#include "llvm/Support/MathExtras.h"
using namespace llvm;
+namespace llvm {
+TLSModel::Model getTLSModel(const GlobalValue *GV, Reloc::Model reloc) {
+ bool isLocal = GV->hasLocalLinkage();
+ bool isDeclaration = GV->isDeclaration();
+ // FIXME: what should we do for protected and internal visibility?
+ // For variables, is internal different from hidden?
+ bool isHidden = GV->hasHiddenVisibility();
+
+ if (reloc == Reloc::PIC_) {
+ if (isLocal || isHidden)
+ return TLSModel::LocalDynamic;
+ else
+ return TLSModel::GeneralDynamic;
+ } else {
+ if (!isDeclaration || isHidden)
+ return TLSModel::LocalExec;
+ else
+ return TLSModel::InitialExec;
+ }
+}
+}
+
/// InitLibcallNames - Set default libcall names.
///
static void InitLibcallNames(const char **Names) {
+ Names[RTLIB::SHL_I16] = "__ashli16";
Names[RTLIB::SHL_I32] = "__ashlsi3";
Names[RTLIB::SHL_I64] = "__ashldi3";
Names[RTLIB::SHL_I128] = "__ashlti3";
+ Names[RTLIB::SRL_I16] = "__lshri16";
Names[RTLIB::SRL_I32] = "__lshrsi3";
Names[RTLIB::SRL_I64] = "__lshrdi3";
Names[RTLIB::SRL_I128] = "__lshrti3";
+ Names[RTLIB::SRA_I16] = "__ashri16";
Names[RTLIB::SRA_I32] = "__ashrsi3";
Names[RTLIB::SRA_I64] = "__ashrdi3";
Names[RTLIB::SRA_I128] = "__ashrti3";
+ Names[RTLIB::MUL_I16] = "__muli16";
Names[RTLIB::MUL_I32] = "__mulsi3";
Names[RTLIB::MUL_I64] = "__muldi3";
Names[RTLIB::MUL_I128] = "__multi3";
TargetLowering::TargetLowering(TargetMachine &tm)
: TM(tm), TD(TM.getTargetData()) {
- assert(ISD::BUILTIN_OP_END <= OpActionsCapacity &&
- "Fixed size array in TargetLowering is not large enough!");
// All operations default to being supported.
memset(OpActions, 0, sizeof(OpActions));
memset(LoadExtActions, 0, sizeof(LoadExtActions));
StackPointerRegisterToSaveRestore = 0;
ExceptionPointerRegister = 0;
ExceptionSelectorRegister = 0;
- SetCCResultContents = UndefinedSetCCResult;
+ BooleanContents = UndefinedBooleanContent;
SchedPreferenceInfo = SchedulingForLatency;
JumpBufSize = 0;
JumpBufAlignment = 0;
IntermediateVT, NumIntermediates,
RegisterVT);
RegisterTypeForVT[i] = RegisterVT;
- TransformToType[i] = MVT::Other; // this isn't actually used
- ValueTypeActions.setTypeAction(VT, Expand);
+
+ // Determine if there is a legal wider type.
+ bool IsLegalWiderType = false;
+ MVT EltVT = VT.getVectorElementType();
+ unsigned NElts = VT.getVectorNumElements();
+ for (unsigned nVT = i+1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
+ MVT SVT = (MVT::SimpleValueType)nVT;
+ if (isTypeLegal(SVT) && SVT.getVectorElementType() == EltVT &&
+ SVT.getVectorNumElements() > NElts) {
+ TransformToType[i] = SVT;
+ ValueTypeActions.setTypeAction(VT, Promote);
+ IsLegalWiderType = true;
+ break;
+ }
+ }
+ if (!IsLegalWiderType) {
+ MVT NVT = VT.getPow2VectorType();
+ if (NVT == VT) {
+ // Type is already a power of 2. The default action is to split.
+ TransformToType[i] = MVT::Other;
+ ValueTypeActions.setTypeAction(VT, Expand);
+ } else {
+ TransformToType[i] = NVT;
+ ValueTypeActions.setTypeAction(VT, Promote);
+ }
+ }
}
}
}
}
-MVT TargetLowering::getSetCCResultType(const SDValue &) const {
+MVT TargetLowering::getSetCCResultType(MVT VT) const {
return getValueType(TD->getIntPtrType());
}
return 1;
}
+/// getWidenVectorType: given a vector type, returns the type to widen to
+/// (e.g., v7i8 to v8i8). If the vector type is legal, it returns itself.
+/// If there is no vector type that we want to widen to, returns MVT::Other
+/// When and where to widen is target dependent based on the cost of
+/// scalarizing vs using the wider vector type.
+MVT TargetLowering::getWidenVectorType(MVT VT) const {
+ assert(VT.isVector());
+ if (isTypeLegal(VT))
+ return VT;
+
+ // Default is not to widen until moved to LegalizeTypes
+ return MVT::Other;
+}
+
/// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
/// function arguments in the caller parameter area. This is the actual
/// alignment, not its logarithm.
SDValue TargetLowering::getPICJumpTableRelocBase(SDValue Table,
SelectionDAG &DAG) const {
if (usesGlobalOffsetTable())
- return DAG.getNode(ISD::GLOBAL_OFFSET_TABLE, getPointerTy());
+ return DAG.getGLOBAL_OFFSET_TABLE(getPointerTy());
return Table;
}
/// constant and return true.
bool TargetLowering::TargetLoweringOpt::ShrinkDemandedConstant(SDValue Op,
const APInt &Demanded) {
+ DebugLoc dl = Op.getDebugLoc();
+
// FIXME: ISD::SELECT, ISD::SELECT_CC
- switch(Op.getOpcode()) {
+ switch (Op.getOpcode()) {
default: break;
- case ISD::AND:
- case ISD::OR:
case ISD::XOR:
- if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1)))
- if (C->getAPIntValue().intersects(~Demanded)) {
- MVT VT = Op.getValueType();
- SDValue New = DAG.getNode(Op.getOpcode(), VT, Op.getOperand(0),
- DAG.getConstant(Demanded &
- C->getAPIntValue(),
- VT));
- return CombineTo(Op, New);
- }
+ case ISD::AND:
+ case ISD::OR: {
+ ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
+ if (!C) return false;
+
+ if (Op.getOpcode() == ISD::XOR &&
+ (C->getAPIntValue() | (~Demanded)).isAllOnesValue())
+ return false;
+
+ // if we can expand it to have all bits set, do it
+ if (C->getAPIntValue().intersects(~Demanded)) {
+ MVT VT = Op.getValueType();
+ SDValue New = DAG.getNode(Op.getOpcode(), dl, VT, Op.getOperand(0),
+ DAG.getConstant(Demanded &
+ C->getAPIntValue(),
+ VT));
+ return CombineTo(Op, New);
+ }
+
break;
}
+ }
+
return false;
}
assert(Op.getValueSizeInBits() == BitWidth &&
"Mask size mismatches value type size!");
APInt NewMask = DemandedMask;
+ DebugLoc dl = Op.getDebugLoc();
// Don't know anything.
