A decade or so ago I would have had a use for this information. I was working on the problem of distinguishing code from data in Windows executables -- which is not always easy because of the presence of computed jumps and multibyte instructions (even if you have a pretty good idea that you're looking at code, you need to figure out exactly which byte it starts on). I wrote a Bayesian analyzer that estimated the likelihood that a given byte was the beginning of an instruction, based on estimated instruction frequencies. It worked very well despite the fact that my frequency estimates were SWAGs [0]. I'm sure it would have worked even better with actual statistics.
[0] https://en.wikipedia.org/wiki/Scientific_Wild-Ass_Guess