X-Git-Url: http://plrg.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FAnalysis%2FScalarEvolution.cpp;h=069f6ec714cc54a01d21df51ad5885a0b0387927;hb=9a2f93121b31bf6345d1552bdc43037f89714d86;hp=0ffe79e5394cc25c3d3b5eeb3dabce9ab245d56c;hpb=6128690c629380291b7185f09cd82ed5244c8a41;p=oota-llvm.git diff --git a/lib/Analysis/ScalarEvolution.cpp b/lib/Analysis/ScalarEvolution.cpp index 0ffe79e5394..069f6ec714c 100644 --- a/lib/Analysis/ScalarEvolution.cpp +++ b/lib/Analysis/ScalarEvolution.cpp @@ -1,10 +1,10 @@ //===- ScalarEvolution.cpp - Scalar Evolution Analysis ----------*- C++ -*-===// -// +// // The LLVM Compiler Infrastructure // // This file was developed by the LLVM research group and is distributed under // the University of Illinois Open Source License. See LICENSE.TXT for details. -// +// //===----------------------------------------------------------------------===// // // This file contains the implementation of the scalar evolution analysis @@ -28,7 +28,7 @@ // have folders that are used to build the *canonical* representation for a // particular expression. These folders are capable of using a variety of // rewrite rules to simplify the expressions. -// +// // Once the folders are defined, we can implement the more interesting // higher-level code, such as the code that recognizes PHI nodes of various // types, computes the execution count of a loop, etc. @@ -59,50 +59,53 @@ // //===----------------------------------------------------------------------===// +#define DEBUG_TYPE "scalar-evolution" #include "llvm/Analysis/ScalarEvolutionExpressions.h" #include "llvm/Constants.h" #include "llvm/DerivedTypes.h" #include "llvm/GlobalVariable.h" #include "llvm/Instructions.h" +#include "llvm/Analysis/ConstantFolding.h" #include "llvm/Analysis/LoopInfo.h" #include "llvm/Assembly/Writer.h" #include "llvm/Transforms/Scalar.h" -#include "llvm/Transforms/Utils/Local.h" #include "llvm/Support/CFG.h" +#include "llvm/Support/CommandLine.h" +#include "llvm/Support/Compiler.h" #include "llvm/Support/ConstantRange.h" #include "llvm/Support/InstIterator.h" -#include "llvm/Support/CommandLine.h" +#include "llvm/Support/ManagedStatic.h" +#include "llvm/Support/MathExtras.h" +#include "llvm/Support/Streams.h" #include "llvm/ADT/Statistic.h" -#include +#include #include +#include using namespace llvm; +STATISTIC(NumBruteForceEvaluations, + "Number of brute force evaluations needed to " + "calculate high-order polynomial exit values"); +STATISTIC(NumArrayLenItCounts, + "Number of trip counts computed with array length"); +STATISTIC(NumTripCountsComputed, + "Number of loops with predictable loop counts"); +STATISTIC(NumTripCountsNotComputed, + "Number of loops without predictable loop counts"); +STATISTIC(NumBruteForceTripCountsComputed, + "Number of loops with trip counts computed by force"); + +cl::opt +MaxBruteForceIterations("scalar-evolution-max-iterations", cl::ReallyHidden, + cl::desc("Maximum number of iterations SCEV will " + "symbolically execute a constant derived loop"), + cl::init(100)); + namespace { - RegisterAnalysis + RegisterPass R("scalar-evolution", "Scalar Evolution Analysis"); - - Statistic<> - NumBruteForceEvaluations("scalar-evolution", - "Number of brute force evaluations needed to " - "calculate high-order polynomial exit values"); - Statistic<> - NumArrayLenItCounts("scalar-evolution", - "Number of trip counts computed with array length"); - Statistic<> - NumTripCountsComputed("scalar-evolution", - "Number of loops with predictable loop counts"); - Statistic<> - NumTripCountsNotComputed("scalar-evolution", - "Number of loops without predictable loop counts"); - Statistic<> - NumBruteForceTripCountsComputed("scalar-evolution", - "Number of loops with trip counts computed by force"); - - cl::opt - MaxBruteForceIterations("scalar-evolution-max-iterations", cl::ReallyHidden, - cl::desc("Maximum number of iterations SCEV will symbolically execute a constant derived loop"), - cl::init(100)); } +char ScalarEvolution::ID = 0; //===----------------------------------------------------------------------===// // SCEV class definitions @@ -113,7 +116,7 @@ namespace { // SCEV::~SCEV() {} void SCEV::dump() const { - print(std::cerr); + print(cerr); } /// getValueRange - Return the tightest constant bounds that this value is @@ -121,9 +124,14 @@ void SCEV::dump() const { ConstantRange SCEV::getValueRange() const { const Type *Ty = getType(); assert(Ty->isInteger() && "Can't get range for a non-integer SCEV!"); - Ty = Ty->getUnsignedVersion(); // Default to a full range if no better information is available. - return ConstantRange(getType()); + return ConstantRange(getBitWidth()); +} + +uint32_t SCEV::getBitWidth() const { + if (const IntegerType* ITy = dyn_cast(getType())) + return ITy->getBitWidth(); + return 0; } @@ -162,27 +170,25 @@ bool SCEVCouldNotCompute::classof(const SCEV *S) { // SCEVConstants - Only allow the creation of one SCEVConstant for any // particular value. Don't use a SCEVHandle here, or else the object will // never be deleted! -static std::map SCEVConstants; - +static ManagedStatic > SCEVConstants; + SCEVConstant::~SCEVConstant() { - SCEVConstants.erase(V); + SCEVConstants->erase(V); } SCEVHandle SCEVConstant::get(ConstantInt *V) { - // Make sure that SCEVConstant instances are all unsigned. - if (V->getType()->isSigned()) { - const Type *NewTy = V->getType()->getUnsignedVersion(); - V = cast(ConstantExpr::getCast(V, NewTy)); - } - - SCEVConstant *&R = SCEVConstants[V]; + SCEVConstant *&R = (*SCEVConstants)[V]; if (R == 0) R = new SCEVConstant(V); return R; } +SCEVHandle SCEVConstant::get(const APInt& Val) { + return get(ConstantInt::get(Val)); +} + ConstantRange SCEVConstant::getValueRange() const { - return ConstantRange(V); + return ConstantRange(V->getValue()); } const Type *SCEVConstant::getType() const { return V->getType(); } @@ -194,23 +200,23 @@ void SCEVConstant::print(std::ostream &OS) const { // SCEVTruncates - Only allow the creation of one SCEVTruncateExpr for any // particular input. Don't use a SCEVHandle here, or else the object will // never be deleted! -static std::map, SCEVTruncateExpr*> SCEVTruncates; +static ManagedStatic, + SCEVTruncateExpr*> > SCEVTruncates; SCEVTruncateExpr::SCEVTruncateExpr(const SCEVHandle &op, const Type *ty) : SCEV(scTruncate), Op(op), Ty(ty) { assert(Op->getType()->isInteger() && Ty->isInteger() && - Ty->isUnsigned() && "Cannot truncate non-integer value!"); - assert(Op->getType()->getPrimitiveSize() > Ty->getPrimitiveSize() && - "This is not a truncating conversion!"); + assert(Op->getType()->getPrimitiveSizeInBits() > Ty->getPrimitiveSizeInBits() + && "This is not a truncating conversion!"); } SCEVTruncateExpr::~SCEVTruncateExpr() { - SCEVTruncates.erase(std::make_pair(Op, Ty)); + SCEVTruncates->erase(std::make_pair(Op, Ty)); } ConstantRange SCEVTruncateExpr::getValueRange() const { - return getOperand()->getValueRange().truncate(getType()); + return getOperand()->getValueRange().truncate(getBitWidth()); } void SCEVTruncateExpr::print(std::ostream &OS) const { @@ -220,40 +226,65 @@ void SCEVTruncateExpr::print(std::ostream &OS) const { // SCEVZeroExtends - Only allow the creation of one SCEVZeroExtendExpr for any // particular input. Don't use a SCEVHandle here, or else the object will never // be deleted! -static std::map, - SCEVZeroExtendExpr*> SCEVZeroExtends; +static ManagedStatic, + SCEVZeroExtendExpr*> > SCEVZeroExtends; SCEVZeroExtendExpr::SCEVZeroExtendExpr(const SCEVHandle &op, const Type *ty) - : SCEV(scTruncate), Op(op), Ty(ty) { + : SCEV(scZeroExtend), Op(op), Ty(ty) { assert(Op->getType()->isInteger() && Ty->isInteger() && - Ty->isUnsigned() && "Cannot zero extend non-integer value!"); - assert(Op->getType()->getPrimitiveSize() < Ty->getPrimitiveSize() && - "This is not an extending conversion!"); + assert(Op->getType()->getPrimitiveSizeInBits() < Ty->getPrimitiveSizeInBits() + && "This is not an extending conversion!"); } SCEVZeroExtendExpr::~SCEVZeroExtendExpr() { - SCEVZeroExtends.erase(std::make_pair(Op, Ty)); + SCEVZeroExtends->erase(std::make_pair(Op, Ty)); } ConstantRange SCEVZeroExtendExpr::getValueRange() const { - return getOperand()->getValueRange().zeroExtend(getType()); + return getOperand()->getValueRange().zeroExtend(getBitWidth()); } void SCEVZeroExtendExpr::print(std::ostream &OS) const { OS << "(zeroextend " << *Op << " to " << *Ty << ")"; } +// SCEVSignExtends - Only allow the creation of one SCEVSignExtendExpr for any +// particular input. Don't use a SCEVHandle here, or else the object will never +// be deleted! +static ManagedStatic, + SCEVSignExtendExpr*> > SCEVSignExtends; + +SCEVSignExtendExpr::SCEVSignExtendExpr(const SCEVHandle &op, const Type *ty) + : SCEV(scSignExtend), Op(op), Ty(ty) { + assert(Op->getType()->isInteger() && Ty->isInteger() && + "Cannot sign extend non-integer value!"); + assert(Op->getType()->getPrimitiveSizeInBits() < Ty->getPrimitiveSizeInBits() + && "This is not an extending conversion!"); +} + +SCEVSignExtendExpr::~SCEVSignExtendExpr() { + SCEVSignExtends->erase(std::make_pair(Op, Ty)); +} + +ConstantRange SCEVSignExtendExpr::getValueRange() const { + return getOperand()->getValueRange().signExtend(getBitWidth()); +} + +void SCEVSignExtendExpr::print(std::ostream &OS) const { + OS << "(signextend " << *Op << " to " << *Ty << ")"; +} + // SCEVCommExprs - Only allow the creation of one SCEVCommutativeExpr for any // particular input. Don't use a SCEVHandle here, or else the object will never // be deleted! -static std::map >, - SCEVCommutativeExpr*> SCEVCommExprs; +static ManagedStatic >, + SCEVCommutativeExpr*> > SCEVCommExprs; SCEVCommutativeExpr::~SCEVCommutativeExpr() { - SCEVCommExprs.erase(std::make_pair(getSCEVType(), - std::vector(Operands.begin(), - Operands.end()))); + SCEVCommExprs->erase(std::make_pair(getSCEVType(), + std::vector(Operands.begin(), + Operands.end()))); } void SCEVCommutativeExpr::print(std::ostream &OS) const { @@ -292,35 +323,34 @@ replaceSymbolicValuesWithConcrete(const SCEVHandle &Sym, } -// SCEVUDivs - Only allow the creation of one SCEVUDivExpr for any particular +// SCEVSDivs - Only allow the creation of one SCEVSDivExpr for any particular // input. Don't use a SCEVHandle here, or else the object will never be // deleted! -static std::map, SCEVUDivExpr*> SCEVUDivs; +static ManagedStatic, + SCEVSDivExpr*> > SCEVSDivs; -SCEVUDivExpr::~SCEVUDivExpr() { - SCEVUDivs.erase(std::make_pair(LHS, RHS)); +SCEVSDivExpr::~SCEVSDivExpr() { + SCEVSDivs->erase(std::make_pair(LHS, RHS)); } -void SCEVUDivExpr::print(std::ostream &OS) const { - OS << "(" << *LHS << " /u " << *RHS << ")"; +void SCEVSDivExpr::print(std::ostream &OS) const { + OS << "(" << *LHS << " /s " << *RHS << ")"; } -const Type *SCEVUDivExpr::getType() const { - const Type *Ty = LHS->getType(); - if (Ty->isSigned()) Ty = Ty->getUnsignedVersion(); - return Ty; +const Type *SCEVSDivExpr::getType() const { + return LHS->getType(); } // SCEVAddRecExprs - Only allow the creation of one SCEVAddRecExpr for any // particular input. Don't use a SCEVHandle here, or else the object will never // be deleted! -static std::map >, - SCEVAddRecExpr*> SCEVAddRecExprs; +static ManagedStatic >, + SCEVAddRecExpr*> > SCEVAddRecExprs; SCEVAddRecExpr::~SCEVAddRecExpr() { - SCEVAddRecExprs.erase(std::make_pair(L, - std::vector(Operands.begin(), - Operands.end()))); + SCEVAddRecExprs->erase(std::make_pair(L, + std::vector(Operands.begin(), + Operands.end()))); } SCEVHandle SCEVAddRecExpr:: @@ -337,7 +367,7 @@ replaceSymbolicValuesWithConcrete(const SCEVHandle &Sym, for (++i; i != e; ++i) NewOps.push_back(getOperand(i)-> replaceSymbolicValuesWithConcrete(Sym, Conc)); - + return get(NewOps, L); } } @@ -347,8 +377,9 @@ replaceSymbolicValuesWithConcrete(const SCEVHandle &Sym, bool SCEVAddRecExpr::isLoopInvariant(const Loop *QueryLoop) const { // This recurrence is invariant w.r.t to QueryLoop iff QueryLoop doesn't - // contain L. - return !QueryLoop->contains(L->getHeader()); + // contain L and if the start is invariant. + return !QueryLoop->contains(L->getHeader()) && + getOperand(0)->isLoopInvariant(QueryLoop); } @@ -362,9 +393,9 @@ void SCEVAddRecExpr::print(std::ostream &OS) const { // SCEVUnknowns - Only allow the creation of one SCEVUnknown for any particular // value. Don't use a SCEVHandle here, or else the object will never be // deleted! -static std::map SCEVUnknowns; +static ManagedStatic > SCEVUnknowns; -SCEVUnknown::~SCEVUnknown() { SCEVUnknowns.erase(V); } +SCEVUnknown::~SCEVUnknown() { SCEVUnknowns->erase(V); } bool SCEVUnknown::isLoopInvariant(const Loop *L) const { // All non-instruction values are loop invariant. All instructions are loop @@ -390,7 +421,7 @@ namespace { /// SCEVComplexityCompare - Return true if the complexity of the LHS is less /// than the complexity of the RHS. This comparator is used to canonicalize /// expressions. - struct SCEVComplexityCompare { + struct VISIBILITY_HIDDEN SCEVComplexityCompare { bool operator()(SCEV *LHS, SCEV *RHS) { return LHS->getSCEVType() < RHS->getSCEVType(); } @@ -451,16 +482,13 @@ static void GroupByComplexity(std::vector &Ops) { /// specified signed integer value and return a SCEV for the constant. SCEVHandle SCEVUnknown::getIntegerSCEV(int Val, const Type *Ty) { Constant *C; - if (Val == 0) + if (Val == 0) C = Constant::getNullValue(Ty); else if (Ty->isFloatingPoint()) - C = ConstantFP::get(Ty, Val); - else if (Ty->isSigned()) - C = ConstantSInt::get(Ty, Val); - else { - C = ConstantSInt::get(Ty->getSignedVersion(), Val); - C = ConstantExpr::getCast(C, Ty); - } + C = ConstantFP::get(Ty, APFloat(Ty==Type::FloatTy ? APFloat::IEEEsingle : + APFloat::IEEEdouble, Val)); + else + C = ConstantInt::get(Ty, Val); return SCEVUnknown::get(C); } @@ -471,9 +499,9 @@ static SCEVHandle getTruncateOrZeroExtend(const SCEVHandle &V, const Type *Ty) { const Type *SrcTy = V->getType(); assert(SrcTy->isInteger() && Ty->isInteger() && "Cannot truncate or zero extend with non-integer arguments!"); - if (SrcTy->getPrimitiveSize() == Ty->getPrimitiveSize()) + if (SrcTy->getPrimitiveSizeInBits() == Ty->getPrimitiveSizeInBits()) return V; // No conversion - if (SrcTy->getPrimitiveSize() > Ty->getPrimitiveSize()) + if (SrcTy->getPrimitiveSizeInBits() > Ty->getPrimitiveSizeInBits()) return SCEVTruncateExpr::get(V, Ty); return SCEVZeroExtendExpr::get(V, Ty); } @@ -483,7 +511,7 @@ static SCEVHandle getTruncateOrZeroExtend(const SCEVHandle &V, const Type *Ty) { SCEVHandle SCEV::getNegativeSCEV(const SCEVHandle &V) { if (SCEVConstant *VC = dyn_cast(V)) return SCEVUnknown::get(ConstantExpr::getNeg(VC->getValue())); - + return SCEVMulExpr::get(V, SCEVUnknown::getIntegerSCEV(-1, V->getType())); } @@ -500,18 +528,17 @@ static SCEVHandle PartialFact(SCEVHandle V, unsigned NumSteps) { // Handle this case efficiently, it is common to have constant iteration // counts while computing loop exit values. if (SCEVConstant *SC = dyn_cast(V)) { - uint64_t Val = SC->getValue()->getRawValue(); - uint64_t Result = 1; + const APInt& Val = SC->getValue()->getValue(); + APInt Result(Val.getBitWidth(), 1); for (; NumSteps; --NumSteps) Result *= Val-(NumSteps-1); - Constant *Res = ConstantUInt::get(Type::ULongTy, Result); - return SCEVUnknown::get(ConstantExpr::getCast(Res, V->getType())); + return SCEVConstant::get(Result); } const Type *Ty = V->getType(); if (NumSteps == 0) return SCEVUnknown::getIntegerSCEV(1, Ty); - + SCEVHandle Result = V; for (unsigned i = 1; i != NumSteps; ++i) Result = SCEVMulExpr::get(Result, SCEV::getMinusSCEV(V, @@ -537,7 +564,7 @@ SCEVHandle SCEVAddRecExpr::evaluateAtIteration(SCEVHandle It) const { for (unsigned i = 1, e = getNumOperands(); i != e; ++i) { SCEVHandle BC = PartialFact(It, i); Divisor *= i; - SCEVHandle Val = SCEVUDivExpr::get(SCEVMulExpr::get(BC, getOperand(i)), + SCEVHandle Val = SCEVSDivExpr::get(SCEVMulExpr::get(BC, getOperand(i)), SCEVUnknown::getIntegerSCEV(Divisor,Ty)); Result = SCEVAddExpr::get(Result, Val); } @@ -551,7 +578,8 @@ SCEVHandle SCEVAddRecExpr::evaluateAtIteration(SCEVHandle It) const { SCEVHandle SCEVTruncateExpr::get(const SCEVHandle &Op, const Type *Ty) { if (SCEVConstant *SC = dyn_cast(Op)) - return SCEVUnknown::get(ConstantExpr::getCast(SC->getValue(), Ty)); + return SCEVUnknown::get( + ConstantExpr::getTrunc(SC->getValue(), Ty)); // If the input value is a chrec scev made out of constants, truncate // all of the constants. @@ -567,25 +595,41 @@ SCEVHandle SCEVTruncateExpr::get(const SCEVHandle &Op, const Type *Ty) { return SCEVAddRecExpr::get(Operands, AddRec->getLoop()); } - SCEVTruncateExpr *&Result = SCEVTruncates[std::make_pair(Op, Ty)]; + SCEVTruncateExpr *&Result = (*SCEVTruncates)[std::make_pair(Op, Ty)]; if (Result == 0) Result = new SCEVTruncateExpr(Op, Ty); return Result; } SCEVHandle SCEVZeroExtendExpr::get(const SCEVHandle &Op, const Type *Ty) { if (SCEVConstant *SC = dyn_cast(Op)) - return SCEVUnknown::get(ConstantExpr::getCast(SC->getValue(), Ty)); + return SCEVUnknown::get( + ConstantExpr::getZExt(SC->getValue(), Ty)); // FIXME: If the input value is a chrec scev, and we can prove that the value // did not overflow the old, smaller, value, we can zero extend all of the // operands (often constants). This would allow analysis of something like // this: for (unsigned char X = 0; X < 100; ++X) { int Y = X; } - SCEVZeroExtendExpr *&Result = SCEVZeroExtends[std::make_pair(Op, Ty)]; + SCEVZeroExtendExpr *&Result = (*SCEVZeroExtends)[std::make_pair(Op, Ty)]; if (Result == 0) Result = new SCEVZeroExtendExpr(Op, Ty); return Result; } +SCEVHandle SCEVSignExtendExpr::get(const SCEVHandle &Op, const Type *Ty) { + if (SCEVConstant *SC = dyn_cast(Op)) + return SCEVUnknown::get( + ConstantExpr::getSExt(SC->getValue(), Ty)); + + // FIXME: If the input value is a chrec scev, and we can prove that the value + // did not overflow the old, smaller, value, we can sign extend all of the + // operands (often constants). This would allow analysis of something like + // this: for (signed char X = 0; X < 100; ++X) { int Y = X; } + + SCEVSignExtendExpr *&Result = (*SCEVSignExtends)[std::make_pair(Op, Ty)]; + if (Result == 0) Result = new SCEVSignExtendExpr(Op, Ty); + return Result; +} + // get - Get a canonical add expression, or something simpler if possible. SCEVHandle SCEVAddExpr::get(std::vector &Ops) { assert(!Ops.empty() && "Cannot get empty add!"); @@ -601,7 +645,8 @@ SCEVHandle SCEVAddExpr::get(std::vector &Ops) { assert(Idx < Ops.size()); while (SCEVConstant *RHSC = dyn_cast(Ops[Idx])) { // We found two constants, fold them together! - Constant *Fold = ConstantExpr::getAdd(LHSC->getValue(), RHSC->getValue()); + Constant *Fold = ConstantInt::get(LHSC->getValue()->getValue() + + RHSC->getValue()->getValue()); if (ConstantInt *CI = dyn_cast(Fold)) { Ops[0] = SCEVConstant::get(CI); Ops.erase(Ops.begin()+1); // Erase the folded element @@ -616,14 +661,14 @@ SCEVHandle SCEVAddExpr::get(std::vector &Ops) { } // If we are left with a constant zero being added, strip it off. - if (cast(Ops[0])->getValue()->isNullValue()) { + if (cast(Ops[0])->getValue()->isZero()) { Ops.erase(Ops.begin()); --Idx; } } if (Ops.size() == 1) return Ops[0]; - + // Okay, check to see if the same value occurs in the operand list twice. If // so, merge them together into an multiply expression. Since we sorted the // list, these values are required to be adjacent. @@ -641,8 +686,11 @@ SCEVHandle SCEVAddExpr::get(std::vector &Ops) { return SCEVAddExpr::get(Ops); } - // Okay, now we know the first non-constant operand. If there are add - // operands they would be next. + // Now we know the first non-constant operand. Skip past any cast SCEVs. + while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddExpr) + ++Idx; + + // If there are add operands they would be next. if (Idx < Ops.size()) { bool DeletedAdd = false; while (SCEVAddExpr *Add = dyn_cast(Ops[Idx])) { @@ -696,7 +744,7 @@ SCEVHandle SCEVAddExpr::get(std::vector &Ops) { Ops.push_back(OuterMul); return SCEVAddExpr::get(Ops); } - + // Check this multiply against other multiplies being added together. for (unsigned OtherMulIdx = Idx+1; OtherMulIdx < Ops.size() && isa(Ops[OtherMulIdx]); @@ -809,8 +857,8 @@ SCEVHandle SCEVAddExpr::get(std::vector &Ops) { // Okay, it looks like we really DO need an add expr. Check to see if we // already have one, otherwise create a new one. std::vector SCEVOps(Ops.begin(), Ops.end()); - SCEVCommutativeExpr *&Result = SCEVCommExprs[std::make_pair(scAddExpr, - SCEVOps)]; + SCEVCommutativeExpr *&Result = (*SCEVCommExprs)[std::make_pair(scAddExpr, + SCEVOps)]; if (Result == 0) Result = new SCEVAddExpr(Ops); return Result; } @@ -838,7 +886,8 @@ SCEVHandle SCEVMulExpr::get(std::vector &Ops) { ++Idx; while (SCEVConstant *RHSC = dyn_cast(Ops[Idx])) { // We found two constants, fold them together! - Constant *Fold = ConstantExpr::getMul(LHSC->getValue(), RHSC->getValue()); + Constant *Fold = ConstantInt::get(LHSC->getValue()->getValue() * + RHSC->getValue()->getValue()); if (ConstantInt *CI = dyn_cast(Fold)) { Ops[0] = SCEVConstant::get(CI); Ops.erase(Ops.begin()+1); // Erase the folded element @@ -856,7 +905,7 @@ SCEVHandle SCEVMulExpr::get(std::vector &Ops) { if (cast(Ops[0])->getValue()->equalsInt(1)) { Ops.erase(Ops.begin()); --Idx; - } else if (cast(Ops[0])->getValue()->isNullValue()) { + } else if (cast(Ops[0])->getValue()->isZero()) { // If we have a multiply of zero, it will always be zero. return Ops[0]; } @@ -868,7 +917,7 @@ SCEVHandle SCEVMulExpr::get(std::vector &Ops) { if (Ops.size() == 1) return Ops[0]; - + // If there are mul operands inline them all into this expression. if (Idx < Ops.size()) { bool DeletedMul = false; @@ -972,36 +1021,31 @@ SCEVHandle SCEVMulExpr::get(std::vector &Ops) { // Okay, it looks like we really DO need an mul expr. Check to see if we // already have one, otherwise create a new one. std::vector SCEVOps(Ops.begin(), Ops.end()); - SCEVCommutativeExpr *&Result = SCEVCommExprs[std::make_pair(scMulExpr, - SCEVOps)]; + SCEVCommutativeExpr *&Result = (*SCEVCommExprs)[std::make_pair(scMulExpr, + SCEVOps)]; if (Result == 0) Result = new SCEVMulExpr(Ops); return Result; } -SCEVHandle SCEVUDivExpr::get(const SCEVHandle &LHS, const SCEVHandle &RHS) { +SCEVHandle SCEVSDivExpr::get(const SCEVHandle &LHS, const SCEVHandle &RHS) { if (SCEVConstant *RHSC = dyn_cast(RHS)) { if (RHSC->getValue()->equalsInt(1)) - return LHS; // X /u 1 --> x + return LHS; // X sdiv 1 --> x if (RHSC->getValue()->isAllOnesValue()) - return SCEV::getNegativeSCEV(LHS); // X /u -1 --> -x + return SCEV::getNegativeSCEV(LHS); // X sdiv -1 --> -x if (SCEVConstant *LHSC = dyn_cast(LHS)) { Constant *LHSCV = LHSC->getValue(); Constant *RHSCV = RHSC->getValue(); - if (LHSCV->getType()->isSigned()) - LHSCV = ConstantExpr::getCast(LHSCV, - LHSCV->getType()->getUnsignedVersion()); - if (RHSCV->getType()->isSigned()) - RHSCV = ConstantExpr::getCast(RHSCV, LHSCV->getType()); - return SCEVUnknown::get(ConstantExpr::getDiv(LHSCV, RHSCV)); + return SCEVUnknown::get(ConstantExpr::getSDiv(LHSCV, RHSCV)); } } // FIXME: implement folding of (X*4)/4 when we know X*4 doesn't overflow. - SCEVUDivExpr *&Result = SCEVUDivs[std::make_pair(LHS, RHS)]; - if (Result == 0) Result = new SCEVUDivExpr(LHS, RHS); + SCEVSDivExpr *&Result = (*SCEVSDivs)[std::make_pair(LHS, RHS)]; + if (Result == 0) Result = new SCEVSDivExpr(LHS, RHS); return Result; } @@ -1030,14 +1074,14 @@ SCEVHandle SCEVAddRecExpr::get(std::vector &Operands, if (Operands.size() == 1) return Operands[0]; if (SCEVConstant *StepC = dyn_cast(Operands.back())) - if (StepC->getValue()->isNullValue()) { + if (StepC->getValue()->isZero()) { Operands.pop_back(); return get(Operands, L); // { X,+,0 } --> X } SCEVAddRecExpr *&Result = - SCEVAddRecExprs[std::make_pair(L, std::vector(Operands.begin(), - Operands.end()))]; + (*SCEVAddRecExprs)[std::make_pair(L, std::vector(Operands.begin(), + Operands.end()))]; if (Result == 0) Result = new SCEVAddRecExpr(Operands, L); return Result; } @@ -1045,7 +1089,7 @@ SCEVHandle SCEVAddRecExpr::get(std::vector &Operands, SCEVHandle SCEVUnknown::get(Value *V) { if (ConstantInt *CI = dyn_cast(V)) return SCEVConstant::get(CI); - SCEVUnknown *&Result = SCEVUnknowns[V]; + SCEVUnknown *&Result = (*SCEVUnknowns)[V]; if (Result == 0) Result = new SCEVUnknown(V); return Result; } @@ -1059,7 +1103,7 @@ SCEVHandle SCEVUnknown::get(Value *V) { /// evolution code. /// namespace { - struct ScalarEvolutionsImpl { + struct VISIBILITY_HIDDEN ScalarEvolutionsImpl { /// F - The function we are analyzing. /// Function &F; @@ -1086,7 +1130,7 @@ namespace { /// properties. An instruction maps to null if we are unable to compute its /// exit value. std::map ConstantEvolutionLoopExitValue; - + public: ScalarEvolutionsImpl(Function &f, LoopInfo &li) : F(f), LI(li), UnknownValue(new SCEVCouldNotCompute()) {} @@ -1095,6 +1139,20 @@ namespace { /// expression and create a new one. SCEVHandle getSCEV(Value *V); + /// hasSCEV - Return true if the SCEV for this value has already been + /// computed. + bool hasSCEV(Value *V) const { + return Scalars.count(V); + } + + /// setSCEV - Insert the specified SCEV into the map of current SCEVs for + /// the specified value. + void setSCEV(Value *V, const SCEVHandle &H) { + bool isNew = Scalars.insert(std::make_pair(V, H)).second; + assert(isNew && "This entry already existed!"); + } + + /// getSCEVAtScope - Compute the value of the specified expression within /// the indicated loop (which may be null to indicate in no loop). If the /// expression cannot be evaluated, return UnknownValue itself. @@ -1110,16 +1168,15 @@ namespace { /// loop without a loop-invariant iteration count. SCEVHandle getIterationCount(const Loop *L); - /// deleteInstructionFromRecords - This method should be called by the - /// client before it removes an instruction from the program, to make sure + /// deleteValueFromRecords - This method should be called by the + /// client before it removes a value from the program, to make sure /// that no dangling references are left around. - void deleteInstructionFromRecords(Instruction *I); + void deleteValueFromRecords(Value *V); private: /// createSCEV - We know that there is no SCEV for the specified value. /// Analyze the expression. SCEVHandle createSCEV(Value *V); - SCEVHandle createNodeForCast(CastInst *CI); /// createNodeForPHI - Provide the special handling we need to analyze PHI /// SCEVs. @@ -1138,11 +1195,11 @@ namespace { SCEVHandle ComputeIterationCount(const Loop *L); /// ComputeLoadConstantCompareIterationCount - Given an exit condition of - /// 'setcc load X, cst', try to se if we can compute the trip count. + /// 'setcc load X, cst', try to see if we can compute the trip count. SCEVHandle ComputeLoadConstantCompareIterationCount(LoadInst *LI, Constant *RHS, const Loop *L, - unsigned SetCCOpcode); + ICmpInst::Predicate p); /// ComputeIterationCountExhaustively - If the trip is known to execute a /// constant number of times (the condition evolves only from constants), @@ -1154,19 +1211,25 @@ namespace { /// HowFarToZero - Return the number of times a backedge comparing the /// specified value to zero will execute. If not computable, return - /// UnknownValue + /// UnknownValue. SCEVHandle HowFarToZero(SCEV *V, const Loop *L); /// HowFarToNonZero - Return the number of times a backedge checking the /// specified value for nonzero will execute. If not computable, return - /// UnknownValue + /// UnknownValue. SCEVHandle HowFarToNonZero(SCEV *V, const Loop *L); + /// HowManyLessThans - Return the number of times a backedge containing the + /// specified less-than comparison will execute. If not computable, return + /// UnknownValue. isSigned specifies whether the less-than is signed. + SCEVHandle HowManyLessThans(SCEV *LHS, SCEV *RHS, const Loop *L, + bool isSigned); + /// getConstantEvolutionLoopExitValue - If we know that the specified Phi is /// in the header of its containing loop, we know the loop executes a /// constant number of times, and the PHI node is just a recurrence /// involving