Tracking Issue for Never Patterns (never_patterns)

[Nadrieril]

This is a tracking issue for the experimental "never patterns" feature (as per the T-lang experimental feature process).
The feature gate for the issue is #![feature(never_patterns)].

RFC: rust-lang/rfcs#3719

This feature introduces a new syntax for patterns: !. In a pattern, ! is valid for any uninhabited type and simply acknowledges that the type is empty. Because a never pattern is statically known to be unreachable, it obeys slightly different rules:

enum Void {}
let result: *const Result<u32, Void> = ...;
unsafe {
    match *result {
        Ok(x) => { foo(x) }
        Err(!), // no need for `=> <expr>`
    }
    let (Ok(x) | Err(!)) = *result; // valid
}

fn blah(!: Void) -> SomeType {} // the function can never be called so it doesn't need a body

This feature does not affect exhaustiveness or pattern reachability; see the exhaustive_patterns and min_exhaustive_patterns features for that.

About tracking issues

Tracking issues are used to record the overall progress of implementation.
They are also used as hubs connecting to other relevant issues, e.g., bugs or open design questions.
A tracking issue is however not meant for large scale discussion, questions, or bug reports about a feature.
Instead, open a dedicated issue for the specific matter and add the relevant feature gate label.

Steps

• Implement the feature
• Write an RFC and have it accepted
• Adjust documentation (see instructions on rustc-dev-guide)
• Formatting for new syntax has been added to the Style Guide (nightly-style-procedure)
• Stabilization PR (see instructions on rustc-dev-guide)

Unresolved Questions

Implementation plans/details

Preview Give feedback

1. ! is parsed as a pattern.
2. Parsing of match expressions allows match <expr> { <pat>, ... } i.e. a pattern without a body. We later check to only allow that if the pattern is a never pattern.
3. The lack of body is desugared into something equivalent to unreachable_unchecked!().
4. Never patterns skip the or-pattern bindings check, to allow Ok(x) | Err(!)
5. Never patterns don't allow bindings, like in (x, !)
6. Allow fn foo(!: Void) -> SomeType {}
7. ! is typechecked like a wildcard; it adds no type constraint
8. In match checking we check that ! is only used on an uninhabited type (modulo match ergonomics).
9. The ! pattern is lowered to MIR.
10. A ! pattern asserts validiy of the matched place
11. Exhaustiveness understands !.
12. Exhaustiveness recommends using never patterns where sensible.
13. Rustfmt understands never patterns
14. Check every bit of the compiler that handles or-patterns to see if it handles never patterns properly
15. Write up the RFC
16. never patterns: ! argument not detected as diverging on async fn #120240
    A-async-await AsyncAwait-Triaged C-bug F-never_patterns T-compiler WG-async requires-incomplete-features +7

Implementation history

Preview Give feedback

1. Add never_patterns feature gate #118157
   F-never_patterns T-compiler +2
   
2. never_patterns: Parse match arms with no body #118527
   F-never_patterns I-lang-nominated T-compiler merged-by-bors perf-regression perf-regression-triaged +6
   
3. Correctly gate the parsing of match arms without body #118868
   A-parser F-never_patterns S-waiting-on-bors T-compiler +4
   
4. never patterns: Check bindings wrt never patterns #119610
   F-never_patterns S-waiting-on-bors T-compiler merged-by-bors perf-regression perf-regression-triaged +6
   
5. never patterns: Document behavior of never patterns with macros-by-example #119622
   A-macros F-never_patterns S-waiting-on-bors T-compiler +4
   
6. never_patterns: typecheck never patterns #120009
   F-never_patterns S-waiting-on-bors T-compiler +3
   
7. Report unreachable subpatterns consistently #120097
   I-lang-nominated S-waiting-on-bors T-compiler +3
   
8. never_patterns: Count ! bindings as diverging #120104
   F-never_patterns S-waiting-on-bors T-compiler +3
   
9. never patterns: It is correct to lower ! to _. #120517
   F-never_patterns S-waiting-on-bors T-compiler +3
   
10. never_patterns: lower a never arm to an unreachable terminator #120758
    F-never_patterns S-waiting-on-author T-compiler +3
    
11. never patterns: Fix liveness analysis in the presence of never patterns #121391
    F-never_patterns S-waiting-on-bors T-compiler +3
    
12. never patterns: suggest ! patterns on non-exhaustive matches #121823
    F-never_patterns S-waiting-on-bors T-compiler +3
    
