1 //===- LoopInterchange.cpp - Loop interchange pass-------------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This Pass handles loop interchange transform.
10 // This pass interchanges loops to provide a more cache-friendly memory access
11 // patterns.
12 //
13 //===----------------------------------------------------------------------===//
14 
15 #include "llvm/Transforms/Scalar/LoopInterchange.h"
16 #include "llvm/ADT/STLExtras.h"
17 #include "llvm/ADT/SmallVector.h"
18 #include "llvm/ADT/Statistic.h"
19 #include "llvm/ADT/StringRef.h"
20 #include "llvm/Analysis/DependenceAnalysis.h"
21 #include "llvm/Analysis/LoopInfo.h"
22 #include "llvm/Analysis/LoopNestAnalysis.h"
23 #include "llvm/Analysis/LoopPass.h"
24 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
25 #include "llvm/Analysis/ScalarEvolution.h"
26 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
27 #include "llvm/IR/BasicBlock.h"
28 #include "llvm/IR/Constants.h"
29 #include "llvm/IR/DiagnosticInfo.h"
30 #include "llvm/IR/Dominators.h"
31 #include "llvm/IR/Function.h"
32 #include "llvm/IR/IRBuilder.h"
33 #include "llvm/IR/InstrTypes.h"
34 #include "llvm/IR/Instruction.h"
35 #include "llvm/IR/Instructions.h"
36 #include "llvm/IR/Type.h"
37 #include "llvm/IR/User.h"
38 #include "llvm/IR/Value.h"
39 #include "llvm/InitializePasses.h"
40 #include "llvm/Pass.h"
41 #include "llvm/Support/Casting.h"
42 #include "llvm/Support/CommandLine.h"
43 #include "llvm/Support/Debug.h"
44 #include "llvm/Support/ErrorHandling.h"
45 #include "llvm/Support/raw_ostream.h"
46 #include "llvm/Transforms/Scalar.h"
47 #include "llvm/Transforms/Utils.h"
48 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
49 #include "llvm/Transforms/Utils/LoopUtils.h"
50 #include <cassert>
51 #include <utility>
52 #include <vector>
53 
54 using namespace llvm;
55 
56 #define DEBUG_TYPE "loop-interchange"
57 
58 STATISTIC(LoopsInterchanged, "Number of loops interchanged");
59 
60 static cl::opt<int> LoopInterchangeCostThreshold(
61     "loop-interchange-threshold", cl::init(0), cl::Hidden,
62     cl::desc("Interchange if you gain more than this number"));
63 
64 namespace {
65 
66 using LoopVector = SmallVector<Loop *, 8>;
67 
68 // TODO: Check if we can use a sparse matrix here.
69 using CharMatrix = std::vector<std::vector<char>>;
70 
71 } // end anonymous namespace
72 
73 // Maximum number of dependencies that can be handled in the dependency matrix.
74 static const unsigned MaxMemInstrCount = 100;
75 
76 // Maximum loop depth supported.
77 static const unsigned MaxLoopNestDepth = 10;
78 
79 #ifdef DUMP_DEP_MATRICIES
80 static void printDepMatrix(CharMatrix &DepMatrix) {
81   for (auto &Row : DepMatrix) {
82     for (auto D : Row)
83       LLVM_DEBUG(dbgs() << D << " ");
84     LLVM_DEBUG(dbgs() << "\n");
85   }
86 }
87 #endif
88 
89 static bool populateDependencyMatrix(CharMatrix &DepMatrix, unsigned Level,
90                                      Loop *L, DependenceInfo *DI) {
91   using ValueVector = SmallVector<Value *, 16>;
92 
93   ValueVector MemInstr;
94 
95   // For each block.
96   for (BasicBlock *BB : L->blocks()) {
97     // Scan the BB and collect legal loads and stores.
98     for (Instruction &I : *BB) {
99       if (!isa<Instruction>(I))
100         return false;
101       if (auto *Ld = dyn_cast<LoadInst>(&I)) {
102         if (!Ld->isSimple())
103           return false;
104         MemInstr.push_back(&I);
105       } else if (auto *St = dyn_cast<StoreInst>(&I)) {
106         if (!St->isSimple())
107           return false;
108         MemInstr.push_back(&I);
109       }
110     }
111   }
112 
113   LLVM_DEBUG(dbgs() << "Found " << MemInstr.size()
114                     << " Loads and Stores to analyze\n");
115 
116   ValueVector::iterator I, IE, J, JE;
117 
118   for (I = MemInstr.begin(), IE = MemInstr.end(); I != IE; ++I) {
119     for (J = I, JE = MemInstr.end(); J != JE; ++J) {
120       std::vector<char> Dep;
121       Instruction *Src = cast<Instruction>(*I);
122       Instruction *Dst = cast<Instruction>(*J);
123       if (Src == Dst)
124         continue;
125       // Ignore Input dependencies.
126       if (isa<LoadInst>(Src) && isa<LoadInst>(Dst))
127         continue;
128       // Track Output, Flow, and Anti dependencies.
129       if (auto D = DI->depends(Src, Dst, true)) {
130         assert(D->isOrdered() && "Expected an output, flow or anti dep.");
131         LLVM_DEBUG(StringRef DepType =
132                        D->isFlow() ? "flow" : D->isAnti() ? "anti" : "output";
133                    dbgs() << "Found " << DepType
134                           << " dependency between Src and Dst\n"
135                           << " Src:" << *Src << "\n Dst:" << *Dst << '\n');
136         unsigned Levels = D->getLevels();
137         char Direction;
138         for (unsigned II = 1; II <= Levels; ++II) {
139           const SCEV *Distance = D->getDistance(II);
140           const SCEVConstant *SCEVConst =
141               dyn_cast_or_null<SCEVConstant>(Distance);
142           if (SCEVConst) {
143             const ConstantInt *CI = SCEVConst->getValue();
144             if (CI->isNegative())
145               Direction = '<';
146             else if (CI->isZero())
147               Direction = '=';
148             else
149               Direction = '>';
150             Dep.push_back(Direction);
151           } else if (D->isScalar(II)) {
152             Direction = 'S';
153             Dep.push_back(Direction);
154           } else {
155             unsigned Dir = D->getDirection(II);
156             if (Dir == Dependence::DVEntry::LT ||
157                 Dir == Dependence::DVEntry::LE)
158               Direction = '<';
159             else if (Dir == Dependence::DVEntry::GT ||
160                      Dir == Dependence::DVEntry::GE)
161               Direction = '>';
162             else if (Dir == Dependence::DVEntry::EQ)
163               Direction = '=';
164             else
165               Direction = '*';
166             Dep.push_back(Direction);
167           }
168         }
169         while (Dep.size() != Level) {
170           Dep.push_back('I');
171         }
172 
173         DepMatrix.push_back(Dep);
174         if (DepMatrix.size() > MaxMemInstrCount) {
175           LLVM_DEBUG(dbgs() << "Cannot handle more than " << MaxMemInstrCount
176                             << " dependencies inside loop\n");
177           return false;
178         }
179       }
180     }
181   }
182 
183   return true;
184 }
185 
186 // A loop is moved from index 'from' to an index 'to'. Update the Dependence
187 // matrix by exchanging the two columns.
188 static void interChangeDependencies(CharMatrix &DepMatrix, unsigned FromIndx,
189                                     unsigned ToIndx) {
190   for (unsigned I = 0, E = DepMatrix.size(); I < E; ++I)
191     std::swap(DepMatrix[I][ToIndx], DepMatrix[I][FromIndx]);
192 }
193 
194 // Checks if outermost non '=','S'or'I' dependence in the dependence matrix is
195 // '>'
196 static bool isOuterMostDepPositive(CharMatrix &DepMatrix, unsigned Row,
197                                    unsigned Column) {
198   for (unsigned i = 0; i <= Column; ++i) {
199     if (DepMatrix[Row][i] == '<')
200       return false;
201     if (DepMatrix[Row][i] == '>')
202       return true;
203   }
204   // All dependencies were '=','S' or 'I'
205   return false;
206 }
207 
208 // Checks if no dependence exist in the dependency matrix in Row before Column.
209 static bool containsNoDependence(CharMatrix &DepMatrix, unsigned Row,
210                                  unsigned Column) {
211   for (unsigned i = 0; i < Column; ++i) {
212     if (DepMatrix[Row][i] != '=' && DepMatrix[Row][i] != 'S' &&
213         DepMatrix[Row][i] != 'I')
214       return false;
215   }
216   return true;
217 }
218 
219 static bool validDepInterchange(CharMatrix &DepMatrix, unsigned Row,
220                                 unsigned OuterLoopId, char InnerDep,
221                                 char OuterDep) {
222   if (isOuterMostDepPositive(DepMatrix, Row, OuterLoopId))
223     return false;
224 
225   if (InnerDep == OuterDep)
226     return true;
227 
228   // It is legal to interchange if and only if after interchange no row has a
229   // '>' direction as the leftmost non-'='.
230 
231   if (InnerDep == '=' || InnerDep == 'S' || InnerDep == 'I')
232     return true;
233 
234   if (InnerDep == '<')
235     return true;
236 
237   if (InnerDep == '>') {
238     // If OuterLoopId represents outermost loop then interchanging will make the
239     // 1st dependency as '>'
240     if (OuterLoopId == 0)
241       return false;
242 
243     // If all dependencies before OuterloopId are '=','S'or 'I'. Then
244     // interchanging will result in this row having an outermost non '='
245     // dependency of '>'
246     if (!containsNoDependence(DepMatrix, Row, OuterLoopId))
247       return true;
248   }
249 
250   return false;
251 }
252 
253 // Checks if it is legal to interchange 2 loops.
254 // [Theorem] A permutation of the loops in a perfect nest is legal if and only
255 // if the direction matrix, after the same permutation is applied to its
256 // columns, has no ">" direction as the leftmost non-"=" direction in any row.
