[WIP] Add downcast_trait and downcast_trait_mut #144363

ivarflakstad · 2025-07-23T14:11:25Z

Tracking issue: #144361

library/coretests/tests/intrinsics.rs

oli-obk · 2025-07-23T14:22:32Z

library/core/src/any.rs

+#[must_use]
+#[unstable(feature = "downcast_trait", issue = "144361")]
+pub const fn downcast_trait_mut<
+    T: Any + 'static,


does the Any bound actually prevent anything?

We'll need some ui tests that actually error when downcast_trait gets mis-used.

Some general test ideas:

passing a dyn Trait for the T where U is also dyn Trait

passing a dyn UnrelatedTrait for T when U is dyn Trait

using a not dyn-safe trait for U

using a dyn Trait<AssocTy = u32> when the type only implements Trait<AssocTy = bool>.

Also: Test out:

downcasting from two disjoint supertrait impls (trait Foo<T, U>: Super<T> + Super<U>).

Downcasting when the type doesn't implement an Auto trait.

This is also I think still unsound because we don't (and cannot) enforce that the lifetime in the input and output are equal. This should allow you to unsoundly cast dyn Foo + 'a to dyn Foo + 'static, right?

The 'static bound maybe prevents potential unsoundness.

Oh, then that really limits the usefulness of this intrinsic as an alternative to specialization as noted in the tracking issue 🤔

As @oli-obk noted in another comment the intrinsic doesn't necessarily need 'static, but downcast_trait* does.

Hm, yeah, that is kind of similar... but the existing downcast is not, not at all.

Doesn't really warrant a new module either, right? Even with type_name

Just for clarity @RalfJung I just plopped it in any so I could keep going. Was the best candidate I saw at a glance, I'm very open to moving it :)

Adding it to unresolved questions in the tracking issue

How about std::convert?

library/core/src/intrinsics/mod.rs

compiler/rustc_const_eval/src/interpret/intrinsics.rs

oli-obk · 2025-07-23T14:37:48Z

library/core/src/any.rs

+    }
+}
+
+#[allow(missing_docs)]


Need some docs before merging

Absolutely. Just appeasing the commit hook ☺️

library/core/src/any.rs

compiler-errors · 2025-07-23T15:01:26Z

compiler/rustc_const_eval/src/interpret/intrinsics.rs

+                    .build_with_typing_env(ty::TypingEnv::fully_monomorphized());
+
+                let type_impls_trait = preds.iter().all(|pred| {
+                    let trait_ref = ty::TraitRef::new(tcx, pred.def_id(), [tp_ty]);


You should definitely be matching on the ExistentialPredicate and using .with_self_ty instead of just constructing a TraitRef here.

As far as I can tell, right now if I try to downcast to dyn Iterator<Item = ()>, I'm only checking that the type implements Iterator, not that it implements the projection predicate of <Self as Iterator>::Item == ().

That should also allow you to remove that def_id function you added to ExistentialPredicate too.

That's great!
This is the first time I've even seen these concepts so I'm glad you can detect mistakes like these hehe

I noticed that when dealing with TraitRefs and .with_self_ty there is often a call to tcx.erase_regions(trait_ref) following.
Inspected the code. Had a look at the documentation. Still not fully certain but it seems to be simplifying the definition you pass in without changing it's actual meaning - is that correct?
I guess it would help if I knew exactly what a region was in this context hehe

yes, you can erase regions as we already checked that all regions are 'static via the bounds.

compiler-errors

This is an interesting idea, but I think before we commit to this experiment or anything, we should definitely do a vibe check about whether this is ever possible to implement in our current codegen backends.

If not, then this doesn't really serve as an alternative to specialization, right?

edit: Oh wait, I understand. This is a constant intrinsic, so we should never need to impl the intrinsic at runtime. I think I see. This whole scheme should be documented regardless, since that's a subtle difference from the way that upcasting is implemented today, which sometimes happens at runtime and thus has backend support.

library/core/src/any.rs

library/core/src/intrinsics/mod.rs

Co-authored-by: Oli Scherer <github35764891676564198441@oli-obk.de>

RalfJung · 2025-07-23T16:08:22Z

edit: Oh wait, I understand. This is a constant intrinsic, so we should never need to impl the intrinsic at runtime. I think I see. This whole scheme should be documented regardless, since that's a subtle difference from the way that upcasting is implemented today, which sometimes happens at runtime and thus has backend support.

