On a side note: some processors are known to do internal optimizations
that are equal or even better then hand crafted key optimizations. I
believe for XSLT 1.0 the .NET processor is one of those highly
optimized ones, and for XSLT 2.0 I believe that Saxon starts to be
increasingly more optimized with each new version.

I'm still quite new to XSLT, so I keep "discovering things". The other
day, I discovered that given two algorithms to solve a problem, which
one performs better may be a question of which processor you use. I also
noticed that Saxon seems to scale best, despite JVM overhead with
respect to LibXSLT or Xalan-C.

For my work, I'm restricted to LibXSLT (so XSL 1.0), as that's what PHP
comes with support for. I'm happy with it, except here: