1 //===- OpenMPIRBuilder.cpp - Builder for LLVM-IR for OpenMP directives ----===// 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 /// \file 9 /// 10 /// This file implements the OpenMPIRBuilder class, which is used as a 11 /// convenient way to create LLVM instructions for OpenMP directives. 12 /// 13 //===----------------------------------------------------------------------===// 14 15 #include "llvm/Frontend/OpenMP/OMPIRBuilder.h" 16 #include "llvm/ADT/SmallSet.h" 17 #include "llvm/ADT/StringRef.h" 18 #include "llvm/Analysis/AssumptionCache.h" 19 #include "llvm/Analysis/CodeMetrics.h" 20 #include "llvm/Analysis/LoopInfo.h" 21 #include "llvm/Analysis/OptimizationRemarkEmitter.h" 22 #include "llvm/Analysis/ScalarEvolution.h" 23 #include "llvm/Analysis/TargetLibraryInfo.h" 24 #include "llvm/IR/CFG.h" 25 #include "llvm/IR/Constants.h" 26 #include "llvm/IR/DebugInfoMetadata.h" 27 #include "llvm/IR/GlobalVariable.h" 28 #include "llvm/IR/IRBuilder.h" 29 #include "llvm/IR/MDBuilder.h" 30 #include "llvm/IR/PassManager.h" 31 #include "llvm/IR/Value.h" 32 #include "llvm/MC/TargetRegistry.h" 33 #include "llvm/Support/CommandLine.h" 34 #include "llvm/Target/TargetMachine.h" 35 #include "llvm/Target/TargetOptions.h" 36 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 37 #include "llvm/Transforms/Utils/CodeExtractor.h" 38 #include "llvm/Transforms/Utils/LoopPeel.h" 39 #include "llvm/Transforms/Utils/UnrollLoop.h" 40 41 #include <cstdint> 42 43 #define DEBUG_TYPE "openmp-ir-builder" 44 45 using namespace llvm; 46 using namespace omp; 47 48 static cl::opt<bool> 49 OptimisticAttributes("openmp-ir-builder-optimistic-attributes", cl::Hidden, 50 cl::desc("Use optimistic attributes describing " 51 "'as-if' properties of runtime calls."), 52 cl::init(false)); 53 54 static cl::opt<double> UnrollThresholdFactor( 55 "openmp-ir-builder-unroll-threshold-factor", cl::Hidden, 56 cl::desc("Factor for the unroll threshold to account for code " 57 "simplifications still taking place"), 58 cl::init(1.5)); 59 60 #ifndef NDEBUG 61 /// Return whether IP1 and IP2 are ambiguous, i.e. that inserting instructions 62 /// at position IP1 may change the meaning of IP2 or vice-versa. This is because 63 /// an InsertPoint stores the instruction before something is inserted. For 64 /// instance, if both point to the same instruction, two IRBuilders alternating 65 /// creating instruction will cause the instructions to be interleaved. 66 static bool isConflictIP(IRBuilder<>::InsertPoint IP1, 67 IRBuilder<>::InsertPoint IP2) { 68 if (!IP1.isSet() || !IP2.isSet()) 69 return false; 70 return IP1.getBlock() == IP2.getBlock() && IP1.getPoint() == IP2.getPoint(); 71 } 72 73 static bool isValidWorkshareLoopScheduleType(OMPScheduleType SchedType) { 74 // Valid ordered/unordered and base algorithm combinations. 75 switch (SchedType & ~OMPScheduleType::MonotonicityMask) { 76 case OMPScheduleType::UnorderedStaticChunked: 77 case OMPScheduleType::UnorderedStatic: 78 case OMPScheduleType::UnorderedDynamicChunked: 79 case OMPScheduleType::UnorderedGuidedChunked: 80 case OMPScheduleType::UnorderedRuntime: 81 case OMPScheduleType::UnorderedAuto: 82 case OMPScheduleType::UnorderedTrapezoidal: 83 case OMPScheduleType::UnorderedGreedy: 84 case OMPScheduleType::UnorderedBalanced: 85 case OMPScheduleType::UnorderedGuidedIterativeChunked: 86 case OMPScheduleType::UnorderedGuidedAnalyticalChunked: 87 case OMPScheduleType::UnorderedSteal: 88 case OMPScheduleType::UnorderedStaticBalancedChunked: 89 case OMPScheduleType::UnorderedGuidedSimd: 90 case OMPScheduleType::UnorderedRuntimeSimd: 91 case OMPScheduleType::OrderedStaticChunked: 92 case OMPScheduleType::OrderedStatic: 93 case OMPScheduleType::OrderedDynamicChunked: 94 case OMPScheduleType::OrderedGuidedChunked: 95 case OMPScheduleType::OrderedRuntime: 96 case OMPScheduleType::OrderedAuto: 97 case OMPScheduleType::OrderdTrapezoidal: 98 case OMPScheduleType::NomergeUnorderedStaticChunked: 99 case OMPScheduleType::NomergeUnorderedStatic: 100 case OMPScheduleType::NomergeUnorderedDynamicChunked: 101 case OMPScheduleType::NomergeUnorderedGuidedChunked: 102 case OMPScheduleType::NomergeUnorderedRuntime: 103 case OMPScheduleType::NomergeUnorderedAuto: 104 case OMPScheduleType::NomergeUnorderedTrapezoidal: 105 case OMPScheduleType::NomergeUnorderedGreedy: 106 case OMPScheduleType::NomergeUnorderedBalanced: 107 case OMPScheduleType::NomergeUnorderedGuidedIterativeChunked: 108 case OMPScheduleType::NomergeUnorderedGuidedAnalyticalChunked: 109 case OMPScheduleType::NomergeUnorderedSteal: 110 case OMPScheduleType::NomergeOrderedStaticChunked: 111 case OMPScheduleType::NomergeOrderedStatic: 112 case OMPScheduleType::NomergeOrderedDynamicChunked: 113 case OMPScheduleType::NomergeOrderedGuidedChunked: 114 case OMPScheduleType::NomergeOrderedRuntime: 115 case OMPScheduleType::NomergeOrderedAuto: 116 case OMPScheduleType::NomergeOrderedTrapezoidal: 117 break; 118 default: 119 return false; 120 } 121 122 // Must not set both monotonicity modifiers at the same time. 123 OMPScheduleType MonotonicityFlags = 124 SchedType & OMPScheduleType::MonotonicityMask; 125 if (MonotonicityFlags == OMPScheduleType::MonotonicityMask) 126 return false; 127 128 return true; 129 } 130 #endif 131 132 /// Determine which scheduling algorithm to use, determined from schedule clause 133 /// arguments. 134 static OMPScheduleType 135 getOpenMPBaseScheduleType(llvm::omp::ScheduleKind ClauseKind, bool HasChunks, 136 bool HasSimdModifier) { 137 // Currently, the default schedule it static. 138 switch (ClauseKind) { 139 case OMP_SCHEDULE_Default: 140 case OMP_SCHEDULE_Static: 141 return HasChunks ? OMPScheduleType::BaseStaticChunked 142 : OMPScheduleType::BaseStatic; 143 case OMP_SCHEDULE_Dynamic: 144 return OMPScheduleType::BaseDynamicChunked; 145 case OMP_SCHEDULE_Guided: 146 return HasSimdModifier ? OMPScheduleType::BaseGuidedSimd 147 : OMPScheduleType::BaseGuidedChunked; 148 case OMP_SCHEDULE_Auto: 149 return llvm::omp::OMPScheduleType::BaseAuto; 150 case OMP_SCHEDULE_Runtime: 151 return HasSimdModifier ? OMPScheduleType::BaseRuntimeSimd 152 : OMPScheduleType::BaseRuntime; 153 } 154 llvm_unreachable("unhandled schedule clause argument"); 155 } 156 157 /// Adds ordering modifier flags to schedule type. 158 static OMPScheduleType 159 getOpenMPOrderingScheduleType(OMPScheduleType BaseScheduleType, 160 bool HasOrderedClause) { 161 assert((BaseScheduleType & OMPScheduleType::ModifierMask) == 162 OMPScheduleType::None && 163 "Must not have ordering nor monotonicity flags already set"); 164 165 OMPScheduleType OrderingModifier = HasOrderedClause 166 ? OMPScheduleType::ModifierOrdered 167 : OMPScheduleType::ModifierUnordered; 168 OMPScheduleType OrderingScheduleType = BaseScheduleType | OrderingModifier; 169 170 // Unsupported combinations 171 if (OrderingScheduleType == 172 (OMPScheduleType::BaseGuidedSimd | OMPScheduleType::ModifierOrdered)) 173 return OMPScheduleType::OrderedGuidedChunked; 174 else if (OrderingScheduleType == (OMPScheduleType::BaseRuntimeSimd | 175 OMPScheduleType::ModifierOrdered)) 176 return OMPScheduleType::OrderedRuntime; 177 178 return OrderingScheduleType; 179 } 180 181 /// Adds monotonicity modifier flags to schedule type. 182 static OMPScheduleType 183 getOpenMPMonotonicityScheduleType(OMPScheduleType ScheduleType, 184 bool HasSimdModifier, bool HasMonotonic, 185 bool HasNonmonotonic, bool HasOrderedClause) { 186 assert((ScheduleType & OMPScheduleType::MonotonicityMask) == 187 OMPScheduleType::None && 188 "Must not have monotonicity flags already set"); 189 assert((!HasMonotonic || !HasNonmonotonic) && 190 "Monotonic and Nonmonotonic are contradicting each other"); 191 192 if (HasMonotonic) { 193 return ScheduleType | OMPScheduleType::ModifierMonotonic; 194 } else if (HasNonmonotonic) { 195 return ScheduleType | OMPScheduleType::ModifierNonmonotonic; 196 } else { 197 // OpenMP 5.1, 2.11.4 Worksharing-Loop Construct, Description. 198 // If the static schedule kind is specified or if the ordered clause is 199 // specified, and if the nonmonotonic modifier is not specified, the 200 // effect is as if the monotonic modifier is specified. Otherwise, unless 201 // the monotonic modifier is specified, the effect is as if the 202 // nonmonotonic modifier is specified. 203 OMPScheduleType BaseScheduleType = 204 ScheduleType & ~OMPScheduleType::ModifierMask; 205 if ((BaseScheduleType == OMPScheduleType::BaseStatic) || 206 (BaseScheduleType == OMPScheduleType::BaseStaticChunked) || 207 HasOrderedClause) { 208 // The monotonic is used by default in openmp runtime library, so no need 209 // to set it. 210 return ScheduleType; 211 } else { 212 return ScheduleType | OMPScheduleType::ModifierNonmonotonic; 213 } 214 } 215 } 216 217 /// Determine the schedule type using schedule and ordering clause arguments. 218 static OMPScheduleType 219 computeOpenMPScheduleType(ScheduleKind ClauseKind, bool HasChunks, 220 bool HasSimdModifier, bool HasMonotonicModifier, 221 bool HasNonmonotonicModifier, bool HasOrderedClause) { 222 OMPScheduleType BaseSchedule = 223 getOpenMPBaseScheduleType(ClauseKind, HasChunks, HasSimdModifier); 224 OMPScheduleType OrderedSchedule = 225 getOpenMPOrderingScheduleType(BaseSchedule, HasOrderedClause); 226 OMPScheduleType Result = getOpenMPMonotonicityScheduleType( 227 OrderedSchedule, HasSimdModifier, HasMonotonicModifier, 228 HasNonmonotonicModifier, HasOrderedClause); 229 230 assert(isValidWorkshareLoopScheduleType(Result)); 231 return Result; 232 } 233 234 /// Make \p Source branch to \p Target. 235 /// 236 /// Handles two situations: 237 /// * \p Source already has an unconditional branch. 238 /// * \p Source is a degenerate block (no terminator because the BB is 239 /// the current head of the IR construction). 240 static void redirectTo(BasicBlock *Source, BasicBlock *Target, DebugLoc DL) { 241 if (Instruction *Term = Source->getTerminator()) { 242 auto *Br = cast<BranchInst>(Term); 243 assert(!Br->isConditional() && 244 "BB's terminator must be an unconditional branch (or degenerate)"); 245 BasicBlock *Succ = Br->getSuccessor(0); 246 Succ->removePredecessor(Source, /*KeepOneInputPHIs=*/true); 247 Br->setSuccessor(0, Target); 248 return; 249 } 250 251 auto *NewBr = BranchInst::Create(Target, Source); 252 NewBr->setDebugLoc(DL); 253 } 254 255 void llvm::spliceBB(IRBuilderBase::InsertPoint IP, BasicBlock *New, 256 bool CreateBranch) { 257 assert(New->getFirstInsertionPt() == New->begin() && 258 "Target BB must not have PHI nodes"); 259 260 // Move instructions to new block. 261 BasicBlock *Old = IP.getBlock(); 262 New->getInstList().splice(New->begin(), Old->getInstList(), IP.getPoint(), 263 Old->end()); 264 265 if (CreateBranch) 266 BranchInst::Create(New, Old); 267 } 268 269 void llvm::spliceBB(IRBuilder<> &Builder, BasicBlock *New, bool CreateBranch) { 270 DebugLoc DebugLoc = Builder.getCurrentDebugLocation(); 271 BasicBlock *Old = Builder.GetInsertBlock(); 272 273 spliceBB(Builder.saveIP(), New, CreateBranch); 274 if (CreateBranch) 275 Builder.SetInsertPoint(Old->getTerminator()); 276 else 277 Builder.SetInsertPoint(Old); 278 279 // SetInsertPoint also updates the Builder's debug location, but we want to 280 // keep the one the Builder was configured to use. 281 Builder.SetCurrentDebugLocation(DebugLoc); 282 } 283 284 BasicBlock *llvm::splitBB(IRBuilderBase::InsertPoint IP, bool CreateBranch, 285 llvm::Twine Name) { 286 BasicBlock *Old = IP.getBlock(); 287 BasicBlock *New = BasicBlock::Create( 288 Old->getContext(), Name.isTriviallyEmpty() ? Old->getName() : Name, 289 Old->getParent(), Old->getNextNode()); 290 spliceBB(IP, New, CreateBranch); 291 New->replaceSuccessorsPhiUsesWith(Old, New); 292 return New; 293 } 294 295 BasicBlock *llvm::splitBB(IRBuilderBase &Builder, bool CreateBranch, 296 llvm::Twine Name) { 297 DebugLoc DebugLoc = Builder.getCurrentDebugLocation(); 298 BasicBlock *New = splitBB(Builder.saveIP(), CreateBranch, Name); 299 if (CreateBranch) 300 Builder.SetInsertPoint(Builder.GetInsertBlock()->getTerminator()); 301 else 302 Builder.SetInsertPoint(Builder.GetInsertBlock()); 303 // SetInsertPoint also updates the Builder's debug location, but we want to 304 // keep the one the Builder was configured to use. 305 Builder.SetCurrentDebugLocation(DebugLoc); 306 return New; 307 } 308 309 BasicBlock *llvm::splitBB(IRBuilder<> &Builder, bool CreateBranch, 310 llvm::Twine Name) { 311 DebugLoc DebugLoc = Builder.getCurrentDebugLocation(); 312 BasicBlock *New = splitBB(Builder.saveIP(), CreateBranch, Name); 313 if (CreateBranch) 314 Builder.SetInsertPoint(Builder.GetInsertBlock()->getTerminator()); 315 else 316 Builder.SetInsertPoint(Builder.GetInsertBlock()); 317 // SetInsertPoint also updates the Builder's debug location, but we want to 318 // keep the one the Builder was configured to use. 319 Builder.SetCurrentDebugLocation(DebugLoc); 320 return New; 321 } 322 323 BasicBlock *llvm::splitBBWithSuffix(IRBuilderBase &Builder, bool CreateBranch, 324 llvm::Twine Suffix) { 325 BasicBlock *Old = Builder.GetInsertBlock(); 326 return splitBB(Builder, CreateBranch, Old->getName() + Suffix); 327 } 328 329 void OpenMPIRBuilder::addAttributes(omp::RuntimeFunction FnID, Function &Fn) { 330 LLVMContext &Ctx = Fn.getContext(); 331 332 // Get the function's current attributes. 333 auto Attrs = Fn.getAttributes(); 334 auto FnAttrs = Attrs.getFnAttrs(); 335 auto RetAttrs = Attrs.getRetAttrs(); 336 SmallVector<AttributeSet, 4> ArgAttrs; 337 for (size_t ArgNo = 0; ArgNo < Fn.arg_size(); ++ArgNo) 338 ArgAttrs.emplace_back(Attrs.getParamAttrs(ArgNo)); 339 340 #define OMP_ATTRS_SET(VarName, AttrSet) AttributeSet VarName = AttrSet; 341 #include "llvm/Frontend/OpenMP/OMPKinds.def" 342 343 // Add attributes to the function declaration. 344 switch (FnID) { 345 #define OMP_RTL_ATTRS(Enum, FnAttrSet, RetAttrSet, ArgAttrSets) \ 346 case Enum: \ 347 FnAttrs = FnAttrs.addAttributes(Ctx, FnAttrSet); \ 348 RetAttrs = RetAttrs.addAttributes(Ctx, RetAttrSet); \ 349 for (size_t ArgNo = 0; ArgNo < ArgAttrSets.size(); ++ArgNo) \ 350 ArgAttrs[ArgNo] = \ 351 ArgAttrs[ArgNo].addAttributes(Ctx, ArgAttrSets[ArgNo]); \ 352 Fn.setAttributes(AttributeList::get(Ctx, FnAttrs, RetAttrs, ArgAttrs)); \ 353 break; 354 #include "llvm/Frontend/OpenMP/OMPKinds.def" 355 default: 356 // Attributes are optional. 357 break; 358 } 359 } 360 361 FunctionCallee 362 OpenMPIRBuilder::getOrCreateRuntimeFunction(Module &M, RuntimeFunction FnID) { 363 FunctionType *FnTy = nullptr; 364 Function *Fn = nullptr; 365 366 // Try to find the declation in the module first. 367 switch (FnID) { 368 #define OMP_RTL(Enum, Str, IsVarArg, ReturnType, ...) \ 369 case Enum: \ 370 FnTy = FunctionType::get(ReturnType, ArrayRef<Type *>{__VA_ARGS__}, \ 371 IsVarArg); \ 372 Fn = M.getFunction(Str); \ 373 break; 374 #include "llvm/Frontend/OpenMP/OMPKinds.def" 375 } 376 377 if (!Fn) { 378 // Create a new declaration if we need one. 379 switch (FnID) { 380 #define OMP_RTL(Enum, Str, ...) \ 381 case Enum: \ 382 Fn = Function::Create(FnTy, GlobalValue::ExternalLinkage, Str, M); \ 383 break; 384 #include "llvm/Frontend/OpenMP/OMPKinds.def" 385 } 386 387 // Add information if the runtime function takes a callback function 388 if (FnID == OMPRTL___kmpc_fork_call || FnID == OMPRTL___kmpc_fork_teams) { 389 if (!Fn->hasMetadata(LLVMContext::MD_callback)) { 390 LLVMContext &Ctx = Fn->getContext(); 391 MDBuilder MDB(Ctx); 392 // Annotate the callback behavior of the runtime function: 393 // - The callback callee is argument number 2 (microtask). 394 // - The first two arguments of the callback callee are unknown (-1). 395 // - All variadic arguments to the runtime function are passed to the 396 // callback callee. 397 Fn->addMetadata( 398 LLVMContext::MD_callback, 399 *MDNode::get(Ctx, {MDB.createCallbackEncoding( 400 2, {-1, -1}, /* VarArgsArePassed */ true)})); 401 } 402 } 403 404 LLVM_DEBUG(dbgs() << "Created OpenMP runtime function " << Fn->getName() 405 << " with type " << *Fn->getFunctionType() << "\n"); 406 addAttributes(FnID, *Fn); 407 408 } else { 409 LLVM_DEBUG(dbgs() << "Found OpenMP runtime function " << Fn->getName() 410 << " with type " << *Fn->getFunctionType() << "\n"); 411 } 412 413 assert(Fn && "Failed to create OpenMP runtime function"); 414 415 // Cast the function to the expected type if necessary 416 Constant *C = ConstantExpr::getBitCast(Fn, FnTy->getPointerTo()); 417 return {FnTy, C}; 418 } 419 420 Function *OpenMPIRBuilder::getOrCreateRuntimeFunctionPtr(RuntimeFunction FnID) { 421 FunctionCallee RTLFn = getOrCreateRuntimeFunction(M, FnID); 422 auto *Fn = dyn_cast<llvm::Function>(RTLFn.getCallee()); 423 assert(Fn && "Failed to create OpenMP runtime function pointer"); 424 return Fn; 425 } 426 427 void OpenMPIRBuilder::initialize() { initializeTypes(M); } 428 429 void OpenMPIRBuilder::finalize(Function *Fn) { 430 SmallPtrSet<BasicBlock *, 32> ParallelRegionBlockSet; 431 SmallVector<BasicBlock *, 32> Blocks; 432 SmallVector<OutlineInfo, 16> DeferredOutlines; 433 for (OutlineInfo &OI : OutlineInfos) { 434 // Skip functions that have not finalized yet; may happen with nested 435 // function generation. 436 if (Fn && OI.getFunction() != Fn) { 437 DeferredOutlines.push_back(OI); 438 continue; 439 } 440 441 ParallelRegionBlockSet.clear(); 442 Blocks.clear(); 443 OI.collectBlocks(ParallelRegionBlockSet, Blocks); 444 445 Function *OuterFn = OI.getFunction(); 446 CodeExtractorAnalysisCache CEAC(*OuterFn); 447 CodeExtractor Extractor(Blocks, /* DominatorTree */ nullptr, 448 /* AggregateArgs */ true, 449 /* BlockFrequencyInfo */ nullptr, 450 /* BranchProbabilityInfo */ nullptr, 451 /* AssumptionCache */ nullptr, 452 /* AllowVarArgs */ true, 453 /* AllowAlloca */ true, 454 /* AllocaBlock*/ OI.OuterAllocaBB, 455 /* Suffix */ ".omp_par"); 456 457 LLVM_DEBUG(dbgs() << "Before outlining: " << *OuterFn << "\n"); 458 LLVM_DEBUG(dbgs() << "Entry " << OI.EntryBB->getName() 459 << " Exit: " << OI.ExitBB->getName() << "\n"); 460 assert(Extractor.isEligible() && 461 "Expected OpenMP outlining to be possible!"); 462 463 for (auto *V : OI.ExcludeArgsFromAggregate) 464 Extractor.excludeArgFromAggregate(V); 465 466 Function *OutlinedFn = Extractor.extractCodeRegion(CEAC); 467 468 LLVM_DEBUG(dbgs() << "After outlining: " << *OuterFn << "\n"); 469 LLVM_DEBUG(dbgs() << " Outlined function: " << *OutlinedFn << "\n"); 470 assert(OutlinedFn->getReturnType()->isVoidTy() && 471 "OpenMP outlined functions should not return a value!"); 472 473 // For compability with the clang CG we move the outlined function after the 474 // one with the parallel region. 475 OutlinedFn->removeFromParent(); 476 M.getFunctionList().insertAfter(OuterFn->getIterator(), OutlinedFn); 477 478 // Remove the artificial entry introduced by the extractor right away, we 479 // made our own entry block after all. 480 { 481 BasicBlock &ArtificialEntry = OutlinedFn->getEntryBlock(); 482 assert(ArtificialEntry.getUniqueSuccessor() == OI.EntryBB); 483 assert(OI.EntryBB->getUniquePredecessor() == &ArtificialEntry); 484 // Move instructions from the to-be-deleted ArtificialEntry to the entry 485 // basic block of the parallel region. CodeExtractor generates 486 // instructions to unwrap the aggregate argument and may sink 487 // allocas/bitcasts for values that are solely used in the outlined region 488 // and do not escape. 489 assert(!ArtificialEntry.empty() && 490 "Expected instructions to add in the outlined region entry"); 491 for (BasicBlock::reverse_iterator It = ArtificialEntry.rbegin(), 492 End = ArtificialEntry.rend(); 493 It != End;) { 494 Instruction &I = *It; 495 It++; 496 497 if (I.isTerminator()) 498 continue; 499 500 I.moveBefore(*OI.EntryBB, OI.EntryBB->getFirstInsertionPt()); 501 } 502 503 OI.EntryBB->moveBefore(&ArtificialEntry); 504 ArtificialEntry.eraseFromParent(); 505 } 506 assert(&OutlinedFn->getEntryBlock() == OI.EntryBB); 507 assert(OutlinedFn && OutlinedFn->getNumUses() == 1); 508 509 // Run a user callback, e.g. to add attributes. 510 if (OI.PostOutlineCB) 511 OI.PostOutlineCB(*OutlinedFn); 512 } 513 514 // Remove work items that have been completed. 515 OutlineInfos = std::move(DeferredOutlines); 516 } 517 518 OpenMPIRBuilder::~OpenMPIRBuilder() { 519 assert(OutlineInfos.empty() && "There must be no outstanding outlinings"); 520 } 521 522 GlobalValue *OpenMPIRBuilder::createGlobalFlag(unsigned Value, StringRef Name) { 523 IntegerType *I32Ty = Type::getInt32Ty(M.getContext()); 524 auto *GV = 525 new GlobalVariable(M, I32Ty, 526 /* isConstant = */ true, GlobalValue::WeakODRLinkage, 527 ConstantInt::get(I32Ty, Value), Name); 528 GV->setVisibility(GlobalValue::HiddenVisibility); 529 530 return GV; 531 } 532 533 Constant *OpenMPIRBuilder::getOrCreateIdent(Constant *SrcLocStr, 534 uint32_t SrcLocStrSize, 535 IdentFlag LocFlags, 536 unsigned Reserve2Flags) { 537 // Enable "C-mode". 538 LocFlags |= OMP_IDENT_FLAG_KMPC; 539 540 Constant *&Ident = 541 IdentMap[{SrcLocStr, uint64_t(LocFlags) << 31 | Reserve2Flags}]; 542 if (!Ident) { 543 Constant *I32Null = ConstantInt::getNullValue(Int32); 544 Constant *IdentData[] = {I32Null, 545 ConstantInt::get(Int32, uint32_t(LocFlags)), 546 ConstantInt::get(Int32, Reserve2Flags), 547 ConstantInt::get(Int32, SrcLocStrSize), SrcLocStr}; 548 Constant *Initializer = 549 ConstantStruct::get(OpenMPIRBuilder::Ident, IdentData); 550 551 // Look for existing encoding of the location + flags, not needed but 552 // minimizes the difference to the existing solution while we transition. 553 for (GlobalVariable &GV : M.getGlobalList()) 554 if (GV.getValueType() == OpenMPIRBuilder::Ident && GV.hasInitializer()) 555 if (GV.getInitializer() == Initializer) 556 Ident = &GV; 557 558 if (!Ident) { 559 auto *GV = new GlobalVariable( 560 M, OpenMPIRBuilder::Ident, 561 /* isConstant = */ true, GlobalValue::PrivateLinkage, Initializer, "", 562 nullptr, GlobalValue::NotThreadLocal, 563 M.getDataLayout().getDefaultGlobalsAddressSpace()); 564 GV->setUnnamedAddr(GlobalValue::UnnamedAddr::Global); 565 GV->setAlignment(Align(8)); 566 Ident = GV; 567 } 568 } 569 570 return ConstantExpr::getPointerBitCastOrAddrSpaceCast(Ident, IdentPtr); 571 } 572 573 Constant *OpenMPIRBuilder::getOrCreateSrcLocStr(StringRef LocStr, 574 uint32_t &SrcLocStrSize) { 575 SrcLocStrSize = LocStr.size(); 576 Constant *&SrcLocStr = SrcLocStrMap[LocStr]; 577 if (!SrcLocStr) { 578 Constant *Initializer = 579 ConstantDataArray::getString(M.getContext(), LocStr); 580 581 // Look for existing encoding of the location, not needed but minimizes the 582 // difference to the existing solution while we transition. 