The question has been asked multiple times, "Why is APR using Incomplete
types?" This document will try to explain that.
Incomplete types are used in APR because they can enforce portability, and
they make the APR developers job easier, as well as allowing APR to use native
types on all platforms. Imagine a scenario where APR wasn't using incomplete
types. The ap_file_t type would have to be defined as:
typedef struct ap_file_t {
ap_pool_t *pool
char *fname;
int eof_hit;
int pipe;
ap_interval_time_t timeout;
#ifdef WIN32
HANDLE file_handle;
DWORD dwFileAttributes;
#elif defined(OS2)
HFILE filedes;
HEV PipeSem
#else
int filedes;
int ungetchar;
#endif
#ifndef WIN32
int buffered;
ap_int32_flags
int isopen;
/* Stuff for buffered mode */
char *buffer;
int bufpos;
unsigned long dataRead;
int direction;
unsigned long filePtr;
ap_lock_t *mutex;
#endif
} ap_file_t;
This captures the essense of what is currently being defined for ap_file_t
using incomplete types. However, using this structure leads developers to
believe that they are safe accessing any of the fields in this structure.
This is not true. On some platforms, such as Windows, about half of the
structure disappears. We could combine some of these definitions with
macros, for example:
#ifdef WIN32
#define filetype HANDLE
#elif OS2
#define filetype HFILE
#else
#define filetype int
#endif
And then in the defintion for ap_file_t, we could say:
filetype filedes;
This gets rid of some of the complexity, by moving it off to the side, but
it is still not safe for a programmers to access the filedes field directly
outside of APR, because the programmer has no way of knowing what the actual
type is. So for example printing the filedes using printf would yield wildly
varying results on Windows and OS2 when compared to Unix.
Another option also presents itself. Stick strictly to POSIX. This means
that all code can be shared on any POSIX compliant platform. The problem
with this is performance. One of the benefits to APR, is that it allows
developers to easily use native types on all platforms with the same code.
This has proven to provide a substantial performance boost on most non-Unix
platforms.
Having said all of that, sometimes incomplete types just don't make sense.
For example, the first implementation of time functions used incomplete types,
which added a layer of complexity that turned out to be unnecessary. If
a platform cannot provide a simple number that represents the number of seconds
elapsed since a specifed date and time, then APR doesn't really want to
provide support for that platform.
APR is trying hard to provide a balance of incomplete and complete types,
but like all things, sometimes the developers make mistakes. If you are
using APR and find that there is an incomplete type that doesn't need to be
an incomplete type, please let us know, we are more than willing to listen
and design parts of APR that do not use incomplete types.