/// Rebuild a new instruction just like 'I' but with the new operands given.
/// In the event of type mismatch, the type of the operands is correct.
static Value *BuildNew(Instruction *I, ArrayRef<Value*> NewOps) {
// We don't want to use the IRBuilder here because we want the replacement
// instructions to appear next to 'I', not the builder's insertion point.
switch (I->getOpcode()) {
case Instruction::Add:
case Instruction::FAdd:
case Instruction::Sub:
case Instruction::FSub:
case Instruction::Mul:
case Instruction::FMul:
case Instruction::UDiv:
case Instruction::SDiv:
case Instruction::FDiv:
case Instruction::URem:
case Instruction::SRem:
case Instruction::FRem:
case Instruction::Shl:
case Instruction::LShr:
case Instruction::AShr:
case Instruction::And:
case Instruction::Or:
case Instruction::Xor: {
BinaryOperator *BO = cast<BinaryOperator>(I);
assert(NewOps.size() == 2 && "binary operator with #ops != 2");
BinaryOperator *New =
BinaryOperator::Create(cast<BinaryOperator>(I)->getOpcode(),
NewOps[0], NewOps[1], "", BO);
if (isa<OverflowingBinaryOperator>(BO)) {
New->setHasNoUnsignedWrap(BO->hasNoUnsignedWrap());
New->setHasNoSignedWrap(BO->hasNoSignedWrap());
}
if (isa<PossiblyExactOperator>(BO)) {
New->setIsExact(BO->isExact());
}
return New;
}
case Instruction::ICmp:
assert(NewOps.size() == 2 && "icmp with #ops != 2");
return new ICmpInst(I, cast<ICmpInst>(I)->getPredicate(),
NewOps[0], NewOps[1]);
case Instruction::FCmp:
assert(NewOps.size() == 2 && "fcmp with #ops != 2");
return new FCmpInst(I, cast<FCmpInst>(I)->getPredicate(),
NewOps[0], NewOps[1]);
case Instruction::Trunc:
case Instruction::ZExt:
case Instruction::SExt:
case Instruction::FPToUI:
case Instruction::FPToSI:
case Instruction::UIToFP:
case Instruction::SIToFP:
case Instruction::FPTrunc:
case Instruction::FPExt: {
// It's possible that the mask has a different number of elements from
// the original cast. We recompute the destination type to match the mask.
Type *DestTy =
VectorType::get(I->getType()->getScalarType(),
NewOps[0]->getType()->getVectorNumElements());
assert(NewOps.size() == 1 && "cast with #ops != 1");
return CastInst::Create(cast<CastInst>(I)->getOpcode(), NewOps[0], DestTy,
"", I);
}
case Instruction::GetElementPtr: {
Value *Ptr = NewOps[0];
ArrayRef<Value*> Idx = NewOps.slice(1);
GetElementPtrInst *GEP = GetElementPtrInst::Create(Ptr, Idx, "", I);
GEP->setIsInBounds(cast<GetElementPtrInst>(I)->isInBounds());
return GEP;
}
}
llvm_unreachable("failed to rebuild vector instructions");
}
Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, Constant *Op1,
BinaryOperator &I) {
bool isLeftShift = I.getOpcode() == Instruction::Shl;
ConstantInt *COp1 = nullptr;
if (ConstantDataVector *CV = dyn_cast<ConstantDataVector>(Op1))
COp1 = dyn_cast_or_null<ConstantInt>(CV->getSplatValue());
else if (ConstantVector *CV = dyn_cast<ConstantVector>(Op1))
COp1 = dyn_cast_or_null<ConstantInt>(CV->getSplatValue());
else
COp1 = dyn_cast<ConstantInt>(Op1);
if (!COp1)
return nullptr;
// See if we can propagate this shift into the input, this covers the trivial
// cast of lshr(shl(x,c1),c2) as well as other more complex cases.
if (I.getOpcode() != Instruction::AShr &&
CanEvaluateShifted(Op0, COp1->getZExtValue(), isLeftShift, *this)) {
DEBUG(dbgs() << "ICE: GetShiftedValue propagating shift through expression"
" to eliminate shift:\n IN: " << *Op0 << "\n SH: " << I <<"\n");
return ReplaceInstUsesWith(I,
GetShiftedValue(Op0, COp1->getZExtValue(), isLeftShift, *this));
}
// See if we can simplify any instructions used by the instruction whose sole
// purpose is to compute bits we don't care about.
