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/ADT/DepthFirstIterator.h"
26 #include "llvm/ADT/EquivalenceClasses.h"
27 #include "llvm/ADT/STLExtras.h"
28 #include "llvm/ADT/Statistic.h"
29 #include "llvm/Analysis/LoopAccessAnalysis.h"
30 #include "llvm/Analysis/LoopInfo.h"
31 #include "llvm/IR/Dominators.h"
32 #include "llvm/Pass.h"
33 #include "llvm/Support/CommandLine.h"
34 #include "llvm/Support/Debug.h"
35 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
36 #include "llvm/Transforms/Utils/Cloning.h"
39 #define LDIST_NAME "loop-distribute"
40 #define DEBUG_TYPE LDIST_NAME
45 LDistVerify("loop-distribute-verify", cl::Hidden,
46 cl::desc("Turn on DominatorTree and LoopInfo verification "
47 "after Loop Distribution"),
50 static cl::opt<bool> DistributeNonIfConvertible(
51 "loop-distribute-non-if-convertible", cl::Hidden,
52 cl::desc("Whether to distribute into a loop that may not be "
53 "if-convertible by the loop vectorizer"),
56 STATISTIC(NumLoopsDistributed, "Number of loops distributed");
58 /// \brief Remaps instructions in a loop including the preheader.
59 static void remapInstructionsInLoop(const SmallVectorImpl<BasicBlock *> &Blocks,
60 ValueToValueMapTy &VMap) {
61 // Rewrite the code to refer to itself.
62 for (auto *BB : Blocks)
63 for (auto &Inst : *BB)
64 RemapInstruction(&Inst, VMap,
65 RF_NoModuleLevelChanges | RF_IgnoreMissingEntries);
68 /// \brief Clones a loop \p OrigLoop. Returns the loop and the blocks in \p
71 /// Updates LoopInfo and DominatorTree assuming the loop is dominated by block
72 /// \p LoopDomBB. Insert the new blocks before block specified in \p Before.
73 static Loop *cloneLoopWithPreheader(BasicBlock *Before, BasicBlock *LoopDomBB,
74 Loop *OrigLoop, ValueToValueMapTy &VMap,
75 const Twine &NameSuffix, LoopInfo *LI,
77 SmallVectorImpl<BasicBlock *> &Blocks) {
78 Function *F = OrigLoop->getHeader()->getParent();
79 Loop *ParentLoop = OrigLoop->getParentLoop();
81 Loop *NewLoop = new Loop();
83 ParentLoop->addChildLoop(NewLoop);
85 LI->addTopLevelLoop(NewLoop);
87 BasicBlock *OrigPH = OrigLoop->getLoopPreheader();
88 BasicBlock *NewPH = CloneBasicBlock(OrigPH, VMap, NameSuffix, F);
89 // To rename the loop PHIs.
91 Blocks.push_back(NewPH);
95 ParentLoop->addBasicBlockToLoop(NewPH, *LI);
97 // Update DominatorTree.
98 DT->addNewBlock(NewPH, LoopDomBB);
100 for (BasicBlock *BB : OrigLoop->getBlocks()) {
101 BasicBlock *NewBB = CloneBasicBlock(BB, VMap, NameSuffix, F);
105 NewLoop->addBasicBlockToLoop(NewBB, *LI);
107 // Update DominatorTree.
108 BasicBlock *IDomBB = DT->getNode(BB)->getIDom()->getBlock();
109 DT->addNewBlock(NewBB, cast<BasicBlock>(VMap[IDomBB]));
111 Blocks.push_back(NewBB);
114 // Move them physically from the end of the block list.
115 F->getBasicBlockList().splice(Before, F->getBasicBlockList(), NewPH);
116 F->getBasicBlockList().splice(Before, F->getBasicBlockList(),
117 NewLoop->getHeader(), F->end());
123 /// \brief Maintains the set of instructions of the loop for a partition before
124 /// cloning. After cloning, it hosts the new loop.
125 class InstPartition {
126 typedef SmallPtrSet<Instruction *, 8> InstructionSet;
129 InstPartition(Instruction *I, Loop *L, bool DepCycle = false)
130 : DepCycle(DepCycle), OrigLoop(L), ClonedLoop(nullptr) {
134 /// \brief Returns whether this partition contains a dependence cycle.
