1 //===---- CGOpenMPRuntimeGPU.cpp - Interface to OpenMP GPU Runtimes ----===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This provides a generalized class for OpenMP runtime code generation 10 // specialized by GPU targets NVPTX and AMDGCN. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "CGOpenMPRuntimeGPU.h" 15 #include "CodeGenFunction.h" 16 #include "clang/AST/Attr.h" 17 #include "clang/AST/DeclOpenMP.h" 18 #include "clang/AST/StmtOpenMP.h" 19 #include "clang/AST/StmtVisitor.h" 20 #include "clang/Basic/Cuda.h" 21 #include "llvm/ADT/SmallPtrSet.h" 22 #include "llvm/Frontend/OpenMP/OMPGridValues.h" 23 #include "llvm/Support/MathExtras.h" 24 25 using namespace clang; 26 using namespace CodeGen; 27 using namespace llvm::omp; 28 29 namespace { 30 /// Pre(post)-action for different OpenMP constructs specialized for NVPTX. 31 class NVPTXActionTy final : public PrePostActionTy { 32 llvm::FunctionCallee EnterCallee = nullptr; 33 ArrayRef<llvm::Value *> EnterArgs; 34 llvm::FunctionCallee ExitCallee = nullptr; 35 ArrayRef<llvm::Value *> ExitArgs; 36 bool Conditional = false; 37 llvm::BasicBlock *ContBlock = nullptr; 38 39 public: 40 NVPTXActionTy(llvm::FunctionCallee EnterCallee, 41 ArrayRef<llvm::Value *> EnterArgs, 42 llvm::FunctionCallee ExitCallee, 43 ArrayRef<llvm::Value *> ExitArgs, bool Conditional = false) 44 : EnterCallee(EnterCallee), EnterArgs(EnterArgs), ExitCallee(ExitCallee), 45 ExitArgs(ExitArgs), Conditional(Conditional) {} 46 void Enter(CodeGenFunction &CGF) override { 47 llvm::Value *EnterRes = CGF.EmitRuntimeCall(EnterCallee, EnterArgs); 48 if (Conditional) { 49 llvm::Value *CallBool = CGF.Builder.CreateIsNotNull(EnterRes); 50 auto *ThenBlock = CGF.createBasicBlock("omp_if.then"); 51 ContBlock = CGF.createBasicBlock("omp_if.end"); 52 // Generate the branch (If-stmt) 53 CGF.Builder.CreateCondBr(CallBool, ThenBlock, ContBlock); 54 CGF.EmitBlock(ThenBlock); 55 } 56 } 57 void Done(CodeGenFunction &CGF) { 58 // Emit the rest of blocks/branches 59 CGF.EmitBranch(ContBlock); 60 CGF.EmitBlock(ContBlock, true); 61 } 62 void Exit(CodeGenFunction &CGF) override { 63 CGF.EmitRuntimeCall(ExitCallee, ExitArgs); 64 } 65 }; 66 67 /// A class to track the execution mode when codegening directives within 68 /// a target region. The appropriate mode (SPMD|NON-SPMD) is set on entry 69 /// to the target region and used by containing directives such as 'parallel' 70 /// to emit optimized code. 71 class ExecutionRuntimeModesRAII { 72 private: 73 CGOpenMPRuntimeGPU::ExecutionMode SavedExecMode = 74 CGOpenMPRuntimeGPU::EM_Unknown; 75 CGOpenMPRuntimeGPU::ExecutionMode &ExecMode; 76 bool SavedRuntimeMode = false; 77 bool *RuntimeMode = nullptr; 78 79 public: 80 /// Constructor for Non-SPMD mode. 81 ExecutionRuntimeModesRAII(CGOpenMPRuntimeGPU::ExecutionMode &ExecMode) 82 : ExecMode(ExecMode) { 83 SavedExecMode = ExecMode; 84 ExecMode = CGOpenMPRuntimeGPU::EM_NonSPMD; 85 } 86 /// Constructor for SPMD mode. 87 ExecutionRuntimeModesRAII(CGOpenMPRuntimeGPU::ExecutionMode &ExecMode, 88 bool &RuntimeMode, bool FullRuntimeMode) 89 : ExecMode(ExecMode), RuntimeMode(&RuntimeMode) { 90 SavedExecMode = ExecMode; 91 SavedRuntimeMode = RuntimeMode; 92 ExecMode = CGOpenMPRuntimeGPU::EM_SPMD; 93 RuntimeMode = FullRuntimeMode; 94 } 95 ~ExecutionRuntimeModesRAII() { 96 ExecMode = SavedExecMode; 97 if (RuntimeMode) 98 *RuntimeMode = SavedRuntimeMode; 99 } 100 }; 101 102 /// GPU Configuration: This information can be derived from cuda registers, 103 /// however, providing compile time constants helps generate more efficient 104 /// code. For all practical purposes this is fine because the configuration 105 /// is the same for all known NVPTX architectures. 106 enum MachineConfiguration : unsigned { 107 /// See "llvm/Frontend/OpenMP/OMPGridValues.h" for various related target 108 /// specific Grid Values like GV_Warp_Size, GV_Slot_Size 109 110 /// Global memory alignment for performance. 111 GlobalMemoryAlignment = 128, 112 113 /// Maximal size of the shared memory buffer. 114 SharedMemorySize = 128, 115 }; 116 117 static const ValueDecl *getPrivateItem(const Expr *RefExpr) { 118 RefExpr = RefExpr->IgnoreParens(); 119 if (const auto *ASE = dyn_cast<ArraySubscriptExpr>(RefExpr)) { 120 const Expr *Base = ASE->getBase()->IgnoreParenImpCasts(); 121 while (const auto *TempASE = dyn_cast<ArraySubscriptExpr>(Base)) 122 Base = TempASE->getBase()->IgnoreParenImpCasts(); 123 RefExpr = Base; 124 } else if (auto *OASE = dyn_cast<OMPArraySectionExpr>(RefExpr)) { 125 const Expr *Base = OASE->getBase()->IgnoreParenImpCasts(); 126 while (const auto *TempOASE = dyn_cast<OMPArraySectionExpr>(Base)) 127 Base = TempOASE->getBase()->IgnoreParenImpCasts(); 128 while (const auto *TempASE = dyn_cast<ArraySubscriptExpr>(Base)) 129 Base = TempASE->getBase()->IgnoreParenImpCasts(); 130 RefExpr = Base; 131 } 132 RefExpr = RefExpr->IgnoreParenImpCasts(); 133 if (const auto *DE = dyn_cast<DeclRefExpr>(RefExpr)) 134 return cast<ValueDecl>(DE->getDecl()->getCanonicalDecl()); 135 const auto *ME = cast<MemberExpr>(RefExpr); 136 return cast<ValueDecl>(ME->getMemberDecl()->getCanonicalDecl()); 137 } 138 139 140 static RecordDecl *buildRecordForGlobalizedVars( 141 ASTContext &C, ArrayRef<const ValueDecl *> EscapedDecls, 142 ArrayRef<const ValueDecl *> EscapedDeclsForTeams, 143 llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> 144 &MappedDeclsFields, int BufSize) { 145 using VarsDataTy = std::pair<CharUnits /*Align*/, const ValueDecl *>; 146 if (EscapedDecls.empty() && EscapedDeclsForTeams.empty()) 147 return nullptr; 148 SmallVector<VarsDataTy, 4> GlobalizedVars; 149 for (const ValueDecl *D : EscapedDecls) 150 GlobalizedVars.emplace_back( 151 CharUnits::fromQuantity(std::max( 152 C.getDeclAlign(D).getQuantity(), 153 static_cast<CharUnits::QuantityType>(GlobalMemoryAlignment))), 154 D); 155 for (const ValueDecl *D : EscapedDeclsForTeams) 156 GlobalizedVars.emplace_back(C.getDeclAlign(D), D); 157 llvm::stable_sort(GlobalizedVars, [](VarsDataTy L, VarsDataTy R) { 158 return L.first > R.first; 159 }); 160 161 // Build struct _globalized_locals_ty { 162 // /* globalized vars */[WarSize] align (max(decl_align, 163 // GlobalMemoryAlignment)) 164 // /* globalized vars */ for EscapedDeclsForTeams 165 // }; 166 RecordDecl *GlobalizedRD = C.buildImplicitRecord("_globalized_locals_ty"); 167 GlobalizedRD->startDefinition(); 168 llvm::SmallPtrSet<const ValueDecl *, 16> SingleEscaped( 169 EscapedDeclsForTeams.begin(), EscapedDeclsForTeams.end()); 170 for (const auto &Pair : GlobalizedVars) { 171 const ValueDecl *VD = Pair.second; 172 QualType Type = VD->getType(); 173 if (Type->isLValueReferenceType()) 174 Type = C.getPointerType(Type.getNonReferenceType()); 175 else 176 Type = Type.getNonReferenceType(); 177 SourceLocation Loc = VD->getLocation(); 178 FieldDecl *Field; 179 if (SingleEscaped.count(VD)) { 180 Field = FieldDecl::Create( 181 C, GlobalizedRD, Loc, Loc, VD->getIdentifier(), Type, 182 C.getTrivialTypeSourceInfo(Type, SourceLocation()), 183 /*BW=*/nullptr, /*Mutable=*/false, 184 /*InitStyle=*/ICIS_NoInit); 185 Field->setAccess(AS_public); 186 if (VD->hasAttrs()) { 187 for (specific_attr_iterator<AlignedAttr> I(VD->getAttrs().begin()), 188 E(VD->getAttrs().end()); 189 I != E; ++I) 190 Field->addAttr(*I); 191 } 192 } else { 193 llvm::APInt ArraySize(32, BufSize); 194 Type = C.getConstantArrayType(Type, ArraySize, nullptr, ArrayType::Normal, 195 0); 196 Field = FieldDecl::Create( 197 C, GlobalizedRD, Loc, Loc, VD->getIdentifier(), Type, 198 C.getTrivialTypeSourceInfo(Type, SourceLocation()), 199 /*BW=*/nullptr, /*Mutable=*/false, 200 /*InitStyle=*/ICIS_NoInit); 201 Field->setAccess(AS_public); 202 llvm::APInt Align(32, std::max(C.getDeclAlign(VD).getQuantity(), 203 static_cast<CharUnits::QuantityType>( 204 GlobalMemoryAlignment))); 205 Field->addAttr(AlignedAttr::CreateImplicit( 206 C, /*IsAlignmentExpr=*/true, 207 IntegerLiteral::Create(C, Align, 208 C.getIntTypeForBitwidth(32, /*Signed=*/0), 209 SourceLocation()), 210 {}, AttributeCommonInfo::AS_GNU, AlignedAttr::GNU_aligned)); 211 } 212 GlobalizedRD->addDecl(Field); 213 MappedDeclsFields.try_emplace(VD, Field); 214 } 215 GlobalizedRD->completeDefinition(); 216 return GlobalizedRD; 217 } 218 219 /// Get the list of variables that can escape their declaration context. 220 class CheckVarsEscapingDeclContext final 221 : public ConstStmtVisitor<CheckVarsEscapingDeclContext> { 222 CodeGenFunction &CGF; 223 llvm::SetVector<const ValueDecl *> EscapedDecls; 224 llvm::SetVector<const ValueDecl *> EscapedVariableLengthDecls; 225 llvm::SmallPtrSet<const Decl *, 4> EscapedParameters; 226 RecordDecl *GlobalizedRD = nullptr; 227 llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> MappedDeclsFields; 228 bool AllEscaped = false; 229 bool IsForCombinedParallelRegion = false; 230 231 void markAsEscaped(const ValueDecl *VD) { 232 // Do not globalize declare target variables. 233 if (!isa<VarDecl>(VD) || 234 OMPDeclareTargetDeclAttr::isDeclareTargetDeclaration(VD)) 235 return; 236 VD = cast<ValueDecl>(VD->getCanonicalDecl()); 237 // Use user-specified allocation. 238 if (VD->hasAttrs() && VD->hasAttr<OMPAllocateDeclAttr>()) 239 return; 240 // Variables captured by value must be globalized. 241 if (auto *CSI = CGF.CapturedStmtInfo) { 242 if (const FieldDecl *FD = CSI->lookup(cast<VarDecl>(VD))) { 243 // Check if need to capture the variable that was already captured by 244 // value in the outer region. 245 if (!IsForCombinedParallelRegion) { 246 if (!FD->hasAttrs()) 247 return; 248 const auto *Attr = FD->getAttr<OMPCaptureKindAttr>(); 249 if (!Attr) 250 return; 251 if (((Attr->getCaptureKind() != OMPC_map) && 252 !isOpenMPPrivate(Attr->getCaptureKind())) || 253 ((Attr->getCaptureKind() == OMPC_map) && 254 !FD->getType()->isAnyPointerType())) 255 return; 256 } 257 if (!FD->getType()->isReferenceType()) { 258 assert(!VD->getType()->isVariablyModifiedType() && 259 "Parameter captured by value with variably modified type"); 260 EscapedParameters.insert(VD); 261 } else if (!IsForCombinedParallelRegion) { 262 return; 263 } 264 } 265 } 266 if ((!CGF.CapturedStmtInfo || 267 (IsForCombinedParallelRegion && CGF.CapturedStmtInfo)) && 268 VD->getType()->isReferenceType()) 269 // Do not globalize variables with reference type. 270 return; 271 if (VD->getType()->isVariablyModifiedType()) 272 EscapedVariableLengthDecls.insert(VD); 273 else 274 EscapedDecls.insert(VD); 275 } 276 277 void VisitValueDecl(const ValueDecl *VD) { 278 if (VD->getType()->isLValueReferenceType()) 279 markAsEscaped(VD); 280 if (const auto *VarD = dyn_cast<VarDecl>(VD)) { 281 if (!isa<ParmVarDecl>(VarD) && VarD->hasInit()) { 282 const bool SavedAllEscaped = AllEscaped; 283 AllEscaped = VD->getType()->isLValueReferenceType(); 284 Visit(VarD->getInit()); 285 AllEscaped = SavedAllEscaped; 286 } 287 } 288 } 289 void VisitOpenMPCapturedStmt(const CapturedStmt *S, 290 ArrayRef<OMPClause *> Clauses, 291 bool IsCombinedParallelRegion) { 292 if (!S) 293 return; 294 for (const CapturedStmt::Capture &C : S->captures()) { 295 if (C.capturesVariable() && !C.capturesVariableByCopy()) { 296 const ValueDecl *VD = C.getCapturedVar(); 297 bool SavedIsForCombinedParallelRegion = IsForCombinedParallelRegion; 298 if (IsCombinedParallelRegion) { 299 // Check if the variable is privatized in the combined construct and 300 // those private copies must be shared in the inner parallel 301 // directive. 302 IsForCombinedParallelRegion = false; 303 for (const OMPClause *C : Clauses) { 304 if (!isOpenMPPrivate(C->getClauseKind()) || 305 C->getClauseKind() == OMPC_reduction || 306 C->getClauseKind() == OMPC_linear || 307 C->getClauseKind() == OMPC_private) 308 continue; 309 ArrayRef<const Expr *> Vars; 310 if (const auto *PC = dyn_cast<OMPFirstprivateClause>(C)) 311 Vars = PC->getVarRefs(); 312 else if (const auto *PC = dyn_cast<OMPLastprivateClause>(C)) 313 Vars = PC->getVarRefs(); 314 else 315 llvm_unreachable("Unexpected clause."); 316 for (const auto *E : Vars) { 317 const Decl *D = 318 cast<DeclRefExpr>(E)->getDecl()->getCanonicalDecl(); 319 if (D == VD->getCanonicalDecl()) { 320 IsForCombinedParallelRegion = true; 321 break; 322 } 323 } 324 if (IsForCombinedParallelRegion) 325 break; 326 } 327 } 328 markAsEscaped(VD); 329 if (isa<OMPCapturedExprDecl>(VD)) 330 VisitValueDecl(VD); 331 IsForCombinedParallelRegion = SavedIsForCombinedParallelRegion; 332 } 333 } 334 } 335 336 void buildRecordForGlobalizedVars(bool IsInTTDRegion) { 337 assert(!GlobalizedRD && 338 "Record for globalized variables is built already."); 339 ArrayRef<const ValueDecl *> EscapedDeclsForParallel, EscapedDeclsForTeams; 340 unsigned WarpSize = CGF.getTarget().getGridValue().GV_Warp_Size; 341 if (IsInTTDRegion) 342 EscapedDeclsForTeams = EscapedDecls.getArrayRef(); 343 else 344 EscapedDeclsForParallel = EscapedDecls.getArrayRef(); 345 GlobalizedRD = ::buildRecordForGlobalizedVars( 346 CGF.getContext(), EscapedDeclsForParallel, EscapedDeclsForTeams, 347 MappedDeclsFields, WarpSize); 348 } 349 350 public: 351 CheckVarsEscapingDeclContext(CodeGenFunction &CGF, 352 ArrayRef<const ValueDecl *> TeamsReductions) 353 : CGF(CGF), EscapedDecls(TeamsReductions.begin(), TeamsReductions.end()) { 354 } 355 virtual ~CheckVarsEscapingDeclContext() = default; 356 void VisitDeclStmt(const DeclStmt *S) { 357 if (!S) 358 return; 359 for (const Decl *D : S->decls()) 360 if (const auto *VD = dyn_cast_or_null<ValueDecl>(D)) 361 VisitValueDecl(VD); 362 } 363 void VisitOMPExecutableDirective(const OMPExecutableDirective *D) { 364 if (!D) 365 return; 366 if (!D->hasAssociatedStmt()) 367 return; 368 if (const auto *S = 369 dyn_cast_or_null<CapturedStmt>(D->getAssociatedStmt())) { 370 // Do not analyze directives that do not actually require capturing, 371 // like `omp for` or `omp simd` directives. 372 llvm::SmallVector<OpenMPDirectiveKind, 4> CaptureRegions; 373 getOpenMPCaptureRegions(CaptureRegions, D->getDirectiveKind()); 374 if (CaptureRegions.size() == 1 && CaptureRegions.back() == OMPD_unknown) { 375 VisitStmt(S->getCapturedStmt()); 376 return; 377 } 378 VisitOpenMPCapturedStmt( 379 S, D->clauses(), 380 CaptureRegions.back() == OMPD_parallel && 381 isOpenMPDistributeDirective(D->getDirectiveKind())); 382 } 383 } 384 void VisitCapturedStmt(const CapturedStmt *S) { 385 if (!S) 386 return; 387 for (const CapturedStmt::Capture &C : S->captures()) { 388 if (C.capturesVariable() && !C.capturesVariableByCopy()) { 389 const ValueDecl *VD = C.getCapturedVar(); 390 markAsEscaped(VD); 391 if (isa<OMPCapturedExprDecl>(VD)) 392 VisitValueDecl(VD); 393 } 394 } 395 } 396 void VisitLambdaExpr(const LambdaExpr *E) { 397 if (!E) 398 return; 399 for (const LambdaCapture &C : E->captures()) { 400 if (C.capturesVariable()) { 401 if (C.getCaptureKind() == LCK_ByRef) { 402 const ValueDecl *VD = C.getCapturedVar(); 403 markAsEscaped(VD); 404 if (E->isInitCapture(&C) || isa<OMPCapturedExprDecl>(VD)) 405 VisitValueDecl(VD); 406 } 407 } 408 } 409 } 410 void VisitBlockExpr(const BlockExpr *E) { 411 if (!E) 412 return; 413 for (const BlockDecl::Capture &C : E->getBlockDecl()->captures()) { 414 if (C.isByRef()) { 415 const VarDecl *VD = C.getVariable(); 416 markAsEscaped(VD); 417 if (isa<OMPCapturedExprDecl>(VD) || VD->isInitCapture()) 418 VisitValueDecl(VD); 419 } 420 } 421 } 422 void VisitCallExpr(const CallExpr *E) { 423 if (!E) 424 return; 425 for (const Expr *Arg : E->arguments()) { 426 if (!Arg) 427 continue; 428 if (Arg->isLValue()) { 429 const bool SavedAllEscaped = AllEscaped; 430 AllEscaped = true; 431 Visit(Arg); 432 AllEscaped = SavedAllEscaped; 433 } else { 434 Visit(Arg); 435 } 436 } 437 Visit(E->getCallee()); 438 } 439 void VisitDeclRefExpr(const DeclRefExpr *E) { 440 if (!E) 441 return; 442 const ValueDecl *VD = E->getDecl(); 443 if (AllEscaped) 444 markAsEscaped(VD); 445 if (isa<OMPCapturedExprDecl>(VD)) 446 VisitValueDecl(VD); 447 else if (const auto *VarD = dyn_cast<VarDecl>(VD)) 448 if (VarD->isInitCapture()) 449 VisitValueDecl(VD); 450 } 451 void VisitUnaryOperator(const UnaryOperator *E) { 452 if (!E) 453 return; 454 if (E->getOpcode() == UO_AddrOf) { 455 const bool SavedAllEscaped = AllEscaped; 456 AllEscaped = true; 457 Visit(E->getSubExpr()); 458 AllEscaped = SavedAllEscaped; 459 } else { 460 Visit(E->getSubExpr()); 461 } 462 } 463 void VisitImplicitCastExpr(const ImplicitCastExpr *E) { 464 if (!E) 465 return; 466 if (E->getCastKind() == CK_ArrayToPointerDecay) { 467 const bool SavedAllEscaped = AllEscaped; 468 AllEscaped = true; 469 Visit(E->getSubExpr()); 470 AllEscaped = SavedAllEscaped; 471 } else { 472 Visit(E->getSubExpr()); 473 } 474 } 475 void VisitExpr(const Expr *E) { 476 if (!E) 477 return; 478 bool SavedAllEscaped = AllEscaped; 479 if (!E->isLValue()) 480 AllEscaped = false; 481 for (const Stmt *Child : E->children()) 482 if (Child) 483 Visit(Child); 484 AllEscaped = SavedAllEscaped; 485 } 486 void VisitStmt(const Stmt *S) { 487 if (!S) 488 return; 489 for (const Stmt *Child : S->children()) 490 if (Child) 491 Visit(Child); 492 } 493 494 /// Returns the record that handles all the escaped local variables and used 495 /// instead of their original storage. 496 const RecordDecl *getGlobalizedRecord(bool IsInTTDRegion) { 497 if (!GlobalizedRD) 498 buildRecordForGlobalizedVars(IsInTTDRegion); 499 return GlobalizedRD; 500 } 501 502 /// Returns the field in the globalized record for the escaped variable. 503 const FieldDecl *getFieldForGlobalizedVar(const ValueDecl *VD) const { 504 assert(GlobalizedRD && 505 "Record for globalized variables must be generated already."); 506 auto I = MappedDeclsFields.find(VD); 507 if (I == MappedDeclsFields.end()) 508 return nullptr; 509 return I->getSecond(); 510 } 511 512 /// Returns the list of the escaped local variables/parameters. 513 ArrayRef<const ValueDecl *> getEscapedDecls() const { 514 return EscapedDecls.getArrayRef(); 515 } 516 517 /// Checks if the escaped local variable is actually a parameter passed by 518 /// value. 519 const llvm::SmallPtrSetImpl<const Decl *> &getEscapedParameters() const { 520 return EscapedParameters; 521 } 522 523 /// Returns the list of the escaped variables with the variably modified 524 /// types. 525 ArrayRef<const ValueDecl *> getEscapedVariableLengthDecls() const { 526 return EscapedVariableLengthDecls.getArrayRef(); 527 } 528 }; 529 } // anonymous namespace 530 531 /// Get the id of the warp in the block. 