// the argument indices.
  int variable_start_index = num_args;
  for (int i = 0; i < num_args; ++i) {
    int index;
    TF_RETURN_IF_ERROR(GetIndexAttr(*args[i], num_args, &index));
    if (args[i]->output_type(0) == DT_RESOURCE &&
        variable_start_index == num_args) {
      variable_start_index = i;
    }
    (*input_permutation)[index] = i;

         { dtype = i32, value = dense<[2]> : tensor<1xi32> } num_args(0)
  %ch2 = tfrt_fallback_async.createop(%ch1) key(1) device("/CPU:0")
         "tf.TestAsyncTfrtAsyncThread"() {T = i32} num_args(1)
  %ch3 = tfrt_fallback_async.createop(%ch2) key(2) device("/CPU:0")
         "tf.TestPrintThreadName"() num_args(0)

  %ch4, %0 = tfrt_fallback_async.executeop.seq(%ch3) key(0) cost(100)

          parser, builder, result, builder.getType<fallback::TFTensorType>(),
          parse_options)))
    return mlir::failure();

  mlir::IntegerAttr num_args;
  if (parser.parseKeyword("num_args") || parser.parseLParen() ||
      parser.parseAttribute(num_args, "num_args", result.attributes) ||
      parser.parseRParen())
    return mlir::failure();

  return mlir::success();
}

// RUN: tf-tfrt-opt %s -inline | FileCheck %s

func.func @_tfrt_fallback_init(%arg0: !tfrt.chain) -> !tfrt.chain {
  %0 = tfrt_fallback_async.createop(%arg0) key(0) device("/device:CPU:0") "tf.Less"() {T = i32} num_args(2)
  tfrt.return %0 : !tfrt.chain
}

func.func @callee(%ch: !tfrt.chain, %arg: !tfrt_fallback.tf_tensor) -> (!tfrt.chain, !tfrt_fallback.tf_tensor) {
  %const = tfrt_fallback_async.const_dense_tensor dense<9> : tensor<i32> {_tfrt_cost = 1 : i64}

func.func @createop(%in_ch: !tfrt.chain) -> !tfrt.chain {

  // CHECK: [[ch:%.*]] = tfrt_fallback_async.createop(%{{.*}}) key(100) device("cpu") "tf.AddV2"() {T = i32} num_args(2)
  %out_ch = tfrt_fallback_async.createop(%in_ch) key(100) device("cpu") "tf.AddV2"() {T = i32} num_args(2)

  // CHECK: tfrt.return [[ch]]
  tfrt.return %out_ch: !tfrt.chain
}

// CHECK-LABEL: func @fallback_resource

    Example:
      %out_ch = tfrt_fallback_async.createop(%in_ch) key(0) device("/CPU:0")
        "some.op"() {attr1 = value, attr2 = value2} {f_attr1 = "fn1", f_attr2 = "fn2"} num_args(1)
  }];

  let arguments = (ins
    TFRT_ChainType:$in_ch,
    I64Attr:$num_args,
    StrAttr:$device,
    ArrayAttr:$op_attrs,
    // TODO(b/173025975): consider using DictionaryAttr after we support
    // BEF conversion for this type.

  // CHECK: %[[VAL_0:.*]] = "tf._FusedConv2D"(%arg2, %arg1, %arg0) <{data_format = "NHWC", dilations = [1, 1, 1, 1], epsilon = 0.000000e+00 : f32, explicit_paddings = [], fused_ops = ["BiasAdd"], num_args = 1 : i64, operandSegmentSizes = array<i32: 1, 1, 1, 0>, padding = "SAME", strides = [1, 1, 1, 1], use_cudnn_on_gpu = true}> {TArgs = [f32]} : (tensor<8x32x32x3xf32>, tensor<1x1x3x128xf32>, tensor<128xf32>) -> tensor<*xf32>

// clang-format on

namespace tensorflow {
namespace tfcompile {
namespace {

void zero_buffers(XlaCompiledCpuFunction* computation) {
  for (int i = 0; i < computation->num_args(); ++i) {
    memset(computation->arg_data(i), 0, computation->arg_size(i));
  }
}

// Trivial test that runs the generated function to ensure it doesn't crash.
TEST(TEST_NAME, NoCrash) {

// CHECK-SAME: sparse_types = [!corert.string, i64]}
// CHECK-SAME: num_args(7)

// CHECK: tfrt_fallback_async.createop(%0) key(1) device("/device:CPU:0") "tf.ReadVariableOp"() {dtype = f32} num_args(1)

// CHECK: tfrt_fallback_async.createop(%1) key(2) device("/device:CPU:0") "tf.MatMul"() {T = f32, transpose_a = false, transpose_b = false} num_args(2)

// CHECK-LABEL: func @main
// CHECK-SAME: {{.*}} !tfrt.chain

                     const CompileResult& compile_result, string* methods) {
  const int num_args = ps.parameters_size();
  // feed_size() + variable_size() is the maximum number of args as an
  // implementation may not create an argument for an unused variable.
  if (config.feed_size() + config.variable_size() < num_args) {
    return errors::InvalidArgument(