I think they meant the other way around, that if you wanted to use it in C/C++, you'd have to either use assembly or some specific SSE construct otherwise the compiler wouldn't bother.

That probably was the case at one point, but I'd be surprised if it's still the case. Though maybe that's part of the reason why the Intel compiler can generate faster code. But I suspect it's more of a case of better optimization by people who have a better understanding of how it works under the hood, and maybe better utilization of newer instruction set extensions.

SSE has been around for a long time and is present in most (all?) x86 chips these days and I'd be very surprised if gcc and other popular compilers don't use it effectively today. Some of the other extensions might be different though.

I recently came across a rust book on how pointers aren't just ints, because of UB.

fn main() {
    a = &1
    b = &2
    a++
    if a == b {
        *a = 3
        print(b)
    }
}

This may either: not print anything, print 3 or print 2.

Depending on the compiler, since b isn't changed at all, it might optimize the print for print(2) instead of print(b). Even though everyone can agree that it should either not print anything or 3, but never 2.

A compiler making assumptions like that about undefined behaviour sounds just like a bug. Maybe the bug is in the spec rather than the compiler, but I can't think of any time it would be better to optimize that code out entirely because UB is detected rather than just throwing an error or warning and otherwise ignoring the edge cases where the behaviour might break. It sounds like the worst possible option exactly for the reasons listed in that blog.

I hope you are not argueing that using assembly is an improvement over using C in that regard....