Shifter ops are not always free. Do not fold them (especially to form

complex load / store addressing mode) when they have higher cost and
when they have more than one use.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@117509 91177308-0d34-0410-b5e6-96231b3b80d8
This commit is contained in:
Evan Cheng
2010-10-27 23:41:30 +00:00
parent de5fa932b9
commit f40deed62f
4 changed files with 183 additions and 25 deletions

View File

@@ -78,8 +78,12 @@ public:
SDNode *Select(SDNode *N);
bool isShifterOpProfitable(const SDValue &Shift,
ARM_AM::ShiftOpc ShOpcVal, unsigned ShAmt);
bool SelectShifterOperandReg(SDValue N, SDValue &A,
SDValue &B, SDValue &C);
bool SelectShiftShifterOperandReg(SDValue N, SDValue &A,
SDValue &B, SDValue &C);
bool SelectAddrModeImm12(SDValue N, SDValue &Base, SDValue &OffImm);
bool SelectLdStSOReg(SDValue N, SDValue &Base, SDValue &Offset, SDValue &Opc);
@@ -246,6 +250,17 @@ static bool isOpcWithIntImmediate(SDNode *N, unsigned Opc, unsigned& Imm) {
}
bool ARMDAGToDAGISel::isShifterOpProfitable(const SDValue &Shift,
ARM_AM::ShiftOpc ShOpcVal,
unsigned ShAmt) {
if (!Subtarget->isCortexA9())
return true;
if (Shift.hasOneUse())
return true;
// R << 2 is free.
return ShOpcVal == ARM_AM::lsl && ShAmt == 2;
}
bool ARMDAGToDAGISel::SelectShifterOperandReg(SDValue N,
SDValue &BaseReg,
SDValue &ShReg,
@@ -261,6 +276,32 @@ bool ARMDAGToDAGISel::SelectShifterOperandReg(SDValue N,
BaseReg = N.getOperand(0);
unsigned ShImmVal = 0;
if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
ShReg = CurDAG->getRegister(0, MVT::i32);
ShImmVal = RHS->getZExtValue() & 31;
} else {
ShReg = N.getOperand(1);
if (!isShifterOpProfitable(N, ShOpcVal, ShImmVal))
return false;
}
Opc = CurDAG->getTargetConstant(ARM_AM::getSORegOpc(ShOpcVal, ShImmVal),
MVT::i32);
return true;
}
bool ARMDAGToDAGISel::SelectShiftShifterOperandReg(SDValue N,
SDValue &BaseReg,
SDValue &ShReg,
SDValue &Opc) {
ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N);
// Don't match base register only case. That is matched to a separate
// lower complexity pattern with explicit register operand.
if (ShOpcVal == ARM_AM::no_shift) return false;
BaseReg = N.getOperand(0);
unsigned ShImmVal = 0;
// Do not check isShifterOpProfitable. This must return true.
if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
ShReg = CurDAG->getRegister(0, MVT::i32);
ShImmVal = RHS->getZExtValue() & 31;
@@ -321,7 +362,8 @@ bool ARMDAGToDAGISel::SelectAddrModeImm12(SDValue N,
bool ARMDAGToDAGISel::SelectLdStSOReg(SDValue N, SDValue &Base, SDValue &Offset,
SDValue &Opc) {
if (N.getOpcode() == ISD::MUL) {
if (N.getOpcode() == ISD::MUL &&
(!Subtarget->isCortexA9() || N.hasOneUse())) {
if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
// X * [3,5,9] -> X + X * [2,4,8] etc.
int RHSC = (int)RHS->getZExtValue();
@@ -357,6 +399,10 @@ bool ARMDAGToDAGISel::SelectLdStSOReg(SDValue N, SDValue &Base, SDValue &Offset,
}
}
if (Subtarget->isCortexA9() && !N.hasOneUse())
// Compute R +/- (R << N) and reuse it.
return false;
// Otherwise this is R +/- [possibly shifted] R.
ARM_AM::AddrOpc AddSub = N.getOpcode() == ISD::ADD ? ARM_AM::add:ARM_AM::sub;
ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOperand(1));
@@ -371,14 +417,20 @@ bool ARMDAGToDAGISel::SelectLdStSOReg(SDValue N, SDValue &Base, SDValue &Offset,
if (ConstantSDNode *Sh =
dyn_cast<ConstantSDNode>(N.getOperand(1).getOperand(1))) {
ShAmt = Sh->getZExtValue();
Offset = N.getOperand(1).getOperand(0);
if (isShifterOpProfitable(Offset, ShOpcVal, ShAmt))
Offset = N.getOperand(1).getOperand(0);
else {
ShAmt = 0;
ShOpcVal = ARM_AM::no_shift;
}
} else {
ShOpcVal = ARM_AM::no_shift;
}
}
// Try matching (R shl C) + (R).
if (N.getOpcode() == ISD::ADD && ShOpcVal == ARM_AM::no_shift) {
if (N.getOpcode() == ISD::ADD && ShOpcVal == ARM_AM::no_shift &&
!(Subtarget->isCortexA9() || N.getOperand(0).hasOneUse())) {
ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOperand(0));
if (ShOpcVal != ARM_AM::no_shift) {
