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/SmallVector.h"
17 #include "llvm/Analysis/AliasAnalysis.h"
18 #include "llvm/Analysis/AssumptionCache.h"
19 #include "llvm/Analysis/BlockFrequencyInfo.h"
20 #include "llvm/Analysis/CodeMetrics.h"
21 #include "llvm/Analysis/DependenceAnalysis.h"
22 #include "llvm/Analysis/LoopInfo.h"
23 #include "llvm/Analysis/LoopIterator.h"
24 #include "llvm/Analysis/LoopPass.h"
25 #include "llvm/Analysis/ScalarEvolution.h"
26 #include "llvm/Analysis/ScalarEvolutionExpander.h"
27 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
28 #include "llvm/Analysis/TargetTransformInfo.h"
29 #include "llvm/Analysis/ValueTracking.h"
30 #include "llvm/IR/Dominators.h"
31 #include "llvm/IR/Function.h"
32 #include "llvm/IR/IRBuilder.h"
33 #include "llvm/IR/InstIterator.h"
34 #include "llvm/IR/IntrinsicInst.h"
35 #include "llvm/IR/Module.h"
36 #include "llvm/Pass.h"
37 #include "llvm/Support/Debug.h"
38 #include "llvm/Support/raw_ostream.h"
39 #include "llvm/Transforms/Scalar.h"
40 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
41 #include "llvm/Transforms/Utils/LoopUtils.h"
45 #define DEBUG_TYPE "loop-interchange"
47 static cl::opt<int> LoopInterchangeCostThreshold(
48 "loop-interchange-threshold", cl::init(0), cl::Hidden,
49 cl::desc("Interchange if you gain more than this number"));
53 typedef SmallVector<Loop *, 8> LoopVector;
55 // TODO: Check if we can use a sparse matrix here.
56 typedef std::vector<std::vector<char>> CharMatrix;
58 // Maximum number of dependencies that can be handled in the dependency matrix.
59 static const unsigned MaxMemInstrCount = 100;
61 // Maximum loop depth supported.
62 static const unsigned MaxLoopNestDepth = 10;
64 struct LoopInterchange;
66 #ifdef DUMP_DEP_MATRICIES
67 void printDepMatrix(CharMatrix &DepMatrix) {
68 for (auto I = DepMatrix.begin(), E = DepMatrix.end(); I != E; ++I) {
69 std::vector<char> Vec = *I;
70 for (auto II = Vec.begin(), EE = Vec.end(); II != EE; ++II)
71 DEBUG(dbgs() << *II << " ");
72 DEBUG(dbgs() << "\n");
77 static bool populateDependencyMatrix(CharMatrix &DepMatrix, unsigned Level,
78 Loop *L, DependenceInfo *DI) {
79 typedef SmallVector<Value *, 16> ValueVector;
83 for (Loop::block_iterator BB = L->block_begin(), BE = L->block_end();
85 // Scan the BB and collect legal loads and stores.
86 for (BasicBlock::iterator I = (*BB)->begin(), E = (*BB)->end(); I != E;
88 if (!isa<Instruction>(I))
90 if (LoadInst *Ld = dyn_cast<LoadInst>(I)) {
93 MemInstr.push_back(&*I);
94 } else if (StoreInst *St = dyn_cast<StoreInst>(I)) {
97 MemInstr.push_back(&*I);
102 DEBUG(dbgs() << "Found " << MemInstr.size()
103 << " Loads and Stores to analyze\n");
105 ValueVector::iterator I, IE, J, JE;
107 for (I = MemInstr.begin(), IE = MemInstr.end(); I != IE; ++I) {
108 for (J = I, JE = MemInstr.end(); J != JE; ++J) {
109 std::vector<char> Dep;
110 Instruction *Src = cast<Instruction>(*I);
111 Instruction *Dst = cast<Instruction>(*J);
114 // Ignore Input dependencies.
115 if (isa<LoadInst>(Src) && isa<LoadInst>(Dst))
117 // Track Output, Flow, and Anti dependencies.
118 if (auto D = DI->depends(Src, Dst, true)) {
119 assert(D->isOrdered() && "Expected an output, flow or anti dep.");
120 DEBUG(StringRef DepType =
121 D->isFlow() ? "flow" : D->isAnti() ? "anti" : "output";
122 dbgs() << "Found " << DepType
123 << " dependency between Src and Dst\n"
124 << " Src:" << *Src << "\n Dst:" << *Dst << '\n');
125 unsigned Levels = D->getLevels();
127 for (unsigned II = 1; II <= Levels; ++II) {
128 const SCEV *Distance = D->getDistance(II);
129 const SCEVConstant *SCEVConst =
130 dyn_cast_or_null<SCEVConstant>(Distance);
132 const ConstantInt *CI = SCEVConst->getValue();
133 if (CI->isNegative())
135 else if (CI->isZero())
139 Dep.push_back(Direction);
140 } else if (D->isScalar(II)) {
142 Dep.push_back(Direction);
144 unsigned Dir = D->getDirection(II);
145 if (Dir == Dependence::DVEntry::LT ||
146 Dir == Dependence::DVEntry::LE)
148 else if (Dir == Dependence::DVEntry::GT ||
149 Dir == Dependence::DVEntry::GE)
151 else if (Dir == Dependence::DVEntry::EQ)
155 Dep.push_back(Direction);
158 while (Dep.size() != Level) {
162 DepMatrix.push_back(Dep);
163 if (DepMatrix.size() > MaxMemInstrCount) {
164 DEBUG(dbgs() << "Cannot handle more than " << MaxMemInstrCount
165 << " dependencies inside loop\n");
172 // We don't have a DepMatrix to check legality return false.
173 if (DepMatrix.size() == 0)
178 // A loop is moved from index 'from' to an index 'to'. Update the Dependence
179 // matrix by exchanging the two columns.
