Bug 1847064

did this work before ?

Does this setup work on different arches?

(In reply to Florian Müllner from comment #7)
> Does this setup work on different arches?

Works on x86_64 on latest RHEL 8.3.

Works on Power and ARM.

This BZ is somewhat reminiscent of several BZs that Ray Strode, Tom
Stellard, and I worked on in 2017:

https://bugs.freedesktop.org/show_bug.cgi?id=100613
Regression in Mesa 17 on s390x

https://bugzilla.redhat.com/show_bug.cgi?id=1438891
Black output on ppc64 and s390x

https://bugzilla.redhat.com/show_bug.cgi?id=1445423
blue shadows on ppc64 and s390x

https://bugzilla.redhat.com/show_bug.cgi?id=1456040
Black output on ppc64 and s390x

Fortunately, at that time glxgears served as a bellweather to pinpoint
bad behavior in Mesa's llvmpipe.

And fortunately, glxgears appears to serve in the same capacity today.

For today's purposes, glxgears exhibits three different behaviors
depending on the underlying version of Mesa:

1) correct display;
2) broken geometry: gears are recognizable, but lighting,
   orientations, and rotation look wrong;
3) totally random-looking 2D trash.

I bisected Mesa between 19.3-branchpoint, which exhibited behavior #2
(broken geometry; later versions 19.3.x show correct behavior), and
20.0-branchpoint; see the attached bisect log.

The bisection pointed to commit e6b2af56cb0 (attached), which enabled
NIR as the default IR (preferred over TGSI).

I did a Mesa 20.1.1 debug build, and ran the following tests:

LP_DEBUG=tgsi_ir GALLIVM_DEBUG="ir asm" glxgears >& tgsi.out
GALLIVM_DEBUG="ir asm" glxgears >& nir.out

The respective output files are posted below.

With NIR (and running metacity as the window manager), glxgears
exhibited behavior #3, i.e. random 2D trash.

But with TGSI, glxgears displayed correctly; AND, as a bonus,
gnome-shell is back to working, too!

The problem of making NIR work for big-endian remains, however.

Created attachment 1698209 [details]
bisect log

Created attachment 1698210 [details]
commit e6b2af56cb0

Created attachment 1698211 [details]
TGSI

Created attachment 1698212 [details]
NIR

Created attachment 1698213 [details]
Possible patch/workaround to go back to TGSI on big-endian arch

Created attachment 1698255 [details]
S390x tri NIR

Here is the NIR output from the Mesa demos trivial/tri program.

This program draws a smooth-shaded triangle on a dim purplish
background (r = 0.3, g = 0.1, b = 0.3, a = 0.0)
using just glBegin/glColor3f/glVertex3f/glEnd to draw the vertices in
counterclockwise order:

x0 = -0.9  y0 = -0.9  z = -30.0  r = 0.8  g = 0    b = 0
x1 =  0.9  y1 = -0.9  z = -30.0  r = 0    g = 0.9  b = 0
x2 =  0.0  y2 =  0.9  z = -30.0  r = 0    g = 0    b = 0.7

So, red on lower left, green on lower right, and blue on top.

Created attachment 1698256 [details]
For comparison, ppc64le tri NIR

For the sake of comparison, here is the NIR output from the same
Mesa demos trivial/tri program.

(In reply to Ben Crocker from comment #18)
> Created attachment 1698255 [details]
> S390x tri NIR
> 
> Here is the NIR output from the Mesa demos trivial/tri program.
> 
> This program draws a smooth-shaded triangle on a dim purplish
> background (r = 0.3, g = 0.1, b = 0.3, a = 0.0)
> using just glBegin/glColor3f/glVertex3f/glEnd to draw the vertices in
> counterclockwise order:
> 
> x0 = -0.9  y0 = -0.9  z = -30.0  r = 0.8  g = 0    b = 0
> x1 =  0.9  y1 = -0.9  z = -30.0  r = 0    g = 0.9  b = 0
> x2 =  0.0  y2 =  0.9  z = -30.0  r = 0    g = 0    b = 0.7
> 
> So, red on lower left, green on lower right, and blue on top.

Sorry, forgot to mention: at this moment, this program only draws
the background, no triangle.

Created attachment 1698860 [details]
Latest S390x NIR output

Note the small number of lines beginning with "input"

Created attachment 1698861 [details]
Latest ppc64le NIR output

Note the LARGE number of lines beginning with "input"

Upstream fix merged:

https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5677

Built locally.

Created attachment 1700331 [details]
s390x vncserver with gnome

I can successfully begin a vncserver on s390x "Server with GUI" without a black screen and it works as expected.

Since the problem described in this bug report should be
resolved in a recent advisory, it has been closed with a
resolution of ERRATA.

For information on the advisory (userspace graphics, xorg-x11, and mesa bug fix and enhancement update), and where to find the updated
files, follow the link below.

If the solution does not work for you, open a new bug report.

https://access.redhat.com/errata/RHEA-2020:4742