Programming

There are 82 entries in this Category.

Enough with the NSStatusItems! Teach your app its place!

I see more and more applications implemented as NSStatusItems in the upper right side of the menu bar. In this posting, I’ll lay out why this is a worrying development, and why you should rarely implement NSStatusItems.

Screen real estate

The menu bar is very precious screen real estate, and the most expensive part of your computer. It takes up a permanent 22 points at the top of your screen (if you have several screens, it may even show up on every screen). The menu bar is fixed in position and size, different from other windows, and no other window can inhabit these sacred pixels. You can’t switch it behind another window. It is always visible, always immediately clickable.

It is also used for an important part of the user interface of the current application. All of an application’s menus have to fit into this area. There is no scrolling, no wrapping to a second line.

Perspective of importance

One of the fundamental rules of UI design is to arrange UI elements by their importance. Things that provide information the user constantly needs to be aware of, or that are constantly used should always be in view/at a single-click range, while things the user uses less can be relegated to less easily reachable spots that might require several clicks to get to.

The document window (or main window in the case of a shoebox application like iPhoto) is the top of this hierarchy. That’s what the user works with most of the time and where her attention is focused. Floating palettes are also near the top.

Things you can’t put directly in front of the user like that go in a menu, where the user needs to click to discover them or trigger them. If something is even less important or needs to display information more complex than is desirable to put in a menu item, it can go in an auxiliary window shown by a menu item.

Popovers, while relatively new to the scene, are kind of halfway between these two. On one hand you need to click to open them, like a menu, on the other hand you can’t have as many of them as you can have menus. They also occupy a half-way position between a menu and a modal window. They can contain more complex controls.

NSStatusItem

So, now that we know how limited room in the menu bar is, and how it is the second go-to location after you’ve run out of main window space, where does NSStatusItem fit in here?

Well, NSStatusItems can show information in their icon, and otherwise work like a menu. They can immediately react to a click (like the “Notifications” icon in the upper right of the screen) or show a menu, or a popover.

They are also visible across all applications. As such, they are a permanent, most reliable fixture in the user interface. Always visible, always clickable. It is prime real estate if there ever was one.

From this follows that it should only hold functions that inhabit exactly this place for the user: Something that is needed no matter what application is frontmost. Something that is constantly needed, not just occasionally when the user is working on one particular project. Or something that indicates some important piece of information, like how long the computer’s battery will last.

The reality of status items

Compare that to the reality we’re living with today: Every Twitter client I’ve used so far had a status item by default. A status item and a dock icon. At the time of this writing I’ve written well over 57’000 tweets, but even I don’t think that Twitter is that important. One dock icon is fine for seeing new tweets and posting a new one. It’s one click away.

I’m sure some users disagree, but really, is that the majority? Does it have to add that status item and take up dock space by default? Can’t it just leave this as a feature that the user can activate if they think it is needed?

Similarly, there are applications that perform periodic clean-up tasks in the background. Maintenance. Do I really need to see those applications’ icons in my menu bar permanently? Couldn’t they just show their icon when they are doing work, then remove it again? Couldn’t they be a GUI front-end with a background helper application that magically does its work? How often do I manually need to trigger a re-scan of my movies folder to see if it contains new files if the application watches the folder for changes anyway? If this really is just a workaround for rare bugs, why not make me launch the GUI front-end to achieve that and stay out of my menu bar?

There are applications that let me run a server, for testing, locally, on my computer. Why can’t they just be a regular GUI front-end with the server as an invisible background process? Why can’t they just add a bookmark file somewhere that I can launch using Spotlight instead of making me use a different item in the precious status item area of the screen to open the URL for that server?

Why does everyone have such an inflated sense of the importance of their app that they need to have an icon in the menu bar?

Cocoa and the Builder Pattern

There’s been a nice discussion about the Builder pattern on Twitter today. The Builder pattern is a nice tool to have, particularly because it addresses a few common problems.

What Builder Pattern?

In short, the Builder pattern is a pattern where you have one object that you configure that then creates another object based on that configuration. The nice thing here is that you can first build your object step by step, like you’d e.g. do with NSMutableString, but then the actual construction of the object happens in one go. Very handy for immutable objects.

Usually, a setter for a Builder object returns self, like retain or autorelease do. That way, you can create something in Java or C++ that almost looks like Objective C:

Image theImage = (new Image.Builder)->SetWidth(100)->SetHeight(80)->SetDepth(8)->Build();

Where the Build() method releases the builder and returns the actual, immutable Image object.

Extending init methods

When you add a parameter to an initializer in Objective-C, it is annoying. You usually add the parameter to the initializer, then create a compatibility version with the old method’s name that calls the newer version with a default value for the extra parameter.

Java and C++ have solved that problem by allowing you to specify default values for parameters, but they don’t maintain binary stability that way. If you add a parameter, you still have to recompile, but at least you don’t need to change your code.