180 static void interChangeDependencies(CharMatrix &DepMatrix, unsigned FromIndx,
182 unsigned numRows = DepMatrix.size();
183 for (unsigned i = 0; i < numRows; ++i) {
184 char TmpVal = DepMatrix[i][ToIndx];
185 DepMatrix[i][ToIndx] = DepMatrix[i][FromIndx];
186 DepMatrix[i][FromIndx] = TmpVal;
190 // Checks if outermost non '=','S'or'I' dependence in the dependence matrix is
192 static bool isOuterMostDepPositive(CharMatrix &DepMatrix, unsigned Row,
194 for (unsigned i = 0; i <= Column; ++i) {
195 if (DepMatrix[Row][i] == '<')
197 if (DepMatrix[Row][i] == '>')
200 // All dependencies were '=','S' or 'I'
204 // Checks if no dependence exist in the dependency matrix in Row before Column.
205 static bool containsNoDependence(CharMatrix &DepMatrix, unsigned Row,
207 for (unsigned i = 0; i < Column; ++i) {
208 if (DepMatrix[Row][i] != '=' && DepMatrix[Row][i] != 'S' &&
209 DepMatrix[Row][i] != 'I')
215 static bool validDepInterchange(CharMatrix &DepMatrix, unsigned Row,
216 unsigned OuterLoopId, char InnerDep,
219 if (isOuterMostDepPositive(DepMatrix, Row, OuterLoopId))
222 if (InnerDep == OuterDep)
225 // It is legal to interchange if and only if after interchange no row has a
226 // '>' direction as the leftmost non-'='.
228 if (InnerDep == '=' || InnerDep == 'S' || InnerDep == 'I')
234 if (InnerDep == '>') {
235 // If OuterLoopId represents outermost loop then interchanging will make the
236 // 1st dependency as '>'
237 if (OuterLoopId == 0)
240 // If all dependencies before OuterloopId are '=','S'or 'I'. Then
241 // interchanging will result in this row having an outermost non '='
243 if (!containsNoDependence(DepMatrix, Row, OuterLoopId))
250 // Checks if it is legal to interchange 2 loops.
251 // [Theorem] A permutation of the loops in a perfect nest is legal if and only
252 // if the direction matrix, after the same permutation is applied to its
253 // columns, has no ">" direction as the leftmost non-"=" direction in any row.
254 static bool isLegalToInterChangeLoops(CharMatrix &DepMatrix,
255 unsigned InnerLoopId,
256 unsigned OuterLoopId) {
258 unsigned NumRows = DepMatrix.size();
259 // For each row check if it is valid to interchange.
260 for (unsigned Row = 0; Row < NumRows; ++Row) {
261 char InnerDep = DepMatrix[Row][InnerLoopId];
262 char OuterDep = DepMatrix[Row][OuterLoopId];
263 if (InnerDep == '*' || OuterDep == '*')
265 if (!validDepInterchange(DepMatrix, Row, OuterLoopId, InnerDep, OuterDep))
271 static void populateWorklist(Loop &L, SmallVector<LoopVector, 8> &V) {
273 DEBUG(dbgs() << "Calling populateWorklist on Func: "
274 << L.getHeader()->getParent()->getName() << " Loop: %"
275 << L.getHeader()->getName() << '\n');
277 Loop *CurrentLoop = &L;
278 const std::vector<Loop *> *Vec = &CurrentLoop->getSubLoops();
279 while (!Vec->empty()) {
280 // The current loop has multiple subloops in it hence it is not tightly
282 // Discard all loops above it added into Worklist.
283 if (Vec->size() != 1) {
287 LoopList.push_back(CurrentLoop);
288 CurrentLoop = Vec->front();
289 Vec = &CurrentLoop->getSubLoops();
291 LoopList.push_back(CurrentLoop);
292 V.push_back(std::move(LoopList));
295 static PHINode *getInductionVariable(Loop *L, ScalarEvolution *SE) {
296 PHINode *InnerIndexVar = L->getCanonicalInductionVariable();
298 return InnerIndexVar;
299 if (L->getLoopLatch() == nullptr || L->getLoopPredecessor() == nullptr)
301 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
302 PHINode *PhiVar = cast<PHINode>(I);
303 Type *PhiTy = PhiVar->getType();
304 if (!PhiTy->isIntegerTy() && !PhiTy->isFloatingPointTy() &&
305 !PhiTy->isPointerTy())
307 const SCEVAddRecExpr *AddRec =
308 dyn_cast<SCEVAddRecExpr>(SE->getSCEV(PhiVar));
309 if (!AddRec || !AddRec->isAffine())
311 const SCEV *Step = AddRec->getStepRecurrence(*SE);
312 if (!isa<SCEVConstant>(Step))
314 // Found the induction variable.
315 // FIXME: Handle loops with more than one induction variable. Note that,
316 // currently, legality makes sure we have only one induction variable.
322 /// LoopInterchangeLegality checks if it is legal to interchange the loop.
323 class LoopInterchangeLegality {
325 LoopInterchangeLegality(Loop *Outer, Loop *Inner, ScalarEvolution *SE,
326 LoopInfo *LI, DominatorTree *DT, bool PreserveLCSSA)
327 : OuterLoop(Outer), InnerLoop(Inner), SE(SE), LI(LI), DT(DT),
328 PreserveLCSSA(PreserveLCSSA), InnerLoopHasReduction(false) {}
330 /// Check if the loops can be interchanged.
331 bool canInterchangeLoops(unsigned InnerLoopId, unsigned OuterLoopId,
332 CharMatrix &DepMatrix);
333 /// Check if the loop structure is understood. We do not handle triangular
335 bool isLoopStructureUnderstood(PHINode *InnerInductionVar);
337 bool currentLimitations();
339 bool hasInnerLoopReduction() { return InnerLoopHasReduction; }
342 bool tightlyNested(Loop *Outer, Loop *Inner);
343 bool containsUnsafeInstructionsInHeader(BasicBlock *BB);
344 bool areAllUsesReductions(Instruction *Ins, Loop *L);
345 bool containsUnsafeInstructionsInLatch(BasicBlock *BB);
346 bool findInductionAndReductions(Loop *L,
347 SmallVector<PHINode *, 8> &Inductions,
348 SmallVector<PHINode *, 8> &Reductions);
357 bool InnerLoopHasReduction;
360 /// LoopInterchangeProfitability checks if it is profitable to interchange the
362 class LoopInterchangeProfitability {
364 LoopInterchangeProfitability(Loop *Outer, Loop *Inner, ScalarEvolution *SE)
365 : OuterLoop(Outer), InnerLoop(Inner), SE(SE) {}
367 /// Check if the loop interchange is profitable.
