Interesting, I've generally found that clang is better at optimising, especially when it comes to auto-vectorisation.