LoopInterchange.cpp revision 360784
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/ADT/STLExtras.h"
16#include "llvm/ADT/SmallVector.h"
17#include "llvm/ADT/Statistic.h"
18#include "llvm/ADT/StringRef.h"
19#include "llvm/Analysis/DependenceAnalysis.h"
20#include "llvm/Analysis/LoopInfo.h"
21#include "llvm/Analysis/LoopPass.h"
22#include "llvm/Analysis/OptimizationRemarkEmitter.h"
23#include "llvm/Analysis/ScalarEvolution.h"
24#include "llvm/Analysis/ScalarEvolutionExpressions.h"
25#include "llvm/IR/BasicBlock.h"
26#include "llvm/IR/Constants.h"
27#include "llvm/IR/DiagnosticInfo.h"
28#include "llvm/IR/Dominators.h"
29#include "llvm/IR/Function.h"
30#include "llvm/IR/InstrTypes.h"
31#include "llvm/IR/Instruction.h"
32#include "llvm/IR/Instructions.h"
33#include "llvm/IR/Type.h"
34#include "llvm/IR/User.h"
35#include "llvm/IR/Value.h"
36#include "llvm/InitializePasses.h"
37#include "llvm/Pass.h"
38#include "llvm/Support/Casting.h"
39#include "llvm/Support/CommandLine.h"
40#include "llvm/Support/Debug.h"
41#include "llvm/Support/ErrorHandling.h"
42#include "llvm/Support/raw_ostream.h"
43#include "llvm/Transforms/Scalar.h"
44#include "llvm/Transforms/Utils.h"
45#include "llvm/Transforms/Utils/BasicBlockUtils.h"
46#include "llvm/Transforms/Utils/LoopUtils.h"
47#include <cassert>
48#include <utility>
49#include <vector>
50
51using namespace llvm;
52
53#define DEBUG_TYPE "loop-interchange"
54
55STATISTIC(LoopsInterchanged, "Number of loops interchanged");
56
57static cl::opt<int> LoopInterchangeCostThreshold(
58    "loop-interchange-threshold", cl::init(0), cl::Hidden,
59    cl::desc("Interchange if you gain more than this number"));
60
61namespace {
62
63using LoopVector = SmallVector<Loop *, 8>;
64
65// TODO: Check if we can use a sparse matrix here.
66using CharMatrix = std::vector<std::vector<char>>;
67
68} // end anonymous namespace
69
70// Maximum number of dependencies that can be handled in the dependency matrix.
71static const unsigned MaxMemInstrCount = 100;
72
73// Maximum loop depth supported.
74static const unsigned MaxLoopNestDepth = 10;
75
76#ifdef DUMP_DEP_MATRICIES
77static void printDepMatrix(CharMatrix &DepMatrix) {
78  for (auto &Row : DepMatrix) {
79    for (auto D : Row)
80      LLVM_DEBUG(dbgs() << D << " ");
81    LLVM_DEBUG(dbgs() << "\n");
82  }
83}
84#endif
85
86static bool populateDependencyMatrix(CharMatrix &DepMatrix, unsigned Level,
87                                     Loop *L, DependenceInfo *DI) {
88  using ValueVector = SmallVector<Value *, 16>;
89
90  ValueVector MemInstr;
91
92  // For each block.
93  for (BasicBlock *BB : L->blocks()) {
94    // Scan the BB and collect legal loads and stores.
95    for (Instruction &I : *BB) {
96      if (!isa<Instruction>(I))
97        return false;
98      if (auto *Ld = dyn_cast<LoadInst>(&I)) {
99        if (!Ld->isSimple())
100          return false;
101        MemInstr.push_back(&I);
102      } else if (auto *St = dyn_cast<StoreInst>(&I)) {
103        if (!St->isSimple())
104          return false;
105        MemInstr.push_back(&I);
106      }
107    }
108  }
109
110  LLVM_DEBUG(dbgs() << "Found " << MemInstr.size()
111                    << " Loads and Stores to analyze\n");
112
113  ValueVector::iterator I, IE, J, JE;
114
115  for (I = MemInstr.begin(), IE = MemInstr.end(); I != IE; ++I) {
116    for (J = I, JE = MemInstr.end(); J != JE; ++J) {
117      std::vector<char> Dep;
118      Instruction *Src = cast<Instruction>(*I);
119      Instruction *Dst = cast<Instruction>(*J);
120      if (Src == Dst)
121        continue;
122      // Ignore Input dependencies.
123      if (isa<LoadInst>(Src) && isa<LoadInst>(Dst))
124        continue;
125      // Track Output, Flow, and Anti dependencies.
126      if (auto D = DI->depends(Src, Dst, true)) {
127        assert(D->isOrdered() && "Expected an output, flow or anti dep.");
128        LLVM_DEBUG(StringRef DepType =
129                       D->isFlow() ? "flow" : D->isAnti() ? "anti" : "output";
130                   dbgs() << "Found " << DepType
131                          << " dependency between Src and Dst\n"
132                          << " Src:" << *Src << "\n Dst:" << *Dst << '\n');
133        unsigned Levels = D->getLevels();
134        char Direction;
135        for (unsigned II = 1; II <= Levels; ++II) {
136          const SCEV *Distance = D->getDistance(II);
137          const SCEVConstant *SCEVConst =
138              dyn_cast_or_null<SCEVConstant>(Distance);
139          if (SCEVConst) {
140            const ConstantInt *CI = SCEVConst->getValue();
141            if (CI->isNegative())
142              Direction = '<';
143            else if (CI->isZero())
144              Direction = '=';
145            else
146              Direction = '>';
147            Dep.push_back(Direction);
148          } else if (D->isScalar(II)) {
149            Direction = 'S';
150            Dep.push_back(Direction);
151          } else {
152            unsigned Dir = D->getDirection(II);
153            if (Dir == Dependence::DVEntry::LT ||
154                Dir == Dependence::DVEntry::LE)
155              Direction = '<';
156            else if (Dir == Dependence::DVEntry::GT ||
157                     Dir == Dependence::DVEntry::GE)
158              Direction = '>';
159            else if (Dir == Dependence::DVEntry::EQ)
160              Direction = '=';
161            else
162              Direction = '*';
163            Dep.push_back(Direction);
164          }
165        }
166        while (Dep.size() != Level) {
167          Dep.push_back('I');
168        }
169
170        DepMatrix.push_back(Dep);
171        if (DepMatrix.size() > MaxMemInstrCount) {
172          LLVM_DEBUG(dbgs() << "Cannot handle more than " << MaxMemInstrCount
173                            << " dependencies inside loop\n");
174          return false;
175        }
176      }
177    }
178  }
179
180  return true;
181}
182
183// A loop is moved from index 'from' to an index 'to'. Update the Dependence
184// matrix by exchanging the two columns.
185static void interChangeDependencies(CharMatrix &DepMatrix, unsigned FromIndx,
186                                    unsigned ToIndx) {
187  unsigned numRows = DepMatrix.size();
188  for (unsigned i = 0; i < numRows; ++i) {
189    char TmpVal = DepMatrix[i][ToIndx];
190    DepMatrix[i][ToIndx] = DepMatrix[i][FromIndx];
191    DepMatrix[i][FromIndx] = TmpVal;
192  }
193}
194
195// Checks if outermost non '=','S'or'I' dependence in the dependence matrix is
196// '>'
197static bool isOuterMostDepPositive(CharMatrix &DepMatrix, unsigned Row,
198                                   unsigned Column) {
199  for (unsigned i = 0; i <= Column; ++i) {
200    if (DepMatrix[Row][i] == '<')
201      return false;
202    if (DepMatrix[Row][i] == '>')
203      return true;
204  }
205  // All dependencies were '=','S' or 'I'
206  return false;
207}
208
209// Checks if no dependence exist in the dependency matrix in Row before Column.
210static bool containsNoDependence(CharMatrix &DepMatrix, unsigned Row,
211                                 unsigned Column) {
212  for (unsigned i = 0; i < Column; ++i) {
213    if (DepMatrix[Row][i] != '=' && DepMatrix[Row][i] != 'S' &&
214        DepMatrix[Row][i] != 'I')
215      return false;
216  }
217  return true;
218}
219
220static bool validDepInterchange(CharMatrix &DepMatrix, unsigned Row,
221                                unsigned OuterLoopId, char InnerDep,
222                                char OuterDep) {
223  if (isOuterMostDepPositive(DepMatrix, Row, OuterLoopId))
224    return false;
225
226  if (InnerDep == OuterDep)
227    return true;
228
229  // It is legal to interchange if and only if after interchange no row has a
230  // '>' direction as the leftmost non-'='.
231
232  if (InnerDep == '=' || InnerDep == 'S' || InnerDep == 'I')
233    return true;
234
235  if (InnerDep == '<')
236    return true;
237
238  if (InnerDep == '>') {
239    // If OuterLoopId represents outermost loop then interchanging will make the
240    // 1st dependency as '>'
241    if (OuterLoopId == 0)
242      return false;
243
244    // If all dependencies before OuterloopId are '=','S'or 'I'. Then
245    // interchanging will result in this row having an outermost non '='
246    // dependency of '>'
247    if (!containsNoDependence(DepMatrix, Row, OuterLoopId))
248      return true;
249  }
250
251  return false;
252}
253
254// Checks if it is legal to interchange 2 loops.
