The reason I wrote this solution was to ask the obvious question about
memoisation. This solution is top-down and will be extremely
inefficient because the maximum depth of subtrees will be
re-calculated many times.
Saxon does have an option to memoise the results of xsl:function, but it's
never done by default. This is one of those situations where there's very
little information available to decide the right trade-off between memory
and compute time. Also, memoising is incorrect if the function returns a
newly constructed node (because identity is significant) and I need to do
more static analysis to spot such situations.
Michael Kay