Action against software patentsGNOME2 LogoW3C logoRed Hat Logo
Made with Libxslt Logo

The XSLT C library for GNOME

Library internals

Main Menu
Related links
API Indexes

Table of contents


This document describes the processing of libxslt, the XSLT C library developed for the GNOME project.

Note: this documentation is by definition incomplete and I am not good at spelling, grammar, so patches and suggestions are really welcome.


XSLT is a transformation language. It takes an input document and a stylesheet document and generates an output document:

the XSLT processing model

Libxslt is written in C. It relies on libxml, the XML C library for GNOME, for the following operations:

  • parsing files
  • building the in-memory DOM structure associated with the documents handled
  • the XPath implementation
  • serializing back the result document to XML and HTML. (Text is handled directly.)

Keep it simple stupid

Libxslt is not very specialized. It is built under the assumption that all nodes from the source and output document can fit in the virtual memory of the system. There is a big trade-off there. It is fine for reasonably sized documents but may not be suitable for large sets of data. The gain is that it can be used in a relatively versatile way. The input or output may never be serialized, but the size of documents it can handle are limited by the size of the memory available.

More specialized memory handling approaches are possible, like building the input tree from a serialization progressively as it is consumed, factoring repetitive patterns, or even on-the-fly generation of the output as the input is parsed but it is possible only for a limited subset of the stylesheets. In general the implementation of libxslt follows the following pattern:

  • KISS (keep it simple stupid)
  • when there is a clear bottleneck optimize on top of this simple framework and refine only as much as is needed to reach the expected result

The result is not that bad, clearly one can do a better job but more specialized too. Most optimization like building the tree on-demand would need serious changes to the libxml XPath framework. An easy step would be to serialize the output directly (or call a set of SAX-like output handler to keep this a flexible interface) and hence avoid the memory consumption of the result.

The libxml nodes

DOM-like trees, as used and generated by libxml and libxslt, are relatively complex. Most node types follow the given structure except a few variations depending on the node type:

description of a libxml node

Nodes carry a name and the node type indicates the kind of node it represents, the most common ones are:

  • document nodes
  • element nodes
  • text nodes

For the XSLT processing, entity nodes should not be generated (i.e. they should be replaced by their content). Most nodes also contains the following "navigation" information:

  • the containing document
  • the parent node
  • the first children node
  • the last children node
  • the previous sibling
  • the following sibling (next)

Elements nodes carries the list of attributes in the properties, an attribute itself holds the navigation pointers and the children list (the attribute value is not represented as a simple string to allow usage of entities references).

The ns points to the namespace declaration for the namespace associated to the node, nsDef is the linked list of namespace declaration present on element nodes.

Most nodes also carry an _private pointer which can be used by the application to hold specific data on this node.

The XSLT processing steps

There are a few steps which are clearly decoupled at the interface level:

  1. parse the stylesheet and generate a DOM tree
  2. take the stylesheet tree and build a compiled version of it (the compilation phase)
  3. take the input and generate a DOM tree
  4. process the stylesheet against the input tree and generate an output tree
  5. serialize the output tree

A few things should be noted here:

  • the steps 1/ 3/ and 5/ are optional: the DOM representing the stylesheet and input can be created by other means, not just by parsing serialized XML documents, and similarly the result tree DOM can be made available to other processeswithout being serialized.
  • the stylesheet obtained at 2/ can be reused by multiple processing 4/ (and this should also work in threaded programs)
  • the tree provided in 2/ should never be freed using xmlFreeDoc, but by freeing the stylesheet.
  • the input tree created in step 3/ is not modified except the _private field which may be used for labelling keys if used by the stylesheet. It's not modified at all in step 4/ to allow parallel processing using a shared precompiled stylesheet.

The XSLT stylesheet compilation

This is the second step described. It takes a stylesheet tree, and "compiles" it. This associates to each node a structure stored in the _private field and containing information computed in the stylesheet:

a compiled XSLT stylesheet

One xsltStylesheet structure is generated per document parsed for the stylesheet. XSLT documents allow includes and imports of other documents, imports are stored in the imports list (hence keeping the tree hierarchy of includes which is very important for a proper XSLT processing model) and includes are stored in the doclist list. An imported stylesheet has a parent link to allow browsing of the tree.

The DOM tree associated to the document is stored in doc. It is preprocessed to remove ignorable empty nodes and all the nodes in the XSLT namespace are subject to precomputing. This usually consist of extracting all the context information from the context tree (attributes, namespaces, XPath expressions), and storing them in an xsltStylePreComp structure associated to the _private field of the node.

A couple of notable exceptions to this are XSLT template nodes (more on this later) and attribute value templates. If they are actually templates, the value cannot be computed at compilation time. (Some preprocessing could be done like isolation and preparsing of the XPath subexpressions but it's not done, yet.)

