1 //===-- SimplifyIndVar.cpp - Induction variable simplification ------------===//
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7 //===----------------------------------------------------------------------===//
9 // This file implements induction variable simplification. It does
10 // not define any actual pass or policy, but provides a single function to
11 // simplify a loop's induction variables based on ScalarEvolution.
13 //===----------------------------------------------------------------------===//
15 #include "llvm/Transforms/Utils/SimplifyIndVar.h"
16 #include "llvm/ADT/SmallVector.h"
17 #include "llvm/ADT/Statistic.h"
18 #include "llvm/Analysis/LoopInfo.h"
19 #include "llvm/IR/Dominators.h"
20 #include "llvm/IR/IRBuilder.h"
21 #include "llvm/IR/Instructions.h"
22 #include "llvm/IR/IntrinsicInst.h"
23 #include "llvm/IR/PatternMatch.h"
24 #include "llvm/Support/Debug.h"
25 #include "llvm/Support/raw_ostream.h"
26 #include "llvm/Transforms/Utils/Local.h"
27 #include "llvm/Transforms/Utils/ScalarEvolutionExpander.h"
31 #define DEBUG_TYPE "indvars"
33 STATISTIC(NumElimIdentity, "Number of IV identities eliminated");
34 STATISTIC(NumElimOperand, "Number of IV operands folded into a use");
35 STATISTIC(NumFoldedUser, "Number of IV users folded into a constant");
36 STATISTIC(NumElimRem , "Number of IV remainder operations eliminated");
39 "Number of IV signed division operations converted to unsigned division");
42 "Number of IV signed remainder operations converted to unsigned remainder");
43 STATISTIC(NumElimCmp , "Number of IV comparisons eliminated");
46 /// This is a utility for simplifying induction variables
47 /// based on ScalarEvolution. It is the primary instrument of the
48 /// IndvarSimplify pass, but it may also be directly invoked to cleanup after
49 /// other loop passes that preserve SCEV.
50 class SimplifyIndvar {
55 const TargetTransformInfo *TTI;
56 SCEVExpander &Rewriter;
57 SmallVectorImpl<WeakTrackingVH> &DeadInsts;
62 SimplifyIndvar(Loop *Loop, ScalarEvolution *SE, DominatorTree *DT,
63 LoopInfo *LI, const TargetTransformInfo *TTI,
64 SCEVExpander &Rewriter,
65 SmallVectorImpl<WeakTrackingVH> &Dead)
66 : L(Loop), LI(LI), SE(SE), DT(DT), TTI(TTI), Rewriter(Rewriter),
68 assert(LI && "IV simplification requires LoopInfo");
71 bool hasChanged() const { return Changed; }
73 /// Iteratively perform simplification on a worklist of users of the
74 /// specified induction variable. This is the top-level driver that applies
75 /// all simplifications to users of an IV.
76 void simplifyUsers(PHINode *CurrIV, IVVisitor *V = nullptr);
78 Value *foldIVUser(Instruction *UseInst, Instruction *IVOperand);
80 bool eliminateIdentitySCEV(Instruction *UseInst, Instruction *IVOperand);
81 bool replaceIVUserWithLoopInvariant(Instruction *UseInst);
83 bool eliminateOverflowIntrinsic(WithOverflowInst *WO);
84 bool eliminateSaturatingIntrinsic(SaturatingInst *SI);
85 bool eliminateTrunc(TruncInst *TI);
86 bool eliminateIVUser(Instruction *UseInst, Instruction *IVOperand);
87 bool makeIVComparisonInvariant(ICmpInst *ICmp, Value *IVOperand);
88 void eliminateIVComparison(ICmpInst *ICmp, Value *IVOperand);
89 void simplifyIVRemainder(BinaryOperator *Rem, Value *IVOperand,
91 void replaceRemWithNumerator(BinaryOperator *Rem);
92 void replaceRemWithNumeratorOrZero(BinaryOperator *Rem);
93 void replaceSRemWithURem(BinaryOperator *Rem);
94 bool eliminateSDiv(BinaryOperator *SDiv);
95 bool strengthenOverflowingOperation(BinaryOperator *OBO, Value *IVOperand);
96 bool strengthenRightShift(BinaryOperator *BO, Value *IVOperand);
100 /// Find a point in code which dominates all given instructions. We can safely
101 /// assume that, whatever fact we can prove at the found point, this fact is
102 /// also true for each of the given instructions.
103 static Instruction *findCommonDominator(ArrayRef<Instruction *> Instructions,
105 Instruction *CommonDom = nullptr;
106 for (auto *Insn : Instructions)
107 if (!CommonDom || DT.dominates(Insn, CommonDom))
109 else if (!DT.dominates(CommonDom, Insn))
110 // If there is no dominance relation, use common dominator.
112 DT.findNearestCommonDominator(CommonDom->getParent(),
113 Insn->getParent())->getTerminator();
114 assert(CommonDom && "Common dominator not found?");
118 /// Fold an IV operand into its use. This removes increments of an
119 /// aligned IV when used by a instruction that ignores the low bits.
121 /// IVOperand is guaranteed SCEVable, but UseInst may not be.
123 /// Return the operand of IVOperand for this induction variable if IVOperand can
124 /// be folded (in case more folding opportunities have been exposed).
125 /// Otherwise return null.
126 Value *SimplifyIndvar::foldIVUser(Instruction *UseInst, Instruction *IVOperand) {
127 Value *IVSrc = nullptr;
128 const unsigned OperIdx = 0;
129 const SCEV *FoldedExpr = nullptr;
130 bool MustDropExactFlag = false;
131 switch (UseInst->getOpcode()) {
134 case Instruction::UDiv:
135 case Instruction::LShr:
136 // We're only interested in the case where we know something about
137 // the numerator and have a constant denominator.
138 if (IVOperand != UseInst->getOperand(OperIdx) ||
139 !isa<ConstantInt>(UseInst->getOperand(1)))
142 // Attempt to fold a binary operator with constant operand.
143 // e.g. ((I + 1) >> 2) => I >> 2
144 if (!isa<BinaryOperator>(IVOperand)
145 || !isa<ConstantInt>(IVOperand->getOperand(1)))
148 IVSrc = IVOperand->getOperand(0);
149 // IVSrc must be the (SCEVable) IV, since the other operand is const.
150 assert(SE->isSCEVable(IVSrc->getType()) && "Expect SCEVable IV operand");
152 ConstantInt *D = cast<ConstantInt>(UseInst->getOperand(1));
153 if (UseInst->getOpcode() == Instruction::LShr) {
154 // Get a constant for the divisor. See createSCEV.
155 uint32_t BitWidth = cast<IntegerType>(UseInst->getType())->getBitWidth();
156 if (D->getValue().uge(BitWidth))
159 D = ConstantInt::get(UseInst->getContext(),
160 APInt::getOneBitSet(BitWidth, D->getZExtValue()));
162 const auto *LHS = SE->getSCEV(IVSrc);
163 const auto *RHS = SE->getSCEV(D);
164 FoldedExpr = SE->getUDivExpr(LHS, RHS);
165 // We might have 'exact' flag set at this point which will no longer be
166 // correct after we make the replacement.
167 if (UseInst->isExact() && LHS != SE->getMulExpr(FoldedExpr, RHS))
168 MustDropExactFlag = true;
170 // We have something that might fold it's operand. Compare SCEVs.
171 if (!SE->isSCEVable(UseInst->getType()))
174 // Bypass the operand if SCEV can prove it has no effect.
175 if (SE->getSCEV(UseInst) != FoldedExpr)
178 LLVM_DEBUG(dbgs() << "INDVARS: Eliminated IV operand: " << *IVOperand
179 << " -> " << *UseInst << '\n');
181 UseInst->setOperand(OperIdx, IVSrc);
182 assert(SE->getSCEV(UseInst) == FoldedExpr && "bad SCEV with folded oper");
184 if (MustDropExactFlag)
185 UseInst->dropPoisonGeneratingFlags();
189 if (IVOperand->use_empty())
190 DeadInsts.emplace_back(IVOperand);
194 bool SimplifyIndvar::makeIVComparisonInvariant(ICmpInst *ICmp,
196 unsigned IVOperIdx = 0;
197 ICmpInst::Predicate Pred = ICmp->getPredicate();
198 if (IVOperand != ICmp->getOperand(0)) {
200 assert(IVOperand == ICmp->getOperand(1) && "Can't find IVOperand");
202 Pred = ICmpInst::getSwappedPredicate(Pred);
205 // Get the SCEVs for the ICmp operands (in the specific context of the
207 const Loop *ICmpLoop = LI->getLoopFor(ICmp->getParent());
208 const SCEV *S = SE->getSCEVAtScope(ICmp->getOperand(IVOperIdx), ICmpLoop);
209 const SCEV *X = SE->getSCEVAtScope(ICmp->getOperand(1 - IVOperIdx), ICmpLoop);
211 auto *PN = dyn_cast<PHINode>(IVOperand);
214 auto LIP = SE->getLoopInvariantPredicate(Pred, S, X, L);
217 ICmpInst::Predicate InvariantPredicate = LIP->Pred;
218 const SCEV *InvariantLHS = LIP->LHS;
219 const SCEV *InvariantRHS = LIP->RHS;
221 // Rewrite the comparison to a loop invariant comparison if it can be done
222 // cheaply, where cheaply means "we don't need to emit any new
225 SmallDenseMap<const SCEV*, Value*> CheapExpansions;
226 CheapExpansions[S] = ICmp->getOperand(IVOperIdx);
227 CheapExpansions[X] = ICmp->getOperand(1 - IVOperIdx);
229 // TODO: Support multiple entry loops? (We currently bail out of these in
230 // the IndVarSimplify pass)
231 if (auto *BB = L->getLoopPredecessor()) {
232 const int Idx = PN->getBasicBlockIndex(BB);
234 Value *Incoming = PN->getIncomingValue(Idx);
235 const SCEV *IncomingS = SE->getSCEV(Incoming);
236 CheapExpansions[IncomingS] = Incoming;
239 Value *NewLHS = CheapExpansions[InvariantLHS];
240 Value *NewRHS = CheapExpansions[InvariantRHS];
243 if (auto *ConstLHS = dyn_cast<SCEVConstant>(InvariantLHS))
244 NewLHS = ConstLHS->getValue();
246 if (auto *ConstRHS = dyn_cast<SCEVConstant>(InvariantRHS))
247 NewRHS = ConstRHS->getValue();
249 if (!NewLHS || !NewRHS)
250 // We could not find an existing value to replace either LHS or RHS.
251 // Generating new instructions has subtler tradeoffs, so avoid doing that
255 LLVM_DEBUG(dbgs() << "INDVARS: Simplified comparison: " << *ICmp << '\n');
256 ICmp->setPredicate(InvariantPredicate);
257 ICmp->setOperand(0, NewLHS);
258 ICmp->setOperand(1, NewRHS);
262 /// SimplifyIVUsers helper for eliminating useless
263 /// comparisons against an induction variable.
