Object-oriented techniques in C

Although the techniques described below won't be very popular nowadays (after all, why bother using C if we're going to write object-oriented code?), they are still quite relevant for embedded systems and other low-level things (kernel development, etc). For example, lots of Linux Kernel subsystems use similar approach.

I work with embedded stuff a lot: that is, I have a microcontroller with limited resources and limited set of toolchains that I can use. Usually it's just C compiler provided by chip manufacturer.

I also have experience in modern techniques, such as Java and Qt, and I think in object-oriented way. Don't get me wrong, I don't consider it as “the only right way” to implement everything: as ever, we have to apply common sense. However, it is a decent tool that comes in handy in many cases in order to build modular software, if we use it right.

So, of course I really want to use object-oriented approach in the embedded world, when it is appropriate, and the techniques I'm going to tell you about worked quite well for me. As an example, here's a class diagram (drawn by PlantUML) of a subsystem of my actual embedded project in C (clickable of course) :

It is considerably large. The resulting system works stably and is flexible enough. I have no idea how would I implement all of this to the same degree of flexibility without object-oriented approach.

There are C++ compilers for some embedded CPUs, but they are uncommon, so I need to stick to C, in the name of portability.

There are already publications on the topic (for example, see this very thorough book), but none of the existing solutions satisfies me. Requirements that I have:

  • Avoid typecasts as much as possible, so we don't lose type safety;
  • Polymorphism: we should be able to use virtual methods, and client of the class should not be aware whether some particular method is virtual or not;
  • Multiple inheritance: I don't use it often, but sometimes I really want some class to implement multiple interfaces (or to extend multiple superclasses).

Most of the existing solutions use typecasts heavily, so we lose type safety: compiler won't help you if you make a mistake. This is completely unacceptable.

So I decided to share my experience. Of course it is not going to be too smooth, but it is robust and architecturally clean, so it helps me a lot in my daily programming life.

Whether to use it or not is a subject to debate (if not a holy war), and I don't expect everyone to agree with me. The rule of thumb for me: don't overuse it. If you need to develop a car multimedia system, it's much better to get a more expensive chip which can run Linux, and then write an application in C++, which will run in userspace. But if it's rather little project on a two-dollar MCU, it might be the way to go. Apply common sense.

Concrete class

Please note that things I'm going to tell you in this section are pretty trivial if you're familiar with basic OO techniques.

Let's start from very simple yet realistic example: the CRC calculator. In fact, it was my first experience of OO in C, ages ago. Back then, I was working on firmware upgrade for device via bluetooth, so I wanted to have some common C module that I can use on both parties (device and host computer).

The usual C approach to calculate things like CRC is to have one-function module that calculates crc of the buffer. Function prototype looks like this:

uint32_t crc32_calc(const uint8_t *buf, size_t size);

This function may work at the host side, where we have lots of RAM. However, such function isn't too useful if we haven't the whole buffer at once. After firmware upgrade is completed, device's bootloader should compare calculated CRC of programmed data with the one received from host. If they match, operation considered successful. Even if I wanted to calculate CRC “at once” by calling crc32_calc(), I couldn't, because flash space on MCU I worked with isn't easily mapped to address space, and of course the MCU hasn't enough RAM so that I could read the whole firmware to RAM and then calculate CRC.

Instead, I'd like to be able to calculate CRC gradually: when next frame (say, 64 bytes) is received from host, the device programs it to flash, then reads it back, and accumulates (“feeds”) this data to “global” CRC. When next frame is received, it is fed to global CRC again, and so on. Eventually, we have CRC of the whole firmware.

For that to be done, OO-approach comes to the rescue. We want to have object Crc which has two methods:

  • byte_feed(uint8_t byte) : feed next byte;
  • value_get() : get current crc32 value.

Since C doesn't support object-oriented programming, we have to manually pass pointer to the object for which method is called. To avoid useless confusion, I use the name me instead of this.

Here's the easiest interface for this, ever:

crc32.h
/*******************************************************************************
 *    CRC32 calculator
 ******************************************************************************/
 
#ifndef _CRC32_H
#define _CRC32_H
 
/*******************************************************************************
 *    INCLUDED FILES
 ******************************************************************************/
 
#include <stdint.h>
 
 
/*******************************************************************************
 *    PUBLIC TYPES
 ******************************************************************************/
 
/**
 * CRC32 object context.
 *
 * In fact, we could even put it into source file instead of header file,
 * but then we can allocate it in heap only, which might not be good idea,
 * especially in embedded world.
 */
struct S_Crc32 {
   uint32_t    crc;
};
 
 
/*******************************************************************************
 *    CONSTRUCTOR, DESTRUCTOR
 ******************************************************************************/
 
void crc32_ctor(struct S_Crc32 *me);
void crc32_dtor(struct S_Crc32 *me);
 
