// There is no native floating point division, but we can convert this to a
// reciprocal/multiply operation. If the first parameter is constant 1.0, then
// just a reciprocal will suffice.
SDValue
VectorProcTargetLowering::LowerFDIV(SDValue Op, SelectionDAG &DAG) const
{
DebugLoc dl = Op.getDebugLoc();
EVT type = Op.getOperand(1).getValueType();
SDValue two = DAG.getConstantFP(2.0, type);
SDValue denom = Op.getOperand(1);
SDValue estimate = DAG.getNode(VectorProcISD::RECIPROCAL_EST, dl, type, denom);
// Perform a series of newton/raphson refinements. Each iteration doubles
// the precision. The initial estimate has 6 bits of precision, so two iteration
// results in 24 bits, which is larger than the significand.
for (int i = 0; i < 2; i++)
{
// trial = x * estimate (ideally, x * 1/x should be 1.0)
// error = 2.0 - trial
// estimate = estimate * error
SDValue trial = DAG.getNode(ISD::FMUL, dl, type, estimate, denom);
SDValue error = DAG.getNode(ISD::FSUB, dl, type, two, trial);
estimate = DAG.getNode(ISD::FMUL, dl, type, estimate, error);
}
// Check if the first parameter is constant 1.0. If so, we don't need
// to multiply.
bool isOne = false;
if (type.isVector())
{
if (isSplatVector(Op.getOperand(0).getNode()))
{
ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Op.getOperand(0).getOperand(0));
isOne = C && C->isExactlyValue(1.0);
}
}
else
{
ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Op.getOperand(0));
isOne = C && C->isExactlyValue(1.0);
}
if (!isOne)
estimate = DAG.getNode(ISD::FMUL, dl, type, Op.getOperand(0), estimate);
return estimate;
}
std::pair<bool, SDNode*> MipsSEDAGToDAGISel::selectNode(SDNode *Node) {
unsigned Opcode = Node->getOpcode();
SDLoc DL(Node);
///
// Instruction Selection not handled by the auto-generated
// tablegen selection should be handled here.
///
SDNode *Result;
switch(Opcode) {
default: break;
case ISD::SUBE: {
SDValue InFlag = Node->getOperand(2);
Result = selectAddESubE(Mips::SUBu, InFlag, InFlag.getOperand(0), DL, Node);
return std::make_pair(true, Result);
}
case ISD::ADDE: {
if (Subtarget.hasDSP()) // Select DSP instructions, ADDSC and ADDWC.
break;
SDValue InFlag = Node->getOperand(2);
Result = selectAddESubE(Mips::ADDu, InFlag, InFlag.getValue(0), DL, Node);
return std::make_pair(true, Result);
}
case ISD::ConstantFP: {
ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(Node);
if (Node->getValueType(0) == MVT::f64 && CN->isExactlyValue(+0.0)) {
if (Subtarget.hasMips64()) {
SDValue Zero = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), DL,
Mips::ZERO_64, MVT::i64);
Result = CurDAG->getMachineNode(Mips::DMTC1, DL, MVT::f64, Zero);
} else {
SDValue Zero = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), DL,
Mips::ZERO, MVT::i32);
Result = CurDAG->getMachineNode(Mips::BuildPairF64, DL, MVT::f64, Zero,
Zero);
}
return std::make_pair(true, Result);
}
break;
}
case ISD::Constant: {
const ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Node);
unsigned Size = CN->getValueSizeInBits(0);
if (Size == 32)
break;
MipsAnalyzeImmediate AnalyzeImm;
int64_t Imm = CN->getSExtValue();
const MipsAnalyzeImmediate::InstSeq &Seq =
AnalyzeImm.Analyze(Imm, Size, false);
MipsAnalyzeImmediate::InstSeq::const_iterator Inst = Seq.begin();
SDLoc DL(CN);
SDNode *RegOpnd;
SDValue ImmOpnd = CurDAG->getTargetConstant(SignExtend64<16>(Inst->ImmOpnd),
MVT::i64);
// The first instruction can be a LUi which is different from other
// instructions (ADDiu, ORI and SLL) in that it does not have a register
// operand.
