RFC 3058: Try trait redesign

lang | libs (error-handling | types-libstd | traits)

Summary

Replace RFC #1859, try_trait, with a new design for the currently-unstable Try trait and corresponding desugaring for the ? operator.

The new design includes support for all intentional interconversions. It proposes removing the accidental interconversions, as a crater run demonstrated that would be feasible, however includes an alternative system that can support them as a low-support-cost edition mechanism if needed.

This is forward-looking to be compatible with other features, like try {} blocks or yeet e expressions or Iterator::try_find, but the statuses of those features are not themselves impacted by this RFC.

Motivation

The motivations from the previous RFC still apply (supporting more types, and restricted interconversion). However, new information has come in since the previous RFC, making people wish for a different approach.

This RFC proposes a solution that mixes the two major options considered last time.

Guide-level explanation

The ops::ControlFlow type

This is a simple enum:

enum ControlFlow<B, C = ()> {
    /// Exit the operation without running subsequent phases.
    Break(B),
    /// Move on to the next phase of the operation as normal.
    Continue(C),
}

It's intended for exposing things (like graph traversals or visitor) where you want the user to be able to choose whether to exit early. Using an enum is clearer than just using a bool -- what did false mean again? -- as well as allows it to carry a value, if desired.

For example, you could use it to expose a simple tree traversal in a way that lets the caller exit early if they want:

impl<T> TreeNode<T> {
    fn traverse_inorder<B>(&self, mut f: impl FnMut(&T) -> ControlFlow<B>) -> ControlFlow<B> {
        if let Some(left) = &self.left {
            left.traverse_inorder(&mut f)?;
        }
        f(&self.value)?;
        if let Some(right) = &self.right {
            right.traverse_inorder(&mut f)?;
        }
        ControlFlow::Continue(())
    }
}

Now, you could write the same thing with Result<(), B> instead. But that would require that the passed-in closure use Err(value) to early-exit the traversal, which can cause mental dissonance when that exit is because it successfully found the value for which it was looking. Using ControlFlow::Break(value) instead avoids that prejudice, the same way that break val in a loop doesn't inherently mean success nor failure.

The Try trait

The ops::Try trait describes a type's behavior when used with the ? operator, like how the ops::Add trait describes its behavior when used with the + operator.

At its core, the ? operator is about splitting a type into its two parts:

(Oxford's definition for a residual is "a quantity remaining after other things have been subtracted or allowed for", thus the use here.)

The Try trait also has facilities for rebuilding a type from either of its parts. This is needed to build the final return value from a function, both in ? and in methods generic over multiple types implementing Try.

Here's a quick overview of a few standard types which implement Try, their corresponding output and residual types, and the functions which convert between them. (Full details will come later; the goal for now is just to get the general idea.)

+-------------+                             +-------------------+                          +-------------------+
| Try::Output |                             |      Try Type     |                          |   Try::Residual   |
+-------------+   Try::branch is Continue   +-------------------+   Try::branch is Break   +-------------------+
|      T      |  <------------------------  |   Result<T, E>    |  --------------------->  |   Result<!, E>    |
|      T      |                             |     Option<T>     |                          |     Option<!>     |
|      C      |  ------------------------>  | ControlFlow<B, C> |  <---------------------  | ControlFlow<B, !> |
+-------------+      Try::from_output       +-------------------+    Try::from_residual    +-------------------+

If you've used ?-on-Result before, that output type is likely unsurprising. Since it's given out directly from the operator, there's not much of a choice.

The residual types, however, are somewhat more interesting. Code using ? doesn't see them directly -- their usage is hidden inside the desugaring -- so there are more possibilities available. So why are we using these ones specifically?

Most importantly, this gives each family of types (Results, Options, ControlFlows) their own distinct residual type. That avoids unrestricted interconversion between the different types, the ability to arbitrarily mix them in the same method. For example, like in the traversal example earlier, just because a ControlFlow::Break is also an early exit, that doesn't mean that it should be allowed to consider it a Result::Err -- it might be a success, conceptually. So by giving ControlFlow<X, _> and Result<_, X> different residual types, it becomes a compilation error to use the ? operator on a ControlFlow in a method which returns a Result, and vice versa. (There are also ways to allow interconversion where it's desirable between a particular pair of types.)

🏗️ Note for those familiar with the previous RFC 🏗️

This is the most critical semantic difference. Structurally this definition of the trait is very similar to the previous -- there's still a method splitting the type into a discriminated union between two associated types, and constructors to rebuild it from them. But by keeping the "result-ness" or "option-ness" in the residual type, it gives extra control over interconversion that wasn't possible before. The changes other than this are comparatively minor, typically either rearrangements to work with that or renamings to change the vocabulary used in the trait.

Using ! is then just a convenient yet efficient way to create those residual types. It's nice as a user, too, not to need to understand an additional type. Just the same "it can't be that one" pattern that's also used in TryFrom, where for example i32::try_from(10_u8) gives a Result<i32, !>, since it's a widening conversion which cannot fail. Note that there's nothing special going on with ! here -- any uninhabited enum would work fine.

How error conversion works

One thing The Book mentions, if you recall, is that error values in ? have From::from called on them, to convert from one error type to another.

