gtk4 4.12.0 failed some tests on i386, which I initially thought were a
regression in 4.12.0, but then 4.10.5+ds-4 failed in the same way.
4.10.5+ds-3 and 4.10.5+ds-4 are essentially the same (the only changes
were unrelated) so this must have been triggered to a change to a
dependency: I would guess fontconfig and pango1.0 are the most likely
candidates.
I don't know whether this results in user-visible regressions to GTK
4 applications on i386; that would be a question for someone who is
routinely using a full i386 system.
These are reftests, which work by rendering GTK's internal representation
of two UIs (loaded from text files) to an image, and comparing them. The
idea is that they're two different ways to achieve the same result, one
simple and one exercising some more complicated code path.
The tests still pass on every other architecture, so I suspect that what's
happening here might be that different rounding from the i387 floating
point interface having extended-precision registers, and as a result of
that different rounding, the text is a very slightly different width?
To reproduce: build gtk4 source package.
Expected (good) result: the two affected reftests pass, and do not leave
output files in debian/build/deb/testsuite/reftests/output/x11/.
Actual (bad) result: they do leave output files. This does not cause FTBFS
or cause them to be marked as failures, but only because we've told the
test suite to ignore them.
Alternative bad result: before version 4.10.5+ds-5, the failures were not
ignored and caused FTBFS.
Logs:
success (slightly older versions of dependencies):
https://buildd.debian.org/status/fetch.php?pkg=gtk4&arch=i386&ver=4.10.5%2Bds-3&stamp=1691406899&raw=0
failure:
https://buildd.debian.org/status/fetch.php?pkg=gtk4&arch=i386&ver=4.10.5%2Bds-4&stamp=1692723794&raw=0
failure ignored:
https://buildd.debian.org/status/fetch.php?pkg=gtk4&arch=i386&ver=4.10.5%2Bds-5&stamp=1692731298&raw=0
smcv