func.func @simple_independent_chains(%arg0: !tf_res, %arg1: !tf_res, %arg2: tensor<f32>) {
  // CHECK: tf_executor.graph {
  tf_executor.graph {
    // CHECK: %[[A_CONTROL:.*]] = tf_executor.island wraps "tf.OpA"() : () -> ()
    %control_A = tf_executor.island wraps "tf.OpA"() : () -> ()
    // CHECK: %[[CHAIN_CONSTANT:.*]], %{{.*}} = tf_executor.island wraps "tf.Const"() <{value = dense<1> : tensor<i32>}> : () -> tensor<i32>

    %control = tf_executor.island() wraps "tf.TPUReplicateMetadata"() {_tpu_replicate = "cluster", device = "/device:TPU:0", num_replicas = 2, topology = "topology"} : () -> ()
    %ri, %c0 = tf_executor.island wraps "tf.TPUReplicatedInput"(%arg0, %arg1) {index = 1 : i64, is_mirrored_variable = false, is_packed = false} : (tensor<i32>, tensor<i32>) -> tensor<i32>

// CHECK: TPUCompile
// CHECK: TPUExecute
// CHECK-NOT: func @_func
    tf_executor.graph {
      %outputs, %control = tf_executor.island wraps "arith.constant"() {value = dense<2.000000e+00> : tensor<f32>} : () -> tensor<f32>
      %outputs_0, %control_1 = tf_executor.island wraps "arith.constant"() {value = dense<3.000000e+00> : tensor<f32>} : () -> tensor<f32>

  func.func @main() {
    tf_executor.graph {
      %0 = tf_executor.island wraps "tf.PartitionedCall"() {Tin = [], Tout = [], config = "", config_proto = "", device = "", executor_type = "", f = @foo, name = "Call_foo"} : () -> ()
      tf_executor.fetch
    }
    func.return
  }
  func.func @foo() {
    tf_executor.graph {
      %0:2 = tf_executor.island {

func.func @multiple_func_body_ops() {
  tf_executor.graph {
    tf_executor.island {
      "tf.NoOp"() : () -> ()
      tf_executor.yield
    }
    tf_executor.fetch
  }
  tf_executor.graph {
    tf_executor.island {
      "tf.NoOp"() : () -> ()
      tf_executor.yield
    }
    tf_executor.fetch
  }
  func.return
}

// -----

}

// -----

// Check that tf_executor.graph can't be nested directly in a tf_executor.graph.
func.func @nested_graph() {
  tf_executor.graph {
    tf_executor.graph {}
// expected-error@-1 {{'tf_executor.graph' op unallowed directly inside another tf_executor.graph}}
  }
  func.return
}

// -----

// Check that a tf_executor.fetch is terminating a tf_executor.graph (custom parser)

      %3 = "tf_device.launch_func"(%2) {device = "/device:test_device:0", func = @_func} : (tensor<?xf32>) -> tensor<?xf32>

      // CHECK: tf_executor.yield %[[CALL_OUTPUT]]
      tf_executor.yield %3 : tensor<?xf32>
    }
    tf_executor.fetch %1#0 : tensor<?xf32>
  }
  func.return %0 : tensor<?xf32>
}

func.func @_func(%arg0: tensor<?xf32>) -> tensor<?xf32> {
  func.return %arg0 : tensor<?xf32>
}

        : (tensor<i1>, !tf_res,
           !tf_res) -> ()
      tf_executor.yield
    }
    tf_executor.fetch %island : !tf_executor.control
  }
  func.return
}

// CHECK-LABEL: func @if_then
func.func @if_then(
  %arg0: !tf_res,
  %arg1: !tf_res) {
  tf_executor.graph {
    // CHECK: tf_executor.island
    %island = tf_executor.island {
      // CHECK-NEXT: "tf.Identity"

#include "tensorflow/compiler/mlir/tensorflow/ir/tf_executor.h"

#define DEBUG_TYPE "tf-functional-to-executor"

namespace mlir {

namespace {
// This pass converts mlir functions consisting of mlir ops into a tf_executor
// dialect as a single island.
// Result like so:
//   func @my_fn(%argi...) -> (result_t) {
//     %results:[[n_args]] = tf_executor.graph {

    %5:2 = tf_executor.island wraps "tf.Identity"(%3#1) : (tensor<i64>) -> tensor<i64>
    %6 = tf_executor.ControlTrigger %4#1, %5#1
    tf_executor.fetch %6 : !tf_executor.control
  }
  func.return
}

// Tests tf_executor.NextIteration.Source/tf_executor.NextIteration.Sink has a
// device when an intermediate op in its loop has a device.