Wednesday, September 12, 2007

Why Objective-C is cool

I've been asked to do an intro about Cocoa. So I thought about what would I tell people about Cocoa if I had some time. Sure I could throw up a quick tutorial on how to code a Cocoa app showing a bare minimum of how Objective-C works. But there are a million of tutorials like that and it doesn't really do the Cocoa justice. I want to give people an idea of why Cocoa or perhaps more specifically Objective-C is cool. I think anybody who has played computer games to some extent know how the game Doom (1993) from id software revolutionized gaming on the PC platform. What few people know was the the game was actually not developed at the PC platform at all. At the time Doom was developed in 1992, Windows 95 didn't exist yet and software development on PC's looked like this: Not very impressive. Instead it was developed on computers from NextStep, the company founded by Steven Jobs after he was kicked out of Apple. As id developer John Romero says:
Why do I care so much about NeXT computers? Because we at id Software developed the groundbreaking titles DOOM and Quake on the NeXTSTEP 3.3 OS running on a variety of hardware for about 4 years. I still remember the wonderful time I had coding DoomEd and QuakeEd in Objective-C; there was nothing like it before and there still is no environment quite like it even today.
To get an impression of how far ahead of its time Objective-C and what is now known as Cocoa was ahead of its time, consider this:
In fact, with the superpower of NeXTSTEP, one of the earliest incarnations of DoomEd had Carmack in his office, me in my office, DoomEd running on both our computers and both of us editing one map together at the same time. I could see John moving entities around on my screen as I drew new walls. Shared memory spaces and distributed objects. Pure magic.
Consider how long time ago that was, CORBA didn't support C++ for remote method invokation until 1996. At the time id did this it was just in its infant stage. Not to mention this was long long before Java RMI made remote objects populare and mainstream. It isn't even common to do this kind of thing today, and yet this was something that was quite trivial to do in Objective-C and Cocoa back in 1992.

Introducing Objective-C

Every time someone introduce you to Objective-C they will tell you about what a small and simple language it is. How easy it is to learn etc. But that was not my impression when I first tried to learn it. At the time I knew C++. When I saw an example of how to declare a class:
@interface MyClass : NSObject
{
  int         mSomeNumber;
  NSString*   mSomeString;
}

- (int)someNumber;
- (void)setSomeNumber:(int)aNum;
@end
I didn't think it looked easy or simple at all. Especially when I saw some code examples:
- (void)drawRect:(NSRect)frameRect
{
  // Fill whole background with white
  [[NSColor whiteColor] set];
  [NSBezierPath fillRect:frameRect];

  // Construct rows of lines
  NSBezierPath *gridLines = [[NSBezierPath alloc] init];
  for (unsigned row = 0; row <= mNoRows; ++row) {
      [gridLines moveToPoint: rowLeft];        
      [gridLines lineToPoint: rowRight];
      rowLeft.y  += rowHeight;
      rowRight.y += rowHeight;            
  }        

  // Set style to draw line in
  [[NSColor blackColor] set];
  [gridLines setLineCapStyle:NSSquareLineCapStyle];
  
  float pattern[2]  = {2.0f, 5.0f};
  [gridLines setLineDash:pattern count:2 phase:0.0f];

  // Draw specified lines in given style
  [gridLines stroke];
  [gridLines release];    
}
The code below is just to give an idea of what Objective-C syntax is like. What the code does is not important. It has been simplified a bit to understand better. It basically draws several rows with black lines on a white background. When I first saw this kind of code I didn't think it looked easy at all. And I thought whoever said that Objective-C was a much simpler language than C++ must have been smoking something that wasn't good for them. The problem was that I was too caught up in the syntax. The syntax is indeed very unusual, but if one looks beyond that, the structure of the syntax is actually quite simple. Perhaps the best way to understand the syntax is to understand, why it looks so strange.

