Loop-invariant code motion

[tmcdonell]

We don't have any form of loop-invariant code-motion implemented. At the scalar level we could assume that the backend compiler (CUDA / LLVM) will do that for us, but at the array level we are on our own (given the current approach of compiling each collective operation individually).

In the mean time, the following little hack will force a thing to be executed before moving on:

licm :: (Arrays a, Arrays b) => Acc a -> (Acc a -> Acc b) -> Acc b
licm x f  = (id >-> f) x