Reply To: Pipelining

Unbounded ILP in loops vs. average iterations in loops?

In several of the talks, Ivan has said words to the effect that,
“loops have unbounded ILP.”

Unbounded ILP would seemingly be true only for unbounded (e.g. while 1 {}) loops; for all others (whether simple counting loops or some form of while loop), the ILP is roughly proportional to the loop’s number of iterations — and thus bounded. Have I missed something/being too pedantic?

I’ll freely concede that loops, including loops repeated enough to benefit from pipelining, are frequent enough in general-purpose code that an architecture that can pipeline virtually all loops should be a genuine boon. So infinite or not ILP, I’m really looking forward to more info on epilogs and LLVM progress for the Mill.