CGM.getContext(), ArgType); |
| 3412 | if (isa<ImplicitParamDecl>(NativeParam)) |
| 3413 | return ImplicitParamDecl::Create( |
| 3414 | CGM.getContext(), /*DC=*/nullptr, NativeParam->getLocation(), |
| 3415 | NativeParam->getIdentifier(), ArgType, ImplicitParamDecl::Other); |
| 3416 | return ParmVarDecl::Create( |
| 3417 | CGM.getContext(), |
| 3418 | const_cast<DeclContext *>(NativeParam->getDeclContext()), |
| 3419 | NativeParam->getBeginLoc(), NativeParam->getLocation(), |
| 3420 | NativeParam->getIdentifier(), ArgType, |
| 3421 | /*TInfo=*/nullptr, SC_None, /*DefArg=*/nullptr); |
| 3422 | } |
| 3423 | |
| 3424 | Address |
| 3425 | CGOpenMPRuntimeGPU::getParameterAddress(CodeGenFunction &CGF, |
| 3426 | const VarDecl *NativeParam, |
| 3427 | const VarDecl *TargetParam) const { |
| 3428 | assert(NativeParam != TargetParam &&(static_cast <bool> (NativeParam != TargetParam && NativeParam->getType()->isReferenceType() && "Native arg must not be the same as target arg." ) ? void (0) : __assert_fail ("NativeParam != TargetParam && NativeParam->getType()->isReferenceType() && \"Native arg must not be the same as target arg.\"" , "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 3430, __extension__ __PRETTY_FUNCTION__)) |
| 3429 | NativeParam->getType()->isReferenceType() &&(static_cast <bool> (NativeParam != TargetParam && NativeParam->getType()->isReferenceType() && "Native arg must not be the same as target arg." ) ? void (0) : __assert_fail ("NativeParam != TargetParam && NativeParam->getType()->isReferenceType() && \"Native arg must not be the same as target arg.\"" , "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 3430, __extension__ __PRETTY_FUNCTION__)) |
| 3430 | "Native arg must not be the same as target arg.")(static_cast <bool> (NativeParam != TargetParam && NativeParam->getType()->isReferenceType() && "Native arg must not be the same as target arg." ) ? void (0) : __assert_fail ("NativeParam != TargetParam && NativeParam->getType()->isReferenceType() && \"Native arg must not be the same as target arg.\"" , "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 3430, __extension__ __PRETTY_FUNCTION__)); |
| 3431 | Address LocalAddr = CGF.GetAddrOfLocalVar(TargetParam); |
| 3432 | QualType NativeParamType = NativeParam->getType(); |
| 3433 | QualifierCollector QC; |
| 3434 | const Type *NonQualTy = QC.strip(NativeParamType); |
| 3435 | QualType NativePointeeTy = cast<ReferenceType>(NonQualTy)->getPointeeType(); |
| 3436 | unsigned NativePointeeAddrSpace = |
| 3437 | CGF.getContext().getTargetAddressSpace(NativePointeeTy); |
| 3438 | QualType TargetTy = TargetParam->getType(); |
| 3439 | llvm::Value *TargetAddr = CGF.EmitLoadOfScalar( |
| 3440 | LocalAddr, /*Volatile=*/false, TargetTy, SourceLocation()); |
| 3441 | // First cast to generic. |
| 3442 | TargetAddr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( |
| 3443 | TargetAddr, llvm::PointerType::getWithSamePointeeType( |
| 3444 | cast<llvm::PointerType>(TargetAddr->getType()), /*AddrSpace=*/0)); |
| 3445 | // Cast from generic to native address space. |
| 3446 | TargetAddr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( |
| 3447 | TargetAddr, llvm::PointerType::getWithSamePointeeType( |
| 3448 | cast<llvm::PointerType>(TargetAddr->getType()), |
| 3449 | NativePointeeAddrSpace)); |
| 3450 | Address NativeParamAddr = CGF.CreateMemTemp(NativeParamType); |
| 3451 | CGF.EmitStoreOfScalar(TargetAddr, NativeParamAddr, /*Volatile=*/false, |
| 3452 | NativeParamType); |
| 3453 | return NativeParamAddr; |
| 3454 | } |
| 3455 | |
| 3456 | void CGOpenMPRuntimeGPU::emitOutlinedFunctionCall( |
| 3457 | CodeGenFunction &CGF, SourceLocation Loc, llvm::FunctionCallee OutlinedFn, |
| 3458 | ArrayRef<llvm::Value *> Args) const { |
| 3459 | SmallVector<llvm::Value *, 4> TargetArgs; |
| 3460 | TargetArgs.reserve(Args.size()); |
| 3461 | auto *FnType = OutlinedFn.getFunctionType(); |
| 3462 | for (unsigned I = 0, E = Args.size(); I < E; ++I) { |
| 3463 | if (FnType->isVarArg() && FnType->getNumParams() <= I) { |
| 3464 | TargetArgs.append(std::next(Args.begin(), I), Args.end()); |
| 3465 | break; |
| 3466 | } |
