1 //===- LoopFuse.cpp - Loop Fusion 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 /// \file
10 /// This file implements the loop fusion pass.
11 /// The implementation is largely based on the following document:
12 ///
13 ///       Code Transformations to Augment the Scope of Loop Fusion in a
14 ///         Production Compiler
15 ///       Christopher Mark Barton
16 ///       MSc Thesis
17 ///       https://webdocs.cs.ualberta.ca/~amaral/thesis/ChristopherBartonMSc.pdf
18 ///
19 /// The general approach taken is to collect sets of control flow equivalent
20 /// loops and test whether they can be fused. The necessary conditions for
21 /// fusion are:
22 ///    1. The loops must be adjacent (there cannot be any statements between
23 ///       the two loops).
24 ///    2. The loops must be conforming (they must execute the same number of
25 ///       iterations).
26 ///    3. The loops must be control flow equivalent (if one loop executes, the
27 ///       other is guaranteed to execute).
28 ///    4. There cannot be any negative distance dependencies between the loops.
29 /// If all of these conditions are satisfied, it is safe to fuse the loops.
30 ///
31 /// This implementation creates FusionCandidates that represent the loop and the
32 /// necessary information needed by fusion. It then operates on the fusion
33 /// candidates, first confirming that the candidate is eligible for fusion. The
34 /// candidates are then collected into control flow equivalent sets, sorted in
35 /// dominance order. Each set of control flow equivalent candidates is then
36 /// traversed, attempting to fuse pairs of candidates in the set. If all
37 /// requirements for fusion are met, the two candidates are fused, creating a
38 /// new (fused) candidate which is then added back into the set to consider for
39 /// additional fusion.
40 ///
41 /// This implementation currently does not make any modifications to remove
42 /// conditions for fusion. Code transformations to make loops conform to each of
43 /// the conditions for fusion are discussed in more detail in the document
44 /// above. These can be added to the current implementation in the future.
45 //===----------------------------------------------------------------------===//
46 
47 #include "llvm/Transforms/Scalar/LoopFuse.h"
48 #include "llvm/ADT/Statistic.h"
49 #include "llvm/Analysis/DependenceAnalysis.h"
50 #include "llvm/Analysis/DomTreeUpdater.h"
51 #include "llvm/Analysis/LoopInfo.h"
52 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
53 #include "llvm/Analysis/PostDominators.h"
54 #include "llvm/Analysis/ScalarEvolution.h"
55 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
56 #include "llvm/IR/Function.h"
57 #include "llvm/IR/Verifier.h"
58 #include "llvm/InitializePasses.h"
59 #include "llvm/Pass.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.h"
65 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
66 #include "llvm/Transforms/Utils/CodeMoverUtils.h"
67 
68 using namespace llvm;
69 
70 #define DEBUG_TYPE "loop-fusion"
71 
72 STATISTIC(FuseCounter, "Loops fused");
73 STATISTIC(NumFusionCandidates, "Number of candidates for loop fusion");
74 STATISTIC(InvalidPreheader, "Loop has invalid preheader");
75 STATISTIC(InvalidHeader, "Loop has invalid header");
76 STATISTIC(InvalidExitingBlock, "Loop has invalid exiting blocks");
77 STATISTIC(InvalidExitBlock, "Loop has invalid exit block");
78 STATISTIC(InvalidLatch, "Loop has invalid latch");
79 STATISTIC(InvalidLoop, "Loop is invalid");
80 STATISTIC(AddressTakenBB, "Basic block has address taken");
81 STATISTIC(MayThrowException, "Loop may throw an exception");
82 STATISTIC(ContainsVolatileAccess, "Loop contains a volatile access");
83 STATISTIC(NotSimplifiedForm, "Loop is not in simplified form");
84 STATISTIC(InvalidDependencies, "Dependencies prevent fusion");
85 STATISTIC(UnknownTripCount, "Loop has unknown trip count");
86 STATISTIC(UncomputableTripCount, "SCEV cannot compute trip count of loop");
87 STATISTIC(NonEqualTripCount, "Loop trip counts are not the same");
88 STATISTIC(NonAdjacent, "Loops are not adjacent");
89 STATISTIC(NonEmptyPreheader, "Loop has a non-empty preheader");
90 STATISTIC(FusionNotBeneficial, "Fusion is not beneficial");
91 STATISTIC(NonIdenticalGuards, "Candidates have different guards");
92 STATISTIC(NonEmptyExitBlock, "Candidate has a non-empty exit block");
93 STATISTIC(NonEmptyGuardBlock, "Candidate has a non-empty guard block");
94 STATISTIC(NotRotated, "Candidate is not rotated");
95 
96 enum FusionDependenceAnalysisChoice {
97   FUSION_DEPENDENCE_ANALYSIS_SCEV,
98   FUSION_DEPENDENCE_ANALYSIS_DA,
99   FUSION_DEPENDENCE_ANALYSIS_ALL,
100 };
101 
102 static cl::opt<FusionDependenceAnalysisChoice> FusionDependenceAnalysis(
103     "loop-fusion-dependence-analysis",
104     cl::desc("Which dependence analysis should loop fusion use?"),
105     cl::values(clEnumValN(FUSION_DEPENDENCE_ANALYSIS_SCEV, "scev",
106                           "Use the scalar evolution interface"),
107                clEnumValN(FUSION_DEPENDENCE_ANALYSIS_DA, "da",
108                           "Use the dependence analysis interface"),
109                clEnumValN(FUSION_DEPENDENCE_ANALYSIS_ALL, "all",
110                           "Use all available analyses")),
111     cl::Hidden, cl::init(FUSION_DEPENDENCE_ANALYSIS_ALL), cl::ZeroOrMore);
112 
113 #ifndef NDEBUG
114 static cl::opt<bool>
115     VerboseFusionDebugging("loop-fusion-verbose-debug",
116                            cl::desc("Enable verbose debugging for Loop Fusion"),
117                            cl::Hidden, cl::init(false), cl::ZeroOrMore);
118 #endif
119 
120 namespace {
121 /// This class is used to represent a candidate for loop fusion. When it is
122 /// constructed, it checks the conditions for loop fusion to ensure that it
123 /// represents a valid candidate. It caches several parts of a loop that are
124 /// used throughout loop fusion (e.g., loop preheader, loop header, etc) instead
125 /// of continually querying the underlying Loop to retrieve these values. It is
126 /// assumed these will not change throughout loop fusion.
127 ///
128 /// The invalidate method should be used to indicate that the FusionCandidate is
129 /// no longer a valid candidate for fusion. Similarly, the isValid() method can
130 /// be used to ensure that the FusionCandidate is still valid for fusion.
131 struct FusionCandidate {
132   /// Cache of parts of the loop used throughout loop fusion. These should not
133   /// need to change throughout the analysis and transformation.
134   /// These parts are cached to avoid repeatedly looking up in the Loop class.
135 
136   /// Preheader of the loop this candidate represents
137   BasicBlock *Preheader;
138   /// Header of the loop this candidate represents
139   BasicBlock *Header;
140   /// Blocks in the loop that exit the loop
141   BasicBlock *ExitingBlock;
142   /// The successor block of this loop (where the exiting blocks go to)
143   BasicBlock *ExitBlock;
144   /// Latch of the loop
145   BasicBlock *Latch;
146   /// The loop that this fusion candidate represents
147   Loop *L;
148   /// Vector of instructions in this loop that read from memory
149   SmallVector<Instruction *, 16> MemReads;
150   /// Vector of instructions in this loop that write to memory
151   SmallVector<Instruction *, 16> MemWrites;
152   /// Are all of the members of this fusion candidate still valid
153   bool Valid;
154   /// Guard branch of the loop, if it exists
155   BranchInst *GuardBranch;
156 
157   /// Dominator and PostDominator trees are needed for the
158   /// FusionCandidateCompare function, required by FusionCandidateSet to
159   /// determine where the FusionCandidate should be inserted into the set. These
160   /// are used to establish ordering of the FusionCandidates based on dominance.
161   const DominatorTree *DT;
162   const PostDominatorTree *PDT;
163 
164   OptimizationRemarkEmitter &ORE;
165 
166   FusionCandidate(Loop *L, const DominatorTree *DT,
167                   const PostDominatorTree *PDT, OptimizationRemarkEmitter &ORE)
168       : Preheader(L->getLoopPreheader()), Header(L->getHeader()),
169         ExitingBlock(L->getExitingBlock()), ExitBlock(L->getExitBlock()),
170         Latch(L->getLoopLatch()), L(L), Valid(true),
171         GuardBranch(L->getLoopGuardBranch()), DT(DT), PDT(PDT), ORE(ORE) {
172 
173     // Walk over all blocks in the loop and check for conditions that may
174     // prevent fusion. For each block, walk over all instructions and collect
175     // the memory reads and writes If any instructions that prevent fusion are
176     // found, invalidate this object and return.
177     for (BasicBlock *BB : L->blocks()) {
178       if (BB->hasAddressTaken()) {
179         invalidate();
180         reportInvalidCandidate(AddressTakenBB);
181         return;
182       }
183 
184       for (Instruction &I : *BB) {
185         if (I.mayThrow()) {
186           invalidate();
187           reportInvalidCandidate(MayThrowException);
188           return;
189         }
190         if (StoreInst *SI = dyn_cast<StoreInst>(&I)) {
191           if (SI->isVolatile()) {
192             invalidate();
193             reportInvalidCandidate(ContainsVolatileAccess);
194             return;
195           }
196         }
197         if (LoadInst *LI = dyn_cast<LoadInst>(&I)) {
198           if (LI->isVolatile()) {
199             invalidate();
200             reportInvalidCandidate(ContainsVolatileAccess);
201             return;
202           }
203         }
204         if (I.mayWriteToMemory())
205           MemWrites.push_back(&I);
206         if (I.mayReadFromMemory())
207           MemReads.push_back(&I);
208       }
209     }
210   }
211 
212   /// Check if all members of the class are valid.
