"derived loop"),
cl::init(100));
-static RegisterPass<ScalarEvolution>
-R("scalar-evolution", "Scalar Evolution Analysis", false, true);
+INITIALIZE_PASS(ScalarEvolution, "scalar-evolution",
+ "Scalar Evolution Analysis", false, true);
char ScalarEvolution::ID = 0;
//===----------------------------------------------------------------------===//
OS << "(";
for (op_iterator I = op_begin(), E = op_end(); I != E; ++I) {
OS << **I;
- if (next(I) != E)
+ if (llvm::next(I) != E)
OS << OpStr;
}
OS << ")";
OS << ">";
}
+void SCEVUnknown::deleted() {
+ // Clear this SCEVUnknown from ValuesAtScopes.
+ SE->ValuesAtScopes.erase(this);
+
+ // Remove this SCEVUnknown from the uniquing map.
+ SE->UniqueSCEVs.RemoveNode(this);
+
+ // Release the value.
+ setValPtr(0);
+}
+
+void SCEVUnknown::allUsesReplacedWith(Value *New) {
+ // Clear this SCEVUnknown from ValuesAtScopes.
+ SE->ValuesAtScopes.erase(this);
+
+ // Remove this SCEVUnknown from the uniquing map.
+ SE->UniqueSCEVs.RemoveNode(this);
+
+ // Update this SCEVUnknown to point to the new value. This is needed
+ // because there may still be outstanding SCEVs which still point to
+ // this SCEVUnknown.
+ setValPtr(New);
+}
+
bool SCEVUnknown::isLoopInvariant(const Loop *L) const {
// All non-instruction values are loop invariant. All instructions are loop
// invariant if they are not contained in the specified loop.
// Instructions are never considered invariant in the function body
// (null loop) because they are defined within the "loop".
- if (Instruction *I = dyn_cast<Instruction>(V))
+ if (Instruction *I = dyn_cast<Instruction>(getValue()))
return L && !L->contains(I);
return true;
}
}
const Type *SCEVUnknown::getType() const {
- return V->getType();
+ return getValue()->getType();
}
bool SCEVUnknown::isSizeOf(const Type *&AllocTy) const {
- if (ConstantExpr *VCE = dyn_cast<ConstantExpr>(V))
+ if (ConstantExpr *VCE = dyn_cast<ConstantExpr>(getValue()))
if (VCE->getOpcode() == Instruction::PtrToInt)
if (ConstantExpr *CE = dyn_cast<ConstantExpr>(VCE->getOperand(0)))
if (CE->getOpcode() == Instruction::GetElementPtr &&
}
bool SCEVUnknown::isAlignOf(const Type *&AllocTy) const {
- if (ConstantExpr *VCE = dyn_cast<ConstantExpr>(V))
+ if (ConstantExpr *VCE = dyn_cast<ConstantExpr>(getValue()))
if (VCE->getOpcode() == Instruction::PtrToInt)
if (ConstantExpr *CE = dyn_cast<ConstantExpr>(VCE->getOperand(0)))
if (CE->getOpcode() == Instruction::GetElementPtr &&
}
bool SCEVUnknown::isOffsetOf(const Type *&CTy, Constant *&FieldNo) const {
- if (ConstantExpr *VCE = dyn_cast<ConstantExpr>(V))
+ if (ConstantExpr *VCE = dyn_cast<ConstantExpr>(getValue()))
if (VCE->getOpcode() == Instruction::PtrToInt)
if (ConstantExpr *CE = dyn_cast<ConstantExpr>(VCE->getOperand(0)))
if (CE->getOpcode() == Instruction::GetElementPtr &&
}
// Otherwise just print it normally.
- WriteAsOperand(OS, V, false);
+ WriteAsOperand(OS, getValue(), false);
}
//===----------------------------------------------------------------------===//
/// than the complexity of the RHS. This comparator is used to canonicalize
/// expressions.
class SCEVComplexityCompare {
- LoopInfo *LI;
+ const LoopInfo *LI;
public:
- explicit SCEVComplexityCompare(LoopInfo *li) : LI(li) {}
+ explicit SCEVComplexityCompare(const LoopInfo *li) : LI(li) {}
bool operator()(const SCEV *LHS, const SCEV *RHS) const {
// Fast-path: SCEVs are uniqued so we can do a quick equality check.
return false;
// Primarily, sort the SCEVs by their getSCEVType().
- if (LHS->getSCEVType() != RHS->getSCEVType())
- return LHS->getSCEVType() < RHS->getSCEVType();
+ unsigned LType = LHS->getSCEVType(), RType = RHS->getSCEVType();
+ if (LType != RType)
+ return LType < RType;
// Aside from the getSCEVType() ordering, the particular ordering
// isn't very important except that it's beneficial to be consistent,
// Order pointer values after integer values. This helps SCEVExpander
// form GEPs.
- if (LU->getType()->isPointerTy() && !RU->getType()->isPointerTy())
- return false;
- if (RU->getType()->isPointerTy() && !LU->getType()->isPointerTy())
- return true;
+ bool LIsPointer = LU->getType()->isPointerTy(),
+ RIsPointer = RU->getType()->isPointerTy();
+ if (LIsPointer != RIsPointer)
+ return RIsPointer;
// Compare getValueID values.
- if (LU->getValue()->getValueID() != RU->getValue()->getValueID())
- return LU->getValue()->getValueID() < RU->getValue()->getValueID();
+ unsigned LID = LU->getValue()->getValueID(),
+ RID = RU->getValue()->getValueID();
+ if (LID != RID)
+ return LID < RID;
// Sort arguments by their position.
if (const Argument *LA = dyn_cast<Argument>(LU->getValue())) {
// For instructions, compare their loop depth, and their opcode.
