+SDValue
+AArch64DAGToDAGISel::getTargetSubregToReg(int SRIdx, SDLoc DL, EVT VT, EVT VTD,
+ SDValue Operand) {
+ SDNode *Reg = CurDAG->getMachineNode(TargetOpcode::SUBREG_TO_REG, DL,
+ VT, VTD, MVT::Other,
+ CurDAG->getTargetConstant(0, MVT::i64),
+ Operand,
+ CurDAG->getTargetConstant(AArch64::sub_64, MVT::i32));
+ return SDValue(Reg, 0);
+}
+
+SDNode *AArch64DAGToDAGISel::SelectVLDDup(SDNode *N, bool isUpdating,
+ unsigned NumVecs,
+ const uint16_t *Opcodes) {
+ assert(NumVecs >=2 && NumVecs <= 4 && "Load Dup NumVecs out-of-range");
+ SDLoc dl(N);
+
+ EVT VT = N->getValueType(0);
+ unsigned OpcodeIndex;
+ bool is64BitVector = VT.is64BitVector();
+ switch (VT.getScalarType().getSizeInBits()) {
+ case 8: OpcodeIndex = is64BitVector ? 0 : 4; break;
+ case 16: OpcodeIndex = is64BitVector ? 1 : 5; break;
+ case 32: OpcodeIndex = is64BitVector ? 2 : 6; break;
+ case 64: OpcodeIndex = is64BitVector ? 3 : 7; break;
+ default: llvm_unreachable("unhandled vector duplicate lane load type");
+ }
+ unsigned Opc = Opcodes[OpcodeIndex];
+
+ SDValue SuperReg;
+ SmallVector<SDValue, 6> Ops;
+ Ops.push_back(N->getOperand(1)); // Push back the Memory Address
+ if (isUpdating) {
+ SDValue Inc = N->getOperand(2);
+ if (!isa<ConstantSDNode>(Inc.getNode())) // Increment in Register
+ Opc = getVLDSTRegisterUpdateOpcode(Opc);
+ Ops.push_back(Inc);
+ }
+ Ops.push_back(N->getOperand(0)); // Push back the Chain
+
+ SmallVector<EVT, 3> ResTys;
+ // Push back the type of return super register
+ if (NumVecs == 3)
+ ResTys.push_back(MVT::Untyped);
+ else {
+ EVT ResTy = EVT::getVectorVT(*CurDAG->getContext(), MVT::i64,
+ is64BitVector ? NumVecs : NumVecs * 2);
+ ResTys.push_back(ResTy);
+ }
+ if (isUpdating)
+ ResTys.push_back(MVT::i64); // Type of the updated register
+ ResTys.push_back(MVT::Other); // Type of the Chain
+ SDNode *VLdDup = CurDAG->getMachineNode(Opc, dl, ResTys, Ops);
+
+ // Transfer memoperands.
+ MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
+ MemOp[0] = cast<MemIntrinsicSDNode>(N)->getMemOperand();
+ cast<MachineSDNode>(VLdDup)->setMemRefs(MemOp, MemOp + 1);
+
+ SuperReg = SDValue(VLdDup, 0);
+ unsigned Sub0 = is64BitVector ? AArch64::dsub_0 : AArch64::qsub_0;
+ // Update uses of each registers in super register
+ for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
+ ReplaceUses(SDValue(N, Vec),
+ CurDAG->getTargetExtractSubreg(Sub0 + Vec, dl, VT, SuperReg));
+ // Update uses of the Chain
+ ReplaceUses(SDValue(N, NumVecs), SDValue(VLdDup, 1));
+ if (isUpdating)
+ ReplaceUses(SDValue(N, NumVecs + 1), SDValue(VLdDup, 2));
+ return NULL;
+}
+
+// We only have 128-bit vector type of load/store lane instructions.
+// If it is 64-bit vector, we also select it to the 128-bit instructions.
+// Just use SUBREG_TO_REG to adapt the input to 128-bit vector and
+// EXTRACT_SUBREG to get the 64-bit vector from the 128-bit vector output.
