Does it know to optimise 68000 integer arithmetic used to implement long multiplication and the like?
It doesn't. It could be added but I don't think many apps spend that much time in multiplications/divisions and the like. Also, if the application is speed critical in this area it typically has an option to choose 68020 or higher CPU, or it automagically switches to faster routines when 68020+ is detected. If this really is an issue you can always use
http://aminet.net/package/util/misc/PatchFor020 manually.
At the very least I would have assumed that 68020 would be your best starting point since it doesn't require chains of instructions to implement certain operations you'd take for granted on 020+.
Indeed 68020+ is the best choice. I meant to say any of 68020/68030/68040/68060 is pretty much the same. I'll fix the post.
It should be pointed out however that since powerpc assembly is much more fine grained than the 680x0, the differences in common 68000 and 68020 code get optimized away. For instance in the following example where 68000 requires 2 instructions over 68020's 1...
68000:
lsl.l #3,d0
move.l 4(a0,d0.l),d0
68020:
move.l 4(a0,d0.l*8),d0
...will result in identical powerpc assembly (note that d0 is overwritten here, so the JIT knows that original d0 value is no longer used. Trance does very sophisticated analysis of the 68k code and can optimize away "unnecessary" code like this). Similarily for instance the new 020+ addressing modes eventually resolve the very similar code, regardless if you use separate 68000 compatible instructions or the new 020+ addressing mode.
But since this doesn't work for everything (for instance the code used for 32/64 bit multiplications and divisions), you're better off by choosing 68020+ version of binaries.