Future-proofing enums/structs with #[non_exhaustive] attribute #2008

clarfonthey · 2017-05-25T22:10:39Z

This is a post-1.0 version of #757, changed to use an attribute instead of a dedicated syntax. This would allow crates to more easily mark enums as "non-exhaustive," requiring downstream crates to add a wildcard arm to matches, so that adding new variants is not a breaking change.

Additionally, this syntax is extended to structs as well.

Rendered

Edit: Updated rendered link to rfcs repo version.

scottmcm · 2017-05-25T22:35:03Z

Another thing in favour of #[non_exhaustive] on structs: the non-pub field prevents construction, whereas the attribute could allow it through FRU. That would allow extensible builders, with struct Builder { a: b, c: d, ..Default::default() }. And it has the same ergonomic advantages as with enum, allowing in-crate matching on the exact set of fields without the bonus non_exhaustive: (), and not needing it on in-crate construction either.

clarfonthey · 2017-05-25T22:55:11Z

Because you mentioned it: if there's enough support for it, I'd be more than happy to promote the struct/trait parts into the main RFC. I've mostly kept what's there small so that we can at least get the enum part in.

Part of the reason why I did that is also ensuring that we choose a name that would work for structs/traits if/when they get added too.

KodrAus · 2017-05-26T00:08:27Z

I like it 👍

This RFC cements the current state of the world with respect to exhaustive matching: where a producer's enum/struct declares itself unexhaustively matchable, and a consumer must respect that with a catch-all pattern. That sounds reasonable to me, because it's the only way for a producer to be sure additions are non-breaking.

If I understand the RFC correctly, the #[non_exhaustive] attribute isn't adding hidden fields or anything. ie it's not making observable changes beyond the match semantics in the consumer, so I think an attribute is a fine way to do this.

sfackler · 2017-05-26T05:35:06Z

I feel a bit uneasy about using an attribute to control this. I'd prefer .. I think. Attributes don't normally change the semantics of a type that significantly.

dlight · 2017-05-26T08:19:04Z

I feel a bit uneasy about using an attribute to control this. I'd prefer .. I think. Attributes don't normally change the semantics of a type that significantly.

A tangential comment: #[async] is exactly an attribute that will change the semantics (of a function) significantly. So I think the naming scheme of attributes #[non_exhaustive] and #[async] (and the possibility of other attributes like this) should be discussed together, for consistency.

seanmonstar · 2017-05-26T08:49:50Z

A tangential comment: #[async] is exactly an attribute that will change the semantics (of a function) significantly.

I'd kind of expect to eventually have keywords async and await, I must assume they are currently an attribute and macro is because that's the easiest to implement and explore. But probably not the place to have this discussion.

kevincox

Excellent writeup. This sounds like a great change to me.

kevincox · 2017-05-26T12:22:16Z

text/0000-non-exhaustive.md

+Should there be a way to warn downstream crate users when their match is
+non-exuhaustive? What if a user wants to add a warning to their matches using
+non-exhaustive enums, to indicate that more code should be added to handle a new
+variant?


I would argue that if you expect users to handle every case then this isn't for you. You should not use this attribute and bump your major version. A warning on not handling a particular case is also incredibly likely to be low-value noise.

That's fair! I mostly carried this over from #757 because it was discussed a lot in the RFC there.

kevincox · 2017-05-26T12:24:41Z

text/0000-non-exhaustive.md

+## Extensions to structs
+
+It makes sense to eventually allow this attribute on structs as well. Unlike
+enums, it's very easy to make a struct non-exhaustive:


This seems like a good idea to me. However in this case non-exhaustive seems a less relevant keyword word because it is usually used to describe matching. incomplete, extensible or will_extend now start to sound better and work for both.

I'd also note here that the term "extensible" was used on the previous RFC.

kevincox · 2017-05-26T12:29:03Z

text/0000-non-exhaustive.md

+
+Tangentially, it also makes sense to have non-exhaustive traits as well, even
+though they'd be non-exhaustive in a different way. Take this example from
+[`byteorder`]:


This concept seems sufficiently different to deserve a different attribute.

