ABSTRACT

The C language is like a carving knife: simple, sharp, and extremely useful in skilled hands. Like any sharp tool, C can injure people who don't know how to handle it.
 
 

0,Introduction	5,Library Functions
1,Lexical Pitfalls	6,The Preprocessor
2,Syntactic Pitfalls	7,Portability Pitfalls
3,Linkage	8,This Space Available
4,Semantic Pitfalls	References

Introduction

The C language and its typical implementations are designed to be used easily by experts. The language is terse and expressive. There are few restrictions to keep the user from blundering. A user who has blundered is often rewarded by an effect that is not obviously related to the cause.

In this paper, we will look at some of these unexpected rewards. Because they are unexpected, it may well be impossible to classify them completely. Nevertheless, we have made a rough effort to do so by looking at what has to happen in order to run a C program. We assume the reader has at least a passing acquaintance with the C language.

Section 1 looks at problems that occur while the program is being broken into tokens. Section 2 follows the program as the compiler groups its tokens into declarations, expressions, and statements. Section 3 recognizes that a C program is often made out of several parts that are compiled separately and bound together. Section 4 deals with misconceptions of meaning: things that happen while the program is actually running. Section 5 examines the relationship between our programs and the library routines they use. In section 6 we note that the program we write is not really the program we run; the preprocessor has gotten at it first. Finally, section 7 discusses portability problems: reasons a program might run on one implementation and not another.

The first part of a compiler is usually called a lexical analyzer. This looks at the sequence of characters that make up the program and breaks them up into tokens. A token is a sequence of one or more characters that have a (relatively) uniform meaning in the language being compiled. In C, for instance, the token -> has a meaning that is quite distinct from that of either of the characters that make it up, and that is independent of the context in which the -> appears.

For another example, consider the statement:

if (x > big) big = x;

In fact, C programs are broken into tokens twice. First the preprocessor reads the program. It must tokenize the program so that it can find the identifiers, some of which may represent macros. It must then replace each macro invocation by the result of evaluating that macro. Finally, the result of the macro replacement is reassembled into a character stream which is given to the compiler proper. The compiler then breaks the stream into tokens a second time.

In this section, we will explore some common misunderstandings about the meanings of tokens and the relationship between tokens and the characters that make them up. We will talk about the preprocessor later.

1.1. = is not ==

Programming languages derived from Algol, such as Pascal and Ada, use := for assignment and = for comparison. C, on the other hand, uses = for assignment and == for comparison. This is because assignment is more frequent than comparison, so the more common meaning is given to the shorter symbol.

Moreover, C treats assignment as an operator, so that multiple assignments (such as a=b=c) can be written easily and assignments can be embedded in larger expressions.

This convenience causes a potential problem: one can inadvertently write an assignment where one intended a comparison. Thus, this statement, which looks like it is checking whether x is equal to y:

if (x = y)
    foo();

while (c == ' ' || c = '\t' || c == '\n')
    c = getc (f);

Some C compilers try to help the user by giving a warning message for conditions of the form e1 = e2. To avoid warning messages from such compilers, when you want to assign a value to a variable and then check whether the variable is zero, consider making the comparison explicit. In other words, instead of:

if (x = y)
    foo();

if ((x = y) != 0)
    foo();

1.2. & and | are not && or ||

It is easy to miss an inadvertent substitution of = for == because so many other languages use = for comparison. It is also easy to interchange & and &&, or | and ||, especially because the & and | operators in C are different from their counterparts in some other languages. We will look at these operators more closely in section 4.

1.3. Multi-character Tokens

Some C tokens, such as /, *, and =, are only one character long. Other C tokens, such as /* and ==, and identifiers, are several characters long. When the C compiler encounters a / followed by an *, it must be able to decide whether to treat these two characters as two separate tokens or as one single token. The C reference manual tells how to decide: "If the input stream has been parsed into tokens up to a given character, the next token is taken to include the longest string of characters which could possibly constitute a token.'' Thus, if a / is the first character of a token, and the / is immediately followed by a *, the two characters begin a comment, regardless of any other context.

The following statement looks like it sets y to the value of x divided by the value pointed to by p:

y = x/*p /* p points at the divisor */;

y = x / *p /* p points at the divisor */;

y = x/(*p) /* p points at the divisor */;

This sort of near-ambiguity can cause trouble in other contexts. For example, older versions of C use =+ to mean what present versions mean by +=. Such a compiler will treat

a=-1;

a =- 1;

a = a - 1;

a = -1;

a=/*b;

a =/ * b ;

1.4. Exceptions

Compound assignment operators such as += are really multiple tokens. Thus,

a + /* strange */ = 1

a += 1

p - > a

p -> a

1.5. Strings and Characters

Single and double quotes mean very different things in C, and there are some contexts in which confusing them will result in surprises rather than error messages.

A character enclosed in single quotes is just another way of writing an integer. The integer is the one that corresponds to the given character in the implementation's collating sequence. Thus, in an ASCII implementation, 'a' means exactly the same thing as 0141 or 97. A string enclosed in double quotes, on the other hand, is a short-hand way of writing a pointer to a nameless array that has been initialized with the characters between the quotes and an extra character whose binary value is zero.

The following two program fragments are equivalent:

printf ("Hello world\n");

char hello[] = {'H', 'e', 'l', 'l', 'o', ' ', 'w', 'o', 'r', 'l', 'd', '\n', 0}; printf (hello);

printf('\n');

printf ("\n");

Because an integer is usually large enough to hold several characters, some C compilers permit multiple characters in a character constant. This means that writing 'yes' instead of "yes" may well go undetected. The latter means "the address of the first of four consecutive memory locations containing y, e, s, and a null character, respectively.'' The former means "an integer that is composed of the values of the characters y, e, and s in some implementation-defined manner.'' Any similarity between these two quantities is purely coincidental.

