Notable improvements:

$ ./python -m timeit -s "from pathlib import Path" "list(Path.cwd().glob('*', follow_symlinks=False))"
2000 loops, best of 5: 180 usec per loop  # before
2000 loops, best of 5: 159 usec per loop  # after
# --> 1.13x faster

$ ./python -m timeit -s "from pathlib import Path" "list(Path.cwd().glob('**/*.py', follow_symlinks=False))"
5 loops, best of 5: 54   msec per loop  # before
5 loops, best of 5: 40.9 msec per loop  # after
# --> 1.32x faster

Everything else is about the same.

For whatever reason, every time I try to review this, I struggle to figure out what the change is doing :D

Since it doesn't require changing any test cases, and I know the tests cases are pretty thorough for this area, I don't think there's any reason to not sign off. Maybe trigger a buildbot run with the tag to make sure it doesn't behave strangely on any of those setups - they can occasionally be a bit unusual and find some edge cases.

Thanks Steve.

For whatever reason, every time I try to review this, I struggle to figure out what the change is doing :D

The algorithm might be worthy of a blog post at this point!

The main change is that we now filter partial paths through a regex corresponding to a partial pattern in _select_recursive, rather than complete paths through a regex corresponding to a complete pattern in PathBase.glob(). We can do this because previous parts have already been filtered by _select_children(), and so there's no need to re-filter them.

The secondary change (which includes the addition of _entry_str()) is to match against os.DirEntry.path directly, which allows us to skip construction of path objects for files that don't match.

Okay, today it made sense :) Guess I'm more awake right now. Reading the changes from the bottom up might have helped as well.

Personally, I don't think you can have too many comments in an algorithm like this, particularly when it's recursive and split between a couple of functions. I'll suggest a few comments that would've helped me, but I don't think there are any code changes needed.