xref: /freebsd/contrib/llvm-project/llvm/lib/Transforms/Scalar/LoopFuse.cpp (revision 0ad011ececb978e22a9bff2acf76633b094f1ff6)
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/Support/CommandLine.h"
61 #include "llvm/Support/Debug.h"
62 #include "llvm/Support/raw_ostream.h"
63 #include "llvm/Transforms/Utils.h"
64 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
65 #include "llvm/Transforms/Utils/CodeMoverUtils.h"
66 #include "llvm/Transforms/Utils/LoopPeel.h"
67 #include "llvm/Transforms/Utils/LoopSimplify.h"
68 
69 using namespace llvm;
70 
71 #define DEBUG_TYPE "loop-fusion"
72 
73 STATISTIC(FuseCounter, "Loops fused");
74 STATISTIC(NumFusionCandidates, "Number of candidates for loop fusion");
75 STATISTIC(InvalidPreheader, "Loop has invalid preheader");
76 STATISTIC(InvalidHeader, "Loop has invalid header");
77 STATISTIC(InvalidExitingBlock, "Loop has invalid exiting blocks");
78 STATISTIC(InvalidExitBlock, "Loop has invalid exit block");
79 STATISTIC(InvalidLatch, "Loop has invalid latch");
80 STATISTIC(InvalidLoop, "Loop is invalid");
81 STATISTIC(AddressTakenBB, "Basic block has address taken");
82 STATISTIC(MayThrowException, "Loop may throw an exception");
83 STATISTIC(ContainsVolatileAccess, "Loop contains a volatile access");
84 STATISTIC(NotSimplifiedForm, "Loop is not in simplified form");
85 STATISTIC(InvalidDependencies, "Dependencies prevent fusion");
86 STATISTIC(UnknownTripCount, "Loop has unknown trip count");
87 STATISTIC(UncomputableTripCount, "SCEV cannot compute trip count of loop");
88 STATISTIC(NonEqualTripCount, "Loop trip counts are not the same");
89 STATISTIC(NonAdjacent, "Loops are not adjacent");
90 STATISTIC(
91     NonEmptyPreheader,
92     "Loop has a non-empty preheader with instructions that cannot be moved");
93 STATISTIC(FusionNotBeneficial, "Fusion is not beneficial");
94 STATISTIC(NonIdenticalGuards, "Candidates have different guards");
95 STATISTIC(NonEmptyExitBlock, "Candidate has a non-empty exit block with "
96                              "instructions that cannot be moved");
97 STATISTIC(NonEmptyGuardBlock, "Candidate has a non-empty guard block with "
98                               "instructions that cannot be moved");
99 STATISTIC(NotRotated, "Candidate is not rotated");
100 STATISTIC(OnlySecondCandidateIsGuarded,
101           "The second candidate is guarded while the first one is not");
102 STATISTIC(NumHoistedInsts, "Number of hoisted preheader instructions.");
103 STATISTIC(NumSunkInsts, "Number of hoisted preheader instructions.");
104 
105 enum FusionDependenceAnalysisChoice {
106   FUSION_DEPENDENCE_ANALYSIS_SCEV,
107   FUSION_DEPENDENCE_ANALYSIS_DA,
108   FUSION_DEPENDENCE_ANALYSIS_ALL,
109 };
110 
111 static cl::opt<FusionDependenceAnalysisChoice> FusionDependenceAnalysis(
112     "loop-fusion-dependence-analysis",
113     cl::desc("Which dependence analysis should loop fusion use?"),
114     cl::values(clEnumValN(FUSION_DEPENDENCE_ANALYSIS_SCEV, "scev",
115                           "Use the scalar evolution interface"),
116                clEnumValN(FUSION_DEPENDENCE_ANALYSIS_DA, "da",
117                           "Use the dependence analysis interface"),
118                clEnumValN(FUSION_DEPENDENCE_ANALYSIS_ALL, "all",
119                           "Use all available analyses")),
120     cl::Hidden, cl::init(FUSION_DEPENDENCE_ANALYSIS_ALL));
121 
122 static cl::opt<unsigned> FusionPeelMaxCount(
123     "loop-fusion-peel-max-count", cl::init(0), cl::Hidden,
124     cl::desc("Max number of iterations to be peeled from a loop, such that "
125              "fusion can take place"));
126 
127 #ifndef NDEBUG
128 static cl::opt<bool>
129     VerboseFusionDebugging("loop-fusion-verbose-debug",
130                            cl::desc("Enable verbose debugging for Loop Fusion"),
131                            cl::Hidden, cl::init(false));
132 #endif
133 
134 namespace {
135 /// This class is used to represent a candidate for loop fusion. When it is
136 /// constructed, it checks the conditions for loop fusion to ensure that it
137 /// represents a valid candidate. It caches several parts of a loop that are
138 /// used throughout loop fusion (e.g., loop preheader, loop header, etc) instead
139 /// of continually querying the underlying Loop to retrieve these values. It is
140 /// assumed these will not change throughout loop fusion.
141 ///
142 /// The invalidate method should be used to indicate that the FusionCandidate is
143 /// no longer a valid candidate for fusion. Similarly, the isValid() method can
144 /// be used to ensure that the FusionCandidate is still valid for fusion.
145 struct FusionCandidate {
146   /// Cache of parts of the loop used throughout loop fusion. These should not
147   /// need to change throughout the analysis and transformation.
148   /// These parts are cached to avoid repeatedly looking up in the Loop class.
149 
150   /// Preheader of the loop this candidate represents
151   BasicBlock *Preheader;
152   /// Header of the loop this candidate represents
153   BasicBlock *Header;
154   /// Blocks in the loop that exit the loop
155   BasicBlock *ExitingBlock;
156   /// The successor block of this loop (where the exiting blocks go to)
157   BasicBlock *ExitBlock;
158   /// Latch of the loop
159   BasicBlock *Latch;
160   /// The loop that this fusion candidate represents
161   Loop *L;
162   /// Vector of instructions in this loop that read from memory
163   SmallVector<Instruction *, 16> MemReads;
164   /// Vector of instructions in this loop that write to memory
165   SmallVector<Instruction *, 16> MemWrites;
166   /// Are all of the members of this fusion candidate still valid
167   bool Valid;
168   /// Guard branch of the loop, if it exists
169   BranchInst *GuardBranch;
170   /// Peeling Paramaters of the Loop.
171   TTI::PeelingPreferences PP;
172   /// Can you Peel this Loop?
173   bool AbleToPeel;
174   /// Has this loop been Peeled
175   bool Peeled;
176 
177   /// Dominator and PostDominator trees are needed for the
178   /// FusionCandidateCompare function, required by FusionCandidateSet to
179   /// determine where the FusionCandidate should be inserted into the set. These
180   /// are used to establish ordering of the FusionCandidates based on dominance.
181   DominatorTree &DT;
182   const PostDominatorTree *PDT;
183 
184   OptimizationRemarkEmitter &ORE;
185 
186   FusionCandidate(Loop *L, DominatorTree &DT, const PostDominatorTree *PDT,
187                   OptimizationRemarkEmitter &ORE, TTI::PeelingPreferences PP)
188       : Preheader(L->getLoopPreheader()), Header(L->getHeader()),
189         ExitingBlock(L->getExitingBlock()), ExitBlock(L->getExitBlock()),
190         Latch(L->getLoopLatch()), L(L), Valid(true),
191         GuardBranch(L->getLoopGuardBranch()), PP(PP), AbleToPeel(canPeel(L)),
192         Peeled(false), DT(DT), PDT(PDT), ORE(ORE) {
193 
194     // Walk over all blocks in the loop and check for conditions that may
195     // prevent fusion. For each block, walk over all instructions and collect
196     // the memory reads and writes If any instructions that prevent fusion are
197     // found, invalidate this object and return.
198     for (BasicBlock *BB : L->blocks()) {
199       if (BB->hasAddressTaken()) {
200         invalidate();
201         reportInvalidCandidate(AddressTakenBB);
202         return;
203       }
204 
205       for (Instruction &I : *BB) {
206         if (I.mayThrow()) {
207           invalidate();
208           reportInvalidCandidate(MayThrowException);
209           return;
210         }
211         if (StoreInst *SI = dyn_cast<StoreInst>(&I)) {
212           if (SI->isVolatile()) {
213             invalidate();
214             reportInvalidCandidate(ContainsVolatileAccess);
215             return;
216           }
217         }
218         if (LoadInst *LI = dyn_cast<LoadInst>(&I)) {
219           if (LI->isVolatile()) {
220             invalidate();
221             reportInvalidCandidate(ContainsVolatileAccess);
222             return;
223           }
224         }
225         if (I.mayWriteToMemory())
226           MemWrites.push_back(&I);
227         if (I.mayReadFromMemory())
228           MemReads.push_back(&I);
229       }
230     }
231   }
232 
233   /// Check if all members of the class are valid.
234   bool isValid() const {
235     return Preheader && Header && ExitingBlock && ExitBlock && Latch && L &&
236            !L->isInvalid() && Valid;
237   }
238 
239   /// Verify that all members are in sync with the Loop object.
240   void verify() const {
241     assert(isValid() && "Candidate is not valid!!");
242     assert(!L->isInvalid() && "Loop is invalid!");
243     assert(Preheader == L->getLoopPreheader() && "Preheader is out of sync");
244     assert(Header == L->getHeader() && "Header is out of sync");
245     assert(ExitingBlock == L->getExitingBlock() &&
246            "Exiting Blocks is out of sync");
247     assert(ExitBlock == L->getExitBlock() && "Exit block is out of sync");
248     assert(Latch == L->getLoopLatch() && "Latch is out of sync");
249   }
250 
251   /// Get the entry block for this fusion candidate.
252   ///
253   /// If this fusion candidate represents a guarded loop, the entry block is the
254   /// loop guard block. If it represents an unguarded loop, the entry block is
255   /// the preheader of the loop.
256   BasicBlock *getEntryBlock() const {
257     if (GuardBranch)
258       return GuardBranch->getParent();
259     else
260       return Preheader;
261   }
262 
263   /// After Peeling the loop is modified quite a bit, hence all of the Blocks
264   /// need to be updated accordingly.
265   void updateAfterPeeling() {
266     Preheader = L->getLoopPreheader();
267     Header = L->getHeader();
268     ExitingBlock = L->getExitingBlock();
269     ExitBlock = L->getExitBlock();
270     Latch = L->getLoopLatch();
271     verify();
272   }
273 
274   /// Given a guarded loop, get the successor of the guard that is not in the
275   /// loop.
276   ///
277   /// This method returns the successor of the loop guard that is not located
278   /// within the loop (i.e., the successor of the guard that is not the
279   /// preheader).
280   /// This method is only valid for guarded loops.
281   BasicBlock *getNonLoopBlock() const {
282     assert(GuardBranch && "Only valid on guarded loops.");
283     assert(GuardBranch->isConditional() &&
284            "Expecting guard to be a conditional branch.");
285     if (Peeled)
286       return GuardBranch->getSuccessor(1);
287     return (GuardBranch->getSuccessor(0) == Preheader)
288                ? GuardBranch->getSuccessor(1)
289                : GuardBranch->getSuccessor(0);
290   }
291 
292 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
293   LLVM_DUMP_METHOD void dump() const {
294     dbgs() << "\tGuardBranch: ";
295     if (GuardBranch)
296       dbgs() << *GuardBranch;
297     else
298       dbgs() << "nullptr";
299     dbgs() << "\n"
300            << (GuardBranch ? GuardBranch->getName() : "nullptr") << "\n"
301            << "\tPreheader: " << (Preheader ? Preheader->getName() : "nullptr")
302            << "\n"
303            << "\tHeader: " << (Header ? Header->getName() : "nullptr") << "\n"
304            << "\tExitingBB: "
305            << (ExitingBlock ? ExitingBlock->getName() : "nullptr") << "\n"
306            << "\tExitBB: " << (ExitBlock ? ExitBlock->getName() : "nullptr")
307            << "\n"
308            << "\tLatch: " << (Latch ? Latch->getName() : "nullptr") << "\n"
309            << "\tEntryBlock: "
310            << (getEntryBlock() ? getEntryBlock()->getName() : "nullptr")
311            << "\n";
312   }
313 #endif
314 
315   /// Determine if a fusion candidate (representing a loop) is eligible for
316   /// fusion. Note that this only checks whether a single loop can be fused - it
317   /// does not check whether it is *legal* to fuse two loops together.
