addRegisterClass(MVT::v1i64, &AArch64::FPR64RegClass);
addRegisterClass(MVT::v1f32, &AArch64::FPR32RegClass);
addRegisterClass(MVT::v1f64, &AArch64::FPR64RegClass);
- addRegisterClass(MVT::v8i8, &AArch64::FPR64RegClass);
+ addRegisterClass(MVT::v8i8, &AArch64::FPR64RegClass);
addRegisterClass(MVT::v4i16, &AArch64::FPR64RegClass);
addRegisterClass(MVT::v2i32, &AArch64::FPR64RegClass);
addRegisterClass(MVT::v1i64, &AArch64::FPR64RegClass);
setOperationAction(ISD::SETCC, MVT::v4f32, Custom);
setOperationAction(ISD::SETCC, MVT::v1f64, Custom);
setOperationAction(ISD::SETCC, MVT::v2f64, Custom);
+
+ setOperationAction(ISD::FFLOOR, MVT::v2f32, Legal);
+ setOperationAction(ISD::FFLOOR, MVT::v4f32, Legal);
+ setOperationAction(ISD::FFLOOR, MVT::v1f64, Legal);
+ setOperationAction(ISD::FFLOOR, MVT::v2f64, Legal);
+
+ setOperationAction(ISD::FCEIL, MVT::v2f32, Legal);
+ setOperationAction(ISD::FCEIL, MVT::v4f32, Legal);
+ setOperationAction(ISD::FCEIL, MVT::v1f64, Legal);
+ setOperationAction(ISD::FCEIL, MVT::v2f64, Legal);
+
+ setOperationAction(ISD::FTRUNC, MVT::v2f32, Legal);
+ setOperationAction(ISD::FTRUNC, MVT::v4f32, Legal);
+ setOperationAction(ISD::FTRUNC, MVT::v1f64, Legal);
+ setOperationAction(ISD::FTRUNC, MVT::v2f64, Legal);
+
+ setOperationAction(ISD::FRINT, MVT::v2f32, Legal);
+ setOperationAction(ISD::FRINT, MVT::v4f32, Legal);
+ setOperationAction(ISD::FRINT, MVT::v1f64, Legal);
+ setOperationAction(ISD::FRINT, MVT::v2f64, Legal);
+
+ setOperationAction(ISD::FNEARBYINT, MVT::v2f32, Legal);
+ setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Legal);
+ setOperationAction(ISD::FNEARBYINT, MVT::v1f64, Legal);
+ setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Legal);
+
+ setOperationAction(ISD::FROUND, MVT::v2f32, Legal);
+ setOperationAction(ISD::FROUND, MVT::v4f32, Legal);
+ setOperationAction(ISD::FROUND, MVT::v1f64, Legal);
+ setOperationAction(ISD::FROUND, MVT::v2f64, Legal);
}
}
StrOpc = StoreOps[Log2_32(Size)];
}
+// FIXME: AArch64::DTripleRegClass and AArch64::QTripleRegClass don't really
+// have value type mapped, and they are both being defined as MVT::untyped.
+// Without knowing the MVT type, MachineLICM::getRegisterClassIDAndCost
+// would fail to figure out the register pressure correctly.
+std::pair<const TargetRegisterClass*, uint8_t>
+AArch64TargetLowering::findRepresentativeClass(MVT VT) const{
+ const TargetRegisterClass *RRC = 0;
+ uint8_t Cost = 1;
+ switch (VT.SimpleTy) {
+ default:
+ return TargetLowering::findRepresentativeClass(VT);
+ case MVT::v4i64:
+ RRC = &AArch64::QPairRegClass;
+ Cost = 2;
+ break;
+ case MVT::v8i64:
+ RRC = &AArch64::QQuadRegClass;
+ Cost = 4;
+ break;
+ }
+ return std::make_pair(RRC, Cost);
+}
+
MachineBasicBlock *
