GeezerVison

Hey @serhiy-storchaka, just for context: this change doesn't provide much of a speedup in itself, but it enables more significant speedups seen in GH-119186. I've separated it out here because it's an observable change that breaks a test. But I suspect that doesn't matter as the scandir() result order is arbitrary and the order of sibling visits can't be influenced. What do you think? Thank you!

If os.walk(topdown=False) reverses the order of visiting directories, this should be fixed.

It presently visits sibling directories in the same order they appear in dirnames.

What I'm proposing here is that we don't need to visit sibling directories in the same order they appear in dirnames in bottom-up mode, because:

1. The order is arbitrary (it comes from os.scandir())
2. Users aren't able to influence the order mid-walk by re-ordering dirnames (this is only possible in top-down mode)

Perhaps a worked example would be helpful? I can put one together.

Here's an example, first with the current main:

>>> os.mkdir('mydir')
>>> os.mkdir('mydir/subdir1')
>>> os.mkdir('mydir/subdir2')
>>> for entry in os.walk('mydir', topdown=False):
...     print(entry)
...     
('mydir/subdir2', [], [])
('mydir/subdir1', [], [])
('mydir', ['subdir2', 'subdir1'], [])

And with this PR:

>>> for entry in os.walk('mydir', topdown=False):
...     print(entry)
...     
('mydir/subdir1', [], [])
('mydir/subdir2', [], [])
('mydir', ['subdir2', 'subdir1'], [])

Note how the entry for subdir1 is now yielded before subdir2, and how this no longer corresponds to the order in dirnames for mydir.

If we think the new behaviour is acceptable (e.g. because we never guaranteed the order of sibling visits) then it unlocks an optimization: we can add paths to the stack while scanning directories rather than in a separate loop afterwards.

I think that it would be confusing if os.walk(), grep.grep(), or manual recursive use of os.listdir() and os.scandir() produce different results. The order depends on the file system, but at least it is consistent.

The fact that the current code needs to reverse the lists of directories is just an implementation detail. It is not exposed to the user. If the code used deque instead of a list, it would not need to reverse them.

BTW, how much it makes difference? What it in comparison of using deque?

That makes sense, thanks.

BTW, how much it makes difference?

Taking into account the speedups in GH-119186 and GH-121433, it's about 1% for os.walk() and 6% for os.fwalk(). Nothing special.

What it in comparison of using deque?

I'll get back to you.

Hm, the deque would help with a breadth-first implementation, but I don't think it helps here without a lot of code changes. I'll close this PR and its two dependents.