Re: Curious global variable bug? Bad assumptions on my part?

[Tom St Denis <tstdenis at ellipticsemi dot com>]

Matthew Woehlke wrote:
> Eljay Love-Jensen wrote:
> > Adding volatile to the g_counter tells the compiler not to optimize away
> > access to the variable, as the variable can change "at any time".
> >
> > That may sound like just the thing for multi-threaded access to the same
> > variable -- but it's not sufficient.
> >
> > Even though the compiler is no longer going to optimize away access 
> > to the
> > variable, the hardware can still effectively isolate the two threads 
> > from
> > one another, such that neither thread is aware that the g_counter 
> > variable
> > has changed.
> >
> > Why?  Because each thread could run with the variable stored in separate
> > caches.
>
> Huh? I thought SMP hardware is required to be responsible for cache 
> syncing in such an instance...

Not all platforms are your desktop.  Think of embedded systems with no 
snooping, the Nintendo DS is such a model iirc.

Even then, even with snooping, GCC could alias the variable to a 
register [without volatile] in which case it would still not work.


Tom