PROGRAMMING-WITH-STATE

programming using the von Neumannesque ACL2 state object
Major Section:  STATE

This documentation section introduces some common techniques for programming using the ACL2 state object. A prerequisite is thus a basic understanding of that object; see state. We hope this section is useful, and we invite suggestions for improvements and additions.

A supplement to this section is the ACL2 source code, which uses most (and probably all) of the techniques discussed here. That code is thus a source of many examples, which can serve as ``templates'' to guide one's own programming with state.

Recall that ``ACL2'' stands for ``A Computational Logic for Applicative Common Lisp''. In particular, the language is applicative: there are no global variables or side effects. For many purposes this does not feel restrictive; for example, an ACL2 user who is programming in raw Lisp may well be more comfortable coding a factorial function applicatively, using recursion, rather than using iteration with repeated assignment to the same variable.

However, there are situations that call for reading or modifying the system state, such as performing input and output, signalling errors, saving information from one computation for use in a later one, or reading and updating system-level or environmental data. This section provides an introductory guide for writing functions that traffic in state. We emphasize that this guide is intended as an introduction; more complete documentation may often be found by following links to documentation of individual utilities, and again, more examples may be found by searching the ACL2 source code for uses of the functions and macros mentioned below. The rest of this section is organized as follows.

ENABLING PROGRAMMING WITH STATE
STATE GLOBALS
A REMARK ON GUARDS
ERRORS AND ERROR TRIPLES
SEQUENTIAL PROGRAMMING
BINDING VARIABLES USING ERROR TRIPLES
BINDING STATE GLOBAL VARIABLES
INPUT AND OUTPUT
TIMINGS
ENVIRONMENT AND SYSTEM
REMARKS ON EVENTS AND LD
ADVANCED TOPICS

ENABLING PROGRAMMING WITH STATE

In order to submit a definition that takes state as a formal parameter, you must either declare state as a :stobj (see xargs) or first evaluate the following form at the top level: (set-state-ok t).

Consider for example the following trivial definition.

(defun foo (state)
  (mv 3 state))
If you submit the above definition in a fresh ACL2 session, you will get this error message.
  ACL2 Error in ( DEFUN FOO ...):  The variable symbol STATE should not
  be used as a formal parameter of a defined function unless you are
  aware of its unusual status and the restrictions enforced on its use.
  See :DOC set-state-ok.
If first you evaluate (set-state-ok t), you can admit the above definition. Alternatively, you can declare state as a :stobj, as follows.
(defun foo (state)
  (declare (xargs :stobjs state))
  (mv 3 state))
A difference in the two approaches is that for the latter, a guard proof obligation is generated by default. See the section below entitled ``A remark on guards''.

STATE GLOBALS

Recall (see state) that one of the fields of the ACL2 state object is the global-table, which logically is an alist associating symbols, known as ``state globals'' or ``state global variables'', with values. But no such alist actually exists in the implementation. Instead, ACL2 provides utilities for reading state globals -- see @ and see f-get-global -- and utilities for writing them -- see assign and see f-put-global. The following log shows how they work; further explanation follows below.

ACL2 !>(assign my-var (+ 3 4))
 7
ACL2 !>(@ my-var)
7
ACL2 !>(f-put-global 'my-var (+ 1 5) state)
<state>
ACL2 !>(f-get-global 'my-var state)
6
ACL2 !>
Note that the first result is indented by one space. This is ACL2's way to indicate that the assign expression returned an ``error triple'' and that no error was signalled. We discuss error triples in more detail below.

As illustrated above, the output signatures of the utilities for assigning to state globals differ from each other as follows: f-put-global returns state, but assign returns an error triple (mv nil val state) where val is the value assigned to the state global. The output signatures of the utilities for reading, @ and f-get-global, are identical. In fact, the form (f-get-global 'my-var state) is the single-step macroexpansion of the form (@ my-var), as can be confirmed using trans1.

ACL2 !>:trans1 (@ my-var)
 (F-GET-GLOBAL 'MY-VAR STATE)
ACL2 !>

State globals are useful for conveying persistent state information. Consider for example the utility set-inhibit-output-lst. The form (set-inhibit-output-lst '(prove proof-tree)) is approximately equivalent to (assign inhibit-output-lst '(prove proof-tree)). We say ``approximately'' because set-inhibit-output-lst additionally does some error checking to insure that all the tokens in the new list are legal. When deciding whether to print output, the ACL2 system reads the value of state global variable inhibit-output-lst.