318   bool isEligibleForFusion(ScalarEvolution &SE) const {
319     if (!isValid()) {
320       LLVM_DEBUG(dbgs() << "FC has invalid CFG requirements!\n");
321       if (!Preheader)
322         ++InvalidPreheader;
323       if (!Header)
324         ++InvalidHeader;
325       if (!ExitingBlock)
326         ++InvalidExitingBlock;
327       if (!ExitBlock)
328         ++InvalidExitBlock;
329       if (!Latch)
330         ++InvalidLatch;
331       if (L->isInvalid())
332         ++InvalidLoop;
333 
334       return false;
335     }
336 
337     // Require ScalarEvolution to be able to determine a trip count.
338     if (!SE.hasLoopInvariantBackedgeTakenCount(L)) {
339       LLVM_DEBUG(dbgs() << "Loop " << L->getName()
340                         << " trip count not computable!\n");
341       return reportInvalidCandidate(UnknownTripCount);
342     }
343 
344     if (!L->isLoopSimplifyForm()) {
345       LLVM_DEBUG(dbgs() << "Loop " << L->getName()
346                         << " is not in simplified form!\n");
347       return reportInvalidCandidate(NotSimplifiedForm);
348     }
349 
350     if (!L->isRotatedForm()) {
351       LLVM_DEBUG(dbgs() << "Loop " << L->getName() << " is not rotated!\n");
352       return reportInvalidCandidate(NotRotated);
353     }
354 
355     return true;
356   }
357 
358 private:
359   // This is only used internally for now, to clear the MemWrites and MemReads
360   // list and setting Valid to false. I can't envision other uses of this right
361   // now, since once FusionCandidates are put into the FusionCandidateSet they
362   // are immutable. Thus, any time we need to change/update a FusionCandidate,
363   // we must create a new one and insert it into the FusionCandidateSet to
364   // ensure the FusionCandidateSet remains ordered correctly.
365   void invalidate() {
366     MemWrites.clear();
367     MemReads.clear();
368     Valid = false;
369   }
370 
371   bool reportInvalidCandidate(llvm::Statistic &Stat) const {
372     using namespace ore;
373     assert(L && Preheader && "Fusion candidate not initialized properly!");
374 #if LLVM_ENABLE_STATS
375     ++Stat;
376     ORE.emit(OptimizationRemarkAnalysis(DEBUG_TYPE, Stat.getName(),
377                                         L->getStartLoc(), Preheader)
378              << "[" << Preheader->getParent()->getName() << "]: "
379              << "Loop is not a candidate for fusion: " << Stat.getDesc());
380 #endif
381     return false;
382   }
383 };
384 
385 struct FusionCandidateCompare {
386   /// Comparison functor to sort two Control Flow Equivalent fusion candidates
387   /// into dominance order.
388   /// If LHS dominates RHS and RHS post-dominates LHS, return true;
389   /// If RHS dominates LHS and LHS post-dominates RHS, return false;
390   /// If both LHS and RHS are not dominating each other then, non-strictly
391   /// post dominate check will decide the order of candidates. If RHS
392   /// non-strictly post dominates LHS then, return true. If LHS non-strictly
393   /// post dominates RHS then, return false. If both are non-strictly post
394   /// dominate each other then, level in the post dominator tree will decide
395   /// the order of candidates.
396   bool operator()(const FusionCandidate &LHS,
397                   const FusionCandidate &RHS) const {
398     const DominatorTree *DT = &(LHS.DT);
399 
400     BasicBlock *LHSEntryBlock = LHS.getEntryBlock();
401     BasicBlock *RHSEntryBlock = RHS.getEntryBlock();
402 
403     // Do not save PDT to local variable as it is only used in asserts and thus
404     // will trigger an unused variable warning if building without asserts.
405     assert(DT && LHS.PDT && "Expecting valid dominator tree");
406 
407     // Do this compare first so if LHS == RHS, function returns false.
408     if (DT->dominates(RHSEntryBlock, LHSEntryBlock)) {
409       // RHS dominates LHS
410       // Verify LHS post-dominates RHS
411       assert(LHS.PDT->dominates(LHSEntryBlock, RHSEntryBlock));
412       return false;
413     }
414 
415     if (DT->dominates(LHSEntryBlock, RHSEntryBlock)) {
416       // Verify RHS Postdominates LHS
417       assert(LHS.PDT->dominates(RHSEntryBlock, LHSEntryBlock));
418       return true;
419     }
420 
421     // If two FusionCandidates are in the same level of dominator tree,
422     // they will not dominate each other, but may still be control flow
423     // equivalent. To sort those FusionCandidates, nonStrictlyPostDominate()
424     // function is needed.
425     bool WrongOrder =
426         nonStrictlyPostDominate(LHSEntryBlock, RHSEntryBlock, DT, LHS.PDT);
427     bool RightOrder =
428         nonStrictlyPostDominate(RHSEntryBlock, LHSEntryBlock, DT, LHS.PDT);
429     if (WrongOrder && RightOrder) {
430       // If common predecessor of LHS and RHS post dominates both
431       // FusionCandidates then, Order of FusionCandidate can be
432       // identified by its level in post dominator tree.
433       DomTreeNode *LNode = LHS.PDT->getNode(LHSEntryBlock);
434       DomTreeNode *RNode = LHS.PDT->getNode(RHSEntryBlock);
435       return LNode->getLevel() > RNode->getLevel();
436     } else if (WrongOrder)
437       return false;
438     else if (RightOrder)
439       return true;
440 
441     // If LHS does not non-strict Postdominate RHS and RHS does not non-strict
442     // Postdominate LHS then, there is no dominance relationship between the
443     // two FusionCandidates. Thus, they should not be in the same set together.
444     llvm_unreachable(
445         "No dominance relationship between these fusion candidates!");
446   }
447 };
448 
449 using LoopVector = SmallVector<Loop *, 4>;
450 
451 // Set of Control Flow Equivalent (CFE) Fusion Candidates, sorted in dominance
452 // order. Thus, if FC0 comes *before* FC1 in a FusionCandidateSet, then FC0
453 // dominates FC1 and FC1 post-dominates FC0.
454 // std::set was chosen because we want a sorted data structure with stable
455 // iterators. A subsequent patch to loop fusion will enable fusing non-adjacent
456 // loops by moving intervening code around. When this intervening code contains
457 // loops, those loops will be moved also. The corresponding FusionCandidates
458 // will also need to be moved accordingly. As this is done, having stable
459 // iterators will simplify the logic. Similarly, having an efficient insert that
460 // keeps the FusionCandidateSet sorted will also simplify the implementation.
461 using FusionCandidateSet = std::set<FusionCandidate, FusionCandidateCompare>;
462 using FusionCandidateCollection = SmallVector<FusionCandidateSet, 4>;
463 
464 #if !defined(NDEBUG)
465 static llvm::raw_ostream &operator<<(llvm::raw_ostream &OS,
466                                      const FusionCandidate &FC) {
467   if (FC.isValid())
468     OS << FC.Preheader->getName();
469   else
470     OS << "<Invalid>";
471 
472   return OS;
473 }
474 
475 static llvm::raw_ostream &operator<<(llvm::raw_ostream &OS,
476                                      const FusionCandidateSet &CandSet) {
477   for (const FusionCandidate &FC : CandSet)
478     OS << FC << '\n';
479 
480   return OS;
481 }
482 
483 static void
484 printFusionCandidates(const FusionCandidateCollection &FusionCandidates) {
485   dbgs() << "Fusion Candidates: \n";
486   for (const auto &CandidateSet : FusionCandidates) {
487     dbgs() << "*** Fusion Candidate Set ***\n";
488     dbgs() << CandidateSet;
489     dbgs() << "****************************\n";
490   }
491 }
492 #endif
493 
494 /// Collect all loops in function at the same nest level, starting at the
495 /// outermost level.
496 ///
497 /// This data structure collects all loops at the same nest level for a
498 /// given function (specified by the LoopInfo object). It starts at the
499 /// outermost level.
500 struct LoopDepthTree {
501   using LoopsOnLevelTy = SmallVector<LoopVector, 4>;
502   using iterator = LoopsOnLevelTy::iterator;
503   using const_iterator = LoopsOnLevelTy::const_iterator;
504 
505   LoopDepthTree(LoopInfo &LI) : Depth(1) {
506     if (!LI.empty())
507       LoopsOnLevel.emplace_back(LoopVector(LI.rbegin(), LI.rend()));
508   }
509 
510   /// Test whether a given loop has been removed from the function, and thus is
511   /// no longer valid.
512   bool isRemovedLoop(const Loop *L) const { return RemovedLoops.count(L); }
513 
514   /// Record that a given loop has been removed from the function and is no
515   /// longer valid.
516   void removeLoop(const Loop *L) { RemovedLoops.insert(L); }
517 
518   /// Descend the tree to the next (inner) nesting level
519   void descend() {
520     LoopsOnLevelTy LoopsOnNextLevel;
521 
522     for (const LoopVector &LV : *this)
523       for (Loop *L : LV)
524         if (!isRemovedLoop(L) && L->begin() != L->end())
525           LoopsOnNextLevel.emplace_back(LoopVector(L->begin(), L->end()));
526 
527     LoopsOnLevel = LoopsOnNextLevel;
528     RemovedLoops.clear();
529     Depth++;
530   }
531 
532   bool empty() const { return size() == 0; }
533   size_t size() const { return LoopsOnLevel.size() - RemovedLoops.size(); }
534   unsigned getDepth() const { return Depth; }
535 
536   iterator begin() { return LoopsOnLevel.begin(); }
537   iterator end() { return LoopsOnLevel.end(); }
538   const_iterator begin() const { return LoopsOnLevel.begin(); }
539   const_iterator end() const { return LoopsOnLevel.end(); }
540 
541 private:
542   /// Set of loops that have been removed from the function and are no longer
543   /// valid.
544   SmallPtrSet<const Loop *, 8> RemovedLoops;
545 
546   /// Depth of the current level, starting at 1 (outermost loops).
547   unsigned Depth;
548 
549   /// Vector of loops at the current depth level that have the same parent loop
550   LoopsOnLevelTy LoopsOnLevel;
551 };
552 
553 #ifndef NDEBUG
554 static void printLoopVector(const LoopVector &LV) {
555   dbgs() << "****************************\n";
556   for (auto *L : LV)
557     printLoop(*L, dbgs());
558   dbgs() << "****************************\n";
559 }
560 #endif
561 
562 struct LoopFuser {
563 private:
564   // Sets of control flow equivalent fusion candidates for a given nest level.
565   FusionCandidateCollection FusionCandidates;
566 
567   LoopDepthTree LDT;
568   DomTreeUpdater DTU;
569 
570   LoopInfo &LI;
571   DominatorTree &DT;
572   DependenceInfo &DI;
573   ScalarEvolution &SE;
574   PostDominatorTree &PDT;
575   OptimizationRemarkEmitter &ORE;
576   AssumptionCache &AC;
577   const TargetTransformInfo &TTI;
578 
579 public:
580   LoopFuser(LoopInfo &LI, DominatorTree &DT, DependenceInfo &DI,
581             ScalarEvolution &SE, PostDominatorTree &PDT,
582             OptimizationRemarkEmitter &ORE, const DataLayout &DL,
583             AssumptionCache &AC, const TargetTransformInfo &TTI)
584       : LDT(LI), DTU(DT, PDT, DomTreeUpdater::UpdateStrategy::Lazy), LI(LI),
585         DT(DT), DI(DI), SE(SE), PDT(PDT), ORE(ORE), AC(AC), TTI(TTI) {}
586 
587   /// This is the main entry point for loop fusion. It will traverse the
588   /// specified function and collect candidate loops to fuse, starting at the
589   /// outermost nesting level and working inwards.
590   bool fuseLoops(Function &F) {
591 #ifndef NDEBUG
592     if (VerboseFusionDebugging) {
593       LI.print(dbgs());
594     }
595 #endif
596 
597     LLVM_DEBUG(dbgs() << "Performing Loop Fusion on function " << F.getName()
598                       << "\n");
599     bool Changed = false;
600 
601     while (!LDT.empty()) {
602       LLVM_DEBUG(dbgs() << "Got " << LDT.size() << " loop sets for depth "
603                         << LDT.getDepth() << "\n";);
604 
605       for (const LoopVector &LV : LDT) {
606         assert(LV.size() > 0 && "Empty loop set was build!");
607 
608         // Skip singleton loop sets as they do not offer fusion opportunities on
609         // this level.
