Which CPU model you were specifying for the compiler?
I tried -m68020-60 and -m68060 (on both compilers). Only difference against 68000 was the fact gcc aligned structure as the chunk on 4 bytes boundary but the member align was the same.
The whole data size affects the performance too (does it fit into cache) and
I think for that kind of stuff gcc is quite dumb and uses default values
(which programmer can override) which are reasonable for a wide variety of
programs.
Well, I don't know either. Maybe it's really some kind of compromise but I'd very curious why this compromise doesn't appear on x86 ABI as well :)