257 static bool isLegalToInterChangeLoops(CharMatrix &DepMatrix,
258                                       unsigned InnerLoopId,
259                                       unsigned OuterLoopId) {
260   unsigned NumRows = DepMatrix.size();
261   // For each row check if it is valid to interchange.
262   for (unsigned Row = 0; Row < NumRows; ++Row) {
263     char InnerDep = DepMatrix[Row][InnerLoopId];
264     char OuterDep = DepMatrix[Row][OuterLoopId];
265     if (InnerDep == '*' || OuterDep == '*')
266       return false;
267     if (!validDepInterchange(DepMatrix, Row, OuterLoopId, InnerDep, OuterDep))
268       return false;
269   }
270   return true;
271 }
272 
273 static LoopVector populateWorklist(Loop &L) {
274   LLVM_DEBUG(dbgs() << "Calling populateWorklist on Func: "
275                     << L.getHeader()->getParent()->getName() << " Loop: %"
276                     << L.getHeader()->getName() << '\n');
277   LoopVector LoopList;
278   Loop *CurrentLoop = &L;
279   const std::vector<Loop *> *Vec = &CurrentLoop->getSubLoops();
280   while (!Vec->empty()) {
281     // The current loop has multiple subloops in it hence it is not tightly
282     // nested.
283     // Discard all loops above it added into Worklist.
284     if (Vec->size() != 1)
285       return {};
286 
287     LoopList.push_back(CurrentLoop);
288     CurrentLoop = Vec->front();
289     Vec = &CurrentLoop->getSubLoops();
290   }
291   LoopList.push_back(CurrentLoop);
292   return LoopList;
293 }
294 
295 namespace {
296 
297 /// LoopInterchangeLegality checks if it is legal to interchange the loop.
298 class LoopInterchangeLegality {
299 public:
300   LoopInterchangeLegality(Loop *Outer, Loop *Inner, ScalarEvolution *SE,
301                           OptimizationRemarkEmitter *ORE)
302       : OuterLoop(Outer), InnerLoop(Inner), SE(SE), ORE(ORE) {}
303 
304   /// Check if the loops can be interchanged.
305   bool canInterchangeLoops(unsigned InnerLoopId, unsigned OuterLoopId,
306                            CharMatrix &DepMatrix);
307 
308   /// Discover induction PHIs in the header of \p L. Induction
309   /// PHIs are added to \p Inductions.
310   bool findInductions(Loop *L, SmallVectorImpl<PHINode *> &Inductions);
311 
312   /// Check if the loop structure is understood. We do not handle triangular
313   /// loops for now.
314   bool isLoopStructureUnderstood();
315 
316   bool currentLimitations();
317 
318   const SmallPtrSetImpl<PHINode *> &getOuterInnerReductions() const {
319     return OuterInnerReductions;
320   }
321 
322   const SmallVectorImpl<PHINode *> &getInnerLoopInductions() const {
323     return InnerLoopInductions;
324   }
325 
326 private:
327   bool tightlyNested(Loop *Outer, Loop *Inner);
328   bool containsUnsafeInstructions(BasicBlock *BB);
329 
330   /// Discover induction and reduction PHIs in the header of \p L. Induction
331   /// PHIs are added to \p Inductions, reductions are added to
332   /// OuterInnerReductions. When the outer loop is passed, the inner loop needs
333   /// to be passed as \p InnerLoop.
334   bool findInductionAndReductions(Loop *L,
335                                   SmallVector<PHINode *, 8> &Inductions,
336                                   Loop *InnerLoop);
337 
338   Loop *OuterLoop;
339   Loop *InnerLoop;
340 
341   ScalarEvolution *SE;
342 
343   /// Interface to emit optimization remarks.
344   OptimizationRemarkEmitter *ORE;
345 
346   /// Set of reduction PHIs taking part of a reduction across the inner and
347   /// outer loop.
348   SmallPtrSet<PHINode *, 4> OuterInnerReductions;
349 
350   /// Set of inner loop induction PHIs
351   SmallVector<PHINode *, 8> InnerLoopInductions;
352 };
353 
354 /// LoopInterchangeProfitability checks if it is profitable to interchange the
355 /// loop.
356 class LoopInterchangeProfitability {
357 public:
358   LoopInterchangeProfitability(Loop *Outer, Loop *Inner, ScalarEvolution *SE,
359                                OptimizationRemarkEmitter *ORE)
360       : OuterLoop(Outer), InnerLoop(Inner), SE(SE), ORE(ORE) {}
361 
362   /// Check if the loop interchange is profitable.
363   bool isProfitable(unsigned InnerLoopId, unsigned OuterLoopId,
364                     CharMatrix &DepMatrix);
365 
366 private:
367   int getInstrOrderCost();
368 
369   Loop *OuterLoop;
370   Loop *InnerLoop;
371 
372   /// Scev analysis.
373   ScalarEvolution *SE;
374 
375   /// Interface to emit optimization remarks.
376   OptimizationRemarkEmitter *ORE;
377 };
378 
379 /// LoopInterchangeTransform interchanges the loop.
380 class LoopInterchangeTransform {
381 public:
382   LoopInterchangeTransform(Loop *Outer, Loop *Inner, ScalarEvolution *SE,
383                            LoopInfo *LI, DominatorTree *DT,
384                            const LoopInterchangeLegality &LIL)
385       : OuterLoop(Outer), InnerLoop(Inner), SE(SE), LI(LI), DT(DT), LIL(LIL) {}
386 
387   /// Interchange OuterLoop and InnerLoop.
388   bool transform();
389   void restructureLoops(Loop *NewInner, Loop *NewOuter,
390                         BasicBlock *OrigInnerPreHeader,
391                         BasicBlock *OrigOuterPreHeader);
392   void removeChildLoop(Loop *OuterLoop, Loop *InnerLoop);
393 
394 private:
395   bool adjustLoopLinks();
396   bool adjustLoopBranches();
397 
398   Loop *OuterLoop;
399   Loop *InnerLoop;
400 
401   /// Scev analysis.
402   ScalarEvolution *SE;
403 
404   LoopInfo *LI;
405   DominatorTree *DT;
406 
407   const LoopInterchangeLegality &LIL;
408 };
409 
410 struct LoopInterchange {
411   ScalarEvolution *SE = nullptr;
412   LoopInfo *LI = nullptr;
413   DependenceInfo *DI = nullptr;
414   DominatorTree *DT = nullptr;
415 
416   /// Interface to emit optimization remarks.
417   OptimizationRemarkEmitter *ORE;
418 
419   LoopInterchange(ScalarEvolution *SE, LoopInfo *LI, DependenceInfo *DI,
420                   DominatorTree *DT, OptimizationRemarkEmitter *ORE)
421       : SE(SE), LI(LI), DI(DI), DT(DT), ORE(ORE) {}
422 
423   bool run(Loop *L) {
424     if (L->getParentLoop())
425       return false;
426 
427     return processLoopList(populateWorklist(*L));
428   }
429 
430   bool run(LoopNest &LN) {
431     const auto &LoopList = LN.getLoops();
432     for (unsigned I = 1; I < LoopList.size(); ++I)
433       if (LoopList[I]->getParentLoop() != LoopList[I - 1])
434         return false;
435     return processLoopList(LoopList);
436   }
437 
438   bool isComputableLoopNest(ArrayRef<Loop *> LoopList) {
439     for (Loop *L : LoopList) {
440       const SCEV *ExitCountOuter = SE->getBackedgeTakenCount(L);
441       if (isa<SCEVCouldNotCompute>(ExitCountOuter)) {
442         LLVM_DEBUG(dbgs() << "Couldn't compute backedge count\n");
443         return false;
444       }
445       if (L->getNumBackEdges() != 1) {
446         LLVM_DEBUG(dbgs() << "NumBackEdges is not equal to 1\n");
447         return false;
448       }
449       if (!L->getExitingBlock()) {
450         LLVM_DEBUG(dbgs() << "Loop doesn't have unique exit block\n");
451         return false;
452       }
453     }
454     return true;
455   }
456 
457   unsigned selectLoopForInterchange(ArrayRef<Loop *> LoopList) {
458     // TODO: Add a better heuristic to select the loop to be interchanged based
459     // on the dependence matrix. Currently we select the innermost loop.
460     return LoopList.size() - 1;
461   }
462 
463   bool processLoopList(ArrayRef<Loop *> LoopList) {
464     bool Changed = false;
465     unsigned LoopNestDepth = LoopList.size();
466     if (LoopNestDepth < 2) {
467       LLVM_DEBUG(dbgs() << "Loop doesn't contain minimum nesting level.\n");
468       return false;
469     }
470     if (LoopNestDepth > MaxLoopNestDepth) {
471       LLVM_DEBUG(dbgs() << "Cannot handle loops of depth greater than "
472                         << MaxLoopNestDepth << "\n");
473       return false;
474     }
475     if (!isComputableLoopNest(LoopList)) {
476       LLVM_DEBUG(dbgs() << "Not valid loop candidate for interchange\n");
477       return false;
478     }
479 
480     LLVM_DEBUG(dbgs() << "Processing LoopList of size = " << LoopNestDepth
481                       << "\n");
482 
483     CharMatrix DependencyMatrix;
484     Loop *OuterMostLoop = *(LoopList.begin());
485     if (!populateDependencyMatrix(DependencyMatrix, LoopNestDepth,
486                                   OuterMostLoop, DI)) {
487       LLVM_DEBUG(dbgs() << "Populating dependency matrix failed\n");
488       return false;
489     }
490 #ifdef DUMP_DEP_MATRICIES
491     LLVM_DEBUG(dbgs() << "Dependence before interchange\n");
492     printDepMatrix(DependencyMatrix);
493 #endif
494 
495     // Get the Outermost loop exit.
