Enum variant types

[varkor]

Enum variants are to be considered types in their own rights. This allows them to be irrefutably matched upon. Where possible, type inference will infer variant types, but as variant types may always be treated as enum types this does not cause any issues with backwards-compatibility.

enum Either<A, B> { L(A), R(B) }

fn all_right<A, B>(b: B) -> Either<A, B>::R {
    Either::R(b)
}

let Either::R(b) = all_right::<(), _>(1729);
println!("b = {}", b);

Rendered

Thanks to @Centril for providing feedback on this RFC!

[alexreg]

Great work, @varkor. I've been looking forward to this for a long time. Just as a side-point, I'd love to follow this up with an RFC for the ideas in https://internals.rust-lang.org/t/pre-rfc-using-existing-structs-and-tuple-structs-as-enum-variants/7529 once this gets implemented in nightly (or perhaps even before). Since you've worked on this, would appreciate your thoughts at some point.

[bchallenor]

Although sum types are becoming increasingly common in programming languages, most do not choose to allow the variants to be treated as types in their own right (that is, the author has not found any that permit this design pattern).

This is possible in Scala - as in your example, Left and Right are subtypes of Either, and can be referred to independently. Coming from Scala, I miss this feature in Rust, and I am fully in favour of this RFC.

[varkor]

This is possible in Scala - as in your example, Left and Right are subtypes of Either, and can be referred to independently.

Ah, great, I'll add that in, thanks!

[Ixrec]

I think I'm in favor of the proposed functionality and semantics here. Where I'm stumbling is the nomenclature/terminology/teachability(?); it's not clear to me that "introducing a new kind of type: variant types" is the best description of this. In particular, precisely because this proposal feels so lightweight compared to previous ones, it doesn't really "feel" like what we're doing is adding a whole new type kind the way structural records or anonymous enums would be doing. It sounds like it could be equally well described as doing the "duplicating a variant as a standalone struct" workaround automagically, so those extra structs are just always there (except they get a specific layout guarantee and different conversion syntax that regular structs wouldn't get). Is there some detail I overlooked that makes this clearly not a sugar?

I'm guessing this is at least partially ignorance on my part because

The loose distinction between the enum type and its variant types could be confusing to those unfamiliar with variant types.

makes it sound like "variant types" are an actual thing with their own special properties that no other kinds of types have, and I just have no idea what that would be (since being autogenerated, having a certain layout guarantee and different conversion syntax seem like "surface level" properties that aren't really part of the type system per se). Maybe I just need to see some more examples of how these types behave?

[leonardo-m]

Nice RFC.

In all cases, the most specific type (i.e. the variant type if possible) is chosen by the type inference.

Is code like this still allowed, or is the compiler going to tell me that the Sum::B(b) branch of the match is impossible and needs to be removed?

enum Sum { A(u8), B(u8) }
let x = Sum::A(5); // x: Sum::A
match x {
    Sum::A(a) => {},
    Sum::B(b) => {},
}

Both options have advantages and disadvantages.

[varkor]

Is code like this still allowed, or is the compiler going to tell me that the Sum::B(b) branch of the match is impossible and needs to be removed?

This is a good question — I'll make note of it in the RFC. Although matching on variant types permits irrefutable matches, it must also accept the any other variants with the same type — otherwise it's not backwards compatible with existing code.

Where I'm stumbling is the nomenclature/terminology/teachability(?); it's not clear to me that "introducing a new kind of type: variant types" is the best description of this.

since being autogenerated, having a certain layout guarantee and different conversion syntax seem like "surface level" properties that aren't really part of the type system per se

It's quite possible there's a better way to explain this. They are essentially as you say, though they act slightly differently from structs (on top of the points you made) in the way they are pattern-matched (as above in this comment) and their discriminant value. I thought it would be clearer to describe them as an entirely new kind of type, but perhaps calling them special kinds of structs would be more intuitive as you say. I'll think about how to reword the relevant sections.

[nrc]

This was previously proposed in #1450. That was postponed because we were unsure about the general story around type fallback (e.g, integer types, default generic types, etc). Enum variants would add another case of this and so we wanted to be certain that the current approach is good and there are no weird interactions. IIRC, there was also some very minor backwards incompatibility.

This RFC should address those and issues, and summarise how this RFC is different to #1450.

For the sake of completeness, an alternative might be some kind of general refinement type, though I don't think that is a good fit with Rust.

I'm still personally very strongly in favour of this feature! The general mood on #1405 was also positive.

[eddyb]

For the sake of completeness, an alternative might be some kind of general refinement type, though I don't think that is a good fit with Rust.

I'm not sure that would need to be at odds with variant types, if Rust ends up with refinement types I expect variant types to be refinements of their enum.

[Centril]

First, irrespective of what happens with the RFC;
I always appreciate the effort put into well thought out RFCs and this is one of those, so thank you!

I am of two minds and a bit torn about the proposal here.

1. I think it would help immensely to make an API such as syn::Expr more ergonomic and avoid auxiliary structs such as syn::ExprBox. Here you don't need any implementations on the variant types you'd get because the variant types are for the most part just dumb data.

2. At the same time, precisely because this RFC does not permit implementations on variant types, as I think is proper to avoid the pitfalls of Java-OOP-inheritance APIs, it will not allow you to refactor an API such as syn::Lit into one where syn::LitStr is a variant type (i.e. Lit::Str) because the implementations there would not be admitted by the type checker.

