//
//===----------------------------------------------------------------------===//
-
#include "InstCombine.h"
-#include "llvm/Target/TargetData.h"
-#include "llvm/IntrinsicInst.h"
+#include "llvm/IR/DataLayout.h"
+#include "llvm/IR/IntrinsicInst.h"
+#include "llvm/IR/PatternMatch.h"
using namespace llvm;
+using namespace llvm::PatternMatch;
+#define DEBUG_TYPE "instcombine"
-/// ShrinkDemandedConstant - Check to see if the specified operand of the
+/// ShrinkDemandedConstant - Check to see if the specified operand of the
/// specified instruction is a constant integer. If so, check to see if there
/// are any bits set in the constant that are not demanded. If so, shrink the
/// constant and return true.
-static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
+static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
APInt Demanded) {
assert(I && "No instruction?");
assert(OpNo < I->getNumOperands() && "Operand index too large");
if (!OpC) return false;
// If there are no bits set that aren't demanded, nothing to do.
- Demanded.zextOrTrunc(OpC->getValue().getBitWidth());
+ Demanded = Demanded.zextOrTrunc(OpC->getValue().getBitWidth());
if ((~Demanded & OpC->getValue()) == 0)
return false;
unsigned BitWidth = Inst.getType()->getScalarSizeInBits();
APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
APInt DemandedMask(APInt::getAllOnesValue(BitWidth));
-
- Value *V = SimplifyDemandedUseBits(&Inst, DemandedMask,
+
+ Value *V = SimplifyDemandedUseBits(&Inst, DemandedMask,
KnownZero, KnownOne, 0);
- if (V == 0) return false;
+ if (!V) return false;
if (V == &Inst) return true;
ReplaceInstUsesWith(Inst, V);
return true;
/// SimplifyDemandedBits - This form of SimplifyDemandedBits simplifies the
/// specified instruction operand if possible, updating it in place. It returns
/// true if it made any change and false otherwise.
-bool InstCombiner::SimplifyDemandedBits(Use &U, APInt DemandedMask,
+bool InstCombiner::SimplifyDemandedBits(Use &U, APInt DemandedMask,
APInt &KnownZero, APInt &KnownOne,
unsigned Depth) {
Value *NewVal = SimplifyDemandedUseBits(U.get(), DemandedMask,
KnownZero, KnownOne, Depth);
- if (NewVal == 0) return false;
+ if (!NewVal) return false;
U = NewVal;
return true;
}
/// to be one in the expression. KnownZero contains all the bits that are known
/// to be zero in the expression. These are provided to potentially allow the
/// caller (which might recursively be SimplifyDemandedBits itself) to simplify
-/// the expression. KnownOne and KnownZero always follow the invariant that
+/// the expression. KnownOne and KnownZero always follow the invariant that
/// KnownOne & KnownZero == 0. That is, a bit can't be both 1 and 0. Note that
/// the bits in KnownOne and KnownZero may only be accurate for those bits set
/// in DemandedMask. Note also that the bitwidth of V, DemandedMask, KnownZero
Value *InstCombiner::SimplifyDemandedUseBits(Value *V, APInt DemandedMask,
APInt &KnownZero, APInt &KnownOne,
unsigned Depth) {
- assert(V != 0 && "Null pointer of Value???");
+ assert(V != nullptr && "Null pointer of Value???");
assert(Depth <= 6 && "Limit Search Depth");
uint32_t BitWidth = DemandedMask.getBitWidth();
- const Type *VTy = V->getType();
- assert((TD || !VTy->isPointerTy()) &&
+ Type *VTy = V->getType();
+ assert((DL || !VTy->isPointerTy()) &&
"SimplifyDemandedBits needs to know bit widths!");
- assert((!TD || TD->getTypeSizeInBits(VTy->getScalarType()) == BitWidth) &&
+ assert((!DL || DL->getTypeSizeInBits(VTy->getScalarType()) == BitWidth) &&
(!VTy->isIntOrIntVectorTy() ||
VTy->getScalarSizeInBits() == BitWidth) &&
KnownZero.getBitWidth() == BitWidth &&
// We know all of the bits for a constant!
KnownOne = CI->getValue() & DemandedMask;
KnownZero = ~KnownOne & DemandedMask;
- return 0;
+ return nullptr;
}
if (isa<ConstantPointerNull>(V)) {
// We know all of the bits for a constant!
- KnownOne.clear();
+ KnownOne.clearAllBits();
KnownZero = DemandedMask;
- return 0;
+ return nullptr;
}
- KnownZero.clear();
- KnownOne.clear();
+ KnownZero.clearAllBits();
+ KnownOne.clearAllBits();
if (DemandedMask == 0) { // Not demanding any bits from V.
if (isa<UndefValue>(V))
- return 0;
+ return nullptr;
return UndefValue::get(VTy);
}
-
+
if (Depth == 6) // Limit search depth.
- return 0;
-
+ return nullptr;
+
APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
APInt RHSKnownZero(BitWidth, 0), RHSKnownOne(BitWidth, 0);
Instruction *I = dyn_cast<Instruction>(V);
if (!I) {
- ComputeMaskedBits(V, DemandedMask, KnownZero, KnownOne, Depth);
- return 0; // Only analyze instructions.
+ ComputeMaskedBits(V, KnownZero, KnownOne, Depth);
+ return nullptr; // Only analyze instructions.
}
// If there are multiple uses of this value and we aren't at the root, then
// this instruction has a simpler value in that context.
if (I->getOpcode() == Instruction::And) {
// If either the LHS or the RHS are Zero, the result is zero.
- ComputeMaskedBits(I->getOperand(1), DemandedMask,
- RHSKnownZero, RHSKnownOne, Depth+1);
- ComputeMaskedBits(I->getOperand(0), DemandedMask & ~RHSKnownZero,
- LHSKnownZero, LHSKnownOne, Depth+1);
-
+ ComputeMaskedBits(I->getOperand(1), RHSKnownZero, RHSKnownOne, Depth+1);
+ ComputeMaskedBits(I->getOperand(0), LHSKnownZero, LHSKnownOne, Depth+1);
+
// If all of the demanded bits are known 1 on one side, return the other.
// These bits cannot contribute to the result of the 'and' in this
// context.