I guess one fix would be if ObjC supported default arguments to a parameter that would simply result in the creation of a second version of this initializer with the label and parameter removed:

-(id) initWithBanana: (NSBanana*)theBanana curvature: (CGFloat)curvature = 5
{
    // magic happens here
}

Would be the same as writing:

-(id) initWithBanana: (NSBanana*)theBanana curvature: (CGFloat)curvature
{
    // magic happens here
}


-(id) initWithBanana: (NSBanana*)theBanana
{
    return [self initWithBanana: theBanana curvature: 5];
}

Of course, you’d still need at least one parameter, because ObjC has no way of knowing what part of the message is the name, and what is the label for the second (for init there could be special code, I guess, but what for a -exfoliateCow:withSpeed: method?). And defaulting to -initWithBanana if the first parameter has a default is obviously not always desirable either. It would solve the annoyance of telescoping constructors, at the least.

The Builder pattern doesn’t have this problem. Each parameter has a setter that you use to set it. A new builder could have defaults for all parameters when it is created. Then you change the ones you want to customize, and call -build on it to get the new object. If a new setter is added, that’s fine. You don’t call it, you get the default. The maintainers only add the one setter, no compatibility method needed.

Thread safety and immutable objects

The easiest way to get thread safety is to prohibit data from changing. If data is immutable, there is nothing to be synchronized between threads,and no need for one thread to wait for the other. However, immutable objects are also annoying, as they need to be fully specified in their init method.

A case where this is a problem in Cocoa is NSImage. NSImage is an immutable object by convention, but not actually. It is an object that has its own builder built in. You are expected to know that, for an NSImage to be thread safe, you are expected to create it, set its attributes, draw something in it, and then stop messing with it, treating it as an immutable, read-only object from then on.

The problem is, nobody enforces it. NSImage is a perfectly mutable object, with setters and getters. There is no exception thrown when you violate this verbal contract. Of course Apple could have added a “makeImmutable” method to NSImage that causes those exceptions to happen when you try to edit an instance. But then they’d have to add code to each setter that errors (Or at the least use some aspect-oriented-programming mechanism to inject code before every setter that performs this check automatically).

The Builder pattern would solve that: They can have a huge, private constructor on NSImage that changes with every release to add new parameters and initialize that immutable object, while the Builder would present a stable and convenient API to all clients. There would not be any setters on NSImage.

But it is ugly…

Admittedly, it feels a bit inelegant to build an object that builds an object. The way NSImage works is so much nicer. But Mike Lee actually offers a neat approach that works almost as well:

Just pass in a list of properties. This could be a dictionary of properties, or even just a variadic argument list like -dictionaryWithObjectsAndKeys: takes it. You’d define a constant for each possible property (that way if you mis-type the parameter name the compiler tells you, which you don’t get from a raw string). Internally, this constant could even hold the actual name of the property, even if it is never exposed as a method in the public header. So, all your constructor would do is call [self setValue: properties[key] forKey: key] in a loop, once for every element.

You get the same effect as labeled parameters (if you put the keys first, even more so). You also get the same effect as optional parameters. The binary ABI never changes, so that’s good, too. The only downside is you need to pass every parameter as an object, and you lose compile-time type checks. OTOH you gain compile-time errors when you try to change the object after creating it (because it declares no setters).

Is it worth all that work?

Admittedly, I haven’t had to add parameters to the init method of a public class that often. Nonetheless, I think Mike’s approach and the Builder pattern both are useful things to keep in mind if you ever come up with a class that can be created in numerous configurations (and is likely to gain new properties in the future) but should then be immutable. Class clusters and plug-in classes seem like a typical place where you might need this.

Are your rectangles blurry, pale and have rounded corners?

One common problem with drawing code in Cocoa (iOS and Mac OS X) is that people have trouble getting crisp, sharp lines. Often this problem ends up as a question like “How do I get a 1-pixel line from NSBezierPath” or “Why are my UIBezierPath lines fuzzy and transparent” or “Why are there little black dots at the corners of my NSRect”.

The problem here is that coordinates in Quartz are not pixels. They are actually “virtual” coordinates that form a grid. At 1x resolution (i.e. non-Retina), these coordinates, using a unit commonly referred to as “points” to distinguish them from act pixels on a screen (or on a printer!), lie at the intersections between pixels. This is fine when filling a rectangle, because every pixel that lies inside the coordinates gets filled:

filled_rectangle_between_pixels

But lines are technically (mathematically!) invisible. To draw them, Quartz has to actually draw a rectangle with the given line width. This rectangle is centered over the coordinates:

coordinates_between_pixels

So when you ask Quartz to stroke a rectangle with integral coordinates, it has the problem that it can only draw whole pixels. But here you see that we have half pixels. So what it does is it averages the color. For a 50% black (the line color) and 50% white (the background) line, it simply draws each pixel in 50% grey. For the corner pixels, which are 1/4th black and 3/4ths black, you get lighter/darker shades accordingly:

line_drawing_between_pixels

This is where your washed-out drawings, half-transparent and too-wide lines come from. The fix is now obvious: Don’t draw between pixels, and you achieve that by moving your points by half a pixel, so your coordinate is centered over the desired pixel:

coordinates_on_pixels

Now of course just offsetting may not be what you wanted. Because if you compare the filled variant to the stroked one, the stroke is one pixel larger towards the lower right. If you’re e.g. clipping to the rectangle, this will cut off the lower right:

coordinates_on_pixels_cut_off

Since people usually expect the rectangle to stroke inside the specified rectangle, what you usually do is that you offset by 0.5 towards the center, so the lower right effectively moves up one pixel. Alternately, many drawing apps offset by 0.5 away from the center, to avoid overlap between the border and the fill (which can look odd when you’re drawing with transparency).