13. never patterns: lower never patterns to Unreachable in MIR #123332
    F-never_patterns S-waiting-on-bors T-compiler merged-by-bors +4
    

[Alonely0]

enum Void {}
unsafe {
   let ptr: *const Result<u32, Void> = ...;
   match *ptr {
       Ok(x) => { ... }
       Err(_) => { ... } // potentially reachable if `ptr` points to uninitialized data
   }
   match *ptr {
       Ok(x) => { ... }
       Err(!) // indicate explicitly that we ensured the empty case couldn't happen
   }
}

The first match is UB. First, in Rust, any reads on uninitialized memory are considered undefined behavior; and second, if the Err variant cannot be constructed, it can't ever exist. Thus, the check for it can (and should) be removed by the optimizer, wiping away any matching on the Err variant and the code associated with it. Basically, the match would always execute the Ok branch regardless of the existence of valid data on the inner, which is exactly what you're not trying to do there (it would always get optimized down to the second match, which is equivalent to the ones proposed that are not unsafe). As a follow-up, it is not even marked as unsafe, there are many things that can go wrong when reading from a random raw pointer, not just the discriminant of an enum... Has this gone through a pre-RFC of some sort?

I think the fact that we do not want to touch uninitialized memory directly (w/o MaybeUninit) in the language with a 10ft pole is pretty well set in stone, so I'd just remove this part of the proposal and rename it "Allow elision of unreachable patterns".

[Nadrieril]

Hi! I thought like you not that long ago, until I digged into the subtleties of rust operational semantics. Allow me to clarify.

in Rust, any reads on uninitialized memory are considered undefined behavior

This is correct, but that match does not read from the value in the Err case, since a _ pattern does not cause a read.

if the Err variant cannot be constructed, it can't ever exist

While this is a likely possibility for rust opsem to decide, this has not been decided yet. Therefore I am erring on the side of assuming it is possible to read Err with uninitialized data. This is discussed here.

it is not even marked as unsafe

I'm confused, there's an unsafe block in the code you quote. Dereferencing a pointer always requires an unsafe block.

Has this gone through a pre-RFC of some sort?

No, this is an experimental feature gate, meant to iron out the details of the feature before we write the RFC. The lang-team took a quick look at my proposal and approved the experiment. I have discussed details of rust opsem with @RalfJung for a while before I felt confident proposing this. While I believe my approach is sound, this will have to be fully settled in the (to-be-written) RFC.

I think the fact that we do not want to touch uninitialized memory directly (w/o MaybeUninit) in the language with a 10ft pole is pretty well set in stone

This isn't quite correct. We need the possibility of uninitialized data behind a pointer. In fact to initialize a MaybeUninit<T> you have to go through as_mut_ptr which returns a *mut T pointing to possibly-uninitialized memory.

The mem::uninitialized debacle that led to the creation of MaybeUninit wasn't about uninitialized memory in general, it was about the fact that the function would return an uninitialized value, which is indeed instant UB. As long as the uninitialized data is not read (e.g. because it is behind a pointer), we're ok.

I'd just remove this part of the proposal and rename it "Allow elision of unreachable patterns".

This is the opposite of what this proposal is about. "Allow elision of unreachable patterns" is implemented in the exhaustive_patterns feature gate, which hasn't been stabilized because it is unsound, exactly because it allows elision of patterns that are reachable without UB in rust's current operational semantics. The whole point of this proposal is to fix that soundness whole.

[RalfJung]

in Rust, any reads on uninitialized memory are considered undefined behavior

This is not correct. Please consult the reference for what is and is not UB. Here is an example of code that reads uninit memory (for some definition of "reads") and has no UB.

You need to be careful with blanket statements like that.

---

Miri's current implementation means that reading the discriminant of a Result<u32, Void> can indeed never yield Err. The question is:

• Do we want to commit to this and start designing surface-language-level things around it?
• Even if we do, do we want the UB to come about implicitly by omitting match arms?

Never Patterns help because they let us make explicit when we want to exploit that a type is empty. We can then factor the exhaustive_patterns discussion into "how do Never Patterns work" and "when do we implicitly insert Never Patterns", i.e., when can they be elided.

For the case of Result<u32, Void>, given that there is an explicit Ok arm that asks for the discriminant to be read, I could imagine that we will decide "yeah okay we can add Err(!) implicitly". But maybe we don't want to do that when there are raw pointers involved, just to be extra safe.

