1 //===- LoopDistribute.cpp - Loop Distribution Pass ------------------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This file implements the Loop Distribution Pass. Its main focus is to 10 // distribute loops that cannot be vectorized due to dependence cycles. It 11 // tries to isolate the offending dependences into a new loop allowing 12 // vectorization of the remaining parts. 13 // 14 // For dependence analysis, the pass uses the LoopVectorizer's 15 // LoopAccessAnalysis. Because this analysis presumes no change in the order of 16 // memory operations, special care is taken to preserve the lexical order of 17 // these operations. 18 // 19 // Similarly to the Vectorizer, the pass also supports loop versioning to 20 // run-time disambiguate potentially overlapping arrays. 21 // 22 //===----------------------------------------------------------------------===// 23 24 #include "llvm/Transforms/Scalar/LoopDistribute.h" 25 #include "llvm/ADT/DenseMap.h" 26 #include "llvm/ADT/DepthFirstIterator.h" 27 #include "llvm/ADT/EquivalenceClasses.h" 28 #include "llvm/ADT/STLExtras.h" 29 #include "llvm/ADT/SmallPtrSet.h" 30 #include "llvm/ADT/SmallVector.h" 31 #include "llvm/ADT/Statistic.h" 32 #include "llvm/ADT/StringRef.h" 33 #include "llvm/ADT/Twine.h" 34 #include "llvm/ADT/iterator_range.h" 35 #include "llvm/Analysis/AssumptionCache.h" 36 #include "llvm/Analysis/GlobalsModRef.h" 37 #include "llvm/Analysis/LoopAccessAnalysis.h" 38 #include "llvm/Analysis/LoopAnalysisManager.h" 39 #include "llvm/Analysis/LoopInfo.h" 40 #include "llvm/Analysis/OptimizationRemarkEmitter.h" 41 #include "llvm/Analysis/ScalarEvolution.h" 42 #include "llvm/Analysis/TargetLibraryInfo.h" 43 #include "llvm/Analysis/TargetTransformInfo.h" 44 #include "llvm/IR/BasicBlock.h" 45 #include "llvm/IR/Constants.h" 46 #include "llvm/IR/DiagnosticInfo.h" 47 #include "llvm/IR/Dominators.h" 48 #include "llvm/IR/Function.h" 49 #include "llvm/IR/Instruction.h" 50 #include "llvm/IR/Instructions.h" 51 #include "llvm/IR/LLVMContext.h" 52 #include "llvm/IR/Metadata.h" 53 #include "llvm/IR/PassManager.h" 54 #include "llvm/IR/Value.h" 55 #include "llvm/Support/Casting.h" 56 #include "llvm/Support/CommandLine.h" 57 #include "llvm/Support/Debug.h" 58 #include "llvm/Support/raw_ostream.h" 59 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 60 #include "llvm/Transforms/Utils/Cloning.h" 61 #include "llvm/Transforms/Utils/LoopUtils.h" 62 #include "llvm/Transforms/Utils/LoopVersioning.h" 63 #include "llvm/Transforms/Utils/ValueMapper.h" 64 #include <cassert> 65 #include <functional> 66 #include <list> 67 #include <tuple> 68 #include <utility> 69 70 using namespace llvm; 71 72 #define LDIST_NAME "loop-distribute" 73 #define DEBUG_TYPE LDIST_NAME 74 75 /// @{ 76 /// Metadata attribute names 77 static const char *const LLVMLoopDistributeFollowupAll = 78 "llvm.loop.distribute.followup_all"; 79 static const char *const LLVMLoopDistributeFollowupCoincident = 80 "llvm.loop.distribute.followup_coincident"; 81 static const char *const LLVMLoopDistributeFollowupSequential = 82 "llvm.loop.distribute.followup_sequential"; 83 static const char *const LLVMLoopDistributeFollowupFallback = 84 "llvm.loop.distribute.followup_fallback"; 85 /// @} 86 87 static cl::opt<bool> 88 LDistVerify("loop-distribute-verify", cl::Hidden, 89 cl::desc("Turn on DominatorTree and LoopInfo verification " 90 "after Loop Distribution"), 91 cl::init(false)); 92 93 static cl::opt<bool> DistributeNonIfConvertible( 94 "loop-distribute-non-if-convertible", cl::Hidden, 95 cl::desc("Whether to distribute into a loop that may not be " 96 "if-convertible by the loop vectorizer"), 97 cl::init(false)); 98 99 static cl::opt<unsigned> DistributeSCEVCheckThreshold( 100 "loop-distribute-scev-check-threshold", cl::init(8), cl::Hidden, 101 cl::desc("The maximum number of SCEV checks allowed for Loop " 102 "Distribution")); 103 104 static cl::opt<unsigned> PragmaDistributeSCEVCheckThreshold( 105 "loop-distribute-scev-check-threshold-with-pragma", cl::init(128), 106 cl::Hidden, 107 cl::desc( 108 "The maximum number of SCEV checks allowed for Loop " 109 "Distribution for loop marked with #pragma loop distribute(enable)")); 110 111 static cl::opt<bool> EnableLoopDistribute( 112 "enable-loop-distribute", cl::Hidden, 113 cl::desc("Enable the new, experimental LoopDistribution Pass"), 114 cl::init(false)); 115 116 STATISTIC(NumLoopsDistributed, "Number of loops distributed"); 117 118 namespace { 119 120 /// Maintains the set of instructions of the loop for a partition before 121 /// cloning. After cloning, it hosts the new loop. 122 class InstPartition { 123 using InstructionSet = SmallPtrSet<Instruction *, 8>; 124 125 public: 126 InstPartition(Instruction *I, Loop *L, bool DepCycle = false) 127 : DepCycle(DepCycle), OrigLoop(L) { 128 Set.insert(I); 129 } 130 131 /// Returns whether this partition contains a dependence cycle. 