1 //===- LoopDistribute.cpp - Loop Distribution Pass ------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the Loop Distribution Pass. Its main focus is to
11 // distribute loops that cannot be vectorized due to dependence cycles. It
12 // tries to isolate the offending dependences into a new loop allowing
13 // vectorization of the remaining parts.
15 // For dependence analysis, the pass uses the LoopVectorizer's
16 // LoopAccessAnalysis. Because this analysis presumes no change in the order of
17 // memory operations, special care is taken to preserve the lexical order of
20 // Similarly to the Vectorizer, the pass also supports loop versioning to
21 // run-time disambiguate potentially overlapping arrays.
23 //===----------------------------------------------------------------------===//
25 #include "llvm/Transforms/Scalar/LoopDistribute.h"
26 #include "llvm/ADT/DenseMap.h"
27 #include "llvm/ADT/DepthFirstIterator.h"
28 #include "llvm/ADT/EquivalenceClasses.h"
29 #include "llvm/ADT/Optional.h"
30 #include "llvm/ADT/STLExtras.h"
31 #include "llvm/ADT/SmallPtrSet.h"
32 #include "llvm/ADT/SmallVector.h"
33 #include "llvm/ADT/Statistic.h"
34 #include "llvm/ADT/StringRef.h"
35 #include "llvm/ADT/Twine.h"
36 #include "llvm/ADT/iterator_range.h"
37 #include "llvm/Analysis/AliasAnalysis.h"
38 #include "llvm/Analysis/AssumptionCache.h"
39 #include "llvm/Analysis/GlobalsModRef.h"
40 #include "llvm/Analysis/LoopAccessAnalysis.h"
41 #include "llvm/Analysis/LoopAnalysisManager.h"
42 #include "llvm/Analysis/LoopInfo.h"
43 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
44 #include "llvm/Analysis/ScalarEvolution.h"
45 #include "llvm/Analysis/TargetLibraryInfo.h"
46 #include "llvm/Analysis/TargetTransformInfo.h"
47 #include "llvm/IR/BasicBlock.h"
48 #include "llvm/IR/Constants.h"
49 #include "llvm/IR/DiagnosticInfo.h"
50 #include "llvm/IR/Dominators.h"
51 #include "llvm/IR/Function.h"
52 #include "llvm/IR/InstrTypes.h"
53 #include "llvm/IR/Instruction.h"
54 #include "llvm/IR/Instructions.h"
55 #include "llvm/IR/LLVMContext.h"
56 #include "llvm/IR/Metadata.h"
57 #include "llvm/IR/PassManager.h"
58 #include "llvm/IR/Value.h"
59 #include "llvm/Pass.h"
60 #include "llvm/Support/Casting.h"
61 #include "llvm/Support/CommandLine.h"
62 #include "llvm/Support/Debug.h"
63 #include "llvm/Support/raw_ostream.h"
64 #include "llvm/Transforms/Scalar.h"
65 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
66 #include "llvm/Transforms/Utils/Cloning.h"
67 #include "llvm/Transforms/Utils/LoopUtils.h"
68 #include "llvm/Transforms/Utils/LoopVersioning.h"
69 #include "llvm/Transforms/Utils/ValueMapper.h"
78 #define LDIST_NAME "loop-distribute"
79 #define DEBUG_TYPE LDIST_NAME
82 LDistVerify("loop-distribute-verify", cl::Hidden,
83 cl::desc("Turn on DominatorTree and LoopInfo verification "
84 "after Loop Distribution"),
87 static cl::opt<bool> DistributeNonIfConvertible(
88 "loop-distribute-non-if-convertible", cl::Hidden,
89 cl::desc("Whether to distribute into a loop that may not be "
90 "if-convertible by the loop vectorizer"),
93 static cl::opt<unsigned> DistributeSCEVCheckThreshold(
94 "loop-distribute-scev-check-threshold", cl::init(8), cl::Hidden,
95 cl::desc("The maximum number of SCEV checks allowed for Loop "
98 static cl::opt<unsigned> PragmaDistributeSCEVCheckThreshold(
99 "loop-distribute-scev-check-threshold-with-pragma", cl::init(128),
102 "The maximum number of SCEV checks allowed for Loop "
103 "Distribution for loop marked with #pragma loop distribute(enable)"));
105 static cl::opt<bool> EnableLoopDistribute(
106 "enable-loop-distribute", cl::Hidden,
107 cl::desc("Enable the new, experimental LoopDistribution Pass"),
110 STATISTIC(NumLoopsDistributed, "Number of loops distributed");
114 /// \brief Maintains the set of instructions of the loop for a partition before
115 /// cloning. After cloning, it hosts the new loop.
