Coding style

From K5Wiki
Revision as of 12:18, 11 April 2008 by TomYu (Talk | contribs)

Jump to: navigation, search

The C language Coding style described here is based on the BSD coding style, with some additional elements from the GNU coding standards and the SunOS coding standards.


External links

Old version

Old content to be moved elsewhere is below.


The code in the krb5 source tree largely follows BSD KNF (/usr/share/misc/style on NetBSD) except that it uses a four column basic offset. The style described here is a synthesis of BSD KNF and the GNU coding standards for the C language. The formatting described in the "Formatting Your Source Code" section of the GNU coding standards is mostly what we want, except we use BSD brace style and BSD-ish conventions for the spacing around operators.

Formatting style for C code

In general, use a four column basic offset, tab stops at eight columns. Indents should be tabified, i.e. continuous tabs followed by spaces if necessary if the indent isn't a multiple of eight columns. The "bsd" style in emacs cc-mode mostly does the right thing. You can use "M-x c-set-style" "bsd" to get this. Alternatively, you can use the "krb5" style that is included here.

Labels, including case labels, are outdented by four columns. Continuations of statements are indented by an additional four columns. When continuing expressions this way, split the expression so that the newline goes before a binary operator rather than after it.

Continuations of argument lists or parenthesized expressions should line up with the column after the opening parenthesis. Note that this may create width problems if you call a fuction deep in a bunch of nested control flow statements. Regardless, any expression split between lines should stil be split so that the newline goes before a binary operator rather than after it.

The maximum width should be 79 columns. If you need more than this, consider rewriting the code so that it fits in 79 columns, since control flow that is nested deeply enough to require excessive width is also likely to be difficult to understand if not broken up. Exceptions may be made for long strings, though ANSI C string concatenation should work around that as well.

Function names for definitions should start on column zero, on the line following their return type name, e.g.

char *
foo(int a)
    /* ... */

The opening brace of a function definition should also be on column zero.

Braces that open substatements, such as those following "if", "else", "while", "for", "do", and "switch", should be on the same line as the begining of the statement. This is sometimes called "hanging" braces. The corresponding closing brace should be at the same indentation level as the beginning of the statement.

The "while" keyword in a do-while construct should sit on the same line as the closing brace of the substatement following "do":

do {
} while (0);

If there is an "if" statement immediately following an "else" keyword, it should go on the same line immediately after the "else":

if (x) {
} else if (y) {

Comments to the right of code start in column 32. Comments not to the right of code are indented at the prevailing indent for the surrounding code. Make the comments complete sentences. If you need more than one line, make them into block comments, like this:

 * This is a block comment.  It should consist of complete
 * sentences.
 * Paragraphs should be separated by blank lines so that emacs
 * fill commands will work properly.

Really important single-line comments should also be done in block form:

 * This is a really important one-line comment.

In order to get the start and end delimiters for block comments to stay when you use emacs to fill paragraphs in the comments, set both the c-hanging-comment-starter-p and the c-hanging-comment-ender-p variables to nil. This will be done by the tentative "krb5" style for the emacs cc-mode.

Since we are mostly aiming for C '89 compatibility, don't use "//" comments.

Spaces go after keywords, but not after function names. Do not, however, put a space after sizeof. Don't put a space after a cast operator, either. Spaces also do not go before parentheses that are argument lists for function calls even if the function call is through a pointer. Spaces go after commas in argument lists, as well as commas that are comma operators. Spaces also go between parts in a for loop, except for "forever" type loops. Use for statements rather than while statements to create forever loops.

if (x) {
    p = calloc(1024, sizeof(int));
cp = (*elem->fp)(1024);
for (i = 0; i < 10; i++);
for (;;) {
/* ... */

Binary operators get spaces, unary ones do not. Prefix and postfix operators also do not get spaces. The structure member operators "." and "->" count as postfix operators syntactically, not as binary operators.

x = --a + b / c - d++;
y = p->z.v[x];

Put spaces around the "?" and ":" in a conditional expression.

x = y ? f() : g();

In general, do not parenthesize the argument of "return".

