public:
explicit WorkListRemover(DAGCombiner &dc) : DC(dc) {}
- virtual void NodeDeleted(SDNode *N) {
+ virtual void NodeDeleted(SDNode *N, SDNode *E) {
DC.removeFromWorkList(N);
}
AfterLegalize = RunningAfterLegalize;
// Add all the dag nodes to the worklist.
+ WorkList.reserve(DAG.allnodes_size());
for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(),
E = DAG.allnodes_end(); I != E; ++I)
WorkList.push_back(I);
DAG.getNode(ISD::CARRY_FALSE, MVT::Flag));
// canonicalize constant to RHS.
- if (N0C && !N1C) {
- SDOperand Ops[] = { N1, N0 };
- return DAG.getNode(ISD::ADDC, N->getVTList(), Ops, 2);
- }
+ if (N0C && !N1C)
+ return DAG.getNode(ISD::ADDC, N->getVTList(), N1, N0);
// fold (addc x, 0) -> x + no carry out
if (N1C && N1C->isNullValue())
//MVT VT = N0.getValueType();
// canonicalize constant to RHS
- if (N0C && !N1C) {
- SDOperand Ops[] = { N1, N0, CarryIn };
- return DAG.getNode(ISD::ADDE, N->getVTList(), Ops, 3);
- }
+ if (N0C && !N1C)
+ return DAG.getNode(ISD::ADDE, N->getVTList(), N1, N0, CarryIn);
// fold (adde x, y, false) -> (addc x, y)
- if (CarryIn.getOpcode() == ISD::CARRY_FALSE) {
- SDOperand Ops[] = { N1, N0 };
- return DAG.getNode(ISD::ADDC, N->getVTList(), Ops, 2);
- }
+ if (CarryIn.getOpcode() == ISD::CARRY_FALSE)
+ return DAG.getNode(ISD::ADDC, N->getVTList(), N1, N0);
return SDOperand();
}
AddToWorkList(Lo.Val);
SDOperand LoOpt = combine(Lo.Val);
if (LoOpt.Val && LoOpt.Val != Lo.Val &&
- TLI.isOperationLegal(LoOpt.getOpcode(), LoOpt.getValueType()))
+ (!AfterLegalize ||
+ TLI.isOperationLegal(LoOpt.getOpcode(), LoOpt.getValueType())))
return CombineTo(N, LoOpt, LoOpt);
}
AddToWorkList(Hi.Val);
SDOperand HiOpt = combine(Hi.Val);
if (HiOpt.Val && HiOpt != Hi &&
- TLI.isOperationLegal(HiOpt.getOpcode(), HiOpt.getValueType()))
+ (!AfterLegalize ||
+ TLI.isOperationLegal(HiOpt.getOpcode(), HiOpt.getValueType())))
return CombineTo(N, HiOpt, HiOpt);
}
return SDOperand();
unsigned BitWidth = N1.getValueSizeInBits();
if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
BitWidth - EVT.getSizeInBits())) &&
- (!AfterLegalize || TLI.isLoadXLegal(ISD::ZEXTLOAD, EVT))) {
+ ((!AfterLegalize && !LN0->isVolatile()) ||
+ TLI.isLoadXLegal(ISD::ZEXTLOAD, EVT))) {
SDOperand ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, VT, LN0->getChain(),
LN0->getBasePtr(), LN0->getSrcValue(),
LN0->getSrcValueOffset(), EVT,
unsigned BitWidth = N1.getValueSizeInBits();
if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
BitWidth - EVT.getSizeInBits())) &&
- (!AfterLegalize || TLI.isLoadXLegal(ISD::ZEXTLOAD, EVT))) {
+ ((!AfterLegalize && !LN0->isVolatile()) ||
+ TLI.isLoadXLegal(ISD::ZEXTLOAD, EVT))) {
SDOperand ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, VT, LN0->getChain(),
LN0->getBasePtr(), LN0->getSrcValue(),
LN0->getSrcValueOffset(), EVT,
if (N1C && N0.getOpcode() == ISD::LOAD) {
LoadSDNode *LN0 = cast<LoadSDNode>(N0);
if (LN0->getExtensionType() != ISD::SEXTLOAD &&
- LN0->isUnindexed() && N0.hasOneUse()) {
- MVT EVT, LoadedVT;
- if (N1C->getAPIntValue() == 255)
- EVT = MVT::i8;
- else if (N1C->getAPIntValue() == 65535)
- EVT = MVT::i16;
- else if (N1C->getAPIntValue() == ~0U)
- EVT = MVT::i32;
- else
- EVT = MVT::Other;
-
- LoadedVT = LN0->getMemoryVT();
- if (EVT != MVT::Other && LoadedVT > EVT &&
+ LN0->isUnindexed() && N0.hasOneUse() &&
+ // Do not change the width of a volatile load.
