static SDValue LowerVAARG(SDValue Op, SelectionDAG &DAG) {
SDNode *Node = Op.getNode();
EVT VT = Node->getValueType(0);
SDValue InChain = Node->getOperand(0);
SDValue VAListPtr = Node->getOperand(1);
const Value *SV = cast<SrcValueSDNode>(Node->getOperand(2))->getValue();
DebugLoc dl = Node->getDebugLoc();
SDValue VAList = DAG.getLoad(MVT::i32, dl, InChain, VAListPtr, SV, 0,
false, false, 0);
// Increment the pointer, VAList, to the next vaarg
SDValue NextPtr = DAG.getNode(ISD::ADD, dl, MVT::i32, VAList,
DAG.getConstant(VT.getSizeInBits()/8,
MVT::i32));
// Store the incremented VAList to the legalized pointer
InChain = DAG.getStore(VAList.getValue(1), dl, NextPtr,
VAListPtr, SV, 0, false, false, 0);
// Load the actual argument out of the pointer VAList, unless this is an
// f64 load.
if (VT != MVT::f64)
return DAG.getLoad(VT, dl, InChain, VAList, NULL, 0, false, false, 0);
// Otherwise, load it as i64, then do a bitconvert.
SDValue V = DAG.getLoad(MVT::i64, dl, InChain, VAList, NULL, 0,
false, false, 0);
// Bit-Convert the value to f64.
SDValue Ops[2] = {
DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, V),
V.getValue(1)
};
return DAG.getMergeValues(Ops, 2, dl);
}
SDValue MSP430TargetLowering::LowerShifts(SDValue Op,
SelectionDAG &DAG) {
unsigned Opc = Op.getOpcode();
SDNode* N = Op.getNode();
MVT VT = Op.getValueType();
DebugLoc dl = N->getDebugLoc();
// We currently only lower shifts of constant argument.
if (!isa<ConstantSDNode>(N->getOperand(1)))
return SDValue();
uint64_t ShiftAmount = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
// Expand the stuff into sequence of shifts.
// FIXME: for some shift amounts this might be done better!
// E.g.: foo >> (8 + N) => sxt(swpb(foo)) >> N
SDValue Victim = N->getOperand(0);
if (Opc == ISD::SRL && ShiftAmount) {
// Emit a special goodness here:
// srl A, 1 => clrc; rrc A
Victim = DAG.getNode(MSP430ISD::RRC, dl, VT, Victim);
ShiftAmount -= 1;
}
while (ShiftAmount--)
Victim = DAG.getNode((Opc == ISD::SHL ? MSP430ISD::RLA : MSP430ISD::RRA),
dl, VT, Victim);
return Victim;
}
SDNode *SparcDAGToDAGISel::Select(SDValue Op) {
SDNode *N = Op.getNode();
DebugLoc dl = N->getDebugLoc();
if (N->isMachineOpcode())
return NULL; // Already selected.
switch (N->getOpcode()) {
default: break;
case SPISD::GLOBAL_BASE_REG:
return getGlobalBaseReg();
case ISD::SDIV:
case ISD::UDIV: {
// FIXME: should use a custom expander to expose the SRA to the dag.
SDValue DivLHS = N->getOperand(0);
SDValue DivRHS = N->getOperand(1);
// Set the Y register to the high-part.
SDValue TopPart;
if (N->getOpcode() == ISD::SDIV) {
TopPart = SDValue(CurDAG->getMachineNode(SP::SRAri, dl, MVT::i32, DivLHS,
CurDAG->getTargetConstant(31, MVT::i32)), 0);
} else {
TopPart = CurDAG->getRegister(SP::G0, MVT::i32);
}
TopPart = SDValue(CurDAG->getMachineNode(SP::WRYrr, dl, MVT::Flag, TopPart,
CurDAG->getRegister(SP::G0, MVT::i32)), 0);
// FIXME: Handle div by immediate.
unsigned Opcode = N->getOpcode() == ISD::SDIV ? SP::SDIVrr : SP::UDIVrr;
return CurDAG->SelectNodeTo(N, Opcode, MVT::i32, DivLHS, DivRHS,
TopPart);
}
case ISD::MULHU:
case ISD::MULHS: {
// FIXME: Handle mul by immediate.
