Issue TYPE-OF-UNDERCONSTRAINED Writeup

Status:	passed, Jun 89 X3J13


Forum:	Cleanup

Issue:      TYPE-OF-UNDERCONSTRAINED


References: TYPE-OF (CLtL)

            88-002R (Class-of)

            88-005, p 43f, Condition system types

Related issues: SUBTYPEP-TOO-VAGUE

                DATA-TYPE-HIERARCHY-UNDERSPECIFIED

                FUNCTION-TYPE

                ARRAY-TYPE-ELEMENT-TYPE-SEMANTICS

                COERCE-INCOMPLETE

			


Category:       CLARIFICATION/CHANGE


Edit history:   Version 1,  1-Dec-88, Masinter

		    Version 2,  9-Dec-88, Masinter

		    Version 3, 12-Dec-88, Masinter (fix "egregious bug")

		Version 4, 3-14-89 Steele

		Version 5, 16-Mar-89, Masinter (add other amendments)

		Version 6, 22-Jun-89, Masinter (fix as per Moon)


Problem description:


The specification of TYPE-OF in CLtL is so weak as to leave it 

nearly useless, if any implementor actually took the specification

seriously.In particular, there are almost no constraints on the

value that TYPE-OF might return for any of the built in

types. The only constraint placed is that for objects created by the

constructor function of DEFSTRUCT, the result of TYPE-OF is the name of the

defstruct.


This means that implementations could return T for all other objects.


Proposal (TYPE-OF-UNDERCONSTRAINED:ADD-CONSTRAINTS): 


Specify that the result of TYPE-OF satisfies the following:


(a) For any object for which (typep object built-in-type) is true when

built-in-type is one of 


INTEGER, RATIO, FLOAT, COMPLEX, NUMBER, NULL, SYMBOL, 

STRING, BIT-VECTOR, VECTOR, ARRAY, RANDOM-STATE, SEQUENCE,CONS, 

STREAM, PACKAGE, CHARACTER, FUNCTION, READTABLE, HASH-TABLE, PATHNAME,

CONDITION, RESTART,

RATIONAL, SHORT-FLOAT, LONG-FLOAT, SINGLE-FLOAT, DOUBLE-FLOAT


the result of TYPE-OF will be a type specifier for which

(subtypep (type-of object) built-in-type) returns T T, i.e., either the

build-in-type or a subtype of it (a subtype that the 

implementation's SUBTYPEP can recognize.)



(b)  For all objects, (TYPEP object (TYPE-OF object)) will be true.

(this implies that it will not be NIL, since no

object is of type NIL.)


(c) will not be T, or use SATISFIES, AND, OR, NOT or VALUES.


(d) The type returned by TYPE-OF is always a subtype of the class

returned by CLASS-OF, and is a subtype that the implementation's

SUBTYPEP can recognize.


(e) For objects of metaclass STRUCTURE-CLASS or STANDARD-CLASS,

TYPE-OF returns the proper name of the class returned by CLASS-OF

if it has a proper name, and otherwise returns the class itself.


In particular, for objects created by the "constructor" function

of a structure defined with DEFSTRUCT without a :TYPE option,

TYPE-OF will return the structure name.


This proposal is intended to be consistent with 88-002R, 

and not to conflict with any of the definitions in that document.


Examples:


It is legal for (TYPE-OF "abc") to return (SIMPLE-STRING 3), or just

STRING. It is legal for (TYPE-OF 112312) to return SI:MEDIUM-SIZE-FIXNUM,

as long as (SUBTYPEP 'SI:MEDIUM-SIZE-FIXNUM 'INTEGER).


 Given:


   (defvar *foo* (make-array 5 :element-type t))

   (class-name (class-of *foo*))  =>  SIMPLE-VECTOR

  It is legal for

   (type-of *foo*)                =>  (SIMPLE-VECTOR 5)



Rationale:


The original specification for "TYPE-OF" was written to allow

implementation freedom, but the wording is in fact more vague than was

intended. 


Current practice:


Most Common Lisp implementations seem to implement the proposal, at least

for the types we've tried them on.


Cost to Implementors:


Even if there are implementations that do not currently implement the

proposal, the required changes for them are likely to be minimal. 


Cost to Users:


Apparently none.


Cost of non-adoption:


TYPE-OF would remain potentially inconsistent.


Performance impact:


Likely none.


Benefits:


Bring the specification of TYPE-OF in like with its common

implementation, while requiring it to be generally useful.


Esthetics:


Little effect.


Discussion:


Originally, TYPE-OF was concieved as being useful for information display.

CLOS specified CLASS-OF far more precisely, in that it must return exactly

the class of an object. Unless TYPE-OF is removed from the language, it

seems reasonable to require it to be at least as precise as CLASS-OF.


The accepted CLOS specification does not define CLASS-OF

in terms of TYPE-OF, but an earlier draft did. 


This proposal presumes the (passed) proposals

DATA-TYPES-HIERARCHY-UNDERSPECIFIED:DISJOINT, FUNCTION-TYPE, and

accomodates the Condition System (version 18) and CLOS.


If ARRAY-TYPE-ELEMENT-TYPE-SEMANTICS does not pass, 

and this proposal does pass, it may be that the only way an 

implementation can satisfy (TYPEP array (TYPE-OF array))

would be to have (TYPE-OF array) return VECTOR or ARRAY

as appropriate, i.e., with no element type qualification.


It is possible to constrain TYPE-OF further, e.g., to eliminate

the possibility that it might return a non-portable 

implementation-specific value. For example, 

(TYPE-OF 112312) should not return SI:MEDIUM-SIZE-FIXNUM, but rather some

thing like (SIGNED-BYTE 17) [if that's what SI:MEDIUM-SIZE-FIXNUM means.]


We would like to make this as an implementation recommendation,

however, rather than as a constraint, at this point.


About part (c): the restriction on T is meaningless, since the CLASS-OF

restriction dominates it. The restriction on MEMBER is OK, just to keep

TYPE-OF from being the silly definition that (type-of x) = `(member ,x).  I

think we're running out of good reasons to leave out SATISFIES, AND, OR,

NOT, and VALUES; they're probably no more useful, but we're probably

overspecifying for no good reason.


Some types are left out of the list in (a), including:


BIGNUM and FIXNUM were left out because their division was implementation

dependent.


KEYWORD was left out because under odd circumstances it is possible to

dynamically change the 'type' (e.g., by UNINTERN).


STANDARD-CHAR and STRING-CHAR were left out for the same reasons they

aren't built-in classes.


(These reasonas are not 100% compelling.)