http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45375
--- Comment #149 from Jan Hubicka <hubicka at ucw dot cz> 2012-12-02 15:05:52 UTC --- > > Please try to reduce HOT_BB_COUNT_WS_PERMILLE to 990. I also see some > > regressions > > on some SPEC benchmarks (such as GCC) and this helps. If it doesn't it > > would be > > nice to know what value is needed for comparable size. > > Unfortunately it doesn't help much, because with "--param > hot-bb-count-ws-permille=990" the size is only 0.25% smaller: > (With --param) : 42098856 > (Without ) : 42204584 > > I will try smaller values later. Hmm, that sounds like quite bad news - the histogram code was supposed to help in such cases. I will try to fix the non-PGO case and lets try to compare how PGO/non-PGO compare first. If you could put somewhere the -fdump-ipa-inline dump, I will try to check if there is something obviously wrong. In worst case we can resort to combining both heuristics - i.e. keeping the hot_bb_fraction in addition to histogram code. In fact I planned to do that this way but Teresa removed the old code and I did not see good reason why to keep it. Honza