610         if (LV.size() == 1)
611           continue;
612 #ifndef NDEBUG
613         if (VerboseFusionDebugging) {
614           LLVM_DEBUG({
615             dbgs() << "  Visit loop set (#" << LV.size() << "):\n";
616             printLoopVector(LV);
617           });
618         }
619 #endif
620 
621         collectFusionCandidates(LV);
622         Changed |= fuseCandidates();
623       }
624 
625       // Finished analyzing candidates at this level.
626       // Descend to the next level and clear all of the candidates currently
627       // collected. Note that it will not be possible to fuse any of the
628       // existing candidates with new candidates because the new candidates will
629       // be at a different nest level and thus not be control flow equivalent
630       // with all of the candidates collected so far.
631       LLVM_DEBUG(dbgs() << "Descend one level!\n");
632       LDT.descend();
633       FusionCandidates.clear();
634     }
635 
636     if (Changed)
637       LLVM_DEBUG(dbgs() << "Function after Loop Fusion: \n"; F.dump(););
638 
639 #ifndef NDEBUG
640     assert(DT.verify());
641     assert(PDT.verify());
642     LI.verify(DT);
643     SE.verify();
644 #endif
645 
646     LLVM_DEBUG(dbgs() << "Loop Fusion complete\n");
647     return Changed;
648   }
649 
650 private:
651   /// Determine if two fusion candidates are control flow equivalent.
652   ///
653   /// Two fusion candidates are control flow equivalent if when one executes,
654   /// the other is guaranteed to execute. This is determined using dominators
655   /// and post-dominators: if A dominates B and B post-dominates A then A and B
656   /// are control-flow equivalent.
657   bool isControlFlowEquivalent(const FusionCandidate &FC0,
658                                const FusionCandidate &FC1) const {
659     assert(FC0.Preheader && FC1.Preheader && "Expecting valid preheaders");
660 
661     return ::isControlFlowEquivalent(*FC0.getEntryBlock(), *FC1.getEntryBlock(),
662                                      DT, PDT);
663   }
664 
665   /// Iterate over all loops in the given loop set and identify the loops that
666   /// are eligible for fusion. Place all eligible fusion candidates into Control
667   /// Flow Equivalent sets, sorted by dominance.
668   void collectFusionCandidates(const LoopVector &LV) {
669     for (Loop *L : LV) {
670       TTI::PeelingPreferences PP =
671           gatherPeelingPreferences(L, SE, TTI, std::nullopt, std::nullopt);
672       FusionCandidate CurrCand(L, DT, &PDT, ORE, PP);
673       if (!CurrCand.isEligibleForFusion(SE))
674         continue;
675 
676       // Go through each list in FusionCandidates and determine if L is control
677       // flow equivalent with the first loop in that list. If it is, append LV.
678       // If not, go to the next list.
679       // If no suitable list is found, start another list and add it to
680       // FusionCandidates.
681       bool FoundSet = false;
682 
683       for (auto &CurrCandSet : FusionCandidates) {
684         if (isControlFlowEquivalent(*CurrCandSet.begin(), CurrCand)) {
685           CurrCandSet.insert(CurrCand);
686           FoundSet = true;
687 #ifndef NDEBUG
688           if (VerboseFusionDebugging)
689             LLVM_DEBUG(dbgs() << "Adding " << CurrCand
690                               << " to existing candidate set\n");
691 #endif
692           break;
693         }
694       }
695       if (!FoundSet) {
696         // No set was found. Create a new set and add to FusionCandidates
697 #ifndef NDEBUG
698         if (VerboseFusionDebugging)
699           LLVM_DEBUG(dbgs() << "Adding " << CurrCand << " to new set\n");
700 #endif
701         FusionCandidateSet NewCandSet;
702         NewCandSet.insert(CurrCand);
703         FusionCandidates.push_back(NewCandSet);
704       }
705       NumFusionCandidates++;
706     }
707   }
708 
709   /// Determine if it is beneficial to fuse two loops.
710   ///
711   /// For now, this method simply returns true because we want to fuse as much
712   /// as possible (primarily to test the pass). This method will evolve, over
713   /// time, to add heuristics for profitability of fusion.
714   bool isBeneficialFusion(const FusionCandidate &FC0,
715                           const FusionCandidate &FC1) {
716     return true;
717   }
718 
719   /// Determine if two fusion candidates have the same trip count (i.e., they
720   /// execute the same number of iterations).
721   ///
722   /// This function will return a pair of values. The first is a boolean,
723   /// stating whether or not the two candidates are known at compile time to
724   /// have the same TripCount. The second is the difference in the two
725   /// TripCounts. This information can be used later to determine whether or not
726   /// peeling can be performed on either one of the candidates.
727   std::pair<bool, std::optional<unsigned>>
728   haveIdenticalTripCounts(const FusionCandidate &FC0,
729                           const FusionCandidate &FC1) const {
730     const SCEV *TripCount0 = SE.getBackedgeTakenCount(FC0.L);
731     if (isa<SCEVCouldNotCompute>(TripCount0)) {
732       UncomputableTripCount++;
733       LLVM_DEBUG(dbgs() << "Trip count of first loop could not be computed!");
734       return {false, std::nullopt};
735     }
736 
737     const SCEV *TripCount1 = SE.getBackedgeTakenCount(FC1.L);
738     if (isa<SCEVCouldNotCompute>(TripCount1)) {
739       UncomputableTripCount++;
740       LLVM_DEBUG(dbgs() << "Trip count of second loop could not be computed!");
741       return {false, std::nullopt};
742     }
743 
744     LLVM_DEBUG(dbgs() << "\tTrip counts: " << *TripCount0 << " & "
745                       << *TripCount1 << " are "
746                       << (TripCount0 == TripCount1 ? "identical" : "different")
747                       << "\n");
748 
749     if (TripCount0 == TripCount1)
750       return {true, 0};
751 
752     LLVM_DEBUG(dbgs() << "The loops do not have the same tripcount, "
753                          "determining the difference between trip counts\n");
754 
755     // Currently only considering loops with a single exit point
756     // and a non-constant trip count.
757     const unsigned TC0 = SE.getSmallConstantTripCount(FC0.L);
758     const unsigned TC1 = SE.getSmallConstantTripCount(FC1.L);
759 
760     // If any of the tripcounts are zero that means that loop(s) do not have
761     // a single exit or a constant tripcount.
762     if (TC0 == 0 || TC1 == 0) {
763       LLVM_DEBUG(dbgs() << "Loop(s) do not have a single exit point or do not "
764                            "have a constant number of iterations. Peeling "
765                            "is not benefical\n");
766       return {false, std::nullopt};
767     }
768 
769     std::optional<unsigned> Difference;
770     int Diff = TC0 - TC1;
771 
772     if (Diff > 0)
773       Difference = Diff;
774     else {
775       LLVM_DEBUG(
776           dbgs() << "Difference is less than 0. FC1 (second loop) has more "
777                     "iterations than the first one. Currently not supported\n");
778     }
779 
780     LLVM_DEBUG(dbgs() << "Difference in loop trip count is: " << Difference
781                       << "\n");
782 
783     return {false, Difference};
784   }
785 
786   void peelFusionCandidate(FusionCandidate &FC0, const FusionCandidate &FC1,
787                            unsigned PeelCount) {
788     assert(FC0.AbleToPeel && "Should be able to peel loop");
789 
790     LLVM_DEBUG(dbgs() << "Attempting to peel first " << PeelCount
791                       << " iterations of the first loop. \n");
792 
793     ValueToValueMapTy VMap;
794     FC0.Peeled = peelLoop(FC0.L, PeelCount, &LI, &SE, DT, &AC, true, VMap);
795     if (FC0.Peeled) {
796       LLVM_DEBUG(dbgs() << "Done Peeling\n");
797 
798 #ifndef NDEBUG
799       auto IdenticalTripCount = haveIdenticalTripCounts(FC0, FC1);
800 
801       assert(IdenticalTripCount.first && *IdenticalTripCount.second == 0 &&
802              "Loops should have identical trip counts after peeling");
803 #endif
804 
805       FC0.PP.PeelCount += PeelCount;
806 
807       // Peeling does not update the PDT
808       PDT.recalculate(*FC0.Preheader->getParent());
809 
810       FC0.updateAfterPeeling();
811 
812       // In this case the iterations of the loop are constant, so the first
813       // loop will execute completely (will not jump from one of
814       // the peeled blocks to the second loop). Here we are updating the
815       // branch conditions of each of the peeled blocks, such that it will
816       // branch to its successor which is not the preheader of the second loop
817       // in the case of unguarded loops, or the succesors of the exit block of
818       // the first loop otherwise. Doing this update will ensure that the entry
819       // block of the first loop dominates the entry block of the second loop.
820       BasicBlock *BB =
821           FC0.GuardBranch ? FC0.ExitBlock->getUniqueSuccessor() : FC1.Preheader;
822       if (BB) {
823         SmallVector<DominatorTree::UpdateType, 8> TreeUpdates;
824         SmallVector<Instruction *, 8> WorkList;
825         for (BasicBlock *Pred : predecessors(BB)) {
826           if (Pred != FC0.ExitBlock) {
827             WorkList.emplace_back(Pred->getTerminator());
828             TreeUpdates.emplace_back(
829                 DominatorTree::UpdateType(DominatorTree::Delete, Pred, BB));
830           }
831         }
832         // Cannot modify the predecessors inside the above loop as it will cause
833         // the iterators to be nullptrs, causing memory errors.
834         for (Instruction *CurrentBranch : WorkList) {
835           BasicBlock *Succ = CurrentBranch->getSuccessor(0);
836           if (Succ == BB)
837             Succ = CurrentBranch->getSuccessor(1);
838           ReplaceInstWithInst(CurrentBranch, BranchInst::Create(Succ));
839         }
840 
841         DTU.applyUpdates(TreeUpdates);
842         DTU.flush();
843       }
844       LLVM_DEBUG(
845           dbgs() << "Sucessfully peeled " << FC0.PP.PeelCount
846                  << " iterations from the first loop.\n"
847                     "Both Loops have the same number of iterations now.\n");
848     }
849   }
850 
851   /// Walk each set of control flow equivalent fusion candidates and attempt to
852   /// fuse them. This does a single linear traversal of all candidates in the
853   /// set. The conditions for legal fusion are checked at this point. If a pair
854   /// of fusion candidates passes all legality checks, they are fused together
855   /// and a new fusion candidate is created and added to the FusionCandidateSet.
856   /// The original fusion candidates are then removed, as they are no longer
857   /// valid.
858   bool fuseCandidates() {
859     bool Fused = false;
860     LLVM_DEBUG(printFusionCandidates(FusionCandidates));
861     for (auto &CandidateSet : FusionCandidates) {
862       if (CandidateSet.size() < 2)
863         continue;
864 
865       LLVM_DEBUG(dbgs() << "Attempting fusion on Candidate Set:\n"
866                         << CandidateSet << "\n");
867 
868       for (auto FC0 = CandidateSet.begin(); FC0 != CandidateSet.end(); ++FC0) {
869         assert(!LDT.isRemovedLoop(FC0->L) &&
870                "Should not have removed loops in CandidateSet!");
871         auto FC1 = FC0;
872         for (++FC1; FC1 != CandidateSet.end(); ++FC1) {
873           assert(!LDT.isRemovedLoop(FC1->L) &&
874                  "Should not have removed loops in CandidateSet!");
875 
876           LLVM_DEBUG(dbgs() << "Attempting to fuse candidate \n"; FC0->dump();
877                      dbgs() << " with\n"; FC1->dump(); dbgs() << "\n");
878 
879           FC0->verify();
880           FC1->verify();
881 
882           // Check if the candidates have identical tripcounts (first value of
883           // pair), and if not check the difference in the tripcounts between
884           // the loops (second value of pair). The difference is not equal to
885           // std::nullopt iff the loops iterate a constant number of times, and
886           // have a single exit.