135 bool hasDepCycle() const { return DepCycle; }
137 /// \brief Adds an instruction to this partition.
138 void add(Instruction *I) { Set.insert(I); }
140 /// \brief Collection accessors.
141 InstructionSet::iterator begin() { return Set.begin(); }
142 InstructionSet::iterator end() { return Set.end(); }
143 InstructionSet::const_iterator begin() const { return Set.begin(); }
144 InstructionSet::const_iterator end() const { return Set.end(); }
145 bool empty() const { return Set.empty(); }
147 /// \brief Moves this partition into \p Other. This partition becomes empty
149 void moveTo(InstPartition &Other) {
150 Other.Set.insert(Set.begin(), Set.end());
152 Other.DepCycle |= DepCycle;
155 /// \brief Populates the partition with a transitive closure of all the
156 /// instructions that the seeded instructions dependent on.
157 void populateUsedSet() {
158 // FIXME: We currently don't use control-dependence but simply include all
159 // blocks (possibly empty at the end) and let simplifycfg mostly clean this
161 for (auto *B : OrigLoop->getBlocks())
162 Set.insert(B->getTerminator());
164 // Follow the use-def chains to form a transitive closure of all the
165 // instructions that the originally seeded instructions depend on.
166 SmallVector<Instruction *, 8> Worklist(Set.begin(), Set.end());
167 while (!Worklist.empty()) {
168 Instruction *I = Worklist.pop_back_val();
169 // Insert instructions from the loop that we depend on.
170 for (Value *V : I->operand_values()) {
171 auto *I = dyn_cast<Instruction>(V);
172 if (I && OrigLoop->contains(I->getParent()) && Set.insert(I).second)
173 Worklist.push_back(I);
178 /// \brief Clones the original loop.
180 /// Updates LoopInfo and DominatorTree using the information that block \p
181 /// LoopDomBB dominates the loop.
182 Loop *cloneLoopWithPreheader(BasicBlock *InsertBefore, BasicBlock *LoopDomBB,
183 unsigned Index, LoopInfo *LI,
185 ClonedLoop = ::cloneLoopWithPreheader(InsertBefore, LoopDomBB, OrigLoop,
186 VMap, Twine(".ldist") + Twine(Index),
187 LI, DT, ClonedLoopBlocks);
191 /// \brief The cloned loop. If this partition is mapped to the original loop,
193 const Loop *getClonedLoop() const { return ClonedLoop; }
195 /// \brief Returns the loop where this partition ends up after distribution.
196 /// If this partition is mapped to the original loop then use the block from
198 const Loop *getDistributedLoop() const {
199 return ClonedLoop ? ClonedLoop : OrigLoop;
202 /// \brief The VMap that is populated by cloning and then used in
203 /// remapinstruction to remap the cloned instructions.
204 ValueToValueMapTy &getVMap() { return VMap; }
206 /// \brief Remaps the cloned instructions using VMap.
207 void remapInstructions() { remapInstructionsInLoop(ClonedLoopBlocks, VMap); }
209 /// \brief Based on the set of instructions selected for this partition,
210 /// removes the unnecessary ones.
211 void removeUnusedInsts() {
212 SmallVector<Instruction *, 8> Unused;
214 for (auto *Block : OrigLoop->getBlocks())
215 for (auto &Inst : *Block)
216 if (!Set.count(&Inst)) {
217 Instruction *NewInst = &Inst;
219 NewInst = cast<Instruction>(VMap[NewInst]);
221 assert(!isa<BranchInst>(NewInst) &&
222 "Branches are marked used early on");
223 Unused.push_back(NewInst);
226 // Delete the instructions backwards, as it has a reduced likelihood of
227 // having to update as many def-use and use-def chains.
228 for (auto I = Unused.rbegin(), E = Unused.rend(); I != E; ++I) {
231 if (!Inst->use_empty())
232 Inst->replaceAllUsesWith(UndefValue::get(Inst->getType()));
233 Inst->eraseFromParent();
239 dbgs() << " (cycle)\n";
241 // Prefix with the block name.
242 dbgs() << " " << I->getParent()->getName() << ":" << *I << "\n";
245 void printBlocks() const {
246 for (auto *BB : getDistributedLoop()->getBlocks())
251 /// \brief Instructions from OrigLoop selected for this partition.
254 /// \brief Whether this partition contains a dependence cycle.
257 /// \brief The original loop.