116 class InstPartition {
117 using InstructionSet = SmallPtrSet<Instruction *, 8>;
120 InstPartition(Instruction *I, Loop *L, bool DepCycle = false)
121 : DepCycle(DepCycle), OrigLoop(L) {
125 /// \brief Returns whether this partition contains a dependence cycle.
126 bool hasDepCycle() const { return DepCycle; }
128 /// \brief Adds an instruction to this partition.
129 void add(Instruction *I) { Set.insert(I); }
131 /// \brief Collection accessors.
132 InstructionSet::iterator begin() { return Set.begin(); }
133 InstructionSet::iterator end() { return Set.end(); }
134 InstructionSet::const_iterator begin() const { return Set.begin(); }
135 InstructionSet::const_iterator end() const { return Set.end(); }
136 bool empty() const { return Set.empty(); }
138 /// \brief Moves this partition into \p Other. This partition becomes empty
140 void moveTo(InstPartition &Other) {
141 Other.Set.insert(Set.begin(), Set.end());
143 Other.DepCycle |= DepCycle;
146 /// \brief Populates the partition with a transitive closure of all the
147 /// instructions that the seeded instructions dependent on.
148 void populateUsedSet() {
149 // FIXME: We currently don't use control-dependence but simply include all
150 // blocks (possibly empty at the end) and let simplifycfg mostly clean this
152 for (auto *B : OrigLoop->getBlocks())
153 Set.insert(B->getTerminator());
155 // Follow the use-def chains to form a transitive closure of all the
156 // instructions that the originally seeded instructions depend on.
157 SmallVector<Instruction *, 8> Worklist(Set.begin(), Set.end());
158 while (!Worklist.empty()) {
159 Instruction *I = Worklist.pop_back_val();
160 // Insert instructions from the loop that we depend on.
161 for (Value *V : I->operand_values()) {
162 auto *I = dyn_cast<Instruction>(V);
163 if (I && OrigLoop->contains(I->getParent()) && Set.insert(I).second)
164 Worklist.push_back(I);
169 /// \brief Clones the original loop.
171 /// Updates LoopInfo and DominatorTree using the information that block \p
172 /// LoopDomBB dominates the loop.
173 Loop *cloneLoopWithPreheader(BasicBlock *InsertBefore, BasicBlock *LoopDomBB,
174 unsigned Index, LoopInfo *LI,
176 ClonedLoop = ::cloneLoopWithPreheader(InsertBefore, LoopDomBB, OrigLoop,
177 VMap, Twine(".ldist") + Twine(Index),
178 LI, DT, ClonedLoopBlocks);
182 /// \brief The cloned loop. If this partition is mapped to the original loop,
184 const Loop *getClonedLoop() const { return ClonedLoop; }
186 /// \brief Returns the loop where this partition ends up after distribution.
187 /// If this partition is mapped to the original loop then use the block from
189 const Loop *getDistributedLoop() const {
190 return ClonedLoop ? ClonedLoop : OrigLoop;
193 /// \brief The VMap that is populated by cloning and then used in
194 /// remapinstruction to remap the cloned instructions.
195 ValueToValueMapTy &getVMap() { return VMap; }
197 /// \brief Remaps the cloned instructions using VMap.
198 void remapInstructions() {
199 remapInstructionsInBlocks(ClonedLoopBlocks, VMap);
202 /// \brief Based on the set of instructions selected for this partition,
203 /// removes the unnecessary ones.
204 void removeUnusedInsts() {
205 SmallVector<Instruction *, 8> Unused;
207 for (auto *Block : OrigLoop->getBlocks())
208 for (auto &Inst : *Block)
209 if (!Set.count(&Inst)) {
210 Instruction *NewInst = &Inst;
212 NewInst = cast<Instruction>(VMap[NewInst]);
214 assert(!isa<BranchInst>(NewInst) &&
215 "Branches are marked used early on");
216 Unused.push_back(NewInst);
219 // Delete the instructions backwards, as it has a reduced likelihood of
220 // having to update as many def-use and use-def chains.
221 for (auto *Inst : reverse(Unused)) {
222 if (!Inst->use_empty())
223 Inst->replaceAllUsesWith(UndefValue::get(Inst->getType()));
224 Inst->eraseFromParent();
230 dbgs() << " (cycle)\n";
232 // Prefix with the block name.
233 dbgs() << " " << I->getParent()->getName() << ":" << *I << "\n";
236 void printBlocks() const {
237 for (auto *BB : getDistributedLoop()->getBlocks())
242 /// \brief Instructions from OrigLoop selected for this partition.
245 /// \brief Whether this partition contains a dependence cycle.
248 /// \brief The original loop.
251 /// \brief The cloned loop. If this partition is mapped to the original loop,
253 Loop *ClonedLoop = nullptr;
255 /// \brief The blocks of ClonedLoop including the preheader. If this
256 /// partition is mapped to the original loop, this is empty.