 
/*******************************************************************************
 *    PUBLIC METHOD PROTOTYPES
 ******************************************************************************/
 
/**
 * Feed next byte to crc32
 */
void crc32_byte_feed(struct S_Crc32 *me, uint8_t byte);
 
/**
 * Get current crc32 value
 */
uint32_t crc32_value_get(const struct S_Crc32 *me);
 
 
/*******************************************************************************
 *    end of file
 ******************************************************************************/
 
#endif  /* _CRC32_H */

Implementation is pretty straightforward as well:

crc32.c
/*******************************************************************************
 *    INCLUDED FILES
 ******************************************************************************/
 
#include "crc32.h"
 
 
/*******************************************************************************
 *    CONSTRUCTOR, DESTRUCTOR
 ******************************************************************************/
 
void crc32_ctor(struct S_Crc32 *me)
{
   me->crc = 0xffffffff;
}
 
void crc32_dtor(struct S_Crc32 *me)
{
   //-- nothing to do here
}
 
 
/*******************************************************************************
 *    PUBLIC METHODS
 ******************************************************************************/
 
/**
 * Feed next byte to crc32
 */
void crc32_byte_feed(struct S_Crc32 *me, uint8_t byte)
{
   static const uint32_t _crc32_table[16] = {
      0x00000000, 0x1db71064, 0x3b6e20c8, 0x26d930ac,
      0x76dc4190, 0x6b6b51f4, 0x4db26158, 0x5005713c,
      0xedb88320, 0xf00f9344, 0xd6d6a3e8, 0xcb61b38c,
      0x9b64c2b0, 0x86d3d2d4, 0xa00ae278, 0xbdbdf21c,
   };
 
   me->crc ^= byte;
   me->crc = _crc32_table[me->crc & 0x0f] ^ (me->crc >> 4);
   me->crc = _crc32_table[me->crc & 0x0f] ^ (me->crc >> 4);
}
 
/**
 * Get current crc32 value
 */
uint32_t crc32_value_get(const struct S_Crc32 *me)
{
   return me->crc ^ 0xffffffff;
}

Now, we can use crc32 calculator object as follows:

main.c
#include "crc32.h"
#include <stdint.h>
 
int main(void)
{
   //-- Allocate crc_calc object and construct it
   struct S_Crc32 crc_calc;
   crc32_ctor(&crc_calc);
 
   //-- Get bytes from wherever you want, and feed them to crc_calc
   //   (in this example, bytes are just hardcoded)
   crc32_byte_feed(&crc_calc, 0x01);
   crc32_byte_feed(&crc_calc, 0x02);
   crc32_byte_feed(&crc_calc, 0x03);
   crc32_byte_feed(&crc_calc, 0x04);
 
   //-- get crc32 value
   uint32_t crc_val = crc32_value_get(&crc_calc);
 
   //-- destruct crc_calc
   crc32_dtor(&crc_calc);
}

Great, that's just what we need: we can get data from wherever we want, and feed data byte-by-byte to CRC calculator.

By the way, the module for bmp screenshot generation, Bmp writer, uses similar approach, so you might want to check its API as well.

Now, let's proceed to something more interesting.

Inheritance and polymorphism

I use inheritance when I want some objects to share common interface.

Approach that I use in C offers 1-level inheritance quite well, but maintaining hierarchy of more levels becomes too verbose. Luckily, 1-level inheritance is usually quite enough to implement common interface that is shared by multiple objects, so, it's not a major problem for me. However, sometimes I use more levels, when I really need to. The only problem with it is that it becomes too verbose.

When some class (Derived) is derived from another class (Base), then, in terms of C, struct Derived contains struct Base. Like this:

/* base class */
typedef struct S_Base {
   /* ... */
} T_Base;
 
/* derived class */
typedef struct S_Derived {
   /* ... */
   T_Base base;
   /* ... */
} T_Derived

In order to achieve polymorphism, we need to manually do what C++ compiler does behind the scenes: maintain virtual functions table. For each class (but not for each instance of class), we're going to have struct with pointers to virtual functions.

At least, the following items should be virtual:

  • destructor
  • memory deallocator

Virtual destructor is quite common thing: when we destruct an instance of subclass, then destructor of derived class should be called first, and then destructor of base class should be called. We also need virtual memory deallocator, since pointer to derived class might not match the pointer to base class.

I'll go ahead and tell you that I've prepared small tool that generates boilerplate “class” code for you, so that you don't have to type all of the following stuff yourself. Here it is: https://github.com/dimonomid/ooc. At the end of this article, I'll explain usage of it in more detail.