| 3467 | llvm::Type *TargetType = FnType->getParamType(I); |
| 3468 | llvm::Value *NativeArg = Args[I]; |
| 3469 | if (!TargetType->isPointerTy()) { |
| 3470 | TargetArgs.emplace_back(NativeArg); |
| 3471 | continue; |
| 3472 | } |
| 3473 | llvm::Value *TargetArg = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( |
| 3474 | NativeArg, llvm::PointerType::getWithSamePointeeType( |
| 3475 | cast<llvm::PointerType>(NativeArg->getType()), /*AddrSpace*/ 0)); |
| 3476 | TargetArgs.emplace_back( |
| 3477 | CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(TargetArg, TargetType)); |
| 3478 | } |
| 3479 | CGOpenMPRuntime::emitOutlinedFunctionCall(CGF, Loc, OutlinedFn, TargetArgs); |
| 3480 | } |
| 3481 | |
| 3482 | /// Emit function which wraps the outline parallel region |
| 3483 | /// and controls the arguments which are passed to this function. |
| 3484 | /// The wrapper ensures that the outlined function is called |
| 3485 | /// with the correct arguments when data is shared. |
| 3486 | llvm::Function *CGOpenMPRuntimeGPU::createParallelDataSharingWrapper( |
| 3487 | llvm::Function *OutlinedParallelFn, const OMPExecutableDirective &D) { |
| 3488 | ASTContext &Ctx = CGM.getContext(); |
| 3489 | const auto &CS = *D.getCapturedStmt(OMPD_parallel); |
| 3490 | |
| 3491 | // Create a function that takes as argument the source thread. |
| 3492 | FunctionArgList WrapperArgs; |
| 3493 | QualType Int16QTy = |
| 3494 | Ctx.getIntTypeForBitwidth(/*DestWidth=*/16, /*Signed=*/false); |
| 3495 | QualType Int32QTy = |
| 3496 | Ctx.getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/false); |
| 3497 | ImplicitParamDecl ParallelLevelArg(Ctx, /*DC=*/nullptr, D.getBeginLoc(), |
| 3498 | /*Id=*/nullptr, Int16QTy, |
| 3499 | ImplicitParamDecl::Other); |
| 3500 | ImplicitParamDecl WrapperArg(Ctx, /*DC=*/nullptr, D.getBeginLoc(), |
| 3501 | /*Id=*/nullptr, Int32QTy, |
| 3502 | ImplicitParamDecl::Other); |
| 3503 | WrapperArgs.emplace_back(&ParallelLevelArg); |
| 3504 | WrapperArgs.emplace_back(&WrapperArg); |
| 3505 | |
| 3506 | const CGFunctionInfo &CGFI = |
| 3507 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(Ctx.VoidTy, WrapperArgs); |
| 3508 | |
| 3509 | auto *Fn = llvm::Function::Create( |
| 3510 | CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage, |
| 3511 | Twine(OutlinedParallelFn->getName(), "_wrapper"), &CGM.getModule()); |
| 3512 | |
| 3513 | // Ensure we do not inline the function. This is trivially true for the ones |
| 3514 | // passed to __kmpc_fork_call but the ones calles in serialized regions |
| 3515 | // could be inlined. This is not a perfect but it is closer to the invariant |
| 3516 | // we want, namely, every data environment starts with a new function. |
| 3517 | // TODO: We should pass the if condition to the runtime function and do the |
| 3518 | // handling there. Much cleaner code. |
| 3519 | Fn->addFnAttr(llvm::Attribute::NoInline); |
| 3520 | |
| 3521 | CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI); |
| 3522 | Fn->setLinkage(llvm::GlobalValue::InternalLinkage); |
| 3523 | Fn->setDoesNotRecurse(); |
| 3524 | |
| 3525 | CodeGenFunction CGF(CGM, /*suppressNewContext=*/true); |
| 3526 | CGF.StartFunction(GlobalDecl(), Ctx.VoidTy, Fn, CGFI, WrapperArgs, |
| 3527 | D.getBeginLoc(), D.getBeginLoc()); |
| 3528 | |
| 3529 | const auto *RD = CS.getCapturedRecordDecl(); |
| 3530 | auto CurField = RD->field_begin(); |
| 3531 | |
| 3532 | Address ZeroAddr = CGF.CreateDefaultAlignTempAlloca(CGF.Int32Ty, |
| 3533 | /*Name=*/".zero.addr"); |
| 3534 | CGF.Builder.CreateStore(CGF.Builder.getInt32(/*C*/ 0), ZeroAddr); |
| 3535 | // Get the array of arguments. |
| 3536 | SmallVector<llvm::Value *, 8> Args; |
| 3537 | |
| 3538 | Args.emplace_back(CGF.GetAddrOfLocalVar(&WrapperArg).getPointer()); |
| 3539 | Args.emplace_back(ZeroAddr.getPointer()); |
| 3540 | |
| 3541 | CGBuilderTy &Bld = CGF.Builder; |
| 3542 | auto CI = CS.capture_begin(); |
| 3543 | |
| 3544 | // Use global memory for data sharing. |
| 3545 | // Handle passing of global args to workers. |
| 3546 | Address GlobalArgs = |
| 3547 | CGF.CreateDefaultAlignTempAlloca(CGF.VoidPtrPtrTy, "global_args"); |
| 3548 | llvm::Value *GlobalArgsPtr = GlobalArgs.getPointer(); |
