It turns out that after making the changes to improve the type inferences Saxon was making on these queries, I didn't actually pick up the new version of the code for the subsequent measurement runs. So I missed the fact that these changes actually created a very significant improvement: from 10500 to 5900ms in the case of Q11, and from 4800 to 3090ms for Q12. Moreover (and this reinforces the message about the real benefits of schema-aware processing) the improved analysis found errors in queries Q10 and Q20 - these errors have been corrected on the XMark benchmark site at http://www.ins.cwi.nl/projects/xmark/Assets/xmlquery.txt but I hadn't picked up the corrected version. There's one bit of bad news, however: the join optimization on Q9 has stopped working. A performance tuner's work is never done...