After much experimentation, it turned out that the smallest program that would compile and run was only 5 bytes long:
main;
The craziest thing I got gcc (the AVR version, specifically) to compile for real purposes was a preprocessor macro that spit out tens of thousands of asm blocks with memory barriers and nop (do nothing) instructions with "PORTB = 1;" in the middle and "PORTB = 0" at the end. I needed it to bitbang out a clock signal to read an RFID tag on poorly documented hardware. (fun fact: the clock the cpu uses on an arduino uno is considerably worse and less accurate than the one that's on the board for the USB chip.)
It's enough to work for most things. Only annoying thing is, if you use the ICSP header to set the clock to external, you need a working clock to even use ICSP. (the 328pb falls back to the internal clock in this case)
But ultimately I gave up after the third time my computer crashed. I was at length 4 by that point, next step would have been to throttle the CPU usage.
In languages like Haskell you can just make 2+2=5: https://codegolf.stackexchange.com/a/28794