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