*   <li>Each other step is derived from one or more input steps. At each step, zero or more objects
 *       can be captured for later closing.
 *   <li>There is one last step (the root of the tree), from which you can extract the final result
 *       of the computation. After that result is available (or the computation fails), all objects
 *       captured by any of the steps in the pipeline are closed.
 * </ol>
 *
 * <h3>Starting a pipeline</h3>
 *

 *   <li>Each other step is derived from one or more input steps. At each step, zero or more objects
 *       can be captured for later closing.
 *   <li>There is one last step (the root of the tree), from which you can extract the final result
 *       of the computation. After that result is available (or the computation fails), all objects
 *       captured by any of the steps in the pipeline are closed.
 * </ol>
 *
 * <h3>Starting a pipeline</h3>
 *

}

def ClusterConstantSinkingPass : Pass<"tf-device-constant-sinking", "mlir::func::FuncOp"> {
  let summary = "Sinks constants implicitly captured in a tf_device.cluster region.";

  let description = [{
    This pass sinks implicitly captured constants (`tf.Const` ops) used by and into
    a `tf_device.cluster` region. Performing this prior to outlining will reduce the
    number of arguments of the outlined function.

	Subsystem MetricSubsystem `json:"Subsystem"`
	Name      MetricName      `json:"MetricName"`
	Help      string          `json:"Help"`
	Type      MetricTypeV2    `json:"Type"`
}

// MetricV2 captures the details for a metric
type MetricV2 struct {
	Description          MetricDescription `json:"Description"`
	StaticLabels         map[string]string `json:"StaticLabels"`

  // state, Y is the number of tensor components that represent the input
  // elements, and Z is the number of tensor components that represent any
  // captured arguments. Y is determined by the output_shapes of an op that
  // defines the first operand of `op`.

  // TODO(jpienaar): Avoid this lookup.
  auto module = op->getParentOfType<ModuleOp>();

    // For this test case the first 12 inputs to the function f below will be
    // from ParallelMapDataset, while the remaining one is a captured input.
    %75 = "tf.MapDataset"(%74, %arg3) {
      device = "/job:localhost/replica:0/task:0/device:CPU:0",
      f = @__inference_Dataset_map_lambda_96150,
      f._tf_data_function = true,

			expected: []string{ipv6},
			protocol: "udp",
		},
		{
			// We should only capture traffic to servers in /etc/resolv.conf nameservers
			// This checks we do not capture traffic to other servers.
			// This is important for cases like app -> istio dns server -> dnsmasq -> upstream
			// If we captured all DNS traffic, we would loop dnsmasq traffic back to our server.
			name:     "tcp localhost server",

				// Note that this test simulates a net-new pod being discovered during HandlePodCleanups that was not
				// delivered to the pod worker via HandlePodAdditions - there is no *known* scenario that can happen, but
				// we want to capture it in the metric. The more likely scenario is that a static pod with a predefined
				// UID is updated, which causes pod config to deliver DELETE -> ADD while the old pod is still shutting

					},
				},
			},
		),
		WithFrameworkOutOfTreeRegistry(frameworkruntime.Registry{
			"FakeNodeSelector": newFakeNodeSelector,
		}),
	)
	if err != nil {
		t.Fatal(err)
	}

	// Capture the bindings and events' controllers.
	var wg sync.WaitGroup
	wg.Add(2 * len(pods))
	bindings := make(map[string]string)

	Left   int32
	Top    int32
	Right  int32
	Bottom int32
}

type GUIThreadInfo struct {
	Size        uint32
	Flags       uint32
	Active      HWND
	Focus       HWND
	Capture     HWND
	MenuOwner   HWND
	MoveSize    HWND
	CaretHandle HWND
	CaretRect   Rect
}

const (
	DWMWA_NCRENDERING_ENABLED            = 1
	DWMWA_NCRENDERING_POLICY             = 2