Issue DECLARE-MACROS Writeup

Issue:        DECLARE-MACROS 

References:   Declaration Syntax (p154)

Category:     CHANGE

Edit history: 22-Oct-87, Version 1 by Pitman

               9-Nov-87, Version 2 by Masinter

               5-Feb-88, Version 3 by Pitman


Problem Description:


  It is permissible for a macro call to expand into a declaration and be

  recognized as such. This linguistic feature provides some useful

  flexibility, but has a number of disadvantages:


  * Operations on the executable portion of a body of code inside a 

    binding form (such as inserting an additional form) is a complicated

    operation. This is because one or more trial macro expansions must be

    done in order to pass over any declarations or documentation string

    and find the beginning of the body.


  * In order to find the end of the declarations, MACROEXPAND must be

    called until a non-macro form is seen or until a macro does not expand

    into a macro. In some interpreters which do macro expansion on the fly,

    this may cause inefficiency because macro expansion of the first form

    in a body must be done twice. In implementations where this is 

    optimized, the implementor may resent the fact that an optimization is

    needed in the first place.


  * Various code analysis tools have been shown to have been significantly

    slowed down by the need to expand macros in order to determine whether

    a binding is SPECIAL when analyzing a variable binding form. This is

    particularly serious when macro invocations are deeply nested; the

    number of macro expansions can be exponential in the depth of nesting

    unless a macro-expansion caching mechanism is added. 


  * User macros must be very careful about finding declarations in a body

    of code by doing proper macro expansion. Often, however, naive users

    don't realize this and so unknowingly write buggy code. This problem can

    be (and is) defined away as simply a programmer error, but this is a

    place where we could fairly straightforwardly redefine the language to

    better accommodate what has been shown to be a common expectation of the

    naive user.


Proposal (DECLARE-MACROS:FLUSH):


   Under this proposal, it would not be "permissible for a macro call to

   expand into a declaration and be recognized as such, provided that the

   macro call appears where a declaration may legitimately appear." (CLtL

   p. 154). Macros could not legitimately expand into declarations; the only

   valid declarations would be a list whose CAR is the symbol DECLARE.


   It would still be possible for a macro call to expand into a PROCLAIM

   form, however.


Rationale:


  The ability to have a macro form expand into a declaration has been shown

  to be useful in some situations.  More often, however, the presence of

  this feature has been seen to cause problems elsewhere in the language.

  Ultimately, its benefits have not outweighed its costs.


Current Practice:


  Most or all implementations support the old behavior even though few

  user programs probably need it.


  Some user macros are careful about finding declarations in a body of code

  by doing proper macro expansion, but some probably cheat and look just

  for explicit uses of DECLARE. The cheat probably works most of the time.


Cost to implementors:


  The nature of this change is such that implementations which did not

  choose to change would simply be supporting an implementation-dependent

  extension (except for some `minor' worry about destructive modification

  due to macro expanding while *MACROEXPAND-HOOK* is set to something

  which implemented displacing macros).


  In any case, there might be several places in which the interpreter,

  compiler, and system macros had knowledge about doing macro expansion

  before declaration processing. The change is not trivial, but most of

  its complexity is likely to be in finding the places which need change

  and not in making the actual change.


Cost to users:


  Most users probably do not write macros which expand into DECLARE forms

  so most users are probably not affected.


  Users who do exploit this feature probably know that they do. In any

  case, compilers could be made to detect cases where this feature is

  being exploited and warn about it.


  Franz and Gold Hill are notable exceptions to the claim that users may

  not want this. Both claim to make a reasonable amount of use of macros

  which expand into different SPEED and SAFETY declarations, usually

  dependent on a global switch.


  Rewrites must be devised on a case-by-case basis. A common sort of

  rewrite would take the form:


   Old code:  (DEFMACRO SPEEDY ()

		`(DECLARE (OPTIMIZE (SPEED 3) (SAFETY 0))))

   	      (LET (..bindings..) (SPEEDY) ..body..)


   New code:  (DEFMACRO SPEEDY-LET (BVL &BODY FORMS)

		`(LET ,BVL

		   (DECLARE (OPTIMIZE (SPEED 3) (SAFETY 0)))

		   ,@FORMS))

	      (SPEEDY-LET (..bindings..) ..body..)