257 SmallVector<BasicBlock *, 8> ClonedLoopBlocks;
259 /// \brief These gets populated once the set of instructions have been
260 /// finalized. If this partition is mapped to the original loop, these are not
262 ValueToValueMapTy VMap;
265 /// \brief Holds the set of Partitions. It populates them, merges them and then
266 /// clones the loops.
267 class InstPartitionContainer {
268 using InstToPartitionIdT = DenseMap<Instruction *, int>;
271 InstPartitionContainer(Loop *L, LoopInfo *LI, DominatorTree *DT)
272 : L(L), LI(LI), DT(DT) {}
274 /// \brief Returns the number of partitions.
275 unsigned getSize() const { return PartitionContainer.size(); }
277 /// \brief Adds \p Inst into the current partition if that is marked to
278 /// contain cycles. Otherwise start a new partition for it.
279 void addToCyclicPartition(Instruction *Inst) {
280 // If the current partition is non-cyclic. Start a new one.
281 if (PartitionContainer.empty() || !PartitionContainer.back().hasDepCycle())
282 PartitionContainer.emplace_back(Inst, L, /*DepCycle=*/true);
284 PartitionContainer.back().add(Inst);
287 /// \brief Adds \p Inst into a partition that is not marked to contain
288 /// dependence cycles.
290 // Initially we isolate memory instructions into as many partitions as
291 // possible, then later we may merge them back together.
292 void addToNewNonCyclicPartition(Instruction *Inst) {
293 PartitionContainer.emplace_back(Inst, L);
296 /// \brief Merges adjacent non-cyclic partitions.
298 /// The idea is that we currently only want to isolate the non-vectorizable
299 /// partition. We could later allow more distribution among these partition
301 void mergeAdjacentNonCyclic() {
302 mergeAdjacentPartitionsIf(
303 [](const InstPartition *P) { return !P->hasDepCycle(); });
306 /// \brief If a partition contains only conditional stores, we won't vectorize
307 /// it. Try to merge it with a previous cyclic partition.
308 void mergeNonIfConvertible() {
309 mergeAdjacentPartitionsIf([&](const InstPartition *Partition) {
310 if (Partition->hasDepCycle())
313 // Now, check if all stores are conditional in this partition.
314 bool seenStore = false;
316 for (auto *Inst : *Partition)
317 if (isa<StoreInst>(Inst)) {
319 if (!LoopAccessInfo::blockNeedsPredication(Inst->getParent(), L, DT))
326 /// \brief Merges the partitions according to various heuristics.
327 void mergeBeforePopulating() {
328 mergeAdjacentNonCyclic();
329 if (!DistributeNonIfConvertible)
330 mergeNonIfConvertible();
333 /// \brief Merges partitions in order to ensure that no loads are duplicated.
335 /// We can't duplicate loads because that could potentially reorder them.
336 /// LoopAccessAnalysis provides dependency information with the context that
337 /// the order of memory operation is preserved.
339 /// Return if any partitions were merged.
340 bool mergeToAvoidDuplicatedLoads() {
341 using LoadToPartitionT = DenseMap<Instruction *, InstPartition *>;
342 using ToBeMergedT = EquivalenceClasses<InstPartition *>;
344 LoadToPartitionT LoadToPartition;
345 ToBeMergedT ToBeMerged;
347 // Step through the partitions and create equivalence between partitions
348 // that contain the same load. Also put partitions in between them in the
349 // same equivalence class to avoid reordering of memory operations.
350 for (PartitionContainerT::iterator I = PartitionContainer.begin(),
351 E = PartitionContainer.end();
355 // If a load occurs in two partitions PartI and PartJ, merge all
356 // partitions (PartI, PartJ] into PartI.
357 for (Instruction *Inst : *PartI)
358 if (isa<LoadInst>(Inst)) {
360 LoadToPartitionT::iterator LoadToPart;
362 std::tie(LoadToPart, NewElt) =
363 LoadToPartition.insert(std::make_pair(Inst, PartI));
365 DEBUG(dbgs() << "Merging partitions due to this load in multiple "
366 << "partitions: " << PartI << ", "
367 << LoadToPart->second << "\n" << *Inst << "\n");
372 ToBeMerged.unionSets(PartI, &*PartJ);
373 } while (&*PartJ != LoadToPart->second);
377 if (ToBeMerged.empty())
380 // Merge the member of an equivalence class into its class leader. This
381 // makes the members empty.
382 for (ToBeMergedT::iterator I = ToBeMerged.begin(), E = ToBeMerged.end();
387 auto PartI = I->getData();
388 for (auto PartJ : make_range(std::next(ToBeMerged.member_begin(I)),
389 ToBeMerged.member_end())) {
390 PartJ->moveTo(*PartI);
394 // Remove the empty partitions.
