Swift Performance Tips for Busy iOS Developers @BradBroulik | BradBroulik@gmail.com BradBroulik.blogspot.com

2 Performance Timeline (Past) Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec Management demands 2X performance boost by Q4 Developers play video games for 8 months Developers install the latest CPUs Management thanks developers for doubling performance Moore’s Law

3 Performance Timeline (Present) Jan Feb Mar Apr May Jun Management demands 2X performance boost by Q2 Developers optimize their code utilizing every technique possible (GCD, caching, code optimizations, etc) Management thanks developers for doubling performance Moore's Law is Dead. Now What?

4 Dimensions of Performance Slower Dimension Faster Heap Allocation Stack More Reference Counting Less Dynamic Method Dispatch Static Is this instance allocated on the stack or heap?

5 Stack allocation Stack a: Person Int Int x: Person Int Int

6 Heap allocation Person Int Int Stack Heap 2 (reference) (reference) a: x:

7 Dimensions of Performance Slower Dimension Faster Heap Allocation Stack More Reference Counting Less Dynamic Method Dispatch Static When I pass this instance around how much reference counting am I going to incur?

8 Reference counting (less) Stack a: Person Int Int x: Person Int Int

9 Person Int Int Stack Heap 2 (reference) (reference) a: x: Reference counting (more) 1 2

10 Dimensions of Performance Slower Dimension Faster Heap Allocation Stack More Reference Counting Less Dynamic Method Dispatch Static When I call a method on this instance is it statically or dynamically dispatched?

11 Which is faster? Swift Objective-C

12 Optimizing Swift Performance Swift can optimize via Static Dispatch Swift Objective-C Which is faster?

13 Which is faster? map for-in forEach

14 (lower is better) Seconds 0.0 4.3 8.5 12.8 17.0 Convert 1,000 UUIDs to Strings 15.89 0.12 0.07 forEach map for-in Avoid for-in Gist: Swift forEach vs map vs for-in performance test map for-in forEach

15 Which is faster? isEmpty) count)

16 •isEmpty is more efficient because it only evaluates the first element; count > 0 evaluates all elements when the collection does not conform to RandomAccessCollection. •Want to enforce this practice in your code base? SwiftLint has a static analysis rule for it. Prefer isEmpty (lower is better) Milliseconds 0.0 0.5 1.0 1.5 2.0 1 million items in Set 1.41 0.26 isEmpty count Gist: Swift isEmpty vs count performance test isEmpty) count)

17 Regular class) Final class) Which is faster?

18 Slower Dimension Faster Heap Allocation Stack More Reference Counting Less Dynamic Method Dispatch Static •All methods become statically dispatched when a class is final •final may be set per method as needed •private methods are implicitly final •final expresses developer intent more clearly Prefer final classes Final class) Regular class)

19 struct B) struct A) Which is faster?

20 Prefer enums over Strings Slower Dimension Faster Heap Allocation Stack More Reference Counting Less Dynamic Method Dispatch Static •enums are allocated on the stack; String characters are allocated on the heap. •enums do not incur a reference count; Strings incur a reference count. •enums are type safe; Strings are not. •Interested in automatically converting your localizable strings, asset names, or storyboard names to enums? SwiftGen has accomplished this task nicely on my projects. struct B) struct A)

21 struct) class) Which is faster?

22 Prefer structs over classes •structs are allocated on the stack; classes are allocated on the heap. •structs do not incur reference counting (when their properties are all value types) •structs implicitly use static dispatch •structs prevent unintended sharing •A structs member wise initializer is a nice convenience Slower Dimension Faster Heap Allocation Stack More Reference Counting Less Dynamic Method Dispatch Static struct) class)

23 struct) Bonus: Which is faster? class)

24 Bonus: Which is faster? (lower is better) Seconds 15.0 15.5 16.0 16.5 17.0 50,000 copies 16.57 15.71 class struct Gist: Swift struct vs class (with many strings) struct) class)

2 25 Struct with many strings Person String String String Person String String String Class 2 Class 2 Class Stack Stack Heap a: y: Optimizing Swift Performance 1 2 3 4 5 6

26 Class with many strings Person String String String Stack Heap 2 (reference) (reference) b: z: 1 2

27 Prefer Structs over Classes (until you reach their point of diminishing returns) •structs do not incur reference counting (when their properties are all value types) •Caution: A struct property will incur reference counting and heap allocation if the property has a reference type for underlying storage (string characters are stored on the heap). Therefore, structs have a point of diminishing returns in regards to reference counting that multiplies proportionally by their number of properties which are reference types. A struct containing more than 1 reference type begins to incur a higher reference count than a comparative class.

28 Inheritance Extensions Which is faster?

Inheritance Extensions 29 Inheritance uses dynamic dispatch Uses static dispatch Dynamic dispatch definitions Static dispatch definitions Which is faster? Protocol extension method dispatch flow chart

30 Prefer extensions over inheritance •Protocols allow us to define methods with either static or dynamic dispatch; inheritance uses dynamic dispatch. •Methods defined within protocol extensions use static dispatch and methods defined within protocol types are dynamically dispatched •Protocol extensions are not bound to single inheritance •Protocol extensions may be applied to both classes or structs •Unfortunately, protocols can not inherit stored properties but classes can Slower Dimension Faster Heap Allocation Stack More Reference Counting Less Dynamic Method Dispatch Static

31 Which is faster? Protocol Type Generic

32 Which is faster? Protocol Type Generic Dynamic dispatch definition Uses static dispatch Uses dynamic dispatch

33 Generic Specialization compiler optimized

34 Prefer Generics over Protocol Types Understanding Swift Performance •Generic methods will be optimized via generic specialization during Whole Module Optimization •Generic specialization creates a type-specific version of method that enables static dispatch. A type-specific method will be created per type in use. •If generic implementation is a struct you also get allocation on the stack and no reference counting (when struct contains no references) •And you still get all the benefits of polymorphism Slower Dimension Faster Heap Allocation Stack More Reference Counting Less Dynamic Method Dispatch Static

35 Parse JSON on a Background Thread Concurrent Programming with GCD in Swift 3 Main Thread Dispatch Queue User Interface Data Transform Data

36 Concurrent Programming with GCD in Swift 3 Group Data Data Result Grouping work requests

37 Enable Whole Module Optimization Optimizing Swift Performance Builds take longer but the generated binary will run faster because the compiler makes more aggressive generic specialization and will automatically enable static dispatch when no overrides exist

38 Dimensions of Performance Slower Dimension Faster Heap Allocation Stack More Reference Counting Less Dynamic Method Dispatch Static Ask yourself these questions when reading or writing Swift code: 1. Is this instance allocated on the stack or heap? 2. When I pass this instance around how much reference counting am I going to incur? 3. When I call a method on this instance is it statically or dynamically dispatched? Understanding Swift Performance

39 Performance Timeline (Future) Jan Feb Mar Apr May Jun Management demands performance boost by Q2 Swift developer enables Whole Module Optimization Management thanks developers for improving performance Swift developers play video games for 2 months

Slides: bit.ly/swift-performance Thank You! @BradBroulik | BradBroulik@gmail.com BradBroulik.blogspot.com