[f52985]

Related to #1660.

The test "Setting non-function" in test/js-api/table/get-set.any.js expects table.set(0, undefined) to throw a TypeError on an anyfunc table.
As far as I can tell, the specification says that this call is identical with table.set(0), and should instead succeed and store a null reference.

What the specification says

As discussed in #2259, Web IDL's overload resolution turns an explicit undefined passed for an optional argument without a default into the special value "missing", exactly as if the argument had been omitted.

Table.set's second parameter is such an argument:
undefined set(AddressValue index, optional any value);

and its algorithm checks for it:

1. If |value| is missing,
    1. Let |ref| be DefaultValue(|elementtype|).
    1. If |ref| is error, throw a TypeError exception.
1. Otherwise,
    1. Let |ref| be ? ToWebAssemblyValue(|value|, |elementtype|).

So table.set(0, undefined) is the same call as table.set(0): it takes the first branch, and DefaultValue(funcref) is ref.null func, so the table entry becomes null without any exception.
The TypeError would only come from ToWebAssemblyValue(undefined, funcref) in the second branch, which this call never reaches.

Change

• Remove undefined from the list of invalid values in "Setting non-function".
• In "Arguments for anyfunc table set", check that table.set(0, undefined) indeed sets the the entry to null without throwing error.
• Adds new test case for Table.grow: it has the same signature: AddressValue grow(AddressValue delta, optional any value), and the same If |value| is missing step, but grow.any.js never passes an explicit undefined as the value. (The omitted case is already covered by "Basic i32")

Why the test may have been written this way

My guess is that "missing" was read in its ordinary English sense, as "not given".
Under that reading an explicit undefined is a "given" value, goes through ToWebAssemblyValue, and is rejected because it is not a function or null.

A concern

I am not sure this is the right direction though, because current engines already follow the test rather than the specification.
The latest versions of V8, SpiderMonkey, and JavaScriptCore all throw a TypeError for both table.set(0, undefined) and table.grow(2, undefined).

If want the current engines' behavior as the intended one, the alternative is to somehow change the specification instead, for example by bypassing the overload resolution algorithm and checking for undefined explicitly.

[f52985]

For reference: There already exists a test suite that treats an explicit undefined as a missing value.

Global constructor has the same optional any signature and the same "If |v| is missing" step as Table.set and Table.grow,
and global/constructor.any.js expects new WebAssembly.Global({value: "i64"}, undefined) to yield 0n ("Explicit value ${name} for type i64"),
which only holds if undefined becomes "missing" and takes DefaultValue(i64);
converting it with ToWebAssemblyValue would throw a TypeError from ToBigInt64.

[Ms2ger]

Thanks