constants, fold it. - Constant *getConstantEvolutionLoopExitValue(PHINode *PN, uint64_t Its, + Constant *getConstantEvolutionLoopExitValue(PHINode *PN, const APInt& Its, const Loop *L); }; } @@ -1175,13 +1238,32 @@ namespace { // Basic SCEV Analysis and PHI Idiom Recognition Code // -/// deleteInstructionFromRecords - This method should be called by the +/// deleteValueFromRecords - This method should be called by the /// client before it removes an instruction from the program, to make sure /// that no dangling references are left around. -void ScalarEvolutionsImpl::deleteInstructionFromRecords(Instruction *I) { - Scalars.erase(I); - if (PHINode *PN = dyn_cast(I)) - ConstantEvolutionLoopExitValue.erase(PN); +void ScalarEvolutionsImpl::deleteValueFromRecords(Value *V) { + SmallVector Worklist; + + if (Scalars.erase(V)) { + if (PHINode *PN = dyn_cast(V)) + ConstantEvolutionLoopExitValue.erase(PN); + Worklist.push_back(V); + } + + while (!Worklist.empty()) { + Value *VV = Worklist.back(); + Worklist.pop_back(); + + for (Instruction::use_iterator UI = VV->use_begin(), UE = VV->use_end(); + UI != UE; ++UI) { + Instruction *Inst = cast(*UI); + if (Scalars.erase(Inst)) { + if (PHINode *PN = dyn_cast(VV)) + ConstantEvolutionLoopExitValue.erase(PN); + Worklist.push_back(Inst); + } + } + } } @@ -1230,7 +1312,7 @@ SCEVHandle ScalarEvolutionsImpl::createNodeForPHI(PHINode *PN) { // from outside the loop, and one from inside. unsigned IncomingEdge = L->contains(PN->getIncomingBlock(0)); unsigned BackEdge = IncomingEdge^1; - + // While we are analyzing this PHI node, handle its value symbolically. SCEVHandle SymbolicName = SCEVUnknown::get(PN); assert(Scalars.find(PN) == Scalars.end() && @@ -1273,6 +1355,31 @@ SCEVHandle ScalarEvolutionsImpl::createNodeForPHI(PHINode *PN) { SCEVHandle StartVal = getSCEV(PN->getIncomingValue(IncomingEdge)); SCEVHandle PHISCEV = SCEVAddRecExpr::get(StartVal, Accum, L); + // Okay, for the entire analysis of this edge we assumed the PHI + // to be symbolic. We now need to go back and update all of the + // entries for the scalars that use the PHI (except for the PHI + // itself) to use the new analyzed value instead of the "symbolic" + // value. + ReplaceSymbolicValueWithConcrete(PN, SymbolicName, PHISCEV); + return PHISCEV; + } + } + } else if (SCEVAddRecExpr *AddRec = dyn_cast(BEValue)) { + // Otherwise, this could be a loop like this: + // i = 0; for (j = 1; ..; ++j) { .... i = j; } + // In this case, j = {1,+,1} and BEValue is j. + // Because the other in-value of i (0) fits the evolution of BEValue + // i really is an addrec evolution. + if (AddRec->getLoop() == L && AddRec->isAffine()) { + SCEVHandle StartVal = getSCEV(PN->getIncomingValue(IncomingEdge)); + + // If StartVal = j.start - j.stride, we can use StartVal as the + // initial step of the addrec evolution. + if (StartVal == SCEV::getMinusSCEV(AddRec->getOperand(0), + AddRec->getOperand(1))) { + SCEVHandle PHISCEV = + SCEVAddRecExpr::get(StartVal, AddRec->getOperand(1), L); + // Okay, for the entire analysis of this edge we assumed the PHI // to be symbolic. We now need to go back and update all of the // entries for the scalars that use the PHI (except for the PHI @@ -1286,41 +1393,70 @@ SCEVHandle ScalarEvolutionsImpl::createNodeForPHI(PHINode *PN) { return SymbolicName; } - + // If it's not a loop phi, we can't handle it yet. return SCEVUnknown::get(PN); } -/// createNodeForCast - Handle the various forms of casts that we support. -/// -SCEVHandle ScalarEvolutionsImpl::createNodeForCast(CastInst *CI) { - const Type *SrcTy = CI->getOperand(0)->getType(); - const Type *DestTy = CI->getType(); - - // If this is a noop cast (ie, conversion from int to uint), ignore it. - if (SrcTy->isLosslesslyConvertibleTo(DestTy)) - return getSCEV(CI->getOperand(0)); - - if (SrcTy->isInteger() && DestTy->isInteger()) { - // Otherwise, if this is a truncating integer cast, we can represent this - // cast. - if (SrcTy->getPrimitiveSize() > DestTy->getPrimitiveSize()) - return SCEVTruncateExpr::get(getSCEV(CI->getOperand(0)), - CI->getType()->getUnsignedVersion()); - if (SrcTy->isUnsigned() && - SrcTy->getPrimitiveSize() > DestTy->getPrimitiveSize()) - return SCEVZeroExtendExpr::get(getSCEV(CI->getOperand(0)), - CI->getType()->getUnsignedVersion()); +/// GetConstantFactor - Determine the largest constant factor that S has. For +/// example, turn {4,+,8} -> 4. (S umod result) should always equal zero. +static APInt GetConstantFactor(SCEVHandle S) { + if (SCEVConstant *C = dyn_cast(S)) { + const APInt& V = C->getValue()->getValue(); + if (!V.isMinValue()) + return V; + else // Zero is a multiple of everything. + return APInt(C->getBitWidth(), 1).shl(C->getBitWidth()-1); } - // If this is an sign or zero extending cast and we can prove that the value - // will never overflow, we could do similar transformations. + if (SCEVTruncateExpr *T = dyn_cast(S)) { + return GetConstantFactor(T->getOperand()).trunc( + cast(T->getType())->getBitWidth()); + } + if (SCEVZeroExtendExpr *E = dyn_cast(S)) + return GetConstantFactor(E->getOperand()).zext( + cast(E->getType())->getBitWidth()); + if (SCEVSignExtendExpr *E = dyn_cast(S)) + return GetConstantFactor(E->getOperand()).sext( + cast(E->getType())->getBitWidth()); + + if (SCEVAddExpr *A = dyn_cast(S)) { + // The result is the min of all operands. + APInt Res(GetConstantFactor(A->getOperand(0))); + for (unsigned i = 1, e = A->getNumOperands(); + i != e && Res.ugt(APInt(Res.getBitWidth(),1)); ++i) { + APInt Tmp(GetConstantFactor(A->getOperand(i))); + Res = APIntOps::umin(Res, Tmp); + } + return Res; + } - // Otherwise, we can't handle this cast! - return SCEVUnknown::get(CI); + if (SCEVMulExpr *M = dyn_cast(S)) { + // The result is the product of all the operands. + APInt Res(GetConstantFactor(M->getOperand(0))); + for (unsigned i = 1, e = M->getNumOperands(); i != e; ++i) { + APInt Tmp(GetConstantFactor(M->getOperand(i))); + Res *= Tmp; + } + return Res; + } + + if (SCEVAddRecExpr *A = dyn_cast(S)) { + // For now, we just handle linear expressions. + if (A->getNumOperands() == 2) { + // We want the GCD between the start and the stride value. + APInt Start(GetConstantFactor(A->getOperand(0))); + if (Start == 1) + return Start; + APInt Stride(GetConstantFactor(A->getOperand(1))); + return APIntOps::GreatestCommonDivisor(Start, Stride); + } + } + + // SCEVSDivExpr, SCEVUnknown. + return APInt(S->getBitWidth(), 1); } - /// createSCEV - We know that there is no SCEV for the specified value. /// Analyze the expression. /// @@ -1333,36 +1469,65 @@ SCEVHandle ScalarEvolutionsImpl::createSCEV(Value *V) { case Instruction::Mul: return SCEVMulExpr::get(getSCEV(I->getOperand(0)), getSCEV(I->getOperand(1))); - case Instruction::Div: - if (V->getType()->isInteger() && V->getType()->isUnsigned()) - return SCEVUDivExpr::get(getSCEV(I->getOperand(0)), - getSCEV(I->getOperand(1))); + case Instruction::SDiv: + return SCEVSDivExpr::get(getSCEV(I->getOperand(0)), + getSCEV(I->getOperand(1))); break; case Instruction::Sub: return SCEV::getMinusSCEV(getSCEV(I->getOperand(0)), getSCEV(I->getOperand(1))); + case Instruction::Or: + // If the RHS of the Or is a constant, we may have something like: + // X*4+1 which got turned into X*4|1. Handle this as an add so loop + // optimizations will transparently handle this case. + if (ConstantInt *CI = dyn_cast(I->getOperand(1))) { + SCEVHandle LHS = getSCEV(I->getOperand(0)); + APInt CommonFact(GetConstantFactor(LHS)); + assert(!CommonFact.isMinValue() && + "Common factor should at least be 1!"); + if (CommonFact.ugt(CI->getValue())) { + // If the LHS is a multiple that is larger than the RHS, use +. + return SCEVAddExpr::get(LHS, + getSCEV(I->getOperand(1))); + } + } + break; + case Instruction::Xor: + // If the RHS of the xor is a signbit, then this is just an add. + // Instcombine turns add of signbit into xor as a strength reduction step. + if (ConstantInt *CI = dyn_cast(I->getOperand(1))) { + if (CI->getValue().isSignBit()) + return SCEVAddExpr::get(getSCEV(I->getOperand(0)), + getSCEV(I->getOperand(1))); + } + break; case Instruction::Shl: // Turn shift left of a constant amount into a multiply. if (ConstantInt *SA = dyn_cast(I->getOperand(1))) { - Constant *X = ConstantInt::get(V->getType(), 1); - X = ConstantExpr::getShl(X, SA); + uint32_t BitWidth = cast(V->getType())->getBitWidth(); + Constant *X = ConstantInt::get( + APInt(BitWidth, 1).shl(SA->getLimitedValue(BitWidth))); return SCEVMulExpr::get(getSCEV(I->getOperand(0)), getSCEV(X)); } break; - case Instruction::Shr: - if (ConstantUInt *SA = dyn_cast(I->getOperand(1))) - if (V->getType()->isUnsigned()) { - Constant *X = ConstantInt::get(V->getType(), 1); - X = ConstantExpr::getShl(X, SA); - return SCEVUDivExpr::get(getSCEV(I->getOperand(0)), getSCEV(X)); - } - break; + case Instruction::Trunc: + return SCEVTruncateExpr::get(getSCEV(I->getOperand(0)), I->getType()); + + case Instruction::ZExt: + return SCEVZeroExtendExpr::get(getSCEV(I->getOperand(0)), I->getType()); - case Instruction::Cast: - return createNodeForCast(cast(I)); + case Instruction::SExt: + return SCEVSignExtendExpr::get(getSCEV(I->getOperand(0)), I->getType()); + + case Instruction::BitCast: + // BitCasts are no-op casts so we just eliminate the cast. + if (I->getType()->isInteger() && + I->getOperand(0)->getType()->isInteger()) + return