// This is pretty loose.
- if (Instruction *LV = dyn_cast<Instruction>(LU->getValue())) {
- Instruction *RV = cast<Instruction>(RU->getValue());
+ if (const Instruction *LV = dyn_cast<Instruction>(LU->getValue())) {
+ const Instruction *RV = cast<Instruction>(RU->getValue());
// Compare loop depths.
- if (LI->getLoopDepth(LV->getParent()) !=
- LI->getLoopDepth(RV->getParent()))
- return LI->getLoopDepth(LV->getParent()) <
- LI->getLoopDepth(RV->getParent());
-
- // Compare opcodes.
- if (LV->getOpcode() != RV->getOpcode())
- return LV->getOpcode() < RV->getOpcode();
+ unsigned LDepth = LI->getLoopDepth(LV->getParent()),
+ RDepth = LI->getLoopDepth(RV->getParent());
+ if (LDepth != RDepth)
+ return LDepth < RDepth;
// Compare the number of operands.
- if (LV->getNumOperands() != RV->getNumOperands())
- return LV->getNumOperands() < RV->getNumOperands();
+ unsigned LNumOps = LV->getNumOperands(),
+ RNumOps = RV->getNumOperands();
+ if (LNumOps != RNumOps)
+ return LNumOps < RNumOps;
}
return false;
// Compare constant values.
if (const SCEVConstant *LC = dyn_cast<SCEVConstant>(LHS)) {
const SCEVConstant *RC = cast<SCEVConstant>(RHS);
- if (LC->getValue()->getBitWidth() != RC->getValue()->getBitWidth())
- return LC->getValue()->getBitWidth() < RC->getValue()->getBitWidth();
- return LC->getValue()->getValue().ult(RC->getValue()->getValue());
+ const ConstantInt *LCC = LC->getValue();
+ const ConstantInt *RCC = RC->getValue();
+ unsigned LBitWidth = LCC->getBitWidth(), RBitWidth = RCC->getBitWidth();
+ if (LBitWidth != RBitWidth)
+ return LBitWidth < RBitWidth;
+ return LCC->getValue().ult(RCC->getValue());
}
// Compare addrec loop depths.
if (const SCEVAddRecExpr *LA = dyn_cast<SCEVAddRecExpr>(LHS)) {
const SCEVAddRecExpr *RA = cast<SCEVAddRecExpr>(RHS);
- if (LA->getLoop()->getLoopDepth() != RA->getLoop()->getLoopDepth())
- return LA->getLoop()->getLoopDepth() < RA->getLoop()->getLoopDepth();
+ unsigned LDepth = LA->getLoop()->getLoopDepth(),
+ RDepth = RA->getLoop()->getLoopDepth();
+ if (LDepth != RDepth)
+ return LDepth < RDepth;
}
// Lexicographically compare n-ary expressions.
if (const SCEVNAryExpr *LC = dyn_cast<SCEVNAryExpr>(LHS)) {
const SCEVNAryExpr *RC = cast<SCEVNAryExpr>(RHS);
- for (unsigned i = 0, e = LC->getNumOperands(); i != e; ++i) {
- if (i >= RC->getNumOperands())
+ unsigned LNumOps = LC->getNumOperands(), RNumOps = RC->getNumOperands();
+ for (unsigned i = 0; i != LNumOps; ++i) {
+ if (i >= RNumOps)
return false;
- if (operator()(LC->getOperand(i), RC->getOperand(i)))
+ const SCEV *LOp = LC->getOperand(i), *ROp = RC->getOperand(i);
+ if (operator()(LOp, ROp))
return true;
- if (operator()(RC->getOperand(i), LC->getOperand(i)))
+ if (operator()(ROp, LOp))
return false;
}
- return LC->getNumOperands() < RC->getNumOperands();
+ return LNumOps < RNumOps;
}
// Lexicographically compare udiv expressions.
if (const SCEVUDivExpr *LC = dyn_cast<SCEVUDivExpr>(LHS)) {
const SCEVUDivExpr *RC = cast<SCEVUDivExpr>(RHS);
- if (operator()(LC->getLHS(), RC->getLHS()))
+ const SCEV *LL = LC->getLHS(), *LR = LC->getRHS(),
+ *RL = RC->getLHS(), *RR = RC->getRHS();
+ if (operator()(LL, RL))
return true;
- if (operator()(RC->getLHS(), LC->getLHS()))
+ if (operator()(RL, LL))
return false;
- if (operator()(LC->getRHS(), RC->getRHS()))
+ if (operator()(LR, RR))
return true;
- if (operator()(RC->getRHS(), LC->getRHS()))
+ if (operator()(RR, LR))
return false;
return false;
}
CalculationBits);
const SCEV *Dividend = SE.getTruncateOrZeroExtend(It, CalculationTy);
for (unsigned i = 1; i != K; ++i) {
- const SCEV *S = SE.getMinusSCEV(It, SE.getIntegerSCEV(i, It->getType()));
+ const SCEV *S = SE.getMinusSCEV(It, SE.getConstant(It->getType(), i));
Dividend = SE.getMulExpr(Dividend,
SE.getTruncateOrZeroExtend(S, CalculationTy));
}
// Fold if the operand is constant.
if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
return getConstant(
- cast<ConstantInt>(ConstantExpr::getTrunc(SC->getValue(), Ty)));
+ cast<ConstantInt>(ConstantExpr::getTrunc(SC->getValue(),
+ getEffectiveSCEVType(Ty))));
// trunc(trunc(x)) --> trunc(x)
if (const SCEVTruncateExpr *ST = dyn_cast<SCEVTruncateExpr>(Op))
return getAddRecExpr(Operands, AddRec->getLoop());
}
- // The cast wasn't folded; create an explicit cast node.
