xref: /freebsd/contrib/llvm-project/llvm/lib/Transforms/Scalar/LoopInterchange.cpp (revision 0fca6ea1d4eea4c934cfff25ac9ee8ad6fe95583)
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/LoopCacheAnalysis.h"
22 #include "llvm/Analysis/LoopInfo.h"
23 #include "llvm/Analysis/LoopNestAnalysis.h"
24 #include "llvm/Analysis/LoopPass.h"
25 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
26 #include "llvm/Analysis/ScalarEvolution.h"
27 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
28 #include "llvm/IR/BasicBlock.h"
29 #include "llvm/IR/Constants.h"
30 #include "llvm/IR/DiagnosticInfo.h"
31 #include "llvm/IR/Dominators.h"
32 #include "llvm/IR/Function.h"
33 #include "llvm/IR/InstrTypes.h"
34 #include "llvm/IR/Instruction.h"
35 #include "llvm/IR/Instructions.h"
36 #include "llvm/IR/User.h"
37 #include "llvm/IR/Value.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/LoopPassManager.h"
44 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
45 #include "llvm/Transforms/Utils/LoopUtils.h"
46 #include <cassert>
47 #include <utility>
48 #include <vector>
49 
50 using namespace llvm;
51 
52 #define DEBUG_TYPE "loop-interchange"
53 
54 STATISTIC(LoopsInterchanged, "Number of loops interchanged");
55 
56 static cl::opt<int> LoopInterchangeCostThreshold(
57     "loop-interchange-threshold", cl::init(0), cl::Hidden,
58     cl::desc("Interchange if you gain more than this number"));
59 
60 namespace {
61 
62 using LoopVector = SmallVector<Loop *, 8>;
63 
64 // TODO: Check if we can use a sparse matrix here.
65 using CharMatrix = std::vector<std::vector<char>>;
66 
67 } // end anonymous namespace
68 
69 // Maximum number of dependencies that can be handled in the dependency matrix.
70 static const unsigned MaxMemInstrCount = 100;
71 
72 // Maximum loop depth supported.
73 static const unsigned MaxLoopNestDepth = 10;
74 
75 #ifdef DUMP_DEP_MATRICIES
printDepMatrix(CharMatrix & DepMatrix)76 static void printDepMatrix(CharMatrix &DepMatrix) {
77   for (auto &Row : DepMatrix) {
78     for (auto D : Row)
79       LLVM_DEBUG(dbgs() << D << " ");
80     LLVM_DEBUG(dbgs() << "\n");
81   }
82 }
83 #endif
84 
populateDependencyMatrix(CharMatrix & DepMatrix,unsigned Level,Loop * L,DependenceInfo * DI,ScalarEvolution * SE)85 static bool populateDependencyMatrix(CharMatrix &DepMatrix, unsigned Level,
86                                      Loop *L, DependenceInfo *DI,
87                                      ScalarEvolution *SE) {
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       // Ignore Input dependencies.
121       if (isa<LoadInst>(Src) && isa<LoadInst>(Dst))
122         continue;
123       // Track Output, Flow, and Anti dependencies.
124       if (auto D = DI->depends(Src, Dst, true)) {
125         assert(D->isOrdered() && "Expected an output, flow or anti dep.");
126         // If the direction vector is negative, normalize it to
127         // make it non-negative.
128         if (D->normalize(SE))
129           LLVM_DEBUG(dbgs() << "Negative dependence vector normalized.\n");
130         LLVM_DEBUG(StringRef DepType =
131                        D->isFlow() ? "flow" : D->isAnti() ? "anti" : "output";
132                    dbgs() << "Found " << DepType
133                           << " dependency between Src and Dst\n"
134                           << " Src:" << *Src << "\n Dst:" << *Dst << '\n');
135         unsigned Levels = D->getLevels();
136         char Direction;
137         for (unsigned II = 1; II <= Levels; ++II) {
138           if (D->isScalar(II)) {
139             Direction = 'S';
140             Dep.push_back(Direction);
141           } else {
142             unsigned Dir = D->getDirection(II);
143             if (Dir == Dependence::DVEntry::LT ||
144                 Dir == Dependence::DVEntry::LE)
145               Direction = '<';
146             else if (Dir == Dependence::DVEntry::GT ||
147                      Dir == Dependence::DVEntry::GE)
148               Direction = '>';
149             else if (Dir == Dependence::DVEntry::EQ)
150               Direction = '=';
151             else
152               Direction = '*';
153             Dep.push_back(Direction);
154           }
155         }
156         while (Dep.size() != Level) {
157           Dep.push_back('I');
158         }
159 
160         DepMatrix.push_back(Dep);
161         if (DepMatrix.size() > MaxMemInstrCount) {
162           LLVM_DEBUG(dbgs() << "Cannot handle more than " << MaxMemInstrCount
163                             << " dependencies inside loop\n");
164           return false;
165         }
166       }
167     }
168   }
169 
170   return true;
171 }
172 
173 // A loop is moved from index 'from' to an index 'to'. Update the Dependence
174 // matrix by exchanging the two columns.
interChangeDependencies(CharMatrix & DepMatrix,unsigned FromIndx,unsigned ToIndx)175 static void interChangeDependencies(CharMatrix &DepMatrix, unsigned FromIndx,
176                                     unsigned ToIndx) {
177   for (unsigned I = 0, E = DepMatrix.size(); I < E; ++I)
178     std::swap(DepMatrix[I][ToIndx], DepMatrix[I][FromIndx]);
179 }
180 
181 // After interchanging, check if the direction vector is valid.
182 // [Theorem] A permutation of the loops in a perfect nest is legal if and only
183 // if the direction matrix, after the same permutation is applied to its
184 // columns, has no ">" direction as the leftmost non-"=" direction in any row.
isLexicographicallyPositive(std::vector<char> & DV)185 static bool isLexicographicallyPositive(std::vector<char> &DV) {
186   for (unsigned char Direction : DV) {
187     if (Direction == '<')
188       return true;
189     if (Direction == '>' || Direction == '*')
190       return false;
191   }
192   return true;
193 }
194 
195 // Checks if it is legal to interchange 2 loops.
isLegalToInterChangeLoops(CharMatrix & DepMatrix,unsigned InnerLoopId,unsigned OuterLoopId)196 static bool isLegalToInterChangeLoops(CharMatrix &DepMatrix,
197                                       unsigned InnerLoopId,
198                                       unsigned OuterLoopId) {
199   unsigned NumRows = DepMatrix.size();
200   std::vector<char> Cur;
201   // For each row check if it is valid to interchange.
202   for (unsigned Row = 0; Row < NumRows; ++Row) {
203     // Create temporary DepVector check its lexicographical order
204     // before and after swapping OuterLoop vs InnerLoop
205     Cur = DepMatrix[Row];
206     if (!isLexicographicallyPositive(Cur))
207       return false;
208     std::swap(Cur[InnerLoopId], Cur[OuterLoopId]);
209     if (!isLexicographicallyPositive(Cur))
210       return false;
211   }
212   return true;
213 }
214 
populateWorklist(Loop & L,LoopVector & LoopList)215 static void populateWorklist(Loop &L, LoopVector &LoopList) {
216   LLVM_DEBUG(dbgs() << "Calling populateWorklist on Func: "
217                     << L.getHeader()->getParent()->getName() << " Loop: %"
218                     << L.getHeader()->getName() << '\n');
219   assert(LoopList.empty() && "LoopList should initially be empty!");
220   Loop *CurrentLoop = &L;
221   const std::vector<Loop *> *Vec = &CurrentLoop->getSubLoops();
222   while (!Vec->empty()) {
223     // The current loop has multiple subloops in it hence it is not tightly
224     // nested.
225     // Discard all loops above it added into Worklist.
226     if (Vec->size() != 1) {
227       LoopList = {};
228       return;
229     }
230 
231     LoopList.push_back(CurrentLoop);
232     CurrentLoop = Vec->front();
233     Vec = &CurrentLoop->getSubLoops();
234   }
235   LoopList.push_back(CurrentLoop);
236 }
237 
238 namespace {
239 
240 /// LoopInterchangeLegality checks if it is legal to interchange the loop.
241 class LoopInterchangeLegality {
242 public:
LoopInterchangeLegality(Loop * Outer,Loop * Inner,ScalarEvolution * SE,OptimizationRemarkEmitter * ORE)243   LoopInterchangeLegality(Loop *Outer, Loop *Inner, ScalarEvolution *SE,
244                           OptimizationRemarkEmitter *ORE)
245       : OuterLoop(Outer), InnerLoop(Inner), SE(SE), ORE(ORE) {}
246 
247   /// Check if the loops can be interchanged.
248   bool canInterchangeLoops(unsigned InnerLoopId, unsigned OuterLoopId,
249                            CharMatrix &DepMatrix);
250 
251   /// Discover induction PHIs in the header of \p L. Induction
252   /// PHIs are added to \p Inductions.
253   bool findInductions(Loop *L, SmallVectorImpl<PHINode *> &Inductions);
254 
255   /// Check if the loop structure is understood. We do not handle triangular
256   /// loops for now.
257   bool isLoopStructureUnderstood();
258 
259   bool currentLimitations();
260 
getOuterInnerReductions() const261   const SmallPtrSetImpl<PHINode *> &getOuterInnerReductions() const {
262     return OuterInnerReductions;
263   }
264 
getInnerLoopInductions() const265   const SmallVectorImpl<PHINode *> &getInnerLoopInductions() const {
266     return InnerLoopInductions;
267   }
268 
269 private:
270   bool tightlyNested(Loop *Outer, Loop *Inner);
271   bool containsUnsafeInstructions(BasicBlock *BB);
272 
273   /// Discover induction and reduction PHIs in the header of \p L. Induction
274   /// PHIs are added to \p Inductions, reductions are added to
275   /// OuterInnerReductions. When the outer loop is passed, the inner loop needs
276   /// to be passed as \p InnerLoop.
277   bool findInductionAndReductions(Loop *L,
278                                   SmallVector<PHINode *, 8> &Inductions,
279                                   Loop *InnerLoop);
280 
281   Loop *OuterLoop;
282   Loop *InnerLoop;
283 
284   ScalarEvolution *SE;
285 
286   /// Interface to emit optimization remarks.
