Ghostscript C coding guidelines

typedef struct xxx_s {
   ... members ...
} xxx_t;

Unions

Restrictions:

• Use unions only as components of structures, not as typedefs in their own right.
• Don't attempt to initialize unions: not all compilers support this, even though it is in the 1989 ANSI standard.

Expressions

Restrictions:

• Don't assign a larger integer data type to a smaller one without a cast (int_x = long_y).
• It's OK to use the address of a structure or array element (&p->e, &a[i]) locally, or pass it to a procedure that won't store it, but don't store such an address in allocated storage unless you're very sure of what you're doing.
• Don't use conditional expressions with structure or union values. (Pointers to structures or unions are OK.)
• For calling a variable or parameter procedure, use ptr->func(...). Some old code uses explicit indirection, (*ptr->func)(...): don't use this in new code.
• Don't write expressions that depend on order of evaluation, unless the order is created explicitly by use of ||, &&, ?:, ,, or function nesting (the arguments of a function must be evaluated before the function is called). In particular, don't assume that the arguments of a function will be evaluated left-to-right, or that the left side of an assignment will be evaluated before the right.
• Don't mix integer and enumerated types ad lib: treat enumerated types as distinct from integer types, and use casts to convert between the two. (Some compilers generate warnings if you do not do this.)

Statements

Restrictions:

• If you use an expression as a statement, other than an assignment or a function call with void return value, enclose it explicitly in DISCARD().
• The type of the operand of a switch must match the type of the case labels, whether the labels are ints or the members of an enum type. (Use a cast if necessary.)
• It is OK for one case of a switch to "fall through" into another (i.e., for the statement just before a case label not to be a control transfer), but a comment /* falls through */ is required.
• If you are returning an error code specified explicitly (e.g., return gs_error_rangecheck or return e_rangecheck), use return_error() rather than plain return. However, if the program is simply propagating an error code generated elsewhere, as opposed to generating the error, use return (e.g., if (code < 0) return code).

Procedures

Restrictions:

• Provide a prototype for every procedure, and make sure the prototype is available at every call site. If the procedure is local to a file (private), the prototype should precede the procedure, in the same file; if the procedure is global, the prototype should be in a header file.
• If a procedure parameter is itself a procedure, do list its parameter types rather than just using (). For example,

  int foo(int (*callback)(int, int));

  rather than just

  int foo(int (*callback)());

• Don't use the P* macros in new code. (See the Procedures section of Language extensions below for more information.)
• Always provide an explicit return type for procedures, in both the prototype and the definition: don't rely on the implicit declaration as int.
• Don't use float as the return type of a procedure, unless there's a special reason. Floating point hardware typically does everything in double precision internally and has to do extra work to convert between double and single precision.
• Don't declare parameters as being of type float, short, or char. If you do this and forget to include the prototype at a call site, ANSI compilers will generate incompatible calling sequences. Use floatp (a synonym for double, mnemonic for "float parameter") instead of float, and use int or uint instead of short or char.

Standard library

Restrictions:

• Only use library features that are documented in the established ANSI standard (e.g., Harbison & Steele's book). Do not use procedures that are "standards" promulgated by Microsoft (e.g., stricmp), or originate in BSD Unix (e.g., strcasecmp), or were added in later versions of the standard such as C 9X.
• Do not use any features from stdio.h that assume the existence of stdin, stdout, or stderr. See gsio.h for the full list. Instead, use gs_stdin et al.

Language extensions

Scoping

inline

inline is available even if the compiler does not support it. Be aware, however, that it may have no effect. In particular, do not use inline in header files. Instead, use the extern_inline facility described just below.

extern_inline

Compilers that do support inline vary in how they decide whether to (also) compile a closed-code copy of the procedure. Because of this, putting an inline procedure in a header file may produce multiple closed copies, causing duplicate name errors at link time. extern_inline provides a safe way to put inline procedures in header files, regardless of compiler. Unfortunately, the only way we've found to make this fully portable involves a fair amount of boilerplate. For details, please see stdpre.h.

private

Use private instead of static for all file-local procedures, and also for file-local variables defined at the outermost level. However, use static, not private, for variables defined within a procedure.

private is normally #define'd as static. However, it can also be #define'd as empty, which allows profilers to measure all procedures, and 'nm' to list all interesting statically allocated variables, not just public ones.

