%coerce-callable-to-fun lack of error checking sometimes

[Douglas Katzman]

SYMBOL-FUNCTION (and fdefinition) is specified to return an arbitrary implementation-chosen value for macros and special operators.
Historically SBCL (and CMUCL) return a function which when called complains that it's not really a function. A consequence of that choice is that an error is almost never trapped if no funcall occurs, such as in:

* (mapcar (symbol-function 'throw) '())
NIL
* (mapcar (symbol-function 'and) '())
NIL

Most other lisps return a magic value from symbol-function. We could do that also.
The desired outcome is that there is no extra cost to rejecting macros wherever the compiler has optimized a higher-order function to using %coerce-callable eagerly (i.e. before trying an actual funcall). We definitely don't want %coerce-callable to check some arbitrary property of the specified function to see if it's really supposed to be called. (We do that if the arg to %coerce-to-callable was a symbol, but in that event it already costs something to lookup the function)

It may work to have macros and symbols look internally like un-fboundp symbols, and also impart behavior to symbol-function, fdefinition, fboundp that don't signal an error. This would fix the %coerce-callable situation efficiently.
Granted it's only a problem for sequence traversal operations on empty sequences. Maybe there are some other situations, but this caught my attention.

[Stas Boukarev (stassats)]

It's kinda similar to (mapcar 'car nil nil) not signaling an error because it's never applied. The standard says it returns an implementation defined value, and returning a function is such a value. Then it's valid to pass functions to mapcar.

So, it's not a standard compliance nor a safety issue, just catching bad values early.

[Douglas Katzman (dougk)]

Well, I think there may be a minor compliance issue if we assume an unknown name refers to a function. Sharp quote can not take a macro but we treat #' and QUOTE the same.
Let's agree that FOO1 gets a free pass here; it's harder to justify that (FOO2 NIL) is legal.

(defun foo1 (x) (mapcar (symbol-function 'something) x)) ; dubious but OK
(defun foo2 (x) (mapcar #'something x)) ; "technically" should barf
(defmacro something () ) ; only after misusing #'something
;; FOO2 _assumed_ that #' is ok but ZOOK knows that it's not ok
(defun zook () #'something) ; and this rightly fails to compile

[Stas Boukarev (stassats)]

It says "SOMETHING is being redefined as a macro when it was previously assumed to be a function."