I just tested my benchmark code to see if it is doing what I think that is doing.

My eyes can certainly be deceived, but why does generating the list of keys in the sub and passing copy of that list to the caller appear to be faster than passing a ref to the caller and having him generate this list his own?