func.return %0 : tensor<1xi32>
}

// CHECK-LABEL: func @main
func.func @main(%arg0: tensor<1xi32> {tf_saved_model.index_path = ["input"]}) -> (tensor<1xi32> {tf_saved_model.index_path = ["r"]})
  attributes {tf_saved_model.exported_names = ["main"]} {

  // CHECK-DAG: %[[INDICES_3:.*]] = "tf.Mul"(%[[INDICES_2]], %arg0) : (tensor<5xi32>, tensor<5xi32>) -> tensor<5xi32>
  // CHECK-DAG: %[[INDICES_4:.*]] = "tf.TensorScatterAdd"(%[[cst_3]], %0, %[[UPDATES]]) : (tensor<5xi32>, tensor<5x1xi32>, tensor<5xi32>) -> tensor<5xi32>
  // CHECK-DAG: %[[INDICES_5:.*]] = "tf.AddV2"(%[[INDICES_3]], %[[INDICES_4]]) : (tensor<5xi32>, tensor<5xi32>) -> tensor<5xi32>

    %4 = "tf.LayerNorm"(%a1, %a2, %a3, %a4) {_tfl_quant_trait = "fully_quantizable", device = ""} : (tensor<128x128xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xi32>) -> tensor<128x128xf32>
   "tfl.yield"(%4) : (tensor<128x128xf32>) -> ()
  }) {_tfl_quant_trait = "fully_quantizable", device = ""} : (tensor<128x128xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xi32>) -> tensor<128x128xf32>

    %4 = "tf.Const"() {value = dense<> : tensor<0xi32>} : () -> tensor<0xi32>
    %5 = "tf.Const"() {value = dense<2.500000e-01> : tensor<f32>} : () -> tensor<f32>
    %6 = "tf.Const"() {value = dense<0> : tensor<1xi32>} : () -> tensor<1xi32>
    %7 = "tf.Const"() {value = dense<[-1, 1001]> : tensor<2xi32>} : () -> tensor<2xi32>
    %8 = "tf.Const"() {value = dense<[1, 2]> : tensor<2xi32>} : () -> tensor<2xi32>

  // CHECK-NEXT: tf_mlrt.executeop([[y]]
  %v = "tf.AddV2"(%y, %z) {__op_key = 3: i32}: (tensor<1xi32>, tensor<1xi32>) -> tensor<1xi32>
  return %w, %u, %v : tensor<1xi32>, tensor<1xi32>, tensor<1xi32>
}

// -----

// Test node names are preserved.

// CHECK-LABEL: func @main
func.func @main(%x: tensor<i32>) -> tensor<i32> {
  // CHECK: tf_mlrt.executeop

    %1 = "tf.Sub"(%0, %input_zp) : (tensor<*xi32>, tensor<*xi32>) -> tensor<*xi32>

    // Use identity op to avoid the weight being constant-folded.
    %identity = "tf.Identity"(%filter) : (tensor<*xi8>) -> tensor<*xi8>
    %2 = "tf.Cast"(%identity) {Truncate = false} : (tensor<*xi8>) -> tensor<*xi32>
    %3 = "tf.Sub"(%2, %weight_zp) : (tensor<*xi32>, tensor<*xi32>) -> tensor<*xi32>

    %5 = "tf.MatMul"(%1, %3) {

  // CHECK-DAG: %[[IND_SLICE0_START:.*]] = "tf.Const"() <{value = dense<0> : tensor<1xi32>}> : () -> tensor<1xi32>
  // CHECK-DAG: %[[IND_SLICE0_SIZE:.*]] = "tf.Const"() <{value = dense<1> : tensor<1xi32>}> : () -> tensor<1xi32>
  // CHECK: %[[IND_SLICE0:.*]] = "tf.Slice"(%[[INDS]], %[[IND_SLICE0_START]], %[[IND_SLICE0_SIZE]]) : (tensor<2xi32>, tensor<1xi32>, tensor<1xi32>) -> tensor<1xi32>

    func.func @tpu_partitioned_call_func(%arg0: tensor<?xi32>) -> (tensor<?xi32>) {
      func.return %arg0 : tensor<?xi32>
    }

    func.func @main(%arg0: tensor<20xi32>, %arg1: tensor<?xi32>) -> tensor<*xi32> {
      %2 = "tf.TPUPartitionedCall"(%arg0, %arg1) {f = @tpu_partitioned_call_func} : (tensor<20xi32>, tensor<?xi32>) -> tensor<*xi32>
      func.return %2 : tensor<*xi32>
    }
  })";

  module attributes {tf.versions = {bad_consumers = [], min_consumer = 0 : i32, producer = 268 : i32}} {
    func.func @main() -> tensor<1xi32> {
      %0 = "tf.Const"() {value = dense<1000> : tensor<1xi32>} : () -> tensor<1xi32>
      func.return %0 : tensor<1xi32>
    }
  })";

absl::StatusOr<std::string> CompileMlirModule(bool compile_to_xla_hlo,
                                              const char* module_str) {

// CHECK-SAME: (tensor<1x2x2x3xi8>, tensor<2x2x3x2xi8>, tensor<2xi32>, tensor<f32>, tensor<i32>, tensor<2xf32>, tensor<2xi32>, tensor<2xf32>, tensor<2xi32>, tensor<f32>, tensor<i32>) -> tensor<*xf32>
// CHECK: return %[[conv_quant]]

// CHECK-LABEL: func private @quantized_conv2d_with_bias_and_relu6_float_output_fn_0