AArch64TargetLowering::emitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
unsigned Size,
return "AArch64ISD::NEON_VDUP";
case AArch64ISD::NEON_VDUPLANE:
return "AArch64ISD::NEON_VDUPLANE";
+ case AArch64ISD::NEON_REV16:
+ return "AArch64ISD::NEON_REV16";
+ case AArch64ISD::NEON_REV32:
+ return "AArch64ISD::NEON_REV32";
+ case AArch64ISD::NEON_REV64:
+ return "AArch64ISD::NEON_REV64";
+ case AArch64ISD::NEON_UZP1:
+ return "AArch64ISD::NEON_UZP1";
+ case AArch64ISD::NEON_UZP2:
+ return "AArch64ISD::NEON_UZP2";
+ case AArch64ISD::NEON_ZIP1:
+ return "AArch64ISD::NEON_ZIP1";
+ case AArch64ISD::NEON_ZIP2:
+ return "AArch64ISD::NEON_ZIP2";
+ case AArch64ISD::NEON_TRN1:
+ return "AArch64ISD::NEON_TRN1";
+ case AArch64ISD::NEON_TRN2:
+ return "AArch64ISD::NEON_TRN2";
case AArch64ISD::NEON_LD1_UPD:
return "AArch64ISD::NEON_LD1_UPD";
case AArch64ISD::NEON_LD2_UPD:
return "AArch64ISD::NEON_ST3_UPD";
case AArch64ISD::NEON_ST4_UPD:
return "AArch64ISD::NEON_ST4_UPD";
+ case AArch64ISD::NEON_LD1x2_UPD:
+ return "AArch64ISD::NEON_LD1x2_UPD";
+ case AArch64ISD::NEON_LD1x3_UPD:
+ return "AArch64ISD::NEON_LD1x3_UPD";
+ case AArch64ISD::NEON_LD1x4_UPD:
+ return "AArch64ISD::NEON_LD1x4_UPD";
+ case AArch64ISD::NEON_ST1x2_UPD:
+ return "AArch64ISD::NEON_ST1x2_UPD";
+ case AArch64ISD::NEON_ST1x3_UPD:
+ return "AArch64ISD::NEON_ST1x3_UPD";
+ case AArch64ISD::NEON_ST1x4_UPD:
+ return "AArch64ISD::NEON_ST1x4_UPD";
+ case AArch64ISD::NEON_LD2DUP:
+ return "AArch64ISD::NEON_LD2DUP";
+ case AArch64ISD::NEON_LD3DUP:
+ return "AArch64ISD::NEON_LD3DUP";
+ case AArch64ISD::NEON_LD4DUP:
+ return "AArch64ISD::NEON_LD4DUP";
+ case AArch64ISD::NEON_LD2DUP_UPD:
+ return "AArch64ISD::NEON_LD2DUP_UPD";
+ case AArch64ISD::NEON_LD3DUP_UPD:
+ return "AArch64ISD::NEON_LD3DUP_UPD";
+ case AArch64ISD::NEON_LD4DUP_UPD:
+ return "AArch64ISD::NEON_LD4DUP_UPD";
+ case AArch64ISD::NEON_LD2LN_UPD:
+ return "AArch64ISD::NEON_LD2LN_UPD";
+ case AArch64ISD::NEON_LD3LN_UPD:
+ return "AArch64ISD::NEON_LD3LN_UPD";
+ case AArch64ISD::NEON_LD4LN_UPD:
+ return "AArch64ISD::NEON_LD4LN_UPD";
+ case AArch64ISD::NEON_ST2LN_UPD:
+ return "AArch64ISD::NEON_ST2LN_UPD";
+ case AArch64ISD::NEON_ST3LN_UPD:
+ return "AArch64ISD::NEON_ST3LN_UPD";
+ case AArch64ISD::NEON_ST4LN_UPD:
+ return "AArch64ISD::NEON_ST4LN_UPD";
+ case AArch64ISD::NEON_VEXTRACT:
+ return "AArch64ISD::NEON_VEXTRACT";
default:
return NULL;
}
return SDValue();
SelectionDAG &DAG = DCI.DAG;
- unsigned AddrOpIdx = 2;
+ bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
+ N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
+ unsigned AddrOpIdx = (isIntrinsic ? 2 : 1);
SDValue Addr = N->getOperand(AddrOpIdx);
// Search for a use of the address operand that is an increment.
// Find the new opcode for the updating load/store.
