@JelleZijlstra I can't assign a reviewer but you merged the prior PR related toos.walk so I decided that it makes sense to ping you here. I hope that's ok :))

Do you have a benchmark that shows this actually makes it faster? We lose a tuple allocation but gain a function call, so it's not completely obvious this should be faster.

I made this MR because my initial benchmarking withpathlib.Path.walk andos.walk was quite obvious. However, after timing it a few more times I have gotten inconsistent results.

Seems like the isinstance version is just a tiny bit (insignificantly) faster but I would still argue that it makes more sense than the original because it is slightly simpler.

But what do you think?

Thanks, seems like it's about a wash but since the new code is simpler and we never released the old version, seems fine to change this.