xref: /freebsd/contrib/llvm-project/llvm/lib/Transforms/Scalar/LoopDistribute.cpp (revision 5def4c47d4bd90b209b9b4a4ba9faec15846d8fd)
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/InstrTypes.h"
51 #include "llvm/IR/Instruction.h"
52 #include "llvm/IR/Instructions.h"
53 #include "llvm/IR/LLVMContext.h"
54 #include "llvm/IR/Metadata.h"
55 #include "llvm/IR/PassManager.h"
56 #include "llvm/IR/Value.h"
57 #include "llvm/InitializePasses.h"
58 #include "llvm/Pass.h"
59 #include "llvm/Support/Casting.h"
60 #include "llvm/Support/CommandLine.h"
61 #include "llvm/Support/Debug.h"
62 #include "llvm/Support/raw_ostream.h"
63 #include "llvm/Transforms/Scalar.h"
64 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
65 #include "llvm/Transforms/Utils/Cloning.h"
66 #include "llvm/Transforms/Utils/LoopUtils.h"
67 #include "llvm/Transforms/Utils/LoopVersioning.h"
68 #include "llvm/Transforms/Utils/ValueMapper.h"
69 #include <cassert>
70 #include <functional>
71 #include <list>
72 #include <tuple>
73 #include <utility>
74 
75 using namespace llvm;
76 
77 #define LDIST_NAME "loop-distribute"
78 #define DEBUG_TYPE LDIST_NAME
79 
80 /// @{
81 /// Metadata attribute names
82 static const char *const LLVMLoopDistributeFollowupAll =
83     "llvm.loop.distribute.followup_all";
84 static const char *const LLVMLoopDistributeFollowupCoincident =
85     "llvm.loop.distribute.followup_coincident";
86 static const char *const LLVMLoopDistributeFollowupSequential =
87     "llvm.loop.distribute.followup_sequential";
88 static const char *const LLVMLoopDistributeFollowupFallback =
89     "llvm.loop.distribute.followup_fallback";
90 /// @}
91 
92 static cl::opt<bool>
93     LDistVerify("loop-distribute-verify", cl::Hidden,
94                 cl::desc("Turn on DominatorTree and LoopInfo verification "
95                          "after Loop Distribution"),
96                 cl::init(false));
97 
98 static cl::opt<bool> DistributeNonIfConvertible(
99     "loop-distribute-non-if-convertible", cl::Hidden,
100     cl::desc("Whether to distribute into a loop that may not be "
101              "if-convertible by the loop vectorizer"),
102     cl::init(false));
103 
104 static cl::opt<unsigned> DistributeSCEVCheckThreshold(
105     "loop-distribute-scev-check-threshold", cl::init(8), cl::Hidden,
106     cl::desc("The maximum number of SCEV checks allowed for Loop "
107              "Distribution"));
108 
109 static cl::opt<unsigned> PragmaDistributeSCEVCheckThreshold(
110     "loop-distribute-scev-check-threshold-with-pragma", cl::init(128),
111     cl::Hidden,
112     cl::desc(
113         "The maximum number of SCEV checks allowed for Loop "
114         "Distribution for loop marked with #pragma loop distribute(enable)"));
115 
116 static cl::opt<bool> EnableLoopDistribute(
117     "enable-loop-distribute", cl::Hidden,
118     cl::desc("Enable the new, experimental LoopDistribution Pass"),
119     cl::init(false));
120 
121 STATISTIC(NumLoopsDistributed, "Number of loops distributed");
122 
123 namespace {
124 
125 /// Maintains the set of instructions of the loop for a partition before
126 /// cloning.  After cloning, it hosts the new loop.
127 class InstPartition {
128   using InstructionSet = SmallPtrSet<Instruction *, 8>;
129 
130 public:
131   InstPartition(Instruction *I, Loop *L, bool DepCycle = false)
132       : DepCycle(DepCycle), OrigLoop(L) {
133     Set.insert(I);
134   }
135 
136   /// Returns whether this partition contains a dependence cycle.
137   bool hasDepCycle() const { return DepCycle; }
138 
139   /// Adds an instruction to this partition.
140   void add(Instruction *I) { Set.insert(I); }
141 
142   /// Collection accessors.
143   InstructionSet::iterator begin() { return Set.begin(); }
144   InstructionSet::iterator end() { return Set.end(); }
145   InstructionSet::const_iterator begin() const { return Set.begin(); }
146   InstructionSet::const_iterator end() const { return Set.end(); }
147   bool empty() const { return Set.empty(); }
148 
149   /// Moves this partition into \p Other.  This partition becomes empty
150   /// after this.
151   void moveTo(InstPartition &Other) {
152     Other.Set.insert(Set.begin(), Set.end());
153     Set.clear();
154     Other.DepCycle |= DepCycle;
155   }
156 
157   /// Populates the partition with a transitive closure of all the
158   /// instructions that the seeded instructions dependent on.
159   void populateUsedSet() {
160     // FIXME: We currently don't use control-dependence but simply include all
161     // blocks (possibly empty at the end) and let simplifycfg mostly clean this
162     // up.
163     for (auto *B : OrigLoop->getBlocks())
164       Set.insert(B->getTerminator());
165 
166     // Follow the use-def chains to form a transitive closure of all the
167     // instructions that the originally seeded instructions depend on.
168     SmallVector<Instruction *, 8> Worklist(Set.begin(), Set.end());
169     while (!Worklist.empty()) {
170       Instruction *I = Worklist.pop_back_val();
171       // Insert instructions from the loop that we depend on.
