Outer alias of 0 to impure definitions

[bvisness]

The component model explainer has this to say about outer aliases, emphasis mine:

Because components, like modules, are pure values, outer aliases that reach across component boundaries are restricted to only refer to pure definitions: modules, components, and types that do not transitively refer to a resource type.

It also makes it clear, though, that 0 is a valid instance index for an outer alias:

In particular, the first u32 can be 0, in which case the outer alias refers to the current component/type.

These two together would seemingly imply that it is valid to do the following:

(component
  (core module $m (func (export "foo")))
  (alias core export $m "foo" (core func $foo))
  (alias outer 0 $foo (core func $fooAgain))
)

But wasm-tools rejects this:

error: unexpected token, expected `module` or `type`
     --> stoopid.wat:4:29
      |
    4 |   (alias outer 0 $foo (core func $fooAgain))
      |                             ^

It seems like wasm-tools has restricted all outer aliases to pure definitions, not just those that reach across component boundaries. Is this a bug in wasm-tools or the component spec?

[bvisness]

For what it's worth, I don't even understand why you would need outer aliases with a depth of zero. Their existence is kind of a plague upon my component code. If the answer to this question was "forbid zero", I would be very happy. If the answer was instead "zero is allowed but outer aliases can only refer to pure definitions", I would be slightly less happy but still happy.

Otherwise I see absolutely no point in allowing arbitrary copies of items you already have in your current index space. It's a pain to validate, causes unnecessary indirection, and I can't imagine why any tool would ever want to generate one.

[alexcrichton]

Currently the binary format validation notes:

For outer aliases, validation restricts the sort to one of type, module or component.

wasm-tools isn't quite compliant here where this is rejected during parsing rather than validation so the error is a bit wonky

Otherwise though I don't think wit-component uses index 0 right now so this might be possible to tweak if we want.

[alexcrichton]

Oh oops sorry I thought this was in the component-model repository, apologies!

cc @lukewagner on this as well, do you have thoughts on this perhaps?

[lukewagner]

IIRC, index=0/same-component outer aliases were not originally allowed, but someone working on producer tooling had a case where it'd be useful to be able to add a second index to an existing definition (not sure why). It might be gone now, so if we're not using it anywhere and it creates trouble, I'd be happy to re-disallow it. But if we keep it, it also makes sense not to special case the index=0 case and have it follow the same "can only refer to pure definitions" rule as the index>1 cases.

[alexcrichton]

The easiest thing for now might be to tweak the spec wording to refer to only pure definitions, even for index=0. I don't personally have much of an opinion about allowing 0 or not myself.

[lukewagner]

Ok, filed component-model/#675 which should hopefully clarify (I had forgotten there are several different cases to consider here, but all do in fact reject func). (The index=0 case is kept.)

[alexcrichton]

Thanks! I'll close this on this side then since I believe wasm-tools is already in-line with that change