255// [Theorem] A permutation of the loops in a perfect nest is legal if and only
256// if the direction matrix, after the same permutation is applied to its
257// columns, has no ">" direction as the leftmost non-"=" direction in any row.
258static bool isLegalToInterChangeLoops(CharMatrix &DepMatrix,
259                                      unsigned InnerLoopId,
260                                      unsigned OuterLoopId) {
261  unsigned NumRows = DepMatrix.size();
262  // For each row check if it is valid to interchange.
263  for (unsigned Row = 0; Row < NumRows; ++Row) {
264    char InnerDep = DepMatrix[Row][InnerLoopId];
265    char OuterDep = DepMatrix[Row][OuterLoopId];
266    if (InnerDep == '*' || OuterDep == '*')
267      return false;
268    if (!validDepInterchange(DepMatrix, Row, OuterLoopId, InnerDep, OuterDep))
269      return false;
270  }
271  return true;
272}
273
274static LoopVector populateWorklist(Loop &L) {
275  LLVM_DEBUG(dbgs() << "Calling populateWorklist on Func: "
276                    << L.getHeader()->getParent()->getName() << " Loop: %"
277                    << L.getHeader()->getName() << '\n');
278  LoopVector LoopList;
279  Loop *CurrentLoop = &L;
280  const std::vector<Loop *> *Vec = &CurrentLoop->getSubLoops();
281  while (!Vec->empty()) {
282    // The current loop has multiple subloops in it hence it is not tightly
283    // nested.
284    // Discard all loops above it added into Worklist.
285    if (Vec->size() != 1)
286      return {};
287
288    LoopList.push_back(CurrentLoop);
289    CurrentLoop = Vec->front();
290    Vec = &CurrentLoop->getSubLoops();
291  }
292  LoopList.push_back(CurrentLoop);
293  return LoopList;
294}
295
296static PHINode *getInductionVariable(Loop *L, ScalarEvolution *SE) {
297  PHINode *InnerIndexVar = L->getCanonicalInductionVariable();
298  if (InnerIndexVar)
299    return InnerIndexVar;
300  if (L->getLoopLatch() == nullptr || L->getLoopPredecessor() == nullptr)
301    return nullptr;
302  for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
303    PHINode *PhiVar = cast<PHINode>(I);
304    Type *PhiTy = PhiVar->getType();
305    if (!PhiTy->isIntegerTy() && !PhiTy->isFloatingPointTy() &&
306        !PhiTy->isPointerTy())
307      return nullptr;
308    const SCEVAddRecExpr *AddRec =
309        dyn_cast<SCEVAddRecExpr>(SE->getSCEV(PhiVar));
310    if (!AddRec || !AddRec->isAffine())
311      continue;
312    const SCEV *Step = AddRec->getStepRecurrence(*SE);
313    if (!isa<SCEVConstant>(Step))
314      continue;
315    // Found the induction variable.
316    // FIXME: Handle loops with more than one induction variable. Note that,
317    // currently, legality makes sure we have only one induction variable.
318    return PhiVar;
319  }
320  return nullptr;
321}
322
323namespace {
324
325/// LoopInterchangeLegality checks if it is legal to interchange the loop.
326class LoopInterchangeLegality {
327public:
328  LoopInterchangeLegality(Loop *Outer, Loop *Inner, ScalarEvolution *SE,
329                          OptimizationRemarkEmitter *ORE)
330      : OuterLoop(Outer), InnerLoop(Inner), SE(SE), ORE(ORE) {}
331
332  /// Check if the loops can be interchanged.
333  bool canInterchangeLoops(unsigned InnerLoopId, unsigned OuterLoopId,
334                           CharMatrix &DepMatrix);
335
336  /// Check if the loop structure is understood. We do not handle triangular
337  /// loops for now.
338  bool isLoopStructureUnderstood(PHINode *InnerInductionVar);
339
340  bool currentLimitations();
341
342  const SmallPtrSetImpl<PHINode *> &getOuterInnerReductions() const {
343    return OuterInnerReductions;
344  }
345
346private:
347  bool tightlyNested(Loop *Outer, Loop *Inner);
348  bool containsUnsafeInstructions(BasicBlock *BB);
349
350  /// Discover induction and reduction PHIs in the header of \p L. Induction
351  /// PHIs are added to \p Inductions, reductions are added to
352  /// OuterInnerReductions. When the outer loop is passed, the inner loop needs
353  /// to be passed as \p InnerLoop.
354  bool findInductionAndReductions(Loop *L,
355                                  SmallVector<PHINode *, 8> &Inductions,
356                                  Loop *InnerLoop);
357
358  Loop *OuterLoop;
359  Loop *InnerLoop;
360
361  ScalarEvolution *SE;
362
363  /// Interface to emit optimization remarks.
364  OptimizationRemarkEmitter *ORE;
365
366  /// Set of reduction PHIs taking part of a reduction across the inner and
367  /// outer loop.
368  SmallPtrSet<PHINode *, 4> OuterInnerReductions;
369};
370
371/// LoopInterchangeProfitability checks if it is profitable to interchange the
372/// loop.
373class LoopInterchangeProfitability {
374public:
375  LoopInterchangeProfitability(Loop *Outer, Loop *Inner, ScalarEvolution *SE,
376                               OptimizationRemarkEmitter *ORE)
377      : OuterLoop(Outer), InnerLoop(Inner), SE(SE), ORE(ORE) {}
378
379  /// Check if the loop interchange is profitable.
380  bool isProfitable(unsigned InnerLoopId, unsigned OuterLoopId,
381                    CharMatrix &DepMatrix);
382
383private:
384  int getInstrOrderCost();
385
386  Loop *OuterLoop;
387  Loop *InnerLoop;
388
389  /// Scev analysis.
390  ScalarEvolution *SE;
391
392  /// Interface to emit optimization remarks.
393  OptimizationRemarkEmitter *ORE;
394};
395
396/// LoopInterchangeTransform interchanges the loop.
397class LoopInterchangeTransform {
398public:
399  LoopInterchangeTransform(Loop *Outer, Loop *Inner, ScalarEvolution *SE,
400                           LoopInfo *LI, DominatorTree *DT,
401                           BasicBlock *LoopNestExit,
402                           const LoopInterchangeLegality &LIL)
403      : OuterLoop(Outer), InnerLoop(Inner), SE(SE), LI(LI), DT(DT),
404        LoopExit(LoopNestExit), LIL(LIL) {}
405
406  /// Interchange OuterLoop and InnerLoop.
407  bool transform();
408  void restructureLoops(Loop *NewInner, Loop *NewOuter,
409                        BasicBlock *OrigInnerPreHeader,
410                        BasicBlock *OrigOuterPreHeader);
411  void removeChildLoop(Loop *OuterLoop, Loop *InnerLoop);
412
413private:
414  bool adjustLoopLinks();
415  void adjustLoopPreheaders();
416  bool adjustLoopBranches();
417
418  Loop *OuterLoop;
419  Loop *InnerLoop;
420
421  /// Scev analysis.
422  ScalarEvolution *SE;
423
424  LoopInfo *LI;
425  DominatorTree *DT;
426  BasicBlock *LoopExit;
427
428  const LoopInterchangeLegality &LIL;
429};
430
431// Main LoopInterchange Pass.
432struct LoopInterchange : public LoopPass {
433  static char ID;
434  ScalarEvolution *SE = nullptr;
435  LoopInfo *LI = nullptr;
436  DependenceInfo *DI = nullptr;
437  DominatorTree *DT = nullptr;
438
439  /// Interface to emit optimization remarks.
440  OptimizationRemarkEmitter *ORE;
441
442  LoopInterchange() : LoopPass(ID) {
443    initializeLoopInterchangePass(*PassRegistry::getPassRegistry());
444  }
445
446  void getAnalysisUsage(AnalysisUsage &AU) const override {
447    AU.addRequired<DependenceAnalysisWrapperPass>();
448    AU.addRequired<OptimizationRemarkEmitterWrapperPass>();
449
450    getLoopAnalysisUsage(AU);
451  }
452
453  bool runOnLoop(Loop *L, LPPassManager &LPM) override {
454    if (skipLoop(L) || L->getParentLoop())
455      return false;
456
457    SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
458    LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
459    DI = &getAnalysis<DependenceAnalysisWrapperPass>().getDI();
460    DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
461    ORE = &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE();
462
463    return processLoopList(populateWorklist(*L));
464  }
465
466  bool isComputableLoopNest(LoopVector LoopList) {
467    for (Loop *L : LoopList) {
468      const SCEV *ExitCountOuter = SE->getBackedgeTakenCount(L);
469      if (ExitCountOuter == SE->getCouldNotCompute()) {
470        LLVM_DEBUG(dbgs() << "Couldn't compute backedge count\n");
471        return false;
472      }
473      if (L->getNumBackEdges() != 1) {
474        LLVM_DEBUG(dbgs() << "NumBackEdges is not equal to 1\n");
475        return false;
476      }
477      if (!L->getExitingBlock()) {
478        LLVM_DEBUG(dbgs() << "Loop doesn't have unique exit block\n");
479        return false;
480      }
481    }
482    return true;
483  }
484
485  unsigned selectLoopForInterchange(const LoopVector &LoopList) {
486    // TODO: Add a better heuristic to select the loop to be interchanged based
487    // on the dependence matrix. Currently we select the innermost loop.
