Re: [PATCH 3/7] Remove atomic_bit_set/bit_test_set

["H.J. Lu" <hjl.tools@gmail.com>]

On Wed, Jul 6, 2022 at 12:15 PM Noah Goldstein via Libc-alpha
<libc-alpha@sourceware.org> wrote:
>
> On Wed, Jul 6, 2022 at 12:00 PM Wilco Dijkstra <Wilco.Dijkstra@arm.com> wrote:
> >
> > Hi Noah,
> >
> > > > -  if (atomic_bit_test_set (&pd->cancelhandling, TERMINATED_BIT) == 0)
> > > > +  if ((atomic_fetch_or_acquire (&pd->cancelhandling, 1 << TERMINATED_BIT)
> > > > +      & (1 << TERMINATED_BIT)) == 0)
> >
> > > Can these remain macros that setup the bit_test_set properly but
> > > use compiler intrinsics?
> >
> > The fetch_or above is already written in a way that allows GCC to recognize it
> > (it's identical to your do_atomic_or0), so I don't see any issue here. However
> > there is a TERMINATED_BITMASK which might make it more readable.
>
> For non-constant shift values it needs to be like atomic_or4/atomic_or5.
>
> Since this needs to be set up in a particular way it seems worth it to
> keep it as a macro that will do it properly. Fine with replacing the underlying
> atomic op with a compiler intrinsic.

GCC can optimize some atomic operations to locked bit operations.
Have we checked if the new assembly codes are as efficient as the
old ones on x86-64?

-- 
H.J.