264 void SimplifyIndvar::eliminateIVComparison(ICmpInst *ICmp, Value *IVOperand) {
265 unsigned IVOperIdx = 0;
266 ICmpInst::Predicate Pred = ICmp->getPredicate();
267 ICmpInst::Predicate OriginalPred = Pred;
268 if (IVOperand != ICmp->getOperand(0)) {
270 assert(IVOperand == ICmp->getOperand(1) && "Can't find IVOperand");
272 Pred = ICmpInst::getSwappedPredicate(Pred);
275 // Get the SCEVs for the ICmp operands (in the specific context of the
277 const Loop *ICmpLoop = LI->getLoopFor(ICmp->getParent());
278 const SCEV *S = SE->getSCEVAtScope(ICmp->getOperand(IVOperIdx), ICmpLoop);
279 const SCEV *X = SE->getSCEVAtScope(ICmp->getOperand(1 - IVOperIdx), ICmpLoop);
281 // If the condition is always true or always false in the given context,
282 // replace it with a constant value.
283 SmallVector<Instruction *, 4> Users;
284 for (auto *U : ICmp->users())
285 Users.push_back(cast<Instruction>(U));
286 const Instruction *CtxI = findCommonDominator(Users, *DT);
287 if (auto Ev = SE->evaluatePredicateAt(Pred, S, X, CtxI)) {
288 ICmp->replaceAllUsesWith(ConstantInt::getBool(ICmp->getContext(), *Ev));
289 DeadInsts.emplace_back(ICmp);
290 LLVM_DEBUG(dbgs() << "INDVARS: Eliminated comparison: " << *ICmp << '\n');
291 } else if (makeIVComparisonInvariant(ICmp, IVOperand)) {
292 // fallthrough to end of function
293 } else if (ICmpInst::isSigned(OriginalPred) &&
294 SE->isKnownNonNegative(S) && SE->isKnownNonNegative(X)) {
295 // If we were unable to make anything above, all we can is to canonicalize
296 // the comparison hoping that it will open the doors for other
297 // optimizations. If we find out that we compare two non-negative values,
298 // we turn the instruction's predicate to its unsigned version. Note that
299 // we cannot rely on Pred here unless we check if we have swapped it.
300 assert(ICmp->getPredicate() == OriginalPred && "Predicate changed?");
301 LLVM_DEBUG(dbgs() << "INDVARS: Turn to unsigned comparison: " << *ICmp
303 ICmp->setPredicate(ICmpInst::getUnsignedPredicate(OriginalPred));
311 bool SimplifyIndvar::eliminateSDiv(BinaryOperator *SDiv) {
312 // Get the SCEVs for the ICmp operands.
313 auto *N = SE->getSCEV(SDiv->getOperand(0));
314 auto *D = SE->getSCEV(SDiv->getOperand(1));
316 // Simplify unnecessary loops away.
317 const Loop *L = LI->getLoopFor(SDiv->getParent());
318 N = SE->getSCEVAtScope(N, L);
319 D = SE->getSCEVAtScope(D, L);
321 // Replace sdiv by udiv if both of the operands are non-negative
322 if (SE->isKnownNonNegative(N) && SE->isKnownNonNegative(D)) {
323 auto *UDiv = BinaryOperator::Create(
324 BinaryOperator::UDiv, SDiv->getOperand(0), SDiv->getOperand(1),
325 SDiv->getName() + ".udiv", SDiv);
326 UDiv->setIsExact(SDiv->isExact());
327 SDiv->replaceAllUsesWith(UDiv);
328 LLVM_DEBUG(dbgs() << "INDVARS: Simplified sdiv: " << *SDiv << '\n');
331 DeadInsts.push_back(SDiv);
338 // i %s n -> i %u n if i >= 0 and n >= 0
339 void SimplifyIndvar::replaceSRemWithURem(BinaryOperator *Rem) {
340 auto *N = Rem->getOperand(0), *D = Rem->getOperand(1);
341 auto *URem = BinaryOperator::Create(BinaryOperator::URem, N, D,
342 Rem->getName() + ".urem", Rem);
343 Rem->replaceAllUsesWith(URem);
344 LLVM_DEBUG(dbgs() << "INDVARS: Simplified srem: " << *Rem << '\n');
347 DeadInsts.emplace_back(Rem);
350 // i % n --> i if i is in [0,n).
351 void SimplifyIndvar::replaceRemWithNumerator(BinaryOperator *Rem) {
352 Rem->replaceAllUsesWith(Rem->getOperand(0));
353 LLVM_DEBUG(dbgs() << "INDVARS: Simplified rem: " << *Rem << '\n');
356 DeadInsts.emplace_back(Rem);
359 // (i+1) % n --> (i+1)==n?0:(i+1) if i is in [0,n).
360 void SimplifyIndvar::replaceRemWithNumeratorOrZero(BinaryOperator *Rem) {
361 auto *T = Rem->getType();
362 auto *N = Rem->getOperand(0), *D = Rem->getOperand(1);
363 ICmpInst *ICmp = new ICmpInst(Rem, ICmpInst::ICMP_EQ, N, D);
365 SelectInst::Create(ICmp, ConstantInt::get(T, 0), N, "iv.rem", Rem);
366 Rem->replaceAllUsesWith(Sel);
367 LLVM_DEBUG(dbgs() << "INDVARS: Simplified rem: " << *Rem << '\n');
370 DeadInsts.emplace_back(Rem);
373 /// SimplifyIVUsers helper for eliminating useless remainder operations
374 /// operating on an induction variable or replacing srem by urem.
375 void SimplifyIndvar::simplifyIVRemainder(BinaryOperator *Rem, Value *IVOperand,
377 auto *NValue = Rem->getOperand(0);
378 auto *DValue = Rem->getOperand(1);
379 // We're only interested in the case where we know something about
380 // the numerator, unless it is a srem, because we want to replace srem by urem
382 bool UsedAsNumerator = IVOperand == NValue;
383 if (!UsedAsNumerator && !IsSigned)
386 const SCEV *N = SE->getSCEV(NValue);
388 // Simplify unnecessary loops away.
389 const Loop *ICmpLoop = LI->getLoopFor(Rem->getParent());
390 N = SE->getSCEVAtScope(N, ICmpLoop);
392 bool IsNumeratorNonNegative = !IsSigned || SE->isKnownNonNegative(N);
394 // Do not proceed if the Numerator may be negative
395 if (!IsNumeratorNonNegative)
398 const SCEV *D = SE->getSCEV(DValue);
399 D = SE->getSCEVAtScope(D, ICmpLoop);
401 if (UsedAsNumerator) {
402 auto LT = IsSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
403 if (SE->isKnownPredicate(LT, N, D)) {
404 replaceRemWithNumerator(Rem);
408 auto *T = Rem->getType();
409 const auto *NLessOne = SE->getMinusSCEV(N, SE->getOne(T));
410 if (SE->isKnownPredicate(LT, NLessOne, D)) {
411 replaceRemWithNumeratorOrZero(Rem);
416 // Try to replace SRem with URem, if both N and D are known non-negative.
417 // Since we had already check N, we only need to check D now
418 if (!IsSigned || !SE->isKnownNonNegative(D))
421 replaceSRemWithURem(Rem);
424 bool SimplifyIndvar::eliminateOverflowIntrinsic(WithOverflowInst *WO) {
425 const SCEV *LHS = SE->getSCEV(WO->getLHS());
426 const SCEV *RHS = SE->getSCEV(WO->getRHS());
427 if (!SE->willNotOverflow(WO->getBinaryOp(), WO->isSigned(), LHS, RHS))
430 // Proved no overflow, nuke the overflow check and, if possible, the overflow
431 // intrinsic as well.
433 BinaryOperator *NewResult = BinaryOperator::Create(
434 WO->getBinaryOp(), WO->getLHS(), WO->getRHS(), "", WO);
437 NewResult->setHasNoSignedWrap(true);
439 NewResult->setHasNoUnsignedWrap(true);
441 SmallVector<ExtractValueInst *, 4> ToDelete;
443 for (auto *U : WO->users()) {
444 if (auto *EVI = dyn_cast<ExtractValueInst>(U)) {
445 if (EVI->getIndices()[0] == 1)
446 EVI->replaceAllUsesWith(ConstantInt::getFalse(WO->getContext()));
448 assert(EVI->getIndices()[0] == 0 && "Only two possibilities!");
449 EVI->replaceAllUsesWith(NewResult);
451 ToDelete.push_back(EVI);
455 for (auto *EVI : ToDelete)
456 EVI->eraseFromParent();
459 WO->eraseFromParent();
465 bool SimplifyIndvar::eliminateSaturatingIntrinsic(SaturatingInst *SI) {
466 const SCEV *LHS = SE->getSCEV(SI->getLHS());
467 const SCEV *RHS = SE->getSCEV(SI->getRHS());
468 if (!SE->willNotOverflow(SI->getBinaryOp(), SI->isSigned(), LHS, RHS))
471 BinaryOperator *BO = BinaryOperator::Create(
472 SI->getBinaryOp(), SI->getLHS(), SI->getRHS(), SI->getName(), SI);
474 BO->setHasNoSignedWrap();
476 BO->setHasNoUnsignedWrap();
478 SI->replaceAllUsesWith(BO);
479 DeadInsts.emplace_back(SI);
484 bool SimplifyIndvar::eliminateTrunc(TruncInst *TI) {
485 // It is always legal to replace
486 // icmp <pred> i32 trunc(iv), n
488 // icmp <pred> i64 sext(trunc(iv)), sext(n), if pred is signed predicate.
490 // icmp <pred> i64 zext(trunc(iv)), zext(n), if pred is unsigned predicate.
491 // Or with either of these if pred is an equality predicate.
493 // If we can prove that iv == sext(trunc(iv)) or iv == zext(trunc(iv)) for
494 // every comparison which uses trunc, it means that we can replace each of
495 // them with comparison of iv against sext/zext(n). We no longer need trunc
498 // TODO: Should we do this if we can widen *some* comparisons, but not all
499 // of them? Sometimes it is enough to enable other optimizations, but the
500 // trunc instruction will stay in the loop.
501 Value *IV = TI->getOperand(0);
502 Type *IVTy = IV->getType();
503 const SCEV *IVSCEV = SE->getSCEV(IV);
504 const SCEV *TISCEV = SE->getSCEV(TI);
506 // Check if iv == zext(trunc(iv)) and if iv == sext(trunc(iv)). If so, we can
508 bool DoesSExtCollapse = false;
509 bool DoesZExtCollapse = false;
510 if (IVSCEV == SE->getSignExtendExpr(TISCEV, IVTy))
511 DoesSExtCollapse = true;
512 if (IVSCEV == SE->getZeroExtendExpr(TISCEV, IVTy))
513 DoesZExtCollapse = true;
515 // If neither sext nor zext does collapse, it is not profitable to do any
517 if (!DoesSExtCollapse && !DoesZExtCollapse)
520 // Collect users of the trunc that look like comparisons against invariants.
