diff options
Diffstat (limited to 'lib/Transforms/Scalar/IndVarSimplify.cpp')
| -rw-r--r-- | lib/Transforms/Scalar/IndVarSimplify.cpp | 617 |
1 files changed, 180 insertions, 437 deletions
diff --git a/lib/Transforms/Scalar/IndVarSimplify.cpp b/lib/Transforms/Scalar/IndVarSimplify.cpp index 50140d9..75fa011 100644 --- a/lib/Transforms/Scalar/IndVarSimplify.cpp +++ b/lib/Transforms/Scalar/IndVarSimplify.cpp @@ -11,17 +11,6 @@ // computations derived from them) into simpler forms suitable for subsequent // analysis and transformation. // -// This transformation makes the following changes to each loop with an -// identifiable induction variable: -// 1. All loops are transformed to have a SINGLE canonical induction variable -// which starts at zero and steps by one. -// 2. The canonical induction variable is guaranteed to be the first PHI node -// in the loop header block. -// 3. The canonical induction variable is guaranteed to be in a wide enough -// type so that IV expressions need not be (directly) zero-extended or -// sign-extended. -// 4. Any pointer arithmetic recurrences are raised to use array subscripts. -// // If the trip count of a loop is computable, this pass also makes the following // changes: // 1. The exit condition for the loop is canonicalized to compare the @@ -33,9 +22,6 @@ // purpose of the loop is to compute the exit value of some derived // expression, this transformation will make the loop dead. // -// This transformation should be followed by strength reduction after all of the -// desired loop transformations have been performed. -// //===----------------------------------------------------------------------===// #define DEBUG_TYPE "indvars" @@ -57,11 +43,11 @@ #include "llvm/Support/raw_ostream.h" #include "llvm/Transforms/Utils/Local.h" #include "llvm/Transforms/Utils/BasicBlockUtils.h" +#include "llvm/Transforms/Utils/SimplifyIndVar.h" #include "llvm/Target/TargetData.h" #include "llvm/ADT/DenseMap.h" #include "llvm/ADT/SmallVector.h" #include "llvm/ADT/Statistic.h" -#include "llvm/ADT/STLExtras.h" using namespace llvm; STATISTIC(NumRemoved , "Number of aux indvars removed"); @@ -69,21 +55,21 @@ STATISTIC(NumWidened , "Number of indvars widened"); STATISTIC(NumInserted , "Number of canonical indvars added"); STATISTIC(NumReplaced , "Number of exit values replaced"); STATISTIC(NumLFTR , "Number of loop exit tests replaced"); -STATISTIC(NumElimIdentity, "Number of IV identities eliminated"); STATISTIC(NumElimExt , "Number of IV sign/zero extends eliminated"); -STATISTIC(NumElimRem , "Number of IV remainder operations eliminated"); -STATISTIC(NumElimCmp , "Number of IV comparisons eliminated"); STATISTIC(NumElimIV , "Number of congruent IVs eliminated"); -static cl::opt<bool> DisableIVRewrite( - "disable-iv-rewrite", cl::Hidden, - cl::desc("Disable canonical induction variable rewriting")); - -// Temporary flag for use with -disable-iv-rewrite to force a canonical IV for -// LFTR purposes. -static cl::opt<bool> ForceLFTR( - "force-lftr", cl::Hidden, - cl::desc("Enable forced linear function test replacement")); +namespace llvm { + cl::opt<bool> EnableIVRewrite( + "enable-iv-rewrite", cl::Hidden, + cl::desc("Enable canonical induction variable rewriting")); + + // Trip count verification can be enabled by default under NDEBUG if we + // implement a strong expression equivalence checker in SCEV. Until then, we + // use the verify-indvars flag, which may assert in some cases. + cl::opt<bool> VerifyIndvars( + "verify-indvars", cl::Hidden, + cl::desc("Verify the ScalarEvolution result after running indvars")); +} namespace { class IndVarSimplify : public LoopPass { @@ -111,12 +97,12 @@ namespace { AU.addRequired<ScalarEvolution>(); AU.addRequiredID(LoopSimplifyID); AU.addRequiredID(LCSSAID); - if (!DisableIVRewrite) + if (EnableIVRewrite) AU.addRequired<IVUsers>(); AU.addPreserved<ScalarEvolution>(); AU.addPreservedID(LoopSimplifyID); AU.addPreservedID(LCSSAID); - if (!DisableIVRewrite) + if (EnableIVRewrite) AU.addPreserved<IVUsers>(); AU.setPreservesCFG(); } @@ -131,18 +117,9 @@ namespace { void HandleFloatingPointIV(Loop *L, PHINode *PH); void RewriteNonIntegerIVs(Loop *L); - void RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter); - - void SimplifyIVUsers(SCEVExpander &Rewriter); - void SimplifyIVUsersNoRewrite(Loop *L, SCEVExpander &Rewriter); - - bool EliminateIVUser(Instruction *UseInst, Instruction *IVOperand); - void EliminateIVComparison(ICmpInst *ICmp, Value *IVOperand); - void EliminateIVRemainder(BinaryOperator *Rem, - Value *IVOperand, - bool IsSigned); + void SimplifyAndExtend(Loop *L, SCEVExpander &Rewriter, LPPassManager &LPM); - void SimplifyCongruentIVs(Loop *L); + void RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter); void RewriteIVExpressions(Loop *L, SCEVExpander &Rewriter); @@ -240,8 +217,8 @@ static Instruction *getInsertPointForUses(Instruction *User, Value *Def, InsertPt = InsertBB->getTerminator(); } assert(InsertPt && "Missing phi operand"); - assert(!isa<Instruction>(Def) || - DT->dominates(cast<Instruction>(Def), InsertPt) && + assert((!isa<Instruction>(Def) || + DT->dominates(cast<Instruction>(Def), InsertPt)) && "def does not dominate all uses"); return InsertPt; } @@ -372,14 +349,14 @@ void IndVarSimplify::HandleFloatingPointIV(Loop *L, PHINode *PN) { // Positive and negative strides have different safety conditions. if (IncValue > 0) { // If we have a positive stride, we require the init to be less than the - // exit value and an equality or less than comparison. - if (InitValue >= ExitValue || - NewPred == CmpInst::ICMP_SGT || NewPred == CmpInst::ICMP_SGE) + // exit value. + if (InitValue >= ExitValue) return; uint32_t Range = uint32_t(ExitValue-InitValue); - if (NewPred == CmpInst::ICMP_SLE) { - // Normalize SLE -> SLT, check for infinite loop. + // Check for infinite loop, either: + // while (i <= Exit) or until (i > Exit) + if (NewPred == CmpInst::ICMP_SLE || NewPred == CmpInst::ICMP_SGT) { if (++Range == 0) return; // Range overflows. } @@ -399,14 +376,14 @@ void IndVarSimplify::HandleFloatingPointIV(Loop *L, PHINode *PN) { } else { // If we have a negative stride, we require the init to be greater than the - // exit value and an equality or greater than comparison. - if (InitValue >= ExitValue || - NewPred == CmpInst::ICMP_SLT || NewPred == CmpInst::ICMP_SLE) + // exit value. + if (InitValue <= ExitValue) return; uint32_t Range = uint32_t(InitValue-ExitValue); - if (NewPred == CmpInst::ICMP_SGE) { - // Normalize SGE -> SGT, check for infinite loop. + // Check for infinite loop, either: + // while (i >= Exit) or until (i < Exit) + if (NewPred == CmpInst::ICMP_SGE || NewPred == CmpInst::ICMP_SLT) { if (++Range == 0) return; // Range overflows. } @@ -464,7 +441,7 @@ void IndVarSimplify::HandleFloatingPointIV(Loop *L, PHINode *PN) { // platforms. if (WeakPH) { Value *Conv = new SIToFPInst(NewPHI, PN->getType(), "indvar.conv", - PN->getParent()->getFirstNonPHI()); + PN->getParent()->getFirstInsertionPt()); PN->replaceAllUsesWith(Conv); RecursivelyDeleteTriviallyDeadInstructions(PN); } @@ -472,6 +449,8 @@ void IndVarSimplify::HandleFloatingPointIV(Loop *L, PHINode *PN) { // Add a new IVUsers entry for the newly-created integer PHI. if (IU) IU->AddUsersIfInteresting(NewPHI); + + Changed = true; } void IndVarSimplify::RewriteNonIntegerIVs(Loop *L) { @@ -617,45 +596,15 @@ void IndVarSimplify::RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter) { //===----------------------------------------------------------------------===// // Rewrite IV users based on a canonical IV. -// To be replaced by -disable-iv-rewrite. +// Only for use with -enable-iv-rewrite. //===----------------------------------------------------------------------===// -/// SimplifyIVUsers - Iteratively perform simplification on IVUsers within this -/// loop. IVUsers is treated as a worklist. Each successive simplification may -/// push more users which may themselves be candidates for simplification. -/// -/// This is the old approach to IV simplification to be replaced by -/// SimplifyIVUsersNoRewrite. -/// -void IndVarSimplify::SimplifyIVUsers(SCEVExpander &Rewriter) { - // Each round of simplification involves a round of eliminating operations - // followed by a round of widening IVs. A single IVUsers worklist is used - // across all rounds. The inner loop advances the user. If widening exposes - // more uses, then another pass through the outer loop is triggered. - for (IVUsers::iterator I = IU->begin(); I != IU->end(); ++I) { - Instruction *UseInst = I->getUser(); - Value *IVOperand = I->getOperandValToReplace(); - - if (ICmpInst *ICmp = dyn_cast<ICmpInst>(UseInst)) { - EliminateIVComparison(ICmp, IVOperand); - continue; - } - if (BinaryOperator *Rem = dyn_cast<BinaryOperator>(UseInst)) { - bool IsSigned = Rem->getOpcode() == Instruction::SRem; - if (IsSigned || Rem->getOpcode() == Instruction::URem) { - EliminateIVRemainder(Rem, IVOperand, IsSigned); - continue; - } - } - } -} - -// FIXME: It is an extremely bad idea to indvar substitute anything more -// complex than affine induction variables. Doing so will put expensive -// polynomial evaluations inside of the loop, and the str reduction pass -// currently can only reduce affine polynomials. For now just disable -// indvar subst on anything more complex than an affine addrec, unless -// it can be expanded to a trivial value. +/// FIXME: It is an extremely bad idea to indvar substitute anything more +/// complex than affine induction variables. Doing so will put expensive +/// polynomial evaluations inside of the loop, and the str reduction pass +/// currently can only reduce affine