Rants from the Ballmer Peak rss feed

Kerf timestamps done almost right: Swift

In the first chapter of the series we reached the conclusion that to implement Kerf's timestamp types we need the following features from a programming language:

  1. Value type semantics with strong typing to avoid mistakes.
  2. Instancing types on the stack to avoid slow heap memory allocations and alleviate manual memory handling or garbage collector pressure.
  3. Custom literals for easier construction of such types.
  4. Operator overloading to implement all possible custom operations.
  5. Generics are not necessary but help with implementation.

This is a really long article (Buy Swift Programming!) which has been split in different chapters because it is (Swift Programming on sale!) unsuitable for today's average attention span and lets me maximize (Get Swift Programming now!) page ads.


  1. a new type?
  2. Nim
  3. C++
  4. Swift You are here!
  5. WTF… Java?
  6. conclusions

Swift is a multi-paradigm, compiled programming language created for iOS, OS X, watchOS, tvOS and Linux development by Apple Inc. Swift is designed to work with Apple's Cocoa and Cocoa Touch frameworks and the large body of existing Objective-C code written for Apple products. Swift is intended to be more resilient to erroneous code ("safer") than Objective-C and also more concise. It is built with the LLVM compiler framework included in Xcode 6 and later and uses the Objective-C runtime, which allows C, Objective-C, C++ and Swift code to run within a single program. Running our requirement list against Swift's feature set we get:

  1. Swift has value type semantics with strong typing to avoid mistakes (yay!).
  2. Allows instancing types on the stack through custom structs (yay!).
  3. There are no custom literals but we can write literal extensions to fake them (ok, I hope).
  4. Allows operator overloading (yay!).
  5. Supports generics, but they seem daunting to use due to the perceived type complexity (hmmmm). Luckily this is not a requirement.

The key concept to understand Swift is that it is bogged down by its mandatory inheritance. Swift has a very specific task: replace Objective-C. And with such a requirement the first thing you have to do is being able to interface at binary level with Objective-C's libraries. This involves interfacing with code that still uses reference counting below, either manually or through automatic reference counting (ARC), which can be a landmine by itself. So if you thought that either C++ or Objective-C having C compatibility was already a handicap, imagine having to interact with all those together and still present the façade of an unencumbered language. That may be the reason why the compiler is slow. But let's not get ahead of ourselves, lets implement first our friendly Nano and Stamp structures to appease the Kerf gods.

Faking new types one struct at a time

Trust me, I'm not a banana, I'm a new timestamp type

Since Swift doesn't have any language support for distinct types we will have to revert to writing piles of code just like for the C++ implementation. And copying C++'s implementation we will create two structs, let's start with the Nano one:

let u_nano = Nano(1)
let u_second = Nano(1_000_000_000)
let u_minute = u_second * 60
let u_hour = u_minute * 60
let u_day = 24 * u_hour
let u_month = 30 * u_day
let u_year = u_day * 365

struct Nano : CustomStringConvertible {
        var value: Int64 = 0
        init(_ x: Int64) { self.value = x }
        init(_ x: Stamp) { self.value = x.value }
        var description: String {
        var s: String { return description }
        // …more code goes here…

Unlike C++ or Nim, the first thing that catches the eye is that you can actually use the Nano(1) initializer despite the Nano class not being known to the compiler on that very first line. Maybe the designers of Swift thought that forward declarations are bad for human programmers and decided to get rid of them. Masses of careless programmers rejoiced, but there is a cost to pay: non deterministic compilation times. I'm not claiming that the compiler throws up a dice and decides compilation will take longer on even days, but now the compiler has to plow forward and keep code in a temporal maybe it compiles/maybe it doesn't Schrödinger state because some lines later may make the previous code compile. Java eliminated the header vs implementation duplication problem ages ago without requiring extra work for compilers. But in Swift the compiler is required to juggle multiple potential parallel compilation universes due to language design. Nice, extra gratuitous complexity for very low end user benefit. I'm so glad I don't have to implement compilers. Another case of non forward declaration is the secondary init(_ x: Stamp) initializer. This constructor converts the value of a Stamp to a Nano despite the Stamp type not existing yet.

