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/Optional.h" 29 #include "llvm/ADT/STLExtras.h" 30 #include "llvm/ADT/SmallPtrSet.h" 31 #include "llvm/ADT/SmallVector.h" 32 #include "llvm/ADT/Statistic.h" 33 #include "llvm/ADT/StringRef.h" 34 #include "llvm/ADT/Twine.h" 35 #include "llvm/ADT/iterator_range.h" 36 #include "llvm/Analysis/AssumptionCache.h" 37 #include "llvm/Analysis/GlobalsModRef.h" 38 #include "llvm/Analysis/LoopAccessAnalysis.h" 39 #include "llvm/Analysis/LoopAnalysisManager.h" 40 #include "llvm/Analysis/LoopInfo.h" 41 #include "llvm/Analysis/OptimizationRemarkEmitter.h" 42 #include "llvm/Analysis/ScalarEvolution.h" 43 #include "llvm/Analysis/TargetLibraryInfo.h" 44 #include "llvm/Analysis/TargetTransformInfo.h" 45 #include "llvm/IR/BasicBlock.h" 46 #include "llvm/IR/Constants.h" 47 #include "llvm/IR/DiagnosticInfo.h" 48 #include "llvm/IR/Dominators.h" 49 #include "llvm/IR/Function.h" 50 #include "llvm/IR/Instruction.h" 51 #include "llvm/IR/Instructions.h" 52 #include "llvm/IR/LLVMContext.h" 53 #include "llvm/IR/Metadata.h" 54 #include "llvm/IR/PassManager.h" 55 #include "llvm/IR/Value.h" 56 #include "llvm/InitializePasses.h" 57 #include "llvm/Pass.h" 58 #include "llvm/Support/Casting.h" 59 #include "llvm/Support/CommandLine.h" 60 #include "llvm/Support/Debug.h" 61 #include "llvm/Support/raw_ostream.h" 62 #include "llvm/Transforms/Scalar.h" 63 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 64 #include "llvm/Transforms/Utils/Cloning.h" 65 #include "llvm/Transforms/Utils/LoopUtils.h" 66 #include "llvm/Transforms/Utils/LoopVersioning.h" 67 #include "llvm/Transforms/Utils/ValueMapper.h" 68 #include <cassert> 69 #include <functional> 70 #include <list> 71 #include <tuple> 72 #include <utility> 73 74 using namespace llvm; 75 76 #define LDIST_NAME "loop-distribute" 77 #define DEBUG_TYPE LDIST_NAME 78 79 /// @{ 80 /// Metadata attribute names 81 static const char *const LLVMLoopDistributeFollowupAll = 82 "llvm.loop.distribute.followup_all"; 83 static const char *const LLVMLoopDistributeFollowupCoincident = 84 "llvm.loop.distribute.followup_coincident"; 85 static const char *const LLVMLoopDistributeFollowupSequential = 86 "llvm.loop.distribute.followup_sequential"; 87 static const char *const LLVMLoopDistributeFollowupFallback = 88 "llvm.loop.distribute.followup_fallback"; 89 /// @} 90 91 static cl::opt<bool> 92 LDistVerify("loop-distribute-verify", cl::Hidden, 93 cl::desc("Turn on DominatorTree and LoopInfo verification " 94 "after Loop Distribution"), 95 cl::init(false)); 96 97 static cl::opt<bool> DistributeNonIfConvertible( 98 "loop-distribute-non-if-convertible", cl::Hidden, 99 cl::desc("Whether to distribute into a loop that may not be " 100 "if-convertible by the loop vectorizer"), 101 cl::init(false)); 102 103 static cl::opt<unsigned> DistributeSCEVCheckThreshold( 104 "loop-distribute-scev-check-threshold", cl::init(8), cl::Hidden, 105 cl::desc("The maximum number of SCEV checks allowed for Loop " 106 "Distribution")); 107 108 static cl::opt<unsigned> PragmaDistributeSCEVCheckThreshold( 109 "loop-distribute-scev-check-threshold-with-pragma", cl::init(128), 110 cl::Hidden, 111 cl::desc( 112 "The maximum number of SCEV checks allowed for Loop " 113 "Distribution for loop marked with #pragma loop distribute(enable)")); 114 115 static cl::opt<bool> EnableLoopDistribute( 116 "enable-loop-distribute", cl::Hidden, 117 cl::desc("Enable the new, experimental LoopDistribution Pass"), 118 cl::init(false)); 119 120 STATISTIC(NumLoopsDistributed, "Number of loops distributed"); 121 122 namespace { 123 124 /// Maintains the set of instructions of the loop for a partition before 125 /// cloning. After cloning, it hosts the new loop. 126 class InstPartition { 127 using InstructionSet = SmallPtrSet<Instruction *, 8>; 128 129 public: 130 InstPartition(Instruction *I, Loop *L, bool DepCycle = false) 131 : DepCycle(DepCycle), OrigLoop(L) { 132 Set.insert(I); 133 } 134 135 /// Returns whether this partition contains a dependence cycle. 136 bool hasDepCycle() const { return DepCycle; } 137 138 /// Adds an instruction to this partition. 139 void add(Instruction *I) { Set.insert(I); } 140 141 /// Collection accessors. 