https://gcc.gnu.org/bugzilla/show_bug.cgi?id=117875
Jan Hubicka <hubicka at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Ever confirmed|0 |1
Status|UNCONFIRMED |NEW
Last reconfirmed| |2024-12-03
--- Comment #3 from Jan Hubicka <hubicka at gcc dot gnu.org> ---
I saw this on LNT testers too. Unforutnately hmmer was not working for a while,
so there is relatively big interval about when this can happen.
Last year I did some work on hmmer and loop splitting. It is was very sensitive
about profile update. I added logic to splitting so if the first loop has
constant number of iterations, we update it correctly, but did not implement
pattern matching needed to check that the second loop has constant number of
iteration. The problem is that w/o intermediate FRE iv analysis is not able to
analyze it from the code we produce in second loop. It takes output from the
first loop and it is not obviously clear what the values are...
I guess what is new is that we manage to vectorize the second loop and
completely lose track of resulting code before working out it is a non-loop :)