521 // Bail if we find something different.
522 SmallVector<ICmpInst *, 4> ICmpUsers;
523 for (auto *U : TI->users()) {
524 // We don't care about users in unreachable blocks.
525 if (isa<Instruction>(U) &&
526 !DT->isReachableFromEntry(cast<Instruction>(U)->getParent()))
528 ICmpInst *ICI = dyn_cast<ICmpInst>(U);
529 if (!ICI) return false;
530 assert(L->contains(ICI->getParent()) && "LCSSA form broken?");
531 if (!(ICI->getOperand(0) == TI && L->isLoopInvariant(ICI->getOperand(1))) &&
532 !(ICI->getOperand(1) == TI && L->isLoopInvariant(ICI->getOperand(0))))
534 // If we cannot get rid of trunc, bail.
535 if (ICI->isSigned() && !DoesSExtCollapse)
537 if (ICI->isUnsigned() && !DoesZExtCollapse)
539 // For equality, either signed or unsigned works.
540 ICmpUsers.push_back(ICI);
543 auto CanUseZExt = [&](ICmpInst *ICI) {
544 // Unsigned comparison can be widened as unsigned.
545 if (ICI->isUnsigned())
547 // Is it profitable to do zext?
548 if (!DoesZExtCollapse)
550 // For equality, we can safely zext both parts.
551 if (ICI->isEquality())
553 // Otherwise we can only use zext when comparing two non-negative or two
554 // negative values. But in practice, we will never pass DoesZExtCollapse
555 // check for a negative value, because zext(trunc(x)) is non-negative. So
556 // it only make sense to check for non-negativity here.
557 const SCEV *SCEVOP1 = SE->getSCEV(ICI->getOperand(0));
558 const SCEV *SCEVOP2 = SE->getSCEV(ICI->getOperand(1));
559 return SE->isKnownNonNegative(SCEVOP1) && SE->isKnownNonNegative(SCEVOP2);
561 // Replace all comparisons against trunc with comparisons against IV.
562 for (auto *ICI : ICmpUsers) {
563 bool IsSwapped = L->isLoopInvariant(ICI->getOperand(0));
564 auto *Op1 = IsSwapped ? ICI->getOperand(0) : ICI->getOperand(1);
565 Instruction *Ext = nullptr;
566 // For signed/unsigned predicate, replace the old comparison with comparison
567 // of immediate IV against sext/zext of the invariant argument. If we can
568 // use either sext or zext (i.e. we are dealing with equality predicate),
569 // then prefer zext as a more canonical form.
570 // TODO: If we see a signed comparison which can be turned into unsigned,
571 // we can do it here for canonicalization purposes.
572 ICmpInst::Predicate Pred = ICI->getPredicate();
573 if (IsSwapped) Pred = ICmpInst::getSwappedPredicate(Pred);
574 if (CanUseZExt(ICI)) {
575 assert(DoesZExtCollapse && "Unprofitable zext?");
576 Ext = new ZExtInst(Op1, IVTy, "zext", ICI);
577 Pred = ICmpInst::getUnsignedPredicate(Pred);
579 assert(DoesSExtCollapse && "Unprofitable sext?");
580 Ext = new SExtInst(Op1, IVTy, "sext", ICI);
581 assert(Pred == ICmpInst::getSignedPredicate(Pred) && "Must be signed!");
584 L->makeLoopInvariant(Ext, Changed);
586 ICmpInst *NewICI = new ICmpInst(ICI, Pred, IV, Ext);
587 ICI->replaceAllUsesWith(NewICI);
588 DeadInsts.emplace_back(ICI);
591 // Trunc no longer needed.
592 TI->replaceAllUsesWith(UndefValue::get(TI->getType()));
593 DeadInsts.emplace_back(TI);
597 /// Eliminate an operation that consumes a simple IV and has no observable
598 /// side-effect given the range of IV values. IVOperand is guaranteed SCEVable,
599 /// but UseInst may not be.
600 bool SimplifyIndvar::eliminateIVUser(Instruction *UseInst,
601 Instruction *IVOperand) {
602 if (ICmpInst *ICmp = dyn_cast<ICmpInst>(UseInst)) {
603 eliminateIVComparison(ICmp, IVOperand);
606 if (BinaryOperator *Bin = dyn_cast<BinaryOperator>(UseInst)) {
607 bool IsSRem = Bin->getOpcode() == Instruction::SRem;
608 if (IsSRem || Bin->getOpcode() == Instruction::URem) {
609 simplifyIVRemainder(Bin, IVOperand, IsSRem);
613 if (Bin->getOpcode() == Instruction::SDiv)
614 return eliminateSDiv(Bin);
617 if (auto *WO = dyn_cast<WithOverflowInst>(UseInst))
618 if (eliminateOverflowIntrinsic(WO))
621 if (auto *SI = dyn_cast<SaturatingInst>(UseInst))
622 if (eliminateSaturatingIntrinsic(SI))
625 if (auto *TI = dyn_cast<TruncInst>(UseInst))
626 if (eliminateTrunc(TI))
629 if (eliminateIdentitySCEV(UseInst, IVOperand))
635 static Instruction *GetLoopInvariantInsertPosition(Loop *L, Instruction *Hint) {
636 if (auto *BB = L->getLoopPreheader())
637 return BB->getTerminator();
642 /// Replace the UseInst with a loop invariant expression if it is safe.
643 bool SimplifyIndvar::replaceIVUserWithLoopInvariant(Instruction *I) {
644 if (!SE->isSCEVable(I->getType()))
647 // Get the symbolic expression for this instruction.
648 const SCEV *S = SE->getSCEV(I);
650 if (!SE->isLoopInvariant(S, L))
653 // Do not generate something ridiculous even if S is loop invariant.
654 if (Rewriter.isHighCostExpansion(S, L, SCEVCheapExpansionBudget, TTI, I))
657 auto *IP = GetLoopInvariantInsertPosition(L, I);
659 if (!isSafeToExpandAt(S, IP, *SE)) {
660 LLVM_DEBUG(dbgs() << "INDVARS: Can not replace IV user: " << *I
661 << " with non-speculable loop invariant: " << *S << '\n');
665 auto *Invariant = Rewriter.expandCodeFor(S, I->getType(), IP);
667 I->replaceAllUsesWith(Invariant);
668 LLVM_DEBUG(dbgs() << "INDVARS: Replace IV user: " << *I
669 << " with loop invariant: " << *S << '\n');
672 DeadInsts.emplace_back(I);
676 /// Eliminate any operation that SCEV can prove is an identity function.
677 bool SimplifyIndvar::eliminateIdentitySCEV(Instruction *UseInst,
678 Instruction *IVOperand) {
679 if (!SE->isSCEVable(UseInst->getType()) ||
680 (UseInst->getType() != IVOperand->getType()) ||
681 (SE->getSCEV(UseInst) != SE->getSCEV(IVOperand)))
684 // getSCEV(X) == getSCEV(Y) does not guarantee that X and Y are related in the
685 // dominator tree, even if X is an operand to Y. For instance, in
687 // %iv = phi i32 {0,+,1}
688 // br %cond, label %left, label %merge
691 // %X = add i32 %iv, 0
695 // %M = phi (%X, %iv)
697 // getSCEV(%M) == getSCEV(%X) == {0,+,1}, but %X does not dominate %M, and
698 // %M.replaceAllUsesWith(%X) would be incorrect.
700 if (isa<PHINode>(UseInst))
701 // If UseInst is not a PHI node then we know that IVOperand dominates
702 // UseInst directly from the legality of SSA.
703 if (!DT || !DT->dominates(IVOperand, UseInst))
706 if (!LI->replacementPreservesLCSSAForm(UseInst, IVOperand))
709 LLVM_DEBUG(dbgs() << "INDVARS: Eliminated identity: " << *UseInst << '\n');
711 UseInst->replaceAllUsesWith(IVOperand);
714 DeadInsts.emplace_back(UseInst);
718 /// Annotate BO with nsw / nuw if it provably does not signed-overflow /
719 /// unsigned-overflow. Returns true if anything changed, false otherwise.
720 bool SimplifyIndvar::strengthenOverflowingOperation(BinaryOperator *BO,
722 SCEV::NoWrapFlags Flags;
724 std::tie(Flags, Deduced) = SE->getStrengthenedNoWrapFlagsFromBinOp(
725 cast<OverflowingBinaryOperator>(BO));
730 BO->setHasNoUnsignedWrap(ScalarEvolution::maskFlags(Flags, SCEV::FlagNUW) ==
732 BO->setHasNoSignedWrap(ScalarEvolution::maskFlags(Flags, SCEV::FlagNSW) ==
735 // The getStrengthenedNoWrapFlagsFromBinOp() check inferred additional nowrap
736 // flags on addrecs while performing zero/sign extensions. We could call
737 // forgetValue() here to make sure those flags also propagate to any other
738 // SCEV expressions based on the addrec. However, this can have pathological
739 // compile-time impact, see https://bugs.llvm.org/show_bug.cgi?id=50384.
743 /// Annotate the Shr in (X << IVOperand) >> C as exact using the
744 /// information from the IV's range. Returns true if anything changed, false
746 bool SimplifyIndvar::strengthenRightShift(BinaryOperator *BO,
748 using namespace llvm::PatternMatch;
750 if (BO->getOpcode() == Instruction::Shl) {
751 bool Changed = false;
752 ConstantRange IVRange = SE->getUnsignedRange(SE->getSCEV(IVOperand));
753 for (auto *U : BO->users()) {
756 m_AShr(m_Shl(m_Value(), m_Specific(IVOperand)), m_APInt(C))) ||
758 m_LShr(m_Shl(m_Value(), m_Specific(IVOperand)), m_APInt(C)))) {
759 BinaryOperator *Shr = cast<BinaryOperator>(U);
760 if (!Shr->isExact() && IVRange.getUnsignedMin().uge(*C)) {
761 Shr->setIsExact(true);
772 /// Add all uses of Def to the current IV's worklist.
773 static void pushIVUsers(
774 Instruction *Def, Loop *L,
775 SmallPtrSet<Instruction*,16> &Simplified,
776 SmallVectorImpl< std::pair<Instruction*,Instruction*> > &SimpleIVUsers) {
778 for (User *U : Def->users()) {
779 Instruction *UI = cast<Instruction>(U);
781 // Avoid infinite or exponential worklist processing.
782 // Also ensure unique worklist users.
783 // If Def is a LoopPhi, it may not be in the Simplified set, so check for
788 // Only change the current Loop, do not change the other parts (e.g. other
790 if (!L->contains(UI))
793 // Do not push the same instruction more than once.
794 if (!Simplified.insert(UI).second)
797 SimpleIVUsers.push_back(std::make_pair(UI, Def));
801 /// Return true if this instruction generates a simple SCEV
802 /// expression in terms of that IV.
