bool CombineToPreIndexedLoadStore(SDNode *N);
bool CombineToPostIndexedLoadStore(SDNode *N);
+ void ReplaceLoadWithPromotedLoad(SDNode *Load, SDNode *ExtLoad);
+ SDValue PromoteOperand(SDValue Op, EVT PVT, bool &Replace);
+ SDValue SExtPromoteOperand(SDValue Op, EVT PVT);
+ SDValue ZExtPromoteOperand(SDValue Op, EVT PVT);
+ SDValue PromoteIntBinOp(SDValue Op);
+ SDValue PromoteIntShiftOp(SDValue Op);
+ SDValue PromoteExtend(SDValue Op);
+ bool PromoteLoad(SDValue Op);
/// combine - call the node-specific routine that knows how to fold each
/// particular type of node. If that doesn't do anything, try the
/// looking for a better chain (aliasing node.)
SDValue FindBetterChain(SDNode *N, SDValue Chain);
- /// getShiftAmountTy - Returns a type large enough to hold any valid
- /// shift amount - before type legalization these can be huge.
- EVT getShiftAmountTy() {
- return LegalTypes ? TLI.getShiftAmountTy() : TLI.getPointerTy();
- }
-
-public:
+ public:
DAGCombiner(SelectionDAG &D, AliasAnalysis &A, CodeGenOpt::Level OL)
- : DAG(D),
- TLI(D.getTargetLoweringInfo()),
- Level(Unrestricted),
- OptLevel(OL),
- LegalOperations(false),
- LegalTypes(false),
- AA(A) {}
+ : DAG(D), TLI(D.getTargetLoweringInfo()), Level(Unrestricted),
+ OptLevel(OL), LegalOperations(false), LegalTypes(false), AA(A) {}
/// Run - runs the dag combiner on all nodes in the work list
void Run(CombineLevel AtLevel);
+
+ SelectionDAG &getDAG() const { return DAG; }
+
+ /// getShiftAmountTy - Returns a type large enough to hold any valid
+ /// shift amount - before type legalization these can be huge.
+ EVT getShiftAmountTy() {
+ return LegalTypes ? TLI.getShiftAmountTy() : TLI.getPointerTy();
+ }
+
+ /// isTypeLegal - This method returns true if we are running before type
+ /// legalization or if the specified VT is legal.
+ bool isTypeLegal(const EVT &VT) {
+ if (!LegalTypes) return true;
+ return TLI.isTypeLegal(VT);
+ }
};
}
return SDValue(N, 0);
}
-void
-DAGCombiner::CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &
- TLO) {
+void DAGCombiner::
+CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) {
// Replace all uses. If any nodes become isomorphic to other nodes and
// are deleted, make sure to remove them from our worklist.
WorkListRemover DeadNodes(*this);
/// it can be simplified or if things it uses can be simplified by bit
/// propagation. If so, return true.
bool DAGCombiner::SimplifyDemandedBits(SDValue Op, const APInt &Demanded) {
- TargetLowering::TargetLoweringOpt TLO(DAG);
+ TargetLowering::TargetLoweringOpt TLO(DAG, LegalTypes, LegalOperations);
APInt KnownZero, KnownOne;
if (!TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
return false;
return true;
}
+void DAGCombiner::ReplaceLoadWithPromotedLoad(SDNode *Load, SDNode *ExtLoad) {
+ DebugLoc dl = Load->getDebugLoc();
+ EVT VT = Load->getValueType(0);
+ SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, VT, SDValue(ExtLoad, 0));
+
+ DEBUG(dbgs() << "\nReplacing.9 ";
+ Load->dump(&DAG);
+ dbgs() << "\nWith: ";
+ Trunc.getNode()->dump(&DAG);
+ dbgs() << '\n');
+ WorkListRemover DeadNodes(*this);
+ DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 0), Trunc, &DeadNodes);
+ DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 1), SDValue(ExtLoad, 1),
+ &DeadNodes);
+ removeFromWorkList(Load);
+ DAG.DeleteNode(Load);
+}
+
+SDValue DAGCombiner::PromoteOperand(SDValue Op, EVT PVT, bool &Replace) {
+ Replace = false;
+ DebugLoc dl = Op.getDebugLoc();
+ if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Op)) {
+ ISD::LoadExtType ExtType =
+ ISD::isNON_EXTLoad(LD) ? ISD::EXTLOAD : LD->getExtensionType();
+ Replace = true;
+ return DAG.getExtLoad(ExtType, dl, PVT,
+ LD->getChain(), LD->getBasePtr(),
+ LD->getSrcValue(), LD->getSrcValueOffset(),
+ LD->getMemoryVT(), LD->isVolatile(),
+ LD->isNonTemporal(), LD->getAlignment());
+ }
+
+ unsigned Opc = Op.getOpcode();
+ switch (Opc) {
+ default: break;
+ case ISD::AssertSext:
+ return DAG.getNode(ISD::AssertSext, dl, PVT,
+ SExtPromoteOperand(Op.getOperand(0), PVT),
+ Op.getOperand(1));
+ case ISD::AssertZext:
+ return DAG.getNode(ISD::AssertZext, dl, PVT,
+ ZExtPromoteOperand(Op.getOperand(0), PVT),
+ Op.getOperand(1));
+ case ISD::Constant: {
+ unsigned ExtOpc =
+ Op.getValueType().isByteSized() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
+ return DAG.getNode(ExtOpc, dl, PVT, Op);
+ }
+ }
+
+ if (!TLI.isOperationLegal(ISD::ANY_EXTEND, PVT))
+ return SDValue();
+ return DAG.getNode(ISD::ANY_EXTEND, dl, PVT, Op);
+}
+
+SDValue DAGCombiner::SExtPromoteOperand(SDValue Op, EVT PVT) {
+ if (!TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, PVT))
+ return SDValue();
+ EVT OldVT = Op.getValueType();
+ DebugLoc dl = Op.getDebugLoc();
+ bool Replace = false;
+ SDValue NewOp = PromoteOperand(Op, PVT, Replace);
+ if (NewOp.getNode() == 0)
+ return SDValue();
+
+ if (Replace)
+ ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode());
+ return DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, NewOp.getValueType(), NewOp,
+ DAG.getValueType(OldVT));
+}
+
+SDValue DAGCombiner::ZExtPromoteOperand(SDValue Op, EVT PVT) {
+ EVT OldVT = Op.getValueType();
+ DebugLoc dl = Op.getDebugLoc();
+ bool Replace = false;
+ SDValue NewOp = PromoteOperand(Op, PVT, Replace);
+ if (NewOp.getNode() == 0)
+ return SDValue();
+
+ if (Replace)
+ ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode());
+ return DAG.getZeroExtendInReg(NewOp, dl, OldVT);
+}
+
+/// PromoteIntBinOp - Promote the specified integer binary operation if the
+/// target indicates it is beneficial. e.g. On x86, it's usually better to
+/// promote i16 operations to i32 since i16 instructions are longer.
+SDValue DAGCombiner::PromoteIntBinOp(SDValue Op) {
+ if (!LegalOperations)
+ return SDValue();
+
+ EVT VT = Op.getValueType();
+ if (VT.isVector() || !VT.isInteger())
+ return SDValue();
+
+ // If operation type is 'undesirable', e.g. i16 on x86, consider
+ // promoting it.
+ unsigned Opc = Op.getOpcode();
+ if (TLI.isTypeDesirableForOp(Opc, VT))
+ return SDValue();
+
+ EVT PVT = VT;
+ // Consult target whether it is a good idea to promote this operation and
+ // what's the right type to promote it to.
