1 //===- LoopInterchange.cpp - Loop interchange pass-------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This Pass handles loop interchange transform.
11 // This pass interchanges loops to provide a more cache-friendly memory access
14 //===----------------------------------------------------------------------===//
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/LoopPass.h"
23 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
24 #include "llvm/Analysis/ScalarEvolution.h"
25 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
26 #include "llvm/IR/BasicBlock.h"
27 #include "llvm/IR/Constants.h"
28 #include "llvm/IR/DiagnosticInfo.h"
29 #include "llvm/IR/Dominators.h"
30 #include "llvm/IR/Function.h"
31 #include "llvm/IR/InstrTypes.h"
32 #include "llvm/IR/Instruction.h"
33 #include "llvm/IR/Instructions.h"
34 #include "llvm/IR/Type.h"
35 #include "llvm/IR/User.h"
36 #include "llvm/IR/Value.h"
37 #include "llvm/Pass.h"
38 #include "llvm/Support/Casting.h"
39 #include "llvm/Support/CommandLine.h"
40 #include "llvm/Support/Debug.h"
41 #include "llvm/Support/ErrorHandling.h"
42 #include "llvm/Support/raw_ostream.h"
43 #include "llvm/Transforms/Scalar.h"
44 #include "llvm/Transforms/Utils.h"
45 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
46 #include "llvm/Transforms/Utils/LoopUtils.h"
53 #define DEBUG_TYPE "loop-interchange"
55 STATISTIC(LoopsInterchanged, "Number of loops interchanged");
57 static cl::opt<int> LoopInterchangeCostThreshold(
58 "loop-interchange-threshold", cl::init(0), cl::Hidden,
59 cl::desc("Interchange if you gain more than this number"));
63 using LoopVector = SmallVector<Loop *, 8>;
65 // TODO: Check if we can use a sparse matrix here.
66 using CharMatrix = std::vector<std::vector<char>>;
68 } // end anonymous namespace
70 // Maximum number of dependencies that can be handled in the dependency matrix.
71 static const unsigned MaxMemInstrCount = 100;
73 // Maximum loop depth supported.
74 static const unsigned MaxLoopNestDepth = 10;
76 #ifdef DUMP_DEP_MATRICIES
77 static void printDepMatrix(CharMatrix &DepMatrix) {
78 for (auto &Row : DepMatrix) {
80 LLVM_DEBUG(dbgs() << D << " ");
81 LLVM_DEBUG(dbgs() << "\n");
86 static bool populateDependencyMatrix(CharMatrix &DepMatrix, unsigned Level,
87 Loop *L, DependenceInfo *DI) {
88 using ValueVector = SmallVector<Value *, 16>;
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))
98 if (auto *Ld = dyn_cast<LoadInst>(&I)) {
101 MemInstr.push_back(&I);
102 } else if (auto *St = dyn_cast<StoreInst>(&I)) {
105 MemInstr.push_back(&I);
110 LLVM_DEBUG(dbgs() << "Found " << MemInstr.size()
111 << " Loads and Stores to analyze\n");
113 ValueVector::iterator I, IE, J, JE;
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);
122 // Ignore Input dependencies.
123 if (isa<LoadInst>(Src) && isa<LoadInst>(Dst))
125 // Track Output, Flow, and Anti dependencies.
126 if (auto D = DI->depends(Src, Dst, true)) {
127 assert(D->isOrdered() && "Expected an output, flow or anti dep.");
128 LLVM_DEBUG(StringRef DepType =
129 D->isFlow() ? "flow" : D->isAnti() ? "anti" : "output";
130 dbgs() << "Found " << DepType
131 << " dependency between Src and Dst\n"
132 << " Src:" << *Src << "\n Dst:" << *Dst << '\n');
133 unsigned Levels = D->getLevels();
135 for (unsigned II = 1; II <= Levels; ++II) {
136 const SCEV *Distance = D->getDistance(II);
137 const SCEVConstant *SCEVConst =
138 dyn_cast_or_null<SCEVConstant>(Distance);
140 const ConstantInt *CI = SCEVConst->getValue();
141 if (CI->isNegative())
143 else if (CI->isZero())
147 Dep.push_back(Direction);
148 } else if (D->isScalar(II)) {
150 Dep.push_back(Direction);
152 unsigned Dir = D->getDirection(II);
153 if (Dir == Dependence::DVEntry::LT ||
154 Dir == Dependence::DVEntry::LE)
156 else if (Dir == Dependence::DVEntry::GT ||
157 Dir == Dependence::DVEntry::GE)
159 else if (Dir == Dependence::DVEntry::EQ)
163 Dep.push_back(Direction);
166 while (Dep.size() != Level) {
170 DepMatrix.push_back(Dep);
171 if (DepMatrix.size() > MaxMemInstrCount) {
172 LLVM_DEBUG(dbgs() << "Cannot handle more than " << MaxMemInstrCount
173 << " dependencies inside loop\n");
183 // A loop is moved from index 'from' to an index 'to'. Update the Dependence
184 // matrix by exchanging the two columns.
185 static void interChangeDependencies(CharMatrix &DepMatrix, unsigned FromIndx,
187 unsigned numRows = DepMatrix.size();
188 for (unsigned i = 0; i < numRows; ++i) {
189 char TmpVal = DepMatrix[i][ToIndx];
190 DepMatrix[i][ToIndx] = DepMatrix[i][FromIndx];
191 DepMatrix[i][FromIndx] = TmpVal;
195 // Checks if outermost non '=','S'or'I' dependence in the dependence matrix is
197 static bool isOuterMostDepPositive(CharMatrix &DepMatrix, unsigned Row,
199 for (unsigned i = 0; i <= Column; ++i) {
200 if (DepMatrix[Row][i] == '<')
202 if (DepMatrix[Row][i] == '>')
205 // All dependencies were '=','S' or 'I'
209 // Checks if no dependence exist in the dependency matrix in Row before Column.
210 static bool containsNoDependence(CharMatrix &DepMatrix, unsigned Row,
212 for (unsigned i = 0; i < Column; ++i) {
213 if (DepMatrix[Row][i] != '=' && DepMatrix[Row][i] != 'S' &&
214 DepMatrix[Row][i] != 'I')
220 static bool validDepInterchange(CharMatrix &DepMatrix, unsigned Row,
221 unsigned OuterLoopId, char InnerDep,
223 if (isOuterMostDepPositive(DepMatrix, Row, OuterLoopId))
226 if (InnerDep == OuterDep)
229 // It is legal to interchange if and only if after interchange no row has a
230 // '>' direction as the leftmost non-'='.
232 if (InnerDep == '=' || InnerDep == 'S' || InnerDep == 'I')
238 if (InnerDep == '>') {
239 // If OuterLoopId represents outermost loop then interchanging will make the
240 // 1st dependency as '>'
241 if (OuterLoopId == 0)
244 // If all dependencies before OuterloopId are '=','S'or 'I'. Then
245 // interchanging will result in this row having an outermost non '='
247 if (!containsNoDependence(DepMatrix, Row, OuterLoopId))
254 // Checks if it is legal to interchange 2 loops.
255 // [Theorem] A permutation of the loops in a perfect nest is legal if and only
256 // if the direction matrix, after the same permutation is applied to its
257 // columns, has no ">" direction as the leftmost non-"=" direction in any row.
