1 //===-- SimplifyIndVar.cpp - Induction variable simplification ------------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
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.
13 //
14 //===----------------------------------------------------------------------===//
15 
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/ScalarEvolutionExpander.h"
22 #include "llvm/IR/DataLayout.h"
23 #include "llvm/IR/Dominators.h"
24 #include "llvm/IR/IRBuilder.h"
25 #include "llvm/IR/Instructions.h"
26 #include "llvm/IR/PatternMatch.h"
27 #include "llvm/Support/Debug.h"
28 #include "llvm/Support/raw_ostream.h"
29 #include "llvm/Transforms/Utils/Local.h"
30 
31 using namespace llvm;
32 
33 #define DEBUG_TYPE "indvars"
34 
35 STATISTIC(NumElimIdentity, "Number of IV identities eliminated");
36 STATISTIC(NumElimOperand,  "Number of IV operands folded into a use");
37 STATISTIC(NumFoldedUser, "Number of IV users folded into a constant");
38 STATISTIC(NumElimRem     , "Number of IV remainder operations eliminated");
39 STATISTIC(
40     NumSimplifiedSDiv,
41     "Number of IV signed division operations converted to unsigned division");
42 STATISTIC(
43     NumSimplifiedSRem,
44     "Number of IV signed remainder operations converted to unsigned remainder");
45 STATISTIC(NumElimCmp     , "Number of IV comparisons eliminated");
46 
47 namespace {
48   /// This is a utility for simplifying induction variables
49   /// based on ScalarEvolution. It is the primary instrument of the
50   /// IndvarSimplify pass, but it may also be directly invoked to cleanup after
51   /// other loop passes that preserve SCEV.
52   class SimplifyIndvar {
53     Loop             *L;
54     LoopInfo         *LI;
55     ScalarEvolution  *SE;
56     DominatorTree    *DT;
57     SCEVExpander     &Rewriter;
58     SmallVectorImpl<WeakTrackingVH> &DeadInsts;
59 
60     bool Changed;
61 
62   public:
SimplifyIndvar(Loop * Loop,ScalarEvolution * SE,DominatorTree * DT,LoopInfo * LI,SCEVExpander & Rewriter,SmallVectorImpl<WeakTrackingVH> & Dead)63     SimplifyIndvar(Loop *Loop, ScalarEvolution *SE, DominatorTree *DT,
64                    LoopInfo *LI, SCEVExpander &Rewriter,
65                    SmallVectorImpl<WeakTrackingVH> &Dead)
66         : L(Loop), LI(LI), SE(SE), DT(DT), Rewriter(Rewriter), DeadInsts(Dead),
67           Changed(false) {
68       assert(LI && "IV simplification requires LoopInfo");
69     }
70 
hasChanged() const71     bool hasChanged() const { return Changed; }
72 
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);
77 
78     Value *foldIVUser(Instruction *UseInst, Instruction *IVOperand);
79 
80     bool eliminateIdentitySCEV(Instruction *UseInst, Instruction *IVOperand);
81     bool replaceIVUserWithLoopInvariant(Instruction *UseInst);
82 
83     bool eliminateOverflowIntrinsic(CallInst *CI);
84     bool eliminateTrunc(TruncInst *TI);
85     bool eliminateIVUser(Instruction *UseInst, Instruction *IVOperand);
86     bool makeIVComparisonInvariant(ICmpInst *ICmp, Value *IVOperand);
87     void eliminateIVComparison(ICmpInst *ICmp, Value *IVOperand);
88     void simplifyIVRemainder(BinaryOperator *Rem, Value *IVOperand,
89                              bool IsSigned);
90     void replaceRemWithNumerator(BinaryOperator *Rem);
91     void replaceRemWithNumeratorOrZero(BinaryOperator *Rem);
92     void replaceSRemWithURem(BinaryOperator *Rem);
93     bool eliminateSDiv(BinaryOperator *SDiv);
94     bool strengthenOverflowingOperation(BinaryOperator *OBO, Value *IVOperand);
95     bool strengthenRightShift(BinaryOperator *BO, Value *IVOperand);
96   };
97 }
98 
99 /// Fold an IV operand into its use.  This removes increments of an
100 /// aligned IV when used by a instruction that ignores the low bits.
101 ///
102 /// IVOperand is guaranteed SCEVable, but UseInst may not be.
103 ///
104 /// Return the operand of IVOperand for this induction variable if IVOperand can
105 /// be folded (in case more folding opportunities have been exposed).
106 /// Otherwise return null.
foldIVUser(Instruction * UseInst,Instruction * IVOperand)107 Value *SimplifyIndvar::foldIVUser(Instruction *UseInst, Instruction *IVOperand) {
108   Value *IVSrc = nullptr;
109   unsigned OperIdx = 0;
110   const SCEV *FoldedExpr = nullptr;
111   switch (UseInst->getOpcode()) {
112   default:
113     return nullptr;
114   case Instruction::UDiv:
115   case Instruction::LShr:
116     // We're only interested in the case where we know something about
117     // the numerator and have a constant denominator.
118     if (IVOperand != UseInst->getOperand(OperIdx) ||
119         !isa<ConstantInt>(UseInst->getOperand(1)))
120       return nullptr;
121 
122     // Attempt to fold a binary operator with constant operand.
123     // e.g. ((I + 1) >> 2) => I >> 2
124     if (!isa<BinaryOperator>(IVOperand)
125         || !isa<ConstantInt>(IVOperand->getOperand(1)))
126       return nullptr;
127 
128     IVSrc = IVOperand->getOperand(0);
129     // IVSrc must be the (SCEVable) IV, since the other operand is const.
