[Sesse__]

In many languages, you can't just optimize + b - b willy-nilly, as there could be side effects and non-obvious interactions abound. For instance, in JavaScript, where everything is a 64-bit double, a + b - b is definitely not the same as a, given large enough or small enough a or b. In LLVM for floats as well, certainly.

[fear91]

It's an example of how trivial some of this low hanging fruit is, the above is a concrete case that I have personally implemented (arithmetic of in-register 64/32-bit integers). You can get into semantics and restrictions, but I think the point I'm raising is clear.