Note that this only holds true for 1x screens. 2x Retina screens exhibit this problem differently, because each of the pixels below is actually drawn by 4 Retina pixels, which means they can actually draw the half-pixels needed for a 1 point wide line:

coordinates_between_pixels_retina

However, you still have this problem if you want to draw a line that is even thinner (e.g. 0.5 points or 1 device pixel). Also, since Apple may in the future introduce other Retina screens where e.g. every pixel could be made up of 9 Retina pixels (3x), you should really not rely on fixed numbers. Instead, there are now API calls to convert rectangles to “backing aligned”, which do this for you, no matter whether you’re running 1x, 2x, or a fictitious 3x. Otherwise, you may be moving things off pixels that would have displayed just fine:

coordinates_on_and_between_pixels_future_retina

And that’s pretty much all there is to sharp drawing with Quartz.

The fast road to unit tests with Xcode

Supposedly Xcode has unit test support. I’ve never seen that work for more than two Xcode revisions. So I’ve come up with a minimal unit test scheme that works reliably.

1) Add a “command line tool” target (Foundation application, C++ application, whatever makes sense). Put your test code in its main.m or whatever. After each test, print out a line starting with “error: ” if the test failed. If you want to be able to see the successes as well, start them with “note: “. Keep a counter of failed tests (e.g. in a global). Use the number as the app’s return value of your main().

2) Add a “Run Shell Script” build phase to this target, at the very end. Set it to run ${TARGET_BUILD_DIR}/${PRODUCT_NAME}. Yes, that’s right, we make it build the unit test app, then immediately run it. Xcode will see the “error: ” and “note: ” lines and format them correctly, including making the build fail.

3) Optionally, if you want these tests to run with every build, make that command line tool target a dependency of your main app, so it runs before every build. Otherwise, just make sure your build tester regularly builds this test target.

4) Add a preprocessor switch to the tests that lets you change all “error:” lines into “warning:” instead. Otherwise, when a test fails, you won’t be able to run it in the debugger to see what’s actually going wrong.

How to write a compiler

A bytecode interpreter feeding instructions through it.Since there isn’t that much beginner info out there about this topic, here’s a very rough run down of what I know about the basics of writing your own compiler, in particular about how the CPU works and how to generate code for it.

CPU/bytecode interpreter

A bytecode interpreter works just like a CPU, the difference being that it is in software, while the CPU is actual hardware. So all a fake or real CPU does is take a list of instructions and fetch them one by one.

To properly do that, there is one variable (in a real CPU, this is a register) that contains the position of the current instruction. This is called the program counter or PC for short, and is basically just the memory address of whatever command it is to execute.

Once an instruction finishes, the CPU adds to the PC to make the pointer point at the next instruction (or in the case of a conditional or loop, it rewinds the PC back to the start of the loop, or jumps over the ‘else’ section, or whatever.

So it’s fairly easy to create a bytecode interpreter. It’s just a loop:

#define NO_OP        0
#define PRINT        1
#define END          2
struct Instruction { int instructionType; int param1; int param2; };
Instruction *currentInstruction = LoadCodeFromDisk();

while( currentInstruction )
{
    if( instructionType == NO_OP )
        currentInstruction++;
    else if( instructionType == PRINT )
    {
        DoPrint( currentInstruction->param1, currentInstruction->param2 );
        currentInstruction++;
    }
    else if( instructionType == END )
        currentInstruction = NULL;
    else
        exit(1); // UNKNOWN INSTRUCTION! BLOW UP!
}

So all that generating machine code or byte code is, is really adding items to an array of structs. Of course, if you want to generate Intel machine code it’s a little more complicated, because instructions can be different size, so you can’t use a classic array, but you can write the raw data to a file or memory block just the same.

