Experimental flag -fdicts-cheap
This experimental flag, -fdicts-cheap, makes a let-binding that bind a value of dictionary type look cheap. That in turn leads to more eta expansion. Instead of f = /\a. \(d1:Ord a). let d2:Ord [a] = dfOrd a d1 in \(x:a). <stuff> which has arity 1, you get f = /\a. \(d1:Ord a). \(x:a). let d2:Ord [a] = dfOrd a d1 in <stuff> Now f has arity 2. This can cretainly waste dictionary-construction work, if f is partially applied to its dictionary argument. However it has knock-on effects. Because f has arity 2, we won't float (f Int d) out of \x. h (f Int d) Floating f out of this lambda makes it impossible for an h/f fusion rule to fire; and this unexpected loss of RULE application was the immediate reason for implementing this flag. (Roman Leshchinskiy came across this when working on array fusion.) I've implemented the change only in CoreUtils.arityType, which only affects eta expansion. I thought of putting the change in exprIsCheap, which is a more systematic place (the former calls the latter) but a) I wanted this under flag control, and the flags are not readily available to all callers of exprIsCheap b) I'm not 100% convinced that this change is a good idea, so it's reasonable to do the narrowest change that solves the immediate problem.