Re: volatile qualifier hurts single-threaded optimized case

["Richard Guenther" <richard dot guenther at gmail dot com>]

On 8/30/06, Paolo Carlini <pcarlini@suse.de> wrote:
> Benjamin Kosnik wrote:
>
> >   static inline _Atomic_word
> >   __attribute__ ((__unused__))
> >-  __exchange_and_add_dispatch(volatile _Atomic_word* __mem, int __val)
> >+  __exchange_and_add_dispatch(_Atomic_word* __mem, int __val)
> >
> >
> Agreed, but doing this means that in the improvements to rope that
> Richard (and Hans) also want we have to call __exchange_and_add and
> __atomic_add directly,  because in that case the involved memory
> locations are volatile and have to remain like that per Hans desiderata,
> I gather. More generally, the original __exchange_and_add and
> __atomic_add could be used also together with volatile memory and now
> calling the "old" functions becomes necessary, we should document that
> somewhere (if it weren't for this reason, everyone should always go
> through the dispatchers).

Luckily rope is stupid and does not use atomic ops at all, but uses a mutex
and manual increment/decrement.  So rope is not an issue (until it
gets atomic ops).

Richard.