@connorjward I would like to revert this commit that adds a bunch of atrocities on top of a recent refactor that makes the loopy callables into dataclasses. All of loopy works OK without the atrociities (see the CI). The atrocities are being committed in order to keep Firedrake working. The one I like perhaps the least is that the dataclasses cannot be frozen (immutable), mainly because of this line:

https://github.com/OP2/PyOP2/blob/31471a606a852aed250b05574d1fc2a2874eec31/pyop2/codegen/rep2loopy.py#L119

Replacing that line with

object.__setattr__(self, "name_in_target", name_in_target if name_in_target else self.name)

would already unblock things a bit.

Beyond that, it'd be nice if all that extra code (see the diff in this PR) could go. For that, Firedrake's derived loopy callables would have to become dataclasses. One notable constraint is that frozen dataclasses cannot inherit from non-frozen, so the transition to frozen would likely have to happen first.

As you can tell, all this requires a bit of coordination. I'm hoping you're open to serving as the point of contact here, but feel free to LMK if someone else would be better. Thanks!

Thanks @inducer for the heads up and detailed information. I will hopefully find the time to address this by the end of the week.

Note that we are currently moving PyOP2 into the Firedrake repository so this fix will have to wait until that is merged.

Fixing this is now quite a high priority. Turns out that what you have here isn't quite sufficient for us. Specifically _all_attrs() needs to be ordered for us to have consistent hashes across ranks.

I will try to get things working on this branch.

I am currently working on firedrakeproject/firedrake#3898 which uses this branch.