Tuesday, October 2, 2012

JavaScript's Pseudo Classical Inheritance diagram

The following is a chart of JavaScript Pseudo Classical Inheritance.  The constructor
Foo is just a class name for an imaginary class.  The foo object is an instance of Foo.

Note that the prototype in Foo.prototype is not to form a prototype chain.  Foo.prototype points to some where in a prototype chain, but this prototype property of Foo is not to form the prototype chain.  What constitute a prototype chain are the __proto__ pointing up the chain, and the objects pointed to by __proto__, such as going from foo.__proto__, going up to foo.__proto__.__proto__, and so forth, until null is reached.

JavaScript's Pseudo Classical Inheritance works like this way: I am a constructor, and I am just a function, and I hold a prototype reference, and whenever foo = new Foo() is called, I will let foo.__proto__ point to my prototype object.  So Foo.prototype and obj.__proto__ are two different concepts.  Foo.prototype indicates that, when an object of Foo is created, this is the point where the prototype chain of the new object should point to -- that is, foo.__proto__ should point to where Foo.prototype is pointing at.

In the ECMA-262 Edition 5.1 spec, the term [[Prototype]] is used.  And that's the same as __proto__.  It is often mentioned as the "[[Prototype]] internal property".  And don't confuse this with a function's prototype property.  One of the key points regarding [[Prototype]] is: "All objects have an internal property called [[Prototype]]. The value of this property is either null or a reference to an object and is used for implementing inheritance."

And now we can see from the diagram why when we inherit Dog from Animal, we would do:

    function Dog() {}    // the usual constructor function
    Dog.prototype = new Animal();
    Dog.prototype.constructor = Dog;

It is to say: first of all, define a function Dog.  This function is an object, as every function in JavaScript is an object.  Now, add a property prototype to the object Dog.  And then, new Animal() will create a new object (and it will become part of the prototype chain), and since this new object is created using the Animal constructor, this new object's __proto__ will point to where Animal.prototype is pointing to.  Now, make Dog.prototype point to this new object.  And that's how the object pointed to by Dog.prototype is created as shown in the diagram.

Next, since that new object was created by the Animal constructor, the new object's constructor property will point to Animal (an object created by constructor Foo will have a constructor property pointing to Foo).  This constructor property is not Dog.prototype's own property -- it is the own property of Dog.prototype.__proto__, so it is an inherited property.  Note that in this case, we actually want the Dog.prototype object's constructor property to point to Dog, and that why we have the second line of code above: Dog.prototype.constructor = Dog;

Note that we can use an empty function F() to set up the above relationship as well:

    function Dog() {}    // the usual constructor function
    function F() {}
    F.prototype = Animal.prototype;
    Dog.prototype = new F();
    Dog.prototype.constructor = Dog;

Because Animal() may take a longer time or more complex logic to run, and it can also create properties in that new object that we don't need.  A way to set up the Dog.__proto__ to point at Animal.prototype is what we need, and F() can already accomplish that.

If there are 3 Dog instances, they would point to the middle of that long prototype chain.  It is still a complete prototype chain, but a shorter one:

Now we can understand why when we add a method to the Animal class, we would use

    Animal.prototype.move = function() { ... };

That's because when we say


If woofie the object doesn't have the move method, it will go up the prototype chain, just like any prototypal inheritance scenario, first to the object pointed to by woofie.__proto__, which is the same as the object that Dog.prototype refers to.  If the method move is not a property of that object (meaning that the Dog class doesn't have a method move), go up one level in the prototype chain, which is woofie.__proto__.__proto__, or the same as Animal.prototype.  Remember we already did

    Animal.prototype.move = function() { ... };

earlier, and so now move is found, and the method can be invoked.

Note again that this is how prototypal inheritance works, and see how "classical inheritance" is simulated: by the help of prototypal inheritance.

Using the diagram, we can also see the working of instanceof.  For foo instanceof Animal, it is true, because we take foo and look at the whole prototype chain, and the Animal.prototype object is part of that chain.  Therefore, it returns true.  woofie instanceof Animal is true for the similar reason: take woofie's whole prototype chain, and the Animal.prototype object is part of that chain.  woofie instanceof Bichon is false because the Bichon.prototype is not part of that chain.  Note that woofie.__proto__ instanceof Animal is true, the same as Dog.prototype instanceof Animal, because instanceof checks for whether the right operand's prototype object is part of the left operand's prototype chain. (Note that Dog.prototype instanceof Dog used to be true, but it has changed in later implementation of JavaScript: so it will go up to see if Dog.prototype is part of the chain, but it won't include the object itself (the left operand) to check against Dog.prototype).

Note that in reality, each constructor function has a __proto__ property as well, and if the Function constructor is also shown here,  a more complete picture is:

Even though foo.constructor === Foo, the constructor property is not foo's own property.  It is actually obtained by going up the prototype chain, to where foo.__proto__ is pointing at.  The same is for Function.constructor.  The diagram can be complicated, and sometimes confusing when we see Constructor.prototype, foo.__proto__, Foo.prototype.constructor.  Note that Firefox, Chrome, Safari, and node.js support __proto__, but IE doesn't support it, and it can be obtained by Object.getPrototypeOf(foo).  (IE 9 or above is needed.  Before IE 9, it can be defined as in John Resig's post, and it requires that the constructor property is set properly.)  To verify the diagram, note that even though foo.constructor will show a value, the property constructor is not foo's own property, but is obtained by following up the prototype chain, as foo.hasOwnProperty("constructor") can tell.