getSCEV(I->getOperand(0)); + break; case Instruction::PHI: return createNodeForPHI(cast(I)); @@ -1405,7 +1570,7 @@ SCEVHandle ScalarEvolutionsImpl::getIterationCount(const Loop *L) { /// will iterate. SCEVHandle ScalarEvolutionsImpl::ComputeIterationCount(const Loop *L) { // If the loop has a non-one exit block count, we can't analyze it. - std::vector ExitBlocks; + SmallVector ExitBlocks; L->getExitBlocks(ExitBlocks); if (ExitBlocks.size() != 1) return UnknownValue; @@ -1428,21 +1593,40 @@ SCEVHandle ScalarEvolutionsImpl::ComputeIterationCount(const Loop *L) { // exit. // // FIXME: we should be able to handle switch instructions (with a single exit) - // FIXME: We should handle cast of int to bool as well BranchInst *ExitBr = dyn_cast(ExitingBlock->getTerminator()); if (ExitBr == 0) return UnknownValue; assert(ExitBr->isConditional() && "If unconditional, it can't be in loop!"); - SetCondInst *ExitCond = dyn_cast(ExitBr->getCondition()); - if (ExitCond == 0) // Not a setcc + + // At this point, we know we have a conditional branch that determines whether + // the loop is exited. However, we don't know if the branch is executed each + // time through the loop. If not, then the execution count of the branch will + // not be equal to the trip count of the loop. + // + // Currently we check for this by checking to see if the Exit branch goes to + // the loop header. If so, we know it will always execute the same number of + // times as the loop. We also handle the case where the exit block *is* the + // loop header. This is common for un-rotated loops. More extensive analysis + // could be done to handle more cases here. + if (ExitBr->getSuccessor(0) != L->getHeader() && + ExitBr->getSuccessor(1) != L->getHeader() && + ExitBr->getParent() != L->getHeader()) + return UnknownValue; + + ICmpInst *ExitCond = dyn_cast(ExitBr->getCondition()); + + // If its not an integer comparison then compute it the hard way. + // Note that ICmpInst deals with pointer comparisons too so we must check + // the type of the operand. + if (ExitCond == 0 || isa(ExitCond->getOperand(0)->getType())) return ComputeIterationCountExhaustively(L, ExitBr->getCondition(), ExitBr->getSuccessor(0) == ExitBlock); - // If the condition was exit on true, convert the condition to exit on false. - Instruction::BinaryOps Cond; + // If the condition was exit on true, convert the condition to exit on false + ICmpInst::Predicate Cond; if (ExitBr->getSuccessor(1) == ExitBlock) - Cond = ExitCond->getOpcode(); + Cond = ExitCond->getPredicate(); else - Cond = ExitCond->getInverseCondition(); + Cond = ExitCond->getInversePredicate(); // Handle common loops like: for (X = "string"; *X; ++X) if (LoadInst *LI = dyn_cast(ExitCond->getOperand(0))) @@ -1461,12 +1645,12 @@ SCEVHandle ScalarEvolutionsImpl::ComputeIterationCount(const Loop *L) { Tmp = getSCEVAtScope(RHS, L); if (!isa(Tmp)) RHS = Tmp; - // At this point, we would like to compute how many iterations of the loop the - // predicate will return true for these inputs. + // At this point, we would like to compute how many iterations of the + // loop the predicate will return true for these inputs. if (isa(LHS) && !isa(RHS)) { // If there is a constant, force it into the RHS. std::swap(LHS, RHS); - Cond = SetCondInst::getSwappedCondition(Cond); + Cond = ICmpInst::getSwappedPredicate(Cond); } // FIXME: think about handling pointer comparisons! i.e.: @@ -1482,58 +1666,71 @@ SCEVHandle ScalarEvolutionsImpl::ComputeIterationCount(const Loop *L) { // comparison. ConstantInt *CompVal = RHSC->getValue(); const Type *RealTy = ExitCond->getOperand(0)->getType(); - CompVal = dyn_cast(ConstantExpr::getCast(CompVal, RealTy)); + CompVal = dyn_cast( + ConstantExpr::getBitCast(CompVal, RealTy)); if (CompVal) { // Form the constant range. - ConstantRange CompRange(Cond, CompVal); - - // Now that we have it, if it's signed, convert it to an unsigned - // range. - if (CompRange.getLower()->getType()->isSigned()) { - const Type *NewTy = RHSC->getValue()->getType(); - Constant *NewL = ConstantExpr::getCast(CompRange.getLower(), NewTy); - Constant *NewU = ConstantExpr::getCast(CompRange.getUpper(), NewTy); - CompRange = ConstantRange(NewL, NewU); - } - + ConstantRange CompRange( + ICmpInst::makeConstantRange(Cond, CompVal->getValue())); + SCEVHandle Ret = AddRec->getNumIterationsInRange(CompRange); if (!isa(Ret)) return Ret; } } - + switch (Cond) { - case Instruction::SetNE: // while (X != Y) + case ICmpInst::ICMP_NE: { // while (X != Y) // Convert to: while (X-Y != 0) - if (LHS->getType()->isInteger()) { - SCEVHandle TC = HowFarToZero(SCEV::getMinusSCEV(LHS, RHS), L); - if (!isa(TC)) return TC; - } + SCEVHandle TC = HowFarToZero(SCEV::getMinusSCEV(LHS, RHS), L); + if (!isa(TC)) return TC; break; - case Instruction::SetEQ: + } + case ICmpInst::ICMP_EQ: { // Convert to: while (X-Y == 0) // while (X == Y) - if (LHS->getType()->isInteger()) { - SCEVHandle TC = HowFarToNonZero(SCEV::getMinusSCEV(LHS, RHS), L); - if (!isa(TC)) return TC; - } + SCEVHandle TC = HowFarToNonZero(SCEV::getMinusSCEV(LHS, RHS), L); + if (!isa(TC)) return TC; break; + } + case ICmpInst::ICMP_SLT: { + SCEVHandle TC = HowManyLessThans(LHS, RHS, L, true); + if (!isa(TC)) return TC; + break; + } + case ICmpInst::ICMP_SGT: { + SCEVHandle TC = HowManyLessThans(SCEV::getNegativeSCEV(LHS), + SCEV::getNegativeSCEV(RHS), L, true); + if (!isa(TC)) return TC; + break; + } + case ICmpInst::ICMP_ULT: { + SCEVHandle TC = HowManyLessThans(LHS, RHS, L, false); + if (!isa(TC)) return TC; + break; + } + case ICmpInst::ICMP_UGT: { + SCEVHandle TC = HowManyLessThans(SCEV::getNegativeSCEV(LHS), + SCEV::getNegativeSCEV(RHS), L, false); + if (!isa(TC)) return TC; + break; + } default: #if 0 - std::cerr << "ComputeIterationCount "; + cerr << "ComputeIterationCount "; if (ExitCond->getOperand(0)->getType()->isUnsigned()) - std::cerr << "[unsigned] "; - std::cerr << *LHS << " " - << Instruction::getOpcodeName(Cond) << " " << *RHS << "\n"; + cerr << "[unsigned] "; + cerr << *LHS << " " + << Instruction::getOpcodeName(Instruction::ICmp) + << " " << *RHS << "\n"; #endif break; } - return ComputeIterationCountExhaustively(L, ExitCond, - ExitBr->getSuccessor(0) == ExitBlock); + ExitBr->getSuccessor(0) == ExitBlock); } static ConstantInt * -EvaluateConstantChrecAtConstant(const SCEVAddRecExpr *AddRec, Constant *C) { - SCEVHandle InVal = SCEVConstant::get(cast(C)); +EvaluateConstantChrecAtConstant(const SCEVAddRecExpr *AddRec, ConstantInt *C) { + SCEVHandle InVal = SCEVConstant::get(C); SCEVHandle Val = AddRec->evaluateAtIteration(InVal); assert(isa(Val) && "Evaluation of SCEV at constant didn't fold correctly?"); @@ -1545,11 +1742,11 @@ EvaluateConstantChrecAtConstant(const SCEVAddRecExpr *AddRec, Constant *C) { /// the addressed element of the initializer or null if the index expression is /// invalid. static Constant * -GetAddressedElementFromGlobal(GlobalVariable *GV, +GetAddressedElementFromGlobal(GlobalVariable *GV, const std::vector &Indices) { Constant *Init = GV->getInitializer(); for (unsigned i = 0, e = Indices.size(); i != e; ++i) { - uint64_t Idx = Indices[i]->getRawValue(); + uint64_t Idx = Indices[i]->getZExtValue(); if (ConstantStruct *CS = dyn_cast(Init)) { assert(Idx < CS->getNumOperands() && "Bad struct index!"); Init = cast(CS->getOperand(Idx)); @@ -1577,8 +1774,9 @@ GetAddressedElementFromGlobal(GlobalVariable *GV, /// ComputeLoadConstantCompareIterationCount - Given an exit condition of /// 'setcc load X, cst', try to se if we can compute the trip count. SCEVHandle ScalarEvolutionsImpl:: -ComputeLoadConstantCompareIterationCount(LoadInst *LI, Constant *RHS, - const Loop *L, unsigned SetCCOpcode) { +ComputeLoadConstantCompareIterationCount(LoadInst *LI, Constant *RHS, + const Loop *L, + ICmpInst::Predicate predicate) { if (LI->isVolatile()) return UnknownValue; // Check to see if the loaded pointer is a getelementptr of a global. @@ -1623,8 +1821,8 @@ ComputeLoadConstantCompareIterationCount(LoadInst *LI, Constant *RHS, unsigned MaxSteps = MaxBruteForceIterations; for (unsigned IterationNum = 0; IterationNum != MaxSteps; ++IterationNum) { - ConstantUInt *ItCst = - ConstantUInt::get(IdxExpr->getType()->getUnsignedVersion(), IterationNum); + ConstantInt *ItCst = + ConstantInt::get(IdxExpr->getType(), IterationNum); ConstantInt *Val = EvaluateConstantChrecAtConstant(IdxExpr, ItCst); // Form the GEP offset. @@ -1634,13 +1832,13 @@ ComputeLoadConstantCompareIterationCount(LoadInst *LI, Constant *RHS, if (Result == 0) break; // Cannot compute! // Evaluate the condition for this iteration. - Result = ConstantExpr::get(SetCCOpcode, Result, RHS); - if (!isa(Result)) break; // Couldn't decide for sure - if (Result == ConstantBool::False) { + Result = ConstantExpr::getICmp(predicate, Result, RHS); + if (!isa(Result)) break; // Couldn't decide for sure + if (cast(Result)->getValue().isMinValue()) { #if 0 - std::cerr << "\n***\n*** Computed loop count " << *ItCst - << "\n*** From global " << *GV << "*** BB: " << *L->getHeader() - << "***\n"; + cerr << "\n***\n*** Computed loop count " << *ItCst + << "\n*** From global " << *GV << "*** BB: " << *L->getHeader() + << "***\n"; #endif ++NumArrayLenItCounts; return SCEVConstant::get(ItCst); // Found terminating iteration! @@ -1653,44 +1851,16 @@ ComputeLoadConstantCompareIterationCount(LoadInst *LI, Constant *RHS, /// CanConstantFold - Return true if we can constant fold an instruction of the /// specified type, assuming that all operands were constants. static bool CanConstantFold(const Instruction *I) { - if (isa(I) || isa(I) || + if (isa(I) || isa(I) || isa(I) || isa(I) || isa(I)) return true; - + if (const CallInst *CI = dyn_cast(I)) if (const Function *F = CI->getCalledFunction()) return canConstantFoldCallTo((Function*)F); // FIXME: elim cast return false; } -/// ConstantFold - Constant fold an instruction of the specified type with the -/// specified constant operands. This function may modify the operands vector. -static Constant *ConstantFold(const Instruction *I, - std::vector &Operands) { - if (isa(I) || isa(I)) - return ConstantExpr::get(I->getOpcode(), Operands[0], Operands[1]); - - switch (I->getOpcode()) { - case Instruction::Cast: - return ConstantExpr::getCast(Operands[0], I->getType()); - case Instruction::Select: - return ConstantExpr::getSelect(Operands[0], Operands[1], Operands[2]); - case Instruction::Call: - if (Function *GV = dyn_cast(Operands[0])) { - Operands.erase(Operands.begin()); - return ConstantFoldCall(cast(GV), Operands); - } - - return 0; - case Instruction::GetElementPtr: - Constant *Base = Operands[0]; - Operands.erase(Operands.begin()); - return ConstantExpr::getGetElementPtr(Base, Operands); - } - return 0; -} - - /// getConstantEvolvingPHI - Given an LLVM value and a loop, return a PHI node /// in the loop that V is derived from. We allow arbitrary operations along the /// way, but the operands of an operation must either be constants or a value @@ -1713,7 +1883,7 @@ static PHINode *getConstantEvolvingPHI(Value *V, const Loop *L) { // If we won't be able to constant fold this expression even if the operands // are constants, return early. if (!CanConstantFold(I)) return 0; - + // Otherwise, we can evaluate this instruction if all of its operands are // constant or derived from a PHI node themselves. PHINode *PHI = 0; @@ -1751,7 +1921,7 @@ static Constant *EvaluateExpression(Value *V, Constant *PHIVal) { if (Operands[i] == 0) return 0; } - return ConstantFold(I, Operands); + return ConstantFoldInstOperands(I, &Operands[0], Operands.size()); } /// getConstantEvolutionLoopExitValue - If we know that the specified Phi is @@ -1759,13 +1929,13 @@ static Constant *EvaluateExpression(Value *V, Constant *PHIVal) { /// constant number of times, and the PHI node is just a recurrence /// involving constants, fold it. Constant *ScalarEvolutionsImpl:: -getConstantEvolutionLoopExitValue(PHINode *PN, uint64_t Its, const Loop *L) { +getConstantEvolutionLoopExitValue(PHINode *PN, const APInt& Its, const Loop *L){ std::map::iterator I = ConstantEvolutionLoopExitValue.find(PN); if (I != ConstantEvolutionLoopExitValue.end()) return I->second; - if (Its > MaxBruteForceIterations) + if (Its.ugt(APInt(Its.getBitWidth(),MaxBruteForceIterations))) return ConstantEvolutionLoopExitValue[PN] = 0; // Not going to evaluate it. Constant *&RetVal = ConstantEvolutionLoopExitValue[PN]; @@ -1785,11 +1955,11 @@ getConstantEvolutionLoopExitValue(PHINode *PN, uint64_t Its, const Loop *L) { return RetVal = 0; // Not derived from same PHI. // Execute the loop symbolically to determine the exit value. - unsigned IterationNum = 0; - unsigned NumIterations = Its; - if (NumIterations != Its) - return RetVal = 0; // More than 2^32 iterations?? + if (Its.getActiveBits() >= 32) + return RetVal = 0; // More than 2^32-1 iterations?? Not doing it! + unsigned NumIterations = Its.getZExtValue(); // must be in range + unsigned IterationNum = 0; for (Constant *PHIVal = StartCST; ; ++IterationNum) { if (IterationNum == NumIterations) return RetVal = PHIVal; // Got exit value! @@ -1833,16 +2003,18 @@ ComputeIterationCountExhaustively(const Loop *L, Value *Cond, bool ExitWhen) { unsigned MaxIterations = MaxBruteForceIterations; // Limit analysis. for (Constant *PHIVal = StartCST; IterationNum != MaxIterations; ++IterationNum) { - ConstantBool *CondVal = - dyn_cast_or_null(EvaluateExpression(Cond, PHIVal)); - if (!CondVal) return UnknownValue; // Couldn't symbolically evaluate. + ConstantInt *CondVal = + dyn_cast_or_null(EvaluateExpression(Cond, PHIVal)); - if (CondVal->getValue() == ExitWhen) { + // Couldn't symbolically evaluate. + if (!CondVal) return UnknownValue; + + if (CondVal->getValue() == uint64_t(ExitWhen)) { ConstantEvolutionLoopExitValue[PN] = PHIVal; ++NumBruteForceTripCountsComputed; - return SCEVConstant::get(ConstantUInt::get(Type::UIntTy, IterationNum)); + return SCEVConstant::get(ConstantInt::get(Type::Int32Ty, IterationNum)); } - + // Compute the value of the PHI node for the next iteration. Constant *NextPHI = EvaluateExpression(BEValue, PHIVal); if (NextPHI == 0 || NextPHI == PHIVal) @@ -1861,7 +2033,7 @@ SCEVHandle ScalarEvolutionsImpl::getSCEVAtScope(SCEV *V, const Loop *L) { // FIXME: this should be turned into a virtual method on SCEV! if (isa(V)) return V; - + // If this instruction is evolves from a constant-evolving PHI, compute the // exit value from the loop without using SCEVs. if (SCEVUnknown *SU = dyn_cast(V)) { @@ -1879,15 +2051,15 @@ SCEVHandle ScalarEvolutionsImpl::getSCEVAtScope(SCEV *V, const Loop *L) { // this is a constant evolving PHI node, get the final value at // the specified iteration number. Constant *RV = getConstantEvolutionLoopExitValue(PN, - ICC->getValue()->getRawValue(), + ICC->getValue()->getValue(), LI); if (RV) return SCEVUnknown::get(RV); } } - // Okay, this is a some expression that we cannot symbolically evaluate + // Okay, this is an expression that we cannot symbolically evaluate // into a SCEV. Check to see if it's possible to symbolically evaluate - // the arguments into constants, and if see, try to constant propagate the + // the arguments into constants, and if so, try to constant propagate the // result. This is particularly useful for computing loop exit values. if (CanConstantFold(I)) { std::vector Operands; @@ -1899,11 +2071,14 @@ SCEVHandle ScalarEvolutionsImpl::getSCEVAtScope(SCEV *V, const Loop *L) { } else { SCEVHandle OpV = getSCEVAtScope(getSCEV(Op), L); if (SCEVConstant *SC = dyn_cast(OpV)) - Operands.push_back(ConstantExpr::getCast(SC->getValue(), - Op->getType())); + Operands.push_back(ConstantExpr::getIntegerCast(SC->getValue(), + Op->getType(), + false)); else if (SCEVUnknown *SU = dyn_cast(OpV)) { if (Constant *C = dyn_cast(SU->getValue())) - Operands.push_back(ConstantExpr::getCast(C, Op->getType())); + Operands.push_back(ConstantExpr::getIntegerCast(C, + Op->getType(), + false)); else return V; } else { @@ -1911,7 +2086,8 @@ SCEVHandle ScalarEvolutionsImpl::getSCEVAtScope(SCEV *V, const Loop *L) { } } } - return SCEVUnknown::get(ConstantFold(I, Operands)); + Constant *C =ConstantFoldInstOperands(I, &Operands[0], Operands.size()); + return SCEVUnknown::get(C); } } @@ -1946,14 +2122,14 @@ SCEVHandle ScalarEvolutionsImpl::getSCEVAtScope(SCEV *V, const Loop *L) { return Comm; } - if (SCEVUDivExpr *UDiv = dyn_cast(V)) { - SCEVHandle LHS = getSCEVAtScope(UDiv->getLHS(), L); + if (SCEVSDivExpr *Div = dyn_cast(V)) { + SCEVHandle LHS = getSCEVAtScope(Div->getLHS(), L); if (LHS == UnknownValue) return LHS; - SCEVHandle RHS = getSCEVAtScope(UDiv->getRHS(), L); + SCEVHandle RHS = getSCEVAtScope(Div->getRHS(), L); if (RHS == UnknownValue) return RHS; - if (LHS == UDiv->getLHS() && RHS == UDiv->getRHS()) - return UDiv; // must be loop invariant - return SCEVUDivExpr::get(LHS, RHS); + if (LHS == Div->getLHS() && RHS == Div->getRHS()) + return Div; // must be loop invariant + return SCEVSDivExpr::get(LHS, RHS); } // If this is a loop recurrence for a loop that does not contain L, then we @@ -1966,7 +2142,7 @@ SCEVHandle ScalarEvolutionsImpl::getSCEVAtScope(SCEV *V, const Loop *L) { if (IterationCount == UnknownValue) return UnknownValue; IterationCount = getTruncateOrZeroExtend(IterationCount, AddRec->getType()); - + // If the value is affine, simplify the expression evaluation to just // Start + Step*IterationCount. if (AddRec->isAffine()) @@ -1992,65 +2168,53 @@ SCEVHandle ScalarEvolutionsImpl::getSCEVAtScope(SCEV *V, const Loop *L) { static std::pair SolveQuadraticEquation(const SCEVAddRecExpr *AddRec) { assert(AddRec->getNumOperands() == 3 && "This is not a quadratic chrec!"); - SCEVConstant *L = dyn_cast(AddRec->getOperand(0)); - SCEVConstant *M = dyn_cast(AddRec->getOperand(1)); - SCEVConstant *N = dyn_cast(AddRec->getOperand(2)); - - // We currently can only solve this if the coefficients are constants. - if (!L || !M || !N) { - SCEV *CNC = new SCEVCouldNotCompute(); - return std::make_pair(CNC, CNC); - } + SCEVConstant *LC = dyn_cast(AddRec->getOperand(0)); + SCEVConstant *MC = dyn_cast(AddRec->getOperand(1)); + SCEVConstant *NC = dyn_cast(AddRec->getOperand(2)); - Constant *Two = ConstantInt::get(L->getValue()->getType(), 2); - - // Convert from chrec coefficients to polynomial coefficients AX^2+BX+C - Constant *C = L->getValue(); - // The B coefficient is M-N/2 - Constant *B = ConstantExpr::getSub(M->getValue(), - ConstantExpr::getDiv(N->getValue(), - Two)); - // The A coefficient is N/2 - Constant *A = ConstantExpr::getDiv(N->getValue(), Two); - - // Compute the B^2-4ac term. - Constant *SqrtTerm = - ConstantExpr::getMul(ConstantInt::get(C->getType(), 4), - ConstantExpr::getMul(A, C)); - SqrtTerm = ConstantExpr::getSub(ConstantExpr::getMul(B, B), SqrtTerm); - - // Compute floor(sqrt(B^2-4ac)) - ConstantUInt *SqrtVal = - cast(ConstantExpr::getCast(SqrtTerm, - SqrtTerm->getType()->getUnsignedVersion())); - uint64_t SqrtValV = SqrtVal->getValue(); - uint64_t SqrtValV2 = (uint64_t)sqrt((double)SqrtValV); - // The square root might not be precise for arbitrary 64-bit integer - // values. Do some sanity checks to ensure it's correct. - if (SqrtValV2*SqrtValV2 > SqrtValV || - (SqrtValV2+1)*(SqrtValV2+1) <= SqrtValV) { + // We currently can only solve this if the coefficients are constants. + if (!LC || !MC || !NC) { SCEV *CNC = new SCEVCouldNotCompute(); return std::make_pair(CNC, CNC); } - SqrtVal = ConstantUInt::get(Type::ULongTy, SqrtValV2); - SqrtTerm = ConstantExpr::getCast(SqrtVal, SqrtTerm->getType()); - - Constant *NegB = ConstantExpr::getNeg(B); - Constant *TwoA = ConstantExpr::getMul(A, Two); - - // The divisions must be performed as signed divisions. - const Type *SignedTy = NegB->getType()->getSignedVersion(); - NegB = ConstantExpr::getCast(NegB, SignedTy); - TwoA = ConstantExpr::getCast(TwoA, SignedTy); - SqrtTerm = ConstantExpr::getCast(SqrtTerm, SignedTy); - - Constant *Solution1 = - ConstantExpr::getDiv(ConstantExpr::getAdd(NegB, SqrtTerm), TwoA); - Constant *Solution2 = - ConstantExpr::getDiv(ConstantExpr::getSub(NegB, SqrtTerm), TwoA); - return std::make_pair(SCEVUnknown::get(Solution1), - SCEVUnknown::get(Solution2)); + uint32_t BitWidth = LC->getValue()->getValue().getBitWidth(); + const APInt &L = LC->getValue()->getValue(); + const APInt &M = MC->getValue()->getValue(); + const APInt &N = NC->getValue()->getValue(); + APInt Two(BitWidth, 2); + APInt Four(BitWidth, 4); + + { + using namespace APIntOps; + const APInt& C = L; + // Convert from chrec coefficients to polynomial coefficients AX^2+BX+C + // The B coefficient is M-N/2 + APInt B(M); + B -= sdiv(N,Two); + + // The A coefficient is N/2 + APInt A(N.sdiv(Two)); + + // Compute the B^2-4ac term. + APInt SqrtTerm(B); + SqrtTerm *= B; + SqrtTerm -= Four * (A * C); + + // Compute sqrt(B^2-4ac). This is guaranteed to be the nearest + // integer value or else APInt::sqrt() will assert. + APInt SqrtVal(SqrtTerm.sqrt()); + + // Compute the two solutions for the quadratic formula. + // The divisions must be performed as signed divisions. + APInt NegB(-B); + APInt TwoA( A << 1 ); + ConstantInt *Solution1 = ConstantInt::get((NegB + SqrtVal).sdiv(TwoA)); + ConstantInt *Solution2 = ConstantInt::get((NegB - SqrtVal).sdiv(TwoA)); + + return std::make_pair(SCEVConstant::get(Solution1), + SCEVConstant::get(Solution2)); + } // end APIntOps namespace } /// HowFarToZero - Return the number of times a backedge comparing the specified @@ -2059,7 +2223,7 @@ SCEVHandle ScalarEvolutionsImpl::HowFarToZero(SCEV *V, const Loop *L) { // If the value is a constant if (SCEVConstant *C = dyn_cast(V)) { // If the value is already zero, the branch will execute zero times. - if (C->getValue()->isNullValue()) return C; + if (C->getValue()->isZero()) return C; return UnknownValue; // Otherwise it will loop infinitely. } @@ -2092,9 +2256,9 @@ SCEVHandle ScalarEvolutionsImpl::HowFarToZero(SCEV *V, const Loop *L) { if (SCEVConstant *StartC = dyn_cast(Start)) { ConstantInt *StartCC = StartC->getValue(); Constant *StartNegC = ConstantExpr::getNeg(StartCC); - Constant *Rem = ConstantExpr::getRem(StartNegC, StepC->getValue()); + Constant *Rem = ConstantExpr::getSRem(StartNegC, StepC->getValue()); if (Rem->isNullValue()) { - Constant *Result =ConstantExpr::getDiv(StartNegC,StepC->getValue()); + Constant *Result =ConstantExpr::getSDiv(StartNegC,StepC->getValue()); return SCEVUnknown::get(Result); } } @@ -2107,28 +2271,27 @@ SCEVHandle ScalarEvolutionsImpl::HowFarToZero(SCEV *V, const Loop *L) { SCEVConstant *R2 = dyn_cast(Roots.second); if (R1) { #if 0 - std::cerr << "HFTZ: " << *V << " - sol#1: " << *R1 - << " sol#2: " << *R2 << "\n"; + cerr << "HFTZ: " << *V << " - sol#1: " << *R1 + << " sol#2: " << *R2 << "\n"; #endif // Pick the smallest positive root value. - assert(R1->getType()->isUnsigned()&&"Didn't canonicalize to unsigned?"); - if (ConstantBool *CB = - dyn_cast(ConstantExpr::getSetLT(R1->getValue(), - R2->getValue()))) { - if (CB != ConstantBool::True) + if (ConstantInt *CB = + dyn_cast(ConstantExpr::getICmp(ICmpInst::ICMP_ULT, + R1->getValue(), R2->getValue()))) { + if (CB->getZExtValue() == false) std::swap(R1, R2); // R1 is the minimum root now. - + // We can only use this value if the chrec ends up with an exact zero // value at this index. When solving for "X*X != 5", for example, we // should not accept a root of 2. SCEVHandle Val = AddRec->evaluateAtIteration(R1); if (SCEVConstant *EvalVal = dyn_cast(Val)) - if (EvalVal->getValue()->isNullValue()) + if (EvalVal->getValue()->isZero()) return R1; // We found a quadratic root! } } } - + return UnknownValue; } @@ -2139,22 +2302,124 @@ SCEVHandle ScalarEvolutionsImpl::HowFarToNonZero(SCEV *V, const Loop *L) { // Loops that look like: while (X == 0) are very strange indeed. We don't // handle them yet except for the trivial case. This could be expanded in the // future as needed. - + // If the value is a constant, check to see if it is known to be non-zero // already. If so, the backedge will execute zero times. if (SCEVConstant *C = dyn_cast(V)) { Constant *Zero = Constant::getNullValue(C->getValue()->getType()); - Constant *NonZero = ConstantExpr::getSetNE(C->getValue(), Zero); - if (NonZero == ConstantBool::True) + Constant *NonZero = + ConstantExpr::getICmp(ICmpInst::ICMP_NE, C->getValue(), Zero); + if (NonZero == ConstantInt::getTrue()) return getSCEV(Zero); return UnknownValue; // Otherwise it will loop infinitely. } - + // We could implement others, but I really doubt anyone writes loops like // this, and if they did, they would already be constant folded. return UnknownValue; } +/// HowManyLessThans - Return the number of times a backedge containing the +/// specified less-than comparison will execute. If not computable, return +/// UnknownValue. +SCEVHandle ScalarEvolutionsImpl:: +HowManyLessThans(SCEV *LHS, SCEV *RHS, const Loop *L, bool isSigned) { + // Only handle: "ADDREC < LoopInvariant". + if (!RHS->isLoopInvariant(L)) return UnknownValue; + + SCEVAddRecExpr *AddRec = dyn_cast(LHS); + if (!AddRec || AddRec->getLoop() != L) + return UnknownValue; + + if (AddRec->isAffine()) { + // FORNOW: We only support unit strides. + SCEVHandle One = SCEVUnknown::getIntegerSCEV(1, RHS->getType()); + if (AddRec->getOperand(1) != One) + return UnknownValue; + + // The number of iterations for "[n,+,1] < m", is m-n. However, we don't + // know that m is >= n on input to the loop. If it is, the condition return + // true zero times. What we really should return, for full generality, is + // SMAX(0, m-n). Since we cannot check this, we will instead check for a + // canonical loop form: most do-loops will have a check that dominates the + // loop, that only enters the loop if [n-1]= n. + + // Search for the check. + BasicBlock *Preheader = L->getLoopPreheader(); + BasicBlock *PreheaderDest = L->getHeader(); + if (Preheader == 0) return UnknownValue; + + BranchInst *LoopEntryPredicate = + dyn_cast(Preheader->getTerminator()); + if (!LoopEntryPredicate) return UnknownValue; + + // This might be a critical edge broken out. If the loop preheader ends in + // an unconditional branch to the loop, check to see if the preheader has a + // single predecessor, and if so, look for its terminator. + while (LoopEntryPredicate->isUnconditional()) { + PreheaderDest = Preheader; + Preheader = Preheader->getSinglePredecessor(); + if (!Preheader) return UnknownValue; // Multiple preds. + + LoopEntryPredicate = + dyn_cast(Preheader->getTerminator()); + if (!LoopEntryPredicate) return UnknownValue; + } + + // Now that we found a conditional branch that dominates the loop, check to + // see if it is the comparison we are looking for. + if (ICmpInst *ICI = dyn_cast(LoopEntryPredicate->getCondition())){ + Value *PreCondLHS = ICI->getOperand(0); + Value *PreCondRHS = ICI->getOperand(1); + ICmpInst::Predicate Cond; + if (LoopEntryPredicate->getSuccessor(0) == PreheaderDest) + Cond = ICI->getPredicate(); + else + Cond = ICI->getInversePredicate(); + + switch (Cond) { + case ICmpInst::ICMP_UGT: + if (isSigned) return UnknownValue; + std::swap(PreCondLHS, PreCondRHS); + Cond = ICmpInst::ICMP_ULT; + break; + case ICmpInst::ICMP_SGT: + if (!isSigned) return UnknownValue; + std::swap(PreCondLHS, PreCondRHS); + Cond = ICmpInst::ICMP_SLT; + break; + case ICmpInst::ICMP_ULT: + if (isSigned) return UnknownValue; + break; + case ICmpInst::ICMP_SLT: + if (!isSigned) return UnknownValue; + break; + default: + return UnknownValue; + } + + if (PreCondLHS->getType()->isInteger()) { + if (RHS != getSCEV(PreCondRHS)) + return UnknownValue; // Not a comparison against 'm'. + + if (SCEV::getMinusSCEV(AddRec->getOperand(0), One) + != getSCEV(PreCondLHS)) + return UnknownValue; // Not a comparison against 'n-1'. + } + else return UnknownValue; + + // cerr << "Computed Loop Trip Count as: " + // << // *SCEV::getMinusSCEV(RHS, AddRec->getOperand(0)) << "\n"; + return SCEV::getMinusSCEV(RHS, AddRec->getOperand(0)); + } + else + return UnknownValue; + } + + return UnknownValue; +} + /// getNumIterationsInRange - Return the number of iterations of