return MulExt.slt(Min) || MulExt.sgt(Max);
}
+/// \brief A helper routine of InstCombiner::visitMul().
+///
+/// If C is a vector of known powers of 2, then this function returns
+/// a new vector obtained from C replacing each element with its logBase2.
+/// Return a null pointer otherwise.
+static Constant *getLogBase2Vector(ConstantDataVector *CV) {
+ const APInt *IVal;
+ SmallVector<Constant *, 4> Elts;
+
+ for (unsigned I = 0, E = CV->getNumElements(); I != E; ++I) {
+ Constant *Elt = CV->getElementAsConstant(I);
+ if (!match(Elt, m_APInt(IVal)) || !IVal->isPowerOf2())
+ return 0;
+ Elts.push_back(ConstantInt::get(Elt->getType(), IVal->logBase2()));
+ }
+
+ return ConstantVector::get(Elts);
+}
+
Instruction *InstCombiner::visitMul(BinaryOperator &I) {
bool Changed = SimplifyAssociativeOrCommutative(I);
Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
if (match(Op1, m_AllOnes())) // X * -1 == 0 - X
return BinaryOperator::CreateNeg(Op0, I.getName());
- if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
-
- // ((X << C1)*C2) == (X * (C2 << C1))
- if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
- if (SI->getOpcode() == Instruction::Shl)
- if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
- return BinaryOperator::CreateMul(SI->getOperand(0),
- ConstantExpr::getShl(CI, ShOp));
-
- const APInt &Val = CI->getValue();
- if (Val.isPowerOf2()) { // Replace X*(2^C) with X << C
- Constant *NewCst = ConstantInt::get(Op0->getType(), Val.logBase2());
- BinaryOperator *Shl = BinaryOperator::CreateShl(Op0, NewCst);
- if (I.hasNoSignedWrap()) Shl->setHasNoSignedWrap();
- if (I.hasNoUnsignedWrap()) Shl->setHasNoUnsignedWrap();
- return Shl;
+ // Also allow combining multiply instructions on vectors.
+ {
+ Value *NewOp;
+ Constant *C1, *C2;
+ const APInt *IVal;
+ if (match(&I, m_Mul(m_Shl(m_Value(NewOp), m_Constant(C2)),
+ m_Constant(C1))) &&
+ match(C1, m_APInt(IVal)))
+ // ((X << C1)*C2) == (X * (C2 << C1))
+ return BinaryOperator::CreateMul(NewOp, ConstantExpr::getShl(C1, C2));
+
+ if (match(&I, m_Mul(m_Value(NewOp), m_Constant(C1)))) {
+ Constant *NewCst = 0;
+ if (match(C1, m_APInt(IVal)) && IVal->isPowerOf2())
+ // Replace X*(2^C) with X << C, where C is either a scalar or a splat.
+ NewCst = ConstantInt::get(NewOp->getType(), IVal->logBase2());
+ else if (ConstantDataVector *CV = dyn_cast<ConstantDataVector>(C1))
+ // Replace X*(2^C) with X << C, where C is a vector of known
+ // constant powers of 2.
+ NewCst = getLogBase2Vector(CV);
+
+ if (NewCst) {
+ BinaryOperator *Shl = BinaryOperator::CreateShl(NewOp, NewCst);
+ if (I.hasNoSignedWrap()) Shl->setHasNoSignedWrap();
+ if (I.hasNoUnsignedWrap()) Shl->setHasNoUnsignedWrap();
+ return Shl;
+ }
}
+ }
+ if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
// Canonicalize (X+C1)*CI -> X*CI+C1*CI.
{ Value *X; ConstantInt *C1;
if (Op0->hasOneUse() &&
if (C0 && C1)
return false;
- return (C0 && C0->getValueAPF().isNormal()) ||
- (C1 && C1->getValueAPF().isNormal());
+ return (C0 && C0->getValueAPF().isFiniteNonZero()) ||
+ (C1 && C1->getValueAPF().isFiniteNonZero());
}
static bool isNormalFp(const ConstantFP *C) {
const APFloat &Flt = C->getValueAPF();
- return Flt.isNormal() && !Flt.isDenormal();
+ return Flt.isNormal();
}
/// foldFMulConst() is a helper routine of InstCombiner::visitFMul().
