// CHECK-SAME:       _xla_compile_device_type = "TPU"
    // CHECK:            "tf.Yield"(%[[FUNC_CALL]]#0, %[[FUNC_CALL]]#1)
    // CHECK:            "tf.Yield"(%[[ARG_11]], %[[ARG_12]])
    // CHECK:          "tf.Yield"(%[[HAS_VALUE]], %[[IF]]#0, %[[IF]]#1, %[[ARG_13]], %[[ARG_14]])

 = "args_7"}, %arg8: tensor<?x?xf32> {tf._user_specified_name = "args_8"}, %arg9: tensor<?x300xi32> {tf._user_specified_name = "args_9"}, %arg10: tensor<?x300xi32> {tf._user_specified_name = "args_10"}, %arg11: tensor<?x300xf32> {tf._user_specified_name = "args_11"}, %arg12: tensor<?x300xf32> {tf._user_specified_name = "args_12"}, %arg13: tensor<?x!tf_type.string> {tf._user_specified_name = "args_13"}, %arg14: tensor<?x!tf_type.string> {tf._user_specified_name = "args_14"}, %arg15: tensor<!tf_ty...

        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: _xla_compile_device_type = "GPU"
# 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"

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 {

node {
  name: "args_0"
  op: "_Arg"
  device: "/CPU:0"
  attr {
    key: "T"
    value {
      type: DT_FLOAT
    }
  }
  attr {
    key: "index"
    value {
      i: 0
    }
  }
}
node {
  name: "args_1"
  op: "_Arg"
  attr {
    key: "T"

  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 {

// 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.

  func.func @main(%arg0: tensor<*xf32> {tf.aliasing_output = 0 : i64}, %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 {

// accesses in between.

// CHECK-LABEL: func @interfering_accesses
// 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<32xf32>
// CHECK-SAME: %[[ARG_4:.*]]: tensor<*x!tf_type.resource<tensor<8xf32>>>
// CHECK-SAME: %[[ARG_5:.*]]: tensor<*x!tf_type.resource<tensor<2xf32>>>
// CHECK-SAME: %[[ARG_6:.*]]: tensor<2xf32>