488    return LoopList.size() - 1;
489  }
490
491  bool processLoopList(LoopVector LoopList) {
492    bool Changed = false;
493    unsigned LoopNestDepth = LoopList.size();
494    if (LoopNestDepth < 2) {
495      LLVM_DEBUG(dbgs() << "Loop doesn't contain minimum nesting level.\n");
496      return false;
497    }
498    if (LoopNestDepth > MaxLoopNestDepth) {
499      LLVM_DEBUG(dbgs() << "Cannot handle loops of depth greater than "
500                        << MaxLoopNestDepth << "\n");
501      return false;
502    }
503    if (!isComputableLoopNest(LoopList)) {
504      LLVM_DEBUG(dbgs() << "Not valid loop candidate for interchange\n");
505      return false;
506    }
507
508    LLVM_DEBUG(dbgs() << "Processing LoopList of size = " << LoopNestDepth
509                      << "\n");
510
511    CharMatrix DependencyMatrix;
512    Loop *OuterMostLoop = *(LoopList.begin());
513    if (!populateDependencyMatrix(DependencyMatrix, LoopNestDepth,
514                                  OuterMostLoop, DI)) {
515      LLVM_DEBUG(dbgs() << "Populating dependency matrix failed\n");
516      return false;
517    }
518#ifdef DUMP_DEP_MATRICIES
519    LLVM_DEBUG(dbgs() << "Dependence before interchange\n");
520    printDepMatrix(DependencyMatrix);
521#endif
522
523    // Get the Outermost loop exit.
524    BasicBlock *LoopNestExit = OuterMostLoop->getExitBlock();
525    if (!LoopNestExit) {
526      LLVM_DEBUG(dbgs() << "OuterMostLoop needs an unique exit block");
527      return false;
528    }
529
530    unsigned SelecLoopId = selectLoopForInterchange(LoopList);
531    // Move the selected loop outwards to the best possible position.
532    for (unsigned i = SelecLoopId; i > 0; i--) {
533      bool Interchanged =
534          processLoop(LoopList, i, i - 1, LoopNestExit, DependencyMatrix);
535      if (!Interchanged)
536        return Changed;
537      // Loops interchanged reflect the same in LoopList
538      std::swap(LoopList[i - 1], LoopList[i]);
539
540      // Update the DependencyMatrix
541      interChangeDependencies(DependencyMatrix, i, i - 1);
542#ifdef DUMP_DEP_MATRICIES
543      LLVM_DEBUG(dbgs() << "Dependence after interchange\n");
544      printDepMatrix(DependencyMatrix);
545#endif
546      Changed |= Interchanged;
547    }
548    return Changed;
549  }
550
551  bool processLoop(LoopVector LoopList, unsigned InnerLoopId,
552                   unsigned OuterLoopId, BasicBlock *LoopNestExit,
553                   std::vector<std::vector<char>> &DependencyMatrix) {
554    LLVM_DEBUG(dbgs() << "Processing Inner Loop Id = " << InnerLoopId
555                      << " and OuterLoopId = " << OuterLoopId << "\n");
556    Loop *InnerLoop = LoopList[InnerLoopId];
557    Loop *OuterLoop = LoopList[OuterLoopId];
558
559    LoopInterchangeLegality LIL(OuterLoop, InnerLoop, SE, ORE);
560    if (!LIL.canInterchangeLoops(InnerLoopId, OuterLoopId, DependencyMatrix)) {
561      LLVM_DEBUG(dbgs() << "Not interchanging loops. Cannot prove legality.\n");
562      return false;
563    }
564    LLVM_DEBUG(dbgs() << "Loops are legal to interchange\n");
565    LoopInterchangeProfitability LIP(OuterLoop, InnerLoop, SE, ORE);
566    if (!LIP.isProfitable(InnerLoopId, OuterLoopId, DependencyMatrix)) {
567      LLVM_DEBUG(dbgs() << "Interchanging loops not profitable.\n");
568      return false;
569    }
570
571    ORE->emit([&]() {
572      return OptimizationRemark(DEBUG_TYPE, "Interchanged",
573                                InnerLoop->getStartLoc(),
574                                InnerLoop->getHeader())
575             << "Loop interchanged with enclosing loop.";
576    });
577
578    LoopInterchangeTransform LIT(OuterLoop, InnerLoop, SE, LI, DT, LoopNestExit,
579                                 LIL);
580    LIT.transform();
581    LLVM_DEBUG(dbgs() << "Loops interchanged.\n");
582    LoopsInterchanged++;
583    return true;
584  }
585};
586
587} // end anonymous namespace
588
589bool LoopInterchangeLegality::containsUnsafeInstructions(BasicBlock *BB) {
590  return any_of(*BB, [](const Instruction &I) {
591    return I.mayHaveSideEffects() || I.mayReadFromMemory();
592  });
593}
594
595bool LoopInterchangeLegality::tightlyNested(Loop *OuterLoop, Loop *InnerLoop) {
596  BasicBlock *OuterLoopHeader = OuterLoop->getHeader();
597  BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader();
598  BasicBlock *OuterLoopLatch = OuterLoop->getLoopLatch();
599
600  LLVM_DEBUG(dbgs() << "Checking if loops are tightly nested\n");
601
602  // A perfectly nested loop will not have any branch in between the outer and
603  // inner block i.e. outer header will branch to either inner preheader and
604  // outerloop latch.
605  BranchInst *OuterLoopHeaderBI =
606      dyn_cast<BranchInst>(OuterLoopHeader->getTerminator());
607  if (!OuterLoopHeaderBI)
608    return false;
609
610  for (BasicBlock *Succ : successors(OuterLoopHeaderBI))
611    if (Succ != InnerLoopPreHeader && Succ != InnerLoop->getHeader() &&
612        Succ != OuterLoopLatch)
613      return false;
614
615  LLVM_DEBUG(dbgs() << "Checking instructions in Loop header and Loop latch\n");
616  // We do not have any basic block in between now make sure the outer header
617  // and outer loop latch doesn't contain any unsafe instructions.
618  if (containsUnsafeInstructions(OuterLoopHeader) ||
619      containsUnsafeInstructions(OuterLoopLatch))
620    return false;
621
622  LLVM_DEBUG(dbgs() << "Loops are perfectly nested\n");
623  // We have a perfect loop nest.
624  return true;
625}
626
627bool LoopInterchangeLegality::isLoopStructureUnderstood(
628    PHINode *InnerInduction) {
629  unsigned Num = InnerInduction->getNumOperands();
630  BasicBlock *InnerLoopPreheader = InnerLoop->getLoopPreheader();
631  for (unsigned i = 0; i < Num; ++i) {
632    Value *Val = InnerInduction->getOperand(i);
633    if (isa<Constant>(Val))
634      continue;
635    Instruction *I = dyn_cast<Instruction>(Val);
636    if (!I)
637      return false;
638    // TODO: Handle triangular loops.
639    // e.g. for(int i=0;i<N;i++)
640    //        for(int j=i;j<N;j++)
641    unsigned IncomBlockIndx = PHINode::getIncomingValueNumForOperand(i);
642    if (InnerInduction->getIncomingBlock(IncomBlockIndx) ==
643            InnerLoopPreheader &&
644        !OuterLoop->isLoopInvariant(I)) {
645      return false;
646    }
647  }
648  return true;
649}
650
651// If SV is a LCSSA PHI node with a single incoming value, return the incoming
652// value.
653static Value *followLCSSA(Value *SV) {
654  PHINode *PHI = dyn_cast<PHINode>(SV);
655  if (!PHI)
656    return SV;
657
658  if (PHI->getNumIncomingValues() != 1)
659    return SV;
660  return followLCSSA(PHI->getIncomingValue(0));
661}
662
663// Check V's users to see if it is involved in a reduction in L.
