Add support for optional numbers

[afdw]

Fixes #624.

[alexcrichton]

Thanks for this!

It looks like the in-memory representation of optional types is a u32 indicating whether it's Some or None and then the payload itself? That seems plausible to me!

@fitzgen was experimenting using the "global stack" as well for the payload and/or discriminant, as this I believe would enable a blanket implementation for all Option<T> rather than needing a per-type solution (like we need today)

Another option I was thinking of is that if we want to go the super-specialized route we could support Option<T> for all T that aren't f64 or u64 by using f64 as the ABI value, choosing an out-of-range sentinel for None for each type (e.g. None::<u32> becomes NaN, None::<f32> becomes Some(f64::max_value()), etc)

I'm not sure what the best option is here, but I'm curious what @fitzgen thinks of this too!

[afdw]

Other approaches are possible, yes. But implementing manually means that it's possible to do it better, more compact (like I do with optional i8, u8, i16 and u16 types, using a single u32 argument/return type everywhere). And we don't have too many primitives.

Is passing (and converting) float significantly cheaper than passing 2 integers / integer and 32-bit float?

[alexcrichton]

True yeah! I personally like the idea of using more compact representations as well, and I think using u32 where possible is a great example of that.

In terms of cost I'm not actually sure, but I'd imagine that it's pretty negligible for wasm-bindgen's purposes

[alexcrichton]

Could this perhaps be named slightly differently to something like is_wasm_native_type or something like that?

[alexcrichton]

Could this perhaps be renamed to is_abi_as_u32?

[alexcrichton]

Nowadays all the \n\ markers here should be fine to clean up, the auto-formatting will strip them all away anyway!

[fitzgen]

I don't have any concrete work in progress, so don't block this on me!

Regarding blanket vs targeted impls: I also doubt that there is observable overhead to putting a discriminant on the global stack. But targeted impls mean that we have to write many impls instead of a single one. I suspect that we will want to move to a blanket impl eventually. But it is also an implementation detail that can be changed at a later date.

[afdw]

I think it's now implemented for all of the primitive types.

[alexcrichton]

Thanks so much @afdw!