+ if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
+ assert(PVT != VT && "Don't know what type to promote to!");
+
+ bool Replace0 = false;
+ SDValue N0 = Op.getOperand(0);
+ SDValue NN0 = PromoteOperand(N0, PVT, Replace0);
+ if (NN0.getNode() == 0)
+ return SDValue();
+
+ bool Replace1 = false;
+ SDValue N1 = Op.getOperand(1);
+ SDValue NN1 = PromoteOperand(N1, PVT, Replace1);
+ if (NN1.getNode() == 0)
+ return SDValue();
+
+ AddToWorkList(NN0.getNode());
+ AddToWorkList(NN1.getNode());
+
+ if (Replace0)
+ ReplaceLoadWithPromotedLoad(N0.getNode(), NN0.getNode());
+ if (Replace1)
+ ReplaceLoadWithPromotedLoad(N1.getNode(), NN1.getNode());
+
+ DebugLoc dl = Op.getDebugLoc();
+ return DAG.getNode(ISD::TRUNCATE, dl, VT,
+ DAG.getNode(Opc, dl, PVT, NN0, NN1));
+ }
+ return SDValue();
+}
+
+/// PromoteIntShiftOp - Promote the specified integer shift operation if the
+/// target indicates it is beneficial. e.g. On x86, it's usually better to
+/// promote i16 operations to i32 since i16 instructions are longer.
+SDValue DAGCombiner::PromoteIntShiftOp(SDValue Op) {
+ if (!LegalOperations)
+ return SDValue();
+
+ EVT VT = Op.getValueType();
+ if (VT.isVector() || !VT.isInteger())
+ return SDValue();
+
+ // If operation type is 'undesirable', e.g. i16 on x86, consider
+ // promoting it.
+ unsigned Opc = Op.getOpcode();
+ if (TLI.isTypeDesirableForOp(Opc, VT))
+ return SDValue();
+
+ EVT PVT = VT;
+ // Consult target whether it is a good idea to promote this operation and
+ // what's the right type to promote it to.
+ if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
+ assert(PVT != VT && "Don't know what type to promote to!");
+
+ bool Replace = false;
+ SDValue N0 = Op.getOperand(0);
+ if (Opc == ISD::SRA)
+ N0 = SExtPromoteOperand(Op.getOperand(0), PVT);
+ else if (Opc == ISD::SRL)
+ N0 = ZExtPromoteOperand(Op.getOperand(0), PVT);
+ else
+ N0 = PromoteOperand(N0, PVT, Replace);
+ if (N0.getNode() == 0)
+ return SDValue();
+
+ AddToWorkList(N0.getNode());
+ if (Replace)
+ ReplaceLoadWithPromotedLoad(Op.getOperand(0).getNode(), N0.getNode());
+
+ DebugLoc dl = Op.getDebugLoc();
+ return DAG.getNode(ISD::TRUNCATE, dl, VT,
+ DAG.getNode(Opc, dl, PVT, N0, Op.getOperand(1)));
+ }
+ return SDValue();
+}
+
+SDValue DAGCombiner::PromoteExtend(SDValue Op) {
+ if (!LegalOperations)
+ return SDValue();
+
+ EVT VT = Op.getValueType();
+ if (VT.isVector() || !VT.isInteger())
+ return SDValue();
+
+ // If operation type is 'undesirable', e.g. i16 on x86, consider
+ // promoting it.
+ unsigned Opc = Op.getOpcode();
+ if (TLI.isTypeDesirableForOp(Opc, VT))
+ return SDValue();
+
+ EVT PVT = VT;
+ // Consult target whether it is a good idea to promote this operation and
+ // what's the right type to promote it to.
+ if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
+ assert(PVT != VT && "Don't know what type to promote to!");
+ // fold (aext (aext x)) -> (aext x)
+ // fold (aext (zext x)) -> (zext x)
+ // fold (aext (sext x)) -> (sext x)
+ return DAG.getNode(Op.getOpcode(), Op.getDebugLoc(), VT, Op.getOperand(0));
+ }
+ return SDValue();
+}
+
+bool DAGCombiner::PromoteLoad(SDValue Op) {
+ if (!LegalOperations)
+ return false;
+
+ EVT VT = Op.getValueType();
+ if (VT.isVector() || !VT.isInteger())
+ return false;
+
+ // If operation type is 'undesirable', e.g. i16 on x86, consider
+ // promoting it.
+ unsigned Opc = Op.getOpcode();
+ if (TLI.isTypeDesirableForOp(Opc, VT))
+ return false;
+
+ EVT PVT = VT;
+ // Consult target whether it is a good idea to promote this operation and
+ // what's the right type to promote it to.
+ if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
+ assert(PVT != VT && "Don't know what type to promote to!");
+
+ DebugLoc dl = Op.getDebugLoc();
+ SDNode *N = Op.getNode();
+ LoadSDNode *LD = cast<LoadSDNode>(N);
+ ISD::LoadExtType ExtType =
+ ISD::isNON_EXTLoad(LD) ? ISD::EXTLOAD : LD->getExtensionType();
+ SDValue NewLD = DAG.getExtLoad(ExtType, dl, PVT,
+ LD->getChain(), LD->getBasePtr(),
+ LD->getSrcValue(), LD->getSrcValueOffset(),
+ LD->getMemoryVT(), LD->isVolatile(),
+ LD->isNonTemporal(), LD->getAlignment());
+ SDValue Result = DAG.getNode(ISD::TRUNCATE, dl, VT, NewLD);
+
+ DEBUG(dbgs() << "\nPromoting ";
+ N->dump(&DAG);
+ dbgs() << "\nTo: ";
+ Result.getNode()->dump(&DAG);
+ dbgs() << '\n');
+ WorkListRemover DeadNodes(*this);
+ DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result, &DeadNodes);
+ DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), NewLD.getValue(1), &DeadNodes);
+ removeFromWorkList(N);
+ DAG.DeleteNode(N);
+ return true;
+ }
+ return false;
+}
+
+
//===----------------------------------------------------------------------===//
// Main DAG Combiner implementation
//===----------------------------------------------------------------------===//
if (VT.isInteger() && !VT.isVector()) {
APInt LHSZero, LHSOne;
APInt RHSZero, RHSOne;
- APInt Mask = APInt::getAllOnesValue(VT.getSizeInBits());
+ APInt Mask = APInt::getAllOnesValue(VT.getScalarType().getSizeInBits());
DAG.ComputeMaskedBits(N0, Mask, LHSZero, LHSOne);
if (LHSZero.getBoolValue()) {
if (Result.getNode()) return Result;
}
- return SDValue();
+ // fold (add x, shl(0 - y, n)) -> sub(x, shl(y, n))
+ if (N1.getOpcode() == ISD::SHL &&
+ N1.getOperand(0).getOpcode() == ISD::SUB)
+ if (ConstantSDNode *C =
+ dyn_cast<ConstantSDNode>(N1.getOperand(0).getOperand(0)))
+ if (C->getAPIntValue() == 0)
+ return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N0,
+ DAG.getNode(ISD::SHL, N->getDebugLoc(), VT,
+ N1.getOperand(0).getOperand(1),
+ N1.getOperand(1)));
+ if (N0.getOpcode() == ISD::SHL &&
+ N0.getOperand(0).getOpcode() == ISD::SUB)
+ if (ConstantSDNode *C =
+ dyn_cast<ConstantSDNode>(N0.getOperand(0).getOperand(0)))
+ if (C->getAPIntValue() == 0)
+ return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N1,
+ DAG.getNode(ISD::SHL, N->getDebugLoc(), VT,
+ N0.getOperand(0).getOperand(1),
+ N0.getOperand(1)));
+
+ return PromoteIntBinOp(SDValue(N, 0));
}
SDValue DAGCombiner::visitADDC(SDNode *N) {
// fold (addc a, b) -> (or a, b), CARRY_FALSE iff a and b share no bits.