887           std::pair<bool, std::optional<unsigned>> IdenticalTripCountRes =
888               haveIdenticalTripCounts(*FC0, *FC1);
889           bool SameTripCount = IdenticalTripCountRes.first;
890           std::optional<unsigned> TCDifference = IdenticalTripCountRes.second;
891 
892           // Here we are checking that FC0 (the first loop) can be peeled, and
893           // both loops have different tripcounts.
894           if (FC0->AbleToPeel && !SameTripCount && TCDifference) {
895             if (*TCDifference > FusionPeelMaxCount) {
896               LLVM_DEBUG(dbgs()
897                          << "Difference in loop trip counts: " << *TCDifference
898                          << " is greater than maximum peel count specificed: "
899                          << FusionPeelMaxCount << "\n");
900             } else {
901               // Dependent on peeling being performed on the first loop, and
902               // assuming all other conditions for fusion return true.
903               SameTripCount = true;
904             }
905           }
906 
907           if (!SameTripCount) {
908             LLVM_DEBUG(dbgs() << "Fusion candidates do not have identical trip "
909                                  "counts. Not fusing.\n");
910             reportLoopFusion<OptimizationRemarkMissed>(*FC0, *FC1,
911                                                        NonEqualTripCount);
912             continue;
913           }
914 
915           if (!isAdjacent(*FC0, *FC1)) {
916             LLVM_DEBUG(dbgs()
917                        << "Fusion candidates are not adjacent. Not fusing.\n");
918             reportLoopFusion<OptimizationRemarkMissed>(*FC0, *FC1, NonAdjacent);
919             continue;
920           }
921 
922           if ((!FC0->GuardBranch && FC1->GuardBranch) ||
923               (FC0->GuardBranch && !FC1->GuardBranch)) {
924             LLVM_DEBUG(dbgs() << "The one of candidate is guarded while the "
925                                  "another one is not. Not fusing.\n");
926             reportLoopFusion<OptimizationRemarkMissed>(
927                 *FC0, *FC1, OnlySecondCandidateIsGuarded);
928             continue;
929           }
930 
931           // Ensure that FC0 and FC1 have identical guards.
932           // If one (or both) are not guarded, this check is not necessary.
933           if (FC0->GuardBranch && FC1->GuardBranch &&
934               !haveIdenticalGuards(*FC0, *FC1) && !TCDifference) {
935             LLVM_DEBUG(dbgs() << "Fusion candidates do not have identical "
936                                  "guards. Not Fusing.\n");
937             reportLoopFusion<OptimizationRemarkMissed>(*FC0, *FC1,
938                                                        NonIdenticalGuards);
939             continue;
940           }
941 
942           if (FC0->GuardBranch) {
943             assert(FC1->GuardBranch && "Expecting valid FC1 guard branch");
944 
945             if (!isSafeToMoveBefore(*FC0->ExitBlock,
946                                     *FC1->ExitBlock->getFirstNonPHIOrDbg(), DT,
947                                     &PDT, &DI)) {
948               LLVM_DEBUG(dbgs() << "Fusion candidate contains unsafe "
949                                    "instructions in exit block. Not fusing.\n");
950               reportLoopFusion<OptimizationRemarkMissed>(*FC0, *FC1,
951                                                          NonEmptyExitBlock);
952               continue;
953             }
954 
955             if (!isSafeToMoveBefore(
956                     *FC1->GuardBranch->getParent(),
957                     *FC0->GuardBranch->getParent()->getTerminator(), DT, &PDT,
958                     &DI)) {
959               LLVM_DEBUG(dbgs()
960                          << "Fusion candidate contains unsafe "
961                             "instructions in guard block. Not fusing.\n");
962               reportLoopFusion<OptimizationRemarkMissed>(*FC0, *FC1,
963                                                          NonEmptyGuardBlock);
964               continue;
965             }
966           }
967 
968           // Check the dependencies across the loops and do not fuse if it would
969           // violate them.
970           if (!dependencesAllowFusion(*FC0, *FC1)) {
971             LLVM_DEBUG(dbgs() << "Memory dependencies do not allow fusion!\n");
972             reportLoopFusion<OptimizationRemarkMissed>(*FC0, *FC1,
973                                                        InvalidDependencies);
974             continue;
975           }
976 
977           // If the second loop has instructions in the pre-header, attempt to
978           // hoist them up to the first loop's pre-header or sink them into the
979           // body of the second loop.
980           SmallVector<Instruction *, 4> SafeToHoist;
981           SmallVector<Instruction *, 4> SafeToSink;
982           // At this point, this is the last remaining legality check.
983           // Which means if we can make this pre-header empty, we can fuse
984           // these loops
985           if (!isEmptyPreheader(*FC1)) {
986             LLVM_DEBUG(dbgs() << "Fusion candidate does not have empty "
987                                  "preheader.\n");
988 
989             // If it is not safe to hoist/sink all instructions in the
990             // pre-header, we cannot fuse these loops.
991             if (!collectMovablePreheaderInsts(*FC0, *FC1, SafeToHoist,
992                                               SafeToSink)) {
993               LLVM_DEBUG(dbgs() << "Could not hoist/sink all instructions in "
994                                    "Fusion Candidate Pre-header.\n"
995                                 << "Not Fusing.\n");
996               reportLoopFusion<OptimizationRemarkMissed>(*FC0, *FC1,
997                                                          NonEmptyPreheader);
998               continue;
999             }
1000           }
1001 
1002           bool BeneficialToFuse = isBeneficialFusion(*FC0, *FC1);
1003           LLVM_DEBUG(dbgs()
1004                      << "\tFusion appears to be "
1005                      << (BeneficialToFuse ? "" : "un") << "profitable!\n");
1006           if (!BeneficialToFuse) {
1007             reportLoopFusion<OptimizationRemarkMissed>(*FC0, *FC1,
1008                                                        FusionNotBeneficial);
1009             continue;
1010           }
1011           // All analysis has completed and has determined that fusion is legal
1012           // and profitable. At this point, start transforming the code and
1013           // perform fusion.
1014 
1015           // Execute the hoist/sink operations on preheader instructions
1016           movePreheaderInsts(*FC0, *FC1, SafeToHoist, SafeToSink);
1017 
1018           LLVM_DEBUG(dbgs() << "\tFusion is performed: " << *FC0 << " and "
1019                             << *FC1 << "\n");
1020 
1021           FusionCandidate FC0Copy = *FC0;
1022           // Peel the loop after determining that fusion is legal. The Loops
1023           // will still be safe to fuse after the peeling is performed.
1024           bool Peel = TCDifference && *TCDifference > 0;
1025           if (Peel)
1026             peelFusionCandidate(FC0Copy, *FC1, *TCDifference);
1027 
1028           // Report fusion to the Optimization Remarks.
1029           // Note this needs to be done *before* performFusion because
1030           // performFusion will change the original loops, making it not
1031           // possible to identify them after fusion is complete.
1032           reportLoopFusion<OptimizationRemark>((Peel ? FC0Copy : *FC0), *FC1,
1033                                                FuseCounter);
1034 
1035           FusionCandidate FusedCand(
1036               performFusion((Peel ? FC0Copy : *FC0), *FC1), DT, &PDT, ORE,
1037               FC0Copy.PP);
1038           FusedCand.verify();
1039           assert(FusedCand.isEligibleForFusion(SE) &&
1040                  "Fused candidate should be eligible for fusion!");
1041 
1042           // Notify the loop-depth-tree that these loops are not valid objects
1043           LDT.removeLoop(FC1->L);
1044 
1045           CandidateSet.erase(FC0);
1046           CandidateSet.erase(FC1);
1047 
1048           auto InsertPos = CandidateSet.insert(FusedCand);
1049 
1050           assert(InsertPos.second &&
1051                  "Unable to insert TargetCandidate in CandidateSet!");
1052 
1053           // Reset FC0 and FC1 the new (fused) candidate. Subsequent iterations
1054           // of the FC1 loop will attempt to fuse the new (fused) loop with the
1055           // remaining candidates in the current candidate set.
1056           FC0 = FC1 = InsertPos.first;
1057 
1058           LLVM_DEBUG(dbgs() << "Candidate Set (after fusion): " << CandidateSet
1059                             << "\n");
1060 
1061           Fused = true;
1062         }
1063       }
1064     }
1065     return Fused;
1066   }
1067 
1068   // Returns true if the instruction \p I can be hoisted to the end of the
1069   // preheader of \p FC0. \p SafeToHoist contains the instructions that are
1070   // known to be safe to hoist. The instructions encountered that cannot be
1071   // hoisted are in \p NotHoisting.
1072   // TODO: Move functionality into CodeMoverUtils
1073   bool canHoistInst(Instruction &I,
1074                     const SmallVector<Instruction *, 4> &SafeToHoist,
1075                     const SmallVector<Instruction *, 4> &NotHoisting,
1076                     const FusionCandidate &FC0) const {
1077     const BasicBlock *FC0PreheaderTarget = FC0.Preheader->getSingleSuccessor();
1078     assert(FC0PreheaderTarget &&
1079            "Expected single successor for loop preheader.");
1080 
1081     for (Use &Op : I.operands()) {
1082       if (auto *OpInst = dyn_cast<Instruction>(Op)) {
1083         bool OpHoisted = is_contained(SafeToHoist, OpInst);
1084         // Check if we have already decided to hoist this operand. In this
1085         // case, it does not dominate FC0 *yet*, but will after we hoist it.
1086         if (!(OpHoisted || DT.dominates(OpInst, FC0PreheaderTarget))) {
1087           return false;
1088         }
1089       }
1090     }
1091 
1092     // PHIs in FC1's header only have FC0 blocks as predecessors. PHIs
1093     // cannot be hoisted and should be sunk to the exit of the fused loop.
1094     if (isa<PHINode>(I))
1095       return false;
1096 
1097     // If this isn't a memory inst, hoisting is safe
1098     if (!I.mayReadOrWriteMemory())
1099       return true;
1100 
1101     LLVM_DEBUG(dbgs() << "Checking if this mem inst can be hoisted.\n");
1102     for (Instruction *NotHoistedInst : NotHoisting) {
1103       if (auto D = DI.depends(&I, NotHoistedInst, true)) {
1104         // Dependency is not read-before-write, write-before-read or
1105         // write-before-write
1106         if (D->isFlow() || D->isAnti() || D->isOutput()) {
1107           LLVM_DEBUG(dbgs() << "Inst depends on an instruction in FC1's "
1108                                "preheader that is not being hoisted.\n");
1109           return false;
1110         }
1111       }
1112     }
1113 
1114     for (Instruction *ReadInst : FC0.MemReads) {
1115       if (auto D = DI.depends(ReadInst, &I, true)) {
1116         // Dependency is not read-before-write
1117         if (D->isAnti()) {
1118           LLVM_DEBUG(dbgs() << "Inst depends on a read instruction in FC0.\n");
1119           return false;
1120         }
1121       }
1122     }
1123 
1124     for (Instruction *WriteInst : FC0.MemWrites) {
1125       if (auto D = DI.depends(WriteInst, &I, true)) {
1126         // Dependency is not write-before-read or write-before-write
1127         if (D->isFlow() || D->isOutput()) {
1128           LLVM_DEBUG(dbgs() << "Inst depends on a write instruction in FC0.\n");
1129           return false;
1130         }
1131       }
1132     }
1133     return true;
1134   }
1135 
1136   // Returns true if the instruction \p I can be sunk to the top of the exit
1137   // block of \p FC1.
1138   // TODO: Move functionality into CodeMoverUtils
1139   bool canSinkInst(Instruction &I, const FusionCandidate &FC1) const {
1140     for (User *U : I.users()) {
1141       if (auto *UI{dyn_cast<Instruction>(U)}) {
1142         // Cannot sink if user in loop
1143         // If FC1 has phi users of this value, we cannot sink it into FC1.