132 bool hasDepCycle() const { return DepCycle; } 133 134 /// Adds an instruction to this partition. 135 void add(Instruction *I) { Set.insert(I); } 136 137 /// Collection accessors. 138 InstructionSet::iterator begin() { return Set.begin(); } 139 InstructionSet::iterator end() { return Set.end(); } 140 InstructionSet::const_iterator begin() const { return Set.begin(); } 141 InstructionSet::const_iterator end() const { return Set.end(); } 142 bool empty() const { return Set.empty(); } 143 144 /// Moves this partition into \p Other. This partition becomes empty 145 /// after this. 146 void moveTo(InstPartition &Other) { 147 Other.Set.insert(Set.begin(), Set.end()); 148 Set.clear(); 149 Other.DepCycle |= DepCycle; 150 } 151 152 /// Populates the partition with a transitive closure of all the 153 /// instructions that the seeded instructions dependent on. 154 void populateUsedSet() { 155 // FIXME: We currently don't use control-dependence but simply include all 156 // blocks (possibly empty at the end) and let simplifycfg mostly clean this 157 // up. 158 for (auto *B : OrigLoop->getBlocks()) 159 Set.insert(B->getTerminator()); 160 161 // Follow the use-def chains to form a transitive closure of all the 162 // instructions that the originally seeded instructions depend on. 163 SmallVector<Instruction *, 8> Worklist(Set.begin(), Set.end()); 164 while (!Worklist.empty()) { 165 Instruction *I = Worklist.pop_back_val(); 166 // Insert instructions from the loop that we depend on. 167 for (Value *V : I->operand_values()) { 168 auto *I = dyn_cast<Instruction>(V); 169 if (I && OrigLoop->contains(I->getParent()) && Set.insert(I).second) 170 Worklist.push_back(I); 171 } 172 } 173 } 174 175 /// Clones the original loop. 176 /// 177 /// Updates LoopInfo and DominatorTree using the information that block \p 178 /// LoopDomBB dominates the loop. 179 Loop *cloneLoopWithPreheader(BasicBlock *InsertBefore, BasicBlock *LoopDomBB, 180 unsigned Index, LoopInfo *LI, 181 DominatorTree *DT) { 182 ClonedLoop = ::cloneLoopWithPreheader(InsertBefore, LoopDomBB, OrigLoop, 183 VMap, Twine(".ldist") + Twine(Index), 184 LI, DT, ClonedLoopBlocks); 185 return ClonedLoop; 186 } 187 188 /// The cloned loop. If this partition is mapped to the original loop, 189 /// this is null. 190 const Loop *getClonedLoop() const { return ClonedLoop; } 191 192 /// Returns the loop where this partition ends up after distribution. 193 /// If this partition is mapped to the original loop then use the block from 194 /// the loop. 195 Loop *getDistributedLoop() const { 196 return ClonedLoop ? ClonedLoop : OrigLoop; 197 } 198 199 /// The VMap that is populated by cloning and then used in 200 /// remapinstruction to remap the cloned instructions. 201 ValueToValueMapTy &getVMap() { return VMap; } 202 203 /// Remaps the cloned instructions using VMap. 204 void remapInstructions() { 205 remapInstructionsInBlocks(ClonedLoopBlocks, VMap); 206 } 207 208 /// Based on the set of instructions selected for this partition, 209 /// removes the unnecessary ones. 210 void removeUnusedInsts() { 211 SmallVector<Instruction *, 8> Unused; 212 213 for (auto *Block : OrigLoop->getBlocks()) 214 for (auto &Inst : *Block) 215 if (!Set.count(&Inst)) { 216 Instruction *NewInst = &Inst; 217 if (!VMap.empty()) 218 NewInst = cast<Instruction>(VMap[NewInst]); 219 220 assert(!isa<BranchInst>(NewInst) && 221 "Branches are marked used early on"); 222 Unused.push_back(NewInst); 223 } 224 225 // Delete the instructions backwards, as it has a reduced likelihood of 226 // having to update as many def-use and use-def chains. 227 for (auto *Inst : reverse(Unused)) { 228 if (!Inst->use_empty()) 229 Inst->replaceAllUsesWith(PoisonValue::get(Inst->getType())); 230 Inst->eraseFromParent(); 231 } 232 } 233 234 void print() const { 235 if (DepCycle) 236 dbgs() << " (cycle)\n"; 237 for (auto *I : Set) 238 // Prefix with the block name. 239 dbgs() << " " << I->getParent()->getName() << ":" << *I << "\n"; 240 } 241 242 void printBlocks() const { 243 for (auto *BB : getDistributedLoop()->getBlocks()) 244 dbgs() << *BB; 245 } 246 247 private: 248 /// Instructions from OrigLoop selected for this partition. 249 InstructionSet Set; 250 251 /// Whether this partition contains a dependence cycle. 