//
// 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 is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
STATISTIC(NumBruteForceTripCountsComputed,
"Number of loops with trip counts computed by force");
-cl::opt<unsigned>
+static cl::opt<unsigned>
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 {
- RegisterPass<ScalarEvolution>
- R("scalar-evolution", "Scalar Evolution Analysis");
-}
+static RegisterPass<ScalarEvolution>
+R("scalar-evolution", "Scalar Evolution Analysis", false, true);
char ScalarEvolution::ID = 0;
//===----------------------------------------------------------------------===//
SCEVHandle SCEVCouldNotCompute::
replaceSymbolicValuesWithConcrete(const SCEVHandle &Sym,
- const SCEVHandle &Conc) const {
+ const SCEVHandle &Conc,
+ ScalarEvolution &SE) const {
return this;
}
SCEVConstants->erase(V);
}
-SCEVHandle SCEVConstant::get(ConstantInt *V) {
+SCEVHandle ScalarEvolution::getConstant(ConstantInt *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));
+SCEVHandle ScalarEvolution::getConstant(const APInt& Val) {
+ return getConstant(ConstantInt::get(Val));
}
ConstantRange SCEVConstant::getValueRange() const {
SCEVHandle SCEVCommutativeExpr::
replaceSymbolicValuesWithConcrete(const SCEVHandle &Sym,
- const SCEVHandle &Conc) const {
+ const SCEVHandle &Conc,
+ ScalarEvolution &SE) const {
for (unsigned i = 0, e = getNumOperands(); i != e; ++i) {
- SCEVHandle H = getOperand(i)->replaceSymbolicValuesWithConcrete(Sym, Conc);
+ SCEVHandle H =
+ getOperand(i)->replaceSymbolicValuesWithConcrete(Sym, Conc, SE);
if (H != getOperand(i)) {
std::vector<SCEVHandle> NewOps;
NewOps.reserve(getNumOperands());
NewOps.push_back(H);
for (++i; i != e; ++i)
NewOps.push_back(getOperand(i)->
- replaceSymbolicValuesWithConcrete(Sym, Conc));
+ replaceSymbolicValuesWithConcrete(Sym, Conc, SE));
if (isa<SCEVAddExpr>(this))
- return SCEVAddExpr::get(NewOps);
+ return SE.getAddExpr(NewOps);
else if (isa<SCEVMulExpr>(this))
- return SCEVMulExpr::get(NewOps);
+ return SE.getMulExpr(NewOps);
+ else if (isa<SCEVSMaxExpr>(this))
+ return SE.getSMaxExpr(NewOps);
+ else if (isa<SCEVUMaxExpr>(this))
+ return SE.getUMaxExpr(NewOps);
else
assert(0 && "Unknown commutative expr!");
}
}
-// SCEVSDivs - Only allow the creation of one SCEVSDivExpr for any particular
+// SCEVUDivs - Only allow the creation of one SCEVUDivExpr for any particular
// input. Don't use a SCEVHandle here, or else the object will never be
// deleted!
static ManagedStatic<std::map<std::pair<SCEV*, SCEV*>,
- SCEVSDivExpr*> > SCEVSDivs;
+ SCEVUDivExpr*> > SCEVUDivs;
-SCEVSDivExpr::~SCEVSDivExpr() {
- SCEVSDivs->erase(std::make_pair(LHS, RHS));
+SCEVUDivExpr::~SCEVUDivExpr() {
+ SCEVUDivs->erase(std::make_pair(LHS, RHS));
}
-void SCEVSDivExpr::print(std::ostream &OS) const {
- OS << "(" << *LHS << " /s " << *RHS << ")";
+void SCEVUDivExpr::print(std::ostream &OS) const {
+ OS << "(" << *LHS << " /u " << *RHS << ")";
}
-const Type *SCEVSDivExpr::getType() const {
+const Type *SCEVUDivExpr::getType() const {
return LHS->getType();
}
SCEVHandle SCEVAddRecExpr::
replaceSymbolicValuesWithConcrete(const SCEVHandle &Sym,
- const SCEVHandle &Conc) const {
+ const SCEVHandle &Conc,
+ ScalarEvolution &SE) const {
for (unsigned i = 0, e = getNumOperands(); i != e; ++i) {
- SCEVHandle H = getOperand(i)->replaceSymbolicValuesWithConcrete(Sym, Conc);
+ SCEVHandle H =
+ getOperand(i)->replaceSymbolicValuesWithConcrete(Sym, Conc, SE);
if (H != getOperand(i)) {
std::vector<SCEVHandle> NewOps;
NewOps.reserve(getNumOperands());
NewOps.push_back(H);
for (++i; i != e; ++i)
NewOps.push_back(getOperand(i)->
- replaceSymbolicValuesWithConcrete(Sym, Conc));
+ replaceSymbolicValuesWithConcrete(Sym, Conc, SE));
- return get(NewOps, L);
+ return SE.getAddRecExpr(NewOps, L);
}
}
return this;
/// than the complexity of the RHS. This comparator is used to canonicalize
/// expressions.
struct VISIBILITY_HIDDEN SCEVComplexityCompare {
- bool operator()(SCEV *LHS, SCEV *RHS) {
+ bool operator()(const SCEV *LHS, const SCEV *RHS) const {
return LHS->getSCEVType() < RHS->getSCEVType();
}
};
if (Ops.size() == 2) {
// This is the common case, which also happens to be trivially simple.
// Special case it.
- if (Ops[0]->getSCEVType() > Ops[1]->getSCEVType())
+ if (SCEVComplexityCompare()(Ops[1], Ops[0]))
std::swap(Ops[0], Ops[1]);
return;
}
/// getIntegerSCEV - Given an integer or FP type, create a constant for the
/// specified signed integer value and return a SCEV for the constant.
-SCEVHandle SCEVUnknown::getIntegerSCEV(int Val, const Type *Ty) {
+SCEVHandle ScalarEvolution::getIntegerSCEV(int Val, const Type *Ty) {
Constant *C;
if (Val == 0)
C = Constant::getNullValue(Ty);
else if (Ty->isFloatingPoint())
- C = ConstantFP::get(Ty, Val);
+ C = ConstantFP::get(APFloat(Ty==Type::FloatTy ? APFloat::IEEEsingle :
+ APFloat::IEEEdouble, Val));
else
C = ConstantInt::get(Ty, Val);
- return SCEVUnknown::get(C);
+ return getUnknown(C);
}
/// getTruncateOrZeroExtend - Return a SCEV corresponding to a conversion of the
/// input value to the specified type. If the type must be extended, it is zero
/// extended.
-static SCEVHandle getTruncateOrZeroExtend(const SCEVHandle &V, const Type *Ty) {
+static SCEVHandle getTruncateOrZeroExtend(const SCEVHandle &V, const Type *Ty,
+ ScalarEvolution &SE) {
const Type *SrcTy = V->getType();
assert(SrcTy->isInteger() && Ty->isInteger() &&
"Cannot truncate or zero extend with non-integer arguments!");
if (SrcTy->getPrimitiveSizeInBits() == Ty->getPrimitiveSizeInBits())
return V; // No conversion
if (SrcTy->getPrimitiveSizeInBits() > Ty->getPrimitiveSizeInBits())
- return SCEVTruncateExpr::get(V, Ty);
- return SCEVZeroExtendExpr::get(V, Ty);
+ return SE.getTruncateExpr(V, Ty);
+ return SE.getZeroExtendExpr(V, Ty);
}
/// getNegativeSCEV - Return a SCEV corresponding to -V = -1*V
///
-SCEVHandle SCEV::getNegativeSCEV(const SCEVHandle &V) {
+SCEVHandle ScalarEvolution::getNegativeSCEV(const SCEVHandle &V) {
+ if (SCEVConstant *VC = dyn_cast<SCEVConstant>(V))
+ return getUnknown(ConstantExpr::getNeg(VC->getValue()));
+
+ return getMulExpr(V, getConstant(ConstantInt::getAllOnesValue(V->getType())));
+}
+
+/// getNotSCEV - Return a SCEV corresponding to ~V = -1-V
+SCEVHandle ScalarEvolution::getNotSCEV(const SCEVHandle &V) {
if (SCEVConstant *VC = dyn_cast<SCEVConstant>(V))
- return SCEVUnknown::get(ConstantExpr::getNeg(VC->getValue()));
+ return getUnknown(ConstantExpr::getNot(VC->getValue()));
- return SCEVMulExpr::get(V, SCEVUnknown::getIntegerSCEV(-1, V->getType()));
+ SCEVHandle AllOnes = getConstant(ConstantInt::getAllOnesValue(V->getType()));
+ return getMinusSCEV(AllOnes, V);
}
/// getMinusSCEV - Return a SCEV corresponding to LHS - RHS.
///
-SCEVHandle SCEV::getMinusSCEV(const SCEVHandle &LHS, const SCEVHandle &RHS) {
+SCEVHandle ScalarEvolution::getMinusSCEV(const SCEVHandle &LHS,
+ const SCEVHandle &RHS) {
// X - Y --> X + -Y
- return SCEVAddExpr::get(LHS, SCEV::getNegativeSCEV(RHS));
+ return getAddExpr(LHS, getNegativeSCEV(RHS));
}
-/// PartialFact - Compute V!/(V-NumSteps)!
-static SCEVHandle PartialFact(SCEVHandle V, unsigned NumSteps) {
+/// BinomialCoefficient - Compute BC(It, K). The result is of the same type as
+/// It. Assume, K > 0.
