Motes, singles, and other issues of array theory for LISP

[GLS at SU-AI (Guy Steele)]

Here are a few thoughts regarding the treatment of motes
and singles in LISP for the purposes of the implementation
of More-type array-theoretic functions.  I would appreciate
commants with regard to suitability of concept and requirements
of implementation.  I bring this up because the design of
of the appropriate data types is a pending issue for NIL.

Arrays in general in NIL ought to work for boundary cases and
have no artificial limits if possible.  Thus, arrays should
be allowed to have axes of zero length, and to have zero
axes.  The latter situation is given an interpretation by array
theory by considering it to be a cell, for which there is
precisely one valid index tuple, namely the empty index tuple.
(If one realizes that the number of elements in a rectangular
array is the product of the axis lengths, then a zero-axis
array clearly has one element, and the index tuple for that element
has no indices.)  In array theory, More calls such a zero-axis
array a *single*.

Such zero-axis arrays could be implemented by a special type,
and used for various interesting purposes.  For example,
a dynamic variable holding a flonum cell (a single whose
element is always a flonum) could solve the PDLNMK-with-
SPECIAL-variables problem.  Alternatively, it could just be
in the standard array format.  (RMS has suggested that one-dimensional
arrays, presumably with standard bounds 0:n-1, be implemented
as simple vectors.  This sounds fine to me except for one glitch,
namely that one-dimensional arrays with leaders or other funny
stuff would not be interchangeable with arrays (i.e. vectors)
which did not have such junk.)

Now array theory takes all primitive data types (numbers,
characters, etc.) to be classes of objects each of which
is a single whose sole element is itself.  (Such objects
More calls *motes*.)  Thus, for purposes
of generalized indexing, element extraction, etc.,
a number or character is considered to be a zero-dimensional
array.  This causes no problem for array-theoretic functions
implemented in NIL, which can just check for such data types
and treat them specially.

What worries me is the treatment of motes and singles in
a way that is implementationally and computationally convenient,
especially in the presence of array assignment, and yet
consistent with the treatment of array theory.  For example,
a single containing a mote is considered to be equivalent
to the mote itself.  Suppose such a single A is created.
As an array, its element can be assigned to (using STORE
or whatever).  If a mote is assigned to it, there is no
convenient way to make it be the mote itself, at least
in the sense of EQ or even EQUAL.  Alternatively, suppose
one were to say (STORE (AR-0 43) 'foo), i.e. store 'foo
into the single (mote) 43.  It is clearly not desirable
to allow the alteration of 43!  I guess it may suffice
merely to decree that array assignment to motes is illegal.

By the way, motes remind me an awful lot of self-lists.
In a cdr-coded scheme, such a self-list could occupy
one cell, and we could decree that the car of a self-list
would be itself and its cdr be ().  Then singles (cells)
would be the generalization of this analogous to the
definition of 1-hunks on the PDP-10.

It is unfortunate that More chose to use the word *list*
for a purpose in array theory.  Aside from that there are
almost no conflicts with LISP terminology.  This in turn
raises the question of whether LISP lists are to be
considered motes by the array-theoretic functions or not.

-------