Coding practices for C

Assume, for most purposes, working ANSI/ISO C ('89, not '99) support, both for internal use and for applications compiling against Kerberos header files and libraries. Some exceptions are noted below.

Do not use assignments as truth values. Rather than this:

/* bad style */
if ((retval = krb5_foo()))
    /* ... */;

do this:

/* better style */
retval = krb5_foo();
if (retval)
    /* ... */;

This makes the code easier to read, and also makes it easier to use debuggers. It may be excusable to put assignments into the conditional espression of a "while" statement, though, like:

while ((ch = getopt(argc, argv, "abn")) != -1)
    /* ... */;

Using assignments as truth values in conditional expressions may make code particularly impenetrable.

There are at least three types of "zero" known to C. These are the integer zero (0), the null pointer constant (NULL), and the character constant zero ('\0'). Yes, these are usually all technically the integer zero. Use them in their correct contexts. (Purists will point out that 0 is a valid null pointer constant; still, do not use 0 to specify a null pointer constant. For further unconfusion, read the section on null pointer constants in the C FAQ.) Do not use a lone variable as a truth value unless it's of integer type. Thus:

int i;
char *cp;
/* ... */
if (i)
    /* ... */;
if (cp != NULL) {
    while (*cp != '\0')
        /* ... */;

Do not cast uses of NULL unless you're calling a function with a variable number of arguments, in which case you should cast it to to the appropriate pointer type. Likewise, do not cast the return value from malloc() and friends; the prototype should declare them properly as returning a void * and thus shouldn't require an explicit cast.

Do not assume that realloc(NULL, size) will do the right thing, or that free(NULL) will do the right thing. ANSI guarantees that it will, but some old libraries (hopefully becoming obsolete) don't. Also, don't assume that malloc(0) will return a non-NULL pointer. Typically, though, the output of malloc(0) will be safe to pass to realloc() and free().

In any case, reading the section in the C FAQ on null pointers is highly recommended to remove confusion regarding null pointers in C, since this is a subject of much confusion to even experienced programmers. In particular, if you do not understand why using calloc() to allocate a struct that contains pointer members or why calling memset() to initialize such a struct to all-bytes-zero is wrong, reread that section again. (Note that there are *lots* of examples of code in the krb5 source tree that erroneously calls memset() to zero a struct, and we should fix these somehow eventually.)

Control flow statements that have a single statement as their body should nevertheless have braces around their bodies if the body is more than one line long, especially in the case of stacked multiple if-else clauses; use:

if (x) {
    if (y)

instead of:

/* bad style */
if (x)
    if (y)

which, while legible to the compiler, may confuse human readers and make the code less maintainable, especially if new branches get added to any of the clauses.

Also, you should almost never intersperse conditional compilation directives with control flow statements, as some combination of

  1. define'd symbols may result in statements getting eaten by dangling

bits of control flow statements. When it is not possible to avoid this questionable practice (you really should rewrite the relevant code section), make use of redundant braces to ensure that a compiler error will result in preference to incorrect runtime behavior (such as inadvertantly providing someone with a root shell).

Do not intersperse conditional compilation directives with control flow statements in such a way that confuses emacs cc-mode. Not only does emacs get confused, but the code becomes more difficult to read and maintain. Therefore, avoid code like this:

    /* bad style */
    if (x) {
#ifdef FOO
    else if (y) {
    else {

Put comments after conditional compilation directives such as "#else" and "#endif". Make them correspond to the sense of the value that controls the compilation of the section they are closing, i.e.

