loperf: performance regressions

Hi there,

I was checking loperf results [1]
and found one huge performance regression (21 -> 102 bn pseudo cycles),
for a .doc document [2] between commits
947feaa896168430c78dd08dabee1447133f2740..8dd624128bf12e013c17b3b24c0b24135296bb8e
so I suspect it has something to do with SwTextBoxHelper.
Callgrind data:
http://dev-builds.libreoffice.org/callgrind_report/traces/onload-012_fdo
%2338513.doc-convert.log.gz

In that commit range, there are also smaller regressions for
- 061_fdo#68750.doc-convert
  - probably the same issue, only not so much visible
- 004_empty.odp-convert
  - unfortunately we don't have any non empty .odp files there,
there might be some regression in saving .odp files but probably not so
important to check and we can wait for bug report?

Btw, loperf already helped me once to notice regression caused by my
commit [3], fixed later [4], so it's quite useful I'd say.

Do you think it makes sense to let you know if I notice any regressions
as in this e-mail?
Presumably yes, hopefully people are interested in fixing them, so I
will try to help with that.

Best,
Matus

[1] http://dev-builds.libreoffice.org/callgrind_report
[2]
https://gerrit.libreoffice.org/gitweb?p=test-files.git;a=blob;f=loperf/012_fdo%2338513.doc;h=302e9173938cccd1da378bf3d45baa16bece94d0;hb=HEAD
[3]
http://cgit.freedesktop.org/libreoffice/core/commit/?id=a6b00d16eb27a5e7e31c721671001a909ecef960
[4]
http://cgit.freedesktop.org/libreoffice/core/commit/?id=16a62079018aea0e72636bdb00576487b4e830b9