+static SCEVHandle BinomialCoefficient(SCEVHandle It, unsigned K,
+ ScalarEvolution &SE) {
+ // We are using the following formula for BC(It, K):
+ //
+ // BC(It, K) = (It * (It - 1) * ... * (It - K + 1)) / K!
+ //
+ // Suppose, W is the bitwidth of It (and of the return value as well). We
+ // must be prepared for overflow. Hence, we must assure that the result of
+ // our computation is equal to the accurate one modulo 2^W. Unfortunately,
+ // division isn't safe in modular arithmetic. This means we must perform the
+ // whole computation accurately and then truncate the result to W bits.
+ //
+ // The dividend of the formula is a multiplication of K integers of bitwidth
+ // W. K*W bits suffice to compute it accurately.
+ //
+ // FIXME: We assume the divisor can be accurately computed using 16-bit
+ // unsigned integer type. It is true up to K = 8 (AddRecs of length 9). In
+ // future we may use APInt to use the minimum number of bits necessary to
+ // compute it accurately.
+ //
+ // It is safe to use unsigned division here: the dividend is nonnegative and
+ // the divisor is positive.
+
+ // Handle the simplest case efficiently.
+ if (K == 1)
+ return It;
+
+ assert(K < 9 && "We cannot handle such long AddRecs yet.");
+
+ // FIXME: A temporary hack to remove in future. Arbitrary precision integers
+ // aren't supported by the code generator yet. For the dividend, the bitwidth
+ // we use is the smallest power of 2 greater or equal to K*W and less or equal
+ // to 64. Note that setting the upper bound for bitwidth may still lead to
+ // miscompilation in some cases.
+ unsigned DividendBits = 1U << Log2_32_Ceil(K * It->getBitWidth());
+ if (DividendBits > 64)
+ DividendBits = 64;
+#if 0 // Waiting for the APInt support in the code generator...
+ unsigned DividendBits = K * It->getBitWidth();
+#endif
+
+ const IntegerType *DividendTy = IntegerType::get(DividendBits);
+ const SCEVHandle ExIt = SE.getZeroExtendExpr(It, DividendTy);
+
+ // The final number of bits we need to perform the division is the maximum of
+ // dividend and divisor bitwidths.
+ const IntegerType *DivisionTy =
+ IntegerType::get(std::max(DividendBits, 16U));
+
+ // Compute K! We know K >= 2 here.
+ unsigned F = 2;
+ for (unsigned i = 3; i <= K; ++i)
+ F *= i;
+ APInt Divisor(DivisionTy->getBitWidth(), F);
+
// Handle this case efficiently, it is common to have constant iteration
// counts while computing loop exit values.
- if (SCEVConstant *SC = dyn_cast<SCEVConstant>(V)) {
- const APInt& Val = SC->getValue()->getValue();
- APInt Result(Val.getBitWidth(), 1);
- for (; NumSteps; --NumSteps)
- Result *= Val-(NumSteps-1);
- return SCEVConstant::get(Result);
+ if (SCEVConstant *SC = dyn_cast<SCEVConstant>(ExIt)) {
+ const APInt& N = SC->getValue()->getValue();
+ APInt Dividend(N.getBitWidth(), 1);
+ for (; K; --K)
+ Dividend *= N-(K-1);
+ if (DividendTy != DivisionTy)
+ Dividend = Dividend.zext(DivisionTy->getBitWidth());
+ return SE.getConstant(Dividend.udiv(Divisor).trunc(It->getBitWidth()));
}
-
- 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,
- SCEVUnknown::getIntegerSCEV(i, Ty)));
- return Result;
+
+ SCEVHandle Dividend = ExIt;
+ for (unsigned i = 1; i != K; ++i)
+ Dividend =
+ SE.getMulExpr(Dividend,
+ SE.getMinusSCEV(ExIt, SE.getIntegerSCEV(i, DividendTy)));
+ if (DividendTy != DivisionTy)
+ Dividend = SE.getZeroExtendExpr(Dividend, DivisionTy);
+ return
+ SE.getTruncateExpr(SE.getUDivExpr(Dividend, SE.getConstant(Divisor)),
+ It->getType());
}
-
/// evaluateAtIteration - Return the value of this chain of recurrences at
/// the specified iteration number. We can evaluate this recurrence by
/// multiplying each element in the chain by the binomial coefficient
/// corresponding to it. In other words, we can evaluate {A,+,B,+,C,+,D} as:
///
-/// A*choose(It, 0) + B*choose(It, 1) + C*choose(It, 2) + D*choose(It, 3)
+/// A*BC(It, 0) + B*BC(It, 1) + C*BC(It, 2) + D*BC(It, 3)
///
-/// FIXME/VERIFY: I don't trust that this is correct in the face of overflow.
-/// Is the binomial equation safe using modular arithmetic??
+/// where BC(It, k) stands for binomial coefficient.
///
-SCEVHandle SCEVAddRecExpr::evaluateAtIteration(SCEVHandle It) const {
+SCEVHandle SCEVAddRecExpr::evaluateAtIteration(SCEVHandle It,
+ ScalarEvolution &SE) const {
SCEVHandle Result = getStart();
- int Divisor = 1;
- const Type *Ty = It->getType();
for (unsigned i = 1, e = getNumOperands(); i != e; ++i) {
- SCEVHandle BC = PartialFact(It, i);
- Divisor *= i;
- SCEVHandle Val = SCEVSDivExpr::get(SCEVMulExpr::get(BC, getOperand(i)),
- SCEVUnknown::getIntegerSCEV(Divisor,Ty));
- Result = SCEVAddExpr::get(Result, Val);
+ // The computation is correct in the face of overflow provided that the
+ // multiplication is performed _after_ the evaluation of the binomial
+ // coefficient.
+ SCEVHandle Val = SE.getMulExpr(getOperand(i),
+ BinomialCoefficient(It, i, SE));
+ Result = SE.getAddExpr(Result, Val);
}
return Result;
}
-
//===----------------------------------------------------------------------===//
// SCEV Expression folder implementations
//===----------------------------------------------------------------------===//
-SCEVHandle SCEVTruncateExpr::get(const SCEVHandle &Op, const Type *Ty) {
+SCEVHandle ScalarEvolution::getTruncateExpr(const SCEVHandle &Op, const Type *Ty) {
if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
- return SCEVUnknown::get(
+ return getUnknown(
ConstantExpr::getTrunc(SC->getValue(), Ty));
// If the input value is a chrec scev made out of constants, truncate
for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i)
// FIXME: This should allow truncation of other expression types!
if (isa<SCEVConstant>(AddRec->getOperand(i)))
- Operands.push_back(get(AddRec->getOperand(i), Ty));
+ Operands.push_back(getTruncateExpr(AddRec->getOperand(i), Ty));
else
break;
if (Operands.size() == AddRec->getNumOperands())
- return SCEVAddRecExpr::get(Operands, AddRec->getLoop());
+ return getAddRecExpr(Operands, AddRec->getLoop());
}
SCEVTruncateExpr *&Result = (*SCEVTruncates)[std::make_pair(Op, Ty)];
return Result;
}
-SCEVHandle SCEVZeroExtendExpr::get(const SCEVHandle &Op, const Type *Ty) {
+SCEVHandle ScalarEvolution::getZeroExtendExpr(const SCEVHandle &Op, const Type *Ty) {
if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
- return SCEVUnknown::get(
+ return getUnknown(
ConstantExpr::getZExt(SC->getValue(), Ty));
// FIXME: If the input value is a chrec scev, and we can prove that the value
return Result;
}
-SCEVHandle SCEVSignExtendExpr::get(const SCEVHandle &Op, const Type *Ty) {
+SCEVHandle ScalarEvolution::getSignExtendExpr(const SCEVHandle &Op, const Type *Ty) {
if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
- return SCEVUnknown::get(
+ return getUnknown(
ConstantExpr::getSExt(SC->getValue(), Ty));
// FIXME: If the input value is a chrec scev, and we can prove that the value
}
// get - Get a canonical add expression, or something simpler if possible.
-SCEVHandle SCEVAddExpr::get(std::vector<SCEVHandle> &Ops) {
+SCEVHandle ScalarEvolution::getAddExpr(std::vector<SCEVHandle> &Ops) {
assert(!Ops.empty() && "Cannot get empty add!");
if (Ops.size() == 1) return Ops[0];
assert(Idx < Ops.size());
while (SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
// We found two constants, fold them together!
- Constant *Fold = ConstantInt::get(LHSC->getValue()->getValue() +
- RHSC->getValue()->getValue());
- if (ConstantInt *CI = dyn_cast<ConstantInt>(Fold)) {
- Ops[0] = SCEVConstant::get(CI);
- Ops.erase(Ops.begin()+1); // Erase the folded element
- if (Ops.size() == 1) return Ops[0];
- LHSC = cast<SCEVConstant>(Ops[0]);
- } else {
- // If we couldn't fold the expression, move to the next constant. Note
- // that this is impossible to happen in practice because we always
- // constant fold constant ints to constant ints.
- ++Idx;
- }
+ ConstantInt *Fold = ConstantInt::get(LHSC->getValue()->getValue() +
+ RHSC->getValue()->getValue());
+ Ops[0] = getConstant(Fold);
+ Ops.erase(Ops.begin()+1); // Erase the folded element
+ if (Ops.size() == 1) return Ops[0];
+ LHSC = cast<SCEVConstant>(Ops[0]);
}
// If we are left with a constant zero being added, strip it off.
if (Ops[i] == Ops[i+1]) { // X + Y + Y --> X + Y*2
// Found a match, merge the two values into a multiply, and add any
// remaining values to the result.
- SCEVHandle Two = SCEVUnknown::getIntegerSCEV(2, Ty);
- SCEVHandle Mul = SCEVMulExpr::get(Ops[i], Two);
+ SCEVHandle Two = getIntegerSCEV(2, Ty);
+ SCEVHandle Mul = getMulExpr(Ops[i], Two);
if (Ops.size() == 2)
return Mul;
Ops.erase(Ops.begin()+i, Ops.begin()+i+2);
Ops.push_back(Mul);
- return SCEVAddExpr::get(Ops);
+ return getAddExpr(Ops);
}
// Now we know the first non-constant operand. Skip past any cast SCEVs.
// and they are not necessarily sorted. Recurse to resort and resimplify
// any operands we just aquired.
if (DeletedAdd)
- return get(Ops);
+ return getAddExpr(Ops);
}
// Skip over the add expression until we get to a multiply.
// Y*Z term.
std::vector<SCEVHandle> MulOps(Mul->op_begin(), Mul->op_end());
MulOps.erase(MulOps.begin()+MulOp);
- InnerMul = SCEVMulExpr::get(MulOps);
+ InnerMul = getMulExpr(MulOps);
}
- SCEVHandle One = SCEVUnknown::getIntegerSCEV(1, Ty);
- SCEVHandle AddOne = SCEVAddExpr::get(InnerMul, One);
- SCEVHandle OuterMul = SCEVMulExpr::get(AddOne, Ops[AddOp]);
+ SCEVHandle One = getIntegerSCEV(1, Ty);
+ SCEVHandle AddOne = getAddExpr(InnerMul, One);
+ SCEVHandle OuterMul = getMulExpr(AddOne, Ops[AddOp]);
if (Ops.size() == 2) return OuterMul;
if (AddOp < Idx) {
Ops.erase(Ops.begin()+AddOp);
Ops.erase(Ops.begin()+AddOp-1);
}
Ops.push_back(OuterMul);
- return SCEVAddExpr::get(Ops);
+ return getAddExpr(Ops);
}
// Check this multiply against other multiplies being added together.
if (Mul->getNumOperands() != 2) {
std::vector<SCEVHandle> MulOps(Mul->op_begin(), Mul->op_end());
MulOps.erase(MulOps.begin()+MulOp);
- InnerMul1 = SCEVMulExpr::get(MulOps);
+ InnerMul1 = getMulExpr(MulOps);
}
SCEVHandle InnerMul2 = OtherMul->getOperand(OMulOp == 0);
if (OtherMul->getNumOperands() != 2) {
std::vector<SCEVHandle> MulOps(OtherMul->op_begin(),
OtherMul->op_end());
MulOps.erase(MulOps.begin()+OMulOp);
- InnerMul2 = SCEVMulExpr::get(MulOps);
+ InnerMul2 = getMulExpr(MulOps);
}
- SCEVHandle InnerMulSum = SCEVAddExpr::get(InnerMul1,InnerMul2);
- SCEVHandle OuterMul = SCEVMulExpr::get(MulOpSCEV, InnerMulSum);
+ SCEVHandle InnerMulSum = getAddExpr(InnerMul1,InnerMul2);
+ SCEVHandle OuterMul = getMulExpr(MulOpSCEV, InnerMulSum);
if (Ops.size() == 2) return OuterMul;
Ops.erase(Ops.begin()+Idx);
Ops.erase(Ops.begin()+OtherMulIdx-1);
Ops.push_back(OuterMul);
- return SCEVAddExpr::get(Ops);
+ return getAddExpr(Ops);
}
}
}
LIOps.push_back(AddRec->getStart());
std::vector<SCEVHandle> AddRecOps(AddRec->op_begin(), AddRec->op_end());
- AddRecOps[0] = SCEVAddExpr::get(LIOps);
+ AddRecOps[0] = getAddExpr(LIOps);
- SCEVHandle NewRec = SCEVAddRecExpr::get(AddRecOps, AddRec->getLoop());
+ SCEVHandle NewRec = getAddRecExpr(AddRecOps, AddRec->getLoop());
// If all of the other operands were loop invariant, we are done.
if (Ops.size() == 1) return NewRec;
Ops[i] = NewRec;
break;
}
- return SCEVAddExpr::get(Ops);
+ return getAddExpr(Ops);
}
// Okay, if there weren't any loop invariants to be folded, check to see if
OtherAddRec->op_end());
break;
}
- NewOps[i] = SCEVAddExpr::get(NewOps[i], OtherAddRec->getOperand(i));
+ NewOps[i] = getAddExpr(NewOps[i], OtherAddRec->getOperand(i));
}
- SCEVHandle NewAddRec = SCEVAddRecExpr::get(NewOps, AddRec->getLoop());
+ SCEVHandle NewAddRec = getAddRecExpr(NewOps, AddRec->getLoop());
if (Ops.size() == 2) return NewAddRec;
Ops.erase(Ops.begin()+Idx);
Ops.erase(Ops.begin()+OtherIdx-1);
Ops.push_back(NewAddRec);
- return SCEVAddExpr::get(Ops);
+ return getAddExpr(Ops);
}
}
}
-SCEVHandle SCEVMulExpr::get(std::vector<SCEVHandle> &Ops) {
+SCEVHandle ScalarEvolution::getMulExpr(std::vector<SCEVHandle> &Ops) {
assert(!Ops.empty() && "Cannot get empty mul!");
// Sort by complexity, this groups all similar expression types together.
if (SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[1]))
if (Add->getNumOperands() == 2 &&
isa<SCEVConstant>(Add->getOperand(0)))
- return SCEVAddExpr::get(SCEVMulExpr::get(LHSC, Add->getOperand(0)),
- SCEVMulExpr::get(LHSC, Add->getOperand(1)));
+ return getAddExpr(getMulExpr(LHSC, Add->getOperand(0)),
+ getMulExpr(LHSC, Add->getOperand(1)));
++Idx;
while (SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
// We found two constants, fold them together!
- Constant *Fold = ConstantInt::get(LHSC->getValue()->getValue() *
- RHSC->getValue()->getValue());
- if (ConstantInt *CI = dyn_cast<ConstantInt>(Fold)) {
- Ops[0] = SCEVConstant::get(CI);
- Ops.erase(Ops.begin()+1); // Erase the folded element
- if (Ops.size() == 1) return Ops[0];
- LHSC = cast<SCEVConstant>(Ops[0]);
- } else {
- // If we couldn't fold the expression, move to the next constant. Note
- // that this is impossible to happen in practice because we always
- // constant fold constant ints to constant ints.
- ++Idx;
- }
+ ConstantInt *Fold = ConstantInt::get(LHSC->getValue()->getValue() *
+ RHSC->getValue()->getValue());
+ Ops[0] = getConstant(Fold);
+ Ops.erase(Ops.begin()+1); // Erase the folded element
+ if (Ops.size() == 1) return Ops[0];
+ LHSC = cast<SCEVConstant>(Ops[0]);
}
// If we are left with a constant one being multiplied, strip it off.
// and they are not necessarily sorted. Recurse to resort and resimplify
// any operands we just aquired.
if (DeletedMul)
- return get(Ops);
+ return getMulExpr(Ops);
}
// If there are any add recurrences in the operands list, see if any other
if (LIOps.size() == 1) {
SCEV *Scale = LIOps[0];
for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i)
- NewOps.push_back(SCEVMulExpr::get(Scale, AddRec->getOperand(i)));
+ NewOps.push_back(getMulExpr(Scale, AddRec->getOperand(i)));
} else {
for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i) {
std::vector<SCEVHandle> MulOps(LIOps);
MulOps.push_back(AddRec->getOperand(i));
- NewOps.push_back(SCEVMulExpr::get(MulOps));
+ NewOps.push_back(getMulExpr(MulOps));
}
}
- SCEVHandle NewRec = SCEVAddRecExpr::get(NewOps, AddRec->getLoop());
+ SCEVHandle NewRec = getAddRecExpr(NewOps, AddRec->getLoop());
// If all of the other operands were loop invariant, we are done.
