// Get block size for the first convolution.
      int64_t block_size = GetConv2DBlockSize(conv2d);
      auto arg_num = arg_num_and_num_users.value().arg_num;
      auto num_users = arg_num_and_num_users.value().num_users;
      argnum_and_convolutions[arg_num].emplace_back(conv2d, block_size);
      argnum_num_users[arg_num] = num_users;
      return WalkResult::interrupt();
    }
    return WalkResult::advance();
  });

    int arg_num = compilation_result->input_mapping[i];
    CHECK_GE(arg_num, missing_ctx_input_prefix);
    const xla::Shape& device_shape = compilation_result->xla_input_shapes[i];
    const xla::Shape& host_shape =
        xla::ShapeUtil::DeviceShapeToHostShape(device_shape);

    auto resource_var_it = resource_vars.find(arg_num);

//   temp bytes aligned: {{TEMP_BYTES_ALIGNED}}
class {{CLASS}} final : public tensorflow::XlaCompiledCpuFunction {
 public:
  // Number of input arguments for the compiled computation.
  static constexpr size_t kNumArgs = {{ARG_NUM}};

  static constexpr size_t kNumResults = {{RESULT_NUM}};

  // Number of variables for the compiled computation.
  static constexpr size_t kNumVariables = {{VARIABLE_NUM}};

	// LSRV <Wd>, <Wn>, <Wm>
	{0xffe0fc00, 0x1ac02400, LSRV, instArgs{arg_Wd, arg_Wn, arg_Wm}, nil},
	// LSR <Xd>, <Xn>, <Xm>
	{0xffe0fc00, 0x9ac02400, LSR, instArgs{arg_Xd, arg_Xn, arg_Xm}, nil},
	// LSRV <Xd>, <Xn>, <Xm>
	{0xffe0fc00, 0x9ac02400, LSRV, instArgs{arg_Xd, arg_Xn, arg_Xm}, nil},
	// MUL <Wd>, <Wn>, <Wm>
	{0xffe0fc00, 0x1b007c00, MUL, instArgs{arg_Wd, arg_Wn, arg_Wm}, nil},

//     one register with arrangement:H and element index encoded in "Q:S:size<1>".

type instArg uint16

const (
	_ instArg = iota
	arg_Bt
	arg_Cm
	arg_Cn
	arg_cond_AllowALNV_Normal
	arg_conditional
	arg_cond_NotAllowALNV_Invert
	arg_Da
	arg_Dd
	arg_Dm
	arg_Dn
	arg_Dt
	arg_Dt2
	arg_Hd
	arg_Hn
	arg_Ht
	arg_IAddSub
	arg_immediate_0_127_CRm_op2
	arg_immediate_0_15_CRm

	argNum := state.argNums[formatArg]
	if argNum <= 0 {
		// Shouldn't happen, so catch it with prejudice.
		panic("negative arg num")
	}
	if argNum < len(call.Args)-1 {
		return true // Always OK.
	}
	if call.Ellipsis.IsValid() {
		return false // We just can't tell; there could be many more arguments.
	}
	if argNum < len(call.Args) {
		return true
	}

		w.err, _ = a[p.wrappedErrs[0]].(error)
		err = w
	default:
		if p.reordered {
			slices.Sort(p.wrappedErrs)
		}
		var errs []error
		for i, argNum := range p.wrappedErrs {
			if i > 0 && p.wrappedErrs[i-1] == argNum {
				continue
			}
			if e, ok := a[argNum].(error); ok {
				errs = append(errs, e)
			}
		}
		err = &wrapErrors{s, errs}
	}
	p.free()
	return err
}

type wrapError struct {

func intFromArg(a []any, argNum int) (num int, isInt bool, newArgNum int) {
	newArgNum = argNum
	if argNum < len(a) {
		num, isInt = a[argNum].(int) // Almost always OK.
		if !isInt {
			// Work harder.
			switch v := reflect.ValueOf(a[argNum]); v.Kind() {
			case reflect.Int, reflect.Int8, reflect.Int16, reflect.Int32, reflect.Int64:
				n := v.Int()

	if ARM64.HasCPUID {
		midr := getMIDR()
		part_num := uint16((midr >> 4) & 0xfff)
		implementer := byte((midr >> 24) & 0xff)

		// d0c - NeoverseN1
		// d40 - NeoverseV1
		// d49 - NeoverseN2
		// d4f - NeoverseV2
		if implementer == 'A' && (part_num == 0xd0c || part_num == 0xd40 ||
			part_num == 0xd49 || part_num == 0xd4f) {
			ARM64.IsNeoverse = true
		}
	}
}

		zb0001--
		field, err = dc.ReadMapKeyPtr()
		if err != nil {
			err = msgp.WrapError(err)
			return
		}
		switch msgp.UnsafeString(field) {
		case "OrgNum":
			z.OrgNum, err = dc.ReadInt()
			if err != nil {
				err = msgp.WrapError(err, "OrgNum")
				return
			}
		case "OrgString":
			z.OrgString, err = dc.ReadString()
			if err != nil {
				err = msgp.WrapError(err, "OrgString")
				return