1 //===-- SimplifyIndVar.cpp - Induction variable simplification ------------===//
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 file implements induction variable simplification. It does
11 // not define any actual pass or policy, but provides a single function to
12 // simplify a loop's induction variables based on ScalarEvolution.
14 //===----------------------------------------------------------------------===//
16 #include "llvm/Transforms/Utils/SimplifyIndVar.h"
17 #include "llvm/ADT/STLExtras.h"
18 #include "llvm/ADT/SmallVector.h"
19 #include "llvm/ADT/Statistic.h"
20 #include "llvm/Analysis/LoopInfo.h"
21 #include "llvm/Analysis/LoopPass.h"
22 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
23 #include "llvm/IR/DataLayout.h"
24 #include "llvm/IR/Dominators.h"
25 #include "llvm/IR/IRBuilder.h"
26 #include "llvm/IR/Instructions.h"
27 #include "llvm/IR/IntrinsicInst.h"
28 #include "llvm/Support/CommandLine.h"
29 #include "llvm/Support/Debug.h"
30 #include "llvm/Support/raw_ostream.h"
34 #define DEBUG_TYPE "indvars"
36 STATISTIC(NumElimIdentity, "Number of IV identities eliminated");
37 STATISTIC(NumElimOperand, "Number of IV operands folded into a use");
38 STATISTIC(NumElimRem , "Number of IV remainder operations eliminated");
39 STATISTIC(NumElimCmp , "Number of IV comparisons eliminated");
42 /// This is a utility for simplifying induction variables
43 /// based on ScalarEvolution. It is the primary instrument of the
44 /// IndvarSimplify pass, but it may also be directly invoked to cleanup after
45 /// other loop passes that preserve SCEV.
46 class SimplifyIndvar {
51 SmallVectorImpl<WeakVH> &DeadInsts;
56 SimplifyIndvar(Loop *Loop, ScalarEvolution *SE, LoopInfo *LI,
57 SmallVectorImpl<WeakVH> &Dead)
58 : L(Loop), LI(LI), SE(SE), DeadInsts(Dead), Changed(false) {
59 assert(LI && "IV simplification requires LoopInfo");
62 bool hasChanged() const { return Changed; }
64 /// Iteratively perform simplification on a worklist of users of the
65 /// specified induction variable. This is the top-level driver that applies
66 /// all simplicitions to users of an IV.
67 void simplifyUsers(PHINode *CurrIV, IVVisitor *V = nullptr);
69 Value *foldIVUser(Instruction *UseInst, Instruction *IVOperand);
71 bool eliminateIVUser(Instruction *UseInst, Instruction *IVOperand);
72 void eliminateIVComparison(ICmpInst *ICmp, Value *IVOperand);
73 void eliminateIVRemainder(BinaryOperator *Rem, Value *IVOperand,
75 bool strengthenOverflowingOperation(BinaryOperator *OBO, Value *IVOperand);
77 Instruction *splitOverflowIntrinsic(Instruction *IVUser,
78 const DominatorTree *DT);
82 /// Fold an IV operand into its use. This removes increments of an
83 /// aligned IV when used by a instruction that ignores the low bits.
85 /// IVOperand is guaranteed SCEVable, but UseInst may not be.
87 /// Return the operand of IVOperand for this induction variable if IVOperand can
88 /// be folded (in case more folding opportunities have been exposed).
89 /// Otherwise return null.
90 Value *SimplifyIndvar::foldIVUser(Instruction *UseInst, Instruction *IVOperand) {
91 Value *IVSrc = nullptr;
93 const SCEV *FoldedExpr = nullptr;
94 switch (UseInst->getOpcode()) {
97 case Instruction::UDiv:
98 case Instruction::LShr:
99 // We're only interested in the case where we know something about
100 // the numerator and have a constant denominator.
101 if (IVOperand != UseInst->getOperand(OperIdx) ||
102 !isa<ConstantInt>(UseInst->getOperand(1)))
105 // Attempt to fold a binary operator with constant operand.
106 // e.g. ((I + 1) >> 2) => I >> 2
107 if (!isa<BinaryOperator>(IVOperand)
108 || !isa<ConstantInt>(IVOperand->getOperand(1)))
111 IVSrc = IVOperand->getOperand(0);
112 // IVSrc must be the (SCEVable) IV, since the other operand is const.
113 assert(SE->isSCEVable(IVSrc->getType()) && "Expect SCEVable IV operand");
115 ConstantInt *D = cast<ConstantInt>(UseInst->getOperand(1));
116 if (UseInst->getOpcode() == Instruction::LShr) {
117 // Get a constant for the divisor. See createSCEV.
118 uint32_t BitWidth = cast<IntegerType>(UseInst->getType())->getBitWidth();
119 if (D->getValue().uge(BitWidth))
122 D = ConstantInt::get(UseInst->getContext(),
123 APInt::getOneBitSet(BitWidth, D->getZExtValue()));
125 FoldedExpr = SE->getUDivExpr(SE->getSCEV(IVSrc), SE->getSCEV(D));
127 // We have something that might fold it's operand. Compare SCEVs.
128 if (!SE->isSCEVable(UseInst->getType()))
131 // Bypass the operand if SCEV can prove it has no effect.
132 if (SE->getSCEV(UseInst) != FoldedExpr)
135 DEBUG(dbgs() << "INDVARS: Eliminated IV operand: " << *IVOperand
136 << " -> " << *UseInst << '\n');
138 UseInst->setOperand(OperIdx, IVSrc);
139 assert(SE->getSCEV(UseInst) == FoldedExpr && "bad SCEV with folded oper");
143 if (IVOperand->use_empty())
144 DeadInsts.emplace_back(IVOperand);
148 /// SimplifyIVUsers helper for eliminating useless
149 /// comparisons against an induction variable.
150 void SimplifyIndvar::eliminateIVComparison(ICmpInst *ICmp, Value *IVOperand) {
151 unsigned IVOperIdx = 0;
152 ICmpInst::Predicate Pred = ICmp->getPredicate();
153 if (IVOperand != ICmp->getOperand(0)) {
155 assert(IVOperand == ICmp->getOperand(1) && "Can't find IVOperand");
157 Pred = ICmpInst::getSwappedPredicate(Pred);
160 // Get the SCEVs for the ICmp operands.
161 const SCEV *S = SE->getSCEV(ICmp->getOperand(IVOperIdx));
162 const SCEV *X = SE->getSCEV(ICmp->getOperand(1 - IVOperIdx));
164 // Simplify unnecessary loops away.
165 const Loop *ICmpLoop = LI->getLoopFor(ICmp->getParent());
166 S = SE->getSCEVAtScope(S, ICmpLoop);
167 X = SE->getSCEVAtScope(X, ICmpLoop);
169 ICmpInst::Predicate InvariantPredicate;
170 const SCEV *InvariantLHS, *InvariantRHS;
172 const char *Verb = nullptr;
174 // If the condition is always true or always false, replace it with
176 if (SE->isKnownPredicate(Pred, S, X)) {
177 ICmp->replaceAllUsesWith(ConstantInt::getTrue(ICmp->getContext()));
178 DeadInsts.emplace_back(ICmp);
180 } else if (SE->isKnownPredicate(ICmpInst::getInversePredicate(Pred), S, X)) {
181 ICmp->replaceAllUsesWith(ConstantInt::getFalse(ICmp->getContext()));
182 DeadInsts.emplace_back(ICmp);
184 } else if (isa<PHINode>(IVOperand) &&
185 SE->isLoopInvariantPredicate(Pred, S, X, ICmpLoop,
186 InvariantPredicate, InvariantLHS,
189 // Rewrite the comparision to a loop invariant comparision if it can be done
190 // cheaply, where cheaply means "we don't need to emit any new
193 Value *NewLHS = nullptr, *NewRHS = nullptr;
195 if (S == InvariantLHS || X == InvariantLHS)
197 ICmp->getOperand(S == InvariantLHS ? IVOperIdx : (1 - IVOperIdx));
199 if (S == InvariantRHS || X == InvariantRHS)
201 ICmp->getOperand(S == InvariantRHS ? IVOperIdx : (1 - IVOperIdx));
203 for (Value *Incoming : cast<PHINode>(IVOperand)->incoming_values()) {
204 if (NewLHS && NewRHS)
207 const SCEV *IncomingS = SE->getSCEV(Incoming);
209 if (!NewLHS && IncomingS == InvariantLHS)
211 if (!NewRHS && IncomingS == InvariantRHS)
215 if (!NewLHS || !NewRHS)
216 // We could not find an existing value to replace either LHS or RHS.
