// Set up the register classes
addRegisterClass(MVT::i32, &Mips::GPR32RegClass);
- if (isGP64bit())
+ if (Subtarget->isGP64bit())
addRegisterClass(MVT::i64, &Mips::GPR64RegClass);
if (Subtarget->hasDSP() || Subtarget->hasMSA()) {
setTargetDAGCombine(ISD::XOR);
}
- if (!Subtarget->mipsSEUsesSoftFloat()) {
+ if (!Subtarget->abiUsesSoftFloat()) {
addRegisterClass(MVT::f32, &Mips::FGR32RegClass);
// When dealing with single precision only, use libcalls
if (Subtarget->hasCnMips())
setOperationAction(ISD::MUL, MVT::i64, Legal);
- else if (isGP64bit())
+ else if (Subtarget->isGP64bit())
setOperationAction(ISD::MUL, MVT::i64, Custom);
- if (isGP64bit()) {
+ if (Subtarget->isGP64bit()) {
setOperationAction(ISD::MULHS, MVT::i64, Custom);
setOperationAction(ISD::MULHU, MVT::i64, Custom);
}
setOperationAction(ISD::STORE, MVT::f64, Custom);
}
+ if (Subtarget->hasMips32r6()) {
+ // MIPS32r6 replaces the accumulator-based multiplies with a three register
+ // instruction
+ setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
+ setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
+ setOperationAction(ISD::MUL, MVT::i32, Legal);
+ setOperationAction(ISD::MULHS, MVT::i32, Legal);
+ setOperationAction(ISD::MULHU, MVT::i32, Legal);
+
+ // MIPS32r6 replaces the accumulator-based division/remainder with separate
+ // three register division and remainder instructions.
+ setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
+ setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
+ setOperationAction(ISD::SDIV, MVT::i32, Legal);
+ setOperationAction(ISD::UDIV, MVT::i32, Legal);
+ setOperationAction(ISD::SREM, MVT::i32, Legal);
+ setOperationAction(ISD::UREM, MVT::i32, Legal);
+
+ // MIPS32r6 replaces conditional moves with an equivalent that removes the
+ // need for three GPR read ports.
+ setOperationAction(ISD::SETCC, MVT::i32, Legal);
+ setOperationAction(ISD::SELECT, MVT::i32, Legal);
+ setOperationAction(ISD::SELECT_CC, MVT::i32, Expand);
+
+ setOperationAction(ISD::SETCC, MVT::f32, Legal);
+ setOperationAction(ISD::SELECT, MVT::f32, Legal);
+ setOperationAction(ISD::SELECT_CC, MVT::f32, Expand);
+
+ assert(Subtarget->isFP64bit() && "FR=1 is required for MIPS32r6");
+ setOperationAction(ISD::SETCC, MVT::f64, Legal);
+ setOperationAction(ISD::SELECT, MVT::f64, Legal);
+ setOperationAction(ISD::SELECT_CC, MVT::f64, Expand);
+
+ setOperationAction(ISD::BRCOND, MVT::Other, Legal);
+
+ // Floating point > and >= are supported via < and <=
+ setCondCodeAction(ISD::SETOGE, MVT::f32, Expand);
+ setCondCodeAction(ISD::SETOGT, MVT::f32, Expand);
+ setCondCodeAction(ISD::SETUGE, MVT::f32, Expand);
+ setCondCodeAction(ISD::SETUGT, MVT::f32, Expand);
+
+ setCondCodeAction(ISD::SETOGE, MVT::f64, Expand);
+ setCondCodeAction(ISD::SETOGT, MVT::f64, Expand);
+ setCondCodeAction(ISD::SETUGE, MVT::f64, Expand);
+ setCondCodeAction(ISD::SETUGT, MVT::f64, Expand);
+ }
+
+ if (Subtarget->hasMips64r6()) {
+ // MIPS64r6 replaces the accumulator-based multiplies with a three register
+ // instruction
+ setOperationAction(ISD::MUL, MVT::i64, Legal);
+ setOperationAction(ISD::MULHS, MVT::i64, Legal);
+ setOperationAction(ISD::MULHU, MVT::i64, Legal);
+
+ // MIPS32r6 replaces the accumulator-based division/remainder with separate
+ // three register division and remainder instructions.
+ setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
+ setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
+ setOperationAction(ISD::SDIV, MVT::i64, Legal);
+ setOperationAction(ISD::UDIV, MVT::i64, Legal);
+ setOperationAction(ISD::SREM, MVT::i64, Legal);
+ setOperationAction(ISD::UREM, MVT::i64, Legal);
+
+ // MIPS64r6 replaces conditional moves with an equivalent that removes the
+ // need for three GPR read ports.
+ setOperationAction(ISD::SETCC, MVT::i64, Legal);
+ setOperationAction(ISD::SELECT, MVT::i64, Legal);
+ setOperationAction(ISD::SELECT_CC, MVT::i64, Expand);
+ }
+
computeRegisterProperties();
}
return new MipsSETargetLowering(TM);
}
+const TargetRegisterClass *
+MipsSETargetLowering::getRepRegClassFor(MVT VT) const {
+ if (VT == MVT::Untyped)
+ return Subtarget->hasDSP() ? &Mips::ACC64DSPRegClass : &Mips::ACC64RegClass;
+
+ return TargetLowering::getRepRegClassFor(VT);
+}
+
// Enable MSA support for the given integer type and Register class.
void MipsSETargetLowering::
addMSAIntType(MVT::SimpleValueType Ty, const TargetRegisterClass *RC) {
bool *Fast) const {
MVT::SimpleValueType SVT = VT.getSimpleVT().SimpleTy;
+ if (Subtarget->systemSupportsUnalignedAccess()) {
+ // MIPS32r6/MIPS64r6 is required to support unaligned access. It's
+ // implementation defined whether this is handled by hardware, software, or
+ // a hybrid of the two but it's expected that most implementations will
+ // handle the majority of cases in hardware.
+ if (Fast)
+ *Fast = true;
+ return true;
+ }
+
switch (SVT) {
case MVT::i64:
case MVT::i32:
if (DCI.isBeforeLegalize())
return SDValue();
- if (Subtarget->hasMips32() && N->getValueType(0) == MVT::i32 &&
- selectMADD(N, &DAG))
+ if (Subtarget->hasMips32() && !Subtarget->hasMips32r6() &&
+ N->getValueType(0) == MVT::i32 && selectMADD(N, &DAG))
return SDValue(N, 0);
return SDValue();
Log2 == ExtendTySize) {
SDValue Ops[] = { Op0->getOperand(0), Op0->getOperand(1), Op0Op2 };
DAG.MorphNodeTo(Op0.getNode(), MipsISD::VEXTRACT_ZEXT_ELT,
- Op0->getVTList(), Ops, Op0->getNumOperands());
+ Op0->getVTList(),
+ makeArrayRef(Ops, Op0->getNumOperands()));
return Op0;
}
}
SDValue Ops[] = { Op0Op0->getOperand(0), Op0Op0->getOperand(1),
Op0Op0->getOperand(2) };
DAG.MorphNodeTo(Op0Op0.getNode(), MipsISD::VEXTRACT_SEXT_ELT,
- Op0Op0->getVTList(), Ops, Op0Op0->getNumOperands());
+ Op0Op0->getVTList(),
+ makeArrayRef(Ops, Op0Op0->getNumOperands()));
return Op0Op0;
}
}
return emitINSERT_FW(MI, BB);
case Mips::INSERT_FD_PSEUDO:
return emitINSERT_FD(MI, BB);
+ case Mips::INSERT_B_VIDX_PSEUDO:
+ return emitINSERT_DF_VIDX(MI, BB, 1, false);
+ case Mips::INSERT_H_VIDX_PSEUDO:
+ return emitINSERT_DF_VIDX(MI, BB, 2, false);
+ case Mips::INSERT_W_VIDX_PSEUDO:
