Issue KEYWORD-ARGUMENT-NAME-PACKAGE Writeup

Version 6 conditionally passed X3J13/Jun87. Version 8 distributed in hardcopy

form X3J13/Nov87.


!

Issue:        KEYWORD-ARGUMENT-NAME-PACKAGE

References:   Lambda Expressions (CLtL pp60-64)

Category:     CLARIFICATION/CHANGE

Edit history: 20-Apr-87, Version 1 by Moon

              29-Apr-87, Version 2 by Pitman

              11-May-87, Version 3 by Moon

              29-May-87, Version 4 by Masinter 

               5-Jun-87, Version 5 by Masinter

              11-Jun-87, Version 6 by Masinter

              23-Oct-87, Version 7 by Masinter

               8-Nov-87, Version 8 by Moon


Problem Description:


CLtL says that only keyword symbols can be used as keyword-names in

&key parameter specifiers (page 60, "each -keyword- must be a

keyword symbol, such as :start.")


As Common Lisp is currently defined, if someone wants to define a

function that accepts keyword (rather than positional) arguments whose

keyword-names are symbols in packages other than the KEYWORD package,

they cannot use &KEY. Instead, they have to duplicate the &KEY mechanism

using &REST, GETF, and (if they want error checking of argument names)

DO.


Some applications (including the draft proposal for the Common Lisp

Object System (CLOS)) require this capability. [See Rationale below.]


Proposal (KEYWORD-ARGUMENT-NAME-PACKAGE:ANY)


Remove restrictions on the package of the keyword-names of keyword

parameters; allow any symbol. That is:


If, following an &KEY, a variable appears alone or in a (variable

default-value) pair, the behavior specified in CLtL is unchanged: a

keyword symbol with the same print name as the variable is created and

is used as the keyword-name when matching arguments to parameter

specifiers.  A keyword-name that is not a keyword symbol can be

specified with the ((-keyword- -var-) ...) syntax of parameter

specifier. The -keyword- can be any symbol, not just a keyword.


A future specification of Common Lisp could be written with revised

terminology that did not use the term "keyword" to refer to three

different things: symbols in the KEYWORD package, symbols beginning

with & that have special meaning in lambda-lists, and keyword-names

used to match function arguments to keyword parameter specifiers.

However, this proposal does not propose to change any terminology.


Test case:


(DEFUN RESULT (&KEY ((SECRET-KEYWORD SECRET) NIL) AMOUNT)

    (FORMAT NIL "You ~A $~D" (if SECRET "win" "lose") AMOUNT))


(RESULT :AMOUNT 100) => "You lose $100"

(RESULT :AMOUNT 100 'SECRET-KEYWORD T) => "You win $100"



Rationale:


The "rationale" box on p.62 of CLtL is an argument in favor of requiring

keyword-names to be symbols, and disallowing numbers, but does not

speak to the issue of whether or not those symbols should be further

restricted to be in the KEYWORD package.


The desire for keyword parameters whose keyword-names are not in the

KEYWORD package arises when the set of keyword-names accepted by a

function is the union of the sets of keyword-names accepted by several

other functions, rather than being enumerated in a single place.  In

this case, it becomes desirable to use packages to prevent accidental

name clashes among keyword-names of different functions.


One example of a Common Lisp application that requires this capability

is the draft proposal for an object-oriented programming standard

(CLOS).  It will have generic functions that accept arguments and pass

them on to one or more applicable methods, with each method defining its

own set of keyword-names that it is interested in.  If this proposal is

not adopted, either the keyword-names will be required to be keywords,

which will require the methods to have non-modular knowledge of each

other in order to avoid name clashes, or the methods will have to be

defined with an ad hoc mechanism that duplicates the essential

functionality of &key but removes the restriction.


A second example of a Common Lisp application that requires this

capability is private communication channels between functions.  Suppose

a public routine MAKE-FOO needs to accept arbitrary arguments from the

caller and passes those arguments along to an internal routine with

additional arguments of its own, and suppose that keyword parameters

are used to receive these arguments.


 (IN-PACKAGE 'FOOLAND)

 (DEFUN MAKE-FOO (&REST NAME-VALUE-PAIRS &KEY &ALLOW-OTHER-KEYS)

   (APPLY #'MAKE-FOO-INTERNAL 'EXPLICIT T NAME-VALUE-PAIRS))


This could be done without fear that the use of EXPLICIT T would

override some argument in NAME-VALUE-PAIRS, since the only way

that could happen is if someone had done (MAKE-FOO 'FOOLAND::EXPLICIT

NIL), or if the user was programming explicitly in the FOOLAND package,

either of which is an implicit admission of willingness to violate

FOOLAND's modularity.


Documentation Impact:


Some careful rewording of the existing language in CLtL is necessary in

the standard to avoid confusion between "keyword", indicating a symbol

in the KEYWORD package, and "keyword name", indicating a syntactic part

of a keyword parameter specifier.  It is likely that this is best served

by changing those instances of "keyword" to "named argument" when the

specification is discussing the indicator which introduces an actual

parameter in a call to a function defined with &KEY.


The wording which refers to named arguments as being introduced by

keyword symbols would change to simply refer to those arguments being

introduced by symbols. For example, in the middle of p.60, the sentence:

   ... each -keyword- must be a keyword symbol, such as :start.

 would become

   ... each named argument name must be a symbol.


The word "keyword" in the first complete sentence on p.62 would be

changed to "symbol" for similar reasons.


Extra wording would have to be added on p.60 to explain that by

convention keyword symbols are normally used as the names for named

arguments, and that all functions built into the Common Lisp language

follow that convention.


Examples would be useful. On p.64 the following examples might be added:


    ((lambda (a b &key ((:sea c)) d) (list a b c d)) 1 2 :sea 6)

    => (1 2 6 NIL)


    ((lambda (a b &key ((c c)) d) (list a b c d)) 1 2 'c 6)

    => (1 2 6 NIL)


Current Practice:


We do not currently know of an implementation that enforces the

restriction that this proposal seeks to remove.


Some implementations have bugs that prevent NIL from working as a

keyword argument name, but allow all non-NIL symbols. (One Symbolics

version that was checked had this bug.)


Cost to implementors:


Some implementors might have to rearrange their error checking slightly,

but it should be very easy.


Cost to users:


None--no existing programs will stop working.  


Benefits:


This will help with the object-oriented programming standard, among

other things.


Aesthetics:


The restriction of &key to only keyword symbols is arbitrary and

unnecessary.


There will probably be an argument about whether the restriction is more

aesthetic or less aesthetic than the freedom, but in either case the

aesthetic effect is slight.


In any case, users who do not want to use the extended functionality can

generally avoid it.


Discussion:


The cleanup committee generally supports this extension. 


Moon was under the impression that this proposal was actually adopted

around December 1985 (although no formal mechanism for adopting

proposals existed at that time), but may be mistaken.


If Common Lisp truly has a restriction that only keyword symbols can be

used as keyword names in calls to functions that take keyword arguments,

it will be more difficult to come up with an object-oriented programming

standard that fits within Common Lisp.


The cleanup committee considered, but did not adopt, a proposal to

exclude NIL as a legal indicator. It might catch some errors, but is

otherwise an odd restriction.