496     BasicBlock *LoopNestExit = OuterMostLoop->getExitBlock();
497     if (!LoopNestExit) {
498       LLVM_DEBUG(dbgs() << "OuterMostLoop needs an unique exit block");
499       return false;
500     }
501 
502     unsigned SelecLoopId = selectLoopForInterchange(LoopList);
503     // Move the selected loop outwards to the best possible position.
504     Loop *LoopToBeInterchanged = LoopList[SelecLoopId];
505     for (unsigned i = SelecLoopId; i > 0; i--) {
506       bool Interchanged = processLoop(LoopToBeInterchanged, LoopList[i - 1], i,
507                                       i - 1, DependencyMatrix);
508       if (!Interchanged)
509         return Changed;
510       // Update the DependencyMatrix
511       interChangeDependencies(DependencyMatrix, i, i - 1);
512 #ifdef DUMP_DEP_MATRICIES
513       LLVM_DEBUG(dbgs() << "Dependence after interchange\n");
514       printDepMatrix(DependencyMatrix);
515 #endif
516       Changed |= Interchanged;
517     }
518     return Changed;
519   }
520 
521   bool processLoop(Loop *InnerLoop, Loop *OuterLoop, unsigned InnerLoopId,
522                    unsigned OuterLoopId,
523                    std::vector<std::vector<char>> &DependencyMatrix) {
524     LLVM_DEBUG(dbgs() << "Processing InnerLoopId = " << InnerLoopId
525                       << " and OuterLoopId = " << OuterLoopId << "\n");
526     LoopInterchangeLegality LIL(OuterLoop, InnerLoop, SE, ORE);
527     if (!LIL.canInterchangeLoops(InnerLoopId, OuterLoopId, DependencyMatrix)) {
528       LLVM_DEBUG(dbgs() << "Not interchanging loops. Cannot prove legality.\n");
529       return false;
530     }
531     LLVM_DEBUG(dbgs() << "Loops are legal to interchange\n");
532     LoopInterchangeProfitability LIP(OuterLoop, InnerLoop, SE, ORE);
533     if (!LIP.isProfitable(InnerLoopId, OuterLoopId, DependencyMatrix)) {
534       LLVM_DEBUG(dbgs() << "Interchanging loops not profitable.\n");
535       return false;
536     }
537 
538     ORE->emit([&]() {
539       return OptimizationRemark(DEBUG_TYPE, "Interchanged",
540                                 InnerLoop->getStartLoc(),
541                                 InnerLoop->getHeader())
542              << "Loop interchanged with enclosing loop.";
543     });
544 
545     LoopInterchangeTransform LIT(OuterLoop, InnerLoop, SE, LI, DT, LIL);
546     LIT.transform();
547     LLVM_DEBUG(dbgs() << "Loops interchanged.\n");
548     LoopsInterchanged++;
549 
550     assert(InnerLoop->isLCSSAForm(*DT) &&
551            "Inner loop not left in LCSSA form after loop interchange!");
552     assert(OuterLoop->isLCSSAForm(*DT) &&
553            "Outer loop not left in LCSSA form after loop interchange!");
554 
555     return true;
556   }
557 };
558 
559 } // end anonymous namespace
560 
561 bool LoopInterchangeLegality::containsUnsafeInstructions(BasicBlock *BB) {
562   return any_of(*BB, [](const Instruction &I) {
563     return I.mayHaveSideEffects() || I.mayReadFromMemory();
564   });
565 }
566 
567 bool LoopInterchangeLegality::tightlyNested(Loop *OuterLoop, Loop *InnerLoop) {
568   BasicBlock *OuterLoopHeader = OuterLoop->getHeader();
569   BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader();
570   BasicBlock *OuterLoopLatch = OuterLoop->getLoopLatch();
571 
572   LLVM_DEBUG(dbgs() << "Checking if loops are tightly nested\n");
573 
574   // A perfectly nested loop will not have any branch in between the outer and
575   // inner block i.e. outer header will branch to either inner preheader and
576   // outerloop latch.
577   BranchInst *OuterLoopHeaderBI =
578       dyn_cast<BranchInst>(OuterLoopHeader->getTerminator());
579   if (!OuterLoopHeaderBI)
580     return false;
581 
582   for (BasicBlock *Succ : successors(OuterLoopHeaderBI))
583     if (Succ != InnerLoopPreHeader && Succ != InnerLoop->getHeader() &&
584         Succ != OuterLoopLatch)
585       return false;
586 
587   LLVM_DEBUG(dbgs() << "Checking instructions in Loop header and Loop latch\n");
588   // We do not have any basic block in between now make sure the outer header
589   // and outer loop latch doesn't contain any unsafe instructions.
590   if (containsUnsafeInstructions(OuterLoopHeader) ||
591       containsUnsafeInstructions(OuterLoopLatch))
592     return false;
593 
594   // Also make sure the inner loop preheader does not contain any unsafe
595   // instructions. Note that all instructions in the preheader will be moved to
596   // the outer loop header when interchanging.
597   if (InnerLoopPreHeader != OuterLoopHeader &&
598       containsUnsafeInstructions(InnerLoopPreHeader))
599     return false;
600 
601   BasicBlock *InnerLoopExit = InnerLoop->getExitBlock();
602   // Ensure the inner loop exit block flows to the outer loop latch possibly
603   // through empty blocks.
604   const BasicBlock &SuccInner =
605       LoopNest::skipEmptyBlockUntil(InnerLoopExit, OuterLoopLatch);
606   if (&SuccInner != OuterLoopLatch) {
607     LLVM_DEBUG(dbgs() << "Inner loop exit block " << *InnerLoopExit
608                       << " does not lead to the outer loop latch.\n";);
609     return false;
610   }
611   // The inner loop exit block does flow to the outer loop latch and not some
612   // other BBs, now make sure it contains safe instructions, since it will be
613   // moved into the (new) inner loop after interchange.
614   if (containsUnsafeInstructions(InnerLoopExit))
615     return false;
616 
617   LLVM_DEBUG(dbgs() << "Loops are perfectly nested\n");
618   // We have a perfect loop nest.
619   return true;
620 }
621 
622 bool LoopInterchangeLegality::isLoopStructureUnderstood() {
623   BasicBlock *InnerLoopPreheader = InnerLoop->getLoopPreheader();
624   for (PHINode *InnerInduction : InnerLoopInductions) {
625     unsigned Num = InnerInduction->getNumOperands();
626     for (unsigned i = 0; i < Num; ++i) {
627       Value *Val = InnerInduction->getOperand(i);
628       if (isa<Constant>(Val))
629         continue;
630       Instruction *I = dyn_cast<Instruction>(Val);
631       if (!I)
632         return false;
633       // TODO: Handle triangular loops.
634       // e.g. for(int i=0;i<N;i++)
635       //        for(int j=i;j<N;j++)
636       unsigned IncomBlockIndx = PHINode::getIncomingValueNumForOperand(i);
637       if (InnerInduction->getIncomingBlock(IncomBlockIndx) ==
638               InnerLoopPreheader &&
639           !OuterLoop->isLoopInvariant(I)) {
640         return false;
641       }
642     }
643   }
644 
645   // TODO: Handle triangular loops of another form.
646   // e.g. for(int i=0;i<N;i++)
647   //        for(int j=0;j<i;j++)
648   // or,
649   //      for(int i=0;i<N;i++)
650   //        for(int j=0;j*i<N;j++)
651   BasicBlock *InnerLoopLatch = InnerLoop->getLoopLatch();
652   BranchInst *InnerLoopLatchBI =
653       dyn_cast<BranchInst>(InnerLoopLatch->getTerminator());
654   if (!InnerLoopLatchBI->isConditional())
655     return false;
656   if (CmpInst *InnerLoopCmp =
657           dyn_cast<CmpInst>(InnerLoopLatchBI->getCondition())) {
658     Value *Op0 = InnerLoopCmp->getOperand(0);
659     Value *Op1 = InnerLoopCmp->getOperand(1);
660 
661     // LHS and RHS of the inner loop exit condition, e.g.,
662     // in "for(int j=0;j<i;j++)", LHS is j and RHS is i.
663     Value *Left = nullptr;
664     Value *Right = nullptr;
665 
666     // Check if V only involves inner loop induction variable.
667     // Return true if V is InnerInduction, or a cast from
668     // InnerInduction, or a binary operator that involves
669     // InnerInduction and a constant.
670     std::function<bool(Value *)> IsPathToInnerIndVar;
671     IsPathToInnerIndVar = [this, &IsPathToInnerIndVar](const Value *V) -> bool {
672       if (llvm::is_contained(InnerLoopInductions, V))
673         return true;
674       if (isa<Constant>(V))
675         return true;
676       const Instruction *I = dyn_cast<Instruction>(V);
677       if (!I)
678         return false;
679       if (isa<CastInst>(I))
680         return IsPathToInnerIndVar(I->getOperand(0));
681       if (isa<BinaryOperator>(I))
682         return IsPathToInnerIndVar(I->getOperand(0)) &&
683                IsPathToInnerIndVar(I->getOperand(1));
684       return false;
685     };
686 
687     // In case of multiple inner loop indvars, it is okay if LHS and RHS
688     // are both inner indvar related variables.
689     if (IsPathToInnerIndVar(Op0) && IsPathToInnerIndVar(Op1))
690       return true;
691 
692     // Otherwise we check if the cmp instruction compares an inner indvar
693     // related variable (Left) with a outer loop invariant (Right).
694     if (IsPathToInnerIndVar(Op0) && !isa<Constant>(Op0)) {
695       Left = Op0;
696       Right = Op1;
697     } else if (IsPathToInnerIndVar(Op1) && !isa<Constant>(Op1)) {
698       Left = Op1;
699       Right = Op0;
700     }
701 
702     if (Left == nullptr)
703       return false;
704 
705     const SCEV *S = SE->getSCEV(Right);
706     if (!SE->isLoopInvariant(S, OuterLoop))
707       return false;
708   }
709 
710   return true;
711 }
712 
713 // If SV is a LCSSA PHI node with a single incoming value, return the incoming
714 // value.
