//===----------------------------------------------------------------------===//
#include "InstCombine.h"
+#include "llvm/IntrinsicInst.h"
#include "llvm/Support/PatternMatch.h"
using namespace llvm;
using namespace PatternMatch;
if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
return Res;
+
+ // X shift (A srem B) -> X shift (A and B-1) iff B is a power of 2.
+ // Because shifts by negative values are undefined.
+ if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op1))
+ if (BO->hasOneUse() && BO->getOpcode() == Instruction::SRem)
+ if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1)))
+ if (CI->getValue().isPowerOf2()) {
+ Constant *C = ConstantInt::get(BO->getType(), CI->getValue()-1);
+ Value *Rem = Builder->CreateAnd(BO->getOperand(0), C, BO->getName());
+ I.setOperand(1, Rem);
+ return &I;
+ }
+
return 0;
}
+/// CanEvaluateShifted - See if we can compute the specified value, but shifted
+/// logically to the left or right by some number of bits. This should return
+/// true if the expression can be computed for the same cost as the current
+/// expression tree. This is used to eliminate extraneous shifting from things
+/// like:
+/// %C = shl i128 %A, 64
+/// %D = shl i128 %B, 96
+/// %E = or i128 %C, %D
+/// %F = lshr i128 %E, 64
+/// where the client will ask if E can be computed shifted right by 64-bits. If
+/// this succeeds, the GetShiftedValue function will be called to produce the
+/// value.
+static bool CanEvaluateShifted(Value *V, unsigned NumBits, bool isLeftShift,
+ InstCombiner &IC) {
+ // We can always evaluate constants shifted.
+ if (isa<Constant>(V))
+ return true;
+
+ Instruction *I = dyn_cast<Instruction>(V);
+ if (!I) return false;
+
+ // If this is the opposite shift, we can directly reuse the input of the shift
+ // if the needed bits are already zero in the input. This allows us to reuse
+ // the value which means that we don't care if the shift has multiple uses.
+ // TODO: Handle opposite shift by exact value.
+ ConstantInt *CI;
+ if ((isLeftShift && match(I, m_LShr(m_Value(), m_ConstantInt(CI)))) ||
+ (!isLeftShift && match(I, m_Shl(m_Value(), m_ConstantInt(CI))))) {
+ if (CI->getZExtValue() == NumBits) {
+ // TODO: Check that the input bits are already zero with MaskedValueIsZero
+#if 0
+ // If this is a truncate of a logical shr, we can truncate it to a smaller
+ // lshr iff we know that the bits we would otherwise be shifting in are
+ // already zeros.
+ uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
+ uint32_t BitWidth = Ty->getScalarSizeInBits();
+ if (MaskedValueIsZero(I->getOperand(0),
+ APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
+ CI->getLimitedValue(BitWidth) < BitWidth) {
+ return CanEvaluateTruncated(I->getOperand(0), Ty);
+ }
+#endif
+
+ }
+ }
+
+ // We can't mutate something that has multiple uses: doing so would
+ // require duplicating the instruction in general, which isn't profitable.
+ if (!I->hasOneUse()) return false;
+
+ switch (I->getOpcode()) {
+ default: return false;
+ case Instruction::And:
+ case Instruction::Or:
+ case Instruction::Xor:
+ // Bitwise operators can all arbitrarily be arbitrarily evaluated shifted.
+ return CanEvaluateShifted(I->getOperand(0), NumBits, isLeftShift, IC) &&
+ CanEvaluateShifted(I->getOperand(1), NumBits, isLeftShift, IC);
+
+ case Instruction::Shl: {
+ // We can often fold the shift into shifts-by-a-constant.
+ CI = dyn_cast<ConstantInt>(I->getOperand(1));
+ if (CI == 0) return false;
+
+ // We can always fold shl(c1)+shl(c2) -> shl(c1+c2).
+ if (isLeftShift) return true;
+
+ // We can always turn shl(c)+shr(c) -> and(c2).
+ if (CI->getValue() == NumBits) return true;
+
+ unsigned TypeWidth = I->getType()->getScalarSizeInBits();
+
+ // We can turn shl(c1)+shr(c2) -> shl(c3)+and(c4), but it isn't
+ // profitable unless we know the and'd out bits are already zero.
