[js-api] Are Wasm-gc abstract heap types included in the JS API of e.g. WebAssembly.Global?

[Liedtke]

The WebAssembly JS-API spec mentions the following types for globals:

enum ValueType {
  "i32",
  "i64",
  "f32",
  "f64",
  "v128",
  "externref",
  "anyfunc",
};

However, both V8 and SpiderMonkey support multiple more values like "anyref" or "i31ref". After some code-search in V8, I found my commit which added it in 2022 when anyref was introduced for wasm-gc as a separate type hierarchy vs. the existing externref. In later changes we added more of these new types like i31ref probably assuming some kind of consistency to the already present anyref.

It seems that JavaScriptCore doesn't accept these values.

The wasm-gc JS API document does not list any extensions to these existing APIs. So from a spec perspective, it looks like this is a bug in V8 and SpiderMonkey? Was it a conscious decision to not add support for the new abstract heap types to the existing JS APIs?

[tlively]

I don't think this was a conscious decision, and I do think it would make sense to add the other abstract heap types. Supporting non-nullable references and references to defined heap types would require much more design work, of course, but I don't think that's a good reason to hold off on making the obvious extensions to the current scheme.

[Liedtke]

Notably I didn't find "externref" as a new value in the JS-API documented anywhere in the typed function references proposal.

The proposals also don't seem to link a js-api spec with their proposed changes? So while we probably shouldn't have added anyref, I'm wondering whether there was any documentation that advised us to add externref in the first place? I also couldn't find a spec test for an externref-global defined via the JS API.

A few more notes:

1. A similar issue applies to tables where new WebAssembly.Table({element: 'anyref', initial: 4, maximum: 4}) will just pass in V8 as well and produce a table of anyref. And yes, for the same assumed consistency, also tables of i31ref and other abstract heap types are possible.
2. While my proposal would be to add "the other abstract heap types" to the APIs, this does not include the null types! There isn't any use case for things like new WebAssembly.Table({element: 'nullref', initial: 4, maximum: 4}), so we shouldn't spec it.

[bvisness]

This whole space is further muddied by the type reflection proposal, which is still languishing in phase 3. I think we are probably long overdue to revisit it.

FWIW, this is the function in Firefox that handles this.

[tlively]

2. While my proposal would be to add "the other abstract heap types" to the APIs, this does not include the null types! There isn't any use case for things like new WebAssembly.Table({element: 'nullref', initial: 4, maximum: 4}), so we shouldn't spec it.

Yes, that seems fine. I assume that whatever long-term type reflection mechanism we end up with would be general enough to allow creating tables and globals of nulls, but no need to support them eagerly.

[rossberg]

I have a vague recollection that we discussed supporting the respective top types, i.e., add anyref, as part of the minimal JS API MVP, but defer the rest to later.

@Liedtke:

Notably I didn't find "externref" as a new value in the JS-API documented anywhere in the typed function references proposal.

The externref type was already part of Wasm 2.0, so is independent of typed func refs or GC types.

[Liedtke]

The externref type was already part of Wasm 2.0, so is independent of typed func refs or GC types.

Thanks for clarifying! I added a spec test for externref globals in the JS API as that is already spec'ed.

Would it be realistic to add more reference-types outside a completely new proposal? The spec is in a weird state right now, where e.g. the algorithm references exnref but it isn't an accepted ValueType, so it seems like the assumption was that exnref would be allowed as value, but then be explicitly rejected, similar to new WebAssembly.Global({value: "v128"}), so arguably there seem to be some assumptions also by spec authors that more abstract heap types would be allowed as ValueType?

2. Let valuetype be ToValueType(descriptor["value"]).
3. If valuetype matches v128 or exnref, 3.1. Throw a TypeError exception.

Really sounds to me like the expectation is that exnref was accepted by ToValueType .

[rossberg]

The latter probably is the result of semi-mechanically adding exnref to all the places where v128 is ruled out. In some of them it is redundant (though harmless, and potentially a bit more robust wrt future additions?).

[eqrion]

I think we definitely should support the top-types defined by wasm-gc in the ValueType enum. I think it was just an oversight that these weren't added to the JS-API spec. We use them quite a bit in our internal testing, so I'd really not want to remove them unless there is a good reason to.

[tlively]

So what is the full set of strings we should accept after fixing these oversights?

Definitely these:

• externref ✅
• anyfunc ✅
• anyref
• exnref

Should we add "funcref" as an alias of "anyfunc"?

Should we add "i31ref", "eqref", "structref", and "arrayref" as well?

[rossberg]

I'd definitely include anyref and funcref. Not sure about exnref, given that it cannot be passed to JS in the first place, but I guess it would be consistent with v128?

I'd avoid adding anything else without consolidating the JS type reflection proposal.