uint32_t TypeBits = Op0->getType()->getScalarSizeInBits();
assert(!COp1->uge(TypeBits) &&
"Shift over the type width should have been removed already");
// ((X*C1) << C2) == (X * (C1 << C2))
if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
if (BO->getOpcode() == Instruction::Mul && isLeftShift)
if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
return BinaryOperator::CreateMul(BO->getOperand(0),
ConstantExpr::getShl(BOOp, Op1));
// Try to fold constant and into select arguments.
if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
if (Instruction *R = FoldOpIntoSelect(I, SI))
return R;
if (isa<PHINode>(Op0))
if (Instruction *NV = FoldOpIntoPhi(I))
return NV;
// Fold shift2(trunc(shift1(x,c1)), c2) -> trunc(shift2(shift1(x,c1),c2))
if (TruncInst *TI = dyn_cast<TruncInst>(Op0)) {
Instruction *TrOp = dyn_cast<Instruction>(TI->getOperand(0));
// If 'shift2' is an ashr, we would have to get the sign bit into a funny
// place. Don't try to do this transformation in this case. Also, we
// require that the input operand is a shift-by-constant so that we have
// confidence that the shifts will get folded together. We could do this
// xform in more cases, but it is unlikely to be profitable.
if (TrOp && I.isLogicalShift() && TrOp->isShift() &&
isa<ConstantInt>(TrOp->getOperand(1))) {
// Okay, we'll do this xform. Make the shift of shift.
Constant *ShAmt = ConstantExpr::getZExt(COp1, TrOp->getType());
// (shift2 (shift1 & 0x00FF), c2)
Value *NSh = Builder->CreateBinOp(I.getOpcode(), TrOp, ShAmt,I.getName());
// For logical shifts, the truncation has the effect of making the high
// part of the register be zeros. Emulate this by inserting an AND to
// clear the top bits as needed. This 'and' will usually be zapped by
// other xforms later if dead.
unsigned SrcSize = TrOp->getType()->getScalarSizeInBits();
unsigned DstSize = TI->getType()->getScalarSizeInBits();
APInt MaskV(APInt::getLowBitsSet(SrcSize, DstSize));
// The mask we constructed says what the trunc would do if occurring
// between the shifts. We want to know the effect *after* the second
// shift. We know that it is a logical shift by a constant, so adjust the
// mask as appropriate.
if (I.getOpcode() == Instruction::Shl)
MaskV <<= COp1->getZExtValue();
else {
assert(I.getOpcode() == Instruction::LShr && "Unknown logical shift");
MaskV = MaskV.lshr(COp1->getZExtValue());
}
// shift1 & 0x00FF
Value *And = Builder->CreateAnd(NSh,
ConstantInt::get(I.getContext(), MaskV),
TI->getName());
// Return the value truncated to the interesting size.
return new TruncInst(And, I.getType());
}
}
if (Op0->hasOneUse()) {
if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
// Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
Value *V1, *V2;
ConstantInt *CC;
switch (Op0BO->getOpcode()) {
default: break;
case Instruction::Add:
case Instruction::And:
case Instruction::Or:
case Instruction::Xor: {
// These operators commute.
// Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
match(Op0BO->getOperand(1), m_Shr(m_Value(V1),
m_Specific(Op1)))) {
Value *YS = // (Y << C)
Builder->CreateShl(Op0BO->getOperand(0), Op1, Op0BO->getName());
// (X + (Y << C))
Value *X = Builder->CreateBinOp(Op0BO->getOpcode(), YS, V1,
Op0BO->getOperand(1)->getName());
uint32_t Op1Val = COp1->getLimitedValue(TypeBits);
APInt Bits = APInt::getHighBitsSet(TypeBits, TypeBits - Op1Val);
Constant *Mask = ConstantInt::get(I.getContext(), Bits);
if (VectorType *VT = dyn_cast<VectorType>(X->getType()))
Mask = ConstantVector::getSplat(VT->getNumElements(), Mask);
return BinaryOperator::CreateAnd(X, Mask);
}
// Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
Value *Op0BOOp1 = Op0BO->getOperand(1);
if (isLeftShift && Op0BOOp1->hasOneUse() &&
match(Op0BOOp1,
m_And(m_OneUse(m_Shr(m_Value(V1), m_Specific(Op1))),
m_ConstantInt(CC)))) {
Value *YS = // (Y << C)
Builder->CreateShl(Op0BO->getOperand(0), Op1,
Op0BO->getName());
// X & (CC << C)
Value *XM = Builder->CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
V1->getName()+".mask");
return BinaryOperator::Create(Op0BO->getOpcode(), YS, XM);
}
}
// FALL THROUGH.