To understand a C program, it is not enough to understand the tokens that make it up. One must also understand how the tokens combine to form declarations, expressions, statements, and programs. While these combinations are usually well-defined, the definitions are sometimes counter-intuitive or confusing.

In this section, we look at some syntactic constructions that are less than obvious.

2.1. Understanding Declarations

I once talked to someone who was writing a C program that was going to run stand-alone in a small microprocessor. When this machine was switched on, the hardware would call the subroutine whose address was stored in location 0.

In order to simulate turning power on, we had to devise a C statement that would call this subroutine explicitly. After some thought, we came up with the following:

(*(void(*)())0)();

Every C variable declaration has two parts: a type and a list of stylized expressions that are expected to evaluate to that type. The simplest such expression is a variable:

float f, g;

float ((f));

Similar logic applies to function and pointer types. For example,

float ff();

float *pf;

These forms combine in declarations the same way they do in expressions. Thus

float *g(), (*h)();

Once we know how to declare a variable of a given type, it is easy to write a cast for that type: just remove the variable name and the semicolon from the declaration and enclose the whole thing in parentheses. Thus, since

float *g();

Armed with this knowledge, we are now prepared to tackle (*(void(*)())0)(). We can analyze this statement in two parts. First, suppose that we have a variable fp that contains a function pointer and we want to call the function to which fp points. That is done this way:

(*fp)();

This problem is the second part of our analysis. If C could read our mind about types, we could write:

(*0)();

If fp is a pointer to a function returning void, then (*fp)() is a void value, and its declaration would look like this:

void (*fp)();

void (*fp)();
(*fp)();

(void(*)())0

(*(void(*)())0)();

At the time we tackled this problem, there was no such thing as a typedef declaration. Using it, we could have solved the problem more clearly:

typedef void (*funcptr)();
(* (funcptr) 0)();

Suppose that the manifest constant FLAG is an integer with exactly one bit turned on in its binary representation (in other words, a power of two), and you want to test whether the integer variable flags has that bit turned on. The usual way to write this is:

if (flags & FLAG) ...

if (flags & FLAG != 0) ...

if (flags & (FLAG != 0)) ...

Suppose you have two integer variables, h and l, whose values are between 0 and 15 inclusive, and you want to set r to an 8-bit value whose low-order bits are those of l and whose high-order bits are those of h. The natural way to do this is to write:

r = h<<4 + l;

r = h << (4 + l);

r = (h << 4) + l;
r = h << 4 | l;

The operators that bind the most tightly are the ones that aren't really operators: subscripting, function calls, and structure selection. These all associate to the left.

Next come the unary operators. These have the highest precedence of any of the true operators. Because function calls bind more tightly than unary operators, you must write (*p)() to call a function pointed to by p; *p() implies that p is a function that returns a pointer. Casts are unary operators and have the same precedence as any other unary operator. Unary operators are right-associative, so *p++ is interpreted as *(p++) and not as (*p)++.

Next come the true binary operators. The arithmetic operators have the highest precedence, then the shift operators, the relational operators, the logical operators, the assignment operators, and finally the conditional operator. The two most important things to keep in mind are:
1. Every logical operator has lower precedence than every relational operator.
2. The shift operators bind more tightly than the relational operators but less tightly than the arithmetic operators.

Within the various operator classes, there are few surprises. Multiplication, division, and remainder have the same precedence, addition and subtraction have the same precedence, and the two shift operators have the same precedence.

One small surprise is that the six relational operators do not all have the same precedence: == and != bind less tightly than the other relational operators. This allows us, for instance, to see if a and b are in the same relative order as c and d by the expression

a < b == c < d

The ternary conditional operator has lower precedence than any we have mentioned so far. This permits the selection expression to contain logical combinations of relational operators, as in

z = a < b && b < c ? d : e

a = b = c

b = c; a = b;

Assignment is another operator often involved in precedence mixups. Consider, for example, the following loop intended to copy one file to another:

while (c=getc(in) != EOF)
    putc(c,out);

It is not too hard to see that the example above should be written:

while ((c=getc(in)) != EOF)
    putc(c,out);

if( (t=BTYPE(pt1?aty)==STRTY) || t==UNIONTY ){

The precedence of the C logical operators comes about for historical reasons. B, the predecessor of C, had logical operators that corresponded rougly to C's & and | operators. Although they were defined to act on bits, the compiler would treat them as && and || if they were in a conditional context. When the two usages were split apart in C, it was deemed too dangerous to change the precedence much.

2.3. Watch Those Semicolons!

An extra semicolon in a C program usually makes little difference: either it is a null statement, which has no effect, or it elicits a diagnostic message from the compiler, which makes it easy to remove. One important exception is after an if or while clause, which must be followed by exactly one statement. Consider this example:

if (x[i] > big);
    big = x[i];

if (x[i] > big)
    big = x[i];

if (x[i] > big) { }
    big = x[i];

big = x[i];

Another place that a semicolon can make a big difference is at the end of a declaration just before a
function definition. Consider the following fragment:

struct foo {
    int x;
}

f()
{
    . . .
}

2.4. The Switch Statement

C is unusual in that the cases in its switch statement can flow into each other. Consider, for example, the following program fragments in C and Pascal:

switch (color) {
case 1: printf ("red");
    break;
case 2: printf ("yellow");
    break;
case 3: printf ("blue");
    break;
}
case color of
1: write ('red');
2: write ('yellow');
3: write ('blue')
end

Looking at it another way, suppose the C fragment looked more like the Pascal fragment:

switch (color) {
case 1: printf ("red");
case 2: printf ("yellow");
case 3: printf ("blue");
}

This is both a strength and a weakness of C switch statements. It is a weakness because leaving out a break statement is easy to do, and often gives rise to obscure program misbehavior. It is a strength because by leaving out a break statement deliberately, one can readily express a control structure that is inconvenient to implement otherwise. Specifically, in large switch statements, one often finds that the processing for one of the cases reduces to some other case after a relatively small amount of special handling.