  Another tactic would be:


   Old code: (EVAL-WHEN (EVAL COMPILE LOAD)

	       (DEFVAR *SPEEDY* NIL))

	     (DEFMACRO USE-STANDARD-SPEED-AND-SAFETY ()

	       (IF *SPEEDY*

		   `(DECLARE (OPTIMIZE (SPEED 3) (SAFETY 0)))

		   `(DECLARE (OPTIMIZE (SPEED 0) (SAFETY 3)))))

	     (DEFUN FOO ()

	       (USE-STANDARD-SPEED-AND-SAFETY)

	       ...)

   New code: (EVAL-WHEN (EVAL COMPILE LOAD)

	       (DEFVAR *STANDARD-SPEED-AND-SAFETY* '((SPEED 0) (SAFETY 3))))

	     (DEFUN FOO ()

	       (DECLARE (OPTIMIZE #.*STANDARD-SPEED-AND-SAFETY*))

	       ...)


  Still a third tactic would be to actually shadow DEFUN, LET, etc. with

  variants that process macro expansions and then to build code in a

  package that used the revised DEFUN, LET, etc. eg,


    (DEFUN PARSE-BODY (BODY ENV)

      (LET ((DECLS '())

	    (DOC   '()))

	(DO () ((NULL (CDR BODY)))

	  (LET ((EXP (MACROEXPAND (POP BODY) ENV)))

	    (COND ((AND (STRINGP EXP) BODY)

		   (UNLESS (NULL DOC)

		     (WARN "More than one documentation string was seen."))

		   (PUSH EXP DOC))

		  ((AND (NOT (ATOM EXP)) (EQ (CAR EXP) 'DECLARE))

		   (PUSH EXP DECLS))

		  (T

		   (PUSH EXP BODY)

		   (RETURN NIL)))))

	(VALUES BODY (NREVERSE DECLS) (NREVERSE DOC))))


   (DEFMACRO MY:DEFUN (NAME LAMBDA-LIST &BODY DECLS-DOC-AND-FORMS

		       &ENVIRONMENT ENV)

     (MULTIPLE-VALUE-BIND (FORMS DECLS DOC-STRINGS)

	 (PARSE-BODY DECLS-DOC-AND-FORMS ENV)

       `(DEFUN ,NAME ,BVL ,@DECLS ,@DOC-STRINGS ,@FORMS)))


   (DEFMACRO MY:LET (BINDINGS &BODY DECLS-DOC-AND-FORMS

		     &ENVIRONMENT ENV)

     (MULTIPLE-VALUE-BIND (FORMS DECLS DOC-STRINGS)

	 (PARSE-BODY DECLS-DOC-AND-FORMS ENV)

       `(LET ,BINDINGS ,@FORMS)))


   ...etc.


  LAMBDA cannot be done this way, of course, since it is not a macro (or

  even a special form). Support for expanding declarations in a LAMBDA

  would have to be provided either by using implementation-specific support

  (such as Zetalisp's ``lambda macros'') or by a workaround such as a

  macro like:


   (DEFMACRO LAMBDA (LAMBDA-LIST &BODY DECLS-DOC-AND-FORMS

		     &ENVIRONMENT ENV)

     (MULTIPLE-VALUE-BIND (FORMS DECLS DOC-STRINGS)

	 (PARSE-BODY DECLS-DOC-AND-FORMS ENV)

       `#'(LAMBDA ,BINDINGS ,@FORMS)))


  Note that unlike the other examples, LAMBDA need not be (and in fact,

  may not be) in the "MY" package in order for this to work since the

  FUNCTION special form will generally only recognize LISP:LAMBDA.


Benefits:


  The efficiency of some tools may be improved.

  User macros which must do minor surgery on bodies of code will be

  easier to write.


Aesthetics:


  This change simplifies the semantics of the language slightly and

  probably tends to better support the assumptions of naive programmers

  writing macros.


  In some cases involving complicated extensions to declarations, it

  may be slightly harder to express such extensions in a modular way.

  Experience thus far has shown such cases to be rare, however.


Discussion:


  Symbolics took an in-house poll of people who take advantage of the

  feature and it was generally agreed that in most cases where this

  feature is used at all, that it would be just as easy to work around

  using the sample rewrite techniques cited above.


  Moon `credits' Pitman for (against some opposition) pushing this

  `feature' down everyone's throats in the original CL design process.

  Pitman admits this was an expensive mistake. Moon and Pitman support

  this change as an important simplification to the language.


  The cleanup committee unanimously endorsed this proposal.


  In discussion at the Nov-87 X3J13 meeting, Franz and Gold Hill

  mentioned that they use this feature a lot and were not entirely

  happy about its going away.