130     assert(SE->isSCEVable(IVSrc->getType()) && "Expect SCEVable IV operand");
131 
132     ConstantInt *D = cast<ConstantInt>(UseInst->getOperand(1));
133     if (UseInst->getOpcode() == Instruction::LShr) {
134       // Get a constant for the divisor. See createSCEV.
135       uint32_t BitWidth = cast<IntegerType>(UseInst->getType())->getBitWidth();
136       if (D->getValue().uge(BitWidth))
137         return nullptr;
138 
139       D = ConstantInt::get(UseInst->getContext(),
140                            APInt::getOneBitSet(BitWidth, D->getZExtValue()));
141     }
142     FoldedExpr = SE->getUDivExpr(SE->getSCEV(IVSrc), SE->getSCEV(D));
143   }
144   // We have something that might fold it's operand. Compare SCEVs.
145   if (!SE->isSCEVable(UseInst->getType()))
146     return nullptr;
147 
148   // Bypass the operand if SCEV can prove it has no effect.
149   if (SE->getSCEV(UseInst) != FoldedExpr)
150     return nullptr;
151 
152   LLVM_DEBUG(dbgs() << "INDVARS: Eliminated IV operand: " << *IVOperand
153                     << " -> " << *UseInst << '\n');
154 
155   UseInst->setOperand(OperIdx, IVSrc);
156   assert(SE->getSCEV(UseInst) == FoldedExpr && "bad SCEV with folded oper");
157 
158   ++NumElimOperand;
159   Changed = true;
160   if (IVOperand->use_empty())
161     DeadInsts.emplace_back(IVOperand);
162   return IVSrc;
163 }
164 
makeIVComparisonInvariant(ICmpInst * ICmp,Value * IVOperand)165 bool SimplifyIndvar::makeIVComparisonInvariant(ICmpInst *ICmp,
166                                                Value *IVOperand) {
167   unsigned IVOperIdx = 0;
168   ICmpInst::Predicate Pred = ICmp->getPredicate();
169   if (IVOperand != ICmp->getOperand(0)) {
170     // Swapped
171     assert(IVOperand == ICmp->getOperand(1) && "Can't find IVOperand");
172     IVOperIdx = 1;
173     Pred = ICmpInst::getSwappedPredicate(Pred);
174   }
175 
176   // Get the SCEVs for the ICmp operands (in the specific context of the
177   // current loop)
178   const Loop *ICmpLoop = LI->getLoopFor(ICmp->getParent());
179   const SCEV *S = SE->getSCEVAtScope(ICmp->getOperand(IVOperIdx), ICmpLoop);
180   const SCEV *X = SE->getSCEVAtScope(ICmp->getOperand(1 - IVOperIdx), ICmpLoop);
181 
182   ICmpInst::Predicate InvariantPredicate;
183   const SCEV *InvariantLHS, *InvariantRHS;
184 
185   auto *PN = dyn_cast<PHINode>(IVOperand);
186   if (!PN)
187     return false;
188   if (!SE->isLoopInvariantPredicate(Pred, S, X, L, InvariantPredicate,
189                                     InvariantLHS, InvariantRHS))
190     return false;
191 
192   // Rewrite the comparison to a loop invariant comparison if it can be done
193   // cheaply, where cheaply means "we don't need to emit any new
194   // instructions".
195 
196   SmallDenseMap<const SCEV*, Value*> CheapExpansions;
197   CheapExpansions[S] = ICmp->getOperand(IVOperIdx);
198   CheapExpansions[X] = ICmp->getOperand(1 - IVOperIdx);
199 
200   // TODO: Support multiple entry loops?  (We currently bail out of these in
201   // the IndVarSimplify pass)
202   if (auto *BB = L->getLoopPredecessor()) {
203     const int Idx = PN->getBasicBlockIndex(BB);
204     if (Idx >= 0) {
205       Value *Incoming = PN->getIncomingValue(Idx);
206       const SCEV *IncomingS = SE->getSCEV(Incoming);
207       CheapExpansions[IncomingS] = Incoming;
208     }
209   }
210   Value *NewLHS = CheapExpansions[InvariantLHS];
211   Value *NewRHS = CheapExpansions[InvariantRHS];
212 
213   if (!NewLHS)
214     if (auto *ConstLHS = dyn_cast<SCEVConstant>(InvariantLHS))
215       NewLHS = ConstLHS->getValue();
216   if (!NewRHS)
217     if (auto *ConstRHS = dyn_cast<SCEVConstant>(InvariantRHS))
218       NewRHS = ConstRHS->getValue();
219 
220   if (!NewLHS || !NewRHS)
221     // We could not find an existing value to replace either LHS or RHS.
222     // Generating new instructions has subtler tradeoffs, so avoid doing that
223     // for now.
224     return false;
225 
226   LLVM_DEBUG(dbgs() << "INDVARS: Simplified comparison: " << *ICmp << '\n');
227   ICmp->setPredicate(InvariantPredicate);
228   ICmp->setOperand(0, NewLHS);
229   ICmp->setOperand(1, NewRHS);
230   return true;
231 }
232 
233 /// SimplifyIVUsers helper for eliminating useless
234 /// comparisons against an induction variable.
eliminateIVComparison(ICmpInst * ICmp,Value * IVOperand)235 void SimplifyIndvar::eliminateIVComparison(ICmpInst *ICmp, Value *IVOperand) {
236   unsigned IVOperIdx = 0;
237   ICmpInst::Predicate Pred = ICmp->getPredicate();
238   ICmpInst::Predicate OriginalPred = Pred;
239   if (IVOperand != ICmp->getOperand(0)) {
240     // Swapped
241     assert(IVOperand == ICmp->getOperand(1) && "Can't find IVOperand");
242     IVOperIdx = 1;
243     Pred = ICmpInst::getSwappedPredicate(Pred);
244   }
245 
246   // Get the SCEVs for the ICmp operands (in the specific context of the
247   // current loop)
248   const Loop *ICmpLoop = LI->getLoopFor(ICmp->getParent());
249   const SCEV *S = SE->getSCEVAtScope(ICmp->getOperand(IVOperIdx), ICmpLoop);
250   const SCEV *X = SE->getSCEVAtScope(ICmp->getOperand(1 - IVOperIdx), ICmpLoop);
251 
252   // If the condition is always true or always false, replace it with
253   // a constant value.
