CLHS: Issue LOAD-OBJECTS Writeup

Issue LOAD-OBJECTS Writeup

Issue: LOAD-OBJECTS

References: none

Related issues: LOAD-TIME-EVAL,

CONSTANT-COMPILABLE-TYPES,

CONSTANT-CIRCULAR-COMPILATION

Category: ADDITION

Forum: Cleanup

Edit history: Version 1, 2-Jan-89, by Moon (for discussion)

Version 2, 13-Jan-89, by Moon (draft updated from discussion)

Version 3, 9-Mar-89, by Moon (changes suggested by discussion)

Version 4, 4-Apr-89, by Pitman (changes per X3J13 Mar 89;

MAKE-LOAD-FORM-USING-SLOTS => MAKE-LOAD-FORM-SAVING-SLOTS)

Status: Accepted by an 18-0 vote, March 1989.

Problem description:

Common Lisp doesn't provide any way to use an object of a user-defined

type (defined with DEFCLASS or DEFSTRUCT) as a constant in a program

compiled with COMPILE-FILE. The problem is that LOAD has to be able

to "reconstruct" an equivalent object when the compiled-code file is

loaded, but the programmer has no way to tell LOAD how to do that.

Proposal (LOAD-OBJECTS:MAKE-LOAD-FORM):

Define a new generic function named MAKE-LOAD-FORM, which takes one

argument and returns two values. The argument is an object that is

referenced as a constant or as a self-evaluating form in a file being

compiled by COMPILE-FILE. The objective is to enable LOAD to

construct an equivalent object.

The first value, called the "creation form," is a form that, when

evaluated at load time, should return an object that is equivalent to

the argument. The exact meaning of "equivalent" depends on the type

of object and is up to the programmer who defines a method for

MAKE-LOAD-FORM. This is the same type of equivalence discussed

in issue CONSTANT-COMPILABLE-TYPES.

The second value, called the "initialization form," is a form that,

when evaluated at load time, should perform further initialization of

the object. The value returned by the initialization form is ignored.

If the MAKE-LOAD-FORM method returns only one value, the

initialization form is NIL, which has no effect. If the object used

as the argument to MAKE-LOAD-FORM appears as a constant in the

initialization form, at load time it will be replaced by the

equivalent object constructed by the creation form; this is how the

further initialization gains access to the object.

Both the creation form and the initialization form can contain

references to objects of user-defined types (defined precisely below).

However, there must not be any circular dependencies in creation forms.

An example of a circular dependency is when the creation form for the

object X contains a reference to the object Y, and the creation form

for the object Y contains a reference to the object X. A simpler

example would be when the creation form for the object X contains

a reference to X itself. Initialization forms are not subject to

any restriction against circular dependencies, which is the entire

reason that initialization forms exist. See the example of circular

data structures below.

The creation form for an object is always evaluated before the

initialization form for that object. When either the creation form or

the initialization form references other objects of user-defined types

that have not been referenced earlier in the COMPILE-FILE, the

compiler collects all of the creation and initialization forms. Each

initialization form is evaluated as soon as possible after its

creation form, as determined by data flow. If the initialization form

for an object does not reference any other objects of user-defined

types that have not been referenced earlier in the COMPILE-FILE, the

initialization form is evaluated immediately after the creation form.

If a creation or initialization form F references other objects of

user-defined types that have not been referenced earlier in the

COMPILE-FILE, the creation forms for those other objects are evaluated

before F, and the initialization forms for those other objects are

also evaluated before F whenever they do not depend on the object

created or initialized by F. Where the above rules do not uniquely

determine an order of evaluation, which of the possible orders of

evaluation is chosen is unspecified.

While these creation and initialization forms are being evaluated, the

objects are possibly in an uninitialized state, analogous to the state

of an object between the time it has been created by ALLOCATE-INSTANCE

and it has been processed fully by INITIALIZE-INSTANCE. Programmers

writing methods for MAKE-LOAD-FORM must take care in manipulating

objects not to depend on slots that have not yet been initialized.

It is unspecified whether LOAD calls EVAL on the forms or does some

other operation that has an equivalent effect. For example, the

forms might be translated into different but equivalent forms and

then evaluated, they might be compiled and the resulting functions

called by LOAD, or they might be interpreted by a special-purpose

interpreter different from EVAL. All that is required is that the

effect be equivalent to evaluating the forms.

COMPILE-FILE calls MAKE-LOAD-FORM on any object that is referenced as

a constant or as a self-evaluating form, if the object's metaclass is

STANDARD-CLASS, STRUCTURE-CLASS, any user-defined metaclass (not a

subclass of BUILT-IN-CLASS), or any of a possibly-empty

implementation-defined list of other metaclasses. COMPILE-FILE will

only call MAKE-LOAD-FORM once for any given object (compared with EQ)

within a single file.

It is valid for user programs to call MAKE-LOAD-FORM in other

circumstances, providing the argument's metaclass is not BUILT-IN-CLASS

or a subclass of BUILT-IN-CLASS.

Define a new function named MAKE-LOAD-FORM-SAVING-SLOTS, which takes

one required argument and one optional argument and returns two

values. This can be useful in user-written MAKE-LOAD-FORM methods.

The first argument is the object. The optional second argument is a

list of the names of the slots to preserve; it defaults to all of the

local slots. MAKE-LOAD-FORM-SAVING-SLOTS returns forms that construct

an equivalent object using MAKE-INSTANCE and SETF of SLOT-VALUE for

slots with values, or SLOT-MAKUNBOUND for slots without values, or

using other functions of equivalent effect.