Generating code that jumps

If you’ve ever programmed BASIC, you’ve probably seen the following program:

Text with arrows indicating the progression from Print to Goto back to Print1 PRINT "Hello World"
2 GOTO 1

This is an endless loop. Line 1 prints some text to the screen, and line 2 jumps back to line 1. Once line 1 is done, we continue to line 2 again, which jumps back to line 1, forever and ever until someone turns off the computer. So all GOTO does is change the currentInstruction from our above example, the program counter.

currentInstruction = 1;

You can implement GOTO the same way in your bytecode interpreter. However, since you usually don’t know what address your code will be loaded at (and it definitely won’t be address 1), you will generally write your code so it jumps relative to the current instruction’s location. So our version of GOTO, the JUMPBY instruction, would be

currentInstruction += currentInstruction->param1;

For our pseudo-machine-code:

PRINT "Hello World"
JUMPBY -1

With this instruction under your belt, you can quickly implement conditional statements like if. An if-instruction is simply an instruction that looks at a location in memory (whose address could be provided as param2) and if that location is 1 (true), jumps by param1. Otherwise it does the usual currentInstruction++.

The conditional GOTO is the basic building block of all flow control. If/else:

Flow from 1,2,5,6 for true case,  2,3,4,6 for false case.1 FOO=1
2 GOTO 5 IF FOO
3 PRINT "Foo is false."
4 GOTO 6
5 PRINT "Foo is true."
6 END

loops:

1 FOO=0
2 GOTO 6 IF FOO
3 PRINT "Repeating."
4 DO_SOMETHING_THAT_COULD_CHANGE_FOO
5 GOTO 2
6 PRINT "End of loop"
7 END
While loop execution order: 1,2,3,4,5,2 and then if FOO is 1, from there to 6 and 7, otherwise on to 3 again etc.

Note that bytecode has no operators, no expressions, no precedence. You provide operations in the order it is supposed to execute them. If you want to compare two strings, you do so in the instruction at the top of the loop, save its result to FOO, then loop over FOO:

1 FOO=COMPARE("a","b")
2 GOTO 6 IF FOO
3 PRINT "Repeating."
4 DO_SOMETHING_THAT_COULD_CHANGE_FOO
5 GOTO 1
6 PRINT "End of loop"
7 END

(Note how line 5 jumps to line *1* here, i.e. every time through the loop, the condition is evaluated, then the conditional GOTO tests it)

Retroactive code generation

Now how do you generate code for this? How do I know, before I have read and generated the individual instructions, what line the GOTO in line 2 will have to jump to?

Well, you don’t. Instead, what you do is write out the GOTO as

2 GOTO 0 IF FOO

and then later, when you reach the end of the loop in line 5, where you write out the GOTO that jumps back to the condition, you simply change the destination of line 2 after the fact. This is usually fairly easy if you write a function for parsing every syntax element. Take the following C-like program:

strcpy( txt, "a" );
while( compare(txt,"b") )
{
    print("repeating");
    do_something_that_could_change_txt();
}
print( "End of loop" );

You’d have a function CompileOneLine() that reads one line and looks at the first word. If it is “if” it calls CompileIfLine(), if it is “while” it calls CompileWhileLine(), “print” – CompilePrintLine() etc.

CompileWhileLine would look something like:

void CompileWhileLine()
{
    int conditionOffset = ReadAndCompileExpression( "FOO" );
    int conditionalGotoOffset = WriteInstruction( "GOTO IF", 0, "FOO" );
    if( NextChar() == '{' )
    {
        SkipChar('{');
        while( NextChar() != '}' )
            CompileOneLine();
    }
    else
        CompileOneLine();
    int gotoOffset = WriteInstruction( "GOTO", conditionOffset );
    SetDestinationOfGoto( conditionalGotoOffset, gotoOffset );
}

And since we call CompileOneLine() again to read the lines inside this while statement, we can nest while statements.

Variables

As demonstrated, this byte-code has one big downside: How do we do variables? We can’t put the variables in with the code, because that would mean that when a function calls itself, it would use the same variables as its previous iteration. So we need some sort of stack to keep track of the most recent function’s variables.

And this is what the stack in programming languages like C is: The code for a function has a prolog and an epilog. That is, a few instructions at the start, before any actual commands that the function contains, where it makes the stack larger to reserve space for the variables we need, and a few more after the end to get rid of them again. In our BASIC-like pseudocode, this could look like:

PUSH 0 // Reserve space for 3 int variables:
PUSH 0
PUSH 0
// ... actual code goes here
POP // Get rid of no longer needed 3 int variables
POP
POP

Now since we sometimes need a variable just for a short while (e.g. FOO for holding the loop condition’s result to hand off to the conditional GOTO instruction), it would be kind of awkward to find our variables by counting from the back of the stack. So for that we have a base pointer. A base pointer is another variable/register, like our program counter before. Before we push our variables on the stack, we write the size of the stack into the base pointer.

SavedBasePointerPUSH basePointer // Save the old basePointer (for whoever called us)
SET_BASEPOINTER // Write current stack size to basePointer
PUSH 0
PUSH 0
PUSH 0
// ... actual code goes here
POP
POP
POP
POP_BASEPOINTER // Restore the saved base pointer into the basePointer

Now, to use a variable, we can simply access it relative to the basePointer, i.e. basePointer[0]=1. No matter how many variables we add, these numbers stay the same for the lifetime of our function.