Background

Unlike C++, Objective-C wasn't really a new language at all originally. There was no Objective-C compiler. Instead Objective-C was just plain old C, with an added preprocessor. As any C/C++ developer worth his/her salt should know all statements starting with # specifies a preprocessor directive in C/C++. E.g. #define and #include. The preprocessor runs before the compiler and replaces the directives with actual C/C++ code. What the makers of Objective-C did was to add another preprocessor, but instead of marking the directives with #, they marked them with @. That way their special preprocessor could find all the new directives easily. While this does make Objective-C look like some alien entity inside C, it does make it trivial to distinguish between pure C code and the add-ons from Objective-C. Unlike C++, where what is C and what is C++ is blurred. Actually C++ isn't a strict superset of C like Objective-C although it usually compiles C code. The reason being among other things that C++ reinterprets a lot of regular C code into new C++ concepts. E.g. a struct isn't just a struct anymore but actually a class in C++.

Syntax inspiration from smalltalk

The second stumble block in order to understand Objective-C syntax is to realize it is derived from Smalltalk, while C/C++ syntax is derived from Algol. Algol based languages separate arguments with comma, while smalltalk separates with name of argument and colon. Below is a code snippet that demonstrates setting the position and dimension of a rectangle object.
// Smalltalk
rectangle setX: 10 y: 10 width: 20 height: 20

// Objective-C
[rectangle setX: 10 y: 10 width: 20 height: 20];
[rectangle setX1: 10 y1: 10 x2: 20 y2: 20];

// C++
rectangle->set(10, 10, 20, 20);
The Smalltalk/Objective-C syntax improves readability of code. When specifying a rectangle some libraries use the start and end coordinates while others use start coordinates and size. With Smalltalk/Objective-C it is made quite clear what is done. While with C++ it is not clear whether the first or the second line of Objective-C code is used. Unlike Smalltalk Objective-C methods are called by enclosing call with []. This was originally to aid the preprocessor in extracting what was regular C code and what was Objective-C specific code.

A small and simple language

What do we exactly mean by saying that Objective-C is a small an simple language. C++ looks simpler syntax wise for the novice. However as said before syntax in deceiving. C++ add a host of new features to the C language: classes, virtual methods, non-virtual methods, constness, templates, friend classes, multiple inheritance, pure virtual functions, operator overloading, function overloading, private, public and protected members, constructors etc. Objective-C on the other hand add very little, it is just classes, methods, categories and protocols. There is only one way to create classes and to inherit form them. You can specify whether inheritance is public or private e.g. There is only one kind of methods. There is no distinction between virtual and non-virtual. They can't be const or not const. The concept of constructors and destructors don't exist. Instead these are just normal methods. As mentioned Objective-C is so simple they didn't even need to create a new compiler. A preprocessor was enough. This can make Objective-C seem overly simple. How can you do much with a language like that? The answer is the Objective-C runtime. The runtime is in fact what makes most of the Objective-C magic happen. Most of Objective-Cs features is provided at runtime and not done at compile time.

Methods and Selectors

Objective-C differs from C++ in that one distinguish between methods and selectors. To call a method on an Objective-C method you send a message. The method is not called directly. Instead Objective-C determines based on the selector which method to call. Conceptually one can think of a message (selector) as simply a string with a list of arguments. The code below e.g. send the message setX:y:width:height: to object rectangle. This will invoke a method on rectangle if it understands the message.
[rectangle setX: 10 y: 10 width: 20 height: 20];
If we think of Objective-C as just a preprocessor as it originally was, all message sending is replaced with a call to a C function:
id objc_msgSend(id theReceiver, SEL theSelector, ...)
So when the preprocessor encounters our rectangle message it is translated into something like this:
objc_msgSend(rectangle, "setX:y:width:height:", 10, 10, 20, 20);
objc_msgSend queries rectangle for its class and then queries the class for methods it contains. Then it finds out which method corresponds to the given selector. The method is nothing but a function pointer of the form:
id (*IMP)(id, SEL, ...)
Of course this is a simplified explanation. In reality every selector is registered. Meaning we don't pass newly created strings to objc_msgSend each time we invoke a method but a pointer to a string. This pointer has to be unique. So we can't pass any string pointer. So if we got å string we can find the unique pointer to this string by using the function:
sel_registerName(const char *str)
Which returns the pointer if it exist or registers the selector as a unique string pointer if it doesn't exist. The benefit of this is that selectors can be looked up quickly and compared quickly by just comparing their pointer addresses rather than comparing each character of the string.