this loop that /// produce values in the specified constant range. Another way of looking at /// this is that it returns the first iteration number where the value is not in @@ -2166,13 +2431,13 @@ SCEVHandle SCEVAddRecExpr::getNumIterationsInRange(ConstantRange Range) const { // If the start is a non-zero constant, shift the range to simplify things. if (SCEVConstant *SC = dyn_cast(getStart())) - if (!SC->getValue()->isNullValue()) { + if (!SC->getValue()->isZero()) { std::vector Operands(op_begin(), op_end()); Operands[0] = SCEVUnknown::getIntegerSCEV(0, SC->getType()); SCEVHandle Shifted = SCEVAddRecExpr::get(Operands, getLoop()); if (SCEVAddRecExpr *ShiftedAddRec = dyn_cast(Shifted)) return ShiftedAddRec->getNumIterationsInRange( - Range.subtract(SC->getValue())); + Range.subtract(SC->getValue()->getValue())); // This is strange and shouldn't happen. return new SCEVCouldNotCompute(); } @@ -2189,48 +2454,45 @@ SCEVHandle SCEVAddRecExpr::getNumIterationsInRange(ConstantRange Range) const { // First check to see if the range contains zero. If not, the first // iteration exits. - ConstantInt *Zero = ConstantInt::get(getType(), 0); - if (!Range.contains(Zero)) return SCEVConstant::get(Zero); - + if (!Range.contains(APInt(getBitWidth(),0))) + return SCEVConstant::get(ConstantInt::get(getType(),0)); + if (isAffine()) { // If this is an affine expression then we have this situation: // Solve {0,+,A} in Range === Ax in Range - // Since we know that zero is in the range, we know that the upper value of - // the range must be the first possible exit value. Also note that we - // already checked for a full range. - ConstantInt *Upper = cast(Range.getUpper()); - ConstantInt *A = cast(getOperand(1))->getValue(); - ConstantInt *One = ConstantInt::get(getType(), 1); - - // The exit value should be (Upper+A-1)/A. - Constant *ExitValue = Upper; - if (A != One) { - ExitValue = ConstantExpr::getSub(ConstantExpr::getAdd(Upper, A), One); - ExitValue = ConstantExpr::getDiv(ExitValue, A); - } - assert(isa(ExitValue) && - "Constant folding of integers not implemented?"); + // We know that zero is in the range. If A is positive then we know that + // the upper value of the range must be the first possible exit value. + // If A is negative then the lower of the range is the last possible loop + // value. Also note that we already checked for a full range. + APInt One(getBitWidth(),1); + APInt A = cast(getOperand(1))->getValue()->getValue(); + APInt End = A.sge(One) ? (Range.getUpper() - One) : Range.getLower(); + + // The exit value should be (End+A)/A. + APInt ExitVal = (End + A).udiv(A); + ConstantInt *ExitValue = ConstantInt::get(ExitVal); // Evaluate at the exit value. If we really did fall out of the valid // range, then we computed our trip count, otherwise wrap around or other // things must have happened. ConstantInt *Val = EvaluateConstantChrecAtConstant(this, ExitValue); - if (Range.contains(Val)) + if (Range.contains(Val->getValue())) return new SCEVCouldNotCompute(); // Something strange happened // Ensure that the previous value is in the range. This is a sanity check. - assert(Range.contains(EvaluateConstantChrecAtConstant(this, - ConstantExpr::getSub(ExitValue, One))) && + assert(Range.contains( + EvaluateConstantChrecAtConstant(this, + ConstantInt::get(ExitVal - One))->getValue()) && "Linear scev computation is off in a bad way!"); - return SCEVConstant::get(cast(ExitValue)); + return SCEVConstant::get(ExitValue); } else if (isQuadratic()) { // If this is a quadratic (3-term) AddRec {L,+,M,+,N}, find the roots of the // quadratic equation to solve it. To do this, we must frame our problem in // terms of figuring out when zero is crossed, instead of when // Range.getUpper() is crossed. std::vector NewOps(op_begin(), op_end()); - NewOps[0] = SCEV::getNegativeSCEV(SCEVUnknown::get(Range.getUpper())); + NewOps[0] = SCEV::getNegativeSCEV(SCEVConstant::get(Range.getUpper())); SCEVHandle NewAddRec = SCEVAddRecExpr::get(NewOps, getLoop()); // Next, solve the constructed addrec @@ -2240,37 +2502,32 @@ SCEVHandle SCEVAddRecExpr::getNumIterationsInRange(ConstantRange Range) const { SCEVConstant *R2 = dyn_cast(Roots.second); if (R1) { // Pick the smallest positive root value. - assert(R1->getType()->isUnsigned() && "Didn't canonicalize to unsigned?"); - if (ConstantBool *CB = - dyn_cast(ConstantExpr::getSetLT(R1->getValue(), - R2->getValue()))) { - if (CB != ConstantBool::True) + if (ConstantInt *CB = + dyn_cast(ConstantExpr::getICmp(ICmpInst::ICMP_ULT, + R1->getValue(), R2->getValue()))) { + if (CB->getZExtValue() == false) std::swap(R1, R2); // R1 is the minimum root now. - + // Make sure the root is not off by one. The returned iteration should // not be in the range, but the previous one should be. When solving // for "X*X < 5", for example, we should not return a root of 2. ConstantInt *R1Val = EvaluateConstantChrecAtConstant(this, R1->getValue()); - if (Range.contains(R1Val)) { + if (Range.contains(R1Val->getValue())) { // The next iteration must be out of the range... - Constant *NextVal = - ConstantExpr::getAdd(R1->getValue(), - ConstantInt::get(R1->getType(), 1)); - + ConstantInt *NextVal = ConstantInt::get(R1->getValue()->getValue()+1); + R1Val = EvaluateConstantChrecAtConstant(this, NextVal); - if (!Range.contains(R1Val)) - return SCEVUnknown::get(NextVal); + if (!Range.contains(R1Val->getValue())) + return SCEVConstant::get(NextVal); return new SCEVCouldNotCompute(); // Something strange happened } - + // If R1 was not in the range, then it is a good return value. Make // sure that R1-1 WAS in the range though, just in case. - Constant *NextVal = - ConstantExpr::getSub(R1->getValue(), - ConstantInt::get(R1->getType(), 1)); + ConstantInt *NextVal = ConstantInt::get(R1->getValue()->getValue()-1); R1Val = EvaluateConstantChrecAtConstant(this, NextVal); - if (Range.contains(R1Val)) + if (Range.contains(R1Val->getValue())) return R1; return new SCEVCouldNotCompute(); // Something strange happened } @@ -2283,7 +2540,6 @@ SCEVHandle SCEVAddRecExpr::getNumIterationsInRange(ConstantRange Range) const { // incredibly important, we will be able to simplify the exit test a lot, and // we are almost guaranteed to get a trip count in this case. ConstantInt *TestVal = ConstantInt::get(getType(), 0); - ConstantInt *One = ConstantInt::get(getType(), 1); ConstantInt *EndVal = TestVal; // Stop when we wrap around. do { ++NumBruteForceEvaluations; @@ -2292,13 +2548,13 @@ SCEVHandle SCEVAddRecExpr::getNumIterationsInRange(ConstantRange Range) const { return new SCEVCouldNotCompute(); // Check to see if we found the value! - if (!Range.contains(cast(Val)->getValue())) + if (!Range.contains(cast(Val)->getValue()->getValue())) return SCEVConstant::get(TestVal); // Increment to test the next index. - TestVal = cast(ConstantExpr::getAdd(TestVal, One)); + TestVal = ConstantInt::get(TestVal->getValue()+1); } while (TestVal != EndVal); - + return new SCEVCouldNotCompute(); } @@ -2327,6 +2583,20 @@ SCEVHandle ScalarEvolution::getSCEV(Value *V) const { return ((ScalarEvolutionsImpl*)Impl)->getSCEV(V); } +/// hasSCEV - Return true if the SCEV for this value has already been +/// computed. +bool ScalarEvolution::hasSCEV(Value *V) const { + return ((ScalarEvolutionsImpl*)Impl)->hasSCEV(V); +} + + +/// setSCEV - Insert the specified SCEV into the map of current SCEVs for +/// the specified value. +void ScalarEvolution::setSCEV(Value *V, const SCEVHandle &H) { + ((ScalarEvolutionsImpl*)Impl)->setSCEV(V, H); +} + + SCEVHandle ScalarEvolution::getIterationCount(const Loop *L) const { return ((ScalarEvolutionsImpl*)Impl)->getIterationCount(L); } @@ -2339,30 +2609,30 @@ SCEVHandle ScalarEvolution::getSCEVAtScope(Value *V, const Loop *L) const { return ((ScalarEvolutionsImpl*)Impl)->getSCEVAtScope(getSCEV(V), L); } -void ScalarEvolution::deleteInstructionFromRecords(Instruction *I) const { - return ((ScalarEvolutionsImpl*)Impl)->deleteInstructionFromRecords(I); +void ScalarEvolution::deleteValueFromRecords(Value *V) const { + return ((ScalarEvolutionsImpl*)Impl)->deleteValueFromRecords(V); } -static void PrintLoopInfo(std::ostream &OS, const ScalarEvolution *SE, +static void PrintLoopInfo(std::ostream &OS, const ScalarEvolution *SE, const Loop *L) { // Print all inner loops first for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I) PrintLoopInfo(OS, SE, *I); - - std::cerr << "Loop " << L->getHeader()->getName() << ": "; - std::vector ExitBlocks; + cerr << "Loop " << L->getHeader()->getName() << ": "; + + SmallVector ExitBlocks; L->getExitBlocks(ExitBlocks); if (ExitBlocks.size() != 1) - std::cerr << " "; + cerr << " "; if (SE->hasLoopInvariantIterationCount(L)) { - std::cerr << *SE->getIterationCount(L) << " iterations! "; + cerr << *SE->getIterationCount(L) << " iterations! "; } else { - std::cerr << "Unpredictable iteration count. "; + cerr << "Unpredictable iteration count. "; } - std::cerr << "\n"; + cerr << "\n"; } void ScalarEvolution::print(std::ostream &OS, const Module* ) const { @@ -2377,8 +2647,8 @@ void ScalarEvolution::print(std::ostream &OS, const Module* ) const { SCEVHandle SV = getSCEV(&*I); SV->print(OS); OS << "\t\t"; - - if ((*I).getType()->isIntegral()) { + + if ((*I).getType()->isInteger()) { ConstantRange Bounds = SV->getValueRange(); if (!Bounds.isFullSet()) OS << "Bounds: " << Bounds << " ";