172       for (Value *V : I->operand_values()) {
173         auto *I = dyn_cast<Instruction>(V);
174         if (I && OrigLoop->contains(I->getParent()) && Set.insert(I).second)
175           Worklist.push_back(I);
176       }
177     }
178   }
179 
180   /// Clones the original loop.
181   ///
182   /// Updates LoopInfo and DominatorTree using the information that block \p
183   /// LoopDomBB dominates the loop.
184   Loop *cloneLoopWithPreheader(BasicBlock *InsertBefore, BasicBlock *LoopDomBB,
185                                unsigned Index, LoopInfo *LI,
186                                DominatorTree *DT) {
187     ClonedLoop = ::cloneLoopWithPreheader(InsertBefore, LoopDomBB, OrigLoop,
188                                           VMap, Twine(".ldist") + Twine(Index),
189                                           LI, DT, ClonedLoopBlocks);
190     return ClonedLoop;
191   }
192 
193   /// The cloned loop.  If this partition is mapped to the original loop,
194   /// this is null.
195   const Loop *getClonedLoop() const { return ClonedLoop; }
196 
197   /// Returns the loop where this partition ends up after distribution.
198   /// If this partition is mapped to the original loop then use the block from
199   /// the loop.
200   Loop *getDistributedLoop() const {
201     return ClonedLoop ? ClonedLoop : OrigLoop;
202   }
203 
204   /// The VMap that is populated by cloning and then used in
205   /// remapinstruction to remap the cloned instructions.
206   ValueToValueMapTy &getVMap() { return VMap; }
207 
208   /// Remaps the cloned instructions using VMap.
209   void remapInstructions() {
210     remapInstructionsInBlocks(ClonedLoopBlocks, VMap);
211   }
212 
213   /// Based on the set of instructions selected for this partition,
214   /// removes the unnecessary ones.
215   void removeUnusedInsts() {
216     SmallVector<Instruction *, 8> Unused;
217 
218     for (auto *Block : OrigLoop->getBlocks())
219       for (auto &Inst : *Block)
220         if (!Set.count(&Inst)) {
221           Instruction *NewInst = &Inst;
222           if (!VMap.empty())
223             NewInst = cast<Instruction>(VMap[NewInst]);
224 
225           assert(!isa<BranchInst>(NewInst) &&
226                  "Branches are marked used early on");
227           Unused.push_back(NewInst);
228         }
229 
230     // Delete the instructions backwards, as it has a reduced likelihood of
231     // having to update as many def-use and use-def chains.
232     for (auto *Inst : reverse(Unused)) {
233       if (!Inst->use_empty())
234         Inst->replaceAllUsesWith(UndefValue::get(Inst->getType()));
235       Inst->eraseFromParent();
236     }
237   }
238 
239   void print() const {
240     if (DepCycle)
241       dbgs() << "  (cycle)\n";
242     for (auto *I : Set)
243       // Prefix with the block name.
244       dbgs() << "  " << I->getParent()->getName() << ":" << *I << "\n";
245   }
246 
247   void printBlocks() const {
248     for (auto *BB : getDistributedLoop()->getBlocks())
249       dbgs() << *BB;
250   }
251 
252 private:
253   /// Instructions from OrigLoop selected for this partition.
254   InstructionSet Set;
255 
256   /// Whether this partition contains a dependence cycle.
257   bool DepCycle;
258 
259   /// The original loop.
260   Loop *OrigLoop;
261 
262   /// The cloned loop.  If this partition is mapped to the original loop,
263   /// this is null.
264   Loop *ClonedLoop = nullptr;
265 
266   /// The blocks of ClonedLoop including the preheader.  If this
267   /// partition is mapped to the original loop, this is empty.
268   SmallVector<BasicBlock *, 8> ClonedLoopBlocks;
269 
270   /// These gets populated once the set of instructions have been
271   /// finalized. If this partition is mapped to the original loop, these are not
272   /// set.
273   ValueToValueMapTy VMap;
274 };
275 
276 /// Holds the set of Partitions.  It populates them, merges them and then
277 /// clones the loops.
278 class InstPartitionContainer {
279   using InstToPartitionIdT = DenseMap<Instruction *, int>;
280 
281 public:
282   InstPartitionContainer(Loop *L, LoopInfo *LI, DominatorTree *DT)
283       : L(L), LI(LI), DT(DT) {}
284 
285   /// Returns the number of partitions.
286   unsigned getSize() const { return PartitionContainer.size(); }
287 
288   /// Adds \p Inst into the current partition if that is marked to
289   /// contain cycles.  Otherwise start a new partition for it.
290   void addToCyclicPartition(Instruction *Inst) {
291     // If the current partition is non-cyclic.  Start a new one.
292     if (PartitionContainer.empty() || !PartitionContainer.back().hasDepCycle())
293       PartitionContainer.emplace_back(Inst, L, /*DepCycle=*/true);
294     else
295       PartitionContainer.back().add(Inst);
296   }
297 
298   /// Adds \p Inst into a partition that is not marked to contain
299   /// dependence cycles.
300   ///
301   //  Initially we isolate memory instructions into as many partitions as
302   //  possible, then later we may merge them back together.
303   void addToNewNonCyclicPartition(Instruction *Inst) {
304     PartitionContainer.emplace_back(Inst, L);
305   }
306 
307   /// Merges adjacent non-cyclic partitions.
308   ///
309   /// The idea is that we currently only want to isolate the non-vectorizable
310   /// partition.  We could later allow more distribution among these partition
311   /// too.