287   OptimizationRemarkEmitter *ORE;
288 
289   /// Set of reduction PHIs taking part of a reduction across the inner and
290   /// outer loop.
291   SmallPtrSet<PHINode *, 4> OuterInnerReductions;
292 
293   /// Set of inner loop induction PHIs
294   SmallVector<PHINode *, 8> InnerLoopInductions;
295 };
296 
297 /// LoopInterchangeProfitability checks if it is profitable to interchange the
298 /// loop.
299 class LoopInterchangeProfitability {
300 public:
LoopInterchangeProfitability(Loop * Outer,Loop * Inner,ScalarEvolution * SE,OptimizationRemarkEmitter * ORE)301   LoopInterchangeProfitability(Loop *Outer, Loop *Inner, ScalarEvolution *SE,
302                                OptimizationRemarkEmitter *ORE)
303       : OuterLoop(Outer), InnerLoop(Inner), SE(SE), ORE(ORE) {}
304 
305   /// Check if the loop interchange is profitable.
306   bool isProfitable(const Loop *InnerLoop, const Loop *OuterLoop,
307                     unsigned InnerLoopId, unsigned OuterLoopId,
308                     CharMatrix &DepMatrix,
309                     const DenseMap<const Loop *, unsigned> &CostMap,
310                     std::unique_ptr<CacheCost> &CC);
311 
312 private:
313   int getInstrOrderCost();
314   std::optional<bool> isProfitablePerLoopCacheAnalysis(
315       const DenseMap<const Loop *, unsigned> &CostMap,
316       std::unique_ptr<CacheCost> &CC);
317   std::optional<bool> isProfitablePerInstrOrderCost();
318   std::optional<bool> isProfitableForVectorization(unsigned InnerLoopId,
319                                                    unsigned OuterLoopId,
320                                                    CharMatrix &DepMatrix);
321   Loop *OuterLoop;
322   Loop *InnerLoop;
323 
324   /// Scev analysis.
325   ScalarEvolution *SE;
326 
327   /// Interface to emit optimization remarks.
328   OptimizationRemarkEmitter *ORE;
329 };
330 
331 /// LoopInterchangeTransform interchanges the loop.
332 class LoopInterchangeTransform {
333 public:
LoopInterchangeTransform(Loop * Outer,Loop * Inner,ScalarEvolution * SE,LoopInfo * LI,DominatorTree * DT,const LoopInterchangeLegality & LIL)334   LoopInterchangeTransform(Loop *Outer, Loop *Inner, ScalarEvolution *SE,
335                            LoopInfo *LI, DominatorTree *DT,
336                            const LoopInterchangeLegality &LIL)
337       : OuterLoop(Outer), InnerLoop(Inner), SE(SE), LI(LI), DT(DT), LIL(LIL) {}
338 
339   /// Interchange OuterLoop and InnerLoop.
340   bool transform();
341   void restructureLoops(Loop *NewInner, Loop *NewOuter,
342                         BasicBlock *OrigInnerPreHeader,
343                         BasicBlock *OrigOuterPreHeader);
344   void removeChildLoop(Loop *OuterLoop, Loop *InnerLoop);
345 
346 private:
347   bool adjustLoopLinks();
348   bool adjustLoopBranches();
349 
350   Loop *OuterLoop;
351   Loop *InnerLoop;
352 
353   /// Scev analysis.
354   ScalarEvolution *SE;
355 
356   LoopInfo *LI;
357   DominatorTree *DT;
358 
359   const LoopInterchangeLegality &LIL;
360 };
361 
362 struct LoopInterchange {
363   ScalarEvolution *SE = nullptr;
364   LoopInfo *LI = nullptr;
365   DependenceInfo *DI = nullptr;
366   DominatorTree *DT = nullptr;
367   std::unique_ptr<CacheCost> CC = nullptr;
368 
369   /// Interface to emit optimization remarks.
370   OptimizationRemarkEmitter *ORE;
371 
LoopInterchange__anon815ea8750211::LoopInterchange372   LoopInterchange(ScalarEvolution *SE, LoopInfo *LI, DependenceInfo *DI,
373                   DominatorTree *DT, std::unique_ptr<CacheCost> &CC,
374                   OptimizationRemarkEmitter *ORE)
375       : SE(SE), LI(LI), DI(DI), DT(DT), CC(std::move(CC)), ORE(ORE) {}
376 
run__anon815ea8750211::LoopInterchange377   bool run(Loop *L) {
378     if (L->getParentLoop())
379       return false;
380     SmallVector<Loop *, 8> LoopList;
381     populateWorklist(*L, LoopList);
382     return processLoopList(LoopList);
383   }
384 
run__anon815ea8750211::LoopInterchange385   bool run(LoopNest &LN) {
386     SmallVector<Loop *, 8> LoopList(LN.getLoops().begin(), LN.getLoops().end());
387     for (unsigned I = 1; I < LoopList.size(); ++I)
388       if (LoopList[I]->getParentLoop() != LoopList[I - 1])
389         return false;
390     return processLoopList(LoopList);
391   }
392 
isComputableLoopNest__anon815ea8750211::LoopInterchange393   bool isComputableLoopNest(ArrayRef<Loop *> LoopList) {
394     for (Loop *L : LoopList) {
395       const SCEV *ExitCountOuter = SE->getBackedgeTakenCount(L);
396       if (isa<SCEVCouldNotCompute>(ExitCountOuter)) {
397         LLVM_DEBUG(dbgs() << "Couldn't compute backedge count\n");
398         return false;
399       }
400       if (L->getNumBackEdges() != 1) {
401         LLVM_DEBUG(dbgs() << "NumBackEdges is not equal to 1\n");
402         return false;
403       }
404       if (!L->getExitingBlock()) {
405         LLVM_DEBUG(dbgs() << "Loop doesn't have unique exit block\n");
406         return false;
407       }
408     }
409     return true;
410   }
411 
selectLoopForInterchange__anon815ea8750211::LoopInterchange412   unsigned selectLoopForInterchange(ArrayRef<Loop *> LoopList) {
413     // TODO: Add a better heuristic to select the loop to be interchanged based
414     // on the dependence matrix. Currently we select the innermost loop.
415     return LoopList.size() - 1;
416   }
417 
processLoopList__anon815ea8750211::LoopInterchange418   bool processLoopList(SmallVectorImpl<Loop *> &LoopList) {
419     bool Changed = false;
420     unsigned LoopNestDepth = LoopList.size();
421     if (LoopNestDepth < 2) {
422       LLVM_DEBUG(dbgs() << "Loop doesn't contain minimum nesting level.\n");
423       return false;
424     }
425     if (LoopNestDepth > MaxLoopNestDepth) {
426       LLVM_DEBUG(dbgs() << "Cannot handle loops of depth greater than "
427                         << MaxLoopNestDepth << "\n");
428       return false;
429     }
430     if (!isComputableLoopNest(LoopList)) {
431       LLVM_DEBUG(dbgs() << "Not valid loop candidate for interchange\n");
432       return false;
433     }
434 
435     LLVM_DEBUG(dbgs() << "Processing LoopList of size = " << LoopNestDepth
436                       << "\n");
437 
438     CharMatrix DependencyMatrix;
439     Loop *OuterMostLoop = *(LoopList.begin());
440     if (!populateDependencyMatrix(DependencyMatrix, LoopNestDepth,
441                                   OuterMostLoop, DI, SE)) {
442       LLVM_DEBUG(dbgs() << "Populating dependency matrix failed\n");
443       return false;
444     }
445 #ifdef DUMP_DEP_MATRICIES
446     LLVM_DEBUG(dbgs() << "Dependence before interchange\n");
447     printDepMatrix(DependencyMatrix);
448 #endif
449 
450     // Get the Outermost loop exit.
451     BasicBlock *LoopNestExit = OuterMostLoop->getExitBlock();
452     if (!LoopNestExit) {
453       LLVM_DEBUG(dbgs() << "OuterMostLoop needs an unique exit block");
454       return false;
455     }
456 
457     unsigned SelecLoopId = selectLoopForInterchange(LoopList);
458     // Obtain the loop vector returned from loop cache analysis beforehand,
459     // and put each <Loop, index> pair into a map for constant time query
460     // later. Indices in loop vector reprsent the optimal order of the
461     // corresponding loop, e.g., given a loopnest with depth N, index 0
462     // indicates the loop should be placed as the outermost loop and index N
463     // indicates the loop should be placed as the innermost loop.
464     //
465     // For the old pass manager CacheCost would be null.
466     DenseMap<const Loop *, unsigned> CostMap;
467     if (CC != nullptr) {
468       const auto &LoopCosts = CC->getLoopCosts();
469       for (unsigned i = 0; i < LoopCosts.size(); i++) {
470         CostMap[LoopCosts[i].first] = i;
471       }
472     }
473     // We try to achieve the globally optimal memory access for the loopnest,
474     // and do interchange based on a bubble-sort fasion. We start from
475     // the innermost loop, move it outwards to the best possible position
476     // and repeat this process.
477     for (unsigned j = SelecLoopId; j > 0; j--) {
478       bool ChangedPerIter = false;
479       for (unsigned i = SelecLoopId; i > SelecLoopId - j; i--) {
480         bool Interchanged = processLoop(LoopList[i], LoopList[i - 1], i, i - 1,
481                                         DependencyMatrix, CostMap);
482         if (!Interchanged)
483           continue;
484         // Loops interchanged, update LoopList accordingly.
485         std::swap(LoopList[i - 1], LoopList[i]);
486         // Update the DependencyMatrix
487         interChangeDependencies(DependencyMatrix, i, i - 1);
488 #ifdef DUMP_DEP_MATRICIES
489         LLVM_DEBUG(dbgs() << "Dependence after interchange\n");
490         printDepMatrix(DependencyMatrix);
491 #endif
492         ChangedPerIter |= Interchanged;
493         Changed |= Interchanged;
494       }
495       // Early abort if there was no interchange during an entire round of
496       // moving loops outwards.
