Tracking Issue for `ops::Residual` (feature `try_trait_v2_residual`)

[scottmcm]

Feature gate: #![feature(try_trait_v2_residual)]

This is a tracking issue for the ops::Residual trait. cc rust-lang/rfcs#3721

This is used by try_* APIs that need to change from one member of a family to another, such as

• https://doc.rust-lang.org/nightly/std/array/fn.try_from_fn.html Tracking issue for try_array_from_fn #89379
• https://doc.rust-lang.org/nightly/std/iter/trait.Iterator.html#method.try_find Tracking Issue for Iterator::try_find #63178
• https://doc.rust-lang.org/nightly/std/primitive.array.html#method.try_map Tracking issue for array::try_map #79711

For example, the closure passed to Iterator::try_find returns Foo<bool>, but the method wants to be able to return Foo<Option<<Self as Iterator>::Item>>.

Public API

// ops::Residual

trait Residual<O> {
    type TryTrait: Try<Output = O, Residual = Self>;
}

// with implementations for `Result`, `Option`, and `ControlFlow`.

Steps / History

• Implementation: Make array::{try_from_fn, try_map} and Iterator::try_find generic over Try #91286
• Final comment period (FCP)
• Stabilization PR

Unresolved Questions

• If Try::Residual is renamed (an open question in Tracking Issue for try_trait_v2, A new design for the ? desugaring (RFC#3058) #84277), this this probably should be too.

[BGR360]

@rustbot label +F-try_trait_v2

Now that GATs are stable, would it be possible to move the generic to the associated type? It would eliminate the need for ChangeOutputType and simplify usage.

trait Residual {
    type TryType<O>: Try<Output = O, Residual = Self>;
}

// core::array

fn try_from_fn<T, const N: usize, R, F>(f: F) -> R::TryType<[T; N]>
where
    R: Residual,
    F: FnMut(usize) -> R::TryType<T>,
{ .. }

// core::array

fn try_map<U, R, F>(self, f: F) -> R::TryType<[U; N]>
where
    R: Residual,
    F: FnMut(T) -> R::TryType<U>,
{ .. }

[danielhenrymantilla]

Yeah, the current Residual trait reads rather awkwardly due to that generic O parameter it must be fed eagerly:

fn try_from_fn<R, const N: usize, F>(cb: F)
  -> ChangeOutputType<R, [R::Output; N]>
where
    F : FnMut(usize) -> R,
    R : Try,
    R::Residual : Residual<[R::Output; N]>,

Notice that : Residual<[R::Output; N]> clause. It can't really be spelled out/read out loud in English: "is a residual of an array / can residual an array" doesn't make that much sense.

In a way, the type alias internally used by the stdlib already provides a way nicer name: ChangeOutputType.

So in this case, taking that name and using it on the trait itself, we would rather imagine:

R::Residual : CanWrapOutputType<[R::Output; N]>,

• (or {Feed,Apply,Change,Set,With}OutputType)

Now we get something that does read better, although it talks a bit too much about type-level operations, which is something the stdlib doesn't do that often.

So, rather than focusing on the how, if we focus on the what instead, we can stick to the "being a Residual" property, but this time with no generic Output parameter yet, by delegating it to a GAT, which features the proper quantification of "it can be fed any output type":

R::Residual : Residual, // <- Bonus: this wouldn't even be needed, since now we could eagerly add this bound to `Try`!

and then using <R::Residual as Residual>::TryType<[T; N]>.

• Or even TryTypeWithOutput<[T; N]>. I'll be using this new name for the remainder of the post

Notice that bonus of being able to eagerly require that Try::Residual types always implement Residual, which we can't do with the current design since that would need a for<Output> kind of quantification. EDIT: not all Residuals / Try types may want to be able to wrap any kind of T, as pointed out by @h4x5 in the comment just below.

I think it would be confusing to have some Try types not be usable with array_from_fn just because of the Residual associated type happens not to meet part of the implicitly required API contract. Having it explicitly required seems like a definite win, in that regard.

---

A technical remark, however, @h4x5: we can't use Res : Residual and then an impl Fn… -> Res::TryTypeWithOutput<T> closure, and then expect Res to be inferrable from context. Indeed, we'd have a "preïmage" situation regarding the type FeedT<Res> = Res::TryType<T>; operation, which is not an injective one, and thus won't be solvable by type inference.

So, while keeping Res seems convenient for the sake of the signature (Res::TryType<…>), we'd have to instead say that the output of the closure implements Try<Residual = R>:

fn try_from_fn<T, const N: usize, F, Ret, Res>(f: F)
  -> Res::TryTypeWithOutput<[T; N]>
where
    F : FnMut(usize) -> Ret,
    Ret : Try<Output = T, Residual = Res>,
    Res : Residual, // EDIT

• Notice how now we are getting the associated "image"/type from Ret to Ret::Residual in order to figure out Res.

Or we could get rid of that Res altogether:

fn try_from_fn<T, const N: usize, R, F>(f: F)
// -> <R::Residual as Residual>::TryTypeWithOutput<[T; N]>
   -> ChangeOutputType<R, [T; N]>
where
    F : FnMut(usize) -> R,
    R : Try<Output = T>,
    R::Residual : Residual, // EDIT

Thanks! I'll try and draft a PR today to change it.

Side note:
This trait really only applies to the generic container Try types (Result, Option, ControlFlow, etc), not to the simpler, non-generic ones like the RFC's ResultCode example. Thus, adding Try::Residual: Residual would be overly restrictive and disallow using try with types like ResultCode.

[danielhenrymantilla]

Oh, good catch! (I've edited my post accordingly).

• _{^{🤔 In that case I very much feel like that trait ought not to be named Residual. Given R : Try, then having R::Residual : Residual not hold by default and have to be explicitly spelled out feels quite weird, tbh. We'd be back to ChangeOutput or things along those lines, etc.}}

I've submitted #104128, an attempt at changing this.