if (Ops.size() == 1) return NewRec;
Ops[i] = NewRec;
break;
}
- return SCEVMulExpr::get(Ops);
+ return getMulExpr(Ops);
}
// Okay, if there weren't any loop invariants to be folded, check to see if
if (AddRec->getLoop() == OtherAddRec->getLoop()) {
// F * G --> {A,+,B} * {C,+,D} --> {A*C,+,F*D + G*B + B*D}
SCEVAddRecExpr *F = AddRec, *G = OtherAddRec;
- SCEVHandle NewStart = SCEVMulExpr::get(F->getStart(),
+ SCEVHandle NewStart = getMulExpr(F->getStart(),
G->getStart());
- SCEVHandle B = F->getStepRecurrence();
- SCEVHandle D = G->getStepRecurrence();
- SCEVHandle NewStep = SCEVAddExpr::get(SCEVMulExpr::get(F, D),
- SCEVMulExpr::get(G, B),
- SCEVMulExpr::get(B, D));
- SCEVHandle NewAddRec = SCEVAddRecExpr::get(NewStart, NewStep,
- F->getLoop());
+ SCEVHandle B = F->getStepRecurrence(*this);
+ SCEVHandle D = G->getStepRecurrence(*this);
+ SCEVHandle NewStep = getAddExpr(getMulExpr(F, D),
+ getMulExpr(G, B),
+ getMulExpr(B, D));
+ SCEVHandle NewAddRec = getAddRecExpr(NewStart, NewStep,
+ F->getLoop());
if (Ops.size() == 2) return NewAddRec;
Ops.erase(Ops.begin()+Idx);
Ops.erase(Ops.begin()+OtherIdx-1);
Ops.push_back(NewAddRec);
- return SCEVMulExpr::get(Ops);
+ return getMulExpr(Ops);
}
}
return Result;
}
-SCEVHandle SCEVSDivExpr::get(const SCEVHandle &LHS, const SCEVHandle &RHS) {
+SCEVHandle ScalarEvolution::getUDivExpr(const SCEVHandle &LHS, const SCEVHandle &RHS) {
if (SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS)) {
if (RHSC->getValue()->equalsInt(1))
- return LHS; // X sdiv 1 --> x
- if (RHSC->getValue()->isAllOnesValue())
- return SCEV::getNegativeSCEV(LHS); // X sdiv -1 --> -x
+ return LHS; // X udiv 1 --> x
if (SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS)) {
Constant *LHSCV = LHSC->getValue();
Constant *RHSCV = RHSC->getValue();
- return SCEVUnknown::get(ConstantExpr::getSDiv(LHSCV, RHSCV));
+ return getUnknown(ConstantExpr::getUDiv(LHSCV, RHSCV));
}
}
// FIXME: implement folding of (X*4)/4 when we know X*4 doesn't overflow.
- SCEVSDivExpr *&Result = (*SCEVSDivs)[std::make_pair(LHS, RHS)];
- if (Result == 0) Result = new SCEVSDivExpr(LHS, RHS);
+ SCEVUDivExpr *&Result = (*SCEVUDivs)[std::make_pair(LHS, RHS)];
+ if (Result == 0) Result = new SCEVUDivExpr(LHS, RHS);
return Result;
}
/// SCEVAddRecExpr::get - Get a add recurrence expression for the
/// specified loop. Simplify the expression as much as possible.
-SCEVHandle SCEVAddRecExpr::get(const SCEVHandle &Start,
+SCEVHandle ScalarEvolution::getAddRecExpr(const SCEVHandle &Start,
const SCEVHandle &Step, const Loop *L) {
std::vector<SCEVHandle> Operands;
Operands.push_back(Start);
if (StepChrec->getLoop() == L) {
Operands.insert(Operands.end(), StepChrec->op_begin(),
StepChrec->op_end());
- return get(Operands, L);
+ return getAddRecExpr(Operands, L);
}
Operands.push_back(Step);
- return get(Operands, L);
+ return getAddRecExpr(Operands, L);
}
/// SCEVAddRecExpr::get - Get a add recurrence expression for the
/// specified loop. Simplify the expression as much as possible.
-SCEVHandle SCEVAddRecExpr::get(std::vector<SCEVHandle> &Operands,
+SCEVHandle ScalarEvolution::getAddRecExpr(std::vector<SCEVHandle> &Operands,
const Loop *L) {
if (Operands.size() == 1) return Operands[0];
if (SCEVConstant *StepC = dyn_cast<SCEVConstant>(Operands.back()))
if (StepC->getValue()->isZero()) {
Operands.pop_back();
- return get(Operands, L); // { X,+,0 } --> X
+ return getAddRecExpr(Operands, L); // { X,+,0 } --> X
}
SCEVAddRecExpr *&Result =
return Result;
}
-SCEVHandle SCEVUnknown::get(Value *V) {
+SCEVHandle ScalarEvolution::getSMaxExpr(const SCEVHandle &LHS,
+ const SCEVHandle &RHS) {
+ std::vector<SCEVHandle> Ops;
+ Ops.push_back(LHS);
+ Ops.push_back(RHS);
+ return getSMaxExpr(Ops);
+}
+
+SCEVHandle ScalarEvolution::getSMaxExpr(std::vector<SCEVHandle> Ops) {
+ assert(!Ops.empty() && "Cannot get empty smax!");
+ if (Ops.size() == 1) return Ops[0];
+
+ // Sort by complexity, this groups all similar expression types together.
+ GroupByComplexity(Ops);
+
+ // If there are any constants, fold them together.
+ unsigned Idx = 0;
+ if (SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
+ ++Idx;
+ assert(Idx < Ops.size());
+ while (SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
+ // We found two constants, fold them together!
+ ConstantInt *Fold = ConstantInt::get(
+ APIntOps::smax(LHSC->getValue()->getValue(),
+ RHSC->getValue()->getValue()));
+ Ops[0] = getConstant(Fold);
+ Ops.erase(Ops.begin()+1); // Erase the folded element
+ if (Ops.size() == 1) return Ops[0];
+ LHSC = cast<SCEVConstant>(Ops[0]);
+ }
+
+ // If we are left with a constant -inf, strip it off.
+ if (cast<SCEVConstant>(Ops[0])->getValue()->isMinValue(true)) {
+ Ops.erase(Ops.begin());
+ --Idx;
+ }
+ }
+
+ if (Ops.size() == 1) return Ops[0];
+
+ // Find the first SMax
+ while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scSMaxExpr)
+ ++Idx;
+
+ // Check to see if one of the operands is an SMax. If so, expand its operands
+ // onto our operand list, and recurse to simplify.
+ if (Idx < Ops.size()) {
+ bool DeletedSMax = false;
+ while (SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(Ops[Idx])) {
+ Ops.insert(Ops.end(), SMax->op_begin(), SMax->op_end());
+ Ops.erase(Ops.begin()+Idx);
+ DeletedSMax = true;
+ }
+
+ if (DeletedSMax)
+ return getSMaxExpr(Ops);
+ }
+
+ // Okay, check to see if the same value occurs in the operand list twice. If
+ // so, delete one. Since we sorted the list, these values are required to
+ // be adjacent.
+ for (unsigned i = 0, e = Ops.size()-1; i != e; ++i)
+ if (Ops[i] == Ops[i+1]) { // X smax Y smax Y --> X smax Y
+ Ops.erase(Ops.begin()+i, Ops.begin()+i+1);
+ --i; --e;
+ }
+
+ if (Ops.size() == 1) return Ops[0];
+
+ assert(!Ops.empty() && "Reduced smax down to nothing!");
+
+ // Okay, it looks like we really DO need an smax expr. Check to see if we
+ // already have one, otherwise create a new one.
+ std::vector<SCEV*> SCEVOps(Ops.begin(), Ops.end());
+ SCEVCommutativeExpr *&Result = (*SCEVCommExprs)[std::make_pair(scSMaxExpr,
+ SCEVOps)];
+ if (Result == 0) Result = new SCEVSMaxExpr(Ops);
+ return Result;
+}
+
+SCEVHandle ScalarEvolution::getUMaxExpr(const SCEVHandle &LHS,
+ const SCEVHandle &RHS) {
+ std::vector<SCEVHandle> Ops;
+ Ops.push_back(LHS);
+ Ops.push_back(RHS);
+ return getUMaxExpr(Ops);
+}
+
+SCEVHandle ScalarEvolution::getUMaxExpr(std::vector<SCEVHandle> Ops) {
+ assert(!Ops.empty() && "Cannot get empty umax!");
+ if (Ops.size() == 1) return Ops[0];
+
+ // Sort by complexity, this groups all similar expression types together.
+ GroupByComplexity(Ops);
+
+ // If there are any constants, fold them together.
+ unsigned Idx = 0;
+ if (SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
+ ++Idx;
+ assert(Idx < Ops.size());
+ while (SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) {
+ // We found two constants, fold them together!
+ ConstantInt *Fold = ConstantInt::get(
+ APIntOps::umax(LHSC->getValue()->getValue(),
+ RHSC->getValue()->getValue()));
+ Ops[0] = getConstant(Fold);
+ Ops.erase(Ops.begin()+1); // Erase the folded element
+ if (Ops.size() == 1) return Ops[0];
+ LHSC = cast<SCEVConstant>(Ops[0]);
+ }
+
+ // If we are left with a constant zero, strip it off.
+ if (cast<SCEVConstant>(Ops[0])->getValue()->isMinValue(false)) {
+ Ops.erase(Ops.begin());
+ --Idx;
+ }
+ }
+
+ if (Ops.size() == 1) return Ops[0];
+
+ // Find the first UMax
+ while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scUMaxExpr)
+ ++Idx;
+
+ // Check to see if one of the operands is a UMax. If so, expand its operands
+ // onto our operand list, and recurse to simplify.
+ if (Idx < Ops.size()) {
+ bool DeletedUMax = false;
+ while (SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(Ops[Idx])) {
+ Ops.insert(Ops.end(), UMax->op_begin(), UMax->op_end());
+ Ops.erase(Ops.begin()+Idx);
+ DeletedUMax = true;
+ }
+
+ if (DeletedUMax)
+ return getUMaxExpr(Ops);
+ }
+
+ // Okay, check to see if the same value occurs in the operand list twice. If
+ // so, delete one. Since we sorted the list, these values are required to
+ // be adjacent.