213   bool isValid() const {
214     return Preheader && Header && ExitingBlock && ExitBlock && Latch && L &&
215            !L->isInvalid() && Valid;
216   }
217 
218   /// Verify that all members are in sync with the Loop object.
219   void verify() const {
220     assert(isValid() && "Candidate is not valid!!");
221     assert(!L->isInvalid() && "Loop is invalid!");
222     assert(Preheader == L->getLoopPreheader() && "Preheader is out of sync");
223     assert(Header == L->getHeader() && "Header is out of sync");
224     assert(ExitingBlock == L->getExitingBlock() &&
225            "Exiting Blocks is out of sync");
226     assert(ExitBlock == L->getExitBlock() && "Exit block is out of sync");
227     assert(Latch == L->getLoopLatch() && "Latch is out of sync");
228   }
229 
230   /// Get the entry block for this fusion candidate.
231   ///
232   /// If this fusion candidate represents a guarded loop, the entry block is the
233   /// loop guard block. If it represents an unguarded loop, the entry block is
234   /// the preheader of the loop.
235   BasicBlock *getEntryBlock() const {
236     if (GuardBranch)
237       return GuardBranch->getParent();
238     else
239       return Preheader;
240   }
241 
242   /// Given a guarded loop, get the successor of the guard that is not in the
243   /// loop.
244   ///
245   /// This method returns the successor of the loop guard that is not located
246   /// within the loop (i.e., the successor of the guard that is not the
247   /// preheader).
248   /// This method is only valid for guarded loops.
249   BasicBlock *getNonLoopBlock() const {
250     assert(GuardBranch && "Only valid on guarded loops.");
251     assert(GuardBranch->isConditional() &&
252            "Expecting guard to be a conditional branch.");
253     return (GuardBranch->getSuccessor(0) == Preheader)
254                ? GuardBranch->getSuccessor(1)
255                : GuardBranch->getSuccessor(0);
256   }
257 
258 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
259   LLVM_DUMP_METHOD void dump() const {
260     dbgs() << "\tGuardBranch: ";
261     if (GuardBranch)
262       dbgs() << *GuardBranch;
263     else
264       dbgs() << "nullptr";
265     dbgs() << "\n"
266            << (GuardBranch ? GuardBranch->getName() : "nullptr") << "\n"
267            << "\tPreheader: " << (Preheader ? Preheader->getName() : "nullptr")
268            << "\n"
269            << "\tHeader: " << (Header ? Header->getName() : "nullptr") << "\n"
270            << "\tExitingBB: "
271            << (ExitingBlock ? ExitingBlock->getName() : "nullptr") << "\n"
272            << "\tExitBB: " << (ExitBlock ? ExitBlock->getName() : "nullptr")
273            << "\n"
274            << "\tLatch: " << (Latch ? Latch->getName() : "nullptr") << "\n"
275            << "\tEntryBlock: "
276            << (getEntryBlock() ? getEntryBlock()->getName() : "nullptr")
277            << "\n";
278   }
279 #endif
280 
281   /// Determine if a fusion candidate (representing a loop) is eligible for
282   /// fusion. Note that this only checks whether a single loop can be fused - it
283   /// does not check whether it is *legal* to fuse two loops together.
284   bool isEligibleForFusion(ScalarEvolution &SE) const {
285     if (!isValid()) {
286       LLVM_DEBUG(dbgs() << "FC has invalid CFG requirements!\n");
287       if (!Preheader)
288         ++InvalidPreheader;
289       if (!Header)
290         ++InvalidHeader;
291       if (!ExitingBlock)
292         ++InvalidExitingBlock;
293       if (!ExitBlock)
294         ++InvalidExitBlock;
295       if (!Latch)
296         ++InvalidLatch;
297       if (L->isInvalid())
298         ++InvalidLoop;
299 
300       return false;
301     }
302 
303     // Require ScalarEvolution to be able to determine a trip count.
304     if (!SE.hasLoopInvariantBackedgeTakenCount(L)) {
305       LLVM_DEBUG(dbgs() << "Loop " << L->getName()
306                         << " trip count not computable!\n");
307       return reportInvalidCandidate(UnknownTripCount);
308     }
309 
310     if (!L->isLoopSimplifyForm()) {
311       LLVM_DEBUG(dbgs() << "Loop " << L->getName()
312                         << " is not in simplified form!\n");
313       return reportInvalidCandidate(NotSimplifiedForm);
314     }
315 
316     if (!L->isRotatedForm()) {
317       LLVM_DEBUG(dbgs() << "Loop " << L->getName() << " is not rotated!\n");
318       return reportInvalidCandidate(NotRotated);
319     }
320 
321     return true;
322   }
323 
324 private:
325   // This is only used internally for now, to clear the MemWrites and MemReads
326   // list and setting Valid to false. I can't envision other uses of this right
327   // now, since once FusionCandidates are put into the FusionCandidateSet they
328   // are immutable. Thus, any time we need to change/update a FusionCandidate,
329   // we must create a new one and insert it into the FusionCandidateSet to
330   // ensure the FusionCandidateSet remains ordered correctly.
331   void invalidate() {
332     MemWrites.clear();
333     MemReads.clear();
334     Valid = false;
335   }
336 
337   bool reportInvalidCandidate(llvm::Statistic &Stat) const {
338     using namespace ore;
339     assert(L && Preheader && "Fusion candidate not initialized properly!");
340     ++Stat;
341     ORE.emit(OptimizationRemarkAnalysis(DEBUG_TYPE, Stat.getName(),
342                                         L->getStartLoc(), Preheader)
343              << "[" << Preheader->getParent()->getName() << "]: "
344              << "Loop is not a candidate for fusion: " << Stat.getDesc());
345     return false;
346   }
347 };
348 
349 struct FusionCandidateCompare {
350   /// Comparison functor to sort two Control Flow Equivalent fusion candidates
351   /// into dominance order.
352   /// If LHS dominates RHS and RHS post-dominates LHS, return true;
353   /// IF RHS dominates LHS and LHS post-dominates RHS, return false;
354   bool operator()(const FusionCandidate &LHS,
355                   const FusionCandidate &RHS) const {
356     const DominatorTree *DT = LHS.DT;
357 
358     BasicBlock *LHSEntryBlock = LHS.getEntryBlock();
359     BasicBlock *RHSEntryBlock = RHS.getEntryBlock();
360 
361     // Do not save PDT to local variable as it is only used in asserts and thus
362     // will trigger an unused variable warning if building without asserts.
363     assert(DT && LHS.PDT && "Expecting valid dominator tree");
364 
365     // Do this compare first so if LHS == RHS, function returns false.
366     if (DT->dominates(RHSEntryBlock, LHSEntryBlock)) {
367       // RHS dominates LHS
368       // Verify LHS post-dominates RHS
369       assert(LHS.PDT->dominates(LHSEntryBlock, RHSEntryBlock));
370       return false;
371     }
372 
373     if (DT->dominates(LHSEntryBlock, RHSEntryBlock)) {
374       // Verify RHS Postdominates LHS
375       assert(LHS.PDT->dominates(RHSEntryBlock, LHSEntryBlock));
376       return true;
377     }
378 
379     // If LHS does not dominate RHS and RHS does not dominate LHS then there is
380     // no dominance relationship between the two FusionCandidates. Thus, they
381     // should not be in the same set together.
382     llvm_unreachable(
383         "No dominance relationship between these fusion candidates!");
384   }
385 };
386 
387 using LoopVector = SmallVector<Loop *, 4>;
388 
389 // Set of Control Flow Equivalent (CFE) Fusion Candidates, sorted in dominance
390 // order. Thus, if FC0 comes *before* FC1 in a FusionCandidateSet, then FC0
391 // dominates FC1 and FC1 post-dominates FC0.
392 // std::set was chosen because we want a sorted data structure with stable
393 // iterators. A subsequent patch to loop fusion will enable fusing non-ajdacent
394 // loops by moving intervening code around. When this intervening code contains
395 // loops, those loops will be moved also. The corresponding FusionCandidates
396 // will also need to be moved accordingly. As this is done, having stable
397 // iterators will simplify the logic. Similarly, having an efficient insert that
398 // keeps the FusionCandidateSet sorted will also simplify the implementation.
399 using FusionCandidateSet = std::set<FusionCandidate, FusionCandidateCompare>;
400 using FusionCandidateCollection = SmallVector<FusionCandidateSet, 4>;
401 
402 #if !defined(NDEBUG)
403 static llvm::raw_ostream &operator<<(llvm::raw_ostream &OS,
404                                      const FusionCandidate &FC) {
405   if (FC.isValid())
406     OS << FC.Preheader->getName();
407   else
408     OS << "<Invalid>";
409 
410   return OS;
411 }
412 
413 static llvm::raw_ostream &operator<<(llvm::raw_ostream &OS,
414                                      const FusionCandidateSet &CandSet) {
415   for (const FusionCandidate &FC : CandSet)
416     OS << FC << '\n';
417 
418   return OS;
419 }
420 
421 static void
422 printFusionCandidates(const FusionCandidateCollection &FusionCandidates) {
423   dbgs() << "Fusion Candidates: \n";
424   for (const auto &CandidateSet : FusionCandidates) {
425     dbgs() << "*** Fusion Candidate Set ***\n";
426     dbgs() << CandidateSet;
427     dbgs() << "****************************\n";
428   }
429 }
430 #endif
431 
432 /// Collect all loops in function at the same nest level, starting at the
433 /// outermost level.
434 ///
435 /// This data structure collects all loops at the same nest level for a
436 /// given function (specified by the LoopInfo object). It starts at the
437 /// outermost level.
438 struct LoopDepthTree {
439   using LoopsOnLevelTy = SmallVector<LoopVector, 4>;
440   using iterator = LoopsOnLevelTy::iterator;
441   using const_iterator = LoopsOnLevelTy::const_iterator;
442 
443   LoopDepthTree(LoopInfo &LI) : Depth(1) {
444     if (!LI.empty())
445       LoopsOnLevel.emplace_back(LoopVector(LI.rbegin(), LI.rend()));
446   }
447 
448   /// Test whether a given loop has been removed from the function, and thus is
449   /// no longer valid.
