xref: /netbsd-src/external/apache2/llvm/dist/llvm/lib/Transforms/Scalar/LoopInterchange.cpp (revision 82d56013d7b633d116a93943de88e08335357a7c)
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
printDepMatrix(CharMatrix & DepMatrix)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 
populateDependencyMatrix(CharMatrix & DepMatrix,unsigned Level,Loop * L,DependenceInfo * DI)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.
interChangeDependencies(CharMatrix & DepMatrix,unsigned FromIndx,unsigned ToIndx)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 // '>'
isOuterMostDepPositive(CharMatrix & DepMatrix,unsigned Row,unsigned Column)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.
containsNoDependence(CharMatrix & DepMatrix,unsigned Row,unsigned 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 
validDepInterchange(CharMatrix & DepMatrix,unsigned Row,unsigned OuterLoopId,char InnerDep,char OuterDep)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.
isLegalToInterChangeLoops(CharMatrix & DepMatrix,unsigned InnerLoopId,unsigned OuterLoopId)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 
populateWorklist(Loop & L)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 
getInductionVariable(Loop * L,ScalarEvolution * SE)295 static PHINode *getInductionVariable(Loop *L, ScalarEvolution *SE) {
296   PHINode *InnerIndexVar = L->getCanonicalInductionVariable();
297   if (InnerIndexVar)
298     return InnerIndexVar;
299   if (L->getLoopLatch() == nullptr || L->getLoopPredecessor() == nullptr)
300     return nullptr;
301   for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
302     PHINode *PhiVar = cast<PHINode>(I);
303     Type *PhiTy = PhiVar->getType();
304     if (!PhiTy->isIntegerTy() && !PhiTy->isFloatingPointTy() &&
305         !PhiTy->isPointerTy())
306       return nullptr;
307     const SCEVAddRecExpr *AddRec =
308         dyn_cast<SCEVAddRecExpr>(SE->getSCEV(PhiVar));
309     if (!AddRec || !AddRec->isAffine())
310       continue;
311     const SCEV *Step = AddRec->getStepRecurrence(*SE);
312     if (!isa<SCEVConstant>(Step))
313       continue;
314     // Found the induction variable.
315     // FIXME: Handle loops with more than one induction variable. Note that,
316     // currently, legality makes sure we have only one induction variable.
317     return PhiVar;
318   }
319   return nullptr;
320 }
321 
322 namespace {
323 
324 /// LoopInterchangeLegality checks if it is legal to interchange the loop.
325 class LoopInterchangeLegality {
326 public:
LoopInterchangeLegality(Loop * Outer,Loop * Inner,ScalarEvolution * SE,OptimizationRemarkEmitter * ORE)327   LoopInterchangeLegality(Loop *Outer, Loop *Inner, ScalarEvolution *SE,
328                           OptimizationRemarkEmitter *ORE)
329       : OuterLoop(Outer), InnerLoop(Inner), SE(SE), ORE(ORE) {}
330 
331   /// Check if the loops can be interchanged.
332   bool canInterchangeLoops(unsigned InnerLoopId, unsigned OuterLoopId,
333                            CharMatrix &DepMatrix);
334 
335   /// Check if the loop structure is understood. We do not handle triangular
336   /// loops for now.
337   bool isLoopStructureUnderstood(PHINode *InnerInductionVar);
338 
339   bool currentLimitations();
340 
getOuterInnerReductions() const341   const SmallPtrSetImpl<PHINode *> &getOuterInnerReductions() const {
342     return OuterInnerReductions;
343   }
344 
345 private:
346   bool tightlyNested(Loop *Outer, Loop *Inner);
347   bool containsUnsafeInstructions(BasicBlock *BB);
348 
349   /// Discover induction and reduction PHIs in the header of \p L. Induction
350   /// PHIs are added to \p Inductions, reductions are added to
351   /// OuterInnerReductions. When the outer loop is passed, the inner loop needs
352   /// to be passed as \p InnerLoop.
353   bool findInductionAndReductions(Loop *L,
354                                   SmallVector<PHINode *, 8> &Inductions,
355                                   Loop *InnerLoop);
356 
357   Loop *OuterLoop;
358   Loop *InnerLoop;
359 
360   ScalarEvolution *SE;
361 
362   /// Interface to emit optimization remarks.
363   OptimizationRemarkEmitter *ORE;
364 
365   /// Set of reduction PHIs taking part of a reduction across the inner and
366   /// outer loop.
367   SmallPtrSet<PHINode *, 4> OuterInnerReductions;
368 };
369 
370 /// LoopInterchangeProfitability checks if it is profitable to interchange the
371 /// loop.
372 class LoopInterchangeProfitability {
373 public:
LoopInterchangeProfitability(Loop * Outer,Loop * Inner,ScalarEvolution * SE,OptimizationRemarkEmitter * ORE)374   LoopInterchangeProfitability(Loop *Outer, Loop *Inner, ScalarEvolution *SE,
375                                OptimizationRemarkEmitter *ORE)
376       : OuterLoop(Outer), InnerLoop(Inner), SE(SE), ORE(ORE) {}
377 
378   /// Check if the loop interchange is profitable.
379   bool isProfitable(unsigned InnerLoopId, unsigned OuterLoopId,
380                     CharMatrix &DepMatrix);
381 
382 private:
383   int getInstrOrderCost();
384 
385   Loop *OuterLoop;
386   Loop *InnerLoop;
387 
388   /// Scev analysis.
389   ScalarEvolution *SE;
390 
391   /// Interface to emit optimization remarks.
392   OptimizationRemarkEmitter *ORE;
393 };
394 
395 /// LoopInterchangeTransform interchanges the loop.
396 class LoopInterchangeTransform {
397 public:
LoopInterchangeTransform(Loop * Outer,Loop * Inner,ScalarEvolution * SE,LoopInfo * LI,DominatorTree * DT,const LoopInterchangeLegality & LIL)398   LoopInterchangeTransform(Loop *Outer, Loop *Inner, ScalarEvolution *SE,
399                            LoopInfo *LI, DominatorTree *DT,
400                            const LoopInterchangeLegality &LIL)
401       : OuterLoop(Outer), InnerLoop(Inner), SE(SE), LI(LI), DT(DT), LIL(LIL) {}
402 
403   /// Interchange OuterLoop and InnerLoop.
404   bool transform();
405   void restructureLoops(Loop *NewInner, Loop *NewOuter,
406                         BasicBlock *OrigInnerPreHeader,
407                         BasicBlock *OrigOuterPreHeader);
408   void removeChildLoop(Loop *OuterLoop, Loop *InnerLoop);
409 
410 private:
411   bool adjustLoopLinks();
412   bool adjustLoopBranches();
413 
414   Loop *OuterLoop;
415   Loop *InnerLoop;
416 
417   /// Scev analysis.
418   ScalarEvolution *SE;
419 
420   LoopInfo *LI;
421   DominatorTree *DT;
422 
423   const LoopInterchangeLegality &LIL;
424 };
425 
426 struct LoopInterchange {
427   ScalarEvolution *SE = nullptr;
428   LoopInfo *LI = nullptr;
429   DependenceInfo *DI = nullptr;
430   DominatorTree *DT = nullptr;
431 
432   /// Interface to emit optimization remarks.
433   OptimizationRemarkEmitter *ORE;
434 
LoopInterchange__anonc2e4e0200211::LoopInterchange435   LoopInterchange(ScalarEvolution *SE, LoopInfo *LI, DependenceInfo *DI,
436                   DominatorTree *DT, OptimizationRemarkEmitter *ORE)
437       : SE(SE), LI(LI), DI(DI), DT(DT), ORE(ORE) {}
438 
run__anonc2e4e0200211::LoopInterchange439   bool run(Loop *L) {
440     if (L->getParentLoop())
441       return false;
442 
443     return processLoopList(populateWorklist(*L));
444   }
445 
run__anonc2e4e0200211::LoopInterchange446   bool run(LoopNest &LN) {
447     const auto &LoopList = LN.getLoops();
448     for (unsigned I = 1; I < LoopList.size(); ++I)
449       if (LoopList[I]->getParentLoop() != LoopList[I - 1])
450         return false;
451     return processLoopList(LoopList);
452   }
453 
isComputableLoopNest__anonc2e4e0200211::LoopInterchange454   bool isComputableLoopNest(ArrayRef<Loop *> LoopList) {
455     for (Loop *L : LoopList) {
456       const SCEV *ExitCountOuter = SE->getBackedgeTakenCount(L);
457       if (isa<SCEVCouldNotCompute>(ExitCountOuter)) {
458         LLVM_DEBUG(dbgs() << "Couldn't compute backedge count\n");
459         return false;
460       }
461       if (L->getNumBackEdges() != 1) {
462         LLVM_DEBUG(dbgs() << "NumBackEdges is not equal to 1\n");
463         return false;
464       }
465       if (!L->getExitingBlock()) {
466         LLVM_DEBUG(dbgs() << "Loop doesn't have unique exit block\n");
467         return false;
468       }
469     }
470     return true;
471   }
472 
selectLoopForInterchange__anonc2e4e0200211::LoopInterchange473   unsigned selectLoopForInterchange(ArrayRef<Loop *> LoopList) {
474     // TODO: Add a better heuristic to select the loop to be interchanged based
475     // on the dependence matrix. Currently we select the innermost loop.