Unique dynamic features of Objective-C

I don't intend to go into every minor of the Objective-C runtime library but the example above should give an idea of how the dynamic features of Objective-C works.

Classes

In C++ classes don't exist passed compile time. Or at least in modern C++ compilers which support runtime type identification classes exist in a watered down sense in that one can query if two classes have a relationship. In Objective-C it is almost opposite. Classes don't exist at compile time but rather are runtime entities. Classes are registered at runtime with function:
void objc_addClass(Class myClass);
Likewise methods and selectors are registered at runtime. So classes can in fact be modified at runtime. Of course users don't call objc_addClass. These methods along with the ones that registers methods are generated by the preprocessor from the class definition provided by the programmer. But it is this fact that classes and methods exist as structures in memory at runtime that allows the programmer at runtime to query classes about their member variables and functions and whether they respond to a selector or not. In fact an Objective-C developer could create an application which could let user specify classes to call and functions to call. User could just type in the name of a class. Then developer could use C function NSClassFromString() to get corresponding class. NSSelectorFromString() could be used to retrieve the selector. With this one could query class further about arguments existing for selector etc. To learn more about Objective-C look at wikipedia

Saturday, September 08, 2007

The right tool for the job: C++ vs Objective-C

I read Alexei's blog entry about C++ vs Objective-C. I agree fully with his statement that:

The argument (which you can go and read yourself) boils down to “C++ can do everything Objective-C can” versus “sure it can, but not easily, usably or understandibly.” The usual argument comes down to object models: With Objective-C, you can send any object any message, and if it understands it, it will respond. C++ is restricted by its compile-time type system, so that even if you have arbitrary objects that all implement the member function foo, there is no way to call that method on a set of them unless they all inherit from the same base class. Except that you can, as exemplified boost::any and boost::variant.

Actually I pretty much agrees with everything Alexei says and the comments put forth by different people. C++ templates is a wonderful thing while often complicated to understand and use. My major issue with them is that they can't be extended into runtime.

So anyway what I want to emphasis is that templates is not substitute for Objective-C's dynamic features. I work daily on a C++ application with +4 million lines of code. The application have a natural fit with OO design because it has a number of data objects which can be inspected, manipulated and visualized in many different sorts of views at the same time. E.g. one view can show the object in 3D while another can show an intersection in 2D. So Model-View-Controller is a natural fit for it.

The more I work on it the clearer Greenspuns Tenth Rule of programming becomes to me:

Any sufficiently complicated C or Fortran program contains an ad hoc, informally-specified, bug-ridden, slow implementation of half of Common Lisp.

Bottom line is that the kind of structure a large application like this would typically have is what Objective-C was designed for. And thus when I look at the app, all we do in C++ is making a bug-ridden slow implementation of Objective-C.

Creation of views for data objects has to be abstracted because we might create new views we never thought about when the data objects were designed. Thus construction of views has to be abstracted. In Objective-C this is easy. Classes can be passed around like objects and instantiated. C++ on the other hand provides no abstraction of object creation so we ended up making this elaborate scheme with Class ids and factory classes.

We created reference counting, a system for sending messages between object for notifications etc. All stuff that Objective-C was designed for. We end up with huge amount of code for the overall architecture. And it is not even working that well.

But we are not the only ones. Every time I look at a large C++ framework or toolkit I realize that they are just re-implementing Objective-C features in a non standard and buggy way. When I say non-standard I mean that every toolkit has those features but they are done differently in each one and thus not compatible.

