https://gcc.gnu.org/bugzilla/show_bug.cgi?id=122827
Andrew Pinski <pinskia at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
See Also| |https://gcc.gnu.org/bugzill
| |a/show_bug.cgi?id=123285
--- Comment #8 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
(In reply to ktkachov from comment #7)
> (In reply to Andrew Pinski from comment #6)
> > (In reply to ktkachov from comment #5)
> > > My bisection landed on g:b191e8bdecf btw, which was added to address
> > > PR117477.
> >
> > Hmm, this would mean it is either a register allocator issue or a pass after
> > ra which is causing the issue. But aarch64, does -mearly-ra=none (or
> > -mearly-ra=strided) cause the failure to go away? I am trying to see if the
> > issue is related to doing ra early for floating point or not.
>
> Indeed, -mearly-ra=none makes the run pass, so it does look related. Can you
> point me to your patch to try please?
It was committed as r16-7687 (PR123285).
I wonder if `-fno-caller-saves` or `-fno-ipa-ra` helps.
PR 121495 might be a similar issue where early ra either goes wrong and causes
the RA afterwards to go wrong. The problem with PR 121495 is that it is only
reproducible in GCC 14 (due to it going latent). I am going to do a few more
steps to understand what is going wrong in PR 121495 this week.
