Experiences comparing performance of XSLT vs. "raw Jav

I know this is a completely open question, but what kind of experience
have people had comparing the runtime performance and throughput of XSLT
vs. "raw Java code"?  I would guess the only scope limitation I could
give for this would be using Xalan and JDK 1.4.2.  Is there any
reasonable data, either within these constraints, or outside of them?  I
personally feel that XSLT will perform perfectly fine against raw Java,
even ignoring other advantages, like maintenance issues with complex
transformation code in Java, but I seem to be hearing resistance from
some people who claim it's a "known fact" that XSLT is slower than pure
Java.  That may be true in trivial examples, but I think real
transformations will likely show that XSLT is just as fast, or faster.