+ for (unsigned i = 0, e = Ops.size()-1; i != e; ++i)
+ if (Ops[i] == Ops[i+1]) { // X umax Y umax Y --> X umax Y
+ Ops.erase(Ops.begin()+i, Ops.begin()+i+1);
+ --i; --e;
+ }
+
+ if (Ops.size() == 1) return Ops[0];
+
+ assert(!Ops.empty() && "Reduced umax down to nothing!");
+
+ // Okay, it looks like we really DO need a umax expr. Check to see if we
+ // already have one, otherwise create a new one.
+ std::vector<SCEV*> SCEVOps(Ops.begin(), Ops.end());
+ SCEVCommutativeExpr *&Result = (*SCEVCommExprs)[std::make_pair(scUMaxExpr,
+ SCEVOps)];
+ if (Result == 0) Result = new SCEVUMaxExpr(Ops);
+ return Result;
+}
+
+SCEVHandle ScalarEvolution::getUnknown(Value *V) {
if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
- return SCEVConstant::get(CI);
+ return getConstant(CI);
SCEVUnknown *&Result = (*SCEVUnknowns)[V];
if (Result == 0) Result = new SCEVUnknown(V);
return Result;
///
namespace {
struct VISIBILITY_HIDDEN ScalarEvolutionsImpl {
+ /// SE - A reference to the public ScalarEvolution object.
+ ScalarEvolution &SE;
+
/// F - The function we are analyzing.
///
Function &F;
std::map<PHINode*, Constant*> ConstantEvolutionLoopExitValue;
public:
- ScalarEvolutionsImpl(Function &f, LoopInfo &li)
- : F(f), LI(li), UnknownValue(new SCEVCouldNotCompute()) {}
+ ScalarEvolutionsImpl(ScalarEvolution &se, Function &f, LoopInfo &li)
+ : SE(se), F(f), LI(li), UnknownValue(new SCEVCouldNotCompute()) {}
/// getSCEV - Return an existing SCEV if it exists, otherwise analyze the
/// expression and create a new one.
SCEVHandle ComputeIterationCount(const Loop *L);
/// ComputeLoadConstantCompareIterationCount - Given an exit condition of
- /// 'setcc load X, cst', try to see if we can compute the trip count.
+ /// 'icmp op load X, cst', try to see if we can compute the trip count.
SCEVHandle ComputeLoadConstantCompareIterationCount(LoadInst *LI,
Constant *RHS,
const Loop *L,
/// HowManyLessThans - Return the number of times a backedge containing the
/// specified less-than comparison will execute. If not computable, return
- /// UnknownValue.
- SCEVHandle HowManyLessThans(SCEV *LHS, SCEV *RHS, const Loop *L);
+ /// 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
if (SI == Scalars.end()) return;
SCEVHandle NV =
- SI->second->replaceSymbolicValuesWithConcrete(SymName, NewVal);
+ SI->second->replaceSymbolicValuesWithConcrete(SymName, NewVal, SE);
if (NV == SI->second) return; // No change.
SI->second = NV; // Update the scalars map!
unsigned BackEdge = IncomingEdge^1;
// While we are analyzing this PHI node, handle its value symbolically.
- SCEVHandle SymbolicName = SCEVUnknown::get(PN);
+ SCEVHandle SymbolicName = SE.getUnknown(PN);
assert(Scalars.find(PN) == Scalars.end() &&
"PHI node already processed?");
Scalars.insert(std::make_pair(PN, SymbolicName));
for (unsigned i = 0, e = Add->getNumOperands(); i != e; ++i)
if (i != FoundIndex)
Ops.push_back(Add->getOperand(i));
- SCEVHandle Accum = SCEVAddExpr::get(Ops);
+ SCEVHandle Accum = SE.getAddExpr(Ops);
// This is not a valid addrec if the step amount is varying each
// loop iteration, but is not itself an addrec in this loop.
(isa<SCEVAddRecExpr>(Accum) &&
cast<SCEVAddRecExpr>(Accum)->getLoop() == L)) {
SCEVHandle StartVal = getSCEV(PN->getIncomingValue(IncomingEdge));
- SCEVHandle PHISCEV = SCEVAddRecExpr::get(StartVal, Accum, L);
+ SCEVHandle PHISCEV = SE.getAddRecExpr(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
// 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))) {
+ if (StartVal == SE.getMinusSCEV(AddRec->getOperand(0),
+ AddRec->getOperand(1))) {
SCEVHandle PHISCEV =
- SCEVAddRecExpr::get(StartVal, AddRec->getOperand(1), L);
+ SE.getAddRecExpr(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
}
// If it's not a loop phi, we can't handle it yet.
- return SCEVUnknown::get(PN);
-}
-
-/// 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<SCEVConstant>(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);
+ return SE.getUnknown(PN);
+}
+
+/// GetMinTrailingZeros - Determine the minimum number of zero bits that S is
+/// guaranteed to end in (at every loop iteration). It is, at the same time,
+/// the minimum number of times S is divisible by 2. For example, given {4,+,8}
+/// it returns 2. If S is guaranteed to be 0, it returns the bitwidth of S.
+static uint32_t GetMinTrailingZeros(SCEVHandle S) {
+ if (SCEVConstant *C = dyn_cast<SCEVConstant>(S))
+ return C->getValue()->getValue().countTrailingZeros();
+
+ if (SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(S))
+ return std::min(GetMinTrailingZeros(T->getOperand()), T->getBitWidth());
+
+ if (SCEVZeroExtendExpr *E = dyn_cast<SCEVZeroExtendExpr>(S)) {
+ uint32_t OpRes = GetMinTrailingZeros(E->getOperand());
+ return OpRes == E->getOperand()->getBitWidth() ? E->getBitWidth() : OpRes;
}
- if (SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(S)) {
- return GetConstantFactor(T->getOperand()).trunc(
- cast<IntegerType>(T->getType())->getBitWidth());
+ if (SCEVSignExtendExpr *E = dyn_cast<SCEVSignExtendExpr>(S)) {
+ uint32_t OpRes = GetMinTrailingZeros(E->getOperand());
+ return OpRes == E->getOperand()->getBitWidth() ? E->getBitWidth() : OpRes;
}
- if (SCEVZeroExtendExpr *E = dyn_cast<SCEVZeroExtendExpr>(S))
- return GetConstantFactor(E->getOperand()).zext(
- cast<IntegerType>(E->getType())->getBitWidth());
- if (SCEVSignExtendExpr *E = dyn_cast<SCEVSignExtendExpr>(S))
- return GetConstantFactor(E->getOperand()).sext(
- cast<IntegerType>(E->getType())->getBitWidth());
-
+
if (SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(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;
+ // The result is the min of all operands results.
+ uint32_t MinOpRes = GetMinTrailingZeros(A->getOperand(0));
+ for (unsigned i = 1, e = A->getNumOperands(); MinOpRes && i != e; ++i)
+ MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(A->getOperand(i)));
+ return MinOpRes;
}
if (SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(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;
+ // The result is the sum of all operands results.
+ uint32_t SumOpRes = GetMinTrailingZeros(M->getOperand(0));
+ uint32_t BitWidth = M->getBitWidth();
+ for (unsigned i = 1, e = M->getNumOperands();
+ SumOpRes != BitWidth && i != e; ++i)
+ SumOpRes = std::min(SumOpRes + GetMinTrailingZeros(M->getOperand(i)),
+ BitWidth);
+ return SumOpRes;
}
-
+
if (SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(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);
- }
+ // The result is the min of all operands results.
+ uint32_t MinOpRes = GetMinTrailingZeros(A->getOperand(0));
+ for (unsigned i = 1, e = A->getNumOperands(); MinOpRes && i != e; ++i)
+ MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(A->getOperand(i)));
+ return MinOpRes;
}
-
- // SCEVSDivExpr, SCEVUnknown.
- return APInt(S->getBitWidth(), 1);
+
+ if (SCEVSMaxExpr *M = dyn_cast<SCEVSMaxExpr>(S)) {
+ // The result is the min of all operands results.
+ uint32_t MinOpRes = GetMinTrailingZeros(M->getOperand(0));
+ for (unsigned i = 1, e = M->getNumOperands(); MinOpRes && i != e; ++i)
+ MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(M->getOperand(i)));
+ return MinOpRes;
+ }
+
+ if (SCEVUMaxExpr *M = dyn_cast<SCEVUMaxExpr>(S)) {
+ // The result is the min of all operands results.
+ uint32_t MinOpRes = GetMinTrailingZeros(M->getOperand(0));
+ for (unsigned i = 1, e = M->getNumOperands(); MinOpRes && i != e; ++i)
+ MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(M->getOperand(i)));
+ return MinOpRes;
+ }
+
+ // SCEVUDivExpr, SCEVUnknown
+ return 0;
}
/// createSCEV - We know that there is no SCEV for the specified value.
/// Analyze the expression.
///
SCEVHandle ScalarEvolutionsImpl::createSCEV(Value *V) {
+ if (!isa<IntegerType>(V->getType()))
+ return SE.getUnknown(V);
+
if (Instruction *I = dyn_cast<Instruction>(V)) {
switch (I->getOpcode()) {
case Instruction::Add:
- return SCEVAddExpr::get(getSCEV(I->getOperand(0)),
- getSCEV(I->getOperand(1)));
+ return SE.getAddExpr(getSCEV(I->getOperand(0)),
+ getSCEV(I->getOperand(1)));
case Instruction::Mul:
- return SCEVMulExpr::get(getSCEV(I->getOperand(0)),
- getSCEV(I->getOperand(1)));
- case Instruction::SDiv:
- return SCEVSDivExpr::get(getSCEV(I->getOperand(0)),
- getSCEV(I->getOperand(1)));
- break;
-
+ return SE.getMulExpr(getSCEV(I->getOperand(0)),
+ getSCEV(I->getOperand(1)));
+ case Instruction::UDiv:
+ return SE.getUDivExpr(getSCEV(I->getOperand(0)),
+ getSCEV(I->getOperand(1)));
case Instruction::Sub:
- return SCEV::getMinusSCEV(getSCEV(I->getOperand(0)),
- getSCEV(I->getOperand(1)));
+ return SE.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
+ // X*4+1 which got turned into X*4|1. Handle this as an Add so loop
// optimizations will transparently handle this case.
