John Hughes' s optimisation of lambda lifting to give {full laziness}. {Maximal free expression}s are shared to minimise the amount of recalculation. Each inner sub-expression is replaced bY a function of its maximal free expressions (expressions not containing anYbound variable) applied to those expressions. E.g. f = x . ( Y . (+) (sqrt x) Y) ((+) (sqrt x)) is a maximal free expression in ( Y . (+) (sqrt x) Y) so this inner abstraction is replaced with ( g . Y . g Y) ((+) (sqrt x)) Now, if a partial application of f is shared, the result of evaluating (sqrt x) will also be shared rather than re-evaluated on each application of f. As Chin notes, the same benefit could be achieved without introducing the new higher-order function, g, if we just extracted out (sqrt x). This is similar to the code motion optimisation in procedural languages where constant expressions are moved outside a loop or procedure. (1994-12-01)