check-all target runs much less regression tests if check-runtimes fails

[DimitryAndric]

Now that releases via the llvm/utils/release/test-release.sh script are using LLVM_ENABLE_RUNTIMES for e.g. compiler-rt and libcxx, it turns out that the check-all target may run much less regression tests. This is most likely because check-all depends (via CMake's add_dependencies()) on check-runtimes, if runtimes are enabled. I.e., if check-runtimes results in any failure whatsoever, none of the other tests under check-all are even tried.

We might need to make check-runtimes failures not block the rest of check-all to run, or maybe splice all the tests from check-runtimes and check-all into one big bucket?

See also https://discourse.llvm.org/t/14-0-0-rc2-has-been-tagged/60621/3

[ldionne]

I'm looking at this, however I'm struggling to make a fix that actually works. I'll ping Petr Hosek for assistance.

[petrhosek]

I came up with a solution but I need a bit more time to clean it up.

[tstellar]

@petrhosek Any progress on a fix? @ldionne if we don't get a fix in time, can we fix this at least for the release testers by reverting c8bb177?

[ldionne]

See https://reviews.llvm.org/D121258 for the revert on release/14.x.

[tstellar]

See https://reviews.llvm.org/D121258 for the revert on release/14.x.

I've merged this patch to the release/14.x branch. We can switch it back to the new runtime builds for 14.0.1 when this is fixed.

[petrhosek]

https://reviews.llvm.org/D121276 is a candidate solution.