142 InstructionSet::iterator begin() { return Set.begin(); } 143 InstructionSet::iterator end() { return Set.end(); } 144 InstructionSet::const_iterator begin() const { return Set.begin(); } 145 InstructionSet::const_iterator end() const { return Set.end(); } 146 bool empty() const { return Set.empty(); } 147 148 /// Moves this partition into \p Other. This partition becomes empty 149 /// after this. 150 void moveTo(InstPartition &Other) { 151 Other.Set.insert(Set.begin(), Set.end()); 152 Set.clear(); 153 Other.DepCycle |= DepCycle; 154 } 155 156 /// Populates the partition with a transitive closure of all the 157 /// instructions that the seeded instructions dependent on. 158 void populateUsedSet() { 159 // FIXME: We currently don't use control-dependence but simply include all 160 // blocks (possibly empty at the end) and let simplifycfg mostly clean this 161 // up. 162 for (auto *B : OrigLoop->getBlocks()) 163 Set.insert(B->getTerminator()); 164 165 // Follow the use-def chains to form a transitive closure of all the 166 // instructions that the originally seeded instructions depend on. 167 SmallVector<Instruction *, 8> Worklist(Set.begin(), Set.end()); 168 while (!Worklist.empty()) { 169 Instruction *I = Worklist.pop_back_val(); 170 // Insert instructions from the loop that we depend on. 171 for (Value *V : I->operand_values()) { 172 auto *I = dyn_cast<Instruction>(V); 173 if (I && OrigLoop->contains(I->getParent()) && Set.insert(I).second) 174 Worklist.push_back(I); 175 } 176 } 177 } 178 179 /// Clones the original loop. 180 /// 181 /// Updates LoopInfo and DominatorTree using the information that block \p 182 /// LoopDomBB dominates the loop. 183 Loop *cloneLoopWithPreheader(BasicBlock *InsertBefore, BasicBlock *LoopDomBB, 184 unsigned Index, LoopInfo *LI, 185 DominatorTree *DT) { 186 ClonedLoop = ::cloneLoopWithPreheader(InsertBefore, LoopDomBB, OrigLoop, 187 VMap, Twine(".ldist") + Twine(Index), 188 LI, DT, ClonedLoopBlocks); 189 return ClonedLoop; 190 } 191 192 /// The cloned loop. If this partition is mapped to the original loop, 193 /// this is null. 194 const Loop *getClonedLoop() const { return ClonedLoop; } 195 196 /// Returns the loop where this partition ends up after distribution. 197 /// If this partition is mapped to the original loop then use the block from 198 /// the loop. 199 Loop *getDistributedLoop() const { 200 return ClonedLoop ? ClonedLoop : OrigLoop; 201 } 202 203 /// The VMap that is populated by cloning and then used in 204 /// remapinstruction to remap the cloned instructions. 205 ValueToValueMapTy &getVMap() { return VMap; } 206 207 /// Remaps the cloned instructions using VMap. 208 void remapInstructions() { 209 remapInstructionsInBlocks(ClonedLoopBlocks, VMap); 210 } 211 212 /// Based on the set of instructions selected for this partition, 213 /// removes the unnecessary ones. 214 void removeUnusedInsts() { 215 SmallVector<Instruction *, 8> Unused; 216 217 for (auto *Block : OrigLoop->getBlocks()) 218 for (auto &Inst : *Block) 219 if (!Set.count(&Inst)) { 220 Instruction *NewInst = &Inst; 221 if (!VMap.empty()) 222 NewInst = cast<Instruction>(VMap[NewInst]); 223 224 assert(!isa<BranchInst>(NewInst) && 225 "Branches are marked used early on"); 226 Unused.push_back(NewInst); 227 } 228 229 // Delete the instructions backwards, as it has a reduced likelihood of 230 // having to update as many def-use and use-def chains. 231 for (auto *Inst : reverse(Unused)) { 232 if (!Inst->use_empty()) 233 Inst->replaceAllUsesWith(PoisonValue::get(Inst->getType())); 234 Inst->eraseFromParent(); 235 } 236 } 237 238 void print() const { 239 if (DepCycle) 240 dbgs() << " (cycle)\n"; 241 for (auto *I : Set) 242 // Prefix with the block name. 243 dbgs() << " " << I->getParent()->getName() << ":" << *I << "\n"; 244 } 245 246 void printBlocks() const { 247 for (auto *BB : getDistributedLoop()->getBlocks()) 248 dbgs() << *BB; 249 } 250 251 private: 252 /// Instructions from OrigLoop selected for this partition. 253 InstructionSet Set; 254 255 /// Whether this partition contains a dependence cycle. 256 bool DepCycle; 257 258 /// The original loop. 259 Loop *OrigLoop; 260 261 /// The cloned loop. If this partition is mapped to the original loop, 262 /// this is null. 263 Loop *ClonedLoop = nullptr; 264 265 /// The blocks of ClonedLoop including the preheader. If this 266 /// partition is mapped to the original loop, this is empty. 267 SmallVector<BasicBlock *, 8> ClonedLoopBlocks; 268 269 /// These gets populated once the set of instructions have been 270 /// finalized. If this partition is mapped to the original loop, these are not 271 /// set. 272 ValueToValueMapTy VMap; 273 }; 274 275 /// Holds the set of Partitions. It populates them, merges them and then 276 /// clones the loops. 277 class InstPartitionContainer { 278 using InstToPartitionIdT = DenseMap<Instruction *, int>; 279 280 public: 281 InstPartitionContainer(Loop *L, LoopInfo *LI, DominatorTree *DT) 282 : L(L), LI(LI), DT(DT) {} 283 284 /// Returns the number of partitions. 285 unsigned getSize() const { return PartitionContainer.size(); } 286 287 /// Adds \p Inst into the current partition if that is marked to 288 /// contain cycles. Otherwise start a new partition for it. 289 void addToCyclicPartition(Instruction *Inst) { 290 // If the current partition is non-cyclic. Start a new one. 291 if (PartitionContainer.empty() || !PartitionContainer.back().hasDepCycle()) 292 PartitionContainer.emplace_back(Inst, L, /*DepCycle=*/true); 293 else 294 PartitionContainer.back().add(Inst); 295 } 296 297 /// Adds \p Inst into a partition that is not marked to contain 298 /// dependence cycles. 