addMSAFloatType(MVT::v4f32, &Mips::MSA128WRegClass);
addMSAFloatType(MVT::v2f64, &Mips::MSA128DRegClass);
+ setTargetDAGCombine(ISD::AND);
+ setTargetDAGCombine(ISD::SRA);
+ setTargetDAGCombine(ISD::VSELECT);
setTargetDAGCombine(ISD::XOR);
}
setOperationAction(ISD::BITCAST, Ty, Legal);
setOperationAction(ISD::LOAD, Ty, Legal);
setOperationAction(ISD::STORE, Ty, Legal);
+ setOperationAction(ISD::EXTRACT_VECTOR_ELT, Ty, Custom);
+ setOperationAction(ISD::INSERT_VECTOR_ELT, Ty, Legal);
setOperationAction(ISD::BUILD_VECTOR, Ty, Custom);
setOperationAction(ISD::ADD, Ty, Legal);
setOperationAction(ISD::AND, Ty, Legal);
setOperationAction(ISD::CTLZ, Ty, Legal);
+ setOperationAction(ISD::CTPOP, Ty, Legal);
setOperationAction(ISD::MUL, Ty, Legal);
setOperationAction(ISD::OR, Ty, Legal);
setOperationAction(ISD::SDIV, Ty, Legal);
setOperationAction(ISD::SRL, Ty, Legal);
setOperationAction(ISD::SUB, Ty, Legal);
setOperationAction(ISD::UDIV, Ty, Legal);
+ setOperationAction(ISD::VSELECT, Ty, Legal);
setOperationAction(ISD::XOR, Ty, Legal);
+
+ setOperationAction(ISD::SETCC, Ty, Legal);
+ setCondCodeAction(ISD::SETNE, Ty, Expand);
+ setCondCodeAction(ISD::SETGE, Ty, Expand);
+ setCondCodeAction(ISD::SETGT, Ty, Expand);
+ setCondCodeAction(ISD::SETUGE, Ty, Expand);
+ setCondCodeAction(ISD::SETUGT, Ty, Expand);
}
// Enable MSA support for the given floating-point type and Register class.
setOperationAction(ISD::LOAD, Ty, Legal);
setOperationAction(ISD::STORE, Ty, Legal);
setOperationAction(ISD::BITCAST, Ty, Legal);
+ setOperationAction(ISD::EXTRACT_VECTOR_ELT, Ty, Legal);
if (Ty != MVT::v8f16) {
+ setOperationAction(ISD::FABS, Ty, Legal);
setOperationAction(ISD::FADD, Ty, Legal);
setOperationAction(ISD::FDIV, Ty, Legal);
setOperationAction(ISD::FLOG2, Ty, Legal);
setOperationAction(ISD::FRINT, Ty, Legal);
setOperationAction(ISD::FSQRT, Ty, Legal);
setOperationAction(ISD::FSUB, Ty, Legal);
+ setOperationAction(ISD::VSELECT, Ty, Legal);
+
+ setOperationAction(ISD::SETCC, Ty, Legal);
+ setCondCodeAction(ISD::SETOGE, Ty, Expand);
+ setCondCodeAction(ISD::SETOGT, Ty, Expand);
+ setCondCodeAction(ISD::SETUGE, Ty, Expand);
+ setCondCodeAction(ISD::SETUGT, Ty, Expand);
+ setCondCodeAction(ISD::SETGE, Ty, Expand);
+ setCondCodeAction(ISD::SETGT, Ty, Expand);
}
}
case ISD::INTRINSIC_WO_CHAIN: return lowerINTRINSIC_WO_CHAIN(Op, DAG);
case ISD::INTRINSIC_W_CHAIN: return lowerINTRINSIC_W_CHAIN(Op, DAG);
case ISD::INTRINSIC_VOID: return lowerINTRINSIC_VOID(Op, DAG);
+ case ISD::EXTRACT_VECTOR_ELT: return lowerEXTRACT_VECTOR_ELT(Op, DAG);
case ISD::BUILD_VECTOR: return lowerBUILD_VECTOR(Op, DAG);
}
return SDValue();
}
+// Fold zero extensions into MipsISD::VEXTRACT_[SZ]EXT_ELT
+//
+// Performs the following transformations:
+// - Changes MipsISD::VEXTRACT_[SZ]EXT_ELT to zero extension if its
+// sign/zero-extension is completely overwritten by the new one performed by
+// the ISD::AND.
+// - Removes redundant zero extensions performed by an ISD::AND.
