LD-REDEFINITION-ACTION

to allow redefinition without undoing
Major Section:  MISCELLANEOUS

Ld-redefinition-action is an ld special (see ld). The accessor is (ld-redefinition-action state) and the updater is (set-ld-redefinition-action val state).

WARNING! If ld-redefinition-action is non-nil then ACL2 is liable to be made unsafe or unsound either by ill-considered definitions, or because redefining a macro or inlined function called in the body of a function, g, may not cause the new definition to be called by g. The keyword command :redef will set ld-redefinition-action to a convenient setting allowing unsound redefinition. See below.

When ld-redefinition-action is nil, redefinition is prohibited. In that case, an error message is printed upon any attempt to introduce a name that is already in use. There is one exception to this rule. It is permitted to redefine a function symbol in :program mode to be a function symbol in :logic mode provided the formals and body remain the same. This is the standard way a function ``comes into'' logical existence.

Throughout the rest of this discussion we exclude from our meaning of ``redefinition'' the case in which a function in :program mode is identically redefined in :logic mode. At one time, ACL2 freely permitted the signature-preserving redefinition of :program mode functions but it no longer does. See redefining-programs.

When ld-redefinition-action is non-nil, you are allowed to redefine a name that is already in use. The system may be rendered unsound by such an act. It is important to understand how dangerous redefinition is. Suppose fn is a function symbol that is called from within some other function, say g. Suppose fn is redefined so that its arity changes. Then the definition of g is rendered syntactically ill-formed by the redefinition. This can be devastating since the entire ACL2 system assumes that terms in its database are well-formed. For example, if ACL2 executes g by running the corresponding function in raw Common Lisp the redefinition of fn may cause raw lisp to break in irreparable ways. As Lisp programmers we live with this all the time by following the simple rule: after changing the syntax of a function don't run any function that calls it via its old syntax. This rule also works in the context of the evaluation of ACL2 functions, but it is harder to follow in the context of ACL2 deductions, since it is hard to know whether the database contains a path leading the theorem prover from facts about one function to facts about another. Finally, of course, even if the database is still syntactically well-formed there is no assurance that all the rules stored in it are valid. For example, theorems proved about g survive the redefinition of fn but may have crucially depended on the properties of the old fn. In summary, we repeat the warning: all bets are off if you set ld-redefinition-action to non-nil.

ACL2 provides some enforcement of the concern above, by disabling certify-book if any world-changing events exist in the certification world that were executed with a non-nil value of 'ld-redefinition-action. (This value is checked at the end of each top-level command, but the value does not change during evaluation of embedded event forms; see embedded-event-form.)

If at any point in a session you wish to see the list of all names that have been redefined, see redefined-names.

That said, we'll give you enough rope to hang yourself. When ld-redefinition-action is non-nil, it must be a pair, (a . b). The value of a determines how the system interacts with you when a redefinition is submitted. The value of b allows you to specify how the property list of the redefined name is to be ``renewed'' before the redefinition.

There are several dimensions to the space of possibilities controlled by part a: Do you want to be queried each time you redefine a name, so you can confirm your intention? (We sometimes make typing mistakes or simply forget we have used that name already.) Do you want to see a warning stating that the name has been redefined? Do you want ACL2 system functions given special protection from possible redefinition? Below are the choices for part a:

:query -- every attempt to redefine a name will produce a query. The query will allow you to abort the redefinition or proceed. It will will also allow you to specify the part b for this redefinition. :Query is the recommended setting for users who wish to dabble in redefinition.

:warn -- any user-defined function may be redefined but a post-redefinition warning is printed. The attempt to redefine a system name produces a query. If you are prototyping and testing a big system in ACL2 this is probably the desired setting for part a.

:doit -- any user-defined function may be redefined silently (without query or warning) but when an attempt is made to redefine a system function, a query is made. This setting is recommended when you start making massive changes to your prototyped system (and tire of even the warning messages issued by :warn).

In support of our own ACL2 systems programming there are two other settings. We suggest ordinary users not use them.

:warn! -- every attempt to redefine a name produces a warning but no query. Since ACL2 system functions can be redefined this way, this setting should be used by the only-slightly-less-than supremely confident ACL2 system hacker.

:doit! -- this setting allows any name to be redefined silently (without query or warnings). ACL2 system functions are fair game. This setting is reserved for the supremely confident ACL2 system hacker. (Actually, this setting is used when we are loading massively modified versions of the ACL2 source files.)

Part b of ld-redefinition-action tells the system how to ``renew'' the property list of the name being redefined. There are two choices:

:erase -- erase all properties stored under the name, or

:overwrite -- preserve existing properties and let the redefining overwrite them.

It should be stressed that neither of these b settings is guaranteed to result in an entirely satisfactory state of affairs after the redefinition. Roughly speaking, :erase returns the property list of the name to the state it was in when the name was first introduced. Lemmas, type information, etc., stored under that name are lost. Is that what you wanted? Sometimes it is, as when the old definition is ``completely wrong.'' But other times the old definition was ``almost right'' in the sense that some of the work done with it is still (intended to be) valid. In that case, :overwrite might be the correct b setting. For example if fn was a function and is being re-defun'd with the same signature, then the properties stored by the new defun should overwrite those stored by the old defun but the properties stored by defthms will be preserved.

In addition, neither setting will cause ACL2 to erase properties stored under other symbols! Thus, if FOO names a rewrite rule which rewrites a term beginning with the function symbol BAR and you then redefine FOO to rewrite a term beginning with the function symbol BAZ, then the old version of FOO is still available (because the rule itself was added to the rewrite rules for BAR, whose property list was not cleared by redefining FOO).

The b setting is only used as the default action when no query is made. If you choose a setting for part a that produces a query then you will have the opportunity, for each redefinition, to specify whether the property list is to be erased or overwritten.

The keyword command :redef sets ld-redefinition-action to the pair (:query . :overwrite). Since the resulting query will give you the chance to specify :erase instead of :overwrite, this setting is quite convenient. But when you are engaged in heavy-duty prototyping, you may wish to use a setting such as :warn or even :doit. For that you will have to invoke a form such as:

(set-ld-redefinition-action '(:doit . :overwrite) state) .

Encapsulate causes somewhat odd interaction with the user if redefinition occurs within the encapsulation because the encapsulated event list is processed several times. For example, if the redefinition action causes a query and a non-local definition is actually a redefinition, then the query will be posed twice, once during each pass. C'est la vie.

Finally, it should be stressed again that redefinition is dangerous because not all of the rules about a name are stored on the property list of the name. Thus, redefinition can render ill-formed terms stored elsewhere in the database or can preserve now-invalid rules. See redundant-events, in particular the section ``Note About Unfortunate Redundancies,'' for more discussion of potential pitfalls of redefinition.