SDValue MulLHS = N->getOperand(0);
SDValue MulRHS = N->getOperand(1);
unsigned Opcode = N->getOpcode() == ISD::MULHU ? SP::UMULrr : SP::SMULrr;
SDNode *Mul = CurDAG->getMachineNode(Opcode, dl, MVT::i32, MVT::Flag,
MulLHS, MulRHS);
// The high part is in the Y register.
return CurDAG->SelectNodeTo(N, SP::RDY, MVT::i32, SDValue(Mul, 1));
return NULL;
}
}
return SelectCode(Op);
}
SDValue MSP430TargetLowering::LowerShifts(SDValue Op,
SelectionDAG &DAG) const {
unsigned Opc = Op.getOpcode();
SDNode* N = Op.getNode();
EVT VT = Op.getValueType();
DebugLoc dl = N->getDebugLoc();
// Expand non-constant shifts to loops:
if (!isa<ConstantSDNode>(N->getOperand(1)))
switch (Opc) {
default:
assert(0 && "Invalid shift opcode!");
case ISD::SHL:
return DAG.getNode(MSP430ISD::SHL, dl,
VT, N->getOperand(0), N->getOperand(1));
case ISD::SRA:
return DAG.getNode(MSP430ISD::SRA, dl,
VT, N->getOperand(0), N->getOperand(1));
case ISD::SRL:
return DAG.getNode(MSP430ISD::SRL, dl,
VT, N->getOperand(0), N->getOperand(1));
}
uint64_t ShiftAmount = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
// Expand the stuff into sequence of shifts.
// FIXME: for some shift amounts this might be done better!
// E.g.: foo >> (8 + N) => sxt(swpb(foo)) >> N
SDValue Victim = N->getOperand(0);
if (Opc == ISD::SRL && ShiftAmount) {
// Emit a special goodness here:
// srl A, 1 => clrc; rrc A
Victim = DAG.getNode(MSP430ISD::RRC, dl, VT, Victim);
ShiftAmount -= 1;
}
while (ShiftAmount--)
Victim = DAG.getNode((Opc == ISD::SHL ? MSP430ISD::RLA : MSP430ISD::RRA),
dl, VT, Victim);
return Victim;
}
// By default CONCAT_VECTORS is lowered by ExpandVectorBuildThroughStack()
// (see LegalizeDAG.cpp). This is slow and uses local memory.
// We use extract/insert/build vector just as what LegalizeOp() does in llvm 2.5
SDValue NVPTXTargetLowering::
LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const {
SDNode *Node = Op.getNode();
DebugLoc dl = Node->getDebugLoc();
SmallVector<SDValue, 8> Ops;
unsigned NumOperands = Node->getNumOperands();
for (unsigned i=0; i < NumOperands; ++i) {
SDValue SubOp = Node->getOperand(i);
EVT VVT = SubOp.getNode()->getValueType(0);
EVT EltVT = VVT.getVectorElementType();
unsigned NumSubElem = VVT.getVectorNumElements();
for (unsigned j=0; j < NumSubElem; ++j) {
Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, SubOp,
DAG.getIntPtrConstant(j)));
}
}
return DAG.getNode(ISD::BUILD_VECTOR, dl, Node->getValueType(0),
&Ops[0], Ops.size());
}
SDValue AVRTargetLowering::LowerShifts(SDValue Op,
SelectionDAG &DAG) const {
unsigned Opc = Op.getOpcode();
SDNode* N = Op.getNode();
EVT VT = Op.getValueType();
DebugLoc dl = N->getDebugLoc();
// Expand non-constant shifts to loops:
if (!isa<ConstantSDNode>(N->getOperand(1)))
switch (Opc) {
default:
assert(0 && "Invalid shift opcode!");
case ISD::SHL:
return DAG.getNode(AVRISD::SHL, dl,
VT, N->getOperand(0), N->getOperand(1));
case ISD::SRA:
return DAG.getNode(AVRISD::SRA, dl,
VT, N->getOperand(0), N->getOperand(1));
case ISD::SRL:
return DAG.getNode(AVRISD::SRL, dl,
VT, N->getOperand(0), N->getOperand(1));
}
uint64_t ShiftAmount = N->getConstantOperandVal(1);
// Expand the stuff into sequence of shifts.
// FIXME: for some shift amounts this might be done better!
