EIEIO and CL

EIEIO and CL

[Chong Yidong]

Currently, the EIEIO library requires CL at runtime.  This is mainly due
to needing `deftype' and `typep' (EIEIO implements a subset of the CLOS
spec, which demand their existence).  If possible, it would be nice to
reduce this to a compile-time dependency, so that loading Semantic or
EDE does not load the CL library.

The `typep' requirement is solvable; the `deftype' requirement is the
more interesting.  Currently, eieio-defclass contains this code:

   ;; When using typep, (typep OBJ 'myclass) returns t for objects which
   ;; are subclasses of myclass.  For our predicates, however, it is
   ;; important for EIEIO to be backwards compatible, where
   ;; myobject-p, and myobject-child-p are different.
   ;; "cl" uses this technique to specify symbols with specific typep
   ;; test, so we can let typep have the CLOS documented behavior
   ;; while keeping our above predicate clean.
   (eval `(deftype ,cname ()
            '(satisfies
              ,(intern (concat (symbol-name cname) "-child-p")))))

The problem here is that the `cname' argument to the deftype CL macro
is, in general, determined at run-time.

Another possibility is to change eieio-defclass so that it defers this
deftype form till CL is loaded, by putting it on cl-macs-load-hook.
This assumes that the deftype is only useful if some later code calls
typep or typecase; if this assumption is wrong, that could lead to
subtle bugs.

Or, there may be a way to change `deftype' (or to provide a different CL
macro) that can do the right thing at compile-time.

Any suggestions?

Re: EIEIO and CL

[Eric M. Ludlam]

On Fri, 2009-10-09 at 12:51 -0400, Chong Yidong wrote:
> Currently, the EIEIO library requires CL at runtime.  This is mainly due
> to needing `deftype' and `typep' (EIEIO implements a subset of the CLOS
> spec, which demand their existence).  If possible, it would be nice to
> reduce this to a compile-time dependency, so that loading Semantic or
> EDE does not load the CL library.
> 
> The `typep' requirement is solvable; the `deftype' requirement is the
> more interesting.  Currently, eieio-defclass contains this code:
> 
>    ;; When using typep, (typep OBJ 'myclass) returns t for objects which
>    ;; are subclasses of myclass.  For our predicates, however, it is
>    ;; important for EIEIO to be backwards compatible, where
>    ;; myobject-p, and myobject-child-p are different.
>    ;; "cl" uses this technique to specify symbols with specific typep
>    ;; test, so we can let typep have the CLOS documented behavior
>    ;; while keeping our above predicate clean.
>    (eval `(deftype ,cname ()
>             '(satisfies
>               ,(intern (concat (symbol-name cname) "-child-p")))))
> 
> The problem here is that the `cname' argument to the deftype CL macro
> is, in general, determined at run-time.
> 
> Another possibility is to change eieio-defclass so that it defers this
> deftype form till CL is loaded, by putting it on cl-macs-load-hook.
> This assumes that the deftype is only useful if some later code calls
> typep or typecase; if this assumption is wrong, that could lead to
> subtle bugs.
> 
> Or, there may be a way to change `deftype' (or to provide a different CL
> macro) that can do the right thing at compile-time.
> 
> Any suggestions?

Hi,

The defclass operation in EIEIO is a macro that calls out to a helper
function.  This is because I have found that I'm just not that good at
writing macros.  I would guess that if the entirety of eieio-defclass
were turned into a macro, then the whole eval thing you quoted above
would become a compile time requirement only.  Would it then be a
requirement during compilation of code that uses defclass, as opposed to
a compile time requirement of eieio.el?

Alternately, it would be cool if typep, or something similar was a part
of Emacs core.  Wouldn't the widget/custom setup like that?

Eric

Re: EIEIO and CL

[Chong Yidong]

"Eric M. Ludlam" <eric@siege-engine.com> writes:

> The defclass operation in EIEIO is a macro that calls out to a helper
> function.  This is because I have found that I'm just not that good at
> writing macros.  I would guess that if the entirety of eieio-defclass
> were turned into a macro, then the whole eval thing you quoted above
> would become a compile time requirement only.

Yes, that would probably work.

> Alternately, it would be cool if typep, or something similar was a part
> of Emacs core.  Wouldn't the widget/custom setup like that?

It's feasible to move typep into core, but deftype involves a layer of
CL-implementation gunk that would need to be cleaned up first; not
feasible for 23.2, I think.

RE: EIEIO and CL

[S+*n_Pe*rm*n]

Doesn't this little bit at the top of cl-macs get you most of the way?

,---- :FILE "/emacs/lisp/emacs-lisp/cl-macs.el"
| ;; This kludge allows macros which use cl-transform-function-property
| ;; to be called at compile-time.
|
| (require
|  (progn
|    (or (fboundp 'cl-transform-function-property)
|        (defalias 'cl-transform-function-property
| 	 (function (lambda (n p f)
| 		     (list 'put (list 'quote n) (list 'quote p)
| 			   (list 'function (cons 'lambda f)))))))
|    (car (or features (setq features (list 'cl-kludge))))))
`----

s_P