Control flow

In computer science and in computer programming, statements in pseudocode or in a program are normally obeyed one after the other in the order in which they are written (sequential flow of control). Most programming languages have control flow statements which allow variations in this sequential order:

statements may only be obeyed under certain conditions (choice),
statements may be obeyed repeatedly (loops),
a group of remote statements may be obeyed (subroutines).

The use of subroutines does not normally cause any control flow problems, but see the discussions below on early return, error recovery, and labels as parameters.

At the machine/assembly language level, it is usually the case that the only instructions available for handling choice and/or loops are goto and conditional goto (often known as variations of jump and/or branch). Compilers for high-level programming languages must translate all control-flow statements into these primitives.

Primitives

Labels

In a few programming languages (e.g. Fortran, BASIC), a label is just a whole number which appears at the beginning of a statement, e.g.

    1234 X = 3

In many programming languages, a label is an identifier, which is attached to a statement by using a colon (:), e.g.

    Success:print("target has been found")

Historical note: Algol 60 allowed both whole numbers and identifiers as labels (both attached by colons to statements), but few if any implementations allowed whole numbers.

Goto

The most common form for the unconditional transfer of control is just

    goto label

Conditional transfer of control varies from language to language, e.g.

    IF test THEN label
    IF (test) GOTO label
    if test then goto label;
    if (test) goto label;

For a fuller discussion on the drawbacks of goto, see Goto. In brief, undisciplined use of goto leads to spaghetti code which tends to be unmaintainable; see Edsger Dijkstra's comments in Go To Statement Considered Harmful . However, Donald Knuth has shown in Structured Programming with goto Statements that disciplined use of goto may be necessary to emulate missing control-flow structures.

A number of authors have pointed out that using goto is often acceptable, provided that control is transferred to some later statement (forward jump) and that control is not transferred into the middle of some other structured statement. Some of the control-flow statements available in high-level programming languages are effectively disguised gotos which comply with these conditions, e.g. break, continue, return as found in C/C++.

Subroutines

The terminology for subroutines varies; they may alternatively be known as routines, procedures, or sometimes methods. If they can be used in an expression and return a single result, they may also be known as functions.

In the 1950's, computer memories were very small by current standards so subroutines were used primarily to reduce program size; a piece of code was written once and then used many times from various other places in the program. Nowadays, subroutines are more fequently used to help make a program more structured, e.g. by isolating some particular algorithm or hiding some particular data access method. If many programmers are working on a single program, subroutines can be used to help split up the work.

Subroutines can be made much more useful by providing them with parameters, e.g. many programming langauges have a built-in square root subroutine whose parameter is the number you wish to find the square root of.

Some programming languages allow recursion, i.e. subroutines can call themselves directly or indirectly. Certain algorithms such as Quicksort and various tree-traversals are very much easier to express in recursive form than in non-recursive form.

The use of subroutines does slow a program down slightly, due to the overhead of passing parameters, calling the subroutine, entering the subroutine (may involve saving information on a stack), and returning. The actual overhead depends on both the hardware instructions available and on any software conventions which are used; excluding parameters, the overhead may range from 2 to 14 instructions, or worse. Some compilers may effectively insert the code of a subroutine inline at the point of call to remove this overhead.

In some programming languages, the only way of returning from a subroutine is by reaching the physical end of the subroutine. Other languages have a return statement. This is equivalent to a forward jump to the physical end of the subroutine and so does not complicate the control flow situation. There may be several such statements within a subroutine if required.

In most cases, a call to a subroutine is only a temporary diversion to the sequential flow of control, and so causes no problems for control flow analysis. A few languages allow labels to be passed as parameters, in which case understanding the control flow becomes very much more complicated, since you may then need to understand the subroutine to figure out what might happen.

