Friendly Finite State Machine Syntax for Swift on macOS, iOS, tvOS and watchOS
Inspired by Uncle Bob's SMC syntax, Swift FSM is a pure Swift DSL for declaring and operating a Finite State Machine (FSM).
This guide presumes some familiarity with FSMs and specifically the SMC syntax linked above. Swift FSM makes liberal use of @resultBuilder blocks, operator overloads, callAsFunction(), and trailing closures, all in combination with one another - familiarity with these concepts will also be helpful.
Swift FSM is a Swift Package, importable through the Swift Package Manager, and requires macOS 13, iOS 16, tvOS 16 and watchOS 9 or later, alongside Swift 5.8 or later. It supports Swift Concurrency, and is tested and compiled against Swift 6 concurrency rules.
It has two dependencies - Apple’s Algorithms, and (in one small corner) my own Reflective Equality.
Borrowing from SMC, we will use the example of a subway turnstile system. This turnstile has two possible states: Locked, and Unlocked, and two possible events: Coin, and Pass.
The logic is as follows (from Uncle Bob, emphasis added):
- Given we are in the Locked state, when we get a Coin event, then we transition to the Unlocked state and invoke the unlock action.
- Given we are in the Locked state, when we get a Pass event, then we stay in the Locked state and invoke the alarm action.
- Given we are in the Unlocked state, when we get a Coin event, then we stay in the Unlocked state and invoke the thankyou action.
- GIven we are in the Unlocked state, when we get a Pass event, then we transition to the Locked state and invoke the lock action.
Following Uncle Bob’s examples, we will build up our table bit by bit to demonstrate the different syntactic possibilities of Swift FSM and how they compare to SMC:
SMC:
Initial: Locked
FSM: Turnstile
{
Locked {
Coin Unlocked unlock
Pass Locked alarm
}
Unlocked {
Coin Unlocked thankyou
Pass Locked lock
}
}
Swift FSM (with additional code for context):
import SwiftFSM
class MyClass: SyntaxBuilder {
enum State { case locked, unlocked }
enum Event { case coin, pass }
let fsm = FSM<State, Event>(initialState: .locked)
func myMethod() throws {
try fsm.buildTable {
define(.locked) {
when(.coin) | then(.unlocked) | unlock
when(.pass) | then(.locked) | alarm
}
define(.unlocked) {
when(.coin) | then(.unlocked) | thankyou
when(.pass) | then(.locked) | lock
}
}
try fsm.handleEvent(.coin)
}
}class MyClass: SyntaxBuilder {
The SyntaxBuilder protocol provides the methods define, when, and then necessary to build the transition table. It has two associated types, State and Event, which must be Hashable & Sendable.
let fsm = FSM<State, Event>(initialState: .locked)
FSM is generic over State and Event. As with SyntaxBuilder, State and Event must be Hashable. Here we have used an Enum, specifying the initial state of the FSM as .locked.
try fsm.buildTable {
fsm.buildTable is a throwing function - though the type system will prevent various illogical statements, there are some issues that can only be detected at runtime.
define(.locked) {
The define statement roughly corresponds to the ‘Given’ keyword in the natural language description of the FSM. It is expected however that you will only write one define per state.
define takes two arguments - a State instance, and a Swift @resultBuilder block.
when(.coin) | then(.unlocked) | unlock
As we are inside a define block, we take the .locked state as a given. We can now list our transitions, with each line representing a single transition. In this case, when we receive a .coin event, we will then transition to the .unlocked state and call unlock.
unlock is a function, also declarable as follows:
when(.coin) | then(.unlocked) | { unlock() //; otherFunction(); etc. }
(see Arrays of Actions for other syntax variants)
The | (pipe) operator binds transitions together. It feeds the output of the left hand side into the input of the right hand side, as you might expect in a terminal.
try fsm.handleEvent(.coin)
The FSM instance will look up the appropriate transition for its current state, call the associated function, and transition to the associated next state. In this case, the FSM will call the unlock function and transition to the unlocked state. If no transition is found, it will do nothing, and if compiled for debugging, will print a warning message.
Currently, in order to work reasonably in a SwiftUI world, the FSM runs its critical loops on the Main Actor, and expects that the actions you pass it will do so as well. Internally, the method handleEvent and all actions are therefore annotated @MainActor.