395 PartitionContainer.remove_if(
396 [](const InstPartition &P) { return P.empty(); });
401 /// \brief Sets up the mapping between instructions to partitions. If the
402 /// instruction is duplicated across multiple partitions, set the entry to -1.
403 void setupPartitionIdOnInstructions() {
405 for (const auto &Partition : PartitionContainer) {
406 for (Instruction *Inst : Partition) {
408 InstToPartitionIdT::iterator Iter;
410 std::tie(Iter, NewElt) =
411 InstToPartitionId.insert(std::make_pair(Inst, PartitionID));
419 /// \brief Populates the partition with everything that the seeding
420 /// instructions require.
421 void populateUsedSet() {
422 for (auto &P : PartitionContainer)
426 /// \brief This performs the main chunk of the work of cloning the loops for
429 BasicBlock *OrigPH = L->getLoopPreheader();
430 // At this point the predecessor of the preheader is either the memcheck
431 // block or the top part of the original preheader.
432 BasicBlock *Pred = OrigPH->getSinglePredecessor();
433 assert(Pred && "Preheader does not have a single predecessor");
434 BasicBlock *ExitBlock = L->getExitBlock();
435 assert(ExitBlock && "No single exit block");
438 assert(!PartitionContainer.empty() && "at least two partitions expected");
439 // We're cloning the preheader along with the loop so we already made sure
441 assert(&*OrigPH->begin() == OrigPH->getTerminator() &&
442 "preheader not empty");
444 // Create a loop for each partition except the last. Clone the original
445 // loop before PH along with adding a preheader for the cloned loop. Then
446 // update PH to point to the newly added preheader.
447 BasicBlock *TopPH = OrigPH;
448 unsigned Index = getSize() - 1;
449 for (auto I = std::next(PartitionContainer.rbegin()),
450 E = PartitionContainer.rend();
451 I != E; ++I, --Index, TopPH = NewLoop->getLoopPreheader()) {
454 NewLoop = Part->cloneLoopWithPreheader(TopPH, Pred, Index, LI, DT);
456 Part->getVMap()[ExitBlock] = TopPH;
457 Part->remapInstructions();
459 Pred->getTerminator()->replaceUsesOfWith(OrigPH, TopPH);
461 // Now go in forward order and update the immediate dominator for the
462 // preheaders with the exiting block of the previous loop. Dominance
463 // within the loop is updated in cloneLoopWithPreheader.
464 for (auto Curr = PartitionContainer.cbegin(),
465 Next = std::next(PartitionContainer.cbegin()),
466 E = PartitionContainer.cend();
467 Next != E; ++Curr, ++Next)
468 DT->changeImmediateDominator(
469 Next->getDistributedLoop()->getLoopPreheader(),
470 Curr->getDistributedLoop()->getExitingBlock());
473 /// \brief Removes the dead instructions from the cloned loops.
474 void removeUnusedInsts() {
475 for (auto &Partition : PartitionContainer)
476 Partition.removeUnusedInsts();
479 /// \brief For each memory pointer, it computes the partitionId the pointer is
482 /// This returns an array of int where the I-th entry corresponds to I-th
483 /// entry in LAI.getRuntimePointerCheck(). If the pointer is used in multiple
484 /// partitions its entry is set to -1.
486 computePartitionSetForPointers(const LoopAccessInfo &LAI) {
487 const RuntimePointerChecking *RtPtrCheck = LAI.getRuntimePointerChecking();
489 unsigned N = RtPtrCheck->Pointers.size();
490 SmallVector<int, 8> PtrToPartitions(N);
491 for (unsigned I = 0; I < N; ++I) {
492 Value *Ptr = RtPtrCheck->Pointers[I].PointerValue;
494 LAI.getInstructionsForAccess(Ptr, RtPtrCheck->Pointers[I].IsWritePtr);
496 int &Partition = PtrToPartitions[I];
497 // First set it to uninitialized.
499 for (Instruction *Inst : Instructions) {
500 // Note that this could be -1 if Inst is duplicated across multiple
502 int ThisPartition = this->InstToPartitionId[Inst];
504 Partition = ThisPartition;
505 // -1 means belonging to multiple partitions.
506 else if (Partition == -1)
508 else if (Partition != (int)ThisPartition)
511 assert(Partition != -2 && "Pointer not belonging to any partition");
514 return PtrToPartitions;
517 void print(raw_ostream &OS) const {
519 for (const auto &P : PartitionContainer) {
520 OS << "Partition " << Index++ << " (" << &P << "):\n";
525 void dump() const { print(dbgs()); }
528 friend raw_ostream &operator<<(raw_ostream &OS,
529 const InstPartitionContainer &Partitions) {
530 Partitions.print(OS);
535 void printBlocks() const {
537 for (const auto &P : PartitionContainer) {
538 dbgs() << "\nPartition " << Index++ << " (" << &P << "):\n";
544 using PartitionContainerT = std::list<InstPartition>;
546 /// \brief List of partitions.
