• 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/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/DiagnosticInfo.h"
32 #include "llvm/IR/Dominators.h"
33 #include "llvm/Pass.h"
34 #include "llvm/Support/CommandLine.h"
35 #include "llvm/Support/Debug.h"
36 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
37 #include "llvm/Transforms/Utils/Cloning.h"
38 #include "llvm/Transforms/Utils/LoopUtils.h"
39 #include "llvm/Transforms/Utils/LoopVersioning.h"
40 #include <list>
41 
42 #define LDIST_NAME "loop-distribute"
43 #define DEBUG_TYPE LDIST_NAME
44 
45 using namespace llvm;
46 
47 static cl::opt<bool>
48     LDistVerify("loop-distribute-verify", cl::Hidden,
49                 cl::desc("Turn on DominatorTree and LoopInfo verification "
50                          "after Loop Distribution"),
51                 cl::init(false));
52 
53 static cl::opt<bool> DistributeNonIfConvertible(
54     "loop-distribute-non-if-convertible", cl::Hidden,
55     cl::desc("Whether to distribute into a loop that may not be "
56              "if-convertible by the loop vectorizer"),
57     cl::init(false));
58 
59 static cl::opt<unsigned> DistributeSCEVCheckThreshold(
60     "loop-distribute-scev-check-threshold", cl::init(8), cl::Hidden,
61     cl::desc("The maximum number of SCEV checks allowed for Loop "
62              "Distribution"));
63 
64 static cl::opt<unsigned> PragmaDistributeSCEVCheckThreshold(
65     "loop-distribute-scev-check-threshold-with-pragma", cl::init(128),
66     cl::Hidden,
67     cl::desc(
68         "The maximum number of SCEV checks allowed for Loop "
69         "Distribution for loop marked with #pragma loop distribute(enable)"));
70 
71 // Note that the initial value for this depends on whether the pass is invoked
72 // directly or from the optimization pipeline.
73 static cl::opt<bool> EnableLoopDistribute(
74     "enable-loop-distribute", cl::Hidden,
75     cl::desc("Enable the new, experimental LoopDistribution Pass"));
76 
77 STATISTIC(NumLoopsDistributed, "Number of loops distributed");
78 
79 namespace {
80 /// \brief Maintains the set of instructions of the loop for a partition before
81 /// cloning.  After cloning, it hosts the new loop.
82 class InstPartition {
83   typedef SmallPtrSet<Instruction *, 8> InstructionSet;
84 
85 public:
InstPartition(Instruction * I,Loop * L,bool DepCycle=false)86   InstPartition(Instruction *I, Loop *L, bool DepCycle = false)
87       : DepCycle(DepCycle), OrigLoop(L), ClonedLoop(nullptr) {
88     Set.insert(I);
89   }
90 
91   /// \brief Returns whether this partition contains a dependence cycle.
hasDepCycle() const92   bool hasDepCycle() const { return DepCycle; }
93 
94   /// \brief Adds an instruction to this partition.
add(Instruction * I)95   void add(Instruction *I) { Set.insert(I); }
96 
97   /// \brief Collection accessors.
begin()98   InstructionSet::iterator begin() { return Set.begin(); }
end()99   InstructionSet::iterator end() { return Set.end(); }
begin() const100   InstructionSet::const_iterator begin() const { return Set.begin(); }
end() const101   InstructionSet::const_iterator end() const { return Set.end(); }
empty() const102   bool empty() const { return Set.empty(); }
103 
104   /// \brief Moves this partition into \p Other.  This partition becomes empty
105   /// after this.
moveTo(InstPartition & Other)106   void moveTo(InstPartition &Other) {
107     Other.Set.insert(Set.begin(), Set.end());
108     Set.clear();
109     Other.DepCycle |= DepCycle;
110   }
111 
112   /// \brief Populates the partition with a transitive closure of all the
113   /// instructions that the seeded instructions dependent on.
populateUsedSet()114   void populateUsedSet() {
115     // FIXME: We currently don't use control-dependence but simply include all
116     // blocks (possibly empty at the end) and let simplifycfg mostly clean this
117     // up.
118     for (auto *B : OrigLoop->getBlocks())
119       Set.insert(B->getTerminator());
120 
121     // Follow the use-def chains to form a transitive closure of all the
122     // instructions that the originally seeded instructions depend on.
123     SmallVector<Instruction *, 8> Worklist(Set.begin(), Set.end());
124     while (!Worklist.empty()) {
125       Instruction *I = Worklist.pop_back_val();
126       // Insert instructions from the loop that we depend on.
127       for (Value *V : I->operand_values()) {
128         auto *I = dyn_cast<Instruction>(V);
129         if (I && OrigLoop->contains(I->getParent()) && Set.insert(I).second)
130           Worklist.push_back(I);
131       }
132     }
133   }
134 
135   /// \brief Clones the original loop.
