}
  TF_RETURN_IF_ERROR(scope.status());

  std::vector<Output> merge_outputs(num_loop_vars);
  for (size_t i = 0; i < num_loop_vars; ++i) {
    TF_RETURN_IF_ERROR(
        CreateMerge(scope, i, enter_outputs[i], &merge_outputs[i]));
  }

  Output cond_out;
  TF_RETURN_IF_ERROR(CreateCond(scope, cond, merge_outputs, &cond_out));
  if (cond_output != nullptr) *cond_output = cond_out;

}

// An implementation of TensorFlow bidrectional_dynamic_rnn with RNNCell.
table BidirectionalSequenceRNNOptions {
  time_major:bool;
  fused_activation_function:ActivationFunctionType;
  merge_outputs: bool;
  asymmetric_quantize_inputs:bool;
}

enum FullyConnectedOptionsWeightsFormat: byte {
  DEFAULT = 0,
  SHUFFLED4x16INT8 = 1,
}

}

// An implementation of TensorFlow bidrectional_dynamic_rnn with RNNCell.
table BidirectionalSequenceRNNOptions {
  time_major:bool;
  fused_activation_function:ActivationFunctionType;
  merge_outputs: bool;
  asymmetric_quantize_inputs:bool;
}

enum FullyConnectedOptionsWeightsFormat: byte {
  DEFAULT = 0,
  SHUFFLED4x16INT8 = 1,
}

}

// An implementation of TensorFlow bidrectional_dynamic_rnn with RNNCell.
table BidirectionalSequenceRNNOptions {
  time_major:bool;
  fused_activation_function:ActivationFunctionType;
  merge_outputs: bool;
  asymmetric_quantize_inputs:bool;
}

// LINT.IfChange
enum FullyConnectedOptionsWeightsFormat: byte {
  DEFAULT = 0,
  SHUFFLED4x16INT8 = 1,
}

  auto _fused_activation_function = _o->fused_activation_function;
  auto _merge_outputs = _o->merge_outputs;
  auto _asymmetric_quantize_inputs = _o->asymmetric_quantize_inputs;
  return tflite::CreateBidirectionalSequenceRNNOptions(
      _fbb,
      _time_major,
      _fused_activation_function,
      _merge_outputs,
      _asymmetric_quantize_inputs);
}

                                  Output(old_node, oidx), new_output);
      merged_output = merged_outputs[oidx] = xla_merge_op.output;
    }

    Node* dst = e->dst();
    int dst_idx = e->dst_input();

    s.graph()->RemoveEdge(e);
    s.graph()->AddEdge(merged_output.node(), merged_output.index(), dst,
                       dst_idx);
  }
}

%arg26, %arg27, %arg28, %arg29, %arg30, %arg31, %arg32, %arg33, %arg34, %arg35, %arg36, %arg37, %arg38, %arg39, %arg40, %arg41, %arg42, %arg43, %arg44, %arg45, %arg46, %arg47) <{cell_clip = 1.000000e+00 : f32, fused_activation_function = "NONE", merge_outputs = true, time_major = false}> : (tensor<?x?x?xf32>, tensor<?x?xf32>, tensor<?x?xf32>, tensor<?x?xf32>, tensor<?x?xf32>, tensor<?x?xf32>, tensor<?x?xf32>, tensor<?x?xf32>, tensor<?x?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>,...

    ConfinedAttr<DefaultValuedOptionalAttr<F32Attr, "0.0f">, [TFL_FloatNonNegative]>:$cell_clip,
    ConfinedAttr<DefaultValuedOptionalAttr<F32Attr, "0.0f">, [TFL_FloatNonNegative]>:$proj_clip,
    BoolAttr:$merge_outputs,
    BoolAttr:$time_major,
    // Used in post-training dynamic range quantization. If the value is true,
    // input activations are asymmetrically quantized.
    OptionalAttr<BoolAttr>:$asymmetric_quantize_inputs

  VerifyFDefEdges(fdef,
                  {{"feed1", "test_loop/Enter:0"},
                   {"test_loop/Enter:output:0", "test_loop/Merge:0"},
                   {"test_loop/Merge:output:0", "test_loop/Switch:0"},
                   {"test_loop/Switch:output_false:0", "test_loop/Exit:0"},
                   {"test_loop/Exit:output:0", "test_loop_exit"}},
                  {}, false);
}