};

// Clustering policy for `test.OpA` and `test.OpB` operations;
class TestOpsClusteringPolicy : public ClusteringPolicy {
  LogicalResult MatchAndUpdateConstraints(
      Operation* op, const ValuesConstraintSet& results,
      ValuesConstraintSet& operands) const final {
    // Check if operation is `test.OpA` or `test.OpB`.
    bool is_op_a = op->getName().getStringRef() == "test.OpA";

	if err != nil {
		return false, err
	}

	// Handle large OPA responses when OPA URL is of
	// form http://localhost:8181/v1/data/httpapi/authz
	type opaResultAllow struct {
		Result struct {
			Allow bool `json:"allow"`
		} `json:"result"`
	}

	// Handle simpler OPA responses when OPA URL is of
	// form http://localhost:8181/v1/data/httpapi/authz/allow
	type opaResult struct {

  // CHECK-SAME: n = 2
  // CHECK:   tf_device.return
}

// -----

// CHECK-LABEL: func @replicate_with_multiple_operands
func.func @replicate_with_multiple_operands() {
  %0 = "tf.opA"() : () -> tensor<*xi1>
  %1 = "tf.opB"() : () -> tensor<*xi1>
  %2 = "tf.opC"() : () -> tensor<*xi1>
  %3 = "tf.opD"() : () -> tensor<*xi32>
  %4 = "tf.opE"() : () -> tensor<*xi32>
  %5 = "tf.opF"() : () -> tensor<*xi32>

		cfg.Compression.MimeTypes = strings.Split(compress.DefaultMimeTypes, config.ValueSeparator)
	case "30":
		// V30 -> V31
		cfg.OpenID = openid.Config{}
		cfg.Policy.OPA = opa.Args{
			URL:       &xnet.URL{},
			AuthToken: "",
		}
	case "31":
		// V31 -> V32
		cfg.Notify.NSQ = make(map[string]target.NSQArgs)
		cfg.Notify.NSQ["1"] = target.NSQArgs{}
	}

  // CHECK-NEXT:     "tf.OpA"(%[[REDUCED]]) : (tensor<f32>) -> ()
  tf_device.replicate {n = 4 : i32} {
    "tf_device.cluster"() ({
      "tf.OpA"(%arg0) : (tensor<f32>) -> ()
      tf_device.return

    tf_executor.graph {
      // expected-error@+1 {{functions must be of a single Graph with single op Islands: tf_executor.island must perfectly wrap a single op}}
      %island = tf_executor.island {
        "tf.OpA"() : () -> ()
        "tf.OpB"() : () -> ()
        tf_executor.yield
      }
      tf_executor.fetch
    }
    func.return
  }
}

// -----

    during TPURewritePass.

    For example, the tf.OpA with the `mini_batch_splits` attribute will be
    moved after _TPUCompileMlir and the first input will use the
    _TPUCompileMlir program output:

    ```mlir
    "tf_device.launch"() ({
     %cst_0 = "tf.Const"() {value = dense<""> : tensor<1x!tf_type.string>} : () -> tensor<1x!tf_type.string>
     "tf.OpA"(%cst_0) { mini_batch_splits = ""} : (tensor<1x!tf_type.string>) -> ()

// guaranteed to dominate all live-in values. On the other hand, it is still
// possible that live-out values don't dominate the region. For example:
//
// ```
// %0 = "tf.OpA"()
// %1 = "tf.OpB"(%0)
// %2 = "tf.OpC"(%0)
// ```
//
// Assuming `tf.OpA` and `tf.OpC` are clustered together, the region will be
// inserted right after `tf.OpC`. The live-out `%0`, however, is used by

//
// tf_executor.island wraps "tf.OpA"
// tf_executor.island {
//  "tf_device.replicate" {n = 2} {
//    "tf.OpB"
//    "tf_device.parallel_execute"() ({
//      "tf.OpC"
//    }, {
//      "tf.OpD"
//    })
//  }
//
// The above IR will be flattened after `ReplicateToIslandPass` and
// `ParallelExecuteToIslandsPass` as follows:
//
// tf_executor.island wraps "tf.OpA"

    For example, the following:

    ```mlir
    func @single_op_launch() {
      %a = "tf.opA"() {device = "CPU:0"} : () -> tensor<i1>
      return %a
    }
    ```

    will be transformed into:

    ```mlir
    func @single_op_launch() {
      %1 = tf_device.launch() ( {
        %a = "tf.opA"() : () -> tensor<i1>
        tf_device.return %a
      }) {device = "CPU:0"} : () -> tensor<i1>