368 bool isProfitable(unsigned InnerLoopId, unsigned OuterLoopId,
369 CharMatrix &DepMatrix);
372 int getInstrOrderCost();
381 /// LoopInterchangeTransform interchanges the loop.
382 class LoopInterchangeTransform {
384 LoopInterchangeTransform(Loop *Outer, Loop *Inner, ScalarEvolution *SE,
385 LoopInfo *LI, DominatorTree *DT,
386 BasicBlock *LoopNestExit,
387 bool InnerLoopContainsReductions)
388 : OuterLoop(Outer), InnerLoop(Inner), SE(SE), LI(LI), DT(DT),
389 LoopExit(LoopNestExit),
390 InnerLoopHasReduction(InnerLoopContainsReductions) {}
392 /// Interchange OuterLoop and InnerLoop.
394 void restructureLoops(Loop *InnerLoop, Loop *OuterLoop);
395 void removeChildLoop(Loop *OuterLoop, Loop *InnerLoop);
398 void splitInnerLoopLatch(Instruction *);
399 void splitInnerLoopHeader();
400 bool adjustLoopLinks();
401 void adjustLoopPreheaders();
402 bool adjustLoopBranches();
403 void updateIncomingBlock(BasicBlock *CurrBlock, BasicBlock *OldPred,
404 BasicBlock *NewPred);
413 BasicBlock *LoopExit;
414 bool InnerLoopHasReduction;
417 // Main LoopInterchange Pass.
418 struct LoopInterchange : public FunctionPass {
426 : FunctionPass(ID), SE(nullptr), LI(nullptr), DI(nullptr), DT(nullptr) {
427 initializeLoopInterchangePass(*PassRegistry::getPassRegistry());
430 void getAnalysisUsage(AnalysisUsage &AU) const override {
431 AU.addRequired<ScalarEvolutionWrapperPass>();
432 AU.addRequired<AAResultsWrapperPass>();
433 AU.addRequired<DominatorTreeWrapperPass>();
434 AU.addRequired<LoopInfoWrapperPass>();
435 AU.addRequired<DependenceAnalysisWrapperPass>();
436 AU.addRequiredID(LoopSimplifyID);
437 AU.addRequiredID(LCSSAID);
440 bool runOnFunction(Function &F) override {
444 SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
445 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
446 DI = &getAnalysis<DependenceAnalysisWrapperPass>().getDI();
447 auto *DTWP = getAnalysisIfAvailable<DominatorTreeWrapperPass>();
448 DT = DTWP ? &DTWP->getDomTree() : nullptr;
449 PreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
451 // Build up a worklist of loop pairs to analyze.
452 SmallVector<LoopVector, 8> Worklist;
455 populateWorklist(*L, Worklist);
457 DEBUG(dbgs() << "Worklist size = " << Worklist.size() << "\n");
459 while (!Worklist.empty()) {
460 LoopVector LoopList = Worklist.pop_back_val();
461 Changed = processLoopList(LoopList, F);
466 bool isComputableLoopNest(LoopVector LoopList) {
467 for (Loop *L : LoopList) {
468 const SCEV *ExitCountOuter = SE->getBackedgeTakenCount(L);
469 if (ExitCountOuter == SE->getCouldNotCompute()) {
470 DEBUG(dbgs() << "Couldn't compute backedge count\n");
473 if (L->getNumBackEdges() != 1) {
474 DEBUG(dbgs() << "NumBackEdges is not equal to 1\n");
477 if (!L->getExitingBlock()) {
478 DEBUG(dbgs() << "Loop doesn't have unique exit block\n");
485 unsigned selectLoopForInterchange(const LoopVector &LoopList) {
486 // TODO: Add a better heuristic to select the loop to be interchanged based
487 // on the dependence matrix. Currently we select the innermost loop.
488 return LoopList.size() - 1;
491 bool processLoopList(LoopVector LoopList, Function &F) {
493 bool Changed = false;
494 unsigned LoopNestDepth = LoopList.size();
495 if (LoopNestDepth < 2) {
496 DEBUG(dbgs() << "Loop doesn't contain minimum nesting level.\n");
499 if (LoopNestDepth > MaxLoopNestDepth) {
500 DEBUG(dbgs() << "Cannot handle loops of depth greater than "
501 << MaxLoopNestDepth << "\n");
504 if (!isComputableLoopNest(LoopList)) {
505 DEBUG(dbgs() << "Not valid loop candidate for interchange\n");
509 DEBUG(dbgs() << "Processing LoopList of size = " << LoopNestDepth << "\n");
511 CharMatrix DependencyMatrix;
512 Loop *OuterMostLoop = *(LoopList.begin());
513 if (!populateDependencyMatrix(DependencyMatrix, LoopNestDepth,
514 OuterMostLoop, DI)) {
515 DEBUG(dbgs() << "Populating dependency matrix failed\n");
518 #ifdef DUMP_DEP_MATRICIES
519 DEBUG(dbgs() << "Dependence before interchange\n");
520 printDepMatrix(DependencyMatrix);
523 BasicBlock *OuterMostLoopLatch = OuterMostLoop->getLoopLatch();
524 BranchInst *OuterMostLoopLatchBI =
525 dyn_cast<BranchInst>(OuterMostLoopLatch->getTerminator());
526 if (!OuterMostLoopLatchBI)
529 // Since we currently do not handle LCSSA PHI's any failure in loop
530 // condition will now branch to LoopNestExit.
531 // TODO: This should be removed once we handle LCSSA PHI nodes.
533 // Get the Outermost loop exit.