if (Inst->Opc == Mips::LUi64)
RegOpnd = CurDAG->getMachineNode(Inst->Opc, DL, MVT::i64, ImmOpnd);
else
RegOpnd =
CurDAG->getMachineNode(Inst->Opc, DL, MVT::i64,
CurDAG->getRegister(Mips::ZERO_64, MVT::i64),
ImmOpnd);
// The remaining instructions in the sequence are handled here.
for (++Inst; Inst != Seq.end(); ++Inst) {
ImmOpnd = CurDAG->getTargetConstant(SignExtend64<16>(Inst->ImmOpnd),
MVT::i64);
RegOpnd = CurDAG->getMachineNode(Inst->Opc, DL, MVT::i64,
SDValue(RegOpnd, 0), ImmOpnd);
}
return std::make_pair(true, RegOpnd);
}
case ISD::INTRINSIC_W_CHAIN: {
switch (cast<ConstantSDNode>(Node->getOperand(1))->getZExtValue()) {
default:
break;
case Intrinsic::mips_cfcmsa: {
SDValue ChainIn = Node->getOperand(0);
SDValue RegIdx = Node->getOperand(2);
SDValue Reg = CurDAG->getCopyFromReg(ChainIn, DL,
getMSACtrlReg(RegIdx), MVT::i32);
return std::make_pair(true, Reg.getNode());
}
}
break;
}
case ISD::INTRINSIC_WO_CHAIN: {
switch (cast<ConstantSDNode>(Node->getOperand(0))->getZExtValue()) {
default:
break;
case Intrinsic::mips_move_v:
// Like an assignment but will always produce a move.v even if
// unnecessary.
return std::make_pair(true,
CurDAG->getMachineNode(Mips::MOVE_V, DL,
Node->getValueType(0),
Node->getOperand(1)));
}
break;
}
case ISD::INTRINSIC_VOID: {
switch (cast<ConstantSDNode>(Node->getOperand(1))->getZExtValue()) {
default:
break;
case Intrinsic::mips_ctcmsa: {
SDValue ChainIn = Node->getOperand(0);
SDValue RegIdx = Node->getOperand(2);
SDValue Value = Node->getOperand(3);
SDValue ChainOut = CurDAG->getCopyToReg(ChainIn, DL,
getMSACtrlReg(RegIdx), Value);
return std::make_pair(true, ChainOut.getNode());
}
}
break;
}
case MipsISD::ThreadPointer: {
EVT PtrVT = getTargetLowering()->getPointerTy();
unsigned RdhwrOpc, DestReg;
if (PtrVT == MVT::i32) {
RdhwrOpc = Mips::RDHWR;
DestReg = Mips::V1;
} else {
RdhwrOpc = Mips::RDHWR64;
DestReg = Mips::V1_64;
}
SDNode *Rdhwr =
CurDAG->getMachineNode(RdhwrOpc, SDLoc(Node),
Node->getValueType(0),
CurDAG->getRegister(Mips::HWR29, MVT::i32));
SDValue Chain = CurDAG->getCopyToReg(CurDAG->getEntryNode(), DL, DestReg,
SDValue(Rdhwr, 0));
SDValue ResNode = CurDAG->getCopyFromReg(Chain, DL, DestReg, PtrVT);
ReplaceUses(SDValue(Node, 0), ResNode);
return std::make_pair(true, ResNode.getNode());
}
case ISD::BUILD_VECTOR: {
// Select appropriate ldi.[bhwd] instructions for constant splats of
// 128-bit when MSA is enabled. Fixup any register class mismatches that
// occur as a result.
//
// This allows the compiler to use a wider range of immediates than would
// otherwise be allowed. If, for example, v4i32 could only use ldi.h then
// it would not be possible to load { 0x01010101, 0x01010101, 0x01010101,
// 0x01010101 } without using a constant pool. This would be sub-optimal
// when // 'ldi.b wd, 1' is capable of producing that bit-pattern in the
// same set/ of registers. Similarly, ldi.h isn't capable of producing {
// 0x00000000, 0x00000001, 0x00000000, 0x00000001 } but 'ldi.d wd, 1' can.
BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Node);
APInt SplatValue, SplatUndef;
unsigned SplatBitSize;
bool HasAnyUndefs;
unsigned LdiOp;
EVT ResVecTy = BVN->getValueType(0);
EVT ViaVecTy;
if (!Subtarget.hasMSA() || !BVN->getValueType(0).is128BitVector())
return std::make_pair(false, (SDNode*)NULL);
if (!BVN->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
HasAnyUndefs, 8,
!Subtarget.isLittle()))
return std::make_pair(false, (SDNode*)NULL);
switch (SplatBitSize) {
default:
return std::make_pair(false, (SDNode*)NULL);
case 8:
LdiOp = Mips::LDI_B;
ViaVecTy = MVT::v16i8;
break;
case 16:
LdiOp = Mips::LDI_H;
ViaVecTy = MVT::v8i16;
break;
case 32:
LdiOp = Mips::LDI_W;
ViaVecTy = MVT::v4i32;
break;
case 64:
LdiOp = Mips::LDI_D;
ViaVecTy = MVT::v2i64;
break;
}
if (!SplatValue.isSignedIntN(10))
return std::make_pair(false, (SDNode*)NULL);
SDValue Imm = CurDAG->getTargetConstant(SplatValue,
ViaVecTy.getVectorElementType());
SDNode *Res = CurDAG->getMachineNode(LdiOp, SDLoc(Node), ViaVecTy, Imm);
if (ResVecTy != ViaVecTy) {
// If LdiOp is writing to a different register class to ResVecTy, then
// fix it up here. This COPY_TO_REGCLASS should never cause a move.v
// since the source and destination register sets contain the same
// registers.
const TargetLowering *TLI = getTargetLowering();
MVT ResVecTySimple = ResVecTy.getSimpleVT();
const TargetRegisterClass *RC = TLI->getRegClassFor(ResVecTySimple);
Res = CurDAG->getMachineNode(Mips::COPY_TO_REGCLASS, SDLoc(Node),
ResVecTy, SDValue(Res, 0),
CurDAG->getTargetConstant(RC->getID(),
MVT::i32));
}
return std::make_pair(true, Res);
}
}
return std::make_pair(false, (SDNode*)NULL);
}
std::pair<bool, SDNode*> MipsSEDAGToDAGISel::selectNode(SDNode *Node) {
unsigned Opcode = Node->getOpcode();
SDLoc DL(Node);
///
// Instruction Selection not handled by the auto-generated
// tablegen selection should be handled here.
///
SDNode *Result;
switch(Opcode) {
default: break;
case ISD::SUBE: {
SDValue InFlag = Node->getOperand(2);
Result = selectAddESubE(Mips::SUBu, InFlag, InFlag.getOperand(0), DL, Node);
return std::make_pair(true, Result);
}
case ISD::ADDE: {
if (Subtarget.hasDSP()) // Select DSP instructions, ADDSC and ADDWC.
break;
SDValue InFlag = Node->getOperand(2);
Result = selectAddESubE(Mips::ADDu, InFlag, InFlag.getValue(0), DL, Node);
return std::make_pair(true, Result);
}
case ISD::ConstantFP: {
ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(Node);
if (Node->getValueType(0) == MVT::f64 && CN->isExactlyValue(+0.0)) {
if (Subtarget.hasMips64()) {
SDValue Zero = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), DL,
Mips::ZERO_64, MVT::i64);
Result = CurDAG->getMachineNode(Mips::DMTC1, DL, MVT::f64, Zero);
} else {
SDValue Zero = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), DL,
Mips::ZERO, MVT::i32);
Result = CurDAG->getMachineNode(Mips::BuildPairF64, DL, MVT::f64, Zero,
Zero);
}
return std::make_pair(true, Result);
}
break;
}
case ISD::Constant: {
const ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Node);
unsigned Size = CN->getValueSizeInBits(0);
if (Size == 32)
break;
MipsAnalyzeImmediate AnalyzeImm;
int64_t Imm = CN->getSExtValue();
const MipsAnalyzeImmediate::InstSeq &Seq =
AnalyzeImm.Analyze(Imm, Size, false);
MipsAnalyzeImmediate::InstSeq::const_iterator Inst = Seq.begin();
SDLoc DL(CN);
SDNode *RegOpnd;
SDValue ImmOpnd = CurDAG->getTargetConstant(SignExtend64<16>(Inst->ImmOpnd),
MVT::i64);
// The first instruction can be a LUi which is different from other
// instructions (ADDiu, ORI and SLL) in that it does not have a register
// operand.
