functions with both inline expansions and compiler optimizations

[Nikodemus Siivola]

(do-all-symbols (s)
  (flet ((scan-name (s)
           (let* ((inline-exp (sb-int:info :function :inline-expansion-designator s))
                  (inlinep (sb-int:info :function :inlinep s))
                  (info (sb-int:info :function :info s))
                  (transforms (when info (and (sb-c::fun-info-transforms info) t)))
                  (templates (when info (and (sb-c::fun-info-templates info) t)))
                  #+nil
                  (pred (when info (and (sb-c::fun-info-predicate-type info)) t)))
             (when (and inline-exp
                        (or transforms templates #+nil pred))
               (print (list s inlinep
                            :transforms transforms
                            :templates templates))))))
    (scan-name s)
    (scan-name `(setf ,s))))

finds the following

(%POW :INLINE :TRANSFORMS NIL :TEMPLATES T)
(FLOOR :MAYBE-INLINE :TRANSFORMS T :TEMPLATES NIL)
(MOD :INLINE :TRANSFORMS T :TEMPLATES NIL)
(NTHCDR :MAYBE-INLINE :TRANSFORMS T :TEMPLATES NIL)
(CEILING :MAYBE-INLINE :TRANSFORMS T :TEMPLATES NIL)
(ARRAY-HAS-FILL-POINTER-P :INLINE :TRANSFORMS T :TEMPLATES NIL)
(KEYWORDP :MAYBE-INLINE :TRANSFORMS T :TEMPLATES NIL)
(REM :INLINE :TRANSFORMS T :TEMPLATES NIL)

The problem is that inlining defeats the compiler transforms, which could probably do a better job.

For some just nuking the inline expansion is probably TRT, but for some we may want to keep a %FOO version available for inlining if all transforms fails and no templates apply. Who knows, maybe it's the compiler magic that should go in some cases.