func.func @main(%arg0: tensor<*x!tf_type.resource>, %arg1: tensor<*x!tf_type.resource<tensor<3x3x1x32xf32>>>, %arg2: tensor<*xf32>, %arg3: tensor<2x4x6x8xi32>) -> (tensor<f32>, tensor<f32>)
attributes {tf.entry_function = {inputs = "args_0,args_1,args_2,args_3", outputs = "rets_0_RetVal,rets_1_RetVal"}} {
  %graph:2 = tf_executor.graph {

        type: DT_INT32
        type: DT_INT64
      }
    }
  }
}
node {
  name: "MultiDeviceIteratorGetNextFromShard"
  op: "MultiDeviceIteratorGetNextFromShard"
  input: "args_2"
  input: "args_3"
  input: "args_4"
  attr {
    key: "output_shapes"
    value {
      list {
        shape {
          dim {
            size: 5
          }
          dim {
            size: 40

# CHECK-SAME: allow_soft_placement
# CHECK-SAME: control_outputs = ""
# CHECK-SAME: inputs = "args_0,args_1,args_2,args_3"
# CHECK-SAME: outputs = "rets_0,rets_1"
# CHECK:          %[[ISLAND_0:.*]], %[[ISLAND_0_control:.*]] = tf_executor.island wraps "tf.Const"
# CHECK:          %[[ISLAND_1:.*]], %[[ISLAND_1_control:.*]] = tf_executor.island wraps "tf.Identity"(%[[ISLAND_0]])

    // CHECK-NEXT: "tf.MakeIterator"(%[[ARG_0]], %[[ANON_ITER]])
    // CHECK-NEXT: %[[COND:.*]] = "tf.Const"
    // CHECK-NEXT: %[[WHILE:[0-9]*]]:5 = "tf.WhileRegion"(%[[COND]], %[[ARG_1]], %[[ARG_2]], %[[ARG_3]], %[[ARG_4]])
    // CHECK-NEXT:   ^bb0(%[[ARG_5:.*]]: tensor<i1>, %[[ARG_6:.*]]: tensor<i64>, %[[ARG_7:.*]]: tensor<i32>, %[[ARG_8:.*]]: tensor<!tf_type.resource<tensor<64xf32>>>, %[[ARG_9:.*]]: tensor<!tf_type.resource<tensor<128xf32>>>)

%arg2: tensor<i32> {tf._user_specified_name = "args_2"}, %arg3: tensor<i32> {tf._user_specified_name = "args_3"}, %arg4: tensor<?x!tf_type.string> {tf._user_specified_name = "args_4"}, %arg5: tensor<?x!tf_type.string> {tf._user_specified_name = "args_5"}, %arg6: tensor<?x!tf_type.string> {tf._user_specified_name = "args_6"}, %arg7: tensor<?x?x512xf32> {tf._user_specified_name = "args_7"}, %arg8: tensor<?x?xf32> {tf._user_specified_name = "args_8"}, %arg9: tensor<?x300xi32> {tf._user_specified_name = "args_9"},...

// Start step 0
C_0 = cond(args_0)
N_0 = non_tpu(args_0)
if (C_0) {
   F_0 = forward(args_0, N_0)
   T_0 = core_tpu(args_0, N_0, F_0)
   // B_0 = backward() is not evaluated here.
}

args_1 = update_args(args_0, N_0, T_0)

// Start step 1
C_1 = cond(args_1)
N_1 = non_tpu(args_1)
if (C_1) {
   F_1 = forward(args_1, N_1)
   // T_1 = core_tpu() is not evaluated here.

	// FMADD <Sd>, <Sn>, <Sm>, <Sa>
	{0xffe08000, 0x1f000000, FMADD, instArgs{arg_Sd, arg_Sn, arg_Sm, arg_Sa}, nil},
	// FMADD <Dd>, <Dn>, <Dm>, <Da>
	{0xffe08000, 0x1f400000, FMADD, instArgs{arg_Dd, arg_Dn, arg_Dm, arg_Da}, nil},
	// FMAX <Sd>, <Sn>, <Sm>
	{0xffe0fc00, 0x1e204800, FMAX, instArgs{arg_Sd, arg_Sn, arg_Sm}, nil},
	// FMAX <Dd>, <Dn>, <Dm>

# Verify arg and ret devices are added as arg and ret attributes.

# CHECK-LABEL: func @main
# CHECK-SAME:  (%[[ARG_0:[a-z0-9]+]]: tensor<*xf32> {tf.device = "/CPU:0"}, %[[ARG_1:[a-z0-9]+]]: tensor<2x4x6x8xi32>) -> (tensor<*xf32>, tensor<*xi32> {tf.device = "/CPU:1"})

node {
  name: "args_0"
  op: "_Arg"
  device: "/CPU:0"
  attr {
    key: "T"
    value {
      type: DT_FLOAT
    }
  }
  attr {

  func.func @main(%arg0: tensor<*xf32> {tf.device = "/CPU:0"}, %arg1: tensor<2x4x6x8xi32>) -> (tensor<*xf32>, tensor<2x4x6x8xi32> {tf.device = "/CPU:1"})
  attributes  {tf.entry_function = {inputs = "args_0,args_1", outputs = "rets_0,rets_1"}} {
    %0:2 = tf_executor.graph {

			continue Search
		}

		var args Args
		for j, aop := range f.args {
			if aop == 0 {
				break
			}
			arg := decodeArg(aop, x)
			if arg == nil { // cannot decode argument
				continue Search
			}
			args[j] = arg
		}

		decoderCover[i] = true

		inst = Inst{
			Op:   op,
			Args: args,
			Enc:  x,
			Len:  4,
		}
		priority = f.priority