I would say #[extensible] applies here just as well--as in extending the set of supertraits. In case we go for .. for enum and struct (which I think is a good idea), we could go with:

trait A: B + C + .. {}

You are right. extensible does sound good. Not only for extending traits but also extending the provided methods. And when I do look at it that way it does seem fairly related.

You're right that they seem sufficiently different, but I see them all as solving the same problem: instead of preventing a user from implementing/using something by creating a hidden field, this simply states that not all of the dependencies can be listed by the user.

clarfonthey · 2017-05-26T16:12:10Z

I feel a bit uneasy about using an attribute to control this. I'd prefer .. I think. Attributes don't normally change the semantics of a type that significantly.

That is fair! I mostly changed it to an attribute because at the time of the original RFC (#757), it was decided that adding extra syntax was not worth the benefits, and someone suggested that an attribute might be better.

That said, a lot has happened since the initial 1.0 release, and I think that people are more willing to accept a syntax addition for this now than when the original RFC was written.

kennytm · 2017-05-26T16:15:12Z

👍 extending this to structs, 👎 extending this to traits. The semantic change applied to traits is different enough that should take a different attribute or syntax (e.g. #[sealed]); I was expecting #[non_exhaustive] applied to traits means "this trait may gain additional methods in the future".

kennytm · 2017-05-26T16:22:29Z

If this RFC is accepted and stabilized, there should be a (clippy?) lint that suggests any C-like enum returned from extern function (-> E or &mut E) be annotated #[non_exhaustive], to prevent rust-lang/rust#36927 ("Match on repr(C) enum returned from C library with unknown value leads to UB").

leoyvens · 2017-05-28T17:57:06Z

Another thing in favour of #[non_exhaustive] on structs: the non-pub field prevents construction, whereas the attribute could allow it through FRU.

@scottmcm I don't see how the attribute would allow FRU, to the contrary it should forbid FRU outside the module that defines the struct. This should be in the RFC.

strega-nil · 2017-05-28T21:00:15Z

@kennytm that doesn't really make much sense. A match on a non-exhaustive enum outside the bounds would still be UB, and it's not uncommon for C enums to be treated like Rust enums - i.e., only within the bounds is okay.

kennytm · 2017-05-28T21:53:57Z

@ubsan a match on a non-exhaustive C-like enum, not arbitrary enum. This particular case should be definable.

clarfonthey · 2017-05-28T22:00:22Z

I'm pretty sure that casting an arbitrary integer to an enum, C-like or not, is undefined behaviour unless it corresponds to a valid variant. I don't think that annotating the enum differently is going to change this problem.

scottmcm · 2017-05-29T05:40:55Z

@leodasvacas Oops, you're absolutely correct. Allowing FRU would need it to use the alternative desugar. ("jFransham's Super-FRUs", I remember it being called...)

le-jzr · 2017-05-29T10:38:41Z

@clarcharr What makes it UB for C-like enums is that it fails the assumption that an exhaustive match will always execute one of its branches. If exhaustive match is impossible, missing labels pose no problem.

le-jzr · 2017-05-29T10:53:39Z

Or to put it simply, putting #[non_exhaustive] on C-like enum could be a statement that every value of the underlying integer type is valid, even if those values don't have labels. Insofar as that statement applies to all code, even within the same crate, which of course is not what's being proposed in this RFC.

clarfonthey · 2017-05-29T16:12:46Z

Note that the attribute does not say "all bit patterns of this enum are valid," only "more variants may be added in the future."

I don't think that this is a good solution to the problem you're suggesting.

le-jzr · 2017-05-29T16:29:34Z

The attribute says what the RFC says it says, and RFC is not merged yet. ;)

But yes, it's somewhat different, even if intimately related. The solution to C enum problem requires #[non_exhaustive], but it's a somewhat stronger guarantee. In particular, match code generation needs to ensure that all unnamed values fall into the wildcard branch, which is not necessarily true right now (as far as I know).