1144         if (FC1.L->contains(UI)) {
1145           // Cannot hoist or sink this instruction. No hoisting/sinking
1146           // should take place, loops should not fuse
1147           return false;
1148         }
1149       }
1150     }
1151 
1152     // If this isn't a memory inst, sinking is safe
1153     if (!I.mayReadOrWriteMemory())
1154       return true;
1155 
1156     for (Instruction *ReadInst : FC1.MemReads) {
1157       if (auto D = DI.depends(&I, ReadInst, true)) {
1158         // Dependency is not write-before-read
1159         if (D->isFlow()) {
1160           LLVM_DEBUG(dbgs() << "Inst depends on a read instruction in FC1.\n");
1161           return false;
1162         }
1163       }
1164     }
1165 
1166     for (Instruction *WriteInst : FC1.MemWrites) {
1167       if (auto D = DI.depends(&I, WriteInst, true)) {
1168         // Dependency is not write-before-write or read-before-write
1169         if (D->isOutput() || D->isAnti()) {
1170           LLVM_DEBUG(dbgs() << "Inst depends on a write instruction in FC1.\n");
1171           return false;
1172         }
1173       }
1174     }
1175 
1176     return true;
1177   }
1178 
1179   /// Collect instructions in the \p FC1 Preheader that can be hoisted
1180   /// to the \p FC0 Preheader or sunk into the \p FC1 Body
1181   bool collectMovablePreheaderInsts(
1182       const FusionCandidate &FC0, const FusionCandidate &FC1,
1183       SmallVector<Instruction *, 4> &SafeToHoist,
1184       SmallVector<Instruction *, 4> &SafeToSink) const {
1185     BasicBlock *FC1Preheader = FC1.Preheader;
1186     // Save the instructions that are not being hoisted, so we know not to hoist
1187     // mem insts that they dominate.
1188     SmallVector<Instruction *, 4> NotHoisting;
1189 
1190     for (Instruction &I : *FC1Preheader) {
1191       // Can't move a branch
1192       if (&I == FC1Preheader->getTerminator())
1193         continue;
1194       // If the instruction has side-effects, give up.
1195       // TODO: The case of mayReadFromMemory we can handle but requires
1196       // additional work with a dependence analysis so for now we give
1197       // up on memory reads.
1198       if (I.mayThrow() || !I.willReturn()) {
1199         LLVM_DEBUG(dbgs() << "Inst: " << I << " may throw or won't return.\n");
1200         return false;
1201       }
1202 
1203       LLVM_DEBUG(dbgs() << "Checking Inst: " << I << "\n");
1204 
1205       if (I.isAtomic() || I.isVolatile()) {
1206         LLVM_DEBUG(
1207             dbgs() << "\tInstruction is volatile or atomic. Cannot move it.\n");
1208         return false;
1209       }
1210 
1211       if (canHoistInst(I, SafeToHoist, NotHoisting, FC0)) {
1212         SafeToHoist.push_back(&I);
1213         LLVM_DEBUG(dbgs() << "\tSafe to hoist.\n");
1214       } else {
1215         LLVM_DEBUG(dbgs() << "\tCould not hoist. Trying to sink...\n");
1216         NotHoisting.push_back(&I);
1217 
1218         if (canSinkInst(I, FC1)) {
1219           SafeToSink.push_back(&I);
1220           LLVM_DEBUG(dbgs() << "\tSafe to sink.\n");
1221         } else {
1222           LLVM_DEBUG(dbgs() << "\tCould not sink.\n");
1223           return false;
1224         }
1225       }
1226     }
1227     LLVM_DEBUG(
1228         dbgs() << "All preheader instructions could be sunk or hoisted!\n");
1229     return true;
1230   }
1231 
1232   /// Rewrite all additive recurrences in a SCEV to use a new loop.
1233   class AddRecLoopReplacer : public SCEVRewriteVisitor<AddRecLoopReplacer> {
1234   public:
1235     AddRecLoopReplacer(ScalarEvolution &SE, const Loop &OldL, const Loop &NewL,
1236                        bool UseMax = true)
1237         : SCEVRewriteVisitor(SE), Valid(true), UseMax(UseMax), OldL(OldL),
1238           NewL(NewL) {}
1239 
1240     const SCEV *visitAddRecExpr(const SCEVAddRecExpr *Expr) {
1241       const Loop *ExprL = Expr->getLoop();
1242       SmallVector<const SCEV *, 2> Operands;
1243       if (ExprL == &OldL) {
1244         append_range(Operands, Expr->operands());
1245         return SE.getAddRecExpr(Operands, &NewL, Expr->getNoWrapFlags());
1246       }
1247 
1248       if (OldL.contains(ExprL)) {
1249         bool Pos = SE.isKnownPositive(Expr->getStepRecurrence(SE));
1250         if (!UseMax || !Pos || !Expr->isAffine()) {
1251           Valid = false;
1252           return Expr;
1253         }
1254         return visit(Expr->getStart());
1255       }
1256 
1257       for (const SCEV *Op : Expr->operands())
1258         Operands.push_back(visit(Op));
1259       return SE.getAddRecExpr(Operands, ExprL, Expr->getNoWrapFlags());
1260     }
1261 
1262     bool wasValidSCEV() const { return Valid; }
1263 
1264   private:
1265     bool Valid, UseMax;
1266     const Loop &OldL, &NewL;
1267   };
1268 
1269   /// Return false if the access functions of \p I0 and \p I1 could cause
1270   /// a negative dependence.
1271   bool accessDiffIsPositive(const Loop &L0, const Loop &L1, Instruction &I0,
1272                             Instruction &I1, bool EqualIsInvalid) {
1273     Value *Ptr0 = getLoadStorePointerOperand(&I0);
1274     Value *Ptr1 = getLoadStorePointerOperand(&I1);
1275     if (!Ptr0 || !Ptr1)
1276       return false;
1277 
1278     const SCEV *SCEVPtr0 = SE.getSCEVAtScope(Ptr0, &L0);
1279     const SCEV *SCEVPtr1 = SE.getSCEVAtScope(Ptr1, &L1);
1280 #ifndef NDEBUG
1281     if (VerboseFusionDebugging)
1282       LLVM_DEBUG(dbgs() << "    Access function check: " << *SCEVPtr0 << " vs "
1283                         << *SCEVPtr1 << "\n");
1284 #endif
1285     AddRecLoopReplacer Rewriter(SE, L0, L1);
1286     SCEVPtr0 = Rewriter.visit(SCEVPtr0);
1287 #ifndef NDEBUG
1288     if (VerboseFusionDebugging)
1289       LLVM_DEBUG(dbgs() << "    Access function after rewrite: " << *SCEVPtr0
1290                         << " [Valid: " << Rewriter.wasValidSCEV() << "]\n");
1291 #endif
1292     if (!Rewriter.wasValidSCEV())
1293       return false;
1294 
1295     // TODO: isKnownPredicate doesnt work well when one SCEV is loop carried (by
1296     //       L0) and the other is not. We could check if it is monotone and test
1297     //       the beginning and end value instead.
1298 
1299     BasicBlock *L0Header = L0.getHeader();
1300     auto HasNonLinearDominanceRelation = [&](const SCEV *S) {
1301       const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(S);
1302       if (!AddRec)
1303         return false;
1304       return !DT.dominates(L0Header, AddRec->getLoop()->getHeader()) &&
1305              !DT.dominates(AddRec->getLoop()->getHeader(), L0Header);
1306     };
1307     if (SCEVExprContains(SCEVPtr1, HasNonLinearDominanceRelation))
1308       return false;
1309 
1310     ICmpInst::Predicate Pred =
1311         EqualIsInvalid ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_SGE;
1312     bool IsAlwaysGE = SE.isKnownPredicate(Pred, SCEVPtr0, SCEVPtr1);
1313 #ifndef NDEBUG
1314     if (VerboseFusionDebugging)
1315       LLVM_DEBUG(dbgs() << "    Relation: " << *SCEVPtr0
1316                         << (IsAlwaysGE ? "  >=  " : "  may <  ") << *SCEVPtr1
1317                         << "\n");
1318 #endif
1319     return IsAlwaysGE;
1320   }
1321 
1322   /// Return true if the dependences between @p I0 (in @p L0) and @p I1 (in
1323   /// @p L1) allow loop fusion of @p L0 and @p L1. The dependence analyses
1324   /// specified by @p DepChoice are used to determine this.
1325   bool dependencesAllowFusion(const FusionCandidate &FC0,
1326                               const FusionCandidate &FC1, Instruction &I0,
1327                               Instruction &I1, bool AnyDep,
1328                               FusionDependenceAnalysisChoice DepChoice) {
1329 #ifndef NDEBUG
1330     if (VerboseFusionDebugging) {
1331       LLVM_DEBUG(dbgs() << "Check dep: " << I0 << " vs " << I1 << " : "
1332                         << DepChoice << "\n");
1333     }
1334 #endif
1335     switch (DepChoice) {
1336     case FUSION_DEPENDENCE_ANALYSIS_SCEV:
1337       return accessDiffIsPositive(*FC0.L, *FC1.L, I0, I1, AnyDep);
1338     case FUSION_DEPENDENCE_ANALYSIS_DA: {
1339       auto DepResult = DI.depends(&I0, &I1, true);
1340       if (!DepResult)
1341         return true;
1342 #ifndef NDEBUG
1343       if (VerboseFusionDebugging) {
1344         LLVM_DEBUG(dbgs() << "DA res: "; DepResult->dump(dbgs());
1345                    dbgs() << " [#l: " << DepResult->getLevels() << "][Ordered: "
1346                           << (DepResult->isOrdered() ? "true" : "false")
1347                           << "]\n");
1348         LLVM_DEBUG(dbgs() << "DepResult Levels: " << DepResult->getLevels()
1349                           << "\n");
1350       }
1351 #endif
1352 
1353       if (DepResult->getNextPredecessor() || DepResult->getNextSuccessor())
1354         LLVM_DEBUG(
1355             dbgs() << "TODO: Implement pred/succ dependence handling!\n");
1356 
1357       // TODO: Can we actually use the dependence info analysis here?
1358       return false;
1359     }
1360 
1361     case FUSION_DEPENDENCE_ANALYSIS_ALL:
1362       return dependencesAllowFusion(FC0, FC1, I0, I1, AnyDep,
1363                                     FUSION_DEPENDENCE_ANALYSIS_SCEV) ||
1364              dependencesAllowFusion(FC0, FC1, I0, I1, AnyDep,
1365                                     FUSION_DEPENDENCE_ANALYSIS_DA);
1366     }
1367 
1368     llvm_unreachable("Unknown fusion dependence analysis choice!");
1369   }
1370 
1371   /// Perform a dependence check and return if @p FC0 and @p FC1 can be fused.
1372   bool dependencesAllowFusion(const FusionCandidate &FC0,
1373                               const FusionCandidate &FC1) {
1374     LLVM_DEBUG(dbgs() << "Check if " << FC0 << " can be fused with " << FC1
1375                       << "\n");
1376     assert(FC0.L->getLoopDepth() == FC1.L->getLoopDepth());
1377     assert(DT.dominates(FC0.getEntryBlock(), FC1.getEntryBlock()));
1378 
1379     for (Instruction *WriteL0 : FC0.MemWrites) {
1380       for (Instruction *WriteL1 : FC1.MemWrites)
1381         if (!dependencesAllowFusion(FC0, FC1, *WriteL0, *WriteL1,
1382                                     /* AnyDep */ false,
1383                                     FusionDependenceAnalysis)) {
1384           InvalidDependencies++;
1385           return false;
1386         }
1387       for (Instruction *ReadL1 : FC1.MemReads)
1388         if (!dependencesAllowFusion(FC0, FC1, *WriteL0, *ReadL1,
1389                                     /* AnyDep */ false,
1390                                     FusionDependenceAnalysis)) {
1391           InvalidDependencies++;
1392           return false;
1393         }
1394     }
1395 
1396     for (Instruction *WriteL1 : FC1.MemWrites) {
1397       for (Instruction *WriteL0 : FC0.MemWrites)
1398         if (!dependencesAllowFusion(FC0, FC1, *WriteL0, *WriteL1,
1399                                     /* AnyDep */ false,
1400                                     FusionDependenceAnalysis)) {
1401           InvalidDependencies++;
1402           return false;
1403         }
1404       for (Instruction *ReadL0 : FC0.MemReads)
1405         if (!dependencesAllowFusion(FC0, FC1, *ReadL0, *WriteL1,
1406                                     /* AnyDep */ false,
1407                                     FusionDependenceAnalysis)) {
1408           InvalidDependencies++;
1409           return false;
1410         }
1411     }
1412 
1413     // Walk through all uses in FC1. For each use, find the reaching def. If the
1414     // def is located in FC0 then it is is not safe to fuse.