547 PartitionContainerT PartitionContainer;
549 /// \brief Mapping from Instruction to partition Id. If the instruction
550 /// belongs to multiple partitions the entry contains -1.
551 InstToPartitionIdT InstToPartitionId;
557 /// \brief The control structure to merge adjacent partitions if both satisfy
558 /// the \p Predicate.
559 template <class UnaryPredicate>
560 void mergeAdjacentPartitionsIf(UnaryPredicate Predicate) {
561 InstPartition *PrevMatch = nullptr;
562 for (auto I = PartitionContainer.begin(); I != PartitionContainer.end();) {
563 auto DoesMatch = Predicate(&*I);
564 if (PrevMatch == nullptr && DoesMatch) {
567 } else if (PrevMatch != nullptr && DoesMatch) {
568 I->moveTo(*PrevMatch);
569 I = PartitionContainer.erase(I);
578 /// \brief For each memory instruction, this class maintains difference of the
579 /// number of unsafe dependences that start out from this instruction minus
580 /// those that end here.
582 /// By traversing the memory instructions in program order and accumulating this
583 /// number, we know whether any unsafe dependence crosses over a program point.
584 class MemoryInstructionDependences {
585 using Dependence = MemoryDepChecker::Dependence;
590 unsigned NumUnsafeDependencesStartOrEnd = 0;
592 Entry(Instruction *Inst) : Inst(Inst) {}
595 using AccessesType = SmallVector<Entry, 8>;
597 AccessesType::const_iterator begin() const { return Accesses.begin(); }
598 AccessesType::const_iterator end() const { return Accesses.end(); }
600 MemoryInstructionDependences(
601 const SmallVectorImpl<Instruction *> &Instructions,
602 const SmallVectorImpl<Dependence> &Dependences) {
603 Accesses.append(Instructions.begin(), Instructions.end());
605 DEBUG(dbgs() << "Backward dependences:\n");
606 for (auto &Dep : Dependences)
607 if (Dep.isPossiblyBackward()) {
608 // Note that the designations source and destination follow the program
609 // order, i.e. source is always first. (The direction is given by the
611 ++Accesses[Dep.Source].NumUnsafeDependencesStartOrEnd;
612 --Accesses[Dep.Destination].NumUnsafeDependencesStartOrEnd;
614 DEBUG(Dep.print(dbgs(), 2, Instructions));
619 AccessesType Accesses;
622 /// \brief The actual class performing the per-loop work.
623 class LoopDistributeForLoop {
625 LoopDistributeForLoop(Loop *L, Function *F, LoopInfo *LI, DominatorTree *DT,
626 ScalarEvolution *SE, OptimizationRemarkEmitter *ORE)
627 : L(L), F(F), LI(LI), DT(DT), SE(SE), ORE(ORE) {
631 /// \brief Try to distribute an inner-most loop.
632 bool processLoop(std::function<const LoopAccessInfo &(Loop &)> &GetLAA) {
633 assert(L->empty() && "Only process inner loops.");
635 DEBUG(dbgs() << "\nLDist: In \"" << L->getHeader()->getParent()->getName()
636 << "\" checking " << *L << "\n");
638 if (!L->getExitBlock())
639 return fail("MultipleExitBlocks", "multiple exit blocks");
640 if (!L->isLoopSimplifyForm())
641 return fail("NotLoopSimplifyForm",
642 "loop is not in loop-simplify form");
644 BasicBlock *PH = L->getLoopPreheader();
646 // LAA will check that we only have a single exiting block.
649 // Currently, we only distribute to isolate the part of the loop with
650 // dependence cycles to enable partial vectorization.
651 if (LAI->canVectorizeMemory())
652 return fail("MemOpsCanBeVectorized",
653 "memory operations are safe for vectorization");
655 auto *Dependences = LAI->getDepChecker().getDependences();
656 if (!Dependences || Dependences->empty())
657 return fail("NoUnsafeDeps", "no unsafe dependences to isolate");
659 InstPartitionContainer Partitions(L, LI, DT);
661 // First, go through each memory operation and assign them to consecutive
662 // partitions (the order of partitions follows program order). Put those
663 // with unsafe dependences into "cyclic" partition otherwise put each store
664 // in its own "non-cyclic" partition (we'll merge these later).