MAKE-LOAD-FORM-SAVING-SLOTS returns two values, thus it can deal with

circular structures. MAKE-LOAD-FORM-SAVING-SLOTS works for any object

of metaclass STANDARD-CLASS or STRUCTURE-CLASS. Whether the result is

useful in an application depends on whether the object's type and slot

contents fully capture the application's idea of the object's state.

MAKE-LOAD-FORM of an object of metaclass STANDARD-CLASS or

STRUCTURE-CLASS for which no user-defined method is applicable signals

an error. It is valid to implement this either by defining default

methods on STANDARD-OBJECT and STRUCTURE-OBJECT that signal an error

or by having no applicable method for those classes.

Examples:

;; Example 1

(defclass my-class ()

((a :initarg :a :reader my-a)

(b :initarg :b :reader my-b)

(c :accessor my-c)))

(defmethod shared-initialize ((self my-class) ignore &rest ignore)

(unless (slot-boundp self 'c)

(setf (my-c self) (some-computation (my-a self) (my-b self)))))

(defmethod make-load-form ((self my-class))

`(make-instance ',(class-name (class-of self))

:a ',(my-a self) :b ',(my-b self)))

In this example, an equivalent instance of my-class is reconstructed

by using the values of two of its slots. The value of the third slot

is derived from those two values.

Another way to write the last form in the above example would have been

(defmethod make-load-form ((self my-class))

(make-load-form-saving-slots self '(a b)))

;; Example 2

(defclass my-frob ()

((name :initarg :name :reader my-name)))

(defmethod make-load-form ((self my-frob))

`(find-my-frob ',(my-name self) :if-does-not-exist :create))

In this example, instances of my-frob are "interned" in some way.

An equivalent instance is reconstructed by using the value of the

name slot as a key for searching existing objects. In this case

the programmer has chosen to create a new object if no existing

object is found; alternatively she could have chosen to signal an

error in that case.

;; Example 3

(defclass tree-with-parent () ((parent :accessor tree-parent)

(children :initarg :children)))

(defmethod make-load-form ((x tree-with-parent))

(values

;; creation form

`(make-instance ',(class-of x) :children ',(slot-value x 'children))

;; initialization form

`(setf (tree-parent ',x) ',(slot-value x 'parent))))

In this example, the data structure to be dumped is circular, because

each parent has a list of its children and each child has a reference

back to its parent. Suppose make-load-form is called on one object in

such a structure. The creation form creates an equivalent object and

fills in the children slot, which forces creation of equivalent

objects for all of its children, grandchildren, etc. At this point

none of the parent slots have been filled in. The initialization form

fills in the parent slot, which forces creation of an equivalent

object for the parent if it was not already created. Thus the entire

tree is recreated at load time. At compile time, MAKE-LOAD-FORM is

called once for each object in the true. All of the creation forms

are evaluated, in unspecified order, and then all of the

initialization forms are evaluated, also in unspecified order.

;; Example 4

(defstruct my-struct a b c)

(defmethod make-load-form ((s my-struct))

(make-load-form-saving-slots s))

In this example, the data structure to be dumped has no special

properties and an equivalent structure can be reconstructed

simply by reconstructing the slots' contents.

Rationale:

Only the programmer who designed a class can know the correct

way to reconstruct objects of that class at load time, therefore

the reconstruction should be controlled by a generic function.

Using EVAL as the interface for telling LOAD what to do provides

full generality.

MAKE-LOAD-FORM returns two values so that circular structures can

be handled. If CONSTANT-CIRCULAR-COMPILATION is rejected,

MAKE-LOAD-FORM will only return one value, although implementations

that make an extension to support circular constants will probably

also make the extension to accept two values from MAKE-LOAD-FORM.

The default for class objects and structures is to signal an error,

rather than picking some particular object reconstruction technique,

because no reconstruction technique is appropriate for all objects.

It only takes two lines of code, as in example 4, to instruct the

compiler to use the technique that most often has been suggested

as the default.

MAKE-LOAD-FORM has a natural resemblance to PRINT-OBJECT, as a hook

for the programmer to control the system's actions.

The order of evaluation rules for creation and initialization forms

eliminate the possibility of partially initialized objects in the

absence of circular structures, and reduce it to the minimum possible

in the presence of circular structures. This allows nodes in

non-circular structures to be built out of fully initialized subparts.

Current practice:

Symbolics Flavors has something like this, but under a different name.

The name Symbolics uses is not suitable for standardization.

JonL reports that Lucid is getting more and more requests for this.

Cost to Implementors:

This seems like only a few one-line changes in the compiled-code

file writer and reader. MAKE-LOAD-FORM-SAVING-SLOTS is a couple

dozen lines of code, assuming the presence of the CLOS metaobject

protocol or an implementation-dependent equivalent.

Cost to Users:

None.

Cost of non-adoption:

Serious impairment of the ability to use extended-type objects. Each

implementation will probably make up its own version of this as an

extension.

Performance impact:

None.

Benefits:

See Cost of non-adoption.

Esthetics:

No significant positive or negative impact.

Discussion:

It would be possible to define an additional level of protocol that

allows multiple classes to contribute to the reconstruction of an

object, combining initialization arguments contributed by each class.

Since a user can easily define that in terms of MAKE-LOAD-FORM without

modifying the Lisp system, it is not being proposed now.

Any type that has a read syntax is likely to appear as a quoted

constant or inside a quoted constant. Pathnames are one example, user

programs often define others. Also many implementations provide a way

to create a compiled-code file full of data (rather than compiled Lisp

programs), and such data probably include extended-type objects.

Moon supports this. David Gray and John Rose made major contributions

to the discussion that produced this improved version 2 proposal.