534 BasicBlock *LoopNestExit;
535 if (OuterMostLoopLatchBI->getSuccessor(0) == OuterMostLoop->getHeader())
536 LoopNestExit = OuterMostLoopLatchBI->getSuccessor(1);
538 LoopNestExit = OuterMostLoopLatchBI->getSuccessor(0);
540 if (isa<PHINode>(LoopNestExit->begin())) {
541 DEBUG(dbgs() << "PHI Nodes in loop nest exit is not handled for now "
542 "since on failure all loops branch to loop nest exit.\n");
546 unsigned SelecLoopId = selectLoopForInterchange(LoopList);
547 // Move the selected loop outwards to the best possible position.
548 for (unsigned i = SelecLoopId; i > 0; i--) {
550 processLoop(LoopList, i, i - 1, LoopNestExit, DependencyMatrix);
553 // Loops interchanged reflect the same in LoopList
554 std::swap(LoopList[i - 1], LoopList[i]);
556 // Update the DependencyMatrix
557 interChangeDependencies(DependencyMatrix, i, i - 1);
559 #ifdef DUMP_DEP_MATRICIES
560 DEBUG(dbgs() << "Dependence after interchange\n");
561 printDepMatrix(DependencyMatrix);
563 Changed |= Interchanged;
568 bool processLoop(LoopVector LoopList, unsigned InnerLoopId,
569 unsigned OuterLoopId, BasicBlock *LoopNestExit,
570 std::vector<std::vector<char>> &DependencyMatrix) {
572 DEBUG(dbgs() << "Processing Inner Loop Id = " << InnerLoopId
573 << " and OuterLoopId = " << OuterLoopId << "\n");
574 Loop *InnerLoop = LoopList[InnerLoopId];
575 Loop *OuterLoop = LoopList[OuterLoopId];
577 LoopInterchangeLegality LIL(OuterLoop, InnerLoop, SE, LI, DT,
579 if (!LIL.canInterchangeLoops(InnerLoopId, OuterLoopId, DependencyMatrix)) {
580 DEBUG(dbgs() << "Not interchanging Loops. Cannot prove legality\n");
583 DEBUG(dbgs() << "Loops are legal to interchange\n");
584 LoopInterchangeProfitability LIP(OuterLoop, InnerLoop, SE);
585 if (!LIP.isProfitable(InnerLoopId, OuterLoopId, DependencyMatrix)) {
586 DEBUG(dbgs() << "Interchanging loops not profitable\n");
590 LoopInterchangeTransform LIT(OuterLoop, InnerLoop, SE, LI, DT,
591 LoopNestExit, LIL.hasInnerLoopReduction());
593 DEBUG(dbgs() << "Loops interchanged\n");
598 } // end of namespace
599 bool LoopInterchangeLegality::areAllUsesReductions(Instruction *Ins, Loop *L) {
600 return none_of(Ins->users(), [=](User *U) -> bool {
601 auto *UserIns = dyn_cast<PHINode>(U);
602 RecurrenceDescriptor RD;
603 return !UserIns || !RecurrenceDescriptor::isReductionPHI(UserIns, L, RD);
607 bool LoopInterchangeLegality::containsUnsafeInstructionsInHeader(
609 for (auto I = BB->begin(), E = BB->end(); I != E; ++I) {
610 // Load corresponding to reduction PHI's are safe while concluding if
612 if (LoadInst *L = dyn_cast<LoadInst>(I)) {
613 if (!areAllUsesReductions(L, InnerLoop))
615 } else if (I->mayHaveSideEffects() || I->mayReadFromMemory())
621 bool LoopInterchangeLegality::containsUnsafeInstructionsInLatch(
623 for (auto I = BB->begin(), E = BB->end(); I != E; ++I) {
624 // Stores corresponding to reductions are safe while concluding if tightly
626 if (StoreInst *L = dyn_cast<StoreInst>(I)) {
627 if (!isa<PHINode>(L->getOperand(0)))
629 } else if (I->mayHaveSideEffects() || I->mayReadFromMemory())
635 bool LoopInterchangeLegality::tightlyNested(Loop *OuterLoop, Loop *InnerLoop) {
636 BasicBlock *OuterLoopHeader = OuterLoop->getHeader();
637 BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader();
638 BasicBlock *OuterLoopLatch = OuterLoop->getLoopLatch();
640 DEBUG(dbgs() << "Checking if loops are tightly nested\n");
642 // A perfectly nested loop will not have any branch in between the outer and
643 // inner block i.e. outer header will branch to either inner preheader and
645 BranchInst *OuterLoopHeaderBI =
646 dyn_cast<BranchInst>(OuterLoopHeader->getTerminator());
647 if (!OuterLoopHeaderBI)
650 for (unsigned i = 0, e = OuterLoopHeaderBI->getNumSuccessors(); i < e; ++i) {
651 if (OuterLoopHeaderBI->getSuccessor(i) != InnerLoopPreHeader &&
652 OuterLoopHeaderBI->getSuccessor(i) != OuterLoopLatch)
656 DEBUG(dbgs() << "Checking instructions in Loop header and Loop latch\n");
657 // We do not have any basic block in between now make sure the outer header
658 // and outer loop latch doesn't contain any unsafe instructions.
659 if (containsUnsafeInstructionsInHeader(OuterLoopHeader) ||
660 containsUnsafeInstructionsInLatch(OuterLoopLatch))
663 DEBUG(dbgs() << "Loops are perfectly nested\n");
664 // We have a perfect loop nest.
669 bool LoopInterchangeLegality::isLoopStructureUnderstood(
670 PHINode *InnerInduction) {
672 unsigned Num = InnerInduction->getNumOperands();
673 BasicBlock *InnerLoopPreheader = InnerLoop->getLoopPreheader();
674 for (unsigned i = 0; i < Num; ++i) {
675 Value *Val = InnerInduction->getOperand(i);
676 if (isa<Constant>(Val))
678 Instruction *I = dyn_cast<Instruction>(Val);
681 // TODO: Handle triangular loops.