ParametersOnStackAnd once we have this, we can also implement parameters. Parameters are simply variables at a negative offset. Since whoever calls us pushes parameters on the stack right before calling us, they all can be found right before our saved basePointer. So basePointer[-2] is parameter 1, basePointer[-3] parameter 2 etc. This means the parameters need to be pushed in reverse order, but that’s all there is to it.

Returning

Given the above, it comes as no surprise that you can’t just return from your function. You need to make sure that, before you return, your clean-up code runs. So what we usually do is, we make a note of the position at which the clean-up code starts, put a RETURN statement right at its end, and then whenever the code wants to return, we just write the return value somewhere on the stack, GOTO the clean-up code and let it return.

Of course, you’ll have to remember all spots that GOTO that code, because when you write them, the clean-up code won’t have been written yet, and fill it out in retrospect. But if you got this far, that’s all old hat to you.

Strings and other (bigger) data

Illustration of a few instructions followed by a block of data with arrows from the instructions to their particular stringsAs you may have noticed, our instructions only have two ints as parameters. What if you need to e.g. provide a format string to printf? Well, you’ll need an instruction that loads a string constant. Pushes its address on the stack, which the print command can then grab and display.

Commonly, what one does is simply write the strings in the same file (or memory block) as the bytecode, e.g. right after the actual instructions, and then load the whole thing into RAM. Since you know how big your block of code is after you’ve written it, you can retroactively fill out the offset from the instruction to the string it is supposed to push on the stack, and at runtime, the instruction can use currentInstruction +n to get a pointer to it and push that on the stack.

And that’s pretty much how you create a bytecode interpreter and generate bytecode (and in a simplified form, this is how you generate machine code).

Cocoa: String comparisons and the optimizer

Woman in front of a mirrorA while ago, a friend came to me with this bit of code:

NSString *a = @"X";
NSString *b = @"X";
if( a == b )
{
    NSLog(@"Same!");
}

“How come it works with the == operator? Didn’t you have to call isEqualToString: in the old days?”

Before we answer his question, let’s go into what he implicitly already knew:

Why wouldn’t == work on objects?

By default, C compares two pointers by simply comparing the addresses. That is logical, fast, and useful. However, it is also a little annoying with strings, arrays and other collections, because you may have two collections that still contain identical objects.

If you have the phone books from 2013 and 2014, do you just want to compare the numbers 2013 and 2014 and be told: “No that’s not the same phone book”, or are you actually interested in whether their contents are different? If nobody’s phone book entry changed in a particular city, wouldn’t you want to know that and save yourself the trip to the phone company to pick up a new phone book?

Since all Objective-C objects are pointers, the only way to do more than compare the addresses needs some special syntax. So NSString offers the isEqualToString: method, which, if the pointers do not match, goes on to check their contents. It compares each character to the same position in the second string to find out whether even though they’re not the same slip of paper, they at least have the same writing on it.

So why does the code above think they’re the same?

After all that, why does the code above think they are the same object after all? Doesn’t a point to the @"X" in the first line, b to the @"X" in the second line?

That is what is conceptually true, what a naïve compiler would do. However, most compilers these days are smart. Compilers know that a string constant can never change. And they see that the contents of both string objects pointed to by a and b are the same. So they just create one constant object to save memory, and make both point to the same object.

There is no difference for your program’s functionality. You get an immutable string containing “X”.

However, note that there is no guarantee that this will happen. Some compilers perform this optimization for identical strings in one file, but not across files. Others are smarter, and give you the same object even across source files in the same project. On some platforms, the string class keeps track of all strings it has created so far, and if you ask for one it already did, gives you that, to save RAM. In most, if you load a dynamic library (like a framework) it gets its own copy of each string, because the compiler can not know whether the surrounding application actually already has that string (it might be loaded into any arbitrary app.

Mutability is important

This is a neat trick compilers use that works only with immutable objects, like NSString or NSDictionary. It does not work with NSMutableString. Why? Because if it put the same mutable string into a and b, and you call appendString: on a, b would change as well. And of course we wouldn’t want to change our program’s behaviour that way.

For the same reason, NSString may be optimized so that copy is implemented like this:

-(id) copy
{
    return [self retain];
}

That’s right. It gives you the same object, just with the reference count bumped, because you can’t change this string once it has been created. From the outside it looks the same. Copy gives you the object with its retain count bumped, so you can release it safely once you’re done with it. It behaves just like a copy. The only hint you may have that this happened is that instead of an NSString with reference count owned solely by you, you get one with a reference count of 2 whose ownership you share with another object. But that’s what shared ownership is about after all.

Of course, this optimization doesn’t work with NSMutableString.

What I take away from this

So if someone walks up to you and shows you code that uses the == operator where it should really be checking for content equality, and argues that “it works, so it is correct”, now you’ll know why it just happens to work:

It’s a fluke, and if Apple decides to switch compilers or finds a better way to optimize performance or memory usage that requires them to no longer perform this optimization, they might just remove it, and this code will break, because it relied on a side effect. And we don’t want our code to break.

