This doesn't work well when XML is not hand-edited, but rather a dump of some in-memory structure - there are many cases where e.g. order is not important, so it can be essentially random during serialization, and that then shows up as spurious diffs.