https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98782
--- Comment #22 from Jan Hubicka <hubicka at gcc dot gnu.org> --- We discussed with Richard the option to biass the IRA spilling heuristics to spill into blocks containing calls (since they are slow anyway) bypassing some of the guessed frequencies. Thinking on it, perhaps I should note that I believe that even with profile feedback we seem to have IRA problem (it used to be possible to get faster code w/o -fprofile-use then with), so the default heuristics seems to be somehow misbehaving. With current guessed profile the frequency of BB containing the recursive call is about 0.9, while reality is about 0.99 which is quite close.