Andreas Klöckner

For context, this happened because I was working on https://github.com/inducer/pytato/pull/393 and had a need to have the persistent hash key builder just "understand" expression values. Loopy has some absurd hack...

I'm guessing I may revert this to dataclasses, for exactly this reason. Not sure.

I've changed this to `dataclasses`, and the tests in pymbolic itself seem to be passing. We'll see how the downstreams do.

Most everything seems to be working, except pytential has a puzzling failure.

Thanks for finding the smaller reproducer! That helped quite a bit. The issue was the missing rewriting from `None` to `cse_scope.EVALUATION` in `CommonSubexpression`. It also helped expose https://github.com/inducer/pytential/issues/209, which, as...

Super weird. I wonder why the pytential run is getting canceled. Out of memory for some reason? If so, why would this PR make a difference?

Identical PR on Gitlab: https://gitlab.tiker.net/inducer/pymbolic/-/merge_requests/64

Tried to investigate more. Here's the time profile of the pytential tests, first with this patch: ``` ============================================================================================================ slowest 10 durations ============================================================================================================= 1357.39s call test/test_linalg_skeletonization.py::test_skeletonize_by_proxy_convergence[-case3] 1355.53s call test/test_beltrami.py::test_beltrami_convergence[-operator7-solution7] 914.25s call...

In a way I think it's not a surprise that the runners get canceled. They have about 7G of memory available, plus 4G swap, so the 12G peak allocation is...

OK, so https://github.com/inducer/sumpy/issues/178 seems to be a different issue.