583 for (GlobalVariable &GV : M.getGlobalList()) 584 if (GV.isConstant() && GV.hasInitializer() && 585 GV.getInitializer() == Initializer) 586 return SrcLocStr = ConstantExpr::getPointerCast(&GV, Int8Ptr); 587 588 SrcLocStr = Builder.CreateGlobalStringPtr(LocStr, /* Name */ "", 589 /* AddressSpace */ 0, &M); 590 } 591 return SrcLocStr; 592 } 593 594 Constant *OpenMPIRBuilder::getOrCreateSrcLocStr(StringRef FunctionName, 595 StringRef FileName, 596 unsigned Line, unsigned Column, 597 uint32_t &SrcLocStrSize) { 598 SmallString<128> Buffer; 599 Buffer.push_back(';'); 600 Buffer.append(FileName); 601 Buffer.push_back(';'); 602 Buffer.append(FunctionName); 603 Buffer.push_back(';'); 604 Buffer.append(std::to_string(Line)); 605 Buffer.push_back(';'); 606 Buffer.append(std::to_string(Column)); 607 Buffer.push_back(';'); 608 Buffer.push_back(';'); 609 return getOrCreateSrcLocStr(Buffer.str(), SrcLocStrSize); 610 } 611 612 Constant * 613 OpenMPIRBuilder::getOrCreateDefaultSrcLocStr(uint32_t &SrcLocStrSize) { 614 StringRef UnknownLoc = ";unknown;unknown;0;0;;"; 615 return getOrCreateSrcLocStr(UnknownLoc, SrcLocStrSize); 616 } 617 618 Constant *OpenMPIRBuilder::getOrCreateSrcLocStr(DebugLoc DL, 619 uint32_t &SrcLocStrSize, 620 Function *F) { 621 DILocation *DIL = DL.get(); 622 if (!DIL) 623 return getOrCreateDefaultSrcLocStr(SrcLocStrSize); 624 StringRef FileName = M.getName(); 625 if (DIFile *DIF = DIL->getFile()) 626 if (Optional<StringRef> Source = DIF->getSource()) 627 FileName = *Source; 628 StringRef Function = DIL->getScope()->getSubprogram()->getName(); 629 if (Function.empty() && F) 630 Function = F->getName(); 631 return getOrCreateSrcLocStr(Function, FileName, DIL->getLine(), 632 DIL->getColumn(), SrcLocStrSize); 633 } 634 635 Constant *OpenMPIRBuilder::getOrCreateSrcLocStr(const LocationDescription &Loc, 636 uint32_t &SrcLocStrSize) { 637 return getOrCreateSrcLocStr(Loc.DL, SrcLocStrSize, 638 Loc.IP.getBlock()->getParent()); 639 } 640 641 Value *OpenMPIRBuilder::getOrCreateThreadID(Value *Ident) { 642 return Builder.CreateCall( 643 getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_global_thread_num), Ident, 644 "omp_global_thread_num"); 645 } 646 647 OpenMPIRBuilder::InsertPointTy 648 OpenMPIRBuilder::createBarrier(const LocationDescription &Loc, Directive DK, 649 bool ForceSimpleCall, bool CheckCancelFlag) { 650 if (!updateToLocation(Loc)) 651 return Loc.IP; 652 return emitBarrierImpl(Loc, DK, ForceSimpleCall, CheckCancelFlag); 653 } 654 655 OpenMPIRBuilder::InsertPointTy 656 OpenMPIRBuilder::emitBarrierImpl(const LocationDescription &Loc, Directive Kind, 657 bool ForceSimpleCall, bool CheckCancelFlag) { 658 // Build call __kmpc_cancel_barrier(loc, thread_id) or 659 // __kmpc_barrier(loc, thread_id); 660 661 IdentFlag BarrierLocFlags; 662 switch (Kind) { 663 case OMPD_for: 664 BarrierLocFlags = OMP_IDENT_FLAG_BARRIER_IMPL_FOR; 665 break; 666 case OMPD_sections: 667 BarrierLocFlags = OMP_IDENT_FLAG_BARRIER_IMPL_SECTIONS; 668 break; 669 case OMPD_single: 670 BarrierLocFlags = OMP_IDENT_FLAG_BARRIER_IMPL_SINGLE; 671 break; 672 case OMPD_barrier: 673 BarrierLocFlags = OMP_IDENT_FLAG_BARRIER_EXPL; 674 break; 675 default: 676 BarrierLocFlags = OMP_IDENT_FLAG_BARRIER_IMPL; 677 break; 678 } 679 680 uint32_t SrcLocStrSize; 681 Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize); 682 Value *Args[] = { 683 getOrCreateIdent(SrcLocStr, SrcLocStrSize, BarrierLocFlags), 684 getOrCreateThreadID(getOrCreateIdent(SrcLocStr, SrcLocStrSize))}; 685 686 // If we are in a cancellable parallel region, barriers are cancellation 687 // points. 688 // TODO: Check why we would force simple calls or to ignore the cancel flag. 689 bool UseCancelBarrier = 690 !ForceSimpleCall && isLastFinalizationInfoCancellable(OMPD_parallel); 691 692 Value *Result = 693 Builder.CreateCall(getOrCreateRuntimeFunctionPtr( 694 UseCancelBarrier ? OMPRTL___kmpc_cancel_barrier 695 : OMPRTL___kmpc_barrier), 696 Args); 697 698 if (UseCancelBarrier && CheckCancelFlag) 699 emitCancelationCheckImpl(Result, OMPD_parallel); 700 701 return Builder.saveIP(); 702 } 703 704 OpenMPIRBuilder::InsertPointTy 705 OpenMPIRBuilder::createCancel(const LocationDescription &Loc, 706 Value *IfCondition, 707 omp::Directive CanceledDirective) { 708 if (!updateToLocation(Loc)) 709 return Loc.IP; 710 711 // LLVM utilities like blocks with terminators. 712 auto *UI = Builder.CreateUnreachable(); 713 714 Instruction *ThenTI = UI, *ElseTI = nullptr; 715 if (IfCondition) 716 SplitBlockAndInsertIfThenElse(IfCondition, UI, &ThenTI, &ElseTI); 717 Builder.SetInsertPoint(ThenTI); 718 719 Value *CancelKind = nullptr; 720 switch (CanceledDirective) { 721 #define OMP_CANCEL_KIND(Enum, Str, DirectiveEnum, Value) \ 722 case DirectiveEnum: \ 723 CancelKind = Builder.getInt32(Value); \ 724 break; 725 #include "llvm/Frontend/OpenMP/OMPKinds.def" 726 default: 727 llvm_unreachable("Unknown cancel kind!"); 728 } 729 730 uint32_t SrcLocStrSize; 731 Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize); 732 Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize); 733 Value *Args[] = {Ident, getOrCreateThreadID(Ident), CancelKind}; 734 Value *Result = Builder.CreateCall( 735 getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_cancel), Args); 736 auto ExitCB = [this, CanceledDirective, Loc](InsertPointTy IP) { 737 if (CanceledDirective == OMPD_parallel) { 738 IRBuilder<>::InsertPointGuard IPG(Builder); 739 Builder.restoreIP(IP); 740 createBarrier(LocationDescription(Builder.saveIP(), Loc.DL), 741 omp::Directive::OMPD_unknown, /* ForceSimpleCall */ false, 742 /* CheckCancelFlag */ false); 743 } 744 }; 745 746 // The actual cancel logic is shared with others, e.g., cancel_barriers. 747 emitCancelationCheckImpl(Result, CanceledDirective, ExitCB); 748 749 // Update the insertion point and remove the terminator we introduced. 750 Builder.SetInsertPoint(UI->getParent()); 751 UI->eraseFromParent(); 752 753 return Builder.saveIP(); 754 } 755 756 void OpenMPIRBuilder::emitOffloadingEntry(Constant *Addr, StringRef Name, 757 uint64_t Size, int32_t Flags, 758 StringRef SectionName) { 759 Type *Int8PtrTy = Type::getInt8PtrTy(M.getContext()); 760 Type *Int32Ty = Type::getInt32Ty(M.getContext()); 761 Type *SizeTy = M.getDataLayout().getIntPtrType(M.getContext()); 762 763 Constant *AddrName = ConstantDataArray::getString(M.getContext(), Name); 764 765 // Create the constant string used to look up the symbol in the device. 766 auto *Str = 767 new llvm::GlobalVariable(M, AddrName->getType(), /*isConstant=*/true, 768 llvm::GlobalValue::InternalLinkage, AddrName, 769 ".omp_offloading.entry_name"); 770 Str->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Global); 771 772 // Construct the offloading entry. 773 Constant *EntryData[] = { 774 ConstantExpr::getPointerBitCastOrAddrSpaceCast(Addr, Int8PtrTy), 775 ConstantExpr::getPointerBitCastOrAddrSpaceCast(Str, Int8PtrTy), 776 ConstantInt::get(SizeTy, Size), 777 ConstantInt::get(Int32Ty, Flags), 778 ConstantInt::get(Int32Ty, 0), 779 }; 780 Constant *EntryInitializer = 781 ConstantStruct::get(OpenMPIRBuilder::OffloadEntry, EntryData); 782 783 auto *Entry = new GlobalVariable( 784 M, OpenMPIRBuilder::OffloadEntry, 785 /* isConstant = */ true, GlobalValue::WeakAnyLinkage, EntryInitializer, 786 ".omp_offloading.entry." + Name, nullptr, GlobalValue::NotThreadLocal, 787 M.getDataLayout().getDefaultGlobalsAddressSpace()); 788 789 // The entry has to be created in the section the linker expects it to be. 790 Entry->setSection(SectionName); 791 Entry->setAlignment(Align(1)); 792 } 793 794 OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::emitTargetKernel( 795 const LocationDescription &Loc, Value *&Return, Value *Ident, 796 Value *DeviceID, Value *NumTeams, Value *NumThreads, Value *HostPtr, 797 ArrayRef<Value *> KernelArgs, ArrayRef<Value *> NoWaitArgs) { 798 if (!updateToLocation(Loc)) 799 return Loc.IP; 800 801 auto *KernelArgsPtr = 802 Builder.CreateAlloca(OpenMPIRBuilder::KernelArgs, nullptr, "kernel_args"); 803 for (unsigned I = 0, Size = KernelArgs.size(); I != Size; ++I) { 804 llvm::Value *Arg = 805 Builder.CreateStructGEP(OpenMPIRBuilder::KernelArgs, KernelArgsPtr, I); 806 Builder.CreateAlignedStore( 807 KernelArgs[I], Arg, 808 M.getDataLayout().getPrefTypeAlign(KernelArgs[I]->getType())); 809 } 810 811 bool HasNoWait = !NoWaitArgs.empty(); 812 SmallVector<Value *> OffloadingArgs{Ident, DeviceID, NumTeams, 813 NumThreads, HostPtr, KernelArgsPtr}; 814 if (HasNoWait) 815 OffloadingArgs.append(NoWaitArgs.begin(), NoWaitArgs.end()); 816 817 Return = Builder.CreateCall( 818 HasNoWait 819 ? getOrCreateRuntimeFunction(M, OMPRTL___tgt_target_kernel_nowait) 820 : getOrCreateRuntimeFunction(M, OMPRTL___tgt_target_kernel), 821 OffloadingArgs); 822 823 return Builder.saveIP(); 824 } 825 826 void OpenMPIRBuilder::emitCancelationCheckImpl(Value *CancelFlag, 827 omp::Directive CanceledDirective, 828 FinalizeCallbackTy ExitCB) { 829 assert(isLastFinalizationInfoCancellable(CanceledDirective) && 830 "Unexpected cancellation!"); 831 832 // For a cancel barrier we create two new blocks. 833 BasicBlock *BB = Builder.GetInsertBlock(); 834 BasicBlock *NonCancellationBlock; 835 if (Builder.GetInsertPoint() == BB->end()) { 836 // TODO: This branch will not be needed once we moved to the 837 // OpenMPIRBuilder codegen completely. 838 NonCancellationBlock = BasicBlock::Create( 839 BB->getContext(), BB->getName() + ".cont", BB->getParent()); 840 } else { 841 NonCancellationBlock = SplitBlock(BB, &*Builder.GetInsertPoint()); 842 BB->getTerminator()->eraseFromParent(); 843 Builder.SetInsertPoint(BB); 844 } 845 BasicBlock *CancellationBlock = BasicBlock::Create( 846 BB->getContext(), BB->getName() + ".cncl", BB->getParent()); 847 848 // Jump to them based on the return value. 849 Value *Cmp = Builder.CreateIsNull(CancelFlag); 850 Builder.CreateCondBr(Cmp, NonCancellationBlock, CancellationBlock, 851 /* TODO weight */ nullptr, nullptr); 852 853 // From the cancellation block we finalize all variables and go to the 854 // post finalization block that is known to the FiniCB callback. 855 Builder.SetInsertPoint(CancellationBlock); 856 if (ExitCB) 857 ExitCB(Builder.saveIP()); 858 auto &FI = FinalizationStack.back(); 859 FI.FiniCB(Builder.saveIP()); 860 861 // The continuation block is where code generation continues. 862 Builder.SetInsertPoint(NonCancellationBlock, NonCancellationBlock->begin()); 863 } 864 865 IRBuilder<>::InsertPoint OpenMPIRBuilder::createParallel( 866 const LocationDescription &Loc, InsertPointTy OuterAllocaIP, 867 BodyGenCallbackTy BodyGenCB, PrivatizeCallbackTy PrivCB, 868 FinalizeCallbackTy FiniCB, Value *IfCondition, Value *NumThreads, 869 omp::ProcBindKind ProcBind, bool IsCancellable) { 870 assert(!isConflictIP(Loc.IP, OuterAllocaIP) && "IPs must not be ambiguous"); 871 872 if (!updateToLocation(Loc)) 873 return Loc.IP; 874 875 uint32_t SrcLocStrSize; 876 Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize); 877 Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize); 878 Value *ThreadID = getOrCreateThreadID(Ident); 879 880 if (NumThreads) { 881 // Build call __kmpc_push_num_threads(&Ident, global_tid, num_threads) 882 Value *Args[] = { 883 Ident, ThreadID, 884 Builder.CreateIntCast(NumThreads, Int32, /*isSigned*/ false)}; 885 Builder.CreateCall( 886 getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_push_num_threads), Args); 887 } 888 889 if (ProcBind != OMP_PROC_BIND_default) { 890 // Build call __kmpc_push_proc_bind(&Ident, global_tid, proc_bind) 891 Value *Args[] = { 892 Ident, ThreadID, 893 ConstantInt::get(Int32, unsigned(ProcBind), /*isSigned=*/true)}; 894 Builder.CreateCall( 895 getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_push_proc_bind), Args); 896 } 897 898 BasicBlock *InsertBB = Builder.GetInsertBlock(); 899 Function *OuterFn = InsertBB->getParent(); 900 901 // Save the outer alloca block because the insertion iterator may get 902 // invalidated and we still need this later. 903 BasicBlock *OuterAllocaBlock = OuterAllocaIP.getBlock(); 904 905 // Vector to remember instructions we used only during the modeling but which 906 // we want to delete at the end. 907 SmallVector<Instruction *, 4> ToBeDeleted; 908 909 // Change the location to the outer alloca insertion point to create and 910 // initialize the allocas we pass into the parallel region. 911 Builder.restoreIP(OuterAllocaIP); 912 AllocaInst *TIDAddr = Builder.CreateAlloca(Int32, nullptr, "tid.addr"); 913 AllocaInst *ZeroAddr = Builder.CreateAlloca(Int32, nullptr, "zero.addr"); 914 915 // If there is an if condition we actually use the TIDAddr and ZeroAddr in the 916 // program, otherwise we only need them for modeling purposes to get the 917 // associated arguments in the outlined function. In the former case, 918 // initialize the allocas properly, in the latter case, delete them later. 919 if (IfCondition) { 920 Builder.CreateStore(Constant::getNullValue(Int32), TIDAddr); 921 Builder.CreateStore(Constant::getNullValue(Int32), ZeroAddr); 922 } else { 923 ToBeDeleted.push_back(TIDAddr); 924 ToBeDeleted.push_back(ZeroAddr); 925 } 926 927 // Create an artificial insertion point that will also ensure the blocks we 928 // are about to split are not degenerated. 929 auto *UI = new UnreachableInst(Builder.getContext(), InsertBB); 930 931 Instruction *ThenTI = UI, *ElseTI = nullptr; 932 if (IfCondition) 933 SplitBlockAndInsertIfThenElse(IfCondition, UI, &ThenTI, &ElseTI); 934 935 BasicBlock *ThenBB = ThenTI->getParent(); 936 BasicBlock *PRegEntryBB = ThenBB->splitBasicBlock(ThenTI, "omp.par.entry"); 937 BasicBlock *PRegBodyBB = 938 PRegEntryBB->splitBasicBlock(ThenTI, "omp.par.region"); 939 BasicBlock *PRegPreFiniBB = 940 PRegBodyBB->splitBasicBlock(ThenTI, "omp.par.pre_finalize"); 941 BasicBlock *PRegExitBB = 942 PRegPreFiniBB->splitBasicBlock(ThenTI, "omp.par.exit"); 943 944 auto FiniCBWrapper = [&](InsertPointTy IP) { 945 // Hide "open-ended" blocks from the given FiniCB by setting the right jump 946 // target to the region exit block. 947 if (IP.getBlock()->end() == IP.getPoint()) { 948 IRBuilder<>::InsertPointGuard IPG(Builder); 949 Builder.restoreIP(IP); 950 Instruction *I = Builder.CreateBr(PRegExitBB); 951 IP = InsertPointTy(I->getParent(), I->getIterator()); 952 } 953 assert(IP.getBlock()->getTerminator()->getNumSuccessors() == 1 && 954 IP.getBlock()->getTerminator()->getSuccessor(0) == PRegExitBB && 955 "Unexpected insertion point for finalization call!"); 956 return FiniCB(IP); 957 }; 958 959 FinalizationStack.push_back({FiniCBWrapper, OMPD_parallel, IsCancellable}); 960 961 // Generate the privatization allocas in the block that will become the entry 962 // of the outlined function. 963 Builder.SetInsertPoint(PRegEntryBB->getTerminator()); 964 InsertPointTy InnerAllocaIP = Builder.saveIP(); 965 966 AllocaInst *PrivTIDAddr = 967 Builder.CreateAlloca(Int32, nullptr, "tid.addr.local"); 968 Instruction *PrivTID = Builder.CreateLoad(Int32, PrivTIDAddr, "tid"); 969 970 // Add some fake uses for OpenMP provided arguments. 971 ToBeDeleted.push_back(Builder.CreateLoad(Int32, TIDAddr, "tid.addr.use")); 972 Instruction *ZeroAddrUse = 973 Builder.CreateLoad(Int32, ZeroAddr, "zero.addr.use"); 974 ToBeDeleted.push_back(ZeroAddrUse); 975 976 // ThenBB 977 // | 978 // V 979 // PRegionEntryBB <- Privatization allocas are placed here. 980 // | 981 // V 982 // PRegionBodyBB <- BodeGen is invoked here. 983 // | 984 // V 985 // PRegPreFiniBB <- The block we will start finalization from. 986 // | 987 // V 988 // PRegionExitBB <- A common exit to simplify block collection. 989 // 990 991 LLVM_DEBUG(dbgs() << "Before body codegen: " << *OuterFn << "\n"); 992 993 // Let the caller create the body. 994 assert(BodyGenCB && "Expected body generation callback!"); 995 InsertPointTy CodeGenIP(PRegBodyBB, PRegBodyBB->begin()); 996 BodyGenCB(InnerAllocaIP, CodeGenIP); 997 998 LLVM_DEBUG(dbgs() << "After body codegen: " << *OuterFn << "\n"); 999 1000 FunctionCallee RTLFn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_fork_call); 1001 if (auto *F = dyn_cast<llvm::Function>(RTLFn.getCallee())) { 1002 if (!F->hasMetadata(llvm::LLVMContext::MD_callback)) { 1003 llvm::LLVMContext &Ctx = F->getContext(); 1004 MDBuilder MDB(Ctx); 1005 // Annotate the callback behavior of the __kmpc_fork_call: 1006 // - The callback callee is argument number 2 (microtask). 1007 // - The first two arguments of the callback callee are unknown (-1). 1008 // - All variadic arguments to the __kmpc_fork_call are passed to the 1009 // callback callee. 1010 F->addMetadata( 1011 llvm::LLVMContext::MD_callback, 1012 *llvm::MDNode::get( 1013 Ctx, {MDB.createCallbackEncoding(2, {-1, -1}, 1014 /* VarArgsArePassed */ true)})); 1015 } 1016 } 1017 1018 OutlineInfo OI; 1019 OI.PostOutlineCB = [=](Function &OutlinedFn) { 1020 // Add some known attributes. 1021 OutlinedFn.addParamAttr(0, Attribute::NoAlias); 1022 OutlinedFn.addParamAttr(1, Attribute::NoAlias); 1023 OutlinedFn.addFnAttr(Attribute::NoUnwind); 1024 OutlinedFn.addFnAttr(Attribute::NoRecurse); 1025 1026 assert(OutlinedFn.arg_size() >= 2 && 1027 "Expected at least tid and bounded tid as arguments"); 1028 unsigned NumCapturedVars = 1029 OutlinedFn.arg_size() - /* tid & bounded tid */ 2; 1030 1031 CallInst *CI = cast<CallInst>(OutlinedFn.user_back()); 1032 CI->getParent()->setName("omp_parallel"); 1033 Builder.SetInsertPoint(CI); 1034 1035 // Build call __kmpc_fork_call(Ident, n, microtask, var1, .., varn); 1036 Value *ForkCallArgs[] = { 1037 Ident, Builder.getInt32(NumCapturedVars), 1038 Builder.CreateBitCast(&OutlinedFn, ParallelTaskPtr)}; 1039 1040 SmallVector<Value *, 16> RealArgs; 1041 RealArgs.append(std::begin(ForkCallArgs), std::end(ForkCallArgs)); 1042 RealArgs.append(CI->arg_begin() + /* tid & bound tid */ 2, CI->arg_end()); 1043 1044 Builder.CreateCall(RTLFn, RealArgs); 1045 1046 LLVM_DEBUG(dbgs() << "With fork_call placed: " 1047 << *Builder.GetInsertBlock()->getParent() << "\n"); 1048 1049 InsertPointTy ExitIP(PRegExitBB, PRegExitBB->end()); 1050 1051 // Initialize the local TID stack location with the argument value. 1052 Builder.SetInsertPoint(PrivTID); 1053 Function::arg_iterator OutlinedAI = OutlinedFn.arg_begin(); 1054 Builder.CreateStore(Builder.CreateLoad(Int32, OutlinedAI), PrivTIDAddr); 1055 1056 // If no "if" clause was present we do not need the call created during 1057 // outlining, otherwise we reuse it in the serialized parallel region. 1058 if (!ElseTI) { 1059 CI->eraseFromParent(); 1060 } else { 1061 1062 // If an "if" clause was present we are now generating the serialized 1063 // version into the "else" branch. 1064 Builder.SetInsertPoint(ElseTI); 1065 1066 // Build calls __kmpc_serialized_parallel(&Ident, GTid); 1067 Value *SerializedParallelCallArgs[] = {Ident, ThreadID}; 1068 Builder.CreateCall( 1069 getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_serialized_parallel), 1070 SerializedParallelCallArgs); 1071 1072 // OutlinedFn(>id, &zero, CapturedStruct); 1073 CI->removeFromParent(); 1074 Builder.Insert(CI); 1075 1076 // __kmpc_end_serialized_parallel(&Ident, GTid); 1077 Value *EndArgs[] = {Ident, ThreadID}; 1078 Builder.CreateCall( 1079 getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_end_serialized_parallel), 1080 EndArgs); 1081 1082 LLVM_DEBUG(dbgs() << "With serialized parallel region: " 1083 << *Builder.GetInsertBlock()->getParent() << "\n"); 1084 } 1085 1086 for (Instruction *I : ToBeDeleted) 1087 I->eraseFromParent(); 1088 }; 1089 1090 // Adjust the finalization stack, verify the adjustment, and call the 1091 // finalize function a last time to finalize values between the pre-fini 1092 // block and the exit block if we left the parallel "the normal way". 1093 auto FiniInfo = FinalizationStack.pop_back_val(); 1094 (void)FiniInfo; 1095 assert(FiniInfo.DK == OMPD_parallel && 1096 "Unexpected finalization stack state!"); 1097 1098 Instruction *PRegPreFiniTI = PRegPreFiniBB->getTerminator(); 1099 1100 InsertPointTy PreFiniIP(PRegPreFiniBB, PRegPreFiniTI->getIterator()); 1101 FiniCB(PreFiniIP); 1102 1103 OI.OuterAllocaBB = OuterAllocaBlock; 1104 OI.EntryBB = PRegEntryBB; 1105 OI.ExitBB = PRegExitBB; 1106 1107 SmallPtrSet<BasicBlock *, 32> ParallelRegionBlockSet; 1108 SmallVector<BasicBlock *, 32> Blocks; 1109 OI.collectBlocks(ParallelRegionBlockSet, Blocks); 1110 1111 // Ensure a single exit node for the outlined region by creating one. 1112 // We might have multiple incoming edges to the exit now due to finalizations, 1113 // e.g., cancel calls that cause the control flow to leave the region. 1114 BasicBlock *PRegOutlinedExitBB = PRegExitBB; 1115 PRegExitBB = SplitBlock(PRegExitBB, &*PRegExitBB->getFirstInsertionPt()); 1116 PRegOutlinedExitBB->setName("omp.par.outlined.exit"); 1117 Blocks.push_back(PRegOutlinedExitBB); 1118 1119 CodeExtractorAnalysisCache CEAC(*OuterFn); 1120 CodeExtractor Extractor(Blocks, /* DominatorTree */ nullptr, 1121 /* AggregateArgs */ false, 1122 /* BlockFrequencyInfo */ nullptr, 1123 /* BranchProbabilityInfo */ nullptr, 1124 /* AssumptionCache */ nullptr, 1125 /* AllowVarArgs */ true, 1126 /* AllowAlloca */ true, 1127 /* AllocationBlock */ OuterAllocaBlock, 1128 /* Suffix */ ".omp_par"); 1129 1130 // Find inputs to, outputs from the code region. 1131 BasicBlock *CommonExit = nullptr; 1132 SetVector<Value *> Inputs, Outputs, SinkingCands, HoistingCands; 1133 Extractor.findAllocas(CEAC, SinkingCands, HoistingCands, CommonExit); 1134 Extractor.findInputsOutputs(Inputs, Outputs, SinkingCands); 1135 1136 LLVM_DEBUG(dbgs() << "Before privatization: " << *OuterFn << "\n"); 1137 1138 FunctionCallee TIDRTLFn = 1139 getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_global_thread_num); 1140 1141 auto PrivHelper = [&](Value &V) { 1142 if (&V == TIDAddr || &V == ZeroAddr) { 1143 OI.ExcludeArgsFromAggregate.push_back(&V); 1144 return; 1145 } 1146 1147 SetVector<Use *> Uses; 1148 for (Use &U : V.uses()) 1149 if (auto *UserI = dyn_cast<Instruction>(U.getUser())) 1150 if (ParallelRegionBlockSet.count(UserI->getParent())) 1151 Uses.insert(&U); 1152 1153 // __kmpc_fork_call expects extra arguments as pointers. If the input 1154 // already has a pointer type, everything is fine. Otherwise, store the 1155 // value onto stack and load it back inside the to-be-outlined region. This 1156 // will ensure only the pointer will be passed to the function. 1157 // FIXME: if there are more than 15 trailing arguments, they must be 1158 // additionally packed in a struct. 1159 Value *Inner = &V; 1160 if (!V.getType()->isPointerTy()) { 1161 IRBuilder<>::InsertPointGuard Guard(Builder); 1162 LLVM_DEBUG(llvm::dbgs() << "Forwarding input as pointer: " << V << "\n"); 1163 1164 Builder.restoreIP(OuterAllocaIP); 1165 Value *Ptr = 1166 Builder.CreateAlloca(V.getType(), nullptr, V.getName() + ".reloaded"); 1167 1168 // Store to stack at end of the block that currently branches to the entry 1169 // block of the to-be-outlined region. 1170 Builder.SetInsertPoint(InsertBB, 1171 InsertBB->getTerminator()->getIterator()); 1172 Builder.CreateStore(&V, Ptr); 1173 1174 // Load back next to allocations in the to-be-outlined region. 