The previous section actually lied to you slightly: there are two traits involved, not just one. The from_residual method is on FromResidual, which is generic so that the implementation on Result can add that extra conversion. Specifically, the trait looks like this:

trait FromResidual<Residual = <Self as Try>::Residual> {
    fn from_residual(r: Residual) -> Self;
}

And while we're showing code, here's the exact definition of the Try trait:

trait Try: FromResidual {
    type Output;
    type Residual;
    fn branch(self) -> ControlFlow<Self::Residual, Self::Output>;
    fn from_output(o: Self::Output) -> Self;
}

The fact that it's a super-trait like that is why I don't feel bad about the slight lie: Every T: Try always has a from_residual function from T::Residual to T. It's just that some types might offer more.

Here's how Result implements FromResidual to do error-conversions:

impl<T, E, F: From<E>> FromResidual<Result<!, E>> for Result<T, F> {
    fn from_residual(x: Result<!, E>) -> Self {
        match x {
            Err(e) => Err(From::from(e)),
        }
    }
}

But Option doesn't need to do anything exciting, so just has a simple implementation, taking advantage of the default parameter:

impl<T> FromResidual for Option<T> {
    fn from_residual(x: Self::Residual) -> Self {
        match x {
            None => None,
        }
    }
}

In your own types, it's up to you to decide how much freedom is appropriate. You can even enable interconversion by defining implementations from the residual types of other families if you'd like. But just supporting your one residual type is ok too.

🏗️ Note for those familiar with the previous RFC 🏗️

This is another notable difference: The From::from is up to the trait implementation, not part of the desugaring.

Implementing Try for a non-generic type

The examples in the standard library are all generic, so serve as good examples of that, but non-generic implementations are also possible.

Suppose we're working on migrating some C code to Rust, and it's still using the common "zero is success; non-zero is an error" pattern. Maybe we're using a simple type like this to stay ABI-compatible:

#[derive(Debug, Copy, Clone, Eq, PartialEq)]
#[repr(transparent)]
pub struct ResultCode(pub i32);
impl ResultCode {
    const SUCCESS: Self = ResultCode(0);
}

We can implement Try for that type to simplify the code without changing the error model.

First, we'll need a residual type. We can make this a simple newtype, and conveniently there's a type with a niche for exactly the value that this can't hold. This is only used inside the desugaring, so we can leave it opaque -- nobody but us will need to create or inspect it.

use std::num::NonZeroI32;
pub struct ResultCodeResidual(NonZeroI32);

With that, it's straight-forward to implement the traits. NonZeroI32's constructor even does exactly the check we need in Try::branch:

impl Try for ResultCode {
    type Output = ();
    type Residual = ResultCodeResidual;
    fn branch(self) -> ControlFlow<Self::Residual> {
        match NonZeroI32::new(self.0) {
            Some(r) => ControlFlow::Break(ResultCodeResidual(r)),
            None => ControlFlow::Continue(()),
        }
    }
    fn from_output((): ()) -> Self {
        ResultCode::SUCCESS
    }
}

impl FromResidual for ResultCode {
    fn from_residual(r: ResultCodeResidual) -> Self {
        ResultCode(r.0.into())
    }
}

Aside: As a nice bonus, the use of a NonZero type in the residual means that <ResultCode as Try>::branch compiles down to a nop on the current nightly. Thanks, enum layout optimizations!

Now, this is all great for keeping the interface that the other unmigrated C code expects, and can even work in no_std if we want. But it might also be nice to give other Rust code that uses it the option to convert things into a Result with a more detailed error.

For expository purposes, we'll use this error type:

#[derive(Debug, Clone)]
pub struct FancyError(String);

(A real one would probably be more complicated and have a better name, but this will work for what we need here -- it's bigger and needs non-core things to work.)

We can allow ? on a ResultCode in a method returning Result with an implementation like this:

impl<T, E: From<FancyError>> FromResidual<ResultCodeResidual> for Result<T, E> {
    fn from_residual(r: ResultCodeResidual) -> Self {
        Err(FancyError(format!("Something fancy about {} at {:?}", r.0, std::time::SystemTime::now())).into())
    }
}

The split between different error strategies in this section is inspired by windows-rs, which has both ErrorCode -- a simple newtype over u32 -- and Error -- a richer type that can capture a stack trace, has an Error trait implementation, and can carry additional debugging information -- where the former can be converted into the latter.

Using these traits in generic code

Iterator::try_fold has been stable to call (but not to implement) for a while now. To illustrate the flow through the traits in this RFC, let's implement our own version.

As a reminder, an infallible version of a fold looks something like this:

fn simple_fold<A, T>(
    iter: impl Iterator<Item = T>,
    mut accum: A,
    mut f: impl FnMut(A, T) -> A,
) -> A {
    for x in iter {
        accum = f(accum, x);
    }
    accum
}

So instead of f returning just an A, we'll need it to return some other type that produces an A in the "don't short circuit" path. Conveniently, that's also the type we need to return from the function.

