// CHECK: entry function `main` must be present

func.func @const() {
  tf_executor.graph {
    %0:2 = tf_executor.island wraps "tf.Const"() {device = "TPU:0", name = "const", dtype = "tfdtype$DT_INT32", value = dense<[1, 2]> : tensor<2xi32>} : () -> tensor<2xi32>
    tf_executor.fetch
  }
  func.return

func.func @controls_per_replica() {
  tf_executor.graph {
    %1 = tf_executor.ControlTrigger {}
    %2 = tf_executor.ControlTrigger {}
    %3 = tf_executor.island(%1, %2) {
      tf_device.replicate {n = 2 : i32} {
        tf_device.return
      }
      tf_executor.yield
    }
    tf_executor.fetch
  }
  func.return
}

// CHECK: %[[CT_0:.*]] = tf_executor.ControlTrigger

    tf_executor.fetch
  }
  func.return
}

// -----

// Check that we don't report a dependency between two stateful ops.
func.func @two_stateful_ops() -> () {
  tf_executor.graph {
    %island1 = tf_executor.island {
      "tf.OpA"() {is_stateless = false}: () -> ()
      tf_executor.yield
    }
    %island2 = tf_executor.island(%island1) {
      "tf.OpB"() {is_stateless = false}: () -> ()
      tf_executor.yield
    }

    %2, %3 = tf_executor.island(%1) wraps "tf.TPUCompilationResult"() : () -> tensor<!tf_type.string>
    tf_executor.fetch %0 : !tf_executor.control
  }
  func.return
}

// CHECK-LABEL: func @preserve_unreachable_tpu_replicated_input
func.func @preserve_unreachable_tpu_replicated_input(%arg0: tensor<i1>) {
  tf_executor.graph {
    %0 = tf_executor.ControlTrigger {}
    // CHECK: "tf.NoOp"
    %1 = tf_executor.island wraps "tf.NoOp"() : () -> ()

    tf_executor.fetch %outputs_17, %control_14, %control_25 : tensor<20x4096xf32>, !tf_executor.control, !tf_executor.control
  }
  return %0 : tensor<20x4096xf32>
}
// CHECK-LABEL: @merge_duplicate_variable
// CHECK: %[[OUT_0:.*]], %[[CTL_0:.*]] = tf_executor.island wraps "tf.VarHandleOp"()
// CHECK: %[[OUT_1:.*]], %[[CTL_1:.*]] = tf_executor.island wraps "tf.RestoreV2"

// CHECK-DAG: %[[CTL_2:.*]] = tf_executor.island(%[[CTL_0]]) wraps "tf.NoOp"()
// CHECK-DAG: %[[CTL_3:.*]] = tf_executor.island(%[[CTL_1]]) wraps "tf.NoOp"()

// CHECK: tf_executor.fetch
// CHECK-SAME: !tf_executor.control, !tf_executor.control
// CHECK-NEXT: }
// CHECK-NEXT: return

// Checks that the location for the init op is properly set.
// CHECK-LOC-LABEL: func.func @main

func.func @unsupported_op() {
  tf_executor.graph {
    // expected-error@+1 {{'tf_executor.ControlTrigger' op is not supported for lifting out of tf_executor.graph, expected tf_executor.island}}
    %control = tf_executor.ControlTrigger {}
    tf_executor.fetch
  }
  func.return
}

// -----

// CHECK-LABEL: func @empty_graph
// CHECK-NEXT: return

      tf_executor.yield
    }
    tf_executor.fetch
  }
  func.return
}

// CHECK: functions must be of a single Graph with single op Islands: tf_executor.island must perfectly wrap a single op

// -----

func.func @main() {
  tf_executor.graph {
    %0 = tf_executor.island {
      %1 = "tf.Const"() {value = dense<1> : tensor<i32>} : () -> tensor<i32>
      tf_executor.yield
    }
    tf_executor.fetch
  }

# `tf._resource_arg_unique_id` argument attribute of the function
# @test_func_name0.

# CHECK:  func @main
# CHECK:    tf_executor.graph
# CHECK:      "tf.VarHandleOp"()
# CHECK:      "tf.LegacyCall"
# CHECK-SAME:   <{_disable_call_shape_inference = true, f = @test_func_name0}> {device = ""}
# CHECK:      tf_executor.fetch
# CHECK:    return
# CHECK:  func private @test_func_name0
# CHECK-SAME:   tf._resource_arg_unique_id = 0

```
### `-tf-executor-to-functional-conversion`

_Lifts tf_executor.island inner ops from a tf_executor.graph_

This pass converts tf_executor.graphs consisting of only tf_executor.islands and
a tf_executor.fetch into a sea of nodes consisting of TensorFlow Dialect ops by
lifting such ops out of a tf_executor.graph's tf_executor.islands. If V1 control
flow ops are present in a tf_executor.graph, an error will be returned.