1175 Builder.restoreIP(InnerAllocaIP); 1176 Inner = Builder.CreateLoad(V.getType(), Ptr); 1177 } 1178 1179 Value *ReplacementValue = nullptr; 1180 CallInst *CI = dyn_cast<CallInst>(&V); 1181 if (CI && CI->getCalledFunction() == TIDRTLFn.getCallee()) { 1182 ReplacementValue = PrivTID; 1183 } else { 1184 Builder.restoreIP( 1185 PrivCB(InnerAllocaIP, Builder.saveIP(), V, *Inner, ReplacementValue)); 1186 assert(ReplacementValue && 1187 "Expected copy/create callback to set replacement value!"); 1188 if (ReplacementValue == &V) 1189 return; 1190 } 1191 1192 for (Use *UPtr : Uses) 1193 UPtr->set(ReplacementValue); 1194 }; 1195 1196 // Reset the inner alloca insertion as it will be used for loading the values 1197 // wrapped into pointers before passing them into the to-be-outlined region. 1198 // Configure it to insert immediately after the fake use of zero address so 1199 // that they are available in the generated body and so that the 1200 // OpenMP-related values (thread ID and zero address pointers) remain leading 1201 // in the argument list. 1202 InnerAllocaIP = IRBuilder<>::InsertPoint( 1203 ZeroAddrUse->getParent(), ZeroAddrUse->getNextNode()->getIterator()); 1204 1205 // Reset the outer alloca insertion point to the entry of the relevant block 1206 // in case it was invalidated. 1207 OuterAllocaIP = IRBuilder<>::InsertPoint( 1208 OuterAllocaBlock, OuterAllocaBlock->getFirstInsertionPt()); 1209 1210 for (Value *Input : Inputs) { 1211 LLVM_DEBUG(dbgs() << "Captured input: " << *Input << "\n"); 1212 PrivHelper(*Input); 1213 } 1214 LLVM_DEBUG({ 1215 for (Value *Output : Outputs) 1216 LLVM_DEBUG(dbgs() << "Captured output: " << *Output << "\n"); 1217 }); 1218 assert(Outputs.empty() && 1219 "OpenMP outlining should not produce live-out values!"); 1220 1221 LLVM_DEBUG(dbgs() << "After privatization: " << *OuterFn << "\n"); 1222 LLVM_DEBUG({ 1223 for (auto *BB : Blocks) 1224 dbgs() << " PBR: " << BB->getName() << "\n"; 1225 }); 1226 1227 // Register the outlined info. 1228 addOutlineInfo(std::move(OI)); 1229 1230 InsertPointTy AfterIP(UI->getParent(), UI->getParent()->end()); 1231 UI->eraseFromParent(); 1232 1233 return AfterIP; 1234 } 1235 1236 void OpenMPIRBuilder::emitFlush(const LocationDescription &Loc) { 1237 // Build call void __kmpc_flush(ident_t *loc) 1238 uint32_t SrcLocStrSize; 1239 Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize); 1240 Value *Args[] = {getOrCreateIdent(SrcLocStr, SrcLocStrSize)}; 1241 1242 Builder.CreateCall(getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_flush), Args); 1243 } 1244 1245 void OpenMPIRBuilder::createFlush(const LocationDescription &Loc) { 1246 if (!updateToLocation(Loc)) 1247 return; 1248 emitFlush(Loc); 1249 } 1250 1251 void OpenMPIRBuilder::emitTaskwaitImpl(const LocationDescription &Loc) { 1252 // Build call kmp_int32 __kmpc_omp_taskwait(ident_t *loc, kmp_int32 1253 // global_tid); 1254 uint32_t SrcLocStrSize; 1255 Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize); 1256 Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize); 1257 Value *Args[] = {Ident, getOrCreateThreadID(Ident)}; 1258 1259 // Ignore return result until untied tasks are supported. 1260 Builder.CreateCall(getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_omp_taskwait), 1261 Args); 1262 } 1263 1264 void OpenMPIRBuilder::createTaskwait(const LocationDescription &Loc) { 1265 if (!updateToLocation(Loc)) 1266 return; 1267 emitTaskwaitImpl(Loc); 1268 } 1269 1270 void OpenMPIRBuilder::emitTaskyieldImpl(const LocationDescription &Loc) { 1271 // Build call __kmpc_omp_taskyield(loc, thread_id, 0); 1272 uint32_t SrcLocStrSize; 1273 Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize); 1274 Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize); 1275 Constant *I32Null = ConstantInt::getNullValue(Int32); 1276 Value *Args[] = {Ident, getOrCreateThreadID(Ident), I32Null}; 1277 1278 Builder.CreateCall(getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_omp_taskyield), 1279 Args); 1280 } 1281 1282 void OpenMPIRBuilder::createTaskyield(const LocationDescription &Loc) { 1283 if (!updateToLocation(Loc)) 1284 return; 1285 emitTaskyieldImpl(Loc); 1286 } 1287 1288 OpenMPIRBuilder::InsertPointTy 1289 OpenMPIRBuilder::createTask(const LocationDescription &Loc, 1290 InsertPointTy AllocaIP, BodyGenCallbackTy BodyGenCB, 1291 bool Tied, Value *Final) { 1292 if (!updateToLocation(Loc)) 1293 return InsertPointTy(); 1294 1295 uint32_t SrcLocStrSize; 1296 Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize); 1297 Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize); 1298 // The current basic block is split into four basic blocks. After outlining, 1299 // they will be mapped as follows: 1300 // ``` 1301 // def current_fn() { 1302 // current_basic_block: 1303 // br label %task.exit 1304 // task.exit: 1305 // ; instructions after task 1306 // } 1307 // def outlined_fn() { 1308 // task.alloca: 1309 // br label %task.body 1310 // task.body: 1311 // ret void 1312 // } 1313 // ``` 1314 BasicBlock *TaskExitBB = splitBB(Builder, /*CreateBranch=*/true, "task.exit"); 1315 BasicBlock *TaskBodyBB = splitBB(Builder, /*CreateBranch=*/true, "task.body"); 1316 BasicBlock *TaskAllocaBB = 1317 splitBB(Builder, /*CreateBranch=*/true, "task.alloca"); 1318 1319 OutlineInfo OI; 1320 OI.EntryBB = TaskAllocaBB; 1321 OI.OuterAllocaBB = AllocaIP.getBlock(); 1322 OI.ExitBB = TaskExitBB; 1323 OI.PostOutlineCB = [this, Ident, Tied, Final](Function &OutlinedFn) { 1324 // The input IR here looks like the following- 1325 // ``` 1326 // func @current_fn() { 1327 // outlined_fn(%args) 1328 // } 1329 // func @outlined_fn(%args) { ... } 1330 // ``` 1331 // 1332 // This is changed to the following- 1333 // 1334 // ``` 1335 // func @current_fn() { 1336 // runtime_call(..., wrapper_fn, ...) 1337 // } 1338 // func @wrapper_fn(..., %args) { 1339 // outlined_fn(%args) 1340 // } 1341 // func @outlined_fn(%args) { ... } 1342 // ``` 1343 1344 // The stale call instruction will be replaced with a new call instruction 1345 // for runtime call with a wrapper function. 1346 assert(OutlinedFn.getNumUses() == 1 && 1347 "there must be a single user for the outlined function"); 1348 CallInst *StaleCI = cast<CallInst>(OutlinedFn.user_back()); 1349 1350 // HasTaskData is true if any variables are captured in the outlined region, 1351 // false otherwise. 1352 bool HasTaskData = StaleCI->arg_size() > 0; 1353 Builder.SetInsertPoint(StaleCI); 1354 1355 // Gather the arguments for emitting the runtime call for 1356 // @__kmpc_omp_task_alloc 1357 Function *TaskAllocFn = 1358 getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_omp_task_alloc); 1359 1360 // Arguments - `loc_ref` (Ident) and `gtid` (ThreadID) 1361 // call. 1362 Value *ThreadID = getOrCreateThreadID(Ident); 1363 1364 // Argument - `flags` 1365 // Task is tied iff (Flags & 1) == 1. 1366 // Task is untied iff (Flags & 1) == 0. 1367 // Task is final iff (Flags & 2) == 2. 1368 // Task is not final iff (Flags & 2) == 0. 1369 // TODO: Handle the other flags. 1370 Value *Flags = Builder.getInt32(Tied); 1371 if (Final) { 1372 Value *FinalFlag = 1373 Builder.CreateSelect(Final, Builder.getInt32(2), Builder.getInt32(0)); 1374 Flags = Builder.CreateOr(FinalFlag, Flags); 1375 } 1376 1377 // Argument - `sizeof_kmp_task_t` (TaskSize) 1378 // Tasksize refers to the size in bytes of kmp_task_t data structure 1379 // including private vars accessed in task. 1380 Value *TaskSize = Builder.getInt64(0); 1381 if (HasTaskData) { 1382 AllocaInst *ArgStructAlloca = 1383 dyn_cast<AllocaInst>(StaleCI->getArgOperand(0)); 1384 assert(ArgStructAlloca && 1385 "Unable to find the alloca instruction corresponding to arguments " 1386 "for extracted function"); 1387 StructType *ArgStructType = 1388 dyn_cast<StructType>(ArgStructAlloca->getAllocatedType()); 1389 assert(ArgStructType && "Unable to find struct type corresponding to " 1390 "arguments for extracted function"); 1391 TaskSize = 1392 Builder.getInt64(M.getDataLayout().getTypeStoreSize(ArgStructType)); 1393 } 1394 1395 // TODO: Argument - sizeof_shareds 1396 1397 // Argument - task_entry (the wrapper function) 1398 // If the outlined function has some captured variables (i.e. HasTaskData is 1399 // true), then the wrapper function will have an additional argument (the 1400 // struct containing captured variables). Otherwise, no such argument will 1401 // be present. 1402 SmallVector<Type *> WrapperArgTys{Builder.getInt32Ty()}; 1403 if (HasTaskData) 1404 WrapperArgTys.push_back(OutlinedFn.getArg(0)->getType()); 1405 FunctionCallee WrapperFuncVal = M.getOrInsertFunction( 1406 (Twine(OutlinedFn.getName()) + ".wrapper").str(), 1407 FunctionType::get(Builder.getInt32Ty(), WrapperArgTys, false)); 1408 Function *WrapperFunc = dyn_cast<Function>(WrapperFuncVal.getCallee()); 1409 PointerType *WrapperFuncBitcastType = 1410 FunctionType::get(Builder.getInt32Ty(), 1411 {Builder.getInt32Ty(), Builder.getInt8PtrTy()}, false) 1412 ->getPointerTo(); 1413 Value *WrapperFuncBitcast = 1414 ConstantExpr::getBitCast(WrapperFunc, WrapperFuncBitcastType); 1415 1416 // Emit the @__kmpc_omp_task_alloc runtime call 1417 // The runtime call returns a pointer to an area where the task captured 1418 // variables must be copied before the task is run (NewTaskData) 1419 CallInst *NewTaskData = Builder.CreateCall( 1420 TaskAllocFn, 1421 {/*loc_ref=*/Ident, /*gtid=*/ThreadID, /*flags=*/Flags, 1422 /*sizeof_task=*/TaskSize, /*sizeof_shared=*/Builder.getInt64(0), 1423 /*task_func=*/WrapperFuncBitcast}); 1424 1425 // Copy the arguments for outlined function 1426 if (HasTaskData) { 1427 Value *TaskData = StaleCI->getArgOperand(0); 1428 Align Alignment = TaskData->getPointerAlignment(M.getDataLayout()); 1429 Builder.CreateMemCpy(NewTaskData, Alignment, TaskData, Alignment, 1430 TaskSize); 1431 } 1432 1433 // Emit the @__kmpc_omp_task runtime call to spawn the task 1434 Function *TaskFn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_omp_task); 1435 Builder.CreateCall(TaskFn, {Ident, ThreadID, NewTaskData}); 1436 1437 StaleCI->eraseFromParent(); 1438 1439 // Emit the body for wrapper function 1440 BasicBlock *WrapperEntryBB = 1441 BasicBlock::Create(M.getContext(), "", WrapperFunc); 1442 Builder.SetInsertPoint(WrapperEntryBB); 1443 if (HasTaskData) 1444 Builder.CreateCall(&OutlinedFn, {WrapperFunc->getArg(1)}); 1445 else 1446 Builder.CreateCall(&OutlinedFn); 1447 Builder.CreateRet(Builder.getInt32(0)); 1448 }; 1449 1450 addOutlineInfo(std::move(OI)); 1451 1452 InsertPointTy TaskAllocaIP = 1453 InsertPointTy(TaskAllocaBB, TaskAllocaBB->begin()); 1454 InsertPointTy TaskBodyIP = InsertPointTy(TaskBodyBB, TaskBodyBB->begin()); 1455 BodyGenCB(TaskAllocaIP, TaskBodyIP); 1456 Builder.SetInsertPoint(TaskExitBB, TaskExitBB->begin()); 1457 1458 return Builder.saveIP(); 1459 } 1460 1461 OpenMPIRBuilder::InsertPointTy 1462 OpenMPIRBuilder::createTaskgroup(const LocationDescription &Loc, 1463 InsertPointTy AllocaIP, 1464 BodyGenCallbackTy BodyGenCB) { 1465 if (!updateToLocation(Loc)) 1466 return InsertPointTy(); 1467 1468 uint32_t SrcLocStrSize; 1469 Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize); 1470 Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize); 1471 Value *ThreadID = getOrCreateThreadID(Ident); 1472 1473 // Emit the @__kmpc_taskgroup runtime call to start the taskgroup 1474 Function *TaskgroupFn = 1475 getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_taskgroup); 1476 Builder.CreateCall(TaskgroupFn, {Ident, ThreadID}); 1477 1478 BasicBlock *TaskgroupExitBB = splitBB(Builder, true, "taskgroup.exit"); 1479 BodyGenCB(AllocaIP, Builder.saveIP()); 1480 1481 Builder.SetInsertPoint(TaskgroupExitBB); 1482 // Emit the @__kmpc_end_taskgroup runtime call to end the taskgroup 1483 Function *EndTaskgroupFn = 1484 getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_end_taskgroup); 1485 Builder.CreateCall(EndTaskgroupFn, {Ident, ThreadID}); 1486 1487 return Builder.saveIP(); 1488 } 1489 1490 OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::createSections( 1491 const LocationDescription &Loc, InsertPointTy AllocaIP, 1492 ArrayRef<StorableBodyGenCallbackTy> SectionCBs, PrivatizeCallbackTy PrivCB, 1493 FinalizeCallbackTy FiniCB, bool IsCancellable, bool IsNowait) { 1494 assert(!isConflictIP(AllocaIP, Loc.IP) && "Dedicated IP allocas required"); 1495 1496 if (!updateToLocation(Loc)) 1497 return Loc.IP; 1498 1499 auto FiniCBWrapper = [&](InsertPointTy IP) { 1500 if (IP.getBlock()->end() != IP.getPoint()) 1501 return FiniCB(IP); 1502 // This must be done otherwise any nested constructs using FinalizeOMPRegion 1503 // will fail because that function requires the Finalization Basic Block to 1504 // have a terminator, which is already removed by EmitOMPRegionBody. 1505 // IP is currently at cancelation block. 1506 // We need to backtrack to the condition block to fetch 1507 // the exit block and create a branch from cancelation 1508 // to exit block. 1509 IRBuilder<>::InsertPointGuard IPG(Builder); 1510 Builder.restoreIP(IP); 1511 auto *CaseBB = IP.getBlock()->getSinglePredecessor(); 1512 auto *CondBB = CaseBB->getSinglePredecessor()->getSinglePredecessor(); 1513 auto *ExitBB = CondBB->getTerminator()->getSuccessor(1); 1514 Instruction *I = Builder.CreateBr(ExitBB); 1515 IP = InsertPointTy(I->getParent(), I->getIterator()); 1516 return FiniCB(IP); 1517 }; 1518 1519 FinalizationStack.push_back({FiniCBWrapper, OMPD_sections, IsCancellable}); 1520 1521 // Each section is emitted as a switch case 1522 // Each finalization callback is handled from clang.EmitOMPSectionDirective() 1523 // -> OMP.createSection() which generates the IR for each section 1524 // Iterate through all sections and emit a switch construct: 1525 // switch (IV) { 1526 // case 0: 1527 // <SectionStmt[0]>; 1528 // break; 1529 // ... 1530 // case <NumSection> - 1: 1531 // <SectionStmt[<NumSection> - 1]>; 1532 // break; 1533 // } 1534 // ... 1535 // section_loop.after: 1536 // <FiniCB>; 1537 auto LoopBodyGenCB = [&](InsertPointTy CodeGenIP, Value *IndVar) { 1538 Builder.restoreIP(CodeGenIP); 1539 BasicBlock *Continue = 1540 splitBBWithSuffix(Builder, /*CreateBranch=*/false, ".sections.after"); 1541 Function *CurFn = Continue->getParent(); 1542 SwitchInst *SwitchStmt = Builder.CreateSwitch(IndVar, Continue); 1543 1544 unsigned CaseNumber = 0; 1545 for (auto SectionCB : SectionCBs) { 1546 BasicBlock *CaseBB = BasicBlock::Create( 1547 M.getContext(), "omp_section_loop.body.case", CurFn, Continue); 1548 SwitchStmt->addCase(Builder.getInt32(CaseNumber), CaseBB); 1549 Builder.SetInsertPoint(CaseBB); 1550 BranchInst *CaseEndBr = Builder.CreateBr(Continue); 1551 SectionCB(InsertPointTy(), 1552 {CaseEndBr->getParent(), CaseEndBr->getIterator()}); 1553 CaseNumber++; 1554 } 1555 // remove the existing terminator from body BB since there can be no 1556 // terminators after switch/case 1557 }; 1558 // Loop body ends here 1559 // LowerBound, UpperBound, and STride for createCanonicalLoop 1560 Type *I32Ty = Type::getInt32Ty(M.getContext()); 1561 Value *LB = ConstantInt::get(I32Ty, 0); 1562 Value *UB = ConstantInt::get(I32Ty, SectionCBs.size()); 1563 Value *ST = ConstantInt::get(I32Ty, 1); 1564 llvm::CanonicalLoopInfo *LoopInfo = createCanonicalLoop( 1565 Loc, LoopBodyGenCB, LB, UB, ST, true, false, AllocaIP, "section_loop"); 1566 InsertPointTy AfterIP = 1567 applyStaticWorkshareLoop(Loc.DL, LoopInfo, AllocaIP, !IsNowait); 1568 1569 // Apply the finalization callback in LoopAfterBB 1570 auto FiniInfo = FinalizationStack.pop_back_val(); 1571 assert(FiniInfo.DK == OMPD_sections && 1572 "Unexpected finalization stack state!"); 1573 if (FinalizeCallbackTy &CB = FiniInfo.FiniCB) { 1574 Builder.restoreIP(AfterIP); 1575 BasicBlock *FiniBB = 1576 splitBBWithSuffix(Builder, /*CreateBranch=*/true, "sections.fini"); 1577 CB(Builder.saveIP()); 1578 AfterIP = {FiniBB, FiniBB->begin()}; 1579 } 1580 1581 return AfterIP; 1582 } 1583 1584 OpenMPIRBuilder::InsertPointTy 1585 OpenMPIRBuilder::createSection(const LocationDescription &Loc, 1586 BodyGenCallbackTy BodyGenCB, 1587 FinalizeCallbackTy FiniCB) { 1588 if (!updateToLocation(Loc)) 1589 return Loc.IP; 1590 1591 auto FiniCBWrapper = [&](InsertPointTy IP) { 1592 if (IP.getBlock()->end() != IP.getPoint()) 1593 return FiniCB(IP); 1594 // This must be done otherwise any nested constructs using FinalizeOMPRegion 1595 // will fail because that function requires the Finalization Basic Block to 1596 // have a terminator, which is already removed by EmitOMPRegionBody. 1597 // IP is currently at cancelation block. 1598 // We need to backtrack to the condition block to fetch 1599 // the exit block and create a branch from cancelation 1600 // to exit block. 1601 IRBuilder<>::InsertPointGuard IPG(Builder); 1602 Builder.restoreIP(IP); 1603 auto *CaseBB = Loc.IP.getBlock(); 1604 auto *CondBB = CaseBB->getSinglePredecessor()->getSinglePredecessor(); 1605 auto *ExitBB = CondBB->getTerminator()->getSuccessor(1); 1606 Instruction *I = Builder.CreateBr(ExitBB); 1607 IP = InsertPointTy(I->getParent(), I->getIterator()); 1608 return FiniCB(IP); 1609 }; 1610 1611 Directive OMPD = Directive::OMPD_sections; 1612 // Since we are using Finalization Callback here, HasFinalize 1613 // and IsCancellable have to be true 1614 return EmitOMPInlinedRegion(OMPD, nullptr, nullptr, BodyGenCB, FiniCBWrapper, 1615 /*Conditional*/ false, /*hasFinalize*/ true, 1616 /*IsCancellable*/ true); 1617 } 1618 1619 /// Create a function with a unique name and a "void (i8*, i8*)" signature in 1620 /// the given module and return it. 1621 Function *getFreshReductionFunc(Module &M) { 1622 Type *VoidTy = Type::getVoidTy(M.getContext()); 1623 Type *Int8PtrTy = Type::getInt8PtrTy(M.getContext()); 1624 auto *FuncTy = 1625 FunctionType::get(VoidTy, {Int8PtrTy, Int8PtrTy}, /* IsVarArg */ false); 1626 return Function::Create(FuncTy, GlobalVariable::InternalLinkage, 1627 M.getDataLayout().getDefaultGlobalsAddressSpace(), 1628 ".omp.reduction.func", &M); 1629 } 1630 1631 OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::createReductions( 1632 const LocationDescription &Loc, InsertPointTy AllocaIP, 1633 ArrayRef<ReductionInfo> ReductionInfos, bool IsNoWait) { 1634 for (const ReductionInfo &RI : ReductionInfos) { 1635 (void)RI; 1636 assert(RI.Variable && "expected non-null variable"); 1637 assert(RI.PrivateVariable && "expected non-null private variable"); 1638 assert(RI.ReductionGen && "expected non-null reduction generator callback"); 1639 assert(RI.Variable->getType() == RI.PrivateVariable->getType() && 1640 "expected variables and their private equivalents to have the same " 1641 "type"); 1642 assert(RI.Variable->getType()->isPointerTy() && 1643 "expected variables to be pointers"); 1644 } 1645 1646 if (!updateToLocation(Loc)) 1647 return InsertPointTy(); 1648 1649 BasicBlock *InsertBlock = Loc.IP.getBlock(); 1650 BasicBlock *ContinuationBlock = 1651 InsertBlock->splitBasicBlock(Loc.IP.getPoint(), "reduce.finalize"); 1652 InsertBlock->getTerminator()->eraseFromParent(); 1653 1654 // Create and populate array of type-erased pointers to private reduction 1655 // values. 1656 unsigned NumReductions = ReductionInfos.size(); 1657 Type *RedArrayTy = ArrayType::get(Builder.getInt8PtrTy(), NumReductions); 1658 Builder.restoreIP(AllocaIP); 1659 Value *RedArray = Builder.CreateAlloca(RedArrayTy, nullptr, "red.array"); 1660 1661 Builder.SetInsertPoint(InsertBlock, InsertBlock->end()); 1662 1663 for (auto En : enumerate(ReductionInfos)) { 1664 unsigned Index = En.index(); 1665 const ReductionInfo &RI = En.value(); 1666 Value *RedArrayElemPtr = Builder.CreateConstInBoundsGEP2_64( 1667 RedArrayTy, RedArray, 0, Index, "red.array.elem." + Twine(Index)); 1668 Value *Casted = 1669 Builder.CreateBitCast(RI.PrivateVariable, Builder.getInt8PtrTy(), 1670 "private.red.var." + Twine(Index) + ".casted"); 1671 Builder.CreateStore(Casted, RedArrayElemPtr); 1672 } 1673 1674 // Emit a call to the runtime function that orchestrates the reduction. 1675 // Declare the reduction function in the process. 1676 Function *Func = Builder.GetInsertBlock()->getParent(); 1677 Module *Module = Func->getParent(); 1678 Value *RedArrayPtr = 1679 Builder.CreateBitCast(RedArray, Builder.getInt8PtrTy(), "red.array.ptr"); 1680 uint32_t SrcLocStrSize; 1681 Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize); 1682 bool CanGenerateAtomic = 1683 llvm::all_of(ReductionInfos, [](const ReductionInfo &RI) { 1684 return RI.AtomicReductionGen; 1685 }); 1686 Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize, 1687 CanGenerateAtomic 1688 ? IdentFlag::OMP_IDENT_FLAG_ATOMIC_REDUCE 1689 : IdentFlag(0)); 1690 Value *ThreadId = getOrCreateThreadID(Ident); 1691 Constant *NumVariables = Builder.getInt32(NumReductions); 1692 const DataLayout &DL = Module->getDataLayout(); 1693 unsigned RedArrayByteSize = DL.getTypeStoreSize(RedArrayTy); 1694 Constant *RedArraySize = Builder.getInt64(RedArrayByteSize); 1695 Function *ReductionFunc = getFreshReductionFunc(*Module); 1696 Value *Lock = getOMPCriticalRegionLock(".reduction"); 1697 Function *ReduceFunc = getOrCreateRuntimeFunctionPtr( 1698 IsNoWait ? RuntimeFunction::OMPRTL___kmpc_reduce_nowait 1699 : RuntimeFunction::OMPRTL___kmpc_reduce); 1700 CallInst *ReduceCall = 1701 Builder.CreateCall(ReduceFunc, 1702 {Ident, ThreadId, NumVariables, RedArraySize, 1703 RedArrayPtr, ReductionFunc, Lock}, 1704 "reduce"); 1705 1706 // Create final reduction entry blocks for the atomic and non-atomic case. 1707 // Emit IR that dispatches control flow to one of the blocks based on the 1708 // reduction supporting the atomic mode. 1709 BasicBlock *NonAtomicRedBlock = 1710 BasicBlock::Create(Module->getContext(), "reduce.switch.nonatomic", Func); 1711 BasicBlock *AtomicRedBlock = 1712 BasicBlock::Create(Module->getContext(), "reduce.switch.atomic", Func); 1713 SwitchInst *Switch = 1714 Builder.CreateSwitch(ReduceCall, ContinuationBlock, /* NumCases */ 2); 1715 Switch->addCase(Builder.getInt32(1), NonAtomicRedBlock); 1716 Switch->addCase(Builder.getInt32(2), AtomicRedBlock); 1717 1718 // Populate the non-atomic reduction using the elementwise reduction function. 1719 // This loads the elements from the global and private variables and reduces 1720 // them before storing back the result to the global variable. 