- // Recompute the insert position, as it may have been invalidated.
- if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
+ // As a special case, fold trunc(undef) to undef. We don't want to
+ // know too much about SCEVUnknowns, but this special case is handy
+ // and harmless.
+ if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Op))
+ if (isa<UndefValue>(U->getValue()))
+ return getSCEV(UndefValue::get(Ty));
+
+ // The cast wasn't folded; create an explicit cast node. We can reuse
+ // the existing insert position since if we get here, we won't have
+ // made any changes which would invalidate it.
SCEV *S = new (SCEVAllocator) SCEVTruncateExpr(ID.Intern(SCEVAllocator),
Op, Ty);
UniqueSCEVs.InsertNode(S, IP);
Ty = getEffectiveSCEVType(Ty);
// Fold if the operand is constant.
- if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op)) {
- const Type *IntTy = getEffectiveSCEVType(Ty);
- Constant *C = ConstantExpr::getZExt(SC->getValue(), IntTy);
- if (IntTy != Ty) C = ConstantExpr::getIntToPtr(C, Ty);
- return getConstant(cast<ConstantInt>(C));
- }
+ if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
+ return getConstant(
+ cast<ConstantInt>(ConstantExpr::getZExt(SC->getValue(),
+ getEffectiveSCEVType(Ty))));
// zext(zext(x)) --> zext(x)
if (const SCEVZeroExtendExpr *SZ = dyn_cast<SCEVZeroExtendExpr>(Op))
} else if (isKnownNegative(Step)) {
const SCEV *N = getConstant(APInt::getMaxValue(BitWidth) -
getSignedRange(Step).getSignedMin());
- if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_UGT, AR, N) &&
- (isLoopEntryGuardedByCond(L, ICmpInst::ICMP_UGT, Start, N) ||
+ if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_UGT, AR, N) ||
+ (isLoopEntryGuardedByCond(L, ICmpInst::ICMP_UGT, Start, N) &&
isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_UGT,
AR->getPostIncExpr(*this), N)))
// Return the expression with the addrec on the outside.
Ty = getEffectiveSCEVType(Ty);
// Fold if the operand is constant.
- if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op)) {
- const Type *IntTy = getEffectiveSCEVType(Ty);
- Constant *C = ConstantExpr::getSExt(SC->getValue(), IntTy);
- if (IntTy != Ty) C = ConstantExpr::getIntToPtr(C, Ty);
- return getConstant(cast<ConstantInt>(C));
- }
+ if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op))
+ return getConstant(
+ cast<ConstantInt>(ConstantExpr::getSExt(SC->getValue(),
+ getEffectiveSCEVType(Ty))));
// sext(sext(x)) --> sext(x)
if (const SCEVSignExtendExpr *SS = dyn_cast<SCEVSignExtendExpr>(Op))
return getAddRecExpr(Ops, AR->getLoop());
}
+ // As a special case, fold anyext(undef) to undef. We don't want to
+ // know too much about SCEVUnknowns, but this special case is handy
+ // and harmless.
+ if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Op))
+ if (isa<UndefValue>(U->getValue()))
+ return getSCEV(UndefValue::get(Ty));
+
// If the expression is obviously signed, use the sext cast value.
if (isa<SCEVSMaxExpr>(Op))
return SExt;
ScalarEvolution &SE) {
bool Interesting = false;
- // Iterate over the add operands.
- for (unsigned i = 0, e = NumOperands; i != e; ++i) {
+ // Iterate over the add operands. They are sorted, with constants first.
+ unsigned i = 0;
+ while (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[i])) {
+ ++i;
+ // Pull a buried constant out to the outside.
+ if (Scale != 1 || AccumulatedConstant != 0 || C->getValue()->isZero())
+ Interesting = true;
+ AccumulatedConstant += Scale * C->getValue()->getValue();
+ }
+
+ // Next comes everything else. We're especially interested in multiplies
+ // here, but they're in the middle, so just visit the rest with one loop.
+ for (; i != NumOperands; ++i) {
const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Ops[i]);
if (Mul && isa<SCEVConstant>(Mul->getOperand(0))) {
APInt NewScale =
Interesting = true;
}
}
- } else if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[i])) {
- // Pull a buried constant out to the outside.
- if (Scale != 1 || AccumulatedConstant != 0 || C->getValue()->isZero())
- Interesting = true;
- AccumulatedConstant += Scale * C->getValue()->getValue();
} else {
// An ordinary operand. Update the map.
std::pair<DenseMap<const SCEV *, APInt>::iterator, bool> Pair =
assert(!Ops.empty() && "Cannot get empty add!");
if (Ops.size() == 1) return Ops[0];
#ifndef NDEBUG
+ const Type *ETy = getEffectiveSCEVType(Ops[0]->getType());
for (unsigned i = 1, e = Ops.size(); i != e; ++i)
- assert(getEffectiveSCEVType(Ops[i]->getType()) ==
- getEffectiveSCEVType(Ops[0]->getType()) &&
+ assert(getEffectiveSCEVType(Ops[i]->getType()) == ETy &&
"SCEVAddExpr operand types don't match!");
#endif
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.
- const SCEV *Two = getIntegerSCEV(2, Ty);
+ const SCEV *Two = getConstant(Ty, 2);
const SCEV *Mul = getMulExpr(Ops[i], Two);
if (Ops.size() == 2)
return Mul;
while (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[Idx])) {
// If we have an add, expand the add operands onto the end of the operands
// list.
