const int num_inputs = input_shapes->num_items;
  NodeDef node_def;
  tensorflow::ImmediateExecutionOperation* op = tensorflow::unwrap(tfe_op);
  node_def.set_name(op->Name());
  node_def.set_op(op->Name());
  for (int i = 0; i < num_inputs; ++i) {
    node_def.add_input("dummy_input");
  }
  OperationFromInterface(op)->Attrs().FillAttrValueMap(node_def.mutable_attr());

	if n.TypeWord == nil || n.TypeWord.Op() == ir.OADDR && n.TypeWord.(*ir.AddrExpr).X.Op() == ir.OLINKSYMOFFSET {
		// Fixed (not loaded from a dictionary) type.
		ta := ir.NewTypeAssertExpr(base.Pos, c, toType)
		ta.SetOp(ir.ODOTTYPE2)
		// Allocate a descriptor for this conversion to pass to the runtime.
		ta.Descriptor = makeTypeAssertDescriptor(toType, true)
		rhs = ta
	} else {

	// same type as the variable on the lhs.
	if ok := n.Lhs[1]; !ir.IsBlank(ok) && ok.Type().IsBoolean() {
		call.Type().Field(1).Type = ok.Type()
	}
	n.Rhs = []ir.Node{call}
	n.SetOp(ir.OAS2FUNC)

	// don't generate a = *var if a is _
	if ir.IsBlank(a) {
		return walkExpr(typecheck.Stmt(n), init)
	}

	var_ := typecheck.TempAt(base.Pos, ir.CurFunc, types.NewPtr(t.Elem()))

	// because walkExpr will lower the string
	// conversion into a runtime call.
	// See issue 24937 for more discussion.
	if cond.Op() == ir.OBYTES2STR && allCaseExprsAreSideEffectFree(sw) {
		cond := cond.(*ir.ConvExpr)
		cond.SetOp(ir.OBYTES2STRTMP)
	}

	cond = walkExpr(cond, sw.PtrInit())
	if cond.Op() != ir.OLITERAL && cond.Op() != ir.ONIL {
		cond = copyExpr(cond, cond.Type(), &sw.Compiled)
	}

	base.Pos = lno

	s := exprSwitch{

	if n.RType != nil && n.RType.Op() == ir.OADDR {
		addr := n.RType.(*ir.AddrExpr)
		if addr.X.Op() == ir.OLINKSYMOFFSET {
			r := ir.NewTypeAssertExpr(n.Pos(), n.X, n.Type())
			if n.Op() == ir.ODYNAMICDOTTYPE2 {
				r.SetOp(ir.ODOTTYPE2)
			}
			r.SetType(n.Type())
			r.SetTypecheck(1)
			return walkExpr(r, init)
		}
	}
	return n
}

// walkIndex walks an OINDEX node.

    bool reuse_existing_functions, FunctionLibraryDefinition* library) {
  // name_in is copied here because name may be modified below if
  // rewrite_subgraph_fn is true.
  string name = name_in;
  call_node_def_.set_op(name);
  call_node_def_.set_name(name);
  call_node_def_.set_device(device_);

  if (rewrite_subgraph_fn) {
    std::vector<OutputTensor> arg_source_tensors(args_by_src_.size());

  *flib_def.add_function() = std::move(func);
  TF_ASSERT_OK(root.graph()->AddFunctionLibrary(flib_def));
  NodeDef call_node;
  call_node.set_name("fn_call");
  call_node.set_op("Stateful_func");
  Status status;
  Node* call = root.graph()->AddNode(call_node, &status);
  TF_ASSERT_OK(status);

  Output shape = Output(call, 0);
  Output reshape_input =

    std::vector<int>* output_permutation, NodeDef* node_def) {
  string old_name = node_def->op();
  string new_name =
      absl::StrCat(xla_cluster_name_, "_", new_function_name_, "_", old_name);
  node_def->set_op(new_name);
  node_def->set_name(new_name);

  // Later we will run PruneForReverseReachability(), so make sure all original
  // nodes are reachable from sink node and won't be removed.

    EXPECT_FALSE(TF_TensorIsAligned(a));
  }
  TF_DeleteTensor(a);
}

TEST(CAPI, MessageBufferConversion) {
  NodeDef node_in, node_out;
  node_in.set_name("Test name");
  node_in.set_op("Test op");

  TF_Buffer* buffer = TF_NewBuffer();
  TF_CHECK_OK(MessageToBuffer(node_in, buffer));
  TF_CHECK_OK(BufferToMessage(buffer, &node_out));
  TF_DeleteBuffer(buffer);

    // Calculate the scaling that needs to be applied to the iota.
    auto step_numerator = rewriter.create<chlo::BroadcastSubOp>(
        op.getLoc(), op.getStart().getType(), op.getStop(), op.getStart(),
        hlo::getBroadcastDimensionsAttr(&rewriter, op.getStop(),
                                        op.getStart()));
    Value step_denominator = rewriter.create<ConvertOp>(
        op.getLoc(), op.getNum(), result_type.getElementType());