476     return LoopList.size() - 1;
477   }
478 
processLoopList__anonc2e4e0200211::LoopInterchange479   bool processLoopList(ArrayRef<Loop *> LoopList) {
480     bool Changed = false;
481     unsigned LoopNestDepth = LoopList.size();
482     if (LoopNestDepth < 2) {
483       LLVM_DEBUG(dbgs() << "Loop doesn't contain minimum nesting level.\n");
484       return false;
485     }
486     if (LoopNestDepth > MaxLoopNestDepth) {
487       LLVM_DEBUG(dbgs() << "Cannot handle loops of depth greater than "
488                         << MaxLoopNestDepth << "\n");
489       return false;
490     }
491     if (!isComputableLoopNest(LoopList)) {
492       LLVM_DEBUG(dbgs() << "Not valid loop candidate for interchange\n");
493       return false;
494     }
495 
496     LLVM_DEBUG(dbgs() << "Processing LoopList of size = " << LoopNestDepth
497                       << "\n");
498 
499     CharMatrix DependencyMatrix;
500     Loop *OuterMostLoop = *(LoopList.begin());
501     if (!populateDependencyMatrix(DependencyMatrix, LoopNestDepth,
502                                   OuterMostLoop, DI)) {
503       LLVM_DEBUG(dbgs() << "Populating dependency matrix failed\n");
504       return false;
505     }
506 #ifdef DUMP_DEP_MATRICIES
507     LLVM_DEBUG(dbgs() << "Dependence before interchange\n");
508     printDepMatrix(DependencyMatrix);
509 #endif
510 
511     // Get the Outermost loop exit.
512     BasicBlock *LoopNestExit = OuterMostLoop->getExitBlock();
513     if (!LoopNestExit) {
514       LLVM_DEBUG(dbgs() << "OuterMostLoop needs an unique exit block");
515       return false;
516     }
517 
518     unsigned SelecLoopId = selectLoopForInterchange(LoopList);
519     // Move the selected loop outwards to the best possible position.
520     Loop *LoopToBeInterchanged = LoopList[SelecLoopId];
521     for (unsigned i = SelecLoopId; i > 0; i--) {
522       bool Interchanged = processLoop(LoopToBeInterchanged, LoopList[i - 1], i,
523                                       i - 1, DependencyMatrix);
524       if (!Interchanged)
525         return Changed;
526       // Update the DependencyMatrix
527       interChangeDependencies(DependencyMatrix, i, i - 1);
528 #ifdef DUMP_DEP_MATRICIES
529       LLVM_DEBUG(dbgs() << "Dependence after interchange\n");
530       printDepMatrix(DependencyMatrix);
531 #endif
532       Changed |= Interchanged;
533     }
534     return Changed;
535   }
536 
processLoop__anonc2e4e0200211::LoopInterchange537   bool processLoop(Loop *InnerLoop, Loop *OuterLoop, unsigned InnerLoopId,
538                    unsigned OuterLoopId,
539                    std::vector<std::vector<char>> &DependencyMatrix) {
540     LLVM_DEBUG(dbgs() << "Processing InnerLoopId = " << InnerLoopId
541                       << " and OuterLoopId = " << OuterLoopId << "\n");
542     LoopInterchangeLegality LIL(OuterLoop, InnerLoop, SE, ORE);
543     if (!LIL.canInterchangeLoops(InnerLoopId, OuterLoopId, DependencyMatrix)) {
544       LLVM_DEBUG(dbgs() << "Not interchanging loops. Cannot prove legality.\n");
545       return false;
546     }
547     LLVM_DEBUG(dbgs() << "Loops are legal to interchange\n");
548     LoopInterchangeProfitability LIP(OuterLoop, InnerLoop, SE, ORE);
549     if (!LIP.isProfitable(InnerLoopId, OuterLoopId, DependencyMatrix)) {
550       LLVM_DEBUG(dbgs() << "Interchanging loops not profitable.\n");
551       return false;
552     }
553 
554     ORE->emit([&]() {
555       return OptimizationRemark(DEBUG_TYPE, "Interchanged",
556                                 InnerLoop->getStartLoc(),
557                                 InnerLoop->getHeader())
558              << "Loop interchanged with enclosing loop.";
559     });
560 
561     LoopInterchangeTransform LIT(OuterLoop, InnerLoop, SE, LI, DT, LIL);
562     LIT.transform();
563     LLVM_DEBUG(dbgs() << "Loops interchanged.\n");
564     LoopsInterchanged++;
565 
566     assert(InnerLoop->isLCSSAForm(*DT) &&
567            "Inner loop not left in LCSSA form after loop interchange!");
568     assert(OuterLoop->isLCSSAForm(*DT) &&
569            "Outer loop not left in LCSSA form after loop interchange!");
570 
571     return true;
572   }
573 };
574 
575 } // end anonymous namespace
576 
containsUnsafeInstructions(BasicBlock * BB)577 bool LoopInterchangeLegality::containsUnsafeInstructions(BasicBlock *BB) {
578   return any_of(*BB, [](const Instruction &I) {
579     return I.mayHaveSideEffects() || I.mayReadFromMemory();
580   });
581 }
582 
tightlyNested(Loop * OuterLoop,Loop * InnerLoop)583 bool LoopInterchangeLegality::tightlyNested(Loop *OuterLoop, Loop *InnerLoop) {
584   BasicBlock *OuterLoopHeader = OuterLoop->getHeader();
585   BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader();
586   BasicBlock *OuterLoopLatch = OuterLoop->getLoopLatch();
587 
588   LLVM_DEBUG(dbgs() << "Checking if loops are tightly nested\n");
589 
590   // A perfectly nested loop will not have any branch in between the outer and
591   // inner block i.e. outer header will branch to either inner preheader and
592   // outerloop latch.
593   BranchInst *OuterLoopHeaderBI =
594       dyn_cast<BranchInst>(OuterLoopHeader->getTerminator());
595   if (!OuterLoopHeaderBI)
596     return false;
597 
598   for (BasicBlock *Succ : successors(OuterLoopHeaderBI))
599     if (Succ != InnerLoopPreHeader && Succ != InnerLoop->getHeader() &&
600         Succ != OuterLoopLatch)
601       return false;
602 
603   LLVM_DEBUG(dbgs() << "Checking instructions in Loop header and Loop latch\n");
604   // We do not have any basic block in between now make sure the outer header
605   // and outer loop latch doesn't contain any unsafe instructions.
606   if (containsUnsafeInstructions(OuterLoopHeader) ||
607       containsUnsafeInstructions(OuterLoopLatch))
608     return false;
609 
610   // Also make sure the inner loop preheader does not contain any unsafe
611   // instructions. Note that all instructions in the preheader will be moved to
612   // the outer loop header when interchanging.
613   if (InnerLoopPreHeader != OuterLoopHeader &&
614       containsUnsafeInstructions(InnerLoopPreHeader))
615     return false;
616 
617   BasicBlock *InnerLoopExit = InnerLoop->getExitBlock();
618   // Ensure the inner loop exit block flows to the outer loop latch possibly
619   // through empty blocks.
620   const BasicBlock &SuccInner =
621       LoopNest::skipEmptyBlockUntil(InnerLoopExit, OuterLoopLatch);
622   if (&SuccInner != OuterLoopLatch) {
623     LLVM_DEBUG(dbgs() << "Inner loop exit block " << *InnerLoopExit
624                       << " does not lead to the outer loop latch.\n";);
625     return false;
626   }
627   // The inner loop exit block does flow to the outer loop latch and not some
628   // other BBs, now make sure it contains safe instructions, since it will be
629   // moved into the (new) inner loop after interchange.
630   if (containsUnsafeInstructions(InnerLoopExit))
631     return false;
632 
633   LLVM_DEBUG(dbgs() << "Loops are perfectly nested\n");
634   // We have a perfect loop nest.
635   return true;
636 }
637 
isLoopStructureUnderstood(PHINode * InnerInduction)638 bool LoopInterchangeLegality::isLoopStructureUnderstood(
639     PHINode *InnerInduction) {
640   unsigned Num = InnerInduction->getNumOperands();
641   BasicBlock *InnerLoopPreheader = InnerLoop->getLoopPreheader();
642   for (unsigned i = 0; i < Num; ++i) {
643     Value *Val = InnerInduction->getOperand(i);
644     if (isa<Constant>(Val))
645       continue;
646     Instruction *I = dyn_cast<Instruction>(Val);
647     if (!I)
648       return false;
649     // TODO: Handle triangular loops.
650     // e.g. for(int i=0;i<N;i++)
651     //        for(int j=i;j<N;j++)
652     unsigned IncomBlockIndx = PHINode::getIncomingValueNumForOperand(i);
653     if (InnerInduction->getIncomingBlock(IncomBlockIndx) ==
654             InnerLoopPreheader &&
655         !OuterLoop->isLoopInvariant(I)) {
656       return false;
657     }
658   }
659 
660   // TODO: Handle triangular loops of another form.