252 bool DepCycle; 253 254 /// The original loop. 255 Loop *OrigLoop; 256 257 /// The cloned loop. If this partition is mapped to the original loop, 258 /// this is null. 259 Loop *ClonedLoop = nullptr; 260 261 /// The blocks of ClonedLoop including the preheader. If this 262 /// partition is mapped to the original loop, this is empty. 263 SmallVector<BasicBlock *, 8> ClonedLoopBlocks; 264 265 /// These gets populated once the set of instructions have been 266 /// finalized. If this partition is mapped to the original loop, these are not 267 /// set. 268 ValueToValueMapTy VMap; 269 }; 270 271 /// Holds the set of Partitions. It populates them, merges them and then 272 /// clones the loops. 273 class InstPartitionContainer { 274 using InstToPartitionIdT = DenseMap<Instruction *, int>; 275 276 public: 277 InstPartitionContainer(Loop *L, LoopInfo *LI, DominatorTree *DT) 278 : L(L), LI(LI), DT(DT) {} 279 280 /// Returns the number of partitions. 281 unsigned getSize() const { return PartitionContainer.size(); } 282 283 /// Adds \p Inst into the current partition if that is marked to 284 /// contain cycles. Otherwise start a new partition for it. 285 void addToCyclicPartition(Instruction *Inst) { 286 // If the current partition is non-cyclic. Start a new one. 287 if (PartitionContainer.empty() || !PartitionContainer.back().hasDepCycle()) 288 PartitionContainer.emplace_back(Inst, L, /*DepCycle=*/true); 289 else 290 PartitionContainer.back().add(Inst); 291 } 292 293 /// Adds \p Inst into a partition that is not marked to contain 294 /// dependence cycles. 295 /// 296 // Initially we isolate memory instructions into as many partitions as 297 // possible, then later we may merge them back together. 298 void addToNewNonCyclicPartition(Instruction *Inst) { 299 PartitionContainer.emplace_back(Inst, L); 300 } 301 302 /// Merges adjacent non-cyclic partitions. 303 /// 304 /// The idea is that we currently only want to isolate the non-vectorizable 305 /// partition. We could later allow more distribution among these partition 306 /// too. 307 void mergeAdjacentNonCyclic() { 308 mergeAdjacentPartitionsIf( 309 [](const InstPartition *P) { return !P->hasDepCycle(); }); 310 } 311 312 /// If a partition contains only conditional stores, we won't vectorize 313 /// it. Try to merge it with a previous cyclic partition. 314 void mergeNonIfConvertible() { 315 mergeAdjacentPartitionsIf([&](const InstPartition *Partition) { 316 if (Partition->hasDepCycle()) 317 return true; 318 319 // Now, check if all stores are conditional in this partition. 320 bool seenStore = false; 321 322 for (auto *Inst : *Partition) 323 if (isa<StoreInst>(Inst)) { 324 seenStore = true; 325 if (!LoopAccessInfo::blockNeedsPredication(Inst->getParent(), L, DT)) 326 return false; 327 } 328 return seenStore; 329 }); 330 } 331 332 /// Merges the partitions according to various heuristics. 333 void mergeBeforePopulating() { 334 mergeAdjacentNonCyclic(); 335 if (!DistributeNonIfConvertible) 336 mergeNonIfConvertible(); 337 } 338 339 /// Merges partitions in order to ensure that no loads are duplicated. 340 /// 341 /// We can't duplicate loads because that could potentially reorder them. 342 /// LoopAccessAnalysis provides dependency information with the context that 343 /// the order of memory operation is preserved. 344 /// 345 /// Return if any partitions were merged. 346 bool mergeToAvoidDuplicatedLoads() { 347 using LoadToPartitionT = DenseMap<Instruction *, InstPartition *>; 348 using ToBeMergedT = EquivalenceClasses<InstPartition *>; 349 350 LoadToPartitionT LoadToPartition; 351 ToBeMergedT ToBeMerged; 352 353 // Step through the partitions and create equivalence between partitions 354 // that contain the same load. Also put partitions in between them in the 355 // same equivalence class to avoid reordering of memory operations. 356 for (PartitionContainerT::iterator I = PartitionContainer.begin(), 357 E = PartitionContainer.end(); 358 I != E; ++I) { 359 auto *PartI = &*I; 360 361 // If a load occurs in two partitions PartI and PartJ, merge all 362 // partitions (PartI, PartJ] into PartI. 363 for (Instruction *Inst : *PartI) 364 if (isa<LoadInst>(Inst)) { 365 bool NewElt; 366 LoadToPartitionT::iterator LoadToPart; 367 368 std::tie(LoadToPart, NewElt) = 369 LoadToPartition.insert(std::make_pair(Inst, PartI)); 370 if (!NewElt) { 371 LLVM_DEBUG(dbgs() 372 << "Merging partitions due to this load in multiple " 373 << "partitions: " << PartI << ", " << LoadToPart->second 374 << "\n" 375 << *Inst << "\n"); 376 377 auto PartJ = I; 378 do { 379 --PartJ; 380 ToBeMerged.unionSets(PartI, &*PartJ); 381 } while (&*PartJ != LoadToPart->second); 382 } 383 } 384 } 385 if (ToBeMerged.empty()) 386 return false; 387 388 // Merge the member of an equivalence class into its class leader. This 389 // makes the members empty. 