- if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
+ if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
(DemandedMask & ~LHSKnownZero))
return I->getOperand(0);
- if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
+ if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
(DemandedMask & ~RHSKnownZero))
return I->getOperand(1);
-
+
// If all of the demanded bits in the inputs are known zeros, return zero.
if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
return Constant::getNullValue(VTy);
-
+
} else if (I->getOpcode() == Instruction::Or) {
// We can simplify (X|Y) -> X or Y in the user's context if we know that
// only bits from X or Y are demanded.
-
+
// If either the LHS or the RHS are One, the result is One.
- ComputeMaskedBits(I->getOperand(1), DemandedMask,
- RHSKnownZero, RHSKnownOne, Depth+1);
- ComputeMaskedBits(I->getOperand(0), DemandedMask & ~RHSKnownOne,
- LHSKnownZero, LHSKnownOne, Depth+1);
-
+ ComputeMaskedBits(I->getOperand(1), RHSKnownZero, RHSKnownOne, Depth+1);
+ ComputeMaskedBits(I->getOperand(0), LHSKnownZero, LHSKnownOne, Depth+1);
+
// If all of the demanded bits are known zero on one side, return the
// other. These bits cannot contribute to the result of the 'or' in this
// context.
- if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
+ if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
(DemandedMask & ~LHSKnownOne))
return I->getOperand(0);
- if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
+ if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
(DemandedMask & ~RHSKnownOne))
return I->getOperand(1);
-
+
// If all of the potentially set bits on one side are known to be set on
// the other side, just use the 'other' side.
- if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
+ if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
(DemandedMask & (~RHSKnownZero)))
return I->getOperand(0);
- if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
+ if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
(DemandedMask & (~LHSKnownZero)))
return I->getOperand(1);
+ } else if (I->getOpcode() == Instruction::Xor) {
+ // We can simplify (X^Y) -> X or Y in the user's context if we know that
+ // only bits from X or Y are demanded.
+
+ ComputeMaskedBits(I->getOperand(1), RHSKnownZero, RHSKnownOne, Depth+1);
+ ComputeMaskedBits(I->getOperand(0), LHSKnownZero, LHSKnownOne, Depth+1);
+
+ // If all of the demanded bits are known zero on one side, return the
+ // other.
+ if ((DemandedMask & RHSKnownZero) == DemandedMask)
+ return I->getOperand(0);
+ if ((DemandedMask & LHSKnownZero) == DemandedMask)
+ return I->getOperand(1);
}
-
+
// Compute the KnownZero/KnownOne bits to simplify things downstream.
- ComputeMaskedBits(I, DemandedMask, KnownZero, KnownOne, Depth);
- return 0;
+ ComputeMaskedBits(I, KnownZero, KnownOne, Depth);
+ return nullptr;
}
-
+
// If this is the root being simplified, allow it to have multiple uses,
// just set the DemandedMask to all bits so that we can try to simplify the
// operands. This allows visitTruncInst (for example) to simplify the
// operand of a trunc without duplicating all the logic below.
if (Depth == 0 && !V->hasOneUse())
DemandedMask = APInt::getAllOnesValue(BitWidth);
-
+
switch (I->getOpcode()) {
default:
- ComputeMaskedBits(I, DemandedMask, KnownZero, KnownOne, Depth);
+ ComputeMaskedBits(I, KnownZero, KnownOne, Depth);
break;
case Instruction::And:
// If either the LHS or the RHS are Zero, the result is zero.
SimplifyDemandedBits(I->getOperandUse(0), DemandedMask & ~RHSKnownZero,
LHSKnownZero, LHSKnownOne, Depth+1))
return I;
- assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
- assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
+ assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
+ assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
// If all of the demanded bits are known 1 on one side, return the other.
// These bits cannot contribute to the result of the 'and'.
- if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
+ if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
(DemandedMask & ~LHSKnownZero))
return I->getOperand(0);
- if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
+ if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
(DemandedMask & ~RHSKnownZero))
return I->getOperand(1);
-
+
// If all of the demanded bits in the inputs are known zeros, return zero.
if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
return Constant::getNullValue(VTy);
-
+
// If the RHS is a constant, see if we can simplify it.
if (ShrinkDemandedConstant(I, 1, DemandedMask & ~LHSKnownZero))
return I;
-
+
// Output known-1 bits are only known if set in both the LHS & RHS.
KnownOne = RHSKnownOne & LHSKnownOne;
// Output known-0 are known to be clear if zero in either the LHS | RHS.
break;
case Instruction::Or:
// If either the LHS or the RHS are One, the result is One.
- if (SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
+ if (SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
RHSKnownZero, RHSKnownOne, Depth+1) ||
- SimplifyDemandedBits(I->getOperandUse(0), DemandedMask & ~RHSKnownOne,
+ SimplifyDemandedBits(I->getOperandUse(0), DemandedMask & ~RHSKnownOne,
LHSKnownZero, LHSKnownOne, Depth+1))
return I;
- assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
- assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
-
+ assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
+ assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
+
// If all of the demanded bits are known zero on one side, return the other.
// These bits cannot contribute to the result of the 'or'.
- if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
+ if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
(DemandedMask & ~LHSKnownOne))
return I->getOperand(0);
- if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
+ if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
(DemandedMask & ~RHSKnownOne))
return I->getOperand(1);
// If all of the potentially set bits on one side are known to be set on
// the other side, just use the 'other' side.
- if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
+ if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
(DemandedMask & (~RHSKnownZero)))
return I->getOperand(0);
- if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
+ if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
(DemandedMask & (~LHSKnownZero)))
return I->getOperand(1);
-
+
// If the RHS is a constant, see if we can simplify it.
if (ShrinkDemandedConstant(I, 1, DemandedMask))
return I;
-
+
// Output known-0 bits are only known if clear in both the LHS & RHS.
KnownZero = RHSKnownZero & LHSKnownZero;
// Output known-1 are known to be set if set in either the LHS | RHS.
case Instruction::Xor: {
if (SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
RHSKnownZero, RHSKnownOne, Depth+1) ||
- SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
+ SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
LHSKnownZero, LHSKnownOne, Depth+1))
return I;
- assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
- assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
-
+ assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
+ assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
+
// If all of the demanded bits are known zero on one side, return the other.