260 /// \brief The cloned loop. If this partition is mapped to the original loop,
264 /// \brief The blocks of ClonedLoop including the preheader. If this
265 /// partition is mapped to the original loop, this is empty.
266 SmallVector<BasicBlock *, 8> ClonedLoopBlocks;
268 /// \brief These gets populated once the set of instructions have been
269 /// finalized. If this partition is mapped to the original loop, these are not
271 ValueToValueMapTy VMap;
274 /// \brief Holds the set of Partitions. It populates them, merges them and then
275 /// clones the loops.
276 class InstPartitionContainer {
277 typedef DenseMap<Instruction *, int> InstToPartitionIdT;
280 InstPartitionContainer(Loop *L, LoopInfo *LI, DominatorTree *DT)
281 : L(L), LI(LI), DT(DT) {}
283 /// \brief Returns the number of partitions.
284 unsigned getSize() const { return PartitionContainer.size(); }
286 /// \brief Adds \p Inst into the current partition if that is marked to
287 /// contain cycles. Otherwise start a new partition for it.
288 void addToCyclicPartition(Instruction *Inst) {
289 // If the current partition is non-cyclic. Start a new one.
290 if (PartitionContainer.empty() || !PartitionContainer.back().hasDepCycle())
291 PartitionContainer.emplace_back(Inst, L, /*DepCycle=*/true);
293 PartitionContainer.back().add(Inst);
296 /// \brief Adds \p Inst into a partition that is not marked to contain
297 /// dependence cycles.
299 // Initially we isolate memory instructions into as many partitions as
300 // possible, then later we may merge them back together.
301 void addToNewNonCyclicPartition(Instruction *Inst) {
302 PartitionContainer.emplace_back(Inst, L);
305 /// \brief Merges adjacent non-cyclic partitions.
307 /// The idea is that we currently only want to isolate the non-vectorizable
308 /// partition. We could later allow more distribution among these partition
310 void mergeAdjacentNonCyclic() {
311 mergeAdjacentPartitionsIf(
312 [](const InstPartition *P) { return !P->hasDepCycle(); });
315 /// \brief If a partition contains only conditional stores, we won't vectorize
316 /// it. Try to merge it with a previous cyclic partition.
317 void mergeNonIfConvertible() {
318 mergeAdjacentPartitionsIf([&](const InstPartition *Partition) {
319 if (Partition->hasDepCycle())
322 // Now, check if all stores are conditional in this partition.
323 bool seenStore = false;
325 for (auto *Inst : *Partition)
326 if (isa<StoreInst>(Inst)) {
328 if (!LoopAccessInfo::blockNeedsPredication(Inst->getParent(), L, DT))
335 /// \brief Merges the partitions according to various heuristics.
336 void mergeBeforePopulating() {
337 mergeAdjacentNonCyclic();
338 if (!DistributeNonIfConvertible)
339 mergeNonIfConvertible();
342 /// \brief Merges partitions in order to ensure that no loads are duplicated.
344 /// We can't duplicate loads because that could potentially reorder them.
345 /// LoopAccessAnalysis provides dependency information with the context that
346 /// the order of memory operation is preserved.
348 /// Return if any partitions were merged.
349 bool mergeToAvoidDuplicatedLoads() {
350 typedef DenseMap<Instruction *, InstPartition *> LoadToPartitionT;
351 typedef EquivalenceClasses<InstPartition *> ToBeMergedT;
353 LoadToPartitionT LoadToPartition;
354 ToBeMergedT ToBeMerged;
356 // Step through the partitions and create equivalence between partitions
357 // that contain the same load. Also put partitions in between them in the
358 // same equivalence class to avoid reordering of memory operations.
359 for (PartitionContainerT::iterator I = PartitionContainer.begin(),
360 E = PartitionContainer.end();
364 // If a load occurs in two partitions PartI and PartJ, merge all
365 // partitions (PartI, PartJ] into PartI.
366 for (Instruction *Inst : *PartI)
367 if (isa<LoadInst>(Inst)) {
369 LoadToPartitionT::iterator LoadToPart;
371 std::tie(LoadToPart, NewElt) =
372 LoadToPartition.insert(std::make_pair(Inst, PartI));
374 DEBUG(dbgs() << "Merging partitions due to this load in multiple "
375 << "partitions: " << PartI << ", "
376 << LoadToPart->second << "\n" << *Inst << "\n");
381 ToBeMerged.unionSets(PartI, &*PartJ);
382 } while (&*PartJ != LoadToPart->second);
386 if (ToBeMerged.empty())
389 // Merge the member of an equivalence class into its class leader. This
390 // makes the members empty.
