Re: [xslt] slow key lookup in libxslt-1.1.22

["William M. Brack" <wbrack mmm com hk>]

Hi Josef,

Josef Urban wrote:
>
> Hi,
>
> I am getting a huge slowdown after moving from libxslt-1.1.15 to
> version 1.1.22 . I compiled both versions on the same machine,
> against the
> same libxml2 (2.6.30). The .xsl, .xml, profiles, and other files
> needed
> for re-running are at http://kti.ms.mff.cuni.cz/~urban/zfmodel1_ex/
> .
>
> The timings are:
>
> time ../xsltproc.1.1.22 --load-trace --profile addabsrefs.xsl
> zfmodel1.xml > /dev/null 2>profile.1.1.22
>
> real    1m40.778s
> user    1m40.650s
> sys     0m0.120s
>
> time ../xsltproc.1.1.15 --load-trace --profile addabsrefs.xsl
> zfmodel1.xml > /dev/null 2>profile.1.1.15
>
> real    0m3.679s
> user    0m3.560s
> sys     0m0.110s
>
> About 30 times slower. The profiles (and experimenting) suggest a
> bottleneck in a key lookup in the "abs" function, which basically
> just
> takes a name of a key $k (one letter code) and looks up an element
> using
> its number $nr: "key($k,$nr)".
>
> Any ideas what might be going on?
>
> Thanks a lot,
> Josef Urban

I spent quite a bit of time debugging this problem, including
low-level (gcc / gprof) profiling of both libxslt and libxml2, and I
think I have found and fixed it.

Since libxslt-1.1.15, there were many enhancements made to libxslt
(and libxml2), mostly aiming to increase efficiency, but also fixing
some problems.  The version which introduced this trouble was
libxslt-1.1.17 (released in June 2006), when it was decided to only
initialize "keys" for a document at the time they were first used. 
Unfortunately, there was a bug in the new code which resulted in the
problem you encountered.  Instead of only initializing the keys
once, with your stylesheet they were being "initialized" a very
large number of times!

My fixes are in SVN, and I would greatly appreciate it if you could
try them and confirm that everything is now okay.  According to my
tests, when compared to xsltproc from 1.1.15 the SVN version has a
16% improvement in "real" time, 5% improvement in "system", and 19%
in "user" :-).

Thanks very much for your report.

Bill