Saxon

Thanks, I have reproduced the issue so it shouldn't be difficult to solve.

I believe the correct answer is the one without optimization (5) - none of the elements has a preceding element named TEST2, therefore generate-id(preceding::TEST2) should return a zero-length string, so the predicate [generate-id(preceding::TEST2[1]) = generate-id(current()/preceding::TEST2[1])] is equivalent to ["" = ""] which is of course true.

The optimizer is turning the comparison generate-id(X) = generate-id(Y) into X is Y, which is incorrect in the case where both X and Y are empty sequences.

The class IdentityComparison has a flag generateIdEmulation which should be set when it is being used to compare generateId results rather than for a pure "is" test. It seems that this flag is set, but it is being ignored when the expression is "elaborated".

In passing I note that this flag doesn't seem to be reflected in the SEF export file -- which probably means that there are tests missing in this area.

I also observe that the filter expression isn't being classified as a "simple boolean filter", which means it doesn't take a fast path for such cases.

Now fixed except in the "export" case.

I've added XSLT test cases expression-0805 and expression-0806 that demonstrate this.