From enh at google.com via openbsd-tech mailing list via pfg@: The existing test is wrong for LP64, where size_t has twice as many relevant bits as int, not just one. (Found by inspection by rprichard.)