[cpp-threads] Should atomic_add support pointers?

[Peter Dimov]

Doug Lea wrote:
> Peter Dimov wrote:
>> I'm trying to nail down the semantics of an atomic ops library and I
>> encountered the following interesting issue:
>>
>> Should atomic_add support pointer addition? That is:
>>
>> int * p;
>>
>> atomic_add( &p, 4 ); // allowed? sensible?
>>
>> What do you think?
>
> In
>   atomic_add( &p, 3 );
> would the "3" mean 3 or does it mean 12 = 3 * sizeof(int)?
> (Either way seems defensible.)

I'm interested in the p += 3 case (3 * sizeof(int) on a byte-addressable 
machine, just 3 on a hypothetical non-extinct word-addressable machine), 
because it has a direct language equivalent.

> Also I suspect that people like Maged Michael who use marked-pointers
> a lot would like to be able to do:
>    atomic_or(&p, (intptr_t)1);
>    atomic_and(&p, ~(intptr_t)1);

This certainly seems useful, but since it's inherently non-portable (or even 
less portable, if you like), the specification can leave its semantics (even 
more) implementation defined. Value bits vs storage bits, also an 
interesting issue.