258 static bool isLegalToInterChangeLoops(CharMatrix &DepMatrix,
259 unsigned InnerLoopId,
260 unsigned OuterLoopId) {
261 unsigned NumRows = DepMatrix.size();
262 // For each row check if it is valid to interchange.
263 for (unsigned Row = 0; Row < NumRows; ++Row) {
264 char InnerDep = DepMatrix[Row][InnerLoopId];
265 char OuterDep = DepMatrix[Row][OuterLoopId];
266 if (InnerDep == '*' || OuterDep == '*')
268 if (!validDepInterchange(DepMatrix, Row, OuterLoopId, InnerDep, OuterDep))
274 static LoopVector populateWorklist(Loop &L) {
275 LLVM_DEBUG(dbgs() << "Calling populateWorklist on Func: "
276 << L.getHeader()->getParent()->getName() << " Loop: %"
277 << L.getHeader()->getName() << '\n');
279 Loop *CurrentLoop = &L;
280 const std::vector<Loop *> *Vec = &CurrentLoop->getSubLoops();
281 while (!Vec->empty()) {
282 // The current loop has multiple subloops in it hence it is not tightly
284 // Discard all loops above it added into Worklist.
285 if (Vec->size() != 1)
288 LoopList.push_back(CurrentLoop);
289 CurrentLoop = Vec->front();
290 Vec = &CurrentLoop->getSubLoops();
292 LoopList.push_back(CurrentLoop);
296 static PHINode *getInductionVariable(Loop *L, ScalarEvolution *SE) {
297 PHINode *InnerIndexVar = L->getCanonicalInductionVariable();
299 return InnerIndexVar;
300 if (L->getLoopLatch() == nullptr || L->getLoopPredecessor() == nullptr)
302 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
303 PHINode *PhiVar = cast<PHINode>(I);
304 Type *PhiTy = PhiVar->getType();
305 if (!PhiTy->isIntegerTy() && !PhiTy->isFloatingPointTy() &&
306 !PhiTy->isPointerTy())
308 const SCEVAddRecExpr *AddRec =
309 dyn_cast<SCEVAddRecExpr>(SE->getSCEV(PhiVar));
310 if (!AddRec || !AddRec->isAffine())
312 const SCEV *Step = AddRec->getStepRecurrence(*SE);
313 if (!isa<SCEVConstant>(Step))
315 // Found the induction variable.
316 // FIXME: Handle loops with more than one induction variable. Note that,
317 // currently, legality makes sure we have only one induction variable.
325 /// LoopInterchangeLegality checks if it is legal to interchange the loop.
326 class LoopInterchangeLegality {
328 LoopInterchangeLegality(Loop *Outer, Loop *Inner, ScalarEvolution *SE,
329 OptimizationRemarkEmitter *ORE)
330 : OuterLoop(Outer), InnerLoop(Inner), SE(SE), ORE(ORE) {}
332 /// Check if the loops can be interchanged.
333 bool canInterchangeLoops(unsigned InnerLoopId, unsigned OuterLoopId,
334 CharMatrix &DepMatrix);
336 /// Check if the loop structure is understood. We do not handle triangular
338 bool isLoopStructureUnderstood(PHINode *InnerInductionVar);
340 bool currentLimitations();
342 const SmallPtrSetImpl<PHINode *> &getOuterInnerReductions() const {
343 return OuterInnerReductions;
347 bool tightlyNested(Loop *Outer, Loop *Inner);
348 bool containsUnsafeInstructions(BasicBlock *BB);
350 /// Discover induction and reduction PHIs in the header of \p L. Induction
351 /// PHIs are added to \p Inductions, reductions are added to
352 /// OuterInnerReductions. When the outer loop is passed, the inner loop needs
353 /// to be passed as \p InnerLoop.
354 bool findInductionAndReductions(Loop *L,
355 SmallVector<PHINode *, 8> &Inductions,
363 /// Interface to emit optimization remarks.
364 OptimizationRemarkEmitter *ORE;
366 /// Set of reduction PHIs taking part of a reduction across the inner and
368 SmallPtrSet<PHINode *, 4> OuterInnerReductions;
371 /// LoopInterchangeProfitability checks if it is profitable to interchange the
373 class LoopInterchangeProfitability {
375 LoopInterchangeProfitability(Loop *Outer, Loop *Inner, ScalarEvolution *SE,
376 OptimizationRemarkEmitter *ORE)
377 : OuterLoop(Outer), InnerLoop(Inner), SE(SE), ORE(ORE) {}
379 /// Check if the loop interchange is profitable.
380 bool isProfitable(unsigned InnerLoopId, unsigned OuterLoopId,
381 CharMatrix &DepMatrix);
384 int getInstrOrderCost();
392 /// Interface to emit optimization remarks.
393 OptimizationRemarkEmitter *ORE;
396 /// LoopInterchangeTransform interchanges the loop.
397 class LoopInterchangeTransform {
399 LoopInterchangeTransform(Loop *Outer, Loop *Inner, ScalarEvolution *SE,
400 LoopInfo *LI, DominatorTree *DT,
401 BasicBlock *LoopNestExit,
402 const LoopInterchangeLegality &LIL)
403 : OuterLoop(Outer), InnerLoop(Inner), SE(SE), LI(LI), DT(DT),
404 LoopExit(LoopNestExit), LIL(LIL) {}
406 /// Interchange OuterLoop and InnerLoop.
408 void restructureLoops(Loop *NewInner, Loop *NewOuter,
409 BasicBlock *OrigInnerPreHeader,
410 BasicBlock *OrigOuterPreHeader);
411 void removeChildLoop(Loop *OuterLoop, Loop *InnerLoop);
414 void splitInnerLoopLatch(Instruction *);
415 void splitInnerLoopHeader();
416 bool adjustLoopLinks();
417 void adjustLoopPreheaders();
418 bool adjustLoopBranches();
428 BasicBlock *LoopExit;
430 const LoopInterchangeLegality &LIL;
433 // Main LoopInterchange Pass.
434 struct LoopInterchange : public LoopPass {
436 ScalarEvolution *SE = nullptr;
437 LoopInfo *LI = nullptr;
438 DependenceInfo *DI = nullptr;
439 DominatorTree *DT = nullptr;
441 /// Interface to emit optimization remarks.
442 OptimizationRemarkEmitter *ORE;
444 LoopInterchange() : LoopPass(ID) {
445 initializeLoopInterchangePass(*PassRegistry::getPassRegistry());
448 void getAnalysisUsage(AnalysisUsage &AU) const override {
449 AU.addRequired<DependenceAnalysisWrapperPass>();
450 AU.addRequired<OptimizationRemarkEmitterWrapperPass>();
452 getLoopAnalysisUsage(AU);
455 bool runOnLoop(Loop *L, LPPassManager &LPM) override {
456 if (skipLoop(L) || L->getParentLoop())
459 SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
460 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
461 DI = &getAnalysis<DependenceAnalysisWrapperPass>().getDI();
462 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
463 ORE = &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE();
465 return processLoopList(populateWorklist(*L));
468 bool isComputableLoopNest(LoopVector LoopList) {
469 for (Loop *L : LoopList) {
470 const SCEV *ExitCountOuter = SE->getBackedgeTakenCount(L);
471 if (ExitCountOuter == SE->getCouldNotCompute()) {
472 LLVM_DEBUG(dbgs() << "Couldn't compute backedge count\n");
475 if (L->getNumBackEdges() != 1) {
476 LLVM_DEBUG(dbgs() << "NumBackEdges is not equal to 1\n");
479 if (!L->getExitingBlock()) {
480 LLVM_DEBUG(dbgs() << "Loop doesn't have unique exit block\n");
487 unsigned selectLoopForInterchange(const LoopVector &LoopList) {
488 // TODO: Add a better heuristic to select the loop to be interchanged based
489 // on the dependence matrix. Currently we select the innermost loop.
