Much and more has been written about Apple’s new programming language, Swift. I even did a talk about it at Yow: Connected 2014. If you saw me there, you’d know that I’m not the biggest fan of Swift. I think the rollout of it has been an absolute disaster: the compiler is a shambles, the language design is inconsistent in places1 and it’s stupidly difficult to get working in libraries, particularly if you want to support iOS 7. Surely though, at the very least its performance would be all right. After all, it’s designed to be more static and do more checking at compile-time and is meant to be made with the LLVM optimiser in mind…

Let’s talk about performance in Swift, then.

Like many people I was lulled into a false sense of security about Swift’s performance by a combination of factors:

• This post. It shows Swift as approaching the performance of C when it comes to sorting arrays. Very impressive. Surely the rest of Swift is this fast too, right?
• Apple’s own marketing about it, which claims an impressive “2.6x” faster than Objective-C and “8.4x” faster than Python 2.7.
• The name. Come on, they called it Swift. They wouldn’t call it that if it was slow.

I got about 1,000 lines into my project before thinking to myself that I should really performance test what I’ve written on the off-chance it wasn’t fast enough. At the time, the project was entirely Swift, and it made use of the pretty nice JSONHelper library to make the deserialisers look nicer. An example of what my Swift-like solution looked like:

So far, so nice.

At this point I devised a test, which would parse about 500kb of JSON, which is pretty far into “worst-case scenario” territory for our app, but isn’t outside the realm of possibility. It’s indicative JSON, very much like something the API would really return. You can find a copy of the sample JSON here. It shows a user who has a membership in 1,000 conversations. The model parser should read this JSON, create Membership objects which point directly to the corresponding instance of User and Convo, and create (or update) a heap of Convo objects based on the contents of the convos key. In the the Convos, the creator key should point to the corresponding instance of User. Nothing here’s unusual for a relatively sophisticated production iOS app.

I wrote a performance test using XCTest’s new performance testing feature. The test itself looked like this:

Making sure I had -O turned on in the build settings, I pulled out my iPod Touch 5th generation, which is realistically the slowest device we’ll be supporting, since it runs iOS 8 and has the same dual-core A5 as the iPhone 4S (which I don’t have). This made for a reasonable worst-case scenario: huge JSON payload, slow device.

Want to know how Swift did?

Guess.

Nope, higher. Try again.

Nope, higher, try again!

Oh fine I’ll just tell you! It took 1.42 seconds. One-point-four-two _seconds_. That can’t be right! Horrified by this, I decided to quickly whip up an Objective-C version of the same performance test. It was just the bare minimum code to implement the test, written in a distinctly Objective-C style, just like the Swift code was written in a distinctly Swift style.

An example of what the Objective-C looked like:

There were also implementations for CSMembership and CSConvo among others, and superclasses too. It was a pretty representative parser. So I whipped out my trusty XCTest case and created a new performance test which used the Objective-C classes instead, again making sure that -Os was enabled in the build settings.

You’ll never guess. This version runs on my device in just 0.09 seconds. Which makes the Objective-C version about 15x faster than the Swift version. Remember, the LLVM optimiser is turned on for both of these cases.

Ridiculous! I’d wasted most of my weekend at this point, so I figured why not waste some more. I wasn’t sure if Swift was slow or if it was just JSONHelper that was slow. The only way to rule it out for sure would be to write a version of the Swift code that is basically a line-for-line translation of the Objective-C so I could do a true Apples-to-Apples test. This is absolutely not how you’d write Swift, since it was full of NSDictionary references everywhere instead of Swift dictionaries, to name an example.

An example of what this looked like:

Now, one caveat about the results here. I simply couldn’t get the Swift to run with -O turned on without segfaulting. So, I figured to be fair, I’d turn off the optimiser for Objective-C this time around too. Here’s the result:

• Objective-C: 0.06 seconds
• Objective-C-like Swift: 0.29 seconds

I have no idea why the Objective-C runs faster with the optimiser turned off, but OK, sure. This proves that Swift is only acceptably fast if you basically write it exactly like you’d write Objective-C, but if you do, you’ll probably get segfaults that can’t be fixed, and it’ll still be about 5x slower anyway.

For one last test just for fun, I decided to rewrite the Objective-C test in Ruby by using RubyMotion. If you haven’t heard of RubyMotion, it allows you to write iOS and Android apps using Ruby, which compiles down into the same native code that you’d get when compiling Swift or Objective-C. It’s meant to be pretty fast, but I’ve never really benchmarked it with real production-like code before. I’ve always suspected it was significantly slower than Objective-C since it can’t do the crazy objc_msgSend tricks that the Objective-C compiler can do thanks to it doing all sorts of weird static analysis on method dispatches to see if it can be sped up. This is just because Ruby is a really dynamic language. Anyway…

Here’s an example of what that Ruby looked like:

RubyMotion doesn’t seem to have any options for the optimiser. It seems to just default to having optmisation on if you build for device, which would be in keeping with the convention-over-configuration mentality.

Here are the final results:

Style Optimisation Time
Objective-C -Os 0.09
Objective-C-like Swift -Onone 0.29
Swift -O 1.42
RubyMotion ??? 0.21

So RubyMotion is faster than Swift, then? The same Ruby, a language notorious in the programmer community for being quite slow2 is faster than a language called Swift which is marketed as being a fast language?

If your language is slower than a language that isn’t even trying to be fast, I’m sorry, but you’re just too far up shit creek. At this point, I wave goodbye to Swift. The Objective-C models I’ve written make for a pretty good starting point, so for now I’m just going to build that out.

I would love to provide all the source code used in this post, but unfortunately our product isn’t open source, and it will in fact be our production code someday. Feel free to leave a comment if you want to know more about what the code did, and I’ll tell you what I can.

UPDATE: I should point out that all these tests were run on Xcode 6.1.1 (6A2006), with whatever Swift compiler comes with that. RubyMotion was at 2.38.

1. Like subscripting, which uses get and set syntax just like a property, but they are declared with a function style syntax? There’s no good reason for this that I can see…

2. This isn’t a criticism. I understand that it’s just an extension of the idea that computer time is less valuable than human time, and Ruby lets you write programs faster to save human time. I appreciate, and even respect that world-view.