I've looked into this now - I was expecting sort to be smart enough to do that manouvre itself. For those who don't understand what it's doing, it is forcing the sort routine to perform the -M operation only once per file, but at the overhead of a hash lookup per sort iteration (there are geometrically upon n sort iterations per n files). sort + block could be optimised for minimal calculation of each side of the <=> operator (or cmp as appropriate) -- after all it ought to know it's going to be doing that irrespective of what's in the block, so it's a safe optimisation for sort to be doing already. But the orcish manouevre appears to be necessary because no such optimisation has been implemented. Perhaps