There are two versions of handleEvent, the first as shown above and in all the examples, and the second named handleEventAsync which must be called with await.
The four function signatures for actions that are accepted are as follows:
@MainActor () -> Void
@MainActor () async -> Void
@MainActor (Event) -> Void
@MainActor (Event) async -> VoidThese are handled interchangeably without any additional syntax. The tradeoff is that you must choose the appropriate handleEvent function to call when using the FSM. The synchronous handleEvent function must be called with try, because it will throw if it is asked to call an async action. Though the asynchronous handleEventAsync function must be called with await, it does not throw any errors as it is a valid context from which to call both synchronous and asynchronous actions.
Note also that there are action signatures that take an event as an argument. This can be useful in situations where you wish to pass an associated value along with an event enum that can then be received by your callback function (see Using Events to Pass Values for more details on how to implement this) .
If you wish to pass an array of actions anywhere where such arrays are accepted, you will need to use a special & operator in order to enable mixing and matching of the four different action signatures:
when(.coin) | then(.unlocked) | first & secondAsync & thirdWithEvent ...This is equivalent to the more verbose (but equally valid):
when(.coin) | then(.unlocked) | { event in first(); await secondAsync(); thirdWithEvent(event) ... }Now let's add an Alarming state that must be reset by a repairman:
SMC:
Initial: Locked
FSM: Turnstile
{
Locked {
Coin Unlocked unlock
Pass Alarming alarmOn
Reset - {alarmOff lock}
}
Unlocked {
Reset Locked {alarmOff lock}
Coin Unlocked thankyou
Pass Locked lock
}
Alarming {
Coin - -
Pass - -
Reset Locked {alarmOff lock}
}
}
Swift FSM:
try fsm.buildTable {
define(.locked) {
when(.coin) | then(.unlocked) | unlock
when(.pass) | then(.alarming) | alarmOn
when(.reset) | then() | alarmOff & lock
}
define(.unlocked) {
when(.reset) | then(.locked) | alarmOff & lock
when(.coin) | then(.unlocked) | thankyou
when(.pass) | then(.locked) | lock
}
define(.alarming) {
when(.coin) | then()
when(.pass) | then()
when(.reset) | then(.locked) | alarmOff & lock
}
}then() with no argument means ‘no change’, and the FSM will remain in the current state. The actions pipe is also optional - if a transition performs no actions, it can be omitted.
Notice the duplication of the Reset transition. In all three states the Reset event does the same thing. It transitions to the Locked state and it invokes the lock and alarmOff actions. This duplication can be eliminated by using a Super State as follows:
SMC:
Initial: Locked
FSM: Turnstile
{
// This is an abstract super state.
(Resetable) {
Reset Locked {alarmOff lock}
}
Locked : Resetable {
Coin Unlocked unlock
Pass Alarming alarmOn
}
Unlocked : Resetable {
Coin Unlocked thankyou
Pass Locked lock
}
Alarming : Resetable { // inherits all it's transitions from Resetable.
}
}
Swift FSM:
try fsm.buildTable {
let resetable = SuperState {
when(.reset) | then(.locked) | alarmOff & lock
}
define(.locked, adopts: resetable) {
when(.coin) | then(.unlocked) | unlock
when(.pass) | then(.alarming) | alarmOn
}
define(.unlocked, adopts: resetable) {
when(.coin) | then(.unlocked) | thankyou
when(.pass) | then(.locked) | lock
}
define(.alarming, adopts: resetable)
}SuperState takes the same @resultBuilder as define, however it does not take a starting state. The starting state is taken from the define statement to which it is passed. Passing SuperState instances to a define call will add the transitions declared in each of the SuperState instances before the other transitions declared in the define.
If a SuperState instance is passed to define, the @resultBuilder argument is optional.