Friday, September 14, 2012

Creating Bitmap Context for Retina and regular iOS devices

When it is a Retina display iOS device, the resolution is 4 times as a regular display.  How is a bitmap context created in this case that can make use of the higher resolution, while for a regular display, a regular bitmap context is used so that memory isn't wasted?

The answer is using the [[UIScreen mainScreen] scale], and create the bitmap context accordingly.  But will any drawing routine also take special care to draw on this bitmap context because now the pixels on x and y-axis have both doubled?

The solution is that we can just do a transform, and everything will be taken care of.  By doing this, any drawing routine will not need to tailor to any particular size.  Moving to a point at (300, 300) will be actually moving to pixel (600, 600), but the drawing can just use (300, 300) for both a regular and Retina device.  The solution is:

float scaleFactor = [[UIScreen mainScreen] scale];
CGSize size = CGSizeMake(768, 768);
CGColorSpaceRef colorSpace = CGColorSpaceCreateDeviceRGB();
CGContextRef context = CGBitmapContextCreate(NULL, 
                           size.width * scaleFactor, size.height * scaleFactor, 
                           8, size.width * scaleFactor * 4, colorSpace, 
CGContextScaleCTM(context, scaleFactor, scaleFactor);

note that the last line, the CGContextScaleCTM is important.  It does the work of making (300, 300) to be the actually pixel (600, 600) on a Retina device.  The line that does the CGSizeMake(768, 768) is how big you'd like the bitmap context to be.  It works on a regular display and is automatically scaled up for a Retina display in the code above.

Thursday, September 13, 2012

Objective-C Manual Retain Release

The Manual Retain Release in Objective-C is not really that hard, but with the following precise rules:

Our motivation is:
  • We would like to hold onto an object, when at least one reference is pointing to it
  • We would like to free an object's memory space, when there is 0 reference pointing to it.

The mechanism is:
  • We use retain count to make this work
  • When an object is alloc'ed by [Foo alloc], the retain count is 1
  • When an object is created by [Foo new], it is the same as [[Foo alloc] init], and the retain count is also 1
  • When an object is copied, the retain count of the new object is 1
  • When the retain count is 1 or greater, that means the object should stay around.
  • When we send the retain message to an object, the retain count is incremented by 1.  This is how we send the retain message to obj: [obj retain];
  • What about decrementing the count?  It is by sending the release message to the object: [obj release];
  • When this release message is sent to the object, first, the retain count is decremented by 1, and when this retain count reaches 0, the system will send the dealloc message to the object.  That is, the system will do [obj dealloc];  for you.  So in the object's dealloc method that you define, make sure to clean up any other objects you keep around for the current object.  Then, call [super dealloc], so that the superclass will clean up objects at each higher level of the class hierarchy.  Then when it reaches the last one: [NSObject dealloc]; the actually memory (RAM or think of it as virtual memory) is released (freed up), and becomes available for other apps or your app to use again (RAM / virtual memory).

That's it.  So match the alloc, new, copy, retain, release, so that there is a balance.  Don't retain too much, and don't release too much (or too early).

Somethings to note:
  • Never call dealloc yourself, except to call super class's dealloc: [super dealloc].  The system will call dealloc for you when performing [obj release] and found the retain count to be 0 after that release.
  • You can check the retain count by [obj retainCount], although you should never use this number to do memory management.  It is only for understanding the mechanism and for experimenting and checking to see how the retain count increased or decreased.

For factory methods:
  • factory methods, such as stringWithFormat, needs to alloc a string, and return it.  So this method cannot do a [str release] because the object cannot be freed up yet.  But alloc has to be matched with a release, so how can this be solved?  This is the way:
  • There is an autorelease pool at each iteration of the app's main loop.  When we do [[[str alloc] initWithFormat: ...] autorelease] the retain count of the str object remain as 1, but the object is added to the autorelease pool.  When the caller gets back the string, it will perform the retain to hold onto the object, so that the object is not freed up.  At this point, the retain count is 2.  And when all application events are handled, the system will drain the autorelease pool, and make the retain count of str become 1, and now everything is in good order, and the system later will start an autorelease pool again, and handle all app related events, and at the end of this iteration, drain the autorelease pool again.
  • Note that every time you send the autorelease to that object, the autorelease pool will keep a number as to how many times to send the release message to the object when the pool drains.  So for example, if [[obj retain] autorelease]; is done 10 times, the retain count of the object will increase by 10, and when the autorelease pool drains, the object will be sent the release message 10 times.  So autorelease doesn't decrease the retain count immediately.  It decreases the retain count later.  A good way to think of autorelease is to think of it the same as a release, but deferred.

For @property:

  • If the property attribute is retain, copy, (or strong, but that is part of ARC), then the instance variable will automatically hold onto this object (the object will be retained once).  The previous object pointed to by this property will be released once.
  • If the property attribute is assign, unsafe_unretained, (or weak, which is ARC), then the instance variable will not cause the object's retain count to increase, as this property is not trying to hold onto the object (no ownership claimed).

Retain count with Cocoa and Cocoa-Touch (iOS) frameworks:

  • The frameworks work naturally with the retain count, so that when an object is added to a collection, such as an NSMutableArray, the retain count is increased by 1, to let the array hold onto the object.
  • In UIKit, such as when a UIView object is added as a subview, its retain count is increased by 1.  When this subview is removed from the superview, the retain count of the subview is decreased by 1.
  • When a certain array element is replaced by another object reference, the old object is released once, while at the same time, the new object is retained once.  This works the same way as a retain property.