Retina for Masochists

Today we released an update for xScope that supports the Retina display. As I alluded to on Episode 14 of The Talk Show, this update was harder than most. The 68k to PowerPC, Carbon to Cocoa, and PowerPC to Intel transitions were no walk in the park, but this update really kicked my butt. Here’s hoping that sharing some of the things I learned along the way will help you with your own Retina work.

For most developers who are working strictly in window points, an update for the Retina display is a fairly straightforward process. xScope, however, does a lot of work with both pixels and points. And that’s where the fun begins…

Mouse Input

The first gotcha I encountered while doing the Retina update was with mouse input using NSEvent’s +mouseLocation. The team at Apple has done some amazing work making sure output looks stunning on the Retina display, but being able to get high-resolution input is definitely lacking.

There are two problems at play here. The first is that mouse coordinates can be reported for coordinates that do not exist on any attached screen. The second is that the NSPoint does not contain enough resolution to address every pixel on screen.

To deal with the first problem, I used an NSEvent category that clamps +mouseLocation results to valid coordinates.

For the second problem, the only workable solution was to capture the +mouseLocation and then track -keyDown: events so the arrow keys can home in on the destination pixel. Yes, kids, that’s what we call a painful fricken’ hack.

Window Positioning

The next big headache was caused because you can’t set a window’s frame using non-integral points.

xScope does this a lot. The best example is with the Ruler tool: the origin of the ruler can be positioned to both even and odd pixels on screen. The red position indicators are also small windows that point to individual pixels in the ruler window.

The workaround is to make an NSWindow that’s larger than you need and then adjust the bounds origin of the NSViews it contains. The pain here is that it immediately introduces a dependency between windows and views. For example, the position indicator windows need to know if the view they’re hovering over has had its bounds origin shifted.

Pixel Alignment

There are many cases where xScope has to align to a pixel boundary. I found myself using this pattern many times throughout NSView -drawRect: code:

CGFloat backingScaleFactor = [self.window backingScaleFactor];
CGFloat pixelWidth = 1.0 / backingScaleFactor;
CGFloat pointOffset = lineWidth / 2.0;

The pixelWidth tells you how wide a single pixel is in points, while the pointOffset can be used to align a coordinate so that it straddles the Quartz drawing point:

NSBezierPath *line = [NSBezierPath bezierPath];
[line setLineWidth:pixelWidth];
[line moveToPoint:NSMakePoint(point.x + pointOffset, 0.0);
[line lineToPoint:NSMakePoint(point.x + pointOffset, 100.0);
[line stroke];

Another common pattern was to use the backingScaleFactor to align a coordinate to the nearest pixel in the view:

NSPoint point;
point.x = floor(x * backingScaleFactor) / backingScaleFactor;
point.y = floor(y * backingScaleFactor) / backingScaleFactor;

Of course you can do much the same thing with NSView’s -centerScanRect:, but in my experience it’s much more common to need aligned NSPoint values when you’re doing custom drawing. Creating an NSRect just to align the origin is a pain.

Flipped Coordinates

As Cocoa developers, we’re used to the pain and suffering caused by flipped view coordinates. Retina support can add a new dimension to this headache when you’re dealing with pixel coordinates.

Say you have a flipped NSView with an origin of 0,0 and dimensions of 1440 x 900 (e.g it covers the entire Retina screen in points.) Y coordinates on the screen can range in value from 0.0 to 899.5. When those coordinates are flipped, the range of values then become 0.5 to 900.0: which is off by a pixel (half point) if you’re trying to address the full range of view coordinates. The solution is to adjust the flipped coordinates like this:

NSRect windowRect = [self.window convertRectFromScreen:NSMakeRect(screenPoint.x, screenPoint.y, 0.0, 0.0)];
NSPoint viewPoint = [self convertPoint:windowRect.origin fromView:nil];
	
CGFloat backingScaleFactor = [self.window backingScaleFactor];
CGFloat pixelWidth = 1.0 / backingScaleFactor;
viewPoint.y -= pixelWidth;

I’ve always wondered why there’s this odd note in the +mouseLocation documentation:

Note: The y coordinate of the returned point will never be less than 1.

Even though it’s a lie, the extra pixel does make the coordinate flip work correctly—and now you have to take care because that pixel can be a fractional point.

Summary

Hopefully my trials and tribulations will help you in your own development. The good news is that the beautiful results on the Retina display make all this hard work worthwhile.