532 /// We assume that the warp size is 32, which is always the case 533 /// on the NVPTX device, to generate more efficient code. 534 static llvm::Value *getNVPTXWarpID(CodeGenFunction &CGF) { 535 CGBuilderTy &Bld = CGF.Builder; 536 unsigned LaneIDBits = 537 llvm::Log2_32(CGF.getTarget().getGridValue().GV_Warp_Size); 538 auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()); 539 return Bld.CreateAShr(RT.getGPUThreadID(CGF), LaneIDBits, "nvptx_warp_id"); 540 } 541 542 /// Get the id of the current lane in the Warp. 543 /// We assume that the warp size is 32, which is always the case 544 /// on the NVPTX device, to generate more efficient code. 545 static llvm::Value *getNVPTXLaneID(CodeGenFunction &CGF) { 546 CGBuilderTy &Bld = CGF.Builder; 547 unsigned LaneIDBits = 548 llvm::Log2_32(CGF.getTarget().getGridValue().GV_Warp_Size); 549 unsigned LaneIDMask = ~0u >> (32u - LaneIDBits); 550 auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()); 551 return Bld.CreateAnd(RT.getGPUThreadID(CGF), Bld.getInt32(LaneIDMask), 552 "nvptx_lane_id"); 553 } 554 555 CGOpenMPRuntimeGPU::ExecutionMode 556 CGOpenMPRuntimeGPU::getExecutionMode() const { 557 return CurrentExecutionMode; 558 } 559 560 static CGOpenMPRuntimeGPU::DataSharingMode 561 getDataSharingMode(CodeGenModule &CGM) { 562 return CGM.getLangOpts().OpenMPCUDAMode ? CGOpenMPRuntimeGPU::CUDA 563 : CGOpenMPRuntimeGPU::Generic; 564 } 565 566 /// Check for inner (nested) SPMD construct, if any 567 static bool hasNestedSPMDDirective(ASTContext &Ctx, 568 const OMPExecutableDirective &D) { 569 const auto *CS = D.getInnermostCapturedStmt(); 570 const auto *Body = 571 CS->getCapturedStmt()->IgnoreContainers(/*IgnoreCaptured=*/true); 572 const Stmt *ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body); 573 574 if (const auto *NestedDir = 575 dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) { 576 OpenMPDirectiveKind DKind = NestedDir->getDirectiveKind(); 577 switch (D.getDirectiveKind()) { 578 case OMPD_target: 579 if (isOpenMPParallelDirective(DKind)) 580 return true; 581 if (DKind == OMPD_teams) { 582 Body = NestedDir->getInnermostCapturedStmt()->IgnoreContainers( 583 /*IgnoreCaptured=*/true); 584 if (!Body) 585 return false; 586 ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body); 587 if (const auto *NND = 588 dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) { 589 DKind = NND->getDirectiveKind(); 590 if (isOpenMPParallelDirective(DKind)) 591 return true; 592 } 593 } 594 return false; 595 case OMPD_target_teams: 596 return isOpenMPParallelDirective(DKind); 597 case OMPD_target_simd: 598 case OMPD_target_parallel: 599 case OMPD_target_parallel_for: 600 case OMPD_target_parallel_for_simd: 601 case OMPD_target_teams_distribute: 602 case OMPD_target_teams_distribute_simd: 603 case OMPD_target_teams_distribute_parallel_for: 604 case OMPD_target_teams_distribute_parallel_for_simd: 605 case OMPD_parallel: 606 case OMPD_for: 607 case OMPD_parallel_for: 608 case OMPD_parallel_master: 609 case OMPD_parallel_sections: 610 case OMPD_for_simd: 611 case OMPD_parallel_for_simd: 612 case OMPD_cancel: 613 case OMPD_cancellation_point: 614 case OMPD_ordered: 615 case OMPD_threadprivate: 616 case OMPD_allocate: 617 case OMPD_task: 618 case OMPD_simd: 619 case OMPD_sections: 620 case OMPD_section: 621 case OMPD_single: 622 case OMPD_master: 623 case OMPD_critical: 624 case OMPD_taskyield: 625 case OMPD_barrier: 626 case OMPD_taskwait: 627 case OMPD_taskgroup: 628 case OMPD_atomic: 629 case OMPD_flush: 630 case OMPD_depobj: 631 case OMPD_scan: 632 case OMPD_teams: 633 case OMPD_target_data: 634 case OMPD_target_exit_data: 635 case OMPD_target_enter_data: 636 case OMPD_distribute: 637 case OMPD_distribute_simd: 638 case OMPD_distribute_parallel_for: 639 case OMPD_distribute_parallel_for_simd: 640 case OMPD_teams_distribute: 641 case OMPD_teams_distribute_simd: 642 case OMPD_teams_distribute_parallel_for: 643 case OMPD_teams_distribute_parallel_for_simd: 644 case OMPD_target_update: 645 case OMPD_declare_simd: 646 case OMPD_declare_variant: 647 case OMPD_begin_declare_variant: 648 case OMPD_end_declare_variant: 649 case OMPD_declare_target: 650 case OMPD_end_declare_target: 651 case OMPD_declare_reduction: 652 case OMPD_declare_mapper: 653 case OMPD_taskloop: 654 case OMPD_taskloop_simd: 655 case OMPD_master_taskloop: 656 case OMPD_master_taskloop_simd: 657 case OMPD_parallel_master_taskloop: 658 case OMPD_parallel_master_taskloop_simd: 659 case OMPD_requires: 660 case OMPD_unknown: 661 default: 662 llvm_unreachable("Unexpected directive."); 663 } 664 } 665 666 return false; 667 } 668 669 static bool supportsSPMDExecutionMode(ASTContext &Ctx, 670 const OMPExecutableDirective &D) { 671 OpenMPDirectiveKind DirectiveKind = D.getDirectiveKind(); 672 switch (DirectiveKind) { 673 case OMPD_target: 674 case OMPD_target_teams: 675 return hasNestedSPMDDirective(Ctx, D); 676 case OMPD_target_parallel: 677 case OMPD_target_parallel_for: 678 case OMPD_target_parallel_for_simd: 679 case OMPD_target_teams_distribute_parallel_for: 680 case OMPD_target_teams_distribute_parallel_for_simd: 681 case OMPD_target_simd: 682 case OMPD_target_teams_distribute_simd: 683 return true; 684 case OMPD_target_teams_distribute: 685 return false; 686 case OMPD_parallel: 687 case OMPD_for: 688 case OMPD_parallel_for: 689 case OMPD_parallel_master: 690 case OMPD_parallel_sections: 691 case OMPD_for_simd: 692 case OMPD_parallel_for_simd: 693 case OMPD_cancel: 694 case OMPD_cancellation_point: 695 case OMPD_ordered: 696 case OMPD_threadprivate: 697 case OMPD_allocate: 698 case OMPD_task: 699 case OMPD_simd: 700 case OMPD_sections: 701 case OMPD_section: 702 case OMPD_single: 703 case OMPD_master: 704 case OMPD_critical: 705 case OMPD_taskyield: 706 case OMPD_barrier: 707 case OMPD_taskwait: 708 case OMPD_taskgroup: 709 case OMPD_atomic: 710 case OMPD_flush: 711 case OMPD_depobj: 712 case OMPD_scan: 713 case OMPD_teams: 714 case OMPD_target_data: 715 case OMPD_target_exit_data: 716 case OMPD_target_enter_data: 717 case OMPD_distribute: 718 case OMPD_distribute_simd: 719 case OMPD_distribute_parallel_for: 720 case OMPD_distribute_parallel_for_simd: 721 case OMPD_teams_distribute: 722 case OMPD_teams_distribute_simd: 723 case OMPD_teams_distribute_parallel_for: 724 case OMPD_teams_distribute_parallel_for_simd: 725 case OMPD_target_update: 726 case OMPD_declare_simd: 727 case OMPD_declare_variant: 728 case OMPD_begin_declare_variant: 729 case OMPD_end_declare_variant: 730 case OMPD_declare_target: 731 case OMPD_end_declare_target: 732 case OMPD_declare_reduction: 733 case OMPD_declare_mapper: 734 case OMPD_taskloop: 735 case OMPD_taskloop_simd: 736 case OMPD_master_taskloop: 737 case OMPD_master_taskloop_simd: 738 case OMPD_parallel_master_taskloop: 739 case OMPD_parallel_master_taskloop_simd: 740 case OMPD_requires: 741 case OMPD_unknown: 742 default: 743 break; 744 } 745 llvm_unreachable( 746 "Unknown programming model for OpenMP directive on NVPTX target."); 747 } 748 749 /// Check if the directive is loops based and has schedule clause at all or has 750 /// static scheduling. 751 static bool hasStaticScheduling(const OMPExecutableDirective &D) { 752 assert(isOpenMPWorksharingDirective(D.getDirectiveKind()) && 753 isOpenMPLoopDirective(D.getDirectiveKind()) && 754 "Expected loop-based directive."); 755 return !D.hasClausesOfKind<OMPOrderedClause>() && 756 (!D.hasClausesOfKind<OMPScheduleClause>() || 757 llvm::any_of(D.getClausesOfKind<OMPScheduleClause>(), 758 [](const OMPScheduleClause *C) { 759 return C->getScheduleKind() == OMPC_SCHEDULE_static; 760 })); 761 } 762 763 /// Check for inner (nested) lightweight runtime construct, if any 764 static bool hasNestedLightweightDirective(ASTContext &Ctx, 765 const OMPExecutableDirective &D) { 766 assert(supportsSPMDExecutionMode(Ctx, D) && "Expected SPMD mode directive."); 767 const auto *CS = D.getInnermostCapturedStmt(); 768 const auto *Body = 769 CS->getCapturedStmt()->IgnoreContainers(/*IgnoreCaptured=*/true); 770 const Stmt *ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body); 771 772 if (const auto *NestedDir = 773 dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) { 774 OpenMPDirectiveKind DKind = NestedDir->getDirectiveKind(); 775 switch (D.getDirectiveKind()) { 776 case OMPD_target: 777 if (isOpenMPParallelDirective(DKind) && 778 isOpenMPWorksharingDirective(DKind) && isOpenMPLoopDirective(DKind) && 779 hasStaticScheduling(*NestedDir)) 780 return true; 781 if (DKind == OMPD_teams_distribute_simd || DKind == OMPD_simd) 782 return true; 783 if (DKind == OMPD_parallel) { 784 Body = NestedDir->getInnermostCapturedStmt()->IgnoreContainers( 785 /*IgnoreCaptured=*/true); 786 if (!Body) 787 return false; 788 ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body); 789 if (const auto *NND = 790 dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) { 791 DKind = NND->getDirectiveKind(); 792 if (isOpenMPWorksharingDirective(DKind) && 793 isOpenMPLoopDirective(DKind) && hasStaticScheduling(*NND)) 794 return true; 795 } 796 } else if (DKind == OMPD_teams) { 797 Body = NestedDir->getInnermostCapturedStmt()->IgnoreContainers( 798 /*IgnoreCaptured=*/true); 799 if (!Body) 800 return false; 801 ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body); 802 if (const auto *NND = 803 dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) { 804 DKind = NND->getDirectiveKind(); 805 if (isOpenMPParallelDirective(DKind) && 806 isOpenMPWorksharingDirective(DKind) && 807 isOpenMPLoopDirective(DKind) && hasStaticScheduling(*NND)) 808 return true; 809 if (DKind == OMPD_parallel) { 810 Body = NND->getInnermostCapturedStmt()->IgnoreContainers( 811 /*IgnoreCaptured=*/true); 812 if (!Body) 813 return false; 814 ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body); 815 if (const auto *NND = 816 dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) { 817 DKind = NND->getDirectiveKind(); 818 if (isOpenMPWorksharingDirective(DKind) && 819 isOpenMPLoopDirective(DKind) && hasStaticScheduling(*NND)) 820 return true; 821 } 822 } 823 } 824 } 825 return false; 826 case OMPD_target_teams: 827 if (isOpenMPParallelDirective(DKind) && 828 isOpenMPWorksharingDirective(DKind) && isOpenMPLoopDirective(DKind) && 829 hasStaticScheduling(*NestedDir)) 830 return true; 831 if (DKind == OMPD_distribute_simd || DKind == OMPD_simd) 832 return true; 833 if (DKind == OMPD_parallel) { 834 Body = NestedDir->getInnermostCapturedStmt()->IgnoreContainers( 835 /*IgnoreCaptured=*/true); 836 if (!Body) 837 return false; 838 ChildStmt = CGOpenMPRuntime::getSingleCompoundChild(Ctx, Body); 839 if (const auto *NND = 840 dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) { 841 DKind = NND->getDirectiveKind(); 842 if (isOpenMPWorksharingDirective(DKind) && 843 isOpenMPLoopDirective(DKind) && hasStaticScheduling(*NND)) 844 return true; 845 } 846 } 847 return false; 848 case OMPD_target_parallel: 849 if (DKind == OMPD_simd) 850 return true; 851 return isOpenMPWorksharingDirective(DKind) && 852 isOpenMPLoopDirective(DKind) && hasStaticScheduling(*NestedDir); 853 case OMPD_target_teams_distribute: 854 case OMPD_target_simd: 855 case OMPD_target_parallel_for: 856 case OMPD_target_parallel_for_simd: 857 case OMPD_target_teams_distribute_simd: 858 case OMPD_target_teams_distribute_parallel_for: 859 case OMPD_target_teams_distribute_parallel_for_simd: 860 case OMPD_parallel: 861 case OMPD_for: 862 case OMPD_parallel_for: 863 case OMPD_parallel_master: 864 case OMPD_parallel_sections: 865 case OMPD_for_simd: 866 case OMPD_parallel_for_simd: 867 case OMPD_cancel: 868 case OMPD_cancellation_point: 869 case OMPD_ordered: 870 case OMPD_threadprivate: 871 case OMPD_allocate: 872 case OMPD_task: 873 case OMPD_simd: 874 case OMPD_sections: 875 case OMPD_section: 876 case OMPD_single: 877 case OMPD_master: 878 case OMPD_critical: 879 case OMPD_taskyield: 880 case OMPD_barrier: 881 case OMPD_taskwait: 882 case OMPD_taskgroup: 883 case OMPD_atomic: 884 case OMPD_flush: 885 case OMPD_depobj: 886 case OMPD_scan: 887 case OMPD_teams: 888 case OMPD_target_data: 889 case OMPD_target_exit_data: 890 case OMPD_target_enter_data: 891 case OMPD_distribute: 892 case OMPD_distribute_simd: 893 case OMPD_distribute_parallel_for: 894 case OMPD_distribute_parallel_for_simd: 895 case OMPD_teams_distribute: 896 case OMPD_teams_distribute_simd: 897 case OMPD_teams_distribute_parallel_for: 898 case OMPD_teams_distribute_parallel_for_simd: 899 case OMPD_target_update: 900 case OMPD_declare_simd: 901 case OMPD_declare_variant: 902 case OMPD_begin_declare_variant: 903 case OMPD_end_declare_variant: 904 case OMPD_declare_target: 905 case OMPD_end_declare_target: 906 case OMPD_declare_reduction: 907 case OMPD_declare_mapper: 908 case OMPD_taskloop: 909 case OMPD_taskloop_simd: 910 case OMPD_master_taskloop: 911 case OMPD_master_taskloop_simd: 912 case OMPD_parallel_master_taskloop: 913 case OMPD_parallel_master_taskloop_simd: 914 case OMPD_requires: 915 case OMPD_unknown: 916 default: 917 llvm_unreachable("Unexpected directive."); 918 } 919 } 920 921 return false; 922 } 923 924 /// Checks if the construct supports lightweight runtime. It must be SPMD 925 /// construct + inner loop-based construct with static scheduling. 926 static bool supportsLightweightRuntime(ASTContext &Ctx, 927 const OMPExecutableDirective &D) { 928 if (!supportsSPMDExecutionMode(Ctx, D)) 929 return false; 930 OpenMPDirectiveKind DirectiveKind = D.getDirectiveKind(); 931 switch (DirectiveKind) { 932 case OMPD_target: 933 case OMPD_target_teams: 934 case OMPD_target_parallel: 935 return hasNestedLightweightDirective(Ctx, D); 936 case OMPD_target_parallel_for: 937 case OMPD_target_parallel_for_simd: 938 case OMPD_target_teams_distribute_parallel_for: 939 case OMPD_target_teams_distribute_parallel_for_simd: 940 // (Last|First)-privates must be shared in parallel region. 941 return hasStaticScheduling(D); 942 case OMPD_target_simd: 943 case OMPD_target_teams_distribute_simd: 944 return true; 945 case OMPD_target_teams_distribute: 946 return false; 947 case OMPD_parallel: 948 case OMPD_for: 949 case OMPD_parallel_for: 950 case OMPD_parallel_master: 951 case OMPD_parallel_sections: 952 case OMPD_for_simd: 953 case OMPD_parallel_for_simd: 954 case OMPD_cancel: 955 case OMPD_cancellation_point: 956 case OMPD_ordered: 957 case OMPD_threadprivate: 958 case OMPD_allocate: 959 case OMPD_task: 960 case OMPD_simd: 961 case OMPD_sections: 962 case OMPD_section: 963 case OMPD_single: 964 case OMPD_master: 965 case OMPD_critical: 966 case OMPD_taskyield: 967 case OMPD_barrier: 968 case OMPD_taskwait: 969 case OMPD_taskgroup: 970 case OMPD_atomic: 971 case OMPD_flush: 972 case OMPD_depobj: 973 case OMPD_scan: 974 case OMPD_teams: 975 case OMPD_target_data: 976 case OMPD_target_exit_data: 977 case OMPD_target_enter_data: 978 case OMPD_distribute: 979 case OMPD_distribute_simd: 980 case OMPD_distribute_parallel_for: 981 case OMPD_distribute_parallel_for_simd: 982 case OMPD_teams_distribute: 983 case OMPD_teams_distribute_simd: 984 case OMPD_teams_distribute_parallel_for: 985 case OMPD_teams_distribute_parallel_for_simd: 986 case OMPD_target_update: 987 case OMPD_declare_simd: 988 case OMPD_declare_variant: 989 case OMPD_begin_declare_variant: 990 case OMPD_end_declare_variant: 991 case OMPD_declare_target: 992 case OMPD_end_declare_target: 993 case OMPD_declare_reduction: 994 case OMPD_declare_mapper: 995 case OMPD_taskloop: 996 case OMPD_taskloop_simd: 997 case OMPD_master_taskloop: 998 case OMPD_master_taskloop_simd: 999 case OMPD_parallel_master_taskloop: 1000 case OMPD_parallel_master_taskloop_simd: 1001 case OMPD_requires: 1002 case OMPD_unknown: 1003 default: 1004 break; 1005 } 1006 llvm_unreachable( 1007 "Unknown programming model for OpenMP directive on NVPTX target."); 1008 } 1009 1010 void CGOpenMPRuntimeGPU::emitNonSPMDKernel(const OMPExecutableDirective &D, 1011 StringRef ParentName, 1012 llvm::Function *&OutlinedFn, 1013 llvm::Constant *&OutlinedFnID, 1014 bool IsOffloadEntry, 1015 const RegionCodeGenTy &CodeGen) { 1016 ExecutionRuntimeModesRAII ModeRAII(CurrentExecutionMode); 1017 EntryFunctionState EST; 1018 WrapperFunctionsMap.clear(); 1019 1020 // Emit target region as a standalone region. 1021 class NVPTXPrePostActionTy : public PrePostActionTy { 1022 CGOpenMPRuntimeGPU::EntryFunctionState &EST; 1023 1024 public: 1025 NVPTXPrePostActionTy(CGOpenMPRuntimeGPU::EntryFunctionState &EST) 1026 : EST(EST) {} 1027 void Enter(CodeGenFunction &CGF) override { 1028 auto &RT = 1029 static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()); 1030 RT.emitKernelInit(CGF, EST, /* IsSPMD */ false); 1031 // Skip target region initialization. 1032 RT.setLocThreadIdInsertPt(CGF, /*AtCurrentPoint=*/true); 1033 } 1034 void Exit(CodeGenFunction &CGF) override { 1035 auto &RT = 1036 static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()); 1037 RT.clearLocThreadIdInsertPt(CGF); 1038 RT.emitKernelDeinit(CGF, EST, /* IsSPMD */ false); 1039 } 1040 } Action(EST); 1041 CodeGen.setAction(Action); 1042 IsInTTDRegion = true; 1043 emitTargetOutlinedFunctionHelper(D, ParentName, OutlinedFn, OutlinedFnID, 1044 IsOffloadEntry, CodeGen); 1045 IsInTTDRegion = false; 1046 } 1047 1048 void CGOpenMPRuntimeGPU::emitKernelInit(CodeGenFunction &CGF, 1049 EntryFunctionState &EST, bool IsSPMD) { 1050 CGBuilderTy &Bld = CGF.Builder; 1051 Bld.restoreIP(OMPBuilder.createTargetInit(Bld, IsSPMD, requiresFullRuntime())); 1052 IsInTargetMasterThreadRegion = IsSPMD; 1053 if (!IsSPMD) 1054 emitGenericVarsProlog(CGF, EST.Loc); 1055 } 1056 1057 void CGOpenMPRuntimeGPU::emitKernelDeinit(CodeGenFunction &CGF, 1058 EntryFunctionState &EST, 1059 bool IsSPMD) { 1060 if (!IsSPMD) 1061 emitGenericVarsEpilog(CGF); 1062 1063 CGBuilderTy &Bld = CGF.Builder; 1064 OMPBuilder.createTargetDeinit(Bld, IsSPMD, requiresFullRuntime()); 1065 } 1066 1067 void CGOpenMPRuntimeGPU::emitSPMDKernel(const OMPExecutableDirective &D, 1068 StringRef ParentName, 1069 llvm::Function *&OutlinedFn, 1070 llvm::Constant *&OutlinedFnID, 1071 bool IsOffloadEntry, 1072 const RegionCodeGenTy &CodeGen) { 1073 ExecutionRuntimeModesRAII ModeRAII( 1074 CurrentExecutionMode, RequiresFullRuntime, 1075 CGM.getLangOpts().OpenMPCUDAForceFullRuntime || 1076 !supportsLightweightRuntime(CGM.getContext(), D)); 1077 EntryFunctionState EST; 1078 1079 // Emit target region as a standalone region. 1080 class NVPTXPrePostActionTy : public PrePostActionTy { 1081 CGOpenMPRuntimeGPU &RT; 1082 CGOpenMPRuntimeGPU::EntryFunctionState &EST; 1083 1084 public: 1085 NVPTXPrePostActionTy(CGOpenMPRuntimeGPU &RT, 1086 CGOpenMPRuntimeGPU::EntryFunctionState &EST) 1087 : RT(RT), EST(EST) {} 1088 void Enter(CodeGenFunction &CGF) override { 1089 RT.emitKernelInit(CGF, EST, /* IsSPMD */ true); 1090 // Skip target region initialization. 1091 RT.setLocThreadIdInsertPt(CGF, /*AtCurrentPoint=*/true); 1092 } 1093 void Exit(CodeGenFunction &CGF) override { 1094 RT.clearLocThreadIdInsertPt(CGF); 1095 RT.emitKernelDeinit(CGF, EST, /* IsSPMD */ true); 1096 } 1097 } Action(*this, EST); 1098 CodeGen.setAction(Action); 1099 IsInTTDRegion = true; 1100 emitTargetOutlinedFunctionHelper(D, ParentName, OutlinedFn, OutlinedFnID, 1101 IsOffloadEntry, CodeGen); 1102 IsInTTDRegion = false; 1103 } 1104 1105 // Create a unique global variable to indicate the execution mode of this target 1106 // region. The execution mode is either 'generic', or 'spmd' depending on the 1107 // target directive. This variable is picked up by the offload library to setup 1108 // the device appropriately before kernel launch. If the execution mode is 1109 // 'generic', the runtime reserves one warp for the master, otherwise, all 1110 // warps participate in parallel work. 1111 static void setPropertyExecutionMode(CodeGenModule &CGM, StringRef Name, 1112 bool Mode) { 1113 auto *GVMode = new llvm::GlobalVariable( 1114 CGM.getModule(), CGM.Int8Ty, /*isConstant=*/true, 1115 llvm::GlobalValue::WeakAnyLinkage, 1116 llvm::ConstantInt::get(CGM.Int8Ty, Mode ? OMP_TGT_EXEC_MODE_SPMD 1117 : OMP_TGT_EXEC_MODE_GENERIC), 1118 Twine(Name, "_exec_mode")); 1119 CGM.addCompilerUsedGlobal(GVMode); 1120 } 1121 1122 void CGOpenMPRuntimeGPU::createOffloadEntry(llvm::Constant *ID, 1123 llvm::Constant *Addr, 1124 uint64_t Size, int32_t, 1125 llvm::GlobalValue::LinkageTypes) { 1126 // TODO: Add support for global variables on the device after declare target 1127 // support. 1128 llvm::Function *Fn = dyn_cast<llvm::Function>(Addr); 1129 if (!Fn) 1130 return; 1131 1132 llvm::Module &M = CGM.getModule(); 1133 llvm::LLVMContext &Ctx = CGM.getLLVMContext(); 1134 1135 // Get "nvvm.annotations" metadata node. 1136 llvm::NamedMDNode *MD = M.getOrInsertNamedMetadata("nvvm.annotations"); 1137 1138 llvm::Metadata *MDVals[] = { 1139 llvm::ConstantAsMetadata::get(Fn), llvm::MDString::get(Ctx, "kernel"), 1140 llvm::ConstantAsMetadata::get( 1141 llvm::ConstantInt::get(llvm::Type::getInt32Ty(Ctx), 1))}; 1142 // Append metadata to nvvm.annotations. 1143 MD->addOperand(llvm::MDNode::get(Ctx, MDVals)); 1144 1145 // Add a function attribute for the kernel. 1146 Fn->addFnAttr(llvm::Attribute::get(Ctx, "kernel")); 1147 } 1148 1149 void CGOpenMPRuntimeGPU::emitTargetOutlinedFunction( 1150 const OMPExecutableDirective &D, StringRef ParentName, 1151 llvm::Function *&OutlinedFn, llvm::Constant *&OutlinedFnID, 1152 bool IsOffloadEntry, const RegionCodeGenTy &CodeGen) { 1153 if (!IsOffloadEntry) // Nothing to do. 1154 return; 1155 1156 assert(!ParentName.empty() && "Invalid target region parent name!"); 1157 1158 bool Mode = supportsSPMDExecutionMode(CGM.getContext(), D); 1159 if (Mode) 1160 emitSPMDKernel(D, ParentName, OutlinedFn, OutlinedFnID, IsOffloadEntry, 1161 CodeGen); 1162 else 1163 emitNonSPMDKernel(D, ParentName, OutlinedFn, OutlinedFnID, IsOffloadEntry, 1164 CodeGen); 1165 1166 setPropertyExecutionMode(CGM, OutlinedFn->getName(), Mode); 1167 } 1168 1169 namespace { 1170 LLVM_ENABLE_BITMASK_ENUMS_IN_NAMESPACE(); 1171 /// Enum for accesseing the reserved_2 field of the ident_t struct. 