rustdoc: Hide #text in doc-tests

[jyn514]

Since #![attr] and #[attr] are the only valid syntax that start with #, we can just special case those two tokens.

Fixes #83284.

[rust-highfive]

r? @GuillaumeGomez

(rust-highfive has picked a reviewer for you, use r? to override)

[GuillaumeGomez]

Not really, for example:

let s = "hello
#because why not?";

[jyn514]

@GuillaumeGomez that's already broken today if you use

let s = "hello
# because why not?";

[jyn514]

Really the problem is that rustdoc chose an ambiguous syntax, it needs to both use something that's both an invalid token normally and also follow string open-closing marks. The string thing might be possible but it will be a real pain. I don't this should be blocked on that improvement given that's there's other valid code that's already broken by the # syntax.

[GuillaumeGomez]

Can't we use the ast parser to get some help there?

[jyn514]

Can't we use the ast parser to get some help there?

@GuillaumeGomez I think this may be possible, but it will be a lot of hard work for very little benefit. If you're worried about breakage, maybe it makes sense to do a crater run instead? I don't think the exact whitespace you can use with # was ever documented, and I would hate to accidentally stabilize it when we don't need to.

[jyn514]

@bors try (in case you want a crater run)

[bors]

⌛ Trying commit af6c320 with merge 9a35232...

[GuillaumeGomez]

No, my problem is more like that this isn't a real fix, more like a patchwork over a few others. I don't mind merging as is. But maybe in the long run, we should try to take a long at this to have a "definitive" fix.

[jyn514]

Ok, I opened #84478 for the proper fix. I still think this is a good change in the meantime that makes the behavior less surprising.

[GuillaumeGomez]

Fine by me, thanks for opening the issue and thanks for this PR!

@bors: r+ rollup

[bors]

📌 Commit af6c320 has been approved by GuillaumeGomez

[bors]

☀️ Test successful - checks-actions
Approved by: GuillaumeGomez
Pushing 9a35232 to master...

[SergioBenitez]

This change is not backwards compatible and has broken Rocket's test suite. Please revert.

Rocket has:

// This (long, contrived) form string...
# assert_form_parses! { Foo,
"[k:top_key][i][k:sub_key]name=Bobert&\
[k:top_key][i][k:sub_key]age=22&\
[k:top_key][i][sub_key]=1337&\
[top_key][7]name=Builder&\
[top_key][7]age=99",

// We could also set the top-level value's key explicitly:
// [top_key][k:7]=7
# "[k:top_key][i][k:sub_key]name=Bobert&\
# [k:top_key][i][k:sub_key]age=22&\
# [top_key][k:7]=7&\
# [k:top_key][i][sub_key]=1337&\
# [top_key][7]name=Builder&\
# [top_key][7]age=99",
# =>

Previously, the #s in the second string were stripped by rustdoc, and so the string passed to the macro did not contain the leading #s. Now, they are not stripped, and so a string with #s gets passed to the macro, causing the test to fail.

[m-ou-se]

What bors did here was.. interesting. The try build was still running when this PR was approved. And then it merged the try commit (the one with r=<try> instead of r=$reviewer) into the main branch without running the full CI for it first.

cc @rust-lang/infra

[kennytm]

feels like rust-lang/homu#47 😢