[PATCH wayland] wayland-egl: Ignore underscored symbols in ABI check

[Daniel Stone]

On 20 March 2018 at 11:55, Emil Velikov <emil.l.velikov at gmail.com> wrote:
> On 20 March 2018 at 11:46, Daniel Stone <daniel at fooishbar.org> wrote:
>> Sure. As on IRC though, we definitely need to add at least _ftext for
>> MIPS anyway:
>> https://gitlab.gnome.org/GNOME/glib/commit/ad12142943e0f20ed9583c9d6bf50f6262110c74
>>
>> And probably some more for ARM toolchains using other linkers:
>> http://infocenter.arm.com/help/index.jsp?topic=/com.arm.doc.dui0474k/pge1362066045758.html
>>
>> Doing a quick check across all the architectures on Debian shows that
>> your updated list is also missing _fbss, _fdata, and _ftext. So we'd
>> need to respin for those as well, but I think at this point the
>> inconvenience of maintaining a list of every linker's implementation
>> detail on every platform, outweighs the risk of an exported
>> underscore-prefixed symbol slipping through review.
>
> I think you're preemptively worried about this. Plus there's no need
> to add every symbol a search could find.
>
> Even then, ask yourself the question:
> Which is better - updating an ugly looking list once in a blue moon or
> having apps use internal API?
>
> As said above: _if_ it turns out to be labour intensive - fine remove it.
> We've already spend more time researching than what the [expected]
> maintenance for 1 year would be ;-)

It all depends on how you view scope and probability.

I don't agree my concern is preemptive: we've just shipped a beta
where 'make check' fails on a few architectures, and the proposed fix
(adding several symbols to the platform list) will still fail on
architectures Debian ships on, until it also adds _fbss, _fdata and
_ftext. I think my suggestion is reasonable: it fixes the very real
problem, in a way which avoids expanding the scope of our ABI checker
to include the implementation details of every linker/architecture
pair that people use Wayland on. The cost of this is the probability
that someone manages to add a new underscored symbol marked with
WL_EXPORT into a library which is currently 105 LoC, and have no-one
notice that in review. I think that's a reasonable tradeoff.

Cheers,
Daniel