497       if (!ChangedPerIter)
498         break;
499     }
500     return Changed;
501   }
502 
processLoop__anon815ea8750211::LoopInterchange503   bool processLoop(Loop *InnerLoop, Loop *OuterLoop, unsigned InnerLoopId,
504                    unsigned OuterLoopId,
505                    std::vector<std::vector<char>> &DependencyMatrix,
506                    const DenseMap<const Loop *, unsigned> &CostMap) {
507     LLVM_DEBUG(dbgs() << "Processing InnerLoopId = " << InnerLoopId
508                       << " and OuterLoopId = " << OuterLoopId << "\n");
509     LoopInterchangeLegality LIL(OuterLoop, InnerLoop, SE, ORE);
510     if (!LIL.canInterchangeLoops(InnerLoopId, OuterLoopId, DependencyMatrix)) {
511       LLVM_DEBUG(dbgs() << "Not interchanging loops. Cannot prove legality.\n");
512       return false;
513     }
514     LLVM_DEBUG(dbgs() << "Loops are legal to interchange\n");
515     LoopInterchangeProfitability LIP(OuterLoop, InnerLoop, SE, ORE);
516     if (!LIP.isProfitable(InnerLoop, OuterLoop, InnerLoopId, OuterLoopId,
517                           DependencyMatrix, CostMap, CC)) {
518       LLVM_DEBUG(dbgs() << "Interchanging loops not profitable.\n");
519       return false;
520     }
521 
522     ORE->emit([&]() {
523       return OptimizationRemark(DEBUG_TYPE, "Interchanged",
524                                 InnerLoop->getStartLoc(),
525                                 InnerLoop->getHeader())
526              << "Loop interchanged with enclosing loop.";
527     });
528 
529     LoopInterchangeTransform LIT(OuterLoop, InnerLoop, SE, LI, DT, LIL);
530     LIT.transform();
531     LLVM_DEBUG(dbgs() << "Loops interchanged.\n");
532     LoopsInterchanged++;
533 
534     llvm::formLCSSARecursively(*OuterLoop, *DT, LI, SE);
535     return true;
536   }
537 };
538 
539 } // end anonymous namespace
540 
containsUnsafeInstructions(BasicBlock * BB)541 bool LoopInterchangeLegality::containsUnsafeInstructions(BasicBlock *BB) {
542   return any_of(*BB, [](const Instruction &I) {
543     return I.mayHaveSideEffects() || I.mayReadFromMemory();
544   });
545 }
546 
tightlyNested(Loop * OuterLoop,Loop * InnerLoop)547 bool LoopInterchangeLegality::tightlyNested(Loop *OuterLoop, Loop *InnerLoop) {
548   BasicBlock *OuterLoopHeader = OuterLoop->getHeader();
549   BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader();
550   BasicBlock *OuterLoopLatch = OuterLoop->getLoopLatch();
551 
552   LLVM_DEBUG(dbgs() << "Checking if loops are tightly nested\n");
553 
554   // A perfectly nested loop will not have any branch in between the outer and
555   // inner block i.e. outer header will branch to either inner preheader and
556   // outerloop latch.
557   BranchInst *OuterLoopHeaderBI =
558       dyn_cast<BranchInst>(OuterLoopHeader->getTerminator());
559   if (!OuterLoopHeaderBI)
560     return false;
561 
562   for (BasicBlock *Succ : successors(OuterLoopHeaderBI))
563     if (Succ != InnerLoopPreHeader && Succ != InnerLoop->getHeader() &&
564         Succ != OuterLoopLatch)
565       return false;
566 
567   LLVM_DEBUG(dbgs() << "Checking instructions in Loop header and Loop latch\n");
568   // We do not have any basic block in between now make sure the outer header
569   // and outer loop latch doesn't contain any unsafe instructions.
570   if (containsUnsafeInstructions(OuterLoopHeader) ||
571       containsUnsafeInstructions(OuterLoopLatch))
572     return false;
573 
574   // Also make sure the inner loop preheader does not contain any unsafe
575   // instructions. Note that all instructions in the preheader will be moved to
576   // the outer loop header when interchanging.
577   if (InnerLoopPreHeader != OuterLoopHeader &&
578       containsUnsafeInstructions(InnerLoopPreHeader))
579     return false;
580 
581   BasicBlock *InnerLoopExit = InnerLoop->getExitBlock();
582   // Ensure the inner loop exit block flows to the outer loop latch possibly
583   // through empty blocks.
584   const BasicBlock &SuccInner =
585       LoopNest::skipEmptyBlockUntil(InnerLoopExit, OuterLoopLatch);
586   if (&SuccInner != OuterLoopLatch) {
587     LLVM_DEBUG(dbgs() << "Inner loop exit block " << *InnerLoopExit
588                       << " does not lead to the outer loop latch.\n";);
589     return false;
590   }
591   // The inner loop exit block does flow to the outer loop latch and not some
592   // other BBs, now make sure it contains safe instructions, since it will be
593   // moved into the (new) inner loop after interchange.
594   if (containsUnsafeInstructions(InnerLoopExit))
595     return false;
596 
597   LLVM_DEBUG(dbgs() << "Loops are perfectly nested\n");
598   // We have a perfect loop nest.
599   return true;
600 }
601 
isLoopStructureUnderstood()602 bool LoopInterchangeLegality::isLoopStructureUnderstood() {
603   BasicBlock *InnerLoopPreheader = InnerLoop->getLoopPreheader();
604   for (PHINode *InnerInduction : InnerLoopInductions) {
605     unsigned Num = InnerInduction->getNumOperands();
606     for (unsigned i = 0; i < Num; ++i) {
607       Value *Val = InnerInduction->getOperand(i);
608       if (isa<Constant>(Val))
609         continue;
610       Instruction *I = dyn_cast<Instruction>(Val);
611       if (!I)
612         return false;
613       // TODO: Handle triangular loops.
614       // e.g. for(int i=0;i<N;i++)
615       //        for(int j=i;j<N;j++)
616       unsigned IncomBlockIndx = PHINode::getIncomingValueNumForOperand(i);
617       if (InnerInduction->getIncomingBlock(IncomBlockIndx) ==
618               InnerLoopPreheader &&
619           !OuterLoop->isLoopInvariant(I)) {
620         return false;
621       }
622     }
623   }
624 
625   // TODO: Handle triangular loops of another form.
626   // e.g. for(int i=0;i<N;i++)
627   //        for(int j=0;j<i;j++)
628   // or,
629   //      for(int i=0;i<N;i++)
630   //        for(int j=0;j*i<N;j++)
631   BasicBlock *InnerLoopLatch = InnerLoop->getLoopLatch();
632   BranchInst *InnerLoopLatchBI =
633       dyn_cast<BranchInst>(InnerLoopLatch->getTerminator());
634   if (!InnerLoopLatchBI->isConditional())
635     return false;
636   if (CmpInst *InnerLoopCmp =
637           dyn_cast<CmpInst>(InnerLoopLatchBI->getCondition())) {
638     Value *Op0 = InnerLoopCmp->getOperand(0);
639     Value *Op1 = InnerLoopCmp->getOperand(1);
640 
641     // LHS and RHS of the inner loop exit condition, e.g.,
642     // in "for(int j=0;j<i;j++)", LHS is j and RHS is i.
643     Value *Left = nullptr;
644     Value *Right = nullptr;
645 
646     // Check if V only involves inner loop induction variable.
647     // Return true if V is InnerInduction, or a cast from
648     // InnerInduction, or a binary operator that involves
649     // InnerInduction and a constant.
650     std::function<bool(Value *)> IsPathToInnerIndVar;
651     IsPathToInnerIndVar = [this, &IsPathToInnerIndVar](const Value *V) -> bool {
652       if (llvm::is_contained(InnerLoopInductions, V))
653         return true;
654       if (isa<Constant>(V))
655         return true;
656       const Instruction *I = dyn_cast<Instruction>(V);
657       if (!I)
658         return false;
659       if (isa<CastInst>(I))
660         return IsPathToInnerIndVar(I->getOperand(0));
661       if (isa<BinaryOperator>(I))
662         return IsPathToInnerIndVar(I->getOperand(0)) &&
663                IsPathToInnerIndVar(I->getOperand(1));
664       return false;
665     };
666 
667     // In case of multiple inner loop indvars, it is okay if LHS and RHS
668     // are both inner indvar related variables.
669     if (IsPathToInnerIndVar(Op0) && IsPathToInnerIndVar(Op1))
670       return true;
671 
672     // Otherwise we check if the cmp instruction compares an inner indvar
673     // related variable (Left) with a outer loop invariant (Right).
674     if (IsPathToInnerIndVar(Op0) && !isa<Constant>(Op0)) {
675       Left = Op0;
676       Right = Op1;
677     } else if (IsPathToInnerIndVar(Op1) && !isa<Constant>(Op1)) {
678       Left = Op1;
679       Right = Op0;
680     }
681 
682     if (Left == nullptr)
683       return false;
684 
685     const SCEV *S = SE->getSCEV(Right);
686     if (!SE->isLoopInvariant(S, OuterLoop))
687       return false;
688   }
689 
690   return true;
691 }
692 
693 // If SV is a LCSSA PHI node with a single incoming value, return the incoming
694 // value.
followLCSSA(Value * SV)695 static Value *followLCSSA(Value *SV) {
696   PHINode *PHI = dyn_cast<PHINode>(SV);
697   if (!PHI)
698     return SV;
699 
700   if (PHI->getNumIncomingValues() != 1)
701     return SV;
702   return followLCSSA(PHI->getIncomingValue(0));
703 }
704 
705 // Check V's users to see if it is involved in a reduction in L.
findInnerReductionPhi(Loop * L,Value * V)706 static PHINode *findInnerReductionPhi(Loop *L, Value *V) {
707   // Reduction variables cannot be constants.
708   if (isa<Constant>(V))
709     return nullptr;
710 
711   for (Value *User : V->users()) {
712     if (PHINode *PHI = dyn_cast<PHINode>(User)) {
713       if (PHI->getNumIncomingValues() == 1)
714         continue;
715       RecurrenceDescriptor RD;
716       if (RecurrenceDescriptor::isReductionPHI(PHI, L, RD)) {
717         // Detect floating point reduction only when it can be reordered.