// These bits cannot contribute to the result of the 'xor'.
if ((DemandedMask & RHSKnownZero) == DemandedMask)
return I->getOperand(0);
if ((DemandedMask & LHSKnownZero) == DemandedMask)
return I->getOperand(1);
-
+
// If all of the demanded bits are known to be zero on one side or the
// other, turn this into an *inclusive* or.
// e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
if ((DemandedMask & ~RHSKnownZero & ~LHSKnownZero) == 0) {
- Instruction *Or =
+ Instruction *Or =
BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
I->getName());
- return InsertNewInstBefore(Or, *I);
+ return InsertNewInstWith(Or, *I);
}
-
+
// If all of the demanded bits on one side are known, and all of the set
// bits on that side are also known to be set on the other side, turn this
// into an AND, as we know the bits will be cleared.
// e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
- if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask) {
+ if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask) {
// all known
if ((RHSKnownOne & LHSKnownOne) == RHSKnownOne) {
Constant *AndC = Constant::getIntegerValue(VTy,
~RHSKnownOne & DemandedMask);
- Instruction *And =
- BinaryOperator::CreateAnd(I->getOperand(0), AndC, "tmp");
- return InsertNewInstBefore(And, *I);
+ Instruction *And = BinaryOperator::CreateAnd(I->getOperand(0), AndC);
+ return InsertNewInstWith(And, *I);
}
}
-
+
// If the RHS is a constant, see if we can simplify it.
// FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
if (ShrinkDemandedConstant(I, 1, DemandedMask))
return I;
-
+
// If our LHS is an 'and' and if it has one use, and if any of the bits we
// are flipping are known to be set, then the xor is just resetting those
// bits to zero. We can just knock out bits from the 'and' and the 'xor',
ConstantInt *AndRHS = cast<ConstantInt>(LHSInst->getOperand(1));
ConstantInt *XorRHS = cast<ConstantInt>(I->getOperand(1));
APInt NewMask = ~(LHSKnownOne & RHSKnownOne & DemandedMask);
-
+
Constant *AndC =
ConstantInt::get(I->getType(), NewMask & AndRHS->getValue());
- Instruction *NewAnd =
- BinaryOperator::CreateAnd(I->getOperand(0), AndC, "tmp");
- InsertNewInstBefore(NewAnd, *I);
-
+ Instruction *NewAnd = BinaryOperator::CreateAnd(I->getOperand(0), AndC);
+ InsertNewInstWith(NewAnd, *I);
+
Constant *XorC =
ConstantInt::get(I->getType(), NewMask & XorRHS->getValue());
- Instruction *NewXor =
- BinaryOperator::CreateXor(NewAnd, XorC, "tmp");
- return InsertNewInstBefore(NewXor, *I);
+ Instruction *NewXor = BinaryOperator::CreateXor(NewAnd, XorC);
+ return InsertNewInstWith(NewXor, *I);
}
// Output known-0 bits are known if clear or set in both the LHS & RHS.
case Instruction::Select:
if (SimplifyDemandedBits(I->getOperandUse(2), DemandedMask,
RHSKnownZero, RHSKnownOne, Depth+1) ||
- SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
+ SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
LHSKnownZero, LHSKnownOne, Depth+1))
return I;
- assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
- assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
-
+ assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
+ assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
+
// If the operands are constants, see if we can simplify them.
if (ShrinkDemandedConstant(I, 1, DemandedMask) ||
ShrinkDemandedConstant(I, 2, DemandedMask))
return I;
-
+
// Only known if known in both the LHS and RHS.
KnownOne = RHSKnownOne & LHSKnownOne;
KnownZero = RHSKnownZero & LHSKnownZero;
break;
case Instruction::Trunc: {
unsigned truncBf = I->getOperand(0)->getType()->getScalarSizeInBits();
- DemandedMask.zext(truncBf);
- KnownZero.zext(truncBf);
- KnownOne.zext(truncBf);
- if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
+ DemandedMask = DemandedMask.zext(truncBf);
+ KnownZero = KnownZero.zext(truncBf);
+ KnownOne = KnownOne.zext(truncBf);
+ if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
KnownZero, KnownOne, Depth+1))
return I;
- DemandedMask.trunc(BitWidth);
- KnownZero.trunc(BitWidth);
- KnownOne.trunc(BitWidth);
- assert(!(KnownZero & KnownOne) && "Bits known to be one AND zero?");
+ DemandedMask = DemandedMask.trunc(BitWidth);
+ KnownZero = KnownZero.trunc(BitWidth);
+ KnownOne = KnownOne.trunc(BitWidth);
+ assert(!(KnownZero & KnownOne) && "Bits known to be one AND zero?");
break;
}
case Instruction::BitCast:
if (!I->getOperand(0)->getType()->isIntOrIntVectorTy())
- return 0; // vector->int or fp->int?
+ return nullptr; // vector->int or fp->int?
- if (const VectorType *DstVTy = dyn_cast<VectorType>(I->getType())) {
- if (const VectorType *SrcVTy =
+ if (VectorType *DstVTy = dyn_cast<VectorType>(I->getType())) {
+ if (VectorType *SrcVTy =
dyn_cast<VectorType>(I->getOperand(0)->getType())) {
if (DstVTy->getNumElements() != SrcVTy->getNumElements())
// Don't touch a bitcast between vectors of different element counts.
- return 0;
+ return nullptr;
} else
// Don't touch a scalar-to-vector bitcast.
- return 0;
+ return nullptr;
} else if (I->getOperand(0)->getType()->isVectorTy())
// Don't touch a vector-to-scalar bitcast.