715 static Value *followLCSSA(Value *SV) {
716   PHINode *PHI = dyn_cast<PHINode>(SV);
717   if (!PHI)
718     return SV;
719 
720   if (PHI->getNumIncomingValues() != 1)
721     return SV;
722   return followLCSSA(PHI->getIncomingValue(0));
723 }
724 
725 // Check V's users to see if it is involved in a reduction in L.
726 static PHINode *findInnerReductionPhi(Loop *L, Value *V) {
727   // Reduction variables cannot be constants.
728   if (isa<Constant>(V))
729     return nullptr;
730 
731   for (Value *User : V->users()) {
732     if (PHINode *PHI = dyn_cast<PHINode>(User)) {
733       if (PHI->getNumIncomingValues() == 1)
734         continue;
735       RecurrenceDescriptor RD;
736       if (RecurrenceDescriptor::isReductionPHI(PHI, L, RD))
737         return PHI;
738       return nullptr;
739     }
740   }
741 
742   return nullptr;
743 }
744 
745 bool LoopInterchangeLegality::findInductionAndReductions(
746     Loop *L, SmallVector<PHINode *, 8> &Inductions, Loop *InnerLoop) {
747   if (!L->getLoopLatch() || !L->getLoopPredecessor())
748     return false;
749   for (PHINode &PHI : L->getHeader()->phis()) {
750     RecurrenceDescriptor RD;
751     InductionDescriptor ID;
752     if (InductionDescriptor::isInductionPHI(&PHI, L, SE, ID))
753       Inductions.push_back(&PHI);
754     else {
755       // PHIs in inner loops need to be part of a reduction in the outer loop,
756       // discovered when checking the PHIs of the outer loop earlier.
757       if (!InnerLoop) {
758         if (!OuterInnerReductions.count(&PHI)) {
759           LLVM_DEBUG(dbgs() << "Inner loop PHI is not part of reductions "
760                                "across the outer loop.\n");
761           return false;
762         }
763       } else {
764         assert(PHI.getNumIncomingValues() == 2 &&
765                "Phis in loop header should have exactly 2 incoming values");
766         // Check if we have a PHI node in the outer loop that has a reduction
767         // result from the inner loop as an incoming value.
768         Value *V = followLCSSA(PHI.getIncomingValueForBlock(L->getLoopLatch()));
769         PHINode *InnerRedPhi = findInnerReductionPhi(InnerLoop, V);
770         if (!InnerRedPhi ||
771             !llvm::is_contained(InnerRedPhi->incoming_values(), &PHI)) {
772           LLVM_DEBUG(
773               dbgs()
774               << "Failed to recognize PHI as an induction or reduction.\n");
775           return false;
776         }
777         OuterInnerReductions.insert(&PHI);
778         OuterInnerReductions.insert(InnerRedPhi);
779       }
780     }
781   }
782   return true;
783 }
784 
785 // This function indicates the current limitations in the transform as a result
786 // of which we do not proceed.
787 bool LoopInterchangeLegality::currentLimitations() {
788   BasicBlock *InnerLoopLatch = InnerLoop->getLoopLatch();
789 
790   // transform currently expects the loop latches to also be the exiting
791   // blocks.
792   if (InnerLoop->getExitingBlock() != InnerLoopLatch ||
793       OuterLoop->getExitingBlock() != OuterLoop->getLoopLatch() ||
794       !isa<BranchInst>(InnerLoopLatch->getTerminator()) ||
795       !isa<BranchInst>(OuterLoop->getLoopLatch()->getTerminator())) {
796     LLVM_DEBUG(
797         dbgs() << "Loops where the latch is not the exiting block are not"
798                << " supported currently.\n");
799     ORE->emit([&]() {
800       return OptimizationRemarkMissed(DEBUG_TYPE, "ExitingNotLatch",
801                                       OuterLoop->getStartLoc(),
802                                       OuterLoop->getHeader())
803              << "Loops where the latch is not the exiting block cannot be"
804                 " interchange currently.";
805     });
806     return true;
807   }
808 
809   SmallVector<PHINode *, 8> Inductions;
810   if (!findInductionAndReductions(OuterLoop, Inductions, InnerLoop)) {
811     LLVM_DEBUG(
812         dbgs() << "Only outer loops with induction or reduction PHI nodes "
813                << "are supported currently.\n");
814     ORE->emit([&]() {
815       return OptimizationRemarkMissed(DEBUG_TYPE, "UnsupportedPHIOuter",
816                                       OuterLoop->getStartLoc(),
817                                       OuterLoop->getHeader())
818              << "Only outer loops with induction or reduction PHI nodes can be"
819                 " interchanged currently.";
820     });
821     return true;
822   }
823 
824   Inductions.clear();
825   if (!findInductionAndReductions(InnerLoop, Inductions, nullptr)) {
826     LLVM_DEBUG(
827         dbgs() << "Only inner loops with induction or reduction PHI nodes "
828                << "are supported currently.\n");
829     ORE->emit([&]() {
830       return OptimizationRemarkMissed(DEBUG_TYPE, "UnsupportedPHIInner",
831                                       InnerLoop->getStartLoc(),
832                                       InnerLoop->getHeader())
833              << "Only inner loops with induction or reduction PHI nodes can be"
834                 " interchange currently.";
835     });
836     return true;
837   }
838 
839   // TODO: Triangular loops are not handled for now.
840   if (!isLoopStructureUnderstood()) {
841     LLVM_DEBUG(dbgs() << "Loop structure not understood by pass\n");
842     ORE->emit([&]() {
843       return OptimizationRemarkMissed(DEBUG_TYPE, "UnsupportedStructureInner",
844                                       InnerLoop->getStartLoc(),
845                                       InnerLoop->getHeader())
846              << "Inner loop structure not understood currently.";
847     });
848     return true;
849   }
850 
851   return false;
852 }
853 
854 bool LoopInterchangeLegality::findInductions(
855     Loop *L, SmallVectorImpl<PHINode *> &Inductions) {
856   for (PHINode &PHI : L->getHeader()->phis()) {
857     InductionDescriptor ID;
858     if (InductionDescriptor::isInductionPHI(&PHI, L, SE, ID))
859       Inductions.push_back(&PHI);
860   }
861   return !Inductions.empty();
862 }
863 
864 // We currently only support LCSSA PHI nodes in the inner loop exit, if their
865 // users are either reduction PHIs or PHIs outside the outer loop (which means
866 // the we are only interested in the final value after the loop).
867 static bool
868 areInnerLoopExitPHIsSupported(Loop *InnerL, Loop *OuterL,
869                               SmallPtrSetImpl<PHINode *> &Reductions) {
870   BasicBlock *InnerExit = OuterL->getUniqueExitBlock();
871   for (PHINode &PHI : InnerExit->phis()) {
872     // Reduction lcssa phi will have only 1 incoming block that from loop latch.
873     if (PHI.getNumIncomingValues() > 1)
874       return false;
875     if (any_of(PHI.users(), [&Reductions, OuterL](User *U) {
876           PHINode *PN = dyn_cast<PHINode>(U);
877           return !PN ||
878                  (!Reductions.count(PN) && OuterL->contains(PN->getParent()));
879         })) {
880       return false;
881     }
882   }
883   return true;
884 }
885 
886 // We currently support LCSSA PHI nodes in the outer loop exit, if their
887 // incoming values do not come from the outer loop latch or if the
888 // outer loop latch has a single predecessor. In that case, the value will
889 // be available if both the inner and outer loop conditions are true, which
890 // will still be true after interchanging. If we have multiple predecessor,
891 // that may not be the case, e.g. because the outer loop latch may be executed
892 // if the inner loop is not executed.
893 static bool areOuterLoopExitPHIsSupported(Loop *OuterLoop, Loop *InnerLoop) {
894   BasicBlock *LoopNestExit = OuterLoop->getUniqueExitBlock();
895   for (PHINode &PHI : LoopNestExit->phis()) {
896     //  FIXME: We currently are not able to detect floating point reductions
897     //         and have to use floating point PHIs as a proxy to prevent
898     //         interchanging in the presence of floating point reductions.
899     if (PHI.getType()->isFloatingPointTy())
900       return false;
901     for (unsigned i = 0; i < PHI.getNumIncomingValues(); i++) {
902      Instruction *IncomingI = dyn_cast<Instruction>(PHI.getIncomingValue(i));
903      if (!IncomingI || IncomingI->getParent() != OuterLoop->getLoopLatch())
904        continue;
905 
906      // The incoming value is defined in the outer loop latch. Currently we
907      // only support that in case the outer loop latch has a single predecessor.
908      // This guarantees that the outer loop latch is executed if and only if
909      // the inner loop is executed (because tightlyNested() guarantees that the
910      // outer loop header only branches to the inner loop or the outer loop
911      // latch).
912      // FIXME: We could weaken this logic and allow multiple predecessors,
913      //        if the values are produced outside the loop latch. We would need
914      //        additional logic to update the PHI nodes in the exit block as
915      //        well.
916      if (OuterLoop->getLoopLatch()->getUniquePredecessor() == nullptr)
917        return false;
918     }
919   }
920   return true;
921 }
922 
923 // In case of multi-level nested loops, it may occur that lcssa phis exist in
924 // the latch of InnerLoop, i.e., when defs of the incoming values are further
925 // inside the loopnest. Sometimes those incoming values are not available
926 // after interchange, since the original inner latch will become the new outer
927 // latch which may have predecessor paths that do not include those incoming
928 // values.
929 // TODO: Handle transformation of lcssa phis in the InnerLoop latch in case of
930 // multi-level loop nests.