SuperState instances themselves can adopt other SuperState instances, and will combine them together in the same way as define:
let s1 = SuperState { when(.coin) | then(.unlocked) | unlock }
let s2 = SuperState { when(.pass) | then(.alarming) | alarmOn }
let s3 = SuperState(adopts: s1, s2)
// s3 is equivalent to:
let s4 = SuperState {
when(.coin) | then(.unlocked) | unlock
when(.pass) | then(.alarming) | alarmOn
}By default, transitions declared in a SuperState cannot be overridden by their adopters. The following code is therefore assumed to be accidental and throws:
let s1 = SuperState { when(.coin) | then(.unlocked) | unlock }
let s2 = SuperState(adopts: s1) {
when(.coin) | then(.locked) | beGrumpy // 💥 error: clashing transitions
}
define(.locked, adopts: s1) {
when(.coin) | then(.locked) | beGrumpy // 💥 error: clashing transitions
}If you wish to override a SuperState transition, you must make this explicit using the overriding { } block:
let s1 = SuperState { when(.coin) | then(.unlocked) | unlock }
let s2 = SuperState(adopts: s1) {
overriding {
when(.coin) | then(.locked) | beGrumpy // ✅ overrides inherited transition
}
}
define(.locked, adopts: s1) {
overriding {
when(.coin) | then(.locked) | beGrumpy // ✅ overrides inherited transition
}
}The overriding block indicates to Swift FSM that any transitions contained within it override any inherited transitions with the same initial states and events.
As multiple inheritance is allowed, overrides replace all matching transitions:
let s1 = SuperState { when(.coin) | then(.unlocked) | doSomething }
let s2 = SuperState { when(.coin) | then(.unlocked) | doSomethingElse }
define(.locked, adopts: s1, s2) {
overriding {
when(.coin) | then(.locked) | doYetAnotherThing // ✅ overrides both inherited transitions
}
}Without the overriding, this multiple inheritance would otherwise create duplicate transitions:
let s1 = SuperState { when(.coin) | then(.unlocked) | doSomething }
let s2 = SuperState { when(.coin) | then(.unlocked) | doSomethingElse }
define(.locked, adopts: s1, s2) // 💥 error: duplicate transitionsIf overriding is used where there is nothing to override, the FSM will throw:
define(.locked) {
overriding {
when(.coin) | then(.locked) | beGrumpy // 💥 error: nothing to override
}
}Writing overriding in the parent rather than the child will throw:
let s1 = SuperState {
overriding {
when(.coin) | then(.locked) | beGrumpy
}
}
let s2 = SuperState(adopts: s1) { when(.coin) | then(.unlocked) | unlock }
// 💥 error: overrides are out of orderAttempting to override within the same SuperState { } or define { } will also throw:
define(.locked) {
when(.coin) | then(.locked) | doSomething
overriding {
when(.coin) | then(.locked) | doSomethingElse
}
}
// 💥 error: duplicate transitionsIn this scope, the word override has no meaning and therefore is ignored by the error handler. What remains is therefore two duplicate transitions, resulting in an error.
Overrides in Swift FSM follow the usual rules of inheritance. In a chain of overrides, it is the final transition in that chain that takes precedence:
let s1 = SuperState { when(.coin) | then(.unlocked) | a1 }
let s2 = SuperState(adopts: s1) { overriding { when(.coin) | then(.unlocked) | a2 } }
let s3 = SuperState(adopts: s2) { overriding { when(.coin) | then(.unlocked) | a3 } }
let s4 = SuperState(adopts: s3) { overriding { when(.coin) | then(.unlocked) | a4 } }
define(.locked, adopts: s4) {
overriding { when(.coin) | then(.unlocked) | a5 } // ✅ overrides all others
}
fsm.handleEvent(.coin) // 'a5' is calledIn the previous example, the fact that the alarm is turned on every time the Alarming state is entered and is turned off every time the Alarming state is exited, is hidden within the logic of several different transitions. We can make it explicit by using entry actions and exit actions.
SMC:
Initial: Locked
FSM: Turnstile
{
(Resetable) {
Reset Locked -
}
Locked : Resetable <lock {
Coin Unlocked -
Pass Alarming -
}
Unlocked : Resetable <unlock {
Coin Unlocked thankyou
Pass Locked -
}
Alarming : Resetable <alarmOn >alarmOff - - -
}
Swift FSM:
try fsm.buildTable {
let resetable = SuperState {
when(.reset) | then(.locked)
}
define(.locked, adopts: resetable, onEntry: lock*) {
when(.coin) | then(.unlocked)
when(.pass) | then(.alarming)
}
define(.unlocked, adopts: resetable, onEntry: unlock*) {
when(.coin) | then(.unlocked) | thankyou
when(.pass) | then(.locked)
}
define(.alarming, adopts: resetable, onEntry: alarmOn*, onExit: alarmOff*)
}onEntry and onExit are the final arguments to define and specify an array of entry and exit actions to be performed when entering or leaving the defined state. Note that these require array syntax rather than varargs, as a work around for limitations in Swift’s matching algorithm for functions that take multiple closure arguments.