490 return LoopList.size() - 1;
493 bool processLoopList(LoopVector LoopList) {
494 bool Changed = false;
495 unsigned LoopNestDepth = LoopList.size();
496 if (LoopNestDepth < 2) {
497 LLVM_DEBUG(dbgs() << "Loop doesn't contain minimum nesting level.\n");
500 if (LoopNestDepth > MaxLoopNestDepth) {
501 LLVM_DEBUG(dbgs() << "Cannot handle loops of depth greater than "
502 << MaxLoopNestDepth << "\n");
505 if (!isComputableLoopNest(LoopList)) {
506 LLVM_DEBUG(dbgs() << "Not valid loop candidate for interchange\n");
510 LLVM_DEBUG(dbgs() << "Processing LoopList of size = " << LoopNestDepth
513 CharMatrix DependencyMatrix;
514 Loop *OuterMostLoop = *(LoopList.begin());
515 if (!populateDependencyMatrix(DependencyMatrix, LoopNestDepth,
516 OuterMostLoop, DI)) {
517 LLVM_DEBUG(dbgs() << "Populating dependency matrix failed\n");
520 #ifdef DUMP_DEP_MATRICIES
521 LLVM_DEBUG(dbgs() << "Dependence before interchange\n");
522 printDepMatrix(DependencyMatrix);
525 // Get the Outermost loop exit.
526 BasicBlock *LoopNestExit = OuterMostLoop->getExitBlock();
528 LLVM_DEBUG(dbgs() << "OuterMostLoop needs an unique exit block");
532 unsigned SelecLoopId = selectLoopForInterchange(LoopList);
533 // Move the selected loop outwards to the best possible position.
534 for (unsigned i = SelecLoopId; i > 0; i--) {
536 processLoop(LoopList, i, i - 1, LoopNestExit, DependencyMatrix);
539 // Loops interchanged reflect the same in LoopList
540 std::swap(LoopList[i - 1], LoopList[i]);
542 // Update the DependencyMatrix
543 interChangeDependencies(DependencyMatrix, i, i - 1);
544 #ifdef DUMP_DEP_MATRICIES
545 LLVM_DEBUG(dbgs() << "Dependence after interchange\n");
546 printDepMatrix(DependencyMatrix);
548 Changed |= Interchanged;
553 bool processLoop(LoopVector LoopList, unsigned InnerLoopId,
554 unsigned OuterLoopId, BasicBlock *LoopNestExit,
555 std::vector<std::vector<char>> &DependencyMatrix) {
556 LLVM_DEBUG(dbgs() << "Processing Inner Loop Id = " << InnerLoopId
557 << " and OuterLoopId = " << OuterLoopId << "\n");
558 Loop *InnerLoop = LoopList[InnerLoopId];
559 Loop *OuterLoop = LoopList[OuterLoopId];
561 LoopInterchangeLegality LIL(OuterLoop, InnerLoop, SE, ORE);
562 if (!LIL.canInterchangeLoops(InnerLoopId, OuterLoopId, DependencyMatrix)) {
563 LLVM_DEBUG(dbgs() << "Not interchanging loops. Cannot prove legality.\n");
566 LLVM_DEBUG(dbgs() << "Loops are legal to interchange\n");
567 LoopInterchangeProfitability LIP(OuterLoop, InnerLoop, SE, ORE);
568 if (!LIP.isProfitable(InnerLoopId, OuterLoopId, DependencyMatrix)) {
569 LLVM_DEBUG(dbgs() << "Interchanging loops not profitable.\n");
574 return OptimizationRemark(DEBUG_TYPE, "Interchanged",
575 InnerLoop->getStartLoc(),
576 InnerLoop->getHeader())
577 << "Loop interchanged with enclosing loop.";
580 LoopInterchangeTransform LIT(OuterLoop, InnerLoop, SE, LI, DT, LoopNestExit,
583 LLVM_DEBUG(dbgs() << "Loops interchanged.\n");
589 } // end anonymous namespace
591 bool LoopInterchangeLegality::containsUnsafeInstructions(BasicBlock *BB) {
592 return any_of(*BB, [](const Instruction &I) {
593 return I.mayHaveSideEffects() || I.mayReadFromMemory();
597 bool LoopInterchangeLegality::tightlyNested(Loop *OuterLoop, Loop *InnerLoop) {
598 BasicBlock *OuterLoopHeader = OuterLoop->getHeader();
599 BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader();
600 BasicBlock *OuterLoopLatch = OuterLoop->getLoopLatch();
602 LLVM_DEBUG(dbgs() << "Checking if loops are tightly nested\n");
604 // A perfectly nested loop will not have any branch in between the outer and
605 // inner block i.e. outer header will branch to either inner preheader and
607 BranchInst *OuterLoopHeaderBI =
608 dyn_cast<BranchInst>(OuterLoopHeader->getTerminator());
609 if (!OuterLoopHeaderBI)
612 for (BasicBlock *Succ : successors(OuterLoopHeaderBI))
613 if (Succ != InnerLoopPreHeader && Succ != InnerLoop->getHeader() &&
614 Succ != OuterLoopLatch)
617 LLVM_DEBUG(dbgs() << "Checking instructions in Loop header and Loop latch\n");
618 // We do not have any basic block in between now make sure the outer header
619 // and outer loop latch doesn't contain any unsafe instructions.
620 if (containsUnsafeInstructions(OuterLoopHeader) ||
621 containsUnsafeInstructions(OuterLoopLatch))
624 LLVM_DEBUG(dbgs() << "Loops are perfectly nested\n");
625 // We have a perfect loop nest.
629 bool LoopInterchangeLegality::isLoopStructureUnderstood(
630 PHINode *InnerInduction) {
631 unsigned Num = InnerInduction->getNumOperands();
632 BasicBlock *InnerLoopPreheader = InnerLoop->getLoopPreheader();
633 for (unsigned i = 0; i < Num; ++i) {
634 Value *Val = InnerInduction->getOperand(i);
635 if (isa<Constant>(Val))
637 Instruction *I = dyn_cast<Instruction>(Val);
640 // TODO: Handle triangular loops.
641 // e.g. for(int i=0;i<N;i++)
642 // for(int j=i;j<N;j++)
643 unsigned IncomBlockIndx = PHINode::getIncomingValueNumForOperand(i);
644 if (InnerInduction->getIncomingBlock(IncomBlockIndx) ==
645 InnerLoopPreheader &&
646 !OuterLoop->isLoopInvariant(I)) {
653 // If SV is a LCSSA PHI node with a single incoming value, return the incoming
655 static Value *followLCSSA(Value *SV) {
656 PHINode *PHI = dyn_cast<PHINode>(SV);
660 if (PHI->getNumIncomingValues() != 1)
662 return followLCSSA(PHI->getIncomingValue(0));
665 // Check V's users to see if it is involved in a reduction in L.
