Revert "fixup: remove boehmgc patch"

[roberth]

Motivation

The patch is still necessary.

Should fix #7644

Context

The patch is still necessary.
@lheckemann @bb010g @edolstra, please do your research, or just ask the author (which happens to be me, but that doesn't matter).

An evaluator like this is not an environment where "it compiles, so it works" will ever hold.

This reverts commit 1c40182.

Checklist for maintainers

Maintainers: tick if completed or explain if not relevant

• agreed on idea
  • n/a
• agreed on implementation strategy
  • n/a
• tests, as appropriate
  • functional tests - tests/**.sh
  • unit tests - src/*/tests
  • integration tests
• documentation in the manual
  • n/a
• code and comments are self-explanatory
  • can explain later if desired
• commit message explains why the change was made
• new feature or bug fix: updated release notes
  • fix; only after Release notes automation #7665

[roberth]

Draft for now because the patch doesn't actually apply completely; the reverted commit doesn't cover the whole change.

[cole-h]

I don't know how easy it would be, but is there a way to test this that we could integrate into the test suite (or even as a hydra job)? As in, a way to validate that the patch is still required (not that it ever might not be required, but mostly to just avoid this situation where removing it seems to be fine, but causes issues later on), or that the patch is not properly applied...?

[roberth]

We could run some large evaluations, but Nixpkgs is unlikely to be GC-ed while a source filtering coroutine is active.
So for me personally it's not a done deal and I'd rather work to improve the GC/coroutine integration in a way that works with upstream. That said, upstream is iterating on a redesign of the interfaces to better support such an integration, so for now I'm going to leave it be.
If I'm wrong to be pessimistic about finding a reliable reproducer, I'd be happy to include it in the test suite of course.

[cole-h]

Upstream integration would be even better!

[github-actions]

Successfully created backport PR for 2.13-maintenance:

• [Backport 2.13-maintenance] Revert "fixup: remove boehmgc patch" #7680

[Ericson2314]

Perhaps we could make some sort of unit test that manually triggers a GC. Hard/impossible to do externally I agree, but perhaps possible internally.

This would be a nice thing to have even if we don't need the patch because vanilla Boehm GC handles such things.

[roberth]

Actually I did add a triggerGC primop originally, but Eelco rejected that. An identity function with a side effect, not unlike trace. You could trigger it in a source filter predicate to make sure a coroutine is active during GC.

[Ericson2314]

Let's have a nix-testing experimental feature so we can make such things for testing purposes without qualms.