136   ///
137   /// Updates LoopInfo and DominatorTree using the information that block \p
138   /// LoopDomBB dominates the loop.
cloneLoopWithPreheader(BasicBlock * InsertBefore,BasicBlock * LoopDomBB,unsigned Index,LoopInfo * LI,DominatorTree * DT)139   Loop *cloneLoopWithPreheader(BasicBlock *InsertBefore, BasicBlock *LoopDomBB,
140                                unsigned Index, LoopInfo *LI,
141                                DominatorTree *DT) {
142     ClonedLoop = ::cloneLoopWithPreheader(InsertBefore, LoopDomBB, OrigLoop,
143                                           VMap, Twine(".ldist") + Twine(Index),
144                                           LI, DT, ClonedLoopBlocks);
145     return ClonedLoop;
146   }
147 
148   /// \brief The cloned loop.  If this partition is mapped to the original loop,
149   /// this is null.
getClonedLoop() const150   const Loop *getClonedLoop() const { return ClonedLoop; }
151 
152   /// \brief Returns the loop where this partition ends up after distribution.
153   /// If this partition is mapped to the original loop then use the block from
154   /// the loop.
getDistributedLoop() const155   const Loop *getDistributedLoop() const {
156     return ClonedLoop ? ClonedLoop : OrigLoop;
157   }
158 
159   /// \brief The VMap that is populated by cloning and then used in
160   /// remapinstruction to remap the cloned instructions.
getVMap()161   ValueToValueMapTy &getVMap() { return VMap; }
162 
163   /// \brief Remaps the cloned instructions using VMap.
remapInstructions()164   void remapInstructions() {
165     remapInstructionsInBlocks(ClonedLoopBlocks, VMap);
166   }
167 
168   /// \brief Based on the set of instructions selected for this partition,
169   /// removes the unnecessary ones.
removeUnusedInsts()170   void removeUnusedInsts() {
171     SmallVector<Instruction *, 8> Unused;
172 
173     for (auto *Block : OrigLoop->getBlocks())
174       for (auto &Inst : *Block)
175         if (!Set.count(&Inst)) {
176           Instruction *NewInst = &Inst;
177           if (!VMap.empty())
178             NewInst = cast<Instruction>(VMap[NewInst]);
179 
180           assert(!isa<BranchInst>(NewInst) &&
181                  "Branches are marked used early on");
182           Unused.push_back(NewInst);
183         }
184 
185     // Delete the instructions backwards, as it has a reduced likelihood of
186     // having to update as many def-use and use-def chains.
187     for (auto *Inst : reverse(Unused)) {
188       if (!Inst->use_empty())
189         Inst->replaceAllUsesWith(UndefValue::get(Inst->getType()));
190       Inst->eraseFromParent();
191     }
192   }
193 
print() const194   void print() const {
195     if (DepCycle)
196       dbgs() << "  (cycle)\n";
197     for (auto *I : Set)
198       // Prefix with the block name.
199       dbgs() << "  " << I->getParent()->getName() << ":" << *I << "\n";
200   }
201 
printBlocks() const202   void printBlocks() const {
203     for (auto *BB : getDistributedLoop()->getBlocks())
204       dbgs() << *BB;
205   }
206 
207 private:
208   /// \brief Instructions from OrigLoop selected for this partition.
209   InstructionSet Set;
210 
211   /// \brief Whether this partition contains a dependence cycle.
212   bool DepCycle;
213 
214   /// \brief The original loop.
215   Loop *OrigLoop;
216 
217   /// \brief The cloned loop.  If this partition is mapped to the original loop,
218   /// this is null.
219   Loop *ClonedLoop;
220 
221   /// \brief The blocks of ClonedLoop including the preheader.  If this
222   /// partition is mapped to the original loop, this is empty.
223   SmallVector<BasicBlock *, 8> ClonedLoopBlocks;
224 
225   /// \brief These gets populated once the set of instructions have been
226   /// finalized. If this partition is mapped to the original loop, these are not
227   /// set.
228   ValueToValueMapTy VMap;
229 };
230 
231 /// \brief Holds the set of Partitions.  It populates them, merges them and then
232 /// clones the loops.
233 class InstPartitionContainer {
234   typedef DenseMap<Instruction *, int> InstToPartitionIdT;
235 
236 public:
InstPartitionContainer(Loop * L,LoopInfo * LI,DominatorTree * DT)237   InstPartitionContainer(Loop *L, LoopInfo *LI, DominatorTree *DT)
238       : L(L), LI(LI), DT(DT) {}
239 
240   /// \brief Returns the number of partitions.
getSize() const241   unsigned getSize() const { return PartitionContainer.size(); }
242 
243   /// \brief Adds \p Inst into the current partition if that is marked to
244   /// contain cycles.  Otherwise start a new partition for it.
addToCyclicPartition(Instruction * Inst)245   void addToCyclicPartition(Instruction *Inst) {
246     // If the current partition is non-cyclic.  Start a new one.
247     if (PartitionContainer.empty() || !PartitionContainer.back().hasDepCycle())
248       PartitionContainer.emplace_back(Inst, L, /*DepCycle=*/true);
249     else
250       PartitionContainer.back().add(Inst);
251   }
252 
253   /// \brief Adds \p Inst into a partition that is not marked to contain
254   /// dependence cycles.