APInt LHSZero, LHSOne;
APInt RHSZero, RHSOne;
- APInt Mask = APInt::getAllOnesValue(VT.getSizeInBits());
+ APInt Mask = APInt::getAllOnesValue(VT.getScalarType().getSizeInBits());
DAG.ComputeMaskedBits(N0, Mask, LHSZero, LHSOne);
if (LHSZero.getBoolValue()) {
if (N1C)
return DAG.getNode(ISD::ADD, N->getDebugLoc(), VT, N0,
DAG.getConstant(-N1C->getAPIntValue(), VT));
+ // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1)
+ if (N0C && N0C->isAllOnesValue())
+ return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT, N1, N0);
// fold (A+B)-A -> B
if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1)
return N0.getOperand(1);
VT);
}
- return SDValue();
+ return PromoteIntBinOp(SDValue(N, 0));
}
SDValue DAGCombiner::visitMUL(SDNode *N) {
if (RMUL.getNode() != 0)
return RMUL;
- return SDValue();
+ return PromoteIntBinOp(SDValue(N, 0));
}
SDValue DAGCombiner::visitSDIV(SDNode *N) {
// into a vsetcc.
EVT Op0VT = N0.getOperand(0).getValueType();
if ((N0.getOpcode() == ISD::ZERO_EXTEND ||
- N0.getOpcode() == ISD::ANY_EXTEND ||
N0.getOpcode() == ISD::SIGN_EXTEND ||
+ // Avoid infinite looping with PromoteIntBinOp.
+ (N0.getOpcode() == ISD::ANY_EXTEND &&
+ (!LegalTypes || TLI.isTypeDesirableForOp(N->getOpcode(), Op0VT))) ||
(N0.getOpcode() == ISD::TRUNCATE && TLI.isTypeLegal(Op0VT))) &&
!VT.isVector() &&
Op0VT == N1.getOperand(0).getValueType() &&
ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
EVT VT = N1.getValueType();
- unsigned BitWidth = VT.getSizeInBits();
+ unsigned BitWidth = VT.getScalarType().getSizeInBits();
// fold vector ops
if (VT.isVector()) {
SDValue RAND = ReassociateOps(ISD::AND, N->getDebugLoc(), N0, N1);
if (RAND.getNode() != 0)
return RAND;
- // fold (and (or x, 0xFFFF), 0xFF) -> 0xFF
+ // fold (and (or x, C), D) -> D if (C & D) == D
if (N1C && N0.getOpcode() == ISD::OR)
if (ConstantSDNode *ORI = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
if ((ORI->getAPIntValue() & N1C->getAPIntValue()) == N1C->getAPIntValue())
EVT MemVT = LN0->getMemoryVT();
// If we zero all the possible extended bits, then we can turn this into
// a zextload if we are running before legalize or the operation is legal.
- unsigned BitWidth = N1.getValueSizeInBits();
+ unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits();
if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
- BitWidth - MemVT.getSizeInBits())) &&
+ BitWidth - MemVT.getScalarType().getSizeInBits())) &&
((!LegalOperations && !LN0->isVolatile()) ||
TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT))) {
SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, N0.getDebugLoc(), VT,
LN0->getChain(), LN0->getBasePtr(),
LN0->getSrcValue(),
LN0->getSrcValueOffset(), MemVT,
- LN0->isVolatile(), LN0->getAlignment());
+ LN0->isVolatile(), LN0->isNonTemporal(),
+ LN0->getAlignment());
AddToWorkList(N);
CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
return SDValue(N, 0); // Return N so it doesn't get rechecked!
EVT MemVT = LN0->getMemoryVT();
// If we zero all the possible extended bits, then we can turn this into
// a zextload if we are running before legalize or the operation is legal.
- unsigned BitWidth = N1.getValueSizeInBits();
+ unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits();
if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
- BitWidth - MemVT.getSizeInBits())) &&
+ BitWidth - MemVT.getScalarType().getSizeInBits())) &&
((!LegalOperations && !LN0->isVolatile()) ||
TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT))) {
SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, N0.getDebugLoc(), VT,
LN0->getChain(),
LN0->getBasePtr(), LN0->getSrcValue(),
LN0->getSrcValueOffset(), MemVT,
- LN0->isVolatile(), LN0->getAlignment());
+ LN0->isVolatile(), LN0->isNonTemporal(),
+ LN0->getAlignment());
AddToWorkList(N);
CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
return SDValue(N, 0); // Return N so it doesn't get rechecked!
DAG.getExtLoad(ISD::ZEXTLOAD, LN0->getDebugLoc(), LoadResultTy,
LN0->getChain(), LN0->getBasePtr(),
LN0->getSrcValue(), LN0->getSrcValueOffset(),
- ExtVT, LN0->isVolatile(), LN0->getAlignment());
+ ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
+ LN0->getAlignment());
AddToWorkList(N);
CombineTo(LN0, NewLoad, NewLoad.getValue(1));
return SDValue(N, 0); // Return N so it doesn't get rechecked!
DAG.getExtLoad(ISD::ZEXTLOAD, LN0->getDebugLoc(), LoadResultTy,
LN0->getChain(), NewPtr,
LN0->getSrcValue(), LN0->getSrcValueOffset(),
- ExtVT, LN0->isVolatile(), Alignment);
+ ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
+ Alignment);
AddToWorkList(N);
CombineTo(LN0, Load, Load.getValue(1));
return SDValue(N, 0); // Return N so it doesn't get rechecked!
}
}
- return SDValue();
+ return PromoteIntBinOp(SDValue(N, 0));
}
SDValue DAGCombiner::visitOR(SDNode *N) {
if (ROR.getNode() != 0)
return ROR;
// Canonicalize (or (and X, c1), c2) -> (and (or X, c2), c1|c2)
+ // iff (c1 & c2) == 0.
if (N1C && N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
isa<ConstantSDNode>(N0.getOperand(1))) {
ConstantSDNode *C1 = cast<ConstantSDNode>(N0.getOperand(1));
- return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
- DAG.getNode(ISD::OR, N0.getDebugLoc(), VT,
- N0.getOperand(0), N1),
- DAG.FoldConstantArithmetic(ISD::OR, VT, N1C, C1));
+ if ((C1->getAPIntValue() & N1C->getAPIntValue()) != 0)
+ return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
+ DAG.getNode(ISD::OR, N0.getDebugLoc(), VT,
+ N0.getOperand(0), N1),
+ DAG.FoldConstantArithmetic(ISD::OR, VT, N1C, C1));
}
// fold (or (setcc x), (setcc y)) -> (setcc (or x, y))
if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
if (SDNode *Rot = MatchRotate(N0, N1, N->getDebugLoc()))
return SDValue(Rot, 0);
- return SDValue();
+ return PromoteIntBinOp(SDValue(N, 0));
}
/// MatchRotateHalf - Match "(X shl/srl V1) & V2" where V2 may not be present.
