Compiler optimizations, CPU pipeline and branch prediction would be my guess, although I'm sure someone else can jump in with more detail. XOR is basically immune to optimized tricks.