804 /// This is similar to IVUsers' isInteresting() but processes each instruction
805 /// non-recursively when the operand is already known to be a simpleIVUser.
807 static bool isSimpleIVUser(Instruction *I, const Loop *L, ScalarEvolution *SE) {
808 if (!SE->isSCEVable(I->getType()))
811 // Get the symbolic expression for this instruction.
812 const SCEV *S = SE->getSCEV(I);
814 // Only consider affine recurrences.
815 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S);
816 if (AR && AR->getLoop() == L)
822 /// Iteratively perform simplification on a worklist of users
823 /// of the specified induction variable. Each successive simplification may push
824 /// more users which may themselves be candidates for simplification.
826 /// This algorithm does not require IVUsers analysis. Instead, it simplifies
827 /// instructions in-place during analysis. Rather than rewriting induction
828 /// variables bottom-up from their users, it transforms a chain of IVUsers
829 /// top-down, updating the IR only when it encounters a clear optimization
832 /// Once DisableIVRewrite is default, LSR will be the only client of IVUsers.
834 void SimplifyIndvar::simplifyUsers(PHINode *CurrIV, IVVisitor *V) {
835 if (!SE->isSCEVable(CurrIV->getType()))
838 // Instructions processed by SimplifyIndvar for CurrIV.
839 SmallPtrSet<Instruction*,16> Simplified;
841 // Use-def pairs if IV users waiting to be processed for CurrIV.
842 SmallVector<std::pair<Instruction*, Instruction*>, 8> SimpleIVUsers;
844 // Push users of the current LoopPhi. In rare cases, pushIVUsers may be
845 // called multiple times for the same LoopPhi. This is the proper thing to
846 // do for loop header phis that use each other.
847 pushIVUsers(CurrIV, L, Simplified, SimpleIVUsers);
849 while (!SimpleIVUsers.empty()) {
850 std::pair<Instruction*, Instruction*> UseOper =
851 SimpleIVUsers.pop_back_val();
852 Instruction *UseInst = UseOper.first;
854 // If a user of the IndVar is trivially dead, we prefer just to mark it dead
855 // rather than try to do some complex analysis or transformation (such as
856 // widening) basing on it.
857 // TODO: Propagate TLI and pass it here to handle more cases.
858 if (isInstructionTriviallyDead(UseInst, /* TLI */ nullptr)) {
859 DeadInsts.emplace_back(UseInst);
863 // Bypass back edges to avoid extra work.
864 if (UseInst == CurrIV) continue;
866 // Try to replace UseInst with a loop invariant before any other
868 if (replaceIVUserWithLoopInvariant(UseInst))
871 Instruction *IVOperand = UseOper.second;
872 for (unsigned N = 0; IVOperand; ++N) {
873 assert(N <= Simplified.size() && "runaway iteration");
876 Value *NewOper = foldIVUser(UseInst, IVOperand);
878 break; // done folding
879 IVOperand = dyn_cast<Instruction>(NewOper);
884 if (eliminateIVUser(UseInst, IVOperand)) {
885 pushIVUsers(IVOperand, L, Simplified, SimpleIVUsers);
889 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(UseInst)) {
890 if ((isa<OverflowingBinaryOperator>(BO) &&
891 strengthenOverflowingOperation(BO, IVOperand)) ||
892 (isa<ShlOperator>(BO) && strengthenRightShift(BO, IVOperand))) {
893 // re-queue uses of the now modified binary operator and fall
894 // through to the checks that remain.
895 pushIVUsers(IVOperand, L, Simplified, SimpleIVUsers);
899 CastInst *Cast = dyn_cast<CastInst>(UseInst);
904 if (isSimpleIVUser(UseInst, L, SE)) {
905 pushIVUsers(UseInst, L, Simplified, SimpleIVUsers);
912 void IVVisitor::anchor() { }
914 /// Simplify instructions that use this induction variable
915 /// by using ScalarEvolution to analyze the IV's recurrence.
916 bool simplifyUsersOfIV(PHINode *CurrIV, ScalarEvolution *SE, DominatorTree *DT,
917 LoopInfo *LI, const TargetTransformInfo *TTI,
918 SmallVectorImpl<WeakTrackingVH> &Dead,
919 SCEVExpander &Rewriter, IVVisitor *V) {
920 SimplifyIndvar SIV(LI->getLoopFor(CurrIV->getParent()), SE, DT, LI, TTI,
922 SIV.simplifyUsers(CurrIV, V);
923 return SIV.hasChanged();
926 /// Simplify users of induction variables within this
927 /// loop. This does not actually change or add IVs.
928 bool simplifyLoopIVs(Loop *L, ScalarEvolution *SE, DominatorTree *DT,
929 LoopInfo *LI, const TargetTransformInfo *TTI,
930 SmallVectorImpl<WeakTrackingVH> &Dead) {
931 SCEVExpander Rewriter(*SE, SE->getDataLayout(), "indvars");
933 Rewriter.setDebugType(DEBUG_TYPE);
935 bool Changed = false;
936 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
938 simplifyUsersOfIV(cast<PHINode>(I), SE, DT, LI, TTI, Dead, Rewriter);
946 //===----------------------------------------------------------------------===//
947 // Widen Induction Variables - Extend the width of an IV to cover its
949 //===----------------------------------------------------------------------===//
962 // Does the module have any calls to the llvm.experimental.guard intrinsic
963 // at all? If not we can avoid scanning instructions looking for guards.
966 bool UsePostIncrementRanges;
969 unsigned NumElimExt = 0;
970 unsigned NumWidened = 0;
973 PHINode *WidePhi = nullptr;
974 Instruction *WideInc = nullptr;
975 const SCEV *WideIncExpr = nullptr;
976 SmallVectorImpl<WeakTrackingVH> &DeadInsts;
978 SmallPtrSet<Instruction *,16> Widened;
980 enum ExtendKind { ZeroExtended, SignExtended, Unknown };
982 // A map tracking the kind of extension used to widen each narrow IV
983 // and narrow IV user.
984 // Key: pointer to a narrow IV or IV user.
985 // Value: the kind of extension used to widen this Instruction.
986 DenseMap<AssertingVH<Instruction>, ExtendKind> ExtendKindMap;
988 using DefUserPair = std::pair<AssertingVH<Value>, AssertingVH<Instruction>>;
990 // A map with control-dependent ranges for post increment IV uses. The key is
991 // a pair of IV def and a use of this def denoting the context. The value is
992 // a ConstantRange representing possible values of the def at the given
994 DenseMap<DefUserPair, ConstantRange> PostIncRangeInfos;
996 Optional<ConstantRange> getPostIncRangeInfo(Value *Def,
998 DefUserPair Key(Def, UseI);
999 auto It = PostIncRangeInfos.find(Key);
1000 return It == PostIncRangeInfos.end()
1001 ? Optional<ConstantRange>(None)
1002 : Optional<ConstantRange>(It->second);
1005 void calculatePostIncRanges(PHINode *OrigPhi);
1006 void calculatePostIncRange(Instruction *NarrowDef, Instruction *NarrowUser);
1008 void updatePostIncRangeInfo(Value *Def, Instruction *UseI, ConstantRange R) {
1009 DefUserPair Key(Def, UseI);
1010 auto It = PostIncRangeInfos.find(Key);
1011 if (It == PostIncRangeInfos.end())
1012 PostIncRangeInfos.insert({Key, R});
1014 It->second = R.intersectWith(It->second);
1018 /// Record a link in the Narrow IV def-use chain along with the WideIV that
1019 /// computes the same value as the Narrow IV def. This avoids caching Use*
1021 struct NarrowIVDefUse {
1022 Instruction *NarrowDef = nullptr;
1023 Instruction *NarrowUse = nullptr;
1024 Instruction *WideDef = nullptr;
1026 // True if the narrow def is never negative. Tracking this information lets
1027 // us use a sign extension instead of a zero extension or vice versa, when
1028 // profitable and legal.
1029 bool NeverNegative = false;
1031 NarrowIVDefUse(Instruction *ND, Instruction *NU, Instruction *WD,
1033 : NarrowDef(ND), NarrowUse(NU), WideDef(WD),
1034 NeverNegative(NeverNegative) {}
1037 WidenIV(const WideIVInfo &WI, LoopInfo *LInfo, ScalarEvolution *SEv,
1038 DominatorTree *DTree, SmallVectorImpl<WeakTrackingVH> &DI,
1039 bool HasGuards, bool UsePostIncrementRanges = true);
1041 PHINode *createWideIV(SCEVExpander &Rewriter);
1043 unsigned getNumElimExt() { return NumElimExt; };
1044 unsigned getNumWidened() { return NumWidened; };
1047 Value *createExtendInst(Value *NarrowOper, Type *WideType, bool IsSigned,
1050 Instruction *cloneIVUser(NarrowIVDefUse DU, const SCEVAddRecExpr *WideAR);
1051 Instruction *cloneArithmeticIVUser(NarrowIVDefUse DU,
1052 const SCEVAddRecExpr *WideAR);
1053 Instruction *cloneBitwiseIVUser(NarrowIVDefUse DU);
1055 ExtendKind getExtendKind(Instruction *I);
1057 using WidenedRecTy = std::pair<const SCEVAddRecExpr *, ExtendKind>;
1059 WidenedRecTy getWideRecurrence(NarrowIVDefUse DU);
1061 WidenedRecTy getExtendedOperandRecurrence(NarrowIVDefUse DU);
1063 const SCEV *getSCEVByOpCode(const SCEV *LHS, const SCEV *RHS,
1064 unsigned OpCode) const;
1066 Instruction *widenIVUse(NarrowIVDefUse DU, SCEVExpander &Rewriter);
1068 bool widenLoopCompare(NarrowIVDefUse DU);
1069 bool widenWithVariantUse(NarrowIVDefUse DU);
1071 void pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef);
1074 SmallVector<NarrowIVDefUse, 8> NarrowIVUsers;
1078 /// Determine the insertion point for this user. By default, insert immediately
1079 /// before the user. SCEVExpander or LICM will hoist loop invariants out of the
1080 /// loop. For PHI nodes, there may be multiple uses, so compute the nearest
1081 /// common dominator for the incoming blocks. A nullptr can be returned if no
1082 /// viable location is found: it may happen if User is a PHI and Def only comes
1083 /// to this PHI from unreachable blocks.
1084 static Instruction *getInsertPointForUses(Instruction *User, Value *Def,
1085 DominatorTree *DT, LoopInfo *LI) {
1086 PHINode *PHI = dyn_cast<PHINode>(User);
1090 Instruction *InsertPt = nullptr;
1091 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i) {
1092 if (PHI->getIncomingValue(i) != Def)
1095 BasicBlock *InsertBB = PHI->getIncomingBlock(i);
1097 if (!DT->isReachableFromEntry(InsertBB))
1101 InsertPt = InsertBB->getTerminator();
1104 InsertBB = DT->findNearestCommonDominator(InsertPt->getParent(), InsertBB);
1105 InsertPt = InsertBB->getTerminator();
1108 // If we have skipped all inputs, it means that Def only comes to Phi from
1109 // unreachable blocks.
