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/AssumptionCache.h"
50 #include "llvm/Analysis/DependenceAnalysis.h"
51 #include "llvm/Analysis/DomTreeUpdater.h"
52 #include "llvm/Analysis/LoopInfo.h"
53 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
54 #include "llvm/Analysis/PostDominators.h"
55 #include "llvm/Analysis/ScalarEvolution.h"
56 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
57 #include "llvm/Analysis/TargetTransformInfo.h"
58 #include "llvm/IR/Function.h"
59 #include "llvm/IR/Verifier.h"
60 #include "llvm/InitializePasses.h"
61 #include "llvm/Pass.h"
62 #include "llvm/Support/CommandLine.h"
63 #include "llvm/Support/Debug.h"
64 #include "llvm/Support/raw_ostream.h"
65 #include "llvm/Transforms/Scalar.h"
66 #include "llvm/Transforms/Utils.h"
67 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
68 #include "llvm/Transforms/Utils/CodeMoverUtils.h"
69 #include "llvm/Transforms/Utils/LoopPeel.h"
70 #include "llvm/Transforms/Utils/LoopSimplify.h"
71 
72 using namespace llvm;
73 
74 #define DEBUG_TYPE "loop-fusion"
75 
76 STATISTIC(FuseCounter, "Loops fused");
77 STATISTIC(NumFusionCandidates, "Number of candidates for loop fusion");
78 STATISTIC(InvalidPreheader, "Loop has invalid preheader");
79 STATISTIC(InvalidHeader, "Loop has invalid header");
80 STATISTIC(InvalidExitingBlock, "Loop has invalid exiting blocks");
81 STATISTIC(InvalidExitBlock, "Loop has invalid exit block");
82 STATISTIC(InvalidLatch, "Loop has invalid latch");
83 STATISTIC(InvalidLoop, "Loop is invalid");
84 STATISTIC(AddressTakenBB, "Basic block has address taken");
85 STATISTIC(MayThrowException, "Loop may throw an exception");
86 STATISTIC(ContainsVolatileAccess, "Loop contains a volatile access");
87 STATISTIC(NotSimplifiedForm, "Loop is not in simplified form");
88 STATISTIC(InvalidDependencies, "Dependencies prevent fusion");
89 STATISTIC(UnknownTripCount, "Loop has unknown trip count");
90 STATISTIC(UncomputableTripCount, "SCEV cannot compute trip count of loop");
91 STATISTIC(NonEqualTripCount, "Loop trip counts are not the same");
92 STATISTIC(NonAdjacent, "Loops are not adjacent");
93 STATISTIC(
94     NonEmptyPreheader,
95     "Loop has a non-empty preheader with instructions that cannot be moved");
96 STATISTIC(FusionNotBeneficial, "Fusion is not beneficial");
97 STATISTIC(NonIdenticalGuards, "Candidates have different guards");
98 STATISTIC(NonEmptyExitBlock, "Candidate has a non-empty exit block with "
99                              "instructions that cannot be moved");
100 STATISTIC(NonEmptyGuardBlock, "Candidate has a non-empty guard block with "
101                               "instructions that cannot be moved");
102 STATISTIC(NotRotated, "Candidate is not rotated");
103 STATISTIC(OnlySecondCandidateIsGuarded,
104           "The second candidate is guarded while the first one is not");
105 STATISTIC(NumHoistedInsts, "Number of hoisted preheader instructions.");
106 STATISTIC(NumSunkInsts, "Number of hoisted preheader instructions.");
107 
108 enum FusionDependenceAnalysisChoice {
109   FUSION_DEPENDENCE_ANALYSIS_SCEV,
110   FUSION_DEPENDENCE_ANALYSIS_DA,
111   FUSION_DEPENDENCE_ANALYSIS_ALL,
112 };
113 
114 static cl::opt<FusionDependenceAnalysisChoice> FusionDependenceAnalysis(
115     "loop-fusion-dependence-analysis",
116     cl::desc("Which dependence analysis should loop fusion use?"),
117     cl::values(clEnumValN(FUSION_DEPENDENCE_ANALYSIS_SCEV, "scev",
118                           "Use the scalar evolution interface"),
119                clEnumValN(FUSION_DEPENDENCE_ANALYSIS_DA, "da",
120                           "Use the dependence analysis interface"),
121                clEnumValN(FUSION_DEPENDENCE_ANALYSIS_ALL, "all",
122                           "Use all available analyses")),
123     cl::Hidden, cl::init(FUSION_DEPENDENCE_ANALYSIS_ALL));
124 
125 static cl::opt<unsigned> FusionPeelMaxCount(
126     "loop-fusion-peel-max-count", cl::init(0), cl::Hidden,
127     cl::desc("Max number of iterations to be peeled from a loop, such that "
128              "fusion can take place"));
129 
130 #ifndef NDEBUG
131 static cl::opt<bool>
132     VerboseFusionDebugging("loop-fusion-verbose-debug",
133                            cl::desc("Enable verbose debugging for Loop Fusion"),
134                            cl::Hidden, cl::init(false));
135 #endif
136 
137 namespace {
138 /// This class is used to represent a candidate for loop fusion. When it is
139 /// constructed, it checks the conditions for loop fusion to ensure that it
140 /// represents a valid candidate. It caches several parts of a loop that are
141 /// used throughout loop fusion (e.g., loop preheader, loop header, etc) instead
142 /// of continually querying the underlying Loop to retrieve these values. It is
143 /// assumed these will not change throughout loop fusion.
144 ///
145 /// The invalidate method should be used to indicate that the FusionCandidate is
146 /// no longer a valid candidate for fusion. Similarly, the isValid() method can
147 /// be used to ensure that the FusionCandidate is still valid for fusion.
148 struct FusionCandidate {
149   /// Cache of parts of the loop used throughout loop fusion. These should not
150   /// need to change throughout the analysis and transformation.
151   /// These parts are cached to avoid repeatedly looking up in the Loop class.
152 
153   /// Preheader of the loop this candidate represents
154   BasicBlock *Preheader;
155   /// Header of the loop this candidate represents
156   BasicBlock *Header;
157   /// Blocks in the loop that exit the loop
158   BasicBlock *ExitingBlock;
159   /// The successor block of this loop (where the exiting blocks go to)
160   BasicBlock *ExitBlock;
161   /// Latch of the loop
162   BasicBlock *Latch;
163   /// The loop that this fusion candidate represents
164   Loop *L;
165   /// Vector of instructions in this loop that read from memory
166   SmallVector<Instruction *, 16> MemReads;
167   /// Vector of instructions in this loop that write to memory
168   SmallVector<Instruction *, 16> MemWrites;
169   /// Are all of the members of this fusion candidate still valid
170   bool Valid;
171   /// Guard branch of the loop, if it exists
172   BranchInst *GuardBranch;
173   /// Peeling Paramaters of the Loop.
174   TTI::PeelingPreferences PP;
175   /// Can you Peel this Loop?
176   bool AbleToPeel;
177   /// Has this loop been Peeled
178   bool Peeled;
179 
180   /// Dominator and PostDominator trees are needed for the
181   /// FusionCandidateCompare function, required by FusionCandidateSet to
182   /// determine where the FusionCandidate should be inserted into the set. These
183   /// are used to establish ordering of the FusionCandidates based on dominance.
184   DominatorTree &DT;
185   const PostDominatorTree *PDT;
186 
187   OptimizationRemarkEmitter &ORE;
188 
189   FusionCandidate(Loop *L, DominatorTree &DT, const PostDominatorTree *PDT,
190                   OptimizationRemarkEmitter &ORE, TTI::PeelingPreferences PP)
191       : Preheader(L->getLoopPreheader()), Header(L->getHeader()),
192         ExitingBlock(L->getExitingBlock()), ExitBlock(L->getExitBlock()),
193         Latch(L->getLoopLatch()), L(L), Valid(true),
194         GuardBranch(L->getLoopGuardBranch()), PP(PP), AbleToPeel(canPeel(L)),
195         Peeled(false), DT(DT), PDT(PDT), ORE(ORE) {
196 
197     // Walk over all blocks in the loop and check for conditions that may
198     // prevent fusion. For each block, walk over all instructions and collect
199     // the memory reads and writes If any instructions that prevent fusion are
200     // found, invalidate this object and return.
201     for (BasicBlock *BB : L->blocks()) {
202       if (BB->hasAddressTaken()) {
203         invalidate();
204         reportInvalidCandidate(AddressTakenBB);
205         return;
206       }
207 
208       for (Instruction &I : *BB) {
209         if (I.mayThrow()) {
210           invalidate();
211           reportInvalidCandidate(MayThrowException);
212           return;
213         }
214         if (StoreInst *SI = dyn_cast<StoreInst>(&I)) {
215           if (SI->isVolatile()) {
216             invalidate();
217             reportInvalidCandidate(ContainsVolatileAccess);
218             return;
219           }
220         }
221         if (LoadInst *LI = dyn_cast<LoadInst>(&I)) {
222           if (LI->isVolatile()) {
223             invalidate();
224             reportInvalidCandidate(ContainsVolatileAccess);
225             return;
226           }
227         }
228         if (I.mayWriteToMemory())
229           MemWrites.push_back(&I);
230         if (I.mayReadFromMemory())
231           MemReads.push_back(&I);
232       }
233     }
234   }
235 
236   /// Check if all members of the class are valid.
237   bool isValid() const {
238     return Preheader && Header && ExitingBlock && ExitBlock && Latch && L &&
239            !L->isInvalid() && Valid;
240   }
241 
242   /// Verify that all members are in sync with the Loop object.
243   void verify() const {
244     assert(isValid() && "Candidate is not valid!!");
245     assert(!L->isInvalid() && "Loop is invalid!");
246     assert(Preheader == L->getLoopPreheader() && "Preheader is out of sync");
247     assert(Header == L->getHeader() && "Header is out of sync");
248     assert(ExitingBlock == L->getExitingBlock() &&
249            "Exiting Blocks is out of sync");
250     assert(ExitBlock == L->getExitBlock() && "Exit block is out of sync");
251     assert(Latch == L->getLoopLatch() && "Latch is out of sync");
252   }
253 
254   /// Get the entry block for this fusion candidate.
255   ///
256   /// If this fusion candidate represents a guarded loop, the entry block is the
257   /// loop guard block. If it represents an unguarded loop, the entry block is
258   /// the preheader of the loop.
259   BasicBlock *getEntryBlock() const {
260     if (GuardBranch)
261       return GuardBranch->getParent();
262     else
263       return Preheader;
264   }
265 
266   /// After Peeling the loop is modified quite a bit, hence all of the Blocks
267   /// need to be updated accordingly.
268   void updateAfterPeeling() {
269     Preheader = L->getLoopPreheader();
270     Header = L->getHeader();
271     ExitingBlock = L->getExitingBlock();
272     ExitBlock = L->getExitBlock();
273     Latch = L->getLoopLatch();
274     verify();
275   }
276 
277   /// Given a guarded loop, get the successor of the guard that is not in the
278   /// loop.
279   ///
280   /// This method returns the successor of the loop guard that is not located
281   /// within the loop (i.e., the successor of the guard that is not the
282   /// preheader).
283   /// This method is only valid for guarded loops.
284   BasicBlock *getNonLoopBlock() const {
285     assert(GuardBranch && "Only valid on guarded loops.");
286     assert(GuardBranch->isConditional() &&
287            "Expecting guard to be a conditional branch.");
288     if (Peeled)
289       return GuardBranch->getSuccessor(1);
290     return (GuardBranch->getSuccessor(0) == Preheader)
291                ? GuardBranch->getSuccessor(1)
292                : GuardBranch->getSuccessor(0);
293   }
294 
295 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
296   LLVM_DUMP_METHOD void dump() const {
297     dbgs() << "\tGuardBranch: ";
298     if (GuardBranch)
299       dbgs() << *GuardBranch;
300     else
301       dbgs() << "nullptr";
302     dbgs() << "\n"
303            << (GuardBranch ? GuardBranch->getName() : "nullptr") << "\n"
304            << "\tPreheader: " << (Preheader ? Preheader->getName() : "nullptr")
305            << "\n"
306            << "\tHeader: " << (Header ? Header->getName() : "nullptr") << "\n"
307            << "\tExitingBB: "
308            << (ExitingBlock ? ExitingBlock->getName() : "nullptr") << "\n"
309            << "\tExitBB: " << (ExitBlock ? ExitBlock->getName() : "nullptr")
310            << "\n"
311            << "\tLatch: " << (Latch ? Latch->getName() : "nullptr") << "\n"
312            << "\tEntryBlock: "
313            << (getEntryBlock() ? getEntryBlock()->getName() : "nullptr")
314            << "\n";
315   }
316 #endif
317 
318   /// Determine if a fusion candidate (representing a loop) is eligible for
319   /// fusion. Note that this only checks whether a single loop can be fused - it
320   /// does not check whether it is *legal* to fuse two loops together.
