Why can't CLISP call certain functions with uninterned names?

Question

I've written an ad hoc parser generator that creates code to convert an old and little known 7-bit character set into unicode. The call to the parser generator expands into a bunch of defuns enclosed in a progn, which then get compiled. I only want to expose one of the generated defuns--the top-level one--to the rest of the system; all the others are internal to the parser and only get called from within the dynamic scope of the top-level one. Therefore, the other defuns generated have uninterned names (created with gensym). This strategy works fine with SBCL, but I recently tested it for the first time with CLISP, and I get errors like:

*** - FUNCALL: undefined function #:G16985

It seems that CLISP can't handle functions with uninterned names. (Interestingly enough, the system compiled without a problem.) EDIT: It seems that it can handle functions with uninterned names in most cases. See the answer by Rörd below.

My questions is: Is this a problem with CLISP, or is it a limitation of Common Lisp that certain implementations (e.g. SBCL) happen to overcome?

EDIT:

For example, the macro expansion of the top-level generated function (called parse) has an expression like this:

(PRINC (#:G75735 #:G75731 #:G75733 #:G75734) #:G75732)

Evaluating this expression (by calling parse) causes an error like the one above, even though the function is definitely defined within the very same macro expansion:

(DEFUN #:G75735 (#:G75742 #:G75743 #:G75744) (DECLARE (OPTIMIZE (DEBUG 2)))
 (DECLARE (LEXER #:G75742) (CONS #:G75743 #:G75744))
 (MULTIPLE-VALUE-BIND (#:G75745 #:G75746) (POP-TOKEN #:G75742)
 ...

The two instances of #:G75735 are definitely the same symbol--not two different symbols with the same name. As I said, this works with SBCL, but not with CLISP.

EDIT:

SO user Joshua Taylor has pointed out that this is due to a long standing CLISP bug.

Answer 1

You don't show one of the lines that give you the error, so I can only guess, but the only thing that could cause this problem as far as I can see is that you are referring to the name of the symbol instead of the symbol itself when trying to call it.

If you were referring to the symbol itself, all your lisp implementation would have to do is lookup that symbol's symbol-function. Whether it's interned or not couldn't possibly matter.

May I ask why you haven't considered another way to hide the functions, i.e. a labels statement or defining the functions within a new package that exports only the one external function?

EDIT: The following example is copied literally from an interaction with the CLISP prompt.

As you can see, calling the function named by a gensym is working as expected.

[1]> (defmacro test ()
(let ((name (gensym)))
`(progn
(defun ,name () (format t "Hello!"))
(,name))))
TEST
[2]> (test)
Hello!
NIL

Maybe your code that's trying to call the function gets evaluated before the defun? If there's any code in the macro expansion besides the various defuns, it may be implementation-dependent what gets evaluated first, and so the behaviour of SBCL and CLISP may differ without any of them violating the standard.

EDIT 2: Some further investigation shows that CLISP's behaviour varies depending upon whether the code is interpreted directly or whether it's first compiled and then interpreted. You can see the difference by either directly loading a Lisp file in CLISP or by first calling compile-file on it and then loading the FASL.

You can see what's going on by looking at the first restart that CLISP offers. It says something like "Input a value to be used instead of (FDEFINITION '#:G3219)." So for compiled code, CLISP quotes the symbol and refers to it by name.

It seems though that this behaviour is standard-conforming. The following definition can be found in the HyperSpec:

function designator n. a designator for a function; that is, an object that denotes a function and that is one of: a symbol (denoting the function named by that symbol in the global environment), or a function (denoting itself). The consequences are undefined if a symbol is used as a function designator but it does not have a global definition as a function, or it has a global definition as a macro or a special form. See also extended function designator.

I think an uninterned symbol matches the "a symbol is used as a function designator but it does not have a global definition as a function" case for unspecified consequences.

EDIT 3: (I can agree that I'm not sure whether CLISP's behaviour is a bug or not. Someone more experienced with details of the standard's terminology should judge this. It comes down to whether the function cell of an uninterned symbol - i.e. a symbol that cannot be referred to by name, only by having a direct hold on the symbol object - would be considered a "global definition" or not)

Anyway, here's an example solution that solves the problem in CLISP by interning the symbols in a throwaway package, avoiding the matter of uninterned symbols:

(defmacro test ()
  (let* ((pkg (make-package (gensym)))
         (name (intern (symbol-name (gensym)) pkg)))
    `(progn
       (defun ,name () (format t "Hello!"))
       (,name))))

(test)

EDIT 4: As Joshua Taylor notes in a comment to the question, this seems to be a case of the (10 year old) CLISP bug #180.

I've tested both workarounds suggested in that bug report and found that replacing the progn with locally actually doesn't help, but replacing it with let () does.

Answer 2

You can most certainly define functions whose names are uninterned symbols. For instance:

CL-USER> (defun #:foo (x)
           (list x))
#:FOO
CL-USER> (defparameter *name-of-function* *)
*NAME-OF-FUNCTION*
CL-USER> *name-of-function*
#:FOO
CL-USER> (funcall *name-of-function* 3)
(3)

However, the sharpsign colon syntax introduces a new symbol each time such a form is read read:

#: introduces an uninterned symbol whose name is symbol-name. Every time this syntax is encountered, a distinct uninterned symbol is created. The symbol-name must have the syntax of a symbol with no package prefix.

