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/OpenMPClause.h" 19 #include "clang/AST/StmtOpenMP.h" 20 #include "clang/AST/StmtVisitor.h" 21 #include "clang/Basic/Cuda.h" 22 #include "llvm/ADT/SmallPtrSet.h" 23 #include "llvm/Frontend/OpenMP/OMPGridValues.h" 24 #include "llvm/Support/MathExtras.h" 25 26 using namespace clang; 27 using namespace CodeGen; 28 using namespace llvm::omp; 29 30 namespace { 31 /// Pre(post)-action for different OpenMP constructs specialized for NVPTX. 32 class NVPTXActionTy final : public PrePostActionTy { 33 llvm::FunctionCallee EnterCallee = nullptr; 34 ArrayRef<llvm::Value *> EnterArgs; 35 llvm::FunctionCallee ExitCallee = nullptr; 36 ArrayRef<llvm::Value *> ExitArgs; 37 bool Conditional = false; 38 llvm::BasicBlock *ContBlock = nullptr; 39 40 public: 41 NVPTXActionTy(llvm::FunctionCallee EnterCallee, 42 ArrayRef<llvm::Value *> EnterArgs, 43 llvm::FunctionCallee ExitCallee, 44 ArrayRef<llvm::Value *> ExitArgs, bool Conditional = false) 45 : EnterCallee(EnterCallee), EnterArgs(EnterArgs), ExitCallee(ExitCallee), 46 ExitArgs(ExitArgs), Conditional(Conditional) {} 47 void Enter(CodeGenFunction &CGF) override { 48 llvm::Value *EnterRes = CGF.EmitRuntimeCall(EnterCallee, EnterArgs); 49 if (Conditional) { 50 llvm::Value *CallBool = CGF.Builder.CreateIsNotNull(EnterRes); 51 auto *ThenBlock = CGF.createBasicBlock("omp_if.then"); 52 ContBlock = CGF.createBasicBlock("omp_if.end"); 53 // Generate the branch (If-stmt) 54 CGF.Builder.CreateCondBr(CallBool, ThenBlock, ContBlock); 55 CGF.EmitBlock(ThenBlock); 56 } 57 } 58 void Done(CodeGenFunction &CGF) { 59 // Emit the rest of blocks/branches 60 CGF.EmitBranch(ContBlock); 61 CGF.EmitBlock(ContBlock, true); 62 } 63 void Exit(CodeGenFunction &CGF) override { 64 CGF.EmitRuntimeCall(ExitCallee, ExitArgs); 65 } 66 }; 67 68 /// A class to track the execution mode when codegening directives within 69 /// a target region. The appropriate mode (SPMD|NON-SPMD) is set on entry 70 /// to the target region and used by containing directives such as 'parallel' 71 /// to emit optimized code. 72 class ExecutionRuntimeModesRAII { 73 private: 74 CGOpenMPRuntimeGPU::ExecutionMode SavedExecMode = 75 CGOpenMPRuntimeGPU::EM_Unknown; 76 CGOpenMPRuntimeGPU::ExecutionMode &ExecMode; 77 78 public: 79 ExecutionRuntimeModesRAII(CGOpenMPRuntimeGPU::ExecutionMode &ExecMode, 80 CGOpenMPRuntimeGPU::ExecutionMode EntryMode) 81 : ExecMode(ExecMode) { 82 SavedExecMode = ExecMode; 83 ExecMode = EntryMode; 84 } 85 ~ExecutionRuntimeModesRAII() { ExecMode = SavedExecMode; } 86 }; 87 88 static const ValueDecl *getPrivateItem(const Expr *RefExpr) { 89 RefExpr = RefExpr->IgnoreParens(); 90 if (const auto *ASE = dyn_cast<ArraySubscriptExpr>(RefExpr)) { 91 const Expr *Base = ASE->getBase()->IgnoreParenImpCasts(); 92 while (const auto *TempASE = dyn_cast<ArraySubscriptExpr>(Base)) 93 Base = TempASE->getBase()->IgnoreParenImpCasts(); 94 RefExpr = Base; 95 } else if (auto *OASE = dyn_cast<OMPArraySectionExpr>(RefExpr)) { 96 const Expr *Base = OASE->getBase()->IgnoreParenImpCasts(); 97 while (const auto *TempOASE = dyn_cast<OMPArraySectionExpr>(Base)) 98 Base = TempOASE->getBase()->IgnoreParenImpCasts(); 99 while (const auto *TempASE = dyn_cast<ArraySubscriptExpr>(Base)) 100 Base = TempASE->getBase()->IgnoreParenImpCasts(); 101 RefExpr = Base; 102 } 103 RefExpr = RefExpr->IgnoreParenImpCasts(); 104 if (const auto *DE = dyn_cast<DeclRefExpr>(RefExpr)) 105 return cast<ValueDecl>(DE->getDecl()->getCanonicalDecl()); 106 const auto *ME = cast<MemberExpr>(RefExpr); 107 return cast<ValueDecl>(ME->getMemberDecl()->getCanonicalDecl()); 108 } 109 110 static RecordDecl *buildRecordForGlobalizedVars( 111 ASTContext &C, ArrayRef<const ValueDecl *> EscapedDecls, 112 ArrayRef<const ValueDecl *> EscapedDeclsForTeams, 113 llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> 114 &MappedDeclsFields, 115 int BufSize) { 116 using VarsDataTy = std::pair<CharUnits /*Align*/, const ValueDecl *>; 117 if (EscapedDecls.empty() && EscapedDeclsForTeams.empty()) 118 return nullptr; 119 SmallVector<VarsDataTy, 4> GlobalizedVars; 120 for (const ValueDecl *D : EscapedDecls) 121 GlobalizedVars.emplace_back(C.getDeclAlign(D), D); 122 for (const ValueDecl *D : EscapedDeclsForTeams) 123 GlobalizedVars.emplace_back(C.getDeclAlign(D), D); 124 125 // Build struct _globalized_locals_ty { 126 // /* globalized vars */[WarSize] align (decl_align) 127 // /* globalized vars */ for EscapedDeclsForTeams 128 // }; 129 RecordDecl *GlobalizedRD = C.buildImplicitRecord("_globalized_locals_ty"); 130 GlobalizedRD->startDefinition(); 131 llvm::SmallPtrSet<const ValueDecl *, 16> SingleEscaped( 132 EscapedDeclsForTeams.begin(), EscapedDeclsForTeams.end()); 133 for (const auto &Pair : GlobalizedVars) { 134 const ValueDecl *VD = Pair.second; 135 QualType Type = VD->getType(); 136 if (Type->isLValueReferenceType()) 137 Type = C.getPointerType(Type.getNonReferenceType()); 138 else 139 Type = Type.getNonReferenceType(); 140 SourceLocation Loc = VD->getLocation(); 141 FieldDecl *Field; 142 if (SingleEscaped.count(VD)) { 143 Field = FieldDecl::Create( 144 C, GlobalizedRD, Loc, Loc, VD->getIdentifier(), Type, 145 C.getTrivialTypeSourceInfo(Type, SourceLocation()), 146 /*BW=*/nullptr, /*Mutable=*/false, 147 /*InitStyle=*/ICIS_NoInit); 148 Field->setAccess(AS_public); 149 if (VD->hasAttrs()) { 150 for (specific_attr_iterator<AlignedAttr> I(VD->getAttrs().begin()), 151 E(VD->getAttrs().end()); 152 I != E; ++I) 153 Field->addAttr(*I); 154 } 155 } else { 156 if (BufSize > 1) { 157 llvm::APInt ArraySize(32, BufSize); 158 Type = C.getConstantArrayType(Type, ArraySize, nullptr, 159 ArraySizeModifier::Normal, 0); 160 } 161 Field = FieldDecl::Create( 162 C, GlobalizedRD, Loc, Loc, VD->getIdentifier(), Type, 163 C.getTrivialTypeSourceInfo(Type, SourceLocation()), 164 /*BW=*/nullptr, /*Mutable=*/false, 165 /*InitStyle=*/ICIS_NoInit); 166 Field->setAccess(AS_public); 167 llvm::APInt Align(32, Pair.first.getQuantity()); 168 Field->addAttr(AlignedAttr::CreateImplicit( 169 C, /*IsAlignmentExpr=*/true, 170 IntegerLiteral::Create(C, Align, 171 C.getIntTypeForBitwidth(32, /*Signed=*/0), 172 SourceLocation()), 173 {}, AlignedAttr::GNU_aligned)); 174 } 175 GlobalizedRD->addDecl(Field); 176 MappedDeclsFields.try_emplace(VD, Field); 177 } 178 GlobalizedRD->completeDefinition(); 179 return GlobalizedRD; 180 } 181 182 /// Get the list of variables that can escape their declaration context. 183 class CheckVarsEscapingDeclContext final 184 : public ConstStmtVisitor<CheckVarsEscapingDeclContext> { 185 CodeGenFunction &CGF; 186 llvm::SetVector<const ValueDecl *> EscapedDecls; 187 llvm::SetVector<const ValueDecl *> EscapedVariableLengthDecls; 188 llvm::SetVector<const ValueDecl *> DelayedVariableLengthDecls; 189 llvm::SmallPtrSet<const Decl *, 4> EscapedParameters; 190 RecordDecl *GlobalizedRD = nullptr; 191 llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> MappedDeclsFields; 192 bool AllEscaped = false; 193 bool IsForCombinedParallelRegion = false; 194 195 void markAsEscaped(const ValueDecl *VD) { 196 // Do not globalize declare target variables. 197 if (!isa<VarDecl>(VD) || 198 OMPDeclareTargetDeclAttr::isDeclareTargetDeclaration(VD)) 199 return; 200 VD = cast<ValueDecl>(VD->getCanonicalDecl()); 201 // Use user-specified allocation. 202 if (VD->hasAttrs() && VD->hasAttr<OMPAllocateDeclAttr>()) 203 return; 204 // Variables captured by value must be globalized. 205 bool IsCaptured = false; 206 if (auto *CSI = CGF.CapturedStmtInfo) { 207 if (const FieldDecl *FD = CSI->lookup(cast<VarDecl>(VD))) { 208 // Check if need to capture the variable that was already captured by 209 // value in the outer region. 210 IsCaptured = true; 211 if (!IsForCombinedParallelRegion) { 212 if (!FD->hasAttrs()) 213 return; 214 const auto *Attr = FD->getAttr<OMPCaptureKindAttr>(); 215 if (!Attr) 216 return; 217 if (((Attr->getCaptureKind() != OMPC_map) && 218 !isOpenMPPrivate(Attr->getCaptureKind())) || 219 ((Attr->getCaptureKind() == OMPC_map) && 220 !FD->getType()->isAnyPointerType())) 221 return; 222 } 223 if (!FD->getType()->isReferenceType()) { 224 assert(!VD->getType()->isVariablyModifiedType() && 225 "Parameter captured by value with variably modified type"); 226 EscapedParameters.insert(VD); 227 } else if (!IsForCombinedParallelRegion) { 228 return; 229 } 230 } 231 } 232 if ((!CGF.CapturedStmtInfo || 233 (IsForCombinedParallelRegion && CGF.CapturedStmtInfo)) && 234 VD->getType()->isReferenceType()) 235 // Do not globalize variables with reference type. 236 return; 237 if (VD->getType()->isVariablyModifiedType()) { 238 // If not captured at the target region level then mark the escaped 239 // variable as delayed. 240 if (IsCaptured) 241 EscapedVariableLengthDecls.insert(VD); 242 else 243 DelayedVariableLengthDecls.insert(VD); 244 } else 245 EscapedDecls.insert(VD); 246 } 247 248 void VisitValueDecl(const ValueDecl *VD) { 249 if (VD->getType()->isLValueReferenceType()) 250 markAsEscaped(VD); 251 if (const auto *VarD = dyn_cast<VarDecl>(VD)) { 252 if (!isa<ParmVarDecl>(VarD) && VarD->hasInit()) { 253 const bool SavedAllEscaped = AllEscaped; 254 AllEscaped = VD->getType()->isLValueReferenceType(); 255 Visit(VarD->getInit()); 256 AllEscaped = SavedAllEscaped; 257 } 258 } 259 } 260 void VisitOpenMPCapturedStmt(const CapturedStmt *S, 261 ArrayRef<OMPClause *> Clauses, 262 bool IsCombinedParallelRegion) { 263 if (!S) 264 return; 265 for (const CapturedStmt::Capture &C : S->captures()) { 266 if (C.capturesVariable() && !C.capturesVariableByCopy()) { 267 const ValueDecl *VD = C.getCapturedVar(); 268 bool SavedIsForCombinedParallelRegion = IsForCombinedParallelRegion; 269 if (IsCombinedParallelRegion) { 270 // Check if the variable is privatized in the combined construct and 271 // those private copies must be shared in the inner parallel 272 // directive. 273 IsForCombinedParallelRegion = false; 274 for (const OMPClause *C : Clauses) { 275 if (!isOpenMPPrivate(C->getClauseKind()) || 276 C->getClauseKind() == OMPC_reduction || 277 C->getClauseKind() == OMPC_linear || 278 C->getClauseKind() == OMPC_private) 279 continue; 280 ArrayRef<const Expr *> Vars; 281 if (const auto *PC = dyn_cast<OMPFirstprivateClause>(C)) 282 Vars = PC->getVarRefs(); 283 else if (const auto *PC = dyn_cast<OMPLastprivateClause>(C)) 284 Vars = PC->getVarRefs(); 285 else 286 llvm_unreachable("Unexpected clause."); 287 for (const auto *E : Vars) { 288 const Decl *D = 289 cast<DeclRefExpr>(E)->getDecl()->getCanonicalDecl(); 290 if (D == VD->getCanonicalDecl()) { 291 IsForCombinedParallelRegion = true; 292 break; 293 } 294 } 295 if (IsForCombinedParallelRegion) 296 break; 297 } 298 } 299 markAsEscaped(VD); 300 if (isa<OMPCapturedExprDecl>(VD)) 301 VisitValueDecl(VD); 302 IsForCombinedParallelRegion = SavedIsForCombinedParallelRegion; 303 } 304 } 305 } 306 307 void buildRecordForGlobalizedVars(bool IsInTTDRegion) { 308 assert(!GlobalizedRD && 309 "Record for globalized variables is built already."); 310 ArrayRef<const ValueDecl *> EscapedDeclsForParallel, EscapedDeclsForTeams; 311 unsigned WarpSize = CGF.getTarget().getGridValue().GV_Warp_Size; 312 if (IsInTTDRegion) 313 EscapedDeclsForTeams = EscapedDecls.getArrayRef(); 314 else 315 EscapedDeclsForParallel = EscapedDecls.getArrayRef(); 316 GlobalizedRD = ::buildRecordForGlobalizedVars( 317 CGF.getContext(), EscapedDeclsForParallel, EscapedDeclsForTeams, 318 MappedDeclsFields, WarpSize); 319 } 320 321 public: 322 CheckVarsEscapingDeclContext(CodeGenFunction &CGF, 323 ArrayRef<const ValueDecl *> TeamsReductions) 324 : CGF(CGF), EscapedDecls(TeamsReductions.begin(), TeamsReductions.end()) { 325 } 326 virtual ~CheckVarsEscapingDeclContext() = default; 327 void VisitDeclStmt(const DeclStmt *S) { 328 if (!S) 329 return; 330 for (const Decl *D : S->decls()) 331 if (const auto *VD = dyn_cast_or_null<ValueDecl>(D)) 332 VisitValueDecl(VD); 333 } 334 void VisitOMPExecutableDirective(const OMPExecutableDirective *D) { 335 if (!D) 336 return; 337 if (!D->hasAssociatedStmt()) 338 return; 339 if (const auto *S = 340 dyn_cast_or_null<CapturedStmt>(D->getAssociatedStmt())) { 341 // Do not analyze directives that do not actually require capturing, 342 // like `omp for` or `omp simd` directives. 343 llvm::SmallVector<OpenMPDirectiveKind, 4> CaptureRegions; 344 getOpenMPCaptureRegions(CaptureRegions, D->getDirectiveKind()); 345 if (CaptureRegions.size() == 1 && CaptureRegions.back() == OMPD_unknown) { 346 VisitStmt(S->getCapturedStmt()); 347 return; 348 } 349 VisitOpenMPCapturedStmt( 350 S, D->clauses(), 351 CaptureRegions.back() == OMPD_parallel && 352 isOpenMPDistributeDirective(D->getDirectiveKind())); 353 } 354 } 355 void VisitCapturedStmt(const CapturedStmt *S) { 356 if (!S) 357 return; 358 for (const CapturedStmt::Capture &C : S->captures()) { 359 if (C.capturesVariable() && !C.capturesVariableByCopy()) { 360 const ValueDecl *VD = C.getCapturedVar(); 361 markAsEscaped(VD); 362 if (isa<OMPCapturedExprDecl>(VD)) 363 VisitValueDecl(VD); 364 } 365 } 366 } 367 void VisitLambdaExpr(const LambdaExpr *E) { 368 if (!E) 369 return; 370 for (const LambdaCapture &C : E->captures()) { 371 if (C.capturesVariable()) { 372 if (C.getCaptureKind() == LCK_ByRef) { 373 const ValueDecl *VD = C.getCapturedVar(); 374 markAsEscaped(VD); 375 if (E->isInitCapture(&C) || isa<OMPCapturedExprDecl>(VD)) 376 VisitValueDecl(VD); 377 } 378 } 379 } 380 } 381 void VisitBlockExpr(const BlockExpr *E) { 382 if (!E) 383 return; 384 for (const BlockDecl::Capture &C : E->getBlockDecl()->captures()) { 385 if (C.isByRef()) { 386 const VarDecl *VD = C.getVariable(); 387 markAsEscaped(VD); 388 if (isa<OMPCapturedExprDecl>(VD) || VD->isInitCapture()) 389 VisitValueDecl(VD); 390 } 391 } 392 } 393 void VisitCallExpr(const CallExpr *E) { 394 if (!E) 395 return; 396 for (const Expr *Arg : E->arguments()) { 397 if (!Arg) 398 continue; 399 if (Arg->isLValue()) { 400 const bool SavedAllEscaped = AllEscaped; 401 AllEscaped = true; 402 Visit(Arg); 403 AllEscaped = SavedAllEscaped; 404 } else { 405 Visit(Arg); 406 } 407 } 408 Visit(E->getCallee()); 409 } 410 void VisitDeclRefExpr(const DeclRefExpr *E) { 411 if (!E) 412 return; 413 const ValueDecl *VD = E->getDecl(); 414 if (AllEscaped) 415 markAsEscaped(VD); 416 if (isa<OMPCapturedExprDecl>(VD)) 417 VisitValueDecl(VD); 418 else if (VD->isInitCapture()) 419 VisitValueDecl(VD); 420 } 421 void VisitUnaryOperator(const UnaryOperator *E) { 422 if (!E) 423 return; 424 if (E->getOpcode() == UO_AddrOf) { 425 const bool SavedAllEscaped = AllEscaped; 426 AllEscaped = true; 427 Visit(E->getSubExpr()); 428 AllEscaped = SavedAllEscaped; 429 } else { 430 Visit(E->getSubExpr()); 431 } 432 } 433 void VisitImplicitCastExpr(const ImplicitCastExpr *E) { 434 if (!E) 435 return; 436 if (E->getCastKind() == CK_ArrayToPointerDecay) { 437 const bool SavedAllEscaped = AllEscaped; 438 AllEscaped = true; 439 Visit(E->getSubExpr()); 440 AllEscaped = SavedAllEscaped; 441 } else { 442 Visit(E->getSubExpr()); 443 } 444 } 445 void VisitExpr(const Expr *E) { 446 if (!E) 447 return; 448 bool SavedAllEscaped = AllEscaped; 449 if (!E->isLValue()) 450 AllEscaped = false; 451 for (const Stmt *Child : E->children()) 452 if (Child) 453 Visit(Child); 454 AllEscaped = SavedAllEscaped; 455 } 456 void VisitStmt(const Stmt *S) { 457 if (!S) 458 return; 459 for (const Stmt *Child : S->children()) 460 if (Child) 461 Visit(Child); 462 } 463 464 /// Returns the record that handles all the escaped local variables and used 465 /// instead of their original storage. 466 const RecordDecl *getGlobalizedRecord(bool IsInTTDRegion) { 467 if (!GlobalizedRD) 468 buildRecordForGlobalizedVars(IsInTTDRegion); 469 return GlobalizedRD; 470 } 471 472 /// Returns the field in the globalized record for the escaped variable. 473 const FieldDecl *getFieldForGlobalizedVar(const ValueDecl *VD) const { 474 assert(GlobalizedRD && 475 "Record for globalized variables must be generated already."); 476 return MappedDeclsFields.lookup(VD); 477 } 478 479 /// Returns the list of the escaped local variables/parameters. 480 ArrayRef<const ValueDecl *> getEscapedDecls() const { 481 return EscapedDecls.getArrayRef(); 482 } 483 484 /// Checks if the escaped local variable is actually a parameter passed by 485 /// value. 486 const llvm::SmallPtrSetImpl<const Decl *> &getEscapedParameters() const { 487 return EscapedParameters; 488 } 489 490 /// Returns the list of the escaped variables with the variably modified 491 /// types. 492 ArrayRef<const ValueDecl *> getEscapedVariableLengthDecls() const { 493 return EscapedVariableLengthDecls.getArrayRef(); 494 } 495 496 /// Returns the list of the delayed variables with the variably modified 497 /// types. 498 ArrayRef<const ValueDecl *> getDelayedVariableLengthDecls() const { 499 return DelayedVariableLengthDecls.getArrayRef(); 500 } 501 }; 502 } // anonymous namespace 503 504 /// Get the id of the warp in the block. 