// E.g.: foo >> (8 + N) => sxt(swpb(foo)) >> N
SDValue Victim = N->getOperand(0);
unsigned int TargetOpcode = 0;
switch(Opc) {
case ISD::SHL: TargetOpcode = AVRISD::SHLC; break;
case ISD::SRA: TargetOpcode = AVRISD::SRAC; break;
case ISD::SRL: TargetOpcode = AVRISD::SRLC; break;
}
while (ShiftAmount--)
Victim = DAG.getNode(TargetOpcode, dl, VT, Victim);
return Victim;
}
SDNode *IA64DAGToDAGISel::SelectDIV(SDValue Op) {
SDNode *N = Op.getNode();
SDValue Chain = N->getOperand(0);
SDValue Tmp1 = N->getOperand(0);
SDValue Tmp2 = N->getOperand(1);
DebugLoc dl = N->getDebugLoc();
bool isFP=false;
if(Tmp1.getValueType().isFloatingPoint())
isFP=true;
bool isModulus=false; // is it a division or a modulus?
bool isSigned=false;
switch(N->getOpcode()) {
case ISD::FDIV:
case ISD::SDIV:
isModulus=false;
isSigned=true;
break;
case ISD::UDIV:
isModulus=false;
isSigned=false;
break;
case ISD::FREM:
case ISD::SREM:
isModulus=true;
isSigned=true;
break;
case ISD::UREM:
isModulus=true;
isSigned=false;
break;
}
// TODO: check for integer divides by powers of 2 (or other simple patterns?)
SDValue TmpPR, TmpPR2;
SDValue TmpF1, TmpF2, TmpF3, TmpF4, TmpF5, TmpF6, TmpF7, TmpF8;
SDValue TmpF9, TmpF10,TmpF11,TmpF12,TmpF13,TmpF14,TmpF15;
SDNode *Result;
// we'll need copies of F0 and F1
SDValue F0 = CurDAG->getRegister(IA64::F0, MVT::f64);
SDValue F1 = CurDAG->getRegister(IA64::F1, MVT::f64);
// OK, emit some code:
if(!isFP) {
// first, load the inputs into FP regs.
TmpF1 =
SDValue(CurDAG->getTargetNode(IA64::SETFSIG, dl, MVT::f64, Tmp1), 0);
Chain = TmpF1.getValue(1);
TmpF2 =
SDValue(CurDAG->getTargetNode(IA64::SETFSIG, dl, MVT::f64, Tmp2), 0);
Chain = TmpF2.getValue(1);
// next, convert the inputs to FP
if(isSigned) {
TmpF3 =
SDValue(CurDAG->getTargetNode(IA64::FCVTXF, dl, MVT::f64, TmpF1), 0);
Chain = TmpF3.getValue(1);
TmpF4 =
SDValue(CurDAG->getTargetNode(IA64::FCVTXF, dl, MVT::f64, TmpF2), 0);
Chain = TmpF4.getValue(1);
} else { // is unsigned
TmpF3 =
SDValue(CurDAG->getTargetNode(IA64::FCVTXUFS1, dl, MVT::f64, TmpF1),
0);
Chain = TmpF3.getValue(1);
TmpF4 =
SDValue(CurDAG->getTargetNode(IA64::FCVTXUFS1, dl, MVT::f64, TmpF2),
0);
Chain = TmpF4.getValue(1);
}
} else { // this is an FP divide/remainder, so we 'leak' some temp
// regs and assign TmpF3=Tmp1, TmpF4=Tmp2
TmpF3=Tmp1;
TmpF4=Tmp2;
}
// we start by computing an approximate reciprocal (good to 9 bits?)