321   bool isEligibleForFusion(ScalarEvolution &SE) const {
322     if (!isValid()) {
323       LLVM_DEBUG(dbgs() << "FC has invalid CFG requirements!\n");
324       if (!Preheader)
325         ++InvalidPreheader;
326       if (!Header)
327         ++InvalidHeader;
328       if (!ExitingBlock)
329         ++InvalidExitingBlock;
330       if (!ExitBlock)
331         ++InvalidExitBlock;
332       if (!Latch)
333         ++InvalidLatch;
334       if (L->isInvalid())
335         ++InvalidLoop;
336 
337       return false;
338     }
339 
340     // Require ScalarEvolution to be able to determine a trip count.
341     if (!SE.hasLoopInvariantBackedgeTakenCount(L)) {
342       LLVM_DEBUG(dbgs() << "Loop " << L->getName()
343                         << " trip count not computable!\n");
344       return reportInvalidCandidate(UnknownTripCount);
345     }
346 
347     if (!L->isLoopSimplifyForm()) {
348       LLVM_DEBUG(dbgs() << "Loop " << L->getName()
349                         << " is not in simplified form!\n");
350       return reportInvalidCandidate(NotSimplifiedForm);
351     }
352 
353     if (!L->isRotatedForm()) {
354       LLVM_DEBUG(dbgs() << "Loop " << L->getName() << " is not rotated!\n");
355       return reportInvalidCandidate(NotRotated);
356     }
357 
358     return true;
359   }
360 
361 private:
362   // This is only used internally for now, to clear the MemWrites and MemReads
363   // list and setting Valid to false. I can't envision other uses of this right
364   // now, since once FusionCandidates are put into the FusionCandidateSet they
365   // are immutable. Thus, any time we need to change/update a FusionCandidate,
366   // we must create a new one and insert it into the FusionCandidateSet to
367   // ensure the FusionCandidateSet remains ordered correctly.
368   void invalidate() {
369     MemWrites.clear();
370     MemReads.clear();
371     Valid = false;
372   }
373 
374   bool reportInvalidCandidate(llvm::Statistic &Stat) const {
375     using namespace ore;
376     assert(L && Preheader && "Fusion candidate not initialized properly!");
377 #if LLVM_ENABLE_STATS
378     ++Stat;
379     ORE.emit(OptimizationRemarkAnalysis(DEBUG_TYPE, Stat.getName(),
380                                         L->getStartLoc(), Preheader)
381              << "[" << Preheader->getParent()->getName() << "]: "
382              << "Loop is not a candidate for fusion: " << Stat.getDesc());
383 #endif
384     return false;
385   }
386 };
387 
388 struct FusionCandidateCompare {
389   /// Comparison functor to sort two Control Flow Equivalent fusion candidates
390   /// into dominance order.
391   /// If LHS dominates RHS and RHS post-dominates LHS, return true;
392   /// If RHS dominates LHS and LHS post-dominates RHS, return false;
393   /// If both LHS and RHS are not dominating each other then, non-strictly
394   /// post dominate check will decide the order of candidates. If RHS
395   /// non-strictly post dominates LHS then, return true. If LHS non-strictly
396   /// post dominates RHS then, return false. If both are non-strictly post
397   /// dominate each other then, level in the post dominator tree will decide
398   /// the order of candidates.
399   bool operator()(const FusionCandidate &LHS,
400                   const FusionCandidate &RHS) const {
401     const DominatorTree *DT = &(LHS.DT);
402 
403     BasicBlock *LHSEntryBlock = LHS.getEntryBlock();
404     BasicBlock *RHSEntryBlock = RHS.getEntryBlock();
405 
406     // Do not save PDT to local variable as it is only used in asserts and thus
407     // will trigger an unused variable warning if building without asserts.
408     assert(DT && LHS.PDT && "Expecting valid dominator tree");
409 
410     // Do this compare first so if LHS == RHS, function returns false.
411     if (DT->dominates(RHSEntryBlock, LHSEntryBlock)) {
412       // RHS dominates LHS
413       // Verify LHS post-dominates RHS
414       assert(LHS.PDT->dominates(LHSEntryBlock, RHSEntryBlock));
415       return false;
416     }
417 
418     if (DT->dominates(LHSEntryBlock, RHSEntryBlock)) {
419       // Verify RHS Postdominates LHS
420       assert(LHS.PDT->dominates(RHSEntryBlock, LHSEntryBlock));
421       return true;
422     }
423 
424     // If two FusionCandidates are in the same level of dominator tree,
425     // they will not dominate each other, but may still be control flow
426     // equivalent. To sort those FusionCandidates, nonStrictlyPostDominate()
427     // function is needed.
428     bool WrongOrder =
429         nonStrictlyPostDominate(LHSEntryBlock, RHSEntryBlock, DT, LHS.PDT);
430     bool RightOrder =
431         nonStrictlyPostDominate(RHSEntryBlock, LHSEntryBlock, DT, LHS.PDT);
432     if (WrongOrder && RightOrder) {
433       // If common predecessor of LHS and RHS post dominates both
434       // FusionCandidates then, Order of FusionCandidate can be
435       // identified by its level in post dominator tree.
436       DomTreeNode *LNode = LHS.PDT->getNode(LHSEntryBlock);
437       DomTreeNode *RNode = LHS.PDT->getNode(RHSEntryBlock);
438       return LNode->getLevel() > RNode->getLevel();
439     } else if (WrongOrder)
440       return false;
441     else if (RightOrder)
442       return true;
443 
444     // If LHS does not non-strict Postdominate RHS and RHS does not non-strict
445     // Postdominate LHS then, there is no dominance relationship between the
446     // two FusionCandidates. Thus, they should not be in the same set together.
447     llvm_unreachable(
448         "No dominance relationship between these fusion candidates!");
449   }
450 };
451 
452 using LoopVector = SmallVector<Loop *, 4>;
453 
454 // Set of Control Flow Equivalent (CFE) Fusion Candidates, sorted in dominance
455 // order. Thus, if FC0 comes *before* FC1 in a FusionCandidateSet, then FC0
456 // dominates FC1 and FC1 post-dominates FC0.
457 // std::set was chosen because we want a sorted data structure with stable
458 // iterators. A subsequent patch to loop fusion will enable fusing non-adjacent
459 // loops by moving intervening code around. When this intervening code contains
460 // loops, those loops will be moved also. The corresponding FusionCandidates
461 // will also need to be moved accordingly. As this is done, having stable
462 // iterators will simplify the logic. Similarly, having an efficient insert that
463 // keeps the FusionCandidateSet sorted will also simplify the implementation.
464 using FusionCandidateSet = std::set<FusionCandidate, FusionCandidateCompare>;
465 using FusionCandidateCollection = SmallVector<FusionCandidateSet, 4>;
466 
467 #if !defined(NDEBUG)
468 static llvm::raw_ostream &operator<<(llvm::raw_ostream &OS,
469                                      const FusionCandidate &FC) {
470   if (FC.isValid())
471     OS << FC.Preheader->getName();
472   else
473     OS << "<Invalid>";
474 
475   return OS;
476 }
477 
478 static llvm::raw_ostream &operator<<(llvm::raw_ostream &OS,
479                                      const FusionCandidateSet &CandSet) {
480   for (const FusionCandidate &FC : CandSet)
481     OS << FC << '\n';
482 
483   return OS;
484 }
485 
486 static void
487 printFusionCandidates(const FusionCandidateCollection &FusionCandidates) {
488   dbgs() << "Fusion Candidates: \n";
489   for (const auto &CandidateSet : FusionCandidates) {
490     dbgs() << "*** Fusion Candidate Set ***\n";
491     dbgs() << CandidateSet;
492     dbgs() << "****************************\n";
493   }
494 }
495 #endif
496 
497 /// Collect all loops in function at the same nest level, starting at the
498 /// outermost level.
499 ///
500 /// This data structure collects all loops at the same nest level for a
501 /// given function (specified by the LoopInfo object). It starts at the
502 /// outermost level.
503 struct LoopDepthTree {
504   using LoopsOnLevelTy = SmallVector<LoopVector, 4>;
505   using iterator = LoopsOnLevelTy::iterator;
506   using const_iterator = LoopsOnLevelTy::const_iterator;
507 
508   LoopDepthTree(LoopInfo &LI) : Depth(1) {
509     if (!LI.empty())
510       LoopsOnLevel.emplace_back(LoopVector(LI.rbegin(), LI.rend()));
511   }
512 
513   /// Test whether a given loop has been removed from the function, and thus is
514   /// no longer valid.
515   bool isRemovedLoop(const Loop *L) const { return RemovedLoops.count(L); }
516 
517   /// Record that a given loop has been removed from the function and is no
518   /// longer valid.
519   void removeLoop(const Loop *L) { RemovedLoops.insert(L); }
520 
521   /// Descend the tree to the next (inner) nesting level
522   void descend() {
523     LoopsOnLevelTy LoopsOnNextLevel;
524 
525     for (const LoopVector &LV : *this)
526       for (Loop *L : LV)
527         if (!isRemovedLoop(L) && L->begin() != L->end())
528           LoopsOnNextLevel.emplace_back(LoopVector(L->begin(), L->end()));
529 
530     LoopsOnLevel = LoopsOnNextLevel;
531     RemovedLoops.clear();
532     Depth++;
533   }
534 
535   bool empty() const { return size() == 0; }
536   size_t size() const { return LoopsOnLevel.size() - RemovedLoops.size(); }
537   unsigned getDepth() const { return Depth; }
538 
539   iterator begin() { return LoopsOnLevel.begin(); }
540   iterator end() { return LoopsOnLevel.end(); }
541   const_iterator begin() const { return LoopsOnLevel.begin(); }
542   const_iterator end() const { return LoopsOnLevel.end(); }
543 
544 private:
545   /// Set of loops that have been removed from the function and are no longer
546   /// valid.
547   SmallPtrSet<const Loop *, 8> RemovedLoops;
548 
549   /// Depth of the current level, starting at 1 (outermost loops).
550   unsigned Depth;
551 
552   /// Vector of loops at the current depth level that have the same parent loop
553   LoopsOnLevelTy LoopsOnLevel;
554 };
555 
556 #ifndef NDEBUG
557 static void printLoopVector(const LoopVector &LV) {
558   dbgs() << "****************************\n";
559   for (auto *L : LV)
560     printLoop(*L, dbgs());
561   dbgs() << "****************************\n";
562 }
563 #endif
564 
565 struct LoopFuser {
566 private:
567   // Sets of control flow equivalent fusion candidates for a given nest level.
568   FusionCandidateCollection FusionCandidates;
569 
570   LoopDepthTree LDT;
571   DomTreeUpdater DTU;
572 
573   LoopInfo &LI;
574   DominatorTree &DT;
575   DependenceInfo &DI;
576   ScalarEvolution &SE;
577   PostDominatorTree &PDT;
578   OptimizationRemarkEmitter &ORE;
579   AssumptionCache &AC;
580   const TargetTransformInfo &TTI;
581 
582 public:
583   LoopFuser(LoopInfo &LI, DominatorTree &DT, DependenceInfo &DI,
584             ScalarEvolution &SE, PostDominatorTree &PDT,
585             OptimizationRemarkEmitter &ORE, const DataLayout &DL,
586             AssumptionCache &AC, const TargetTransformInfo &TTI)
587       : LDT(LI), DTU(DT, PDT, DomTreeUpdater::UpdateStrategy::Lazy), LI(LI),
588         DT(DT), DI(DI), SE(SE), PDT(PDT), ORE(ORE), AC(AC), TTI(TTI) {}
589 
590   /// This is the main entry point for loop fusion. It will traverse the
591   /// specified function and collect candidate loops to fuse, starting at the
592   /// outermost nesting level and working inwards.
593   bool fuseLoops(Function &F) {
594 #ifndef NDEBUG
595     if (VerboseFusionDebugging) {
596       LI.print(dbgs());
597     }
598 #endif
599 
600     LLVM_DEBUG(dbgs() << "Performing Loop Fusion on function " << F.getName()
601                       << "\n");
602     bool Changed = false;
603 
604     while (!LDT.empty()) {
605       LLVM_DEBUG(dbgs() << "Got " << LDT.size() << " loop sets for depth "
606                         << LDT.getDepth() << "\n";);
607 
608       for (const LoopVector &LV : LDT) {
609         assert(LV.size() > 0 && "Empty loop set was build!");
610 
611         // Skip singleton loop sets as they do not offer fusion opportunities on
612         // this level.