Ericson2314 · 2017-05-29T22:26:01Z

I do want some "real" solution for this. I want to be able to machine-check semver comparability, and the "doc trick" will be ugly to implement for any tool which does that.

Virtlink · 2017-05-31T08:47:00Z

This solution would be welcome, as it's a non-breaking change for Rust 1.x. But I think it's the wrong way around.

Non-exhaustive should be the default. This allows a crate author to add new enum variants and add new public and private fields to a struct, without breaking user's code. Of course, this forces users to handle the case where enums are extended, prevents users from using the default constructor to construct a struct, and also prevent users from destructuring the struct.

If the author really wants to enable the user to construct and destruct structs and exhaustively match enums, they have to promise that they won't add members to the struct or enum by finalizing it through some keyword or attribute (or otherwise risk breaking user's code).

Maybe for Rust 2.0?

bjorn3 · 2017-05-31T15:18:46Z

text/0000-non-exhaustive.md

+This pattern could again be solved by using `#[non_exhaustive]`:
+
+```rust
+#[non_exhaustive]


I thinkno_extern_impl is a better name, because exhaustivenes feels weird for traits.

petrochenkov · 2017-08-26T19:03:47Z

text/0000-non-exhaustive.md

@@ -417,6 +417,10 @@ Then we the only valid way of matching will be:
 let Config { 0: width, 1: height, .. } = config;
 ```

+We can think of this as lowering the visibility of the constructor to
+`pub(crate)` if it is marked as `pub`, then applying the standard structure
+rules.


This needs to apply to unit structs as well.

Will do; one sec.

petrochenkov · 2017-08-26T19:07:18Z

text/0000-non-exhaustive.md

+
+```
+#[non_exhaustive]
+pub struct Unit {}


Well, this seems to be equivalent to 3d56889#r135395315, but I'd still prefer to use the same wording for tuple and unit structs.

petrochenkov · 2017-08-26T19:18:10Z

LGTM now 👍

scottmcm · 2017-08-26T23:02:53Z

Huzzah! The @rust-lang/lang team has decided to accept this RFC.

To track further discussion, subscribe to the tracking issue here: rust-lang/rust#44109

Thanks for the RFC, @clarcharr!

@nikomatsakis

RFC 2008: Future-proofing enums/structs with #[non_exhaustive] attribute This work-in-progress pull request contains my changes to implement [RFC 2008](rust-lang/rfcs#2008). The related tracking issue is #44109. As of writing, enum-related functionality is not included and there are some issues related to tuple/unit structs. Enum related tests are currently ignored. WIP PR requested by @nikomatsakis [in Gitter](https://gitter.im/rust-impl-period/WG-compiler-middle?at=59e90e6297cedeb0482ade3e).

Add non_exhaustive RFC.

ac2b8be

clarfonthey mentioned this pull request May 25, 2017

Wishlist: marker for open-ended enums #943

Closed

aturon assigned nikomatsakis May 25, 2017

aturon added the T-lang Relevant to the language team, which will review and decide on the RFC. label May 25, 2017

kevincox approved these changes May 26, 2017

View reviewed changes

Incorporate non-exhaustive structs into main RFC.

25d8980

clarfonthey changed the title ~~Future-proofing enums with #[non_exhaustive] attribute~~ Future-proofing enums/structs with #[non_exhaustive] attribute May 26, 2017

Reword summary.

aae8a31

Add note on FRU.

c9c82cf

bjorn3 reviewed May 31, 2017

View reviewed changes

petrochenkov reviewed Aug 26, 2017

View reviewed changes

Add note on unit struct constructors.

aa2502f

petrochenkov reviewed Aug 26, 2017

View reviewed changes

clarfonthey added 2 commits August 26, 2017 15:14

Fix FRU.

d7c62c4

Wording.

09eeb1c

scottmcm mentioned this pull request Aug 26, 2017