As the array is heterogeneous (it can include any of the four function types accepted by SwiftFSM as valid actions), a special postfix operator * is provided to convert a single one of these into an array of AnyAction.
let _ = unlock* // preferred syntax, same as...
let _ = Array(unlock) // same as...
let _ = [AnyAction(unlock)]
let _ = unlock & thankyou // preferred syntax, same as...
let _ = AnyAction(unlock) & thankyou // same as...
let _ = AnyAction(unlock) & AnyAction(thankyou) // same as...
let _ = [AnyAction(unlock), AnyAction(thankyou)]SuperState instances can also accept entry and exit actions:
let resetable = SuperState(onEntry: lock*) {
when(.reset) | then(.locked)
}
define(.locked, adopts: resetable) {
when(.coin) | then(.unlocked)
when(.pass) | then(.alarming)
}
// equivalent to:
define(.locked, onEntry: lock*) {
when(.reset) | then(.locked)
when(.coin) | then(.unlocked)
when(.pass) | then(.alarming)
}SuperState instances also inherit entry and exit actions from their superstates:
let s1 = SuperState(onEntry: unlock*) { when(.coin) | then(.unlocked) }
let s2 = SuperState(onEntry: alarmOn*) { when(.pass) | then(.alarming) }
let s3 = SuperState(adopts: s1, s2)
// s3 is equivalent to:
let s4 = SuperState(onEntry: [unlock, alarmOn]) {
when(.coin) | then(.unlocked)
when(.pass) | then(.alarming)
}In SMC, entry and exit actions are invoked even if the state does not change. In the example above, the unlock entry action would be called on all transitions into the Unlocked state, even if the FSM is already in the Unlocked state.
In contrast, Swift FSM’s default behaviour is to invoke entry and exit actions only if there is a state change. In the example above, this means that, in the .unlocked state, after a .coin event, unlock will not be called.
This policy is configurable, by passing .executeAlways as the second argument to FSM.init:
FSM<State, Event>(initialState: .locked, actionsPolicy: .executeAlways)This setting replicates SMC entry/exit action behaviour. The default is .executeOnChangeOnly and is not a required argument.
All statements must be made in the form define { when | then | actions }. See Expanded Syntax below for exceptions to this rule.
when statements accept vararg Event instances for convenience.
define(.locked) {
when(.coin, or: .pass, ...) | then(.unlocked) | unlock
}
// equivalent to:
define(.locked) {
when(.coin) | then(.unlocked) | unlock
when(.pass) | then(.unlocked) | unlock
...
}Actions can receive the event that resulted in them being called, which can be useful if that event includes a value you may wish to pass on. SwiftFSM requires a special struct FSMValue<T> and protocol EventWithValues that work together to enable you to do this.
enum Event: EventWithValues {
case .coin(FSMValue<Int>), ...
var coinValue: Int? {
guard case .coin(let amount) = event else { return nil }
return amount.wrappedValue
}
}
func main() throws {
try fsm.buildTable(initialState: .locked) {
define(.locked) {
when(.coin(.any)) | then(.verifyingPayment) | verifyPayment
// here we use .any to match any value
}
}
try fsm.handleEvent(.coin(50))
// here we pass a specific value that will be matched by .any
}
func verifyPayment(_ event: Event) {
// here we receive the actual value passed to handleEvent: .coin(50)
if let amount = event.coinValue {
if amount >= requiredAmount {
letThemThrough()
} else {
insufficientPayment(shortfall: requiredAmount - amount)
}
}
}when(.coin(.any)) works polymorphically. We wish to write a single transition row that will match against any value inside .coin(someValue) and pass that value on to the verifyPayment function. Without the combination of EventWithValues and FSMValue<T>, the table would have to be written as follows:
try fsm.buildTable(initialState: .locked) {
define(.locked) {
when(.coin(1)) | then(.verifyingPayment) | verifyPayment
when(.coin(2)) | then(.verifyingPayment) | verifyPayment
when(.coin(3)) | then(.verifyingPayment) | verifyPayment
when(.coin(4)) | then(.verifyingPayment) | verifyPayment
... // and so on for all relevant values
}
}By using .any, the transition to .verifyingPayment will be activated when a .coin event is received, no matter what the wrapped value is. That wrapped value is then passed into the verifyPayment function where it can be examined. FSMValue provides a convenience var wrappedValue: T?, which returns an optional value (potentially nil if it is called on a .any instance or if T is optional and nil).