For example, consider a program that is an interpreter for some kind of imaginary machine. Such a program might contain a switch statement to handle each of the various operation codes. On such a machine, it is often true that a subtract operation is identical to an add operation after the sign of the second operand has been inverted. Thus, it is nice to be able to write something like this:

case SUBTRACT:
    opnd2 = -opnd2;
    /* no break */
case ADD:
    . . .

case '\n':
    linecount++;
    /* no break */
case '\t':
case ' ':
    . . .

Unlike some other programming languages, C requires a function call to have an argument list, even if there are no arguments. Thus, if f is a function,

f();

f;

2.6. The Dangling else Problem

We would be remiss in leaving any discussion of syntactic pitfalls without mentioning this one. Although it is not unique to C, it has bitten C programmers with many years of experience.

Consider the following program fragment:

if (x == 0)
    if (y == 0) error();
else {
    z = x + y;
    f (&z);
}

However, the program fragment actually does something quite different. The reason is the rule that an else is always associated with the closest unmatched if. If we were to indent this fragment the way it is actually executed, it would look like this:

if (x == 0) {
    if (y == 0)
        error();
    else {
        z = x + y;
        f (&z);
    }
}

if (x == 0) {
    if (y == 0)
        error();
} else {
    z = x + y;
    f (&z);
}

A C program may consist of several parts that are compiled separately and then bound together by a program usually called a linker, linkage editor, or loader. Because the compiler normally sees only one file at a time, it cannot detect errors whose recognition would require knowledge of several source program files at once.

In this section, we look at some errors of that type. Some C implementations, but not all, have a program called lint that catches many of these errors. It is impossible to overemphasize the importance of using such a program if it is available.

3.1. You Must Check External Types Yourself

Suppose you have a C program divided into two files. One file contains the declaration:

int n;

long n;

What actually happens when this program is run? There are many possibilities:

1. The implementation is clever enough to detect the type clash. One would then expect to see a diagnostic message explaining that the type of n was given differently in two different files.

2. You are using an implementation in which int and long are really the same type. This is typically true of machines in which 32-bit arithmetic comes most naturally. In this case, your program will probably work as if you had said long (or int) in both declarations. This would be a good example of a program that works only by coincidence.

3. The two instances of n require different amounts of storage, but they happen to share storage in such a way that the values assigned to one are valid for the other. This might happen, for example, if the linker arranged for the int to share storage with the low-order part of the long. Whether or not this happens is obviously machine-and system-dependent. This is an even better example of a program that works only by coincidence.

4. The two instances of n share storage in such a way that assigning a value to one has the effect of apparently assigning a different value to the other. In this case, the program will probably fail.

Another example of this sort of thing happens surprisingly often. One file of a program will contain a declaration like:

char filename[] = "/etc/passwd";

char *filename;

In the second declaration, filename is the name of a pointer. That pointer points wherever the programmer makes it point. If the programmer doesn't give it a value, it will have a zero (null) value by default.

The two declarations of filename use storage in different ways; they cannot coexist.

One way to avoid type clashes of this sort is to use a tool like lint if it is available. In order to be able to check for type clashes between separately compiled parts of a program, some program must be able to see all the parts at once. The typical compiler does not do this, but lint does.

Another way to avoid these problems is to put external declarations into include files. That way, the type of an external object only appears once.

A sentence can be perfectly spelled and written with impeccable grammar and still be meaningless. In this section, we will look at ways of writing programs that look like they mean one thing but actually mean something quite different.

We will also discuss contexts in which things that look reasonable on the surface actually give undefined results. We will limit ourselves here to things that are not guaranteed to work on any C implementation.

We will leave those that might work on some implementations but not others until section 7, which looks at portability problems.

4.1. Expression Evaluation Sequence

Some C operators always evaluate their operands in a known, specified order. Others don't. Consider, for instance, the following expression:

a < b && c < d

To evaluate a<b, on the other hand, the compiler may evaluate either a or b first. On some machines, it may even evaluate them in parallel.

Only the four C operators &&, ||, ?:, and , specify an order of evaluation. && and || evaluate the left operand first, and the right operand only if necessary. The ?: operator takes three operands: a?b:c evaluates a first, and then evaluates either b or c, depending on the value of a. The , operator evaluates its left operand and discards its value, then evaluates its right operand.

All other C operators evaluate their operands in undefined order. In particular, the assignment operators do not make any guarantees about evaluation order.

For this reason, the following way of copying the first n elements of array x to array y doesn't work:

i = 0;
while (i < n)
    y[i] = x[i++];

On some implementations, it will; on others, it won't. This similar version fails for the same reason:

i = 0;
while (i < n)
    y[i++] = x[i];

i = 0;
while (i < n) {
    y[i] = x[i];
    i++;
}

for (i = 0; i < n; i++)
    y[i] = x[i];

4.2. The &&, ||, and ! Operators

C has two classes of logical operators that are occasionally interchangeable: the bitwise operators &, |, and ? and the logical operators &&, ||, and !. A programmer who substitutes one of these operators for the corresponding operator from the other class may be in for a surprise: the program may appear to work correctly after such an interchange but may actually be working only by coincidence.

The &, |, and ?operators treat their operands as a sequence of bits and work on each bit separately. For example, 10&12 is 8 (1000), because & looks at the binary representations of 10 (1010) and 12 (1100) and produces a result that has a bit turned on for each bit that is on in the same position in both operands. Similarly, 10|12 is 14 (1110) and ?0 is --11 (11...110101), at least on a 2's complement machine.