Let's add a new generic parameter R for that type, and bound it to the output type that we want:

fn simple_try_fold_1<A, T, R: Try<Output = A>>(
    iter: impl Iterator<Item = T>,
    mut accum: A,
    mut f: impl FnMut(A, T) -> R,
) -> R {
    todo!()
}

Try is also the trait we need to get the updated accumulator from f's return value and return the result if we manage to get through the entire iterator:

fn simple_try_fold_2<A, T, R: Try<Output = A>>(
    iter: impl Iterator<Item = T>,
    mut accum: A,
    mut f: impl FnMut(A, T) -> R,
) -> R {
    for x in iter {
        let cf = f(accum, x).branch();
        match cf {
            ControlFlow::Continue(a) => accum = a,
            ControlFlow::Break(_) => todo!(),
        }
    }
    R::from_output(accum)
}

We'll also need FromResidual::from_residual to turn the residual back into the original type. But because it's a supertrait of Try, we don't need to mention it in the bounds. All types which implement Try can always be recreated from their corresponding residual, so we'll just call it:

pub fn simple_try_fold_3<A, T, R: Try<Output = A>>(
    iter: impl Iterator<Item = T>,
    mut accum: A,
    mut f: impl FnMut(A, T) -> R,
) -> R {
    for x in iter {
        let cf = f(accum, x).branch();
        match cf {
            ControlFlow::Continue(a) => accum = a,
            ControlFlow::Break(r) => return R::from_residual(r),
        }
    }
    R::from_output(accum)
}

But this "call branch, then match on it, and return if it was a Break" is exactly what happens inside the ? operator. So rather than do all this manually, we can just use ? instead:

fn simple_try_fold<A, T, R: Try<Output = A>>(
    iter: impl Iterator<Item = T>,
    mut accum: A,
    mut f: impl FnMut(A, T) -> R,
) -> R {
    for x in iter {
        accum = f(accum, x)?;
    }
    R::from_output(accum)
}

Reference-level explanation

ops::ControlFlow

#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum ControlFlow<B, C = ()> {
    /// Exit the operation without running subsequent phases.
    Break(B),
    /// Move on to the next phase of the operation as normal.
    Continue(C),
}

The traits

pub trait Try: FromResidual {
    /// The type of the value consumed or produced when not short-circuiting.
    type Output;

    /// A type that "colours" the short-circuit value so it can stay associated
    /// with the type constructor from which it came.
    type Residual;

    /// Used in `try{}` blocks to wrap the result of the block.
    fn from_output(x: Self::Output) -> Self;

    /// Determine whether to short-circuit (by returning `ControlFlow::Break`)
    /// or continue executing (by returning `ControlFlow::Continue`).
    fn branch(self) -> ControlFlow<Self::Residual, Self::Output>;
}

pub trait FromResidual<Residual = <Self as Try>::Residual> {
    /// Recreate the type implementing `Try` from a related residual
    fn from_residual(x: Residual) -> Self;
}

Expected laws

What comes out is what you put in:

You can recreate what you split up:

Desugaring ?

The previous desugaring of x? was

match Try::into_result(x) {
    Ok(v) => v,
    Err(e) => return Try::from_error(From::from(e)),
}

The new one is very similar:

match Try::branch(x) {
    ControlFlow::Continue(v) => v,
    ControlFlow::Break(r) => return FromResidual::from_residual(r),
}

The critical difference is that conversion (such as From::from) is left up to the implementation instead of forcing it in the desugar.

Standard implementations

Result

impl<T, E> ops::Try for Result<T, E> {
    type Output = T;
    type Residual = Result<!, E>;

    #[inline]
    fn from_output(c: T) -> Self {
        Ok(c)
    }

    #[inline]
    fn branch(self) -> ControlFlow<Self::Residual, T> {
        match self {
            Ok(c) => ControlFlow::Continue(c),
            Err(e) => ControlFlow::Break(Err(e)),
        }
    }
}

impl<T, E, F: From<E>> ops::FromResidual<Result<!, E>> for Result<T, F> {
    fn from_residual(x: Result<!, E>) -> Self {
        match x {
            Err(e) => Err(From::from(e)),
        }
    }
}

Option

impl<T> ops::Try for Option<T> {
    type Output = T;
    type Residual = Option<!>;

    #[inline]
    fn from_output(c: T) -> Self {
        Some(c)
    }

    #[inline]
    fn branch(self) -> ControlFlow<Self::Residual, T> {
        match self {
            Some(c) => ControlFlow::Continue(c),
            None => ControlFlow::Break(None),
        }
    }
}

impl<T> ops::FromResidual for Option<T> {
    fn from_residual(x: <Self as ops::Try>::Residual) -> Self {
        match x {
            None => None,
        }
    }
}

Poll

These reuse Result's residual type, and thus interconversion between Poll and Result is allowed without needing additional FromResidual implementations on Result.

impl<T, E> ops::Try for Poll<Result<T, E>> {
    type Output = Poll<T>;
    type Residual = <Result<T, E> as ops::Try>::Residual;

    fn from_output(c: Self::Output) -> Self {
        c.map(Ok)
    }

    fn branch(self) -> ControlFlow<Self::Residual, Self::Output> {
        match self {
            Poll::Ready(Ok(x)) => ControlFlow::Continue(Poll::Ready(x)),
            Poll::Ready(Err(e)) => ControlFlow::Break(Err(e)),
            Poll::Pending => ControlFlow::Continue(Poll::Pending),
        }
    }
}

impl<T, E, F: From<E>> ops::FromResidual<Result<!, E>> for Poll<Result<T, F>> {
    fn from_residual(x: Result<!, E>) -> Self {
        match x {
            Err(e) => Poll::Ready(Err(From::from(e))),
        }
    }
}
impl<T, E> ops::Try for Poll<Option<Result<T, E>>> {
    type Output = Poll<Option<T>>;
    type Residual = <Result<T, E> as ops::Try>::Residual;

    fn from_output(c: Self::Output) -> Self {
        c.map(|x| x.map(Ok))
    }

    fn branch(self) -> ControlFlow<Self::Residual, Self::Output> {
        match self {
            Poll::Ready(Some(Ok(x))) => ControlFlow::Continue(Poll::Ready(Some(x))),
            Poll::Ready(Some(Err(e))) => ControlFlow::Break(Err(e)),
            Poll::Ready(None) => ControlFlow::Continue(Poll::Ready(None)),
            Poll::Pending => ControlFlow::Continue(Poll::Pending),
        }
    }
}

impl<T, E, F: From<E>> ops::FromResidual<Result<!, E>> for Poll<Option<Result<T, F>>> {
    fn from_residual(x: Result<!, E>) -> Self {
        match x {
            Err(e) => Poll::Ready(Some(Err(From::from(e)))),
        }
    }
}