299 /// 300 // Initially we isolate memory instructions into as many partitions as 301 // possible, then later we may merge them back together. 302 void addToNewNonCyclicPartition(Instruction *Inst) { 303 PartitionContainer.emplace_back(Inst, L); 304 } 305 306 /// Merges adjacent non-cyclic partitions. 307 /// 308 /// The idea is that we currently only want to isolate the non-vectorizable 309 /// partition. We could later allow more distribution among these partition 310 /// too. 311 void mergeAdjacentNonCyclic() { 312 mergeAdjacentPartitionsIf( 313 [](const InstPartition *P) { return !P->hasDepCycle(); }); 314 } 315 316 /// If a partition contains only conditional stores, we won't vectorize 317 /// it. Try to merge it with a previous cyclic partition. 318 void mergeNonIfConvertible() { 319 mergeAdjacentPartitionsIf([&](const InstPartition *Partition) { 320 if (Partition->hasDepCycle()) 321 return true; 322 323 // Now, check if all stores are conditional in this partition. 324 bool seenStore = false; 325 326 for (auto *Inst : *Partition) 327 if (isa<StoreInst>(Inst)) { 328 seenStore = true; 329 if (!LoopAccessInfo::blockNeedsPredication(Inst->getParent(), L, DT)) 330 return false; 331 } 332 return seenStore; 333 }); 334 } 335 336 /// Merges the partitions according to various heuristics. 337 void mergeBeforePopulating() { 338 mergeAdjacentNonCyclic(); 339 if (!DistributeNonIfConvertible) 340 mergeNonIfConvertible(); 341 } 342 343 /// Merges partitions in order to ensure that no loads are duplicated. 344 /// 345 /// We can't duplicate loads because that could potentially reorder them. 346 /// LoopAccessAnalysis provides dependency information with the context that 347 /// the order of memory operation is preserved. 348 /// 349 /// Return if any partitions were merged. 350 bool mergeToAvoidDuplicatedLoads() { 351 using LoadToPartitionT = DenseMap<Instruction *, InstPartition *>; 352 using ToBeMergedT = EquivalenceClasses<InstPartition *>; 353 354 LoadToPartitionT LoadToPartition; 355 ToBeMergedT ToBeMerged; 356 357 // Step through the partitions and create equivalence between partitions 358 // that contain the same load. Also put partitions in between them in the 359 // same equivalence class to avoid reordering of memory operations. 360 for (PartitionContainerT::iterator I = PartitionContainer.begin(), 361 E = PartitionContainer.end(); 362 I != E; ++I) { 363 auto *PartI = &*I; 364 365 // If a load occurs in two partitions PartI and PartJ, merge all 366 // partitions (PartI, PartJ] into PartI. 367 for (Instruction *Inst : *PartI) 368 if (isa<LoadInst>(Inst)) { 369 bool NewElt; 370 LoadToPartitionT::iterator LoadToPart; 371 372 std::tie(LoadToPart, NewElt) = 373 LoadToPartition.insert(std::make_pair(Inst, PartI)); 374 if (!NewElt) { 375 LLVM_DEBUG(dbgs() 376 << "Merging partitions due to this load in multiple " 377 << "partitions: " << PartI << ", " << LoadToPart->second 378 << "\n" 379 << *Inst << "\n"); 380 381 auto PartJ = I; 382 do { 383 --PartJ; 384 ToBeMerged.unionSets(PartI, &*PartJ); 385 } while (&*PartJ != LoadToPart->second); 386 } 387 } 388 } 389 if (ToBeMerged.empty()) 390 return false; 391 392 // Merge the member of an equivalence class into its class leader. This 393 // makes the members empty. 394 for (ToBeMergedT::iterator I = ToBeMerged.begin(), E = ToBeMerged.end(); 395 I != E; ++I) { 396 if (!I->isLeader()) 397 continue; 398 399 auto PartI = I->getData(); 400 for (auto PartJ : make_range(std::next(ToBeMerged.member_begin(I)), 401 ToBeMerged.member_end())) { 402 PartJ->moveTo(*PartI); 403 } 404 } 405 406 // Remove the empty partitions. 407 PartitionContainer.remove_if( 408 [](const InstPartition &P) { return P.empty(); }); 409 410 return true; 411 } 412 413 /// Sets up the mapping between instructions to partitions. If the 414 /// instruction is duplicated across multiple partitions, set the entry to -1. 415 void setupPartitionIdOnInstructions() { 416 int PartitionID = 0; 417 for (const auto &Partition : PartitionContainer) { 418 for (Instruction *Inst : Partition) { 419 bool NewElt; 420 InstToPartitionIdT::iterator Iter; 421 422 std::tie(Iter, NewElt) = 423 InstToPartitionId.insert(std::make_pair(Inst, PartitionID)); 424 if (!NewElt) 425 Iter->second = -1; 426 } 427 ++PartitionID; 428 } 429 } 430 431 /// Populates the partition with everything that the seeding 432 /// instructions require. 433 void populateUsedSet() { 434 for (auto &P : PartitionContainer) 435 P.populateUsedSet(); 436 } 437 438 /// This performs the main chunk of the work of cloning the loops for 439 /// the partitions. 440 void cloneLoops() { 441 BasicBlock *OrigPH = L->getLoopPreheader(); 442 // At this point the predecessor of the preheader is either the memcheck 443 // block or the top part of the original preheader. 