+ //
+ // In order for this transformation to be safe, the LHS must be of the
+ // form X*(2^n) and the Or constant must be less than 2^n.
if (ConstantInt *CI = dyn_cast<ConstantInt>(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)));
- }
+ const APInt &CIVal = CI->getValue();
+ if (GetMinTrailingZeros(LHS) >=
+ (CIVal.getBitWidth() - CIVal.countLeadingZeros()))
+ return SE.getAddExpr(LHS, getSCEV(I->getOperand(1)));
}
break;
case Instruction::Xor:
// Instcombine turns add of signbit into xor as a strength reduction step.
if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
if (CI->getValue().isSignBit())
- return SCEVAddExpr::get(getSCEV(I->getOperand(0)),
- getSCEV(I->getOperand(1)));
+ return SE.getAddExpr(getSCEV(I->getOperand(0)),
+ getSCEV(I->getOperand(1)));
+ else if (CI->isAllOnesValue())
+ return SE.getNotSCEV(getSCEV(I->getOperand(0)));
}
break;
uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
Constant *X = ConstantInt::get(
APInt(BitWidth, 1).shl(SA->getLimitedValue(BitWidth)));
- return SCEVMulExpr::get(getSCEV(I->getOperand(0)), getSCEV(X));
+ return SE.getMulExpr(getSCEV(I->getOperand(0)), getSCEV(X));
}
break;
case Instruction::Trunc:
- return SCEVTruncateExpr::get(getSCEV(I->getOperand(0)), I->getType());
+ return SE.getTruncateExpr(getSCEV(I->getOperand(0)), I->getType());
case Instruction::ZExt:
- return SCEVZeroExtendExpr::get(getSCEV(I->getOperand(0)), I->getType());
+ return SE.getZeroExtendExpr(getSCEV(I->getOperand(0)), I->getType());
case Instruction::SExt:
- return SCEVSignExtendExpr::get(getSCEV(I->getOperand(0)), I->getType());
+ return SE.getSignExtendExpr(getSCEV(I->getOperand(0)), I->getType());
case Instruction::BitCast:
// BitCasts are no-op casts so we just eliminate the cast.
case Instruction::PHI:
return createNodeForPHI(cast<PHINode>(I));
+ case Instruction::Select:
+ // This could be a smax or umax that was lowered earlier.
+ // Try to recover it.
+ if (ICmpInst *ICI = dyn_cast<ICmpInst>(I->getOperand(0))) {
+ Value *LHS = ICI->getOperand(0);
+ Value *RHS = ICI->getOperand(1);
+ switch (ICI->getPredicate()) {
+ case ICmpInst::ICMP_SLT:
+ case ICmpInst::ICMP_SLE:
+ std::swap(LHS, RHS);
+ // fall through
+ case ICmpInst::ICMP_SGT:
+ case ICmpInst::ICMP_SGE:
+ if (LHS == I->getOperand(1) && RHS == I->getOperand(2))
+ return SE.getSMaxExpr(getSCEV(LHS), getSCEV(RHS));
+ else if (LHS == I->getOperand(2) && RHS == I->getOperand(1))
+ // -smax(-x, -y) == smin(x, y).
+ return SE.getNegativeSCEV(SE.getSMaxExpr(
+ SE.getNegativeSCEV(getSCEV(LHS)),
+ SE.getNegativeSCEV(getSCEV(RHS))));
+ break;
+ case ICmpInst::ICMP_ULT:
+ case ICmpInst::ICMP_ULE:
+ std::swap(LHS, RHS);
+ // fall through
+ case ICmpInst::ICMP_UGT:
+ case ICmpInst::ICMP_UGE:
+ if (LHS == I->getOperand(1) && RHS == I->getOperand(2))
+ return SE.getUMaxExpr(getSCEV(LHS), getSCEV(RHS));
+ else if (LHS == I->getOperand(2) && RHS == I->getOperand(1))
+ // ~umax(~x, ~y) == umin(x, y)
+ return SE.getNotSCEV(SE.getUMaxExpr(SE.getNotSCEV(getSCEV(LHS)),
+ SE.getNotSCEV(getSCEV(RHS))));
+ break;
+ default:
+ break;
+ }
+ }
+
default: // We cannot analyze this expression.
break;
}
}
- return SCEVUnknown::get(V);
+ return SE.getUnknown(V);
}
/// 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<BasicBlock*> ExitBlocks;
+ SmallVector<BasicBlock*, 8> ExitBlocks;
L->getExitBlocks(ExitBlocks);
if (ExitBlocks.size() != 1) return UnknownValue;
ICmpInst *ExitCond = dyn_cast<ICmpInst>(ExitBr->getCondition());
- // If its not an integer comparison then compute it the hard way.
+ // If it's 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<PointerType>(ExitCond->getOperand(0)->getType()))
ConstantRange CompRange(
ICmpInst::makeConstantRange(Cond, CompVal->getValue()));
- SCEVHandle Ret = AddRec->getNumIterationsInRange(CompRange,
- false /*Always treat as unsigned range*/);
+ SCEVHandle Ret = AddRec->getNumIterationsInRange(CompRange, SE);
if (!isa<SCEVCouldNotCompute>(Ret)) return Ret;
}
}
switch (Cond) {
case ICmpInst::ICMP_NE: { // while (X != Y)
// Convert to: while (X-Y != 0)
- SCEVHandle TC = HowFarToZero(SCEV::getMinusSCEV(LHS, RHS), L);
+ SCEVHandle TC = HowFarToZero(SE.getMinusSCEV(LHS, RHS), L);
if (!isa<SCEVCouldNotCompute>(TC)) return TC;
break;
}
case ICmpInst::ICMP_EQ: {
// Convert to: while (X-Y == 0) // while (X == Y)
- SCEVHandle TC = HowFarToNonZero(SCEV::getMinusSCEV(LHS, RHS), L);
+ SCEVHandle TC = HowFarToNonZero(SE.getMinusSCEV(LHS, RHS), L);
if (!isa<SCEVCouldNotCompute>(TC)) return TC;
break;
}
case ICmpInst::ICMP_SLT: {
- SCEVHandle TC = HowManyLessThans(LHS, RHS, L);
+ SCEVHandle TC = HowManyLessThans(LHS, RHS, L, true);
if (!isa<SCEVCouldNotCompute>(TC)) return TC;
break;
}
case ICmpInst::ICMP_SGT: {
- SCEVHandle TC = HowManyLessThans(RHS, LHS, L);
+ SCEVHandle TC = HowManyLessThans(SE.getNegativeSCEV(LHS),
+ SE.getNegativeSCEV(RHS), L, true);
+ if (!isa<SCEVCouldNotCompute>(TC)) return TC;
+ break;
+ }
+ case ICmpInst::ICMP_ULT: {
+ SCEVHandle TC = HowManyLessThans(LHS, RHS, L, false);
+ if (!isa<SCEVCouldNotCompute>(TC)) return TC;
+ break;
+ }
+ case ICmpInst::ICMP_UGT: {
+ SCEVHandle TC = HowManyLessThans(SE.getNotSCEV(LHS),
+ SE.getNotSCEV(RHS), L, false);
if (!isa<SCEVCouldNotCompute>(TC)) return TC;
break;
}
}
static ConstantInt *
-EvaluateConstantChrecAtConstant(const SCEVAddRecExpr *AddRec, ConstantInt *C) {
- SCEVHandle InVal = SCEVConstant::get(C);
- SCEVHandle Val = AddRec->evaluateAtIteration(InVal);
+EvaluateConstantChrecAtConstant(const SCEVAddRecExpr *AddRec, ConstantInt *C,
+ ScalarEvolution &SE) {
+ SCEVHandle InVal = SE.getConstant(C);
+ SCEVHandle Val = AddRec->evaluateAtIteration(InVal, SE);
assert(isa<SCEVConstant>(Val) &&
"Evaluation of SCEV at constant didn't fold correctly?");
return cast<SCEVConstant>(Val)->getValue();
}
/// ComputeLoadConstantCompareIterationCount - Given an exit condition of
-/// 'setcc load X, cst', try to se if we can compute the trip count.
+/// 'icmp op load X, cst', try to see if we can compute the trip count.
SCEVHandle ScalarEvolutionsImpl::
ComputeLoadConstantCompareIterationCount(LoadInst *LI, Constant *RHS,
const Loop *L,
for (unsigned IterationNum = 0; IterationNum != MaxSteps; ++IterationNum) {
ConstantInt *ItCst =
ConstantInt::get(IdxExpr->getType(), IterationNum);
- ConstantInt *Val = EvaluateConstantChrecAtConstant(IdxExpr, ItCst);
+ ConstantInt *Val = EvaluateConstantChrecAtConstant(IdxExpr, ItCst, SE);
// Form the GEP offset.