254   if (SE->isKnownPredicate(Pred, S, X)) {
255     ICmp->replaceAllUsesWith(ConstantInt::getTrue(ICmp->getContext()));
256     DeadInsts.emplace_back(ICmp);
257     LLVM_DEBUG(dbgs() << "INDVARS: Eliminated comparison: " << *ICmp << '\n');
258   } else if (SE->isKnownPredicate(ICmpInst::getInversePredicate(Pred), S, X)) {
259     ICmp->replaceAllUsesWith(ConstantInt::getFalse(ICmp->getContext()));
260     DeadInsts.emplace_back(ICmp);
261     LLVM_DEBUG(dbgs() << "INDVARS: Eliminated comparison: " << *ICmp << '\n');
262   } else if (makeIVComparisonInvariant(ICmp, IVOperand)) {
263     // fallthrough to end of function
264   } else if (ICmpInst::isSigned(OriginalPred) &&
265              SE->isKnownNonNegative(S) && SE->isKnownNonNegative(X)) {
266     // If we were unable to make anything above, all we can is to canonicalize
267     // the comparison hoping that it will open the doors for other
268     // optimizations. If we find out that we compare two non-negative values,
269     // we turn the instruction's predicate to its unsigned version. Note that
270     // we cannot rely on Pred here unless we check if we have swapped it.
271     assert(ICmp->getPredicate() == OriginalPred && "Predicate changed?");
272     LLVM_DEBUG(dbgs() << "INDVARS: Turn to unsigned comparison: " << *ICmp
273                       << '\n');
274     ICmp->setPredicate(ICmpInst::getUnsignedPredicate(OriginalPred));
275   } else
276     return;
277 
278   ++NumElimCmp;
279   Changed = true;
280 }
281 
eliminateSDiv(BinaryOperator * SDiv)282 bool SimplifyIndvar::eliminateSDiv(BinaryOperator *SDiv) {
283   // Get the SCEVs for the ICmp operands.
284   auto *N = SE->getSCEV(SDiv->getOperand(0));
285   auto *D = SE->getSCEV(SDiv->getOperand(1));
286 
287   // Simplify unnecessary loops away.
288   const Loop *L = LI->getLoopFor(SDiv->getParent());
289   N = SE->getSCEVAtScope(N, L);
290   D = SE->getSCEVAtScope(D, L);
291 
292   // Replace sdiv by udiv if both of the operands are non-negative
293   if (SE->isKnownNonNegative(N) && SE->isKnownNonNegative(D)) {
294     auto *UDiv = BinaryOperator::Create(
295         BinaryOperator::UDiv, SDiv->getOperand(0), SDiv->getOperand(1),
296         SDiv->getName() + ".udiv", SDiv);
297     UDiv->setIsExact(SDiv->isExact());
298     SDiv->replaceAllUsesWith(UDiv);
299     LLVM_DEBUG(dbgs() << "INDVARS: Simplified sdiv: " << *SDiv << '\n');
300     ++NumSimplifiedSDiv;
301     Changed = true;
302     DeadInsts.push_back(SDiv);
303     return true;
304   }
305 
306   return false;
307 }
308 
309 // i %s n -> i %u n if i >= 0 and n >= 0
replaceSRemWithURem(BinaryOperator * Rem)310 void SimplifyIndvar::replaceSRemWithURem(BinaryOperator *Rem) {
311   auto *N = Rem->getOperand(0), *D = Rem->getOperand(1);
312   auto *URem = BinaryOperator::Create(BinaryOperator::URem, N, D,
313                                       Rem->getName() + ".urem", Rem);
314   Rem->replaceAllUsesWith(URem);
315   LLVM_DEBUG(dbgs() << "INDVARS: Simplified srem: " << *Rem << '\n');
316   ++NumSimplifiedSRem;
317   Changed = true;
318   DeadInsts.emplace_back(Rem);
319 }
320 
321 // i % n  -->  i  if i is in [0,n).
replaceRemWithNumerator(BinaryOperator * Rem)322 void SimplifyIndvar::replaceRemWithNumerator(BinaryOperator *Rem) {
323   Rem->replaceAllUsesWith(Rem->getOperand(0));
324   LLVM_DEBUG(dbgs() << "INDVARS: Simplified rem: " << *Rem << '\n');
325   ++NumElimRem;
326   Changed = true;
327   DeadInsts.emplace_back(Rem);
328 }
329 
330 // (i+1) % n  -->  (i+1)==n?0:(i+1)  if i is in [0,n).
replaceRemWithNumeratorOrZero(BinaryOperator * Rem)331 void SimplifyIndvar::replaceRemWithNumeratorOrZero(BinaryOperator *Rem) {
332   auto *T = Rem->getType();
333   auto *N = Rem->getOperand(0), *D = Rem->getOperand(1);
334   ICmpInst *ICmp = new ICmpInst(Rem, ICmpInst::ICMP_EQ, N, D);
335   SelectInst *Sel =
336       SelectInst::Create(ICmp, ConstantInt::get(T, 0), N, "iv.rem", Rem);
337   Rem->replaceAllUsesWith(Sel);
338   LLVM_DEBUG(dbgs() << "INDVARS: Simplified rem: " << *Rem << '\n');
339   ++NumElimRem;
340   Changed = true;
341   DeadInsts.emplace_back(Rem);
342 }
343 
344 /// SimplifyIVUsers helper for eliminating useless remainder operations
345 /// operating on an induction variable or replacing srem by urem.
simplifyIVRemainder(BinaryOperator * Rem,Value * IVOperand,bool IsSigned)346 void SimplifyIndvar::simplifyIVRemainder(BinaryOperator *Rem, Value *IVOperand,
347                                          bool IsSigned) {
348   auto *NValue = Rem->getOperand(0);
349   auto *DValue = Rem->getOperand(1);
350   // We're only interested in the case where we know something about
351   // the numerator, unless it is a srem, because we want to replace srem by urem
352   // in general.