3. All in all, I think the benefits of this proposal are well motivated and the costs in terms of understanding are not that great. I think this proposal is something that a subset of users would naturally expect; the user also doesn't have to do much, expressiveness is given for free to them.

4. The RFC interacts well with #1806 as well as goals and plans for gradual struct initialization; in fact, it provides the "missing link" that makes gradual initialization more generally applicable in the type system. This is a wonderful thing.

5. Thus on balance I think the RFC is a good idea.

(Feel free to integrate any points that you found relevant into the text of the RFC)

---

@nrc

This RFC should address those and issues, and summarise how this RFC is different to #1450.

👍

For the sake of completeness, an alternative might be some kind of general refinement type, though I don't think that is a good fit with Rust.

(Aside, but let's not go too deeply into this: I personally think that refinement / dependent typing is both a good idea, a good fit for Rust's general aim for correctness and type system power for library authors -- and RFC 2000 is sort of dependent types anyways so it's sort of sunk cost wrt. complexity -- the use cases for dependent/refinement types are sort of different than the goal here; With dependent types we wish to express things like { x: usize | x < 10 })

---

@eddyb

I'm not sure that would need to be at odds with variant types, if Rust ends up with refinement types I expect variant types to be refinements of their enum.

I agree; I think you can think of variant types in the general framework of refinement / dependent types;
With the notation due to @petrochenkov for pattern matching as a boolean operator, we can think of variant types as:

type FooVar = { x: Foo | x is Foo::Variant(...) };

[burdges]

We do want formal verification of rust code eventually, and afaik doing that well requires refinement types. I'm not saying rust itself needs refinement types per se, but rust should eventually have a type system plugin/fork/preprocessor for formal verification features, like refinement types. I do like this feature of course, but ideally the syntax here should avoid conflicts with refinement types.

[Centril]

@burdges

I do like this feature of course, but ideally the syntax here should avoid conflicts with refinement types.

Are there any such conflicts in your view?

[Centril]

(Or to elaborate; if there are any conflicts with the RFC as proposed with refinement typing, then stable Rust as is has that conflict since the RFC does not introduce any new syntax...)

[ExpHP]

Note that because a variant type, e.g. Sum::A, is not a subtype of the enum type (rather, it can simply be coerced to the enum type), a type like VecSum::A is not a subtype of Vec. (However, this should not pose a problem as it should generally be convenient to coerce Sum::A to Sum upon either formation or use.)

So, if I understand correctly, all existing code that uses enums now have coercions all over the place in order to ensure they continue functioning? I'm really not sure this works...

let mut x:

x = None;

// At this point the compiler knows the type
// of x is Option<?0>::None.

// But Option<_>::Some cannot be coerced to None
x = Some(1); // type error?

[leonardo-m]

let mut x:
x = None;
// At this point the compiler knows the type
// of x is Option<?0>::None.
// But Option<_>::Some cannot be coerced to None
x = Some(1); // type error?

I think in theory the type system should infer x to be of type Option<i32> because it sees both assignments.

But I think we need a formalization of the involved type system rules, to assure soundness, before implementing this proposal...

[mark-i-m]

I would rather frame this as follows:

• the type is Option
• the compiler tracks the most specific variant the value is if it can be known statically. This could be done with a standard dataflow analysis.

[leonardo-m]

We do want formal verification of rust code eventually, and afaik doing that well requires refinement types. I'm not saying rust itself needs refinement types per se, but rust should eventually have a type system plugin/fork/preprocessor for formal verification features, like refinement types.

While I don't dislike LiquidHaskell-like refinement typing, lately for the future of Rust I prefer a style of verification as in the Why3 language ( http://why3.lri.fr/ , that is also related to the Ada-SPARK verification style). We'll need a pre-RFC for this.

[leonardo-m]

I hope this syntax is also supported (I suggest to add it to the RFC):

enum Sum { A(u32), B, C }
fn print_a1(Sum::A(x): Sum::A) {}

A question regarding the ABI: is the print_a1() function receiving the Sum discriminant too as argument?

And in future it could also be supported the more DRY syntax (I think suggested by Centril):

fn print_a2(Sum::A(x)) {}

You could also add a new (silly) example to this RFC that shows the purposes of this type system improvement:

enum List { Nil, Succ(u32, Box<List>) }

fn prepend(x: u32, lst: List) -> List {
    List::Succ(x, box lst)
}

With this improvement you can write instead:

fn prepend(x: u32, lst: List) -> List::Succ {
    List::Succ(x, box lst)
}

Then you can define a list_head_succ() function that returns the head of the result of prepend() without a unwraps or Option result:

fn list_head(lst: List) -> Option<u32> {
    match lst {
        List::Succ(x, _) => x,
        List::Nil => None,
    }
}

fn list_head_succ(List::Succ(x, _): List::Succ) -> u32 { x }

{ x: usize | x < 10 }

For the common case of integer intervals for Rust I sometimes prefer a shorter and simpler syntax like:

type Small = usize[.. 10];

[shepmaster]

"Overhead"

I'm mostly interested in this RFC from the point-of-view of "enums of lots of standalone other types". The biggest example I have is the AST expressed in fuzzy-pickles, a Rust parser which uses this pattern extensively:

pub enum Item {
    AttributeContaining(AttributeContaining),
    Const(Const),
    Enum(Enum),
    // ...

Unfortunately, I don't see this as being a large win for such a case due to the "forced overhead" of each enum variant still being the same size as all the other variants. It's an understandable decision, just not one that I see as helping as much as it could.

This is mentioned in the alternatives section, but I want to make sure the point is reiterated.