- Ops.insert(Ops.end(), Add->op_begin(), Add->op_end());
Ops.erase(Ops.begin()+Idx);
+ Ops.append(Add->op_begin(), Add->op_end());
DeletedAdd = true;
}
Ops.push_back(getMulExpr(getConstant(I->first),
getAddExpr(I->second)));
if (Ops.empty())
- return getIntegerSCEV(0, Ty);
+ return getConstant(Ty, 0);
if (Ops.size() == 1)
return Ops[0];
return getAddExpr(Ops);
MulOps.erase(MulOps.begin()+MulOp);
InnerMul = getMulExpr(MulOps);
}
- const SCEV *One = getIntegerSCEV(1, Ty);
+ const SCEV *One = getConstant(Ty, 1);
const SCEV *AddOne = getAddExpr(InnerMul, One);
const SCEV *OuterMul = getMulExpr(AddOne, Ops[AddOp]);
if (Ops.size() == 2) return OuterMul;
AddRec->op_end());
AddRecOps[0] = getAddExpr(LIOps);
- // It's tempting to propagate NUW/NSW flags here, but nuw/nsw addition
- // is not associative so this isn't necessarily safe.
- const SCEV *NewRec = getAddRecExpr(AddRecOps, AddRecLoop);
+ // Build the new addrec. Propagate the NUW and NSW flags if both the
+ // outer add and the inner addrec are guaranteed to have no overflow.
+ const SCEV *NewRec = getAddRecExpr(AddRecOps, AddRecLoop,
+ HasNUW && AddRec->hasNoUnsignedWrap(),
+ HasNSW && AddRec->hasNoSignedWrap());
// If all of the other operands were loop invariant, we are done.
if (Ops.size() == 1) return NewRec;
AddRec->op_end());
for (unsigned i = 0, e = OtherAddRec->getNumOperands(); i != e; ++i) {
if (i >= NewOps.size()) {
- NewOps.insert(NewOps.end(), OtherAddRec->op_begin()+i,
+ NewOps.append(OtherAddRec->op_begin()+i,
OtherAddRec->op_end());
break;
}
while (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Ops[Idx])) {
// If we have an mul, expand the mul operands onto the end of the operands
// list.
- Ops.insert(Ops.end(), Mul->op_begin(), Mul->op_end());
Ops.erase(Ops.begin()+Idx);
+ Ops.append(Mul->op_begin(), Mul->op_end());
DeletedMul = true;
}
// NLI * LI * {Start,+,Step} --> NLI * {LI*Start,+,LI*Step}
SmallVector<const SCEV *, 4> NewOps;
NewOps.reserve(AddRec->getNumOperands());
- if (LIOps.size() == 1) {
- const SCEV *Scale = LIOps[0];
- for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i)
- NewOps.push_back(getMulExpr(Scale, AddRec->getOperand(i)));
- } else {
- for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i) {
- SmallVector<const SCEV *, 4> MulOps(LIOps.begin(), LIOps.end());
- MulOps.push_back(AddRec->getOperand(i));
- NewOps.push_back(getMulExpr(MulOps));
- }
- }
+ const SCEV *Scale = getMulExpr(LIOps);
+ for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i)
+ NewOps.push_back(getMulExpr(Scale, AddRec->getOperand(i)));
- // It's tempting to propagate the NSW flag here, but nsw multiplication
- // is not associative so this isn't necessarily safe.
+ // Build the new addrec. Propagate the NUW and NSW flags if both the
+ // outer mul and the inner addrec are guaranteed to have no overflow.
const SCEV *NewRec = getAddRecExpr(NewOps, AddRec->getLoop(),
HasNUW && AddRec->hasNoUnsignedWrap(),
- /*HasNSW=*/false);
+ HasNSW && AddRec->hasNoSignedWrap());
// If all of the other operands were loop invariant, we are done.
