• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 //===- LoopDistribute.cpp - Loop Distribution Pass ------------------------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
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.
14 //
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
18 // these operations.
19 //
20 // Similarly to the Vectorizer, the pass also supports loop versioning to
21 // run-time disambiguate potentially overlapping arrays.
22 //
23 //===----------------------------------------------------------------------===//
24 
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"
70 #include <cassert>
71 #include <functional>
72 #include <list>
73 #include <tuple>
74 #include <utility>
75 
76 using namespace llvm;
77 
78 #define LDIST_NAME "loop-distribute"
79 #define DEBUG_TYPE LDIST_NAME
80 
81 static cl::opt<bool>
82     LDistVerify("loop-distribute-verify", cl::Hidden,
83                 cl::desc("Turn on DominatorTree and LoopInfo verification "
84                          "after Loop Distribution"),
85                 cl::init(false));
86 
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"),
91     cl::init(false));
92 
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 "
96              "Distribution"));
97 
98 static cl::opt<unsigned> PragmaDistributeSCEVCheckThreshold(
99     "loop-distribute-scev-check-threshold-with-pragma", cl::init(128),
100     cl::Hidden,
101     cl::desc(
102         "The maximum number of SCEV checks allowed for Loop "
103         "Distribution for loop marked with #pragma loop distribute(enable)"));
104 
105 static cl::opt<bool> EnableLoopDistribute(
106     "enable-loop-distribute", cl::Hidden,
107     cl::desc("Enable the new, experimental LoopDistribution Pass"),
108     cl::init(false));
109 
110 STATISTIC(NumLoopsDistributed, "Number of loops distributed");
111 
112 namespace {
113 
114 /// 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>;
118 
119 public:
InstPartition(Instruction * I,Loop * L,bool DepCycle=false)120   InstPartition(Instruction *I, Loop *L, bool DepCycle = false)
121       : DepCycle(DepCycle), OrigLoop(L) {
122     Set.insert(I);
123   }
124 
125   /// Returns whether this partition contains a dependence cycle.
hasDepCycle() const126   bool hasDepCycle() const { return DepCycle; }
127 
128   /// Adds an instruction to this partition.
add(Instruction * I)129   void add(Instruction *I) { Set.insert(I); }
130 
131   /// Collection accessors.
begin()132   InstructionSet::iterator begin() { return Set.begin(); }
end()133   InstructionSet::iterator end() { return Set.end(); }
begin() const134   InstructionSet::const_iterator begin() const { return Set.begin(); }
end() const135   InstructionSet::const_iterator end() const { return Set.end(); }
empty() const136   bool empty() const { return Set.empty(); }
137 
138   /// Moves this partition into \p Other.  This partition becomes empty
139   /// after this.
moveTo(InstPartition & Other)140   void moveTo(InstPartition &Other) {
141     Other.Set.insert(Set.begin(), Set.end());
142     Set.clear();
143     Other.DepCycle |= DepCycle;
144   }
145 
146   /// Populates the partition with a transitive closure of all the
147   /// instructions that the seeded instructions dependent on.
populateUsedSet()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
151     // up.
152     for (auto *B : OrigLoop->getBlocks())
153       Set.insert(B->getTerminator());
154 
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);
165       }
166     }
167   }
168 
169   /// Clones the original loop.
170   ///
171   /// Updates LoopInfo and DominatorTree using the information that block \p
172   /// LoopDomBB dominates the loop.
cloneLoopWithPreheader(BasicBlock * InsertBefore,BasicBlock * LoopDomBB,unsigned Index,LoopInfo * LI,DominatorTree * DT)173   Loop *cloneLoopWithPreheader(BasicBlock *InsertBefore, BasicBlock *LoopDomBB,
174                                unsigned Index, LoopInfo *LI,
175                                DominatorTree *DT) {
176     ClonedLoop = ::cloneLoopWithPreheader(InsertBefore, LoopDomBB, OrigLoop,
177                                           VMap, Twine(".ldist") + Twine(Index),
178                                           LI, DT, ClonedLoopBlocks);
179     return ClonedLoop;
180   }
181 
182   /// The cloned loop.  If this partition is mapped to the original loop,
183   /// this is null.
getClonedLoop() const184   const Loop *getClonedLoop() const { return ClonedLoop; }
185 
186   /// 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
188   /// the loop.
getDistributedLoop() const189   const Loop *getDistributedLoop() const {
190     return ClonedLoop ? ClonedLoop : OrigLoop;
191   }
192 
193   /// The VMap that is populated by cloning and then used in
194   /// remapinstruction to remap the cloned instructions.
getVMap()195   ValueToValueMapTy &getVMap() { return VMap; }
196 
197   /// Remaps the cloned instructions using VMap.
remapInstructions()198   void remapInstructions() {
199     remapInstructionsInBlocks(ClonedLoopBlocks, VMap);
200   }
201 
202   /// Based on the set of instructions selected for this partition,
203   /// removes the unnecessary ones.
removeUnusedInsts()204   void removeUnusedInsts() {
205     SmallVector<Instruction *, 8> Unused;
206 
207     for (auto *Block : OrigLoop->getBlocks())
208       for (auto &Inst : *Block)
209         if (!Set.count(&Inst)) {
210           Instruction *NewInst = &Inst;
211           if (!VMap.empty())
212             NewInst = cast<Instruction>(VMap[NewInst]);
213 
214           assert(!isa<BranchInst>(NewInst) &&
215                  "Branches are marked used early on");
216           Unused.push_back(NewInst);
217         }
218 
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();
225     }
226   }
227 
print() const228   void print() const {
229     if (DepCycle)
230       dbgs() << "  (cycle)\n";
231     for (auto *I : Set)
232       // Prefix with the block name.