- return 0;
+ return nullptr;
if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
KnownZero, KnownOne, Depth+1))
return I;
- assert(!(KnownZero & KnownOne) && "Bits known to be one AND zero?");
+ assert(!(KnownZero & KnownOne) && "Bits known to be one AND zero?");
break;
case Instruction::ZExt: {
// Compute the bits in the result that are not present in the input.
unsigned SrcBitWidth =I->getOperand(0)->getType()->getScalarSizeInBits();
-
- DemandedMask.trunc(SrcBitWidth);
- KnownZero.trunc(SrcBitWidth);
- KnownOne.trunc(SrcBitWidth);
+
+ DemandedMask = DemandedMask.trunc(SrcBitWidth);
+ KnownZero = KnownZero.trunc(SrcBitWidth);
+ KnownOne = KnownOne.trunc(SrcBitWidth);
if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
KnownZero, KnownOne, Depth+1))
return I;
- DemandedMask.zext(BitWidth);
- KnownZero.zext(BitWidth);
- KnownOne.zext(BitWidth);
- assert(!(KnownZero & KnownOne) && "Bits known to be one AND zero?");
+ DemandedMask = DemandedMask.zext(BitWidth);
+ KnownZero = KnownZero.zext(BitWidth);
+ KnownOne = KnownOne.zext(BitWidth);
+ assert(!(KnownZero & KnownOne) && "Bits known to be one AND zero?");
// The top bits are known to be zero.
KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
break;
case Instruction::SExt: {
// Compute the bits in the result that are not present in the input.
unsigned SrcBitWidth =I->getOperand(0)->getType()->getScalarSizeInBits();
-
- APInt InputDemandedBits = DemandedMask &
+
+ APInt InputDemandedBits = DemandedMask &
APInt::getLowBitsSet(BitWidth, SrcBitWidth);
APInt NewBits(APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth));
// If any of the sign extended bits are demanded, we know that the sign
// bit is demanded.
if ((NewBits & DemandedMask) != 0)
- InputDemandedBits.set(SrcBitWidth-1);
-
- InputDemandedBits.trunc(SrcBitWidth);
- KnownZero.trunc(SrcBitWidth);
- KnownOne.trunc(SrcBitWidth);
+ InputDemandedBits.setBit(SrcBitWidth-1);
+
+ InputDemandedBits = InputDemandedBits.trunc(SrcBitWidth);
+ KnownZero = KnownZero.trunc(SrcBitWidth);
+ KnownOne = KnownOne.trunc(SrcBitWidth);
if (SimplifyDemandedBits(I->getOperandUse(0), InputDemandedBits,
KnownZero, KnownOne, Depth+1))
return I;
- InputDemandedBits.zext(BitWidth);
- KnownZero.zext(BitWidth);
- KnownOne.zext(BitWidth);
- assert(!(KnownZero & KnownOne) && "Bits known to be one AND zero?");
-
+ InputDemandedBits = InputDemandedBits.zext(BitWidth);
+ KnownZero = KnownZero.zext(BitWidth);
+ KnownOne = KnownOne.zext(BitWidth);
+ assert(!(KnownZero & KnownOne) && "Bits known to be one AND zero?");
+
// If the sign bit of the input is known set or clear, then we know the
// top bits of the result.
if (KnownZero[SrcBitWidth-1] || (NewBits & ~DemandedMask) == NewBits) {
// Convert to ZExt cast
CastInst *NewCast = new ZExtInst(I->getOperand(0), VTy, I->getName());
- return InsertNewInstBefore(NewCast, *I);
+ return InsertNewInstWith(NewCast, *I);
} else if (KnownOne[SrcBitWidth-1]) { // Input sign bit known set
KnownOne |= NewBits;
}
// are not demanded, then the add doesn't demand them from its input
// either.
unsigned NLZ = DemandedMask.countLeadingZeros();
-
+
// If there is a constant on the RHS, there are a variety of xformations
// we can do.
if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
// won't work if the RHS is zero.
if (RHS->isZero())
break;
-
+
// If the top bit of the output is demanded, demand everything from the
// input. Otherwise, we demand all the input bits except NLZ top bits.
APInt InDemandedBits(APInt::getLowBitsSet(BitWidth, BitWidth - NLZ));
// Find information about known zero/one bits in the input.
- if (SimplifyDemandedBits(I->getOperandUse(0), InDemandedBits,
+ if (SimplifyDemandedBits(I->getOperandUse(0), InDemandedBits,
LHSKnownZero, LHSKnownOne, Depth+1))
return I;
// the constant.
if (ShrinkDemandedConstant(I, 1, InDemandedBits))
return I;
-
+
// Avoid excess work.
if (LHSKnownZero == 0 && LHSKnownOne == 0)
break;
-
+
// Turn it into OR if input bits are zero.
if ((LHSKnownZero & RHS->getValue()) == RHS->getValue()) {
Instruction *Or =
BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
I->getName());
- return InsertNewInstBefore(Or, *I);
+ return InsertNewInstWith(Or, *I);
}
-
+
// We can say something about the output known-zero and known-one bits,
// depending on potential carries from the input constant and the
// unknowns. For example if the LHS is known to have at most the 0x0F0F0
// bits set and the RHS constant is 0x01001, then we know we have a known
// one mask of 0x00001 and a known zero mask of 0xE0F0E.
-
+
// To compute this, we first compute the potential carry bits. These are
// the bits which may be modified. I'm not aware of a better way to do
// this scan.
const APInt &RHSVal = RHS->getValue();
APInt CarryBits((~LHSKnownZero + RHSVal) ^ (~LHSKnownZero ^ RHSVal));
-
+
// Now that we know which bits have carries, compute the known-1/0 sets.
-
+
// Bits are known one if they are known zero in one operand and one in the
// other, and there is no input carry.
- KnownOne = ((LHSKnownZero & RHSVal) |
+ KnownOne = ((LHSKnownZero & RHSVal) |
(LHSKnownOne & ~RHSVal)) & ~CarryBits;
-
+
// Bits are known zero if they are known zero in both operands and there
// is no input carry.
KnownZero = LHSKnownZero & ~RHSVal & ~CarryBits;
LHSKnownZero, LHSKnownOne, Depth+1))
return I;
}
+
// Otherwise just hand the sub off to ComputeMaskedBits to fill in
// the known zeros and ones.
- ComputeMaskedBits(V, DemandedMask, KnownZero, KnownOne, Depth);
+ ComputeMaskedBits(V, KnownZero, KnownOne, Depth);
+
+ // Turn this into a xor if LHS is 2^n-1 and the remaining bits are known
+ // zero.