Base class

Assume we have base class MyBase, which is going to have some derived classes. Apart from destructor and memory deallocator, it has two virtual methods: some_method(int a, int b) and other_method(void). We need to define prototypes for all virtual methods:

/*******************************************************************************
 *    VTABLE
 ******************************************************************************/
 
/**
 * Destructor
 */
typedef void (T_MyBase_Dtor)                (T_MyBase *me);
 
/**
 * Memory deallocator. Needs to be overridden because pointer to subclass not
 * necessarily matches pointer to superclass.
 */
typedef void (T_MyBase_Free)                (T_MyBase *me);
 
/**
 * Some method
 */
typedef void (T_MyBase_SomeMethod)          (T_MyBase *me, int a, int b);
 
/**
 * Other method
 */
typedef void (T_MyBase_OtherMethod)          (T_MyBase *me);

Note that each of them explicitly takes pointer to the object for which the method is called. Now, create virtual functions table (vtable) :

/**
 * Virtual functions table (vtable)
 */
typedef struct S_MyBase_VTable {
   T_MyBase_Dtor            *p_dtor;
   T_MyBase_Free            *p_free;
   T_MyBase_SomeMethod      *p_some_method;
   T_MyBase_OtherMethod     *p_other_method;
} T_MyBase_VTable;

And object context of MyBase should contain a pointer to that vtable:

struct S_MyBase {
   /* ... */
   struct S_MyBase_VTable *p_vtable;
   /* ... */
};

Obviously, we should implement these virtual methods for MyBase:

static void _some_method(T_MyBase *me, int a, int b)
{
   /* ... */
}
 
static void _other_method(T_MyBase *me)
{
   /* ... */
}
 
static void _dtor(T_MyBase *me)
{
   /* ... */
}
 
static void _free(T_MyBase *me)
{
   free(me);
}

The question now is: how to initialize vtable. First of all, I decided to make it const, like this:

/* not so good: */
static const T_MyBase_VTable _vtable = {
   .p_dtor         = _dtor,
   .p_free         = _free,
   .p_some_method  = _some_method,
   .p_other_method = _other_method,
};

This approach works, but it has serious drawback for derived classes: when we define vtable for derived class, we might or might not want to override some particular method. If we do, that's no problem: we just specify pointer to function that overrides that from base class. But if we don't, then:

  • We need to get pointer to base function somehow. So, these functions can no longer be static, and all subclasses should have access to them;
  • When we add some new virtual method to vtable, we need to manually modify vtable of all subclasses. This practice is error-prone;
  • It violates the DRY principle (Don't Repeat Yourself).

After some research, I decided to make it non-const, and use lazy initialization. This approach worked for me for years, and I'm quite happy with it. So, for MyBase, we're going to have vtable and bool variable indicating status of vtable (initialized or not) :

static T_MyBase_VTable _vtable;
static bool _bool_vtable_initialized = 0;

We need function that initializes vtable if needed:

static void _vtable_init()
{
   if (!_bool_vtable_initialized){
      _vtable.p_dtor           = _dtor;
      _vtable.p_free           = _free;
 
      _vtable.p_some_method    = _some_method;
      _vtable.p_other_method   = _other_method;
 
      _bool_vtable_initialized = true;
   }
}

And, finally, this function needs to be called from constructor. So, when first instance of the class MyBase is created, its vtable gets initialized. When more instances are created, vtable isn't modified anymore.

Let's get to constructor. The constructor usually needs to have some parameters. Personally I find it useful to create special structure for constructor params, and implement constructor so that it takes pointer to that struct. That way, we can easily add new params to ctor later, and we don't need to modify function signatures for that. So, initially the params struct is empty:

typedef struct S_MyBase_CtorParams {
   //-- add your ctor params here
} T_MyBase_CtorParams;

And constructor might look like this:

/**
 * Constructor
 */
void mybase__ctor(T_MyBase *me, const T_MyBase_CtorParams *p_params)
{
   memset(me, 0x00, sizeof(T_MyBase));
 
   //-- init vtable
   _vtable_init();
   me->p_vtable = &_vtable;
 
   //-- some construct code
}

We can now define non-virtual methods easily:

void mybase__my_method(T_MyBase *me)
{
   /* ... */
}

But how we're going to call virtual methods? We definitely don't want to impose a burden of dealing with vtable to the client of MyBase. So, for each virtual method (except memory deallocator) we just define inline function in the header:

/**
 * Desctructor
 */
static inline void mybase__dtor(T_MyBase *me)
{
   me->p_vtable->p_dtor(me);
}
 
/**
 * Some method
 */
static inline void mybase__some_method(T_MyBase *me, int a, int b)
{
   me->p_vtable->p_some_method(me, a, b);
}
 
/**
 * Other method
 */
static inline void mybase__other_method(T_MyBase *me)
{
   me->p_vtable->p_other_method(me);
}

Now, client of MyBase can equally call mybase__some_method() or mybase__other_method(), and he/she doesn't need to care about whether the method is virtual or not.