255   ///
256   //  Initially we isolate memory instructions into as many partitions as
257   //  possible, then later we may merge them back together.
addToNewNonCyclicPartition(Instruction * Inst)258   void addToNewNonCyclicPartition(Instruction *Inst) {
259     PartitionContainer.emplace_back(Inst, L);
260   }
261 
262   /// \brief Merges adjacent non-cyclic partitions.
263   ///
264   /// The idea is that we currently only want to isolate the non-vectorizable
265   /// partition.  We could later allow more distribution among these partition
266   /// too.
mergeAdjacentNonCyclic()267   void mergeAdjacentNonCyclic() {
268     mergeAdjacentPartitionsIf(
269         [](const InstPartition *P) { return !P->hasDepCycle(); });
270   }
271 
272   /// \brief If a partition contains only conditional stores, we won't vectorize
273   /// it.  Try to merge it with a previous cyclic partition.
mergeNonIfConvertible()274   void mergeNonIfConvertible() {
275     mergeAdjacentPartitionsIf([&](const InstPartition *Partition) {
276       if (Partition->hasDepCycle())
277         return true;
278 
279       // Now, check if all stores are conditional in this partition.
280       bool seenStore = false;
281 
282       for (auto *Inst : *Partition)
283         if (isa<StoreInst>(Inst)) {
284           seenStore = true;
285           if (!LoopAccessInfo::blockNeedsPredication(Inst->getParent(), L, DT))
286             return false;
287         }
288       return seenStore;
289     });
290   }
291 
292   /// \brief Merges the partitions according to various heuristics.
mergeBeforePopulating()293   void mergeBeforePopulating() {
294     mergeAdjacentNonCyclic();
295     if (!DistributeNonIfConvertible)
296       mergeNonIfConvertible();
297   }
298 
299   /// \brief Merges partitions in order to ensure that no loads are duplicated.
300   ///
301   /// We can't duplicate loads because that could potentially reorder them.
302   /// LoopAccessAnalysis provides dependency information with the context that
303   /// the order of memory operation is preserved.
304   ///
305   /// Return if any partitions were merged.
mergeToAvoidDuplicatedLoads()306   bool mergeToAvoidDuplicatedLoads() {
307     typedef DenseMap<Instruction *, InstPartition *> LoadToPartitionT;
308     typedef EquivalenceClasses<InstPartition *> ToBeMergedT;
309 
310     LoadToPartitionT LoadToPartition;
311     ToBeMergedT ToBeMerged;
312 
313     // Step through the partitions and create equivalence between partitions
314     // that contain the same load.  Also put partitions in between them in the
315     // same equivalence class to avoid reordering of memory operations.
316     for (PartitionContainerT::iterator I = PartitionContainer.begin(),
317                                        E = PartitionContainer.end();
318          I != E; ++I) {
319       auto *PartI = &*I;
320 
321       // If a load occurs in two partitions PartI and PartJ, merge all
322       // partitions (PartI, PartJ] into PartI.
323       for (Instruction *Inst : *PartI)
324         if (isa<LoadInst>(Inst)) {
325           bool NewElt;
326           LoadToPartitionT::iterator LoadToPart;
327 
328           std::tie(LoadToPart, NewElt) =
329               LoadToPartition.insert(std::make_pair(Inst, PartI));
330           if (!NewElt) {
331             DEBUG(dbgs() << "Merging partitions due to this load in multiple "
332                          << "partitions: " << PartI << ", "
333                          << LoadToPart->second << "\n" << *Inst << "\n");
334 
335             auto PartJ = I;
336             do {
337               --PartJ;
338               ToBeMerged.unionSets(PartI, &*PartJ);
339             } while (&*PartJ != LoadToPart->second);
340           }
341         }
342     }
343     if (ToBeMerged.empty())
344       return false;
345 
346     // Merge the member of an equivalence class into its class leader.  This
347     // makes the members empty.
348     for (ToBeMergedT::iterator I = ToBeMerged.begin(), E = ToBeMerged.end();
349          I != E; ++I) {
350       if (!I->isLeader())
351         continue;
352 
353       auto PartI = I->getData();
354       for (auto PartJ : make_range(std::next(ToBeMerged.member_begin(I)),
355                                    ToBeMerged.member_end())) {
356         PartJ->moveTo(*PartI);
357       }
358     }
359 
360     // Remove the empty partitions.
361     PartitionContainer.remove_if(
362         [](const InstPartition &P) { return P.empty(); });
363 
364     return true;
365   }
366 
367   /// \brief Sets up the mapping between instructions to partitions.  If the
368   /// instruction is duplicated across multiple partitions, set the entry to -1.
setupPartitionIdOnInstructions()369   void setupPartitionIdOnInstructions() {
370     int PartitionID = 0;
371     for (const auto &Partition : PartitionContainer) {
372       for (Instruction *Inst : Partition) {
373         bool NewElt;
374         InstToPartitionIdT::iterator Iter;
375 
376         std::tie(Iter, NewElt) =
377             InstToPartitionId.insert(std::make_pair(Inst, PartitionID));
378         if (!NewElt)
379           Iter->second = -1;
380       }
381       ++PartitionID;
382     }
383   }
384 
385   /// \brief Populates the partition with everything that the seeding
386   /// instructions require.
populateUsedSet()387   void populateUsedSet() {
388     for (auto &P : PartitionContainer)
389       P.populateUsedSet();
390   }
391 
392   /// \brief This performs the main chunk of the work of cloning the loops for
393   /// the partitions.
cloneLoops()394   void cloneLoops() {
395     BasicBlock *OrigPH = L->getLoopPreheader();
396     // At this point the predecessor of the preheader is either the memcheck
397     // block or the top part of the original preheader.