The universal catch-all singleton

Personified application delegate creating objects by cracking eggsOne bad habit I see time and time again is filling up your application delegate with junk that has no business being in there.

Avoid putting stuff in your app delegate.

What is the application delegate?

By definition, the application delegate is a controller. Most model classes are standard Apple classes. Those that aren’t are slightly smarter collections of these classes. Most view classes just display the model and forward a few IBActions to the controller in a dynamic way, so are inherently reusable as well (even if not always the particular arrangement of their instances).

The controllers, on the other hand, aren’t really reusable. They glue all this stuff together. They’re what makes your application your application, along maybe with a few bindings. So, again by definition, the application delegate is the least reusable part of an application. It’s the part that kicks everything off, creates all the other controllers and has them load the model and views.

Reusability is best!

The whole point behind OOP was to reduce bugs, speed up development, and help structure your code by keeping it grouped in reusable components. The best way to maintain this separation of components and permit re-using parts of it in different projects, is to keep the boundaries between components (e.g. objects) clean and distinct.

Objects have a clear hierarchy. The top creates objects lower down and gives it the information they need to operate correctly for your application. Nobody reaches up the hierarchy, except maybe to notify whoever created it of occurrences in the form of delegate messages. That way, the more application-specific your code gets, the fewer other objects know about it. The further down you get, the more reusable.

Moving operations or instance variables that are shared by several objects in your application into the application delegate, and having other objects reach straight up through NSApplication.sharedApplication.delegate to get at it, goes head-on against this desire, and turns your carefully separated code into an inseparable glob of molten sludge. Suddenly *everybody* includes the most application-specific header your application contains.

Don’t lie to yourself

The application delegate is one of the singletons every application contains. Its name is misleading and fuzzy. If you see it as a place to hold code “relevant to the application as a whole”, there is pretty much nothing that is off-topic for it. It is the universal, catch-all singleton.

So why not be honest to yourself: Whenever you add code to the application delegate, and you’re not just doing it to react to a delegate method from NSApplication and create a controller to perform the actual action in response, what you are really doing is create a singleton.

As we all know, singletons have a use, but having many singletons is a code smell. So avoid them if you can, but if you feel you can’t, be honest to yourself and actually make it a separate singleton (or find one whose purpose these operations fit).

Just say no to the application delegate as the universal singleton.

Update: If this hasn’t convinced you, here’s another blogger with a more CoreData-centric discussion of the issue, coming to the same conclusion: Don’t abuse the app delegate.

What a block really is

BlocksLegoBrickAfter quite a while of thinking that Objective-C blocks did some mean magic on the stack, it simply took me seriously using C++’s lambdas (their implementation of the concept) that I realized what blocks are.

Effectively, a block is simply a declaration of a class, plus an instantiation of one instance of that class, hidden under syntactic sugar. Don’t believe me? Well, let’s have a look at C++ lambdas to clear things up:

MyVisitorPattern( [localVariableToCapture]( MyObject* objectToVisit ) { objectToVisit->Print(localVariableToCapture); }, 15 );

The red part is a C++ block. It’s pretty much the same as an Objective-C block, with two differences:

  1. You explicitly specify which local variables to capture in square brackets.
  2. Instead of the ^-operator, you use those square brackets to indicate that this is a block.

Seeing the captured variables specified explicitly listed here, like parameters to a constructor, made me realize that that’s really all that a block is. In-line syntax to declare a subclass of a functor (i.e. an object whose entire purpose is to call a single of its methods), and return you an instance of that class. In ObjC-like pseudo-code, you could rewrite the above statement as:

@interface MYBlockSubClass : NSBlock
{
    int localVariableToCapture;
}

-(id) initWithLocalVar: (int)inLocalVariableToCapture;

-(void) runForObject: (MyObject*)objectToVisit;

@end

@implementation MYBlockSubClass
-(id) initWithLocalVar: (int)inLocalVariableToCapture
{
    self = [super init];
    if( self )
        localVariableToCature = inLocalVariableToCapture;
    return self;
}

-(void) runForObject: (MyObject*)objectToVisit
{
    objectToVisit->Print(localVariableToCapture);
}
@end

and at the actual call site:

MyVisitorPattern( [[MYBlockSubClass alloc] initWithLocalVar: localVariableToCapture], 15 );

The difference is that C++ (and even more so Objective-C) automatically declare the class for you, create the instance variables and constructor for the variables you want to capture, pick a unique class name (which you can see in the stack backtraces if you stop the debugger inside a block) and instantiate the class all in a single line of code.

So there you see it, blocks aren’t really black magic, they’re 99% syntactic sugar. Delicious, readability-enhancing syntactic sugar. Mmmmmh…

PS – Of course I’m simplifying. Objective-C blocks are actually Objective-C objects created on the stack, which you usually can’t do in plain Objective-C, though it can be done with some clever C code if you insist.

