This paper appears in the proceedings for the
First European Conference on the Practical Application of Lisp (EUROPAL'90),
Churchill College, Cambridge, UK, March 27-29, 1990.
It won a "Best Paper" award (Tools & Techniques category).
The substance of the original text appears in normal fonting,
though a small number of out-and-out typos were corrected.
Oddities of spelling that were the custom of the time
(either generally, or just for me) were left alone.
Some formatting of headings and tables was adjusted slightly for HTML.
Any new text that has been added appears
bracketed and in color green; such text is intended to
help clarify the historical context, since considerable time
passed between the time this paper was published and the time I
converted it to HTML.
--Kent Pitman, 17-Oct-1998.
Annotated original document follows.
Click here for an index of other titles by Kent Pitman.
Kent M. Pitman
[address at time of publication]
8 New England Executive Park, East
Burlington, MA 01803-5007 U.S.A.
It is useful to partition the description of a program's behavior into two parts--what happens in normal situations, and what happens in exceptional situations. This paper surveys the issues involved in the description of program behavior during exceptional situations.
Keywords: Conditions, Errors, Exceptions, Handling, Lisp, Signalling.
Copyright © 1990, 1985 Kent M. Pitman.
All Rights Reserved.
[HTML document Copyright © 1999 Kent M. Pitman. All Rights Reserved.]
During a computation, functions communicate with one another by passing arguments, by returning values, and by side-effect upon shared structures. Consider the following definitions:
|(+ n1 n2)||returns the sum of its arguments, n1 and n2.|
|returns the next character on the input stream, instream. This function performs an implicit side-effect upon instream, typically incrementing some internal scan pointer, in order to assure that subsequent attempts to read from the same stream do not try to re-read the same character.|
The descriptions of + and READ-CHAR given above describe their behavior in what we shall call the normal situation. Sometimes, however, an exceptional situation will arise which does not fit neatly into the normal description of a function. For example, + might receive an argument which is not a number, or READ-CHAR might receive a stream which has no more available characters.
This paper examines the phenomenon of exceptional situations and the programming techniques which can be employed to deal with such situations.
It is important to recognize that this distinction between normal and exceptional situations is in some sense contrived. Making this distinction does not change the way programs behave; it simply changes the way we reason about programs--hopefully for the better.
We normally think of + as a function which adds two numbers, but imagine for a moment that when it receives a non-numeric argument, it just returns false. We could define it to be "normal" for + to return "either the sum of its two arguments if they are numbers or false if they are not." This wouldn't change what + did, only how we thought about it. Rather than just being a function which returned a number, + would become a function which returned either a number or NIL. In general, any program which could potentially be described in terms of normal and exceptional situations can also be described with only a normal description by simply taking some union of those normal and exceptional descriptions and declaring that union to be normal.
However, if we were to claim that it was "normal" to use + with a non-numeric argument, then we would have to tolerate as normal a usage such as
(+ 'A 'B)
where a programmer called + explicitly with non-numeric because he planned to take advantage of the fact that it would yield false. (Perhaps the programmer doesn't know that the variable NIL holds the false object and this expression encodes his desire to produce NIL as a value.) Somehow we want to discourage the programmer from using + in this fashion, and one way to do so is simply to describe such a use as "not normal."
In the rest of this section, we will survey some of the specific advantages that come from making this admittedly contrived distinction between normal and exceptional situations.
One reason for distinguishing between normal and exceptional situations shows itself in the use of simplifying assumptions that typically distinguish a prototype from a production quality program. A programmer may speed the development of a program by assuming that the program will be needed only for some normal case. For example, the tokenizer for a parser might be initially coded to simply read characters from a stream until a token break, without handing the exceptional situation that occurs at the end of file when there is no next character. In so doing, the programmer pushes off worrying about a level of detail that would only be distracting during the prototyping phase. Later, after it has been shown that the program works satisfactorily in normal situations, the programmer can "tighten up" the program to handle exceptional situations. This idea is explored in detail in [Rich 81] and [Lieberman 82].