The &&, ||, and ! operators, on the other hand, treat their arguments as if they are either "true'' or "false,'' with the convention that 0 represents "false'' and any other value represents "true.'' These operators return 1 for ``true'' and 0 for ``false,'' and the && and || operators do not even evaluate their right-hand operands if their results can be determined from their left-hand operands.

Thus !10 is zero, because 10 is nonzero, 10&&12 is 1, because both 10 and 12 are nonzero, and 10||12 is also 1, because 10 is nonzero. Moreover, 12 is not even evaluated in the latter expression, nor is f() in 10||f().

Consider the following program fragment to look for a particular element in a table:

 
i = 0;
while (i < tabsize && tab[i] != x)
    i++;

Suppose that the && were inadvertently replaced by & in this example. Then the loop would probably still appear to work, but would do so only because of two lucky breaks.

The first is that both comparisons in this example are of a sort that yield 0 if the condition is false and 1 if the condition is true. As long as x and y are both 1 or 0, x&y and x&&y will always have the same value. However, if one of the comparisons were to be replaced by one that uses some non-zero value other than 1 to represent ``true,'' then the loop would stop working.

The second lucky break is that looking just one element off the end of an array is usually harmless, provided that the program doesn't change that element. The modified program looks past the end of the array because &, unlike &&, must always evaluate both of its operands. Thus in the last iteration of the loop, the value of tab[i] will be fetched even though i is equal to tabsize. If tabsize is the number of elements in tab, this will fetch a non-existent element of tab.

4.3. Subscripts Start from Zero

In most languages, an array with n elements normally has those elements numbered with subscripts ranging from 1 to n inclusive. Not so in C.

A C array with n elements does not have an element with a subscript of n, as the elements are numbered from 0 through n?. Because of this, programmers coming from other languages must be especially careful when using arrays:

int i, a[10];
for (i=1; i<=10; i++)
    a[i] = 0;

4.4. C Doesn't Always Cast Actual Parameters

The following simple program fragment fails for two reasons:

double s;
s = sqrt (2);
printf ("%g\n", s);

double s, sqrt();
s = sqrt (2.0);
printf ("%g\n", s);

Therefore, a programmer who uses a function like sqrt, whose parameter is a double, must be careful to pass arguments that are of float or double type only. The constant 2 is an int and is therefore of the wrong type.

When the value of a function is used in an expression, that value is automatically cast to an appropriate type. However, the compiler must know the actual type returned by the function in order to be able to do this. Functions used without further declaration are assumed to return an int, so declarations for such functions are unnecessary. However, sqrt returns a double, so it must be declared as such before it can be used successfully.

In practice, C implementations generally provide a file that can be brought in with an include statement that contains declarations for library functions like sqrt, but writing declarations is still necessary for programmers who write their own functions -- in other words, for anyone who writes non-trivial C programs.

Here is a more spectacular example:

main()
{
    int i;
    char c;
    for (i=0; i<5; i++) {
        scanf ("%d", &c);
        printf ("%d ", i);
    }
    printf ("\n");
}

Why? The key is the declaration of c as a char rather than as an int. When you ask scanf to read an integer, it expects a pointer to an integer. What it gets in this case is a pointer to a character. Scanf has no way to tell that it didn't get what it expected: it treats its input as an integer pointer and stores an integer there. Since an integer takes up more memory than a character, this steps on some of the memory near c.

Exactly what is near c is the compiler's business; in this case it turned out to be the low-order part of i. Therefore, each time a value was read for c, it reset i to zero. When the program finally reached end of file, scanf stopped trying to put new values into c, so i could be incremented normally to end the loop.

4.5. Pointers are not Arrays

C programs often adopt the convention that a character string is stored as an array of characters, followed by a null character. Suppose we have two such strings s and t, and we want to concatenate them into a single string r. To do this, we have the usual library functions strcpy and strcat. The following obvious method doesn't work:

char *r;
strcpy (r, s);
strcat (r, t);

Let's try again, allocating some memory for r:

char r[100];
strcpy (r, s);
strcat (r, t);

char *r, *malloc();
r = malloc (strlen(s) + strlen(t));
strcpy (r, s);
strcat (r, t);

Second, and much more important, is that the call to malloc doesn't allocate quite enough memory. Recall the convention that a string is terminated by a null character. The strlen function returns the number of characters in the argument string, excluding the null character at the end. Therefore, if strlen(s) is n, s really requires n+1 characters to contain it. We must therefore allocate one extra character for r. After doing this and checking that malloc worked, we get:

char *r, *malloc();
r = malloc (strlen(s) + strlen(t) + 1);
if (!r) {
    complain();
    exit (1);
}
strcpy (r, s);
strcat (r, t);

A synecdoche (sin-ECK-duh-key) is a literary device, somewhat like a simile or a metaphor, in which, according to the Oxford English Dictionary, "a more comprehensive term is used for a less comprehensive or vice versa; as whole for part or part for whole, genus for species or species for genus, etc.''

This exactly describes the common C pitfall of confusing a pointer with the data to which it points. This is most common for character strings. For instance:

char *p, *q;
p = "xyz";

q = p;

p
|
x y z '\0'
|
q

Thus, if after this we were to execute

q[1] = 'Y';

4.7. The Null Pointer is Not the Null String

The result of converting an integer to a pointer is implementation-dependent, with one important exception. That exception is the constant 0, which is guaranteed to be converted to a pointer that is unequal to any valid pointer. For documentation, this value is often given symbolically:

#define NULL 0

if (p == (char *) 0) ...

if (strcmp (p, (char *) 0) == 0) ...