if (Inst->Opc == Mips::LUi64)
RegOpnd = CurDAG->getMachineNode(Inst->Opc, DL, MVT::i64, ImmOpnd);
else
RegOpnd =
CurDAG->getMachineNode(Inst->Opc, DL, MVT::i64,
CurDAG->getRegister(Mips::ZERO_64, MVT::i64),
ImmOpnd);
// The remaining instructions in the sequence are handled here.
for (++Inst; Inst != Seq.end(); ++Inst) {
ImmOpnd = CurDAG->getTargetConstant(SignExtend64<16>(Inst->ImmOpnd),
MVT::i64);
RegOpnd = CurDAG->getMachineNode(Inst->Opc, DL, MVT::i64,
SDValue(RegOpnd, 0), ImmOpnd);
}
return std::make_pair(true, RegOpnd);
}
case MipsISD::ThreadPointer: {
EVT PtrVT = getTargetLowering()->getPointerTy();
unsigned RdhwrOpc, SrcReg, DestReg;
if (PtrVT == MVT::i32) {
RdhwrOpc = Mips::RDHWR;
SrcReg = Mips::HWR29;
DestReg = Mips::V1;
} else {
RdhwrOpc = Mips::RDHWR64;
SrcReg = Mips::HWR29_64;
DestReg = Mips::V1_64;
}
SDNode *Rdhwr =
CurDAG->getMachineNode(RdhwrOpc, SDLoc(Node),
Node->getValueType(0),
CurDAG->getRegister(SrcReg, PtrVT));
SDValue Chain = CurDAG->getCopyToReg(CurDAG->getEntryNode(), DL, DestReg,
SDValue(Rdhwr, 0));
SDValue ResNode = CurDAG->getCopyFromReg(Chain, DL, DestReg, PtrVT);
ReplaceUses(SDValue(Node, 0), ResNode);
return std::make_pair(true, ResNode.getNode());
}
case MipsISD::InsertLOHI: {
unsigned RCID = Subtarget.hasDSP() ? Mips::ACRegsDSPRegClassID :
Mips::ACRegsRegClassID;
SDValue RegClass = CurDAG->getTargetConstant(RCID, MVT::i32);
SDValue LoIdx = CurDAG->getTargetConstant(Mips::sub_lo, MVT::i32);
SDValue HiIdx = CurDAG->getTargetConstant(Mips::sub_hi, MVT::i32);
const SDValue Ops[] = { RegClass, Node->getOperand(0), LoIdx,
Node->getOperand(1), HiIdx };
SDNode *Res = CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, DL,
MVT::Untyped, Ops);
return std::make_pair(true, Res);
}
}
return std::make_pair(false, (SDNode*)NULL);
}
std::pair<bool, SDNode*> MipsSEDAGToDAGISel::selectNode(SDNode *Node) {
unsigned Opcode = Node->getOpcode();
DebugLoc DL = Node->getDebugLoc();
///
// Instruction Selection not handled by the auto-generated
// tablegen selection should be handled here.
///
EVT NodeTy = Node->getValueType(0);
SDNode *Result;
unsigned MultOpc;
switch(Opcode) {
default: break;
case ISD::SUBE: {
SDValue InFlag = Node->getOperand(2);
Result = selectAddESubE(Mips::SUBu, InFlag, InFlag.getOperand(0), DL, Node);
return std::make_pair(true, Result);
}
case ISD::ADDE: {
SDValue InFlag = Node->getOperand(2);
Result = selectAddESubE(Mips::ADDu, InFlag, InFlag.getValue(0), DL, Node);
return std::make_pair(true, Result);
}
/// Mul with two results
case ISD::SMUL_LOHI:
case ISD::UMUL_LOHI: {
if (NodeTy == MVT::i32)
MultOpc = (Opcode == ISD::UMUL_LOHI ? Mips::MULTu : Mips::MULT);
else
MultOpc = (Opcode == ISD::UMUL_LOHI ? Mips::DMULTu : Mips::DMULT);
std::pair<SDNode*, SDNode*> LoHi = selectMULT(Node, MultOpc, DL, NodeTy,
true, true);
if (!SDValue(Node, 0).use_empty())
ReplaceUses(SDValue(Node, 0), SDValue(LoHi.first, 0));
if (!SDValue(Node, 1).use_empty())
ReplaceUses(SDValue(Node, 1), SDValue(LoHi.second, 0));
return std::make_pair(true, (SDNode*)NULL);
}
/// Special Muls
case ISD::MUL: {
// Mips32 has a 32-bit three operand mul instruction.