+ return emitINSERT_DF_VIDX(MI, BB, 4, false);
+ case Mips::INSERT_D_VIDX_PSEUDO:
+ return emitINSERT_DF_VIDX(MI, BB, 8, false);
+ case Mips::INSERT_FW_VIDX_PSEUDO:
+ return emitINSERT_DF_VIDX(MI, BB, 4, true);
+ case Mips::INSERT_FD_VIDX_PSEUDO:
+ return emitINSERT_DF_VIDX(MI, BB, 8, true);
case Mips::FILL_FW_PSEUDO:
return emitFILL_FW(MI, BB);
case Mips::FILL_FD_PSEUDO:
SDValue BP = DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64, Lo, Hi);
SDValue Ops[2] = {BP, Hi.getValue(1)};
- return DAG.getMergeValues(Ops, 2, DL);
+ return DAG.getMergeValues(Ops, DL);
}
SDValue MipsSETargetLowering::lowerSTORE(SDValue Op, SelectionDAG &DAG) const {
SDValue MipsSETargetLowering::lowerMulDiv(SDValue Op, unsigned NewOpc,
bool HasLo, bool HasHi,
SelectionDAG &DAG) const {
+ // MIPS32r6/MIPS64r6 removed accumulator based multiplies.
+ assert(!Subtarget->hasMips32r6());
+
EVT Ty = Op.getOperand(0).getValueType();
SDLoc DL(Op);
SDValue Mult = DAG.getNode(NewOpc, DL, MVT::Untyped,
return HasLo ? Lo : Hi;
SDValue Vals[] = { Lo, Hi };
- return DAG.getMergeValues(Vals, 2, DL);
+ return DAG.getMergeValues(Vals, DL);
}
ResTys.push_back((*I == MVT::i64) ? MVT::Untyped : *I);
// Create node.
- SDValue Val = DAG.getNode(Opc, DL, ResTys, &Ops[0], Ops.size());
+ SDValue Val = DAG.getNode(Opc, DL, ResTys, Ops);
SDValue Out = (ResTys[0] == MVT::Untyped) ? extractLOHI(Val, DL, DAG) : Val;
if (!HasChainIn)
assert(Val->getValueType(1) == MVT::Other);
SDValue Vals[] = { Out, SDValue(Val.getNode(), 1) };
- return DAG.getMergeValues(Vals, 2, DL);
+ return DAG.getMergeValues(Vals, DL);
}
// Lower an MSA copy intrinsic into the specified SelectionDAG node
SDValue Ops[16] = { LaneA, LaneB, LaneA, LaneB, LaneA, LaneB, LaneA, LaneB,
LaneA, LaneB, LaneA, LaneB, LaneA, LaneB, LaneA, LaneB };
- SDValue Result = DAG.getNode(ISD::BUILD_VECTOR, DL, ViaVecTy, Ops,
- ViaVecTy.getVectorNumElements());
+ SDValue Result = DAG.getNode(ISD::BUILD_VECTOR, DL, ViaVecTy,
+ makeArrayRef(Ops, ViaVecTy.getVectorNumElements()));
if (ViaVecTy != ResVecTy)
Result = DAG.getNode(ISD::BITCAST, DL, ResVecTy, Result);
SplatValueA, SplatValueB, SplatValueA, SplatValueB,
SplatValueA, SplatValueB, SplatValueA, SplatValueB };
- SDValue Result = DAG.getNode(ISD::BUILD_VECTOR, DL, ViaVecTy, Ops,
- ViaVecTy.getVectorNumElements());
+ SDValue Result = DAG.getNode(ISD::BUILD_VECTOR, DL, ViaVecTy,
+ makeArrayRef(Ops, ViaVecTy.getVectorNumElements()));
if (VecTy != ViaVecTy)
Result = DAG.getNode(ISD::BITCAST, DL, VecTy, Result);
case Intrinsic::mips_copy_s_w:
return lowerMSACopyIntr(Op, DAG, MipsISD::VEXTRACT_SEXT_ELT);
case Intrinsic::mips_copy_s_d:
- if (hasMips64())
+ if (Subtarget->hasMips64())
// Lower directly into VEXTRACT_SEXT_ELT since i64 is legal on Mips64.
