Isn’t the whole thrust of this thread that most normal algorithms see little to no speed up from things like avx, and therefore multiversioning those things that do makes more sense that compiling the whole OS for a newer set of cpu features?