312   void mergeAdjacentNonCyclic() {
313     mergeAdjacentPartitionsIf(
314         [](const InstPartition *P) { return !P->hasDepCycle(); });
315   }
316 
317   /// If a partition contains only conditional stores, we won't vectorize
318   /// it.  Try to merge it with a previous cyclic partition.
319   void mergeNonIfConvertible() {
320     mergeAdjacentPartitionsIf([&](const InstPartition *Partition) {
321       if (Partition->hasDepCycle())
322         return true;
323 
324       // Now, check if all stores are conditional in this partition.
325       bool seenStore = false;
326 
327       for (auto *Inst : *Partition)
328         if (isa<StoreInst>(Inst)) {
329           seenStore = true;
330           if (!LoopAccessInfo::blockNeedsPredication(Inst->getParent(), L, DT))
331             return false;
332         }
333       return seenStore;
334     });
335   }
336 
337   /// Merges the partitions according to various heuristics.
338   void mergeBeforePopulating() {
339     mergeAdjacentNonCyclic();
340     if (!DistributeNonIfConvertible)
341       mergeNonIfConvertible();
342   }
343 
344   /// Merges partitions in order to ensure that no loads are duplicated.
345   ///
346   /// We can't duplicate loads because that could potentially reorder them.
347   /// LoopAccessAnalysis provides dependency information with the context that
348   /// the order of memory operation is preserved.
349   ///
350   /// Return if any partitions were merged.
351   bool mergeToAvoidDuplicatedLoads() {
352     using LoadToPartitionT = DenseMap<Instruction *, InstPartition *>;
353     using ToBeMergedT = EquivalenceClasses<InstPartition *>;
354 
355     LoadToPartitionT LoadToPartition;
356     ToBeMergedT ToBeMerged;
357 
358     // Step through the partitions and create equivalence between partitions
359     // that contain the same load.  Also put partitions in between them in the
360     // same equivalence class to avoid reordering of memory operations.
361     for (PartitionContainerT::iterator I = PartitionContainer.begin(),
362                                        E = PartitionContainer.end();
363          I != E; ++I) {
364       auto *PartI = &*I;
365 
366       // If a load occurs in two partitions PartI and PartJ, merge all
367       // partitions (PartI, PartJ] into PartI.
368       for (Instruction *Inst : *PartI)
369         if (isa<LoadInst>(Inst)) {
370           bool NewElt;
371           LoadToPartitionT::iterator LoadToPart;
372 
373           std::tie(LoadToPart, NewElt) =
374               LoadToPartition.insert(std::make_pair(Inst, PartI));
375           if (!NewElt) {
376             LLVM_DEBUG(dbgs()
377                        << "Merging partitions due to this load in multiple "
378                        << "partitions: " << PartI << ", " << LoadToPart->second
379                        << "\n"
380                        << *Inst << "\n");
381 
382             auto PartJ = I;
383             do {
384               --PartJ;
385               ToBeMerged.unionSets(PartI, &*PartJ);
386             } while (&*PartJ != LoadToPart->second);
387           }
388         }
389     }
390     if (ToBeMerged.empty())
391       return false;
392 
393     // Merge the member of an equivalence class into its class leader.  This
394     // makes the members empty.
395     for (ToBeMergedT::iterator I = ToBeMerged.begin(), E = ToBeMerged.end();
396          I != E; ++I) {
397       if (!I->isLeader())
398         continue;
399 
400       auto PartI = I->getData();
401       for (auto PartJ : make_range(std::next(ToBeMerged.member_begin(I)),
402                                    ToBeMerged.member_end())) {
403         PartJ->moveTo(*PartI);
404       }
405     }
406 
407     // Remove the empty partitions.
408     PartitionContainer.remove_if(
409         [](const InstPartition &P) { return P.empty(); });
410 
411     return true;
412   }
413 
414   /// Sets up the mapping between instructions to partitions.  If the
415   /// instruction is duplicated across multiple partitions, set the entry to -1.
416   void setupPartitionIdOnInstructions() {
417     int PartitionID = 0;
418     for (const auto &Partition : PartitionContainer) {
419       for (Instruction *Inst : Partition) {
420         bool NewElt;
421         InstToPartitionIdT::iterator Iter;
422 
423         std::tie(Iter, NewElt) =
424             InstToPartitionId.insert(std::make_pair(Inst, PartitionID));
425         if (!NewElt)
426           Iter->second = -1;
427       }
428       ++PartitionID;
429     }
430   }
431 
432   /// Populates the partition with everything that the seeding
433   /// instructions require.
434   void populateUsedSet() {
435     for (auto &P : PartitionContainer)
436       P.populateUsedSet();
437   }
438 
439   /// This performs the main chunk of the work of cloning the loops for
440   /// the partitions.
441   void cloneLoops() {
442     BasicBlock *OrigPH = L->getLoopPreheader();
443     // At this point the predecessor of the preheader is either the memcheck
444     // block or the top part of the original preheader.
445     BasicBlock *Pred = OrigPH->getSinglePredecessor();
446     assert(Pred && "Preheader does not have a single predecessor");
447     BasicBlock *ExitBlock = L->getExitBlock();
448     assert(ExitBlock && "No single exit block");
449     Loop *NewLoop;
450 
451     assert(!PartitionContainer.empty() && "at least two partitions expected");
452     // We're cloning the preheader along with the loop so we already made sure
453     // it was empty.
454     assert(&*OrigPH->begin() == OrigPH->getTerminator() &&
455            "preheader not empty");
456 
457     // Preserve the original loop ID for use after the transformation.
