Category Archives: iOS

F-Script Anywhere With LLDB

Ever since the start of my career at Apple, working with the venerable Macsbug, I have prided myself on making the most of whatever debugging facilities are at my disposal. I came to be quite capable with Macsbug, adding custom commands and data templates that helped me speed through crucial debugging sessions that would have otherwise taken much longer.

When I moved from the classic Mac OS team to Mac OS X, I was forced, only slightly earlier than every other Mac developer, to adapt to gdb. I did so with modest aplomb, adding custom commands that made it easy to, for example, continue until the next branch instruction (as it happens, also the first real post on the Red Sweater Blog).

When Apple started shifting away from gdb to lldb a few years ago, I realized I would have to throw out all my old tricks and start building new ones. To my great shame, progress on this front has been slower than I wished. The shame is made greater by the fact that lldb is so delightfully extensible, it practically begs for a nerd like me to go town adding every manner of finessing to suit my needs.

The nut of lldb’s extensibility is that much of its functionality is actually implemented in Python, and developers such as ourselves are invited to extend that functionality by providing Python modules of our own.

I finally decided to break the ice with lldb’s extensibility by adding a shortcut command for something I often want to do, but frequently put off because it’s too cumbersome: injecting F-Script into an arbitrary application running on my Mac. F-Script is a novel dynamic programming interface that lets you query the runtime of a Cocoa app using a custom language. It also features a handy tool for drilling down into an app’s view hierarchy and then navigating the various superviews and subviews, along with all their attributes. In some respects it’s very similar to a “web inspector,” only for native Objective-C applications on the Mac (and sadly, with far fewer features).

There are Automator workflows that aim to automate the process of injecting F-Script into a target app, by running the required commands, via gdb or lldb, to make the injection work seamlessly. For some reason, these workflows have never worked so seamlessly for me, so I’m always reduced to attaching to the process with lldb, and running the required commands manually to get the framework loaded.

Fortunately for me, “having to run lldb” is not such a big deal. Usually when I want to poke around at an app, I’m in lldb anyway, trying to break on a specific function or method, or examining the application’s windows and views via the command line. Once I’m attached to a process with lldb, getting F-Script to inject itself is as easy as running these two commands:

expr (void) [[NSBundle bundleWithPath:@"/Library/Frameworks/FScript.framework"] load]
expr (void) [FScriptMenuItem insertInMainMenu]

That’s all well and good but to do that I always have to find the memo I took with the specific commands, then copy and paste them individually into lldb. Far too often, I wind up imagining the struggle of this work and put it off until I’ve spent minutes if not hours doing things “the harder way” until I finally relent and load F-Script.

Today I decided that I need to stop manually copying and pasting these commands, and I need to finally learn the slightest bit about lldb’s Python-based script commands. The fruit of this effort is summarized in this GitHub gist. Follow the directions on that link, and you’ll be no further away than typing “fsa” in lldb from having all of its utility instantly injected into the attached app from lldb.

Even if you’re not interested in F-Script Anywhere, the gist is a concise example of what it takes to install a simple, Python-backed macro command into lldb. Enjoy!

Brent’s Coding Nits

I agree with every one of Brent Simmons’s coding nits, even if I don’t practice what he preaches completely. Everybody has to make some mistakes to keep life interesting, right?

There’s also one point he makes:

All of your class names should have a prefix. All of them.

Which I think is particularly applicable to shared code such as the kind he’s reviewing on GitHub when compiling this list of nits. It’s not nearly so important, and in fact not really that important at all, for a project’s private classes to have prefixes. For example if you make a new app called “Motorcycle Derby,” you should be forgiven for calling your app’s delegate class “AppDelegate.m” if that is what suits you. By this point in history, Apple takes virtual namespacing seriously enough that you’re unlikely to clash with an unprefixed name from them, and so long as everybody distributing shared library code follow’s Brent’s rule above, you won’t clash with them either.

Still, you can’t go wrong by adding prefixes to be extra sure that you won’t conflict with anybody.

Transplanting Constraints

Over the past few months I have become quite taken by Auto Layout, Apple’s powerful layout specification framework for Mac and iOS.

For the past few years I’ve heard both that Auto Layout is brilliant and that it has a frustrating learning curve. I can now attest that both of these are true.