KnownZero = KnownOne = APInt(BitWidth, 0);
} else if (DemandedMask == 0) {
// Not demanding any bits from Op.
if (Op.getOpcode() != ISD::UNDEF)
- return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::UNDEF, Op.getValueType()));
+ return TLO.CombineTo(Op, TLO.DAG.getUNDEF(Op.getValueType()));
return false;
} else if (Depth == 6) { // Limit search depth.
return false;
// (but not both) turn this into an *inclusive* or.
// e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
if ((NewMask & ~KnownZero & ~KnownZero2) == 0)
- return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::OR, Op.getValueType(),
+ return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::OR, dl, Op.getValueType(),
Op.getOperand(0),
Op.getOperand(1)));
if ((KnownOne & KnownOne2) == KnownOne) {
MVT VT = Op.getValueType();
SDValue ANDC = TLO.DAG.getConstant(~KnownOne & NewMask, VT);
- return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::AND, VT, Op.getOperand(0),
- ANDC));
+ return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::AND, dl, VT,
+ Op.getOperand(0), ANDC));
}
}
if (Expanded.isAllOnesValue()) {
if (Expanded != C->getAPIntValue()) {
MVT VT = Op.getValueType();
- SDValue New = TLO.DAG.getNode(Op.getOpcode(), VT, Op.getOperand(0),
+ SDValue New = TLO.DAG.getNode(Op.getOpcode(), dl,VT, Op.getOperand(0),
TLO.DAG.getConstant(Expanded, VT));
return TLO.CombineTo(Op, New);
}
SDValue NewSA =
TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType());
MVT VT = Op.getValueType();
- return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, VT,
+ return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT,
InOp.getOperand(0), NewSA));
}
}
SDValue NewSA =
TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType());
- return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, VT,
+ return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT,
InOp.getOperand(0), NewSA));
}
}
}
break;
case ISD::SRA:
+ // If this is an arithmetic shift right and only the low-bit is set, we can
+ // always convert this into a logical shr, even if the shift amount is
+ // variable. The low bit of the shift cannot be an input sign bit unless
+ // the shift amount is >= the size of the datatype, which is undefined.
+ if (DemandedMask == 1)
+ return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl, Op.getValueType(),
+ Op.getOperand(0), Op.getOperand(1)));
+
if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
MVT VT = Op.getValueType();
unsigned ShAmt = SA->getZExtValue();
// 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 (KnownZero.intersects(SignBit) || (HighBits & ~NewMask) == HighBits) {
- return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, VT, Op.getOperand(0),
+ return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl, VT,
+ Op.getOperand(0),
Op.getOperand(1)));
} else if (KnownOne.intersects(SignBit)) { // New bits are known one.
KnownOne |= HighBits;
// If the input sign bit is known zero, convert this into a zero extension.
if (KnownZero.intersects(InSignBit))
return TLO.CombineTo(Op,
- TLO.DAG.getZeroExtendInReg(Op.getOperand(0), EVT));
+ TLO.DAG.getZeroExtendInReg(Op.getOperand(0),dl,EVT));
if (KnownOne.intersects(InSignBit)) { // Input sign bit known set
KnownOne |= NewBits;
APInt NewBits =
APInt::getHighBitsSet(BitWidth, BitWidth - OperandBitWidth) & NewMask;
if (!NewBits.intersects(NewMask))
- return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ANY_EXTEND,
+ return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ANY_EXTEND, dl,
Op.getValueType(),
Op.getOperand(0)));
// If none of the top bits are demanded, convert this into an any_extend.
if (NewBits == 0)
- return TLO.CombineTo(Op,TLO.DAG.getNode(ISD::ANY_EXTEND,Op.getValueType(),
- Op.getOperand(0)));
+ return TLO.CombineTo(Op,TLO.DAG.getNode(ISD::ANY_EXTEND, dl,
+ Op.getValueType(),
+ Op.getOperand(0)));
// Since some of the sign extended bits are demanded, we know that the sign
// bit is demanded.
// If the sign bit is known zero, convert this to a zero extend.
if (KnownZero.intersects(InSignBit))
- return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ZERO_EXTEND,
+ return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ZERO_EXTEND, dl,
Op.getValueType(),
Op.getOperand(0)));
if (ShAmt->getZExtValue() < BitWidth && !(HighBits & NewMask)) {
// None of the shifted in bits are needed. Add a truncate of the
// shift input, then shift it.
- SDValue NewTrunc = TLO.DAG.getNode(ISD::TRUNCATE,
+ SDValue NewTrunc = TLO.DAG.getNode(ISD::TRUNCATE, dl,
Op.getValueType(),
In.getOperand(0));
- return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL,Op.getValueType(),
- NewTrunc, In.getOperand(1)));
+ return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl,
+ Op.getValueType(),
+ NewTrunc,
+ In.getOperand(1)));
}
}
break;
return 1;
}
+/// ValueHasExactlyOneBitSet - Test if the given value is known to have exactly
+/// one bit set. This differs from ComputeMaskedBits in that it doesn't need to
+/// determine which bit is set.
+///
+static bool ValueHasExactlyOneBitSet(SDValue Val, const SelectionDAG &DAG) {
+ // A left-shift of a constant one will have exactly one bit set, because
+ // shifting the bit off the end is undefined.
+ if (Val.getOpcode() == ISD::SHL)
+ if (ConstantSDNode *C =
+ dyn_cast<ConstantSDNode>(Val.getNode()->getOperand(0)))
+ if (C->getAPIntValue() == 1)
+ return true;
+
+ // Similarly, a right-shift of a constant sign-bit will have exactly
+ // one bit set.
+ if (Val.getOpcode() == ISD::SRL)
+ if (ConstantSDNode *C =
+ dyn_cast<ConstantSDNode>(Val.getNode()->getOperand(0)))
+ if (C->getAPIntValue().isSignBit())
+ return true;
+
+ // More could be done here, though the above checks are enough
+ // to handle some common cases.
+
+ // Fall back to ComputeMaskedBits to catch other known cases.
+ MVT OpVT = Val.getValueType();
+ unsigned BitWidth = OpVT.getSizeInBits();
+ APInt Mask = APInt::getAllOnesValue(BitWidth);
+ APInt KnownZero, KnownOne;
+ DAG.ComputeMaskedBits(Val, Mask, KnownZero, KnownOne);
+ return (KnownZero.countPopulation() == BitWidth - 1) &&
+ (KnownOne.countPopulation() == 1);
+}
/// SimplifySetCC - Try to simplify a setcc built with the specified operands
/// and cc. If it is unable to simplify it, return a null SDValue.
