[cpp-threads] Some concerns with N2324

[Peter Dimov]

Lawrence Crowl wrote:
> On 7/16/07, Peter Dimov <pdimov at mmltd.net> wrote:

...

>> [Note: two possible definitions of _Relaxed and __relaxed might be
>>
>> struct _Relaxed {};
>> #define __relaxed _Relaxed()
>>
>> and
>>
>> enum _Relaxed { __relaxed };
>>
>> --end note]
>
> Syntactically, this works, but the ABI for the two are
> generally not the same, and we need the same ABI.

The ABI for the constraints is only a concern when one has a C function

void f( memory_order m, ... );

that is also used from C++ (where it's declared extern "C") - or vice versa. 
N2195 pretty much prohibits this use case since there is no containing 
'memory_order' type to which all constraints are convertible, and since 
there is no overloading in C, there's no need to pass a constraint argument 
at all.

I don't know how important is this use case in practice. Most existing 
atomic libraries support no constraint arguments at all (at least in its C 
portion) so they obviously don't deliver the functionality. The function is 
hard to inline, so the performance is likely to suffer on most platforms.

The atomic_* functions themselves can map to different names in C and C++ 
mode (if they exist as functions at all; N2195 says that they need not, not 
sure about N2324 - I don't see a prohibition in it to take their addresses, 
so intrinsics and macros seem outlawed).