Multiple variants

I didn't see any mention of if multiple variants would be supported:

#[derive(Debug)]
enum Count {
    Zero,
    One,
    Many(usize),
}

fn example(c: Count) {
    use Count::*;
    match c {
        x @ Zero | x @ One => println!("{:?}", x), // what is the type of `x` here?
        x => println!("{:?}", x),
    }
} 

It may also be worth explicitly calling out what the type is for those catch-all patterns as well as in cases of match guards.

foo @

This may be swerving into refinement type territory, but I naturally wanted to not type the foo @ in the previous example:

match c {
    Zero | One => println!("{:?}", c),
    // ...

I feel this is a pretty hidden and uncommon aspect of patterns, and it'd be nice to just be able to intuit the type based on the pattern without adding the explicit binding. That might even mean we could do:

if let Count::Many(..) = c {
    println!("{}", c.0);
}

[shepmaster]

"Overhead"

I'm mostly interested in this RFC from the point-of-view of "enums of lots of standalone other types". The biggest example I have is the AST expressed in fuzzy-pickles, a Rust parser which uses this pattern extensively:

pub enum Item {
    AttributeContaining(AttributeContaining),
    Const(Const),
    Enum(Enum),
    // ...

Unfortunately, I don't see this as being a large win for such a case due to the "forced overhead" of each enum variant still being the same size as all the other variants. It's an understandable decision, just not one that I see as helping as much as it could.

This is mentioned in the alternatives section, but I want to make sure the point is reiterated.

Multiple variants

I didn't see any mention of if multiple variants would be supported:

#[derive(Debug)]
enum Count {
    Zero,
    One,
    Many(usize),
}

fn example(c: Count) {
    use Count::*;
    match c {
        x @ Zero | x @ One => println!("{:?}", x), // what is the type of `x` here?
        x => println!("{:?}", x),
    }
} 

It may also be worth explicitly calling out what the type is for those catch-all patterns as well as in cases of match guards.

foo @

This may be swerving into refinement type territory, but I naturally wanted to not type the foo @ in the previous example:

match c {
    Zero | One => println!("{:?}", c),
    // ...

I feel this is a pretty hidden and uncommon aspect of patterns, and it'd be nice to just be able to intuit the type based on the pattern without adding the explicit binding. That might even mean we could do:

if let Count::Many(..) = c {
    println!("{}", c.0);
}

[shepmaster]

I'm a fan of the proposed style, but it might be worth stating why Rust the language wants to dissuade this pattern.

[shepmaster]

I disagree that this goal is going to be as widely achieved by this RFC as I would like due to the following point:

the variant types proposed here have identical representations to their enums

That means that if I have an enum with large variants:

enum Thing {
    One([u8; 128]),
    Two(u8),
}

Even the "small" variants (e.g. Thing::Two) are still going to take "a lot" of space.

[eddyb]

If space is a concern then we could have it so variant types only convert to their enum by-value, so e.g. a &Thing::Two wouldn't be a valid &Thing.

That's weaker than something more akin to refinement typing, but maybe it's enough?

[Centril]

@eddyb I think that's already the case; the RFC doesn't state anywhere, as far as I can tell, that &Thing::Two is a valid &Thing. Also note that the RFC explicitly states that Thing::Two and Thing having the same layout is not a guarantee so we could change the layout to be more space efficient.

[H2CO3]

Performance. Casting it once means that it becomes a concreet type and you can do static dispatch.

This can already be achieved by just matching on the enum once and using the contents of its variants directly. By the way, traits as bounds to type parameters result in static dispatch. Unless you specifically ask for a trait object, there's no dynamic dispatch going on by default.

Traits in Rust are much harder to deal with than concrete types and there's a lot of thing you can't do. If the the trait is not object safe, you can't return it etc...

That is not true anymore because there is impl Trait.

Of course traits are different from concrete types – that's kind of their raison d'être. Generic programming is a quite high level of abstraction, indeed – so people not familiar with it might need some time until they get used to programming against abstract interfaces rather than concrete types. I wouldn't say that this means that "traits are harder to deal with" in general, nor that this would warrant conflating certain, very different aspects of the language.

[thibaultdelor]

This can already be achieved by just matching on the enum once and using the contents of its variants directly.

Well then you are not using traits at all, that's a completely different story. You were making a point with abstraction and traits.

That is not true anymore because there is impl Trait.

impl trait is not magic, it's possible only when you know the concrete type at compile time. impl trait is more related to encapsulation than abstraction.

I wouldn't say that this means that "traits are harder to deal with" in general

Traits have additional restriction, can be monomorphised, require knowledge about trait objects, require to be boxed in some scenario...
They are harder to deal with, not because of abstraction, because of technical limitations due to Rust design (which is about zero cost abstraction etc.).

[alercah]

My musings there eventually lead me to think that we should consider forcing all the variants to have the same size/layout so that they are reference-compatible, because I think that there's a decent chance the differently-sized version has an alternative solution.

What exactly do you mean by "reference compatible" here?

I mean that &E::V can be safely cast to &E, which requires that the layouts match.

it means that all of these partial functions have to pattern match to do dispatch

I might be missing something, but doesn't anything dynamic need to perform some sort of runtime check to get to the correct function/type(/anything associated with them)? A pattern match is a comparison of the enum discriminant against a constant, and a branch, or an indirect branch indexed by the discriminant (depending on how the compiler happens to codegen it), modulo optimizations. But a method call through a trait object isn't free either: it's an offset against the vtable pointer and an indirect call (again, modulo devirtualization).

