hobel (do.delete@this.not.reply) on February 25, 2019 9:48 am wrote:
> Michael S (already5chosen.delete@this.yahoo.com) on February 25, 2019 1:52 am wrote:
> >
> > gcc is actually damn interesting case.
> > I have a strong feeling (not my field at all, so just feeling) that the same
> > deep learning techniques that in last couple of years produced excellent
> > results in go and chess* have big potential in compiler optimization.
> > Unfortunately, gcc code base is probably too convoluted to incorparate something that different quickly.
>
>
> But being deterministic is a nice property of a compiler, and not so important for a game
> playing software. Finding out why a network does not deliver the wanted result seems to be
> even less fun than debugging compilers, I have some doubts, but I see what you mean...

My reading of the suggestion isn't that you would use a neural network to do the optimizations, that would indeed be crazy, but rather it would suggest which transformation are most likely to improve performance. And those transformations would be the normal compiler ones that are currently chosen via heuristics and provably don't affect the program output.

Of course, all this might mean performance being harder to predict but that can be hard right now.