If p is a null pointer, it is not even valid to say:

printf (p);

printf ("%s", p);

The C language definition is very specific about what happens when an integer operation overflows or underflows.

If either operand is unsigned, the result is unsigned, and is defined to be modulo 2 n , where n is the word size. If both operands are signed, the result is undefined.

Suppose, for example, that a and b are two integer variables, known to be non-negative, and you want to test whether a+b might overflow. One obvious way to do it looks something like this:

if (a + b < 0)
    complain();

The point is that once a+b has overflowed, all bets are off as to what the result will be. For example, on some machines, an addition operation sets an internal register to one of four states: positive, negative, zero, or overflow. On such a machine, the compiler would have every right to implement the example given above by adding a and b and checking whether this internal register was in negative state afterwards.

If the operation overflowed, the register would be in overflow state, and the test would fail.

One correct way of doing this particular test relies on the fact that unsigned arithmetic is well-defined for all values, as are the conversions between signed and unsigned values:

if ((int) ((unsigned) a + (unsigned) b) < 0)
    complain();

Two questions seem to cause trouble for people who use shift operators:

1. In a right shift, are vacated bits filled with zeroes or copies of the sign bit?
2. What values are permitted for the shift count?

The answer to the first question is simple but sometimes implementation-dependent. If the item being shifted is unsigned, zeroes are shifted in. If the item is signed, the implementation is permitted to fill vacated bit positions either with zeroes or with copies of the sign bit. If you care about vacated bits in a right shift, declare the variable in question as unsigned. You are then entitled to assume that vacated bits will be set to zero.

The answer to the second question is also simple: if the item being shifted is n bits long, then the shift count must be greater than or equal to zero and strictly less than n. Thus, it is not possible to shift all the bits out of a value in a single operation.

For example, if an int is 32 bits, and n is an int, it is legal to write n<<31 and n<<0 but not n<<32 or n<<-1.

Note that a right shift of a signed integer is generally not equivalent to division by a power of two, even if the implementation copies the sign into vacated bits. To prove this, consider that the value of (-1)>>1 cannot possibly be zero.

Every useful C program must use library functions, because there is no way of doing input or output built into the language. In this section, we look at some cases where some widely-available library functions behave in ways that the programmer might not expect.
 

5.1. Getc Returns an Integer

Consider the following program:

#include <stdio.h>
main()
{
    char c;
    while ((c = getchar()) != EOF)
        putchar (c);
}

The reason is that c is declared as a character rather than as an integer. This means that it is impossible for c to hold every possible character as well as EOF.

Thus there are two possibilities. Either some legitimate input character will cause c to take on the same value as EOF, or it will be impossible for c to have the value EOF at all. In the former case, the program will stop copying in the middle of certain files. In the latter case, the program will go into an infinite loop.

Actually, there is a third case: the program may work by coincidence. The C Reference Manual defines the result of the expression

((c = getchar()) != EOF)

When a longer integer is converted to a shorter or to a char, it is truncated on the left; excess bits are simply discarded.

Section 7.14 states:

There are a number of assignment operators, all of which group right-to-left. All require an lvalue as their left operand, and the type of an assignment expression is that of its left operand. The value is the value stored in the left operand after the assignment has taken place.

The combined effect of these two sections is to require that the result of getchar be truncated to a character value by discarding the high-order bits, and that this truncated value then be compared with EOF. As part of this comparison, the value of c must be extended to an integer, either by padding on the left with zero bits or by sign extension, as appropriate.

However, some compilers do not implement this expression correctly. They properly assign the low-order bits of the value of getchar to c. However, instead of then comparing c to EOF, they compare the entire value of getchar! A compiler that does this will make the sample program shown above appear to work "correctly.''

5.2. Buffered Output and Memory Allocation

When a program produces output, how important is it that a human be able to see that output immediately? It depends on the program.

For example, if the output is going to a terminal and is asking the person sitting at that terminal to answer a question, it is crucial that the person see the output in order to be able to know what to type. On the other hand, if the output is going to a file, from where it will eventually be sent to a line printer, it is only important that all the output get there eventually.

It is often more expensive to arrange for output to appear immediately than it is to save it up for a while and write it later on in a large chunk. For this reason, C implementations typically afford the programmer some control over how much output is to be produced before it is actually written.

That control is often vested in a library function called setbuf. If buf is a character array of appropriate size, then

setbuf (stdout, buf);

Thus, the following program illustrates the obvious way to use setbuf in a program that copies its standard input to its standard output:

#include <stdio.h>
main()
{
    int c;
    char buf[BUFSIZ];
    setbuf (stdout, buf);
    while ((c = getchar()) != EOF)
        putchar (c);
}

To see where the trouble lies, ask when the buffer is flushed for the last time. Answer: after the main program has finished, as part of the cleaning up that the library does before handing control back to the operating system. But by that time, the buffer has already been freed!

There are two ways to prevent this sort of trouble.

First, make the buffer static, either by declaring it explicitly as static:

static char buf[BUFSIZ];

Another possibility is to allocate the buffer dynamically and never free it:

char *malloc();
setbuf (stdout, malloc (BUFSIZ));

The programs we run are not the programs we write: they are first transformed by the C preprocessor.
The preprocessor gives us a way of abbreviating things that is important for two major reasons (and several minor ones).

First, we may want to be able to change all instances of a particular quantity, such as the size of a table, by changing one number and recompiling the program.

Second, we may want to define things that appear to be functions but do not have the execution overhead normally associated with a function call. For example, getchar and putchar are usually implemented as macros to avoid having to call a function for each character of input or output.