217 // Generating new instructions has subtler tradeoffs, so avoid doing that
222 ICmp->setPredicate(InvariantPredicate);
223 ICmp->setOperand(0, NewLHS);
224 ICmp->setOperand(1, NewRHS);
228 DEBUG(dbgs() << "INDVARS: " << Verb << " comparison: " << *ICmp << '\n');
233 /// SimplifyIVUsers helper for eliminating useless
234 /// remainder operations operating on an induction variable.
235 void SimplifyIndvar::eliminateIVRemainder(BinaryOperator *Rem,
238 // We're only interested in the case where we know something about
240 if (IVOperand != Rem->getOperand(0))
243 // Get the SCEVs for the ICmp operands.
244 const SCEV *S = SE->getSCEV(Rem->getOperand(0));
245 const SCEV *X = SE->getSCEV(Rem->getOperand(1));
247 // Simplify unnecessary loops away.
248 const Loop *ICmpLoop = LI->getLoopFor(Rem->getParent());
249 S = SE->getSCEVAtScope(S, ICmpLoop);
250 X = SE->getSCEVAtScope(X, ICmpLoop);
252 // i % n --> i if i is in [0,n).
253 if ((!IsSigned || SE->isKnownNonNegative(S)) &&
254 SE->isKnownPredicate(IsSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT,
256 Rem->replaceAllUsesWith(Rem->getOperand(0));
258 // (i+1) % n --> (i+1)==n?0:(i+1) if i is in [0,n).
259 const SCEV *LessOne =
260 SE->getMinusSCEV(S, SE->getConstant(S->getType(), 1));
261 if (IsSigned && !SE->isKnownNonNegative(LessOne))
264 if (!SE->isKnownPredicate(IsSigned ?
265 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT,
269 ICmpInst *ICmp = new ICmpInst(Rem, ICmpInst::ICMP_EQ,
270 Rem->getOperand(0), Rem->getOperand(1));
272 SelectInst::Create(ICmp,
273 ConstantInt::get(Rem->getType(), 0),
274 Rem->getOperand(0), "tmp", Rem);
275 Rem->replaceAllUsesWith(Sel);
278 DEBUG(dbgs() << "INDVARS: Simplified rem: " << *Rem << '\n');
281 DeadInsts.emplace_back(Rem);
284 /// Eliminate an operation that consumes a simple IV and has
285 /// no observable side-effect given the range of IV values.
286 /// IVOperand is guaranteed SCEVable, but UseInst may not be.
287 bool SimplifyIndvar::eliminateIVUser(Instruction *UseInst,
288 Instruction *IVOperand) {
289 if (ICmpInst *ICmp = dyn_cast<ICmpInst>(UseInst)) {
290 eliminateIVComparison(ICmp, IVOperand);
293 if (BinaryOperator *Rem = dyn_cast<BinaryOperator>(UseInst)) {
294 bool IsSigned = Rem->getOpcode() == Instruction::SRem;
295 if (IsSigned || Rem->getOpcode() == Instruction::URem) {
296 eliminateIVRemainder(Rem, IVOperand, IsSigned);
301 // Eliminate any operation that SCEV can prove is an identity function.
302 if (!SE->isSCEVable(UseInst->getType()) ||
303 (UseInst->getType() != IVOperand->getType()) ||
304 (SE->getSCEV(UseInst) != SE->getSCEV(IVOperand)))
307 DEBUG(dbgs() << "INDVARS: Eliminated identity: " << *UseInst << '\n');
309 UseInst->replaceAllUsesWith(IVOperand);
312 DeadInsts.emplace_back(UseInst);
316 /// Annotate BO with nsw / nuw if it provably does not signed-overflow /
317 /// unsigned-overflow. Returns true if anything changed, false otherwise.
318 bool SimplifyIndvar::strengthenOverflowingOperation(BinaryOperator *BO,
321 // Fastpath: we don't have any work to do if `BO` is `nuw` and `nsw`.