613         if (LV.size() == 1)
614           continue;
615 #ifndef NDEBUG
616         if (VerboseFusionDebugging) {
617           LLVM_DEBUG({
618             dbgs() << "  Visit loop set (#" << LV.size() << "):\n";
619             printLoopVector(LV);
620           });
621         }
622 #endif
623 
624         collectFusionCandidates(LV);
625         Changed |= fuseCandidates();
626       }
627 
628       // Finished analyzing candidates at this level.
629       // Descend to the next level and clear all of the candidates currently
630       // collected. Note that it will not be possible to fuse any of the
631       // existing candidates with new candidates because the new candidates will
632       // be at a different nest level and thus not be control flow equivalent
633       // with all of the candidates collected so far.
634       LLVM_DEBUG(dbgs() << "Descend one level!\n");
635       LDT.descend();
636       FusionCandidates.clear();
637     }
638 
639     if (Changed)
640       LLVM_DEBUG(dbgs() << "Function after Loop Fusion: \n"; F.dump(););
641 
642 #ifndef NDEBUG
643     assert(DT.verify());
644     assert(PDT.verify());
645     LI.verify(DT);
646     SE.verify();
647 #endif
648 
649     LLVM_DEBUG(dbgs() << "Loop Fusion complete\n");
650     return Changed;
651   }
652 
653 private:
654   /// Determine if two fusion candidates are control flow equivalent.
655   ///
656   /// Two fusion candidates are control flow equivalent if when one executes,
657   /// the other is guaranteed to execute. This is determined using dominators
658   /// and post-dominators: if A dominates B and B post-dominates A then A and B
659   /// are control-flow equivalent.
660   bool isControlFlowEquivalent(const FusionCandidate &FC0,
661                                const FusionCandidate &FC1) const {
662     assert(FC0.Preheader && FC1.Preheader && "Expecting valid preheaders");
663 
664     return ::isControlFlowEquivalent(*FC0.getEntryBlock(), *FC1.getEntryBlock(),
665                                      DT, PDT);
666   }
667 
668   /// Iterate over all loops in the given loop set and identify the loops that
669   /// are eligible for fusion. Place all eligible fusion candidates into Control
670   /// Flow Equivalent sets, sorted by dominance.
671   void collectFusionCandidates(const LoopVector &LV) {
672     for (Loop *L : LV) {
673       TTI::PeelingPreferences PP =
674           gatherPeelingPreferences(L, SE, TTI, std::nullopt, std::nullopt);
675       FusionCandidate CurrCand(L, DT, &PDT, ORE, PP);
676       if (!CurrCand.isEligibleForFusion(SE))
677         continue;
678 
679       // Go through each list in FusionCandidates and determine if L is control
680       // flow equivalent with the first loop in that list. If it is, append LV.
681       // If not, go to the next list.
682       // If no suitable list is found, start another list and add it to
683       // FusionCandidates.
684       bool FoundSet = false;
685 
686       for (auto &CurrCandSet : FusionCandidates) {
687         if (isControlFlowEquivalent(*CurrCandSet.begin(), CurrCand)) {
688           CurrCandSet.insert(CurrCand);
689           FoundSet = true;
690 #ifndef NDEBUG
691           if (VerboseFusionDebugging)
692             LLVM_DEBUG(dbgs() << "Adding " << CurrCand
693                               << " to existing candidate set\n");
694 #endif
695           break;
696         }
697       }
698       if (!FoundSet) {
699         // No set was found. Create a new set and add to FusionCandidates
700 #ifndef NDEBUG
701         if (VerboseFusionDebugging)
702           LLVM_DEBUG(dbgs() << "Adding " << CurrCand << " to new set\n");
703 #endif
704         FusionCandidateSet NewCandSet;
705         NewCandSet.insert(CurrCand);
706         FusionCandidates.push_back(NewCandSet);
707       }
708       NumFusionCandidates++;
709     }
710   }
711 
712   /// Determine if it is beneficial to fuse two loops.
713   ///
714   /// For now, this method simply returns true because we want to fuse as much
715   /// as possible (primarily to test the pass). This method will evolve, over
716   /// time, to add heuristics for profitability of fusion.
717   bool isBeneficialFusion(const FusionCandidate &FC0,
718                           const FusionCandidate &FC1) {
719     return true;
720   }
721 
722   /// Determine if two fusion candidates have the same trip count (i.e., they
723   /// execute the same number of iterations).
724   ///
725   /// This function will return a pair of values. The first is a boolean,
726   /// stating whether or not the two candidates are known at compile time to
727   /// have the same TripCount. The second is the difference in the two
728   /// TripCounts. This information can be used later to determine whether or not
729   /// peeling can be performed on either one of the candidates.
730   std::pair<bool, std::optional<unsigned>>
731   haveIdenticalTripCounts(const FusionCandidate &FC0,
732                           const FusionCandidate &FC1) const {
733     const SCEV *TripCount0 = SE.getBackedgeTakenCount(FC0.L);
734     if (isa<SCEVCouldNotCompute>(TripCount0)) {
735       UncomputableTripCount++;
736       LLVM_DEBUG(dbgs() << "Trip count of first loop could not be computed!");
737       return {false, std::nullopt};
738     }
739 
740     const SCEV *TripCount1 = SE.getBackedgeTakenCount(FC1.L);
741     if (isa<SCEVCouldNotCompute>(TripCount1)) {
742       UncomputableTripCount++;
743       LLVM_DEBUG(dbgs() << "Trip count of second loop could not be computed!");
744       return {false, std::nullopt};
745     }
746 
747     LLVM_DEBUG(dbgs() << "\tTrip counts: " << *TripCount0 << " & "
748                       << *TripCount1 << " are "
749                       << (TripCount0 == TripCount1 ? "identical" : "different")
750                       << "\n");
751 
752     if (TripCount0 == TripCount1)
753       return {true, 0};
754 
755     LLVM_DEBUG(dbgs() << "The loops do not have the same tripcount, "
756                          "determining the difference between trip counts\n");
757 
758     // Currently only considering loops with a single exit point
759     // and a non-constant trip count.
760     const unsigned TC0 = SE.getSmallConstantTripCount(FC0.L);
761     const unsigned TC1 = SE.getSmallConstantTripCount(FC1.L);
762 
763     // If any of the tripcounts are zero that means that loop(s) do not have
764     // a single exit or a constant tripcount.
765     if (TC0 == 0 || TC1 == 0) {
766       LLVM_DEBUG(dbgs() << "Loop(s) do not have a single exit point or do not "
767                            "have a constant number of iterations. Peeling "
768                            "is not benefical\n");
769       return {false, std::nullopt};
770     }
771 
772     std::optional<unsigned> Difference;
773     int Diff = TC0 - TC1;
774 
775     if (Diff > 0)
776       Difference = Diff;
777     else {
778       LLVM_DEBUG(
779           dbgs() << "Difference is less than 0. FC1 (second loop) has more "
780                     "iterations than the first one. Currently not supported\n");
781     }
782 
783     LLVM_DEBUG(dbgs() << "Difference in loop trip count is: " << Difference
784                       << "\n");
785 
786     return {false, Difference};
787   }
788 
789   void peelFusionCandidate(FusionCandidate &FC0, const FusionCandidate &FC1,
790                            unsigned PeelCount) {
791     assert(FC0.AbleToPeel && "Should be able to peel loop");
792 
793     LLVM_DEBUG(dbgs() << "Attempting to peel first " << PeelCount
794                       << " iterations of the first loop. \n");
795 
796     ValueToValueMapTy VMap;
797     FC0.Peeled = peelLoop(FC0.L, PeelCount, &LI, &SE, DT, &AC, true, VMap);
798     if (FC0.Peeled) {
799       LLVM_DEBUG(dbgs() << "Done Peeling\n");
800 
801 #ifndef NDEBUG
802       auto IdenticalTripCount = haveIdenticalTripCounts(FC0, FC1);
803 
804       assert(IdenticalTripCount.first && *IdenticalTripCount.second == 0 &&
805              "Loops should have identical trip counts after peeling");
806 #endif
807 
808       FC0.PP.PeelCount += PeelCount;
809 
810       // Peeling does not update the PDT
811       PDT.recalculate(*FC0.Preheader->getParent());
812 
813       FC0.updateAfterPeeling();
814 
815       // In this case the iterations of the loop are constant, so the first
816       // loop will execute completely (will not jump from one of
817       // the peeled blocks to the second loop). Here we are updating the
818       // branch conditions of each of the peeled blocks, such that it will
819       // branch to its successor which is not the preheader of the second loop
820       // in the case of unguarded loops, or the succesors of the exit block of
821       // the first loop otherwise. Doing this update will ensure that the entry
822       // block of the first loop dominates the entry block of the second loop.
823       BasicBlock *BB =
824           FC0.GuardBranch ? FC0.ExitBlock->getUniqueSuccessor() : FC1.Preheader;
825       if (BB) {
826         SmallVector<DominatorTree::UpdateType, 8> TreeUpdates;
827         SmallVector<Instruction *, 8> WorkList;
828         for (BasicBlock *Pred : predecessors(BB)) {
829           if (Pred != FC0.ExitBlock) {
830             WorkList.emplace_back(Pred->getTerminator());
831             TreeUpdates.emplace_back(
832                 DominatorTree::UpdateType(DominatorTree::Delete, Pred, BB));
833           }
834         }
835         // Cannot modify the predecessors inside the above loop as it will cause
836         // the iterators to be nullptrs, causing memory errors.
837         for (Instruction *CurrentBranch : WorkList) {
838           BasicBlock *Succ = CurrentBranch->getSuccessor(0);
839           if (Succ == BB)
840             Succ = CurrentBranch->getSuccessor(1);
841           ReplaceInstWithInst(CurrentBranch, BranchInst::Create(Succ));
842         }
843 
844         DTU.applyUpdates(TreeUpdates);
845         DTU.flush();
846       }
847       LLVM_DEBUG(
848           dbgs() << "Sucessfully peeled " << FC0.PP.PeelCount
849                  << " iterations from the first loop.\n"
850                     "Both Loops have the same number of iterations now.\n");
851     }
852   }
853 
854   /// Walk each set of control flow equivalent fusion candidates and attempt to
855   /// fuse them. This does a single linear traversal of all candidates in the
856   /// set. The conditions for legal fusion are checked at this point. If a pair
857   /// of fusion candidates passes all legality checks, they are fused together
858   /// and a new fusion candidate is created and added to the FusionCandidateSet.
859   /// The original fusion candidates are then removed, as they are no longer
860   /// valid.
861   bool fuseCandidates() {
862     bool Fused = false;
863     LLVM_DEBUG(printFusionCandidates(FusionCandidates));
864     for (auto &CandidateSet : FusionCandidates) {
865       if (CandidateSet.size() < 2)
866         continue;
867 
868       LLVM_DEBUG(dbgs() << "Attempting fusion on Candidate Set:\n"
869                         << CandidateSet << "\n");
870 
871       for (auto FC0 = CandidateSet.begin(); FC0 != CandidateSet.end(); ++FC0) {
872         assert(!LDT.isRemovedLoop(FC0->L) &&
873                "Should not have removed loops in CandidateSet!");
874         auto FC1 = FC0;
875         for (++FC1; FC1 != CandidateSet.end(); ++FC1) {
876           assert(!LDT.isRemovedLoop(FC1->L) &&
877                  "Should not have removed loops in CandidateSet!");
878 
879           LLVM_DEBUG(dbgs() << "Attempting to fuse candidate \n"; FC0->dump();
880                      dbgs() << " with\n"; FC1->dump(); dbgs() << "\n");
881 
882           FC0->verify();
883           FC1->verify();
884 
885           // Check if the candidates have identical tripcounts (first value of
886           // pair), and if not check the difference in the tripcounts between
887           // the loops (second value of pair). The difference is not equal to
888           // std::nullopt iff the loops iterate a constant number of times, and
889           // have a single exit.
890           std::pair<bool, std::optional<unsigned>> IdenticalTripCountRes =
891               haveIdenticalTripCounts(*FC0, *FC1);
892           bool SameTripCount = IdenticalTripCountRes.first;
893           std::optional<unsigned> TCDifference = IdenticalTripCountRes.second;
894 
895           // Here we are checking that FC0 (the first loop) can be peeled, and
896           // both loops have different tripcounts.
897           if (FC0->AbleToPeel && !SameTripCount && TCDifference) {
898             if (*TCDifference > FusionPeelMaxCount) {
899               LLVM_DEBUG(dbgs()
900                          << "Difference in loop trip counts: " << *TCDifference
901                          << " is greater than maximum peel count specificed: "
902                          << FusionPeelMaxCount << "\n");
903             } else {
904               // Dependent on peeling being performed on the first loop, and
905               // assuming all other conditions for fusion return true.