The xsltStylePreComp structure also allows storing of the precompiled form of an XPath expression that can be associated to an XSLT element (more on this later).

The XSLT template compilation

A proper handling of templates lookup is one of the keys of fast XSLT processing. (Given a node in the source document this is the process of finding which templates should be applied to this node.) Libxslt follows the hint suggested in the 5.2 Patterns section of the XSLT Recommendation, i.e. it doesn't evaluate it as an XPath expression but tokenizes it and compiles it as a set of rules to be evaluated on a candidate node. There usually is an indication of the node name in the last step of this evaluation and this is used as a key check for the match. As a result libxslt builds a relatively more complex set of structures for the templates:

The templates related structure

Let's describe a bit more closely what is built. First the xsltStylesheet structure holds a pointer to the template hash table. All the XSLT patterns compiled in this stylesheet are indexed by the value of the the target element (or attribute, pi ...) name, so when a element or an attribute "foo" needs to be processed the lookup is done using the name as a key.

Each of the patterns is compiled into an xsltCompMatch (i.e. an ''XSLT compiled match') structure. It holds the set of rules based on the tokenization of the pattern stored in reverse order (matching is easier this way).

The xsltCompMatch are then stored in the hash table, the clash list is itself sorted by priority of the template to implement "naturally" the XSLT priority rules.

Associated to the compiled pattern is the xsltTemplate itself containing the information required for the processing of the pattern including, of course, a pointer to the list of elements used for building the pattern result.

Last but not least a number of patterns do not fit in the hash table because they are not associated to a name, this is the case for patterns applying to the root, any element, any attributes, text nodes, pi nodes, keys etc. Those are stored independently in the stylesheet structure as separate linked lists of xsltCompMatch.

The processing itself

The processing is defined by the XSLT specification (the basis of the algorithm is explained in the Introduction section). Basically it works by taking the root of the input document as the cureent node and applying the following algorithm:

  1. Finding the template applying to current node. This is a lookup in the template hash table, walking the hash list until the node satisfies all the steps of the pattern, then checking the appropriate global template(s) (i.e. templates applying to a node type) to see if there isn't a higher priority rule to apply
  2. If there is no template, apply the default rule (recurse on the children as the current node)
  3. else walk the content list of the selected templates, for each of them:
    • if the node is in the XSLT namespace then the node has a _private field pointing to the preprocessed values, jump to the specific code
    • if the node is in an extension namespace, look up the associated behavior
    • otherwise copy the node.

    The closure is usually done through the XSLT apply-templatesconstruct, which invokes this process recursively starting at step 1, to find the appropriate template for the nodes selected by the 'select' attribute of the apply-templates instruction (default: the children of the node currently being processed)

Note that large parts of the input tree may not be processed by a given stylesheet and that conversely some may be processed multiple times. (This often is the case when a Table of Contents is built).

The module transform.c is the one implementing most of this logic. xsltApplyStylesheet() is the entry point, it allocates an xsltTransformContext containing the following:

  • a pointer to the stylesheet being processed
  • a stack of templates
  • a stack of variables and parameters
  • an XPath context
  • the template mode
  • current document
  • current input node
  • current selected node list
  • the current insertion points in the output document
  • a couple of hash tables for extension elements and functions

Then a new document gets allocated (HTML or XML depending on the type of output), the user parameters and global variables and parameters are evaluated. Then xsltProcessOneNode() which implements the 1-2-3 algorithm is called on the docuemnt node of the input. Step 1/ is implemented by calling xsltGetTemplate(), step 2/ is implemented by xsltDefaultProcessOneNode() and step 3/ is implemented by xsltApplyOneTemplate().

XPath expression compilation

The XPath support is actually implemented in the libxml module (where it is reused by the XPointer implementation). XPath is a relatively classic expression language. The only uncommon feature is that it is working on XML trees and hence has specific syntax and types to handle them.

XPath expressions are compiled using xmlXPathCompile(). It will take an expression string in input and generate a structure containing the parsed expression tree, for example the expression:


will be compiled as

Compiled Expression : 10 elements
    COLLECT  'child' 'name' 'node' chapter
      COLLECT  'child' 'name' 'node' doc
          EQUAL =
            COLLECT  'child' 'name' 'node' title
            ELEM Object is a string : Introduction
              COLLECT  'child' 'name' 'node' title

This can be tested using the testXPath command (in the libxml codebase) using the --tree option.

Again, the KISS approach is used. No optimization is done. This could be an interesting thing to add. Michael Kay describes a lot of possible and interesting optimizations done in Saxon which would be possible at this level. I'm unsure they would provide much gain since the expressions tends to be relatively simple in general and stylesheets are still hand generated. Optimizations at the interpretation sounds likely to be more efficient.