SDValue
TargetLowering::SimplifySetCC(MVT VT, SDValue N0, SDValue N1,
ISD::CondCode Cond, bool foldBooleans,
- DAGCombinerInfo &DCI) const {
+ DAGCombinerInfo &DCI, DebugLoc dl) const {
SelectionDAG &DAG = DCI.DAG;
// These setcc operations always fold.
if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode())) {
const APInt &C1 = N1C->getAPIntValue();
if (isa<ConstantSDNode>(N0.getNode())) {
- return DAG.FoldSetCC(VT, N0, N1, Cond);
+ return DAG.FoldSetCC(VT, N0, N1, Cond, dl);
} else {
// If the LHS is '(srl (ctlz x), 5)', the RHS is 0/1, and this is an
// equality comparison, then we're just comparing whether X itself is
Cond = ISD::SETEQ;
}
SDValue Zero = DAG.getConstant(0, N0.getValueType());
- return DAG.getSetCC(VT, N0.getOperand(0).getOperand(0),
+ return DAG.getSetCC(dl, VT, N0.getOperand(0).getOperand(0),
Zero, Cond);
}
}
-
+
+ // If the LHS is '(and load, const)', the RHS is 0,
+ // the test is for equality or unsigned, and all 1 bits of the const are
+ // in the same partial word, see if we can shorten the load.
+ if (DCI.isBeforeLegalize() &&
+ N0.getOpcode() == ISD::AND && C1 == 0 &&
+ isa<LoadSDNode>(N0.getOperand(0)) &&
+ N0.getOperand(0).getNode()->hasOneUse() &&
+ isa<ConstantSDNode>(N0.getOperand(1))) {
+ LoadSDNode *Lod = cast<LoadSDNode>(N0.getOperand(0));
+ uint64_t Mask = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
+ uint64_t bestMask = 0;
+ unsigned bestWidth = 0, bestOffset = 0;
+ if (!Lod->isVolatile() && Lod->isUnindexed()) {
+ unsigned origWidth = N0.getValueType().getSizeInBits();
+ // We can narrow (e.g.) 16-bit extending loads on 32-bit target to
+ // 8 bits, but have to be careful...
+ if (Lod->getExtensionType() != ISD::NON_EXTLOAD)
+ origWidth = Lod->getMemoryVT().getSizeInBits();
+ for (unsigned width = origWidth / 2; width>=8; width /= 2) {
+ uint64_t newMask = (1ULL << width) - 1;
+ for (unsigned offset=0; offset<origWidth/width; offset++) {
+ if ((newMask & Mask)==Mask) {
+ if (!TD->isLittleEndian())
+ bestOffset = (origWidth/width - offset - 1) * (width/8);
+ else
+ bestOffset = (uint64_t)offset * (width/8);
+ bestMask = Mask >> (offset * (width/8) * 8);
+ bestWidth = width;
+ break;
+ }
+ newMask = newMask << width;
+ }
+ }
+ }
+ if (bestWidth) {
+ MVT newVT = MVT::getIntegerVT(bestWidth);
+ if (newVT.isRound()) {
+ MVT PtrType = Lod->getOperand(1).getValueType();
+ SDValue Ptr = Lod->getBasePtr();
+ if (bestOffset != 0)
+ Ptr = DAG.getNode(ISD::ADD, dl, PtrType, Lod->getBasePtr(),
+ DAG.getConstant(bestOffset, PtrType));
+ unsigned NewAlign = MinAlign(Lod->getAlignment(), bestOffset);
+ SDValue NewLoad = DAG.getLoad(newVT, dl, Lod->getChain(), Ptr,
+ Lod->getSrcValue(),
+ Lod->getSrcValueOffset() + bestOffset,
+ false, NewAlign);
+ return DAG.getSetCC(dl, VT,
+ DAG.getNode(ISD::AND, dl, newVT, NewLoad,
+ DAG.getConstant(bestMask, newVT)),
+ DAG.getConstant(0LL, newVT), Cond);
+ }
+ }
+ }
+
// If the LHS is a ZERO_EXTEND, perform the comparison on the input.
if (N0.getOpcode() == ISD::ZERO_EXTEND) {
unsigned InSize = N0.getOperand(0).getValueType().getSizeInBits();
case ISD::SETUGE:
case ISD::SETULT:
case ISD::SETULE:
- return DAG.getSetCC(VT, N0.getOperand(0),
+ return DAG.getSetCC(dl, VT, N0.getOperand(0),
DAG.getConstant(APInt(C1).trunc(InSize),
N0.getOperand(0).getValueType()),
Cond);
ZextOp = N0.getOperand(0);
} else {
APInt Imm = APInt::getLowBitsSet(ExtDstTyBits, ExtSrcTyBits);
- ZextOp = DAG.getNode(ISD::AND, Op0Ty, N0.getOperand(0),
+ ZextOp = DAG.getNode(ISD::AND, dl, Op0Ty, N0.getOperand(0),
DAG.getConstant(Imm, Op0Ty));
}
if (!DCI.isCalledByLegalizer())
DCI.AddToWorklist(ZextOp.getNode());
// Otherwise, make this a use of a zext.
- return DAG.getSetCC(VT, ZextOp,
+ return DAG.getSetCC(dl, VT, ZextOp,
DAG.getConstant(C1 & APInt::getLowBitsSet(
ExtDstTyBits,
ExtSrcTyBits),
ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
CC = ISD::getSetCCInverse(CC,
N0.getOperand(0).getValueType().isInteger());
- return DAG.getSetCC(VT, N0.getOperand(0), N0.getOperand(1), CC);
+ return DAG.getSetCC(dl, VT, N0.getOperand(0), N0.getOperand(1), CC);
}
if ((N0.getOpcode() == ISD::XOR ||
assert(N0.getOpcode() == ISD::AND &&
N0.getOperand(0).getOpcode() == ISD::XOR);
// ((X^1)&1)^1 -> X & 1
- Val = DAG.getNode(ISD::AND, N0.getValueType(),
+ Val = DAG.getNode(ISD::AND, dl, N0.getValueType(),
N0.getOperand(0).getOperand(0),
N0.getOperand(1));
}
- return DAG.getSetCC(VT, Val, N1,
+ return DAG.getSetCC(dl, VT, Val, N1,
Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ);
}
}
if (Cond == ISD::SETGE || Cond == ISD::SETUGE) {
if (C1 == MinVal) return DAG.getConstant(1, VT); // X >= MIN --> true
// X >= C0 --> X > (C0-1)
- return DAG.getSetCC(VT, N0, DAG.getConstant(C1-1, N1.getValueType()),
- (Cond == ISD::SETGE) ? ISD::SETGT : ISD::SETUGT);
+ return DAG.getSetCC(dl, VT, N0,
+ DAG.getConstant(C1-1, N1.getValueType()),
+ (Cond == ISD::SETGE) ? ISD::SETGT : ISD::SETUGT);
}
if (Cond == ISD::SETLE || Cond == ISD::SETULE) {
if (C1 == MaxVal) return DAG.getConstant(1, VT); // X <= MAX --> true
// X <= C0 --> X < (C0+1)
- return DAG.getSetCC(VT, N0, DAG.getConstant(C1+1, N1.getValueType()),
- (Cond == ISD::SETLE) ? ISD::SETLT : ISD::SETULT);
+ return DAG.getSetCC(dl, VT, N0,
+ DAG.getConstant(C1+1, N1.getValueType()),
+ (Cond == ISD::SETLE) ? ISD::SETLT : ISD::SETULT);
}
if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal)
// Canonicalize setgt X, Min --> setne X, Min
if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MinVal)
- return DAG.getSetCC(VT, N0, N1, ISD::SETNE);
+ return DAG.getSetCC(dl, VT, N0, N1, ISD::SETNE);
// Canonicalize setlt X, Max --> setne X, Max
if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MaxVal)
- return DAG.getSetCC(VT, N0, N1, ISD::SETNE);
+ return DAG.getSetCC(dl, VT, N0, N1, ISD::SETNE);
// If we have setult X, 1, turn it into seteq X, 0
if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal+1)
- return DAG.getSetCC(VT, N0, DAG.getConstant(MinVal, N0.getValueType()),
- ISD::SETEQ);
+ return DAG.getSetCC(dl, VT, N0,
+ DAG.getConstant(MinVal, N0.getValueType()),
+ ISD::SETEQ);
// If we have setugt X, Max-1, turn it into seteq X, Max
else if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal-1)
- return DAG.getSetCC(VT, N0, DAG.getConstant(MaxVal, N0.getValueType()),
- ISD::SETEQ);
+ return DAG.getSetCC(dl, VT, N0,
+ DAG.getConstant(MaxVal, N0.getValueType()),
+ ISD::SETEQ);
// If we have "setcc X, C0", check to see if we can shrink the immediate
// by changing cc.
