> But what do you mean with the second pass. Do I have to invoke the 
> transformer with another xsl or is it possible to invoke this second 
> transformation within the original xsl file?

It's surprising how rarely this technique is taught and discussed. Splitting
complex transformations up into a pipeline of simple transformations is
something that ought to be a standard design pattern used by every XSLT
developer.

There are two ways of doing it: multiple stylesheets, and multiple phases
within a single stylesheet. I use both approaches, often within the same
pipeline.

Multiple stylesheets can be linked into a pipeline in a number of ways:

* with your own custom Java code, e.g. using the JAXP interfaces

* with a pipeline processor such as Orbeon

* from a shell script

* Saxon has a custom extension, saxon:next-in-chain, that allows one
stylesheet to direct its output to be processed by another stylesheet

Within a single stylesheet, a pipeline is expressed as a series of
variables:

<xsl:variable name="phase-1-output">
  <xsl:apply-templates select="/" mode="phase-1"/>
</xsl:variable>

<xsl:variable name="phase-2-output">
  <xsl:apply-templates select="$phase-1-output" mode="phase-2"/>
</xsl:variable>

<xsl:template match="/">
  <xsl:copy-of select="$phase-8-output"/>
</xsl:template>

To do this in XSLT 1.0, you need the xx:node-set() extension.

Using multiple stylesheets gives you greater modularity and reusability, but
is a bit more complex to deploy. Importantly, it also allows you to
incorporate steps into the pipeline [I've never been sure what a step in a
pipeline should be called!] that are implemented using technologies other
than XSLT - for example, STX, XQuery, Java SAX filters, Perl scripts.

Michael Kay
http://www.saxonica.com/