1721 Builder.SetInsertPoint(NonAtomicRedBlock); 1722 for (auto En : enumerate(ReductionInfos)) { 1723 const ReductionInfo &RI = En.value(); 1724 Type *ValueType = RI.ElementType; 1725 Value *RedValue = Builder.CreateLoad(ValueType, RI.Variable, 1726 "red.value." + Twine(En.index())); 1727 Value *PrivateRedValue = 1728 Builder.CreateLoad(ValueType, RI.PrivateVariable, 1729 "red.private.value." + Twine(En.index())); 1730 Value *Reduced; 1731 Builder.restoreIP( 1732 RI.ReductionGen(Builder.saveIP(), RedValue, PrivateRedValue, Reduced)); 1733 if (!Builder.GetInsertBlock()) 1734 return InsertPointTy(); 1735 Builder.CreateStore(Reduced, RI.Variable); 1736 } 1737 Function *EndReduceFunc = getOrCreateRuntimeFunctionPtr( 1738 IsNoWait ? RuntimeFunction::OMPRTL___kmpc_end_reduce_nowait 1739 : RuntimeFunction::OMPRTL___kmpc_end_reduce); 1740 Builder.CreateCall(EndReduceFunc, {Ident, ThreadId, Lock}); 1741 Builder.CreateBr(ContinuationBlock); 1742 1743 // Populate the atomic reduction using the atomic elementwise reduction 1744 // function. There are no loads/stores here because they will be happening 1745 // inside the atomic elementwise reduction. 1746 Builder.SetInsertPoint(AtomicRedBlock); 1747 if (CanGenerateAtomic) { 1748 for (const ReductionInfo &RI : ReductionInfos) { 1749 Builder.restoreIP(RI.AtomicReductionGen(Builder.saveIP(), RI.ElementType, 1750 RI.Variable, RI.PrivateVariable)); 1751 if (!Builder.GetInsertBlock()) 1752 return InsertPointTy(); 1753 } 1754 Builder.CreateBr(ContinuationBlock); 1755 } else { 1756 Builder.CreateUnreachable(); 1757 } 1758 1759 // Populate the outlined reduction function using the elementwise reduction 1760 // function. Partial values are extracted from the type-erased array of 1761 // pointers to private variables. 1762 BasicBlock *ReductionFuncBlock = 1763 BasicBlock::Create(Module->getContext(), "", ReductionFunc); 1764 Builder.SetInsertPoint(ReductionFuncBlock); 1765 Value *LHSArrayPtr = Builder.CreateBitCast(ReductionFunc->getArg(0), 1766 RedArrayTy->getPointerTo()); 1767 Value *RHSArrayPtr = Builder.CreateBitCast(ReductionFunc->getArg(1), 1768 RedArrayTy->getPointerTo()); 1769 for (auto En : enumerate(ReductionInfos)) { 1770 const ReductionInfo &RI = En.value(); 1771 Value *LHSI8PtrPtr = Builder.CreateConstInBoundsGEP2_64( 1772 RedArrayTy, LHSArrayPtr, 0, En.index()); 1773 Value *LHSI8Ptr = Builder.CreateLoad(Builder.getInt8PtrTy(), LHSI8PtrPtr); 1774 Value *LHSPtr = Builder.CreateBitCast(LHSI8Ptr, RI.Variable->getType()); 1775 Value *LHS = Builder.CreateLoad(RI.ElementType, LHSPtr); 1776 Value *RHSI8PtrPtr = Builder.CreateConstInBoundsGEP2_64( 1777 RedArrayTy, RHSArrayPtr, 0, En.index()); 1778 Value *RHSI8Ptr = Builder.CreateLoad(Builder.getInt8PtrTy(), RHSI8PtrPtr); 1779 Value *RHSPtr = 1780 Builder.CreateBitCast(RHSI8Ptr, RI.PrivateVariable->getType()); 1781 Value *RHS = Builder.CreateLoad(RI.ElementType, RHSPtr); 1782 Value *Reduced; 1783 Builder.restoreIP(RI.ReductionGen(Builder.saveIP(), LHS, RHS, Reduced)); 1784 if (!Builder.GetInsertBlock()) 1785 return InsertPointTy(); 1786 Builder.CreateStore(Reduced, LHSPtr); 1787 } 1788 Builder.CreateRetVoid(); 1789 1790 Builder.SetInsertPoint(ContinuationBlock); 1791 return Builder.saveIP(); 1792 } 1793 1794 OpenMPIRBuilder::InsertPointTy 1795 OpenMPIRBuilder::createMaster(const LocationDescription &Loc, 1796 BodyGenCallbackTy BodyGenCB, 1797 FinalizeCallbackTy FiniCB) { 1798 1799 if (!updateToLocation(Loc)) 1800 return Loc.IP; 1801 1802 Directive OMPD = Directive::OMPD_master; 1803 uint32_t SrcLocStrSize; 1804 Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize); 1805 Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize); 1806 Value *ThreadId = getOrCreateThreadID(Ident); 1807 Value *Args[] = {Ident, ThreadId}; 1808 1809 Function *EntryRTLFn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_master); 1810 Instruction *EntryCall = Builder.CreateCall(EntryRTLFn, Args); 1811 1812 Function *ExitRTLFn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_end_master); 1813 Instruction *ExitCall = Builder.CreateCall(ExitRTLFn, Args); 1814 1815 return EmitOMPInlinedRegion(OMPD, EntryCall, ExitCall, BodyGenCB, FiniCB, 1816 /*Conditional*/ true, /*hasFinalize*/ true); 1817 } 1818 1819 OpenMPIRBuilder::InsertPointTy 1820 OpenMPIRBuilder::createMasked(const LocationDescription &Loc, 1821 BodyGenCallbackTy BodyGenCB, 1822 FinalizeCallbackTy FiniCB, Value *Filter) { 1823 if (!updateToLocation(Loc)) 1824 return Loc.IP; 1825 1826 Directive OMPD = Directive::OMPD_masked; 1827 uint32_t SrcLocStrSize; 1828 Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize); 1829 Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize); 1830 Value *ThreadId = getOrCreateThreadID(Ident); 1831 Value *Args[] = {Ident, ThreadId, Filter}; 1832 Value *ArgsEnd[] = {Ident, ThreadId}; 1833 1834 Function *EntryRTLFn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_masked); 1835 Instruction *EntryCall = Builder.CreateCall(EntryRTLFn, Args); 1836 1837 Function *ExitRTLFn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_end_masked); 1838 Instruction *ExitCall = Builder.CreateCall(ExitRTLFn, ArgsEnd); 1839 1840 return EmitOMPInlinedRegion(OMPD, EntryCall, ExitCall, BodyGenCB, FiniCB, 1841 /*Conditional*/ true, /*hasFinalize*/ true); 1842 } 1843 1844 CanonicalLoopInfo *OpenMPIRBuilder::createLoopSkeleton( 1845 DebugLoc DL, Value *TripCount, Function *F, BasicBlock *PreInsertBefore, 1846 BasicBlock *PostInsertBefore, const Twine &Name) { 1847 Module *M = F->getParent(); 1848 LLVMContext &Ctx = M->getContext(); 1849 Type *IndVarTy = TripCount->getType(); 1850 1851 // Create the basic block structure. 1852 BasicBlock *Preheader = 1853 BasicBlock::Create(Ctx, "omp_" + Name + ".preheader", F, PreInsertBefore); 1854 BasicBlock *Header = 1855 BasicBlock::Create(Ctx, "omp_" + Name + ".header", F, PreInsertBefore); 1856 BasicBlock *Cond = 1857 BasicBlock::Create(Ctx, "omp_" + Name + ".cond", F, PreInsertBefore); 1858 BasicBlock *Body = 1859 BasicBlock::Create(Ctx, "omp_" + Name + ".body", F, PreInsertBefore); 1860 BasicBlock *Latch = 1861 BasicBlock::Create(Ctx, "omp_" + Name + ".inc", F, PostInsertBefore); 1862 BasicBlock *Exit = 1863 BasicBlock::Create(Ctx, "omp_" + Name + ".exit", F, PostInsertBefore); 1864 BasicBlock *After = 1865 BasicBlock::Create(Ctx, "omp_" + Name + ".after", F, PostInsertBefore); 1866 1867 // Use specified DebugLoc for new instructions. 1868 Builder.SetCurrentDebugLocation(DL); 1869 1870 Builder.SetInsertPoint(Preheader); 1871 Builder.CreateBr(Header); 1872 1873 Builder.SetInsertPoint(Header); 1874 PHINode *IndVarPHI = Builder.CreatePHI(IndVarTy, 2, "omp_" + Name + ".iv"); 1875 IndVarPHI->addIncoming(ConstantInt::get(IndVarTy, 0), Preheader); 1876 Builder.CreateBr(Cond); 1877 1878 Builder.SetInsertPoint(Cond); 1879 Value *Cmp = 1880 Builder.CreateICmpULT(IndVarPHI, TripCount, "omp_" + Name + ".cmp"); 1881 Builder.CreateCondBr(Cmp, Body, Exit); 1882 1883 Builder.SetInsertPoint(Body); 1884 Builder.CreateBr(Latch); 1885 1886 Builder.SetInsertPoint(Latch); 1887 Value *Next = Builder.CreateAdd(IndVarPHI, ConstantInt::get(IndVarTy, 1), 1888 "omp_" + Name + ".next", /*HasNUW=*/true); 1889 Builder.CreateBr(Header); 1890 IndVarPHI->addIncoming(Next, Latch); 1891 1892 Builder.SetInsertPoint(Exit); 1893 Builder.CreateBr(After); 1894 1895 // Remember and return the canonical control flow. 1896 LoopInfos.emplace_front(); 1897 CanonicalLoopInfo *CL = &LoopInfos.front(); 1898 1899 CL->Header = Header; 1900 CL->Cond = Cond; 1901 CL->Latch = Latch; 1902 CL->Exit = Exit; 1903 1904 #ifndef NDEBUG 1905 CL->assertOK(); 1906 #endif 1907 return CL; 1908 } 1909 1910 CanonicalLoopInfo * 1911 OpenMPIRBuilder::createCanonicalLoop(const LocationDescription &Loc, 1912 LoopBodyGenCallbackTy BodyGenCB, 1913 Value *TripCount, const Twine &Name) { 1914 BasicBlock *BB = Loc.IP.getBlock(); 1915 BasicBlock *NextBB = BB->getNextNode(); 1916 1917 CanonicalLoopInfo *CL = createLoopSkeleton(Loc.DL, TripCount, BB->getParent(), 1918 NextBB, NextBB, Name); 1919 BasicBlock *After = CL->getAfter(); 1920 1921 // If location is not set, don't connect the loop. 1922 if (updateToLocation(Loc)) { 1923 // Split the loop at the insertion point: Branch to the preheader and move 1924 // every following instruction to after the loop (the After BB). Also, the 1925 // new successor is the loop's after block. 1926 spliceBB(Builder, After, /*CreateBranch=*/false); 1927 Builder.CreateBr(CL->getPreheader()); 1928 } 1929 1930 // Emit the body content. We do it after connecting the loop to the CFG to 1931 // avoid that the callback encounters degenerate BBs. 1932 BodyGenCB(CL->getBodyIP(), CL->getIndVar()); 1933 1934 #ifndef NDEBUG 1935 CL->assertOK(); 1936 #endif 1937 return CL; 1938 } 1939 1940 CanonicalLoopInfo *OpenMPIRBuilder::createCanonicalLoop( 1941 const LocationDescription &Loc, LoopBodyGenCallbackTy BodyGenCB, 1942 Value *Start, Value *Stop, Value *Step, bool IsSigned, bool InclusiveStop, 1943 InsertPointTy ComputeIP, const Twine &Name) { 1944 1945 // Consider the following difficulties (assuming 8-bit signed integers): 1946 // * Adding \p Step to the loop counter which passes \p Stop may overflow: 1947 // DO I = 1, 100, 50 1948 /// * A \p Step of INT_MIN cannot not be normalized to a positive direction: 1949 // DO I = 100, 0, -128 1950 1951 // Start, Stop and Step must be of the same integer type. 1952 auto *IndVarTy = cast<IntegerType>(Start->getType()); 1953 assert(IndVarTy == Stop->getType() && "Stop type mismatch"); 1954 assert(IndVarTy == Step->getType() && "Step type mismatch"); 1955 1956 LocationDescription ComputeLoc = 1957 ComputeIP.isSet() ? LocationDescription(ComputeIP, Loc.DL) : Loc; 1958 updateToLocation(ComputeLoc); 1959 1960 ConstantInt *Zero = ConstantInt::get(IndVarTy, 0); 1961 ConstantInt *One = ConstantInt::get(IndVarTy, 1); 1962 1963 // Like Step, but always positive. 1964 Value *Incr = Step; 1965 1966 // Distance between Start and Stop; always positive. 1967 Value *Span; 1968 1969 // Condition whether there are no iterations are executed at all, e.g. because 1970 // UB < LB. 1971 Value *ZeroCmp; 1972 1973 if (IsSigned) { 1974 // Ensure that increment is positive. If not, negate and invert LB and UB. 1975 Value *IsNeg = Builder.CreateICmpSLT(Step, Zero); 1976 Incr = Builder.CreateSelect(IsNeg, Builder.CreateNeg(Step), Step); 1977 Value *LB = Builder.CreateSelect(IsNeg, Stop, Start); 1978 Value *UB = Builder.CreateSelect(IsNeg, Start, Stop); 1979 Span = Builder.CreateSub(UB, LB, "", false, true); 1980 ZeroCmp = Builder.CreateICmp( 1981 InclusiveStop ? CmpInst::ICMP_SLT : CmpInst::ICMP_SLE, UB, LB); 1982 } else { 1983 Span = Builder.CreateSub(Stop, Start, "", true); 1984 ZeroCmp = Builder.CreateICmp( 1985 InclusiveStop ? CmpInst::ICMP_ULT : CmpInst::ICMP_ULE, Stop, Start); 1986 } 1987 1988 Value *CountIfLooping; 1989 if (InclusiveStop) { 1990 CountIfLooping = Builder.CreateAdd(Builder.CreateUDiv(Span, Incr), One); 1991 } else { 1992 // Avoid incrementing past stop since it could overflow. 1993 Value *CountIfTwo = Builder.CreateAdd( 1994 Builder.CreateUDiv(Builder.CreateSub(Span, One), Incr), One); 1995 Value *OneCmp = Builder.CreateICmp( 1996 InclusiveStop ? CmpInst::ICMP_ULT : CmpInst::ICMP_ULE, Span, Incr); 1997 CountIfLooping = Builder.CreateSelect(OneCmp, One, CountIfTwo); 1998 } 1999 Value *TripCount = Builder.CreateSelect(ZeroCmp, Zero, CountIfLooping, 2000 "omp_" + Name + ".tripcount"); 2001 2002 auto BodyGen = [=](InsertPointTy CodeGenIP, Value *IV) { 2003 Builder.restoreIP(CodeGenIP); 2004 Value *Span = Builder.CreateMul(IV, Step); 2005 Value *IndVar = Builder.CreateAdd(Span, Start); 2006 BodyGenCB(Builder.saveIP(), IndVar); 2007 }; 2008 LocationDescription LoopLoc = ComputeIP.isSet() ? Loc.IP : Builder.saveIP(); 2009 return createCanonicalLoop(LoopLoc, BodyGen, TripCount, Name); 2010 } 2011 2012 // Returns an LLVM function to call for initializing loop bounds using OpenMP 2013 // static scheduling depending on `type`. Only i32 and i64 are supported by the 2014 // runtime. Always interpret integers as unsigned similarly to 2015 // CanonicalLoopInfo. 2016 static FunctionCallee getKmpcForStaticInitForType(Type *Ty, Module &M, 2017 OpenMPIRBuilder &OMPBuilder) { 2018 unsigned Bitwidth = Ty->getIntegerBitWidth(); 2019 if (Bitwidth == 32) 2020 return OMPBuilder.getOrCreateRuntimeFunction( 2021 M, omp::RuntimeFunction::OMPRTL___kmpc_for_static_init_4u); 2022 if (Bitwidth == 64) 2023 return OMPBuilder.getOrCreateRuntimeFunction( 2024 M, omp::RuntimeFunction::OMPRTL___kmpc_for_static_init_8u); 2025 llvm_unreachable("unknown OpenMP loop iterator bitwidth"); 2026 } 2027 2028 OpenMPIRBuilder::InsertPointTy 2029 OpenMPIRBuilder::applyStaticWorkshareLoop(DebugLoc DL, CanonicalLoopInfo *CLI, 2030 InsertPointTy AllocaIP, 2031 bool NeedsBarrier) { 2032 assert(CLI->isValid() && "Requires a valid canonical loop"); 2033 assert(!isConflictIP(AllocaIP, CLI->getPreheaderIP()) && 2034 "Require dedicated allocate IP"); 2035 2036 // Set up the source location value for OpenMP runtime. 2037 Builder.restoreIP(CLI->getPreheaderIP()); 2038 Builder.SetCurrentDebugLocation(DL); 2039 2040 uint32_t SrcLocStrSize; 2041 Constant *SrcLocStr = getOrCreateSrcLocStr(DL, SrcLocStrSize); 2042 Value *SrcLoc = getOrCreateIdent(SrcLocStr, SrcLocStrSize); 2043 2044 // Declare useful OpenMP runtime functions. 2045 Value *IV = CLI->getIndVar(); 2046 Type *IVTy = IV->getType(); 2047 FunctionCallee StaticInit = getKmpcForStaticInitForType(IVTy, M, *this); 2048 FunctionCallee StaticFini = 2049 getOrCreateRuntimeFunction(M, omp::OMPRTL___kmpc_for_static_fini); 2050 2051 // Allocate space for computed loop bounds as expected by the "init" function. 2052 Builder.restoreIP(AllocaIP); 2053 Type *I32Type = Type::getInt32Ty(M.getContext()); 2054 Value *PLastIter = Builder.CreateAlloca(I32Type, nullptr, "p.lastiter"); 2055 Value *PLowerBound = Builder.CreateAlloca(IVTy, nullptr, "p.lowerbound"); 2056 Value *PUpperBound = Builder.CreateAlloca(IVTy, nullptr, "p.upperbound"); 2057 Value *PStride = Builder.CreateAlloca(IVTy, nullptr, "p.stride"); 2058 2059 // At the end of the preheader, prepare for calling the "init" function by 2060 // storing the current loop bounds into the allocated space. A canonical loop 2061 // always iterates from 0 to trip-count with step 1. Note that "init" expects 2062 // and produces an inclusive upper bound. 2063 Builder.SetInsertPoint(CLI->getPreheader()->getTerminator()); 2064 Constant *Zero = ConstantInt::get(IVTy, 0); 2065 Constant *One = ConstantInt::get(IVTy, 1); 2066 Builder.CreateStore(Zero, PLowerBound); 2067 Value *UpperBound = Builder.CreateSub(CLI->getTripCount(), One); 2068 Builder.CreateStore(UpperBound, PUpperBound); 2069 Builder.CreateStore(One, PStride); 2070 2071 Value *ThreadNum = getOrCreateThreadID(SrcLoc); 2072 2073 Constant *SchedulingType = ConstantInt::get( 2074 I32Type, static_cast<int>(OMPScheduleType::UnorderedStatic)); 2075 2076 // Call the "init" function and update the trip count of the loop with the 2077 // value it produced. 2078 Builder.CreateCall(StaticInit, 2079 {SrcLoc, ThreadNum, SchedulingType, PLastIter, PLowerBound, 2080 PUpperBound, PStride, One, Zero}); 2081 Value *LowerBound = Builder.CreateLoad(IVTy, PLowerBound); 2082 Value *InclusiveUpperBound = Builder.CreateLoad(IVTy, PUpperBound); 2083 Value *TripCountMinusOne = Builder.CreateSub(InclusiveUpperBound, LowerBound); 2084 Value *TripCount = Builder.CreateAdd(TripCountMinusOne, One); 2085 CLI->setTripCount(TripCount); 2086 2087 // Update all uses of the induction variable except the one in the condition 2088 // block that compares it with the actual upper bound, and the increment in 2089 // the latch block. 2090 2091 CLI->mapIndVar([&](Instruction *OldIV) -> Value * { 2092 Builder.SetInsertPoint(CLI->getBody(), 2093 CLI->getBody()->getFirstInsertionPt()); 2094 Builder.SetCurrentDebugLocation(DL); 2095 return Builder.CreateAdd(OldIV, LowerBound); 2096 }); 2097 2098 // In the "exit" block, call the "fini" function. 2099 Builder.SetInsertPoint(CLI->getExit(), 2100 CLI->getExit()->getTerminator()->getIterator()); 2101 Builder.CreateCall(StaticFini, {SrcLoc, ThreadNum}); 2102 2103 // Add the barrier if requested. 2104 if (NeedsBarrier) 2105 createBarrier(LocationDescription(Builder.saveIP(), DL), 2106 omp::Directive::OMPD_for, /* ForceSimpleCall */ false, 2107 /* CheckCancelFlag */ false); 2108 2109 InsertPointTy AfterIP = CLI->getAfterIP(); 2110 CLI->invalidate(); 2111 2112 return AfterIP; 2113 } 2114 2115 OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::applyStaticChunkedWorkshareLoop( 2116 DebugLoc DL, CanonicalLoopInfo *CLI, InsertPointTy AllocaIP, 2117 bool NeedsBarrier, Value *ChunkSize) { 2118 assert(CLI->isValid() && "Requires a valid canonical loop"); 2119 assert(ChunkSize && "Chunk size is required"); 2120 2121 LLVMContext &Ctx = CLI->getFunction()->getContext(); 2122 Value *IV = CLI->getIndVar(); 2123 Value *OrigTripCount = CLI->getTripCount(); 2124 Type *IVTy = IV->getType(); 2125 assert(IVTy->getIntegerBitWidth() <= 64 && 2126 "Max supported tripcount bitwidth is 64 bits"); 2127 Type *InternalIVTy = IVTy->getIntegerBitWidth() <= 32 ? Type::getInt32Ty(Ctx) 2128 : Type::getInt64Ty(Ctx); 2129 Type *I32Type = Type::getInt32Ty(M.getContext()); 2130 Constant *Zero = ConstantInt::get(InternalIVTy, 0); 2131 Constant *One = ConstantInt::get(InternalIVTy, 1); 2132 2133 // Declare useful OpenMP runtime functions. 2134 FunctionCallee StaticInit = 2135 getKmpcForStaticInitForType(InternalIVTy, M, *this); 2136 FunctionCallee StaticFini = 2137 getOrCreateRuntimeFunction(M, omp::OMPRTL___kmpc_for_static_fini); 2138 2139 // Allocate space for computed loop bounds as expected by the "init" function. 2140 Builder.restoreIP(AllocaIP); 2141 Builder.SetCurrentDebugLocation(DL); 2142 Value *PLastIter = Builder.CreateAlloca(I32Type, nullptr, "p.lastiter"); 2143 Value *PLowerBound = 2144 Builder.CreateAlloca(InternalIVTy, nullptr, "p.lowerbound"); 2145 Value *PUpperBound = 2146 Builder.CreateAlloca(InternalIVTy, nullptr, "p.upperbound"); 2147 Value *PStride = Builder.CreateAlloca(InternalIVTy, nullptr, "p.stride"); 2148 2149 // Set up the source location value for the OpenMP runtime. 2150 Builder.restoreIP(CLI->getPreheaderIP()); 2151 Builder.SetCurrentDebugLocation(DL); 2152 2153 // TODO: Detect overflow in ubsan or max-out with current tripcount. 2154 Value *CastedChunkSize = 2155 Builder.CreateZExtOrTrunc(ChunkSize, InternalIVTy, "chunksize"); 2156 Value *CastedTripCount = 2157 Builder.CreateZExt(OrigTripCount, InternalIVTy, "tripcount"); 2158 2159 Constant *SchedulingType = ConstantInt::get( 2160 I32Type, static_cast<int>(OMPScheduleType::UnorderedStaticChunked)); 2161 Builder.CreateStore(Zero, PLowerBound); 2162 Value *OrigUpperBound = Builder.CreateSub(CastedTripCount, One); 2163 Builder.CreateStore(OrigUpperBound, PUpperBound); 2164 Builder.CreateStore(One, PStride); 2165 2166 // Call the "init" function and update the trip count of the loop with the 2167 // value it produced. 2168 uint32_t SrcLocStrSize; 2169 Constant *SrcLocStr = getOrCreateSrcLocStr(DL, SrcLocStrSize); 2170 Value *SrcLoc = getOrCreateIdent(SrcLocStr, SrcLocStrSize); 2171 Value *ThreadNum = getOrCreateThreadID(SrcLoc); 2172 Builder.CreateCall(StaticInit, 2173 {/*loc=*/SrcLoc, /*global_tid=*/ThreadNum, 2174 /*schedtype=*/SchedulingType, /*plastiter=*/PLastIter, 2175 /*plower=*/PLowerBound, /*pupper=*/PUpperBound, 2176 /*pstride=*/PStride, /*incr=*/One, 2177 /*chunk=*/CastedChunkSize}); 2178 2179 // Load values written by the "init" function. 2180 Value *FirstChunkStart = 2181 Builder.CreateLoad(InternalIVTy, PLowerBound, "omp_firstchunk.lb"); 2182 Value *FirstChunkStop = 2183 Builder.CreateLoad(InternalIVTy, PUpperBound, "omp_firstchunk.ub"); 2184 Value *FirstChunkEnd = Builder.CreateAdd(FirstChunkStop, One); 2185 Value *ChunkRange = 2186 Builder.CreateSub(FirstChunkEnd, FirstChunkStart, "omp_chunk.range"); 2187 Value *NextChunkStride = 2188 Builder.CreateLoad(InternalIVTy, PStride, "omp_dispatch.stride"); 2189 2190 // Create outer "dispatch" loop for enumerating the chunks. 2191 BasicBlock *DispatchEnter = splitBB(Builder, true); 2192 Value *DispatchCounter; 2193 CanonicalLoopInfo *DispatchCLI = createCanonicalLoop( 2194 {Builder.saveIP(), DL}, 2195 [&](InsertPointTy BodyIP, Value *Counter) { DispatchCounter = Counter; }, 2196 FirstChunkStart, CastedTripCount, NextChunkStride, 2197 /*IsSigned=*/false, /*InclusiveStop=*/false, /*ComputeIP=*/{}, 2198 "dispatch"); 2199 2200 // Remember the BasicBlocks of the dispatch loop we need, then invalidate to 2201 // not have to preserve the canonical invariant. 2202 BasicBlock *DispatchBody = DispatchCLI->getBody(); 2203 BasicBlock *DispatchLatch = DispatchCLI->getLatch(); 2204 BasicBlock *DispatchExit = DispatchCLI->getExit(); 2205 BasicBlock *DispatchAfter = DispatchCLI->getAfter(); 2206 DispatchCLI->invalidate(); 2207 2208 // Rewire the original loop to become the chunk loop inside the dispatch loop. 2209 redirectTo(DispatchAfter, CLI->getAfter(), DL); 2210 redirectTo(CLI->getExit(), DispatchLatch, DL); 2211 redirectTo(DispatchBody, DispatchEnter, DL); 2212 2213 // Prepare the prolog of the chunk loop. 2214 Builder.restoreIP(CLI->getPreheaderIP()); 2215 Builder.SetCurrentDebugLocation(DL); 2216 2217 // Compute the number of iterations of the chunk loop. 2218 Builder.SetInsertPoint(CLI->getPreheader()->getTerminator()); 2219 Value *ChunkEnd = Builder.CreateAdd(DispatchCounter, ChunkRange); 2220 Value *IsLastChunk = 2221 Builder.CreateICmpUGE(ChunkEnd, CastedTripCount, "omp_chunk.is_last"); 2222 Value *CountUntilOrigTripCount = 2223 Builder.CreateSub(CastedTripCount, DispatchCounter); 2224 Value *ChunkTripCount = Builder.CreateSelect( 2225 IsLastChunk, CountUntilOrigTripCount, ChunkRange, "omp_chunk.tripcount"); 2226 Value *BackcastedChunkTC = 2227 Builder.CreateTrunc(ChunkTripCount, IVTy, "omp_chunk.tripcount.trunc"); 2228 CLI->setTripCount(BackcastedChunkTC); 2229 2230 // Update all uses of the induction variable except the one in the condition 2231 // block that compares it with the actual upper bound, and the increment in 2232 // the latch block. 2233 Value *BackcastedDispatchCounter = 2234 Builder.CreateTrunc(DispatchCounter, IVTy, "omp_dispatch.iv.trunc"); 2235 CLI->mapIndVar([&](Instruction *) -> Value * { 2236 Builder.restoreIP(CLI->getBodyIP()); 2237 return Builder.CreateAdd(IV, BackcastedDispatchCounter); 2238 }); 2239 2240 // In the "exit" block, call the "fini" function. 2241 Builder.SetInsertPoint(DispatchExit, DispatchExit->getFirstInsertionPt()); 2242 Builder.CreateCall(StaticFini, {SrcLoc, ThreadNum}); 2243 2244 // Add the barrier if requested. 