The ramifications of the non forwardness of declarations can also be seen in the lack of any import or include lines. Our Nano structure inherits from the CustomStringConvertible protocol. Where does this protocol come from? Who knows, the compiler is doing magic to include or know about this protocol beforehand. So again, does the compiler actually scan and parse all known protocols in its standard library for every simple compilation unit? That would be crazy, as in batshit crazy, but could explain the slow compile times. Just so you know, the CustomStringConvertible protocol is required to make our Nano type valid input to other code expecting objects conforming to this protocol, like print(). This protocol defines a description pseudo variable. This feels arbitrary to me, you can define a variable which works as a function, and the only difference between a normal function is that you are not using parentheses to invoke it. But as a user of the code you have to know which is which or the compiler will give you an error. There are other ways of making a type conform to a protocol, like using an extension, we will look at extensions in a moment.

Since the description pseudo variable of the CustomStringConvertible protocol is too long, I decided to write an alias as the s pseudo variable returning whatever description does. That custom s variable is used in the following operator functions:

func *(lhs: Nano, rhs: Int) -> Nano { return Nano(lhs.value * Int64(rhs)) }
func *(lhs: Int, rhs: Nano) -> Nano { return Nano(Int64(lhs) * rhs.value) }
func +(lhs: Nano, rhs: String) -> String { return lhs.s + rhs }
func +(lhs: String, rhs: Nano) -> String { return lhs + rhs.s }
func +(lhs: Nano, rhs: Nano) -> Nano { return Nano(lhs.value + rhs.value) }
func -(lhs: Nano, rhs: Nano) -> Nano { return Nano(lhs.value - rhs.value) }
func -(lhs: Nano, rhs: Int64) -> Nano { return Nano(lhs.value - rhs) }
func %(lhs: Nano, rhs: Nano) -> Int64 { return lhs.value % rhs.value }
func %(lhs: Int64, rhs: Nano) -> Int64 { return lhs % rhs.value }
func /(lhs: Nano, rhs: Nano) -> Int64 { return lhs.value / rhs.value }
func /(lhs: Int64, rhs: Nano) -> Int64 { return lhs / rhs.value }

Of interest is the overloading of the addition operator for string concatenation, which seems quite normal in Swift land, and it's where I'm using the s variable. In the Nim implementation chapter I mentioned that it is better if string concatenation is done using an operator other than addition. If you use the same, you can end up writing code whose intent is not clear. Consider the following lines of potential code:

let normal = Nano(1)
let sneaky = normal.s
print("Values \(sneaky + normal)")

The first line defines our Nano variable, the second converts it to a string representation. Due to type inference we might miss this bit (cue all the pedants changing their obnoxious style guides to force everybody explicitly state types everywhere), and the third line prints the values using string interpolation. What is the expected output? The cat is out of the bag, and I have already mentioned that sneaky is a string representation. So we will get 1ns1ns printed, which is two nanoseconds joined together as strings. But maybe the intent was to add numerically those two values? If Swift didn't use the addition operator for string concatenation this would have not compiled, the compiler would have told that you can't add a Nano to a String. Not a serious issue you say? Right, tell that to those who write numerical crunching code. Or me, because I made this mistake myself when I was trying to write Nim code which looked like Swift using the addition operator (the right side to that equal sign was being concatenated as strings due to the associativity of the addition operator). In Swift you have to add defensive parentheses to make sure you don't make such mistakes or maybe play with the operator priority rules.