+ if (CI->getZExtValue() > NumBits) {
+ unsigned LowBits = TypeWidth - CI->getZExtValue();
+ if (MaskedValueIsZero(I->getOperand(0),
+ APInt::getLowBitsSet(TypeWidth, NumBits) << LowBits))
+ return true;
+ }
+
+ return false;
+ }
+ case Instruction::LShr: {
+ // We can often fold the shift into shifts-by-a-constant.
+ CI = dyn_cast<ConstantInt>(I->getOperand(1));
+ if (CI == 0) return false;
+
+ // We can always fold lshr(c1)+lshr(c2) -> lshr(c1+c2).
+ if (!isLeftShift) return true;
+
+ // We can always turn lshr(c)+shl(c) -> and(c2).
+ if (CI->getValue() == NumBits) return true;
+
+ unsigned TypeWidth = I->getType()->getScalarSizeInBits();
+
+ // We can always turn lshr(c1)+shl(c2) -> lshr(c3)+and(c4), but it isn't
+ // profitable unless we know the and'd out bits are already zero.
+ if (CI->getZExtValue() > NumBits) {
+ unsigned LowBits = CI->getZExtValue() - NumBits;
+ if (MaskedValueIsZero(I->getOperand(0),
+ APInt::getLowBitsSet(TypeWidth, NumBits) << LowBits))
+ return true;
+ }
+
+ return false;
+ }
+ case Instruction::Select: {
+ SelectInst *SI = cast<SelectInst>(I);
+ return CanEvaluateShifted(SI->getTrueValue(), NumBits, isLeftShift, IC) &&
+ CanEvaluateShifted(SI->getFalseValue(), NumBits, isLeftShift, IC);
+ }
+ case Instruction::PHI: {
+ // We can change a phi if we can change all operands. Note that we never
+ // get into trouble with cyclic PHIs here because we only consider
+ // instructions with a single use.
+ PHINode *PN = cast<PHINode>(I);
+ for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
+ if (!CanEvaluateShifted(PN->getIncomingValue(i), NumBits, isLeftShift,IC))
+ return false;
+ return true;
+ }
+ }
+}
+
+/// GetShiftedValue - When CanEvaluateShifted returned true for an expression,
+/// this value inserts the new computation that produces the shifted value.
+static Value *GetShiftedValue(Value *V, unsigned NumBits, bool isLeftShift,
+ InstCombiner &IC) {
+ // We can always evaluate constants shifted.
+ if (Constant *C = dyn_cast<Constant>(V)) {
+ if (isLeftShift)
+ V = IC.Builder->CreateShl(C, NumBits);
+ else
+ V = IC.Builder->CreateLShr(C, NumBits);
+ // If we got a constantexpr back, try to simplify it with TD info.
+ if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
+ V = ConstantFoldConstantExpression(CE, IC.getTargetData());
+ return V;
+ }
+
+ Instruction *I = cast<Instruction>(V);
+ IC.Worklist.Add(I);
+
+ switch (I->getOpcode()) {
+ default: assert(0 && "Inconsistency with CanEvaluateShifted");
+ case Instruction::And:
+ case Instruction::Or:
+ case Instruction::Xor:
+ // Bitwise operators can all arbitrarily be arbitrarily evaluated shifted.
+ I->setOperand(0, GetShiftedValue(I->getOperand(0), NumBits,isLeftShift,IC));
+ I->setOperand(1, GetShiftedValue(I->getOperand(1), NumBits,isLeftShift,IC));
+ return I;
+
+ case Instruction::Shl: {
+ unsigned TypeWidth = I->getType()->getScalarSizeInBits();
+
+ // We only accept shifts-by-a-constant in CanEvaluateShifted.
+ ConstantInt *CI = cast<ConstantInt>(I->getOperand(1));
+
+ // We can always fold shl(c1)+shl(c2) -> shl(c1+c2).
+ if (isLeftShift) {
+ // If this is oversized composite shift, then unsigned shifts get 0.
