final

The final attribute is the converse of the structural attribute. It configures how wasm-bindgen will generate JS imports to call the imported function. Notably a function imported by final never changes after it was imported, whereas a function imported by default (or with structural) is subject to runtime lookup rules such as walking the prototype chain of an object. Note that final is not suitable for accessing data descriptor properties of JS objects; to accomplish this, use the structural attribute.

The final attribute is intended to be purely related to performance. It ideally has no user-visible effect, and structural imports (the default) should be able to transparently switch to final eventually.

The eventual performance aspect is that with the component model proposal then wasm-bindgen will need to generate far fewer JS function shims to import than it does today. For example, consider this import today:

#![allow(unused)]
fn main() {
#[wasm_bindgen]
extern "C" {
    type Foo;
    #[wasm_bindgen(method)]
    fn bar(this: &Foo, argument: &str) -> JsValue;
}
}

Without the final attribute the generated JS looks like this:

// without `final`
export function __wbg_bar_a81456386e6b526f(arg0, arg1, arg2) {
    let varg1 = getStringFromWasm(arg1, arg2);
    return addHeapObject(getObject(arg0).bar(varg1));
}

We can see here that this JS function shim is required, but it's all relatively self-contained. It does, however, execute the bar method in a duck-type-y fashion in the sense that it never validates getObject(arg0) is of type Foo to actually call the Foo.prototype.bar method.

If we instead, however, write this:

#![allow(unused)]
fn main() {
#[wasm_bindgen]
extern "C" {
    type Foo;
    #[wasm_bindgen(method, final)] // note the change here
    fn bar(this: &Foo, argument: &str) -> JsValue;
}
}

it generates this JS glue (roughly):

const __wbg_bar_target = Foo.prototype.bar;

export function __wbg_bar_a81456386e6b526f(arg0, arg1, arg2) {
    let varg1 = getStringFromWasm(arg1, arg2);
    return addHeapObject(__wbg_bar_target.call(getObject(arg0), varg1));
}

The difference here is pretty subtle, but we can see how the function being called is hoisted out of the generated shim and is bound to always be Foo.prototype.bar. This then uses the Function.call method to invoke that function with getObject(arg0) as the receiver.

But wait, there's still a JS function shim here even with final! That's true, and this is simply a fact of future WebAssembly proposals not being implemented yet. The semantics, though, match the future component model proposal because the method being called is determined exactly once, and it's located on the prototype chain rather than being resolved at runtime when the function is called.

Interaction with future proposals

If you're curious to see how our JS function shim will be eliminated entirely, let's take a look at the generated bindings. We're starting off with this:

const __wbg_bar_target = Foo.prototype.bar;

export function __wbg_bar_a81456386e6b526f(arg0, arg1, arg2) {
    let varg1 = getStringFromWasm(arg1, arg2);
    return addHeapObject(__wbg_bar_target.call(getObject(arg0), varg1));
}

... and once the reference types proposal is implemented then we won't need some of these pesky functions. That'll transform our generated JS shim to look like:

const __wbg_bar_target = Foo.prototype.bar;

export function __wbg_bar_a81456386e6b526f(arg0, arg1, arg2) {
    let varg1 = getStringFromWasm(arg1, arg2);
    return __wbg_bar_target.call(arg0, varg1);
}

Getting better! Next up we need the component model proposal. Note that the proposal is undergoing some changes right now so it's tough to link to reference documentation, but it suffices to say that it'll empower us with at least two different features.

First, component model promises to provide the concept of "argument conversions". The arg1 and arg2 values here are actually a pointer and a length to a utf-8 encoded string, and with component model we'll be able to annotate that this import should take those two arguments and convert them to a JS string (that is, the host should do this, the WebAssembly engine). Using that feature we can further trim this down to:

const __wbg_bar_target = Foo.prototype.bar;

export function __wbg_bar_a81456386e6b526f(arg0, varg1) {
    return __wbg_bar_target.call(arg0, varg1);
}

And finally, the second promise of the component model proposal is that we can flag a function call to indicate the first argument is the this binding of the function call. Today the this value of all called imported functions is undefined, and this flag (configured with component model) will indicate the first argument here is actually the this.

With that in mind we can further transform this to:

export const __wbg_bar_a81456386e6b526f = Foo.prototype.bar;

and voila! We, with reference types and component model, now have no JS function shim at all necessary to call the imported function. Additionally future Wasm proposals to the ES module system may also mean that don't even need the export const ... here too.

It's also worth pointing out that with all these Wasm proposals implemented the default way to import the bar function (aka structural) would generate a JS function shim that looks like:

export function __wbg_bar_a81456386e6b526f(varg1) {
    return this.bar(varg1);
}

where this import is still subject to runtime prototype chain lookups and such.

  翻译：