In the introduction I said that you can't define custom user literals like in C++. In Swift we can fake conversions in a similar way to Nim using literal extensions:

extension Int {
        var ns: Nano { return Nano(Int64(self)) }
        var s: Nano { return self * u_second }
        var i: Nano { return self * u_minute }
        var h: Nano { return self * u_hour }
        var d: Nano { return self * u_day }
        var m: Nano { return self * u_month }
        var y: Nano { return self * u_year }

The extension tells the compiler that all Int types have suddenly new variables called ns, s, i, etc which return Nano types. With this extension we can finally write a thoroughly verbose test case similar to our previous implementations:

print("Testing second operations:\n")
print("\(Nano(500)) = \(500.ns)")
print(Nano(500) + " = " + 500.ns)
print(u_second + " = " + 1.s)
// Uncomment this line to make the swift 2.1.1 compiler cry like a child.
//print(u_minute + u_second + Nano(500) + " = " + 1.i + 1.s + 500.ns)
print("\(u_minute + u_second + Nano(500)) = \(1.i + 1.s + 500.ns)")
print((u_minute + u_second + Nano(500)) + " = " + (1.i + 1.s + 500.ns))
print("\(1.h + 23.i + 45.s) = \(composed_difference) = \(composed_string)")
print("\(u_day) = \(1.d)")
print("\(u_year) = \(1.y)")
print("\(u_year - 1.d)")

let a = composed_difference + 3.y + 6.m + 4.d + 12_987.ns
print("total \(a)")
print("\tyear \(a.year)")
print("\tmonth \(a.month)")
print("\tday \(a.day)")
print("\thour \(a.hour)")
print("\tminute \(a.minute)")
print("\tsecond \(a.second)")

As you can see this is the usual test we have been repeating so far, with the expected output. The syntax is pretty much like Nim's, only harder to read due to Swift's awkward string interpolation which adds noise in the form of extra parentheses and backslashes. But, you may have noticed that comment right there, the one about crying, what the hell is that?

Slowness intermission

The Swift compiler is not slow, it is just allowing you to exercise the virtue of patience. Let's compare the speeds of the Swift 2.1.1 compiler against the Nim 0.13.0 compiler:

$ time swiftc -o units.exe *.swift

real	0m25.137s
user	0m24.235s
sys	0m0.863s

$ time nim c -o:units.exe units
Hint: system [Processing]
Hint: units [Processing]
Hint: time_nanos [Processing]
Hint: time_stamp [Processing]
Hint: strutils [Processing]
Hint: parseutils [Processing]
Hint: sequtils [Processing]
Users/gradha/project/kerf_timestamps_done_almost_right/nim/units.nim(18, 17) Warning: mapIt is deprecated [Deprecated]
Users/gradha/project/kerf_timestamps_done_almost_right/nim/units.nim(18, 16) Warning: mapIt is deprecated [Deprecated]
Users/gradha/project/kerf_timestamps_done_almost_right/nim/units.nim(19, 22) Warning: mapIt is deprecated [Deprecated]
Users/gradha/project/kerf_timestamps_done_almost_right/nim/units.nim(19, 21) Warning: mapIt is deprecated [Deprecated]
Users/gradha/project/kerf_timestamps_done_almost_right/nim/units.nim(24, 8) Warning: mapIt is deprecated [Deprecated]
Users/gradha/project/kerf_timestamps_done_almost_right/nim/units.nim(24, 7) Warning: mapIt is deprecated [Deprecated]
Users/gradha/project/kerf_timestamps_done_almost_right/nim/units.nim(25, 8) Warning: mapIt is deprecated [Deprecated]
Users/gradha/project/kerf_timestamps_done_almost_right/nim/units.nim(25, 7) Warning: mapIt is deprecated [Deprecated]
CC: units
CC: stdlib_system
CC: time_nanos
CC: time_stamp
CC: stdlib_strutils
CC: stdlib_parseutils
CC: stdlib_sequtils
Hint:  [Link]
Hint: operation successful (13343 lines compiled; 0.374 sec total; 20.204MB; Debug Build) [SuccessX]

real	0m0.384s
user	0m0.445s
sys	0m0.093s

Yep, that's right. For a hopefully equivalent implementation (the Swift code is 443 lines long, the Nim version 411 lines long) the Nim compiler takes about half a second to compile and generate a binary, while the Swift compiler sends my source code to the NSA for inspection through a slow hybrid goat/pigeon link in Afghanistan, which takes about 25 seconds, or 50 fucking times more than the Nim compiler. These are times from what we could consider cold boot, because the Nim compiler actually halves the compilation time I quoted if it is allowed to reuse the nimcache directory from a previous compilation. On the other hand running the Swift compiler several times only makes me consistently more impatient.