Yes, it does. But here I meant that you're doing this repeatedly because it's done on every single method, rather than doing it only once: if you pattern match on an enum to destructure it and then access the fields of the variant, you don't need to do the dispatch every time you access a field. Trait objects basically do have to do dynamic dispatch each time, but it's nicely hidden by the compiler.

This is a good approach.

Isn't that particular use case of ASTs solved by just implementing the trait on the member structs, though?

On the variant types you mean? The RFC proposes to ban that.

This isn't uncommon, because you'll want to treat FuncDecl and VarDecl separately in many cases, in a way that simply trying to farm out to traits isn't going to work for.

But with newtype variants around structs, you still have the concrete types of the wrapped structs, so you can just call different methods (possibly from different traits) on them. I'll explain this below in more detail, but I think if you want to treat two types differently, you simply shouldn't try to achieve that by means of a common trait. Just use the concrete types, or perhaps different traits.

Yes, you will. But sometimes you'll have a function which takes a NamedDecl and then wants to behave differently based on the concrete type. I don't think there's any problem with that personally, but it does require some way to destructure the NamedDecl into one of the concrete types.

but it's worth not[h]ing that this will likely make this a bigger pain point.

Going from a trait to a specific type induces feelings of object-oriented "downcasting" in me. (No surprise, the hack that is known as Any uses the same approach on the surface.) This direction always felt just wrong to me. If you are abstracting over a certain property (i.e. you expect a trait), why would you try to get to the guts of the concrete type? If the concrete type cares about its own internals, it should implement the trait accordingly, instead of various callers of the trait method switching on it from the outside.

I've long enjoyed Rust's enums because of the beauty, simplicity, and obvious behavior of pattern matching, and I'd really prefer not to spoil it by having arbitrary downcasts over (partial) subsets of variants. I see how this might result in some writing convenience, but it's awful for readability (which, I argue, should be weighted a lot more in evaluating language features. We spend much more time reading code than power-typing). By the way I've done a lot of AST manipulation in the past too, and this never was a real issue. Yes, it means that when the represented syntax changes, you have to restructure the AST nodes associated with it. I'd consider that to be inevitable by nature, rather than a "problem".

I dislike downcasting too, but destructuring an enum into its variant is not really different from downcasting, especially over a sealed trait (where you know every possible concrete type). I think the moral argument here is about when you expect consumers of a trait to be (morally) parametric. For enums you generally don't, but for unsealed traits you generally do. For a sealed trait, I think it starts to blur the line---and I think that taking advantage of that blurring in language design is valuable.

[H2CO3]

@alercah

Yes, it does. But here I meant that you're doing this repeatedly because it's done on every single method, rather than doing it only once:

Can you please provide a more specific example? The quote above seems to go against your former argument. I suspect we might be talking about different things… What I meant is, given an enum:

enum Foo {
    VarOne(ContentOne),
    VarTwo(ContentTwo),
}

you can implement a method on it, which calls into methods on the variants' associated data:

impl Foo {
    fn frobnicate(&self) {
        match *self { // 1 runtime check per method call against discriminant
            VarOne(ref content) => content.method_1(), // no dynamic dispatch here
            VarTwo(ref content) => content.method_2(), // no dynamic dispatch here
        }
    }
}

This isn't any worse, in terms of number of runtime checks, than:

trait Frobnicator { fn frobnicate(&self); }

impl Frobnicator for ContentOne {
    fn frobnicate(&self) { // 1 virtual call per method when used as a trait object
        self.method_1() // no dynamic dispatch here
    }
}

impl Frobnicator for ContentOne {
    fn frobnicate(&self) { // 1 virtual call per method when used as a trait object
        self.method_2() // no dynamic dispatch here
    }
}

let f: &dyn Frobnicator = &ContentOne::new();
f.frobnicate(); // the virtual call happens here

On the variant types you mean? The RFC proposes to ban that.

No, I meant on the types of the associated values wrapped in each newtype variant.

Yes, you will. But sometimes you'll have a function which takes a NamedDecl and then wants to behave differently based on the concrete type. I don't think there's any problem with that personally

Well, then we just disagree here. I think if consumer code wants to behave differently based on concrete types, an enum with public variants is a perfectly fine solution, and such an enum around the different types should be provided. There is no point in consuming a trait if its abstraction capabilities are thrown away because one ends up inspecting the concrete type anyway. That only results in unnecessary complication and an unclear mixture of concepts, which are hard to understand because they belong to neither approach (concrete types or abstract interfaces).

To clarify, I'm not saying that switching on an enum is in itself bad; rather, switching on an enum and taking advantage of the heterogeneous types, while pretending to program against a homogeneous interface is bad.

I dislike downcasting too, but destructuring an enum into its variant is not really different from downcasting

Indeed, it isn't from a technical point of view. However, enums are expected to be used like that, while one of the primary purposes of a trait is the exact opposite: to provide abstraction via information hiding. Therefore…

I think that taking advantage of that blurring in language design is valuable.

…I disagree with this statement too. This is exactly what I'm trying to avoid, as mentioned in the previous point.

[H2CO3]

@thibaultdelor

impl trait is not magic, it's possible only when you know the concrete type at compile time.

I am well aware that impl Trait is backed by a concrete type. What's a specific example when you couldn't usefully return an impl Trait because "you don't know the concrete type at compile time"? And how could a proposal around (even more) static dispatch resolve such a problem?

impl trait is more related to encapsulation than abstraction.