| 3549 | llvm::Value *DataSharingArgs[] = {GlobalArgsPtr}; |
| 3550 | CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( |
| 3551 | CGM.getModule(), OMPRTL___kmpc_get_shared_variables), |
| 3552 | DataSharingArgs); |
| 3553 | |
| 3554 | // Retrieve the shared variables from the list of references returned |
| 3555 | // by the runtime. Pass the variables to the outlined function. |
| 3556 | Address SharedArgListAddress = Address::invalid(); |
| 3557 | if (CS.capture_size() > 0 || |
| 3558 | isOpenMPLoopBoundSharingDirective(D.getDirectiveKind())) { |
| 3559 | SharedArgListAddress = CGF.EmitLoadOfPointer( |
| 3560 | GlobalArgs, CGF.getContext() |
| 3561 | .getPointerType(CGF.getContext().VoidPtrTy) |
| 3562 | .castAs<PointerType>()); |
| 3563 | } |
| 3564 | unsigned Idx = 0; |
| 3565 | if (isOpenMPLoopBoundSharingDirective(D.getDirectiveKind())) { |
| 3566 | Address Src = Bld.CreateConstInBoundsGEP(SharedArgListAddress, Idx); |
| 3567 | Address TypedAddress = Bld.CreatePointerBitCastOrAddrSpaceCast( |
| 3568 | Src, CGF.SizeTy->getPointerTo(), CGF.SizeTy); |
| 3569 | llvm::Value *LB = CGF.EmitLoadOfScalar( |
| 3570 | TypedAddress, |
| 3571 | /*Volatile=*/false, |
| 3572 | CGF.getContext().getPointerType(CGF.getContext().getSizeType()), |
| 3573 | cast<OMPLoopDirective>(D).getLowerBoundVariable()->getExprLoc()); |
| 3574 | Args.emplace_back(LB); |
| 3575 | ++Idx; |
| 3576 | Src = Bld.CreateConstInBoundsGEP(SharedArgListAddress, Idx); |
| 3577 | TypedAddress = Bld.CreatePointerBitCastOrAddrSpaceCast( |
| 3578 | Src, CGF.SizeTy->getPointerTo(), CGF.SizeTy); |
| 3579 | llvm::Value *UB = CGF.EmitLoadOfScalar( |
| 3580 | TypedAddress, |
| 3581 | /*Volatile=*/false, |
| 3582 | CGF.getContext().getPointerType(CGF.getContext().getSizeType()), |
| 3583 | cast<OMPLoopDirective>(D).getUpperBoundVariable()->getExprLoc()); |
| 3584 | Args.emplace_back(UB); |
| 3585 | ++Idx; |
| 3586 | } |
| 3587 | if (CS.capture_size() > 0) { |
| 3588 | ASTContext &CGFContext = CGF.getContext(); |
| 3589 | for (unsigned I = 0, E = CS.capture_size(); I < E; ++I, ++CI, ++CurField) { |
| 3590 | QualType ElemTy = CurField->getType(); |
| 3591 | Address Src = Bld.CreateConstInBoundsGEP(SharedArgListAddress, I + Idx); |
| 3592 | Address TypedAddress = Bld.CreatePointerBitCastOrAddrSpaceCast( |
| 3593 | Src, CGF.ConvertTypeForMem(CGFContext.getPointerType(ElemTy)), |
| 3594 | CGF.ConvertTypeForMem(ElemTy)); |
| 3595 | llvm::Value *Arg = CGF.EmitLoadOfScalar(TypedAddress, |
| 3596 | /*Volatile=*/false, |
| 3597 | CGFContext.getPointerType(ElemTy), |
| 3598 | CI->getLocation()); |
| 3599 | if (CI->capturesVariableByCopy() && |
| 3600 | !CI->getCapturedVar()->getType()->isAnyPointerType()) { |
| 3601 | Arg = castValueToType(CGF, Arg, ElemTy, CGFContext.getUIntPtrType(), |
| 3602 | CI->getLocation()); |
| 3603 | } |
| 3604 | Args.emplace_back(Arg); |
| 3605 | } |
| 3606 | } |
| 3607 | |
| 3608 | emitOutlinedFunctionCall(CGF, D.getBeginLoc(), OutlinedParallelFn, Args); |
| 3609 | CGF.FinishFunction(); |
| 3610 | return Fn; |
| 3611 | } |
| 3612 | |
| 3613 | void CGOpenMPRuntimeGPU::emitFunctionProlog(CodeGenFunction &CGF, |
| 3614 | const Decl *D) { |
| 3615 | if (getDataSharingMode(CGM) != CGOpenMPRuntimeGPU::Generic) |
| 3616 | return; |
| 3617 | |
| 3618 | assert(D && "Expected function or captured|block decl.")(static_cast <bool> (D && "Expected function or captured|block decl." ) ? void (0) : __assert_fail ("D && \"Expected function or captured|block decl.\"" , "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 3618, __extension__ __PRETTY_FUNCTION__)); |
| 3619 | assert(FunctionGlobalizedDecls.count(CGF.CurFn) == 0 &&(static_cast <bool> (FunctionGlobalizedDecls.count(CGF. CurFn) == 0 && "Function is registered already.") ? void (0) : __assert_fail ("FunctionGlobalizedDecls.count(CGF.CurFn) == 0 && \"Function is registered already.\"" , "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 3620, __extension__ __PRETTY_FUNCTION__)) |