One of the problems people have complained most about with respect to Auto Layout is the extent to which Xcode’s Interface Builder falls short in providing assistance specifying constraints. As many people have noticed, Apple is addressing these complaints slowly but surely. Xcode 5’s UI for adding constraints and debugging layout issues is dramatically superior to the functionality in Xcode 4.

Still, there is much room for improvement.

One frustrating behavior arises when one deigns to move a large number of views from one position in a view hierarchy to another. For example, the simple and common task of collecting a number of views and embedding them in a new superview. This task is so common that Apple provides a variety of helpful tools under Editor -> Embed In to streamline the task.

Here’s the big downer with respect to constraints: whenever you move a view from one superview to another, all of the constraints attached to the old superview, constraints that you may have laboriously fine-tuned over hours or days, simply disappear. Poof!

This isn’t such a big deal when your constraints happen to match what Interface Builder suggests for you. But even very simple interfaces may have a fairly large number of constraints. Consider this contrived example, in which three buttons are arranged to roughly share the width of a container view:

SimpleButtons 1

Nine constraints, and the removal or misconfiguration of any one will lead to incorrect layout in my app. Yet simply embedding the views in a custom view wipes them all out:

TestView2 xib 1

This problem is bad enough in the contrived scenario about, but in my much more complicated interfaces, a collection of views might comprise 50 or more customized constraints. Here’s a “simple” subsection of MarsEdit’s post editor side panel:


Having to piece those all together again just because I want to rearrange some views, well it makes me mad. And when I get mad? I get … innovative!

A Pattern For Transplanting Constraints

Thanks to recent changes in Interface Builder’s file-format for xib files, it’s more straight-forward than ever to hand-tune the contents of a xib file outside of Xcode. It should go without saying that in doing so, you take your fate into your hands, etc., etc. But if you’re anything like me, a little hand-editing in BBEdit is worth the risk if it saves hours of much more intricate hand-editing back in Interface Builder. You’ll save valuable time and also reduce the very real risk of missing some nuanced detail as you try to reimplement everything by hand.

So without further ado, here are steps you can follow to transplant a set of views in a xib file such that the constraints from the old view follow over to the new view:

  1. Make a backup of your .xib file. You’re going to screw this up at least once, so you’ll want something “sane” to fall back on when you do.
  2. In Interface Builder, create the parent view if it doesn’t exist already. Give it a real obvious name like “New Parent View” so you’ll be able to spot it later:

    NewParentView xib

  3. Save changes in IB to make sure the .xib file is up-to-date.
  4. Open the .xib file in a text editor such as BBEdit, or right-click the file in Xcode and select Edit As -> Source Code to edit as text right in Xcode.
  5. Locate the new parent view by searching on the name you gave it. For example, in my sample project the view looks like this in the text file:
    <customView ... id="5M5-9Q-zMt" userLabel="New Parent View">
  6. Locate the old parent view. If you have trouble, you may want to give it a custom name as well before saving again in IB. In my trivial example, the old parent is the first and only top-level view in the xib file, so it looks like this:
    <customView id="1">
  7. Take note of the id for the old parent view and the new parent view. We’re going to need these in a minute to tie up some loose ends.
  8. Locate the constraints from the old parent view, cut them, and paste them into the new parent view’s XML content. Again in my case it’s trivial because I want all the constraints from the old parent view. I cut them out of the old and into the new so things looks something like this:
    <customView ... id="5M5-9Q-zMt" userLabel="New Parent View">
                    <constraint firstItem="rfg-hN-1Il" firstAttribute="leading" secondItem="1" secondAttribute="leading" constant="20" symbolic="YES" id="LOu-nX-awU"/>
                    <constraint firstItem="8Ju-hM-RbA" firstAttribute="baseline" secondItem="Sgd-MR-FMw" secondAttribute="baseline" id="Mwc-6y-uaP"/>
  9. Locate the subviews themselves from the old parent view, and cut and paste them in the same way, making sure they reside in a <subviews> node in the new parent view. You should now have a new parent view whose xml topology looks something like this:
    <customView ... id="5M5-9Q-zMt" userLabel="New Parent View">
    	<rect ... />
    	<autoresizingMask ... />
    		... your transplanted subviews here ...
    		... your transplanted constraints here ...