Minimal structured control flow

In May 1966, Böhm and Jacopini published an article in Communications of the ACM which showed that any program with gotos could be transformed into a goto-free form involving only choice (IF THEN ELSE) and loops (WHILE condition DO xxx), possibly with duplicated code and/or the addition of Boolean variables (true/false flags). Can someone check for a possible 1964 publication of this in Italian? Later authors have shown that choice can be replaced by loops (and yet more Boolean variables).

The fact that such minimalism is possible does not necessarily mean that it is desirable; after all, computers theoretically only need one machine instruction (subtract one number from another and branch if the result is negative), but practical computers have dozens or even hundreds of machine instructions.

What Böhm and Jacopini's article showed was that all programs could be goto-free. Other research showed that control structures with one entry and one exit were much easier to understand than any other form, primarily because they could be used anywhere as a statement without disrupting the control flow.

Control structures in practice

Most programming languages with control structures have an initial keyword which indicates the type of control structure involved (Smalltalk is an exception). Languages then divide as to whether or not control structures have a final keyword.

No final keyword: Algol 60, Pascal, C, C++, Java, PL/1.
Such languages need some way of grouping statements together, e.g. begin end for Algol 60 and Pascal, curly brackets { } for C, C++, Java.

Final keyword: Algol 68, Modula-2, Fortran (77 onwards). The forms of the final keyword vary:
Algol 68: initial keyword backwards e.g. if fi, case esac,
Modula-2: same final keyword end for everything (now thought not to be good idea),
Fortran 77: final keyword is end + initial keyword, IF ENDIF, DO ENDDO

Languages which have a final keyword tend to have less debate regarding layout and indentation. Languages whose final keyword is of the form: end + initial keyword tend to easier to learn.

Choice

Choice using arbitrary tests

These are usually known as if statements. Note that if the language has an endif, then it usually has elseif as well, in order to avoid a large number of endifs for multiple tests.

    if test then statementTrue else statementFalse;

    if (test) statementTrue else statementFalse;

    if (test1) statementTrue1
    else if (test2) statement2True
    else if (test3) statement3True
    else            statementAllFalse;

    IF (test1) THEN
       xxx1True
    ELSEIF (test2) THEN
       xxx2True
    ELSEIF (test3) THEN
       xxx3True
    ELSE
       xxxAllFalse
    ENDIF

Choice based on specific constant values

These are usually known as case or switch statements. The effect is to compare a given value with specified constants and take action according to the first constant to match. If the constants form a compact range then this can be implemented very efficiently as if it were a choce based on whole numbers.

    case someChar of                switch (someChar) {
       'a': actionOnA;                 case 'a': actionOnA;
       'x': actionOnX;                     break;
       'y','z':actionOnYandZ;          case 'x': actionOnX;
    end;                                   break;
                                       case 'y':
                                       case 'z': actionOnYandZ;
                                           break;
                                       default: actionOnNoMatch;
                                    }

Choice based on whole numbers 1..N

Relatively few programming languages have these constructions but it can be implemented very efficiently using a computed goto.

    GOTO (label1,label2,label3), I
    outOfRangeAction

    case I in action1, action2, action3 out outOfRangeAction esac

Loops

A loop is a sequence of statements which is specified once but which may be carried out several times in succession. The code "inside" the loop (the body of the loop, shown below as xxx) is obeyed a specified number of times, or once for each of a collection of items, or until some condition is met.

A few languages do not have any constructions for looping (e.g. Lisp Scheme) and use tail recursion instead.

Count-controlled loops

Most programming languages have constructions for repeating a loop a certain number of times. Note that if N is less than 1 in these examples then the body is skipped completely. In most cases counting can go downwards instead of upwards and step sizes other than 1 can be used.

    FOR I = 1 TO N            for I := 1 to N do begin
        xxx                       xxx
    NEXT I                    end;

    DO I = 1,N                for ( I=1; I<=N; ++I ) {
        xxx                       xxx
    END DO                    }

Condition-controlled loops

Again, most programming languages have constructions for repeating a loop until some condition changes. Note that some variations place the test at the start of the loop, while others have the test at the end of the loop. In the former case the body may be skipped completely, while in the latter case the body is always obeyed at least once.