case Instruction::Sub: {
// Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
match(Op0BO->getOperand(0), m_Shr(m_Value(V1),
m_Specific(Op1)))) {
Value *YS = // (Y << C)
Builder->CreateShl(Op0BO->getOperand(1), Op1, Op0BO->getName());
// (X + (Y << C))
Value *X = Builder->CreateBinOp(Op0BO->getOpcode(), V1, YS,
Op0BO->getOperand(0)->getName());
uint32_t Op1Val = COp1->getLimitedValue(TypeBits);
APInt Bits = APInt::getHighBitsSet(TypeBits, TypeBits - Op1Val);
Constant *Mask = ConstantInt::get(I.getContext(), Bits);
if (VectorType *VT = dyn_cast<VectorType>(X->getType()))
Mask = ConstantVector::getSplat(VT->getNumElements(), Mask);
return BinaryOperator::CreateAnd(X, Mask);
}
// Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
match(Op0BO->getOperand(0),
m_And(m_OneUse(m_Shr(m_Value(V1), m_Value(V2))),
m_ConstantInt(CC))) && V2 == Op1) {
Value *YS = // (Y << C)
Builder->CreateShl(Op0BO->getOperand(1), Op1, Op0BO->getName());
// X & (CC << C)
Value *XM = Builder->CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
V1->getName()+".mask");
return BinaryOperator::Create(Op0BO->getOpcode(), XM, YS);
}
break;
}
}
// If the operand is an bitwise operator with a constant RHS, and the
// shift is the only use, we can pull it out of the shift.
if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
bool isValid = true; // Valid only for And, Or, Xor
bool highBitSet = false; // Transform if high bit of constant set?
switch (Op0BO->getOpcode()) {
default: isValid = false; break; // Do not perform transform!
case Instruction::Add:
isValid = isLeftShift;
break;
case Instruction::Or:
case Instruction::Xor:
highBitSet = false;
break;
case Instruction::And:
highBitSet = true;
break;
}
// If this is a signed shift right, and the high bit is modified
// by the logical operation, do not perform the transformation.
// The highBitSet boolean indicates the value of the high bit of
// the constant which would cause it to be modified for this
// operation.
//
if (isValid && I.getOpcode() == Instruction::AShr)
isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
if (isValid) {
Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
Value *NewShift =
Builder->CreateBinOp(I.getOpcode(), Op0BO->getOperand(0), Op1);
NewShift->takeName(Op0BO);
return BinaryOperator::Create(Op0BO->getOpcode(), NewShift,
NewRHS);
}
}
}
}
// Find out if this is a shift of a shift by a constant.
BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
if (ShiftOp && !ShiftOp->isShift())
ShiftOp = nullptr;
if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
// This is a constant shift of a constant shift. Be careful about hiding
// shl instructions behind bit masks. They are used to represent multiplies
// by a constant, and it is important that simple arithmetic expressions
// are still recognizable by scalar evolution.
//
// The transforms applied to shl are very similar to the transforms applied
// to mul by constant. We can be more aggressive about optimizing right
// shifts.
//
// Combinations of right and left shifts will still be optimized in
// DAGCombine where scalar evolution no longer applies.
ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
uint32_t ShiftAmt2 = COp1->getLimitedValue(TypeBits);
assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
if (ShiftAmt1 == 0) return nullptr; // Will be simplified in the future.
Value *X = ShiftOp->getOperand(0);
IntegerType *Ty = cast<IntegerType>(I.getType());
// Check for (X << c1) << c2 and (X >> c1) >> c2
if (I.getOpcode() == ShiftOp->getOpcode()) {
uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
// If this is oversized composite shift, then unsigned shifts get 0, ashr
// saturates.
if (AmtSum >= TypeBits) {
if (I.getOpcode() != Instruction::AShr)
return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
AmtSum = TypeBits-1; // Saturate to 31 for i32 ashr.