450   bool isRemovedLoop(const Loop *L) const { return RemovedLoops.count(L); }
451 
452   /// Record that a given loop has been removed from the function and is no
453   /// longer valid.
454   void removeLoop(const Loop *L) { RemovedLoops.insert(L); }
455 
456   /// Descend the tree to the next (inner) nesting level
457   void descend() {
458     LoopsOnLevelTy LoopsOnNextLevel;
459 
460     for (const LoopVector &LV : *this)
461       for (Loop *L : LV)
462         if (!isRemovedLoop(L) && L->begin() != L->end())
463           LoopsOnNextLevel.emplace_back(LoopVector(L->begin(), L->end()));
464 
465     LoopsOnLevel = LoopsOnNextLevel;
466     RemovedLoops.clear();
467     Depth++;
468   }
469 
470   bool empty() const { return size() == 0; }
471   size_t size() const { return LoopsOnLevel.size() - RemovedLoops.size(); }
472   unsigned getDepth() const { return Depth; }
473 
474   iterator begin() { return LoopsOnLevel.begin(); }
475   iterator end() { return LoopsOnLevel.end(); }
476   const_iterator begin() const { return LoopsOnLevel.begin(); }
477   const_iterator end() const { return LoopsOnLevel.end(); }
478 
479 private:
480   /// Set of loops that have been removed from the function and are no longer
481   /// valid.
482   SmallPtrSet<const Loop *, 8> RemovedLoops;
483 
484   /// Depth of the current level, starting at 1 (outermost loops).
485   unsigned Depth;
486 
487   /// Vector of loops at the current depth level that have the same parent loop
488   LoopsOnLevelTy LoopsOnLevel;
489 };
490 
491 #ifndef NDEBUG
492 static void printLoopVector(const LoopVector &LV) {
493   dbgs() << "****************************\n";
494   for (auto L : LV)
495     printLoop(*L, dbgs());
496   dbgs() << "****************************\n";
497 }
498 #endif
499 
500 struct LoopFuser {
501 private:
502   // Sets of control flow equivalent fusion candidates for a given nest level.
503   FusionCandidateCollection FusionCandidates;
504 
505   LoopDepthTree LDT;
506   DomTreeUpdater DTU;
507 
508   LoopInfo &LI;
509   DominatorTree &DT;
510   DependenceInfo &DI;
511   ScalarEvolution &SE;
512   PostDominatorTree &PDT;
513   OptimizationRemarkEmitter &ORE;
514 
515 public:
516   LoopFuser(LoopInfo &LI, DominatorTree &DT, DependenceInfo &DI,
517             ScalarEvolution &SE, PostDominatorTree &PDT,
518             OptimizationRemarkEmitter &ORE, const DataLayout &DL)
519       : LDT(LI), DTU(DT, PDT, DomTreeUpdater::UpdateStrategy::Lazy), LI(LI),
520         DT(DT), DI(DI), SE(SE), PDT(PDT), ORE(ORE) {}
521 
522   /// This is the main entry point for loop fusion. It will traverse the
523   /// specified function and collect candidate loops to fuse, starting at the
524   /// outermost nesting level and working inwards.
525   bool fuseLoops(Function &F) {
526 #ifndef NDEBUG
527     if (VerboseFusionDebugging) {
528       LI.print(dbgs());
529     }
530 #endif
531 
532     LLVM_DEBUG(dbgs() << "Performing Loop Fusion on function " << F.getName()
533                       << "\n");
534     bool Changed = false;
535 
536     while (!LDT.empty()) {
537       LLVM_DEBUG(dbgs() << "Got " << LDT.size() << " loop sets for depth "
538                         << LDT.getDepth() << "\n";);
539 
540       for (const LoopVector &LV : LDT) {
541         assert(LV.size() > 0 && "Empty loop set was build!");
542 
543         // Skip singleton loop sets as they do not offer fusion opportunities on
544         // this level.
545         if (LV.size() == 1)
546           continue;
547 #ifndef NDEBUG
548         if (VerboseFusionDebugging) {
549           LLVM_DEBUG({
550             dbgs() << "  Visit loop set (#" << LV.size() << "):\n";
551             printLoopVector(LV);
552           });
553         }
554 #endif
555 
556         collectFusionCandidates(LV);
557         Changed |= fuseCandidates();
558       }
559 
560       // Finished analyzing candidates at this level.
561       // Descend to the next level and clear all of the candidates currently
562       // collected. Note that it will not be possible to fuse any of the
563       // existing candidates with new candidates because the new candidates will
564       // be at a different nest level and thus not be control flow equivalent
565       // with all of the candidates collected so far.
566       LLVM_DEBUG(dbgs() << "Descend one level!\n");
567       LDT.descend();
568       FusionCandidates.clear();
569     }
570 
571     if (Changed)
572       LLVM_DEBUG(dbgs() << "Function after Loop Fusion: \n"; F.dump(););
573 
574 #ifndef NDEBUG
575     assert(DT.verify());
576     assert(PDT.verify());
577     LI.verify(DT);
578     SE.verify();
579 #endif
580 
581     LLVM_DEBUG(dbgs() << "Loop Fusion complete\n");
582     return Changed;
583   }
584 
585 private:
586   /// Determine if two fusion candidates are control flow equivalent.
587   ///
588   /// Two fusion candidates are control flow equivalent if when one executes,
589   /// the other is guaranteed to execute. This is determined using dominators
590   /// and post-dominators: if A dominates B and B post-dominates A then A and B
591   /// are control-flow equivalent.
592   bool isControlFlowEquivalent(const FusionCandidate &FC0,
593                                const FusionCandidate &FC1) const {
594     assert(FC0.Preheader && FC1.Preheader && "Expecting valid preheaders");
595 
596     return ::isControlFlowEquivalent(*FC0.getEntryBlock(), *FC1.getEntryBlock(),
597                                      DT, PDT);
598   }
599 
600   /// Iterate over all loops in the given loop set and identify the loops that
601   /// are eligible for fusion. Place all eligible fusion candidates into Control
602   /// Flow Equivalent sets, sorted by dominance.
603   void collectFusionCandidates(const LoopVector &LV) {
604     for (Loop *L : LV) {
605       FusionCandidate CurrCand(L, &DT, &PDT, ORE);
606       if (!CurrCand.isEligibleForFusion(SE))
607         continue;
608 
609       // Go through each list in FusionCandidates and determine if L is control
610       // flow equivalent with the first loop in that list. If it is, append LV.
611       // If not, go to the next list.
612       // If no suitable list is found, start another list and add it to
613       // FusionCandidates.
614       bool FoundSet = false;
615 
616       for (auto &CurrCandSet : FusionCandidates) {
617         if (isControlFlowEquivalent(*CurrCandSet.begin(), CurrCand)) {
618           CurrCandSet.insert(CurrCand);
619           FoundSet = true;
620 #ifndef NDEBUG
621           if (VerboseFusionDebugging)
622             LLVM_DEBUG(dbgs() << "Adding " << CurrCand
623                               << " to existing candidate set\n");
624 #endif
625           break;
626         }
627       }
628       if (!FoundSet) {
629         // No set was found. Create a new set and add to FusionCandidates
630 #ifndef NDEBUG
631         if (VerboseFusionDebugging)
632           LLVM_DEBUG(dbgs() << "Adding " << CurrCand << " to new set\n");
633 #endif
634         FusionCandidateSet NewCandSet;
635         NewCandSet.insert(CurrCand);
636         FusionCandidates.push_back(NewCandSet);
637       }
638       NumFusionCandidates++;
639     }
640   }
641 
642   /// Determine if it is beneficial to fuse two loops.
643   ///
644   /// For now, this method simply returns true because we want to fuse as much
645   /// as possible (primarily to test the pass). This method will evolve, over
646   /// time, to add heuristics for profitability of fusion.
647   bool isBeneficialFusion(const FusionCandidate &FC0,
648                           const FusionCandidate &FC1) {
649     return true;
650   }
651 
652   /// Determine if two fusion candidates have the same trip count (i.e., they
653   /// execute the same number of iterations).
654   ///
655   /// Note that for now this method simply returns a boolean value because there
656   /// are no mechanisms in loop fusion to handle different trip counts. In the
657   /// future, this behaviour can be extended to adjust one of the loops to make
658   /// the trip counts equal (e.g., loop peeling). When this is added, this
659   /// interface may need to change to return more information than just a
660   /// boolean value.
661   bool identicalTripCounts(const FusionCandidate &FC0,
662                            const FusionCandidate &FC1) const {
663     const SCEV *TripCount0 = SE.getBackedgeTakenCount(FC0.L);
664     if (isa<SCEVCouldNotCompute>(TripCount0)) {
665       UncomputableTripCount++;
666       LLVM_DEBUG(dbgs() << "Trip count of first loop could not be computed!");
667       return false;
668     }
669 
670     const SCEV *TripCount1 = SE.getBackedgeTakenCount(FC1.L);
671     if (isa<SCEVCouldNotCompute>(TripCount1)) {
672       UncomputableTripCount++;
673       LLVM_DEBUG(dbgs() << "Trip count of second loop could not be computed!");
674       return false;
675     }
676     LLVM_DEBUG(dbgs() << "\tTrip counts: " << *TripCount0 << " & "
677                       << *TripCount1 << " are "
678                       << (TripCount0 == TripCount1 ? "identical" : "different")
679                       << "\n");
680 
681     return (TripCount0 == TripCount1);
682   }
683 
684   /// Walk each set of control flow equivalent fusion candidates and attempt to
685   /// fuse them. This does a single linear traversal of all candidates in the
686   /// set. The conditions for legal fusion are checked at this point. If a pair
687   /// of fusion candidates passes all legality checks, they are fused together
688   /// and a new fusion candidate is created and added to the FusionCandidateSet.
689   /// The original fusion candidates are then removed, as they are no longer
690   /// valid.
