[PATCH 05/10] Remove __need macros from errno.h (__need_Emath, __need_error_t).

[Florian Weimer]

On 05/10/2017 01:47 PM, Joseph Myers wrote:
> On Wed, 10 May 2017, Florian Weimer wrote:
> 
>> On 05/09/2017 11:38 PM, Joseph Myers wrote:
>>> On Tue, 9 May 2017, Zack Weinberg wrote:
>>>
>>>> diff --git a/misc/sys/cdefs.h b/misc/sys/cdefs.h
>>>> index 74f9a49b05..075c996f28 100644
>>>> --- a/misc/sys/cdefs.h
>>>> +++ b/misc/sys/cdefs.h
>>>> @@ -326,6 +326,13 @@
>>>>    # define __attribute_artificial__ /* Ignore */
>>>>    #endif
>>>>    +/* Force use of the initial-exec TLS model.  */
>>>> +#if __GNUC_PREREQ (3,3)
>>>> +# define __attribute_tls_model_ie__ __attribute__ ((tls_model
>>>> ("initial-exec")))
>>>
>>> This is not namespace-clean for an installed header (needs to be
>>> __tls_model__).
>>
>> The larger problem is that we need to error out if __GLIBC_USE (TLS_ERRNO) and
>> the compiler does not support the initial-exec TLS model because it cannot
>> produce the required ABI for the errno symbol.
> 
> You don't need initial-exec TLS for errno.  If a port wishes to export TLS
> errno directly, it's perfectly valid for the headers to do
> 
> extern _Thread_local int errno;
> #define errno errno
> 
> with any C11 compiler, and if the compiler only supports GD model TLS,
> that's fine for accessing an IE variable (it may or may not get optimized
> to IE by the linker).  Declaring the TLS model is simply an optimization.

That's true if we want to create a port which is effectively C11-only. 
(The NaCl header generator uses __thread, though, so it's still 
non-standard.)

However, I still think that the NaCl port should be aligned with the 
rest of glibc in this regard, and that Zack's cleanup should not 
introduce a new extension point for it.  Particularly since we don't 
have a working port right now which would exercise this functionality, 
so it's likely to be broken anyway.

Thanks,
Florian