664static PHINode *findInnerReductionPhi(Loop *L, Value *V) {
665  for (Value *User : V->users()) {
666    if (PHINode *PHI = dyn_cast<PHINode>(User)) {
667      if (PHI->getNumIncomingValues() == 1)
668        continue;
669      RecurrenceDescriptor RD;
670      if (RecurrenceDescriptor::isReductionPHI(PHI, L, RD))
671        return PHI;
672      return nullptr;
673    }
674  }
675
676  return nullptr;
677}
678
679bool LoopInterchangeLegality::findInductionAndReductions(
680    Loop *L, SmallVector<PHINode *, 8> &Inductions, Loop *InnerLoop) {
681  if (!L->getLoopLatch() || !L->getLoopPredecessor())
682    return false;
683  for (PHINode &PHI : L->getHeader()->phis()) {
684    RecurrenceDescriptor RD;
685    InductionDescriptor ID;
686    if (InductionDescriptor::isInductionPHI(&PHI, L, SE, ID))
687      Inductions.push_back(&PHI);
688    else {
689      // PHIs in inner loops need to be part of a reduction in the outer loop,
690      // discovered when checking the PHIs of the outer loop earlier.
691      if (!InnerLoop) {
692        if (OuterInnerReductions.find(&PHI) == OuterInnerReductions.end()) {
693          LLVM_DEBUG(dbgs() << "Inner loop PHI is not part of reductions "
694                               "across the outer loop.\n");
695          return false;
696        }
697      } else {
698        assert(PHI.getNumIncomingValues() == 2 &&
699               "Phis in loop header should have exactly 2 incoming values");
700        // Check if we have a PHI node in the outer loop that has a reduction
701        // result from the inner loop as an incoming value.
702        Value *V = followLCSSA(PHI.getIncomingValueForBlock(L->getLoopLatch()));
703        PHINode *InnerRedPhi = findInnerReductionPhi(InnerLoop, V);
704        if (!InnerRedPhi ||
705            !llvm::any_of(InnerRedPhi->incoming_values(),
706                          [&PHI](Value *V) { return V == &PHI; })) {
707          LLVM_DEBUG(
708              dbgs()
709              << "Failed to recognize PHI as an induction or reduction.\n");
710          return false;
711        }
712        OuterInnerReductions.insert(&PHI);
713        OuterInnerReductions.insert(InnerRedPhi);
714      }
715    }
716  }
717  return true;
718}
719
720// This function indicates the current limitations in the transform as a result
721// of which we do not proceed.
722bool LoopInterchangeLegality::currentLimitations() {
723  BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader();
724  BasicBlock *InnerLoopLatch = InnerLoop->getLoopLatch();
725
726  // transform currently expects the loop latches to also be the exiting
727  // blocks.
728  if (InnerLoop->getExitingBlock() != InnerLoopLatch ||
729      OuterLoop->getExitingBlock() != OuterLoop->getLoopLatch() ||
730      !isa<BranchInst>(InnerLoopLatch->getTerminator()) ||
731      !isa<BranchInst>(OuterLoop->getLoopLatch()->getTerminator())) {
732    LLVM_DEBUG(
733        dbgs() << "Loops where the latch is not the exiting block are not"
734               << " supported currently.\n");
735    ORE->emit([&]() {
736      return OptimizationRemarkMissed(DEBUG_TYPE, "ExitingNotLatch",
737                                      OuterLoop->getStartLoc(),
738                                      OuterLoop->getHeader())
739             << "Loops where the latch is not the exiting block cannot be"
740                " interchange currently.";
741    });
742    return true;
743  }
744
745  PHINode *InnerInductionVar;
746  SmallVector<PHINode *, 8> Inductions;
747  if (!findInductionAndReductions(OuterLoop, Inductions, InnerLoop)) {
748    LLVM_DEBUG(
749        dbgs() << "Only outer loops with induction or reduction PHI nodes "
750               << "are supported currently.\n");
751    ORE->emit([&]() {
752      return OptimizationRemarkMissed(DEBUG_TYPE, "UnsupportedPHIOuter",
753                                      OuterLoop->getStartLoc(),
754                                      OuterLoop->getHeader())
755             << "Only outer loops with induction or reduction PHI nodes can be"
756                " interchanged currently.";
757    });
758    return true;
759  }
760
761  // TODO: Currently we handle only loops with 1 induction variable.
762  if (Inductions.size() != 1) {
763    LLVM_DEBUG(dbgs() << "Loops with more than 1 induction variables are not "
764                      << "supported currently.\n");
765    ORE->emit([&]() {
766      return OptimizationRemarkMissed(DEBUG_TYPE, "MultiIndutionOuter",
767                                      OuterLoop->getStartLoc(),
768                                      OuterLoop->getHeader())
769             << "Only outer loops with 1 induction variable can be "
770                "interchanged currently.";
771    });
772    return true;
773  }
774
775  Inductions.clear();
776  if (!findInductionAndReductions(InnerLoop, Inductions, nullptr)) {
777    LLVM_DEBUG(
778        dbgs() << "Only inner loops with induction or reduction PHI nodes "
779               << "are supported currently.\n");
780    ORE->emit([&]() {
781      return OptimizationRemarkMissed(DEBUG_TYPE, "UnsupportedPHIInner",
782                                      InnerLoop->getStartLoc(),
783                                      InnerLoop->getHeader())
784             << "Only inner loops with induction or reduction PHI nodes can be"
785                " interchange currently.";
786    });
787    return true;
788  }
789
790  // TODO: Currently we handle only loops with 1 induction variable.
791  if (Inductions.size() != 1) {
792    LLVM_DEBUG(
793        dbgs() << "We currently only support loops with 1 induction variable."
794               << "Failed to interchange due to current limitation\n");
795    ORE->emit([&]() {
796      return OptimizationRemarkMissed(DEBUG_TYPE, "MultiInductionInner",
797                                      InnerLoop->getStartLoc(),
798                                      InnerLoop->getHeader())
799             << "Only inner loops with 1 induction variable can be "
800                "interchanged currently.";
801    });
802    return true;
803  }
804  InnerInductionVar = Inductions.pop_back_val();
805
806  // TODO: Triangular loops are not handled for now.
807  if (!isLoopStructureUnderstood(InnerInductionVar)) {
808    LLVM_DEBUG(dbgs() << "Loop structure not understood by pass\n");
809    ORE->emit([&]() {
810      return OptimizationRemarkMissed(DEBUG_TYPE, "UnsupportedStructureInner",
811                                      InnerLoop->getStartLoc(),
812                                      InnerLoop->getHeader())
813             << "Inner loop structure not understood currently.";
814    });
815    return true;
816  }
817
818  // TODO: Current limitation: Since we split the inner loop latch at the point
819  // were induction variable is incremented (induction.next); We cannot have
820  // more than 1 user of induction.next since it would result in broken code
821  // after split.
822  // e.g.
823  // for(i=0;i<N;i++) {
824  //    for(j = 0;j<M;j++) {
825  //      A[j+1][i+2] = A[j][i]+k;
826  //  }
827  // }
828  Instruction *InnerIndexVarInc = nullptr;
829  if (InnerInductionVar->getIncomingBlock(0) == InnerLoopPreHeader)
830    InnerIndexVarInc =
831        dyn_cast<Instruction>(InnerInductionVar->getIncomingValue(1));
832  else
833    InnerIndexVarInc =
834        dyn_cast<Instruction>(InnerInductionVar->getIncomingValue(0));
835
836  if (!InnerIndexVarInc) {
837    LLVM_DEBUG(
838        dbgs() << "Did not find an instruction to increment the induction "
839               << "variable.\n");
840    ORE->emit([&]() {
841      return OptimizationRemarkMissed(DEBUG_TYPE, "NoIncrementInInner",
842                                      InnerLoop->getStartLoc(),
843                                      InnerLoop->getHeader())
844             << "The inner loop does not increment the induction variable.";
845    });
846    return true;
847  }
848
849  // Since we split the inner loop latch on this induction variable. Make sure
850  // we do not have any instruction between the induction variable and branch
851  // instruction.
852
853  bool FoundInduction = false;
854  for (const Instruction &I :
855       llvm::reverse(InnerLoopLatch->instructionsWithoutDebug())) {
856    if (isa<BranchInst>(I) || isa<CmpInst>(I) || isa<TruncInst>(I) ||
857        isa<ZExtInst>(I))
858      continue;
859
860    // We found an instruction. If this is not induction variable then it is not
861    // safe to split this loop latch.
862    if (!I.isIdenticalTo(InnerIndexVarInc)) {
863      LLVM_DEBUG(dbgs() << "Found unsupported instructions between induction "
864                        << "variable increment and branch.\n");
865      ORE->emit([&]() {
866        return OptimizationRemarkMissed(
867                   DEBUG_TYPE, "UnsupportedInsBetweenInduction",
868                   InnerLoop->getStartLoc(), InnerLoop->getHeader())
869               << "Found unsupported instruction between induction variable "
870                  "increment and branch.";
871      });
872      return true;
873    }
874
875    FoundInduction = true;
876    break;
877  }
878  // The loop latch ended and we didn't find the induction variable return as
879  // current limitation.
880  if (!FoundInduction) {
881    LLVM_DEBUG(dbgs() << "Did not find the induction variable.\n");
882    ORE->emit([&]() {
883      return OptimizationRemarkMissed(DEBUG_TYPE, "NoIndutionVariable",
884                                      InnerLoop->getStartLoc(),
885                                      InnerLoop->getHeader())
886             << "Did not find the induction variable.";
887    });
888    return true;
889  }
890  return false;
891}
892
893// We currently only support LCSSA PHI nodes in the inner loop exit, if their
894// users are either reduction PHIs or PHIs outside the outer loop (which means
895// the we are only interested in the final value after the loop).