} else {
if (C0) {
// (C0 / X) * C => (C0 * C) / X
- ConstantFP *F = cast<ConstantFP>(ConstantExpr::getFMul(C0, C));
- if (isNormalFp(F))
- R = BinaryOperator::CreateFDiv(F, Opnd1);
+ if (FMulOrDiv->hasOneUse()) {
+ // It would otherwise introduce another div.
+ ConstantFP *F = cast<ConstantFP>(ConstantExpr::getFMul(C0, C));
+ if (isNormalFp(F))
+ R = BinaryOperator::CreateFDiv(F, Opnd1);
+ }
} else {
// (X / C1) * C => X * (C/C1) if C/C1 is not a denormal
ConstantFP *F = cast<ConstantFP>(ConstantExpr::getFDiv(C, C1));
return NV;
ConstantFP *C = dyn_cast<ConstantFP>(Op1);
- if (C && AllowReassociate && C->getValueAPF().isNormal()) {
+ if (C && AllowReassociate && C->getValueAPF().isFiniteNonZero()) {
// Let MDC denote an expression in one of these forms:
// X * C, C/X, X/C, where C is a constant.
//
Swap = true;
}
- if (C1 && C1->getValueAPF().isNormal() &&
+ if (C1 && C1->getValueAPF().isFiniteNonZero() &&
isFMulOrFDivWithConstant(Opnd0)) {
Value *M1 = ConstantExpr::getFMul(C1, C);
Value *M0 = isNormalFp(cast<ConstantFP>(M1)) ?
if (Swap && FAddSub->getOpcode() == Instruction::FSub)
std::swap(M0, M1);
- Value *R = (FAddSub->getOpcode() == Instruction::FAdd) ?
- BinaryOperator::CreateFAdd(M0, M1) :
- BinaryOperator::CreateFSub(M0, M1);
- Instruction *RI = cast<Instruction>(R);
+ Instruction *RI = (FAddSub->getOpcode() == Instruction::FAdd)
+ ? BinaryOperator::CreateFAdd(M0, M1)
+ : BinaryOperator::CreateFSub(M0, M1);
RI->copyFastMathFlags(&I);
return RI;
}
}
// if pattern detected emit alternate sequence
if (OpX && OpY) {
+ BuilderTy::FastMathFlagGuard Guard(*Builder);
+ Builder->SetFastMathFlags(Log2->getFastMathFlags());
Log2->setArgOperand(0, OpY);
Value *FMulVal = Builder->CreateFMul(OpX, Log2);
- Instruction *FMul = cast<Instruction>(FMulVal);
- FMul->copyFastMathFlags(Log2);
- Instruction *FSub = BinaryOperator::CreateFSub(FMulVal, OpX);
- FSub->copyFastMathFlags(Log2);
- return FSub;
+ Value *FSub = Builder->CreateFSub(FMulVal, OpX);
+ FSub->takeName(&I);
+ return ReplaceInstUsesWith(I, FSub);
}
}
for (int i = 0; i < 2; i++) {
bool IgnoreZeroSign = I.hasNoSignedZeros();
if (BinaryOperator::isFNeg(Opnd0, IgnoreZeroSign)) {
+ BuilderTy::FastMathFlagGuard Guard(*Builder);
+ Builder->SetFastMathFlags(I.getFastMathFlags());
+
Value *N0 = dyn_castFNegVal(Opnd0, IgnoreZeroSign);
Value *N1 = dyn_castFNegVal(Opnd1, IgnoreZeroSign);
if (Opnd0->hasOneUse()) {
// -X * Y => -(X*Y) (Promote negation as high as possible)
Value *T = Builder->CreateFMul(N0, Opnd1);
- cast<Instruction>(T)->setDebugLoc(I.getDebugLoc());
- Instruction *Neg = BinaryOperator::CreateFNeg(T);
- if (I.getFastMathFlags().any()) {
