[PATCH] meson: Disable optimizations if CLang doesn't support -fsemantic-interposition

[Daniel P. Berrangé]

On Tue, Mar 21, 2023 at 05:34:45PM +0100, Michal Prívozník wrote:
> On 3/21/23 17:04, Pavel Hrdina wrote:
> > On Tue, Mar 21, 2023 at 03:33:24PM +0000, Daniel P. Berrangé wrote:
> >> On Tue, Mar 21, 2023 at 04:31:00PM +0100, Michal Prívozník wrote:
> >>> On 3/21/23 16:25, Daniel P. Berrangé wrote:
> >>>> On Tue, Mar 21, 2023 at 04:11:33PM +0100, Michal Privoznik wrote:
> >>>>> <snip/>
> >>>>
> >>>> I don't like the idea of forcing -O0 for the production builds, just to
> >>>> work around the problem of our broken tests. Can we approach it from the
> >>>> opposite POV and disable building of tests, if we see meson optimization
> >>>> level is > 0
> >>>>
> >>>> eg something roughly like this:
> >>>>
> >>>>   with_tests = true
> >>>>   if cc.get_id() == 'clang' and
> >>>>      not supported_cc_flags.contains('-fsemantic-interposition') and
> >>>>      get_option('optimization') != 0
> >>>>      with_tests = false
> >>>>   endif
> >>>>
> >>>>   if with_tests
> >>>>     subdir('tests')
> >>>>   endif
> >>>>
> >>>>
> >>>> So people can choose to have tests work or not
> >>>
> >>> That could work too, yeah. My reasoning for going with -O0 was that it's
> >>> very rare that somebody would use such old CLang, but I guess disabling
> >>> tests is less invasive. I'll send v2 shortly.
> >>
> >> It isn't just old Clang. Latest clang lacks -fsemantic-interposition
> >> on non-x86_64, which means current macOS M1 platform today. For our
> >> CI, I guess we'll want to request non-optimized builds for macOS
> >> and FreeBSD 12, so we still run unit tests.
> > 
> > We can also utilize the meson 'buildtype' option [1] to figure out if we
> > need to disable/enable tests or modify the optimization that we use when
> > building.
> > 
> > When building from git the default value is 'debugoptimized' but when
> > building using RPM it changes to 'plain'. Not sure what FreeBSD and
> > macOS are using.
> > 
> > The 'buildtype' affects what value will be stored in 'debug' and
> > 'optimization' options. That could allow us to run tests from git builds
> > where we could disable optimizations and disable tests when libvirt is
> > compiled using package manager.
> 
> To summarize what we want:
> 
> if CLang doesn't support -fsemantic-interposition:
>   if building from a git checkout:
>     enforce -O0
>   else
>     disable tests
> 
> What worries me about this approach is that some distros might leave
> .git/ dir in the checkout (e.g. live ebuilds from Gentoo) in which case
> we detect the build as from git. But such distros surely pass
> --buildtype=...
> 
> Now, we need to distinguish two scenarios: when the default value from
> the beginning of the meson.build file was used (and thus we can enforce
> -O0), OR when --buildtype=debugoptimized (or any other option that sets
> optimization level) was passed on the cmd line (in which case we need to
> disable tests). But I don't think there's a way to detect that.

I don't believe we need to check 'buildtype'. A get_option("optimization")
call should reflect the results of the 'buildtype' option IIRC>

> But at the same time, disabling optimization whenever .git/ dir is found
> feels very, very wrong.

Yeah, I wouldn't enforce -O0, just disable tests, and print a message
at end of meson execution to use 'buildtype' to turn off optimization
or use a newer clang, if they want to re-enable tests.

With regards,
Daniel
-- 
|: https://berrange.com      -o-    https://www.flickr.com/photos/dberrange :|
|: https://libvirt.org         -o-            https://fstop138.berrange.com :|
|: https://entangle-photo.org    -o-    https://www.instagram.com/dberrange :|