// Check to see if the RHS of the shift is a constant, if not, we can't
@@ -386,8 +438,15 @@ bool ARMDAGToDAGISel::SelectLdStSOReg(SDValue N, SDValue &Base, SDValue &Offset,
if (ConstantSDNode *Sh =
dyn_cast<ConstantSDNode>(N.getOperand(0).getOperand(1))) {
ShAmt = Sh->getZExtValue();
Offset = N.getOperand(0).getOperand(0);
Base = N.getOperand(1);
if (!Subtarget->isCortexA9() ||
(N.hasOneUse() &&
isShifterOpProfitable(N.getOperand(0), ShOpcVal, ShAmt))) {
Offset = N.getOperand(0).getOperand(0);
Base = N.getOperand(1);
} else {
ShAmt = 0;
ShOpcVal = ARM_AM::no_shift;
}
} else {
ShOpcVal = ARM_AM::no_shift;
}
@@ -408,7 +467,8 @@ AddrMode2Type ARMDAGToDAGISel::SelectAddrMode2Worker(SDValue N,
SDValue &Base,
SDValue &Offset,
SDValue &Opc) {
if (N.getOpcode() == ISD::MUL) {
if (N.getOpcode() == ISD::MUL &&
(!Subtarget->isCortexA9() || N.hasOneUse())) {
if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
// X * [3,5,9] -> X + X * [2,4,8] etc.
int RHSC = (int)RHS->getZExtValue();
@@ -474,6 +534,16 @@ AddrMode2Type ARMDAGToDAGISel::SelectAddrMode2Worker(SDValue N,
}
}
if (Subtarget->isCortexA9() && !N.hasOneUse()) {
// Compute R +/- (R << N) and reuse it.
Base = N;
Offset = CurDAG->getRegister(0, MVT::i32);
Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(ARM_AM::add, 0,
ARM_AM::no_shift),
MVT::i32);
return AM2_BASE;
}
// Otherwise this is R +/- [possibly shifted] R.
ARM_AM::AddrOpc AddSub = N.getOpcode() == ISD::ADD ? ARM_AM::add:ARM_AM::sub;
ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOperand(1));
@@ -488,14 +558,20 @@ AddrMode2Type ARMDAGToDAGISel::SelectAddrMode2Worker(SDValue N,
if (ConstantSDNode *Sh =
dyn_cast<ConstantSDNode>(N.getOperand(1).getOperand(1))) {
ShAmt = Sh->getZExtValue();
Offset = N.getOperand(1).getOperand(0);
if (isShifterOpProfitable(Offset, ShOpcVal, ShAmt))
Offset = N.getOperand(1).getOperand(0);
else {
ShAmt = 0;
ShOpcVal = ARM_AM::no_shift;
}
} else {
ShOpcVal = ARM_AM::no_shift;
}
}
// Try matching (R shl C) + (R).
if (N.getOpcode() == ISD::ADD && ShOpcVal == ARM_AM::no_shift) {
if (N.getOpcode() == ISD::ADD && ShOpcVal == ARM_AM::no_shift &&
!(Subtarget->isCortexA9() || N.getOperand(0).hasOneUse())) {
ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOperand(0));
if (ShOpcVal != ARM_AM::no_shift) {
// Check to see if the RHS of the shift is a constant, if not, we can't
@@ -503,8 +579,15 @@ AddrMode2Type ARMDAGToDAGISel::SelectAddrMode2Worker(SDValue N,
if (ConstantSDNode *Sh =
dyn_cast<ConstantSDNode>(N.getOperand(0).getOperand(1))) {
ShAmt = Sh->getZExtValue();
Offset = N.getOperand(0).getOperand(0);
Base = N.getOperand(1);
if (!Subtarget->isCortexA9() ||
(N.hasOneUse() &&
isShifterOpProfitable(N.getOperand(0), ShOpcVal, ShAmt))) {
Offset = N.getOperand(0).getOperand(0);
Base = N.getOperand(1);
} else {
ShAmt = 0;
ShOpcVal = ARM_AM::no_shift;
}
} else {
ShOpcVal = ARM_AM::no_shift;
}
@@ -543,7 +626,12 @@ bool ARMDAGToDAGISel::SelectAddrMode2Offset(SDNode *Op, SDValue N,
// it.
if (ConstantSDNode *Sh = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
ShAmt = Sh->getZExtValue();
Offset = N.getOperand(0);
if (isShifterOpProfitable(N, ShOpcVal, ShAmt))
Offset = N.getOperand(0);
else {
ShAmt = 0;
ShOpcVal = ARM_AM::no_shift;
}
} else {
ShOpcVal = ARM_AM::no_shift;
}
@@ -959,6 +1047,12 @@ bool ARMDAGToDAGISel::SelectT2AddrModeSoReg(SDValue N,
return false;
}
if (Subtarget->isCortexA9() && !N.hasOneUse()) {
// Compute R + (R << [1,2,3]) and reuse it.
Base = N;
return false;
}
// Look for (R + R) or (R + (R << [1,2,3])).
unsigned ShAmt = 0;
Base = N.getOperand(0);
@@ -977,11 +1071,12 @@ bool ARMDAGToDAGISel::SelectT2AddrModeSoReg(SDValue N,
// it.
if (ConstantSDNode *Sh = dyn_cast<ConstantSDNode>(OffReg.getOperand(1))) {
ShAmt = Sh->getZExtValue();
if (ShAmt >= 4) {
if (ShAmt < 4 && isShifterOpProfitable(OffReg, ShOpcVal, ShAmt))
OffReg = OffReg.getOperand(0);
else {
ShAmt = 0;
ShOpcVal = ARM_AM::no_shift;
} else
OffReg = OffReg.getOperand(0);
}
} else {
ShOpcVal = ARM_AM::no_shift;
}