Adding a comment next to #![feature(ptr_metadata,)] silences failing doctests

[raldone01]

Reproduce

https://github.com/ink-feather-org/trait_cast_rs/tree/80a0a0994dfdbf5e8b6978879ce3205188401df1

1:

git clone https://github.com/ink-feather-org/trait_cast_rs.git
git checkout 80a0a0994dfdbf5e8b6978879ce3205188401df1
cargo test --all-features

All tests pass!

2:

Remove the comment in line 43 of the README.md:

#![feature(
  ptr_metadata, // <- Remove this comment
)]
use trait_cast_rs::{
  make_trait_castable, TraitcastableAny, TraitcastableAnyInfra, TraitcastableAnyInfraExt,
};
#[make_trait_castable(Print)]
struct Source(i32);
...

Next run the tests again:

cargo test --all-features

Oh no! The test fails with E0119.

What is going on?!? I did not expect the comment breaking anything.

Meta

rustc --version --verbose:

rustc 1.85.0-nightly (21fe748be 2024-12-11)
binary: rustc
commit-hash: 21fe748be15271ea5804e0507cd699b675efe038
commit-date: 2024-12-11
host: x86_64-unknown-linux-gnu
release: 1.85.0-nightly
LLVM version: 19.1.5

[cyrgani]

The bug appears to be not that your test fails, but rather that it suceeds when you add the comment.
Here is a reduced example:

/*!
```rust
#![feature(
  foo, //
)]
```
*/

This as a sole lib.rs should fail after cargo test because foo isn't a feature, but it silently suceeds with the comment. Removing it causes the test to fail as expected.

In 1.81, this code still errored correctly, but it was accepted in 1.82 and above.
@rustbot label: +regression-from-stable-to-stable

[jieyouxu]

Would be good to get a more precise bisection and a more self-contained MCVE, looks not ideal.

[jieyouxu]

Wait on hold, @cyrgani to double-check, do you mean this as a doctest?

[jieyouxu]

searched nightlies: from nightly-2024-07-20 to nightly-2024-08-30
regressed nightly: nightly-2024-08-15
searched commit range: 80eb5a8...13a5289
regressed commit: e5b3e68

bisected with cargo-bisect-rustc v0.6.9

Host triple: x86_64-unknown-linux-gnu
Reproduce with:

cargo bisect-rustc --script=./check.sh --start=1.81.0 --end=2024-08-30 --regress=success

[jieyouxu]

cc @GuillaumeGomez as this bisects to #126245, do you know if this is intended behavior change?

[raldone01]

Intended behavior change

Good one. 🫠

It's true my doctest was broken but the // in the feature expression silenced the failure.
If I had kept the comment I would have never noticed that the doctest was failing.

---

Can't get a doctest to pass?

Add a nonexistent #[feature()] with //.

[onestacked]

Great now I know how to get all my doctests to pass!

[cyrgani]

Note that there seem to be 2 underlying problems:

1. Any doctest starting with a malformed inner attribute is silently accepted, even just

//! ```rust
//! #![
//! ```

2. I guess that line breaks inside the attribute are removed before comments are removed, so adding the comment creates a malformed attribute and 1. applies.

[GuillaumeGomez]

No it's not an intended behaviour change. ^^'

Gonna check what's going on.

[GuillaumeGomez]

Opened #134260.