SimplifyDemandedBits(SDValue(N, 0)))
return SDValue(N, 0);
- return SDValue();
+ return PromoteIntBinOp(SDValue(N, 0));
}
/// visitShiftByConstant - Handle transforms common to the three shifts, when
HiBitsMask);
}
- return N1C ? visitShiftByConstant(N, N1C->getZExtValue()) : SDValue();
+ if (N1C) {
+ SDValue NewSHL = visitShiftByConstant(N, N1C->getZExtValue());
+ if (NewSHL.getNode())
+ return NewSHL;
+ }
+
+ return PromoteIntShiftOp(SDValue(N, 0));
}
SDValue DAGCombiner::visitSRA(SDNode *N) {
// 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,
+ // on that type, and the truncate to that type is both legal and free,
// perform the transform.
if ((ShiftAmt > 0) &&
TLI.isOperationLegalOrCustom(ISD::SIGN_EXTEND, TruncVT) &&
if (DAG.SignBitIsZero(N0))
return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0, N1);
- return N1C ? visitShiftByConstant(N, N1C->getZExtValue()) : SDValue();
+ if (N1C) {
+ SDValue NewSRA = visitShiftByConstant(N, N1C->getZExtValue());
+ if (NewSRA.getNode())
+ return NewSRA;
+ }
+
+ return PromoteIntShiftOp(SDValue(N, 0));
}
SDValue DAGCombiner::visitSRL(SDNode *N) {
return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0.getOperand(0),
DAG.getConstant(c1 + c2, N1.getValueType()));
}
+
+ // fold (srl (shl x, c), c) -> (and x, cst2)
+ if (N1C && N0.getOpcode() == ISD::SHL && N0.getOperand(1) == N1 &&
+ N0.getValueSizeInBits() <= 64) {
+ uint64_t ShAmt = N1C->getZExtValue()+64-N0.getValueSizeInBits();
+ return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0.getOperand(0),
+ DAG.getConstant(~0ULL >> ShAmt, VT));
+ }
+
// fold (srl (anyextend x), c) -> (anyextend (srl x, c))
if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
if (N1C->getZExtValue() >= SmallVT.getSizeInBits())
return DAG.getUNDEF(VT);
- SDValue SmallShift = DAG.getNode(ISD::SRL, N0.getDebugLoc(), SmallVT,
- N0.getOperand(0), N1);
- AddToWorkList(SmallShift.getNode());
- return DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, SmallShift);
+ if (!LegalTypes || TLI.isTypeDesirableForOp(ISD::SRL, SmallVT)) {
+ SDValue SmallShift = DAG.getNode(ISD::SRL, N0.getDebugLoc(), SmallVT,
+ N0.getOperand(0), N1);
+ AddToWorkList(SmallShift.getNode());
+ return DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, SmallShift);
+ }
}
// fold (srl (sra X, Y), 31) -> (srl X, 31). This srl only looks at the sign
if (N1C && N0.getOpcode() == ISD::CTLZ &&
N1C->getAPIntValue() == Log2_32(VT.getSizeInBits())) {
APInt KnownZero, KnownOne;
- APInt Mask = APInt::getAllOnesValue(VT.getSizeInBits());
+ APInt Mask = APInt::getAllOnesValue(VT.getScalarType().getSizeInBits());
DAG.ComputeMaskedBits(N0.getOperand(0), Mask, KnownZero, KnownOne);
// If any of the input bits are KnownOne, then the input couldn't be all
}
}
- return SDValue();
+ return PromoteIntShiftOp(SDValue(N, 0));
}
SDValue DAGCombiner::visitCTLZ(SDNode *N) {
LN0->getBasePtr(), LN0->getSrcValue(),
LN0->getSrcValueOffset(),
N0.getValueType(),
- LN0->isVolatile(), LN0->getAlignment());
+ LN0->isVolatile(), LN0->isNonTemporal(),
+ LN0->getAlignment());
CombineTo(N, ExtLoad);
SDValue Trunc = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
N0.getValueType(), ExtLoad);
LN0->getChain(),
LN0->getBasePtr(), LN0->getSrcValue(),
LN0->getSrcValueOffset(), MemVT,
- LN0->isVolatile(), LN0->getAlignment());
+ LN0->isVolatile(), LN0->isNonTemporal(),
+ LN0->getAlignment());
CombineTo(N, ExtLoad);
CombineTo(N0.getNode(),
DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
}
// sext(setcc x, y, cc) -> (select_cc x, y, -1, 0, cc)
+ unsigned ElementWidth = VT.getScalarType().getSizeInBits();
SDValue NegOne =
- DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), VT);
+ DAG.getConstant(APInt::getAllOnesValue(ElementWidth), VT);
SDValue SCC =
SimplifySelectCC(N->getDebugLoc(), N0.getOperand(0), N0.getOperand(1),
NegOne, DAG.getConstant(0, VT),
cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
if (SCC.getNode()) return SCC;
+ if (!LegalOperations ||
+ TLI.isOperationLegal(ISD::SETCC, TLI.getSetCCResultType(VT)))
+ return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
+ DAG.getSetCC(N->getDebugLoc(),
+ TLI.getSetCCResultType(VT),
+ N0.getOperand(0), N0.getOperand(1),
+ cast<CondCodeSDNode>(N0.getOperand(2))->get()),
+ NegOne, DAG.getConstant(0, VT));
}
DAG.SignBitIsZero(N0))
return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT, N0);
- return SDValue();
+ return PromoteExtend(SDValue(N, 0));
}
SDValue DAGCombiner::visitZERO_EXTEND(SDNode *N) {
LN0->getBasePtr(), LN0->getSrcValue(),
LN0->getSrcValueOffset(),
N0.getValueType(),
- LN0->isVolatile(), LN0->getAlignment());
+ LN0->isVolatile(), LN0->isNonTemporal(),
+ LN0->getAlignment());
CombineTo(N, ExtLoad);
SDValue Trunc = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
N0.getValueType(), ExtLoad);
LN0->getChain(),
LN0->getBasePtr(), LN0->getSrcValue(),
LN0->getSrcValueOffset(), MemVT,
- LN0->isVolatile(), LN0->getAlignment());
+ LN0->isVolatile(), LN0->isNonTemporal(),
+ LN0->getAlignment());
CombineTo(N, ExtLoad);
CombineTo(N0.getNode(),
DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(), N0.getValueType(),
N0.getOperand(1)));
}
- return SDValue();
+ return PromoteExtend(SDValue(N, 0));
}
SDValue DAGCombiner::visitANY_EXTEND(SDNode *N) {
LN0->getBasePtr(), LN0->getSrcValue(),
LN0->getSrcValueOffset(),
N0.getValueType(),
- LN0->isVolatile(), LN0->getAlignment());
+ LN0->isVolatile(), LN0->isNonTemporal(),
+ LN0->getAlignment());
CombineTo(N, ExtLoad);
SDValue Trunc = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
N0.getValueType(), ExtLoad);
VT, LN0->getChain(), LN0->getBasePtr(),
LN0->getSrcValue(),
LN0->getSrcValueOffset(), MemVT,
- LN0->isVolatile(), LN0->getAlignment());
+ LN0->isVolatile(), LN0->isNonTemporal(),
+ LN0->getAlignment());
CombineTo(N, ExtLoad);
CombineTo(N0.getNode(),
DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
return SCC;
}
- return SDValue();
+ return PromoteExtend(SDValue(N, 0));
}
/// GetDemandedBits - See if the specified operand can be simplified with the
// 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 (isa<LoadSDNode>(N0) && N0.hasOneUse() && ExtVT.isRound() &&
- cast<LoadSDNode>(N0)->getMemoryVT().getSizeInBits() > EVTBits &&
+ cast<LoadSDNode>(N0)->getMemoryVT().getSizeInBits() >= EVTBits &&
// Do not change the width of a volatile load.