931 static bool areInnerLoopLatchPHIsSupported(Loop *OuterLoop, Loop *InnerLoop) {
932   if (InnerLoop->getSubLoops().empty())
933     return true;
934   // If the original outer latch has only one predecessor, then values defined
935   // further inside the looploop, e.g., in the innermost loop, will be available
936   // at the new outer latch after interchange.
937   if (OuterLoop->getLoopLatch()->getUniquePredecessor() != nullptr)
938     return true;
939 
940   // The outer latch has more than one predecessors, i.e., the inner
941   // exit and the inner header.
942   // PHI nodes in the inner latch are lcssa phis where the incoming values
943   // are defined further inside the loopnest. Check if those phis are used
944   // in the original inner latch. If that is the case then bail out since
945   // those incoming values may not be available at the new outer latch.
946   BasicBlock *InnerLoopLatch = InnerLoop->getLoopLatch();
947   for (PHINode &PHI : InnerLoopLatch->phis()) {
948     for (auto *U : PHI.users()) {
949       Instruction *UI = cast<Instruction>(U);
950       if (InnerLoopLatch == UI->getParent())
951         return false;
952     }
953   }
954   return true;
955 }
956 
957 bool LoopInterchangeLegality::canInterchangeLoops(unsigned InnerLoopId,
958                                                   unsigned OuterLoopId,
959                                                   CharMatrix &DepMatrix) {
960   if (!isLegalToInterChangeLoops(DepMatrix, InnerLoopId, OuterLoopId)) {
961     LLVM_DEBUG(dbgs() << "Failed interchange InnerLoopId = " << InnerLoopId
962                       << " and OuterLoopId = " << OuterLoopId
963                       << " due to dependence\n");
964     ORE->emit([&]() {
965       return OptimizationRemarkMissed(DEBUG_TYPE, "Dependence",
966                                       InnerLoop->getStartLoc(),
967                                       InnerLoop->getHeader())
968              << "Cannot interchange loops due to dependences.";
969     });
970     return false;
971   }
972   // Check if outer and inner loop contain legal instructions only.
973   for (auto *BB : OuterLoop->blocks())
974     for (Instruction &I : BB->instructionsWithoutDebug())
975       if (CallInst *CI = dyn_cast<CallInst>(&I)) {
976         // readnone functions do not prevent interchanging.
977         if (CI->onlyWritesMemory())
978           continue;
979         LLVM_DEBUG(
980             dbgs() << "Loops with call instructions cannot be interchanged "
981                    << "safely.");
982         ORE->emit([&]() {
983           return OptimizationRemarkMissed(DEBUG_TYPE, "CallInst",
984                                           CI->getDebugLoc(),
985                                           CI->getParent())
986                  << "Cannot interchange loops due to call instruction.";
987         });
988 
989         return false;
990       }
991 
992   if (!findInductions(InnerLoop, InnerLoopInductions)) {
993     LLVM_DEBUG(dbgs() << "Cound not find inner loop induction variables.\n");
994     return false;
995   }
996 
997   if (!areInnerLoopLatchPHIsSupported(OuterLoop, InnerLoop)) {
998     LLVM_DEBUG(dbgs() << "Found unsupported PHI nodes in inner loop latch.\n");
999     ORE->emit([&]() {
1000       return OptimizationRemarkMissed(DEBUG_TYPE, "UnsupportedInnerLatchPHI",
1001                                       InnerLoop->getStartLoc(),
1002                                       InnerLoop->getHeader())
1003              << "Cannot interchange loops because unsupported PHI nodes found "
1004                 "in inner loop latch.";
1005     });
1006     return false;
1007   }
1008 
1009   // TODO: The loops could not be interchanged due to current limitations in the
1010   // transform module.
1011   if (currentLimitations()) {
1012     LLVM_DEBUG(dbgs() << "Not legal because of current transform limitation\n");
1013     return false;
1014   }
1015 
1016   // Check if the loops are tightly nested.
1017   if (!tightlyNested(OuterLoop, InnerLoop)) {
1018     LLVM_DEBUG(dbgs() << "Loops not tightly nested\n");
1019     ORE->emit([&]() {
1020       return OptimizationRemarkMissed(DEBUG_TYPE, "NotTightlyNested",
1021                                       InnerLoop->getStartLoc(),
1022                                       InnerLoop->getHeader())
1023              << "Cannot interchange loops because they are not tightly "
1024                 "nested.";
1025     });
1026     return false;
1027   }
1028 
1029   if (!areInnerLoopExitPHIsSupported(OuterLoop, InnerLoop,
1030                                      OuterInnerReductions)) {
1031     LLVM_DEBUG(dbgs() << "Found unsupported PHI nodes in inner loop exit.\n");
1032     ORE->emit([&]() {
1033       return OptimizationRemarkMissed(DEBUG_TYPE, "UnsupportedExitPHI",
1034                                       InnerLoop->getStartLoc(),
1035                                       InnerLoop->getHeader())
1036              << "Found unsupported PHI node in loop exit.";
1037     });
1038     return false;
1039   }
1040 
1041   if (!areOuterLoopExitPHIsSupported(OuterLoop, InnerLoop)) {
1042     LLVM_DEBUG(dbgs() << "Found unsupported PHI nodes in outer loop exit.\n");
1043     ORE->emit([&]() {
1044       return OptimizationRemarkMissed(DEBUG_TYPE, "UnsupportedExitPHI",
1045                                       OuterLoop->getStartLoc(),
1046                                       OuterLoop->getHeader())
1047              << "Found unsupported PHI node in loop exit.";
1048     });
1049     return false;
1050   }
1051 
1052   return true;
1053 }
1054 
1055 int LoopInterchangeProfitability::getInstrOrderCost() {
1056   unsigned GoodOrder, BadOrder;
1057   BadOrder = GoodOrder = 0;
1058   for (BasicBlock *BB : InnerLoop->blocks()) {
1059     for (Instruction &Ins : *BB) {
1060       if (const GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(&Ins)) {
1061         unsigned NumOp = GEP->getNumOperands();
1062         bool FoundInnerInduction = false;
1063         bool FoundOuterInduction = false;
1064         for (unsigned i = 0; i < NumOp; ++i) {
1065           // Skip operands that are not SCEV-able.
1066           if (!SE->isSCEVable(GEP->getOperand(i)->getType()))
1067             continue;
1068 
1069           const SCEV *OperandVal = SE->getSCEV(GEP->getOperand(i));
1070           const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(OperandVal);
1071           if (!AR)
1072             continue;
1073 
1074           // If we find the inner induction after an outer induction e.g.
1075           // for(int i=0;i<N;i++)
1076           //   for(int j=0;j<N;j++)
1077           //     A[i][j] = A[i-1][j-1]+k;
1078           // then it is a good order.
1079           if (AR->getLoop() == InnerLoop) {
1080             // We found an InnerLoop induction after OuterLoop induction. It is
1081             // a good order.
1082             FoundInnerInduction = true;
1083             if (FoundOuterInduction) {
1084               GoodOrder++;
1085               break;
1086             }
1087           }
1088           // If we find the outer induction after an inner induction e.g.
1089           // for(int i=0;i<N;i++)
1090           //   for(int j=0;j<N;j++)
1091           //     A[j][i] = A[j-1][i-1]+k;
1092           // then it is a bad order.
1093           if (AR->getLoop() == OuterLoop) {
1094             // We found an OuterLoop induction after InnerLoop induction. It is
1095             // a bad order.
1096             FoundOuterInduction = true;
1097             if (FoundInnerInduction) {
1098               BadOrder++;
1099               break;
1100             }
1101           }
1102         }
1103       }
1104     }
1105   }
1106   return GoodOrder - BadOrder;
1107 }
1108 
1109 static bool isProfitableForVectorization(unsigned InnerLoopId,
1110                                          unsigned OuterLoopId,
1111                                          CharMatrix &DepMatrix) {
1112   // TODO: Improve this heuristic to catch more cases.
1113   // If the inner loop is loop independent or doesn't carry any dependency it is
1114   // profitable to move this to outer position.
1115   for (auto &Row : DepMatrix) {
1116     if (Row[InnerLoopId] != 'S' && Row[InnerLoopId] != 'I')
1117       return false;
1118     // TODO: We need to improve this heuristic.
1119     if (Row[OuterLoopId] != '=')
1120       return false;
1121   }
1122   // If outer loop has dependence and inner loop is loop independent then it is
1123   // profitable to interchange to enable parallelism.
1124   // If there are no dependences, interchanging will not improve anything.
1125   return !DepMatrix.empty();
1126 }
1127 
1128 bool LoopInterchangeProfitability::isProfitable(unsigned InnerLoopId,
1129                                                 unsigned OuterLoopId,
1130                                                 CharMatrix &DepMatrix) {
1131   // TODO: Add better profitability checks.
1132   // e.g
1133   // 1) Construct dependency matrix and move the one with no loop carried dep
1134   //    inside to enable vectorization.
1135 
1136   // This is rough cost estimation algorithm. It counts the good and bad order
1137   // of induction variables in the instruction and allows reordering if number
1138   // of bad orders is more than good.
1139   int Cost = getInstrOrderCost();
1140   LLVM_DEBUG(dbgs() << "Cost = " << Cost << "\n");
1141   if (Cost < -LoopInterchangeCostThreshold)
1142     return true;
1143 
1144   // It is not profitable as per current cache profitability model. But check if
1145   // we can move this loop outside to improve parallelism.