398     BasicBlock *Pred = OrigPH->getSinglePredecessor();
399     assert(Pred && "Preheader does not have a single predecessor");
400     BasicBlock *ExitBlock = L->getExitBlock();
401     assert(ExitBlock && "No single exit block");
402     Loop *NewLoop;
403 
404     assert(!PartitionContainer.empty() && "at least two partitions expected");
405     // We're cloning the preheader along with the loop so we already made sure
406     // it was empty.
407     assert(&*OrigPH->begin() == OrigPH->getTerminator() &&
408            "preheader not empty");
409 
410     // Create a loop for each partition except the last.  Clone the original
411     // loop before PH along with adding a preheader for the cloned loop.  Then
412     // update PH to point to the newly added preheader.
413     BasicBlock *TopPH = OrigPH;
414     unsigned Index = getSize() - 1;
415     for (auto I = std::next(PartitionContainer.rbegin()),
416               E = PartitionContainer.rend();
417          I != E; ++I, --Index, TopPH = NewLoop->getLoopPreheader()) {
418       auto *Part = &*I;
419 
420       NewLoop = Part->cloneLoopWithPreheader(TopPH, Pred, Index, LI, DT);
421 
422       Part->getVMap()[ExitBlock] = TopPH;
423       Part->remapInstructions();
424     }
425     Pred->getTerminator()->replaceUsesOfWith(OrigPH, TopPH);
426 
427     // Now go in forward order and update the immediate dominator for the
428     // preheaders with the exiting block of the previous loop.  Dominance
429     // within the loop is updated in cloneLoopWithPreheader.
430     for (auto Curr = PartitionContainer.cbegin(),
431               Next = std::next(PartitionContainer.cbegin()),
432               E = PartitionContainer.cend();
433          Next != E; ++Curr, ++Next)
434       DT->changeImmediateDominator(
435           Next->getDistributedLoop()->getLoopPreheader(),
436           Curr->getDistributedLoop()->getExitingBlock());
437   }
438 
439   /// \brief Removes the dead instructions from the cloned loops.
removeUnusedInsts()440   void removeUnusedInsts() {
441     for (auto &Partition : PartitionContainer)
442       Partition.removeUnusedInsts();
443   }
444 
445   /// \brief For each memory pointer, it computes the partitionId the pointer is
446   /// used in.
447   ///
448   /// This returns an array of int where the I-th entry corresponds to I-th
449   /// entry in LAI.getRuntimePointerCheck().  If the pointer is used in multiple
450   /// partitions its entry is set to -1.
451   SmallVector<int, 8>
computePartitionSetForPointers(const LoopAccessInfo & LAI)452   computePartitionSetForPointers(const LoopAccessInfo &LAI) {
453     const RuntimePointerChecking *RtPtrCheck = LAI.getRuntimePointerChecking();
454 
455     unsigned N = RtPtrCheck->Pointers.size();
456     SmallVector<int, 8> PtrToPartitions(N);
457     for (unsigned I = 0; I < N; ++I) {
458       Value *Ptr = RtPtrCheck->Pointers[I].PointerValue;
459       auto Instructions =
460           LAI.getInstructionsForAccess(Ptr, RtPtrCheck->Pointers[I].IsWritePtr);
461 
462       int &Partition = PtrToPartitions[I];
463       // First set it to uninitialized.
464       Partition = -2;
465       for (Instruction *Inst : Instructions) {
466         // Note that this could be -1 if Inst is duplicated across multiple
467         // partitions.
468         int ThisPartition = this->InstToPartitionId[Inst];
469         if (Partition == -2)
470           Partition = ThisPartition;
471         // -1 means belonging to multiple partitions.
472         else if (Partition == -1)
473           break;
474         else if (Partition != (int)ThisPartition)
475           Partition = -1;
476       }
477       assert(Partition != -2 && "Pointer not belonging to any partition");
478     }
479 
480     return PtrToPartitions;
481   }
482 
print(raw_ostream & OS) const483   void print(raw_ostream &OS) const {
484     unsigned Index = 0;
485     for (const auto &P : PartitionContainer) {
486       OS << "Partition " << Index++ << " (" << &P << "):\n";
487       P.print();
488     }
489   }
490 
dump() const491   void dump() const { print(dbgs()); }
492 
493 #ifndef NDEBUG
operator <<(raw_ostream & OS,const InstPartitionContainer & Partitions)494   friend raw_ostream &operator<<(raw_ostream &OS,
495                                  const InstPartitionContainer &Partitions) {
496     Partitions.print(OS);
497     return OS;
498   }
499 #endif
500 
printBlocks() const501   void printBlocks() const {
502     unsigned Index = 0;
503     for (const auto &P : PartitionContainer) {
504       dbgs() << "\nPartition " << Index++ << " (" << &P << "):\n";
505       P.printBlocks();
506     }
507   }
508 
509 private:
510   typedef std::list<InstPartition> PartitionContainerT;
511 
512   /// \brief List of partitions.