| 3620 | "Function is registered already.")(static_cast <bool> (FunctionGlobalizedDecls.count(CGF. CurFn) == 0 && "Function is registered already.") ? void (0) : __assert_fail ("FunctionGlobalizedDecls.count(CGF.CurFn) == 0 && \"Function is registered already.\"" , "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 3620, __extension__ __PRETTY_FUNCTION__)); |
| 3621 | assert((!TeamAndReductions.first || TeamAndReductions.first == D) &&(static_cast <bool> ((!TeamAndReductions.first || TeamAndReductions .first == D) && "Team is set but not processed.") ? void (0) : __assert_fail ("(!TeamAndReductions.first || TeamAndReductions.first == D) && \"Team is set but not processed.\"" , "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 3622, __extension__ __PRETTY_FUNCTION__)) |
| 3622 | "Team is set but not processed.")(static_cast <bool> ((!TeamAndReductions.first || TeamAndReductions .first == D) && "Team is set but not processed.") ? void (0) : __assert_fail ("(!TeamAndReductions.first || TeamAndReductions.first == D) && \"Team is set but not processed.\"" , "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 3622, __extension__ __PRETTY_FUNCTION__)); |
| 3623 | const Stmt *Body = nullptr; |
| 3624 | bool NeedToDelayGlobalization = false; |
| 3625 | if (const auto *FD = dyn_cast<FunctionDecl>(D)) { |
| 3626 | Body = FD->getBody(); |
| 3627 | } else if (const auto *BD = dyn_cast<BlockDecl>(D)) { |
| 3628 | Body = BD->getBody(); |
| 3629 | } else if (const auto *CD = dyn_cast<CapturedDecl>(D)) { |
| 3630 | Body = CD->getBody(); |
| 3631 | NeedToDelayGlobalization = CGF.CapturedStmtInfo->getKind() == CR_OpenMP; |
| 3632 | if (NeedToDelayGlobalization && |
| 3633 | getExecutionMode() == CGOpenMPRuntimeGPU::EM_SPMD) |
| 3634 | return; |
| 3635 | } |
| 3636 | if (!Body) |
| 3637 | return; |
| 3638 | CheckVarsEscapingDeclContext VarChecker(CGF, TeamAndReductions.second); |
| 3639 | VarChecker.Visit(Body); |
| 3640 | const RecordDecl *GlobalizedVarsRecord = |
| 3641 | VarChecker.getGlobalizedRecord(IsInTTDRegion); |
| 3642 | TeamAndReductions.first = nullptr; |
| 3643 | TeamAndReductions.second.clear(); |
| 3644 | ArrayRef<const ValueDecl *> EscapedVariableLengthDecls = |
| 3645 | VarChecker.getEscapedVariableLengthDecls(); |
| 3646 | if (!GlobalizedVarsRecord && EscapedVariableLengthDecls.empty()) |
| 3647 | return; |
| 3648 | auto I = FunctionGlobalizedDecls.try_emplace(CGF.CurFn).first; |
| 3649 | I->getSecond().MappedParams = |
| 3650 | std::make_unique<CodeGenFunction::OMPMapVars>(); |
| 3651 | I->getSecond().EscapedParameters.insert( |
| 3652 | VarChecker.getEscapedParameters().begin(), |
| 3653 | VarChecker.getEscapedParameters().end()); |
| 3654 | I->getSecond().EscapedVariableLengthDecls.append( |
| 3655 | EscapedVariableLengthDecls.begin(), EscapedVariableLengthDecls.end()); |
| 3656 | DeclToAddrMapTy &Data = I->getSecond().LocalVarData; |
| 3657 | for (const ValueDecl *VD : VarChecker.getEscapedDecls()) { |
| 3658 | assert(VD->isCanonicalDecl() && "Expected canonical declaration")(static_cast <bool> (VD->isCanonicalDecl() && "Expected canonical declaration") ? void (0) : __assert_fail ("VD->isCanonicalDecl() && \"Expected canonical declaration\"" , "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 3658, __extension__ __PRETTY_FUNCTION__)); |
| 3659 | Data.insert(std::make_pair(VD, MappedVarData())); |
| 3660 | } |
| 3661 | if (!IsInTTDRegion && !NeedToDelayGlobalization && !IsInParallelRegion) { |
| 3662 | CheckVarsEscapingDeclContext VarChecker(CGF, llvm::None); |
| 3663 | VarChecker.Visit(Body); |
| 3664 | I->getSecond().SecondaryLocalVarData.emplace(); |
| 3665 | DeclToAddrMapTy &Data = I->getSecond().SecondaryLocalVarData.getValue(); |
| 3666 | for (const ValueDecl *VD : VarChecker.getEscapedDecls()) { |
| 3667 | assert(VD->isCanonicalDecl() && "Expected canonical declaration")(static_cast <bool> (VD->isCanonicalDecl() && "Expected canonical declaration") ? void (0) : __assert_fail ("VD->isCanonicalDecl() && \"Expected canonical declaration\"" , "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 3667, __extension__ __PRETTY_FUNCTION__)); |
| 3668 | Data.insert(std::make_pair(VD, MappedVarData())); |
| 3669 | } |
| 3670 | } |
| 3671 | if (!NeedToDelayGlobalization) { |
| 3672 | emitGenericVarsProlog(CGF, D->getBeginLoc(), /*WithSPMDCheck=*/true); |
