Re: Improving GNUstep tooling with Clang

[Frederik Seiffert]

Am 28.11.2019 um 22:06 schrieb Niels Grewe <niels.grewe@halbordnung.de>:
> 
> On 28.11.19 17:15, Frederik Seiffert wrote:
>> - Set up CI testing for the 2.0 runtime version, which currently doesn’t 
>> seem to be tested with libs-base.
>> I have submitted a pull request for this 
>> (https://github.com/gnustep/libs-base/pull/88).
> 
> Unfortunately, when I tried to integrate that into my libdispatch stuff,
> I discovered that this doesn't actually test the 2.0 runtime. The CI
> script is still using a Ubuntu 14.04 builder, which ships with a clang
> version that is too old to know about the 2.0 runtime version
> (defaulting to something else, presumably). Once run under 18.04 (baby
> steps…), with clang 7, we're getting failures in the configure script
> because the new objc_init struct references non-existent sections in the
> binary (those conftest things don't declare any classes and such).

Oh, good catch! How did you find out that the Clang version didn’t support the 
2.0 runtime? Is there a warning or similar that I can watch out for in the 
future?

> Upgrading to clang 9 fixes that, but now I'm seeing a load of test
> failures that I haven't had the time to investigate further:
> 
> https://travis-ci.org/gnustep/libs-base/builds/618352009

Ok wow – to be honest, based on our experience with the 2.0 runtime, I was kind 
of surprised when all the tests were previously passing. It sounds like 
something fundamental is broken here though and you’re onto it 
(https://github.com/gnustep/libobjc2/issues/136). Let me know if I can give a 
hand with this in any way.

Frederik