https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116713
Oleg Endo <olegendo at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|UNCONFIRMED |NEW
Last reconfirmed| |2024-09-20
Ever confirmed|0 |1
--- Comment #2 from Oleg Endo <olegendo at gcc dot gnu.org> ---
(In reply to pietro from comment #1)
> (In reply to Oleg Endo from comment #0)
> > In some cases __builtin_prefetch will be eliminated and thus it can't be
> > reliably used for SH4 store queue writes.
> >
> > An example here: https://godbolt.org/z/TGsed8cnq
>
> It doesn't get eliminated on SH2A/SH3/SH4 on that example. It does get moved
> to before the writes to sq_part between O1 and O2 though:
>
> https://godbolt.org/z/sP8PPKz3K
Thanks!
It looks like the problem can be "fixed" by inserting a '__atomic_thread_fence
(1);' before the '__builtin_prefetch', which kinda makes sense.
So maybe a new SH specific builtin function could be added which would insert
the memory fence automatically.