return lowerMSACopyIntr(Op, DAG, MipsISD::VEXTRACT_SEXT_ELT);
else {
case Intrinsic::mips_copy_u_w:
return lowerMSACopyIntr(Op, DAG, MipsISD::VEXTRACT_ZEXT_ELT);
case Intrinsic::mips_copy_u_d:
- if (hasMips64())
+ if (Subtarget->hasMips64())
// Lower directly into VEXTRACT_ZEXT_ELT since i64 is legal on Mips64.
return lowerMSACopyIntr(Op, DAG, MipsISD::VEXTRACT_ZEXT_ELT);
else {
// If ResTy is v2i64 then the type legalizer will break this node down into
// an equivalent v4i32.
- return DAG.getNode(ISD::BUILD_VECTOR, DL, ResTy, &Ops[0], Ops.size());
+ return DAG.getNode(ISD::BUILD_VECTOR, DL, ResTy, Ops);
}
case Intrinsic::mips_fexp2_w:
case Intrinsic::mips_fexp2_d: {
++I)
Ops.push_back(DAG.getTargetConstant(*I, MaskEltTy));
- SDValue MaskVec = DAG.getNode(ISD::BUILD_VECTOR, DL, MaskVecTy, &Ops[0],
- Ops.size());
+ SDValue MaskVec = DAG.getNode(ISD::BUILD_VECTOR, DL, MaskVecTy, Ops);
if (Using1stVec && Using2ndVec) {
Op0 = Op->getOperand(0);
return BB;
}
+// Emit the INSERT_([BHWD]|F[WD])_VIDX pseudo instruction.
+//
+// For integer:
+// (INSERT_([BHWD]|F[WD])_PSEUDO $wd, $wd_in, $n, $rs)
+// =>
+// (SLL $lanetmp1, $lane, <log2size)
+// (SLD_B $wdtmp1, $wd_in, $wd_in, $lanetmp1)
+// (INSERT_[BHWD], $wdtmp2, $wdtmp1, 0, $rs)
+// (NEG $lanetmp2, $lanetmp1)
+// (SLD_B $wd, $wdtmp2, $wdtmp2, $lanetmp2)
+//
+// For floating point:
+// (INSERT_([BHWD]|F[WD])_PSEUDO $wd, $wd_in, $n, $fs)
+// =>
+// (SUBREG_TO_REG $wt, $fs, <subreg>)
+// (SLL $lanetmp1, $lane, <log2size)
+// (SLD_B $wdtmp1, $wd_in, $wd_in, $lanetmp1)
+// (INSVE_[WD], $wdtmp2, 0, $wdtmp1, 0)
+// (NEG $lanetmp2, $lanetmp1)
+// (SLD_B $wd, $wdtmp2, $wdtmp2, $lanetmp2)
+MachineBasicBlock *
+MipsSETargetLowering::emitINSERT_DF_VIDX(MachineInstr *MI,
+ MachineBasicBlock *BB,
+ unsigned EltSizeInBytes,
+ bool IsFP) const {
+ const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
+ MachineRegisterInfo &RegInfo = BB->getParent()->getRegInfo();
+ DebugLoc DL = MI->getDebugLoc();
+ unsigned Wd = MI->getOperand(0).getReg();
+ unsigned SrcVecReg = MI->getOperand(1).getReg();
+ unsigned LaneReg = MI->getOperand(2).getReg();
+ unsigned SrcValReg = MI->getOperand(3).getReg();
+
+ const TargetRegisterClass *VecRC = nullptr;
+ const TargetRegisterClass *GPRRC =
+ Subtarget->isGP64bit() ? &Mips::GPR64RegClass : &Mips::GPR32RegClass;
+ unsigned EltLog2Size;
+ unsigned InsertOp = 0;
+ unsigned InsveOp = 0;