A more magical approach to blocks

That said, there is a fancier way for a compiler developer to implement blocks that also makes them 100% compatible with regular C functions:

If you implement a function in assembler, you can stick additional data onto the end of a function and calculate an offset between an instruction and the end of the function (e.g. by just filling the end of the function with a bunch of 1-byte No-Ops). This means that if someone duplicates a block, they’ll duplicate this data section as well. So what you can do is declare a struct equivalent to the captured variables, and implement your code with (pseudocode):

void    MyBlock( void )
{
struct CapturedIVars * capturedIVars = NULL;

currentInstruction:
    capturedIVars = pointer_to_current_instruction + ivarsSection-currentInstruction;

    // Block's code goes here.
    
    goto ivarsSectionEnd; // Jump over ivars so we don't try to execute our data.
    
ivarsSection:
    assembler_magic_to_reserve_some_space;
ivarsSectionEnd:
}

Now you can use the capturedIVars pointer to access the data attached to your function, but to any caller, MyBlock is just a plain old function that takes no arguments. But if you look at it from a distance, this is simply an object prefixed with a stub that looks like a function, so our general theory of blocks being just special syntax for objects holds.

I presume this is how Swift implements its blocks, because it really doesn’t distinguish between blocks and functions.

Mapping Strings to Selectors

MappingStringsToSelectorsSketchBack in the old days of Carbon, when you wanted to handle a button press, you set up a command ID on your button, which was a simple integer, and then implemented a central command-handling function on your window that received the command ID and used a big switch statement to dispatch it to the right action.

In Cocoa, thanks to message sending and target/action, we don’t have this issue anymore. Each button knows the message to send and the object to send it to, and just triggers the action directly. No gigantic switch statement.

However, we still have a similar issue in key-value observing: When you call addObserver:forKeyPath:options:context, all key-value-observing notifications go through the one bottleneck: observeValueForKeyPath:ofObject:change:context:. So, to detect which property was changed, you have to chain several if statements together and check whether the key path is the one you registered for (and check the ‘context’ parameter so you’re sure this is not just a KVO notification your superclass or subclass requested), and then dispatch it to a method that actually reacts to it.

It would be much nicer if Apple just called a method that already contained the name of the key-value-path, wouldn’t it? E.g. if the key-path you are observing is passwordField.text, why doesn’t it call observerValueOfPasswordField_TextOfObject:change:context:?

But there is a common Cocoa coding pattern that can help us with this: Mapping strings to selectors. The centerpiece of this method is the NSSelectorFromString function. So imagine you just implemented observeValueForKeyPath:ofObject:change:context: like this:

-(void) observeValueForKeyPath: (NSString*)keyPath ofObject: (id)observedObject change: (NSDictionary*)changeInfo context: (void*)context
{
    NSString *sanitizedKeyPath = [keyPath stringByReplacingOccurrencesOfString: @"." withString: @"_"];
    NSString *selName = [NSString stringWithFormat: @"observeValueOf%@OfObject:change:context:", sanitizedKeyPath];
    SEL      action = NSSelectorFromString(selName);
    if( [self respondsToSelector: action] )
    {
        NSInvocation * inv = [NSInvocation invocationWithMethodSignature: [self methodSignatureForSelector: action]];
        [inv setTarget: self]; // Argument 0
        [inv setSelector: action]; // Argument 1
        [inv setArgument: &observedObject atIndex: 2];
        [inv setArgument: &changeInfo atIndex: 3];
        [inv setArgument: &context atIndex: 4];
        [inv invoke]
    }
    else
        [super observeValueForKeyPath: keyPath ofObject: observedObject change: changeInfo context: context];
}

We build a string that includes the name of the key-path, turn it into an actual selector, and then use -performSelector:withObject:, or in more complex cases like this one NSInvocation, to actually call it on ourselves.

For cases that have no clear mapping like this, you can always maintain an NSMutableDictionary where the key is whatever string your input is and the value the selector name for your output, and then use that to translate between the two. When you make whatever call equivalent to addObserver: you have in that case, it would add an entry to the dictionary. That’s probably how NSNotificationCenter does it internally.

Update:
As Peter Hosey pointed out, another good use case for this pattern is -validateMenuItem: where one could turn the menu item’s action into a string and concatenate that with ‘validate’.

Fuzzy words in Programming

UnreadableDictionaryThe hardest problem in programming is not proving correctness, but naming things. Frequently, I find myself looking at a new codebase, or a section of a larger codebase I haven’t been exposed to so far, and I have problems determining what this code is supposed to do.

Sometimes this is due to 400-line-functions that should have had their different code paths split up into separate functions ages ago, but other times, the code is well-structured and I’m familiar with the domain’s terminology, but the naming of the variables and functions and classes still actively hinders understanding. Why is that? How can this be?

Class of word

One thing I’ve seen is that identifier names don’t follow the grammar of a sentence or don’t use the proper class of word. If a method triggers an actual process, an action as the user would understand it, it should be a verb. setFoo: applies a change. If a method named commandProcessor actually runs a loop that processes commands, it should be renamed runCommandProcessingLoop instead. That way, when I read it somewhere, even if I haven’t read its contents yet, I have a general idea what it does.