505 /// We assume that the warp size is 32, which is always the case 506 /// on the NVPTX device, to generate more efficient code. 507 static llvm::Value *getNVPTXWarpID(CodeGenFunction &CGF) { 508 CGBuilderTy &Bld = CGF.Builder; 509 unsigned LaneIDBits = 510 llvm::Log2_32(CGF.getTarget().getGridValue().GV_Warp_Size); 511 auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()); 512 return Bld.CreateAShr(RT.getGPUThreadID(CGF), LaneIDBits, "nvptx_warp_id"); 513 } 514 515 /// Get the id of the current lane in the Warp. 516 /// We assume that the warp size is 32, which is always the case 517 /// on the NVPTX device, to generate more efficient code. 518 static llvm::Value *getNVPTXLaneID(CodeGenFunction &CGF) { 519 CGBuilderTy &Bld = CGF.Builder; 520 unsigned LaneIDBits = 521 llvm::Log2_32(CGF.getTarget().getGridValue().GV_Warp_Size); 522 assert(LaneIDBits < 32 && "Invalid LaneIDBits size in NVPTX device."); 523 unsigned LaneIDMask = ~0u >> (32u - LaneIDBits); 524 auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()); 525 return Bld.CreateAnd(RT.getGPUThreadID(CGF), Bld.getInt32(LaneIDMask), 526 "nvptx_lane_id"); 527 } 528 529 CGOpenMPRuntimeGPU::ExecutionMode 530 CGOpenMPRuntimeGPU::getExecutionMode() const { 531 return CurrentExecutionMode; 532 } 533 534 CGOpenMPRuntimeGPU::DataSharingMode 535 CGOpenMPRuntimeGPU::getDataSharingMode() const { 536 return CurrentDataSharingMode; 537 } 538 539 /// Check for inner (nested) SPMD construct, if any 540 static bool hasNestedSPMDDirective(ASTContext &Ctx, 541 const OMPExecutableDirective &D) { 542 const auto *CS = D.getInnermostCapturedStmt(); 543 const auto *Body = 544 CS->getCapturedStmt()->IgnoreContainers(/*IgnoreCaptured=*/true); 545 const Stmt *ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body); 546 547 if (const auto *NestedDir = 548 dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) { 549 OpenMPDirectiveKind DKind = NestedDir->getDirectiveKind(); 550 switch (D.getDirectiveKind()) { 551 case OMPD_target: 552 if (isOpenMPParallelDirective(DKind)) 553 return true; 554 if (DKind == OMPD_teams) { 555 Body = NestedDir->getInnermostCapturedStmt()->IgnoreContainers( 556 /*IgnoreCaptured=*/true); 557 if (!Body) 558 return false; 559 ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body); 560 if (const auto *NND = 561 dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) { 562 DKind = NND->getDirectiveKind(); 563 if (isOpenMPParallelDirective(DKind)) 564 return true; 565 } 566 } 567 return false; 568 case OMPD_target_teams: 569 return isOpenMPParallelDirective(DKind); 570 case OMPD_target_simd: 571 case OMPD_target_parallel: 572 case OMPD_target_parallel_for: 573 case OMPD_target_parallel_for_simd: 574 case OMPD_target_teams_distribute: 575 case OMPD_target_teams_distribute_simd: 576 case OMPD_target_teams_distribute_parallel_for: 577 case OMPD_target_teams_distribute_parallel_for_simd: 578 case OMPD_parallel: 579 case OMPD_for: 580 case OMPD_parallel_for: 581 case OMPD_parallel_master: 582 case OMPD_parallel_sections: 583 case OMPD_for_simd: 584 case OMPD_parallel_for_simd: 585 case OMPD_cancel: 586 case OMPD_cancellation_point: 587 case OMPD_ordered: 588 case OMPD_threadprivate: 589 case OMPD_allocate: 590 case OMPD_task: 591 case OMPD_simd: 592 case OMPD_sections: 593 case OMPD_section: 594 case OMPD_single: 595 case OMPD_master: 596 case OMPD_critical: 597 case OMPD_taskyield: 598 case OMPD_barrier: 599 case OMPD_taskwait: 600 case OMPD_taskgroup: 601 case OMPD_atomic: 602 case OMPD_flush: 603 case OMPD_depobj: 604 case OMPD_scan: 605 case OMPD_teams: 606 case OMPD_target_data: 607 case OMPD_target_exit_data: 608 case OMPD_target_enter_data: 609 case OMPD_distribute: 610 case OMPD_distribute_simd: 611 case OMPD_distribute_parallel_for: 612 case OMPD_distribute_parallel_for_simd: 613 case OMPD_teams_distribute: 614 case OMPD_teams_distribute_simd: 615 case OMPD_teams_distribute_parallel_for: 616 case OMPD_teams_distribute_parallel_for_simd: 617 case OMPD_target_update: 618 case OMPD_declare_simd: 619 case OMPD_declare_variant: 620 case OMPD_begin_declare_variant: 621 case OMPD_end_declare_variant: 622 case OMPD_declare_target: 623 case OMPD_end_declare_target: 624 case OMPD_declare_reduction: 625 case OMPD_declare_mapper: 626 case OMPD_taskloop: 627 case OMPD_taskloop_simd: 628 case OMPD_master_taskloop: 629 case OMPD_master_taskloop_simd: 630 case OMPD_parallel_master_taskloop: 631 case OMPD_parallel_master_taskloop_simd: 632 case OMPD_requires: 633 case OMPD_unknown: 634 default: 635 llvm_unreachable("Unexpected directive."); 636 } 637 } 638 639 return false; 640 } 641 642 static bool supportsSPMDExecutionMode(ASTContext &Ctx, 643 const OMPExecutableDirective &D) { 644 OpenMPDirectiveKind DirectiveKind = D.getDirectiveKind(); 645 switch (DirectiveKind) { 646 case OMPD_target: 647 case OMPD_target_teams: 648 return hasNestedSPMDDirective(Ctx, D); 649 case OMPD_target_teams_loop: 650 case OMPD_target_parallel_loop: 651 case OMPD_target_parallel: 652 case OMPD_target_parallel_for: 653 case OMPD_target_parallel_for_simd: 654 case OMPD_target_teams_distribute_parallel_for: 655 case OMPD_target_teams_distribute_parallel_for_simd: 656 case OMPD_target_simd: 657 case OMPD_target_teams_distribute_simd: 658 return true; 659 case OMPD_target_teams_distribute: 660 return false; 661 case OMPD_parallel: 662 case OMPD_for: 663 case OMPD_parallel_for: 664 case OMPD_parallel_master: 665 case OMPD_parallel_sections: 666 case OMPD_for_simd: 667 case OMPD_parallel_for_simd: 668 case OMPD_cancel: 669 case OMPD_cancellation_point: 670 case OMPD_ordered: 671 case OMPD_threadprivate: 672 case OMPD_allocate: 673 case OMPD_task: 674 case OMPD_simd: 675 case OMPD_sections: 676 case OMPD_section: 677 case OMPD_single: 678 case OMPD_master: 679 case OMPD_critical: 680 case OMPD_taskyield: 681 case OMPD_barrier: 682 case OMPD_taskwait: 683 case OMPD_taskgroup: 684 case OMPD_atomic: 685 case OMPD_flush: 686 case OMPD_depobj: 687 case OMPD_scan: 688 case OMPD_teams: 689 case OMPD_target_data: 690 case OMPD_target_exit_data: 691 case OMPD_target_enter_data: 692 case OMPD_distribute: 693 case OMPD_distribute_simd: 694 case OMPD_distribute_parallel_for: 695 case OMPD_distribute_parallel_for_simd: 696 case OMPD_teams_distribute: 697 case OMPD_teams_distribute_simd: 698 case OMPD_teams_distribute_parallel_for: 699 case OMPD_teams_distribute_parallel_for_simd: 700 case OMPD_target_update: 701 case OMPD_declare_simd: 702 case OMPD_declare_variant: 703 case OMPD_begin_declare_variant: 704 case OMPD_end_declare_variant: 705 case OMPD_declare_target: 706 case OMPD_end_declare_target: 707 case OMPD_declare_reduction: 708 case OMPD_declare_mapper: 709 case OMPD_taskloop: 710 case OMPD_taskloop_simd: 711 case OMPD_master_taskloop: 712 case OMPD_master_taskloop_simd: 713 case OMPD_parallel_master_taskloop: 714 case OMPD_parallel_master_taskloop_simd: 715 case OMPD_requires: 716 case OMPD_unknown: 717 default: 718 break; 719 } 720 llvm_unreachable( 721 "Unknown programming model for OpenMP directive on NVPTX target."); 722 } 723 724 void CGOpenMPRuntimeGPU::emitNonSPMDKernel(const OMPExecutableDirective &D, 725 StringRef ParentName, 726 llvm::Function *&OutlinedFn, 727 llvm::Constant *&OutlinedFnID, 728 bool IsOffloadEntry, 729 const RegionCodeGenTy &CodeGen) { 730 ExecutionRuntimeModesRAII ModeRAII(CurrentExecutionMode, EM_NonSPMD); 731 EntryFunctionState EST; 732 WrapperFunctionsMap.clear(); 733 734 [[maybe_unused]] bool IsBareKernel = D.getSingleClause<OMPXBareClause>(); 735 assert(!IsBareKernel && "bare kernel should not be at generic mode"); 736 737 // Emit target region as a standalone region. 738 class NVPTXPrePostActionTy : public PrePostActionTy { 739 CGOpenMPRuntimeGPU::EntryFunctionState &EST; 740 const OMPExecutableDirective &D; 741 742 public: 743 NVPTXPrePostActionTy(CGOpenMPRuntimeGPU::EntryFunctionState &EST, 744 const OMPExecutableDirective &D) 745 : EST(EST), D(D) {} 746 void Enter(CodeGenFunction &CGF) override { 747 auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()); 748 RT.emitKernelInit(D, CGF, EST, /* IsSPMD */ false); 749 // Skip target region initialization. 750 RT.setLocThreadIdInsertPt(CGF, /*AtCurrentPoint=*/true); 751 } 752 void Exit(CodeGenFunction &CGF) override { 753 auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()); 754 RT.clearLocThreadIdInsertPt(CGF); 755 RT.emitKernelDeinit(CGF, EST, /* IsSPMD */ false); 756 } 757 } Action(EST, D); 758 CodeGen.setAction(Action); 759 IsInTTDRegion = true; 760 emitTargetOutlinedFunctionHelper(D, ParentName, OutlinedFn, OutlinedFnID, 761 IsOffloadEntry, CodeGen); 762 IsInTTDRegion = false; 763 } 764 765 void CGOpenMPRuntimeGPU::emitKernelInit(const OMPExecutableDirective &D, 766 CodeGenFunction &CGF, 767 EntryFunctionState &EST, bool IsSPMD) { 768 int32_t MinThreadsVal = 1, MaxThreadsVal = -1, MinTeamsVal = 1, 769 MaxTeamsVal = -1; 770 computeMinAndMaxThreadsAndTeams(D, CGF, MinThreadsVal, MaxThreadsVal, 771 MinTeamsVal, MaxTeamsVal); 772 773 CGBuilderTy &Bld = CGF.Builder; 774 Bld.restoreIP(OMPBuilder.createTargetInit( 775 Bld, IsSPMD, MinThreadsVal, MaxThreadsVal, MinTeamsVal, MaxTeamsVal)); 776 if (!IsSPMD) 777 emitGenericVarsProlog(CGF, EST.Loc); 778 } 779 780 void CGOpenMPRuntimeGPU::emitKernelDeinit(CodeGenFunction &CGF, 781 EntryFunctionState &EST, 782 bool IsSPMD) { 783 if (!IsSPMD) 784 emitGenericVarsEpilog(CGF); 785 786 // This is temporary until we remove the fixed sized buffer. 787 ASTContext &C = CGM.getContext(); 788 RecordDecl *StaticRD = C.buildImplicitRecord( 789 "_openmp_teams_reduction_type_$_", RecordDecl::TagKind::Union); 790 StaticRD->startDefinition(); 791 for (const RecordDecl *TeamReductionRec : TeamsReductions) { 792 QualType RecTy = C.getRecordType(TeamReductionRec); 793 auto *Field = FieldDecl::Create( 794 C, StaticRD, SourceLocation(), SourceLocation(), nullptr, RecTy, 795 C.getTrivialTypeSourceInfo(RecTy, SourceLocation()), 796 /*BW=*/nullptr, /*Mutable=*/false, 797 /*InitStyle=*/ICIS_NoInit); 798 Field->setAccess(AS_public); 799 StaticRD->addDecl(Field); 800 } 801 StaticRD->completeDefinition(); 802 QualType StaticTy = C.getRecordType(StaticRD); 803 llvm::Type *LLVMReductionsBufferTy = 804 CGM.getTypes().ConvertTypeForMem(StaticTy); 805 const auto &DL = CGM.getModule().getDataLayout(); 806 uint64_t ReductionDataSize = 807 TeamsReductions.empty() 808 ? 0 809 : DL.getTypeAllocSize(LLVMReductionsBufferTy).getFixedValue(); 810 CGBuilderTy &Bld = CGF.Builder; 811 OMPBuilder.createTargetDeinit(Bld, ReductionDataSize, 812 C.getLangOpts().OpenMPCUDAReductionBufNum); 813 TeamsReductions.clear(); 814 } 815 816 void CGOpenMPRuntimeGPU::emitSPMDKernel(const OMPExecutableDirective &D, 817 StringRef ParentName, 818 llvm::Function *&OutlinedFn, 819 llvm::Constant *&OutlinedFnID, 820 bool IsOffloadEntry, 821 const RegionCodeGenTy &CodeGen) { 822 ExecutionRuntimeModesRAII ModeRAII(CurrentExecutionMode, EM_SPMD); 823 EntryFunctionState EST; 824 825 bool IsBareKernel = D.getSingleClause<OMPXBareClause>(); 826 827 // Emit target region as a standalone region. 828 class NVPTXPrePostActionTy : public PrePostActionTy { 829 CGOpenMPRuntimeGPU &RT; 830 CGOpenMPRuntimeGPU::EntryFunctionState &EST; 831 bool IsBareKernel; 832 DataSharingMode Mode; 833 const OMPExecutableDirective &D; 834 835 public: 836 NVPTXPrePostActionTy(CGOpenMPRuntimeGPU &RT, 837 CGOpenMPRuntimeGPU::EntryFunctionState &EST, 838 bool IsBareKernel, const OMPExecutableDirective &D) 839 : RT(RT), EST(EST), IsBareKernel(IsBareKernel), 840 Mode(RT.CurrentDataSharingMode), D(D) {} 841 void Enter(CodeGenFunction &CGF) override { 842 if (IsBareKernel) { 843 RT.CurrentDataSharingMode = DataSharingMode::DS_CUDA; 844 return; 845 } 846 RT.emitKernelInit(D, CGF, EST, /* IsSPMD */ true); 847 // Skip target region initialization. 848 RT.setLocThreadIdInsertPt(CGF, /*AtCurrentPoint=*/true); 849 } 850 void Exit(CodeGenFunction &CGF) override { 851 if (IsBareKernel) { 852 RT.CurrentDataSharingMode = Mode; 853 return; 854 } 855 RT.clearLocThreadIdInsertPt(CGF); 856 RT.emitKernelDeinit(CGF, EST, /* IsSPMD */ true); 857 } 858 } Action(*this, EST, IsBareKernel, D); 859 CodeGen.setAction(Action); 860 IsInTTDRegion = true; 861 emitTargetOutlinedFunctionHelper(D, ParentName, OutlinedFn, OutlinedFnID, 862 IsOffloadEntry, CodeGen); 863 IsInTTDRegion = false; 864 } 865 866 void CGOpenMPRuntimeGPU::emitTargetOutlinedFunction( 867 const OMPExecutableDirective &D, StringRef ParentName, 868 llvm::Function *&OutlinedFn, llvm::Constant *&OutlinedFnID, 869 bool IsOffloadEntry, const RegionCodeGenTy &CodeGen) { 870 if (!IsOffloadEntry) // Nothing to do. 871 return; 872 873 assert(!ParentName.empty() && "Invalid target region parent name!"); 874 875 bool Mode = supportsSPMDExecutionMode(CGM.getContext(), D); 876 bool IsBareKernel = D.getSingleClause<OMPXBareClause>(); 877 if (Mode || IsBareKernel) 878 emitSPMDKernel(D, ParentName, OutlinedFn, OutlinedFnID, IsOffloadEntry, 879 CodeGen); 880 else 881 emitNonSPMDKernel(D, ParentName, OutlinedFn, OutlinedFnID, IsOffloadEntry, 882 CodeGen); 883 } 884 885 CGOpenMPRuntimeGPU::CGOpenMPRuntimeGPU(CodeGenModule &CGM) 886 : CGOpenMPRuntime(CGM) { 887 llvm::OpenMPIRBuilderConfig Config( 888 CGM.getLangOpts().OpenMPIsTargetDevice, isGPU(), 889 CGM.getLangOpts().OpenMPOffloadMandatory, 890 /*HasRequiresReverseOffload*/ false, /*HasRequiresUnifiedAddress*/ false, 891 hasRequiresUnifiedSharedMemory(), /*HasRequiresDynamicAllocators*/ false); 892 OMPBuilder.setConfig(Config); 893 894 if (!CGM.getLangOpts().OpenMPIsTargetDevice) 895 llvm_unreachable("OpenMP can only handle device code."); 896 897 if (CGM.getLangOpts().OpenMPCUDAMode) 898 CurrentDataSharingMode = CGOpenMPRuntimeGPU::DS_CUDA; 899 900 llvm::OpenMPIRBuilder &OMPBuilder = getOMPBuilder(); 901 if (CGM.getLangOpts().NoGPULib || CGM.getLangOpts().OMPHostIRFile.empty()) 902 return; 903 904 OMPBuilder.createGlobalFlag(CGM.getLangOpts().OpenMPTargetDebug, 905 "__omp_rtl_debug_kind"); 906 OMPBuilder.createGlobalFlag(CGM.getLangOpts().OpenMPTeamSubscription, 907 "__omp_rtl_assume_teams_oversubscription"); 908 OMPBuilder.createGlobalFlag(CGM.getLangOpts().OpenMPThreadSubscription, 909 "__omp_rtl_assume_threads_oversubscription"); 910 OMPBuilder.createGlobalFlag(CGM.getLangOpts().OpenMPNoThreadState, 911 "__omp_rtl_assume_no_thread_state"); 912 OMPBuilder.createGlobalFlag(CGM.getLangOpts().OpenMPNoNestedParallelism, 913 "__omp_rtl_assume_no_nested_parallelism"); 914 } 915 916 void CGOpenMPRuntimeGPU::emitProcBindClause(CodeGenFunction &CGF, 917 ProcBindKind ProcBind, 918 SourceLocation Loc) { 919 // Nothing to do. 920 } 921 922 void CGOpenMPRuntimeGPU::emitNumThreadsClause(CodeGenFunction &CGF, 923 llvm::Value *NumThreads, 924 SourceLocation Loc) { 925 // Nothing to do. 926 } 927 928 void CGOpenMPRuntimeGPU::emitNumTeamsClause(CodeGenFunction &CGF, 929 const Expr *NumTeams, 930 const Expr *ThreadLimit, 931 SourceLocation Loc) {} 932 933 llvm::Function *CGOpenMPRuntimeGPU::emitParallelOutlinedFunction( 934 CodeGenFunction &CGF, const OMPExecutableDirective &D, 935 const VarDecl *ThreadIDVar, OpenMPDirectiveKind InnermostKind, 936 const RegionCodeGenTy &CodeGen) { 937 // Emit target region as a standalone region. 938 bool PrevIsInTTDRegion = IsInTTDRegion; 939 IsInTTDRegion = false; 940 auto *OutlinedFun = 941 cast<llvm::Function>(CGOpenMPRuntime::emitParallelOutlinedFunction( 942 CGF, D, ThreadIDVar, InnermostKind, CodeGen)); 943 IsInTTDRegion = PrevIsInTTDRegion; 944 if (getExecutionMode() != CGOpenMPRuntimeGPU::EM_SPMD) { 945 llvm::Function *WrapperFun = 946 createParallelDataSharingWrapper(OutlinedFun, D); 947 WrapperFunctionsMap[OutlinedFun] = WrapperFun; 948 } 949 950 return OutlinedFun; 951 } 952 953 /// Get list of lastprivate variables from the teams distribute ... or 954 /// teams {distribute ...} directives. 955 static void 956 getDistributeLastprivateVars(ASTContext &Ctx, const OMPExecutableDirective &D, 957 llvm::SmallVectorImpl<const ValueDecl *> &Vars) { 958 assert(isOpenMPTeamsDirective(D.getDirectiveKind()) && 959 "expected teams directive."); 960 const OMPExecutableDirective *Dir = &D; 961 if (!isOpenMPDistributeDirective(D.getDirectiveKind())) { 962 if (const Stmt *S = CGOpenMPRuntime::getSingleCompoundChild( 963 Ctx, 964 D.getInnermostCapturedStmt()->getCapturedStmt()->IgnoreContainers( 965 /*IgnoreCaptured=*/true))) { 966 Dir = dyn_cast_or_null<OMPExecutableDirective>(S); 967 if (Dir && !isOpenMPDistributeDirective(Dir->getDirectiveKind())) 968 Dir = nullptr; 969 } 970 } 971 if (!Dir) 972 return; 973 for (const auto *C : Dir->getClausesOfKind<OMPLastprivateClause>()) { 974 for (const Expr *E : C->getVarRefs()) 975 Vars.push_back(getPrivateItem(E)); 976 } 977 } 978 979 /// Get list of reduction variables from the teams ... directives. 980 static void 981 getTeamsReductionVars(ASTContext &Ctx, const OMPExecutableDirective &D, 982 llvm::SmallVectorImpl<const ValueDecl *> &Vars) { 983 assert(isOpenMPTeamsDirective(D.getDirectiveKind()) && 984 "expected teams directive."); 985 for (const auto *C : D.getClausesOfKind<OMPReductionClause>()) { 986 for (const Expr *E : C->privates()) 987 Vars.push_back(getPrivateItem(E)); 988 } 989 } 990 991 llvm::Function *CGOpenMPRuntimeGPU::emitTeamsOutlinedFunction( 992 CodeGenFunction &CGF, const OMPExecutableDirective &D, 993 const VarDecl *ThreadIDVar, OpenMPDirectiveKind InnermostKind, 994 const RegionCodeGenTy &CodeGen) { 995 SourceLocation Loc = D.getBeginLoc(); 996 997 const RecordDecl *GlobalizedRD = nullptr; 998 llvm::SmallVector<const ValueDecl *, 4> LastPrivatesReductions; 999 llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> MappedDeclsFields; 1000 unsigned WarpSize = CGM.getTarget().getGridValue().GV_Warp_Size; 1001 // Globalize team reductions variable unconditionally in all modes. 1002 if (getExecutionMode() != CGOpenMPRuntimeGPU::EM_SPMD) 1003 getTeamsReductionVars(CGM.getContext(), D, LastPrivatesReductions); 1004 if (getExecutionMode() == CGOpenMPRuntimeGPU::EM_SPMD) { 1005 getDistributeLastprivateVars(CGM.getContext(), D, LastPrivatesReductions); 1006 if (!LastPrivatesReductions.empty()) { 1007 GlobalizedRD = ::buildRecordForGlobalizedVars( 1008 CGM.getContext(), std::nullopt, LastPrivatesReductions, 1009 MappedDeclsFields, WarpSize); 1010 } 1011 } else if (!LastPrivatesReductions.empty()) { 1012 assert(!TeamAndReductions.first && 1013 "Previous team declaration is not expected."); 1014 TeamAndReductions.first = D.getCapturedStmt(OMPD_teams)->getCapturedDecl(); 1015 std::swap(TeamAndReductions.second, LastPrivatesReductions); 1016 } 1017 1018 // Emit target region as a standalone region. 