[cpp-threads] modes

[Peter Dimov]

Doug Lea wrote:

> The minimal list seems to be along the lines of
>   read-with-acquire
>   write-with-release
>   write-preserving-order
>   CAS-with-acquire-release
>   CAS-with-release
> (plus the default ordinary-read and ordinary-write)
>
> We'd still need to spell out the details of exactly what acquire,
> release, and write-ordering mean, especially wrt the highly
> inconvenient sequence-point wording of the rest of the C++ spec.

The good thing about acquire and release is that a write-with-release on a 
memory location X happens-before a subsequent load-with-acquire on the same 
memory location. :-)

But there are two problems with the above set.

The first problem is that it's missing unconstrained atomics. Ordinary reads 
and writes are not atomic or race-resistant, and there is no unconstrained 
CAS in the set (and by extension, no unconstrained read-modify-writes.) This 
will penalize reference counting by up to 100% because it doesn't need 
synchronization on increment, only on decrement.

The second problem is that finer-grained constraints can eliminate some of 
the barriers. For example, read with acquire is

    read
    #LoadLoad | #LoadStore

but read with hoist-load barrier is

    read
    #LoadLoad

and a compiler is unable to exploit this without being explicitly told that 
only loads need to be constrained.

However finer-grained constraints seem unexpressible in a sequential 
consistency or a happens-before model. Or if they are expressible, I haven't 
figured out how.