391 for (ToBeMergedT::iterator I = ToBeMerged.begin(), E = ToBeMerged.end();
396 auto PartI = I->getData();
397 for (auto PartJ : make_range(std::next(ToBeMerged.member_begin(I)),
398 ToBeMerged.member_end())) {
399 PartJ->moveTo(*PartI);
403 // Remove the empty partitions.
404 PartitionContainer.remove_if(
405 [](const InstPartition &P) { return P.empty(); });
410 /// \brief Sets up the mapping between instructions to partitions. If the
411 /// instruction is duplicated across multiple partitions, set the entry to -1.
412 void setupPartitionIdOnInstructions() {
414 for (const auto &Partition : PartitionContainer) {
415 for (Instruction *Inst : Partition) {
417 InstToPartitionIdT::iterator Iter;
419 std::tie(Iter, NewElt) =
420 InstToPartitionId.insert(std::make_pair(Inst, PartitionID));
428 /// \brief Populates the partition with everything that the seeding
429 /// instructions require.
430 void populateUsedSet() {
431 for (auto &P : PartitionContainer)
435 /// \brief This performs the main chunk of the work of cloning the loops for
437 void cloneLoops(Pass *P) {
438 BasicBlock *OrigPH = L->getLoopPreheader();
439 // At this point the predecessor of the preheader is either the memcheck
440 // block or the top part of the original preheader.
441 BasicBlock *Pred = OrigPH->getSinglePredecessor();
442 assert(Pred && "Preheader does not have a single predecessor");
443 BasicBlock *ExitBlock = L->getExitBlock();
444 assert(ExitBlock && "No single exit block");
447 assert(!PartitionContainer.empty() && "at least two partitions expected");
448 // We're cloning the preheader along with the loop so we already made sure
450 assert(&*OrigPH->begin() == OrigPH->getTerminator() &&
451 "preheader not empty");
453 // Create a loop for each partition except the last. Clone the original
454 // loop before PH along with adding a preheader for the cloned loop. Then
455 // update PH to point to the newly added preheader.
456 BasicBlock *TopPH = OrigPH;
457 unsigned Index = getSize() - 1;
458 for (auto I = std::next(PartitionContainer.rbegin()),
459 E = PartitionContainer.rend();
460 I != E; ++I, --Index, TopPH = NewLoop->getLoopPreheader()) {
463 NewLoop = Part->cloneLoopWithPreheader(TopPH, Pred, Index, LI, DT);
465 Part->getVMap()[ExitBlock] = TopPH;
466 Part->remapInstructions();
468 Pred->getTerminator()->replaceUsesOfWith(OrigPH, TopPH);
470 // Now go in forward order and update the immediate dominator for the
471 // preheaders with the exiting block of the previous loop. Dominance
472 // within the loop is updated in cloneLoopWithPreheader.
473 for (auto Curr = PartitionContainer.cbegin(),
474 Next = std::next(PartitionContainer.cbegin()),
475 E = PartitionContainer.cend();
476 Next != E; ++Curr, ++Next)
477 DT->changeImmediateDominator(
478 Next->getDistributedLoop()->getLoopPreheader(),
479 Curr->getDistributedLoop()->getExitingBlock());
482 /// \brief Removes the dead instructions from the cloned loops.
483 void removeUnusedInsts() {
484 for (auto &Partition : PartitionContainer)
485 Partition.removeUnusedInsts();
488 /// \brief For each memory pointer, it computes the partitionId the pointer is
491 /// This returns an array of int where the I-th entry corresponds to I-th
492 /// entry in LAI.getRuntimePointerCheck(). If the pointer is used in multiple
493 /// partitions its entry is set to -1.
495 computePartitionSetForPointers(const LoopAccessInfo &LAI) {
496 const LoopAccessInfo::RuntimePointerCheck *RtPtrCheck =
497 LAI.getRuntimePointerCheck();
499 unsigned N = RtPtrCheck->Pointers.size();
500 SmallVector<int, 8> PtrToPartitions(N);
501 for (unsigned I = 0; I < N; ++I) {
502 Value *Ptr = RtPtrCheck->Pointers[I];
504 LAI.getInstructionsForAccess(Ptr, RtPtrCheck->IsWritePtr[I]);
506 int &Partition = PtrToPartitions[I];
507 // First set it to uninitialized.