!cast<LoadSDNode>(N0)->isVolatile()) {
LoadSDNode *LN0 = cast<LoadSDNode>(N0);
SDValue Load = (ExtType == ISD::NON_EXTLOAD)
? DAG.getLoad(VT, N0.getDebugLoc(), LN0->getChain(), NewPtr,
LN0->getSrcValue(), LN0->getSrcValueOffset() + PtrOff,
- LN0->isVolatile(), NewAlign)
+ LN0->isVolatile(), LN0->isNonTemporal(), NewAlign)
: DAG.getExtLoad(ExtType, N0.getDebugLoc(), VT, LN0->getChain(), NewPtr,
LN0->getSrcValue(), LN0->getSrcValueOffset() + PtrOff,
- ExtVT, LN0->isVolatile(), NewAlign);
+ ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
+ NewAlign);
// Replace the old load's chain with the new load's chain.
WorkListRemover DeadNodes(*this);
// if x is small enough.
if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND) {
SDValue N00 = N0.getOperand(0);
- if (N00.getValueType().getScalarType().getSizeInBits() < EVTBits)
+ if (N00.getValueType().getScalarType().getSizeInBits() <= EVTBits &&
+ (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND, VT)))
return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT, N00, N1);
}
LN0->getChain(),
LN0->getBasePtr(), LN0->getSrcValue(),
LN0->getSrcValueOffset(), EVT,
- LN0->isVolatile(), LN0->getAlignment());
+ LN0->isVolatile(), LN0->isNonTemporal(),
+ LN0->getAlignment());
CombineTo(N, ExtLoad);
CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
return SDValue(N, 0); // Return N so it doesn't get rechecked!
LN0->getChain(),
LN0->getBasePtr(), LN0->getSrcValue(),
LN0->getSrcValueOffset(), EVT,
- LN0->isVolatile(), LN0->getAlignment());
+ LN0->isVolatile(), LN0->isNonTemporal(),
+ LN0->getAlignment());
CombineTo(N, ExtLoad);
CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
return SDValue(N, 0); // Return N so it doesn't get rechecked!
if (N0.getOpcode() == ISD::TRUNCATE)
return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, N0.getOperand(0));
// fold (truncate (ext x)) -> (ext x) or (truncate x) or x
- if (N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::SIGN_EXTEND||
+ if (N0.getOpcode() == ISD::ZERO_EXTEND ||
+ N0.getOpcode() == ISD::SIGN_EXTEND ||
N0.getOpcode() == ISD::ANY_EXTEND) {
if (N0.getOperand(0).getValueType().bitsLT(VT))
// if the source is smaller than the dest, we still need an extend
// fold (truncate (load x)) -> (smaller load x)
// fold (truncate (srl (load x), c)) -> (smaller load (x+c/evtbits))
- return ReduceLoadWidth(N);
+ if (!LegalTypes || TLI.isTypeDesirableForOp(N0.getOpcode(), VT))
+ return ReduceLoadWidth(N);
+ return SDValue();
}
static SDNode *getBuildPairElt(SDNode *N, unsigned i) {
(!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT)))
return DAG.getLoad(VT, N->getDebugLoc(), LD1->getChain(),
LD1->getBasePtr(), LD1->getSrcValue(),
- LD1->getSrcValueOffset(), false, Align);
+ LD1->getSrcValueOffset(), false, false, Align);
}
return SDValue();
SDValue Load = DAG.getLoad(VT, N->getDebugLoc(), LN0->getChain(),
LN0->getBasePtr(),
LN0->getSrcValue(), LN0->getSrcValueOffset(),
- LN0->isVolatile(), OrigAlign);
+ LN0->isVolatile(), LN0->isNonTemporal(),
+ OrigAlign);
AddToWorkList(N);
CombineTo(N0.getNode(),
DAG.getNode(ISD::BIT_CONVERT, N0.getDebugLoc(),
VT.isInteger() && !VT.isVector()) {
unsigned OrigXWidth = N0.getOperand(1).getValueType().getSizeInBits();
EVT IntXVT = EVT::getIntegerVT(*DAG.getContext(), OrigXWidth);
- if (TLI.isTypeLegal(IntXVT) || !LegalTypes) {
+ if (isTypeLegal(IntXVT)) {
SDValue X = DAG.getNode(ISD::BIT_CONVERT, N0.getDebugLoc(),
IntXVT, N0.getOperand(1));
AddToWorkList(X.getNode());
if (Op.getOpcode() == ISD::UNDEF) continue;
EltIsUndef = false;
- NewBits |= (APInt(cast<ConstantSDNode>(Op)->getAPIntValue()).
- zextOrTrunc(SrcBitSize).zext(DstBitSize));
+ NewBits |= APInt(cast<ConstantSDNode>(Op)->getAPIntValue()).
+ zextOrTrunc(SrcBitSize).zext(DstBitSize);
}
if (EltIsUndef)
ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
// fold (fp_round_inreg c1fp) -> c1fp
- if (N0CFP && (TLI.isTypeLegal(EVT) || !LegalTypes)) {
+ if (N0CFP && isTypeLegal(EVT)) {
SDValue Round = DAG.getConstantFP(*N0CFP->getConstantFPValue(), EVT);
return DAG.getNode(ISD::FP_EXTEND, N->getDebugLoc(), VT, Round);
}
LN0->getBasePtr(), LN0->getSrcValue(),
LN0->getSrcValueOffset(),
N0.getValueType(),
- LN0->isVolatile(), LN0->getAlignment());
+ LN0->isVolatile(), LN0->isNonTemporal(),
+ LN0->getAlignment());
CombineTo(N, ExtLoad);
CombineTo(N0.getNode(),
DAG.getNode(ISD::FP_ROUND, N0.getDebugLoc(),
SDNode *Trunc = 0;
if (N1.getOpcode() == ISD::TRUNCATE && N1.hasOneUse()) {
- // Look pass truncate.
+ // Look past truncate.
Trunc = N1.getNode();
N1 = N1.getOperand(0);
}
DAG.DeleteNode(Trunc);
}
// Replace the uses of SRL with SETCC
- DAG.ReplaceAllUsesOfValueWith(N1, SetCC);
+ WorkListRemover DeadNodes(*this);
+ DAG.ReplaceAllUsesOfValueWith(N1, SetCC, &DeadNodes);
removeFromWorkList(N1.getNode());
DAG.DeleteNode(N1.getNode());
return SDValue(N, 0); // Return N so it doesn't get rechecked!
}
}
}
+
+ // Transform br(xor(x, y)) -> br(x != y)
+ // Transform br(xor(xor(x,y), 1)) -> br (x == y)
+ if (N1.hasOneUse() && N1.getOpcode() == ISD::XOR) {
+ SDNode *TheXor = N1.getNode();
+ SDValue Op0 = TheXor->getOperand(0);
+ SDValue Op1 = TheXor->getOperand(1);
+ if (Op0.getOpcode() == Op1.getOpcode()) {
+ // Avoid missing important xor optimizations.