390 for (ToBeMergedT::iterator I = ToBeMerged.begin(), E = ToBeMerged.end(); 391 I != E; ++I) { 392 if (!I->isLeader()) 393 continue; 394 395 auto PartI = I->getData(); 396 for (auto *PartJ : make_range(std::next(ToBeMerged.member_begin(I)), 397 ToBeMerged.member_end())) { 398 PartJ->moveTo(*PartI); 399 } 400 } 401 402 // Remove the empty partitions. 403 PartitionContainer.remove_if( 404 [](const InstPartition &P) { return P.empty(); }); 405 406 return true; 407 } 408 409 /// Sets up the mapping between instructions to partitions. If the 410 /// instruction is duplicated across multiple partitions, set the entry to -1. 411 void setupPartitionIdOnInstructions() { 412 int PartitionID = 0; 413 for (const auto &Partition : PartitionContainer) { 414 for (Instruction *Inst : Partition) { 415 bool NewElt; 416 InstToPartitionIdT::iterator Iter; 417 418 std::tie(Iter, NewElt) = 419 InstToPartitionId.insert(std::make_pair(Inst, PartitionID)); 420 if (!NewElt) 421 Iter->second = -1; 422 } 423 ++PartitionID; 424 } 425 } 426 427 /// Populates the partition with everything that the seeding 428 /// instructions require. 429 void populateUsedSet() { 430 for (auto &P : PartitionContainer) 431 P.populateUsedSet(); 432 } 433 434 /// This performs the main chunk of the work of cloning the loops for 435 /// the partitions. 436 void cloneLoops() { 437 BasicBlock *OrigPH = L->getLoopPreheader(); 438 // At this point the predecessor of the preheader is either the memcheck 439 // block or the top part of the original preheader. 440 BasicBlock *Pred = OrigPH->getSinglePredecessor(); 441 assert(Pred && "Preheader does not have a single predecessor"); 442 BasicBlock *ExitBlock = L->getExitBlock(); 443 assert(ExitBlock && "No single exit block"); 444 Loop *NewLoop; 445 446 assert(!PartitionContainer.empty() && "at least two partitions expected"); 447 // We're cloning the preheader along with the loop so we already made sure 448 // it was empty. 449 assert(&*OrigPH->begin() == OrigPH->getTerminator() && 450 "preheader not empty"); 451 452 // Preserve the original loop ID for use after the transformation. 453 MDNode *OrigLoopID = L->getLoopID(); 454 455 // Create a loop for each partition except the last. Clone the original 456 // loop before PH along with adding a preheader for the cloned loop. Then 457 // update PH to point to the newly added preheader. 458 BasicBlock *TopPH = OrigPH; 459 unsigned Index = getSize() - 1; 460 for (auto &Part : llvm::drop_begin(llvm::reverse(PartitionContainer))) { 461 NewLoop = Part.cloneLoopWithPreheader(TopPH, Pred, Index, LI, DT); 462 463 Part.getVMap()[ExitBlock] = TopPH; 464 Part.remapInstructions(); 465 setNewLoopID(OrigLoopID, &Part); 466 --Index; 467 TopPH = NewLoop->getLoopPreheader(); 468 } 469 Pred->getTerminator()->replaceUsesOfWith(OrigPH, TopPH); 470 471 // Also set a new loop ID for the last loop. 472 setNewLoopID(OrigLoopID, &PartitionContainer.back()); 473 474 // Now go in forward order and update the immediate dominator for the 475 // preheaders with the exiting block of the previous loop. Dominance 476 // within the loop is updated in cloneLoopWithPreheader. 477 for (auto Curr = PartitionContainer.cbegin(), 478 Next = std::next(PartitionContainer.cbegin()), 479 E = PartitionContainer.cend(); 480 Next != E; ++Curr, ++Next) 481 DT->changeImmediateDominator( 482 Next->getDistributedLoop()->getLoopPreheader(), 483 Curr->getDistributedLoop()->getExitingBlock()); 484 } 485 486 /// Removes the dead instructions from the cloned loops. 487 void removeUnusedInsts() { 488 for (auto &Partition : PartitionContainer) 489 Partition.removeUnusedInsts(); 490 } 491 492 /// For each memory pointer, it computes the partitionId the pointer is 493 /// used in. 494 /// 495 /// This returns an array of int where the I-th entry corresponds to I-th 496 /// entry in LAI.getRuntimePointerCheck(). If the pointer is used in multiple 497 /// partitions its entry is set to -1. 498 SmallVector<int, 8> 499 computePartitionSetForPointers(const LoopAccessInfo &LAI) { 500 const RuntimePointerChecking *RtPtrCheck = LAI.getRuntimePointerChecking(); 501 502 unsigned N = RtPtrCheck->Pointers.size(); 503 SmallVector<int, 8> PtrToPartitions(N); 504 for (unsigned I = 0; I < N; ++I) { 505 Value *Ptr = RtPtrCheck->Pointers[I].PointerValue; 506 auto Instructions = 507 LAI.getInstructionsForAccess(Ptr, RtPtrCheck->Pointers[I].IsWritePtr); 508 509 int &Partition = PtrToPartitions[I]; 510 // First set it to uninitialized. 