Why not write about some of the books I read in January? Honestly, thinking about writing a post about some of my reading prompted me to read more, so it's a good thing and I'm planning to keep up the habit.
“Functional Swift” by Chris Eidhof, Florian Kugler and Wouter Swiersta of objc.io
I believe it was called “Functional programming in Swift” recently, but when the book was updated in December 2015 for Swift 2.1 (which I read) it got renamed as well, and for the better.
This one read a lot like the Apple's official Swift book for me: clear, concise, to the point and at the same time introductory. The book consists of several sections exploring different topics e.g. writing a wrapper for Core Image or implementing a test suite which generates tests automatically, all of them revolving around the functional approach to solving problems.
If you're interested in the actual functional programming, reading “Learn you a Haskell” would be more beneficial, but if you're curious how to think differently while writing in Swift or learn some new approaches to common problems, it's a great introductory book that you can later supplement with more reading on the topic, having the primaries down.
“Turning Pro” by Steven Pressfield
This is a small but intense book intended for everyone who is (or should be) hard at work. I read it last time three years ago, worked a lot since then, and this time it stood out even brighter than before. It's a great reminder why you would jump out of bed each morning and work hard. A motivational book, so to say, but not in a happy way — a sobering one: the work is hard and it's you who has to do it, no exceptions.
Here's my slightly longer review on Goodreads if you're interested.
The final two books are (surprise!) about financial markets and specifically, high-frequency trading. Why read them over crunchy programming books? December saw the launch of Stockfighter, a game where you basically write a trading bot to try and game the simulated financial market. Yes, this is supposed to be someone's idea of having a good time, and incidentally I had a good time designing and (partly) writing my own library for it. Anyway, on to the books.
“Flash Boys” by Michael Lewis
Lewis is a good writer and writes a good story, but when it comes to the technical details… not so much. Some glaring holes right there, and far more than once I was close to making a facepalm.
Still, it's probably the best book to start with on this topic as it paints the picture in large strokes and gives you a good idea of the emotions behind what people were doing. You can read any number of technical books on the subject, there are many, but first you have to dive in and look around, and this is a really good introduction that starts slow and holds you by the hand all the way in.
Yes, some claims are outrageous and plain wrong, but it's entertaining and Lewis gradually explains most of the buzzwords you're likely to encounter in other books about the subject.
Don't confuse it with “Flash Boys: Not So Fast” by Peter Kovac which is sort of a bashing reply to Lewis. Kovac too makes quite a lot of mistakes, judging by the reviews, so I haven't read it yet.
“Dark Pools” by Scott Patterson
This is another “introductory”, or rather, non-technical book, which looks at the same process described in Lewis' “Flash Boys”, but from a different angle. Where Lewis mostly follows the story of a banker playing catch-up to high-frequency trading firms, Patterson describes the lives of several prominent people who directly influenced the rise of HFT from the inside. Small shops which eventually grew to enormous size and changed the market altogether.
Arguably, “Dark Pools” is far more technical of the two and I really enjoyed reading it. It's more technical, but it's still a story, so it's perfectly readable and just as entertaining: Patterson goes into a lot of detail on how the markets worked since the early 90s up to the current moment. It also turned out to be more applicable to Stockfighter: quite a few of the actual techniques he describes used by high-frequency traders I can see used in the game verbatim.
Reading about HFT is surely nice, but I decided that I should stop reading and beat a few levels of the game first :)
After reading “Profiling your Swift compilation times” I decided to try it for our project, as it's about 80% Swift and a couple of Objective-C frameworks. Frankly, dry compile time (after cleaning) is just really bad, pushing 3 minutes. Most of the time compilation is incremental and takes much less time, but the project has to be cleaned several times a day anyway, and the compile times add up.