- cast<Instruction>(T)->copyFastMathFlags(&I);
- Neg->copyFastMathFlags(&I);
- }
- return Neg;
+ Value *Neg = Builder->CreateFNeg(T);
+ Neg->takeName(&I);
+ return ReplaceInstUsesWith(I, Neg);
}
}
Y = Opnd0_0;
if (Y) {
- Instruction *T = cast<Instruction>(Builder->CreateFMul(Opnd1, Opnd1));
- T->copyFastMathFlags(&I);
- T->setDebugLoc(I.getDebugLoc());
+ BuilderTy::FastMathFlagGuard Guard(*Builder);
+ Builder->SetFastMathFlags(I.getFastMathFlags());
+ Value *T = Builder->CreateFMul(Opnd1, Opnd1);
- Instruction *R = BinaryOperator::CreateFMul(T, Y);
- R->copyFastMathFlags(&I);
- return R;
+ Value *R = Builder->CreateFMul(T, Y);
+ R->takeName(&I);
+ return ReplaceInstUsesWith(I, R);
}
}
}
if (I.hasNoNaNs() && I.hasNoInfs() && I.hasNoSignedZeros()) {
Value *LHS = Op0, *RHS = Op1;
Value *B, *C;
- if (!match(RHS, m_UIToFp(m_Value(C))))
+ if (!match(RHS, m_UIToFP(m_Value(C))))
std::swap(LHS, RHS);
- if (match(RHS, m_UIToFp(m_Value(C))) && C->getType()->isIntegerTy(1)) {
+ if (match(RHS, m_UIToFP(m_Value(C))) && C->getType()->isIntegerTy(1)) {
B = LHS;
Value *Zero = ConstantFP::getNegativeZero(B->getType());
return SelectInst::Create(C, B, Zero);
if (I.hasNoNaNs() && I.hasNoInfs() && I.hasNoSignedZeros()) {
Value *LHS = Op0, *RHS = Op1;
Value *A, *C;
- if (!match(RHS, m_FSub(m_FPOne(), m_UIToFp(m_Value(C)))))
+ if (!match(RHS, m_FSub(m_FPOne(), m_UIToFP(m_Value(C)))))
std::swap(LHS, RHS);
- if (match(RHS, m_FSub(m_FPOne(), m_UIToFp(m_Value(C)))) &&
+ if (match(RHS, m_FSub(m_FPOne(), m_UIToFP(m_Value(C)))) &&
C->getType()->isIntegerTy(1)) {
A = LHS;
Value *Zero = ConstantFP::getNegativeZero(A->getType());
*I = SI->getOperand(NonNullOperand);
Worklist.Add(BBI);
} else if (*I == SelectCond) {
- *I = NonNullOperand == 1 ? ConstantInt::getTrue(BBI->getContext()) :
- ConstantInt::getFalse(BBI->getContext());
+ *I = Builder->getInt1(NonNullOperand == 1);
Worklist.Add(BBI);
}
}
return 0;
}
-Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
- Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
+namespace {
+const unsigned MaxDepth = 6;
+typedef Instruction *(*FoldUDivOperandCb)(Value *Op0, Value *Op1,
+ const BinaryOperator &I,
+ InstCombiner &IC);
+
+/// \brief Used to maintain state for visitUDivOperand().
+struct UDivFoldAction {
+ FoldUDivOperandCb FoldAction; ///< Informs visitUDiv() how to fold this
+ ///< operand. This can be zero if this action
+ ///< joins two actions together.
+
+ Value *OperandToFold; ///< Which operand to fold.
+ union {
+ Instruction *FoldResult; ///< The instruction returned when FoldAction is
+ ///< invoked.
+
+ size_t SelectLHSIdx; ///< Stores the LHS action index if this action
+ ///< joins two actions together.