+static SDValue performANDCombine(SDNode *N, SelectionDAG &DAG,
+ TargetLowering::DAGCombinerInfo &DCI,
+ const MipsSubtarget *Subtarget) {
+ if (!Subtarget->hasMSA())
+ return SDValue();
+
+ SDValue Op0 = N->getOperand(0);
+ SDValue Op1 = N->getOperand(1);
+ unsigned Op0Opcode = Op0->getOpcode();
+
+ // (and (MipsVExtract[SZ]Ext $a, $b, $c), imm:$d)
+ // where $d + 1 == 2^n and n == 32
+ // or $d + 1 == 2^n and n <= 32 and ZExt
+ // -> (MipsVExtractZExt $a, $b, $c)
+ if (Op0Opcode == MipsISD::VEXTRACT_SEXT_ELT ||
+ Op0Opcode == MipsISD::VEXTRACT_ZEXT_ELT) {
+ ConstantSDNode *Mask = dyn_cast<ConstantSDNode>(Op1);
+
+ if (!Mask)
+ return SDValue();
+
+ int32_t Log2IfPositive = (Mask->getAPIntValue() + 1).exactLogBase2();
+
+ if (Log2IfPositive <= 0)
+ return SDValue(); // Mask+1 is not a power of 2
+
+ SDValue Op0Op2 = Op0->getOperand(2);
+ EVT ExtendTy = cast<VTSDNode>(Op0Op2)->getVT();
+ unsigned ExtendTySize = ExtendTy.getSizeInBits();
+ unsigned Log2 = Log2IfPositive;
+
+ if ((Op0Opcode == MipsISD::VEXTRACT_ZEXT_ELT && Log2 >= ExtendTySize) ||
+ Log2 == ExtendTySize) {
+ SDValue Ops[] = { Op0->getOperand(0), Op0->getOperand(1), Op0Op2 };
+ DAG.MorphNodeTo(Op0.getNode(), MipsISD::VEXTRACT_ZEXT_ELT,
+ Op0->getVTList(), Ops, Op0->getNumOperands());
+ return Op0;
+ }
+ }
+
+ return SDValue();
+}
+
static SDValue performSUBECombine(SDNode *N, SelectionDAG &DAG,
TargetLowering::DAGCombinerInfo &DCI,
const MipsSubtarget *Subtarget) {
return performDSPShiftCombine(MipsISD::SHLL_DSP, N, Ty, DAG, Subtarget);
}
+// Fold sign-extensions into MipsISD::VEXTRACT_[SZ]EXT_ELT for MSA and fold
+// constant splats into MipsISD::SHRA_DSP for DSPr2.
+//
+// Performs the following transformations:
+// - Changes MipsISD::VEXTRACT_[SZ]EXT_ELT to sign extension if its
+// sign/zero-extension is completely overwritten by the new one performed by
+// the ISD::SRA and ISD::SHL nodes.
+// - Removes redundant sign extensions performed by an ISD::SRA and ISD::SHL
+// sequence.
+//
+// See performDSPShiftCombine for more information about the transformation
+// used for DSPr2.
static SDValue performSRACombine(SDNode *N, SelectionDAG &DAG,
TargetLowering::DAGCombinerInfo &DCI,
const MipsSubtarget *Subtarget) {
EVT Ty = N->getValueType(0);
+ if (Subtarget->hasMSA()) {
+ SDValue Op0 = N->getOperand(0);
+ SDValue Op1 = N->getOperand(1);
+
+ // (sra (shl (MipsVExtract[SZ]Ext $a, $b, $c), imm:$d), imm:$d)
+ // where $d + sizeof($c) == 32
+ // or $d + sizeof($c) <= 32 and SExt
+ // -> (MipsVExtractSExt $a, $b, $c)
+ if (Op0->getOpcode() == ISD::SHL && Op1 == Op0->getOperand(1)) {
+ SDValue Op0Op0 = Op0->getOperand(0);
+ ConstantSDNode *ShAmount = dyn_cast<ConstantSDNode>(Op1);
+
+ if (!ShAmount)
+ return SDValue();
+
+ EVT ExtendTy = cast<VTSDNode>(Op0Op0->getOperand(2))->getVT();
+ unsigned TotalBits = ShAmount->getZExtValue() + ExtendTy.getSizeInBits();
+
+ if (TotalBits == 32 ||
+ (Op0Op0->getOpcode() == MipsISD::VEXTRACT_SEXT_ELT &&
+ TotalBits <= 32)) {
+ SDValue Ops[] = { Op0Op0->getOperand(0), Op0Op0->getOperand(1),
+ Op0Op0->getOperand(2) };
+ DAG.MorphNodeTo(Op0Op0.getNode(), MipsISD::VEXTRACT_SEXT_ELT,
+ Op0Op0->getVTList(), Ops, Op0Op0->getNumOperands());
+ return Op0Op0;
+ }
+ }
+ }
+
if ((Ty != MVT::v2i16) && ((Ty != MVT::v4i8) || !Subtarget->hasDSPR2()))
return SDValue();
static SDValue performVSELECTCombine(SDNode *N, SelectionDAG &DAG) {
EVT Ty = N->getValueType(0);
- if ((Ty != MVT::v2i16) && (Ty != MVT::v4i8))
- return SDValue();
+ if (Ty.is128BitVector() && Ty.isInteger()) {
+ // Try the following combines:
+ // (vselect (setcc $a, $b, SETLT), $b, $a)) -> (vsmax $a, $b)
+ // (vselect (setcc $a, $b, SETLE), $b, $a)) -> (vsmax $a, $b)
+ // (vselect (setcc $a, $b, SETLT), $a, $b)) -> (vsmin $a, $b)
+ // (vselect (setcc $a, $b, SETLE), $a, $b)) -> (vsmin $a, $b)
+ // (vselect (setcc $a, $b, SETULT), $b, $a)) -> (vumax $a, $b)
+ // (vselect (setcc $a, $b, SETULE), $b, $a)) -> (vumax $a, $b)
+ // (vselect (setcc $a, $b, SETULT), $a, $b)) -> (vumin $a, $b)
+ // (vselect (setcc $a, $b, SETULE), $a, $b)) -> (vumin $a, $b)
+ // SETGT/SETGE/SETUGT/SETUGE variants of these will show up initially but
+ // will be expanded to equivalent SETLT/SETLE/SETULT/SETULE versions by the
+ // legalizer.