511 Partition = -2; 512 for (Instruction *Inst : Instructions) { 513 // Note that this could be -1 if Inst is duplicated across multiple 514 // partitions. 515 int ThisPartition = this->InstToPartitionId[Inst]; 516 if (Partition == -2) 517 Partition = ThisPartition; 518 // -1 means belonging to multiple partitions. 519 else if (Partition == -1) 520 break; 521 else if (Partition != (int)ThisPartition) 522 Partition = -1; 523 } 524 assert(Partition != -2 && "Pointer not belonging to any partition"); 525 } 526 527 return PtrToPartitions; 528 } 529 530 void print(raw_ostream &OS) const { 531 unsigned Index = 0; 532 for (const auto &P : PartitionContainer) { 533 OS << "Partition " << Index++ << " (" << &P << "):\n"; 534 P.print(); 535 } 536 } 537 538 void dump() const { print(dbgs()); } 539 540 #ifndef NDEBUG 541 friend raw_ostream &operator<<(raw_ostream &OS, 542 const InstPartitionContainer &Partitions) { 543 Partitions.print(OS); 544 return OS; 545 } 546 #endif 547 548 void printBlocks() const { 549 unsigned Index = 0; 550 for (const auto &P : PartitionContainer) { 551 dbgs() << "\nPartition " << Index++ << " (" << &P << "):\n"; 552 P.printBlocks(); 553 } 554 } 555 556 private: 557 using PartitionContainerT = std::list<InstPartition>; 558 559 /// List of partitions. 560 PartitionContainerT PartitionContainer; 561 562 /// Mapping from Instruction to partition Id. If the instruction 563 /// belongs to multiple partitions the entry contains -1. 564 InstToPartitionIdT InstToPartitionId; 565 566 Loop *L; 567 LoopInfo *LI; 568 DominatorTree *DT; 569 570 /// The control structure to merge adjacent partitions if both satisfy 571 /// the \p Predicate. 572 template <class UnaryPredicate> 573 void mergeAdjacentPartitionsIf(UnaryPredicate Predicate) { 574 InstPartition *PrevMatch = nullptr; 575 for (auto I = PartitionContainer.begin(); I != PartitionContainer.end();) { 576 auto DoesMatch = Predicate(&*I); 577 if (PrevMatch == nullptr && DoesMatch) { 578 PrevMatch = &*I; 579 ++I; 580 } else if (PrevMatch != nullptr && DoesMatch) { 581 I->moveTo(*PrevMatch); 582 I = PartitionContainer.erase(I); 583 } else { 584 PrevMatch = nullptr; 585 ++I; 586 } 587 } 588 } 589 590 /// Assign new LoopIDs for the partition's cloned loop. 591 void setNewLoopID(MDNode *OrigLoopID, InstPartition *Part) { 592 std::optional<MDNode *> PartitionID = makeFollowupLoopID( 593 OrigLoopID, 594 {LLVMLoopDistributeFollowupAll, 595 Part->hasDepCycle() ? LLVMLoopDistributeFollowupSequential 596 : LLVMLoopDistributeFollowupCoincident}); 597 if (PartitionID) { 598 Loop *NewLoop = Part->getDistributedLoop(); 599 NewLoop->setLoopID(*PartitionID); 600 } 601 } 602 }; 603 604 /// For each memory instruction, this class maintains difference of the 605 /// number of unsafe dependences that start out from this instruction minus 606 /// those that end here. 607 /// 608 /// By traversing the memory instructions in program order and accumulating this 609 /// number, we know whether any unsafe dependence crosses over a program point. 610 class MemoryInstructionDependences { 611 using Dependence = MemoryDepChecker::Dependence; 612 613 public: 614 struct Entry { 615 Instruction *Inst; 616 unsigned NumUnsafeDependencesStartOrEnd = 0; 617 618 Entry(Instruction *Inst) : Inst(Inst) {} 619 }; 620 621 using AccessesType = SmallVector<Entry, 8>; 622 623 AccessesType::const_iterator begin() const { return Accesses.begin(); } 624 AccessesType::const_iterator end() const { return Accesses.end(); } 625 626 MemoryInstructionDependences( 627 const SmallVectorImpl<Instruction *> &Instructions, 628 const SmallVectorImpl<Dependence> &Dependences) { 629 Accesses.append(Instructions.begin(), Instructions.end()); 630 631 LLVM_DEBUG(dbgs() << "Backward dependences:\n"); 632 for (const auto &Dep : Dependences) 633 if (Dep.isPossiblyBackward()) { 634 // Note that the designations source and destination follow the program 635 // order, i.e. source is always first. (The direction is given by the 636 // DepType.) 637 ++Accesses[Dep.Source].NumUnsafeDependencesStartOrEnd; 638 --Accesses[Dep.Destination].NumUnsafeDependencesStartOrEnd; 639 640 LLVM_DEBUG(Dep.print(dbgs(), 2, Instructions)); 641 } 642 } 643 644 private: 645 AccessesType Accesses; 646 }; 647 648 /// The actual class performing the per-loop work. 649 class LoopDistributeForLoop { 650 public: 651 LoopDistributeForLoop(Loop *L, Function *F, LoopInfo *LI, DominatorTree *DT, 652 ScalarEvolution *SE, LoopAccessInfoManager &LAIs, 653 OptimizationRemarkEmitter *ORE) 654 : L(L), F(F), LI(LI), DT(DT), SE(SE), LAIs(LAIs), ORE(ORE) { 655 setForced(); 656 } 657 658 /// Try to distribute an inner-most loop. 659 bool processLoop() { 660 assert(L->isInnermost() && "Only process inner loops."); 661 662 LLVM_DEBUG(dbgs() << "\nLDist: In \"" 663 << L->getHeader()->getParent()->getName() 664 << "\" checking " << *L << "\n"); 665 666 // Having a single exit block implies there's also one exiting block. 