Simplifying assumptions can also be important after a program has been developed--when it is being read by another programmer. The person reading the code may want to first understand what it does in the normal situation and then later refine his understanding of it by inspecting its behavior in exceptional situations. If the code for dealing with the normal situation is thickly interspersed with code attempting to recognize and deal with exceptional situations, the reader's ability to decipher the intent of the program may be greatly diminished.
This problem may be dealt with in at least two ways. One way would be to provide constructs within the language which encourage the separation of code to deal with normal and exceptional situations. Another way would be to provide editing technology which allowed the code for the exceptional situations to be hidden while the normal situation is examined on its own. Even in the case where editing technology is used, it helps significantly if there is language-level support for this separation to serve as a guide for the editing tools.
The idea that the language should encourage separation of normal and exceptional situations has appeal for another reason. It frequently happens that the same exceptional situation can occur in many different places in a program; and, in many cases, the way it should be handled is the same throughout the program.
For example, consider a program which is compiling a file from a remote machine. If the remote machine crashes, the compilation will probably want to simply fail no matter what routine was running. Languages should provide some way of centralizing the information about a program's behavior in such an exceptional situation so that it doesn't have to be needlessly repeated throughout the code for the program.
In languages which have no built-in support for condition handling, information must be propagated back from every program to indicate success or failure. The constant need to deal explicitly with error codes is at best tedious and distracting--and at worst very dangerous. Some programmers are prone to get sloppy about checking such codes in all cases where they should, and this can lead to unreliable connectivity between programs.
In some cases, there may be efficiency reasons for considering some cases to be exceptional. For example, the compiled code for + will be less efficient if it must contain explicit checks for its arguments being non-numbers. In Maclisp [Pitman 83] , for example, the interpreted + function will "signal an error" if invoked with bad arguments, but compiled calls to + will just "do the wrong thing" if that situation arises .
It would seem a useful goal for a compiler to be permitted to do this sort of optimization, but a compiler cannot be permitted to sacrifice proper semantics for efficiency.
To avoid any chance that the actions of a good compiler might be seen as unpredictable or irrational, we must make it a goal to describe functions in such a way that compiler optimizations do not violate programmer expectations. By distinguishing between normal and exceptional situations, we can accomplish this goal. The definition of + is potentially compromised only in the exceptional situation--its semantics will always be maintained for the normal situation, even in highly optimized code. If we had not made this distinction, we would have to either make + do extra work in compiled code or describe the process of compiling + as "unreliable."
When an exceptional situation arises in a program, there are a number of possible actions that can be taken.
Our description of exceptional situations is general enough to include situations which are not normal but which the program does not recognize as such. The example of how compiled calls to + behave in Maclisp is an example of this behavior. Hence, one possible way in which programs can deal with exceptional situations is to fail to recognize them.
While such failure might not be something to encourage as a programming style, it is important to recognize that it does occur--frequently, in fact. It can happen any time an assumption of any kind is made by code but not mechanically verified (either at compile time or runtime).
In some situations, the type of the return value is highly constrained. For example, since we know that many operations on numbers yield only other numbers, we could use non-numeric return values from such functions to convey other kinds of information. Our earlier suggestion that + might return false when passed non-numeric arguments is an example of this technique. While this technique is not totally general, it is clearly useful in a statistically interesting number of cases.
A slightly more complicated variation of this technique arises with the Maclisp ERRSET primitive. ERRSET evaluates its argument, "trapping" any errors. The designers of ERRSET wanted to return the result of the computation if there was no error, but also to be able to detect whether there was an error or not. Since evaluation could return any single object, their challenge was to devise a return value convention which expressed both of these pieces of information unambiguously. Their solution was to return NIL if an error had occurred or a list the first element of which was the result of the evaluation if no error occurred.