// note, this instruction writes _both_ TmpF5 (answer) and TmpPR (predicate)
if(isFP)
TmpF5 = SDValue(CurDAG->getTargetNode(IA64::FRCPAS0, dl, MVT::f64,
MVT::i1, TmpF3, TmpF4), 0);
else
TmpF5 = SDValue(CurDAG->getTargetNode(IA64::FRCPAS1, dl, MVT::f64,
MVT::i1, TmpF3, TmpF4), 0);
TmpPR = TmpF5.getValue(1);
Chain = TmpF5.getValue(2);
SDValue minusB;
if(isModulus) { // for remainders, it'll be handy to have
// copies of -input_b
minusB = SDValue(CurDAG->getTargetNode(IA64::SUB, dl, MVT::i64,
CurDAG->getRegister(IA64::r0, MVT::i64), Tmp2), 0);
Chain = minusB.getValue(1);
}
SDValue TmpE0, TmpY1, TmpE1, TmpY2;
SDValue OpsE0[] = { TmpF4, TmpF5, F1, TmpPR };
TmpE0 = SDValue(CurDAG->getTargetNode(IA64::CFNMAS1, dl, MVT::f64,
OpsE0, 4), 0);
Chain = TmpE0.getValue(1);
SDValue OpsY1[] = { TmpF5, TmpE0, TmpF5, TmpPR };
TmpY1 = SDValue(CurDAG->getTargetNode(IA64::CFMAS1, dl, MVT::f64,
OpsY1, 4), 0);
Chain = TmpY1.getValue(1);
SDValue OpsE1[] = { TmpE0, TmpE0, F0, TmpPR };
TmpE1 = SDValue(CurDAG->getTargetNode(IA64::CFMAS1, dl, MVT::f64,
OpsE1, 4), 0);
Chain = TmpE1.getValue(1);
SDValue OpsY2[] = { TmpY1, TmpE1, TmpY1, TmpPR };
TmpY2 = SDValue(CurDAG->getTargetNode(IA64::CFMAS1, dl, MVT::f64,
OpsY2, 4), 0);
Chain = TmpY2.getValue(1);
if(isFP) { // if this is an FP divide, we finish up here and exit early
if(isModulus)
assert(0 && "Sorry, try another FORTRAN compiler.");
SDValue TmpE2, TmpY3, TmpQ0, TmpR0;
SDValue OpsE2[] = { TmpE1, TmpE1, F0, TmpPR };
TmpE2 = SDValue(CurDAG->getTargetNode(IA64::CFMAS1, dl, MVT::f64,
OpsE2, 4), 0);
Chain = TmpE2.getValue(1);
SDValue OpsY3[] = { TmpY2, TmpE2, TmpY2, TmpPR };
TmpY3 = SDValue(CurDAG->getTargetNode(IA64::CFMAS1, dl, MVT::f64,
OpsY3, 4), 0);
Chain = TmpY3.getValue(1);
SDValue OpsQ0[] = { Tmp1, TmpY3, F0, TmpPR };
TmpQ0 =
SDValue(CurDAG->getTargetNode(IA64::CFMADS1, dl, // double prec!
MVT::f64, OpsQ0, 4), 0);
Chain = TmpQ0.getValue(1);
SDValue OpsR0[] = { Tmp2, TmpQ0, Tmp1, TmpPR };
TmpR0 =
SDValue(CurDAG->getTargetNode(IA64::CFNMADS1, dl, // double prec!
MVT::f64, OpsR0, 4), 0);
Chain = TmpR0.getValue(1);
// we want Result to have the same target register as the frcpa, so
// we two-address hack it. See the comment "for this to work..." on
// page 48 of Intel application note #245415
SDValue Ops[] = { TmpF5, TmpY3, TmpR0, TmpQ0, TmpPR };
Result = CurDAG->getTargetNode(IA64::TCFMADS0, dl, // d.p. s0 rndg!
MVT::f64, Ops, 5);
Chain = SDValue(Result, 1);
return Result; // XXX: early exit!
} else { // this is *not* an FP divide, so there's a bit left to do:
SDValue TmpQ2, TmpR2, TmpQ3, TmpQ;
SDValue OpsQ2[] = { TmpF3, TmpY2, F0, TmpPR };
TmpQ2 = SDValue(CurDAG->getTargetNode(IA64::CFMAS1, dl, MVT::f64,
OpsQ2, 4), 0);
Chain = TmpQ2.getValue(1);
SDValue OpsR2[] = { TmpF4, TmpQ2, TmpF3, TmpPR };
TmpR2 = SDValue(CurDAG->getTargetNode(IA64::CFNMAS1, dl, MVT::f64,
OpsR2, 4), 0);
Chain = TmpR2.getValue(1);
// we want TmpQ3 to have the same target register as the frcpa? maybe we
// should two-address hack it. See the comment "for this to work..." on page
// 48 of Intel application note #245415
SDValue OpsQ3[] = { TmpF5, TmpR2, TmpY2, TmpQ2, TmpPR };
TmpQ3 = SDValue(CurDAG->getTargetNode(IA64::TCFMAS1, dl, MVT::f64,
OpsQ3, 5), 0);
Chain = TmpQ3.getValue(1);
// STORY: without these two-address instructions (TCFMAS1 and TCFMADS0)
// the FPSWA won't be able to help out in the case of large/tiny
// arguments. Other fun bugs may also appear, e.g. 0/x = x, not 0.
