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

[Martin Sebor via Libc-alpha <libc-alpha@sourceware.org>]

On 4/26/21 10:41 PM, Florian Weimer wrote:
> * Martin Sebor via Libc-alpha:
> 
>> diff --git a/misc/sys/cdefs.h b/misc/sys/cdefs.h
>> index 8e244a77cf..ac56be4d87 100644
>> --- a/misc/sys/cdefs.h
>> +++ b/misc/sys/cdefs.h
>> @@ -592,8 +592,14 @@ _Static_assert (0, "IEEE 128-bits long double requires redirection on this platf
>>      array according to access mode, or at least one element when
>>      size-index is not provided:
>>        access (access-mode, <ref-index> [, <size-index>])  */
>> -#define __attr_access(x) __attribute__ ((__access__ x))
>> +#  define __attr_access(x) __attribute__ ((__access__ x))
>> +#  if __GNUC_PREREQ (11, 0)
>> +#    define __attr_access_none(pos) __attribute__ ((__access__ (__none__, pos)))
>> +#  endif
>>   #else
>>   #  define __attr_access(x)
>> +#  define __attr_access_none(pos)
>> +#endif
> 
> I don't think this works because __attr_access_none is not defined for
> GCC 10 due to the way the definitions are nested.  I think you should
> move the __GNUC_PREREQ (11, 0) check to the top level.

You're right.  I had inadvertently reverted the pthread.h changes so
my testing didn't expose it.  Too many patches in the same tree...
I've committed a1561c3bbe with the missing definition (and retesting
the macro).

> 
> It might be consistent with fewer attribute access extensions to write
> __attr_access_none ((__none__, 2)) instead of __attr_access_none (2).

Repeating the "none" part doesn't seem in any way useful to me.  It
wouldn't reduce the number of macros, existing or new.

Martin

> 
> Thanks,
> Florian
>