+ if (ConstantInt *C0 = dyn_cast<ConstantInt>(I->getOperand(0))) {
+ APInt I0 = C0->getValue();
+ if ((I0 + 1).isPowerOf2() && (I0 | KnownZero).isAllOnesValue()) {
+ Instruction *Xor = BinaryOperator::CreateXor(I->getOperand(1), C0);
+ return InsertNewInstWith(Xor, *I);
+ }
+ }
break;
case Instruction::Shl:
if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
- uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
+ {
+ Value *VarX; ConstantInt *C1;
+ if (match(I->getOperand(0), m_Shr(m_Value(VarX), m_ConstantInt(C1)))) {
+ Instruction *Shr = cast<Instruction>(I->getOperand(0));
+ Value *R = SimplifyShrShlDemandedBits(Shr, I, DemandedMask,
+ KnownZero, KnownOne);
+ if (R)
+ return R;
+ }
+ }
+
+ uint64_t ShiftAmt = SA->getLimitedValue(BitWidth-1);
APInt DemandedMaskIn(DemandedMask.lshr(ShiftAmt));
- if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMaskIn,
+
+ // If the shift is NUW/NSW, then it does demand the high bits.
+ ShlOperator *IOp = cast<ShlOperator>(I);
+ if (IOp->hasNoSignedWrap())
+ DemandedMaskIn |= APInt::getHighBitsSet(BitWidth, ShiftAmt+1);
+ else if (IOp->hasNoUnsignedWrap())
+ DemandedMaskIn |= APInt::getHighBitsSet(BitWidth, ShiftAmt);
+
+ if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMaskIn,
KnownZero, KnownOne, Depth+1))
return I;
assert(!(KnownZero & KnownOne) && "Bits known to be one AND zero?");
case Instruction::LShr:
// For a logical shift right
if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
- uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
-
+ uint64_t ShiftAmt = SA->getLimitedValue(BitWidth-1);
+
// Unsigned shift right.
APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
+
+ // If the shift is exact, then it does demand the low bits (and knows that
+ // they are zero).
+ if (cast<LShrOperator>(I)->isExact())
+ DemandedMaskIn |= APInt::getLowBitsSet(BitWidth, ShiftAmt);
+
if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMaskIn,
KnownZero, KnownOne, Depth+1))
return I;
// Perform the logical shift right.
Instruction *NewVal = BinaryOperator::CreateLShr(
I->getOperand(0), I->getOperand(1), I->getName());
- return InsertNewInstBefore(NewVal, *I);
- }
+ return InsertNewInstWith(NewVal, *I);
+ }
// If the sign bit is the only bit demanded by this ashr, then there is no
// need to do it, the shift doesn't change the high bit.
if (DemandedMask.isSignBit())
return I->getOperand(0);
-
+
if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
- uint32_t ShiftAmt = SA->getLimitedValue(BitWidth);
-
+ uint32_t ShiftAmt = SA->getLimitedValue(BitWidth-1);
+
// Signed shift right.
APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
// If any of the "high bits" are demanded, we should set the sign bit as
// demanded.
if (DemandedMask.countLeadingZeros() <= ShiftAmt)
- DemandedMaskIn.set(BitWidth-1);
+ DemandedMaskIn.setBit(BitWidth-1);
+
+ // If the shift is exact, then it does demand the low bits (and knows that
+ // they are zero).
+ if (cast<AShrOperator>(I)->isExact())
+ DemandedMaskIn |= APInt::getLowBitsSet(BitWidth, ShiftAmt);
+
if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMaskIn,
KnownZero, KnownOne, Depth+1))
return I;
APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
-
+
// Handle the sign bits.
APInt SignBit(APInt::getSignBit(BitWidth));
// Adjust to where it is now in the mask.
- SignBit = APIntOps::lshr(SignBit, ShiftAmt);
-
+ SignBit = APIntOps::lshr(SignBit, ShiftAmt);
+
// If the input sign bit is known to be zero, or if none of the top bits
// are demanded, turn this into an unsigned shift right.
- if (BitWidth <= ShiftAmt || KnownZero[BitWidth-ShiftAmt-1] ||
+ if (BitWidth <= ShiftAmt || KnownZero[BitWidth-ShiftAmt-1] ||
(HighBits & ~DemandedMask) == HighBits) {
// Perform the logical shift right.
- Instruction *NewVal = BinaryOperator::CreateLShr(
- I->getOperand(0), SA, I->getName());
- return InsertNewInstBefore(NewVal, *I);
+ BinaryOperator *NewVal = BinaryOperator::CreateLShr(I->getOperand(0),
+ SA, I->getName());
+ NewVal->setIsExact(cast<BinaryOperator>(I)->isExact());
+ return InsertNewInstWith(NewVal, *I);
} else if ((KnownOne & SignBit) != 0) { // New bits are known one.
KnownOne |= HighBits;
}
break;
case Instruction::SRem:
if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
+ // X % -1 demands all the bits because we don't want to introduce
+ // INT_MIN % -1 (== undef) by accident.
+ if (Rem->isAllOnesValue())
+ break;
APInt RA = Rem->getValue().abs();
if (RA.isPowerOf2()) {
if (DemandedMask.ult(RA)) // srem won't affect demanded bits
if (LHSKnownOne[BitWidth-1] && ((LHSKnownOne & LowBits) != 0))
KnownOne |= ~LowBits;
- assert(!(KnownZero & KnownOne) && "Bits known to be one AND zero?");
+ assert(!(KnownZero & KnownOne) && "Bits known to be one AND zero?");
}
}
+
+ // The sign bit is the LHS's sign bit, except when the result of the
+ // remainder is zero.
+ if (DemandedMask.isNegative() && KnownZero.isNonNegative()) {
+ APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
+ ComputeMaskedBits(I->getOperand(0), LHSKnownZero, LHSKnownOne, Depth+1);
+ // If it's known zero, our sign bit is also zero.
+ if (LHSKnownZero.isNegative())
+ KnownZero.setBit(KnownZero.getBitWidth() - 1);
+ }
break;
case Instruction::URem: {
APInt KnownZero2(BitWidth, 0), KnownOne2(BitWidth, 0);
// just shift the input byte into position to eliminate the bswap.
unsigned NLZ = DemandedMask.countLeadingZeros();
unsigned NTZ = DemandedMask.countTrailingZeros();
-
+
// Round NTZ down to the next byte. If we have 11 trailing zeros, then
// we need all the bits down to bit 8. Likewise, round NLZ. If we
// have 14 leading zeros, round to 8.
if (BitWidth-NLZ-NTZ == 8) {
unsigned ResultBit = NTZ;
unsigned InputBit = BitWidth-NTZ-8;
-
+
// Replace this with either a left or right shift to get the byte into
// the right place.