661   // e.g. for(int i=0;i<N;i++)
662   //        for(int j=0;j<i;j++)
663   // or,
664   //      for(int i=0;i<N;i++)
665   //        for(int j=0;j*i<N;j++)
666   BasicBlock *InnerLoopLatch = InnerLoop->getLoopLatch();
667   BranchInst *InnerLoopLatchBI =
668       dyn_cast<BranchInst>(InnerLoopLatch->getTerminator());
669   if (!InnerLoopLatchBI->isConditional())
670     return false;
671   if (CmpInst *InnerLoopCmp =
672           dyn_cast<CmpInst>(InnerLoopLatchBI->getCondition())) {
673     Value *Op0 = InnerLoopCmp->getOperand(0);
674     Value *Op1 = InnerLoopCmp->getOperand(1);
675 
676     // LHS and RHS of the inner loop exit condition, e.g.,
677     // in "for(int j=0;j<i;j++)", LHS is j and RHS is i.
678     Value *Left = nullptr;
679     Value *Right = nullptr;
680 
681     // Check if V only involves inner loop induction variable.
682     // Return true if V is InnerInduction, or a cast from
683     // InnerInduction, or a binary operator that involves
684     // InnerInduction and a constant.
685     std::function<bool(Value *)> IsPathToIndVar;
686     IsPathToIndVar = [&InnerInduction, &IsPathToIndVar](Value *V) -> bool {
687       if (V == InnerInduction)
688         return true;
689       if (isa<Constant>(V))
690         return true;
691       Instruction *I = dyn_cast<Instruction>(V);
692       if (!I)
693         return false;
694       if (isa<CastInst>(I))
695         return IsPathToIndVar(I->getOperand(0));
696       if (isa<BinaryOperator>(I))
697         return IsPathToIndVar(I->getOperand(0)) &&
698                IsPathToIndVar(I->getOperand(1));
699       return false;
700     };
701 
702     if (IsPathToIndVar(Op0) && !isa<Constant>(Op0)) {
703       Left = Op0;
704       Right = Op1;
705     } else if (IsPathToIndVar(Op1) && !isa<Constant>(Op1)) {
706       Left = Op1;
707       Right = Op0;
708     }
709 
710     if (Left == nullptr)
711       return false;
712 
713     const SCEV *S = SE->getSCEV(Right);
714     if (!SE->isLoopInvariant(S, OuterLoop))
715       return false;
716   }
717 
718   return true;
719 }
720 
721 // If SV is a LCSSA PHI node with a single incoming value, return the incoming
722 // value.
followLCSSA(Value * SV)723 static Value *followLCSSA(Value *SV) {
724   PHINode *PHI = dyn_cast<PHINode>(SV);
725   if (!PHI)
726     return SV;
727 
728   if (PHI->getNumIncomingValues() != 1)
729     return SV;
730   return followLCSSA(PHI->getIncomingValue(0));
731 }
732 
733 // Check V's users to see if it is involved in a reduction in L.
findInnerReductionPhi(Loop * L,Value * V)734 static PHINode *findInnerReductionPhi(Loop *L, Value *V) {
735   // Reduction variables cannot be constants.
736   if (isa<Constant>(V))
737     return nullptr;
738 
739   for (Value *User : V->users()) {
740     if (PHINode *PHI = dyn_cast<PHINode>(User)) {
741       if (PHI->getNumIncomingValues() == 1)
742         continue;
743       RecurrenceDescriptor RD;
744       if (RecurrenceDescriptor::isReductionPHI(PHI, L, RD))
745         return PHI;
746       return nullptr;
747     }
748   }
749 
750   return nullptr;
751 }
752 
findInductionAndReductions(Loop * L,SmallVector<PHINode *,8> & Inductions,Loop * InnerLoop)753 bool LoopInterchangeLegality::findInductionAndReductions(
754     Loop *L, SmallVector<PHINode *, 8> &Inductions, Loop *InnerLoop) {
755   if (!L->getLoopLatch() || !L->getLoopPredecessor())
756     return false;
757   for (PHINode &PHI : L->getHeader()->phis()) {
758     RecurrenceDescriptor RD;
759     InductionDescriptor ID;
760     if (InductionDescriptor::isInductionPHI(&PHI, L, SE, ID))
761       Inductions.push_back(&PHI);
762     else {
763       // PHIs in inner loops need to be part of a reduction in the outer loop,
764       // discovered when checking the PHIs of the outer loop earlier.
765       if (!InnerLoop) {
766         if (!OuterInnerReductions.count(&PHI)) {
767           LLVM_DEBUG(dbgs() << "Inner loop PHI is not part of reductions "
768                                "across the outer loop.\n");
769           return false;
770         }
771       } else {
772         assert(PHI.getNumIncomingValues() == 2 &&
773                "Phis in loop header should have exactly 2 incoming values");
774         // Check if we have a PHI node in the outer loop that has a reduction
775         // result from the inner loop as an incoming value.
776         Value *V = followLCSSA(PHI.getIncomingValueForBlock(L->getLoopLatch()));
777         PHINode *InnerRedPhi = findInnerReductionPhi(InnerLoop, V);
778         if (!InnerRedPhi ||
779             !llvm::is_contained(InnerRedPhi->incoming_values(), &PHI)) {
780           LLVM_DEBUG(
781               dbgs()
782               << "Failed to recognize PHI as an induction or reduction.\n");
783           return false;
784         }
785         OuterInnerReductions.insert(&PHI);
786         OuterInnerReductions.insert(InnerRedPhi);
787       }
788     }
789   }
790   return true;
791 }
792 
793 // This function indicates the current limitations in the transform as a result
794 // of which we do not proceed.
currentLimitations()795 bool LoopInterchangeLegality::currentLimitations() {
796   BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader();
797   BasicBlock *InnerLoopLatch = InnerLoop->getLoopLatch();
798 
799   // transform currently expects the loop latches to also be the exiting
800   // blocks.
801   if (InnerLoop->getExitingBlock() != InnerLoopLatch ||
802       OuterLoop->getExitingBlock() != OuterLoop->getLoopLatch() ||
803       !isa<BranchInst>(InnerLoopLatch->getTerminator()) ||
804       !isa<BranchInst>(OuterLoop->getLoopLatch()->getTerminator())) {
805     LLVM_DEBUG(
806         dbgs() << "Loops where the latch is not the exiting block are not"
807                << " supported currently.\n");
808     ORE->emit([&]() {
809       return OptimizationRemarkMissed(DEBUG_TYPE, "ExitingNotLatch",
810                                       OuterLoop->getStartLoc(),
811                                       OuterLoop->getHeader())
812              << "Loops where the latch is not the exiting block cannot be"
813                 " interchange currently.";
814     });
815     return true;
816   }
817 
818   PHINode *InnerInductionVar;
819   SmallVector<PHINode *, 8> Inductions;
820   if (!findInductionAndReductions(OuterLoop, Inductions, InnerLoop)) {
821     LLVM_DEBUG(
822         dbgs() << "Only outer loops with induction or reduction PHI nodes "
823                << "are supported currently.\n");
824     ORE->emit([&]() {
825       return OptimizationRemarkMissed(DEBUG_TYPE, "UnsupportedPHIOuter",
826                                       OuterLoop->getStartLoc(),
827                                       OuterLoop->getHeader())
828              << "Only outer loops with induction or reduction PHI nodes can be"
829                 " interchanged currently.";
830     });
831     return true;
832   }
833 
834   // TODO: Currently we handle only loops with 1 induction variable.
835   if (Inductions.size() != 1) {
836     LLVM_DEBUG(dbgs() << "Loops with more than 1 induction variables are not "
837                       << "supported currently.\n");
838     ORE->emit([&]() {
839       return OptimizationRemarkMissed(DEBUG_TYPE, "MultiIndutionOuter",
840                                       OuterLoop->getStartLoc(),
841                                       OuterLoop->getHeader())
842              << "Only outer loops with 1 induction variable can be "
843                 "interchanged currently.";
844     });
845     return true;
846   }
847 
848   Inductions.clear();
849   if (!findInductionAndReductions(InnerLoop, Inductions, nullptr)) {
850     LLVM_DEBUG(
851         dbgs() << "Only inner loops with induction or reduction PHI nodes "
852                << "are supported currently.\n");
853     ORE->emit([&]() {
854       return OptimizationRemarkMissed(DEBUG_TYPE, "UnsupportedPHIInner",
855                                       InnerLoop->getStartLoc(),
856                                       InnerLoop->getHeader())
857              << "Only inner loops with induction or reduction PHI nodes can be"
858                 " interchange currently.";
859     });
860     return true;
861   }
862 
863   // TODO: Currently we handle only loops with 1 induction variable.
864   if (Inductions.size() != 1) {
865     LLVM_DEBUG(
866         dbgs() << "We currently only support loops with 1 induction variable."
867                << "Failed to interchange due to current limitation\n");
868     ORE->emit([&]() {
869       return OptimizationRemarkMissed(DEBUG_TYPE, "MultiInductionInner",
870                                       InnerLoop->getStartLoc(),
871                                       InnerLoop->getHeader())
872              << "Only inner loops with 1 induction variable can be "
873                 "interchanged currently.";
874     });
875     return true;
876   }
877   InnerInductionVar = Inductions.pop_back_val();
878 
879   // TODO: Triangular loops are not handled for now.
