recent KahanSum change causes a new test failure on ppc64le

Dan Horák dan at danny.cz
Wed Sep 13 08:32:35 UTC 2023


On Thu, 7 Sep 2023 11:36:33 +0200
Stephan Bergmann <sbergman at redhat.com> wrote:

> On 9/7/23 08:25, Dan Horák wrote:
> > On Thu, 7 Sep 2023 08:16:28 +0200
> > Stephan Bergmann <sbergman at redhat.com> wrote:
> >> On 9/5/23 18:53, Dan Horák wrote:
> >>> seems the recent change [1] to KahanSum handling causes a test failure
> >>> on ppc64le.
> >>>
> >>> Running scope as unit: -home-sharkcz-projects-libreoffice-workdir-CppunitTest-sc_statistical_functions_test.test:20230905152639:2378561.scope
> >>> [_RUN_____] StatisticalFunctionsTest::testStatisticalFormulasFODS
> >>> Testing load file:///home/sharkcz/projects/libreoffice//sc/qa/unit/data/functions/statistical/fods/KahanSum.fods:
> >>> /home/sharkcz/projects/libreoffice/sc/qa/unit/functions_test.cxx:49:StatisticalFunctionsTest::testStatisticalFormulasFODS
> >>> forced failure
> >>> - Testing file:///home/sharkcz/projects/libreoffice//sc/qa/unit/data/functions/statistical/fods/KahanSum.fods failed, Sheet2.A90 '=SUM(F3:F102)' result: 6.6, expected: 5
> >>
> >> Interesting; I also saw that failure with my latest local build on macOS
> >> aarch64 against Clang trunk.  (But didn't debug it further and wrote it
> >> off as maybe some intermittent Clang trunk bug.)
> >>
> > 
> > someone on IRC reported the same issue, also from macOS I believe
> 
> So, wild speculation, maybe it's a difference between x86-64 and all 
> other architectures, given how that 
> <https://gerrit.libreoffice.org/c/core/+/156253/> "Resolves: tdf#156985 
> Treat adding two KahanSum differently" kept failing for (32-bit) 
> <https://ci.libreoffice.org/job/gerrit_windows/> up until patch set 7. 
> (And where patch set 8 reverted back to the original code for _WIN32, 
> maybe in a misguided attempt to fix something that was seen failing on 
> Windows 32-bit x86, but not known to (not) fail on Windows x86-64.)
> 
> Eike, any thoughts?
> 

and for the record, I have just confirmed Fedora aarch64 and s390x
systems have the same issue


		Dan


More information about the LibreOffice mailing list