444 BasicBlock *Pred = OrigPH->getSinglePredecessor(); 445 assert(Pred && "Preheader does not have a single predecessor"); 446 BasicBlock *ExitBlock = L->getExitBlock(); 447 assert(ExitBlock && "No single exit block"); 448 Loop *NewLoop; 449 450 assert(!PartitionContainer.empty() && "at least two partitions expected"); 451 // We're cloning the preheader along with the loop so we already made sure 452 // it was empty. 453 assert(&*OrigPH->begin() == OrigPH->getTerminator() && 454 "preheader not empty"); 455 456 // Preserve the original loop ID for use after the transformation. 457 MDNode *OrigLoopID = L->getLoopID(); 458 459 // Create a loop for each partition except the last. Clone the original 460 // loop before PH along with adding a preheader for the cloned loop. Then 461 // update PH to point to the newly added preheader. 462 BasicBlock *TopPH = OrigPH; 463 unsigned Index = getSize() - 1; 464 for (auto I = std::next(PartitionContainer.rbegin()), 465 E = PartitionContainer.rend(); 466 I != E; ++I, --Index, TopPH = NewLoop->getLoopPreheader()) { 467 auto *Part = &*I; 468 469 NewLoop = Part->cloneLoopWithPreheader(TopPH, Pred, Index, LI, DT); 470 471 Part->getVMap()[ExitBlock] = TopPH; 472 Part->remapInstructions(); 473 setNewLoopID(OrigLoopID, Part); 474 } 475 Pred->getTerminator()->replaceUsesOfWith(OrigPH, TopPH); 476 477 // Also set a new loop ID for the last loop. 478 setNewLoopID(OrigLoopID, &PartitionContainer.back()); 479 480 // Now go in forward order and update the immediate dominator for the 481 // preheaders with the exiting block of the previous loop. Dominance 482 // within the loop is updated in cloneLoopWithPreheader. 483 for (auto Curr = PartitionContainer.cbegin(), 484 Next = std::next(PartitionContainer.cbegin()), 485 E = PartitionContainer.cend(); 486 Next != E; ++Curr, ++Next) 487 DT->changeImmediateDominator( 488 Next->getDistributedLoop()->getLoopPreheader(), 489 Curr->getDistributedLoop()->getExitingBlock()); 490 } 491 492 /// Removes the dead instructions from the cloned loops. 493 void removeUnusedInsts() { 494 for (auto &Partition : PartitionContainer) 495 Partition.removeUnusedInsts(); 496 } 497 498 /// For each memory pointer, it computes the partitionId the pointer is 499 /// used in. 500 /// 501 /// This returns an array of int where the I-th entry corresponds to I-th 502 /// entry in LAI.getRuntimePointerCheck(). If the pointer is used in multiple 503 /// partitions its entry is set to -1. 504 SmallVector<int, 8> 505 computePartitionSetForPointers(const LoopAccessInfo &LAI) { 506 const RuntimePointerChecking *RtPtrCheck = LAI.getRuntimePointerChecking(); 507 508 unsigned N = RtPtrCheck->Pointers.size(); 509 SmallVector<int, 8> PtrToPartitions(N); 510 for (unsigned I = 0; I < N; ++I) { 511 Value *Ptr = RtPtrCheck->Pointers[I].PointerValue; 512 auto Instructions = 513 LAI.getInstructionsForAccess(Ptr, RtPtrCheck->Pointers[I].IsWritePtr); 514 515 int &Partition = PtrToPartitions[I]; 516 // First set it to uninitialized. 517 Partition = -2; 518 for (Instruction *Inst : Instructions) { 519 // Note that this could be -1 if Inst is duplicated across multiple 520 // partitions. 521 int ThisPartition = this->InstToPartitionId[Inst]; 522 if (Partition == -2) 523 Partition = ThisPartition; 524 // -1 means belonging to multiple partitions. 525 else if (Partition == -1) 526 break; 527 else if (Partition != (int)ThisPartition) 528 Partition = -1; 529 } 530 assert(Partition != -2 && "Pointer not belonging to any partition"); 531 } 532 533 return PtrToPartitions; 534 } 535 536 void print(raw_ostream &OS) const { 537 unsigned Index = 0; 538 for (const auto &P : PartitionContainer) { 539 OS << "Partition " << Index++ << " (" << &P << "):\n"; 540 P.print(); 541 } 542 } 543 544 void dump() const { print(dbgs()); } 545 546 #ifndef NDEBUG 547 friend raw_ostream &operator<<(raw_ostream &OS, 548 const InstPartitionContainer &Partitions) { 549 Partitions.print(OS); 550 return OS; 551 } 552 #endif 553 554 void printBlocks() const { 555 unsigned Index = 0; 556 for (const auto &P : PartitionContainer) { 557 dbgs() << "\nPartition " << Index++ << " (" << &P << "):\n"; 558 P.printBlocks(); 559 } 560 } 561 562 private: 563 using PartitionContainerT = std::list<InstPartition>; 564 565 /// List of partitions. 566 PartitionContainerT PartitionContainer; 567 568 /// Mapping from Instruction to partition Id. If the instruction 569 /// belongs to multiple partitions the entry contains -1. 570 InstToPartitionIdT InstToPartitionId; 571 572 Loop *L; 573 LoopInfo *LI; 574 DominatorTree *DT; 575 576 /// The control structure to merge adjacent partitions if both satisfy 577 /// the \p Predicate. 