1019 class NVPTXPrePostActionTy : public PrePostActionTy { 1020 SourceLocation &Loc; 1021 const RecordDecl *GlobalizedRD; 1022 llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> 1023 &MappedDeclsFields; 1024 1025 public: 1026 NVPTXPrePostActionTy( 1027 SourceLocation &Loc, const RecordDecl *GlobalizedRD, 1028 llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> 1029 &MappedDeclsFields) 1030 : Loc(Loc), GlobalizedRD(GlobalizedRD), 1031 MappedDeclsFields(MappedDeclsFields) {} 1032 void Enter(CodeGenFunction &CGF) override { 1033 auto &Rt = 1034 static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()); 1035 if (GlobalizedRD) { 1036 auto I = Rt.FunctionGlobalizedDecls.try_emplace(CGF.CurFn).first; 1037 I->getSecond().MappedParams = 1038 std::make_unique<CodeGenFunction::OMPMapVars>(); 1039 DeclToAddrMapTy &Data = I->getSecond().LocalVarData; 1040 for (const auto &Pair : MappedDeclsFields) { 1041 assert(Pair.getFirst()->isCanonicalDecl() && 1042 "Expected canonical declaration"); 1043 Data.insert(std::make_pair(Pair.getFirst(), MappedVarData())); 1044 } 1045 } 1046 Rt.emitGenericVarsProlog(CGF, Loc); 1047 } 1048 void Exit(CodeGenFunction &CGF) override { 1049 static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()) 1050 .emitGenericVarsEpilog(CGF); 1051 } 1052 } Action(Loc, GlobalizedRD, MappedDeclsFields); 1053 CodeGen.setAction(Action); 1054 llvm::Function *OutlinedFun = CGOpenMPRuntime::emitTeamsOutlinedFunction( 1055 CGF, D, ThreadIDVar, InnermostKind, CodeGen); 1056 1057 return OutlinedFun; 1058 } 1059 1060 void CGOpenMPRuntimeGPU::emitGenericVarsProlog(CodeGenFunction &CGF, 1061 SourceLocation Loc) { 1062 if (getDataSharingMode() != CGOpenMPRuntimeGPU::DS_Generic) 1063 return; 1064 1065 CGBuilderTy &Bld = CGF.Builder; 1066 1067 const auto I = FunctionGlobalizedDecls.find(CGF.CurFn); 1068 if (I == FunctionGlobalizedDecls.end()) 1069 return; 1070 1071 for (auto &Rec : I->getSecond().LocalVarData) { 1072 const auto *VD = cast<VarDecl>(Rec.first); 1073 bool EscapedParam = I->getSecond().EscapedParameters.count(Rec.first); 1074 QualType VarTy = VD->getType(); 1075 1076 // Get the local allocation of a firstprivate variable before sharing 1077 llvm::Value *ParValue; 1078 if (EscapedParam) { 1079 LValue ParLVal = 1080 CGF.MakeAddrLValue(CGF.GetAddrOfLocalVar(VD), VD->getType()); 1081 ParValue = CGF.EmitLoadOfScalar(ParLVal, Loc); 1082 } 1083 1084 // Allocate space for the variable to be globalized 1085 llvm::Value *AllocArgs[] = {CGF.getTypeSize(VD->getType())}; 1086 llvm::CallBase *VoidPtr = 1087 CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( 1088 CGM.getModule(), OMPRTL___kmpc_alloc_shared), 1089 AllocArgs, VD->getName()); 1090 // FIXME: We should use the variables actual alignment as an argument. 1091 VoidPtr->addRetAttr(llvm::Attribute::get( 1092 CGM.getLLVMContext(), llvm::Attribute::Alignment, 1093 CGM.getContext().getTargetInfo().getNewAlign() / 8)); 1094 1095 // Cast the void pointer and get the address of the globalized variable. 1096 llvm::PointerType *VarPtrTy = CGF.ConvertTypeForMem(VarTy)->getPointerTo(); 1097 llvm::Value *CastedVoidPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( 1098 VoidPtr, VarPtrTy, VD->getName() + "_on_stack"); 1099 LValue VarAddr = CGF.MakeNaturalAlignAddrLValue(CastedVoidPtr, VarTy); 1100 Rec.second.PrivateAddr = VarAddr.getAddress(CGF); 1101 Rec.second.GlobalizedVal = VoidPtr; 1102 1103 // Assign the local allocation to the newly globalized location. 1104 if (EscapedParam) { 1105 CGF.EmitStoreOfScalar(ParValue, VarAddr); 1106 I->getSecond().MappedParams->setVarAddr(CGF, VD, VarAddr.getAddress(CGF)); 1107 } 1108 if (auto *DI = CGF.getDebugInfo()) 1109 VoidPtr->setDebugLoc(DI->SourceLocToDebugLoc(VD->getLocation())); 1110 } 1111 1112 for (const auto *ValueD : I->getSecond().EscapedVariableLengthDecls) { 1113 const auto *VD = cast<VarDecl>(ValueD); 1114 std::pair<llvm::Value *, llvm::Value *> AddrSizePair = 1115 getKmpcAllocShared(CGF, VD); 1116 I->getSecond().EscapedVariableLengthDeclsAddrs.emplace_back(AddrSizePair); 1117 LValue Base = CGF.MakeAddrLValue(AddrSizePair.first, VD->getType(), 1118 CGM.getContext().getDeclAlign(VD), 1119 AlignmentSource::Decl); 1120 I->getSecond().MappedParams->setVarAddr(CGF, VD, Base.getAddress(CGF)); 1121 } 1122 I->getSecond().MappedParams->apply(CGF); 1123 } 1124 1125 bool CGOpenMPRuntimeGPU::isDelayedVariableLengthDecl(CodeGenFunction &CGF, 1126 const VarDecl *VD) const { 1127 const auto I = FunctionGlobalizedDecls.find(CGF.CurFn); 1128 if (I == FunctionGlobalizedDecls.end()) 1129 return false; 1130 1131 // Check variable declaration is delayed: 1132 return llvm::is_contained(I->getSecond().DelayedVariableLengthDecls, VD); 1133 } 1134 1135 std::pair<llvm::Value *, llvm::Value *> 1136 CGOpenMPRuntimeGPU::getKmpcAllocShared(CodeGenFunction &CGF, 1137 const VarDecl *VD) { 1138 CGBuilderTy &Bld = CGF.Builder; 1139 1140 // Compute size and alignment. 1141 llvm::Value *Size = CGF.getTypeSize(VD->getType()); 1142 CharUnits Align = CGM.getContext().getDeclAlign(VD); 1143 Size = Bld.CreateNUWAdd( 1144 Size, llvm::ConstantInt::get(CGF.SizeTy, Align.getQuantity() - 1)); 1145 llvm::Value *AlignVal = 1146 llvm::ConstantInt::get(CGF.SizeTy, Align.getQuantity()); 1147 Size = Bld.CreateUDiv(Size, AlignVal); 1148 Size = Bld.CreateNUWMul(Size, AlignVal); 1149 1150 // Allocate space for this VLA object to be globalized. 1151 llvm::Value *AllocArgs[] = {Size}; 1152 llvm::CallBase *VoidPtr = 1153 CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( 1154 CGM.getModule(), OMPRTL___kmpc_alloc_shared), 1155 AllocArgs, VD->getName()); 1156 VoidPtr->addRetAttr(llvm::Attribute::get( 1157 CGM.getLLVMContext(), llvm::Attribute::Alignment, Align.getQuantity())); 1158 1159 return std::make_pair(VoidPtr, Size); 1160 } 1161 1162 void CGOpenMPRuntimeGPU::getKmpcFreeShared( 1163 CodeGenFunction &CGF, 1164 const std::pair<llvm::Value *, llvm::Value *> &AddrSizePair) { 1165 // Deallocate the memory for each globalized VLA object 1166 CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( 1167 CGM.getModule(), OMPRTL___kmpc_free_shared), 1168 {AddrSizePair.first, AddrSizePair.second}); 1169 } 1170 1171 void CGOpenMPRuntimeGPU::emitGenericVarsEpilog(CodeGenFunction &CGF) { 1172 if (getDataSharingMode() != CGOpenMPRuntimeGPU::DS_Generic) 1173 return; 1174 1175 const auto I = FunctionGlobalizedDecls.find(CGF.CurFn); 1176 if (I != FunctionGlobalizedDecls.end()) { 1177 // Deallocate the memory for each globalized VLA object that was 1178 // globalized in the prolog (i.e. emitGenericVarsProlog). 1179 for (const auto &AddrSizePair : 1180 llvm::reverse(I->getSecond().EscapedVariableLengthDeclsAddrs)) { 1181 CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( 1182 CGM.getModule(), OMPRTL___kmpc_free_shared), 1183 {AddrSizePair.first, AddrSizePair.second}); 1184 } 1185 // Deallocate the memory for each globalized value 1186 for (auto &Rec : llvm::reverse(I->getSecond().LocalVarData)) { 1187 const auto *VD = cast<VarDecl>(Rec.first); 1188 I->getSecond().MappedParams->restore(CGF); 1189 1190 llvm::Value *FreeArgs[] = {Rec.second.GlobalizedVal, 1191 CGF.getTypeSize(VD->getType())}; 1192 CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( 1193 CGM.getModule(), OMPRTL___kmpc_free_shared), 1194 FreeArgs); 1195 } 1196 } 1197 } 1198 1199 void CGOpenMPRuntimeGPU::emitTeamsCall(CodeGenFunction &CGF, 1200 const OMPExecutableDirective &D, 1201 SourceLocation Loc, 1202 llvm::Function *OutlinedFn, 1203 ArrayRef<llvm::Value *> CapturedVars) { 1204 if (!CGF.HaveInsertPoint()) 1205 return; 1206 1207 bool IsBareKernel = D.getSingleClause<OMPXBareClause>(); 1208 1209 Address ZeroAddr = CGF.CreateDefaultAlignTempAlloca(CGF.Int32Ty, 1210 /*Name=*/".zero.addr"); 1211 CGF.Builder.CreateStore(CGF.Builder.getInt32(/*C*/ 0), ZeroAddr); 1212 llvm::SmallVector<llvm::Value *, 16> OutlinedFnArgs; 1213 // We don't emit any thread id function call in bare kernel, but because the 1214 // outlined function has a pointer argument, we emit a nullptr here. 1215 if (IsBareKernel) 1216 OutlinedFnArgs.push_back(llvm::ConstantPointerNull::get(CGM.VoidPtrTy)); 1217 else 1218 OutlinedFnArgs.push_back(emitThreadIDAddress(CGF, Loc).getPointer()); 1219 OutlinedFnArgs.push_back(ZeroAddr.getPointer()); 1220 OutlinedFnArgs.append(CapturedVars.begin(), CapturedVars.end()); 1221 emitOutlinedFunctionCall(CGF, Loc, OutlinedFn, OutlinedFnArgs); 1222 } 1223 1224 void CGOpenMPRuntimeGPU::emitParallelCall(CodeGenFunction &CGF, 1225 SourceLocation Loc, 1226 llvm::Function *OutlinedFn, 1227 ArrayRef<llvm::Value *> CapturedVars, 1228 const Expr *IfCond, 1229 llvm::Value *NumThreads) { 1230 if (!CGF.HaveInsertPoint()) 1231 return; 1232 1233 auto &&ParallelGen = [this, Loc, OutlinedFn, CapturedVars, IfCond, 1234 NumThreads](CodeGenFunction &CGF, 1235 PrePostActionTy &Action) { 1236 CGBuilderTy &Bld = CGF.Builder; 1237 llvm::Value *NumThreadsVal = NumThreads; 1238 llvm::Function *WFn = WrapperFunctionsMap[OutlinedFn]; 1239 llvm::Value *ID = llvm::ConstantPointerNull::get(CGM.Int8PtrTy); 1240 if (WFn) 1241 ID = Bld.CreateBitOrPointerCast(WFn, CGM.Int8PtrTy); 1242 llvm::Value *FnPtr = Bld.CreateBitOrPointerCast(OutlinedFn, CGM.Int8PtrTy); 1243 1244 // Create a private scope that will globalize the arguments 1245 // passed from the outside of the target region. 1246 // TODO: Is that needed? 1247 CodeGenFunction::OMPPrivateScope PrivateArgScope(CGF); 1248 1249 Address CapturedVarsAddrs = CGF.CreateDefaultAlignTempAlloca( 1250 llvm::ArrayType::get(CGM.VoidPtrTy, CapturedVars.size()), 1251 "captured_vars_addrs"); 1252 // There's something to share. 1253 if (!CapturedVars.empty()) { 1254 // Prepare for parallel region. Indicate the outlined function. 1255 ASTContext &Ctx = CGF.getContext(); 1256 unsigned Idx = 0; 1257 for (llvm::Value *V : CapturedVars) { 1258 Address Dst = Bld.CreateConstArrayGEP(CapturedVarsAddrs, Idx); 1259 llvm::Value *PtrV; 1260 if (V->getType()->isIntegerTy()) 1261 PtrV = Bld.CreateIntToPtr(V, CGF.VoidPtrTy); 1262 else 1263 PtrV = Bld.CreatePointerBitCastOrAddrSpaceCast(V, CGF.VoidPtrTy); 1264 CGF.EmitStoreOfScalar(PtrV, Dst, /*Volatile=*/false, 1265 Ctx.getPointerType(Ctx.VoidPtrTy)); 1266 ++Idx; 1267 } 1268 } 1269 1270 llvm::Value *IfCondVal = nullptr; 1271 if (IfCond) 1272 IfCondVal = Bld.CreateIntCast(CGF.EvaluateExprAsBool(IfCond), CGF.Int32Ty, 1273 /* isSigned */ false); 1274 else 1275 IfCondVal = llvm::ConstantInt::get(CGF.Int32Ty, 1); 1276 1277 if (!NumThreadsVal) 1278 NumThreadsVal = llvm::ConstantInt::get(CGF.Int32Ty, -1); 1279 else 1280 NumThreadsVal = Bld.CreateZExtOrTrunc(NumThreadsVal, CGF.Int32Ty), 1281 1282 assert(IfCondVal && "Expected a value"); 1283 llvm::Value *RTLoc = emitUpdateLocation(CGF, Loc); 1284 llvm::Value *Args[] = { 1285 RTLoc, 1286 getThreadID(CGF, Loc), 1287 IfCondVal, 1288 NumThreadsVal, 1289 llvm::ConstantInt::get(CGF.Int32Ty, -1), 1290 FnPtr, 1291 ID, 1292 Bld.CreateBitOrPointerCast(CapturedVarsAddrs.getPointer(), 1293 CGF.VoidPtrPtrTy), 1294 llvm::ConstantInt::get(CGM.SizeTy, CapturedVars.size())}; 1295 CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( 1296 CGM.getModule(), OMPRTL___kmpc_parallel_51), 1297 Args); 1298 }; 1299 1300 RegionCodeGenTy RCG(ParallelGen); 1301 RCG(CGF); 1302 } 1303 1304 void CGOpenMPRuntimeGPU::syncCTAThreads(CodeGenFunction &CGF) { 1305 // Always emit simple barriers! 1306 if (!CGF.HaveInsertPoint()) 1307 return; 1308 // Build call __kmpc_barrier_simple_spmd(nullptr, 0); 1309 // This function does not use parameters, so we can emit just default values. 1310 llvm::Value *Args[] = { 1311 llvm::ConstantPointerNull::get( 1312 cast<llvm::PointerType>(getIdentTyPointerTy())), 1313 llvm::ConstantInt::get(CGF.Int32Ty, /*V=*/0, /*isSigned=*/true)}; 1314 CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( 1315 CGM.getModule(), OMPRTL___kmpc_barrier_simple_spmd), 1316 Args); 1317 } 1318 1319 void CGOpenMPRuntimeGPU::emitBarrierCall(CodeGenFunction &CGF, 1320 SourceLocation Loc, 1321 OpenMPDirectiveKind Kind, bool, 1322 bool) { 1323 // Always emit simple barriers! 1324 if (!CGF.HaveInsertPoint()) 1325 return; 1326 // Build call __kmpc_cancel_barrier(loc, thread_id); 1327 unsigned Flags = getDefaultFlagsForBarriers(Kind); 1328 llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc, Flags), 1329 getThreadID(CGF, Loc)}; 1330 1331 CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( 1332 CGM.getModule(), OMPRTL___kmpc_barrier), 1333 Args); 1334 } 1335 1336 void CGOpenMPRuntimeGPU::emitCriticalRegion( 1337 CodeGenFunction &CGF, StringRef CriticalName, 1338 const RegionCodeGenTy &CriticalOpGen, SourceLocation Loc, 1339 const Expr *Hint) { 1340 llvm::BasicBlock *LoopBB = CGF.createBasicBlock("omp.critical.loop"); 1341 llvm::BasicBlock *TestBB = CGF.createBasicBlock("omp.critical.test"); 1342 llvm::BasicBlock *SyncBB = CGF.createBasicBlock("omp.critical.sync"); 1343 llvm::BasicBlock *BodyBB = CGF.createBasicBlock("omp.critical.body"); 1344 llvm::BasicBlock *ExitBB = CGF.createBasicBlock("omp.critical.exit"); 1345 1346 auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()); 1347 1348 // Get the mask of active threads in the warp. 1349 llvm::Value *Mask = CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( 1350 CGM.getModule(), OMPRTL___kmpc_warp_active_thread_mask)); 1351 // Fetch team-local id of the thread. 1352 llvm::Value *ThreadID = RT.getGPUThreadID(CGF); 1353 1354 // Get the width of the team. 1355 llvm::Value *TeamWidth = RT.getGPUNumThreads(CGF); 1356 1357 // Initialize the counter variable for the loop. 1358 QualType Int32Ty = 1359 CGF.getContext().getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/0); 1360 Address Counter = CGF.CreateMemTemp(Int32Ty, "critical_counter"); 1361 LValue CounterLVal = CGF.MakeAddrLValue(Counter, Int32Ty); 1362 CGF.EmitStoreOfScalar(llvm::Constant::getNullValue(CGM.Int32Ty), CounterLVal, 1363 /*isInit=*/true); 1364 1365 // Block checks if loop counter exceeds upper bound. 1366 CGF.EmitBlock(LoopBB); 1367 llvm::Value *CounterVal = CGF.EmitLoadOfScalar(CounterLVal, Loc); 1368 llvm::Value *CmpLoopBound = CGF.Builder.CreateICmpSLT(CounterVal, TeamWidth); 1369 CGF.Builder.CreateCondBr(CmpLoopBound, TestBB, ExitBB); 1370 1371 // Block tests which single thread should execute region, and which threads 1372 // should go straight to synchronisation point. 1373 CGF.EmitBlock(TestBB); 1374 CounterVal = CGF.EmitLoadOfScalar(CounterLVal, Loc); 1375 llvm::Value *CmpThreadToCounter = 1376 CGF.Builder.CreateICmpEQ(ThreadID, CounterVal); 1377 CGF.Builder.CreateCondBr(CmpThreadToCounter, BodyBB, SyncBB); 1378 1379 // Block emits the body of the critical region. 1380 CGF.EmitBlock(BodyBB); 1381 1382 // Output the critical statement. 1383 CGOpenMPRuntime::emitCriticalRegion(CGF, CriticalName, CriticalOpGen, Loc, 1384 Hint); 1385 1386 // After the body surrounded by the critical region, the single executing 1387 // thread will jump to the synchronisation point. 1388 // Block waits for all threads in current team to finish then increments the 1389 // counter variable and returns to the loop. 1390 CGF.EmitBlock(SyncBB); 1391 // Reconverge active threads in the warp. 1392 (void)CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( 1393 CGM.getModule(), OMPRTL___kmpc_syncwarp), 1394 Mask); 1395 1396 llvm::Value *IncCounterVal = 1397 CGF.Builder.CreateNSWAdd(CounterVal, CGF.Builder.getInt32(1)); 1398 CGF.EmitStoreOfScalar(IncCounterVal, CounterLVal); 1399 CGF.EmitBranch(LoopBB); 1400 1401 // Block that is reached when all threads in the team complete the region. 1402 CGF.EmitBlock(ExitBB, /*IsFinished=*/true); 1403 } 1404 1405 /// Cast value to the specified type. 1406 static llvm::Value *castValueToType(CodeGenFunction &CGF, llvm::Value *Val, 1407 QualType ValTy, QualType CastTy, 1408 SourceLocation Loc) { 1409 assert(!CGF.getContext().getTypeSizeInChars(CastTy).isZero() && 1410 "Cast type must sized."); 1411 assert(!CGF.getContext().getTypeSizeInChars(ValTy).isZero() && 1412 "Val type must sized."); 1413 llvm::Type *LLVMCastTy = CGF.ConvertTypeForMem(CastTy); 1414 if (ValTy == CastTy) 1415 return Val; 1416 if (CGF.getContext().getTypeSizeInChars(ValTy) == 1417 CGF.getContext().getTypeSizeInChars(CastTy)) 1418 return CGF.Builder.CreateBitCast(Val, LLVMCastTy); 1419 if (CastTy->isIntegerType() && ValTy->isIntegerType()) 1420 return CGF.Builder.CreateIntCast(Val, LLVMCastTy, 1421 CastTy->hasSignedIntegerRepresentation()); 1422 Address CastItem = CGF.CreateMemTemp(CastTy); 1423 Address ValCastItem = CastItem.withElementType(Val->getType()); 1424 CGF.EmitStoreOfScalar(Val, ValCastItem, /*Volatile=*/false, ValTy, 1425 LValueBaseInfo(AlignmentSource::Type), 1426 TBAAAccessInfo()); 1427 return CGF.EmitLoadOfScalar(CastItem, /*Volatile=*/false, CastTy, Loc, 1428 LValueBaseInfo(AlignmentSource::Type), 1429 TBAAAccessInfo()); 1430 } 1431 1432 /// This function creates calls to one of two shuffle functions to copy 1433 /// variables between lanes in a warp. 1434 static llvm::Value *createRuntimeShuffleFunction(CodeGenFunction &CGF, 1435 llvm::Value *Elem, 1436 QualType ElemType, 1437 llvm::Value *Offset, 1438 SourceLocation Loc) { 1439 CodeGenModule &CGM = CGF.CGM; 1440 CGBuilderTy &Bld = CGF.Builder; 1441 CGOpenMPRuntimeGPU &RT = 1442 *(static_cast<CGOpenMPRuntimeGPU *>(&CGM.getOpenMPRuntime())); 1443 llvm::OpenMPIRBuilder &OMPBuilder = RT.getOMPBuilder(); 1444 1445 CharUnits Size = CGF.getContext().getTypeSizeInChars(ElemType); 1446 assert(Size.getQuantity() <= 8 && 1447 "Unsupported bitwidth in shuffle instruction."); 1448 1449 RuntimeFunction ShuffleFn = Size.getQuantity() <= 4 1450 ? OMPRTL___kmpc_shuffle_int32 1451 : OMPRTL___kmpc_shuffle_int64; 1452 1453 // Cast all types to 32- or 64-bit values before calling shuffle routines. 1454 QualType CastTy = CGF.getContext().getIntTypeForBitwidth( 1455 Size.getQuantity() <= 4 ? 