Move calls to test_main into one function

[GuillaumeGomez]

Fixes #86254.

cc @jyn514
r? @camelid

[jyn514]

@bors r+ rollup

@GuillaumeGomez is there a reason you cc me on all these PRs?

[bors]

📌 Commit ef953b0 has been approved by jyn514

[GuillaumeGomez]

On this one it's simply because I know you like to keep an eye on rustdoc "backend" changes. But I can stop if the notifications annoy you, as you prefer. :)

[jyn514]

ahh, that makes sense :) no worries, the pings are fine

[camelid]

Thank you, this is definitely an improvement! It still seems surprising to me that there are two separate doctest-running functions (markdown::test and doctest::run), with significantly different contents. Why are they different, and should this PR still be marked as closing #86254?

[jyn514]

@camelid they're different entry points: doctest is used by rustdoc --test and markdown is used as a library by error_index_generator. I want to get rid of error index's dependence on rustdoc at some point so it should go away when that happens.

[camelid]

@camelid they're different entry points: doctest is used by rustdoc --test and markdown is used as a library by error_index_generator. I want to get rid of error index's dependence on rustdoc at some point so it should go away when that happens.

Hmm, Guillaume mentioned previously that markdown is also used by rustdoc foo.md (i.e., calling rustdoc on a Markdown file instead of a Rust file), so I'm not sure if we'd be able to get rid of it.

[Mark-Simulacrum]

Both of them are invoked from main rustdoc (rustdoc supports rustdoc --test foo.md on the CLI); it seems like the difference is whether the input source text is Rust code (and so needs to be parsed & expanded to get at the doc tests) or Markdown (so needs to be parsed, but in a completely different way).

I don't think either could go away, but it does seem a little suspicious that this PR is not unifying the (seemingly overlooked?) support for unused externs in markdown tests. I at least would expect that to work just as well regardless of the input type...

[GuillaumeGomez]

Just like @camelid said: one is run from rust code (we iterate through hir items and then check into their documentation if they have tests) whereas the second directly check from a markdown file (which is used by the error index as @jyn514 said). So we need to keep both, the goal here is to simply to merge the parts which can be merged.

[GuillaumeGomez]

Why are they different, and should this PR still be marked as closing #86254?

And yes, this is the part of the code which I referred to when I opened this issue. ;)