Scalar types

bool, true, false

bool is intended as a Boolean type, with canonical values true and false. In a more reasonable language, such as Java, bool is an enumerated type requiring an explicit cast to or from int; however, because C's conditionals are defined as producing int values, we can't even define bool as a C enum without provoking compiler warnings.

Even though bool is a synonym for int, treat them as conceptually different types:

• Initialize or set bool variables to true or false, not 0 or 1.
• Use the Boolean operators !, &&, and || only with Booleans. Don't use the idiom !!x to create a Boolean that is true iff x != 0: use x != 0.
• Use an explicit (int) cast to convert a Boolean to an integer.

byte, ushort, uint, ulong

These types are simply shorthands for unsigned char, short, int, long.

floatp

This is a synonym for double. It should be used for, and only for, procedure parameters that would otherwise be float. (As noted above, procedure parameters should not be declared as float.)

bits8, bits16, bits32

These are unsigned integer types of the given width. Use them wherever the actual width matters: do not, for example, use short assuming that it is 16 bits wide.

bits64

****** NOT IMPLEMENTED YET ****** This is an unsigned 64-bit integer type, but it may not be available on all platforms. Any code that uses this type should be surrounded by #if ARCH_HAS_BITS64.

Procedures

P0(), P1(p1), ..., P16(p1, ..., p16) [deprecated]

Nearly all existing code uses these macros for the parameter lists of procedure prototypes, e.g., int proc(P2(int a, int b)), to provide compatibility with pre-ANSI, a.k.a. "traditional" or "K & R", compilers. Since only ANSI compilers are now supported, don't use these macros in new code: use the ANSI standard syntax int proc(int a, int b).

Stylistic conventions

Ghostscript's coding rules cover not only the use of the language, but also many stylistic issues like the choice of names and the use of whitespace. The stylistic rules are meant to produce code that is easy to read. It's important to observe them as much as possible in order to maintain a consistent style, but if you find these rules getting in your way or producing ugly-looking results once in a while, it's OK to break it.

Formatting

Indentation

We've formatted all of our code using the GNU indent program.

indent -bad -nbap -nsob -br -ce -cli4 -npcs -ncs \
   -i4 -di0 -psl -lp -lps somefile.c

does a 98% accurate job of producing our preferred style. Unfortunately, there are bugs in all versions of GNU indent, requiring both pre- and post-processing of the code. The gsindent script in the Ghostscript fileset contains the necessary workarounds.

Put indentation points every 4 spaces, with 8 spaces = 1 tab stop.

For assignments (including chain assignments), put the entire statement on one line if it will fit; if not, break it after a = and indent all the following lines. I.e., format like this:

var1 = value;
var1 = var2 = value;
var1 =
    value;
var1 =
    var2 = value;
var1 = var2 =
    value;

But not like this:

var1 =
var2 = value;

Indent in-line blocks thus:

{
   ... declarations ...
   {{ blank line if any declarations above }}
   ... statements ...
}

Similarly, indent procedures thus:

return_type
proc_name(... arguments ...)
{
   ... declarations ...
   {{ blank line if any declarations above }}
   ... statements ...
}

If a control construct (if, do, while, or for) has a one-line body, use this:

... control construct ...
   ... subordinate simple statement ...

If it has a multi-line body, use this:

... control construct ... {
   ... subordinate code ...
}

If the subordinate code has declarations, see blocks above.

For if-else statements, do this:

if ( ... ) {
   ... subordinate code ...
} else if ( ... ) {
   ... subordinate code ...
} else {
   ... subordinate code ...
}

When there are more than two alternatives, as in the example above, use the above ("parallel") syntax rather than the following ("nested") syntax:

if ( ... ) {
   ... subordinate code ...
} else {
   if ( ... ) {
      ... subordinate code ...
   } else {
      ... subordinate code ...
   }
}

Similarly, for do-while statements, do this:

do {
   ... body ...
} while ( ... condition ... );

Spaces

Do put a space:

• after every comma and semicolon, unless it ends a line;
• around every binary operator other than "->" and ".", although you can omit the spaces around the innermost operator in a nested expression if you like;
• on both sides of the parentheses of an if, for, or while.