If the linked post ever goes away, you can profile your project simply by adding the following to your Other Swift Flags in Build Settings for the intended target:
After that you do the following to get the output file culprits.txt with sorted function compile times:
xcodebuild -workspace Project.xcworkspace -scheme Project clean build | grep [1-9].[0-9]ms | sort -nr > culprits.txt
Here's an edited excerpt from our project before I started optimizing:
1172.0ms Views/BeaconMeasurementDataTableViewCell.swift:17:10 func configure(beacon: FilteredBeaconData) 1053.3ms Controllers/ClientViewController.swift:855:20 @IBAction @objc func canvasViewTapped(sender: UITapGestureRecognizer) 994.2ms Helpers/CLLocationDistance.swift:35:17 public func region() -> String? 978.2ms Helpers/CLLocationDistance.swift:35:17 public func region() -> String? 928.3ms Controllers/ClientViewController.swift:855:20 @IBAction @objc func canvasViewTapped(sender: UITapGestureRecognizer) 782.2ms Helpers/CGRectSquare.swift:24:10 func canBeInsetBy(x: CGFloat, y: CGFloat) -> Bool
As you can see, the worst function takes just over a second, and some functions get repeated, so you can edit them once for multiplied benefit. In total, there were about 70 lines with functions taking longer than 100 ms to compile. Here's the before:
$ time xcodebuild -workspace Project.xcworkspace -scheme Project clean build real 15m41.535s user 24m43.824s sys 2m59.478s
I decided to not spend much time on this as we had a different situation than the linked post's author and were not likely to get much improvement. So I edited about 10 worst functions (and found a couple of missed bugs in the process, which is an added benefit). Then I ran the test again:
real 13m7.661s user 22m24.623s sys 2m46.180s
Not much, but that is still a measurable difference: 16.3% for real, 9.4% for user and 7.4% for sys, so about a 10% improvement in total. Not bad for an hour and a half of work.
I also tried comparing the difference in compile time from Xcode when running the app on the simulator, but couldn't come to a definite conclusion: even after cleaning, Xcode compile times vary a lot and the total time stayed around the same 3 minutes. Of course, it's much faster when the libraries and some modules are already cached, usually the build takes no more than 10-15 seconds.
I encourage you to test this for your Swift project if you're disappointed with compile times, it may well be a quick and big win.
Behind the scenes it is using swift build — the build option available only in the open-source toolchain so far. The errors are the same that you can find in Xcode, and Syntastic makes them available with line marks and both command window and location list. Though I have an option set which prevents the location list from opening automatically, it can only close if there are no more errors:
let g:syntastic_auto_loc_list = 2
Sadly there is no autocompletion yet. There is decent support via SourceKitten but it's only working for Xcode projects. I only have Package.swift :)
I wrote a UISplitViewController category for our app USpace which makes it behave in a slightly different way: it collapses to the right and has a wide “master” portion. When expanded on iPad, it looks like this:
I chose UISplitViewController over a custom solution because the app is universal and has to support everything from the same codebase: iPhone, iPad and iPad split-screen mode, and switch and rotate seamlessly. Another constraint is that it has to be rotation-locked on iPhone (i.e. the app should always be in portrait mode) but at the same time show a specific controller in landscape mode, which prevents us from using the Info.plist interface orientation restrictions.
Most of the custom solutions on GitHub that implement a slide-out horizontal menu do it in a very primitive way, by having a parent controller that slides two views around. USpace view hierarchy is much more variable and complicated to use such a solution, and on top of that, most of them only support iPhone. Besides, I'm a big proponent of rolling your own solution first, to test the water and figure out what features you would actually need. More often than not, you quickly end up with a minimal solution that a) fulfills all your needs b) works perfectly c) is easy to maintain.
We move pretty fast, and my first task was just to create a slide-out controller not unlike one you see on iPad when you swipe from the right side of the screen. It's called Slide Over if I'm not mistaken. So I just presented a full-screen overlay with a thin controller on the right, and with a bit of work it was working both on iPad and iPhone and was customized to our needs. We quickly grew out of it, setting up the expected view hierarchy was finicky, so I decided to switch the whole app to a UISplitViewController. It took a few days to iron out all the problems (did you know you could put a navigation controller inside another navigation controller in a specific way and have two working navigation bars on top of each other?) but finally it worked, and worked beautifully.