FSMValue conforms to ExpressibleByIntegerLiteral, ExpressibleByFloatLiteral, ExpressibleByArrayLiteral, ExpressibleByDictionaryLiteral, ExpressionByNilLiteral, and ExpressionByStringLiteral forwarding to the wrapped type where relevant. It also forwards conformances to Equatable, Comparable, and AdditiveArithmetic where relevant, as well as RandomAccessCollection and its parent protocols for Arrays, and subscript access for Dictionaries. Finally, it forwards CustomStringConvertible, which also covers most uses of ExpressibleByStringInterpolation.
A few examples:
let s: FSMValue<String> = "1" // equivalent to .some("1")
let i: FSMValue<Int> = 1 // equivalent to .some(1)
let ai: FSMValue<[Int]> = [1] // equivalent to .some([1])
let _ = s + "1" // "11"
let _ = i + 1 // 2
let _ = ai[0] // 1
let _ = ai[0] == i // true
let _ = ai[0] > i // false
let _ = "\(i)\(s)" // "11"Warning: where forward operations are available on the wrapped type, be aware that this will crash if you attempt to access a value on a .any instance (as there is no instance of the wrapped type). .any should only appear inside a define statement - there are no circumstances in which it would be useful or meaningful to pass such an event with FSMValue.any to handleEvent.
Equally, you should always unwrap FSMValue<T> instances before continuing - indeed, all convenience methods that return a value return an instance of T and not of FSMValue<T>. Comparing or hashing FSMValue instances is not recommended - they are specialised DTOs used by the FSM and should not escape.
The @resultBuilder blocks in SwiftFSM do not support control flow logic. Though is it possible to enable such logic, it would be misleading:
define(.locked) {
if something { // ⛔️ does not compile
when(.pass) | then(.unlocked) | unlock
} else {
when(.pass) | then(.alarming) | alarmOn
}
...
}If the if/else block were evaluated by the FSM at transition time, this would be a useful addition. However what we are doing inside these blocks is compiling our state transition table. The use of if and else in this manner is more akin to the conditional compilation statements #if/#else - based on a value defined at compile time, only one transition or the other will be added to the table.
If you do have a use for this kind of conditional compilation, please open an issue. See Expanded Syntax for alternative ways to evaluate conditional statements at transition time rather than compile time.
Warning - most Swift FSM function calls and initialisers take additional parameters file: String = #file and line: Int = #line. This is similar to XCTest assertions, and allows Swift FSM to produce errors that pinpoint the location of problematic statement/s.
As these cannot be hidden, please note that there are unlikely to be any circumstances in which it would be useful or necessary to override these default arguments with alternate values.
All blocks must contain at least one transition, otherwise an error will be thrown:
try fsm.buildTable { } //💥 error: empty table
try fsm.buildTable {
define(.locked) { } // 💥 error: empty block
}Transitions are duplicates if they share the same start state, event, and next state:
try fsm.buildTable {
define(.locked) {
when(.coin) | then(.unlocked) | unlock
when(.coin) | then(.unlocked) | lock
}
}
// 💥 error: duplicate transitionsA logical clash occurs when transitions share the same start state and event, but their next states differ:
try fsm.buildTable {
define(.locked) {
when(.coin) | then(.unlocked) | unlock
when(.coin) | then(.locked) | lock
}
}
// 💥 error: logical clashThough the two transitions are distinct from one another, logically they cannot co-exist - the .coin event must lead either to the .unlocked state or to the .locked state. It cannot lead to both.
Because .any matches all cases, the following would throw an error:
try fsm.buildTable(initialState: .locked) {
define(.locked) {
when(.coin(.any)) | then(.verifyingPayment) | verifyPayment
when(.coin(50) | then(.unlocked) | pass
}
}
//💥 error: logical clashThe .any case already includes .some(50) and this specific case is therefore referenced ambiguously. It would however be possible to write the following:
try fsm.buildTable(initialState: .locked) {
define(.locked) {
when(.coin(20) | then(.verifyingPayment) | verifyPayment
when(.coin(50) | then(.unlocked) | pass
}
}
// ✅ transitions are logically distinctAdditional calls to fsm.buildTable { } will throw a TableAlreadyBuiltError.