1415     for (BasicBlock *BB : FC1.L->blocks())
1416       for (Instruction &I : *BB)
1417         for (auto &Op : I.operands())
1418           if (Instruction *Def = dyn_cast<Instruction>(Op))
1419             if (FC0.L->contains(Def->getParent())) {
1420               InvalidDependencies++;
1421               return false;
1422             }
1423 
1424     return true;
1425   }
1426 
1427   /// Determine if two fusion candidates are adjacent in the CFG.
1428   ///
1429   /// This method will determine if there are additional basic blocks in the CFG
1430   /// between the exit of \p FC0 and the entry of \p FC1.
1431   /// If the two candidates are guarded loops, then it checks whether the
1432   /// non-loop successor of the \p FC0 guard branch is the entry block of \p
1433   /// FC1. If not, then the loops are not adjacent. If the two candidates are
1434   /// not guarded loops, then it checks whether the exit block of \p FC0 is the
1435   /// preheader of \p FC1.
1436   bool isAdjacent(const FusionCandidate &FC0,
1437                   const FusionCandidate &FC1) const {
1438     // If the successor of the guard branch is FC1, then the loops are adjacent
1439     if (FC0.GuardBranch)
1440       return FC0.getNonLoopBlock() == FC1.getEntryBlock();
1441     else
1442       return FC0.ExitBlock == FC1.getEntryBlock();
1443   }
1444 
1445   bool isEmptyPreheader(const FusionCandidate &FC) const {
1446     return FC.Preheader->size() == 1;
1447   }
1448 
1449   /// Hoist \p FC1 Preheader instructions to \p FC0 Preheader
1450   /// and sink others into the body of \p FC1.
1451   void movePreheaderInsts(const FusionCandidate &FC0,
1452                           const FusionCandidate &FC1,
1453                           SmallVector<Instruction *, 4> &HoistInsts,
1454                           SmallVector<Instruction *, 4> &SinkInsts) const {
1455     // All preheader instructions except the branch must be hoisted or sunk
1456     assert(HoistInsts.size() + SinkInsts.size() == FC1.Preheader->size() - 1 &&
1457            "Attempting to sink and hoist preheader instructions, but not all "
1458            "the preheader instructions are accounted for.");
1459 
1460     NumHoistedInsts += HoistInsts.size();
1461     NumSunkInsts += SinkInsts.size();
1462 
1463     LLVM_DEBUG(if (VerboseFusionDebugging) {
1464       if (!HoistInsts.empty())
1465         dbgs() << "Hoisting: \n";
1466       for (Instruction *I : HoistInsts)
1467         dbgs() << *I << "\n";
1468       if (!SinkInsts.empty())
1469         dbgs() << "Sinking: \n";
1470       for (Instruction *I : SinkInsts)
1471         dbgs() << *I << "\n";
1472     });
1473 
1474     for (Instruction *I : HoistInsts) {
1475       assert(I->getParent() == FC1.Preheader);
1476       I->moveBefore(FC0.Preheader->getTerminator());
1477     }
1478     // insert instructions in reverse order to maintain dominance relationship
1479     for (Instruction *I : reverse(SinkInsts)) {
1480       assert(I->getParent() == FC1.Preheader);
1481       I->moveBefore(&*FC1.ExitBlock->getFirstInsertionPt());
1482     }
1483   }
1484 
1485   /// Determine if two fusion candidates have identical guards
1486   ///
1487   /// This method will determine if two fusion candidates have the same guards.
1488   /// The guards are considered the same if:
1489   ///   1. The instructions to compute the condition used in the compare are
1490   ///      identical.
1491   ///   2. The successors of the guard have the same flow into/around the loop.
1492   /// If the compare instructions are identical, then the first successor of the
1493   /// guard must go to the same place (either the preheader of the loop or the
1494   /// NonLoopBlock). In other words, the the first successor of both loops must
1495   /// both go into the loop (i.e., the preheader) or go around the loop (i.e.,
1496   /// the NonLoopBlock). The same must be true for the second successor.
1497   bool haveIdenticalGuards(const FusionCandidate &FC0,
1498                            const FusionCandidate &FC1) const {
1499     assert(FC0.GuardBranch && FC1.GuardBranch &&
1500            "Expecting FC0 and FC1 to be guarded loops.");
1501 
1502     if (auto FC0CmpInst =
1503             dyn_cast<Instruction>(FC0.GuardBranch->getCondition()))
1504       if (auto FC1CmpInst =
1505               dyn_cast<Instruction>(FC1.GuardBranch->getCondition()))
1506         if (!FC0CmpInst->isIdenticalTo(FC1CmpInst))
1507           return false;
1508 
1509     // The compare instructions are identical.
1510     // Now make sure the successor of the guards have the same flow into/around
1511     // the loop
1512     if (FC0.GuardBranch->getSuccessor(0) == FC0.Preheader)
1513       return (FC1.GuardBranch->getSuccessor(0) == FC1.Preheader);
1514     else
1515       return (FC1.GuardBranch->getSuccessor(1) == FC1.Preheader);
1516   }
1517 
1518   /// Modify the latch branch of FC to be unconditional since successors of the
1519   /// branch are the same.
1520   void simplifyLatchBranch(const FusionCandidate &FC) const {
1521     BranchInst *FCLatchBranch = dyn_cast<BranchInst>(FC.Latch->getTerminator());
1522     if (FCLatchBranch) {
1523       assert(FCLatchBranch->isConditional() &&
1524              FCLatchBranch->getSuccessor(0) == FCLatchBranch->getSuccessor(1) &&
1525              "Expecting the two successors of FCLatchBranch to be the same");
1526       BranchInst *NewBranch =
1527           BranchInst::Create(FCLatchBranch->getSuccessor(0));
1528       ReplaceInstWithInst(FCLatchBranch, NewBranch);
1529     }
1530   }
1531 
1532   /// Move instructions from FC0.Latch to FC1.Latch. If FC0.Latch has an unique
1533   /// successor, then merge FC0.Latch with its unique successor.
1534   void mergeLatch(const FusionCandidate &FC0, const FusionCandidate &FC1) {
1535     moveInstructionsToTheBeginning(*FC0.Latch, *FC1.Latch, DT, PDT, DI);
1536     if (BasicBlock *Succ = FC0.Latch->getUniqueSuccessor()) {
1537       MergeBlockIntoPredecessor(Succ, &DTU, &LI);
1538       DTU.flush();
1539     }
1540   }
1541 
1542   /// Fuse two fusion candidates, creating a new fused loop.
1543   ///
1544   /// This method contains the mechanics of fusing two loops, represented by \p
1545   /// FC0 and \p FC1. It is assumed that \p FC0 dominates \p FC1 and \p FC1
1546   /// postdominates \p FC0 (making them control flow equivalent). It also
1547   /// assumes that the other conditions for fusion have been met: adjacent,
1548   /// identical trip counts, and no negative distance dependencies exist that
1549   /// would prevent fusion. Thus, there is no checking for these conditions in
1550   /// this method.
1551   ///
1552   /// Fusion is performed by rewiring the CFG to update successor blocks of the
1553   /// components of tho loop. Specifically, the following changes are done:
1554   ///
1555   ///   1. The preheader of \p FC1 is removed as it is no longer necessary
1556   ///   (because it is currently only a single statement block).
1557   ///   2. The latch of \p FC0 is modified to jump to the header of \p FC1.
1558   ///   3. The latch of \p FC1 i modified to jump to the header of \p FC0.
1559   ///   4. All blocks from \p FC1 are removed from FC1 and added to FC0.
1560   ///
1561   /// All of these modifications are done with dominator tree updates, thus
1562   /// keeping the dominator (and post dominator) information up-to-date.
1563   ///
1564   /// This can be improved in the future by actually merging blocks during
1565   /// fusion. For example, the preheader of \p FC1 can be merged with the
1566   /// preheader of \p FC0. This would allow loops with more than a single
1567   /// statement in the preheader to be fused. Similarly, the latch blocks of the
1568   /// two loops could also be fused into a single block. This will require
1569   /// analysis to prove it is safe to move the contents of the block past
1570   /// existing code, which currently has not been implemented.
1571   Loop *performFusion(const FusionCandidate &FC0, const FusionCandidate &FC1) {
1572     assert(FC0.isValid() && FC1.isValid() &&
1573            "Expecting valid fusion candidates");
1574 
1575     LLVM_DEBUG(dbgs() << "Fusion Candidate 0: \n"; FC0.dump();
1576                dbgs() << "Fusion Candidate 1: \n"; FC1.dump(););
1577 
1578     // Move instructions from the preheader of FC1 to the end of the preheader
1579     // of FC0.
1580     moveInstructionsToTheEnd(*FC1.Preheader, *FC0.Preheader, DT, PDT, DI);
1581 
1582     // Fusing guarded loops is handled slightly differently than non-guarded
1583     // loops and has been broken out into a separate method instead of trying to
1584     // intersperse the logic within a single method.
1585     if (FC0.GuardBranch)
1586       return fuseGuardedLoops(FC0, FC1);
1587 
1588     assert(FC1.Preheader ==
1589            (FC0.Peeled ? FC0.ExitBlock->getUniqueSuccessor() : FC0.ExitBlock));
1590     assert(FC1.Preheader->size() == 1 &&
1591            FC1.Preheader->getSingleSuccessor() == FC1.Header);
1592 
1593     // Remember the phi nodes originally in the header of FC0 in order to rewire
1594     // them later. However, this is only necessary if the new loop carried
1595     // values might not dominate the exiting branch. While we do not generally
1596     // test if this is the case but simply insert intermediate phi nodes, we
1597     // need to make sure these intermediate phi nodes have different
1598     // predecessors. To this end, we filter the special case where the exiting
1599     // block is the latch block of the first loop. Nothing needs to be done
1600     // anyway as all loop carried values dominate the latch and thereby also the
1601     // exiting branch.
1602     SmallVector<PHINode *, 8> OriginalFC0PHIs;
1603     if (FC0.ExitingBlock != FC0.Latch)
1604       for (PHINode &PHI : FC0.Header->phis())
1605         OriginalFC0PHIs.push_back(&PHI);
1606 
1607     // Replace incoming blocks for header PHIs first.
1608     FC1.Preheader->replaceSuccessorsPhiUsesWith(FC0.Preheader);
1609     FC0.Latch->replaceSuccessorsPhiUsesWith(FC1.Latch);
1610 
1611     // Then modify the control flow and update DT and PDT.
1612     SmallVector<DominatorTree::UpdateType, 8> TreeUpdates;
1613 
1614     // The old exiting block of the first loop (FC0) has to jump to the header
1615     // of the second as we need to execute the code in the second header block
1616     // regardless of the trip count. That is, if the trip count is 0, so the
1617     // back edge is never taken, we still have to execute both loop headers,
1618     // especially (but not only!) if the second is a do-while style loop.
1619     // However, doing so might invalidate the phi nodes of the first loop as
1620     // the new values do only need to dominate their latch and not the exiting
1621     // predicate. To remedy this potential problem we always introduce phi
1622     // nodes in the header of the second loop later that select the loop carried
1623     // value, if the second header was reached through an old latch of the
1624     // first, or undef otherwise. This is sound as exiting the first implies the
1625     // second will exit too, __without__ taking the back-edge. [Their
1626     // trip-counts are equal after all.
1627     // KB: Would this sequence be simpler to just just make FC0.ExitingBlock go
1628     // to FC1.Header? I think this is basically what the three sequences are
1629     // trying to accomplish; however, doing this directly in the CFG may mean
1630     // the DT/PDT becomes invalid
1631     if (!FC0.Peeled) {
1632       FC0.ExitingBlock->getTerminator()->replaceUsesOfWith(FC1.Preheader,
1633                                                            FC1.Header);
1634       TreeUpdates.emplace_back(DominatorTree::UpdateType(
1635           DominatorTree::Delete, FC0.ExitingBlock, FC1.Preheader));
1636       TreeUpdates.emplace_back(DominatorTree::UpdateType(
1637           DominatorTree::Insert, FC0.ExitingBlock, FC1.Header));
1638     } else {
1639       TreeUpdates.emplace_back(DominatorTree::UpdateType(
1640           DominatorTree::Delete, FC0.ExitBlock, FC1.Preheader));
1641 
1642       // Remove the ExitBlock of the first Loop (also not needed)
1643       FC0.ExitingBlock->getTerminator()->replaceUsesOfWith(FC0.ExitBlock,
1644                                                            FC1.Header);
1645       TreeUpdates.emplace_back(DominatorTree::UpdateType(
1646           DominatorTree::Delete, FC0.ExitingBlock, FC0.ExitBlock));
1647       FC0.ExitBlock->getTerminator()->eraseFromParent();
1648       TreeUpdates.emplace_back(DominatorTree::UpdateType(
1649           DominatorTree::Insert, FC0.ExitingBlock, FC1.Header));
1650       new UnreachableInst(FC0.ExitBlock->getContext(), FC0.ExitBlock);
1651     }
1652 
1653     // The pre-header of L1 is not necessary anymore.