896static bool
897areInnerLoopExitPHIsSupported(Loop *InnerL, Loop *OuterL,
898                              SmallPtrSetImpl<PHINode *> &Reductions) {
899  BasicBlock *InnerExit = OuterL->getUniqueExitBlock();
900  for (PHINode &PHI : InnerExit->phis()) {
901    // Reduction lcssa phi will have only 1 incoming block that from loop latch.
902    if (PHI.getNumIncomingValues() > 1)
903      return false;
904    if (any_of(PHI.users(), [&Reductions, OuterL](User *U) {
905          PHINode *PN = dyn_cast<PHINode>(U);
906          return !PN || (Reductions.find(PN) == Reductions.end() &&
907                         OuterL->contains(PN->getParent()));
908        })) {
909      return false;
910    }
911  }
912  return true;
913}
914
915// We currently support LCSSA PHI nodes in the outer loop exit, if their
916// incoming values do not come from the outer loop latch or if the
917// outer loop latch has a single predecessor. In that case, the value will
918// be available if both the inner and outer loop conditions are true, which
919// will still be true after interchanging. If we have multiple predecessor,
920// that may not be the case, e.g. because the outer loop latch may be executed
921// if the inner loop is not executed.
922static bool areOuterLoopExitPHIsSupported(Loop *OuterLoop, Loop *InnerLoop) {
923  BasicBlock *LoopNestExit = OuterLoop->getUniqueExitBlock();
924  for (PHINode &PHI : LoopNestExit->phis()) {
925    //  FIXME: We currently are not able to detect floating point reductions
926    //         and have to use floating point PHIs as a proxy to prevent
927    //         interchanging in the presence of floating point reductions.
928    if (PHI.getType()->isFloatingPointTy())
929      return false;
930    for (unsigned i = 0; i < PHI.getNumIncomingValues(); i++) {
931     Instruction *IncomingI = dyn_cast<Instruction>(PHI.getIncomingValue(i));
932     if (!IncomingI || IncomingI->getParent() != OuterLoop->getLoopLatch())
933       continue;
934
935     // The incoming value is defined in the outer loop latch. Currently we
936     // only support that in case the outer loop latch has a single predecessor.
937     // This guarantees that the outer loop latch is executed if and only if
938     // the inner loop is executed (because tightlyNested() guarantees that the
939     // outer loop header only branches to the inner loop or the outer loop
940     // latch).
941     // FIXME: We could weaken this logic and allow multiple predecessors,
942     //        if the values are produced outside the loop latch. We would need
943     //        additional logic to update the PHI nodes in the exit block as
944     //        well.
945     if (OuterLoop->getLoopLatch()->getUniquePredecessor() == nullptr)
946       return false;
947    }
948  }
949  return true;
950}
951
952bool LoopInterchangeLegality::canInterchangeLoops(unsigned InnerLoopId,
953                                                  unsigned OuterLoopId,
954                                                  CharMatrix &DepMatrix) {
955  if (!isLegalToInterChangeLoops(DepMatrix, InnerLoopId, OuterLoopId)) {
956    LLVM_DEBUG(dbgs() << "Failed interchange InnerLoopId = " << InnerLoopId
957                      << " and OuterLoopId = " << OuterLoopId
958                      << " due to dependence\n");
959    ORE->emit([&]() {
960      return OptimizationRemarkMissed(DEBUG_TYPE, "Dependence",
961                                      InnerLoop->getStartLoc(),
962                                      InnerLoop->getHeader())
963             << "Cannot interchange loops due to dependences.";
964    });
965    return false;
966  }
967  // Check if outer and inner loop contain legal instructions only.
968  for (auto *BB : OuterLoop->blocks())
969    for (Instruction &I : BB->instructionsWithoutDebug())
970      if (CallInst *CI = dyn_cast<CallInst>(&I)) {
971        // readnone functions do not prevent interchanging.
972        if (CI->doesNotReadMemory())
973          continue;
974        LLVM_DEBUG(
975            dbgs() << "Loops with call instructions cannot be interchanged "
976                   << "safely.");
977        ORE->emit([&]() {
978          return OptimizationRemarkMissed(DEBUG_TYPE, "CallInst",
979                                          CI->getDebugLoc(),
980                                          CI->getParent())
981                 << "Cannot interchange loops due to call instruction.";
982        });
983
984        return false;
985      }
986
987  // TODO: The loops could not be interchanged due to current limitations in the
988  // transform module.
989  if (currentLimitations()) {
990    LLVM_DEBUG(dbgs() << "Not legal because of current transform limitation\n");
991    return false;
992  }
993
994  // Check if the loops are tightly nested.
995  if (!tightlyNested(OuterLoop, InnerLoop)) {
996    LLVM_DEBUG(dbgs() << "Loops not tightly nested\n");
997    ORE->emit([&]() {
998      return OptimizationRemarkMissed(DEBUG_TYPE, "NotTightlyNested",
999                                      InnerLoop->getStartLoc(),
1000                                      InnerLoop->getHeader())
1001             << "Cannot interchange loops because they are not tightly "
1002                "nested.";
1003    });
1004    return false;
1005  }
1006
1007  if (!areInnerLoopExitPHIsSupported(OuterLoop, InnerLoop,
1008                                     OuterInnerReductions)) {
1009    LLVM_DEBUG(dbgs() << "Found unsupported PHI nodes in inner loop exit.\n");
1010    ORE->emit([&]() {
1011      return OptimizationRemarkMissed(DEBUG_TYPE, "UnsupportedExitPHI",
1012                                      InnerLoop->getStartLoc(),
1013                                      InnerLoop->getHeader())
1014             << "Found unsupported PHI node in loop exit.";
1015    });
1016    return false;
1017  }
1018
1019  if (!areOuterLoopExitPHIsSupported(OuterLoop, InnerLoop)) {
1020    LLVM_DEBUG(dbgs() << "Found unsupported PHI nodes in outer loop exit.\n");
1021    ORE->emit([&]() {
1022      return OptimizationRemarkMissed(DEBUG_TYPE, "UnsupportedExitPHI",
1023                                      OuterLoop->getStartLoc(),
1024                                      OuterLoop->getHeader())
1025             << "Found unsupported PHI node in loop exit.";
1026    });
1027    return false;
1028  }
1029
1030  return true;
1031}
1032
1033int LoopInterchangeProfitability::getInstrOrderCost() {
1034  unsigned GoodOrder, BadOrder;
1035  BadOrder = GoodOrder = 0;
1036  for (BasicBlock *BB : InnerLoop->blocks()) {
1037    for (Instruction &Ins : *BB) {
1038      if (const GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(&Ins)) {
1039        unsigned NumOp = GEP->getNumOperands();
1040        bool FoundInnerInduction = false;
1041        bool FoundOuterInduction = false;
1042        for (unsigned i = 0; i < NumOp; ++i) {
1043          const SCEV *OperandVal = SE->getSCEV(GEP->getOperand(i));
1044          const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(OperandVal);
1045          if (!AR)
1046            continue;
1047
1048          // If we find the inner induction after an outer induction e.g.
1049          // for(int i=0;i<N;i++)
1050          //   for(int j=0;j<N;j++)
1051          //     A[i][j] = A[i-1][j-1]+k;
1052          // then it is a good order.
1053          if (AR->getLoop() == InnerLoop) {
1054            // We found an InnerLoop induction after OuterLoop induction. It is
1055            // a good order.
1056            FoundInnerInduction = true;
1057            if (FoundOuterInduction) {
1058              GoodOrder++;
1059              break;
1060            }
1061          }
1062          // If we find the outer induction after an inner induction e.g.
1063          // for(int i=0;i<N;i++)
1064          //   for(int j=0;j<N;j++)
1065          //     A[j][i] = A[j-1][i-1]+k;
1066          // then it is a bad order.
1067          if (AR->getLoop() == OuterLoop) {
1068            // We found an OuterLoop induction after InnerLoop induction. It is
1069            // a bad order.
1070            FoundOuterInduction = true;
1071            if (FoundInnerInduction) {
1072              BadOrder++;
1073              break;
1074            }
1075          }
1076        }
1077      }
1078    }
1079  }
1080  return GoodOrder - BadOrder;
1081}
1082
1083static bool isProfitableForVectorization(unsigned InnerLoopId,
1084                                         unsigned OuterLoopId,
1085                                         CharMatrix &DepMatrix) {
1086  // TODO: Improve this heuristic to catch more cases.
1087  // If the inner loop is loop independent or doesn't carry any dependency it is
1088  // profitable to move this to outer position.
1089  for (auto &Row : DepMatrix) {
1090    if (Row[InnerLoopId] != 'S' && Row[InnerLoopId] != 'I')
1091      return false;
1092    // TODO: We need to improve this heuristic.