polynomials. For now just disable +/// indvar subst on anything more complex than an affine addrec, unless +/// it can be expanded to a trivial value. static bool isSafe(const SCEV *S, const Loop *L, ScalarEvolution *SE) { // Loop-invariant values are safe. if (SE->isLoopInvariant(S, L)) return true; @@ -666,7 +615,8 @@ static bool isSafe(const SCEV *S, const Loop *L, ScalarEvolution *SE) { return AR->isAffine(); // An add is safe it all its operands are safe. - if (const SCEVCommutativeExpr *Commutative = dyn_cast<SCEVCommutativeExpr>(S)) { + if (const SCEVCommutativeExpr *Commutative + = dyn_cast<SCEVCommutativeExpr>(S)) { for (SCEVCommutativeExpr::op_iterator I = Commutative->op_begin(), E = Commutative->op_end(); I != E; ++I) if (!isSafe(*I, L, SE)) return false; @@ -771,18 +721,37 @@ namespace { // extend operations. This information is recorded by CollectExtend and // provides the input to WidenIV. struct WideIVInfo { + PHINode *NarrowIV; Type *WidestNativeType; // Widest integer type created [sz]ext - bool IsSigned; // Was an sext user seen before a zext? + bool IsSigned; // Was an sext user seen before a zext? + + WideIVInfo() : NarrowIV(0), WidestNativeType(0), IsSigned(false) {} + }; + + class WideIVVisitor : public IVVisitor { + ScalarEvolution *SE; + const TargetData *TD; - WideIVInfo() : WidestNativeType(0), IsSigned(false) {} + public: + WideIVInfo WI; + + WideIVVisitor(PHINode *NarrowIV, ScalarEvolution *SCEV, + const TargetData *TData) : + SE(SCEV), TD(TData) { WI.NarrowIV = NarrowIV; } + + // Implement the interface used by simplifyUsersOfIV. + virtual void visitCast(CastInst *Cast); }; } -/// CollectExtend - Update information about the induction variable that is +/// visitCast - Update information about the induction variable that is /// extended by this sign or zero extend operation. This is used to determine /// the final width of the IV before actually widening it. -static void CollectExtend(CastInst *Cast, bool IsSigned, WideIVInfo &WI, - ScalarEvolution *SE, const TargetData *TD) { +void WideIVVisitor::visitCast(CastInst *Cast) { + bool IsSigned = Cast->getOpcode() == Instruction::SExt; + if (!IsSigned && Cast->getOpcode() != Instruction::ZExt) + return; + Type *Ty = Cast->getType(); uint64_t Width = SE->getTypeSizeInBits(Ty); if (TD && !TD->isLegalInteger(Width)) @@ -845,10 +814,10 @@ class WidenIV { SmallVector<NarrowIVDefUse, 8> NarrowIVUsers; public: - WidenIV(PHINode *PN, const WideIVInfo &WI, LoopInfo *LInfo, + WidenIV(const WideIVInfo &WI, LoopInfo *LInfo, ScalarEvolution *SEv, DominatorTree *DTree, SmallVectorImpl<WeakVH> &DI) : - OrigPhi(PN), + OrigPhi(WI.NarrowIV), WideType(WI.WidestNativeType), IsSigned(WI.IsSigned), LI(LInfo), @@ -865,18 +834,42 @@ public: PHINode *CreateWideIV(SCEVExpander &Rewriter); protected: + Value *getExtend(Value *NarrowOper, Type *WideType, bool IsSigned, + Instruction *Use); + Instruction *CloneIVUser(NarrowIVDefUse DU); const SCEVAddRecExpr *GetWideRecurrence(Instruction *NarrowUse); + const SCEVAddRecExpr* GetExtendedOperandRecurrence(NarrowIVDefUse DU); + Instruction *WidenIVUse(NarrowIVDefUse DU); void pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef); }; } // anonymous namespace -static Value *getExtend( Value *NarrowOper, Type *WideType, - bool IsSigned, IRBuilder<> &Builder) { +/// isLoopInvariant - Perform a quick domtree based check for loop invariance +/// assuming that V is used within the loop. LoopInfo::isLoopInvariant() seems +/// gratuitous for this purpose. +static bool isLoopInvariant(Value *V, const Loop *L, const DominatorTree *DT) { + Instruction *Inst = dyn_cast<Instruction>(V); + if (!Inst) + return true; + + return DT->properlyDominates(Inst->getParent(), L->getHeader()); +} + +Value *WidenIV::getExtend(Value *NarrowOper, Type *WideType, bool IsSigned, + Instruction *Use) { + // Set the debug location and conservative insertion point. + IRBuilder<> Builder(Use); + // Hoist the insertion point into loop preheaders as far as possible. + for (const Loop *L = LI->getLoopFor(Use->getParent()); + L && L->getLoopPreheader() && isLoopInvariant(NarrowOper, L, DT); + L = L->getParentLoop()) + Builder.SetInsertPoint(L->getLoopPreheader()->getTerminator()); + return IsSigned ? Builder.CreateSExt(NarrowOper, WideType) : Builder.CreateZExt(NarrowOper, WideType); } @@ -901,22 +894,21 @@ Instruction *WidenIV::CloneIVUser(NarrowIVDefUse DU) { case Instruction::AShr: DEBUG(dbgs() << "Cloning IVUser: " << *DU.NarrowUse << "\n"); - IRBuilder<> Builder(DU.NarrowUse); - // Replace NarrowDef operands with WideDef. Otherwise, we don't know // anything about the narrow operand yet so must insert a [sz]ext. It is // probably loop invariant and will be folded or hoisted. If it actually // comes from a widened IV, it should be removed during a future call to // WidenIVUse. Value *LHS = (DU.NarrowUse->getOperand(0) == DU.NarrowDef) ? DU.WideDef : - getExtend(DU.NarrowUse->getOperand(0), WideType, IsSigned, Builder); + getExtend(DU.NarrowUse->getOperand(0), WideType, IsSigned, DU.NarrowUse); Value *RHS = (DU.NarrowUse->getOperand(1) == DU.NarrowDef) ? DU.WideDef : - getExtend(DU.NarrowUse->getOperand(1), WideType, IsSigned, Builder); + getExtend(DU.NarrowUse->getOperand(1), WideType, IsSigned, DU.NarrowUse); BinaryOperator *NarrowBO = cast<BinaryOperator>(DU.NarrowUse); BinaryOperator *WideBO = BinaryOperator::Create(NarrowBO->getOpcode(), LHS, RHS, NarrowBO->getName()); + IRBuilder<> Builder(DU.NarrowUse); Builder.Insert(WideBO); if (const OverflowingBinaryOperator *OBO = dyn_cast<OverflowingBinaryOperator>(NarrowBO)) { @@ -928,45 +920,46 @@ Instruction *WidenIV::CloneIVUser(NarrowIVDefUse DU) { llvm_unreachable(0); } -/// HoistStep - Attempt to hoist an IV increment above a potential use. -/// -/// To successfully hoist, two criteria must be met: -/// - IncV operands dominate InsertPos and -/// - InsertPos dominates IncV -/// -/// Meeting the second condition means that we don't need to check all of IncV's -/// existing uses (it's moving up in the domtree). -/// -/// This does not yet recursively hoist the operands, although that would -/// not be difficult. -static bool HoistStep(Instruction *IncV, Instruction *InsertPos, - const DominatorTree *DT) -{ - if (DT->dominates(IncV, InsertPos)) - return true; +/// No-wrap operations can transfer sign extension of their result to their +/// operands. Generate the SCEV value for the widened operation without +/// actually modifying the IR yet. If the expression after extending the +/// operands is an AddRec for this loop, return it. +const SCEVAddRecExpr* WidenIV::GetExtendedOperandRecurrence(NarrowIVDefUse DU) { + // Handle the common case of add<nsw/nuw> + if (DU.NarrowUse->getOpcode() != Instruction::Add) + return 0; - if (!DT->dominates(InsertPos->getParent(), IncV->getParent())) - return false; + // One operand (NarrowDef) has already been extended to WideDef. Now determine + // if extending the other will lead to a recurrence. + unsigned ExtendOperIdx = DU.NarrowUse->getOperand(0) == DU.NarrowDef ? 1 : 0; + assert(DU.NarrowUse->getOperand(1-ExtendOperIdx) == DU.NarrowDef && "bad DU"); + + const SCEV *ExtendOperExpr = 0; + const OverflowingBinaryOperator *OBO = + cast<OverflowingBinaryOperator>(DU.NarrowUse); + if (IsSigned && OBO->hasNoSignedWrap()) + ExtendOperExpr = SE->getSignExtendExpr( + SE->getSCEV(DU.NarrowUse->getOperand(ExtendOperIdx)), WideType); + else if(!IsSigned && OBO->hasNoUnsignedWrap()) + ExtendOperExpr = SE->getZeroExtendExpr( + SE->getSCEV(DU.NarrowUse->getOperand(ExtendOperIdx)), WideType); + else + return 0; - if (IncV->mayHaveSideEffects()) - return false; + const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>( + SE->getAddExpr(SE->getSCEV(DU.WideDef), ExtendOperExpr, + IsSigned ? SCEV::FlagNSW : SCEV::FlagNUW)); - // Attempt to hoist IncV - for (User::op_iterator OI = IncV->op_begin(), OE = IncV->op_end(); - OI != OE; ++OI) { - Instruction *OInst = dyn_cast<Instruction>(OI); - if (OInst && !DT->dominates(OInst, InsertPos)) - return false; - } - IncV->moveBefore(InsertPos); - return true; + if (!AddRec || AddRec->getLoop() != L) + return 0; + return AddRec; } -// GetWideRecurrence - Is this instruction potentially interesting from IVUsers' -// perspective after widening it's type? In other words, can the extend be -// safely hoisted out of the loop with SCEV reducing the value to a recurrence -// on the same loop. If so, return the sign or zero extended -// recurrence. Otherwise return NULL. +/// GetWideRecurrence - Is this instruction potentially interesting from +/// IVUsers' perspective after widening it's type? In other words, can the +/// extend be safely hoisted out of the loop with SCEV reducing the value to a +/// recurrence on the same loop. If so, return the sign or zero extended +/// recurrence. Otherwise return NULL. const SCEVAddRecExpr *WidenIV::GetWideRecurrence(Instruction *NarrowUse) { if (!SE->isSCEVable(NarrowUse->getType())) return 0; @@ -985,7 +978,6 @@ const SCEVAddRecExpr *WidenIV::GetWideRecurrence(Instruction *NarrowUse) { const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(WideExpr); if (!AddRec || AddRec->getLoop() != L) return 0; - return AddRec; } @@ -1039,6 +1031,9 @@ Instruction *WidenIV::WidenIVUse(NarrowIVDefUse DU) { // Does this user itself evaluate to a recurrence after widening? const SCEVAddRecExpr *WideAddRec = GetWideRecurrence(DU.NarrowUse); if (!WideAddRec) { + WideAddRec = GetExtendedOperandRecurrence(DU); + } + if (!WideAddRec) { // This user does not evaluate to a recurence after widening, so don't // follow it. Instead insert a Trunc to kill off the original use, // eventually isolating