Memory model

[Doug Lea]

> For some of the details, I completely agree.  But I think we need to
> quickly resolve whether or not we want to define semantics to programs with
> races between ordinary variables.  That appears to me to affect a lot
> of other things.  We'd have to go down two completely different paths.

I'm a little diappointed to agree with you, but I do. Trying to define
semantics of C++ programs with races will probably encounter too many
obstacles to do right. Except...

> 1) I suspect that we could leave the semantics of programs that
> have races under a sequentially consistent interpretation undefined,
> provided our definition of "race" is such that the calls into
> the atomic operations library don't count, or are viewed as
> synchronization operations. 

We can also allow reads/writes to volatiles not to count as
races, as did the JMM.

-Doug