gentext.template eats 1/3 of CPU

[Stefan Kost <ensonic@hora-obscura.de>]

hi,

I have a first version of a callgraph profiler for libxml/libxslt
(https://bugzilla.gnome.org/show_bug.cgi?id=612186). With this I looked
at profiles for applying the docbook stylesheets. These are the topmost
entries:

index % time    self  children    called     name
                0.043    0.380   1947/3287      gentext.template.exists [4]
                0.002    0.397    234/3287      *[object.title.template]
[49]
                0.380    0.380   1071/3287      gentext.template [0]
                0.000    0.397     17/3287     
part[object.title.template] [90]
                0.000    0.397     18/3287     
chapter[object.title.template] [102]
[0]    33.26    0.380    0.380   3287     gentext.template [0]
                0.380    0.380   1071/1071      gentext.template [0]
-----------------------------------------------
                0.009    0.149    786/3967      *[find.chunks] [19]
                0.010    0.154    738/3967      href.target.uri [18]
                0.012    0.241    754/3967      *[chunk-filename] [16]
                0.056    0.202   1638/3967     
*[recursive-chunk-filename] [2]
                0.001    0.136     35/3967      in.other.chunk [76]
                0.001    0.173     16/3967      process-chunk [55]
[1]    11.84    0.135    0.000   3967     chunk [1]
-----------------------------------------------

lines with [x] in begin are the templates. lines above the callers,
lines below the callees. Any idea how to make gentext.template less
slow? Or make it called less often. There sees to be many:
  match="*" mode="object.title.template"
and
  match="*" mode="object.title.markup"
leading to it, but I can't see any template that uses e.g.
mode="object.title.markup"? Any ideas?

Stefan