682 // e.g. for(int i=0;i<N;i++)
683 // for(int j=i;j<N;j++)
684 unsigned IncomBlockIndx = PHINode::getIncomingValueNumForOperand(i);
685 if (InnerInduction->getIncomingBlock(IncomBlockIndx) ==
686 InnerLoopPreheader &&
687 !OuterLoop->isLoopInvariant(I)) {
694 bool LoopInterchangeLegality::findInductionAndReductions(
695 Loop *L, SmallVector<PHINode *, 8> &Inductions,
696 SmallVector<PHINode *, 8> &Reductions) {
697 if (!L->getLoopLatch() || !L->getLoopPredecessor())
699 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
700 RecurrenceDescriptor RD;
701 InductionDescriptor ID;
702 PHINode *PHI = cast<PHINode>(I);
703 if (InductionDescriptor::isInductionPHI(PHI, L, SE, ID))
704 Inductions.push_back(PHI);
705 else if (RecurrenceDescriptor::isReductionPHI(PHI, L, RD))
706 Reductions.push_back(PHI);
709 dbgs() << "Failed to recognize PHI as an induction or reduction.\n");
716 static bool containsSafePHI(BasicBlock *Block, bool isOuterLoopExitBlock) {
717 for (auto I = Block->begin(); isa<PHINode>(I); ++I) {
718 PHINode *PHI = cast<PHINode>(I);
719 // Reduction lcssa phi will have only 1 incoming block that from loop latch.
720 if (PHI->getNumIncomingValues() > 1)
722 Instruction *Ins = dyn_cast<Instruction>(PHI->getIncomingValue(0));
725 // Incoming value for lcssa phi's in outer loop exit can only be inner loop
726 // exits lcssa phi else it would not be tightly nested.
727 if (!isa<PHINode>(Ins) && isOuterLoopExitBlock)
733 static BasicBlock *getLoopLatchExitBlock(BasicBlock *LatchBlock,
734 BasicBlock *LoopHeader) {
735 if (BranchInst *BI = dyn_cast<BranchInst>(LatchBlock->getTerminator())) {
736 unsigned Num = BI->getNumSuccessors();
738 for (unsigned i = 0; i < Num; ++i) {
739 if (BI->getSuccessor(i) == LoopHeader)
741 return BI->getSuccessor(i);
747 // This function indicates the current limitations in the transform as a result
748 // of which we do not proceed.
749 bool LoopInterchangeLegality::currentLimitations() {
751 BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader();
752 BasicBlock *InnerLoopHeader = InnerLoop->getHeader();
753 BasicBlock *InnerLoopLatch = InnerLoop->getLoopLatch();
754 BasicBlock *OuterLoopLatch = OuterLoop->getLoopLatch();
755 BasicBlock *OuterLoopHeader = OuterLoop->getHeader();
757 PHINode *InnerInductionVar;
758 SmallVector<PHINode *, 8> Inductions;
759 SmallVector<PHINode *, 8> Reductions;
760 if (!findInductionAndReductions(InnerLoop, Inductions, Reductions))
763 // TODO: Currently we handle only loops with 1 induction variable.
764 if (Inductions.size() != 1) {
765 DEBUG(dbgs() << "We currently only support loops with 1 induction variable."
766 << "Failed to interchange due to current limitation\n");
769 if (Reductions.size() > 0)
770 InnerLoopHasReduction = true;
772 InnerInductionVar = Inductions.pop_back_val();
774 if (!findInductionAndReductions(OuterLoop, Inductions, Reductions))
777 // Outer loop cannot have reduction because then loops will not be tightly
779 if (!Reductions.empty())
781 // TODO: Currently we handle only loops with 1 induction variable.
782 if (Inductions.size() != 1)
785 // TODO: Triangular loops are not handled for now.
786 if (!isLoopStructureUnderstood(InnerInductionVar)) {
787 DEBUG(dbgs() << "Loop structure not understood by pass\n");
791 // TODO: We only handle LCSSA PHI's corresponding to reduction for now.
792 BasicBlock *LoopExitBlock =
793 getLoopLatchExitBlock(OuterLoopLatch, OuterLoopHeader);
794 if (!LoopExitBlock || !containsSafePHI(LoopExitBlock, true))
797 LoopExitBlock = getLoopLatchExitBlock(InnerLoopLatch, InnerLoopHeader);
798 if (!LoopExitBlock || !containsSafePHI(LoopExitBlock, false))
801 // TODO: Current limitation: Since we split the inner loop latch at the point
802 // were induction variable is incremented (induction.next); We cannot have
803 // more than 1 user of induction.next since it would result in broken code
806 // for(i=0;i<N;i++) {
807 // for(j = 0;j<M;j++) {
808 // A[j+1][i+2] = A[j][i]+k;
811 Instruction *InnerIndexVarInc = nullptr;
812 if (InnerInductionVar->getIncomingBlock(0) == InnerLoopPreHeader)
814 dyn_cast<Instruction>(InnerInductionVar->getIncomingValue(1));
817 dyn_cast<Instruction>(InnerInductionVar->getIncomingValue(0));
819 if (!InnerIndexVarInc)
822 // Since we split the inner loop latch on this induction variable. Make sure
823 // we do not have any instruction between the induction variable and branch
826 bool FoundInduction = false;
827 for (const Instruction &I : reverse(*InnerLoopLatch)) {
828 if (isa<BranchInst>(I) || isa<CmpInst>(I) || isa<TruncInst>(I))
830 // We found an instruction. If this is not induction variable then it is not
831 // safe to split this loop latch.
832 if (!I.isIdenticalTo(InnerIndexVarInc))
835 FoundInduction = true;
838 // The loop latch ended and we didn't find the induction variable return as
839 // current limitation.
846 bool LoopInterchangeLegality::canInterchangeLoops(unsigned InnerLoopId,
847 unsigned OuterLoopId,
848 CharMatrix &DepMatrix) {
850 if (!isLegalToInterChangeLoops(DepMatrix, InnerLoopId, OuterLoopId)) {
851 DEBUG(dbgs() << "Failed interchange InnerLoopId = " << InnerLoopId
852 << " and OuterLoopId = " << OuterLoopId
853 << " due to dependence\n");
857 // Create unique Preheaders if we already do not have one.