+SDNode *AArch64DAGToDAGISel::SelectVLDSTLane(SDNode *N, bool IsLoad,
+ bool isUpdating, unsigned NumVecs,
+ const uint16_t *Opcodes) {
+ assert(NumVecs >= 2 && NumVecs <= 4 && "VLDSTLane NumVecs out-of-range");
+ SDLoc dl(N);
+ unsigned AddrOpIdx = isUpdating ? 1 : 2;
+ unsigned Vec0Idx = 3;
+
+ SDValue Chain = N->getOperand(0);
+ unsigned Lane =
+ cast<ConstantSDNode>(N->getOperand(Vec0Idx + NumVecs))->getZExtValue();
+ EVT VT = N->getOperand(Vec0Idx).getValueType();
+ bool is64BitVector = VT.is64BitVector();
+ EVT VT64; // 64-bit Vector Type
+
+ if (is64BitVector) {
+ VT64 = VT;
+ VT = EVT::getVectorVT(*CurDAG->getContext(), VT.getVectorElementType(),
+ VT.getVectorNumElements() * 2);
+ }
+
+ unsigned OpcodeIndex;
+ switch (VT.getScalarType().getSizeInBits()) {
+ case 8: OpcodeIndex = 0; break;
+ case 16: OpcodeIndex = 1; break;
+ case 32: OpcodeIndex = 2; break;
+ case 64: OpcodeIndex = 3; break;
+ default: llvm_unreachable("unhandled vector lane load/store type");
+ }
+ unsigned Opc = Opcodes[OpcodeIndex];
+
+ SmallVector<EVT, 3> ResTys;
+ if (IsLoad) {
+ // Push back the type of return super register
+ if (NumVecs == 3)
+ ResTys.push_back(MVT::Untyped);
+ else {
+ EVT ResTy = EVT::getVectorVT(*CurDAG->getContext(), MVT::i64,
+ is64BitVector ? NumVecs : NumVecs * 2);
+ ResTys.push_back(ResTy);
+ }
+ }
+ if (isUpdating)
+ ResTys.push_back(MVT::i64); // Type of the updated register
+ ResTys.push_back(MVT::Other); // Type of Chain
+ SmallVector<SDValue, 5> Ops;
+ Ops.push_back(N->getOperand(AddrOpIdx)); // Push back the Memory Address
+ if (isUpdating) {
+ SDValue Inc = N->getOperand(AddrOpIdx + 1);
+ if (!isa<ConstantSDNode>(Inc.getNode())) // Increment in Register
+ Opc = getVLDSTRegisterUpdateOpcode(Opc);
+ Ops.push_back(Inc);
+ }
+
+ SmallVector<SDValue, 4> Regs(N->op_begin() + Vec0Idx,
+ N->op_begin() + Vec0Idx + NumVecs);
+ if (is64BitVector)
+ for (unsigned i = 0; i < Regs.size(); i++)
+ Regs[i] = getTargetSubregToReg(AArch64::sub_64, dl, VT, VT64, Regs[i]);
+ SDValue SuperReg = createQTuple(Regs);
+
+ Ops.push_back(SuperReg); // Source Reg
+ SDValue LaneValue = CurDAG->getTargetConstant(Lane, MVT::i32);
+ Ops.push_back(LaneValue);
+ Ops.push_back(Chain); // Push back the Chain
+
+ SDNode *VLdLn = CurDAG->getMachineNode(Opc, dl, ResTys, Ops);
+ MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
+ MemOp[0] = cast<MemIntrinsicSDNode>(N)->getMemOperand();
+ cast<MachineSDNode>(VLdLn)->setMemRefs(MemOp, MemOp + 1);
+ if (!IsLoad)
+ return VLdLn;
+
+ // Extract the subregisters.
+ SuperReg = SDValue(VLdLn, 0);
+ unsigned Sub0 = AArch64::qsub_0;
+ // Update uses of each registers in super register
+ for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
+ SDValue SUB0 = CurDAG->getTargetExtractSubreg(Sub0 + Vec, dl, VT, SuperReg);
+ if (is64BitVector) {
+ SUB0 = CurDAG->getTargetExtractSubreg(AArch64::sub_64, dl, VT64, SUB0);
+ }
+ ReplaceUses(SDValue(N, Vec), SUB0);
+ }
+ ReplaceUses(SDValue(N, NumVecs), SDValue(VLdLn, 1));
+ if (isUpdating)
+ ReplaceUses(SDValue(N, NumVecs + 1), SDValue(VLdLn, 2));
+ return NULL;
+}
+