667 if (!L->getExitBlock()) 668 return fail("MultipleExitBlocks", "multiple exit blocks"); 669 if (!L->isLoopSimplifyForm()) 670 return fail("NotLoopSimplifyForm", 671 "loop is not in loop-simplify form"); 672 if (!L->isRotatedForm()) 673 return fail("NotBottomTested", "loop is not bottom tested"); 674 675 BasicBlock *PH = L->getLoopPreheader(); 676 677 LAI = &LAIs.getInfo(*L); 678 679 // Currently, we only distribute to isolate the part of the loop with 680 // dependence cycles to enable partial vectorization. 681 if (LAI->canVectorizeMemory()) 682 return fail("MemOpsCanBeVectorized", 683 "memory operations are safe for vectorization"); 684 685 auto *Dependences = LAI->getDepChecker().getDependences(); 686 if (!Dependences || Dependences->empty()) 687 return fail("NoUnsafeDeps", "no unsafe dependences to isolate"); 688 689 InstPartitionContainer Partitions(L, LI, DT); 690 691 // First, go through each memory operation and assign them to consecutive 692 // partitions (the order of partitions follows program order). Put those 693 // with unsafe dependences into "cyclic" partition otherwise put each store 694 // in its own "non-cyclic" partition (we'll merge these later). 695 // 696 // Note that a memory operation (e.g. Load2 below) at a program point that 697 // has an unsafe dependence (Store3->Load1) spanning over it must be 698 // included in the same cyclic partition as the dependent operations. This 699 // is to preserve the original program order after distribution. E.g.: 700 // 701 // NumUnsafeDependencesStartOrEnd NumUnsafeDependencesActive 702 // Load1 -. 1 0->1 703 // Load2 | /Unsafe/ 0 1 704 // Store3 -' -1 1->0 705 // Load4 0 0 706 // 707 // NumUnsafeDependencesActive > 0 indicates this situation and in this case 708 // we just keep assigning to the same cyclic partition until 709 // NumUnsafeDependencesActive reaches 0. 710 const MemoryDepChecker &DepChecker = LAI->getDepChecker(); 711 MemoryInstructionDependences MID(DepChecker.getMemoryInstructions(), 712 *Dependences); 713 714 int NumUnsafeDependencesActive = 0; 715 for (const auto &InstDep : MID) { 716 Instruction *I = InstDep.Inst; 717 // We update NumUnsafeDependencesActive post-instruction, catch the 718 // start of a dependence directly via NumUnsafeDependencesStartOrEnd. 719 if (NumUnsafeDependencesActive || 720 InstDep.NumUnsafeDependencesStartOrEnd > 0) 721 Partitions.addToCyclicPartition(I); 722 else 723 Partitions.addToNewNonCyclicPartition(I); 724 NumUnsafeDependencesActive += InstDep.NumUnsafeDependencesStartOrEnd; 725 assert(NumUnsafeDependencesActive >= 0 && 726 "Negative number of dependences active"); 727 } 728 729 // Add partitions for values used outside. These partitions can be out of 730 // order from the original program order. This is OK because if the 731 // partition uses a load we will merge this partition with the original 732 // partition of the load that we set up in the previous loop (see 733 // mergeToAvoidDuplicatedLoads). 734 auto DefsUsedOutside = findDefsUsedOutsideOfLoop(L); 735 for (auto *Inst : DefsUsedOutside) 736 Partitions.addToNewNonCyclicPartition(Inst); 737 738 LLVM_DEBUG(dbgs() << "Seeded partitions:\n" << Partitions); 739 if (Partitions.getSize() < 2) 740 return fail("CantIsolateUnsafeDeps", 741 "cannot isolate unsafe dependencies"); 742 743 // Run the merge heuristics: Merge non-cyclic adjacent partitions since we 744 // should be able to vectorize these together. 745 Partitions.mergeBeforePopulating(); 746 LLVM_DEBUG(dbgs() << "\nMerged partitions:\n" << Partitions); 747 if (Partitions.getSize() < 2) 748 return fail("CantIsolateUnsafeDeps", 749 "cannot isolate unsafe dependencies"); 750 751 // Now, populate the partitions with non-memory operations. 752 Partitions.populateUsedSet(); 753 LLVM_DEBUG(dbgs() << "\nPopulated partitions:\n" << Partitions); 754 755 // In order to preserve original lexical order for loads, keep them in the 756 // partition that we set up in the MemoryInstructionDependences loop. 757 if (Partitions.mergeToAvoidDuplicatedLoads()) { 758 LLVM_DEBUG(dbgs() << "\nPartitions merged to ensure unique loads:\n" 759 << Partitions); 760 if (Partitions.getSize() < 2) 761 return fail("CantIsolateUnsafeDeps", 762 "cannot isolate unsafe dependencies"); 763 } 764 765 // Don't distribute the loop if we need too many SCEV run-time checks, or 766 // any if it's illegal. 