353   bool UsedAsNumerator = IVOperand == NValue;
354   if (!UsedAsNumerator && !IsSigned)
355     return;
356 
357   const SCEV *N = SE->getSCEV(NValue);
358 
359   // Simplify unnecessary loops away.
360   const Loop *ICmpLoop = LI->getLoopFor(Rem->getParent());
361   N = SE->getSCEVAtScope(N, ICmpLoop);
362 
363   bool IsNumeratorNonNegative = !IsSigned || SE->isKnownNonNegative(N);
364 
365   // Do not proceed if the Numerator may be negative
366   if (!IsNumeratorNonNegative)
367     return;
368 
369   const SCEV *D = SE->getSCEV(DValue);
370   D = SE->getSCEVAtScope(D, ICmpLoop);
371 
372   if (UsedAsNumerator) {
373     auto LT = IsSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
374     if (SE->isKnownPredicate(LT, N, D)) {
375       replaceRemWithNumerator(Rem);
376       return;
377     }
378 
379     auto *T = Rem->getType();
380     const auto *NLessOne = SE->getMinusSCEV(N, SE->getOne(T));
381     if (SE->isKnownPredicate(LT, NLessOne, D)) {
382       replaceRemWithNumeratorOrZero(Rem);
383       return;
384     }
385   }
386 
387   // Try to replace SRem with URem, if both N and D are known non-negative.
388   // Since we had already check N, we only need to check D now
389   if (!IsSigned || !SE->isKnownNonNegative(D))
390     return;
391 
392   replaceSRemWithURem(Rem);
393 }
394 
eliminateOverflowIntrinsic(CallInst * CI)395 bool SimplifyIndvar::eliminateOverflowIntrinsic(CallInst *CI) {
396   auto *F = CI->getCalledFunction();
397   if (!F)
398     return false;
399 
400   typedef const SCEV *(ScalarEvolution::*OperationFunctionTy)(
401       const SCEV *, const SCEV *, SCEV::NoWrapFlags, unsigned);
402   typedef const SCEV *(ScalarEvolution::*ExtensionFunctionTy)(
403       const SCEV *, Type *, unsigned);
404 
405   OperationFunctionTy Operation;
406   ExtensionFunctionTy Extension;
407 
408   Instruction::BinaryOps RawOp;
409 
410   // We always have exactly one of nsw or nuw.  If NoSignedOverflow is false, we
411   // have nuw.
412   bool NoSignedOverflow;
413 
414   switch (F->getIntrinsicID()) {
415   default:
416     return false;
417 
418   case Intrinsic::sadd_with_overflow:
419     Operation = &ScalarEvolution::getAddExpr;
420     Extension = &ScalarEvolution::getSignExtendExpr;
421     RawOp = Instruction::Add;
422     NoSignedOverflow = true;
423     break;
424 
425   case Intrinsic::uadd_with_overflow:
426     Operation = &ScalarEvolution::getAddExpr;
427     Extension = &ScalarEvolution::getZeroExtendExpr;
428     RawOp = Instruction::Add;
429     NoSignedOverflow = false;
430     break;
431 
432   case Intrinsic::ssub_with_overflow:
433     Operation = &ScalarEvolution::getMinusSCEV;
434     Extension = &ScalarEvolution::getSignExtendExpr;
435     RawOp = Instruction::Sub;
436     NoSignedOverflow = true;
437     break;
438 
439   case Intrinsic::usub_with_overflow:
440     Operation = &ScalarEvolution::getMinusSCEV;
441     Extension = &ScalarEvolution::getZeroExtendExpr;
442     RawOp = Instruction::Sub;
443     NoSignedOverflow = false;
444     break;
445   }
446 
447   const SCEV *LHS = SE->getSCEV(CI->getArgOperand(0));
448   const SCEV *RHS = SE->getSCEV(CI->getArgOperand(1));
449 
450   auto *NarrowTy = cast<IntegerType>(LHS->getType());
451   auto *WideTy =
452     IntegerType::get(NarrowTy->getContext(), NarrowTy->getBitWidth() * 2);
453 
454   const SCEV *A =
455       (SE->*Extension)((SE->*Operation)(LHS, RHS, SCEV::FlagAnyWrap, 0),
456                        WideTy, 0);
457   const SCEV *B =
458       (SE->*Operation)((SE->*Extension)(LHS, WideTy, 0),
459                        (SE->*Extension)(RHS, WideTy, 0), SCEV::FlagAnyWrap, 0);
460 
461   if (A != B)
462     return false;
463 
464   // Proved no overflow, nuke the overflow check and, if possible, the overflow
465   // intrinsic as well.