| 3673 | struct GlobalizationScope final : EHScopeStack::Cleanup { |
| 3674 | GlobalizationScope() = default; |
| 3675 | |
| 3676 | void Emit(CodeGenFunction &CGF, Flags flags) override { |
| 3677 | static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()) |
| 3678 | .emitGenericVarsEpilog(CGF, /*WithSPMDCheck=*/true); |
| 3679 | } |
| 3680 | }; |
| 3681 | CGF.EHStack.pushCleanup<GlobalizationScope>(NormalAndEHCleanup); |
| 3682 | } |
| 3683 | } |
| 3684 | |
| 3685 | Address CGOpenMPRuntimeGPU::getAddressOfLocalVariable(CodeGenFunction &CGF, |
| 3686 | const VarDecl *VD) { |
| 3687 | if (VD && VD->hasAttr<OMPAllocateDeclAttr>()) { |
| 3688 | const auto *A = VD->getAttr<OMPAllocateDeclAttr>(); |
| 3689 | auto AS = LangAS::Default; |
| 3690 | switch (A->getAllocatorType()) { |
| 3691 | // Use the default allocator here as by default local vars are |
| 3692 | // threadlocal. |
| 3693 | case OMPAllocateDeclAttr::OMPNullMemAlloc: |
| 3694 | case OMPAllocateDeclAttr::OMPDefaultMemAlloc: |
| 3695 | case OMPAllocateDeclAttr::OMPThreadMemAlloc: |
| 3696 | case OMPAllocateDeclAttr::OMPHighBWMemAlloc: |
| 3697 | case OMPAllocateDeclAttr::OMPLowLatMemAlloc: |
| 3698 | // Follow the user decision - use default allocation. |
| 3699 | return Address::invalid(); |
| 3700 | case OMPAllocateDeclAttr::OMPUserDefinedMemAlloc: |
| 3701 | // TODO: implement aupport for user-defined allocators. |
| 3702 | return Address::invalid(); |
| 3703 | case OMPAllocateDeclAttr::OMPConstMemAlloc: |
| 3704 | AS = LangAS::cuda_constant; |
| 3705 | break; |
| 3706 | case OMPAllocateDeclAttr::OMPPTeamMemAlloc: |
| 3707 | AS = LangAS::cuda_shared; |
| 3708 | break; |
| 3709 | case OMPAllocateDeclAttr::OMPLargeCapMemAlloc: |
| 3710 | case OMPAllocateDeclAttr::OMPCGroupMemAlloc: |
| 3711 | break; |
| 3712 | } |
| 3713 | llvm::Type *VarTy = CGF.ConvertTypeForMem(VD->getType()); |
| 3714 | auto *GV = new llvm::GlobalVariable( |
| 3715 | CGM.getModule(), VarTy, /*isConstant=*/false, |
| 3716 | llvm::GlobalValue::InternalLinkage, llvm::Constant::getNullValue(VarTy), |
| 3717 | VD->getName(), |
| 3718 | /*InsertBefore=*/nullptr, llvm::GlobalValue::NotThreadLocal, |
| 3719 | CGM.getContext().getTargetAddressSpace(AS)); |
| 3720 | CharUnits Align = CGM.getContext().getDeclAlign(VD); |
| 3721 | GV->setAlignment(Align.getAsAlign()); |
| 3722 | return Address( |
| 3723 | CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( |
| 3724 | GV, VarTy->getPointerTo(CGM.getContext().getTargetAddressSpace( |
| 3725 | VD->getType().getAddressSpace()))), |
| 3726 | VarTy, Align); |
| 3727 | } |
| 3728 | |
| 3729 | if (getDataSharingMode(CGM) != CGOpenMPRuntimeGPU::Generic) |
| 3730 | return Address::invalid(); |
| 3731 | |
| 3732 | VD = VD->getCanonicalDecl(); |
| 3733 | auto I = FunctionGlobalizedDecls.find(CGF.CurFn); |
| 3734 | if (I == FunctionGlobalizedDecls.end()) |
| 3735 | return Address::invalid(); |
| 3736 | auto VDI = I->getSecond().LocalVarData.find(VD); |
| 3737 | if (VDI != I->getSecond().LocalVarData.end()) |
| 3738 | return VDI->second.PrivateAddr; |
| 3739 | if (VD->hasAttrs()) { |
| 3740 | for (specific_attr_iterator<OMPReferencedVarAttr> IT(VD->attr_begin()), |
| 3741 | E(VD->attr_end()); |
| 3742 | IT != E; ++IT) { |
| 3743 | auto VDI = I->getSecond().LocalVarData.find( |
| 3744 | cast<VarDecl>(cast<DeclRefExpr>(IT->getRef())->getDecl()) |
| 3745 | ->getCanonicalDecl()); |
| 3746 | if (VDI != I->getSecond().LocalVarData.end()) |
| 3747 | return VDI->second.PrivateAddr; |
| 3748 | } |
| 3749 | } |
| 3750 | |
| 3751 | return Address::invalid(); |
| 3752 | } |
| 3753 | |
| 3754 | void CGOpenMPRuntimeGPU::functionFinished(CodeGenFunction &CGF) { |
| 3755 | FunctionGlobalizedDecls.erase(CGF.CurFn); |
| 3756 | CGOpenMPRuntime::functionFinished(CGF); |
| 3757 | } |
| 3758 | |
| 3759 | void CGOpenMPRuntimeGPU::getDefaultDistScheduleAndChunk( |
| 3760 | CodeGenFunction &CGF, const OMPLoopDirective &S, |
| 3761 | OpenMPDistScheduleClauseKind &ScheduleKind, |
| 3762 | llvm::Value *&Chunk) const { |
| 3763 | auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()); |
| 3764 | if (getExecutionMode() == CGOpenMPRuntimeGPU::EM_SPMD) { |