233       dbgs() << "  " << I->getParent()->getName() << ":" << *I << "\n";
234   }
235 
printBlocks() const236   void printBlocks() const {
237     for (auto *BB : getDistributedLoop()->getBlocks())
238       dbgs() << *BB;
239   }
240 
241 private:
242   /// Instructions from OrigLoop selected for this partition.
243   InstructionSet Set;
244 
245   /// Whether this partition contains a dependence cycle.
246   bool DepCycle;
247 
248   /// The original loop.
249   Loop *OrigLoop;
250 
251   /// The cloned loop.  If this partition is mapped to the original loop,
252   /// this is null.
253   Loop *ClonedLoop = nullptr;
254 
255   /// 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;
258 
259   /// These gets populated once the set of instructions have been
260   /// finalized. If this partition is mapped to the original loop, these are not
261   /// set.
262   ValueToValueMapTy VMap;
263 };
264 
265 /// 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>;
269 
270 public:
InstPartitionContainer(Loop * L,LoopInfo * LI,DominatorTree * DT)271   InstPartitionContainer(Loop *L, LoopInfo *LI, DominatorTree *DT)
272       : L(L), LI(LI), DT(DT) {}
273 
274   /// Returns the number of partitions.
getSize() const275   unsigned getSize() const { return PartitionContainer.size(); }
276 
277   /// Adds \p Inst into the current partition if that is marked to
278   /// contain cycles.  Otherwise start a new partition for it.
addToCyclicPartition(Instruction * Inst)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);
283     else
284       PartitionContainer.back().add(Inst);
285   }
286 
287   /// Adds \p Inst into a partition that is not marked to contain
288   /// dependence cycles.
289   ///
290   //  Initially we isolate memory instructions into as many partitions as
291   //  possible, then later we may merge them back together.
addToNewNonCyclicPartition(Instruction * Inst)292   void addToNewNonCyclicPartition(Instruction *Inst) {
293     PartitionContainer.emplace_back(Inst, L);
294   }
295 
296   /// Merges adjacent non-cyclic partitions.
297   ///
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
300   /// too.
mergeAdjacentNonCyclic()301   void mergeAdjacentNonCyclic() {
302     mergeAdjacentPartitionsIf(
303         [](const InstPartition *P) { return !P->hasDepCycle(); });
304   }
305 
306   /// If a partition contains only conditional stores, we won't vectorize
307   /// it.  Try to merge it with a previous cyclic partition.
mergeNonIfConvertible()308   void mergeNonIfConvertible() {
309     mergeAdjacentPartitionsIf([&](const InstPartition *Partition) {
310       if (Partition->hasDepCycle())
311         return true;
312 
313       // Now, check if all stores are conditional in this partition.
314       bool seenStore = false;
315 
316       for (auto *Inst : *Partition)
317         if (isa<StoreInst>(Inst)) {
318           seenStore = true;
319           if (!LoopAccessInfo::blockNeedsPredication(Inst->getParent(), L, DT))
320             return false;
321         }
322       return seenStore;
323     });
324   }
325 
326   /// Merges the partitions according to various heuristics.
mergeBeforePopulating()327   void mergeBeforePopulating() {
328     mergeAdjacentNonCyclic();
329     if (!DistributeNonIfConvertible)
330       mergeNonIfConvertible();
331   }
332 
333   /// Merges partitions in order to ensure that no loads are duplicated.
334   ///
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.
338   ///
339   /// Return if any partitions were merged.
mergeToAvoidDuplicatedLoads()340   bool mergeToAvoidDuplicatedLoads() {
341     using LoadToPartitionT = DenseMap<Instruction *, InstPartition *>;
342     using ToBeMergedT = EquivalenceClasses<InstPartition *>;
343 
344     LoadToPartitionT LoadToPartition;
345     ToBeMergedT ToBeMerged;
346 
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();
352          I != E; ++I) {
353       auto *PartI = &*I;
354 
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)) {
359           bool NewElt;
360           LoadToPartitionT::iterator LoadToPart;
361 
362           std::tie(LoadToPart, NewElt) =
363               LoadToPartition.insert(std::make_pair(Inst, PartI));
364           if (!NewElt) {
365             LLVM_DEBUG(dbgs()
366                        << "Merging partitions due to this load in multiple "
367                        << "partitions: " << PartI << ", " << LoadToPart->second
368                        << "\n"
369                        << *Inst << "\n");
370 
371             auto PartJ = I;
372             do {
373               --PartJ;
374               ToBeMerged.unionSets(PartI, &*PartJ);
375             } while (&*PartJ != LoadToPart->second);
376           }
377         }
378     }
379     if (ToBeMerged.empty())
380       return false;
381 
382     // Merge the member of an equivalence class into its class leader.  This
383     // makes the members empty.
