Tracking Issue for `core::mem::variant_count`

[cofibrant]

The feature gate for the issue is #![feature(variant_count)].

Steps

• Implement the intrinsic
• Adjust documentation (see instructions on rustc-dev-guide)
• Stabilization PR (see instructions on rustc-dev-guide)

Unresolved Questions

None

Implementation history

• Add unstable core::mem::variant_count intrinsic #73418
• Add note about non_exhaustive to variant_count #92568

[avl]

@RalfJung Responding to your response here instead, as per kennytm:s request.

The reason I asked about if it returned "largest discriminant + 1" is because such a behavior would be kind of reasonable with your motivation, about using the enum as an index into a Vec. There could be other situations where knowing the maximum discriminant value could actually be more important than knowing the number of elements. That said, maybe there should be two functions, one to get the maximum discriminant and one to get the number of elements. Or maybe the number of variants really is the only thing anyone ever cares about. Just wanted to bring up the question :-) .

The other point I was making was that perhaps an obligatory u64 is not necessary for the count. Other counts in rust use usize, and though I understand that you could in theory have 2^40 enum variants on a 32 bit machine, I don't think the compiler can really compile this even on a 64 bit machine right now. Also, I think having 2^40 enum variants on a 32-bit machine is sufficiently 'weird' that it seems reasonable for some friction to occur. I personally cannot imagine a scenario where an application would need more enum variants than the number of addressable bytes on the platform, though of course the fact I can't imagine it doesn't mean it doesn't exist :-) .

Really, there are some problems which are too big for some machines. As an example, a problem requiring a 100MB lookup-table is simply too big to process on a 16 bit machine. My argument is that a problem requiring a 2^40 variant enum is simply too big to process on a 32-bit machine. That said, a reasonable person could certainly disagree here!

Knowing the largest discriminant of an enum could be useful in these cases:

• Choosing a representation during serialization. Knowing the largest discriminant means being able to tell how many bits will be needed to serialize it.
• If an array is used as a sort of lookup-table with discriminants as keys, knowing the largest discriminant could be useful to be able to statically size the array correctly. Of course, such an approach only works well if the manually assigned discriminants are assigned in a mostly dense way.

[RalfJung]

I see, thanks.

I personally have no skin in this game, I doubt I will ever use that API.^^ I was just involved in reviewing. @doctorn would be the one to ask about motivation. :)

[Flying-Toast]

We should probably add a deny lint to rustc that prevents calling variant_count::<T>() where T is not an enum.

Regarding the API itself though, IMO it feels a bit janky:

• Calling variant_count where T isn't an enum is unspecified behavior (a lint would help with this)
• Getting the variant count of an enum with > usize::MAX variants is unspecified (although practically impossible, it seems strange to forbid something that's technically allowed by the language)
• Returns a usize, even if the repr is something else

What would be the drawbacks to doing this with a trait?

trait Enum {
    type Repr;
    /// Returns the maximum possible index into an array containing every variant of this enum.
    /// Returns `None` if the enum has no variants.
    const fn max_variant() -> Option<Repr>;
}

The compiler could automatically implement this trait for all enums, and it would be used like:

enum Foo { /* ... */ }

<Foo as core::intrinsics::Enum>::max_variant()

[Flying-Toast]

RE: @doctorn's comment in 73418:

you can’t use <T as DiscriminantKind>::Discriminant because if I have an enum with 256 variants the discriminant type will be u8 which can’t represent 256.

We could also use a similar API like my previous comment, but without a trait:

/// Returns the maximum possible index into an array containing every variant of the enum `T`.
/// Returns `None` if the enum has no variants.
fn max_variant<T>() -> Option<<T as DiscriminantKind>::Discriminant>

Combined with a rustc lint for non-enum Ts, I think that would get rid of all unspecified behavior. Thoughts?

[cofibrant]

Honestly the Enum trait sounded better to me. I much prefer having the number of variants returned rather than the maximum repr.

[Flying-Toast]

Honestly the Enum trait sounded better to me. I much prefer having the number of variants returned rather than the maximum repr.

Ok.

One more thing - the discriminant_value intrinsic seems to use u64s for discriminants; maybe variant_count should return a u64 to be consistent with that?

[lcnr]

the discriminant_value intrinsic seems to use u64s for discriminants;

https://doc.rust-lang.org/nightly/std/intrinsics/fn.discriminant_value.html

it does not, the docs are outdated on stable though '^^