458     MDNode *OrigLoopID = L->getLoopID();
459 
460     // Create a loop for each partition except the last.  Clone the original
461     // loop before PH along with adding a preheader for the cloned loop.  Then
462     // update PH to point to the newly added preheader.
463     BasicBlock *TopPH = OrigPH;
464     unsigned Index = getSize() - 1;
465     for (auto I = std::next(PartitionContainer.rbegin()),
466               E = PartitionContainer.rend();
467          I != E; ++I, --Index, TopPH = NewLoop->getLoopPreheader()) {
468       auto *Part = &*I;
469 
470       NewLoop = Part->cloneLoopWithPreheader(TopPH, Pred, Index, LI, DT);
471 
472       Part->getVMap()[ExitBlock] = TopPH;
473       Part->remapInstructions();
474       setNewLoopID(OrigLoopID, Part);
475     }
476     Pred->getTerminator()->replaceUsesOfWith(OrigPH, TopPH);
477 
478     // Also set a new loop ID for the last loop.
479     setNewLoopID(OrigLoopID, &PartitionContainer.back());
480 
481     // Now go in forward order and update the immediate dominator for the
482     // preheaders with the exiting block of the previous loop.  Dominance
483     // within the loop is updated in cloneLoopWithPreheader.
484     for (auto Curr = PartitionContainer.cbegin(),
485               Next = std::next(PartitionContainer.cbegin()),
486               E = PartitionContainer.cend();
487          Next != E; ++Curr, ++Next)
488       DT->changeImmediateDominator(
489           Next->getDistributedLoop()->getLoopPreheader(),
490           Curr->getDistributedLoop()->getExitingBlock());
491   }
492 
493   /// Removes the dead instructions from the cloned loops.
494   void removeUnusedInsts() {
495     for (auto &Partition : PartitionContainer)
496       Partition.removeUnusedInsts();
497   }
498 
499   /// For each memory pointer, it computes the partitionId the pointer is
500   /// used in.
501   ///
502   /// This returns an array of int where the I-th entry corresponds to I-th
503   /// entry in LAI.getRuntimePointerCheck().  If the pointer is used in multiple
504   /// partitions its entry is set to -1.
505   SmallVector<int, 8>
506   computePartitionSetForPointers(const LoopAccessInfo &LAI) {
507     const RuntimePointerChecking *RtPtrCheck = LAI.getRuntimePointerChecking();
508 
509     unsigned N = RtPtrCheck->Pointers.size();
510     SmallVector<int, 8> PtrToPartitions(N);
511     for (unsigned I = 0; I < N; ++I) {
512       Value *Ptr = RtPtrCheck->Pointers[I].PointerValue;
513       auto Instructions =
514           LAI.getInstructionsForAccess(Ptr, RtPtrCheck->Pointers[I].IsWritePtr);
515 
516       int &Partition = PtrToPartitions[I];
517       // First set it to uninitialized.
518       Partition = -2;
519       for (Instruction *Inst : Instructions) {
520         // Note that this could be -1 if Inst is duplicated across multiple
521         // partitions.
522         int ThisPartition = this->InstToPartitionId[Inst];
523         if (Partition == -2)
524           Partition = ThisPartition;
525         // -1 means belonging to multiple partitions.
526         else if (Partition == -1)
527           break;
528         else if (Partition != (int)ThisPartition)
529           Partition = -1;
530       }
531       assert(Partition != -2 && "Pointer not belonging to any partition");
532     }
533 
534     return PtrToPartitions;
535   }
536 
537   void print(raw_ostream &OS) const {
538     unsigned Index = 0;
539     for (const auto &P : PartitionContainer) {
540       OS << "Partition " << Index++ << " (" << &P << "):\n";
541       P.print();
542     }
543   }
544 
545   void dump() const { print(dbgs()); }
546 
547 #ifndef NDEBUG
548   friend raw_ostream &operator<<(raw_ostream &OS,
549                                  const InstPartitionContainer &Partitions) {
550     Partitions.print(OS);
551     return OS;
552   }
553 #endif
554 
555   void printBlocks() const {
556     unsigned Index = 0;
557     for (const auto &P : PartitionContainer) {
558       dbgs() << "\nPartition " << Index++ << " (" << &P << "):\n";
559       P.printBlocks();
560     }
561   }
562 
563 private:
564   using PartitionContainerT = std::list<InstPartition>;
565 
566   /// List of partitions.
567   PartitionContainerT PartitionContainer;
568 
569   /// Mapping from Instruction to partition Id.  If the instruction
570   /// belongs to multiple partitions the entry contains -1.
571   InstToPartitionIdT InstToPartitionId;
572 
573   Loop *L;
574   LoopInfo *LI;
575   DominatorTree *DT;
576 
577   /// The control structure to merge adjacent partitions if both satisfy
578   /// the \p Predicate.
579   template <class UnaryPredicate>
580   void mergeAdjacentPartitionsIf(UnaryPredicate Predicate) {
581     InstPartition *PrevMatch = nullptr;
582     for (auto I = PartitionContainer.begin(); I != PartitionContainer.end();) {
583       auto DoesMatch = Predicate(&*I);
584       if (PrevMatch == nullptr && DoesMatch) {
585         PrevMatch = &*I;
586         ++I;
587       } else if (PrevMatch != nullptr && DoesMatch) {
588         I->moveTo(*PrevMatch);
589         I = PartitionContainer.erase(I);
590       } else {
591         PrevMatch = nullptr;
592         ++I;
593       }
594     }
595   }
596 
597   /// Assign new LoopIDs for the partition's cloned loop.
