This note is an attempt to summarize the discussion JONL and I

[GLS at SU-AI (Guy Steele)]

and others had about the nature of variable bindings and references
in NIL.  If I have made any mistakes, or failed to represent the
consensus of opinion, please let me know.  I am going to try to
mention the alternatives that were considered, and also contrast
NIL, MacLISP, LISP 1.5, and SCHEME.
"Variables" in NIL are used for at least three different purposes:
[1] as the names of "globally" defined procedures, e.g. CONS.
[2] as "local" names of data objects.
[3] as "global" (i.e. dynamic) names of data objects.
Now the names of functions are normally thought of as being
"constant", while global names of data objects are though of
as varying over time.  It may be a useful notion, we have said,
to allow binding of global function cells.  Conversely, LISP 1.5
had a notion of a globally constant data object name (APVALs).
Now one problem with APVALs was that they were checked for BEFORE
the a-list was searched -- there was no way to "shadow" them.
Besides the issue of binding function names, there is also the
issue of wanting to treat functions per se as data objects.

In any case, we assume there are three sorts of variables (at least
-- one could for example complete the symmetry and imagine local
functional variables as well as global ones).  Now if gross screws
are to be avoided, there must be a way for the user to precisely
specify which variable is to be bound and which variable is to be
referenced at any point in his program.

In MacLISP, functions are not really first class data objects.
The model is that there are two distinct "evaluation contexts"
for names.  One is the function position of the combination,
and the other is essentially all other places.  In the function
position, an atomic symbol is interpreted as a reference to
the function cell, unless undefined, in which case the dynamic
value is used (modulo EVALPUNT, and the fact that a function as such
cannot live in a value cell, but only a symbol).  I have no idea
what happens when you write (FUNCTION symbol) in the function
position, either interpreted or compiled.  There is no good way
to refer to a function object in a local or dynamic variable
from the functional context -- instead, the FUNCALL construction
must be used.
In non-functional contexts, the "function cell" can be referred
to by writing the construct (FUNCTION symbol), and a plain atomic
symbol refers to a dynamic value.  (MacLISP has no consistent notion
of a local variable.  The interpreter cannot handle them correctly.
The compiler assumes that variables are by default local, leading
to a non-robustness, because by default interpreted and compiled
versions of the same program are likely to have radically differing
behaviors.)
MacLISP has only one kind of variable binding (dynamic, local
being excluded as above).  There is no way to bind a function
cell.

In SCHEME, there is (or ought to be) a consistent notion of
a global function cell, a dynamic value cell, and local variables.
SCHEME provides no facility for binding function cells; the model
is that the function cells provide the outermost "contour" for
the lexical environment.  There is only a single evaluation context
for names.  In that context, one writes an atomic symbol to refer
to a lexical variable, and (FLUID symbol) to refer to a dynamic
variable.  There is no way as such to refer directly to a global
function cell, because a function cell can always be lexically
shadowed.  This is seldom a problem for user programs, but can
pose problems for macros.
For binding, SCHEME distinguishes fluid bindings from lexical ones.
There is no way to bind a global function cell.  Thus function cells
can be shadowed lexically but not dynamically.

I believe what was decided, at last meeting, for NIL was
that the notion of two distinct "evaluation contexts" for names
be preserved, the model being that function names are usually thought
of as constants in some sense.  However, there may be a way
to shadow such definitions dynamically by binding the function
cell.  On the third hand, such names cannot be shadowed lexically.
Thus, in the "functional context", only a small class of expressions
are legitimate: atomic symbols (referring to the global, possibly
dynamic, function cell), and a few special expressions such as
LAMBDA-expressions.  (Are macro calls to be allowed to work in functional
context?)  There is no way to refer to a dynamic variable or a local
variable in function context -- one must use FUNCALL, as in MacLISP.
(It would be possible to define constructs which would legitimately
refer to such other variables in function context, but it is not
clear whether this is better or worse as a user syntax than FUNCALL.)
In other evaluation contexts, the three kinds of variable can be
referred to by the constructs:
	symbol			[local or dynamic variable]
	(DYNAMIC symbol)	[dynamic variable]
	(FUNCTION symbol)	[global function cell]
Now as we discussed, there is a set of nasty cases where an ambiguity
arises from the fact that an atomic symbol can be viewed as either
a local or a dynamic variable, and so one must have a way to force
the dynamic interpretation; hence the provision of the DYNAMIC construct.
There is still some doubt in my mind as to whether we should simply
tell the user always to use the DYNAMIC construct, which always but
always works (in which case a macro character similar to ' for QUOTE
should be provided), or tell him it mostly works not to use it
and describe the kind of hairy situations where it will lose.