2245 if (NeedsBarrier) 2246 createBarrier(LocationDescription(Builder.saveIP(), DL), OMPD_for, 2247 /*ForceSimpleCall=*/false, /*CheckCancelFlag=*/false); 2248 2249 #ifndef NDEBUG 2250 // Even though we currently do not support applying additional methods to it, 2251 // the chunk loop should remain a canonical loop. 2252 CLI->assertOK(); 2253 #endif 2254 2255 return {DispatchAfter, DispatchAfter->getFirstInsertionPt()}; 2256 } 2257 2258 OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::applyWorkshareLoop( 2259 DebugLoc DL, CanonicalLoopInfo *CLI, InsertPointTy AllocaIP, 2260 bool NeedsBarrier, llvm::omp::ScheduleKind SchedKind, 2261 llvm::Value *ChunkSize, bool HasSimdModifier, bool HasMonotonicModifier, 2262 bool HasNonmonotonicModifier, bool HasOrderedClause) { 2263 OMPScheduleType EffectiveScheduleType = computeOpenMPScheduleType( 2264 SchedKind, ChunkSize, HasSimdModifier, HasMonotonicModifier, 2265 HasNonmonotonicModifier, HasOrderedClause); 2266 2267 bool IsOrdered = (EffectiveScheduleType & OMPScheduleType::ModifierOrdered) == 2268 OMPScheduleType::ModifierOrdered; 2269 switch (EffectiveScheduleType & ~OMPScheduleType::ModifierMask) { 2270 case OMPScheduleType::BaseStatic: 2271 assert(!ChunkSize && "No chunk size with static-chunked schedule"); 2272 if (IsOrdered) 2273 return applyDynamicWorkshareLoop(DL, CLI, AllocaIP, EffectiveScheduleType, 2274 NeedsBarrier, ChunkSize); 2275 // FIXME: Monotonicity ignored? 2276 return applyStaticWorkshareLoop(DL, CLI, AllocaIP, NeedsBarrier); 2277 2278 case OMPScheduleType::BaseStaticChunked: 2279 if (IsOrdered) 2280 return applyDynamicWorkshareLoop(DL, CLI, AllocaIP, EffectiveScheduleType, 2281 NeedsBarrier, ChunkSize); 2282 // FIXME: Monotonicity ignored? 2283 return applyStaticChunkedWorkshareLoop(DL, CLI, AllocaIP, NeedsBarrier, 2284 ChunkSize); 2285 2286 case OMPScheduleType::BaseRuntime: 2287 case OMPScheduleType::BaseAuto: 2288 case OMPScheduleType::BaseGreedy: 2289 case OMPScheduleType::BaseBalanced: 2290 case OMPScheduleType::BaseSteal: 2291 case OMPScheduleType::BaseGuidedSimd: 2292 case OMPScheduleType::BaseRuntimeSimd: 2293 assert(!ChunkSize && 2294 "schedule type does not support user-defined chunk sizes"); 2295 LLVM_FALLTHROUGH; 2296 case OMPScheduleType::BaseDynamicChunked: 2297 case OMPScheduleType::BaseGuidedChunked: 2298 case OMPScheduleType::BaseGuidedIterativeChunked: 2299 case OMPScheduleType::BaseGuidedAnalyticalChunked: 2300 case OMPScheduleType::BaseStaticBalancedChunked: 2301 return applyDynamicWorkshareLoop(DL, CLI, AllocaIP, EffectiveScheduleType, 2302 NeedsBarrier, ChunkSize); 2303 2304 default: 2305 llvm_unreachable("Unknown/unimplemented schedule kind"); 2306 } 2307 } 2308 2309 /// Returns an LLVM function to call for initializing loop bounds using OpenMP 2310 /// dynamic scheduling depending on `type`. Only i32 and i64 are supported by 2311 /// the runtime. Always interpret integers as unsigned similarly to 2312 /// CanonicalLoopInfo. 2313 static FunctionCallee 2314 getKmpcForDynamicInitForType(Type *Ty, Module &M, OpenMPIRBuilder &OMPBuilder) { 2315 unsigned Bitwidth = Ty->getIntegerBitWidth(); 2316 if (Bitwidth == 32) 2317 return OMPBuilder.getOrCreateRuntimeFunction( 2318 M, omp::RuntimeFunction::OMPRTL___kmpc_dispatch_init_4u); 2319 if (Bitwidth == 64) 2320 return OMPBuilder.getOrCreateRuntimeFunction( 2321 M, omp::RuntimeFunction::OMPRTL___kmpc_dispatch_init_8u); 2322 llvm_unreachable("unknown OpenMP loop iterator bitwidth"); 2323 } 2324 2325 /// Returns an LLVM function to call for updating the next loop using OpenMP 2326 /// dynamic scheduling depending on `type`. Only i32 and i64 are supported by 2327 /// the runtime. Always interpret integers as unsigned similarly to 2328 /// CanonicalLoopInfo. 2329 static FunctionCallee 2330 getKmpcForDynamicNextForType(Type *Ty, Module &M, OpenMPIRBuilder &OMPBuilder) { 2331 unsigned Bitwidth = Ty->getIntegerBitWidth(); 2332 if (Bitwidth == 32) 2333 return OMPBuilder.getOrCreateRuntimeFunction( 2334 M, omp::RuntimeFunction::OMPRTL___kmpc_dispatch_next_4u); 2335 if (Bitwidth == 64) 2336 return OMPBuilder.getOrCreateRuntimeFunction( 2337 M, omp::RuntimeFunction::OMPRTL___kmpc_dispatch_next_8u); 2338 llvm_unreachable("unknown OpenMP loop iterator bitwidth"); 2339 } 2340 2341 /// Returns an LLVM function to call for finalizing the dynamic loop using 2342 /// depending on `type`. Only i32 and i64 are supported by the runtime. Always 2343 /// interpret integers as unsigned similarly to CanonicalLoopInfo. 2344 static FunctionCallee 2345 getKmpcForDynamicFiniForType(Type *Ty, Module &M, OpenMPIRBuilder &OMPBuilder) { 2346 unsigned Bitwidth = Ty->getIntegerBitWidth(); 2347 if (Bitwidth == 32) 2348 return OMPBuilder.getOrCreateRuntimeFunction( 2349 M, omp::RuntimeFunction::OMPRTL___kmpc_dispatch_fini_4u); 2350 if (Bitwidth == 64) 2351 return OMPBuilder.getOrCreateRuntimeFunction( 2352 M, omp::RuntimeFunction::OMPRTL___kmpc_dispatch_fini_8u); 2353 llvm_unreachable("unknown OpenMP loop iterator bitwidth"); 2354 } 2355 2356 OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::applyDynamicWorkshareLoop( 2357 DebugLoc DL, CanonicalLoopInfo *CLI, InsertPointTy AllocaIP, 2358 OMPScheduleType SchedType, bool NeedsBarrier, Value *Chunk) { 2359 assert(CLI->isValid() && "Requires a valid canonical loop"); 2360 assert(!isConflictIP(AllocaIP, CLI->getPreheaderIP()) && 2361 "Require dedicated allocate IP"); 2362 assert(isValidWorkshareLoopScheduleType(SchedType) && 2363 "Require valid schedule type"); 2364 2365 bool Ordered = (SchedType & OMPScheduleType::ModifierOrdered) == 2366 OMPScheduleType::ModifierOrdered; 2367 2368 // Set up the source location value for OpenMP runtime. 2369 Builder.SetCurrentDebugLocation(DL); 2370 2371 uint32_t SrcLocStrSize; 2372 Constant *SrcLocStr = getOrCreateSrcLocStr(DL, SrcLocStrSize); 2373 Value *SrcLoc = getOrCreateIdent(SrcLocStr, SrcLocStrSize); 2374 2375 // Declare useful OpenMP runtime functions. 2376 Value *IV = CLI->getIndVar(); 2377 Type *IVTy = IV->getType(); 2378 FunctionCallee DynamicInit = getKmpcForDynamicInitForType(IVTy, M, *this); 2379 FunctionCallee DynamicNext = getKmpcForDynamicNextForType(IVTy, M, *this); 2380 2381 // Allocate space for computed loop bounds as expected by the "init" function. 2382 Builder.restoreIP(AllocaIP); 2383 Type *I32Type = Type::getInt32Ty(M.getContext()); 2384 Value *PLastIter = Builder.CreateAlloca(I32Type, nullptr, "p.lastiter"); 2385 Value *PLowerBound = Builder.CreateAlloca(IVTy, nullptr, "p.lowerbound"); 2386 Value *PUpperBound = Builder.CreateAlloca(IVTy, nullptr, "p.upperbound"); 2387 Value *PStride = Builder.CreateAlloca(IVTy, nullptr, "p.stride"); 2388 2389 // At the end of the preheader, prepare for calling the "init" function by 2390 // storing the current loop bounds into the allocated space. A canonical loop 2391 // always iterates from 0 to trip-count with step 1. Note that "init" expects 2392 // and produces an inclusive upper bound. 2393 BasicBlock *PreHeader = CLI->getPreheader(); 2394 Builder.SetInsertPoint(PreHeader->getTerminator()); 2395 Constant *One = ConstantInt::get(IVTy, 1); 2396 Builder.CreateStore(One, PLowerBound); 2397 Value *UpperBound = CLI->getTripCount(); 2398 Builder.CreateStore(UpperBound, PUpperBound); 2399 Builder.CreateStore(One, PStride); 2400 2401 BasicBlock *Header = CLI->getHeader(); 2402 BasicBlock *Exit = CLI->getExit(); 2403 BasicBlock *Cond = CLI->getCond(); 2404 BasicBlock *Latch = CLI->getLatch(); 2405 InsertPointTy AfterIP = CLI->getAfterIP(); 2406 2407 // The CLI will be "broken" in the code below, as the loop is no longer 2408 // a valid canonical loop. 2409 2410 if (!Chunk) 2411 Chunk = One; 2412 2413 Value *ThreadNum = getOrCreateThreadID(SrcLoc); 2414 2415 Constant *SchedulingType = 2416 ConstantInt::get(I32Type, static_cast<int>(SchedType)); 2417 2418 // Call the "init" function. 2419 Builder.CreateCall(DynamicInit, 2420 {SrcLoc, ThreadNum, SchedulingType, /* LowerBound */ One, 2421 UpperBound, /* step */ One, Chunk}); 2422 2423 // An outer loop around the existing one. 2424 BasicBlock *OuterCond = BasicBlock::Create( 2425 PreHeader->getContext(), Twine(PreHeader->getName()) + ".outer.cond", 2426 PreHeader->getParent()); 2427 // This needs to be 32-bit always, so can't use the IVTy Zero above. 2428 Builder.SetInsertPoint(OuterCond, OuterCond->getFirstInsertionPt()); 2429 Value *Res = 2430 Builder.CreateCall(DynamicNext, {SrcLoc, ThreadNum, PLastIter, 2431 PLowerBound, PUpperBound, PStride}); 2432 Constant *Zero32 = ConstantInt::get(I32Type, 0); 2433 Value *MoreWork = Builder.CreateCmp(CmpInst::ICMP_NE, Res, Zero32); 2434 Value *LowerBound = 2435 Builder.CreateSub(Builder.CreateLoad(IVTy, PLowerBound), One, "lb"); 2436 Builder.CreateCondBr(MoreWork, Header, Exit); 2437 2438 // Change PHI-node in loop header to use outer cond rather than preheader, 2439 // and set IV to the LowerBound. 2440 Instruction *Phi = &Header->front(); 2441 auto *PI = cast<PHINode>(Phi); 2442 PI->setIncomingBlock(0, OuterCond); 2443 PI->setIncomingValue(0, LowerBound); 2444 2445 // Then set the pre-header to jump to the OuterCond 2446 Instruction *Term = PreHeader->getTerminator(); 2447 auto *Br = cast<BranchInst>(Term); 2448 Br->setSuccessor(0, OuterCond); 2449 2450 // Modify the inner condition: 2451 // * Use the UpperBound returned from the DynamicNext call. 2452 // * jump to the loop outer loop when done with one of the inner loops. 2453 Builder.SetInsertPoint(Cond, Cond->getFirstInsertionPt()); 2454 UpperBound = Builder.CreateLoad(IVTy, PUpperBound, "ub"); 2455 Instruction *Comp = &*Builder.GetInsertPoint(); 2456 auto *CI = cast<CmpInst>(Comp); 2457 CI->setOperand(1, UpperBound); 2458 // Redirect the inner exit to branch to outer condition. 2459 Instruction *Branch = &Cond->back(); 2460 auto *BI = cast<BranchInst>(Branch); 2461 assert(BI->getSuccessor(1) == Exit); 2462 BI->setSuccessor(1, OuterCond); 2463 2464 // Call the "fini" function if "ordered" is present in wsloop directive. 2465 if (Ordered) { 2466 Builder.SetInsertPoint(&Latch->back()); 2467 FunctionCallee DynamicFini = getKmpcForDynamicFiniForType(IVTy, M, *this); 2468 Builder.CreateCall(DynamicFini, {SrcLoc, ThreadNum}); 2469 } 2470 2471 // Add the barrier if requested. 2472 if (NeedsBarrier) { 2473 Builder.SetInsertPoint(&Exit->back()); 2474 createBarrier(LocationDescription(Builder.saveIP(), DL), 2475 omp::Directive::OMPD_for, /* ForceSimpleCall */ false, 2476 /* CheckCancelFlag */ false); 2477 } 2478 2479 CLI->invalidate(); 2480 return AfterIP; 2481 } 2482 2483 /// Redirect all edges that branch to \p OldTarget to \p NewTarget. That is, 2484 /// after this \p OldTarget will be orphaned. 2485 static void redirectAllPredecessorsTo(BasicBlock *OldTarget, 2486 BasicBlock *NewTarget, DebugLoc DL) { 2487 for (BasicBlock *Pred : make_early_inc_range(predecessors(OldTarget))) 2488 redirectTo(Pred, NewTarget, DL); 2489 } 2490 2491 /// Determine which blocks in \p BBs are reachable from outside and remove the 2492 /// ones that are not reachable from the function. 2493 static void removeUnusedBlocksFromParent(ArrayRef<BasicBlock *> BBs) { 2494 SmallPtrSet<BasicBlock *, 6> BBsToErase{BBs.begin(), BBs.end()}; 2495 auto HasRemainingUses = [&BBsToErase](BasicBlock *BB) { 2496 for (Use &U : BB->uses()) { 2497 auto *UseInst = dyn_cast<Instruction>(U.getUser()); 2498 if (!UseInst) 2499 continue; 2500 if (BBsToErase.count(UseInst->getParent())) 2501 continue; 2502 return true; 2503 } 2504 return false; 2505 }; 2506 2507 while (true) { 2508 bool Changed = false; 2509 for (BasicBlock *BB : make_early_inc_range(BBsToErase)) { 2510 if (HasRemainingUses(BB)) { 2511 BBsToErase.erase(BB); 2512 Changed = true; 2513 } 2514 } 2515 if (!Changed) 2516 break; 2517 } 2518 2519 SmallVector<BasicBlock *, 7> BBVec(BBsToErase.begin(), BBsToErase.end()); 2520 DeleteDeadBlocks(BBVec); 2521 } 2522 2523 CanonicalLoopInfo * 2524 OpenMPIRBuilder::collapseLoops(DebugLoc DL, ArrayRef<CanonicalLoopInfo *> Loops, 2525 InsertPointTy ComputeIP) { 2526 assert(Loops.size() >= 1 && "At least one loop required"); 2527 size_t NumLoops = Loops.size(); 2528 2529 // Nothing to do if there is already just one loop. 2530 if (NumLoops == 1) 2531 return Loops.front(); 2532 2533 CanonicalLoopInfo *Outermost = Loops.front(); 2534 CanonicalLoopInfo *Innermost = Loops.back(); 2535 BasicBlock *OrigPreheader = Outermost->getPreheader(); 2536 BasicBlock *OrigAfter = Outermost->getAfter(); 2537 Function *F = OrigPreheader->getParent(); 2538 2539 // Loop control blocks that may become orphaned later. 2540 SmallVector<BasicBlock *, 12> OldControlBBs; 2541 OldControlBBs.reserve(6 * Loops.size()); 2542 for (CanonicalLoopInfo *Loop : Loops) 2543 Loop->collectControlBlocks(OldControlBBs); 2544 2545 // Setup the IRBuilder for inserting the trip count computation. 2546 Builder.SetCurrentDebugLocation(DL); 2547 if (ComputeIP.isSet()) 2548 Builder.restoreIP(ComputeIP); 2549 else 2550 Builder.restoreIP(Outermost->getPreheaderIP()); 2551 2552 // Derive the collapsed' loop trip count. 2553 // TODO: Find common/largest indvar type. 2554 Value *CollapsedTripCount = nullptr; 2555 for (CanonicalLoopInfo *L : Loops) { 2556 assert(L->isValid() && 2557 "All loops to collapse must be valid canonical loops"); 2558 Value *OrigTripCount = L->getTripCount(); 2559 if (!CollapsedTripCount) { 2560 CollapsedTripCount = OrigTripCount; 2561 continue; 2562 } 2563 2564 // TODO: Enable UndefinedSanitizer to diagnose an overflow here. 2565 CollapsedTripCount = Builder.CreateMul(CollapsedTripCount, OrigTripCount, 2566 {}, /*HasNUW=*/true); 2567 } 2568 2569 // Create the collapsed loop control flow. 2570 CanonicalLoopInfo *Result = 2571 createLoopSkeleton(DL, CollapsedTripCount, F, 2572 OrigPreheader->getNextNode(), OrigAfter, "collapsed"); 2573 2574 // Build the collapsed loop body code. 2575 // Start with deriving the input loop induction variables from the collapsed 2576 // one, using a divmod scheme. To preserve the original loops' order, the 2577 // innermost loop use the least significant bits. 2578 Builder.restoreIP(Result->getBodyIP()); 2579 2580 Value *Leftover = Result->getIndVar(); 2581 SmallVector<Value *> NewIndVars; 2582 NewIndVars.resize(NumLoops); 2583 for (int i = NumLoops - 1; i >= 1; --i) { 2584 Value *OrigTripCount = Loops[i]->getTripCount(); 2585 2586 Value *NewIndVar = Builder.CreateURem(Leftover, OrigTripCount); 2587 NewIndVars[i] = NewIndVar; 2588 2589 Leftover = Builder.CreateUDiv(Leftover, OrigTripCount); 2590 } 2591 // Outermost loop gets all the remaining bits. 2592 NewIndVars[0] = Leftover; 2593 2594 // Construct the loop body control flow. 2595 // We progressively construct the branch structure following in direction of 2596 // the control flow, from the leading in-between code, the loop nest body, the 2597 // trailing in-between code, and rejoining the collapsed loop's latch. 2598 // ContinueBlock and ContinuePred keep track of the source(s) of next edge. If 2599 // the ContinueBlock is set, continue with that block. If ContinuePred, use 2600 // its predecessors as sources. 2601 BasicBlock *ContinueBlock = Result->getBody(); 2602 BasicBlock *ContinuePred = nullptr; 2603 auto ContinueWith = [&ContinueBlock, &ContinuePred, DL](BasicBlock *Dest, 2604 BasicBlock *NextSrc) { 2605 if (ContinueBlock) 2606 redirectTo(ContinueBlock, Dest, DL); 2607 else 2608 redirectAllPredecessorsTo(ContinuePred, Dest, DL); 2609 2610 ContinueBlock = nullptr; 2611 ContinuePred = NextSrc; 2612 }; 2613 2614 // The code before the nested loop of each level. 2615 // Because we are sinking it into the nest, it will be executed more often 2616 // that the original loop. More sophisticated schemes could keep track of what 2617 // the in-between code is and instantiate it only once per thread. 2618 for (size_t i = 0; i < NumLoops - 1; ++i) 2619 ContinueWith(Loops[i]->getBody(), Loops[i + 1]->getHeader()); 2620 2621 // Connect the loop nest body. 2622 ContinueWith(Innermost->getBody(), Innermost->getLatch()); 2623 2624 // The code after the nested loop at each level. 2625 for (size_t i = NumLoops - 1; i > 0; --i) 2626 ContinueWith(Loops[i]->getAfter(), Loops[i - 1]->getLatch()); 2627 2628 // Connect the finished loop to the collapsed loop latch. 2629 ContinueWith(Result->getLatch(), nullptr); 2630 2631 // Replace the input loops with the new collapsed loop. 2632 redirectTo(Outermost->getPreheader(), Result->getPreheader(), DL); 2633 redirectTo(Result->getAfter(), Outermost->getAfter(), DL); 2634 2635 // Replace the input loop indvars with the derived ones. 2636 for (size_t i = 0; i < NumLoops; ++i) 2637 Loops[i]->getIndVar()->replaceAllUsesWith(NewIndVars[i]); 2638 2639 // Remove unused parts of the input loops. 2640 removeUnusedBlocksFromParent(OldControlBBs); 2641 2642 for (CanonicalLoopInfo *L : Loops) 2643 L->invalidate(); 2644 2645 #ifndef NDEBUG 2646 Result->assertOK(); 2647 #endif 2648 return Result; 2649 } 2650 2651 std::vector<CanonicalLoopInfo *> 2652 OpenMPIRBuilder::tileLoops(DebugLoc DL, ArrayRef<CanonicalLoopInfo *> Loops, 2653 ArrayRef<Value *> TileSizes) { 2654 assert(TileSizes.size() == Loops.size() && 2655 "Must pass as many tile sizes as there are loops"); 2656 int NumLoops = Loops.size(); 2657 assert(NumLoops >= 1 && "At least one loop to tile required"); 2658 2659 CanonicalLoopInfo *OutermostLoop = Loops.front(); 2660 CanonicalLoopInfo *InnermostLoop = Loops.back(); 2661 Function *F = OutermostLoop->getBody()->getParent(); 2662 BasicBlock *InnerEnter = InnermostLoop->getBody(); 2663 BasicBlock *InnerLatch = InnermostLoop->getLatch(); 2664 2665 // Loop control blocks that may become orphaned later. 2666 SmallVector<BasicBlock *, 12> OldControlBBs; 2667 OldControlBBs.reserve(6 * Loops.size()); 2668 for (CanonicalLoopInfo *Loop : Loops) 2669 Loop->collectControlBlocks(OldControlBBs); 2670 2671 // Collect original trip counts and induction variable to be accessible by 2672 // index. Also, the structure of the original loops is not preserved during 2673 // the construction of the tiled loops, so do it before we scavenge the BBs of 2674 // any original CanonicalLoopInfo. 2675 SmallVector<Value *, 4> OrigTripCounts, OrigIndVars; 2676 for (CanonicalLoopInfo *L : Loops) { 2677 assert(L->isValid() && "All input loops must be valid canonical loops"); 2678 OrigTripCounts.push_back(L->getTripCount()); 2679 OrigIndVars.push_back(L->getIndVar()); 2680 } 2681 2682 // Collect the code between loop headers. These may contain SSA definitions 2683 // that are used in the loop nest body. To be usable with in the innermost 2684 // body, these BasicBlocks will be sunk into the loop nest body. That is, 2685 // these instructions may be executed more often than before the tiling. 2686 // TODO: It would be sufficient to only sink them into body of the 2687 // corresponding tile loop. 2688 SmallVector<std::pair<BasicBlock *, BasicBlock *>, 4> InbetweenCode; 2689 for (int i = 0; i < NumLoops - 1; ++i) { 2690 CanonicalLoopInfo *Surrounding = Loops[i]; 2691 CanonicalLoopInfo *Nested = Loops[i + 1]; 2692 2693 BasicBlock *EnterBB = Surrounding->getBody(); 2694 BasicBlock *ExitBB = Nested->getHeader(); 2695 InbetweenCode.emplace_back(EnterBB, ExitBB); 2696 } 2697 2698 // Compute the trip counts of the floor loops. 2699 Builder.SetCurrentDebugLocation(DL); 2700 Builder.restoreIP(OutermostLoop->getPreheaderIP()); 2701 SmallVector<Value *, 4> FloorCount, FloorRems; 2702 for (int i = 0; i < NumLoops; ++i) { 2703 Value *TileSize = TileSizes[i]; 2704 Value *OrigTripCount = OrigTripCounts[i]; 2705 Type *IVType = OrigTripCount->getType(); 2706 2707 Value *FloorTripCount = Builder.CreateUDiv(OrigTripCount, TileSize); 2708 Value *FloorTripRem = Builder.CreateURem(OrigTripCount, TileSize); 2709 2710 // 0 if tripcount divides the tilesize, 1 otherwise. 2711 // 1 means we need an additional iteration for a partial tile. 2712 // 2713 // Unfortunately we cannot just use the roundup-formula 2714 // (tripcount + tilesize - 1)/tilesize 2715 // because the summation might overflow. We do not want introduce undefined 2716 // behavior when the untiled loop nest did not. 2717 Value *FloorTripOverflow = 2718 Builder.CreateICmpNE(FloorTripRem, ConstantInt::get(IVType, 0)); 2719 2720 FloorTripOverflow = Builder.CreateZExt(FloorTripOverflow, IVType); 2721 FloorTripCount = 2722 Builder.CreateAdd(FloorTripCount, FloorTripOverflow, 2723 "omp_floor" + Twine(i) + ".tripcount", true); 2724 2725 // Remember some values for later use. 2726 FloorCount.push_back(FloorTripCount); 2727 FloorRems.push_back(FloorTripRem); 2728 } 2729 2730 // Generate the new loop nest, from the outermost to the innermost. 2731 std::vector<CanonicalLoopInfo *> Result; 2732 Result.reserve(NumLoops * 2); 2733 2734 // The basic block of the surrounding loop that enters the nest generated 2735 // loop. 2736 BasicBlock *Enter = OutermostLoop->getPreheader(); 2737 2738 // The basic block of the surrounding loop where the inner code should 2739 // continue. 2740 BasicBlock *Continue = OutermostLoop->getAfter(); 2741 2742 // Where the next loop basic block should be inserted. 2743 BasicBlock *OutroInsertBefore = InnermostLoop->getExit(); 2744 2745 auto EmbeddNewLoop = 2746 [this, DL, F, InnerEnter, &Enter, &Continue, &OutroInsertBefore]( 2747 Value *TripCount, const Twine &Name) -> CanonicalLoopInfo * { 2748 CanonicalLoopInfo *EmbeddedLoop = createLoopSkeleton( 2749 DL, TripCount, F, InnerEnter, OutroInsertBefore, Name); 2750 redirectTo(Enter, EmbeddedLoop->getPreheader(), DL); 2751 redirectTo(EmbeddedLoop->getAfter(), Continue, DL); 2752 2753 // Setup the position where the next embedded loop connects to this loop. 2754 Enter = EmbeddedLoop->getBody(); 2755 Continue = EmbeddedLoop->getLatch(); 2756 OutroInsertBefore = EmbeddedLoop->getLatch(); 2757 return EmbeddedLoop; 2758 }; 2759 2760 auto EmbeddNewLoops = [&Result, &EmbeddNewLoop](ArrayRef<Value *> TripCounts, 2761 const Twine &NameBase) { 2762 for (auto P : enumerate(TripCounts)) { 2763 CanonicalLoopInfo *EmbeddedLoop = 2764 EmbeddNewLoop(P.value(), NameBase + Twine(P.index())); 2765 Result.push_back(EmbeddedLoop); 2766 } 2767 }; 2768 2769 EmbeddNewLoops(FloorCount, "floor"); 2770 2771 // Within the innermost floor loop, emit the code that computes the tile 2772 // sizes. 2773 Builder.SetInsertPoint(Enter->getTerminator()); 2774 SmallVector<Value *, 4> TileCounts; 2775 for (int i = 0; i < NumLoops; ++i) { 2776 CanonicalLoopInfo *FloorLoop = Result[i]; 2777 Value *TileSize = TileSizes[i]; 2778 2779 Value *FloorIsEpilogue = 2780 Builder.CreateICmpEQ(FloorLoop->getIndVar(), FloorCount[i]); 2781 Value *TileTripCount = 2782 Builder.CreateSelect(FloorIsEpilogue, FloorRems[i], TileSize); 2783 2784 TileCounts.push_back(TileTripCount); 2785 } 2786 2787 // Create the tile loops. 2788 EmbeddNewLoops(TileCounts, "tile"); 2789 2790 // Insert the inbetween code into the body. 2791 BasicBlock *BodyEnter = Enter; 2792 BasicBlock *BodyEntered = nullptr; 2793 for (std::pair<BasicBlock *, BasicBlock *> P : InbetweenCode) { 2794 BasicBlock *EnterBB = P.first; 2795 BasicBlock *ExitBB = P.second; 2796 2797 if (BodyEnter) 2798 redirectTo(BodyEnter, EnterBB, DL); 2799 else 2800 redirectAllPredecessorsTo(BodyEntered, EnterBB, DL); 2801 2802 BodyEnter = nullptr; 2803 BodyEntered = ExitBB; 2804 } 2805 2806 // Append the original loop nest body into the generated loop nest body. 2807 if (BodyEnter) 2808 redirectTo(BodyEnter, InnerEnter, DL); 2809 else 2810 redirectAllPredecessorsTo(BodyEntered, InnerEnter, DL); 2811 redirectAllPredecessorsTo(InnerLatch, Continue, DL); 2812 2813 // Replace the original induction variable with an induction variable computed 2814 // from the tile and floor induction variables. 