Further research shows that no bot uses “all” targets; the bot clang-x64-windows-msvc is listed as running debuginfo-tests, but that ninja target became invalid after the move to cross-project-tests. It’s the only bot I see in the zorg config file that has that target. So, apparently the missing piece of work is to get that bot switched over to the new target (and make sure it’s still passing).
@rnk is listed as the owner of that bot. With some notice, we (Sony) should be able to provide assistance if the switch doesn’t Just Work.
Sure, that’d probably be good. I think Reid’d be OK if you want to commit that zorg change and check the buildbot - either undo the change or xfail the tests in some way to clean it up?
I want to apologize for all the fuss and bother here. We do want things to be working. Is it worth filing a PR for this? We have an internal ticket already to get this taken care of.
Not too fussed - however it’s tracked. Mostly would like there not to be standing failures - so if you folks want to XFAIL them in some way in the interim. If most of the existing test coverage is on Darwin where Dexter doesn’t run (can it be made to run) and one Windows buildbot - might be worth Sony working with other buildbot/configurations to get this included, and/or bringing up a public buildbot that does something similar to your internal one?
(maybe I missed it - but was your internal buildbot failing with the same problems I was seeing externally (that look like they’ve been there for a month or so?)? Or was it covering some different configuration that didn’t hit these issues?)
From: llvm-dev <email@example.com> On Behalf Of via llvm-dev
Sent: Thursday, October 28, 2021 9:18 AM
To: firstname.lastname@example.org; Cazalet-Hyams, Orlando <email@example.com>; firstname.lastname@example.org
Cc: Weaver, Tom <Tom.Weaver@sony.com>
Subject: Re: [llvm-dev] dexter failures in cross-project-tests
without any buildbots running them
Perhaps naively, I thought cross-project-tests (which includes dexter) was part of check-all, and so would be running on any bots that used check-all. Do bots not use check-all? Internally it looks like our Linux build does exactly that, and I see dexter tests passing there.
“targets” is a bit ambiguous - mostly when I think of a buildbot and which targets they use I’m thinking of which LLVM target backends are enabled in the build.
In any case, no, I’m not sure whether most bots use “check-all” and while that’s part of the issue, the other part is probably whether or not they enable the subproject in LLVM_ENABLE_PROJECTS in the cmake configuration (not all bots build all projects - so their “check-all” is limited to whichever projects they enable). It’s possible to use “all” in LLVM_ENABLE_PROJECTS, but I doubt many folks do that - it pulls in a lot of stuff and would make for quite a slow buildbot, though I suppose the nature of cross-project-tests is that it’s meant to use a lot of the other subprojects and so it may be appropriate to have a buildbot with “all” projects enabled and running “check-all”, which should cover everything - but I think that’d take some work to cleanup - I believe (based on some recent practical experience trying to enable libcxx, libcxxabi, lldb, and cross-project-tests and running check-all and getting failures due to missing dependencies in some parts of that.
I guess partly the responsibility here also goes to Apple who added the debuginfo-tests ( +Adrian Prantl +Jonas Devlieghere ) - maybe they run that on green dragon. Perhaps they lost test coverage when the project was renamed & are no longer testing it? Or Perhaps debuginfo-tests had only/predominantly darwin coverage (because it was introduced by Apple folks and that’s their priority) & then the dexter stuff was added as non-darwin, so came in without significant test coverage?