Rustc asks for Reflect, which is unstable, when it could be asking for Any which is stable

[tinco]

Hi,

I had a latenight coding session and ran into this error:

src/components/mod.rs:35:69: 35:83 error: the trait `core::marker::Reflect` is not implemented for the type `T` [E0277]
src/components/mod.rs:35    let component_list : &'static SharedMutex<T> = *component_list_any.downcast_ref().expect("Invalid type for components list");
                                                                                               ^~~~~~~~~~~~~~

Seems like an easy fix, add Reflect to the type constraint of the T, but that yields this exception:

src/components/mod.rs:12:5: 12:25 error: use of unstable library feature 'reflect_marker': requires RFC and more experience (see issue #27749)
src/components/mod.rs:12 use std::marker::Reflect;
                             ^~~~~~~~~~~~~~~~~~~~

That #27749 looks like it won't soon be solved and probably #31844 will supersede it in some way. So this advice from the compiler of adding Reflect is not very nice.

Perhaps we could have it recommend implementing Any instead? Or at least mention Any would do it as well? Most rookies won't know that Any is basically Reflect+'static.

Here's my forum question about this topic with the code that threw the error.

[durka]

There is supposed to be a note:

rust/src/libcore/marker.rs

Lines 456 to 457 in 34fd686

	#[rustc_on_unimplemented = "`{Self}` does not implement `Any`; \
	ensure all type parameters are bounded by `Any`"]

But it's not triggering. It used to work on stable, but is broken on beta. cc @GuillaumeGomez

[GuillaumeGomez]

@durka: @arielb1 removed my implementation. I'll try to fix this case.

[GuillaumeGomez]

@tinco: Do you have a shorter code sample by any chance?

[jonas-schievink]

Reproduction:

use std::any::Any;

fn log<T>(value: &T) {
    let value_any = value as &Any;

    // try to convert our value to a String.  If successful, we want to
    // output the String's length as well as its value.  If not, it's a
    // different type: just print it out unadorned.
    match value_any.downcast_ref::<String>() {
        Some(as_string) => {
        }
        None => {
        }
    }
}

fn main() {}

<anon>:4:21: 4:26 error: the trait bound `T: std::marker::Reflect` is not satisfied [E0277]
<anon>:4     let value_any = value as &Any;
                             ^~~~~
<anon>:4:21: 4:26 help: see the detailed explanation for E0277
<anon>:4:21: 4:26 help: consider adding a `where T: std::marker::Reflect` bound
<anon>:4:21: 4:26 note: required for the cast to the object type `std::any::Any + 'static`
error: aborting due to previous error

Looks like it does mention Any.

[GuillaumeGomez]

That's what I got as well so I concluded it wasn't what he did. But maybe this is already solved in nightly?

[durka]

It mentions Any but the help says to add a T: Reflect bound which is unstable and the subject of this issue. There's a rustc_on_unimplemented note about it that isn't getting printed (see my source link above).

[durka]

This was tagged regression-from-stable-to-beta but that was a cycle ago so it is now a regression-from-stable-to-stable (albeit a quite minor one).

rustc 1.8.0 prints this in response to @jonas-schievink's code:

foo.rs:4:21: 4:26 error: the trait `core::marker::Reflect` is not implemented for the type `T` [E0277]
foo.rs:4     let value_any = value as &Any;
                             ^~~~~
foo.rs:4:21: 4:26 help: run `rustc --explain E0277` to see a detailed explanation
foo.rs:4:21: 4:26 note: `T` does not implement `Any`; ensure all type parameters are bounded by `Any`
foo.rs:4:21: 4:26 note: required for the cast to the object type `core::any::Any + 'static`
error: aborting due to previous error

The fifth line (first note) is not printed by 1.9.0 or later versions.

[GuillaumeGomez]

@durka: I updated the tag.

[nrc]

@rust-lang/compiler and/or @durka is this likely to get fixed or should we give up on this? Does seem a shame to live with this regression, but given that it's been on stable for a while (and Reflect will probably disappear anyway) perhaps it is not worth fixing?

[eddyb]

@nrc It's good idea to track the underlying diagnostics regression (@jonathandturner may be interested).

[durka]

Reflect is going to disappear but there is some underlying problem here with the #[rustc_on_unimplemented] getting dropped on the floor.

[nikomatsakis]

Do we have any code that actually reproduces the problem?

[arielb1]

current error message:

error[E0277]: the trait bound `T: std::marker::Reflect` is not satisfied
 --> <anon>:6:21
  |
6 |     let value_any = value as &Any;
  |                     ^^^^^
  |
  = help: consider adding a `where T: std::marker::Reflect` bound
  = note: required because of the requirements on the impl of `std::any::Any` for `T`
  = note: required for the cast to the object type `std::any::Any + 'static`

[nikomatsakis]

Conclusion for @rust-lang/compiler meeting: current behavior appears to be correct, so we believe problem is fixed, can call this I-needs-test.

[nikomatsakis]

triage: P-low

[nikomatsakis]

Actually, this was misunderstood. The problem is that at some point we decided (perhaps incorrectly) that this kind of message we see now is higher priority that #[rustc_on_unimplemented] messages, and that is why we do not see the custom message anymore. I am not convinced this is correct, but it was deliberate.

[durka]

OK, that's weird to me (why not print both), but if it was deliberate then this can be closed.

[brson]

Closing per previous comments.