384     for (ToBeMergedT::iterator I = ToBeMerged.begin(), E = ToBeMerged.end();
385          I != E; ++I) {
386       if (!I->isLeader())
387         continue;
388 
389       auto PartI = I->getData();
390       for (auto PartJ : make_range(std::next(ToBeMerged.member_begin(I)),
391                                    ToBeMerged.member_end())) {
392         PartJ->moveTo(*PartI);
393       }
394     }
395 
396     // Remove the empty partitions.
397     PartitionContainer.remove_if(
398         [](const InstPartition &P) { return P.empty(); });
399 
400     return true;
401   }
402 
403   /// Sets up the mapping between instructions to partitions.  If the
404   /// instruction is duplicated across multiple partitions, set the entry to -1.
setupPartitionIdOnInstructions()405   void setupPartitionIdOnInstructions() {
406     int PartitionID = 0;
407     for (const auto &Partition : PartitionContainer) {
408       for (Instruction *Inst : Partition) {
409         bool NewElt;
410         InstToPartitionIdT::iterator Iter;
411 
412         std::tie(Iter, NewElt) =
413             InstToPartitionId.insert(std::make_pair(Inst, PartitionID));
414         if (!NewElt)
415           Iter->second = -1;
416       }
417       ++PartitionID;
418     }
419   }
420 
421   /// Populates the partition with everything that the seeding
422   /// instructions require.
populateUsedSet()423   void populateUsedSet() {
424     for (auto &P : PartitionContainer)
425       P.populateUsedSet();
426   }
427 
428   /// This performs the main chunk of the work of cloning the loops for
429   /// the partitions.
cloneLoops()430   void cloneLoops() {
431     BasicBlock *OrigPH = L->getLoopPreheader();
432     // At this point the predecessor of the preheader is either the memcheck
433     // block or the top part of the original preheader.
434     BasicBlock *Pred = OrigPH->getSinglePredecessor();
435     assert(Pred && "Preheader does not have a single predecessor");
436     BasicBlock *ExitBlock = L->getExitBlock();
437     assert(ExitBlock && "No single exit block");
438     Loop *NewLoop;
439 
440     assert(!PartitionContainer.empty() && "at least two partitions expected");
441     // We're cloning the preheader along with the loop so we already made sure
442     // it was empty.
443     assert(&*OrigPH->begin() == OrigPH->getTerminator() &&
444            "preheader not empty");
445 
446     // Create a loop for each partition except the last.  Clone the original
447     // loop before PH along with adding a preheader for the cloned loop.  Then
448     // update PH to point to the newly added preheader.
449     BasicBlock *TopPH = OrigPH;
450     unsigned Index = getSize() - 1;
451     for (auto I = std::next(PartitionContainer.rbegin()),
452               E = PartitionContainer.rend();
453          I != E; ++I, --Index, TopPH = NewLoop->getLoopPreheader()) {
454       auto *Part = &*I;
455 
456       NewLoop = Part->cloneLoopWithPreheader(TopPH, Pred, Index, LI, DT);
457 
458       Part->getVMap()[ExitBlock] = TopPH;
459       Part->remapInstructions();
460     }
461     Pred->getTerminator()->replaceUsesOfWith(OrigPH, TopPH);
462 
463     // Now go in forward order and update the immediate dominator for the
464     // preheaders with the exiting block of the previous loop.  Dominance
465     // within the loop is updated in cloneLoopWithPreheader.
466     for (auto Curr = PartitionContainer.cbegin(),
467               Next = std::next(PartitionContainer.cbegin()),
468               E = PartitionContainer.cend();
469          Next != E; ++Curr, ++Next)
470       DT->changeImmediateDominator(
471           Next->getDistributedLoop()->getLoopPreheader(),
472           Curr->getDistributedLoop()->getExitingBlock());
473   }
474 
475   /// Removes the dead instructions from the cloned loops.
removeUnusedInsts()476   void removeUnusedInsts() {
477     for (auto &Partition : PartitionContainer)
478       Partition.removeUnusedInsts();
479   }
480 
481   /// For each memory pointer, it computes the partitionId the pointer is
482   /// used in.
483   ///
484   /// This returns an array of int where the I-th entry corresponds to I-th
485   /// entry in LAI.getRuntimePointerCheck().  If the pointer is used in multiple
486   /// partitions its entry is set to -1.
