#1153522 mesa: 26.2.6 appears to regress on riscv64

Package:
src:mesa
Source:
src:mesa
Submitter:
Jeremy Bícha
Date:
2026-11-19 03:45:01 UTC
Severity:
normal
Tags:
#1153522#5
Date:
2026-11-09 14:00:51 UTC
From:
To:
mesa 26.2.6 appears to regress on riscv64 as seen in autopkgtests
triggered by mesa itself:
https://qa.debian.org/excuses.php?package=mesa

In qt6-pim's build tests, this error is seen on riscv64:
https://buildd.debian.org/status/logs.php?pkg=qt6-pim&arch=riscv64

25: qt.glx: qglx_findConfig: Failed to finding matching FBConfig for
QSurfaceFormat(version 2.0, options
QFlags<QSurfaceFormat::FormatOption>(), depthBufferSize -1,
redBufferSize 1, greenBufferSize 1, blueBufferSize 1, alphaBufferSize
-1, stencilBufferSize -1, samples -1, swapBehavior
QSurfaceFormat::SingleBuffer, swapInterval 1, colorSpace
QColorSpace(), profile  QSurfaceFormat::NoProfile)
25: qt.glx: qglx_findConfig: Failed to finding matching FBConfig for
QSurfaceFormat(version 2.0, options
QFlags<QSurfaceFormat::FormatOption>(), depthBufferSize -1,
redBufferSize 1, greenBufferSize 1, blueBufferSize 1, alphaBufferSize
-1, stencilBufferSize -1, samples -1, swapBehavior
QSurfaceFormat::SingleBuffer, swapInterval 1, colorSpace
QColorSpace(), profile  QSurfaceFormat::NoProfile)
25: qt.glx: qglx_findConfig: Failed to finding matching FBConfig for
QSurfaceFormat(version 2.0, options
QFlags<QSurfaceFormat::FormatOption>(), depthBufferSize -1,
redBufferSize 1, greenBufferSize 1, blueBufferSize 1, alphaBufferSize
-1, stencilBufferSize -1, samples -1, swapBehavior
QSurfaceFormat::SingleBuffer, swapInterval 1, colorSpace
QColorSpace(), profile  QSurfaceFormat::NoProfile)
25: Could not initialize GLX

This error will likely cause build tests for many packages across
Debian to fail. In other words, FTBFS.

Thank you,
Jeremy Bícha

#1153522#12
Date:
2026-11-11 19:16:24 UTC
From:
To:
Hi,

I confirm the issue. It is due to binutils 2.47.50.20261021-1 which
miscompiles mesa-libgallium. I still have to investigate if a newer
binutils snapshot fixes the issue, and if not try to understand if the
issue needs to be fixed on the mesa side or on the binutils side.

Regards
Aurelien

#1153522#19
Date:
2026-11-16 05:37:38 UTC
From:
To:
Hi,

2c4ccb8103d73917d769cf6c35fa8b71424f20ca is the first 'bad' commit
commit 2c4ccb8103d73917d769cf6c35fa8b71424f20ca
Author: Tatsuyuki Ishi <ishitatsuyuki@gmail.com>
Date:   Wed Feb 21 02:55:56 2024 +0900

    RISC-V: Introduce TLSDESC relaxation.

I'll report upstream and see if I can find a workaround for the mesa
package until it get fixed.

Regards
Aurelien

[1] https://sourceware.org/git/?p=binutils-gdb.git;a=commit;h=2c4ccb8103d73917d769cf6c35fa8b71424f20ca

#1153522#24
Date:
2026-11-16 17:52:47 UTC
From:
To:
control: clone -1 -2
control: reassign -2 src:binutils
control: retitle -2 binutils: TLSDESC relaxation breaks Mesa libgallium on riscv64
control: forwarded -2 https://sourceware.org/bugzilla/show_bug.cgi?id=35009
control: affects -2 src:mesa
control: tag -1 + patch

Hi,
https://sourceware.org/bugzilla/show_bug.cgi?id=35009

Until this is fixed in binutils, the following workaround is enough to
at least get the qt6-pim tests working again:
--- mesa-26.2.6/debian/rules
+++ mesa-26.2.6/debian/rules
@@ -36,6 +36,10 @@
   # Workaround for https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83143
   DEB_CFLAGS_MAINT_APPEND += -freorder-blocks-algorithm=simple
   DEB_CXXFLAGS_MAINT_APPEND += -freorder-blocks-algorithm=simple
+else ifneq (,$(filter $(DEB_HOST_ARCH), riscv64))
+  # Workaround for https://sourceware.org/bugzilla/show_bug.cgi?id=35009
+  DEB_CFLAGS_MAINT_APPEND += -mtls-dialect=trad
+  DEB_CXXFLAGS_MAINT_APPEND += -mtls-dialect=trad
 endif
 buildflags = $(shell \
 	DEB_CFLAGS_MAINT_APPEND='$(DEB_CFLAGS_MAINT_APPEND)' \

Regards
Aurelien

#1153522#33
Date:
2026-11-16 22:27:46 UTC
From:
To:
Hi,

As pointed out in the binutils bug tracker this workaround is definitely
correct as it was the value used with binutils 2.47-6 [1]:

  Checking if "-mtls-dialect=gnu2" runs: DID NOT COMPILE
  Checking if "-mtls-dialect=desc" runs: NO (127)

With binutils 2.47.50.20261021-1 this got change to [2]:

  Checking if "-mtls-dialect=gnu2" runs: DID NOT COMPILE
  Checking if "-mtls-dialect=desc" runs: YES

Regards
Aurelien

[1] https://buildd.debian.org/status/fetch.php?pkg=mesa&arch=riscv64&ver=26.2.5-1&stamp=1792602107&raw=0
[2] https://buildd.debian.org/status/fetch.php?pkg=mesa&arch=riscv64&ver=26.2.6-1&stamp=1793899311&raw=0

#1153522#38
Date:
2026-11-17 05:34:03 UTC
From:
To:
control: forwarded -1 https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/45593

Hi,

Note that while the workaround is perfectly fine, this appears to not be
a binutils issue after all, but a wrong detection of TLSDESC
availability in Mesa. A fix is proposed here:
https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/45593

I have tested that it fixes the issue.

Regards
Aurelien