That is really vague; encapsulation is a form of/tool for abstraction. I also fail to see why hiding a concrete type under an interface doesn't qualify as "abstraction". But I'm not particularly interested in arguing over definitions; I was merely falsifying your (incorrect) claim that a type implementing a non-object-safe trait can't be returned.

Traits have additional restriction, can be monomorphised, require knowledge about trait objects, require to be boxed in some scenario...
They are harder to deal with, not because of abstraction, because of technical limitations due to Rust design (which is about zero cost abstraction etc.).

Again, these claims are very general and hard to concretize. Yes, generic functions and types with trait bounds can be monomorphized. Is that bad? Yes, traits can and sometimes need to be boxed, just like other types. Is that bad? When you make a trait object, it becomes a concrete type and can for all purposes be treated as one. Also based on this, I'm not sure about the "additional restrictions". You're really making an apples to oranges comparison here. Traits, generics, and trait objects all have their specific purpose, and they are used differently from (concrete, unrelated) types. I don't really see that as a problem. They are distinct language features solving different problems, which is why they all exist and are all useful in different ways.

[thibaultdelor]

What's a specific example when you couldn't usefully return an impl Trait because "you don't know the concrete type at compile time"?

Are you actually suggesting that it rarely occurs?!? Doing something like the following is pretty common IMO :

if something {
  return this;
} else {
  return that;
}

(where this and that both implement a trait)
I have plenty of concrete example in minds, I am just unsure what's your point here...

And how could a proposal around (even more) static dispatch resolve such a problem?

Which problem?!? Which proposal?!?

I was merely falsifying your (incorrect) claim that a type implementing a non-object-safe trait can't be returned.

You are arguing with a straw-man... A non object-safe trait can't be return as such. If you know the concrete type then you can but it's not much different conceptually than returning the concrete type and it's not my point. When talking about abstraction, the case where you abstract over a single type is not really the most relevant one.

Is that bad? [...] Is that bad? [...]

Who said it was? I said harder to deal with, because it has technical implications and limitations and forces you to think (like boxing) about concepts that you don't have to with concrete types.

When you make a trait object [...]

You can't always do that, that was my point.

I am going to stop replying, the discussion isn't constructive.

[H2CO3]

Are you actually suggesting that it rarely occurs?!?

No need to yell. I'm not suggesting that this "rarely occurs", it was just unclear what you precisely meant by "knowing the concrete type".

Your example is legitimate. However, it still does need some sort of runtime check. If the returned concrete type depends on a parameter that is only known at runtime, there is no way a runtime check could be avoided.

the discussion isn't constructive.

Agreed.

[CAD97]

ping from not-triage-but-this-was-mentioned-on-internals: what's the status of this RFC?

It seems that rust-lang/rust#57644 is a future-compat lint for syntax space for this (or an equivalent) feature; what's between this and proposal-merge?

[ExpHP]

Okay, let's take a look at the new stuff about type inference.

• If the value is treated as a single variant (and possibly additionally as the enum), we choose
  the single variant type. For example:

let x = Sum::A(5); // x: Sum::A
println!("x is {}", x.0);

• If the type is treated as multiple different variants, we choose the enum type.

let mut x = Sum::A(5); // x: Sum
println!("x is {}", x.0); // error: no field `0` on type `Sum`
x = Sum::B;
println!("x is not numeric");

• In a case where the type variable is unknown, we default to the enum.

I suspect this can't work for the same reason that we get type annotations needed all the time when trying to call methods; the type checker works strictly in order. It wants to be able to resolve the type of x.0 as soon as it sees it, not several statements later.

Without that restriction, we get the following madness:

#[derive(Debug)]
struct Struct {
    field: i8,
}

#[derive(Debug)]
enum Enum {
    A { field: Struct },
    B { field: Struct },
}

impl std::ops::Deref for Enum {
    type Target = Struct;
    
    fn deref(&self) -> &Struct {
        match self {
            Enum::A { field } => field,
            Enum::B { field } => field,
        }
    }
}

fn main() {
    let mut a = Enum::A { field: Struct { field: 0 } };
    
    println!("`{:?}`", a.field);
    
    // Commenting out the following line changes the output of the
    // above println from: `0`
    //                 to: `Struct { field: 0 }`
    a = Enum::B { field: Struct { field: 0 } };
}

(by the way, because the above snippet does currently compile and print 0 when you comment out the second line, this constitutes a breaking change)

[varkor]

I suspect this can't work for the same reason that we get type annotations needed all the time when trying to call methods; the type checker works strictly in order. It wants to be able to resolve the type of x.0 as soon as it sees it, not several statements later.

(Edited: see @alercah's comment below.)

// Commenting out the following line changes the output of the
// above println from: `0`
//                 to: `Struct { field: 0 }`

Enum::A still implements Deref, so I would expect this to print `0`, regardless of whether that line was commented out or not.

[alercah]

That doesn't sound right to me; types can be inferred due to unification which may well occur several statements later:

fn main() {
    let x = vec![1, 2, 3].into_iter().collect();
    let y = &x;
    let _z : &Vec<u64> = y;
}

[golddranks]

@alercah I think @ExpHP meant cases like this, since they are talking about "trying to call methods":

fn return_vec() -> Vec<u64> {
    let mut x = vec![1, 2, 3].into_iter().collect();
    
    // uncomment to make it compile
    // let mut x: Vec<u64> = x;
    
    x.push(4);
    
    x
}

fn main() {
    println!("{:?}", return_vec());
}

https://play.rust-lang.org/?version=stable&mode=debug&edition=2018&gist=940235f3b5743e92086a19465d5b8607

[ExpHP]

@golddranks is right.