+ !LN0->isVolatile()) {
+ MVT EVT = MVT::Other;
+ uint32_t ActiveBits = N1C->getAPIntValue().getActiveBits();
+ if (ActiveBits > 0 && APIntOps::isMask(ActiveBits, N1C->getAPIntValue()))
+ EVT = MVT::getIntegerVT(ActiveBits);
+
+ MVT LoadedVT = LN0->getMemoryVT();
+ // Do not generate loads of non-round integer types since these can
+ // be expensive (and would be wrong if the type is not byte sized).
+ if (EVT != MVT::Other && LoadedVT.bitsGT(EVT) && EVT.isRound() &&
(!AfterLegalize || TLI.isLoadXLegal(ISD::ZEXTLOAD, EVT))) {
MVT PtrType = N0.getOperand(1).getValueType();
// For big endian targets, we need to add an offset to the pointer to
// idioms for rotate, and if the target supports rotation instructions, generate
// a rot[lr].
SDNode *DAGCombiner::MatchRotate(SDOperand LHS, SDOperand RHS) {
- // Must be a legal type. Expanded an promoted things won't work with rotates.
+ // Must be a legal type. Expanded 'n promoted things won't work with rotates.
MVT VT = LHS.getValueType();
if (!TLI.isTypeLegal(VT)) return 0;
bool HasROTL = TLI.isOperationLegal(ISD::ROTL, VT);
bool HasROTR = TLI.isOperationLegal(ISD::ROTR, VT);
if (!HasROTL && !HasROTR) return 0;
-
+
// Match "(X shl/srl V1) & V2" where V2 may not be present.
SDOperand LHSShift; // The shift.
SDOperand LHSMask; // AND value if any.
// sext_inreg.
if (N1C && N0.getOpcode() == ISD::SHL && N1 == N0.getOperand(1)) {
unsigned LowBits = VT.getSizeInBits() - (unsigned)N1C->getValue();
- MVT EVT;
- switch (LowBits) {
- default: EVT = MVT::Other; break;
- case 1: EVT = MVT::i1; break;
- case 8: EVT = MVT::i8; break;
- case 16: EVT = MVT::i16; break;
- case 32: EVT = MVT::i32; break;
- }
- if (EVT > MVT::Other && TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, EVT))
+ MVT EVT = MVT::getIntegerVT(LowBits);
+ if (EVT.isSimple() && // TODO: remove when apint codegen support lands.
+ (!AfterLegalize || TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, EVT)))
return DAG.getNode(ISD::SIGN_EXTEND_INREG, VT, N0.getOperand(0),
DAG.getValueType(EVT));
}
-
+
// fold (sra (sra x, c1), c2) -> (sra x, c1+c2)
if (N1C && N0.getOpcode() == ISD::SRA) {
if (ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
MVT::getIntegerVT(VTValSize - N1C->getValue());
// Determine the residual right-shift amount.
unsigned ShiftAmt = N1C->getValue() - N01C->getValue();
-
+
// If the shift is not a no-op (in which case this should be just a sign
// extend already), the truncated to type is legal, sign_extend is legal
// on that type, and the the truncate to that type is both legal and free,
// perform the transform.
if (ShiftAmt &&
- TLI.isTypeLegal(TruncVT) &&
TLI.isOperationLegal(ISD::SIGN_EXTEND, TruncVT) &&
TLI.isOperationLegal(ISD::TRUNCATE, VT) &&
TLI.isTruncateFree(VT, TruncVT)) {
if (VT == VT0)
return XORNode;
AddToWorkList(XORNode.Val);
- if (VT.getSizeInBits() > VT0.getSizeInBits())
+ if (VT.bitsGT(VT0))
return DAG.getNode(ISD::ZERO_EXTEND, VT, XORNode);
return DAG.getNode(ISD::TRUNCATE, VT, XORNode);
}
// If we can fold this based on the true/false value, do so.
if (SimplifySelectOps(N, N1, N2))
return SDOperand(N, 0); // Don't revisit N.