1146   if (isProfitableForVectorization(InnerLoopId, OuterLoopId, DepMatrix))
1147     return true;
1148 
1149   ORE->emit([&]() {
1150     return OptimizationRemarkMissed(DEBUG_TYPE, "InterchangeNotProfitable",
1151                                     InnerLoop->getStartLoc(),
1152                                     InnerLoop->getHeader())
1153            << "Interchanging loops is too costly (cost="
1154            << ore::NV("Cost", Cost) << ", threshold="
1155            << ore::NV("Threshold", LoopInterchangeCostThreshold)
1156            << ") and it does not improve parallelism.";
1157   });
1158   return false;
1159 }
1160 
1161 void LoopInterchangeTransform::removeChildLoop(Loop *OuterLoop,
1162                                                Loop *InnerLoop) {
1163   for (Loop *L : *OuterLoop)
1164     if (L == InnerLoop) {
1165       OuterLoop->removeChildLoop(L);
1166       return;
1167     }
1168   llvm_unreachable("Couldn't find loop");
1169 }
1170 
1171 /// Update LoopInfo, after interchanging. NewInner and NewOuter refer to the
1172 /// new inner and outer loop after interchanging: NewInner is the original
1173 /// outer loop and NewOuter is the original inner loop.
1174 ///
1175 /// Before interchanging, we have the following structure
1176 /// Outer preheader
1177 //  Outer header
1178 //    Inner preheader
1179 //    Inner header
1180 //      Inner body
1181 //      Inner latch
1182 //   outer bbs
1183 //   Outer latch
1184 //
1185 // After interchanging:
1186 // Inner preheader
1187 // Inner header
1188 //   Outer preheader
1189 //   Outer header
1190 //     Inner body
1191 //     outer bbs
1192 //     Outer latch
1193 //   Inner latch
1194 void LoopInterchangeTransform::restructureLoops(
1195     Loop *NewInner, Loop *NewOuter, BasicBlock *OrigInnerPreHeader,
1196     BasicBlock *OrigOuterPreHeader) {
1197   Loop *OuterLoopParent = OuterLoop->getParentLoop();
1198   // The original inner loop preheader moves from the new inner loop to
1199   // the parent loop, if there is one.
1200   NewInner->removeBlockFromLoop(OrigInnerPreHeader);
1201   LI->changeLoopFor(OrigInnerPreHeader, OuterLoopParent);
1202 
1203   // Switch the loop levels.
1204   if (OuterLoopParent) {
1205     // Remove the loop from its parent loop.
1206     removeChildLoop(OuterLoopParent, NewInner);
1207     removeChildLoop(NewInner, NewOuter);
1208     OuterLoopParent->addChildLoop(NewOuter);
1209   } else {
1210     removeChildLoop(NewInner, NewOuter);
1211     LI->changeTopLevelLoop(NewInner, NewOuter);
1212   }
1213   while (!NewOuter->isInnermost())
1214     NewInner->addChildLoop(NewOuter->removeChildLoop(NewOuter->begin()));
1215   NewOuter->addChildLoop(NewInner);
1216 
1217   // BBs from the original inner loop.
1218   SmallVector<BasicBlock *, 8> OrigInnerBBs(NewOuter->blocks());
1219 
1220   // Add BBs from the original outer loop to the original inner loop (excluding
1221   // BBs already in inner loop)
1222   for (BasicBlock *BB : NewInner->blocks())
1223     if (LI->getLoopFor(BB) == NewInner)
1224       NewOuter->addBlockEntry(BB);
1225 
1226   // Now remove inner loop header and latch from the new inner loop and move
1227   // other BBs (the loop body) to the new inner loop.
1228   BasicBlock *OuterHeader = NewOuter->getHeader();
1229   BasicBlock *OuterLatch = NewOuter->getLoopLatch();
1230   for (BasicBlock *BB : OrigInnerBBs) {
1231     // Nothing will change for BBs in child loops.
1232     if (LI->getLoopFor(BB) != NewOuter)
1233       continue;
1234     // Remove the new outer loop header and latch from the new inner loop.
1235     if (BB == OuterHeader || BB == OuterLatch)
1236       NewInner->removeBlockFromLoop(BB);
1237     else
1238       LI->changeLoopFor(BB, NewInner);
1239   }
1240 
1241   // The preheader of the original outer loop becomes part of the new
1242   // outer loop.
1243   NewOuter->addBlockEntry(OrigOuterPreHeader);
1244   LI->changeLoopFor(OrigOuterPreHeader, NewOuter);
1245 
1246   // Tell SE that we move the loops around.
1247   SE->forgetLoop(NewOuter);
1248   SE->forgetLoop(NewInner);
1249 }
1250 
1251 bool LoopInterchangeTransform::transform() {
1252   bool Transformed = false;
1253 
1254   if (InnerLoop->getSubLoops().empty()) {
1255     BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader();
1256     LLVM_DEBUG(dbgs() << "Splitting the inner loop latch\n");
1257     auto &InductionPHIs = LIL.getInnerLoopInductions();
1258     if (InductionPHIs.empty()) {
1259       LLVM_DEBUG(dbgs() << "Failed to find the point to split loop latch \n");
1260       return false;
1261     }
1262 
1263     SmallVector<Instruction *, 8> InnerIndexVarList;
1264     for (PHINode *CurInductionPHI : InductionPHIs) {
1265       if (CurInductionPHI->getIncomingBlock(0) == InnerLoopPreHeader)
1266         InnerIndexVarList.push_back(
1267             dyn_cast<Instruction>(CurInductionPHI->getIncomingValue(1)));
1268       else
1269         InnerIndexVarList.push_back(
1270             dyn_cast<Instruction>(CurInductionPHI->getIncomingValue(0)));
1271     }
1272 
1273     // Create a new latch block for the inner loop. We split at the
1274     // current latch's terminator and then move the condition and all
1275     // operands that are not either loop-invariant or the induction PHI into the
1276     // new latch block.
1277     BasicBlock *NewLatch =
1278         SplitBlock(InnerLoop->getLoopLatch(),
1279                    InnerLoop->getLoopLatch()->getTerminator(), DT, LI);
1280 
1281     SmallSetVector<Instruction *, 4> WorkList;
1282     unsigned i = 0;
1283     auto MoveInstructions = [&i, &WorkList, this, &InductionPHIs, NewLatch]() {
1284       for (; i < WorkList.size(); i++) {
1285         // Duplicate instruction and move it the new latch. Update uses that
1286         // have been moved.
1287         Instruction *NewI = WorkList[i]->clone();
1288         NewI->insertBefore(NewLatch->getFirstNonPHI());
1289         assert(!NewI->mayHaveSideEffects() &&
1290                "Moving instructions with side-effects may change behavior of "
1291                "the loop nest!");
1292         for (Use &U : llvm::make_early_inc_range(WorkList[i]->uses())) {
1293           Instruction *UserI = cast<Instruction>(U.getUser());
1294           if (!InnerLoop->contains(UserI->getParent()) ||
1295               UserI->getParent() == NewLatch ||
1296               llvm::is_contained(InductionPHIs, UserI))
1297             U.set(NewI);
1298         }
1299         // Add operands of moved instruction to the worklist, except if they are
1300         // outside the inner loop or are the induction PHI.
1301         for (Value *Op : WorkList[i]->operands()) {
1302           Instruction *OpI = dyn_cast<Instruction>(Op);
1303           if (!OpI ||
1304               this->LI->getLoopFor(OpI->getParent()) != this->InnerLoop ||
1305               llvm::is_contained(InductionPHIs, OpI))
1306             continue;
1307           WorkList.insert(OpI);
1308         }
1309       }
1310     };
1311 
1312     // FIXME: Should we interchange when we have a constant condition?
1313     Instruction *CondI = dyn_cast<Instruction>(
1314         cast<BranchInst>(InnerLoop->getLoopLatch()->getTerminator())
1315             ->getCondition());
1316     if (CondI)
1317       WorkList.insert(CondI);
1318     MoveInstructions();
1319     for (Instruction *InnerIndexVar : InnerIndexVarList)
1320       WorkList.insert(cast<Instruction>(InnerIndexVar));
1321     MoveInstructions();
1322 
1323     // Splits the inner loops phi nodes out into a separate basic block.
1324     BasicBlock *InnerLoopHeader = InnerLoop->getHeader();
1325     SplitBlock(InnerLoopHeader, InnerLoopHeader->getFirstNonPHI(), DT, LI);
1326     LLVM_DEBUG(dbgs() << "splitting InnerLoopHeader done\n");
1327   }
1328 
1329   // Instructions in the original inner loop preheader may depend on values
1330   // defined in the outer loop header. Move them there, because the original
1331   // inner loop preheader will become the entry into the interchanged loop nest.
1332   // Currently we move all instructions and rely on LICM to move invariant
1333   // instructions outside the loop nest.
1334   BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader();
1335   BasicBlock *OuterLoopHeader = OuterLoop->getHeader();
1336   if (InnerLoopPreHeader != OuterLoopHeader) {
1337     SmallPtrSet<Instruction *, 4> NeedsMoving;
1338     for (Instruction &I :
1339          make_early_inc_range(make_range(InnerLoopPreHeader->begin(),
1340                                          std::prev(InnerLoopPreHeader->end()))))
1341       I.moveBefore(OuterLoopHeader->getTerminator());
1342   }
1343 
1344   Transformed |= adjustLoopLinks();
1345   if (!Transformed) {
1346     LLVM_DEBUG(dbgs() << "adjustLoopLinks failed\n");
1347     return false;
1348   }
1349 
1350   return true;
1351 }
1352 
1353 /// \brief Move all instructions except the terminator from FromBB right before
1354 /// InsertBefore
1355 static void moveBBContents(BasicBlock *FromBB, Instruction *InsertBefore) {
1356   auto &ToList = InsertBefore->getParent()->getInstList();
1357   auto &FromList = FromBB->getInstList();
1358 
1359   ToList.splice(InsertBefore->getIterator(), FromList, FromList.begin(),
1360                 FromBB->getTerminator()->getIterator());
1361 }
1362 
1363 /// Swap instructions between \p BB1 and \p BB2 but keep terminators intact.
1364 static void swapBBContents(BasicBlock *BB1, BasicBlock *BB2) {
1365   // Save all non-terminator instructions of BB1 into TempInstrs and unlink them
1366   // from BB1 afterwards.