767 const SCEVPredicate &Pred = LAI->getPSE().getPredicate(); 768 if (LAI->hasConvergentOp() && !Pred.isAlwaysTrue()) { 769 return fail("RuntimeCheckWithConvergent", 770 "may not insert runtime check with convergent operation"); 771 } 772 773 if (Pred.getComplexity() > (IsForced.value_or(false) 774 ? PragmaDistributeSCEVCheckThreshold 775 : DistributeSCEVCheckThreshold)) 776 return fail("TooManySCEVRuntimeChecks", 777 "too many SCEV run-time checks needed.\n"); 778 779 if (!IsForced.value_or(false) && hasDisableAllTransformsHint(L)) 780 return fail("HeuristicDisabled", "distribution heuristic disabled"); 781 782 LLVM_DEBUG(dbgs() << "\nDistributing loop: " << *L << "\n"); 783 // We're done forming the partitions set up the reverse mapping from 784 // instructions to partitions. 785 Partitions.setupPartitionIdOnInstructions(); 786 787 // If we need run-time checks, version the loop now. 788 auto PtrToPartition = Partitions.computePartitionSetForPointers(*LAI); 789 const auto *RtPtrChecking = LAI->getRuntimePointerChecking(); 790 const auto &AllChecks = RtPtrChecking->getChecks(); 791 auto Checks = includeOnlyCrossPartitionChecks(AllChecks, PtrToPartition, 792 RtPtrChecking); 793 794 if (LAI->hasConvergentOp() && !Checks.empty()) { 795 return fail("RuntimeCheckWithConvergent", 796 "may not insert runtime check with convergent operation"); 797 } 798 799 // To keep things simple have an empty preheader before we version or clone 800 // the loop. (Also split if this has no predecessor, i.e. entry, because we 801 // rely on PH having a predecessor.) 802 if (!PH->getSinglePredecessor() || &*PH->begin() != PH->getTerminator()) 803 SplitBlock(PH, PH->getTerminator(), DT, LI); 804 805 if (!Pred.isAlwaysTrue() || !Checks.empty()) { 806 assert(!LAI->hasConvergentOp() && "inserting illegal loop versioning"); 807 808 MDNode *OrigLoopID = L->getLoopID(); 809 810 LLVM_DEBUG(dbgs() << "\nPointers:\n"); 811 LLVM_DEBUG(LAI->getRuntimePointerChecking()->printChecks(dbgs(), Checks)); 812 LoopVersioning LVer(*LAI, Checks, L, LI, DT, SE); 813 LVer.versionLoop(DefsUsedOutside); 814 LVer.annotateLoopWithNoAlias(); 815 816 // The unversioned loop will not be changed, so we inherit all attributes 817 // from the original loop, but remove the loop distribution metadata to 818 // avoid to distribute it again. 819 MDNode *UnversionedLoopID = *makeFollowupLoopID( 820 OrigLoopID, 821 {LLVMLoopDistributeFollowupAll, LLVMLoopDistributeFollowupFallback}, 822 "llvm.loop.distribute.", true); 823 LVer.getNonVersionedLoop()->setLoopID(UnversionedLoopID); 824 } 825 826 // Create identical copies of the original loop for each partition and hook 827 // them up sequentially. 828 Partitions.cloneLoops(); 829 830 // Now, we remove the instruction from each loop that don't belong to that 831 // partition. 832 Partitions.removeUnusedInsts(); 833 LLVM_DEBUG(dbgs() << "\nAfter removing unused Instrs:\n"); 834 LLVM_DEBUG(Partitions.printBlocks()); 835 836 if (LDistVerify) { 837 LI->verify(*DT); 838 assert(DT->verify(DominatorTree::VerificationLevel::Fast)); 839 } 840 841 ++NumLoopsDistributed; 842 // Report the success. 843 ORE->emit([&]() { 844 return OptimizationRemark(LDIST_NAME, "Distribute", L->getStartLoc(), 845 L->getHeader()) 846 << "distributed loop"; 847 }); 848 return true; 849 } 850 851 /// Provide diagnostics then \return with false. 852 bool fail(StringRef RemarkName, StringRef Message) { 853 LLVMContext &Ctx = F->getContext(); 854 bool Forced = isForced().value_or(false); 855 856 LLVM_DEBUG(dbgs() << "Skipping; " << Message << "\n"); 857 858 // With Rpass-missed report that distribution failed. 859 ORE->emit([&]() { 860 return OptimizationRemarkMissed(LDIST_NAME, "NotDistributed", 861 L->getStartLoc(), L->getHeader()) 862 << "loop not distributed: use -Rpass-analysis=loop-distribute for " 863 "more " 864 "info"; 865 }); 866 867 // With Rpass-analysis report why. This is on by default if distribution 868 // was requested explicitly. 869 ORE->emit(OptimizationRemarkAnalysis( 870 Forced ? OptimizationRemarkAnalysis::AlwaysPrint : LDIST_NAME, 871 RemarkName, L->getStartLoc(), L->getHeader()) 872 << "loop not distributed: " << Message); 873 874 // Also issue a warning if distribution was requested explicitly but it 875 // failed. 876 if (Forced) 877 Ctx.diagnose(DiagnosticInfoOptimizationFailure( 878 *F, L->getStartLoc(), "loop not distributed: failed " 879 "explicitly specified loop distribution")); 880 881 return false; 882 } 883 884 /// Return if distribution forced to be enabled/disabled for the loop. 885 /// 886 /// If the optional has a value, it indicates whether distribution was forced 887 /// to be enabled (true) or disabled (false). If the optional has no value 888 /// distribution was not forced either way. 889 const std::optional<bool> &isForced() const { return IsForced; } 890 891 private: 892 /// Filter out checks between pointers from the same partition. 