If we need to use heap, we also need to define allocator (which will call constructor as well) and deallocator (which will call destructor). I prefer to name them like this in order to replicate new and delete operators from C++ :

/* allocator */
T_MyBase *new_mybase(const T_MyBase_CtorParams *p_params)
{
   T_MyBase *me = (T_MyBase *)malloc( sizeof(T_MyBase) );
   mybase__ctor(me, p_params);
   return me;
}
 
/* deallocator */
void delete_mybase(T_MyBase *me){
   mybase__dtor(me);
   me->p_vtable->p_free(me);
}

Now, we can use our class as follows:

T_MyBase_CtorParams params = {};
 
//-- in stack
{
   T_MyBase mybase;
   mybase__ctor(&mybase, &params);
   mybase__some_method(&mybase, 1, 2);
   mybase__dtor(&mybase);
}
 
//-- in heap
{
   T_MyBase *p_mybase = new_mybase(&params);
   mybase__some_method(p_mybase, 1, 2);
   delete_mybase(p_mybase);
}

So if we create two instances of T_MyBase: my_base_1 and my_base_2, the data and code will be arranged as follows:

Please note that even though we're bringing some object stuff to C, we can't write in C in the same manner as in C++: in C, we have to manually call new_…() / delete_…() (or constructor / destructor), but in C++, we should avoid plain delete as much as possible, and rely on RAII instead.

Ok, base class seems to be done to a first approximation, let's proceed to derived class now.

Derived class

As I already mentioned, I use inheritance when I want some objects to share common interface. So, the interface is already done (for base class), now we should make derived class accessible through that interface.

First of all, let's specify structure for derived class context:

struct S_MyDerived {
   /* ... */
   struct {
      T_MyBase mybase;
   } super;
   /* ... */
};

I'd like to put base class(es) to the separate struct super, even though it is more verbose. If you like, you may omit it, as follows:

struct S_MyDerived {
   /* ... */
   T_MyBase mybase;
   /* ... */
};

In the rest of the text, I assume the first approach. We need a way to convert between pointer to base class and pointer to derived class. The derived-to-base is trivial, and I'd put it into the header as static inline function:

static inline T_MyBase *myderived__mybase__get(T_MyDerived *me)
{
   return &(me->super.mybase);
}

But for reverse conversion we need to perform some tricks.

Firstly, we'd like to ensure that the given pointer to base class actually points to an instance of some particular derived class. Unfortunately, it's not so easy to implement it correctly without a kind of RTTI (run-time type information). I haven't bothered yet about RTTI in C. But, as I already mentioned, most of the time I'm happy with 1-level inheritance. And then, we have very simple solution: since derived class has its own vtable, we can just check that pointer to vtable in the given instance actually points to vtable of some particular subclass:

/*
 * NOTE: works for 1-level hierarchy only!
 */
bool instanceof_myderived(const T_MyBase *me_super)
{
   //-- here we just check vtable pointer.
   return (me_super->p_vtable == &_super_vtable);
}

_super_vtable will be explained later, but I bet you've already guessed how it is defined.

Secondly, we can't just typecast from T_MyBase * to T_MyDerived *, since mybase might be not a first field of T_MyDerived. At least that inevitably happens with multiple inheritance. We need to be smarter.

I've created the header file ooc.h, in which I put common things useful for object-oriented approach in C. At least, there are a couple of macros:

/**
 * Calculates offset of member inside the struct
 */
#define OOC_OFFSETOF(_struct_, _member_)  (size_t)&(((_struct_ *)0)->_member_)
 
/**
 * Calculates pointer to subclass (i.e. containing struct) from pointer to
 * superclass (i.e. contained struct).
 */
#define OOC_GET_CONTAINER_PT(                                              \
      _T_Subclass_,                                                        \
      _superclass_field_name_,                                             \
      _superclass_pt_                                                      \
      )                                                                    \
                                                                           \
   ((_T_Subclass_ *)(                                                      \
      (unsigned char *)(_superclass_pt_)                                   \
      - OOC_OFFSETOF(_T_Subclass_, _superclass_field_name_)                \
      )                                                                    \
    )

Given macro OOC_GET_CONTAINER_PT(), we can calculate offset of base class field within derived class struct.

At the beginning of this article, I mentioned that I want to avoid typecasts as much as possible, but in the macro above you clearly see typecast to subclass. This is the only typecast I use in my OOC approach, and it is localized to well-defined macro, so that we aren't likely to make a mistake here. Even more, the code that uses this macro is autogenerated from template (see https://github.com/dimonomid/ooc), so you don't have to write it manually for every new subclass. I consider it quite acceptable (taking into account that there is no other way to convert from base to derived).