578 template <class UnaryPredicate> 579 void mergeAdjacentPartitionsIf(UnaryPredicate Predicate) { 580 InstPartition *PrevMatch = nullptr; 581 for (auto I = PartitionContainer.begin(); I != PartitionContainer.end();) { 582 auto DoesMatch = Predicate(&*I); 583 if (PrevMatch == nullptr && DoesMatch) { 584 PrevMatch = &*I; 585 ++I; 586 } else if (PrevMatch != nullptr && DoesMatch) { 587 I->moveTo(*PrevMatch); 588 I = PartitionContainer.erase(I); 589 } else { 590 PrevMatch = nullptr; 591 ++I; 592 } 593 } 594 } 595 596 /// Assign new LoopIDs for the partition's cloned loop. 597 void setNewLoopID(MDNode *OrigLoopID, InstPartition *Part) { 598 Optional<MDNode *> PartitionID = makeFollowupLoopID( 599 OrigLoopID, 600 {LLVMLoopDistributeFollowupAll, 601 Part->hasDepCycle() ? LLVMLoopDistributeFollowupSequential 602 : LLVMLoopDistributeFollowupCoincident}); 603 if (PartitionID) { 604 Loop *NewLoop = Part->getDistributedLoop(); 605 NewLoop->setLoopID(PartitionID.value()); 606 } 607 } 608 }; 609 610 /// For each memory instruction, this class maintains difference of the 611 /// number of unsafe dependences that start out from this instruction minus 612 /// those that end here. 613 /// 614 /// By traversing the memory instructions in program order and accumulating this 615 /// number, we know whether any unsafe dependence crosses over a program point. 616 class MemoryInstructionDependences { 617 using Dependence = MemoryDepChecker::Dependence; 618 619 public: 620 struct Entry { 621 Instruction *Inst; 622 unsigned NumUnsafeDependencesStartOrEnd = 0; 623 624 Entry(Instruction *Inst) : Inst(Inst) {} 625 }; 626 627 using AccessesType = SmallVector<Entry, 8>; 628 629 AccessesType::const_iterator begin() const { return Accesses.begin(); } 630 AccessesType::const_iterator end() const { return Accesses.end(); } 631 632 MemoryInstructionDependences( 633 const SmallVectorImpl<Instruction *> &Instructions, 634 const SmallVectorImpl<Dependence> &Dependences) { 635 Accesses.append(Instructions.begin(), Instructions.end()); 636 637 LLVM_DEBUG(dbgs() << "Backward dependences:\n"); 638 for (auto &Dep : Dependences) 639 if (Dep.isPossiblyBackward()) { 640 // Note that the designations source and destination follow the program 641 // order, i.e. source is always first. (The direction is given by the 642 // DepType.) 643 ++Accesses[Dep.Source].NumUnsafeDependencesStartOrEnd; 644 --Accesses[Dep.Destination].NumUnsafeDependencesStartOrEnd; 645 646 LLVM_DEBUG(Dep.print(dbgs(), 2, Instructions)); 647 } 648 } 649 650 private: 651 AccessesType Accesses; 652 }; 653 654 /// The actual class performing the per-loop work. 655 class LoopDistributeForLoop { 656 public: 657 LoopDistributeForLoop(Loop *L, Function *F, LoopInfo *LI, DominatorTree *DT, 658 ScalarEvolution *SE, OptimizationRemarkEmitter *ORE) 659 : L(L), F(F), LI(LI), DT(DT), SE(SE), ORE(ORE) { 660 setForced(); 661 } 662 663 /// Try to distribute an inner-most loop. 664 bool processLoop(std::function<const LoopAccessInfo &(Loop &)> &GetLAA) { 665 assert(L->isInnermost() && "Only process inner loops."); 666 667 LLVM_DEBUG(dbgs() << "\nLDist: In \"" 668 << L->getHeader()->getParent()->getName() 669 << "\" checking " << *L << "\n"); 670 671 // Having a single exit block implies there's also one exiting block. 672 if (!L->getExitBlock()) 673 return fail("MultipleExitBlocks", "multiple exit blocks"); 674 if (!L->isLoopSimplifyForm()) 675 return fail("NotLoopSimplifyForm", 676 "loop is not in loop-simplify form"); 677 if (!L->isRotatedForm()) 678 return fail("NotBottomTested", "loop is not bottom tested"); 679 680 BasicBlock *PH = L->getLoopPreheader(); 681 682 LAI = &GetLAA(*L); 683 684 // Currently, we only distribute to isolate the part of the loop with 685 // dependence cycles to enable partial vectorization. 686 if (LAI->canVectorizeMemory()) 687 return fail("MemOpsCanBeVectorized", 688 "memory operations are safe for vectorization"); 689 690 auto *Dependences = LAI->getDepChecker().getDependences(); 691 if (!Dependences || Dependences->empty()) 692 return fail("NoUnsafeDeps", "no unsafe dependences to isolate"); 693 694 InstPartitionContainer Partitions(L, LI, DT); 695 696 // First, go through each memory operation and assign them to consecutive 697 // partitions (the order of partitions follows program order). Put those 698 // with unsafe dependences into "cyclic" partition otherwise put each store 699 // in its own "non-cyclic" partition (we'll merge these later). 700 // 701 // Note that a memory operation (e.g. Load2 below) at a program point that 702 // has an unsafe dependence (Store3->Load1) spanning over it must be 703 // included in the same cyclic partition as the dependent operations. This 704 // is to preserve the original program order after distribution. E.g.: 705 // 706 // NumUnsafeDependencesStartOrEnd NumUnsafeDependencesActive 707 // Load1 -. 1 0->1 708 // Load2 | /Unsafe/ 0 1 709 // Store3 -' -1 1->0 710 // Load4 0 0 711 // 712 // NumUnsafeDependencesActive > 0 indicates this situation and in this case 713 // we just keep assigning to the same cyclic partition until 714 // NumUnsafeDependencesActive reaches 0. 