if (Ops.size() == 1) return NewRec;
Operands.push_back(Start);
if (const SCEVAddRecExpr *StepChrec = dyn_cast<SCEVAddRecExpr>(Step))
if (StepChrec->getLoop() == L) {
- Operands.insert(Operands.end(), StepChrec->op_begin(),
- StepChrec->op_end());
+ Operands.append(StepChrec->op_begin(), StepChrec->op_end());
return getAddRecExpr(Operands, L);
}
if (Idx < Ops.size()) {
bool DeletedSMax = false;
while (const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(Ops[Idx])) {
- Ops.insert(Ops.end(), SMax->op_begin(), SMax->op_end());
Ops.erase(Ops.begin()+Idx);
+ Ops.append(SMax->op_begin(), SMax->op_end());
DeletedSMax = true;
}
if (Idx < Ops.size()) {
bool DeletedUMax = false;
while (const SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(Ops[Idx])) {
- Ops.insert(Ops.end(), UMax->op_begin(), UMax->op_end());
Ops.erase(Ops.begin()+Idx);
+ Ops.append(UMax->op_begin(), UMax->op_end());
DeletedUMax = true;
}
Constant *C = ConstantExpr::getSizeOf(AllocTy);
if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
- C = ConstantFoldConstantExpression(CE, TD);
+ if (Constant *Folded = ConstantFoldConstantExpression(CE, TD))
+ C = Folded;
const Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(AllocTy));
return getTruncateOrZeroExtend(getSCEV(C), Ty);
}
const SCEV *ScalarEvolution::getAlignOfExpr(const Type *AllocTy) {
Constant *C = ConstantExpr::getAlignOf(AllocTy);
if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
- C = ConstantFoldConstantExpression(CE, TD);
+ if (Constant *Folded = ConstantFoldConstantExpression(CE, TD))
+ C = Folded;
const Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(AllocTy));
return getTruncateOrZeroExtend(getSCEV(C), Ty);
}
Constant *C = ConstantExpr::getOffsetOf(STy, FieldNo);
if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
- C = ConstantFoldConstantExpression(CE, TD);
+ if (Constant *Folded = ConstantFoldConstantExpression(CE, TD))
+ C = Folded;
const Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(STy));
return getTruncateOrZeroExtend(getSCEV(C), Ty);
}
Constant *FieldNo) {
Constant *C = ConstantExpr::getOffsetOf(CTy, FieldNo);
if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
- C = ConstantFoldConstantExpression(CE, TD);
+ if (Constant *Folded = ConstantFoldConstantExpression(CE, TD))
+ C = Folded;
const Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(CTy));
return getTruncateOrZeroExtend(getSCEV(C), Ty);
}
ID.AddInteger(scUnknown);
ID.AddPointer(V);
void *IP = 0;
- if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;
- SCEV *S = new (SCEVAllocator) SCEVUnknown(ID.Intern(SCEVAllocator), V);
+ if (SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) {
+ assert(cast<SCEVUnknown>(S)->getValue() == V &&
+ "Stale SCEVUnknown in uniquing map!");
+ return S;
+ }
+ SCEV *S = new (SCEVAllocator) SCEVUnknown(ID.Intern(SCEVAllocator), V, this,
+ FirstUnknown);
+ FirstUnknown = cast<SCEVUnknown>(S);
UniqueSCEVs.InsertNode(S, IP);
return S;
}
return S;
}
-/// getIntegerSCEV - Given a SCEVable type, create a constant for the
-/// specified signed integer value and return a SCEV for the constant.
-const SCEV *ScalarEvolution::getIntegerSCEV(int64_t Val, const Type *Ty) {
- const IntegerType *ITy = cast<IntegerType>(getEffectiveSCEVType(Ty));
- return getConstant(ConstantInt::get(ITy, Val));
-}
-
/// getNegativeSCEV - Return a SCEV corresponding to -V = -1*V
///
const SCEV *ScalarEvolution::getNegativeSCEV(const SCEV *V) {
///
const SCEV *ScalarEvolution::getMinusSCEV(const SCEV *LHS,
const SCEV *RHS) {
+ // Fast path: X - X --> 0.
+ if (LHS == RHS)
+ return getConstant(LHS->getType(), 0);
+
// X - Y --> X + -Y
return getAddExpr(LHS, getNegativeSCEV(RHS));
}
// Push the def-use children onto the Worklist stack.
for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
UI != UE; ++UI)
- Worklist.push_back(cast<Instruction>(UI));
+ Worklist.push_back(cast<Instruction>(*UI));
}
/// ForgetSymbolicValue - This looks up computed SCEV values for all
///
const SCEV *ScalarEvolution::createNodeForGEP(GEPOperator *GEP) {
- bool InBounds = GEP->isInBounds();
+ // Don't blindly transfer the inbounds flag from the GEP instruction to the
+ // Add expression, because the Instruction may be guarded by control flow
+ // and the no-overflow bits may not be valid for the expression in any
+ // context.
+
const Type *IntPtrTy = getEffectiveSCEVType(GEP->getType());
Value *Base = GEP->getOperand(0);
// Don't attempt to analyze GEPs over unsized objects.
if (!cast<PointerType>(Base->getType())->getElementType()->isSized())
return getUnknown(GEP);
- const SCEV *TotalOffset = getIntegerSCEV(0, IntPtrTy);
+ const SCEV *TotalOffset = getConstant(IntPtrTy, 0);
gep_type_iterator GTI = gep_type_begin(GEP);
- for (GetElementPtrInst::op_iterator I = next(GEP->op_begin()),
+ for (GetElementPtrInst::op_iterator I = llvm::next(GEP->op_begin()),
E = GEP->op_end();
I != E; ++I) {
Value *Index = *I;
if (const StructType *STy = dyn_cast<StructType>(*GTI++)) {
// For a struct, add the member offset.
unsigned FieldNo = cast<ConstantInt>(Index)->getZExtValue();
- TotalOffset = getAddExpr(TotalOffset,
- getOffsetOfExpr(STy, FieldNo),
- /*HasNUW=*/false, /*HasNSW=*/InBounds);
+ const SCEV *FieldOffset = getOffsetOfExpr(STy, FieldNo);
+
+ // Add the field offset to the running total offset.
+ TotalOffset = getAddExpr(TotalOffset, FieldOffset);
} else {
// For an array, add the element offset, explicitly scaled.
- const SCEV *LocalOffset = getSCEV(Index);
+ const SCEV *ElementSize = getSizeOfExpr(*GTI);
+ const SCEV *IndexS = getSCEV(Index);
// Getelementptr indices are signed.
- LocalOffset = getTruncateOrSignExtend(LocalOffset, IntPtrTy);
- // Lower "inbounds" GEPs to NSW arithmetic.
- LocalOffset = getMulExpr(LocalOffset, getSizeOfExpr(*GTI),
- /*HasNUW=*/false, /*HasNSW=*/InBounds);
- TotalOffset = getAddExpr(TotalOffset, LocalOffset,
- /*HasNUW=*/false, /*HasNSW=*/InBounds);
+ IndexS = getTruncateOrSignExtend(IndexS, IntPtrTy);
+
+ // Multiply the index by the element size to compute the element offset.
