This commit removes, in retrospect, a silly use of `unsafe`. In particular,
to extract a file name extension (distinct from how `std` implements it),
we were transmuting an OsStr to its underlying WTF-8 byte representation
and then searching that. This required `unsafe` and relied on an
undocumented std API, so it was a bad choice to make, but everything gets
sacrificed at the Alter of Performance.
The thing I didn't seem to realize at the time was that:
1. On Unix, you can already get the raw byte representation in a manner
that has zero cost.
2. On Windows, paths are already being encoded and copied every which
way. So doing a UTF-8 check and, in rare cases (for invalid UTF-8),
an extra copy, doesn't seem like that much more of an added expense.
Thus, rewrite the extension extraction using safe APIs. On Unix, this
should have identical performance characteristics as the previous
implementation. On Windows, we do pay a higher cost in the UTF-8
check, but Windows is already paying a similar cost a few times over
anyway.
This adds a few tests that check for bugs reported here:
https://github.com/rust-lang/cargo/issues/4268
The bugs reported in the aforementioned issue are probably caused by not
enabling the `literal_separator` option in `GlobBuilder`. Enabling that
in the tests under question fixes the issue.
Closes#773
This threads the original glob given by end users through all of the
glob parsing errors. This was slightly trickier than it might appear
because the gitignore implementation actually modifies the glob before
compiling it. So in order to get better glob error messages everywhere,
we need to track the original glob both in the glob parser and in the
higher-level abstractions in the `ignore` crate.
Fixes#444
This commit completes the initial move of glob matching to an external
crate, including fixing up cross platform support, polishing the
external crate for others to use and fixing a number of bugs in the
process.
Fixes#87, #127, #131