Status: passed, as amended, Mar 89 X3J13.Forum: Cleanup
Issue: DYNAMIC-EXTENT
References: Scope and Extent
Category: ADDITION
Edit history: 27-Jun-88, Version 1 by Pitman (as issue STACK-LET)
15-Nov-88, Version 2 by Pitman (issue renamed, major revision)
11-Jan-89, Version 3 by Masinter (Moon's proposal)
05-Apr-89, Version 4 by Pitman and Steele (changes per X3J13)
Related-Issues: REST-ARGUMENT-EXTENT, WITH-DYNAMIC-EXTENT
Problem Description:
Sometimes a programmer knows that a particular data structure
will have only dynamic extent. In some implementations, it is
possible to allocate such structures in a way that will make them
easier to reclaim than by general purpose garbage collection
(eg, on the stack or in some temporary area). Currently, however,
there is no way to request the use of such an allocation mechanism.
Proposal (DYNAMIC-EXTENT:NEW-DECLARATION):
Introduce a new declaration called DYNAMIC-EXTENT. The arguments to
this declaration are names of variables.
It is permissible for an implementation to simply ignore this declaration.
In implementations which do not ignore it, the compiler (or interpreter)
is free to make whatever optimizations are appropriate given this
information; the most common optimization is to stack-allocate the
initial value of the object. What data types (if any) can have dynamic
extent will can vary from implementation to implementation.
Definition: Object <x> is an ``otherwise inaccessible part'' (OIP)
of <y> iff making <y> inaccessible would make <x> inaccessible.
(Note that every object is an OIP of itself.)
Suppose that construct <c> contains a DYNAMIC-EXTENT declaration for
variable <v> (which need not be bound by <c>). Consider the values
<w1>, ..., <wN> taken on by <v> during the course of some execution of
<c>. The declaration asserts that if object <x> is an OIP of <wI>
when <wI> ever becomes the value of <v>, then just after execution of
<c> terminates <x> will be either inaccessible or still an OIP of <v>.
If the assertion is ever violated, the conseqeuences are undefined.
Examples:
Since stack allocation of the initial value entails knowing at the
object's creation time that the object can be stack-allocated, it is
not generally useful to declare DYNAMIC-EXTENT for variables for
which have no lexically apparent initial value. For example,
(DEFUN F ()
(DECLARE (DYNAMIC-EXTENT X))
...))
would permit those compilers which wish to do so to stack-allocate the
list in X. However,
(DEFUN G (X) (DECLARE (DYNAMIC-EXTENT X)) ...)
could not typically permit a similar optimization in G because it would
be a modularity violation for the compiler to assume facts about G from
within F. Only an implementation which was willing to be responsible for
recompiling F if G's definition changed incompatibly could stack-allocate
the list argument to G in F.
Other interesting cases are:
(DEFUN G (X) (DECLARE (DYNAMIC-EXTENT X)) ...)
and (DEFUN F ()
(FLET ((G (X) (DECLARE (DYNAMIC-EXTENT X)) ...))
(G (LIST 1 2 3))))
where some compilers might realize the optimization was possible and others
might not.
An interesting variant of this is the so-called `stack allocated rest list'
which can be achieved (in implementations supporting the optimization) by:
(DECLARE (DYNAMIC-EXTENT X))
...)
Note here that although the initial value of X is not explicit, the F
function is responsible for assembling the list X from the passed arguments,
so the F function can be optimized by the compiler to construct a
stack-allocated list instead of a heap-allocated list in implementations
which support such.
In
(Y (CONS 'A2 (CONS 'B2 (CONS 'C2 NIL)))))
(DECLARE (DYNAMIC-EXTENT X Y))
...)
The OIP's of X are three conses, and the OIP's of Y are three other
conses. None of the symbols A1, B1, C1, A2, B2, C2, or NIL is an
OIP of X or Y. However, if a freshly allocated uninterned symbol had
been used, it would have been an OIP.
- - - - - - - -
(DOTIMES (I N)
(DECLARE (DYNAMIC-EXTENT I))
This is particularly instructive. Since I is an integer by the
definition of DOTIMES, but EQ and EQL are not necessarily equivalent for
integers, what are the OIP's of I, which this declaration
requires the body of the DOTIMES not to "save"? If the value of I is 3,
and the body does (SETQ FOO 3), is that an error? The answer is no, but
the interesting thing is that it depends on the implementation-dependent
behavior of EQ on numbers. In an implementation where EQ and EQL are
equivalent for 3, then 3 is not an OIP because (EQ I (+ 2 1)) is true,
and therefore there is another way to access the object besides
going through I. On the other hand, in an implementation where EQ and
EQL are not equivalent for 3, then the particular 3 that is the value of
I is an OIP, but any other 3 is not. Thus (SETQ FOO 3) is valid
but (SETQ FOO I) is erroneous. Since (SETQ FOO I) is erroneous in some
implementations, it is erroneous in all portable programs, but some other
implementations may not be able to detect the error.
- - - - - - - -
(DECLARE (DYNAMIC-EXTENT X))
(PRINT X)
NIL)
PRINT does not "save" any part of its input.
