Some of them appear to be previously undetected problems with bad RPATHs for asan (this should be fixed by https://reviews.llvm.org/D24184), but I've been investigating a different error with asan unittests on OS X. See http://lists.llvm.org/pipermail/llvm-dev/2016-September/104431.html for more details.
Downgraded to a warning in: [llvm] r280674 - [lit] Downgrade error to warning on gtest crashes during discovery.
Minor update: I was able to reproduce the asan unittest error I was hitting, but only on OS X 10.10 machines, not 10.11 or 10.12. This seems related to r241487: for some reason, we don't DYLD_INSERT_LIBRARIES the asan dylib in the unittests, and on 10.10, that disables interception, causing inconsistent asan errors. Manually adding it lets the unittests run successfully. I suspect we need OS X DYLD_INSERT_LIBRARIES logic similar to LD_LIBRARY_PATH in the asan lit config.
I narrowed the asan problem down to r263695: it disabled the re-exec support for all unittests (it previously was only disabled for noinst tests).
Can you tell me what exactly is failing? The mentioned email thread discusses a failure on a Linux bot, but here you’re saying that this is a OS X 10.10-specific failure. I do have a 10.10 machine and it “check-all” seems to work fine.
Lit used to ignore crashing unittests, but now errors out. This masked several different issues. I downgraded the lit error to a warning in r280674, so we're back to masking the test failures. This is why check-all passes. The various Linux bot failures should all have been fixed by r280791. The asan 10.10 issue is different, and is not fixed yet. For check-asan, you'll see lit warnings with "unable to discover google-tests" with cryptic asan errors. I believe that's caused by r263695 disabling the re-exec on missing DYLD_INSERT_LIBRARIES for the instrumented tests. Does that make sense?
You’re right, I can reproduce this. It’s a bit unfortunate that the instrumented and non-instrumented tests work very differently. Submitted <https://reviews.llvm.org/D24699> for review.
I assume this is fixed or obsolete?