+ SDValue Tmp = visitXOR(TheXor);
+ if (Tmp.getNode() && Tmp.getNode() != TheXor) {
+ DEBUG(dbgs() << "\nReplacing.8 ";
+ TheXor->dump(&DAG);
+ dbgs() << "\nWith: ";
+ Tmp.getNode()->dump(&DAG);
+ dbgs() << '\n');
+ WorkListRemover DeadNodes(*this);
+ DAG.ReplaceAllUsesOfValueWith(N1, Tmp, &DeadNodes);
+ removeFromWorkList(TheXor);
+ DAG.DeleteNode(TheXor);
+ return DAG.getNode(ISD::BRCOND, N->getDebugLoc(),
+ MVT::Other, Chain, Tmp, N2);
+ }
+ }
+
+ if (Op0.getOpcode() != ISD::SETCC && Op1.getOpcode() != ISD::SETCC) {
+ bool Equal = false;
+ if (ConstantSDNode *RHSCI = dyn_cast<ConstantSDNode>(Op0))
+ if (RHSCI->getAPIntValue() == 1 && Op0.hasOneUse() &&
+ Op0.getOpcode() == ISD::XOR) {
+ TheXor = Op0.getNode();
+ Equal = true;
+ }
+
+ SDValue NodeToReplace = Trunc ? SDValue(Trunc, 0) : N1;
+
+ EVT SetCCVT = NodeToReplace.getValueType();
+ if (LegalTypes)
+ SetCCVT = TLI.getSetCCResultType(SetCCVT);
+ SDValue SetCC = DAG.getSetCC(TheXor->getDebugLoc(),
+ SetCCVT,
+ Op0, Op1,
+ Equal ? ISD::SETEQ : ISD::SETNE);
+ // Replace the uses of XOR with SETCC
+ WorkListRemover DeadNodes(*this);
+ DAG.ReplaceAllUsesOfValueWith(NodeToReplace, SetCC, &DeadNodes);
+ removeFromWorkList(NodeToReplace.getNode());
+ DAG.DeleteNode(NodeToReplace.getNode());
+ return DAG.getNode(ISD::BRCOND, N->getDebugLoc(),
+ MVT::Other, Chain, SetCC, N2);
+ }
+ }
return SDValue();
}
SDValue Chain = LD->getChain();
SDValue Ptr = LD->getBasePtr();
- // Try to infer better alignment information than the load already has.
- if (OptLevel != CodeGenOpt::None && LD->isUnindexed()) {
- if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
- if (Align > LD->getAlignment())
- return DAG.getExtLoad(LD->getExtensionType(), N->getDebugLoc(),
- LD->getValueType(0),
- Chain, Ptr, LD->getSrcValue(),
- LD->getSrcValueOffset(), LD->getMemoryVT(),
- LD->isVolatile(), Align);
- }
- }
-
// If load is not volatile and there are no uses of the loaded value (and
// the updated indexed value in case of indexed loads), change uses of the
// chain value into uses of the chain input (i.e. delete the dead load).
assert(N->getValueType(2) == MVT::Other && "Malformed indexed loads?");
if (N->hasNUsesOfValue(0, 0) && N->hasNUsesOfValue(0, 1)) {
SDValue Undef = DAG.getUNDEF(N->getValueType(0));
- DEBUG(dbgs() << "\nReplacing.6 ";
+ DEBUG(dbgs() << "\nReplacing.7 ";
N->dump(&DAG);
dbgs() << "\nWith: ";
Undef.getNode()->dump(&DAG);
}
}
+ // Try to infer better alignment information than the load already has.
+ if (OptLevel != CodeGenOpt::None && LD->isUnindexed()) {
+ if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
+ if (Align > LD->getAlignment())
+ return DAG.getExtLoad(LD->getExtensionType(), N->getDebugLoc(),
+ LD->getValueType(0),
+ Chain, Ptr, LD->getSrcValue(),
+ LD->getSrcValueOffset(), LD->getMemoryVT(),
+ LD->isVolatile(), LD->isNonTemporal(), Align);
+ }
+ }
+
if (CombinerAA) {
// Walk up chain skipping non-aliasing memory nodes.
SDValue BetterChain = FindBetterChain(N, Chain);
ReplLoad = DAG.getLoad(N->getValueType(0), LD->getDebugLoc(),
BetterChain, Ptr,
LD->getSrcValue(), LD->getSrcValueOffset(),
- LD->isVolatile(), LD->getAlignment());
+ LD->isVolatile(), LD->isNonTemporal(),
+ LD->getAlignment());
} else {
ReplLoad = DAG.getExtLoad(LD->getExtensionType(), LD->getDebugLoc(),
LD->getValueType(0),
LD->getSrcValueOffset(),
LD->getMemoryVT(),
LD->isVolatile(),
+ LD->isNonTemporal(),
LD->getAlignment());
}
if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
return SDValue(N, 0);
+ if (PromoteLoad(SDValue(N, 0)))
+ return SDValue(N, 0);
return SDValue();
}
+/// CheckForMaskedLoad - Check to see if V is (and load (ptr), imm), where the
+/// load is having specific bytes cleared out. If so, return the byte size
+/// being masked out and the shift amount.
+static std::pair<unsigned, unsigned>
+CheckForMaskedLoad(SDValue V, SDValue Ptr, SDValue Chain) {
+ std::pair<unsigned, unsigned> Result(0, 0);
+
+ // Check for the structure we're looking for.
+ if (V->getOpcode() != ISD::AND ||
+ !isa<ConstantSDNode>(V->getOperand(1)) ||
+ !ISD::isNormalLoad(V->getOperand(0).getNode()))
+ return Result;
+
+ // Check the chain and pointer.
+ LoadSDNode *LD = cast<LoadSDNode>(V->getOperand(0));
+ if (LD->getBasePtr() != Ptr) return Result; // Not from same pointer.
+
+ // The store should be chained directly to the load or be an operand of a
+ // tokenfactor.
+ if (LD == Chain.getNode())
+ ; // ok.
+ else if (Chain->getOpcode() != ISD::TokenFactor)
+ return Result; // Fail.
+ else {
+ bool isOk = false;
+ for (unsigned i = 0, e = Chain->getNumOperands(); i != e; ++i)
+ if (Chain->getOperand(i).getNode() == LD) {
+ isOk = true;
+ break;
+ }
+ if (!isOk) return Result;
+ }
+
+ // This only handles simple types.
+ if (V.getValueType() != MVT::i16 &&
+ V.getValueType() != MVT::i32 &&
+ V.getValueType() != MVT::i64)
+ return Result;
+
+ // Check the constant mask. Invert it so that the bits being masked out are
+ // 0 and the bits being kept are 1. Use getSExtValue so that leading bits
+ // follow the sign bit for uniformity.
+ uint64_t NotMask = ~cast<ConstantSDNode>(V->getOperand(1))->getSExtValue();
+ unsigned NotMaskLZ = CountLeadingZeros_64(NotMask);
+ if (NotMaskLZ & 7) return Result; // Must be multiple of a byte.
+ unsigned NotMaskTZ = CountTrailingZeros_64(NotMask);
+ if (NotMaskTZ & 7) return Result; // Must be multiple of a byte.
+ if (NotMaskLZ == 64) return Result; // All zero mask.
+
+ // See if we have a continuous run of bits. If so, we have 0*1+0*
+ if (CountTrailingOnes_64(NotMask >> NotMaskTZ)+NotMaskTZ+NotMaskLZ != 64)
+ return Result;
+
+ // Adjust NotMaskLZ down to be from the actual size of the int instead of i64.
+ if (V.getValueType() != MVT::i64 && NotMaskLZ)
+ NotMaskLZ -= 64-V.getValueSizeInBits();
+
+ unsigned MaskedBytes = (V.getValueSizeInBits()-NotMaskLZ-NotMaskTZ)/8;
+ switch (MaskedBytes) {
+ case 1:
+ case 2:
+ case 4: break;
+ default: return Result; // All one mask, or 5-byte mask.
+ }
+
+ // Verify that the first bit starts at a multiple of mask so that the access
+ // is aligned the same as the access width.
