Performance regression for $x[$i]
Saxon 9.5 shows a performance regression when evaluating an expression such as
for $i in 1 to $n return $x[$i]
The cause is that the subscript expression is evaluated using a TailIterator, whose strategy depends on whether the iterator for $x is "grounded" or not. In this case the iterator is a ProgressiveIterator, which 9.4 decides is grounded, and 9.5 decides is not. In this instance the 9.4 strategy works better, despite forcing $x to be eagerly evaluated.
Reported by Gunther Rademacher on saxon-help list. See also MHK response on list.
#1 Updated by Michael Kay about 3 years ago
- Status changed from In Progress to Resolved
I've made a number of changes.
First, the MemoClosure ProgressiveIterator now reports itself as GROUNDED, which means that when subscripting access is made, the MemoClosure will be materialized in memory. This gets us back to the 9.4 measurements.
Second, to avoid eager evaluation of the MemoClosure, the subscript expression now checks whether the iterator is a ProgressiveIterator, and if it is, calls its itemAt() method which causes it to be evaluated as far as the requested item and no further.
Third, this logic is now replicated in the bytecode, which was previously not testing run-time conditions as carefully as the interpreted code was.
Finally, the bytecode is now improved in the case where it is statically known that the subscript is an integer; it avoids boxing the value and avoids the runtime isWholeNumber check, as well as the check against Integer.MAX_VALUE if the bounds are known statically.
The attached document contains the query and some measurements for different values of n.