715 const MemoryDepChecker &DepChecker = LAI->getDepChecker(); 716 MemoryInstructionDependences MID(DepChecker.getMemoryInstructions(), 717 *Dependences); 718 719 int NumUnsafeDependencesActive = 0; 720 for (auto &InstDep : MID) { 721 Instruction *I = InstDep.Inst; 722 // We update NumUnsafeDependencesActive post-instruction, catch the 723 // start of a dependence directly via NumUnsafeDependencesStartOrEnd. 724 if (NumUnsafeDependencesActive || 725 InstDep.NumUnsafeDependencesStartOrEnd > 0) 726 Partitions.addToCyclicPartition(I); 727 else 728 Partitions.addToNewNonCyclicPartition(I); 729 NumUnsafeDependencesActive += InstDep.NumUnsafeDependencesStartOrEnd; 730 assert(NumUnsafeDependencesActive >= 0 && 731 "Negative number of dependences active"); 732 } 733 734 // Add partitions for values used outside. These partitions can be out of 735 // order from the original program order. This is OK because if the 736 // partition uses a load we will merge this partition with the original 737 // partition of the load that we set up in the previous loop (see 738 // mergeToAvoidDuplicatedLoads). 739 auto DefsUsedOutside = findDefsUsedOutsideOfLoop(L); 740 for (auto *Inst : DefsUsedOutside) 741 Partitions.addToNewNonCyclicPartition(Inst); 742 743 LLVM_DEBUG(dbgs() << "Seeded partitions:\n" << Partitions); 744 if (Partitions.getSize() < 2) 745 return fail("CantIsolateUnsafeDeps", 746 "cannot isolate unsafe dependencies"); 747 748 // Run the merge heuristics: Merge non-cyclic adjacent partitions since we 749 // should be able to vectorize these together. 750 Partitions.mergeBeforePopulating(); 751 LLVM_DEBUG(dbgs() << "\nMerged partitions:\n" << Partitions); 752 if (Partitions.getSize() < 2) 753 return fail("CantIsolateUnsafeDeps", 754 "cannot isolate unsafe dependencies"); 755 756 // Now, populate the partitions with non-memory operations. 757 Partitions.populateUsedSet(); 758 LLVM_DEBUG(dbgs() << "\nPopulated partitions:\n" << Partitions); 759 760 // In order to preserve original lexical order for loads, keep them in the 761 // partition that we set up in the MemoryInstructionDependences loop. 762 if (Partitions.mergeToAvoidDuplicatedLoads()) { 763 LLVM_DEBUG(dbgs() << "\nPartitions merged to ensure unique loads:\n" 764 << Partitions); 765 if (Partitions.getSize() < 2) 766 return fail("CantIsolateUnsafeDeps", 767 "cannot isolate unsafe dependencies"); 768 } 769 770 // Don't distribute the loop if we need too many SCEV run-time checks, or 771 // any if it's illegal. 772 const SCEVPredicate &Pred = LAI->getPSE().getPredicate(); 773 if (LAI->hasConvergentOp() && !Pred.isAlwaysTrue()) { 774 return fail("RuntimeCheckWithConvergent", 775 "may not insert runtime check with convergent operation"); 776 } 777 778 if (Pred.getComplexity() > (IsForced.value_or(false) 779 ? PragmaDistributeSCEVCheckThreshold 780 : DistributeSCEVCheckThreshold)) 781 return fail("TooManySCEVRuntimeChecks", 782 "too many SCEV run-time checks needed.\n"); 783 784 if (!IsForced.value_or(false) && hasDisableAllTransformsHint(L)) 785 return fail("HeuristicDisabled", "distribution heuristic disabled"); 786 787 LLVM_DEBUG(dbgs() << "\nDistributing loop: " << *L << "\n"); 788 // We're done forming the partitions set up the reverse mapping from 789 // instructions to partitions. 790 Partitions.setupPartitionIdOnInstructions(); 791 792 // If we need run-time checks, version the loop now. 793 auto PtrToPartition = Partitions.computePartitionSetForPointers(*LAI); 794 const auto *RtPtrChecking = LAI->getRuntimePointerChecking(); 795 const auto &AllChecks = RtPtrChecking->getChecks(); 796 auto Checks = includeOnlyCrossPartitionChecks(AllChecks, PtrToPartition, 797 RtPtrChecking); 798 799 if (LAI->hasConvergentOp() && !Checks.empty()) { 800 return fail("RuntimeCheckWithConvergent", 801 "may not insert runtime check with convergent operation"); 802 } 803 804 // To keep things simple have an empty preheader before we version or clone 805 // the loop. (Also split if this has no predecessor, i.e. entry, because we 806 // rely on PH having a predecessor.) 807 if (!PH->getSinglePredecessor() || &*PH->begin() != PH->getTerminator()) 808 SplitBlock(PH, PH->getTerminator(), DT, LI); 809 810 if (!Pred.isAlwaysTrue() || !Checks.empty()) { 811 assert(!LAI->hasConvergentOp() && "inserting illegal loop versioning"); 812 813 MDNode *OrigLoopID = L->getLoopID(); 814 815 LLVM_DEBUG(dbgs() << "\nPointers:\n"); 816 LLVM_DEBUG(LAI->getRuntimePointerChecking()->printChecks(dbgs(), Checks)); 817 LoopVersioning LVer(*LAI, Checks, L, LI, DT, SE); 818 LVer.versionLoop(DefsUsedOutside); 819 LVer.annotateLoopWithNoAlias(); 820 821 // The unversioned loop will not be changed, so we inherit all attributes 822 // from the original loop, but remove the loop distribution metadata to 823 // avoid to distribute it again. 