Saxon

The cost of matching a positional pattern such as

match="XYZ[P]/node()[last()]

appears to have increased substantially between 9.5 and 9.6, at least in the case where the predicate P is expensive to evaluate. In 9.5, P is evaluated only if the node being tested is the last child of an XYZ, but in 9.6, P is tested before testing whether the node is the last child.

Both strategies appear to b correct and it's not clear that one is superior to the other in all cases. In the absence of any cost-based optimization, or learning approaches, or parallel threads so we can report a mismatch as soon as either condition is found to be false, it's not clear what the best approach is.