Indexes[VarIdxNum] = Val;
<< "***\n";
#endif
++NumArrayLenItCounts;
- return SCEVConstant::get(ItCst); // Found terminating iteration!
+ return SE.getConstant(ItCst); // Found terminating iteration!
}
}
return UnknownValue;
if (const CallInst *CI = dyn_cast<CallInst>(I))
if (const Function *F = CI->getCalledFunction())
- return canConstantFoldCallTo((Function*)F); // FIXME: elim cast
+ return canConstantFoldCallTo(F);
return false;
}
Instruction *I = dyn_cast<Instruction>(V);
if (I == 0 || !L->contains(I->getParent())) return 0;
- if (PHINode *PN = dyn_cast<PHINode>(I))
+ if (PHINode *PN = dyn_cast<PHINode>(I)) {
if (L->getHeader() == I->getParent())
return PN;
else
// We don't currently keep track of the control flow needed to evaluate
// PHIs, so we cannot handle PHIs inside of loops.
return 0;
+ }
// If we won't be able to constant fold this expression even if the operands
// are constants, return early.
/// reason, return null.
static Constant *EvaluateExpression(Value *V, Constant *PHIVal) {
if (isa<PHINode>(V)) return PHIVal;
- if (GlobalValue *GV = dyn_cast<GlobalValue>(V))
- return GV;
if (Constant *C = dyn_cast<Constant>(V)) return C;
Instruction *I = cast<Instruction>(V);
if (Operands[i] == 0) return 0;
}
- return ConstantFoldInstOperands(I, &Operands[0], Operands.size());
+ if (const CmpInst *CI = dyn_cast<CmpInst>(I))
+ return ConstantFoldCompareInstOperands(CI->getPredicate(),
+ &Operands[0], Operands.size());
+ else
+ return ConstantFoldInstOperands(I->getOpcode(), I->getType(),
+ &Operands[0], Operands.size());
}
/// getConstantEvolutionLoopExitValue - If we know that the specified Phi is
if (CondVal->getValue() == uint64_t(ExitWhen)) {
ConstantEvolutionLoopExitValue[PN] = PHIVal;
++NumBruteForceTripCountsComputed;
- return SCEVConstant::get(ConstantInt::get(Type::Int32Ty, IterationNum));
+ return SE.getConstant(ConstantInt::get(Type::Int32Ty, IterationNum));
}
// Compute the value of the PHI node for the next iteration.
if (isa<SCEVConstant>(V)) return V;
- // If this instruction is evolves from a constant-evolving PHI, compute the
+ // If this instruction is evolved from a constant-evolving PHI, compute the
// exit value from the loop without using SCEVs.
if (SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V)) {
if (Instruction *I = dyn_cast<Instruction>(SU->getValue())) {
Constant *RV = getConstantEvolutionLoopExitValue(PN,
ICC->getValue()->getValue(),
LI);
- if (RV) return SCEVUnknown::get(RV);
+ if (RV) return SE.getUnknown(RV);
}
}
if (Constant *C = dyn_cast<Constant>(Op)) {
Operands.push_back(C);
} else {
+ // If any of the operands is non-constant and if they are
+ // non-integer, don't even try to analyze them with scev techniques.
+ if (!isa<IntegerType>(Op->getType()))
+ return V;
+
SCEVHandle OpV = getSCEVAtScope(getSCEV(Op), L);
if (SCEVConstant *SC = dyn_cast<SCEVConstant>(OpV))
Operands.push_back(ConstantExpr::getIntegerCast(SC->getValue(),
}
}
}
- Constant *C =ConstantFoldInstOperands(I, &Operands[0], Operands.size());
- return SCEVUnknown::get(C);
+
+ Constant *C;
+ if (const CmpInst *CI = dyn_cast<CmpInst>(I))
+ C = ConstantFoldCompareInstOperands(CI->getPredicate(),
+ &Operands[0], Operands.size());
+ else
+ C = ConstantFoldInstOperands(I->getOpcode(), I->getType(),
+ &Operands[0], Operands.size());
+ return SE.getUnknown(C);
}
}
NewOps.push_back(OpAtScope);
}
if (isa<SCEVAddExpr>(Comm))
- return SCEVAddExpr::get(NewOps);
- assert(isa<SCEVMulExpr>(Comm) && "Only know about add and mul!");
- return SCEVMulExpr::get(NewOps);
+ return SE.getAddExpr(NewOps);
+ if (isa<SCEVMulExpr>(Comm))
+ return SE.getMulExpr(NewOps);
+ if (isa<SCEVSMaxExpr>(Comm))
+ return SE.getSMaxExpr(NewOps);
+ if (isa<SCEVUMaxExpr>(Comm))
+ return SE.getUMaxExpr(NewOps);
+ assert(0 && "Unknown commutative SCEV type!");
}
}
// If we got here, all operands are loop invariant.
return Comm;
}
- if (SCEVSDivExpr *Div = dyn_cast<SCEVSDivExpr>(V)) {
+ if (SCEVUDivExpr *Div = dyn_cast<SCEVUDivExpr>(V)) {
SCEVHandle LHS = getSCEVAtScope(Div->getLHS(), L);
if (LHS == UnknownValue) return LHS;
SCEVHandle RHS = getSCEVAtScope(Div->getRHS(), L);
if (RHS == UnknownValue) return RHS;
if (LHS == Div->getLHS() && RHS == Div->getRHS())
return Div; // must be loop invariant
- return SCEVSDivExpr::get(LHS, RHS);
+ return SE.getUDivExpr(LHS, RHS);
}
// If this is a loop recurrence for a loop that does not contain L, then we
SCEVHandle IterationCount = getIterationCount(AddRec->getLoop());
if (IterationCount == UnknownValue) return UnknownValue;
IterationCount = getTruncateOrZeroExtend(IterationCount,
- AddRec->getType());
+ AddRec->getType(), SE);
// If the value is affine, simplify the expression evaluation to just
// Start + Step*IterationCount.
if (AddRec->isAffine())
- return SCEVAddExpr::get(AddRec->getStart(),
- SCEVMulExpr::get(IterationCount,
- AddRec->getOperand(1)));
+ return SE.getAddExpr(AddRec->getStart(),
+ SE.getMulExpr(IterationCount,
+ AddRec->getOperand(1)));
// Otherwise, evaluate it the hard way.
- return AddRec->evaluateAtIteration(IterationCount);
+ return AddRec->evaluateAtIteration(IterationCount, SE);
}
return UnknownValue;
}
/// might be the same) or two SCEVCouldNotCompute objects.
///
static std::pair<SCEVHandle,SCEVHandle>
-SolveQuadraticEquation(const SCEVAddRecExpr *AddRec) {
+SolveQuadraticEquation(const SCEVAddRecExpr *AddRec, ScalarEvolution &SE) {
assert(AddRec->getNumOperands() == 3 && "This is not a quadratic chrec!");
SCEVConstant *LC = dyn_cast<SCEVConstant>(AddRec->getOperand(0));
SCEVConstant *MC = dyn_cast<SCEVConstant>(AddRec->getOperand(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));
+ return std::make_pair(SE.getConstant(Solution1),
+ SE.getConstant(Solution2));
} // end APIntOps namespace
}
// FIXME: We should add DivExpr and RemExpr operations to our AST.
if (SCEVConstant *StepC = dyn_cast<SCEVConstant>(Step)) {
if (StepC->getValue()->equalsInt(1)) // N % 1 == 0
- return SCEV::getNegativeSCEV(Start); // 0 - Start/1 == -Start
+ return SE.getNegativeSCEV(Start); // 0 - Start/1 == -Start
if (StepC->getValue()->isAllOnesValue()) // N % -1 == 0
return Start; // 0 - Start/-1 == Start
Constant *Rem = ConstantExpr::getSRem(StartNegC, StepC->getValue());
if (Rem->isNullValue()) {
Constant *Result =ConstantExpr::getSDiv(StartNegC,StepC->getValue());
- return SCEVUnknown::get(Result);
+ return SE.getUnknown(Result);
}
}
}
} else if (AddRec->isQuadratic() && AddRec->getType()->isInteger()) {
// If this is a quadratic (3-term) AddRec {L,+,M,+,N}, find the roots of
// the quadratic equation to solve it.
- std::pair<SCEVHandle,SCEVHandle> Roots = SolveQuadraticEquation(AddRec);
+ std::pair<SCEVHandle,SCEVHandle> Roots = SolveQuadraticEquation(AddRec, SE);
SCEVConstant *R1 = dyn_cast<SCEVConstant>(Roots.first);
SCEVConstant *R2 = dyn_cast<SCEVConstant>(Roots.second);
if (R1) {
// 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);
+ SCEVHandle Val = AddRec->evaluateAtIteration(R1, SE);
if (SCEVConstant *EvalVal = dyn_cast<SCEVConstant>(Val))
if (EvalVal->getValue()->isZero())
return R1; // We found a quadratic root!
// 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<SCEVConstant>(V)) {
- Constant *Zero = Constant::getNullValue(C->getValue()->getType());
- Constant *NonZero =
- ConstantExpr::getICmp(ICmpInst::ICMP_NE, C->getValue(), Zero);
- if (NonZero == ConstantInt::getTrue())
- return getSCEV(Zero);
+ if (!C->getValue()->isNullValue())
+ return SE.getIntegerSCEV(0, C->getType());
return UnknownValue; // Otherwise it will loop infinitely.