Now it is interesting to note some similarities and differences between
the NIL proposal and the SCHEME proposal (which does not fully exist
either as an implementation).  Both implicitly acknowledge the existence
of three distinct kinds of variable in terms of which the user may wish
to think.  Both try to squeeze the three kinds into two sorts, because
the user may wish to mix sorts in his mind occasionally.  In SCHEME,
the global functional and local (lexical) types of variables are allowed
to be confused: atomic symbols are conditionally treated first
as local, and then as global functional if that fails.  This seems to
be good for styles of programming in which it is likely that globally
defined functions will be referred to as objects to be passed as
arguments;  it was of course just such styles that SCHEME was in part
designed to explore.  The dynamic variables are kept apart, because
they behave differently from the other two.  (The relationship between
global "data" variables and global "function" variables is not apparent
in SCHEME because there is no facility for dynamic binding of global
functional variables.  In writing "The Art of the Interpreter" I first
realized how important this notion might be to SCHEME.)
In the NIL proposal, the three kinds of variable are also squeezed into
two, but using a different division: here it is the global functional
which is the loner, and the other two are confused.  This suits a style
of programming in which functions are seldom treated as data objects,
and so it is mostly useful to separate names for functions from names
for other kinds of data objects.  It may be useful to confuse local
with dynamic variables for some purposes.  The lookup discipline is
(as I understand it) to check for a local definition first, and if
that fails, use the dynamic value cell.
Both proposals have the serious disadvantage that whichever kind of
variable is confused with local variables can be accidentally shadowed
by a lexical variable (note the similarity between the conditional
lookup disciplines in the two proposals as I have (I hope accurately)
described them).  This poses little problem to the user, but can
make things tricky for macros which cannot see their precise binding
contexts.  In each case this necessitates the introduction of a special
construct for eliminating the confusion.  In the case of NIL, it is
the dynamic construct; I am considering likewise proposing the
FUNCTION construct for SCHEME to eliminate the analogous ambiguity.

As may perhaps be well known, I still happen to favor the SCHEME
proposal slightly, because it simplifies the description of the language
(requiring only one kind of name evaluation context) while maintaining
that standard "Cambridge Polish" functional notation for the simple
cases; i.e. (CONS A B) means applying the function names by "CONS"
to two arguments named by "A" and "B".  The FUNCALL notation is not needed.
On the other hand, for those used to using one name as both the name
of a function as as the name of a data object (and there are often
good reasons for this -- e.g. LIST and EXP!), the separation inherited
from MacLISP contained in the NIL proposal can be valuable.
In conclusion, I repeat what I said earlier: the NIL proposal, the
SCHEME proposal, and yet other possibilities are each appropriate to
a certain style of programming.  Certainly the choice of model and
discipline will have the effect of encouraging certain styles of
programming and discouraging others.  This choice is therefore not
to be made lightly.  I favor a rather liberal approach, namely to
try to encompass as many styles comfortably as is possible without
conflict, other things being equal.

Now some issues that have not yet been touched upon here or in our
meetings include that of local functions, encompassing both
closures of various sorts and LABEL or LABELS or whatever.  I think
it is true that the MacLISP handling of LABEL is not acceptable.
What can be done to fix this?  Also, I recommend the LABELS syntax
as being analogous to LET in allowing mutually recursive definitions;
indeed, it is a subset of LETREC (cf. Landin).  A useful borderline
case between the use of a LABELS-like construct and packages is the
construction of OWN variables by some such piece of code as:
	(PROG (OWN0 OWN1 OWN2 ...)
	      (DEFUN FUNNYFN1 ...)
	      (DEFUN FUNNYFN2 ...)
	      ...)
Gerry and I have been considering simply defining DEFINE to be a macro
which expands into an ASET', i.e. not to hide so carefully the fact that
it simply clobbers afunctional value into a certain variable (though the
user can ignore this fact mostly if he likes):
	(DEFINE FOO (X Y) BARF)  =>  (ASET' FOO (LAMBDA (X Y) BARF))
Do we similarly want to simply define DEFUN as a macro defined in
terms of FSETQ or SETQ or whatever?

-------