+ const SCEV *LocalOffset = getMulExpr(IndexS, ElementSize);
+
+ // Add the element offset to the running total offset.
+ TotalOffset = getAddExpr(TotalOffset, LocalOffset);
}
}
- return getAddExpr(getSCEV(Base), TotalOffset,
- /*HasNUW=*/false, /*HasNSW=*/InBounds);
+
+ // Get the SCEV for the GEP base.
+ const SCEV *BaseS = getSCEV(Base);
+
+ // Add the total offset from all the GEP indices to the base.
+ return getAddExpr(BaseS, TotalOffset);
}
/// GetMinTrailingZeros - Determine the minimum number of zero bits that S is
if (const SCEVConstant *C = dyn_cast<SCEVConstant>(AddRec->getStart()))
if (!C->getValue()->isZero())
ConservativeResult =
- ConstantRange(C->getValue()->getValue(), APInt(BitWidth, 0));
+ ConservativeResult.intersectWith(
+ ConstantRange(C->getValue()->getValue(), APInt(BitWidth, 0)));
// TODO: non-affine addrec
if (AddRec->isAffine()) {
else if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
return getConstant(CI);
else if (isa<ConstantPointerNull>(V))
- return getIntegerSCEV(0, V->getType());
+ return getConstant(V->getType(), 0);
else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V))
return GA->mayBeOverridden() ? getUnknown(V) : getSCEV(GA->getAliasee());
else
Operator *U = cast<Operator>(V);
switch (Opcode) {
case Instruction::Add:
- // Don't transfer the NSW and NUW bits from the Add instruction to the
- // Add expression, because the Instruction may be guarded by control
- // flow and the no-overflow bits may not be valid for the expression in
- // any context.
return getAddExpr(getSCEV(U->getOperand(0)),
getSCEV(U->getOperand(1)));
case Instruction::Mul:
- // Don't transfer the NSW and NUW bits from the Mul instruction to the
- // Mul expression, as with Add.
return getMulExpr(getSCEV(U->getOperand(0)),
getSCEV(U->getOperand(1)));
case Instruction::UDiv:
return getCouldNotCompute();
else
// The backedge is never taken.
- return getIntegerSCEV(0, CI->getType());
+ return getConstant(CI->getType(), 0);
}
// If it's not an integer or pointer comparison then compute it the hard way.
// constant or derived from a PHI node themselves.
PHINode *PHI = 0;
for (unsigned Op = 0, e = I->getNumOperands(); Op != e; ++Op)
- if (!(isa<Constant>(I->getOperand(Op)) ||
- isa<GlobalValue>(I->getOperand(Op)))) {
+ if (!isa<Constant>(I->getOperand(Op))) {
PHINode *P = getConstantEvolvingPHI(I->getOperand(Op), L);
if (P == 0) return 0; // Not evolving from PHI
if (PHI == 0)
const TargetData *TD) {
if (isa<PHINode>(V)) return PHIVal;
if (Constant *C = dyn_cast<Constant>(V)) return C;
- if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) return GV;
Instruction *I = cast<Instruction>(V);
- std::vector<Constant*> Operands;
- Operands.resize(I->getNumOperands());
+ std::vector<Constant*> Operands(I->getNumOperands());
for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
Operands[i] = EvaluateExpression(I->getOperand(i), PHIVal, TD);
return RetVal = 0; // Must be a constant.
Value *BEValue = PN->getIncomingValue(SecondIsBackedge);
- PHINode *PN2 = getConstantEvolvingPHI(BEValue, L);
- if (PN2 != PN)
+ if (getConstantEvolvingPHI(BEValue, L) != PN &&
+ !isa<Constant>(BEValue))
return RetVal = 0; // Not derived from same PHI.
// Execute the loop symbolically to determine the exit value.
PHINode *PN = getConstantEvolvingPHI(Cond, L);
if (PN == 0) return getCouldNotCompute();
- // Since the loop is canonicalized, the PHI node must have two entries. One
- // entry must be a constant (coming in from outside of the loop), and the
+ // If the loop is canonicalized, the PHI will have exactly two entries.
+ // That's the only form we support here.
+ if (PN->getNumIncomingValues() != 2) return getCouldNotCompute();
+
+ // One entry must be a constant (coming in from outside of the loop), and the
// second must be derived from the same PHI.
bool SecondIsBackedge = L->contains(PN->getIncomingBlock(1));
Constant *StartCST =
if (StartCST == 0) return getCouldNotCompute(); // Must be a constant.
Value *BEValue = PN->getIncomingValue(SecondIsBackedge);
- PHINode *PN2 = getConstantEvolvingPHI(BEValue, L);
- if (PN2 != PN) return getCouldNotCompute(); // Not derived from same PHI.
+ if (getConstantEvolvingPHI(BEValue, L) != PN &&
+ !isa<Constant>(BEValue))
+ return getCouldNotCompute(); // Not derived from same PHI.
// Okay, we find a PHI node that defines the trip count of this loop. Execute
// the loop symbolically to determine when the condition gets a value of
// 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<Constant*> Operands;
- Operands.reserve(I->getNumOperands());
+ SmallVector<Constant *, 4> Operands;
+ bool MadeImprovement = false;
for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
Value *Op = I->getOperand(i);
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 and non-pointer, don't even try to analyze them
- // with scev techniques.
