On Fri, Jul 21, 2023 at 7:34 PM Jan Hubicka <hubi...@ucw.cz> wrote: > > > On Mon, Jul 17, 2023 at 12:36 PM Jan Hubicka via Gcc-patches > > <gcc-patches@gcc.gnu.org> wrote: > > > > > > Hi, > > > While looking into sphinx3 regression I noticed that vectorizer produces > > > BBs with overall probability count 120%. This patch fixes it. > > > Richi, I don't know how to create a testcase, but having one would > > > be nice. > > > > > > Bootstrapped/regtested x86_64-linux, commited last night (sorry for > > > late email) > > > > This should trigger with sth like > > > > for (i) > > if (cond[i]) > > out[i] = 1.; > > > > so a masked store and then using AVX2+. ISTR we disable AVX masked > > stores on zen (but not AVX512). > > Richard, > if we know probability of if (cond[i]) to be p, > then we know that the combined conditional is somewhere between > low = p (the strategy packing true and falses into VF sized > blocks) > and > high = min (p*vf,1) > (the stragegy doing only one true per block if possible) > Likely value is > > likely = 1-pow(1-p, vf) > > I wonder if we can work out p at least in common cases. > Making store unlikely as we do right now will place it offline with > extra jump. Making it likely is better unless p is very small. > > I think if p is close to 0 or 1 which may be common case the analysis > above may be useful. If range [low...high] is small, we can use likely > and keep it as reliable. > If it is high, we can probably just end up with guessed value close but > above 50% so the store stays inline.
I'd say we want to keep the store inline in all cases (we likely lost any explicit profile info during if-conversion), not sure what we gain with providing a better "guess" here. So I think we should simply go with 'likely' derived from statistical independent events. If in future we can tie if-conversion and vectorization even closer we might be able to preserve profile data here. Richard. > > Honza