So, base-to-derived conversion will be done as follows:

T_MyDerived *myderived__get_by_mybase(const T_MyBase *me_super)
{
   T_MyDerived *p_derived = NULL;
 
   if (instanceof_myderived(me_super)){
      p_derived = OOC_GET_CONTAINER_PT(
            T_MyDerived,
            super.mybase,
            me_super
            );
   } else {
      //-- TODO: probably add some run-time error
   }
 
   return p_derived;
}

Let's get to virtual methods implementation for derived class. The derived class needs access to the vtable of base class. So, we should add function to base class for that:

const T_MyBase_VTable *_mybase__vtable__get(void)
{
   _vtable_init();
   return &_vtable;
}

This function is considered as “protected”, i.e. it is expected to be called from subclasses only. I found it useful to add the underscore prefix to “protected” functions.

Now, subclass can call _mybase__vtable__get to get pointer to the vtable of base class. Let's implement virtual methods:

static void _some_method(T_MyBase *me_super, int a, int b)
{
   //-- call method of superclass (if we need implementation inheritance as well)
   _mybase__vtable__get()->p_some_method(me_super, a, b);
 
   T_MyDerived *me = myderived__get_by_mybase(me_super);
 
   /* ... */
}
 
/**
 * Destructor (virtual)
 */
static void _dtor(T_MyBase *me_super)
{
   T_MyDerived *me = myderived__get_by_mybase(me_super);
 
   /* ... some destruct code ... */
 
   // NOTE: this is a subclass, so that after performing destruction code,
   // we should call desctructor of superclass:
   _mybase__vtable__get()->p_dtor(me_super);
}
 
/**
 * Memory deallocator (virtual)
 */
static void _free(T_MyBase *me_super)
{
   //-- we need to free pointer to "outer" struct, not to "inner" one.
   T_MyDerived *me = myderived__get_by_mybase(me_super);
   free(me);
}

As you see, in these implementations we use myderived__get_by_mybase() heavily, since these functions take pointer to base class, not to derived class (see prototypes of virtual functions above: T_MyBase_Dtor, etc). Additionally, we should call method of base class where appropriate. Say, in _dtor it is mandatory (because we clearly want base class to perform its needed cleanup), but in _some_method() we might not want to inherit implementation, so we could omit call of base method there.

Now we need to create vtable for our derived class. Here it is:

static T_MyBase_VTable _super_vtable;
static OOC_BOOL _bool_vtable_initialized = 0;

And the function that initializes it, it's similar to that of base class, but here we first copy vtable from base, and then override what we want:

static void _vtable_init()
{
   if (!_bool_vtable_initialized){
      //-- firstly, just copy vtable of base class
      _super_vtable = *_mybase__vtable__get();
 
      //-- and then, specify what we need to override.
      //   There are _dtor and _free, inevitably.
      _super_vtable.p_dtor           = _dtor;
      _super_vtable.p_free           = _free;
 
      //-- and then, our own virtual methods. If we don't override them here,
      //   it's ok: then, methods of base class will be called.
      _super_vtable.p_some_method    = _some_method;
 
      //-- remember that vtable is already initialized.
      _bool_vtable_initialized = 1;
   }
}

Please note that it's quite fine not to override some method (other than _dtor and _free) : you might have noticed that we've left p_other_method unchanged, so, it points to that of base class. We couldn't achieve this convenience if vtable is const.

Proceed to constructor now. For derived class, we create its own struct with params, which will of course contain struct with base params:

typedef struct S_MyDerived_CtorParams {
   T_MyBase_CtorParams super_params;
 
   //-- add your ctor params here
} T_MyDerived_CtorParams;

And constructor looks as follows:

void myderived__ctor(T_MyDerived *me, const T_MyDerived_CtorParams *p_params)
{
   memset(me, 0x00, sizeof(T_MyDerived));
   mybase__ctor(&me->super.mybase, &p_params->super_params);
 
   //-- init virtual methods
   _vtable_init();
   me->super.mybase.p_vtable = &_super_vtable;
 
   //-- some construct code
}

We're almost done. What is missing is an allocator for our derived class:

T_MyDerived *new_myderived(const T_MyDerived_CtorParams *p_params)
{
   T_MyDerived *me = (T_MyDerived *)malloc( sizeof(T_MyDerived) );
   myderived__ctor(me, p_params);
   return me;
}

Note that we don't need deallocator: for that, deallocator of base class should be used (see “Usage” section below).

So if we create two instances of T_MyBase: my_base_1 and my_base_2, and one instance of T_MyDerived: my_derived_1, the data and code will be arranged as follows:

Some conclusions

Note that each instance of the class (either base or derived) has a pointer to the vtable, not the vtable itself.

As an example, let's imagine that we have one base class Shape and three derived classes: Circle, Triangle, Rectangle. We would probably have lots of virtual methods for these classes, for drawing, resizing, converting them, etc. And we use them in some GUI-extensive application with lots of shapes: let it be 100 Circles, 100 Triangles and 100 Rectangles.

