The performance bug seems to be fixed now.
As I expected several other benchmarks benefit
from the fix as well (decoy, patterns).
The new magic numbers on my machine are
XSLTC 1192
XT 471
Saxon 298
XalanJ 197
The infamous 'dbonerow' test doesn't need rewriting
into "match/if" anymore and 10 iterations take about 1 second
(900 MHz, J2SDK1.4) instead of 130 seconds as before.
That's not the end of the story 'cause I see bunches
of opportunities for further improvements.
That translets are about 20 times faster now
on 'dbonerow' than XT/Saxon/XalanJ is no big surprise:
since FORTRAN times compiled code ran faster than interpreted.
When eg. Saxon works on 'dbonerow'
the only bytecodes that are actually executed are the bytecodes
Mike Kay put into saxon.jar. In the case of XSLTC
some bytecodes come from prepackaged libraries,
but (hopefully) most of the "action" is handled by
bytecodes 'freshly' generated ('personalized'/specialized)
for the stylesheet at hand.
The new 'xsltc.jar' is available from me.
--Jacek
_________________________________________________________
Do You Yahoo!?
Get your free @yahoo.com address at http://mail.yahoo.com
---------------------------------------------------------------------
To unsubscribe, e-mail: [EMAIL PROTECTED]
For additional commands, email: [EMAIL PROTECTED]