(ERRSET (OPEN "NOSUCHFILE.TXT")) -> NIL (ERRSET (OPEN "FILE.TXT")) -> (#FILE-IN-|FILE.TXT|-70766)
The solution to this "ERRSET problem" in Zetalisp [Weinreb 81] is somewhat more graceful because a function can simply return multiple values. Zetalisp provides a macro called IGNORE-ERRORS which evaluates its argument and returns the result (or NIL if an error occurrs) as its "first return value." Since the case of NIL being returned in this position is ambiguous (the form might have returned NIL or an error might have occurred), it also returns a "second return value" which is NIL if no error occurred or non-NIL otherwise.
(IGNORE-ERRORS (OPEN "NOSUCHFILE.TXT")) -> NIL, T (IGNORE-ERRORS (OPEN "FILE.TXT")) -> #<INPUT-STREAM "FILE.TXT" 3370754>, NIL
Especially in Scheme [Steele 78], a popular way to deal with exceptional situations is simply to pass multiple continuations [Steele 76].
For example, we could imagine that our earlier case of READ-CHAR could be reformulated by making READ-CHAR take additional arguments of functions to be called in the case that either there is or is not a "next character." Consider a hypothetical Scheme definition:
(DEFINE (READ-CHAR STREAM SUCCESS FAIL) (IF (STREAM-EMPTY? STREAM) (FAIL) (SUCCESS (STREAM-FIRST STREAM) (STREAM-REST STREAM))))
which would allow the caller to write:
(DEFINE (VIEW-STREAM STREAM) (READ-CHAR STREAM (LAMBDA (FIRST-CHAR REST-OF-STREAM) (WRITE-CHAR FIRST-CHAR) (VIEW-STREAM REST-OF-STREAM)) (LAMBDA () 'DONE)))
In this case, the normal and exceptional situations are factored out into separate functions so that they can be studied independently.
Sometimes a convention may have been established between a group of programs so that when certain exceptional situations are detected, control is transferred to a dynamically established point in the call chain. Such a non-local transfer of control is generally accomplished by the THROW primitive in Lisp.
Also in this category are the mechanisms some languages provide for aborting and restarting computations. These are typically just "syntactic sugar" for the same kind of non-local transfer of control already provided by THROW. For example, the Maclisp ^G  function "throws" to a point which restarts the toplevel Read-Eval-Print loop.
In modern Lisp dialects, we encourage the use of DEFSTRUCT and DEFCLASS and their associated accessors and constructors rather than CAR, CDR, and CONS because they allow us to get away from thinking about how objects are implemented and enable us to speak directly in terms of more abstract properties of objects. As we have seen, Lisp primitively provides many ways to implement treatment of exceptional situations, but it seems equally important that we develop proper abstractions for describing and reasoning about exceptions.
When a program detects an exceptional situation, it will typically want to "announce" the presence of the situation, allowing any special code which has been set up for dealing with that situation to be run. We will refer to this process of announcing an exceptional situation as signalling and to the piece of code which announces the situation as the signaller.
The class of exceptional situations which are recognized and signalled will henceforth be referred to as conditions.
We will think of signalling as a way for a program to "stop" and request advice on which of several ways to "restart." These possible ways to restart will be called restart options .
The pieces of code which are asked to select a restart option will be called handlers. If a handler performs a non-local transfer of control (beyond the point of the signal), it will be said to have handled the situation. If it does not, it will be said to have declined.
Some conditions are non-fatal in that the failure to handle them will not affect the correct behavior of the program if it simply continues on. Some common examples of places where non-fatal conditions might want to be signalled are "end of line" and "end of page" exceptions on output devices, the entry and exit of major and minor modes in an editor, the successful completion of subtasks in a multi-task project, or if-needed and if-used situations in a knowledge-representation system. These non-fatal conditions are sometimes called "hooks" because there is a well-defined behavior that can occur if no advice is supplied, but they offer a place to "hang" advice which will augment or override the default behavior.
Some situations may be considered fatal in the sense that code should not restart beyond the point where they are signalled unless the condition is "corrected" by some form of external intervention. We will call such a fatal situation an error situation, or error.
Error situations, sometimes called "bugs," may be divided into two classes: errors which are detected and signalled, which we shall call error conditions; and errors that go unsignalled (usually only to cause confusion later in the program's execution).