1093    if (Row[OuterLoopId] != '=')
1094      return false;
1095  }
1096  // If outer loop has dependence and inner loop is loop independent then it is
1097  // profitable to interchange to enable parallelism.
1098  // If there are no dependences, interchanging will not improve anything.
1099  return !DepMatrix.empty();
1100}
1101
1102bool LoopInterchangeProfitability::isProfitable(unsigned InnerLoopId,
1103                                                unsigned OuterLoopId,
1104                                                CharMatrix &DepMatrix) {
1105  // TODO: Add better profitability checks.
1106  // e.g
1107  // 1) Construct dependency matrix and move the one with no loop carried dep
1108  //    inside to enable vectorization.
1109
1110  // This is rough cost estimation algorithm. It counts the good and bad order
1111  // of induction variables in the instruction and allows reordering if number
1112  // of bad orders is more than good.
1113  int Cost = getInstrOrderCost();
1114  LLVM_DEBUG(dbgs() << "Cost = " << Cost << "\n");
1115  if (Cost < -LoopInterchangeCostThreshold)
1116    return true;
1117
1118  // It is not profitable as per current cache profitability model. But check if
1119  // we can move this loop outside to improve parallelism.
1120  if (isProfitableForVectorization(InnerLoopId, OuterLoopId, DepMatrix))
1121    return true;
1122
1123  ORE->emit([&]() {
1124    return OptimizationRemarkMissed(DEBUG_TYPE, "InterchangeNotProfitable",
1125                                    InnerLoop->getStartLoc(),
1126                                    InnerLoop->getHeader())
1127           << "Interchanging loops is too costly (cost="
1128           << ore::NV("Cost", Cost) << ", threshold="
1129           << ore::NV("Threshold", LoopInterchangeCostThreshold)
1130           << ") and it does not improve parallelism.";
1131  });
1132  return false;
1133}
1134
1135void LoopInterchangeTransform::removeChildLoop(Loop *OuterLoop,
1136                                               Loop *InnerLoop) {
1137  for (Loop *L : *OuterLoop)
1138    if (L == InnerLoop) {
1139      OuterLoop->removeChildLoop(L);
1140      return;
1141    }
1142  llvm_unreachable("Couldn't find loop");
1143}
1144
1145/// Update LoopInfo, after interchanging. NewInner and NewOuter refer to the
1146/// new inner and outer loop after interchanging: NewInner is the original
1147/// outer loop and NewOuter is the original inner loop.
1148///
1149/// Before interchanging, we have the following structure
1150/// Outer preheader
1151//  Outer header
1152//    Inner preheader
1153//    Inner header
1154//      Inner body
1155//      Inner latch
1156//   outer bbs
1157//   Outer latch
1158//
1159// After interchanging:
1160// Inner preheader
1161// Inner header
1162//   Outer preheader
1163//   Outer header
1164//     Inner body
1165//     outer bbs
1166//     Outer latch
1167//   Inner latch
1168void LoopInterchangeTransform::restructureLoops(
1169    Loop *NewInner, Loop *NewOuter, BasicBlock *OrigInnerPreHeader,
1170    BasicBlock *OrigOuterPreHeader) {
1171  Loop *OuterLoopParent = OuterLoop->getParentLoop();
1172  // The original inner loop preheader moves from the new inner loop to
1173  // the parent loop, if there is one.
1174  NewInner->removeBlockFromLoop(OrigInnerPreHeader);
1175  LI->changeLoopFor(OrigInnerPreHeader, OuterLoopParent);
1176
1177  // Switch the loop levels.
1178  if (OuterLoopParent) {
1179    // Remove the loop from its parent loop.
1180    removeChildLoop(OuterLoopParent, NewInner);
1181    removeChildLoop(NewInner, NewOuter);
1182    OuterLoopParent->addChildLoop(NewOuter);
1183  } else {
1184    removeChildLoop(NewInner, NewOuter);
1185    LI->changeTopLevelLoop(NewInner, NewOuter);
1186  }
1187  while (!NewOuter->empty())
1188    NewInner->addChildLoop(NewOuter->removeChildLoop(NewOuter->begin()));
1189  NewOuter->addChildLoop(NewInner);
1190
1191  // BBs from the original inner loop.
1192  SmallVector<BasicBlock *, 8> OrigInnerBBs(NewOuter->blocks());
1193
1194  // Add BBs from the original outer loop to the original inner loop (excluding
1195  // BBs already in inner loop)
1196  for (BasicBlock *BB : NewInner->blocks())
1197    if (LI->getLoopFor(BB) == NewInner)
1198      NewOuter->addBlockEntry(BB);
1199
1200  // Now remove inner loop header and latch from the new inner loop and move
1201  // other BBs (the loop body) to the new inner loop.
1202  BasicBlock *OuterHeader = NewOuter->getHeader();
1203  BasicBlock *OuterLatch = NewOuter->getLoopLatch();
1204  for (BasicBlock *BB : OrigInnerBBs) {
1205    // Nothing will change for BBs in child loops.
1206    if (LI->getLoopFor(BB) != NewOuter)
1207      continue;
1208    // Remove the new outer loop header and latch from the new inner loop.
1209    if (BB == OuterHeader || BB == OuterLatch)
1210      NewInner->removeBlockFromLoop(BB);
1211    else
1212      LI->changeLoopFor(BB, NewInner);
1213  }
1214
1215  // The preheader of the original outer loop becomes part of the new
1216  // outer loop.
1217  NewOuter->addBlockEntry(OrigOuterPreHeader);
1218  LI->changeLoopFor(OrigOuterPreHeader, NewOuter);
1219
1220  // Tell SE that we move the loops around.
1221  SE->forgetLoop(NewOuter);
1222  SE->forgetLoop(NewInner);
1223}
1224
1225bool LoopInterchangeTransform::transform() {
1226  bool Transformed = false;
1227  Instruction *InnerIndexVar;
1228
1229  if (InnerLoop->getSubLoops().empty()) {
1230    BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader();
1231    LLVM_DEBUG(dbgs() << "Splitting the inner loop latch\n");
1232    PHINode *InductionPHI = getInductionVariable(InnerLoop, SE);
1233    if (!InductionPHI) {
1234      LLVM_DEBUG(dbgs() << "Failed to find the point to split loop latch \n");
1235      return false;
1236    }
1237
1238    if (InductionPHI->getIncomingBlock(0) == InnerLoopPreHeader)
1239      InnerIndexVar = dyn_cast<Instruction>(InductionPHI->getIncomingValue(1));
1240    else
1241      InnerIndexVar = dyn_cast<Instruction>(InductionPHI->getIncomingValue(0));
1242
1243    // Ensure that InductionPHI is the first Phi node.
1244    if (&InductionPHI->getParent()->front() != InductionPHI)
1245      InductionPHI->moveBefore(&InductionPHI->getParent()->front());
1246
1247    // Create a new latch block for the inner loop. We split at the
1248    // current latch's terminator and then move the condition and all
1249    // operands that are not either loop-invariant or the induction PHI into the
1250    // new latch block.
1251    BasicBlock *NewLatch =
1252        SplitBlock(InnerLoop->getLoopLatch(),
1253                   InnerLoop->getLoopLatch()->getTerminator(), DT, LI);
1254
1255    SmallSetVector<Instruction *, 4> WorkList;
1256    unsigned i = 0;
1257    auto MoveInstructions = [&i, &WorkList, this, InductionPHI, NewLatch]() {
1258      for (; i < WorkList.size(); i++) {
1259        // Duplicate instruction and move it the new latch. Update uses that
1260        // have been moved.
1261        Instruction *NewI = WorkList[i]->clone();
1262        NewI->insertBefore(NewLatch->getFirstNonPHI());
1263        assert(!NewI->mayHaveSideEffects() &&
1264               "Moving instructions with side-effects may change behavior of "
1265               "the loop nest!");
1266        for (auto UI = WorkList[i]->use_begin(), UE = WorkList[i]->use_end();
1267             UI != UE;) {
1268          Use &U = *UI++;
1269          Instruction *UserI = cast<Instruction>(U.getUser());
1270          if (!InnerLoop->contains(UserI->getParent()) ||
1271              UserI->getParent() == NewLatch || UserI == InductionPHI)
1272            U.set(NewI);
1273        }
1274        // Add operands of moved instruction to the worklist, except if they are
1275        // outside the inner loop or are the induction PHI.
1276        for (Value *Op : WorkList[i]->operands()) {
1277          Instruction *OpI = dyn_cast<Instruction>(Op);
1278          if (!OpI ||
1279              this->LI->getLoopFor(OpI->getParent()) != this->InnerLoop ||
1280              OpI == InductionPHI)
1281            continue;
1282          WorkList.insert(OpI);
1283        }
1284      }
1285    };
1286
1287    // FIXME: Should we interchange when we have a constant condition?