487   SmallVector<int, 8>
computePartitionSetForPointers(const LoopAccessInfo & LAI)488   computePartitionSetForPointers(const LoopAccessInfo &LAI) {
489     const RuntimePointerChecking *RtPtrCheck = LAI.getRuntimePointerChecking();
490 
491     unsigned N = RtPtrCheck->Pointers.size();
492     SmallVector<int, 8> PtrToPartitions(N);
493     for (unsigned I = 0; I < N; ++I) {
494       Value *Ptr = RtPtrCheck->Pointers[I].PointerValue;
495       auto Instructions =
496           LAI.getInstructionsForAccess(Ptr, RtPtrCheck->Pointers[I].IsWritePtr);
497 
498       int &Partition = PtrToPartitions[I];
499       // First set it to uninitialized.
500       Partition = -2;
501       for (Instruction *Inst : Instructions) {
502         // Note that this could be -1 if Inst is duplicated across multiple
503         // partitions.
504         int ThisPartition = this->InstToPartitionId[Inst];
505         if (Partition == -2)
506           Partition = ThisPartition;
507         // -1 means belonging to multiple partitions.
508         else if (Partition == -1)
509           break;
510         else if (Partition != (int)ThisPartition)
511           Partition = -1;
512       }
513       assert(Partition != -2 && "Pointer not belonging to any partition");
514     }
515 
516     return PtrToPartitions;
517   }
518 
print(raw_ostream & OS) const519   void print(raw_ostream &OS) const {
520     unsigned Index = 0;
521     for (const auto &P : PartitionContainer) {
522       OS << "Partition " << Index++ << " (" << &P << "):\n";
523       P.print();
524     }
525   }
526 
dump() const527   void dump() const { print(dbgs()); }
528 
529 #ifndef NDEBUG
operator <<(raw_ostream & OS,const InstPartitionContainer & Partitions)530   friend raw_ostream &operator<<(raw_ostream &OS,
531                                  const InstPartitionContainer &Partitions) {
532     Partitions.print(OS);
533     return OS;
534   }
535 #endif
536 
printBlocks() const537   void printBlocks() const {
538     unsigned Index = 0;
539     for (const auto &P : PartitionContainer) {
540       dbgs() << "\nPartition " << Index++ << " (" << &P << "):\n";
541       P.printBlocks();
542     }
543   }
544 
545 private:
546   using PartitionContainerT = std::list<InstPartition>;
547 
548   /// List of partitions.
549   PartitionContainerT PartitionContainer;
550 
551   /// Mapping from Instruction to partition Id.  If the instruction
552   /// belongs to multiple partitions the entry contains -1.
553   InstToPartitionIdT InstToPartitionId;
554 
555   Loop *L;
556   LoopInfo *LI;
557   DominatorTree *DT;
558 
559   /// The control structure to merge adjacent partitions if both satisfy
560   /// the \p Predicate.
561   template <class UnaryPredicate>
mergeAdjacentPartitionsIf(UnaryPredicate Predicate)562   void mergeAdjacentPartitionsIf(UnaryPredicate Predicate) {
563     InstPartition *PrevMatch = nullptr;
564     for (auto I = PartitionContainer.begin(); I != PartitionContainer.end();) {
565       auto DoesMatch = Predicate(&*I);
566       if (PrevMatch == nullptr && DoesMatch) {
567         PrevMatch = &*I;
568         ++I;
569       } else if (PrevMatch != nullptr && DoesMatch) {
570         I->moveTo(*PrevMatch);
571         I = PartitionContainer.erase(I);
572       } else {
573         PrevMatch = nullptr;
574         ++I;
575       }
576     }
577   }
578 };
579 
580 /// For each memory instruction, this class maintains difference of the
581 /// number of unsafe dependences that start out from this instruction minus
582 /// those that end here.
583 ///
584 /// By traversing the memory instructions in program order and accumulating this
585 /// number, we know whether any unsafe dependence crosses over a program point.
586 class MemoryInstructionDependences {
587   using Dependence = MemoryDepChecker::Dependence;
588 
589 public:
590   struct Entry {
591     Instruction *Inst;
592     unsigned NumUnsafeDependencesStartOrEnd = 0;
593 
Entry__anonce78c04f0111::MemoryInstructionDependences::Entry594     Entry(Instruction *Inst) : Inst(Inst) {}
595   };
596 
597   using AccessesType = SmallVector<Entry, 8>;
598 
begin() const599   AccessesType::const_iterator begin() const { return Accesses.begin(); }
end() const600   AccessesType::const_iterator end() const { return Accesses.end(); }
601 
MemoryInstructionDependences(const SmallVectorImpl<Instruction * > & Instructions,const SmallVectorImpl<Dependence> & Dependences)602   MemoryInstructionDependences(
603       const SmallVectorImpl<Instruction *> &Instructions,
604       const SmallVectorImpl<Dependence> &Dependences) {
605     Accesses.append(Instructions.begin(), Instructions.end());
606 
607     LLVM_DEBUG(dbgs() << "Backward dependences:\n");
608     for (auto &Dep : Dependences)
609       if (Dep.isPossiblyBackward()) {
610         // Note that the designations source and destination follow the program
611         // order, i.e. source is always first.  (The direction is given by the
612         // DepType.)
