Major Section: ACL2 Documentation
Example Form (from community book finite-set-theory/total-ordering.lisp): (defthm <<-trichotomy (implies (and (ordinaryp x) (ordinaryp y)) (or (<< x y) (equal x y) (<< y x))) :rule-classes ((:rewrite :corollary (implies (and (ordinaryp x) (ordinaryp y) (not (<< x y)) (not (equal x y))) (<< y x))))) General Form: a true list of rule class objects as defined below Special Cases: a symbol abbreviating a single rule class object
When defthm
is used to prove a named theorem, rules may be derived from
the proved formula and stored in the database. The user specifies which
kinds of rules are to be built, by providing a list of rule class names
or, more generally, rule class objects, which name the kind of rule to
build and optionally specify varioius attributes of the desired rule.
The rule class names are :
REWRITE
, :
BUILT-IN-CLAUSE
,
:
CLAUSE-PROCESSOR
, :
COMPOUND-RECOGNIZER
,
:
CONGRUENCE
, :
DEFINITION
, :
ELIM
,
:
EQUIVALENCE
, :
FORWARD-CHAINING
, :
GENERALIZE
,
:
INDUCTION
, :
LINEAR
, :
META
,
:
REFINEMENT
, :
TAU-SYSTEM
, :
TYPE-PRESCRIPTION
,
:
TYPE-SET-INVERTER
, and :
WELL-FOUNDED-RELATION
. Some
classes require the user-specification of certain class-specific
attributes. Each class of rule affects the theorem prover's behavior in a
different way, as discussed in the corresponding documentation topic. In
this topic we discuss the various attributes that may be attached to rule
classes.
A rule class object is either one of the :class
keywords or else is a
list of the form shown below. Those fields marked with ``(!)'' are required
when the :class
is as indicated.
(:class :COROLLARY term :TRIGGER-FNS (fn1 ... fnk) ; provided :class = :META (!) :TRIGGER-TERMS (t1 ... tk) ; provided :class = :FORWARD-CHAINING ; or :class = :LINEAR :TYPE-SET n ; provided :class = :TYPE-SET-INVERTER :TYPED-TERM term ; provided :class = :TYPE-PRESCRIPTION :CLIQUE (fn1 ... fnk) ; provided :class = :DEFINITION :CONTROLLER-ALIST alist ; provided :class = :DEFINITION :INSTALL-BODY directive ; provided :class = :DEFINITION :LOOP-STOPPER alist ; provided :class = :REWRITE :PATTERN term ; provided :class = :INDUCTION (!) :CONDITION term ; provided :class = :INDUCTION :SCHEME term ; provided :class = :INDUCTION (!) :MATCH-FREE all-or-once ; provided :class = :REWRITE or :class = :LINEAR or :class = :FORWARD-CHAINING :BACKCHAIN-LIMIT-LST limit ; provided :class = :REWRITE or :class = :META or :class = :LINEAR or :class = :TYPE-PRESCRIPTION :HINTS hints ; provided instrs = nil :INSTRUCTIONS instrs ; provided hints = nil :OTF-FLG flg)When rule class objects are provided by the user, most of the fields are optional and their values are computed in a context sensitive way. When a
:class
keyword is used as a rule class object, all relevant fields are
determined contextually. Each rule class object in :rule-classes
causes
one or more rules to be added to the database. The :class
keywords are
documented individually under the following names. Note that when one of
these names is used as a :class
, it is expected to be in the keyword
package (i.e., the names below should be preceded by a colon but the ACL2
documentation facilities do not permit us to use keywords below).
:clause-processor
rule (goal-level simplifier)
:meta
rule (a hand-written simplifier)
:rewrite
rules (possibly conditional ones)
See also force
, case-split, syntaxp
, and bind-free
for ``pragmas'' one can wrap around individual hypotheses of linear
and
rewrite
rules to affect how the hypothesis is relieved.
Before we get into the discussion of rule classes, let us return to an
important point. In spite of the large variety of rule classes available, at
present we recommend that new ACL2 users rely almost exclusively on
(conditional) rewrite rules. A reasonable but slightly bolder approach is to
use :
type-prescription
and :
forward-chaining
rules for
``type-theoretic'' rules, especially ones whose top-level function symbol is
a common one like true-listp
or consp
; see type-prescription and
see forward-chaining. However, the rest of the rule classes are really not
intended for widespread use, but rather are mainly for experts.
We expect that we will write more about the question of which kind of rule to
use. For now: when in doubt, use a :
rewrite
rule.
:Rule-classes
is an optional keyword argument of the defthm
(and
defaxiom
) event. In the following, let name
be the name of the
event and let thm
be the formula to be proved or added as an axiom.