[RalfJung]

The summary of this issue doesn't entirely capture this factoring I described though, it could be improved. Currently it mixes up the concept of Never Patterns with the concept of eliding Never Patterns, while I think these concepts should be discussed completely separately. In that sense gathering feedback on a separate writeup might be helpful. The experimental process means we don't have to wait until the RFC is all set and done, but it can still help to have the RFC at least sketched so that key stakeholders can even understand what the experimental change is doing without reverse engineering that from the implementation. It can also serve as a guideline to the reviewer.

The first example in this issue does not use a never pattern, making it a strange example for explaining never patterns.

[Nadrieril]

This is not correct.

Oh well, every time I think I get it I discover I'm missing something. Thx for pointing it out.

Even if we do, do we want the UB to come about implicitly by omitting match arms?

Currently it mixes up the concept of Never Patterns with the concept of eliding Never Patterns

No, my proposal is different. I'm intentionally moving away from this "implicit/explicit" frame that was in Niko's original never patterns blog post. Maybe this will turn out to be mistaken, but my current stance is: if a branch is reachable without UB, then it can't be omitted. That would be exhaustiveness being unsound. That's it, no elided never patterns or such things. Then I add the never patterns syntax as a simpler alternative to Err(x) => match x {}.

I find that a lot easier to understand and explain than a contextual desugaring. The drawback is that exhaustiveness now depends on the source of the scrutinee place and not just its type.

it can still help to have the RFC at least sketched

I linked this sketch in the OP, have you read it?

The first example in this issue does not use a never pattern, making it a strange example for explaining never patterns.

Well yeah, that's because this proposal has two parts: there's a refinement of exhaustiveness checking, and there's the never patterns syntax. I do intend to split off the exhaustiveness part into its own feature that we might be able to stabilize sooner.

[Alonely0]

in Rust, any reads on uninitialized memory are considered undefined behavior

This is not correct. Please consult the reference for what is and is not UB. Here is an example of code that reads uninit memory (for some definition of "reads") and has no UB.

You need to be careful with blanket statements like that.

I consulted the Rustonomicon, sorry if I was misinformed.
https://doc.rust-lang.org/nomicon/uninitialized.html

EDIT: It seems that in this context (enums) I was right, but my statement without it wasn't, thanks for pointing that out. I wasn't aware of the following:

Note: Uninitialized memory is also implicitly invalid for any type that has a restricted set of valid values. In other words, the only cases in which reading uninitialized memory is permitted are inside unions and in "padding" (the gaps between the fields/elements of a type).

[Alonely0]

Hi! I thought like you not that long ago, until I digged into the subtleties of rust operational semantics. Allow me to clarify.

in Rust, any reads on uninitialized memory are considered undefined behavior

This is correct, but that match does not read from the value in the Err case, since a _ pattern does not cause a read.

But how do you know that it is an Err? Because you read the discriminant, which is either valid (Ok/Err) or invalid and UB to operate with it. You can't say "it is Err because it was not Ok", you're assuming something the optimizer hasn't yet and might not do. If, for example, it decides to use jump tables, you could be jumping to dead code for all you know. The unsafe thus should mean that at least, you ensure that the enum has a valid representation, and if it has it, it still has to be only Ok because the Err is uninhabited (more on that on the next paragraph).

if the Err variant cannot be constructed, it can't ever exist

While this is a likely possibility for rust opsem to decide, this has not been decided yet. Therefore I am erring on the side of assuming it is possible to read Err with uninitialized data. This is discussed here.

But that post is about the layout and how we manage it internally, and I'm talking about the semantics of uninhabited (never) types. According to the reference, casting uninhabited (never) types out of thin air is unsound, and that's the same as having the discriminant be Err on that value, you're saying that whatever is next (in this case, zero bytes), is a constructed uninhabited type, which is equivalent to a singularity. The Err variant is an invalid (ergo unsound) state of the program. So you have to extend the meaning of that unsafe statement to mean "I'm sure it is an Ok", but it appears to me that's exactly the opposite of what you're trying to do...

it is not even marked as unsafe

I'm confused, there's an unsafe block in the code you quote. Dereferencing a pointer always requires an unsafe block.

Now I see it, I was looking for it, but for some reason I couldn't find it yesterday, sorry for that, mea culpa.

Has this gone through a pre-RFC of some sort?