613         ++Accesses[Dep.Source].NumUnsafeDependencesStartOrEnd;
614         --Accesses[Dep.Destination].NumUnsafeDependencesStartOrEnd;
615 
616         LLVM_DEBUG(Dep.print(dbgs(), 2, Instructions));
617       }
618   }
619 
620 private:
621   AccessesType Accesses;
622 };
623 
624 /// The actual class performing the per-loop work.
625 class LoopDistributeForLoop {
626 public:
LoopDistributeForLoop(Loop * L,Function * F,LoopInfo * LI,DominatorTree * DT,ScalarEvolution * SE,OptimizationRemarkEmitter * ORE)627   LoopDistributeForLoop(Loop *L, Function *F, LoopInfo *LI, DominatorTree *DT,
628                         ScalarEvolution *SE, OptimizationRemarkEmitter *ORE)
629       : L(L), F(F), LI(LI), DT(DT), SE(SE), ORE(ORE) {
630     setForced();
631   }
632 
633   /// Try to distribute an inner-most loop.
processLoop(std::function<const LoopAccessInfo & (Loop &)> & GetLAA)634   bool processLoop(std::function<const LoopAccessInfo &(Loop &)> &GetLAA) {
635     assert(L->empty() && "Only process inner loops.");
636 
637     LLVM_DEBUG(dbgs() << "\nLDist: In \""
638                       << L->getHeader()->getParent()->getName()
639                       << "\" checking " << *L << "\n");
640 
641     if (!L->getExitBlock())
642       return fail("MultipleExitBlocks", "multiple exit blocks");
643     if (!L->isLoopSimplifyForm())
644       return fail("NotLoopSimplifyForm",
645                   "loop is not in loop-simplify form");
646 
647     BasicBlock *PH = L->getLoopPreheader();
648 
649     // LAA will check that we only have a single exiting block.
650     LAI = &GetLAA(*L);
651 
652     // Currently, we only distribute to isolate the part of the loop with
653     // dependence cycles to enable partial vectorization.
654     if (LAI->canVectorizeMemory())
655       return fail("MemOpsCanBeVectorized",
656                   "memory operations are safe for vectorization");
657 
658     auto *Dependences = LAI->getDepChecker().getDependences();
659     if (!Dependences || Dependences->empty())
660       return fail("NoUnsafeDeps", "no unsafe dependences to isolate");
661 
662     InstPartitionContainer Partitions(L, LI, DT);
663 
664     // First, go through each memory operation and assign them to consecutive
665     // partitions (the order of partitions follows program order).  Put those
666     // with unsafe dependences into "cyclic" partition otherwise put each store
667     // in its own "non-cyclic" partition (we'll merge these later).
668     //
669     // Note that a memory operation (e.g. Load2 below) at a program point that
670     // has an unsafe dependence (Store3->Load1) spanning over it must be
671     // included in the same cyclic partition as the dependent operations.  This
672     // is to preserve the original program order after distribution.  E.g.:
673     //
674     //                NumUnsafeDependencesStartOrEnd  NumUnsafeDependencesActive
675     //  Load1   -.                     1                       0->1
676     //  Load2    | /Unsafe/            0                       1
677     //  Store3  -'                    -1                       1->0
678     //  Load4                          0                       0
679     //
680     // NumUnsafeDependencesActive > 0 indicates this situation and in this case
681     // we just keep assigning to the same cyclic partition until
682     // NumUnsafeDependencesActive reaches 0.
683     const MemoryDepChecker &DepChecker = LAI->getDepChecker();
684     MemoryInstructionDependences MID(DepChecker.getMemoryInstructions(),
685                                      *Dependences);
686 
687     int NumUnsafeDependencesActive = 0;
688     for (auto &InstDep : MID) {
689       Instruction *I = InstDep.Inst;
690       // We update NumUnsafeDependencesActive post-instruction, catch the
691       // start of a dependence directly via NumUnsafeDependencesStartOrEnd.
692       if (NumUnsafeDependencesActive ||
693           InstDep.NumUnsafeDependencesStartOrEnd > 0)
694         Partitions.addToCyclicPartition(I);
695       else
696         Partitions.addToNewNonCyclicPartition(I);
697       NumUnsafeDependencesActive += InstDep.NumUnsafeDependencesStartOrEnd;
698       assert(NumUnsafeDependencesActive >= 0 &&
699              "Negative number of dependences active");
700     }
701 
702     // Add partitions for values used outside.  These partitions can be out of
703     // order from the original program order.  This is OK because if the
704     // partition uses a load we will merge this partition with the original
705     // partition of the load that we set up in the previous loop (see
706     // mergeToAvoidDuplicatedLoads).
707     auto DefsUsedOutside = findDefsUsedOutsideOfLoop(L);
708     for (auto *Inst : DefsUsedOutside)
709       Partitions.addToNewNonCyclicPartition(Inst);
710 
711     LLVM_DEBUG(dbgs() << "Seeded partitions:\n" << Partitions);
712     if (Partitions.getSize() < 2)
713       return fail("CantIsolateUnsafeDeps",
714                   "cannot isolate unsafe dependencies");
715 
716     // Run the merge heuristics: Merge non-cyclic adjacent partitions since we
717     // should be able to vectorize these together.