If it doesn’t contain a verb, I’ll have to make up my own. What could it be? createCommandProcessor? resetCommandProcessor?

Of course, there are fuzzy spots. If NSViewController‘s view method implicitly loads and creates the view, shouldn’t it be called loadView instead? Or is this an implementation detail that I can leave out? In general, I tend to count caches and lazy-loading that happens as needed as implementation details. So your property does not have to be named as a verb.

Cases and modifiers

Another thing that I see a lot, particularly in codebases that have a lot of non-native English speakers working on them, are problems with cases. This can take the shape of reversed word-order that has the defining word of a compound at the wrong end, like ListChannel instead of ChannelList.

This can be very unfortunate, as in English, many verbs are identical to their nouns, and only differ in their position in the sentence. So when I see ListChannel I don’t just think this is a single channel that somehow contains or presents a list, but rather I might read it as a verb. When someone declares a variable of this type, my interpretation, in order of likelihood is therefore:

  1. A block or functor type that somehow lists channels.
  2. A channel that somehow lists something.
  3. A mis-named list of channels.

Where the right one, in our example, is the 3rd and least likely choice.

Another common mistake is excessive use of plurals, like MyChannelsList. This may look harmless at first, but can in some cases lead to ambiguity. E.g. if your project convention is to name typed arrays as the plural of your element type, this would be a list of Channels objects, which in turn are lists of Channel objects. Now, ignoring the (lack of) wisdom of having a one-character difference in names, excessive plurals like that make me mistrust the code and force me to check whether there is a reason for this extraneous plural or not.

Fuzzy words

The third class of bad names that I see in projects are “fuzzy” words. The most common case are class and file names that don’t have a clearly delineated meaning and therefore attract functionality like one of those magnetic cranes in a junk yard.

Utilities.h
NetworkExtras
MYOtherDebugMenuSection
StringExtensions
NSURL+Helpers.h

I’m sure you can come up with more examples, but they all have one thing in common: They are so generic that almost no method or function is off-topic for this file. And when you’re actually searching for a particular method or function in a foreign codebase, any of these files is probably a good candidate.

While Utilities.h is obviously a catch-all into which everyone can dump whatever they want when they are too lazy to add a new file to a project “just for this one function”, or want to avoid having to merge project files if someone else also just added another file to the project, the others are less obvious, but just as magnetic:

The network is a big place, with many layers. Is this just a wrapper to convert certain data types before they’re sent over a socket? Are these actual protocol implementations? URL string-formatting functions? So much could be valid for a class named NetworkExtras.

I worked on one project that contained a debug menu that scrolled for 3 screens, and where none of the menu items could be re-used in other projects easily because all the code for them was in the application delegate, causing that to need a recompile every time one of them changed.

So for the next project, I created a MYDebugMenuSection class whose +load method adds it to a list which is then used to build the debug menu. Each of these classes is one submenu. So I now have a generic MYViewsDebugMenuSection that turns on/off lots of view-related things (like drawing their frames, flashing updates etc.) or logs the view hierarchy. To every new project, we can just add that header and source file, and get the same debug features we have in every other app.

Then someone found that too onerous, and created an “Other” submenu, and shortly, another developer added another bunch of items to this menu instead of creating their own section for their submodule. All we gained was that you now had to navigate to a submenu before you were forced to scroll through your 3 screens full of items. So I split them, and renamed them, to avoid the magnetism of the fuzzy “Other”.

StringExtensions? Same problem. There are lots of things you do with strings. Name it StringURLFormattingExtensions, or be prepared to have to split the file every few weeks. Same with NSURL+Helpers.

Magnetism

Some of you may not believe me. And I don’t blame you, it’s easy to overshoot the target. To create headers that are too concrete and really only contain one function. I agree, that’s bad. Avoiding that requires common sense, too. But if you don’t believe that crap is magnetic, go to an underpass that is not publicly visible from the outside and has a big empty flat wall that has been freshly painted. Visit it regularly.

At first, it’ll be virginally white. Then, someone will paint a tag on it, or stick a sticker on it. Then it’ll stay like that for a while. Then the next person will come along, and she’ll think: “Hey, nobody removed the tag. Let’s put mine on there as well!” And then there’ll be more and more tags, until finally someone will slap a gorgeous graffiti over it. And then someone else will add another one, and a few taggers will tag in empty spots on the graffiti …

This even works with advertisements. Here in Munich, under the Heimeranplatz train overpass, there was an insurance advertisement that, as part of its picture, contained a few scrawls that, blown up to billboard size, looked just like tags. Within a month, the billboard had accumulated another dozen or two tags on “what was obviously a derelict billboard”.

It works with garbage: One person stuffs a paper cup in a hole in the wall, other people throw their trash there as well.

It works with code. Send the bin man round every few weeks. You’ll love your code’s fresh smell.