691   bool fuseCandidates() {
692     bool Fused = false;
693     LLVM_DEBUG(printFusionCandidates(FusionCandidates));
694     for (auto &CandidateSet : FusionCandidates) {
695       if (CandidateSet.size() < 2)
696         continue;
697 
698       LLVM_DEBUG(dbgs() << "Attempting fusion on Candidate Set:\n"
699                         << CandidateSet << "\n");
700 
701       for (auto FC0 = CandidateSet.begin(); FC0 != CandidateSet.end(); ++FC0) {
702         assert(!LDT.isRemovedLoop(FC0->L) &&
703                "Should not have removed loops in CandidateSet!");
704         auto FC1 = FC0;
705         for (++FC1; FC1 != CandidateSet.end(); ++FC1) {
706           assert(!LDT.isRemovedLoop(FC1->L) &&
707                  "Should not have removed loops in CandidateSet!");
708 
709           LLVM_DEBUG(dbgs() << "Attempting to fuse candidate \n"; FC0->dump();
710                      dbgs() << " with\n"; FC1->dump(); dbgs() << "\n");
711 
712           FC0->verify();
713           FC1->verify();
714 
715           if (!identicalTripCounts(*FC0, *FC1)) {
716             LLVM_DEBUG(dbgs() << "Fusion candidates do not have identical trip "
717                                  "counts. Not fusing.\n");
718             reportLoopFusion<OptimizationRemarkMissed>(*FC0, *FC1,
719                                                        NonEqualTripCount);
720             continue;
721           }
722 
723           if (!isAdjacent(*FC0, *FC1)) {
724             LLVM_DEBUG(dbgs()
725                        << "Fusion candidates are not adjacent. Not fusing.\n");
726             reportLoopFusion<OptimizationRemarkMissed>(*FC0, *FC1, NonAdjacent);
727             continue;
728           }
729 
730           // Ensure that FC0 and FC1 have identical guards.
731           // If one (or both) are not guarded, this check is not necessary.
732           if (FC0->GuardBranch && FC1->GuardBranch &&
733               !haveIdenticalGuards(*FC0, *FC1)) {
734             LLVM_DEBUG(dbgs() << "Fusion candidates do not have identical "
735                                  "guards. Not Fusing.\n");
736             reportLoopFusion<OptimizationRemarkMissed>(*FC0, *FC1,
737                                                        NonIdenticalGuards);
738             continue;
739           }
740 
741           // The following three checks look for empty blocks in FC0 and FC1. If
742           // any of these blocks are non-empty, we do not fuse. This is done
743           // because we currently do not have the safety checks to determine if
744           // it is safe to move the blocks past other blocks in the loop. Once
745           // these checks are added, these conditions can be relaxed.
746           if (!isEmptyPreheader(*FC1)) {
747             LLVM_DEBUG(dbgs() << "Fusion candidate does not have empty "
748                                  "preheader. Not fusing.\n");
749             reportLoopFusion<OptimizationRemarkMissed>(*FC0, *FC1,
750                                                        NonEmptyPreheader);
751             continue;
752           }
753 
754           if (FC0->GuardBranch && !isEmptyExitBlock(*FC0)) {
755             LLVM_DEBUG(dbgs() << "Fusion candidate does not have empty exit "
756                                  "block. Not fusing.\n");
757             reportLoopFusion<OptimizationRemarkMissed>(*FC0, *FC1,
758                                                        NonEmptyExitBlock);
759             continue;
760           }
761 
762           if (FC1->GuardBranch && !isEmptyGuardBlock(*FC1)) {
763             LLVM_DEBUG(dbgs() << "Fusion candidate does not have empty guard "
764                                  "block. Not fusing.\n");
765             reportLoopFusion<OptimizationRemarkMissed>(*FC0, *FC1,
766                                                        NonEmptyGuardBlock);
767             continue;
768           }
769 
770           // Check the dependencies across the loops and do not fuse if it would
771           // violate them.
772           if (!dependencesAllowFusion(*FC0, *FC1)) {
773             LLVM_DEBUG(dbgs() << "Memory dependencies do not allow fusion!\n");
774             reportLoopFusion<OptimizationRemarkMissed>(*FC0, *FC1,
775                                                        InvalidDependencies);
776             continue;
777           }
778 
779           bool BeneficialToFuse = isBeneficialFusion(*FC0, *FC1);
780           LLVM_DEBUG(dbgs()
781                      << "\tFusion appears to be "
782                      << (BeneficialToFuse ? "" : "un") << "profitable!\n");
783           if (!BeneficialToFuse) {
784             reportLoopFusion<OptimizationRemarkMissed>(*FC0, *FC1,
785                                                        FusionNotBeneficial);
786             continue;
787           }
788           // All analysis has completed and has determined that fusion is legal
789           // and profitable. At this point, start transforming the code and
790           // perform fusion.
791 
792           LLVM_DEBUG(dbgs() << "\tFusion is performed: " << *FC0 << " and "
793                             << *FC1 << "\n");
794 
795           // Report fusion to the Optimization Remarks.
796           // Note this needs to be done *before* performFusion because
797           // performFusion will change the original loops, making it not
798           // possible to identify them after fusion is complete.
799           reportLoopFusion<OptimizationRemark>(*FC0, *FC1, FuseCounter);
800 
801           FusionCandidate FusedCand(performFusion(*FC0, *FC1), &DT, &PDT, ORE);
802           FusedCand.verify();
803           assert(FusedCand.isEligibleForFusion(SE) &&
804                  "Fused candidate should be eligible for fusion!");
805 
806           // Notify the loop-depth-tree that these loops are not valid objects
807           LDT.removeLoop(FC1->L);
808 
809           CandidateSet.erase(FC0);
810           CandidateSet.erase(FC1);
811 
812           auto InsertPos = CandidateSet.insert(FusedCand);
813 
814           assert(InsertPos.second &&
815                  "Unable to insert TargetCandidate in CandidateSet!");
816 
817           // Reset FC0 and FC1 the new (fused) candidate. Subsequent iterations
818           // of the FC1 loop will attempt to fuse the new (fused) loop with the
819           // remaining candidates in the current candidate set.
820           FC0 = FC1 = InsertPos.first;
821 
822           LLVM_DEBUG(dbgs() << "Candidate Set (after fusion): " << CandidateSet
823                             << "\n");
824 
825           Fused = true;
826         }
827       }
828     }
829     return Fused;
830   }
831 
832   /// Rewrite all additive recurrences in a SCEV to use a new loop.
833   class AddRecLoopReplacer : public SCEVRewriteVisitor<AddRecLoopReplacer> {
834   public:
835     AddRecLoopReplacer(ScalarEvolution &SE, const Loop &OldL, const Loop &NewL,
836                        bool UseMax = true)
837         : SCEVRewriteVisitor(SE), Valid(true), UseMax(UseMax), OldL(OldL),
838           NewL(NewL) {}
839 
840     const SCEV *visitAddRecExpr(const SCEVAddRecExpr *Expr) {
841       const Loop *ExprL = Expr->getLoop();
842       SmallVector<const SCEV *, 2> Operands;
843       if (ExprL == &OldL) {
844         Operands.append(Expr->op_begin(), Expr->op_end());
845         return SE.getAddRecExpr(Operands, &NewL, Expr->getNoWrapFlags());
846       }
847 
848       if (OldL.contains(ExprL)) {
849         bool Pos = SE.isKnownPositive(Expr->getStepRecurrence(SE));
850         if (!UseMax || !Pos || !Expr->isAffine()) {
851           Valid = false;
852           return Expr;
853         }
854         return visit(Expr->getStart());
855       }
856 
857       for (const SCEV *Op : Expr->operands())
858         Operands.push_back(visit(Op));
859       return SE.getAddRecExpr(Operands, ExprL, Expr->getNoWrapFlags());
860     }
861 
862     bool wasValidSCEV() const { return Valid; }
863 
864   private:
865     bool Valid, UseMax;
866     const Loop &OldL, &NewL;
867   };
868 
869   /// Return false if the access functions of \p I0 and \p I1 could cause
870   /// a negative dependence.
871   bool accessDiffIsPositive(const Loop &L0, const Loop &L1, Instruction &I0,
872                             Instruction &I1, bool EqualIsInvalid) {
873     Value *Ptr0 = getLoadStorePointerOperand(&I0);
874     Value *Ptr1 = getLoadStorePointerOperand(&I1);
875     if (!Ptr0 || !Ptr1)
876       return false;
877 
878     const SCEV *SCEVPtr0 = SE.getSCEVAtScope(Ptr0, &L0);
879     const SCEV *SCEVPtr1 = SE.getSCEVAtScope(Ptr1, &L1);
880 #ifndef NDEBUG
881     if (VerboseFusionDebugging)
882       LLVM_DEBUG(dbgs() << "    Access function check: " << *SCEVPtr0 << " vs "
883                         << *SCEVPtr1 << "\n");
884 #endif
885     AddRecLoopReplacer Rewriter(SE, L0, L1);
886     SCEVPtr0 = Rewriter.visit(SCEVPtr0);
887 #ifndef NDEBUG
888     if (VerboseFusionDebugging)
889       LLVM_DEBUG(dbgs() << "    Access function after rewrite: " << *SCEVPtr0
890                         << " [Valid: " << Rewriter.wasValidSCEV() << "]\n");
891 #endif
892     if (!Rewriter.wasValidSCEV())
893       return false;
894 
895     // TODO: isKnownPredicate doesnt work well when one SCEV is loop carried (by
896     //       L0) and the other is not. We could check if it is monotone and test
897     //       the beginning and end value instead.