718         if (RD.getExactFPMathInst() != nullptr)
719           return nullptr;
720         return PHI;
721       }
722       return nullptr;
723     }
724   }
725 
726   return nullptr;
727 }
728 
findInductionAndReductions(Loop * L,SmallVector<PHINode *,8> & Inductions,Loop * InnerLoop)729 bool LoopInterchangeLegality::findInductionAndReductions(
730     Loop *L, SmallVector<PHINode *, 8> &Inductions, Loop *InnerLoop) {
731   if (!L->getLoopLatch() || !L->getLoopPredecessor())
732     return false;
733   for (PHINode &PHI : L->getHeader()->phis()) {
734     InductionDescriptor ID;
735     if (InductionDescriptor::isInductionPHI(&PHI, L, SE, ID))
736       Inductions.push_back(&PHI);
737     else {
738       // PHIs in inner loops need to be part of a reduction in the outer loop,
739       // discovered when checking the PHIs of the outer loop earlier.
740       if (!InnerLoop) {
741         if (!OuterInnerReductions.count(&PHI)) {
742           LLVM_DEBUG(dbgs() << "Inner loop PHI is not part of reductions "
743                                "across the outer loop.\n");
744           return false;
745         }
746       } else {
747         assert(PHI.getNumIncomingValues() == 2 &&
748                "Phis in loop header should have exactly 2 incoming values");
749         // Check if we have a PHI node in the outer loop that has a reduction
750         // result from the inner loop as an incoming value.
751         Value *V = followLCSSA(PHI.getIncomingValueForBlock(L->getLoopLatch()));
752         PHINode *InnerRedPhi = findInnerReductionPhi(InnerLoop, V);
753         if (!InnerRedPhi ||
754             !llvm::is_contained(InnerRedPhi->incoming_values(), &PHI)) {
755           LLVM_DEBUG(
756               dbgs()
757               << "Failed to recognize PHI as an induction or reduction.\n");
758           return false;
759         }
760         OuterInnerReductions.insert(&PHI);
761         OuterInnerReductions.insert(InnerRedPhi);
762       }
763     }
764   }
765   return true;
766 }
767 
768 // This function indicates the current limitations in the transform as a result
769 // of which we do not proceed.
currentLimitations()770 bool LoopInterchangeLegality::currentLimitations() {
771   BasicBlock *InnerLoopLatch = InnerLoop->getLoopLatch();
772 
773   // transform currently expects the loop latches to also be the exiting
774   // blocks.
775   if (InnerLoop->getExitingBlock() != InnerLoopLatch ||
776       OuterLoop->getExitingBlock() != OuterLoop->getLoopLatch() ||
777       !isa<BranchInst>(InnerLoopLatch->getTerminator()) ||
778       !isa<BranchInst>(OuterLoop->getLoopLatch()->getTerminator())) {
779     LLVM_DEBUG(
780         dbgs() << "Loops where the latch is not the exiting block are not"
781                << " supported currently.\n");
782     ORE->emit([&]() {
783       return OptimizationRemarkMissed(DEBUG_TYPE, "ExitingNotLatch",
784                                       OuterLoop->getStartLoc(),
785                                       OuterLoop->getHeader())
786              << "Loops where the latch is not the exiting block cannot be"
787                 " interchange currently.";
788     });
789     return true;
790   }
791 
792   SmallVector<PHINode *, 8> Inductions;
793   if (!findInductionAndReductions(OuterLoop, Inductions, InnerLoop)) {
794     LLVM_DEBUG(
795         dbgs() << "Only outer loops with induction or reduction PHI nodes "
796                << "are supported currently.\n");
797     ORE->emit([&]() {
798       return OptimizationRemarkMissed(DEBUG_TYPE, "UnsupportedPHIOuter",
799                                       OuterLoop->getStartLoc(),
800                                       OuterLoop->getHeader())
801              << "Only outer loops with induction or reduction PHI nodes can be"
802                 " interchanged currently.";
803     });
804     return true;
805   }
806 
807   Inductions.clear();
808   // For multi-level loop nests, make sure that all phi nodes for inner loops
809   // at all levels can be recognized as a induction or reduction phi. Bail out
810   // if a phi node at a certain nesting level cannot be properly recognized.
811   Loop *CurLevelLoop = OuterLoop;
812   while (!CurLevelLoop->getSubLoops().empty()) {
813     // We already made sure that the loop nest is tightly nested.
814     CurLevelLoop = CurLevelLoop->getSubLoops().front();
815     if (!findInductionAndReductions(CurLevelLoop, Inductions, nullptr)) {
816       LLVM_DEBUG(
817           dbgs() << "Only inner loops with induction or reduction PHI nodes "
818                 << "are supported currently.\n");
819       ORE->emit([&]() {
820         return OptimizationRemarkMissed(DEBUG_TYPE, "UnsupportedPHIInner",
821                                         CurLevelLoop->getStartLoc(),
822                                         CurLevelLoop->getHeader())
823               << "Only inner loops with induction or reduction PHI nodes can be"
824                   " interchange currently.";
825       });
826       return true;
827     }
828   }
829 
830   // TODO: Triangular loops are not handled for now.
831   if (!isLoopStructureUnderstood()) {
832     LLVM_DEBUG(dbgs() << "Loop structure not understood by pass\n");
833     ORE->emit([&]() {
834       return OptimizationRemarkMissed(DEBUG_TYPE, "UnsupportedStructureInner",
835                                       InnerLoop->getStartLoc(),
836                                       InnerLoop->getHeader())
837              << "Inner loop structure not understood currently.";
838     });
839     return true;
840   }
841 
842   return false;
843 }
844 
findInductions(Loop * L,SmallVectorImpl<PHINode * > & Inductions)845 bool LoopInterchangeLegality::findInductions(
846     Loop *L, SmallVectorImpl<PHINode *> &Inductions) {
847   for (PHINode &PHI : L->getHeader()->phis()) {
848     InductionDescriptor ID;
849     if (InductionDescriptor::isInductionPHI(&PHI, L, SE, ID))
850       Inductions.push_back(&PHI);
851   }
852   return !Inductions.empty();
853 }
854 
855 // We currently only support LCSSA PHI nodes in the inner loop exit, if their
856 // users are either reduction PHIs or PHIs outside the outer loop (which means
857 // the we are only interested in the final value after the loop).
858 static bool
areInnerLoopExitPHIsSupported(Loop * InnerL,Loop * OuterL,SmallPtrSetImpl<PHINode * > & Reductions)859 areInnerLoopExitPHIsSupported(Loop *InnerL, Loop *OuterL,
860                               SmallPtrSetImpl<PHINode *> &Reductions) {
861   BasicBlock *InnerExit = OuterL->getUniqueExitBlock();
862   for (PHINode &PHI : InnerExit->phis()) {
863     // Reduction lcssa phi will have only 1 incoming block that from loop latch.
864     if (PHI.getNumIncomingValues() > 1)
865       return false;
866     if (any_of(PHI.users(), [&Reductions, OuterL](User *U) {
867           PHINode *PN = dyn_cast<PHINode>(U);
868           return !PN ||
869                  (!Reductions.count(PN) && OuterL->contains(PN->getParent()));
870         })) {
871       return false;
872     }
873   }
874   return true;
875 }
876 
877 // We currently support LCSSA PHI nodes in the outer loop exit, if their
878 // incoming values do not come from the outer loop latch or if the
879 // outer loop latch has a single predecessor. In that case, the value will
880 // be available if both the inner and outer loop conditions are true, which
881 // will still be true after interchanging. If we have multiple predecessor,
882 // that may not be the case, e.g. because the outer loop latch may be executed
883 // if the inner loop is not executed.
areOuterLoopExitPHIsSupported(Loop * OuterLoop,Loop * InnerLoop)884 static bool areOuterLoopExitPHIsSupported(Loop *OuterLoop, Loop *InnerLoop) {
885   BasicBlock *LoopNestExit = OuterLoop->getUniqueExitBlock();
886   for (PHINode &PHI : LoopNestExit->phis()) {
887     for (unsigned i = 0; i < PHI.getNumIncomingValues(); i++) {
888       Instruction *IncomingI = dyn_cast<Instruction>(PHI.getIncomingValue(i));
889       if (!IncomingI || IncomingI->getParent() != OuterLoop->getLoopLatch())
890         continue;
891 
892       // The incoming value is defined in the outer loop latch. Currently we
893       // only support that in case the outer loop latch has a single predecessor.
894       // This guarantees that the outer loop latch is executed if and only if
895       // the inner loop is executed (because tightlyNested() guarantees that the
896       // outer loop header only branches to the inner loop or the outer loop
897       // latch).
898       // FIXME: We could weaken this logic and allow multiple predecessors,
899       //        if the values are produced outside the loop latch. We would need
900       //        additional logic to update the PHI nodes in the exit block as
901       //        well.
902       if (OuterLoop->getLoopLatch()->getUniquePredecessor() == nullptr)
903         return false;
904     }
905   }
906   return true;
907 }
908 
909 // In case of multi-level nested loops, it may occur that lcssa phis exist in
910 // the latch of InnerLoop, i.e., when defs of the incoming values are further
911 // inside the loopnest. Sometimes those incoming values are not available
912 // after interchange, since the original inner latch will become the new outer
913 // latch which may have predecessor paths that do not include those incoming
914 // values.
915 // TODO: Handle transformation of lcssa phis in the InnerLoop latch in case of
916 // multi-level loop nests.
areInnerLoopLatchPHIsSupported(Loop * OuterLoop,Loop * InnerLoop)917 static bool areInnerLoopLatchPHIsSupported(Loop *OuterLoop, Loop *InnerLoop) {
918   if (InnerLoop->getSubLoops().empty())
919     return true;
920   // If the original outer latch has only one predecessor, then values defined
921   // further inside the looploop, e.g., in the innermost loop, will be available
922   // at the new outer latch after interchange.
923   if (OuterLoop->getLoopLatch()->getUniquePredecessor() != nullptr)
924     return true;
925 
926   // The outer latch has more than one predecessors, i.e., the inner
927   // exit and the inner header.
928   // PHI nodes in the inner latch are lcssa phis where the incoming values
929   // are defined further inside the loopnest. Check if those phis are used
930   // in the original inner latch. If that is the case then bail out since
931   // those incoming values may not be available at the new outer latch.