466 
467   BinaryOperator *NewResult = BinaryOperator::Create(
468       RawOp, CI->getArgOperand(0), CI->getArgOperand(1), "", CI);
469 
470   if (NoSignedOverflow)
471     NewResult->setHasNoSignedWrap(true);
472   else
473     NewResult->setHasNoUnsignedWrap(true);
474 
475   SmallVector<ExtractValueInst *, 4> ToDelete;
476 
477   for (auto *U : CI->users()) {
478     if (auto *EVI = dyn_cast<ExtractValueInst>(U)) {
479       if (EVI->getIndices()[0] == 1)
480         EVI->replaceAllUsesWith(ConstantInt::getFalse(CI->getContext()));
481       else {
482         assert(EVI->getIndices()[0] == 0 && "Only two possibilities!");
483         EVI->replaceAllUsesWith(NewResult);
484       }
485       ToDelete.push_back(EVI);
486     }
487   }
488 
489   for (auto *EVI : ToDelete)
490     EVI->eraseFromParent();
491 
492   if (CI->use_empty())
493     CI->eraseFromParent();
494 
495   return true;
496 }
497 
eliminateTrunc(TruncInst * TI)498 bool SimplifyIndvar::eliminateTrunc(TruncInst *TI) {
499   // It is always legal to replace
500   //   icmp <pred> i32 trunc(iv), n
501   // with
502   //   icmp <pred> i64 sext(trunc(iv)), sext(n), if pred is signed predicate.
503   // Or with
504   //   icmp <pred> i64 zext(trunc(iv)), zext(n), if pred is unsigned predicate.
505   // Or with either of these if pred is an equality predicate.
506   //
507   // If we can prove that iv == sext(trunc(iv)) or iv == zext(trunc(iv)) for
508   // every comparison which uses trunc, it means that we can replace each of
509   // them with comparison of iv against sext/zext(n). We no longer need trunc
510   // after that.
511   //
512   // TODO: Should we do this if we can widen *some* comparisons, but not all
513   // of them? Sometimes it is enough to enable other optimizations, but the
514   // trunc instruction will stay in the loop.
515   Value *IV = TI->getOperand(0);
516   Type *IVTy = IV->getType();
517   const SCEV *IVSCEV = SE->getSCEV(IV);
518   const SCEV *TISCEV = SE->getSCEV(TI);
519 
520   // Check if iv == zext(trunc(iv)) and if iv == sext(trunc(iv)). If so, we can
521   // get rid of trunc
522   bool DoesSExtCollapse = false;
523   bool DoesZExtCollapse = false;
524   if (IVSCEV == SE->getSignExtendExpr(TISCEV, IVTy))
525     DoesSExtCollapse = true;
526   if (IVSCEV == SE->getZeroExtendExpr(TISCEV, IVTy))
527     DoesZExtCollapse = true;
528 
529   // If neither sext nor zext does collapse, it is not profitable to do any
530   // transform. Bail.
531   if (!DoesSExtCollapse && !DoesZExtCollapse)
532     return false;
533 
534   // Collect users of the trunc that look like comparisons against invariants.
535   // Bail if we find something different.
536   SmallVector<ICmpInst *, 4> ICmpUsers;
537   for (auto *U : TI->users()) {
538     // We don't care about users in unreachable blocks.
539     if (isa<Instruction>(U) &&
540         !DT->isReachableFromEntry(cast<Instruction>(U)->getParent()))
541       continue;
542     if (ICmpInst *ICI = dyn_cast<ICmpInst>(U)) {
543       if (ICI->getOperand(0) == TI && L->isLoopInvariant(ICI->getOperand(1))) {
544         assert(L->contains(ICI->getParent()) && "LCSSA form broken?");
545         // If we cannot get rid of trunc, bail.
546         if (ICI->isSigned() && !DoesSExtCollapse)
547           return false;
548         if (ICI->isUnsigned() && !DoesZExtCollapse)
549           return false;
550         // For equality, either signed or unsigned works.
551         ICmpUsers.push_back(ICI);
552       } else
553         return false;
554     } else
555       return false;
556   }
557 
558   auto CanUseZExt = [&](ICmpInst *ICI) {
559     // Unsigned comparison can be widened as unsigned.
560     if (ICI->isUnsigned())
561       return true;
562     // Is it profitable to do zext?
563     if (!DoesZExtCollapse)
564       return false;
565     // For equality, we can safely zext both parts.
566     if (ICI->isEquality())
567       return true;
568     // Otherwise we can only use zext when comparing two non-negative or two
569     // negative values. But in practice, we will never pass DoesZExtCollapse
570     // check for a negative value, because zext(trunc(x)) is non-negative. So
571     // it only make sense to check for non-negativity here.
572     const SCEV *SCEVOP1 = SE->getSCEV(ICI->getOperand(0));
573     const SCEV *SCEVOP2 = SE->getSCEV(ICI->getOperand(1));
574     return SE->isKnownNonNegative(SCEVOP1) && SE->isKnownNonNegative(SCEVOP2);
575   };
576   // Replace all comparisons against trunc with comparisons against IV.
577   for (auto *ICI : ICmpUsers) {
578     auto *Op1 = ICI->getOperand(1);
579     Instruction *Ext = nullptr;
580     // For signed/unsigned predicate, replace the old comparison with comparison
581     // of immediate IV against sext/zext of the invariant argument. If we can
582     // use either sext or zext (i.e. we are dealing with equality predicate),
583     // then prefer zext as a more canonical form.
584     // TODO: If we see a signed comparison which can be turned into unsigned,
585     // we can do it here for canonicalization purposes.
586     ICmpInst::Predicate Pred = ICI->getPredicate();
587     if (CanUseZExt(ICI)) {
588       assert(DoesZExtCollapse && "Unprofitable zext?");
589       Ext = new ZExtInst(Op1, IVTy, "zext", ICI);
590       Pred = ICmpInst::getUnsignedPredicate(Pred);
591     } else {
592       assert(DoesSExtCollapse && "Unprofitable sext?");
593       Ext = new SExtInst(Op1, IVTy, "sext", ICI);
594       assert(Pred == ICmpInst::getSignedPredicate(Pred) && "Must be signed!");
595     }
596     bool Changed;
597     L->makeLoopInvariant(Ext, Changed);
598     (void)Changed;
599     ICmpInst *NewICI = new ICmpInst(ICI, Pred, IV, Ext);
600     ICI->replaceAllUsesWith(NewICI);
601     DeadInsts.emplace_back(ICI);
602   }
603 
604   // Trunc no longer needed.