906               SameTripCount = true;
907             }
908           }
909 
910           if (!SameTripCount) {
911             LLVM_DEBUG(dbgs() << "Fusion candidates do not have identical trip "
912                                  "counts. Not fusing.\n");
913             reportLoopFusion<OptimizationRemarkMissed>(*FC0, *FC1,
914                                                        NonEqualTripCount);
915             continue;
916           }
917 
918           if (!isAdjacent(*FC0, *FC1)) {
919             LLVM_DEBUG(dbgs()
920                        << "Fusion candidates are not adjacent. Not fusing.\n");
921             reportLoopFusion<OptimizationRemarkMissed>(*FC0, *FC1, NonAdjacent);
922             continue;
923           }
924 
925           if ((!FC0->GuardBranch && FC1->GuardBranch) ||
926               (FC0->GuardBranch && !FC1->GuardBranch)) {
927             LLVM_DEBUG(dbgs() << "The one of candidate is guarded while the "
928                                  "another one is not. Not fusing.\n");
929             reportLoopFusion<OptimizationRemarkMissed>(
930                 *FC0, *FC1, OnlySecondCandidateIsGuarded);
931             continue;
932           }
933 
934           // Ensure that FC0 and FC1 have identical guards.
935           // If one (or both) are not guarded, this check is not necessary.
936           if (FC0->GuardBranch && FC1->GuardBranch &&
937               !haveIdenticalGuards(*FC0, *FC1) && !TCDifference) {
938             LLVM_DEBUG(dbgs() << "Fusion candidates do not have identical "
939                                  "guards. Not Fusing.\n");
940             reportLoopFusion<OptimizationRemarkMissed>(*FC0, *FC1,
941                                                        NonIdenticalGuards);
942             continue;
943           }
944 
945           if (FC0->GuardBranch) {
946             assert(FC1->GuardBranch && "Expecting valid FC1 guard branch");
947 
948             if (!isSafeToMoveBefore(*FC0->ExitBlock,
949                                     *FC1->ExitBlock->getFirstNonPHIOrDbg(), DT,
950                                     &PDT, &DI)) {
951               LLVM_DEBUG(dbgs() << "Fusion candidate contains unsafe "
952                                    "instructions in exit block. Not fusing.\n");
953               reportLoopFusion<OptimizationRemarkMissed>(*FC0, *FC1,
954                                                          NonEmptyExitBlock);
955               continue;
956             }
957 
958             if (!isSafeToMoveBefore(
959                     *FC1->GuardBranch->getParent(),
960                     *FC0->GuardBranch->getParent()->getTerminator(), DT, &PDT,
961                     &DI)) {
962               LLVM_DEBUG(dbgs()
963                          << "Fusion candidate contains unsafe "
964                             "instructions in guard block. Not fusing.\n");
965               reportLoopFusion<OptimizationRemarkMissed>(*FC0, *FC1,
966                                                          NonEmptyGuardBlock);
967               continue;
968             }
969           }
970 
971           // Check the dependencies across the loops and do not fuse if it would
972           // violate them.
973           if (!dependencesAllowFusion(*FC0, *FC1)) {
974             LLVM_DEBUG(dbgs() << "Memory dependencies do not allow fusion!\n");
975             reportLoopFusion<OptimizationRemarkMissed>(*FC0, *FC1,
976                                                        InvalidDependencies);
977             continue;
978           }
979 
980           // If the second loop has instructions in the pre-header, attempt to
981           // hoist them up to the first loop's pre-header or sink them into the
982           // body of the second loop.
983           SmallVector<Instruction *, 4> SafeToHoist;
984           SmallVector<Instruction *, 4> SafeToSink;
985           // At this point, this is the last remaining legality check.
986           // Which means if we can make this pre-header empty, we can fuse
987           // these loops
988           if (!isEmptyPreheader(*FC1)) {
989             LLVM_DEBUG(dbgs() << "Fusion candidate does not have empty "
990                                  "preheader.\n");
991 
992             // If it is not safe to hoist/sink all instructions in the
993             // pre-header, we cannot fuse these loops.
994             if (!collectMovablePreheaderInsts(*FC0, *FC1, SafeToHoist,
995                                               SafeToSink)) {
996               LLVM_DEBUG(dbgs() << "Could not hoist/sink all instructions in "
997                                    "Fusion Candidate Pre-header.\n"
998                                 << "Not Fusing.\n");
999               reportLoopFusion<OptimizationRemarkMissed>(*FC0, *FC1,
1000                                                          NonEmptyPreheader);
1001               continue;
1002             }
1003           }
1004 
1005           bool BeneficialToFuse = isBeneficialFusion(*FC0, *FC1);
1006           LLVM_DEBUG(dbgs()
1007                      << "\tFusion appears to be "
1008                      << (BeneficialToFuse ? "" : "un") << "profitable!\n");
1009           if (!BeneficialToFuse) {
1010             reportLoopFusion<OptimizationRemarkMissed>(*FC0, *FC1,
1011                                                        FusionNotBeneficial);
1012             continue;
1013           }
1014           // All analysis has completed and has determined that fusion is legal
1015           // and profitable. At this point, start transforming the code and
1016           // perform fusion.
1017 
1018           // Execute the hoist/sink operations on preheader instructions
1019           movePreheaderInsts(*FC0, *FC1, SafeToHoist, SafeToSink);
1020 
1021           LLVM_DEBUG(dbgs() << "\tFusion is performed: " << *FC0 << " and "
1022                             << *FC1 << "\n");
1023 
1024           FusionCandidate FC0Copy = *FC0;
1025           // Peel the loop after determining that fusion is legal. The Loops
1026           // will still be safe to fuse after the peeling is performed.
1027           bool Peel = TCDifference && *TCDifference > 0;
1028           if (Peel)
1029             peelFusionCandidate(FC0Copy, *FC1, *TCDifference);
1030 
1031           // Report fusion to the Optimization Remarks.
1032           // Note this needs to be done *before* performFusion because
1033           // performFusion will change the original loops, making it not
1034           // possible to identify them after fusion is complete.
1035           reportLoopFusion<OptimizationRemark>((Peel ? FC0Copy : *FC0), *FC1,
1036                                                FuseCounter);
1037 
1038           FusionCandidate FusedCand(
1039               performFusion((Peel ? FC0Copy : *FC0), *FC1), DT, &PDT, ORE,
1040               FC0Copy.PP);
1041           FusedCand.verify();
1042           assert(FusedCand.isEligibleForFusion(SE) &&
1043                  "Fused candidate should be eligible for fusion!");
1044 
1045           // Notify the loop-depth-tree that these loops are not valid objects
1046           LDT.removeLoop(FC1->L);
1047 
1048           CandidateSet.erase(FC0);
1049           CandidateSet.erase(FC1);
1050 
1051           auto InsertPos = CandidateSet.insert(FusedCand);
1052 
1053           assert(InsertPos.second &&
1054                  "Unable to insert TargetCandidate in CandidateSet!");
1055 
1056           // Reset FC0 and FC1 the new (fused) candidate. Subsequent iterations
1057           // of the FC1 loop will attempt to fuse the new (fused) loop with the
1058           // remaining candidates in the current candidate set.
1059           FC0 = FC1 = InsertPos.first;
1060 
1061           LLVM_DEBUG(dbgs() << "Candidate Set (after fusion): " << CandidateSet
1062                             << "\n");
1063 
1064           Fused = true;
1065         }
1066       }
1067     }
1068     return Fused;
1069   }
1070 
1071   // Returns true if the instruction \p I can be hoisted to the end of the
1072   // preheader of \p FC0. \p SafeToHoist contains the instructions that are
1073   // known to be safe to hoist. The instructions encountered that cannot be
1074   // hoisted are in \p NotHoisting.
1075   // TODO: Move functionality into CodeMoverUtils
1076   bool canHoistInst(Instruction &I,
1077                     const SmallVector<Instruction *, 4> &SafeToHoist,
1078                     const SmallVector<Instruction *, 4> &NotHoisting,
1079                     const FusionCandidate &FC0) const {
1080     const BasicBlock *FC0PreheaderTarget = FC0.Preheader->getSingleSuccessor();
1081     assert(FC0PreheaderTarget &&
1082            "Expected single successor for loop preheader.");
1083 
1084     for (Use &Op : I.operands()) {
1085       if (auto *OpInst = dyn_cast<Instruction>(Op)) {
1086         bool OpHoisted = is_contained(SafeToHoist, OpInst);
1087         // Check if we have already decided to hoist this operand. In this
1088         // case, it does not dominate FC0 *yet*, but will after we hoist it.
1089         if (!(OpHoisted || DT.dominates(OpInst, FC0PreheaderTarget))) {
1090           return false;
1091         }
1092       }
1093     }
1094 
1095     // PHIs in FC1's header only have FC0 blocks as predecessors. PHIs
1096     // cannot be hoisted and should be sunk to the exit of the fused loop.
1097     if (isa<PHINode>(I))
1098       return false;
1099 
1100     // If this isn't a memory inst, hoisting is safe
1101     if (!I.mayReadOrWriteMemory())
1102       return true;
1103 
1104     LLVM_DEBUG(dbgs() << "Checking if this mem inst can be hoisted.\n");
1105     for (Instruction *NotHoistedInst : NotHoisting) {
1106       if (auto D = DI.depends(&I, NotHoistedInst, true)) {
1107         // Dependency is not read-before-write, write-before-read or
1108         // write-before-write
1109         if (D->isFlow() || D->isAnti() || D->isOutput()) {
1110           LLVM_DEBUG(dbgs() << "Inst depends on an instruction in FC1's "
1111                                "preheader that is not being hoisted.\n");
1112           return false;
1113         }
1114       }
1115     }
1116 
1117     for (Instruction *ReadInst : FC0.MemReads) {
1118       if (auto D = DI.depends(ReadInst, &I, true)) {
1119         // Dependency is not read-before-write
1120         if (D->isAnti()) {
1121           LLVM_DEBUG(dbgs() << "Inst depends on a read instruction in FC0.\n");
1122           return false;
1123         }
1124       }
1125     }
1126 
1127     for (Instruction *WriteInst : FC0.MemWrites) {
1128       if (auto D = DI.depends(WriteInst, &I, true)) {
1129         // Dependency is not write-before-read or write-before-write
1130         if (D->isFlow() || D->isOutput()) {
1131           LLVM_DEBUG(dbgs() << "Inst depends on a write instruction in FC0.\n");
1132           return false;
1133         }
1134       }
1135     }
1136     return true;
1137   }
1138 
1139   // Returns true if the instruction \p I can be sunk to the top of the exit
1140   // block of \p FC1.
1141   // TODO: Move functionality into CodeMoverUtils
1142   bool canSinkInst(Instruction &I, const FusionCandidate &FC1) const {
1143     for (User *U : I.users()) {
1144       if (auto *UI{dyn_cast<Instruction>(U)}) {
1145         // Cannot sink if user in loop
1146         // If FC1 has phi users of this value, we cannot sink it into FC1.
1147         if (FC1.L->contains(UI)) {
1148           // Cannot hoist or sink this instruction. No hoisting/sinking
1149           // should take place, loops should not fuse
1150           return false;
1151         }
1152       }
1153     }
1154 
1155     // If this isn't a memory inst, sinking is safe
1156     if (!I.mayReadOrWriteMemory())
1157       return true;
1158 
1159     for (Instruction *ReadInst : FC1.MemReads) {
1160       if (auto D = DI.depends(&I, ReadInst, true)) {
1161         // Dependency is not write-before-read
1162         if (D->isFlow()) {
1163           LLVM_DEBUG(dbgs() << "Inst depends on a read instruction in FC1.\n");
1164           return false;
1165         }
1166       }
1167     }
1168 
1169     for (Instruction *WriteInst : FC1.MemWrites) {
1170       if (auto D = DI.depends(&I, WriteInst, true)) {
1171         // Dependency is not write-before-write or read-before-write
1172         if (D->isOutput() || D->isAnti()) {
1173           LLVM_DEBUG(dbgs() << "Inst depends on a write instruction in FC1.\n");
1174           return false;
1175         }
1176       }
1177     }
1178 
1179     return true;
1180   }
1181 
1182   /// Collect instructions in the \p FC1 Preheader that can be hoisted
1183   /// to the \p FC0 Preheader or sunk into the \p FC1 Body
1184   bool collectMovablePreheaderInsts(
1185       const FusionCandidate &FC0, const FusionCandidate &FC1,
1186       SmallVector<Instruction *, 4> &SafeToHoist,
1187       SmallVector<Instruction *, 4> &SafeToSink) const {
1188     BasicBlock *FC1Preheader = FC1.Preheader;
1189     // Save the instructions that are not being hoisted, so we know not to hoist
1190     // mem insts that they dominate.