32 : 64, /*Signed=*/1); 1456 llvm::Value *ElemCast = castValueToType(CGF, Elem, ElemType, CastTy, Loc); 1457 llvm::Value *WarpSize = 1458 Bld.CreateIntCast(RT.getGPUWarpSize(CGF), CGM.Int16Ty, /*isSigned=*/true); 1459 1460 llvm::Value *ShuffledVal = CGF.EmitRuntimeCall( 1461 OMPBuilder.getOrCreateRuntimeFunction(CGM.getModule(), ShuffleFn), 1462 {ElemCast, Offset, WarpSize}); 1463 1464 return castValueToType(CGF, ShuffledVal, CastTy, ElemType, Loc); 1465 } 1466 1467 static void shuffleAndStore(CodeGenFunction &CGF, Address SrcAddr, 1468 Address DestAddr, QualType ElemType, 1469 llvm::Value *Offset, SourceLocation Loc) { 1470 CGBuilderTy &Bld = CGF.Builder; 1471 1472 CharUnits Size = CGF.getContext().getTypeSizeInChars(ElemType); 1473 // Create the loop over the big sized data. 1474 // ptr = (void*)Elem; 1475 // ptrEnd = (void*) Elem + 1; 1476 // Step = 8; 1477 // while (ptr + Step < ptrEnd) 1478 // shuffle((int64_t)*ptr); 1479 // Step = 4; 1480 // while (ptr + Step < ptrEnd) 1481 // shuffle((int32_t)*ptr); 1482 // ... 1483 Address ElemPtr = DestAddr; 1484 Address Ptr = SrcAddr; 1485 Address PtrEnd = Bld.CreatePointerBitCastOrAddrSpaceCast( 1486 Bld.CreateConstGEP(SrcAddr, 1), CGF.VoidPtrTy, CGF.Int8Ty); 1487 for (int IntSize = 8; IntSize >= 1; IntSize /= 2) { 1488 if (Size < CharUnits::fromQuantity(IntSize)) 1489 continue; 1490 QualType IntType = CGF.getContext().getIntTypeForBitwidth( 1491 CGF.getContext().toBits(CharUnits::fromQuantity(IntSize)), 1492 /*Signed=*/1); 1493 llvm::Type *IntTy = CGF.ConvertTypeForMem(IntType); 1494 Ptr = Bld.CreatePointerBitCastOrAddrSpaceCast(Ptr, IntTy->getPointerTo(), 1495 IntTy); 1496 ElemPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( 1497 ElemPtr, IntTy->getPointerTo(), IntTy); 1498 if (Size.getQuantity() / IntSize > 1) { 1499 llvm::BasicBlock *PreCondBB = CGF.createBasicBlock(".shuffle.pre_cond"); 1500 llvm::BasicBlock *ThenBB = CGF.createBasicBlock(".shuffle.then"); 1501 llvm::BasicBlock *ExitBB = CGF.createBasicBlock(".shuffle.exit"); 1502 llvm::BasicBlock *CurrentBB = Bld.GetInsertBlock(); 1503 CGF.EmitBlock(PreCondBB); 1504 llvm::PHINode *PhiSrc = 1505 Bld.CreatePHI(Ptr.getType(), /*NumReservedValues=*/2); 1506 PhiSrc->addIncoming(Ptr.getPointer(), CurrentBB); 1507 llvm::PHINode *PhiDest = 1508 Bld.CreatePHI(ElemPtr.getType(), /*NumReservedValues=*/2); 1509 PhiDest->addIncoming(ElemPtr.getPointer(), CurrentBB); 1510 Ptr = Address(PhiSrc, Ptr.getElementType(), Ptr.getAlignment()); 1511 ElemPtr = 1512 Address(PhiDest, ElemPtr.getElementType(), ElemPtr.getAlignment()); 1513 llvm::Value *PtrDiff = Bld.CreatePtrDiff( 1514 CGF.Int8Ty, PtrEnd.getPointer(), 1515 Bld.CreatePointerBitCastOrAddrSpaceCast(Ptr.getPointer(), 1516 CGF.VoidPtrTy)); 1517 Bld.CreateCondBr(Bld.CreateICmpSGT(PtrDiff, Bld.getInt64(IntSize - 1)), 1518 ThenBB, ExitBB); 1519 CGF.EmitBlock(ThenBB); 1520 llvm::Value *Res = createRuntimeShuffleFunction( 1521 CGF, 1522 CGF.EmitLoadOfScalar(Ptr, /*Volatile=*/false, IntType, Loc, 1523 LValueBaseInfo(AlignmentSource::Type), 1524 TBAAAccessInfo()), 1525 IntType, Offset, Loc); 1526 CGF.EmitStoreOfScalar(Res, ElemPtr, /*Volatile=*/false, IntType, 1527 LValueBaseInfo(AlignmentSource::Type), 1528 TBAAAccessInfo()); 1529 Address LocalPtr = Bld.CreateConstGEP(Ptr, 1); 1530 Address LocalElemPtr = Bld.CreateConstGEP(ElemPtr, 1); 1531 PhiSrc->addIncoming(LocalPtr.getPointer(), ThenBB); 1532 PhiDest->addIncoming(LocalElemPtr.getPointer(), ThenBB); 1533 CGF.EmitBranch(PreCondBB); 1534 CGF.EmitBlock(ExitBB); 1535 } else { 1536 llvm::Value *Res = createRuntimeShuffleFunction( 1537 CGF, 1538 CGF.EmitLoadOfScalar(Ptr, /*Volatile=*/false, IntType, Loc, 1539 LValueBaseInfo(AlignmentSource::Type), 1540 TBAAAccessInfo()), 1541 IntType, Offset, Loc); 1542 CGF.EmitStoreOfScalar(Res, ElemPtr, /*Volatile=*/false, IntType, 1543 LValueBaseInfo(AlignmentSource::Type), 1544 TBAAAccessInfo()); 1545 Ptr = Bld.CreateConstGEP(Ptr, 1); 1546 ElemPtr = Bld.CreateConstGEP(ElemPtr, 1); 1547 } 1548 Size = Size % IntSize; 1549 } 1550 } 1551 1552 namespace { 1553 enum CopyAction : unsigned { 1554 // RemoteLaneToThread: Copy over a Reduce list from a remote lane in 1555 // the warp using shuffle instructions. 1556 RemoteLaneToThread, 1557 // ThreadCopy: Make a copy of a Reduce list on the thread's stack. 1558 ThreadCopy, 1559 }; 1560 } // namespace 1561 1562 struct CopyOptionsTy { 1563 llvm::Value *RemoteLaneOffset; 1564 llvm::Value *ScratchpadIndex; 1565 llvm::Value *ScratchpadWidth; 1566 }; 1567 1568 /// Emit instructions to copy a Reduce list, which contains partially 1569 /// aggregated values, in the specified direction. 1570 static void emitReductionListCopy( 1571 CopyAction Action, CodeGenFunction &CGF, QualType ReductionArrayTy, 1572 ArrayRef<const Expr *> Privates, Address SrcBase, Address DestBase, 1573 CopyOptionsTy CopyOptions = {nullptr, nullptr, nullptr}) { 1574 1575 CodeGenModule &CGM = CGF.CGM; 1576 ASTContext &C = CGM.getContext(); 1577 CGBuilderTy &Bld = CGF.Builder; 1578 1579 llvm::Value *RemoteLaneOffset = CopyOptions.RemoteLaneOffset; 1580 1581 // Iterates, element-by-element, through the source Reduce list and 1582 // make a copy. 1583 unsigned Idx = 0; 1584 for (const Expr *Private : Privates) { 1585 Address SrcElementAddr = Address::invalid(); 1586 Address DestElementAddr = Address::invalid(); 1587 Address DestElementPtrAddr = Address::invalid(); 1588 // Should we shuffle in an element from a remote lane? 1589 bool ShuffleInElement = false; 1590 // Set to true to update the pointer in the dest Reduce list to a 1591 // newly created element. 1592 bool UpdateDestListPtr = false; 1593 QualType PrivatePtrType = C.getPointerType(Private->getType()); 1594 llvm::Type *PrivateLlvmPtrType = CGF.ConvertType(PrivatePtrType); 1595 1596 switch (Action) { 1597 case RemoteLaneToThread: { 1598 // Step 1.1: Get the address for the src element in the Reduce list. 1599 Address SrcElementPtrAddr = Bld.CreateConstArrayGEP(SrcBase, Idx); 1600 SrcElementAddr = CGF.EmitLoadOfPointer( 1601 SrcElementPtrAddr.withElementType(PrivateLlvmPtrType), 1602 PrivatePtrType->castAs<PointerType>()); 1603 1604 // Step 1.2: Create a temporary to store the element in the destination 1605 // Reduce list. 1606 DestElementPtrAddr = Bld.CreateConstArrayGEP(DestBase, Idx); 1607 DestElementAddr = 1608 CGF.CreateMemTemp(Private->getType(), ".omp.reduction.element"); 1609 ShuffleInElement = true; 1610 UpdateDestListPtr = true; 1611 break; 1612 } 1613 case ThreadCopy: { 1614 // Step 1.1: Get the address for the src element in the Reduce list. 1615 Address SrcElementPtrAddr = Bld.CreateConstArrayGEP(SrcBase, Idx); 1616 SrcElementAddr = CGF.EmitLoadOfPointer( 1617 SrcElementPtrAddr.withElementType(PrivateLlvmPtrType), 1618 PrivatePtrType->castAs<PointerType>()); 1619 1620 // Step 1.2: Get the address for dest element. The destination 1621 // element has already been created on the thread's stack. 1622 DestElementPtrAddr = Bld.CreateConstArrayGEP(DestBase, Idx); 1623 DestElementAddr = CGF.EmitLoadOfPointer( 1624 DestElementPtrAddr.withElementType(PrivateLlvmPtrType), 1625 PrivatePtrType->castAs<PointerType>()); 1626 break; 1627 } 1628 } 1629 1630 // Regardless of src and dest of copy, we emit the load of src 1631 // element as this is required in all directions 1632 SrcElementAddr = SrcElementAddr.withElementType( 1633 CGF.ConvertTypeForMem(Private->getType())); 1634 DestElementAddr = 1635 DestElementAddr.withElementType(SrcElementAddr.getElementType()); 1636 1637 // Now that all active lanes have read the element in the 1638 // Reduce list, shuffle over the value from the remote lane. 1639 if (ShuffleInElement) { 1640 shuffleAndStore(CGF, SrcElementAddr, DestElementAddr, Private->getType(), 1641 RemoteLaneOffset, Private->getExprLoc()); 1642 } else { 1643 switch (CGF.getEvaluationKind(Private->getType())) { 1644 case TEK_Scalar: { 1645 llvm::Value *Elem = CGF.EmitLoadOfScalar( 1646 SrcElementAddr, /*Volatile=*/false, Private->getType(), 1647 Private->getExprLoc(), LValueBaseInfo(AlignmentSource::Type), 1648 TBAAAccessInfo()); 1649 // Store the source element value to the dest element address. 1650 CGF.EmitStoreOfScalar( 1651 Elem, DestElementAddr, /*Volatile=*/false, Private->getType(), 1652 LValueBaseInfo(AlignmentSource::Type), TBAAAccessInfo()); 1653 break; 1654 } 1655 case TEK_Complex: { 1656 CodeGenFunction::ComplexPairTy Elem = CGF.EmitLoadOfComplex( 1657 CGF.MakeAddrLValue(SrcElementAddr, Private->getType()), 1658 Private->getExprLoc()); 1659 CGF.EmitStoreOfComplex( 1660 Elem, CGF.MakeAddrLValue(DestElementAddr, Private->getType()), 1661 /*isInit=*/false); 1662 break; 1663 } 1664 case TEK_Aggregate: 1665 CGF.EmitAggregateCopy( 1666 CGF.MakeAddrLValue(DestElementAddr, Private->getType()), 1667 CGF.MakeAddrLValue(SrcElementAddr, Private->getType()), 1668 Private->getType(), AggValueSlot::DoesNotOverlap); 1669 break; 1670 } 1671 } 1672 1673 // Step 3.1: Modify reference in dest Reduce list as needed. 1674 // Modifying the reference in Reduce list to point to the newly 1675 // created element. The element is live in the current function 1676 // scope and that of functions it invokes (i.e., reduce_function). 1677 // RemoteReduceData[i] = (void*)&RemoteElem 1678 if (UpdateDestListPtr) { 1679 CGF.EmitStoreOfScalar(Bld.CreatePointerBitCastOrAddrSpaceCast( 1680 DestElementAddr.getPointer(), CGF.VoidPtrTy), 1681 DestElementPtrAddr, /*Volatile=*/false, 1682 C.VoidPtrTy); 1683 } 1684 1685 ++Idx; 1686 } 1687 } 1688 1689 /// This function emits a helper that gathers Reduce lists from the first 1690 /// lane of every active warp to lanes in the first warp. 1691 /// 1692 /// void inter_warp_copy_func(void* reduce_data, num_warps) 1693 /// shared smem[warp_size]; 1694 /// For all data entries D in reduce_data: 1695 /// sync 1696 /// If (I am the first lane in each warp) 1697 /// Copy my local D to smem[warp_id] 1698 /// sync 1699 /// if (I am the first warp) 1700 /// Copy smem[thread_id] to my local D 1701 static llvm::Value *emitInterWarpCopyFunction(CodeGenModule &CGM, 1702 ArrayRef<const Expr *> Privates, 1703 QualType ReductionArrayTy, 1704 SourceLocation Loc) { 1705 ASTContext &C = CGM.getContext(); 1706 llvm::Module &M = CGM.getModule(); 1707 1708 // ReduceList: thread local Reduce list. 1709 // At the stage of the computation when this function is called, partially 1710 // aggregated values reside in the first lane of every active warp. 1711 ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, 1712 C.VoidPtrTy, ImplicitParamKind::Other); 1713 // NumWarps: number of warps active in the parallel region. This could 1714 // be smaller than 32 (max warps in a CTA) for partial block reduction. 1715 ImplicitParamDecl NumWarpsArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, 1716 C.getIntTypeForBitwidth(32, /* Signed */ true), 1717 ImplicitParamKind::Other); 1718 FunctionArgList Args; 1719 Args.push_back(&ReduceListArg); 1720 Args.push_back(&NumWarpsArg); 1721 1722 const CGFunctionInfo &CGFI = 1723 CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); 1724 auto *Fn = llvm::Function::Create(CGM.getTypes().GetFunctionType(CGFI), 1725 llvm::GlobalValue::InternalLinkage, 1726 "_omp_reduction_inter_warp_copy_func", &M); 1727 CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI); 1728 Fn->setDoesNotRecurse(); 1729 CodeGenFunction CGF(CGM); 1730 CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc); 1731 1732 CGBuilderTy &Bld = CGF.Builder; 1733 1734 // This array is used as a medium to transfer, one reduce element at a time, 1735 // the data from the first lane of every warp to lanes in the first warp 1736 // in order to perform the final step of a reduction in a parallel region 1737 // (reduction across warps). The array is placed in NVPTX __shared__ memory 1738 // for reduced latency, as well as to have a distinct copy for concurrently 1739 // executing target regions. The array is declared with common linkage so 1740 // as to be shared across compilation units. 1741 StringRef TransferMediumName = 1742 "__openmp_nvptx_data_transfer_temporary_storage"; 1743 llvm::GlobalVariable *TransferMedium = 1744 M.getGlobalVariable(TransferMediumName); 1745 unsigned WarpSize = CGF.getTarget().getGridValue().GV_Warp_Size; 1746 if (!TransferMedium) { 1747 auto *Ty = llvm::ArrayType::get(CGM.Int32Ty, WarpSize); 1748 unsigned SharedAddressSpace = C.getTargetAddressSpace(LangAS::cuda_shared); 1749 TransferMedium = new llvm::GlobalVariable( 1750 M, Ty, /*isConstant=*/false, llvm::GlobalVariable::WeakAnyLinkage, 1751 llvm::UndefValue::get(Ty), TransferMediumName, 1752 /*InsertBefore=*/nullptr, llvm::GlobalVariable::NotThreadLocal, 1753 SharedAddressSpace); 1754 CGM.addCompilerUsedGlobal(TransferMedium); 1755 } 1756 1757 auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()); 1758 // Get the CUDA thread id of the current OpenMP thread on the GPU. 1759 llvm::Value *ThreadID = RT.getGPUThreadID(CGF); 1760 // nvptx_lane_id = nvptx_id % warpsize 1761 llvm::Value *LaneID = getNVPTXLaneID(CGF); 1762 // nvptx_warp_id = nvptx_id / warpsize 1763 llvm::Value *WarpID = getNVPTXWarpID(CGF); 1764 1765 Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg); 1766 llvm::Type *ElemTy = CGF.ConvertTypeForMem(ReductionArrayTy); 1767 Address LocalReduceList( 1768 Bld.CreatePointerBitCastOrAddrSpaceCast( 1769 CGF.EmitLoadOfScalar( 1770 AddrReduceListArg, /*Volatile=*/false, C.VoidPtrTy, Loc, 1771 LValueBaseInfo(AlignmentSource::Type), TBAAAccessInfo()), 1772 ElemTy->getPointerTo()), 1773 ElemTy, CGF.getPointerAlign()); 1774 1775 unsigned Idx = 0; 1776 for (const Expr *Private : Privates) { 1777 // 1778 // Warp master copies reduce element to transfer medium in __shared__ 1779 // memory. 1780 // 1781 unsigned RealTySize = 1782 C.getTypeSizeInChars(Private->getType()) 1783 .alignTo(C.getTypeAlignInChars(Private->getType())) 1784 .getQuantity(); 1785 for (unsigned TySize = 4; TySize > 0 && RealTySize > 0; TySize /=2) { 1786 unsigned NumIters = RealTySize / TySize; 1787 if (NumIters == 0) 1788 continue; 1789 QualType CType = C.getIntTypeForBitwidth( 1790 C.toBits(CharUnits::fromQuantity(TySize)), /*Signed=*/1); 1791 llvm::Type *CopyType = CGF.ConvertTypeForMem(CType); 1792 CharUnits Align = CharUnits::fromQuantity(TySize); 1793 llvm::Value *Cnt = nullptr; 1794 Address CntAddr = Address::invalid(); 1795 llvm::BasicBlock *PrecondBB = nullptr; 1796 llvm::BasicBlock *ExitBB = nullptr; 1797 if (NumIters > 1) { 1798 CntAddr = CGF.CreateMemTemp(C.IntTy, ".cnt.addr"); 1799 CGF.EmitStoreOfScalar(llvm::Constant::getNullValue(CGM.IntTy), CntAddr, 1800 /*Volatile=*/false, C.IntTy); 1801 PrecondBB = CGF.createBasicBlock("precond"); 1802 ExitBB = CGF.createBasicBlock("exit"); 1803 llvm::BasicBlock *BodyBB = CGF.createBasicBlock("body"); 1804 // There is no need to emit line number for unconditional branch. 1805 (void)ApplyDebugLocation::CreateEmpty(CGF); 1806 CGF.EmitBlock(PrecondBB); 1807 Cnt = CGF.EmitLoadOfScalar(CntAddr, /*Volatile=*/false, C.IntTy, Loc); 1808 llvm::Value *Cmp = 1809 Bld.CreateICmpULT(Cnt, llvm::ConstantInt::get(CGM.IntTy, NumIters)); 1810 Bld.CreateCondBr(Cmp, BodyBB, ExitBB); 1811 CGF.EmitBlock(BodyBB); 1812 } 1813 // kmpc_barrier. 1814 CGM.getOpenMPRuntime().emitBarrierCall(CGF, Loc, OMPD_unknown, 1815 /*EmitChecks=*/false, 1816 /*ForceSimpleCall=*/true); 1817 llvm::BasicBlock *ThenBB = CGF.createBasicBlock("then"); 1818 llvm::BasicBlock *ElseBB = CGF.createBasicBlock("else"); 1819 llvm::BasicBlock *MergeBB = CGF.createBasicBlock("ifcont"); 1820 1821 // if (lane_id == 0) 1822 llvm::Value *IsWarpMaster = Bld.CreateIsNull(LaneID, "warp_master"); 1823 Bld.CreateCondBr(IsWarpMaster, ThenBB, ElseBB); 1824 CGF.EmitBlock(ThenBB); 1825 1826 // Reduce element = LocalReduceList[i] 1827 Address ElemPtrPtrAddr = Bld.CreateConstArrayGEP(LocalReduceList, Idx); 1828 llvm::Value *ElemPtrPtr = CGF.EmitLoadOfScalar( 1829 ElemPtrPtrAddr, /*Volatile=*/false, C.VoidPtrTy, SourceLocation()); 1830 // elemptr = ((CopyType*)(elemptrptr)) + I 1831 Address ElemPtr(ElemPtrPtr, CopyType, Align); 1832 if (NumIters > 1) 1833 ElemPtr = Bld.CreateGEP(ElemPtr, Cnt); 1834 1835 // Get pointer to location in transfer medium. 1836 // MediumPtr = &medium[warp_id] 1837 llvm::Value *MediumPtrVal = Bld.CreateInBoundsGEP( 1838 TransferMedium->getValueType(), TransferMedium, 1839 {llvm::Constant::getNullValue(CGM.Int64Ty), WarpID}); 1840 // Casting to actual data type. 1841 // MediumPtr = (CopyType*)MediumPtrAddr; 1842 Address MediumPtr(MediumPtrVal, CopyType, Align); 1843 1844 // elem = *elemptr 1845 //*MediumPtr = elem 1846 llvm::Value *Elem = CGF.EmitLoadOfScalar( 1847 ElemPtr, /*Volatile=*/false, CType, Loc, 1848 LValueBaseInfo(AlignmentSource::Type), TBAAAccessInfo()); 1849 // Store the source element value to the dest element address. 1850 CGF.EmitStoreOfScalar(Elem, MediumPtr, /*Volatile=*/true, CType, 1851 LValueBaseInfo(AlignmentSource::Type), 1852 TBAAAccessInfo()); 1853 1854 Bld.CreateBr(MergeBB); 1855 1856 CGF.EmitBlock(ElseBB); 1857 Bld.CreateBr(MergeBB); 1858 1859 CGF.EmitBlock(MergeBB); 1860 1861 // kmpc_barrier. 1862 CGM.getOpenMPRuntime().emitBarrierCall(CGF, Loc, OMPD_unknown, 1863 /*EmitChecks=*/false, 1864 /*ForceSimpleCall=*/true); 1865 1866 // 1867 // Warp 0 copies reduce element from transfer medium. 1868 // 1869 llvm::BasicBlock *W0ThenBB = CGF.createBasicBlock("then"); 1870 llvm::BasicBlock *W0ElseBB = CGF.createBasicBlock("else"); 1871 llvm::BasicBlock *W0MergeBB = CGF.createBasicBlock("ifcont"); 1872 1873 Address AddrNumWarpsArg = CGF.GetAddrOfLocalVar(&NumWarpsArg); 1874 llvm::Value *NumWarpsVal = CGF.EmitLoadOfScalar( 1875 AddrNumWarpsArg, /*Volatile=*/false, C.IntTy, Loc); 1876 1877 // Up to 32 threads in warp 0 are active. 