| 3765 | ScheduleKind = OMPC_DIST_SCHEDULE_static; |
| 3766 | Chunk = CGF.EmitScalarConversion( |
| 3767 | RT.getGPUNumThreads(CGF), |
| 3768 | CGF.getContext().getIntTypeForBitwidth(32, /*Signed=*/0), |
| 3769 | S.getIterationVariable()->getType(), S.getBeginLoc()); |
| 3770 | return; |
| 3771 | } |
| 3772 | CGOpenMPRuntime::getDefaultDistScheduleAndChunk( |
| 3773 | CGF, S, ScheduleKind, Chunk); |
| 3774 | } |
| 3775 | |
| 3776 | void CGOpenMPRuntimeGPU::getDefaultScheduleAndChunk( |
| 3777 | CodeGenFunction &CGF, const OMPLoopDirective &S, |
| 3778 | OpenMPScheduleClauseKind &ScheduleKind, |
| 3779 | const Expr *&ChunkExpr) const { |
| 3780 | ScheduleKind = OMPC_SCHEDULE_static; |
| 3781 | // Chunk size is 1 in this case. |
| 3782 | llvm::APInt ChunkSize(32, 1); |
| 3783 | ChunkExpr = IntegerLiteral::Create(CGF.getContext(), ChunkSize, |
| 3784 | CGF.getContext().getIntTypeForBitwidth(32, /*Signed=*/0), |
| 3785 | SourceLocation()); |
| 3786 | } |
| 3787 | |
| 3788 | void CGOpenMPRuntimeGPU::adjustTargetSpecificDataForLambdas( |
| 3789 | CodeGenFunction &CGF, const OMPExecutableDirective &D) const { |
| 3790 | assert(isOpenMPTargetExecutionDirective(D.getDirectiveKind()) &&(static_cast <bool> (isOpenMPTargetExecutionDirective(D .getDirectiveKind()) && " Expected target-based directive." ) ? void (0) : __assert_fail ("isOpenMPTargetExecutionDirective(D.getDirectiveKind()) && \" Expected target-based directive.\"" , "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 3791, __extension__ __PRETTY_FUNCTION__)) |
| 3791 | " Expected target-based directive.")(static_cast <bool> (isOpenMPTargetExecutionDirective(D .getDirectiveKind()) && " Expected target-based directive." ) ? void (0) : __assert_fail ("isOpenMPTargetExecutionDirective(D.getDirectiveKind()) && \" Expected target-based directive.\"" , "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 3791, __extension__ __PRETTY_FUNCTION__)); |
| 3792 | const CapturedStmt *CS = D.getCapturedStmt(OMPD_target); |
| 3793 | for (const CapturedStmt::Capture &C : CS->captures()) { |
| 3794 | // Capture variables captured by reference in lambdas for target-based |
| 3795 | // directives. |
| 3796 | if (!C.capturesVariable()) |
| 3797 | continue; |
| 3798 | const VarDecl *VD = C.getCapturedVar(); |
| 3799 | const auto *RD = VD->getType() |
| 3800 | .getCanonicalType() |
| 3801 | .getNonReferenceType() |
| 3802 | ->getAsCXXRecordDecl(); |
| 3803 | if (!RD || !RD->isLambda()) |
| 3804 | continue; |
| 3805 | Address VDAddr = CGF.GetAddrOfLocalVar(VD); |
| 3806 | LValue VDLVal; |
| 3807 | if (VD->getType().getCanonicalType()->isReferenceType()) |
| 3808 | VDLVal = CGF.EmitLoadOfReferenceLValue(VDAddr, VD->getType()); |
| 3809 | else |
| 3810 | VDLVal = CGF.MakeAddrLValue( |
| 3811 | VDAddr, VD->getType().getCanonicalType().getNonReferenceType()); |
| 3812 | llvm::DenseMap<const VarDecl *, FieldDecl *> Captures; |
| 3813 | FieldDecl *ThisCapture = nullptr; |
| 3814 | RD->getCaptureFields(Captures, ThisCapture); |
| 3815 | if (ThisCapture && CGF.CapturedStmtInfo->isCXXThisExprCaptured()) { |
| 3816 | LValue ThisLVal = |
| 3817 | CGF.EmitLValueForFieldInitialization(VDLVal, ThisCapture); |
| 3818 | llvm::Value *CXXThis = CGF.LoadCXXThis(); |
| 3819 | CGF.EmitStoreOfScalar(CXXThis, ThisLVal); |
| 3820 | } |
| 3821 | for (const LambdaCapture &LC : RD->captures()) { |
| 3822 | if (LC.getCaptureKind() != LCK_ByRef) |
| 3823 | continue; |
| 3824 | const VarDecl *VD = LC.getCapturedVar(); |
| 3825 | if (!CS->capturesVariable(VD)) |
| 3826 | continue; |
| 3827 | auto It = Captures.find(VD); |
| 3828 | assert(It != Captures.end() && "Found lambda capture without field.")(static_cast <bool> (It != Captures.end() && "Found lambda capture without field." ) ? void (0) : __assert_fail ("It != Captures.end() && \"Found lambda capture without field.\"" , "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp", 3828, __extension__ __PRETTY_FUNCTION__)); |
| 3829 | LValue VarLVal = CGF.EmitLValueForFieldInitialization(VDLVal, It->second); |
| 3830 | Address VDAddr = CGF.GetAddrOfLocalVar(VD); |
| 3831 | if (VD->getType().getCanonicalType()->isReferenceType()) |
