Issue SUBTYPEP-TOO-VAGUE, version 5

[Guy Steele <gls@Think.COM>]

This is a proposed amendment to version 4 passed in January 1989 at Kauai.

Issue:		SUBTYPEP-TOO-VAGUE
References:	CLtL p. 72-73
Category:	CLARIFICATION
Edit History:   Version 1, 11 Jul 1988 (Sandra Loosemore)
                Version 2, 19 Jul 1988 (Sandra Loosemore)
                Version 3,  6-Oct-88 (Masinter)
                Version 4,  7-Oct-88 (Masinter, per Moon's comments)

Problem Description:

[From version 4]

The description of SUBTYPEP allows it to return a second value of NIL
when the relationship between the two types cannot be determined.  In
some cases this is a reasonable thing to do because it is impossible
to tell (if the SATISFIES type specifier is involved), and in other
cases the relationships between types are not well-defined (for
example, the VALUES type specifier or the list form of the FUNCTION
type specifier). 

Some implementations, however, have apparently interpreted this to
mean that it is permissible for SUBTYPEP to "give up" and return a
second value of NIL in some cases where it actually would be possible
to determine the relationship.  This makes it difficult to depend on
subtype relationships in portable code.

[Addition for version 5]

There are two problems with version 4.  First is that of the first three
bulleted points in the version 4 proposal:

    * Clarify that SUBTYPEP will return a second value of NIL
    only when either of the type specifiers involves the SATISFIES, NOT, 
    AND, OR, MEMBER. SUBTYPEP will not return a second
    value of NIL when both arguments involve only the words in Table 4-1, or
    names of DEFSTRUCT- or DEFCLASS-defined types, or user-defined deftypes
    that expand into only those words and/or names.

    * SUBTYPEP should signal an error when handed (for either argument)
    a type specifier that involves VALUES or the list form of the FUNCTION
    type.

    * SUBTYPEP must always return values T T in the case where the two
    type specifiers (or their expansions) are EQUAL.

any two have significant overlap, and indeed all three can overlap;
version 4 contained no indication of how this conflict should be resolved.

Second is that version 4 calls for SUBTYPEP to signal an error (at least at
high safety)even when the arguments are valid type specifiers, but this can
make it harder to use SUBTYPEP.  These are cases that returning NIL NIL
was supposed to cover.

This version replaces the three bulleted points above with a single point
and some observations about its consequences.  This version eliminates
the requirement to signal an error.


Proposal: SUBTYPEP-TOO-VAGUE:CLARIFY-MORE

A type specifier "involves" a word like SATISFIES, MEMBER, NOT, etc.
if it either contains it directly or as the result of expansion of a 
DEFTYPE  defined type specifier. 

* Clarify that SUBTYPEP is permitted to return NIL NIL only when
  at least one argument involves SATISFIES, AND, OR, NOT, MEMBER,
  VALUES, or the list form of FUNCTION.

  Note that one consequence of this is that if neither argument
  involves any of these type specifiers, then SUBTYPEP is obliged
  to determine the relationship accurately.  In particular, SUBTYPEP
  must return T T if the arguments are EQUAL and do not involve
  any of the above-stated type specifiers.

* Clarify that the relationships between types reflected by SUBTYPEP
are those specific to the particular implementation.  For example, if
an implementation supports only a single type of floating-point numbers,
in that implementation (SUBTYPEP 'FLOAT 'LONG-FLOAT) would return T T
(since the two types would be identical).

Rationale:

Specifying the behavior of SUBTYPEP makes it more useful. Otherwise,
programs cannot rely on any more than NIL NIL as return values.

It is generally conceded that it is impossible to determine the
relationships between types defined with the SATISFIES specifier.
MEMBER, AND, OR, and NOT are messy to deal with.   

Current Practice:

The implementation of SUBTYPEP in (the original) HPCL does not try to
expand type specifiers defined with DEFTYPE and does not recognize
EQUAL type specifiers as being equivalent.  Most other implementations
appear to be substantially in conformance with the proposal.

Cost to implementors:

Some implementations will have to rewrite and/or extend parts of SUBTYPEP.

Cost to users:

Its hard to imagine a portable program that depends heavily
on SUBTYPEP. This proposal does not require any implementation
to "handle" fewer cases of SUBTYPEP.

Benefits:

An area of confusion in the language is cleared up.  Usages of SUBTYPEP
will be more portable.

Discussion:

The handling of FLOAT and SINGLE-FLOAT  appeared to be the 
consensus from a discussion on the common-lisp mailing list some
 time ago.

It would not be too onerous to require implementations to handle
the cases where one but not the other type specifier involves
OR, AND, NOT or MEMBER, but the specification becomes 
cumbersome.

A related issue is clarifying what kinds of type specifiers must be
recognized by functions such as MAKE-SEQUENCE and COERCE.  For example,
HPCL complains that (SIMPLE-ARRAY (UNSIGNED-BYTE *) (*)) is not a valid
sequence type when passed to MAKE-SEQUENCE, although SUBTYPEP does
recognize it to be a subtype of SEQUENCE.  Should this proposal be
extended to deal with these issues, or is another proposal in order?

The rules for comparing the various type specifiers (such as ARRAY)
need to be spelled out in detail.