666 // Note that a memory operation (e.g. Load2 below) at a program point that
667 // has an unsafe dependence (Store3->Load1) spanning over it must be
668 // included in the same cyclic partition as the dependent operations. This
669 // is to preserve the original program order after distribution. E.g.:
671 // NumUnsafeDependencesStartOrEnd NumUnsafeDependencesActive
673 // Load2 | /Unsafe/ 0 1
677 // NumUnsafeDependencesActive > 0 indicates this situation and in this case
678 // we just keep assigning to the same cyclic partition until
679 // NumUnsafeDependencesActive reaches 0.
680 const MemoryDepChecker &DepChecker = LAI->getDepChecker();
681 MemoryInstructionDependences MID(DepChecker.getMemoryInstructions(),
684 int NumUnsafeDependencesActive = 0;
685 for (auto &InstDep : MID) {
686 Instruction *I = InstDep.Inst;
687 // We update NumUnsafeDependencesActive post-instruction, catch the
688 // start of a dependence directly via NumUnsafeDependencesStartOrEnd.
689 if (NumUnsafeDependencesActive ||
690 InstDep.NumUnsafeDependencesStartOrEnd > 0)
691 Partitions.addToCyclicPartition(I);
693 Partitions.addToNewNonCyclicPartition(I);
694 NumUnsafeDependencesActive += InstDep.NumUnsafeDependencesStartOrEnd;
695 assert(NumUnsafeDependencesActive >= 0 &&
696 "Negative number of dependences active");
699 // Add partitions for values used outside. These partitions can be out of
700 // order from the original program order. This is OK because if the
701 // partition uses a load we will merge this partition with the original
702 // partition of the load that we set up in the previous loop (see
703 // mergeToAvoidDuplicatedLoads).
704 auto DefsUsedOutside = findDefsUsedOutsideOfLoop(L);
705 for (auto *Inst : DefsUsedOutside)
706 Partitions.addToNewNonCyclicPartition(Inst);
708 DEBUG(dbgs() << "Seeded partitions:\n" << Partitions);
709 if (Partitions.getSize() < 2)
710 return fail("CantIsolateUnsafeDeps",
711 "cannot isolate unsafe dependencies");
713 // Run the merge heuristics: Merge non-cyclic adjacent partitions since we
714 // should be able to vectorize these together.
715 Partitions.mergeBeforePopulating();
716 DEBUG(dbgs() << "\nMerged partitions:\n" << Partitions);
717 if (Partitions.getSize() < 2)
718 return fail("CantIsolateUnsafeDeps",
719 "cannot isolate unsafe dependencies");
721 // Now, populate the partitions with non-memory operations.
722 Partitions.populateUsedSet();
723 DEBUG(dbgs() << "\nPopulated partitions:\n" << Partitions);
725 // In order to preserve original lexical order for loads, keep them in the
726 // partition that we set up in the MemoryInstructionDependences loop.
727 if (Partitions.mergeToAvoidDuplicatedLoads()) {
728 DEBUG(dbgs() << "\nPartitions merged to ensure unique loads:\n"
730 if (Partitions.getSize() < 2)
731 return fail("CantIsolateUnsafeDeps",
732 "cannot isolate unsafe dependencies");
735 // Don't distribute the loop if we need too many SCEV run-time checks.
736 const SCEVUnionPredicate &Pred = LAI->getPSE().getUnionPredicate();
737 if (Pred.getComplexity() > (IsForced.getValueOr(false)
738 ? PragmaDistributeSCEVCheckThreshold
739 : DistributeSCEVCheckThreshold))
740 return fail("TooManySCEVRuntimeChecks",
741 "too many SCEV run-time checks needed.\n");
743 DEBUG(dbgs() << "\nDistributing loop: " << *L << "\n");
744 // We're done forming the partitions set up the reverse mapping from
745 // instructions to partitions.
746 Partitions.setupPartitionIdOnInstructions();
748 // To keep things simple have an empty preheader before we version or clone
749 // the loop. (Also split if this has no predecessor, i.e. entry, because we
750 // rely on PH having a predecessor.)
751 if (!PH->getSinglePredecessor() || &*PH->begin() != PH->getTerminator())
752 SplitBlock(PH, PH->getTerminator(), DT, LI);
754 // If we need run-time checks, version the loop now.
755 auto PtrToPartition = Partitions.computePartitionSetForPointers(*LAI);
756 const auto *RtPtrChecking = LAI->getRuntimePointerChecking();
757 const auto &AllChecks = RtPtrChecking->getChecks();
758 auto Checks = includeOnlyCrossPartitionChecks(AllChecks, PtrToPartition,
761 if (!Pred.isAlwaysTrue() || !Checks.empty()) {
762 DEBUG(dbgs() << "\nPointers:\n");
763 DEBUG(LAI->getRuntimePointerChecking()->printChecks(dbgs(), Checks));
764 LoopVersioning LVer(*LAI, L, LI, DT, SE, false);
765 LVer.setAliasChecks(std::move(Checks));
766 LVer.setSCEVChecks(LAI->getPSE().getUnionPredicate());
767 LVer.versionLoop(DefsUsedOutside);
768 LVer.annotateLoopWithNoAlias();
771 // Create identical copies of the original loop for each partition and hook
772 // them up sequentially.