// SETUGT X, SINTMAX -> SETLT X, 0
- if (Cond == ISD::SETUGT && OperandBitSize != 1 &&
- C1 == (~0ULL >> (65-OperandBitSize)))
- return DAG.getSetCC(VT, N0, DAG.getConstant(0, N1.getValueType()),
+ if (Cond == ISD::SETUGT &&
+ C1 == APInt::getSignedMaxValue(OperandBitSize))
+ return DAG.getSetCC(dl, VT, N0,
+ DAG.getConstant(0, N1.getValueType()),
ISD::SETLT);
- // FIXME: Implement the rest of these.
+ // SETULT X, SINTMIN -> SETGT X, -1
+ if (Cond == ISD::SETULT &&
+ C1 == APInt::getSignedMinValue(OperandBitSize)) {
+ SDValue ConstMinusOne =
+ DAG.getConstant(APInt::getAllOnesValue(OperandBitSize),
+ N1.getValueType());
+ return DAG.getSetCC(dl, VT, N0, ConstMinusOne, ISD::SETGT);
+ }
// Fold bit comparisons when we can.
if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
VT == N0.getValueType() && N0.getOpcode() == ISD::AND)
if (ConstantSDNode *AndRHS =
dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
+ MVT ShiftTy = DCI.isBeforeLegalize() ?
+ getPointerTy() : getShiftAmountTy();
if (Cond == ISD::SETNE && C1 == 0) {// (X & 8) != 0 --> (X & 8) >> 3
// Perform the xform if the AND RHS is a single bit.
if (isPowerOf2_64(AndRHS->getZExtValue())) {
- return DAG.getNode(ISD::SRL, VT, N0,
- DAG.getConstant(Log2_64(AndRHS->getZExtValue()),
- getShiftAmountTy()));
+ return DAG.getNode(ISD::SRL, dl, VT, N0,
+ DAG.getConstant(Log2_64(AndRHS->getZExtValue()),
+ ShiftTy));
}
} else if (Cond == ISD::SETEQ && C1 == AndRHS->getZExtValue()) {
// (X & 8) == 8 --> (X & 8) >> 3
// Perform the xform if C1 is a single bit.
if (C1.isPowerOf2()) {
- return DAG.getNode(ISD::SRL, VT, N0,
- DAG.getConstant(C1.logBase2(), getShiftAmountTy()));
+ return DAG.getNode(ISD::SRL, dl, VT, N0,
+ DAG.getConstant(C1.logBase2(), ShiftTy));
}
}
}
}
} else if (isa<ConstantSDNode>(N0.getNode())) {
// Ensure that the constant occurs on the RHS.
- return DAG.getSetCC(VT, N1, N0, ISD::getSetCCSwappedOperands(Cond));
+ return DAG.getSetCC(dl, VT, N1, N0, ISD::getSetCCSwappedOperands(Cond));
}
if (isa<ConstantFPSDNode>(N0.getNode())) {
// Constant fold or commute setcc.
- SDValue O = DAG.FoldSetCC(VT, N0, N1, Cond);
+ SDValue O = DAG.FoldSetCC(VT, N0, N1, Cond, dl);
if (O.getNode()) return O;
} else if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1.getNode())) {
// If the RHS of an FP comparison is a constant, simplify it away in
case 1: // Known true.
return DAG.getConstant(1, VT);
case 2: // Undefined.
- return DAG.getNode(ISD::UNDEF, VT);
+ return DAG.getUNDEF(VT);
}
}
// have SETO(x,x) instead of SETO(x, 0.0) because this avoids having to
// materialize 0.0.
if (Cond == ISD::SETO || Cond == ISD::SETUO)
- return DAG.getSetCC(VT, N0, N0, Cond);
+ return DAG.getSetCC(dl, VT, N0, N0, Cond);
}
if (N0 == N1) {
// if it is not already.
ISD::CondCode NewCond = UOF == 0 ? ISD::SETO : ISD::SETUO;
if (NewCond != Cond)
- return DAG.getSetCC(VT, N0, N1, NewCond);
+ return DAG.getSetCC(dl, VT, N0, N1, NewCond);
}
if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
// Simplify (X+Y) == (X+Z) --> Y == Z
if (N0.getOpcode() == N1.getOpcode()) {
if (N0.getOperand(0) == N1.getOperand(0))
- return DAG.getSetCC(VT, N0.getOperand(1), N1.getOperand(1), Cond);
+ return DAG.getSetCC(dl, VT, N0.getOperand(1), N1.getOperand(1), Cond);
if (N0.getOperand(1) == N1.getOperand(1))
- return DAG.getSetCC(VT, N0.getOperand(0), N1.getOperand(0), Cond);
+ return DAG.getSetCC(dl, VT, N0.getOperand(0), N1.getOperand(0), Cond);
if (DAG.isCommutativeBinOp(N0.getOpcode())) {
// If X op Y == Y op X, try other combinations.
if (N0.getOperand(0) == N1.getOperand(1))
- return DAG.getSetCC(VT, N0.getOperand(1), N1.getOperand(0), Cond);
+ return DAG.getSetCC(dl, VT, N0.getOperand(1), N1.getOperand(0),
+ Cond);
if (N0.getOperand(1) == N1.getOperand(0))
- return DAG.getSetCC(VT, N0.getOperand(0), N1.getOperand(1), Cond);
+ return DAG.getSetCC(dl, VT, N0.getOperand(0), N1.getOperand(1),
+ Cond);
}
}
if (ConstantSDNode *LHSR = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
// Turn (X+C1) == C2 --> X == C2-C1
if (N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse()) {
- return DAG.getSetCC(VT, N0.getOperand(0),
+ return DAG.getSetCC(dl, VT, N0.getOperand(0),
DAG.getConstant(RHSC->getAPIntValue()-
LHSR->getAPIntValue(),
N0.getValueType()), Cond);
// performing the inversion.