If :rule-classes
is not specified in a defthm
(or defaxiom
)
event, it is as though what was specified was to make one or more
:
rewrite
rules, i.e., as though :rule-classes
((:rewrite))
had been used. Use :rule-classes
nil
to specify that no rules are to
be generated.
If :rule-classes
class is specified, where class is a non-nil
symbol,
it is as though :rule-classes
((class))
had been used. Thus,
:rule-classes
:
forward-chaining
is equivalent to
:rule-classes
((:forward-chaining))
.
We therefore now consider :rule-classes
as a true list. If any element
of that list is a keyword, replace it by the singleton list containing that
keyword. Thus, :rule-classes
(:rewrite :elim)
is the same as
:rule-classes
((:rewrite) (:elim))
.
Each element of the expanded value of :rule-classes
must be a true list
whose car
is one of the rule class keyword tokens listed above, e.g.,
:
rewrite
, :
elim
, etc., and whose cdr
is a ``keyword
alist'' alternately listing keywords and values. The keywords in this alist
must be taken from those shown below. They may be listed in any order and
most may be omitted, as specified below.
:
Corollary
-- its value,term
, must be a term. If omitted, this field defaults tothm
. The:
corollary
of a rule class object is the formula actually used to justify the rule created and thus determines the form of the rule. Nqthm provided no similar capability: each rule was determined bythm
, the theorem or axiom added. ACL2 permitsthm
to be stated ``elegantly'' and then allows the:
corollary
of a rule class object to specify how that elegant statement is to be interpreted as a rule. For the rule class object to be well-formed, its (defaulted):
corollary
,term
, must follow fromthm
. Unlessterm
follows trivially fromthm
using little more than propositional logic, the formula(implies thm term)
is submitted to the theorem prover and the proof attempt must be successful. During that proof attempt the values of:
hints
,:
instructions
, and:
otf-flg
, as provided in the rule class object, are provided as arguments to the prover. Such auxiliary proofs give the sort of output that one expects from the prover. However, as noted above, corollaries that follow trivially are not submitted to the prover; thus, such corollaries cause no prover output.Note that before
term
is stored, all calls of macros in it are expanded away. See trans.
:
Hints
,:
instructions
,:
otf-flg
-- the values of these fields must satisfy the same restrictions placed on the fields of the same names indefthm
. These values are passed to the recursive call of the prover used to establish that the:
corollary
of the rule class object follows from the theorem or axiomthm
.
:
Type-set
-- this field may be supplied only if the:class
is:
type-set-inverter
. When provided, the value must be a type-set, an integer in a certain range. If not provided, an attempt is made to compute it from the corollary. See type-set-inverter.
:Typed-term
-- this field may be supplied only if the:class
is:
type-prescription
. When provided, the value is the term for which the:
corollary
is a type-prescription lemma. If no:typed-term
is provided in a:
type-prescription
rule class object, we try to compute heuristically an acceptable term. See type-prescription.
:Trigger-terms
-- this field may be supplied only if the:class
is:
forward-chaining
or:
linear
. When provided, the value is a list of terms, each of which is to trigger the attempted application of the rule. If no:trigger-terms
is provided, we attempt to compute heuristically an appropriate set of triggers. See forward-chaining or see linear.
:Trigger-fns
-- this field must (and may only) be supplied if the:class
is:
meta
. Its value must be a list of function symbols (except that a macro alias can stand in for a function symbol; see add-macro-alias). Terms with these symbols trigger the application of the rule. See meta.
:Clique
and:controller-alist
-- these two fields may only be supplied if the:class
is:
definition
. If they are omitted, then ACL2 will attempt to guess them. Suppose the:
corollary
of the rule is(implies hyp (equiv (fn a1 ... an) body))
. The value of the:clique
field should be a true list of function symbols, and if non-nil
must includefn
. These symbols are all the members of the mutually recursive clique containing this definition offn
. That is, a call of any function in:clique
is considered a ``recursive call'' for purposes of the expansion heuristics. The value of the:controller-alist
field should be an alist that maps each function symbol in the:clique
to a list oft
's andnil
's of length equal to the arity of the function. For example, if:clique
consists of just two symbols,fn1
andfn2
, of arities2
and3
respectively, then((fn1 t nil) (fn2 nil t t))
is a legal value of:controller-alist
. The value associated with a function symbol in this alist is a ``mask'' specifying which argument slots of the function ``control'' the recursion for heuristic purposes. Sloppy choice of:clique
or:controller-alist
can result in infinite expansion and stack overflow.