A REMARK ON GUARDS

For a function definition (see defun), if state is specified as a stobj as with the form (declare (xargs :stobjs state)), then the guard for that function is considered to include the condition (state-p state). By default, guard verification will then be performed.

We can illustrate this point by modifying the example above as follows, to read the value of state global gag-mode.

(defun foo (state)
  (declare (xargs :stobjs state))
  (f-get-global 'gag-mode state))
If you try this in a fresh ACL2 session, the proof will fail with the following key checkpoint, which says that the state global gag-mode is bound in the global-table of the state.
  (IMPLIES (STATE-P1 STATE)
           (ASSOC-EQUAL 'GAG-MODE (NTH 2 STATE)))

How can we deal with this proof failure? One way is simply to ignore the issue by defining the function in :program mode, as follows.

(defun foo (state)
  (declare (xargs :stobjs state
                  :mode :program))
  (f-get-global 'gag-mode state))
Perhaps a better way is to strengthen the guard to assert that the indicated state global is bound, as follows.
(defun foo (state)
  (declare (xargs :guard (boundp-global 'gag-mode state)
                  :stobjs state))
  (f-get-global 'gag-mode state))
Also see guard-miscellany for a discussion of how guards are generated from xargs fields of declare forms, specifically, for keywords :guard and :stobjs.

ERRORS AND ERROR TRIPLES

When evaluation returns three values, where the first two are ordinary objects and the third is the ACL2 state, the result may be called an ``error triple''. (Whether it is treated as an error triple depends on the programmer.) Error triples are often denoted (mv erp val state), and common ACL2 programming idioms treat erp as a flag indicating whether an error is being signalled and val as the ``value'' computed.

Even ACL2 users who are not programmers encounter error triples, because these are the values returned by evaluation of ACL2 events. Consider the following log, where the only user input is the defun form following the prompt.

ACL2 !>(defun foo (x) x)

Since FOO is non-recursive, its admission is trivial.  We observe that
the type of FOO is described by the theorem (EQUAL (FOO X) X).  

Summary
Form:  ( DEFUN FOO ...)
Rules: NIL
Time:  0.00 seconds (prove: 0.00, print: 0.00, other: 0.00)
 FOO
ACL2 !>
All output above results from explicit calls of output functions, except for the next-to-last line, which contains FOO. Notice the single-space indentation preceding FOO. That space indicates that in fact, the value returned by evaluation of the defun form is the error triple whose error flag is nil and whose computed value is FOO. By default, ACL2 prints any error triple (mv nil val state) by inserting a space before printing val. You can change the default by setting state global ld-post-eval-print to t; notice how the same result is printed below.
ACL2 !>:u
          0:x(EXIT-BOOT-STRAP-MODE)
ACL2 !>(set-ld-post-eval-print t state)
(NIL T <state>)
ACL2 !>(defun foo (x) x)

Since FOO is non-recursive, its admission is trivial.  We observe that
the type of FOO is described by the theorem (EQUAL (FOO X) X).  

Summary
Form:  ( DEFUN FOO ...)
Rules: NIL
Time:  0.00 seconds (prove: 0.00, print: 0.00, other: 0.00)
(NIL FOO <state>)
ACL2 !>

The way error triples are printed by ld is controlled not only by state global ld-post-eval-print, but also by state global ld-error-triples. These are examples of ``ld specials''; see ld, see ld-post-eval-print, and see ld-error-triples.

It is so common to produce an error triple whose first (error flag) component is nil that ACL2 provides a handy macro, value, for this purpose. Thus, (value <expression>) is equivalent to (mv nil <expression> state). Also see value-triple for a similar construct that is a legal event form.

We turn now to the topic of errors. The macro ER ``causes'' an error, but there are really two quite different kinds of errors: ``soft'' and ``hard'' errors. We use the term ``soft error'' to refer to a form that returns an error triple (mv erp val state) for which erp is non-nil. Soft errors do not interrupt the normal flow of evaluation: the error triple is returned to the caller which interprets the erp flag and val as directed by the programmer. Macros discussed below make it convenient to think about soft errors as short-circuiting the computation. Hard errors, on the other hand, do actually rip control away from the current evaluation and return it to the top-level loop. Logically speaking, expressions that cause hard errors return nil in the error case, but the nil is never seen in actual evaluation because control does not return to the caller.

Note that the function abort!, which you can invoke by typing :a!, always returns to the top level. Note that ACL2 can prove that (abort!) returns nil but that this cannot be confirmed by computation.

ACL2 !>(thm (equal (abort!) nil))

Q.E.D.

