[PATCH v1 2/4] nptl: Continue use arch prefered atomic exchange in spinlock loop

[H.J. Lu]

On Thu, Sep 29, 2022 at 11:40 AM Noah Goldstein via Libc-alpha
<libc-alpha@sourceware.org> wrote:
>
> On Thu, Sep 29, 2022 at 9:35 AM Wilco Dijkstra <Wilco.Dijkstra@arm.com> wrote:
> >
> > Hi Noah,
> >
> > Did you try building it both ways? I don't think this could ever compile:
> >
> > +  if (__glibc_likely (pthread_spin_lock_grab_lock (lock, &val, 1)))
> >
> > and:
> >
> > +#  define pthread_spin_lock_grab_lock(mem, val, c) \
> > +    atomic_compare_exchange_weak_acquire (lock, &val, 1))
> >
> > The define uses 'lock' and 'mem' inconsistently and the use of the macro
> > expands into &&val...
> >
> > Apart from that there is the question whether we should keep the weird
> > ATOMIC_EXCHANGE_USES_CAS setting - I have removed it in my atomic
> > patch series since most targets appear confused as to what it means (so
> > are likely to have the wrong setting). Also there is no evidence it is actually
> > faster. So using exchange in both cases is easier (and less error prone!).
> >
> > Also you do realize that no matter how much you change this code, it
> > won't make a difference on x86, right?
>
> Why's that?

x86-64 uses sysdeps/x86_64/nptl/pthread_spin_lock.S

-- 
H.J.