-
+
// fold selects based on a setcc into other things, such as min/max/abs
if (N0.getOpcode() == ISD::SETCC) {
// FIXME:
// fold (sext (truncate x)) -> (sextinreg x).
if (!AfterLegalize || TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG,
N0.getValueType())) {
- if (Op.getValueType() < VT)
+ if (Op.getValueType().bitsLT(VT))
Op = DAG.getNode(ISD::ANY_EXTEND, VT, Op);
- else if (Op.getValueType() > VT)
+ else if (Op.getValueType().bitsGT(VT))
Op = DAG.getNode(ISD::TRUNCATE, VT, Op);
return DAG.getNode(ISD::SIGN_EXTEND_INREG, VT, Op,
DAG.getValueType(N0.getValueType()));
// fold (sext (load x)) -> (sext (truncate (sextload x)))
if (ISD::isNON_EXTLoad(N0.Val) &&
- (!AfterLegalize||TLI.isLoadXLegal(ISD::SEXTLOAD, N0.getValueType()))){
+ ((!AfterLegalize && !cast<LoadSDNode>(N0)->isVolatile()) ||
+ TLI.isLoadXLegal(ISD::SEXTLOAD, N0.getValueType()))) {
bool DoXform = true;
SmallVector<SDNode*, 4> SetCCs;
if (!N0.hasOneUse())
ISD::isUNINDEXEDLoad(N0.Val) && N0.hasOneUse()) {
LoadSDNode *LN0 = cast<LoadSDNode>(N0);
MVT EVT = LN0->getMemoryVT();
- if (!AfterLegalize || TLI.isLoadXLegal(ISD::SEXTLOAD, EVT)) {
+ if ((!AfterLegalize && !LN0->isVolatile()) ||
+ TLI.isLoadXLegal(ISD::SEXTLOAD, EVT)) {
SDOperand ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, VT, LN0->getChain(),
LN0->getBasePtr(), LN0->getSrcValue(),
LN0->getSrcValueOffset(), EVT,
if (N0.getOpcode() == ISD::TRUNCATE &&
(!AfterLegalize || TLI.isOperationLegal(ISD::AND, VT))) {
SDOperand Op = N0.getOperand(0);
- if (Op.getValueType() < VT) {
+ if (Op.getValueType().bitsLT(VT)) {
Op = DAG.getNode(ISD::ANY_EXTEND, VT, Op);
- } else if (Op.getValueType() > VT) {
+ } else if (Op.getValueType().bitsGT(VT)) {
Op = DAG.getNode(ISD::TRUNCATE, VT, Op);
}
return DAG.getZeroExtendInReg(Op, N0.getValueType());
N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
N0.getOperand(1).getOpcode() == ISD::Constant) {
SDOperand X = N0.getOperand(0).getOperand(0);
- if (X.getValueType() < VT) {
+ if (X.getValueType().bitsLT(VT)) {
X = DAG.getNode(ISD::ANY_EXTEND, VT, X);
- } else if (X.getValueType() > VT) {
+ } else if (X.getValueType().bitsGT(VT)) {
X = DAG.getNode(ISD::TRUNCATE, VT, X);
}
APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
// fold (zext (load x)) -> (zext (truncate (zextload x)))
if (ISD::isNON_EXTLoad(N0.Val) &&
- (!AfterLegalize||TLI.isLoadXLegal(ISD::ZEXTLOAD, N0.getValueType()))) {
+ ((!AfterLegalize && !cast<LoadSDNode>(N0)->isVolatile()) ||
+ TLI.isLoadXLegal(ISD::ZEXTLOAD, N0.getValueType()))) {
bool DoXform = true;
SmallVector<SDNode*, 4> SetCCs;
if (!N0.hasOneUse())
ISD::isUNINDEXEDLoad(N0.Val) && N0.hasOneUse()) {
LoadSDNode *LN0 = cast<LoadSDNode>(N0);
MVT EVT = LN0->getMemoryVT();
- SDOperand ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, VT, LN0->getChain(),
- LN0->getBasePtr(), LN0->getSrcValue(),
- LN0->getSrcValueOffset(), EVT,
- LN0->isVolatile(),
- LN0->getAlignment());
- CombineTo(N, ExtLoad);
- CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad),
- ExtLoad.getValue(1));
- return SDOperand(N, 0); // Return N so it doesn't get rechecked!
