This inserts dup and drop primitives according to a caller dup, callee drop policy. The value returned by a function should be passed back referenced (ownership transfers from the callee back to the caller).

The dup : a -> a primitive behaves like the identity function, evaluating and returning its first argument and increasing the reference count on the result. It is meant to be wrapped around function arguments.

The drop : a -> b -> a primitive evaluates and returns its first argument. It decrements the reference count to its second argument after it has evaluated its first argument. It is meant to be wrapped around function bodies that need to use and then de-reference their arguments.

Thus, something like

add a b = a + b

becomes

add a b =
  drop
    (drop
       ((dup a) + (dup b))
       b)
    a

Arguments a and b to the + primitive are duplicated and the result of + is duplicated internally, so add does not need to duplicate its result. Both arguments a and b are dropped.

Try running sslc --dump-ir-final on an example to see the inserted dup and drop constructs.

Our approach was inspired by Perceus https://www.microsoft.com/en-us/research/publication/perceus-garbage-free-reference-counting-with-reuse/