322 if (BO->hasNoUnsignedWrap() && BO->hasNoSignedWrap())
325 const SCEV *(ScalarEvolution::*GetExprForBO)(const SCEV *, const SCEV *,
328 switch (BO->getOpcode()) {
332 case Instruction::Add:
333 GetExprForBO = &ScalarEvolution::getAddExpr;
336 case Instruction::Sub:
337 GetExprForBO = &ScalarEvolution::getMinusSCEV;
340 case Instruction::Mul:
341 GetExprForBO = &ScalarEvolution::getMulExpr;
345 unsigned BitWidth = cast<IntegerType>(BO->getType())->getBitWidth();
346 Type *WideTy = IntegerType::get(BO->getContext(), BitWidth * 2);
347 const SCEV *LHS = SE->getSCEV(BO->getOperand(0));
348 const SCEV *RHS = SE->getSCEV(BO->getOperand(1));
350 bool Changed = false;
352 if (!BO->hasNoUnsignedWrap()) {
353 const SCEV *ExtendAfterOp = SE->getZeroExtendExpr(SE->getSCEV(BO), WideTy);
354 const SCEV *OpAfterExtend = (SE->*GetExprForBO)(
355 SE->getZeroExtendExpr(LHS, WideTy), SE->getZeroExtendExpr(RHS, WideTy),
357 if (ExtendAfterOp == OpAfterExtend) {
358 BO->setHasNoUnsignedWrap();
364 if (!BO->hasNoSignedWrap()) {
365 const SCEV *ExtendAfterOp = SE->getSignExtendExpr(SE->getSCEV(BO), WideTy);
366 const SCEV *OpAfterExtend = (SE->*GetExprForBO)(
367 SE->getSignExtendExpr(LHS, WideTy), SE->getSignExtendExpr(RHS, WideTy),
369 if (ExtendAfterOp == OpAfterExtend) {
370 BO->setHasNoSignedWrap();
379 /// \brief Split sadd.with.overflow into add + sadd.with.overflow to allow
380 /// analysis and optimization.
382 /// \return A new value representing the non-overflowing add if possible,
383 /// otherwise return the original value.
384 Instruction *SimplifyIndvar::splitOverflowIntrinsic(Instruction *IVUser,
385 const DominatorTree *DT) {
386 IntrinsicInst *II = dyn_cast<IntrinsicInst>(IVUser);
387 if (!II || II->getIntrinsicID() != Intrinsic::sadd_with_overflow)
390 // Find a branch guarded by the overflow check.
391 BranchInst *Branch = nullptr;
392 Instruction *AddVal = nullptr;
393 for (User *U : II->users()) {
394 if (ExtractValueInst *ExtractInst = dyn_cast<ExtractValueInst>(U)) {
395 if (ExtractInst->getNumIndices() != 1)
397 if (ExtractInst->getIndices()[0] == 0)
398 AddVal = ExtractInst;
399 else if (ExtractInst->getIndices()[0] == 1 && ExtractInst->hasOneUse())
400 Branch = dyn_cast<BranchInst>(ExtractInst->user_back());
403 if (!AddVal || !Branch)
406 BasicBlock *ContinueBB = Branch->getSuccessor(1);
407 if (std::next(pred_begin(ContinueBB)) != pred_end(ContinueBB))
410 // Check if all users of the add are provably NSW.
412 for (Use &U : AddVal->uses()) {
413 if (Instruction *UseInst = dyn_cast<Instruction>(U.getUser())) {
414 BasicBlock *UseBB = UseInst->getParent();
415 if (PHINode *PHI = dyn_cast<PHINode>(UseInst))
416 UseBB = PHI->getIncomingBlock(U);
417 if (!DT->dominates(ContinueBB, UseBB)) {
427 IRBuilder<> Builder(IVUser);
428 Instruction *AddInst = dyn_cast<Instruction>(
429 Builder.CreateNSWAdd(II->getOperand(0), II->getOperand(1)));
431 // The caller expects the new add to have the same form as the intrinsic. The
432 // IV operand position must be the same.
433 assert((AddInst->getOpcode() == Instruction::Add &&
434 AddInst->getOperand(0) == II->getOperand(0)) &&
435 "Bad add instruction created from overflow intrinsic.");
437 AddVal->replaceAllUsesWith(AddInst);
438 DeadInsts.emplace_back(AddVal);
442 /// Add all uses of Def to the current IV's worklist.
443 static void pushIVUsers(
445 SmallPtrSet<Instruction*,16> &Simplified,
446 SmallVectorImpl< std::pair<Instruction*,Instruction*> > &SimpleIVUsers) {
448 for (User *U : Def->users()) {
449 Instruction *UI = cast<Instruction>(U);
451 // Avoid infinite or exponential worklist processing.
452 // Also ensure unique worklist users.
453 // If Def is a LoopPhi, it may not be in the Simplified set, so check for
455 if (UI != Def && Simplified.insert(UI).second)
456 SimpleIVUsers.push_back(std::make_pair(UI, Def));
460 /// Return true if this instruction generates a simple SCEV
461 /// expression in terms of that IV.
