> Yeah, I think it doesn't belong into this series at least. I have > also > (multiple times now ...), done the flag_wrapv = 1; before the last reassoc > pass idea. -fwrapv is what we have on RTL so it made some sense to > try transitioning to that earlier. But "late" reassoc is quite early, so it > regresses things a lot, and we now use ranger during RTL expansion > even. Also it catches the actual expression simplifications only late.
I tried enabling it in a third reassoc pass just before expand. While that didn't regress a lot, it was too late for the followup optimizations that would make my example work (and likely others). -- Regards Robin