+ unsigned NewShAmt = NumBits+CI->getZExtValue();
+ if (NewShAmt >= TypeWidth)
+ return Constant::getNullValue(I->getType());
+
+ I->setOperand(1, ConstantInt::get(I->getType(), NewShAmt));
+ return I;
+ }
+
+ // We turn shl(c)+lshr(c) -> and(c2) if the input doesn't already have
+ // zeros.
+ if (CI->getValue() == NumBits) {
+ APInt Mask(APInt::getLowBitsSet(TypeWidth, TypeWidth - NumBits));
+ V = IC.Builder->CreateAnd(I->getOperand(0),
+ ConstantInt::get(I->getContext(), Mask));
+ if (Instruction *VI = dyn_cast<Instruction>(V)) {
+ VI->moveBefore(I);
+ VI->takeName(I);
+ }
+ return V;
+ }
+
+ // We turn shl(c1)+shr(c2) -> shl(c3)+and(c4), but only when we know that
+ // the and won't be needed.
+ assert(CI->getZExtValue() > NumBits);
+ I->setOperand(1, ConstantInt::get(I->getType(),
+ CI->getZExtValue() - NumBits));
+ return I;
+ }
+ case Instruction::LShr: {
+ unsigned TypeWidth = I->getType()->getScalarSizeInBits();
+ // We only accept shifts-by-a-constant in CanEvaluateShifted.
+ ConstantInt *CI = cast<ConstantInt>(I->getOperand(1));
+
+ // We can always fold lshr(c1)+lshr(c2) -> lshr(c1+c2).
+ if (!isLeftShift) {
+ // If this is oversized composite shift, then unsigned shifts get 0.
+ unsigned NewShAmt = NumBits+CI->getZExtValue();
+ if (NewShAmt >= TypeWidth)
+ return Constant::getNullValue(I->getType());
+
+ I->setOperand(1, ConstantInt::get(I->getType(), NewShAmt));
+ return I;
+ }
+
+ // We turn lshr(c)+shl(c) -> and(c2) if the input doesn't already have
+ // zeros.
+ if (CI->getValue() == NumBits) {
+ APInt Mask(APInt::getHighBitsSet(TypeWidth, TypeWidth - NumBits));
+ V = IC.Builder->CreateAnd(I->getOperand(0),
+ ConstantInt::get(I->getContext(), Mask));
+ if (Instruction *VI = dyn_cast<Instruction>(V)) {
+ VI->moveBefore(I);
+ VI->takeName(I);
+ }
+ return V;
+ }
+
+ // We turn lshr(c1)+shl(c2) -> lshr(c3)+and(c4), but only when we know that
+ // the and won't be needed.
+ assert(CI->getZExtValue() > NumBits);
+ I->setOperand(1, ConstantInt::get(I->getType(),
+ CI->getZExtValue() - NumBits));
+ return I;
+ }
+
+ case Instruction::Select:
+ I->setOperand(1, GetShiftedValue(I->getOperand(1), NumBits,isLeftShift,IC));
+ I->setOperand(2, GetShiftedValue(I->getOperand(2), NumBits,isLeftShift,IC));
+ return I;
+ case Instruction::PHI: {
+ // We can change a phi if we can change all operands. Note that we never
+ // get into trouble with cyclic PHIs here because we only consider
+ // instructions with a single use.