+ };
+
+ UDivFoldAction(FoldUDivOperandCb FA, Value *InputOperand)
+ : FoldAction(FA), OperandToFold(InputOperand), FoldResult(0) {}
+ UDivFoldAction(FoldUDivOperandCb FA, Value *InputOperand, size_t SLHS)
+ : FoldAction(FA), OperandToFold(InputOperand), SelectLHSIdx(SLHS) {}
+};
+}
- if (Value *V = SimplifyUDivInst(Op0, Op1, TD))
- return ReplaceInstUsesWith(I, V);
+// X udiv 2^C -> X >> C
+static Instruction *foldUDivPow2Cst(Value *Op0, Value *Op1,
+ const BinaryOperator &I, InstCombiner &IC) {
+ const APInt &C = cast<Constant>(Op1)->getUniqueInteger();
+ BinaryOperator *LShr = BinaryOperator::CreateLShr(
+ Op0, ConstantInt::get(Op0->getType(), C.logBase2()));
+ if (I.isExact()) LShr->setIsExact();
+ return LShr;
+}
- // Handle the integer div common cases
- if (Instruction *Common = commonIDivTransforms(I))
- return Common;
+// X udiv C, where C >= signbit
+static Instruction *foldUDivNegCst(Value *Op0, Value *Op1,
+ const BinaryOperator &I, InstCombiner &IC) {
+ Value *ICI = IC.Builder->CreateICmpULT(Op0, cast<ConstantInt>(Op1));
- {
- // X udiv 2^C -> X >> C
- // Check to see if this is an unsigned division with an exact power of 2,
- // if so, convert to a right shift.
- const APInt *C;
- if (match(Op1, m_Power2(C))) {
- BinaryOperator *LShr =
- BinaryOperator::CreateLShr(Op0,
- ConstantInt::get(Op0->getType(),
- C->logBase2()));
- if (I.isExact()) LShr->setIsExact();
- return LShr;
- }
+ return SelectInst::Create(ICI, Constant::getNullValue(I.getType()),
+ ConstantInt::get(I.getType(), 1));
+}
+
+// X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
+static Instruction *foldUDivShl(Value *Op0, Value *Op1, const BinaryOperator &I,
+ InstCombiner &IC) {
+ Instruction *ShiftLeft = cast<Instruction>(Op1);
+ if (isa<ZExtInst>(ShiftLeft))
+ ShiftLeft = cast<Instruction>(ShiftLeft->getOperand(0));
+
+ const APInt &CI =
+ cast<Constant>(ShiftLeft->getOperand(0))->getUniqueInteger();
+ Value *N = ShiftLeft->getOperand(1);
+ if (CI != 1)
+ N = IC.Builder->CreateAdd(N, ConstantInt::get(N->getType(), CI.logBase2()));
+ if (ZExtInst *Z = dyn_cast<ZExtInst>(Op1))
+ N = IC.Builder->CreateZExt(N, Z->getDestTy());
+ BinaryOperator *LShr = BinaryOperator::CreateLShr(Op0, N);
+ if (I.isExact()) LShr->setIsExact();
+ return LShr;
+}
+
+// \brief Recursively visits the possible right hand operands of a udiv
+// instruction, seeing through select instructions, to determine if we can
+// replace the udiv with something simpler. If we find that an operand is not
+// able to simplify the udiv, we abort the entire transformation.
+static size_t visitUDivOperand(Value *Op0, Value *Op1, const BinaryOperator &I,
+ SmallVectorImpl<UDivFoldAction> &Actions,
+ unsigned Depth = 0) {
+ // Check to see if this is an unsigned division with an exact power of 2,
+ // if so, convert to a right shift.