if (DAG.MaskedValueIsZero(N0.getOperand(0), ~LHSR->getAPIntValue()))
return
- DAG.getSetCC(VT, N0.getOperand(0),
+ DAG.getSetCC(dl, VT, N0.getOperand(0),
DAG.getConstant(LHSR->getAPIntValue() ^
RHSC->getAPIntValue(),
N0.getValueType()),
if (ConstantSDNode *SUBC = dyn_cast<ConstantSDNode>(N0.getOperand(0))) {
if (N0.getOpcode() == ISD::SUB && N0.getNode()->hasOneUse()) {
return
- DAG.getSetCC(VT, N0.getOperand(1),
+ DAG.getSetCC(dl, VT, N0.getOperand(1),
DAG.getConstant(SUBC->getAPIntValue() -
RHSC->getAPIntValue(),
N0.getValueType()),
// Simplify (X+Z) == X --> Z == 0
if (N0.getOperand(0) == N1)
- return DAG.getSetCC(VT, N0.getOperand(1),
+ return DAG.getSetCC(dl, VT, N0.getOperand(1),
DAG.getConstant(0, N0.getValueType()), Cond);
if (N0.getOperand(1) == N1) {
if (DAG.isCommutativeBinOp(N0.getOpcode()))
- return DAG.getSetCC(VT, N0.getOperand(0),
+ return DAG.getSetCC(dl, VT, N0.getOperand(0),
DAG.getConstant(0, N0.getValueType()), Cond);
else if (N0.getNode()->hasOneUse()) {
assert(N0.getOpcode() == ISD::SUB && "Unexpected operation!");
// (Z-X) == X --> Z == X<<1
- SDValue SH = DAG.getNode(ISD::SHL, N1.getValueType(),
+ SDValue SH = DAG.getNode(ISD::SHL, dl, N1.getValueType(),
N1,
DAG.getConstant(1, getShiftAmountTy()));
if (!DCI.isCalledByLegalizer())
DCI.AddToWorklist(SH.getNode());
- return DAG.getSetCC(VT, N0.getOperand(0), SH, Cond);
+ return DAG.getSetCC(dl, VT, N0.getOperand(0), SH, Cond);
}
}
}
N1.getOpcode() == ISD::XOR) {
// Simplify X == (X+Z) --> Z == 0
if (N1.getOperand(0) == N0) {
- return DAG.getSetCC(VT, N1.getOperand(1),
+ return DAG.getSetCC(dl, VT, N1.getOperand(1),
DAG.getConstant(0, N1.getValueType()), Cond);
} else if (N1.getOperand(1) == N0) {
if (DAG.isCommutativeBinOp(N1.getOpcode())) {
- return DAG.getSetCC(VT, N1.getOperand(0),
+ return DAG.getSetCC(dl, VT, N1.getOperand(0),
DAG.getConstant(0, N1.getValueType()), Cond);
} else if (N1.getNode()->hasOneUse()) {
assert(N1.getOpcode() == ISD::SUB && "Unexpected operation!");
// X == (Z-X) --> X<<1 == Z
- SDValue SH = DAG.getNode(ISD::SHL, N1.getValueType(), N0,
+ SDValue SH = DAG.getNode(ISD::SHL, dl, N1.getValueType(), N0,
DAG.getConstant(1, getShiftAmountTy()));
if (!DCI.isCalledByLegalizer())
DCI.AddToWorklist(SH.getNode());
- return DAG.getSetCC(VT, SH, N1.getOperand(0), Cond);
+ return DAG.getSetCC(dl, VT, SH, N1.getOperand(0), Cond);
}
}
}
+
+ // Simplify x&y == y to x&y != 0 if y has exactly one bit set.
+ // Note that where y is variable and is known to have at most
+ // one bit set (for example, if it is z&1) we cannot do this;
+ // the expressions are not equivalent when y==0.
+ if (N0.getOpcode() == ISD::AND)
+ if (N0.getOperand(0) == N1 || N0.getOperand(1) == N1) {
+ if (ValueHasExactlyOneBitSet(N1, DAG)) {
+ Cond = ISD::getSetCCInverse(Cond, /*isInteger=*/true);
+ SDValue Zero = DAG.getConstant(0, N1.getValueType());
+ return DAG.getSetCC(dl, VT, N0, Zero, Cond);
+ }
+ }
+ if (N1.getOpcode() == ISD::AND)
+ if (N1.getOperand(0) == N0 || N1.getOperand(1) == N0) {
+ if (ValueHasExactlyOneBitSet(N0, DAG)) {
+ Cond = ISD::getSetCCInverse(Cond, /*isInteger=*/true);
+ SDValue Zero = DAG.getConstant(0, N0.getValueType());
+ return DAG.getSetCC(dl, VT, N1, Zero, Cond);
+ }
+ }
}
// Fold away ALL boolean setcc's.