6.1. Macros are not Functions

Because macros can be made to appear almost as if they were functions, programmers are sometimes tempted to regard them as truly equivalent. Thus, one sees things like this:

#define max(a,b) ((a)>(b)?(a):(b))

Not only can this be inefficient, it can also be wrong:

biggest = x[0];
i = 1;
while (i < n)
   biggest = max (biggest, x[i++]);

biggest = ((biggest)>(x[i++])?(biggest):(x[i++]));

Because the relation is false, the value of x[i++] is now assigned to biggest. However, i is now 2, so the value assigned to biggest is the value of x[2], which is 1.

One way around these worries is to ensure that the arguments to the max macro do not have any side effects:

biggest = x[0];
for (i = 1; i < n; i++)
    biggest = max (biggest, x[i]);

#define putc(x,p) (­­(p)?_cnt>=0?(*(p)?_ptr++=(x)):_flsbuf(x,p))

Some C implementations are less careful. For instance, not everyone handles putc(*c++,f) correctly. As another example, consider the toupper function that appears in many C libraries. It translates a lower-case letter to the corresponding upper-case letter while leaving other characters unchanged. If we assume that all the lower-case letters and all the upper-case letters are contiguous (with a possible gap between the cases), we get the following function:

toupper(c)
{
    if (c >= 'a' && c <= 'z')
        c += 'A' ?'a';
    return c;
}

#define toupper(c) ((c)>='a' && (c)<='z'? (c)+('A'?a'): (c))

Another thing to watch out for when using macros is that they may generate very large expressions indeed. For example, look again at the definition of max:

#define max(a,b) ((a)>(b)?(a):(b))

((a)>(((b)>(((c)>(d)?(c):(d)))?(b):(((c)>(d)?(c):(d)))))?
(a):(((b)>(((c)>(d)?(c):(d)))?(b):(((c)>(d)?(c):(d))))))

max(max(a,b),max(c,d))

((((a)>(b)?(a):(b)))>(((c)>(d)?(c):(d)))?
(((a)>(b)?(a):(b))):(((c)>(d)?(c):(d))))

biggest = a;
if (biggest < b) biggest = b;
if (biggest < c) biggest = c;
if (biggest < d) biggest = d;

One common use of macros is to permit several things in diverse places to be the same type:

#define FOOTYPE struct foo
FOOTYPE a;
FOOTYPE b, c;

Using a macro definition for this has the advantage of portability -- any C compiler supports it. Most C compilers also support another way of doing this:

typedef struct foo FOOTYPE;

These two ways of naming a type may appear to be equivalent, but the typedef is more flexible. Consider, for example, the following:

#define T1 struct foo *
typedef struct foo *T2;

T1 a, b;
T2 c, d;

struct foo * a, b;

C has been implemented by many people to run on many machines. Indeed, one of the reasons to write programs in C in the first place is that it is easy to move them from one programming environment to another.

However, because there are so many implementors, they do not all talk to each other. Moreover, different systems have different requirements, so it is reasonable to expect C implementations to differ slightly between one machine and another.

Because so many of the early C implementations were associated with the UNIX operating system, the nature of many of these functions was shaped by that system. When people started implementing C under other systems, they tried to make the library behave in ways that would be familiar to programmers used to the UNIX system.

They did not always succeed. What is more, as more people in different parts of the world started working on different versions of the UNIX system, the exact nature of some of the library functions inevitably diverged. Today, a C programmer who wishes to write programs useful in someone else's environment must know about many of these subtle differences.

7.1. What's in a Name?

Some C compilers treat all the characters of an identifier as being significant. Others ignore characters past some limit when storing identifiers. C compilers usually produce object programs that must then be processed by loaders in order to be able to access library subroutines. Loaders, in turn, often impose their own restrictions on the kinds of names they can handle.

One common loader restriction is that letters in external names must be in upper case only. When faced with such a restriction, it is reasonable for a C implementor to force all external names to upper case. Restrictions of this sort are blessed by section 2.1 the C reference manual:

An identifier is a sequence of letters and digits; the first character must be a letter. The underscore _ counts as as a letter. Upper and lower case letters are different. No more than the first eight characters are significant, although more may be used. External identifiers, which are used by various assemblers and loaders, are more restricted:

Here, the reference manual goes on to give examples of various implementations that restrict external identifiers to a single case, or to fewer than eight characters, or both.

Because of all this, it is important to be careful when choosing identifiers in programs intended to be portable. Having two subroutines named, say print_fields and print_float would not be a very good idea.

As a striking example, consider the following function:

char * Malloc (n)
unsigned n;
{
    char *p, *malloc();
    p = malloc (n);
    if (p == NULL)
        panic ("out of memory");
    return p;
}

Consider, however, what happens when this function is used on a system that ignores case distinctions in external identifiers. In effect, the names malloc and Malloc become equivalent. In other words, the library function malloc is effectively replaced by the Malloc function above, which when it calls malloc is really calling itself. The result, of course, is that the first attempt to allocate memory results in a recursion loop and consequent mayhem, even though the function will work on an implementation that preserves case distinctions.

7.2. How Big is an Integer?

C provides the programmer with three sizes of integers: ordinary, short, and long, and with characters, which behave as if they were small integers. The language definition does not guarantee much about the relative sizes of the various kinds of integer:

1. The four sizes of integers are non-decreasing.
2. An ordinary integer is large enough to contain any array subscript.
3. The size of a character is natural for the particular hardware.

Most modern machines have 8-bit characters, though a few have 7-or 9-bit characters, so characters are usually 7, 8, or 9 bits.

Long integers are usually at least 32 bits long, so that a long integer can be used to represent the size of a file.

Ordinary integers are usually at least 16 bits long, because shorter integers would impose too much of a restriction on the maximum size of an array.

Short integers are almost always exactly 16 bits long.

