The XQuery FLWOR expression has semantics that are defined in terms of a sequence of tuples of variable bindings. This strange concept has never really had any direct representation in Saxon, because it simply isn't needed. The vast majority of XQuery 1.0 FLWOR expressions are equivalent to a nested loop delivering a sequence of items, like nested xsl:for-each instructions in XSLT. There's only one case in XQuery 1.0 that can't be handled using this model, and that's a FLWOR expression that has two or more "for" clauses, and an "order by" clause that doesn't break down into a major key that depends on the outer "for" variable, and a minor key that depends on the inner "for" variable. It's actually quite hard to construct a realistic example that exploits this feature, so the fact that the Saxon implementation is somewhat bizarre doesn't really matter too much, because it is very rarely needed.

XQuery 3.0, however, introduces several new clauses to the FLWOR expression. In fact, it's no longer a FLWOR expression because it now has an optional "count" clause, an optional "group by" clause, and an optional "window" clause - so it's now a FLWCGWOR expression. Not a term that is very likely to catch on. However, all these new constructs are defined essentially as operations that take a stream of tuples as input, and deliver a stream of tuples as output. So to implement these features, Saxon really needs to have some proper machinery for handling tuples. It's an odd state of affairs, having these composable constructs ("clauses") that aren't actually expressions, manipulaling things ("tuples") that aren't actually values, but that's the way XQuery works. It seems to have something to do with a relational algebra heritage, where operations defined in terms of tuples are of course the core stuff of the data model.

To understand what these strange things called tuples are, it might help to start with simple examples and gradually increase the complexity. Consider the expression for $x in //a where $x[@b] return ($x+1) Here there's only one variable, so the tuple is a 1-tuple. The formal model of FLWOR expressions is that there is a sucession of values of the variable $x (as many as there are elements selected by //a) acting as input to the "where" clause, and the output is another succession of values of $x (this time containing as many values as match the predicate [@b]). The "return" clause then takes this stream of 1-tuples as input, and turns it into a sequence of items that are now regular values which can be stored in variables, passed to functions, and so on. Saxon actually implements this succession of 1-tuples as a sequence of items (with pipelined evaluation, of course, so it never holds the whole sequence in memory); because they are 1-tuples, they can be treated as regular values and don't require any special treatment.

Now consider this expression: for $d in //dept, $e in //empl order by $e/salary return $d/name This is where things start getting weird, and where we need to understand tuples. The semantics are defined as delivering a sequence of ($d, $e) pairs (one pair for each department, employee combination), sorting this sequence of pairs by $e/salary, and then for each pair computing $d/name. Saxon does this by exploiting the "ObjectValue" class, which is a kind of XPath Item that encapsulates any Java object. It was designed to allow Java extension functions to return values that can be passed transparently to other Java extension functions, but it allows us to pack any kind of data we like into an Item, and this enables us to reuse all the machinery for handling sequences of items when we want to handle a sequence of tuples. In fact Saxon doesn't just pack the values of ($d, $e) into this composite pseudo-item, it also adds the values of the sort keys, and indeed the return value. Rather than sorting the input stream and then computing the return value for each tuple in the sorted input, we compute the return values for every item in the input before doing the sort. So we end up in effect with a sequence of values each of which has the form (salary, name); we sort this sequence on the value of salary, and then deliver the value of name. In contrast with the model described in the spec where the tuple holds the input variables ($d, $e), in the Saxon implementation the tuple holds the precomputed value of the return clause. Which means it doesn't actually need to contain the input values, because they will not be needed again. Precomputing the return value, and sorting all the return values along with the input tuples, is probably rather inefficient, at any rate in its use of memory. Also, it only works because a FLOWR expression can only have a single ORDER BY clause and a single RETURN clause. We therefore never need to implement a pipeline of operations working on the tuple stream; we only ever do one operation, which is sorting.

To implement the extended FLWOR expression, we probably need to do something that is closer to the model used to describe the semantics. In this implementation the "stream of tuples" can be represented simply as a pair of integers, the location on stack where the values of $d and $e are held. When we ask for the next tuple, the contents of these variables change, but we don't actually need to return anything: it's more of an advance() operation with side-effects than a get-next() operation. The evaluation of the FLWOR expression as a whole can now be modelled like this (it's easiest to think of it in push mode, though Saxon will more often be using pull): until (exhausted) { advance(); return compute-result(); } where advance() advances the tuple stream causing the values of the relevant variables $d and $e to be updated on the stack, and compute-result() evaluates the return clause based on these values of the variables. In the case where there's an "order by" clause, the "order by" pseudo-expression works like this: until exhausted() { advance(); compute-sort-keys(); add-to-list(L); } sort(L) and when it in turn is called to advance(), it gets the next tuple from the sorted list, and updates the slots for $d and $e in the local stack frame. The sorted list needs to contain the values of the variables ($d, $e) and the sort keys, as at present, but it does not need to hold the result value; that should be computed after sorting.

This looks like quite a substantial change, and it's because it feels fairly daunting that we haven't implemented this part of XQuery 3.0 in Saxon yet (the GROUP BY subset that's implemented in Saxon 9.3 is essentially the subset that can be implemented using conventional sequences of items, without resorting to the complexities of tuples. It also happens to match what XSLT implements, and what 99% of users actually need). However, I think I'm starting to understand what needs to be done to make this happen. Excuse me airing my thoughts in public, it always helps me to get my ideas straight to try and explain them to the world! I had been thinking of putting most of this off beyond Saxon 9.4, but I'm starting to feel more confident in tackling some of the changes sooner.

The fact that other XQuery WG members have started contributing test material is an encouragement: implementation is always far easier if there are test cases available before you start.

POSTSCRIPT: 1 August 2011 I have now implemented a tuple stream pipeline that handles the for, let, where, order-by, group-by, and count clauses of XQuery 3.0 - everything, in fact, except windowing. It all works quite neatly. It's a pull pipeline only at the moment, and there is currently no optimization. Rather than re-implement all existing optimizations of FLWOR expressions, in the short term I shall probably take the new FLWOR expression and rewrite it where possible into familiar forms that the optimizer already knows how to handle, so that the new tuple stream evaluation strategy is used only in cases where tuples are actually necessary, for example anything involving "group-by", anything with "order-by" over more than one range variable, anything with a "count" clause, and so on. Adding a push version of the pipeline should not be difficult, and is certainly worth doing for cases where the results of the FLWOR expression are written straight to the serializer. Michael Kay