1654     assert(pred_empty(FC1.Preheader));
1655     FC1.Preheader->getTerminator()->eraseFromParent();
1656     new UnreachableInst(FC1.Preheader->getContext(), FC1.Preheader);
1657     TreeUpdates.emplace_back(DominatorTree::UpdateType(
1658         DominatorTree::Delete, FC1.Preheader, FC1.Header));
1659 
1660     // Moves the phi nodes from the second to the first loops header block.
1661     while (PHINode *PHI = dyn_cast<PHINode>(&FC1.Header->front())) {
1662       if (SE.isSCEVable(PHI->getType()))
1663         SE.forgetValue(PHI);
1664       if (PHI->hasNUsesOrMore(1))
1665         PHI->moveBefore(&*FC0.Header->getFirstInsertionPt());
1666       else
1667         PHI->eraseFromParent();
1668     }
1669 
1670     // Introduce new phi nodes in the second loop header to ensure
1671     // exiting the first and jumping to the header of the second does not break
1672     // the SSA property of the phis originally in the first loop. See also the
1673     // comment above.
1674     Instruction *L1HeaderIP = &FC1.Header->front();
1675     for (PHINode *LCPHI : OriginalFC0PHIs) {
1676       int L1LatchBBIdx = LCPHI->getBasicBlockIndex(FC1.Latch);
1677       assert(L1LatchBBIdx >= 0 &&
1678              "Expected loop carried value to be rewired at this point!");
1679 
1680       Value *LCV = LCPHI->getIncomingValue(L1LatchBBIdx);
1681 
1682       PHINode *L1HeaderPHI = PHINode::Create(
1683           LCV->getType(), 2, LCPHI->getName() + ".afterFC0", L1HeaderIP);
1684       L1HeaderPHI->addIncoming(LCV, FC0.Latch);
1685       L1HeaderPHI->addIncoming(UndefValue::get(LCV->getType()),
1686                                FC0.ExitingBlock);
1687 
1688       LCPHI->setIncomingValue(L1LatchBBIdx, L1HeaderPHI);
1689     }
1690 
1691     // Replace latch terminator destinations.
1692     FC0.Latch->getTerminator()->replaceUsesOfWith(FC0.Header, FC1.Header);
1693     FC1.Latch->getTerminator()->replaceUsesOfWith(FC1.Header, FC0.Header);
1694 
1695     // Modify the latch branch of FC0 to be unconditional as both successors of
1696     // the branch are the same.
1697     simplifyLatchBranch(FC0);
1698 
1699     // If FC0.Latch and FC0.ExitingBlock are the same then we have already
1700     // performed the updates above.
1701     if (FC0.Latch != FC0.ExitingBlock)
1702       TreeUpdates.emplace_back(DominatorTree::UpdateType(
1703           DominatorTree::Insert, FC0.Latch, FC1.Header));
1704 
1705     TreeUpdates.emplace_back(DominatorTree::UpdateType(DominatorTree::Delete,
1706                                                        FC0.Latch, FC0.Header));
1707     TreeUpdates.emplace_back(DominatorTree::UpdateType(DominatorTree::Insert,
1708                                                        FC1.Latch, FC0.Header));
1709     TreeUpdates.emplace_back(DominatorTree::UpdateType(DominatorTree::Delete,
1710                                                        FC1.Latch, FC1.Header));
1711 
1712     // Update DT/PDT
1713     DTU.applyUpdates(TreeUpdates);
1714 
1715     LI.removeBlock(FC1.Preheader);
1716     DTU.deleteBB(FC1.Preheader);
1717     if (FC0.Peeled) {
1718       LI.removeBlock(FC0.ExitBlock);
1719       DTU.deleteBB(FC0.ExitBlock);
1720     }
1721 
1722     DTU.flush();
1723 
1724     // Is there a way to keep SE up-to-date so we don't need to forget the loops
1725     // and rebuild the information in subsequent passes of fusion?
1726     // Note: Need to forget the loops before merging the loop latches, as
1727     // mergeLatch may remove the only block in FC1.
1728     SE.forgetLoop(FC1.L);
1729     SE.forgetLoop(FC0.L);
1730     SE.forgetLoopDispositions();
1731 
1732     // Move instructions from FC0.Latch to FC1.Latch.
1733     // Note: mergeLatch requires an updated DT.
1734     mergeLatch(FC0, FC1);
1735 
1736     // Merge the loops.
1737     SmallVector<BasicBlock *, 8> Blocks(FC1.L->blocks());
1738     for (BasicBlock *BB : Blocks) {
1739       FC0.L->addBlockEntry(BB);
1740       FC1.L->removeBlockFromLoop(BB);
1741       if (LI.getLoopFor(BB) != FC1.L)
1742         continue;
1743       LI.changeLoopFor(BB, FC0.L);
1744     }
1745     while (!FC1.L->isInnermost()) {
1746       const auto &ChildLoopIt = FC1.L->begin();
1747       Loop *ChildLoop = *ChildLoopIt;
1748       FC1.L->removeChildLoop(ChildLoopIt);
1749       FC0.L->addChildLoop(ChildLoop);
1750     }
1751 
1752     // Delete the now empty loop L1.
1753     LI.erase(FC1.L);
1754 
1755 #ifndef NDEBUG
1756     assert(!verifyFunction(*FC0.Header->getParent(), &errs()));
1757     assert(DT.verify(DominatorTree::VerificationLevel::Fast));
1758     assert(PDT.verify());
1759     LI.verify(DT);
1760     SE.verify();
1761 #endif
1762 
1763     LLVM_DEBUG(dbgs() << "Fusion done:\n");
1764 
1765     return FC0.L;
1766   }
1767 
1768   /// Report details on loop fusion opportunities.
1769   ///
1770   /// This template function can be used to report both successful and missed
1771   /// loop fusion opportunities, based on the RemarkKind. The RemarkKind should
1772   /// be one of:
1773   ///   - OptimizationRemarkMissed to report when loop fusion is unsuccessful
1774   ///     given two valid fusion candidates.
1775   ///   - OptimizationRemark to report successful fusion of two fusion
1776   ///     candidates.
1777   /// The remarks will be printed using the form:
1778   ///    <path/filename>:<line number>:<column number>: [<function name>]:
1779   ///       <Cand1 Preheader> and <Cand2 Preheader>: <Stat Description>
1780   template <typename RemarkKind>
1781   void reportLoopFusion(const FusionCandidate &FC0, const FusionCandidate &FC1,
1782                         llvm::Statistic &Stat) {
1783     assert(FC0.Preheader && FC1.Preheader &&
1784            "Expecting valid fusion candidates");
1785     using namespace ore;
1786 #if LLVM_ENABLE_STATS
1787     ++Stat;
1788     ORE.emit(RemarkKind(DEBUG_TYPE, Stat.getName(), FC0.L->getStartLoc(),
1789                         FC0.Preheader)
1790              << "[" << FC0.Preheader->getParent()->getName()
1791              << "]: " << NV("Cand1", StringRef(FC0.Preheader->getName()))
1792              << " and " << NV("Cand2", StringRef(FC1.Preheader->getName()))
1793              << ": " << Stat.getDesc());
1794 #endif
1795   }
1796 
1797   /// Fuse two guarded fusion candidates, creating a new fused loop.
1798   ///
1799   /// Fusing guarded loops is handled much the same way as fusing non-guarded
1800   /// loops. The rewiring of the CFG is slightly different though, because of
1801   /// the presence of the guards around the loops and the exit blocks after the
1802   /// loop body. As such, the new loop is rewired as follows:
1803   ///    1. Keep the guard branch from FC0 and use the non-loop block target
1804   /// from the FC1 guard branch.
1805   ///    2. Remove the exit block from FC0 (this exit block should be empty
1806   /// right now).
1807   ///    3. Remove the guard branch for FC1
1808   ///    4. Remove the preheader for FC1.
1809   /// The exit block successor for the latch of FC0 is updated to be the header
1810   /// of FC1 and the non-exit block successor of the latch of FC1 is updated to
1811   /// be the header of FC0, thus creating the fused loop.
1812   Loop *fuseGuardedLoops(const FusionCandidate &FC0,
1813                          const FusionCandidate &FC1) {
1814     assert(FC0.GuardBranch && FC1.GuardBranch && "Expecting guarded loops");
1815 
1816     BasicBlock *FC0GuardBlock = FC0.GuardBranch->getParent();
1817     BasicBlock *FC1GuardBlock = FC1.GuardBranch->getParent();
1818     BasicBlock *FC0NonLoopBlock = FC0.getNonLoopBlock();
1819     BasicBlock *FC1NonLoopBlock = FC1.getNonLoopBlock();
1820     BasicBlock *FC0ExitBlockSuccessor = FC0.ExitBlock->getUniqueSuccessor();
1821 
1822     // Move instructions from the exit block of FC0 to the beginning of the exit
1823     // block of FC1, in the case that the FC0 loop has not been peeled. In the
1824     // case that FC0 loop is peeled, then move the instructions of the successor
1825     // of the FC0 Exit block to the beginning of the exit block of FC1.
1826     moveInstructionsToTheBeginning(
1827         (FC0.Peeled ? *FC0ExitBlockSuccessor : *FC0.ExitBlock), *FC1.ExitBlock,
1828         DT, PDT, DI);
1829 
1830     // Move instructions from the guard block of FC1 to the end of the guard
1831     // block of FC0.
1832     moveInstructionsToTheEnd(*FC1GuardBlock, *FC0GuardBlock, DT, PDT, DI);
1833 
1834     assert(FC0NonLoopBlock == FC1GuardBlock && "Loops are not adjacent");
1835 
1836     SmallVector<DominatorTree::UpdateType, 8> TreeUpdates;
1837 
1838     ////////////////////////////////////////////////////////////////////////////
1839     // Update the Loop Guard
1840     ////////////////////////////////////////////////////////////////////////////
1841     // The guard for FC0 is updated to guard both FC0 and FC1. This is done by
1842     // changing the NonLoopGuardBlock for FC0 to the NonLoopGuardBlock for FC1.
1843     // Thus, one path from the guard goes to the preheader for FC0 (and thus
1844     // executes the new fused loop) and the other path goes to the NonLoopBlock
1845     // for FC1 (where FC1 guard would have gone if FC1 was not executed).
1846     FC1NonLoopBlock->replacePhiUsesWith(FC1GuardBlock, FC0GuardBlock);
1847     FC0.GuardBranch->replaceUsesOfWith(FC0NonLoopBlock, FC1NonLoopBlock);
1848 
1849     BasicBlock *BBToUpdate = FC0.Peeled ? FC0ExitBlockSuccessor : FC0.ExitBlock;
1850     BBToUpdate->getTerminator()->replaceUsesOfWith(FC1GuardBlock, FC1.Header);
1851 
1852     // The guard of FC1 is not necessary anymore.
