[bazel] Fix lit tests with python 3.11+

[keith]

In python3.11 there is a new environment variable PYTHONSAFEPATH which
stops python from setting the current directory as the first entry in
sys.path. Bazel started setting this to ensure that python targets
don't accidentally access things that aren't in their dependency tree.
This resulted in lit tests breaking because sys.path didn't include the
directory to the lit source files. This is fixed by adding the lit
binary to the dependency tree and propagating the import path from it.

Fixes #75963

[rupprecht]

Excellent! Fixes #75963.

[keith]

oh cool, thanks for the link! I realized that I don't think we need to accept deps at all in the macro, so I just removed it. None of the ones in the tree right now are so I assume that's fine. I didn't do that originally just because I wasn't sure what kwargs were being passed, but I doubt it makes sense to pass deps to the underlying py_test in general.

[rupprecht]

I realized that I don't think we need to accept deps at all in the macro, so I just removed it. None of the ones in the tree right now are so I assume that's fine.

I also don't see anything out of tree. Users will pass their own lit configs/tests via data, nobody should need to touch deps. If someone is doing that, and it's legitimate, we can use what you had before. The error message should be obvious in pointing to this -- something along the lines of "you're passing deps twice".

[fzakaria]

Awesome.

How did you debug this ? I'm always keen to learn debugging workflows from others.

[keith]

I actually wish I had found this issue first since that would have helped a lot.

I was lucky enough to hit the failure on my macOS machine, but had a Linux machine where I wasn't hitting it (but I didn't know why yet). From there I threw a print(sys.path) in the first line of the lit.py script where I spotted a difference. I also happened to know that it's been an issue in bazel before that sys.path changes have broken things. From there I found bazelbuild/bazel#9239 which through some cross links eventually lead me to bazelbuild/bazel#15701 once it was clear it was python version related, and that it was in the name of better hermiticity, I just started looking for a fix.