Summary
Form:  ( THM ...)
Rules: ((:FAKE-RUNE-FOR-TYPE-SET NIL)
        (:TYPE-PRESCRIPTION ABORT!))
Time:  0.00 seconds (prove: 0.00, print: 0.00, other: 0.00)

Proof succeeded.
ACL2 !>(equal (abort!) nil)
Abort to ACL2 top-level
...
ACL2 !>

(What actually happens with a hard error, including non-default cases, is a bit subtle; most readers will probably want to skip this paragraph. The read-eval-print loop implemented by ld is implemented by a call of the ACL2 evaluator function, trans-eval, on each input form. If a hard error occurs during evaluation of an input form, its trans-eval call will return with a soft error. Ld, in turn handles that soft error appropriately; see ld-error-action.)

The most common way to signal errors is the macro er, which prints a formatted error message and returns a soft or hard error as specified by the call. Note however that soft errors are signalled using :program mode functions.

Since the output signatures of soft and hard errors are different -- hard errors ``return'' a single value while soft errors return a triple -- mixing them in an expression requires embedding the hard error form in (an irrelevant) triple, as illustrated below. All branches of the expression must produce an error triple if any branch does.

ACL2 !>(defun chk-find-or-abort (e x state)
         (declare (xargs :mode :program))
         (if (endp x)
             (value                          ; Note use of VALUE!
               (er hard 'chk-find-or-abort
                   "Did not find ~x0!"
                    e))
             (if (not (integerp (car x)))
                 (er soft 'chk-find-or-abort
                     "Non-integer, ~x0, in list!"
                     (car x))
                 (if (eql (car x) e)
                     (value x)
                     (chk-find-or-abort e (cdr x) state)))))

Summary
Form:  ( DEFUN CHK-FIND-OR-ABORT ...)
Rules: NIL
Time:  0.00 seconds (prove: 0.00, print: 0.00, other: 0.00)
 CHK-FIND-OR-ABORT
ACL2 !>(chk-find-or-abort 3 '(1 2 3 4 5) state)
 (3 4 5)
ACL2 !>(chk-find-or-abort 3 '(1 A 3 4 5) state)


ACL2 Error in CHK-FIND-OR-ABORT:  Non-integer, A, in list!

ACL2 !>(chk-find-or-abort 3 '(1 2 4 5) state)

HARD ACL2 ERROR in CHK-FIND-OR-ABORT:  Did not find 3!
...
ACL2 !>

See er for further discussion of errors. For some other individual topics related to errors see assert$, see break-on-error, see error1, see hard-error, see illegal, and see ld-error-triples.

In the next section we discuss soft errors further, in the context of programming.

SEQUENTIAL PROGRAMMING

This section describes handy ways to modify state in steps, using macros that implement a sequence of let or mv-let bindings. For example, suppose you want to assign the values 1 and 2 to two state globals one-var and two-var, respectively. Because of ACL2's syntactic restrictions on state, it is not legal simply to write (f-put-global 'two-var 2 (f-put-global 'one-var 1 state)). However, let comes to the rescue as follows.

(let ((state (f-put-global 'one-var 1 state)))
  (let ((state (f-put-global 'two-var 2 state)))
    state))
It is so common to bind state successively in such a manner that ACL2 provides a macro, pprogn, for this purpose. Thus, an equivalent solution to the problem above is
(pprogn (f-put-global 'one-var 1 state)
        (f-put-global 'two-var 2 state)
        state)
or, more simply, as follows.
(pprogn (f-put-global 'one-var 1 state)
        (f-put-global 'two-var 2 state))
See pprogn. Note that the last form is allowed to return multiple values; the only requirement on the last form is that its value include state.

It is also common to update the state using a sequence of forms such that each returns an error triple, where the intention is for evaluation to short-circuit immediately if a soft error is encountered. Suppose <expr1> and <expr2> are expressions that return error triples, where the state components of the error triples might be updated, and one wishes to evaluate <expr1> and then <expr2>, returning the (multiple) values returned by <expr2> unless the error triple returned by <expr1> is a soft error, in which case that error triple is returned. One can of course do so as follows.

(mv-let (erp val state)
        <expr1>
        (cond (erp (mv erp val state))
              (t <expr2>)))
But ACL2 provides a handy macro, er-progn, for this purpose. The following code is equivalent to the code just above.
(er-progn <expr1> <expr2>)
See er-progn for more details. Note that unlike pprogn, the return signature for the last expression must be the same as that of the others: an error triple.

Let's consider how to use pprogn and er-progn together. In the following example f1 and f2 both return state, while each of g1 and g2 returns an error triple. The following code modifies state by executing these in the order f1, g1, f2, and finally g2, returning (mv nil val state) where val is the value component of the error triple returned by g2 -- except we return a soft error if g1 or g2 returns a soft error.