    We’re close! But not quite finished. If you save and try to use the .xib now, you’ll find that Interface Builder rejects it as corrupted. What’s wrong? The constraints we transplanted mostly reference only the other views that we transplanted, but some of them also reference the old parent view.. To fix the integrity of these constraints, we need to update them to reference the new parent view instead.

  10. Refer back to the Interface Builder “id” values you noted in step 7. We need to locate any reference to the old parent view and adjust it so it references the new parent view. In our example, the old parent view id is “1” and the new parent view id is “5M5-9Q-zMt”. Specifically, we’re looking for attributes on our transplanted constraints where the “firstItem” or “secondItem” references the old parent ID:
    <constraint firstItem="rfg-hN-1Il" firstAttribute="leading" secondItem="1" secondAttribute="leading" constant="20" symbolic="YES" id="LOu-nX-awU"/>

    Change the value secondItem=”1″ to secondItem=”5M5-9Q-zMt”, and repeat for any other instances where the old parent view is referenced.

  11. Save the text-formatted .xib file, cross your fingers, and hope you didn’t make any mistakes.
  12. Reopen the .xib file in Interface Builder, or if you’re already in Xcode’s text editor, right-click the file and select Open As -> Interface Builder.

If your combination of luck and skill paid off as planned, then you’ll see something beautiful like this:

TestView xib

All of my views, now situated within the new parent view, and the desired constraints in-tact.

I hope this helps serve as a specific reference for folks in the same boat as I am in, wanting to shuffle views around without losing the hard work I’ve put into my constraints. And I hope it also serves to inspire you to think beyond the limitations of our tools. As great as Xcode, Interface Builder, and a host of other essential technologies are, they often fall short of desired behavior. When they do, it’s often in our power to work around the issues and carry on developing software as effectively as we know how.

Static Analysis

I am thus far primarily a Mac developer, though I have dipped my toes in the iOS development arena many times in the — sheesh! — 5 years since iOS 2.0 shipped with its developer-facing SDK.

My first, and only shipping app for iOS is Shush, a static noise generator that was inspired by my son Henry’s birth. He was born in August, 2008, months after iOS had been opened to the public. As you might imagine, I didn’t have a lot of spare time to play around with iOS programming, but I did have a screaming baby. For those of you who don’t know, static noise is famously soothing to small babies. Shush 1.0 was my bare-bones solution for dispensing infinite, soothing static noise from the magical device I could hold in my hands:

Shush1 0

Pretty hot, huh? It did the trick. I would hold crying, months-old Henry against my chest and, with the iPhone quietly shushing in my hand, he would drift off to sleep.

I mostly forgot about Shush after Henry got old enough to no longer benefit from it. Fast-forward 3 years and my second son, Matthew was about to be born. I realized I was going to need to dust off the old soothing machine, and it seemed like a great excuse to finally brush up the UI a little.

There isn’t much reason to look at the screen while using Shush: its primary purpose is generating audio noise. But despite Shush 1.0’s extremely minimalist design, I had always imagined the app was a prime candidate for a skeuomorophic design. In the old days of analog television, a common technique for generating this sound was simply to tune to an unused station. While the audio of the room filled with static white noise, the screen similarly rumbled with visual black-and-white “snow.” I thought it would be pretty cool to simulate this on the iPhone, as a throwback to those nostalgic days and so a Shush user would have something vaguely interesting to gaze at if they chose to.

It turns out, simulating the static television snow of my analog youth is extremely challenging to do, even on a fancy iPhone. Generating audio white noise is relatively easy: you can get close to the desired output by simply taking random numbers and feeding them to the audio system as samples. It seems reasonable to assume you could do the same for video. For each frame, you could simply generate a random grey between 0.0 and 1.0 for each pixel, rendering the result to an image:

This yields a pretty TV-snow-like image:

Static snow image

The problem is it’s incredibly CPU intensive to calculate that many random numbers and construct an image. Even testing this naive approach again today on my relatively speedy iPhone 5, the naive approach produces an animation where the frames only alternate every 3 seconds or so. Clearly, this would not do.