+ SDValue Op0 = N->getOperand(0);
- SDValue SetCC = N->getOperand(0);
+ if (Op0->getOpcode() != ISD::SETCC)
+ return SDValue();
- if (SetCC.getOpcode() != MipsISD::SETCC_DSP)
- return SDValue();
+ ISD::CondCode CondCode = cast<CondCodeSDNode>(Op0->getOperand(2))->get();
+ bool Signed;
- return DAG.getNode(MipsISD::SELECT_CC_DSP, SDLoc(N), Ty,
- SetCC.getOperand(0), SetCC.getOperand(1), N->getOperand(1),
- N->getOperand(2), SetCC.getOperand(2));
+ if (CondCode == ISD::SETLT || CondCode == ISD::SETLE)
+ Signed = true;
+ else if (CondCode == ISD::SETULT || CondCode == ISD::SETULE)
+ Signed = false;
+ else
+ return SDValue();
+
+ SDValue Op1 = N->getOperand(1);
+ SDValue Op2 = N->getOperand(2);
+ SDValue Op0Op0 = Op0->getOperand(0);
+ SDValue Op0Op1 = Op0->getOperand(1);
+
+ if (Op1 == Op0Op0 && Op2 == Op0Op1)
+ return DAG.getNode(Signed ? MipsISD::VSMIN : MipsISD::VUMIN, SDLoc(N),
+ Ty, Op1, Op2);
+ else if (Op1 == Op0Op1 && Op2 == Op0Op0)
+ return DAG.getNode(Signed ? MipsISD::VSMAX : MipsISD::VUMAX, SDLoc(N),
+ Ty, Op1, Op2);
+ } else if ((Ty == MVT::v2i16) || (Ty == MVT::v4i8)) {
+ SDValue SetCC = N->getOperand(0);
+
+ if (SetCC.getOpcode() != MipsISD::SETCC_DSP)
+ return SDValue();
+
+ return DAG.getNode(MipsISD::SELECT_CC_DSP, SDLoc(N), Ty,
+ SetCC.getOperand(0), SetCC.getOperand(1),
+ N->getOperand(1), N->getOperand(2), SetCC.getOperand(2));
+ }
+
+ return SDValue();
}
static SDValue performXORCombine(SDNode *N, SelectionDAG &DAG,
SDValue Op0 = N->getOperand(0);
SDValue Op1 = N->getOperand(1);
SDValue NotOp;
- ConstantSDNode *Const;
if (ISD::isBuildVectorAllOnes(Op0.getNode()))
NotOp = Op1;
else if (ISD::isBuildVectorAllOnes(Op1.getNode()))
NotOp = Op0;
- else if ((Op0->getOpcode() == MipsISD::VSPLAT ||
- Op0->getOpcode() == MipsISD::VSPLATD) &&
- (Const = dyn_cast<ConstantSDNode>(Op0->getOperand(0))) &&
- Const->isAllOnesValue())
- NotOp = Op1;
- else if ((Op1->getOpcode() == MipsISD::VSPLAT ||
- Op1->getOpcode() == MipsISD::VSPLATD) &&
- (Const = dyn_cast<ConstantSDNode>(Op1->getOperand(0))) &&
- Const->isAllOnesValue())
- NotOp = Op0;
else
return SDValue();
switch (N->getOpcode()) {
case ISD::ADDE:
return performADDECombine(N, DAG, DCI, Subtarget);
+ case ISD::AND:
+ Val = performANDCombine(N, DAG, DCI, Subtarget);
+ break;
case ISD::SUBE:
return performSUBECombine(N, DAG, DCI, Subtarget);
case ISD::MUL:
return Result;
}
+static SDValue lowerMSABinaryImmIntr(SDValue Op, SelectionDAG &DAG,
+ unsigned Opc, SDValue RHS) {
+ SDValue LHS = Op->getOperand(1);
+ EVT ResTy = Op->getValueType(0);
+
+ return DAG.getNode(Opc, SDLoc(Op), ResTy, LHS, RHS);
+}
+
static SDValue lowerMSABranchIntr(SDValue Op, SelectionDAG &DAG, unsigned Opc) {
SDLoc DL(Op);
SDValue Value = Op->getOperand(1);
return Result;
}
+// Lower an MSA copy intrinsic into the specified SelectionDAG node
+static SDValue lowerMSACopyIntr(SDValue Op, SelectionDAG &DAG, unsigned Opc) {
+ SDLoc DL(Op);
+ SDValue Vec = Op->getOperand(1);
+ SDValue Idx = Op->getOperand(2);
+ EVT ResTy = Op->getValueType(0);
+ EVT EltTy = Vec->getValueType(0).getVectorElementType();
+
+ SDValue Result = DAG.getNode(Opc, DL, ResTy, Vec, Idx,
+ DAG.getValueType(EltTy));
+
+ return Result;
+}
+
+// Lower an MSA insert intrinsic into the specified SelectionDAG node
+static SDValue lowerMSAInsertIntr(SDValue Op, SelectionDAG &DAG, unsigned Opc) {
+ SDLoc DL(Op);
+ SDValue Op0 = Op->getOperand(1);