ControlFlow

impl<B, C> ops::Try for ControlFlow<B, C> {
    type Output = C;
    type Residual = ControlFlow<B, !>;

    fn from_output(c: C) -> Self {
        ControlFlow::Continue(c)
    }

    fn branch(self) -> ControlFlow<Self::Residual, C> {
        match self {
            ControlFlow::Continue(c) => ControlFlow::Continue(c),
            ControlFlow::Break(b) => ControlFlow::Break(ControlFlow::Break(b)),
        }
    }
}

impl<B, C> ops::FromResidual for ControlFlow<B, C> {
    fn from_residual(x: <Self as ops::Try>::Residual) -> Self {
        match x {
            ControlFlow::Break(r) => ControlFlow::Break(r),
        }
    }
}

Use in Iterator

The provided implementation of try_fold is already just using ? and try{}, so doesn't change. The only difference is the name of the associated type in the bound:

fn try_fold<B, F, R>(&mut self, init: B, mut f: F) -> R
where
    Self: Sized,
    F: FnMut(B, Self::Item) -> R,
    R: Try<Output = B>,
{
    let mut accum = init;
    while let Some(x) = self.next() {
        accum = f(accum, x)?;
    }
    try { accum }
}

Drawbacks

Rationale and alternatives

Why ControlFlow pulls its weight

The previous RFC discussed having such a type, but ended up deciding that defining a new type for the desugar wasn't worth it, and just used Result.

This RFC does use a new type because one already exists in nightly under the control_flow_enum feature gate. It's being used in the library and the compiler, demonstrating that it's useful beyond just this desugaring, so the desugar might as well use it too for extra clarity. There are also ecosystem changes waiting on something like it, so it's not just a compiler-internal need.

Methods on ControlFlow

On nightly there are a variety of methods available on ControlFlow. However, none of them are needed for the stabilization of the traits, so they left out of this RFC. They can be considered by libs at a later point.

There's a basic set of simple ones that could be included if desired, though:

impl<B, C> ControlFlow<B, C> {
    fn is_break(&self) -> bool;
    fn is_continue(&self) -> bool;
    fn break_value(self) -> Option<B>;
    fn continue_value(self) -> Option<C>;
}

Traits for ControlFlow

ControlFlow derives a variety of traits where they have obvious behaviour. It does not, however, derive PartialOrd/Ord. They're left out as it's unclear which order, if any, makes sense between the variants.

For Options, None < Some(_), but for Results, Ok(_) < Err(_). So there's no definition for ControlFlow that's consistent with the isomorphism to both types.

Leaving it out also leaves us free to change the ordering of the variants in the definition in case doing so can allow us to optimize the ? operator. (For a similar previous experiment, see PR #49499.)

Naming the variants on ControlFlow

The variants are given those names as they serve the same purpose as the corresponding keywords when used in Iterator::try_fold or Iterator::try_for_each.

For example, this (admittedly contrived) loop

let mut sum = 0;
for x in iter {
    if x % 2 == 0 { continue }
    sum += x;
    if sum > 100 { break }
    continue
}

can be written as

let mut sum = 0;
iter.try_for_each(|x| {
    if x % 2 == 0 { return ControlFlow::Continue(()) }
    sum += x;
    if sum > 100 { return ControlFlow::Break(()) }
    ControlFlow::Continue(())
});

(Of course, one wouldn't normally use the continue keyword at the end of a for loop like that, but I've included it here to emphasize that even the ControlFlow::Continue(()) as the final expression of the block it ends up working like the keyword would.)

Why ControlFlow has C = ()

The type that eventually became ControlFlow was originally added way back in 2017 as the internal-only type LoopState used to make some default implementations in Iterator easier to read. It had no type parameter defaults.

Issue #75744 in 2020 started the process of exposing it, coming out of the observation that Iterator::try_fold isn't a great replacement for the deprecated-at-the-time Itertools::fold_while since using Err for a conceptual success makes code hard to read.

The compiler actually had its own version of the type in librustc_data_structures at the time:

pub enum ControlFlow<T> {
    Break(T),
    Continue,
}

The compiler was moved over to the newly-exposed type, and that inspired the creation of MCP#374, TypeVisitor: use ops::ControlFlow instead of bool. Experience from that lead to flipping the type arguments in PR#76614 -- which also helped the original use cases in Iterator, where things like default implementation of find also want C = (). And these were so successful that it lead to MCP#383, TypeVisitor: do not hard-code a ControlFlow<()> result, having the visitors use ControlFlow<Self::BreakTy>.