No, this is an experimental feature gate, meant to iron out the details of the feature before we write the RFC. The lang-team took a quick look at my proposal and approved the experiment. I have discussed details of rust opsem with @RalfJung for a while before I felt confident proposing this. While I believe my approach is sound, this will have to be fully settled in the (to-be-written) RFC.

I think the fact that we do not want to touch uninitialized memory directly (w/o MaybeUninit) in the language with a 10ft pole is pretty well set in stone

This isn't quite correct. We need the possibility of uninitialized data behind a pointer. In fact to initialize a MaybeUninit<T> you have to go through as_mut_ptr which returns a *mut T pointing to possibly-uninitialized memory.

Indeed, raw pointers indeed can point to uninitialized data, but that was not my point. My point is that you're reading a value that you don't know if it is initialized and in a valid state (the discriminant), and you're checking that validity using semantics that assume the value is in a valid state, so the checks may get removed altogether.

The mem::uninitialized debacle that led to the creation of MaybeUninit wasn't about uninitialized memory in general, it was about the fact that the function would return an uninitialized value, which is indeed instant UB. As long as the uninitialized data is not read (e.g. because it is behind a pointer), we're ok.

I'd just remove this part of the proposal and rename it "Allow elision of unreachable patterns".

This is the opposite of what this proposal is about. "Allow elision of unreachable patterns" is implemented in the exhaustive_patterns feature gate, which hasn't been stabilized because it is unsound, exactly because it allows elision of patterns that are reachable without UB in rust's current operational semantics. The whole point of this proposal is to fix that soundness whole.

Oh, great then.

[Nadrieril]

I feel like we're talking a bit past each other, I ask your patience as we try to find what we're disagreeing on.

casting uninhabited (never) types out of thin air is unsound, and that's the same as having the discriminant be Err on that value

That doesn't seem correct. First I can very well make a ptr: *const Result<T, !>. And until opsem rules it out, I could find a way to write to ptr in a way that initializes the discriminant to Err. So far, no uninhabited type has ever been read so we're good. Then if I read the discriminant, I get Err which is still good. The case that would be UB is if I then read the data part of the enum.

Note the very important fact that all of this happens behind a pointer. If it was instead a plain let x: Result<T, !> = ...;, then x must be valid, which indeed forbids the Err case.

Also I'm not sure what we're arguing about because it is already the case on stable rust that the Err(_) case must be included. This proposal doesn't change that.

You can't say "it is Err because it was not Ok", you're assuming something the optimizer hasn't yet and might not do. If, for example, it decides to use jump tables, you could be jumping to dead code for all you know.

That is not what I'm doing. If the match expression was compiled that way I'm pretty sure it would be a bug. What the Err(_) case means is: "if we read the discriminant, and it returns Err, take this branch". If it turns out that it's UB to construct the Err case, then that's fine, the discriminant read will always return Ok and the Err branch will just never be taken.

If on the other hand I allowed match ... { Ok(_) => {} } to be exhaustive, then if Err is ever read we have UB. That would mean match exhaustiveness is unsound! I insist that requiring a branch for Err is the safer alternative. Unless that's not what you were arguing?

Thus, the check for it can (and should) be removed by the optimizer, wiping away any matching on the Err variant and the code associated with it

Until opsem decides for sure that Err is not constructible, then the optimized MUSN'T remove this case. Today's rust doesn't optimize this away, and this proposal does not change that. Moreover, if this optimization was ever correct, then what's the problem? The optimizer optimized away some dead code, I don't see how that causes any UB. I'm saying the same thing as above again, but I want to be sure we're on the same page.

[Nadrieril]

Maybe I should be clearer about what's currently required on stable rust. In both stable rust and under the proposal, the first match below is accepted, and the second match below gives a "non-exhaustive" error.

enum Void {}
unsafe {
    let ptr: *const Result<u32, Void> = ...;
    match *ptr {
        Ok(x) => { ... }
        Err(_) => { ... }
    }
    match *ptr { // ERROR: non-exhaustive: `Err` not covered
        Ok(x) => { ... }
    }
}

As long as we're behind a pointer/reference/union, there's no changes compared to stable. Where it does change is in the other cases, e.g. a plain value. In stable rust, we get:

enum Void {}
let val: Result<u32, Void> = ...;
match val { // accepted
    Ok(x) => ...
    Err(_) => ...
}
match val { // ERROR: non-exhaustive: `Err` not covered
    Ok(x) => ...
}

With never_patterns, we get:

enum Void {}
let val: Result<u32, Void> = ...;
match val {
    Ok(x) => ...
    Err(_) => ... // WARN: unreachable
}
match val { // accepted
    Ok(x) => ...
}

[Alonely0]

That doesn't seem correct. First I can very well make a ptr: *const Result<T, !>. And until opsem rules it out, I could find a way to write to ptr in a way that initializes the discriminant to Err. So far, no uninhabited type has ever been read so we're good. Then if I read the discriminant, I get Err which is still good. The case that would be UB is if I then read the data part of the enum.

