Wrapping or unwrapping already-wrapped functions

[caitp]

Suppose you have the following scenario:

let r = new ShadowRealm;
let wrapped1 = r.evaluate("(function(a, b) { return a(b); })");
let wrapped2 = r.evaluate("(function() { return globalThis.specialSauce++; })"

let result = wrapped1(wrapped2, 7);

Invoking wrapped1 results in re-wrapping the already wrapped wrapped2, but did any wrapping really need to happen? Would it make sense to simply pass the unwrapped target function as a parameter?

This difference would be observable, but I don't think it enables an avenue of cross-realm information sharing, and it could provide some heap savings

[caitp]

Also, multiply-wrapped functions can be a pain to query, so it might be nice if we could guarantee that a wrapped function is at most singly wrapped

[mhofman]

Besides the identity observability, which is better served by symbols, what would be the benefit to specifying an unwrapping or limiting the number of wrapping? The current spec doesn't prevent implementations from optimizing the calls of multi-wrapped functions and in effect flatten invocations.

See #293 for previous discussion on the topic.

[caitp]

I do believe identity observability of functions created within the same shadow realm is more intuitive and easier to understand, and simpler to work with, for users.

But that isn't really the point, on the runtime side, I'm not a huge fan of allocating a new function wrapper every time you want to pass a function returned from a shadow realm, as an argument to a function within that shadow realm. I'm of the opinion that wrapping should only be done for cross-realm operations, but when passed a wrapper of a function originating in the same shadow realm, adding another layer of indirection doesn't make sense

[ljharb]

Can't the wrapper have properties added in the new realm?

[caitp]

Can't the wrapper have properties added in the new realm?

ShadowRealm provides a function to the Incubator realm (this is wrapped, the incubator realm can't operate on the inner function)
Incubator realm passes wrapped function back to ShadowRealm as an argument (This is unwrapped, the ShadowRealm doesn't see anything added to the wrapper object by the IncubatorRealm)

I don't see a way for properties on the wrapper to be exposed to the ShadowRealm, or properties on the original object to be exposed to the incubator realm or different ShadowRealm -- I think, as far as each realm is concerned, they all are operating with distinct objects representing the same underlying function, and there's no real mechanism for sharing information with it --- but as far as one realm is concerned, a reference to that realm's functions always matches the identity of that realm's functions

[ljharb]

ok - so currently, if i have function F in the main realm, and i pass it into the shadow realm, and the shadow realm passes it back as G, and another round trip gives me H. F !== G && F !== H && G !== H.

With your proposal, would G === H?

[mhofman]

Identity stability of function through the boundary is not a design goal of this API, and I believe the champions would be against it.

In particular passing the same function twice from realm A -> B would result in 2 different wrappers, and doing otherwise would require the implementation to keep a WeakMap of callables and their wrapper. Since there is already identity instability in that case, there is no reason to have identity stability in round trips.

[caitp]

If I understood your example:

let r = new ShadowRealm;
let wrap = r.evaluate("(function(f) { return f; })");

let f = function() {}
let g = wrap(f);
let h = wrap(g);

// f == g == h

[mhofman]

My point is that you already have the following:

let f = function() {};
let r = new ShadowRealm;
const g = r.evaluate("(function(f1, f2) { return f1 === f2; })");

assert(g(f, f) === false);

[ljharb]

@caitp and here:

let r = new ShadowRealm;
let wrap = r.evaluate("globalThis.cache ||= new Set(); (function(f) { globalThis.cache.add(f); return f; })");

let f = function() {}
let g = wrap(f);
let h = wrap(g);

r.evaluate('globalThis.cache.size'); // ?

what would be the result?

[caitp]

r.evaluate('globalThis.cache.size'); // ?

My thing here is concerned with the way things are treated which originate from and are returned to a particular realm --- those would preserve their identity, but wrappers would still be created when crossing the realm boundary, so the globalThis cache would have 2 items in it at the end, and you'd really want it to be a WeakMap 👀

[caitp]

My point is that you already have the following:

let f = function() {};
let r = new ShadowRealm;
const g = r.evaluate("(function(f1, f2) { return f1 === f2; })");

assert(g(f, f) === false);

I think this would still be the case with the idea I presented here, this doesn't change

[ljharb]

WeakMaps don't have a size, otherwise i'd have used that :-)

[caitp]

so, it would enable a level of querying of the unwrapped identity of an object that wouldn't be possible currently, I guess there are some security implications. They do seem like the sorts of things which have other solutions though

[caridy]

We have discussed the identity of wrapped functions extensibly in various forms. I will close this assuming that there are no more open questions.