880   if (!isLoopStructureUnderstood(InnerInductionVar)) {
881     LLVM_DEBUG(dbgs() << "Loop structure not understood by pass\n");
882     ORE->emit([&]() {
883       return OptimizationRemarkMissed(DEBUG_TYPE, "UnsupportedStructureInner",
884                                       InnerLoop->getStartLoc(),
885                                       InnerLoop->getHeader())
886              << "Inner loop structure not understood currently.";
887     });
888     return true;
889   }
890 
891   // TODO: Current limitation: Since we split the inner loop latch at the point
892   // were induction variable is incremented (induction.next); We cannot have
893   // more than 1 user of induction.next since it would result in broken code
894   // after split.
895   // e.g.
896   // for(i=0;i<N;i++) {
897   //    for(j = 0;j<M;j++) {
898   //      A[j+1][i+2] = A[j][i]+k;
899   //  }
900   // }
901   Instruction *InnerIndexVarInc = nullptr;
902   if (InnerInductionVar->getIncomingBlock(0) == InnerLoopPreHeader)
903     InnerIndexVarInc =
904         dyn_cast<Instruction>(InnerInductionVar->getIncomingValue(1));
905   else
906     InnerIndexVarInc =
907         dyn_cast<Instruction>(InnerInductionVar->getIncomingValue(0));
908 
909   if (!InnerIndexVarInc) {
910     LLVM_DEBUG(
911         dbgs() << "Did not find an instruction to increment the induction "
912                << "variable.\n");
913     ORE->emit([&]() {
914       return OptimizationRemarkMissed(DEBUG_TYPE, "NoIncrementInInner",
915                                       InnerLoop->getStartLoc(),
916                                       InnerLoop->getHeader())
917              << "The inner loop does not increment the induction variable.";
918     });
919     return true;
920   }
921 
922   // Since we split the inner loop latch on this induction variable. Make sure
923   // we do not have any instruction between the induction variable and branch
924   // instruction.
925 
926   bool FoundInduction = false;
927   for (const Instruction &I :
928        llvm::reverse(InnerLoopLatch->instructionsWithoutDebug())) {
929     if (isa<BranchInst>(I) || isa<CmpInst>(I) || isa<TruncInst>(I) ||
930         isa<ZExtInst>(I))
931       continue;
932 
933     // We found an instruction. If this is not induction variable then it is not
934     // safe to split this loop latch.
935     if (!I.isIdenticalTo(InnerIndexVarInc)) {
936       LLVM_DEBUG(dbgs() << "Found unsupported instructions between induction "
937                         << "variable increment and branch.\n");
938       ORE->emit([&]() {
939         return OptimizationRemarkMissed(
940                    DEBUG_TYPE, "UnsupportedInsBetweenInduction",
941                    InnerLoop->getStartLoc(), InnerLoop->getHeader())
942                << "Found unsupported instruction between induction variable "
943                   "increment and branch.";
944       });
945       return true;
946     }
947 
948     FoundInduction = true;
949     break;
950   }
951   // The loop latch ended and we didn't find the induction variable return as
952   // current limitation.
953   if (!FoundInduction) {
954     LLVM_DEBUG(dbgs() << "Did not find the induction variable.\n");
955     ORE->emit([&]() {
956       return OptimizationRemarkMissed(DEBUG_TYPE, "NoIndutionVariable",
957                                       InnerLoop->getStartLoc(),
958                                       InnerLoop->getHeader())
959              << "Did not find the induction variable.";
960     });
961     return true;
962   }
963   return false;
964 }
965 
966 // We currently only support LCSSA PHI nodes in the inner loop exit, if their
967 // users are either reduction PHIs or PHIs outside the outer loop (which means
968 // the we are only interested in the final value after the loop).
969 static bool
areInnerLoopExitPHIsSupported(Loop * InnerL,Loop * OuterL,SmallPtrSetImpl<PHINode * > & Reductions)970 areInnerLoopExitPHIsSupported(Loop *InnerL, Loop *OuterL,
971                               SmallPtrSetImpl<PHINode *> &Reductions) {
972   BasicBlock *InnerExit = OuterL->getUniqueExitBlock();
973   for (PHINode &PHI : InnerExit->phis()) {
974     // Reduction lcssa phi will have only 1 incoming block that from loop latch.
975     if (PHI.getNumIncomingValues() > 1)
976       return false;
977     if (any_of(PHI.users(), [&Reductions, OuterL](User *U) {
978           PHINode *PN = dyn_cast<PHINode>(U);
979           return !PN ||
980                  (!Reductions.count(PN) && OuterL->contains(PN->getParent()));
981         })) {
982       return false;
983     }
984   }
985   return true;
986 }
987 
988 // We currently support LCSSA PHI nodes in the outer loop exit, if their
989 // incoming values do not come from the outer loop latch or if the
990 // outer loop latch has a single predecessor. In that case, the value will
991 // be available if both the inner and outer loop conditions are true, which
992 // will still be true after interchanging. If we have multiple predecessor,
993 // that may not be the case, e.g. because the outer loop latch may be executed
994 // if the inner loop is not executed.
areOuterLoopExitPHIsSupported(Loop * OuterLoop,Loop * InnerLoop)995 static bool areOuterLoopExitPHIsSupported(Loop *OuterLoop, Loop *InnerLoop) {
996   BasicBlock *LoopNestExit = OuterLoop->getUniqueExitBlock();
997   for (PHINode &PHI : LoopNestExit->phis()) {
998     //  FIXME: We currently are not able to detect floating point reductions
999     //         and have to use floating point PHIs as a proxy to prevent
1000     //         interchanging in the presence of floating point reductions.
1001     if (PHI.getType()->isFloatingPointTy())
1002       return false;
1003     for (unsigned i = 0; i < PHI.getNumIncomingValues(); i++) {
1004      Instruction *IncomingI = dyn_cast<Instruction>(PHI.getIncomingValue(i));
1005      if (!IncomingI || IncomingI->getParent() != OuterLoop->getLoopLatch())
1006        continue;
1007 
1008      // The incoming value is defined in the outer loop latch. Currently we
1009      // only support that in case the outer loop latch has a single predecessor.
1010      // This guarantees that the outer loop latch is executed if and only if
1011      // the inner loop is executed (because tightlyNested() guarantees that the
1012      // outer loop header only branches to the inner loop or the outer loop
1013      // latch).
1014      // FIXME: We could weaken this logic and allow multiple predecessors,
1015      //        if the values are produced outside the loop latch. We would need
1016      //        additional logic to update the PHI nodes in the exit block as
1017      //        well.
1018      if (OuterLoop->getLoopLatch()->getUniquePredecessor() == nullptr)
1019        return false;
1020     }
1021   }
1022   return true;
1023 }
1024 
canInterchangeLoops(unsigned InnerLoopId,unsigned OuterLoopId,CharMatrix & DepMatrix)1025 bool LoopInterchangeLegality::canInterchangeLoops(unsigned InnerLoopId,
1026                                                   unsigned OuterLoopId,
1027                                                   CharMatrix &DepMatrix) {
1028   if (!isLegalToInterChangeLoops(DepMatrix, InnerLoopId, OuterLoopId)) {
1029     LLVM_DEBUG(dbgs() << "Failed interchange InnerLoopId = " << InnerLoopId
1030                       << " and OuterLoopId = " << OuterLoopId
1031                       << " due to dependence\n");
1032     ORE->emit([&]() {
1033       return OptimizationRemarkMissed(DEBUG_TYPE, "Dependence",
1034                                       InnerLoop->getStartLoc(),
1035                                       InnerLoop->getHeader())
1036              << "Cannot interchange loops due to dependences.";
1037     });
1038     return false;
1039   }
1040   // Check if outer and inner loop contain legal instructions only.
1041   for (auto *BB : OuterLoop->blocks())
1042     for (Instruction &I : BB->instructionsWithoutDebug())
1043       if (CallInst *CI = dyn_cast<CallInst>(&I)) {
1044         // readnone functions do not prevent interchanging.
1045         if (CI->doesNotReadMemory())
1046           continue;
1047         LLVM_DEBUG(
1048             dbgs() << "Loops with call instructions cannot be interchanged "
1049                    << "safely.");
1050         ORE->emit([&]() {
1051           return OptimizationRemarkMissed(DEBUG_TYPE, "CallInst",
1052                                           CI->getDebugLoc(),
1053                                           CI->getParent())
1054                  << "Cannot interchange loops due to call instruction.";
1055         });
1056 
1057         return false;
1058       }
1059 
1060   // TODO: The loops could not be interchanged due to current limitations in the
1061   // transform module.
1062   if (currentLimitations()) {
1063     LLVM_DEBUG(dbgs() << "Not legal because of current transform limitation\n");
1064     return false;
1065   }
1066 
1067   // Check if the loops are tightly nested.