| 3832 | VDAddr = CGF.EmitLoadOfReferenceLValue(VDAddr, |
| 3833 | VD->getType().getCanonicalType()) |
| 3834 | .getAddress(CGF); |
| 3835 | CGF.EmitStoreOfScalar(VDAddr.getPointer(), VarLVal); |
| 3836 | } |
| 3837 | } |
| 3838 | } |
| 3839 | |
| 3840 | bool CGOpenMPRuntimeGPU::hasAllocateAttributeForGlobalVar(const VarDecl *VD, |
| 3841 | LangAS &AS) { |
| 3842 | if (!VD || !VD->hasAttr<OMPAllocateDeclAttr>()) |
| 3843 | return false; |
| 3844 | const auto *A = VD->getAttr<OMPAllocateDeclAttr>(); |
| 3845 | switch(A->getAllocatorType()) { |
| 3846 | case OMPAllocateDeclAttr::OMPNullMemAlloc: |
| 3847 | case OMPAllocateDeclAttr::OMPDefaultMemAlloc: |
| 3848 | // Not supported, fallback to the default mem space. |
| 3849 | case OMPAllocateDeclAttr::OMPThreadMemAlloc: |
| 3850 | case OMPAllocateDeclAttr::OMPLargeCapMemAlloc: |
| 3851 | case OMPAllocateDeclAttr::OMPCGroupMemAlloc: |
| 3852 | case OMPAllocateDeclAttr::OMPHighBWMemAlloc: |
| 3853 | case OMPAllocateDeclAttr::OMPLowLatMemAlloc: |
| 3854 | AS = LangAS::Default; |
| 3855 | return true; |
| 3856 | case OMPAllocateDeclAttr::OMPConstMemAlloc: |
| 3857 | AS = LangAS::cuda_constant; |
| 3858 | return true; |
| 3859 | case OMPAllocateDeclAttr::OMPPTeamMemAlloc: |
| 3860 | AS = LangAS::cuda_shared; |
| 3861 | return true; |
| 3862 | case OMPAllocateDeclAttr::OMPUserDefinedMemAlloc: |
| 3863 | llvm_unreachable("Expected predefined allocator for the variables with the "::llvm::llvm_unreachable_internal("Expected predefined allocator for the variables with the " "static storage.", "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp" , 3864) |
| 3864 | "static storage.")::llvm::llvm_unreachable_internal("Expected predefined allocator for the variables with the " "static storage.", "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp" , 3864); |
| 3865 | } |
| 3866 | return false; |
| 3867 | } |
| 3868 | |
| 3869 | // Get current CudaArch and ignore any unknown values |
| 3870 | static CudaArch getCudaArch(CodeGenModule &CGM) { |
| 3871 | if (!CGM.getTarget().hasFeature("ptx")) |
| 3872 | return CudaArch::UNKNOWN; |
| 3873 | for (const auto &Feature : CGM.getTarget().getTargetOpts().FeatureMap) { |
| 3874 | if (Feature.getValue()) { |
| 3875 | CudaArch Arch = StringToCudaArch(Feature.getKey()); |
| 3876 | if (Arch != CudaArch::UNKNOWN) |
| 3877 | return Arch; |
| 3878 | } |
| 3879 | } |
| 3880 | return CudaArch::UNKNOWN; |
| 3881 | } |
| 3882 | |
| 3883 | /// Check to see if target architecture supports unified addressing which is |
| 3884 | /// a restriction for OpenMP requires clause "unified_shared_memory". |
| 3885 | void CGOpenMPRuntimeGPU::processRequiresDirective( |
| 3886 | const OMPRequiresDecl *D) { |
| 3887 | for (const OMPClause *Clause : D->clauselists()) { |
| 3888 | if (Clause->getClauseKind() == OMPC_unified_shared_memory) { |
| 3889 | CudaArch Arch = getCudaArch(CGM); |
| 3890 | switch (Arch) { |
| 3891 | case CudaArch::SM_20: |
| 3892 | case CudaArch::SM_21: |
| 3893 | case CudaArch::SM_30: |
| 3894 | case CudaArch::SM_32: |
| 3895 | case CudaArch::SM_35: |
| 3896 | case CudaArch::SM_37: |
| 3897 | case CudaArch::SM_50: |
| 3898 | case CudaArch::SM_52: |
| 3899 | case CudaArch::SM_53: { |
| 3900 | SmallString<256> Buffer; |
| 3901 | llvm::raw_svector_ostream Out(Buffer); |
| 3902 | Out << "Target architecture " << CudaArchToString(Arch) |
| 3903 | << " does not support unified addressing"; |
| 3904 | CGM.Error(Clause->getBeginLoc(), Out.str()); |
| 3905 | return; |
| 3906 | } |
| 3907 | case CudaArch::SM_60: |
| 3908 | case CudaArch::SM_61: |
| 3909 | case CudaArch::SM_62: |
| 3910 | case CudaArch::SM_70: |
| 3911 | case CudaArch::SM_72: |
| 3912 | case CudaArch::SM_75: |
| 3913 | case CudaArch::SM_80: |
| 3914 | case CudaArch::SM_86: |
| 3915 | case CudaArch::GFX600: |
| 3916 | case CudaArch::GFX601: |
| 3917 | case CudaArch::GFX602: |
| 3918 | case CudaArch::GFX700: |
| 3919 | case CudaArch::GFX701: |
| 3920 | case CudaArch::GFX702: |
| 3921 | case CudaArch::GFX703: |
| 3922 | case CudaArch::GFX704: |
| 3923 | case CudaArch::GFX705: |
| 3924 | case CudaArch::GFX801: |
| 3925 | case CudaArch::GFX802: |
| 3926 | case CudaArch::GFX803: |