1367   auto Iter = map_range(*BB1, [](Instruction &I) { return &I; });
1368   SmallVector<Instruction *, 4> TempInstrs(Iter.begin(), std::prev(Iter.end()));
1369   for (Instruction *I : TempInstrs)
1370     I->removeFromParent();
1371 
1372   // Move instructions from BB2 to BB1.
1373   moveBBContents(BB2, BB1->getTerminator());
1374 
1375   // Move instructions from TempInstrs to BB2.
1376   for (Instruction *I : TempInstrs)
1377     I->insertBefore(BB2->getTerminator());
1378 }
1379 
1380 // Update BI to jump to NewBB instead of OldBB. Records updates to the
1381 // dominator tree in DTUpdates. If \p MustUpdateOnce is true, assert that
1382 // \p OldBB  is exactly once in BI's successor list.
1383 static void updateSuccessor(BranchInst *BI, BasicBlock *OldBB,
1384                             BasicBlock *NewBB,
1385                             std::vector<DominatorTree::UpdateType> &DTUpdates,
1386                             bool MustUpdateOnce = true) {
1387   assert((!MustUpdateOnce ||
1388           llvm::count_if(successors(BI),
1389                          [OldBB](BasicBlock *BB) {
1390                            return BB == OldBB;
1391                          }) == 1) && "BI must jump to OldBB exactly once.");
1392   bool Changed = false;
1393   for (Use &Op : BI->operands())
1394     if (Op == OldBB) {
1395       Op.set(NewBB);
1396       Changed = true;
1397     }
1398 
1399   if (Changed) {
1400     DTUpdates.push_back(
1401         {DominatorTree::UpdateKind::Insert, BI->getParent(), NewBB});
1402     DTUpdates.push_back(
1403         {DominatorTree::UpdateKind::Delete, BI->getParent(), OldBB});
1404   }
1405   assert(Changed && "Expected a successor to be updated");
1406 }
1407 
1408 // Move Lcssa PHIs to the right place.
1409 static void moveLCSSAPhis(BasicBlock *InnerExit, BasicBlock *InnerHeader,
1410                           BasicBlock *InnerLatch, BasicBlock *OuterHeader,
1411                           BasicBlock *OuterLatch, BasicBlock *OuterExit,
1412                           Loop *InnerLoop, LoopInfo *LI) {
1413 
1414   // Deal with LCSSA PHI nodes in the exit block of the inner loop, that are
1415   // defined either in the header or latch. Those blocks will become header and
1416   // latch of the new outer loop, and the only possible users can PHI nodes
1417   // in the exit block of the loop nest or the outer loop header (reduction
1418   // PHIs, in that case, the incoming value must be defined in the inner loop
1419   // header). We can just substitute the user with the incoming value and remove
1420   // the PHI.
1421   for (PHINode &P : make_early_inc_range(InnerExit->phis())) {
1422     assert(P.getNumIncomingValues() == 1 &&
1423            "Only loops with a single exit are supported!");
1424 
1425     // Incoming values are guaranteed be instructions currently.
1426     auto IncI = cast<Instruction>(P.getIncomingValueForBlock(InnerLatch));
1427     // Skip phis with incoming values from the inner loop body, excluding the
1428     // header and latch.
1429     if (IncI->getParent() != InnerLatch && IncI->getParent() != InnerHeader)
1430       continue;
1431 
1432     assert(all_of(P.users(),
1433                   [OuterHeader, OuterExit, IncI, InnerHeader](User *U) {
1434                     return (cast<PHINode>(U)->getParent() == OuterHeader &&
1435                             IncI->getParent() == InnerHeader) ||
1436                            cast<PHINode>(U)->getParent() == OuterExit;
1437                   }) &&
1438            "Can only replace phis iff the uses are in the loop nest exit or "
1439            "the incoming value is defined in the inner header (it will "
1440            "dominate all loop blocks after interchanging)");
1441     P.replaceAllUsesWith(IncI);
1442     P.eraseFromParent();
1443   }
1444 
1445   SmallVector<PHINode *, 8> LcssaInnerExit;
1446   for (PHINode &P : InnerExit->phis())
1447     LcssaInnerExit.push_back(&P);
1448 
1449   SmallVector<PHINode *, 8> LcssaInnerLatch;
1450   for (PHINode &P : InnerLatch->phis())
1451     LcssaInnerLatch.push_back(&P);
1452 
1453   // Lcssa PHIs for values used outside the inner loop are in InnerExit.
1454   // If a PHI node has users outside of InnerExit, it has a use outside the
1455   // interchanged loop and we have to preserve it. We move these to
1456   // InnerLatch, which will become the new exit block for the innermost
1457   // loop after interchanging.
1458   for (PHINode *P : LcssaInnerExit)
1459     P->moveBefore(InnerLatch->getFirstNonPHI());
1460 
1461   // If the inner loop latch contains LCSSA PHIs, those come from a child loop
1462   // and we have to move them to the new inner latch.
1463   for (PHINode *P : LcssaInnerLatch)
1464     P->moveBefore(InnerExit->getFirstNonPHI());
1465 
1466   // Deal with LCSSA PHI nodes in the loop nest exit block. For PHIs that have
1467   // incoming values defined in the outer loop, we have to add a new PHI
1468   // in the inner loop latch, which became the exit block of the outer loop,
1469   // after interchanging.
1470   if (OuterExit) {
1471     for (PHINode &P : OuterExit->phis()) {
1472       if (P.getNumIncomingValues() != 1)
1473         continue;
1474       // Skip Phis with incoming values defined in the inner loop. Those should
1475       // already have been updated.
1476       auto I = dyn_cast<Instruction>(P.getIncomingValue(0));
1477       if (!I || LI->getLoopFor(I->getParent()) == InnerLoop)
1478         continue;
1479 
1480       PHINode *NewPhi = dyn_cast<PHINode>(P.clone());
1481       NewPhi->setIncomingValue(0, P.getIncomingValue(0));
1482       NewPhi->setIncomingBlock(0, OuterLatch);
1483       // We might have incoming edges from other BBs, i.e., the original outer
1484       // header.
1485       for (auto *Pred : predecessors(InnerLatch)) {
1486         if (Pred == OuterLatch)
1487           continue;
1488         NewPhi->addIncoming(P.getIncomingValue(0), Pred);
1489       }
1490       NewPhi->insertBefore(InnerLatch->getFirstNonPHI());
1491       P.setIncomingValue(0, NewPhi);
1492     }
1493   }
1494 
1495   // Now adjust the incoming blocks for the LCSSA PHIs.
1496   // For PHIs moved from Inner's exit block, we need to replace Inner's latch
1497   // with the new latch.
1498   InnerLatch->replacePhiUsesWith(InnerLatch, OuterLatch);
1499 }
1500 
1501 bool LoopInterchangeTransform::adjustLoopBranches() {
1502   LLVM_DEBUG(dbgs() << "adjustLoopBranches called\n");
1503   std::vector<DominatorTree::UpdateType> DTUpdates;
1504 
1505   BasicBlock *OuterLoopPreHeader = OuterLoop->getLoopPreheader();
1506   BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader();
1507 
1508   assert(OuterLoopPreHeader != OuterLoop->getHeader() &&
1509          InnerLoopPreHeader != InnerLoop->getHeader() && OuterLoopPreHeader &&
1510          InnerLoopPreHeader && "Guaranteed by loop-simplify form");
1511   // Ensure that both preheaders do not contain PHI nodes and have single
1512   // predecessors. This allows us to move them easily. We use
1513   // InsertPreHeaderForLoop to create an 'extra' preheader, if the existing
1514   // preheaders do not satisfy those conditions.
1515   if (isa<PHINode>(OuterLoopPreHeader->begin()) ||
1516       !OuterLoopPreHeader->getUniquePredecessor())
1517     OuterLoopPreHeader =
1518         InsertPreheaderForLoop(OuterLoop, DT, LI, nullptr, true);
1519   if (InnerLoopPreHeader == OuterLoop->getHeader())
1520     InnerLoopPreHeader =
1521         InsertPreheaderForLoop(InnerLoop, DT, LI, nullptr, true);
1522 
1523   // Adjust the loop preheader
1524   BasicBlock *InnerLoopHeader = InnerLoop->getHeader();
1525   BasicBlock *OuterLoopHeader = OuterLoop->getHeader();
1526   BasicBlock *InnerLoopLatch = InnerLoop->getLoopLatch();
1527   BasicBlock *OuterLoopLatch = OuterLoop->getLoopLatch();
1528   BasicBlock *OuterLoopPredecessor = OuterLoopPreHeader->getUniquePredecessor();
1529   BasicBlock *InnerLoopLatchPredecessor =
1530       InnerLoopLatch->getUniquePredecessor();
1531   BasicBlock *InnerLoopLatchSuccessor;
1532   BasicBlock *OuterLoopLatchSuccessor;
1533 
1534   BranchInst *OuterLoopLatchBI =
1535       dyn_cast<BranchInst>(OuterLoopLatch->getTerminator());
1536   BranchInst *InnerLoopLatchBI =
1537       dyn_cast<BranchInst>(InnerLoopLatch->getTerminator());
1538   BranchInst *OuterLoopHeaderBI =
1539       dyn_cast<BranchInst>(OuterLoopHeader->getTerminator());
1540   BranchInst *InnerLoopHeaderBI =
1541       dyn_cast<BranchInst>(InnerLoopHeader->getTerminator());
1542 
1543   if (!OuterLoopPredecessor || !InnerLoopLatchPredecessor ||
1544       !OuterLoopLatchBI || !InnerLoopLatchBI || !OuterLoopHeaderBI ||
1545       !InnerLoopHeaderBI)
1546     return false;
1547 
1548   BranchInst *InnerLoopLatchPredecessorBI =
1549       dyn_cast<BranchInst>(InnerLoopLatchPredecessor->getTerminator());
1550   BranchInst *OuterLoopPredecessorBI =
1551       dyn_cast<BranchInst>(OuterLoopPredecessor->getTerminator());
1552 
1553   if (!OuterLoopPredecessorBI || !InnerLoopLatchPredecessorBI)
1554     return false;
1555   BasicBlock *InnerLoopHeaderSuccessor = InnerLoopHeader->getUniqueSuccessor();
1556   if (!InnerLoopHeaderSuccessor)
1557     return false;
1558 
1559   // Adjust Loop Preheader and headers.