1068   if (!tightlyNested(OuterLoop, InnerLoop)) {
1069     LLVM_DEBUG(dbgs() << "Loops not tightly nested\n");
1070     ORE->emit([&]() {
1071       return OptimizationRemarkMissed(DEBUG_TYPE, "NotTightlyNested",
1072                                       InnerLoop->getStartLoc(),
1073                                       InnerLoop->getHeader())
1074              << "Cannot interchange loops because they are not tightly "
1075                 "nested.";
1076     });
1077     return false;
1078   }
1079 
1080   if (!areInnerLoopExitPHIsSupported(OuterLoop, InnerLoop,
1081                                      OuterInnerReductions)) {
1082     LLVM_DEBUG(dbgs() << "Found unsupported PHI nodes in inner loop exit.\n");
1083     ORE->emit([&]() {
1084       return OptimizationRemarkMissed(DEBUG_TYPE, "UnsupportedExitPHI",
1085                                       InnerLoop->getStartLoc(),
1086                                       InnerLoop->getHeader())
1087              << "Found unsupported PHI node in loop exit.";
1088     });
1089     return false;
1090   }
1091 
1092   if (!areOuterLoopExitPHIsSupported(OuterLoop, InnerLoop)) {
1093     LLVM_DEBUG(dbgs() << "Found unsupported PHI nodes in outer loop exit.\n");
1094     ORE->emit([&]() {
1095       return OptimizationRemarkMissed(DEBUG_TYPE, "UnsupportedExitPHI",
1096                                       OuterLoop->getStartLoc(),
1097                                       OuterLoop->getHeader())
1098              << "Found unsupported PHI node in loop exit.";
1099     });
1100     return false;
1101   }
1102 
1103   return true;
1104 }
1105 
getInstrOrderCost()1106 int LoopInterchangeProfitability::getInstrOrderCost() {
1107   unsigned GoodOrder, BadOrder;
1108   BadOrder = GoodOrder = 0;
1109   for (BasicBlock *BB : InnerLoop->blocks()) {
1110     for (Instruction &Ins : *BB) {
1111       if (const GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(&Ins)) {
1112         unsigned NumOp = GEP->getNumOperands();
1113         bool FoundInnerInduction = false;
1114         bool FoundOuterInduction = false;
1115         for (unsigned i = 0; i < NumOp; ++i) {
1116           // Skip operands that are not SCEV-able.
1117           if (!SE->isSCEVable(GEP->getOperand(i)->getType()))
1118             continue;
1119 
1120           const SCEV *OperandVal = SE->getSCEV(GEP->getOperand(i));
1121           const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(OperandVal);
1122           if (!AR)
1123             continue;
1124 
1125           // If we find the inner induction after an outer induction e.g.
1126           // for(int i=0;i<N;i++)
1127           //   for(int j=0;j<N;j++)
1128           //     A[i][j] = A[i-1][j-1]+k;
1129           // then it is a good order.
1130           if (AR->getLoop() == InnerLoop) {
1131             // We found an InnerLoop induction after OuterLoop induction. It is
1132             // a good order.
1133             FoundInnerInduction = true;
1134             if (FoundOuterInduction) {
1135               GoodOrder++;
1136               break;
1137             }
1138           }
1139           // If we find the outer induction after an inner induction e.g.
1140           // for(int i=0;i<N;i++)
1141           //   for(int j=0;j<N;j++)
1142           //     A[j][i] = A[j-1][i-1]+k;
1143           // then it is a bad order.
1144           if (AR->getLoop() == OuterLoop) {
1145             // We found an OuterLoop induction after InnerLoop induction. It is
1146             // a bad order.
1147             FoundOuterInduction = true;
1148             if (FoundInnerInduction) {
1149               BadOrder++;
1150               break;
1151             }
1152           }
1153         }
1154       }
1155     }
1156   }
1157   return GoodOrder - BadOrder;
1158 }
1159 
isProfitableForVectorization(unsigned InnerLoopId,unsigned OuterLoopId,CharMatrix & DepMatrix)1160 static bool isProfitableForVectorization(unsigned InnerLoopId,
1161                                          unsigned OuterLoopId,
1162                                          CharMatrix &DepMatrix) {
1163   // TODO: Improve this heuristic to catch more cases.
1164   // If the inner loop is loop independent or doesn't carry any dependency it is
1165   // profitable to move this to outer position.
1166   for (auto &Row : DepMatrix) {
1167     if (Row[InnerLoopId] != 'S' && Row[InnerLoopId] != 'I')
1168       return false;
1169     // TODO: We need to improve this heuristic.
1170     if (Row[OuterLoopId] != '=')
1171       return false;
1172   }
1173   // If outer loop has dependence and inner loop is loop independent then it is
1174   // profitable to interchange to enable parallelism.
1175   // If there are no dependences, interchanging will not improve anything.
1176   return !DepMatrix.empty();
1177 }
1178 
isProfitable(unsigned InnerLoopId,unsigned OuterLoopId,CharMatrix & DepMatrix)1179 bool LoopInterchangeProfitability::isProfitable(unsigned InnerLoopId,
1180                                                 unsigned OuterLoopId,
1181                                                 CharMatrix &DepMatrix) {
1182   // TODO: Add better profitability checks.
1183   // e.g
1184   // 1) Construct dependency matrix and move the one with no loop carried dep
1185   //    inside to enable vectorization.
1186 
1187   // This is rough cost estimation algorithm. It counts the good and bad order
1188   // of induction variables in the instruction and allows reordering if number
1189   // of bad orders is more than good.
1190   int Cost = getInstrOrderCost();
1191   LLVM_DEBUG(dbgs() << "Cost = " << Cost << "\n");
1192   if (Cost < -LoopInterchangeCostThreshold)
1193     return true;
1194 
1195   // It is not profitable as per current cache profitability model. But check if
1196   // we can move this loop outside to improve parallelism.
1197   if (isProfitableForVectorization(InnerLoopId, OuterLoopId, DepMatrix))
1198     return true;
1199 
1200   ORE->emit([&]() {
1201     return OptimizationRemarkMissed(DEBUG_TYPE, "InterchangeNotProfitable",
1202                                     InnerLoop->getStartLoc(),
1203                                     InnerLoop->getHeader())
1204            << "Interchanging loops is too costly (cost="
1205            << ore::NV("Cost", Cost) << ", threshold="
1206            << ore::NV("Threshold", LoopInterchangeCostThreshold)
1207            << ") and it does not improve parallelism.";
1208   });
1209   return false;
1210 }
1211 
removeChildLoop(Loop * OuterLoop,Loop * InnerLoop)1212 void LoopInterchangeTransform::removeChildLoop(Loop *OuterLoop,
1213                                                Loop *InnerLoop) {
1214   for (Loop *L : *OuterLoop)
1215     if (L == InnerLoop) {
1216       OuterLoop->removeChildLoop(L);
1217       return;
1218     }
1219   llvm_unreachable("Couldn't find loop");
1220 }
1221 
1222 /// Update LoopInfo, after interchanging. NewInner and NewOuter refer to the
1223 /// new inner and outer loop after interchanging: NewInner is the original
1224 /// outer loop and NewOuter is the original inner loop.
1225 ///
1226 /// Before interchanging, we have the following structure
1227 /// Outer preheader
1228 //  Outer header
1229 //    Inner preheader
1230 //    Inner header
1231 //      Inner body
1232 //      Inner latch
1233 //   outer bbs
1234 //   Outer latch
1235 //
1236 // After interchanging:
1237 // Inner preheader
1238 // Inner header
1239 //   Outer preheader
1240 //   Outer header
1241 //     Inner body
1242 //     outer bbs
1243 //     Outer latch
1244 //   Inner latch
restructureLoops(Loop * NewInner,Loop * NewOuter,BasicBlock * OrigInnerPreHeader,BasicBlock * OrigOuterPreHeader)1245 void LoopInterchangeTransform::restructureLoops(
1246     Loop *NewInner, Loop *NewOuter, BasicBlock *OrigInnerPreHeader,
1247     BasicBlock *OrigOuterPreHeader) {
1248   Loop *OuterLoopParent = OuterLoop->getParentLoop();
1249   // The original inner loop preheader moves from the new inner loop to
1250   // the parent loop, if there is one.
1251   NewInner->removeBlockFromLoop(OrigInnerPreHeader);
1252   LI->changeLoopFor(OrigInnerPreHeader, OuterLoopParent);
1253 
1254   // Switch the loop levels.
1255   if (OuterLoopParent) {
1256     // Remove the loop from its parent loop.
1257     removeChildLoop(OuterLoopParent, NewInner);
1258     removeChildLoop(NewInner, NewOuter);
1259     OuterLoopParent->addChildLoop(NewOuter);
1260   } else {
1261     removeChildLoop(NewInner, NewOuter);
1262     LI->changeTopLevelLoop(NewInner, NewOuter);
1263   }
1264   while (!NewOuter->isInnermost())
1265     NewInner->addChildLoop(NewOuter->removeChildLoop(NewOuter->begin()));
1266   NewOuter->addChildLoop(NewInner);
1267 
1268   // BBs from the original inner loop.
1269   SmallVector<BasicBlock *, 8> OrigInnerBBs(NewOuter->blocks());
1270 
1271   // Add BBs from the original outer loop to the original inner loop (excluding
1272   // BBs already in inner loop)
1273   for (BasicBlock *BB : NewInner->blocks())
1274     if (LI->getLoopFor(BB) == NewInner)
1275       NewOuter->addBlockEntry(BB);
1276 
1277   // Now remove inner loop header and latch from the new inner loop and move
1278   // other BBs (the loop body) to the new inner loop.
