> I don't understand your benchmark. Do you mean that os.walk() is slower
> with fast_bottom-up.patch because islink() is called or because I replaced
> "for entry in scandir(top):" with "entry = next(scandir_it)"?

No, sorry, I was making two separate comments: 1) the code's gotten quite a bit more complex (and if it needs to be that way for correctness, I'm okay with that), and 2) I'm surprised at how fast it still is.

> Are you testing the top-bottom or bottom-up?

My benchmark.py calls os.walk() with topdown=True, which is the default. I was testing the Python 3.4 version of os.walk() via listdir against your fast_bottom-up.patch.

I'm keen to look into this a bit further, but it won't be today.