An interesting subtle property of where clauses with super traits

undecidable instances, traits, super-traits, where-clause

2020-05-02 23:50:00 UTC, by Dimitri Sabadie — feed


Lately, I’ve been hacking on the next version of luminance, luminance-0.40. It should be out “soon-ish” but in the meantime, I’ve been struggling a bit with some highly and strongly typed code. I want to share something interesting I discovered with rustc. Especially, I haven’t seen a mention of that property in the book, so I’m happily sharing.

What is a super trait?

The weird situation is about super traits. A super trait is a trait that must be implemented for another trait to be usable, because it’s relied on. Traits can be thought of as constraints, so a super trait is a bit like a dependency when implementing a trait, and an implication when using a trait on which a super trait is declared. A trait can have zero to several super traits (added with the + operator). For instance, imagine you have a trait Alive:

trait Alive {
  fn get_health(&self) -> Health;
}

Now imagine you want a trait to move someone or something around:

trait Move: Alive {
  fn go(&mut self, direction: Direction);
}

Here, you can see that:

So now that we understand what super traits are, let’s get to the weird stuff.

The weird stuff

When you implement a trait which has a super trait, do you think that:

  1. Your implementation is valid when the super trait is implemented? After all, you could simply assume it is implemented. If it’s not, instances of your trait won’t be pickable.
  2. Or your implementation requires the super trait to be implemented?

The distinction is important. (1.) doesn’t require rustc to prove when implementing a trait that the super trait is implemented. That will be required when using the trait. With (2.), rustc will have to prove that the super trait is, first, implemented, before even considering the implementation of the trait you’re making.

Rust currently uses (2.). If you impl Move for Foo, impl Alive for Foo must be in scope for that implementor to be possible.

But… it would be interesting to actually have (1.). Imagine that you want to implement Move for something a bit complex, like Creature<T>, but not all Creature<T> are alive. Only a subset of them, and you can’t tell exactly when — i.e. you just cannot assume anything about T. So what are you going to write?

impl<T> Move for Creature<T> {
  fn go(&mut self, direction: Direction) {
    // …
  }
}

This code will not compile, because you haven’t implemented Alive for Creature<T>. Remember that the trait solver requires to prove super traits. However, and this is where all the interesting / weird stuff happens:

impl<T> Move for Creature<T> where Self: Alive {
  fn go(&mut self, direction: Direction) {
    // …
  }
}

This compiles. It compiles because the where clause tells rustc that your implementor will be valid if used with Creature<T>: Alive. The distinction is really subtle, but is, to my opinion, very powerful. With the where clause, you state that Move is implemented for any Creature<T> that is also Alive, but you don’t require them all to be Alive! You could implement Alive for a bunch of creatures, like Creature<Vampire> and Creature<CloudDog>.

So, I remember having read somewhere (maybe in some Rust book, but I’m not quite sure) that the where Self: _ clause was not really useful, but in our case, you can see that it allows to express a completely different semantics.

You could also have used Creature<T>: Alive in place of Self: Alive, as here, Self = Creature<T>.

A small disgression to Haskell land

In Haskell, that code requires to use UndecidableInstances. I don’t know exactly why, but GHC states that the constraint (Alive (Creature a)) is no smaller than the instance head (Move (Creature a)), and this is not permitted, as being undecidable. Enabling the UndecidableInstances GHC extension will make it possible to compile:

class Alive a where
  getHealth :: a -> Health

-- The super class is declared on the left side of the => (the parenthesis are
-- optional, but I’m used to put them all the time, as they are required when you
-- have more constraints).
class (Alive a) => Move a where
  go :: Direction -> a -> a

-- This instance requires UndecidableInstances to compile.
instance (Alive (Creature a)) => Move (Creature a) where
  go = -- …

instance Alive (Creature Vampire) where
  getHealth = -- …

instance Alive (Creature CloudDog) where
  getHealth = -- …

I’m not quite sure why this very exact situation requires UndecidableInstances though. In this case, it seems fine.

I hope you’ll have learned something with this small yet fun type theory candy. Keep the vibes!