[PATCH] D68708: [RFC] Adopt Dexter and use it to run debuginfo-tests

Tue Oct 29 11:13:59 PDT 2019

jmorse added subscribers: aprantl, vsk.
jmorse added a comment.

Add some debuginfo subscribers; please add more people who might be interested, I didn't get everyones name at the conference. Alternately this might be more appropriate on the mailing list, I can move it there.

All the feedback at the conference was very positive, thanks all. The question now is how to land this, on my mind are four topics:

- Python python python python
- Buildbots
- Code maintenance
- Test convergence

Python: We focused on python 3 only, seeing how python 2 expires in 63 days. This might put Dexter in the unfortunate position of pushing people to upgrade. Please observe the contents of CMakeLists.txt: in its current state, this patch will require python3 to run the debuginfo tests, but should otherwise work alongside LLVM using python2 elsewhere. AFAIUI usage of the debuginfo-tests is currently low, my major concern is the buildbots.

Buildbots: From a quick look, I believe only green-dragon and clang-x64-windows-msvc (@rnk s?) run the debuginfo tests. I know little about the windows buildbot, but it should be straight forwards (TM) for it to use python3. green-dragon definitely has python3 installed because it builds lldb with it. There's a difficulty with green dragon though, I believe it checks debuginfo-tests _before_ building lldb. In its current form, Dexter will pick up the system lldb and use that if there isn't another available. If the system lldb wasn't built for python3, things will go wrong. (I imagine the only way to find all buildbot related problems is going to be pushing up this patch and seeing what breaks).

Code maintenance: Do we need to designate a code owner here? Contributions from the rest of the community is exactly what we want, and tools like llvm-lit don't seem to have code owners. On the other hand having a code owner guarantees someone will review patches, which we / I are happy to do.

Test convergence: right now all the test are pinned to a specific OS and debugger. I'm confident that these limits can be removed and debuggers autoselected, with XFails / unsupporteds added where there are known limitations to platform debuginfo / limitations to the debugger. I'd prefer pushing up the pinned tests as they are then expanding them, as that allows an incremental approach. Note that everything marked "system-linux" is purely because we don't have a mac immediately available to test these things on, we could probably push them up supporting Darwin too.

tl;dr: I think this is good to go, but as we're adding python things there'll probably be buildbot breakage. I'd be happy to push/revert until we've incrementally found all the issues and it sticks.

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D68708/new/

https://reviews.llvm.org/D68708