1113 auto *DefI = dyn_cast<Instruction>(Def);
1117 assert(DT->dominates(DefI, InsertPt) && "def does not dominate all uses");
1119 auto *L = LI->getLoopFor(DefI->getParent());
1120 assert(!L || L->contains(LI->getLoopFor(InsertPt->getParent())));
1122 for (auto *DTN = (*DT)[InsertPt->getParent()]; DTN; DTN = DTN->getIDom())
1123 if (LI->getLoopFor(DTN->getBlock()) == L)
1124 return DTN->getBlock()->getTerminator();
1126 llvm_unreachable("DefI dominates InsertPt!");
1129 WidenIV::WidenIV(const WideIVInfo &WI, LoopInfo *LInfo, ScalarEvolution *SEv,
1130 DominatorTree *DTree, SmallVectorImpl<WeakTrackingVH> &DI,
1131 bool HasGuards, bool UsePostIncrementRanges)
1132 : OrigPhi(WI.NarrowIV), WideType(WI.WidestNativeType), LI(LInfo),
1133 L(LI->getLoopFor(OrigPhi->getParent())), SE(SEv), DT(DTree),
1134 HasGuards(HasGuards), UsePostIncrementRanges(UsePostIncrementRanges),
1136 assert(L->getHeader() == OrigPhi->getParent() && "Phi must be an IV");
1137 ExtendKindMap[OrigPhi] = WI.IsSigned ? SignExtended : ZeroExtended;
1140 Value *WidenIV::createExtendInst(Value *NarrowOper, Type *WideType,
1141 bool IsSigned, Instruction *Use) {
1142 // Set the debug location and conservative insertion point.
1143 IRBuilder<> Builder(Use);
1144 // Hoist the insertion point into loop preheaders as far as possible.
1145 for (const Loop *L = LI->getLoopFor(Use->getParent());
1146 L && L->getLoopPreheader() && L->isLoopInvariant(NarrowOper);
1147 L = L->getParentLoop())
1148 Builder.SetInsertPoint(L->getLoopPreheader()->getTerminator());
1150 return IsSigned ? Builder.CreateSExt(NarrowOper, WideType) :
1151 Builder.CreateZExt(NarrowOper, WideType);
1154 /// Instantiate a wide operation to replace a narrow operation. This only needs
1155 /// to handle operations that can evaluation to SCEVAddRec. It can safely return
1156 /// 0 for any operation we decide not to clone.
1157 Instruction *WidenIV::cloneIVUser(WidenIV::NarrowIVDefUse DU,
1158 const SCEVAddRecExpr *WideAR) {
1159 unsigned Opcode = DU.NarrowUse->getOpcode();
1163 case Instruction::Add:
1164 case Instruction::Mul:
1165 case Instruction::UDiv:
1166 case Instruction::Sub:
1167 return cloneArithmeticIVUser(DU, WideAR);
1169 case Instruction::And:
1170 case Instruction::Or:
1171 case Instruction::Xor:
1172 case Instruction::Shl:
1173 case Instruction::LShr:
1174 case Instruction::AShr:
1175 return cloneBitwiseIVUser(DU);
1179 Instruction *WidenIV::cloneBitwiseIVUser(WidenIV::NarrowIVDefUse DU) {
1180 Instruction *NarrowUse = DU.NarrowUse;
1181 Instruction *NarrowDef = DU.NarrowDef;
1182 Instruction *WideDef = DU.WideDef;
1184 LLVM_DEBUG(dbgs() << "Cloning bitwise IVUser: " << *NarrowUse << "\n");
1186 // Replace NarrowDef operands with WideDef. Otherwise, we don't know anything
1187 // about the narrow operand yet so must insert a [sz]ext. It is probably loop
1188 // invariant and will be folded or hoisted. If it actually comes from a
1189 // widened IV, it should be removed during a future call to widenIVUse.
1190 bool IsSigned = getExtendKind(NarrowDef) == SignExtended;
1191 Value *LHS = (NarrowUse->getOperand(0) == NarrowDef)
1193 : createExtendInst(NarrowUse->getOperand(0), WideType,
1194 IsSigned, NarrowUse);
1195 Value *RHS = (NarrowUse->getOperand(1) == NarrowDef)
1197 : createExtendInst(NarrowUse->getOperand(1), WideType,
1198 IsSigned, NarrowUse);
1200 auto *NarrowBO = cast<BinaryOperator>(NarrowUse);
1201 auto *WideBO = BinaryOperator::Create(NarrowBO->getOpcode(), LHS, RHS,
1202 NarrowBO->getName());
1203 IRBuilder<> Builder(NarrowUse);
1204 Builder.Insert(WideBO);
1205 WideBO->copyIRFlags(NarrowBO);
1209 Instruction *WidenIV::cloneArithmeticIVUser(WidenIV::NarrowIVDefUse DU,
1210 const SCEVAddRecExpr *WideAR) {
1211 Instruction *NarrowUse = DU.NarrowUse;
1212 Instruction *NarrowDef = DU.NarrowDef;
1213 Instruction *WideDef = DU.WideDef;
1215 LLVM_DEBUG(dbgs() << "Cloning arithmetic IVUser: " << *NarrowUse << "\n");
1217 unsigned IVOpIdx = (NarrowUse->getOperand(0) == NarrowDef) ? 0 : 1;
1219 // We're trying to find X such that
1221 // Widen(NarrowDef `op` NonIVNarrowDef) == WideAR == WideDef `op.wide` X
1223 // We guess two solutions to X, sext(NonIVNarrowDef) and zext(NonIVNarrowDef),
1224 // and check using SCEV if any of them are correct.
1226 // Returns true if extending NonIVNarrowDef according to `SignExt` is a
1227 // correct solution to X.
1228 auto GuessNonIVOperand = [&](bool SignExt) {
1229 const SCEV *WideLHS;
1230 const SCEV *WideRHS;
1232 auto GetExtend = [this, SignExt](const SCEV *S, Type *Ty) {
1234 return SE->getSignExtendExpr(S, Ty);
1235 return SE->getZeroExtendExpr(S, Ty);
1239 WideLHS = SE->getSCEV(WideDef);
1240 const SCEV *NarrowRHS = SE->getSCEV(NarrowUse->getOperand(1));
1241 WideRHS = GetExtend(NarrowRHS, WideType);
1243 const SCEV *NarrowLHS = SE->getSCEV(NarrowUse->getOperand(0));
1244 WideLHS = GetExtend(NarrowLHS, WideType);
1245 WideRHS = SE->getSCEV(WideDef);
1248 // WideUse is "WideDef `op.wide` X" as described in the comment.
1249 const SCEV *WideUse =
1250 getSCEVByOpCode(WideLHS, WideRHS, NarrowUse->getOpcode());
1252 return WideUse == WideAR;
1255 bool SignExtend = getExtendKind(NarrowDef) == SignExtended;
1256 if (!GuessNonIVOperand(SignExtend)) {
1257 SignExtend = !SignExtend;
1258 if (!GuessNonIVOperand(SignExtend))
1262 Value *LHS = (NarrowUse->getOperand(0) == NarrowDef)
1264 : createExtendInst(NarrowUse->getOperand(0), WideType,
1265 SignExtend, NarrowUse);
1266 Value *RHS = (NarrowUse->getOperand(1) == NarrowDef)
1268 : createExtendInst(NarrowUse->getOperand(1), WideType,
1269 SignExtend, NarrowUse);
1271 auto *NarrowBO = cast<BinaryOperator>(NarrowUse);
1272 auto *WideBO = BinaryOperator::Create(NarrowBO->getOpcode(), LHS, RHS,
1273 NarrowBO->getName());
1275 IRBuilder<> Builder(NarrowUse);
1276 Builder.Insert(WideBO);
1277 WideBO->copyIRFlags(NarrowBO);
1281 WidenIV::ExtendKind WidenIV::getExtendKind(Instruction *I) {
1282 auto It = ExtendKindMap.find(I);
1283 assert(It != ExtendKindMap.end() && "Instruction not yet extended!");
1287 const SCEV *WidenIV::getSCEVByOpCode(const SCEV *LHS, const SCEV *RHS,
1288 unsigned OpCode) const {
1290 case Instruction::Add:
1291 return SE->getAddExpr(LHS, RHS);
1292 case Instruction::Sub:
1293 return SE->getMinusSCEV(LHS, RHS);
1294 case Instruction::Mul:
1295 return SE->getMulExpr(LHS, RHS);
1296 case Instruction::UDiv:
1297 return SE->getUDivExpr(LHS, RHS);
1299 llvm_unreachable("Unsupported opcode.");
1303 /// No-wrap operations can transfer sign extension of their result to their
1304 /// operands. Generate the SCEV value for the widened operation without
1305 /// actually modifying the IR yet. If the expression after extending the
1306 /// operands is an AddRec for this loop, return the AddRec and the kind of
1308 WidenIV::WidenedRecTy
1309 WidenIV::getExtendedOperandRecurrence(WidenIV::NarrowIVDefUse DU) {
1310 // Handle the common case of add<nsw/nuw>
1311 const unsigned OpCode = DU.NarrowUse->getOpcode();
1312 // Only Add/Sub/Mul instructions supported yet.
1313 if (OpCode != Instruction::Add && OpCode != Instruction::Sub &&
1314 OpCode != Instruction::Mul)
1315 return {nullptr, Unknown};
1317 // One operand (NarrowDef) has already been extended to WideDef. Now determine
1318 // if extending the other will lead to a recurrence.
1319 const unsigned ExtendOperIdx =
1320 DU.NarrowUse->getOperand(0) == DU.NarrowDef ? 1 : 0;
1321 assert(DU.NarrowUse->getOperand(1-ExtendOperIdx) == DU.NarrowDef && "bad DU");
1323 const SCEV *ExtendOperExpr = nullptr;
1324 const OverflowingBinaryOperator *OBO =
1325 cast<OverflowingBinaryOperator>(DU.NarrowUse);
1326 ExtendKind ExtKind = getExtendKind(DU.NarrowDef);
1327 if (ExtKind == SignExtended && OBO->hasNoSignedWrap())
1328 ExtendOperExpr = SE->getSignExtendExpr(
1329 SE->getSCEV(DU.NarrowUse->getOperand(ExtendOperIdx)), WideType);
1330 else if(ExtKind == ZeroExtended && OBO->hasNoUnsignedWrap())
1331 ExtendOperExpr = SE->getZeroExtendExpr(
1332 SE->getSCEV(DU.NarrowUse->getOperand(ExtendOperIdx)), WideType);
1334 return {nullptr, Unknown};
1336 // When creating this SCEV expr, don't apply the current operations NSW or NUW
1337 // flags. This instruction may be guarded by control flow that the no-wrap
1338 // behavior depends on. Non-control-equivalent instructions can be mapped to
1339 // the same SCEV expression, and it would be incorrect to transfer NSW/NUW
1340 // semantics to those operations.