It is a common misconception that type inference can use information later in the function body to resolve types earlier in the body. Type checking works almost entirely strictly forwards in a function body. The only thing that type inference actually does is allow the compiler to reason about types generically, using type inference variables to delay filling in the details that it does not yet need to know.

@alercah your example goes like this:

let x = vec![1, 2, 3].into_iter().collect();
let y = &x;
let _z : &Vec<u64> = y;

1. The temporary vec![1, 2, 3] receives a type of Vec<?int_0>, where ?int_0 is an integer-flavored type inference variable
2. .into_iter() and .collect() are resolved to IntoIter::into_iter and Iterator::collect because there exist general enough impls for these traits to apply to Vec<?int_0>.
3. x ends up with type ?1, with the obligation ?1: FromIterator<?int_0>. (To be checked later)
4. y gets type &?1.
5. _z has type &Vec<u64>. When the compiler unifies the type of the expression y with the type of _z, it determines that ?int_0 = u64.

If a x.push() is inserted before let _z = then the compiler rejects it, because no methods are known for ?1. (it has no outermost type constructor, so inherent methods cannot be sought; and it is not known to satisfy any traits with a push associated method)

Here's some more fun examples to help convince you:

Example 1

struct Wrapped<T>(T);

impl Wrapped<Vec<u8>> {
    fn boo(&self) { println!("Vec<u8>") }
}

trait Boo {
    fn boo(&self) { println!("Trait method") }
}

impl<T> Boo for Vec<T> {}

fn main() {
    let mut a = Wrapped(vec![]);

    a.boo();
    a.0.push(());
}

Does this compile? If so, what does it print? If not, what is the error?

(Hint: This would compile if the compiler could determine that a has type Vec<()>. But... can it?)

Solution

   Compiling playground v0.0.1 (/playground)
error[E0308]: mismatched types
  --> src/main.rs:17:14
   |
17 |     a.0.push(());
   |              ^^ expected u8, found ()
   |
   = note: expected type `u8`
              found type `()`

Somebody who thinks type inference can flow information backwards might think that a will have type Vec<()> and that this should therefore compile and print "Trait method". Such is not the case.

a initially has type Wrapped<Vec<?0>>. At a.boo(), the compiler needs to know immediately what method is being called. It searches inherent methods of Wrapped<Vec<_>> first because these take precedence over trait methods. There is a single match, <Wrapped<Vec<u8>>>::boo, so it selects this, and determines that ?0 = u8.

Example 2

struct Wrapped<T>(T);

impl Wrapped<Vec<u8>> {
    fn boo(&self) { println!("Vec<u8>") }
}

// !!!!! This was added to example 1
impl Wrapped<Vec<()>> {
    fn boo(&self) { println!("Vec<())>") }
}

// !!!!!! Trait was removed
// (it was already shown to be irrelevant)

fn main() {
    let mut a = Wrapped(vec![]);

    a.boo();
    a.0.push(());
}

Does this compile? If so, what does it print? If not, what is the error?

Solution

   Compiling playground v0.0.1 (/playground)
error[E0034]: multiple applicable items in scope
  --> src/main.rs:21:7
   |
21 |     a.boo();
   |       ^^^ multiple `boo` found
   |
note: candidate #1 is defined in an impl for the type `Wrapped<std::vec::Vec<u8>>`
  --> src/main.rs:4:5
   |
4  |     fn boo(&self) { println!("Vec<u8>") }
   |     ^^^^^^^^^^^^^
note: candidate #2 is defined in an impl for the type `Wrapped<std::vec::Vec<()>>`
  --> src/main.rs:9:5
   |
9  |     fn boo(&self) { println!("Vec<())>") }
   |     ^^^^^^^^^^^^^

Type checking here begins similarly to problem 1, but when it searches the inherent methods of Wrapped<Vec<_>>, it finds two candidates and refuses to continue.

Why does it give up? Well, let's consider a more sinister example. Above, the signatures just so happen to match, but what if they didn't?

struct Wrapped<T>(T);

impl Wrapped<Vec<u8>> {
    fn boo(&self) -> Self { Wrapped(self.0.to_vec()) }
}

impl Wrapped<Vec<()>> {
    fn boo(&self) -> (u32, u32) { (2, 3) }
}

fn main() {
    let mut a = Wrapped(vec![]);

    println!("{:?}", a.boo().0.rotate_left(3));
    
    // a.0.push(());  // (1)
    // a.0.push(1u8); // (2)
}

If the compiler did not give up when type-checking a.boo(), then when it checks the println there are two completely different interpretations: (notice that one of these interpretations even inserts auto-refs and coercions not present in the other)

// if a is Vec<u8>
let tmp: &Wrapped<Vec<u8>> = &a;
let tmp: Wrapped<Vec<u8>> = <Wrapped<Vec<u8>>>::boo(&a);
let tmp: () = <&mut [u8]>::rotate_left(&mut *tmp.0, 3_usize);
println!("{:?}", tmp);

// if a is Vec<()>
let tmp: (u32, u32) = <Wrapped<Vec<()>>>::boo(&a);
let tmp: u32 = <u32>::rotate_left(tmp.0, 3_u32);
println!("{:?}", tmp);

Before reaching either (1) or (2), how would it decide which interpretation to use for this line? Scenarios like this are likely why the compiler gives up so willingly.