Debugging Core Data Objects

If you’re working on an app that uses Core Data, it’s inevitable that you’ll end up in the debugger and need to dig around in the object graph. You’ll also quickly realize that Core Data’s -description of an object isn’t terribly helpful:

(lldb) po myListObj
(List *) $19 = 0x08054ac0 <List: 0x8054ac0> (entity: List; id: 0x8074280 <x-coredata://29B10357-0723-4950-9EB6-E6D7AD6269B9/List/p175> ; data: <fault>)

Core Data’s documentation is excellent, but surprisingly doesn’t cover some of the tricks you can use to examine managed objects in the debugger.

The first trick is to fire a fault on the object using -willAccessValueForKey. After that, you can see what’s really there:

(lldb) po [myListObj willAccessValueForKey:nil]
(id) $20 = 0x08054ac0 <List: 0x8054ac0> (entity: List; id: 0x8074280 <x-coredata://29B10357-0723-4950-9EB6-E6D7AD6269B9/List/p175> ; data: {
    containerId = "8E4652D3-5516-4186-B1C9-DDBE41E108CF";
    createdAt = "2009-10-08 15:17:53 +0000";
    itemContainers = "<relationship fault: 0x8020850 'itemContainers'>";
})

You might also be surprised when you try to access one of the properties of the object:

(lldb) po myListObj.containerId
error: property 'containerId' not found on object of type 'List *'
error: 1 errors parsing expression

Remember that these properties are defined as @dynamic and there’s a lot of work done by Core Data at runtime to provide the implementation. The solution here is to use the KVC accessor -valueForKey: to get the object’s value:

(lldb) po [myListObj valueForKey:@"containerId"]
(id) $26 = 0x08069b60 8E4652D3-5516-4186-B1C9-DDBE41E108CF

Often, you’ll want to examine the relationships between objects. As you can see in the output above, the attribute itemContainers, which is a to-many relationship, is a fault. To fire the fault, get all the objects from the set:

(lldb) po [[myListObj valueForKey:@"itemContainers"] allObjects]
(id) $23 = 0x0d0667b0 <__NSArrayI 0xd0667b0>(
<Container: 0x807a180> (entity: Container; id: 0x801bc50 <x-coredata://29B10357-0723-4950-9EB6-E6D7AD6269B9/Container/p1> ; data: {
    containerId = TestContainer;
    items = "<relationship fault: 0x805b100 'items'>";
    state = "(...not nil..)";
    type = 4;
})
)

Finally, you may be using Transformable attribute types. The state attribute above is an example. If you’d like more information than (...not nil...), use the KVC getter and you’ll see that it’s an empty NSDictionary:

(lldb) po [[[[myListObj valueForKey:@"itemContainers"] allObjects] lastObject] valueForKey:@"state"]
(id) $29 = 0x0807dd30 {
}

(lldb) po [[[[[myListObj valueForKey:@"itemContainers"] allObjects] lastObject] valueForKey:@"state"] class]
(id) $30 = 0x01978e0c __NSCFDictionary

It’s taken me months to learn these tricks, so hopefully this short essay helps you come up to speed more quickly than I did!

ARC and copy

Like many of you, I’ve recently starting coming to terms with automatic reference counting (ARC) in Objective-C. For the most part, it’s gone remarkably smoothly. The only hard part is remembering to not type autorelease!

ARC lets us get rid of the retain/release pairs in our code. But can you spot the bug in the following code?

@interface MyObject : NSObject

@property (copy) id ivar;

@end


@implementation MyObject

@synthesize ivar = _ivar;

- (id)initWithIvar:(id)ivar
{
  self = [super init];
  if (self != nil) {
    _ivar = ivar;
  }
  return self;
}

@end

The ivar instance variable is a strong reference, not a copy. Since I think it’s a bad idea to use accessors during -init the copy semantics defined by the @property are never used and ARC happily retains the reference instead of copying it. Also of note: the static analyzer doesn’t think there’s anything wrong with the code above, so you won’t see any warnings that you’ve screwed up.

In my case, this initializer led to a bug where several worker threads were modifying the same instance variable and, well, you know how that goes. The fix, as usual, was painfully simple:

    _ivar = [ivar copy];

It’s important to remember that ARC primarily affects how we use -retain and -release. As Matt Drance points out the way -copy works hasn’t changed significantly. Just because we can forget about typing “retain” and “release”, doesn’t necessarily mean that we can forget to type “copy”, too.