898 
899     BasicBlock *L0Header = L0.getHeader();
900     auto HasNonLinearDominanceRelation = [&](const SCEV *S) {
901       const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(S);
902       if (!AddRec)
903         return false;
904       return !DT.dominates(L0Header, AddRec->getLoop()->getHeader()) &&
905              !DT.dominates(AddRec->getLoop()->getHeader(), L0Header);
906     };
907     if (SCEVExprContains(SCEVPtr1, HasNonLinearDominanceRelation))
908       return false;
909 
910     ICmpInst::Predicate Pred =
911         EqualIsInvalid ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_SGE;
912     bool IsAlwaysGE = SE.isKnownPredicate(Pred, SCEVPtr0, SCEVPtr1);
913 #ifndef NDEBUG
914     if (VerboseFusionDebugging)
915       LLVM_DEBUG(dbgs() << "    Relation: " << *SCEVPtr0
916                         << (IsAlwaysGE ? "  >=  " : "  may <  ") << *SCEVPtr1
917                         << "\n");
918 #endif
919     return IsAlwaysGE;
920   }
921 
922   /// Return true if the dependences between @p I0 (in @p L0) and @p I1 (in
923   /// @p L1) allow loop fusion of @p L0 and @p L1. The dependence analyses
924   /// specified by @p DepChoice are used to determine this.
925   bool dependencesAllowFusion(const FusionCandidate &FC0,
926                               const FusionCandidate &FC1, Instruction &I0,
927                               Instruction &I1, bool AnyDep,
928                               FusionDependenceAnalysisChoice DepChoice) {
929 #ifndef NDEBUG
930     if (VerboseFusionDebugging) {
931       LLVM_DEBUG(dbgs() << "Check dep: " << I0 << " vs " << I1 << " : "
932                         << DepChoice << "\n");
933     }
934 #endif
935     switch (DepChoice) {
936     case FUSION_DEPENDENCE_ANALYSIS_SCEV:
937       return accessDiffIsPositive(*FC0.L, *FC1.L, I0, I1, AnyDep);
938     case FUSION_DEPENDENCE_ANALYSIS_DA: {
939       auto DepResult = DI.depends(&I0, &I1, true);
940       if (!DepResult)
941         return true;
942 #ifndef NDEBUG
943       if (VerboseFusionDebugging) {
944         LLVM_DEBUG(dbgs() << "DA res: "; DepResult->dump(dbgs());
945                    dbgs() << " [#l: " << DepResult->getLevels() << "][Ordered: "
946                           << (DepResult->isOrdered() ? "true" : "false")
947                           << "]\n");
948         LLVM_DEBUG(dbgs() << "DepResult Levels: " << DepResult->getLevels()
949                           << "\n");
950       }
951 #endif
952 
953       if (DepResult->getNextPredecessor() || DepResult->getNextSuccessor())
954         LLVM_DEBUG(
955             dbgs() << "TODO: Implement pred/succ dependence handling!\n");
956 
957       // TODO: Can we actually use the dependence info analysis here?
958       return false;
959     }
960 
961     case FUSION_DEPENDENCE_ANALYSIS_ALL:
962       return dependencesAllowFusion(FC0, FC1, I0, I1, AnyDep,
963                                     FUSION_DEPENDENCE_ANALYSIS_SCEV) ||
964              dependencesAllowFusion(FC0, FC1, I0, I1, AnyDep,
965                                     FUSION_DEPENDENCE_ANALYSIS_DA);
966     }
967 
968     llvm_unreachable("Unknown fusion dependence analysis choice!");
969   }
970 
971   /// Perform a dependence check and return if @p FC0 and @p FC1 can be fused.
972   bool dependencesAllowFusion(const FusionCandidate &FC0,
973                               const FusionCandidate &FC1) {
974     LLVM_DEBUG(dbgs() << "Check if " << FC0 << " can be fused with " << FC1
975                       << "\n");
976     assert(FC0.L->getLoopDepth() == FC1.L->getLoopDepth());
977     assert(DT.dominates(FC0.getEntryBlock(), FC1.getEntryBlock()));
978 
979     for (Instruction *WriteL0 : FC0.MemWrites) {
980       for (Instruction *WriteL1 : FC1.MemWrites)
981         if (!dependencesAllowFusion(FC0, FC1, *WriteL0, *WriteL1,
982                                     /* AnyDep */ false,
983                                     FusionDependenceAnalysis)) {
984           InvalidDependencies++;
985           return false;
986         }
987       for (Instruction *ReadL1 : FC1.MemReads)
988         if (!dependencesAllowFusion(FC0, FC1, *WriteL0, *ReadL1,
989                                     /* AnyDep */ false,
990                                     FusionDependenceAnalysis)) {
991           InvalidDependencies++;
992           return false;
993         }
994     }
995 
996     for (Instruction *WriteL1 : FC1.MemWrites) {
997       for (Instruction *WriteL0 : FC0.MemWrites)
998         if (!dependencesAllowFusion(FC0, FC1, *WriteL0, *WriteL1,
999                                     /* AnyDep */ false,
1000                                     FusionDependenceAnalysis)) {
1001           InvalidDependencies++;
1002           return false;
1003         }
1004       for (Instruction *ReadL0 : FC0.MemReads)
1005         if (!dependencesAllowFusion(FC0, FC1, *ReadL0, *WriteL1,
1006                                     /* AnyDep */ false,
1007                                     FusionDependenceAnalysis)) {
1008           InvalidDependencies++;
1009           return false;
1010         }
1011     }
1012 
1013     // Walk through all uses in FC1. For each use, find the reaching def. If the
1014     // def is located in FC0 then it is is not safe to fuse.
1015     for (BasicBlock *BB : FC1.L->blocks())
1016       for (Instruction &I : *BB)
1017         for (auto &Op : I.operands())
1018           if (Instruction *Def = dyn_cast<Instruction>(Op))
1019             if (FC0.L->contains(Def->getParent())) {
1020               InvalidDependencies++;
1021               return false;
1022             }
1023 
1024     return true;
1025   }
1026 
1027   /// Determine if two fusion candidates are adjacent in the CFG.
1028   ///
1029   /// This method will determine if there are additional basic blocks in the CFG
1030   /// between the exit of \p FC0 and the entry of \p FC1.
1031   /// If the two candidates are guarded loops, then it checks whether the
1032   /// non-loop successor of the \p FC0 guard branch is the entry block of \p
1033   /// FC1. If not, then the loops are not adjacent. If the two candidates are
1034   /// not guarded loops, then it checks whether the exit block of \p FC0 is the
1035   /// preheader of \p FC1.
1036   bool isAdjacent(const FusionCandidate &FC0,
1037                   const FusionCandidate &FC1) const {
1038     // If the successor of the guard branch is FC1, then the loops are adjacent
1039     if (FC0.GuardBranch)
1040       return FC0.getNonLoopBlock() == FC1.getEntryBlock();
1041     else
1042       return FC0.ExitBlock == FC1.getEntryBlock();
1043   }
1044 
1045   /// Determine if two fusion candidates have identical guards
1046   ///
1047   /// This method will determine if two fusion candidates have the same guards.
1048   /// The guards are considered the same if:
1049   ///   1. The instructions to compute the condition used in the compare are
1050   ///      identical.
1051   ///   2. The successors of the guard have the same flow into/around the loop.
1052   /// If the compare instructions are identical, then the first successor of the
1053   /// guard must go to the same place (either the preheader of the loop or the
1054   /// NonLoopBlock). In other words, the the first successor of both loops must
1055   /// both go into the loop (i.e., the preheader) or go around the loop (i.e.,
1056   /// the NonLoopBlock). The same must be true for the second successor.
1057   bool haveIdenticalGuards(const FusionCandidate &FC0,
1058                            const FusionCandidate &FC1) const {
1059     assert(FC0.GuardBranch && FC1.GuardBranch &&
1060            "Expecting FC0 and FC1 to be guarded loops.");
1061 
1062     if (auto FC0CmpInst =
1063             dyn_cast<Instruction>(FC0.GuardBranch->getCondition()))
1064       if (auto FC1CmpInst =
1065               dyn_cast<Instruction>(FC1.GuardBranch->getCondition()))
1066         if (!FC0CmpInst->isIdenticalTo(FC1CmpInst))
1067           return false;
1068 
1069     // The compare instructions are identical.
1070     // Now make sure the successor of the guards have the same flow into/around
1071     // the loop
1072     if (FC0.GuardBranch->getSuccessor(0) == FC0.Preheader)
1073       return (FC1.GuardBranch->getSuccessor(0) == FC1.Preheader);
1074     else
1075       return (FC1.GuardBranch->getSuccessor(1) == FC1.Preheader);
1076   }
1077 
1078   /// Check that the guard for \p FC *only* contains the cmp/branch for the
1079   /// guard.
1080   /// Once we are able to handle intervening code, any code in the guard block
1081   /// for FC1 will need to be treated as intervening code and checked whether
1082   /// it can safely move around the loops.
1083   bool isEmptyGuardBlock(const FusionCandidate &FC) const {
1084     assert(FC.GuardBranch && "Expecting a fusion candidate with guard branch.");
1085     if (auto *CmpInst = dyn_cast<Instruction>(FC.GuardBranch->getCondition())) {
1086       auto *GuardBlock = FC.GuardBranch->getParent();
1087       // If the generation of the cmp value is in GuardBlock, then the size of
1088       // the guard block should be 2 (cmp + branch). If the generation of the
1089       // cmp value is in a different block, then the size of the guard block
1090       // should only be 1.
1091       if (CmpInst->getParent() == GuardBlock)
1092         return GuardBlock->size() == 2;
1093       else
1094         return GuardBlock->size() == 1;
1095     }
1096 
1097     return false;
1098   }
1099 
1100   bool isEmptyPreheader(const FusionCandidate &FC) const {
1101     assert(FC.Preheader && "Expecting a valid preheader");
1102     return FC.Preheader->size() == 1;
1103   }
1104 
1105   bool isEmptyExitBlock(const FusionCandidate &FC) const {
1106     assert(FC.ExitBlock && "Expecting a valid exit block");
1107     return FC.ExitBlock->size() == 1;
1108   }
1109 
1110   /// Simplify the condition of the latch branch of \p FC to true, when both of
1111   /// its successors are the same.