With the techniques described in this article, we end up with just 4 vtables, not 300 vtables, since vtable is created for each class, not for each instance. Effectively, this is approximately what the C++ compiler does for you behind the scenes when you use virtual methods (well, in fact, C++ compiler can be even smarter: it can eliminate pointers to non-overridden methods, but maintaining that manually in C is hardly a way to go).

So, this approach doesn't waste your RAM, which is always insufficient in the embedded world.

Usage

There is a very important design principle: “program to an interface, not an implementation”. If you haven't heard of it, you probably should read the Design Patterns book, or at least this conversation with its author.

Actually, that's why I bothered implementing all of this in the first place: I want most of my code not to care about exact object classes with which it works; instead, the code should work with some common interfaces.

Say, we might have some function that works with base class and then deletes it:

void my_func(T_MyBase *p_base)
{
   /* ... */
   mybase__some_method(p_base, 1, 2);
   /* ... */
   delete_mybase(p_base);
}

Then, we can equally use base or derived class for that. Consider:

int main(void){
   //-- use base class
   {
      const T_MyBase_CtorParams params = {/* ... */};
 
      T_MyBase *p_mybase = new_mybase(&params);
      my_func(p_mybase);
   }
 
   //-- use derived class
   {
      const T_MyDerived_CtorParams params = {/* ... */};
 
      T_MyDerived *p_myderived = new_myderived(&params);
      T_MyBase *p_mybase_inner = myderived__mybase__get(p_myderived);
      my_func(p_mybase_inner);
   }
}

In either case (with base or derived class), everything will work as expected: appropriate version of some_method() will be called, and when my_func() deletes object, needed destructors will be called, and memory will be freed correctly. That is, my_func() is completely unaware of exact type of object it works with; it knows the interface.

Of course we may write it more compact, but I wanted to elaborate on the process more, so I've written in in verbose manner.

Note that when we call mybase__some_method() on an instance of T_MyDerived, the actual pointer given to _some_method() of derived class is a pointer to base class, wrapped into the derived one. That's why it is named me_super. And then, we can obtain a pointer to the wrapper derived instance by calling myderived__get_by_mybase(me_super).

Consider this call chain, which happens when we call my_func(p_mybase_inner) :

Of course, compiler will optimize a call to inline mybase__some_method() away, but the idea is still the same.

Boilerplate code generator

Writing all this mess manually for each new class, or copy-pasting from existing classes all the time would be tedious and error-prone, so I've written a little utility that generates base and derived class templates with given names.

Here it is: https://github.com/dimonomid/ooc

The script ooc_gen.rb is written in ruby, so you need ruby for that to work.

Usage is trivial:

Generate base class Shape and subclass Circle:

$ ruby ooc_gen.rb new class shape circle

Generate base class Shape and two subclasses: Circle and MyShape:

$ ruby ooc_gen.rb new class shape circle myshape:MyShape

After either command, the directory shape containing all generated files will be created in the current directory.

So that, each class name may be specified as just single lowecased (underscore-separated) word, or as two words separated by colon.

If camelized version isn't provided by user, it is autogenerated, which isn't always what you want (say, for myshape autogenerated version is Myshape. However, if you provide my_shape, then it will be converted to MyShape).

In order to use generated classes, you also need for ooc.c and ooc.h, and configuration file ooc_cfg.h (initially, just copy default ooc_cfg_default.h as ooc_cfg.h in your project)


You might as well be interested in:

Discuss on reddit: or on Hacker News: Vote on HN

Discussion

Kenneth Kasajian, 2015/09/22 23:34

Also take a look at Simply Object-Oriented C http://www.codeproject.com/Articles/22139/Simply-Object-Oriented-C

Dmitry Frank, 2015/09/22 23:38

Thanks, as I mentioned in the article, there are quite a few existing publications on the topic. About this one: I don't like these macro-heavy solutions. But anyway, interesting read.

Tom Biskupic, 2015/10/01 05:37

Hello,

Nice article. A few suggestions however:

1) I wasn't sure if you were trying to avoid dynamic memory allocation but I see you aren't. In that case make your class type opaque and have the ctor method return it (you can typedef a void *). This way your callers aren't tempted to tinker, changes to the library cause minimal re-compilation of the client code and the implementation dependencies don't become dependencies of the client.

2) If instead of putting the function pointers in a vtable, make them just part of the struct. Then when you derive all you do is make the first part of the derived struct the parent struct. (i.e. struct Dervied { struct Base base; int additionalMember…};) This enables you to cast the struct into the parent struct type and allows the parent functions to operate on it as normal. To override a function in the base you just overwrite the pointer.

Tom

Dmitry Frank, 2015/10/01 08:21

Hi.

1) Sometimes I do just that, but rather rarely however, since making a type opaque makes it impossible to allocate instances in any way other than in the heap, which is often unacceptable in embedded projects. More, subclasses still need to access the structure itself, so that we should put it to some kind of “protected” header, for subclasses only. If in some application this is not a problem, I might do that.

