// Move child processes of the ParallelExecute that do not change. These are all
// children except for the child with the ClusterFunc.
// Returns the index of the child with the ClusterFunc.
int MovePreservedParallelExecuteChildren(
    int num_cores_per_replica,
    llvm::SmallVector<mlir::Type, 8>& concatenated_output_types,
    mlir::OpBuilder* builder, mlir::tf_device::ClusterFuncOp cluster_func,

                 << " is not empty.";
      return mlir::failure();
    }
    old_parallel_execute->erase();
  }
  return mlir::success();
}

int MovePreservedParallelExecuteChildren(
    int num_cores_per_replica,
    llvm::SmallVector<mlir::Type, 8>& concatenated_output_types,
    mlir::OpBuilder* builder, mlir::tf_device::ClusterFuncOp cluster_func,

        &(*cluster_to_core_index)[core]);
    if (failed(result)) return failure();

    for (Type t : output_types) concatenated_output_types.emplace_back(t);
  }

  *cluster_idx = tensorflow::MovePreservedParallelExecuteChildren(
      num_cores_per_replica, concatenated_output_types, builder, cluster_func,
      old_parallel_execute, new_parallel_execute);

  // Extract inputs for each block of the parallel_execute op. The i-th