598   void setNewLoopID(MDNode *OrigLoopID, InstPartition *Part) {
599     Optional<MDNode *> PartitionID = makeFollowupLoopID(
600         OrigLoopID,
601         {LLVMLoopDistributeFollowupAll,
602          Part->hasDepCycle() ? LLVMLoopDistributeFollowupSequential
603                              : LLVMLoopDistributeFollowupCoincident});
604     if (PartitionID.hasValue()) {
605       Loop *NewLoop = Part->getDistributedLoop();
606       NewLoop->setLoopID(PartitionID.getValue());
607     }
608   }
609 };
610 
611 /// For each memory instruction, this class maintains difference of the
612 /// number of unsafe dependences that start out from this instruction minus
613 /// those that end here.
614 ///
615 /// By traversing the memory instructions in program order and accumulating this
616 /// number, we know whether any unsafe dependence crosses over a program point.
617 class MemoryInstructionDependences {
618   using Dependence = MemoryDepChecker::Dependence;
619 
620 public:
621   struct Entry {
622     Instruction *Inst;
623     unsigned NumUnsafeDependencesStartOrEnd = 0;
624 
625     Entry(Instruction *Inst) : Inst(Inst) {}
626   };
627 
628   using AccessesType = SmallVector<Entry, 8>;
629 
630   AccessesType::const_iterator begin() const { return Accesses.begin(); }
631   AccessesType::const_iterator end() const { return Accesses.end(); }
632 
633   MemoryInstructionDependences(
634       const SmallVectorImpl<Instruction *> &Instructions,
635       const SmallVectorImpl<Dependence> &Dependences) {
636     Accesses.append(Instructions.begin(), Instructions.end());
637 
638     LLVM_DEBUG(dbgs() << "Backward dependences:\n");
639     for (auto &Dep : Dependences)
640       if (Dep.isPossiblyBackward()) {
641         // Note that the designations source and destination follow the program
642         // order, i.e. source is always first.  (The direction is given by the
643         // DepType.)
644         ++Accesses[Dep.Source].NumUnsafeDependencesStartOrEnd;
645         --Accesses[Dep.Destination].NumUnsafeDependencesStartOrEnd;
646 
647         LLVM_DEBUG(Dep.print(dbgs(), 2, Instructions));
648       }
649   }
650 
651 private:
652   AccessesType Accesses;
653 };
654 
655 /// The actual class performing the per-loop work.
656 class LoopDistributeForLoop {
657 public:
658   LoopDistributeForLoop(Loop *L, Function *F, LoopInfo *LI, DominatorTree *DT,
659                         ScalarEvolution *SE, OptimizationRemarkEmitter *ORE)
660       : L(L), F(F), LI(LI), DT(DT), SE(SE), ORE(ORE) {
661     setForced();
662   }
663 
664   /// Try to distribute an inner-most loop.
665   bool processLoop(std::function<const LoopAccessInfo &(Loop &)> &GetLAA) {
666     assert(L->isInnermost() && "Only process inner loops.");
667 
668     LLVM_DEBUG(dbgs() << "\nLDist: In \""
669                       << L->getHeader()->getParent()->getName()
670                       << "\" checking " << *L << "\n");
671 
672     // Having a single exit block implies there's also one exiting block.
673     if (!L->getExitBlock())
674       return fail("MultipleExitBlocks", "multiple exit blocks");
675     if (!L->isLoopSimplifyForm())
676       return fail("NotLoopSimplifyForm",
677                   "loop is not in loop-simplify form");
678     if (!L->isRotatedForm())
679       return fail("NotBottomTested", "loop is not bottom tested");
680 
681     BasicBlock *PH = L->getLoopPreheader();
682 
683     LAI = &GetLAA(*L);
684 
685     // Currently, we only distribute to isolate the part of the loop with
686     // dependence cycles to enable partial vectorization.
687     if (LAI->canVectorizeMemory())
688       return fail("MemOpsCanBeVectorized",
689                   "memory operations are safe for vectorization");
690 
691     auto *Dependences = LAI->getDepChecker().getDependences();
692     if (!Dependences || Dependences->empty())
693       return fail("NoUnsafeDeps", "no unsafe dependences to isolate");
694 
695     InstPartitionContainer Partitions(L, LI, DT);
696 
697     // First, go through each memory operation and assign them to consecutive
698     // partitions (the order of partitions follows program order).  Put those
699     // with unsafe dependences into "cyclic" partition otherwise put each store
700     // in its own "non-cyclic" partition (we'll merge these later).
701     //
702     // Note that a memory operation (e.g. Load2 below) at a program point that
703     // has an unsafe dependence (Store3->Load1) spanning over it must be
704     // included in the same cyclic partition as the dependent operations.  This
705     // is to preserve the original program order after distribution.  E.g.:
706     //
707     //                NumUnsafeDependencesStartOrEnd  NumUnsafeDependencesActive
708     //  Load1   -.                     1                       0->1
709     //  Load2    | /Unsafe/            0                       1
710     //  Store3  -'                    -1                       1->0
711     //  Load4                          0                       0
712     //
713     // NumUnsafeDependencesActive > 0 indicates this situation and in this case
714     // we just keep assigning to the same cyclic partition until
715     // NumUnsafeDependencesActive reaches 0.
716     const MemoryDepChecker &DepChecker = LAI->getDepChecker();
717     MemoryInstructionDependences MID(DepChecker.getMemoryInstructions(),
718                                      *Dependences);
719 
720     int NumUnsafeDependencesActive = 0;
721     for (auto &InstDep : MID) {
722       Instruction *I = InstDep.Inst;
723       // We update NumUnsafeDependencesActive post-instruction, catch the
724       // start of a dependence directly via NumUnsafeDependencesStartOrEnd.