As an additional anecdote that C = () is particularly common, Hytak mentioned the following on Discord in response to seeing a draft of this RFC:

i didn't read your proposal in depth, but this reminds me of a recursive search function i experimented with a few days ago. It used a Result type as output, where Err(value) meant that it found the value and Ok(()) meant that it didn't find the value. That way i could use the ? to exit early

So when thinking about ControlFlow, it's often best to think of it not like Result, but like an Option which short-circuits the other variant. While it can flow a Continue value, that seems to be a fairly uncommon use in practice.

Was this considered last time?

Interestingly, a previous version of RFC #1859 did actually mention a two-trait solution, splitting the "associated type for ok" and "generic type for error" like is done here. It's no longer mentioned in the version that was merged. To speculate, it may have been unpopular due to a thought that an extra traits just for the associated type wasn't worth it.

Current desires for the solution, however, have more requirements than were included in the RFC at the time of that version. Notably, the stabilized Iterator::try_fold method depends on being able to create a Try type from the accumulator. Including such a constructor on the trait with the associated type helps that separate trait provide value.

Also, ok-wrapping was decided in #70941, which needs such a constructor, making this "much more appealing".

Why not make the output a generic type?

It's helpful that type information can flow both ways through ?.

Similar scenarios exist for try, though of course they're not yet stable:

Why does FromResidual take a generic type?

The simplest case is that the already-stable error conversions require a generic somewhere in the error path in the desugaring. In the RFC #1859 implementation, that generic comes from using From::from in the desugaring.

However, more experience with trying to use Try for scenarios other than "the early exit is an error" have shown that forcing this on everything is inappropriate. ControlFlow, for example, would rather not have it, for the same kinds of reasons that return and break-from-loop don't implicitly call it. Option may not care, as it only ever gets applied for NoneNone, but that's not really a glowing endorsement.

But even for the error path, forcing From causes problems, notably because of its identity impl. anyhow's Error type, for example, doesn't implement std::error::Error because that would prevent it from being From-convertible from any E: std::error::Error type. The error handling project group under libs has experimented with a prototype toolchain with this RFC implemented, and is excited at the possibilities that could come from being free of this restriction:

my mind is exploding, the possibility of all error types implementing error the way they actually should has such massive implications for the rest of the error reporting stuff we've been working on

As a bonus, moving conversion (if any) into the FromResidual implementation may actually speed up the compiler -- the simpler desugar means generating less HIR, and thus less work for everything thereafter (up to LLVM optimizations, at least). The serde crate has their own macro for error propagation which omits From-conversion as they see a "significant improvement" from doing so.

Why not merge Try and FromResidual?

This RFC treats them as conceptually the same trait -- there are no types proposed here to implement FromResidual<_> which don't also implement Try -- so one might wonder why they're not merged into one Try<R>. After all, that would seem to remove the duplication between the associated type and the generic type, as something like

trait Try<Residual> {
    type Output;
    fn branch(self) -> ControlFlow<Residual, Self::Output>;
    fn from_residual(r: Residual) -> Self;
    fn from_output(x: Self::Output) -> Self;
}

This, however, is technically too much freedom. Looking at the error propagation case, it would end up calling both Try<?R1>::branch and Try<?R2>::from_residual. With the implementation for Result, where those inference variables go through From, there's no way to pick what they should be, similar to how .into().into() doesn't compile. And even outside the desugaring, this would make Try::from_output(x) no longer work, since the compiler would (correctly) insist that the desired residual type be specified.

And even for a human, it's not clear that this freedom is helpful. While any trait can be implemented weirdly, one good part of RFC #1859 that this one hopes to retain is that one doesn't need to know contextual information to understand what comes out of ?. Whereas any design that puts branch on a generic trait would mean it'd be possible for ? to return different things depending on that generic type parameter -- unless the associated type were split out into a separate trait, but that just reopens the "why are they different traits" conversation again, without solving the other issues.

Naming the ?-related traits and associated types

This RFC introduces the residual concept as it was helpful to have a name to talk about in the guide section. (A previous version proved unclear, perhaps in part due to it being difficult to discuss something without naming it.) But the fn branch(self) -> ControlFlow<Self::Residual, Self::Output> API is not necessarily obvious.

A different scheme might be clearer for people. For example, there's some elegance to matching the variant names by using fn branch(self) -> ControlFlow<Self::Break, Self::Continue>. Or perhaps there are more descriptive names, like KeepGoing/ShortCircuit.

As a sketch, one of those alternatives might look something like this:

trait Try: FromBreak {
    type Break;
    type Continue;
    fn branch(self) -> ControlFlow<Self::Break, Self::Continue>;
    fn from_continue(c: Self::Continue) -> Self;
}
trait FromBreak<B = <Self as Try>::Break> {
    fn from_break(b: B) -> Self;
}

However the "boring" Output name does have the advantage that one doesn't need to remember a special name, as it's the same as the other operator traits. (For precedent, it's Add::Output and Div::Output even if one could argue that Add::Sum or Div::Quotient would be more "correct", in a sense.)

ℹ Per feedback from T-libs, this is left as an unresolved question for the RFC, to be resolved in nightly.

Splitting up Try more

This RFC encourages one to think of a Try type holistically, as something that supports all three of the core operations, with expected rules between them.

That's not necessarily the way it should go. It could be different, like there's no guarantee that Add and AddAssign work consistently, nor that Add and Sub are inverses.

Notably, the this proposal has both an introduction rule (Try::from_output) and elimination rule (Try::branch), in the Gentzian sense, on the same trait. That means that an implementor will need to support both, which could restrict the set of type with which ? (and try and yeet) could be used.