1172 enum ModeFlagsTy : unsigned { 1173 /// Bit set to 1 when in SPMD mode. 1174 KMP_IDENT_SPMD_MODE = 0x01, 1175 /// Bit set to 1 when a simplified runtime is used. 1176 KMP_IDENT_SIMPLE_RT_MODE = 0x02, 1177 LLVM_MARK_AS_BITMASK_ENUM(/*LargestValue=*/KMP_IDENT_SIMPLE_RT_MODE) 1178 }; 1179 1180 /// Special mode Undefined. Is the combination of Non-SPMD mode + SimpleRuntime. 1181 static const ModeFlagsTy UndefinedMode = 1182 (~KMP_IDENT_SPMD_MODE) & KMP_IDENT_SIMPLE_RT_MODE; 1183 } // anonymous namespace 1184 1185 unsigned CGOpenMPRuntimeGPU::getDefaultLocationReserved2Flags() const { 1186 switch (getExecutionMode()) { 1187 case EM_SPMD: 1188 if (requiresFullRuntime()) 1189 return KMP_IDENT_SPMD_MODE & (~KMP_IDENT_SIMPLE_RT_MODE); 1190 return KMP_IDENT_SPMD_MODE | KMP_IDENT_SIMPLE_RT_MODE; 1191 case EM_NonSPMD: 1192 assert(requiresFullRuntime() && "Expected full runtime."); 1193 return (~KMP_IDENT_SPMD_MODE) & (~KMP_IDENT_SIMPLE_RT_MODE); 1194 case EM_Unknown: 1195 return UndefinedMode; 1196 } 1197 llvm_unreachable("Unknown flags are requested."); 1198 } 1199 1200 CGOpenMPRuntimeGPU::CGOpenMPRuntimeGPU(CodeGenModule &CGM) 1201 : CGOpenMPRuntime(CGM, "_", "$") { 1202 if (!CGM.getLangOpts().OpenMPIsDevice) 1203 llvm_unreachable("OpenMP can only handle device code."); 1204 1205 llvm::OpenMPIRBuilder &OMPBuilder = getOMPBuilder(); 1206 if (CGM.getLangOpts().NoGPULib || CGM.getLangOpts().OMPHostIRFile.empty()) 1207 return; 1208 1209 OMPBuilder.createGlobalFlag(CGM.getLangOpts().OpenMPTargetDebug, 1210 "__omp_rtl_debug_kind"); 1211 OMPBuilder.createGlobalFlag(CGM.getLangOpts().OpenMPTeamSubscription, 1212 "__omp_rtl_assume_teams_oversubscription"); 1213 OMPBuilder.createGlobalFlag(CGM.getLangOpts().OpenMPThreadSubscription, 1214 "__omp_rtl_assume_threads_oversubscription"); 1215 OMPBuilder.createGlobalFlag(CGM.getLangOpts().OpenMPNoThreadState, 1216 "__omp_rtl_assume_no_thread_state"); 1217 } 1218 1219 void CGOpenMPRuntimeGPU::emitProcBindClause(CodeGenFunction &CGF, 1220 ProcBindKind ProcBind, 1221 SourceLocation Loc) { 1222 // Do nothing in case of SPMD mode and L0 parallel. 1223 if (getExecutionMode() == CGOpenMPRuntimeGPU::EM_SPMD) 1224 return; 1225 1226 CGOpenMPRuntime::emitProcBindClause(CGF, ProcBind, Loc); 1227 } 1228 1229 void CGOpenMPRuntimeGPU::emitNumThreadsClause(CodeGenFunction &CGF, 1230 llvm::Value *NumThreads, 1231 SourceLocation Loc) { 1232 // Nothing to do. 1233 } 1234 1235 void CGOpenMPRuntimeGPU::emitNumTeamsClause(CodeGenFunction &CGF, 1236 const Expr *NumTeams, 1237 const Expr *ThreadLimit, 1238 SourceLocation Loc) {} 1239 1240 llvm::Function *CGOpenMPRuntimeGPU::emitParallelOutlinedFunction( 1241 const OMPExecutableDirective &D, const VarDecl *ThreadIDVar, 1242 OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen) { 1243 // Emit target region as a standalone region. 1244 class NVPTXPrePostActionTy : public PrePostActionTy { 1245 bool &IsInParallelRegion; 1246 bool PrevIsInParallelRegion; 1247 1248 public: 1249 NVPTXPrePostActionTy(bool &IsInParallelRegion) 1250 : IsInParallelRegion(IsInParallelRegion) {} 1251 void Enter(CodeGenFunction &CGF) override { 1252 PrevIsInParallelRegion = IsInParallelRegion; 1253 IsInParallelRegion = true; 1254 } 1255 void Exit(CodeGenFunction &CGF) override { 1256 IsInParallelRegion = PrevIsInParallelRegion; 1257 } 1258 } Action(IsInParallelRegion); 1259 CodeGen.setAction(Action); 1260 bool PrevIsInTTDRegion = IsInTTDRegion; 1261 IsInTTDRegion = false; 1262 bool PrevIsInTargetMasterThreadRegion = IsInTargetMasterThreadRegion; 1263 IsInTargetMasterThreadRegion = false; 1264 auto *OutlinedFun = 1265 cast<llvm::Function>(CGOpenMPRuntime::emitParallelOutlinedFunction( 1266 D, ThreadIDVar, InnermostKind, CodeGen)); 1267 IsInTargetMasterThreadRegion = PrevIsInTargetMasterThreadRegion; 1268 IsInTTDRegion = PrevIsInTTDRegion; 1269 if (getExecutionMode() != CGOpenMPRuntimeGPU::EM_SPMD && 1270 !IsInParallelRegion) { 1271 llvm::Function *WrapperFun = 1272 createParallelDataSharingWrapper(OutlinedFun, D); 1273 WrapperFunctionsMap[OutlinedFun] = WrapperFun; 1274 } 1275 1276 return OutlinedFun; 1277 } 1278 1279 /// Get list of lastprivate variables from the teams distribute ... or 1280 /// teams {distribute ...} directives. 1281 static void 1282 getDistributeLastprivateVars(ASTContext &Ctx, const OMPExecutableDirective &D, 1283 llvm::SmallVectorImpl<const ValueDecl *> &Vars) { 1284 assert(isOpenMPTeamsDirective(D.getDirectiveKind()) && 1285 "expected teams directive."); 1286 const OMPExecutableDirective *Dir = &D; 1287 if (!isOpenMPDistributeDirective(D.getDirectiveKind())) { 1288 if (const Stmt *S = CGOpenMPRuntime::getSingleCompoundChild( 1289 Ctx, 1290 D.getInnermostCapturedStmt()->getCapturedStmt()->IgnoreContainers( 1291 /*IgnoreCaptured=*/true))) { 1292 Dir = dyn_cast_or_null<OMPExecutableDirective>(S); 1293 if (Dir && !isOpenMPDistributeDirective(Dir->getDirectiveKind())) 1294 Dir = nullptr; 1295 } 1296 } 1297 if (!Dir) 1298 return; 1299 for (const auto *C : Dir->getClausesOfKind<OMPLastprivateClause>()) { 1300 for (const Expr *E : C->getVarRefs()) 1301 Vars.push_back(getPrivateItem(E)); 1302 } 1303 } 1304 1305 /// Get list of reduction variables from the teams ... directives. 1306 static void 1307 getTeamsReductionVars(ASTContext &Ctx, const OMPExecutableDirective &D, 1308 llvm::SmallVectorImpl<const ValueDecl *> &Vars) { 1309 assert(isOpenMPTeamsDirective(D.getDirectiveKind()) && 1310 "expected teams directive."); 1311 for (const auto *C : D.getClausesOfKind<OMPReductionClause>()) { 1312 for (const Expr *E : C->privates()) 1313 Vars.push_back(getPrivateItem(E)); 1314 } 1315 } 1316 1317 llvm::Function *CGOpenMPRuntimeGPU::emitTeamsOutlinedFunction( 1318 const OMPExecutableDirective &D, const VarDecl *ThreadIDVar, 1319 OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen) { 1320 SourceLocation Loc = D.getBeginLoc(); 1321 1322 const RecordDecl *GlobalizedRD = nullptr; 1323 llvm::SmallVector<const ValueDecl *, 4> LastPrivatesReductions; 1324 llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> MappedDeclsFields; 1325 unsigned WarpSize = CGM.getTarget().getGridValue().GV_Warp_Size; 1326 // Globalize team reductions variable unconditionally in all modes. 1327 if (getExecutionMode() != CGOpenMPRuntimeGPU::EM_SPMD) 1328 getTeamsReductionVars(CGM.getContext(), D, LastPrivatesReductions); 1329 if (getExecutionMode() == CGOpenMPRuntimeGPU::EM_SPMD) { 1330 getDistributeLastprivateVars(CGM.getContext(), D, LastPrivatesReductions); 1331 if (!LastPrivatesReductions.empty()) { 1332 GlobalizedRD = ::buildRecordForGlobalizedVars( 1333 CGM.getContext(), llvm::None, LastPrivatesReductions, 1334 MappedDeclsFields, WarpSize); 1335 } 1336 } else if (!LastPrivatesReductions.empty()) { 1337 assert(!TeamAndReductions.first && 1338 "Previous team declaration is not expected."); 1339 TeamAndReductions.first = D.getCapturedStmt(OMPD_teams)->getCapturedDecl(); 1340 std::swap(TeamAndReductions.second, LastPrivatesReductions); 1341 } 1342 1343 // Emit target region as a standalone region. 1344 class NVPTXPrePostActionTy : public PrePostActionTy { 1345 SourceLocation &Loc; 1346 const RecordDecl *GlobalizedRD; 1347 llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> 1348 &MappedDeclsFields; 1349 1350 public: 1351 NVPTXPrePostActionTy( 1352 SourceLocation &Loc, const RecordDecl *GlobalizedRD, 1353 llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> 1354 &MappedDeclsFields) 1355 : Loc(Loc), GlobalizedRD(GlobalizedRD), 1356 MappedDeclsFields(MappedDeclsFields) {} 1357 void Enter(CodeGenFunction &CGF) override { 1358 auto &Rt = 1359 static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()); 1360 if (GlobalizedRD) { 1361 auto I = Rt.FunctionGlobalizedDecls.try_emplace(CGF.CurFn).first; 1362 I->getSecond().MappedParams = 1363 std::make_unique<CodeGenFunction::OMPMapVars>(); 1364 DeclToAddrMapTy &Data = I->getSecond().LocalVarData; 1365 for (const auto &Pair : MappedDeclsFields) { 1366 assert(Pair.getFirst()->isCanonicalDecl() && 1367 "Expected canonical declaration"); 1368 Data.insert(std::make_pair(Pair.getFirst(), MappedVarData())); 1369 } 1370 } 1371 Rt.emitGenericVarsProlog(CGF, Loc); 1372 } 1373 void Exit(CodeGenFunction &CGF) override { 1374 static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()) 1375 .emitGenericVarsEpilog(CGF); 1376 } 1377 } Action(Loc, GlobalizedRD, MappedDeclsFields); 1378 CodeGen.setAction(Action); 1379 llvm::Function *OutlinedFun = CGOpenMPRuntime::emitTeamsOutlinedFunction( 1380 D, ThreadIDVar, InnermostKind, CodeGen); 1381 1382 return OutlinedFun; 1383 } 1384 1385 void CGOpenMPRuntimeGPU::emitGenericVarsProlog(CodeGenFunction &CGF, 1386 SourceLocation Loc, 1387 bool WithSPMDCheck) { 1388 if (getDataSharingMode(CGM) != CGOpenMPRuntimeGPU::Generic && 1389 getExecutionMode() != CGOpenMPRuntimeGPU::EM_SPMD) 1390 return; 1391 1392 CGBuilderTy &Bld = CGF.Builder; 1393 1394 const auto I = FunctionGlobalizedDecls.find(CGF.CurFn); 1395 if (I == FunctionGlobalizedDecls.end()) 1396 return; 1397 1398 for (auto &Rec : I->getSecond().LocalVarData) { 1399 const auto *VD = cast<VarDecl>(Rec.first); 1400 bool EscapedParam = I->getSecond().EscapedParameters.count(Rec.first); 1401 QualType VarTy = VD->getType(); 1402 1403 // Get the local allocation of a firstprivate variable before sharing 1404 llvm::Value *ParValue; 1405 if (EscapedParam) { 1406 LValue ParLVal = 1407 CGF.MakeAddrLValue(CGF.GetAddrOfLocalVar(VD), VD->getType()); 1408 ParValue = CGF.EmitLoadOfScalar(ParLVal, Loc); 1409 } 1410 1411 // Allocate space for the variable to be globalized 1412 llvm::Value *AllocArgs[] = {CGF.getTypeSize(VD->getType())}; 1413 llvm::CallBase *VoidPtr = 1414 CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( 1415 CGM.getModule(), OMPRTL___kmpc_alloc_shared), 1416 AllocArgs, VD->getName()); 1417 // FIXME: We should use the variables actual alignment as an argument. 1418 VoidPtr->addRetAttr(llvm::Attribute::get( 1419 CGM.getLLVMContext(), llvm::Attribute::Alignment, 1420 CGM.getContext().getTargetInfo().getNewAlign() / 8)); 1421 1422 // Cast the void pointer and get the address of the globalized variable. 1423 llvm::PointerType *VarPtrTy = CGF.ConvertTypeForMem(VarTy)->getPointerTo(); 1424 llvm::Value *CastedVoidPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( 1425 VoidPtr, VarPtrTy, VD->getName() + "_on_stack"); 1426 LValue VarAddr = CGF.MakeNaturalAlignAddrLValue(CastedVoidPtr, VarTy); 1427 Rec.second.PrivateAddr = VarAddr.getAddress(CGF); 1428 Rec.second.GlobalizedVal = VoidPtr; 1429 1430 // Assign the local allocation to the newly globalized location. 1431 if (EscapedParam) { 1432 CGF.EmitStoreOfScalar(ParValue, VarAddr); 1433 I->getSecond().MappedParams->setVarAddr(CGF, VD, VarAddr.getAddress(CGF)); 1434 } 1435 if (auto *DI = CGF.getDebugInfo()) 1436 VoidPtr->setDebugLoc(DI->SourceLocToDebugLoc(VD->getLocation())); 1437 } 1438 for (const auto *VD : I->getSecond().EscapedVariableLengthDecls) { 1439 // Use actual memory size of the VLA object including the padding 1440 // for alignment purposes. 1441 llvm::Value *Size = CGF.getTypeSize(VD->getType()); 1442 CharUnits Align = CGM.getContext().getDeclAlign(VD); 1443 Size = Bld.CreateNUWAdd( 1444 Size, llvm::ConstantInt::get(CGF.SizeTy, Align.getQuantity() - 1)); 1445 llvm::Value *AlignVal = 1446 llvm::ConstantInt::get(CGF.SizeTy, Align.getQuantity()); 1447 1448 Size = Bld.CreateUDiv(Size, AlignVal); 1449 Size = Bld.CreateNUWMul(Size, AlignVal); 1450 1451 // Allocate space for this VLA object to be globalized. 1452 llvm::Value *AllocArgs[] = {CGF.getTypeSize(VD->getType())}; 1453 llvm::CallBase *VoidPtr = 1454 CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( 1455 CGM.getModule(), OMPRTL___kmpc_alloc_shared), 1456 AllocArgs, VD->getName()); 1457 VoidPtr->addRetAttr( 1458 llvm::Attribute::get(CGM.getLLVMContext(), llvm::Attribute::Alignment, 1459 CGM.getContext().getTargetInfo().getNewAlign())); 1460 1461 I->getSecond().EscapedVariableLengthDeclsAddrs.emplace_back( 1462 std::pair<llvm::Value *, llvm::Value *>( 1463 {VoidPtr, CGF.getTypeSize(VD->getType())})); 1464 LValue Base = CGF.MakeAddrLValue(VoidPtr, VD->getType(), 1465 CGM.getContext().getDeclAlign(VD), 1466 AlignmentSource::Decl); 1467 I->getSecond().MappedParams->setVarAddr(CGF, cast<VarDecl>(VD), 1468 Base.getAddress(CGF)); 1469 } 1470 I->getSecond().MappedParams->apply(CGF); 1471 } 1472 1473 void CGOpenMPRuntimeGPU::emitGenericVarsEpilog(CodeGenFunction &CGF, 1474 bool WithSPMDCheck) { 1475 if (getDataSharingMode(CGM) != CGOpenMPRuntimeGPU::Generic && 1476 getExecutionMode() != CGOpenMPRuntimeGPU::EM_SPMD) 1477 return; 1478 1479 const auto I = FunctionGlobalizedDecls.find(CGF.CurFn); 1480 if (I != FunctionGlobalizedDecls.end()) { 1481 // Deallocate the memory for each globalized VLA object 1482 for (auto AddrSizePair : 1483 llvm::reverse(I->getSecond().EscapedVariableLengthDeclsAddrs)) { 1484 CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( 1485 CGM.getModule(), OMPRTL___kmpc_free_shared), 1486 {AddrSizePair.first, AddrSizePair.second}); 1487 } 1488 // Deallocate the memory for each globalized value 1489 for (auto &Rec : llvm::reverse(I->getSecond().LocalVarData)) { 1490 const auto *VD = cast<VarDecl>(Rec.first); 1491 I->getSecond().MappedParams->restore(CGF); 1492 1493 llvm::Value *FreeArgs[] = {Rec.second.GlobalizedVal, 1494 CGF.getTypeSize(VD->getType())}; 1495 CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( 1496 CGM.getModule(), OMPRTL___kmpc_free_shared), 1497 FreeArgs); 1498 } 1499 } 1500 } 1501 1502 void CGOpenMPRuntimeGPU::emitTeamsCall(CodeGenFunction &CGF, 1503 const OMPExecutableDirective &D, 1504 SourceLocation Loc, 1505 llvm::Function *OutlinedFn, 1506 ArrayRef<llvm::Value *> CapturedVars) { 1507 if (!CGF.HaveInsertPoint()) 1508 return; 1509 1510 Address ZeroAddr = CGF.CreateDefaultAlignTempAlloca(CGF.Int32Ty, 1511 /*Name=*/".zero.addr"); 1512 CGF.Builder.CreateStore(CGF.Builder.getInt32(/*C*/ 0), ZeroAddr); 1513 llvm::SmallVector<llvm::Value *, 16> OutlinedFnArgs; 1514 OutlinedFnArgs.push_back(emitThreadIDAddress(CGF, Loc).getPointer()); 1515 OutlinedFnArgs.push_back(ZeroAddr.getPointer()); 1516 OutlinedFnArgs.append(CapturedVars.begin(), CapturedVars.end()); 1517 emitOutlinedFunctionCall(CGF, Loc, OutlinedFn, OutlinedFnArgs); 1518 } 1519 1520 void CGOpenMPRuntimeGPU::emitParallelCall(CodeGenFunction &CGF, 1521 SourceLocation Loc, 1522 llvm::Function *OutlinedFn, 1523 ArrayRef<llvm::Value *> CapturedVars, 1524 const Expr *IfCond, 1525 llvm::Value *NumThreads) { 1526 if (!CGF.HaveInsertPoint()) 1527 return; 1528 1529 auto &&ParallelGen = [this, Loc, OutlinedFn, CapturedVars, IfCond, 1530 NumThreads](CodeGenFunction &CGF, 1531 PrePostActionTy &Action) { 1532 CGBuilderTy &Bld = CGF.Builder; 1533 llvm::Value *NumThreadsVal = NumThreads; 1534 llvm::Function *WFn = WrapperFunctionsMap[OutlinedFn]; 1535 llvm::Value *ID = llvm::ConstantPointerNull::get(CGM.Int8PtrTy); 1536 if (WFn) 1537 ID = Bld.CreateBitOrPointerCast(WFn, CGM.Int8PtrTy); 1538 llvm::Value *FnPtr = Bld.CreateBitOrPointerCast(OutlinedFn, CGM.Int8PtrTy); 1539 1540 // Create a private scope that will globalize the arguments 1541 // passed from the outside of the target region. 1542 // TODO: Is that needed? 1543 CodeGenFunction::OMPPrivateScope PrivateArgScope(CGF); 1544 1545 Address CapturedVarsAddrs = CGF.CreateDefaultAlignTempAlloca( 1546 llvm::ArrayType::get(CGM.VoidPtrTy, CapturedVars.size()), 1547 "captured_vars_addrs"); 1548 // There's something to share. 1549 if (!CapturedVars.empty()) { 1550 // Prepare for parallel region. Indicate the outlined function. 1551 ASTContext &Ctx = CGF.getContext(); 1552 unsigned Idx = 0; 1553 for (llvm::Value *V : CapturedVars) { 1554 Address Dst = Bld.CreateConstArrayGEP(CapturedVarsAddrs, Idx); 1555 llvm::Value *PtrV; 1556 if (V->getType()->isIntegerTy()) 1557 PtrV = Bld.CreateIntToPtr(V, CGF.VoidPtrTy); 1558 else 1559 PtrV = Bld.CreatePointerBitCastOrAddrSpaceCast(V, CGF.VoidPtrTy); 1560 CGF.EmitStoreOfScalar(PtrV, Dst, /*Volatile=*/false, 1561 Ctx.getPointerType(Ctx.VoidPtrTy)); 1562 ++Idx; 1563 } 1564 } 1565 1566 llvm::Value *IfCondVal = nullptr; 1567 if (IfCond) 1568 IfCondVal = Bld.CreateIntCast(CGF.EvaluateExprAsBool(IfCond), CGF.Int32Ty, 1569 /* isSigned */ false); 1570 else 1571 IfCondVal = llvm::ConstantInt::get(CGF.Int32Ty, 1); 1572 1573 if (!NumThreadsVal) 1574 NumThreadsVal = llvm::ConstantInt::get(CGF.Int32Ty, -1); 1575 else 1576 NumThreadsVal = Bld.CreateZExtOrTrunc(NumThreadsVal, CGF.Int32Ty), 1577 1578 assert(IfCondVal && "Expected a value"); 1579 llvm::Value *RTLoc = emitUpdateLocation(CGF, Loc); 1580 llvm::Value *Args[] = { 1581 RTLoc, 1582 getThreadID(CGF, Loc), 1583 IfCondVal, 1584 NumThreadsVal, 1585 llvm::ConstantInt::get(CGF.Int32Ty, -1), 1586 FnPtr, 1587 ID, 1588 Bld.CreateBitOrPointerCast(CapturedVarsAddrs.getPointer(), 1589 CGF.VoidPtrPtrTy), 1590 llvm::ConstantInt::get(CGM.SizeTy, CapturedVars.size())}; 1591 CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( 1592 CGM.getModule(), OMPRTL___kmpc_parallel_51), 1593 Args); 1594 }; 1595 1596 RegionCodeGenTy RCG(ParallelGen); 1597 RCG(CGF); 1598 } 1599 1600 void CGOpenMPRuntimeGPU::syncCTAThreads(CodeGenFunction &CGF) { 1601 // Always emit simple barriers! 1602 if (!CGF.HaveInsertPoint()) 1603 return; 1604 // Build call __kmpc_barrier_simple_spmd(nullptr, 0); 1605 // This function does not use parameters, so we can emit just default values. 1606 llvm::Value *Args[] = { 1607 llvm::ConstantPointerNull::get( 1608 cast<llvm::PointerType>(getIdentTyPointerTy())), 1609 llvm::ConstantInt::get(CGF.Int32Ty, /*V=*/0, /*isSigned=*/true)}; 1610 CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( 1611 CGM.getModule(), OMPRTL___kmpc_barrier_simple_spmd), 1612 Args); 1613 } 1614 1615 void CGOpenMPRuntimeGPU::emitBarrierCall(CodeGenFunction &CGF, 1616 SourceLocation Loc, 1617 OpenMPDirectiveKind Kind, bool, 1618 bool) { 1619 // Always emit simple barriers! 1620 if (!CGF.HaveInsertPoint()) 1621 return; 1622 // Build call __kmpc_cancel_barrier(loc, thread_id); 1623 unsigned Flags = getDefaultFlagsForBarriers(Kind); 1624 llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc, Flags), 1625 getThreadID(CGF, Loc)}; 1626 1627 CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( 1628 CGM.getModule(), OMPRTL___kmpc_barrier), 1629 Args); 1630 } 1631 1632 void CGOpenMPRuntimeGPU::emitCriticalRegion( 1633 CodeGenFunction &CGF, StringRef CriticalName, 1634 const RegionCodeGenTy &CriticalOpGen, SourceLocation Loc, 1635 const Expr *Hint) { 1636 llvm::BasicBlock *LoopBB = CGF.createBasicBlock("omp.critical.loop"); 1637 llvm::BasicBlock *TestBB = CGF.createBasicBlock("omp.critical.test"); 1638 llvm::BasicBlock *SyncBB = CGF.createBasicBlock("omp.critical.sync"); 1639 llvm::BasicBlock *BodyBB = CGF.createBasicBlock("omp.critical.body"); 1640 llvm::BasicBlock *ExitBB = CGF.createBasicBlock("omp.critical.exit"); 1641 1642 auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()); 1643 1644 // Get the mask of active threads in the warp. 1645 llvm::Value *Mask = CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( 1646 CGM.getModule(), OMPRTL___kmpc_warp_active_thread_mask)); 1647 // Fetch team-local id of the thread. 1648 llvm::Value *ThreadID = RT.getGPUThreadID(CGF); 1649 1650 // Get the width of the team. 1651 llvm::Value *TeamWidth = RT.getGPUNumThreads(CGF); 1652 1653 // Initialize the counter variable for the loop. 1654 QualType Int32Ty = 1655 CGF.getContext().getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/0); 1656 Address Counter = CGF.CreateMemTemp(Int32Ty, "critical_counter"); 1657 LValue CounterLVal = CGF.MakeAddrLValue(Counter, Int32Ty); 1658 CGF.EmitStoreOfScalar(llvm::Constant::getNullValue(CGM.Int32Ty), CounterLVal, 1659 /*isInit=*/true); 1660 1661 // Block checks if loop counter exceeds upper bound. 1662 CGF.EmitBlock(LoopBB); 1663 llvm::Value *CounterVal = CGF.EmitLoadOfScalar(CounterLVal, Loc); 1664 llvm::Value *CmpLoopBound = CGF.Builder.CreateICmpSLT(CounterVal, TeamWidth); 1665 CGF.Builder.CreateCondBr(CmpLoopBound, TestBB, ExitBB); 1666 1667 // Block tests which single thread should execute region, and which threads 1668 // should go straight to synchronisation point. 