if(isSigned)
TmpQ = SDValue(CurDAG->getTargetNode(IA64::FCVTFXTRUNCS1, dl,
MVT::f64, TmpQ3), 0);
else
TmpQ = SDValue(CurDAG->getTargetNode(IA64::FCVTFXUTRUNCS1, dl,
MVT::f64, TmpQ3), 0);
Chain = TmpQ.getValue(1);
if(isModulus) {
SDValue FPminusB =
SDValue(CurDAG->getTargetNode(IA64::SETFSIG, dl, MVT::f64, minusB),
0);
Chain = FPminusB.getValue(1);
SDValue Remainder =
SDValue(CurDAG->getTargetNode(IA64::XMAL, dl, MVT::f64,
TmpQ, FPminusB, TmpF1), 0);
Chain = Remainder.getValue(1);
Result = CurDAG->getTargetNode(IA64::GETFSIG, dl, MVT::i64, Remainder);
Chain = SDValue(Result, 1);
} else { // just an integer divide
Result = CurDAG->getTargetNode(IA64::GETFSIG, dl, MVT::i64, TmpQ);
Chain = SDValue(Result, 1);
}
return Result;
} // wasn't an FP divide
}
SDNode *XCoreDAGToDAGISel::Select(SDValue Op) {
SDNode *N = Op.getNode();
DebugLoc dl = N->getDebugLoc();
EVT NVT = N->getValueType(0);
if (NVT == MVT::i32) {
switch (N->getOpcode()) {
default: break;
case ISD::Constant: {
if (Predicate_immMskBitp(N)) {
SDValue MskSize = Transform_msksize_xform(N);
return CurDAG->getTargetNode(XCore::MKMSK_rus, dl, MVT::i32, MskSize);
}
else if (! Predicate_immU16(N)) {
unsigned Val = cast<ConstantSDNode>(N)->getZExtValue();
SDValue CPIdx =
CurDAG->getTargetConstantPool(ConstantInt::get(
Type::getInt32Ty(*CurDAG->getContext()), Val),
TLI.getPointerTy());
return CurDAG->getTargetNode(XCore::LDWCP_lru6, dl, MVT::i32,
MVT::Other, CPIdx,
CurDAG->getEntryNode());
}
break;
}
case ISD::SMUL_LOHI: {
// FIXME fold addition into the macc instruction
if (!Subtarget.isXS1A()) {
SDValue Zero(CurDAG->getTargetNode(XCore::LDC_ru6, dl, MVT::i32,
CurDAG->getTargetConstant(0, MVT::i32)), 0);
SDValue Ops[] = { Zero, Zero, Op.getOperand(0), Op.getOperand(1) };
SDNode *ResNode = CurDAG->getTargetNode(XCore::MACCS_l4r, dl,
MVT::i32, MVT::i32, Ops, 4);
ReplaceUses(SDValue(N, 0), SDValue(ResNode, 1));
ReplaceUses(SDValue(N, 1), SDValue(ResNode, 0));
return NULL;
}
break;
}
case ISD::UMUL_LOHI: {
// FIXME fold addition into the macc / lmul instruction
SDValue Zero(CurDAG->getTargetNode(XCore::LDC_ru6, dl, MVT::i32,
CurDAG->getTargetConstant(0, MVT::i32)), 0);
SDValue Ops[] = { Op.getOperand(0), Op.getOperand(1),
Zero, Zero };
SDNode *ResNode = CurDAG->getTargetNode(XCore::LMUL_l6r, dl, MVT::i32,
MVT::i32, Ops, 4);
ReplaceUses(SDValue(N, 0), SDValue(ResNode, 1));
ReplaceUses(SDValue(N, 1), SDValue(ResNode, 0));
return NULL;
}
case XCoreISD::LADD: {
if (!Subtarget.isXS1A()) {
SDValue Ops[] = { Op.getOperand(0), Op.getOperand(1),
Op.getOperand(2) };
return CurDAG->getTargetNode(XCore::LADD_l5r, dl, MVT::i32, MVT::i32,
Ops, 3);
}
break;
}
case XCoreISD::LSUB: {
if (!Subtarget.isXS1A()) {
SDValue Ops[] = { Op.getOperand(0), Op.getOperand(1),
Op.getOperand(2) };
return CurDAG->getTargetNode(XCore::LSUB_l5r, dl, MVT::i32, MVT::i32,
Ops, 3);
}
break;
}
// Other cases are autogenerated.
}
}
return SelectCode(Op);
}
