[cpp-threads] Yet another visibility question

[Peter Dimov]

Boehm, Hans wrote:

> We could disallow the thread 5 behavior by explicitly requiring that
> stores to each single atomic object x occur in a single total order
> "store-order-for-x", and that all such orders be included in the
> "happens-before" relation.

Do you mean store-releases here (ordering raw stores seems too strict)? And 
if so, isn't this already the case? If it isn't, consider this example:

// thread 1

fetchadd_release( &x, 1 );

// thread 2

fetchadd_release( &x, 1 );

// thread 3

r1 = load_acquire( &x );
r2 = load_acquire( &x );

assert( r1 <= r2 );

A model in which thread 3 can see r1 == 2 and r2 == 1 would be a challenge 
to explain.