I've noticed that people often write things like

all (x in X) {f x (doSomethingComplex Y)}

Instead of trying to convince people that they should always lift the complex computation on Y out of the loop, it's not so difficult to do it automatically.

This would need to be an optimization performed before lambda-lifiting. It would identify maximial non-immediate subexpressions in the lambda-body which don't contain the lambda-bound variables and lift them out.

Since the above is syntactic sugar for:

all \{x := f x (doSomethingComplex Y)} X

the transformation would optimize it to

all (let y := doSomethingComplex Y in \{x := f x y}) X

In the optimized version, doSomethingComplex is performed only once before passing the lambda-function to all.