As I read this, if the compiler determines that (with the #define set) the file is effectively empty, it knows it can just check the #define next time rather than re-reading. Without this optimization, you get the same result slower. It doesn't depend on well-behaved anything to get a result just as correct as without the optimization.

With the single possible exception of a file that changes on disk between the multiple includes in the same build step. That is not a situation I have ever encountered and I am not at all sure it is supported and there aren't a bazillion other things that would break it in any given compiler chain (certainly at least line directives would be wrong, but that's arguably minor I guess).