Example 3

struct Wrapped<T>(T);

trait Boo {
    type Assoc;
    fn boo(&self) -> Self::Assoc;
}

// !!!!!! The inherent methods from Problem 2
//        have been replaced with trait impls
impl Boo for Wrapped<Vec<u8>> {
    type Assoc = Self;
    fn boo(&self) -> Self { Wrapped(self.0.to_vec()) }
}

impl Boo for Wrapped<Vec<()>> {
    type Assoc = (u32, u32);
    fn boo(&self) -> (u32, u32) { (2, 3) }
}

fn main() {
    let mut a = Wrapped(vec![]);

    println!("{:?}", a.boo());
    // println!("{:?}", a.boo().0.rotate_left(3));
    
    a.0.push(());
}

Does this compile? If so, what does it print? If not, what is the error?

What if we uncomment the second println!?

(note: for now just assume that println!("{:?}", x) expands to code containing a call to fmt::Debug::fmt(&x))

Solution

   Compiling playground v0.0.1 (/playground)
    Finished dev [unoptimized + debuginfo] target(s) in 0.58s
     Running `target/debug/playground`
(2, 3)

In contrast to the multiple inherent methods in Problem 2, the compiler allows multiple impls from the same trait to match. Basically:

• a has type Wrapped<Vec<?0>>
• The output of a.boo() can be written generically as <Wrapped<Vec<?0>> as Boo>::Assoc.
• This is borrowed and fed to Debug::fmt, creating the obligation that <Wrapped<Vec<?0>> as Boo>::Assoc: Debug. Trait obligations are checked lazily, so it's okay that we still don't know this type!
• a.0.push(()) tells it that ?0 = ().
• At some point later it verifies that <Wrapped<Vec<()>> as Boo>::Assoc implements Debug.

Of course, if you uncomment the second println, you instead get

error[E0282]: type annotations needed
  --> src/main.rs:25:22
   |
25 |     println!("{:?}", a.boo().0.rotate_left(3));
   |                      ^^^^^^^ cannot infer type
   |
   = note: type must be known at this point

for reasons already discussed in the solution to problem 2.

---

Enum::A still implements Deref, so I would expect this to print 0, regardless of whether that line was commented out or not.

So auto-deref takes precedence over "inherent" fields of the variant? If so that's quite surprising and should be mentioned in the RFC.

[alercah]

Yes, I believe you're correct in general, but unless I'm mistaken, nothing prevents the type checker from letting `x.0` be typed as a variable until later.