1878 llvm::Value *IsActiveThread = 1879 Bld.CreateICmpULT(ThreadID, NumWarpsVal, "is_active_thread"); 1880 Bld.CreateCondBr(IsActiveThread, W0ThenBB, W0ElseBB); 1881 1882 CGF.EmitBlock(W0ThenBB); 1883 1884 // SrcMediumPtr = &medium[tid] 1885 llvm::Value *SrcMediumPtrVal = Bld.CreateInBoundsGEP( 1886 TransferMedium->getValueType(), TransferMedium, 1887 {llvm::Constant::getNullValue(CGM.Int64Ty), ThreadID}); 1888 // SrcMediumVal = *SrcMediumPtr; 1889 Address SrcMediumPtr(SrcMediumPtrVal, CopyType, Align); 1890 1891 // TargetElemPtr = (CopyType*)(SrcDataAddr[i]) + I 1892 Address TargetElemPtrPtr = Bld.CreateConstArrayGEP(LocalReduceList, Idx); 1893 llvm::Value *TargetElemPtrVal = CGF.EmitLoadOfScalar( 1894 TargetElemPtrPtr, /*Volatile=*/false, C.VoidPtrTy, Loc); 1895 Address TargetElemPtr(TargetElemPtrVal, CopyType, Align); 1896 if (NumIters > 1) 1897 TargetElemPtr = Bld.CreateGEP(TargetElemPtr, Cnt); 1898 1899 // *TargetElemPtr = SrcMediumVal; 1900 llvm::Value *SrcMediumValue = 1901 CGF.EmitLoadOfScalar(SrcMediumPtr, /*Volatile=*/true, CType, Loc); 1902 CGF.EmitStoreOfScalar(SrcMediumValue, TargetElemPtr, /*Volatile=*/false, 1903 CType); 1904 Bld.CreateBr(W0MergeBB); 1905 1906 CGF.EmitBlock(W0ElseBB); 1907 Bld.CreateBr(W0MergeBB); 1908 1909 CGF.EmitBlock(W0MergeBB); 1910 1911 if (NumIters > 1) { 1912 Cnt = Bld.CreateNSWAdd(Cnt, llvm::ConstantInt::get(CGM.IntTy, /*V=*/1)); 1913 CGF.EmitStoreOfScalar(Cnt, CntAddr, /*Volatile=*/false, C.IntTy); 1914 CGF.EmitBranch(PrecondBB); 1915 (void)ApplyDebugLocation::CreateEmpty(CGF); 1916 CGF.EmitBlock(ExitBB); 1917 } 1918 RealTySize %= TySize; 1919 } 1920 ++Idx; 1921 } 1922 1923 CGF.FinishFunction(); 1924 return Fn; 1925 } 1926 1927 /// Emit a helper that reduces data across two OpenMP threads (lanes) 1928 /// in the same warp. It uses shuffle instructions to copy over data from 1929 /// a remote lane's stack. The reduction algorithm performed is specified 1930 /// by the fourth parameter. 1931 /// 1932 /// Algorithm Versions. 1933 /// Full Warp Reduce (argument value 0): 1934 /// This algorithm assumes that all 32 lanes are active and gathers 1935 /// data from these 32 lanes, producing a single resultant value. 1936 /// Contiguous Partial Warp Reduce (argument value 1): 1937 /// This algorithm assumes that only a *contiguous* subset of lanes 1938 /// are active. This happens for the last warp in a parallel region 1939 /// when the user specified num_threads is not an integer multiple of 1940 /// 32. This contiguous subset always starts with the zeroth lane. 1941 /// Partial Warp Reduce (argument value 2): 1942 /// This algorithm gathers data from any number of lanes at any position. 1943 /// All reduced values are stored in the lowest possible lane. The set 1944 /// of problems every algorithm addresses is a super set of those 1945 /// addressable by algorithms with a lower version number. Overhead 1946 /// increases as algorithm version increases. 1947 /// 1948 /// Terminology 1949 /// Reduce element: 1950 /// Reduce element refers to the individual data field with primitive 1951 /// data types to be combined and reduced across threads. 1952 /// Reduce list: 1953 /// Reduce list refers to a collection of local, thread-private 1954 /// reduce elements. 1955 /// Remote Reduce list: 1956 /// Remote Reduce list refers to a collection of remote (relative to 1957 /// the current thread) reduce elements. 1958 /// 1959 /// We distinguish between three states of threads that are important to 1960 /// the implementation of this function. 1961 /// Alive threads: 1962 /// Threads in a warp executing the SIMT instruction, as distinguished from 1963 /// threads that are inactive due to divergent control flow. 1964 /// Active threads: 1965 /// The minimal set of threads that has to be alive upon entry to this 1966 /// function. The computation is correct iff active threads are alive. 1967 /// Some threads are alive but they are not active because they do not 1968 /// contribute to the computation in any useful manner. Turning them off 1969 /// may introduce control flow overheads without any tangible benefits. 1970 /// Effective threads: 1971 /// In order to comply with the argument requirements of the shuffle 1972 /// function, we must keep all lanes holding data alive. But at most 1973 /// half of them perform value aggregation; we refer to this half of 1974 /// threads as effective. The other half is simply handing off their 1975 /// data. 1976 /// 1977 /// Procedure 1978 /// Value shuffle: 1979 /// In this step active threads transfer data from higher lane positions 1980 /// in the warp to lower lane positions, creating Remote Reduce list. 1981 /// Value aggregation: 1982 /// In this step, effective threads combine their thread local Reduce list 1983 /// with Remote Reduce list and store the result in the thread local 1984 /// Reduce list. 1985 /// Value copy: 1986 /// In this step, we deal with the assumption made by algorithm 2 1987 /// (i.e. contiguity assumption). When we have an odd number of lanes 1988 /// active, say 2k+1, only k threads will be effective and therefore k 1989 /// new values will be produced. However, the Reduce list owned by the 1990 /// (2k+1)th thread is ignored in the value aggregation. Therefore 1991 /// we copy the Reduce list from the (2k+1)th lane to (k+1)th lane so 1992 /// that the contiguity assumption still holds. 1993 static llvm::Function *emitShuffleAndReduceFunction( 1994 CodeGenModule &CGM, ArrayRef<const Expr *> Privates, 1995 QualType ReductionArrayTy, llvm::Function *ReduceFn, SourceLocation Loc) { 1996 ASTContext &C = CGM.getContext(); 1997 1998 // Thread local Reduce list used to host the values of data to be reduced. 1999 ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, 2000 C.VoidPtrTy, ImplicitParamKind::Other); 2001 // Current lane id; could be logical. 2002 ImplicitParamDecl LaneIDArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.ShortTy, 2003 ImplicitParamKind::Other); 2004 // Offset of the remote source lane relative to the current lane. 2005 ImplicitParamDecl RemoteLaneOffsetArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, 2006 C.ShortTy, ImplicitParamKind::Other); 2007 // Algorithm version. This is expected to be known at compile time. 2008 ImplicitParamDecl AlgoVerArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, 2009 C.ShortTy, ImplicitParamKind::Other); 2010 FunctionArgList Args; 2011 Args.push_back(&ReduceListArg); 2012 Args.push_back(&LaneIDArg); 2013 Args.push_back(&RemoteLaneOffsetArg); 2014 Args.push_back(&AlgoVerArg); 2015 2016 const CGFunctionInfo &CGFI = 2017 CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); 2018 auto *Fn = llvm::Function::Create( 2019 CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage, 2020 "_omp_reduction_shuffle_and_reduce_func", &CGM.getModule()); 2021 CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI); 2022 Fn->setDoesNotRecurse(); 2023 2024 CodeGenFunction CGF(CGM); 2025 CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc); 2026 2027 CGBuilderTy &Bld = CGF.Builder; 2028 2029 Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg); 2030 llvm::Type *ElemTy = CGF.ConvertTypeForMem(ReductionArrayTy); 2031 Address LocalReduceList( 2032 Bld.CreatePointerBitCastOrAddrSpaceCast( 2033 CGF.EmitLoadOfScalar(AddrReduceListArg, /*Volatile=*/false, 2034 C.VoidPtrTy, SourceLocation()), 2035 ElemTy->getPointerTo()), 2036 ElemTy, CGF.getPointerAlign()); 2037 2038 Address AddrLaneIDArg = CGF.GetAddrOfLocalVar(&LaneIDArg); 2039 llvm::Value *LaneIDArgVal = CGF.EmitLoadOfScalar( 2040 AddrLaneIDArg, /*Volatile=*/false, C.ShortTy, SourceLocation()); 2041 2042 Address AddrRemoteLaneOffsetArg = CGF.GetAddrOfLocalVar(&RemoteLaneOffsetArg); 2043 llvm::Value *RemoteLaneOffsetArgVal = CGF.EmitLoadOfScalar( 2044 AddrRemoteLaneOffsetArg, /*Volatile=*/false, C.ShortTy, SourceLocation()); 2045 2046 Address AddrAlgoVerArg = CGF.GetAddrOfLocalVar(&AlgoVerArg); 2047 llvm::Value *AlgoVerArgVal = CGF.EmitLoadOfScalar( 2048 AddrAlgoVerArg, /*Volatile=*/false, C.ShortTy, SourceLocation()); 2049 2050 // Create a local thread-private variable to host the Reduce list 2051 // from a remote lane. 2052 Address RemoteReduceList = 2053 CGF.CreateMemTemp(ReductionArrayTy, ".omp.reduction.remote_reduce_list"); 2054 2055 // This loop iterates through the list of reduce elements and copies, 2056 // element by element, from a remote lane in the warp to RemoteReduceList, 2057 // hosted on the thread's stack. 2058 emitReductionListCopy(RemoteLaneToThread, CGF, ReductionArrayTy, Privates, 2059 LocalReduceList, RemoteReduceList, 2060 {/*RemoteLaneOffset=*/RemoteLaneOffsetArgVal, 2061 /*ScratchpadIndex=*/nullptr, 2062 /*ScratchpadWidth=*/nullptr}); 2063 2064 // The actions to be performed on the Remote Reduce list is dependent 2065 // on the algorithm version. 2066 // 2067 // if (AlgoVer==0) || (AlgoVer==1 && (LaneId < Offset)) || (AlgoVer==2 && 2068 // LaneId % 2 == 0 && Offset > 0): 2069 // do the reduction value aggregation 2070 // 2071 // The thread local variable Reduce list is mutated in place to host the 2072 // reduced data, which is the aggregated value produced from local and 2073 // remote lanes. 2074 // 2075 // Note that AlgoVer is expected to be a constant integer known at compile 2076 // time. 2077 // When AlgoVer==0, the first conjunction evaluates to true, making 2078 // the entire predicate true during compile time. 2079 // When AlgoVer==1, the second conjunction has only the second part to be 2080 // evaluated during runtime. Other conjunctions evaluates to false 2081 // during compile time. 2082 // When AlgoVer==2, the third conjunction has only the second part to be 2083 // evaluated during runtime. Other conjunctions evaluates to false 2084 // during compile time. 2085 llvm::Value *CondAlgo0 = Bld.CreateIsNull(AlgoVerArgVal); 2086 2087 llvm::Value *Algo1 = Bld.CreateICmpEQ(AlgoVerArgVal, Bld.getInt16(1)); 2088 llvm::Value *CondAlgo1 = Bld.CreateAnd( 2089 Algo1, Bld.CreateICmpULT(LaneIDArgVal, RemoteLaneOffsetArgVal)); 2090 2091 llvm::Value *Algo2 = Bld.CreateICmpEQ(AlgoVerArgVal, Bld.getInt16(2)); 2092 llvm::Value *CondAlgo2 = Bld.CreateAnd( 2093 Algo2, Bld.CreateIsNull(Bld.CreateAnd(LaneIDArgVal, Bld.getInt16(1)))); 2094 CondAlgo2 = Bld.CreateAnd( 2095 CondAlgo2, Bld.CreateICmpSGT(RemoteLaneOffsetArgVal, Bld.getInt16(0))); 2096 2097 llvm::Value *CondReduce = Bld.CreateOr(CondAlgo0, CondAlgo1); 2098 CondReduce = Bld.CreateOr(CondReduce, CondAlgo2); 2099 2100 llvm::BasicBlock *ThenBB = CGF.createBasicBlock("then"); 2101 llvm::BasicBlock *ElseBB = CGF.createBasicBlock("else"); 2102 llvm::BasicBlock *MergeBB = CGF.createBasicBlock("ifcont"); 2103 Bld.CreateCondBr(CondReduce, ThenBB, ElseBB); 2104 2105 CGF.EmitBlock(ThenBB); 2106 // reduce_function(LocalReduceList, RemoteReduceList) 2107 llvm::Value *LocalReduceListPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( 2108 LocalReduceList.getPointer(), CGF.VoidPtrTy); 2109 llvm::Value *RemoteReduceListPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( 2110 RemoteReduceList.getPointer(), CGF.VoidPtrTy); 2111 CGM.getOpenMPRuntime().emitOutlinedFunctionCall( 2112 CGF, Loc, ReduceFn, {LocalReduceListPtr, RemoteReduceListPtr}); 2113 Bld.CreateBr(MergeBB); 2114 2115 CGF.EmitBlock(ElseBB); 2116 Bld.CreateBr(MergeBB); 2117 2118 CGF.EmitBlock(MergeBB); 2119 2120 // if (AlgoVer==1 && (LaneId >= Offset)) copy Remote Reduce list to local 2121 // Reduce list. 2122 Algo1 = Bld.CreateICmpEQ(AlgoVerArgVal, Bld.getInt16(1)); 2123 llvm::Value *CondCopy = Bld.CreateAnd( 2124 Algo1, Bld.CreateICmpUGE(LaneIDArgVal, RemoteLaneOffsetArgVal)); 2125 2126 llvm::BasicBlock *CpyThenBB = CGF.createBasicBlock("then"); 2127 llvm::BasicBlock *CpyElseBB = CGF.createBasicBlock("else"); 2128 llvm::BasicBlock *CpyMergeBB = CGF.createBasicBlock("ifcont"); 2129 Bld.CreateCondBr(CondCopy, CpyThenBB, CpyElseBB); 2130 2131 CGF.EmitBlock(CpyThenBB); 2132 emitReductionListCopy(ThreadCopy, CGF, ReductionArrayTy, Privates, 2133 RemoteReduceList, LocalReduceList); 2134 Bld.CreateBr(CpyMergeBB); 2135 2136 CGF.EmitBlock(CpyElseBB); 2137 Bld.CreateBr(CpyMergeBB); 2138 2139 CGF.EmitBlock(CpyMergeBB); 2140 2141 CGF.FinishFunction(); 2142 return Fn; 2143 } 2144 2145 /// This function emits a helper that copies all the reduction variables from 2146 /// the team into the provided global buffer for the reduction variables. 2147 /// 2148 /// void list_to_global_copy_func(void *buffer, int Idx, void *reduce_data) 2149 /// For all data entries D in reduce_data: 2150 /// Copy local D to buffer.D[Idx] 2151 static llvm::Value *emitListToGlobalCopyFunction( 2152 CodeGenModule &CGM, ArrayRef<const Expr *> Privates, 2153 QualType ReductionArrayTy, SourceLocation Loc, 2154 const RecordDecl *TeamReductionRec, 2155 const llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> 2156 &VarFieldMap) { 2157 ASTContext &C = CGM.getContext(); 2158 2159 // Buffer: global reduction buffer. 2160 ImplicitParamDecl BufferArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, 2161 C.VoidPtrTy, ImplicitParamKind::Other); 2162 // Idx: index of the buffer. 2163 ImplicitParamDecl IdxArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.IntTy, 2164 ImplicitParamKind::Other); 2165 // ReduceList: thread local Reduce list. 2166 ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, 2167 C.VoidPtrTy, ImplicitParamKind::Other); 2168 FunctionArgList Args; 2169 Args.push_back(&BufferArg); 2170 Args.push_back(&IdxArg); 2171 Args.push_back(&ReduceListArg); 2172 2173 const CGFunctionInfo &CGFI = 2174 CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); 2175 auto *Fn = llvm::Function::Create( 2176 CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage, 2177 "_omp_reduction_list_to_global_copy_func", &CGM.getModule()); 2178 CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI); 2179 Fn->setDoesNotRecurse(); 2180 CodeGenFunction CGF(CGM); 2181 CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc); 2182 2183 CGBuilderTy &Bld = CGF.Builder; 2184 2185 Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg); 2186 Address AddrBufferArg = CGF.GetAddrOfLocalVar(&BufferArg); 2187 llvm::Type *ElemTy = CGF.ConvertTypeForMem(ReductionArrayTy); 2188 Address LocalReduceList( 2189 Bld.CreatePointerBitCastOrAddrSpaceCast( 2190 CGF.EmitLoadOfScalar(AddrReduceListArg, /*Volatile=*/false, 2191 C.VoidPtrTy, Loc), 2192 ElemTy->getPointerTo()), 2193 ElemTy, CGF.getPointerAlign()); 2194 QualType StaticTy = C.getRecordType(TeamReductionRec); 2195 llvm::Type *LLVMReductionsBufferTy = 2196 CGM.getTypes().ConvertTypeForMem(StaticTy); 2197 llvm::Value *BufferArrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( 2198 CGF.EmitLoadOfScalar(AddrBufferArg, /*Volatile=*/false, C.VoidPtrTy, Loc), 2199 LLVMReductionsBufferTy->getPointerTo()); 2200 llvm::Value *Idxs[] = {CGF.EmitLoadOfScalar(CGF.GetAddrOfLocalVar(&IdxArg), 2201 /*Volatile=*/false, C.IntTy, 2202 Loc)}; 2203 unsigned Idx = 0; 2204 for (const Expr *Private : Privates) { 2205 // Reduce element = LocalReduceList[i] 2206 Address ElemPtrPtrAddr = Bld.CreateConstArrayGEP(LocalReduceList, Idx); 2207 llvm::Value *ElemPtrPtr = CGF.EmitLoadOfScalar( 2208 ElemPtrPtrAddr, /*Volatile=*/false, C.VoidPtrTy, SourceLocation()); 2209 // elemptr = ((CopyType*)(elemptrptr)) + I 2210 ElemTy = CGF.ConvertTypeForMem(Private->getType()); 2211 ElemPtrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( 2212 ElemPtrPtr, ElemTy->getPointerTo()); 2213 Address ElemPtr = 2214 Address(ElemPtrPtr, ElemTy, C.getTypeAlignInChars(Private->getType())); 2215 const ValueDecl *VD = cast<DeclRefExpr>(Private)->getDecl(); 2216 // Global = Buffer.VD[Idx]; 2217 const FieldDecl *FD = VarFieldMap.lookup(VD); 2218 llvm::Value *BufferPtr = 2219 Bld.CreateInBoundsGEP(LLVMReductionsBufferTy, BufferArrPtr, Idxs); 2220 LValue GlobLVal = CGF.EmitLValueForField( 2221 CGF.MakeNaturalAlignAddrLValue(BufferPtr, StaticTy), FD); 2222 Address GlobAddr = GlobLVal.getAddress(CGF); 2223 GlobLVal.setAddress(Address(GlobAddr.getPointer(), 2224 CGF.ConvertTypeForMem(Private->getType()), 2225 GlobAddr.getAlignment())); 2226 switch (CGF.getEvaluationKind(Private->getType())) { 2227 case TEK_Scalar: { 2228 llvm::Value *V = CGF.EmitLoadOfScalar( 2229 ElemPtr, /*Volatile=*/false, Private->getType(), Loc, 2230 LValueBaseInfo(AlignmentSource::Type), TBAAAccessInfo()); 2231 CGF.EmitStoreOfScalar(V, GlobLVal); 2232 break; 2233 } 2234 case TEK_Complex: { 2235 CodeGenFunction::ComplexPairTy V = CGF.EmitLoadOfComplex( 2236 CGF.MakeAddrLValue(ElemPtr, Private->getType()), Loc); 2237 CGF.EmitStoreOfComplex(V, GlobLVal, /*isInit=*/false); 2238 break; 2239 } 2240 case TEK_Aggregate: 2241 CGF.EmitAggregateCopy(GlobLVal, 2242 CGF.MakeAddrLValue(ElemPtr, Private->getType()), 2243 Private->getType(), AggValueSlot::DoesNotOverlap); 2244 break; 2245 } 2246 ++Idx; 2247 } 2248 2249 CGF.FinishFunction(); 2250 return Fn; 2251 } 2252 2253 /// This function emits a helper that reduces all the reduction variables from 2254 /// the team into the provided global buffer for the reduction variables. 2255 /// 2256 /// void list_to_global_reduce_func(void *buffer, int Idx, void *reduce_data) 2257 /// void *GlobPtrs[]; 2258 /// GlobPtrs[0] = (void*)&buffer.D0[Idx]; 2259 /// ... 2260 /// GlobPtrs[N] = (void*)&buffer.DN[Idx]; 2261 /// reduce_function(GlobPtrs, reduce_data); 2262 static llvm::Value *emitListToGlobalReduceFunction( 2263 CodeGenModule &CGM, ArrayRef<const Expr *> Privates, 2264 QualType ReductionArrayTy, SourceLocation Loc, 2265 const RecordDecl *TeamReductionRec, 2266 const llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> 2267 &VarFieldMap, 2268 llvm::Function *ReduceFn) { 2269 ASTContext &C = CGM.getContext(); 2270 2271 // Buffer: global reduction buffer. 2272 ImplicitParamDecl BufferArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, 2273 C.VoidPtrTy, ImplicitParamKind::Other); 2274 // Idx: index of the buffer. 