932   BasicBlock *InnerLoopLatch = InnerLoop->getLoopLatch();
933   for (PHINode &PHI : InnerLoopLatch->phis()) {
934     for (auto *U : PHI.users()) {
935       Instruction *UI = cast<Instruction>(U);
936       if (InnerLoopLatch == UI->getParent())
937         return false;
938     }
939   }
940   return true;
941 }
942 
canInterchangeLoops(unsigned InnerLoopId,unsigned OuterLoopId,CharMatrix & DepMatrix)943 bool LoopInterchangeLegality::canInterchangeLoops(unsigned InnerLoopId,
944                                                   unsigned OuterLoopId,
945                                                   CharMatrix &DepMatrix) {
946   if (!isLegalToInterChangeLoops(DepMatrix, InnerLoopId, OuterLoopId)) {
947     LLVM_DEBUG(dbgs() << "Failed interchange InnerLoopId = " << InnerLoopId
948                       << " and OuterLoopId = " << OuterLoopId
949                       << " due to dependence\n");
950     ORE->emit([&]() {
951       return OptimizationRemarkMissed(DEBUG_TYPE, "Dependence",
952                                       InnerLoop->getStartLoc(),
953                                       InnerLoop->getHeader())
954              << "Cannot interchange loops due to dependences.";
955     });
956     return false;
957   }
958   // Check if outer and inner loop contain legal instructions only.
959   for (auto *BB : OuterLoop->blocks())
960     for (Instruction &I : BB->instructionsWithoutDebug())
961       if (CallInst *CI = dyn_cast<CallInst>(&I)) {
962         // readnone functions do not prevent interchanging.
963         if (CI->onlyWritesMemory())
964           continue;
965         LLVM_DEBUG(
966             dbgs() << "Loops with call instructions cannot be interchanged "
967                    << "safely.");
968         ORE->emit([&]() {
969           return OptimizationRemarkMissed(DEBUG_TYPE, "CallInst",
970                                           CI->getDebugLoc(),
971                                           CI->getParent())
972                  << "Cannot interchange loops due to call instruction.";
973         });
974 
975         return false;
976       }
977 
978   if (!findInductions(InnerLoop, InnerLoopInductions)) {
979     LLVM_DEBUG(dbgs() << "Could not find inner loop induction variables.\n");
980     return false;
981   }
982 
983   if (!areInnerLoopLatchPHIsSupported(OuterLoop, InnerLoop)) {
984     LLVM_DEBUG(dbgs() << "Found unsupported PHI nodes in inner loop latch.\n");
985     ORE->emit([&]() {
986       return OptimizationRemarkMissed(DEBUG_TYPE, "UnsupportedInnerLatchPHI",
987                                       InnerLoop->getStartLoc(),
988                                       InnerLoop->getHeader())
989              << "Cannot interchange loops because unsupported PHI nodes found "
990                 "in inner loop latch.";
991     });
992     return false;
993   }
994 
995   // TODO: The loops could not be interchanged due to current limitations in the
996   // transform module.
997   if (currentLimitations()) {
998     LLVM_DEBUG(dbgs() << "Not legal because of current transform limitation\n");
999     return false;
1000   }
1001 
1002   // Check if the loops are tightly nested.
1003   if (!tightlyNested(OuterLoop, InnerLoop)) {
1004     LLVM_DEBUG(dbgs() << "Loops not tightly nested\n");
1005     ORE->emit([&]() {
1006       return OptimizationRemarkMissed(DEBUG_TYPE, "NotTightlyNested",
1007                                       InnerLoop->getStartLoc(),
1008                                       InnerLoop->getHeader())
1009              << "Cannot interchange loops because they are not tightly "
1010                 "nested.";
1011     });
1012     return false;
1013   }
1014 
1015   if (!areInnerLoopExitPHIsSupported(OuterLoop, InnerLoop,
1016                                      OuterInnerReductions)) {
1017     LLVM_DEBUG(dbgs() << "Found unsupported PHI nodes in inner loop exit.\n");
1018     ORE->emit([&]() {
1019       return OptimizationRemarkMissed(DEBUG_TYPE, "UnsupportedExitPHI",
1020                                       InnerLoop->getStartLoc(),
1021                                       InnerLoop->getHeader())
1022              << "Found unsupported PHI node in loop exit.";
1023     });
1024     return false;
1025   }
1026 
1027   if (!areOuterLoopExitPHIsSupported(OuterLoop, InnerLoop)) {
1028     LLVM_DEBUG(dbgs() << "Found unsupported PHI nodes in outer loop exit.\n");
1029     ORE->emit([&]() {
1030       return OptimizationRemarkMissed(DEBUG_TYPE, "UnsupportedExitPHI",
1031                                       OuterLoop->getStartLoc(),
1032                                       OuterLoop->getHeader())
1033              << "Found unsupported PHI node in loop exit.";
1034     });
1035     return false;
1036   }
1037 
1038   return true;
1039 }
1040 
getInstrOrderCost()1041 int LoopInterchangeProfitability::getInstrOrderCost() {
1042   unsigned GoodOrder, BadOrder;
1043   BadOrder = GoodOrder = 0;
1044   for (BasicBlock *BB : InnerLoop->blocks()) {
1045     for (Instruction &Ins : *BB) {
1046       if (const GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(&Ins)) {
1047         unsigned NumOp = GEP->getNumOperands();
1048         bool FoundInnerInduction = false;
1049         bool FoundOuterInduction = false;
1050         for (unsigned i = 0; i < NumOp; ++i) {
1051           // Skip operands that are not SCEV-able.
1052           if (!SE->isSCEVable(GEP->getOperand(i)->getType()))
1053             continue;
1054 
1055           const SCEV *OperandVal = SE->getSCEV(GEP->getOperand(i));
1056           const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(OperandVal);
1057           if (!AR)
1058             continue;
1059 
1060           // If we find the inner induction after an outer induction e.g.
1061           // for(int i=0;i<N;i++)
1062           //   for(int j=0;j<N;j++)
1063           //     A[i][j] = A[i-1][j-1]+k;
1064           // then it is a good order.
1065           if (AR->getLoop() == InnerLoop) {
1066             // We found an InnerLoop induction after OuterLoop induction. It is
1067             // a good order.
1068             FoundInnerInduction = true;
1069             if (FoundOuterInduction) {
1070               GoodOrder++;
1071               break;
1072             }
1073           }
1074           // If we find the outer induction after an inner induction e.g.
1075           // for(int i=0;i<N;i++)
1076           //   for(int j=0;j<N;j++)
1077           //     A[j][i] = A[j-1][i-1]+k;
1078           // then it is a bad order.
1079           if (AR->getLoop() == OuterLoop) {
1080             // We found an OuterLoop induction after InnerLoop induction. It is
1081             // a bad order.
1082             FoundOuterInduction = true;
1083             if (FoundInnerInduction) {
1084               BadOrder++;
1085               break;
1086             }
1087           }
1088         }
1089       }
1090     }
1091   }
1092   return GoodOrder - BadOrder;
1093 }
1094 
1095 std::optional<bool>
isProfitablePerLoopCacheAnalysis(const DenseMap<const Loop *,unsigned> & CostMap,std::unique_ptr<CacheCost> & CC)1096 LoopInterchangeProfitability::isProfitablePerLoopCacheAnalysis(
1097     const DenseMap<const Loop *, unsigned> &CostMap,
1098     std::unique_ptr<CacheCost> &CC) {
1099   // This is the new cost model returned from loop cache analysis.
1100   // A smaller index means the loop should be placed an outer loop, and vice
1101   // versa.
1102   if (CostMap.contains(InnerLoop) && CostMap.contains(OuterLoop)) {
1103     unsigned InnerIndex = 0, OuterIndex = 0;
1104     InnerIndex = CostMap.find(InnerLoop)->second;
1105     OuterIndex = CostMap.find(OuterLoop)->second;
1106     LLVM_DEBUG(dbgs() << "InnerIndex = " << InnerIndex
1107                       << ", OuterIndex = " << OuterIndex << "\n");
1108     if (InnerIndex < OuterIndex)
1109       return std::optional<bool>(true);
1110     assert(InnerIndex != OuterIndex && "CostMap should assign unique "
1111                                        "numbers to each loop");
1112     if (CC->getLoopCost(*OuterLoop) == CC->getLoopCost(*InnerLoop))
1113       return std::nullopt;
1114     return std::optional<bool>(false);
1115   }
1116   return std::nullopt;
1117 }
1118 
1119 std::optional<bool>
isProfitablePerInstrOrderCost()1120 LoopInterchangeProfitability::isProfitablePerInstrOrderCost() {
1121   // Legacy cost model: this is rough cost estimation algorithm. It counts the
1122   // good and bad order of induction variables in the instruction and allows
1123   // reordering if number of bad orders is more than good.
1124   int Cost = getInstrOrderCost();
1125   LLVM_DEBUG(dbgs() << "Cost = " << Cost << "\n");
1126   if (Cost < 0 && Cost < LoopInterchangeCostThreshold)
1127     return std::optional<bool>(true);
1128 
1129   return std::nullopt;
1130 }
1131 
isProfitableForVectorization(unsigned InnerLoopId,unsigned OuterLoopId,CharMatrix & DepMatrix)1132 std::optional<bool> LoopInterchangeProfitability::isProfitableForVectorization(
1133     unsigned InnerLoopId, unsigned OuterLoopId, CharMatrix &DepMatrix) {
1134   for (auto &Row : DepMatrix) {
1135     // If the inner loop is loop independent or doesn't carry any dependency
1136     // it is not profitable to move this to outer position, since we are
1137     // likely able to do inner loop vectorization already.
1138     if (Row[InnerLoopId] == 'I' || Row[InnerLoopId] == '=')
1139       return std::optional<bool>(false);
1140 
1141     // If the outer loop is not loop independent it is not profitable to move
1142     // this to inner position, since doing so would not enable inner loop
1143     // parallelism.
1144     if (Row[OuterLoopId] != 'I' && Row[OuterLoopId] != '=')
1145       return std::optional<bool>(false);
1146   }
1147   // If inner loop has dependence and outer loop is loop independent then it
1148   // is/ profitable to interchange to enable inner loop parallelism.
1149   // If there are no dependences, interchanging will not improve anything.
1150   return std::optional<bool>(!DepMatrix.empty());
1151 }
1152 
isProfitable(const Loop * InnerLoop,const Loop * OuterLoop,unsigned InnerLoopId,unsigned OuterLoopId,CharMatrix & DepMatrix,const DenseMap<const Loop *,unsigned> & CostMap,std::unique_ptr<CacheCost> & CC)1153 bool LoopInterchangeProfitability::isProfitable(
1154     const Loop *InnerLoop, const Loop *OuterLoop, unsigned InnerLoopId,
1155     unsigned OuterLoopId, CharMatrix &DepMatrix,
1156     const DenseMap<const Loop *, unsigned> &CostMap,
1157     std::unique_ptr<CacheCost> &CC) {
1158   // isProfitable() is structured to avoid endless loop interchange.