+ if (match(Op1, m_Power2())) {
+ Actions.push_back(UDivFoldAction(foldUDivPow2Cst, Op1));
+ return Actions.size();
}
- if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
+ if (ConstantInt *C = dyn_cast<ConstantInt>(Op1))
// X udiv C, where C >= signbit
if (C->getValue().isNegative()) {
- Value *IC = Builder->CreateICmpULT(Op0, C);
- return SelectInst::Create(IC, Constant::getNullValue(I.getType()),
- ConstantInt::get(I.getType(), 1));
+ Actions.push_back(UDivFoldAction(foldUDivNegCst, C));
+ return Actions.size();
}
+
+ // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
+ if (match(Op1, m_Shl(m_Power2(), m_Value())) ||
+ match(Op1, m_ZExt(m_Shl(m_Power2(), m_Value())))) {
+ Actions.push_back(UDivFoldAction(foldUDivShl, Op1));
+ return Actions.size();
}
+ // The remaining tests are all recursive, so bail out if we hit the limit.
+ if (Depth++ == MaxDepth)
+ return 0;
+
+ if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
+ if (size_t LHSIdx = visitUDivOperand(Op0, SI->getOperand(1), I, Actions))
+ if (visitUDivOperand(Op0, SI->getOperand(2), I, Actions)) {
+ Actions.push_back(UDivFoldAction((FoldUDivOperandCb)0, Op1, LHSIdx-1));
+ return Actions.size();
+ }
+
+ return 0;
+}
+
+Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
+ Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
+
+ if (Value *V = SimplifyUDivInst(Op0, Op1, TD))
+ return ReplaceInstUsesWith(I, V);
+
+ // Handle the integer div common cases
+ if (Instruction *Common = commonIDivTransforms(I))
+ return Common;
+
// (x lshr C1) udiv C2 --> x udiv (C2 << C1)
if (ConstantInt *C2 = dyn_cast<ConstantInt>(Op1)) {
Value *X;
}
}
- // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
- { const APInt *CI; Value *N;
- if (match(Op1, m_Shl(m_Power2(CI), m_Value(N))) ||
- match(Op1, m_ZExt(m_Shl(m_Power2(CI), m_Value(N))))) {
- if (*CI != 1)
- N = Builder->CreateAdd(N,
- ConstantInt::get(N->getType(), CI->logBase2()));
- if (ZExtInst *Z = dyn_cast<ZExtInst>(Op1))
- N = Builder->CreateZExt(N, Z->getDestTy());
- if (I.isExact())
- return BinaryOperator::CreateExactLShr(Op0, N);
- return BinaryOperator::CreateLShr(Op0, N);
- }
- }
-
- // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
- // where C1&C2 are powers of two.
- { Value *Cond; const APInt *C1, *C2;
- if (match(Op1, m_Select(m_Value(Cond), m_Power2(C1), m_Power2(C2)))) {
- // Construct the "on true" case of the select
- Value *TSI = Builder->CreateLShr(Op0, C1->logBase2(), Op1->getName()+".t",
- I.isExact());
-
- // Construct the "on false" case of the select
- Value *FSI = Builder->CreateLShr(Op0, C2->logBase2(), Op1->getName()+".f",
- I.isExact());
-
- // construct the select instruction and return it.
- return SelectInst::Create(Cond, TSI, FSI);
- }
- }
-
// (zext A) udiv (zext B) --> zext (A udiv B)
if (ZExtInst *ZOp0 = dyn_cast<ZExtInst>(Op0))
if (Value *ZOp1 = dyn_castZExtVal(Op1, ZOp0->getSrcTy()))
I.isExact()),
I.getType());
+ // (LHS udiv (select (select (...)))) -> (LHS >> (select (select (...))))
+ SmallVector<UDivFoldAction, 6> UDivActions;
+ if (visitUDivOperand(Op0, Op1, I, UDivActions))
+ for (unsigned i = 0, e = UDivActions.size(); i != e; ++i) {
+ FoldUDivOperandCb Action = UDivActions[i].FoldAction;
+ Value *ActionOp1 = UDivActions[i].OperandToFold;
+ Instruction *Inst;
+ if (Action)
+ Inst = Action(Op0, ActionOp1, I, *this);
+ else {
+ // This action joins two actions together. The RHS of this action is
+ // simply the last action we processed, we saved the LHS action index in
+ // the joining action.
