You think gcc cannot compile efficient m68k binaries?
I don't think it can't, I know it can't. At least it never uses scaled indexed addressing where it would be faster than using a scratch register to calculate the offset and adding it to the base pointer. This was certainly the case in 2.95 and as far as I can tell all versions of 3. Might be fixed in 4, but I wouldn't bet on it.
It also has a propensity (in C++ mode) to add expensive exception context state save operations where they'd never actually be caught in unwinding, nor does it bother to check properly whether or not it actually needs to save off the entire FPU register set when doing so. Older versions emitted multiple copies of virtual function tables for the same abstract classes which were never link optimised away. I could go on...
Luckily, nothing I haven't found workarounds for
