js-api tests for externref

[yurydelendik]

I'm trying to find (and add?) tests for init arguments for Table/Global constructors/methods that were introduced by reference-types proposal.

The https://webassembly.github.io/spec/js-api/index.html#dom-table-table defines the initial ref value to be ToWebAssemblyValue(undefined, externref), which is resolved to some ref.extern (not null).

Is this a mistake and the above default value has to be null? If it is not a mistake, what the following test must produce?

  const argument = { "element": "externref", "initial": 3 };
  const table = new WebAssembly.Table(argument);
  assert_equals(table.length, 3);
  assert_equals(table.get(0), null); // ??

[eqrion]

There's also a test case here for default JS-API values for globals:

const global = new WebAssembly.Global({value: "externref", mutable: true});
assert_equals(global.value, ???);

My reading of the JS-API spec is that the default value of externref is 'undefined' because that's what omitted parameters in JS give, and that's a valid non-null externref value. So I'd expect in both of these test cases that the default value would be undefined, not null.

However Chrome/Firefox/Safari currently give different results:

	V8	SM	JSC
Global	null	null	undefined
Table	undefined	undefined	undefined

[Ms2ger]

Let's look at the spec: both Global and Table have:

1. If v is missing,
   1. Let value be DefaultValue(valuetype).

DefaultValue:

6. If valuetype equals externref, return ToWebAssemblyValue(undefined, valuetype).

and ToWebAssemblyValue returns a non-null externref for undefined.

Unless there's a compelling reason to change, I'd say this seems fine and the bugs in implementation should be fixed.

I'm surprised to see there's no tests for either case; they should be in global/constructor.any.js and table/constructor.any.js. It would be good to add them regardless of the behavior we decide upon.

[rossberg]

I'd argue this definition of DefaultValue is a bug, because it doesn't match what the core spec defines as default value.

[Ms2ger]

Oh, I recalled one good reason to keep the current behavior - the following are defined to be equivalent in IDL:

  const table = new WebAssembly.Table(argument);
  const table = new WebAssembly.Table(argument, undefined);

so if you treated undefined as ref.null, you couldn't create a table (global) with initial value undefined.

[rossberg]

Yes, that's why the undefined value is such a broken concept. One shouldn't ever be using it as a first-class thing. Once you assume that's a thing, default arguments of generic type basically become impossible.