2275 ImplicitParamDecl IdxArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.IntTy, 2276 ImplicitParamKind::Other); 2277 // ReduceList: thread local Reduce list. 2278 ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, 2279 C.VoidPtrTy, ImplicitParamKind::Other); 2280 FunctionArgList Args; 2281 Args.push_back(&BufferArg); 2282 Args.push_back(&IdxArg); 2283 Args.push_back(&ReduceListArg); 2284 2285 const CGFunctionInfo &CGFI = 2286 CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); 2287 auto *Fn = llvm::Function::Create( 2288 CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage, 2289 "_omp_reduction_list_to_global_reduce_func", &CGM.getModule()); 2290 CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI); 2291 Fn->setDoesNotRecurse(); 2292 CodeGenFunction CGF(CGM); 2293 CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc); 2294 2295 CGBuilderTy &Bld = CGF.Builder; 2296 2297 Address AddrBufferArg = CGF.GetAddrOfLocalVar(&BufferArg); 2298 QualType StaticTy = C.getRecordType(TeamReductionRec); 2299 llvm::Type *LLVMReductionsBufferTy = 2300 CGM.getTypes().ConvertTypeForMem(StaticTy); 2301 llvm::Value *BufferArrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( 2302 CGF.EmitLoadOfScalar(AddrBufferArg, /*Volatile=*/false, C.VoidPtrTy, Loc), 2303 LLVMReductionsBufferTy->getPointerTo()); 2304 2305 // 1. Build a list of reduction variables. 2306 // void *RedList[<n>] = {<ReductionVars>[0], ..., <ReductionVars>[<n>-1]}; 2307 Address ReductionList = 2308 CGF.CreateMemTemp(ReductionArrayTy, ".omp.reduction.red_list"); 2309 auto IPriv = Privates.begin(); 2310 llvm::Value *Idxs[] = {CGF.EmitLoadOfScalar(CGF.GetAddrOfLocalVar(&IdxArg), 2311 /*Volatile=*/false, C.IntTy, 2312 Loc)}; 2313 unsigned Idx = 0; 2314 for (unsigned I = 0, E = Privates.size(); I < E; ++I, ++IPriv, ++Idx) { 2315 Address Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx); 2316 // Global = Buffer.VD[Idx]; 2317 const ValueDecl *VD = cast<DeclRefExpr>(*IPriv)->getDecl(); 2318 const FieldDecl *FD = VarFieldMap.lookup(VD); 2319 llvm::Value *BufferPtr = 2320 Bld.CreateInBoundsGEP(LLVMReductionsBufferTy, BufferArrPtr, Idxs); 2321 LValue GlobLVal = CGF.EmitLValueForField( 2322 CGF.MakeNaturalAlignAddrLValue(BufferPtr, StaticTy), FD); 2323 Address GlobAddr = GlobLVal.getAddress(CGF); 2324 CGF.EmitStoreOfScalar(GlobAddr.getPointer(), Elem, /*Volatile=*/false, 2325 C.VoidPtrTy); 2326 if ((*IPriv)->getType()->isVariablyModifiedType()) { 2327 // Store array size. 2328 ++Idx; 2329 Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx); 2330 llvm::Value *Size = CGF.Builder.CreateIntCast( 2331 CGF.getVLASize( 2332 CGF.getContext().getAsVariableArrayType((*IPriv)->getType())) 2333 .NumElts, 2334 CGF.SizeTy, /*isSigned=*/false); 2335 CGF.Builder.CreateStore(CGF.Builder.CreateIntToPtr(Size, CGF.VoidPtrTy), 2336 Elem); 2337 } 2338 } 2339 2340 // Call reduce_function(GlobalReduceList, ReduceList) 2341 llvm::Value *GlobalReduceList = ReductionList.getPointer(); 2342 Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg); 2343 llvm::Value *ReducedPtr = CGF.EmitLoadOfScalar( 2344 AddrReduceListArg, /*Volatile=*/false, C.VoidPtrTy, Loc); 2345 CGM.getOpenMPRuntime().emitOutlinedFunctionCall( 2346 CGF, Loc, ReduceFn, {GlobalReduceList, ReducedPtr}); 2347 CGF.FinishFunction(); 2348 return Fn; 2349 } 2350 2351 /// This function emits a helper that copies all the reduction variables from 2352 /// the team into the provided global buffer for the reduction variables. 2353 /// 2354 /// void list_to_global_copy_func(void *buffer, int Idx, void *reduce_data) 2355 /// For all data entries D in reduce_data: 2356 /// Copy buffer.D[Idx] to local D; 2357 static llvm::Value *emitGlobalToListCopyFunction( 2358 CodeGenModule &CGM, ArrayRef<const Expr *> Privates, 2359 QualType ReductionArrayTy, SourceLocation Loc, 2360 const RecordDecl *TeamReductionRec, 2361 const llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> 2362 &VarFieldMap) { 2363 ASTContext &C = CGM.getContext(); 2364 2365 // Buffer: global reduction buffer. 2366 ImplicitParamDecl BufferArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, 2367 C.VoidPtrTy, ImplicitParamKind::Other); 2368 // Idx: index of the buffer. 2369 ImplicitParamDecl IdxArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.IntTy, 2370 ImplicitParamKind::Other); 2371 // ReduceList: thread local Reduce list. 2372 ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, 2373 C.VoidPtrTy, ImplicitParamKind::Other); 2374 FunctionArgList Args; 2375 Args.push_back(&BufferArg); 2376 Args.push_back(&IdxArg); 2377 Args.push_back(&ReduceListArg); 2378 2379 const CGFunctionInfo &CGFI = 2380 CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); 2381 auto *Fn = llvm::Function::Create( 2382 CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage, 2383 "_omp_reduction_global_to_list_copy_func", &CGM.getModule()); 2384 CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI); 2385 Fn->setDoesNotRecurse(); 2386 CodeGenFunction CGF(CGM); 2387 CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc); 2388 2389 CGBuilderTy &Bld = CGF.Builder; 2390 2391 Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg); 2392 Address AddrBufferArg = CGF.GetAddrOfLocalVar(&BufferArg); 2393 llvm::Type *ElemTy = CGF.ConvertTypeForMem(ReductionArrayTy); 2394 Address LocalReduceList( 2395 Bld.CreatePointerBitCastOrAddrSpaceCast( 2396 CGF.EmitLoadOfScalar(AddrReduceListArg, /*Volatile=*/false, 2397 C.VoidPtrTy, Loc), 2398 ElemTy->getPointerTo()), 2399 ElemTy, CGF.getPointerAlign()); 2400 QualType StaticTy = C.getRecordType(TeamReductionRec); 2401 llvm::Type *LLVMReductionsBufferTy = 2402 CGM.getTypes().ConvertTypeForMem(StaticTy); 2403 llvm::Value *BufferArrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( 2404 CGF.EmitLoadOfScalar(AddrBufferArg, /*Volatile=*/false, C.VoidPtrTy, Loc), 2405 LLVMReductionsBufferTy->getPointerTo()); 2406 2407 llvm::Value *Idxs[] = {CGF.EmitLoadOfScalar(CGF.GetAddrOfLocalVar(&IdxArg), 2408 /*Volatile=*/false, C.IntTy, 2409 Loc)}; 2410 unsigned Idx = 0; 2411 for (const Expr *Private : Privates) { 2412 // Reduce element = LocalReduceList[i] 2413 Address ElemPtrPtrAddr = Bld.CreateConstArrayGEP(LocalReduceList, Idx); 2414 llvm::Value *ElemPtrPtr = CGF.EmitLoadOfScalar( 2415 ElemPtrPtrAddr, /*Volatile=*/false, C.VoidPtrTy, SourceLocation()); 2416 // elemptr = ((CopyType*)(elemptrptr)) + I 2417 ElemTy = CGF.ConvertTypeForMem(Private->getType()); 2418 ElemPtrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( 2419 ElemPtrPtr, ElemTy->getPointerTo()); 2420 Address ElemPtr = 2421 Address(ElemPtrPtr, ElemTy, C.getTypeAlignInChars(Private->getType())); 2422 const ValueDecl *VD = cast<DeclRefExpr>(Private)->getDecl(); 2423 // Global = Buffer.VD[Idx]; 2424 const FieldDecl *FD = VarFieldMap.lookup(VD); 2425 llvm::Value *BufferPtr = 2426 Bld.CreateInBoundsGEP(LLVMReductionsBufferTy, BufferArrPtr, Idxs); 2427 LValue GlobLVal = CGF.EmitLValueForField( 2428 CGF.MakeNaturalAlignAddrLValue(BufferPtr, StaticTy), FD); 2429 Address GlobAddr = GlobLVal.getAddress(CGF); 2430 GlobLVal.setAddress(Address(GlobAddr.getPointer(), 2431 CGF.ConvertTypeForMem(Private->getType()), 2432 GlobAddr.getAlignment())); 2433 switch (CGF.getEvaluationKind(Private->getType())) { 2434 case TEK_Scalar: { 2435 llvm::Value *V = CGF.EmitLoadOfScalar(GlobLVal, Loc); 2436 CGF.EmitStoreOfScalar(V, ElemPtr, /*Volatile=*/false, Private->getType(), 2437 LValueBaseInfo(AlignmentSource::Type), 2438 TBAAAccessInfo()); 2439 break; 2440 } 2441 case TEK_Complex: { 2442 CodeGenFunction::ComplexPairTy V = CGF.EmitLoadOfComplex(GlobLVal, Loc); 2443 CGF.EmitStoreOfComplex(V, CGF.MakeAddrLValue(ElemPtr, Private->getType()), 2444 /*isInit=*/false); 2445 break; 2446 } 2447 case TEK_Aggregate: 2448 CGF.EmitAggregateCopy(CGF.MakeAddrLValue(ElemPtr, Private->getType()), 2449 GlobLVal, Private->getType(), 2450 AggValueSlot::DoesNotOverlap); 2451 break; 2452 } 2453 ++Idx; 2454 } 2455 2456 CGF.FinishFunction(); 2457 return Fn; 2458 } 2459 2460 /// This function emits a helper that reduces all the reduction variables from 2461 /// the team into the provided global buffer for the reduction variables. 2462 /// 2463 /// void global_to_list_reduce_func(void *buffer, int Idx, void *reduce_data) 2464 /// void *GlobPtrs[]; 2465 /// GlobPtrs[0] = (void*)&buffer.D0[Idx]; 2466 /// ... 2467 /// GlobPtrs[N] = (void*)&buffer.DN[Idx]; 2468 /// reduce_function(reduce_data, GlobPtrs); 2469 static llvm::Value *emitGlobalToListReduceFunction( 2470 CodeGenModule &CGM, ArrayRef<const Expr *> Privates, 2471 QualType ReductionArrayTy, SourceLocation Loc, 2472 const RecordDecl *TeamReductionRec, 2473 const llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> 2474 &VarFieldMap, 2475 llvm::Function *ReduceFn) { 2476 ASTContext &C = CGM.getContext(); 2477 2478 // Buffer: global reduction buffer. 2479 ImplicitParamDecl BufferArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, 2480 C.VoidPtrTy, ImplicitParamKind::Other); 2481 // Idx: index of the buffer. 2482 ImplicitParamDecl IdxArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.IntTy, 2483 ImplicitParamKind::Other); 2484 // ReduceList: thread local Reduce list. 2485 ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, 2486 C.VoidPtrTy, ImplicitParamKind::Other); 2487 FunctionArgList Args; 2488 Args.push_back(&BufferArg); 2489 Args.push_back(&IdxArg); 2490 Args.push_back(&ReduceListArg); 2491 2492 const CGFunctionInfo &CGFI = 2493 CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); 2494 auto *Fn = llvm::Function::Create( 2495 CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage, 2496 "_omp_reduction_global_to_list_reduce_func", &CGM.getModule()); 2497 CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI); 2498 Fn->setDoesNotRecurse(); 2499 CodeGenFunction CGF(CGM); 2500 CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc); 2501 2502 CGBuilderTy &Bld = CGF.Builder; 2503 2504 Address AddrBufferArg = CGF.GetAddrOfLocalVar(&BufferArg); 2505 QualType StaticTy = C.getRecordType(TeamReductionRec); 2506 llvm::Type *LLVMReductionsBufferTy = 2507 CGM.getTypes().ConvertTypeForMem(StaticTy); 2508 llvm::Value *BufferArrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( 2509 CGF.EmitLoadOfScalar(AddrBufferArg, /*Volatile=*/false, C.VoidPtrTy, Loc), 2510 LLVMReductionsBufferTy->getPointerTo()); 2511 2512 // 1. Build a list of reduction variables. 2513 // void *RedList[<n>] = {<ReductionVars>[0], ..., <ReductionVars>[<n>-1]}; 2514 Address ReductionList = 2515 CGF.CreateMemTemp(ReductionArrayTy, ".omp.reduction.red_list"); 2516 auto IPriv = Privates.begin(); 2517 llvm::Value *Idxs[] = {CGF.EmitLoadOfScalar(CGF.GetAddrOfLocalVar(&IdxArg), 2518 /*Volatile=*/false, C.IntTy, 2519 Loc)}; 2520 unsigned Idx = 0; 2521 for (unsigned I = 0, E = Privates.size(); I < E; ++I, ++IPriv, ++Idx) { 2522 Address Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx); 2523 // Global = Buffer.VD[Idx]; 2524 const ValueDecl *VD = cast<DeclRefExpr>(*IPriv)->getDecl(); 2525 const FieldDecl *FD = VarFieldMap.lookup(VD); 2526 llvm::Value *BufferPtr = 2527 Bld.CreateInBoundsGEP(LLVMReductionsBufferTy, BufferArrPtr, Idxs); 2528 LValue GlobLVal = CGF.EmitLValueForField( 2529 CGF.MakeNaturalAlignAddrLValue(BufferPtr, StaticTy), FD); 2530 Address GlobAddr = GlobLVal.getAddress(CGF); 2531 CGF.EmitStoreOfScalar(GlobAddr.getPointer(), Elem, /*Volatile=*/false, 2532 C.VoidPtrTy); 2533 if ((*IPriv)->getType()->isVariablyModifiedType()) { 2534 // Store array size. 2535 ++Idx; 2536 Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx); 2537 llvm::Value *Size = CGF.Builder.CreateIntCast( 2538 CGF.getVLASize( 2539 CGF.getContext().getAsVariableArrayType((*IPriv)->getType())) 2540 .NumElts, 2541 CGF.SizeTy, /*isSigned=*/false); 2542 CGF.Builder.CreateStore(CGF.Builder.CreateIntToPtr(Size, CGF.VoidPtrTy), 2543 Elem); 2544 } 2545 } 2546 2547 // Call reduce_function(ReduceList, GlobalReduceList) 2548 llvm::Value *GlobalReduceList = ReductionList.getPointer(); 2549 Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg); 2550 llvm::Value *ReducedPtr = CGF.EmitLoadOfScalar( 2551 AddrReduceListArg, /*Volatile=*/false, C.VoidPtrTy, Loc); 2552 CGM.getOpenMPRuntime().emitOutlinedFunctionCall( 2553 CGF, Loc, ReduceFn, {ReducedPtr, GlobalReduceList}); 2554 CGF.FinishFunction(); 2555 return Fn; 2556 } 2557 2558 /// 2559 /// Design of OpenMP reductions on the GPU 2560 /// 2561 /// Consider a typical OpenMP program with one or more reduction 2562 /// clauses: 2563 /// 2564 /// float foo; 2565 /// double bar; 2566 /// #pragma omp target teams distribute parallel for \ 2567 /// reduction(+:foo) reduction(*:bar) 2568 /// for (int i = 0; i < N; i++) { 2569 /// foo += A[i]; bar *= B[i]; 2570 /// } 2571 /// 2572 /// where 'foo' and 'bar' are reduced across all OpenMP threads in 2573 /// all teams. In our OpenMP implementation on the NVPTX device an 2574 /// OpenMP team is mapped to a CUDA threadblock and OpenMP threads 2575 /// within a team are mapped to CUDA threads within a threadblock. 2576 /// Our goal is to efficiently aggregate values across all OpenMP 2577 /// threads such that: 2578 /// 2579 /// - the compiler and runtime are logically concise, and 2580 /// - the reduction is performed efficiently in a hierarchical 2581 /// manner as follows: within OpenMP threads in the same warp, 2582 /// across warps in a threadblock, and finally across teams on 2583 /// the NVPTX device. 2584 /// 2585 /// Introduction to Decoupling 2586 /// 2587 /// We would like to decouple the compiler and the runtime so that the 2588 /// latter is ignorant of the reduction variables (number, data types) 2589 /// and the reduction operators. This allows a simpler interface 2590 /// and implementation while still attaining good performance. 2591 /// 2592 /// Pseudocode for the aforementioned OpenMP program generated by the 2593 /// compiler is as follows: 2594 /// 2595 /// 1. Create private copies of reduction variables on each OpenMP 2596 /// thread: 'foo_private', 'bar_private' 2597 /// 2. Each OpenMP thread reduces the chunk of 'A' and 'B' assigned 2598 /// to it and writes the result in 'foo_private' and 'bar_private' 2599 /// respectively. 2600 /// 3. Call the OpenMP runtime on the GPU to reduce within a team 2601 /// and store the result on the team master: 2602 /// 2603 /// __kmpc_nvptx_parallel_reduce_nowait_v2(..., 2604 /// reduceData, shuffleReduceFn, interWarpCpyFn) 2605 /// 2606 /// where: 2607 /// struct ReduceData { 2608 /// double *foo; 2609 /// double *bar; 2610 /// } reduceData 2611 /// reduceData.foo = &foo_private 2612 /// reduceData.bar = &bar_private 2613 /// 2614 /// 'shuffleReduceFn' and 'interWarpCpyFn' are pointers to two 2615 /// auxiliary functions generated by the compiler that operate on 2616 /// variables of type 'ReduceData'. They aid the runtime perform 2617 /// algorithmic steps in a data agnostic manner. 2618 /// 2619 /// 'shuffleReduceFn' is a pointer to a function that reduces data 2620 /// of type 'ReduceData' across two OpenMP threads (lanes) in the 2621 /// same warp. It takes the following arguments as input: 2622 /// 2623 /// a. variable of type 'ReduceData' on the calling lane, 2624 /// b. its lane_id, 2625 /// c. an offset relative to the current lane_id to generate a 2626 /// remote_lane_id. The remote lane contains the second 2627 /// variable of type 'ReduceData' that is to be reduced. 2628 /// d. an algorithm version parameter determining which reduction 2629 /// algorithm to use. 2630 /// 2631 /// 'shuffleReduceFn' retrieves data from the remote lane using 2632 /// efficient GPU shuffle intrinsics and reduces, using the 2633 /// algorithm specified by the 4th parameter, the two operands 2634 /// element-wise. The result is written to the first operand. 2635 /// 2636 /// Different reduction algorithms are implemented in different 2637 /// runtime functions, all calling 'shuffleReduceFn' to perform 2638 /// the essential reduction step. Therefore, based on the 4th 2639 /// parameter, this function behaves slightly differently to 2640 /// cooperate with the runtime to ensure correctness under 2641 /// different circumstances. 2642 /// 2643 /// 'InterWarpCpyFn' is a pointer to a function that transfers 2644 /// reduced variables across warps. It tunnels, through CUDA 2645 /// shared memory, the thread-private data of type 'ReduceData' 2646 /// from lane 0 of each warp to a lane in the first warp. 2647 /// 4. Call the OpenMP runtime on the GPU to reduce across teams. 2648 /// The last team writes the global reduced value to memory. 2649 /// 2650 /// ret = __kmpc_nvptx_teams_reduce_nowait(..., 2651 /// reduceData, shuffleReduceFn, interWarpCpyFn, 2652 /// scratchpadCopyFn, loadAndReduceFn) 2653 /// 2654 /// 'scratchpadCopyFn' is a helper that stores reduced 2655 /// data from the team master to a scratchpad array in 2656 /// global memory. 2657 /// 2658 /// 'loadAndReduceFn' is a helper that loads data from 2659 /// the scratchpad array and reduces it with the input 2660 /// operand. 2661 /// 2662 /// These compiler generated functions hide address 2663 /// calculation and alignment information from the runtime. 2664 /// 5. if ret == 1: 2665 /// The team master of the last team stores the reduced 2666 /// result to the globals in memory. 2667 /// foo += reduceData.foo; bar *= reduceData.bar 2668 /// 2669 /// 2670 /// Warp Reduction Algorithms 2671 /// 2672 /// On the warp level, we have three algorithms implemented in the 2673 /// OpenMP runtime depending on the number of active lanes: 2674 /// 2675 /// Full Warp Reduction 2676 /// 2677 /// The reduce algorithm within a warp where all lanes are active 2678 /// is implemented in the runtime as follows: 2679 /// 2680 /// full_warp_reduce(void *reduce_data, 2681 /// kmp_ShuffleReductFctPtr ShuffleReduceFn) { 2682 /// for (int offset = WARPSIZE/2; offset > 0; offset /= 2) 2683 /// ShuffleReduceFn(reduce_data, 0, offset, 0); 2684 /// } 2685 /// 2686 /// The algorithm completes in log(2, WARPSIZE) steps. 2687 /// 2688 /// 'ShuffleReduceFn' is used here with lane_id set to 0 because it is 2689 /// not used therefore we save instructions by not retrieving lane_id 2690 /// from the corresponding special registers. The 4th parameter, which 2691 /// represents the version of the algorithm being used, is set to 0 to 2692 /// signify full warp reduction. 