605   TI->replaceAllUsesWith(UndefValue::get(TI->getType()));
606   DeadInsts.emplace_back(TI);
607   return true;
608 }
609 
610 /// Eliminate an operation that consumes a simple IV and has no observable
611 /// side-effect given the range of IV values.  IVOperand is guaranteed SCEVable,
612 /// but UseInst may not be.
eliminateIVUser(Instruction * UseInst,Instruction * IVOperand)613 bool SimplifyIndvar::eliminateIVUser(Instruction *UseInst,
614                                      Instruction *IVOperand) {
615   if (ICmpInst *ICmp = dyn_cast<ICmpInst>(UseInst)) {
616     eliminateIVComparison(ICmp, IVOperand);
617     return true;
618   }
619   if (BinaryOperator *Bin = dyn_cast<BinaryOperator>(UseInst)) {
620     bool IsSRem = Bin->getOpcode() == Instruction::SRem;
621     if (IsSRem || Bin->getOpcode() == Instruction::URem) {
622       simplifyIVRemainder(Bin, IVOperand, IsSRem);
623       return true;
624     }
625 
626     if (Bin->getOpcode() == Instruction::SDiv)
627       return eliminateSDiv(Bin);
628   }
629 
630   if (auto *CI = dyn_cast<CallInst>(UseInst))
631     if (eliminateOverflowIntrinsic(CI))
632       return true;
633 
634   if (auto *TI = dyn_cast<TruncInst>(UseInst))
635     if (eliminateTrunc(TI))
636       return true;
637 
638   if (eliminateIdentitySCEV(UseInst, IVOperand))
639     return true;
640 
641   return false;
642 }
643 
GetLoopInvariantInsertPosition(Loop * L,Instruction * Hint)644 static Instruction *GetLoopInvariantInsertPosition(Loop *L, Instruction *Hint) {
645   if (auto *BB = L->getLoopPreheader())
646     return BB->getTerminator();
647 
648   return Hint;
649 }
650 
651 /// Replace the UseInst with a constant if possible.
replaceIVUserWithLoopInvariant(Instruction * I)652 bool SimplifyIndvar::replaceIVUserWithLoopInvariant(Instruction *I) {
653   if (!SE->isSCEVable(I->getType()))
654     return false;
655 
656   // Get the symbolic expression for this instruction.
657   const SCEV *S = SE->getSCEV(I);
658 
659   if (!SE->isLoopInvariant(S, L))
660     return false;
661 
662   // Do not generate something ridiculous even if S is loop invariant.
663   if (Rewriter.isHighCostExpansion(S, L, I))
664     return false;
665 
666   auto *IP = GetLoopInvariantInsertPosition(L, I);
667   auto *Invariant = Rewriter.expandCodeFor(S, I->getType(), IP);
668 
669   I->replaceAllUsesWith(Invariant);
670   LLVM_DEBUG(dbgs() << "INDVARS: Replace IV user: " << *I
671                     << " with loop invariant: " << *S << '\n');
672   ++NumFoldedUser;
673   Changed = true;
674   DeadInsts.emplace_back(I);
675   return true;
676 }
677 
678 /// Eliminate any operation that SCEV can prove is an identity function.
eliminateIdentitySCEV(Instruction * UseInst,Instruction * IVOperand)679 bool SimplifyIndvar::eliminateIdentitySCEV(Instruction *UseInst,
680                                            Instruction *IVOperand) {
681   if (!SE->isSCEVable(UseInst->getType()) ||
682       (UseInst->getType() != IVOperand->getType()) ||
683       (SE->getSCEV(UseInst) != SE->getSCEV(IVOperand)))
684     return false;
685 
686   // getSCEV(X) == getSCEV(Y) does not guarantee that X and Y are related in the
687   // dominator tree, even if X is an operand to Y.  For instance, in
688   //
689   //     %iv = phi i32 {0,+,1}
690   //     br %cond, label %left, label %merge
691   //
692   //   left:
693   //     %X = add i32 %iv, 0
694   //     br label %merge
695   //
696   //   merge:
697   //     %M = phi (%X, %iv)
698   //
699   // getSCEV(%M) == getSCEV(%X) == {0,+,1}, but %X does not dominate %M, and
700   // %M.replaceAllUsesWith(%X) would be incorrect.
701 
702   if (isa<PHINode>(UseInst))
703     // If UseInst is not a PHI node then we know that IVOperand dominates
704     // UseInst directly from the legality of SSA.
705     if (!DT || !DT->dominates(IVOperand, UseInst))
706       return false;
707 
708   if (!LI->replacementPreservesLCSSAForm(UseInst, IVOperand))
709     return false;
710 
711   LLVM_DEBUG(dbgs() << "INDVARS: Eliminated identity: " << *UseInst << '\n');
712 
713   UseInst->replaceAllUsesWith(IVOperand);
714   ++NumElimIdentity;
715   Changed = true;
716   DeadInsts.emplace_back(UseInst);
717   return true;
718 }
719 
720 /// Annotate BO with nsw / nuw if it provably does not signed-overflow /
721 /// unsigned-overflow.  Returns true if anything changed, false otherwise.
strengthenOverflowingOperation(BinaryOperator * BO,Value * IVOperand)722 bool SimplifyIndvar::strengthenOverflowingOperation(BinaryOperator *BO,
723                                                     Value *IVOperand) {
724 
725   // Fastpath: we don't have any work to do if `BO` is `nuw` and `nsw`.
