jplehr wrote: It seems doing a complete "clean build" did help on one of the bots. That is a bit surprising to me, as I think the ones that are still failing should "always clean". I'll have to investigate that part. https://github.com/llvm/llvm-project/pull/154447