Don't put a space:

• at the end of a line;
• before a comma or semicolon;
• after unary prefix operators;
• before the parenthesis of a macro or procedure call.

Parentheses

Parentheses are important in only a few places:

• Around the inner subexpressions in expressions that mix && and ||, even if they are not required by precedence, for example

  (xx && yy) || zz

• Similarly around inner subexpressions in expressions that mix &, |, or shifts, especially if mixing these with other operators, for instance

  (x << 3) | (y >> 5)

• In macro definitions around every use of an argument that logically could be an expression, for example

  ((x) * (x) + (y) * (y))

Anywhere else, given the choice, use fewer parentheses.

For stylistic consistency with the existing Ghostscript code, put parentheses around conditional expressions even if they aren't syntactically required, unless you really dislike doing this. Note that the parentheses should go around the entire expression, not the condition. For instance, instead of

hpgl_add_point_to_path(pgls, arccoord_x, arccoord_y,
   (pgls->g.pen_down) ? gs_lineto : gs_moveto);

use

hpgl_add_point_to_path(pgls, arccoord_x, arccoord_y,
   (pgls->g.pen_down ? gs_lineto : gs_moveto));

Preprocessor

Conditionals

Using preprocessor conditionals can easily lead to unreadable code, since the eye really wants to read linearly rather than having to parse the conditionals just to figure out what code is relevant. It's OK to use conditionals that have small scope and that don't change the structure or logic of the program (typically, they select between different sets of values depending on some configuration parameter), but where possible, break up source modules rather than use conditionals that affect the structure or logic.

Macros

Ghostscript code uses macros heavily to effectively extend the rather weak abstraction capabilities of the C language, specifically in the area of memory management and garbage collection: in order to read Ghostscript code effectively, you simply have to learn some of these macros as though they were part of the language. The current code also uses macros heavily for other purposes, but we are trying to phase these out as rapidly as possible, because they make the code harder to read and debug, and to use the rules that follow consistently in new code.

Define macros in the smallest scope you can manage (procedure, file, or .h file), and #undef them at the end of that scope: that way, someone reading the code can see the definitions easily when reading the uses. If that isn't appropriate, define them in as large a scope as possible, so that they effectively become part of the language. This places an additional burden on the reader, but it can be amortized over reading larger amounts of code.

Try hard to use procedures instead of macros. Use "inline" if you really think the extra speed is needed, but only within a .c file: don't put inline procedures in .h files, because most compilers don't honor "inline" and you'll wind up with a copy of the procedure in every .c file that includes the .h file.

If you define a macro that looks like a procedure, make sure it will work wherever a procedure will work. In particular, put parentheses around every use of an argument within the macro body, so that the macro will parse correctly if some of the arguments are expressions, and put parentheses around the entire macro body. (This is still subject to the problem that an argument may be evaluated more than once, but there is no way around this in C, since C doesn't provide for local variables within expressions.)

If you define macros for special loop control structures, make their uses look somewhat like ordinary loops, for instance:

BEGIN_RECT(xx, yy) {
  ... body indented one position ...
} END_RECT(xx, yy);

If at all possible, don't use free variables in macros -- that is, variables that aren't apparent as arguments of the macro. If you must use free variables, list them all in a comment at the point where the macro is defined.

Comments

The most important descriptive comments are ones in header files that describe structures, including invariants; but every procedure or structure declaration, or group of other declarations, should have a comment. Don't spend a lot of time commenting executable code unless something unusual or subtle is going on.

Naming

Use fully spelled-out English words in names, rather than contractions. This is most important for procedure and macro names, global variables and constants, values of #define and enum, struct and other typedef names, and structure member names, and for argument and variable names which have uninformative types like int. It's not very important for arguments or local variables of distinctive types, or for local index or count variables.

Avoid names that run English words together: "hpgl_compute_arc_center" is better than "hpgl_compute_arccenter". However, for terms drawn from some predefined source, like the names of PostScript operators, use a term in its well-known form (for instance, gs_setlinewidth rather than gs_set_line_width).

Procedures, variables, and structures visible outside a single .c file should generally have prefixes that indicate what subsystem they belong to (in the case of Ghostscript, gs_ or gx_). This rule isn't followed very consistently.