Tracking issue for RFC 2008: Future-proofing enums/structs with #[non_exhaustive] attribute rust-lang/rust#44109

Closed

7 tasks

RFC 2008: Future-proofing enums/structs with #[non_exhaustive] attribute

60cf588

scottmcm merged commit edbb821 into rust-lang:master Aug 26, 2017

clarfonthey deleted the non_exhaustive branch August 27, 2017 03:13

petrochenkov mentioned this pull request Sep 1, 2017

Default Struct Field Values #1806

Closed

fstirlitz mentioned this pull request Sep 8, 2017

Support defining C-compatible variadic functions in Rust #2137

Merged

davidtwco mentioned this pull request Oct 19, 2017

RFC 2008: Future-proofing enums/structs with #[non_exhaustive] attribute rust-lang/rust#45394

Merged

Centril removed the S-waiting-on-team label Feb 24, 2018

tobz1000 mentioned this pull request Jul 10, 2018

Error type redesign hyperium/hyper#1131

Open

davidtwco mentioned this pull request Jul 19, 2018

RFC 2008 non-exhaustive enums/structs: Rustdoc rust-lang/rust#51854

Merged

Centril added A-data-types RFCs about data-types A-typesystem Type system related proposals & ideas A-privacy Privacy related proposals & ideas labels Nov 23, 2018

This was referenced May 12, 2019

Add #[non_exhaustive] documentation to TRPL. rust-lang/book#1955

Closed

Add non_exhaustive to reference. rust-lang/reference#609

Merged

daurnimator mentioned this pull request May 20, 2019

non-exhaustive enums ziglang/zig#2524

Closed

pheki mentioned this pull request Feb 24, 2020

Emit #[non_exhaustive] for Rustified enums? rust-lang/rust-bindgen#1554

Closed

kim mentioned this pull request Mar 12, 2020

Error Handling radicle-dev/radicle-decisions#2

Merged

wooorm mentioned this pull request Dec 11, 2022

Future-proof Options before 1.0 wooorm/markdown-rs#33

Open

ehuss mentioned this pull request Sep 26, 2024

Unable to match non_exhaustive enum tuple variants with rest pattern rust-lang/rust#130891

Open

bjorn3 mentioned this pull request Dec 12, 2024

Cannot use ..Default::default() with structs having private fields despite implementing a custom Default trait rust-lang/rust#134167

Closed

Future-proofing enums/structs with #[non_exhaustive] attribute #2008

Future-proofing enums/structs with #[non_exhaustive] attribute #2008

Conversation

clarfonthey commented May 25, 2017 • edited by scottmcm Loading

scottmcm commented May 25, 2017 • edited Loading

clarfonthey commented May 25, 2017 • edited Loading

KodrAus commented May 26, 2017

sfackler commented May 26, 2017

dlight commented May 26, 2017

seanmonstar commented May 26, 2017

kevincox left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

whitequark May 26, 2017 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

clarfonthey commented May 26, 2017 • edited Loading

kennytm commented May 26, 2017

kennytm commented May 26, 2017

leoyvens commented May 28, 2017 • edited Loading

strega-nil commented May 28, 2017

kennytm commented May 28, 2017

clarfonthey commented May 28, 2017

scottmcm commented May 29, 2017

le-jzr commented May 29, 2017

le-jzr commented May 29, 2017

clarfonthey commented May 29, 2017

le-jzr commented May 29, 2017 • edited Loading

Ericson2314 commented May 29, 2017

Virtlink commented May 31, 2017

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

petrochenkov commented Aug 26, 2017

scottmcm commented Aug 26, 2017

clarfonthey commented May 25, 2017 •

edited by scottmcm

Loading

scottmcm commented May 25, 2017 •

edited

Loading

clarfonthey commented May 25, 2017 •

edited

Loading

whitequark May 26, 2017 •

edited

Loading

clarfonthey commented May 26, 2017 •

edited

Loading

leoyvens commented May 28, 2017 •

edited

Loading

le-jzr commented May 29, 2017 •

edited

Loading