XPath interpretation

The interpreter is implemented by xmlXPathCompiledEval() which is the front-end to xmlXPathCompOpEval() the function implementing the evaluation of the expression tree. This evaluation follows the KISS approach again. It's recursive and calls xmlXPathNodeCollectAndTest() to collect a set of nodes when evaluating a COLLECT node.

An evaluation is done within the framework of an XPath context stored in an xmlXPathContext structure, in the framework of a transformation the context is maintained within the XSLT context. Its content follows the requirements from the XPath specification:

  • the current document
  • the current node
  • a hash table of defined variables (but not used by XSLT, which uses its own stack frame for variables, described below)
  • a hash table of defined functions
  • the proximity position (the place of the node in the current node list)
  • the context size (the size of the current node list)
  • the array of namespace declarations in scope (there also is a namespace hash table but it is not used in the XSLT transformation).

For the purpose of XSLT an extra pointer has been added allowing to retrieve the XSLT transformation context. When an XPath evaluation is about to be performed, an XPath parser context is allocated containing an XPath object stack (this is actually an XPath evaluation context, this is a relic of the time where there was no separate parsing and evaluation phase in the XPath implementation). Here is an overview of the set of contexts associated to an XPath evaluation within an XSLT transformation:

The set of contexts associated

Clearly this is a bit too complex and confusing and should be refactored at the next set of binary incompatible releases of libxml. For example the xmlXPathCtxt has a lot of unused parts and should probably be merged with xmlXPathParserCtxt.

Description of XPath Objects

An XPath expression manipulates XPath objects. XPath defines the default types boolean, numbers, strings and node sets. XSLT adds the result tree fragment type which is basically an unmodifiable node set.

Implementation-wise, libxml follows again a KISS approach, the xmlXPathObject is a structure containing a type description and the various possibilities. (Using an enum could have gained some bytes.) In the case of node sets (or result tree fragments), it points to a separate xmlNodeSet object which contains the list of pointers to the document nodes:

An Node set object pointing to

The XPath API (and its 'internal' part) includes a number of functions to create, copy, compare, convert or free XPath objects.

XPath functions

All the XPath functions available to the interpreter are registered in the function hash table linked from the XPath context. They all share the same signature:

void xmlXPathFunc (xmlXPathParserContextPtr ctxt, int nargs);

The first argument is the XPath interpretation context, holding the interpretation stack. The second argument defines the number of objects passed on the stack for the function to consume (last argument is on top of the stack).

Basically an XPath function does the following:

  • check nargs for proper handling of errors or functions with variable numbers of parameters
  • pop the parameters from the stack using obj = valuePop(ctxt);
  • do the function specific computation
  • push the result parameter on the stack using valuePush(ctxt, res);
  • free up the input parameters with xmlXPathFreeObject(obj);
  • return

Sometime the work can be done directly by modifying in-situ the top object on the stack ctxt->value.

The XSLT variables stack frame

Not to be confused with XPath object stack, this stack holds the XSLT variables and parameters as they are defined through the recursive calls of call-template, apply-templates and default templates. This is used to define the scope of variables being called.

This part seems to be one needing most work , first it is done in a very inefficient way since the location of the variables and parameters within the stylesheet tree is still done at run time (it really should be done statically at compile time), and I am still unsure that my understanding of the template variables and parameter scope is actually right.

This part of the documentation is still to be written once this part of the code will be stable. TODO

Extension support

There is a separate document explaining how the extension support works.

Further reading

Michael Kay wrote a really interesting article on Saxon internals and the work he did on performance issues. I wish I had read it before starting libxslt design (I would probably have avoided a few mistakes and progressed faster). A lot of the ideas in his papers should be implemented or at least tried in libxslt.

The libxml documentation, especially the I/O interfaces and the memory management.


redesign the XSLT stack frame handling. Far too much work is done at execution time. Similarly for the attribute value templates handling, at least the embedded subexpressions ought to be precompiled.

Allow output to be saved to a SAX like output (this notion of SAX like API for output should be added directly to libxml).

Implement and test some of the optimization explained by Michael Kay especially:

  • static slot allocation on the stack frame
  • specific boolean interpretation of an XPath expression
  • some of the sorting optimization
  • Lazy evaluation of location path. (this may require more changes but sounds really interesting. XT does this too.)
  • Optimization of an expression tree (This could be done as a completely independent module.)

Error reporting, there is a lot of case where the XSLT specification specify that a given construct is an error are not checked adequately by libxslt. Basically one should do a complete pass on the XSLT spec again and add all tests to the stylesheet compilation. Using the DTD provided in the appendix and making direct checks using the libxml validation API sounds a good idea too (though one should take care of not raising errors for elements/attributes in different namespaces).

Double check all the places where the stylesheet compiled form might be modified at run time (extra removal of blanks nodes, hint on the xsltCompMatch).


Thanks to Michael Sperberg-McQueen for various fixes and clarifications on this document!

Daniel Veillard