858 BasicBlock *OuterLoopPreHeader = OuterLoop->getLoopPreheader();
859 BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader();
861 // Create a unique outer preheader -
862 // 1) If OuterLoop preheader is not present.
863 // 2) If OuterLoop Preheader is same as OuterLoop Header
864 // 3) If OuterLoop Preheader is same as Header of the previous loop.
865 // 4) If OuterLoop Preheader is Entry node.
866 if (!OuterLoopPreHeader || OuterLoopPreHeader == OuterLoop->getHeader() ||
867 isa<PHINode>(OuterLoopPreHeader->begin()) ||
868 !OuterLoopPreHeader->getUniquePredecessor()) {
870 InsertPreheaderForLoop(OuterLoop, DT, LI, PreserveLCSSA);
873 if (!InnerLoopPreHeader || InnerLoopPreHeader == InnerLoop->getHeader() ||
874 InnerLoopPreHeader == OuterLoop->getHeader()) {
876 InsertPreheaderForLoop(InnerLoop, DT, LI, PreserveLCSSA);
879 // TODO: The loops could not be interchanged due to current limitations in the
881 if (currentLimitations()) {
882 DEBUG(dbgs() << "Not legal because of current transform limitation\n");
886 // Check if the loops are tightly nested.
887 if (!tightlyNested(OuterLoop, InnerLoop)) {
888 DEBUG(dbgs() << "Loops not tightly nested\n");
895 int LoopInterchangeProfitability::getInstrOrderCost() {
896 unsigned GoodOrder, BadOrder;
897 BadOrder = GoodOrder = 0;
898 for (auto BI = InnerLoop->block_begin(), BE = InnerLoop->block_end();
900 for (Instruction &Ins : **BI) {
901 if (const GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(&Ins)) {
902 unsigned NumOp = GEP->getNumOperands();
903 bool FoundInnerInduction = false;
904 bool FoundOuterInduction = false;
905 for (unsigned i = 0; i < NumOp; ++i) {
906 const SCEV *OperandVal = SE->getSCEV(GEP->getOperand(i));
907 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(OperandVal);
911 // If we find the inner induction after an outer induction e.g.
912 // for(int i=0;i<N;i++)
913 // for(int j=0;j<N;j++)
914 // A[i][j] = A[i-1][j-1]+k;
915 // then it is a good order.
916 if (AR->getLoop() == InnerLoop) {
917 // We found an InnerLoop induction after OuterLoop induction. It is
919 FoundInnerInduction = true;
920 if (FoundOuterInduction) {
925 // If we find the outer induction after an inner induction e.g.
926 // for(int i=0;i<N;i++)
927 // for(int j=0;j<N;j++)
928 // A[j][i] = A[j-1][i-1]+k;
929 // then it is a bad order.
930 if (AR->getLoop() == OuterLoop) {
931 // We found an OuterLoop induction after InnerLoop induction. It is
933 FoundOuterInduction = true;
934 if (FoundInnerInduction) {
943 return GoodOrder - BadOrder;
946 static bool isProfitableForVectorization(unsigned InnerLoopId,
947 unsigned OuterLoopId,
948 CharMatrix &DepMatrix) {
949 // TODO: Improve this heuristic to catch more cases.
950 // If the inner loop is loop independent or doesn't carry any dependency it is
951 // profitable to move this to outer position.
952 unsigned Row = DepMatrix.size();
953 for (unsigned i = 0; i < Row; ++i) {
954 if (DepMatrix[i][InnerLoopId] != 'S' && DepMatrix[i][InnerLoopId] != 'I')
956 // TODO: We need to improve this heuristic.
957 if (DepMatrix[i][OuterLoopId] != '=')
960 // If outer loop has dependence and inner loop is loop independent then it is
961 // profitable to interchange to enable parallelism.
965 bool LoopInterchangeProfitability::isProfitable(unsigned InnerLoopId,
966 unsigned OuterLoopId,
967 CharMatrix &DepMatrix) {
969 // TODO: Add better profitability checks.
971 // 1) Construct dependency matrix and move the one with no loop carried dep
972 // inside to enable vectorization.
974 // This is rough cost estimation algorithm. It counts the good and bad order
975 // of induction variables in the instruction and allows reordering if number
976 // of bad orders is more than good.
977 int Cost = getInstrOrderCost();
978 DEBUG(dbgs() << "Cost = " << Cost << "\n");
979 if (Cost < -LoopInterchangeCostThreshold)
982 // It is not profitable as per current cache profitability model. But check if
983 // we can move this loop outside to improve parallelism.
985 isProfitableForVectorization(InnerLoopId, OuterLoopId, DepMatrix);
989 void LoopInterchangeTransform::removeChildLoop(Loop *OuterLoop,
991 for (Loop::iterator I = OuterLoop->begin(), E = OuterLoop->end(); I != E;
993 if (*I == InnerLoop) {
994 OuterLoop->removeChildLoop(I);
998 llvm_unreachable("Couldn't find loop");
1001 void LoopInterchangeTransform::restructureLoops(Loop *InnerLoop,
1003 Loop *OuterLoopParent = OuterLoop->getParentLoop();
1004 if (OuterLoopParent) {
1005 // Remove the loop from its parent loop.
1006 removeChildLoop(OuterLoopParent, OuterLoop);
1007 removeChildLoop(OuterLoop, InnerLoop);
1008 OuterLoopParent->addChildLoop(InnerLoop);
1010 removeChildLoop(OuterLoop, InnerLoop);
1011 LI->changeTopLevelLoop(OuterLoop, InnerLoop);
1014 while (!InnerLoop->empty())
1015 OuterLoop->addChildLoop(InnerLoop->removeChildLoop(InnerLoop->begin()));
1017 InnerLoop->addChildLoop(OuterLoop);
1020 bool LoopInterchangeTransform::transform() {
1021 bool Transformed = false;
1022 Instruction *InnerIndexVar;
1024 if (InnerLoop->getSubLoops().size() == 0) {
1025 BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader();
1026 DEBUG(dbgs() << "Calling Split Inner Loop\n");
1027 PHINode *InductionPHI = getInductionVariable(InnerLoop, SE);
1028 if (!InductionPHI) {
1029 DEBUG(dbgs() << "Failed to find the point to split loop latch \n");
1033 if (InductionPHI->getIncomingBlock(0) == InnerLoopPreHeader)
1034 InnerIndexVar = dyn_cast<Instruction>(InductionPHI->getIncomingValue(1));
1036 InnerIndexVar = dyn_cast<Instruction>(InductionPHI->getIncomingValue(0));
1039 // Split at the place were the induction variable is
1040 // incremented/decremented.
