Object.assign and __proto__ property

Nathan Wall nathan.wall at live.com
Mon Oct 21 19:28:43 PDT 2013

> Allen Wirfs-Brock wrote:
> Actually, this is a good point. As currently specified Object.assign
> of with an own __proto__ property on the RHS object will trigger a
> [[SetPrototypeOf]] on the LHS object. Is that what we want. It is a
> direct fallout of specifying Object.assign as the equivalent of a
> sequence of property assignments from the RHS to the LHS. "__proto__"
> could be special cased. But should it?
> Object.mixin doesn't have this issue because, as Rick points out, it
> uses GetOwnProperty/DefineProperty instead of Get/Set.
> Anybody want to argue that Object.assign shouldn't trigger
> [[SetPrototypeOf]]?
> Allen

Given that `__proto__` is defined as an accessor on `Object.prototype`, I think this should be expected. Any other setter defined on the object or its prototype chain will be invoked in the same manner. It's directly triggering `[[SetPrototypeOf]]`, it's just invoking the `__proto__` setter which happens to trigger `[[SetPrototypeOf]]`. Any number of accessor properties could do weird things:

    Object.defineProperty(Object.prototype, 'bloikle', {
        set(value) { Object.freeze(this); }

    // Somewhere else far off in the code...
    Object.assign(foo, { bloikle: 'bar' });

    // Now foo is strangely frozen...

For consistency, `__proto__` shouldn't be hidden from `Object.assign`.

Plus, there's no reason this shouldn't work:

    var info = JSON.stringify(
        '{ "__proto__": "A description of __proto__",' +
        '"hasOwnProperty": "A description of hasOwnProperty" }'

    var otherInfo = JSON.stringify(
        '{ "propertyIsEnumerable": "A description of propertyIsEnumerable" }'

    Object.assign(otherInfo, info);

Exploring other options.. A possible alternative consistent option would be to expose some mechanism to hide properties from `Object.assign`, so that the bloikle property could be hidden in the above... Doing this well would probably require a way to make it work cross-realm, which at first thought sounds pretty difficult. And it probably wouldn't play nice with the above `info` example.

I think I agree with Andrea on this.  Should `Object.assign` be left out of ES6 for libraries to implement? ... especially given that this `__proto__` dilemma feels "messy".


More information about the es-discuss mailing list