1191     SmallVector<Instruction *, 4> NotHoisting;
1192 
1193     for (Instruction &I : *FC1Preheader) {
1194       // Can't move a branch
1195       if (&I == FC1Preheader->getTerminator())
1196         continue;
1197       // If the instruction has side-effects, give up.
1198       // TODO: The case of mayReadFromMemory we can handle but requires
1199       // additional work with a dependence analysis so for now we give
1200       // up on memory reads.
1201       if (I.mayThrow() || !I.willReturn()) {
1202         LLVM_DEBUG(dbgs() << "Inst: " << I << " may throw or won't return.\n");
1203         return false;
1204       }
1205 
1206       LLVM_DEBUG(dbgs() << "Checking Inst: " << I << "\n");
1207 
1208       if (I.isAtomic() || I.isVolatile()) {
1209         LLVM_DEBUG(
1210             dbgs() << "\tInstruction is volatile or atomic. Cannot move it.\n");
1211         return false;
1212       }
1213 
1214       if (canHoistInst(I, SafeToHoist, NotHoisting, FC0)) {
1215         SafeToHoist.push_back(&I);
1216         LLVM_DEBUG(dbgs() << "\tSafe to hoist.\n");
1217       } else {
1218         LLVM_DEBUG(dbgs() << "\tCould not hoist. Trying to sink...\n");
1219         NotHoisting.push_back(&I);
1220 
1221         if (canSinkInst(I, FC1)) {
1222           SafeToSink.push_back(&I);
1223           LLVM_DEBUG(dbgs() << "\tSafe to sink.\n");
1224         } else {
1225           LLVM_DEBUG(dbgs() << "\tCould not sink.\n");
1226           return false;
1227         }
1228       }
1229     }
1230     LLVM_DEBUG(
1231         dbgs() << "All preheader instructions could be sunk or hoisted!\n");
1232     return true;
1233   }
1234 
1235   /// Rewrite all additive recurrences in a SCEV to use a new loop.
1236   class AddRecLoopReplacer : public SCEVRewriteVisitor<AddRecLoopReplacer> {
1237   public:
1238     AddRecLoopReplacer(ScalarEvolution &SE, const Loop &OldL, const Loop &NewL,
1239                        bool UseMax = true)
1240         : SCEVRewriteVisitor(SE), Valid(true), UseMax(UseMax), OldL(OldL),
1241           NewL(NewL) {}
1242 
1243     const SCEV *visitAddRecExpr(const SCEVAddRecExpr *Expr) {
1244       const Loop *ExprL = Expr->getLoop();
1245       SmallVector<const SCEV *, 2> Operands;
1246       if (ExprL == &OldL) {
1247         append_range(Operands, Expr->operands());
1248         return SE.getAddRecExpr(Operands, &NewL, Expr->getNoWrapFlags());
1249       }
1250 
1251       if (OldL.contains(ExprL)) {
1252         bool Pos = SE.isKnownPositive(Expr->getStepRecurrence(SE));
1253         if (!UseMax || !Pos || !Expr->isAffine()) {
1254           Valid = false;
1255           return Expr;
1256         }
1257         return visit(Expr->getStart());
1258       }
1259 
1260       for (const SCEV *Op : Expr->operands())
1261         Operands.push_back(visit(Op));
1262       return SE.getAddRecExpr(Operands, ExprL, Expr->getNoWrapFlags());
1263     }
1264 
1265     bool wasValidSCEV() const { return Valid; }
1266 
1267   private:
1268     bool Valid, UseMax;
1269     const Loop &OldL, &NewL;
1270   };
1271 
1272   /// Return false if the access functions of \p I0 and \p I1 could cause
1273   /// a negative dependence.
1274   bool accessDiffIsPositive(const Loop &L0, const Loop &L1, Instruction &I0,
1275                             Instruction &I1, bool EqualIsInvalid) {
1276     Value *Ptr0 = getLoadStorePointerOperand(&I0);
1277     Value *Ptr1 = getLoadStorePointerOperand(&I1);
1278     if (!Ptr0 || !Ptr1)
1279       return false;
1280 
1281     const SCEV *SCEVPtr0 = SE.getSCEVAtScope(Ptr0, &L0);
1282     const SCEV *SCEVPtr1 = SE.getSCEVAtScope(Ptr1, &L1);
1283 #ifndef NDEBUG
1284     if (VerboseFusionDebugging)
1285       LLVM_DEBUG(dbgs() << "    Access function check: " << *SCEVPtr0 << " vs "
1286                         << *SCEVPtr1 << "\n");
1287 #endif
1288     AddRecLoopReplacer Rewriter(SE, L0, L1);
1289     SCEVPtr0 = Rewriter.visit(SCEVPtr0);
1290 #ifndef NDEBUG
1291     if (VerboseFusionDebugging)
1292       LLVM_DEBUG(dbgs() << "    Access function after rewrite: " << *SCEVPtr0
1293                         << " [Valid: " << Rewriter.wasValidSCEV() << "]\n");
1294 #endif
1295     if (!Rewriter.wasValidSCEV())
1296       return false;
1297 
1298     // TODO: isKnownPredicate doesnt work well when one SCEV is loop carried (by
1299     //       L0) and the other is not. We could check if it is monotone and test
1300     //       the beginning and end value instead.
1301 
1302     BasicBlock *L0Header = L0.getHeader();
1303     auto HasNonLinearDominanceRelation = [&](const SCEV *S) {
1304       const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(S);
1305       if (!AddRec)
1306         return false;
1307       return !DT.dominates(L0Header, AddRec->getLoop()->getHeader()) &&
1308              !DT.dominates(AddRec->getLoop()->getHeader(), L0Header);
1309     };
1310     if (SCEVExprContains(SCEVPtr1, HasNonLinearDominanceRelation))
1311       return false;
1312 
1313     ICmpInst::Predicate Pred =
1314         EqualIsInvalid ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_SGE;
1315     bool IsAlwaysGE = SE.isKnownPredicate(Pred, SCEVPtr0, SCEVPtr1);
1316 #ifndef NDEBUG
1317     if (VerboseFusionDebugging)
1318       LLVM_DEBUG(dbgs() << "    Relation: " << *SCEVPtr0
1319                         << (IsAlwaysGE ? "  >=  " : "  may <  ") << *SCEVPtr1
1320                         << "\n");
1321 #endif
1322     return IsAlwaysGE;
1323   }
1324 
1325   /// Return true if the dependences between @p I0 (in @p L0) and @p I1 (in
1326   /// @p L1) allow loop fusion of @p L0 and @p L1. The dependence analyses
1327   /// specified by @p DepChoice are used to determine this.
1328   bool dependencesAllowFusion(const FusionCandidate &FC0,
1329                               const FusionCandidate &FC1, Instruction &I0,
1330                               Instruction &I1, bool AnyDep,
1331                               FusionDependenceAnalysisChoice DepChoice) {
1332 #ifndef NDEBUG
1333     if (VerboseFusionDebugging) {
1334       LLVM_DEBUG(dbgs() << "Check dep: " << I0 << " vs " << I1 << " : "
1335                         << DepChoice << "\n");
1336     }
1337 #endif
1338     switch (DepChoice) {
1339     case FUSION_DEPENDENCE_ANALYSIS_SCEV:
1340       return accessDiffIsPositive(*FC0.L, *FC1.L, I0, I1, AnyDep);
1341     case FUSION_DEPENDENCE_ANALYSIS_DA: {
1342       auto DepResult = DI.depends(&I0, &I1, true);
1343       if (!DepResult)
1344         return true;
1345 #ifndef NDEBUG
1346       if (VerboseFusionDebugging) {
1347         LLVM_DEBUG(dbgs() << "DA res: "; DepResult->dump(dbgs());
1348                    dbgs() << " [#l: " << DepResult->getLevels() << "][Ordered: "
1349                           << (DepResult->isOrdered() ? "true" : "false")
1350                           << "]\n");
1351         LLVM_DEBUG(dbgs() << "DepResult Levels: " << DepResult->getLevels()
1352                           << "\n");
1353       }
1354 #endif
1355 
1356       if (DepResult->getNextPredecessor() || DepResult->getNextSuccessor())
1357         LLVM_DEBUG(
1358             dbgs() << "TODO: Implement pred/succ dependence handling!\n");
1359 
1360       // TODO: Can we actually use the dependence info analysis here?
1361       return false;
1362     }
1363 
1364     case FUSION_DEPENDENCE_ANALYSIS_ALL:
1365       return dependencesAllowFusion(FC0, FC1, I0, I1, AnyDep,
1366                                     FUSION_DEPENDENCE_ANALYSIS_SCEV) ||
1367              dependencesAllowFusion(FC0, FC1, I0, I1, AnyDep,
1368                                     FUSION_DEPENDENCE_ANALYSIS_DA);
1369     }
1370 
1371     llvm_unreachable("Unknown fusion dependence analysis choice!");
1372   }
1373 
1374   /// Perform a dependence check and return if @p FC0 and @p FC1 can be fused.
1375   bool dependencesAllowFusion(const FusionCandidate &FC0,
1376                               const FusionCandidate &FC1) {
1377     LLVM_DEBUG(dbgs() << "Check if " << FC0 << " can be fused with " << FC1
1378                       << "\n");
1379     assert(FC0.L->getLoopDepth() == FC1.L->getLoopDepth());
1380     assert(DT.dominates(FC0.getEntryBlock(), FC1.getEntryBlock()));
1381 
1382     for (Instruction *WriteL0 : FC0.MemWrites) {
1383       for (Instruction *WriteL1 : FC1.MemWrites)
1384         if (!dependencesAllowFusion(FC0, FC1, *WriteL0, *WriteL1,
1385                                     /* AnyDep */ false,
1386                                     FusionDependenceAnalysis)) {
1387           InvalidDependencies++;
1388           return false;
1389         }
1390       for (Instruction *ReadL1 : FC1.MemReads)
1391         if (!dependencesAllowFusion(FC0, FC1, *WriteL0, *ReadL1,
1392                                     /* AnyDep */ false,
1393                                     FusionDependenceAnalysis)) {
1394           InvalidDependencies++;
1395           return false;
1396         }
1397     }
1398 
1399     for (Instruction *WriteL1 : FC1.MemWrites) {
1400       for (Instruction *WriteL0 : FC0.MemWrites)
1401         if (!dependencesAllowFusion(FC0, FC1, *WriteL0, *WriteL1,
1402                                     /* AnyDep */ false,
1403                                     FusionDependenceAnalysis)) {
1404           InvalidDependencies++;
1405           return false;
1406         }
1407       for (Instruction *ReadL0 : FC0.MemReads)
1408         if (!dependencesAllowFusion(FC0, FC1, *ReadL0, *WriteL1,
1409                                     /* AnyDep */ false,
1410                                     FusionDependenceAnalysis)) {
1411           InvalidDependencies++;
1412           return false;
1413         }
1414     }
1415 
1416     // Walk through all uses in FC1. For each use, find the reaching def. If the
1417     // def is located in FC0 then it is is not safe to fuse.
1418     for (BasicBlock *BB : FC1.L->blocks())
1419       for (Instruction &I : *BB)
1420         for (auto &Op : I.operands())
1421           if (Instruction *Def = dyn_cast<Instruction>(Op))
1422             if (FC0.L->contains(Def->getParent())) {
1423               InvalidDependencies++;
1424               return false;
1425             }
1426 
1427     return true;
1428   }
1429 
1430   /// Determine if two fusion candidates are adjacent in the CFG.
1431   ///
1432   /// This method will determine if there are additional basic blocks in the CFG
1433   /// between the exit of \p FC0 and the entry of \p FC1.
1434   /// If the two candidates are guarded loops, then it checks whether the
1435   /// non-loop successor of the \p FC0 guard branch is the entry block of \p
1436   /// FC1. If not, then the loops are not adjacent. If the two candidates are
1437   /// not guarded loops, then it checks whether the exit block of \p FC0 is the
1438   /// preheader of \p FC1.
1439   bool isAdjacent(const FusionCandidate &FC0,
1440                   const FusionCandidate &FC1) const {
1441     // If the successor of the guard branch is FC1, then the loops are adjacent
1442     if (FC0.GuardBranch)
1443       return FC0.getNonLoopBlock() == FC1.getEntryBlock();
1444     else
1445       return FC0.ExitBlock == FC1.getEntryBlock();
1446   }
1447 
1448   bool isEmptyPreheader(const FusionCandidate &FC) const {
1449     return FC.Preheader->size() == 1;
1450   }
1451 
1452   /// Hoist \p FC1 Preheader instructions to \p FC0 Preheader
1453   /// and sink others into the body of \p FC1.
