https://gcc.gnu.org/bugzilla/show_bug.cgi?id=118634
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Last reconfirmed| |2025-01-24
Status|UNCONFIRMED |NEW
Ever confirmed|0 |1
--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
It's indeed missed unrolling / jump threading and the resulting optimization.
The cunroll size estimate doesn't properly "simulate" through the body, there's
also
<bb 3> [local count: 715863672]:
# counter_30 = PHI <counter_25(17), 0(2)>
# iter_22 = PHI <iter_26(17), _1(2)>
# ivtmp_55 = PHI <ivtmp_33(17), 2(2)>
if (counter_30 != 2)
goto <bb 4>; [50.00%]
which I think is statically known to evaluate to false. What's a bit
unfortunate is that we run VRP before loop header copying but not after.
I do have somewhere a prototype to make cunroll size estimates more
precise by simulating the loop (to account for unreachable parts).
It's all optimized away with -O3.