1288    Instruction *CondI = dyn_cast<Instruction>(
1289        cast<BranchInst>(InnerLoop->getLoopLatch()->getTerminator())
1290            ->getCondition());
1291    if (CondI)
1292      WorkList.insert(CondI);
1293    MoveInstructions();
1294    WorkList.insert(cast<Instruction>(InnerIndexVar));
1295    MoveInstructions();
1296
1297    // Splits the inner loops phi nodes out into a separate basic block.
1298    BasicBlock *InnerLoopHeader = InnerLoop->getHeader();
1299    SplitBlock(InnerLoopHeader, InnerLoopHeader->getFirstNonPHI(), DT, LI);
1300    LLVM_DEBUG(dbgs() << "splitting InnerLoopHeader done\n");
1301  }
1302
1303  Transformed |= adjustLoopLinks();
1304  if (!Transformed) {
1305    LLVM_DEBUG(dbgs() << "adjustLoopLinks failed\n");
1306    return false;
1307  }
1308
1309  return true;
1310}
1311
1312/// \brief Move all instructions except the terminator from FromBB right before
1313/// InsertBefore
1314static void moveBBContents(BasicBlock *FromBB, Instruction *InsertBefore) {
1315  auto &ToList = InsertBefore->getParent()->getInstList();
1316  auto &FromList = FromBB->getInstList();
1317
1318  ToList.splice(InsertBefore->getIterator(), FromList, FromList.begin(),
1319                FromBB->getTerminator()->getIterator());
1320}
1321
1322// Update BI to jump to NewBB instead of OldBB. Records updates to the
1323// dominator tree in DTUpdates. If \p MustUpdateOnce is true, assert that
1324// \p OldBB  is exactly once in BI's successor list.
1325static void updateSuccessor(BranchInst *BI, BasicBlock *OldBB,
1326                            BasicBlock *NewBB,
1327                            std::vector<DominatorTree::UpdateType> &DTUpdates,
1328                            bool MustUpdateOnce = true) {
1329  assert((!MustUpdateOnce ||
1330          llvm::count_if(successors(BI),
1331                         [OldBB](BasicBlock *BB) {
1332                           return BB == OldBB;
1333                         }) == 1) && "BI must jump to OldBB exactly once.");
1334  bool Changed = false;
1335  for (Use &Op : BI->operands())
1336    if (Op == OldBB) {
1337      Op.set(NewBB);
1338      Changed = true;
1339    }
1340
1341  if (Changed) {
1342    DTUpdates.push_back(
1343        {DominatorTree::UpdateKind::Insert, BI->getParent(), NewBB});
1344    DTUpdates.push_back(
1345        {DominatorTree::UpdateKind::Delete, BI->getParent(), OldBB});
1346  }
1347  assert(Changed && "Expected a successor to be updated");
1348}
1349
1350// Move Lcssa PHIs to the right place.
1351static void moveLCSSAPhis(BasicBlock *InnerExit, BasicBlock *InnerHeader,
1352                          BasicBlock *InnerLatch, BasicBlock *OuterHeader,
1353                          BasicBlock *OuterLatch, BasicBlock *OuterExit,
1354                          Loop *InnerLoop, LoopInfo *LI) {
1355
1356  // Deal with LCSSA PHI nodes in the exit block of the inner loop, that are
1357  // defined either in the header or latch. Those blocks will become header and
1358  // latch of the new outer loop, and the only possible users can PHI nodes
1359  // in the exit block of the loop nest or the outer loop header (reduction
1360  // PHIs, in that case, the incoming value must be defined in the inner loop
1361  // header). We can just substitute the user with the incoming value and remove
1362  // the PHI.
1363  for (PHINode &P : make_early_inc_range(InnerExit->phis())) {
1364    assert(P.getNumIncomingValues() == 1 &&
1365           "Only loops with a single exit are supported!");
1366
1367    // Incoming values are guaranteed be instructions currently.
1368    auto IncI = cast<Instruction>(P.getIncomingValueForBlock(InnerLatch));
1369    // Skip phis with incoming values from the inner loop body, excluding the
1370    // header and latch.
1371    if (IncI->getParent() != InnerLatch && IncI->getParent() != InnerHeader)
1372      continue;
1373
1374    assert(all_of(P.users(),
1375                  [OuterHeader, OuterExit, IncI, InnerHeader](User *U) {
1376                    return (cast<PHINode>(U)->getParent() == OuterHeader &&
1377                            IncI->getParent() == InnerHeader) ||
1378                           cast<PHINode>(U)->getParent() == OuterExit;
1379                  }) &&
1380           "Can only replace phis iff the uses are in the loop nest exit or "
1381           "the incoming value is defined in the inner header (it will "
1382           "dominate all loop blocks after interchanging)");
1383    P.replaceAllUsesWith(IncI);
1384    P.eraseFromParent();
1385  }
1386
1387  SmallVector<PHINode *, 8> LcssaInnerExit;
1388  for (PHINode &P : InnerExit->phis())
1389    LcssaInnerExit.push_back(&P);
1390
1391  SmallVector<PHINode *, 8> LcssaInnerLatch;
1392  for (PHINode &P : InnerLatch->phis())
1393    LcssaInnerLatch.push_back(&P);
1394
1395  // Lcssa PHIs for values used outside the inner loop are in InnerExit.
1396  // If a PHI node has users outside of InnerExit, it has a use outside the
1397  // interchanged loop and we have to preserve it. We move these to
1398  // InnerLatch, which will become the new exit block for the innermost
1399  // loop after interchanging.
1400  for (PHINode *P : LcssaInnerExit)
1401    P->moveBefore(InnerLatch->getFirstNonPHI());
1402
1403  // If the inner loop latch contains LCSSA PHIs, those come from a child loop
1404  // and we have to move them to the new inner latch.
1405  for (PHINode *P : LcssaInnerLatch)
1406    P->moveBefore(InnerExit->getFirstNonPHI());
1407
1408  // Deal with LCSSA PHI nodes in the loop nest exit block. For PHIs that have
1409  // incoming values defined in the outer loop, we have to add a new PHI
1410  // in the inner loop latch, which became the exit block of the outer loop,
1411  // after interchanging.
1412  if (OuterExit) {
1413    for (PHINode &P : OuterExit->phis()) {
1414      if (P.getNumIncomingValues() != 1)
1415        continue;
1416      // Skip Phis with incoming values defined in the inner loop. Those should
1417      // already have been updated.
1418      auto I = dyn_cast<Instruction>(P.getIncomingValue(0));
1419      if (!I || LI->getLoopFor(I->getParent()) == InnerLoop)
1420        continue;
1421
1422      PHINode *NewPhi = dyn_cast<PHINode>(P.clone());
1423      NewPhi->setIncomingValue(0, P.getIncomingValue(0));
1424      NewPhi->setIncomingBlock(0, OuterLatch);
1425      NewPhi->insertBefore(InnerLatch->getFirstNonPHI());
1426      P.setIncomingValue(0, NewPhi);
1427    }
1428  }
1429
1430  // Now adjust the incoming blocks for the LCSSA PHIs.
1431  // For PHIs moved from Inner's exit block, we need to replace Inner's latch
1432  // with the new latch.
1433  InnerLatch->replacePhiUsesWith(InnerLatch, OuterLatch);
1434}
1435
1436bool LoopInterchangeTransform::adjustLoopBranches() {
1437  LLVM_DEBUG(dbgs() << "adjustLoopBranches called\n");
1438  std::vector<DominatorTree::UpdateType> DTUpdates;
1439
1440  BasicBlock *OuterLoopPreHeader = OuterLoop->getLoopPreheader();
1441  BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader();
1442
1443  assert(OuterLoopPreHeader != OuterLoop->getHeader() &&
1444         InnerLoopPreHeader != InnerLoop->getHeader() && OuterLoopPreHeader &&
1445         InnerLoopPreHeader && "Guaranteed by loop-simplify form");
1446  // Ensure that both preheaders do not contain PHI nodes and have single
1447  // predecessors. This allows us to move them easily. We use
1448  // InsertPreHeaderForLoop to create an 'extra' preheader, if the existing
1449  // preheaders do not satisfy those conditions.