1159   // If loop cache analysis could decide the profitability then,
1160   // profitability check will stop and return the analysis result.
1161   // If cache analysis failed to analyze the loopnest (e.g.,
1162   // due to delinearization issues) then only check whether it is
1163   // profitable for InstrOrderCost. Likewise, if InstrOrderCost failed to
1164   // analysis the profitability then only, isProfitableForVectorization
1165   // will decide.
1166   std::optional<bool> shouldInterchange =
1167       isProfitablePerLoopCacheAnalysis(CostMap, CC);
1168   if (!shouldInterchange.has_value()) {
1169     shouldInterchange = isProfitablePerInstrOrderCost();
1170     if (!shouldInterchange.has_value())
1171       shouldInterchange =
1172           isProfitableForVectorization(InnerLoopId, OuterLoopId, DepMatrix);
1173   }
1174   if (!shouldInterchange.has_value()) {
1175     ORE->emit([&]() {
1176       return OptimizationRemarkMissed(DEBUG_TYPE, "InterchangeNotProfitable",
1177                                       InnerLoop->getStartLoc(),
1178                                       InnerLoop->getHeader())
1179              << "Insufficient information to calculate the cost of loop for "
1180                 "interchange.";
1181     });
1182     return false;
1183   } else if (!shouldInterchange.value()) {
1184     ORE->emit([&]() {
1185       return OptimizationRemarkMissed(DEBUG_TYPE, "InterchangeNotProfitable",
1186                                       InnerLoop->getStartLoc(),
1187                                       InnerLoop->getHeader())
1188              << "Interchanging loops is not considered to improve cache "
1189                 "locality nor vectorization.";
1190     });
1191     return false;
1192   }
1193   return true;
1194 }
1195 
removeChildLoop(Loop * OuterLoop,Loop * InnerLoop)1196 void LoopInterchangeTransform::removeChildLoop(Loop *OuterLoop,
1197                                                Loop *InnerLoop) {
1198   for (Loop *L : *OuterLoop)
1199     if (L == InnerLoop) {
1200       OuterLoop->removeChildLoop(L);
1201       return;
1202     }
1203   llvm_unreachable("Couldn't find loop");
1204 }
1205 
1206 /// Update LoopInfo, after interchanging. NewInner and NewOuter refer to the
1207 /// new inner and outer loop after interchanging: NewInner is the original
1208 /// outer loop and NewOuter is the original inner loop.
1209 ///
1210 /// Before interchanging, we have the following structure
1211 /// Outer preheader
1212 //  Outer header
1213 //    Inner preheader
1214 //    Inner header
1215 //      Inner body
1216 //      Inner latch
1217 //   outer bbs
1218 //   Outer latch
1219 //
1220 // After interchanging:
1221 // Inner preheader
1222 // Inner header
1223 //   Outer preheader
1224 //   Outer header
1225 //     Inner body
1226 //     outer bbs
1227 //     Outer latch
1228 //   Inner latch
restructureLoops(Loop * NewInner,Loop * NewOuter,BasicBlock * OrigInnerPreHeader,BasicBlock * OrigOuterPreHeader)1229 void LoopInterchangeTransform::restructureLoops(
1230     Loop *NewInner, Loop *NewOuter, BasicBlock *OrigInnerPreHeader,
1231     BasicBlock *OrigOuterPreHeader) {
1232   Loop *OuterLoopParent = OuterLoop->getParentLoop();
1233   // The original inner loop preheader moves from the new inner loop to
1234   // the parent loop, if there is one.
1235   NewInner->removeBlockFromLoop(OrigInnerPreHeader);
1236   LI->changeLoopFor(OrigInnerPreHeader, OuterLoopParent);
1237 
1238   // Switch the loop levels.
1239   if (OuterLoopParent) {
1240     // Remove the loop from its parent loop.
1241     removeChildLoop(OuterLoopParent, NewInner);
1242     removeChildLoop(NewInner, NewOuter);
1243     OuterLoopParent->addChildLoop(NewOuter);
1244   } else {
1245     removeChildLoop(NewInner, NewOuter);
1246     LI->changeTopLevelLoop(NewInner, NewOuter);
1247   }
1248   while (!NewOuter->isInnermost())
1249     NewInner->addChildLoop(NewOuter->removeChildLoop(NewOuter->begin()));
1250   NewOuter->addChildLoop(NewInner);
1251 
1252   // BBs from the original inner loop.
1253   SmallVector<BasicBlock *, 8> OrigInnerBBs(NewOuter->blocks());
1254 
1255   // Add BBs from the original outer loop to the original inner loop (excluding
1256   // BBs already in inner loop)
1257   for (BasicBlock *BB : NewInner->blocks())
1258     if (LI->getLoopFor(BB) == NewInner)
1259       NewOuter->addBlockEntry(BB);
1260 
1261   // Now remove inner loop header and latch from the new inner loop and move
1262   // other BBs (the loop body) to the new inner loop.
1263   BasicBlock *OuterHeader = NewOuter->getHeader();
1264   BasicBlock *OuterLatch = NewOuter->getLoopLatch();
1265   for (BasicBlock *BB : OrigInnerBBs) {
1266     // Nothing will change for BBs in child loops.
1267     if (LI->getLoopFor(BB) != NewOuter)
1268       continue;
1269     // Remove the new outer loop header and latch from the new inner loop.
1270     if (BB == OuterHeader || BB == OuterLatch)
1271       NewInner->removeBlockFromLoop(BB);
1272     else
1273       LI->changeLoopFor(BB, NewInner);
1274   }
1275 
1276   // The preheader of the original outer loop becomes part of the new
1277   // outer loop.
1278   NewOuter->addBlockEntry(OrigOuterPreHeader);
1279   LI->changeLoopFor(OrigOuterPreHeader, NewOuter);
1280 
1281   // Tell SE that we move the loops around.
1282   SE->forgetLoop(NewOuter);
1283 }
1284 
transform()1285 bool LoopInterchangeTransform::transform() {
1286   bool Transformed = false;
1287 
1288   if (InnerLoop->getSubLoops().empty()) {
1289     BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader();
1290     LLVM_DEBUG(dbgs() << "Splitting the inner loop latch\n");
1291     auto &InductionPHIs = LIL.getInnerLoopInductions();
1292     if (InductionPHIs.empty()) {
1293       LLVM_DEBUG(dbgs() << "Failed to find the point to split loop latch \n");
1294       return false;
1295     }
1296 
1297     SmallVector<Instruction *, 8> InnerIndexVarList;
1298     for (PHINode *CurInductionPHI : InductionPHIs) {
1299       if (CurInductionPHI->getIncomingBlock(0) == InnerLoopPreHeader)
1300         InnerIndexVarList.push_back(
1301             dyn_cast<Instruction>(CurInductionPHI->getIncomingValue(1)));
1302       else
1303         InnerIndexVarList.push_back(
1304             dyn_cast<Instruction>(CurInductionPHI->getIncomingValue(0)));
1305     }
1306 
1307     // Create a new latch block for the inner loop. We split at the
1308     // current latch's terminator and then move the condition and all
1309     // operands that are not either loop-invariant or the induction PHI into the
1310     // new latch block.
1311     BasicBlock *NewLatch =
1312         SplitBlock(InnerLoop->getLoopLatch(),
1313                    InnerLoop->getLoopLatch()->getTerminator(), DT, LI);
1314 
1315     SmallSetVector<Instruction *, 4> WorkList;
1316     unsigned i = 0;
1317     auto MoveInstructions = [&i, &WorkList, this, &InductionPHIs, NewLatch]() {
1318       for (; i < WorkList.size(); i++) {
1319         // Duplicate instruction and move it the new latch. Update uses that
1320         // have been moved.
1321         Instruction *NewI = WorkList[i]->clone();
1322         NewI->insertBefore(NewLatch->getFirstNonPHI());
1323         assert(!NewI->mayHaveSideEffects() &&
1324                "Moving instructions with side-effects may change behavior of "
1325                "the loop nest!");
1326         for (Use &U : llvm::make_early_inc_range(WorkList[i]->uses())) {
1327           Instruction *UserI = cast<Instruction>(U.getUser());
1328           if (!InnerLoop->contains(UserI->getParent()) ||
1329               UserI->getParent() == NewLatch ||
1330               llvm::is_contained(InductionPHIs, UserI))
1331             U.set(NewI);
1332         }
1333         // Add operands of moved instruction to the worklist, except if they are
1334         // outside the inner loop or are the induction PHI.
1335         for (Value *Op : WorkList[i]->operands()) {
1336           Instruction *OpI = dyn_cast<Instruction>(Op);
1337           if (!OpI ||
1338               this->LI->getLoopFor(OpI->getParent()) != this->InnerLoop ||
1339               llvm::is_contained(InductionPHIs, OpI))
1340             continue;
1341           WorkList.insert(OpI);
1342         }
1343       }
1344     };
1345 
1346     // FIXME: Should we interchange when we have a constant condition?
1347     Instruction *CondI = dyn_cast<Instruction>(
1348         cast<BranchInst>(InnerLoop->getLoopLatch()->getTerminator())
1349             ->getCondition());
1350     if (CondI)
1351       WorkList.insert(CondI);
1352     MoveInstructions();
1353     for (Instruction *InnerIndexVar : InnerIndexVarList)
1354       WorkList.insert(cast<Instruction>(InnerIndexVar));
1355     MoveInstructions();
1356   }
1357 
1358   // Ensure the inner loop phi nodes have a separate basic block.
1359   BasicBlock *InnerLoopHeader = InnerLoop->getHeader();
1360   if (InnerLoopHeader->getFirstNonPHI() != InnerLoopHeader->getTerminator()) {
1361     SplitBlock(InnerLoopHeader, InnerLoopHeader->getFirstNonPHI(), DT, LI);
1362     LLVM_DEBUG(dbgs() << "splitting InnerLoopHeader done\n");
1363   }
1364 
1365   // Instructions in the original inner loop preheader may depend on values
1366   // defined in the outer loop header. Move them there, because the original
1367   // inner loop preheader will become the entry into the interchanged loop nest.