1341 const SCEV *lhs = SE->getSCEV(DU.WideDef);
1342 const SCEV *rhs = ExtendOperExpr;
1344 // Let's swap operands to the initial order for the case of non-commutative
1345 // operations, like SUB. See PR21014.
1346 if (ExtendOperIdx == 0)
1347 std::swap(lhs, rhs);
1348 const SCEVAddRecExpr *AddRec =
1349 dyn_cast<SCEVAddRecExpr>(getSCEVByOpCode(lhs, rhs, OpCode));
1351 if (!AddRec || AddRec->getLoop() != L)
1352 return {nullptr, Unknown};
1354 return {AddRec, ExtKind};
1357 /// Is this instruction potentially interesting for further simplification after
1358 /// widening it's type? In other words, can the extend be safely hoisted out of
1359 /// the loop with SCEV reducing the value to a recurrence on the same loop. If
1360 /// so, return the extended recurrence and the kind of extension used. Otherwise
1361 /// return {nullptr, Unknown}.
1362 WidenIV::WidenedRecTy WidenIV::getWideRecurrence(WidenIV::NarrowIVDefUse DU) {
1363 if (!DU.NarrowUse->getType()->isIntegerTy())
1364 return {nullptr, Unknown};
1366 const SCEV *NarrowExpr = SE->getSCEV(DU.NarrowUse);
1367 if (SE->getTypeSizeInBits(NarrowExpr->getType()) >=
1368 SE->getTypeSizeInBits(WideType)) {
1369 // NarrowUse implicitly widens its operand. e.g. a gep with a narrow
1370 // index. So don't follow this use.
1371 return {nullptr, Unknown};
1374 const SCEV *WideExpr;
1376 if (DU.NeverNegative) {
1377 WideExpr = SE->getSignExtendExpr(NarrowExpr, WideType);
1378 if (isa<SCEVAddRecExpr>(WideExpr))
1379 ExtKind = SignExtended;
1381 WideExpr = SE->getZeroExtendExpr(NarrowExpr, WideType);
1382 ExtKind = ZeroExtended;
1384 } else if (getExtendKind(DU.NarrowDef) == SignExtended) {
1385 WideExpr = SE->getSignExtendExpr(NarrowExpr, WideType);
1386 ExtKind = SignExtended;
1388 WideExpr = SE->getZeroExtendExpr(NarrowExpr, WideType);
1389 ExtKind = ZeroExtended;
1391 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(WideExpr);
1392 if (!AddRec || AddRec->getLoop() != L)
1393 return {nullptr, Unknown};
1394 return {AddRec, ExtKind};
1397 /// This IV user cannot be widened. Replace this use of the original narrow IV
1398 /// with a truncation of the new wide IV to isolate and eliminate the narrow IV.
1399 static void truncateIVUse(WidenIV::NarrowIVDefUse DU, DominatorTree *DT,
1401 auto *InsertPt = getInsertPointForUses(DU.NarrowUse, DU.NarrowDef, DT, LI);
1404 LLVM_DEBUG(dbgs() << "INDVARS: Truncate IV " << *DU.WideDef << " for user "
1405 << *DU.NarrowUse << "\n");
1406 IRBuilder<> Builder(InsertPt);
1407 Value *Trunc = Builder.CreateTrunc(DU.WideDef, DU.NarrowDef->getType());
1408 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, Trunc);
1411 /// If the narrow use is a compare instruction, then widen the compare
1412 // (and possibly the other operand). The extend operation is hoisted into the
1413 // loop preheader as far as possible.
1414 bool WidenIV::widenLoopCompare(WidenIV::NarrowIVDefUse DU) {
1415 ICmpInst *Cmp = dyn_cast<ICmpInst>(DU.NarrowUse);
1419 // We can legally widen the comparison in the following two cases:
1421 // - The signedness of the IV extension and comparison match
1423 // - The narrow IV is always positive (and thus its sign extension is equal
1424 // to its zero extension). For instance, let's say we're zero extending
1425 // %narrow for the following use
1427 // icmp slt i32 %narrow, %val ... (A)
1429 // and %narrow is always positive. Then
1431 // (A) == icmp slt i32 sext(%narrow), sext(%val)
1432 // == icmp slt i32 zext(%narrow), sext(%val)
1433 bool IsSigned = getExtendKind(DU.NarrowDef) == SignExtended;
1434 if (!(DU.NeverNegative || IsSigned == Cmp->isSigned()))
1437 Value *Op = Cmp->getOperand(Cmp->getOperand(0) == DU.NarrowDef ? 1 : 0);
1438 unsigned CastWidth = SE->getTypeSizeInBits(Op->getType());
1439 unsigned IVWidth = SE->getTypeSizeInBits(WideType);
1440 assert(CastWidth <= IVWidth && "Unexpected width while widening compare.");
1442 // Widen the compare instruction.
1443 auto *InsertPt = getInsertPointForUses(DU.NarrowUse, DU.NarrowDef, DT, LI);
1446 IRBuilder<> Builder(InsertPt);
1447 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, DU.WideDef);
1449 // Widen the other operand of the compare, if necessary.
1450 if (CastWidth < IVWidth) {
1451 Value *ExtOp = createExtendInst(Op, WideType, Cmp->isSigned(), Cmp);
1452 DU.NarrowUse->replaceUsesOfWith(Op, ExtOp);
1457 // The widenIVUse avoids generating trunc by evaluating the use as AddRec, this
1458 // will not work when:
1459 // 1) SCEV traces back to an instruction inside the loop that SCEV can not
1460 // expand, eg. add %indvar, (load %addr)
1461 // 2) SCEV finds a loop variant, eg. add %indvar, %loopvariant
1462 // While SCEV fails to avoid trunc, we can still try to use instruction
1463 // combining approach to prove trunc is not required. This can be further
1464 // extended with other instruction combining checks, but for now we handle the
1465 // following case (sub can be "add" and "mul", "nsw + sext" can be "nus + zext")
1468 // %c = sub nsw %b, %indvar
1469 // %d = sext %c to i64
1471 // %indvar.ext1 = sext %indvar to i64
1472 // %m = sext %b to i64
1473 // %d = sub nsw i64 %m, %indvar.ext1
1474 // Therefore, as long as the result of add/sub/mul is extended to wide type, no
1475 // trunc is required regardless of how %b is generated. This pattern is common
1476 // when calculating address in 64 bit architecture
1477 bool WidenIV::widenWithVariantUse(WidenIV::NarrowIVDefUse DU) {
1478 Instruction *NarrowUse = DU.NarrowUse;
1479 Instruction *NarrowDef = DU.NarrowDef;
1480 Instruction *WideDef = DU.WideDef;
1482 // Handle the common case of add<nsw/nuw>
1483 const unsigned OpCode = NarrowUse->getOpcode();
1484 // Only Add/Sub/Mul instructions are supported.
1485 if (OpCode != Instruction::Add && OpCode != Instruction::Sub &&
1486 OpCode != Instruction::Mul)
1489 // The operand that is not defined by NarrowDef of DU. Let's call it the
1491 assert((NarrowUse->getOperand(0) == NarrowDef ||
1492 NarrowUse->getOperand(1) == NarrowDef) &&
1495 const OverflowingBinaryOperator *OBO =
1496 cast<OverflowingBinaryOperator>(NarrowUse);
1497 ExtendKind ExtKind = getExtendKind(NarrowDef);
1498 bool CanSignExtend = ExtKind == SignExtended && OBO->hasNoSignedWrap();
1499 bool CanZeroExtend = ExtKind == ZeroExtended && OBO->hasNoUnsignedWrap();
1500 auto AnotherOpExtKind = ExtKind;
1502 // Check that all uses are either:
1503 // - narrow def (in case of we are widening the IV increment);
1504 // - single-input LCSSA Phis;
1505 // - comparison of the chosen type;
1506 // - extend of the chosen type (raison d'etre).
1507 SmallVector<Instruction *, 4> ExtUsers;
1508 SmallVector<PHINode *, 4> LCSSAPhiUsers;
1509 SmallVector<ICmpInst *, 4> ICmpUsers;
1510 for (Use &U : NarrowUse->uses()) {
1511 Instruction *User = cast<Instruction>(U.getUser());
1512 if (User == NarrowDef)
1514 if (!L->contains(User)) {
1515 auto *LCSSAPhi = cast<PHINode>(User);
1516 // Make sure there is only 1 input, so that we don't have to split
1518 if (LCSSAPhi->getNumOperands() != 1)
1520 LCSSAPhiUsers.push_back(LCSSAPhi);
1523 if (auto *ICmp = dyn_cast<ICmpInst>(User)) {
1524 auto Pred = ICmp->getPredicate();
1525 // We have 3 types of predicates: signed, unsigned and equality
1526 // predicates. For equality, it's legal to widen icmp for either sign and
1527 // zero extend. For sign extend, we can also do so for signed predicates,
1528 // likeweise for zero extend we can widen icmp for unsigned predicates.
1529 if (ExtKind == ZeroExtended && ICmpInst::isSigned(Pred))
1531 if (ExtKind == SignExtended && ICmpInst::isUnsigned(Pred))
1533 ICmpUsers.push_back(ICmp);
1536 if (ExtKind == SignExtended)
1537 User = dyn_cast<SExtInst>(User);
1539 User = dyn_cast<ZExtInst>(User);
1540 if (!User || User->getType() != WideType)
1542 ExtUsers.push_back(User);
1544 if (ExtUsers.empty()) {
1545 DeadInsts.emplace_back(NarrowUse);
1549 // We'll prove some facts that should be true in the context of ext users. If
1550 // there is no users, we are done now. If there are some, pick their common
1551 // dominator as context.
1552 const Instruction *CtxI = findCommonDominator(ExtUsers, *DT);
1554 if (!CanSignExtend && !CanZeroExtend) {
1555 // Because InstCombine turns 'sub nuw' to 'add' losing the no-wrap flag, we
1556 // will most likely not see it. Let's try to prove it.
1557 if (OpCode != Instruction::Add)
1559 if (ExtKind != ZeroExtended)
1561 const SCEV *LHS = SE->getSCEV(OBO->getOperand(0));
1562 const SCEV *RHS = SE->getSCEV(OBO->getOperand(1));
1563 // TODO: Support case for NarrowDef = NarrowUse->getOperand(1).
1564 if (NarrowUse->getOperand(0) != NarrowDef)
1566 if (!SE->isKnownNegative(RHS))
1568 bool ProvedSubNUW = SE->isKnownPredicateAt(ICmpInst::ICMP_UGE, LHS,
1569 SE->getNegativeSCEV(RHS), CtxI);
1572 // In fact, our 'add' is 'sub nuw'. We will need to widen the 2nd operand as
1573 // neg(zext(neg(op))), which is basically sext(op).
