Re: volatile qualifier hurts single-threaded optimized case

[Paolo Carlini <pcarlini at suse dot de>]

Paolo Carlini wrote:

> Benjamin Kosnik wrote:
>
> > > #ifdef __arch64__
> > >  typedef long _Atomic_word;
> > > #else
> > >  typedef int _Atomic_word;
> > > #endif  
> >
> > I'm pretty sure: I scoped out the baseline files for what is exported.
> > However, if there is an issue, I'll fix it up.
>
> If __arch64__ is defined the signatures should be:
>
>    _ZN9__gnu_cxx12__atomic_addEPVli
>    _ZN9__gnu_cxx18__exchange_and_addEPVli
>
> therefore we could use the usual:
>
>     _ZN9__gnu_cxx12__atomic_addEPV[il]i
>    _ZN9__gnu_cxx18__exchange_and_addEPV[il]i

Ah, in the meanwhile I tried to understand why the baseline for sparc 
uses ii also for sparc64 and testresults are fine: it seems from the 
documentation that *by default* the sparc back end always generates code 
for v7. Then the problem arises if the user starts generating code for 
v9 and passes -m64.

To be super-safe, I think we should "respect" the macro above and export 
also the long variant...

Paolo.