Transforming XML Streams with References

Many useful XML transformations can be formulated through deterministic top-down tree transducers. As soon as transducers process parts of the input repeatedly or in an order deviating from the document order, such transductions cannot be realized over the XML document stream with constant or even depthbounded memory. Here we show that by enriching streams by forward references every such transformation can be compiled into a stream processor with a space consumption depending only on the transducer and the depth of the XML document. This is remarkable because tree transducers have rich restructuring capabilities and can process copies of input subdocuments independently. Also, references allow one to produce output in a compressed form which is guaranteed to be linear in the size of the input (up to the space required for labels). Our model is designed so that without decompression, the output may again serve as the input of a subsequent transducer. In order to reduce the extra overhead incurred by references, we investigate three optimizations: (i) reference reuse to save space, (ii) multi-labels to avoid chains of references and (iii) inlining to limit the use of references to deviations of the document order. 1998 ACM Subject Classification F.1.2 Modes of Computation