On Mon, Mar 16, 2026 at 11:16:40AM +0000, Soumya AR wrote:
> Hi,
>
> Pinging again with updated patches after some fixes and cleanup.
>
> This work is deferred till GCC17 but sending these here to prevent
> redundant review on the older patches.
Does it really have to be a separate new pass (especially always enabled)?
Every new pass which has to walk the whole IL is costly (especially when
done also at -O0 and -Og cases), can't be done inside of some other pass
which already walks the IL?
If it can't (but I seriously doubt that), at least there should be some flag
on struct function whether the current function has any of such IFN calls
(it can be also just whether it could have them, can be just set
conservatively, but should rule out the 99.95% of functions which don't have
such calls). Also, naming it about atomic fetch minmax when you plan to
add some other atomics that should be lowered too doesn't seem to be a good
idea, are you going to add another pass for floating point atomic add?
I saw at least one case of {}s around single statement body, that is just
a formatting nit.
Jakub
