also, avoid the "encoding_sort" transformation, if it can be done
directly in the printer.

On the same example as in the previous commits, this gives 5x
acceleration together with some memory usage reduction.