In principle we can do this for all transformations, but performance for arbitrary transformations is still very slow.