725       if (NumUnsafeDependencesActive ||
726           InstDep.NumUnsafeDependencesStartOrEnd > 0)
727         Partitions.addToCyclicPartition(I);
728       else
729         Partitions.addToNewNonCyclicPartition(I);
730       NumUnsafeDependencesActive += InstDep.NumUnsafeDependencesStartOrEnd;
731       assert(NumUnsafeDependencesActive >= 0 &&
732              "Negative number of dependences active");
733     }
734 
735     // Add partitions for values used outside.  These partitions can be out of
736     // order from the original program order.  This is OK because if the
737     // partition uses a load we will merge this partition with the original
738     // partition of the load that we set up in the previous loop (see
739     // mergeToAvoidDuplicatedLoads).
740     auto DefsUsedOutside = findDefsUsedOutsideOfLoop(L);
741     for (auto *Inst : DefsUsedOutside)
742       Partitions.addToNewNonCyclicPartition(Inst);
743 
744     LLVM_DEBUG(dbgs() << "Seeded partitions:\n" << Partitions);
745     if (Partitions.getSize() < 2)
746       return fail("CantIsolateUnsafeDeps",
747                   "cannot isolate unsafe dependencies");
748 
749     // Run the merge heuristics: Merge non-cyclic adjacent partitions since we
750     // should be able to vectorize these together.
751     Partitions.mergeBeforePopulating();
752     LLVM_DEBUG(dbgs() << "\nMerged partitions:\n" << Partitions);
753     if (Partitions.getSize() < 2)
754       return fail("CantIsolateUnsafeDeps",
755                   "cannot isolate unsafe dependencies");
756 
757     // Now, populate the partitions with non-memory operations.
758     Partitions.populateUsedSet();
759     LLVM_DEBUG(dbgs() << "\nPopulated partitions:\n" << Partitions);
760 
761     // In order to preserve original lexical order for loads, keep them in the
762     // partition that we set up in the MemoryInstructionDependences loop.
763     if (Partitions.mergeToAvoidDuplicatedLoads()) {
764       LLVM_DEBUG(dbgs() << "\nPartitions merged to ensure unique loads:\n"
765                         << Partitions);
766       if (Partitions.getSize() < 2)
767         return fail("CantIsolateUnsafeDeps",
768                     "cannot isolate unsafe dependencies");
769     }
770 
771     // Don't distribute the loop if we need too many SCEV run-time checks, or
772     // any if it's illegal.
773     const SCEVUnionPredicate &Pred = LAI->getPSE().getUnionPredicate();
774     if (LAI->hasConvergentOp() && !Pred.isAlwaysTrue()) {
775       return fail("RuntimeCheckWithConvergent",
776                   "may not insert runtime check with convergent operation");
777     }
778 
779     if (Pred.getComplexity() > (IsForced.getValueOr(false)
780                                     ? PragmaDistributeSCEVCheckThreshold
781                                     : DistributeSCEVCheckThreshold))
782       return fail("TooManySCEVRuntimeChecks",
783                   "too many SCEV run-time checks needed.\n");
784 
785     if (!IsForced.getValueOr(false) && hasDisableAllTransformsHint(L))
786       return fail("HeuristicDisabled", "distribution heuristic disabled");
787 
788     LLVM_DEBUG(dbgs() << "\nDistributing loop: " << *L << "\n");
789     // We're done forming the partitions set up the reverse mapping from
790     // instructions to partitions.
791     Partitions.setupPartitionIdOnInstructions();
792 
793     // If we need run-time checks, version the loop now.
794     auto PtrToPartition = Partitions.computePartitionSetForPointers(*LAI);
795     const auto *RtPtrChecking = LAI->getRuntimePointerChecking();
796     const auto &AllChecks = RtPtrChecking->getChecks();
797     auto Checks = includeOnlyCrossPartitionChecks(AllChecks, PtrToPartition,
798                                                   RtPtrChecking);
799 
800     if (LAI->hasConvergentOp() && !Checks.empty()) {
801       return fail("RuntimeCheckWithConvergent",
802                   "may not insert runtime check with convergent operation");
803     }
804 
805     // To keep things simple have an empty preheader before we version or clone
806     // the loop.  (Also split if this has no predecessor, i.e. entry, because we
807     // rely on PH having a predecessor.)
808     if (!PH->getSinglePredecessor() || &*PH->begin() != PH->getTerminator())
809       SplitBlock(PH, PH->getTerminator(), DT, LI);
810 
811     if (!Pred.isAlwaysTrue() || !Checks.empty()) {
812       assert(!LAI->hasConvergentOp() && "inserting illegal loop versioning");
813 
814       MDNode *OrigLoopID = L->getLoopID();
815 
816       LLVM_DEBUG(dbgs() << "\nPointers:\n");
817       LLVM_DEBUG(LAI->getRuntimePointerChecking()->printChecks(dbgs(), Checks));
818       LoopVersioning LVer(*LAI, Checks, L, LI, DT, SE);
819       LVer.versionLoop(DefsUsedOutside);
820       LVer.annotateLoopWithNoAlias();
821 
822       // The unversioned loop will not be changed, so we inherit all attributes
823       // from the original loop, but remove the loop distribution metadata to
824       // avoid to distribute it again.