1454   void movePreheaderInsts(const FusionCandidate &FC0,
1455                           const FusionCandidate &FC1,
1456                           SmallVector<Instruction *, 4> &HoistInsts,
1457                           SmallVector<Instruction *, 4> &SinkInsts) const {
1458     // All preheader instructions except the branch must be hoisted or sunk
1459     assert(HoistInsts.size() + SinkInsts.size() == FC1.Preheader->size() - 1 &&
1460            "Attempting to sink and hoist preheader instructions, but not all "
1461            "the preheader instructions are accounted for.");
1462 
1463     NumHoistedInsts += HoistInsts.size();
1464     NumSunkInsts += SinkInsts.size();
1465 
1466     LLVM_DEBUG(if (VerboseFusionDebugging) {
1467       if (!HoistInsts.empty())
1468         dbgs() << "Hoisting: \n";
1469       for (Instruction *I : HoistInsts)
1470         dbgs() << *I << "\n";
1471       if (!SinkInsts.empty())
1472         dbgs() << "Sinking: \n";
1473       for (Instruction *I : SinkInsts)
1474         dbgs() << *I << "\n";
1475     });
1476 
1477     for (Instruction *I : HoistInsts) {
1478       assert(I->getParent() == FC1.Preheader);
1479       I->moveBefore(FC0.Preheader->getTerminator());
1480     }
1481     // insert instructions in reverse order to maintain dominance relationship
1482     for (Instruction *I : reverse(SinkInsts)) {
1483       assert(I->getParent() == FC1.Preheader);
1484       I->moveBefore(&*FC1.ExitBlock->getFirstInsertionPt());
1485     }
1486   }
1487 
1488   /// Determine if two fusion candidates have identical guards
1489   ///
1490   /// This method will determine if two fusion candidates have the same guards.
1491   /// The guards are considered the same if:
1492   ///   1. The instructions to compute the condition used in the compare are
1493   ///      identical.
1494   ///   2. The successors of the guard have the same flow into/around the loop.
1495   /// If the compare instructions are identical, then the first successor of the
1496   /// guard must go to the same place (either the preheader of the loop or the
1497   /// NonLoopBlock). In other words, the the first successor of both loops must
1498   /// both go into the loop (i.e., the preheader) or go around the loop (i.e.,
1499   /// the NonLoopBlock). The same must be true for the second successor.
1500   bool haveIdenticalGuards(const FusionCandidate &FC0,
1501                            const FusionCandidate &FC1) const {
1502     assert(FC0.GuardBranch && FC1.GuardBranch &&
1503            "Expecting FC0 and FC1 to be guarded loops.");
1504 
1505     if (auto FC0CmpInst =
1506             dyn_cast<Instruction>(FC0.GuardBranch->getCondition()))
1507       if (auto FC1CmpInst =
1508               dyn_cast<Instruction>(FC1.GuardBranch->getCondition()))
1509         if (!FC0CmpInst->isIdenticalTo(FC1CmpInst))
1510           return false;
1511 
1512     // The compare instructions are identical.
1513     // Now make sure the successor of the guards have the same flow into/around
1514     // the loop
1515     if (FC0.GuardBranch->getSuccessor(0) == FC0.Preheader)
1516       return (FC1.GuardBranch->getSuccessor(0) == FC1.Preheader);
1517     else
1518       return (FC1.GuardBranch->getSuccessor(1) == FC1.Preheader);
1519   }
1520 
1521   /// Modify the latch branch of FC to be unconditional since successors of the
1522   /// branch are the same.
1523   void simplifyLatchBranch(const FusionCandidate &FC) const {
1524     BranchInst *FCLatchBranch = dyn_cast<BranchInst>(FC.Latch->getTerminator());
1525     if (FCLatchBranch) {
1526       assert(FCLatchBranch->isConditional() &&
1527              FCLatchBranch->getSuccessor(0) == FCLatchBranch->getSuccessor(1) &&
1528              "Expecting the two successors of FCLatchBranch to be the same");
1529       BranchInst *NewBranch =
1530           BranchInst::Create(FCLatchBranch->getSuccessor(0));
1531       ReplaceInstWithInst(FCLatchBranch, NewBranch);
1532     }
1533   }
1534 
1535   /// Move instructions from FC0.Latch to FC1.Latch. If FC0.Latch has an unique
1536   /// successor, then merge FC0.Latch with its unique successor.
1537   void mergeLatch(const FusionCandidate &FC0, const FusionCandidate &FC1) {
1538     moveInstructionsToTheBeginning(*FC0.Latch, *FC1.Latch, DT, PDT, DI);
1539     if (BasicBlock *Succ = FC0.Latch->getUniqueSuccessor()) {
1540       MergeBlockIntoPredecessor(Succ, &DTU, &LI);
1541       DTU.flush();
1542     }
1543   }
1544 
1545   /// Fuse two fusion candidates, creating a new fused loop.
1546   ///
1547   /// This method contains the mechanics of fusing two loops, represented by \p
1548   /// FC0 and \p FC1. It is assumed that \p FC0 dominates \p FC1 and \p FC1
1549   /// postdominates \p FC0 (making them control flow equivalent). It also
1550   /// assumes that the other conditions for fusion have been met: adjacent,
1551   /// identical trip counts, and no negative distance dependencies exist that
1552   /// would prevent fusion. Thus, there is no checking for these conditions in
1553   /// this method.
1554   ///
1555   /// Fusion is performed by rewiring the CFG to update successor blocks of the
1556   /// components of tho loop. Specifically, the following changes are done:
1557   ///
1558   ///   1. The preheader of \p FC1 is removed as it is no longer necessary
1559   ///   (because it is currently only a single statement block).
1560   ///   2. The latch of \p FC0 is modified to jump to the header of \p FC1.
1561   ///   3. The latch of \p FC1 i modified to jump to the header of \p FC0.
1562   ///   4. All blocks from \p FC1 are removed from FC1 and added to FC0.
1563   ///
1564   /// All of these modifications are done with dominator tree updates, thus
1565   /// keeping the dominator (and post dominator) information up-to-date.
1566   ///
1567   /// This can be improved in the future by actually merging blocks during
1568   /// fusion. For example, the preheader of \p FC1 can be merged with the
1569   /// preheader of \p FC0. This would allow loops with more than a single
1570   /// statement in the preheader to be fused. Similarly, the latch blocks of the
1571   /// two loops could also be fused into a single block. This will require
1572   /// analysis to prove it is safe to move the contents of the block past
1573   /// existing code, which currently has not been implemented.
1574   Loop *performFusion(const FusionCandidate &FC0, const FusionCandidate &FC1) {
1575     assert(FC0.isValid() && FC1.isValid() &&
1576            "Expecting valid fusion candidates");
1577 
1578     LLVM_DEBUG(dbgs() << "Fusion Candidate 0: \n"; FC0.dump();
1579                dbgs() << "Fusion Candidate 1: \n"; FC1.dump(););
1580 
1581     // Move instructions from the preheader of FC1 to the end of the preheader
1582     // of FC0.
1583     moveInstructionsToTheEnd(*FC1.Preheader, *FC0.Preheader, DT, PDT, DI);
1584 
1585     // Fusing guarded loops is handled slightly differently than non-guarded
1586     // loops and has been broken out into a separate method instead of trying to
1587     // intersperse the logic within a single method.
1588     if (FC0.GuardBranch)
1589       return fuseGuardedLoops(FC0, FC1);
1590 
1591     assert(FC1.Preheader ==
1592            (FC0.Peeled ? FC0.ExitBlock->getUniqueSuccessor() : FC0.ExitBlock));
1593     assert(FC1.Preheader->size() == 1 &&
1594            FC1.Preheader->getSingleSuccessor() == FC1.Header);
1595 
1596     // Remember the phi nodes originally in the header of FC0 in order to rewire
1597     // them later. However, this is only necessary if the new loop carried
1598     // values might not dominate the exiting branch. While we do not generally
1599     // test if this is the case but simply insert intermediate phi nodes, we
1600     // need to make sure these intermediate phi nodes have different
1601     // predecessors. To this end, we filter the special case where the exiting
1602     // block is the latch block of the first loop. Nothing needs to be done
1603     // anyway as all loop carried values dominate the latch and thereby also the
1604     // exiting branch.
1605     SmallVector<PHINode *, 8> OriginalFC0PHIs;
1606     if (FC0.ExitingBlock != FC0.Latch)
1607       for (PHINode &PHI : FC0.Header->phis())
1608         OriginalFC0PHIs.push_back(&PHI);
1609 
1610     // Replace incoming blocks for header PHIs first.
1611     FC1.Preheader->replaceSuccessorsPhiUsesWith(FC0.Preheader);
1612     FC0.Latch->replaceSuccessorsPhiUsesWith(FC1.Latch);
1613 
1614     // Then modify the control flow and update DT and PDT.
1615     SmallVector<DominatorTree::UpdateType, 8> TreeUpdates;
1616 
1617     // The old exiting block of the first loop (FC0) has to jump to the header
1618     // of the second as we need to execute the code in the second header block
1619     // regardless of the trip count. That is, if the trip count is 0, so the
1620     // back edge is never taken, we still have to execute both loop headers,
1621     // especially (but not only!) if the second is a do-while style loop.
1622     // However, doing so might invalidate the phi nodes of the first loop as
1623     // the new values do only need to dominate their latch and not the exiting
1624     // predicate. To remedy this potential problem we always introduce phi
1625     // nodes in the header of the second loop later that select the loop carried
1626     // value, if the second header was reached through an old latch of the
1627     // first, or undef otherwise. This is sound as exiting the first implies the
1628     // second will exit too, __without__ taking the back-edge. [Their
1629     // trip-counts are equal after all.
1630     // KB: Would this sequence be simpler to just just make FC0.ExitingBlock go
1631     // to FC1.Header? I think this is basically what the three sequences are
1632     // trying to accomplish; however, doing this directly in the CFG may mean
1633     // the DT/PDT becomes invalid
1634     if (!FC0.Peeled) {
1635       FC0.ExitingBlock->getTerminator()->replaceUsesOfWith(FC1.Preheader,
1636                                                            FC1.Header);
1637       TreeUpdates.emplace_back(DominatorTree::UpdateType(
1638           DominatorTree::Delete, FC0.ExitingBlock, FC1.Preheader));
1639       TreeUpdates.emplace_back(DominatorTree::UpdateType(
1640           DominatorTree::Insert, FC0.ExitingBlock, FC1.Header));
1641     } else {
1642       TreeUpdates.emplace_back(DominatorTree::UpdateType(
1643           DominatorTree::Delete, FC0.ExitBlock, FC1.Preheader));
1644 
1645       // Remove the ExitBlock of the first Loop (also not needed)
1646       FC0.ExitingBlock->getTerminator()->replaceUsesOfWith(FC0.ExitBlock,
1647                                                            FC1.Header);
1648       TreeUpdates.emplace_back(DominatorTree::UpdateType(
1649           DominatorTree::Delete, FC0.ExitingBlock, FC0.ExitBlock));
1650       FC0.ExitBlock->getTerminator()->eraseFromParent();
1651       TreeUpdates.emplace_back(DominatorTree::UpdateType(
1652           DominatorTree::Insert, FC0.ExitingBlock, FC1.Header));
1653       new UnreachableInst(FC0.ExitBlock->getContext(), FC0.ExitBlock);
1654     }
1655 
1656     // The pre-header of L1 is not necessary anymore.
1657     assert(pred_empty(FC1.Preheader));
1658     FC1.Preheader->getTerminator()->eraseFromParent();
1659     new UnreachableInst(FC1.Preheader->getContext(), FC1.Preheader);
1660     TreeUpdates.emplace_back(DominatorTree::UpdateType(
1661         DominatorTree::Delete, FC1.Preheader, FC1.Header));
1662 
1663     // Moves the phi nodes from the second to the first loops header block.
1664     while (PHINode *PHI = dyn_cast<PHINode>(&FC1.Header->front())) {
1665       if (SE.isSCEVable(PHI->getType()))
1666         SE.forgetValue(PHI);
1667       if (PHI->hasNUsesOrMore(1))
1668         PHI->moveBefore(&*FC0.Header->getFirstInsertionPt());
1669       else
1670         PHI->eraseFromParent();
1671     }
1672 
1673     // Introduce new phi nodes in the second loop header to ensure
1674     // exiting the first and jumping to the header of the second does not break
1675     // the SSA property of the phis originally in the first loop. See also the
1676     // comment above.