1669 CGF.EmitBlock(TestBB); 1670 CounterVal = CGF.EmitLoadOfScalar(CounterLVal, Loc); 1671 llvm::Value *CmpThreadToCounter = 1672 CGF.Builder.CreateICmpEQ(ThreadID, CounterVal); 1673 CGF.Builder.CreateCondBr(CmpThreadToCounter, BodyBB, SyncBB); 1674 1675 // Block emits the body of the critical region. 1676 CGF.EmitBlock(BodyBB); 1677 1678 // Output the critical statement. 1679 CGOpenMPRuntime::emitCriticalRegion(CGF, CriticalName, CriticalOpGen, Loc, 1680 Hint); 1681 1682 // After the body surrounded by the critical region, the single executing 1683 // thread will jump to the synchronisation point. 1684 // Block waits for all threads in current team to finish then increments the 1685 // counter variable and returns to the loop. 1686 CGF.EmitBlock(SyncBB); 1687 // Reconverge active threads in the warp. 1688 (void)CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( 1689 CGM.getModule(), OMPRTL___kmpc_syncwarp), 1690 Mask); 1691 1692 llvm::Value *IncCounterVal = 1693 CGF.Builder.CreateNSWAdd(CounterVal, CGF.Builder.getInt32(1)); 1694 CGF.EmitStoreOfScalar(IncCounterVal, CounterLVal); 1695 CGF.EmitBranch(LoopBB); 1696 1697 // Block that is reached when all threads in the team complete the region. 1698 CGF.EmitBlock(ExitBB, /*IsFinished=*/true); 1699 } 1700 1701 /// Cast value to the specified type. 1702 static llvm::Value *castValueToType(CodeGenFunction &CGF, llvm::Value *Val, 1703 QualType ValTy, QualType CastTy, 1704 SourceLocation Loc) { 1705 assert(!CGF.getContext().getTypeSizeInChars(CastTy).isZero() && 1706 "Cast type must sized."); 1707 assert(!CGF.getContext().getTypeSizeInChars(ValTy).isZero() && 1708 "Val type must sized."); 1709 llvm::Type *LLVMCastTy = CGF.ConvertTypeForMem(CastTy); 1710 if (ValTy == CastTy) 1711 return Val; 1712 if (CGF.getContext().getTypeSizeInChars(ValTy) == 1713 CGF.getContext().getTypeSizeInChars(CastTy)) 1714 return CGF.Builder.CreateBitCast(Val, LLVMCastTy); 1715 if (CastTy->isIntegerType() && ValTy->isIntegerType()) 1716 return CGF.Builder.CreateIntCast(Val, LLVMCastTy, 1717 CastTy->hasSignedIntegerRepresentation()); 1718 Address CastItem = CGF.CreateMemTemp(CastTy); 1719 Address ValCastItem = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( 1720 CastItem, Val->getType()->getPointerTo(CastItem.getAddressSpace()), 1721 Val->getType()); 1722 CGF.EmitStoreOfScalar(Val, ValCastItem, /*Volatile=*/false, ValTy, 1723 LValueBaseInfo(AlignmentSource::Type), 1724 TBAAAccessInfo()); 1725 return CGF.EmitLoadOfScalar(CastItem, /*Volatile=*/false, CastTy, Loc, 1726 LValueBaseInfo(AlignmentSource::Type), 1727 TBAAAccessInfo()); 1728 } 1729 1730 /// This function creates calls to one of two shuffle functions to copy 1731 /// variables between lanes in a warp. 1732 static llvm::Value *createRuntimeShuffleFunction(CodeGenFunction &CGF, 1733 llvm::Value *Elem, 1734 QualType ElemType, 1735 llvm::Value *Offset, 1736 SourceLocation Loc) { 1737 CodeGenModule &CGM = CGF.CGM; 1738 CGBuilderTy &Bld = CGF.Builder; 1739 CGOpenMPRuntimeGPU &RT = 1740 *(static_cast<CGOpenMPRuntimeGPU *>(&CGM.getOpenMPRuntime())); 1741 llvm::OpenMPIRBuilder &OMPBuilder = RT.getOMPBuilder(); 1742 1743 CharUnits Size = CGF.getContext().getTypeSizeInChars(ElemType); 1744 assert(Size.getQuantity() <= 8 && 1745 "Unsupported bitwidth in shuffle instruction."); 1746 1747 RuntimeFunction ShuffleFn = Size.getQuantity() <= 4 1748 ? OMPRTL___kmpc_shuffle_int32 1749 : OMPRTL___kmpc_shuffle_int64; 1750 1751 // Cast all types to 32- or 64-bit values before calling shuffle routines. 1752 QualType CastTy = CGF.getContext().getIntTypeForBitwidth( 1753 Size.getQuantity() <= 4 ? 32 : 64, /*Signed=*/1); 1754 llvm::Value *ElemCast = castValueToType(CGF, Elem, ElemType, CastTy, Loc); 1755 llvm::Value *WarpSize = 1756 Bld.CreateIntCast(RT.getGPUWarpSize(CGF), CGM.Int16Ty, /*isSigned=*/true); 1757 1758 llvm::Value *ShuffledVal = CGF.EmitRuntimeCall( 1759 OMPBuilder.getOrCreateRuntimeFunction(CGM.getModule(), ShuffleFn), 1760 {ElemCast, Offset, WarpSize}); 1761 1762 return castValueToType(CGF, ShuffledVal, CastTy, ElemType, Loc); 1763 } 1764 1765 static void shuffleAndStore(CodeGenFunction &CGF, Address SrcAddr, 1766 Address DestAddr, QualType ElemType, 1767 llvm::Value *Offset, SourceLocation Loc) { 1768 CGBuilderTy &Bld = CGF.Builder; 1769 1770 CharUnits Size = CGF.getContext().getTypeSizeInChars(ElemType); 1771 // Create the loop over the big sized data. 1772 // ptr = (void*)Elem; 1773 // ptrEnd = (void*) Elem + 1; 1774 // Step = 8; 1775 // while (ptr + Step < ptrEnd) 1776 // shuffle((int64_t)*ptr); 1777 // Step = 4; 1778 // while (ptr + Step < ptrEnd) 1779 // shuffle((int32_t)*ptr); 1780 // ... 1781 Address ElemPtr = DestAddr; 1782 Address Ptr = SrcAddr; 1783 Address PtrEnd = Bld.CreatePointerBitCastOrAddrSpaceCast( 1784 Bld.CreateConstGEP(SrcAddr, 1), CGF.VoidPtrTy, CGF.Int8Ty); 1785 for (int IntSize = 8; IntSize >= 1; IntSize /= 2) { 1786 if (Size < CharUnits::fromQuantity(IntSize)) 1787 continue; 1788 QualType IntType = CGF.getContext().getIntTypeForBitwidth( 1789 CGF.getContext().toBits(CharUnits::fromQuantity(IntSize)), 1790 /*Signed=*/1); 1791 llvm::Type *IntTy = CGF.ConvertTypeForMem(IntType); 1792 Ptr = Bld.CreatePointerBitCastOrAddrSpaceCast(Ptr, IntTy->getPointerTo(), 1793 IntTy); 1794 ElemPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( 1795 ElemPtr, IntTy->getPointerTo(), IntTy); 1796 if (Size.getQuantity() / IntSize > 1) { 1797 llvm::BasicBlock *PreCondBB = CGF.createBasicBlock(".shuffle.pre_cond"); 1798 llvm::BasicBlock *ThenBB = CGF.createBasicBlock(".shuffle.then"); 1799 llvm::BasicBlock *ExitBB = CGF.createBasicBlock(".shuffle.exit"); 1800 llvm::BasicBlock *CurrentBB = Bld.GetInsertBlock(); 1801 CGF.EmitBlock(PreCondBB); 1802 llvm::PHINode *PhiSrc = 1803 Bld.CreatePHI(Ptr.getType(), /*NumReservedValues=*/2); 1804 PhiSrc->addIncoming(Ptr.getPointer(), CurrentBB); 1805 llvm::PHINode *PhiDest = 1806 Bld.CreatePHI(ElemPtr.getType(), /*NumReservedValues=*/2); 1807 PhiDest->addIncoming(ElemPtr.getPointer(), CurrentBB); 1808 Ptr = Address(PhiSrc, Ptr.getElementType(), Ptr.getAlignment()); 1809 ElemPtr = 1810 Address(PhiDest, ElemPtr.getElementType(), ElemPtr.getAlignment()); 1811 llvm::Value *PtrDiff = Bld.CreatePtrDiff( 1812 CGF.Int8Ty, PtrEnd.getPointer(), 1813 Bld.CreatePointerBitCastOrAddrSpaceCast(Ptr.getPointer(), 1814 CGF.VoidPtrTy)); 1815 Bld.CreateCondBr(Bld.CreateICmpSGT(PtrDiff, Bld.getInt64(IntSize - 1)), 1816 ThenBB, ExitBB); 1817 CGF.EmitBlock(ThenBB); 1818 llvm::Value *Res = createRuntimeShuffleFunction( 1819 CGF, 1820 CGF.EmitLoadOfScalar(Ptr, /*Volatile=*/false, IntType, Loc, 1821 LValueBaseInfo(AlignmentSource::Type), 1822 TBAAAccessInfo()), 1823 IntType, Offset, Loc); 1824 CGF.EmitStoreOfScalar(Res, ElemPtr, /*Volatile=*/false, IntType, 1825 LValueBaseInfo(AlignmentSource::Type), 1826 TBAAAccessInfo()); 1827 Address LocalPtr = Bld.CreateConstGEP(Ptr, 1); 1828 Address LocalElemPtr = Bld.CreateConstGEP(ElemPtr, 1); 1829 PhiSrc->addIncoming(LocalPtr.getPointer(), ThenBB); 1830 PhiDest->addIncoming(LocalElemPtr.getPointer(), ThenBB); 1831 CGF.EmitBranch(PreCondBB); 1832 CGF.EmitBlock(ExitBB); 1833 } else { 1834 llvm::Value *Res = createRuntimeShuffleFunction( 1835 CGF, 1836 CGF.EmitLoadOfScalar(Ptr, /*Volatile=*/false, IntType, Loc, 1837 LValueBaseInfo(AlignmentSource::Type), 1838 TBAAAccessInfo()), 1839 IntType, Offset, Loc); 1840 CGF.EmitStoreOfScalar(Res, ElemPtr, /*Volatile=*/false, IntType, 1841 LValueBaseInfo(AlignmentSource::Type), 1842 TBAAAccessInfo()); 1843 Ptr = Bld.CreateConstGEP(Ptr, 1); 1844 ElemPtr = Bld.CreateConstGEP(ElemPtr, 1); 1845 } 1846 Size = Size % IntSize; 1847 } 1848 } 1849 1850 namespace { 1851 enum CopyAction : unsigned { 1852 // RemoteLaneToThread: Copy over a Reduce list from a remote lane in 1853 // the warp using shuffle instructions. 1854 RemoteLaneToThread, 1855 // ThreadCopy: Make a copy of a Reduce list on the thread's stack. 1856 ThreadCopy, 1857 // ThreadToScratchpad: Copy a team-reduced array to the scratchpad. 1858 ThreadToScratchpad, 1859 // ScratchpadToThread: Copy from a scratchpad array in global memory 1860 // containing team-reduced data to a thread's stack. 1861 ScratchpadToThread, 1862 }; 1863 } // namespace 1864 1865 struct CopyOptionsTy { 1866 llvm::Value *RemoteLaneOffset; 1867 llvm::Value *ScratchpadIndex; 1868 llvm::Value *ScratchpadWidth; 1869 }; 1870 1871 /// Emit instructions to copy a Reduce list, which contains partially 1872 /// aggregated values, in the specified direction. 1873 static void emitReductionListCopy( 1874 CopyAction Action, CodeGenFunction &CGF, QualType ReductionArrayTy, 1875 ArrayRef<const Expr *> Privates, Address SrcBase, Address DestBase, 1876 CopyOptionsTy CopyOptions = {nullptr, nullptr, nullptr}) { 1877 1878 CodeGenModule &CGM = CGF.CGM; 1879 ASTContext &C = CGM.getContext(); 1880 CGBuilderTy &Bld = CGF.Builder; 1881 1882 llvm::Value *RemoteLaneOffset = CopyOptions.RemoteLaneOffset; 1883 llvm::Value *ScratchpadIndex = CopyOptions.ScratchpadIndex; 1884 llvm::Value *ScratchpadWidth = CopyOptions.ScratchpadWidth; 1885 1886 // Iterates, element-by-element, through the source Reduce list and 1887 // make a copy. 1888 unsigned Idx = 0; 1889 unsigned Size = Privates.size(); 1890 for (const Expr *Private : Privates) { 1891 Address SrcElementAddr = Address::invalid(); 1892 Address DestElementAddr = Address::invalid(); 1893 Address DestElementPtrAddr = Address::invalid(); 1894 // Should we shuffle in an element from a remote lane? 1895 bool ShuffleInElement = false; 1896 // Set to true to update the pointer in the dest Reduce list to a 1897 // newly created element. 1898 bool UpdateDestListPtr = false; 1899 // Increment the src or dest pointer to the scratchpad, for each 1900 // new element. 1901 bool IncrScratchpadSrc = false; 1902 bool IncrScratchpadDest = false; 1903 QualType PrivatePtrType = C.getPointerType(Private->getType()); 1904 llvm::Type *PrivateLlvmPtrType = CGF.ConvertType(PrivatePtrType); 1905 1906 switch (Action) { 1907 case RemoteLaneToThread: { 1908 // Step 1.1: Get the address for the src element in the Reduce list. 1909 Address SrcElementPtrAddr = Bld.CreateConstArrayGEP(SrcBase, Idx); 1910 SrcElementAddr = 1911 CGF.EmitLoadOfPointer(CGF.Builder.CreateElementBitCast( 1912 SrcElementPtrAddr, PrivateLlvmPtrType), 1913 PrivatePtrType->castAs<PointerType>()); 1914 1915 // Step 1.2: Create a temporary to store the element in the destination 1916 // Reduce list. 1917 DestElementPtrAddr = Bld.CreateConstArrayGEP(DestBase, Idx); 1918 DestElementAddr = 1919 CGF.CreateMemTemp(Private->getType(), ".omp.reduction.element"); 1920 ShuffleInElement = true; 1921 UpdateDestListPtr = true; 1922 break; 1923 } 1924 case ThreadCopy: { 1925 // Step 1.1: Get the address for the src element in the Reduce list. 1926 Address SrcElementPtrAddr = Bld.CreateConstArrayGEP(SrcBase, Idx); 1927 SrcElementAddr = 1928 CGF.EmitLoadOfPointer(CGF.Builder.CreateElementBitCast( 1929 SrcElementPtrAddr, PrivateLlvmPtrType), 1930 PrivatePtrType->castAs<PointerType>()); 1931 1932 // Step 1.2: Get the address for dest element. The destination 1933 // element has already been created on the thread's stack. 1934 DestElementPtrAddr = Bld.CreateConstArrayGEP(DestBase, Idx); 1935 DestElementAddr = 1936 CGF.EmitLoadOfPointer(CGF.Builder.CreateElementBitCast( 1937 DestElementPtrAddr, PrivateLlvmPtrType), 1938 PrivatePtrType->castAs<PointerType>()); 1939 break; 1940 } 1941 case ThreadToScratchpad: { 1942 // Step 1.1: Get the address for the src element in the Reduce list. 1943 Address SrcElementPtrAddr = Bld.CreateConstArrayGEP(SrcBase, Idx); 1944 SrcElementAddr = 1945 CGF.EmitLoadOfPointer(CGF.Builder.CreateElementBitCast( 1946 SrcElementPtrAddr, PrivateLlvmPtrType), 1947 PrivatePtrType->castAs<PointerType>()); 1948 1949 // Step 1.2: Get the address for dest element: 1950 // address = base + index * ElementSizeInChars. 1951 llvm::Value *ElementSizeInChars = CGF.getTypeSize(Private->getType()); 1952 llvm::Value *CurrentOffset = 1953 Bld.CreateNUWMul(ElementSizeInChars, ScratchpadIndex); 1954 llvm::Value *ScratchPadElemAbsolutePtrVal = 1955 Bld.CreateNUWAdd(DestBase.getPointer(), CurrentOffset); 1956 ScratchPadElemAbsolutePtrVal = 1957 Bld.CreateIntToPtr(ScratchPadElemAbsolutePtrVal, CGF.VoidPtrTy); 1958 DestElementAddr = Address(ScratchPadElemAbsolutePtrVal, CGF.Int8Ty, 1959 C.getTypeAlignInChars(Private->getType())); 1960 IncrScratchpadDest = true; 1961 break; 1962 } 1963 case ScratchpadToThread: { 1964 // Step 1.1: Get the address for the src element in the scratchpad. 1965 // address = base + index * ElementSizeInChars. 1966 llvm::Value *ElementSizeInChars = CGF.getTypeSize(Private->getType()); 1967 llvm::Value *CurrentOffset = 1968 Bld.CreateNUWMul(ElementSizeInChars, ScratchpadIndex); 1969 llvm::Value *ScratchPadElemAbsolutePtrVal = 1970 Bld.CreateNUWAdd(SrcBase.getPointer(), CurrentOffset); 1971 ScratchPadElemAbsolutePtrVal = 1972 Bld.CreateIntToPtr(ScratchPadElemAbsolutePtrVal, CGF.VoidPtrTy); 1973 SrcElementAddr = Address(ScratchPadElemAbsolutePtrVal, CGF.Int8Ty, 1974 C.getTypeAlignInChars(Private->getType())); 1975 IncrScratchpadSrc = true; 1976 1977 // Step 1.2: Create a temporary to store the element in the destination 1978 // Reduce list. 1979 DestElementPtrAddr = Bld.CreateConstArrayGEP(DestBase, Idx); 1980 DestElementAddr = 1981 CGF.CreateMemTemp(Private->getType(), ".omp.reduction.element"); 1982 UpdateDestListPtr = true; 1983 break; 1984 } 1985 } 1986 1987 // Regardless of src and dest of copy, we emit the load of src 1988 // element as this is required in all directions 1989 SrcElementAddr = Bld.CreateElementBitCast( 1990 SrcElementAddr, CGF.ConvertTypeForMem(Private->getType())); 1991 DestElementAddr = Bld.CreateElementBitCast(DestElementAddr, 1992 SrcElementAddr.getElementType()); 1993 1994 // Now that all active lanes have read the element in the 1995 // Reduce list, shuffle over the value from the remote lane. 1996 if (ShuffleInElement) { 1997 shuffleAndStore(CGF, SrcElementAddr, DestElementAddr, Private->getType(), 1998 RemoteLaneOffset, Private->getExprLoc()); 1999 } else { 2000 switch (CGF.getEvaluationKind(Private->getType())) { 2001 case TEK_Scalar: { 2002 llvm::Value *Elem = CGF.EmitLoadOfScalar( 2003 SrcElementAddr, /*Volatile=*/false, Private->getType(), 2004 Private->getExprLoc(), LValueBaseInfo(AlignmentSource::Type), 2005 TBAAAccessInfo()); 2006 // Store the source element value to the dest element address. 2007 CGF.EmitStoreOfScalar( 2008 Elem, DestElementAddr, /*Volatile=*/false, Private->getType(), 2009 LValueBaseInfo(AlignmentSource::Type), TBAAAccessInfo()); 2010 break; 2011 } 2012 case TEK_Complex: { 2013 CodeGenFunction::ComplexPairTy Elem = CGF.EmitLoadOfComplex( 2014 CGF.MakeAddrLValue(SrcElementAddr, Private->getType()), 2015 Private->getExprLoc()); 2016 CGF.EmitStoreOfComplex( 2017 Elem, CGF.MakeAddrLValue(DestElementAddr, Private->getType()), 2018 /*isInit=*/false); 2019 break; 2020 } 2021 case TEK_Aggregate: 2022 CGF.EmitAggregateCopy( 2023 CGF.MakeAddrLValue(DestElementAddr, Private->getType()), 2024 CGF.MakeAddrLValue(SrcElementAddr, Private->getType()), 2025 Private->getType(), AggValueSlot::DoesNotOverlap); 2026 break; 2027 } 2028 } 2029 2030 // Step 3.1: Modify reference in dest Reduce list as needed. 2031 // Modifying the reference in Reduce list to point to the newly 2032 // created element. The element is live in the current function 2033 // scope and that of functions it invokes (i.e., reduce_function). 2034 // RemoteReduceData[i] = (void*)&RemoteElem 2035 if (UpdateDestListPtr) { 2036 CGF.EmitStoreOfScalar(Bld.CreatePointerBitCastOrAddrSpaceCast( 2037 DestElementAddr.getPointer(), CGF.VoidPtrTy), 2038 DestElementPtrAddr, /*Volatile=*/false, 2039 C.VoidPtrTy); 2040 } 2041 2042 // Step 4.1: Increment SrcBase/DestBase so that it points to the starting 2043 // address of the next element in scratchpad memory, unless we're currently 2044 // processing the last one. Memory alignment is also taken care of here. 2045 if ((IncrScratchpadDest || IncrScratchpadSrc) && (Idx + 1 < Size)) { 2046 // FIXME: This code doesn't make any sense, it's trying to perform 2047 // integer arithmetic on pointers. 2048 llvm::Value *ScratchpadBasePtr = 2049 IncrScratchpadDest ? DestBase.getPointer() : SrcBase.getPointer(); 2050 llvm::Value *ElementSizeInChars = CGF.getTypeSize(Private->getType()); 2051 ScratchpadBasePtr = Bld.CreateNUWAdd( 2052 ScratchpadBasePtr, 2053 Bld.CreateNUWMul(ScratchpadWidth, ElementSizeInChars)); 2054 2055 // Take care of global memory alignment for performance 2056 ScratchpadBasePtr = Bld.CreateNUWSub( 2057 ScratchpadBasePtr, llvm::ConstantInt::get(CGM.SizeTy, 1)); 2058 ScratchpadBasePtr = Bld.CreateUDiv( 2059 ScratchpadBasePtr, 2060 llvm::ConstantInt::get(CGM.SizeTy, GlobalMemoryAlignment)); 2061 ScratchpadBasePtr = Bld.CreateNUWAdd( 2062 ScratchpadBasePtr, llvm::ConstantInt::get(CGM.SizeTy, 1)); 2063 ScratchpadBasePtr = Bld.CreateNUWMul( 2064 ScratchpadBasePtr, 2065 llvm::ConstantInt::get(CGM.SizeTy, GlobalMemoryAlignment)); 2066 2067 if (IncrScratchpadDest) 2068 DestBase = 2069 Address(ScratchpadBasePtr, CGF.VoidPtrTy, CGF.getPointerAlign()); 2070 else /* IncrScratchpadSrc = true */ 2071 SrcBase = 2072 Address(ScratchpadBasePtr, CGF.VoidPtrTy, CGF.getPointerAlign()); 2073 } 2074 2075 ++Idx; 2076 } 2077 } 2078 2079 /// This function emits a helper that gathers Reduce lists from the first 2080 /// lane of every active warp to lanes in the first warp. 2081 /// 2082 /// void inter_warp_copy_func(void* reduce_data, num_warps) 2083 /// shared smem[warp_size]; 2084 /// For all data entries D in reduce_data: 2085 /// sync 2086 /// If (I am the first lane in each warp) 2087 /// Copy my local D to smem[warp_id] 2088 /// sync 2089 /// if (I am the first warp) 2090 /// Copy smem[thread_id] to my local D 2091 static llvm::Value *emitInterWarpCopyFunction(CodeGenModule &CGM, 2092 ArrayRef<const Expr *> Privates, 2093 QualType ReductionArrayTy, 2094 SourceLocation Loc) { 2095 ASTContext &C = CGM.getContext(); 2096 llvm::Module &M = CGM.getModule(); 2097 2098 // ReduceList: thread local Reduce list. 2099 // At the stage of the computation when this function is called, partially 2100 // aggregated values reside in the first lane of every active warp. 2101 ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, 2102 C.VoidPtrTy, ImplicitParamDecl::Other); 2103 // NumWarps: number of warps active in the parallel region. This could 2104 // be smaller than 32 (max warps in a CTA) for partial block reduction. 2105 ImplicitParamDecl NumWarpsArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, 2106 C.getIntTypeForBitwidth(32, /* Signed */ true), 2107 ImplicitParamDecl::Other); 2108 FunctionArgList Args; 2109 Args.push_back(&ReduceListArg); 2110 Args.push_back(&NumWarpsArg); 2111 2112 const CGFunctionInfo &CGFI = 2113 CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); 2114 auto *Fn = llvm::Function::Create(CGM.getTypes().GetFunctionType(CGFI), 2115 llvm::GlobalValue::InternalLinkage, 2116 "_omp_reduction_inter_warp_copy_func", &M); 2117 CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI); 2118 Fn->setDoesNotRecurse(); 2119 CodeGenFunction CGF(CGM); 2120 CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc); 2121 2122 CGBuilderTy &Bld = CGF.Builder; 2123 2124 // This array is used as a medium to transfer, one reduce element at a time, 2125 // the data from the first lane of every warp to lanes in the first warp 2126 // in order to perform the final step of a reduction in a parallel region 2127 // (reduction across warps). The array is placed in NVPTX __shared__ memory 2128 // for reduced latency, as well as to have a distinct copy for concurrently 2129 // executing target regions. The array is declared with common linkage so 2130 // as to be shared across compilation units. 2131 StringRef TransferMediumName = 2132 "__openmp_nvptx_data_transfer_temporary_storage"; 2133 llvm::GlobalVariable *TransferMedium = 2134 M.getGlobalVariable(TransferMediumName); 2135 unsigned WarpSize = CGF.getTarget().getGridValue().GV_Warp_Size; 2136 if (!TransferMedium) { 2137 auto *Ty = llvm::ArrayType::get(CGM.Int32Ty, WarpSize); 2138 unsigned SharedAddressSpace = C.getTargetAddressSpace(LangAS::cuda_shared); 2139 TransferMedium = new llvm::GlobalVariable( 2140 M, Ty, /*isConstant=*/false, llvm::GlobalVariable::WeakAnyLinkage, 2141 llvm::UndefValue::get(Ty), TransferMediumName, 2142 /*InsertBefore=*/nullptr, llvm::GlobalVariable::NotThreadLocal, 2143 SharedAddressSpace); 2144 CGM.addCompilerUsedGlobal(TransferMedium); 2145 } 2146 2147 auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()); 2148 // Get the CUDA thread id of the current OpenMP thread on the GPU. 2149 llvm::Value *ThreadID = RT.getGPUThreadID(CGF); 2150 // nvptx_lane_id = nvptx_id % warpsize 2151 llvm::Value *LaneID = getNVPTXLaneID(CGF); 2152 // nvptx_warp_id = nvptx_id / warpsize 2153 llvm::Value *WarpID = getNVPTXWarpID(CGF); 2154 2155 Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg); 2156 llvm::Type *ElemTy = CGF.ConvertTypeForMem(ReductionArrayTy); 2157 Address LocalReduceList( 2158 Bld.CreatePointerBitCastOrAddrSpaceCast( 2159 CGF.EmitLoadOfScalar( 2160 AddrReduceListArg, /*Volatile=*/false, C.VoidPtrTy, Loc, 2161 LValueBaseInfo(AlignmentSource::Type), TBAAAccessInfo()), 2162 ElemTy->getPointerTo()), 2163 ElemTy, CGF.getPointerAlign()); 2164 2165 unsigned Idx = 0; 2166 for (const Expr *Private : Privates) { 2167 // 2168 // Warp master copies reduce element to transfer medium in __shared__ 2169 // memory. 2170 // 2171 unsigned RealTySize = 2172 C.getTypeSizeInChars(Private->getType()) 2173 .alignTo(C.getTypeAlignInChars(Private->getType())) 2174 .getQuantity(); 2175 for (unsigned TySize = 4; TySize > 0 && RealTySize > 0; TySize /=2) { 2176 unsigned NumIters = RealTySize / TySize; 2177 if (NumIters == 0) 2178 continue; 2179 QualType CType = C.getIntTypeForBitwidth( 2180 C.toBits(CharUnits::fromQuantity(TySize)), /*Signed=*/1); 2181 llvm::Type *CopyType = CGF.ConvertTypeForMem(CType); 2182 CharUnits Align = CharUnits::fromQuantity(TySize); 2183 llvm::Value *Cnt = nullptr; 2184 Address CntAddr = Address::invalid(); 2185 llvm::BasicBlock *PrecondBB = nullptr; 2186 llvm::BasicBlock *ExitBB = nullptr; 2187 if (NumIters > 1) { 2188 CntAddr = CGF.CreateMemTemp(C.IntTy, ".cnt.addr"); 2189 CGF.EmitStoreOfScalar(llvm::Constant::getNullValue(CGM.IntTy), CntAddr, 2190 /*Volatile=*/false, C.IntTy); 2191 PrecondBB = CGF.createBasicBlock("precond"); 2192 ExitBB = CGF.createBasicBlock("exit"); 2193 llvm::BasicBlock *BodyBB = CGF.createBasicBlock("body"); 2194 // There is no need to emit line number for unconditional branch. 2195 (void)ApplyDebugLocation::CreateEmpty(CGF); 2196 CGF.EmitBlock(PrecondBB); 2197 Cnt = CGF.EmitLoadOfScalar(CntAddr, /*Volatile=*/false, C.IntTy, Loc); 2198 llvm::Value *Cmp = 2199 Bld.CreateICmpULT(Cnt, llvm::ConstantInt::get(CGM.IntTy, NumIters)); 2200 Bld.CreateCondBr(Cmp, BodyBB, ExitBB); 2201 CGF.EmitBlock(BodyBB); 2202 } 2203 // kmpc_barrier. 2204 CGM.getOpenMPRuntime().emitBarrierCall(CGF, Loc, OMPD_unknown, 2205 /*EmitChecks=*/false, 2206 /*ForceSimpleCall=*/true); 2207 llvm::BasicBlock *ThenBB = CGF.createBasicBlock("then"); 2208 llvm::BasicBlock *ElseBB = CGF.createBasicBlock("else"); 2209 llvm::BasicBlock *MergeBB = CGF.createBasicBlock("ifcont"); 2210 2211 // if (lane_id == 0) 2212 llvm::Value *IsWarpMaster = Bld.CreateIsNull(LaneID, "warp_master"); 2213 Bld.CreateCondBr(IsWarpMaster, ThenBB, ElseBB); 2214 CGF.EmitBlock(ThenBB); 2215 2216 // Reduce element = LocalReduceList[i] 2217 Address ElemPtrPtrAddr = Bld.CreateConstArrayGEP(LocalReduceList, Idx); 2218 llvm::Value *ElemPtrPtr = CGF.EmitLoadOfScalar( 2219 ElemPtrPtrAddr, /*Volatile=*/false, C.VoidPtrTy, SourceLocation()); 2220 // elemptr = ((CopyType*)(elemptrptr)) + I 2221 Address ElemPtr(ElemPtrPtr, CGF.Int8Ty, Align); 2222 ElemPtr = Bld.CreateElementBitCast(ElemPtr, CopyType); 2223 if (NumIters > 1) 2224 ElemPtr = Bld.CreateGEP(ElemPtr, Cnt); 2225 2226 // Get pointer to location in transfer medium. 2227 // MediumPtr = &medium[warp_id] 2228 llvm::Value *MediumPtrVal = Bld.CreateInBoundsGEP( 2229 TransferMedium->getValueType(), TransferMedium, 2230 {llvm::Constant::getNullValue(CGM.Int64Ty), WarpID}); 2231 // Casting to actual data type. 2232 // MediumPtr = (CopyType*)MediumPtrAddr; 2233 Address MediumPtr( 2234 Bld.CreateBitCast( 2235 MediumPtrVal, 2236 CopyType->getPointerTo( 2237 MediumPtrVal->getType()->getPointerAddressSpace())), 2238 CopyType, Align); 2239 2240 // elem = *elemptr 2241 //*MediumPtr = elem 2242 llvm::Value *Elem = CGF.EmitLoadOfScalar( 2243 ElemPtr, /*Volatile=*/false, CType, Loc, 2244 LValueBaseInfo(AlignmentSource::Type), TBAAAccessInfo()); 2245 // Store the source element value to the dest element address. 2246 CGF.EmitStoreOfScalar(Elem, MediumPtr, /*Volatile=*/true, CType, 2247 LValueBaseInfo(AlignmentSource::Type), 2248 TBAAAccessInfo()); 2249 2250 Bld.CreateBr(MergeBB); 2251 2252 CGF.EmitBlock(ElseBB); 2253 Bld.CreateBr(MergeBB); 2254 2255 CGF.EmitBlock(MergeBB); 2256 2257 // kmpc_barrier. 2258 CGM.getOpenMPRuntime().emitBarrierCall(CGF, Loc, OMPD_unknown, 2259 /*EmitChecks=*/false, 2260 /*ForceSimpleCall=*/true); 2261 2262 // 2263 // Warp 0 copies reduce element from transfer medium. 2264 // 2265 llvm::BasicBlock *W0ThenBB = CGF.createBasicBlock("then"); 2266 llvm::BasicBlock *W0ElseBB = CGF.createBasicBlock("else"); 2267 llvm::BasicBlock *W0MergeBB = CGF.createBasicBlock("ifcont"); 2268 2269 Address AddrNumWarpsArg = CGF.GetAddrOfLocalVar(&NumWarpsArg); 2270 llvm::Value *NumWarpsVal = CGF.EmitLoadOfScalar( 2271 AddrNumWarpsArg, /*Volatile=*/false, C.IntTy, Loc); 2272 2273 // Up to 32 threads in warp 0 are active. 2274 llvm::Value *IsActiveThread = 2275 Bld.CreateICmpULT(ThreadID, NumWarpsVal, "is_active_thread"); 2276 Bld.CreateCondBr(IsActiveThread, W0ThenBB, W0ElseBB); 2277 2278 CGF.EmitBlock(W0ThenBB); 2279 2280 // SrcMediumPtr = &medium[tid] 2281 llvm::Value *SrcMediumPtrVal = Bld.CreateInBoundsGEP( 2282 TransferMedium->getValueType(), TransferMedium, 2283 {llvm::Constant::getNullValue(CGM.Int64Ty), ThreadID}); 2284 // SrcMediumVal = *SrcMediumPtr; 2285 Address SrcMediumPtr( 2286 Bld.CreateBitCast( 2287 SrcMediumPtrVal, 2288 CopyType->getPointerTo( 2289 SrcMediumPtrVal->getType()->getPointerAddressSpace())), 2290 CopyType, Align); 2291 2292 // TargetElemPtr = (CopyType*)(SrcDataAddr[i]) + I 2293 Address TargetElemPtrPtr = Bld.CreateConstArrayGEP(LocalReduceList, Idx); 2294 llvm::Value *TargetElemPtrVal = CGF.EmitLoadOfScalar( 2295 TargetElemPtrPtr, /*Volatile=*/false, C.VoidPtrTy, Loc); 2296 Address TargetElemPtr(TargetElemPtrVal, CGF.Int8Ty, Align); 2297 TargetElemPtr = Bld.CreateElementBitCast(TargetElemPtr, CopyType); 2298 if (NumIters > 1) 2299 TargetElemPtr = Bld.CreateGEP(TargetElemPtr, Cnt); 2300 2301 // *TargetElemPtr = SrcMediumVal; 2302 llvm::Value *SrcMediumValue = 2303 CGF.EmitLoadOfScalar(SrcMediumPtr, /*Volatile=*/true, CType, Loc); 2304 CGF.EmitStoreOfScalar(SrcMediumValue, TargetElemPtr, /*Volatile=*/false, 2305 CType); 2306 Bld.CreateBr(W0MergeBB); 2307 2308 CGF.EmitBlock(W0ElseBB); 2309 Bld.CreateBr(W0MergeBB); 2310 2311 CGF.EmitBlock(W0MergeBB); 2312 2313 if (NumIters > 1) { 2314 Cnt = Bld.CreateNSWAdd(Cnt, llvm::ConstantInt::get(CGM.IntTy, /*V=*/1)); 2315 CGF.EmitStoreOfScalar(Cnt, CntAddr, /*Volatile=*/false, C.IntTy); 2316 CGF.EmitBranch(PrecondBB); 2317 (void)ApplyDebugLocation::CreateEmpty(CGF); 2318 CGF.EmitBlock(ExitBB); 2319 } 2320 RealTySize %= TySize; 2321 } 2322 ++Idx; 2323 } 2324 2325 CGF.FinishFunction(); 2326 return Fn; 2327 } 2328 2329 /// Emit a helper that reduces data across two OpenMP threads (lanes) 2330 /// in the same warp. It uses shuffle instructions to copy over data from 2331 /// a remote lane's stack. The reduction algorithm performed is specified 2332 /// by the fourth parameter. 2333 /// 2334 /// Algorithm Versions. 2335 /// Full Warp Reduce (argument value 0): 2336 /// This algorithm assumes that all 32 lanes are active and gathers 2337 /// data from these 32 lanes, producing a single resultant value. 2338 /// Contiguous Partial Warp Reduce (argument value 1): 2339 /// This algorithm assumes that only a *contiguous* subset of lanes 2340 /// are active. This happens for the last warp in a parallel region 2341 /// when the user specified num_threads is not an integer multiple of 2342 /// 32. This contiguous subset always starts with the zeroth lane. 2343 /// Partial Warp Reduce (argument value 2): 2344 /// This algorithm gathers data from any number of lanes at any position. 2345 /// All reduced values are stored in the lowest possible lane. The set 2346 /// of problems every algorithm addresses is a super set of those 2347 /// addressable by algorithms with a lower version number. Overhead 2348 /// increases as algorithm version increases. 2349 /// 2350 /// Terminology 2351 /// Reduce element: 2352 /// Reduce element refers to the individual data field with primitive 2353 /// data types to be combined and reduced across threads. 2354 /// Reduce list: 2355 /// Reduce list refers to a collection of local, thread-private 2356 /// reduce elements. 2357 /// Remote Reduce list: 2358 /// Remote Reduce list refers to a collection of remote (relative to 2359 /// the current thread) reduce elements. 2360 /// 2361 /// We distinguish between three states of threads that are important to 2362 /// the implementation of this function. 2363 /// Alive threads: 2364 /// Threads in a warp executing the SIMT instruction, as distinguished from 2365 /// threads that are inactive due to divergent control flow. 2366 /// Active threads: 2367 /// The minimal set of threads that has to be alive upon entry to this 2368 /// function. The computation is correct iff active threads are alive. 2369 /// Some threads are alive but they are not active because they do not 2370 /// contribute to the computation in any useful manner. Turning them off 2371 /// may introduce control flow overheads without any tangible benefits. 2372 /// Effective threads: 2373 /// In order to comply with the argument requirements of the shuffle 2374 /// function, we must keep all lanes holding data alive. But at most 2375 /// half of them perform value aggregation; we refer to this half of 2376 /// threads as effective. The other half is simply handing off their 2377 /// data. 2378 /// 2379 /// Procedure 2380 /// Value shuffle: 2381 /// In this step active threads transfer data from higher lane positions 2382 /// in the warp to lower lane positions, creating Remote Reduce list. 2383 /// Value aggregation: 2384 /// In this step, effective threads combine their thread local Reduce list 2385 /// with Remote Reduce list and store the result in the thread local 2386 /// Reduce list. 2387 /// Value copy: 2388 /// In this step, we deal with the assumption made by algorithm 2 2389 /// (i.e. contiguity assumption). When we have an odd number of lanes 2390 /// active, say 2k+1, only k threads will be effective and therefore k 2391 /// new values will be produced. However, the Reduce list owned by the 2392 /// (2k+1)th thread is ignored in the value aggregation. Therefore 2393 /// we copy the Reduce list from the (2k+1)th lane to (k+1)th lane so 2394 /// that the contiguity assumption still holds. 2395 static llvm::Function *emitShuffleAndReduceFunction( 2396 CodeGenModule &CGM, ArrayRef<const Expr *> Privates, 2397 QualType ReductionArrayTy, llvm::Function *ReduceFn, SourceLocation Loc) { 2398 ASTContext &C = CGM.getContext(); 2399 2400 // Thread local Reduce list used to host the values of data to be reduced. 2401 ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, 2402 C.VoidPtrTy, ImplicitParamDecl::Other); 2403 // Current lane id; could be logical. 2404 ImplicitParamDecl LaneIDArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.ShortTy, 2405 ImplicitParamDecl::Other); 2406 // Offset of the remote source lane relative to the current lane. 2407 ImplicitParamDecl RemoteLaneOffsetArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, 2408 C.ShortTy, ImplicitParamDecl::Other); 2409 // Algorithm version. This is expected to be known at compile time. 2410 ImplicitParamDecl AlgoVerArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, 2411 C.ShortTy, ImplicitParamDecl::Other); 2412 FunctionArgList Args; 2413 Args.push_back(&ReduceListArg); 2414 Args.push_back(&LaneIDArg); 2415 Args.push_back(&RemoteLaneOffsetArg); 2416 Args.push_back(&AlgoVerArg); 2417 2418 const CGFunctionInfo &CGFI = 2419 CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); 2420 auto *Fn = llvm::Function::Create( 2421 CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage, 2422 "_omp_reduction_shuffle_and_reduce_func", &CGM.getModule()); 2423 CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI); 2424 Fn->setDoesNotRecurse(); 2425 2426 CodeGenFunction CGF(CGM); 2427 CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc); 2428 2429 CGBuilderTy &Bld = CGF.Builder; 2430 2431 Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg); 2432 llvm::Type *ElemTy = CGF.ConvertTypeForMem(ReductionArrayTy); 2433 Address LocalReduceList( 2434 Bld.CreatePointerBitCastOrAddrSpaceCast( 2435 CGF.EmitLoadOfScalar(AddrReduceListArg, /*Volatile=*/false, 2436 C.VoidPtrTy, SourceLocation()), 2437 ElemTy->getPointerTo()), 2438 ElemTy, CGF.getPointerAlign()); 2439 2440 Address AddrLaneIDArg = CGF.GetAddrOfLocalVar(&LaneIDArg); 2441 llvm::Value *LaneIDArgVal = CGF.EmitLoadOfScalar( 2442 AddrLaneIDArg, /*Volatile=*/false, C.ShortTy, SourceLocation()); 2443 2444 Address AddrRemoteLaneOffsetArg = CGF.GetAddrOfLocalVar(&RemoteLaneOffsetArg); 2445 llvm::Value *RemoteLaneOffsetArgVal = CGF.EmitLoadOfScalar( 2446 AddrRemoteLaneOffsetArg, /*Volatile=*/false, C.ShortTy, SourceLocation()); 2447 2448 Address AddrAlgoVerArg = CGF.GetAddrOfLocalVar(&AlgoVerArg); 2449 llvm::Value *AlgoVerArgVal = CGF.EmitLoadOfScalar( 2450 AddrAlgoVerArg, /*Volatile=*/false, C.ShortTy, SourceLocation()); 2451 2452 // Create a local thread-private variable to host the Reduce list 2453 // from a remote lane. 2454 Address RemoteReduceList = 2455 CGF.CreateMemTemp(ReductionArrayTy, ".omp.reduction.remote_reduce_list"); 2456 2457 // This loop iterates through the list of reduce elements and copies, 2458 // element by element, from a remote lane in the warp to RemoteReduceList, 2459 // hosted on the thread's stack. 2460 emitReductionListCopy(RemoteLaneToThread, CGF, ReductionArrayTy, Privates, 2461 LocalReduceList, RemoteReduceList, 2462 {/*RemoteLaneOffset=*/RemoteLaneOffsetArgVal, 2463 /*ScratchpadIndex=*/nullptr, 2464 /*ScratchpadWidth=*/nullptr}); 2465 2466 // The actions to be performed on the Remote Reduce list is dependent 2467 // on the algorithm version. 2468 // 2469 // if (AlgoVer==0) || (AlgoVer==1 && (LaneId < Offset)) || (AlgoVer==2 && 2470 // LaneId % 2 == 0 && Offset > 0): 2471 // do the reduction value aggregation 2472 // 2473 // The thread local variable Reduce list is mutated in place to host the 2474 // reduced data, which is the aggregated value produced from local and 2475 // remote lanes. 2476 // 2477 // Note that AlgoVer is expected to be a constant integer known at compile 2478 // time. 2479 // When AlgoVer==0, the first conjunction evaluates to true, making 2480 // the entire predicate true during compile time. 2481 // When AlgoVer==1, the second conjunction has only the second part to be 2482 // evaluated during runtime. Other conjunctions evaluates to false 2483 // during compile time. 2484 // When AlgoVer==2, the third conjunction has only the second part to be 2485 // evaluated during runtime. Other conjunctions evaluates to false 2486 // during compile time. 2487 llvm::Value *CondAlgo0 = Bld.CreateIsNull(AlgoVerArgVal); 2488 2489 llvm::Value *Algo1 = Bld.CreateICmpEQ(AlgoVerArgVal, Bld.getInt16(1)); 2490 llvm::Value *CondAlgo1 = Bld.CreateAnd( 2491 Algo1, Bld.CreateICmpULT(LaneIDArgVal, RemoteLaneOffsetArgVal)); 2492 2493 llvm::Value *Algo2 = Bld.CreateICmpEQ(AlgoVerArgVal, Bld.getInt16(2)); 2494 llvm::Value *CondAlgo2 = Bld.CreateAnd( 2495 Algo2, Bld.CreateIsNull(Bld.CreateAnd(LaneIDArgVal, Bld.getInt16(1)))); 2496 CondAlgo2 = Bld.CreateAnd( 2497 CondAlgo2, Bld.CreateICmpSGT(RemoteLaneOffsetArgVal, Bld.getInt16(0))); 2498 2499 llvm::Value *CondReduce = Bld.CreateOr(CondAlgo0, CondAlgo1); 2500 CondReduce = Bld.CreateOr(CondReduce, CondAlgo2); 2501 2502 llvm::BasicBlock *ThenBB = CGF.createBasicBlock("then"); 2503 llvm::BasicBlock *ElseBB = CGF.createBasicBlock("else"); 2504 llvm::BasicBlock *MergeBB = CGF.createBasicBlock("ifcont"); 2505 Bld.CreateCondBr(CondReduce, ThenBB, ElseBB); 2506 2507 CGF.EmitBlock(ThenBB); 2508 // reduce_function(LocalReduceList, RemoteReduceList) 2509 llvm::Value *LocalReduceListPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( 2510 LocalReduceList.getPointer(), CGF.VoidPtrTy); 2511 llvm::Value *RemoteReduceListPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( 2512 RemoteReduceList.getPointer(), CGF.VoidPtrTy); 2513 CGM.getOpenMPRuntime().emitOutlinedFunctionCall( 2514 CGF, Loc, ReduceFn, {LocalReduceListPtr, RemoteReduceListPtr}); 2515 Bld.CreateBr(MergeBB); 2516 2517 CGF.EmitBlock(ElseBB); 2518 Bld.CreateBr(MergeBB); 2519 2520 CGF.EmitBlock(MergeBB); 2521 2522 // if (AlgoVer==1 && (LaneId >= Offset)) copy Remote Reduce list to local 2523 // Reduce list. 2524 Algo1 = Bld.CreateICmpEQ(AlgoVerArgVal, Bld.getInt16(1)); 2525 llvm::Value *CondCopy = Bld.CreateAnd( 2526 Algo1, Bld.CreateICmpUGE(LaneIDArgVal, RemoteLaneOffsetArgVal)); 2527 2528 llvm::BasicBlock *CpyThenBB = CGF.createBasicBlock("then"); 2529 llvm::BasicBlock *CpyElseBB = CGF.createBasicBlock("else"); 2530 llvm::BasicBlock *CpyMergeBB = CGF.createBasicBlock("ifcont"); 2531 Bld.CreateCondBr(CondCopy, CpyThenBB, CpyElseBB); 2532 2533 CGF.EmitBlock(CpyThenBB); 2534 emitReductionListCopy(ThreadCopy, CGF, ReductionArrayTy, Privates, 2535 RemoteReduceList, LocalReduceList); 2536 Bld.CreateBr(CpyMergeBB); 2537 2538 CGF.EmitBlock(CpyElseBB); 2539 Bld.CreateBr(CpyMergeBB); 2540 2541 CGF.EmitBlock(CpyMergeBB); 2542 2543 CGF.FinishFunction(); 2544 return Fn; 2545 } 2546 2547 /// This function emits a helper that copies all the reduction variables from 2548 /// the team into the provided global buffer for the reduction variables. 2549 /// 2550 /// void list_to_global_copy_func(void *buffer, int Idx, void *reduce_data) 2551 /// For all data entries D in reduce_data: 2552 /// Copy local D to buffer.D[Idx] 2553 static llvm::Value *emitListToGlobalCopyFunction( 2554 CodeGenModule &CGM, ArrayRef<const Expr *> Privates, 2555 QualType ReductionArrayTy, SourceLocation Loc, 2556 const RecordDecl *TeamReductionRec, 2557 const llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> 2558 &VarFieldMap) { 2559 ASTContext &C = CGM.getContext(); 2560 2561 // Buffer: global reduction buffer. 2562 ImplicitParamDecl BufferArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, 2563 C.VoidPtrTy, ImplicitParamDecl::Other); 2564 // Idx: index of the buffer. 