https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107822
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Last reconfirmed|2022-12-21 00:00:00 |2023-2-27
--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
Re-confirmed. More importantly it's not only -Os which regresses but also -O2.
All VRP passes are now "EVRP" style, doing a single walk of the CFG
via the dominator walk in substitute_and_fold and triggering ranger
"on demand" from there.
That's a fundamental regression but it was a concious one maybe not
fully anticipating the effect. SCEV can only help in a subset (but
probably the most important set) of cases here.
It might be possible to have ranger itself mimic what the old SSA propagator
VRP did for SSA SCCs (which was tuned to be good enough but not too
expensive).