773 Partitions.cloneLoops();
775 // Now, we remove the instruction from each loop that don't belong to that
777 Partitions.removeUnusedInsts();
778 DEBUG(dbgs() << "\nAfter removing unused Instrs:\n");
779 DEBUG(Partitions.printBlocks());
786 ++NumLoopsDistributed;
787 // Report the success.
789 return OptimizationRemark(LDIST_NAME, "Distribute", L->getStartLoc(),
791 << "distributed loop";
796 /// \brief Provide diagnostics then \return with false.
797 bool fail(StringRef RemarkName, StringRef Message) {
798 LLVMContext &Ctx = F->getContext();
799 bool Forced = isForced().getValueOr(false);
801 DEBUG(dbgs() << "Skipping; " << Message << "\n");
803 // With Rpass-missed report that distribution failed.
805 return OptimizationRemarkMissed(LDIST_NAME, "NotDistributed",
806 L->getStartLoc(), L->getHeader())
807 << "loop not distributed: use -Rpass-analysis=loop-distribute for "
812 // With Rpass-analysis report why. This is on by default if distribution
813 // was requested explicitly.
814 ORE->emit(OptimizationRemarkAnalysis(
815 Forced ? OptimizationRemarkAnalysis::AlwaysPrint : LDIST_NAME,
816 RemarkName, L->getStartLoc(), L->getHeader())
817 << "loop not distributed: " << Message);
819 // Also issue a warning if distribution was requested explicitly but it
822 Ctx.diagnose(DiagnosticInfoOptimizationFailure(
823 *F, L->getStartLoc(), "loop not distributed: failed "
824 "explicitly specified loop distribution"));
829 /// \brief Return if distribution forced to be enabled/disabled for the loop.
831 /// If the optional has a value, it indicates whether distribution was forced
832 /// to be enabled (true) or disabled (false). If the optional has no value
833 /// distribution was not forced either way.
834 const Optional<bool> &isForced() const { return IsForced; }
837 /// \brief Filter out checks between pointers from the same partition.
839 /// \p PtrToPartition contains the partition number for pointers. Partition
840 /// number -1 means that the pointer is used in multiple partitions. In this
841 /// case we can't safely omit the check.
842 SmallVector<RuntimePointerChecking::PointerCheck, 4>
843 includeOnlyCrossPartitionChecks(
844 const SmallVectorImpl<RuntimePointerChecking::PointerCheck> &AllChecks,
845 const SmallVectorImpl<int> &PtrToPartition,
846 const RuntimePointerChecking *RtPtrChecking) {
847 SmallVector<RuntimePointerChecking::PointerCheck, 4> Checks;
849 copy_if(AllChecks, std::back_inserter(Checks),
850 [&](const RuntimePointerChecking::PointerCheck &Check) {
851 for (unsigned PtrIdx1 : Check.first->Members)
852 for (unsigned PtrIdx2 : Check.second->Members)
853 // Only include this check if there is a pair of pointers
854 // that require checking and the pointers fall into
855 // separate partitions.
857 // (Note that we already know at this point that the two
858 // pointer groups need checking but it doesn't follow
859 // that each pair of pointers within the two groups need
862 // In other words we don't want to include a check just
863 // because there is a pair of pointers between the two
864 // pointer groups that require checks and a different
865 // pair whose pointers fall into different partitions.)
866 if (RtPtrChecking->needsChecking(PtrIdx1, PtrIdx2) &&
867 !RuntimePointerChecking::arePointersInSamePartition(
868 PtrToPartition, PtrIdx1, PtrIdx2))
876 /// \brief Check whether the loop metadata is forcing distribution to be
877 /// enabled/disabled.
879 Optional<const MDOperand *> Value =
880 findStringMetadataForLoop(L, "llvm.loop.distribute.enable");
884 const MDOperand *Op = *Value;
885 assert(Op && mdconst::hasa<ConstantInt>(*Op) && "invalid metadata");
886 IsForced = mdconst::extract<ConstantInt>(*Op)->getZExtValue();
894 const LoopAccessInfo *LAI = nullptr;
897 OptimizationRemarkEmitter *ORE;
899 /// \brief Indicates whether distribution is forced to be enabled/disabled for
902 /// If the optional has a value, it indicates whether distribution was forced
903 /// to be enabled (true) or disabled (false). If the optional has no value
904 /// distribution was not forced either way.