2) Disagreed. This suggestion makes me think you haven't read the article carefully. First of all, making function pointers a part of the struct bumps the size of instances uselessly. If our class has 5 virtual methods, they take 20 bytes for each instance. And we have 50 instances, it is a whole 1 kB. In embedded world, RAM is a very expensive. With the separate vtable approach, each vtable is allocated not for each instance, but for each class/subclass: this approach saves a lot of RAM. See the Some conclusions section.

This enables you to cast the struct into the parent struct type

If you've read an article, you might have noticed that one of my goals was to avoid typecasts as much as possible. I consider casting from derived to parent as a very bad practice; more, it doesn't allow us to use multiple inheritance.

Developer Francis, 2015/10/10 12:38

Great write-up and awesome insight - thank you :-)

Keroronsk, 2016/09/27 09:40

Looks interesting, but a bit complex for me so far. Could you provide some small project, which I could build? Maybe with shape and cirlcle class, or somrthing like this.

Dmitry Frank, 2016/09/27 09:50

Thanks for the comment. Yeah probably a small example would be useful; I'll let you know once I get some time for it.

Marco I., 2018/03/22 08:15

Is there already a small project available? :D

Dmitry Frank, 2018/03/22 09:14

Nope sorry, not yet.

Marco I., 2018/03/22 08:17

Is there already a small project available?

Dmitry Frank, 2018/08/14 22:55

Not yet, sorry.

Marco I, 2018/04/15 16:43

Hello Dmitry

I tried out your generator tool today and I saw that it generates some ceedling stuff. How can I get it to run and what must I do to test the code in the classes? Must I create the test source files manual or is there a way to do it automatic “(this is done by ceedling automatically)”?

Can you maybe describe in a new article how to use/combine ceedling in/with your OO-stuff/generator tool … Pleaseee? ´´´°.°'`

And some other question:

- What is the best way if “Base” should inherit from a new base afterward?

- How can I generate SubSub-Clas which inherits from an (existing) Subclass (like in your UML)

Dmitry Frank, 2018/08/14 23:08
tried out your generator tool today and I saw that it generates some ceedling stuff.

The tool doesn't generate any ceedling stuff; it just uses ceedling stuff to make sure that the generator itself (and the templates it uses) works fine.

Can you maybe describe in a new article how to use/combine ceedling in/with your OO-stuff/generator tool

Not sure I'll be able to get time for that in a near future, sorry. Thanks for the request though.

How can I generate SubSub-Clas which inherits from an (existing) Subclass (like in your UML)

The generator doesn't support sub-subclasses. As I mentioned in the article, maintaining hierarchy like that manually in C is a lot more burdensome than having just 1-level inheritance. If you really-really need that, you'd have to write it manually. By the way, where in my UML did you see 2-level inheritance?

What is the best way if “Base” should inherit from a new base afterward?

Not sure I understand the question. If it's about refactoring existing 1-level inheritance into 2-level inheritance, then again, you'd have to implement it manually.

M. Israel, 2018/08/26 01:26

The reason why I asked about 2-level inheritance or inheritance from a second base class is because I'm think about how to handle exceptions in C. I was thinking to write an exception handler class which makes available stuff to subclasses.Like constructor and destructor, every object must know itself how to handle its exception in detail. But the techniques are equal to all objects (like exception handler functions to overwrite). Every object has a “status” like “ok” or an exception status. I could put this stuff into the first (real) base class. But I think it should be separated into an extra class.

Currently I don't know really how to deal with exceptions. With callbacks or overwritten function or us “setjump” and “longjump”? How you handle exceptions? Do you use the “CExceptions” or “CExcept” frameworks? It would be cool if you tell me your way of exception handling.

Kind regards

Dmitry Frank, 2018/08/28 13:40

First, I don't think that exceptions in the first place are such a great idea. Even in the languages which do support exceptions, it just makes the code more obscure, because I have to think all the time: can an exception be thrown here or not? and of which type? etc etc. It's often not obvious by just reading the code.

As an example, you can look at Go language (which I'm a huge fan of): https://tour.golang.com , it's a relatively new language, and it doesn't have exceptions.

The Go's approach is to have a second value returned from a function, so it's easy to return an error; obviously it's not as flawless in C where a function can only return a single value, but we can work it around by e.g. passing a pointer to where the error should be stored.