513   PartitionContainerT PartitionContainer;
514 
515   /// \brief Mapping from Instruction to partition Id.  If the instruction
516   /// belongs to multiple partitions the entry contains -1.
517   InstToPartitionIdT InstToPartitionId;
518 
519   Loop *L;
520   LoopInfo *LI;
521   DominatorTree *DT;
522 
523   /// \brief The control structure to merge adjacent partitions if both satisfy
524   /// the \p Predicate.
525   template <class UnaryPredicate>
mergeAdjacentPartitionsIf(UnaryPredicate Predicate)526   void mergeAdjacentPartitionsIf(UnaryPredicate Predicate) {
527     InstPartition *PrevMatch = nullptr;
528     for (auto I = PartitionContainer.begin(); I != PartitionContainer.end();) {
529       auto DoesMatch = Predicate(&*I);
530       if (PrevMatch == nullptr && DoesMatch) {
531         PrevMatch = &*I;
532         ++I;
533       } else if (PrevMatch != nullptr && DoesMatch) {
534         I->moveTo(*PrevMatch);
535         I = PartitionContainer.erase(I);
536       } else {
537         PrevMatch = nullptr;
538         ++I;
539       }
540     }
541   }
542 };
543 
544 /// \brief For each memory instruction, this class maintains difference of the
545 /// number of unsafe dependences that start out from this instruction minus
546 /// those that end here.
547 ///
548 /// By traversing the memory instructions in program order and accumulating this
549 /// number, we know whether any unsafe dependence crosses over a program point.
550 class MemoryInstructionDependences {
551   typedef MemoryDepChecker::Dependence Dependence;
552 
553 public:
554   struct Entry {
555     Instruction *Inst;
556     unsigned NumUnsafeDependencesStartOrEnd;
557 
Entry__anon191a722d0111::MemoryInstructionDependences::Entry558     Entry(Instruction *Inst) : Inst(Inst), NumUnsafeDependencesStartOrEnd(0) {}
559   };
560 
561   typedef SmallVector<Entry, 8> AccessesType;
562 
begin() const563   AccessesType::const_iterator begin() const { return Accesses.begin(); }
end() const564   AccessesType::const_iterator end() const { return Accesses.end(); }
565 
MemoryInstructionDependences(const SmallVectorImpl<Instruction * > & Instructions,const SmallVectorImpl<Dependence> & Dependences)566   MemoryInstructionDependences(
567       const SmallVectorImpl<Instruction *> &Instructions,
568       const SmallVectorImpl<Dependence> &Dependences) {
569     Accesses.append(Instructions.begin(), Instructions.end());
570 
571     DEBUG(dbgs() << "Backward dependences:\n");
572     for (auto &Dep : Dependences)
573       if (Dep.isPossiblyBackward()) {
574         // Note that the designations source and destination follow the program
575         // order, i.e. source is always first.  (The direction is given by the
576         // DepType.)
577         ++Accesses[Dep.Source].NumUnsafeDependencesStartOrEnd;
578         --Accesses[Dep.Destination].NumUnsafeDependencesStartOrEnd;
579 
580         DEBUG(Dep.print(dbgs(), 2, Instructions));
581       }
582   }
583 
584 private:
585   AccessesType Accesses;
586 };
587 
588 /// \brief The actual class performing the per-loop work.
589 class LoopDistributeForLoop {
590 public:
LoopDistributeForLoop(Loop * L,Function * F,LoopInfo * LI,DominatorTree * DT,ScalarEvolution * SE)591   LoopDistributeForLoop(Loop *L, Function *F, LoopInfo *LI, DominatorTree *DT,
592                         ScalarEvolution *SE)
593       : L(L), F(F), LI(LI), LAI(nullptr), DT(DT), SE(SE) {
594     setForced();
595   }
596 
597   /// \brief Try to distribute an inner-most loop.
processLoop(LoopAccessLegacyAnalysis * LAA)598   bool processLoop(LoopAccessLegacyAnalysis *LAA) {
599     assert(L->empty() && "Only process inner loops.");
600 
601     DEBUG(dbgs() << "\nLDist: In \"" << L->getHeader()->getParent()->getName()
602                  << "\" checking " << *L << "\n");
603 
604     BasicBlock *PH = L->getLoopPreheader();
605     if (!PH)
606       return fail("no preheader");
607     if (!L->getExitBlock())
608       return fail("multiple exit blocks");
609 
610     // LAA will check that we only have a single exiting block.
611     LAI = &LAA->getInfo(L);
612 
613     // Currently, we only distribute to isolate the part of the loop with
614     // dependence cycles to enable partial vectorization.