893 /// 894 /// \p PtrToPartition contains the partition number for pointers. Partition 895 /// number -1 means that the pointer is used in multiple partitions. In this 896 /// case we can't safely omit the check. 897 SmallVector<RuntimePointerCheck, 4> includeOnlyCrossPartitionChecks( 898 const SmallVectorImpl<RuntimePointerCheck> &AllChecks, 899 const SmallVectorImpl<int> &PtrToPartition, 900 const RuntimePointerChecking *RtPtrChecking) { 901 SmallVector<RuntimePointerCheck, 4> Checks; 902 903 copy_if(AllChecks, std::back_inserter(Checks), 904 [&](const RuntimePointerCheck &Check) { 905 for (unsigned PtrIdx1 : Check.first->Members) 906 for (unsigned PtrIdx2 : Check.second->Members) 907 // Only include this check if there is a pair of pointers 908 // that require checking and the pointers fall into 909 // separate partitions. 910 // 911 // (Note that we already know at this point that the two 912 // pointer groups need checking but it doesn't follow 913 // that each pair of pointers within the two groups need 914 // checking as well. 915 // 916 // In other words we don't want to include a check just 917 // because there is a pair of pointers between the two 918 // pointer groups that require checks and a different 919 // pair whose pointers fall into different partitions.) 920 if (RtPtrChecking->needsChecking(PtrIdx1, PtrIdx2) && 921 !RuntimePointerChecking::arePointersInSamePartition( 922 PtrToPartition, PtrIdx1, PtrIdx2)) 923 return true; 924 return false; 925 }); 926 927 return Checks; 928 } 929 930 /// Check whether the loop metadata is forcing distribution to be 931 /// enabled/disabled. 932 void setForced() { 933 std::optional<const MDOperand *> Value = 934 findStringMetadataForLoop(L, "llvm.loop.distribute.enable"); 935 if (!Value) 936 return; 937 938 const MDOperand *Op = *Value; 939 assert(Op && mdconst::hasa<ConstantInt>(*Op) && "invalid metadata"); 940 IsForced = mdconst::extract<ConstantInt>(*Op)->getZExtValue(); 941 } 942 943 Loop *L; 944 Function *F; 945 946 // Analyses used. 947 LoopInfo *LI; 948 const LoopAccessInfo *LAI = nullptr; 949 DominatorTree *DT; 950 ScalarEvolution *SE; 951 LoopAccessInfoManager &LAIs; 952 OptimizationRemarkEmitter *ORE; 953 954 /// Indicates whether distribution is forced to be enabled/disabled for 955 /// the loop. 956 /// 957 /// If the optional has a value, it indicates whether distribution was forced 958 /// to be enabled (true) or disabled (false). If the optional has no value 959 /// distribution was not forced either way. 960 std::optional<bool> IsForced; 961 }; 962 963 } // end anonymous namespace 964 965 /// Shared implementation between new and old PMs. 966 static bool runImpl(Function &F, LoopInfo *LI, DominatorTree *DT, 967 ScalarEvolution *SE, OptimizationRemarkEmitter *ORE, 968 LoopAccessInfoManager &LAIs) { 969 // Build up a worklist of inner-loops to vectorize. This is necessary as the 970 // act of distributing a loop creates new loops and can invalidate iterators 971 // across the loops. 972 SmallVector<Loop *, 8> Worklist; 973 974 for (Loop *TopLevelLoop : *LI) 975 for (Loop *L : depth_first(TopLevelLoop)) 976 // We only handle inner-most loops. 977 if (L->isInnermost()) 978 Worklist.push_back(L); 979 980 // Now walk the identified inner loops. 981 bool Changed = false; 982 for (Loop *L : Worklist) { 983 LoopDistributeForLoop LDL(L, &F, LI, DT, SE, LAIs, ORE); 984 985 // If distribution was forced for the specific loop to be 986 // enabled/disabled, follow that. Otherwise use the global flag. 987 if (LDL.isForced().value_or(EnableLoopDistribute)) 988 Changed |= LDL.processLoop(); 989 } 990 991 // Process each loop nest in the function. 992 return Changed; 993 } 994 995 PreservedAnalyses LoopDistributePass::run(Function &F, 996 FunctionAnalysisManager &AM) { 997 auto &LI = AM.getResult<LoopAnalysis>(F); 998 auto &DT = AM.getResult<DominatorTreeAnalysis>(F); 999 auto &SE = AM.getResult<ScalarEvolutionAnalysis>(F); 1000 auto &ORE = AM.getResult<OptimizationRemarkEmitterAnalysis>(F); 1001 1002 LoopAccessInfoManager &LAIs = AM.getResult<LoopAccessAnalysis>(F); 1003 bool Changed = runImpl(F, &LI, &DT, &SE, &ORE, LAIs); 1004 if (!Changed) 1005 return PreservedAnalyses::all(); 1006 PreservedAnalyses PA; 1007 PA.preserve<LoopAnalysis>(); 1008 PA.preserve<DominatorTreeAnalysis>(); 1009 return PA; 1010 } 1011