Updated May 4th, 2012: A Radar about the static analyzer not issuing a warning has been filed: rdar://11386493

iTunes Manglement

I think we can all agree that iTunes is in need of a major overhaul. So why isn’t it happening?

Apple has shown no fear of rethinking and innovating with the user interfaces that manage our own personal data. Both the iMovie and iPhoto apps are great examples of this: they’ve gotten much simpler to use over the years (especially as they’ve moved to the iOS platform.)

This makes me think that there may be another factor that’s holding back iTunes; and I fear that it’s contractual.

Much of iTunes functionality is based around content that Apple or the user doesn’t own. And as we all know, the media companies that own the content are particularly paranoid about how digital assets are managed. In the 10+ years that iTunes has been in existence, I’m sure there’s a tangled web of legal obligations that makes improvements a huge technical headache.

To give you an idea of how painful this must be, imagine being a developer at Apple and having to consult this before implementing or improving a feature in iTunes. And when you’re done wrapping your head around those conditions, make sure you have thought about restrictions in other parts of the world. Having fun yet?

Core Data without Fetch Requests

If you follow me on Twitter, you’re probably aware of a really nasty problem I encountered with versioned Core Data models. Suffice it to say that was two days of hell caused by a single Fetch Request in a .xcdatamodeld file.

The irony of it all is that I no longer use Fetch Requests this way: the queries defined in the model file weren’t being used so I was happy to delete them all.

In my opinion, fetch requests that are stored as a part of the managed object model have some downsides:

  • You need to carry around an instance of NSManagedObjectModel in order to access the requests.
  • A text-based search (such as Find in Workspace from Xcode’s Search Navigator) doesn’t find the requests.

Instead, I use a category for NSManagedObjectContext that adds some nice functionality for creating fetches and executing them purely with code. The original idea and implementation comes from Matt Gallagher.

A picture’s worth a thousand words, so let’s look at how this works in code. Instead of creating an “existingPartner” fetch configured as:

Fetch all: Partner objects where:
Expression: name == $NAME

I write this code:

NSArray *objects = [managedObjectContext fetchObjectArrayForEntityName:@"Partner" withPredicateFormat:@"name == %@", name];

If I need the managed objects in sorted order, there’s this variant:

NSArray *sortDescriptors = [NSArray arrayWithObject:[NSSortDescriptor sortDescriptorWithKey:@"name" ascending:YES]];

NSArray *objects = [managedObjectContext fetchObjectArrayForEntityName:@"Partner" usingSortDescriptors:sortDescriptors withPredicateFormat:@"name LIKE %@", pattern];

Getting count entity objects is just as easy:

NSNumber *salaryLevel = [NSNumber numberWithInteger:1000000];
NSUInteger count = [managedObjectContext countForEntityName:@"Employee" withPredicateFormat:@"salary > %@", salaryLevel];

After reading about how NSExpression can be used to significantly speed up fetches against attribute values, I extended Matt’s NSManagedObjectContext+FetchAdditions category. This lets you do things like this:

NSNumber *marketingPayroll = [managedObjectContext fetchObjectArrayForEntityName:@"Employee" usingAttribute:@"salary" andFunction:@"sum:" withPredicateFormat:@"department == %@", marketing];

One thing to keep in mind when you’re using NSExpression against a SQLite store: it only works on data that has been persisted. Since the underlying implementation relies on SQL functions such as min(), max() and sum(), any managed objects that you have in memory (because the context hasn’t received a -save) won’t be included in the results.

One final trick I use is to keep the code for these fetches as close to the objects and attributes as possible. I’ve found it very handy to create class methods in the NSManagedObject subclasses. For example, using our partner name example from above, I’d implement the following class method for the Partner object:

+ (Partner *)fetchInManagedObjectContext:managedObjectContext withName:(NSString *)name
{
	Partner *result = nil;
	
	NSArray *results = [managedObjectContext fetchObjectArrayForEntityName:@"Partner" withPredicateFormat:@"name == %@", name];
	if ([results count] > 0) {
		result = [results lastObject];
	}
	
	return result;
}

Then when I need a Partner with the given name, I use:

Partner *partner = [Partner fetchInManagedObjectContext:managedObjectContext withName:@"CHOCK GLOBAL INDUSTRYS INK"];

I hope this category cleans up your code as much as it has mine. Feel free to use this code in any way you see fit.