But you are reading it! That's where our disagreement comes from, and it's not even the root of it. You're reading it when you dereference the pointer, because you're not reborrowing it (you then discard/drop it later with _, but that's irrelevant). The thing is, when the compiler & the optimizer see the Err variant, they see that in order for the Err to exist, a never type had to be created beforehand, because it is a dependency of Err (unlike C, where everything is wrapped in a MaybeUninit). The semantics of the never type are simple, nothing can exist after it is created; but, however, if we go back to the definition of the Err variant, the never type had to exist before, so the Err exists after the never type! This violates the semantics of the never type, so the path where the Err variant matches is unsound; Q.E.D. You can indeed write the ptr so that the data is invalid, but then even thinking about it as that enum is unsound (and even if unused, unused data must also be valid). That's proof by contradiction, this is proof by miri (the playground is down at the time of writing, so I'll just paste here the code):

fn main() {
    let mut x = Foo::<i32, Uninhabited>::X(1);
    assert_eq!(std::mem::size_of_val(&x), std::mem::size_of::<i32>() * 2); // Uninhabited is a ZST.
    assert_eq!(std::mem::size_of::<()>(), std::mem::size_of::<Uninhabited>()); // () is also a ZST.
    let mut y = Foo::<i32, ()>::X(1);
    assert_eq!(x, unsafe { std::mem::transmute(y) }); // Ergo, reprs are the same.
    y = Foo::Y(());
    let mut x = unsafe { std::mem::transmute::<_, FooLayout<i32, Uninhabited>>(x) };
    x.tag = 1; // Y's tag.
    let x = unsafe { std::mem::transmute::<_, Foo<i32, Uninhabited>>(x) };
}

#[derive(Copy, Clone, Debug, PartialEq)]
enum Uninhabited {}

#[repr(C, i32)]
#[derive(Debug, PartialEq)]
enum Foo<A, B> where A: Copy, B: Copy {
    X(A) = 0,
    Y(B) = 1,
}

#[repr(C)]
union UntaggedFoo<A, B> where A: Copy, B: Copy {
    x: A,
    y: B,
}

#[repr(C)]
struct FooLayout<A, B> where A: Copy, B: Copy {
    tag: i32,
    union: UntaggedFoo<A, B>,
}

When run with miri, notice that the write is, indeed, valid, but even if unused or unread, thinking about it as a Foo is already UB. On your initial piece of code you're checking if the discriminant is valid, and if it is not, you execute some code. However, when you do that, the potentially invalid data escapes the raw pointer, which is unsound.

[Nadrieril]

I do not understand your explanation. Anyway, what do you concretely propose we should do differently?

[Nadrieril]

I propose to split off the exhaustiveness-affecting changes into their own feature gate: #118803. If that gets accepted, then never_patterns will become exclusively about ! patterns (which makes a lot of sense).

[RalfJung]

@RalfJung you did? On this document right? I don't see any comments, where should I be looking?

That comment disappeared. So I guess you found them? :)

[Nadrieril]

I didn't but I realized hackmd was overall very buggy y'day for me so that was probably the cause

[RalfJung]

There should be a button for "comments" at the top:

If you click that it shows all the comments; you can click the comment to scroll to the place in the document it refers to.

[Nadrieril]

I see it! And it definitely wasn't there yesterday x)

[Nadrieril]

Hey sorry, I quote:

A tracking issue is however not meant for large scale discussion, questions, or bug reports about a feature.
Instead, open a dedicated issue for the specific matter and add the relevant feature gate label.

Please open a dedicated issue or come ask your question on Zulip.

[Nadrieril]

I have split off the exhaustiveness-affecting parts of this feature to min_exhaustive_patterns. This feature is now strictly about adding the ! pattern.

[Nadrieril]

RFC is up: rust-lang/rfcs#3719