One unknown question here is whether this is important for any FFI scenarios. Often error APIs come in pairs (like Win32's GetLastError and SetLastError), but some libraries may only give them out without allowing updating them to a custom value. It's unclear whether such a thing would want to be exposed as ? on some ZST, and thus would need a trait split to work, or whether it would be sufficient to load such things into a ?-supporting type where supporting from_residual would be simple.

In pure rust, one could also imagine types where it might be interesting to allow introduction rules but not elimination rules. With try blocks, one could perhaps have something like

let _: IgnoreAllErrors = try {
    foo()?;
    bar()?;
    qux()?;
};

which works by allowing from_residual from any Result<_, _>::Residual, as well as from_output from (). On such a type there's no real use in allowing ? on the result, but at the same time it wouldn't be a hardship to offer it.

The split currently in the proposal, though it's there for other reasons, would allow a small version of this: it would be possible to add an implementation like impl FromResidual<Result<!, !>> for (), which would allow code like u64::try_from(123_u16)? even in a method that returns unit. That has a number of issues, however, like only supporting -> () and not other things like -> i32 where one would probably also expect it to work, and it could not be a generic implementation without some form of specialization, as it would conflict with the desired implementation on Result. And even if it did work, it's not clear that allowing ? here is the clearest option -- other options such as an always_ok method on Result<T, !> might be superior anyway.

Another downside of the flexibility is that the structure of the traits would be somewhat more complicated.

The simplest split would just move each method to its own trait,

trait Branch {
    type Output;
    type Residual;
    fn branch(self) -> ControlFlow<Self::Residual, Self::Output>;
}
trait FromOutput {
    type Output;
    fn from_output(x: Self::Output) -> Self;
}
trait FromResidual<R> {
    fn from_residual(x: R) -> Self;
}

but that loses the desired property that the returned-by-? and expected-by-try types match for types which do implement both.

One way to fix that would be to add another trait for that associated type, perhaps something like

trait TryBase {
    type Output;
}
trait Branch: TryBase {
    type Residual;
    fn branch(self) -> ControlFlow<Self::Residual, Self::Output>;
}
trait FromOutput: TryBase {
    fn from_output(x: Self::Output) -> Self;
}
trait FromResidual<R> {
    fn from_residual(x: R) -> Self;
}

But this has still lost the simplicity of the R: Try bound for use in simple cases like try_fold. (And, in fact, all designs that allow types to choose them independently have that issue.) That may mean that it would also be useful to add yet another item, a trait alias to tie everything together in the "usual" way again. Perhaps it would look something like this:

trait Try = Branch + FromOutput + FromResidual<<Self as Branch>::Residual>;

There are probably also useful intermediary designs here. Perhaps the IgnoreAllErrors example above suggests that introduction on its own is reasonable, but elimination should require that both be supported. That's also the direction that would make sense for ? in infallible functions: it's absolutely undesirable for ()????? to compile, but it might be fine for all return types to support something like T: FromResidual<!> eventually.

ℹ Per feedback from T-libs, this is left as an unresolved question for the RFC, to be resolved in nightly.

Why a "residual" type is better than an "error" type

Most importantly, for any type generic in its "output type" it's easy to produce a residual type using an uninhabited type. That works for Option -- no NoneError residual type needed -- as well as for the StrandFail<T> type from the experience report. And thanks to enum layout optimizations, there's no space overhead to doing this: Option<!> is a ZST, and Result<!, E> is no larger than E itself. So most of the time one will not need to define anything additional.

In those cases where a separate type is needed, it's still easier to make a residual type because they're transient and thus can be opaque: there's no point at which a user is expected to do anything with a residual type other than convert it back into a known Try type. This is different from the previous design, where less-restrictive interconversion meant that anything could be exposed via a Result. That has lead to requests, such as for NoneError to implement Error, that are perfectly understandable given that the instances are exposed in Results. As residual types aren't ever exposed like that, it would be fine for them to implement nothing but FromResidual (and probably Debug), making them cheap to define and maintain.

Use of !

This RFC uses ! to be concise. It would work fine with convert::Infallible instead if ! has not yet stabilized, though a few more match arms would be needed in the implementations. (For example, Option::from_residual would need Some(c) => match c {}.)

Why FromResidual is the supertrait

It's nicer for try_fold implementations to just mention the simpler Try name. It being the subtrait means that code needing only the basic scenario can just bound on Try and know that both from_output and from_residual are available.

Default Residual on FromResidual

The default here is provided to make the basic case simple. It means that when implementing the trait, the simple case (like in Option) doesn't need to think about it -- similar to how you can impl Add for Foo for the homogeneous case even though that trait also has a generic parameter.

FromResidual::from_residual vs Residual::into_try

Either of these directions could be made to work. Indeed, an early experiment while drafting this had a method on a required trait for the residual that created the type implementing Try (not just the associated type). However that method was removed as unnecessary once from_residual was added, and then the whole trait was moved to future work in order to descope the RFC, as it proved unnecessary for the essential ?/try_fold functionality.

A major advantage of the FromResidual::from_residual direction is that it's more flexible with coherence when it comes to allowing other things to be converted into a new type being defined. That does come at the cost of higher restriction on allowing the new type to be converted into other things, but reusing a residual can also be used for that scenario.