bool isLoad = true;
+ bool isLaneOp = false;
unsigned NewOpc = 0;
unsigned NumVecs = 0;
- unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
- switch (IntNo) {
- default: llvm_unreachable("unexpected intrinsic for Neon base update");
- case Intrinsic::arm_neon_vld1: NewOpc = AArch64ISD::NEON_LD1_UPD;
- NumVecs = 1; break;
- case Intrinsic::arm_neon_vld2: NewOpc = AArch64ISD::NEON_LD2_UPD;
- NumVecs = 2; break;
- case Intrinsic::arm_neon_vld3: NewOpc = AArch64ISD::NEON_LD3_UPD;
- NumVecs = 3; break;
- case Intrinsic::arm_neon_vld4: NewOpc = AArch64ISD::NEON_LD4_UPD;
- NumVecs = 4; break;
- case Intrinsic::arm_neon_vst1: NewOpc = AArch64ISD::NEON_ST1_UPD;
- NumVecs = 1; isLoad = false; break;
- case Intrinsic::arm_neon_vst2: NewOpc = AArch64ISD::NEON_ST2_UPD;
- NumVecs = 2; isLoad = false; break;
- case Intrinsic::arm_neon_vst3: NewOpc = AArch64ISD::NEON_ST3_UPD;
- NumVecs = 3; isLoad = false; break;
- case Intrinsic::arm_neon_vst4: NewOpc = AArch64ISD::NEON_ST4_UPD;
- NumVecs = 4; isLoad = false; break;
+ if (isIntrinsic) {
+ unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
+ switch (IntNo) {
+ default: llvm_unreachable("unexpected intrinsic for Neon base update");
+ case Intrinsic::arm_neon_vld1: NewOpc = AArch64ISD::NEON_LD1_UPD;
+ NumVecs = 1; break;
+ case Intrinsic::arm_neon_vld2: NewOpc = AArch64ISD::NEON_LD2_UPD;
+ NumVecs = 2; break;
+ case Intrinsic::arm_neon_vld3: NewOpc = AArch64ISD::NEON_LD3_UPD;
+ NumVecs = 3; break;
+ case Intrinsic::arm_neon_vld4: NewOpc = AArch64ISD::NEON_LD4_UPD;
+ NumVecs = 4; break;
+ case Intrinsic::arm_neon_vst1: NewOpc = AArch64ISD::NEON_ST1_UPD;
+ NumVecs = 1; isLoad = false; break;
+ case Intrinsic::arm_neon_vst2: NewOpc = AArch64ISD::NEON_ST2_UPD;
+ NumVecs = 2; isLoad = false; break;
+ case Intrinsic::arm_neon_vst3: NewOpc = AArch64ISD::NEON_ST3_UPD;
+ NumVecs = 3; isLoad = false; break;
+ case Intrinsic::arm_neon_vst4: NewOpc = AArch64ISD::NEON_ST4_UPD;
+ NumVecs = 4; isLoad = false; break;
+ case Intrinsic::aarch64_neon_vld1x2: NewOpc = AArch64ISD::NEON_LD1x2_UPD;
+ NumVecs = 2; break;
+ case Intrinsic::aarch64_neon_vld1x3: NewOpc = AArch64ISD::NEON_LD1x3_UPD;
+ NumVecs = 3; break;
+ case Intrinsic::aarch64_neon_vld1x4: NewOpc = AArch64ISD::NEON_LD1x4_UPD;
+ NumVecs = 4; break;
+ case Intrinsic::aarch64_neon_vst1x2: NewOpc = AArch64ISD::NEON_ST1x2_UPD;
+ NumVecs = 2; isLoad = false; break;
+ case Intrinsic::aarch64_neon_vst1x3: NewOpc = AArch64ISD::NEON_ST1x3_UPD;
+ NumVecs = 3; isLoad = false; break;
+ case Intrinsic::aarch64_neon_vst1x4: NewOpc = AArch64ISD::NEON_ST1x4_UPD;
+ NumVecs = 4; isLoad = false; break;
+ case Intrinsic::arm_neon_vld2lane: NewOpc = AArch64ISD::NEON_LD2LN_UPD;
+ NumVecs = 2; isLaneOp = true; break;
+ case Intrinsic::arm_neon_vld3lane: NewOpc = AArch64ISD::NEON_LD3LN_UPD;
+ NumVecs = 3; isLaneOp = true; break;
+ case Intrinsic::arm_neon_vld4lane: NewOpc = AArch64ISD::NEON_LD4LN_UPD;
+ NumVecs = 4; isLaneOp = true; break;
+ case Intrinsic::arm_neon_vst2lane: NewOpc = AArch64ISD::NEON_ST2LN_UPD;
+ NumVecs = 2; isLoad = false; isLaneOp = true; break;
+ case Intrinsic::arm_neon_vst3lane: NewOpc = AArch64ISD::NEON_ST3LN_UPD;
+ NumVecs = 3; isLoad = false; isLaneOp = true; break;
+ case Intrinsic::arm_neon_vst4lane: NewOpc = AArch64ISD::NEON_ST4LN_UPD;
+ NumVecs = 4; isLoad = false; isLaneOp = true; break;
+ }
+ } else {
+ isLaneOp = true;
+ switch (N->getOpcode()) {
+ default: llvm_unreachable("unexpected opcode for Neon base update");
+ case AArch64ISD::NEON_LD2DUP: NewOpc = AArch64ISD::NEON_LD2DUP_UPD;
+ NumVecs = 2; break;
+ case AArch64ISD::NEON_LD3DUP: NewOpc = AArch64ISD::NEON_LD3DUP_UPD;
+ NumVecs = 3; break;
+ case AArch64ISD::NEON_LD4DUP: NewOpc = AArch64ISD::NEON_LD4DUP_UPD;
+ NumVecs = 4; break;
+ }
}
// Find the size of memory referenced by the load/store.
else
VecTy = N->getOperand(AddrOpIdx + 1).getValueType();
unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
+ if (isLaneOp)
+ NumBytes /= VecTy.getVectorNumElements();
// If the increment is a constant, it must match the memory ref size.
SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
return SDValue();
}
+/// For a VDUPLANE node N, check if its source operand is a vldN-lane (N > 1)
+/// intrinsic, and if all the other uses of that intrinsic are also VDUPLANEs.