905 Optional<bool> IsForced;
908 } // end anonymous namespace
910 /// Shared implementation between new and old PMs.
911 static bool runImpl(Function &F, LoopInfo *LI, DominatorTree *DT,
912 ScalarEvolution *SE, OptimizationRemarkEmitter *ORE,
913 std::function<const LoopAccessInfo &(Loop &)> &GetLAA) {
914 // Build up a worklist of inner-loops to vectorize. This is necessary as the
915 // act of distributing a loop creates new loops and can invalidate iterators
917 SmallVector<Loop *, 8> Worklist;
919 for (Loop *TopLevelLoop : *LI)
920 for (Loop *L : depth_first(TopLevelLoop))
921 // We only handle inner-most loops.
923 Worklist.push_back(L);
925 // Now walk the identified inner loops.
926 bool Changed = false;
927 for (Loop *L : Worklist) {
928 LoopDistributeForLoop LDL(L, &F, LI, DT, SE, ORE);
930 // If distribution was forced for the specific loop to be
931 // enabled/disabled, follow that. Otherwise use the global flag.
932 if (LDL.isForced().getValueOr(EnableLoopDistribute))
933 Changed |= LDL.processLoop(GetLAA);
936 // Process each loop nest in the function.
942 /// \brief The pass class.
943 class LoopDistributeLegacy : public FunctionPass {
947 LoopDistributeLegacy() : FunctionPass(ID) {
948 // The default is set by the caller.
949 initializeLoopDistributeLegacyPass(*PassRegistry::getPassRegistry());
952 bool runOnFunction(Function &F) override {
956 auto *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
957 auto *LAA = &getAnalysis<LoopAccessLegacyAnalysis>();
958 auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
959 auto *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
960 auto *ORE = &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE();
961 std::function<const LoopAccessInfo &(Loop &)> GetLAA =
962 [&](Loop &L) -> const LoopAccessInfo & { return LAA->getInfo(&L); };
964 return runImpl(F, LI, DT, SE, ORE, GetLAA);
967 void getAnalysisUsage(AnalysisUsage &AU) const override {
968 AU.addRequired<ScalarEvolutionWrapperPass>();
969 AU.addRequired<LoopInfoWrapperPass>();
970 AU.addPreserved<LoopInfoWrapperPass>();
971 AU.addRequired<LoopAccessLegacyAnalysis>();
972 AU.addRequired<DominatorTreeWrapperPass>();
973 AU.addPreserved<DominatorTreeWrapperPass>();
974 AU.addRequired<OptimizationRemarkEmitterWrapperPass>();
975 AU.addPreserved<GlobalsAAWrapperPass>();
979 } // end anonymous namespace
981 PreservedAnalyses LoopDistributePass::run(Function &F,
982 FunctionAnalysisManager &AM) {
983 auto &LI = AM.getResult<LoopAnalysis>(F);
984 auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
985 auto &SE = AM.getResult<ScalarEvolutionAnalysis>(F);
986 auto &ORE = AM.getResult<OptimizationRemarkEmitterAnalysis>(F);
988 // We don't directly need these analyses but they're required for loop
989 // analyses so provide them below.
990 auto &AA = AM.getResult<AAManager>(F);
991 auto &AC = AM.getResult<AssumptionAnalysis>(F);
992 auto &TTI = AM.getResult<TargetIRAnalysis>(F);
993 auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
995 auto &LAM = AM.getResult<LoopAnalysisManagerFunctionProxy>(F).getManager();
996 std::function<const LoopAccessInfo &(Loop &)> GetLAA =
997 [&](Loop &L) -> const LoopAccessInfo & {
998 LoopStandardAnalysisResults AR = {AA, AC, DT, LI, SE, TLI, TTI, nullptr};
999 return LAM.getResult<LoopAccessAnalysis>(L, AR);
1002 bool Changed = runImpl(F, &LI, &DT, &SE, &ORE, GetLAA);
1004 return PreservedAnalyses::all();
1005 PreservedAnalyses PA;
1006 PA.preserve<LoopAnalysis>();
1007 PA.preserve<DominatorTreeAnalysis>();
1008 PA.preserve<GlobalsAA>();
1012 char LoopDistributeLegacy::ID;
1014 static const char ldist_name[] = "Loop Distribution";
1016 INITIALIZE_PASS_BEGIN(LoopDistributeLegacy, LDIST_NAME, ldist_name, false,
1018 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
1019 INITIALIZE_PASS_DEPENDENCY(LoopAccessLegacyAnalysis)
1020 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
1021 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
1022 INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass)
1023 INITIALIZE_PASS_END(LoopDistributeLegacy, LDIST_NAME, ldist_name, false, false)
1025 FunctionPass *llvm::createLoopDistributePass() { return new LoopDistributeLegacy(); }