+ SDValue Op1 = Op->getOperand(2);
+ SDValue Op2 = Op->getOperand(3);
+ EVT ResTy = Op->getValueType(0);
+
+ SDValue Result = DAG.getNode(Opc, DL, ResTy, Op0, Op2, Op1);
+
+ return Result;
+}
+
+static SDValue lowerMSASplatImm(SDValue Op, SDValue ImmOp, SelectionDAG &DAG) {
+ EVT ResTy = Op->getValueType(0);
+ EVT ViaVecTy = ResTy;
+ SmallVector<SDValue, 16> Ops;
+ SDValue ImmHiOp;
+ SDLoc DL(Op);
+
+ if (ViaVecTy == MVT::v2i64) {
+ ImmHiOp = DAG.getNode(ISD::SRA, DL, MVT::i32, ImmOp,
+ DAG.getConstant(31, MVT::i32));
+ for (unsigned i = 0; i < ViaVecTy.getVectorNumElements(); ++i) {
+ Ops.push_back(ImmHiOp);
+ Ops.push_back(ImmOp);
+ }
+ ViaVecTy = MVT::v4i32;
+ } else {
+ for (unsigned i = 0; i < ResTy.getVectorNumElements(); ++i)
+ Ops.push_back(ImmOp);
+ }
+
+ SDValue Result = DAG.getNode(ISD::BUILD_VECTOR, DL, ViaVecTy, &Ops[0],
+ Ops.size());
+
+ if (ResTy != ViaVecTy)
+ Result = DAG.getNode(ISD::BITCAST, DL, ResTy, Result);
+
+ return Result;
+}
+
+static SDValue
+lowerMSASplatImm(SDValue Op, unsigned ImmOp, SelectionDAG &DAG) {
+ return lowerMSASplatImm(Op, Op->getOperand(ImmOp), DAG);
+}
+
static SDValue lowerMSAUnaryIntr(SDValue Op, SelectionDAG &DAG, unsigned Opc) {
SDLoc DL(Op);
SDValue Value = Op->getOperand(1);
case Intrinsic::mips_addv_w:
case Intrinsic::mips_addv_d:
return lowerMSABinaryIntr(Op, DAG, ISD::ADD);
+ case Intrinsic::mips_addvi_b:
+ case Intrinsic::mips_addvi_h:
+ case Intrinsic::mips_addvi_w:
+ case Intrinsic::mips_addvi_d:
+ return lowerMSABinaryImmIntr(Op, DAG, ISD::ADD,
+ lowerMSASplatImm(Op, 2, DAG));
case Intrinsic::mips_and_v:
return lowerMSABinaryIntr(Op, DAG, ISD::AND);
+ case Intrinsic::mips_andi_b:
+ return lowerMSABinaryImmIntr(Op, DAG, ISD::AND,
+ lowerMSASplatImm(Op, 2, DAG));
case Intrinsic::mips_bnz_b:
case Intrinsic::mips_bnz_h:
case Intrinsic::mips_bnz_w:
return lowerMSABranchIntr(Op, DAG, MipsISD::VALL_NONZERO);
case Intrinsic::mips_bnz_v:
return lowerMSABranchIntr(Op, DAG, MipsISD::VANY_NONZERO);
+ case Intrinsic::mips_bsel_v:
+ return DAG.getNode(ISD::VSELECT, SDLoc(Op), Op->getValueType(0),
+ Op->getOperand(1), Op->getOperand(2),
+ Op->getOperand(3));
+ case Intrinsic::mips_bseli_b:
+ return DAG.getNode(ISD::VSELECT, SDLoc(Op), Op->getValueType(0),
+ Op->getOperand(1), Op->getOperand(2),
+ lowerMSASplatImm(Op, 3, DAG));
case Intrinsic::mips_bz_b:
case Intrinsic::mips_bz_h:
case Intrinsic::mips_bz_w:
return lowerMSABranchIntr(Op, DAG, MipsISD::VALL_ZERO);
case Intrinsic::mips_bz_v:
return lowerMSABranchIntr(Op, DAG, MipsISD::VANY_ZERO);
+ case Intrinsic::mips_ceq_b:
+ case Intrinsic::mips_ceq_h:
+ case Intrinsic::mips_ceq_w:
+ case Intrinsic::mips_ceq_d:
+ return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
+ Op->getOperand(2), ISD::SETEQ);
+ case Intrinsic::mips_ceqi_b:
+ case Intrinsic::mips_ceqi_h:
+ case Intrinsic::mips_ceqi_w:
+ case Intrinsic::mips_ceqi_d:
+ return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
+ lowerMSASplatImm(Op, 2, DAG), ISD::SETEQ);
+ case Intrinsic::mips_cle_s_b:
+ case Intrinsic::mips_cle_s_h:
+ case Intrinsic::mips_cle_s_w:
+ case Intrinsic::mips_cle_s_d:
+ return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
+ Op->getOperand(2), ISD::SETLE);
+ case Intrinsic::mips_clei_s_b:
+ case Intrinsic::mips_clei_s_h:
+ case Intrinsic::mips_clei_s_w:
+ case Intrinsic::mips_clei_s_d:
+ return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
+ lowerMSASplatImm(Op, 2, DAG), ISD::SETLE);
+ case Intrinsic::mips_cle_u_b:
+ case Intrinsic::mips_cle_u_h:
+ case Intrinsic::mips_cle_u_w:
+ case Intrinsic::mips_cle_u_d:
+ return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
+ Op->getOperand(2), ISD::SETULE);
+ case Intrinsic::mips_clei_u_b:
+ case Intrinsic::mips_clei_u_h:
+ case Intrinsic::mips_clei_u_w:
+ case Intrinsic::mips_clei_u_d:
+ return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
+ lowerMSASplatImm(Op, 2, DAG), ISD::SETULE);
+ case Intrinsic::mips_clt_s_b:
+ case Intrinsic::mips_clt_s_h:
+ case Intrinsic::mips_clt_s_w:
+ case Intrinsic::mips_clt_s_d:
+ return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
+ Op->getOperand(2), ISD::SETLT);
+ case Intrinsic::mips_clti_s_b:
+ case Intrinsic::mips_clti_s_h:
+ case Intrinsic::mips_clti_s_w:
+ case Intrinsic::mips_clti_s_d:
+ return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
+ lowerMSASplatImm(Op, 2, DAG), ISD::SETLT);
+ case Intrinsic::mips_clt_u_b:
+ case Intrinsic::mips_clt_u_h:
+ case Intrinsic::mips_clt_u_w:
+ case Intrinsic::mips_clt_u_d:
+ return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
+ Op->getOperand(2), ISD::SETULT);
+ case Intrinsic::mips_clti_u_b:
+ case Intrinsic::mips_clti_u_h:
+ case Intrinsic::mips_clti_u_w:
+ case Intrinsic::mips_clti_u_d:
+ return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
+ lowerMSASplatImm(Op, 2, DAG), ISD::SETULT);
+ case Intrinsic::mips_copy_s_b:
+ case Intrinsic::mips_copy_s_h:
+ case Intrinsic::mips_copy_s_w:
+ return lowerMSACopyIntr(Op, DAG, MipsISD::VEXTRACT_SEXT_ELT);
+ case Intrinsic::mips_copy_u_b:
+ case Intrinsic::mips_copy_u_h:
+ case Intrinsic::mips_copy_u_w:
+ return lowerMSACopyIntr(Op, DAG, MipsISD::VEXTRACT_ZEXT_ELT);
case Intrinsic::mips_div_s_b:
case Intrinsic::mips_div_s_h:
case Intrinsic::mips_div_s_w:
case Intrinsic::mips_fadd_w:
case Intrinsic::mips_fadd_d:
return lowerMSABinaryIntr(Op, DAG, ISD::FADD);
+ // Don't lower mips_fcaf_[wd] since LLVM folds SETFALSE condcodes away
+ case Intrinsic::mips_fceq_w:
+ case Intrinsic::mips_fceq_d:
+ return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
+ Op->getOperand(2), ISD::SETOEQ);
+ case Intrinsic::mips_fcle_w:
+ case Intrinsic::mips_fcle_d:
+ return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
+ Op->getOperand(2), ISD::SETOLE);
+ case Intrinsic::mips_fclt_w:
+ case Intrinsic::mips_fclt_d:
+ return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
+ Op->getOperand(2), ISD::SETOLT);
+ case Intrinsic::mips_fcne_w:
+ case Intrinsic::mips_fcne_d:
+ return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
+ Op->getOperand(2), ISD::SETONE);
+ case Intrinsic::mips_fcor_w:
+ case Intrinsic::mips_fcor_d:
+ return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
+ Op->getOperand(2), ISD::SETO);
+ case Intrinsic::mips_fcueq_w:
+ case Intrinsic::mips_fcueq_d:
+ return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
+ Op->getOperand(2), ISD::SETUEQ);
+ case Intrinsic::mips_fcule_w:
+ case Intrinsic::mips_fcule_d:
+ return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
+ Op->getOperand(2), ISD::SETULE);
+ case Intrinsic::mips_fcult_w:
+ case Intrinsic::mips_fcult_d:
+ return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
+ Op->getOperand(2), ISD::SETULT);
+ case Intrinsic::mips_fcun_w:
+ case Intrinsic::mips_fcun_d:
+ return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
+ Op->getOperand(2), ISD::SETUO);
+ case Intrinsic::mips_fcune_w:
+ case Intrinsic::mips_fcune_d:
+ return DAG.getSetCC(SDLoc(Op), Op->getValueType(0), Op->getOperand(1),
+ Op->getOperand(2), ISD::SETUNE);
case Intrinsic::mips_fdiv_w:
case Intrinsic::mips_fdiv_d:
return lowerMSABinaryIntr(Op, DAG, ISD::FDIV);
case Intrinsic::mips_fill_b:
case Intrinsic::mips_fill_h:
- case Intrinsic::mips_fill_w:
- return lowerMSAUnaryIntr(Op, DAG, MipsISD::VSPLAT);
+ case Intrinsic::mips_fill_w: {
+ SmallVector<SDValue, 16> Ops;
+ EVT ResTy = Op->getValueType(0);
+
+ for (unsigned i = 0; i < ResTy.getVectorNumElements(); ++i)
+ Ops.push_back(Op->getOperand(1));
+
+ return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), ResTy, &Ops[0],
+ Ops.size());
+ }
case Intrinsic::mips_flog2_w:
case Intrinsic::mips_flog2_d:
return lowerMSAUnaryIntr(Op, DAG, ISD::FLOG2);
case Intrinsic::mips_fsub_w:
case Intrinsic::mips_fsub_d:
return lowerMSABinaryIntr(Op, DAG, ISD::FSUB);
+ case Intrinsic::mips_insert_b:
+ case Intrinsic::mips_insert_h:
+ case Intrinsic::mips_insert_w:
+ return lowerMSAInsertIntr(Op, DAG, ISD::INSERT_VECTOR_ELT);
case Intrinsic::mips_ldi_b:
case Intrinsic::mips_ldi_h:
case Intrinsic::mips_ldi_w:
case Intrinsic::mips_ldi_d:
- return lowerMSAUnaryIntr(Op, DAG, MipsISD::VSPLAT);
+ return lowerMSASplatImm(Op, 1, DAG);
+ case Intrinsic::mips_max_s_b:
+ case Intrinsic::mips_max_s_h:
+ case Intrinsic::mips_max_s_w:
+ case Intrinsic::mips_max_s_d:
+ return lowerMSABinaryIntr(Op, DAG, MipsISD::VSMAX);
+ case Intrinsic::mips_max_u_b:
+ case Intrinsic::mips_max_u_h:
+ case Intrinsic::mips_max_u_w:
+ case Intrinsic::mips_max_u_d:
+ return lowerMSABinaryIntr(Op, DAG, MipsISD::VUMAX);
+ case Intrinsic::mips_maxi_s_b:
+ case Intrinsic::mips_maxi_s_h:
+ case Intrinsic::mips_maxi_s_w:
+ case Intrinsic::mips_maxi_s_d:
+ return lowerMSABinaryImmIntr(Op, DAG, MipsISD::VSMAX,
+ lowerMSASplatImm(Op, 2, DAG));
+ case Intrinsic::mips_maxi_u_b:
+ case Intrinsic::mips_maxi_u_h:
+ case Intrinsic::mips_maxi_u_w:
+ case Intrinsic::mips_maxi_u_d:
+ return lowerMSABinaryImmIntr(Op, DAG, MipsISD::VUMAX,
+ lowerMSASplatImm(Op, 2, DAG));
+ case Intrinsic::mips_min_s_b:
+ case Intrinsic::mips_min_s_h:
+ case Intrinsic::mips_min_s_w:
+ case Intrinsic::mips_min_s_d:
+ return lowerMSABinaryIntr(Op, DAG, MipsISD::VSMIN);
+ case Intrinsic::mips_min_u_b:
+ case Intrinsic::mips_min_u_h:
+ case Intrinsic::mips_min_u_w:
+ case Intrinsic::mips_min_u_d:
+ return lowerMSABinaryIntr(Op, DAG, MipsISD::VUMIN);
+ case Intrinsic::mips_mini_s_b:
+ case Intrinsic::mips_mini_s_h:
+ case Intrinsic::mips_mini_s_w:
+ case Intrinsic::mips_mini_s_d:
+ return lowerMSABinaryImmIntr(Op, DAG, MipsISD::VSMIN,
+ lowerMSASplatImm(Op, 2, DAG));
+ case Intrinsic::mips_mini_u_b:
+ case Intrinsic::mips_mini_u_h:
+ case Intrinsic::mips_mini_u_w:
+ case Intrinsic::mips_mini_u_d:
+ return lowerMSABinaryImmIntr(Op, DAG, MipsISD::VUMIN,
+ lowerMSASplatImm(Op, 2, DAG));
case Intrinsic::mips_mulv_b:
case Intrinsic::mips_mulv_h:
case Intrinsic::mips_mulv_w:
SDValue Res = lowerMSABinaryIntr(Op, DAG, ISD::OR);
return DAG.getNOT(SDLoc(Op), Res, Res->getValueType(0));
}
+ case Intrinsic::mips_nori_b: {
+ SDValue Res = lowerMSABinaryImmIntr(Op, DAG, ISD::OR,
+ lowerMSASplatImm(Op, 2, DAG));