1368   // Currently we move all instructions and rely on LICM to move invariant
1369   // instructions outside the loop nest.
1370   BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader();
1371   BasicBlock *OuterLoopHeader = OuterLoop->getHeader();
1372   if (InnerLoopPreHeader != OuterLoopHeader) {
1373     SmallPtrSet<Instruction *, 4> NeedsMoving;
1374     for (Instruction &I :
1375          make_early_inc_range(make_range(InnerLoopPreHeader->begin(),
1376                                          std::prev(InnerLoopPreHeader->end()))))
1377       I.moveBeforePreserving(OuterLoopHeader->getTerminator());
1378   }
1379 
1380   Transformed |= adjustLoopLinks();
1381   if (!Transformed) {
1382     LLVM_DEBUG(dbgs() << "adjustLoopLinks failed\n");
1383     return false;
1384   }
1385 
1386   return true;
1387 }
1388 
1389 /// \brief Move all instructions except the terminator from FromBB right before
1390 /// InsertBefore
moveBBContents(BasicBlock * FromBB,Instruction * InsertBefore)1391 static void moveBBContents(BasicBlock *FromBB, Instruction *InsertBefore) {
1392   BasicBlock *ToBB = InsertBefore->getParent();
1393 
1394   ToBB->splice(InsertBefore->getIterator(), FromBB, FromBB->begin(),
1395                FromBB->getTerminator()->getIterator());
1396 }
1397 
1398 /// Swap instructions between \p BB1 and \p BB2 but keep terminators intact.
swapBBContents(BasicBlock * BB1,BasicBlock * BB2)1399 static void swapBBContents(BasicBlock *BB1, BasicBlock *BB2) {
1400   // Save all non-terminator instructions of BB1 into TempInstrs and unlink them
1401   // from BB1 afterwards.
1402   auto Iter = map_range(*BB1, [](Instruction &I) { return &I; });
1403   SmallVector<Instruction *, 4> TempInstrs(Iter.begin(), std::prev(Iter.end()));
1404   for (Instruction *I : TempInstrs)
1405     I->removeFromParent();
1406 
1407   // Move instructions from BB2 to BB1.
1408   moveBBContents(BB2, BB1->getTerminator());
1409 
1410   // Move instructions from TempInstrs to BB2.
1411   for (Instruction *I : TempInstrs)
1412     I->insertBefore(BB2->getTerminator());
1413 }
1414 
1415 // Update BI to jump to NewBB instead of OldBB. Records updates to the
1416 // dominator tree in DTUpdates. If \p MustUpdateOnce is true, assert that
1417 // \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)1418 static void updateSuccessor(BranchInst *BI, BasicBlock *OldBB,
1419                             BasicBlock *NewBB,
1420                             std::vector<DominatorTree::UpdateType> &DTUpdates,
1421                             bool MustUpdateOnce = true) {
1422   assert((!MustUpdateOnce ||
1423           llvm::count_if(successors(BI),
1424                          [OldBB](BasicBlock *BB) {
1425                            return BB == OldBB;
1426                          }) == 1) && "BI must jump to OldBB exactly once.");
1427   bool Changed = false;
1428   for (Use &Op : BI->operands())
1429     if (Op == OldBB) {
1430       Op.set(NewBB);
1431       Changed = true;
1432     }
1433 
1434   if (Changed) {
1435     DTUpdates.push_back(
1436         {DominatorTree::UpdateKind::Insert, BI->getParent(), NewBB});
1437     DTUpdates.push_back(
1438         {DominatorTree::UpdateKind::Delete, BI->getParent(), OldBB});
1439   }
1440   assert(Changed && "Expected a successor to be updated");
1441 }
1442 
1443 // Move Lcssa PHIs to the right place.
moveLCSSAPhis(BasicBlock * InnerExit,BasicBlock * InnerHeader,BasicBlock * InnerLatch,BasicBlock * OuterHeader,BasicBlock * OuterLatch,BasicBlock * OuterExit,Loop * InnerLoop,LoopInfo * LI)1444 static void moveLCSSAPhis(BasicBlock *InnerExit, BasicBlock *InnerHeader,
1445                           BasicBlock *InnerLatch, BasicBlock *OuterHeader,
1446                           BasicBlock *OuterLatch, BasicBlock *OuterExit,
1447                           Loop *InnerLoop, LoopInfo *LI) {
1448 
1449   // Deal with LCSSA PHI nodes in the exit block of the inner loop, that are
1450   // defined either in the header or latch. Those blocks will become header and
1451   // latch of the new outer loop, and the only possible users can PHI nodes
1452   // in the exit block of the loop nest or the outer loop header (reduction
1453   // PHIs, in that case, the incoming value must be defined in the inner loop
1454   // header). We can just substitute the user with the incoming value and remove
1455   // the PHI.
1456   for (PHINode &P : make_early_inc_range(InnerExit->phis())) {
1457     assert(P.getNumIncomingValues() == 1 &&
1458            "Only loops with a single exit are supported!");
1459 
1460     // Incoming values are guaranteed be instructions currently.
1461     auto IncI = cast<Instruction>(P.getIncomingValueForBlock(InnerLatch));
1462     // In case of multi-level nested loops, follow LCSSA to find the incoming
1463     // value defined from the innermost loop.
1464     auto IncIInnerMost = cast<Instruction>(followLCSSA(IncI));
1465     // Skip phis with incoming values from the inner loop body, excluding the
1466     // header and latch.
1467     if (IncIInnerMost->getParent() != InnerLatch &&
1468         IncIInnerMost->getParent() != InnerHeader)
1469       continue;
1470 
1471     assert(all_of(P.users(),
1472                   [OuterHeader, OuterExit, IncI, InnerHeader](User *U) {
1473                     return (cast<PHINode>(U)->getParent() == OuterHeader &&
1474                             IncI->getParent() == InnerHeader) ||
1475                            cast<PHINode>(U)->getParent() == OuterExit;
1476                   }) &&
1477            "Can only replace phis iff the uses are in the loop nest exit or "
1478            "the incoming value is defined in the inner header (it will "
1479            "dominate all loop blocks after interchanging)");
1480     P.replaceAllUsesWith(IncI);
1481     P.eraseFromParent();
1482   }
1483 
1484   SmallVector<PHINode *, 8> LcssaInnerExit;
1485   for (PHINode &P : InnerExit->phis())
1486     LcssaInnerExit.push_back(&P);
1487 
1488   SmallVector<PHINode *, 8> LcssaInnerLatch;
1489   for (PHINode &P : InnerLatch->phis())
1490     LcssaInnerLatch.push_back(&P);
1491 
1492   // Lcssa PHIs for values used outside the inner loop are in InnerExit.
1493   // If a PHI node has users outside of InnerExit, it has a use outside the
1494   // interchanged loop and we have to preserve it. We move these to
1495   // InnerLatch, which will become the new exit block for the innermost
1496   // loop after interchanging.
1497   for (PHINode *P : LcssaInnerExit)
1498     P->moveBefore(InnerLatch->getFirstNonPHI());
1499 
1500   // If the inner loop latch contains LCSSA PHIs, those come from a child loop
1501   // and we have to move them to the new inner latch.
1502   for (PHINode *P : LcssaInnerLatch)
1503     P->moveBefore(InnerExit->getFirstNonPHI());
1504 
1505   // Deal with LCSSA PHI nodes in the loop nest exit block. For PHIs that have
1506   // incoming values defined in the outer loop, we have to add a new PHI
1507   // in the inner loop latch, which became the exit block of the outer loop,
1508   // after interchanging.
1509   if (OuterExit) {
1510     for (PHINode &P : OuterExit->phis()) {
1511       if (P.getNumIncomingValues() != 1)
1512         continue;
1513       // Skip Phis with incoming values defined in the inner loop. Those should
1514       // already have been updated.
1515       auto I = dyn_cast<Instruction>(P.getIncomingValue(0));
1516       if (!I || LI->getLoopFor(I->getParent()) == InnerLoop)
1517         continue;
1518 
1519       PHINode *NewPhi = dyn_cast<PHINode>(P.clone());
1520       NewPhi->setIncomingValue(0, P.getIncomingValue(0));
1521       NewPhi->setIncomingBlock(0, OuterLatch);
1522       // We might have incoming edges from other BBs, i.e., the original outer
1523       // header.
1524       for (auto *Pred : predecessors(InnerLatch)) {
1525         if (Pred == OuterLatch)
1526           continue;
1527         NewPhi->addIncoming(P.getIncomingValue(0), Pred);
1528       }
1529       NewPhi->insertBefore(InnerLatch->getFirstNonPHI());
1530       P.setIncomingValue(0, NewPhi);
1531     }
1532   }
1533 
1534   // Now adjust the incoming blocks for the LCSSA PHIs.
1535   // For PHIs moved from Inner's exit block, we need to replace Inner's latch
1536   // with the new latch.
1537   InnerLatch->replacePhiUsesWith(InnerLatch, OuterLatch);
1538 }
1539 
adjustLoopBranches()1540 bool LoopInterchangeTransform::adjustLoopBranches() {
1541   LLVM_DEBUG(dbgs() << "adjustLoopBranches called\n");
1542   std::vector<DominatorTree::UpdateType> DTUpdates;
1543 
1544   BasicBlock *OuterLoopPreHeader = OuterLoop->getLoopPreheader();
1545   BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader();
1546 
1547   assert(OuterLoopPreHeader != OuterLoop->getHeader() &&
1548          InnerLoopPreHeader != InnerLoop->getHeader() && OuterLoopPreHeader &&
1549          InnerLoopPreHeader && "Guaranteed by loop-simplify form");
1550   // Ensure that both preheaders do not contain PHI nodes and have single
1551   // predecessors. This allows us to move them easily. We use
1552   // InsertPreHeaderForLoop to create an 'extra' preheader, if the existing
1553   // preheaders do not satisfy those conditions.