+ if ((!AfterLegalize && !LN0->isVolatile()) ||
+ TLI.isLoadXLegal(ISD::ZEXTLOAD, EVT)) {
+ SDOperand ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, VT, LN0->getChain(),
+ LN0->getBasePtr(), LN0->getSrcValue(),
+ LN0->getSrcValueOffset(), EVT,
+ LN0->isVolatile(),
+ LN0->getAlignment());
+ CombineTo(N, ExtLoad);
+ CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad),
+ ExtLoad.getValue(1));
+ return SDOperand(N, 0); // Return N so it doesn't get rechecked!
+ }
}
// zext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
SDOperand TruncOp = N0.getOperand(0);
if (TruncOp.getValueType() == VT)
return TruncOp; // x iff x size == zext size.
- if (TruncOp.getValueType() > VT)
+ if (TruncOp.getValueType().bitsGT(VT))
return DAG.getNode(ISD::TRUNCATE, VT, TruncOp);
return DAG.getNode(ISD::ANY_EXTEND, VT, TruncOp);
}
N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
N0.getOperand(1).getOpcode() == ISD::Constant) {
SDOperand X = N0.getOperand(0).getOperand(0);
- if (X.getValueType() < VT) {
+ if (X.getValueType().bitsLT(VT)) {
X = DAG.getNode(ISD::ANY_EXTEND, VT, X);
- } else if (X.getValueType() > VT) {
+ } else if (X.getValueType().bitsGT(VT)) {
X = DAG.getNode(ISD::TRUNCATE, VT, X);
}
APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
// fold (aext (load x)) -> (aext (truncate (extload x)))
if (ISD::isNON_EXTLoad(N0.Val) && N0.hasOneUse() &&
- (!AfterLegalize||TLI.isLoadXLegal(ISD::EXTLOAD, N0.getValueType()))) {
+ ((!AfterLegalize && !cast<LoadSDNode>(N0)->isVolatile()) ||
+ TLI.isLoadXLegal(ISD::EXTLOAD, N0.getValueType()))) {
LoadSDNode *LN0 = cast<LoadSDNode>(N0);
SDOperand ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, VT, LN0->getChain(),
LN0->getBasePtr(), LN0->getSrcValue(),
}
}
- if (ISD::isNON_EXTLoad(N0.Val) && N0.hasOneUse() &&
- // Do not allow folding to i1 here. i1 is implicitly stored in memory in
- // zero extended form: by shrinking the load, we lose track of the fact
- // that it is already zero extended.
- // FIXME: This should be reevaluated.
- VT != MVT::i1) {
+ // Do not generate loads of non-round integer types since these can
+ // be expensive (and would be wrong if the type is not byte sized).
+ if (ISD::isNON_EXTLoad(N0.Val) && N0.hasOneUse() && VT.isRound() &&
+ // Do not change the width of a volatile load.
+ !cast<LoadSDNode>(N0)->isVolatile()) {
assert(N0.getValueType().getSizeInBits() > EVTBits &&
"Cannot truncate to larger type!");
LoadSDNode *LN0 = cast<LoadSDNode>(N0);
// fold (sext_in_reg (sext_in_reg x, VT2), VT1) -> (sext_in_reg x, minVT) pt2
if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
- EVT < cast<VTSDNode>(N0.getOperand(1))->getVT()) {
+ EVT.bitsLT(cast<VTSDNode>(N0.getOperand(1))->getVT())) {
return DAG.getNode(ISD::SIGN_EXTEND_INREG, VT, N0.getOperand(0), N1);
}
if (ISD::isEXTLoad(N0.Val) &&
ISD::isUNINDEXEDLoad(N0.Val) &&
EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
- (!AfterLegalize || TLI.isLoadXLegal(ISD::SEXTLOAD, EVT))) {
+ ((!AfterLegalize && !cast<LoadSDNode>(N0)->isVolatile()) ||
+ TLI.isLoadXLegal(ISD::SEXTLOAD, EVT))) {
LoadSDNode *LN0 = cast<LoadSDNode>(N0);
SDOperand ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, VT, LN0->getChain(),
LN0->getBasePtr(), LN0->getSrcValue(),
if (ISD::isZEXTLoad(N0.Val) && ISD::isUNINDEXEDLoad(N0.Val) &&
N0.hasOneUse() &&
EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
- (!AfterLegalize || TLI.isLoadXLegal(ISD::SEXTLOAD, EVT))) {
+ ((!AfterLegalize && !cast<LoadSDNode>(N0)->isVolatile()) ||
+ TLI.isLoadXLegal(ISD::SEXTLOAD, EVT))) {
LoadSDNode *LN0 = cast<LoadSDNode>(N0);
SDOperand ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, VT, LN0->getChain(),
LN0->getBasePtr(), LN0->getSrcValue(),
// fold (truncate (ext x)) -> (ext x) or (truncate x) or x
if (N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::SIGN_EXTEND||
N0.getOpcode() == ISD::ANY_EXTEND) {
- if (N0.getOperand(0).getValueType() < VT)
+ if (N0.getOperand(0).getValueType().bitsLT(VT))
// if the source is smaller than the dest, we still need an extend
return DAG.getNode(N0.getOpcode(), VT, N0.getOperand(0));
- else if (N0.getOperand(0).getValueType() > VT)
+ else if (N0.getOperand(0).getValueType().bitsGT(VT))
// if the source is larger than the dest, than we just need the truncate
return DAG.getNode(ISD::TRUNCATE, VT, N0.getOperand(0));
else
const MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
if (ISD::isNON_EXTLoad(LD2) &&
LD2->hasOneUse() &&
+ // If both are volatile this would reduce the number of volatile loads.
