So, ping-pong with Master Yoda, King of Optimization (aka Campbell 😛 ) kept going…
I must say I’m rather amazed and happy with the results! In latest patch, we have another 50% or so gain in performance, and we now only use 8 bytes per edge and 4 bytes per loop!
Btw, found out FBX export currently exports tessellated faces, which is bad and will need to be fixed before adding split normals export to this format. Will try to tackle this in the next days (if optimizations suggestions stop filling up my available time! 😉 ).
mont29 
Hello there. Really appreciate this feature you are working on.
Just wanted to send this link http://wiki.polycount.net/VertexNormal that shows an interesting use of normal handling (I know it’s not the target to hand manipulate the normals, but I thought would be interesting to show this page, perhaps you already knew this.
Either way, looking forward to use this on Blender. Keep up the good work!
Thanks for all! 🙂