666 static PHINode *findInnerReductionPhi(Loop *L, Value *V) {
667 for (Value *User : V->users()) {
668 if (PHINode *PHI = dyn_cast<PHINode>(User)) {
669 if (PHI->getNumIncomingValues() == 1)
671 RecurrenceDescriptor RD;
672 if (RecurrenceDescriptor::isReductionPHI(PHI, L, RD))
681 bool LoopInterchangeLegality::findInductionAndReductions(
682 Loop *L, SmallVector<PHINode *, 8> &Inductions, Loop *InnerLoop) {
683 if (!L->getLoopLatch() || !L->getLoopPredecessor())
685 for (PHINode &PHI : L->getHeader()->phis()) {
686 RecurrenceDescriptor RD;
687 InductionDescriptor ID;
688 if (InductionDescriptor::isInductionPHI(&PHI, L, SE, ID))
689 Inductions.push_back(&PHI);
691 // PHIs in inner loops need to be part of a reduction in the outer loop,
692 // discovered when checking the PHIs of the outer loop earlier.
694 if (OuterInnerReductions.find(&PHI) == OuterInnerReductions.end()) {
695 LLVM_DEBUG(dbgs() << "Inner loop PHI is not part of reductions "
696 "across the outer loop.\n");
700 assert(PHI.getNumIncomingValues() == 2 &&
701 "Phis in loop header should have exactly 2 incoming values");
702 // Check if we have a PHI node in the outer loop that has a reduction
703 // result from the inner loop as an incoming value.
704 Value *V = followLCSSA(PHI.getIncomingValueForBlock(L->getLoopLatch()));
705 PHINode *InnerRedPhi = findInnerReductionPhi(InnerLoop, V);
707 !llvm::any_of(InnerRedPhi->incoming_values(),
708 [&PHI](Value *V) { return V == &PHI; })) {
711 << "Failed to recognize PHI as an induction or reduction.\n");
714 OuterInnerReductions.insert(&PHI);
715 OuterInnerReductions.insert(InnerRedPhi);
722 static bool containsSafePHI(BasicBlock *Block, bool isOuterLoopExitBlock) {
723 for (PHINode &PHI : Block->phis()) {
724 // Reduction lcssa phi will have only 1 incoming block that from loop latch.
725 if (PHI.getNumIncomingValues() > 1)
727 Instruction *Ins = dyn_cast<Instruction>(PHI.getIncomingValue(0));
730 // Incoming value for lcssa phi's in outer loop exit can only be inner loop
731 // exits lcssa phi else it would not be tightly nested.
732 if (!isa<PHINode>(Ins) && isOuterLoopExitBlock)
738 // This function indicates the current limitations in the transform as a result
739 // of which we do not proceed.
740 bool LoopInterchangeLegality::currentLimitations() {
741 BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader();
742 BasicBlock *InnerLoopLatch = InnerLoop->getLoopLatch();
744 // transform currently expects the loop latches to also be the exiting
746 if (InnerLoop->getExitingBlock() != InnerLoopLatch ||
747 OuterLoop->getExitingBlock() != OuterLoop->getLoopLatch() ||
748 !isa<BranchInst>(InnerLoopLatch->getTerminator()) ||
749 !isa<BranchInst>(OuterLoop->getLoopLatch()->getTerminator())) {
751 dbgs() << "Loops where the latch is not the exiting block are not"
752 << " supported currently.\n");
754 return OptimizationRemarkMissed(DEBUG_TYPE, "ExitingNotLatch",
755 OuterLoop->getStartLoc(),
756 OuterLoop->getHeader())
757 << "Loops where the latch is not the exiting block cannot be"
758 " interchange currently.";
763 PHINode *InnerInductionVar;
764 SmallVector<PHINode *, 8> Inductions;
765 if (!findInductionAndReductions(OuterLoop, Inductions, InnerLoop)) {
767 dbgs() << "Only outer loops with induction or reduction PHI nodes "
768 << "are supported currently.\n");
770 return OptimizationRemarkMissed(DEBUG_TYPE, "UnsupportedPHIOuter",
771 OuterLoop->getStartLoc(),
772 OuterLoop->getHeader())
773 << "Only outer loops with induction or reduction PHI nodes can be"
774 " interchanged currently.";
779 // TODO: Currently we handle only loops with 1 induction variable.
780 if (Inductions.size() != 1) {
781 LLVM_DEBUG(dbgs() << "Loops with more than 1 induction variables are not "
782 << "supported currently.\n");
784 return OptimizationRemarkMissed(DEBUG_TYPE, "MultiIndutionOuter",
785 OuterLoop->getStartLoc(),
786 OuterLoop->getHeader())
787 << "Only outer loops with 1 induction variable can be "
788 "interchanged currently.";
794 if (!findInductionAndReductions(InnerLoop, Inductions, nullptr)) {
796 dbgs() << "Only inner loops with induction or reduction PHI nodes "
797 << "are supported currently.\n");
799 return OptimizationRemarkMissed(DEBUG_TYPE, "UnsupportedPHIInner",
800 InnerLoop->getStartLoc(),
801 InnerLoop->getHeader())
802 << "Only inner loops with induction or reduction PHI nodes can be"
803 " interchange currently.";
808 // TODO: Currently we handle only loops with 1 induction variable.
809 if (Inductions.size() != 1) {
811 dbgs() << "We currently only support loops with 1 induction variable."
812 << "Failed to interchange due to current limitation\n");
814 return OptimizationRemarkMissed(DEBUG_TYPE, "MultiInductionInner",
815 InnerLoop->getStartLoc(),
816 InnerLoop->getHeader())
817 << "Only inner loops with 1 induction variable can be "
818 "interchanged currently.";
822 InnerInductionVar = Inductions.pop_back_val();
824 // TODO: Triangular loops are not handled for now.
825 if (!isLoopStructureUnderstood(InnerInductionVar)) {
826 LLVM_DEBUG(dbgs() << "Loop structure not understood by pass\n");
828 return OptimizationRemarkMissed(DEBUG_TYPE, "UnsupportedStructureInner",
829 InnerLoop->getStartLoc(),
830 InnerLoop->getHeader())
831 << "Inner loop structure not understood currently.";
836 // TODO: We only handle LCSSA PHI's corresponding to reduction for now.