+/// If so, combine them to a vldN-dup operation and return true.
+static SDValue CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
+ SelectionDAG &DAG = DCI.DAG;
+ EVT VT = N->getValueType(0);
+
+ // Check if the VDUPLANE operand is a vldN-dup intrinsic.
+ SDNode *VLD = N->getOperand(0).getNode();
+ if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
+ return SDValue();
+ unsigned NumVecs = 0;
+ unsigned NewOpc = 0;
+ unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
+ if (IntNo == Intrinsic::arm_neon_vld2lane) {
+ NumVecs = 2;
+ NewOpc = AArch64ISD::NEON_LD2DUP;
+ } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
+ NumVecs = 3;
+ NewOpc = AArch64ISD::NEON_LD3DUP;
+ } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
+ NumVecs = 4;
+ NewOpc = AArch64ISD::NEON_LD4DUP;
+ } else {
+ return SDValue();
+ }
+
+ // First check that all the vldN-lane uses are VDUPLANEs and that the lane
+ // numbers match the load.
+ unsigned VLDLaneNo =
+ cast<ConstantSDNode>(VLD->getOperand(NumVecs + 3))->getZExtValue();
+ for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
+ UI != UE; ++UI) {
+ // Ignore uses of the chain result.
+ if (UI.getUse().getResNo() == NumVecs)
+ continue;
+ SDNode *User = *UI;
+ if (User->getOpcode() != AArch64ISD::NEON_VDUPLANE ||
+ VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
+ return SDValue();
+ }
+
+ // Create the vldN-dup node.
+ EVT Tys[5];
+ unsigned n;
+ for (n = 0; n < NumVecs; ++n)
+ Tys[n] = VT;
+ Tys[n] = MVT::Other;
+ SDVTList SDTys = DAG.getVTList(Tys, NumVecs + 1);
+ SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
+ MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
+ SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, SDLoc(VLD), SDTys, Ops, 2,
+ VLDMemInt->getMemoryVT(),
+ VLDMemInt->getMemOperand());
+
+ // Update the uses.
+ for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
+ UI != UE; ++UI) {
+ unsigned ResNo = UI.getUse().getResNo();
+ // Ignore uses of the chain result.
+ if (ResNo == NumVecs)
+ continue;
+ SDNode *User = *UI;
+ DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
+ }
+
+ // Now the vldN-lane intrinsic is dead except for its chain result.
+ // Update uses of the chain.
+ std::vector<SDValue> VLDDupResults;
+ for (unsigned n = 0; n < NumVecs; ++n)
+ VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
+ VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
+ DCI.CombineTo(VLD, VLDDupResults);
+
+ return SDValue(N, 0);
+}
+
SDValue
AArch64TargetLowering::PerformDAGCombine(SDNode *N,
DAGCombinerInfo &DCI) const {
return PerformShiftCombine(N, DCI, getSubtarget());
case ISD::INTRINSIC_WO_CHAIN:
return PerformIntrinsicCombine(N, DCI.DAG);
+ case AArch64ISD::NEON_VDUPLANE:
+ return CombineVLDDUP(N, DCI);
+ case AArch64ISD::NEON_LD2DUP:
+ case AArch64ISD::NEON_LD3DUP:
+ case AArch64ISD::NEON_LD4DUP:
+ return CombineBaseUpdate(N, DCI);
case ISD::INTRINSIC_VOID:
case ISD::INTRINSIC_W_CHAIN:
switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
case Intrinsic::arm_neon_vst2:
case Intrinsic::arm_neon_vst3:
case Intrinsic::arm_neon_vst4:
+ case Intrinsic::arm_neon_vld2lane:
+ case Intrinsic::arm_neon_vld3lane:
+ case Intrinsic::arm_neon_vld4lane:
+ case Intrinsic::aarch64_neon_vld1x2:
+ case Intrinsic::aarch64_neon_vld1x3:
+ case Intrinsic::aarch64_neon_vld1x4:
+ case Intrinsic::aarch64_neon_vst1x2:
+ case Intrinsic::aarch64_neon_vst1x3:
+ case Intrinsic::aarch64_neon_vst1x4:
+ case Intrinsic::arm_neon_vst2lane:
+ case Intrinsic::arm_neon_vst3lane:
+ case Intrinsic::arm_neon_vst4lane:
return CombineBaseUpdate(N, DCI);
default:
break;
return false;
}
+// Check whether a Build Vector could be presented as Shuffle Vector. If yes,
+// try to call LowerVECTOR_SHUFFLE to lower it.
+bool AArch64TargetLowering::isKnownShuffleVector(SDValue Op, SelectionDAG &DAG,
+ SDValue &Res) const {
+ SDLoc DL(Op);
+ EVT VT = Op.getValueType();
+ unsigned NumElts = VT.getVectorNumElements();
+ unsigned V0NumElts = 0;
+ int Mask[16];
+ SDValue V0, V1;
+
+ // Check if all elements are extracted from less than 3 vectors.