+ // If one is volatile it might be ok, but play conservative and bail out.
+ !cast<LoadSDNode>(LD1)->isVolatile() &&
+ !cast<LoadSDNode>(LD2)->isVolatile() &&
TLI.isConsecutiveLoad(LD2, LD1, LD1VT.getSizeInBits()/8, 1, MFI)) {
LoadSDNode *LD = cast<LoadSDNode>(LD1);
unsigned Align = LD->getAlignment();
unsigned NewAlign = TLI.getTargetMachine().getTargetData()->
getABITypeAlignment(VT.getTypeForMVT());
- if ((!AfterLegalize || TLI.isTypeLegal(VT)) &&
- TLI.isOperationLegal(ISD::LOAD, VT) && NewAlign <= Align)
+ if (NewAlign <= Align &&
+ (!AfterLegalize || TLI.isOperationLegal(ISD::LOAD, VT)))
return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(),
LD->getSrcValue(), LD->getSrcValueOffset(),
- LD->isVolatile(), Align);
+ false, Align);
}
return SDOperand();
}
// fold (conv (load x)) -> (load (conv*)x)
// If the resultant load doesn't need a higher alignment than the original!
if (ISD::isNormalLoad(N0.Val) && N0.hasOneUse() &&
- TLI.isOperationLegal(ISD::LOAD, VT)) {
+ // Do not change the width of a volatile load.
+ !cast<LoadSDNode>(N0)->isVolatile() &&
+ (!AfterLegalize || TLI.isOperationLegal(ISD::LOAD, VT))) {
LoadSDNode *LN0 = cast<LoadSDNode>(N0);
unsigned Align = TLI.getTargetMachine().getTargetData()->
getABITypeAlignment(VT.getTypeForMVT());
if (Align <= OrigAlign) {
SDOperand Load = DAG.getLoad(VT, LN0->getChain(), LN0->getBasePtr(),
LN0->getSrcValue(), LN0->getSrcValueOffset(),
- LN0->isVolatile(), Align);
+ LN0->isVolatile(), OrigAlign);
AddToWorkList(N);
CombineTo(N0.Val, DAG.getNode(ISD::BIT_CONVERT, N0.getValueType(), Load),
Load.getValue(1));
return Load;
}
}
-
+
// Fold bitconvert(fneg(x)) -> xor(bitconvert(x), signbit)
// Fold bitconvert(fabs(x)) -> and(bitconvert(x), ~signbit)
// This often reduces constant pool loads.
// Convert the input float vector to a int vector where the elements are the
// same sizes.
assert((SrcEltVT == MVT::f32 || SrcEltVT == MVT::f64) && "Unknown FP VT!");
- MVT IntVT = SrcEltVT == MVT::f32 ? MVT::i32 : MVT::i64;
+ MVT IntVT = MVT::getIntegerVT(SrcEltVT.getSizeInBits());
BV = ConstantFoldBIT_CONVERTofBUILD_VECTOR(BV, IntVT).Val;
SrcEltVT = IntVT;
}
// convert to integer first, then to FP of the right size.
if (DstEltVT.isFloatingPoint()) {
assert((DstEltVT == MVT::f32 || DstEltVT == MVT::f64) && "Unknown FP VT!");
- MVT TmpVT = DstEltVT == MVT::f32 ? MVT::i32 : MVT::i64;
+ MVT TmpVT = MVT::getIntegerVT(DstEltVT.getSizeInBits());
SDNode *Tmp = ConstantFoldBIT_CONVERTofBUILD_VECTOR(BV, TmpVT).Val;
// Next, convert to FP elements of the same size.
