A postscript concerning point (c) in my conclusions: as a result of implementing this improving the value comparison logic to create a customized Comparator at compile time, I'm seeing a doubling in the speed of a couple of the XMark benchmark queries that do a lot of numeric comparison. The biggest benefit is in the schema-aware version of the queries where sufficient type information is available to enable this. So the promise that schema-awareness should enable improved optimization is slowly starting to come true.