718     Partitions.mergeBeforePopulating();
719     LLVM_DEBUG(dbgs() << "\nMerged partitions:\n" << Partitions);
720     if (Partitions.getSize() < 2)
721       return fail("CantIsolateUnsafeDeps",
722                   "cannot isolate unsafe dependencies");
723 
724     // Now, populate the partitions with non-memory operations.
725     Partitions.populateUsedSet();
726     LLVM_DEBUG(dbgs() << "\nPopulated partitions:\n" << Partitions);
727 
728     // In order to preserve original lexical order for loads, keep them in the
729     // partition that we set up in the MemoryInstructionDependences loop.
730     if (Partitions.mergeToAvoidDuplicatedLoads()) {
731       LLVM_DEBUG(dbgs() << "\nPartitions merged to ensure unique loads:\n"
732                         << Partitions);
733       if (Partitions.getSize() < 2)
734         return fail("CantIsolateUnsafeDeps",
735                     "cannot isolate unsafe dependencies");
736     }
737 
738     // Don't distribute the loop if we need too many SCEV run-time checks.
739     const SCEVUnionPredicate &Pred = LAI->getPSE().getUnionPredicate();
740     if (Pred.getComplexity() > (IsForced.getValueOr(false)
741                                     ? PragmaDistributeSCEVCheckThreshold
742                                     : DistributeSCEVCheckThreshold))
743       return fail("TooManySCEVRuntimeChecks",
744                   "too many SCEV run-time checks needed.\n");
745 
746     LLVM_DEBUG(dbgs() << "\nDistributing loop: " << *L << "\n");
747     // We're done forming the partitions set up the reverse mapping from
748     // instructions to partitions.
749     Partitions.setupPartitionIdOnInstructions();
750 
751     // To keep things simple have an empty preheader before we version or clone
752     // the loop.  (Also split if this has no predecessor, i.e. entry, because we
753     // rely on PH having a predecessor.)
754     if (!PH->getSinglePredecessor() || &*PH->begin() != PH->getTerminator())
755       SplitBlock(PH, PH->getTerminator(), DT, LI);
756 
757     // If we need run-time checks, version the loop now.
758     auto PtrToPartition = Partitions.computePartitionSetForPointers(*LAI);
759     const auto *RtPtrChecking = LAI->getRuntimePointerChecking();
760     const auto &AllChecks = RtPtrChecking->getChecks();
761     auto Checks = includeOnlyCrossPartitionChecks(AllChecks, PtrToPartition,
762                                                   RtPtrChecking);
763 
764     if (!Pred.isAlwaysTrue() || !Checks.empty()) {
765       LLVM_DEBUG(dbgs() << "\nPointers:\n");
766       LLVM_DEBUG(LAI->getRuntimePointerChecking()->printChecks(dbgs(), Checks));
767       LoopVersioning LVer(*LAI, L, LI, DT, SE, false);
768       LVer.setAliasChecks(std::move(Checks));
769       LVer.setSCEVChecks(LAI->getPSE().getUnionPredicate());
770       LVer.versionLoop(DefsUsedOutside);
771       LVer.annotateLoopWithNoAlias();
772     }
773 
774     // Create identical copies of the original loop for each partition and hook
775     // them up sequentially.
776     Partitions.cloneLoops();
777 
778     // Now, we remove the instruction from each loop that don't belong to that
779     // partition.
780     Partitions.removeUnusedInsts();
781     LLVM_DEBUG(dbgs() << "\nAfter removing unused Instrs:\n");
782     LLVM_DEBUG(Partitions.printBlocks());
783 
784     if (LDistVerify) {
785       LI->verify(*DT);
786       assert(DT->verify(DominatorTree::VerificationLevel::Fast));
787     }
788 
789     ++NumLoopsDistributed;
790     // Report the success.
791     ORE->emit([&]() {
792       return OptimizationRemark(LDIST_NAME, "Distribute", L->getStartLoc(),
793                                 L->getHeader())
794              << "distributed loop";
795     });
796     return true;
797   }
798 
799   /// Provide diagnostics then \return with false.
fail(StringRef RemarkName,StringRef Message)800   bool fail(StringRef RemarkName, StringRef Message) {
801     LLVMContext &Ctx = F->getContext();
802     bool Forced = isForced().getValueOr(false);
803 
804     LLVM_DEBUG(dbgs() << "Skipping; " << Message << "\n");
805 
806     // With Rpass-missed report that distribution failed.
807     ORE->emit([&]() {
808       return OptimizationRemarkMissed(LDIST_NAME, "NotDistributed",
809                                       L->getStartLoc(), L->getHeader())
810              << "loop not distributed: use -Rpass-analysis=loop-distribute for "
811                 "more "
812                 "info";
813     });
814 
815     // With Rpass-analysis report why.  This is on by default if distribution
816     // was requested explicitly.