509 for (Instruction *Inst : Instructions) {
510 // Note that this could be -1 if Inst is duplicated across multiple
512 int ThisPartition = this->InstToPartitionId[Inst];
514 Partition = ThisPartition;
515 // -1 means belonging to multiple partitions.
516 else if (Partition == -1)
518 else if (Partition != (int)ThisPartition)
521 assert(Partition != -2 && "Pointer not belonging to any partition");
524 return PtrToPartitions;
527 void print(raw_ostream &OS) const {
529 for (const auto &P : PartitionContainer) {
530 OS << "Partition " << Index++ << " (" << &P << "):\n";
535 void dump() const { print(dbgs()); }
538 friend raw_ostream &operator<<(raw_ostream &OS,
539 const InstPartitionContainer &Partitions) {
540 Partitions.print(OS);
545 void printBlocks() const {
547 for (const auto &P : PartitionContainer) {
548 dbgs() << "\nPartition " << Index++ << " (" << &P << "):\n";
554 typedef std::list<InstPartition> PartitionContainerT;
556 /// \brief List of partitions.
557 PartitionContainerT PartitionContainer;
559 /// \brief Mapping from Instruction to partition Id. If the instruction
560 /// belongs to multiple partitions the entry contains -1.
561 InstToPartitionIdT InstToPartitionId;
567 /// \brief The control structure to merge adjacent partitions if both satisfy
568 /// the \p Predicate.
569 template <class UnaryPredicate>
570 void mergeAdjacentPartitionsIf(UnaryPredicate Predicate) {
571 InstPartition *PrevMatch = nullptr;
572 for (auto I = PartitionContainer.begin(); I != PartitionContainer.end();) {
573 auto DoesMatch = Predicate(&*I);
574 if (PrevMatch == nullptr && DoesMatch) {
577 } else if (PrevMatch != nullptr && DoesMatch) {
578 I->moveTo(*PrevMatch);
579 I = PartitionContainer.erase(I);
588 /// \brief For each memory instruction, this class maintains difference of the
589 /// number of unsafe dependences that start out from this instruction minus
590 /// those that end here.
592 /// By traversing the memory instructions in program order and accumulating this
593 /// number, we know whether any unsafe dependence crosses over a program point.
594 class MemoryInstructionDependences {
595 typedef MemoryDepChecker::Dependence Dependence;
600 unsigned NumUnsafeDependencesStartOrEnd;
602 Entry(Instruction *Inst) : Inst(Inst), NumUnsafeDependencesStartOrEnd(0) {}
605 typedef SmallVector<Entry, 8> AccessesType;
607 AccessesType::const_iterator begin() const { return Accesses.begin(); }
608 AccessesType::const_iterator end() const { return Accesses.end(); }
610 MemoryInstructionDependences(
611 const SmallVectorImpl<Instruction *> &Instructions,
612 const SmallVectorImpl<Dependence> &InterestingDependences) {
613 Accesses.append(Instructions.begin(), Instructions.end());
615 DEBUG(dbgs() << "Backward dependences:\n");
616 for (auto &Dep : InterestingDependences)
617 if (Dep.isPossiblyBackward()) {
618 // Note that the designations source and destination follow the program
619 // order, i.e. source is always first. (The direction is given by the
621 ++Accesses[Dep.Source].NumUnsafeDependencesStartOrEnd;
622 --Accesses[Dep.Destination].NumUnsafeDependencesStartOrEnd;
624 DEBUG(Dep.print(dbgs(), 2, Instructions));
629 AccessesType Accesses;
632 /// \brief Handles the loop versioning based on memchecks.
633 class RuntimeCheckEmitter {
635 RuntimeCheckEmitter(const LoopAccessInfo &LAI, Loop *L, LoopInfo *LI,
637 const SmallVector<int, 8> *PtrToPartition = nullptr)
638 : OrigLoop(L), NonDistributedLoop(nullptr),
639 PtrToPartition(PtrToPartition), LAI(LAI), LI(LI), DT(DT) {}
641 /// \brief Returns true if we need memchecks to distribute the loop.
