http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54776
Richard Guenther <rguenth at gcc dot gnu.org> changed: What |Removed |Added ---------------------------------------------------------------------------- Status|UNCONFIRMED |NEW Last reconfirmed| |2012-10-02 CC| |hubicka at gcc dot gnu.org, | |rguenth at gcc dot gnu.org Target Milestone|--- |4.8.0 Ever Confirmed|0 |1 --- Comment #1 from Richard Guenther <rguenth at gcc dot gnu.org> 2012-10-02 09:52:17 UTC --- Confirmed. ISTR the inline predicate stuff is responsible - the flatten numbers are still ok, likewise the results with profile feedback. As for -flto, using -fwhole-program should be enough to get all possible speedup (and a nice reality check if -flto works for single-TU as well as -fwhole-program). Looks like tramp3d is no longer our primary benchmark focus ;)