2565 ImplicitParamDecl IdxArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.IntTy, 2566 ImplicitParamDecl::Other); 2567 // ReduceList: thread local Reduce list. 2568 ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, 2569 C.VoidPtrTy, ImplicitParamDecl::Other); 2570 FunctionArgList Args; 2571 Args.push_back(&BufferArg); 2572 Args.push_back(&IdxArg); 2573 Args.push_back(&ReduceListArg); 2574 2575 const CGFunctionInfo &CGFI = 2576 CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); 2577 auto *Fn = llvm::Function::Create( 2578 CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage, 2579 "_omp_reduction_list_to_global_copy_func", &CGM.getModule()); 2580 CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI); 2581 Fn->setDoesNotRecurse(); 2582 CodeGenFunction CGF(CGM); 2583 CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc); 2584 2585 CGBuilderTy &Bld = CGF.Builder; 2586 2587 Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg); 2588 Address AddrBufferArg = CGF.GetAddrOfLocalVar(&BufferArg); 2589 llvm::Type *ElemTy = CGF.ConvertTypeForMem(ReductionArrayTy); 2590 Address LocalReduceList( 2591 Bld.CreatePointerBitCastOrAddrSpaceCast( 2592 CGF.EmitLoadOfScalar(AddrReduceListArg, /*Volatile=*/false, 2593 C.VoidPtrTy, Loc), 2594 ElemTy->getPointerTo()), 2595 ElemTy, CGF.getPointerAlign()); 2596 QualType StaticTy = C.getRecordType(TeamReductionRec); 2597 llvm::Type *LLVMReductionsBufferTy = 2598 CGM.getTypes().ConvertTypeForMem(StaticTy); 2599 llvm::Value *BufferArrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( 2600 CGF.EmitLoadOfScalar(AddrBufferArg, /*Volatile=*/false, C.VoidPtrTy, Loc), 2601 LLVMReductionsBufferTy->getPointerTo()); 2602 llvm::Value *Idxs[] = {llvm::ConstantInt::getNullValue(CGF.Int32Ty), 2603 CGF.EmitLoadOfScalar(CGF.GetAddrOfLocalVar(&IdxArg), 2604 /*Volatile=*/false, C.IntTy, 2605 Loc)}; 2606 unsigned Idx = 0; 2607 for (const Expr *Private : Privates) { 2608 // Reduce element = LocalReduceList[i] 2609 Address ElemPtrPtrAddr = Bld.CreateConstArrayGEP(LocalReduceList, Idx); 2610 llvm::Value *ElemPtrPtr = CGF.EmitLoadOfScalar( 2611 ElemPtrPtrAddr, /*Volatile=*/false, C.VoidPtrTy, SourceLocation()); 2612 // elemptr = ((CopyType*)(elemptrptr)) + I 2613 ElemTy = CGF.ConvertTypeForMem(Private->getType()); 2614 ElemPtrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( 2615 ElemPtrPtr, ElemTy->getPointerTo()); 2616 Address ElemPtr = 2617 Address(ElemPtrPtr, ElemTy, C.getTypeAlignInChars(Private->getType())); 2618 const ValueDecl *VD = cast<DeclRefExpr>(Private)->getDecl(); 2619 // Global = Buffer.VD[Idx]; 2620 const FieldDecl *FD = VarFieldMap.lookup(VD); 2621 LValue GlobLVal = CGF.EmitLValueForField( 2622 CGF.MakeNaturalAlignAddrLValue(BufferArrPtr, StaticTy), FD); 2623 Address GlobAddr = GlobLVal.getAddress(CGF); 2624 llvm::Value *BufferPtr = Bld.CreateInBoundsGEP(GlobAddr.getElementType(), 2625 GlobAddr.getPointer(), Idxs); 2626 GlobLVal.setAddress(Address(BufferPtr, 2627 CGF.ConvertTypeForMem(Private->getType()), 2628 GlobAddr.getAlignment())); 2629 switch (CGF.getEvaluationKind(Private->getType())) { 2630 case TEK_Scalar: { 2631 llvm::Value *V = CGF.EmitLoadOfScalar( 2632 ElemPtr, /*Volatile=*/false, Private->getType(), Loc, 2633 LValueBaseInfo(AlignmentSource::Type), TBAAAccessInfo()); 2634 CGF.EmitStoreOfScalar(V, GlobLVal); 2635 break; 2636 } 2637 case TEK_Complex: { 2638 CodeGenFunction::ComplexPairTy V = CGF.EmitLoadOfComplex( 2639 CGF.MakeAddrLValue(ElemPtr, Private->getType()), Loc); 2640 CGF.EmitStoreOfComplex(V, GlobLVal, /*isInit=*/false); 2641 break; 2642 } 2643 case TEK_Aggregate: 2644 CGF.EmitAggregateCopy(GlobLVal, 2645 CGF.MakeAddrLValue(ElemPtr, Private->getType()), 2646 Private->getType(), AggValueSlot::DoesNotOverlap); 2647 break; 2648 } 2649 ++Idx; 2650 } 2651 2652 CGF.FinishFunction(); 2653 return Fn; 2654 } 2655 2656 /// This function emits a helper that reduces all the reduction variables from 2657 /// the team into the provided global buffer for the reduction variables. 2658 /// 2659 /// void list_to_global_reduce_func(void *buffer, int Idx, void *reduce_data) 2660 /// void *GlobPtrs[]; 2661 /// GlobPtrs[0] = (void*)&buffer.D0[Idx]; 2662 /// ... 2663 /// GlobPtrs[N] = (void*)&buffer.DN[Idx]; 2664 /// reduce_function(GlobPtrs, reduce_data); 2665 static llvm::Value *emitListToGlobalReduceFunction( 2666 CodeGenModule &CGM, ArrayRef<const Expr *> Privates, 2667 QualType ReductionArrayTy, SourceLocation Loc, 2668 const RecordDecl *TeamReductionRec, 2669 const llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> 2670 &VarFieldMap, 2671 llvm::Function *ReduceFn) { 2672 ASTContext &C = CGM.getContext(); 2673 2674 // Buffer: global reduction buffer. 2675 ImplicitParamDecl BufferArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, 2676 C.VoidPtrTy, ImplicitParamDecl::Other); 2677 // Idx: index of the buffer. 2678 ImplicitParamDecl IdxArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.IntTy, 2679 ImplicitParamDecl::Other); 2680 // ReduceList: thread local Reduce list. 2681 ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, 2682 C.VoidPtrTy, ImplicitParamDecl::Other); 2683 FunctionArgList Args; 2684 Args.push_back(&BufferArg); 2685 Args.push_back(&IdxArg); 2686 Args.push_back(&ReduceListArg); 2687 2688 const CGFunctionInfo &CGFI = 2689 CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); 2690 auto *Fn = llvm::Function::Create( 2691 CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage, 2692 "_omp_reduction_list_to_global_reduce_func", &CGM.getModule()); 2693 CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI); 2694 Fn->setDoesNotRecurse(); 2695 CodeGenFunction CGF(CGM); 2696 CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc); 2697 2698 CGBuilderTy &Bld = CGF.Builder; 2699 2700 Address AddrBufferArg = CGF.GetAddrOfLocalVar(&BufferArg); 2701 QualType StaticTy = C.getRecordType(TeamReductionRec); 2702 llvm::Type *LLVMReductionsBufferTy = 2703 CGM.getTypes().ConvertTypeForMem(StaticTy); 2704 llvm::Value *BufferArrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( 2705 CGF.EmitLoadOfScalar(AddrBufferArg, /*Volatile=*/false, C.VoidPtrTy, Loc), 2706 LLVMReductionsBufferTy->getPointerTo()); 2707 2708 // 1. Build a list of reduction variables. 2709 // void *RedList[<n>] = {<ReductionVars>[0], ..., <ReductionVars>[<n>-1]}; 2710 Address ReductionList = 2711 CGF.CreateMemTemp(ReductionArrayTy, ".omp.reduction.red_list"); 2712 auto IPriv = Privates.begin(); 2713 llvm::Value *Idxs[] = {llvm::ConstantInt::getNullValue(CGF.Int32Ty), 2714 CGF.EmitLoadOfScalar(CGF.GetAddrOfLocalVar(&IdxArg), 2715 /*Volatile=*/false, C.IntTy, 2716 Loc)}; 2717 unsigned Idx = 0; 2718 for (unsigned I = 0, E = Privates.size(); I < E; ++I, ++IPriv, ++Idx) { 2719 Address Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx); 2720 // Global = Buffer.VD[Idx]; 2721 const ValueDecl *VD = cast<DeclRefExpr>(*IPriv)->getDecl(); 2722 const FieldDecl *FD = VarFieldMap.lookup(VD); 2723 LValue GlobLVal = CGF.EmitLValueForField( 2724 CGF.MakeNaturalAlignAddrLValue(BufferArrPtr, StaticTy), FD); 2725 Address GlobAddr = GlobLVal.getAddress(CGF); 2726 llvm::Value *BufferPtr = Bld.CreateInBoundsGEP( 2727 GlobAddr.getElementType(), GlobAddr.getPointer(), Idxs); 2728 llvm::Value *Ptr = CGF.EmitCastToVoidPtr(BufferPtr); 2729 CGF.EmitStoreOfScalar(Ptr, Elem, /*Volatile=*/false, C.VoidPtrTy); 2730 if ((*IPriv)->getType()->isVariablyModifiedType()) { 2731 // Store array size. 2732 ++Idx; 2733 Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx); 2734 llvm::Value *Size = CGF.Builder.CreateIntCast( 2735 CGF.getVLASize( 2736 CGF.getContext().getAsVariableArrayType((*IPriv)->getType())) 2737 .NumElts, 2738 CGF.SizeTy, /*isSigned=*/false); 2739 CGF.Builder.CreateStore(CGF.Builder.CreateIntToPtr(Size, CGF.VoidPtrTy), 2740 Elem); 2741 } 2742 } 2743 2744 // Call reduce_function(GlobalReduceList, ReduceList) 2745 llvm::Value *GlobalReduceList = 2746 CGF.EmitCastToVoidPtr(ReductionList.getPointer()); 2747 Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg); 2748 llvm::Value *ReducedPtr = CGF.EmitLoadOfScalar( 2749 AddrReduceListArg, /*Volatile=*/false, C.VoidPtrTy, Loc); 2750 CGM.getOpenMPRuntime().emitOutlinedFunctionCall( 2751 CGF, Loc, ReduceFn, {GlobalReduceList, ReducedPtr}); 2752 CGF.FinishFunction(); 2753 return Fn; 2754 } 2755 2756 /// This function emits a helper that copies all the reduction variables from 2757 /// the team into the provided global buffer for the reduction variables. 2758 /// 2759 /// void list_to_global_copy_func(void *buffer, int Idx, void *reduce_data) 2760 /// For all data entries D in reduce_data: 2761 /// Copy buffer.D[Idx] to local D; 2762 static llvm::Value *emitGlobalToListCopyFunction( 2763 CodeGenModule &CGM, ArrayRef<const Expr *> Privates, 2764 QualType ReductionArrayTy, SourceLocation Loc, 2765 const RecordDecl *TeamReductionRec, 2766 const llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> 2767 &VarFieldMap) { 2768 ASTContext &C = CGM.getContext(); 2769 2770 // Buffer: global reduction buffer. 2771 ImplicitParamDecl BufferArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, 2772 C.VoidPtrTy, ImplicitParamDecl::Other); 2773 // Idx: index of the buffer. 2774 ImplicitParamDecl IdxArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.IntTy, 2775 ImplicitParamDecl::Other); 2776 // ReduceList: thread local Reduce list. 2777 ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, 2778 C.VoidPtrTy, ImplicitParamDecl::Other); 2779 FunctionArgList Args; 2780 Args.push_back(&BufferArg); 2781 Args.push_back(&IdxArg); 2782 Args.push_back(&ReduceListArg); 2783 2784 const CGFunctionInfo &CGFI = 2785 CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); 2786 auto *Fn = llvm::Function::Create( 2787 CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage, 2788 "_omp_reduction_global_to_list_copy_func", &CGM.getModule()); 2789 CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI); 2790 Fn->setDoesNotRecurse(); 2791 CodeGenFunction CGF(CGM); 2792 CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc); 2793 2794 CGBuilderTy &Bld = CGF.Builder; 2795 2796 Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg); 2797 Address AddrBufferArg = CGF.GetAddrOfLocalVar(&BufferArg); 2798 llvm::Type *ElemTy = CGF.ConvertTypeForMem(ReductionArrayTy); 2799 Address LocalReduceList( 2800 Bld.CreatePointerBitCastOrAddrSpaceCast( 2801 CGF.EmitLoadOfScalar(AddrReduceListArg, /*Volatile=*/false, 2802 C.VoidPtrTy, Loc), 2803 ElemTy->getPointerTo()), 2804 ElemTy, CGF.getPointerAlign()); 2805 QualType StaticTy = C.getRecordType(TeamReductionRec); 2806 llvm::Type *LLVMReductionsBufferTy = 2807 CGM.getTypes().ConvertTypeForMem(StaticTy); 2808 llvm::Value *BufferArrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( 2809 CGF.EmitLoadOfScalar(AddrBufferArg, /*Volatile=*/false, C.VoidPtrTy, Loc), 2810 LLVMReductionsBufferTy->getPointerTo()); 2811 2812 llvm::Value *Idxs[] = {llvm::ConstantInt::getNullValue(CGF.Int32Ty), 2813 CGF.EmitLoadOfScalar(CGF.GetAddrOfLocalVar(&IdxArg), 2814 /*Volatile=*/false, C.IntTy, 2815 Loc)}; 2816 unsigned Idx = 0; 2817 for (const Expr *Private : Privates) { 2818 // Reduce element = LocalReduceList[i] 2819 Address ElemPtrPtrAddr = Bld.CreateConstArrayGEP(LocalReduceList, Idx); 2820 llvm::Value *ElemPtrPtr = CGF.EmitLoadOfScalar( 2821 ElemPtrPtrAddr, /*Volatile=*/false, C.VoidPtrTy, SourceLocation()); 2822 // elemptr = ((CopyType*)(elemptrptr)) + I 2823 ElemTy = CGF.ConvertTypeForMem(Private->getType()); 2824 ElemPtrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( 2825 ElemPtrPtr, ElemTy->getPointerTo()); 2826 Address ElemPtr = 2827 Address(ElemPtrPtr, ElemTy, C.getTypeAlignInChars(Private->getType())); 2828 const ValueDecl *VD = cast<DeclRefExpr>(Private)->getDecl(); 2829 // Global = Buffer.VD[Idx]; 2830 const FieldDecl *FD = VarFieldMap.lookup(VD); 2831 LValue GlobLVal = CGF.EmitLValueForField( 2832 CGF.MakeNaturalAlignAddrLValue(BufferArrPtr, StaticTy), FD); 2833 Address GlobAddr = GlobLVal.getAddress(CGF); 2834 llvm::Value *BufferPtr = Bld.CreateInBoundsGEP(GlobAddr.getElementType(), 2835 GlobAddr.getPointer(), Idxs); 2836 GlobLVal.setAddress(Address(BufferPtr, 2837 CGF.ConvertTypeForMem(Private->getType()), 2838 GlobAddr.getAlignment())); 2839 switch (CGF.getEvaluationKind(Private->getType())) { 2840 case TEK_Scalar: { 2841 llvm::Value *V = CGF.EmitLoadOfScalar(GlobLVal, Loc); 2842 CGF.EmitStoreOfScalar(V, ElemPtr, /*Volatile=*/false, Private->getType(), 2843 LValueBaseInfo(AlignmentSource::Type), 2844 TBAAAccessInfo()); 2845 break; 2846 } 2847 case TEK_Complex: { 2848 CodeGenFunction::ComplexPairTy V = CGF.EmitLoadOfComplex(GlobLVal, Loc); 2849 CGF.EmitStoreOfComplex(V, CGF.MakeAddrLValue(ElemPtr, Private->getType()), 2850 /*isInit=*/false); 2851 break; 2852 } 2853 case TEK_Aggregate: 2854 CGF.EmitAggregateCopy(CGF.MakeAddrLValue(ElemPtr, Private->getType()), 2855 GlobLVal, Private->getType(), 2856 AggValueSlot::DoesNotOverlap); 2857 break; 2858 } 2859 ++Idx; 2860 } 2861 2862 CGF.FinishFunction(); 2863 return Fn; 2864 } 2865 2866 /// This function emits a helper that reduces all the reduction variables from 2867 /// the team into the provided global buffer for the reduction variables. 2868 /// 2869 /// void global_to_list_reduce_func(void *buffer, int Idx, void *reduce_data) 2870 /// void *GlobPtrs[]; 2871 /// GlobPtrs[0] = (void*)&buffer.D0[Idx]; 2872 /// ... 2873 /// GlobPtrs[N] = (void*)&buffer.DN[Idx]; 2874 /// reduce_function(reduce_data, GlobPtrs); 2875 static llvm::Value *emitGlobalToListReduceFunction( 2876 CodeGenModule &CGM, ArrayRef<const Expr *> Privates, 2877 QualType ReductionArrayTy, SourceLocation Loc, 2878 const RecordDecl *TeamReductionRec, 2879 const llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> 2880 &VarFieldMap, 2881 llvm::Function *ReduceFn) { 2882 ASTContext &C = CGM.getContext(); 2883 2884 // Buffer: global reduction buffer. 2885 ImplicitParamDecl BufferArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, 2886 C.VoidPtrTy, ImplicitParamDecl::Other); 2887 // Idx: index of the buffer. 2888 ImplicitParamDecl IdxArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.IntTy, 2889 ImplicitParamDecl::Other); 2890 // ReduceList: thread local Reduce list. 2891 ImplicitParamDecl ReduceListArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, 2892 C.VoidPtrTy, ImplicitParamDecl::Other); 2893 FunctionArgList Args; 2894 Args.push_back(&BufferArg); 2895 Args.push_back(&IdxArg); 2896 Args.push_back(&ReduceListArg); 2897 2898 const CGFunctionInfo &CGFI = 2899 CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); 2900 auto *Fn = llvm::Function::Create( 2901 CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage, 2902 "_omp_reduction_global_to_list_reduce_func", &CGM.getModule()); 2903 CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI); 2904 Fn->setDoesNotRecurse(); 2905 CodeGenFunction CGF(CGM); 2906 CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc); 2907 2908 CGBuilderTy &Bld = CGF.Builder; 2909 2910 Address AddrBufferArg = CGF.GetAddrOfLocalVar(&BufferArg); 2911 QualType StaticTy = C.getRecordType(TeamReductionRec); 2912 llvm::Type *LLVMReductionsBufferTy = 2913 CGM.getTypes().ConvertTypeForMem(StaticTy); 2914 llvm::Value *BufferArrPtr = Bld.CreatePointerBitCastOrAddrSpaceCast( 2915 CGF.EmitLoadOfScalar(AddrBufferArg, /*Volatile=*/false, C.VoidPtrTy, Loc), 2916 LLVMReductionsBufferTy->getPointerTo()); 2917 2918 // 1. Build a list of reduction variables. 2919 // void *RedList[<n>] = {<ReductionVars>[0], ..., <ReductionVars>[<n>-1]}; 2920 Address ReductionList = 2921 CGF.CreateMemTemp(ReductionArrayTy, ".omp.reduction.red_list"); 2922 auto IPriv = Privates.begin(); 2923 llvm::Value *Idxs[] = {llvm::ConstantInt::getNullValue(CGF.Int32Ty), 2924 CGF.EmitLoadOfScalar(CGF.GetAddrOfLocalVar(&IdxArg), 2925 /*Volatile=*/false, C.IntTy, 2926 Loc)}; 2927 unsigned Idx = 0; 2928 for (unsigned I = 0, E = Privates.size(); I < E; ++I, ++IPriv, ++Idx) { 2929 Address Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx); 2930 // Global = Buffer.VD[Idx]; 2931 const ValueDecl *VD = cast<DeclRefExpr>(*IPriv)->getDecl(); 2932 const FieldDecl *FD = VarFieldMap.lookup(VD); 2933 LValue GlobLVal = CGF.EmitLValueForField( 2934 CGF.MakeNaturalAlignAddrLValue(BufferArrPtr, StaticTy), FD); 2935 Address GlobAddr = GlobLVal.getAddress(CGF); 2936 llvm::Value *BufferPtr = Bld.CreateInBoundsGEP( 2937 GlobAddr.getElementType(), GlobAddr.getPointer(), Idxs); 2938 llvm::Value *Ptr = CGF.EmitCastToVoidPtr(BufferPtr); 2939 CGF.EmitStoreOfScalar(Ptr, Elem, /*Volatile=*/false, C.VoidPtrTy); 2940 if ((*IPriv)->getType()->isVariablyModifiedType()) { 2941 // Store array size. 2942 ++Idx; 2943 Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx); 2944 llvm::Value *Size = CGF.Builder.CreateIntCast( 2945 CGF.getVLASize( 2946 CGF.getContext().getAsVariableArrayType((*IPriv)->getType())) 2947 .NumElts, 2948 CGF.SizeTy, /*isSigned=*/false); 2949 CGF.Builder.CreateStore(CGF.Builder.CreateIntToPtr(Size, CGF.VoidPtrTy), 2950 Elem); 2951 } 2952 } 2953 2954 // Call reduce_function(ReduceList, GlobalReduceList) 2955 llvm::Value *GlobalReduceList = 2956 CGF.EmitCastToVoidPtr(ReductionList.getPointer()); 2957 Address AddrReduceListArg = CGF.GetAddrOfLocalVar(&ReduceListArg); 2958 llvm::Value *ReducedPtr = CGF.EmitLoadOfScalar( 2959 AddrReduceListArg, /*Volatile=*/false, C.VoidPtrTy, Loc); 2960 CGM.getOpenMPRuntime().emitOutlinedFunctionCall( 2961 CGF, Loc, ReduceFn, {ReducedPtr, GlobalReduceList}); 2962 CGF.FinishFunction(); 2963 return Fn; 2964 } 2965 2966 /// 2967 /// Design of OpenMP reductions on the GPU 2968 /// 2969 /// Consider a typical OpenMP program with one or more reduction 2970 /// clauses: 2971 /// 2972 /// float foo; 2973 /// double bar; 2974 /// #pragma omp target teams distribute parallel for \ 2975 /// reduction(+:foo) reduction(*:bar) 2976 /// for (int i = 0; i < N; i++) { 2977 /// foo += A[i]; bar *= B[i]; 2978 /// } 2979 /// 2980 /// where 'foo' and 'bar' are reduced across all OpenMP threads in 2981 /// all teams. In our OpenMP implementation on the NVPTX device an 2982 /// OpenMP team is mapped to a CUDA threadblock and OpenMP threads 2983 /// within a team are mapped to CUDA threads within a threadblock. 2984 /// Our goal is to efficiently aggregate values across all OpenMP 2985 /// threads such that: 2986 /// 2987 /// - the compiler and runtime are logically concise, and 2988 /// - the reduction is performed efficiently in a hierarchical 2989 /// manner as follows: within OpenMP threads in the same warp, 2990 /// across warps in a threadblock, and finally across teams on 2991 /// the NVPTX device. 2992 /// 2993 /// Introduction to Decoupling 2994 /// 2995 /// We would like to decouple the compiler and the runtime so that the 2996 /// latter is ignorant of the reduction variables (number, data types) 2997 /// and the reduction operators. This allows a simpler interface 2998 /// and implementation while still attaining good performance. 2999 /// 3000 /// Pseudocode for the aforementioned OpenMP program generated by the 3001 /// compiler is as follows: 3002 /// 3003 /// 1. Create private copies of reduction variables on each OpenMP 3004 /// thread: 'foo_private', 'bar_private' 3005 /// 2. Each OpenMP thread reduces the chunk of 'A' and 'B' assigned 3006 /// to it and writes the result in 'foo_private' and 'bar_private' 3007 /// respectively. 3008 /// 3. Call the OpenMP runtime on the GPU to reduce within a team 3009 /// and store the result on the team master: 3010 /// 3011 /// __kmpc_nvptx_parallel_reduce_nowait_v2(..., 3012 /// reduceData, shuffleReduceFn, interWarpCpyFn) 3013 /// 3014 /// where: 3015 /// struct ReduceData { 3016 /// double *foo; 3017 /// double *bar; 3018 /// } reduceData 3019 /// reduceData.foo = &foo_private 3020 /// reduceData.bar = &bar_private 3021 /// 3022 /// 'shuffleReduceFn' and 'interWarpCpyFn' are pointers to two 3023 /// auxiliary functions generated by the compiler that operate on 3024 /// variables of type 'ReduceData'. They aid the runtime perform 3025 /// algorithmic steps in a data agnostic manner. 3026 /// 3027 /// 'shuffleReduceFn' is a pointer to a function that reduces data 3028 /// of type 'ReduceData' across two OpenMP threads (lanes) in the 3029 /// same warp. It takes the following arguments as input: 3030 /// 3031 /// a. variable of type 'ReduceData' on the calling lane, 3032 /// b. its lane_id, 3033 /// c. an offset relative to the current lane_id to generate a 3034 /// remote_lane_id. The remote lane contains the second 3035 /// variable of type 'ReduceData' that is to be reduced. 3036 /// d. an algorithm version parameter determining which reduction 3037 /// algorithm to use. 3038 /// 3039 /// 'shuffleReduceFn' retrieves data from the remote lane using 3040 /// efficient GPU shuffle intrinsics and reduces, using the 3041 /// algorithm specified by the 4th parameter, the two operands 3042 /// element-wise. The result is written to the first operand. 3043 /// 3044 /// Different reduction algorithms are implemented in different 3045 /// runtime functions, all calling 'shuffleReduceFn' to perform 3046 /// the essential reduction step. Therefore, based on the 4th 3047 /// parameter, this function behaves slightly differently to 3048 /// cooperate with the runtime to ensure correctness under 3049 /// different circumstances. 