Common Lisp [Steele 84] is careful to distinguish between two kinds of error situations by using the terminology "is an error" when an error might be signalled (but is not guaranteed to be) and "signals an error" to refer to situations where an error is guaranteed to be signalled. These terms correspond fairly directly to our terms "error situation" and "error condition."
We now turn to the question of what kind of information needs to accompany a signal. It may help here to appeal to a fairly concrete example.
Let us imagine a situation in the control program for a robot butler which is about to put food on the table where the robot notices that the eggs it is about to serve are green. We might imagine that the signaller's code would look something like:
To Serve a Tray-of-Food: For each Food in the Tray-of-Food, If the Food's Color is not the Food's Expected-Color, Signal a Bad-Food-Color condition specifying the Food, its Color, and its Expected-Color. Carry the Tray-of-Food to the Dining-Room. Place the Tray-of-Food on the Table.
Unfortunately, the "code" does not provide enough information for a potential handler to advise it about how to continue beyond the point where the error is signalled. Since the handler may not have access to the data structures necessary for correcting the problem and since the signaller has not provided "code" to implement any corrections, multiple correction strategies will not in general be possible.
For this reason, when a condition is signalled, we should remember that two kinds of information may need to be provided: a description of the condition and a description of what, if anything, the signaller is prepared to do in order to restart the stopped computation.
In the case that restarting is possible, the above code does not suffice. What would be needed might look more like:
To Serve a Tray-of-Food: For each Food in the Tray-of-Food If the Food's Color is not the Food's Expected-Color, Signal a Bad-Food-Color condition specifying the Food, its Color, and its Expected-Color and heeding advice on which of the following ways to restart: To Add-Food-Coloring of a given Color: Put Food-Coloring of the given Color into the Food; To Serve-Food-Anyway: Continue; To Throw-Food-Away: Remove the Food from the Tray-of-Food. Carry the Tray-of-Food to the Dining-Room. Place the Tray-of-Food on the Table.
When a condition is signalled, handlers may exist which could potentially handle the error. We will defer for the moment the question of how such handlers are located and for now concern ourselves with what those handlers will want to do once they are located.
Informally, we may think of a handler as analogous to a piece of sage advice. But like any advice, there may be times when it is appropriate and times when it is not. Some of those times may be partly determined by issues of scope which will be discussed later; others will be determined by inspecting the description of the condition and the options the signaller has suggested are available.
Some appropriate means will have to be provided for inspecting the condition. In Maclisp, the means provided  is a function to examine the condition state at the current (or a given) stack frame. This made it quite difficult to test condition handlers out of context. Experience with Zetalisp suggests that an object-based approach (where the handler receives an argument representing the condition's description and is provided with a means of inspecting that object) is more flexible.
After inspecting the description, the handler might want to do any of a number of things. For example, to continue our earlier robot scenario:
These items involve giving advice about how to restart. In order to give such advice, the handler would have to have verified that the particular restart option it was suggesting was acceptable to the signaller.
Regardless of the signaller's situation, the handler may always decline to handle the condition.
Also regardless of the situation, the handler may always handle the situation by bypassing it entirely. For example:
The demolition of the building containing our hypothetical food problem is analogous to what happens during a non-local transfer of control from a stack frame that is signalling an error. The situation is resolved by obliterating some outer context in which the situation arose in the first place.
Some advice may want to be given provisionally. For example, the handler might want to say: "If you can find nothing better to guide you, I have some advice. However, I would prefer that you ask around for other advice first and come back to me only as a last resort."
In Zetalisp, the CONDITION-BIND-DEFAULT macro addresses this issue, but forces the decision about whether advice is to be provisional or not to be decided at coding time (when the user must select between CONDITION-BIND and CONDITION-BIND-DEFAULT) rather than allowing the user to defer this decision until runtime.