+ return DAG.getNOT(SDLoc(Op), Res, Res->getValueType(0));
+ }
case Intrinsic::mips_or_v:
return lowerMSABinaryIntr(Op, DAG, ISD::OR);
+ case Intrinsic::mips_ori_b:
+ return lowerMSABinaryImmIntr(Op, DAG, ISD::OR,
+ lowerMSASplatImm(Op, 2, DAG));
+ case Intrinsic::mips_pcnt_b:
+ case Intrinsic::mips_pcnt_h:
+ case Intrinsic::mips_pcnt_w:
+ case Intrinsic::mips_pcnt_d:
+ return lowerMSAUnaryIntr(Op, DAG, ISD::CTPOP);
case Intrinsic::mips_sll_b:
case Intrinsic::mips_sll_h:
case Intrinsic::mips_sll_w:
case Intrinsic::mips_sll_d:
return lowerMSABinaryIntr(Op, DAG, ISD::SHL);
+ case Intrinsic::mips_slli_b:
+ case Intrinsic::mips_slli_h:
+ case Intrinsic::mips_slli_w:
+ case Intrinsic::mips_slli_d:
+ return lowerMSABinaryImmIntr(Op, DAG, ISD::SHL,
+ lowerMSASplatImm(Op, 2, DAG));
case Intrinsic::mips_sra_b:
case Intrinsic::mips_sra_h:
case Intrinsic::mips_sra_w:
case Intrinsic::mips_sra_d:
return lowerMSABinaryIntr(Op, DAG, ISD::SRA);
+ case Intrinsic::mips_srai_b:
+ case Intrinsic::mips_srai_h:
+ case Intrinsic::mips_srai_w:
+ case Intrinsic::mips_srai_d:
+ return lowerMSABinaryImmIntr(Op, DAG, ISD::SRA,
+ lowerMSASplatImm(Op, 2, DAG));
case Intrinsic::mips_srl_b:
case Intrinsic::mips_srl_h:
case Intrinsic::mips_srl_w:
case Intrinsic::mips_srl_d:
return lowerMSABinaryIntr(Op, DAG, ISD::SRL);
+ case Intrinsic::mips_srli_b:
+ case Intrinsic::mips_srli_h:
+ case Intrinsic::mips_srli_w:
+ case Intrinsic::mips_srli_d:
+ return lowerMSABinaryImmIntr(Op, DAG, ISD::SRL,
+ lowerMSASplatImm(Op, 2, DAG));
case Intrinsic::mips_subv_b:
case Intrinsic::mips_subv_h:
case Intrinsic::mips_subv_w:
case Intrinsic::mips_subv_d:
return lowerMSABinaryIntr(Op, DAG, ISD::SUB);
+ case Intrinsic::mips_subvi_b:
+ case Intrinsic::mips_subvi_h:
+ case Intrinsic::mips_subvi_w:
+ case Intrinsic::mips_subvi_d:
+ return lowerMSABinaryImmIntr(Op, DAG, ISD::SUB,
+ lowerMSASplatImm(Op, 2, DAG));
case Intrinsic::mips_xor_v:
return lowerMSABinaryIntr(Op, DAG, ISD::XOR);
+ case Intrinsic::mips_xori_b:
+ return lowerMSABinaryImmIntr(Op, DAG, ISD::XOR,
+ lowerMSASplatImm(Op, 2, DAG));
}
}
return true;
}
+// Lower ISD::EXTRACT_VECTOR_ELT into MipsISD::VEXTRACT_SEXT_ELT.
+//
+// The non-value bits resulting from ISD::EXTRACT_VECTOR_ELT are undefined. We
+// choose to sign-extend but we could have equally chosen zero-extend. The
+// DAGCombiner will fold any sign/zero extension of the ISD::EXTRACT_VECTOR_ELT
+// result into this node later (possibly changing it to a zero-extend in the
+// process).
+SDValue MipsSETargetLowering::
+lowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const {
+ SDLoc DL(Op);
+ EVT ResTy = Op->getValueType(0);
+ SDValue Op0 = Op->getOperand(0);
+ SDValue Op1 = Op->getOperand(1);
+ EVT EltTy = Op0->getValueType(0).getVectorElementType();
+ return DAG.getNode(MipsISD::VEXTRACT_SEXT_ELT, DL, ResTy, Op0, Op1,
+ DAG.getValueType(EltTy));
+}
+
+static bool isConstantOrUndef(const SDValue Op) {
+ if (Op->getOpcode() == ISD::UNDEF)
+ return true;
+ if (dyn_cast<ConstantSDNode>(Op))
+ return true;
+ if (dyn_cast<ConstantFPSDNode>(Op))
+ return true;
+ return false;
+}
+
+static bool isConstantOrUndefBUILD_VECTOR(const BuildVectorSDNode *Op) {
+ for (unsigned i = 0; i < Op->getNumOperands(); ++i)
+ if (isConstantOrUndef(Op->getOperand(i)))
+ return true;
+ return false;
+}
+
// Lowers ISD::BUILD_VECTOR into appropriate SelectionDAG nodes for the
// backend.