615     if (LAI->canVectorizeMemory())
616       return fail("memory operations are safe for vectorization");
617 
618     auto *Dependences = LAI->getDepChecker().getDependences();
619     if (!Dependences || Dependences->empty())
620       return fail("no unsafe dependences to isolate");
621 
622     InstPartitionContainer Partitions(L, LI, DT);
623 
624     // First, go through each memory operation and assign them to consecutive
625     // partitions (the order of partitions follows program order).  Put those
626     // with unsafe dependences into "cyclic" partition otherwise put each store
627     // in its own "non-cyclic" partition (we'll merge these later).
628     //
629     // Note that a memory operation (e.g. Load2 below) at a program point that
630     // has an unsafe dependence (Store3->Load1) spanning over it must be
631     // included in the same cyclic partition as the dependent operations.  This
632     // is to preserve the original program order after distribution.  E.g.:
633     //
634     //                NumUnsafeDependencesStartOrEnd  NumUnsafeDependencesActive
635     //  Load1   -.                     1                       0->1
636     //  Load2    | /Unsafe/            0                       1
637     //  Store3  -'                    -1                       1->0
638     //  Load4                          0                       0
639     //
640     // NumUnsafeDependencesActive > 0 indicates this situation and in this case
641     // we just keep assigning to the same cyclic partition until
642     // NumUnsafeDependencesActive reaches 0.
643     const MemoryDepChecker &DepChecker = LAI->getDepChecker();
644     MemoryInstructionDependences MID(DepChecker.getMemoryInstructions(),
645                                      *Dependences);
646 
647     int NumUnsafeDependencesActive = 0;
648     for (auto &InstDep : MID) {
649       Instruction *I = InstDep.Inst;
650       // We update NumUnsafeDependencesActive post-instruction, catch the
651       // start of a dependence directly via NumUnsafeDependencesStartOrEnd.
652       if (NumUnsafeDependencesActive ||
653           InstDep.NumUnsafeDependencesStartOrEnd > 0)
654         Partitions.addToCyclicPartition(I);
655       else
656         Partitions.addToNewNonCyclicPartition(I);
657       NumUnsafeDependencesActive += InstDep.NumUnsafeDependencesStartOrEnd;
658       assert(NumUnsafeDependencesActive >= 0 &&
659              "Negative number of dependences active");
660     }
661 
662     // Add partitions for values used outside.  These partitions can be out of
663     // order from the original program order.  This is OK because if the
664     // partition uses a load we will merge this partition with the original
665     // partition of the load that we set up in the previous loop (see
666     // mergeToAvoidDuplicatedLoads).
667     auto DefsUsedOutside = findDefsUsedOutsideOfLoop(L);
668     for (auto *Inst : DefsUsedOutside)
669       Partitions.addToNewNonCyclicPartition(Inst);
670 
671     DEBUG(dbgs() << "Seeded partitions:\n" << Partitions);
672     if (Partitions.getSize() < 2)
673       return fail("cannot isolate unsafe dependencies");
674 
675     // Run the merge heuristics: Merge non-cyclic adjacent partitions since we
676     // should be able to vectorize these together.
677     Partitions.mergeBeforePopulating();
678     DEBUG(dbgs() << "\nMerged partitions:\n" << Partitions);
679     if (Partitions.getSize() < 2)
680       return fail("cannot isolate unsafe dependencies");
681 
682     // Now, populate the partitions with non-memory operations.
683     Partitions.populateUsedSet();
684     DEBUG(dbgs() << "\nPopulated partitions:\n" << Partitions);
685 
686     // In order to preserve original lexical order for loads, keep them in the
687     // partition that we set up in the MemoryInstructionDependences loop.
688     if (Partitions.mergeToAvoidDuplicatedLoads()) {
689       DEBUG(dbgs() << "\nPartitions merged to ensure unique loads:\n"
690                    << Partitions);
691       if (Partitions.getSize() < 2)
692         return fail("cannot isolate unsafe dependencies");
693     }
694 
695     // Don't distribute the loop if we need too many SCEV run-time checks.
696     const SCEVUnionPredicate &Pred = LAI->getPSE().getUnionPredicate();
697     if (Pred.getComplexity() > (IsForced.getValueOr(false)
698                                     ? PragmaDistributeSCEVCheckThreshold
699                                     : DistributeSCEVCheckThreshold))
700       return fail("too many SCEV run-time checks needed.\n");
701 
702     DEBUG(dbgs() << "\nDistributing loop: " << *L << "\n");
703     // We're done forming the partitions set up the reverse mapping from
704     // instructions to partitions.
705     Partitions.setupPartitionIdOnInstructions();
706 
707     // To keep things simple have an empty preheader before we version or clone
708     // the loop.  (Also split if this has no predecessor, i.e. entry, because we
709     // rely on PH having a predecessor.)
710     if (!PH->getSinglePredecessor() || &*PH->begin() != PH->getTerminator())
711       SplitBlock(PH, PH->getTerminator(), DT, LI);
712 
713     // If we need run-time checks, version the loop now.