837 BasicBlock *InnerExit = InnerLoop->getExitBlock();
838 if (!containsSafePHI(InnerExit, false)) {
840 dbgs() << "Can only handle LCSSA PHIs in inner loops currently.\n");
842 return OptimizationRemarkMissed(DEBUG_TYPE, "NoLCSSAPHIOuterInner",
843 InnerLoop->getStartLoc(),
844 InnerLoop->getHeader())
845 << "Only inner loops with LCSSA PHIs can be interchange "
851 // TODO: Current limitation: Since we split the inner loop latch at the point
852 // were induction variable is incremented (induction.next); We cannot have
853 // more than 1 user of induction.next since it would result in broken code
856 // for(i=0;i<N;i++) {
857 // for(j = 0;j<M;j++) {
858 // A[j+1][i+2] = A[j][i]+k;
861 Instruction *InnerIndexVarInc = nullptr;
862 if (InnerInductionVar->getIncomingBlock(0) == InnerLoopPreHeader)
864 dyn_cast<Instruction>(InnerInductionVar->getIncomingValue(1));
867 dyn_cast<Instruction>(InnerInductionVar->getIncomingValue(0));
869 if (!InnerIndexVarInc) {
871 dbgs() << "Did not find an instruction to increment the induction "
874 return OptimizationRemarkMissed(DEBUG_TYPE, "NoIncrementInInner",
875 InnerLoop->getStartLoc(),
876 InnerLoop->getHeader())
877 << "The inner loop does not increment the induction variable.";
882 // Since we split the inner loop latch on this induction variable. Make sure
883 // we do not have any instruction between the induction variable and branch
886 bool FoundInduction = false;
887 for (const Instruction &I :
888 llvm::reverse(InnerLoopLatch->instructionsWithoutDebug())) {
889 if (isa<BranchInst>(I) || isa<CmpInst>(I) || isa<TruncInst>(I) ||
893 // We found an instruction. If this is not induction variable then it is not
894 // safe to split this loop latch.
895 if (!I.isIdenticalTo(InnerIndexVarInc)) {
896 LLVM_DEBUG(dbgs() << "Found unsupported instructions between induction "
897 << "variable increment and branch.\n");
899 return OptimizationRemarkMissed(
900 DEBUG_TYPE, "UnsupportedInsBetweenInduction",
901 InnerLoop->getStartLoc(), InnerLoop->getHeader())
902 << "Found unsupported instruction between induction variable "
903 "increment and branch.";
908 FoundInduction = true;
911 // The loop latch ended and we didn't find the induction variable return as
912 // current limitation.
913 if (!FoundInduction) {
914 LLVM_DEBUG(dbgs() << "Did not find the induction variable.\n");
916 return OptimizationRemarkMissed(DEBUG_TYPE, "NoIndutionVariable",
917 InnerLoop->getStartLoc(),
918 InnerLoop->getHeader())
919 << "Did not find the induction variable.";
926 // We currently support LCSSA PHI nodes in the outer loop exit, if their
927 // incoming values do not come from the outer loop latch or if the
928 // outer loop latch has a single predecessor. In that case, the value will
929 // be available if both the inner and outer loop conditions are true, which
930 // will still be true after interchanging. If we have multiple predecessor,
931 // that may not be the case, e.g. because the outer loop latch may be executed
932 // if the inner loop is not executed.
933 static bool areLoopExitPHIsSupported(Loop *OuterLoop, Loop *InnerLoop) {
934 BasicBlock *LoopNestExit = OuterLoop->getUniqueExitBlock();
935 for (PHINode &PHI : LoopNestExit->phis()) {
936 // FIXME: We currently are not able to detect floating point reductions
937 // and have to use floating point PHIs as a proxy to prevent
938 // interchanging in the presence of floating point reductions.
939 if (PHI.getType()->isFloatingPointTy())
941 for (unsigned i = 0; i < PHI.getNumIncomingValues(); i++) {
942 Instruction *IncomingI = dyn_cast<Instruction>(PHI.getIncomingValue(i));
943 if (!IncomingI || IncomingI->getParent() != OuterLoop->getLoopLatch())
946 // The incoming value is defined in the outer loop latch. Currently we
947 // only support that in case the outer loop latch has a single predecessor.
948 // This guarantees that the outer loop latch is executed if and only if
949 // the inner loop is executed (because tightlyNested() guarantees that the
950 // outer loop header only branches to the inner loop or the outer loop
952 // FIXME: We could weaken this logic and allow multiple predecessors,
953 // if the values are produced outside the loop latch. We would need
954 // additional logic to update the PHI nodes in the exit block as
956 if (OuterLoop->getLoopLatch()->getUniquePredecessor() == nullptr)
963 bool LoopInterchangeLegality::canInterchangeLoops(unsigned InnerLoopId,
964 unsigned OuterLoopId,
965 CharMatrix &DepMatrix) {
966 if (!isLegalToInterChangeLoops(DepMatrix, InnerLoopId, OuterLoopId)) {
967 LLVM_DEBUG(dbgs() << "Failed interchange InnerLoopId = " << InnerLoopId
968 << " and OuterLoopId = " << OuterLoopId
969 << " due to dependence\n");
971 return OptimizationRemarkMissed(DEBUG_TYPE, "Dependence",
972 InnerLoop->getStartLoc(),
973 InnerLoop->getHeader())
974 << "Cannot interchange loops due to dependences.";
978 // Check if outer and inner loop contain legal instructions only.
979 for (auto *BB : OuterLoop->blocks())
980 for (Instruction &I : BB->instructionsWithoutDebug())
981 if (CallInst *CI = dyn_cast<CallInst>(&I)) {
982 // readnone functions do not prevent interchanging.
983 if (CI->doesNotReadMemory())
986 dbgs() << "Loops with call instructions cannot be interchanged "
989 return OptimizationRemarkMissed(DEBUG_TYPE, "CallInst",
992 << "Cannot interchange loops due to call instruction.";
998 // TODO: The loops could not be interchanged due to current limitations in the
1000 if (currentLimitations()) {
1001 LLVM_DEBUG(dbgs() << "Not legal because of current transform limitation\n");
1005 // Check if the loops are tightly nested.
1006 if (!tightlyNested(OuterLoop, InnerLoop)) {
1007 LLVM_DEBUG(dbgs() << "Loops not tightly nested\n");
1009 return OptimizationRemarkMissed(DEBUG_TYPE, "NotTightlyNested",
1010 InnerLoop->getStartLoc(),
1011 InnerLoop->getHeader())
1012 << "Cannot interchange loops because they are not tightly "
1018 if (!areLoopExitPHIsSupported(OuterLoop, InnerLoop)) {
1019 LLVM_DEBUG(dbgs() << "Found unsupported PHI nodes in outer loop exit.\n");
1021 return OptimizationRemarkMissed(DEBUG_TYPE, "UnsupportedExitPHI",
1022 OuterLoop->getStartLoc(),
1023 OuterLoop->getHeader())
1024 << "Found unsupported PHI node in loop exit.";
1032 int LoopInterchangeProfitability::getInstrOrderCost() {
1033 unsigned GoodOrder, BadOrder;
1034 BadOrder = GoodOrder = 0;
1035 for (BasicBlock *BB : InnerLoop->blocks()) {
1036 for (Instruction &Ins : *BB) {
1037 if (const GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(&Ins)) {
1038 unsigned NumOp = GEP->getNumOperands();
1039 bool FoundInnerInduction = false;
1040 bool FoundOuterInduction = false;
1041 for (unsigned i = 0; i < NumOp; ++i) {
1042 const SCEV *OperandVal = SE->getSCEV(GEP->getOperand(i));
1043 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(OperandVal);
1047 // If we find the inner induction after an outer induction e.g.
1048 // for(int i=0;i<N;i++)
1049 // for(int j=0;j<N;j++)
1050 // A[i][j] = A[i-1][j-1]+k;
1051 // then it is a good order.