- if (!isSCEVable(Op->getType()))
- return V;
-
- const SCEV *OpV = getSCEVAtScope(Op, L);
- if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(OpV)) {
- Constant *C = SC->getValue();
- if (C->getType() != Op->getType())
- C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
- Op->getType(),
- false),
- C, Op->getType());
- Operands.push_back(C);
- } else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(OpV)) {
- if (Constant *C = dyn_cast<Constant>(SU->getValue())) {
- if (C->getType() != Op->getType())
- C =
- ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
- Op->getType(),
- false),
- C, Op->getType());
- Operands.push_back(C);
- } else
- return V;
- } else {
- return V;
- }
+ continue;
}
+
+ // If any of the operands is non-constant and if they are
+ // non-integer and non-pointer, don't even try to analyze them
+ // with scev techniques.
+ if (!isSCEVable(Op->getType()))
+ return V;
+
+ const SCEV *OrigV = getSCEV(Op);
+ const SCEV *OpV = getSCEVAtScope(OrigV, L);
+ MadeImprovement |= OrigV != OpV;
+
+ Constant *C = 0;
+ if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(OpV))
+ C = SC->getValue();
+ if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(OpV))
+ C = dyn_cast<Constant>(SU->getValue());
+ if (!C) return V;
+ if (C->getType() != Op->getType())
+ C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
+ Op->getType(),
+ false),
+ C, Op->getType());
+ Operands.push_back(C);
}
- Constant *C = 0;
- if (const CmpInst *CI = dyn_cast<CmpInst>(I))
- C = ConstantFoldCompareInstOperands(CI->getPredicate(),
- Operands[0], Operands[1], TD);
- else
- C = ConstantFoldInstOperands(I->getOpcode(), I->getType(),
- &Operands[0], Operands.size(), TD);
- if (C)
+ // Check to see if getSCEVAtScope actually made an improvement.
+ if (MadeImprovement) {
+ Constant *C = 0;
+ if (const CmpInst *CI = dyn_cast<CmpInst>(I))
+ C = ConstantFoldCompareInstOperands(CI->getPredicate(),
+ Operands[0], Operands[1], TD);
+ else
+ C = ConstantFoldInstOperands(I->getOpcode(), I->getType(),
+ &Operands[0], Operands.size(), TD);
+ if (!C) return V;
return getSCEV(C);
+ }
}
}
// If this is a loop recurrence for a loop that does not contain L, then we
// are dealing with the final value computed by the loop.
if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(V)) {
- if (!L || !AddRec->getLoop()->contains(L)) {
+ // First, attempt to evaluate each operand.
+ // Avoid performing the look-up in the common case where the specified
+ // expression has no loop-variant portions.
+ for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i) {
+ const SCEV *OpAtScope = getSCEVAtScope(AddRec->getOperand(i), L);
+ if (OpAtScope == AddRec->getOperand(i))
+ continue;
+
+ // Okay, at least one of these operands is loop variant but might be
+ // foldable. Build a new instance of the folded commutative expression.
+ SmallVector<const SCEV *, 8> NewOps(AddRec->op_begin(),
+ AddRec->op_begin()+i);
+ NewOps.push_back(OpAtScope);
+ for (++i; i != e; ++i)
+ NewOps.push_back(getSCEVAtScope(AddRec->getOperand(i), L));
+
+ AddRec = cast<SCEVAddRecExpr>(getAddRecExpr(NewOps, AddRec->getLoop()));
+ break;
+ }
+
+ // If the scope is outside the addrec's loop, evaluate it by using the
+ // loop exit value of the addrec.
+ if (!AddRec->getLoop()->contains(L)) {
// To evaluate this recurrence, we need to know how many times the AddRec
// loop iterates. Compute this now.
const SCEV *BackedgeTakenCount = getBackedgeTakenCount(AddRec->getLoop());
// Then, evaluate the AddRec.
return AddRec->evaluateAtIteration(BackedgeTakenCount, *this);
}
+
return AddRec;
}
// already. If so, the backedge will execute zero times.
if (const SCEVConstant *C = dyn_cast<SCEVConstant>(V)) {
if (!C->getValue()->isNullValue())
- return getIntegerSCEV(0, C->getType());
+ return getConstant(C->getType(), 0);
return getCouldNotCompute(); // Otherwise it will loop infinitely.
}
return getCouldNotCompute();
}
-/// getLoopPredecessor - If the given loop's header has exactly one unique
-/// predecessor outside the loop, return it. Otherwise return null.
-/// This is less strict that the loop "preheader" concept, which requires
-/// the predecessor to have only one single successor.
-///
-BasicBlock *ScalarEvolution::getLoopPredecessor(const Loop *L) {
- BasicBlock *Header = L->getHeader();
- BasicBlock *Pred = 0;
- for (pred_iterator PI = pred_begin(Header), E = pred_end(Header);
- PI != E; ++PI)
- if (!L->contains(*PI)) {
- if (Pred && Pred != *PI) return 0; // Multiple predecessors.
- Pred = *PI;
- }
- return Pred;
-}
-
/// getPredecessorWithUniqueSuccessorForBB - Return a predecessor of BB
/// (which may not be an immediate predecessor) which has exactly one
/// successor from which BB is reachable, or null if no such block is
// If the header has a unique predecessor outside the loop, it must be
// a block that has exactly one successor that can reach the loop.
if (Loop *L = LI->getLoopFor(BB))
- return std::make_pair(getLoopPredecessor(L), L->getHeader());
+ return std::make_pair(L->getLoopPredecessor(), L->getHeader());
return std::pair<BasicBlock *, BasicBlock *>();
}
if (isLoopEntryGuardedByCond(
AR->getLoop(), Pred, AR->getStart(), RHS) &&
isLoopBackedgeGuardedByCond(
- AR->getLoop(), Pred,
- getAddExpr(AR, AR->getStepRecurrence(*this)), RHS))
+ AR->getLoop(), Pred, AR->getPostIncExpr(*this), RHS))
return true;
if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(RHS))
if (isLoopEntryGuardedByCond(
AR->getLoop(), Pred, LHS, AR->getStart()) &&
isLoopBackedgeGuardedByCond(
- AR->getLoop(), Pred,
- LHS, getAddExpr(AR, AR->getStepRecurrence(*this))))
+ AR->getLoop(), Pred, LHS, AR->getPostIncExpr(*this)))
return true;
// Otherwise see what can be done with known constant ranges.