I experimented with a variety of approaches to speed up the rendering. What if I didn’t generate a wholly random number, but just alternated between 0 and 1? Also, do I really need to generate a random value for every pixel? What if I clump the pixels together to cover more ground? I tried a variety of techniques to speed up the view drawing:

The result was faster, but still not fast enough. What’s worse? It looked more like an homage to a Commodore 64 than to a vintage analog television set:

Another snow image

I was about ready to throw in the towel. Maybe this was simply not possible on an iPhone. I did some research on the web and it was not promising: not only is this a hard problem on an iPhone, it’s a hard problem everywhere. I learned that I’m not the first person who tried to generate an approximation of visual static on a digital computer, and that most people eventually resort to using a canned video animation of static. Modern television sets typically do this to give you the old-timey sense of “nothing is plugged in,” but if you look closely you can see repeating patterns in the static. In other words, it’s not really random. It’s not really static. If I couldn’t make this strange, skeuomorphic homage look more or less like real TV static, I was not interested in the challenge.

Up to this point all of my efforts had been at the level of the CPU: how can I fill this image buffer with random pixels faster? Having no experience with OpenGL or directly programming a GPU, I hadn’t even considered the possibility of approaching that. But a conversation with my friend Mike Ash put the idea in my head, and I ran with it. Since iOS devices are famously optimized for leveraging the GPU, I figured it might be a simple matter of asking the GPU to generate the random pixels for each frame on the fly, obviating the need for the generation of any CPU-bound image data.

I gave myself a crash course education in OpenGL, learning the bare minimum I needed to know before tackling the problem. To give you an idea of where I was starting, I had heard the term “shader” before, but honestly had no idea what it did. I eventually learned that probably what I needed was a bit of OpenGL magic called a “pixel shader.” Essentially it’s a chunk of code that runs on the video card and gets to choose what color each pixel in a given scene should be. For my scenario, I would be setting up OpenGL with a sheer surface pointed directly at the camera, so as to appear 2D. The shader’s job would be to fill that 2D surface up with random gray pixels.

Using Apple’s GLKViewController, I was able to skip over much of the hardcore OpenGL setup, and skip right to the work on the shaders. I used some boilerplate code to get my GLKViewController wired to my pixel shader, and was able to e.g. demonstrate my ability to fill the surface with a specific color:

It works! And while I was working out how to make OpenGL do my bidding, I put some work into the TV frame appearance for the app:

Skitched 20130927 153929

At this point, it feels like I’m almost home. I just need to swap out the constant RGB values for ones that insert random values of gray. What’s that you say? There’s no random generator in OpenGL? Well, I’ll be damned.

Once again things appeared hopeless. I played around with the addition of a “vertex shader,” which is a shader that has access to additional information about the scene. Using the fact that a vertex shader and pixel shader can communicate with each other, I was able to incorporate the specific coordinate for the pixel being shaded. Scouring the web, I found example code for OpenGL that would take a varying number like this and “fuzz” it sufficiently that it appeared to be somewhat random. Thus, my next effort involved taking the x and y coordinates for the current pixel and transforming them into a seemingly random shade of gray:


Oh my god! It’s beautiful. We’re done, this is exactly what I’ve been striving to do for weeks, now. Except… it doesn’t animate. It’s just a rendered scene of random grey pixels in which the random grey pixels are always exactly the same as before. Why? Because the inputs to the pseudo-random fuzz function are always the same: the coordinates of each pixel in the scene.

My final stroke of insight was to inject “just enough randomness” into the scene by hooking up a value that the pixel shader obtains from the client app. If I can supply random numbers to shader, you may ask, what’s the big deal? Why not just supply all the random numbers? Because the facility for injecting values into the shader only gives the client app access once per complete rendering. Once it starts rendering, the determination of values for each of the pixels in the scene is completely up to the shader itself. But by combining the pseudo-random generation based on pixel coordinate, and further fuzzing that value with a random value injected once per rendering, the results are as in the image above, but beautifully, quickly rendered (video capture doesn’t do it full justice).

Here is a final code snippet of both the vertex and fragment shaders. You can pop them into a project in Apple’s OpenGL Shader Builder to get a better feel for how they work.

On January 17, 2012, I released Shush 2.0. The next day, Matthew was born. It worked great for the few months that I needed it, and just as I did before, I have since mostly stopped using the app myself. However, it was a great exercise in pushing the limits of what the iPhone seemed capable of doing. Hopefully this experience will inspire you to look deeper for solutions to the problems that vex you while working with these fascinating, limited devices.