liballoc doc tests are "silently" failing

[glandium]

I noticed this in the output running python x.py test locally:

   Doc-tests alloc
error: no global memory allocator found but one is required; link to std or add #[global_allocator] to a static item that implements the GlobalAlloc trait.

and that didn't make tests fail.

Cc: @SimonSapin

[glandium]

Same with panic_unwind:

   Doc-tests panic_unwind
error: no global memory allocator found but one is required; link to std or add #[global_allocator] to a static item that implements the GlobalAlloc trait.

[SimonSapin]

This error message is printed when running rustdoc --test src/liballoc/lib.rs […], but before running any test. The code printing it is in src/librustc_metadata/creader.rs. It looks like rustdoc is using that code in order to load the crate and find doctests in it. But since it’s not actually compiling this crate, not having an allocator is not a problem.

So I think there are two issues:

• rustdoc keeps doing its thing even though there are errors. It probably should call something like abort_if_errors on its rustc::session::Session instance at some point.
• This error should not occur in the first place. One way to inhibit it would be to tell creader that we’re "compiling" an rlib (CrateTypeRlib) rather than some other crate type. rlibs do not need (yet) to have an allocator defined. This is possibly this line:

rust/src/librustdoc/test.rs

Line 76 in 4700e11

crate_types: vec![config::CrateTypeDylib],

The first issue could in some other situation hide a real problem, but in this case it makes the second one mostly harmless. The only bad consequence is some noise in the output of a test run.

[RalfJung]

Related: #54010

[jethrogb]

@rustbot modify labels: +T-rustdoc

[rustbot]

Error: Parsing label command in comment failed: ...ify labels|error: must have : or to as label starter at >| +T-rustdo...

Please let @rust-lang/release know if you're having trouble with this bot.