+ for (unsigned i = 0; i < NumElts; ++i) {
+ SDValue Elt = Op.getOperand(i);
+ if (Elt.getOpcode() != ISD::EXTRACT_VECTOR_ELT)
+ return false;
+
+ if (V0.getNode() == 0) {
+ V0 = Elt.getOperand(0);
+ V0NumElts = V0.getValueType().getVectorNumElements();
+ }
+ if (Elt.getOperand(0) == V0) {
+ Mask[i] = (cast<ConstantSDNode>(Elt->getOperand(1))->getZExtValue());
+ continue;
+ } else if (V1.getNode() == 0) {
+ V1 = Elt.getOperand(0);
+ }
+ if (Elt.getOperand(0) == V1) {
+ unsigned Lane = cast<ConstantSDNode>(Elt->getOperand(1))->getZExtValue();
+ Mask[i] = (Lane + V0NumElts);
+ continue;
+ } else {
+ return false;
+ }
+ }
+
+ if (!V1.getNode() && V0NumElts == NumElts * 2) {
+ V1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V0,
+ DAG.getConstant(NumElts, MVT::i64));
+ V0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V0,
+ DAG.getConstant(0, MVT::i64));
+ V0NumElts = V0.getValueType().getVectorNumElements();
+ }
+
+ if (V1.getNode() && NumElts == V0NumElts &&
+ V0NumElts == V1.getValueType().getVectorNumElements()) {
+ SDValue Shuffle = DAG.getVectorShuffle(VT, DL, V0, V1, Mask);
+ Res = LowerVECTOR_SHUFFLE(Shuffle, DAG);
+ return true;
+ } else
+ return false;
+}
+
// If this is a case we can't handle, return null and let the default
// expansion code take care of it.
SDValue
return DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VT, Value);
unsigned EltSize = VT.getVectorElementType().getSizeInBits();
- // Use VDUP for non-constant splats.
if (hasDominantValue && EltSize <= 64) {
+ // Use VDUP for non-constant splats.
if (!isConstant) {
SDValue N;
SmallVector<SDValue, 3> Ops;
Ops.push_back(N);
Ops.push_back(Op.getOperand(I));
- Ops.push_back(DAG.getConstant(I, MVT::i32));
+ Ops.push_back(DAG.getConstant(I, MVT::i64));
N = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, &Ops[0], 3);
}
}
if (isConstant)
return SDValue();
+ // Try to lower this in lowering ShuffleVector way.
+ SDValue Shuf;
+ if (isKnownShuffleVector(Op, DAG, Shuf))
+ return Shuf;
+
// If all else fails, just use a sequence of INSERT_VECTOR_ELT when we
// know the default expansion would otherwise fall back on something even
// worse. For a vector with one or two non-undef values, that's
SDValue V = Op.getOperand(i);
if (V.getOpcode() == ISD::UNDEF)
continue;
- SDValue LaneIdx = DAG.getConstant(i, MVT::i32);
+ SDValue LaneIdx = DAG.getConstant(i, MVT::i64);
Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, Vec, V, LaneIdx);
}
return Vec;
return SDValue();
}
+/// isREVMask - Check if a vector shuffle corresponds to a REV
+/// instruction with the specified blocksize. (The order of the elements
+/// within each block of the vector is reversed.)
+static bool isREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
+ assert((BlockSize == 16 || BlockSize == 32 || BlockSize == 64) &&
+ "Only possible block sizes for REV are: 16, 32, 64");
+
+ unsigned EltSz = VT.getVectorElementType().getSizeInBits();
+ if (EltSz == 64)
+ return false;
+
+ unsigned NumElts = VT.getVectorNumElements();
+ unsigned BlockElts = M[0] + 1;
+ // If the first shuffle index is UNDEF, be optimistic.
+ if (M[0] < 0)
+ BlockElts = BlockSize / EltSz;
+
+ if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
+ return false;
+
+ for (unsigned i = 0; i < NumElts; ++i) {
+ if (M[i] < 0)
+ continue; // ignore UNDEF indices
+ if ((unsigned)M[i] != (i - i % BlockElts) + (BlockElts - 1 - i % BlockElts))
+ return false;
+ }
+
+ return true;
+}
+
+// isPermuteMask - Check whether the vector shuffle matches to UZP, ZIP and
+// TRN instruction.