Types

Many older structure names don't have _t on the end, but this suffix should be used in all new code. (The _s structure name is needed only to satisfy some debuggers. No code other than the structure declaration should refer to it.)

Declare structure types that contain pointers to other instances of themselves like this:

typedef struct xxx_s xxx_t;
struct xxx_s {
   ... members ...
   xxx_t *ptr_member_name;
   ... members ...
};

If, to maintain data abstraction and avoid including otherwise unnecessary header files, you find that you want the type xxx_t to be available in a header file that doesn't include the definition of the structure xxx_s, use this approach:

#ifndef xxx_DEFINED
#  define xxx_DEFINED
typedef struct xxx_s xxx_t;
#endif
struct xxx_s {
   ... members ...
};

You can then copy the first 4 lines in other header files. (Don't ever include them in an executable code file.)

Don't bother using const for anything other than with pointers as described below. However, in those places where it is necessary to cast a pointer of type const T * to type T *, always include a comment that explains why you are "breaking const".

Pointers

Use const for pointer referents (that is, const T *) wherever possible and appropriate.

If you find yourself wanting to use void *, try to find an alternative using unions or (in the case of super- and subclasses) casts, unless you're writing something like a memory manager that really treats memory as opaque.

Procedures

In general, don't create procedures that are private and only called from one place. However, if a compound statement (especially an arm of a conditional) is too long for the eye easily to match its enclosing braces "{...}" -- that is, longer than 10 or 15 lines -- and it doesn't use or set a lot of state held in outer local variables, it may be more readable if you put it in a procedure.

Miscellany

Local variables

Don't assign new values to procedure parameters. It makes debugging very confusing when the parameter values printed for a procedure are not the ones actually supplied by the caller. Instead use a separate local variable initialized to the value of the parameter.

If a local variable is only assigned a value once, assign it that value at its declaration, if possible. For example,

int x = some expression ;

rather than

int x;
...
x = some expression ;

Use a local pointer variable like this to "narrow" pointer types:

int
someproc(... gx_device *dev ...)
{
   gx_device_printer *const pdev = (gx_device_printer *)dev;
   ...
}

Compiler warnings

The following section refers to the warnings produced by gcc: your compiler may produce different ones.

It's OK if compilation produces the following warnings:

• <name> might be used uninitialized in this function
• cast discards `const' from pointer target type

The first of these often occurs because the compiler isn't aware of control flow restrictions that guarantee the variable will be initialized before use: if it occurs in new code, check the code carefully, but don't worry about the message. The second is often unavoidable in code that initializes or frees a structure that is otherwise const during its lifetime.

Do work hard to eliminate all warnings other than these, since they often indicate the possible presence of coding errors. In particular, get rid of warnings about parameter passing or initialization that discards const, by using explicit casts.

File structuring

All files

Keep file names within the "8.3" format for portability:

• Use only letters, digits, dash, and underscore in file names.
• Don't assume upper and lower case letters are distinct.
• Put no more than 8 characters before the ".", if any.
• If there is a ".", put between 1 and 3 characters after the ".".

For files other than documentation files, use only lower case letters in the names; for HTML documentation files, capitalize the first letter.

Every code file should start with comments containing

1. a copyright notice,
2. the name of the file in the form of an RCS Id:

   /*Id$: filename.ext $*/

   (using the comment convention appropriate to the language of the file), and

3. a summary, no more than one line, of what the file contains.

If you create a file by copying the beginning of another file, be sure to update the copyright year and change the file name.

Makefiles

Use the extension .mak for makefiles.

For each

#include "xxx.h"

make sure there is a dependency on $(xxx_h) in the makefile. If xxx ends with a "_", this rule still holds, so that if you code

#include "math_.h"

the makefile must contain a dependency on "$(math__h)" (note the two underscores "__").

List the dependencies bottom-to-top, like the #include statements themselves; within each level, list them alphabetically. Do this also with #include statements themselves whenever possible (but sometimes there are inter-header dependencies that require bending this rule).

For compatibility with the build utilities on OpenVMS, always put a space before the colon that separates the target(s) of a rule from the dependents.