726   if (BO->hasNoUnsignedWrap() && BO->hasNoSignedWrap())
727     return false;
728 
729   const SCEV *(ScalarEvolution::*GetExprForBO)(const SCEV *, const SCEV *,
730                                                SCEV::NoWrapFlags, unsigned);
731   switch (BO->getOpcode()) {
732   default:
733     return false;
734 
735   case Instruction::Add:
736     GetExprForBO = &ScalarEvolution::getAddExpr;
737     break;
738 
739   case Instruction::Sub:
740     GetExprForBO = &ScalarEvolution::getMinusSCEV;
741     break;
742 
743   case Instruction::Mul:
744     GetExprForBO = &ScalarEvolution::getMulExpr;
745     break;
746   }
747 
748   unsigned BitWidth = cast<IntegerType>(BO->getType())->getBitWidth();
749   Type *WideTy = IntegerType::get(BO->getContext(), BitWidth * 2);
750   const SCEV *LHS = SE->getSCEV(BO->getOperand(0));
751   const SCEV *RHS = SE->getSCEV(BO->getOperand(1));
752 
753   bool Changed = false;
754 
755   if (!BO->hasNoUnsignedWrap()) {
756     const SCEV *ExtendAfterOp = SE->getZeroExtendExpr(SE->getSCEV(BO), WideTy);
757     const SCEV *OpAfterExtend = (SE->*GetExprForBO)(
758       SE->getZeroExtendExpr(LHS, WideTy), SE->getZeroExtendExpr(RHS, WideTy),
759       SCEV::FlagAnyWrap, 0u);
760     if (ExtendAfterOp == OpAfterExtend) {
761       BO->setHasNoUnsignedWrap();
762       SE->forgetValue(BO);
763       Changed = true;
764     }
765   }
766 
767   if (!BO->hasNoSignedWrap()) {
768     const SCEV *ExtendAfterOp = SE->getSignExtendExpr(SE->getSCEV(BO), WideTy);
769     const SCEV *OpAfterExtend = (SE->*GetExprForBO)(
770       SE->getSignExtendExpr(LHS, WideTy), SE->getSignExtendExpr(RHS, WideTy),
771       SCEV::FlagAnyWrap, 0u);
772     if (ExtendAfterOp == OpAfterExtend) {
773       BO->setHasNoSignedWrap();
774       SE->forgetValue(BO);
775       Changed = true;
776     }
777   }
778 
779   return Changed;
780 }
781 
782 /// Annotate the Shr in (X << IVOperand) >> C as exact using the
783 /// information from the IV's range. Returns true if anything changed, false
784 /// otherwise.
strengthenRightShift(BinaryOperator * BO,Value * IVOperand)785 bool SimplifyIndvar::strengthenRightShift(BinaryOperator *BO,
786                                           Value *IVOperand) {
787   using namespace llvm::PatternMatch;
788 
789   if (BO->getOpcode() == Instruction::Shl) {
790     bool Changed = false;
791     ConstantRange IVRange = SE->getUnsignedRange(SE->getSCEV(IVOperand));
792     for (auto *U : BO->users()) {
793       const APInt *C;
794       if (match(U,
795                 m_AShr(m_Shl(m_Value(), m_Specific(IVOperand)), m_APInt(C))) ||
796           match(U,
797                 m_LShr(m_Shl(m_Value(), m_Specific(IVOperand)), m_APInt(C)))) {
798         BinaryOperator *Shr = cast<BinaryOperator>(U);
799         if (!Shr->isExact() && IVRange.getUnsignedMin().uge(*C)) {
800           Shr->setIsExact(true);
801           Changed = true;
802         }
803       }
804     }
805     return Changed;
806   }
807 
808   return false;
809 }
810 
811 /// Add all uses of Def to the current IV's worklist.
pushIVUsers(Instruction * Def,Loop * L,SmallPtrSet<Instruction *,16> & Simplified,SmallVectorImpl<std::pair<Instruction *,Instruction * >> & SimpleIVUsers)812 static void pushIVUsers(
813   Instruction *Def, Loop *L,
814   SmallPtrSet<Instruction*,16> &Simplified,
815   SmallVectorImpl< std::pair<Instruction*,Instruction*> > &SimpleIVUsers) {
816 
817   for (User *U : Def->users()) {
818     Instruction *UI = cast<Instruction>(U);
819 
820     // Avoid infinite or exponential worklist processing.
821     // Also ensure unique worklist users.
822     // If Def is a LoopPhi, it may not be in the Simplified set, so check for
823     // self edges first.
824     if (UI == Def)
825       continue;
826 
827     // Only change the current Loop, do not change the other parts (e.g. other
828     // Loops).
829     if (!L->contains(UI))
830       continue;
831 
832     // Do not push the same instruction more than once.
833     if (!Simplified.insert(UI).second)
834       continue;
835 
836     SimpleIVUsers.push_back(std::make_pair(UI, Def));
837   }
838 }
839 
840 /// Return true if this instruction generates a simple SCEV
841 /// expression in terms of that IV.
842 ///
843 /// This is similar to IVUsers' isInteresting() but processes each instruction
844 /// non-recursively when the operand is already known to be a simpleIVUser.
845 ///
isSimpleIVUser(Instruction * I,const Loop * L,ScalarEvolution * SE)846 static bool isSimpleIVUser(Instruction *I, const Loop *L, ScalarEvolution *SE) {
847   if (!SE->isSCEVable(I->getType()))
848     return false;
849 
850   // Get the symbolic expression for this instruction.
851   const SCEV *S = SE->getSCEV(I);
852 
853   // Only consider affine recurrences.
854   const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S);
855   if (AR && AR->getLoop() == L)
856     return true;
857 
858   return false;
859 }
860 
861 /// Iteratively perform simplification on a worklist of users
862 /// of the specified induction variable. Each successive simplification may push
863 /// more users which may themselves be candidates for simplification.
