Associated error type for RngCore

[dhardy]

Note: PR is built on top of #306 code.

Interesting idea, but one major problem comes up: implementation-agnostic dynamic dispatch is impossible, i.e. no full type erasure:

// we are required to specify the Error type here:
let mut r = &mut rng as &mut RngCore<Error=Void>;

@burdges I don't think this is avoidable?

Possibly we could try something like this I suppose:

impl<R: RngCore> RngCore<Error=Error> for R
where Error: From<<R as RngCore>::Error> {...}

[dhardy]

No, we can't use that impl, because it would require R to implement RngCore multiple times.

We may instead be able to use type parameters, but I'm not sure I want to go there:

impl<E, R: RngCore<E>> RngCore<Error> for R where Error: From<E> {...}

[burdges]

It might be because &mut RngCore<Error=Void> is a trait object, which may require all types to be known sooner. It possible monomorphisation can infer more if you avoid the trait object. It appears you only use the trait object in tests though, so maybe no big deal.

[dhardy]

It would be nice if it is still possible for users to write a function like this, though maybe the idea is overrated, I don't know:

// accepts any RNG type
fn foo(rng: &mut RngCore);

[burdges]

We encountered this issue with associated error types not being so ergonomic for trait objects before I suppose. And it likely contributed to the "just panic" decision before.

[dhardy]

Okay, I'm going to close it. It isn't like there wasn't already a huge amount of thought put into the design of RngCore. I do wonder whether specialisation or negative type bounds would solve the problems with this design, but lets not wait forever on Rustc features.