On Tue, Jun 05, 2018 at 04:14:16PM +0200, Peter Zijlstra wrote: > On Tue, Jun 05, 2018 at 02:58:23PM +0100, Mark Rutland wrote: > > > Sure, it all works, it's just less than optimal as above, and also means > > that we have to duplicate the ifdeffery for optional atomics -- once in > > the instrumented atomics, then in the "real" atomics. > > > > Whereas if we filled in the raw atomics atop of the arch atomics, > > everything above that can assume the whole API is present, no ifdeffery > > required. > > Aah, I see your point now. I don't think performance is a particular > concern when you enable K*SAN, but getting rid of a fair bunch of > ifdeffery is always nice.
I agree that performance isn't a concern there when debugging, but I would like to keep the overhead down when fuzzing. Regardless, we'd have to move arches over to arch_atomic_* first, and once that's done the raw_atomic_* stuff is fairly easy to implement by reworking the scripts. Thanks, Mark.