Defunctorization helps, but from talking with Matthew Fluet, most of the perf comes from the combination of monomorphization and whole-program compilation. You get to avoid all the hackery involved with trying to mix essential inlining (e.g., map, foldl) with separate compilation and the somewhat unpredictable performance that results when a user accidentally writes their own map function but puts it in a separate source file without magical incantations for the inliner.
Also, we're working to get the rights for the definition back from MIT press so we can both push out a free PDF version and update the bugs. There are several corner cases that Harper, Tofte, and MacQueen consider mistakes in the '97 version of the definition, but haven't really taken the time to push out an updated definition, after the ML2000 effort went nowhere. Follow up with me offline (contact info in my profile) if you're morbidly curious.