2815 Builder.restoreIP(Result.back()->getBodyIP()); 2816 for (int i = 0; i < NumLoops; ++i) { 2817 CanonicalLoopInfo *FloorLoop = Result[i]; 2818 CanonicalLoopInfo *TileLoop = Result[NumLoops + i]; 2819 Value *OrigIndVar = OrigIndVars[i]; 2820 Value *Size = TileSizes[i]; 2821 2822 Value *Scale = 2823 Builder.CreateMul(Size, FloorLoop->getIndVar(), {}, /*HasNUW=*/true); 2824 Value *Shift = 2825 Builder.CreateAdd(Scale, TileLoop->getIndVar(), {}, /*HasNUW=*/true); 2826 OrigIndVar->replaceAllUsesWith(Shift); 2827 } 2828 2829 // Remove unused parts of the original loops. 2830 removeUnusedBlocksFromParent(OldControlBBs); 2831 2832 for (CanonicalLoopInfo *L : Loops) 2833 L->invalidate(); 2834 2835 #ifndef NDEBUG 2836 for (CanonicalLoopInfo *GenL : Result) 2837 GenL->assertOK(); 2838 #endif 2839 return Result; 2840 } 2841 2842 /// Attach loop metadata \p Properties to the loop described by \p Loop. If the 2843 /// loop already has metadata, the loop properties are appended. 2844 static void addLoopMetadata(CanonicalLoopInfo *Loop, 2845 ArrayRef<Metadata *> Properties) { 2846 assert(Loop->isValid() && "Expecting a valid CanonicalLoopInfo"); 2847 2848 // Nothing to do if no property to attach. 2849 if (Properties.empty()) 2850 return; 2851 2852 LLVMContext &Ctx = Loop->getFunction()->getContext(); 2853 SmallVector<Metadata *> NewLoopProperties; 2854 NewLoopProperties.push_back(nullptr); 2855 2856 // If the loop already has metadata, prepend it to the new metadata. 2857 BasicBlock *Latch = Loop->getLatch(); 2858 assert(Latch && "A valid CanonicalLoopInfo must have a unique latch"); 2859 MDNode *Existing = Latch->getTerminator()->getMetadata(LLVMContext::MD_loop); 2860 if (Existing) 2861 append_range(NewLoopProperties, drop_begin(Existing->operands(), 1)); 2862 2863 append_range(NewLoopProperties, Properties); 2864 MDNode *LoopID = MDNode::getDistinct(Ctx, NewLoopProperties); 2865 LoopID->replaceOperandWith(0, LoopID); 2866 2867 Latch->getTerminator()->setMetadata(LLVMContext::MD_loop, LoopID); 2868 } 2869 2870 /// Attach llvm.access.group metadata to the memref instructions of \p Block 2871 static void addSimdMetadata(BasicBlock *Block, MDNode *AccessGroup, 2872 LoopInfo &LI) { 2873 for (Instruction &I : *Block) { 2874 if (I.mayReadOrWriteMemory()) { 2875 // TODO: This instruction may already have access group from 2876 // other pragmas e.g. #pragma clang loop vectorize. Append 2877 // so that the existing metadata is not overwritten. 2878 I.setMetadata(LLVMContext::MD_access_group, AccessGroup); 2879 } 2880 } 2881 } 2882 2883 void OpenMPIRBuilder::unrollLoopFull(DebugLoc, CanonicalLoopInfo *Loop) { 2884 LLVMContext &Ctx = Builder.getContext(); 2885 addLoopMetadata( 2886 Loop, {MDNode::get(Ctx, MDString::get(Ctx, "llvm.loop.unroll.enable")), 2887 MDNode::get(Ctx, MDString::get(Ctx, "llvm.loop.unroll.full"))}); 2888 } 2889 2890 void OpenMPIRBuilder::unrollLoopHeuristic(DebugLoc, CanonicalLoopInfo *Loop) { 2891 LLVMContext &Ctx = Builder.getContext(); 2892 addLoopMetadata( 2893 Loop, { 2894 MDNode::get(Ctx, MDString::get(Ctx, "llvm.loop.unroll.enable")), 2895 }); 2896 } 2897 2898 void OpenMPIRBuilder::applySimd(CanonicalLoopInfo *CanonicalLoop, 2899 ConstantInt *Simdlen) { 2900 LLVMContext &Ctx = Builder.getContext(); 2901 2902 Function *F = CanonicalLoop->getFunction(); 2903 2904 FunctionAnalysisManager FAM; 2905 FAM.registerPass([]() { return DominatorTreeAnalysis(); }); 2906 FAM.registerPass([]() { return LoopAnalysis(); }); 2907 FAM.registerPass([]() { return PassInstrumentationAnalysis(); }); 2908 2909 LoopAnalysis LIA; 2910 LoopInfo &&LI = LIA.run(*F, FAM); 2911 2912 Loop *L = LI.getLoopFor(CanonicalLoop->getHeader()); 2913 2914 SmallSet<BasicBlock *, 8> Reachable; 2915 2916 // Get the basic blocks from the loop in which memref instructions 2917 // can be found. 2918 // TODO: Generalize getting all blocks inside a CanonicalizeLoopInfo, 2919 // preferably without running any passes. 2920 for (BasicBlock *Block : L->getBlocks()) { 2921 if (Block == CanonicalLoop->getCond() || 2922 Block == CanonicalLoop->getHeader()) 2923 continue; 2924 Reachable.insert(Block); 2925 } 2926 2927 // Add access group metadata to memory-access instructions. 2928 MDNode *AccessGroup = MDNode::getDistinct(Ctx, {}); 2929 for (BasicBlock *BB : Reachable) 2930 addSimdMetadata(BB, AccessGroup, LI); 2931 2932 // Use the above access group metadata to create loop level 2933 // metadata, which should be distinct for each loop. 2934 ConstantAsMetadata *BoolConst = 2935 ConstantAsMetadata::get(ConstantInt::getTrue(Type::getInt1Ty(Ctx))); 2936 // TODO: If the loop has existing parallel access metadata, have 2937 // to combine two lists. 2938 addLoopMetadata( 2939 CanonicalLoop, 2940 {MDNode::get(Ctx, {MDString::get(Ctx, "llvm.loop.parallel_accesses"), 2941 AccessGroup}), 2942 MDNode::get(Ctx, {MDString::get(Ctx, "llvm.loop.vectorize.enable"), 2943 BoolConst})}); 2944 if (Simdlen != nullptr) 2945 addLoopMetadata( 2946 CanonicalLoop, 2947 MDNode::get(Ctx, {MDString::get(Ctx, "llvm.loop.vectorize.width"), 2948 ConstantAsMetadata::get(Simdlen)})); 2949 } 2950 2951 /// Create the TargetMachine object to query the backend for optimization 2952 /// preferences. 2953 /// 2954 /// Ideally, this would be passed from the front-end to the OpenMPBuilder, but 2955 /// e.g. Clang does not pass it to its CodeGen layer and creates it only when 2956 /// needed for the LLVM pass pipline. We use some default options to avoid 2957 /// having to pass too many settings from the frontend that probably do not 2958 /// matter. 2959 /// 2960 /// Currently, TargetMachine is only used sometimes by the unrollLoopPartial 2961 /// method. If we are going to use TargetMachine for more purposes, especially 2962 /// those that are sensitive to TargetOptions, RelocModel and CodeModel, it 2963 /// might become be worth requiring front-ends to pass on their TargetMachine, 2964 /// or at least cache it between methods. Note that while fontends such as Clang 2965 /// have just a single main TargetMachine per translation unit, "target-cpu" and 2966 /// "target-features" that determine the TargetMachine are per-function and can 2967 /// be overrided using __attribute__((target("OPTIONS"))). 2968 static std::unique_ptr<TargetMachine> 2969 createTargetMachine(Function *F, CodeGenOpt::Level OptLevel) { 2970 Module *M = F->getParent(); 2971 2972 StringRef CPU = F->getFnAttribute("target-cpu").getValueAsString(); 2973 StringRef Features = F->getFnAttribute("target-features").getValueAsString(); 2974 const std::string &Triple = M->getTargetTriple(); 2975 2976 std::string Error; 2977 const llvm::Target *TheTarget = TargetRegistry::lookupTarget(Triple, Error); 2978 if (!TheTarget) 2979 return {}; 2980 2981 llvm::TargetOptions Options; 2982 return std::unique_ptr<TargetMachine>(TheTarget->createTargetMachine( 2983 Triple, CPU, Features, Options, /*RelocModel=*/None, /*CodeModel=*/None, 2984 OptLevel)); 2985 } 2986 2987 /// Heuristically determine the best-performant unroll factor for \p CLI. This 2988 /// depends on the target processor. We are re-using the same heuristics as the 2989 /// LoopUnrollPass. 2990 static int32_t computeHeuristicUnrollFactor(CanonicalLoopInfo *CLI) { 2991 Function *F = CLI->getFunction(); 2992 2993 // Assume the user requests the most aggressive unrolling, even if the rest of 2994 // the code is optimized using a lower setting. 2995 CodeGenOpt::Level OptLevel = CodeGenOpt::Aggressive; 2996 std::unique_ptr<TargetMachine> TM = createTargetMachine(F, OptLevel); 2997 2998 FunctionAnalysisManager FAM; 2999 FAM.registerPass([]() { return TargetLibraryAnalysis(); }); 3000 FAM.registerPass([]() { return AssumptionAnalysis(); }); 3001 FAM.registerPass([]() { return DominatorTreeAnalysis(); }); 3002 FAM.registerPass([]() { return LoopAnalysis(); }); 3003 FAM.registerPass([]() { return ScalarEvolutionAnalysis(); }); 3004 FAM.registerPass([]() { return PassInstrumentationAnalysis(); }); 3005 TargetIRAnalysis TIRA; 3006 if (TM) 3007 TIRA = TargetIRAnalysis( 3008 [&](const Function &F) { return TM->getTargetTransformInfo(F); }); 3009 FAM.registerPass([&]() { return TIRA; }); 3010 3011 TargetIRAnalysis::Result &&TTI = TIRA.run(*F, FAM); 3012 ScalarEvolutionAnalysis SEA; 3013 ScalarEvolution &&SE = SEA.run(*F, FAM); 3014 DominatorTreeAnalysis DTA; 3015 DominatorTree &&DT = DTA.run(*F, FAM); 3016 LoopAnalysis LIA; 3017 LoopInfo &&LI = LIA.run(*F, FAM); 3018 AssumptionAnalysis ACT; 3019 AssumptionCache &&AC = ACT.run(*F, FAM); 3020 OptimizationRemarkEmitter ORE{F}; 3021 3022 Loop *L = LI.getLoopFor(CLI->getHeader()); 3023 assert(L && "Expecting CanonicalLoopInfo to be recognized as a loop"); 3024 3025 TargetTransformInfo::UnrollingPreferences UP = 3026 gatherUnrollingPreferences(L, SE, TTI, 3027 /*BlockFrequencyInfo=*/nullptr, 3028 /*ProfileSummaryInfo=*/nullptr, ORE, OptLevel, 3029 /*UserThreshold=*/None, 3030 /*UserCount=*/None, 3031 /*UserAllowPartial=*/true, 3032 /*UserAllowRuntime=*/true, 3033 /*UserUpperBound=*/None, 3034 /*UserFullUnrollMaxCount=*/None); 3035 3036 UP.Force = true; 3037 3038 // Account for additional optimizations taking place before the LoopUnrollPass 3039 // would unroll the loop. 3040 UP.Threshold *= UnrollThresholdFactor; 3041 UP.PartialThreshold *= UnrollThresholdFactor; 3042 3043 // Use normal unroll factors even if the rest of the code is optimized for 3044 // size. 3045 UP.OptSizeThreshold = UP.Threshold; 3046 UP.PartialOptSizeThreshold = UP.PartialThreshold; 3047 3048 LLVM_DEBUG(dbgs() << "Unroll heuristic thresholds:\n" 3049 << " Threshold=" << UP.Threshold << "\n" 3050 << " PartialThreshold=" << UP.PartialThreshold << "\n" 3051 << " OptSizeThreshold=" << UP.OptSizeThreshold << "\n" 3052 << " PartialOptSizeThreshold=" 3053 << UP.PartialOptSizeThreshold << "\n"); 3054 3055 // Disable peeling. 3056 TargetTransformInfo::PeelingPreferences PP = 3057 gatherPeelingPreferences(L, SE, TTI, 3058 /*UserAllowPeeling=*/false, 3059 /*UserAllowProfileBasedPeeling=*/false, 3060 /*UnrollingSpecficValues=*/false); 3061 3062 SmallPtrSet<const Value *, 32> EphValues; 3063 CodeMetrics::collectEphemeralValues(L, &AC, EphValues); 3064 3065 // Assume that reads and writes to stack variables can be eliminated by 3066 // Mem2Reg, SROA or LICM. That is, don't count them towards the loop body's 3067 // size. 3068 for (BasicBlock *BB : L->blocks()) { 3069 for (Instruction &I : *BB) { 3070 Value *Ptr; 3071 if (auto *Load = dyn_cast<LoadInst>(&I)) { 3072 Ptr = Load->getPointerOperand(); 3073 } else if (auto *Store = dyn_cast<StoreInst>(&I)) { 3074 Ptr = Store->getPointerOperand(); 3075 } else 3076 continue; 3077 3078 Ptr = Ptr->stripPointerCasts(); 3079 3080 if (auto *Alloca = dyn_cast<AllocaInst>(Ptr)) { 3081 if (Alloca->getParent() == &F->getEntryBlock()) 3082 EphValues.insert(&I); 3083 } 3084 } 3085 } 3086 3087 unsigned NumInlineCandidates; 3088 bool NotDuplicatable; 3089 bool Convergent; 3090 InstructionCost LoopSizeIC = 3091 ApproximateLoopSize(L, NumInlineCandidates, NotDuplicatable, Convergent, 3092 TTI, EphValues, UP.BEInsns); 3093 LLVM_DEBUG(dbgs() << "Estimated loop size is " << LoopSizeIC << "\n"); 3094 3095 // Loop is not unrollable if the loop contains certain instructions. 3096 if (NotDuplicatable || Convergent || !LoopSizeIC.isValid()) { 3097 LLVM_DEBUG(dbgs() << "Loop not considered unrollable\n"); 3098 return 1; 3099 } 3100 unsigned LoopSize = *LoopSizeIC.getValue(); 3101 3102 // TODO: Determine trip count of \p CLI if constant, computeUnrollCount might 3103 // be able to use it. 3104 int TripCount = 0; 3105 int MaxTripCount = 0; 3106 bool MaxOrZero = false; 3107 unsigned TripMultiple = 0; 3108 3109 bool UseUpperBound = false; 3110 computeUnrollCount(L, TTI, DT, &LI, SE, EphValues, &ORE, TripCount, 3111 MaxTripCount, MaxOrZero, TripMultiple, LoopSize, UP, PP, 3112 UseUpperBound); 3113 unsigned Factor = UP.Count; 3114 LLVM_DEBUG(dbgs() << "Suggesting unroll factor of " << Factor << "\n"); 3115 3116 // This function returns 1 to signal to not unroll a loop. 3117 if (Factor == 0) 3118 return 1; 3119 return Factor; 3120 } 3121 3122 void OpenMPIRBuilder::unrollLoopPartial(DebugLoc DL, CanonicalLoopInfo *Loop, 3123 int32_t Factor, 3124 CanonicalLoopInfo **UnrolledCLI) { 3125 assert(Factor >= 0 && "Unroll factor must not be negative"); 3126 3127 Function *F = Loop->getFunction(); 3128 LLVMContext &Ctx = F->getContext(); 3129 3130 // If the unrolled loop is not used for another loop-associated directive, it 3131 // is sufficient to add metadata for the LoopUnrollPass. 3132 if (!UnrolledCLI) { 3133 SmallVector<Metadata *, 2> LoopMetadata; 3134 LoopMetadata.push_back( 3135 MDNode::get(Ctx, MDString::get(Ctx, "llvm.loop.unroll.enable"))); 3136 3137 if (Factor >= 1) { 3138 ConstantAsMetadata *FactorConst = ConstantAsMetadata::get( 3139 ConstantInt::get(Type::getInt32Ty(Ctx), APInt(32, Factor))); 3140 LoopMetadata.push_back(MDNode::get( 3141 Ctx, {MDString::get(Ctx, "llvm.loop.unroll.count"), FactorConst})); 3142 } 3143 3144 addLoopMetadata(Loop, LoopMetadata); 3145 return; 3146 } 3147 3148 // Heuristically determine the unroll factor. 3149 if (Factor == 0) 3150 Factor = computeHeuristicUnrollFactor(Loop); 3151 3152 // No change required with unroll factor 1. 3153 if (Factor == 1) { 3154 *UnrolledCLI = Loop; 3155 return; 3156 } 3157 3158 assert(Factor >= 2 && 3159 "unrolling only makes sense with a factor of 2 or larger"); 3160 3161 Type *IndVarTy = Loop->getIndVarType(); 3162 3163 // Apply partial unrolling by tiling the loop by the unroll-factor, then fully 3164 // unroll the inner loop. 3165 Value *FactorVal = 3166 ConstantInt::get(IndVarTy, APInt(IndVarTy->getIntegerBitWidth(), Factor, 3167 /*isSigned=*/false)); 3168 std::vector<CanonicalLoopInfo *> LoopNest = 3169 tileLoops(DL, {Loop}, {FactorVal}); 3170 assert(LoopNest.size() == 2 && "Expect 2 loops after tiling"); 3171 *UnrolledCLI = LoopNest[0]; 3172 CanonicalLoopInfo *InnerLoop = LoopNest[1]; 3173 3174 // LoopUnrollPass can only fully unroll loops with constant trip count. 3175 // Unroll by the unroll factor with a fallback epilog for the remainder 3176 // iterations if necessary. 3177 ConstantAsMetadata *FactorConst = ConstantAsMetadata::get( 3178 ConstantInt::get(Type::getInt32Ty(Ctx), APInt(32, Factor))); 3179 addLoopMetadata( 3180 InnerLoop, 3181 {MDNode::get(Ctx, MDString::get(Ctx, "llvm.loop.unroll.enable")), 3182 MDNode::get( 3183 Ctx, {MDString::get(Ctx, "llvm.loop.unroll.count"), FactorConst})}); 3184 3185 #ifndef NDEBUG 3186 (*UnrolledCLI)->assertOK(); 3187 #endif 3188 } 3189 3190 OpenMPIRBuilder::InsertPointTy 3191 OpenMPIRBuilder::createCopyPrivate(const LocationDescription &Loc, 3192 llvm::Value *BufSize, llvm::Value *CpyBuf, 3193 llvm::Value *CpyFn, llvm::Value *DidIt) { 3194 if (!updateToLocation(Loc)) 3195 return Loc.IP; 3196 3197 uint32_t SrcLocStrSize; 3198 Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize); 3199 Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize); 3200 Value *ThreadId = getOrCreateThreadID(Ident); 3201 3202 llvm::Value *DidItLD = Builder.CreateLoad(Builder.getInt32Ty(), DidIt); 3203 3204 Value *Args[] = {Ident, ThreadId, BufSize, CpyBuf, CpyFn, DidItLD}; 3205 3206 Function *Fn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_copyprivate); 3207 Builder.CreateCall(Fn, Args); 3208 3209 return Builder.saveIP(); 3210 } 3211 3212 OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::createSingle( 3213 const LocationDescription &Loc, BodyGenCallbackTy BodyGenCB, 3214 FinalizeCallbackTy FiniCB, bool IsNowait, llvm::Value *DidIt) { 3215 3216 if (!updateToLocation(Loc)) 3217 return Loc.IP; 3218 3219 // If needed (i.e. not null), initialize `DidIt` with 0 3220 if (DidIt) { 3221 Builder.CreateStore(Builder.getInt32(0), DidIt); 3222 } 3223 3224 Directive OMPD = Directive::OMPD_single; 3225 uint32_t SrcLocStrSize; 3226 Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize); 3227 Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize); 3228 Value *ThreadId = getOrCreateThreadID(Ident); 3229 Value *Args[] = {Ident, ThreadId}; 3230 3231 Function *EntryRTLFn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_single); 3232 Instruction *EntryCall = Builder.CreateCall(EntryRTLFn, Args); 3233 3234 Function *ExitRTLFn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_end_single); 3235 Instruction *ExitCall = Builder.CreateCall(ExitRTLFn, Args); 3236 3237 // generates the following: 3238 // if (__kmpc_single()) { 3239 // .... single region ... 3240 // __kmpc_end_single 3241 // } 3242 // __kmpc_barrier 3243 3244 EmitOMPInlinedRegion(OMPD, EntryCall, ExitCall, BodyGenCB, FiniCB, 3245 /*Conditional*/ true, 3246 /*hasFinalize*/ true); 3247 if (!IsNowait) 3248 createBarrier(LocationDescription(Builder.saveIP(), Loc.DL), 3249 omp::Directive::OMPD_unknown, /* ForceSimpleCall */ false, 3250 /* CheckCancelFlag */ false); 3251 return Builder.saveIP(); 3252 } 3253 3254 OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::createCritical( 3255 const LocationDescription &Loc, BodyGenCallbackTy BodyGenCB, 3256 FinalizeCallbackTy FiniCB, StringRef CriticalName, Value *HintInst) { 3257 3258 if (!updateToLocation(Loc)) 3259 return Loc.IP; 3260 3261 Directive OMPD = Directive::OMPD_critical; 3262 uint32_t SrcLocStrSize; 3263 Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize); 3264 Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize); 3265 Value *ThreadId = getOrCreateThreadID(Ident); 3266 Value *LockVar = getOMPCriticalRegionLock(CriticalName); 3267 Value *Args[] = {Ident, ThreadId, LockVar}; 3268 3269 SmallVector<llvm::Value *, 4> EnterArgs(std::begin(Args), std::end(Args)); 3270 Function *RTFn = nullptr; 3271 if (HintInst) { 3272 // Add Hint to entry Args and create call 3273 EnterArgs.push_back(HintInst); 3274 RTFn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_critical_with_hint); 3275 } else { 3276 RTFn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_critical); 3277 } 3278 Instruction *EntryCall = Builder.CreateCall(RTFn, EnterArgs); 3279 3280 Function *ExitRTLFn = 3281 getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_end_critical); 3282 Instruction *ExitCall = Builder.CreateCall(ExitRTLFn, Args); 3283 3284 return EmitOMPInlinedRegion(OMPD, EntryCall, ExitCall, BodyGenCB, FiniCB, 3285 /*Conditional*/ false, /*hasFinalize*/ true); 3286 } 3287 3288 OpenMPIRBuilder::InsertPointTy 3289 OpenMPIRBuilder::createOrderedDepend(const LocationDescription &Loc, 3290 InsertPointTy AllocaIP, unsigned NumLoops, 3291 ArrayRef<llvm::Value *> StoreValues, 3292 const Twine &Name, bool IsDependSource) { 3293 for (size_t I = 0; I < StoreValues.size(); I++) 3294 assert(StoreValues[I]->getType()->isIntegerTy(64) && 3295 "OpenMP runtime requires depend vec with i64 type"); 3296 3297 if (!updateToLocation(Loc)) 3298 return Loc.IP; 3299 3300 // Allocate space for vector and generate alloc instruction. 3301 auto *ArrI64Ty = ArrayType::get(Int64, NumLoops); 3302 Builder.restoreIP(AllocaIP); 3303 AllocaInst *ArgsBase = Builder.CreateAlloca(ArrI64Ty, nullptr, Name); 3304 ArgsBase->setAlignment(Align(8)); 3305 Builder.restoreIP(Loc.IP); 3306 3307 // Store the index value with offset in depend vector. 3308 for (unsigned I = 0; I < NumLoops; ++I) { 3309 Value *DependAddrGEPIter = Builder.CreateInBoundsGEP( 3310 ArrI64Ty, ArgsBase, {Builder.getInt64(0), Builder.getInt64(I)}); 3311 StoreInst *STInst = Builder.CreateStore(StoreValues[I], DependAddrGEPIter); 3312 STInst->setAlignment(Align(8)); 3313 } 3314 3315 Value *DependBaseAddrGEP = Builder.CreateInBoundsGEP( 3316 ArrI64Ty, ArgsBase, {Builder.getInt64(0), Builder.getInt64(0)}); 3317 3318 uint32_t SrcLocStrSize; 3319 Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize); 3320 Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize); 3321 Value *ThreadId = getOrCreateThreadID(Ident); 3322 Value *Args[] = {Ident, ThreadId, DependBaseAddrGEP}; 3323 3324 Function *RTLFn = nullptr; 3325 if (IsDependSource) 3326 RTLFn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_doacross_post); 3327 else 3328 RTLFn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_doacross_wait); 3329 Builder.CreateCall(RTLFn, Args); 3330 3331 return Builder.saveIP(); 3332 } 3333 3334 OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::createOrderedThreadsSimd( 3335 const LocationDescription &Loc, BodyGenCallbackTy BodyGenCB, 3336 FinalizeCallbackTy FiniCB, bool IsThreads) { 3337 if (!updateToLocation(Loc)) 3338 return Loc.IP; 3339 3340 Directive OMPD = Directive::OMPD_ordered; 3341 Instruction *EntryCall = nullptr; 3342 Instruction *ExitCall = nullptr; 3343 3344 if (IsThreads) { 3345 uint32_t SrcLocStrSize; 3346 Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize); 3347 Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize); 3348 Value *ThreadId = getOrCreateThreadID(Ident); 3349 Value *Args[] = {Ident, ThreadId}; 3350 3351 Function *EntryRTLFn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_ordered); 3352 EntryCall = Builder.CreateCall(EntryRTLFn, Args); 3353 3354 Function *ExitRTLFn = 3355 getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_end_ordered); 3356 ExitCall = Builder.CreateCall(ExitRTLFn, Args); 3357 } 3358 3359 return EmitOMPInlinedRegion(OMPD, EntryCall, ExitCall, BodyGenCB, FiniCB, 3360 /*Conditional*/ false, /*hasFinalize*/ true); 3361 } 3362 3363 OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::EmitOMPInlinedRegion( 3364 Directive OMPD, Instruction *EntryCall, Instruction *ExitCall, 3365 BodyGenCallbackTy BodyGenCB, FinalizeCallbackTy FiniCB, bool Conditional, 3366 bool HasFinalize, bool IsCancellable) { 3367 3368 if (HasFinalize) 3369 FinalizationStack.push_back({FiniCB, OMPD, IsCancellable}); 3370 3371 // Create inlined region's entry and body blocks, in preparation 3372 // for conditional creation 3373 BasicBlock *EntryBB = Builder.GetInsertBlock(); 3374 Instruction *SplitPos = EntryBB->getTerminator(); 3375 if (!isa_and_nonnull<BranchInst>(SplitPos)) 3376 SplitPos = new UnreachableInst(Builder.getContext(), EntryBB); 3377 BasicBlock *ExitBB = EntryBB->splitBasicBlock(SplitPos, "omp_region.end"); 3378 BasicBlock *FiniBB = 3379 EntryBB->splitBasicBlock(EntryBB->getTerminator(), "omp_region.finalize"); 3380 3381 Builder.SetInsertPoint(EntryBB->getTerminator()); 3382 emitCommonDirectiveEntry(OMPD, EntryCall, ExitBB, Conditional); 3383 3384 // generate body 3385 BodyGenCB(/* AllocaIP */ InsertPointTy(), 3386 /* CodeGenIP */ Builder.saveIP()); 3387 3388 // emit exit call and do any needed finalization. 