Pushing an empty path to a verbatim path no longer adds a separator

[dylni]

Code

I tried this code:

use std::path::Path;

fn main() {
    let mut path = Path::new(r"C:\Users").to_owned();
    path.push("");
    assert_eq!(r"C:\Users\", path.as_os_str());

    path = Path::new(r"\\?\C:\Users").to_owned();
    path.push("");
    assert_eq!(r"\\?\C:\Users\", path.as_os_str());
}

I expected to see this happen: Both assertions succeed.

Instead, this happened: The second assertion fails, likely due to #89270:

thread 'main' panicked at 'assertion failed: `(left == right)`
  left: `"\\\\?\\C:\\Users\\"`,
 right: `"\\\\?\\C:\\Users"`', src\main.rs:10:5

Although adding a separator when pushing an empty path is debatable, the behavior should probably be consistent for verbatim and non-verbatim paths.

This change caused normpath to break. I can add an inefficient workaround, but is this change intentional?

Version it worked on

It most recently worked on: 1.56.0-beta.4

Version with regression

rustc --version --verbose:

rustc 1.57.0-nightly (485ced56b 2021-10-07)
binary: rustc
commit-hash: 485ced56b8753ec86936903f2a8c95e9be8996a1
commit-date: 2021-10-07
host: x86_64-pc-windows-msvc
release: 1.57.0-nightly
LLVM version: 13.0.0

@rustbot modify labels: +regression-from-stable-to-nightly -regression-untriaged +O-windows

cc @seanyoung

[seanyoung]

This is an interesting case. The normpath crate has its own path push function to work around the issue of verbatim paths not working with PathBuf::push(), which is exactly the same issue that my PR #89270 was trying to solve upstream.

The normpath push() function uses pathbuf push function to append a directory separator by calling .push(""). Since the pathbuf push now already returns normalized paths (for verbatim paths) this no longer works.

I've created a PR on normpath to fix the issue there: dylni/normpath#4

So since verbatim paths have special handling to ensure the returned path normalized, yes there are differences. If you call .push(".") on a verbatim path, then the path will remain unchanged, otherwise it would not be valid any more. We cannot expect push to work the same on verbatim paths as on non-verbatim paths, verbatim paths must be normalized.

[ChrisDenton]

Hm, should this not be special cased so that it works the same for all paths? Otherwise the only way to push a trailing separator is to convert to an OsString, push the separator and then convert back to a PathBuf.

IMHO, ideally this would be fixed in components so that trailing slashes are never normalized away but I'm uncertain if that change will be accepted.

[seanyoung]

We could special case it so that .push("") is a no-op in all cases.

[ChrisDenton]

Would this break existing code though?

[seanyoung]

@ChrisDenton on second thought, maybe you are right. Let me see what I can do.