1677     Instruction *L1HeaderIP = &FC1.Header->front();
1678     for (PHINode *LCPHI : OriginalFC0PHIs) {
1679       int L1LatchBBIdx = LCPHI->getBasicBlockIndex(FC1.Latch);
1680       assert(L1LatchBBIdx >= 0 &&
1681              "Expected loop carried value to be rewired at this point!");
1682 
1683       Value *LCV = LCPHI->getIncomingValue(L1LatchBBIdx);
1684 
1685       PHINode *L1HeaderPHI = PHINode::Create(
1686           LCV->getType(), 2, LCPHI->getName() + ".afterFC0", L1HeaderIP);
1687       L1HeaderPHI->addIncoming(LCV, FC0.Latch);
1688       L1HeaderPHI->addIncoming(UndefValue::get(LCV->getType()),
1689                                FC0.ExitingBlock);
1690 
1691       LCPHI->setIncomingValue(L1LatchBBIdx, L1HeaderPHI);
1692     }
1693 
1694     // Replace latch terminator destinations.
1695     FC0.Latch->getTerminator()->replaceUsesOfWith(FC0.Header, FC1.Header);
1696     FC1.Latch->getTerminator()->replaceUsesOfWith(FC1.Header, FC0.Header);
1697 
1698     // Modify the latch branch of FC0 to be unconditional as both successors of
1699     // the branch are the same.
1700     simplifyLatchBranch(FC0);
1701 
1702     // If FC0.Latch and FC0.ExitingBlock are the same then we have already
1703     // performed the updates above.
1704     if (FC0.Latch != FC0.ExitingBlock)
1705       TreeUpdates.emplace_back(DominatorTree::UpdateType(
1706           DominatorTree::Insert, FC0.Latch, FC1.Header));
1707 
1708     TreeUpdates.emplace_back(DominatorTree::UpdateType(DominatorTree::Delete,
1709                                                        FC0.Latch, FC0.Header));
1710     TreeUpdates.emplace_back(DominatorTree::UpdateType(DominatorTree::Insert,
1711                                                        FC1.Latch, FC0.Header));
1712     TreeUpdates.emplace_back(DominatorTree::UpdateType(DominatorTree::Delete,
1713                                                        FC1.Latch, FC1.Header));
1714 
1715     // Update DT/PDT
1716     DTU.applyUpdates(TreeUpdates);
1717 
1718     LI.removeBlock(FC1.Preheader);
1719     DTU.deleteBB(FC1.Preheader);
1720     if (FC0.Peeled) {
1721       LI.removeBlock(FC0.ExitBlock);
1722       DTU.deleteBB(FC0.ExitBlock);
1723     }
1724 
1725     DTU.flush();
1726 
1727     // Is there a way to keep SE up-to-date so we don't need to forget the loops
1728     // and rebuild the information in subsequent passes of fusion?
1729     // Note: Need to forget the loops before merging the loop latches, as
1730     // mergeLatch may remove the only block in FC1.
1731     SE.forgetLoop(FC1.L);
1732     SE.forgetLoop(FC0.L);
1733     SE.forgetLoopDispositions();
1734 
1735     // Move instructions from FC0.Latch to FC1.Latch.
1736     // Note: mergeLatch requires an updated DT.
1737     mergeLatch(FC0, FC1);
1738 
1739     // Merge the loops.
1740     SmallVector<BasicBlock *, 8> Blocks(FC1.L->blocks());
1741     for (BasicBlock *BB : Blocks) {
1742       FC0.L->addBlockEntry(BB);
1743       FC1.L->removeBlockFromLoop(BB);
1744       if (LI.getLoopFor(BB) != FC1.L)
1745         continue;
1746       LI.changeLoopFor(BB, FC0.L);
1747     }
1748     while (!FC1.L->isInnermost()) {
1749       const auto &ChildLoopIt = FC1.L->begin();
1750       Loop *ChildLoop = *ChildLoopIt;
1751       FC1.L->removeChildLoop(ChildLoopIt);
1752       FC0.L->addChildLoop(ChildLoop);
1753     }
1754 
1755     // Delete the now empty loop L1.
1756     LI.erase(FC1.L);
1757 
1758 #ifndef NDEBUG
1759     assert(!verifyFunction(*FC0.Header->getParent(), &errs()));
1760     assert(DT.verify(DominatorTree::VerificationLevel::Fast));
1761     assert(PDT.verify());
1762     LI.verify(DT);
1763     SE.verify();
1764 #endif
1765 
1766     LLVM_DEBUG(dbgs() << "Fusion done:\n");
1767 
1768     return FC0.L;
1769   }
1770 
1771   /// Report details on loop fusion opportunities.
1772   ///
1773   /// This template function can be used to report both successful and missed
1774   /// loop fusion opportunities, based on the RemarkKind. The RemarkKind should
1775   /// be one of:
1776   ///   - OptimizationRemarkMissed to report when loop fusion is unsuccessful
1777   ///     given two valid fusion candidates.
1778   ///   - OptimizationRemark to report successful fusion of two fusion
1779   ///     candidates.
1780   /// The remarks will be printed using the form:
1781   ///    <path/filename>:<line number>:<column number>: [<function name>]:
1782   ///       <Cand1 Preheader> and <Cand2 Preheader>: <Stat Description>
1783   template <typename RemarkKind>
1784   void reportLoopFusion(const FusionCandidate &FC0, const FusionCandidate &FC1,
1785                         llvm::Statistic &Stat) {
1786     assert(FC0.Preheader && FC1.Preheader &&
1787            "Expecting valid fusion candidates");
1788     using namespace ore;
1789 #if LLVM_ENABLE_STATS
1790     ++Stat;
1791     ORE.emit(RemarkKind(DEBUG_TYPE, Stat.getName(), FC0.L->getStartLoc(),
1792                         FC0.Preheader)
1793              << "[" << FC0.Preheader->getParent()->getName()
1794              << "]: " << NV("Cand1", StringRef(FC0.Preheader->getName()))
1795              << " and " << NV("Cand2", StringRef(FC1.Preheader->getName()))
1796              << ": " << Stat.getDesc());
1797 #endif
1798   }
1799 
1800   /// Fuse two guarded fusion candidates, creating a new fused loop.
1801   ///
1802   /// Fusing guarded loops is handled much the same way as fusing non-guarded
1803   /// loops. The rewiring of the CFG is slightly different though, because of
1804   /// the presence of the guards around the loops and the exit blocks after the
1805   /// loop body. As such, the new loop is rewired as follows:
1806   ///    1. Keep the guard branch from FC0 and use the non-loop block target
1807   /// from the FC1 guard branch.
1808   ///    2. Remove the exit block from FC0 (this exit block should be empty
1809   /// right now).
1810   ///    3. Remove the guard branch for FC1
1811   ///    4. Remove the preheader for FC1.
1812   /// The exit block successor for the latch of FC0 is updated to be the header
1813   /// of FC1 and the non-exit block successor of the latch of FC1 is updated to
1814   /// be the header of FC0, thus creating the fused loop.
1815   Loop *fuseGuardedLoops(const FusionCandidate &FC0,
1816                          const FusionCandidate &FC1) {
1817     assert(FC0.GuardBranch && FC1.GuardBranch && "Expecting guarded loops");
1818 
1819     BasicBlock *FC0GuardBlock = FC0.GuardBranch->getParent();
1820     BasicBlock *FC1GuardBlock = FC1.GuardBranch->getParent();
1821     BasicBlock *FC0NonLoopBlock = FC0.getNonLoopBlock();
1822     BasicBlock *FC1NonLoopBlock = FC1.getNonLoopBlock();
1823     BasicBlock *FC0ExitBlockSuccessor = FC0.ExitBlock->getUniqueSuccessor();
1824 
1825     // Move instructions from the exit block of FC0 to the beginning of the exit
1826     // block of FC1, in the case that the FC0 loop has not been peeled. In the
1827     // case that FC0 loop is peeled, then move the instructions of the successor
1828     // of the FC0 Exit block to the beginning of the exit block of FC1.
1829     moveInstructionsToTheBeginning(
1830         (FC0.Peeled ? *FC0ExitBlockSuccessor : *FC0.ExitBlock), *FC1.ExitBlock,
1831         DT, PDT, DI);
1832 
1833     // Move instructions from the guard block of FC1 to the end of the guard
1834     // block of FC0.
1835     moveInstructionsToTheEnd(*FC1GuardBlock, *FC0GuardBlock, DT, PDT, DI);
1836 
1837     assert(FC0NonLoopBlock == FC1GuardBlock && "Loops are not adjacent");
1838 
1839     SmallVector<DominatorTree::UpdateType, 8> TreeUpdates;
1840 
1841     ////////////////////////////////////////////////////////////////////////////
1842     // Update the Loop Guard
1843     ////////////////////////////////////////////////////////////////////////////
1844     // The guard for FC0 is updated to guard both FC0 and FC1. This is done by
1845     // changing the NonLoopGuardBlock for FC0 to the NonLoopGuardBlock for FC1.
1846     // Thus, one path from the guard goes to the preheader for FC0 (and thus
1847     // executes the new fused loop) and the other path goes to the NonLoopBlock
1848     // for FC1 (where FC1 guard would have gone if FC1 was not executed).
1849     FC1NonLoopBlock->replacePhiUsesWith(FC1GuardBlock, FC0GuardBlock);
1850     FC0.GuardBranch->replaceUsesOfWith(FC0NonLoopBlock, FC1NonLoopBlock);
1851 
1852     BasicBlock *BBToUpdate = FC0.Peeled ? FC0ExitBlockSuccessor : FC0.ExitBlock;
1853     BBToUpdate->getTerminator()->replaceUsesOfWith(FC1GuardBlock, FC1.Header);
1854 
1855     // The guard of FC1 is not necessary anymore.
1856     FC1.GuardBranch->eraseFromParent();
1857     new UnreachableInst(FC1GuardBlock->getContext(), FC1GuardBlock);
1858 
1859     TreeUpdates.emplace_back(DominatorTree::UpdateType(
1860         DominatorTree::Delete, FC1GuardBlock, FC1.Preheader));
1861     TreeUpdates.emplace_back(DominatorTree::UpdateType(
1862         DominatorTree::Delete, FC1GuardBlock, FC1NonLoopBlock));
1863     TreeUpdates.emplace_back(DominatorTree::UpdateType(
1864         DominatorTree::Delete, FC0GuardBlock, FC1GuardBlock));
1865     TreeUpdates.emplace_back(DominatorTree::UpdateType(
1866         DominatorTree::Insert, FC0GuardBlock, FC1NonLoopBlock));
1867 
1868     if (FC0.Peeled) {
1869       // Remove the Block after the ExitBlock of FC0
1870       TreeUpdates.emplace_back(DominatorTree::UpdateType(
1871           DominatorTree::Delete, FC0ExitBlockSuccessor, FC1GuardBlock));
1872       FC0ExitBlockSuccessor->getTerminator()->eraseFromParent();
1873       new UnreachableInst(FC0ExitBlockSuccessor->getContext(),
1874                           FC0ExitBlockSuccessor);
1875     }
1876 
1877     assert(pred_empty(FC1GuardBlock) &&
1878            "Expecting guard block to have no predecessors");
1879     assert(succ_empty(FC1GuardBlock) &&
1880            "Expecting guard block to have no successors");
1881 
1882     // Remember the phi nodes originally in the header of FC0 in order to rewire
1883     // them later. However, this is only necessary if the new loop carried
1884     // values might not dominate the exiting branch. While we do not generally
1885     // test if this is the case but simply insert intermediate phi nodes, we
1886     // need to make sure these intermediate phi nodes have different
1887     // predecessors. To this end, we filter the special case where the exiting
1888     // block is the latch block of the first loop. Nothing needs to be done
1889     // anyway as all loop carried values dominate the latch and thereby also the
1890     // exiting branch.
1891     // KB: This is no longer necessary because FC0.ExitingBlock == FC0.Latch
1892     // (because the loops are rotated. Thus, nothing will ever be added to
1893     // OriginalFC0PHIs.
1894     SmallVector<PHINode *, 8> OriginalFC0PHIs;
1895     if (FC0.ExitingBlock != FC0.Latch)
1896       for (PHINode &PHI : FC0.Header->phis())
1897         OriginalFC0PHIs.push_back(&PHI);
1898 
1899     assert(OriginalFC0PHIs.empty() && "Expecting OriginalFC0PHIs to be empty!");
1900 
1901     // Replace incoming blocks for header PHIs first.
1902     FC1.Preheader->replaceSuccessorsPhiUsesWith(FC0.Preheader);
1903     FC0.Latch->replaceSuccessorsPhiUsesWith(FC1.Latch);
1904 
1905     // The old exiting block of the first loop (FC0) has to jump to the header
1906     // of the second as we need to execute the code in the second header block
1907     // regardless of the trip count. That is, if the trip count is 0, so the
1908     // back edge is never taken, we still have to execute both loop headers,
1909     // especially (but not only!) if the second is a do-while style loop.
