Record: make field ordering deterministic, add test

[matthiasdiener]

This is a proof-of-concept and is ready for a first look @inducer. I'm not sure if there is a better way to do this.

Please squash

[inducer]

• This seems OK, although it is a compatibility break, which I'm not loving.
• We should just deprecate the thing for removal in 2025. I.e. have it spew warnings, and remove it from the docs. (in favor of dataclasses, which do the same thing, but more competently)

[matthiasdiener]

• This seems OK, although it is a compatibility break, which I'm not loving.

Is there a use case that breaks with this change? My goal was to remain compatible with the previous set-based implementation (which is the reason for the conversion from set to dict in __init__).

• We should just deprecate the thing for removal in 2025. I.e. have it spew warnings, and remove it from the docs. (in favor of dataclasses, which do the same thing, but more competently)

What do you think of c32e6d0?

[matthiasdiener]

Setting this as ready for review. In addition to the pytato/loopy tests, mirgecom also seems to run fine with it.

[inducer]

I'll be honest, I underestimated the compatibility impact of this. I've just pushed a revert.

[inducer]

inducer/loopy#785 gets rid of a good bunch more Records in Loopy, which may make this more plausible.

[inducer]

@matthiasdiener Could you please re-make a PR for this?

[matthiasdiener]

Sure, no problem! Sorry for the confusion. Is there a particular error that came up with this?

[inducer]

It was code like this that made me substantially uneasy:

https://github.com/inducer/loopy/blob/d5512d5c6aec12f16a60aa5da31701c23d006bbd/loopy/kernel/instruction.py#L206-L211