Swift FSM will throw an error if your State and/or Event types (or their children) inherit from NSObject.
State and Event instances are hashed to produce keys for the transition Dictionary. These keys are then recreated and reused each time fsm.handleEvent is called. This is not an issue for most Swift types, as Hashable conformance will have to be declared explicitly. NSObject however already conforms to Hashable, and is hashed by instance identity, rather than by value. This would lead to a defunct transition table where all transition lookups fail, and therefore throws an error.
This is an edge case and it is extremely unlikely that you will ever encounter this error. Nonetheless, the check is quite exhaustive - If you would like to know more about the mechanism involved, see Reflective Equality.
Each call to handleEvent() requires a single operation to find the correct transition in the table. Though O(1) is ideal, this table-based system still has 2-3x the basic overhead of a nested switch case statement.
Whilst Swift FSM matches most of the syntax of SMC, it also introduces some new possibilities of its own. None of this additional syntax is required, and is provided purely for convenience.
Let’s imagine an extension to our turnstile rules: under some circumstances, we want to enforce the ‘everyone pays’ rule by entering the alarming state if a .pass is detected when still in the .locked state, yet in others, perhaps at rush hour, we want to be more permissive in order to avoid disruption to other passengers.
We could implement a time of day check somewhere else in the system, perhaps inside the implementation of the alarmOn function to decide what the appropriate behaviour should be:
try fsm.buildTable {
...
define(.locked) {
when(.pass) | then(.alarming) | handleAlarm
}
...
}
// some other file...
enum Enforcement: Predicate { case weak, strong }
let enforcement = Enforcement.weak
func handleAlarm() {
switch enforcement {
case .weak: smile()
case .strong: defconOne()
}
}But we now have some aspects of our state transition logic declared inside the transition table, and other aspects declared elsewhere.
Furthermore, we must transition to the .alarming state, regardless of the Enforcement policy. But what if different policies called for different transitions altogether?
An alternative might be to introduce extra events to differentiate between the new policies:
try fsm.buildTable {
...
define(.locked) {
when(.passWithEnforcement) | then(.alarming) | defconOne
when(.passWithoutEnforcement) | then(.locked) | smile
}
...
}This allows us both to call different functions, and to transition to different states, depending on the enforcement policy, all whilst keeping all of our logic inside the transition table.
In order to make this change however, every transition that originally responded to the .pass event will need to be rewritten twice, once for each of the two new versions of this event, even if they are identical in both cases. In no time at all, the state transition table is going to become unmanageably long, and littered with duplication.
The Swift FSM Solution
import SwiftFSM
class MyClass: ExpandedSyntaxBuilder {
enum State { case locked, unlocked }
enum Event { case coin, pass }
enum Enforcement: Predicate { case weak, strong }
let fsm = FSM<State, Event>(initialState: .locked)
func myMethod() throws {
try fsm.buildTable {
...
define(.locked) {
matching(Enforcement.weak) | when(.pass) | then(.locked) | smile
matching(Enforcement.strong) | when(.pass) | then(.alarming) | defconOne
when(.coin) | then(.unlocked)
}
...
}
fsm.handleEvent(.pass, predicates: Enforcement.weak)
}
}Here we have introduced a new keyword matching, and two new protocols, ExpandedSyntaxBuilder and Predicate.
define(.locked) {
matching(Enforcement.weak) | when(.pass) | then(.locked) | smile
matching(Enforcement.strong) | when(.pass) | then(.alarming) | defconOne
when(.coin) | then(.unlocked) | unlock
}Given that we are in the .locked state:
- If
Enforcementis.weak, when we get a.pass, transition to.lockedandsmile - If
Enforcementis.strong, when we get a.pass, transition to.alarminganddefconOne - Regardless of
Enforcement, when we get a.coin, transition to.unlockedandunlock
In this system, only those statements that depend upon the Enforcement policy need know it has been added, and all other existing statements that do not depend upon it continue to work as they always did.
ExpandedSyntaxBuilder implements SyntaxBuilder, providing all the SMC-equivalent syntax, alongside the new matching statements for working with predicates.