E.g. VTK (Visual Toolkit) has reference counting, message passing and abstraction of object creation. Open Inventor has their own reference counting etc. The Qt GUI toolkit has basically gone around the limitations of the C++ language for dynamic behavior and essentially created a library that allows for object introspection and runtime and dynamic dispatch. Essentially they have reimplemented Objective-C in C++. While they have don't quite a good job and I love the Qt toolkit it is still a kludge and non-standard.

So what exactly is my point? My point is that while people can argue from a theoretical foundation that C++ really is the better language due to all its features the power of templates etc, reality speaks for itself. In practice C++ very frequently fails as a language. It seems quite plain to me that when any large C++ toolkit seem to make a half bad implementation of Objective-C then something is wrong with the language.

Does this mean that I think Objective-C is a better language? No, far from it. I just think that one should use the best tool for the job. Frequently it seems like Objective-C would have been the best tool but C++ was selected instead. While C++ has an obvious advantage in implementing algorithms and high performance code, it doesn't seem like the mistake has been done as frequently at the Objective-C camp.

My speculation would be that this is because Objective-C's deficiencies are so obvious. While C++ deficiencies are not as clear. One can more easily kid oneself into thinking C++ will solve the problem without any hassle. And not at least because C++ is better known, a Objective-C developer will more likely consider using C++ for specific parts of the program than the other way around.

C++ and Objective-C complements each other very well in my opinion and they have different enough syntax from each other that it should be easy to keep the two apart in a program. But too often people think C++ is the silver bullet that can be used equally well for any task.

Sunday, February 18, 2007

Wrapping C++ classes in Lua

Lua is a script language frequently used in game development. Typically the game engine is written in C/C++ and the game itself is defined in Lua. Meaning setting up behaviour, placement of objects like powerups, doors, enemies etc is done in Lua. That way the game can be quickly tweaked without having to go through a time consuming re-compile.

But I am not going to go in depth about the different usage of Lua, nor explain the language itself. An online book at the Lua web page does that very well already. I will here assume you already have some basic knowledge about Lua.

However finding any documentation on how to wrap a C++ class so that it can be used in Lua is difficult. One could of course use one of the ready made bridges. But here we are going to look at how to do it yourself.

The tricky part is deciding on how to do it. Because Lua is such a flexible language there is really a lot of ways you could achieve it.

The naive approach

First lets look at several options, what works and doesn't. My first approach was to light user data to store a pointer to my C++ class. I will use a Sprite class and related classes as examples here as that was the type of classes I was wrapping.

static int newSprite(lua_State *L)
{
    int n = lua_gettop(L);  // Number of arguments
    if (n != 4)
        return luaL_error(L, "Got %d arguments expected 4", n);
  
    double x = luaL_checknumber (L, 1);      
    double y = luaL_checknumber (L, 2);
    double dir = luaL_checknumber (L, 3);      
    double speed = luaL_checknumber (L, 4);

    Sprite *s = new Sprite(Point2(x, y), dir, speed);
    
    lua_pushlightuserdata(L, s);

  return 1;
}

The code snippet above shows a naive implementation of this approach. Unfortunately it doesn't work. The problem is that a light user data is just a simple pointer. Lua does not store any information with it. For instance a metatable which we could use to define the methods the class supports.

An approach with limited functionality

The next approach would be to use user data. Unlike light user data it can store a reference to a metatable.

static int newSprite(lua_State *L)
{
    int n = lua_gettop(L);  // Number of arguments
    if (n != 4)
        return luaL_error(L, "Got %d arguments expected 4", n);
  
    // Allocate memory for a pointer to to object
    Sprite **s = (Sprite **)lua_newuserdata(L, sizeof(Sprite *));  

    double x = luaL_checknumber (L, 1);      
    double y = luaL_checknumber (L, 2);
    double dir = luaL_checknumber (L, 3);      
    double speed = luaL_checknumber (L, 4);

    *s = new Sprite(Point2(x, y), dir, speed);
    
    lua_getglobal(L, "Sprite"); // Use global table 'Sprite' as metatable
    lua_setmetatable(L, -2);       
    
  return 1;
}

For us to be able to use sprite like this we need to register class first. Basically we need to create a table Sprite which contains all the methods that our user data should support.