+ size_t SelectRHSIdx = i - 1;
+ Value *SelectRHS = UDivActions[SelectRHSIdx].FoldResult;
+ size_t SelectLHSIdx = UDivActions[i].SelectLHSIdx;
+ Value *SelectLHS = UDivActions[SelectLHSIdx].FoldResult;
+ Inst = SelectInst::Create(cast<SelectInst>(ActionOp1)->getCondition(),
+ SelectLHS, SelectRHS);
+ }
+
+ // If this is the last action to process, return it to the InstCombiner.
+ // Otherwise, we insert it before the UDiv and record it so that we may
+ // use it as part of a joining action (i.e., a SelectInst).
+ if (e - i != 1) {
+ Inst->insertBefore(&I);
+ UDivActions[i].FoldResult = Inst;
+ } else
+ return Inst;
+ }
+
return 0;
}
/// FP value and:
/// 1) 1/C is exact, or
/// 2) reciprocal is allowed.
-/// If the convertion was successful, the simplified expression "X * 1/C" is
+/// If the conversion was successful, the simplified expression "X * 1/C" is
/// returned; otherwise, NULL is returned.
///
static Instruction *CvtFDivConstToReciprocal(Value *Dividend,
APFloat Reciprocal(FpVal.getSemantics());
bool Cvt = FpVal.getExactInverse(&Reciprocal);
- if (!Cvt && AllowReciprocal && FpVal.isNormal()) {
+ if (!Cvt && AllowReciprocal && FpVal.isFiniteNonZero()) {
Reciprocal = APFloat(FpVal.getSemantics(), 1.0f);
(void)Reciprocal.divide(FpVal, APFloat::rmNearestTiesToEven);
Cvt = !Reciprocal.isDenormal();
if (Value *V = SimplifyFDivInst(Op0, Op1, TD))
return ReplaceInstUsesWith(I, V);
+ if (isa<Constant>(Op0))
+ if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
+ if (Instruction *R = FoldOpIntoSelect(I, SI))
+ return R;
+
bool AllowReassociate = I.hasUnsafeAlgebra();
bool AllowReciprocal = I.hasAllowReciprocal();
if (ConstantFP *Op1C = dyn_cast<ConstantFP>(Op1)) {
+ if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
+ if (Instruction *R = FoldOpIntoSelect(I, SI))
+ return R;
+
if (AllowReassociate) {
ConstantFP *C1 = 0;
ConstantFP *C2 = Op1C;
//
Constant *C = ConstantExpr::getFDiv(C1, C2);
const APFloat &F = cast<ConstantFP>(C)->getValueAPF();
- if (F.isNormal() && !F.isDenormal())
+ if (F.isNormal())
Res = BinaryOperator::CreateFMul(X, C);
} else if (match(Op0, m_FDiv(m_Value(X), m_ConstantFP(C1)))) {
// (X/C1)/C2 => X /(C2*C1) [=> X * 1/(C2*C1) if reciprocal is allowed]
//
Constant *C = ConstantExpr::getFMul(C1, C2);
const APFloat &F = cast<ConstantFP>(C)->getValueAPF();
- if (F.isNormal() && !F.isDenormal()) {
+ if (F.isNormal()) {
Res = CvtFDivConstToReciprocal(X, cast<ConstantFP>(C),
AllowReciprocal);
if (!Res)
if (Fold) {
const APFloat &FoldC = cast<ConstantFP>(Fold)->getValueAPF();
- if (FoldC.isNormal() && !FoldC.isDenormal()) {
+ if (FoldC.isNormal()) {
Instruction *R = CreateDiv ?
BinaryOperator::CreateFDiv(Fold, X) :
BinaryOperator::CreateFMul(X, Fold);
return BinaryOperator::CreateAnd(Op0, Add);
}
+ // 1 urem X -> zext(X != 1)
+ if (match(Op0, m_One())) {
+ Value *Cmp = Builder->CreateICmpNE(Op1, Op0);
+ Value *Ext = Builder->CreateZExt(Cmp, I.getType());
+ return ReplaceInstUsesWith(I, Ext);
+ }
+
return 0;
}