2693 /// 2694 /// In this version, 'ShuffleReduceFn' behaves, per element, as follows: 2695 /// 2696 /// #reduce_elem refers to an element in the local lane's data structure 2697 /// #remote_elem is retrieved from a remote lane 2698 /// remote_elem = shuffle_down(reduce_elem, offset, WARPSIZE); 2699 /// reduce_elem = reduce_elem REDUCE_OP remote_elem; 2700 /// 2701 /// Contiguous Partial Warp Reduction 2702 /// 2703 /// This reduce algorithm is used within a warp where only the first 2704 /// 'n' (n <= WARPSIZE) lanes are active. It is typically used when the 2705 /// number of OpenMP threads in a parallel region is not a multiple of 2706 /// WARPSIZE. The algorithm is implemented in the runtime as follows: 2707 /// 2708 /// void 2709 /// contiguous_partial_reduce(void *reduce_data, 2710 /// kmp_ShuffleReductFctPtr ShuffleReduceFn, 2711 /// int size, int lane_id) { 2712 /// int curr_size; 2713 /// int offset; 2714 /// curr_size = size; 2715 /// mask = curr_size/2; 2716 /// while (offset>0) { 2717 /// ShuffleReduceFn(reduce_data, lane_id, offset, 1); 2718 /// curr_size = (curr_size+1)/2; 2719 /// offset = curr_size/2; 2720 /// } 2721 /// } 2722 /// 2723 /// In this version, 'ShuffleReduceFn' behaves, per element, as follows: 2724 /// 2725 /// remote_elem = shuffle_down(reduce_elem, offset, WARPSIZE); 2726 /// if (lane_id < offset) 2727 /// reduce_elem = reduce_elem REDUCE_OP remote_elem 2728 /// else 2729 /// reduce_elem = remote_elem 2730 /// 2731 /// This algorithm assumes that the data to be reduced are located in a 2732 /// contiguous subset of lanes starting from the first. When there is 2733 /// an odd number of active lanes, the data in the last lane is not 2734 /// aggregated with any other lane's dat but is instead copied over. 2735 /// 2736 /// Dispersed Partial Warp Reduction 2737 /// 2738 /// This algorithm is used within a warp when any discontiguous subset of 2739 /// lanes are active. It is used to implement the reduction operation 2740 /// across lanes in an OpenMP simd region or in a nested parallel region. 2741 /// 2742 /// void 2743 /// dispersed_partial_reduce(void *reduce_data, 2744 /// kmp_ShuffleReductFctPtr ShuffleReduceFn) { 2745 /// int size, remote_id; 2746 /// int logical_lane_id = number_of_active_lanes_before_me() * 2; 2747 /// do { 2748 /// remote_id = next_active_lane_id_right_after_me(); 2749 /// # the above function returns 0 of no active lane 2750 /// # is present right after the current lane. 2751 /// size = number_of_active_lanes_in_this_warp(); 2752 /// logical_lane_id /= 2; 2753 /// ShuffleReduceFn(reduce_data, logical_lane_id, 2754 /// remote_id-1-threadIdx.x, 2); 2755 /// } while (logical_lane_id % 2 == 0 && size > 1); 2756 /// } 2757 /// 2758 /// There is no assumption made about the initial state of the reduction. 2759 /// Any number of lanes (>=1) could be active at any position. The reduction 2760 /// result is returned in the first active lane. 2761 /// 2762 /// In this version, 'ShuffleReduceFn' behaves, per element, as follows: 2763 /// 2764 /// remote_elem = shuffle_down(reduce_elem, offset, WARPSIZE); 2765 /// if (lane_id % 2 == 0 && offset > 0) 2766 /// reduce_elem = reduce_elem REDUCE_OP remote_elem 2767 /// else 2768 /// reduce_elem = remote_elem 2769 /// 2770 /// 2771 /// Intra-Team Reduction 2772 /// 2773 /// This function, as implemented in the runtime call 2774 /// '__kmpc_nvptx_parallel_reduce_nowait_v2', aggregates data across OpenMP 2775 /// threads in a team. It first reduces within a warp using the 2776 /// aforementioned algorithms. We then proceed to gather all such 2777 /// reduced values at the first warp. 2778 /// 2779 /// The runtime makes use of the function 'InterWarpCpyFn', which copies 2780 /// data from each of the "warp master" (zeroth lane of each warp, where 2781 /// warp-reduced data is held) to the zeroth warp. This step reduces (in 2782 /// a mathematical sense) the problem of reduction across warp masters in 2783 /// a block to the problem of warp reduction. 2784 /// 2785 /// 2786 /// Inter-Team Reduction 2787 /// 2788 /// Once a team has reduced its data to a single value, it is stored in 2789 /// a global scratchpad array. Since each team has a distinct slot, this 2790 /// can be done without locking. 2791 /// 2792 /// The last team to write to the scratchpad array proceeds to reduce the 2793 /// scratchpad array. One or more workers in the last team use the helper 2794 /// 'loadAndReduceDataFn' to load and reduce values from the array, i.e., 2795 /// the k'th worker reduces every k'th element. 2796 /// 2797 /// Finally, a call is made to '__kmpc_nvptx_parallel_reduce_nowait_v2' to 2798 /// reduce across workers and compute a globally reduced value. 2799 /// 2800 void CGOpenMPRuntimeGPU::emitReduction( 2801 CodeGenFunction &CGF, SourceLocation Loc, ArrayRef<const Expr *> Privates, 2802 ArrayRef<const Expr *> LHSExprs, ArrayRef<const Expr *> RHSExprs, 2803 ArrayRef<const Expr *> ReductionOps, ReductionOptionsTy Options) { 2804 if (!CGF.HaveInsertPoint()) 2805 return; 2806 2807 bool ParallelReduction = isOpenMPParallelDirective(Options.ReductionKind); 2808 #ifndef NDEBUG 2809 bool TeamsReduction = isOpenMPTeamsDirective(Options.ReductionKind); 2810 #endif 2811 2812 if (Options.SimpleReduction) { 2813 assert(!TeamsReduction && !ParallelReduction && 2814 "Invalid reduction selection in emitReduction."); 2815 CGOpenMPRuntime::emitReduction(CGF, Loc, Privates, LHSExprs, RHSExprs, 2816 ReductionOps, Options); 2817 return; 2818 } 2819 2820 assert((TeamsReduction || ParallelReduction) && 2821 "Invalid reduction selection in emitReduction."); 2822 2823 llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> VarFieldMap; 2824 llvm::SmallVector<const ValueDecl *, 4> PrivatesReductions(Privates.size()); 2825 int Cnt = 0; 2826 for (const Expr *DRE : Privates) { 2827 PrivatesReductions[Cnt] = cast<DeclRefExpr>(DRE)->getDecl(); 2828 ++Cnt; 2829 } 2830 2831 ASTContext &C = CGM.getContext(); 2832 const RecordDecl *ReductionRec = ::buildRecordForGlobalizedVars( 2833 CGM.getContext(), PrivatesReductions, std::nullopt, VarFieldMap, 1); 2834 2835 // Build res = __kmpc_reduce{_nowait}(<gtid>, <n>, sizeof(RedList), 2836 // RedList, shuffle_reduce_func, interwarp_copy_func); 2837 // or 2838 // Build res = __kmpc_reduce_teams_nowait_simple(<loc>, <gtid>, <lck>); 2839 llvm::Value *RTLoc = emitUpdateLocation(CGF, Loc); 2840 2841 llvm::Value *Res; 2842 // 1. Build a list of reduction variables. 2843 // void *RedList[<n>] = {<ReductionVars>[0], ..., <ReductionVars>[<n>-1]}; 2844 auto Size = RHSExprs.size(); 2845 for (const Expr *E : Privates) { 2846 if (E->getType()->isVariablyModifiedType()) 2847 // Reserve place for array size. 2848 ++Size; 2849 } 2850 llvm::APInt ArraySize(/*unsigned int numBits=*/32, Size); 2851 QualType ReductionArrayTy = C.getConstantArrayType( 2852 C.VoidPtrTy, ArraySize, nullptr, ArraySizeModifier::Normal, 2853 /*IndexTypeQuals=*/0); 2854 Address ReductionList = 2855 CGF.CreateMemTemp(ReductionArrayTy, ".omp.reduction.red_list"); 2856 auto IPriv = Privates.begin(); 2857 unsigned Idx = 0; 2858 for (unsigned I = 0, E = RHSExprs.size(); I < E; ++I, ++IPriv, ++Idx) { 2859 Address Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx); 2860 CGF.Builder.CreateStore( 2861 CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( 2862 CGF.EmitLValue(RHSExprs[I]).getPointer(CGF), CGF.VoidPtrTy), 2863 Elem); 2864 if ((*IPriv)->getType()->isVariablyModifiedType()) { 2865 // Store array size. 2866 ++Idx; 2867 Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx); 2868 llvm::Value *Size = CGF.Builder.CreateIntCast( 2869 CGF.getVLASize( 2870 CGF.getContext().getAsVariableArrayType((*IPriv)->getType())) 2871 .NumElts, 2872 CGF.SizeTy, /*isSigned=*/false); 2873 CGF.Builder.CreateStore(CGF.Builder.CreateIntToPtr(Size, CGF.VoidPtrTy), 2874 Elem); 2875 } 2876 } 2877 2878 llvm::Value *RL = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( 2879 ReductionList.getPointer(), CGF.VoidPtrTy); 2880 llvm::Function *ReductionFn = emitReductionFunction( 2881 CGF.CurFn->getName(), Loc, CGF.ConvertTypeForMem(ReductionArrayTy), 2882 Privates, LHSExprs, RHSExprs, ReductionOps); 2883 llvm::Value *ReductionDataSize = 2884 CGF.getTypeSize(C.getRecordType(ReductionRec)); 2885 ReductionDataSize = 2886 CGF.Builder.CreateSExtOrTrunc(ReductionDataSize, CGF.Int64Ty); 2887 llvm::Function *ShuffleAndReduceFn = emitShuffleAndReduceFunction( 2888 CGM, Privates, ReductionArrayTy, ReductionFn, Loc); 2889 llvm::Value *InterWarpCopyFn = 2890 emitInterWarpCopyFunction(CGM, Privates, ReductionArrayTy, Loc); 2891 2892 if (ParallelReduction) { 2893 llvm::Value *Args[] = {RTLoc, ReductionDataSize, RL, ShuffleAndReduceFn, 2894 InterWarpCopyFn}; 2895 2896 Res = CGF.EmitRuntimeCall( 2897 OMPBuilder.getOrCreateRuntimeFunction( 2898 CGM.getModule(), OMPRTL___kmpc_nvptx_parallel_reduce_nowait_v2), 2899 Args); 2900 } else { 2901 assert(TeamsReduction && "expected teams reduction."); 2902 TeamsReductions.push_back(ReductionRec); 2903 auto *KernelTeamsReductionPtr = CGF.EmitRuntimeCall( 2904 OMPBuilder.getOrCreateRuntimeFunction( 2905 CGM.getModule(), OMPRTL___kmpc_reduction_get_fixed_buffer), 2906 {}, "_openmp_teams_reductions_buffer_$_$ptr"); 2907 llvm::Value *GlobalToBufferCpyFn = ::emitListToGlobalCopyFunction( 2908 CGM, Privates, ReductionArrayTy, Loc, ReductionRec, VarFieldMap); 2909 llvm::Value *GlobalToBufferRedFn = ::emitListToGlobalReduceFunction( 2910 CGM, Privates, ReductionArrayTy, Loc, ReductionRec, VarFieldMap, 2911 ReductionFn); 2912 llvm::Value *BufferToGlobalCpyFn = ::emitGlobalToListCopyFunction( 2913 CGM, Privates, ReductionArrayTy, Loc, ReductionRec, VarFieldMap); 2914 llvm::Value *BufferToGlobalRedFn = ::emitGlobalToListReduceFunction( 2915 CGM, Privates, ReductionArrayTy, Loc, ReductionRec, VarFieldMap, 2916 ReductionFn); 2917 2918 llvm::Value *Args[] = { 2919 RTLoc, 2920 KernelTeamsReductionPtr, 2921 CGF.Builder.getInt32(C.getLangOpts().OpenMPCUDAReductionBufNum), 2922 ReductionDataSize, 2923 RL, 2924 ShuffleAndReduceFn, 2925 InterWarpCopyFn, 2926 GlobalToBufferCpyFn, 2927 GlobalToBufferRedFn, 2928 BufferToGlobalCpyFn, 2929 BufferToGlobalRedFn}; 2930 2931 Res = CGF.EmitRuntimeCall( 2932 OMPBuilder.getOrCreateRuntimeFunction( 2933 CGM.getModule(), OMPRTL___kmpc_nvptx_teams_reduce_nowait_v2), 2934 Args); 2935 } 2936 2937 // 5. Build if (res == 1) 2938 llvm::BasicBlock *ExitBB = CGF.createBasicBlock(".omp.reduction.done"); 2939 llvm::BasicBlock *ThenBB = CGF.createBasicBlock(".omp.reduction.then"); 2940 llvm::Value *Cond = CGF.Builder.CreateICmpEQ( 2941 Res, llvm::ConstantInt::get(CGM.Int32Ty, /*V=*/1)); 2942 CGF.Builder.CreateCondBr(Cond, ThenBB, ExitBB); 2943 2944 // 6. Build then branch: where we have reduced values in the master 2945 // thread in each team. 2946 // __kmpc_end_reduce{_nowait}(<gtid>); 2947 // break; 2948 CGF.EmitBlock(ThenBB); 2949 2950 // Add emission of __kmpc_end_reduce{_nowait}(<gtid>); 2951 auto &&CodeGen = [Privates, LHSExprs, RHSExprs, ReductionOps, 2952 this](CodeGenFunction &CGF, PrePostActionTy &Action) { 2953 auto IPriv = Privates.begin(); 2954 auto ILHS = LHSExprs.begin(); 2955 auto IRHS = RHSExprs.begin(); 2956 for (const Expr *E : ReductionOps) { 2957 emitSingleReductionCombiner(CGF, E, *IPriv, cast<DeclRefExpr>(*ILHS), 2958 cast<DeclRefExpr>(*IRHS)); 2959 ++IPriv; 2960 ++ILHS; 2961 ++IRHS; 2962 } 2963 }; 2964 RegionCodeGenTy RCG(CodeGen); 2965 RCG(CGF); 2966 // There is no need to emit line number for unconditional branch. 2967 (void)ApplyDebugLocation::CreateEmpty(CGF); 2968 CGF.EmitBlock(ExitBB, /*IsFinished=*/true); 2969 } 2970 2971 const VarDecl * 2972 CGOpenMPRuntimeGPU::translateParameter(const FieldDecl *FD, 2973 const VarDecl *NativeParam) const { 2974 if (!NativeParam->getType()->isReferenceType()) 2975 return NativeParam; 2976 QualType ArgType = NativeParam->getType(); 2977 QualifierCollector QC; 2978 const Type *NonQualTy = QC.strip(ArgType); 2979 QualType PointeeTy = cast<ReferenceType>(NonQualTy)->getPointeeType(); 2980 if (const auto *Attr = FD->getAttr<OMPCaptureKindAttr>()) { 2981 if (Attr->getCaptureKind() == OMPC_map) { 2982 PointeeTy = CGM.getContext().getAddrSpaceQualType(PointeeTy, 2983 LangAS::opencl_global); 2984 } 2985 } 2986 ArgType = CGM.getContext().getPointerType(PointeeTy); 2987 QC.addRestrict(); 2988 enum { NVPTX_local_addr = 5 }; 2989 QC.addAddressSpace(getLangASFromTargetAS(NVPTX_local_addr)); 2990 ArgType = QC.apply(CGM.getContext(), ArgType); 2991 if (isa<ImplicitParamDecl>(NativeParam)) 2992 return ImplicitParamDecl::Create( 2993 CGM.getContext(), /*DC=*/nullptr, NativeParam->getLocation(), 2994 NativeParam->getIdentifier(), ArgType, ImplicitParamKind::Other); 2995 return ParmVarDecl::Create( 2996 CGM.getContext(), 2997 const_cast<DeclContext *>(NativeParam->getDeclContext()), 2998 NativeParam->getBeginLoc(), NativeParam->getLocation(), 2999 NativeParam->getIdentifier(), ArgType, 3000 /*TInfo=*/nullptr, SC_None, /*DefArg=*/nullptr); 3001 } 3002 3003 Address 3004 CGOpenMPRuntimeGPU::getParameterAddress(CodeGenFunction &CGF, 3005 const VarDecl *NativeParam, 3006 const VarDecl *TargetParam) const { 3007 assert(NativeParam != TargetParam && 3008 NativeParam->getType()->isReferenceType() && 3009 "Native arg must not be the same as target arg."); 3010 Address LocalAddr = CGF.GetAddrOfLocalVar(TargetParam); 3011 QualType NativeParamType = NativeParam->getType(); 3012 QualifierCollector QC; 3013 const Type *NonQualTy = QC.strip(NativeParamType); 3014 QualType NativePointeeTy = cast<ReferenceType>(NonQualTy)->getPointeeType(); 3015 unsigned NativePointeeAddrSpace = 3016 CGF.getTypes().getTargetAddressSpace(NativePointeeTy); 3017 QualType TargetTy = TargetParam->getType(); 3018 llvm::Value *TargetAddr = CGF.EmitLoadOfScalar(LocalAddr, /*Volatile=*/false, 3019 TargetTy, SourceLocation()); 3020 // Cast to native address space. 3021 TargetAddr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( 3022 TargetAddr, 3023 llvm::PointerType::get(CGF.getLLVMContext(), NativePointeeAddrSpace)); 3024 Address NativeParamAddr = CGF.CreateMemTemp(NativeParamType); 3025 CGF.EmitStoreOfScalar(TargetAddr, NativeParamAddr, /*Volatile=*/false, 3026 NativeParamType); 3027 return NativeParamAddr; 3028 } 3029 3030 void CGOpenMPRuntimeGPU::emitOutlinedFunctionCall( 3031 CodeGenFunction &CGF, SourceLocation Loc, llvm::FunctionCallee OutlinedFn, 3032 ArrayRef<llvm::Value *> Args) const { 3033 SmallVector<llvm::Value *, 4> TargetArgs; 3034 TargetArgs.reserve(Args.size()); 3035 auto *FnType = OutlinedFn.getFunctionType(); 3036 for (unsigned I = 0, E = Args.size(); I < E; ++I) { 3037 if (FnType->isVarArg() && FnType->getNumParams() <= I) { 3038 TargetArgs.append(std::next(Args.begin(), I), Args.end()); 3039 break; 3040 } 3041 llvm::Type *TargetType = FnType->getParamType(I); 3042 llvm::Value *NativeArg = Args[I]; 3043 if (!TargetType->isPointerTy()) { 3044 TargetArgs.emplace_back(NativeArg); 3045 continue; 3046 } 3047 TargetArgs.emplace_back( 3048 CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(NativeArg, TargetType)); 3049 } 3050 CGOpenMPRuntime::emitOutlinedFunctionCall(CGF, Loc, OutlinedFn, TargetArgs); 3051 } 3052 3053 /// Emit function which wraps the outline parallel region 3054 /// and controls the arguments which are passed to this function. 3055 /// The wrapper ensures that the outlined function is called 3056 /// with the correct arguments when data is shared. 3057 llvm::Function *CGOpenMPRuntimeGPU::createParallelDataSharingWrapper( 3058 llvm::Function *OutlinedParallelFn, const OMPExecutableDirective &D) { 3059 ASTContext &Ctx = CGM.getContext(); 3060 const auto &CS = *D.getCapturedStmt(OMPD_parallel); 3061 3062 // Create a function that takes as argument the source thread. 3063 FunctionArgList WrapperArgs; 3064 QualType Int16QTy = 3065 Ctx.getIntTypeForBitwidth(/*DestWidth=*/16, /*Signed=*/false); 3066 QualType Int32QTy = 3067 Ctx.getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/false); 3068 ImplicitParamDecl ParallelLevelArg(Ctx, /*DC=*/nullptr, D.getBeginLoc(), 3069 /*Id=*/nullptr, Int16QTy, 3070 ImplicitParamKind::Other); 3071 ImplicitParamDecl WrapperArg(Ctx, /*DC=*/nullptr, D.getBeginLoc(), 3072 /*Id=*/nullptr, Int32QTy, 3073 ImplicitParamKind::Other); 3074 WrapperArgs.emplace_back(&ParallelLevelArg); 3075 WrapperArgs.emplace_back(&WrapperArg); 3076 3077 const CGFunctionInfo &CGFI = 3078 CGM.getTypes().arrangeBuiltinFunctionDeclaration(Ctx.VoidTy, WrapperArgs); 3079 3080 auto *Fn = llvm::Function::Create( 3081 CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage, 3082 Twine(OutlinedParallelFn->getName(), "_wrapper"), &CGM.getModule()); 3083 3084 // Ensure we do not inline the function. This is trivially true for the ones 3085 // passed to __kmpc_fork_call but the ones calles in serialized regions 3086 // could be inlined. This is not a perfect but it is closer to the invariant 3087 // we want, namely, every data environment starts with a new function. 3088 // TODO: We should pass the if condition to the runtime function and do the 3089 // handling there. Much cleaner code. 3090 Fn->addFnAttr(llvm::Attribute::NoInline); 3091 3092 CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI); 3093 Fn->setLinkage(llvm::GlobalValue::InternalLinkage); 3094 Fn->setDoesNotRecurse(); 3095 3096 CodeGenFunction CGF(CGM, /*suppressNewContext=*/true); 3097 CGF.StartFunction(GlobalDecl(), Ctx.VoidTy, Fn, CGFI, WrapperArgs, 3098 D.getBeginLoc(), D.getBeginLoc()); 3099 3100 const auto *RD = CS.getCapturedRecordDecl(); 3101 auto CurField = RD->field_begin(); 3102 3103 Address ZeroAddr = CGF.CreateDefaultAlignTempAlloca(CGF.Int32Ty, 3104 /*Name=*/".zero.addr"); 3105 CGF.Builder.CreateStore(CGF.Builder.getInt32(/*C*/ 0), ZeroAddr); 3106 // Get the array of arguments. 