1554   if (isa<PHINode>(OuterLoopPreHeader->begin()) ||
1555       !OuterLoopPreHeader->getUniquePredecessor())
1556     OuterLoopPreHeader =
1557         InsertPreheaderForLoop(OuterLoop, DT, LI, nullptr, true);
1558   if (InnerLoopPreHeader == OuterLoop->getHeader())
1559     InnerLoopPreHeader =
1560         InsertPreheaderForLoop(InnerLoop, DT, LI, nullptr, true);
1561 
1562   // Adjust the loop preheader
1563   BasicBlock *InnerLoopHeader = InnerLoop->getHeader();
1564   BasicBlock *OuterLoopHeader = OuterLoop->getHeader();
1565   BasicBlock *InnerLoopLatch = InnerLoop->getLoopLatch();
1566   BasicBlock *OuterLoopLatch = OuterLoop->getLoopLatch();
1567   BasicBlock *OuterLoopPredecessor = OuterLoopPreHeader->getUniquePredecessor();
1568   BasicBlock *InnerLoopLatchPredecessor =
1569       InnerLoopLatch->getUniquePredecessor();
1570   BasicBlock *InnerLoopLatchSuccessor;
1571   BasicBlock *OuterLoopLatchSuccessor;
1572 
1573   BranchInst *OuterLoopLatchBI =
1574       dyn_cast<BranchInst>(OuterLoopLatch->getTerminator());
1575   BranchInst *InnerLoopLatchBI =
1576       dyn_cast<BranchInst>(InnerLoopLatch->getTerminator());
1577   BranchInst *OuterLoopHeaderBI =
1578       dyn_cast<BranchInst>(OuterLoopHeader->getTerminator());
1579   BranchInst *InnerLoopHeaderBI =
1580       dyn_cast<BranchInst>(InnerLoopHeader->getTerminator());
1581 
1582   if (!OuterLoopPredecessor || !InnerLoopLatchPredecessor ||
1583       !OuterLoopLatchBI || !InnerLoopLatchBI || !OuterLoopHeaderBI ||
1584       !InnerLoopHeaderBI)
1585     return false;
1586 
1587   BranchInst *InnerLoopLatchPredecessorBI =
1588       dyn_cast<BranchInst>(InnerLoopLatchPredecessor->getTerminator());
1589   BranchInst *OuterLoopPredecessorBI =
1590       dyn_cast<BranchInst>(OuterLoopPredecessor->getTerminator());
1591 
1592   if (!OuterLoopPredecessorBI || !InnerLoopLatchPredecessorBI)
1593     return false;
1594   BasicBlock *InnerLoopHeaderSuccessor = InnerLoopHeader->getUniqueSuccessor();
1595   if (!InnerLoopHeaderSuccessor)
1596     return false;
1597 
1598   // Adjust Loop Preheader and headers.
1599   // The branches in the outer loop predecessor and the outer loop header can
1600   // be unconditional branches or conditional branches with duplicates. Consider
1601   // this when updating the successors.
1602   updateSuccessor(OuterLoopPredecessorBI, OuterLoopPreHeader,
1603                   InnerLoopPreHeader, DTUpdates, /*MustUpdateOnce=*/false);
1604   // The outer loop header might or might not branch to the outer latch.
1605   // We are guaranteed to branch to the inner loop preheader.
1606   if (llvm::is_contained(OuterLoopHeaderBI->successors(), OuterLoopLatch)) {
1607     // In this case the outerLoopHeader should branch to the InnerLoopLatch.
1608     updateSuccessor(OuterLoopHeaderBI, OuterLoopLatch, InnerLoopLatch,
1609                     DTUpdates,
1610                     /*MustUpdateOnce=*/false);
1611   }
1612   updateSuccessor(OuterLoopHeaderBI, InnerLoopPreHeader,
1613                   InnerLoopHeaderSuccessor, DTUpdates,
1614                   /*MustUpdateOnce=*/false);
1615 
1616   // Adjust reduction PHI's now that the incoming block has changed.
1617   InnerLoopHeaderSuccessor->replacePhiUsesWith(InnerLoopHeader,
1618                                                OuterLoopHeader);
1619 
1620   updateSuccessor(InnerLoopHeaderBI, InnerLoopHeaderSuccessor,
1621                   OuterLoopPreHeader, DTUpdates);
1622 
1623   // -------------Adjust loop latches-----------
1624   if (InnerLoopLatchBI->getSuccessor(0) == InnerLoopHeader)
1625     InnerLoopLatchSuccessor = InnerLoopLatchBI->getSuccessor(1);
1626   else
1627     InnerLoopLatchSuccessor = InnerLoopLatchBI->getSuccessor(0);
1628 
1629   updateSuccessor(InnerLoopLatchPredecessorBI, InnerLoopLatch,
1630                   InnerLoopLatchSuccessor, DTUpdates);
1631 
1632   if (OuterLoopLatchBI->getSuccessor(0) == OuterLoopHeader)
1633     OuterLoopLatchSuccessor = OuterLoopLatchBI->getSuccessor(1);
1634   else
1635     OuterLoopLatchSuccessor = OuterLoopLatchBI->getSuccessor(0);
1636 
1637   updateSuccessor(InnerLoopLatchBI, InnerLoopLatchSuccessor,
1638                   OuterLoopLatchSuccessor, DTUpdates);
1639   updateSuccessor(OuterLoopLatchBI, OuterLoopLatchSuccessor, InnerLoopLatch,
1640                   DTUpdates);
1641 
1642   DT->applyUpdates(DTUpdates);
1643   restructureLoops(OuterLoop, InnerLoop, InnerLoopPreHeader,
1644                    OuterLoopPreHeader);
1645 
1646   moveLCSSAPhis(InnerLoopLatchSuccessor, InnerLoopHeader, InnerLoopLatch,
1647                 OuterLoopHeader, OuterLoopLatch, InnerLoop->getExitBlock(),
1648                 InnerLoop, LI);
1649   // For PHIs in the exit block of the outer loop, outer's latch has been
1650   // replaced by Inners'.
1651   OuterLoopLatchSuccessor->replacePhiUsesWith(OuterLoopLatch, InnerLoopLatch);
1652 
1653   auto &OuterInnerReductions = LIL.getOuterInnerReductions();
1654   // Now update the reduction PHIs in the inner and outer loop headers.
1655   SmallVector<PHINode *, 4> InnerLoopPHIs, OuterLoopPHIs;
1656   for (PHINode &PHI : InnerLoopHeader->phis())
1657     if (OuterInnerReductions.contains(&PHI))
1658       InnerLoopPHIs.push_back(&PHI);
1659 
1660   for (PHINode &PHI : OuterLoopHeader->phis())
1661     if (OuterInnerReductions.contains(&PHI))
1662       OuterLoopPHIs.push_back(&PHI);
1663 
1664   // Now move the remaining reduction PHIs from outer to inner loop header and
1665   // vice versa. The PHI nodes must be part of a reduction across the inner and
1666   // outer loop and all the remains to do is and updating the incoming blocks.
1667   for (PHINode *PHI : OuterLoopPHIs) {
1668     LLVM_DEBUG(dbgs() << "Outer loop reduction PHIs:\n"; PHI->dump(););
1669     PHI->moveBefore(InnerLoopHeader->getFirstNonPHI());
1670     assert(OuterInnerReductions.count(PHI) && "Expected a reduction PHI node");
1671   }
1672   for (PHINode *PHI : InnerLoopPHIs) {
1673     LLVM_DEBUG(dbgs() << "Inner loop reduction PHIs:\n"; PHI->dump(););
1674     PHI->moveBefore(OuterLoopHeader->getFirstNonPHI());
1675     assert(OuterInnerReductions.count(PHI) && "Expected a reduction PHI node");
1676   }
1677 
1678   // Update the incoming blocks for moved PHI nodes.
1679   OuterLoopHeader->replacePhiUsesWith(InnerLoopPreHeader, OuterLoopPreHeader);
1680   OuterLoopHeader->replacePhiUsesWith(InnerLoopLatch, OuterLoopLatch);
1681   InnerLoopHeader->replacePhiUsesWith(OuterLoopPreHeader, InnerLoopPreHeader);
1682   InnerLoopHeader->replacePhiUsesWith(OuterLoopLatch, InnerLoopLatch);
1683 
1684   // Values defined in the outer loop header could be used in the inner loop
1685   // latch. In that case, we need to create LCSSA phis for them, because after
1686   // interchanging they will be defined in the new inner loop and used in the
1687   // new outer loop.
1688   SmallVector<Instruction *, 4> MayNeedLCSSAPhis;
1689   for (Instruction &I :
1690        make_range(OuterLoopHeader->begin(), std::prev(OuterLoopHeader->end())))
1691     MayNeedLCSSAPhis.push_back(&I);
1692   formLCSSAForInstructions(MayNeedLCSSAPhis, *DT, *LI, SE);
1693 
1694   return true;
1695 }
1696 
adjustLoopLinks()1697 bool LoopInterchangeTransform::adjustLoopLinks() {
1698   // Adjust all branches in the inner and outer loop.
1699   bool Changed = adjustLoopBranches();
1700   if (Changed) {
1701     // We have interchanged the preheaders so we need to interchange the data in
1702     // the preheaders as well. This is because the content of the inner
1703     // preheader was previously executed inside the outer loop.
1704     BasicBlock *OuterLoopPreHeader = OuterLoop->getLoopPreheader();
1705     BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader();
1706     swapBBContents(OuterLoopPreHeader, InnerLoopPreHeader);
1707   }
1708   return Changed;
1709 }
1710 
run(LoopNest & LN,LoopAnalysisManager & AM,LoopStandardAnalysisResults & AR,LPMUpdater & U)1711 PreservedAnalyses LoopInterchangePass::run(LoopNest &LN,
1712                                            LoopAnalysisManager &AM,
1713                                            LoopStandardAnalysisResults &AR,
1714                                            LPMUpdater &U) {
1715   Function &F = *LN.getParent();
1716 
1717   DependenceInfo DI(&F, &AR.AA, &AR.SE, &AR.LI);
1718   std::unique_ptr<CacheCost> CC =
1719       CacheCost::getCacheCost(LN.getOutermostLoop(), AR, DI);
1720   OptimizationRemarkEmitter ORE(&F);
1721   if (!LoopInterchange(&AR.SE, &AR.LI, &DI, &AR.DT, CC, &ORE).run(LN))
1722     return PreservedAnalyses::all();
1723   U.markLoopNestChanged(true);
1724   return getLoopPassPreservedAnalyses();
1725 }
1726