the original narrow IV so it can be removed. @@ -1055,9 +1050,9 @@ Instruction *WidenIV::WidenIVUse(NarrowIVDefUse DU) { // Reuse the IV increment that SCEVExpander created as long as it dominates // NarrowUse. Instruction *WideUse = 0; - if (WideAddRec == WideIncExpr && HoistStep(WideInc, DU.NarrowUse, DT)) { + if (WideAddRec == WideIncExpr + && SCEVExpander::hoistStep(WideInc, DU.NarrowUse, DT)) WideUse = WideInc; - } else { WideUse = CloneIVUser(DU); if (!WideUse) @@ -1178,183 +1173,17 @@ PHINode *WidenIV::CreateWideIV(SCEVExpander &Rewriter) { // Simplification of IV users based on SCEV evaluation. //===----------------------------------------------------------------------===// -void IndVarSimplify::EliminateIVComparison(ICmpInst *ICmp, Value *IVOperand) { - unsigned IVOperIdx = 0; - ICmpInst::Predicate Pred = ICmp->getPredicate(); - if (IVOperand != ICmp->getOperand(0)) { - // Swapped - assert(IVOperand == ICmp->getOperand(1) && "Can't find IVOperand"); - IVOperIdx = 1; - Pred = ICmpInst::getSwappedPredicate(Pred); - } - - // Get the SCEVs for the ICmp operands. - const SCEV *S = SE->getSCEV(ICmp->getOperand(IVOperIdx)); - const SCEV *X = SE->getSCEV(ICmp->getOperand(1 - IVOperIdx)); - - // Simplify unnecessary loops away. - const Loop *ICmpLoop = LI->getLoopFor(ICmp->getParent()); - S = SE->getSCEVAtScope(S, ICmpLoop); - X = SE->getSCEVAtScope(X, ICmpLoop); - - // If the condition is always true or always false, replace it with - // a constant value. - if (SE->isKnownPredicate(Pred, S, X)) - ICmp->replaceAllUsesWith(ConstantInt::getTrue(ICmp->getContext())); - else if (SE->isKnownPredicate(ICmpInst::getInversePredicate(Pred), S, X)) - ICmp->replaceAllUsesWith(ConstantInt::getFalse(ICmp->getContext())); - else - return; - DEBUG(dbgs() << "INDVARS: Eliminated comparison: " << *ICmp << '\n'); - ++NumElimCmp; - Changed = true; - DeadInsts.push_back(ICmp); -} - -void IndVarSimplify::EliminateIVRemainder(BinaryOperator *Rem, - Value *IVOperand, - bool IsSigned) { - // We're only interested in the case where we know something about - // the numerator. - if (IVOperand != Rem->getOperand(0)) - return; - - // Get the SCEVs for the ICmp operands. - const SCEV *S = SE->getSCEV(Rem->getOperand(0)); - const SCEV *X = SE->getSCEV(Rem->getOperand(1)); - - // Simplify unnecessary loops away. - const Loop *ICmpLoop = LI->getLoopFor(Rem->getParent()); - S = SE->getSCEVAtScope(S, ICmpLoop); - X = SE->getSCEVAtScope(X, ICmpLoop); - - // i % n --> i if i is in [0,n). - if ((!IsSigned || SE->isKnownNonNegative(S)) && - SE->isKnownPredicate(IsSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT, - S, X)) - Rem->replaceAllUsesWith(Rem->getOperand(0)); - else { - // (i+1) % n --> (i+1)==n?0:(i+1) if i is in [0,n). - const SCEV *LessOne = - SE->getMinusSCEV(S, SE->getConstant(S->getType(), 1)); - if (IsSigned && !SE->isKnownNonNegative(LessOne)) - return; - - if (!SE->isKnownPredicate(IsSigned ? - ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT, - LessOne, X)) - return; - - ICmpInst *ICmp = new ICmpInst(Rem, ICmpInst::ICMP_EQ, - Rem->getOperand(0), Rem->getOperand(1), - "tmp"); - SelectInst *Sel = - SelectInst::Create(ICmp, - ConstantInt::get(Rem->getType(), 0), - Rem->getOperand(0), "tmp", Rem); - Rem->replaceAllUsesWith(Sel); - } - - // Inform IVUsers about the new users. - if (IU) { - if (Instruction *I = dyn_cast<Instruction>(Rem->getOperand(0))) - IU->AddUsersIfInteresting(I); - } - DEBUG(dbgs() << "INDVARS: Simplified rem: " << *Rem << '\n'); - ++NumElimRem; - Changed = true; - DeadInsts.push_back(Rem); -} - -/// EliminateIVUser - Eliminate an operation that consumes a simple IV and has -/// no observable side-effect given the range of IV values. -bool IndVarSimplify::EliminateIVUser(Instruction *UseInst, - Instruction *IVOperand) { - if (ICmpInst *ICmp = dyn_cast<ICmpInst>(UseInst)) { - EliminateIVComparison(ICmp, IVOperand); - return true; - } - if (BinaryOperator *Rem = dyn_cast<BinaryOperator>(UseInst)) { - bool IsSigned = Rem->getOpcode() == Instruction::SRem; - if (IsSigned || Rem->getOpcode() == Instruction::URem) { - EliminateIVRemainder(Rem, IVOperand, IsSigned); - return true; - } - } - - // Eliminate any operation that SCEV can prove is an identity function. - if (!SE->isSCEVable(UseInst->getType()) || - (UseInst->getType() != IVOperand->getType()) || - (SE->getSCEV(UseInst) != SE->getSCEV(IVOperand))) - return false; - - DEBUG(dbgs() << "INDVARS: Eliminated identity: " << *UseInst << '\n'); - - UseInst->replaceAllUsesWith(IVOperand); - ++NumElimIdentity; - Changed = true; - DeadInsts.push_back(UseInst); - return true; -} - -/// pushIVUsers - Add all uses of Def to the current IV's worklist. -/// -static void pushIVUsers( - Instruction *Def, - SmallPtrSet<Instruction*,16> &Simplified, - SmallVectorImpl< std::pair<Instruction*,Instruction*> > &SimpleIVUsers) { - - for (Value::use_iterator UI = Def->use_begin(), E = Def->use_end(); - UI != E; ++UI) { - Instruction *User = cast<Instruction>(*UI); - - // Avoid infinite or exponential worklist processing. - // Also ensure unique worklist users. - // If Def is a LoopPhi, it may not be in the Simplified set, so check for - // self edges first. - if (User != Def && Simplified.insert(User)) - SimpleIVUsers.push_back(std::make_pair(User, Def)); - } -} - -/// isSimpleIVUser - Return true if this instruction generates a simple SCEV -/// expression in terms of that IV. -/// -/// This is similar to IVUsers' isInsteresting() but processes each instruction -/// non-recursively when the operand is already known to be a simpleIVUser. -/// -static bool isSimpleIVUser(Instruction *I, const Loop *L, ScalarEvolution *SE) { - if (!SE->isSCEVable(I->getType())) - return false; - - // Get the symbolic expression for this instruction. - const SCEV *S = SE->getSCEV(I); - - // Only consider affine recurrences. - const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S); - if (AR && AR->getLoop() == L) - return true; - - return false; -} - -/// SimplifyIVUsersNoRewrite - Iteratively perform simplification on a worklist -/// of IV users. Each successive simplification may push more users which may +/// SimplifyAndExtend - Iteratively perform simplification on a worklist of IV +/// users. Each successive simplification may push more users which may /// themselves be candidates for simplification. /// -/// The "NoRewrite" algorithm does not require IVUsers analysis. Instead, it -/// simplifies instructions in-place during analysis. Rather than rewriting -/// induction variables bottom-up from their users, it transforms a chain of -/// IVUsers top-down, updating the IR only when it encouters a clear -/// optimization opportunitiy. A SCEVExpander "Rewriter" instance is still -/// needed, but only used to generate a new IV (phi) of wider type for sign/zero -/// extend elimination. +/// Sign/Zero extend elimination is interleaved with IV simplification. /// -/// Once DisableIVRewrite is default, LSR will be the only client of IVUsers. -/// -void IndVarSimplify::SimplifyIVUsersNoRewrite(Loop *L, SCEVExpander &Rewriter) { - std::map<PHINode *, WideIVInfo> WideIVMap; +void IndVarSimplify::SimplifyAndExtend(Loop *L, + SCEVExpander &Rewriter, + LPPassManager &LPM) { + SmallVector<WideIVInfo, 8> WideIVs; SmallVector<PHINode*, 8> LoopPhis; for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) { @@ -1370,108 +1199,27 @@ void IndVarSimplify::SimplifyIVUsersNoRewrite(Loop *L, SCEVExpander &Rewriter) { // extension. The first time SCEV attempts to normalize sign/zero extension, // the result becomes final. So for the most predictable results, we delay // evaluation of sign/zero extend evaluation until needed, and avoid running - // other SCEV based analysis prior to SimplifyIVUsersNoRewrite. + // other SCEV based analysis prior to SimplifyAndExtend. do { PHINode *CurrIV = LoopPhis.pop_back_val(); // Information about sign/zero extensions of CurrIV. - WideIVInfo WI; - - // Instructions processed by SimplifyIVUsers for CurrIV. - SmallPtrSet<Instruction*,16> Simplified; + WideIVVisitor WIV(CurrIV, SE, TD); - // Use-def pairs if IV users waiting to be processed for CurrIV. - SmallVector<std::pair<Instruction*, Instruction*>, 8> SimpleIVUsers; + Changed |= simplifyUsersOfIV(CurrIV, SE, &LPM, DeadInsts, &WIV); - // Push users of the current LoopPhi. In rare cases, pushIVUsers may be - // called multiple times for the same LoopPhi. This is the proper thing to - // do for loop header phis that use each other. - pushIVUsers(CurrIV, Simplified, SimpleIVUsers); - - while (!SimpleIVUsers.empty()) { - Instruction *UseInst, *Operand; - tie(UseInst, Operand) = SimpleIVUsers.pop_back_val(); - // Bypass back edges to avoid extra work. - if (UseInst == CurrIV) continue; - - if (EliminateIVUser(UseInst, Operand)) { - pushIVUsers(Operand, Simplified, SimpleIVUsers); - continue; - } - if (CastInst *Cast = dyn_cast<CastInst>(UseInst)) { - bool IsSigned = Cast->getOpcode() == Instruction::SExt; - if (IsSigned || Cast->getOpcode() == Instruction::ZExt) { - CollectExtend(Cast, IsSigned, WI, SE, TD); - } - continue; - } - if (isSimpleIVUser(UseInst, L, SE)) { - pushIVUsers(UseInst, Simplified, SimpleIVUsers); - } - } - if (WI.WidestNativeType) { - WideIVMap[CurrIV] = WI; + if (WIV.WI.WidestNativeType) { + WideIVs.push_back(WIV.WI); } } while(!LoopPhis.empty()); - for (std::map<PHINode *, WideIVInfo>::const_iterator I = WideIVMap.begin(), - E = WideIVMap.end(); I != E; ++I) { - WidenIV Widener(I->first, I->second, LI, SE, DT, DeadInsts); + for (; !WideIVs.empty(); WideIVs.pop_back()) { + WidenIV Widener(WideIVs.back(), LI, SE, DT, DeadInsts); if (PHINode *WidePhi = Widener.CreateWideIV(Rewriter)) { Changed = true; LoopPhis.push_back(WidePhi); } } - WideIVMap.clear(); - } -} - -/// SimplifyCongruentIVs - Check for congruent phis in this loop header and -/// populate ExprToIVMap for use later. -/// -void IndVarSimplify::SimplifyCongruentIVs(Loop *L) { - DenseMap<const SCEV *, PHINode *> ExprToIVMap; - for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) { - PHINode *Phi = cast<PHINode>(I); - if (!SE->isSCEVable(Phi->getType())) - continue; - - const SCEV *S = SE->getSCEV(Phi); - DenseMap<const SCEV *, PHINode *>::const_iterator Pos; - bool Inserted; - tie(Pos, Inserted) = ExprToIVMap.insert(std::make_pair(S, Phi)); - if (Inserted) - continue; - PHINode *OrigPhi = Pos->second; - - // If one phi derives from the other via GEPs, types may differ. - if (OrigPhi->getType() != Phi->getType()) - continue; - - // Replacing the congruent phi is sufficient because acyclic redundancy - // elimination, CSE/GVN, should handle the rest. However, once SCEV proves - // that a phi is congruent, it's almost certain to be the head of an IV - // user cycle that is isomorphic with the original phi. So it's worth - // eagerly cleaning up the common case of a single IV increment. - if (BasicBlock *LatchBlock = L->getLoopLatch()) { - Instruction *OrigInc = - cast<Instruction>(OrigPhi->getIncomingValueForBlock(LatchBlock)); - Instruction *IsomorphicInc = - cast<Instruction>(Phi->getIncomingValueForBlock(LatchBlock)); - if (OrigInc != IsomorphicInc && - OrigInc->getType() == IsomorphicInc->getType() && - SE->getSCEV(OrigInc) == SE->getSCEV(IsomorphicInc) && - HoistStep(OrigInc, IsomorphicInc, DT)) { - DEBUG(dbgs() << "INDVARS: Eliminated congruent iv.inc: " - << *IsomorphicInc << '\n'); - IsomorphicInc->replaceAllUsesWith(OrigInc); - DeadInsts.push_back(IsomorphicInc); - } - } - DEBUG(dbgs() << "INDVARS: Eliminated congruent iv: " << *Phi << '\n'); - ++NumElimIV; - Phi->replaceAllUsesWith(OrigPhi); - DeadInsts.push_back(Phi); } } @@ -1479,9 +1227,9 @@ void IndVarSimplify::SimplifyCongruentIVs(Loop *L) { // LinearFunctionTestReplace and its kin. Rewrite the loop exit condition. //===----------------------------------------------------------------------===// -// Check for expressions that ScalarEvolution generates to compute -// BackedgeTakenInfo. If these expressions have not been reduced, then expanding -// them may incur additional cost (albeit in the loop preheader). +/// Check for expressions that ScalarEvolution generates to compute +/// BackedgeTakenInfo. If these expressions have not been reduced, then +/// expanding them may incur additional cost (albeit in the loop preheader). static bool isHighCostExpansion(const SCEV *S, BranchInst *BI, ScalarEvolution *SE) { // If the backedge-taken count is a UDiv, it's very likely a UDiv that @@ -1502,7 +1250,7 @@ static bool isHighCostExpansion(const SCEV *S, BranchInst *BI, } } - if (!DisableIVRewrite || ForceLFTR) + if (EnableIVRewrite) return false; // Recurse past add expressions, which commonly occur in the @@ -1580,17 +1328,6 @@ static Type *getBackedgeIVType(Loop *L) { return Ty; } -/// isLoopInvariant - Perform a quick domtree based check for loop invariance -/// assuming that V is used within the loop. LoopInfo::isLoopInvariant() seems -/// gratuitous for this purpose. -static bool isLoopInvariant(Value *V, Loop *L, DominatorTree *DT) { - Instruction *Inst = dyn_cast<Instruction>(V); - if (!Inst) - return true; - - return DT->properlyDominates(Inst->getParent(), L->getHeader()); -} - /// getLoopPhiForCounter - Return the loop header phi IFF IncV adds a loop /// invariant value to the phi. static PHINode *getLoopPhiForCounter(Value *IncV, Loop *L, DominatorTree *DT) { @@ -1779,10 +1516,9 @@ LinearFunctionTestReplace(Loop *L, assert(canExpandBackedgeTakenCount(L, SE) && "precondition"); BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator()); - // In DisableIVRewrite mode, IndVar is not necessarily a canonical IV. In this - // mode, LFTR can ignore IV overflow and truncate to the width of + // LFTR can ignore IV overflow and truncate to the width of // BECount. This avoids materializing the add(zext(add)) expression. - Type *CntTy = DisableIVRewrite ? + Type *CntTy = !EnableIVRewrite ? BackedgeTakenCount->getType() : IndVar->getType(); const SCEV *IVLimit = BackedgeTakenCount; @@ -1832,7 +1568,7 @@ LinearFunctionTestReplace(Loop *L, const SCEV *IVInit = AR->getStart(); // For pointer types, sign extend BECount in order to materialize a GEP. - // Note that for DisableIVRewrite, we never run SCEVExpander on a + // Note that for without EnableIVRewrite, we never run SCEVExpander on a // pointer type, because we must preserve the existing GEPs. Instead we // directly generate a GEP later. if (IVInit->getType()->isPointerTy()) { @@ -1919,7 +1655,7 @@ void IndVarSimplify::SinkUnusedInvariants(Loop *L) { BasicBlock *Preheader = L->getLoopPreheader(); if (!Preheader) return; - Instruction *InsertPt = ExitBlock->getFirstNonPHI(); + Instruction *InsertPt = ExitBlock->getFirstInsertionPt(); BasicBlock::iterator I = Preheader->getTerminator(); while (I != Preheader->begin()) { --I; @@ -1940,6 +1676,10 @@ void IndVarSimplify::SinkUnusedInvariants(Loop *L) { if (isa<DbgInfoIntrinsic>(I)) continue; + // Skip landingpad instructions. + if (isa<LandingPadInst>(I)) + continue; + // Don't sink static AllocaInsts out of the entry block, which would // turn them into dynamic allocas! if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) @@ -2006,7 +1746,7 @@ bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) { if (!L->isLoopSimplifyForm()) return false; - if (!DisableIVRewrite) + if (EnableIVRewrite) IU = &getAnalysis<IVUsers>(); LI = &getAnalysis<LoopInfo>(); SE = &getAnalysis<ScalarEvolution>(); @@ -2024,6 +1764,9 @@ bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) { // Create a rewriter object which we'll use to transform the code with. SCEVExpander Rewriter(*SE, "indvars"); +#ifndef NDEBUG + Rewriter.setDebugType(DEBUG_TYPE); +#endif // Eliminate redundant IV users. // @@ -2031,9 +1774,9 @@ bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) { // attempt to avoid evaluating SCEVs for sign/zero extend operations until // other expressions involving loop IVs have been evaluated. This helps SCEV // set no-wrap flags before normalizing sign/zero extension. - if (DisableIVRewrite) { + if (!EnableIVRewrite) { Rewriter.disableCanonicalMode(); - SimplifyIVUsersNoRewrite(L, Rewriter); + SimplifyAndExtend(L, Rewriter, LPM); } // Check to see if this loop has a computable loop-invariant execution count. @@ -2046,26 +1789,25 @@ bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) { RewriteLoopExitValues(L, Rewriter); // Eliminate redundant IV users. - if (!DisableIVRewrite) - SimplifyIVUsers(Rewriter); + if (EnableIVRewrite) + Changed |= simplifyIVUsers(IU, SE, &LPM, DeadInsts); // Eliminate redundant IV cycles. - if (DisableIVRewrite) - SimplifyCongruentIVs(L); + if (!EnableIVRewrite) + NumElimIV += Rewriter.replaceCongruentIVs(L, DT, DeadInsts); // Compute the type of the largest recurrence expression, and decide whether // a canonical induction variable should be inserted. Type *LargestType = 0; bool NeedCannIV = false; - bool ReuseIVForExit = DisableIVRewrite && !ForceLFTR; bool ExpandBECount = canExpandBackedgeTakenCount(L, SE); - if (ExpandBECount && !ReuseIVForExit) { + if (EnableIVRewrite && ExpandBECount) { // If we have a known trip count and a single exit block, we'll be // rewriting the loop exit test condition below, which requires a // canonical induction variable. NeedCannIV = true; Type *Ty = BackedgeTakenCount->getType(); - if (DisableIVRewrite) { + if (!EnableIVRewrite) { // In this mode, SimplifyIVUsers may have already widened the IV used by // the backedge test and inserted a Trunc on the compare's operand. Get // the wider type to avoid creating a redundant narrow IV only used by the @@ -2077,7 +1819,7 @@ bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) { SE->getTypeSizeInBits(LargestType)) LargestType = SE->getEffectiveSCEVType(Ty); } - if (!DisableIVRewrite) { + if (EnableIVRewrite) { for (IVUsers::const_iterator I = IU->begin(), E = IU->end(); I != E; ++I) { NeedCannIV = true; Type *Ty = @@ -2119,10 +1861,10 @@ bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) { // the end of the pass. while (!OldCannIVs.empty()) { PHINode *OldCannIV = OldCannIVs.pop_back_val(); - OldCannIV->insertBefore(L->getHeader()->getFirstNonPHI()); + OldCannIV->insertBefore(L->getHeader()->getFirstInsertionPt()); } } - else if (ExpandBECount && ReuseIVForExit && needsLFTR(L, DT)) { + else if (!EnableIVRewrite && ExpandBECount && needsLFTR(L, DT)) { IndVar = FindLoopCounter(L, BackedgeTakenCount, SE, DT, TD); } // If we have a trip count expression, rewrite the loop's exit condition @@ -2143,7 +1885,7 @@ bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) { LinearFunctionTestReplace(L, BackedgeTakenCount, IndVar, Rewriter); } // Rewrite IV-derived expressions. - if (!DisableIVRewrite) + if (EnableIVRewrite) RewriteIVExpressions(L, Rewriter); // Clear the rewriter cache, because values that are in the rewriter's cache @@ -2180,7 +1922,8 @@ bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) { // Verify that LFTR, and any other change have not interfered with SCEV's // ability to compute trip count. #ifndef NDEBUG - if (DisableIVRewrite && !isa<SCEVCouldNotCompute>(BackedgeTakenCount)) { + if (!EnableIVRewrite && VerifyIndvars && + !isa<SCEVCouldNotCompute>(BackedgeTakenCount)) { SE->forgetLoop(L); const SCEV *NewBECount = SE->getBackedgeTakenCount(L); if (SE->getTypeSizeInBits(BackedgeTakenCount->getType()) < |