1450  if (isa<PHINode>(OuterLoopPreHeader->begin()) ||
1451      !OuterLoopPreHeader->getUniquePredecessor())
1452    OuterLoopPreHeader =
1453        InsertPreheaderForLoop(OuterLoop, DT, LI, nullptr, true);
1454  if (InnerLoopPreHeader == OuterLoop->getHeader())
1455    InnerLoopPreHeader =
1456        InsertPreheaderForLoop(InnerLoop, DT, LI, nullptr, true);
1457
1458  // Adjust the loop preheader
1459  BasicBlock *InnerLoopHeader = InnerLoop->getHeader();
1460  BasicBlock *OuterLoopHeader = OuterLoop->getHeader();
1461  BasicBlock *InnerLoopLatch = InnerLoop->getLoopLatch();
1462  BasicBlock *OuterLoopLatch = OuterLoop->getLoopLatch();
1463  BasicBlock *OuterLoopPredecessor = OuterLoopPreHeader->getUniquePredecessor();
1464  BasicBlock *InnerLoopLatchPredecessor =
1465      InnerLoopLatch->getUniquePredecessor();
1466  BasicBlock *InnerLoopLatchSuccessor;
1467  BasicBlock *OuterLoopLatchSuccessor;
1468
1469  BranchInst *OuterLoopLatchBI =
1470      dyn_cast<BranchInst>(OuterLoopLatch->getTerminator());
1471  BranchInst *InnerLoopLatchBI =
1472      dyn_cast<BranchInst>(InnerLoopLatch->getTerminator());
1473  BranchInst *OuterLoopHeaderBI =
1474      dyn_cast<BranchInst>(OuterLoopHeader->getTerminator());
1475  BranchInst *InnerLoopHeaderBI =
1476      dyn_cast<BranchInst>(InnerLoopHeader->getTerminator());
1477
1478  if (!OuterLoopPredecessor || !InnerLoopLatchPredecessor ||
1479      !OuterLoopLatchBI || !InnerLoopLatchBI || !OuterLoopHeaderBI ||
1480      !InnerLoopHeaderBI)
1481    return false;
1482
1483  BranchInst *InnerLoopLatchPredecessorBI =
1484      dyn_cast<BranchInst>(InnerLoopLatchPredecessor->getTerminator());
1485  BranchInst *OuterLoopPredecessorBI =
1486      dyn_cast<BranchInst>(OuterLoopPredecessor->getTerminator());
1487
1488  if (!OuterLoopPredecessorBI || !InnerLoopLatchPredecessorBI)
1489    return false;
1490  BasicBlock *InnerLoopHeaderSuccessor = InnerLoopHeader->getUniqueSuccessor();
1491  if (!InnerLoopHeaderSuccessor)
1492    return false;
1493
1494  // Adjust Loop Preheader and headers.
1495  // The branches in the outer loop predecessor and the outer loop header can
1496  // be unconditional branches or conditional branches with duplicates. Consider
1497  // this when updating the successors.
1498  updateSuccessor(OuterLoopPredecessorBI, OuterLoopPreHeader,
1499                  InnerLoopPreHeader, DTUpdates, /*MustUpdateOnce=*/false);
1500  // The outer loop header might or might not branch to the outer latch.
1501  // We are guaranteed to branch to the inner loop preheader.
1502  if (std::find(succ_begin(OuterLoopHeaderBI), succ_end(OuterLoopHeaderBI),
1503                OuterLoopLatch) != succ_end(OuterLoopHeaderBI))
1504    updateSuccessor(OuterLoopHeaderBI, OuterLoopLatch, LoopExit, DTUpdates,
1505                    /*MustUpdateOnce=*/false);
1506  updateSuccessor(OuterLoopHeaderBI, InnerLoopPreHeader,
1507                  InnerLoopHeaderSuccessor, DTUpdates,
1508                  /*MustUpdateOnce=*/false);
1509
1510  // Adjust reduction PHI's now that the incoming block has changed.
1511  InnerLoopHeaderSuccessor->replacePhiUsesWith(InnerLoopHeader,
1512                                               OuterLoopHeader);
1513
1514  updateSuccessor(InnerLoopHeaderBI, InnerLoopHeaderSuccessor,
1515                  OuterLoopPreHeader, DTUpdates);
1516
1517  // -------------Adjust loop latches-----------
1518  if (InnerLoopLatchBI->getSuccessor(0) == InnerLoopHeader)
1519    InnerLoopLatchSuccessor = InnerLoopLatchBI->getSuccessor(1);
1520  else
1521    InnerLoopLatchSuccessor = InnerLoopLatchBI->getSuccessor(0);
1522
1523  updateSuccessor(InnerLoopLatchPredecessorBI, InnerLoopLatch,
1524                  InnerLoopLatchSuccessor, DTUpdates);
1525
1526
1527  if (OuterLoopLatchBI->getSuccessor(0) == OuterLoopHeader)
1528    OuterLoopLatchSuccessor = OuterLoopLatchBI->getSuccessor(1);
1529  else
1530    OuterLoopLatchSuccessor = OuterLoopLatchBI->getSuccessor(0);
1531
1532  updateSuccessor(InnerLoopLatchBI, InnerLoopLatchSuccessor,
1533                  OuterLoopLatchSuccessor, DTUpdates);
1534  updateSuccessor(OuterLoopLatchBI, OuterLoopLatchSuccessor, InnerLoopLatch,
1535                  DTUpdates);
1536
1537  DT->applyUpdates(DTUpdates);
1538  restructureLoops(OuterLoop, InnerLoop, InnerLoopPreHeader,
1539                   OuterLoopPreHeader);
1540
1541  moveLCSSAPhis(InnerLoopLatchSuccessor, InnerLoopHeader, InnerLoopLatch,
1542                OuterLoopHeader, OuterLoopLatch, InnerLoop->getExitBlock(),
1543                InnerLoop, LI);
1544  // For PHIs in the exit block of the outer loop, outer's latch has been
1545  // replaced by Inners'.
1546  OuterLoopLatchSuccessor->replacePhiUsesWith(OuterLoopLatch, InnerLoopLatch);
1547
1548  // Now update the reduction PHIs in the inner and outer loop headers.
1549  SmallVector<PHINode *, 4> InnerLoopPHIs, OuterLoopPHIs;
1550  for (PHINode &PHI : drop_begin(InnerLoopHeader->phis(), 1))
1551    InnerLoopPHIs.push_back(cast<PHINode>(&PHI));
1552  for (PHINode &PHI : drop_begin(OuterLoopHeader->phis(), 1))
1553    OuterLoopPHIs.push_back(cast<PHINode>(&PHI));
1554
1555  auto &OuterInnerReductions = LIL.getOuterInnerReductions();
1556  (void)OuterInnerReductions;
1557
1558  // Now move the remaining reduction PHIs from outer to inner loop header and
1559  // vice versa. The PHI nodes must be part of a reduction across the inner and
1560  // outer loop and all the remains to do is and updating the incoming blocks.
1561  for (PHINode *PHI : OuterLoopPHIs) {
1562    PHI->moveBefore(InnerLoopHeader->getFirstNonPHI());
1563    assert(OuterInnerReductions.find(PHI) != OuterInnerReductions.end() &&
1564           "Expected a reduction PHI node");
1565  }
1566  for (PHINode *PHI : InnerLoopPHIs) {
1567    PHI->moveBefore(OuterLoopHeader->getFirstNonPHI());
1568    assert(OuterInnerReductions.find(PHI) != OuterInnerReductions.end() &&
1569           "Expected a reduction PHI node");
1570  }
1571
1572  // Update the incoming blocks for moved PHI nodes.
1573  OuterLoopHeader->replacePhiUsesWith(InnerLoopPreHeader, OuterLoopPreHeader);
1574  OuterLoopHeader->replacePhiUsesWith(InnerLoopLatch, OuterLoopLatch);
1575  InnerLoopHeader->replacePhiUsesWith(OuterLoopPreHeader, InnerLoopPreHeader);
1576  InnerLoopHeader->replacePhiUsesWith(OuterLoopLatch, InnerLoopLatch);
1577
1578  return true;
1579}
1580
1581void LoopInterchangeTransform::adjustLoopPreheaders() {
1582  // We have interchanged the preheaders so we need to interchange the data in
1583  // the preheader as well.
1584  // This is because the content of inner preheader was previously executed
1585  // inside the outer loop.
1586  BasicBlock *OuterLoopPreHeader = OuterLoop->getLoopPreheader();
1587  BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader();
1588  BasicBlock *OuterLoopHeader = OuterLoop->getHeader();
1589  BranchInst *InnerTermBI =
1590      cast<BranchInst>(InnerLoopPreHeader->getTerminator());
1591
1592  // These instructions should now be executed inside the loop.
1593  // Move instruction into a new block after outer header.
1594  moveBBContents(InnerLoopPreHeader, OuterLoopHeader->getTerminator());
1595  // These instructions were not executed previously in the loop so move them to
1596  // the older inner loop preheader.
1597  moveBBContents(OuterLoopPreHeader, InnerTermBI);
1598}
1599
1600bool LoopInterchangeTransform::adjustLoopLinks() {
1601  // Adjust all branches in the inner and outer loop.
1602  bool Changed = adjustLoopBranches();
1603  if (Changed)
1604    adjustLoopPreheaders();
1605  return Changed;
1606}
1607
1608char LoopInterchange::ID = 0;
1609
1610INITIALIZE_PASS_BEGIN(LoopInterchange, "loop-interchange",
1611                      "Interchanges loops for cache reuse", false, false)
1612INITIALIZE_PASS_DEPENDENCY(LoopPass)
1613INITIALIZE_PASS_DEPENDENCY(DependenceAnalysisWrapperPass)
1614INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass)
1615
1616INITIALIZE_PASS_END(LoopInterchange, "loop-interchange",
1617                    "Interchanges loops for cache reuse", false, false)
1618
1619Pass *llvm::createLoopInterchangePass() { return new LoopInterchange(); }
1620