This prints (1 2 3)
- - - - - - - -
(DO ((L (LIST-ALL-PACKAGES) (CDR L)))
((NULL L))
(DECLARE (DYNAMIC-EXTENT L))
prints all packages; none of the newly-allocated list structures are saved.
- - - - - - - -
(DEFUN ADD (&REST X) (DECLARE (DYNAMIC-EXTENT X)) (APPLY #'+ X))
(ADD 1 2 3) => 6
I.e., useful way to declare that &REST lists have dynamic extent
- - - - - - - -
(DEFUN ZAP (X Y Z)
(DO ((L (LIST X Y Z) (CDR L)))
((NULL L))
(DECLARE (DYNAMIC-EXTENT L))
(ZAP 1 2 3)
prints 123
- - - - - - - -
(DEFUN ZAP (N M)
;; Computes (RANDOM (+ M 1)) at relative speed of roughly O(N).
;; It may be slow, but with a good compiler at least it
;; doesn't waste much heap storage. :-)
(LET ((A (MAKE-ARRAY N)))
(DECLARE (DYNAMIC-EXTENT A))
(DOTIMES (I N)
(DECLARE (DYNAMIC-EXTENT I))
(SETF (AREF A I) (RANDOM (+ I 1))))
(AREF A M)))
(< (ZAP 5 3) 3) => T
- - - - - - - -
The following are in error, since the value of X is used outside of its
extent:
(LENGTH (LIST (LET ((X (LIST 1 2 3))) (DECLARE (DYNAMIC-EXTENT X)) X)))
(PROGN (LET ((X (LIST 1 2 3))) (DECLARE (DYNAMIC-EXTENT X)) X)
NIL)
- - - - - - - -
Rationale:
This permits a programmer to offer advice to an implementation about
what may be stack-allocated for efficiency.
It may be difficult or impossible for a compiler to infer this
same information statically.
Since a number of implementations offer this capability and there
is demand from users for access to the capability, this ``codifies
existing practice.''
Because this approach is purely lexical, it does not interact badly with
other programs in the way that the macro WITH-DYNAMIC-EXTENT (see issue
by same name) would.
Current Practice:
Symbolics Genera and Symbolics Cloe offer stack allocation, though not
in this strategy.
[KMP thinks that] Lucid supports the proposal.
Cost to Implementors:
No cost is forced since implementations are permitted to simply
ignore the DYNAMIC-EXTENT declaration.
Cost to Users:
None. This change is upward compatible.
There may be some hidden costs to debugging using this declaration (or any
feature which permits the user to access dynamic extent objects without
the compiler proving that they are appropriate). If the user misdeclares
something and returns a pointer into the stack (or stores it in the heap),
an undefined situation may result and the integrity of the Lisp storage
mechanism may be compromised. Debugging these situations may be tricky,
but users who have asked for this feature have indicated a willingness
to deal with such costs. Nevertheless, the perils should be clearly
documented and casual users should not be encouraged to use this
declaration.
Cost of Non-Adoption:
Some portable code would be forced to run more slowly (due to
GC overhead), or to use non-portable language features.
Benefits:
The cost of non-adoption is avoided.
Aesthetics:
This declaration allows a fairly low level optimization to work
by asking the user to provide only very high level information.
The alternatives (sharpsign conditionals, some of which may
lead to more bit-picky abstractions) are far less aesthetic.
Discussion:
A previous version of this proposal suggested primitives STACK-LET
and STACK-LET*. Consensus was that the more general declaration facility
would be more popular.
Moon came up with a description of something called a "proper part" which
Steele formalized into the idea of an "otherwise inaccessible part". The
two are essentially interchangeable, but Steele's description was more
rigorous.
KMP: ... it still raises the question of whether we should define
per-function for every CL function whether any of the arguments is
permitted to be "saved" so that CL programs don't get any funny
surprises. If we don't, it ends up being implementor's discretion how
to resolve cases ... and everyone might not agree that all cases are
... obvious ...
JonL: PDP10 MacLisp had a similar problem w.r.t pdlnums. That is why
"identity" functions were so troublsome for it -- in order to
return a guaranteed safe value, it typically had to copy it's
pdlnum argument, thereby making some cases of "fast arithmetic"
code much worse than interpreted code! [Remember PRINT in MacLisp?
it returns T rather than it's argument for just this reason.]
It is necessary for an optimizing compiler to know something about
what happens to the data it passes along to "system" functions; for
example, it could assume that GET doesn't clobber the list given
to it, nor does it retain pointers to any part of it [what was the
terminology in the revised proposal? "saved"? and "proper part"?]
The issue LISP-SYMBOL-REDEFINITION might help here, in that an
implementation's compilers could depend upon it's own internal
database. But it wouldn't hurt at all to have some of these
requirements "up front" in the standard.
It was generally agreed that we would also like to consider a proposal
on dynamic extent functions at the next meeting. (Sandra said she would
prepare one, and has already done so. See issue DYNAMIC-EXTENT-FUNCTION.)