1560   // The branches in the outer loop predecessor and the outer loop header can
1561   // be unconditional branches or conditional branches with duplicates. Consider
1562   // this when updating the successors.
1563   updateSuccessor(OuterLoopPredecessorBI, OuterLoopPreHeader,
1564                   InnerLoopPreHeader, DTUpdates, /*MustUpdateOnce=*/false);
1565   // The outer loop header might or might not branch to the outer latch.
1566   // We are guaranteed to branch to the inner loop preheader.
1567   if (llvm::is_contained(OuterLoopHeaderBI->successors(), OuterLoopLatch)) {
1568     // In this case the outerLoopHeader should branch to the InnerLoopLatch.
1569     updateSuccessor(OuterLoopHeaderBI, OuterLoopLatch, InnerLoopLatch,
1570                     DTUpdates,
1571                     /*MustUpdateOnce=*/false);
1572   }
1573   updateSuccessor(OuterLoopHeaderBI, InnerLoopPreHeader,
1574                   InnerLoopHeaderSuccessor, DTUpdates,
1575                   /*MustUpdateOnce=*/false);
1576 
1577   // Adjust reduction PHI's now that the incoming block has changed.
1578   InnerLoopHeaderSuccessor->replacePhiUsesWith(InnerLoopHeader,
1579                                                OuterLoopHeader);
1580 
1581   updateSuccessor(InnerLoopHeaderBI, InnerLoopHeaderSuccessor,
1582                   OuterLoopPreHeader, DTUpdates);
1583 
1584   // -------------Adjust loop latches-----------
1585   if (InnerLoopLatchBI->getSuccessor(0) == InnerLoopHeader)
1586     InnerLoopLatchSuccessor = InnerLoopLatchBI->getSuccessor(1);
1587   else
1588     InnerLoopLatchSuccessor = InnerLoopLatchBI->getSuccessor(0);
1589 
1590   updateSuccessor(InnerLoopLatchPredecessorBI, InnerLoopLatch,
1591                   InnerLoopLatchSuccessor, DTUpdates);
1592 
1593   if (OuterLoopLatchBI->getSuccessor(0) == OuterLoopHeader)
1594     OuterLoopLatchSuccessor = OuterLoopLatchBI->getSuccessor(1);
1595   else
1596     OuterLoopLatchSuccessor = OuterLoopLatchBI->getSuccessor(0);
1597 
1598   updateSuccessor(InnerLoopLatchBI, InnerLoopLatchSuccessor,
1599                   OuterLoopLatchSuccessor, DTUpdates);
1600   updateSuccessor(OuterLoopLatchBI, OuterLoopLatchSuccessor, InnerLoopLatch,
1601                   DTUpdates);
1602 
1603   DT->applyUpdates(DTUpdates);
1604   restructureLoops(OuterLoop, InnerLoop, InnerLoopPreHeader,
1605                    OuterLoopPreHeader);
1606 
1607   moveLCSSAPhis(InnerLoopLatchSuccessor, InnerLoopHeader, InnerLoopLatch,
1608                 OuterLoopHeader, OuterLoopLatch, InnerLoop->getExitBlock(),
1609                 InnerLoop, LI);
1610   // For PHIs in the exit block of the outer loop, outer's latch has been
1611   // replaced by Inners'.
1612   OuterLoopLatchSuccessor->replacePhiUsesWith(OuterLoopLatch, InnerLoopLatch);
1613 
1614   auto &OuterInnerReductions = LIL.getOuterInnerReductions();
1615   // Now update the reduction PHIs in the inner and outer loop headers.
1616   SmallVector<PHINode *, 4> InnerLoopPHIs, OuterLoopPHIs;
1617   for (PHINode &PHI : InnerLoopHeader->phis())
1618     if (OuterInnerReductions.contains(&PHI))
1619       InnerLoopPHIs.push_back(&PHI);
1620 
1621   for (PHINode &PHI : OuterLoopHeader->phis())
1622     if (OuterInnerReductions.contains(&PHI))
1623       OuterLoopPHIs.push_back(&PHI);
1624 
1625   // Now move the remaining reduction PHIs from outer to inner loop header and
1626   // vice versa. The PHI nodes must be part of a reduction across the inner and
1627   // outer loop and all the remains to do is and updating the incoming blocks.
1628   for (PHINode *PHI : OuterLoopPHIs) {
1629     LLVM_DEBUG(dbgs() << "Outer loop reduction PHIs:\n"; PHI->dump(););
1630     PHI->moveBefore(InnerLoopHeader->getFirstNonPHI());
1631     assert(OuterInnerReductions.count(PHI) && "Expected a reduction PHI node");
1632   }
1633   for (PHINode *PHI : InnerLoopPHIs) {
1634     LLVM_DEBUG(dbgs() << "Inner loop reduction PHIs:\n"; PHI->dump(););
1635     PHI->moveBefore(OuterLoopHeader->getFirstNonPHI());
1636     assert(OuterInnerReductions.count(PHI) && "Expected a reduction PHI node");
1637   }
1638 
1639   // Update the incoming blocks for moved PHI nodes.
1640   OuterLoopHeader->replacePhiUsesWith(InnerLoopPreHeader, OuterLoopPreHeader);
1641   OuterLoopHeader->replacePhiUsesWith(InnerLoopLatch, OuterLoopLatch);
1642   InnerLoopHeader->replacePhiUsesWith(OuterLoopPreHeader, InnerLoopPreHeader);
1643   InnerLoopHeader->replacePhiUsesWith(OuterLoopLatch, InnerLoopLatch);
1644 
1645   // Values defined in the outer loop header could be used in the inner loop
1646   // latch. In that case, we need to create LCSSA phis for them, because after
1647   // interchanging they will be defined in the new inner loop and used in the
1648   // new outer loop.
1649   IRBuilder<> Builder(OuterLoopHeader->getContext());
1650   SmallVector<Instruction *, 4> MayNeedLCSSAPhis;
1651   for (Instruction &I :
1652        make_range(OuterLoopHeader->begin(), std::prev(OuterLoopHeader->end())))
1653     MayNeedLCSSAPhis.push_back(&I);
1654   formLCSSAForInstructions(MayNeedLCSSAPhis, *DT, *LI, SE, Builder);
1655 
1656   return true;
1657 }
1658 
1659 bool LoopInterchangeTransform::adjustLoopLinks() {
1660   // Adjust all branches in the inner and outer loop.
1661   bool Changed = adjustLoopBranches();
1662   if (Changed) {
1663     // We have interchanged the preheaders so we need to interchange the data in
1664     // the preheaders as well. This is because the content of the inner
1665     // preheader was previously executed inside the outer loop.
1666     BasicBlock *OuterLoopPreHeader = OuterLoop->getLoopPreheader();
1667     BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader();
1668     swapBBContents(OuterLoopPreHeader, InnerLoopPreHeader);
1669   }
1670   return Changed;
1671 }
1672 
1673 namespace {
1674 /// Main LoopInterchange Pass.
1675 struct LoopInterchangeLegacyPass : public LoopPass {
1676   static char ID;
1677 
1678   LoopInterchangeLegacyPass() : LoopPass(ID) {
1679     initializeLoopInterchangeLegacyPassPass(*PassRegistry::getPassRegistry());
1680   }
1681 
1682   void getAnalysisUsage(AnalysisUsage &AU) const override {
1683     AU.addRequired<DependenceAnalysisWrapperPass>();
1684     AU.addRequired<OptimizationRemarkEmitterWrapperPass>();
1685 
1686     getLoopAnalysisUsage(AU);
1687   }
1688 
1689   bool runOnLoop(Loop *L, LPPassManager &LPM) override {
1690     if (skipLoop(L))
1691       return false;
1692 
1693     auto *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
1694     auto *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1695     auto *DI = &getAnalysis<DependenceAnalysisWrapperPass>().getDI();
1696     auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1697     auto *ORE = &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE();
1698 
1699     return LoopInterchange(SE, LI, DI, DT, ORE).run(L);
1700   }
1701 };
1702 } // namespace
1703 
1704 char LoopInterchangeLegacyPass::ID = 0;
1705 
1706 INITIALIZE_PASS_BEGIN(LoopInterchangeLegacyPass, "loop-interchange",
1707                       "Interchanges loops for cache reuse", false, false)
1708 INITIALIZE_PASS_DEPENDENCY(LoopPass)
1709 INITIALIZE_PASS_DEPENDENCY(DependenceAnalysisWrapperPass)
1710 INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass)
1711 
1712 INITIALIZE_PASS_END(LoopInterchangeLegacyPass, "loop-interchange",
1713                     "Interchanges loops for cache reuse", false, false)
1714 
1715 Pass *llvm::createLoopInterchangePass() {
1716   return new LoopInterchangeLegacyPass();
1717 }
1718 
1719 PreservedAnalyses LoopInterchangePass::run(LoopNest &LN,
1720                                            LoopAnalysisManager &AM,
1721                                            LoopStandardAnalysisResults &AR,
1722                                            LPMUpdater &U) {
1723   Function &F = *LN.getParent();
1724 
1725   DependenceInfo DI(&F, &AR.AA, &AR.SE, &AR.LI);
1726   OptimizationRemarkEmitter ORE(&F);
1727   if (!LoopInterchange(&AR.SE, &AR.LI, &DI, &AR.DT, &ORE).run(LN))
1728     return PreservedAnalyses::all();
1729   return getLoopPassPreservedAnalyses();
1730 }
1731