[cpp-threads] C++ memory model

[Peter Dimov]

Nelson, Clark wrote:
>>> I'd like to just delete the second senetnce:
>>>
>>> "Implementations shall not move an atomic operation out of an
>>> unbounded loop."
>>
>> I'd like to cast my vote in favor of the sentence. It does indeed
>> carry little formal value, but in practice, it can (easily) settle a
>> hypothetical
>> debate between a compiler user and a compiler writer about whether
>> such an
>> optimization is allowed.
>
> Well, maybe; but then the argument will turn to the whether the loop
> in question is really "unbounded", and since the standard doesn't
> define "unbounded loop", the argument will continue -- without
> bounds.

An unbounded loop is one that has no (finite) bound. A loop has a bound N if 
it never loops more than N times.

> So, for example, is the compiler required to prove that a loop can't
> ever terminate for it to qualify as unbounded?

No. It has to prove that the loop has a bound in order to move atomic 
operations out of it.

> I think it would be helpful to see an example of a program for which
> this sentence might be considered relevant.

    while( x );

?

    while( !x.compare_exchange_weak( &t, t+1 ) );

?