CppunitTest_vcl_skia failing on tb68 with illegal instruction

Michael Stahl mst at libreoffice.org
Tue Mar 9 13:35:46 UTC 2021

On 09.03.21 14:19, Luboš Luňák wrote:
> On Tuesday 09 of March 2021, Stephan Bergmann wrote:
>> (e.g., <https://ci.libreoffice.org/job/gerrit_windows/90702/>).
>> @Luboš:  I assume that external/skia (maybe since
>> <https://git.libreoffice.org/core/+/ad8bff9d2625524999871ace65cfe0382f991f2
>> 4%5E!/> "update Skia to chrome/m90", which reached master on 2021-03-03?) is
>> built with support for instruction set extensions that are not available on
>> tb68.
>   That doesn't make sense to me. The crashing test calls normal Skia code that
> shouldn't be special WRT instruction sets, and it's normally compiled with
> clang-cl without explicit instruction set flags. The only plausible
> explanation I can think of is that clang-cl defaults to a higher instruction
> set default than cl's (or ours) and the relevant code is somehow unlucky to
> have clang-cl to actually use such an instruction there. But even that
> doesn't really make sense, because that code should get called also by some
> of the previous tests. Moreover tb68 is reportedly E5530 CPU, which supports
> up to SSE4.2, so I don't quite see how something would get accidentally
> compiled to something not supported there without people noticing. Even on
> x86_64 clang doesn't default to anything higher than the baseline SSE2.
>   Is it possible to debug the crash?

is that tb68 running some virtualization?  it's possible to restrict the 
result of CPUID via the hypervisor (to enable live migrations to other 
hosts), could that be the case?

More information about the LibreOffice mailing list