This looks like a good metric to track. Looking at some generated asm when I was optimising a no_std program for size, I was surprised how much stack shuffling was going on. Also iirc it had runs of load/store where it seemed like a loop might be better.

Shrinking the size of my library's Error struct seemed to help a lot - I wonder if it's because an error "E" smaller than the Result<T, E>'s "T" can be returned in-place, but a larger one needs a copy...?