url 0.5.10 and 1.5.1 beta regression

[est31]

url versions 0.5.10 and 1.5.1 regressed from stable to beta (full log 1, full log 2):

Aug 04 02:42:21.790 INFO kablam! error[E0560]: struct `__test::test::TestDesc` has no field named `allow_fail`
Aug 04 02:42:21.790 INFO kablam!   --> tests/unit.rs:20:1
Aug 04 02:42:21.790 INFO kablam!    |
Aug 04 02:42:21.790 INFO kablam! 20 | / fn size() {
Aug 04 02:42:21.790 INFO kablam! 21 | |     use std::mem::size_of;
Aug 04 02:42:21.790 INFO kablam! 22 | |     assert_eq!(size_of::<Url>(), size_of::<Option<Url>>());
Aug 04 02:42:21.790 INFO kablam! 23 | | }
Aug 04 02:42:21.790 INFO kablam!    | |_^ `__test::test::TestDesc` does not have this field
Aug 04 02:42:21.790 INFO kablam! 

cc @seanmonstar @SimonSapin @Hoverbear

[arielb1]

wtf? need to investigate

[SimonSapin]

TL;DR: this regression itself is not a bug in Rust. If something here is arguably a bug, it’s the loophole that allowed me to indirectly depend on unstable APIs on the stable channel.

This is servo/rust-url#371. It was fixed in servo/rustc-test#4 and servo/rust-url#372, we haven’t published a new version on crates.io since. It only affects compiling tests, not the library itself.

https://github.com/servo/rustc-test 0.1 is a copy of an old version of Rust’s src/libtest with #[unstable] attributes removed and a some other small changes to make it run on stable. When declaring the former as Cargo dependency, it ends up being used instead of the latter for extern crate test;.

#42219 added a new field to test::TestDesc, so rustc-test 0.2 conditionally compiles to include that field or not based on the rustc version and adds a constructor that works in both cases. This 0.2 change was needed because not all the relevant code is in the test crate. src/libsyntax/test.rs also generates a static &[TestDescAndFn] when rustc --test is used.

We’ve discussed all this before in https://internals.rust-lang.org/t/test-and-external-test-harnesses/3145. I now think that the future of “proper” external test harnesses involves procedural macros 2.0, but I’ve been waiting for them to be closer to stabilization to open that discussion.

[arielb1]

Is this also the cause of #43683 (in unic-idna-mapping)?

[SimonSapin]

Yes.

[arielb1]

The issue looks fixed in rust-url. Closing.