if (N0.getValueType() == MVT::i1 && foldBooleans) {
switch (Cond) {
default: assert(0 && "Unknown integer setcc!");
- case ISD::SETEQ: // X == Y -> (X^Y)^1
- Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, N1);
- N0 = DAG.getNode(ISD::XOR, MVT::i1, Temp, DAG.getConstant(1, MVT::i1));
+ case ISD::SETEQ: // X == Y -> ~(X^Y)
+ Temp = DAG.getNode(ISD::XOR, dl, MVT::i1, N0, N1);
+ N0 = DAG.getNOT(dl, Temp, MVT::i1);
if (!DCI.isCalledByLegalizer())
DCI.AddToWorklist(Temp.getNode());
break;
case ISD::SETNE: // X != Y --> (X^Y)
- N0 = DAG.getNode(ISD::XOR, MVT::i1, N0, N1);
+ N0 = DAG.getNode(ISD::XOR, dl, MVT::i1, N0, N1);
break;
- case ISD::SETGT: // X >s Y --> X == 0 & Y == 1 --> X^1 & Y
- case ISD::SETULT: // X <u Y --> X == 0 & Y == 1 --> X^1 & Y
- Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, DAG.getConstant(1, MVT::i1));
- N0 = DAG.getNode(ISD::AND, MVT::i1, N1, Temp);
+ case ISD::SETGT: // X >s Y --> X == 0 & Y == 1 --> ~X & Y
+ case ISD::SETULT: // X <u Y --> X == 0 & Y == 1 --> ~X & Y
+ Temp = DAG.getNOT(dl, N0, MVT::i1);
+ N0 = DAG.getNode(ISD::AND, dl, MVT::i1, N1, Temp);
if (!DCI.isCalledByLegalizer())
DCI.AddToWorklist(Temp.getNode());
break;
- case ISD::SETLT: // X <s Y --> X == 1 & Y == 0 --> Y^1 & X
- case ISD::SETUGT: // X >u Y --> X == 1 & Y == 0 --> Y^1 & X
- Temp = DAG.getNode(ISD::XOR, MVT::i1, N1, DAG.getConstant(1, MVT::i1));
- N0 = DAG.getNode(ISD::AND, MVT::i1, N0, Temp);
+ case ISD::SETLT: // X <s Y --> X == 1 & Y == 0 --> ~Y & X
+ case ISD::SETUGT: // X >u Y --> X == 1 & Y == 0 --> ~Y & X
+ Temp = DAG.getNOT(dl, N1, MVT::i1);
+ N0 = DAG.getNode(ISD::AND, dl, MVT::i1, N0, Temp);
if (!DCI.isCalledByLegalizer())
DCI.AddToWorklist(Temp.getNode());
break;
- case ISD::SETULE: // X <=u Y --> X == 0 | Y == 1 --> X^1 | Y
- case ISD::SETGE: // X >=s Y --> X == 0 | Y == 1 --> X^1 | Y
- Temp = DAG.getNode(ISD::XOR, MVT::i1, N0, DAG.getConstant(1, MVT::i1));
- N0 = DAG.getNode(ISD::OR, MVT::i1, N1, Temp);
+ case ISD::SETULE: // X <=u Y --> X == 0 | Y == 1 --> ~X | Y
+ case ISD::SETGE: // X >=s Y --> X == 0 | Y == 1 --> ~X | Y
+ Temp = DAG.getNOT(dl, N0, MVT::i1);
+ N0 = DAG.getNode(ISD::OR, dl, MVT::i1, N1, Temp);
if (!DCI.isCalledByLegalizer())
DCI.AddToWorklist(Temp.getNode());
break;
- case ISD::SETUGE: // X >=u Y --> X == 1 | Y == 0 --> Y^1 | X
- case ISD::SETLE: // X <=s Y --> X == 1 | Y == 0 --> Y^1 | X
- Temp = DAG.getNode(ISD::XOR, MVT::i1, N1, DAG.getConstant(1, MVT::i1));
- N0 = DAG.getNode(ISD::OR, MVT::i1, N0, Temp);
+ case ISD::SETUGE: // X >=u Y --> X == 1 | Y == 0 --> ~Y | X
+ case ISD::SETLE: // X <=s Y --> X == 1 | Y == 0 --> ~Y | X
+ Temp = DAG.getNOT(dl, N1, MVT::i1);
+ N0 = DAG.getNode(ISD::OR, dl, MVT::i1, N0, Temp);
break;
}
if (VT != MVT::i1) {
if (!DCI.isCalledByLegalizer())
DCI.AddToWorklist(N0.getNode());
// FIXME: If running after legalize, we probably can't do this.
- N0 = DAG.getNode(ISD::ZERO_EXTEND, VT, N0);
+ N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, N0);
}
return N0;
}
if (C) { // just C, no GV.
// Simple constants are not allowed for 's'.
if (ConstraintLetter != 's') {
- Ops.push_back(DAG.getTargetConstant(C->getAPIntValue(),
- Op.getValueType()));
+ // gcc prints these as sign extended. Sign extend value to 64 bits
+ // now; without this it would get ZExt'd later in
+ // ScheduleDAGSDNodes::EmitNode, which is very generic.
+ Ops.push_back(DAG.getTargetConstant(C->getAPIntValue().getSExtValue(),
+ MVT::i64));
return;
}
}
return true;
}
-// Magic for divide replacement
-
-struct ms {
- int64_t m; // magic number
- int64_t s; // shift amount
-};
-
struct mu {
- uint64_t m; // magic number
- int64_t a; // add indicator
- int64_t s; // shift amount
+ APInt m; // magic number
+ bool a; // add indicator
+ unsigned s; // shift amount
};
-/// magic - calculate the magic numbers required to codegen an integer sdiv as
-/// a sequence of multiply and shifts. Requires that the divisor not be 0, 1,
-/// or -1.
-static ms magic32(int32_t d) {
- int32_t p;
- uint32_t ad, anc, delta, q1, r1, q2, r2, t;
- const uint32_t two31 = 0x80000000U;
- struct ms mag;
-
- ad = abs(d);
- t = two31 + ((uint32_t)d >> 31);
- anc = t - 1 - t%ad; // absolute value of nc
- p = 31; // initialize p
- q1 = two31/anc; // initialize q1 = 2p/abs(nc)
- r1 = two31 - q1*anc; // initialize r1 = rem(2p,abs(nc))
- q2 = two31/ad; // initialize q2 = 2p/abs(d)
- r2 = two31 - q2*ad; // initialize r2 = rem(2p,abs(d))
- do {
- p = p + 1;
- q1 = 2*q1; // update q1 = 2p/abs(nc)
- r1 = 2*r1; // update r1 = rem(2p/abs(nc))
- if (r1 >= anc) { // must be unsigned comparison
- q1 = q1 + 1;
- r1 = r1 - anc;
- }
- q2 = 2*q2; // update q2 = 2p/abs(d)
- r2 = 2*r2; // update r2 = rem(2p/abs(d))
- if (r2 >= ad) { // must be unsigned comparison
- q2 = q2 + 1;
- r2 = r2 - ad;
- }
- delta = ad - r2;
- } while (q1 < delta || (q1 == delta && r1 == 0));
-
- mag.m = (int32_t)(q2 + 1); // make sure to sign extend
- if (d < 0) mag.m = -mag.m; // resulting magic number
- mag.s = p - 32; // resulting shift
- return mag;
-}
-
/// magicu - calculate the magic numbers required to codegen an integer udiv as
/// a sequence of multiply, add and shifts. Requires that the divisor not be 0.