Instruction *NewVal;
if (InputBit > ResultBit)
- NewVal = BinaryOperator::CreateLShr(I->getOperand(1),
+ NewVal = BinaryOperator::CreateLShr(II->getArgOperand(0),
ConstantInt::get(I->getType(), InputBit-ResultBit));
else
- NewVal = BinaryOperator::CreateShl(I->getOperand(1),
+ NewVal = BinaryOperator::CreateShl(II->getArgOperand(0),
ConstantInt::get(I->getType(), ResultBit-InputBit));
NewVal->takeName(I);
- return InsertNewInstBefore(NewVal, *I);
+ return InsertNewInstWith(NewVal, *I);
}
-
+
// TODO: Could compute known zero/one bits based on the input.
break;
}
+ case Intrinsic::x86_sse42_crc32_64_64:
+ KnownZero = APInt::getHighBitsSet(64, 32);
+ return nullptr;
}
}
- ComputeMaskedBits(V, DemandedMask, KnownZero, KnownOne, Depth);
+ ComputeMaskedBits(V, KnownZero, KnownOne, Depth);
break;
}
-
+
// If the client is only demanding bits that we know, return the known
// constant.
if ((DemandedMask & (KnownZero|KnownOne)) == DemandedMask)
return Constant::getIntegerValue(VTy, KnownOne);
- return 0;
+ return nullptr;
}
+/// Helper routine of SimplifyDemandedUseBits. It tries to simplify
+/// "E1 = (X lsr C1) << C2", where the C1 and C2 are constant, into
+/// "E2 = X << (C2 - C1)" or "E2 = X >> (C1 - C2)", depending on the sign
+/// of "C2-C1".
+///
+/// Suppose E1 and E2 are generally different in bits S={bm, bm+1,
+/// ..., bn}, without considering the specific value X is holding.
+/// This transformation is legal iff one of following conditions is hold:
+/// 1) All the bit in S are 0, in this case E1 == E2.
+/// 2) We don't care those bits in S, per the input DemandedMask.
+/// 3) Combination of 1) and 2). Some bits in S are 0, and we don't care the
+/// rest bits.
+///
+/// Currently we only test condition 2).
+///
+/// As with SimplifyDemandedUseBits, it returns NULL if the simplification was
+/// not successful.
+Value *InstCombiner::SimplifyShrShlDemandedBits(Instruction *Shr,
+ Instruction *Shl, APInt DemandedMask, APInt &KnownZero, APInt &KnownOne) {
+
+ const APInt &ShlOp1 = cast<ConstantInt>(Shl->getOperand(1))->getValue();
+ const APInt &ShrOp1 = cast<ConstantInt>(Shr->getOperand(1))->getValue();
+ if (!ShlOp1 || !ShrOp1)
+ return nullptr; // Noop.
+
+ Value *VarX = Shr->getOperand(0);
+ Type *Ty = VarX->getType();
+ unsigned BitWidth = Ty->getIntegerBitWidth();
+ if (ShlOp1.uge(BitWidth) || ShrOp1.uge(BitWidth))
+ return nullptr; // Undef.
+
+ unsigned ShlAmt = ShlOp1.getZExtValue();
+ unsigned ShrAmt = ShrOp1.getZExtValue();
+
+ KnownOne.clearAllBits();
+ KnownZero = APInt::getBitsSet(KnownZero.getBitWidth(), 0, ShlAmt-1);
+ KnownZero &= DemandedMask;
+
+ APInt BitMask1(APInt::getAllOnesValue(BitWidth));
+ APInt BitMask2(APInt::getAllOnesValue(BitWidth));
+
+ bool isLshr = (Shr->getOpcode() == Instruction::LShr);
+ BitMask1 = isLshr ? (BitMask1.lshr(ShrAmt) << ShlAmt) :
+ (BitMask1.ashr(ShrAmt) << ShlAmt);
+
+ if (ShrAmt <= ShlAmt) {
+ BitMask2 <<= (ShlAmt - ShrAmt);
+ } else {
+ BitMask2 = isLshr ? BitMask2.lshr(ShrAmt - ShlAmt):
+ BitMask2.ashr(ShrAmt - ShlAmt);
+ }
+
+ // Check if condition-2 (see the comment to this function) is satified.
+ if ((BitMask1 & DemandedMask) == (BitMask2 & DemandedMask)) {
+ if (ShrAmt == ShlAmt)
+ return VarX;
+
+ if (!Shr->hasOneUse())
+ return nullptr;
+
+ BinaryOperator *New;
+ if (ShrAmt < ShlAmt) {
+ Constant *Amt = ConstantInt::get(VarX->getType(), ShlAmt - ShrAmt);
+ New = BinaryOperator::CreateShl(VarX, Amt);
+ BinaryOperator *Orig = cast<BinaryOperator>(Shl);
+ New->setHasNoSignedWrap(Orig->hasNoSignedWrap());
+ New->setHasNoUnsignedWrap(Orig->hasNoUnsignedWrap());
+ } else {
+ Constant *Amt = ConstantInt::get(VarX->getType(), ShrAmt - ShlAmt);
+ New = isLshr ? BinaryOperator::CreateLShr(VarX, Amt) :
+ BinaryOperator::CreateAShr(VarX, Amt);
+ if (cast<BinaryOperator>(Shr)->isExact())
+ New->setIsExact(true);
+ }
+
+ return InsertNewInstWith(New, *Shl);
+ }
+
+ return nullptr;
+}
/// SimplifyDemandedVectorElts - The specified value produces a vector with
/// any number of elements. DemandedElts contains the set of elements that are
if (isa<UndefValue>(V)) {
// If the entire vector is undefined, just return this info.
UndefElts = EltMask;
- return 0;
+ return nullptr;
}
-
+
if (DemandedElts == 0) { // If nothing is demanded, provide undef.
UndefElts = EltMask;
return UndefValue::get(V->getType());
}
UndefElts = 0;
- if (ConstantVector *CV = dyn_cast<ConstantVector>(V)) {
- const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
+
+ // Handle ConstantAggregateZero, ConstantVector, ConstantDataSequential.