1041 // TODO: This splitting logic may not work always. Fix this.
1042 splitInnerLoopLatch(InnerIndexVar);
1043 DEBUG(dbgs() << "splitInnerLoopLatch done\n");
1045 // Splits the inner loops phi nodes out into a separate basic block.
1046 splitInnerLoopHeader();
1047 DEBUG(dbgs() << "splitInnerLoopHeader done\n");
1050 Transformed |= adjustLoopLinks();
1052 DEBUG(dbgs() << "adjustLoopLinks failed\n");
1056 restructureLoops(InnerLoop, OuterLoop);
1060 void LoopInterchangeTransform::splitInnerLoopLatch(Instruction *Inc) {
1061 BasicBlock *InnerLoopLatch = InnerLoop->getLoopLatch();
1062 BasicBlock *InnerLoopLatchPred = InnerLoopLatch;
1063 InnerLoopLatch = SplitBlock(InnerLoopLatchPred, Inc, DT, LI);
1066 void LoopInterchangeTransform::splitInnerLoopHeader() {
1068 // Split the inner loop header out. Here make sure that the reduction PHI's
1069 // stay in the innerloop body.
1070 BasicBlock *InnerLoopHeader = InnerLoop->getHeader();
1071 BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader();
1072 if (InnerLoopHasReduction) {
1073 // FIXME: Check if the induction PHI will always be the first PHI.
1074 BasicBlock *New = InnerLoopHeader->splitBasicBlock(
1075 ++(InnerLoopHeader->begin()), InnerLoopHeader->getName() + ".split");
1077 if (Loop *L = LI->getLoopFor(InnerLoopHeader))
1078 L->addBasicBlockToLoop(New, *LI);
1080 // Adjust Reduction PHI's in the block.
1081 SmallVector<PHINode *, 8> PHIVec;
1082 for (auto I = New->begin(); isa<PHINode>(I); ++I) {
1083 PHINode *PHI = dyn_cast<PHINode>(I);
1084 Value *V = PHI->getIncomingValueForBlock(InnerLoopPreHeader);
1085 PHI->replaceAllUsesWith(V);
1086 PHIVec.push_back((PHI));
1088 for (PHINode *P : PHIVec) {
1089 P->eraseFromParent();
1092 SplitBlock(InnerLoopHeader, InnerLoopHeader->getFirstNonPHI(), DT, LI);
1095 DEBUG(dbgs() << "Output of splitInnerLoopHeader InnerLoopHeaderSucc & "
1096 "InnerLoopHeader\n");
1099 /// \brief Move all instructions except the terminator from FromBB right before
1101 static void moveBBContents(BasicBlock *FromBB, Instruction *InsertBefore) {
1102 auto &ToList = InsertBefore->getParent()->getInstList();
1103 auto &FromList = FromBB->getInstList();
1105 ToList.splice(InsertBefore->getIterator(), FromList, FromList.begin(),
1106 FromBB->getTerminator()->getIterator());
1109 void LoopInterchangeTransform::updateIncomingBlock(BasicBlock *CurrBlock,
1110 BasicBlock *OldPred,
1111 BasicBlock *NewPred) {
1112 for (auto I = CurrBlock->begin(); isa<PHINode>(I); ++I) {
1113 PHINode *PHI = cast<PHINode>(I);
1114 unsigned Num = PHI->getNumIncomingValues();
1115 for (unsigned i = 0; i < Num; ++i) {
1116 if (PHI->getIncomingBlock(i) == OldPred)
1117 PHI->setIncomingBlock(i, NewPred);
1122 bool LoopInterchangeTransform::adjustLoopBranches() {
1124 DEBUG(dbgs() << "adjustLoopBranches called\n");
1125 // Adjust the loop preheader
1126 BasicBlock *InnerLoopHeader = InnerLoop->getHeader();
1127 BasicBlock *OuterLoopHeader = OuterLoop->getHeader();
1128 BasicBlock *InnerLoopLatch = InnerLoop->getLoopLatch();
1129 BasicBlock *OuterLoopLatch = OuterLoop->getLoopLatch();
1130 BasicBlock *OuterLoopPreHeader = OuterLoop->getLoopPreheader();
1131 BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader();
1132 BasicBlock *OuterLoopPredecessor = OuterLoopPreHeader->getUniquePredecessor();
1133 BasicBlock *InnerLoopLatchPredecessor =
1134 InnerLoopLatch->getUniquePredecessor();
1135 BasicBlock *InnerLoopLatchSuccessor;
1136 BasicBlock *OuterLoopLatchSuccessor;
1138 BranchInst *OuterLoopLatchBI =
1139 dyn_cast<BranchInst>(OuterLoopLatch->getTerminator());
1140 BranchInst *InnerLoopLatchBI =
1141 dyn_cast<BranchInst>(InnerLoopLatch->getTerminator());
1142 BranchInst *OuterLoopHeaderBI =
1143 dyn_cast<BranchInst>(OuterLoopHeader->getTerminator());
1144 BranchInst *InnerLoopHeaderBI =
1145 dyn_cast<BranchInst>(InnerLoopHeader->getTerminator());
1147 if (!OuterLoopPredecessor || !InnerLoopLatchPredecessor ||
1148 !OuterLoopLatchBI || !InnerLoopLatchBI || !OuterLoopHeaderBI ||
1152 BranchInst *InnerLoopLatchPredecessorBI =
1153 dyn_cast<BranchInst>(InnerLoopLatchPredecessor->getTerminator());
1154 BranchInst *OuterLoopPredecessorBI =
1155 dyn_cast<BranchInst>(OuterLoopPredecessor->getTerminator());
1157 if (!OuterLoopPredecessorBI || !InnerLoopLatchPredecessorBI)
1159 BasicBlock *InnerLoopHeaderSuccessor = InnerLoopHeader->getUniqueSuccessor();
1160 if (!InnerLoopHeaderSuccessor)
1163 // Adjust Loop Preheader and headers
1165 unsigned NumSucc = OuterLoopPredecessorBI->getNumSuccessors();
1166 for (unsigned i = 0; i < NumSucc; ++i) {
1167 if (OuterLoopPredecessorBI->getSuccessor(i) == OuterLoopPreHeader)
1168 OuterLoopPredecessorBI->setSuccessor(i, InnerLoopPreHeader);
1171 NumSucc = OuterLoopHeaderBI->getNumSuccessors();
1172 for (unsigned i = 0; i < NumSucc; ++i) {
1173 if (OuterLoopHeaderBI->getSuccessor(i) == OuterLoopLatch)
1174 OuterLoopHeaderBI->setSuccessor(i, LoopExit);
1175 else if (OuterLoopHeaderBI->getSuccessor(i) == InnerLoopPreHeader)
1176 OuterLoopHeaderBI->setSuccessor(i, InnerLoopHeaderSuccessor);
1179 // Adjust reduction PHI's now that the incoming block has changed.