In any system where default handlers may be established, there is the question of how to resolve the situation where two defaults have been asserted. In Zetalisp, when a signal goes unhandled by normal bound handlers (set up by CONDITION-BIND), default handlers (set up by CONDITION-BIND-DEFAULT) are searched from the inside out (along the dynamic call chain) until a handler is found .
The emerging standard for ANSI Common Lisp permits the programmer to resignal a condition from within a handler. This power enables one to achieve the power of Zetalisp's CONDITION-BIND-DEFAULT using a form like :
(HANDLER-BIND ((type1 #'(LAMBDA (C) (SIGNAL C) ... default advice ... ))) form1 form2 ...)
In the above scenario, the use of SIGNAL effectively guards the default advice: A recursive signal of the condition occurs. If the recursive signal is handled, then the default advice will never be reached. If the recursive signal is not handled, the call to SIGNAL returns and the default advice is executed. Unlike the Zetalisp approach, the new Common Lisp approach will permit the decision about whether the advice is "default" to be made dynamically at program-execution time rather than statically at program-writing time.
When a condition handler is trying to determine whether its advice would be applicable in a given situation, it can be considerably helped by the availability of a class hierarchy for condition types. For example, a general piece of advice appropriate for a "file" error can be recognized as appropriate for a more specific "file not found" error using such a mechanism.
Experience with the Zetalisp condition system has shown that this is a situation where the availability of multiple superclasses for condition types is nearly essential. Some errors do not fit neatly into a strict hierarchy; for example, if an "end of file" error is signalled from within the READ function, is this a "parse" error or a "stream" error? If it is possible for a condition type to have multiple superclasses, an "end of file" error can inherit from both of the more general types "parse" error and "stream" error. This feature allows handlers a great deal of flexibility in deciding which conditions they are willing to advise.
Some advice may be more specific than others. For example, suppose that the handler for our Bad-Food-Color condition wants to suggest that the food color to be changed and the signaller is prepared to Add-Food-Coloring. There might want to be some notion of a class hierarchy for restart options such that Add-Food-Coloring is recognized as an appropriate substitute for Change-Color .
When a condition is signalled, there are some questions about how an appropriate handler should be located.
The first decision to be made is whether handlers for conditions are to be "globally assigned" or "locally bound."
The trend in programming seems strongly in the direction of eliminating any notion of "global" assignment in favor of operators that bind things. Hence, recent Lisp dialects are more likely to provide operators with names like CONDITION-BIND which provide handlers for conditions within a certain scope  than operators with names like CONDITION-SET which side-effect some "global" default handler.
In Zetalisp, for example, the CONDITION-BIND macro makes a set of handlers available within the dynamic context of the execution of its body. It also accomplishes the additional service of partitioning the applicability of handlers according to condition type. For example, the form
(CONDITION-BIND ((type1 handler1) (type2 handler2) ...) form1 form2 ...)
will attempt to use a given handler for signals generated within its body only if those signals are of the associated type (or some subclass of that type).
Had the handlers not been partitioned by type, each handler would have been required to determine which its applicability dynamically. In practice, this would probably have resulted in unnecessarily bloated and highly idiomatic code usages such as:
(HYPOTHETICAL-CONDITION-BIND (#'(LAMBDA (CONDITION) (COND ((TYPEP CONDITION 'type1) (FUNCALL handler1 CONDITION)) (T (DECLINE CONDITION)))) #'(LAMBDA (CONDITION) (COND ((TYPEP CONDITION 'type2) (FUNCALL handler2 CONDITION)) (T (DECLINE CONDITION)))) ...) form1 form2 ...)
One could imagine the possibility of associating a newly defined condition type with a global handler for conditions of that type. In Lisp dialects providing some sort of generic function or message passing facility, this might be implemented as a method for objects of that condition type which is called only if no bound handler was found. Among other things, this idea of a global handler could be used to insure that some action (e.g., entry to the debugger) would always occur if a particular type of condition went unhandled.
