Re: [Sbcl-devel] Type inference problem?

[Nikodemus S. <nik...@ra...>]

2009/5/6 Rupert Swarbrick <rsw...@go...>:

> (defun testme (txt)
>  (cl-ppcre:register-groups-bind
>   (a) ("([0-9]+)" txt)
>   (parse-integer a)))

Expanding the REGISTER-GROUPS-BIND inline gives

(defun testme (txt)
  (let ((target-string951 txt))
  (multiple-value-bind
        (match-start952 match-end953 reg-starts954 reg-ends955)
      (cl-ppcre:scan "([0-9]+)" target-string951 :start (or nil 0) :end
                     (or nil (length target-string951)))
    (declare (ignore match-end953))
    (when match-start952
      (let* ((substr-fn957 (if nil #'cl-ppcre::nsubseq #'subseq))
             (a
              (let ((start-index956 (aref reg-starts954 0)))
                (IF START-INDEX956
                    (FUNCALL #'IDENTITY
                             (FUNCALL SUBSTR-FN957 TARGET-STRING951
                                      START-INDEX956 (AREF REG-ENDS955 0)))
                    NIL))))
        (parse-integer a))))))

Notice the upcased IF: the compiler cannot know that START-INDEX956 is
always true if MATCH-START952 is true, so it thinks there is potential
for NIL to be passes to PARSE-INTEGER. That's the warning. (I'm just
assuming that start-index exists when match-start is true -- I don't
know. If it does, the IF is not needed. If it does not, the IF belongs
there, and you should wrap your call to PARSE-INTEGER in a WHEN.)

DEBUG 2 makes type inference more stupid, so the compiler fails to see
the apparent problem. Hence no warning.

Cheers,

 -- Nikodemus