From: Kees Cook
> Sent: Friday, August 18, 2023 9:25 PM
>
> On Fri, Aug 18, 2023 at 04:10:49PM -0400, Jeff Layton wrote:
> > [...]
> > extra checks (supposedly) compile down to nothing. It should be possible
> > to build alternate refcount_t handling functions that are just wrappers
> > around atomic_t with no extra checks, for folks who want to really run
> > "fast and loose".
>
> No -- there's no benefit for this. We already did all this work years
> ago with the fast vs full break-down. All that got tossed out since it
> didn't matter. We did all the performance benchmarking and there was no
> meaningful difference -- refcount _is_ atomic with an added check that
> is branch-predicted away.
Hmmm IIRC recent Intel x86 cpu never do static branch prediction.
So you can't avoid mis-predicted branches in cold code.
David
> Peter Zijlstra and Will Deacon spent a lot of
> time making it run smoothly. :)
>
> -Kees
>
> --
> Kees Cook
-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
Registration No: 1397386 (Wales)