Re: Bug#1077757: gtk4: memorytexture test fails on s390x

Control: retitle -1 gtk4: memorytexture test fails on s390x
Control: severity -1 important

I mitigated this in 4.14.4+ds-8 by skipping the memorytexture test during
autopkgtest as well as during build-time testing, so I'm downgrading this
to important.

On Thu, 01 Aug 2024 at 10:15:11 -0400, Jeremy Bícha wrote:
> There were several upstream fixes to the memorytexture test in June
> and July that aren't in our gtk 4.14.4
> 
> https://gitlab.gnome.org/GNOME/gtk/-/commits/main/testsuite/gdk/memorytexture.c

If you're going to update to 4.15.x in experimental soon *anyway*, then
that could be a good opportunity to find out whether the memorytexture
test might have been fixed on s390x (again). Sorry, but I have already put
more time than I can justify into repeatedly building GTK on porterboxes
in the last couple of weeks, so I am probably not going to chase this
any time soon.

I believe the situation is that it failed in 4.12.x, was fixed in 4.12.x,
was fixed differently in 4.13.x, and then regressed in 4.14.x? Or
something like that?

I'm 95% sure this is because s390x is big-endian and nobody except
us is routinely testing (or even building) GTK on big-endian systems.
Probably one of the mappings between different libraries' colour channel
layouts (GTK, Cairo, OpenGL, Vulkan, ...) is mixing up "first byte is red"
with "least significant 8 bits are red", or similar.

There have been endianness issues in the past that turned out to be a
regression in lower-level libraries like libpixman, so we can't necessarily
be sure that this is really a GTK bug.

> We do have a patch,
> debian/tests-Mark-memorytexture-as-expected-to-fail-on-big-endian.patch
> I'm unsure about the significance: did this test actually pass on
> s390x in 4.12 and then began failing again in 4.14?

Yes. Sebastien Bacher disabled this test in 4.14.4+ds-3
"mark also memorytexture as expected to fail on big endian", without
any reference to a Debian or upstream bug; let's use this bug to
represent that failure.

> gdk/memorytexture.test passes in installed-tests-wayland but not in
> installed-tests-x11

That's perhaps interesting input for someone who understands the finer
points of OpenGL, but I'm sorry, that's not me.

    smcv