#ifdef FOO
/* ... */
#else /* !FOO */
/* ... */
#endif /* !FOO */

Also, in the case of more complex conditional compilation directives, write the comments like this:

#if defined(FOO) || defined(BAR)
/* ... */
#else /* !(defined(FOO) || defined(BAR)) */
/* ... */
#endif /* !(defined(FOO) || defined(BAR)) */

If you are writing a do-while loop that has only one statement in its body, put braces around it anyway, since the while clause may be mistaken for a while loop with an empty body. Don't do this:

/* bad style */
while (x);

Instead, write this:

/* better style */
do {
} while (x);

While it is syntactically correct to call through a function pointer without applying a dereference operator to it, do not write code that does this. It is easier to see that the function call is actually taking place through a function pointer if you write an explicit dereference. However, do not explicitly take the address of a function in order to assign it to a function pointer, since a function name degrades into a pointer. Thus:

int (*fp)(void);
int foofunc(void);
fp = foofunc;
x = (*fp)();

In general, do not take the address of an array. It does not return a pointer to the first element; it returns a pointer to the array itself. These are often identical when cast to an integral type, but they are inherently of different types themselves. Functions that take array types or pointers to array types as arguments can be particularly trouble-prone.

If a function is declared to return a value, do not call "return" without an argument or allow the flow of control to fall off the end of the function.

Always declare the return type of a function, even if it returns int. Yes, this means declaring main() to return int, since main() is required to return int by the standard. If a function is not supposed to return a value, declare it as returning void rather than omitting the return type, which will default the return type to int.

Try to use ANSI C prototype-style function definitions in preference to K&R style definitions. When using K&R style function definitions, declare all the argument types, even those that are int, but beware of any narrow types in the argument list.

Do not declare variables in an inner scope, e.g. inside the compound substatement of an if statement, unless the complexity of the code really demands that the variables be declared that way. In such situations, the function could probably stand to be broken up into smaller chunks anyway. Do not declare variables in an inner scope that shadow ones in an outer scope, since this leads to confusion. Also, some debugging environments, such as gdb under Solaris, can't see variables declared in an inner scope, so declaring such variables will make maintenance more difficult as well.

Parenthesize expressions that may be confusing, particularly where C's precedences are broken. For example, the shift operators have lower precedence than the +, -, *, /, and % operators. Perhaps the most familiar C precedence quirk is that equality and relational operators are of higher precedence than assignment operators. Less well known is that the bitwise operators are of a lower precedence than equality and relational operators.

The sizeof operator takes either a unary expression or a parenthesized type name. It is not necessary to parenthesize the operand of sizeof if it is applied to a unary expression, but still, always parenthesize the operand of the sizeof operator. The sizeof operator does not evaluate its operand if it is a unary expression, so usages such as

s = sizeof(++foo);

should be avoided for the sake of sanity and readability.

Don't pass around structures except by address. We may relax this restriction for non-API function, though.

For new functions, input parameters should go before output parameters in the call signature. There are exceptions, such as a context-like parameter.

Every function should have block comment preceding it describing briefly in complete sentences what it does, what inputs and outputs it has, and what error codes it can return. It should also describe any unsual aspects of the function. At some point we will want to put some of this information into a machine-parsable form.

Macros should have all-uppercase names. If it is necessary to use multiple statements, use braces, and wrap the whole thing in a do-while(0) construct, such as

#define FOOMACRO(x, y) do {                     \
    foo = (x) + (y);                            \
    f(y);                                       \
} while (0)

Leave off the semicolon at the end of a function-like macro, so that it can be mostly used like a call to a function without a return value. Line up the backslashes to make it more readable. Use M-x c-backslash-region in emacs to do neat lined-up backslashes. Parenthesize uses of arguments in the replacement text of a macro in order to prevent weird interactions.

Strive to make your code capable of compiling using "gcc -Wall -Wmissing-prototypes -Wtraditional -Wcast-qual -Wcast-align -Wconversion -Waggregate-return -pedantic" [XXX need to rethink this somewhat] without generating any errors or warnings. Do not, however, compile using the "-ansi" flag to gcc, since that can result in odd behavior with header files on some systems, causing some necessary symbols to not be defined.


