Re: [PATCH] add attribute none to pthread_setspecific (BZ #27714)

[Martin Sebor <msebor@gmail.com>]

On 4/22/21 6:11 PM, Paul Eggert wrote:
> On 4/22/21 2:30 PM, Martin Sebor via Libc-alpha wrote:
> 
>> -    __THROW;
>> +    __THROW __attr_access_none (2);
> 
> Instead of inventing a new __attr_access_none macro that developers will 
> need to remember, why not add support to the existing __attr_access 
> macro? That is, uses can look like this:
> 
>       __THROW __attr_access ((__none__, 2));
> 
> if we define __attr_access with something like the attached patch.

I don't have a preference for how to define the macro.  I went with
a new one because Joseph suggested that approach in hist comments
on the mismatched allocation patch (for the __attr_dealloc macros).
The other approach I considered was using the __attr_access macro
but guarding it for GCC 11 in situ.  Since functions like
pthread_setspecific are exceeingly rare, I don't expect the new
attribute to be used very much at all (if you do know of other
such functions, though, please let me know so I can annotate them
as well).

As for your suggested change. I think we considered variadic macros
when we first introduced the attribute but rejected it for some
reason that I'm not sure I remember.  Maybe because they're a C99
feature and Glibc supports older compilers?

If this isn't the case and others share your preference for this
approach I don't mind changing it.

> 
> Alternatively, one could keep both cdefs.h and the callers simple by 
> doing access attribute checking only for GCC 11 and later. That'd be 
> good enough in the long run.

I'd view disabling the checking with older GCC releases as a QoI
regression, so I'm not in favor of this alternative.

Martin