-static mu magicu32(uint32_t d) {
- int32_t p;
- uint32_t nc, delta, q1, r1, q2, r2;
+static mu magicu(const APInt& d) {
+ unsigned p;
+ APInt nc, delta, q1, r1, q2, r2;
struct mu magu;
magu.a = 0; // initialize "add" indicator
- nc = - 1 - (-d)%d;
- p = 31; // initialize p
- q1 = 0x80000000/nc; // initialize q1 = 2p/nc
- r1 = 0x80000000 - q1*nc; // initialize r1 = rem(2p,nc)
- q2 = 0x7FFFFFFF/d; // initialize q2 = (2p-1)/d
- r2 = 0x7FFFFFFF - q2*d; // initialize r2 = rem((2p-1),d)
+ APInt allOnes = APInt::getAllOnesValue(d.getBitWidth());
+ APInt signedMin = APInt::getSignedMinValue(d.getBitWidth());
+ APInt signedMax = APInt::getSignedMaxValue(d.getBitWidth());
+
+ nc = allOnes - (-d).urem(d);
+ p = d.getBitWidth() - 1; // initialize p
+ q1 = signedMin.udiv(nc); // initialize q1 = 2p/nc
+ r1 = signedMin - q1*nc; // initialize r1 = rem(2p,nc)
+ q2 = signedMax.udiv(d); // initialize q2 = (2p-1)/d
+ r2 = signedMax - q2*d; // initialize r2 = rem((2p-1),d)
do {
p = p + 1;
- if (r1 >= nc - r1 ) {
- q1 = 2*q1 + 1; // update q1
- r1 = 2*r1 - nc; // update r1
+ if (r1.uge(nc - r1)) {
+ q1 = q1 + q1 + 1; // update q1
+ r1 = r1 + r1 - nc; // update r1
}
else {
- q1 = 2*q1; // update q1
- r1 = 2*r1; // update r1
+ q1 = q1+q1; // update q1
+ r1 = r1+r1; // update r1
}
- if (r2 + 1 >= d - r2) {
- if (q2 >= 0x7FFFFFFF) magu.a = 1;
- q2 = 2*q2 + 1; // update q2
- r2 = 2*r2 + 1 - d; // update r2
+ if ((r2 + 1).uge(d - r2)) {
+ if (q2.uge(signedMax)) magu.a = 1;
+ q2 = q2+q2 + 1; // update q2
+ r2 = r2+r2 + 1 - d; // update r2
}
else {
- if (q2 >= 0x80000000) magu.a = 1;
- q2 = 2*q2; // update q2
- r2 = 2*r2 + 1; // update r2
+ if (q2.uge(signedMin)) magu.a = 1;
+ q2 = q2+q2; // update q2
+ r2 = r2+r2 + 1; // update r2
}
delta = d - 1 - r2;
- } while (p < 64 && (q1 < delta || (q1 == delta && r1 == 0)));
+ } while (p < d.getBitWidth()*2 &&
+ (q1.ult(delta) || (q1 == delta && r1 == 0)));
magu.m = q2 + 1; // resulting magic number
- magu.s = p - 32; // resulting shift
+ magu.s = p - d.getBitWidth(); // resulting shift
return magu;
}
+// Magic for divide replacement
+struct ms {
+ APInt m; // magic number
+ unsigned s; // shift amount
+};
+
/// magic - calculate the magic numbers required to codegen an integer sdiv as
/// a sequence of multiply and shifts. Requires that the divisor not be 0, 1,
/// or -1.
-static ms magic64(int64_t d) {
- int64_t p;
- uint64_t ad, anc, delta, q1, r1, q2, r2, t;
- const uint64_t two63 = 9223372036854775808ULL; // 2^63
+static ms magic(const APInt& d) {
+ unsigned p;
+ APInt ad, anc, delta, q1, r1, q2, r2, t;
+ APInt allOnes = APInt::getAllOnesValue(d.getBitWidth());
+ APInt signedMin = APInt::getSignedMinValue(d.getBitWidth());
+ APInt signedMax = APInt::getSignedMaxValue(d.getBitWidth());
struct ms mag;
- ad = d >= 0 ? d : -d;
- t = two63 + ((uint64_t)d >> 63);
- anc = t - 1 - t%ad; // absolute value of nc
- p = 63; // initialize p
- q1 = two63/anc; // initialize q1 = 2p/abs(nc)
- r1 = two63 - q1*anc; // initialize r1 = rem(2p,abs(nc))
- q2 = two63/ad; // initialize q2 = 2p/abs(d)
- r2 = two63 - q2*ad; // initialize r2 = rem(2p,abs(d))
+ ad = d.abs();
+ t = signedMin + (d.lshr(d.getBitWidth() - 1));
+ anc = t - 1 - t.urem(ad); // absolute value of nc
+ p = d.getBitWidth() - 1; // initialize p
+ q1 = signedMin.udiv(anc); // initialize q1 = 2p/abs(nc)
+ r1 = signedMin - q1*anc; // initialize r1 = rem(2p,abs(nc))
+ q2 = signedMin.udiv(ad); // initialize q2 = 2p/abs(d)
+ r2 = signedMin - q2*ad; // initialize r2 = rem(2p,abs(d))
do {
p = p + 1;
- q1 = 2*q1; // update q1 = 2p/abs(nc)
- r1 = 2*r1; // update r1 = rem(2p/abs(nc))
- if (r1 >= anc) { // must be unsigned comparison
+ q1 = q1<<1; // update q1 = 2p/abs(nc)
+ r1 = r1<<1; // update r1 = rem(2p/abs(nc))
+ if (r1.uge(anc)) { // must be unsigned comparison
q1 = q1 + 1;
r1 = r1 - anc;
}
- q2 = 2*q2; // update q2 = 2p/abs(d)
- r2 = 2*r2; // update r2 = rem(2p/abs(d))
- if (r2 >= ad) { // must be unsigned comparison
+ q2 = q2<<1; // update q2 = 2p/abs(d)
+ r2 = r2<<1; // update r2 = rem(2p/abs(d))
+ if (r2.uge(ad)) { // must be unsigned comparison
q2 = q2 + 1;
r2 = r2 - ad;
}
delta = ad - r2;
- } while (q1 < delta || (q1 == delta && r1 == 0));
+ } while (q1.ule(delta) || (q1 == delta && r1 == 0));
mag.m = q2 + 1;
- if (d < 0) mag.m = -mag.m; // resulting magic number
- mag.s = p - 64; // resulting shift
+ if (d.isNegative()) mag.m = -mag.m; // resulting magic number
+ mag.s = p - d.getBitWidth(); // resulting shift
return mag;
}
-/// magicu - calculate the magic numbers required to codegen an integer udiv as
-/// a sequence of multiply, add and shifts. Requires that the divisor not be 0.
-static mu magicu64(uint64_t d)
-{
- int64_t p;
- uint64_t nc, delta, q1, r1, q2, r2;
- struct mu magu;
- magu.a = 0; // initialize "add" indicator
- nc = - 1 - (-d)%d;
- p = 63; // initialize p
- q1 = 0x8000000000000000ull/nc; // initialize q1 = 2p/nc
- r1 = 0x8000000000000000ull - q1*nc; // initialize r1 = rem(2p,nc)
- q2 = 0x7FFFFFFFFFFFFFFFull/d; // initialize q2 = (2p-1)/d
- r2 = 0x7FFFFFFFFFFFFFFFull - q2*d; // initialize r2 = rem((2p-1),d)
- do {
- p = p + 1;
- if (r1 >= nc - r1 ) {
- q1 = 2*q1 + 1; // update q1
- r1 = 2*r1 - nc; // update r1
- }
- else {
- q1 = 2*q1; // update q1
- r1 = 2*r1; // update r1
- }
- if (r2 + 1 >= d - r2) {
- if (q2 >= 0x7FFFFFFFFFFFFFFFull) magu.a = 1;
- q2 = 2*q2 + 1; // update q2
- r2 = 2*r2 + 1 - d; // update r2
- }
- else {
- if (q2 >= 0x8000000000000000ull) magu.a = 1;
- q2 = 2*q2; // update q2
- r2 = 2*r2 + 1; // update r2
- }
- delta = d - 1 - r2;
- } while (p < 128 && (q1 < delta || (q1 == delta && r1 == 0)));
- magu.m = q2 + 1; // resulting magic number
- magu.s = p - 64; // resulting shift
- return magu;
-}
-
/// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant,
/// return a DAG expression to select that will generate the same value by
/// multiplying by a magic number. See:
SDValue TargetLowering::BuildSDIV(SDNode *N, SelectionDAG &DAG,
std::vector<SDNode*>* Created) const {
MVT VT = N->getValueType(0);
+ DebugLoc dl= N->getDebugLoc();
// Check to see if we can do this.