817     ORE->emit(OptimizationRemarkAnalysis(
818                   Forced ? OptimizationRemarkAnalysis::AlwaysPrint : LDIST_NAME,
819                   RemarkName, L->getStartLoc(), L->getHeader())
820               << "loop not distributed: " << Message);
821 
822     // Also issue a warning if distribution was requested explicitly but it
823     // failed.
824     if (Forced)
825       Ctx.diagnose(DiagnosticInfoOptimizationFailure(
826           *F, L->getStartLoc(), "loop not distributed: failed "
827                                 "explicitly specified loop distribution"));
828 
829     return false;
830   }
831 
832   /// Return if distribution forced to be enabled/disabled for the loop.
833   ///
834   /// If the optional has a value, it indicates whether distribution was forced
835   /// to be enabled (true) or disabled (false).  If the optional has no value
836   /// distribution was not forced either way.
isForced() const837   const Optional<bool> &isForced() const { return IsForced; }
838 
839 private:
840   /// Filter out checks between pointers from the same partition.
841   ///
842   /// \p PtrToPartition contains the partition number for pointers.  Partition
843   /// number -1 means that the pointer is used in multiple partitions.  In this
844   /// case we can't safely omit the check.
845   SmallVector<RuntimePointerChecking::PointerCheck, 4>
includeOnlyCrossPartitionChecks(const SmallVectorImpl<RuntimePointerChecking::PointerCheck> & AllChecks,const SmallVectorImpl<int> & PtrToPartition,const RuntimePointerChecking * RtPtrChecking)846   includeOnlyCrossPartitionChecks(
847       const SmallVectorImpl<RuntimePointerChecking::PointerCheck> &AllChecks,
848       const SmallVectorImpl<int> &PtrToPartition,
849       const RuntimePointerChecking *RtPtrChecking) {
850     SmallVector<RuntimePointerChecking::PointerCheck, 4> Checks;
851 
852     copy_if(AllChecks, std::back_inserter(Checks),
853             [&](const RuntimePointerChecking::PointerCheck &Check) {
854               for (unsigned PtrIdx1 : Check.first->Members)
855                 for (unsigned PtrIdx2 : Check.second->Members)
856                   // Only include this check if there is a pair of pointers
857                   // that require checking and the pointers fall into
858                   // separate partitions.
859                   //
860                   // (Note that we already know at this point that the two
861                   // pointer groups need checking but it doesn't follow
862                   // that each pair of pointers within the two groups need
863                   // checking as well.
864                   //
865                   // In other words we don't want to include a check just
866                   // because there is a pair of pointers between the two
867                   // pointer groups that require checks and a different
868                   // pair whose pointers fall into different partitions.)
869                   if (RtPtrChecking->needsChecking(PtrIdx1, PtrIdx2) &&
870                       !RuntimePointerChecking::arePointersInSamePartition(
871                           PtrToPartition, PtrIdx1, PtrIdx2))
872                     return true;
873               return false;
874             });
875 
876     return Checks;
877   }
878 
879   /// Check whether the loop metadata is forcing distribution to be
880   /// enabled/disabled.
setForced()881   void setForced() {
882     Optional<const MDOperand *> Value =
883         findStringMetadataForLoop(L, "llvm.loop.distribute.enable");
884     if (!Value)
885       return;
886 
887     const MDOperand *Op = *Value;
888     assert(Op && mdconst::hasa<ConstantInt>(*Op) && "invalid metadata");
889     IsForced = mdconst::extract<ConstantInt>(*Op)->getZExtValue();
890   }
891 
892   Loop *L;
893   Function *F;
894 
895   // Analyses used.
896   LoopInfo *LI;
897   const LoopAccessInfo *LAI = nullptr;
898   DominatorTree *DT;
899   ScalarEvolution *SE;
900   OptimizationRemarkEmitter *ORE;
901 
902   /// Indicates whether distribution is forced to be enabled/disabled for
903   /// the loop.
904   ///
905   /// If the optional has a value, it indicates whether distribution was forced
906   /// to be enabled (true) or disabled (false).  If the optional has no value
907   /// distribution was not forced either way.
908   Optional<bool> IsForced;
909 };
910 
911 } // end anonymous namespace
912 
913 /// Shared implementation between new and old PMs.
runImpl(Function & F,LoopInfo * LI,DominatorTree * DT,ScalarEvolution * SE,OptimizationRemarkEmitter * ORE,std::function<const LoopAccessInfo & (Loop &)> & GetLAA)914 static bool runImpl(Function &F, LoopInfo *LI, DominatorTree *DT,
915                     ScalarEvolution *SE, OptimizationRemarkEmitter *ORE,
916                     std::function<const LoopAccessInfo &(Loop &)> &GetLAA) {
917   // Build up a worklist of inner-loops to vectorize. This is necessary as the
918   // act of distributing a loop creates new loops and can invalidate iterators
919   // across the loops.