+ if (NotMaskTZ && NotMaskTZ/8 % MaskedBytes) return Result;
+
+ Result.first = MaskedBytes;
+ Result.second = NotMaskTZ/8;
+ return Result;
+}
+
+
+/// ShrinkLoadReplaceStoreWithStore - Check to see if IVal is something that
+/// provides a value as specified by MaskInfo. If so, replace the specified
+/// store with a narrower store of truncated IVal.
+static SDNode *
+ShrinkLoadReplaceStoreWithStore(const std::pair<unsigned, unsigned> &MaskInfo,
+ SDValue IVal, StoreSDNode *St,
+ DAGCombiner *DC) {
+ unsigned NumBytes = MaskInfo.first;
+ unsigned ByteShift = MaskInfo.second;
+ SelectionDAG &DAG = DC->getDAG();
+
+ // Check to see if IVal is all zeros in the part being masked in by the 'or'
+ // that uses this. If not, this is not a replacement.
+ APInt Mask = ~APInt::getBitsSet(IVal.getValueSizeInBits(),
+ ByteShift*8, (ByteShift+NumBytes)*8);
+ if (!DAG.MaskedValueIsZero(IVal, Mask)) return 0;
+
+ // Check that it is legal on the target to do this. It is legal if the new
+ // VT we're shrinking to (i8/i16/i32) is legal or we're still before type
+ // legalization.
+ MVT VT = MVT::getIntegerVT(NumBytes*8);
+ if (!DC->isTypeLegal(VT))
+ return 0;
+
+ // Okay, we can do this! Replace the 'St' store with a store of IVal that is
+ // shifted by ByteShift and truncated down to NumBytes.
+ if (ByteShift)
+ IVal = DAG.getNode(ISD::SRL, IVal->getDebugLoc(), IVal.getValueType(), IVal,
+ DAG.getConstant(ByteShift*8, DC->getShiftAmountTy()));
+
+ // Figure out the offset for the store and the alignment of the access.
+ unsigned StOffset;
+ unsigned NewAlign = St->getAlignment();
+
+ if (DAG.getTargetLoweringInfo().isLittleEndian())
+ StOffset = ByteShift;
+ else
+ StOffset = IVal.getValueType().getStoreSize() - ByteShift - NumBytes;
+
+ SDValue Ptr = St->getBasePtr();
+ if (StOffset) {
+ Ptr = DAG.getNode(ISD::ADD, IVal->getDebugLoc(), Ptr.getValueType(),
+ Ptr, DAG.getConstant(StOffset, Ptr.getValueType()));
+ NewAlign = MinAlign(NewAlign, StOffset);
+ }
+
+ // Truncate down to the new size.
+ IVal = DAG.getNode(ISD::TRUNCATE, IVal->getDebugLoc(), VT, IVal);
+
+ ++OpsNarrowed;
+ return DAG.getStore(St->getChain(), St->getDebugLoc(), IVal, Ptr,
+ St->getSrcValue(), St->getSrcValueOffset()+StOffset,
+ false, false, NewAlign).getNode();
+}
+
/// ReduceLoadOpStoreWidth - Look for sequence of load / op / store where op is
/// one of 'or', 'xor', and 'and' of immediates. If 'op' is only touching some
return SDValue();
unsigned Opc = Value.getOpcode();
+
+ // If this is "store (or X, Y), P" and X is "(and (load P), cst)", where cst
+ // is a byte mask indicating a consecutive number of bytes, check to see if
+ // Y is known to provide just those bytes. If so, we try to replace the
+ // load + replace + store sequence with a single (narrower) store, which makes
+ // the load dead.
+ if (Opc == ISD::OR) {
+ std::pair<unsigned, unsigned> MaskedLoad;
+ MaskedLoad = CheckForMaskedLoad(Value.getOperand(0), Ptr, Chain);
+ if (MaskedLoad.first)
+ if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
+ Value.getOperand(1), ST,this))
+ return SDValue(NewST, 0);
+
+ // Or is commutative, so try swapping X and Y.
+ MaskedLoad = CheckForMaskedLoad(Value.getOperand(1), Ptr, Chain);
+ if (MaskedLoad.first)
+ if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
+ Value.getOperand(0), ST,this))
+ return SDValue(NewST, 0);
+ }
+
if ((Opc != ISD::OR && Opc != ISD::XOR && Opc != ISD::AND) ||
Value.getOperand(1).getOpcode() != ISD::Constant)
return SDValue();
PtrOff = (BitWidth + 7 - NewBW) / 8 - PtrOff;
unsigned NewAlign = MinAlign(LD->getAlignment(), PtrOff);
- if (NewAlign <
- TLI.getTargetData()->getABITypeAlignment(NewVT.getTypeForEVT(*DAG.getContext())))
+ const Type *NewVTTy = NewVT.getTypeForEVT(*DAG.getContext());
+ if (NewAlign < TLI.getTargetData()->getABITypeAlignment(NewVTTy))
return SDValue();
SDValue NewPtr = DAG.getNode(ISD::ADD, LD->getDebugLoc(),
SDValue NewLD = DAG.getLoad(NewVT, N0.getDebugLoc(),
LD->getChain(), NewPtr,
LD->getSrcValue(), LD->getSrcValueOffset(),
- LD->isVolatile(), NewAlign);
+ LD->isVolatile(), LD->isNonTemporal(),
+ NewAlign);
SDValue NewVal = DAG.getNode(Opc, Value.getDebugLoc(), NewVT, NewLD,
DAG.getConstant(NewImm, NewVT));
SDValue NewST = DAG.getStore(Chain, N->getDebugLoc(),
NewVal, NewPtr,
ST->getSrcValue(), ST->getSrcValueOffset(),
- false, NewAlign);
+ false, false, NewAlign);
AddToWorkList(NewPtr.getNode());
AddToWorkList(NewLD.getNode());
SDValue Value = ST->getValue();
SDValue Ptr = ST->getBasePtr();
- // Try to infer better alignment information than the store already has.
- if (OptLevel != CodeGenOpt::None && ST->isUnindexed()) {
- if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
- if (Align > ST->getAlignment())
- return DAG.getTruncStore(Chain, N->getDebugLoc(), Value,
- Ptr, ST->getSrcValue(),
- ST->getSrcValueOffset(), ST->getMemoryVT(),
- 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() &&
TLI.isOperationLegalOrCustom(ISD::STORE, SVT)))
return DAG.getStore(Chain, N->getDebugLoc(), Value.getOperand(0),
Ptr, ST->getSrcValue(),
- ST->getSrcValueOffset(), ST->isVolatile(), OrigAlign);
+ ST->getSrcValueOffset(), ST->isVolatile(),
+ ST->isNonTemporal(), OrigAlign);
}
// Turn 'store float 1.0, Ptr' -> 'store int 0x12345678, Ptr'
case MVT::ppcf128:
break;
case MVT::f32:
- if (((TLI.isTypeLegal(MVT::i32) || !LegalTypes) && !LegalOperations &&
- !ST->isVolatile()) ||
+ if ((isTypeLegal(MVT::i32) && !LegalOperations && !ST->isVolatile()) ||
TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
Tmp = DAG.getConstant((uint32_t)CFP->getValueAPF().
bitcastToAPInt().getZExtValue(), MVT::i32);
return DAG.getStore(Chain, N->getDebugLoc(), Tmp,
Ptr, ST->getSrcValue(),
ST->getSrcValueOffset(), ST->isVolatile(),
- ST->getAlignment());
+ ST->isNonTemporal(), ST->getAlignment());
}
break;
case MVT::f64:
- if (((TLI.isTypeLegal(MVT::i64) || !LegalTypes) && !LegalOperations &&
+ if ((TLI.isTypeLegal(MVT::i64) && !LegalOperations &&
!ST->isVolatile()) ||
TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i64)) {
Tmp = DAG.getConstant(CFP->getValueAPF().bitcastToAPInt().