+ if (Constant *C = dyn_cast<Constant>(V)) {
+ // Check if this is identity. If so, return 0 since we are not simplifying
+ // anything.
+ if (DemandedElts.isAllOnesValue())
+ return nullptr;
+
+ Type *EltTy = cast<VectorType>(V->getType())->getElementType();
Constant *Undef = UndefValue::get(EltTy);
- std::vector<Constant*> Elts;
- for (unsigned i = 0; i != VWidth; ++i)
+ SmallVector<Constant*, 16> Elts;
+ for (unsigned i = 0; i != VWidth; ++i) {
if (!DemandedElts[i]) { // If not demanded, set to undef.
Elts.push_back(Undef);
- UndefElts.set(i);
- } else if (isa<UndefValue>(CV->getOperand(i))) { // Already undef.
+ UndefElts.setBit(i);
+ continue;
+ }
+
+ Constant *Elt = C->getAggregateElement(i);
+ if (!Elt) return nullptr;
+
+ if (isa<UndefValue>(Elt)) { // Already undef.
Elts.push_back(Undef);
- UndefElts.set(i);
+ UndefElts.setBit(i);
} else { // Otherwise, defined.
- Elts.push_back(CV->getOperand(i));
+ Elts.push_back(Elt);
}
+ }
// If we changed the constant, return it.
- Constant *NewCP = ConstantVector::get(Elts);
- return NewCP != CV ? NewCP : 0;
+ Constant *NewCV = ConstantVector::get(Elts);
+ return NewCV != C ? NewCV : nullptr;
}
-
- if (isa<ConstantAggregateZero>(V)) {
- // Simplify the CAZ to a ConstantVector where the non-demanded elements are
- // set to undef.
-
- // Check if this is identity. If so, return 0 since we are not simplifying
- // anything.
- if (DemandedElts.isAllOnesValue())
- return 0;
-
- const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
- Constant *Zero = Constant::getNullValue(EltTy);
- Constant *Undef = UndefValue::get(EltTy);
- std::vector<Constant*> Elts;
- for (unsigned i = 0; i != VWidth; ++i) {
- Constant *Elt = DemandedElts[i] ? Zero : Undef;
- Elts.push_back(Elt);
- }
- UndefElts = DemandedElts ^ EltMask;
- return ConstantVector::get(Elts);
- }
-
+
// Limit search depth.
if (Depth == 10)
- return 0;
+ return nullptr;
- // If multiple users are using the root value, procede with
+ // If multiple users are using the root value, proceed with
// simplification conservatively assuming that all elements
// are needed.
if (!V->hasOneUse()) {
// the main instcombine process.
if (Depth != 0)
// TODO: Just compute the UndefElts information recursively.
- return 0;
+ return nullptr;
// Conservatively assume that all elements are needed.
DemandedElts = EltMask;
}
-
+
Instruction *I = dyn_cast<Instruction>(V);
- if (!I) return 0; // Only analyze instructions.
-
+ if (!I) return nullptr; // Only analyze instructions.
+
bool MadeChange = false;
APInt UndefElts2(VWidth, 0);
Value *TmpV;
switch (I->getOpcode()) {
default: break;
-
+
case Instruction::InsertElement: {
// If this is a variable index, we don't know which element it overwrites.
// demand exactly the same input as we produce.
ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
- if (Idx == 0) {
+ if (!Idx) {
// Note that we can't propagate undef elt info, because we don't know
// which elt is getting updated.
TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
break;
}
-
+
// If this is inserting an element that isn't demanded, remove this
// insertelement.
unsigned IdxNo = Idx->getZExtValue();
Worklist.Add(I);
return I->getOperand(0);
}
-
+
// Otherwise, the element inserted overwrites whatever was there, so the
// input demanded set is simpler than the output set.
APInt DemandedElts2 = DemandedElts;
- DemandedElts2.clear(IdxNo);
+ DemandedElts2.clearBit(IdxNo);
TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts2,
UndefElts, Depth+1);
if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
// The inserted element is defined.
- UndefElts.clear(IdxNo);
+ UndefElts.clearBit(IdxNo);
break;
}
case Instruction::ShuffleVector: {
assert(MaskVal < LHSVWidth * 2 &&
"shufflevector mask index out of range!");
if (MaskVal < LHSVWidth)
- LeftDemanded.set(MaskVal);
+ LeftDemanded.setBit(MaskVal);
else
- RightDemanded.set(MaskVal - LHSVWidth);
+ RightDemanded.setBit(MaskVal - LHSVWidth);
}
}
}
for (unsigned i = 0; i < VWidth; i++) {
unsigned MaskVal = Shuffle->getMaskValue(i);
if (MaskVal == -1u) {
- UndefElts.set(i);
+ UndefElts.setBit(i);
+ } else if (!DemandedElts[i]) {
+ NewUndefElts = true;
+ UndefElts.setBit(i);
} else if (MaskVal < LHSVWidth) {
if (UndefElts4[MaskVal]) {
NewUndefElts = true;
- UndefElts.set(i);
+ UndefElts.setBit(i);
}
} else {
if (UndefElts3[MaskVal - LHSVWidth]) {
NewUndefElts = true;
- UndefElts.set(i);
+ UndefElts.setBit(i);
}
}
}
if (NewUndefElts) {
// Add additional discovered undefs.
- std::vector<Constant*> Elts;
+ SmallVector<Constant*, 16> Elts;
for (unsigned i = 0; i < VWidth; ++i) {
if (UndefElts[i])
Elts.push_back(UndefValue::get(Type::getInt32Ty(I->getContext())));
}
break;
}
+ case Instruction::Select: {
+ APInt LeftDemanded(DemandedElts), RightDemanded(DemandedElts);
+ if (ConstantVector* CV = dyn_cast<ConstantVector>(I->getOperand(0))) {
+ for (unsigned i = 0; i < VWidth; i++) {
+ if (CV->getAggregateElement(i)->isNullValue())
+ LeftDemanded.clearBit(i);
+ else
+ RightDemanded.clearBit(i);
+ }
+ }
+
+ TmpV = SimplifyDemandedVectorElts(I->getOperand(1), LeftDemanded,
+ UndefElts, Depth+1);
+ if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
+
+ TmpV = SimplifyDemandedVectorElts(I->getOperand(2), RightDemanded,
+ UndefElts2, Depth+1);
+ if (TmpV) { I->setOperand(2, TmpV); MadeChange = true; }
+
+ // Output elements are undefined if both are undefined.