SDOperand N0 = N->getOperand(0);
ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
MVT VT = N->getValueType(0);
-
+ MVT OpVT = N0.getValueType();
+
// fold (sint_to_fp c1) -> c1fp
- if (N0C && N0.getValueType() != MVT::ppcf128)
+ if (N0C && OpVT != MVT::ppcf128)
return DAG.getNode(ISD::SINT_TO_FP, VT, N0);
+
+ // If the input is a legal type, and SINT_TO_FP is not legal on this target,
+ // but UINT_TO_FP is legal on this target, try to convert.
+ if (!TLI.isOperationLegal(ISD::SINT_TO_FP, OpVT) &&
+ TLI.isOperationLegal(ISD::UINT_TO_FP, OpVT)) {
+ // If the sign bit is known to be zero, we can change this to UINT_TO_FP.
+ if (DAG.SignBitIsZero(N0))
+ return DAG.getNode(ISD::UINT_TO_FP, VT, N0);
+ }
+
+
return SDOperand();
}
SDOperand N0 = N->getOperand(0);
ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
MVT VT = N->getValueType(0);
+ MVT OpVT = N0.getValueType();
// fold (uint_to_fp c1) -> c1fp
- if (N0C && N0.getValueType() != MVT::ppcf128)
+ if (N0C && OpVT != MVT::ppcf128)
return DAG.getNode(ISD::UINT_TO_FP, VT, N0);
+
+ // If the input is a legal type, and UINT_TO_FP is not legal on this target,
+ // but SINT_TO_FP is legal on this target, try to convert.
+ if (!TLI.isOperationLegal(ISD::UINT_TO_FP, OpVT) &&
+ TLI.isOperationLegal(ISD::SINT_TO_FP, OpVT)) {
+ // If the sign bit is known to be zero, we can change this to SINT_TO_FP.
+ if (DAG.SignBitIsZero(N0))
+ return DAG.getNode(ISD::SINT_TO_FP, VT, N0);
+ }
+
return SDOperand();
}
if (N0.getOpcode() == ISD::FP_ROUND && N0.Val->getConstantOperandVal(1) == 1){
SDOperand In = N0.getOperand(0);
if (In.getValueType() == VT) return In;
- if (VT < In.getValueType())
+ if (VT.bitsLT(In.getValueType()))
return DAG.getNode(ISD::FP_ROUND, VT, In, N0.getOperand(1));
return DAG.getNode(ISD::FP_EXTEND, VT, In);
}
// fold (fpext (load x)) -> (fpext (fptrunc (extload x)))
if (ISD::isNON_EXTLoad(N0.Val) && N0.hasOneUse() &&
- (!AfterLegalize||TLI.isLoadXLegal(ISD::EXTLOAD, N0.getValueType()))) {
+ ((!AfterLegalize && !cast<LoadSDNode>(N0)->isVolatile()) ||
+ TLI.isLoadXLegal(ISD::EXTLOAD, N0.getValueType()))) {
LoadSDNode *LN0 = cast<LoadSDNode>(N0);
SDOperand ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, VT, LN0->getChain(),
LN0->getBasePtr(), LN0->getSrcValue(),
ExtLoad.getValue(1));
return SDOperand(N, 0); // Return N so it doesn't get rechecked!
}
-
-
+
return SDOperand();
}
CondCodeSDNode *CC = cast<CondCodeSDNode>(N->getOperand(1));
SDOperand CondLHS = N->getOperand(2), CondRHS = N->getOperand(3);
- // Use SimplifySetCC to simplify SETCC's.
+ // Use SimplifySetCC to simplify SETCC's.
SDOperand Simp = SimplifySetCC(MVT::i1, CondLHS, CondRHS, CC->get(), false);
if (Simp.Val) AddToWorkList(Simp.Val);
}
-/// CombineToPreIndexedLoadStore - Try turning a load / store and a
-/// pre-indexed load / store when the base pointer is a add or subtract
+/// CombineToPreIndexedLoadStore - Try turning a load / store into a
+/// pre-indexed load / store when the base pointer is an add or subtract
/// and it has other uses besides the load / store. After the
/// transformation, the new indexed load / store has effectively folded
/// the add / subtract in and all of its other uses are redirected to the
return true;
}
-/// CombineToPostIndexedLoadStore - Try combine a load / store with a
+/// CombineToPostIndexedLoadStore - Try to combine a load / store with a
/// add / sub of the base pointer node into a post-indexed load / store.