3050 /// 3051 /// 'InterWarpCpyFn' is a pointer to a function that transfers 3052 /// reduced variables across warps. It tunnels, through CUDA 3053 /// shared memory, the thread-private data of type 'ReduceData' 3054 /// from lane 0 of each warp to a lane in the first warp. 3055 /// 4. Call the OpenMP runtime on the GPU to reduce across teams. 3056 /// The last team writes the global reduced value to memory. 3057 /// 3058 /// ret = __kmpc_nvptx_teams_reduce_nowait(..., 3059 /// reduceData, shuffleReduceFn, interWarpCpyFn, 3060 /// scratchpadCopyFn, loadAndReduceFn) 3061 /// 3062 /// 'scratchpadCopyFn' is a helper that stores reduced 3063 /// data from the team master to a scratchpad array in 3064 /// global memory. 3065 /// 3066 /// 'loadAndReduceFn' is a helper that loads data from 3067 /// the scratchpad array and reduces it with the input 3068 /// operand. 3069 /// 3070 /// These compiler generated functions hide address 3071 /// calculation and alignment information from the runtime. 3072 /// 5. if ret == 1: 3073 /// The team master of the last team stores the reduced 3074 /// result to the globals in memory. 3075 /// foo += reduceData.foo; bar *= reduceData.bar 3076 /// 3077 /// 3078 /// Warp Reduction Algorithms 3079 /// 3080 /// On the warp level, we have three algorithms implemented in the 3081 /// OpenMP runtime depending on the number of active lanes: 3082 /// 3083 /// Full Warp Reduction 3084 /// 3085 /// The reduce algorithm within a warp where all lanes are active 3086 /// is implemented in the runtime as follows: 3087 /// 3088 /// full_warp_reduce(void *reduce_data, 3089 /// kmp_ShuffleReductFctPtr ShuffleReduceFn) { 3090 /// for (int offset = WARPSIZE/2; offset > 0; offset /= 2) 3091 /// ShuffleReduceFn(reduce_data, 0, offset, 0); 3092 /// } 3093 /// 3094 /// The algorithm completes in log(2, WARPSIZE) steps. 3095 /// 3096 /// 'ShuffleReduceFn' is used here with lane_id set to 0 because it is 3097 /// not used therefore we save instructions by not retrieving lane_id 3098 /// from the corresponding special registers. The 4th parameter, which 3099 /// represents the version of the algorithm being used, is set to 0 to 3100 /// signify full warp reduction. 3101 /// 3102 /// In this version, 'ShuffleReduceFn' behaves, per element, as follows: 3103 /// 3104 /// #reduce_elem refers to an element in the local lane's data structure 3105 /// #remote_elem is retrieved from a remote lane 3106 /// remote_elem = shuffle_down(reduce_elem, offset, WARPSIZE); 3107 /// reduce_elem = reduce_elem REDUCE_OP remote_elem; 3108 /// 3109 /// Contiguous Partial Warp Reduction 3110 /// 3111 /// This reduce algorithm is used within a warp where only the first 3112 /// 'n' (n <= WARPSIZE) lanes are active. It is typically used when the 3113 /// number of OpenMP threads in a parallel region is not a multiple of 3114 /// WARPSIZE. The algorithm is implemented in the runtime as follows: 3115 /// 3116 /// void 3117 /// contiguous_partial_reduce(void *reduce_data, 3118 /// kmp_ShuffleReductFctPtr ShuffleReduceFn, 3119 /// int size, int lane_id) { 3120 /// int curr_size; 3121 /// int offset; 3122 /// curr_size = size; 3123 /// mask = curr_size/2; 3124 /// while (offset>0) { 3125 /// ShuffleReduceFn(reduce_data, lane_id, offset, 1); 3126 /// curr_size = (curr_size+1)/2; 3127 /// offset = curr_size/2; 3128 /// } 3129 /// } 3130 /// 3131 /// In this version, 'ShuffleReduceFn' behaves, per element, as follows: 3132 /// 3133 /// remote_elem = shuffle_down(reduce_elem, offset, WARPSIZE); 3134 /// if (lane_id < offset) 3135 /// reduce_elem = reduce_elem REDUCE_OP remote_elem 3136 /// else 3137 /// reduce_elem = remote_elem 3138 /// 3139 /// This algorithm assumes that the data to be reduced are located in a 3140 /// contiguous subset of lanes starting from the first. When there is 3141 /// an odd number of active lanes, the data in the last lane is not 3142 /// aggregated with any other lane's dat but is instead copied over. 3143 /// 3144 /// Dispersed Partial Warp Reduction 3145 /// 3146 /// This algorithm is used within a warp when any discontiguous subset of 3147 /// lanes are active. It is used to implement the reduction operation 3148 /// across lanes in an OpenMP simd region or in a nested parallel region. 3149 /// 3150 /// void 3151 /// dispersed_partial_reduce(void *reduce_data, 3152 /// kmp_ShuffleReductFctPtr ShuffleReduceFn) { 3153 /// int size, remote_id; 3154 /// int logical_lane_id = number_of_active_lanes_before_me() * 2; 3155 /// do { 3156 /// remote_id = next_active_lane_id_right_after_me(); 3157 /// # the above function returns 0 of no active lane 3158 /// # is present right after the current lane. 3159 /// size = number_of_active_lanes_in_this_warp(); 3160 /// logical_lane_id /= 2; 3161 /// ShuffleReduceFn(reduce_data, logical_lane_id, 3162 /// remote_id-1-threadIdx.x, 2); 3163 /// } while (logical_lane_id % 2 == 0 && size > 1); 3164 /// } 3165 /// 3166 /// There is no assumption made about the initial state of the reduction. 3167 /// Any number of lanes (>=1) could be active at any position. The reduction 3168 /// result is returned in the first active lane. 3169 /// 3170 /// In this version, 'ShuffleReduceFn' behaves, per element, as follows: 3171 /// 3172 /// remote_elem = shuffle_down(reduce_elem, offset, WARPSIZE); 3173 /// if (lane_id % 2 == 0 && offset > 0) 3174 /// reduce_elem = reduce_elem REDUCE_OP remote_elem 3175 /// else 3176 /// reduce_elem = remote_elem 3177 /// 3178 /// 3179 /// Intra-Team Reduction 3180 /// 3181 /// This function, as implemented in the runtime call 3182 /// '__kmpc_nvptx_parallel_reduce_nowait_v2', aggregates data across OpenMP 3183 /// threads in a team. It first reduces within a warp using the 3184 /// aforementioned algorithms. We then proceed to gather all such 3185 /// reduced values at the first warp. 3186 /// 3187 /// The runtime makes use of the function 'InterWarpCpyFn', which copies 3188 /// data from each of the "warp master" (zeroth lane of each warp, where 3189 /// warp-reduced data is held) to the zeroth warp. This step reduces (in 3190 /// a mathematical sense) the problem of reduction across warp masters in 3191 /// a block to the problem of warp reduction. 3192 /// 3193 /// 3194 /// Inter-Team Reduction 3195 /// 3196 /// Once a team has reduced its data to a single value, it is stored in 3197 /// a global scratchpad array. Since each team has a distinct slot, this 3198 /// can be done without locking. 3199 /// 3200 /// The last team to write to the scratchpad array proceeds to reduce the 3201 /// scratchpad array. One or more workers in the last team use the helper 3202 /// 'loadAndReduceDataFn' to load and reduce values from the array, i.e., 3203 /// the k'th worker reduces every k'th element. 3204 /// 3205 /// Finally, a call is made to '__kmpc_nvptx_parallel_reduce_nowait_v2' to 3206 /// reduce across workers and compute a globally reduced value. 3207 /// 3208 void CGOpenMPRuntimeGPU::emitReduction( 3209 CodeGenFunction &CGF, SourceLocation Loc, ArrayRef<const Expr *> Privates, 3210 ArrayRef<const Expr *> LHSExprs, ArrayRef<const Expr *> RHSExprs, 3211 ArrayRef<const Expr *> ReductionOps, ReductionOptionsTy Options) { 3212 if (!CGF.HaveInsertPoint()) 3213 return; 3214 3215 bool ParallelReduction = isOpenMPParallelDirective(Options.ReductionKind); 3216 #ifndef NDEBUG 3217 bool TeamsReduction = isOpenMPTeamsDirective(Options.ReductionKind); 3218 #endif 3219 3220 if (Options.SimpleReduction) { 3221 assert(!TeamsReduction && !ParallelReduction && 3222 "Invalid reduction selection in emitReduction."); 3223 CGOpenMPRuntime::emitReduction(CGF, Loc, Privates, LHSExprs, RHSExprs, 3224 ReductionOps, Options); 3225 return; 3226 } 3227 3228 assert((TeamsReduction || ParallelReduction) && 3229 "Invalid reduction selection in emitReduction."); 3230 3231 // Build res = __kmpc_reduce{_nowait}(<gtid>, <n>, sizeof(RedList), 3232 // RedList, shuffle_reduce_func, interwarp_copy_func); 3233 // or 3234 // Build res = __kmpc_reduce_teams_nowait_simple(<loc>, <gtid>, <lck>); 3235 llvm::Value *RTLoc = emitUpdateLocation(CGF, Loc); 3236 llvm::Value *ThreadId = getThreadID(CGF, Loc); 3237 3238 llvm::Value *Res; 3239 ASTContext &C = CGM.getContext(); 3240 // 1. Build a list of reduction variables. 3241 // void *RedList[<n>] = {<ReductionVars>[0], ..., <ReductionVars>[<n>-1]}; 3242 auto Size = RHSExprs.size(); 3243 for (const Expr *E : Privates) { 3244 if (E->getType()->isVariablyModifiedType()) 3245 // Reserve place for array size. 3246 ++Size; 3247 } 3248 llvm::APInt ArraySize(/*unsigned int numBits=*/32, Size); 3249 QualType ReductionArrayTy = 3250 C.getConstantArrayType(C.VoidPtrTy, ArraySize, nullptr, ArrayType::Normal, 3251 /*IndexTypeQuals=*/0); 3252 Address ReductionList = 3253 CGF.CreateMemTemp(ReductionArrayTy, ".omp.reduction.red_list"); 3254 auto IPriv = Privates.begin(); 3255 unsigned Idx = 0; 3256 for (unsigned I = 0, E = RHSExprs.size(); I < E; ++I, ++IPriv, ++Idx) { 3257 Address Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx); 3258 CGF.Builder.CreateStore( 3259 CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( 3260 CGF.EmitLValue(RHSExprs[I]).getPointer(CGF), CGF.VoidPtrTy), 3261 Elem); 3262 if ((*IPriv)->getType()->isVariablyModifiedType()) { 3263 // Store array size. 3264 ++Idx; 3265 Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx); 3266 llvm::Value *Size = CGF.Builder.CreateIntCast( 3267 CGF.getVLASize( 3268 CGF.getContext().getAsVariableArrayType((*IPriv)->getType())) 3269 .NumElts, 3270 CGF.SizeTy, /*isSigned=*/false); 3271 CGF.Builder.CreateStore(CGF.Builder.CreateIntToPtr(Size, CGF.VoidPtrTy), 3272 Elem); 3273 } 3274 } 3275 3276 llvm::Value *RL = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( 3277 ReductionList.getPointer(), CGF.VoidPtrTy); 3278 llvm::Function *ReductionFn = 3279 emitReductionFunction(Loc, CGF.ConvertTypeForMem(ReductionArrayTy), 3280 Privates, LHSExprs, RHSExprs, ReductionOps); 3281 llvm::Value *ReductionArrayTySize = CGF.getTypeSize(ReductionArrayTy); 3282 llvm::Function *ShuffleAndReduceFn = emitShuffleAndReduceFunction( 3283 CGM, Privates, ReductionArrayTy, ReductionFn, Loc); 3284 llvm::Value *InterWarpCopyFn = 3285 emitInterWarpCopyFunction(CGM, Privates, ReductionArrayTy, Loc); 3286 3287 if (ParallelReduction) { 3288 llvm::Value *Args[] = {RTLoc, 3289 ThreadId, 3290 CGF.Builder.getInt32(RHSExprs.size()), 3291 ReductionArrayTySize, 3292 RL, 3293 ShuffleAndReduceFn, 3294 InterWarpCopyFn}; 3295 3296 Res = CGF.EmitRuntimeCall( 3297 OMPBuilder.getOrCreateRuntimeFunction( 3298 CGM.getModule(), OMPRTL___kmpc_nvptx_parallel_reduce_nowait_v2), 3299 Args); 3300 } else { 3301 assert(TeamsReduction && "expected teams reduction."); 3302 llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *> VarFieldMap; 3303 llvm::SmallVector<const ValueDecl *, 4> PrivatesReductions(Privates.size()); 3304 int Cnt = 0; 3305 for (const Expr *DRE : Privates) { 3306 PrivatesReductions[Cnt] = cast<DeclRefExpr>(DRE)->getDecl(); 3307 ++Cnt; 3308 } 3309 const RecordDecl *TeamReductionRec = ::buildRecordForGlobalizedVars( 3310 CGM.getContext(), PrivatesReductions, llvm::None, VarFieldMap, 3311 C.getLangOpts().OpenMPCUDAReductionBufNum); 3312 TeamsReductions.push_back(TeamReductionRec); 3313 if (!KernelTeamsReductionPtr) { 3314 KernelTeamsReductionPtr = new llvm::GlobalVariable( 3315 CGM.getModule(), CGM.VoidPtrTy, /*isConstant=*/true, 3316 llvm::GlobalValue::InternalLinkage, nullptr, 3317 "_openmp_teams_reductions_buffer_$_$ptr"); 3318 } 3319 llvm::Value *GlobalBufferPtr = CGF.EmitLoadOfScalar( 3320 Address(KernelTeamsReductionPtr, CGF.VoidPtrTy, CGM.getPointerAlign()), 3321 /*Volatile=*/false, C.getPointerType(C.VoidPtrTy), Loc); 3322 llvm::Value *GlobalToBufferCpyFn = ::emitListToGlobalCopyFunction( 3323 CGM, Privates, ReductionArrayTy, Loc, TeamReductionRec, VarFieldMap); 3324 llvm::Value *GlobalToBufferRedFn = ::emitListToGlobalReduceFunction( 3325 CGM, Privates, ReductionArrayTy, Loc, TeamReductionRec, VarFieldMap, 3326 ReductionFn); 3327 llvm::Value *BufferToGlobalCpyFn = ::emitGlobalToListCopyFunction( 3328 CGM, Privates, ReductionArrayTy, Loc, TeamReductionRec, VarFieldMap); 3329 llvm::Value *BufferToGlobalRedFn = ::emitGlobalToListReduceFunction( 3330 CGM, Privates, ReductionArrayTy, Loc, TeamReductionRec, VarFieldMap, 3331 ReductionFn); 3332 3333 llvm::Value *Args[] = { 3334 RTLoc, 3335 ThreadId, 3336 GlobalBufferPtr, 3337 CGF.Builder.getInt32(C.getLangOpts().OpenMPCUDAReductionBufNum), 3338 RL, 3339 ShuffleAndReduceFn, 3340 InterWarpCopyFn, 3341 GlobalToBufferCpyFn, 3342 GlobalToBufferRedFn, 3343 BufferToGlobalCpyFn, 3344 BufferToGlobalRedFn}; 3345 3346 Res = CGF.EmitRuntimeCall( 3347 OMPBuilder.getOrCreateRuntimeFunction( 3348 CGM.getModule(), OMPRTL___kmpc_nvptx_teams_reduce_nowait_v2), 3349 Args); 3350 } 3351 3352 // 5. Build if (res == 1) 3353 llvm::BasicBlock *ExitBB = CGF.createBasicBlock(".omp.reduction.done"); 3354 llvm::BasicBlock *ThenBB = CGF.createBasicBlock(".omp.reduction.then"); 3355 llvm::Value *Cond = CGF.Builder.CreateICmpEQ( 3356 Res, llvm::ConstantInt::get(CGM.Int32Ty, /*V=*/1)); 3357 CGF.Builder.CreateCondBr(Cond, ThenBB, ExitBB); 3358 3359 // 6. Build then branch: where we have reduced values in the master 3360 // thread in each team. 3361 // __kmpc_end_reduce{_nowait}(<gtid>); 3362 // break; 3363 CGF.EmitBlock(ThenBB); 3364 3365 // Add emission of __kmpc_end_reduce{_nowait}(<gtid>); 3366 auto &&CodeGen = [Privates, LHSExprs, RHSExprs, ReductionOps, 3367 this](CodeGenFunction &CGF, PrePostActionTy &Action) { 3368 auto IPriv = Privates.begin(); 3369 auto ILHS = LHSExprs.begin(); 3370 auto IRHS = RHSExprs.begin(); 3371 for (const Expr *E : ReductionOps) { 3372 emitSingleReductionCombiner(CGF, E, *IPriv, cast<DeclRefExpr>(*ILHS), 3373 cast<DeclRefExpr>(*IRHS)); 3374 ++IPriv; 3375 ++ILHS; 3376 ++IRHS; 3377 } 3378 }; 3379 llvm::Value *EndArgs[] = {ThreadId}; 3380 RegionCodeGenTy RCG(CodeGen); 3381 NVPTXActionTy Action( 3382 nullptr, llvm::None, 3383 OMPBuilder.getOrCreateRuntimeFunction( 3384 CGM.getModule(), OMPRTL___kmpc_nvptx_end_reduce_nowait), 3385 EndArgs); 3386 RCG.setAction(Action); 3387 RCG(CGF); 3388 // There is no need to emit line number for unconditional branch. 3389 (void)ApplyDebugLocation::CreateEmpty(CGF); 3390 CGF.EmitBlock(ExitBB, /*IsFinished=*/true); 3391 } 3392 3393 const VarDecl * 3394 CGOpenMPRuntimeGPU::translateParameter(const FieldDecl *FD, 3395 const VarDecl *NativeParam) const { 3396 if (!NativeParam->getType()->isReferenceType()) 3397 return NativeParam; 3398 QualType ArgType = NativeParam->getType(); 3399 QualifierCollector QC; 3400 const Type *NonQualTy = QC.strip(ArgType); 3401 QualType PointeeTy = cast<ReferenceType>(NonQualTy)->getPointeeType(); 3402 if (const auto *Attr = FD->getAttr<OMPCaptureKindAttr>()) { 3403 if (Attr->getCaptureKind() == OMPC_map) { 3404 PointeeTy = CGM.getContext().getAddrSpaceQualType(PointeeTy, 3405 LangAS::opencl_global); 3406 } 3407 } 3408 ArgType = CGM.getContext().getPointerType(PointeeTy); 3409 QC.addRestrict(); 3410 enum { NVPTX_local_addr = 5 }; 3411 QC.addAddressSpace(getLangASFromTargetAS(NVPTX_local_addr)); 3412 ArgType = QC.apply(CGM.getContext(), ArgType); 3413 if (isa<ImplicitParamDecl>(NativeParam)) 3414 return ImplicitParamDecl::Create( 3415 CGM.getContext(), /*DC=*/nullptr, NativeParam->getLocation(), 3416 NativeParam->getIdentifier(), ArgType, ImplicitParamDecl::Other); 3417 return ParmVarDecl::Create( 3418 CGM.getContext(), 3419 const_cast<DeclContext *>(NativeParam->getDeclContext()), 3420 NativeParam->getBeginLoc(), NativeParam->getLocation(), 3421 NativeParam->getIdentifier(), ArgType, 3422 /*TInfo=*/nullptr, SC_None, /*DefArg=*/nullptr); 3423 } 3424 3425 Address 3426 CGOpenMPRuntimeGPU::getParameterAddress(CodeGenFunction &CGF, 3427 const VarDecl *NativeParam, 3428 const VarDecl *TargetParam) const { 3429 assert(NativeParam != TargetParam && 3430 NativeParam->getType()->isReferenceType() && 3431 "Native arg must not be the same as target arg."); 3432 Address LocalAddr = CGF.GetAddrOfLocalVar(TargetParam); 3433 QualType NativeParamType = NativeParam->getType(); 3434 QualifierCollector QC; 3435 const Type *NonQualTy = QC.strip(NativeParamType); 3436 QualType NativePointeeTy = cast<ReferenceType>(NonQualTy)->getPointeeType(); 3437 unsigned NativePointeeAddrSpace = 3438 CGF.getContext().getTargetAddressSpace(NativePointeeTy); 3439 QualType TargetTy = TargetParam->getType(); 3440 llvm::Value *TargetAddr = CGF.EmitLoadOfScalar( 3441 LocalAddr, /*Volatile=*/false, TargetTy, SourceLocation()); 3442 // First cast to generic. 3443 TargetAddr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( 3444 TargetAddr, llvm::PointerType::getWithSamePointeeType( 3445 cast<llvm::PointerType>(TargetAddr->getType()), /*AddrSpace=*/0)); 3446 // Cast from generic to native address space. 3447 TargetAddr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( 3448 TargetAddr, llvm::PointerType::getWithSamePointeeType( 3449 cast<llvm::PointerType>(TargetAddr->getType()), 3450 NativePointeeAddrSpace)); 3451 Address NativeParamAddr = CGF.CreateMemTemp(NativeParamType); 3452 CGF.EmitStoreOfScalar(TargetAddr, NativeParamAddr, /*Volatile=*/false, 3453 NativeParamType); 3454 return NativeParamAddr; 3455 } 3456 3457 void CGOpenMPRuntimeGPU::emitOutlinedFunctionCall( 3458 CodeGenFunction &CGF, SourceLocation Loc, llvm::FunctionCallee OutlinedFn, 3459 ArrayRef<llvm::Value *> Args) const { 3460 SmallVector<llvm::Value *, 4> TargetArgs; 3461 TargetArgs.reserve(Args.size()); 3462 auto *FnType = OutlinedFn.getFunctionType(); 3463 for (unsigned I = 0, E = Args.size(); I < E; ++I) { 3464 if (FnType->isVarArg() && FnType->getNumParams() <= I) { 3465 TargetArgs.append(std::next(Args.begin(), I), Args.end()); 3466 break; 3467 } 3468 llvm::Type *TargetType = FnType->getParamType(I); 3469 llvm::Value *NativeArg = Args[I]; 3470 if (!TargetType->isPointerTy()) { 3471 TargetArgs.emplace_back(NativeArg); 3472 continue; 3473 } 3474 llvm::Value *TargetArg = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( 3475 NativeArg, llvm::PointerType::getWithSamePointeeType( 3476 cast<llvm::PointerType>(NativeArg->getType()), /*AddrSpace*/ 0)); 3477 TargetArgs.emplace_back( 3478 CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(TargetArg, TargetType)); 3479 } 3480 CGOpenMPRuntime::emitOutlinedFunctionCall(CGF, Loc, OutlinedFn, TargetArgs); 3481 } 3482 3483 /// Emit function which wraps the outline parallel region 3484 /// and controls the arguments which are passed to this function. 3485 /// The wrapper ensures that the outlined function is called 3486 /// with the correct arguments when data is shared. 3487 llvm::Function *CGOpenMPRuntimeGPU::createParallelDataSharingWrapper( 3488 llvm::Function *OutlinedParallelFn, const OMPExecutableDirective &D) { 3489 ASTContext &Ctx = CGM.getContext(); 3490 const auto &CS = *D.getCapturedStmt(OMPD_parallel); 3491 3492 // Create a function that takes as argument the source thread. 3493 FunctionArgList WrapperArgs; 3494 QualType Int16QTy = 3495 Ctx.getIntTypeForBitwidth(/*DestWidth=*/16, /*Signed=*/false); 3496 QualType Int32QTy = 3497 Ctx.getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/false); 3498 ImplicitParamDecl ParallelLevelArg(Ctx, /*DC=*/nullptr, D.getBeginLoc(), 3499 /*Id=*/nullptr, Int16QTy, 3500 ImplicitParamDecl::Other); 3501 ImplicitParamDecl WrapperArg(Ctx, /*DC=*/nullptr, D.getBeginLoc(), 3502 /*Id=*/nullptr, Int32QTy, 3503 ImplicitParamDecl::Other); 3504 WrapperArgs.emplace_back(&ParallelLevelArg); 3505 WrapperArgs.emplace_back(&WrapperArg); 3506 3507 const CGFunctionInfo &CGFI = 3508 CGM.getTypes().arrangeBuiltinFunctionDeclaration(Ctx.VoidTy, WrapperArgs); 3509 3510 auto *Fn = llvm::Function::Create( 3511 CGM.getTypes().GetFunctionType(CGFI), llvm::GlobalValue::InternalLinkage, 3512 Twine(OutlinedParallelFn->getName(), "_wrapper"), &CGM.getModule()); 3513 3514 // Ensure we do not inline the function. This is trivially true for the ones 3515 // passed to __kmpc_fork_call but the ones calles in serialized regions 3516 // could be inlined. This is not a perfect but it is closer to the invariant 3517 // we want, namely, every data environment starts with a new function. 