+ UndefElts &= UndefElts2;
+ break;
+ }
case Instruction::BitCast: {
// Vector->vector casts only.
- const VectorType *VTy = dyn_cast<VectorType>(I->getOperand(0)->getType());
+ VectorType *VTy = dyn_cast<VectorType>(I->getOperand(0)->getType());
if (!VTy) break;
unsigned InVWidth = VTy->getNumElements();
APInt InputDemandedElts(InVWidth, 0);
} else if (VWidth > InVWidth) {
// Untested so far.
break;
-
+
// If there are more elements in the result than there are in the source,
// then an input element is live if any of the corresponding output
// elements are live.
Ratio = VWidth/InVWidth;
for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx) {
if (DemandedElts[OutIdx])
- InputDemandedElts.set(OutIdx/Ratio);
+ InputDemandedElts.setBit(OutIdx/Ratio);
}
} else {
// Untested so far.
break;
-
+
// If there are more elements in the source than there are in the result,
// then an input element is live if the corresponding output element is
// live.
Ratio = InVWidth/VWidth;
for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
if (DemandedElts[InIdx/Ratio])
- InputDemandedElts.set(InIdx);
+ InputDemandedElts.setBit(InIdx);
}
-
+
// div/rem demand all inputs, because they don't want divide by zero.
TmpV = SimplifyDemandedVectorElts(I->getOperand(0), InputDemandedElts,
UndefElts2, Depth+1);
I->setOperand(0, TmpV);
MadeChange = true;
}
-
+
UndefElts = UndefElts2;
if (VWidth > InVWidth) {
llvm_unreachable("Unimp");
// undef.
for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx)
if (UndefElts2[OutIdx/Ratio])
- UndefElts.set(OutIdx);
+ UndefElts.setBit(OutIdx);
} else if (VWidth < InVWidth) {
llvm_unreachable("Unimp");
// If there are more elements in the source than there are in the result,
UndefElts = ~0ULL >> (64-VWidth); // Start out all undef.
for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
if (!UndefElts2[InIdx]) // Not undef?
- UndefElts.clear(InIdx/Ratio); // Clear undef bit.
+ UndefElts.clearBit(InIdx/Ratio); // Clear undef bit.
}
break;
}
TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
UndefElts2, Depth+1);
if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
-
+
// Output elements are undefined if both are undefined. Consider things
// like undef&0. The result is known zero, not undef.
UndefElts &= UndefElts2;
break;
-
+ case Instruction::FPTrunc:
+ case Instruction::FPExt:
+ TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
+ UndefElts, Depth+1);
+ if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
+ break;
+
case Instruction::Call: {
IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
if (!II) break;
switch (II->getIntrinsicID()) {
default: break;
-
+
// Binary vector operations that work column-wise. A dest element is a
// function of the corresponding input elements from the two inputs.
case Intrinsic::x86_sse_sub_ss:
case Intrinsic::x86_sse2_mul_sd:
case Intrinsic::x86_sse2_min_sd:
case Intrinsic::x86_sse2_max_sd:
- TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
+ TmpV = SimplifyDemandedVectorElts(II->getArgOperand(0), DemandedElts,
UndefElts, Depth+1);
- if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
- TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
+ if (TmpV) { II->setArgOperand(0, TmpV); MadeChange = true; }
+ TmpV = SimplifyDemandedVectorElts(II->getArgOperand(1), DemandedElts,
UndefElts2, Depth+1);
- if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
+ if (TmpV) { II->setArgOperand(1, TmpV); MadeChange = true; }
// If only the low elt is demanded and this is a scalarizable intrinsic,
// scalarize it now.
case Intrinsic::x86_sse2_sub_sd:
case Intrinsic::x86_sse2_mul_sd:
// TODO: Lower MIN/MAX/ABS/etc
- Value *LHS = II->getOperand(1);
- Value *RHS = II->getOperand(2);
+ Value *LHS = II->getArgOperand(0);
+ Value *RHS = II->getArgOperand(1);
// Extract the element as scalars.
- LHS = InsertNewInstBefore(ExtractElementInst::Create(LHS,
+ LHS = InsertNewInstWith(ExtractElementInst::Create(LHS,
ConstantInt::get(Type::getInt32Ty(I->getContext()), 0U)), *II);
- RHS = InsertNewInstBefore(ExtractElementInst::Create(RHS,
+ RHS = InsertNewInstWith(ExtractElementInst::Create(RHS,
ConstantInt::get(Type::getInt32Ty(I->getContext()), 0U)), *II);
-
+
switch (II->getIntrinsicID()) {
default: llvm_unreachable("Case stmts out of sync!");
case Intrinsic::x86_sse_sub_ss:
case Intrinsic::x86_sse2_sub_sd:
- TmpV = InsertNewInstBefore(BinaryOperator::CreateFSub(LHS, RHS,
+ TmpV = InsertNewInstWith(BinaryOperator::CreateFSub(LHS, RHS,
II->getName()), *II);
break;
case Intrinsic::x86_sse_mul_ss:
case Intrinsic::x86_sse2_mul_sd:
- TmpV = InsertNewInstBefore(BinaryOperator::CreateFMul(LHS, RHS,
+ TmpV = InsertNewInstWith(BinaryOperator::CreateFMul(LHS, RHS,
II->getName()), *II);
break;
}
-
+
Instruction *New =
InsertElementInst::Create(
UndefValue::get(II->getType()), TmpV,
ConstantInt::get(Type::getInt32Ty(I->getContext()), 0U, false),
II->getName());
- InsertNewInstBefore(New, *II);
+ InsertNewInstWith(New, *II);
return New;
- }
+ }
}
-
+
// Output elements are undefined if both are undefined. Consider things
// like undef&0. The result is known zero, not undef.
UndefElts &= UndefElts2;
break;
}
}
- return MadeChange ? I : 0;
+ return MadeChange ? I : nullptr;
}