1052 if (AR->getLoop() == InnerLoop) {
1053 // We found an InnerLoop induction after OuterLoop induction. It is
1055 FoundInnerInduction = true;
1056 if (FoundOuterInduction) {
1061 // If we find the outer induction after an inner induction e.g.
1062 // for(int i=0;i<N;i++)
1063 // for(int j=0;j<N;j++)
1064 // A[j][i] = A[j-1][i-1]+k;
1065 // then it is a bad order.
1066 if (AR->getLoop() == OuterLoop) {
1067 // We found an OuterLoop induction after InnerLoop induction. It is
1069 FoundOuterInduction = true;
1070 if (FoundInnerInduction) {
1079 return GoodOrder - BadOrder;
1082 static bool isProfitableForVectorization(unsigned InnerLoopId,
1083 unsigned OuterLoopId,
1084 CharMatrix &DepMatrix) {
1085 // TODO: Improve this heuristic to catch more cases.
1086 // If the inner loop is loop independent or doesn't carry any dependency it is
1087 // profitable to move this to outer position.
1088 for (auto &Row : DepMatrix) {
1089 if (Row[InnerLoopId] != 'S' && Row[InnerLoopId] != 'I')
1091 // TODO: We need to improve this heuristic.
1092 if (Row[OuterLoopId] != '=')
1095 // If outer loop has dependence and inner loop is loop independent then it is
1096 // profitable to interchange to enable parallelism.
1097 // If there are no dependences, interchanging will not improve anything.
1098 return !DepMatrix.empty();
1101 bool LoopInterchangeProfitability::isProfitable(unsigned InnerLoopId,
1102 unsigned OuterLoopId,
1103 CharMatrix &DepMatrix) {
1104 // TODO: Add better profitability checks.
1106 // 1) Construct dependency matrix and move the one with no loop carried dep
1107 // inside to enable vectorization.
1109 // This is rough cost estimation algorithm. It counts the good and bad order
1110 // of induction variables in the instruction and allows reordering if number
1111 // of bad orders is more than good.
1112 int Cost = getInstrOrderCost();
1113 LLVM_DEBUG(dbgs() << "Cost = " << Cost << "\n");
1114 if (Cost < -LoopInterchangeCostThreshold)
1117 // It is not profitable as per current cache profitability model. But check if
1118 // we can move this loop outside to improve parallelism.
1119 if (isProfitableForVectorization(InnerLoopId, OuterLoopId, DepMatrix))
1123 return OptimizationRemarkMissed(DEBUG_TYPE, "InterchangeNotProfitable",
1124 InnerLoop->getStartLoc(),
1125 InnerLoop->getHeader())
1126 << "Interchanging loops is too costly (cost="
1127 << ore::NV("Cost", Cost) << ", threshold="
1128 << ore::NV("Threshold", LoopInterchangeCostThreshold)
1129 << ") and it does not improve parallelism.";
1134 void LoopInterchangeTransform::removeChildLoop(Loop *OuterLoop,
1136 for (Loop *L : *OuterLoop)
1137 if (L == InnerLoop) {
1138 OuterLoop->removeChildLoop(L);
1141 llvm_unreachable("Couldn't find loop");
1144 /// Update LoopInfo, after interchanging. NewInner and NewOuter refer to the
1145 /// new inner and outer loop after interchanging: NewInner is the original
1146 /// outer loop and NewOuter is the original inner loop.
1148 /// Before interchanging, we have the following structure
1158 // After interchanging:
1167 void LoopInterchangeTransform::restructureLoops(
1168 Loop *NewInner, Loop *NewOuter, BasicBlock *OrigInnerPreHeader,
1169 BasicBlock *OrigOuterPreHeader) {
1170 Loop *OuterLoopParent = OuterLoop->getParentLoop();
1171 // The original inner loop preheader moves from the new inner loop to
1172 // the parent loop, if there is one.
1173 NewInner->removeBlockFromLoop(OrigInnerPreHeader);
1174 LI->changeLoopFor(OrigInnerPreHeader, OuterLoopParent);
1176 // Switch the loop levels.
1177 if (OuterLoopParent) {
1178 // Remove the loop from its parent loop.
1179 removeChildLoop(OuterLoopParent, NewInner);
1180 removeChildLoop(NewInner, NewOuter);
1181 OuterLoopParent->addChildLoop(NewOuter);
1183 removeChildLoop(NewInner, NewOuter);
1184 LI->changeTopLevelLoop(NewInner, NewOuter);
1186 while (!NewOuter->empty())
1187 NewInner->addChildLoop(NewOuter->removeChildLoop(NewOuter->begin()));
1188 NewOuter->addChildLoop(NewInner);
1190 // BBs from the original inner loop.
1191 SmallVector<BasicBlock *, 8> OrigInnerBBs(NewOuter->blocks());
1193 // Add BBs from the original outer loop to the original inner loop (excluding
1194 // BBs already in inner loop)
1195 for (BasicBlock *BB : NewInner->blocks())
1196 if (LI->getLoopFor(BB) == NewInner)
1197 NewOuter->addBlockEntry(BB);
1199 // Now remove inner loop header and latch from the new inner loop and move
1200 // other BBs (the loop body) to the new inner loop.
1201 BasicBlock *OuterHeader = NewOuter->getHeader();
1202 BasicBlock *OuterLatch = NewOuter->getLoopLatch();
1203 for (BasicBlock *BB : OrigInnerBBs) {
1204 // Nothing will change for BBs in child loops.
1205 if (LI->getLoopFor(BB) != NewOuter)
1207 // Remove the new outer loop header and latch from the new inner loop.
1208 if (BB == OuterHeader || BB == OuterLatch)
1209 NewInner->removeBlockFromLoop(BB);
1211 LI->changeLoopFor(BB, NewInner);
1214 // The preheader of the original outer loop becomes part of the new
1216 NewOuter->addBlockEntry(OrigOuterPreHeader);
1217 LI->changeLoopFor(OrigOuterPreHeader, NewOuter);
1219 // Tell SE that we move the loops around.
1220 SE->forgetLoop(NewOuter);
1221 SE->forgetLoop(NewInner);
1224 bool LoopInterchangeTransform::transform() {
1225 bool Transformed = false;
1226 Instruction *InnerIndexVar;
1228 if (InnerLoop->getSubLoops().empty()) {
1229 BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader();
1230 LLVM_DEBUG(dbgs() << "Calling Split Inner Loop\n");
1231 PHINode *InductionPHI = getInductionVariable(InnerLoop, SE);
1232 if (!InductionPHI) {
1233 LLVM_DEBUG(dbgs() << "Failed to find the point to split loop latch \n");
1237 if (InductionPHI->getIncomingBlock(0) == InnerLoopPreHeader)
1238 InnerIndexVar = dyn_cast<Instruction>(InductionPHI->getIncomingValue(1));
1240 InnerIndexVar = dyn_cast<Instruction>(InductionPHI->getIncomingValue(0));
1242 // Ensure that InductionPHI is the first Phi node.
1243 if (&InductionPHI->getParent()->front() != InductionPHI)
1244 InductionPHI->moveBefore(&InductionPHI->getParent()->front());
1246 // Split at the place were the induction variable is
1247 // incremented/decremented.
1248 // TODO: This splitting logic may not work always. Fix this.
