[cpp-threads] memory model

[Peter Dimov]

Doug Lea wrote:
>> Yes. A qualifier approach is also error-prone. This can't be
>> designed out. - Checkable by compiler:
>>
>> More checkable than volatile.
>
> I was referring here to the fact that in the qualifier/volatile
> version, the qualifier declares that every load of a variable must
> have acquire semantics (etc.). So you can't forget to do it. No such
> checks seem possible in the function template approach.

Yes, you are right. Alexander's atomic<> template does have this property. 
The function template formulation is razor-blades-extreme-edition.

On an unrelated note, I wonder what the semantics of x = y are supposed to 
be under the new volatile when x and y are volatile variables of a struct 
type that contains two or more pointers (this code is valid today, so it 
can't be made a compile-time error.)