}

  Exporter exporter(&configs, graph.get(), tf_dialect, &symbol_table);

  auto graph_op = llvm::cast<mlir::tf_executor::GraphOp>(block.front());

  // Set input and output names and increment the use counter for them to help
  // generate unique names.
  if (!output_names.empty()) {
    const int num_data_results = graph_op.getNumResults();
    const int64_t output_names_size = output_names.size();

  }
  return changed;
}

bool ShapeInference::InferShapeForNonTFDialectOperation(Operation* op) {
  if (auto graph_op = dyn_cast<tf_executor::GraphOp>(op)) {
    return ForceTypeForPassThroughOperands(graph_op.GetFetch(),
                                           graph_op.GetFetch().getFetches(),
                                           op->getResults());
  }

  options.client = client;
  options.flib_def = flib_def.get();
  tensorflow::XlaCompiler compiler(options);

  std::unique_ptr<tensorflow::Graph> graph_copy(
      new tensorflow::Graph(tensorflow::OpRegistry::Global()));
  tensorflow::CopyGraph(*graph, graph_copy.get());

  tensorflow::XlaCompiler::CompileOptions compile_options;
  tensorflow::XlaCompiler::CompilationResult result;

}

def ExecutorIslandCoarseningPass : Pass<"tf-executor-island-coarsening", "mlir::func::FuncOp"> {
  let summary = "Walks tf_executor::GraphOp and merges individual tf_executor::IslandOps.";
  let description = [{
    This pass performs whole graph analysis for a graph encapsulated into tf_executor::GraphOp.
    The analysis identifies all IslandOps within the graph which could be merged together.
    The goal is to merge as many islands as possible.