return DAG.getStore(Chain, N->getDebugLoc(), Tmp,
Ptr, ST->getSrcValue(),
ST->getSrcValueOffset(), ST->isVolatile(),
- ST->getAlignment());
+ ST->isNonTemporal(), ST->getAlignment());
} else if (!ST->isVolatile() &&
TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
// Many FP stores are not made apparent until after legalize, e.g. for
int SVOffset = ST->getSrcValueOffset();
unsigned Alignment = ST->getAlignment();
bool isVolatile = ST->isVolatile();
+ bool isNonTemporal = ST->isNonTemporal();
SDValue St0 = DAG.getStore(Chain, ST->getDebugLoc(), Lo,
Ptr, ST->getSrcValue(),
ST->getSrcValueOffset(),
- isVolatile, ST->getAlignment());
+ isVolatile, isNonTemporal,
+ ST->getAlignment());
Ptr = DAG.getNode(ISD::ADD, N->getDebugLoc(), Ptr.getValueType(), Ptr,
DAG.getConstant(4, Ptr.getValueType()));
SVOffset += 4;
Alignment = MinAlign(Alignment, 4U);
SDValue St1 = DAG.getStore(Chain, ST->getDebugLoc(), Hi,
Ptr, ST->getSrcValue(),
- SVOffset, isVolatile, Alignment);
+ SVOffset, isVolatile, isNonTemporal,
+ Alignment);
return DAG.getNode(ISD::TokenFactor, N->getDebugLoc(), MVT::Other,
St0, St1);
}
}
}
+ // Try to infer better alignment information than the store already has.
+ if (OptLevel != CodeGenOpt::None && ST->isUnindexed()) {
+ if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
+ if (Align > ST->getAlignment())
+ return DAG.getTruncStore(Chain, N->getDebugLoc(), Value,
+ Ptr, ST->getSrcValue(),
+ ST->getSrcValueOffset(), ST->getMemoryVT(),
+ ST->isVolatile(), ST->isNonTemporal(), Align);
+ }
+ }
+
if (CombinerAA) {
// Walk up chain skipping non-aliasing memory nodes.
SDValue BetterChain = FindBetterChain(N, Chain);
if (ST->isTruncatingStore()) {
ReplStore = DAG.getTruncStore(BetterChain, N->getDebugLoc(), Value, Ptr,
ST->getSrcValue(),ST->getSrcValueOffset(),
- ST->getMemoryVT(),
- ST->isVolatile(), ST->getAlignment());
+ ST->getMemoryVT(), ST->isVolatile(),
+ ST->isNonTemporal(), ST->getAlignment());
} else {
ReplStore = DAG.getStore(BetterChain, N->getDebugLoc(), Value, Ptr,
ST->getSrcValue(), ST->getSrcValueOffset(),
- ST->isVolatile(), ST->getAlignment());
+ ST->isVolatile(), ST->isNonTemporal(),
+ ST->getAlignment());
}
// Create token to keep both nodes around.
return DAG.getTruncStore(Chain, N->getDebugLoc(), Shorter,
Ptr, ST->getSrcValue(),
ST->getSrcValueOffset(), ST->getMemoryVT(),
- ST->isVolatile(), ST->getAlignment());
+ ST->isVolatile(), ST->isNonTemporal(),
+ ST->getAlignment());
// Otherwise, see if we can simplify the operation with
// SimplifyDemandedBits, which only works if the value has a single use.
if (SimplifyDemandedBits(Value,
APInt::getLowBitsSet(
Value.getValueType().getScalarType().getSizeInBits(),
- ST->getMemoryVT().getSizeInBits())))
+ ST->getMemoryVT().getScalarType().getSizeInBits())))
return SDValue(N, 0);
}
return DAG.getTruncStore(Chain, N->getDebugLoc(), Value.getOperand(0),
Ptr, ST->getSrcValue(),
ST->getSrcValueOffset(), ST->getMemoryVT(),
- ST->isVolatile(), ST->getAlignment());
+ ST->isVolatile(), ST->isNonTemporal(),
+ ST->getAlignment());
}
return ReduceLoadOpStoreWidth(N);
SDValue InVec = N->getOperand(0);
if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
- // If the operand is wider than the vector element type then it is implicitly
- // truncated. Make that explicit here.
+ // Check if the result type doesn't match the inserted element type. A
+ // SCALAR_TO_VECTOR may truncate the inserted element and the
+ // EXTRACT_VECTOR_ELT may widen the extracted vector.
EVT EltVT = InVec.getValueType().getVectorElementType();
SDValue InOp = InVec.getOperand(0);
- if (InOp.getValueType() != EltVT)
- return DAG.getNode(ISD::TRUNCATE, InVec.getDebugLoc(), EltVT, InOp);
+ EVT NVT = N->getValueType(0);
+ if (InOp.getValueType() != NVT) {
+ assert(InOp.getValueType().isInteger() && NVT.isInteger());
+ return DAG.getSExtOrTrunc(InOp, InVec.getDebugLoc(), NVT);
+ }
return InOp;
}
InVec = InVec.getOperand(0);
if (ISD::isNormalLoad(InVec.getNode())) {
LN0 = cast<LoadSDNode>(InVec);
- Elt = (Idx < (int)NumElems) ? Idx : Idx - NumElems;
+ Elt = (Idx < (int)NumElems) ? Idx : Idx - (int)NumElems;
}
}
return DAG.getLoad(LVT, N->getDebugLoc(), LN0->getChain(), NewPtr,
LN0->getSrcValue(), LN0->getSrcValueOffset(),
- LN0->isVolatile(), Align);
+ LN0->isVolatile(), LN0->isNonTemporal(), Align);
}
return SDValue();
}
// Add count and size info.
- if (!TLI.isTypeLegal(VT) && LegalTypes)
+ if (!isTypeLegal(VT))
return SDValue();
// Return the new VECTOR_SHUFFLE node.
LLD->getChain(),
Addr, 0, 0,
LLD->isVolatile(),
+ LLD->isNonTemporal(),
LLD->getAlignment());
} else {
Load = DAG.getExtLoad(LLD->getExtensionType(),
LLD->getChain(), Addr, 0, 0,
LLD->getMemoryVT(),
LLD->isVolatile(),
+ LLD->isNonTemporal(),
LLD->getAlignment());
}
CstOffset);
return DAG.getLoad(TV->getValueType(0), DL, DAG.getEntryNode(), CPIdx,
PseudoSourceValue::getConstantPool(), 0, false,
- Alignment);
+ false, Alignment);
}
}
/// FindBaseOffset - Return true if base is a frame index, which is known not
// to alias with anything but itself. Provides base object and offset as results.
static bool FindBaseOffset(SDValue Ptr, SDValue &Base, int64_t &Offset,
- GlobalValue *&GV, void *&CV) {
+ const GlobalValue *&GV, void *&CV) {
// Assume it is a primitive operation.
Base = Ptr; Offset = 0; GV = 0; CV = 0;
// Gather base node and offset information.
SDValue Base1, Base2;
int64_t Offset1, Offset2;
- GlobalValue *GV1, *GV2;
+ const GlobalValue *GV1, *GV2;
void *CV1, *CV2;
bool isFrameIndex1 = FindBaseOffset(Ptr1, Base1, Offset1, GV1, CV1);
bool isFrameIndex2 = FindBaseOffset(Ptr2, Base2, Offset2, GV2, CV2);