:Install-body
-- this field may only be supplied if the:class
is:
definition
. Its value must bet
,nil
, or the default,:normalize
. A value oft
or:normalize
will cause ACL2 to install this rule as the new body of the function being ``defined'' (fn
in the paragraph just above); hence this definition will be installed for future:expand
hints. Furthermore, if this field is omitted or the value is:normalize
, then this definition will be simplified using the so-called ``normalization'' procedure that is used when processing definitions made withdefun
. You must explicitly specify:install-body nil
in the following cases:fn
(as above) is a member of the value of constant*definition-minimal-theory*
, the arguments are not a list of distinct variables,equiv
(as above) is notequal
, or there are free variables in the hypotheses or right-hand side (see free-variables). However, supplying:install-body nil
will not affect the rewriter's application of the:definition
rule, other than to avoid using the rule to apply:expand
hints. If a definition rule equates(f a1 ... ak)
withbody
but there are hypotheses,hyps
, then:expand
hints will replace terms(f term1 ... termk)
by corresponding terms(if hyps body (hide (f term1 ... termk)))
.
:
Loop-stopper
-- this field may only be supplied if the class is:
rewrite
. Its value must be a list of entries each consisting of two variables followed by a (possibly empty) list of functions, for example((x y binary-+) (u v foo bar))
. It will be used to restrict application of rewrite rules by requiring that the list of instances of the second variables must be ``smaller'' than the list of instances of the first variables in a sense related to the corresponding functions listed; see loop-stopper. The list as a whole is allowed to benil
, indicating that no such restriction shall be made. Note that any such entry that contains a variable not being instantiated, i.e., not occurring on the left side of the rewrite rule, will be ignored. However, for simplicity we merely require that every variable mentioned should appear somewhere in the corresponding:
corollary
formula.
:Pattern
,:Condition
,:Scheme
-- the first and last of these fields must (and may only) be supplied if the class is:
induction
.:Condition
is optional but may only be supplied if the class is:
induction
. The values must all be terms and indicate, respectively, the pattern to which a new induction scheme is to be attached, the condition under which the suggestion is to be made, and a term which suggests the new scheme. See induction.
:Match-free
-- this field must be:all
or:once
and may be supplied only if the:class
is either:
rewrite
,:
linear
, or:
forward-chaining
. (This field is not implemented for other rule classes, including the:
type-prescription
rule class.) See free-variables for a description of this field. Note: Although this field is intended to be used for controlling retries of matching free variables in hypotheses, it is legal to supply it even if there are no such free variables. This can simplify the automated generation of rules, but note that when:match-free
is supplied, the warning otherwise provided for the presence of free variables in hypotheses will be suppressed.
:Backchain-limit-lst
-- this field may be supplied only if the:class
is either:
rewrite
,:
meta
,:
linear
, or:
type-prescription
. It is further required either only one rule is generated from the formula or, at least, every such rule has the same list of hypotheses. The value for:backchain-limit-lst
must benil
; a non-negative integer; or, except in the case of:
meta
rules, a true list each element of which is eithernil
or a non-negative integer. If it is a list, its length must be equal to the number of hypotheses of the rule and each item in the list is the ``backchain limit'' associated with the corresponding hypothesis. Ifbackchain-limit-lst
is a non-negative integer, it is defaulted to a list of the appropriate number of repetitions of that integer. The backchain limit of a hypothesis is used to limit the effort that ACL2 will expend when relieving the hypothesis. If it isNIL
, no new limits are imposed; if it is an integer, the hypothesis will be limited to backchaining at most that many times. Note that backchaining may be further limited by a globalbackchain-limit
; see backchain-limit for details. For different ways to reign in the rewriter, see rewrite-stack-limit and see set-prover-step-limit. Jared Davis has pointed out that you can set the:backchain-limit-lst
to 0 to avoid any attempt to relieveforce
d hypotheses, which can lead to a significant speed-up in some cases.
Once thm
has been proved (in the case of defthm
) and each rule
class object has been checked for well-formedness (which might require
additional proofs), we consider each rule class object in turn to generate
and add rules. Let :class
be the class keyword token of the i
th
class object (counting from left to right). Generate the rune
(:class name . x)
, where x
is nil
if there is only one class and
otherwise x
is i
. Then, from the :
corollary
of that
object, generate one or more rules, each of which has the name
(:class name . x)
. See the :
doc
entry for each rule class to
see how formulas determine rules. Note that it is in principle possible for
several rules to share the same name; it happens whenever a
:
corollary
determines more than one rule. This in fact only occurs
for :
rewrite
, :
linear
, and :
forward-chaining
class rules and only then if the :
corollary
is essentially a
conjunction. (See the documentation for rewrite, linear, or
forward-chaining for details.)