/// The transformation folded the add / subtract into the new indexed
/// load / store effectively and all of its uses are redirected to the
ST->isVolatile(), Align);
}
}
-
+
// If this is a store of a bit convert, store the input value if the
// resultant store does not need a higher alignment than the original.
if (Value.getOpcode() == ISD::BIT_CONVERT && !ST->isTruncatingStore() &&
MVT SVT = Value.getOperand(0).getValueType();
unsigned OrigAlign = TLI.getTargetMachine().getTargetData()->
getABITypeAlignment(SVT.getTypeForMVT());
- if (Align <= OrigAlign && TLI.isOperationLegal(ISD::STORE, SVT))
+ if (Align <= OrigAlign &&
+ ((!AfterLegalize && !ST->isVolatile()) ||
+ TLI.isOperationLegal(ISD::STORE, SVT)))
return DAG.getStore(Chain, Value.getOperand(0), Ptr, ST->getSrcValue(),
- ST->getSrcValueOffset(), ST->isVolatile(), Align);
+ ST->getSrcValueOffset(), ST->isVolatile(), OrigAlign);
}
-
+
// Turn 'store float 1.0, Ptr' -> 'store int 0x12345678, Ptr'
if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Value)) {
+ // NOTE: If the original store is volatile, this transform must not increase
+ // the number of stores. For example, on x86-32 an f64 can be stored in one
+ // processor operation but an i64 (which is not legal) requires two. So the
+ // transform should not be done in this case.
if (Value.getOpcode() != ISD::TargetConstantFP) {
SDOperand Tmp;
switch (CFP->getValueType(0).getSimpleVT()) {
case MVT::ppcf128:
break;
case MVT::f32:
- if (!AfterLegalize || TLI.isTypeLegal(MVT::i32)) {
+ if ((!AfterLegalize && !ST->isVolatile()) ||
+ TLI.isOperationLegal(ISD::STORE, MVT::i32)) {
Tmp = DAG.getConstant((uint32_t)CFP->getValueAPF().
convertToAPInt().getZExtValue(), MVT::i32);
return DAG.getStore(Chain, Tmp, Ptr, ST->getSrcValue(),
}
break;
case MVT::f64:
- if (!AfterLegalize || TLI.isTypeLegal(MVT::i64)) {
+ if ((!AfterLegalize && !ST->isVolatile()) ||
+ TLI.isOperationLegal(ISD::STORE, MVT::i64)) {
Tmp = DAG.getConstant(CFP->getValueAPF().convertToAPInt().
getZExtValue(), MVT::i64);
return DAG.getStore(Chain, Tmp, Ptr, ST->getSrcValue(),
ST->getSrcValueOffset(), ST->isVolatile(),
ST->getAlignment());
- } else if (TLI.isTypeLegal(MVT::i32)) {
+ } else if (!ST->isVolatile() &&
+ TLI.isOperationLegal(ISD::STORE, MVT::i32)) {
// Many FP stores are not made apparent until after legalize, e.g. for
// argument passing. Since this is so common, custom legalize the
// 64-bit integer store into two 32-bit stores.
return Chain;
}
}
-
+
// If this is an FP_ROUND or TRUNC followed by a store, fold this into a
// truncating store. We can do this even if this is already a truncstore.
if ((Value.getOpcode() == ISD::FP_ROUND || Value.getOpcode() == ISD::TRUNCATE)
- && TLI.isTypeLegal(Value.getOperand(0).getValueType()) &&
- Value.Val->hasOneUse() && ST->isUnindexed() &&
+ && Value.Val->hasOneUse() && ST->isUnindexed() &&
TLI.isTruncStoreLegal(Value.getOperand(0).getValueType(),
ST->getMemoryVT())) {
return DAG.getTruncStore(Chain, Value.getOperand(0), Ptr, ST->getSrcValue(),
ST->getSrcValueOffset(), ST->getMemoryVT(),
ST->isVolatile(), ST->getAlignment());
}
-
+
return SDOperand();
}
MVT LVT = EVT;
if (InVec.getOpcode() == ISD::BIT_CONVERT) {
MVT BCVT = InVec.getOperand(0).getValueType();
- if (!BCVT.isVector()
- || (EVT.getSizeInBits() >
- BCVT.getVectorElementType().getSizeInBits()))
+ if (!BCVT.isVector() || EVT.bitsGT(BCVT.getVectorElementType()))
return SDOperand();
InVec = InVec.getOperand(0);
EVT = BCVT.getVectorElementType();
Elt = (Idx < NumElems) ? Idx : Idx - NumElems;
}
}
- if (!LN0 || !LN0->hasOneUse())
+ if (!LN0 || !LN0->hasOneUse() || LN0->isVolatile())
return SDOperand();
unsigned Align = LN0->getAlignment();
// original load.