642 bool needsRuntimeChecks() const {
643 return LAI.getRuntimePointerCheck()->needsAnyChecking(PtrToPartition);
646 /// \brief Performs the CFG manipulation part of versioning the loop including
647 /// the DominatorTree and LoopInfo updates.
648 void versionLoop(Pass *P) {
649 Instruction *FirstCheckInst;
650 Instruction *MemRuntimeCheck;
651 // Add the memcheck in the original preheader (this is empty initially).
652 BasicBlock *MemCheckBB = OrigLoop->getLoopPreheader();
653 std::tie(FirstCheckInst, MemRuntimeCheck) =
654 LAI.addRuntimeCheck(MemCheckBB->getTerminator(), PtrToPartition);
655 assert(MemRuntimeCheck && "called even though needsAnyChecking = false");
657 // Rename the block to make the IR more readable.
658 MemCheckBB->setName(OrigLoop->getHeader()->getName() + ".ldist.memcheck");
660 // Create empty preheader for the loop (and after cloning for the
661 // original/nondist loop).
663 SplitBlock(MemCheckBB, MemCheckBB->getTerminator(), DT, LI);
664 PH->setName(OrigLoop->getHeader()->getName() + ".ph");
666 // Clone the loop including the preheader.
668 // FIXME: This does not currently preserve SimplifyLoop because the exit
669 // block is a join between the two loops.
670 SmallVector<BasicBlock *, 8> NonDistributedLoopBlocks;
672 cloneLoopWithPreheader(PH, MemCheckBB, OrigLoop, VMap, ".ldist.nondist",
673 LI, DT, NonDistributedLoopBlocks);
674 remapInstructionsInLoop(NonDistributedLoopBlocks, VMap);
676 // Insert the conditional branch based on the result of the memchecks.
677 Instruction *OrigTerm = MemCheckBB->getTerminator();
678 BranchInst::Create(NonDistributedLoop->getLoopPreheader(),
679 OrigLoop->getLoopPreheader(), MemRuntimeCheck, OrigTerm);
680 OrigTerm->eraseFromParent();
682 // The loops merge in the original exit block. This is now dominated by the
683 // memchecking block.
684 DT->changeImmediateDominator(OrigLoop->getExitBlock(), MemCheckBB);
687 /// \brief Adds the necessary PHI nodes for the versioned loops based on the
688 /// loop-defined values used outside of the loop.
689 void addPHINodes(const SmallVectorImpl<Instruction *> &DefsUsedOutside) {
690 BasicBlock *PHIBlock = OrigLoop->getExitBlock();
691 assert(PHIBlock && "No single successor to loop exit block");
693 for (auto *Inst : DefsUsedOutside) {
694 auto *NonDistInst = cast<Instruction>(VMap[Inst]);
697 // First see if we have a single-operand PHI with the value defined by the
699 for (auto I = PHIBlock->begin(); (PN = dyn_cast<PHINode>(I)); ++I) {
700 assert(PN->getNumOperands() == 1 &&
701 "Exit block should only have on predecessor");
702 if (PN->getIncomingValue(0) == Inst)
707 PN = PHINode::Create(Inst->getType(), 2, Inst->getName() + ".ldist",
709 for (auto *User : Inst->users())
710 if (!OrigLoop->contains(cast<Instruction>(User)->getParent()))
711 User->replaceUsesOfWith(Inst, PN);
712 PN->addIncoming(Inst, OrigLoop->getExitingBlock());
714 // Add the new incoming value from the non-distributed loop.
715 PN->addIncoming(NonDistInst, NonDistributedLoop->getExitingBlock());
720 /// \brief The original loop. This becomes the "versioned" one, i.e. control
721 /// goes if the memchecks all pass.
723 /// \brief The fall-back loop, i.e. if any of the memchecks fail.
724 Loop *NonDistributedLoop;
726 /// \brief For each memory pointer it contains the partitionId it is used in.
727 /// If nullptr, no partitioning is used.
729 /// The I-th entry corresponds to I-th entry in LAI.getRuntimePointerCheck().
730 /// If the pointer is used in multiple partitions the entry is set to -1.
731 const SmallVector<int, 8> *PtrToPartition;
733 /// \brief This maps the instructions from OrigLoop to their counterpart in
734 /// NonDistributedLoop.
735 ValueToValueMapTy VMap;
737 /// \brief Analyses used.