// as there are predecessors that can be found that have unique successors
// leading to the original header.
for (std::pair<BasicBlock *, BasicBlock *>
- Pair(getLoopPredecessor(L), L->getHeader());
+ Pair(L->getLoopPredecessor(), L->getHeader());
Pair.first;
Pair = getPredecessorWithUniqueSuccessorForBB(Pair.first)) {
"This code doesn't handle negative strides yet!");
const Type *Ty = Start->getType();
- const SCEV *NegOne = getIntegerSCEV(-1, Ty);
+ const SCEV *NegOne = getConstant(Ty, (uint64_t)-1);
const SCEV *Diff = getMinusSCEV(End, Start);
const SCEV *RoundUp = getAddExpr(Step, NegOne);
// behavior, so if wrap does occur, the loop could either terminate or
// loop infinitely, but in either case, the loop is guaranteed to
// iterate at least until the iteration where the wrapping occurs.
- const SCEV *One = getIntegerSCEV(1, Step->getType());
+ const SCEV *One = getConstant(Step->getType(), 1);
if (isSigned) {
APInt Max = APInt::getSignedMaxValue(BitWidth);
if ((Max - getSignedRange(getMinusSCEV(Step, One)).getSignedMax())
// This allows the subsequent ceiling division of (N+(step-1))/step to
// compute the correct value.
const SCEV *StepMinusOne = getMinusSCEV(Step,
- getIntegerSCEV(1, Step->getType()));
+ getConstant(Step->getType(), 1));
MaxEnd = isSigned ?
getSMinExpr(MaxEnd,
getMinusSCEV(getConstant(APInt::getSignedMaxValue(BitWidth)),
if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(getStart()))
if (!SC->getValue()->isZero()) {
SmallVector<const SCEV *, 4> Operands(op_begin(), op_end());
- Operands[0] = SE.getIntegerSCEV(0, SC->getType());
+ Operands[0] = SE.getConstant(SC->getType(), 0);
const SCEV *Shifted = SE.getAddRecExpr(Operands, getLoop());
if (const SCEVAddRecExpr *ShiftedAddRec =
dyn_cast<SCEVAddRecExpr>(Shifted))
// iteration exits.
unsigned BitWidth = SE.getTypeSizeInBits(getType());
if (!Range.contains(APInt(BitWidth, 0)))
- return SE.getIntegerSCEV(0, getType());
+ return SE.getConstant(getType(), 0);
if (isAffine()) {
// If this is an affine expression then we have this situation:
// this now dangles!
}
-void ScalarEvolution::SCEVCallbackVH::allUsesReplacedWith(Value *) {
+void ScalarEvolution::SCEVCallbackVH::allUsesReplacedWith(Value *V) {
assert(SE && "SCEVCallbackVH called with a null ScalarEvolution!");
// Forget all the expressions associated with users of the old value,
// so that future queries will recompute the expressions using the new
// value.
+ Value *Old = getValPtr();
SmallVector<User *, 16> Worklist;
SmallPtrSet<User *, 8> Visited;
- Value *Old = getValPtr();
- bool DeleteOld = false;
for (Value::use_iterator UI = Old->use_begin(), UE = Old->use_end();
UI != UE; ++UI)
Worklist.push_back(*UI);
User *U = Worklist.pop_back_val();
// Deleting the Old value will cause this to dangle. Postpone
// that until everything else is done.
- if (U == Old) {
- DeleteOld = true;
+ if (U == Old)
continue;
- }
if (!Visited.insert(U))
continue;
if (PHINode *PN = dyn_cast<PHINode>(U))
UI != UE; ++UI)
Worklist.push_back(*UI);
}
- // Delete the Old value if it (indirectly) references itself.
- if (DeleteOld) {
- if (PHINode *PN = dyn_cast<PHINode>(Old))
- SE->ConstantEvolutionLoopExitValue.erase(PN);
- SE->Scalars.erase(Old);
- // this now dangles!
- }
- // this may dangle!
+ // Delete the Old value.
+ if (PHINode *PN = dyn_cast<PHINode>(Old))
+ SE->ConstantEvolutionLoopExitValue.erase(PN);
+ SE->Scalars.erase(Old);
+ // this now dangles!
}
ScalarEvolution::SCEVCallbackVH::SCEVCallbackVH(Value *V, ScalarEvolution *se)
//===----------------------------------------------------------------------===//
ScalarEvolution::ScalarEvolution()
- : FunctionPass(&ID) {
+ : FunctionPass(&ID), FirstUnknown(0) {
}
bool ScalarEvolution::runOnFunction(Function &F) {
}
void ScalarEvolution::releaseMemory() {
+ // Iterate through all the SCEVUnknown instances and call their
+ // destructors, so that they release their references to their values.
+ for (SCEVUnknown *U = FirstUnknown; U; U = U->Next)
+ U->~SCEVUnknown();
+ FirstUnknown = 0;
+
Scalars.clear();
BackedgeTakenCounts.clear();
ConstantEvolutionLoopExitValue.clear();
projects / oota-llvm.git / blobdiff