(pprogn (f1 x state)
        (er-progn (g1 x state)
                  (pprogn (f2 x state)
                          (g2 x state))))

Finally, consider the events progn and progn!. These have similar behavior to that of er-progn. However, progn and progn! may only be used in event contexts, for example at the top level or immediately underneath a call of encapsulate or progn, while er-progn has no such restriction. So when writing code, use er-progn rather than progn or progn!. In particular, the body of a defun must not have any calls of progn (or of progn! either), and the same restriction holds for any code to be executed, such as the body of a make-event form.

BINDING VARIABLES USING ERROR TRIPLES

In this section we discuss the macro er-let*, which is a variant of the special form, let*, that is useful when programming with state.

The macro er-let* is useful when binding variables to the value components of error triples. It is actually quite similar to er-progn, described above, except that er-let* binds variables. First consider the following example.

(er-let* ((x1 (f1 state))
          (x2 (f2 x1 state)))
  (value (cons x1 x2)))
The code just above is essentially equivalent to writing the following.
(mv-let (erp x1 state)
        (f1 state)
        (cond (erp (mv erp x1 state))
              (t (mv-let (erp x2 state)
                         (f2 x1 state)
                         (cond (erp (mv erp x2 state))
                               (t (value (cons x1 x2))))))))

As suggested by the example above, er-let* has the same syntax as let*, except that declarations are not supported. (But note that ignore declarations are not needed; all variables that are bound are also used, at least in the error case. Consider replacing (cons x1 x2) by nil in the example displayed immediately above, and note that x1 and x2 are still used.) However, unlike let*, er-let* requires that for each binding (var expr), the expression expr must evaluate to an error triple and, moreover, it requires that the second argument (the ``body'') of er-let* must evaluate to an error triple. If one of the variable expressions (e.g., the f1 and f2 calls above) signals an error, its error triple is returned as the value of the er-let*.

Of course, soft errors can be ``caught'' by using mv-let instead of er-let* and simply ignoring the error flag or, more generally, by returning a non-erroneous error triple even if the error flag was on.

BINDING STATE GLOBAL VARIABLES

In this section we introduce a utility, state-global-let*, that is an analogue of let* for state global variables. Consider the following example.

(state-global-let*
 ((inhibit-output-lst (add-to-set-eq 'summary (@ inhibit-output-lst))))
 (thm (equal x x)))
This form binds state global variable inhibit-output-lst to the result of adding the symbol, summary, to the current value of that state global. Thus (see set-inhibit-output-lst), the usual summary is not printed when evaluating this call of thm.

See state-global-let* for more complete documentation.

INPUT AND OUTPUT

In ACL2, most input and output involves the ACL2 state. See io.

TIMINGS

For how to obtain the runtime elapsed since the start of the ACL2 session, see read-run-time.

For a utility for saving times into the ACL2 state and for printing those saved times, see the distributed book misc/save-time.lisp.

To time an evaluation (though this really isn't about state), see time$.

ENVIRONMENT AND SYSTEM

Next, we mention briefly some ways in which ACL2 interacts with its environment using the ACL2 state.

For how to read and write environment variables, see getenv$ and see setenv$.

For how to run a command in the host operating system, see sys-call.

REMARKS ON EVENTS AND LD

In general, undefined or surprising behavior may occur when using ACL2 events or calling ld in your programs. In some cases ACL2 enforces restrictions against these uses. We strongly discourage using ld in programs, as it has been designed to be called only at the top level of a read-eval-print loop.

There is also a restriction on contexts in which make-event may be called: it may only be called in a context where an event is expected, such as the top level, in a book, or as an argument of encapsulate or progn. The reason is that ACL2 does very subtle and careful tracking of make-event expansions; and it is only able to do this in event contexts, where it is able to carry out such tracking accurately.

ADVANCED TOPICS

ACL2 provides the function trans-eval to evaluate an arbitrary form (after translating it to a term, i.e., into internal form). For more information, we refer to reader to comments in the definition of trans-eval in the ACL2 source code. There are also many examples of its use in the ACL2 sources.

For a function that provides the true absolute filename, with soft links resolved, see canonical-pathname.

For a function that returns a check-sum on the characters in a channel, see check-sum.

To obtain a random number, see random$.

If you are programming in raw-mode (see set-raw-mode) or in raw Lisp, use the variable *the-live-state* in place of the variable state.

We invite suggestions for additional advanced topics.