824 MDNode *UnversionedLoopID = 825 makeFollowupLoopID(OrigLoopID, 826 {LLVMLoopDistributeFollowupAll, 827 LLVMLoopDistributeFollowupFallback}, 828 "llvm.loop.distribute.", true) 829 .value(); 830 LVer.getNonVersionedLoop()->setLoopID(UnversionedLoopID); 831 } 832 833 // Create identical copies of the original loop for each partition and hook 834 // them up sequentially. 835 Partitions.cloneLoops(); 836 837 // Now, we remove the instruction from each loop that don't belong to that 838 // partition. 839 Partitions.removeUnusedInsts(); 840 LLVM_DEBUG(dbgs() << "\nAfter removing unused Instrs:\n"); 841 LLVM_DEBUG(Partitions.printBlocks()); 842 843 if (LDistVerify) { 844 LI->verify(*DT); 845 assert(DT->verify(DominatorTree::VerificationLevel::Fast)); 846 } 847 848 ++NumLoopsDistributed; 849 // Report the success. 850 ORE->emit([&]() { 851 return OptimizationRemark(LDIST_NAME, "Distribute", L->getStartLoc(), 852 L->getHeader()) 853 << "distributed loop"; 854 }); 855 return true; 856 } 857 858 /// Provide diagnostics then \return with false. 859 bool fail(StringRef RemarkName, StringRef Message) { 860 LLVMContext &Ctx = F->getContext(); 861 bool Forced = isForced().value_or(false); 862 863 LLVM_DEBUG(dbgs() << "Skipping; " << Message << "\n"); 864 865 // With Rpass-missed report that distribution failed. 866 ORE->emit([&]() { 867 return OptimizationRemarkMissed(LDIST_NAME, "NotDistributed", 868 L->getStartLoc(), L->getHeader()) 869 << "loop not distributed: use -Rpass-analysis=loop-distribute for " 870 "more " 871 "info"; 872 }); 873 874 // With Rpass-analysis report why. This is on by default if distribution 875 // was requested explicitly. 876 ORE->emit(OptimizationRemarkAnalysis( 877 Forced ? OptimizationRemarkAnalysis::AlwaysPrint : LDIST_NAME, 878 RemarkName, L->getStartLoc(), L->getHeader()) 879 << "loop not distributed: " << Message); 880 881 // Also issue a warning if distribution was requested explicitly but it 882 // failed. 883 if (Forced) 884 Ctx.diagnose(DiagnosticInfoOptimizationFailure( 885 *F, L->getStartLoc(), "loop not distributed: failed " 886 "explicitly specified loop distribution")); 887 888 return false; 889 } 890 891 /// Return if distribution forced to be enabled/disabled for the loop. 892 /// 893 /// If the optional has a value, it indicates whether distribution was forced 894 /// to be enabled (true) or disabled (false). If the optional has no value 895 /// distribution was not forced either way. 896 const Optional<bool> &isForced() const { return IsForced; } 897 898 private: 899 /// Filter out checks between pointers from the same partition. 900 /// 901 /// \p PtrToPartition contains the partition number for pointers. Partition 902 /// number -1 means that the pointer is used in multiple partitions. In this 903 /// case we can't safely omit the check. 904 SmallVector<RuntimePointerCheck, 4> includeOnlyCrossPartitionChecks( 905 const SmallVectorImpl<RuntimePointerCheck> &AllChecks, 906 const SmallVectorImpl<int> &PtrToPartition, 907 const RuntimePointerChecking *RtPtrChecking) { 908 SmallVector<RuntimePointerCheck, 4> Checks; 909 910 copy_if(AllChecks, std::back_inserter(Checks), 911 [&](const RuntimePointerCheck &Check) { 912 for (unsigned PtrIdx1 : Check.first->Members) 913 for (unsigned PtrIdx2 : Check.second->Members) 914 // Only include this check if there is a pair of pointers 915 // that require checking and the pointers fall into 916 // separate partitions. 917 // 918 // (Note that we already know at this point that the two 919 // pointer groups need checking but it doesn't follow 920 // that each pair of pointers within the two groups need 921 // checking as well. 922 // 923 // In other words we don't want to include a check just 924 // because there is a pair of pointers between the two 925 // pointer groups that require checks and a different 926 // pair whose pointers fall into different partitions.) 927 if (RtPtrChecking->needsChecking(PtrIdx1, PtrIdx2) && 928 !RuntimePointerChecking::arePointersInSamePartition( 929 PtrToPartition, PtrIdx1, PtrIdx2)) 930 return true; 931 return false; 932 }); 933 934 return Checks; 935 } 936 937 /// Check whether the loop metadata is forcing distribution to be 938 /// enabled/disabled. 