+static unsigned isPermuteMask(ArrayRef<int> M, EVT VT) {
+ unsigned NumElts = VT.getVectorNumElements();
+ if (NumElts < 4)
+ return 0;
+
+ bool ismatch = true;
+
+ // Check UZP1
+ for (unsigned i = 0; i < NumElts; ++i) {
+ if ((unsigned)M[i] != i * 2) {
+ ismatch = false;
+ break;
+ }
+ }
+ if (ismatch)
+ return AArch64ISD::NEON_UZP1;
+
+ // Check UZP2
+ ismatch = true;
+ for (unsigned i = 0; i < NumElts; ++i) {
+ if ((unsigned)M[i] != i * 2 + 1) {
+ ismatch = false;
+ break;
+ }
+ }
+ if (ismatch)
+ return AArch64ISD::NEON_UZP2;
+
+ // Check ZIP1
+ ismatch = true;
+ for (unsigned i = 0; i < NumElts; ++i) {
+ if ((unsigned)M[i] != i / 2 + NumElts * (i % 2)) {
+ ismatch = false;
+ break;
+ }
+ }
+ if (ismatch)
+ return AArch64ISD::NEON_ZIP1;
+
+ // Check ZIP2
+ ismatch = true;
+ for (unsigned i = 0; i < NumElts; ++i) {
+ if ((unsigned)M[i] != (NumElts + i) / 2 + NumElts * (i % 2)) {
+ ismatch = false;
+ break;
+ }
+ }
+ if (ismatch)
+ return AArch64ISD::NEON_ZIP2;
+
+ // Check TRN1
+ ismatch = true;
+ for (unsigned i = 0; i < NumElts; ++i) {
+ if ((unsigned)M[i] != i + (NumElts - 1) * (i % 2)) {
+ ismatch = false;
+ break;
+ }
+ }
+ if (ismatch)
+ return AArch64ISD::NEON_TRN1;
+
+ // Check TRN2
+ ismatch = true;
+ for (unsigned i = 0; i < NumElts; ++i) {
+ if ((unsigned)M[i] != 1 + i + (NumElts - 1) * (i % 2)) {
+ ismatch = false;
+ break;
+ }
+ }
+ if (ismatch)
+ return AArch64ISD::NEON_TRN2;
+
+ return 0;
+}
+
SDValue
AArch64TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
- SelectionDAG &DAG) const {
+ SelectionDAG &DAG) const {
SDValue V1 = Op.getOperand(0);
SDValue V2 = Op.getOperand(1);
SDLoc dl(Op);
ArrayRef<int> ShuffleMask = SVN->getMask();
unsigned EltSize = VT.getVectorElementType().getSizeInBits();
- if (EltSize <= 64) {
- if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
- int Lane = SVN->getSplatIndex();
- // If this is undef splat, generate it via "just" vdup, if possible.
- if (Lane == -1) Lane = 0;
-
- // Test if V1 is a SCALAR_TO_VECTOR.
- if (V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
- return DAG.getNode(AArch64ISD::NEON_VDUP, dl, VT, V1.getOperand(0));
- }
- // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR.
- if (V1.getOpcode() == ISD::BUILD_VECTOR) {
- bool IsScalarToVector = true;
- for (unsigned i = 0, e = V1.getNumOperands(); i != e; ++i)
- if (V1.getOperand(i).getOpcode() != ISD::UNDEF &&
- i != (unsigned)Lane) {
- IsScalarToVector = false;
- break;
- }
- if (IsScalarToVector)
- return DAG.getNode(AArch64ISD::NEON_VDUP, dl, VT,
- V1.getOperand(Lane));
- }
- return DAG.getNode(AArch64ISD::NEON_VDUPLANE, dl, VT, V1,
- DAG.getConstant(Lane, MVT::i64));
+ if (EltSize > 64)
+ return SDValue();
+
+ if (isREVMask(ShuffleMask, VT, 64))
+ return DAG.getNode(AArch64ISD::NEON_REV64, dl, VT, V1);
+ if (isREVMask(ShuffleMask, VT, 32))
+ return DAG.getNode(AArch64ISD::NEON_REV32, dl, VT, V1);
+ if (isREVMask(ShuffleMask, VT, 16))
+ return DAG.getNode(AArch64ISD::NEON_REV16, dl, VT, V1);
+
+ unsigned ISDNo = isPermuteMask(ShuffleMask, VT);
+ if (ISDNo)
+ return DAG.getNode(ISDNo, dl, VT, V1, V2);
+
+ // If the element of shuffle mask are all the same constant, we can
+ // transform it into either NEON_VDUP or NEON_VDUPLANE
+ if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
+ int Lane = SVN->getSplatIndex();
+ // If this is undef splat, generate it via "just" vdup, if possible.
+ if (Lane == -1) Lane = 0;
+
+ // Test if V1 is a SCALAR_TO_VECTOR.
+ if (V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
+ return DAG.getNode(AArch64ISD::NEON_VDUP, dl, VT, V1.getOperand(0));
}
- // For shuffle mask like "0, 1, 2, 3, 4, 5, 13, 7", try to generate insert
- // by element from V2 to V1 .
- // If shuffle mask is like "0, 1, 10, 11, 12, 13, 14, 15", V2 would be a
- // better choice to be inserted than V1 as less insert needed, so we count
- // element to be inserted for both V1 and V2, and select less one as insert
- // target.