1279   BasicBlock *OuterHeader = NewOuter->getHeader();
1280   BasicBlock *OuterLatch = NewOuter->getLoopLatch();
1281   for (BasicBlock *BB : OrigInnerBBs) {
1282     // Nothing will change for BBs in child loops.
1283     if (LI->getLoopFor(BB) != NewOuter)
1284       continue;
1285     // Remove the new outer loop header and latch from the new inner loop.
1286     if (BB == OuterHeader || BB == OuterLatch)
1287       NewInner->removeBlockFromLoop(BB);
1288     else
1289       LI->changeLoopFor(BB, NewInner);
1290   }
1291 
1292   // The preheader of the original outer loop becomes part of the new
1293   // outer loop.
1294   NewOuter->addBlockEntry(OrigOuterPreHeader);
1295   LI->changeLoopFor(OrigOuterPreHeader, NewOuter);
1296 
1297   // Tell SE that we move the loops around.
1298   SE->forgetLoop(NewOuter);
1299   SE->forgetLoop(NewInner);
1300 }
1301 
transform()1302 bool LoopInterchangeTransform::transform() {
1303   bool Transformed = false;
1304   Instruction *InnerIndexVar;
1305 
1306   if (InnerLoop->getSubLoops().empty()) {
1307     BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader();
1308     LLVM_DEBUG(dbgs() << "Splitting the inner loop latch\n");
1309     PHINode *InductionPHI = getInductionVariable(InnerLoop, SE);
1310     if (!InductionPHI) {
1311       LLVM_DEBUG(dbgs() << "Failed to find the point to split loop latch \n");
1312       return false;
1313     }
1314 
1315     if (InductionPHI->getIncomingBlock(0) == InnerLoopPreHeader)
1316       InnerIndexVar = dyn_cast<Instruction>(InductionPHI->getIncomingValue(1));
1317     else
1318       InnerIndexVar = dyn_cast<Instruction>(InductionPHI->getIncomingValue(0));
1319 
1320     // Ensure that InductionPHI is the first Phi node.
1321     if (&InductionPHI->getParent()->front() != InductionPHI)
1322       InductionPHI->moveBefore(&InductionPHI->getParent()->front());
1323 
1324     // Create a new latch block for the inner loop. We split at the
1325     // current latch's terminator and then move the condition and all
1326     // operands that are not either loop-invariant or the induction PHI into the
1327     // new latch block.
1328     BasicBlock *NewLatch =
1329         SplitBlock(InnerLoop->getLoopLatch(),
1330                    InnerLoop->getLoopLatch()->getTerminator(), DT, LI);
1331 
1332     SmallSetVector<Instruction *, 4> WorkList;
1333     unsigned i = 0;
1334     auto MoveInstructions = [&i, &WorkList, this, InductionPHI, NewLatch]() {
1335       for (; i < WorkList.size(); i++) {
1336         // Duplicate instruction and move it the new latch. Update uses that
1337         // have been moved.
1338         Instruction *NewI = WorkList[i]->clone();
1339         NewI->insertBefore(NewLatch->getFirstNonPHI());
1340         assert(!NewI->mayHaveSideEffects() &&
1341                "Moving instructions with side-effects may change behavior of "
1342                "the loop nest!");
1343         for (Use &U : llvm::make_early_inc_range(WorkList[i]->uses())) {
1344           Instruction *UserI = cast<Instruction>(U.getUser());
1345           if (!InnerLoop->contains(UserI->getParent()) ||
1346               UserI->getParent() == NewLatch || UserI == InductionPHI)
1347             U.set(NewI);
1348         }
1349         // Add operands of moved instruction to the worklist, except if they are
1350         // outside the inner loop or are the induction PHI.
1351         for (Value *Op : WorkList[i]->operands()) {
1352           Instruction *OpI = dyn_cast<Instruction>(Op);
1353           if (!OpI ||
1354               this->LI->getLoopFor(OpI->getParent()) != this->InnerLoop ||
1355               OpI == InductionPHI)
1356             continue;
1357           WorkList.insert(OpI);
1358         }
1359       }
1360     };
1361 
1362     // FIXME: Should we interchange when we have a constant condition?
1363     Instruction *CondI = dyn_cast<Instruction>(
1364         cast<BranchInst>(InnerLoop->getLoopLatch()->getTerminator())
1365             ->getCondition());
1366     if (CondI)
1367       WorkList.insert(CondI);
1368     MoveInstructions();
1369     WorkList.insert(cast<Instruction>(InnerIndexVar));
1370     MoveInstructions();
1371 
1372     // Splits the inner loops phi nodes out into a separate basic block.
1373     BasicBlock *InnerLoopHeader = InnerLoop->getHeader();
1374     SplitBlock(InnerLoopHeader, InnerLoopHeader->getFirstNonPHI(), DT, LI);
1375     LLVM_DEBUG(dbgs() << "splitting InnerLoopHeader done\n");
1376   }
1377 
1378   // Instructions in the original inner loop preheader may depend on values
1379   // defined in the outer loop header. Move them there, because the original
1380   // inner loop preheader will become the entry into the interchanged loop nest.
1381   // Currently we move all instructions and rely on LICM to move invariant
1382   // instructions outside the loop nest.
1383   BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader();
1384   BasicBlock *OuterLoopHeader = OuterLoop->getHeader();
1385   if (InnerLoopPreHeader != OuterLoopHeader) {
1386     SmallPtrSet<Instruction *, 4> NeedsMoving;
1387     for (Instruction &I :
1388          make_early_inc_range(make_range(InnerLoopPreHeader->begin(),
1389                                          std::prev(InnerLoopPreHeader->end()))))
1390       I.moveBefore(OuterLoopHeader->getTerminator());
1391   }
1392 
1393   Transformed |= adjustLoopLinks();
1394   if (!Transformed) {
1395     LLVM_DEBUG(dbgs() << "adjustLoopLinks failed\n");
1396     return false;
1397   }
1398 
1399   return true;
1400 }
1401 
1402 /// \brief Move all instructions except the terminator from FromBB right before
1403 /// InsertBefore
moveBBContents(BasicBlock * FromBB,Instruction * InsertBefore)1404 static void moveBBContents(BasicBlock *FromBB, Instruction *InsertBefore) {
1405   auto &ToList = InsertBefore->getParent()->getInstList();
1406   auto &FromList = FromBB->getInstList();
1407 
1408   ToList.splice(InsertBefore->getIterator(), FromList, FromList.begin(),
1409                 FromBB->getTerminator()->getIterator());
1410 }
1411 
1412 /// Swap instructions between \p BB1 and \p BB2 but keep terminators intact.
swapBBContents(BasicBlock * BB1,BasicBlock * BB2)1413 static void swapBBContents(BasicBlock *BB1, BasicBlock *BB2) {
1414   // Save all non-terminator instructions of BB1 into TempInstrs and unlink them
1415   // from BB1 afterwards.
1416   auto Iter = map_range(*BB1, [](Instruction &I) { return &I; });
1417   SmallVector<Instruction *, 4> TempInstrs(Iter.begin(), std::prev(Iter.end()));
1418   for (Instruction *I : TempInstrs)
1419     I->removeFromParent();
1420 
1421   // Move instructions from BB2 to BB1.
1422   moveBBContents(BB2, BB1->getTerminator());
1423 
1424   // Move instructions from TempInstrs to BB2.
1425   for (Instruction *I : TempInstrs)
1426     I->insertBefore(BB2->getTerminator());
1427 }
1428 
1429 // Update BI to jump to NewBB instead of OldBB. Records updates to the
1430 // dominator tree in DTUpdates. If \p MustUpdateOnce is true, assert that
1431 // \p OldBB  is exactly once in BI's successor list.
updateSuccessor(BranchInst * BI,BasicBlock * OldBB,BasicBlock * NewBB,std::vector<DominatorTree::UpdateType> & DTUpdates,bool MustUpdateOnce=true)1432 static void updateSuccessor(BranchInst *BI, BasicBlock *OldBB,
1433                             BasicBlock *NewBB,
1434                             std::vector<DominatorTree::UpdateType> &DTUpdates,
1435                             bool MustUpdateOnce = true) {
1436   assert((!MustUpdateOnce ||
1437           llvm::count_if(successors(BI),
1438                          [OldBB](BasicBlock *BB) {
1439                            return BB == OldBB;
1440                          }) == 1) && "BI must jump to OldBB exactly once.");
1441   bool Changed = false;
1442   for (Use &Op : BI->operands())
1443     if (Op == OldBB) {
1444       Op.set(NewBB);
1445       Changed = true;
1446     }
1447 
1448   if (Changed) {
1449     DTUpdates.push_back(
1450         {DominatorTree::UpdateKind::Insert, BI->getParent(), NewBB});
1451     DTUpdates.push_back(
1452         {DominatorTree::UpdateKind::Delete, BI->getParent(), OldBB});
1453   }
1454   assert(Changed && "Expected a successor to be updated");
1455 }
1456 
1457 // Move Lcssa PHIs to the right place.
moveLCSSAPhis(BasicBlock * InnerExit,BasicBlock * InnerHeader,BasicBlock * InnerLatch,BasicBlock * OuterHeader,BasicBlock * OuterLatch,BasicBlock * OuterExit,Loop * InnerLoop,LoopInfo * LI)1458 static void moveLCSSAPhis(BasicBlock *InnerExit, BasicBlock *InnerHeader,
1459                           BasicBlock *InnerLatch, BasicBlock *OuterHeader,
1460                           BasicBlock *OuterLatch, BasicBlock *OuterExit,
1461                           Loop *InnerLoop, LoopInfo *LI) {
1462 
1463   // Deal with LCSSA PHI nodes in the exit block of the inner loop, that are
1464   // defined either in the header or latch. Those blocks will become header and
1465   // latch of the new outer loop, and the only possible users can PHI nodes
1466   // in the exit block of the loop nest or the outer loop header (reduction
1467   // PHIs, in that case, the incoming value must be defined in the inner loop
1468   // header). We can just substitute the user with the incoming value and remove
1469   // the PHI.
