Re: [PATCH i-g-t 00/25] tools: remove unnecessary shared library

["Dixit, Ashutosh" <ashutosh.dixit@intel.com>]

On Thu, 25 Jun 2026 01:47:13 -0700, Sebastian Brzezinka wrote:
>

Hi Sebastian,

> On Wed Jun 24, 2026 at 6:33 PM CEST, Ashutosh Dixit wrote:
> > On Wed, 24 Jun 2026 01:47:07 -0700, Sebastian Brzezinka wrote:
> >>
> >> Standalone tools linked libigt.so as a side effect, pulling in cairo,
> >> pixman and libX11 even without framebuffer use.
> >>
> >> Following intel_gpu_top's approach, each tool now links only the static
> >> sub-libraries it needs. Five new ones are introduced:
> >>
> >> lib_igt_tools_stub, lib_igt_drm_stub, lib_igt_halffloat,
> >> lib_i915_decode, lib_igt_reg_tools.
> >
> > So I am trying to figure out what is the "real" reason for doing this.
> The motivation for this was explained in 2019, freedesktop bug
> #110249, by Eero Tamminen:
>
> "It's annoying to need to install these redundant dependencies on
> e.g. headless media transcoding server if one just wants to use
> intel_gpu_top to monitor GPU utilization."
>
> Chris Wilson fixed intel_gpu_top back then and said "One down, the rest
> left to an adventurous sole." This series finishes that job.  Tools should
> only link what they actually use. That's basic. intel_stepping does
> not need Cairo. intel_backlight does not need ALSA.  Yet today they
> both pull in 30+ libraries through libigt.so. This series fixes that.
> I shouldn't have to explain why a tool that reads a register shouldn't
> depend on a graphics rendering library.
>
> > If tomorrow the tool uses a function which is not present in the tiny
> > library but in the big library, what are we going to do? So, isn't it
> > better to link against one big library, rather than tiny libraries?
> >
> If a tool needs new functionality in the future, we add it to the
> appropriate sub-library or add a new dependency. That's a one-line meson
> change, done intentionally, with a clear reason.
>
> That's better than the alternative: preemptively linking everything
> against a library that drags in Cairo, ALSA, X11 and 35 other things
> "just in case" someone someday needs something from it.
>
> > Since it should use dynamic linking, it is not that the size of the
> > executable will change one way or the other.
> >
> The point is not executable size. The point is runtime dependencies
> what needs to be installed on the system for the tool to run at all.
>
> With libigt.so as a dynamic dependency, you cannot run intel_stepping
> on a headless server without first installing Cairo, X11, ALSA, freetype
> and 35 other libraries that the tool has absolutely no use for.
>
> After this series, you just run the tool. Nothing extra needed.

OK, because you are linking statically I think.

>
> > So what is driving this change? Maybe I am missing something.
> I appreciate the question, sorry for not including the link in the cover
> letter. This issue was raised by the community back in 2019 and is well
> documented here: https://bugs.freedesktop.org/show_bug.cgi?id=110249

Thanks for the detailed explanations, the series is:

Acked-by: Ashutosh Dixit <ashutosh.dixit@intel.com>

Thanks.
--
Ashutosh