Today I was trying to fix one of the last bugs with this approach, and it was… strange. If you hold the iPhone horizontally (remember that the app is rotation-locked, so everything is sideways) and try opening that one controller which should be displayed in landscape mode, the navigation controller you're pops your current visible controller and only then presents the landscape-controller. In code it's just a simple presentViewController(_:). Here's a sample lldb output:
(lldb) thread backtrace * thread #1: tid = 0xb08ed2, 0x00000001859e2008 UIKit`-[UINavigationController popViewControllerAnimated:], queue = 'com.apple.main-thread', activity = 'send gesture actions', stop reason = breakpoint 1.1 * frame #0: 0x00000001859e2008 UIKit`-[UINavigationController popViewControllerAnimated:] frame #1: 0x0000000185bce11c UIKit`-[UINavigationController separateSecondaryViewControllerForSplitViewController:] + 148 frame #2: 0x0000000185e48104 UIKit`-[UISplitViewController _separateSecondaryViewControllerFromPrimaryViewController:] + 360 … frame #27: 0x000000018592dcec UIKit`-[UIViewController presentViewController:animated:completion:] + 184
I quickly figured out that when trying to present the new controller the app window rotated to landscape orientation and it prompted the wrapper UISplitViewController to split its “secondary” view controller — without asking me, apparently! I made a dive into the docs, and lo and behold, right there on the UISplitViewControllerDelegate page it said:
When you return nil from this method [ splitViewController(_:separateSecondaryViewControllerFromPrimaryViewController:) ], the split view controller calls the primary view controller’s separateSecondaryViewControllerForSplitViewController: method, giving it a chance to designate an appropriate secondary view controller. Most view controllers do nothing by default but the UINavigationController class responds by popping and returning the view controller from the top of its navigation stack.
We're very interested in the last sentence, because that is exactly what was happening. This only happened on iPhone in landscape layout, because on iPad it didn't have to split this way, or rather, it did, but it had no effect because the view hierarchy was different when it was allowed.
Still, that was not the end. I returned nil most of the time from the method mentioned in the docs, but this time I had to return a view controller if I didn't want the navigation controller to pop. I could also subclass the navigation controller (but it's not future-proof and cumbersome) or prevent it from popping in the navigation controller delegate, but it would be far from the place where the action originated, and the split view controller delegate approach allowed me to keep control in the same file, so there you go.
What I ultimately did is returned a “dummy” view controller from the “splitting” method and and did a return true for this case in the “collapsing” method. What return true means is that you tell the system that you have done what was necessary and you need no further action on its part. Basically it gives you a view controller and says: “Hey, I want to collapse this!” and you reply “Don't worry, I did everything myself!” while doing absolutely nothing. This way the “collapsing” view controller is just dumped and you never see it again. Again, if you let the system do its job, it will push this controller on top of your navigation controller in the “master” controller.
All in all, it was a fun bug to resolve. Hopefully my explanation maybe helps someone out there :)
That's how many it takes for Xcode (with XVim installed) to become slow on a file.
As you may have guessed, that is a symptom of a Massive View Controller which is begging to be refactored :) When I started that controller just a few months ago, it was a really thin wrapper about a hundred lines long.
There comes a time during the life of your Swift project when the compile times just don't cut it. There are some tweaks and knob twirling you can do to improve them, but honestly, it doesn't make that much of a difference.
I started looking around and found the deck “Swift compile time is so slow” by Masato Oshima where he claims that concatenating all the files together improved his build time sevenfold, from 6 minutes 55 seconds to a bit under a minute. “Impressive!” I thought and decided to test this approach for our project.
Now, concatenating all files to help the compiler out is not a new method, I know for certain that some game developers use it for C++ and the results are no less staggering, so this is a legitimate thing to try.
Let's get to the test results. At the time of writing our project is about 40k lines of Swift code and a dozen of external frameworks. I used the slightly modified concatenate_swift_files.sh from keychain-swift, their version disposed of the newlines we used in some of our strings so I had to edit the script to handle that. Otherwise, no changes.
The time is from tapping on Run in Xcode to the Simulator appearing.
Concatenated files: 2 minutes 8 seconds.
Normal source files (around 300 of them): 2 minutes 32 seconds.
Result: a measly 16% improvement. Not worth pursuing for our project. On top of that, the logger we use (XCGLogger) showed all debug prints originating in the concatenated file (versus a specific source file), which is logical and expected, but annoying and unhelpful.