This means that even though something like

CL-USER> (list '#:foo '#:foo)
;=> (#:FOO #:FOO) 

shows the same printed representation, you actually have two different symbols, as the following demonstrates:

CL-USER> (eq '#:foo '#:foo)
NIL

This means that if you try to call such a function by typing #: and then the name of the symbol naming the function, you're going to have trouble:

CL-USER> (#:foo 3)
; undefined function #:foo error

So, while you can call the function using something like the first example I gave, you can't do this last one. This can be kind of confusing, because the printed representation makes it look like this is what's happening. For instance, you could write such a factorial function like this:

(defun #1=#:fact (n &optional (acc 1))
  (if (zerop n) acc
      (#1# (1- n) (* acc n))))

using the special reader notation #1=#:fact and #1# to later refer to the same symbol. However, look what happens when you print that same form:

CL-USER> (pprint '(defun #1=#:fact (n &optional (acc 1))
                    (if (zerop n) acc
                        (#1# (1- n) (* acc n)))))

(DEFUN #:FACT (N &OPTIONAL (ACC 1))
  (IF (ZEROP N)
      ACC
      (#:FACT (1- N) (* ACC N))))

If you take that printed output, and try to copy and paste it as a definition, the reader creates two symbols named "FACT" when it comes to the two occurrences of #:FACT, and the function won't work (and you might even get undefined function warnings):

CL-USER> (DEFUN #:FACT (N &OPTIONAL (ACC 1))
           (IF (ZEROP N)
               ACC
               (#:FACT (1- N) (* ACC N))))

; in: DEFUN #:FACT
;     (#:FACT (1- N) (* ACC N))
; 
; caught STYLE-WARNING:
;   undefined function: #:FACT
; 
; compilation unit finished
;   Undefined function:
;     #:FACT
;   caught 1 STYLE-WARNING condition

Answer 3

I hope I get the issue right. For me it works in CLISP.

I tried it like this: using a macro for creating a function with a GENSYM-ed name.

(defmacro test ()  
  (let ((name (gensym)))  
    `(progn  
       (defun ,name (x) (* x x))  
       ',name)))

Now I can get the name (setf x (test)) and call it (funcall x 2).

Answer 4

Yes, it is perfectly fine defining functions that have names that are unintenred symbols. The problem is that you cannot then call them "by name", since you can't fetch the uninterned symbol by name (that is what "uninterned" means, essentially).

You would need to store the uninterned symbol in some sort of data structure, to then be able to fetch the symbol. Alternatively, store the defined function in some sort of data structure.

Answer 5

Surprisingly, CLISP bug 180 isn't actually an ANSI CL conformance bug. Not only that, but evidently, ANSI Common Lisp is itself so broken in this regard that even the progn based workaround is a courtesy of the implementation.

Common Lisp is a language intended for compilation, and compilation produces issues regarding the identity of objects which are placed into compiled files and later loaded ("externalized" objects). ANSI Common Lisp requires that literal objects reproduced from compiled files are only similar to the original objects. (CLHS 3.2.4 Literal Objects in Compiled Files).

Firstly, according to the definition similarity (3.2.4.2.2 Definition of Similarity), the rules for uninterned symbols is that similarity is name based. If we compile code with a literal that contains an uninterned symbol, then when we load the compiled file, we get a symbol which is similar and not (necessarily) the same object: a symbol which has the same name.

What if the same uninterned symbol is inserted into two different top-level forms which are then compiled as a file? When the file is loaded, are those two similar to each other at least? No, there is no such requirement.

But it gets worse: there is also no requirement that two occurrences of the same uninterned symbol in the same form will be externalized in such a way that their relative identity is preserved: that the re-loaded version of that object will have the same symbol object in all the places where the original was. In fact, the definition of similarity contains no provision for preserving the circular structure and substructure sharing. If we have a literal like '#1=(a b . #1#), as a literal in a compiled file, there appears to be no requirement that this be reproduced as a circular object with the same graph structure as the original (a graph isomorphism). The similarity rule for conses is given as naive recursion: two conses are similar if their respective cars and cdrs are similar. (The rule can't even be evaluated for circular objects; it doesn't terminate).

That the above works is because of implementations going beyond what is required in the spec; they are providing an extension consistent with (3.2.4.3 Extensions to Similarity Rules).

Thus, purely according to ANSI CL, we cannot expect to use macros with gensyms in compiled files, at least in some ways. The expectation expressed in code like the following runs afoul of the spec:

(defmacro foo (arg)
   (let ((g (gensym))
         (literal '(blah ,g ,g ,arg)))
      ...))

(defun bar ()
  (foo 42))

The bar function contains a literal with two insertions of a gensym, which according to the similarity rules for conses and symbols need not reproduce as a list containing two occurrences of the same object in the second and third positions.

If the above works as expected, it's due to "extensions to the similarity rules".

So the answer to the "Why can't CLISP ..." question is that although CLISP does provide an extension for similarity which preserves the graph structure of literal forms, it doesn't do it across the entire compiled file, only within individual top level items within that file. (It uses *print-circle* to emit the individual items.) The bug is that CLISP doesn't conform to the best possible behavior users can imagine, or at least to a better behavior exhibited by other implementations.