…

On Sun., Mar. 3, 2019, 14:24 Michael Lamparski, ***@***.***> wrote: @golddranks <https://github.com/golddranks> is right. It is a common misconception that type inference can use information later in the function body to resolve types earlier in the body. Type checking works almost entirely strictly forwards in a function body. The only thing that type inference actually does is allow the compiler to reason about types generically, using type inference variables for details that it does not yet *need* to know. @alercah <https://github.com/alercah> your example goes like this: let x = vec![1, 2, 3].into_iter().collect();let y = &x;let _z : &Vec<u64> = y; 1. The temporary vec![1, 2, 3] receives a type of Vec<?int_0>, where ?int_0 is an integer-flavored type inference variable <https://github.com/rust-lang/rust/blob/c0086b9e8972fef9fd4af24bae20d45021ed06c6/src/librustc/ty/sty.rs#L1257> 2. .into_iter() and .collect() are resolved to IntoIter::into_iter and Iterator::collect because there exist general enough impls for these traits to apply to Vec<?int_0>. 3. x ends up with type ?1, with the obligation ?1: FromIterator<?int_0> . 4. y gets type &?1. 5. _z has type &Vec<u64>. When the compiler unifies the type of the expression y with the type of _z, it determines that ?int_0 = u64. If a x.push() is inserted before let _z = then the compiler rejects it, because no methods are known for ?1. (it has no outermost type constructor, so inherent methods cannot be sought; and it is not known to satisfy any traits with a push associated method) Here's some more fun examples to help convince you: Example 1 struct Wrapped<T>(T); impl Wrapped<Vec<u8>> { fn boo(&self) { println!("Vec<u8>") } } trait Boo { fn boo(&self) { println!("Trait method") } } impl<T> Boo for Vec<T> {} fn main() { let mut a = Wrapped(vec![]); a.boo(); a.0.push(()); } Does this compile? If so, what does it print? If not, what is the error? Solution Compiling playground v0.0.1 (/playground) error[E0308]: mismatched types --> src/main.rs:17:14 |17 | a.0.push(()); | ^^ expected u8, found () | = note: expected type `u8` found type `()` Somebody who thinks type inference can flow information backwards might think that a will have type Vec<()> and that this should therefore compile and print "Trait method". Such is not the case. a initially has type Wrapped<Vec<?0>>. At a.boo(), the compiler needs to know immediately what method is being called. It searches inherent methods of Wrapped<Vec<_>> first because these take precedence over trait methods. There is a single match, <Wrapped<Vec<u8>>>::boo, so it selects this, and determines that ?0 = u8. Example 2 struct Wrapped<T>(T); impl Wrapped<Vec<u8>> { fn boo(&self) { println!("Vec<u8>") } } // !!!!! This was added to example 1impl Wrapped<Vec<()>> { fn boo(&self) { println!("Vec<())>") } } // !!!!!! Trait was removed// (it was already shown to be irrelevant) fn main() { let mut a = Wrapped(vec![]); a.boo(); a.0.push(()); } Solution Compiling playground v0.0.1 (/playground) error[E0034]: multiple applicable items in scope --> src/main.rs:21:7 |21 | a.boo(); | ^^^ multiple `boo` found | note: candidate #1 is defined in an impl for the type `Wrapped<std::vec::Vec<u8>>` --> src/main.rs:4:5 | 4 | fn boo(&self) { println!("Vec<u8>") } | ^^^^^^^^^^^^^ note: candidate #2 is defined in an impl for the type `Wrapped<std::vec::Vec<()>>` --> src/main.rs:9:5 | 9 | fn boo(&self) { println!("Vec<())>") } | ^^^^^^^^^^^^^ Type checking here begins similarly to problem 1, but when it searches the inherent methods of Wrapped<Vec<_>>, it finds two candidates and refuses to continue. Why does it give up? Well, let's consider a more sinister example. Above, the signatures just so happen to match, but what if they didn't? struct Wrapped<T>(T); impl Wrapped<Vec<u8>> { fn boo(&self) -> Self { Wrapped(self.0.to_vec()) } } impl Wrapped<Vec<()>> { fn boo(&self) -> (u32, u32) { (2, 3) } } fn main() { let mut a = Wrapped(vec![]); println!("{:?}", a.boo().0.rotate_left(3)); // a.0.push(()); // (1) // a.0.push(1u8); // (2) } If the compiler did not give up when type-checking a.boo(), then when it checks the println there are two *completely* different interpretations: (notice that one of these interpretations even inserts auto-refs and coercions not present in the other) // if a is Vec<u8>let tmp: &Wrapped<Vec<u8>> = &a;let tmp: Wrapped<Vec<u8>> = <Wrapped<Vec<u8>>>::boo(&a);let tmp: () = <&mut [u8]>::rotate_left(&mut *tmp.0, 3_usize);println!("{:?}", tmp); // if a is Vec<()>let tmp: (u32, u32) = <Wrapped<Vec<()>>>::boo(&a);let tmp: u32 = <u32>::rotate_left(tmp.0, 3_u32);println!("{:?}", tmp); Before reaching either (1) or (2), how would it decide which interpretation to use for this line? Scenarios like this are likely why the compiler gives up so willingly. Example 3 struct Wrapped<T>(T); trait Boo { type Assoc; fn boo(&self) -> Self::Assoc; } // !!!!!! The inherent methods from Problem 2// have been replaced with trait implsimpl Boo for Wrapped<Vec<u8>> { type Assoc = Self; fn boo(&self) -> Self { Wrapped(self.0.to_vec()) } } impl Boo for Wrapped<Vec<()>> { type Assoc = (u32, u32); fn boo(&self) -> (u32, u32) { (2, 3) } } fn main() { let mut a = Wrapped(vec![]); println!("{:?}", a.boo()); // println!("{:?}", a.boo().0.rotate_left(3)); a.0.push(()); } Does this compile? If so, what does it print? If not, what is the error? What if we uncomment the second println!? (note: for now just assume that println!("{:?}", x) expands to code containing a call to fmt::Debug::fmt(&x)) Solution Compiling playground v0.0.1 (/playground) Finished dev [unoptimized + debuginfo] target(s) in 0.58s Running `target/debug/playground` (2, 3) In contrast to the multiple inherent methods in Problem 2, the compiler allows multiple impls from the same trait to match. Basically: - a has type Wrapped<Vec<?0>> - The output of a.boo() can be written generically as <Wrapped<Vec<?0>> as Boo>::Assoc. - This is borrowed and fed to Debug::fmt, creating the obligation that <Wrapped<Vec<?0>> as Boo>::Assoc: Debug. *Trait obligations are checked lazily,* so it's okay that we still don't know this type! - a.0.push(()) tells it that ?0 = (). - At some point later it verifies that <Wrapped<Vec<()>> as Boo>::Assoc implements Debug. Of course, if you uncomment the second println, you get error[E0282]: type annotations needed --> src/main.rs:25:22 | 25 | println!("{:?}", a.boo().0.rotate_left(3)); | ^^^^^^^ cannot infer type | = note: type must be known at this point for reasons already discussed in the solution to problem 2. ------------------------------ Enum::A still implements Deref, so I would expect this to print 0, regardless of whether that line was commented out or not. So deref coercions take precedence over "inherent" fields of the variant? If so that's quite surprising and should be mentioned in the RFC. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#2593 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AT4HVY-FPPexMZRaRRBBhmtgsbhfGZ8sks5vTCFQgaJpZM4YYFVh> .

[ExpHP]

Yes, I believe you're correct in general, but unless I'm mistaken, nothing
prevents the type checker from letting x.0 be typed as a variable until
later.

Even for usage as simple as Debug::fmt(&x.field) (which does not need to know anything more than typeof(x.field): Debug), it would still need to potentially come back and insert a deref to make it Debug::fmt(&(*x).field) after it is already finished checking the expression. If the compiler could do that here, then it could no doubt also solve many of the millions of other paper cuts that all ultimately boil down to these limitations of the type checker.

I've long lost count of how many times I've responded to github issues or posts on URLO to explain my understanding of what's happening behind the scenes in these apparent "bugs" of type inference.... and I am thoroughly convinced at this point that if there was a reasonable and scalable solution, then it'd already be implemented and I wouldn't be here playing the role of the antagonist!

[varkor]

I suspect this can't work for the same reason that we get type annotations needed all the time when trying to call methods; the type checker works strictly in order.

I imagine the right solution, then, is to require explicit annotations for variants and fallback on the general enum type. (This is forwards compatible with a solution that doesn't involve explicit annotations, so I don't think this means losing too much.)