Note that, as it looks, the directive was already broken in previous v5 versions.
So while I solved what looked like a (intentional) regression from the
performance tuning, I actually solved a previous regression as well ;)
I have also added new test cases to the testbench in order to capture such
problems in the future. This version does now look pretty good in shape.
