An issue for the SC: horrible documentation quality of GCC

[Richard Kenner]

    good luck. If you want them be really good, they must all work globally.
    They must alst take into account what any other later optimization could
    do. This is not feasible.

It isn't feasible to do it perfectly, of course not, but if we
properly specify what later optimizations do, it's not that unreasonable
to get something that makes the right decision the vast majority of the time.

    Why it should not? You must gather the data on that you base your
    decison, which definitely is not fast if you want to do it everywhere.

Usually that data is fairly cheap to get, such as a cost function.