op_state.addOperands(op.getOperands());
    op_state.addTypes(op.getResultTypes());

    SmallVector<NamedAttribute, 2> attrs;
    std::string parsed_op_name;
    tensorflow::NodeDef node_def;
    if (failed(ParseCustomOption(op.getCustomOption().getValue(), op.getLoc(),
                                 parsed_op_name, attrs, node_def))) {
      return failure();
    }
    if (parsed_op_name != tf_op_name) {
      return op.emitOpError(

def _impl(ctx):
    classes = ",".join(
        [_AsClassName(x) for x in ctx.attr.srcs],
    )
    ctx.actions.write(output = ctx.outputs.out, content = _OUTPUT % (
        classes,
        ctx.attr.outname,
    ))

_GenSuite = rule(
    attrs = {
        "srcs": attr.label_list(allow_files = True),
        "outname": attr.string(),
    },
    outputs = {"out": "%{name}.java"},
    implementation = _impl,

    std::vector<NamedAttribute> attrs = contraction->getAttrs();
    ArrayAttr fused_ops_attr = ArrayAttr::get(context, fused_ops);
    attrs.push_back(
        NamedAttribute(StringAttr::get(context, "fused_ops"), fused_ops_attr));
    // Epsilon is used only in fusions with the FusedBatchNorm op, so we zero it
    // here.
    Attribute epsilon = rewriter.getF32FloatAttr(0);
    attrs.push_back(

      "to be 0 but found 1");
}

TEST(NodeMatchers, Attrs) {
  Scope root = Scope::NewRootScope().ExitOnError();
  Output enter = ops::internal::Enter(
      root.WithOpName("enter"),
      ops::Placeholder(root.WithOpName("data"), DT_FLOAT), "frame_name",
      ops::internal::Enter::Attrs{}.IsConstant(true));
  EXPECT_THAT(enter.node(), NodeWith(Attr("is_constant", true)));

              value.add_int64_val(dim.size());
            }
          }
          AddNodeAttr("value", value, &const_def);
          for (auto const& attr : n->attrs()) {
            if (*attr.first.begin() == '_') {
              AddNodeAttr(attr.first, attr.second, &const_def);
            }
          }

          TF_ASSIGN_OR_RETURN(Node * const_node, graph->AddNode(const_def));

  const char* key = pair.getName().data();
  const auto attr = pair.getValue();

  if (mlir::isa<::mlir::IntegerAttr>(attr)) {
    fbb->Int(key, mlir::dyn_cast<mlir::IntegerAttr>(attr).getInt());
    return success();
  }

  if (mlir::isa<::mlir::FloatAttr>(attr)) {
    fbb->Double(key, mlir::dyn_cast<mlir::FloatAttr>(attr).getValueAsDouble());
    return success();
  }

        // Same type repeated "repeats" times.
        Attribute attr = attrs_[output_arg.type_attr()];
        if (!attr)
          return InvalidArgument("Missing attribute '", output_arg.type_attr(),
                                 "' required for output '", output_arg.name(),
                                 "'");
        TypedAttr type_attr = mlir::dyn_cast<TypedAttr>(attr);
        if (!type_attr)

# Attrs and Values

An [Attr] is a key-value pair. The Logger output methods accept Attrs as well as
alternating keys and values. The statement

	slog.Info("hello", slog.Int("count", 3))

behaves the same as

	slog.Info("hello", "count", 3)

There are convenience constructors for [Attr] such as [Int], [String], and [Bool]

  let description = [{
    The `attr` operation stores TF op's attribute, which doesn't support
    arithmetic operations.

    Example:

    ```mlir
    %1 = "tfr.constant"() { value: i32 } : () -> !tfr.attr
    %2 = "tfr.constant"() { value: [i32, f32] } : () -> !tfr.attr
    %3 = tfr.constant [i32, f32] -> !tfr.attr
    %4 = tfr.constant f32 -> !tfr.attr
    ```
  }];

  return absl::OkStatus();
}

Status Encapsulator::GetFunctionNameAttr(Node const* node, string* attr) const {
  AttrSlice attrs = node->attrs();
  attr->clear();
  for (const auto& node_attr : attrs) {
    if (node_attr.first == group_attribute_) {
      TF_RETURN_IF_ERROR(AttrValueHasType(node_attr.second, "string"));
      *attr = node_attr.second.s();
      break;
    }
  }
  return absl::OkStatus();
}