+ PHINode *PN = cast<PHINode>(I);
+ for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
+ PN->setIncomingValue(i, GetShiftedValue(PN->getIncomingValue(i),
+ NumBits, isLeftShift, IC));
+ return PN;
+ }
+ }
+}
+
+
+
Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
BinaryOperator &I) {
bool isLeftShift = I.getOpcode() == Instruction::Shl;
-
+
+
+ // 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, Op1->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, Op1->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();
if (Op1->uge(TypeBits)) {
if (I.getOpcode() != Instruction::AShr)
return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
- else {
- I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
- return &I;
- }
+ // ashr i32 X, 32 --> ashr i32 X, 31
+ I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
+ return &I;
}
// ((X*C1) << C2) == (X * (C1 << C2))
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 = Op1->getLimitedValue(TypeBits);
- return BinaryOperator::CreateAnd(X, ConstantInt::get(I.getContext(),
- APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
- }
-
- // 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_Shr(m_Value(V1), m_Specific(Op1)),
- m_ConstantInt(CC))) &&
- cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse()) {
- 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);
- }
+ 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 = Op1->getLimitedValue(TypeBits);
+ return BinaryOperator::CreateAnd(X, ConstantInt::get(I.getContext(),
+ APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
}
+
+ // 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_Shr(m_Value(V1), m_Specific(Op1)),
+ m_ConstantInt(CC))) &&
+ cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse()) {
+ 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 = Op1->getLimitedValue(TypeBits);
+ return BinaryOperator::CreateAnd(X, ConstantInt::get(I.getContext(),
+ APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
+ }
+
+ // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
+ if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
+ match(Op0BO->getOperand(0),
+ m_And(m_Shr(m_Value(V1), m_Value(V2)),
+ m_ConstantInt(CC))) && V2 == Op1 &&
+ cast<BinaryOperator>(Op0BO->getOperand(0))
+ ->getOperand(0)->hasOneUse()) {
+ 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");
- // 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 = Op1->getLimitedValue(TypeBits);
- return BinaryOperator::CreateAnd(X, ConstantInt::get(I.getContext(),
- APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
- }
-
- // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
- if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
- match(Op0BO->getOperand(0),
- m_And(m_Shr(m_Value(V1), m_Value(V2)),
- m_ConstantInt(CC))) && V2 == Op1 &&
- cast<BinaryOperator>(Op0BO->getOperand(0))
- ->getOperand(0)->hasOneUse()) {
- 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;
+ return BinaryOperator::Create(Op0BO->getOpcode(), XM, YS);
}
+
+ break;
+ }
}
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;
+ 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
ConstantInt::get(Ty, AmtSum));
}
- if (ShiftOp->getOpcode() == Instruction::LShr &&
- I.getOpcode() == Instruction::AShr) {
- if (AmtSum >= TypeBits)
- return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
-
- // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
- return BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, AmtSum));
- }
-
- if (ShiftOp->getOpcode() == Instruction::AShr &&
- I.getOpcode() == Instruction::LShr) {
- // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
- if (AmtSum >= TypeBits)
- AmtSum = TypeBits-1;
-
- Value *Shift = Builder->CreateAShr(X, ConstantInt::get(Ty, AmtSum));
-
- APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
- return BinaryOperator::CreateAnd(Shift,
- ConstantInt::get(I.getContext(), Mask));
- }
-
- // Okay, if we get here, one shift must be left, and the other shift must be
- // right. See if the amounts are equal.
if (ShiftAmt1 == ShiftAmt2) {
// If we have ((X >>? C) << C), turn this into X & (-1 << C).
- if (I.getOpcode() == Instruction::Shl) {
+ if (I.getOpcode() == Instruction::Shl &&
+ ShiftOp->getOpcode() != Instruction::Shl) {
APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
return BinaryOperator::CreateAnd(X,
ConstantInt::get(I.getContext(),Mask));
}
// If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
- if (I.getOpcode() == Instruction::LShr) {
+ 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));
}
- // We can simplify ((X << C) >>s C) into a trunc + sext.
- // NOTE: we could do this for any C, but that would make 'unusual' integer
- // types. For now, just stick to ones well-supported by the code
- // generators.
- const Type *SExtType = 0;
- switch (Ty->getBitWidth() - ShiftAmt1) {
- case 1 :
- case 8 :
- case 16 :
- case 32 :
- case 64 :
- case 128:
- SExtType = IntegerType::get(I.getContext(),
- Ty->getBitWidth() - ShiftAmt1);
- break;
- default: break;
- }
- if (SExtType)
- return new SExtInst(Builder->CreateTrunc(X, SExtType, "sext"), Ty);
- // Otherwise, we can't handle it yet.