1470   for (PHINode &P : make_early_inc_range(InnerExit->phis())) {
1471     assert(P.getNumIncomingValues() == 1 &&
1472            "Only loops with a single exit are supported!");
1473 
1474     // Incoming values are guaranteed be instructions currently.
1475     auto IncI = cast<Instruction>(P.getIncomingValueForBlock(InnerLatch));
1476     // Skip phis with incoming values from the inner loop body, excluding the
1477     // header and latch.
1478     if (IncI->getParent() != InnerLatch && IncI->getParent() != InnerHeader)
1479       continue;
1480 
1481     assert(all_of(P.users(),
1482                   [OuterHeader, OuterExit, IncI, InnerHeader](User *U) {
1483                     return (cast<PHINode>(U)->getParent() == OuterHeader &&
1484                             IncI->getParent() == InnerHeader) ||
1485                            cast<PHINode>(U)->getParent() == OuterExit;
1486                   }) &&
1487            "Can only replace phis iff the uses are in the loop nest exit or "
1488            "the incoming value is defined in the inner header (it will "
1489            "dominate all loop blocks after interchanging)");
1490     P.replaceAllUsesWith(IncI);
1491     P.eraseFromParent();
1492   }
1493 
1494   SmallVector<PHINode *, 8> LcssaInnerExit;
1495   for (PHINode &P : InnerExit->phis())
1496     LcssaInnerExit.push_back(&P);
1497 
1498   SmallVector<PHINode *, 8> LcssaInnerLatch;
1499   for (PHINode &P : InnerLatch->phis())
1500     LcssaInnerLatch.push_back(&P);
1501 
1502   // Lcssa PHIs for values used outside the inner loop are in InnerExit.
1503   // If a PHI node has users outside of InnerExit, it has a use outside the
1504   // interchanged loop and we have to preserve it. We move these to
1505   // InnerLatch, which will become the new exit block for the innermost
1506   // loop after interchanging.
1507   for (PHINode *P : LcssaInnerExit)
1508     P->moveBefore(InnerLatch->getFirstNonPHI());
1509 
1510   // If the inner loop latch contains LCSSA PHIs, those come from a child loop
1511   // and we have to move them to the new inner latch.
1512   for (PHINode *P : LcssaInnerLatch)
1513     P->moveBefore(InnerExit->getFirstNonPHI());
1514 
1515   // Deal with LCSSA PHI nodes in the loop nest exit block. For PHIs that have
1516   // incoming values defined in the outer loop, we have to add a new PHI
1517   // in the inner loop latch, which became the exit block of the outer loop,
1518   // after interchanging.
1519   if (OuterExit) {
1520     for (PHINode &P : OuterExit->phis()) {
1521       if (P.getNumIncomingValues() != 1)
1522         continue;
1523       // Skip Phis with incoming values defined in the inner loop. Those should
1524       // already have been updated.
1525       auto I = dyn_cast<Instruction>(P.getIncomingValue(0));
1526       if (!I || LI->getLoopFor(I->getParent()) == InnerLoop)
1527         continue;
1528 
1529       PHINode *NewPhi = dyn_cast<PHINode>(P.clone());
1530       NewPhi->setIncomingValue(0, P.getIncomingValue(0));
1531       NewPhi->setIncomingBlock(0, OuterLatch);
1532       // We might have incoming edges from other BBs, i.e., the original outer
1533       // header.
1534       for (auto *Pred : predecessors(InnerLatch)) {
1535         if (Pred == OuterLatch)
1536           continue;
1537         NewPhi->addIncoming(P.getIncomingValue(0), Pred);
1538       }
1539       NewPhi->insertBefore(InnerLatch->getFirstNonPHI());
1540       P.setIncomingValue(0, NewPhi);
1541     }
1542   }
1543 
1544   // Now adjust the incoming blocks for the LCSSA PHIs.
1545   // For PHIs moved from Inner's exit block, we need to replace Inner's latch
1546   // with the new latch.
1547   InnerLatch->replacePhiUsesWith(InnerLatch, OuterLatch);
1548 }
1549 
adjustLoopBranches()1550 bool LoopInterchangeTransform::adjustLoopBranches() {
1551   LLVM_DEBUG(dbgs() << "adjustLoopBranches called\n");
1552   std::vector<DominatorTree::UpdateType> DTUpdates;
1553 
1554   BasicBlock *OuterLoopPreHeader = OuterLoop->getLoopPreheader();
1555   BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader();
1556 
1557   assert(OuterLoopPreHeader != OuterLoop->getHeader() &&
1558          InnerLoopPreHeader != InnerLoop->getHeader() && OuterLoopPreHeader &&
1559          InnerLoopPreHeader && "Guaranteed by loop-simplify form");
1560   // Ensure that both preheaders do not contain PHI nodes and have single
1561   // predecessors. This allows us to move them easily. We use
1562   // InsertPreHeaderForLoop to create an 'extra' preheader, if the existing
1563   // preheaders do not satisfy those conditions.
1564   if (isa<PHINode>(OuterLoopPreHeader->begin()) ||
1565       !OuterLoopPreHeader->getUniquePredecessor())
1566     OuterLoopPreHeader =
1567         InsertPreheaderForLoop(OuterLoop, DT, LI, nullptr, true);
1568   if (InnerLoopPreHeader == OuterLoop->getHeader())
1569     InnerLoopPreHeader =
1570         InsertPreheaderForLoop(InnerLoop, DT, LI, nullptr, true);
1571 
1572   // Adjust the loop preheader
1573   BasicBlock *InnerLoopHeader = InnerLoop->getHeader();
1574   BasicBlock *OuterLoopHeader = OuterLoop->getHeader();
1575   BasicBlock *InnerLoopLatch = InnerLoop->getLoopLatch();
1576   BasicBlock *OuterLoopLatch = OuterLoop->getLoopLatch();
1577   BasicBlock *OuterLoopPredecessor = OuterLoopPreHeader->getUniquePredecessor();
1578   BasicBlock *InnerLoopLatchPredecessor =
1579       InnerLoopLatch->getUniquePredecessor();
1580   BasicBlock *InnerLoopLatchSuccessor;
1581   BasicBlock *OuterLoopLatchSuccessor;
1582 
1583   BranchInst *OuterLoopLatchBI =
1584       dyn_cast<BranchInst>(OuterLoopLatch->getTerminator());
1585   BranchInst *InnerLoopLatchBI =
1586       dyn_cast<BranchInst>(InnerLoopLatch->getTerminator());
1587   BranchInst *OuterLoopHeaderBI =
1588       dyn_cast<BranchInst>(OuterLoopHeader->getTerminator());
1589   BranchInst *InnerLoopHeaderBI =
1590       dyn_cast<BranchInst>(InnerLoopHeader->getTerminator());
1591 
1592   if (!OuterLoopPredecessor || !InnerLoopLatchPredecessor ||
1593       !OuterLoopLatchBI || !InnerLoopLatchBI || !OuterLoopHeaderBI ||
1594       !InnerLoopHeaderBI)
1595     return false;
1596 
1597   BranchInst *InnerLoopLatchPredecessorBI =
1598       dyn_cast<BranchInst>(InnerLoopLatchPredecessor->getTerminator());
1599   BranchInst *OuterLoopPredecessorBI =
1600       dyn_cast<BranchInst>(OuterLoopPredecessor->getTerminator());
1601 
1602   if (!OuterLoopPredecessorBI || !InnerLoopLatchPredecessorBI)
1603     return false;
1604   BasicBlock *InnerLoopHeaderSuccessor = InnerLoopHeader->getUniqueSuccessor();
1605   if (!InnerLoopHeaderSuccessor)
1606     return false;
1607 
1608   // Adjust Loop Preheader and headers.
1609   // The branches in the outer loop predecessor and the outer loop header can
1610   // be unconditional branches or conditional branches with duplicates. Consider
1611   // this when updating the successors.
1612   updateSuccessor(OuterLoopPredecessorBI, OuterLoopPreHeader,
1613                   InnerLoopPreHeader, DTUpdates, /*MustUpdateOnce=*/false);
1614   // The outer loop header might or might not branch to the outer latch.