3389 auto FinIP = InsertPointTy(FiniBB, FiniBB->getFirstInsertionPt()); 3390 assert(FiniBB->getTerminator()->getNumSuccessors() == 1 && 3391 FiniBB->getTerminator()->getSuccessor(0) == ExitBB && 3392 "Unexpected control flow graph state!!"); 3393 emitCommonDirectiveExit(OMPD, FinIP, ExitCall, HasFinalize); 3394 assert(FiniBB->getUniquePredecessor()->getUniqueSuccessor() == FiniBB && 3395 "Unexpected Control Flow State!"); 3396 MergeBlockIntoPredecessor(FiniBB); 3397 3398 // If we are skipping the region of a non conditional, remove the exit 3399 // block, and clear the builder's insertion point. 3400 assert(SplitPos->getParent() == ExitBB && 3401 "Unexpected Insertion point location!"); 3402 auto merged = MergeBlockIntoPredecessor(ExitBB); 3403 BasicBlock *ExitPredBB = SplitPos->getParent(); 3404 auto InsertBB = merged ? ExitPredBB : ExitBB; 3405 if (!isa_and_nonnull<BranchInst>(SplitPos)) 3406 SplitPos->eraseFromParent(); 3407 Builder.SetInsertPoint(InsertBB); 3408 3409 return Builder.saveIP(); 3410 } 3411 3412 OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::emitCommonDirectiveEntry( 3413 Directive OMPD, Value *EntryCall, BasicBlock *ExitBB, bool Conditional) { 3414 // if nothing to do, Return current insertion point. 3415 if (!Conditional || !EntryCall) 3416 return Builder.saveIP(); 3417 3418 BasicBlock *EntryBB = Builder.GetInsertBlock(); 3419 Value *CallBool = Builder.CreateIsNotNull(EntryCall); 3420 auto *ThenBB = BasicBlock::Create(M.getContext(), "omp_region.body"); 3421 auto *UI = new UnreachableInst(Builder.getContext(), ThenBB); 3422 3423 // Emit thenBB and set the Builder's insertion point there for 3424 // body generation next. Place the block after the current block. 3425 Function *CurFn = EntryBB->getParent(); 3426 CurFn->getBasicBlockList().insertAfter(EntryBB->getIterator(), ThenBB); 3427 3428 // Move Entry branch to end of ThenBB, and replace with conditional 3429 // branch (If-stmt) 3430 Instruction *EntryBBTI = EntryBB->getTerminator(); 3431 Builder.CreateCondBr(CallBool, ThenBB, ExitBB); 3432 EntryBBTI->removeFromParent(); 3433 Builder.SetInsertPoint(UI); 3434 Builder.Insert(EntryBBTI); 3435 UI->eraseFromParent(); 3436 Builder.SetInsertPoint(ThenBB->getTerminator()); 3437 3438 // return an insertion point to ExitBB. 3439 return IRBuilder<>::InsertPoint(ExitBB, ExitBB->getFirstInsertionPt()); 3440 } 3441 3442 OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::emitCommonDirectiveExit( 3443 omp::Directive OMPD, InsertPointTy FinIP, Instruction *ExitCall, 3444 bool HasFinalize) { 3445 3446 Builder.restoreIP(FinIP); 3447 3448 // If there is finalization to do, emit it before the exit call 3449 if (HasFinalize) { 3450 assert(!FinalizationStack.empty() && 3451 "Unexpected finalization stack state!"); 3452 3453 FinalizationInfo Fi = FinalizationStack.pop_back_val(); 3454 assert(Fi.DK == OMPD && "Unexpected Directive for Finalization call!"); 3455 3456 Fi.FiniCB(FinIP); 3457 3458 BasicBlock *FiniBB = FinIP.getBlock(); 3459 Instruction *FiniBBTI = FiniBB->getTerminator(); 3460 3461 // set Builder IP for call creation 3462 Builder.SetInsertPoint(FiniBBTI); 3463 } 3464 3465 if (!ExitCall) 3466 return Builder.saveIP(); 3467 3468 // place the Exitcall as last instruction before Finalization block terminator 3469 ExitCall->removeFromParent(); 3470 Builder.Insert(ExitCall); 3471 3472 return IRBuilder<>::InsertPoint(ExitCall->getParent(), 3473 ExitCall->getIterator()); 3474 } 3475 3476 OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::createCopyinClauseBlocks( 3477 InsertPointTy IP, Value *MasterAddr, Value *PrivateAddr, 3478 llvm::IntegerType *IntPtrTy, bool BranchtoEnd) { 3479 if (!IP.isSet()) 3480 return IP; 3481 3482 IRBuilder<>::InsertPointGuard IPG(Builder); 3483 3484 // creates the following CFG structure 3485 // OMP_Entry : (MasterAddr != PrivateAddr)? 3486 // F T 3487 // | \ 3488 // | copin.not.master 3489 // | / 3490 // v / 3491 // copyin.not.master.end 3492 // | 3493 // v 3494 // OMP.Entry.Next 3495 3496 BasicBlock *OMP_Entry = IP.getBlock(); 3497 Function *CurFn = OMP_Entry->getParent(); 3498 BasicBlock *CopyBegin = 3499 BasicBlock::Create(M.getContext(), "copyin.not.master", CurFn); 3500 BasicBlock *CopyEnd = nullptr; 3501 3502 // If entry block is terminated, split to preserve the branch to following 3503 // basic block (i.e. OMP.Entry.Next), otherwise, leave everything as is. 3504 if (isa_and_nonnull<BranchInst>(OMP_Entry->getTerminator())) { 3505 CopyEnd = OMP_Entry->splitBasicBlock(OMP_Entry->getTerminator(), 3506 "copyin.not.master.end"); 3507 OMP_Entry->getTerminator()->eraseFromParent(); 3508 } else { 3509 CopyEnd = 3510 BasicBlock::Create(M.getContext(), "copyin.not.master.end", CurFn); 3511 } 3512 3513 Builder.SetInsertPoint(OMP_Entry); 3514 Value *MasterPtr = Builder.CreatePtrToInt(MasterAddr, IntPtrTy); 3515 Value *PrivatePtr = Builder.CreatePtrToInt(PrivateAddr, IntPtrTy); 3516 Value *cmp = Builder.CreateICmpNE(MasterPtr, PrivatePtr); 3517 Builder.CreateCondBr(cmp, CopyBegin, CopyEnd); 3518 3519 Builder.SetInsertPoint(CopyBegin); 3520 if (BranchtoEnd) 3521 Builder.SetInsertPoint(Builder.CreateBr(CopyEnd)); 3522 3523 return Builder.saveIP(); 3524 } 3525 3526 CallInst *OpenMPIRBuilder::createOMPAlloc(const LocationDescription &Loc, 3527 Value *Size, Value *Allocator, 3528 std::string Name) { 3529 IRBuilder<>::InsertPointGuard IPG(Builder); 3530 Builder.restoreIP(Loc.IP); 3531 3532 uint32_t SrcLocStrSize; 3533 Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize); 3534 Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize); 3535 Value *ThreadId = getOrCreateThreadID(Ident); 3536 Value *Args[] = {ThreadId, Size, Allocator}; 3537 3538 Function *Fn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_alloc); 3539 3540 return Builder.CreateCall(Fn, Args, Name); 3541 } 3542 3543 CallInst *OpenMPIRBuilder::createOMPFree(const LocationDescription &Loc, 3544 Value *Addr, Value *Allocator, 3545 std::string Name) { 3546 IRBuilder<>::InsertPointGuard IPG(Builder); 3547 Builder.restoreIP(Loc.IP); 3548 3549 uint32_t SrcLocStrSize; 3550 Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize); 3551 Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize); 3552 Value *ThreadId = getOrCreateThreadID(Ident); 3553 Value *Args[] = {ThreadId, Addr, Allocator}; 3554 Function *Fn = getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_free); 3555 return Builder.CreateCall(Fn, Args, Name); 3556 } 3557 3558 CallInst *OpenMPIRBuilder::createOMPInteropInit( 3559 const LocationDescription &Loc, Value *InteropVar, 3560 omp::OMPInteropType InteropType, Value *Device, Value *NumDependences, 3561 Value *DependenceAddress, bool HaveNowaitClause) { 3562 IRBuilder<>::InsertPointGuard IPG(Builder); 3563 Builder.restoreIP(Loc.IP); 3564 3565 uint32_t SrcLocStrSize; 3566 Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize); 3567 Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize); 3568 Value *ThreadId = getOrCreateThreadID(Ident); 3569 if (Device == nullptr) 3570 Device = ConstantInt::get(Int32, -1); 3571 Constant *InteropTypeVal = ConstantInt::get(Int64, (int)InteropType); 3572 if (NumDependences == nullptr) { 3573 NumDependences = ConstantInt::get(Int32, 0); 3574 PointerType *PointerTypeVar = Type::getInt8PtrTy(M.getContext()); 3575 DependenceAddress = ConstantPointerNull::get(PointerTypeVar); 3576 } 3577 Value *HaveNowaitClauseVal = ConstantInt::get(Int32, HaveNowaitClause); 3578 Value *Args[] = { 3579 Ident, ThreadId, InteropVar, InteropTypeVal, 3580 Device, NumDependences, DependenceAddress, HaveNowaitClauseVal}; 3581 3582 Function *Fn = getOrCreateRuntimeFunctionPtr(OMPRTL___tgt_interop_init); 3583 3584 return Builder.CreateCall(Fn, Args); 3585 } 3586 3587 CallInst *OpenMPIRBuilder::createOMPInteropDestroy( 3588 const LocationDescription &Loc, Value *InteropVar, Value *Device, 3589 Value *NumDependences, Value *DependenceAddress, bool HaveNowaitClause) { 3590 IRBuilder<>::InsertPointGuard IPG(Builder); 3591 Builder.restoreIP(Loc.IP); 3592 3593 uint32_t SrcLocStrSize; 3594 Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize); 3595 Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize); 3596 Value *ThreadId = getOrCreateThreadID(Ident); 3597 if (Device == nullptr) 3598 Device = ConstantInt::get(Int32, -1); 3599 if (NumDependences == nullptr) { 3600 NumDependences = ConstantInt::get(Int32, 0); 3601 PointerType *PointerTypeVar = Type::getInt8PtrTy(M.getContext()); 3602 DependenceAddress = ConstantPointerNull::get(PointerTypeVar); 3603 } 3604 Value *HaveNowaitClauseVal = ConstantInt::get(Int32, HaveNowaitClause); 3605 Value *Args[] = { 3606 Ident, ThreadId, InteropVar, Device, 3607 NumDependences, DependenceAddress, HaveNowaitClauseVal}; 3608 3609 Function *Fn = getOrCreateRuntimeFunctionPtr(OMPRTL___tgt_interop_destroy); 3610 3611 return Builder.CreateCall(Fn, Args); 3612 } 3613 3614 CallInst *OpenMPIRBuilder::createOMPInteropUse(const LocationDescription &Loc, 3615 Value *InteropVar, Value *Device, 3616 Value *NumDependences, 3617 Value *DependenceAddress, 3618 bool HaveNowaitClause) { 3619 IRBuilder<>::InsertPointGuard IPG(Builder); 3620 Builder.restoreIP(Loc.IP); 3621 uint32_t SrcLocStrSize; 3622 Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize); 3623 Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize); 3624 Value *ThreadId = getOrCreateThreadID(Ident); 3625 if (Device == nullptr) 3626 Device = ConstantInt::get(Int32, -1); 3627 if (NumDependences == nullptr) { 3628 NumDependences = ConstantInt::get(Int32, 0); 3629 PointerType *PointerTypeVar = Type::getInt8PtrTy(M.getContext()); 3630 DependenceAddress = ConstantPointerNull::get(PointerTypeVar); 3631 } 3632 Value *HaveNowaitClauseVal = ConstantInt::get(Int32, HaveNowaitClause); 3633 Value *Args[] = { 3634 Ident, ThreadId, InteropVar, Device, 3635 NumDependences, DependenceAddress, HaveNowaitClauseVal}; 3636 3637 Function *Fn = getOrCreateRuntimeFunctionPtr(OMPRTL___tgt_interop_use); 3638 3639 return Builder.CreateCall(Fn, Args); 3640 } 3641 3642 CallInst *OpenMPIRBuilder::createCachedThreadPrivate( 3643 const LocationDescription &Loc, llvm::Value *Pointer, 3644 llvm::ConstantInt *Size, const llvm::Twine &Name) { 3645 IRBuilder<>::InsertPointGuard IPG(Builder); 3646 Builder.restoreIP(Loc.IP); 3647 3648 uint32_t SrcLocStrSize; 3649 Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize); 3650 Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize); 3651 Value *ThreadId = getOrCreateThreadID(Ident); 3652 Constant *ThreadPrivateCache = 3653 getOrCreateOMPInternalVariable(Int8PtrPtr, Name); 3654 llvm::Value *Args[] = {Ident, ThreadId, Pointer, Size, ThreadPrivateCache}; 3655 3656 Function *Fn = 3657 getOrCreateRuntimeFunctionPtr(OMPRTL___kmpc_threadprivate_cached); 3658 3659 return Builder.CreateCall(Fn, Args); 3660 } 3661 3662 OpenMPIRBuilder::InsertPointTy 3663 OpenMPIRBuilder::createTargetInit(const LocationDescription &Loc, bool IsSPMD, 3664 bool RequiresFullRuntime) { 3665 if (!updateToLocation(Loc)) 3666 return Loc.IP; 3667 3668 uint32_t SrcLocStrSize; 3669 Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize); 3670 Constant *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize); 3671 ConstantInt *IsSPMDVal = ConstantInt::getSigned( 3672 IntegerType::getInt8Ty(Int8->getContext()), 3673 IsSPMD ? OMP_TGT_EXEC_MODE_SPMD : OMP_TGT_EXEC_MODE_GENERIC); 3674 ConstantInt *UseGenericStateMachine = 3675 ConstantInt::getBool(Int32->getContext(), !IsSPMD); 3676 ConstantInt *RequiresFullRuntimeVal = 3677 ConstantInt::getBool(Int32->getContext(), RequiresFullRuntime); 3678 3679 Function *Fn = getOrCreateRuntimeFunctionPtr( 3680 omp::RuntimeFunction::OMPRTL___kmpc_target_init); 3681 3682 CallInst *ThreadKind = Builder.CreateCall( 3683 Fn, {Ident, IsSPMDVal, UseGenericStateMachine, RequiresFullRuntimeVal}); 3684 3685 Value *ExecUserCode = Builder.CreateICmpEQ( 3686 ThreadKind, ConstantInt::get(ThreadKind->getType(), -1), 3687 "exec_user_code"); 3688 3689 // ThreadKind = __kmpc_target_init(...) 3690 // if (ThreadKind == -1) 3691 // user_code 3692 // else 3693 // return; 3694 3695 auto *UI = Builder.CreateUnreachable(); 3696 BasicBlock *CheckBB = UI->getParent(); 3697 BasicBlock *UserCodeEntryBB = CheckBB->splitBasicBlock(UI, "user_code.entry"); 3698 3699 BasicBlock *WorkerExitBB = BasicBlock::Create( 3700 CheckBB->getContext(), "worker.exit", CheckBB->getParent()); 3701 Builder.SetInsertPoint(WorkerExitBB); 3702 Builder.CreateRetVoid(); 3703 3704 auto *CheckBBTI = CheckBB->getTerminator(); 3705 Builder.SetInsertPoint(CheckBBTI); 3706 Builder.CreateCondBr(ExecUserCode, UI->getParent(), WorkerExitBB); 3707 3708 CheckBBTI->eraseFromParent(); 3709 UI->eraseFromParent(); 3710 3711 // Continue in the "user_code" block, see diagram above and in 3712 // openmp/libomptarget/deviceRTLs/common/include/target.h . 3713 return InsertPointTy(UserCodeEntryBB, UserCodeEntryBB->getFirstInsertionPt()); 3714 } 3715 3716 void OpenMPIRBuilder::createTargetDeinit(const LocationDescription &Loc, 3717 bool IsSPMD, 3718 bool RequiresFullRuntime) { 3719 if (!updateToLocation(Loc)) 3720 return; 3721 3722 uint32_t SrcLocStrSize; 3723 Constant *SrcLocStr = getOrCreateSrcLocStr(Loc, SrcLocStrSize); 3724 Value *Ident = getOrCreateIdent(SrcLocStr, SrcLocStrSize); 3725 ConstantInt *IsSPMDVal = ConstantInt::getSigned( 3726 IntegerType::getInt8Ty(Int8->getContext()), 3727 IsSPMD ? OMP_TGT_EXEC_MODE_SPMD : OMP_TGT_EXEC_MODE_GENERIC); 3728 ConstantInt *RequiresFullRuntimeVal = 3729 ConstantInt::getBool(Int32->getContext(), RequiresFullRuntime); 3730 3731 Function *Fn = getOrCreateRuntimeFunctionPtr( 3732 omp::RuntimeFunction::OMPRTL___kmpc_target_deinit); 3733 3734 Builder.CreateCall(Fn, {Ident, IsSPMDVal, RequiresFullRuntimeVal}); 3735 } 3736 3737 std::string OpenMPIRBuilder::getNameWithSeparators(ArrayRef<StringRef> Parts, 3738 StringRef FirstSeparator, 3739 StringRef Separator) { 3740 SmallString<128> Buffer; 3741 llvm::raw_svector_ostream OS(Buffer); 3742 StringRef Sep = FirstSeparator; 3743 for (StringRef Part : Parts) { 3744 OS << Sep << Part; 3745 Sep = Separator; 3746 } 3747 return OS.str().str(); 3748 } 3749 3750 Constant *OpenMPIRBuilder::getOrCreateOMPInternalVariable( 3751 llvm::Type *Ty, const llvm::Twine &Name, unsigned AddressSpace) { 3752 // TODO: Replace the twine arg with stringref to get rid of the conversion 3753 // logic. However This is taken from current implementation in clang as is. 3754 // Since this method is used in many places exclusively for OMP internal use 3755 // we will keep it as is for temporarily until we move all users to the 3756 // builder and then, if possible, fix it everywhere in one go. 3757 SmallString<256> Buffer; 3758 llvm::raw_svector_ostream Out(Buffer); 3759 Out << Name; 3760 StringRef RuntimeName = Out.str(); 3761 auto &Elem = *InternalVars.try_emplace(RuntimeName, nullptr).first; 3762 if (Elem.second) { 3763 assert(cast<PointerType>(Elem.second->getType()) 3764 ->isOpaqueOrPointeeTypeMatches(Ty) && 3765 "OMP internal variable has different type than requested"); 3766 } else { 3767 // TODO: investigate the appropriate linkage type used for the global 3768 // variable for possibly changing that to internal or private, or maybe 3769 // create different versions of the function for different OMP internal 3770 // variables. 3771 Elem.second = new llvm::GlobalVariable( 3772 M, Ty, /*IsConstant*/ false, llvm::GlobalValue::CommonLinkage, 3773 llvm::Constant::getNullValue(Ty), Elem.first(), 3774 /*InsertBefore=*/nullptr, llvm::GlobalValue::NotThreadLocal, 3775 AddressSpace); 3776 } 3777 3778 return Elem.second; 3779 } 3780 3781 Value *OpenMPIRBuilder::getOMPCriticalRegionLock(StringRef CriticalName) { 3782 std::string Prefix = Twine("gomp_critical_user_", CriticalName).str(); 3783 std::string Name = getNameWithSeparators({Prefix, "var"}, ".", "."); 3784 return getOrCreateOMPInternalVariable(KmpCriticalNameTy, Name); 3785 } 3786 3787 GlobalVariable * 3788 OpenMPIRBuilder::createOffloadMaptypes(SmallVectorImpl<uint64_t> &Mappings, 3789 std::string VarName) { 3790 llvm::Constant *MaptypesArrayInit = 3791 llvm::ConstantDataArray::get(M.getContext(), Mappings); 3792 auto *MaptypesArrayGlobal = new llvm::GlobalVariable( 3793 M, MaptypesArrayInit->getType(), 3794 /*isConstant=*/true, llvm::GlobalValue::PrivateLinkage, MaptypesArrayInit, 3795 VarName); 3796 MaptypesArrayGlobal->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Global); 3797 return MaptypesArrayGlobal; 3798 } 3799 3800 void OpenMPIRBuilder::createMapperAllocas(const LocationDescription &Loc, 3801 InsertPointTy AllocaIP, 3802 unsigned NumOperands, 3803 struct MapperAllocas &MapperAllocas) { 3804 if (!updateToLocation(Loc)) 3805 return; 3806 3807 auto *ArrI8PtrTy = ArrayType::get(Int8Ptr, NumOperands); 3808 auto *ArrI64Ty = ArrayType::get(Int64, NumOperands); 3809 Builder.restoreIP(AllocaIP); 3810 AllocaInst *ArgsBase = Builder.CreateAlloca(ArrI8PtrTy); 3811 AllocaInst *Args = Builder.CreateAlloca(ArrI8PtrTy); 3812 AllocaInst *ArgSizes = Builder.CreateAlloca(ArrI64Ty); 3813 Builder.restoreIP(Loc.IP); 3814 MapperAllocas.ArgsBase = ArgsBase; 3815 MapperAllocas.Args = Args; 3816 MapperAllocas.ArgSizes = ArgSizes; 3817 } 3818 3819 void OpenMPIRBuilder::emitMapperCall(const LocationDescription &Loc, 3820 Function *MapperFunc, Value *SrcLocInfo, 3821 Value *MaptypesArg, Value *MapnamesArg, 3822 struct MapperAllocas &MapperAllocas, 3823 int64_t DeviceID, unsigned NumOperands) { 3824 if (!updateToLocation(Loc)) 3825 return; 3826 3827 auto *ArrI8PtrTy = ArrayType::get(Int8Ptr, NumOperands); 3828 auto *ArrI64Ty = ArrayType::get(Int64, NumOperands); 3829 Value *ArgsBaseGEP = 3830 Builder.CreateInBoundsGEP(ArrI8PtrTy, MapperAllocas.ArgsBase, 3831 {Builder.getInt32(0), Builder.getInt32(0)}); 3832 Value *ArgsGEP = 3833 Builder.CreateInBoundsGEP(ArrI8PtrTy, MapperAllocas.Args, 3834 {Builder.getInt32(0), Builder.getInt32(0)}); 3835 Value *ArgSizesGEP = 3836 Builder.CreateInBoundsGEP(ArrI64Ty, MapperAllocas.ArgSizes, 3837 {Builder.getInt32(0), Builder.getInt32(0)}); 3838 Value *NullPtr = Constant::getNullValue(Int8Ptr->getPointerTo()); 3839 Builder.CreateCall(MapperFunc, 3840 {SrcLocInfo, Builder.getInt64(DeviceID), 3841 Builder.getInt32(NumOperands), ArgsBaseGEP, ArgsGEP, 3842 ArgSizesGEP, MaptypesArg, MapnamesArg, NullPtr}); 3843 } 3844 3845 bool OpenMPIRBuilder::checkAndEmitFlushAfterAtomic( 3846 const LocationDescription &Loc, llvm::AtomicOrdering AO, AtomicKind AK) { 3847 assert(!(AO == AtomicOrdering::NotAtomic || 3848 AO == llvm::AtomicOrdering::Unordered) && 3849 "Unexpected Atomic Ordering."); 3850 3851 bool Flush = false; 3852 llvm::AtomicOrdering FlushAO = AtomicOrdering::Monotonic; 3853 3854 switch (AK) { 3855 case Read: 3856 if (AO == AtomicOrdering::Acquire || AO == AtomicOrdering::AcquireRelease || 3857 AO == AtomicOrdering::SequentiallyConsistent) { 3858 FlushAO = AtomicOrdering::Acquire; 3859 Flush = true; 3860 } 3861 break; 3862 case Write: 3863 case Compare: 3864 case Update: 3865 if (AO == AtomicOrdering::Release || AO == AtomicOrdering::AcquireRelease || 3866 AO == AtomicOrdering::SequentiallyConsistent) { 3867 FlushAO = AtomicOrdering::Release; 3868 Flush = true; 3869 } 3870 break; 3871 case Capture: 3872 switch (AO) { 3873 case AtomicOrdering::Acquire: 3874 FlushAO = AtomicOrdering::Acquire; 3875 Flush = true; 3876 break; 3877 case AtomicOrdering::Release: 3878 FlushAO = AtomicOrdering::Release; 3879 Flush = true; 3880 break; 3881 case AtomicOrdering::AcquireRelease: 3882 case AtomicOrdering::SequentiallyConsistent: 3883 FlushAO = AtomicOrdering::AcquireRelease; 3884 Flush = true; 3885 break; 3886 default: 3887 // do nothing - leave silently. 3888 break; 3889 } 3890 } 3891 3892 if (Flush) { 3893 // Currently Flush RT call still doesn't take memory_ordering, so for when 3894 // that happens, this tries to do the resolution of which atomic ordering 3895 // to use with but issue the flush call 3896 // TODO: pass `FlushAO` after memory ordering support is added 3897 (void)FlushAO; 3898 emitFlush(Loc); 3899 } 3900 3901 // for AO == AtomicOrdering::Monotonic and all other case combinations 3902 // do nothing 3903 return Flush; 3904 } 3905 3906 OpenMPIRBuilder::InsertPointTy 3907 OpenMPIRBuilder::createAtomicRead(const LocationDescription &Loc, 3908 AtomicOpValue &X, AtomicOpValue &V, 3909 AtomicOrdering AO) { 3910 if (!updateToLocation(Loc)) 3911 return Loc.IP; 3912 3913 Type *XTy = X.Var->getType(); 3914 assert(XTy->isPointerTy() && "OMP Atomic expects a pointer to target memory"); 3915 Type *XElemTy = X.ElemTy; 3916 assert((XElemTy->isFloatingPointTy() || XElemTy->isIntegerTy() || 3917 XElemTy->isPointerTy()) && 3918 "OMP atomic read expected a scalar type"); 3919 3920 Value *XRead = nullptr; 3921 3922 if (XElemTy->isIntegerTy()) { 3923 LoadInst *XLD = 3924 Builder.CreateLoad(XElemTy, X.Var, X.IsVolatile, "omp.atomic.read"); 3925 XLD->setAtomic(AO); 3926 XRead = cast<Value>(XLD); 3927 } else { 3928 // We need to bitcast and perform atomic op as integer 3929 unsigned Addrspace = cast<PointerType>(XTy)->getAddressSpace(); 3930 IntegerType *IntCastTy = 3931 IntegerType::get(M.getContext(), XElemTy->getScalarSizeInBits()); 3932 Value *XBCast = Builder.CreateBitCast( 3933 X.Var, IntCastTy->getPointerTo(Addrspace), "atomic.src.int.cast"); 3934 LoadInst *XLoad = 3935 Builder.CreateLoad(IntCastTy, XBCast, X.IsVolatile, "omp.atomic.load"); 3936 XLoad->setAtomic(AO); 3937 if (XElemTy->isFloatingPointTy()) { 3938 XRead = Builder.CreateBitCast(XLoad, XElemTy, "atomic.flt.cast"); 3939 } else { 3940 XRead = Builder.CreateIntToPtr(XLoad, XElemTy, "atomic.ptr.cast"); 3941 } 3942 } 3943 checkAndEmitFlushAfterAtomic(Loc, AO, AtomicKind::Read); 3944 Builder.CreateStore(XRead, V.Var, V.IsVolatile); 3945 return Builder.saveIP(); 3946 } 3947 3948 OpenMPIRBuilder::InsertPointTy 3949 OpenMPIRBuilder::createAtomicWrite(const LocationDescription &Loc, 3950 AtomicOpValue &X, Value *Expr, 3951 AtomicOrdering AO) { 3952 if (!updateToLocation(Loc)) 3953 return Loc.IP; 3954 3955 Type *XTy = X.Var->getType(); 3956 assert(XTy->isPointerTy() && "OMP Atomic expects a pointer to target memory"); 3957 Type *XElemTy = X.ElemTy; 3958 assert((XElemTy->isFloatingPointTy() || XElemTy->isIntegerTy() || 3959 XElemTy->isPointerTy()) && 3960 "OMP atomic write expected a scalar type"); 3961 3962 if (XElemTy->isIntegerTy()) { 3963 StoreInst *XSt = Builder.CreateStore(Expr, X.Var, X.IsVolatile); 3964 XSt->setAtomic(AO); 3965 } else { 3966 // We need to bitcast and perform atomic op as integers 3967 unsigned Addrspace = cast<PointerType>(XTy)->getAddressSpace(); 3968 IntegerType *IntCastTy = 3969 IntegerType::get(M.getContext(), XElemTy->getScalarSizeInBits()); 3970 Value *XBCast = Builder.CreateBitCast( 3971 X.Var, IntCastTy->getPointerTo(Addrspace), "atomic.dst.int.cast"); 3972 Value *ExprCast = 3973 Builder.CreateBitCast(Expr, IntCastTy, "atomic.src.int.cast"); 3974 StoreInst *XSt = Builder.CreateStore(ExprCast, XBCast, X.IsVolatile); 3975 XSt->setAtomic(AO); 3976 } 3977 3978 checkAndEmitFlushAfterAtomic(Loc, AO, AtomicKind::Write); 3979 return Builder.saveIP(); 3980 } 3981 3982 OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::createAtomicUpdate( 3983 const LocationDescription &Loc, InsertPointTy AllocaIP, AtomicOpValue &X, 3984 Value *Expr, AtomicOrdering AO, AtomicRMWInst::BinOp RMWOp, 3985 AtomicUpdateCallbackTy &UpdateOp, bool IsXBinopExpr) { 3986 assert(!isConflictIP(Loc.IP, AllocaIP) && "IPs must not be ambiguous"); 3987 if (!updateToLocation(Loc)) 3988 return Loc.IP; 3989 3990 LLVM_DEBUG({ 3991 Type *XTy = X.Var->getType(); 3992 assert(XTy->isPointerTy() && 3993 "OMP Atomic expects a pointer to target memory"); 3994 Type *XElemTy = X.ElemTy; 3995 assert((XElemTy->isFloatingPointTy() || XElemTy->isIntegerTy() || 3996 XElemTy->isPointerTy()) && 3997 "OMP atomic update expected a scalar type"); 3998 assert((RMWOp != AtomicRMWInst::Max) && (RMWOp != AtomicRMWInst::Min) && 3999 (RMWOp != AtomicRMWInst::UMax) && (RMWOp != AtomicRMWInst::UMin) && 4000 "OpenMP atomic does not support LT or GT operations"); 4001 }); 4002 4003 emitAtomicUpdate(AllocaIP, X.Var, X.ElemTy, Expr, AO, RMWOp, UpdateOp, 4004 X.IsVolatile, IsXBinopExpr); 4005 checkAndEmitFlushAfterAtomic(Loc, AO, AtomicKind::Update); 4006 return Builder.saveIP(); 4007 } 4008 4009 Value *OpenMPIRBuilder::emitRMWOpAsInstruction(Value *Src1, Value *Src2, 4010 AtomicRMWInst::BinOp RMWOp) { 4011 switch (RMWOp) { 4012 case AtomicRMWInst::Add: 4013 return Builder.CreateAdd(Src1, Src2); 4014 case AtomicRMWInst::Sub: 4015 return Builder.CreateSub(Src1, Src2); 4016 case AtomicRMWInst::And: 4017 return Builder.CreateAnd(Src1, Src2); 4018 case AtomicRMWInst::Nand: 4019 return Builder.CreateNeg(Builder.CreateAnd(Src1, Src2)); 4020 case AtomicRMWInst::Or: 4021 return Builder.CreateOr(Src1, Src2); 4022 case AtomicRMWInst::Xor: 4023 return Builder.CreateXor(Src1, Src2); 4024 case AtomicRMWInst::Xchg: 4025 case AtomicRMWInst::FAdd: 4026 case AtomicRMWInst::FSub: 4027 case AtomicRMWInst::BAD_BINOP: 4028 case AtomicRMWInst::Max: 4029 case AtomicRMWInst::Min: 4030 case AtomicRMWInst::UMax: 4031 case AtomicRMWInst::UMin: 4032 case AtomicRMWInst::FMax: 4033 case AtomicRMWInst::FMin: 4034 llvm_unreachable("Unsupported atomic update operation"); 4035 } 4036 llvm_unreachable("Unsupported atomic update operation"); 4037 } 4038 4039 std::pair<Value *, Value *> OpenMPIRBuilder::emitAtomicUpdate( 4040 InsertPointTy AllocaIP, Value *X, Type *XElemTy, Value *Expr, 4041 AtomicOrdering AO, AtomicRMWInst::BinOp RMWOp, 4042 AtomicUpdateCallbackTy &UpdateOp, bool VolatileX, bool IsXBinopExpr) { 4043 // TODO: handle the case where XElemTy is not byte-sized or not a power of 2 4044 // or a complex datatype. 