Fun fact: for simple text diffs, difftastic is now constrained by terminal output performance!
I halved instruction counts for diffing a 300KLOC text file against "" and I only went from 8.2 seconds to 6.5 seconds.
I really like instruction counts for measuring performance. It's not perfect, but it's a very stable metric, even if I'm listening to music whilst developing and benchmarking.
One thing I've learnt from writing a toy compiler is that this instruction usually means you have a bug:
add %al,(%rax)
This is the instruction encoding of all zeroes. It means you've jumped to the wrong place, or screwed up the size of an operand.