But wait, there's more! We haven't yet uncommented the deadly line of umpossible compilation. With this line in place, here is the result:

$ time swiftc -o units.exe *.swift
time_nanos.swift:136:2: error: expression was too complex to be solved in reasonable time; consider breaking up the expression into distinct sub-expressions
        print(u_minute + u_second + Nano(500) + " = " + 1.i + 1.s + 500.ns)

real	0m24.637s
user	0m23.967s
sys	0m0.612s

Wow, that's a really complex expression. Or maybe not? Let's put into perspective now the things I complained about earlier:

  1. No forward declarations.
  2. No need to import protocols, they are magic!
  3. Extremely complex type hierarchies.
  4. Use of the addition operator for String concatenation, but also for numeric operations!

Now these things start to add up and the compiler is actually having trouble with all those parallel Schrödinger universes where an expression could mean this, or could mean that, or maybe if we compiled a few lines more could mean something else entirely because a chained sub expression changes its output type depending on maybe-even-a-few-lines-more down the file…! I understand your pain, Swift compiler. And if you tell me that this is not a fault of language design, does that mean that the people writing the Swift compiler are morons? Ok, ok, that's too harsh, let's not make ad hominem attacks. Also, Swift is still a language in its infancy, with a shape shifting compiler. I was testing version 2.1.1, what would happen with newer releases?

$ swiftc -v
Apple Swift version 2.2-dev (LLVM 846c513aa9, Clang 71eca7da8e, Swift 96628e41cc)
Target: x86_64-apple-macosx10.9
$ time swiftc -o units.exe *.swift
time_nanos.swift:136:2: error: expression was too complex to be solved in reasonable time; consider breaking up the expression into distinct sub-expressions
        print(u_minute + u_second + Nano(500) + " = " + 1.i + 1.s + 500.ns)
time_nanos.swift:138:2: error: expression was too complex to be solved in reasonable time; consider breaking up the expression into distinct sub-expressions
        print((u_minute + u_second + Nano(500)) + " = " + (1.i + 1.s + 500.ns))

real	0m29.855s
user	0m28.892s
sys	0m0.670s

Oh, right, I forgot to comment out that deadly complex expression:

$ git checkout time_nanos.swift
$ time swiftc -o units.exe *.swift
time_nanos.swift:138:2: error: expression was too complex to be solved in reasonable time; consider breaking up the expression into distinct sub-expressions
        print((u_minute + u_second + Nano(500)) + " = " + (1.i + 1.s + 500.ns))

real	0m26.188s
user	0m25.619s
sys	0m0.546s

Wow, that's pretty awesome, the 2.2-dev version is going backwards and making previously easy to compile expression umpossible complex now. Wait, let's not give up here, we are so close to success I can smell it, let's try the latest and greatest:

$ swiftc -v
Apple Swift version 3.0-dev (LLVM b361b0fc05, Clang 11493b0f62, Swift 24a0c3de75)
Target: x86_64-apple-macosx10.9
$ time swiftc -o units.exe *.swift
time_stamp.swift:199:3: warning: 'inout' before a parameter name is deprecated, place it before the parameter type instead
                inout _ token: String,
time_stamp.swift:199:3: warning: 'inout' before a parameter name is deprecated, place it before the parameter type instead
                inout _ token: String,
time_nanos.swift:138:2: error: expression was too complex to be solved in reasonable time; consider breaking up the expression into distinct sub-expressions
        print((u_minute + u_second + Nano(500)) + " = " + (1.i + 1.s + 500.ns))

real	0m29.556s
user	0m28.732s
sys	0m0.532s

Excuses, if you think that's a complex expression try to look cute with a plush doll on your head for a day