1574 AnotherOpExtKind = SignExtended;
1577 // Verifying that Defining operand is an AddRec
1578 const SCEV *Op1 = SE->getSCEV(WideDef);
1579 const SCEVAddRecExpr *AddRecOp1 = dyn_cast<SCEVAddRecExpr>(Op1);
1580 if (!AddRecOp1 || AddRecOp1->getLoop() != L)
1583 LLVM_DEBUG(dbgs() << "Cloning arithmetic IVUser: " << *NarrowUse << "\n");
1585 // Generating a widening use instruction.
1586 Value *LHS = (NarrowUse->getOperand(0) == NarrowDef)
1588 : createExtendInst(NarrowUse->getOperand(0), WideType,
1589 AnotherOpExtKind, NarrowUse);
1590 Value *RHS = (NarrowUse->getOperand(1) == NarrowDef)
1592 : createExtendInst(NarrowUse->getOperand(1), WideType,
1593 AnotherOpExtKind, NarrowUse);
1595 auto *NarrowBO = cast<BinaryOperator>(NarrowUse);
1596 auto *WideBO = BinaryOperator::Create(NarrowBO->getOpcode(), LHS, RHS,
1597 NarrowBO->getName());
1598 IRBuilder<> Builder(NarrowUse);
1599 Builder.Insert(WideBO);
1600 WideBO->copyIRFlags(NarrowBO);
1601 ExtendKindMap[NarrowUse] = ExtKind;
1603 for (Instruction *User : ExtUsers) {
1604 assert(User->getType() == WideType && "Checked before!");
1605 LLVM_DEBUG(dbgs() << "INDVARS: eliminating " << *User << " replaced by "
1606 << *WideBO << "\n");
1608 User->replaceAllUsesWith(WideBO);
1609 DeadInsts.emplace_back(User);
1612 for (PHINode *User : LCSSAPhiUsers) {
1613 assert(User->getNumOperands() == 1 && "Checked before!");
1614 Builder.SetInsertPoint(User);
1616 Builder.CreatePHI(WideBO->getType(), 1, User->getName() + ".wide");
1617 BasicBlock *LoopExitingBlock = User->getParent()->getSinglePredecessor();
1618 assert(LoopExitingBlock && L->contains(LoopExitingBlock) &&
1619 "Not a LCSSA Phi?");
1620 WidePN->addIncoming(WideBO, LoopExitingBlock);
1621 Builder.SetInsertPoint(&*User->getParent()->getFirstInsertionPt());
1622 auto *TruncPN = Builder.CreateTrunc(WidePN, User->getType());
1623 User->replaceAllUsesWith(TruncPN);
1624 DeadInsts.emplace_back(User);
1627 for (ICmpInst *User : ICmpUsers) {
1628 Builder.SetInsertPoint(User);
1629 auto ExtendedOp = [&](Value * V)->Value * {
1632 if (ExtKind == ZeroExtended)
1633 return Builder.CreateZExt(V, WideBO->getType());
1635 return Builder.CreateSExt(V, WideBO->getType());
1637 auto Pred = User->getPredicate();
1638 auto *LHS = ExtendedOp(User->getOperand(0));
1639 auto *RHS = ExtendedOp(User->getOperand(1));
1641 Builder.CreateICmp(Pred, LHS, RHS, User->getName() + ".wide");
1642 User->replaceAllUsesWith(WideCmp);
1643 DeadInsts.emplace_back(User);
1649 /// Determine whether an individual user of the narrow IV can be widened. If so,
1650 /// return the wide clone of the user.
1651 Instruction *WidenIV::widenIVUse(WidenIV::NarrowIVDefUse DU, SCEVExpander &Rewriter) {
1652 assert(ExtendKindMap.count(DU.NarrowDef) &&
1653 "Should already know the kind of extension used to widen NarrowDef");
1655 // Stop traversing the def-use chain at inner-loop phis or post-loop phis.
1656 if (PHINode *UsePhi = dyn_cast<PHINode>(DU.NarrowUse)) {
1657 if (LI->getLoopFor(UsePhi->getParent()) != L) {
1658 // For LCSSA phis, sink the truncate outside the loop.
1659 // After SimplifyCFG most loop exit targets have a single predecessor.
1660 // Otherwise fall back to a truncate within the loop.
1661 if (UsePhi->getNumOperands() != 1)
1662 truncateIVUse(DU, DT, LI);
1664 // Widening the PHI requires us to insert a trunc. The logical place
1665 // for this trunc is in the same BB as the PHI. This is not possible if
1666 // the BB is terminated by a catchswitch.
1667 if (isa<CatchSwitchInst>(UsePhi->getParent()->getTerminator()))
1671 PHINode::Create(DU.WideDef->getType(), 1, UsePhi->getName() + ".wide",
1673 WidePhi->addIncoming(DU.WideDef, UsePhi->getIncomingBlock(0));
1674 IRBuilder<> Builder(&*WidePhi->getParent()->getFirstInsertionPt());
1675 Value *Trunc = Builder.CreateTrunc(WidePhi, DU.NarrowDef->getType());
1676 UsePhi->replaceAllUsesWith(Trunc);
1677 DeadInsts.emplace_back(UsePhi);
1678 LLVM_DEBUG(dbgs() << "INDVARS: Widen lcssa phi " << *UsePhi << " to "
1679 << *WidePhi << "\n");
1685 // This narrow use can be widened by a sext if it's non-negative or its narrow
1686 // def was widended by a sext. Same for zext.
1687 auto canWidenBySExt = [&]() {
1688 return DU.NeverNegative || getExtendKind(DU.NarrowDef) == SignExtended;
1690 auto canWidenByZExt = [&]() {
1691 return DU.NeverNegative || getExtendKind(DU.NarrowDef) == ZeroExtended;
1694 // Our raison d'etre! Eliminate sign and zero extension.
1695 if ((isa<SExtInst>(DU.NarrowUse) && canWidenBySExt()) ||
1696 (isa<ZExtInst>(DU.NarrowUse) && canWidenByZExt())) {
1697 Value *NewDef = DU.WideDef;
1698 if (DU.NarrowUse->getType() != WideType) {
1699 unsigned CastWidth = SE->getTypeSizeInBits(DU.NarrowUse->getType());
1700 unsigned IVWidth = SE->getTypeSizeInBits(WideType);
1701 if (CastWidth < IVWidth) {
1702 // The cast isn't as wide as the IV, so insert a Trunc.
1703 IRBuilder<> Builder(DU.NarrowUse);
1704 NewDef = Builder.CreateTrunc(DU.WideDef, DU.NarrowUse->getType());
1707 // A wider extend was hidden behind a narrower one. This may induce
1708 // another round of IV widening in which the intermediate IV becomes
1709 // dead. It should be very rare.
1710 LLVM_DEBUG(dbgs() << "INDVARS: New IV " << *WidePhi
1711 << " not wide enough to subsume " << *DU.NarrowUse
1713 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, DU.WideDef);
1714 NewDef = DU.NarrowUse;
1717 if (NewDef != DU.NarrowUse) {
1718 LLVM_DEBUG(dbgs() << "INDVARS: eliminating " << *DU.NarrowUse
1719 << " replaced by " << *DU.WideDef << "\n");
1721 DU.NarrowUse->replaceAllUsesWith(NewDef);
1722 DeadInsts.emplace_back(DU.NarrowUse);
1724 // Now that the extend is gone, we want to expose it's uses for potential
1725 // further simplification. We don't need to directly inform SimplifyIVUsers
1726 // of the new users, because their parent IV will be processed later as a
1727 // new loop phi. If we preserved IVUsers analysis, we would also want to
1728 // push the uses of WideDef here.
1730 // No further widening is needed. The deceased [sz]ext had done it for us.
1734 // Does this user itself evaluate to a recurrence after widening?
1735 WidenedRecTy WideAddRec = getExtendedOperandRecurrence(DU);
1736 if (!WideAddRec.first)
1737 WideAddRec = getWideRecurrence(DU);
1739 assert((WideAddRec.first == nullptr) == (WideAddRec.second == Unknown));
1740 if (!WideAddRec.first) {
1741 // If use is a loop condition, try to promote the condition instead of
1742 // truncating the IV first.
1743 if (widenLoopCompare(DU))
1746 // We are here about to generate a truncate instruction that may hurt
1747 // performance because the scalar evolution expression computed earlier
1748 // in WideAddRec.first does not indicate a polynomial induction expression.
1749 // In that case, look at the operands of the use instruction to determine
1750 // if we can still widen the use instead of truncating its operand.
1751 if (widenWithVariantUse(DU))
1754 // This user does not evaluate to a recurrence after widening, so don't
1755 // follow it. Instead insert a Trunc to kill off the original use,
1756 // eventually isolating the original narrow IV so it can be removed.
1757 truncateIVUse(DU, DT, LI);
1761 // Reuse the IV increment that SCEVExpander created as long as it dominates
1763 Instruction *WideUse = nullptr;
1764 if (WideAddRec.first == WideIncExpr &&
1765 Rewriter.hoistIVInc(WideInc, DU.NarrowUse))
1768 WideUse = cloneIVUser(DU, WideAddRec.first);
1772 // Evaluation of WideAddRec ensured that the narrow expression could be
1773 // extended outside the loop without overflow. This suggests that the wide use
1774 // evaluates to the same expression as the extended narrow use, but doesn't
1775 // absolutely guarantee it. Hence the following failsafe check. In rare cases
1776 // where it fails, we simply throw away the newly created wide use.
1777 if (WideAddRec.first != SE->getSCEV(WideUse)) {
1778 LLVM_DEBUG(dbgs() << "Wide use expression mismatch: " << *WideUse << ": "
1779 << *SE->getSCEV(WideUse) << " != " << *WideAddRec.first
1781 DeadInsts.emplace_back(WideUse);
1785 // if we reached this point then we are going to replace
1786 // DU.NarrowUse with WideUse. Reattach DbgValue then.
1787 replaceAllDbgUsesWith(*DU.NarrowUse, *WideUse, *WideUse, *DT);
1789 ExtendKindMap[DU.NarrowUse] = WideAddRec.second;
1790 // Returning WideUse pushes it on the worklist.
1794 /// Add eligible users of NarrowDef to NarrowIVUsers.
1795 void WidenIV::pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef) {
1796 const SCEV *NarrowSCEV = SE->getSCEV(NarrowDef);
1797 bool NonNegativeDef =
1798 SE->isKnownPredicate(ICmpInst::ICMP_SGE, NarrowSCEV,
1799 SE->getZero(NarrowSCEV->getType()));
1800 for (User *U : NarrowDef->users()) {
1801 Instruction *NarrowUser = cast<Instruction>(U);
1803 // Handle data flow merges and bizarre phi cycles.