1112   void simplifyLatchBranch(const FusionCandidate &FC) const {
1113     BranchInst *FCLatchBranch = dyn_cast<BranchInst>(FC.Latch->getTerminator());
1114     if (FCLatchBranch) {
1115       assert(FCLatchBranch->isConditional() &&
1116              FCLatchBranch->getSuccessor(0) == FCLatchBranch->getSuccessor(1) &&
1117              "Expecting the two successors of FCLatchBranch to be the same");
1118       FCLatchBranch->setCondition(
1119           llvm::ConstantInt::getTrue(FCLatchBranch->getCondition()->getType()));
1120     }
1121   }
1122 
1123   /// Move instructions from FC0.Latch to FC1.Latch. If FC0.Latch has an unique
1124   /// successor, then merge FC0.Latch with its unique successor.
1125   void mergeLatch(const FusionCandidate &FC0, const FusionCandidate &FC1) {
1126     moveInstsBottomUp(*FC0.Latch, *FC1.Latch, DT, PDT, DI);
1127     if (BasicBlock *Succ = FC0.Latch->getUniqueSuccessor()) {
1128       MergeBlockIntoPredecessor(Succ, &DTU, &LI);
1129       DTU.flush();
1130     }
1131   }
1132 
1133   /// Fuse two fusion candidates, creating a new fused loop.
1134   ///
1135   /// This method contains the mechanics of fusing two loops, represented by \p
1136   /// FC0 and \p FC1. It is assumed that \p FC0 dominates \p FC1 and \p FC1
1137   /// postdominates \p FC0 (making them control flow equivalent). It also
1138   /// assumes that the other conditions for fusion have been met: adjacent,
1139   /// identical trip counts, and no negative distance dependencies exist that
1140   /// would prevent fusion. Thus, there is no checking for these conditions in
1141   /// this method.
1142   ///
1143   /// Fusion is performed by rewiring the CFG to update successor blocks of the
1144   /// components of tho loop. Specifically, the following changes are done:
1145   ///
1146   ///   1. The preheader of \p FC1 is removed as it is no longer necessary
1147   ///   (because it is currently only a single statement block).
1148   ///   2. The latch of \p FC0 is modified to jump to the header of \p FC1.
1149   ///   3. The latch of \p FC1 i modified to jump to the header of \p FC0.
1150   ///   4. All blocks from \p FC1 are removed from FC1 and added to FC0.
1151   ///
1152   /// All of these modifications are done with dominator tree updates, thus
1153   /// keeping the dominator (and post dominator) information up-to-date.
1154   ///
1155   /// This can be improved in the future by actually merging blocks during
1156   /// fusion. For example, the preheader of \p FC1 can be merged with the
1157   /// preheader of \p FC0. This would allow loops with more than a single
1158   /// statement in the preheader to be fused. Similarly, the latch blocks of the
1159   /// two loops could also be fused into a single block. This will require
1160   /// analysis to prove it is safe to move the contents of the block past
1161   /// existing code, which currently has not been implemented.
1162   Loop *performFusion(const FusionCandidate &FC0, const FusionCandidate &FC1) {
1163     assert(FC0.isValid() && FC1.isValid() &&
1164            "Expecting valid fusion candidates");
1165 
1166     LLVM_DEBUG(dbgs() << "Fusion Candidate 0: \n"; FC0.dump();
1167                dbgs() << "Fusion Candidate 1: \n"; FC1.dump(););
1168 
1169     // Fusing guarded loops is handled slightly differently than non-guarded
1170     // loops and has been broken out into a separate method instead of trying to
1171     // intersperse the logic within a single method.
1172     if (FC0.GuardBranch)
1173       return fuseGuardedLoops(FC0, FC1);
1174 
1175     assert(FC1.Preheader == FC0.ExitBlock);
1176     assert(FC1.Preheader->size() == 1 &&
1177            FC1.Preheader->getSingleSuccessor() == FC1.Header);
1178 
1179     // Remember the phi nodes originally in the header of FC0 in order to rewire
1180     // them later. However, this is only necessary if the new loop carried
1181     // values might not dominate the exiting branch. While we do not generally
1182     // test if this is the case but simply insert intermediate phi nodes, we
1183     // need to make sure these intermediate phi nodes have different
1184     // predecessors. To this end, we filter the special case where the exiting
1185     // block is the latch block of the first loop. Nothing needs to be done
1186     // anyway as all loop carried values dominate the latch and thereby also the
1187     // exiting branch.
1188     SmallVector<PHINode *, 8> OriginalFC0PHIs;
1189     if (FC0.ExitingBlock != FC0.Latch)
1190       for (PHINode &PHI : FC0.Header->phis())
1191         OriginalFC0PHIs.push_back(&PHI);
1192 
1193     // Replace incoming blocks for header PHIs first.
1194     FC1.Preheader->replaceSuccessorsPhiUsesWith(FC0.Preheader);
1195     FC0.Latch->replaceSuccessorsPhiUsesWith(FC1.Latch);
1196 
1197     // Then modify the control flow and update DT and PDT.
1198     SmallVector<DominatorTree::UpdateType, 8> TreeUpdates;
1199 
1200     // The old exiting block of the first loop (FC0) has to jump to the header
1201     // of the second as we need to execute the code in the second header block
1202     // regardless of the trip count. That is, if the trip count is 0, so the
1203     // back edge is never taken, we still have to execute both loop headers,
1204     // especially (but not only!) if the second is a do-while style loop.
1205     // However, doing so might invalidate the phi nodes of the first loop as
1206     // the new values do only need to dominate their latch and not the exiting
1207     // predicate. To remedy this potential problem we always introduce phi
1208     // nodes in the header of the second loop later that select the loop carried
1209     // value, if the second header was reached through an old latch of the
1210     // first, or undef otherwise. This is sound as exiting the first implies the
1211     // second will exit too, __without__ taking the back-edge. [Their
1212     // trip-counts are equal after all.
1213     // KB: Would this sequence be simpler to just just make FC0.ExitingBlock go
1214     // to FC1.Header? I think this is basically what the three sequences are
1215     // trying to accomplish; however, doing this directly in the CFG may mean
1216     // the DT/PDT becomes invalid
1217     FC0.ExitingBlock->getTerminator()->replaceUsesOfWith(FC1.Preheader,
1218                                                          FC1.Header);
1219     TreeUpdates.emplace_back(DominatorTree::UpdateType(
1220         DominatorTree::Delete, FC0.ExitingBlock, FC1.Preheader));
1221     TreeUpdates.emplace_back(DominatorTree::UpdateType(
1222         DominatorTree::Insert, FC0.ExitingBlock, FC1.Header));
1223 
1224     // The pre-header of L1 is not necessary anymore.
1225     assert(pred_begin(FC1.Preheader) == pred_end(FC1.Preheader));
1226     FC1.Preheader->getTerminator()->eraseFromParent();
1227     new UnreachableInst(FC1.Preheader->getContext(), FC1.Preheader);
1228     TreeUpdates.emplace_back(DominatorTree::UpdateType(
1229         DominatorTree::Delete, FC1.Preheader, FC1.Header));
1230 
1231     // Moves the phi nodes from the second to the first loops header block.
1232     while (PHINode *PHI = dyn_cast<PHINode>(&FC1.Header->front())) {
1233       if (SE.isSCEVable(PHI->getType()))
1234         SE.forgetValue(PHI);
1235       if (PHI->hasNUsesOrMore(1))
1236         PHI->moveBefore(&*FC0.Header->getFirstInsertionPt());
1237       else
1238         PHI->eraseFromParent();
1239     }
1240 
1241     // Introduce new phi nodes in the second loop header to ensure
1242     // exiting the first and jumping to the header of the second does not break
1243     // the SSA property of the phis originally in the first loop. See also the
1244     // comment above.
1245     Instruction *L1HeaderIP = &FC1.Header->front();
1246     for (PHINode *LCPHI : OriginalFC0PHIs) {
1247       int L1LatchBBIdx = LCPHI->getBasicBlockIndex(FC1.Latch);
1248       assert(L1LatchBBIdx >= 0 &&
1249              "Expected loop carried value to be rewired at this point!");
1250 
1251       Value *LCV = LCPHI->getIncomingValue(L1LatchBBIdx);
1252 
1253       PHINode *L1HeaderPHI = PHINode::Create(
1254           LCV->getType(), 2, LCPHI->getName() + ".afterFC0", L1HeaderIP);
1255       L1HeaderPHI->addIncoming(LCV, FC0.Latch);
1256       L1HeaderPHI->addIncoming(UndefValue::get(LCV->getType()),
1257                                FC0.ExitingBlock);
1258 
1259       LCPHI->setIncomingValue(L1LatchBBIdx, L1HeaderPHI);
1260     }
1261 
1262     // Replace latch terminator destinations.
1263     FC0.Latch->getTerminator()->replaceUsesOfWith(FC0.Header, FC1.Header);
1264     FC1.Latch->getTerminator()->replaceUsesOfWith(FC1.Header, FC0.Header);
1265 
1266     // Change the condition of FC0 latch branch to true, as both successors of
1267     // the branch are the same.
1268     simplifyLatchBranch(FC0);
1269 
1270     // If FC0.Latch and FC0.ExitingBlock are the same then we have already
1271     // performed the updates above.
1272     if (FC0.Latch != FC0.ExitingBlock)
1273       TreeUpdates.emplace_back(DominatorTree::UpdateType(
1274           DominatorTree::Insert, FC0.Latch, FC1.Header));
1275 
1276     TreeUpdates.emplace_back(DominatorTree::UpdateType(DominatorTree::Delete,
1277                                                        FC0.Latch, FC0.Header));
1278     TreeUpdates.emplace_back(DominatorTree::UpdateType(DominatorTree::Insert,
1279                                                        FC1.Latch, FC0.Header));
1280     TreeUpdates.emplace_back(DominatorTree::UpdateType(DominatorTree::Delete,
1281                                                        FC1.Latch, FC1.Header));
1282 
1283     // Update DT/PDT
1284     DTU.applyUpdates(TreeUpdates);
1285 
1286     LI.removeBlock(FC1.Preheader);
1287     DTU.deleteBB(FC1.Preheader);
1288     DTU.flush();
1289 
1290     // Is there a way to keep SE up-to-date so we don't need to forget the loops
1291     // and rebuild the information in subsequent passes of fusion?
1292     // Note: Need to forget the loops before merging the loop latches, as
1293     // mergeLatch may remove the only block in FC1.
