likely and unlikely exist in different forms as compiler extensions for a while. There is code where they indeed help compilers to produce slightly better code for the hot path. Good use typically involves lots of measurement as developers are often wrong about the impact and about what the real flow is.

What kind of things can the compiler improve? For instance it can arrange the code in a way that the instructions of the hot path are directly behind each other, whereas the unlikely case jumps further away. Or it can arrange code a bit to help the branch predictor.