[cpp-threads] atomic operations package

[Peter Dimov]

Boehm, Hans wrote:

> It still seems to me that if we use templates only at the top level,
> with some predefined instantiations for "template haters" and C
> programmers, and "run-time" parameterization for the sync arguments,
> we should be OK.  And since these functions should normally be
> inlined, there shouldn't be any run-time overhead.  (Unlike
> the template version, we can't guarantee that, but that's minor.)

The "runtime" formulation isn't actually runtime even in its library 
incarnation, it's still a template, because the types of the various msync_* 
values can differ. So instead of

template<class T> class atomic
{
    template<class Msync> T load();
};

you have

template<class T, class Msync> T atomic_load( T * location, Msync msync );

which does have the same two template parameters as the above, it's just the 
syntax that you expose to the outside (possibly-template-unfriendly) world 
that is different.

Of course if atomic_load is a compiler intrinsic, msync_* can just be 
integral constants, since the dispatch can still happen at compile time.