1910     // However, doing so might invalidate the phi nodes of the first loop as
1911     // the new values do only need to dominate their latch and not the exiting
1912     // predicate. To remedy this potential problem we always introduce phi
1913     // nodes in the header of the second loop later that select the loop carried
1914     // value, if the second header was reached through an old latch of the
1915     // first, or undef otherwise. This is sound as exiting the first implies the
1916     // second will exit too, __without__ taking the back-edge (their
1917     // trip-counts are equal after all).
1918     FC0.ExitingBlock->getTerminator()->replaceUsesOfWith(FC0.ExitBlock,
1919                                                          FC1.Header);
1920 
1921     TreeUpdates.emplace_back(DominatorTree::UpdateType(
1922         DominatorTree::Delete, FC0.ExitingBlock, FC0.ExitBlock));
1923     TreeUpdates.emplace_back(DominatorTree::UpdateType(
1924         DominatorTree::Insert, FC0.ExitingBlock, FC1.Header));
1925 
1926     // Remove FC0 Exit Block
1927     // The exit block for FC0 is no longer needed since control will flow
1928     // directly to the header of FC1. Since it is an empty block, it can be
1929     // removed at this point.
1930     // TODO: In the future, we can handle non-empty exit blocks my merging any
1931     // instructions from FC0 exit block into FC1 exit block prior to removing
1932     // the block.
1933     assert(pred_empty(FC0.ExitBlock) && "Expecting exit block to be empty");
1934     FC0.ExitBlock->getTerminator()->eraseFromParent();
1935     new UnreachableInst(FC0.ExitBlock->getContext(), FC0.ExitBlock);
1936 
1937     // Remove FC1 Preheader
1938     // The pre-header of L1 is not necessary anymore.
1939     assert(pred_empty(FC1.Preheader));
1940     FC1.Preheader->getTerminator()->eraseFromParent();
1941     new UnreachableInst(FC1.Preheader->getContext(), FC1.Preheader);
1942     TreeUpdates.emplace_back(DominatorTree::UpdateType(
1943         DominatorTree::Delete, FC1.Preheader, FC1.Header));
1944 
1945     // Moves the phi nodes from the second to the first loops header block.
1946     while (PHINode *PHI = dyn_cast<PHINode>(&FC1.Header->front())) {
1947       if (SE.isSCEVable(PHI->getType()))
1948         SE.forgetValue(PHI);
1949       if (PHI->hasNUsesOrMore(1))
1950         PHI->moveBefore(&*FC0.Header->getFirstInsertionPt());
1951       else
1952         PHI->eraseFromParent();
1953     }
1954 
1955     // Introduce new phi nodes in the second loop header to ensure
1956     // exiting the first and jumping to the header of the second does not break
1957     // the SSA property of the phis originally in the first loop. See also the
1958     // comment above.
1959     Instruction *L1HeaderIP = &FC1.Header->front();
1960     for (PHINode *LCPHI : OriginalFC0PHIs) {
1961       int L1LatchBBIdx = LCPHI->getBasicBlockIndex(FC1.Latch);
1962       assert(L1LatchBBIdx >= 0 &&
1963              "Expected loop carried value to be rewired at this point!");
1964 
1965       Value *LCV = LCPHI->getIncomingValue(L1LatchBBIdx);
1966 
1967       PHINode *L1HeaderPHI = PHINode::Create(
1968           LCV->getType(), 2, LCPHI->getName() + ".afterFC0", L1HeaderIP);
1969       L1HeaderPHI->addIncoming(LCV, FC0.Latch);
1970       L1HeaderPHI->addIncoming(UndefValue::get(LCV->getType()),
1971                                FC0.ExitingBlock);
1972 
1973       LCPHI->setIncomingValue(L1LatchBBIdx, L1HeaderPHI);
1974     }
1975 
1976     // Update the latches
1977 
1978     // Replace latch terminator destinations.
1979     FC0.Latch->getTerminator()->replaceUsesOfWith(FC0.Header, FC1.Header);
1980     FC1.Latch->getTerminator()->replaceUsesOfWith(FC1.Header, FC0.Header);
1981 
1982     // Modify the latch branch of FC0 to be unconditional as both successors of
1983     // the branch are the same.
1984     simplifyLatchBranch(FC0);
1985 
1986     // If FC0.Latch and FC0.ExitingBlock are the same then we have already
1987     // performed the updates above.
1988     if (FC0.Latch != FC0.ExitingBlock)
1989       TreeUpdates.emplace_back(DominatorTree::UpdateType(
1990           DominatorTree::Insert, FC0.Latch, FC1.Header));
1991 
1992     TreeUpdates.emplace_back(DominatorTree::UpdateType(DominatorTree::Delete,
1993                                                        FC0.Latch, FC0.Header));
1994     TreeUpdates.emplace_back(DominatorTree::UpdateType(DominatorTree::Insert,
1995                                                        FC1.Latch, FC0.Header));
1996     TreeUpdates.emplace_back(DominatorTree::UpdateType(DominatorTree::Delete,
1997                                                        FC1.Latch, FC1.Header));
1998 
1999     // All done
2000     // Apply the updates to the Dominator Tree and cleanup.
2001 
2002     assert(succ_empty(FC1GuardBlock) && "FC1GuardBlock has successors!!");
2003     assert(pred_empty(FC1GuardBlock) && "FC1GuardBlock has predecessors!!");
2004 
2005     // Update DT/PDT
2006     DTU.applyUpdates(TreeUpdates);
2007 
2008     LI.removeBlock(FC1GuardBlock);
2009     LI.removeBlock(FC1.Preheader);
2010     LI.removeBlock(FC0.ExitBlock);
2011     if (FC0.Peeled) {
2012       LI.removeBlock(FC0ExitBlockSuccessor);
2013       DTU.deleteBB(FC0ExitBlockSuccessor);
2014     }
2015     DTU.deleteBB(FC1GuardBlock);
2016     DTU.deleteBB(FC1.Preheader);
2017     DTU.deleteBB(FC0.ExitBlock);
2018     DTU.flush();
2019 
2020     // Is there a way to keep SE up-to-date so we don't need to forget the loops
2021     // and rebuild the information in subsequent passes of fusion?
2022     // Note: Need to forget the loops before merging the loop latches, as
2023     // mergeLatch may remove the only block in FC1.
2024     SE.forgetLoop(FC1.L);
2025     SE.forgetLoop(FC0.L);
2026     SE.forgetLoopDispositions();
2027 
2028     // Move instructions from FC0.Latch to FC1.Latch.
2029     // Note: mergeLatch requires an updated DT.
2030     mergeLatch(FC0, FC1);
2031 
2032     // Merge the loops.
2033     SmallVector<BasicBlock *, 8> Blocks(FC1.L->blocks());
2034     for (BasicBlock *BB : Blocks) {
2035       FC0.L->addBlockEntry(BB);
2036       FC1.L->removeBlockFromLoop(BB);
2037       if (LI.getLoopFor(BB) != FC1.L)
2038         continue;
2039       LI.changeLoopFor(BB, FC0.L);
2040     }
2041     while (!FC1.L->isInnermost()) {
2042       const auto &ChildLoopIt = FC1.L->begin();
2043       Loop *ChildLoop = *ChildLoopIt;
2044       FC1.L->removeChildLoop(ChildLoopIt);
2045       FC0.L->addChildLoop(ChildLoop);
2046     }
2047 
2048     // Delete the now empty loop L1.
2049     LI.erase(FC1.L);
2050 
2051 #ifndef NDEBUG
2052     assert(!verifyFunction(*FC0.Header->getParent(), &errs()));
2053     assert(DT.verify(DominatorTree::VerificationLevel::Fast));
2054     assert(PDT.verify());
2055     LI.verify(DT);
2056     SE.verify();
2057 #endif
2058 
2059     LLVM_DEBUG(dbgs() << "Fusion done:\n");
2060 
2061     return FC0.L;
2062   }
2063 };
2064 
2065 struct LoopFuseLegacy : public FunctionPass {
2066 
2067   static char ID;
2068 
2069   LoopFuseLegacy() : FunctionPass(ID) {
2070     initializeLoopFuseLegacyPass(*PassRegistry::getPassRegistry());
2071   }
2072 
2073   void getAnalysisUsage(AnalysisUsage &AU) const override {
2074     AU.addRequiredID(LoopSimplifyID);
2075     AU.addRequired<ScalarEvolutionWrapperPass>();
2076     AU.addRequired<LoopInfoWrapperPass>();
2077     AU.addRequired<DominatorTreeWrapperPass>();
2078     AU.addRequired<PostDominatorTreeWrapperPass>();
2079     AU.addRequired<OptimizationRemarkEmitterWrapperPass>();
2080     AU.addRequired<DependenceAnalysisWrapperPass>();
2081     AU.addRequired<AssumptionCacheTracker>();
2082     AU.addRequired<TargetTransformInfoWrapperPass>();
2083 
2084     AU.addPreserved<ScalarEvolutionWrapperPass>();
2085     AU.addPreserved<LoopInfoWrapperPass>();
2086     AU.addPreserved<DominatorTreeWrapperPass>();
2087     AU.addPreserved<PostDominatorTreeWrapperPass>();
2088   }
2089 
2090   bool runOnFunction(Function &F) override {
2091     if (skipFunction(F))
2092       return false;
2093 
2094     auto &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
2095     auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
2096     auto &DI = getAnalysis<DependenceAnalysisWrapperPass>().getDI();
2097     auto &SE = getAnalysis<ScalarEvolutionWrapperPass>().getSE();
2098     auto &PDT = getAnalysis<PostDominatorTreeWrapperPass>().getPostDomTree();
2099     auto &ORE = getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE();
2100     auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
2101     const TargetTransformInfo &TTI =
2102         getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
2103     const DataLayout &DL = F.getParent()->getDataLayout();
2104 
2105     LoopFuser LF(LI, DT, DI, SE, PDT, ORE, DL, AC, TTI);
2106     return LF.fuseLoops(F);
2107   }
2108 };
2109 } // namespace
2110 
2111 PreservedAnalyses LoopFusePass::run(Function &F, FunctionAnalysisManager &AM) {
2112   auto &LI = AM.getResult<LoopAnalysis>(F);
2113   auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
2114   auto &DI = AM.getResult<DependenceAnalysis>(F);
2115   auto &SE = AM.getResult<ScalarEvolutionAnalysis>(F);
2116   auto &PDT = AM.getResult<PostDominatorTreeAnalysis>(F);
2117   auto &ORE = AM.getResult<OptimizationRemarkEmitterAnalysis>(F);
2118   auto &AC = AM.getResult<AssumptionAnalysis>(F);
2119   const TargetTransformInfo &TTI = AM.getResult<TargetIRAnalysis>(F);
2120   const DataLayout &DL = F.getParent()->getDataLayout();
2121 
2122   // Ensure loops are in simplifed form which is a pre-requisite for loop fusion
2123   // pass. Added only for new PM since the legacy PM has already added
2124   // LoopSimplify pass as a dependency.
2125   bool Changed = false;
2126   for (auto &L : LI) {
2127     Changed |=
2128         simplifyLoop(L, &DT, &LI, &SE, &AC, nullptr, false /* PreserveLCSSA */);
2129   }
2130   if (Changed)
2131     PDT.recalculate(F);
2132 
2133   LoopFuser LF(LI, DT, DI, SE, PDT, ORE, DL, AC, TTI);
2134   Changed |= LF.fuseLoops(F);
2135   if (!Changed)
2136     return PreservedAnalyses::all();
2137 
2138   PreservedAnalyses PA;
2139   PA.preserve<DominatorTreeAnalysis>();
2140   PA.preserve<PostDominatorTreeAnalysis>();
2141   PA.preserve<ScalarEvolutionAnalysis>();
2142   PA.preserve<LoopAnalysis>();
2143   return PA;
2144 }
2145 
2146 char LoopFuseLegacy::ID = 0;
2147 
2148 INITIALIZE_PASS_BEGIN(LoopFuseLegacy, "loop-fusion", "Loop Fusion", false,
2149                       false)
2150 INITIALIZE_PASS_DEPENDENCY(PostDominatorTreeWrapperPass)
2151 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
2152 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
2153 INITIALIZE_PASS_DEPENDENCY(DependenceAnalysisWrapperPass)
2154 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
2155 INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass)
2156 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
2157 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
2158 INITIALIZE_PASS_END(LoopFuseLegacy, "loop-fusion", "Loop Fusion", false, false)
2159 
2160 FunctionPass *llvm::createLoopFusePass() { return new LoopFuseLegacy(); }
2161