    DO WHILE (test)           repeat 
        xxx                       xxx 
    END DO                    until test;

    while (test) {            do
        xxx                       xxx
    }                         while (test);

Collection-controlled loops

A few programming languages (e.g. Perl, Smalltalk, C#) have special constructs which allow you to implicitly loop through all elements of an array, or all members of a set or collection.

    someCollection do: [ :eachElement | xxx ].

    foreach someArray { xxx }

    foreach (string s in myStringCollection) { xxx }

Early exit from loops

When using a count-controlled loop to search through a table, you may wish to stop searching as soon as you have found the required item. Some programming languages provide a statement such as break or exit, whose effect is to terminate the current loop immediately and transfer control to the statement immediately following that loop. Things can get a bit messy if you are searching a multi-dimensional table using nested loops (see Missing Control Structures below).

Potential problems with loops

Count-controlled loops should always use whole numbers or equivalent, since a loop such as
for X := 0.1 step 0.1 to 1.0 do
might be repeated 9 or 10 times, depending on rounding errors and/or the hardware and/or the compiler version.

Condition-controlled loops rely on the test condition being changed in some way within the body of the loop; if not, you get an infinite loop.

Error recovery

Most programming languages have some way in which a program can detect that end-of-file has been reached when reading data, but very few programming languages have any systematic way of handling the situation when something goes wrong or something unusual happens.

PL/1 has some 22 standard conditions (e.g. ZERODIVIDE SUBSCRIPTRANGE ENDFILE) which can be RAISEd and which can be intercepted by: ON condition action; Programmers can also define and use their own named conditions. In many cases a GOTO is needed to decide where flow of control should resume. Unfortunately, some implementations had a substantial overhead in both space and time (especially SUBSCRIPTRANGE), so many programmers tried to avoid using conditions.

Exceptions

C++, Java, and C# have a special construct for exception handling:

    try {
        xxx1                                  // Somewhere in here
        xxx2                                  //     use: throw someValue;
        xxx3
    } catch (someClass & someId) {            // catch value of someClass
        actionForSomeClass 
    } catch (someType & anotherId) {          // catch value of someType
        actionForSomeType
    } catch (...) {                           // catch anything not already caught
        actionForAnythingElse
    }

Any number and variety of catch clauses can be used above. In Java and C#, a finally clause can be added to the try construct. No matter how control leaves the try the code inside the finally clause is guaranteed to execute. This is useful when writing code that must relinquish an expensive resource (such as an opened file or a database connection) when finished processing:

    FileStream stm = null;                    // C# example
    try {
        stm = new FileStream("logfile.txt", FileMode.Create);
        return ProcessStuff(stm);                    // may throw an exception
    } finally {
        if (stm != null)
            stm.Close();
    }

Since this pattern is fairly common, C# has a special syntax that is slightly more readable:

    using (FileStream stm = new FileStream("logfile.txt", FileMode.Create)) {
        return ProcessStuff(stm);                  // may throw an exception
    }

Upon leaving the using-block, the compiler guarantees that the stm object is released.

The C++, Java, and C# languages define standard exceptions and the circumstances under which they are thrown. Users can throw exceptions of their own (in fact C++ allows users to throw and catch almost any type!) If there is no catch matching a particular throw, then control percolates back through subroutine calls and/or nested blocks until a matching catch is found or until the end of the main program is reached, at which point the program is forcibly stopped with a suitable error message.

Can anyone add something helpful about these or other programming languages?

Constructs to be avoided

Some "features" in programming languages tend to result in (very) unstructured code and are best avoided. A few of these are listed below.