General C code

List #include statements from "bottom" to "top", that is, in the following order:

1. System includes ("xxx_.h")
2. gs*.h
3. gx*.h (yes, gs and gx are in the wrong order.)
4. s*.h
5. i*.h (or other interpreter headers that don't start with "i")

Headers (.h files)

In header files, always use the following at the beginning of a header file to prevent double inclusion:

{{ Copyright notice etc. }}

#ifndef <filename>_INCLUDED
#define <filename>_INCLUDED

{{ The contents of the file }}

#endif /* <filename>_INCLUDED */

The header file is the first place that a reader goes for information about procedures, structures, constants, etc., so ensure that every procedure and structure has a comment that says what it does. Divide procedures into meaningful groups set off by some distinguished form of comment.

Source (.c files)

After the initial comments, arrange C files in the following order:

1. #include statements
2. Exported data declarations
3. Explicit externs (if necessary)
4. Forward declarations of procedures
5. Private data declarations
6. Exported procedures
7. Private procedures

Be flexible about the order of the declarations if necessary to improve readability. Many older files don't follow this order, often without good reason.

Ghostscript conventions

Specific names

The Ghostscript code uses certain names consistently for certain kinds of values. Some of the commonest and least obvious are these two:

code

A value to be returned from a procedure:

< 0		An error code defined in gserrors.h (or errors.h)
0		Normal return
> 0		A non-standard but successful return (which must be documented, preferably with the procedure's prototype)

status

A value returned from a stream procedure:

< 0		An exceptional condition as defined in scommon.h
0		Normal return (or, from the "process" procedure, means that more input is needed)
1		More output space is needed (from the "process" procedure)

Structure type descriptors

The Ghostscript memory manager requires run-time type information for every structure. (We don't document this in detail here: see the Structure descriptors section of the developer documentation for details.) Putting the descriptor for a structure next to the structure definition will help keep the two consistent, so immediately after the definition of a structure xxx_s, define its structure descriptor:

struct xxx_s {
   ... members ...
};
#define private_st_xxx()  /* in <filename>.c */\
  gs_private_st_<whatever>(st_xxx, xxx_t,\
    "xxx_t", xxx_enum_ptrs, xxx_reloc_ptrs,\
    ... additional parameters as needed ...)

The file that implements operations on this structure (<filename>.c) should then include, near the beginning, the line:

private_st_xxx();

In much existing code, structure descriptors are declared as public, which allows clients to allocate instances of the structure directly. We now consider this bad design. Instead, structure descriptors should always be private; the implementation file should provide one or more procedures for allocating instances, e.g.,

xxx_t *gs_xxx_alloc(P1(gs_memory_t *mem));

If it is necessary to make a structure descriptor public, it should be declared in its clients as

extern_st(st_xxx);

"Objects"

Ghostscript makes heavy use of object-oriented constructs, including analogues of classes, instances, subclassing, and class-associated procedures. However, these constructs are implemented in C rather than C++, for two reasons:

• The first Ghostscript code was written in 1986, long before C++ was codified or was well supported by tools. Even today, C++ tools rarely support C++ as well as C tools support C.
• C++ imposes its own implementations for virtual procedures, inheritance, run-time type information, and (to some extent) memory management. Ghostscript requires use of its own memory manager, and also sometimes requires the ability to change the virtual procedures of an object dynamically.

Classes

The source code representation of a class is simply a typedef for a C struct. See Structures, above, for details.

Procedures

Ghostscript has no special construct for non-virtual procedures associated with a class. In some cases, the typedef for the class is in a header file but the struct declaration is in the implementation code file: this provides an extra level of opaqueness, since clients then can't see the representation and must make all accesses through procedures. You should use this approach in new code, if it doesn't increase the size of the code too much or require procedure calls for very heavily used accesses.

Ghostscript uses three different approaches for storing and accessing virtual procedures, plus a fourth one that is recommended but not currently used. For exposition, we assume the class (type) is named xxx_t, it has a virtual procedure void (*virtu)(P1(xxx_t *)), and we have a variable declared as xxx_t *pxx.

1. The procedures are stored in a separate, constant structure of type xxx_procs, of which virtu is a member. The structure definition of xxx_t includes a member defined as const xxx_procs *procs (always named procs). The construct for calling the virtual procedure is pxx->procs->virtu(pxx).
2. The procedures are defined in a structure of type xxx_procs as above. The structure definition of xxx_t includes a member defined as xxx_procs procs (always named procs). The construct for calling the virtual procedure is pxx->procs.virtu(pxx).
3. The procedures are not defined in a separate structure: each procedure is a separate member of xxx_t. The construct for calling the virtual procedure is pxx->virtu(pxx).
4. The procedures are defined in a structure of type xxx_procs as above. The structure definition of xxx_t includes a member defined as xxx_procs procs[1] (always named procs). The construct for calling the virtual procedure is again pxx->procs->virtu(pxx).