1804 if (!Widened.insert(NarrowUser).second)
1807 bool NonNegativeUse = false;
1808 if (!NonNegativeDef) {
1809 // We might have a control-dependent range information for this context.
1810 if (auto RangeInfo = getPostIncRangeInfo(NarrowDef, NarrowUser))
1811 NonNegativeUse = RangeInfo->getSignedMin().isNonNegative();
1814 NarrowIVUsers.emplace_back(NarrowDef, NarrowUser, WideDef,
1815 NonNegativeDef || NonNegativeUse);
1819 /// Process a single induction variable. First use the SCEVExpander to create a
1820 /// wide induction variable that evaluates to the same recurrence as the
1821 /// original narrow IV. Then use a worklist to forward traverse the narrow IV's
1822 /// def-use chain. After widenIVUse has processed all interesting IV users, the
1823 /// narrow IV will be isolated for removal by DeleteDeadPHIs.
1825 /// It would be simpler to delete uses as they are processed, but we must avoid
1826 /// invalidating SCEV expressions.
1827 PHINode *WidenIV::createWideIV(SCEVExpander &Rewriter) {
1828 // Is this phi an induction variable?
1829 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(OrigPhi));
1833 // Widen the induction variable expression.
1834 const SCEV *WideIVExpr = getExtendKind(OrigPhi) == SignExtended
1835 ? SE->getSignExtendExpr(AddRec, WideType)
1836 : SE->getZeroExtendExpr(AddRec, WideType);
1838 assert(SE->getEffectiveSCEVType(WideIVExpr->getType()) == WideType &&
1839 "Expect the new IV expression to preserve its type");
1841 // Can the IV be extended outside the loop without overflow?
1842 AddRec = dyn_cast<SCEVAddRecExpr>(WideIVExpr);
1843 if (!AddRec || AddRec->getLoop() != L)
1846 // An AddRec must have loop-invariant operands. Since this AddRec is
1847 // materialized by a loop header phi, the expression cannot have any post-loop
1848 // operands, so they must dominate the loop header.
1850 SE->properlyDominates(AddRec->getStart(), L->getHeader()) &&
1851 SE->properlyDominates(AddRec->getStepRecurrence(*SE), L->getHeader()) &&
1852 "Loop header phi recurrence inputs do not dominate the loop");
1854 // Iterate over IV uses (including transitive ones) looking for IV increments
1855 // of the form 'add nsw %iv, <const>'. For each increment and each use of
1856 // the increment calculate control-dependent range information basing on
1857 // dominating conditions inside of the loop (e.g. a range check inside of the
1858 // loop). Calculated ranges are stored in PostIncRangeInfos map.
1860 // Control-dependent range information is later used to prove that a narrow
1861 // definition is not negative (see pushNarrowIVUsers). It's difficult to do
1862 // this on demand because when pushNarrowIVUsers needs this information some
1863 // of the dominating conditions might be already widened.
1864 if (UsePostIncrementRanges)
1865 calculatePostIncRanges(OrigPhi);
1867 // The rewriter provides a value for the desired IV expression. This may
1868 // either find an existing phi or materialize a new one. Either way, we
1869 // expect a well-formed cyclic phi-with-increments. i.e. any operand not part
1870 // of the phi-SCC dominates the loop entry.
1871 Instruction *InsertPt = &*L->getHeader()->getFirstInsertionPt();
1872 Value *ExpandInst = Rewriter.expandCodeFor(AddRec, WideType, InsertPt);
1873 // If the wide phi is not a phi node, for example a cast node, like bitcast,
1874 // inttoptr, ptrtoint, just skip for now.
1875 if (!(WidePhi = dyn_cast<PHINode>(ExpandInst))) {
1876 // if the cast node is an inserted instruction without any user, we should
1877 // remove it to make sure the pass don't touch the function as we can not
1879 if (ExpandInst->hasNUses(0) &&
1880 Rewriter.isInsertedInstruction(cast<Instruction>(ExpandInst)))
1881 DeadInsts.emplace_back(ExpandInst);
1885 // Remembering the WideIV increment generated by SCEVExpander allows
1886 // widenIVUse to reuse it when widening the narrow IV's increment. We don't
1887 // employ a general reuse mechanism because the call above is the only call to
1888 // SCEVExpander. Henceforth, we produce 1-to-1 narrow to wide uses.
1889 if (BasicBlock *LatchBlock = L->getLoopLatch()) {
1891 cast<Instruction>(WidePhi->getIncomingValueForBlock(LatchBlock));
1892 WideIncExpr = SE->getSCEV(WideInc);
1893 // Propagate the debug location associated with the original loop increment
1894 // to the new (widened) increment.
1896 cast<Instruction>(OrigPhi->getIncomingValueForBlock(LatchBlock));
1897 WideInc->setDebugLoc(OrigInc->getDebugLoc());
1900 LLVM_DEBUG(dbgs() << "Wide IV: " << *WidePhi << "\n");
1903 // Traverse the def-use chain using a worklist starting at the original IV.
1904 assert(Widened.empty() && NarrowIVUsers.empty() && "expect initial state" );
1906 Widened.insert(OrigPhi);
1907 pushNarrowIVUsers(OrigPhi, WidePhi);
1909 while (!NarrowIVUsers.empty()) {
1910 WidenIV::NarrowIVDefUse DU = NarrowIVUsers.pop_back_val();
1912 // Process a def-use edge. This may replace the use, so don't hold a
1913 // use_iterator across it.
1914 Instruction *WideUse = widenIVUse(DU, Rewriter);
1916 // Follow all def-use edges from the previous narrow use.
1918 pushNarrowIVUsers(DU.NarrowUse, WideUse);
1920 // widenIVUse may have removed the def-use edge.
1921 if (DU.NarrowDef->use_empty())
1922 DeadInsts.emplace_back(DU.NarrowDef);
1925 // Attach any debug information to the new PHI.
1926 replaceAllDbgUsesWith(*OrigPhi, *WidePhi, *WidePhi, *DT);
1931 /// Calculates control-dependent range for the given def at the given context
1932 /// by looking at dominating conditions inside of the loop
1933 void WidenIV::calculatePostIncRange(Instruction *NarrowDef,
1934 Instruction *NarrowUser) {
1935 using namespace llvm::PatternMatch;
1937 Value *NarrowDefLHS;
1938 const APInt *NarrowDefRHS;
1939 if (!match(NarrowDef, m_NSWAdd(m_Value(NarrowDefLHS),
1940 m_APInt(NarrowDefRHS))) ||
1941 !NarrowDefRHS->isNonNegative())
1944 auto UpdateRangeFromCondition = [&] (Value *Condition,
1946 CmpInst::Predicate Pred;
1948 if (!match(Condition, m_ICmp(Pred, m_Specific(NarrowDefLHS),
1952 CmpInst::Predicate P =
1953 TrueDest ? Pred : CmpInst::getInversePredicate(Pred);
1955 auto CmpRHSRange = SE->getSignedRange(SE->getSCEV(CmpRHS));
1956 auto CmpConstrainedLHSRange =
1957 ConstantRange::makeAllowedICmpRegion(P, CmpRHSRange);
1958 auto NarrowDefRange = CmpConstrainedLHSRange.addWithNoWrap(
1959 *NarrowDefRHS, OverflowingBinaryOperator::NoSignedWrap);
1961 updatePostIncRangeInfo(NarrowDef, NarrowUser, NarrowDefRange);
1964 auto UpdateRangeFromGuards = [&](Instruction *Ctx) {
1968 for (Instruction &I : make_range(Ctx->getIterator().getReverse(),
1969 Ctx->getParent()->rend())) {
1971 if (match(&I, m_Intrinsic<Intrinsic::experimental_guard>(m_Value(C))))
1972 UpdateRangeFromCondition(C, /*TrueDest=*/true);
1976 UpdateRangeFromGuards(NarrowUser);
1978 BasicBlock *NarrowUserBB = NarrowUser->getParent();
1979 // If NarrowUserBB is statically unreachable asking dominator queries may
1980 // yield surprising results. (e.g. the block may not have a dom tree node)
1981 if (!DT->isReachableFromEntry(NarrowUserBB))
1984 for (auto *DTB = (*DT)[NarrowUserBB]->getIDom();
1985 L->contains(DTB->getBlock());
1986 DTB = DTB->getIDom()) {
1987 auto *BB = DTB->getBlock();
1988 auto *TI = BB->getTerminator();
1989 UpdateRangeFromGuards(TI);
1991 auto *BI = dyn_cast<BranchInst>(TI);
1992 if (!BI || !BI->isConditional())
1995 auto *TrueSuccessor = BI->getSuccessor(0);
1996 auto *FalseSuccessor = BI->getSuccessor(1);
1998 auto DominatesNarrowUser = [this, NarrowUser] (BasicBlockEdge BBE) {
1999 return BBE.isSingleEdge() &&
2000 DT->dominates(BBE, NarrowUser->getParent());
2003 if (DominatesNarrowUser(BasicBlockEdge(BB, TrueSuccessor)))
2004 UpdateRangeFromCondition(BI->getCondition(), /*TrueDest=*/true);
2006 if (DominatesNarrowUser(BasicBlockEdge(BB, FalseSuccessor)))
2007 UpdateRangeFromCondition(BI->getCondition(), /*TrueDest=*/false);
2011 /// Calculates PostIncRangeInfos map for the given IV
2012 void WidenIV::calculatePostIncRanges(PHINode *OrigPhi) {
2013 SmallPtrSet<Instruction *, 16> Visited;
2014 SmallVector<Instruction *, 6> Worklist;
2015 Worklist.push_back(OrigPhi);
2016 Visited.insert(OrigPhi);
2018 while (!Worklist.empty()) {
2019 Instruction *NarrowDef = Worklist.pop_back_val();
2021 for (Use &U : NarrowDef->uses()) {
2022 auto *NarrowUser = cast<Instruction>(U.getUser());
2024 // Don't go looking outside the current loop.
2025 auto *NarrowUserLoop = (*LI)[NarrowUser->getParent()];
2026 if (!NarrowUserLoop || !L->contains(NarrowUserLoop))
2029 if (!Visited.insert(NarrowUser).second)
2032 Worklist.push_back(NarrowUser);
2034 calculatePostIncRange(NarrowDef, NarrowUser);
2039 PHINode *llvm::createWideIV(const WideIVInfo &WI,
2040 LoopInfo *LI, ScalarEvolution *SE, SCEVExpander &Rewriter,
2041 DominatorTree *DT, SmallVectorImpl<WeakTrackingVH> &DeadInsts,
2042 unsigned &NumElimExt, unsigned &NumWidened,
2043 bool HasGuards, bool UsePostIncrementRanges) {
2044 WidenIV Widener(WI, LI, SE, DT, DeadInsts, HasGuards, UsePostIncrementRanges);
2045 PHINode *WidePHI = Widener.createWideIV(Rewriter);
2046 NumElimExt = Widener.getNumElimExt();
2047 NumWidened = Widener.getNumWidened();