463 /// This is similar to IVUsers' isInteresting() but processes each instruction
464 /// non-recursively when the operand is already known to be a simpleIVUser.
466 static bool isSimpleIVUser(Instruction *I, const Loop *L, ScalarEvolution *SE) {
467 if (!SE->isSCEVable(I->getType()))
470 // Get the symbolic expression for this instruction.
471 const SCEV *S = SE->getSCEV(I);
473 // Only consider affine recurrences.
474 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S);
475 if (AR && AR->getLoop() == L)
481 /// Iteratively perform simplification on a worklist of users
482 /// of the specified induction variable. Each successive simplification may push
483 /// more users which may themselves be candidates for simplification.
485 /// This algorithm does not require IVUsers analysis. Instead, it simplifies
486 /// instructions in-place during analysis. Rather than rewriting induction
487 /// variables bottom-up from their users, it transforms a chain of IVUsers
488 /// top-down, updating the IR only when it encouters a clear optimization
491 /// Once DisableIVRewrite is default, LSR will be the only client of IVUsers.
493 void SimplifyIndvar::simplifyUsers(PHINode *CurrIV, IVVisitor *V) {
494 if (!SE->isSCEVable(CurrIV->getType()))
497 // Instructions processed by SimplifyIndvar for CurrIV.
498 SmallPtrSet<Instruction*,16> Simplified;
500 // Use-def pairs if IV users waiting to be processed for CurrIV.
501 SmallVector<std::pair<Instruction*, Instruction*>, 8> SimpleIVUsers;
503 // Push users of the current LoopPhi. In rare cases, pushIVUsers may be
504 // called multiple times for the same LoopPhi. This is the proper thing to
505 // do for loop header phis that use each other.
506 pushIVUsers(CurrIV, Simplified, SimpleIVUsers);
508 while (!SimpleIVUsers.empty()) {
509 std::pair<Instruction*, Instruction*> UseOper =
510 SimpleIVUsers.pop_back_val();
511 Instruction *UseInst = UseOper.first;
513 // Bypass back edges to avoid extra work.
514 if (UseInst == CurrIV) continue;
516 if (V && V->shouldSplitOverflowInstrinsics()) {
517 UseInst = splitOverflowIntrinsic(UseInst, V->getDomTree());
522 Instruction *IVOperand = UseOper.second;
523 for (unsigned N = 0; IVOperand; ++N) {
524 assert(N <= Simplified.size() && "runaway iteration");
526 Value *NewOper = foldIVUser(UseOper.first, IVOperand);
528 break; // done folding
529 IVOperand = dyn_cast<Instruction>(NewOper);
534 if (eliminateIVUser(UseOper.first, IVOperand)) {
535 pushIVUsers(IVOperand, Simplified, SimpleIVUsers);
539 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(UseOper.first)) {
540 if (isa<OverflowingBinaryOperator>(BO) &&
541 strengthenOverflowingOperation(BO, IVOperand)) {
542 // re-queue uses of the now modified binary operator and fall
543 // through to the checks that remain.
544 pushIVUsers(IVOperand, Simplified, SimpleIVUsers);
548 CastInst *Cast = dyn_cast<CastInst>(UseOper.first);
553 if (isSimpleIVUser(UseOper.first, L, SE)) {
554 pushIVUsers(UseOper.first, Simplified, SimpleIVUsers);
561 void IVVisitor::anchor() { }
563 /// Simplify instructions that use this induction variable
564 /// by using ScalarEvolution to analyze the IV's recurrence.
565 bool simplifyUsersOfIV(PHINode *CurrIV, ScalarEvolution *SE, LPPassManager *LPM,
566 SmallVectorImpl<WeakVH> &Dead, IVVisitor *V)
568 LoopInfo *LI = &LPM->getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
569 SimplifyIndvar SIV(LI->getLoopFor(CurrIV->getParent()), SE, LI, Dead);
570 SIV.simplifyUsers(CurrIV, V);
571 return SIV.hasChanged();
574 /// Simplify users of induction variables within this
575 /// loop. This does not actually change or add IVs.
576 bool simplifyLoopIVs(Loop *L, ScalarEvolution *SE, LPPassManager *LPM,
577 SmallVectorImpl<WeakVH> &Dead) {
578 bool Changed = false;
579 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
580 Changed |= simplifyUsersOfIV(cast<PHINode>(I), SE, LPM, Dead);