+ switch (EltSizeInBytes) {
+ default:
+ llvm_unreachable("Unexpected size");
+ case 1:
+ EltLog2Size = 0;
+ InsertOp = Mips::INSERT_B;
+ InsveOp = Mips::INSVE_B;
+ VecRC = &Mips::MSA128BRegClass;
+ break;
+ case 2:
+ EltLog2Size = 1;
+ InsertOp = Mips::INSERT_H;
+ InsveOp = Mips::INSVE_H;
+ VecRC = &Mips::MSA128HRegClass;
+ break;
+ case 4:
+ EltLog2Size = 2;
+ InsertOp = Mips::INSERT_W;
+ InsveOp = Mips::INSVE_W;
+ VecRC = &Mips::MSA128WRegClass;
+ break;
+ case 8:
+ EltLog2Size = 3;
+ InsertOp = Mips::INSERT_D;
+ InsveOp = Mips::INSVE_D;
+ VecRC = &Mips::MSA128DRegClass;
+ break;
+ }
+
+ if (IsFP) {
+ unsigned Wt = RegInfo.createVirtualRegister(VecRC);
+ BuildMI(*BB, MI, DL, TII->get(Mips::SUBREG_TO_REG), Wt)
+ .addImm(0)
+ .addReg(SrcValReg)
+ .addImm(EltSizeInBytes == 8 ? Mips::sub_64 : Mips::sub_lo);
+ SrcValReg = Wt;
+ }
+
+ // Convert the lane index into a byte index
+ if (EltSizeInBytes != 1) {
+ unsigned LaneTmp1 = RegInfo.createVirtualRegister(GPRRC);
+ BuildMI(*BB, MI, DL, TII->get(Mips::SLL), LaneTmp1)
+ .addReg(LaneReg)
+ .addImm(EltLog2Size);
+ LaneReg = LaneTmp1;
+ }
+
+ // Rotate bytes around so that the desired lane is element zero
+ unsigned WdTmp1 = RegInfo.createVirtualRegister(VecRC);
+ BuildMI(*BB, MI, DL, TII->get(Mips::SLD_B), WdTmp1)
+ .addReg(SrcVecReg)
+ .addReg(SrcVecReg)
+ .addReg(LaneReg);
+
+ unsigned WdTmp2 = RegInfo.createVirtualRegister(VecRC);
+ if (IsFP) {
+ // Use insve.df to insert to element zero
+ BuildMI(*BB, MI, DL, TII->get(InsveOp), WdTmp2)
+ .addReg(WdTmp1)
+ .addImm(0)
+ .addReg(SrcValReg)
+ .addImm(0);
+ } else {
+ // Use insert.df to insert to element zero
+ BuildMI(*BB, MI, DL, TII->get(InsertOp), WdTmp2)
+ .addReg(WdTmp1)
+ .addReg(SrcValReg)
+ .addImm(0);
+ }
+
+ // Rotate elements the rest of the way for a full rotation.
+ // sld.df inteprets $rt modulo the number of columns so we only need to negate
+ // the lane index to do this.
+ unsigned LaneTmp2 = RegInfo.createVirtualRegister(GPRRC);
+ BuildMI(*BB, MI, DL, TII->get(Mips::SUB), LaneTmp2)
+ .addReg(Mips::ZERO)
+ .addReg(LaneReg);
+ BuildMI(*BB, MI, DL, TII->get(Mips::SLD_B), Wd)
+ .addReg(WdTmp2)
+ .addReg(WdTmp2)
+ .addReg(LaneTmp2);
+
+ MI->eraseFromParent(); // The pseudo instruction is gone now.
+ return BB;
+}
+
// Emit the FILL_FW pseudo instruction.
//
// fill_fw_pseudo $wd, $fs
projects / oota-llvm.git / blobdiff