Converting a known residual into a generic Try type seems impossible (unless it's uninhabited), but consuming arbitrary residuals could work -- imagine something like

impl<R: std::fmt::Debug> FromResidual<R> for LogAndIgnoreErrors {
    fn from_residual(h: H) -> Self {
        dbg!(h);
        Self
    }
}

(Not that that's necessarily a good idea -- it's plausibly too generic. This RFC definitely isn't proposing it for the standard library.)

And, ignoring the coherence implications, a major difference between the two sides is that the target type is typically typed out visibly (in a return type) whereas the source type (going into the ?) is often the result of some called function. So it's preferable for any behaviour extensions to be on the type that can more easily be seen in the code.

Can we just remove the accidental interconversions?

This depends on how we choose to read the rules around breaking changes.

A crater run on a prototype implementation found that some people are doing this. PRs have been sent to the places that broke, and generally it was agreed that removing the mixing improved the code:

Definitely a good change.

Thanks for spotting that, that was indeed a confusing mix

However another instance is in an abandoned project where the repository has been archived, so will not be fixed. And of course if it happened 3 times, there might be more instances in the wild.

The interesting pattern boils down to this:

.map(|v| Ok(something_returning_option(v)?))

That means it's using ? on an Option, but the closure ends up returning Result<_, NoneError> without needing to name the type as trait resolution discovers that it's the only possibility. It seems reasonable that this could happen accidentally while refactoring. That does mean, however, that the breakage could also be considered "allowed" as an inference change, and hypothetically additional implementations could make it ambiguous in the future. (It's like the normal AsRef breakage, and fits the pattern of "there's a way it could be written that works before and after", though in this case the disambiguated form requires naming an unstable type.)

This RFC thus proposes removing the accidental interconversions.

Compatibility with accidental interconversions (if needed)

If something happens that turns out they need to be supported, the following approach can work.

This would take a multi-step approach:

This keeps them from being visible in the trait system on stable, as FromResidual (the only form that would ever stabilize, or even be mentionable) would not include them.

mod sadness {
    use super::*;

    /// This includes all of the [`ops::FromResidual`] conversions, but
    /// also adds the two interconversions that work in 2015 & 2018.
    /// It will never be stable.
    pub trait FromResidualLegacy<R> {
        fn from_residual_legacy(r: R) -> Self;
    }

    impl<T: ops::FromResidual<R>, R> FromResidualLegacy<R> for T {
        fn from_residual_legacy(r: R) -> Self {
            <Self as ops::FromResidual<R>>::from_residual(r)
        }
    }

    /// This is a remnant of the old `NoneError` which is never going to be stabilized.
    /// It's here as a snapshot of an oversight that allowed this to work in the past,
    /// so we're stuck supporting it even though we'd really rather not.
    /// This will never be stabilized; use [`Option::ok_or`] to mix `Option` and `Result`.
    #[derive(Clone, Copy, PartialEq, PartialOrd, Eq, Ord, Debug, Hash)]
    pub struct LegacyNoneError;

    impl<T, E> ops::FromResidual<Option<!>> for Result<T, E>
    where
        E: From<LegacyNoneError>,
    {
        fn from_residual(x: Option<!>) -> Self {
            match x {
                None => Err(From::from(LegacyNoneError)),
            }
        }
    }


    #[unstable(feature = "try_trait_v2", issue = "42327")]
    impl<T> FromResidualLegacy<Result<!, LegacyNoneError>> for Option<T>
    {
        fn from_residual_legacy(_: Result<!, LegacyNoneError>) -> Self {
            None
        }
    }
}

Prior art

Previous approaches used on nightly

This is definitely monadic. One can define the basic monad operations for the Maybe monad as

use std::ops::Try;

fn monad_unit<T: Try>(x: <T as Try>::Ok) -> T {
    T::from_output(x)
}

fn monad_bind<T1: Try<Residual = R>, T2: Try<Residual = R>, R>(mx: T1, f: impl FnOnce(<T1 as Try>::Ok) -> T2) -> T2 {
    let x = mx?;
    f(x)
}

fn main() {
    let mx: Option<i32> = monad_unit(1);
    let my = monad_bind(mx, |x| Some(x + 1));
    let mz = monad_bind(my, |x| Some(-x));
    assert_eq!(mz, Some(-2));
}

However, like boats described for async.await, using monads directly isn't a great fit for rust. ? desugaring to a return (rather than closures) mixes better with the other control flow constructs, such as break and continue, that don't work through closures. And while the definitions above work fine for Option, they don't allow the error-conversion that's already stable with Result, so any monad-based implementation of ? wouldn't be able to be the normal monad structure regardless.

Unresolved questions

Questions from T-libs to be resolved in nightly:

Implementation and Stabilization Sequencing

Future possibilities

While it isn't directly used in this RFC, a particular residual type can be used to define a "family" of types which all share that residual.

For example, one could define a trait like this one:

pub trait GetCorrespondingTryType<TryOutputType>: Sized {
    /// The type from the original type constructor that also has this residual type,
    /// but has the specified Output type.
    type TryType: Try<Output = TryOutputType, Residual = Self>;
}

With corresponding simple implementations like these:

impl<T> GetCorrespondingTryType<T> for Option<!> {
    type TryType = Option<T>;
}

impl<C, B> ops::GetCorrespondingTryType<C> for ControlFlow<B, !> {
    type TryType = ControlFlow<B, C>;
}

And thus allow code to put whatever value they want into the appropriate type from the same family.

