"Simon Marlow"
If a function is called, then the result has been demanded. There are no situations in which a function has been called but the caller will accept a thunk as the result without further evaluating it. The caller would definitely have to evaluate it, so every eval will need to loop until the value is in HNF.
Indeed. So what we are talking about is a peephole optimisation, where the callee can save some work for the system overall by evaluating its result to HNF, rather than leaving it to the caller to loop and test on every iteration.
So what is the advantage of returning a possibly-unevaluated result?
There is no operational advantage. The results are indistinguishable except in performance terms. My only, rather trivial, point is that the close visual similarity of f x = x and f (C x) = (C x) means that one can /explain/ both cases in identical terms, namely that the value bound by the variable x is potentially not-yet-evaluated, and that function f builds its result without looking at the value of x. And if you implement the language naively, it even works! Just slower than it might be.
And what about updates?
What do you mean? Regards, Malcolm