//
// Lowers according to the following rules:
-// - Vectors of 128-bits may be legal subject to the other rules. Other sizes
-// are not legal.
-// - Non-constant splats are legal and are lowered to MipsISD::VSPLAT.
-// - Constant splats with an element size of 32-bits or less are legal and are
-// lowered to MipsISD::VSPLAT.
-// - Constant splats with an element size of 64-bits but whose value would fit
-// within a 10 bit immediate are legal and are lowered to MipsISD::VSPLATD.
-// - All other ISD::BUILD_VECTORS are not legal
+// - Constant splats are legal as-is as long as the SplatBitSize is a power of
+// 2 less than or equal to 64 and the value fits into a signed 10-bit
+// immediate
+// - Constant splats are lowered to bitconverted BUILD_VECTORs if SplatBitSize
+// is a power of 2 less than or equal to 64 and the value does not fit into a
+// signed 10-bit immediate
+// - Non-constant splats are legal as-is.
+// - Non-constant non-splats are lowered to sequences of INSERT_VECTOR_ELT.
+// - All others are illegal and must be expanded.
SDValue MipsSETargetLowering::lowerBUILD_VECTOR(SDValue Op,
SelectionDAG &DAG) const {
BuildVectorSDNode *Node = cast<BuildVectorSDNode>(Op);
if (Node->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
HasAnyUndefs, 8,
- !Subtarget->isLittle())) {
- SDValue Result;
- EVT TmpVecTy;
- EVT ConstTy = MVT::i32;
- unsigned SplatOp = MipsISD::VSPLAT;
+ !Subtarget->isLittle()) && SplatBitSize <= 64) {
+ // We can only cope with 8, 16, 32, or 64-bit elements
+ if (SplatBitSize != 8 && SplatBitSize != 16 && SplatBitSize != 32 &&
+ SplatBitSize != 64)
+ return SDValue();
+
+ // If the value fits into a simm10 then we can use ldi.[bhwd]
+ if (SplatValue.isSignedIntN(10))
+ return Op;
+
+ EVT ViaVecTy;
switch (SplatBitSize) {
default:
return SDValue();
- case 64:
- TmpVecTy = MVT::v2i64;
-
- // i64 is an illegal type on Mips32, but if it the constant fits into a
- // signed 10-bit value then we can still handle it using VSPLATD and an
- // i32 constant
- if (HasMips64)
- ConstTy = MVT::i64;
- else if (isInt<10>(SplatValue.getSExtValue())) {
- SplatValue = SplatValue.trunc(32);
- SplatOp = MipsISD::VSPLATD;
- } else
- return SDValue();
- break;
- case 32:
- TmpVecTy = MVT::v4i32;
+ case 8:
+ ViaVecTy = MVT::v16i8;
break;
case 16:
- TmpVecTy = MVT::v8i16;
- SplatValue = SplatValue.sext(32);
+ ViaVecTy = MVT::v8i16;
break;
- case 8:
- TmpVecTy = MVT::v16i8;
- SplatValue = SplatValue.sext(32);
+ case 32:
+ ViaVecTy = MVT::v4i32;
break;
+ case 64:
+ // There's no fill.d to fall back on for 64-bit values
+ return SDValue();
}
- Result = DAG.getNode(SplatOp, DL, TmpVecTy,
- DAG.getConstant(SplatValue, ConstTy));
- if (ResTy != Result.getValueType())
- Result = DAG.getNode(ISD::BITCAST, DL, ResTy, Result);
+ SmallVector<SDValue, 16> Ops;
+ SDValue Constant = DAG.getConstant(SplatValue.sextOrSelf(32), MVT::i32);
+
+ for (unsigned i = 0; i < ViaVecTy.getVectorNumElements(); ++i)
+ Ops.push_back(Constant);
+
+ SDValue Result = DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Node), ViaVecTy,
+ &Ops[0], Ops.size());
+
+ if (ViaVecTy != ResTy)
+ Result = DAG.getNode(ISD::BITCAST, SDLoc(Node), ResTy, Result);
return Result;
+ } else if (isSplatVector(Node))
+ return Op;
+ else if (!isConstantOrUndefBUILD_VECTOR(Node)) {
+ // Use INSERT_VECTOR_ELT operations rather than expand to stores.
+ // The resulting code is the same length as the expansion, but it doesn't
+ // use memory operations
+ EVT ResTy = Node->getValueType(0);
+
+ assert(ResTy.isVector());
+
+ unsigned NumElts = ResTy.getVectorNumElements();
+ SDValue Vector = DAG.getUNDEF(ResTy);
+ for (unsigned i = 0; i < NumElts; ++i) {
+ Vector = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, ResTy, Vector,
+ Node->getOperand(i),
+ DAG.getConstant(i, MVT::i32));
+ }
+ return Vector;
}
- else if (isSplatVector(Node))
- return DAG.getNode(MipsISD::VSPLAT, DL, ResTy, Op->getOperand(0));
return SDValue();
}