Add Box<[T; N]>: IntoIterator without any method dispatch hacks

[compiler-errors]

Unlike Box<[T]> (#116607 (comment)), there's a much higher chance that this will not conflict with existing usages since it produces an iterator with the same type before/after this change, but let's test that theory with crater. Any breakage would need to be relying specifically on the expr Box::([...]).into_iter() yielding an iterator specifically of type array::IntoIter<T; N> as it currently does.

Ideally we have fewer migrations that are tangled up in hacks like rustc_skip_during_method_dispatch, so if this crater comes back clean, I'd strongly suggest landing this as-is.

As for the rationale for having this impl at all, I agree (as @clarfonthey pointed out in #124097 (comment)) that it is generally better for any user to not require moving the array out of the box just to turn it into an iterator.

[rustbot]

r? @Nilstrieb

rustbot has assigned @Nilstrieb.
They will have a look at your PR within the next two weeks and either review your PR or reassign to another reviewer.

Use r? to explicitly pick a reviewer

[compiler-errors]

@bors try @rust-timer queue

[compiler-errors]

why did i rust-timer this -- i plan to use crater lol

[bors]

⌛ Trying commit 78d9b1e with merge 2e5b773...

[compiler-errors]

Well -- I guess Box<[T; N]> technically dispatches to the &[T] impl on edition < 2021 because we skip the [T; N] impl. Anyways, let's see what crater says.

[bors]

☀️ Try build successful - checks-actions
Build commit: 2e5b773 (2e5b7739e5167b731ad5eb628f04a54f932f1fe2)

[compiler-errors]

@craterbot check

[craterbot]

👌 Experiment pr-124108 created and queued.
🤖 Automatically detected try build 2e5b773
🔍 You can check out the queue and this experiment's details.

ℹ️ Crater is a tool to run experiments across parts of the Rust ecosystem. Learn more

[craterbot]

🚧 Experiment pr-124108 is now running

ℹ️ Crater is a tool to run experiments across parts of the Rust ecosystem. Learn more

[craterbot]

🎉 Experiment pr-124108 is completed!
📊 4 regressed and 2 fixed (439535 total)
📰 Open the full report.

⚠️ If you notice any spurious failure please add them to the blacklist!
ℹ️ Crater is a tool to run experiments across parts of the Rust ecosystem. Learn more

[compiler-errors]

The only regression is derive-visitor 0.3.0:

[INFO] [stdout]  --> tests/test_containers.rs:8:19
[INFO] [stdout]   |
[INFO] [stdout] 8 | #[derive(Default, Drive, DriveMut)]
[INFO] [stdout]   |                   ^^^^^
[INFO] [stdout]   |
[INFO] [stdout]   = note: multiple `impl`s satisfying `Box<[CountMe1; 5]>: Drive` found in the `derive_visitor` crate:
[INFO] [stdout]           - impl<T> Drive for Box<T>
[INFO] [stdout]             where T: Drive;
[INFO] [stdout]           - impl<T> Drive for T
[INFO] [stdout]             where T: 'static, for<'a> &'a T: IntoIterator, for<'a> <&'a T as IntoIterator>::Item: derive_visitor::DerefAndDrive;
[INFO] [stdout]   = note: this error originates in the derive macro `Drive` (in Nightly builds, run with -Z macro-backtrace for more info)

This will not be fixed by adding any method dispatch hacks, so I think we should just land this as-is. I don't think we should block landing this impl on the one breakage.¹

Nominating this for T-libs-api for confirmation.

Footnotes

1. Also, I'm not confident that the set of impls that are present in derive-visitor are actually even coherent/sound. We've been aware of its breakage in the new trait solver for a while: https://github.com/rust-lang/trait-system-refactor-initiative/issues/52 ↩

[scottmcm]

Hmm, I didn't actually mean a TAIT. Let's see if I can digest it further.

Suppose we had struct alloc::box::IntoIter<…> such that on a Box<[T], A> you get box::IntoIter<[T], A> and on Box<[T; N], A> you get box::IntoIter<[T; N], A>. Then we make vec::IntoIter<T, A> a type alias for box::IntoIter<[T], A>.

The idea would be to think of that [T] vs [T; N] as a tail of the struct, but of course it's not literally a tail of it, so I don't know if it would actually work out. But if you could "unsize" a box::IntoIter<[T; N], A> into a box::IntoIter<[T], A>, that might help people avoid the monomorphization costs of the array if they want, while also letting people avoid the extra usize in the iterator if they want.

(I don't know if that would actually work, though.)

[WaffleLapkin]

@scottmcm what you are describing is actually quite possible: [playground] (finally, my love for unsizing paid off,,,). Basically what you need is just a CoerceUnsized implementation for box::IntoIter which coerces buf: *mut ArrayOrSliceGeneric (so that we can conditionally store length as metadata of the pointer, while allowing unsize coercion) (IIRC there is some work to make this nicer, but you can do that today too) (it currenlty requires a type like IntoIter<[T], T, A>, but I believe this is fixable with a smarter check).

I really dig your idea, this sounds nice :)

(P.S. to be clear I did not mean you meant TAIT, I just meant that it'd be possible to use TAIT)

[compiler-errors]

While that's a clever idea, it also seems like a lot of work to avoid a single usize from the vec::IntoIter impl. Unless someone on T-libs wants this to be implemented, I'd rather we don't, tbh.

[WaffleLapkin]

I don't think it's that much work, given that this would replace the existing vec::IntoIter with something only a bit more complex (most of the things don't need to change, only signatures and drop is different).

[scottmcm]

it also seems like a lot of work

I'm hoping it wouldn't be that bad, actually. Looking at vec::IntoIter,

rust/library/alloc/src/vec/into_iter.rs

Lines 45 to 61 in e59f2c5

	pub struct IntoIter<
	T,
	#[unstable(feature = "allocator_api", issue = "32838")] A: Allocator = Global,
	> {
	pub(super) buf: NonNull<T>,
	pub(super) phantom: PhantomData<T>,
	pub(super) cap: usize,
	// the drop impl reconstructs a RawVec from buf, cap and alloc
	// to avoid dropping the allocator twice we need to wrap it into ManuallyDrop
	pub(super) alloc: ManuallyDrop<A>,
	pub(super) ptr: NonNull<T>,
	/// If T is a ZST, this is actually ptr+len. This encoding is picked so that
	/// ptr == end is a quick test for the Iterator being empty, that works
	/// for both ZST and non-ZST.
	/// For non-ZSTs the pointer is treated as `NonNull<T>`
	pub(super) end: *const T,
	}

if we can "just" change that NonNull<T>+usize to be either NonNull<[T]> or NonNull<[T; N]> accordingly, surprisingly little else would have to change, and it'd save the extra usize. All the iterator stuff working on the ptr+end fields wouldn't have to change at all, for example.

libs-api folks, if the code needed to make it happens turned out to not be too bad, might you be interested? Or is it not worth bothering making a sample PR because you'd not want it anyway?

[rfcbot]

The final comment period, with a disposition to merge, as per the review above, is now complete.

As the automated representative of the governance process, I would like to thank the author for their work and everyone else who contributed.

This will be merged soon.

[compiler-errors]

I believe this is ready for review. I don't think it's worth complicating the implementation by adding a new unsizeable iterator just to get rid of the cap: usize field, but regardless, that should be done in a follow-up PR if so imo.

r? libs-api
@rustbot ready

[scottmcm]

that should be done in a follow-up PR if so imo.

That would be a breaking change to a stable feature, so I don't think it can be a follow-up PR.

Once the trait impl exists returning vec::IntoIter, it's too late to change to a different type.

[compiler-errors]

That would be a breaking change to a stable feature, so I don't think it can be a follow-up PR.

It's effectively equally as breaking as introducing this implementation, and we have until the end of the beta bump to decide whether we want this custom implementation in practice. The only code that would break is code that relies on Box<[T; N]>::IntoIter == vec::Iter<T>, and from the crater run we see that's no code currently.

if we can "just" change that NonNull+usize to be either NonNull<[T]> or NonNull<[T; N]> accordingly, surprisingly little else would have to change, and it'd save the extra usize.

I don't think this is possible given the current rules for how CoerceUnsized works. A correct implementation is going to be necessarily as complicated as the example that @WaffleLapkin shared, and I really don't think that is worthwhile. I believe the primary motivation for the Box<[T; N]>: IntoIterator impl existing at all is to avoid a move-out of the heap, and that we're gaining very little by getting rid of that additional cap: usize field.

[scottmcm]

I did some explorations of the unsizing version, and unfortunately it looks like the checker for what unsizing is allowed to do doesn't normalize associated types, so even if the source and target types share the same generic parameter as the not-being-unsized field, the compiler still doesn't let you do it.

So the clever version appears not currently feasible, but the simpler version of just having a "static capacity or runtime capacity" would be easy.

[Amanieu]

We discussed this in the libs-api meeting today. In the end we concluded that we would prefer for Box<[T; N]>'s iterator to be a separate type which include N as a generic parameter. While re-using vec::IntoIter may bring some benefits in compilation time, these are likely to be insignificant considering how little used this type would be. On the other hand the size reduction from omitting the capacity may be useful for some users in performance-critical code.

[bors]

☔ The latest upstream changes (presumably #124097) made this pull request unmergeable. Please resolve the merge conflicts.

[JohnCSimon]

@compiler-errors

ping from triage - can you post your status on this PR? This PR has not received an update in a few months.

[clarfonthey]

Since compiler-errors took over the larger compiler change, I can take over this one if you need to offload some of the work. Shouldn't be too hard to write up a simple boxed version of array::IntoIter.

(Only if you want, though; just an offer.)

[compiler-errors]

ack, @clarfonthey: I totally missed your message 😭. Go ahead, I probably will not be able to get around to this one... but it also seems to not require any T-compiler involvement.

[clarfonthey]

Right ack at you, will work on my own version of this change.

Not sure whether I want to base it off of vec::IntoIter or array::IntoIter, but I'll figure it out.