if (Subtarget.hasMips32() && NodeTy == MVT::i32)
break;
MultOpc = NodeTy == MVT::i32 ? Mips::MULT : Mips::DMULT;
Result = selectMULT(Node, MultOpc, DL, NodeTy, true, false).first;
return std::make_pair(true, Result);
}
case ISD::MULHS:
case ISD::MULHU: {
if (NodeTy == MVT::i32)
MultOpc = (Opcode == ISD::MULHU ? Mips::MULTu : Mips::MULT);
else
MultOpc = (Opcode == ISD::MULHU ? Mips::DMULTu : Mips::DMULT);
Result = selectMULT(Node, MultOpc, DL, NodeTy, false, true).second;
return std::make_pair(true, Result);
}
case ISD::ConstantFP: {
ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(Node);
if (Node->getValueType(0) == MVT::f64 && CN->isExactlyValue(+0.0)) {
if (Subtarget.hasMips64()) {
SDValue Zero = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), DL,
Mips::ZERO_64, MVT::i64);
Result = CurDAG->getMachineNode(Mips::DMTC1, DL, MVT::f64, Zero);
} else {
SDValue Zero = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), DL,
Mips::ZERO, MVT::i32);
Result = CurDAG->getMachineNode(Mips::BuildPairF64, DL, MVT::f64, Zero,
Zero);
}
return std::make_pair(true, Result);
}
break;
}
case ISD::Constant: {
const ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Node);
unsigned Size = CN->getValueSizeInBits(0);
if (Size == 32)
break;
MipsAnalyzeImmediate AnalyzeImm;
int64_t Imm = CN->getSExtValue();
const MipsAnalyzeImmediate::InstSeq &Seq =
AnalyzeImm.Analyze(Imm, Size, false);
MipsAnalyzeImmediate::InstSeq::const_iterator Inst = Seq.begin();
DebugLoc DL = CN->getDebugLoc();
SDNode *RegOpnd;
SDValue ImmOpnd = CurDAG->getTargetConstant(SignExtend64<16>(Inst->ImmOpnd),
MVT::i64);
// The first instruction can be a LUi which is different from other
// instructions (ADDiu, ORI and SLL) in that it does not have a register
// operand.
if (Inst->Opc == Mips::LUi64)
RegOpnd = CurDAG->getMachineNode(Inst->Opc, DL, MVT::i64, ImmOpnd);
else
RegOpnd =
CurDAG->getMachineNode(Inst->Opc, DL, MVT::i64,
CurDAG->getRegister(Mips::ZERO_64, MVT::i64),
ImmOpnd);
// The remaining instructions in the sequence are handled here.
for (++Inst; Inst != Seq.end(); ++Inst) {
ImmOpnd = CurDAG->getTargetConstant(SignExtend64<16>(Inst->ImmOpnd),
MVT::i64);
RegOpnd = CurDAG->getMachineNode(Inst->Opc, DL, MVT::i64,
SDValue(RegOpnd, 0), ImmOpnd);
}
return std::make_pair(true, RegOpnd);
}
case MipsISD::ThreadPointer: {
EVT PtrVT = TLI.getPointerTy();
unsigned RdhwrOpc, SrcReg, DestReg;
if (PtrVT == MVT::i32) {
RdhwrOpc = Mips::RDHWR;
SrcReg = Mips::HWR29;
DestReg = Mips::V1;
} else {
RdhwrOpc = Mips::RDHWR64;
SrcReg = Mips::HWR29_64;
DestReg = Mips::V1_64;
}
SDNode *Rdhwr =
CurDAG->getMachineNode(RdhwrOpc, Node->getDebugLoc(),
Node->getValueType(0),
CurDAG->getRegister(SrcReg, PtrVT));
SDValue Chain = CurDAG->getCopyToReg(CurDAG->getEntryNode(), DL, DestReg,
SDValue(Rdhwr, 0));
SDValue ResNode = CurDAG->getCopyFromReg(Chain, DL, DestReg, PtrVT);
ReplaceUses(SDValue(Node, 0), ResNode);
return std::make_pair(true, ResNode.getNode());
}
}
return std::make_pair(false, (SDNode*)NULL);
}