What does this all mean in practice? The most important thing is that one cannot count on having any particular precision available. Informally, one can probably expect 16 bits for a short or an ordinary integer, and 32 bits for a long integer, but not even those sizes are guaranteed. One can certainly use ordinary integers to express table sizes and subscripts, but what about a variable that must be able to hold values up to ten million?

The most portable way to do that is probably to define a "new'' type:

typedef long tenmil;

7.3. Are Characters Signed or Unsigned?

Most modern computers support 8-bit characters, so most modern C compilers implement characters as 8-bit integers. However, not all compilers interpret those 8-bit quantities the same way.

The issue becomes important only when converting a char quantity to a larger integer. Going the other way, the results are well-defined: excess bits are simply discarded. But a compiler converting a char to an int has a choice: should it treat the char as a signed or an unsigned quantity? If the former, it should expand the char to an int by replicating the sign bit; if the latter, it should fill the extra bit positions with zeroes.

The results of this decision are important to virtually anyone who deals with characters with their high-order bits turned on. It determines whether 8-bit characters are going to be considered to range from -128 through 127 or from 0 through 255. This, in turn, affects the way the programmer will design things like hash tables and translate tables.

If you care whether a character value with the high-order bit on is treated as a negative number, you should probably declare it as unsigned char. Such values are guaranteed to be zero-extended when converted to integer, whereas ordinary char variables may be signed in one implementation and unsigned in another.

Incidentally, it is a common misconception that if c is a character variable, one can obtain the unsigned integer equivalent of c by writing (unsigned) c. This fails because a char quantity is converted to int before any operator is applied to it, even a cast. Thus c is converted first to a signed integer and then to an unsigned integer, with possibly unexpected results. The right way to do it is (unsigned char) c.

7.4. Are Right Shifts Signed or Unsigned?

This bears repeating: a program that cares how shifts are done had better declare the quantities being shifted as unsigned.

7.5. How Does Division Truncate?

Suppose we divide a by b to give a quotient q and remainder r:

q = a / b;
r = a % b;

What relationships might we want to hold between a, b, p, and q?

1. Most important, we want q*b + r == a, because this is the relation that defines the remainder.
2. If we change the sign of a, we want that to change the sign of q, but not the absolute value.
3. We want to ensure that r>=0 and r<b. For instance, if the remainder is being used as an index to a hash table, it is important to be able to know that it will always be a valid index.

These three properties are clearly desirable for integer division and remainder operations. Unfortunately, they cannot all be true at once.

Consider 3/2, giving a quotient of 1 and a remainder of 1. This satisfies property 1. What should be the value of - 3/2? Property 2 suggests that it should be - 1, but if that is so, the remainder must also be -1, which violates property 3. Alternatively, we can satisfy property 3 by making the remainder 1, in which case property 1 demands that the quotient be - 2. This violates property 2.

Thus C, and any language that implements truncating integer division, must give up at least one of these three principles.

Most programming languages give up number 3, saying instead that the remainder has the same sign as the dividend. This makes it possible to preserve properties 1 and 2. Most C implementations do this in practice, also.

However, the C language definition only guarantees property 1, along with the property that |r|< |b| and that r is not less than 0 whenever a is not less than 0 and b > 0. This property is less restrictive than either property 2 or property 3, and actually permits some rather strange implementations that would be unlikely to occur in practice (such as an implementation that always truncates the quotient away from zero).

Despite its sometimes unwanted flexibility, the C definition is enough that we can usually make integer division do what we want, provided that we know what we want. Suppose, for example, that we have a number n that represents some function of the characters in an identifier, and we want to use division to obtain a hash table entry h such that h is not less than 0 and h<HASHSIZE. If we know that n is ever negative, we simply write

h = n % HASHSIZE;

h = n % HASHSIZE;
if (h < 0)
    h += HASHSIZE;

7.6. How Big is a Random Number?

This size ambiguity has affected library design as well. When the only C implementation ran on the PDP-11 computer, there was a function called rand that returned a (pseudo-) random non-negative integer. PDP-11 integers were 16 bits long, including the sign, so rand would return an integer between 0 and 2^15 - 1.

When C was implemented on the VAX-11, integers were 32 bits long. What was the range of the rand function on the VAX-11?

For their system, the people at the University of California took the view that rand should return a value that ranges over all possible non-negative integers, so their version of rand returns an integer between 0 and 2^31 - 1.

The people at AT&T, on the other hand, decided that a PDP-11 program that expected the result of rand to be less than 2^15 would be easier to transport to a VAX-11 if the rand function returned a value between 0 and 2^15 there, too.

As a result, it is now difficult to write a program that uses rand without tailoring it to the implementation.

7.7. Case Conversion

The toupper and tolower functions have a similar history. They were originally written as macros:

#define toupper(c) ((c)+'A'?a')
#define tolower(c) ((c)+'a'?A')

These macros do have one disadvantage, though: when given something that is not a letter of the appropriate case, they return garbage. Thus, the following innocent program fragment to convert a file to lower case doesn't work with these macros:

int c;
while ((c = getchar()) != EOF)
    putchar (tolower (c));

int c;
while ((c = getchar()) != EOF)
    putchar (isupper (c)? tolower (c): c);

#define toupper(c) ((c) >= 'a' && (c) <= 'z'? (c) + 'A' ?'a': (c))
#define tolower(c) ((c) >= 'A' && (c) <= 'Z'? (c) + 'a' ?'A': (c))

int toupper (c)
int c;
{
    if (c >= 'a' && c <= 'z')
        return c + 'A' ?'a';
    return c;
}

This change had the advantage of robustness, at the cost of introducing function call overhead into each use of these functions. Our hero realized that some people might not be willing to pay the cost of this overhead, so he re-introduced the macros with new names:

#define _toupper(c) ((c)+'A'?a')
#define _tolower(c) ((c)+'a'?A')

There was just one problem in all this: the people at Berkeley never followed suit, nor did some other C implementors. This means that a program written on an AT&T system that uses toupper or tolower, and assumes that it will be able to pass an argument that is not a letter of the appropriate case, may stop working on some other C implementation.

