more aggressive type propagation for local functions' parameter types

[Tobias C. Rittweiler]

The following example

  (defun foo (x fn)
    (declare (optimize speed))
    (declare (function fn))
    (labels ((recurse (x fn)
               (if (atom x)
                   (recurse (list x) fn)
                   (mapcar fn x))))
      (recurse x fn)))

will generate the following note

  ; note: unable to
  ; optimize
  ; because:
  ; optimize away possible call to FDEFINITION at runtime

PKhuong said on #lisp that is due to SBCL's conversatism regaring type
propagation: It assumes the top type and tries to refine that to more exact
type:

  <pkhuong> tcr: our propagation pass is conservative (primal, we keep
            refining known truths). For something like the above to
            work, we'd need to use another paradigm, with a dual
            iterative method. This is similar to the issue with naive
            dead variable analysis (primal induction), versus the
            right way (assume everything is dead and find
            counter-examples, co-induction).

[Nikodemus Siivola (nikodemus)]

  status confirmed
  tag optimization
  importance wishlist
  done