738 const LoopAccessInfo &LAI;
743 /// \brief Returns the instructions that use values defined in the loop.
744 static SmallVector<Instruction *, 8> findDefsUsedOutsideOfLoop(Loop *L) {
745 SmallVector<Instruction *, 8> UsedOutside;
747 for (auto *Block : L->getBlocks())
748 // FIXME: I believe that this could use copy_if if the Inst reference could
749 // be adapted into a pointer.
750 for (auto &Inst : *Block) {
751 auto Users = Inst.users();
752 if (std::any_of(Users.begin(), Users.end(), [&](User *U) {
753 auto *Use = cast<Instruction>(U);
754 return !L->contains(Use->getParent());
756 UsedOutside.push_back(&Inst);
762 /// \brief The pass class.
763 class LoopDistribute : public FunctionPass {
765 LoopDistribute() : FunctionPass(ID) {
766 initializeLoopDistributePass(*PassRegistry::getPassRegistry());
769 bool runOnFunction(Function &F) override {
770 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
771 LAA = &getAnalysis<LoopAccessAnalysis>();
772 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
774 // Build up a worklist of inner-loops to vectorize. This is necessary as the
775 // act of distributing a loop creates new loops and can invalidate iterators
777 SmallVector<Loop *, 8> Worklist;
779 for (Loop *TopLevelLoop : *LI)
780 for (Loop *L : depth_first(TopLevelLoop))
781 // We only handle inner-most loops.
783 Worklist.push_back(L);
785 // Now walk the identified inner loops.
786 bool Changed = false;
787 for (Loop *L : Worklist)
788 Changed |= processLoop(L);
790 // Process each loop nest in the function.
794 void getAnalysisUsage(AnalysisUsage &AU) const override {
795 AU.addRequired<LoopInfoWrapperPass>();
796 AU.addPreserved<LoopInfoWrapperPass>();
797 AU.addRequired<LoopAccessAnalysis>();
798 AU.addRequired<DominatorTreeWrapperPass>();
799 AU.addPreserved<DominatorTreeWrapperPass>();
805 /// \brief Try to distribute an inner-most loop.
806 bool processLoop(Loop *L) {
807 assert(L->empty() && "Only process inner loops.");
809 DEBUG(dbgs() << "\nLDist: In \"" << L->getHeader()->getParent()->getName()
810 << "\" checking " << *L << "\n");
812 BasicBlock *PH = L->getLoopPreheader();
814 DEBUG(dbgs() << "Skipping; no preheader");
817 if (!L->getExitBlock()) {
818 DEBUG(dbgs() << "Skipping; multiple exit blocks");
821 // LAA will check that we only have a single exiting block.
823 const LoopAccessInfo &LAI = LAA->getInfo(L, ValueToValueMap());
825 // Currently, we only distribute to isolate the part of the loop with
826 // dependence cycles to enable partial vectorization.
827 if (LAI.canVectorizeMemory()) {
828 DEBUG(dbgs() << "Skipping; memory operations are safe for vectorization");
831 auto *InterestingDependences =
832 LAI.getDepChecker().getInterestingDependences();
833 if (!InterestingDependences || InterestingDependences->empty()) {
834 DEBUG(dbgs() << "Skipping; No unsafe dependences to isolate");
838 InstPartitionContainer Partitions(L, LI, DT);
840 // First, go through each memory operation and assign them to consecutive
841 // partitions (the order of partitions follows program order). Put those
842 // with unsafe dependences into "cyclic" partition otherwise put each store
843 // in its own "non-cyclic" partition (we'll merge these later).
845 // Note that a memory operation (e.g. Load2 below) at a program point that
846 // has an unsafe dependence (Store3->Load1) spanning over it must be
847 // included in the same cyclic partition as the dependent operations. This
848 // is to preserve the original program order after distribution. E.g.:
850 // NumUnsafeDependencesStartOrEnd NumUnsafeDependencesActive
852 // Load2 | /Unsafe/ 0 1
856 // NumUnsafeDependencesActive > 0 indicates this situation and in this case
857 // we just keep assigning to the same cyclic partition until
858 // NumUnsafeDependencesActive reaches 0.