825       MDNode *UnversionedLoopID =
826           makeFollowupLoopID(OrigLoopID,
827                              {LLVMLoopDistributeFollowupAll,
828                               LLVMLoopDistributeFollowupFallback},
829                              "llvm.loop.distribute.", true)
830               .getValue();
831       LVer.getNonVersionedLoop()->setLoopID(UnversionedLoopID);
832     }
833 
834     // Create identical copies of the original loop for each partition and hook
835     // them up sequentially.
836     Partitions.cloneLoops();
837 
838     // Now, we remove the instruction from each loop that don't belong to that
839     // partition.
840     Partitions.removeUnusedInsts();
841     LLVM_DEBUG(dbgs() << "\nAfter removing unused Instrs:\n");
842     LLVM_DEBUG(Partitions.printBlocks());
843 
844     if (LDistVerify) {
845       LI->verify(*DT);
846       assert(DT->verify(DominatorTree::VerificationLevel::Fast));
847     }
848 
849     ++NumLoopsDistributed;
850     // Report the success.
851     ORE->emit([&]() {
852       return OptimizationRemark(LDIST_NAME, "Distribute", L->getStartLoc(),
853                                 L->getHeader())
854              << "distributed loop";
855     });
856     return true;
857   }
858 
859   /// Provide diagnostics then \return with false.
860   bool fail(StringRef RemarkName, StringRef Message) {
861     LLVMContext &Ctx = F->getContext();
862     bool Forced = isForced().getValueOr(false);
863 
864     LLVM_DEBUG(dbgs() << "Skipping; " << Message << "\n");
865 
866     // With Rpass-missed report that distribution failed.
867     ORE->emit([&]() {
868       return OptimizationRemarkMissed(LDIST_NAME, "NotDistributed",
869                                       L->getStartLoc(), L->getHeader())
870              << "loop not distributed: use -Rpass-analysis=loop-distribute for "
871                 "more "
872                 "info";
873     });
874 
875     // With Rpass-analysis report why.  This is on by default if distribution
876     // was requested explicitly.
877     ORE->emit(OptimizationRemarkAnalysis(
878                   Forced ? OptimizationRemarkAnalysis::AlwaysPrint : LDIST_NAME,
879                   RemarkName, L->getStartLoc(), L->getHeader())
880               << "loop not distributed: " << Message);
881 
882     // Also issue a warning if distribution was requested explicitly but it
883     // failed.
884     if (Forced)
885       Ctx.diagnose(DiagnosticInfoOptimizationFailure(
886           *F, L->getStartLoc(), "loop not distributed: failed "
887                                 "explicitly specified loop distribution"));
888 
889     return false;
890   }
891 
892   /// Return if distribution forced to be enabled/disabled for the loop.
893   ///
894   /// If the optional has a value, it indicates whether distribution was forced
895   /// to be enabled (true) or disabled (false).  If the optional has no value
896   /// distribution was not forced either way.
897   const Optional<bool> &isForced() const { return IsForced; }
898 
899 private:
900   /// Filter out checks between pointers from the same partition.
901   ///
902   /// \p PtrToPartition contains the partition number for pointers.  Partition
903   /// number -1 means that the pointer is used in multiple partitions.  In this
904   /// case we can't safely omit the check.
905   SmallVector<RuntimePointerCheck, 4> includeOnlyCrossPartitionChecks(
906       const SmallVectorImpl<RuntimePointerCheck> &AllChecks,
907       const SmallVectorImpl<int> &PtrToPartition,
908       const RuntimePointerChecking *RtPtrChecking) {
909     SmallVector<RuntimePointerCheck, 4> Checks;
910 
911     copy_if(AllChecks, std::back_inserter(Checks),
912             [&](const RuntimePointerCheck &Check) {
913               for (unsigned PtrIdx1 : Check.first->Members)
914                 for (unsigned PtrIdx2 : Check.second->Members)
915                   // Only include this check if there is a pair of pointers
916                   // that require checking and the pointers fall into
917                   // separate partitions.
918                   //
919                   // (Note that we already know at this point that the two
920                   // pointer groups need checking but it doesn't follow
921                   // that each pair of pointers within the two groups need
922                   // checking as well.
923                   //
924                   // In other words we don't want to include a check just
925                   // because there is a pair of pointers between the two
926                   // pointer groups that require checks and a different
927                   // pair whose pointers fall into different partitions.)
928                   if (RtPtrChecking->needsChecking(PtrIdx1, PtrIdx2) &&
929                       !RuntimePointerChecking::arePointersInSamePartition(
930                           PtrToPartition, PtrIdx1, PtrIdx2))
931                     return true;
932               return false;
933             });
934 
935     return Checks;
936   }
937 
938   /// Check whether the loop metadata is forcing distribution to be
939   /// enabled/disabled.
940   void setForced() {
941     Optional<const MDOperand *> Value =
942         findStringMetadataForLoop(L, "llvm.loop.distribute.enable");
943     if (!Value)
944       return;
945 
946     const MDOperand *Op = *Value;
947     assert(Op && mdconst::hasa<ConstantInt>(*Op) && "invalid metadata");
948     IsForced = mdconst::extract<ConstantInt>(*Op)->getZExtValue();
949   }
950 
951   Loop *L;
952   Function *F;
953 
954   // Analyses used.
955   LoopInfo *LI;
956   const LoopAccessInfo *LAI = nullptr;
957   DominatorTree *DT;
958   ScalarEvolution *SE;
959   OptimizationRemarkEmitter *ORE;
960 
961   /// Indicates whether distribution is forced to be enabled/disabled for
962   /// the loop.
963   ///
964   /// If the optional has a value, it indicates whether distribution was forced
965   /// to be enabled (true) or disabled (false).  If the optional has no value
966   /// distribution was not forced either way.