// Show registration of class
static const luaL_Reg gSpriteFuncs[] = {
  // Creation
  {"new", newSprite},
  {"position", position},
  {"nextPosition", nextPosition},    
  {"setPosition", setPosition},  
  {"render", render},      
  {"update", update},          
  {"collision", collision},   
  {"move", move},    
  {"accelerate", accelerate},      
  {"rotate", rotate},  
  {NULL, NULL}
};

void registerSprite(lua_State *L)
{
  luaL_register(L, "Sprite", gSpriteFuncs);  
  lua_pushvalue(L,-1);
  lua_setfield(L, -2, "__index");    
}

This will allow us to create instances of Sprite and call methods on it in Lua like this:

-- Create an instance of sprite an call some methods
local sprite = Sprite.new(x, y, dir, speed)
sprite:setPosition(x,y)
sprite:render()

The final approach

In most cases this approach is sufficient but it has one major limitation. It does not support inheritance. You can change the methods of Sprite in Lua but that will change the behavior of all instances of Sprite. What you would want to do is to be able to change method on just the instance and then use that instance as a prototype for new Sprite instances, effectively creating a class inheritance system.

To do this we need to change the instance into being a table. How do we access our C++ object then? Simple, we just store the pointer to it as user data in one of the field of the table. You might think that this time light user data will be sufficient. However the problem is that only user data is informed of garbage collection, not tables or light user data. So if you want to delete the C++ object when corresponding lua table is garbage collected you need to use user data.

So we then arrive at our final solution. We will store a pointer to our C++ object as user data on the key __self in the table that represents our instance. __self is an arbitrary selected name. It could be anything. We will not register Sprite as as the metatable for our instance but instead register the first argument to the new function as it. This will allow us to support inheritance. Further the garbage collection function will be register on a separate table which will be used as metatable only for the user data. This is to allow it to be garbage collected.

static int newSprite(lua_State *L)
{
    int n = lua_gettop(L);  // Number of arguments
    if (n != 5)
        return luaL_error(L, "Got %d arguments expected 5 (class, x, y, dir, speed)", n);
    // First argument is now a table that represent the class to instantiate        
    luaL_checktype(L, 1, LUA_TTABLE);   
    
    lua_newtable(L);      // Create table to represent instance

    // Set first argument of new to metatable of instance
    lua_pushvalue(L,1);       
    lua_setmetatable(L, -2);

    // Do function lookups in metatable
    lua_pushvalue(L,1);
    lua_setfield(L, 1, "__index");  

    // Allocate memory for a pointer to to object
    Sprite **s = (Sprite **)lua_newuserdata(L, sizeof(Sprite *));  

    double x = luaL_checknumber (L, 2);      
    double y = luaL_checknumber (L, 3);
    double dir = luaL_checknumber (L, 4);      
    double speed = luaL_checknumber (L, 5);

    *s = new Sprite(Point2(x, y), dir, speed);
    
    // Get metatable 'Lusion.Sprite' store in the registry
    luaL_getmetatable(L, "Lusion.Sprite");

    // Set user data for Sprite to use this metatable
    lua_setmetatable(L, -2);       
    
    // Set field '__self' of instance table to the sprite user data
    lua_setfield(L, -2, "__self");  
    
  return 1;
}

We can now work with sprite instances in Lua like this:

-- Instantiate from Sprite class
local sprite = Sprite:new(x, y, dir, speed)
sprite:setPosition(x,y)
sprite:render()

-- Add method to instance as use it as class
function sprite:doSomething()
  print("do something")
end

local derived = sprite:new(x, y, dir, speed)
derived:render()
derived:doSomething() -- This is now a legal operation

There are still a couple of loose ends. We haven't showed how the methods are registered with this new solution nor how we access C++ object pointer in methods. But this is fairly straight forward as I will show.