4045 bool emitRMWOp = false; 4046 switch (RMWOp) { 4047 case AtomicRMWInst::Add: 4048 case AtomicRMWInst::And: 4049 case AtomicRMWInst::Nand: 4050 case AtomicRMWInst::Or: 4051 case AtomicRMWInst::Xor: 4052 case AtomicRMWInst::Xchg: 4053 emitRMWOp = XElemTy; 4054 break; 4055 case AtomicRMWInst::Sub: 4056 emitRMWOp = (IsXBinopExpr && XElemTy); 4057 break; 4058 default: 4059 emitRMWOp = false; 4060 } 4061 emitRMWOp &= XElemTy->isIntegerTy(); 4062 4063 std::pair<Value *, Value *> Res; 4064 if (emitRMWOp) { 4065 Res.first = Builder.CreateAtomicRMW(RMWOp, X, Expr, llvm::MaybeAlign(), AO); 4066 // not needed except in case of postfix captures. Generate anyway for 4067 // consistency with the else part. Will be removed with any DCE pass. 4068 // AtomicRMWInst::Xchg does not have a coressponding instruction. 4069 if (RMWOp == AtomicRMWInst::Xchg) 4070 Res.second = Res.first; 4071 else 4072 Res.second = emitRMWOpAsInstruction(Res.first, Expr, RMWOp); 4073 } else { 4074 unsigned Addrspace = cast<PointerType>(X->getType())->getAddressSpace(); 4075 IntegerType *IntCastTy = 4076 IntegerType::get(M.getContext(), XElemTy->getScalarSizeInBits()); 4077 Value *XBCast = 4078 Builder.CreateBitCast(X, IntCastTy->getPointerTo(Addrspace)); 4079 LoadInst *OldVal = 4080 Builder.CreateLoad(IntCastTy, XBCast, X->getName() + ".atomic.load"); 4081 OldVal->setAtomic(AO); 4082 // CurBB 4083 // | /---\ 4084 // ContBB | 4085 // | \---/ 4086 // ExitBB 4087 BasicBlock *CurBB = Builder.GetInsertBlock(); 4088 Instruction *CurBBTI = CurBB->getTerminator(); 4089 CurBBTI = CurBBTI ? CurBBTI : Builder.CreateUnreachable(); 4090 BasicBlock *ExitBB = 4091 CurBB->splitBasicBlock(CurBBTI, X->getName() + ".atomic.exit"); 4092 BasicBlock *ContBB = CurBB->splitBasicBlock(CurBB->getTerminator(), 4093 X->getName() + ".atomic.cont"); 4094 ContBB->getTerminator()->eraseFromParent(); 4095 Builder.restoreIP(AllocaIP); 4096 AllocaInst *NewAtomicAddr = Builder.CreateAlloca(XElemTy); 4097 NewAtomicAddr->setName(X->getName() + "x.new.val"); 4098 Builder.SetInsertPoint(ContBB); 4099 llvm::PHINode *PHI = Builder.CreatePHI(OldVal->getType(), 2); 4100 PHI->addIncoming(OldVal, CurBB); 4101 IntegerType *NewAtomicCastTy = 4102 IntegerType::get(M.getContext(), XElemTy->getScalarSizeInBits()); 4103 bool IsIntTy = XElemTy->isIntegerTy(); 4104 Value *NewAtomicIntAddr = 4105 (IsIntTy) 4106 ? NewAtomicAddr 4107 : Builder.CreateBitCast(NewAtomicAddr, 4108 NewAtomicCastTy->getPointerTo(Addrspace)); 4109 Value *OldExprVal = PHI; 4110 if (!IsIntTy) { 4111 if (XElemTy->isFloatingPointTy()) { 4112 OldExprVal = Builder.CreateBitCast(PHI, XElemTy, 4113 X->getName() + ".atomic.fltCast"); 4114 } else { 4115 OldExprVal = Builder.CreateIntToPtr(PHI, XElemTy, 4116 X->getName() + ".atomic.ptrCast"); 4117 } 4118 } 4119 4120 Value *Upd = UpdateOp(OldExprVal, Builder); 4121 Builder.CreateStore(Upd, NewAtomicAddr); 4122 LoadInst *DesiredVal = Builder.CreateLoad(IntCastTy, NewAtomicIntAddr); 4123 Value *XAddr = 4124 (IsIntTy) 4125 ? X 4126 : Builder.CreateBitCast(X, IntCastTy->getPointerTo(Addrspace)); 4127 AtomicOrdering Failure = 4128 llvm::AtomicCmpXchgInst::getStrongestFailureOrdering(AO); 4129 AtomicCmpXchgInst *Result = Builder.CreateAtomicCmpXchg( 4130 XAddr, PHI, DesiredVal, llvm::MaybeAlign(), AO, Failure); 4131 Result->setVolatile(VolatileX); 4132 Value *PreviousVal = Builder.CreateExtractValue(Result, /*Idxs=*/0); 4133 Value *SuccessFailureVal = Builder.CreateExtractValue(Result, /*Idxs=*/1); 4134 PHI->addIncoming(PreviousVal, Builder.GetInsertBlock()); 4135 Builder.CreateCondBr(SuccessFailureVal, ExitBB, ContBB); 4136 4137 Res.first = OldExprVal; 4138 Res.second = Upd; 4139 4140 // set Insertion point in exit block 4141 if (UnreachableInst *ExitTI = 4142 dyn_cast<UnreachableInst>(ExitBB->getTerminator())) { 4143 CurBBTI->eraseFromParent(); 4144 Builder.SetInsertPoint(ExitBB); 4145 } else { 4146 Builder.SetInsertPoint(ExitTI); 4147 } 4148 } 4149 4150 return Res; 4151 } 4152 4153 OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::createAtomicCapture( 4154 const LocationDescription &Loc, InsertPointTy AllocaIP, AtomicOpValue &X, 4155 AtomicOpValue &V, Value *Expr, AtomicOrdering AO, 4156 AtomicRMWInst::BinOp RMWOp, AtomicUpdateCallbackTy &UpdateOp, 4157 bool UpdateExpr, bool IsPostfixUpdate, bool IsXBinopExpr) { 4158 if (!updateToLocation(Loc)) 4159 return Loc.IP; 4160 4161 LLVM_DEBUG({ 4162 Type *XTy = X.Var->getType(); 4163 assert(XTy->isPointerTy() && 4164 "OMP Atomic expects a pointer to target memory"); 4165 Type *XElemTy = X.ElemTy; 4166 assert((XElemTy->isFloatingPointTy() || XElemTy->isIntegerTy() || 4167 XElemTy->isPointerTy()) && 4168 "OMP atomic capture expected a scalar type"); 4169 assert((RMWOp != AtomicRMWInst::Max) && (RMWOp != AtomicRMWInst::Min) && 4170 "OpenMP atomic does not support LT or GT operations"); 4171 }); 4172 4173 // If UpdateExpr is 'x' updated with some `expr` not based on 'x', 4174 // 'x' is simply atomically rewritten with 'expr'. 4175 AtomicRMWInst::BinOp AtomicOp = (UpdateExpr ? RMWOp : AtomicRMWInst::Xchg); 4176 std::pair<Value *, Value *> Result = 4177 emitAtomicUpdate(AllocaIP, X.Var, X.ElemTy, Expr, AO, AtomicOp, UpdateOp, 4178 X.IsVolatile, IsXBinopExpr); 4179 4180 Value *CapturedVal = (IsPostfixUpdate ? Result.first : Result.second); 4181 Builder.CreateStore(CapturedVal, V.Var, V.IsVolatile); 4182 4183 checkAndEmitFlushAfterAtomic(Loc, AO, AtomicKind::Capture); 4184 return Builder.saveIP(); 4185 } 4186 4187 OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::createAtomicCompare( 4188 const LocationDescription &Loc, AtomicOpValue &X, AtomicOpValue &V, 4189 AtomicOpValue &R, Value *E, Value *D, AtomicOrdering AO, 4190 omp::OMPAtomicCompareOp Op, bool IsXBinopExpr, bool IsPostfixUpdate, 4191 bool IsFailOnly) { 4192 4193 if (!updateToLocation(Loc)) 4194 return Loc.IP; 4195 4196 assert(X.Var->getType()->isPointerTy() && 4197 "OMP atomic expects a pointer to target memory"); 4198 // compare capture 4199 if (V.Var) { 4200 assert(V.Var->getType()->isPointerTy() && "v.var must be of pointer type"); 4201 assert(V.ElemTy == X.ElemTy && "x and v must be of same type"); 4202 } 4203 4204 bool IsInteger = E->getType()->isIntegerTy(); 4205 4206 if (Op == OMPAtomicCompareOp::EQ) { 4207 AtomicOrdering Failure = AtomicCmpXchgInst::getStrongestFailureOrdering(AO); 4208 AtomicCmpXchgInst *Result = nullptr; 4209 if (!IsInteger) { 4210 unsigned Addrspace = 4211 cast<PointerType>(X.Var->getType())->getAddressSpace(); 4212 IntegerType *IntCastTy = 4213 IntegerType::get(M.getContext(), X.ElemTy->getScalarSizeInBits()); 4214 Value *XBCast = 4215 Builder.CreateBitCast(X.Var, IntCastTy->getPointerTo(Addrspace)); 4216 Value *EBCast = Builder.CreateBitCast(E, IntCastTy); 4217 Value *DBCast = Builder.CreateBitCast(D, IntCastTy); 4218 Result = Builder.CreateAtomicCmpXchg(XBCast, EBCast, DBCast, MaybeAlign(), 4219 AO, Failure); 4220 } else { 4221 Result = 4222 Builder.CreateAtomicCmpXchg(X.Var, E, D, MaybeAlign(), AO, Failure); 4223 } 4224 4225 if (V.Var) { 4226 Value *OldValue = Builder.CreateExtractValue(Result, /*Idxs=*/0); 4227 if (!IsInteger) 4228 OldValue = Builder.CreateBitCast(OldValue, X.ElemTy); 4229 assert(OldValue->getType() == V.ElemTy && 4230 "OldValue and V must be of same type"); 4231 if (IsPostfixUpdate) { 4232 Builder.CreateStore(OldValue, V.Var, V.IsVolatile); 4233 } else { 4234 Value *SuccessOrFail = Builder.CreateExtractValue(Result, /*Idxs=*/1); 4235 if (IsFailOnly) { 4236 // CurBB---- 4237 // | | 4238 // v | 4239 // ContBB | 4240 // | | 4241 // v | 4242 // ExitBB <- 4243 // 4244 // where ContBB only contains the store of old value to 'v'. 4245 BasicBlock *CurBB = Builder.GetInsertBlock(); 4246 Instruction *CurBBTI = CurBB->getTerminator(); 4247 CurBBTI = CurBBTI ? CurBBTI : Builder.CreateUnreachable(); 4248 BasicBlock *ExitBB = CurBB->splitBasicBlock( 4249 CurBBTI, X.Var->getName() + ".atomic.exit"); 4250 BasicBlock *ContBB = CurBB->splitBasicBlock( 4251 CurBB->getTerminator(), X.Var->getName() + ".atomic.cont"); 4252 ContBB->getTerminator()->eraseFromParent(); 4253 CurBB->getTerminator()->eraseFromParent(); 4254 4255 Builder.CreateCondBr(SuccessOrFail, ExitBB, ContBB); 4256 4257 Builder.SetInsertPoint(ContBB); 4258 Builder.CreateStore(OldValue, V.Var); 4259 Builder.CreateBr(ExitBB); 4260 4261 if (UnreachableInst *ExitTI = 4262 dyn_cast<UnreachableInst>(ExitBB->getTerminator())) { 4263 CurBBTI->eraseFromParent(); 4264 Builder.SetInsertPoint(ExitBB); 4265 } else { 4266 Builder.SetInsertPoint(ExitTI); 4267 } 4268 } else { 4269 Value *CapturedValue = 4270 Builder.CreateSelect(SuccessOrFail, E, OldValue); 4271 Builder.CreateStore(CapturedValue, V.Var, V.IsVolatile); 4272 } 4273 } 4274 } 4275 // The comparison result has to be stored. 4276 if (R.Var) { 4277 assert(R.Var->getType()->isPointerTy() && 4278 "r.var must be of pointer type"); 4279 assert(R.ElemTy->isIntegerTy() && "r must be of integral type"); 4280 4281 Value *SuccessFailureVal = Builder.CreateExtractValue(Result, /*Idxs=*/1); 4282 Value *ResultCast = R.IsSigned 4283 ? Builder.CreateSExt(SuccessFailureVal, R.ElemTy) 4284 : Builder.CreateZExt(SuccessFailureVal, R.ElemTy); 4285 Builder.CreateStore(ResultCast, R.Var, R.IsVolatile); 4286 } 4287 } else { 4288 assert((Op == OMPAtomicCompareOp::MAX || Op == OMPAtomicCompareOp::MIN) && 4289 "Op should be either max or min at this point"); 4290 assert(!IsFailOnly && "IsFailOnly is only valid when the comparison is =="); 4291 4292 // Reverse the ordop as the OpenMP forms are different from LLVM forms. 4293 // Let's take max as example. 4294 // OpenMP form: 4295 // x = x > expr ? expr : x; 4296 // LLVM form: 4297 // *ptr = *ptr > val ? *ptr : val; 4298 // We need to transform to LLVM form. 4299 // x = x <= expr ? x : expr; 4300 AtomicRMWInst::BinOp NewOp; 4301 if (IsXBinopExpr) { 4302 if (IsInteger) { 4303 if (X.IsSigned) 4304 NewOp = Op == OMPAtomicCompareOp::MAX ? AtomicRMWInst::Min 4305 : AtomicRMWInst::Max; 4306 else 4307 NewOp = Op == OMPAtomicCompareOp::MAX ? AtomicRMWInst::UMin 4308 : AtomicRMWInst::UMax; 4309 } else { 4310 NewOp = Op == OMPAtomicCompareOp::MAX ? AtomicRMWInst::FMin 4311 : AtomicRMWInst::FMax; 4312 } 4313 } else { 4314 if (IsInteger) { 4315 if (X.IsSigned) 4316 NewOp = Op == OMPAtomicCompareOp::MAX ? AtomicRMWInst::Max 4317 : AtomicRMWInst::Min; 4318 else 4319 NewOp = Op == OMPAtomicCompareOp::MAX ? AtomicRMWInst::UMax 4320 : AtomicRMWInst::UMin; 4321 } else { 4322 NewOp = Op == OMPAtomicCompareOp::MAX ? AtomicRMWInst::FMax 4323 : AtomicRMWInst::FMin; 4324 } 4325 } 4326 4327 AtomicRMWInst *OldValue = 4328 Builder.CreateAtomicRMW(NewOp, X.Var, E, MaybeAlign(), AO); 4329 if (V.Var) { 4330 Value *CapturedValue = nullptr; 4331 if (IsPostfixUpdate) { 4332 CapturedValue = OldValue; 4333 } else { 4334 CmpInst::Predicate Pred; 4335 switch (NewOp) { 4336 case AtomicRMWInst::Max: 4337 Pred = CmpInst::ICMP_SGT; 4338 break; 4339 case AtomicRMWInst::UMax: 4340 Pred = CmpInst::ICMP_UGT; 4341 break; 4342 case AtomicRMWInst::FMax: 4343 Pred = CmpInst::FCMP_OGT; 4344 break; 4345 case AtomicRMWInst::Min: 4346 Pred = CmpInst::ICMP_SLT; 4347 break; 4348 case AtomicRMWInst::UMin: 4349 Pred = CmpInst::ICMP_ULT; 4350 break; 4351 case AtomicRMWInst::FMin: 4352 Pred = CmpInst::FCMP_OLT; 4353 break; 4354 default: 4355 llvm_unreachable("unexpected comparison op"); 4356 } 4357 Value *NonAtomicCmp = Builder.CreateCmp(Pred, OldValue, E); 4358 CapturedValue = Builder.CreateSelect(NonAtomicCmp, E, OldValue); 4359 } 4360 Builder.CreateStore(CapturedValue, V.Var, V.IsVolatile); 4361 } 4362 } 4363 4364 checkAndEmitFlushAfterAtomic(Loc, AO, AtomicKind::Compare); 4365 4366 return Builder.saveIP(); 4367 } 4368 4369 GlobalVariable * 4370 OpenMPIRBuilder::createOffloadMapnames(SmallVectorImpl<llvm::Constant *> &Names, 4371 std::string VarName) { 4372 llvm::Constant *MapNamesArrayInit = llvm::ConstantArray::get( 4373 llvm::ArrayType::get( 4374 llvm::Type::getInt8Ty(M.getContext())->getPointerTo(), Names.size()), 4375 Names); 4376 auto *MapNamesArrayGlobal = new llvm::GlobalVariable( 4377 M, MapNamesArrayInit->getType(), 4378 /*isConstant=*/true, llvm::GlobalValue::PrivateLinkage, MapNamesArrayInit, 4379 VarName); 4380 return MapNamesArrayGlobal; 4381 } 4382 4383 // Create all simple and struct types exposed by the runtime and remember 4384 // the llvm::PointerTypes of them for easy access later. 4385 void OpenMPIRBuilder::initializeTypes(Module &M) { 4386 LLVMContext &Ctx = M.getContext(); 4387 StructType *T; 4388 #define OMP_TYPE(VarName, InitValue) VarName = InitValue; 4389 #define OMP_ARRAY_TYPE(VarName, ElemTy, ArraySize) \ 4390 VarName##Ty = ArrayType::get(ElemTy, ArraySize); \ 4391 VarName##PtrTy = PointerType::getUnqual(VarName##Ty); 4392 #define OMP_FUNCTION_TYPE(VarName, IsVarArg, ReturnType, ...) \ 4393 VarName = FunctionType::get(ReturnType, {__VA_ARGS__}, IsVarArg); \ 4394 VarName##Ptr = PointerType::getUnqual(VarName); 4395 #define OMP_STRUCT_TYPE(VarName, StructName, ...) \ 4396 T = StructType::getTypeByName(Ctx, StructName); \ 4397 if (!T) \ 4398 T = StructType::create(Ctx, {__VA_ARGS__}, StructName); \ 4399 VarName = T; \ 4400 VarName##Ptr = PointerType::getUnqual(T); 4401 #include "llvm/Frontend/OpenMP/OMPKinds.def" 4402 } 4403 4404 void OpenMPIRBuilder::OutlineInfo::collectBlocks( 4405 SmallPtrSetImpl<BasicBlock *> &BlockSet, 4406 SmallVectorImpl<BasicBlock *> &BlockVector) { 4407 SmallVector<BasicBlock *, 32> Worklist; 4408 BlockSet.insert(EntryBB); 4409 BlockSet.insert(ExitBB); 4410 4411 Worklist.push_back(EntryBB); 4412 while (!Worklist.empty()) { 4413 BasicBlock *BB = Worklist.pop_back_val(); 4414 BlockVector.push_back(BB); 4415 for (BasicBlock *SuccBB : successors(BB)) 4416 if (BlockSet.insert(SuccBB).second) 4417 Worklist.push_back(SuccBB); 4418 } 4419 } 4420 4421 void CanonicalLoopInfo::collectControlBlocks( 4422 SmallVectorImpl<BasicBlock *> &BBs) { 4423 // We only count those BBs as control block for which we do not need to 4424 // reverse the CFG, i.e. not the loop body which can contain arbitrary control 4425 // flow. For consistency, this also means we do not add the Body block, which 4426 // is just the entry to the body code. 4427 BBs.reserve(BBs.size() + 6); 4428 BBs.append({getPreheader(), Header, Cond, Latch, Exit, getAfter()}); 4429 } 4430 4431 BasicBlock *CanonicalLoopInfo::getPreheader() const { 4432 assert(isValid() && "Requires a valid canonical loop"); 4433 for (BasicBlock *Pred : predecessors(Header)) { 4434 if (Pred != Latch) 4435 return Pred; 4436 } 4437 llvm_unreachable("Missing preheader"); 4438 } 4439 4440 void CanonicalLoopInfo::setTripCount(Value *TripCount) { 4441 assert(isValid() && "Requires a valid canonical loop"); 4442 4443 Instruction *CmpI = &getCond()->front(); 4444 assert(isa<CmpInst>(CmpI) && "First inst must compare IV with TripCount"); 4445 CmpI->setOperand(1, TripCount); 4446 4447 #ifndef NDEBUG 4448 assertOK(); 4449 #endif 4450 } 4451 4452 void CanonicalLoopInfo::mapIndVar( 4453 llvm::function_ref<Value *(Instruction *)> Updater) { 4454 assert(isValid() && "Requires a valid canonical loop"); 4455 4456 Instruction *OldIV = getIndVar(); 4457 4458 // Record all uses excluding those introduced by the updater. Uses by the 4459 // CanonicalLoopInfo itself to keep track of the number of iterations are 4460 // excluded. 4461 SmallVector<Use *> ReplacableUses; 4462 for (Use &U : OldIV->uses()) { 4463 auto *User = dyn_cast<Instruction>(U.getUser()); 4464 if (!User) 4465 continue; 4466 if (User->getParent() == getCond()) 4467 continue; 4468 if (User->getParent() == getLatch()) 4469 continue; 4470 ReplacableUses.push_back(&U); 4471 } 4472 4473 // Run the updater that may introduce new uses 4474 Value *NewIV = Updater(OldIV); 4475 4476 // Replace the old uses with the value returned by the updater. 4477 for (Use *U : ReplacableUses) 4478 U->set(NewIV); 4479 4480 #ifndef NDEBUG 4481 assertOK(); 4482 #endif 4483 } 4484 4485 void CanonicalLoopInfo::assertOK() const { 4486 #ifndef NDEBUG 4487 // No constraints if this object currently does not describe a loop. 4488 if (!isValid()) 4489 return; 4490 4491 BasicBlock *Preheader = getPreheader(); 4492 BasicBlock *Body = getBody(); 4493 BasicBlock *After = getAfter(); 4494 4495 // Verify standard control-flow we use for OpenMP loops. 4496 assert(Preheader); 4497 assert(isa<BranchInst>(Preheader->getTerminator()) && 4498 "Preheader must terminate with unconditional branch"); 4499 assert(Preheader->getSingleSuccessor() == Header && 4500 "Preheader must jump to header"); 4501 4502 assert(Header); 4503 assert(isa<BranchInst>(Header->getTerminator()) && 4504 "Header must terminate with unconditional branch"); 4505 assert(Header->getSingleSuccessor() == Cond && 4506 "Header must jump to exiting block"); 4507 4508 assert(Cond); 4509 assert(Cond->getSinglePredecessor() == Header && 4510 "Exiting block only reachable from header"); 4511 4512 assert(isa<BranchInst>(Cond->getTerminator()) && 4513 "Exiting block must terminate with conditional branch"); 4514 assert(size(successors(Cond)) == 2 && 4515 "Exiting block must have two successors"); 4516 assert(cast<BranchInst>(Cond->getTerminator())->getSuccessor(0) == Body && 4517 "Exiting block's first successor jump to the body"); 4518 assert(cast<BranchInst>(Cond->getTerminator())->getSuccessor(1) == Exit && 4519 "Exiting block's second successor must exit the loop"); 4520 4521 assert(Body); 4522 assert(Body->getSinglePredecessor() == Cond && 4523 "Body only reachable from exiting block"); 4524 assert(!isa<PHINode>(Body->front())); 4525 4526 assert(Latch); 4527 assert(isa<BranchInst>(Latch->getTerminator()) && 4528 "Latch must terminate with unconditional branch"); 4529 assert(Latch->getSingleSuccessor() == Header && "Latch must jump to header"); 4530 // TODO: To support simple redirecting of the end of the body code that has 4531 // multiple; introduce another auxiliary basic block like preheader and after. 4532 assert(Latch->getSinglePredecessor() != nullptr); 4533 assert(!isa<PHINode>(Latch->front())); 4534 4535 assert(Exit); 4536 assert(isa<BranchInst>(Exit->getTerminator()) && 4537 "Exit block must terminate with unconditional branch"); 4538 assert(Exit->getSingleSuccessor() == After && 4539 "Exit block must jump to after block"); 4540 4541 assert(After); 4542 assert(After->getSinglePredecessor() == Exit && 4543 "After block only reachable from exit block"); 4544 assert(After->empty() || !isa<PHINode>(After->front())); 4545 4546 Instruction *IndVar = getIndVar(); 4547 assert(IndVar && "Canonical induction variable not found?"); 4548 assert(isa<IntegerType>(IndVar->getType()) && 4549 "Induction variable must be an integer"); 4550 assert(cast<PHINode>(IndVar)->getParent() == Header && 4551 "Induction variable must be a PHI in the loop header"); 4552 assert(cast<PHINode>(IndVar)->getIncomingBlock(0) == Preheader); 4553 assert( 4554 cast<ConstantInt>(cast<PHINode>(IndVar)->getIncomingValue(0))->isZero()); 4555 assert(cast<PHINode>(IndVar)->getIncomingBlock(1) == Latch); 4556 4557 auto *NextIndVar = cast<PHINode>(IndVar)->getIncomingValue(1); 4558 assert(cast<Instruction>(NextIndVar)->getParent() == Latch); 4559 assert(cast<BinaryOperator>(NextIndVar)->getOpcode() == BinaryOperator::Add); 4560 assert(cast<BinaryOperator>(NextIndVar)->getOperand(0) == IndVar); 4561 assert(cast<ConstantInt>(cast<BinaryOperator>(NextIndVar)->getOperand(1)) 4562 ->isOne()); 4563 4564 Value *TripCount = getTripCount(); 4565 assert(TripCount && "Loop trip count not found?"); 4566 assert(IndVar->getType() == TripCount->getType() && 4567 "Trip count and induction variable must have the same type"); 4568 4569 auto *CmpI = cast<CmpInst>(&Cond->front()); 4570 assert(CmpI->getPredicate() == CmpInst::ICMP_ULT && 4571 "Exit condition must be a signed less-than comparison"); 4572 assert(CmpI->getOperand(0) == IndVar && 4573 "Exit condition must compare the induction variable"); 4574 assert(CmpI->getOperand(1) == TripCount && 4575 "Exit condition must compare with the trip count"); 4576 #endif 4577 } 4578 4579 void CanonicalLoopInfo::invalidate() { 4580 Header = nullptr; 4581 Cond = nullptr; 4582 Latch = nullptr; 4583 Exit = nullptr; 4584 } 4585