unsigned NewAlign = TLI.getTargetMachine().getTargetData()->
getABITypeAlignment(LVT.getTypeForMVT());
- if (!TLI.isOperationLegal(ISD::LOAD, LVT) || NewAlign > Align)
+ if (NewAlign > Align || !TLI.isOperationLegal(ISD::LOAD, LVT))
return SDOperand();
Align = NewAlign;
}
// This triggers in things like "select bool X, 10.0, 123.0" after the FP
// constants have been dropped into the constant pool.
if (LHS.getOpcode() == ISD::LOAD &&
+ // Do not let this transformation reduce the number of volatile loads.
+ !cast<LoadSDNode>(LHS)->isVolatile() &&
+ !cast<LoadSDNode>(RHS)->isVolatile() &&
// Token chains must be identical.
LHS.getOperand(0) == RHS.getOperand(0)) {
LoadSDNode *LLD = cast<LoadSDNode>(LHS);
(N1C->getAPIntValue() == 1 && N0 == N2))) { // (a < 1) ? a : 0
MVT XType = N0.getValueType();
MVT AType = N2.getValueType();
- if (XType >= AType) {
+ if (XType.bitsGE(AType)) {
// and (sra X, size(X)-1, A) -> "and (srl X, C2), A" iff A is a
// single-bit constant.
if (N2C && ((N2C->getAPIntValue() & (N2C->getAPIntValue()-1)) == 0)) {
SDOperand ShCt = DAG.getConstant(ShCtV, TLI.getShiftAmountTy());
SDOperand Shift = DAG.getNode(ISD::SRL, XType, N0, ShCt);
AddToWorkList(Shift.Val);
- if (XType > AType) {
+ if (XType.bitsGT(AType)) {
Shift = DAG.getNode(ISD::TRUNCATE, AType, Shift);
AddToWorkList(Shift.Val);
}
DAG.getConstant(XType.getSizeInBits()-1,
TLI.getShiftAmountTy()));
AddToWorkList(Shift.Val);
- if (XType > AType) {
+ if (XType.bitsGT(AType)) {
Shift = DAG.getNode(ISD::TRUNCATE, AType, Shift);
AddToWorkList(Shift.Val);
}
// cast from setcc result type to select result type
if (AfterLegalize) {
SCC = DAG.getSetCC(TLI.getSetCCResultType(N0), N0, N1, CC);
- if (N2.getValueType() < SCC.getValueType())
+ if (N2.getValueType().bitsLT(SCC.getValueType()))
Temp = DAG.getZeroExtendInReg(SCC, N2.getValueType());
else
Temp = DAG.getNode(ISD::ZERO_EXTEND, N2.getValueType(), SCC);
// otherwise, go ahead with the folds.
if (0 && N3C && N3C->isNullValue() && N2C && (N2C->getAPIntValue() == 1ULL)) {
MVT XType = N0.getValueType();
- if (TLI.isOperationLegal(ISD::SETCC, TLI.getSetCCResultType(N0))) {
+ if (!AfterLegalize ||
+ TLI.isOperationLegal(ISD::SETCC, TLI.getSetCCResultType(N0))) {
SDOperand Res = DAG.getSetCC(TLI.getSetCCResultType(N0), N0, N1, CC);
if (Res.getValueType() != VT)
Res = DAG.getNode(ISD::ZERO_EXTEND, VT, Res);
// seteq X, 0 -> srl (ctlz X, log2(size(X)))
if (N1C && N1C->isNullValue() && CC == ISD::SETEQ &&
- TLI.isOperationLegal(ISD::CTLZ, XType)) {
+ (!AfterLegalize ||
+ TLI.isOperationLegal(ISD::CTLZ, XType))) {
SDOperand Ctlz = DAG.getNode(ISD::CTLZ, XType, N0);
return DAG.getNode(ISD::SRL, XType, Ctlz,
DAG.getConstant(Log2_32(XType.getSizeInBits()),