Correctly encode doc attribute metadata

[GuillaumeGomez]

Follow-up of #149645.

The change I made was slightly wrong. Originally the check was:

        if let Some(item_list) = attr.meta_item_list() {
            for item in item_list {
                if !item.has_name(sym::inline) {

So we were checking that there was no doc(inline) attribute. This PR should fix it.

r? @lqd

[rustbot]

Some changes occurred in compiler/rustc_hir/src/attrs

cc @jdonszelmann

[fmease]

Is this regression user observable? If so, it would be nice to add a test exercising it.

[GuillaumeGomez]

Would be nice indeed. Gonna try to trigger a bug with it.

[lqd]

So that should bring back the crate metadata "lost" in the previous PR, right? We could do a perf run :3

r=me with the test if it's possible to craft one

[rustbot]

This PR was rebased onto a different main commit. Here's a range-diff highlighting what actually changed.

Rebasing is a normal part of keeping PRs up to date, so no action is needed—this note is just to help reviewers.

[GuillaumeGomez]

So the regression is actually observable but I can't reproduce it locally: the source links are not present in a lot of std items (like OnceCell). I tried x.py doc library in both stage 1 and stage 2 and they always have their source links. I tried with other approaches (always with reexports) but no luck.

So this PR is a very annoying mix of "believe me it works" and "we'll see once this is merged and std is regenerated if it fixed the issue". But if someone is able to write a reproducer, I'd be EXTREMELY thankful as this is a very big issue.

Anyway, PR should be ready. I implemented Encodable by hand (copied the expanded macro code) to exclude doc(inline) to mimic previous behaviour, but I'm pretty sure the original code was also wrong.

[GuillaumeGomez]

If I don't commit all changes, obviously things can't work...

[lqd]

So the regression is actually observable but I can't reproduce it locally: the source links are not present in a lot of std items (like OnceCell). I tried x.py doc library in both stage 1 and stage 2 and they always have their source links. I tried with other approaches (always with reexports) but no luck.

Would looking at the std docs of a try build reproduce the issue then? (CI or opt-dist may need to be checked to ensure that the docs will be built for a try build).

As far as the current implementation goes, it’s not the prettiest but it seems our hands are somewhat tied due to preserving the possibly unexpected current encoding w/r/t doc inlining. In that sense I don’t mind it too much as it fixes the issue, it seems, but would also appreciate @fmease’s opinion.

[Urgau]

I don't know if it's worth much, but if I take a look at PR CI generated doc (in the "Artifacts" section), I can see that the "Source" links are present.

[GuillaumeGomez]

Can you check for std::cell::OnceCell too please?

[Urgau]

Seems to be present as well.

[GuillaumeGomez]

So seems like the fix is all good. Drives me nuts to be unable to have a local reproducer though... I thought reexports of deps would be enough but doesn't seem so.