//
  // Two ops with the same groups and different branches are considered
  // parallel so are not made dependent. For example if `OpA` has parallel_ids
  //   `{group0:branch0, group1:branch0}`
  // and `OpB` has parallel_ids
  //   `{group0:branch1, graph1:branch0}`
  // then `OpA` and `OpB` are executed in parallel because `group0` is common
  // with a different branch.
  //

  %1 = "tf.ReadVariableOp"(%0) : (tensor<!tf_type.resource<tensor<i32>>>) -> tensor<i32>
  // CHECK:      [[COMPUTATION:%.+]] = "tf_device.parallel_execute"()
  // CHECK:      "tf.OpA"([[READ0]])
  %2 = "tf_device.parallel_execute"() ({
    "tf_device.launch"() ({
      "tf.OpA"(%1) : (tensor<i32>) -> ()
      tf_device.return
    }) {device = "TPU_REPLICATED_HOST_0"} : () -> ()
    tf_device.return
  }, {

  "tf_device.parallel_execute"() ({
// expected-error@-1 {{'tf_device.parallel_execute' op number of output types (3) must match the total number of outputs from all regions (2).}}
    %0 = "tf.opA"() : () -> (tensor<*xi1>)
    %1 = "tf.opB"() : () -> (tensor<*xi32>)
    tf_device.return %0, %1 : tensor<*xi1>, tensor<*xi32>
  },
  {
    %2 = "tf.opC"() : () -> (tensor<*xi1>)
    tf_device.return

      %outputs:4, %control = tf_executor.island {
       "tf.TPUReplicateMetadata"() {_xla_compile_device_type = "TPU", _replication_info = "cluster", device = "device", num_replicas = 1, topology = "topology"} : () -> ()
        %1 = "tf.opA"(%arg0) {_xla_compile_device_type = "TPU", _replication_info = "cluster"} : (tensor<i1>) -> tensor<i32>

      %output, %control = tf_executor.island {
       "tf.TPUReplicateMetadata"() {_xla_compile_device_type = "TPU", _replication_info = "cluster", device = "device", num_replicas = 1, topology = "topology"} : () -> ()
        %index = "tf.opA"(%arg0) {_xla_compile_device_type = "TPU", _replication_info = "cluster"} : (tensor<i1>) -> tensor<i32>
        %input = "tf.opB"(%arg0) {_xla_compile_device_type = "TPU", _replication_info = "cluster"} : (tensor<i1>) -> tensor<i32>

	case config.PolicyOPASubSys:
		// In case legacy OPA config is being set, we treat it as if the
		// AuthZPlugin is being set.
		subSys = config.PolicyPluginSubSys
		fallthrough
	case config.PolicyPluginSubSys:
		if ppargs, err := polplugin.LookupConfig(s, GetDefaultConnSettings(), xhttp.DrainBody); err != nil {
			return err
		} else if ppargs.URL == nil {
			// Check if legacy opa is configured.

      "tf.OpA"(%0) : (tensor<f32>) -> ()
    }
    ```

    will be transformed into:

    ``mlir
    %0 = "tf.ReadVariableOp"(%arg0) : (tensor<*x!tf_type.resource<tensor<f32>>>) -> tensor<f32>
    tf_device.replicate {n = 2 : i32} {
      "tf.OpA"(%0) : (tensor<f32>) -> ()
    }
    ```

      "tf.TPUReplicateMetadata"() {_xla_compile_device_type = "TPU", _replication_info = "cluster", device = "device", num_replicas = 1 : i64, topology = "topology"} : () -> ()
      %0 = "tf.opA"(%arg0) {_xla_compile_device_type = "TPU", _replication_info = "cluster"} : (tensor<i1>) -> tensor<i32>

// `tf_device.parallel_execute`.
//
// For example, the following:
//
//  %0 = tf_executor.island {
//    tf_executor.yield
//  }
//  %1:2 = tf_executor.island {
//    %2 = "tf.opA"(%arg0) : (tensor<i1>) -> tensor<i1>
//      tf_executor.yield %2 : tensor<i1>
//  }
//  %3:2 = tf_executor.island(%0) {
//    %4 = "tf_device.parallel_execute"() ({
//      %5 = "tf.opB"() : () -> tensor<i1>

//               devices = {DEVICE_ALIAS_0 = ["/DEVICE:0", "/DEVICE:1"],
//                          DEVICE_ALIAS_1 = ["/DEVICE:2", "/DEVICE:3"]}} {
//     %a = "tf_device.launch"() ({
//       %2 = "tf.opA"(%ri) : (tensor<i1>) -> tensor<i1>
//       tf_device.return %2 : tensor<i1>
//     }) {device = "DEVICE_ALIAS_0"} : () -> tensor<i1>
//     %b = "tf_device.launch"() ({
//       %3 = "tf.opB"(%a) : (tensor<i1>) -> tensor<i1>