This sort of failure is very hard to trace for someone who does not know this bit of history.

7.8. Free First, then Reallocate

Most C implementations provide users with three memory allocation functions called malloc, realloc, and free. Calling malloc(n) returns a pointer to n characters of newly-allocated memory that the programmer can use. Giving free a pointer to memory previously returned by malloc makes that memory available for re-use. Calling realloc with a pointer to an allocated area and a new size stretchesm or shrinks the memory to the new size, possibly copying it in the process.

Or so one might think. The truth is actually somewhat more subtle. Here is an excerpt from the description of realloc that appears in the System V Interface Definition:

Realloc changes the size of the block pointed to by ptr to size bytes and returns a pointer to the (possibly moved) block. The contents will be unchanged up to the lesser of the new and old sizes.

The Seventh Edition of the reference manual for the UNIX system contains a copy of the same paragraph. In addition, it contains a second paragraph describing realloc:

Realloc also works if ptr points to a block freed since the last call of malloc, realloc, or calloc; thus sequences of free, malloc and realloc can exploit the search strategy of malloc to do storage compaction.

Thus, the following is legal under the Seventh Edition:

free (p);
p = realloc (p, newsize);

for (p = head; p != NULL; p = p->next)
free ((char *) p);

Needless to say, this technique is not recommended, if only because not all C implementations preserve memory long enough after it has been freed. However, the Seventh Edition manual leaves one thing unstated: the original implementation of realloc actually required that the area given to it for reallocation be free first. For this reason, there are many C programs floating around that free memory first and then reallocate it, and this is something to watch out for when moving a C program to another implementation.

7.9. An Example of Portability Problems

Let's take a look at a problem that has been solved many times by many people. The following program takes two arguments: a long integer and a (pointer to a) function. It converts the integer to decimal and calls the given function with each character of the decimal representation.

void printnum (n, p)
long n;
void (*p)();
{
    if (n < 0) {
        (*p) ('?);
        n = -n;
    }
    if (n >= 10)
        printnum (n/10, p);
    (*p) (n % 10 + '0');
}

This program, for all its simplicity, has several portability problems. The first is the method it uses to convert the low-order decimal digit of n to character form. Using n%10 to get the value of the low-order digit is fine, but adding '0' to it to get the corresponding character representation is not. This addition assumes that the machine collating sequence has all the digits in sequence with no gaps, so that '0'+5 has the same value as '5', and so on. This assumption, while true of the ASCII and EBCDIC character sets, might not be true for some machines. The way to avoid that problem is to use a table:

void printnum (n, p)
long n;
void (*p)();
{
    if (n < 0) {
        (*p) ('-');
        n =-n;
    }
    if (n >= 10)
        printnum (n/10, p);
    (*p) ("0123456789"[n % 10]);
}

There are several ways around this problem. The most obvious one is to assign n to an unsigned long value and be done with it. However, some C compilers do not implement unsigned long, so let us see how we can get along without it.

In both 1's complement and 2's complement machines, changing the sign of a positive integer is guaranteed not to overflow. The only trouble comes when changing the sign of a negative value. Therefore, we can avoid trouble by making sure we do not attempt to make n positive.

Of course, once we have printed the sign of a negative value, we would like to be able to treat negative and positive numbers the same way. The way to do that is to force n to be negative after printing the sign, and to do all our arithmetic with negative values. If we do this, we will have to ensure that the part of the program that prints the sign is executed only once; the easiest way to do that is to split the program into two functions:

void printnum (n, p)
long n;
void (*p)();
{
    void printneg();
    if (n < 0) {
        (*p) ('-');
        printneg (n, p);
    } else
    printneg (-n, p);
}

void printneg (n, p)
long n;
void (*p)();
{
    if (n<=-10)
    printneg (n/10, p);
    (*p) ("0123456789"[-(n % 10)]);
}

Or have we? We have used n/10 and n%10 to represent the leading digits and the trailing digit of n (with suitable sign changes). Recall that integer division behaves in a somewhat implementation-dependent way when one of the operands is negative. For that reason, it might actually be that n%10 is positive! In that case, -(n%10) would be negative, and we would run off the end of our digit array.

We cater to this problem by creating two temporary variables to hold the quotient and remainder. After we do the division, we check that the remainder is in range and adjust both variables if not. Printnum has not changed, so we show only printneg:

void printneg (n, p)
long n;
void (*p)();
{
    long q;
    int r;
    q = n / 10;
    r = n % 10;
    if (r > 0) {
        r - =10;
        q++;
    }
    if (n <= -10)
        printneg (q, p);
    (*p) ("0123456789"[-r]);
}

There are many ways for C programmers to go astray that have not been mentioned in this paper. If you find one, please contact the author. It may well be included, with an acknowledging footnote, in a future revision.

The C Programming Language (Kernighan and Ritchie, Prentice-Hall 1978) is the definitive work on C. It contains both an excellent tutorial, aimed at people who are already familiar with other high-level languages, and a reference manual that describes the entire language succinctly. While the language has expanded slightly since 1978, this book is still the last word on most subjects. This book also contains the "C Reference Manual'' we have mentioned several times in this paper.

The C Puzzle Book (Feuer, Prentice-Hall, 1982) is an unusual way to hone one's syntactic skills. The book is a collection of puzzles (and answers) whose solutions test the reader's knowledge of C's fine points.

C: A Reference Manual (Harbison and Steele, Prentice Hall 1984) is mostly intended as a reference source for implementors. Other users may also find it useful, particularly because of its meticulous cross references.