Zetalisp runs each bound handler in an environment where it is no longer visible, so that if an error occurs while running a buggy handler, that handler will not be used to try to handle the error. A key problem with the idea of global handlers is that it is more difficult to devise a mechanism (especially in a shared memory, multi-tasking environment such as Zetalisp uses) for disabling a global handler temporarily (in a way that does not interfere with other "processes").
The "condition systems" provided in existing Lisps  are primarily synchronous. That is, the act of signalling a condition from code causes that code to block pending advice from handlers, which are then run synchronously in the same process.
Any discussion of asynchronous conditions would require new terminology to define what it means to interrupt a running process, and is beyond the scope of this paper. X3J13 has also chosen to leave this an open issue.
The descriptions of programs may be usefully divided into two parts, which separately describe their behavior in normal situations and their behavior in exceptional situations.
This distinction is important when developing programs because it allows programmers to make simplifying assumptions about the nature of programs. It can also have important consequences on the presentation of a program's code, by visually separating the code that handles normal and exceptional situations. Aspects of programs' modularity and efficiency can also be affected by the decision to make this distinction.
Most languages already provide adequate control mechanism for dealing with exceptional situations, including ignoring such situations, varying return value conventions, varying the number of return values, and allowing programs to exit via non-local transfer of control such as THROW.
Unfortunately, although these mechanisms are powerful enough to implement appropriate kinds of control structures, they are not sufficiently abstract to encourage as a language for describing and reasoning about exceptional situations. New vocabulary must be evolved to present existing functionality in a more abstract way.
In this paper, we have surveyed various aspects of existing condition systems  for Lisp and have established terminology which may be used in describing their behavior in an abstract way.
An earlier version of this paper [Pitman 85] was available as Working Paper 268 from the M.I.T Artificial Intelligence Laboratory but was never formally published. In spite of this, the original paper did serve as a basis for the design of the condition system which is part of the Common Lisp standard now  being finalized by American National Standards Institute (ANSI) subcommittee X3J13 in the United States. The concepts expressed in the earlier paper are as valid today as they were five years ago. The paper has been updated primarily to incorporate minor changes in terminology and "frame of reference" since the 1985 paper.
Many of the ideas in this paper were inspired by the "New Error System" for the Symbolics computers [Weinreb 83]. Other ideas descend from discussions with Eugene Ciccarelli, David Moon, and Daniel Weinreb.
Any reference to "the emerging ANSI Common Lisp standard" refers to work in progress by the American National Standards Institute (ANSI) subcommittee X3J13 in the United States. At the time this paper was written (December, 1989), X3J13 had not produced any public document such as a standard or even a draft standard. As such, references to what the ANSI standard might or might not look like are really just snapshots of work in progress and therefore subject to change without notice. Mention of ANSI and its ongoing process is for illustration purposes only--the ideas expressed here are those of the author and are not yet formally endorsed by ANSI.
[Since the time of this paper, of course, ANSI has published
X3.226/1994, informally referred to as ANSI Common Lisp.
The official copy is available only in hardcopy, but a content-equivalent
derivative work is available as
the Common Lisp HyperSpec.
Some terms used in this paper also occur in the documentation for Zetalisp (e.g., [Weinreb 83]) and in the emerging ANSI Common Lisp standard. The usages in this paper have been chosen to be essentially compatible.
|[Conway 75]||R.W. Conway and D. Gries, An Introduction to Programming: A Structured Approach Using PL/I and PL/C-7, Winthrop Publishers, Inc., Cambridge, MA, 1975.|
[The purpose of this reference here is probably confused. I had wanted to cite a reference to PL/1, from which many of the ideas in Weinreb's work descended. However, I think it may be only Honeywell Multics PL/1 that had the relevant features and this referenced item doesn't specifically document that dialect. Oh well. Can't change history. If someone does have a citation for Honeywell Multics PL/1, they should tell me and I'll update this document.