Very few programming languages have all the control structures mentioned in this article, so you can reasonably use goto to emulate the missing structures as required; see Donald Knuth's 1974 article . You should not otherwise use goto, due to the risk of creating spaghetti code.

For reasons of backwards compatibility, Fortran still has some arcane unstructured features which should be avoided, e.g. ASSIGNed GOTO, Arithmetic IF (3-way branch), Logical IF (2-way branch), labels as parameters.

Self-modifying code, i.e. code which alters itself when executing, tends to result in very obscure code. Most assembly languages allow this, as does the ALTER verb in COBOL.

In a spoof Datamation article (December 1973), R. Lawrence Clark suggested that the GOTO statement could be replaced by the COMEFROM statement, and provides some entertaining examples. This was actually implemented in the INTERCAL programming language, a language designed to make programs as obscure as possible.

Missing control structures

In his 1974 article , Donald Knuth identified two situations which were not covered by the control structures listed above, and gave examples of control structures which could handle these situations. Despite their utility, these constructions have not yet found their way into main-stream programming languages.

Loop with test in the middle

This was proposed by Dahl in 1972.

    loop                           loop
        xxx1                           read(char);
    while test;                    while not atEndOfFile;
        xxx2                           write(char);
    repeat;                        repeat;

If xxx1 is omitted we get a loop with the test at the top. If xxx2 is omitted we get a loop with the test at the bottom. If while is omitted we get an infinite loop. Hence this single construction can replace several constructions in most programming languages. A possible variant is to allow more than one while test; within the loop, but the use of exitwhen (see next section) appears to cover this case better.

As the example on the right shows (copying a file one character at a time), there are simple situations where this is exactly the right construction to use in order to avoid duplicated code and/or repeated tests.

Multiple early exit/exit from nested loops

This was proposed by Zahn in 1974. A modified version is presented here.

    exitwhen EventA or EventB or EventC;
        xxx
    exits
        EventA: actionA
        EventB: actionB
        EventC: actionC
    endexit;

exitwhen is used to specify the events which may occur within xxx, their occurrence is indicated by using the name of the event as a statement. When some event does occur, the relevant action is carried out, and then control passes just after endexit. This construction provides a very clear separation between determining that some situation applies, and the action to be taken for that situation.

exitwhen is conceptually similar to the try/catch construct in C++, but is likely to be much more efficient since there is no percolation across subroutine calls and no transfer of arbitrary values. Also, the compiler can check that all specified events do actually occur and have associated actions.

The following simple example involves searching a two-dimensional table for a particular item.

    exitwhen found or missing;
        for I := 1 to N do
            for J := 1 to M do
                if table[I,J] = target then found;
        missing;
    exits
        found:   print("item is in table");
        missing: print("item is not in table");
    endexit;

Anecdotal evidence

The following statistics apply to a 6000-line compiler written in a private language containing the above constructions.

There are 10 condition-controlled loops, of which 6 have the test at the top, 1 has the test at the bottom, and 3 have the test in the middle.

There are 18 exitwhen statements, 5 with 2 events, 11 with 3 events, and 2 with 4 events. When these were first used in the compiler, replacing various flags and tests, the number of source lines increased by 0.1%, the size of the object code decreased by 3%, and the compiler (when compiling itself) was 4% faster. Prior to the introduction of exitwhen, 4 of the condition-controlled loops had more than one while test; and 5 of the count-controlled loops also had a while test;

References

Dahl & Dijkstra & Hoare, "Structured Programming" Academic Press, 1972.
Böhm, Jacopini. Flow diagrams, "Turing Machines and Languages with only Two Formation Rules" Communications of the ACM, 9(5):366-371, May 1966.
Hoare, C. A. R. "Partition: Algorithm 63," "Quicksort: Algorithm 64," and "Find: Algorithm 65." Comm. ACM 4, 321-322, 1961.
Zahn, C. T. "A control statement for natural top-down structured programming" presented at Symposium on Programming Languages, Paris, 1974.