Distributions without direct dynamic dispatch

[dhardy]

This is #256 without #244, i.e. Distribution::sample takes rng: &mut R where R: Rng, not R: Rng + ?Sized.

The advantage of this approach is that it doesn't require #244, which in some ways is less ergonomic for users.

The disadvantages are

1. less ergonomic support for using trait object Rngs
2. if we wanted to add the equivalent of Add trait SampleRng: Rng and related doc #244 + Replace distribution::Sample with Distribution + polymorphism over Rng #256 later, it would break all existing Distribution implementations

[vks]

What about platforms with 16 bit pointers? Does Rust support them?

[dhardy]

I don't think so. There was some discussion of 16-bit CPUs here but ultimately because of the extremely limited memory addressing I think most 16-bit code has to be written expressly for that architecture.

[vks]

Looks good to me!

The advantage of this approach is that it doesn't require #244, which in some ways is less ergonomic for users.

Could you give some example user code contrasting the two approaches?

[dhardy]

Could you give some example user code contrasting the two approaches?

With this PR:

use rand::{thread_rng, Rng};
use rand::distributions::Uniform;

let mut rng = thread_rng();
let mut erased_rng: &mut Rng = &mut rng;
let val: f32 = (&mut erased_rng).sample(Uniform);
println!("f32 from [0,1): {}", val);

With #256 instead:

use rand::{thread_rng, Rng, SampleRng};
use rand::distributions::Uniform;

let mut rng = thread_rng();
let mut erased_rng: &mut Rng = &mut rng;
let val: f32 = erased_rng.sample(Uniform);
println!("f32 from [0,1): {}", val);

As you can see, very little difference — the extra reference to erased_rng vs the extra import (SampleRng). The most significant user difference perhaps is that the former solution is not obvious.

[vks]

Thanks! Why is the extra &mut needed? What is the compiler error if you miss it?

Edit: I missed what you said in #244:

rng.gen() will not work with rng: &mut Rng, although (&mut rng).gen() appears to work — I don't understand why the latter works

[dhardy]

See this comment.

From the error message one wouldn't expect to be able to use generic methods on a type-erased Rng at all; we can add documentation there but of course people may miss it.

        error: the `sample` method cannot be invoked on a trait object
 --> src/distributions/mod.rs:9:27
  |
9 | let val: f32 = erased_rng.sample(Uniform);
  |                           ^^^^^^

[vks]

If I understand it correctly, this is caused by this:

impl<'a> Rng for &'a mut Rng { ... }

Why is this impl necessary?

[dhardy]

Without that impl, the indirect dispatch (&mut rng).gen() won't work either with type-erased Rng.

[burdges]

Also, one writes struct Foo<R: Rng> { rng: R, ... } like everywhere, specifically so that Foo<SomeRng> and Foo<&mut SomeRng> both work.

[dhardy]

This is obsolete now that I've merged #265.