939 void setForced() { 940 Optional<const MDOperand *> Value = 941 findStringMetadataForLoop(L, "llvm.loop.distribute.enable"); 942 if (!Value) 943 return; 944 945 const MDOperand *Op = *Value; 946 assert(Op && mdconst::hasa<ConstantInt>(*Op) && "invalid metadata"); 947 IsForced = mdconst::extract<ConstantInt>(*Op)->getZExtValue(); 948 } 949 950 Loop *L; 951 Function *F; 952 953 // Analyses used. 954 LoopInfo *LI; 955 const LoopAccessInfo *LAI = nullptr; 956 DominatorTree *DT; 957 ScalarEvolution *SE; 958 OptimizationRemarkEmitter *ORE; 959 960 /// Indicates whether distribution is forced to be enabled/disabled for 961 /// the loop. 962 /// 963 /// If the optional has a value, it indicates whether distribution was forced 964 /// to be enabled (true) or disabled (false). If the optional has no value 965 /// distribution was not forced either way. 966 Optional<bool> IsForced; 967 }; 968 969 } // end anonymous namespace 970 971 /// Shared implementation between new and old PMs. 972 static bool runImpl(Function &F, LoopInfo *LI, DominatorTree *DT, 973 ScalarEvolution *SE, OptimizationRemarkEmitter *ORE, 974 std::function<const LoopAccessInfo &(Loop &)> &GetLAA) { 975 // Build up a worklist of inner-loops to vectorize. This is necessary as the 976 // act of distributing a loop creates new loops and can invalidate iterators 977 // across the loops. 978 SmallVector<Loop *, 8> Worklist; 979 980 for (Loop *TopLevelLoop : *LI) 981 for (Loop *L : depth_first(TopLevelLoop)) 982 // We only handle inner-most loops. 983 if (L->isInnermost()) 984 Worklist.push_back(L); 985 986 // Now walk the identified inner loops. 987 bool Changed = false; 988 for (Loop *L : Worklist) { 989 LoopDistributeForLoop LDL(L, &F, LI, DT, SE, ORE); 990 991 // If distribution was forced for the specific loop to be 992 // enabled/disabled, follow that. Otherwise use the global flag. 993 if (LDL.isForced().value_or(EnableLoopDistribute)) 994 Changed |= LDL.processLoop(GetLAA); 995 } 996 997 // Process each loop nest in the function. 998 return Changed; 999 } 1000 1001 namespace { 1002 1003 /// The pass class. 1004 class LoopDistributeLegacy : public FunctionPass { 1005 public: 1006 static char ID; 1007 1008 LoopDistributeLegacy() : FunctionPass(ID) { 1009 // The default is set by the caller. 1010 initializeLoopDistributeLegacyPass(*PassRegistry::getPassRegistry()); 1011 } 1012 1013 bool runOnFunction(Function &F) override { 1014 if (skipFunction(F)) 1015 return false; 1016 1017 auto *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); 1018 auto *LAA = &getAnalysis<LoopAccessLegacyAnalysis>(); 1019 auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 1020 auto *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE(); 1021 auto *ORE = &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE(); 1022 std::function<const LoopAccessInfo &(Loop &)> GetLAA = 1023 [&](Loop &L) -> const LoopAccessInfo & { return LAA->getInfo(&L); }; 1024 1025 return runImpl(F, LI, DT, SE, ORE, GetLAA); 1026 } 1027 1028 void getAnalysisUsage(AnalysisUsage &AU) const override { 1029 AU.addRequired<ScalarEvolutionWrapperPass>(); 1030 AU.addRequired<LoopInfoWrapperPass>(); 1031 AU.addPreserved<LoopInfoWrapperPass>(); 1032 AU.addRequired<LoopAccessLegacyAnalysis>(); 1033 AU.addRequired<DominatorTreeWrapperPass>(); 1034 AU.addPreserved<DominatorTreeWrapperPass>(); 1035 AU.addRequired<OptimizationRemarkEmitterWrapperPass>(); 1036 AU.addPreserved<GlobalsAAWrapperPass>(); 1037 } 1038 }; 1039 1040 } // end anonymous namespace 1041 1042 PreservedAnalyses LoopDistributePass::run(Function &F, 1043 FunctionAnalysisManager &AM) { 1044 auto &LI = AM.getResult<LoopAnalysis>(F); 1045 auto &DT = AM.getResult<DominatorTreeAnalysis>(F); 1046 auto &SE = AM.getResult<ScalarEvolutionAnalysis>(F); 1047 auto &ORE = AM.getResult<OptimizationRemarkEmitterAnalysis>(F); 1048 1049 // We don't directly need these analyses but they're required for loop 1050 // analyses so provide them below. 1051 auto &AA = AM.getResult<AAManager>(F); 1052 auto &AC = AM.getResult<AssumptionAnalysis>(F); 1053 auto &TTI = AM.getResult<TargetIRAnalysis>(F); 1054 auto &TLI = AM.getResult<TargetLibraryAnalysis>(F); 1055 1056 auto &LAM = AM.getResult<LoopAnalysisManagerFunctionProxy>(F).getManager(); 1057 std::function<const LoopAccessInfo &(Loop &)> GetLAA = 1058 [&](Loop &L) -> const LoopAccessInfo & { 1059 LoopStandardAnalysisResults AR = {AA, AC, DT, LI, SE, 1060 TLI, TTI, nullptr, nullptr, nullptr}; 1061 return LAM.getResult<LoopAccessAnalysis>(L, AR); 1062 }; 1063 1064 bool Changed = runImpl(F, &LI, &DT, &SE, &ORE, GetLAA); 1065 if (!Changed) 1066 return PreservedAnalyses::all(); 1067 PreservedAnalyses PA; 1068 PA.preserve<LoopAnalysis>(); 1069 PA.preserve<DominatorTreeAnalysis>(); 1070 return PA; 1071 } 1072 1073 char LoopDistributeLegacy::ID; 1074 1075 static const char ldist_name[] = "Loop Distribution"; 1076 1077 INITIALIZE_PASS_BEGIN(LoopDistributeLegacy, LDIST_NAME, ldist_name, false, 1078 false) 1079 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass) 1080 INITIALIZE_PASS_DEPENDENCY(LoopAccessLegacyAnalysis) 1081 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 1082 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass) 1083 INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass) 1084 INITIALIZE_PASS_END(LoopDistributeLegacy, LDIST_NAME, ldist_name, false, false) 1085 1086 FunctionPass *llvm::createLoopDistributePass() { return new LoopDistributeLegacy(); } 1087