Predicate requires the conformer to be Hashable, Sendable and CaseIterable. It is possible to use any type you wish, as long as your conformance to Hashable and CaseIterable makes logical sense. In practice, this is likely to limit Predicates to Enums without associated types, as these can be automatically conformed to CaseIterable.
when(.coin) | then(.unlocked)In the above, no Predicate is specified, and its full meaning must therefore be inferred from context. The scope for contextual inference is the sum of the builder blocks for all SuperState and define calls inside fsm.buildTable { }.
In our example, the type Enforcement appears in a matching statement elsewhere in the table, and Swift FSM will therefore infer the absent matching statement as follows:
when(.coin) | then(.unlocked)
// is inferred to mean:
matching(Enforcement.weak) | when(.coin) | then(.unlocked)
matching(Enforcement.strong) | when(.coin) | then(.unlocked)Transitions in Swift FSM are are therefore Predicate agnostic by default, matching any given Predicate unless otherwise specified. In this way, matching is an optional modifier that constrains the transition to one or more specific Predicate cases.
There is no limit on the number of Predicate types that can be used in one table (see Predicate Performance for practical limitations). The following (contrived and rather silly) expansion of the original Predicate example remains valid:
enum Enforcement: Predicate { case weak, strong }
enum Reward: Predicate { case positive, negative }
try fsm.buildTable {
...
define(.locked) {
matching(Enforcement.weak) | when(.pass) | then(.locked) | lock
matching(Enforcement.strong) | when(.pass) | then(.alarming) | alarmOn
when(.coin) | then(.unlocked) | unlock
}
define(.unlocked) {
matching(Reward.positive) | when(.coin) | then(.unlocked) | thankyou
matching(Reward.negative) | when(.coin) | then(.unlocked) | idiot
when(.pass) | then(.locked) | lock
}
...
}
fsm.handleEvent(.pass, predicates: Enforcement.weak, Reward.positive)The same inference rules also apply:
when(.coin) | then(.unlocked)
// types Enforcement and Reward appear elsewhere in context
// when(.coin) | then(.unlocked) is now equivalent to:
matching(Enforcement.weak, and: Reward.positive) | when(.coin) | then(.unlocked)
matching(Enforcement.strong, and: Reward.positive) | when(.coin) | then(.unlocked)
matching(Enforcement.weak, and: Reward.negative) | when(.coin) | then(.unlocked)
matching(Enforcement.strong, and: Reward.negative) | when(.coin) | then(.unlocked)As seen in the above example, multiple predicates can be combined in a single matching statement, by using the and: Predicate... and or: Predicate... varargs arguments.
enum A: Predicate { case x, y, z }
enum B: Predicate { case x, y, z }
enum C: Predicate { case x, y, z }
matching(A.x, or: A.y)... // if A.x OR A.y
matching(A.x, or: A.y, A.z)... // if A.x OR A.y OR A.z
matching(A.x, and: B.x)... // if A.x AND B.x
matching(A.x, and: B.x, C.x)... // if A.x AND B.x AND C.x
matching(A.x, or: A.y, A.z, and: B.x, C.x)... // if (A.x OR A.y OR A.z) AND B.x AND C.x
matching(A.x, or: B.x)... // ⛔️ does not compile: OR types must be the same
matching(A.x, and: A.y)... // 💥 error: cannot match A.x AND A.y simultaneously
fsm.handleEvent(.coin, predicates: A.x, B.x, C.x)In Swift FSM, matching(and:) means that we expect both predicates to be present at the same time, whereas mathing(or:) means that we expect any and only one of them to be present.
Swift FSM expects exactly one instance of each Predicate type present in the table to be passed to each call to handleEvent, as in the example above, where fsm.handleEvent(.coin, predicates: A.x, B.x, C.x) contains a single instance of types A, B and C. Accordingly, A.x AND A.y should never occur - only one can be present. Therefore, predicates passed to matching(and:) must all be of a different type. This cannot be checked at compile time, and therefore throws at runtime if violated.
In contrast, matching(or:) specifies multiple possibilities for a single Predicate. Predicates joined by or must therefore all be of the same type, and attempting to pass different Predicate types to matching(or:) will not compile (see Implicit Clashes for more information on this limitation).