859 const MemoryDepChecker &DepChecker = LAI.getDepChecker();
860 MemoryInstructionDependences MID(DepChecker.getMemoryInstructions(),
861 *InterestingDependences);
863 int NumUnsafeDependencesActive = 0;
864 for (auto &InstDep : MID) {
865 Instruction *I = InstDep.Inst;
866 // We update NumUnsafeDependencesActive post-instruction, catch the
867 // start of a dependence directly via NumUnsafeDependencesStartOrEnd.
868 if (NumUnsafeDependencesActive ||
869 InstDep.NumUnsafeDependencesStartOrEnd > 0)
870 Partitions.addToCyclicPartition(I);
872 Partitions.addToNewNonCyclicPartition(I);
873 NumUnsafeDependencesActive += InstDep.NumUnsafeDependencesStartOrEnd;
874 assert(NumUnsafeDependencesActive >= 0 &&
875 "Negative number of dependences active");
878 // Add partitions for values used outside. These partitions can be out of
879 // order from the original program order. This is OK because if the
880 // partition uses a load we will merge this partition with the original
881 // partition of the load that we set up in the previous loop (see
882 // mergeToAvoidDuplicatedLoads).
883 auto DefsUsedOutside = findDefsUsedOutsideOfLoop(L);
884 for (auto *Inst : DefsUsedOutside)
885 Partitions.addToNewNonCyclicPartition(Inst);
887 DEBUG(dbgs() << "Seeded partitions:\n" << Partitions);
888 if (Partitions.getSize() < 2)
891 // Run the merge heuristics: Merge non-cyclic adjacent partitions since we
892 // should be able to vectorize these together.
893 Partitions.mergeBeforePopulating();
894 DEBUG(dbgs() << "\nMerged partitions:\n" << Partitions);
895 if (Partitions.getSize() < 2)
898 // Now, populate the partitions with non-memory operations.
899 Partitions.populateUsedSet();
900 DEBUG(dbgs() << "\nPopulated partitions:\n" << Partitions);
902 // In order to preserve original lexical order for loads, keep them in the
903 // partition that we set up in the MemoryInstructionDependences loop.
904 if (Partitions.mergeToAvoidDuplicatedLoads()) {
905 DEBUG(dbgs() << "\nPartitions merged to ensure unique loads:\n"
907 if (Partitions.getSize() < 2)
911 DEBUG(dbgs() << "\nDistributing loop: " << *L << "\n");
912 // We're done forming the partitions set up the reverse mapping from
913 // instructions to partitions.
914 Partitions.setupPartitionIdOnInstructions();
916 // To keep things simple have an empty preheader before we version or clone
917 // the loop. (Also split if this has no predecessor, i.e. entry, because we
918 // rely on PH having a predecessor.)
919 if (!PH->getSinglePredecessor() || &*PH->begin() != PH->getTerminator())
920 SplitBlock(PH, PH->getTerminator(), DT, LI);
922 // If we need run-time checks to disambiguate pointers are run-time, version
924 auto PtrToPartition = Partitions.computePartitionSetForPointers(LAI);
925 RuntimeCheckEmitter RtCheckEmitter(LAI, L, LI, DT, &PtrToPartition);
926 if (RtCheckEmitter.needsRuntimeChecks()) {
927 DEBUG(dbgs() << "\nPointers:\n");
928 DEBUG(LAI.getRuntimePointerCheck()->print(dbgs(), 0, &PtrToPartition));
929 RtCheckEmitter.versionLoop(this);
930 RtCheckEmitter.addPHINodes(DefsUsedOutside);
933 // Create identical copies of the original loop for each partition and hook
934 // them up sequentially.
935 Partitions.cloneLoops(this);
937 // Now, we remove the instruction from each loop that don't belong to that
939 Partitions.removeUnusedInsts();
940 DEBUG(dbgs() << "\nAfter removing unused Instrs:\n");
941 DEBUG(Partitions.printBlocks());
948 ++NumLoopsDistributed;
954 LoopAccessAnalysis *LAA;
957 } // anonymous namespace
959 char LoopDistribute::ID;
960 static const char ldist_name[] = "Loop Distribition";
962 INITIALIZE_PASS_BEGIN(LoopDistribute, LDIST_NAME, ldist_name, false, false)
963 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
964 INITIALIZE_PASS_DEPENDENCY(LoopAccessAnalysis)
965 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
966 INITIALIZE_PASS_END(LoopDistribute, LDIST_NAME, ldist_name, false, false)
969 FunctionPass *createLoopDistributePass() { return new LoopDistribute(); }