1294     SE.forgetLoop(FC1.L);
1295     SE.forgetLoop(FC0.L);
1296 
1297     // Move instructions from FC0.Latch to FC1.Latch.
1298     // Note: mergeLatch requires an updated DT.
1299     mergeLatch(FC0, FC1);
1300 
1301     // Merge the loops.
1302     SmallVector<BasicBlock *, 8> Blocks(FC1.L->block_begin(),
1303                                         FC1.L->block_end());
1304     for (BasicBlock *BB : Blocks) {
1305       FC0.L->addBlockEntry(BB);
1306       FC1.L->removeBlockFromLoop(BB);
1307       if (LI.getLoopFor(BB) != FC1.L)
1308         continue;
1309       LI.changeLoopFor(BB, FC0.L);
1310     }
1311     while (!FC1.L->empty()) {
1312       const auto &ChildLoopIt = FC1.L->begin();
1313       Loop *ChildLoop = *ChildLoopIt;
1314       FC1.L->removeChildLoop(ChildLoopIt);
1315       FC0.L->addChildLoop(ChildLoop);
1316     }
1317 
1318     // Delete the now empty loop L1.
1319     LI.erase(FC1.L);
1320 
1321 #ifndef NDEBUG
1322     assert(!verifyFunction(*FC0.Header->getParent(), &errs()));
1323     assert(DT.verify(DominatorTree::VerificationLevel::Fast));
1324     assert(PDT.verify());
1325     LI.verify(DT);
1326     SE.verify();
1327 #endif
1328 
1329     LLVM_DEBUG(dbgs() << "Fusion done:\n");
1330 
1331     return FC0.L;
1332   }
1333 
1334   /// Report details on loop fusion opportunities.
1335   ///
1336   /// This template function can be used to report both successful and missed
1337   /// loop fusion opportunities, based on the RemarkKind. The RemarkKind should
1338   /// be one of:
1339   ///   - OptimizationRemarkMissed to report when loop fusion is unsuccessful
1340   ///     given two valid fusion candidates.
1341   ///   - OptimizationRemark to report successful fusion of two fusion
1342   ///     candidates.
1343   /// The remarks will be printed using the form:
1344   ///    <path/filename>:<line number>:<column number>: [<function name>]:
1345   ///       <Cand1 Preheader> and <Cand2 Preheader>: <Stat Description>
1346   template <typename RemarkKind>
1347   void reportLoopFusion(const FusionCandidate &FC0, const FusionCandidate &FC1,
1348                         llvm::Statistic &Stat) {
1349     assert(FC0.Preheader && FC1.Preheader &&
1350            "Expecting valid fusion candidates");
1351     using namespace ore;
1352     ++Stat;
1353     ORE.emit(RemarkKind(DEBUG_TYPE, Stat.getName(), FC0.L->getStartLoc(),
1354                         FC0.Preheader)
1355              << "[" << FC0.Preheader->getParent()->getName()
1356              << "]: " << NV("Cand1", StringRef(FC0.Preheader->getName()))
1357              << " and " << NV("Cand2", StringRef(FC1.Preheader->getName()))
1358              << ": " << Stat.getDesc());
1359   }
1360 
1361   /// Fuse two guarded fusion candidates, creating a new fused loop.
1362   ///
1363   /// Fusing guarded loops is handled much the same way as fusing non-guarded
1364   /// loops. The rewiring of the CFG is slightly different though, because of
1365   /// the presence of the guards around the loops and the exit blocks after the
1366   /// loop body. As such, the new loop is rewired as follows:
1367   ///    1. Keep the guard branch from FC0 and use the non-loop block target
1368   /// from the FC1 guard branch.
1369   ///    2. Remove the exit block from FC0 (this exit block should be empty
1370   /// right now).
1371   ///    3. Remove the guard branch for FC1
1372   ///    4. Remove the preheader for FC1.
1373   /// The exit block successor for the latch of FC0 is updated to be the header
1374   /// of FC1 and the non-exit block successor of the latch of FC1 is updated to
1375   /// be the header of FC0, thus creating the fused loop.
1376   Loop *fuseGuardedLoops(const FusionCandidate &FC0,
1377                          const FusionCandidate &FC1) {
1378     assert(FC0.GuardBranch && FC1.GuardBranch && "Expecting guarded loops");
1379 
1380     BasicBlock *FC0GuardBlock = FC0.GuardBranch->getParent();
1381     BasicBlock *FC1GuardBlock = FC1.GuardBranch->getParent();
1382     BasicBlock *FC0NonLoopBlock = FC0.getNonLoopBlock();
1383     BasicBlock *FC1NonLoopBlock = FC1.getNonLoopBlock();
1384 
1385     assert(FC0NonLoopBlock == FC1GuardBlock && "Loops are not adjacent");
1386 
1387     SmallVector<DominatorTree::UpdateType, 8> TreeUpdates;
1388 
1389     ////////////////////////////////////////////////////////////////////////////
1390     // Update the Loop Guard
1391     ////////////////////////////////////////////////////////////////////////////
1392     // The guard for FC0 is updated to guard both FC0 and FC1. This is done by
1393     // changing the NonLoopGuardBlock for FC0 to the NonLoopGuardBlock for FC1.
1394     // Thus, one path from the guard goes to the preheader for FC0 (and thus
1395     // executes the new fused loop) and the other path goes to the NonLoopBlock
1396     // for FC1 (where FC1 guard would have gone if FC1 was not executed).
1397     FC0.GuardBranch->replaceUsesOfWith(FC0NonLoopBlock, FC1NonLoopBlock);
1398     FC0.ExitBlock->getTerminator()->replaceUsesOfWith(FC1GuardBlock,
1399                                                       FC1.Header);
1400 
1401     // The guard of FC1 is not necessary anymore.
1402     FC1.GuardBranch->eraseFromParent();
1403     new UnreachableInst(FC1GuardBlock->getContext(), FC1GuardBlock);
1404 
1405     TreeUpdates.emplace_back(DominatorTree::UpdateType(
1406         DominatorTree::Delete, FC1GuardBlock, FC1.Preheader));
1407     TreeUpdates.emplace_back(DominatorTree::UpdateType(
1408         DominatorTree::Delete, FC1GuardBlock, FC1NonLoopBlock));
1409     TreeUpdates.emplace_back(DominatorTree::UpdateType(
1410         DominatorTree::Delete, FC0GuardBlock, FC1GuardBlock));
1411     TreeUpdates.emplace_back(DominatorTree::UpdateType(
1412         DominatorTree::Insert, FC0GuardBlock, FC1NonLoopBlock));
1413 
1414     assert(pred_begin(FC1GuardBlock) == pred_end(FC1GuardBlock) &&
1415            "Expecting guard block to have no predecessors");
1416     assert(succ_begin(FC1GuardBlock) == succ_end(FC1GuardBlock) &&
1417            "Expecting guard block to have no successors");
1418 
1419     // Remember the phi nodes originally in the header of FC0 in order to rewire
1420     // them later. However, this is only necessary if the new loop carried
1421     // values might not dominate the exiting branch. While we do not generally
1422     // test if this is the case but simply insert intermediate phi nodes, we
1423     // need to make sure these intermediate phi nodes have different
1424     // predecessors. To this end, we filter the special case where the exiting
1425     // block is the latch block of the first loop. Nothing needs to be done
1426     // anyway as all loop carried values dominate the latch and thereby also the
1427     // exiting branch.
1428     // KB: This is no longer necessary because FC0.ExitingBlock == FC0.Latch
1429     // (because the loops are rotated. Thus, nothing will ever be added to
1430     // OriginalFC0PHIs.
1431     SmallVector<PHINode *, 8> OriginalFC0PHIs;
1432     if (FC0.ExitingBlock != FC0.Latch)
1433       for (PHINode &PHI : FC0.Header->phis())
1434         OriginalFC0PHIs.push_back(&PHI);
1435 
1436     assert(OriginalFC0PHIs.empty() && "Expecting OriginalFC0PHIs to be empty!");
1437 
1438     // Replace incoming blocks for header PHIs first.
1439     FC1.Preheader->replaceSuccessorsPhiUsesWith(FC0.Preheader);
1440     FC0.Latch->replaceSuccessorsPhiUsesWith(FC1.Latch);
1441 
1442     // The old exiting block of the first loop (FC0) has to jump to the header
1443     // of the second as we need to execute the code in the second header block
1444     // regardless of the trip count. That is, if the trip count is 0, so the
1445     // back edge is never taken, we still have to execute both loop headers,
1446     // especially (but not only!) if the second is a do-while style loop.
1447     // However, doing so might invalidate the phi nodes of the first loop as
1448     // the new values do only need to dominate their latch and not the exiting
1449     // predicate. To remedy this potential problem we always introduce phi
1450     // nodes in the header of the second loop later that select the loop carried
1451     // value, if the second header was reached through an old latch of the
1452     // first, or undef otherwise. This is sound as exiting the first implies the
1453     // second will exit too, __without__ taking the back-edge (their
1454     // trip-counts are equal after all).
1455     FC0.ExitingBlock->getTerminator()->replaceUsesOfWith(FC0.ExitBlock,
1456                                                          FC1.Header);
1457 
1458     TreeUpdates.emplace_back(DominatorTree::UpdateType(
1459         DominatorTree::Delete, FC0.ExitingBlock, FC0.ExitBlock));
1460     TreeUpdates.emplace_back(DominatorTree::UpdateType(
1461         DominatorTree::Insert, FC0.ExitingBlock, FC1.Header));
1462 
1463     // Remove FC0 Exit Block
1464     // The exit block for FC0 is no longer needed since control will flow
1465     // directly to the header of FC1. Since it is an empty block, it can be
1466     // removed at this point.
1467     // TODO: In the future, we can handle non-empty exit blocks my merging any
1468     // instructions from FC0 exit block into FC1 exit block prior to removing
1469     // the block.