3107 SmallVector<llvm::Value *, 8> Args; 3108 3109 Args.emplace_back(CGF.GetAddrOfLocalVar(&WrapperArg).getPointer()); 3110 Args.emplace_back(ZeroAddr.getPointer()); 3111 3112 CGBuilderTy &Bld = CGF.Builder; 3113 auto CI = CS.capture_begin(); 3114 3115 // Use global memory for data sharing. 3116 // Handle passing of global args to workers. 3117 Address GlobalArgs = 3118 CGF.CreateDefaultAlignTempAlloca(CGF.VoidPtrPtrTy, "global_args"); 3119 llvm::Value *GlobalArgsPtr = GlobalArgs.getPointer(); 3120 llvm::Value *DataSharingArgs[] = {GlobalArgsPtr}; 3121 CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( 3122 CGM.getModule(), OMPRTL___kmpc_get_shared_variables), 3123 DataSharingArgs); 3124 3125 // Retrieve the shared variables from the list of references returned 3126 // by the runtime. Pass the variables to the outlined function. 3127 Address SharedArgListAddress = Address::invalid(); 3128 if (CS.capture_size() > 0 || 3129 isOpenMPLoopBoundSharingDirective(D.getDirectiveKind())) { 3130 SharedArgListAddress = CGF.EmitLoadOfPointer( 3131 GlobalArgs, CGF.getContext() 3132 .getPointerType(CGF.getContext().VoidPtrTy) 3133 .castAs<PointerType>()); 3134 } 3135 unsigned Idx = 0; 3136 if (isOpenMPLoopBoundSharingDirective(D.getDirectiveKind())) { 3137 Address Src = Bld.CreateConstInBoundsGEP(SharedArgListAddress, Idx); 3138 Address TypedAddress = Bld.CreatePointerBitCastOrAddrSpaceCast( 3139 Src, CGF.SizeTy->getPointerTo(), CGF.SizeTy); 3140 llvm::Value *LB = CGF.EmitLoadOfScalar( 3141 TypedAddress, 3142 /*Volatile=*/false, 3143 CGF.getContext().getPointerType(CGF.getContext().getSizeType()), 3144 cast<OMPLoopDirective>(D).getLowerBoundVariable()->getExprLoc()); 3145 Args.emplace_back(LB); 3146 ++Idx; 3147 Src = Bld.CreateConstInBoundsGEP(SharedArgListAddress, Idx); 3148 TypedAddress = Bld.CreatePointerBitCastOrAddrSpaceCast( 3149 Src, CGF.SizeTy->getPointerTo(), CGF.SizeTy); 3150 llvm::Value *UB = CGF.EmitLoadOfScalar( 3151 TypedAddress, 3152 /*Volatile=*/false, 3153 CGF.getContext().getPointerType(CGF.getContext().getSizeType()), 3154 cast<OMPLoopDirective>(D).getUpperBoundVariable()->getExprLoc()); 3155 Args.emplace_back(UB); 3156 ++Idx; 3157 } 3158 if (CS.capture_size() > 0) { 3159 ASTContext &CGFContext = CGF.getContext(); 3160 for (unsigned I = 0, E = CS.capture_size(); I < E; ++I, ++CI, ++CurField) { 3161 QualType ElemTy = CurField->getType(); 3162 Address Src = Bld.CreateConstInBoundsGEP(SharedArgListAddress, I + Idx); 3163 Address TypedAddress = Bld.CreatePointerBitCastOrAddrSpaceCast( 3164 Src, CGF.ConvertTypeForMem(CGFContext.getPointerType(ElemTy)), 3165 CGF.ConvertTypeForMem(ElemTy)); 3166 llvm::Value *Arg = CGF.EmitLoadOfScalar(TypedAddress, 3167 /*Volatile=*/false, 3168 CGFContext.getPointerType(ElemTy), 3169 CI->getLocation()); 3170 if (CI->capturesVariableByCopy() && 3171 !CI->getCapturedVar()->getType()->isAnyPointerType()) { 3172 Arg = castValueToType(CGF, Arg, ElemTy, CGFContext.getUIntPtrType(), 3173 CI->getLocation()); 3174 } 3175 Args.emplace_back(Arg); 3176 } 3177 } 3178 3179 emitOutlinedFunctionCall(CGF, D.getBeginLoc(), OutlinedParallelFn, Args); 3180 CGF.FinishFunction(); 3181 return Fn; 3182 } 3183 3184 void CGOpenMPRuntimeGPU::emitFunctionProlog(CodeGenFunction &CGF, 3185 const Decl *D) { 3186 if (getDataSharingMode() != CGOpenMPRuntimeGPU::DS_Generic) 3187 return; 3188 3189 assert(D && "Expected function or captured|block decl."); 3190 assert(FunctionGlobalizedDecls.count(CGF.CurFn) == 0 && 3191 "Function is registered already."); 3192 assert((!TeamAndReductions.first || TeamAndReductions.first == D) && 3193 "Team is set but not processed."); 3194 const Stmt *Body = nullptr; 3195 bool NeedToDelayGlobalization = false; 3196 if (const auto *FD = dyn_cast<FunctionDecl>(D)) { 3197 Body = FD->getBody(); 3198 } else if (const auto *BD = dyn_cast<BlockDecl>(D)) { 3199 Body = BD->getBody(); 3200 } else if (const auto *CD = dyn_cast<CapturedDecl>(D)) { 3201 Body = CD->getBody(); 3202 NeedToDelayGlobalization = CGF.CapturedStmtInfo->getKind() == CR_OpenMP; 3203 if (NeedToDelayGlobalization && 3204 getExecutionMode() == CGOpenMPRuntimeGPU::EM_SPMD) 3205 return; 3206 } 3207 if (!Body) 3208 return; 3209 CheckVarsEscapingDeclContext VarChecker(CGF, TeamAndReductions.second); 3210 VarChecker.Visit(Body); 3211 const RecordDecl *GlobalizedVarsRecord = 3212 VarChecker.getGlobalizedRecord(IsInTTDRegion); 3213 TeamAndReductions.first = nullptr; 3214 TeamAndReductions.second.clear(); 3215 ArrayRef<const ValueDecl *> EscapedVariableLengthDecls = 3216 VarChecker.getEscapedVariableLengthDecls(); 3217 ArrayRef<const ValueDecl *> DelayedVariableLengthDecls = 3218 VarChecker.getDelayedVariableLengthDecls(); 3219 if (!GlobalizedVarsRecord && EscapedVariableLengthDecls.empty() && 3220 DelayedVariableLengthDecls.empty()) 3221 return; 3222 auto I = FunctionGlobalizedDecls.try_emplace(CGF.CurFn).first; 3223 I->getSecond().MappedParams = 3224 std::make_unique<CodeGenFunction::OMPMapVars>(); 3225 I->getSecond().EscapedParameters.insert( 3226 VarChecker.getEscapedParameters().begin(), 3227 VarChecker.getEscapedParameters().end()); 3228 I->getSecond().EscapedVariableLengthDecls.append( 3229 EscapedVariableLengthDecls.begin(), EscapedVariableLengthDecls.end()); 3230 I->getSecond().DelayedVariableLengthDecls.append( 3231 DelayedVariableLengthDecls.begin(), DelayedVariableLengthDecls.end()); 3232 DeclToAddrMapTy &Data = I->getSecond().LocalVarData; 3233 for (const ValueDecl *VD : VarChecker.getEscapedDecls()) { 3234 assert(VD->isCanonicalDecl() && "Expected canonical declaration"); 3235 Data.insert(std::make_pair(VD, MappedVarData())); 3236 } 3237 if (!NeedToDelayGlobalization) { 3238 emitGenericVarsProlog(CGF, D->getBeginLoc()); 3239 struct GlobalizationScope final : EHScopeStack::Cleanup { 3240 GlobalizationScope() = default; 3241 3242 void Emit(CodeGenFunction &CGF, Flags flags) override { 3243 static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()) 3244 .emitGenericVarsEpilog(CGF); 3245 } 3246 }; 3247 CGF.EHStack.pushCleanup<GlobalizationScope>(NormalAndEHCleanup); 3248 } 3249 } 3250 3251 Address CGOpenMPRuntimeGPU::getAddressOfLocalVariable(CodeGenFunction &CGF, 3252 const VarDecl *VD) { 3253 if (VD && VD->hasAttr<OMPAllocateDeclAttr>()) { 3254 const auto *A = VD->getAttr<OMPAllocateDeclAttr>(); 3255 auto AS = LangAS::Default; 3256 switch (A->getAllocatorType()) { 3257 // Use the default allocator here as by default local vars are 3258 // threadlocal. 3259 case OMPAllocateDeclAttr::OMPNullMemAlloc: 3260 case OMPAllocateDeclAttr::OMPDefaultMemAlloc: 3261 case OMPAllocateDeclAttr::OMPThreadMemAlloc: 3262 case OMPAllocateDeclAttr::OMPHighBWMemAlloc: 3263 case OMPAllocateDeclAttr::OMPLowLatMemAlloc: 3264 // Follow the user decision - use default allocation. 3265 return Address::invalid(); 3266 case OMPAllocateDeclAttr::OMPUserDefinedMemAlloc: 3267 // TODO: implement aupport for user-defined allocators. 3268 return Address::invalid(); 3269 case OMPAllocateDeclAttr::OMPConstMemAlloc: 3270 AS = LangAS::cuda_constant; 3271 break; 3272 case OMPAllocateDeclAttr::OMPPTeamMemAlloc: 3273 AS = LangAS::cuda_shared; 3274 break; 3275 case OMPAllocateDeclAttr::OMPLargeCapMemAlloc: 3276 case OMPAllocateDeclAttr::OMPCGroupMemAlloc: 3277 break; 3278 } 3279 llvm::Type *VarTy = CGF.ConvertTypeForMem(VD->getType()); 3280 auto *GV = new llvm::GlobalVariable( 3281 CGM.getModule(), VarTy, /*isConstant=*/false, 3282 llvm::GlobalValue::InternalLinkage, llvm::PoisonValue::get(VarTy), 3283 VD->getName(), 3284 /*InsertBefore=*/nullptr, llvm::GlobalValue::NotThreadLocal, 3285 CGM.getContext().getTargetAddressSpace(AS)); 3286 CharUnits Align = CGM.getContext().getDeclAlign(VD); 3287 GV->setAlignment(Align.getAsAlign()); 3288 return Address( 3289 CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( 3290 GV, VarTy->getPointerTo(CGM.getContext().getTargetAddressSpace( 3291 VD->getType().getAddressSpace()))), 3292 VarTy, Align); 3293 } 3294 3295 if (getDataSharingMode() != CGOpenMPRuntimeGPU::DS_Generic) 3296 return Address::invalid(); 3297 3298 VD = VD->getCanonicalDecl(); 3299 auto I = FunctionGlobalizedDecls.find(CGF.CurFn); 3300 if (I == FunctionGlobalizedDecls.end()) 3301 return Address::invalid(); 3302 auto VDI = I->getSecond().LocalVarData.find(VD); 3303 if (VDI != I->getSecond().LocalVarData.end()) 3304 return VDI->second.PrivateAddr; 3305 if (VD->hasAttrs()) { 3306 for (specific_attr_iterator<OMPReferencedVarAttr> IT(VD->attr_begin()), 3307 E(VD->attr_end()); 3308 IT != E; ++IT) { 3309 auto VDI = I->getSecond().LocalVarData.find( 3310 cast<VarDecl>(cast<DeclRefExpr>(IT->getRef())->getDecl()) 3311 ->getCanonicalDecl()); 3312 if (VDI != I->getSecond().LocalVarData.end()) 3313 return VDI->second.PrivateAddr; 3314 } 3315 } 3316 3317 return Address::invalid(); 3318 } 3319 3320 void CGOpenMPRuntimeGPU::functionFinished(CodeGenFunction &CGF) { 3321 FunctionGlobalizedDecls.erase(CGF.CurFn); 3322 CGOpenMPRuntime::functionFinished(CGF); 3323 } 3324 3325 void CGOpenMPRuntimeGPU::getDefaultDistScheduleAndChunk( 3326 CodeGenFunction &CGF, const OMPLoopDirective &S, 3327 OpenMPDistScheduleClauseKind &ScheduleKind, 3328 llvm::Value *&Chunk) const { 3329 auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()); 3330 if (getExecutionMode() == CGOpenMPRuntimeGPU::EM_SPMD) { 3331 ScheduleKind = OMPC_DIST_SCHEDULE_static; 3332 Chunk = CGF.EmitScalarConversion( 3333 RT.getGPUNumThreads(CGF), 3334 CGF.getContext().getIntTypeForBitwidth(32, /*Signed=*/0), 3335 S.getIterationVariable()->getType(), S.getBeginLoc()); 3336 return; 3337 } 3338 CGOpenMPRuntime::getDefaultDistScheduleAndChunk( 3339 CGF, S, ScheduleKind, Chunk); 3340 } 3341 3342 void CGOpenMPRuntimeGPU::getDefaultScheduleAndChunk( 3343 CodeGenFunction &CGF, const OMPLoopDirective &S, 3344 OpenMPScheduleClauseKind &ScheduleKind, 3345 const Expr *&ChunkExpr) const { 3346 ScheduleKind = OMPC_SCHEDULE_static; 3347 // Chunk size is 1 in this case. 3348 llvm::APInt ChunkSize(32, 1); 3349 ChunkExpr = IntegerLiteral::Create(CGF.getContext(), ChunkSize, 3350 CGF.getContext().getIntTypeForBitwidth(32, /*Signed=*/0), 3351 SourceLocation()); 3352 } 3353 3354 void CGOpenMPRuntimeGPU::adjustTargetSpecificDataForLambdas( 3355 CodeGenFunction &CGF, const OMPExecutableDirective &D) const { 3356 assert(isOpenMPTargetExecutionDirective(D.getDirectiveKind()) && 3357 " Expected target-based directive."); 3358 const CapturedStmt *CS = D.getCapturedStmt(OMPD_target); 3359 for (const CapturedStmt::Capture &C : CS->captures()) { 3360 // Capture variables captured by reference in lambdas for target-based 3361 // directives. 3362 if (!C.capturesVariable()) 3363 continue; 3364 const VarDecl *VD = C.getCapturedVar(); 3365 const auto *RD = VD->getType() 3366 .getCanonicalType() 3367 .getNonReferenceType() 3368 ->getAsCXXRecordDecl(); 3369 if (!RD || !RD->isLambda()) 3370 continue; 3371 Address VDAddr = CGF.GetAddrOfLocalVar(VD); 3372 LValue VDLVal; 3373 if (VD->getType().getCanonicalType()->isReferenceType()) 3374 VDLVal = CGF.EmitLoadOfReferenceLValue(VDAddr, VD->getType()); 3375 else 3376 VDLVal = CGF.MakeAddrLValue( 3377 VDAddr, VD->getType().getCanonicalType().getNonReferenceType()); 3378 llvm::DenseMap<const ValueDecl *, FieldDecl *> Captures; 3379 FieldDecl *ThisCapture = nullptr; 3380 RD->getCaptureFields(Captures, ThisCapture); 3381 if (ThisCapture && CGF.CapturedStmtInfo->isCXXThisExprCaptured()) { 3382 LValue ThisLVal = 3383 CGF.EmitLValueForFieldInitialization(VDLVal, ThisCapture); 3384 llvm::Value *CXXThis = CGF.LoadCXXThis(); 3385 CGF.EmitStoreOfScalar(CXXThis, ThisLVal); 3386 } 3387 for (const LambdaCapture &LC : RD->captures()) { 3388 if (LC.getCaptureKind() != LCK_ByRef) 3389 continue; 3390 const ValueDecl *VD = LC.getCapturedVar(); 3391 // FIXME: For now VD is always a VarDecl because OpenMP does not support 3392 // capturing structured bindings in lambdas yet. 3393 if (!CS->capturesVariable(cast<VarDecl>(VD))) 3394 continue; 3395 auto It = Captures.find(VD); 3396 assert(It != Captures.end() && "Found lambda capture without field."); 3397 LValue VarLVal = CGF.EmitLValueForFieldInitialization(VDLVal, It->second); 3398 Address VDAddr = CGF.GetAddrOfLocalVar(cast<VarDecl>(VD)); 3399 if (VD->getType().getCanonicalType()->isReferenceType()) 3400 VDAddr = CGF.EmitLoadOfReferenceLValue(VDAddr, 3401 VD->getType().getCanonicalType()) 3402 .getAddress(CGF); 3403 CGF.EmitStoreOfScalar(VDAddr.getPointer(), VarLVal); 3404 } 3405 } 3406 } 3407 3408 bool CGOpenMPRuntimeGPU::hasAllocateAttributeForGlobalVar(const VarDecl *VD, 3409 LangAS &AS) { 3410 if (!VD || !VD->hasAttr<OMPAllocateDeclAttr>()) 3411 return false; 3412 const auto *A = VD->getAttr<OMPAllocateDeclAttr>(); 3413 switch(A->getAllocatorType()) { 3414 case OMPAllocateDeclAttr::OMPNullMemAlloc: 3415 case OMPAllocateDeclAttr::OMPDefaultMemAlloc: 3416 // Not supported, fallback to the default mem space. 3417 case OMPAllocateDeclAttr::OMPThreadMemAlloc: 3418 case OMPAllocateDeclAttr::OMPLargeCapMemAlloc: 3419 case OMPAllocateDeclAttr::OMPCGroupMemAlloc: 3420 case OMPAllocateDeclAttr::OMPHighBWMemAlloc: 3421 case OMPAllocateDeclAttr::OMPLowLatMemAlloc: 3422 AS = LangAS::Default; 3423 return true; 3424 case OMPAllocateDeclAttr::OMPConstMemAlloc: 3425 AS = LangAS::cuda_constant; 3426 return true; 3427 case OMPAllocateDeclAttr::OMPPTeamMemAlloc: 3428 AS = LangAS::cuda_shared; 3429 return true; 3430 case OMPAllocateDeclAttr::OMPUserDefinedMemAlloc: 3431 llvm_unreachable("Expected predefined allocator for the variables with the " 3432 "static storage."); 3433 } 3434 return false; 3435 } 3436 3437 // Get current CudaArch and ignore any unknown values 3438 static CudaArch getCudaArch(CodeGenModule &CGM) { 3439 if (!CGM.getTarget().hasFeature("ptx")) 3440 return CudaArch::UNKNOWN; 3441 for (const auto &Feature : CGM.getTarget().getTargetOpts().FeatureMap) { 3442 if (Feature.getValue()) { 3443 CudaArch Arch = StringToCudaArch(Feature.getKey()); 3444 if (Arch != CudaArch::UNKNOWN) 3445 return Arch; 3446 } 3447 } 3448 return CudaArch::UNKNOWN; 3449 } 3450 3451 /// Check to see if target architecture supports unified addressing which is 3452 /// a restriction for OpenMP requires clause "unified_shared_memory". 3453 void CGOpenMPRuntimeGPU::processRequiresDirective( 3454 const OMPRequiresDecl *D) { 3455 for (const OMPClause *Clause : D->clauselists()) { 3456 if (Clause->getClauseKind() == OMPC_unified_shared_memory) { 3457 CudaArch Arch = getCudaArch(CGM); 3458 switch (Arch) { 3459 case CudaArch::SM_20: 3460 case CudaArch::SM_21: 3461 case CudaArch::SM_30: 3462 case CudaArch::SM_32: 3463 case CudaArch::SM_35: 3464 case CudaArch::SM_37: 3465 case CudaArch::SM_50: 3466 case CudaArch::SM_52: 3467 case CudaArch::SM_53: { 3468 SmallString<256> Buffer; 3469 llvm::raw_svector_ostream Out(Buffer); 3470 Out << "Target architecture " << CudaArchToString(Arch) 3471 << " does not support unified addressing"; 3472 CGM.Error(Clause->getBeginLoc(), Out.str()); 3473 return; 3474 } 3475 case CudaArch::SM_60: 3476 case CudaArch::SM_61: 3477 case CudaArch::SM_62: 3478 case CudaArch::SM_70: 3479 case CudaArch::SM_72: 3480 case CudaArch::SM_75: 3481 case CudaArch::SM_80: 3482 case CudaArch::SM_86: 3483 case CudaArch::SM_87: 3484 case CudaArch::SM_89: 3485 case CudaArch::SM_90: 3486 case CudaArch::SM_90a: 3487 case CudaArch::GFX600: 3488 case CudaArch::GFX601: 3489 case CudaArch::GFX602: 3490 case CudaArch::GFX700: 3491 case CudaArch::GFX701: 3492 case CudaArch::GFX702: 3493 case CudaArch::GFX703: 3494 case CudaArch::GFX704: 3495 case CudaArch::GFX705: 3496 case CudaArch::GFX801: 3497 case CudaArch::GFX802: 3498 case CudaArch::GFX803: 3499 case CudaArch::GFX805: 3500 case CudaArch::GFX810: 3501 case CudaArch::GFX900: 3502 case CudaArch::GFX902: 3503 case CudaArch::GFX904: 3504 case CudaArch::GFX906: 3505 case CudaArch::GFX908: 3506 case CudaArch::GFX909: 3507 case CudaArch::GFX90a: 3508 case CudaArch::GFX90c: 3509 case CudaArch::GFX940: 3510 case CudaArch::GFX941: 3511 case CudaArch::GFX942: 3512 case CudaArch::GFX1010: 3513 case CudaArch::GFX1011: 3514 case CudaArch::GFX1012: 3515 case CudaArch::GFX1013: 3516 case CudaArch::GFX1030: 3517 case CudaArch::GFX1031: 3518 case CudaArch::GFX1032: 3519 case CudaArch::GFX1033: 3520 case CudaArch::GFX1034: 3521 case CudaArch::GFX1035: 3522 case CudaArch::GFX1036: 3523 case CudaArch::GFX1100: 3524 case CudaArch::GFX1101: 3525 case CudaArch::GFX1102: 3526 case CudaArch::GFX1103: 3527 case CudaArch::GFX1150: 3528 case CudaArch::GFX1151: 3529 case CudaArch::GFX1200: 3530 case CudaArch::GFX1201: 3531 case CudaArch::Generic: 3532 case CudaArch::UNUSED: 3533 case CudaArch::UNKNOWN: 3534 break; 3535 case CudaArch::LAST: 3536 llvm_unreachable("Unexpected Cuda arch."); 3537 } 3538 } 3539 } 3540 CGOpenMPRuntime::processRequiresDirective(D); 3541 } 3542 3543 llvm::Value *CGOpenMPRuntimeGPU::getGPUNumThreads(CodeGenFunction &CGF) { 3544 CGBuilderTy &Bld = CGF.Builder; 3545 llvm::Module *M = &CGF.CGM.getModule(); 3546 const char *LocSize = "__kmpc_get_hardware_num_threads_in_block"; 3547 llvm::Function *F = M->getFunction(LocSize); 3548 if (!F) { 3549 F = llvm::Function::Create( 3550 llvm::FunctionType::get(CGF.Int32Ty, std::nullopt, false), 3551 llvm::GlobalVariable::ExternalLinkage, LocSize, &CGF.CGM.getModule()); 3552 } 3553 return Bld.CreateCall(F, std::nullopt, "nvptx_num_threads"); 3554 } 3555 3556 llvm::Value *CGOpenMPRuntimeGPU::getGPUThreadID(CodeGenFunction &CGF) { 3557 ArrayRef<llvm::Value *> Args{}; 3558 return CGF.EmitRuntimeCall( 3559 OMPBuilder.getOrCreateRuntimeFunction( 3560 CGM.getModule(), OMPRTL___kmpc_get_hardware_thread_id_in_block), 3561 Args); 3562 } 3563 3564 llvm::Value *CGOpenMPRuntimeGPU::getGPUWarpSize(CodeGenFunction &CGF) { 3565 ArrayRef<llvm::Value *> Args{}; 3566 return CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( 3567 CGM.getModule(), OMPRTL___kmpc_get_warp_size), 3568 Args); 3569 } 3570