This can be thought of as the type-level inverse of Try's associated types: It splits them apart, and this puts them back together again.

(Why is this not written using Generic Associated Types (GATs)? Because it allows implementations to work with only specific types, or with generic-but-bounded types. Anything using it can bound to just the specific types needed for that method.)

A previous version of this RFC included a trait along these lines, but it wasn't needed for the stable-at-time-of-writing scenarios. Furthermore, some experiments demonstrated that having a bound in Try requiring it (something like where Self::Residual: GetCorrespondingTryType<Self::Output>) wasn't actually even helpful for unstable scenarios, so there was no need to include it in normative section of the RFC.

Possibilities for try_find

Various library methods, such as try_map for arrays (PR #79713), would like to be able to do HKT-like things to produce their result types. For example, Iterator::try_find wants to be able to return a Foo<Option<Item>> from a predicate that returned a Foo<bool>.

That could be done with an implementation such as the following:

fn try_find<F, R>(
    &mut self,
    f: F,
) -> <R::Residual as ops::GetCorrespondingTryType<Option<Self::Item>>>::TryType
where
    Self: Sized,
    F: FnMut(&Self::Item) -> R,
    R: ops::Try<Output = bool>,
    R::Residual: ops::GetCorrespondingTryType<Option<Self::Item>>,
{
    #[inline]
    fn check<F, T, R>(mut f: F) -> impl FnMut((), T) -> ControlFlow<Result<T, R::Residual>>
    where
        F: FnMut(&T) -> R,
        R: Try<Output = bool>,
    {
        move |(), x| match f(&x).branch() {
            ControlFlow::Continue(false) => ControlFlow::Continue(()),
            ControlFlow::Continue(true) => ControlFlow::Break(Ok(x)),
            ControlFlow::Break(r) => ControlFlow::Break(Err(r)),
        }
    }

    match self.try_fold((), check(f)) {
        ControlFlow::Continue(()) => Try::from_output(None),
        ControlFlow::Break(Ok(x)) => Try::from_output(Some(x)),
        ControlFlow::Break(Err(r)) => <_>::from_residual(r),
    }
}

Similarly, it could allow Try to automatically provide an appropriate map method:

fn map<T>(self, f: impl FnOnce(Self::Output) -> T) -> <Self::Residual as GetCorrespondingTryType<T>>::TryType
where
    Self::Residual: GetCorrespondingTryType<T>,
{
    match self.branch() {
        ControlFlow::Continue(c) => Try::from_output(f(c)),
        ControlFlow::Break(r) => FromResidual::from_residual(r),
    }
}

Possibilities for try{}

A core problem with try blocks as implemented in nightly, is that they require their contextual type to be known.

That is, the following never compiles, no matter the types of x and y:

let _ = try {
    foo(x?);
    bar(y?);
    z
};

This usually isn't a problem on stable, as the ? usually has a contextual type from its function, but can still happen there in closures.

But with something like GetCorrespondingTryType, an alternative desugaring becomes available which takes advantage of how the residual type preserves the "result-ness" (or whatever-ness) of the original value. That might turn the block above into something like the following:

fn helper<C, R: GetCorrespondingTryType<C>>(r: R) -> <R as GetCorrespondingTryType<C>>::TryType
{
    FromResidual::from_residual(h)
}

'block: {
    foo(match Try::branch(x) {
        ControlFlow::Continue(c) => c,
        ControlFlow::Break(r) => break 'block helper(r),
    });
    bar(match Try::branch(y) {
        ControlFlow::Continue(c) => c,
        ControlFlow::Break(r) => break 'block helper(r),
    });
    Try::from_output(z)
}

(It's untested whether the inference engine is smart enough to pick the appropriate C with just that -- the Output associated type is constrained to have a Continue type matching the generic parameter, and that Continue type needs to match that of z, so it's possible. But hopefully this communicates the idea, even if an actual implementation might need to more specifically introduce type variables or something.)

That way it could compile so long as the TryTypes of the residuals matched. For example, these uses in rustc would work without the extra annotation.

Now, of course that wouldn't cover anything. It wouldn't work with anything needing error conversion, for example, but annotation is also unavoidable in those cases -- there's no reasonable way for the compiler to pick "the" type into which all the errors are convertible.

So a future RFC could define a way (syntax, code inspection, heuristics, who knows) to pick which of the desugarings would be best. (As a strawman, one could say that try { ... } uses the "same family" desugaring whereas try as anyhow::Result<_> { ... } uses the contextual desugaring.) This RFC declines to debate those possibilities, however.

Note that the ? desugaring in nightly is already different depending whether it's inside a try {} (since it needs to block-break instead of return), so making it slightly more different shouldn't have excessive implementation cost.

Possibilities for yeet

As previously mentioned, this RFC neither defines nor proposes a yeet operator. However, like the previous design could support one with its Try::from_error, it's important that this design would be sufficient to support it.

yeet is a bikeshed-avoidance name for throw/fail/raise/etc, used because it definitely won't be the final keyword.

Because this "residual" design carries along the "result-ness" or "option-ness" or similar, it means there are two possibilities for a desugaring.

These have various implications -- like yeet None/yeet, yeet Err(ErrorKind::NotFound)/yeet ErrorKind::NotFound.into(), etc -- but thankfully this RFC doesn't need to discuss those. (And please don't do so in the GitHub comments either, to keep things focused, though feel free to start an IRLO or Zulip thread if you're so inspired.)