} else if (ShiftAmt1 < ShiftAmt2) {
uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
// (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
- if (I.getOpcode() == Instruction::Shl) {
+ if (I.getOpcode() == Instruction::Shl &&
+ ShiftOp->getOpcode() != Instruction::Shl) {
assert(ShiftOp->getOpcode() == Instruction::LShr ||
ShiftOp->getOpcode() == Instruction::AShr);
Value *Shift = Builder->CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
}
// (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
- if (I.getOpcode() == Instruction::LShr) {
+ if (I.getOpcode() == Instruction::LShr &&
+ ShiftOp->getOpcode() == Instruction::Shl) {
assert(ShiftOp->getOpcode() == Instruction::Shl);
Value *Shift = Builder->CreateLShr(X, ConstantInt::get(Ty, ShiftDiff));
uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
// (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
- if (I.getOpcode() == Instruction::Shl) {
- assert(ShiftOp->getOpcode() == Instruction::LShr ||
- ShiftOp->getOpcode() == Instruction::AShr);
+ if (I.getOpcode() == Instruction::Shl &&
+ ShiftOp->getOpcode() != Instruction::Shl) {
Value *Shift = Builder->CreateBinOp(ShiftOp->getOpcode(), X,
ConstantInt::get(Ty, ShiftDiff));
}
// (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
- if (I.getOpcode() == Instruction::LShr) {
- assert(ShiftOp->getOpcode() == Instruction::Shl);
+ if (I.getOpcode() == Instruction::LShr &&
+ ShiftOp->getOpcode() == Instruction::Shl) {
Value *Shift = Builder->CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
}
Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
- return commonShiftTransforms(I);
+ if (Instruction *R = commonShiftTransforms(I))
+ return R;
+
+ Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
+
+ if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1))
+ if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Op0)) {
+ unsigned BitWidth = Op0->getType()->getScalarSizeInBits();
+ // ctlz.i32(x)>>5 --> zext(x == 0)
+ // cttz.i32(x)>>5 --> zext(x == 0)
+ // ctpop.i32(x)>>5 --> zext(x == -1)
+ if ((II->getIntrinsicID() == Intrinsic::ctlz ||
+ II->getIntrinsicID() == Intrinsic::cttz ||
+ II->getIntrinsicID() == Intrinsic::ctpop) &&
+ isPowerOf2_32(BitWidth) && Log2_32(BitWidth) == Op1C->getZExtValue()){
+ bool isCtPop = II->getIntrinsicID() == Intrinsic::ctpop;
+ Constant *RHS = ConstantInt::getSigned(Op0->getType(), isCtPop ? -1:0);
+ Value *Cmp = Builder->CreateICmpEQ(II->getArgOperand(0), RHS);
+ return new ZExtInst(Cmp, II->getType());
+ }
+ }
+
+ return 0;
}
Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
// If the input is a SHL by the same constant (ashr (shl X, C), C), then we
- // have a sign-extend idiom. If the input value is known to already be sign
- // extended enough, delete the extension.
+ // have a sign-extend idiom.
Value *X;
- if (match(Op0, m_Shl(m_Value(X), m_Specific(Op1))) &&
- ComputeNumSignBits(X) > Op1C->getZExtValue())
- return ReplaceInstUsesWith(I, X);
+ if (match(Op0, m_Shl(m_Value(X), m_Specific(Op1)))) {
+ // If the input value is known to already be sign extended enough, delete
+ // the extension.
+ if (ComputeNumSignBits(X) > Op1C->getZExtValue())
+ return ReplaceInstUsesWith(I, X);
+
+ // If the input is an extension from the shifted amount value, e.g.
+ // %x = zext i8 %A to i32
+ // %y = shl i32 %x, 24
+ // %z = ashr %y, 24
+ // then turn this into "z = sext i8 A to i32".
+ if (ZExtInst *ZI = dyn_cast<ZExtInst>(X)) {
+ uint32_t SrcBits = ZI->getOperand(0)->getType()->getScalarSizeInBits();
+ uint32_t DestBits = ZI->getType()->getScalarSizeInBits();
+ if (Op1C->getZExtValue() == DestBits-SrcBits)
+ return new SExtInst(ZI->getOperand(0), ZI->getType());
+ }
+ }
}
// See if we can turn a signed shr into an unsigned shr.
projects / oota-llvm.git / blobdiff