// CHECK-DAG: TPUReplicateMetadata
  // CHECK-DAG: TPUCompilationResult
  // CHECK-DAG: TPUReplicatedOutput
  // CHECK: return

  // core_tput should have TPU ops including both replicated inputs and outputs
  // CHECK: func.func private @_func_core_tpu
  // CHECK-DAG: TPUReplicatedInput
  // CHECK-DAG: TPUReplicateMetadata
  // CHECK-DAG: TPUCompilationResult
  // CHECK-DAG: TPUReplicatedOutput
  // CHECK: return

    "tf.TPUReplicateMetadata"() {_has_manual_control_dependencies = true, _replication_info = "repl_info", num_replicas = 1 : i64} : () -> ()
    // expected-error @+1 {{'tf.TPUCompilationResult' op is not part of the replication region "repl_info" vs "wrong_repl_info"}}
    %1 = "tf.TPUCompilationResult"() {_tpu_compilation_status = "wrong_repl_info"} : () -> tensor<!tf_type.string>
    return %arg0 : tensor<i1>
  }

func.func @serving_default(%arg0: tensor<1x3xf32>) -> (tensor<1x3xf32>) {
  %outputs  =  "tf.TPUCompilationResult"() {_tpu_compilation_status = "cluster", device = ""} : () -> tensor<!tf_type.string>

    %outputs_5, %control_6 = tf_executor.island(%control_4) wraps "tf.TPUCompilationResult"() {_tpu_compilation_status = "cluster", device = ""} : () -> tensor<!tf_type.string>

  %1 = "tf.TPUCompilationResult"() {_tpu_compilation_status = "cluster", device = ""} : () -> tensor<!tf_type.string>
  %2 = "tf.Transpose"(%0, %cst_0) {device = ""} : (tensor<1x3x4x3xbf16>, tensor<4xi32>) -> tensor<1x3x3x4xbf16>

  // all graphs should have control dependencies to enforce this but this is
  // currently not the case (see b/177478741).
  const llvm::SmallVector<std::string, 4> ops_to_preserve = {
      "tf.TPUReplicateMetadata", "tf.TPUCompilationResult",
      "tf.TPUReplicatedOutput"};
  bool strict_clusters =
      tensorflow::GetMlirCommonFlags()->tf_mlir_enable_strict_clusters;
  pm.addNestedPass<FuncOp>(

                               llvm::SetVector<Operation*>& merged_set,
                               OpType& found_op) {
  // Find the TPUReplicationMetadata or TPUCompilationResult ops which will be
  // cloned/inserted into each region. We add them to the merged_set so that
  // they're ignored when extracting the four main functions.
  found_op = nullptr;

      %outputs_25, %control_26 = tf_executor.island(%control_18) wraps "tf.TPUCompilationResult"() {_tpu_compilation_status = "cluster_sample_sequence", device = ""} : () -> tensor<!tf_type.string>

// returns `std::nullopt`.
std::optional<llvm::StringRef> GetTpuClusterName(Operation* op) {
  if (auto tpu_status = op->getAttrOfType<StringAttr>(kTpuStatusAttr)) {
    // Borrow cluster name from TPU status (for `TPUCompilationResult` op).
    return tpu_status.getValue();
  }
  auto device_type = op->getAttrOfType<StringAttr>(TF::kCompileDeviceTypeAttr);
  if (!device_type || device_type.getValue() != TF::kTpuDevice) {

    %compile_result = "tf.TPUCompilationResult"() {_tpu_compilation_status = "cluster0"} : () -> tensor<!tf_type.string>
    %compile_result2 = "tf.TPUCompilationResult"() {_tpu_compilation_status = "cluster0"} : () -> tensor<!tf_type.string>

    // CHECK-NOT: "tf.TPUCompilationResult"

    // CHECK: return %[[EXECUTE_OUTPUT]], %[[COMPILE_OUTPUT]]#0, %[[COMPILE_OUTPUT]]#0