- if (!isTypeLegal(VT) || (VT != MVT::i32 && VT != MVT::i64))
- return SDValue(); // BuildSDIV only operates on i32 or i64
+ // FIXME: We should be more aggressive here.
+ if (!isTypeLegal(VT))
+ return SDValue();
- int64_t d = cast<ConstantSDNode>(N->getOperand(1))->getSExtValue();
- ms magics = (VT == MVT::i32) ? magic32(d) : magic64(d);
+ APInt d = cast<ConstantSDNode>(N->getOperand(1))->getAPIntValue();
+ ms magics = magic(d);
// Multiply the numerator (operand 0) by the magic value
+ // FIXME: We should support doing a MUL in a wider type
SDValue Q;
- if (isOperationLegal(ISD::MULHS, VT))
- Q = DAG.getNode(ISD::MULHS, VT, N->getOperand(0),
+ if (isOperationLegalOrCustom(ISD::MULHS, VT))
+ Q = DAG.getNode(ISD::MULHS, dl, VT, N->getOperand(0),
DAG.getConstant(magics.m, VT));
- else if (isOperationLegal(ISD::SMUL_LOHI, VT))
- Q = SDValue(DAG.getNode(ISD::SMUL_LOHI, DAG.getVTList(VT, VT),
+ else if (isOperationLegalOrCustom(ISD::SMUL_LOHI, VT))
+ Q = SDValue(DAG.getNode(ISD::SMUL_LOHI, dl, DAG.getVTList(VT, VT),
N->getOperand(0),
DAG.getConstant(magics.m, VT)).getNode(), 1);
else
return SDValue(); // No mulhs or equvialent
// If d > 0 and m < 0, add the numerator
- if (d > 0 && magics.m < 0) {
- Q = DAG.getNode(ISD::ADD, VT, Q, N->getOperand(0));
+ if (d.isStrictlyPositive() && magics.m.isNegative()) {
+ Q = DAG.getNode(ISD::ADD, dl, VT, Q, N->getOperand(0));
if (Created)
Created->push_back(Q.getNode());
}
// If d < 0 and m > 0, subtract the numerator.
- if (d < 0 && magics.m > 0) {
- Q = DAG.getNode(ISD::SUB, VT, Q, N->getOperand(0));
+ if (d.isNegative() && magics.m.isStrictlyPositive()) {
+ Q = DAG.getNode(ISD::SUB, dl, VT, Q, N->getOperand(0));
if (Created)
Created->push_back(Q.getNode());
}
// Shift right algebraic if shift value is nonzero
if (magics.s > 0) {
- Q = DAG.getNode(ISD::SRA, VT, Q,
+ Q = DAG.getNode(ISD::SRA, dl, VT, Q,
DAG.getConstant(magics.s, getShiftAmountTy()));
if (Created)
Created->push_back(Q.getNode());
}
// Extract the sign bit and add it to the quotient
SDValue T =
- DAG.getNode(ISD::SRL, VT, Q, DAG.getConstant(VT.getSizeInBits()-1,
+ DAG.getNode(ISD::SRL, dl, VT, Q, DAG.getConstant(VT.getSizeInBits()-1,
getShiftAmountTy()));
if (Created)
Created->push_back(T.getNode());
- return DAG.getNode(ISD::ADD, VT, Q, T);
+ return DAG.getNode(ISD::ADD, dl, VT, Q, T);
}
/// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant,
SDValue TargetLowering::BuildUDIV(SDNode *N, SelectionDAG &DAG,
std::vector<SDNode*>* Created) const {
MVT VT = N->getValueType(0);
-
+ DebugLoc dl = N->getDebugLoc();
+
// Check to see if we can do this.
- if (!isTypeLegal(VT) || (VT != MVT::i32 && VT != MVT::i64))
- return SDValue(); // BuildUDIV only operates on i32 or i64
-
- uint64_t d = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
- mu magics = (VT == MVT::i32) ? magicu32(d) : magicu64(d);
-
+ // FIXME: We should be more aggressive here.
+ if (!isTypeLegal(VT))
+ return SDValue();
+
+ // FIXME: We should use a narrower constant when the upper
+ // bits are known to be zero.
+ ConstantSDNode *N1C = cast<ConstantSDNode>(N->getOperand(1));
+ mu magics = magicu(N1C->getAPIntValue());
+
// Multiply the numerator (operand 0) by the magic value
+ // FIXME: We should support doing a MUL in a wider type
SDValue Q;
- if (isOperationLegal(ISD::MULHU, VT))
- Q = DAG.getNode(ISD::MULHU, VT, N->getOperand(0),
+ if (isOperationLegalOrCustom(ISD::MULHU, VT))
+ Q = DAG.getNode(ISD::MULHU, dl, VT, N->getOperand(0),
DAG.getConstant(magics.m, VT));
- else if (isOperationLegal(ISD::UMUL_LOHI, VT))
- Q = SDValue(DAG.getNode(ISD::UMUL_LOHI, DAG.getVTList(VT, VT),
+ else if (isOperationLegalOrCustom(ISD::UMUL_LOHI, VT))
+ Q = SDValue(DAG.getNode(ISD::UMUL_LOHI, dl, DAG.getVTList(VT, VT),
N->getOperand(0),
DAG.getConstant(magics.m, VT)).getNode(), 1);
else
Created->push_back(Q.getNode());
if (magics.a == 0) {
- return DAG.getNode(ISD::SRL, VT, Q,
+ assert(magics.s < N1C->getAPIntValue().getBitWidth() &&
+ "We shouldn't generate an undefined shift!");
+ return DAG.getNode(ISD::SRL, dl, VT, Q,
DAG.getConstant(magics.s, getShiftAmountTy()));
} else {
- SDValue NPQ = DAG.getNode(ISD::SUB, VT, N->getOperand(0), Q);
+ SDValue NPQ = DAG.getNode(ISD::SUB, dl, VT, N->getOperand(0), Q);
if (Created)
Created->push_back(NPQ.getNode());
- NPQ = DAG.getNode(ISD::SRL, VT, NPQ,
+ NPQ = DAG.getNode(ISD::SRL, dl, VT, NPQ,
DAG.getConstant(1, getShiftAmountTy()));
if (Created)
Created->push_back(NPQ.getNode());
- NPQ = DAG.getNode(ISD::ADD, VT, NPQ, Q);
+ NPQ = DAG.getNode(ISD::ADD, dl, VT, NPQ, Q);
if (Created)
Created->push_back(NPQ.getNode());
- return DAG.getNode(ISD::SRL, VT, NPQ,
+ return DAG.getNode(ISD::SRL, dl, VT, NPQ,
DAG.getConstant(magics.s-1, getShiftAmountTy()));
}
}
projects / oota-llvm.git / blobdiff