864 ///
865 /// This algorithm does not require IVUsers analysis. Instead, it simplifies
866 /// instructions in-place during analysis. Rather than rewriting induction
867 /// variables bottom-up from their users, it transforms a chain of IVUsers
868 /// top-down, updating the IR only when it encounters a clear optimization
869 /// opportunity.
870 ///
871 /// Once DisableIVRewrite is default, LSR will be the only client of IVUsers.
872 ///
simplifyUsers(PHINode * CurrIV,IVVisitor * V)873 void SimplifyIndvar::simplifyUsers(PHINode *CurrIV, IVVisitor *V) {
874   if (!SE->isSCEVable(CurrIV->getType()))
875     return;
876 
877   // Instructions processed by SimplifyIndvar for CurrIV.
878   SmallPtrSet<Instruction*,16> Simplified;
879 
880   // Use-def pairs if IV users waiting to be processed for CurrIV.
881   SmallVector<std::pair<Instruction*, Instruction*>, 8> SimpleIVUsers;
882 
883   // Push users of the current LoopPhi. In rare cases, pushIVUsers may be
884   // called multiple times for the same LoopPhi. This is the proper thing to
885   // do for loop header phis that use each other.
886   pushIVUsers(CurrIV, L, Simplified, SimpleIVUsers);
887 
888   while (!SimpleIVUsers.empty()) {
889     std::pair<Instruction*, Instruction*> UseOper =
890       SimpleIVUsers.pop_back_val();
891     Instruction *UseInst = UseOper.first;
892 
893     // If a user of the IndVar is trivially dead, we prefer just to mark it dead
894     // rather than try to do some complex analysis or transformation (such as
895     // widening) basing on it.
896     // TODO: Propagate TLI and pass it here to handle more cases.
897     if (isInstructionTriviallyDead(UseInst, /* TLI */ nullptr)) {
898       DeadInsts.emplace_back(UseInst);
899       continue;
900     }
901 
902     // Bypass back edges to avoid extra work.
903     if (UseInst == CurrIV) continue;
904 
905     // Try to replace UseInst with a loop invariant before any other
906     // simplifications.
907     if (replaceIVUserWithLoopInvariant(UseInst))
908       continue;
909 
910     Instruction *IVOperand = UseOper.second;
911     for (unsigned N = 0; IVOperand; ++N) {
912       assert(N <= Simplified.size() && "runaway iteration");
913 
914       Value *NewOper = foldIVUser(UseInst, IVOperand);
915       if (!NewOper)
916         break; // done folding
917       IVOperand = dyn_cast<Instruction>(NewOper);
918     }
919     if (!IVOperand)
920       continue;
921 
922     if (eliminateIVUser(UseInst, IVOperand)) {
923       pushIVUsers(IVOperand, L, Simplified, SimpleIVUsers);
924       continue;
925     }
926 
927     if (BinaryOperator *BO = dyn_cast<BinaryOperator>(UseInst)) {
928       if ((isa<OverflowingBinaryOperator>(BO) &&
929            strengthenOverflowingOperation(BO, IVOperand)) ||
930           (isa<ShlOperator>(BO) && strengthenRightShift(BO, IVOperand))) {
931         // re-queue uses of the now modified binary operator and fall
932         // through to the checks that remain.
933         pushIVUsers(IVOperand, L, Simplified, SimpleIVUsers);
934       }
935     }
936 
937     CastInst *Cast = dyn_cast<CastInst>(UseInst);
938     if (V && Cast) {
939       V->visitCast(Cast);
940       continue;
941     }
942     if (isSimpleIVUser(UseInst, L, SE)) {
943       pushIVUsers(UseInst, L, Simplified, SimpleIVUsers);
944     }
945   }
946 }
947 
948 namespace llvm {
949 
anchor()950 void IVVisitor::anchor() { }
951 
952 /// Simplify instructions that use this induction variable
953 /// by using ScalarEvolution to analyze the IV's recurrence.
simplifyUsersOfIV(PHINode * CurrIV,ScalarEvolution * SE,DominatorTree * DT,LoopInfo * LI,SmallVectorImpl<WeakTrackingVH> & Dead,SCEVExpander & Rewriter,IVVisitor * V)954 bool simplifyUsersOfIV(PHINode *CurrIV, ScalarEvolution *SE, DominatorTree *DT,
955                        LoopInfo *LI, SmallVectorImpl<WeakTrackingVH> &Dead,
956                        SCEVExpander &Rewriter, IVVisitor *V) {
957   SimplifyIndvar SIV(LI->getLoopFor(CurrIV->getParent()), SE, DT, LI, Rewriter,
958                      Dead);
959   SIV.simplifyUsers(CurrIV, V);
960   return SIV.hasChanged();
961 }
962 
963 /// Simplify users of induction variables within this
964 /// loop. This does not actually change or add IVs.
simplifyLoopIVs(Loop * L,ScalarEvolution * SE,DominatorTree * DT,LoopInfo * LI,SmallVectorImpl<WeakTrackingVH> & Dead)965 bool simplifyLoopIVs(Loop *L, ScalarEvolution *SE, DominatorTree *DT,
966                      LoopInfo *LI, SmallVectorImpl<WeakTrackingVH> &Dead) {
967   SCEVExpander Rewriter(*SE, SE->getDataLayout(), "indvars");
968 #ifndef NDEBUG
969   Rewriter.setDebugType(DEBUG_TYPE);
970 #endif
971   bool Changed = false;
972   for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
973     Changed |= simplifyUsersOfIV(cast<PHINode>(I), SE, DT, LI, Dead, Rewriter);
974   }
975   return Changed;
976 }
977 
978 } // namespace llvm
979