{"Toutputs", absl::Span<const DataType>({DT_FLOAT})},
            {"ancestors", absl::Span<const string>({})},
            {"key", "host_compute_channel_F1_F1_O1"},
            {"send_key", ""},
            {"recv_key", ""},
            {"shape_inference_graph", shape_inference_graph},
            {"tpu_core", 0},
            {"cost_estimate_ns", 1000000},

    // CHECK:            "tf._XlaHostComputeMlir"(%[[A_OUTPUT]])
    // CHECK-SAME:       send_key = "host_compute_channel_0_args"
    // CHECK:            %[[C_OUTPUT:[0-9]*]] = "tf.C"
    // CHECK:            "tf._XlaHostComputeMlir"(%[[C_OUTPUT]])
    // CHECK-SAME:       send_key = "host_compute_channel_1_args"
    %1:2 = tf_device.replicate([%0, %arg0] as %ri_0: tensor<2xi32>) {n = 2 : i32} {

// -----

func.func @testXlaHostComputeMlir(%arg0: tensor<2xf32>) -> () {
  "tf._XlaHostComputeMlir"(%arg0) {send_key="", recv_key="", host_mlir_module=""} : (tensor<2xf32>) -> ()
  func.return
}

// -----

func.func @testXlaHostComputeMlir(%arg0: tensor<2xf32>) -> () {

    // CHECK: "tf._XlaHostComputeMlir"
    // CHECK-SAME: -> tensor<*xf32>
    %0 = "tf._XlaHostComputeMlir"(%arg0) {recv_key = "host_compute_channel_recv", send_key = "host_compute_channel_send", tpu_core = 0, host_mlir_module = ""} : (tensor<2xf32>) -> tensor<*xf32>
    func.return %0 : tensor<*xf32>
  }

  // CHECK-LABEL: func @xla_host_compute_mlir_shape_inferred

    %0 = "tf.A"() : () -> tensor<?xi32>
    "tf._XlaHostComputeMlir"(%0) {host_mlir_module = "", recv_key = "host_compute_channel_0_retvals", send_key = "host_compute_channel_0_args"} : (tensor<?xi32>) -> ()
    %1 = "tf.C"() : () -> tensor<?xi32>
    return %1 : tensor<?xi32>
  }
}

// -----

// The following xla.OpSharding is used: