// CHECK: %[[WHILE_OUT:.*]]:6, %[[WHILE_CONTROL:.*]] = tf_executor.island(%[[A_CONTROL]]) wraps "tf.While"(%[[ARG_0]], %[[ARG_1]], %[[ARG_2]], %[[ARG_2]], %[[CHAIN_CONSTANT]], %[[CHAIN_CONSTANT]])

// CHECK-LABEL: func @merge_same_device_variables
// CHECK-SAME: %[[ARG_0:.*]]: tensor<*x!tf_type.resource<tensor<32xf32>>>
// CHECK-SAME: %[[ARG_1:.*]]: tensor<*x!tf_type.resource<tensor<64xf32>>>
// CHECK-SAME: %[[ARG_2:.*]]: tensor<*x!tf_type.resource<tensor<16xf32>>>
func.func @merge_same_device_variables(

# CHECK: func {{@[a-zA-Z_0-9]+}}(
# CHECK-SAME: %[[ARG_1:.*]]: tensor<3x1xf32> {tf_saved_model.index_path = ["x"]}
# CHECK-SAME: %[[ARG_2:.*]]: tensor<!tf_type.resource<tensor<1x3xf32>>> {tf_saved_model.bound_input = @y}
# CHECK-SAME: -> (tensor<3x3xf32> {tf_saved_model.index_path = ["r"]})
# CHECK-SAME: attributes {{.*}} tf_saved_model.exported_names = ["key"]

// CHECK-LABEL: func @return_multiple_operands
// CHECK-SAME: (%[[ARG_0:[a-z0-9]*]]: tensor<*xf32>, %[[ARG_1:[a-z0-9]*]]: tensor<*xi32>)
func.func @return_multiple_operands(%arg_0: tensor<*xf32>, %arg_1: tensor<*xi32>) {
  %result:2 = "tf_device.launch"() ({
// CHECK:   tf_device.return %[[ARG_0]], %[[ARG_1]] : tensor<*xf32>, tensor<*xi32>
    tf_device.return %arg_0, %arg_1 : tensor<*xf32>, tensor<*xi32>
  }) {device = "device"} : () -> (tensor<*xf32>, tensor<?xi32>)

// CHECK-DAG:  [%[[OP_B]], %[[ARG_1]]] as %[[RI_1:[a-z0-9]*]]: tensor<i32>
// CHECK-SAME: n = 2 : i32
// CHECK-NEXT:   %[[CLUSTER:[0-9]*]]:2 = "tf_device.cluster"() ({
// CHECK:          %[[OP_D:[0-9]*]] = "tf.opD"(%[[RI_0]], %[[RI_1]], %[[ARG_2]], %[[OP_C]])
// CHECK:          %[[OP_E:[0-9]*]] = "tf.opE"(%[[OP_D]], %[[RI_0]], %[[RI_1]], %[[ARG_2]], %[[OP_C]])
// CHECK:          tf_device.return %[[OP_D]], %[[OP_E]]

	{0x0fe00000, 0x02800000, 2, ADD_EQ, 0x14011c04, instArgs{arg_R_12, arg_R_16, arg_const}},                      // ADD{S}<c> <Rd>,<Rn>,#<const> cond:4|0|0|1|0|1|0|0|S|Rn:4|Rd:4|imm12:12

    value {
      type: DT_FLOAT
    }
  }
}
node {
  name: "args_0"
  op: "_Arg"
  attr {
    key: "T"
    value {
      type: DT_RESOURCE
    }
  }
  attr {
    key: "index"
    value {
      i: 0
    }
  }
}
node {
  name: "spc1"
  op: "StatefulPartitionedCall"
  input: "input0"
  input: "args_0"
  attr {
    key: "Tin"
    value {
      list {

    key: "parallel_copy"
    value {
      b: false
    }
  }
}
library {
  function {
    signature {
      name: "__inference_Dataset_map_<lambda>_8"
      input_arg {
        name: "args_0"
        type: DT_INT32
      }
      output_arg {
        name: "identity"
        type: DT_INT32
      }
    }
    node_def {
      name: "mul/y"
      op: "Const"
      attr {

# CHECK:           %[[ADD:.*]], %{{.*}} = tf_executor.island wraps "tf.Add"(%[[ARG_0]], %[[ARG_1]])
# CHECK:           tf_executor.fetch %[[ARG_1]], %[[ARG_0]], %[[ADD]]

# Test when non zero index output tensors are feeds, remaining ops that are
# unreachable are pruned if pruning is enabled.
#
# PRUNE-LABEL: func @main
# PRUNE-SAME:  (%[[ARG_0:.*]]: tensor<f32>, %[[ARG_1:.*]]: tensor<f32>) -> (tensor<f32>, tensor<f32>, tensor<*xf32>)

        assertThat(arg2.getRawClass(), is(sameClass(Set[].class)));

        final ParameterizedClassDesc[] args2 = arg2.getArguments();
        assertThat(args2.length, is(1));

        final ParameterizedClassDesc arg2_1 = args2[0];
        assertThat(arg2_1.getRawClass(), is(sameClass(Integer.class)));
    }

    /**
     * @throws Exception
     */
    @Test