[regression] cannot declare non-inline module inside a block

[Mark-Simulacrum]

[INFO] [stdout] error: Cannot declare a non-inline module inside a block unless it has a path attribute
[INFO] [stdout]  --> src/rcalc/lexer.rs:124:1
[INFO] [stdout]   |
[INFO] [stdout] 4 | mod rcalc;
[INFO] [stdout]   | ^^^^^^^^^^
[INFO] [stdout] 
[INFO] [stdout] error: aborting due to previous error

• https://crater-reports.s3.amazonaws.com/beta-1.44-1/beta-2020-04-26/gh/ayazhafiz.rcalc/log.txt
• https://crater-reports.s3.amazonaws.com/beta-1.44-1/beta-2020-04-26/gh/ayazhafiz.xela-lang/log.txt
• https://crater-reports.s3.amazonaws.com/beta-1.44-1/beta-2020-04-26/reg/rcalc-0.1.2/log.txt

cc @petrochenkov

Possibly related to changes in rustdoc, I believe both cases are inside doctests?

[petrochenkov]

Bisection points to 57e1da5.

[petrochenkov]

That's #69838 "Expansion-driven outline module parsing".

[petrochenkov]

Minimized:

/// ```
/// mod nonexistent;
/// ```
pub fn in_library_crate() {}

cargo test

Passes on stable, produces error: Cannot declare a non-inline module inside a block unless it has a path attribute on nightly.

[petrochenkov]

So, before #69838 Parser had this option recurse_into_file_modules that silently turned mod foo; into mod foo {}, and rustdoc happened to use it in doctests, and now the option is removed, and rustdoc no longer turns mod foo; into mod foo {}, and rustdoc tries to load it from a file and fails.

[petrochenkov]

Seems acceptable, the regression can only happen in doctests and making doctests behave more like regular Rust code is good.

[Mark-Simulacrum]

To be clear, mod foo; was turned into mod foo {} in blocks, right?

If so I agree that this seems like a bug fix that we should accept. I'm going to cc @rust-lang/lang just in case there's disagreement but it seems obvious to me that we should rather reject mod foo; in blocks than essentially silently ignore it.

[petrochenkov]

Everything in doctests is in a block, AFAIK?

However, if it wasn't, rustdoc would still turn mod foo; to mod foo {} due to the way it called parser for doctests.

[nikomatsakis]

I agree this seems like a bug fix, yes.