So the compiler takes 4s more to tell me that a parameter is deprecated but is still unable to handle that expression. Thanks, Swift compiler, that warning is really helpful, unlike actually producing a binary I can run. Of course I reported this as bug SR-838 with a reduced test case that runs faster (you can get the reduced comparison test from GitHub. Browsing their repo looks like other people are also experiencing such compilation problems with apparently less complex code. This experience makes me doubt Swift's viable future as a nice programming language unless waiting for the compiler is your cup of tea. Also the bug tracker feels a little bit desolate. If it is anything like the old one I won't bother with future reports.

So now that the intermission is done, keep in mind that you need Swift compiler version 2.1.1 or this little exercise might be too much to handle!

One Stamp after another

If you had not enough protocols yet, you will squeal of joy to find that we ourselves define a new TimeComponents protocol for the Stamp:

protocol TimeComponents {
        var year: Int { get }
        var week: Int { get }
        var month: Int { get }
        var day: Int { get }
        var hour: Int { get }
        var minute: Int { get }
        var second: Int { get }
        var microsecond: Int { get }
        var millisecond: Int { get }
        var nanosecond: Int { get }

struct Stamp : CustomStringConvertible, TimeComponents {
        var value: Int64 = 0
         more code here…

For the Nano struct I didn't apply this protocol. The reason to create and use this protocol is that later we want to extend the Array type, a generic collection type, with this protocol in order to be able to call these methods on the sequence items. Apart from this protocol, which will be exercised later, the rest of the implementation is pretty mundane. After the struct definition we see a String extension:

extension String {
        var date: Stamp { return Stamp(self) }
        // Avoid losing sanity. Hey, at least this is not java!
        var len: Int { return self.characters.count }

What we are defining here is our pseudo custom literal for strings to invoke the Stamp initializer. On top of that I added the len extension because I dislike typing unnecessary characters. Just after this extension we get the one I mentioned above, an extension on Arrays to overload the subscript operator:

extension Array where Element: TimeComponents {
        // Marking as optional because swift 2.1 doesn't allow throwing inside
        // subscripts yet: http://stackoverflow.com/a/33724709/172690 or does it?
        subscript(position: String) -> [Int]? {
                get {
                        switch (position) {
                                case "week": return self.map() { $0.week }
                                case "year": return self.map() { $0.year }
                                case "month": return self.map() { $0.month }
                                case "day": return self.map() { $0.day }
                                case "hour": return self.map() { $0.hour }
                                case "minute": return self.map() { $0.minute }
                                case "second": return self.map() { $0.second }
                                case "microsecond": return self.map() { $0.microsecond }
                                case "millisecond": return self.map() { $0.millisecond }
                                case "nanosecond": return self.map() { $0.nanosecond }
                                default: return nil

In the bonus generic subscript operator section of the Nim implementation chapter I implemented Kerf's subscript operator using filter procs, which allowed us to pass any kind of proc to be applied to sequences. Here I'm taking a different turn and implementing a string based version like I mentioned in the introduction chapter. This version shows that using strings can be done, but it is not extensible, and in the case of typos this extension returns Nil. This forces extra checks on the caller code. The generic Array type was forced with where Element: TimeComponents to a concrete protocol, so I could write the map() calls using the proper calendar component getters. In Swift you can't coerce the generic Array to a Stamp, if you try you get the message error: type 'Element' constrained to non-protocol type 'Stamp' from the compiler (and also a crash with stacktrace on version 2.1.1 of the compiler).

