impl-everywhere.rs has a weird test

[Munksgaard]

From #54824 (comment):

[ ] impl-everywhere.rs:// @!has foo/fn.foo.html '//section[@id="main"]//pre' "x: &'x impl Foo"

This test looks strange. The generated documentation looks like this (wrapped in pre):

pub fn foo<'x>(
    x: &'x impl Foo
) -> &'x impl Foo

so it seems like the backslash in the test is taken literally which means that it doesn't match. Is that on purpose? @GuillaumeGomez it seems like you created the test, so perhaps you can enlighten me?

Edit: To be clear, removing the backslash from the test causes the test to fail.

My understanding is that test is trying to assert that fn.foo.html doesn't contain the phrase x: &' impl Foo, but because it is erroneously escaping a quote it fails to make that assertion. Furthermore, the generated fn.foo.html file actually does contain the text being checked for, so there is a bug in rustdoc as well.

[Munksgaard]

This issue should be assigned to @GuillaumeGomez, as per his request in #54824 (comment)

[liketechnik]

@rustbot claim

[liketechnik]

I've made the following observations:

• The current output of rustdoc for the test is the same as the output with the initial implementation used with the test (e67bba8) and also seems to match with the initial goal of the implementation in Fix missing impl trait display as ret type #53541. (I don't think the actual output of rustdoc is necessary here, but can provide that if needed).
• The check has not been modified since its creation, I haven't been able to compare the implementations of the relevant parts in rustdoc though, due to my unfamiliarity with rustdocs code.

My interpretation is that there is no bug in rustdoc (since the output stayed the same), but the test simply never validated the output correctly, because

• all checks validate the absence (all checks start with @!has) of an impl Type in parameters or return values, while the original goal of the implementation (see Fix missing impl trait display as ret type #53541) was to add support for impl Type in parameters and return values. Example:

  rust/src/test/rustdoc/impl-everywhere.rs

  Lines 11 to 13 in 430feb2

  	// @!has foo/fn.foo.html '//section[@id="main"]//pre' "x: &\'x impl Foo"
  	// @!has foo/fn.foo.html '//section[@id="main"]//pre' "-> &\'x impl Foo {"
  	pub fn foo<'x>(x: &'x impl Foo) -> &'x impl Foo {

  which would generate pub fn foo<'x>(x: &'x impl Foo) -> &'x impl Foo as documentation
• all checks expect an opening brace { after the return type, which is not generated by rustdoc. This makes the checks work, because now the 'expected' (see point 1) mismatch occurs.
• all checks with a lifetime were erroneously escaping a quote. Again, this makes the checks work, because now the 'expected' mismatch occurs.

If my observations are correct, I would go forward by opening a pr fixing the issues listed above for the test (e. g. making the test fail for possibly incorrect output by rustdoc in the future).

[jyn514]

@GuillaumeGomez can you confirm whether the summary above is correct?

[GuillaumeGomez]

Yes, it seems to be correct. Thanks for the summary @liketechnik ! Please feel free to open the PR. :)