714     auto PtrToPartition = Partitions.computePartitionSetForPointers(*LAI);
715     const auto *RtPtrChecking = LAI->getRuntimePointerChecking();
716     const auto &AllChecks = RtPtrChecking->getChecks();
717     auto Checks = includeOnlyCrossPartitionChecks(AllChecks, PtrToPartition,
718                                                   RtPtrChecking);
719 
720     if (!Pred.isAlwaysTrue() || !Checks.empty()) {
721       DEBUG(dbgs() << "\nPointers:\n");
722       DEBUG(LAI->getRuntimePointerChecking()->printChecks(dbgs(), Checks));
723       LoopVersioning LVer(*LAI, L, LI, DT, SE, false);
724       LVer.setAliasChecks(std::move(Checks));
725       LVer.setSCEVChecks(LAI->getPSE().getUnionPredicate());
726       LVer.versionLoop(DefsUsedOutside);
727       LVer.annotateLoopWithNoAlias();
728     }
729 
730     // Create identical copies of the original loop for each partition and hook
731     // them up sequentially.
732     Partitions.cloneLoops();
733 
734     // Now, we remove the instruction from each loop that don't belong to that
735     // partition.
736     Partitions.removeUnusedInsts();
737     DEBUG(dbgs() << "\nAfter removing unused Instrs:\n");
738     DEBUG(Partitions.printBlocks());
739 
740     if (LDistVerify) {
741       LI->verify();
742       DT->verifyDomTree();
743     }
744 
745     ++NumLoopsDistributed;
746     // Report the success.
747     emitOptimizationRemark(F->getContext(), LDIST_NAME, *F, L->getStartLoc(),
748                            "distributed loop");
749     return true;
750   }
751 
752   /// \brief Provide diagnostics then \return with false.
fail(llvm::StringRef Message)753   bool fail(llvm::StringRef Message) {
754     LLVMContext &Ctx = F->getContext();
755     bool Forced = isForced().getValueOr(false);
756 
757     DEBUG(dbgs() << "Skipping; " << Message << "\n");
758 
759     // With Rpass-missed report that distribution failed.
760     emitOptimizationRemarkMissed(
761         Ctx, LDIST_NAME, *F, L->getStartLoc(),
762         "loop not distributed: use -Rpass-analysis=loop-distribute for more "
763         "info");
764 
765     // With Rpass-analysis report why.  This is on by default if distribution
766     // was requested explicitly.
767     emitOptimizationRemarkAnalysis(
768         Ctx, Forced ? DiagnosticInfoOptimizationRemarkAnalysis::AlwaysPrint
769                     : LDIST_NAME,
770         *F, L->getStartLoc(), Twine("loop not distributed: ") + Message);
771 
772     // Also issue a warning if distribution was requested explicitly but it
773     // failed.
774     if (Forced)
775       Ctx.diagnose(DiagnosticInfoOptimizationFailure(
776           *F, L->getStartLoc(), "loop not disributed: failed "
777                                 "explicitly specified loop distribution"));
778 
779     return false;
780   }
781 
782   /// \brief Return if distribution forced to be enabled/disabled for the loop.
783   ///
784   /// If the optional has a value, it indicates whether distribution was forced
785   /// to be enabled (true) or disabled (false).  If the optional has no value
786   /// distribution was not forced either way.
isForced() const787   const Optional<bool> &isForced() const { return IsForced; }
788 
789 private:
790   /// \brief Filter out checks between pointers from the same partition.
791   ///
792   /// \p PtrToPartition contains the partition number for pointers.  Partition
793   /// number -1 means that the pointer is used in multiple partitions.  In this
794   /// case we can't safely omit the check.
795   SmallVector<RuntimePointerChecking::PointerCheck, 4>
includeOnlyCrossPartitionChecks(const SmallVectorImpl<RuntimePointerChecking::PointerCheck> & AllChecks,const SmallVectorImpl<int> & PtrToPartition,const RuntimePointerChecking * RtPtrChecking)796   includeOnlyCrossPartitionChecks(
797       const SmallVectorImpl<RuntimePointerChecking::PointerCheck> &AllChecks,
798       const SmallVectorImpl<int> &PtrToPartition,
799       const RuntimePointerChecking *RtPtrChecking) {
800     SmallVector<RuntimePointerChecking::PointerCheck, 4> Checks;
801 
802     std::copy_if(AllChecks.begin(), AllChecks.end(), std::back_inserter(Checks),
803                  [&](const RuntimePointerChecking::PointerCheck &Check) {
804                    for (unsigned PtrIdx1 : Check.first->Members)
805                      for (unsigned PtrIdx2 : Check.second->Members)
806                        // Only include this check if there is a pair of pointers
807                        // that require checking and the pointers fall into
808                        // separate partitions.
809                        //
810                        // (Note that we already know at this point that the two
811                        // pointer groups need checking but it doesn't follow
812                        // that each pair of pointers within the two groups need
813                        // checking as well.
814                        //
815                        // In other words we don't want to include a check just
816                        // because there is a pair of pointers between the two
817                        // pointer groups that require checks and a different
818                        // pair whose pointers fall into different partitions.)