1470     assert(pred_begin(FC0.ExitBlock) == pred_end(FC0.ExitBlock) &&
1471            "Expecting exit block to be empty");
1472     FC0.ExitBlock->getTerminator()->eraseFromParent();
1473     new UnreachableInst(FC0.ExitBlock->getContext(), FC0.ExitBlock);
1474 
1475     // Remove FC1 Preheader
1476     // The pre-header of L1 is not necessary anymore.
1477     assert(pred_begin(FC1.Preheader) == pred_end(FC1.Preheader));
1478     FC1.Preheader->getTerminator()->eraseFromParent();
1479     new UnreachableInst(FC1.Preheader->getContext(), FC1.Preheader);
1480     TreeUpdates.emplace_back(DominatorTree::UpdateType(
1481         DominatorTree::Delete, FC1.Preheader, FC1.Header));
1482 
1483     // Moves the phi nodes from the second to the first loops header block.
1484     while (PHINode *PHI = dyn_cast<PHINode>(&FC1.Header->front())) {
1485       if (SE.isSCEVable(PHI->getType()))
1486         SE.forgetValue(PHI);
1487       if (PHI->hasNUsesOrMore(1))
1488         PHI->moveBefore(&*FC0.Header->getFirstInsertionPt());
1489       else
1490         PHI->eraseFromParent();
1491     }
1492 
1493     // Introduce new phi nodes in the second loop header to ensure
1494     // exiting the first and jumping to the header of the second does not break
1495     // the SSA property of the phis originally in the first loop. See also the
1496     // comment above.
1497     Instruction *L1HeaderIP = &FC1.Header->front();
1498     for (PHINode *LCPHI : OriginalFC0PHIs) {
1499       int L1LatchBBIdx = LCPHI->getBasicBlockIndex(FC1.Latch);
1500       assert(L1LatchBBIdx >= 0 &&
1501              "Expected loop carried value to be rewired at this point!");
1502 
1503       Value *LCV = LCPHI->getIncomingValue(L1LatchBBIdx);
1504 
1505       PHINode *L1HeaderPHI = PHINode::Create(
1506           LCV->getType(), 2, LCPHI->getName() + ".afterFC0", L1HeaderIP);
1507       L1HeaderPHI->addIncoming(LCV, FC0.Latch);
1508       L1HeaderPHI->addIncoming(UndefValue::get(LCV->getType()),
1509                                FC0.ExitingBlock);
1510 
1511       LCPHI->setIncomingValue(L1LatchBBIdx, L1HeaderPHI);
1512     }
1513 
1514     // Update the latches
1515 
1516     // Replace latch terminator destinations.
1517     FC0.Latch->getTerminator()->replaceUsesOfWith(FC0.Header, FC1.Header);
1518     FC1.Latch->getTerminator()->replaceUsesOfWith(FC1.Header, FC0.Header);
1519 
1520     // Change the condition of FC0 latch branch to true, as both successors of
1521     // the branch are the same.
1522     simplifyLatchBranch(FC0);
1523 
1524     // If FC0.Latch and FC0.ExitingBlock are the same then we have already
1525     // performed the updates above.
1526     if (FC0.Latch != FC0.ExitingBlock)
1527       TreeUpdates.emplace_back(DominatorTree::UpdateType(
1528           DominatorTree::Insert, FC0.Latch, FC1.Header));
1529 
1530     TreeUpdates.emplace_back(DominatorTree::UpdateType(DominatorTree::Delete,
1531                                                        FC0.Latch, FC0.Header));
1532     TreeUpdates.emplace_back(DominatorTree::UpdateType(DominatorTree::Insert,
1533                                                        FC1.Latch, FC0.Header));
1534     TreeUpdates.emplace_back(DominatorTree::UpdateType(DominatorTree::Delete,
1535                                                        FC1.Latch, FC1.Header));
1536 
1537     // All done
1538     // Apply the updates to the Dominator Tree and cleanup.
1539 
1540     assert(succ_begin(FC1GuardBlock) == succ_end(FC1GuardBlock) &&
1541            "FC1GuardBlock has successors!!");
1542     assert(pred_begin(FC1GuardBlock) == pred_end(FC1GuardBlock) &&
1543            "FC1GuardBlock has predecessors!!");
1544 
1545     // Update DT/PDT
1546     DTU.applyUpdates(TreeUpdates);
1547 
1548     LI.removeBlock(FC1.Preheader);
1549     DTU.deleteBB(FC1.Preheader);
1550     DTU.deleteBB(FC0.ExitBlock);
1551     DTU.flush();
1552 
1553     // Is there a way to keep SE up-to-date so we don't need to forget the loops
1554     // and rebuild the information in subsequent passes of fusion?
1555     // Note: Need to forget the loops before merging the loop latches, as
1556     // mergeLatch may remove the only block in FC1.
1557     SE.forgetLoop(FC1.L);
1558     SE.forgetLoop(FC0.L);
1559 
1560     // Move instructions from FC0.Latch to FC1.Latch.
1561     // Note: mergeLatch requires an updated DT.
1562     mergeLatch(FC0, FC1);
1563 
1564     // Merge the loops.
1565     SmallVector<BasicBlock *, 8> Blocks(FC1.L->block_begin(),
1566                                         FC1.L->block_end());
1567     for (BasicBlock *BB : Blocks) {
1568       FC0.L->addBlockEntry(BB);
1569       FC1.L->removeBlockFromLoop(BB);
1570       if (LI.getLoopFor(BB) != FC1.L)
1571         continue;
1572       LI.changeLoopFor(BB, FC0.L);
1573     }
1574     while (!FC1.L->empty()) {
1575       const auto &ChildLoopIt = FC1.L->begin();
1576       Loop *ChildLoop = *ChildLoopIt;
1577       FC1.L->removeChildLoop(ChildLoopIt);
1578       FC0.L->addChildLoop(ChildLoop);
1579     }
1580 
1581     // Delete the now empty loop L1.
1582     LI.erase(FC1.L);
1583 
1584 #ifndef NDEBUG
1585     assert(!verifyFunction(*FC0.Header->getParent(), &errs()));
1586     assert(DT.verify(DominatorTree::VerificationLevel::Fast));
1587     assert(PDT.verify());
1588     LI.verify(DT);
1589     SE.verify();
1590 #endif
1591 
1592     LLVM_DEBUG(dbgs() << "Fusion done:\n");
1593 
1594     return FC0.L;
1595   }
1596 };
1597 
1598 struct LoopFuseLegacy : public FunctionPass {
1599 
1600   static char ID;
1601 
1602   LoopFuseLegacy() : FunctionPass(ID) {
1603     initializeLoopFuseLegacyPass(*PassRegistry::getPassRegistry());
1604   }
1605 
1606   void getAnalysisUsage(AnalysisUsage &AU) const override {
1607     AU.addRequiredID(LoopSimplifyID);
1608     AU.addRequired<ScalarEvolutionWrapperPass>();
1609     AU.addRequired<LoopInfoWrapperPass>();
1610     AU.addRequired<DominatorTreeWrapperPass>();
1611     AU.addRequired<PostDominatorTreeWrapperPass>();
1612     AU.addRequired<OptimizationRemarkEmitterWrapperPass>();
1613     AU.addRequired<DependenceAnalysisWrapperPass>();
1614 
1615     AU.addPreserved<ScalarEvolutionWrapperPass>();
1616     AU.addPreserved<LoopInfoWrapperPass>();
1617     AU.addPreserved<DominatorTreeWrapperPass>();
1618     AU.addPreserved<PostDominatorTreeWrapperPass>();
1619   }
1620 
1621   bool runOnFunction(Function &F) override {
1622     if (skipFunction(F))
1623       return false;
1624     auto &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1625     auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1626     auto &DI = getAnalysis<DependenceAnalysisWrapperPass>().getDI();
1627     auto &SE = getAnalysis<ScalarEvolutionWrapperPass>().getSE();
1628     auto &PDT = getAnalysis<PostDominatorTreeWrapperPass>().getPostDomTree();
1629     auto &ORE = getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE();
1630 
1631     const DataLayout &DL = F.getParent()->getDataLayout();
1632     LoopFuser LF(LI, DT, DI, SE, PDT, ORE, DL);
1633     return LF.fuseLoops(F);
1634   }
1635 };
1636 } // namespace
1637 
1638 PreservedAnalyses LoopFusePass::run(Function &F, FunctionAnalysisManager &AM) {
1639   auto &LI = AM.getResult<LoopAnalysis>(F);
1640   auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
1641   auto &DI = AM.getResult<DependenceAnalysis>(F);
1642   auto &SE = AM.getResult<ScalarEvolutionAnalysis>(F);
1643   auto &PDT = AM.getResult<PostDominatorTreeAnalysis>(F);
1644   auto &ORE = AM.getResult<OptimizationRemarkEmitterAnalysis>(F);
1645 
1646   const DataLayout &DL = F.getParent()->getDataLayout();
1647   LoopFuser LF(LI, DT, DI, SE, PDT, ORE, DL);
1648   bool Changed = LF.fuseLoops(F);
1649   if (!Changed)
1650     return PreservedAnalyses::all();
1651 
1652   PreservedAnalyses PA;
1653   PA.preserve<DominatorTreeAnalysis>();
1654   PA.preserve<PostDominatorTreeAnalysis>();
1655   PA.preserve<ScalarEvolutionAnalysis>();
1656   PA.preserve<LoopAnalysis>();
1657   return PA;
1658 }
1659 
1660 char LoopFuseLegacy::ID = 0;
1661 
1662 INITIALIZE_PASS_BEGIN(LoopFuseLegacy, "loop-fusion", "Loop Fusion", false,
1663                       false)
1664 INITIALIZE_PASS_DEPENDENCY(PostDominatorTreeWrapperPass)
1665 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
1666 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
1667 INITIALIZE_PASS_DEPENDENCY(DependenceAnalysisWrapperPass)
1668 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
1669 INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass)
1670 INITIALIZE_PASS_END(LoopFuseLegacy, "loop-fusion", "Loop Fusion", false, false)
1671 
1672 FunctionPass *llvm::createLoopFusePass() { return new LoopFuseLegacy(); }
1673