Remove Scene/SceneList impls for unit (), replace with unit struct

[laundmo]

Objective

Having Scene and SceneList implemented for () (unit) can be a source of confusion if users forget a ;. See this already happening before bsn is released here: https://discord.com/channels/691052431525675048/691052431974465548/1498960812662849637
The intention of this is to cause an error when ; is accidentally used at the end of a expression which evaluates to impl Scene/impl SceneList

Solution

The issue was that () is both the value any expression ending in ; evaluates to, and the value for empty the corresponding traits were implemented for by all_tuples!

By raising the start of all_tuples! to 1 instead of 0, and implementing SceneList for a new unit struct, the implementation for () can be avoided. Luckily, everywhere where an empty Scene/SceneList was used in at least Bevy itself, it was done through bsn_list!() which ultimately created the () by calling out to bevy_scene::macro_utils::auto_nest_tuple!, allowing me to replace the () there with the new unit struct.

This means, empty BSN scenes and scene lists now consist of the new EmptyTuple unit struct instead of (), and an accidental ; now raises "error[E0277]: the trait bound (): bevy_scene::Scene is not satisfied"

Testing

• cargo test
• bevy_city example with a bsn_list!() changed to empty
• feathers_gallery example with a bsn_list!() changed to empty
• test with crates already using main?

[alice-i-cecile]

Comments here about why this exists please. This is going to be very mysterious in the code in 5 years!

[laundmo]

fair, will do

[alice-i-cecile]

It might be nice to add a compile failure test for this + the ; case, but I won't block on it.

[alice-i-cecile]

On board with this idea: that's a major footgun that we should avoid.

[cart]

I don't particularly like this pattern. () is semantically an empty tuple and Scene is implemented for tuples.

Having () represent an empty Scene is also useful:

• You can assign it in places like type Scene = ();. Because Scene is implemented for tuples, this is natural.
• It lets you do fn scene() -> impl Scene {} as a form of "todo". Given that fn scene() -> impl Scene{ todo!() } doesn't work anymore (thanks to some semi-recent rust changes), I quite like having this as an option.

A custom EmptyTuple type also just introduces a form of "jank" to the API that makes everything just a little bit harder to understand (as a reader of docs, an author of internals, and a consumer of public APIs).

Yes, bsn! papers over all of this for most people. But Scene is intended to be an API that is usable on its own.

[cart]

I like that you can do this:

fn scene() -> impl Scene {
  ()
}

And then expand into this:

fn scene() -> impl Scene {
  (
    Sprite::patch(|value, context| { value.image = "sprite.png".into() }),
    Transform::patch(|value, context| { }),
  )
}

[laundmo]

So your solution to the issue of accidental ; is to keep it, or are you proposing that theres some other solution besides something like this?

i should have mentioned: the motivation for this was https://discord.com/channels/691052431525675048/691052431974465548/1498960812662849637

[laundmo]

There really is a core tension in Rust that () stands both for "empty tuple" and "no return"

[cart]

Yeah my argument is "yeah this is annoying, but it is a behavior inherent to Rust semantics". This is a problem everywhere the "trait implememented for tuples" pattern is used:

fn my_bundle() -> impl Bundle {
  (Transform::default(), Name::new("hello"));
}

It is also true for any trait that () has an impl for:

fn some_default() -> impl Default {
    vec![10usize];
}