1180 updateIncomingBlock(InnerLoopHeaderSuccessor, InnerLoopHeader,
1183 BranchInst::Create(OuterLoopPreHeader, InnerLoopHeaderBI);
1184 InnerLoopHeaderBI->eraseFromParent();
1186 // -------------Adjust loop latches-----------
1187 if (InnerLoopLatchBI->getSuccessor(0) == InnerLoopHeader)
1188 InnerLoopLatchSuccessor = InnerLoopLatchBI->getSuccessor(1);
1190 InnerLoopLatchSuccessor = InnerLoopLatchBI->getSuccessor(0);
1192 NumSucc = InnerLoopLatchPredecessorBI->getNumSuccessors();
1193 for (unsigned i = 0; i < NumSucc; ++i) {
1194 if (InnerLoopLatchPredecessorBI->getSuccessor(i) == InnerLoopLatch)
1195 InnerLoopLatchPredecessorBI->setSuccessor(i, InnerLoopLatchSuccessor);
1198 // Adjust PHI nodes in InnerLoopLatchSuccessor. Update all uses of PHI with
1199 // the value and remove this PHI node from inner loop.
1200 SmallVector<PHINode *, 8> LcssaVec;
1201 for (auto I = InnerLoopLatchSuccessor->begin(); isa<PHINode>(I); ++I) {
1202 PHINode *LcssaPhi = cast<PHINode>(I);
1203 LcssaVec.push_back(LcssaPhi);
1205 for (PHINode *P : LcssaVec) {
1206 Value *Incoming = P->getIncomingValueForBlock(InnerLoopLatch);
1207 P->replaceAllUsesWith(Incoming);
1208 P->eraseFromParent();
1211 if (OuterLoopLatchBI->getSuccessor(0) == OuterLoopHeader)
1212 OuterLoopLatchSuccessor = OuterLoopLatchBI->getSuccessor(1);
1214 OuterLoopLatchSuccessor = OuterLoopLatchBI->getSuccessor(0);
1216 if (InnerLoopLatchBI->getSuccessor(1) == InnerLoopLatchSuccessor)
1217 InnerLoopLatchBI->setSuccessor(1, OuterLoopLatchSuccessor);
1219 InnerLoopLatchBI->setSuccessor(0, OuterLoopLatchSuccessor);
1221 updateIncomingBlock(OuterLoopLatchSuccessor, OuterLoopLatch, InnerLoopLatch);
1223 if (OuterLoopLatchBI->getSuccessor(0) == OuterLoopLatchSuccessor) {
1224 OuterLoopLatchBI->setSuccessor(0, InnerLoopLatch);
1226 OuterLoopLatchBI->setSuccessor(1, InnerLoopLatch);
1231 void LoopInterchangeTransform::adjustLoopPreheaders() {
1233 // We have interchanged the preheaders so we need to interchange the data in
1234 // the preheader as well.
1235 // This is because the content of inner preheader was previously executed
1236 // inside the outer loop.
1237 BasicBlock *OuterLoopPreHeader = OuterLoop->getLoopPreheader();
1238 BasicBlock *InnerLoopPreHeader = InnerLoop->getLoopPreheader();
1239 BasicBlock *OuterLoopHeader = OuterLoop->getHeader();
1240 BranchInst *InnerTermBI =
1241 cast<BranchInst>(InnerLoopPreHeader->getTerminator());
1243 // These instructions should now be executed inside the loop.
1244 // Move instruction into a new block after outer header.
1245 moveBBContents(InnerLoopPreHeader, OuterLoopHeader->getTerminator());
1246 // These instructions were not executed previously in the loop so move them to
1247 // the older inner loop preheader.
1248 moveBBContents(OuterLoopPreHeader, InnerTermBI);
1251 bool LoopInterchangeTransform::adjustLoopLinks() {
1253 // Adjust all branches in the inner and outer loop.
1254 bool Changed = adjustLoopBranches();
1256 adjustLoopPreheaders();
1260 char LoopInterchange::ID = 0;
1261 INITIALIZE_PASS_BEGIN(LoopInterchange, "loop-interchange",
1262 "Interchanges loops for cache reuse", false, false)
1263 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
1264 INITIALIZE_PASS_DEPENDENCY(DependenceAnalysisWrapperPass)
1265 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
1266 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
1267 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
1268 INITIALIZE_PASS_DEPENDENCY(LCSSAWrapperPass)
1269 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
1271 INITIALIZE_PASS_END(LoopInterchange, "loop-interchange",
1272 "Interchanges loops for cache reuse", false, false)
1274 Pass *llvm::createLoopInterchangePass() { return new LoopInterchange(); }