1249 splitInnerLoopLatch(InnerIndexVar);
1250 LLVM_DEBUG(dbgs() << "splitInnerLoopLatch done\n");
1252 // Splits the inner loops phi nodes out into a separate basic block.
1253 BasicBlock *InnerLoopHeader = InnerLoop->getHeader();
1254 SplitBlock(InnerLoopHeader, InnerLoopHeader->getFirstNonPHI(), DT, LI);
1255 LLVM_DEBUG(dbgs() << "splitting InnerLoopHeader done\n");
1258 Transformed |= adjustLoopLinks();
1260 LLVM_DEBUG(dbgs() << "adjustLoopLinks failed\n");
1267 void LoopInterchangeTransform::splitInnerLoopLatch(Instruction *Inc) {
1268 BasicBlock *InnerLoopLatch = InnerLoop->getLoopLatch();
1269 BasicBlock *InnerLoopLatchPred = InnerLoopLatch;
1270 InnerLoopLatch = SplitBlock(InnerLoopLatchPred, Inc, DT, LI);
1273 /// \brief Move all instructions except the terminator from FromBB right before
1275 static void moveBBContents(BasicBlock *FromBB, Instruction *InsertBefore) {
1276 auto &ToList = InsertBefore->getParent()->getInstList();
1277 auto &FromList = FromBB->getInstList();
1279 ToList.splice(InsertBefore->getIterator(), FromList, FromList.begin(),
1280 FromBB->getTerminator()->getIterator());
1283 static void updateIncomingBlock(BasicBlock *CurrBlock, BasicBlock *OldPred,
1284 BasicBlock *NewPred) {
1285 for (PHINode &PHI : CurrBlock->phis()) {
1286 unsigned Num = PHI.getNumIncomingValues();
1287 for (unsigned i = 0; i < Num; ++i) {
1288 if (PHI.getIncomingBlock(i) == OldPred)
1289 PHI.setIncomingBlock(i, NewPred);
1294 /// Update BI to jump to NewBB instead of OldBB. Records updates to
1295 /// the dominator tree in DTUpdates, if DT should be preserved.
1296 static void updateSuccessor(BranchInst *BI, BasicBlock *OldBB,
1298 std::vector<DominatorTree::UpdateType> &DTUpdates) {
1299 assert(llvm::count_if(successors(BI),
1300 [OldBB](BasicBlock *BB) { return BB == OldBB; }) < 2 &&
1301 "BI must jump to OldBB at most once.");
1302 for (unsigned i = 0, e = BI->getNumSuccessors(); i < e; ++i) {
1303 if (BI->getSuccessor(i) == OldBB) {
1304 BI->setSuccessor(i, NewBB);
1306 DTUpdates.push_back(
1307 {DominatorTree::UpdateKind::Insert, BI->getParent(), NewBB});
1308 DTUpdates.push_back(
1309 {DominatorTree::UpdateKind::Delete, BI->getParent(), OldBB});
1315 // Move Lcssa PHIs to the right place.
1316 static void moveLCSSAPhis(BasicBlock *InnerExit, BasicBlock *InnerLatch,
1317 BasicBlock *OuterLatch) {
1318 SmallVector<PHINode *, 8> LcssaInnerExit;
1319 for (PHINode &P : InnerExit->phis())
1320 LcssaInnerExit.push_back(&P);
1322 SmallVector<PHINode *, 8> LcssaInnerLatch;
1323 for (PHINode &P : InnerLatch->phis())
1324 LcssaInnerLatch.push_back(&P);
1326 // Lcssa PHIs for values used outside the inner loop are in InnerExit.
1327 // If a PHI node has users outside of InnerExit, it has a use outside the
1328 // interchanged loop and we have to preserve it. We move these to
1329 // InnerLatch, which will become the new exit block for the innermost
1330 // loop after interchanging. For PHIs only used in InnerExit, we can just
1331 // replace them with the incoming value.
1332 for (PHINode *P : LcssaInnerExit) {
1333 bool hasUsersOutside = false;
1334 for (auto UI = P->use_begin(), E = P->use_end(); UI != E;) {
1337 auto *Usr = cast<Instruction>(U.getUser());
1338 if (Usr->getParent() != InnerExit) {
1339 hasUsersOutside = true;
1342 U.set(P->getIncomingValueForBlock(InnerLatch));
1344 if (hasUsersOutside)
1345 P->moveBefore(InnerLatch->getFirstNonPHI());
1347 P->eraseFromParent();
1350 // If the inner loop latch contains LCSSA PHIs, those come from a child loop
1351 // and we have to move them to the new inner latch.
1352 for (PHINode *P : LcssaInnerLatch)
1353 P->moveBefore(InnerExit->getFirstNonPHI());
1355 // Now adjust the incoming blocks for the LCSSA PHIs.
1356 // For PHIs moved from Inner's exit block, we need to replace Inner's latch
1357 // with the new latch.
1358 updateIncomingBlock(InnerLatch, InnerLatch, OuterLatch);
1361 bool LoopInterchangeTransform::adjustLoopBranches() {
1362 LLVM_DEBUG(dbgs() << "adjustLoopBranches called\n");
1363 std::vector<DominatorTree::UpdateType> DTUpdates;
1365 BasicBlock *OuterLoopPreHeader = OuterLoop->getLoopPreheader();
1366 BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader();
1368 assert(OuterLoopPreHeader != OuterLoop->getHeader() &&
1369 InnerLoopPreHeader != InnerLoop->getHeader() && OuterLoopPreHeader &&
1370 InnerLoopPreHeader && "Guaranteed by loop-simplify form");
1371 // Ensure that both preheaders do not contain PHI nodes and have single
1372 // predecessors. This allows us to move them easily. We use
1373 // InsertPreHeaderForLoop to create an 'extra' preheader, if the existing
1374 // preheaders do not satisfy those conditions.