Note that in approach 1, the procedures are in a shared constant structure; in approaches 2 - 4, they are in a per-instance structure that can be changed dynamically, which is sometimes important.

In the present Ghostscript code, approach 1 is most common, followed by 2 and 3; 4 is not used at all. For new code, you should use 1 or 4: that way, all virtual procedure calls have the same form, regardless of whether the procedures are shared and constant or per-instance and mutable.

Subclassing

Ghostscript's class mechanism allows for subclasses that can add data members, or can add procedure members if approach 1 or 3 (above) is used. Since C doesn't support subclassing, we use a convention to accomplish it. In the example below, gx_device is the root class; it has a subclass gx_device_forward, which in turn has a subclass gx_device_null. First we define a macro for all the members of the root class, and the root class type. (As for structures in general, classes need a structure descriptor, as discussed in Structures above: we include these in the examples below.)

#define gx_device_common\
    type1 member1;\
    ...
    typeN memberN

typedef struct gx_device_s {
    gx_device_common;
} gx_device;

#define private_st_gx_device()  /* in gsdevice.c */\
  gs_private_st_<whatever>(st_gx_device, gx_device,\
    "gx_device", device_enum_ptrs, device_reloc_ptrs,\
    ... additional parameters as needed ...)

We then define a similar macro and type for the subclass.

#define gx_device_forward_common\
    gx_device_common;\
    gx_device *target

typedef struct gx_device_forward_s {
    gx_device_forward_common;
} gx_device_forward;

#define private_st_device_forward()  /* in gsdevice.c */\
  gs_private_st_suffix_add1(st_device_forward, gx_device_forward,\
    "gx_device_forward", device_forward_enum_ptrs, device_forward_reloc_ptrs,\
    gx_device, target)

Finally, we define a leaf class, which doesn't need a macro because we don't currently subclass it. (We can create the macro later if needed, with no changes anywhere else.) In this particular case, the leaf class has no additional data members, but it could have some.

typedef struct gx_device_null_s {
    gx_device_forward_common;
};

#define private_st_device_null()  /* in gsdevice.c */\
  gs_private_st_suffix_add0_local(st_device_null, gx_device_null,\
    "gx_device_null", device_null_enum_ptrs, device_null_reloc_ptrs,\
    gx_device_forward)

Note that the above example is not the actual definition of the gx_device structure type: the actual type has some additional complications because it has a finalization procedure. See src/gxdevcli.h for the details.

If you add members to a root class (such as gx_device in this example), or change existing members, do this in the gx_device_common macro, not the gx_device structure definition. Similarly, to change the gx_device_forward class, modify the gx_device_forward_common macro, not the structure definition. Only change the structure definition if the class is a leaf class (one with no _common macro and no possibility of subclassing), like gx_device_null.

Error handling

Every caller should check for error returns and, in general, propagate them to its callers. By convention, nearly every procedure returns an int to indicate the outcome of the call:

< 0		Error return
0		Normal return
> 0		Non-error return other than the normal case

To make a procedure generate an error and return it, as opposed to propagating an error generated by a lower procedure, you should use

return_error(error_number);

Sometimes it is more convenient to generate the error in one place and return it in another. In this case, you should use

code = gs_note_error(error_number);
...
return code;

In executables built for debugging, the -E (or -Z#) command line switch causes return_error and gs_note_error to print the error number and the source file and line: this is often helpful for identifying the original cause of an error.

See the file src/gserrors.h for the error return codes used by the graphics library, most of which correspond directly to PostScript error conditions.

Copyright © 1996, 1997, 1998 Aladdin Enterprises. Copyright © 2001 artofcode LLC. All rights reserved.

This file is part of AFPL Ghostscript. See the Aladdin Free Public License (the "License") for full details of the terms of using, copying, modifying, and redistributing AFPL Ghostscript.

Ghostscript version 7.03, 20 October 2001