Marco I, 2018/08/12 11:33

Sorry, next question ;)

if i take a look into the deatail of OOC_GET_CONTAINER_PT:

Is it possible that it only works if the memory grows from low to hight? So like the heap does? What is if I allocate the object struct in the stack which grows from up to low? Must I than change

(unsigned char *)(_superclass_pt_) - OOC_OFFSETOF(_T_Subclass_, _superclass_field_name_)

To

(unsigned char *)(_superclass_pt_) + OOC_OFFSETOF(_T_Subclass_, _superclass_field_name_)

I think it works as follows (please correct me if not) 1. relative offset from relativ adress 0x00 of super inside the substruct

==> (size_t) 0x30, 

2. real_address_of_substruct == real_adress_of_superstruct - relative_offset_inside_substruct

      ==>  0x3300  == 0x3330 - 0x30
Dmitry Frank, 2018/08/14 22:55
memory grows from low to hight

Memory can't grow in another direction. The stack (not “memory”) can grow downwards, but it does not affect the addresses within a struct. E.g. in this struct:

struct foo {
  int a;
  int b;
}

The address of b is guaranteed to be higher on all platforms, independently of whether the stack grows upwards or downwards.

Marco I., 2018/08/15 11:09

Ahh you right. My mistake, sorry. Now it get sense again. And I was thinking forwards and backwards if and how it will work correct. Really thank you. :))

Goeff H, 2019/01/30 11:59

ever hear of Abstract Data Types?

Bosco Nguyen, 2019/10/30 17:00

Thank you for a professorial article. This article helps me started with OOC. I have a trivial question. Why do attach “S_” and “T_” to the structure declaration? It seems that tag name is only used once in the struct declaration, why attache “S_” to it to distinguish it from struct type name?

Dmitry Frank, 2019/10/31 09:39

Hi Bosco, thanks for the comment.

So first of all, I'm actually not sure anymore whether those prefixes S_, T_ etc are a good idea. They are definitely not a common practice, so working in a team it might be hard to use them everywhere. If anything, a _t suffix is sort of accepted (even stdlib uses it), so these days I'd probably write foo_t instead of T_Foo.

Second thing, I tend to not use typedef with structs anymore. I mean, having struct foo *a reads easier than just foo *a or even foo_t *a. Also, sometimes it's actually required to use struct name, if it references itself, like this:

struct foo {
  // ....
  struct foo *next;
}
example, 2021/03/28 20:34

You always have to use the struct keyword exactly twice with a typedef of a same type referencing with a forward typedef.

  typedef struct node node;
  struct node {
      node * prev;
      node * next;
      // data
  };

Inline typedef does not have this advantage, and requires using the keywords more times.

  typedef struct node {
      struct node * prev;
      struct node * next;
      // data
  } node;

The _t is a POSIX type, so it is a violation of standards if used for your own types.

Izabela, 2019/11/05 07:30

Hi Dmitry : D

Thank you very much. I've been struggling to find a good source that explains OOC. Your article helped me understand the concept as a whole. I based my OOC solution heavily on your methods and they work great for me. Great job, thanks again.

Dmitry Frank, 2019/11/05 07:50

Thanks for the comment Izabela, glad it helped :)

Drew Keller, 2020/10/02 21:34

Thanks for the writeup. The link to the utility on bitbucket seems to be dead (404 error). I think this is your github page but I don't see it there, either (https://github.com/dimonomid).

Dmitry Frank, 2020/10/03 10:45

Right, thanks for reporting. It was an old mercurial repo and bitbucket discontinued mercurial support a while ago. Just created a new git repo and pushed on github: https://github.com/dimonomid/ooc.

Drew Keller, 2020/10/05 15:14

Awesome, thanks! Is there a copyright license on the generated classes? Can I use them in a commercial project? I don't know if I will or not, but you know…

Dmitry Frank, 2020/10/05 16:15

Sorry for not adding a license file, but it's BSD, so feel free to use them in any projects.

Aleksandar, 2022/02/26 17:48

Hi,

very nice work. Specially the idea to pack class meta data in its own static struct is very good (I started something similar, but wanted to add VTable to every instance :/ ). I have one question or suggestion:

wouldn't putting the pointer to VTable as first element of every class and class members from second position allow “casting” of data members between different base and inheritzed class much easier? Member variables in every class hierarchy would start on exactly same position. I will try to write some meta code and write it here again.

Dmitry Frank, 2022/02/26 20:16

Hey, thanks.

wouldn't putting the pointer to VTable as first element of every class and class members from second position allow “casting” of data members between different base and inheritzed class much easier?

It's been a while since I touched anything C-related, so I don't remember all the fine details by now, but first of all, the very first requirement that I list above is this: Avoid typecasts as much as possible, so we don't lose type safety. So I'm not a fan of casting in the first place.

Also, relying on something like that means that multiple inheritance isn't possible, and I wanted to still use it occasionally.

Enter your comment (please, English only). Wiki syntax is allowed:
   _  __   ___   ____   _____   ____
  | |/_/  / _ \ / __ \ / ___/  / __/
 _>  <   / // // /_/ // /__   _\ \  
/_/|_|  /____/ \___\_\\___/  /___/
 
articles/oop_in_c.txt · Last modified: 2023/03/09 08:34 by dfrank
Driven by DokuWiki Recent changes RSS feed Valid CSS Valid XHTML 1.0