1853     FC1.GuardBranch->eraseFromParent();
1854     new UnreachableInst(FC1GuardBlock->getContext(), FC1GuardBlock);
1855 
1856     TreeUpdates.emplace_back(DominatorTree::UpdateType(
1857         DominatorTree::Delete, FC1GuardBlock, FC1.Preheader));
1858     TreeUpdates.emplace_back(DominatorTree::UpdateType(
1859         DominatorTree::Delete, FC1GuardBlock, FC1NonLoopBlock));
1860     TreeUpdates.emplace_back(DominatorTree::UpdateType(
1861         DominatorTree::Delete, FC0GuardBlock, FC1GuardBlock));
1862     TreeUpdates.emplace_back(DominatorTree::UpdateType(
1863         DominatorTree::Insert, FC0GuardBlock, FC1NonLoopBlock));
1864 
1865     if (FC0.Peeled) {
1866       // Remove the Block after the ExitBlock of FC0
1867       TreeUpdates.emplace_back(DominatorTree::UpdateType(
1868           DominatorTree::Delete, FC0ExitBlockSuccessor, FC1GuardBlock));
1869       FC0ExitBlockSuccessor->getTerminator()->eraseFromParent();
1870       new UnreachableInst(FC0ExitBlockSuccessor->getContext(),
1871                           FC0ExitBlockSuccessor);
1872     }
1873 
1874     assert(pred_empty(FC1GuardBlock) &&
1875            "Expecting guard block to have no predecessors");
1876     assert(succ_empty(FC1GuardBlock) &&
1877            "Expecting guard block to have no successors");
1878 
1879     // Remember the phi nodes originally in the header of FC0 in order to rewire
1880     // them later. However, this is only necessary if the new loop carried
1881     // values might not dominate the exiting branch. While we do not generally
1882     // test if this is the case but simply insert intermediate phi nodes, we
1883     // need to make sure these intermediate phi nodes have different
1884     // predecessors. To this end, we filter the special case where the exiting
1885     // block is the latch block of the first loop. Nothing needs to be done
1886     // anyway as all loop carried values dominate the latch and thereby also the
1887     // exiting branch.
1888     // KB: This is no longer necessary because FC0.ExitingBlock == FC0.Latch
1889     // (because the loops are rotated. Thus, nothing will ever be added to
1890     // OriginalFC0PHIs.
1891     SmallVector<PHINode *, 8> OriginalFC0PHIs;
1892     if (FC0.ExitingBlock != FC0.Latch)
1893       for (PHINode &PHI : FC0.Header->phis())
1894         OriginalFC0PHIs.push_back(&PHI);
1895 
1896     assert(OriginalFC0PHIs.empty() && "Expecting OriginalFC0PHIs to be empty!");
1897 
1898     // Replace incoming blocks for header PHIs first.
1899     FC1.Preheader->replaceSuccessorsPhiUsesWith(FC0.Preheader);
1900     FC0.Latch->replaceSuccessorsPhiUsesWith(FC1.Latch);
1901 
1902     // The old exiting block of the first loop (FC0) has to jump to the header
1903     // of the second as we need to execute the code in the second header block
1904     // regardless of the trip count. That is, if the trip count is 0, so the
1905     // back edge is never taken, we still have to execute both loop headers,
1906     // especially (but not only!) if the second is a do-while style loop.
1907     // However, doing so might invalidate the phi nodes of the first loop as
1908     // the new values do only need to dominate their latch and not the exiting
1909     // predicate. To remedy this potential problem we always introduce phi
1910     // nodes in the header of the second loop later that select the loop carried
1911     // value, if the second header was reached through an old latch of the
1912     // first, or undef otherwise. This is sound as exiting the first implies the
1913     // second will exit too, __without__ taking the back-edge (their
1914     // trip-counts are equal after all).
1915     FC0.ExitingBlock->getTerminator()->replaceUsesOfWith(FC0.ExitBlock,
1916                                                          FC1.Header);
1917 
1918     TreeUpdates.emplace_back(DominatorTree::UpdateType(
1919         DominatorTree::Delete, FC0.ExitingBlock, FC0.ExitBlock));
1920     TreeUpdates.emplace_back(DominatorTree::UpdateType(
1921         DominatorTree::Insert, FC0.ExitingBlock, FC1.Header));
1922 
1923     // Remove FC0 Exit Block
1924     // The exit block for FC0 is no longer needed since control will flow
1925     // directly to the header of FC1. Since it is an empty block, it can be
1926     // removed at this point.
1927     // TODO: In the future, we can handle non-empty exit blocks my merging any
1928     // instructions from FC0 exit block into FC1 exit block prior to removing
1929     // the block.
1930     assert(pred_empty(FC0.ExitBlock) && "Expecting exit block to be empty");
1931     FC0.ExitBlock->getTerminator()->eraseFromParent();
1932     new UnreachableInst(FC0.ExitBlock->getContext(), FC0.ExitBlock);
1933 
1934     // Remove FC1 Preheader
1935     // The pre-header of L1 is not necessary anymore.
1936     assert(pred_empty(FC1.Preheader));
1937     FC1.Preheader->getTerminator()->eraseFromParent();
1938     new UnreachableInst(FC1.Preheader->getContext(), FC1.Preheader);
1939     TreeUpdates.emplace_back(DominatorTree::UpdateType(
1940         DominatorTree::Delete, FC1.Preheader, FC1.Header));
1941 
1942     // Moves the phi nodes from the second to the first loops header block.
1943     while (PHINode *PHI = dyn_cast<PHINode>(&FC1.Header->front())) {
1944       if (SE.isSCEVable(PHI->getType()))
1945         SE.forgetValue(PHI);
1946       if (PHI->hasNUsesOrMore(1))
1947         PHI->moveBefore(&*FC0.Header->getFirstInsertionPt());
1948       else
1949         PHI->eraseFromParent();
1950     }
1951 
1952     // Introduce new phi nodes in the second loop header to ensure
1953     // exiting the first and jumping to the header of the second does not break
1954     // the SSA property of the phis originally in the first loop. See also the
1955     // comment above.
1956     Instruction *L1HeaderIP = &FC1.Header->front();
1957     for (PHINode *LCPHI : OriginalFC0PHIs) {
1958       int L1LatchBBIdx = LCPHI->getBasicBlockIndex(FC1.Latch);
1959       assert(L1LatchBBIdx >= 0 &&
1960              "Expected loop carried value to be rewired at this point!");
1961 
1962       Value *LCV = LCPHI->getIncomingValue(L1LatchBBIdx);
1963 
1964       PHINode *L1HeaderPHI = PHINode::Create(
1965           LCV->getType(), 2, LCPHI->getName() + ".afterFC0", L1HeaderIP);
1966       L1HeaderPHI->addIncoming(LCV, FC0.Latch);
1967       L1HeaderPHI->addIncoming(UndefValue::get(LCV->getType()),
1968                                FC0.ExitingBlock);
1969 
1970       LCPHI->setIncomingValue(L1LatchBBIdx, L1HeaderPHI);
1971     }
1972 
1973     // Update the latches
1974 
1975     // Replace latch terminator destinations.
1976     FC0.Latch->getTerminator()->replaceUsesOfWith(FC0.Header, FC1.Header);
1977     FC1.Latch->getTerminator()->replaceUsesOfWith(FC1.Header, FC0.Header);
1978 
1979     // Modify the latch branch of FC0 to be unconditional as both successors of
1980     // the branch are the same.
1981     simplifyLatchBranch(FC0);
1982 
1983     // If FC0.Latch and FC0.ExitingBlock are the same then we have already
1984     // performed the updates above.
1985     if (FC0.Latch != FC0.ExitingBlock)
1986       TreeUpdates.emplace_back(DominatorTree::UpdateType(
1987           DominatorTree::Insert, FC0.Latch, FC1.Header));
1988 
1989     TreeUpdates.emplace_back(DominatorTree::UpdateType(DominatorTree::Delete,
1990                                                        FC0.Latch, FC0.Header));
1991     TreeUpdates.emplace_back(DominatorTree::UpdateType(DominatorTree::Insert,
1992                                                        FC1.Latch, FC0.Header));
1993     TreeUpdates.emplace_back(DominatorTree::UpdateType(DominatorTree::Delete,
1994                                                        FC1.Latch, FC1.Header));
1995 
1996     // All done
1997     // Apply the updates to the Dominator Tree and cleanup.
1998 
1999     assert(succ_empty(FC1GuardBlock) && "FC1GuardBlock has successors!!");
2000     assert(pred_empty(FC1GuardBlock) && "FC1GuardBlock has predecessors!!");
2001 
2002     // Update DT/PDT
2003     DTU.applyUpdates(TreeUpdates);
2004 
2005     LI.removeBlock(FC1GuardBlock);
2006     LI.removeBlock(FC1.Preheader);
2007     LI.removeBlock(FC0.ExitBlock);
2008     if (FC0.Peeled) {
2009       LI.removeBlock(FC0ExitBlockSuccessor);
2010       DTU.deleteBB(FC0ExitBlockSuccessor);
2011     }
2012     DTU.deleteBB(FC1GuardBlock);
2013     DTU.deleteBB(FC1.Preheader);
2014     DTU.deleteBB(FC0.ExitBlock);
2015     DTU.flush();
2016 
2017     // Is there a way to keep SE up-to-date so we don't need to forget the loops
2018     // and rebuild the information in subsequent passes of fusion?
2019     // Note: Need to forget the loops before merging the loop latches, as
2020     // mergeLatch may remove the only block in FC1.
2021     SE.forgetLoop(FC1.L);
2022     SE.forgetLoop(FC0.L);
2023     SE.forgetLoopDispositions();
2024 
2025     // Move instructions from FC0.Latch to FC1.Latch.
2026     // Note: mergeLatch requires an updated DT.
2027     mergeLatch(FC0, FC1);
2028 
2029     // Merge the loops.
2030     SmallVector<BasicBlock *, 8> Blocks(FC1.L->blocks());
2031     for (BasicBlock *BB : Blocks) {
2032       FC0.L->addBlockEntry(BB);
2033       FC1.L->removeBlockFromLoop(BB);
2034       if (LI.getLoopFor(BB) != FC1.L)
2035         continue;
2036       LI.changeLoopFor(BB, FC0.L);
2037     }
2038     while (!FC1.L->isInnermost()) {
2039       const auto &ChildLoopIt = FC1.L->begin();
2040       Loop *ChildLoop = *ChildLoopIt;
2041       FC1.L->removeChildLoop(ChildLoopIt);
2042       FC0.L->addChildLoop(ChildLoop);
2043     }
2044 
2045     // Delete the now empty loop L1.
2046     LI.erase(FC1.L);
2047 
2048 #ifndef NDEBUG
2049     assert(!verifyFunction(*FC0.Header->getParent(), &errs()));
2050     assert(DT.verify(DominatorTree::VerificationLevel::Fast));
2051     assert(PDT.verify());
2052     LI.verify(DT);
2053     SE.verify();
2054 #endif
2055 
2056     LLVM_DEBUG(dbgs() << "Fusion done:\n");
2057 
2058     return FC0.L;
2059   }
2060 };
2061 } // namespace
2062 
2063 PreservedAnalyses LoopFusePass::run(Function &F, FunctionAnalysisManager &AM) {
2064   auto &LI = AM.getResult<LoopAnalysis>(F);
2065   auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
2066   auto &DI = AM.getResult<DependenceAnalysis>(F);
2067   auto &SE = AM.getResult<ScalarEvolutionAnalysis>(F);
2068   auto &PDT = AM.getResult<PostDominatorTreeAnalysis>(F);
2069   auto &ORE = AM.getResult<OptimizationRemarkEmitterAnalysis>(F);
2070   auto &AC = AM.getResult<AssumptionAnalysis>(F);
2071   const TargetTransformInfo &TTI = AM.getResult<TargetIRAnalysis>(F);
2072   const DataLayout &DL = F.getParent()->getDataLayout();
2073 
2074   // Ensure loops are in simplifed form which is a pre-requisite for loop fusion
2075   // pass. Added only for new PM since the legacy PM has already added
2076   // LoopSimplify pass as a dependency.
2077   bool Changed = false;
2078   for (auto &L : LI) {
2079     Changed |=
2080         simplifyLoop(L, &DT, &LI, &SE, &AC, nullptr, false /* PreserveLCSSA */);
2081   }
2082   if (Changed)
2083     PDT.recalculate(F);
2084 
2085   LoopFuser LF(LI, DT, DI, SE, PDT, ORE, DL, AC, TTI);
2086   Changed |= LF.fuseLoops(F);
2087   if (!Changed)
2088     return PreservedAnalyses::all();
2089 
2090   PreservedAnalyses PA;
2091   PA.preserve<DominatorTreeAnalysis>();
2092   PA.preserve<PostDominatorTreeAnalysis>();
2093   PA.preserve<ScalarEvolutionAnalysis>();
2094   PA.preserve<LoopAnalysis>();
2095   return PA;
2096 }
2097