|[Lieberman 82]||H. Lieberman, "Seeing What Your Programs Are Doing," Memo 656, MIT Artificial Intelligence Laboratory, Cambridge, MA, February 1982.|
|[Liskov 79]||B. Liskov, et al, CLU Reference Manual, Technical Report 225, MIT Laboratory for Computer Science, Cambridge, MA, October 1979.|
|[Pitman 83]||K. Pitman, The Revised Maclisp Manual (Saturday Evening Edition), Technical Report 295, MIT Laboratory for Computer Science, Cambridge, MA, May 1983.|
|[Pitman 85]||K. Pitman, "Exceptional Situations In Lisp," Working Paper 268, MIT Artificial Intelligence Laboratory, Cambridge, MA, February 1985.|
|[Rich 81]||C. Rich and R.C. Waters, "The Disciplined Use of Simplifying Assumptions," Working Paper 220, MIT Artificial Intelligence Laboratory, Cambridge, MA, December 1981.|
|[Steele 76]||G.L. Steele, Jr., and G.J. Sussman, "LAMBDA, The Ultimate Imperative," Memo 353, MIT Artificial Intelligence Laboratory, Cambridge, MA, March 1976.|
|[Steele 78]||G.L. Steele, Jr., and G.J. Sussman, "The Revised Report on SCHEME: A Dialect of LISP," Memo 452, MIT Artificial Intelligence Laboratory, Cambridge, MA, January 1978.|
|[Steele 84]||G.L. Steele, Jr., Common Lisp: The Language, Digital Press, Burlington, MA, 1984.|
[Subsequent to publication of this paper, Steele published the Second Edition which incorporates the first; this second edition is available on the web. --KMP 04-Apr-1999]
|[Weinreb 81]||D.L. Weinreb and D.A. Moon, Lisp Machine Manual (Fourth Edition), MIT Artificial Intelligence Laboratory, Cambridge, MA, July 1981.|
|[Weinreb 83]||D.L. Weinreb, Signalling and Handling Conditions, Document #990097, Symbolics, Inc., Cambridge, MA, 1983.|
At least one commercially available dialect of Lisp does exactly this.
Maclisp is a dialect of Lisp for DEC PDP-10 and Honeywell Multics machines. It reached peak popularity in the 1970's and had a strong influence on Common Lisp.
Specifically, it will ignore the type information of the arguments, interpreting the bit patterns as if they represented fixnums. The result may be mystifying to novice programmers.
The Maclisp function ^G is so named because it simulates the effect of the Maclisp abort character, Control-G.
The concept which was originally introduced in Symbolics Zetalisp as a "proceed option" is called a "restart" by X3J13. This paper uses X3J13 terminology where possible.
Actually, Maclisp has only an error system, not a condition system, but we will assume that the same principles apply.
The best direction of search is a subject of debate. For example, the author contends that CONDITION-BIND-DEFAULT should search from outside in. However, few systems have ever reached the necessary complexity for this situation to have occurred at all. In time, as systems grow larger and need more fine tuning, better data will be available about what programmers really want and need, and subjective questions like this will be more readily answered.
The emerging ANSI Common Lisp specification uses the term HANDLER-BIND for what Zetalisp calls CONDITION-BIND.
The author is not aware of condition systems for any existing languages that support the notion of a class hierarchy for restart options.
The scope could in principle be either dynamic or lexical, though in practice dynamic scope or some very close variant always seems to be preferred in existing systems.
Original printed text document
Copyright 1985, 1990 Kent M. Pitman. All Rights Reserved.
HTML hypertext version of document
Copyright 1999, Kent M. Pitman. All rights reserved.
The following limited, non-exclusive, revokable licenses are granted:
Browsing of this document (that is, transmission and display of a temporary copy of this document for the ordinary purpose of direct viewing by a human being in the usual manner that hypertext browsers permit such viewing) is expressly permitted, provided that no recopying, redistribution, redisplay, or retransmission is made of any such copy.
Bookmarking of this document (that is, recording only the document's title and Uniform Resource Locator, or URL, but not its content, for the purpose of remembering an association between the document's title and the URL, and/or for the purpose of making a subsequent request for a fresh copy of the content named by that URL) is also expressly permitted.
All other uses require negotiated permission.
Click here for an index of other titles by Kent Pitman.