Rendering is significantly slowed down by complex clip setups - which can be caught like this:

diff --git a/vcl/headless/svpgdi.cxx b/vcl/headless/svpgdi.cxx
index 6f993db..aa43951 100644
--- a/vcl/headless/svpgdi.cxx
+++ b/vcl/headless/svpgdi.cxx
@@ -304,7 +304,10 @@ bool SvpSalGraphics::setClipRegion( const vcl::Region& i_rClip )
         m_bClipSetup = true;
     }
     else
+    {
+        SAL_DEBUG("Complex clipping regions pushed " << aRectangles);
         m_bClipSetup = false;
+    }
 
     return true;
 }

To reproduce run:

./tilebench `pwd` /opt/libreoffice/testfiles/CP41vsAOO401/groupshape.rtf

preferably under callgrind under master - and you'll see the issue.

We -try- hard to defer setting up an alpha clipping mask - since that is usually huge and also horribly expensive =)

Stupidly we do this even for rendering rectangles which -should- be easy to do; and (it seems) the biggest chunk of our time is spent filling those masks.

We should either:

a) stop creating un-necessarily complex clipping setups: I can't see why we would need anything more complex than a simple rectangle ~ever really =) quite possibly the code is just being obtuse and we can fix it higher up.

and/or

b) handle rectangle & line rendering better.

I attach a prototype patch here [ that incidentally screws up all the rectangle rendering - as you can see if you load groupshape.rtf in 'gtktiledrender' ;-] which needs some more thought as to why it breaks ~everything I guess ;-)

Quite possibly we need to do gradient rendering for the outline / border of the document - or somesuch in BitmapDevice itself - only profiling will tell.