DRMacIver
commented
1 month ago (Some of) your wish is my command! If you update to the latest version you'll have:
- Initial comment stripping
- New Python passes for stripping annotations and replacing blocks with an ellipsis.
- Pass ordering should be better (although I'd like a better general solution to this)
Inlining is possible in principle but inlining Python in general is a nightmare. A less nightmarish solution that I've been meaning to implement at some point is to support multi-file reduction. Shrink Ray is more or less designed to be able to do this I just haven't yet.
I think allowing the user to edit the file as shrink ray runs is another nightmare. The current interface for that is to press q, edit the file, and try again, and I think this is probably better than most user interfaces I can try.
Zac-HD
I've been hacking on https://github.com/HypothesisWorks/hypothesis/pull/4034 for a few weeks now, and since that involves taking a test and cutting it down to a minimal example which behaves differently depending on the backend, you can see where Shrinkray comes in 😁
As I write, Shrinkray has been running on something extracted from this CI run, trying to minimize the
Larkclass to whatever kernel actually induces our failure:This has worked remarkably well, though with a ~1s interestingness test it does take tens of minutes per round. My basic workflow is to let it run until the code sample is pretty small (~500 -> ~20 lines, around 20 minutes), then inline one of the remaining imports (pre-processed with this script) and repeat.
wishlist of mostly bad ideas that would make me even happier
- Rip through some initial junk as quickly as possible - if it looks like Python code, delete comments, docstrings, and type annotations; remove unused imports; try deleting the entire body of any function. All very crude but they'd speed up the rest of the passes considerably - Pass ordering seems kinda weird for Python; I'm seeing a lot of initial time in debracket and then later split-on-\n-and-delete-chunks makes way more progress. Maybe a more exploratory early-adaptive phase when e.g. debracket is making progress slowly? - Probably doesn't make sense to put it in shrinkray, but a reduction-pump that could inline non-stdlib imports would be really neat. Does require some reasonable judgement about the import graph though. - Utterly cursed, but... allow the user to propose edits while shrinkray is running, and incorporate them if sucessful? - It'd be somewhere between a fun minigame while you're waiting, and actually pretty nice as a way to get occasional speedups - I often see that some pass has caused an undefined name and thus the whole function body can be deleted. - If the interface is a generic "we shell out to $tool and try the patch" you could use an editor, or tools like isort/autoflake etc.