}
/// specified less-than comparison will execute. If not computable, return
/// UnknownValue.
SCEVHandle ScalarEvolutionsImpl::
-HowManyLessThans(SCEV *LHS, SCEV *RHS, const Loop *L) {
+HowManyLessThans(SCEV *LHS, SCEV *RHS, const Loop *L, bool isSigned) {
// Only handle: "ADDREC < LoopInvariant".
if (!RHS->isLoopInvariant(L)) return UnknownValue;
if (AddRec->isAffine()) {
// FORNOW: We only support unit strides.
- SCEVHandle One = SCEVUnknown::getIntegerSCEV(1, RHS->getType());
+ SCEVHandle One = SE.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]<m. If we can find this check,
- // we know that the SMAX will evaluate to m-n, because we know that m >= n.
-
- // Search for the check.
- BasicBlock *Preheader = L->getLoopPreheader();
- BasicBlock *PreheaderDest = L->getHeader();
- if (Preheader == 0) return UnknownValue;
-
- BranchInst *LoopEntryPredicate =
- dyn_cast<BranchInst>(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<BranchInst>(Preheader->getTerminator());
- if (!LoopEntryPredicate) return UnknownValue;
- }
+ // We know the LHS is of the form {n,+,1} and the RHS is some loop-invariant
+ // m. So, we count the number of iterations in which {n,+,1} < m is true.
+ // Note that we cannot simply return max(m-n,0) because it's not safe to
+ // treat m-n as signed nor unsigned due to overflow possibility.
- // 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<ICmpInst>(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:
- std::swap(PreCondLHS, PreCondRHS);
- Cond = ICmpInst::ICMP_ULT;
- break;
- case ICmpInst::ICMP_SGT:
- std::swap(PreCondLHS, PreCondRHS);
- Cond = ICmpInst::ICMP_SLT;
- break;
- default: break;
- }
+ // First, we get the value of the LHS in the first iteration: n
+ SCEVHandle Start = AddRec->getOperand(0);
- if (Cond == ICmpInst::ICMP_SLT) {
- 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;
- } else if (Cond == ICmpInst::ICMP_ULT)
- return UnknownValue;
+ // Then, we get the value of the LHS in the first iteration in which the
+ // above condition doesn't hold. This equals to max(m,n).
+ SCEVHandle End = isSigned ? SE.getSMaxExpr(RHS, Start)
+ : SE.getUMaxExpr(RHS, Start);
- // cerr << "Computed Loop Trip Count as: "
- // << // *SCEV::getMinusSCEV(RHS, AddRec->getOperand(0)) << "\n";
- return SCEV::getMinusSCEV(RHS, AddRec->getOperand(0));
- }
- else
- return UnknownValue;
+ // Finally, we subtract these two values to get the number of times the
+ // backedge is executed: max(m,n)-n.
+ return SE.getMinusSCEV(End, Start);
}
return UnknownValue;
/// this is that it returns the first iteration number where the value is not in
/// the condition, thus computing the exit count. If the iteration count can't
/// be computed, an instance of SCEVCouldNotCompute is returned.
-SCEVHandle SCEVAddRecExpr::getNumIterationsInRange(ConstantRange Range,
- bool isSigned) const {
+SCEVHandle SCEVAddRecExpr::getNumIterationsInRange(ConstantRange Range,
+ ScalarEvolution &SE) const {
if (Range.isFullSet()) // Infinite loop.
return new SCEVCouldNotCompute();
if (SCEVConstant *SC = dyn_cast<SCEVConstant>(getStart()))
if (!SC->getValue()->isZero()) {
std::vector<SCEVHandle> Operands(op_begin(), op_end());
- Operands[0] = SCEVUnknown::getIntegerSCEV(0, SC->getType());
- SCEVHandle Shifted = SCEVAddRecExpr::get(Operands, getLoop());
+ Operands[0] = SE.getIntegerSCEV(0, SC->getType());
+ SCEVHandle Shifted = SE.getAddRecExpr(Operands, getLoop());
if (SCEVAddRecExpr *ShiftedAddRec = dyn_cast<SCEVAddRecExpr>(Shifted))
return ShiftedAddRec->getNumIterationsInRange(
- Range.subtract(SC->getValue()->getValue()),isSigned);
+ Range.subtract(SC->getValue()->getValue()), SE);
// This is strange and shouldn't happen.
return new SCEVCouldNotCompute();
}
// First check to see if the range contains zero. If not, the first
// iteration exits.
if (!Range.contains(APInt(getBitWidth(),0)))
- return SCEVConstant::get(ConstantInt::get(getType(),0));
+ return SE.getConstant(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.
- const APInt &Upper = Range.getUpper();
- APInt A = cast<SCEVConstant>(getOperand(1))->getValue()->getValue();
+ // 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<SCEVConstant>(getOperand(1))->getValue()->getValue();
+ APInt End = A.sge(One) ? (Range.getUpper() - One) : Range.getLower();
- // The exit value should be (Upper+A-1)/A.
- APInt ExitVal(Upper);
- if (A != One)
- ExitVal = (Upper + A - One).sdiv(A);
+ // 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);
+ ConstantInt *Val = EvaluateConstantChrecAtConstant(this, ExitValue, SE);
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,
- ConstantInt::get(ExitVal - One))->getValue()) &&
+ ConstantInt::get(ExitVal - One), SE)->getValue()) &&
"Linear scev computation is off in a bad way!");
- return SCEVConstant::get(ExitValue);
+ return SE.getConstant(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<SCEVHandle> NewOps(op_begin(), op_end());
- NewOps[0] = SCEV::getNegativeSCEV(SCEVConstant::get(Range.getUpper()));
- SCEVHandle NewAddRec = SCEVAddRecExpr::get(NewOps, getLoop());
+ NewOps[0] = SE.getNegativeSCEV(SE.getConstant(Range.getUpper()));
+ SCEVHandle NewAddRec = SE.getAddRecExpr(NewOps, getLoop());
// Next, solve the constructed addrec
std::pair<SCEVHandle,SCEVHandle> Roots =
- SolveQuadraticEquation(cast<SCEVAddRecExpr>(NewAddRec));
+ SolveQuadraticEquation(cast<SCEVAddRecExpr>(NewAddRec), SE);
SCEVConstant *R1 = dyn_cast<SCEVConstant>(Roots.first);
SCEVConstant *R2 = dyn_cast<SCEVConstant>(Roots.second);
if (R1) {
// 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());
+ R1->getValue(),
+ SE);
if (Range.contains(R1Val->getValue())) {
// The next iteration must be out of the range...
ConstantInt *NextVal = ConstantInt::get(R1->getValue()->getValue()+1);
- R1Val = EvaluateConstantChrecAtConstant(this, NextVal);
+ R1Val = EvaluateConstantChrecAtConstant(this, NextVal, SE);
if (!Range.contains(R1Val->getValue()))
- return SCEVConstant::get(NextVal);
+ return SE.getConstant(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.
ConstantInt *NextVal = ConstantInt::get(R1->getValue()->getValue()-1);
- R1Val = EvaluateConstantChrecAtConstant(this, NextVal);
+ R1Val = EvaluateConstantChrecAtConstant(this, NextVal, SE);
if (Range.contains(R1Val->getValue()))
return R1;
return new SCEVCouldNotCompute(); // Something strange happened
ConstantInt *EndVal = TestVal; // Stop when we wrap around.
do {
++NumBruteForceEvaluations;
- SCEVHandle Val = evaluateAtIteration(SCEVConstant::get(TestVal));
+ SCEVHandle Val = evaluateAtIteration(SE.getConstant(TestVal), SE);
if (!isa<SCEVConstant>(Val)) // This shouldn't happen.
return new SCEVCouldNotCompute();
// Check to see if we found the value!
if (!Range.contains(cast<SCEVConstant>(Val)->getValue()->getValue()))
- return SCEVConstant::get(TestVal);
+ return SE.getConstant(TestVal);
// Increment to test the next index.
TestVal = ConstantInt::get(TestVal->getValue()+1);
//===----------------------------------------------------------------------===//
bool ScalarEvolution::runOnFunction(Function &F) {
- Impl = new ScalarEvolutionsImpl(F, getAnalysis<LoopInfo>());
+ Impl = new ScalarEvolutionsImpl(*this, F, getAnalysis<LoopInfo>());
return false;
}
for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
PrintLoopInfo(OS, SE, *I);
- cerr << "Loop " << L->getHeader()->getName() << ": ";
+ OS << "Loop " << L->getHeader()->getName() << ": ";
- std::vector<BasicBlock*> ExitBlocks;
+ SmallVector<BasicBlock*, 8> ExitBlocks;
L->getExitBlocks(ExitBlocks);
if (ExitBlocks.size() != 1)
- cerr << "<multiple exits> ";
+ OS << "<multiple exits> ";
if (SE->hasLoopInvariantIterationCount(L)) {
- cerr << *SE->getIterationCount(L) << " iterations! ";
+ OS << *SE->getIterationCount(L) << " iterations! ";
} else {
- cerr << "Unpredictable iteration count. ";
+ OS << "Unpredictable iteration count. ";
}
- cerr << "\n";
+ OS << "\n";
}
void ScalarEvolution::print(std::ostream &OS, const Module* ) const {
for (LoopInfo::iterator I = LI.begin(), E = LI.end(); I != E; ++I)
PrintLoopInfo(OS, this, *I);
}
-