RE: [xsl] How smart are the XSLT processors? Are there any XSLT processors that convert tree-recursive functions into efficient iterative procedures?

["Vladimir Nesterovsky" <vladimir@xxxxxxxxxxxxxxxxxxxx>]

This task, in fact, raises a different question about generator functions.
Consider a hypothetic function ex:fibonacci($Fn-1, $Fn-2) that returns an 
infinitive sequence of fibonacci numbers.
To use it, one just accesses its result sequence by index.

<xsl:function name="ex:fibonachi" as="xs:integer*">
  <xsl:param value="Fn-1" as="xs:integer"/>
  <xsl:param value="Fn-2" as="xs:integer"/>

  <xsl:variable name="Fn" as="xs:integer" select="$Fn-1 + $Fn-2"/>

  <xsl:sequence select="$Fn"/>
  <xsl:sequence select="ex:fibonachi($Fn, $Fn-1)"/>
</xsl:function>

Honestly, I'm not entirely sure if it's a legal technique in xslt, but 
definitely it's not supported in saxon,
and it would be good if it were supported, as it allows separation of 
iterator and iteration logic.

See 
http://www.nesterovsky-bros.com/weblog/2010/04/09/GeneratorFunctionsInXslt.a
spx
--
Vladimir Nesterovsky
http://www.nesterovsky-bros.com/


> > Are there any XSLT processors that convert tree recursive 
> > functions into more efficient iterative procedures?
> > 
> > /Roger
> > 
> > -------------------------------------------------
> >      Version #1: Tree Recursion
> > -------------------------------------------------
> >     <xsl:function name="ex:fibonacci">
> >       <xsl:param name="n" />
> > 
> >       <xsl:choose>
> >           <xsl:when test="$n eq 0">
> >               <xsl:value-of select="0" />
> >           </xsl:when>
> >           <xsl:when test="$n eq 1">
> >               <xsl:value-of select="1" />
> >           </xsl:when>
> >           <xsl:otherwise>
> >               <xsl:value-of select="ex:fibonacci($n - 1) + 
ex:fibonacci($n
> - 2)" />
> >           </xsl:otherwise>
> >       </xsl:choose>
> > 
> >     </xsl:function>
> >
> 
> I think it's unlikely that an XSLT implementor would optimize this to 
avoid
> the exponential nature of the algorithm. The reason is that it's 
difficult
> to find optimizations that would do this that are sufficiently general 
to
> optimize a significant number of real user-written programs - there's no
> point in spending effort on optimizations that only help with 
ivory-tower
> examples. 
> 
> The most realistic optimization that one might attempt for this function 
is
> to turn it automatically into a memo-function. That involves a 
space-time
> tradeoff; automating a choice that requires a space-time tradeoff 
requires a
> cost-based optimizer that understands the costs of space and time and is
> able to estimate how much of either will be used. It's easy to look at 
this
> particular function and produce an argument as to why it should be
> implemented as a memo-function. It's not at all easy to generalize that
> reasoning so it can produce a good answer for any function.
> 
> Optimization techniques are advancing all the time (just look at the
> advances with just-in-time approaches that benefit from learning about 
the
> behaviour of the code at run-time). But for this example, it's not there
> yet.
> 
> I'd be interested how the performance of this compares with your other
> implementation if you add saxon:memo-function="yes".
> 
> Regards,
> 
> Michael Kay
> http://www.saxonica.com/
> http://twitter.com/michaelhkay