3518 // TODO: We should pass the if condition to the runtime function and do the 3519 // handling there. Much cleaner code. 3520 Fn->addFnAttr(llvm::Attribute::NoInline); 3521 3522 CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI); 3523 Fn->setLinkage(llvm::GlobalValue::InternalLinkage); 3524 Fn->setDoesNotRecurse(); 3525 3526 CodeGenFunction CGF(CGM, /*suppressNewContext=*/true); 3527 CGF.StartFunction(GlobalDecl(), Ctx.VoidTy, Fn, CGFI, WrapperArgs, 3528 D.getBeginLoc(), D.getBeginLoc()); 3529 3530 const auto *RD = CS.getCapturedRecordDecl(); 3531 auto CurField = RD->field_begin(); 3532 3533 Address ZeroAddr = CGF.CreateDefaultAlignTempAlloca(CGF.Int32Ty, 3534 /*Name=*/".zero.addr"); 3535 CGF.Builder.CreateStore(CGF.Builder.getInt32(/*C*/ 0), ZeroAddr); 3536 // Get the array of arguments. 3537 SmallVector<llvm::Value *, 8> Args; 3538 3539 Args.emplace_back(CGF.GetAddrOfLocalVar(&WrapperArg).getPointer()); 3540 Args.emplace_back(ZeroAddr.getPointer()); 3541 3542 CGBuilderTy &Bld = CGF.Builder; 3543 auto CI = CS.capture_begin(); 3544 3545 // Use global memory for data sharing. 3546 // Handle passing of global args to workers. 3547 Address GlobalArgs = 3548 CGF.CreateDefaultAlignTempAlloca(CGF.VoidPtrPtrTy, "global_args"); 3549 llvm::Value *GlobalArgsPtr = GlobalArgs.getPointer(); 3550 llvm::Value *DataSharingArgs[] = {GlobalArgsPtr}; 3551 CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( 3552 CGM.getModule(), OMPRTL___kmpc_get_shared_variables), 3553 DataSharingArgs); 3554 3555 // Retrieve the shared variables from the list of references returned 3556 // by the runtime. Pass the variables to the outlined function. 3557 Address SharedArgListAddress = Address::invalid(); 3558 if (CS.capture_size() > 0 || 3559 isOpenMPLoopBoundSharingDirective(D.getDirectiveKind())) { 3560 SharedArgListAddress = CGF.EmitLoadOfPointer( 3561 GlobalArgs, CGF.getContext() 3562 .getPointerType(CGF.getContext().VoidPtrTy) 3563 .castAs<PointerType>()); 3564 } 3565 unsigned Idx = 0; 3566 if (isOpenMPLoopBoundSharingDirective(D.getDirectiveKind())) { 3567 Address Src = Bld.CreateConstInBoundsGEP(SharedArgListAddress, Idx); 3568 Address TypedAddress = Bld.CreatePointerBitCastOrAddrSpaceCast( 3569 Src, CGF.SizeTy->getPointerTo(), CGF.SizeTy); 3570 llvm::Value *LB = CGF.EmitLoadOfScalar( 3571 TypedAddress, 3572 /*Volatile=*/false, 3573 CGF.getContext().getPointerType(CGF.getContext().getSizeType()), 3574 cast<OMPLoopDirective>(D).getLowerBoundVariable()->getExprLoc()); 3575 Args.emplace_back(LB); 3576 ++Idx; 3577 Src = Bld.CreateConstInBoundsGEP(SharedArgListAddress, Idx); 3578 TypedAddress = Bld.CreatePointerBitCastOrAddrSpaceCast( 3579 Src, CGF.SizeTy->getPointerTo(), CGF.SizeTy); 3580 llvm::Value *UB = CGF.EmitLoadOfScalar( 3581 TypedAddress, 3582 /*Volatile=*/false, 3583 CGF.getContext().getPointerType(CGF.getContext().getSizeType()), 3584 cast<OMPLoopDirective>(D).getUpperBoundVariable()->getExprLoc()); 3585 Args.emplace_back(UB); 3586 ++Idx; 3587 } 3588 if (CS.capture_size() > 0) { 3589 ASTContext &CGFContext = CGF.getContext(); 3590 for (unsigned I = 0, E = CS.capture_size(); I < E; ++I, ++CI, ++CurField) { 3591 QualType ElemTy = CurField->getType(); 3592 Address Src = Bld.CreateConstInBoundsGEP(SharedArgListAddress, I + Idx); 3593 Address TypedAddress = Bld.CreatePointerBitCastOrAddrSpaceCast( 3594 Src, CGF.ConvertTypeForMem(CGFContext.getPointerType(ElemTy)), 3595 CGF.ConvertTypeForMem(ElemTy)); 3596 llvm::Value *Arg = CGF.EmitLoadOfScalar(TypedAddress, 3597 /*Volatile=*/false, 3598 CGFContext.getPointerType(ElemTy), 3599 CI->getLocation()); 3600 if (CI->capturesVariableByCopy() && 3601 !CI->getCapturedVar()->getType()->isAnyPointerType()) { 3602 Arg = castValueToType(CGF, Arg, ElemTy, CGFContext.getUIntPtrType(), 3603 CI->getLocation()); 3604 } 3605 Args.emplace_back(Arg); 3606 } 3607 } 3608 3609 emitOutlinedFunctionCall(CGF, D.getBeginLoc(), OutlinedParallelFn, Args); 3610 CGF.FinishFunction(); 3611 return Fn; 3612 } 3613 3614 void CGOpenMPRuntimeGPU::emitFunctionProlog(CodeGenFunction &CGF, 3615 const Decl *D) { 3616 if (getDataSharingMode(CGM) != CGOpenMPRuntimeGPU::Generic) 3617 return; 3618 3619 assert(D && "Expected function or captured|block decl."); 3620 assert(FunctionGlobalizedDecls.count(CGF.CurFn) == 0 && 3621 "Function is registered already."); 3622 assert((!TeamAndReductions.first || TeamAndReductions.first == D) && 3623 "Team is set but not processed."); 3624 const Stmt *Body = nullptr; 3625 bool NeedToDelayGlobalization = false; 3626 if (const auto *FD = dyn_cast<FunctionDecl>(D)) { 3627 Body = FD->getBody(); 3628 } else if (const auto *BD = dyn_cast<BlockDecl>(D)) { 3629 Body = BD->getBody(); 3630 } else if (const auto *CD = dyn_cast<CapturedDecl>(D)) { 3631 Body = CD->getBody(); 3632 NeedToDelayGlobalization = CGF.CapturedStmtInfo->getKind() == CR_OpenMP; 3633 if (NeedToDelayGlobalization && 3634 getExecutionMode() == CGOpenMPRuntimeGPU::EM_SPMD) 3635 return; 3636 } 3637 if (!Body) 3638 return; 3639 CheckVarsEscapingDeclContext VarChecker(CGF, TeamAndReductions.second); 3640 VarChecker.Visit(Body); 3641 const RecordDecl *GlobalizedVarsRecord = 3642 VarChecker.getGlobalizedRecord(IsInTTDRegion); 3643 TeamAndReductions.first = nullptr; 3644 TeamAndReductions.second.clear(); 3645 ArrayRef<const ValueDecl *> EscapedVariableLengthDecls = 3646 VarChecker.getEscapedVariableLengthDecls(); 3647 if (!GlobalizedVarsRecord && EscapedVariableLengthDecls.empty()) 3648 return; 3649 auto I = FunctionGlobalizedDecls.try_emplace(CGF.CurFn).first; 3650 I->getSecond().MappedParams = 3651 std::make_unique<CodeGenFunction::OMPMapVars>(); 3652 I->getSecond().EscapedParameters.insert( 3653 VarChecker.getEscapedParameters().begin(), 3654 VarChecker.getEscapedParameters().end()); 3655 I->getSecond().EscapedVariableLengthDecls.append( 3656 EscapedVariableLengthDecls.begin(), EscapedVariableLengthDecls.end()); 3657 DeclToAddrMapTy &Data = I->getSecond().LocalVarData; 3658 for (const ValueDecl *VD : VarChecker.getEscapedDecls()) { 3659 assert(VD->isCanonicalDecl() && "Expected canonical declaration"); 3660 Data.insert(std::make_pair(VD, MappedVarData())); 3661 } 3662 if (!IsInTTDRegion && !NeedToDelayGlobalization && !IsInParallelRegion) { 3663 CheckVarsEscapingDeclContext VarChecker(CGF, llvm::None); 3664 VarChecker.Visit(Body); 3665 I->getSecond().SecondaryLocalVarData.emplace(); 3666 DeclToAddrMapTy &Data = *I->getSecond().SecondaryLocalVarData; 3667 for (const ValueDecl *VD : VarChecker.getEscapedDecls()) { 3668 assert(VD->isCanonicalDecl() && "Expected canonical declaration"); 3669 Data.insert(std::make_pair(VD, MappedVarData())); 3670 } 3671 } 3672 if (!NeedToDelayGlobalization) { 3673 emitGenericVarsProlog(CGF, D->getBeginLoc(), /*WithSPMDCheck=*/true); 3674 struct GlobalizationScope final : EHScopeStack::Cleanup { 3675 GlobalizationScope() = default; 3676 3677 void Emit(CodeGenFunction &CGF, Flags flags) override { 3678 static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()) 3679 .emitGenericVarsEpilog(CGF, /*WithSPMDCheck=*/true); 3680 } 3681 }; 3682 CGF.EHStack.pushCleanup<GlobalizationScope>(NormalAndEHCleanup); 3683 } 3684 } 3685 3686 Address CGOpenMPRuntimeGPU::getAddressOfLocalVariable(CodeGenFunction &CGF, 3687 const VarDecl *VD) { 3688 if (VD && VD->hasAttr<OMPAllocateDeclAttr>()) { 3689 const auto *A = VD->getAttr<OMPAllocateDeclAttr>(); 3690 auto AS = LangAS::Default; 3691 switch (A->getAllocatorType()) { 3692 // Use the default allocator here as by default local vars are 3693 // threadlocal. 3694 case OMPAllocateDeclAttr::OMPNullMemAlloc: 3695 case OMPAllocateDeclAttr::OMPDefaultMemAlloc: 3696 case OMPAllocateDeclAttr::OMPThreadMemAlloc: 3697 case OMPAllocateDeclAttr::OMPHighBWMemAlloc: 3698 case OMPAllocateDeclAttr::OMPLowLatMemAlloc: 3699 // Follow the user decision - use default allocation. 3700 return Address::invalid(); 3701 case OMPAllocateDeclAttr::OMPUserDefinedMemAlloc: 3702 // TODO: implement aupport for user-defined allocators. 3703 return Address::invalid(); 3704 case OMPAllocateDeclAttr::OMPConstMemAlloc: 3705 AS = LangAS::cuda_constant; 3706 break; 3707 case OMPAllocateDeclAttr::OMPPTeamMemAlloc: 3708 AS = LangAS::cuda_shared; 3709 break; 3710 case OMPAllocateDeclAttr::OMPLargeCapMemAlloc: 3711 case OMPAllocateDeclAttr::OMPCGroupMemAlloc: 3712 break; 3713 } 3714 llvm::Type *VarTy = CGF.ConvertTypeForMem(VD->getType()); 3715 auto *GV = new llvm::GlobalVariable( 3716 CGM.getModule(), VarTy, /*isConstant=*/false, 3717 llvm::GlobalValue::InternalLinkage, llvm::Constant::getNullValue(VarTy), 3718 VD->getName(), 3719 /*InsertBefore=*/nullptr, llvm::GlobalValue::NotThreadLocal, 3720 CGM.getContext().getTargetAddressSpace(AS)); 3721 CharUnits Align = CGM.getContext().getDeclAlign(VD); 3722 GV->setAlignment(Align.getAsAlign()); 3723 return Address( 3724 CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( 3725 GV, VarTy->getPointerTo(CGM.getContext().getTargetAddressSpace( 3726 VD->getType().getAddressSpace()))), 3727 VarTy, Align); 3728 } 3729 3730 if (getDataSharingMode(CGM) != CGOpenMPRuntimeGPU::Generic) 3731 return Address::invalid(); 3732 3733 VD = VD->getCanonicalDecl(); 3734 auto I = FunctionGlobalizedDecls.find(CGF.CurFn); 3735 if (I == FunctionGlobalizedDecls.end()) 3736 return Address::invalid(); 3737 auto VDI = I->getSecond().LocalVarData.find(VD); 3738 if (VDI != I->getSecond().LocalVarData.end()) 3739 return VDI->second.PrivateAddr; 3740 if (VD->hasAttrs()) { 3741 for (specific_attr_iterator<OMPReferencedVarAttr> IT(VD->attr_begin()), 3742 E(VD->attr_end()); 3743 IT != E; ++IT) { 3744 auto VDI = I->getSecond().LocalVarData.find( 3745 cast<VarDecl>(cast<DeclRefExpr>(IT->getRef())->getDecl()) 3746 ->getCanonicalDecl()); 3747 if (VDI != I->getSecond().LocalVarData.end()) 3748 return VDI->second.PrivateAddr; 3749 } 3750 } 3751 3752 return Address::invalid(); 3753 } 3754 3755 void CGOpenMPRuntimeGPU::functionFinished(CodeGenFunction &CGF) { 3756 FunctionGlobalizedDecls.erase(CGF.CurFn); 3757 CGOpenMPRuntime::functionFinished(CGF); 3758 } 3759 3760 void CGOpenMPRuntimeGPU::getDefaultDistScheduleAndChunk( 3761 CodeGenFunction &CGF, const OMPLoopDirective &S, 3762 OpenMPDistScheduleClauseKind &ScheduleKind, 3763 llvm::Value *&Chunk) const { 3764 auto &RT = static_cast<CGOpenMPRuntimeGPU &>(CGF.CGM.getOpenMPRuntime()); 3765 if (getExecutionMode() == CGOpenMPRuntimeGPU::EM_SPMD) { 3766 ScheduleKind = OMPC_DIST_SCHEDULE_static; 3767 Chunk = CGF.EmitScalarConversion( 3768 RT.getGPUNumThreads(CGF), 3769 CGF.getContext().getIntTypeForBitwidth(32, /*Signed=*/0), 3770 S.getIterationVariable()->getType(), S.getBeginLoc()); 3771 return; 3772 } 3773 CGOpenMPRuntime::getDefaultDistScheduleAndChunk( 3774 CGF, S, ScheduleKind, Chunk); 3775 } 3776 3777 void CGOpenMPRuntimeGPU::getDefaultScheduleAndChunk( 3778 CodeGenFunction &CGF, const OMPLoopDirective &S, 3779 OpenMPScheduleClauseKind &ScheduleKind, 3780 const Expr *&ChunkExpr) const { 3781 ScheduleKind = OMPC_SCHEDULE_static; 3782 // Chunk size is 1 in this case. 3783 llvm::APInt ChunkSize(32, 1); 3784 ChunkExpr = IntegerLiteral::Create(CGF.getContext(), ChunkSize, 3785 CGF.getContext().getIntTypeForBitwidth(32, /*Signed=*/0), 3786 SourceLocation()); 3787 } 3788 3789 void CGOpenMPRuntimeGPU::adjustTargetSpecificDataForLambdas( 3790 CodeGenFunction &CGF, const OMPExecutableDirective &D) const { 3791 assert(isOpenMPTargetExecutionDirective(D.getDirectiveKind()) && 3792 " Expected target-based directive."); 3793 const CapturedStmt *CS = D.getCapturedStmt(OMPD_target); 3794 for (const CapturedStmt::Capture &C : CS->captures()) { 3795 // Capture variables captured by reference in lambdas for target-based 3796 // directives. 3797 if (!C.capturesVariable()) 3798 continue; 3799 const VarDecl *VD = C.getCapturedVar(); 3800 const auto *RD = VD->getType() 3801 .getCanonicalType() 3802 .getNonReferenceType() 3803 ->getAsCXXRecordDecl(); 3804 if (!RD || !RD->isLambda()) 3805 continue; 3806 Address VDAddr = CGF.GetAddrOfLocalVar(VD); 3807 LValue VDLVal; 3808 if (VD->getType().getCanonicalType()->isReferenceType()) 3809 VDLVal = CGF.EmitLoadOfReferenceLValue(VDAddr, VD->getType()); 3810 else 3811 VDLVal = CGF.MakeAddrLValue( 3812 VDAddr, VD->getType().getCanonicalType().getNonReferenceType()); 3813 llvm::DenseMap<const VarDecl *, FieldDecl *> Captures; 3814 FieldDecl *ThisCapture = nullptr; 3815 RD->getCaptureFields(Captures, ThisCapture); 3816 if (ThisCapture && CGF.CapturedStmtInfo->isCXXThisExprCaptured()) { 3817 LValue ThisLVal = 3818 CGF.EmitLValueForFieldInitialization(VDLVal, ThisCapture); 3819 llvm::Value *CXXThis = CGF.LoadCXXThis(); 3820 CGF.EmitStoreOfScalar(CXXThis, ThisLVal); 3821 } 3822 for (const LambdaCapture &LC : RD->captures()) { 3823 if (LC.getCaptureKind() != LCK_ByRef) 3824 continue; 3825 const VarDecl *VD = LC.getCapturedVar(); 3826 if (!CS->capturesVariable(VD)) 3827 continue; 3828 auto It = Captures.find(VD); 3829 assert(It != Captures.end() && "Found lambda capture without field."); 3830 LValue VarLVal = CGF.EmitLValueForFieldInitialization(VDLVal, It->second); 3831 Address VDAddr = CGF.GetAddrOfLocalVar(VD); 3832 if (VD->getType().getCanonicalType()->isReferenceType()) 3833 VDAddr = CGF.EmitLoadOfReferenceLValue(VDAddr, 3834 VD->getType().getCanonicalType()) 3835 .getAddress(CGF); 3836 CGF.EmitStoreOfScalar(VDAddr.getPointer(), VarLVal); 3837 } 3838 } 3839 } 3840 3841 bool CGOpenMPRuntimeGPU::hasAllocateAttributeForGlobalVar(const VarDecl *VD, 3842 LangAS &AS) { 3843 if (!VD || !VD->hasAttr<OMPAllocateDeclAttr>()) 3844 return false; 3845 const auto *A = VD->getAttr<OMPAllocateDeclAttr>(); 3846 switch(A->getAllocatorType()) { 3847 case OMPAllocateDeclAttr::OMPNullMemAlloc: 3848 case OMPAllocateDeclAttr::OMPDefaultMemAlloc: 3849 // Not supported, fallback to the default mem space. 3850 case OMPAllocateDeclAttr::OMPThreadMemAlloc: 3851 case OMPAllocateDeclAttr::OMPLargeCapMemAlloc: 3852 case OMPAllocateDeclAttr::OMPCGroupMemAlloc: 3853 case OMPAllocateDeclAttr::OMPHighBWMemAlloc: 3854 case OMPAllocateDeclAttr::OMPLowLatMemAlloc: 3855 AS = LangAS::Default; 3856 return true; 3857 case OMPAllocateDeclAttr::OMPConstMemAlloc: 3858 AS = LangAS::cuda_constant; 3859 return true; 3860 case OMPAllocateDeclAttr::OMPPTeamMemAlloc: 3861 AS = LangAS::cuda_shared; 3862 return true; 3863 case OMPAllocateDeclAttr::OMPUserDefinedMemAlloc: 3864 llvm_unreachable("Expected predefined allocator for the variables with the " 3865 "static storage."); 3866 } 3867 return false; 3868 } 3869 3870 // Get current CudaArch and ignore any unknown values 3871 static CudaArch getCudaArch(CodeGenModule &CGM) { 3872 if (!CGM.getTarget().hasFeature("ptx")) 3873 return CudaArch::UNKNOWN; 3874 for (const auto &Feature : CGM.getTarget().getTargetOpts().FeatureMap) { 3875 if (Feature.getValue()) { 3876 CudaArch Arch = StringToCudaArch(Feature.getKey()); 3877 if (Arch != CudaArch::UNKNOWN) 3878 return Arch; 3879 } 3880 } 3881 return CudaArch::UNKNOWN; 3882 } 3883 3884 /// Check to see if target architecture supports unified addressing which is 3885 /// a restriction for OpenMP requires clause "unified_shared_memory". 3886 void CGOpenMPRuntimeGPU::processRequiresDirective( 3887 const OMPRequiresDecl *D) { 3888 for (const OMPClause *Clause : D->clauselists()) { 3889 if (Clause->getClauseKind() == OMPC_unified_shared_memory) { 3890 CudaArch Arch = getCudaArch(CGM); 3891 switch (Arch) { 3892 case CudaArch::SM_20: 3893 case CudaArch::SM_21: 3894 case CudaArch::SM_30: 3895 case CudaArch::SM_32: 3896 case CudaArch::SM_35: 3897 case CudaArch::SM_37: 3898 case CudaArch::SM_50: 3899 case CudaArch::SM_52: 3900 case CudaArch::SM_53: { 3901 SmallString<256> Buffer; 3902 llvm::raw_svector_ostream Out(Buffer); 3903 Out << "Target architecture " << CudaArchToString(Arch) 3904 << " does not support unified addressing"; 3905 CGM.Error(Clause->getBeginLoc(), Out.str()); 3906 return; 3907 } 3908 case CudaArch::SM_60: 3909 case CudaArch::SM_61: 3910 case CudaArch::SM_62: 3911 case CudaArch::SM_70: 3912 case CudaArch::SM_72: 3913 case CudaArch::SM_75: 3914 case CudaArch::SM_80: 3915 case CudaArch::SM_86: 3916 case CudaArch::GFX600: 3917 case CudaArch::GFX601: 3918 case CudaArch::GFX602: 3919 case CudaArch::GFX700: 3920 case CudaArch::GFX701: 3921 case CudaArch::GFX702: 3922 case CudaArch::GFX703: 3923 case CudaArch::GFX704: 3924 case CudaArch::GFX705: 3925 case CudaArch::GFX801: 3926 case CudaArch::GFX802: 3927 case CudaArch::GFX803: 3928 case CudaArch::GFX805: 3929 case CudaArch::GFX810: 3930 case CudaArch::GFX900: 3931 case CudaArch::GFX902: 3932 case CudaArch::GFX904: 3933 case CudaArch::GFX906: 3934 case CudaArch::GFX908: 3935 case CudaArch::GFX909: 3936 case CudaArch::GFX90a: 3937 case CudaArch::GFX90c: 3938 case CudaArch::GFX940: 3939 case CudaArch::GFX1010: 3940 case CudaArch::GFX1011: 3941 case CudaArch::GFX1012: 3942 case CudaArch::GFX1013: 3943 case CudaArch::GFX1030: 3944 case CudaArch::GFX1031: 3945 case CudaArch::GFX1032: 3946 case CudaArch::GFX1033: 3947 case CudaArch::GFX1034: 3948 case CudaArch::GFX1035: 3949 case CudaArch::GFX1036: 3950 case CudaArch::GFX1100: 3951 case CudaArch::GFX1101: 3952 case CudaArch::GFX1102: 3953 case CudaArch::GFX1103: 3954 case CudaArch::Generic: 3955 case CudaArch::UNUSED: 3956 case CudaArch::UNKNOWN: 3957 break; 3958 case CudaArch::LAST: 3959 llvm_unreachable("Unexpected Cuda arch."); 3960 } 3961 } 3962 } 3963 CGOpenMPRuntime::processRequiresDirective(D); 3964 } 3965 3966 void CGOpenMPRuntimeGPU::clear() { 3967 3968 if (!TeamsReductions.empty()) { 3969 ASTContext &C = CGM.getContext(); 3970 RecordDecl *StaticRD = C.buildImplicitRecord( 3971 "_openmp_teams_reduction_type_$_", RecordDecl::TagKind::TTK_Union); 3972 StaticRD->startDefinition(); 3973 for (const RecordDecl *TeamReductionRec : TeamsReductions) { 3974 QualType RecTy = C.getRecordType(TeamReductionRec); 3975 auto *Field = FieldDecl::Create( 3976 C, StaticRD, SourceLocation(), SourceLocation(), nullptr, RecTy, 3977 C.getTrivialTypeSourceInfo(RecTy, SourceLocation()), 3978 /*BW=*/nullptr, /*Mutable=*/false, 3979 /*InitStyle=*/ICIS_NoInit); 3980 Field->setAccess(AS_public); 3981 StaticRD->addDecl(Field); 3982 } 3983 StaticRD->completeDefinition(); 3984 QualType StaticTy = C.getRecordType(StaticRD); 3985 llvm::Type *LLVMReductionsBufferTy = 3986 CGM.getTypes().ConvertTypeForMem(StaticTy); 3987 // FIXME: nvlink does not handle weak linkage correctly (object with the 3988 // different size are reported as erroneous). 3989 // Restore CommonLinkage as soon as nvlink is fixed. 3990 auto *GV = new llvm::GlobalVariable( 3991 CGM.getModule(), LLVMReductionsBufferTy, 3992 /*isConstant=*/false, llvm::GlobalValue::InternalLinkage, 3993 llvm::Constant::getNullValue(LLVMReductionsBufferTy), 3994 "_openmp_teams_reductions_buffer_$_"); 3995 KernelTeamsReductionPtr->setInitializer( 3996 llvm::ConstantExpr::getPointerBitCastOrAddrSpaceCast(GV, 3997 CGM.VoidPtrTy)); 3998 } 3999 CGOpenMPRuntime::clear(); 4000 } 4001 4002 llvm::Value *CGOpenMPRuntimeGPU::getGPUNumThreads(CodeGenFunction &CGF) { 4003 CGBuilderTy &Bld = CGF.Builder; 4004 llvm::Module *M = &CGF.CGM.getModule(); 4005 const char *LocSize = "__kmpc_get_hardware_num_threads_in_block"; 4006 llvm::Function *F = M->getFunction(LocSize); 4007 if (!F) { 4008 F = llvm::Function::Create( 4009 llvm::FunctionType::get(CGF.Int32Ty, llvm::None, false), 4010 llvm::GlobalVariable::ExternalLinkage, LocSize, &CGF.CGM.getModule()); 4011 } 4012 return Bld.CreateCall(F, llvm::None, "nvptx_num_threads"); 4013 } 4014 4015 llvm::Value *CGOpenMPRuntimeGPU::getGPUThreadID(CodeGenFunction &CGF) { 4016 ArrayRef<llvm::Value *> Args{}; 4017 return CGF.EmitRuntimeCall( 4018 OMPBuilder.getOrCreateRuntimeFunction( 4019 CGM.getModule(), OMPRTL___kmpc_get_hardware_thread_id_in_block), 4020 Args); 4021 } 4022 4023 llvm::Value *CGOpenMPRuntimeGPU::getGPUWarpSize(CodeGenFunction &CGF) { 4024 ArrayRef<llvm::Value *> Args{}; 4025 return CGF.EmitRuntimeCall(OMPBuilder.getOrCreateRuntimeFunction( 4026 CGM.getModule(), OMPRTL___kmpc_get_warp_size), 4027 Args); 4028 } 4029