967   Optional<bool> IsForced;
968 };
969 
970 } // end anonymous namespace
971 
972 /// Shared implementation between new and old PMs.
973 static bool runImpl(Function &F, LoopInfo *LI, DominatorTree *DT,
974                     ScalarEvolution *SE, OptimizationRemarkEmitter *ORE,
975                     std::function<const LoopAccessInfo &(Loop &)> &GetLAA) {
976   // Build up a worklist of inner-loops to vectorize. This is necessary as the
977   // act of distributing a loop creates new loops and can invalidate iterators
978   // across the loops.
979   SmallVector<Loop *, 8> Worklist;
980 
981   for (Loop *TopLevelLoop : *LI)
982     for (Loop *L : depth_first(TopLevelLoop))
983       // We only handle inner-most loops.
984       if (L->isInnermost())
985         Worklist.push_back(L);
986 
987   // Now walk the identified inner loops.
988   bool Changed = false;
989   for (Loop *L : Worklist) {
990     LoopDistributeForLoop LDL(L, &F, LI, DT, SE, ORE);
991 
992     // If distribution was forced for the specific loop to be
993     // enabled/disabled, follow that.  Otherwise use the global flag.
994     if (LDL.isForced().getValueOr(EnableLoopDistribute))
995       Changed |= LDL.processLoop(GetLAA);
996   }
997 
998   // Process each loop nest in the function.
999   return Changed;
1000 }
1001 
1002 namespace {
1003 
1004 /// The pass class.
1005 class LoopDistributeLegacy : public FunctionPass {
1006 public:
1007   static char ID;
1008 
1009   LoopDistributeLegacy() : FunctionPass(ID) {
1010     // The default is set by the caller.
1011     initializeLoopDistributeLegacyPass(*PassRegistry::getPassRegistry());
1012   }
1013 
1014   bool runOnFunction(Function &F) override {
1015     if (skipFunction(F))
1016       return false;
1017 
1018     auto *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1019     auto *LAA = &getAnalysis<LoopAccessLegacyAnalysis>();
1020     auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1021     auto *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
1022     auto *ORE = &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE();
1023     std::function<const LoopAccessInfo &(Loop &)> GetLAA =
1024         [&](Loop &L) -> const LoopAccessInfo & { return LAA->getInfo(&L); };
1025 
1026     return runImpl(F, LI, DT, SE, ORE, GetLAA);
1027   }
1028 
1029   void getAnalysisUsage(AnalysisUsage &AU) const override {
1030     AU.addRequired<ScalarEvolutionWrapperPass>();
1031     AU.addRequired<LoopInfoWrapperPass>();
1032     AU.addPreserved<LoopInfoWrapperPass>();
1033     AU.addRequired<LoopAccessLegacyAnalysis>();
1034     AU.addRequired<DominatorTreeWrapperPass>();
1035     AU.addPreserved<DominatorTreeWrapperPass>();
1036     AU.addRequired<OptimizationRemarkEmitterWrapperPass>();
1037     AU.addPreserved<GlobalsAAWrapperPass>();
1038   }
1039 };
1040 
1041 } // end anonymous namespace
1042 
1043 PreservedAnalyses LoopDistributePass::run(Function &F,
1044                                           FunctionAnalysisManager &AM) {
1045   auto &LI = AM.getResult<LoopAnalysis>(F);
1046   auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
1047   auto &SE = AM.getResult<ScalarEvolutionAnalysis>(F);
1048   auto &ORE = AM.getResult<OptimizationRemarkEmitterAnalysis>(F);
1049 
1050   // We don't directly need these analyses but they're required for loop
1051   // analyses so provide them below.
1052   auto &AA = AM.getResult<AAManager>(F);
1053   auto &AC = AM.getResult<AssumptionAnalysis>(F);
1054   auto &TTI = AM.getResult<TargetIRAnalysis>(F);
1055   auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
1056 
1057   auto &LAM = AM.getResult<LoopAnalysisManagerFunctionProxy>(F).getManager();
1058   std::function<const LoopAccessInfo &(Loop &)> GetLAA =
1059       [&](Loop &L) -> const LoopAccessInfo & {
1060     LoopStandardAnalysisResults AR = {AA,  AC,  DT,      LI,     SE,
1061                                       TLI, TTI, nullptr, nullptr};
1062     return LAM.getResult<LoopAccessAnalysis>(L, AR);
1063   };
1064 
1065   bool Changed = runImpl(F, &LI, &DT, &SE, &ORE, GetLAA);
1066   if (!Changed)
1067     return PreservedAnalyses::all();
1068   PreservedAnalyses PA;
1069   PA.preserve<LoopAnalysis>();
1070   PA.preserve<DominatorTreeAnalysis>();
1071   PA.preserve<GlobalsAA>();
1072   return PA;
1073 }
1074 
1075 char LoopDistributeLegacy::ID;
1076 
1077 static const char ldist_name[] = "Loop Distribution";
1078 
1079 INITIALIZE_PASS_BEGIN(LoopDistributeLegacy, LDIST_NAME, ldist_name, false,
1080                       false)
1081 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
1082 INITIALIZE_PASS_DEPENDENCY(LoopAccessLegacyAnalysis)
1083 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
1084 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
1085 INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass)
1086 INITIALIZE_PASS_END(LoopDistributeLegacy, LDIST_NAME, ldist_name, false, false)
1087 
1088 FunctionPass *llvm::createLoopDistributePass() { return new LoopDistributeLegacy(); }
1089