1375 if (isa<PHINode>(OuterLoopPreHeader->begin()) ||
1376 !OuterLoopPreHeader->getUniquePredecessor())
1377 OuterLoopPreHeader = InsertPreheaderForLoop(OuterLoop, DT, LI, true);
1378 if (InnerLoopPreHeader == OuterLoop->getHeader())
1379 InnerLoopPreHeader = InsertPreheaderForLoop(InnerLoop, DT, LI, true);
1381 // Adjust the loop preheader
1382 BasicBlock *InnerLoopHeader = InnerLoop->getHeader();
1383 BasicBlock *OuterLoopHeader = OuterLoop->getHeader();
1384 BasicBlock *InnerLoopLatch = InnerLoop->getLoopLatch();
1385 BasicBlock *OuterLoopLatch = OuterLoop->getLoopLatch();
1386 BasicBlock *OuterLoopPredecessor = OuterLoopPreHeader->getUniquePredecessor();
1387 BasicBlock *InnerLoopLatchPredecessor =
1388 InnerLoopLatch->getUniquePredecessor();
1389 BasicBlock *InnerLoopLatchSuccessor;
1390 BasicBlock *OuterLoopLatchSuccessor;
1392 BranchInst *OuterLoopLatchBI =
1393 dyn_cast<BranchInst>(OuterLoopLatch->getTerminator());
1394 BranchInst *InnerLoopLatchBI =
1395 dyn_cast<BranchInst>(InnerLoopLatch->getTerminator());
1396 BranchInst *OuterLoopHeaderBI =
1397 dyn_cast<BranchInst>(OuterLoopHeader->getTerminator());
1398 BranchInst *InnerLoopHeaderBI =
1399 dyn_cast<BranchInst>(InnerLoopHeader->getTerminator());
1401 if (!OuterLoopPredecessor || !InnerLoopLatchPredecessor ||
1402 !OuterLoopLatchBI || !InnerLoopLatchBI || !OuterLoopHeaderBI ||
1406 BranchInst *InnerLoopLatchPredecessorBI =
1407 dyn_cast<BranchInst>(InnerLoopLatchPredecessor->getTerminator());
1408 BranchInst *OuterLoopPredecessorBI =
1409 dyn_cast<BranchInst>(OuterLoopPredecessor->getTerminator());
1411 if (!OuterLoopPredecessorBI || !InnerLoopLatchPredecessorBI)
1413 BasicBlock *InnerLoopHeaderSuccessor = InnerLoopHeader->getUniqueSuccessor();
1414 if (!InnerLoopHeaderSuccessor)
1417 // Adjust Loop Preheader and headers
1418 updateSuccessor(OuterLoopPredecessorBI, OuterLoopPreHeader,
1419 InnerLoopPreHeader, DTUpdates);
1420 updateSuccessor(OuterLoopHeaderBI, OuterLoopLatch, LoopExit, DTUpdates);
1421 updateSuccessor(OuterLoopHeaderBI, InnerLoopPreHeader,
1422 InnerLoopHeaderSuccessor, DTUpdates);
1424 // Adjust reduction PHI's now that the incoming block has changed.
1425 updateIncomingBlock(InnerLoopHeaderSuccessor, InnerLoopHeader,
1428 updateSuccessor(InnerLoopHeaderBI, InnerLoopHeaderSuccessor,
1429 OuterLoopPreHeader, DTUpdates);
1431 // -------------Adjust loop latches-----------
1432 if (InnerLoopLatchBI->getSuccessor(0) == InnerLoopHeader)
1433 InnerLoopLatchSuccessor = InnerLoopLatchBI->getSuccessor(1);
1435 InnerLoopLatchSuccessor = InnerLoopLatchBI->getSuccessor(0);
1437 updateSuccessor(InnerLoopLatchPredecessorBI, InnerLoopLatch,
1438 InnerLoopLatchSuccessor, DTUpdates);
1441 if (OuterLoopLatchBI->getSuccessor(0) == OuterLoopHeader)
1442 OuterLoopLatchSuccessor = OuterLoopLatchBI->getSuccessor(1);
1444 OuterLoopLatchSuccessor = OuterLoopLatchBI->getSuccessor(0);
1446 updateSuccessor(InnerLoopLatchBI, InnerLoopLatchSuccessor,
1447 OuterLoopLatchSuccessor, DTUpdates);
1448 updateSuccessor(OuterLoopLatchBI, OuterLoopLatchSuccessor, InnerLoopLatch,
1451 DT->applyUpdates(DTUpdates);
1452 restructureLoops(OuterLoop, InnerLoop, InnerLoopPreHeader,
1453 OuterLoopPreHeader);
1455 moveLCSSAPhis(InnerLoopLatchSuccessor, InnerLoopLatch, OuterLoopLatch);
1456 // For PHIs in the exit block of the outer loop, outer's latch has been
1457 // replaced by Inners'.
1458 updateIncomingBlock(OuterLoopLatchSuccessor, OuterLoopLatch, InnerLoopLatch);
1460 // Now update the reduction PHIs in the inner and outer loop headers.
1461 SmallVector<PHINode *, 4> InnerLoopPHIs, OuterLoopPHIs;
1462 for (PHINode &PHI : drop_begin(InnerLoopHeader->phis(), 1))
1463 InnerLoopPHIs.push_back(cast<PHINode>(&PHI));
1464 for (PHINode &PHI : drop_begin(OuterLoopHeader->phis(), 1))
1465 OuterLoopPHIs.push_back(cast<PHINode>(&PHI));
1467 auto &OuterInnerReductions = LIL.getOuterInnerReductions();
1468 (void)OuterInnerReductions;
1470 // Now move the remaining reduction PHIs from outer to inner loop header and
1471 // vice versa. The PHI nodes must be part of a reduction across the inner and
1472 // outer loop and all the remains to do is and updating the incoming blocks.
1473 for (PHINode *PHI : OuterLoopPHIs) {
1474 PHI->moveBefore(InnerLoopHeader->getFirstNonPHI());
1475 assert(OuterInnerReductions.find(PHI) != OuterInnerReductions.end() &&
1476 "Expected a reduction PHI node");
1478 for (PHINode *PHI : InnerLoopPHIs) {
1479 PHI->moveBefore(OuterLoopHeader->getFirstNonPHI());
1480 assert(OuterInnerReductions.find(PHI) != OuterInnerReductions.end() &&
1481 "Expected a reduction PHI node");
1484 // Update the incoming blocks for moved PHI nodes.
1485 updateIncomingBlock(OuterLoopHeader, InnerLoopPreHeader, OuterLoopPreHeader);
1486 updateIncomingBlock(OuterLoopHeader, InnerLoopLatch, OuterLoopLatch);
1487 updateIncomingBlock(InnerLoopHeader, OuterLoopPreHeader, InnerLoopPreHeader);
1488 updateIncomingBlock(InnerLoopHeader, OuterLoopLatch, InnerLoopLatch);
1493 void LoopInterchangeTransform::adjustLoopPreheaders() {
1494 // We have interchanged the preheaders so we need to interchange the data in
1495 // the preheader as well.
1496 // This is because the content of inner preheader was previously executed
1497 // inside the outer loop.
1498 BasicBlock *OuterLoopPreHeader = OuterLoop->getLoopPreheader();
1499 BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader();
1500 BasicBlock *OuterLoopHeader = OuterLoop->getHeader();
1501 BranchInst *InnerTermBI =
1502 cast<BranchInst>(InnerLoopPreHeader->getTerminator());
1504 // These instructions should now be executed inside the loop.
1505 // Move instruction into a new block after outer header.
1506 moveBBContents(InnerLoopPreHeader, OuterLoopHeader->getTerminator());
1507 // These instructions were not executed previously in the loop so move them to
1508 // the older inner loop preheader.
1509 moveBBContents(OuterLoopPreHeader, InnerTermBI);
1512 bool LoopInterchangeTransform::adjustLoopLinks() {
1513 // Adjust all branches in the inner and outer loop.
1514 bool Changed = adjustLoopBranches();
1516 adjustLoopPreheaders();
1520 char LoopInterchange::ID = 0;
1522 INITIALIZE_PASS_BEGIN(LoopInterchange, "loop-interchange",
1523 "Interchanges loops for cache reuse", false, false)
1524 INITIALIZE_PASS_DEPENDENCY(LoopPass)
1525 INITIALIZE_PASS_DEPENDENCY(DependenceAnalysisWrapperPass)
1526 INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass)
1528 INITIALIZE_PASS_END(LoopInterchange, "loop-interchange",
1529 "Interchanges loops for cache reuse", false, false)
1531 Pass *llvm::createLoopInterchangePass() { return new LoopInterchange(); }