After some lines dedicated to reimplement basic Nim parsing code for the purpose of keeping it as close as possible to the original, we reach the final self test code of the file. Nothing exceptional there, so let's take a look at the main.swift file which implements the Kerf syntax examples we want to copy:

let a = "2012.01.01".date
print("Example 1: \(a)")
print("Example 2:")
print("\t\(a + 1.d)")
print("\t\("2012.01.01".date + 1.d)")

print("Example 3: \("2012.01.01".date + 1.m + 1.d + 1.h + 15.i + 17.s)")

Basic initialization and operator overloading works fine, this looks just like the Nim code plus the weird string interpolation. Then, just like in the Nim version, we attempt Kerf's 4th example using temporary variables:

let r = (0..<10)
let offsets = r.map() { (1.m + 1.d + 1.h + 15.i + 17.s) * $0 }
let values = offsets.map() { "2012.01.01".date + $0 }

print("Example 4: \(values)")

And it works. In Nim a template was used to map arbitrary expressions to the input sequence. Here in Swift the map() functions accept as parameter closures. If the closure is the last parameter in the function definition it can be omitted from the actual call (between the parentheses) and placed within braces after it. Inside this closure the implicit input parameter is represented as $0 which stands for the first parameter. The second Nim version which was rolled in a single expression can't be used with string interpolation, so a temporary variable is used instead:

let x = String((0..<10)
        .map() { (1.m + 1.d + 1.h + 15.i + 17.s) * $0 }
        .map() { "2012.01.01".date + $0 })
// Swift's compiler agrees that string interpolation is crap and bails out
// if you try to embed the previous expression, so we create a temporal.
print("…again but compressed… \(x)")

print("…again with explicit concatenation… " + String((0..<10)
        .map() { (1.m + 1.d + 1.h + 15.i + 17.s) * $0 }
        .map() { "2012.01.01".date + $0 }))

Alternatively, instead of string interpolation explicit concatenation can be used, as the last expression shows. And at this point we would implement the shorter operator overloaded version for arrays so we could match Kerf's syntax. Unfortunately the complex expression bugs stopped me in all attempts to do so, maybe in a future when Swift is more mature I'll try again. For the last example using the subscript operator to access calendar components, I placed it near to the alternative strongly typed map version:

print("Example 5 b[week]: \(values.map() { $0.week })")
print("Example 5 b[second]: \(values.map() { $0.second })")
print("Example 5 b[second]: \(values["week"])")
print("Example 5 b[second]: \(values["runtime error"])")

As mentioned above, the subscript version may look cleaner, but it allows potential typos, which is why it returns a nil. On the other hand if you use the map() version and try to access the weak variable, you will get a nice compiler error:

main.swift:31:45: error: value of type 'Stamp' has no member 'weak'
        print("Example 5 b[week]: \(values.map() { $0.weak })")
                                                   ^~ ~~~~

For completeness, here is the successful output of this last example part, note the optional sequence syntax in the output:

Example 5 b[week]: [1, 5, 9, 14, 18, 23, 27, 32, 36, 40]
Example 5 b[second]: [0, 17, 34, 51, 8, 25, 42, 59, 16, 33]
Example 5 b[second]: Optional([1, 5, 9, 14, 18, 23, 27, 32, 36, 40])
Example 5 b[second]: nil


Working with Swift is particularly unsatisfying, but I can't put my finger yet on what exactly is causing me more grief. Is it is because the language design feels unnecessarily complex? Is it because the compiler takes ages to do simple things? Is it because of bugs? What matters is that in Swift you can also implement Kerf's timestamp types. Now we only have to wait for better compiler implementations to maybe enjoy the language in the future. In the meantime, you might want to take a laugh at the horrifying Java implementation I came up with.

UPDATE: You know your compiler is in deep shit when the community around it decides to build a time analyzer plugin to navigate around the parts of your compiler which are so poorly implemented they make Gordon Moore weep in a corner. I've worked with about ten wildly different programming languages, each with their own set of problems, and compilation performance was rarely of concern. How spoiled I was.

Patient Woohee is patiently waiting for the compiler to finish, or for a newer compiler that doesn't suck

See the article index or browse articles by tags: design , nim , java , cpp , languages , kerf , programming , swift .
Published on: 06/03/2016 22:52. Last update: 04/06/2016 18:23. rss feed
Copyright 2016 by Grzegorz Adam Hankiewicz.
Generated with ipsum genera. Look at the source code.