Yeah, the key point is that everything is static. You can't take an &dyn X and ask "does the underlying type implement some trait". You can only do this for a concrete, statically-known type.

That also makes it extremely confusing to put this near the existing downcast methods which work on the actual underlying runtime type.

theemathas · 2025-07-23T16:10:15Z

Does overflow in the trait solver cause a post-mono error?

theemathas · 2025-07-24T05:08:33Z

Edit: Same behavior in 8e9b17b

At commit 0da23e7 (the last commit that builds on my machine), this PR has weird behavior with overflows in the trait solver, seemingly causing an error only if a function makes it to codegen.

Code and error

#![feature(downcast_trait)]
#![recursion_limit = "8"]

use std::any::downcast_trait;

struct Wrap<T>(T);
struct Indirect<T>(T);

trait Trait {}

impl<T> Trait for Wrap<T> where Indirect<Indirect<T>>: Trait {}
impl<T> Trait for Indirect<T> where T: Trait {}
impl Trait for i32 {}

// Uncommenting the below line makes the code compile
// #[inline]
pub fn weird() {
    let x = Wrap(Wrap(Wrap(Wrap(1i32))));
    let b = downcast_trait::<Wrap<Wrap<Wrap<Wrap<i32>>>>, dyn Trait>(&x).is_some();
    println!("{b}");
}

error[E0275]: overflow evaluating the requirement `Indirect<Indirect<Wrap<i32>>>: Trait`
   |
   = help: consider increasing the recursion limit by adding a `#![recursion_limit = "16"]` attribute to your crate (`foo`)
note: required for `Wrap<Wrap<i32>>` to implement `Trait`
  --> src/lib.rs:11:9
   |
11 | impl<T> Trait for Wrap<T> where Indirect<Indirect<T>>: Trait {}
   |         ^^^^^     ^^^^^^^                              ----- unsatisfied trait bound introduced here
   = note: 6 redundant requirements hidden
   = note: required for `Wrap<Wrap<Wrap<Wrap<i32>>>>` to implement `Trait`

For more information about this error, try `rustc --explain E0275`.

ivarflakstad · 2025-07-24T08:02:19Z

@theemathas it's not compiling because I used the gh commit
suggestion functionality, which I am ok with because it's still very much WIP:)

Great catch wrt the trait solver overflow. I'm hoping the suggestions pointed out by @compiler-errors and @oli-obk will solve that.

bors · 2025-07-24T08:09:26Z

☔ The latest upstream changes (presumably #116707) made this pull request unmergeable. Please resolve the merge conflicts.

…ait().unwrap() as expected

…it it

ivarflakstad · 2025-07-24T21:54:29Z

I think the recent changes have addressed the unsoundness pointed out by @theemathas, but I still need to address the recursion issue that causes trait solver overflow.

theemathas · 2025-07-25T04:07:06Z

library/core/src/intrinsics/mod.rs

+#[rustc_intrinsic]
+pub const fn vtable_for<T, U: ptr::Pointee<Metadata = ptr::DynMetadata<U>> + ?Sized>()
+-> Option<ptr::DynMetadata<U>>;
+


Since you replaced the erased lifetimes with 'static, I believe that this function needs a T: 'static and a U: 'static bound.

We could make it unsafe to allow for best-effort checks for optimizations like the Fuse specializations

theemathas · 2025-07-25T17:55:12Z

I just realized that the weirdness with #[inline] isn't specific to this PR, so I've filed issue #144460

ivarflakstad added 9 commits July 13, 2025 15:35

Introduce vtable_for intrinsic. Passes hir analysis, far from functional

ca8e62d

Merge branch 'master' into vtable-for-intrinsic

4a00305

Add def_id() method to ty::Binder<.., ExistentialPredicate>

17b5402

Verify that type implements trait in vtable_for. Return None if not.

d40408e

Add initial downcast_trait impl

7a46b59

Add downcast_trait ui tests

46224d5

Add initial downcast_trait_mut impl

b16ea15

Add downcast_trait_mut ui tests

ff13405

Add actual tracking issue reference

93aa4f2

ivarflakstad mentioned this pull request Jul 23, 2025

Tracking Issue for downcast_trait #144361

Open

3 tasks

oli-obk self-assigned this Jul 23, 2025

oli-obk reviewed Jul 23, 2025

View reviewed changes

library/coretests/tests/intrinsics.rs Outdated Show resolved Hide resolved

This comment has been minimized.

Sign in to view

oli-obk requested changes Jul 23, 2025

View reviewed changes

oli-obk reviewed Jul 23, 2025

View reviewed changes

library/core/src/any.rs Show resolved Hide resolved

Remove redundant #[allow(dead_code)]

1249f38

compiler-errors reviewed Jul 23, 2025

View reviewed changes

RalfJung reviewed Jul 23, 2025

View reviewed changes

library/core/src/any.rs Show resolved Hide resolved

oli-obk reviewed Jul 23, 2025

View reviewed changes

library/core/src/intrinsics/mod.rs Outdated Show resolved Hide resolved

ivarflakstad and others added 6 commits July 23, 2025 17:54

update vtable_for U bounds

ad5fa57

Use existing typing env when building infer context

0da23e7

Co-authored-by: Oli Scherer <github35764891676564198441@oli-obk.de>

Remove redundant upcast

d9b8269

Co-authored-by: Oli Scherer <github35764891676564198441@oli-obk.de>

Document which discriminant we write

b08d57f

Co-authored-by: Oli Scherer <github35764891676564198441@oli-obk.de>

Document behaviour of writing Option<NonNull> pointer

db579d3

Co-authored-by: Oli Scherer <github35764891676564198441@oli-obk.de>

Safety comment

852c332

Co-authored-by: Oli Scherer <github35764891676564198441@oli-obk.de>

This comment has been minimized.

Sign in to view

Make unreachable path trigger ICE with span

8298c84

This comment was marked as resolved.

Sign in to view

ivarflakstad added 3 commits July 24, 2025 11:47

Improve type impls trait check

d9cb965

Add SAFETY comment to downcast_trait_mut

2919430

Merge branch 'master' into vtable-for-and-downcast-trait

8e9b17b

This comment was marked as resolved.

Sign in to view

ivarflakstad added 3 commits July 24, 2025 20:58

Turn predicate 'erased lifetimes into 'static

5b438f8

Add ui tests that were unsound previously and now fail on downcast_tr…

11748bf

…ait().unwrap() as expected

Restructure ui tests to highlight unsound path and how we no longer h…

c38ca5b

…it it

theemathas reviewed Jul 25, 2025

View reviewed changes

theemathas mentioned this pull request Jul 25, 2025

#[inline] affects whether consteval happens #144460

Open

[WIP] Add downcast_trait and downcast_trait_mut #144363

Are you sure you want to change the base?

[WIP] Add downcast_trait and downcast_trait_mut #144363

Conversation

ivarflakstad commented Jul 23, 2025 • edited by rustbot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

This comment has been minimized.

oli-obk Jul 23, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

compiler-errors left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

RalfJung commented Jul 23, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

This comment has been minimized.

theemathas commented Jul 23, 2025

Uh oh!

This comment was marked as resolved.

This comment was marked as resolved.

theemathas commented Jul 24, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

ivarflakstad commented Jul 24, 2025

Uh oh!

bors commented Jul 24, 2025

Uh oh!

This comment was marked as resolved.

ivarflakstad commented Jul 24, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

theemathas Jul 25, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

ivarflakstad commented Jul 23, 2025 •

edited by rustbot

Loading

oli-obk Jul 23, 2025 •

edited

Loading

compiler-errors left a comment •

edited

Loading

RalfJung commented Jul 23, 2025 •

edited

Loading

theemathas commented Jul 24, 2025 •

edited

Loading

ivarflakstad commented Jul 24, 2025 •

edited

Loading

theemathas Jul 25, 2025 •

edited

Loading