}
return BinaryOperator::Create(I.getOpcode(), X,
ConstantInt::get(Ty, AmtSum));
}
if (ShiftAmt1 == ShiftAmt2) {
// If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
if (I.getOpcode() == Instruction::LShr &&
ShiftOp->getOpcode() == Instruction::Shl) {
APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
return BinaryOperator::CreateAnd(X,
ConstantInt::get(I.getContext(), Mask));
}
} else if (ShiftAmt1 < ShiftAmt2) {
uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
// (X >>?,exact C1) << C2 --> X << (C2-C1)
// The inexact version is deferred to DAGCombine so we don't hide shl
// behind a bit mask.
if (I.getOpcode() == Instruction::Shl &&
ShiftOp->getOpcode() != Instruction::Shl &&
ShiftOp->isExact()) {
assert(ShiftOp->getOpcode() == Instruction::LShr ||
ShiftOp->getOpcode() == Instruction::AShr);
ConstantInt *ShiftDiffCst = ConstantInt::get(Ty, ShiftDiff);
BinaryOperator *NewShl = BinaryOperator::Create(Instruction::Shl,
X, ShiftDiffCst);
NewShl->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
NewShl->setHasNoSignedWrap(I.hasNoSignedWrap());
return NewShl;
}
// (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
if (I.getOpcode() == Instruction::LShr &&
ShiftOp->getOpcode() == Instruction::Shl) {
ConstantInt *ShiftDiffCst = ConstantInt::get(Ty, ShiftDiff);
// (X <<nuw C1) >>u C2 --> X >>u (C2-C1)
if (ShiftOp->hasNoUnsignedWrap()) {
BinaryOperator *NewLShr = BinaryOperator::Create(Instruction::LShr,
X, ShiftDiffCst);
NewLShr->setIsExact(I.isExact());
return NewLShr;
}
Value *Shift = Builder->CreateLShr(X, ShiftDiffCst);
APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
return BinaryOperator::CreateAnd(Shift,
ConstantInt::get(I.getContext(),Mask));
}
// We can't handle (X << C1) >>s C2, it shifts arbitrary bits in. However,
// we can handle (X <<nsw C1) >>s C2 since it only shifts in sign bits.
if (I.getOpcode() == Instruction::AShr &&
ShiftOp->getOpcode() == Instruction::Shl) {
if (ShiftOp->hasNoSignedWrap()) {
// (X <<nsw C1) >>s C2 --> X >>s (C2-C1)
ConstantInt *ShiftDiffCst = ConstantInt::get(Ty, ShiftDiff);
BinaryOperator *NewAShr = BinaryOperator::Create(Instruction::AShr,
X, ShiftDiffCst);
NewAShr->setIsExact(I.isExact());
return NewAShr;
}
}
} else {
assert(ShiftAmt2 < ShiftAmt1);
uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
// (X >>?exact C1) << C2 --> X >>?exact (C1-C2)
// The inexact version is deferred to DAGCombine so we don't hide shl
// behind a bit mask.
if (I.getOpcode() == Instruction::Shl &&
ShiftOp->getOpcode() != Instruction::Shl &&
ShiftOp->isExact()) {
ConstantInt *ShiftDiffCst = ConstantInt::get(Ty, ShiftDiff);
BinaryOperator *NewShr = BinaryOperator::Create(ShiftOp->getOpcode(),
X, ShiftDiffCst);
NewShr->setIsExact(true);
return NewShr;
}
// (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
if (I.getOpcode() == Instruction::LShr &&
ShiftOp->getOpcode() == Instruction::Shl) {
ConstantInt *ShiftDiffCst = ConstantInt::get(Ty, ShiftDiff);
if (ShiftOp->hasNoUnsignedWrap()) {
// (X <<nuw C1) >>u C2 --> X <<nuw (C1-C2)
BinaryOperator *NewShl = BinaryOperator::Create(Instruction::Shl,
X, ShiftDiffCst);
NewShl->setHasNoUnsignedWrap(true);
return NewShl;
}
Value *Shift = Builder->CreateShl(X, ShiftDiffCst);
APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
return BinaryOperator::CreateAnd(Shift,
ConstantInt::get(I.getContext(),Mask));
}
// We can't handle (X << C1) >>s C2, it shifts arbitrary bits in. However,
// we can handle (X <<nsw C1) >>s C2 since it only shifts in sign bits.
if (I.getOpcode() == Instruction::AShr &&
ShiftOp->getOpcode() == Instruction::Shl) {
if (ShiftOp->hasNoSignedWrap()) {
// (X <<nsw C1) >>s C2 --> X <<nsw (C1-C2)
ConstantInt *ShiftDiffCst = ConstantInt::get(Ty, ShiftDiff);
BinaryOperator *NewShl = BinaryOperator::Create(Instruction::Shl,
X, ShiftDiffCst);
NewShl->setHasNoSignedWrap(true);
return NewShl;
}
}
}
}
return nullptr;
}