Just FYI, this was a bug in the m68k backend since V0.9d. When copying structures their alignment was no longer correctly determined, so it always generated copy-code with move.b instead of move.l. Fixed now.

Thanks for looking into the problem but are you sure that was the only problem? I don't know what I was looking at in the disassembly without debug info but it looked like extra code for 32x32=32 was being generated which resulted in many extra 16x16=32 multiplies compared to the old code (possibly in parallel paths which would only increase code size). I wish bebbo would have posted his source code which could have been compiled with debug and made testing easier. Some people don't seem to appreciate enough the effort of the only supported Amiga compiler. Did you compare both the Dhrystone score and size of the new executable to the old vbcc generated executable?

On your page you mention a GCC 6.2.0 cross compiler. :shocked Where can that be found? I have been looking anxiously for a recent C++ compiler targeting Amiga for some time (to get C++11 features). :agree

Have a look here http://franke.ms/amiga/gcc.wiki

I haven't tried it myself. Doesn't look like full compiler, (i.e. no libc, stdc++ and I am pretty sure it won't know how to call amiga libraries) but it is still interesting.

@bebbo
I am also curious about this GCC 6.2.0 cross compiler. Your page mentions "Meanwhile I ported the patches to gcc 6.2.0 and it starts to work". Could you make those available ?