819                        if (RtPtrChecking->needsChecking(PtrIdx1, PtrIdx2) &&
820                            !RuntimePointerChecking::arePointersInSamePartition(
821                                PtrToPartition, PtrIdx1, PtrIdx2))
822                          return true;
823                    return false;
824                  });
825 
826     return Checks;
827   }
828 
829   /// \brief Check whether the loop metadata is forcing distribution to be
830   /// enabled/disabled.
setForced()831   void setForced() {
832     Optional<const MDOperand *> Value =
833         findStringMetadataForLoop(L, "llvm.loop.distribute.enable");
834     if (!Value)
835       return;
836 
837     const MDOperand *Op = *Value;
838     assert(Op && mdconst::hasa<ConstantInt>(*Op) && "invalid metadata");
839     IsForced = mdconst::extract<ConstantInt>(*Op)->getZExtValue();
840   }
841 
842   Loop *L;
843   Function *F;
844 
845   // Analyses used.
846   LoopInfo *LI;
847   const LoopAccessInfo *LAI;
848   DominatorTree *DT;
849   ScalarEvolution *SE;
850 
851   /// \brief Indicates whether distribution is forced to be enabled/disabled for
852   /// the loop.
853   ///
854   /// If the optional has a value, it indicates whether distribution was forced
855   /// to be enabled (true) or disabled (false).  If the optional has no value
856   /// distribution was not forced either way.
857   Optional<bool> IsForced;
858 };
859 
860 /// \brief The pass class.
861 class LoopDistribute : public FunctionPass {
862 public:
863   /// \p ProcessAllLoopsByDefault specifies whether loop distribution should be
864   /// performed by default.  Pass -enable-loop-distribute={0,1} overrides this
865   /// default.  We use this to keep LoopDistribution off by default when invoked
866   /// from the optimization pipeline but on when invoked explicitly from opt.
LoopDistribute(bool ProcessAllLoopsByDefault=true)867   LoopDistribute(bool ProcessAllLoopsByDefault = true)
868       : FunctionPass(ID), ProcessAllLoops(ProcessAllLoopsByDefault) {
869     // The default is set by the caller.
870     if (EnableLoopDistribute.getNumOccurrences() > 0)
871       ProcessAllLoops = EnableLoopDistribute;
872     initializeLoopDistributePass(*PassRegistry::getPassRegistry());
873   }
874 
runOnFunction(Function & F)875   bool runOnFunction(Function &F) override {
876     if (skipFunction(F))
877       return false;
878 
879     auto *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
880     auto *LAA = &getAnalysis<LoopAccessLegacyAnalysis>();
881     auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
882     auto *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
883 
884     // Build up a worklist of inner-loops to vectorize. This is necessary as the
885     // act of distributing a loop creates new loops and can invalidate iterators
886     // across the loops.
887     SmallVector<Loop *, 8> Worklist;
888 
889     for (Loop *TopLevelLoop : *LI)
890       for (Loop *L : depth_first(TopLevelLoop))
891         // We only handle inner-most loops.
892         if (L->empty())
893           Worklist.push_back(L);
894 
895     // Now walk the identified inner loops.
896     bool Changed = false;
897     for (Loop *L : Worklist) {
898       LoopDistributeForLoop LDL(L, &F, LI, DT, SE);
899 
900       // If distribution was forced for the specific loop to be
901       // enabled/disabled, follow that.  Otherwise use the global flag.
902       if (LDL.isForced().getValueOr(ProcessAllLoops))
903         Changed |= LDL.processLoop(LAA);
904     }
905 
906     // Process each loop nest in the function.
907     return Changed;
908   }
909 
getAnalysisUsage(AnalysisUsage & AU) const910   void getAnalysisUsage(AnalysisUsage &AU) const override {
911     AU.addRequired<ScalarEvolutionWrapperPass>();
912     AU.addRequired<LoopInfoWrapperPass>();
913     AU.addPreserved<LoopInfoWrapperPass>();
914     AU.addRequired<LoopAccessLegacyAnalysis>();
915     AU.addRequired<DominatorTreeWrapperPass>();
916     AU.addPreserved<DominatorTreeWrapperPass>();
917   }
918 
919   static char ID;
920 
921 private:
922   /// \brief Whether distribution should be on in this function.  The per-loop
923   /// pragma can override this.
924   bool ProcessAllLoops;
925 };
926 } // anonymous namespace
927 
928 char LoopDistribute::ID;
929 static const char ldist_name[] = "Loop Distribition";
930 
931 INITIALIZE_PASS_BEGIN(LoopDistribute, LDIST_NAME, ldist_name, false, false)
932 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
933 INITIALIZE_PASS_DEPENDENCY(LoopAccessLegacyAnalysis)
934 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
935 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
936 INITIALIZE_PASS_END(LoopDistribute, LDIST_NAME, ldist_name, false, false)
937 
938 namespace llvm {
createLoopDistributePass(bool ProcessAllLoopsByDefault)939 FunctionPass *createLoopDistributePass(bool ProcessAllLoopsByDefault) {
940   return new LoopDistribute(ProcessAllLoopsByDefault);
941 }
942 }
943