Matching Moshi ADTs with MLStyle

[visr]

SciML is migrating from Expronicon to Moshi. To make that less breaking I already added some tricks like SciML/SciMLBase.jl#921. Another concern is that Mochi ADTs cannot be matched with MLStyle out of the box. We could of course migrate MLStyle to Moshi but there is a lot of code out there so it would help if we can match Moshi ADTs with MLStyle. Since both support extensible pattern matching I guess that can be made to work, I haven't tried yet. Do you already happen to have code for this? Would you be interested in having such code in a weak dependency, or should we place it in a weakdep of say SciMLBase? See also discussion in SciML/SciMLBase.jl#922 (comment).

https://thautwarm.github.io/MLStyle.jl/latest/syntax/pattern.html#custom-patterns
https://rogerluo.dev/Moshi.jl/start/match/#extensible-pattern-matching

[visr]

Hmm I'm guessing a weakdep with a generic MLStyle.pattern_uncalll method cannot work since each Moshi ADT is distinct with types not subtyping any Moshi type. Still some recipe would be useful. I tried some things but got stuck, the MLStyle docs are quite minimal for this part. Perhaps @thautwarm has ideas.

I was trying something along the lines of this to start with, but I get StackOverflows. This is on the Message ADT from https://rogerluo.dev/Moshi.jl/start/getting-started/#quick-example.

using MLStyle: MLStyle, @match

MLStyle.pattern_uncall(t::Type{<:Message.Variant}, m, type_params, type_args, args) = m(Expr(:call, Symbol(t), args...))

# trying to compile a match like
@match m begin
    Quit() => "time to quit"
    Move(x, y) => x
    Write(s) => s
    ChangeColor(r, g, b) => r
    _ => "other"
end

[Roger-luo]

Moshi was made because there are some design flaws made it impossible to support ADT with MLStyle (see also my blog post) these breaking changes are intentional which was why it end up in a new package...

[Roger-luo]

Expronicon was made for MLStyle and you can see where the limitations etc from there. If you wish to have something compatible with MLStyle I'm afraid you will have to use Expronicon's approach.

[Roger-luo]

I think syntax wise, if you really want to go this route it's still possible to support a specific subset of the syntax. Which limits things to call syntax only. E.g you won't be able to use the namespace syntax in MLStyle, and you need to use $ mark to say what you are evaluating instead of assigning.

Unfortunately I won't have the bandwidth to work on anything else right now as explained in previous issue I'm in a limbo situation in Toronto and I have no idea where I'm gonna live next month at the moment, thanks to US govt...

[visr]

Yeah indeed I'm aware they are quite different, it is mainly that the transition to Moshi would be easier if we can figure out a way to use MLStyle's extensible pattern matching on some Moshi ADTs.

Good luck with the visa situation! I really appreciate your work on Moshi, I can tell a lot of thoughtful design went into it.

[visr]

I managed to get a small subset for a simple ADT to work, as a package extension to SciMLBase in SciML/SciMLBase.jl#922.

Main limitation is that wildcard matches don't work: SciML/SciMLBase.jl#922 (comment)

[visr]

I'll close this since the package extension, though flawed, was enough to provide an upgrade path. The main issue that came up after the switch was #43.