1615   // We are guaranteed to branch to the inner loop preheader.
1616   if (llvm::is_contained(OuterLoopHeaderBI->successors(), OuterLoopLatch)) {
1617     // In this case the outerLoopHeader should branch to the InnerLoopLatch.
1618     updateSuccessor(OuterLoopHeaderBI, OuterLoopLatch, InnerLoopLatch,
1619                     DTUpdates,
1620                     /*MustUpdateOnce=*/false);
1621   }
1622   updateSuccessor(OuterLoopHeaderBI, InnerLoopPreHeader,
1623                   InnerLoopHeaderSuccessor, DTUpdates,
1624                   /*MustUpdateOnce=*/false);
1625 
1626   // Adjust reduction PHI's now that the incoming block has changed.
1627   InnerLoopHeaderSuccessor->replacePhiUsesWith(InnerLoopHeader,
1628                                                OuterLoopHeader);
1629 
1630   updateSuccessor(InnerLoopHeaderBI, InnerLoopHeaderSuccessor,
1631                   OuterLoopPreHeader, DTUpdates);
1632 
1633   // -------------Adjust loop latches-----------
1634   if (InnerLoopLatchBI->getSuccessor(0) == InnerLoopHeader)
1635     InnerLoopLatchSuccessor = InnerLoopLatchBI->getSuccessor(1);
1636   else
1637     InnerLoopLatchSuccessor = InnerLoopLatchBI->getSuccessor(0);
1638 
1639   updateSuccessor(InnerLoopLatchPredecessorBI, InnerLoopLatch,
1640                   InnerLoopLatchSuccessor, DTUpdates);
1641 
1642 
1643   if (OuterLoopLatchBI->getSuccessor(0) == OuterLoopHeader)
1644     OuterLoopLatchSuccessor = OuterLoopLatchBI->getSuccessor(1);
1645   else
1646     OuterLoopLatchSuccessor = OuterLoopLatchBI->getSuccessor(0);
1647 
1648   updateSuccessor(InnerLoopLatchBI, InnerLoopLatchSuccessor,
1649                   OuterLoopLatchSuccessor, DTUpdates);
1650   updateSuccessor(OuterLoopLatchBI, OuterLoopLatchSuccessor, InnerLoopLatch,
1651                   DTUpdates);
1652 
1653   DT->applyUpdates(DTUpdates);
1654   restructureLoops(OuterLoop, InnerLoop, InnerLoopPreHeader,
1655                    OuterLoopPreHeader);
1656 
1657   moveLCSSAPhis(InnerLoopLatchSuccessor, InnerLoopHeader, InnerLoopLatch,
1658                 OuterLoopHeader, OuterLoopLatch, InnerLoop->getExitBlock(),
1659                 InnerLoop, LI);
1660   // For PHIs in the exit block of the outer loop, outer's latch has been
1661   // replaced by Inners'.
1662   OuterLoopLatchSuccessor->replacePhiUsesWith(OuterLoopLatch, InnerLoopLatch);
1663 
1664   // Now update the reduction PHIs in the inner and outer loop headers.
1665   SmallVector<PHINode *, 4> InnerLoopPHIs, OuterLoopPHIs;
1666   for (PHINode &PHI : drop_begin(InnerLoopHeader->phis()))
1667     InnerLoopPHIs.push_back(cast<PHINode>(&PHI));
1668   for (PHINode &PHI : drop_begin(OuterLoopHeader->phis()))
1669     OuterLoopPHIs.push_back(cast<PHINode>(&PHI));
1670 
1671   auto &OuterInnerReductions = LIL.getOuterInnerReductions();
1672   (void)OuterInnerReductions;
1673 
1674   // Now move the remaining reduction PHIs from outer to inner loop header and
1675   // vice versa. The PHI nodes must be part of a reduction across the inner and
1676   // outer loop and all the remains to do is and updating the incoming blocks.
1677   for (PHINode *PHI : OuterLoopPHIs) {
1678     PHI->moveBefore(InnerLoopHeader->getFirstNonPHI());
1679     assert(OuterInnerReductions.count(PHI) && "Expected a reduction PHI node");
1680   }
1681   for (PHINode *PHI : InnerLoopPHIs) {
1682     PHI->moveBefore(OuterLoopHeader->getFirstNonPHI());
1683     assert(OuterInnerReductions.count(PHI) && "Expected a reduction PHI node");
1684   }
1685 
1686   // Update the incoming blocks for moved PHI nodes.
1687   OuterLoopHeader->replacePhiUsesWith(InnerLoopPreHeader, OuterLoopPreHeader);
1688   OuterLoopHeader->replacePhiUsesWith(InnerLoopLatch, OuterLoopLatch);
1689   InnerLoopHeader->replacePhiUsesWith(OuterLoopPreHeader, InnerLoopPreHeader);
1690   InnerLoopHeader->replacePhiUsesWith(OuterLoopLatch, InnerLoopLatch);
1691 
1692   // Values defined in the outer loop header could be used in the inner loop
1693   // latch. In that case, we need to create LCSSA phis for them, because after
1694   // interchanging they will be defined in the new inner loop and used in the
1695   // new outer loop.
1696   IRBuilder<> Builder(OuterLoopHeader->getContext());
1697   SmallVector<Instruction *, 4> MayNeedLCSSAPhis;
1698   for (Instruction &I :
1699        make_range(OuterLoopHeader->begin(), std::prev(OuterLoopHeader->end())))
1700     MayNeedLCSSAPhis.push_back(&I);
1701   formLCSSAForInstructions(MayNeedLCSSAPhis, *DT, *LI, SE, Builder);
1702 
1703   return true;
1704 }
1705 
adjustLoopLinks()1706 bool LoopInterchangeTransform::adjustLoopLinks() {
1707   // Adjust all branches in the inner and outer loop.
1708   bool Changed = adjustLoopBranches();
1709   if (Changed) {
1710     // We have interchanged the preheaders so we need to interchange the data in
1711     // the preheaders as well. This is because the content of the inner
1712     // preheader was previously executed inside the outer loop.
1713     BasicBlock *OuterLoopPreHeader = OuterLoop->getLoopPreheader();
1714     BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader();
1715     swapBBContents(OuterLoopPreHeader, InnerLoopPreHeader);
1716   }
1717   return Changed;
1718 }
1719 
1720 /// Main LoopInterchange Pass.
1721 struct LoopInterchangeLegacyPass : public LoopPass {
1722   static char ID;
1723 
LoopInterchangeLegacyPassLoopInterchangeLegacyPass1724   LoopInterchangeLegacyPass() : LoopPass(ID) {
1725     initializeLoopInterchangeLegacyPassPass(*PassRegistry::getPassRegistry());
1726   }
1727 
getAnalysisUsageLoopInterchangeLegacyPass1728   void getAnalysisUsage(AnalysisUsage &AU) const override {
1729     AU.addRequired<DependenceAnalysisWrapperPass>();
1730     AU.addRequired<OptimizationRemarkEmitterWrapperPass>();
1731 
1732     getLoopAnalysisUsage(AU);
1733   }
1734 
runOnLoopLoopInterchangeLegacyPass1735   bool runOnLoop(Loop *L, LPPassManager &LPM) override {
1736     if (skipLoop(L))
1737       return false;
1738 
1739     auto *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
1740     auto *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1741     auto *DI = &getAnalysis<DependenceAnalysisWrapperPass>().getDI();
1742     auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1743     auto *ORE = &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE();
1744 
1745     return LoopInterchange(SE, LI, DI, DT, ORE).run(L);
1746   }
1747 };
1748 
1749 char LoopInterchangeLegacyPass::ID = 0;
1750 
1751 INITIALIZE_PASS_BEGIN(LoopInterchangeLegacyPass, "loop-interchange",
1752                       "Interchanges loops for cache reuse", false, false)
INITIALIZE_PASS_DEPENDENCY(LoopPass)1753 INITIALIZE_PASS_DEPENDENCY(LoopPass)
1754 INITIALIZE_PASS_DEPENDENCY(DependenceAnalysisWrapperPass)
1755 INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass)
1756 
1757 INITIALIZE_PASS_END(LoopInterchangeLegacyPass, "loop-interchange",
1758                     "Interchanges loops for cache reuse", false, false)
1759 
1760 Pass *llvm::createLoopInterchangePass() {
1761   return new LoopInterchangeLegacyPass();
1762 }
1763 
run(LoopNest & LN,LoopAnalysisManager & AM,LoopStandardAnalysisResults & AR,LPMUpdater & U)1764 PreservedAnalyses LoopInterchangePass::run(LoopNest &LN,
1765                                            LoopAnalysisManager &AM,
1766                                            LoopStandardAnalysisResults &AR,
1767                                            LPMUpdater &U) {
1768   Function &F = *LN.getParent();
1769 
1770   DependenceInfo DI(&F, &AR.AA, &AR.SE, &AR.LI);
1771   OptimizationRemarkEmitter ORE(&F);
1772   if (!LoopInterchange(&AR.SE, &AR.LI, &DI, &AR.DT, &ORE).run(LN))
1773     return PreservedAnalyses::all();
1774   return getLoopPassPreservedAnalyses();
1775 }
1776