Building skia with clang(-cl)

[Luboš Luňák]

On Monday 08 of June 2020, Stephan Bergmann wrote:
> On 23/04/2020 11:46, Miklos Vajna wrote:
> > In short, I would like to hear what others think about using clang(-cl)
> > for building skia.
> >
> > Why? See this video from Lubos, showing skia raster performance on
> > Windows: <https://youtu.be/iSo4wVtGQ2A>. Above is clang-cl, below is
> > MSVC. The reason for the difference is that skia developers focus on
> > good performance with clang, so building their codebase with gcc or msvc
> > works, but the result is not as great as it could be. This would be the
> > benefit.
>
> FYI,
> <https://lists.fedoraproject.org/archives/list/devel@lists.fedoraproject.or
>g/message/HWBXFXPFNNPZSB3OYW6DUKZUDSSGZJ7H/> "Re: Fedora 33 System-Wide
> Change proposal: CompilerPolicy Change":  "If I understand the LO case, it
...
> So at least for GCC the performance degradation when building skia with
> anything but Clang might eventually get addressed.

 In the specific case I saw (fmin/fmax), GCC can actually already perform more 
or less the same, if it gets -ffinite-math-only (without it GCC is not smart 
enough to optimize "fmax(0,fmin(1,x))"). I didn't look into the issues in 
that much detail though.

 But as long as Skia is used only by VCL gen backend this really doesn't 
matter in practice. We could probably even default to --disable-skia on Linux 
and it wouldn't make much of a difference for users.

-- 
 Luboš Luňák
 l.lunak at collabora.com