| 3927 | case CudaArch::GFX805: |
| 3928 | case CudaArch::GFX810: |
| 3929 | case CudaArch::GFX900: |
| 3930 | case CudaArch::GFX902: |
| 3931 | case CudaArch::GFX904: |
| 3932 | case CudaArch::GFX906: |
| 3933 | case CudaArch::GFX908: |
| 3934 | case CudaArch::GFX909: |
| 3935 | case CudaArch::GFX90a: |
| 3936 | case CudaArch::GFX90c: |
| 3937 | case CudaArch::GFX940: |
| 3938 | case CudaArch::GFX1010: |
| 3939 | case CudaArch::GFX1011: |
| 3940 | case CudaArch::GFX1012: |
| 3941 | case CudaArch::GFX1013: |
| 3942 | case CudaArch::GFX1030: |
| 3943 | case CudaArch::GFX1031: |
| 3944 | case CudaArch::GFX1032: |
| 3945 | case CudaArch::GFX1033: |
| 3946 | case CudaArch::GFX1034: |
| 3947 | case CudaArch::GFX1035: |
| 3948 | case CudaArch::GFX1036: |
| 3949 | case CudaArch::Generic: |
| 3950 | case CudaArch::UNUSED: |
| 3951 | case CudaArch::UNKNOWN: |
| 3952 | break; |
| 3953 | case CudaArch::LAST: |
| 3954 | llvm_unreachable("Unexpected Cuda arch.")::llvm::llvm_unreachable_internal("Unexpected Cuda arch.", "clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp" , 3954); |
| 3955 | } |
| 3956 | } |
| 3957 | } |
| 3958 | CGOpenMPRuntime::processRequiresDirective(D); |
| 3959 | } |
| 3960 | |
| 3961 | void CGOpenMPRuntimeGPU::clear() { |
| 3962 | |
| 3963 | if (!TeamsReductions.empty()) { |
| 3964 | ASTContext &C = CGM.getContext(); |
| 3965 | RecordDecl *StaticRD = C.buildImplicitRecord( |
| 3966 | "_openmp_teams_reduction_type_$_", RecordDecl::TagKind::TTK_Union); |
| 3967 | StaticRD->startDefinition(); |
| 3968 | for (const RecordDecl *TeamReductionRec : TeamsReductions) { |
| 3969 | QualType RecTy = C.getRecordType(TeamReductionRec); |
| 3970 | auto *Field = FieldDecl::Create( |
| 3971 | C, StaticRD, SourceLocation(), SourceLocation(), nullptr, RecTy, |
| 3972 | C.getTrivialTypeSourceInfo(RecTy, SourceLocation()), |
| 3973 | /*BW=*/nullptr, /*Mutable=*/false, |
| 3974 | /*InitStyle=*/ICIS_NoInit); |
| 3975 | Field->setAccess(AS_public); |
| 3976 | StaticRD->addDecl(Field); |
| 3977 | } |
| 3978 | StaticRD->completeDefinition(); |
| 3979 | QualType StaticTy = C.getRecordType(StaticRD); |
| 3980 | llvm::Type *LLVMReductionsBufferTy = |
| 3981 | CGM.getTypes().ConvertTypeForMem(StaticTy); |
| 3982 | // FIXME: nvlink does not handle weak linkage correctly (object with the |
| 3983 | // different size are reported as erroneous). |
| 3984 | // Restore CommonLinkage as soon as nvlink is fixed. |
| 3985 | auto *GV = new llvm::GlobalVariable( |
| 3986 | CGM.getModule(), LLVMReductionsBufferTy, |
| 3987 | /*isConstant=*/false, llvm::GlobalValue::InternalLinkage, |
| 3988 | llvm::Constant::getNullValue(LLVMReductionsBufferTy), |
| 3989 | "_openmp_teams_reductions_buffer_$_"); |
| 3990 | KernelTeamsReductionPtr->setInitializer( |
| 3991 | llvm::ConstantExpr::getPointerBitCastOrAddrSpaceCast(GV, |
| 3992 | CGM.VoidPtrTy)); |
| 3993 | } |
| 3994 | CGOpenMPRuntime::clear(); |
| 3995 | } |
| 3996 | |
| 3997 | llvm::Value *CGOpenMPRuntimeGPU::getGPUNumThreads(CodeGenFunction &CGF) { |
| 3998 | CGBuilderTy &Bld = CGF.Builder; |
| 3999 | llvm::Module *M = &CGF.CGM.getModule(); |
| 4000 | const char *LocSize = "__kmpc_get_hardware_num_threads_in_block"; |
| 4001 | llvm::Function *F = M->getFunction(LocSize); |
| 4002 | if (!F) { |
| 4003 | F = llvm::Function::Create( |
| 4004 | llvm::FunctionType::get(CGF.Int32Ty, llvm::None, false), |
| 4005 | llvm::GlobalVariable::ExternalLinkage, LocSize, &CGF.CGM.getModule()); |
| 4006 | } |
| 4007 | return Bld.CreateCall(F, llvm::None, "nvptx_num_threads"); |
| 4008 | } |
| 4009 | |
| 4010 | llvm::Value *CGOpenMPRuntimeGPU::getGPUThreadID(CodeGenFunction &CGF) { |
| 4011 | ArrayRef<llvm::Value *> Args{}; |
| 4012 | return CGF.EmitRuntimeCall( |
| 4013 | OMPBuilder.getOrCreateRuntimeFunction( |
| 4014 | CGM.getModule(), OMPRTL___kmpc_get_hardware_thread_id_in_block), |
| 4015 | Args); |
| 4016 | } |
| 4017 | |
| 4018 | llvm::Value *CGOpenMPRuntimeGPU::getGPUWarpSize(CodeGenFunction &CGF) { |
| 4019 | ArrayRef<llvm::Value *> Args{}; |
| 4020 | return CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( |
| 4021 | CGM.getModule(), OMPRTL___kmpc_get_warp_size), |
| 4022 | Args); |
| 4023 | } |