The C standard reserves a bunch of namespaces for the implementation. Don't stomp on them. For practical purposes, any identifier with a leading underscore should not be used. (Technically, ^_[a-z].* are reserved only for file scope, so should be safe for things smaller than file scope, but it's better to be paranoid in this case.)

POSIX reserves typedef names ending with _t as well.

Recall that errno is a reserved identifier, and is permitted to be a macro. Therefore, do not use errno as the name of a structure member, etc.

Reserved namespaces are somewhat more restricted than this; read the appropriate section of the C standard if you have questions.

If you're writing new library code, pick a short prefix and stick with it for any identifier with external linkage. If for some reason a library needs to have external symbols that should not be visible to the application, pick another (related) prefix to use for the internal globals. This applies to typedef names, tag names, and preprocessor identifiers as well.

For the krb5 library, the prefix for public global symbols is "krb5_". Use "krb5int_" as a prefix for library internal globals. Avoid using "__" in symbol names, as it may confuse C++ implementations. There are admittedly a number of places where we leak thing into the namespace; we should try to fix these.

Header files should also not leak symbols. Usually using the upcased version of the prefix you've picked will suffice, e.g. "KRB5_" as a CPP symbol prefix corresponding to "krb5_". In general, do not define macros that are lowercase, in order to avoid confusion and to prevent namespace collisions.

The C standard only guarantees six case-insensitive characters to be significant in external identifiers; this is largely regarded as obsolescent even in 1989 and we will ignore it. It does, however, only guarantee 31 case-sensitive characters to be signficant in internal identifiers, so do not use identifiers that differ beyond the 31st character. This is unlikely to be a problem, though.

Aspects of C style in GNU coding std but not here

  • redundant parens to force extra indent of operators of different precedences
  • redundant parens to force general extra indent of expressions that are broken between lines
  • use of ^L characters to break up source files into pages
  • nitpicking about capitalization in comments of variable names when their values are meant
  • commenting usages of static variables
  • casts to void
  • separation of word in names with underscores vs case change
  • enum vs #define'd integer constants
  • 14 char filename limits, MS-DOS filename limits
  • portability
  • system library function quirks
  • internationalization
  • mmap()

Aspects of C style in BSD KNF but not here

  • sorting of header files
  • sorting of struct members
  • separating struct tag decl and struct typedef
  • sorting of var decl
  • lining up var names in decls
  • newline after decls
  • usage of __P
  • usage of getopt
  • not initializing vars in decls
  • stdarg/varargs handling

Emacs cc-mode style

Putting the following code in your .emacs file will result in mostly the right thing happening with respect to formatting style. Note that you may want to turn on auto-newline feature of cc-mode, though that seems to have some bugs with brace-elseif-brace handling at least in the version of cc-mode that comes with emacs 20.3.

       (defconst krb5-c-style
            brace-elseif-brace brace-else-brace defun-close-semi)
           (c-comment-continuation-stars . "* ")
           (c-electric-pound-behavior alignleft)
            (class-open after)
            (substatement-open after)
            (block-close . c-snug-do-while)
            (extern-lang-open after))
            (case-label after)
            (label after))
           (c-indent-comments-syntactically-p . t)
           (c-label-minimum-indentation . 0)
       (defun krb5-c-hook ()
         (c-add-style "krb5" krb5-c-style t))
       (add-hook 'c-mode-common-hook 'krb5-c-hook)

indent.pro settings

The following settings for the indent program should produce a reasonable approximation to the C coding style described here, though some manual cleanup may be necessary. Note that the gindent installed in the gnu locker does not currently handle -psl correctly though.

-bap -br -ce -ci4 -cli0 -d0 -di8 -i4 -ip4 -l79 -nbc -ncdb -ndj -nfc1 -lp -npcs -psl -sc -sob


[XXX needs to be written]


[XXX needs to be written]

Personal tools