-
- // Collect elements need to be inserted and their index.
- SmallVector<int, 8> NV1Elt;
- SmallVector<int, 8> N1Index;
- SmallVector<int, 8> NV2Elt;
- SmallVector<int, 8> N2Index;
- int Length = ShuffleMask.size();
- int V1EltNum = V1.getValueType().getVectorNumElements();
- for (int I = 0; I != Length; ++I) {
- if (ShuffleMask[I] != I) {
- NV1Elt.push_back(ShuffleMask[I]);
- N1Index.push_back(I);
- }
+ // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR.
+ if (V1.getOpcode() == ISD::BUILD_VECTOR) {
+ bool IsScalarToVector = true;
+ for (unsigned i = 0, e = V1.getNumOperands(); i != e; ++i)
+ if (V1.getOperand(i).getOpcode() != ISD::UNDEF &&
+ i != (unsigned)Lane) {
+ IsScalarToVector = false;
+ break;
+ }
+ if (IsScalarToVector)
+ return DAG.getNode(AArch64ISD::NEON_VDUP, dl, VT,
+ V1.getOperand(Lane));
}
- for (int I = 0; I != Length; ++I) {
- if (ShuffleMask[I] != (I + V1EltNum)) {
- NV2Elt.push_back(ShuffleMask[I]);
- N2Index.push_back(I);
- }
+
+ // Test if V1 is a EXTRACT_SUBVECTOR.
+ if (V1.getOpcode() == ISD::EXTRACT_SUBVECTOR) {
+ int ExtLane = cast<ConstantSDNode>(V1.getOperand(1))->getZExtValue();
+ return DAG.getNode(AArch64ISD::NEON_VDUPLANE, dl, VT, V1.getOperand(0),
+ DAG.getConstant(Lane + ExtLane, MVT::i64));
}
+ // Test if V1 is a CONCAT_VECTORS.
+ if (V1.getOpcode() == ISD::CONCAT_VECTORS &&
+ V1.getOperand(1).getOpcode() == ISD::UNDEF) {
+ SDValue Op0 = V1.getOperand(0);
+ assert((unsigned)Lane < Op0.getValueType().getVectorNumElements() &&
+ "Invalid vector lane access");
+ return DAG.getNode(AArch64ISD::NEON_VDUPLANE, dl, VT, Op0,
+ DAG.getConstant(Lane, MVT::i64));
+ }
+
+ return DAG.getNode(AArch64ISD::NEON_VDUPLANE, dl, VT, V1,
+ DAG.getConstant(Lane, MVT::i64));
+ }
- // Decide which to be inserted. If all lanes mismatch, neither V1 nor V2
- // will be inserted.
- SDValue InsV = V1;
- SmallVector<int, 8> InsMasks = NV1Elt;
- SmallVector<int, 8> InsIndex = N1Index;
- if ((int)NV1Elt.size() != Length || (int)NV2Elt.size() != Length) {
- if (NV1Elt.size() > NV2Elt.size()) {
- InsV = V2;
- InsMasks = NV2Elt;
- InsIndex = N2Index;
+ int Length = ShuffleMask.size();
+ int V1EltNum = V1.getValueType().getVectorNumElements();
+
+ // If the number of v1 elements is the same as the number of shuffle mask
+ // element and the shuffle masks are sequential values, we can transform
+ // it into NEON_VEXTRACT.
+ if (V1EltNum == Length) {
+ // Check if the shuffle mask is sequential.
+ bool IsSequential = true;
+ int CurMask = ShuffleMask[0];
+ for (int I = 0; I < Length; ++I) {
+ if (ShuffleMask[I] != CurMask) {
+ IsSequential = false;
+ break;
}
- } else {
- InsV = DAG.getNode(ISD::UNDEF, dl, VT);
+ CurMask++;
}
+ if (IsSequential) {
+ assert((EltSize % 8 == 0) && "Bitsize of vector element is incorrect");
+ unsigned VecSize = EltSize * V1EltNum;
+ unsigned Index = (EltSize/8) * ShuffleMask[0];
+ if (VecSize == 64 || VecSize == 128)
+ return DAG.getNode(AArch64ISD::NEON_VEXTRACT, dl, VT, V1, V2,
+ DAG.getConstant(Index, MVT::i64));
+ }
+ }
- SDValue PassN;
+ // For shuffle mask like "0, 1, 2, 3, 4, 5, 13, 7", try to generate insert
+ // by element from V2 to V1 .
+ // If shuffle mask is like "0, 1, 10, 11, 12, 13, 14, 15", V2 would be a
+ // better choice to be inserted than V1 as less insert needed, so we count
+ // element to be inserted for both V1 and V2, and select less one as insert
+ // target.
+
+ // Collect elements need to be inserted and their index.
