[Pkg-opencl-devel] Bug#1065395: spirv-llvm-translator-14: autopkgtest on s390x uses huge amount of disk space

[Andreas Beckmann]

Control: forwarded -1 
https://github.com/KhronosGroup/SPIRV-LLVM-Translator/issues/2397

On 03/03/2024 20.52, Paul Gevers wrote:
> Source: spirv-llvm-translator-14
> Version: 14.0.0-10

> Since a couple of days, our workers on s390x are dying because some test
> is filling up all disk space. Several days ago, I wrongly suspected 
> 
> One of the suspects started to be spirv-llvm-translator-14, so I ran its 
> autopkgtest manually, while logging disk use every 10 seconds (I started 
> slightly delayed because I monitored the wrong partition first). As you 
> can see below, during the test it grows from 17 GB (at the end) to its 
> peak at 179 GB. That's not acceptable on our infrastructure. One file I 
> happened to spot on the way was 
> build/test/test_output/DebugInfo/Generic/Output/two-cus-from-same-file.ll.tmp:
> -rw-r--r-- 1 root root  41G Mar  3 19:18 two-cus-from-same-file.ll.tmp
> 
> I have added spirv-llvm-translator-14 to our reject-list on s390x.
> 
> As this seems to be a rather new issue, I'm wondering if it's due to:
> * Add build-needed autopkgtest for spirv-headers compat check.

Probably.

The buildds report disk usage when building spirv-llvm-translator-* 
between 400MB and 600MB on all architectures except s390x, ppc64, 
sparc64, i.e. all the big-endian ones, where it's slightly above 40GB 
(which very vell corresponds to the file you spotted).
This started with 14.0.0-2 (i.e. 14.0.0-1 was around 500MB on s390x, 
too) which had "* Enable build-time tests, ignore failures on !amd64."

So maybe I should skip the build-time tests on big-endian altogether.

Failure rates:
amd64: 0%
i386: <1%
ppc64el: <2%
most: <10%
s390x: >60%
ppc64: >60%

(Upstream seems to test the testsuite only on amd64, 
https://github.com/KhronosGroup/SPIRV-LLVM-Translator/issues/1964)

Andreas