920   SmallVector<Loop *, 8> Worklist;
921 
922   for (Loop *TopLevelLoop : *LI)
923     for (Loop *L : depth_first(TopLevelLoop))
924       // We only handle inner-most loops.
925       if (L->empty())
926         Worklist.push_back(L);
927 
928   // Now walk the identified inner loops.
929   bool Changed = false;
930   for (Loop *L : Worklist) {
931     LoopDistributeForLoop LDL(L, &F, LI, DT, SE, ORE);
932 
933     // If distribution was forced for the specific loop to be
934     // enabled/disabled, follow that.  Otherwise use the global flag.
935     if (LDL.isForced().getValueOr(EnableLoopDistribute))
936       Changed |= LDL.processLoop(GetLAA);
937   }
938 
939   // Process each loop nest in the function.
940   return Changed;
941 }
942 
943 namespace {
944 
945 /// The pass class.
946 class LoopDistributeLegacy : public FunctionPass {
947 public:
948   static char ID;
949 
LoopDistributeLegacy()950   LoopDistributeLegacy() : FunctionPass(ID) {
951     // The default is set by the caller.
952     initializeLoopDistributeLegacyPass(*PassRegistry::getPassRegistry());
953   }
954 
runOnFunction(Function & F)955   bool runOnFunction(Function &F) override {
956     if (skipFunction(F))
957       return false;
958 
959     auto *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
960     auto *LAA = &getAnalysis<LoopAccessLegacyAnalysis>();
961     auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
962     auto *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
963     auto *ORE = &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE();
964     std::function<const LoopAccessInfo &(Loop &)> GetLAA =
965         [&](Loop &L) -> const LoopAccessInfo & { return LAA->getInfo(&L); };
966 
967     return runImpl(F, LI, DT, SE, ORE, GetLAA);
968   }
969 
getAnalysisUsage(AnalysisUsage & AU) const970   void getAnalysisUsage(AnalysisUsage &AU) const override {
971     AU.addRequired<ScalarEvolutionWrapperPass>();
972     AU.addRequired<LoopInfoWrapperPass>();
973     AU.addPreserved<LoopInfoWrapperPass>();
974     AU.addRequired<LoopAccessLegacyAnalysis>();
975     AU.addRequired<DominatorTreeWrapperPass>();
976     AU.addPreserved<DominatorTreeWrapperPass>();
977     AU.addRequired<OptimizationRemarkEmitterWrapperPass>();
978     AU.addPreserved<GlobalsAAWrapperPass>();
979   }
980 };
981 
982 } // end anonymous namespace
983 
run(Function & F,FunctionAnalysisManager & AM)984 PreservedAnalyses LoopDistributePass::run(Function &F,
985                                           FunctionAnalysisManager &AM) {
986   auto &LI = AM.getResult<LoopAnalysis>(F);
987   auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
988   auto &SE = AM.getResult<ScalarEvolutionAnalysis>(F);
989   auto &ORE = AM.getResult<OptimizationRemarkEmitterAnalysis>(F);
990 
991   // We don't directly need these analyses but they're required for loop
992   // analyses so provide them below.
993   auto &AA = AM.getResult<AAManager>(F);
994   auto &AC = AM.getResult<AssumptionAnalysis>(F);
995   auto &TTI = AM.getResult<TargetIRAnalysis>(F);
996   auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
997 
998   auto &LAM = AM.getResult<LoopAnalysisManagerFunctionProxy>(F).getManager();
999   std::function<const LoopAccessInfo &(Loop &)> GetLAA =
1000       [&](Loop &L) -> const LoopAccessInfo & {
1001     LoopStandardAnalysisResults AR = {AA, AC, DT, LI, SE, TLI, TTI, nullptr};
1002     return LAM.getResult<LoopAccessAnalysis>(L, AR);
1003   };
1004 
1005   bool Changed = runImpl(F, &LI, &DT, &SE, &ORE, GetLAA);
1006   if (!Changed)
1007     return PreservedAnalyses::all();
1008   PreservedAnalyses PA;
1009   PA.preserve<LoopAnalysis>();
1010   PA.preserve<DominatorTreeAnalysis>();
1011   PA.preserve<GlobalsAA>();
1012   return PA;
1013 }
1014 
1015 char LoopDistributeLegacy::ID;
1016 
1017 static const char ldist_name[] = "Loop Distribution";
1018 
INITIALIZE_PASS_BEGIN(LoopDistributeLegacy,LDIST_NAME,ldist_name,false,false)1019 INITIALIZE_PASS_BEGIN(LoopDistributeLegacy, LDIST_NAME, ldist_name, false,
1020                       false)
1021 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
1022 INITIALIZE_PASS_DEPENDENCY(LoopAccessLegacyAnalysis)
1023 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
1024 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
1025 INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass)
1026 INITIALIZE_PASS_END(LoopDistributeLegacy, LDIST_NAME, ldist_name, false, false)
1027 
1028 FunctionPass *llvm::createLoopDistributePass() { return new LoopDistributeLegacy(); }
1029