+ SmallVector<int, 8> NV1Elt;
+ SmallVector<int, 8> N1Index;
+ SmallVector<int, 8> NV2Elt;
+ SmallVector<int, 8> N2Index;
+ for (int I = 0; I != Length; ++I) {
+ if (ShuffleMask[I] != I) {
+ NV1Elt.push_back(ShuffleMask[I]);
+ N1Index.push_back(I);
+ }
+ }
+ for (int I = 0; I != Length; ++I) {
+ if (ShuffleMask[I] != (I + V1EltNum)) {
+ NV2Elt.push_back(ShuffleMask[I]);
+ N2Index.push_back(I);
+ }
+ }
- for (int I = 0, E = InsMasks.size(); I != E; ++I) {
- SDValue ExtV = V1;
- int Mask = InsMasks[I];
- if (Mask > V1EltNum) {
- ExtV = V2;
- Mask -= V1EltNum;
- }
- // Any value type smaller than i32 is illegal in AArch64, and this lower
- // function is called after legalize pass, so we need to legalize
- // the result here.
- EVT EltVT;
- if (VT.getVectorElementType().isFloatingPoint())
- EltVT = (EltSize == 64) ? MVT::f64 : MVT::f32;
- else
- EltVT = (EltSize == 64) ? MVT::i64 : MVT::i32;
-
- PassN = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, ExtV,
- DAG.getConstant(Mask, MVT::i64));
- PassN = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, InsV, PassN,
- DAG.getConstant(InsIndex[I], MVT::i64));
+ // Decide which to be inserted. If all lanes mismatch, neither V1 nor V2
+ // will be inserted.
+ SDValue InsV = V1;
+ SmallVector<int, 8> InsMasks = NV1Elt;
+ SmallVector<int, 8> InsIndex = N1Index;
+ if ((int)NV1Elt.size() != Length || (int)NV2Elt.size() != Length) {
+ if (NV1Elt.size() > NV2Elt.size()) {
+ InsV = V2;
+ InsMasks = NV2Elt;
+ InsIndex = N2Index;
}
- return PassN;
+ } else {
+ InsV = DAG.getNode(ISD::UNDEF, dl, VT);
}
- return SDValue();
+ for (int I = 0, E = InsMasks.size(); I != E; ++I) {
+ SDValue ExtV = V1;
+ int Mask = InsMasks[I];
+ if (Mask >= V1EltNum) {
+ ExtV = V2;
+ Mask -= V1EltNum;
+ }
+ // Any value type smaller than i32 is illegal in AArch64, and this lower
+ // function is called after legalize pass, so we need to legalize
+ // the result here.
+ EVT EltVT;
+ if (VT.getVectorElementType().isFloatingPoint())
+ EltVT = (EltSize == 64) ? MVT::f64 : MVT::f32;
+ else
+ EltVT = (EltSize == 64) ? MVT::i64 : MVT::i32;
+
+ if (Mask >= 0) {
+ ExtV = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, ExtV,
+ DAG.getConstant(Mask, MVT::i64));
+ InsV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, InsV, ExtV,
+ DAG.getConstant(InsIndex[I], MVT::i64));
+ }
+ }
+ return InsV;
}
AArch64TargetLowering::ConstraintType
case Intrinsic::arm_neon_vld1:
case Intrinsic::arm_neon_vld2:
case Intrinsic::arm_neon_vld3:
- case Intrinsic::arm_neon_vld4: {
+ case Intrinsic::arm_neon_vld4:
+ case Intrinsic::aarch64_neon_vld1x2:
+ case Intrinsic::aarch64_neon_vld1x3:
+ case Intrinsic::aarch64_neon_vld1x4:
+ case Intrinsic::arm_neon_vld2lane:
+ case Intrinsic::arm_neon_vld3lane:
+ case Intrinsic::arm_neon_vld4lane: {
Info.opc = ISD::INTRINSIC_W_CHAIN;
// Conservatively set memVT to the entire set of vectors loaded.
uint64_t NumElts = getDataLayout()->getTypeAllocSize(I.getType()) / 8;
case Intrinsic::arm_neon_vst1:
case Intrinsic::arm_neon_vst2:
case Intrinsic::arm_neon_vst3:
- case Intrinsic::arm_neon_vst4: {
+ case Intrinsic::arm_neon_vst4:
+ case Intrinsic::aarch64_neon_vst1x2:
+ case Intrinsic::aarch64_neon_vst1x3:
+ case Intrinsic::aarch64_neon_vst1x4:
+ case Intrinsic::arm_neon_vst2lane:
+ case Intrinsic::arm_neon_vst3lane:
+ case Intrinsic::arm_neon_vst4lane: {
Info.opc = ISD::INTRINSIC_VOID;
// Conservatively set memVT to the entire set of vectors stored.
unsigned NumElts = 0;
projects / oota-llvm.git / blobdiff