void registerSprite(lua_State *L)
{  
  // Register metatable for user data in registry
  luaL_newmetatable(L, "Lusion.Sprite");
  luaL_register(L, 0, gDestroySpriteFuncs);      
  luaL_register(L, 0, gSpriteFuncs);      
  lua_pushvalue(L,-1);
  lua_setfield(L,-2, "__index");  
  
  // Register the base class for instances of Sprite
  luaL_register(L, "Sprite", gSpriteFuncs);  
}

We can then implement a method of Sprite like this

static int setSpeed(lua_State *L)
{
  int n = lua_gettop(L);  // Number of arguments
  
  if (n == 2) {
    Sprite* sprite = checkSprite(L);
    assert(sprite != 0);
    real speed = luaL_checknumber (L, 2);
    sprite->setSpeed(speed);
  }
  else
    luaL_error(L, "Got %d arguments expected 2 (self, speed)", n);
  
  return 0;
}

To extract the pointer to the C++ object and make sure it is of the correct type we use the following code:

Sprite* checkSprite(lua_State* L, int index)
{
  void* ud = 0;
  luaL_checktype(L, index, LUA_TTABLE);
  lua_getfield(L, index, "__self");
  ud = luaL_checkudata(L, index, "Lusion.Sprite");
  luaL_argcheck(L, ud != 0, "`Lusion.Sprite' expected");  
  
  return *((Sprite**)ud);      
}

The only thing left is dealing with garbage collection but I leave that as an exercise. You should already have the basic idea of how to deal with it. Please note that I have not tested the exact same code as written here so why the principles are correct there might be minor errors in the code. In my own code I have separated more of the code into separate functions since the code for creating an instance is almost identical for any class, as well as the code for extracting the __self pointer.

Conclusion

While using a bridge might be better for bigger projects I think it is good to know for yourself exactly what goes on under the hood and when you do it yourself you can more easily fine tune what you export and not and in which way. Typically you would want the Lua interface to be simpler and more limited than the C++ interface to your classes.

Saturday, January 20, 2007

Intersection test assertion failure in CGAL

I have just started using the CGAL library. It is a library for doing computational geometry. It contains classes and functions for different geometric data structures like triangles, line segments, polygons and functions for testing for intersection, calculating convex hull and a lot more. This week I fell into the first trap of computational geometry. I stored polygons in Polygon_2 classes and then performed an intersection test with the do_intersect function. Unfortunately that too often lead to the program crashing with the assertion failure: CGAL error: precondition violation! Expr: Segment_assertions::_assert_is_point_on (p, cv, Has_exact_division()) && compare_xy(cv.left(), p) == SMALLER && compare_xy(cv.right(), p) == LARGER File: /Users/Shared/CGAL-3.2.1/include/CGAL/Arr_segment_traits_2.h Line: 730 or CGAL error: assertion violation! Expr: *slIter == curve File: /Users/Shared/CGAL-3.2.1/include/CGAL/Basic_sweep_line_2.h Line: 591 Eventually I narrowed it down to discover that it happened in specific degenerate cases. For instance when a corner of a polygon intersected an edge of another polygon. Which means two line segments are intersecting another line segment in the exact same spot. Now I know this would normally be a problem for line sweep algorithms but I also know that any computational geometry library with respect for itself should handle such degeneracies. So I could not understand why CGAL should fail. Well it turns out that I didn't read this part of the CGAL manual:
If it is crucial for you that the computation is reliable, the right choice is probably a number type that guarantees exact computation. The Filtered kernel provides a way to apply filtering techniques [BBP01] to achieve a kernel with exact and efficient predicates. Still other people will prefer the built-in type double, because they need speed and can live with approximate results, or even algorithms that, from time to time, crash or compute incorrect results due to accumulated rounding errors.
I had expected this kind of problem. So I thought it was because I used float. I changed to double and got the same problem. It turns out I should have used CGAL::Exact_predicates_exact_constructions_kernel.