Warning - nested matching statements are combined by AND-ing them together, which makes it possible inadvertently to create a conflict.
define(.locked) {
matching(A.x) {
matching(A.y) {
// 💥 error: cannot match A.x AND A.y simultaneously
}
}
}matching(or:) statements are also combined using AND:
define(.locked) {
matching(A.x, or: A.y) {
matching(A.z) {
// 💥 error: cannot match A.x AND A.z simultaneously
// 💥 error: cannot match A.y AND A.z simultaneously
}
}
}Valid nested matching(or:) statements are combined as follows:
define(.locked) {
matching(A.x, or: A.y) {
matching(B.x, or: B.y) {
// ✅ logically matches (A.x OR A.y) AND (B.x OR B.y)
// internally translates to:
// 1. matching(A.x, and: B.x)
// 2. matching(A.x, and: B.y)
// 3. matching(A.y, and: B.x)
// 4. matching(A.y, and: B.y)
}
}
}define(.locked) {
matching(Enforcement.weak) | when(.coin) | then(.unlocked)
matching(Reward.negative) | when(.coin) | then(.locked)
}
// 💥 error: implicit clashThe two transitions above appear to be different from one another, until we reconsider the inference rules for multiple Predicate types:
define(.locked) {
matching(Enforcement.weak) | when(.coin) ...
// inferred as:
matching(Enforcement.weak, and: Reward.positive) | when(.coin) ...
matching(Enforcement.weak, and: Reward.negative) | when(.coin) ... // 💥 clash
matching(Reward.negative) | when(.coin) ...
// inferred as:
matching(Enforcement.weak, and: Reward.negative) | when(.coin) ... // 💥 clash
matching(Enforcement.strong, and: Reward.negative) | when(.coin) ...We can break the deadlock by disambiguating at least one of the statements:
define(.locked) {
matching(Enforcement.weak, and: Reward.positive) | when(.coin) | then(.unlocked)
matching(Reward.negative) | when(.coin) | then(.locked)
}
// ✅ inferred as:
define(.locked) {
matching(Enforcement.weak, and: Reward.positive) | when(.coin) | then(.unlocked)
// matching(Enforcement.weak, and: Reward.negative) ... removed by disambiguation
matching(Enforcement.weak, and: Reward.negative) | when(.coin) | then(.locked)
matching(Enforcement.strong, and: Reward.negative) | when(.coin) | then(.locked)
}In some cases, Swift FSM can break the deadlock without disambiguation:
define(.locked) {
matching(Enforcement.weak, and: Reward.positive) | when(.coin) | then(.unlocked)
matching(Enforcement.weak) | when(.coin) | then(.locked)
}
// ✅ inferred as:
define(.locked) {
matching(Enforcement.weak, and: Reward.positive) | when(.coin) | then(.unlocked)
matching(Enforcement.weak, and: Reward.negative) | when(.coin) | then(.locked)
}Swift FSM defers to the statement that explicitly specifies the greatest number of predicates - in this case, the first statement matching(Enforcement.weak, and: Reward.positive), which specifies two predicates, versus the second statement’s single predicate matching(Enforcement.weak).
Following the inference logic, connecting different types using the word ‘or’ is not allowed:
define(.locked) {
matching(Enforcement.weak, or: Reward.negative) | when(.coin) | then(.unlocked)
}
// ⛔️ does not compile, because it implies:
define(.locked) {
matching(Enforcement.weak) | when(.coin) | then(.unlocked)
matching(Reward.negative) | when(.coin) | then(.unlocked)
}
// 💥 error: implicit clashIf we were to call handleEvent(.coin, predicates: Enforcement.weak, Reward.negative) with such a table, there would be no reasonable way to decide which transition to perform. Unlike between-predicates implicit clashes, within-predicates clashes are eliminated through the type system.
matching(Enforcement.weak) | when(.pass) /* duplication */ | then(.locked)
matching(Enforcement.strong) | when(.pass) /* duplication */ | then(.alarming)In this example, when(.pass) is duplicated. We can factor this out using a context block:
when(.pass) {
matching(Enforcement.weak) | then(.locked)
matching(Enforcement.strong) | then(.alarming)
}The full example would now be:
try fsm.buildTable {
define(.locked) {
when(.pass) {
matching(Enforcement.weak) | then(.locked)
matching(Enforcement.strong) | then(.alarming)
}
when(.coin) | then(.unlocked)
}
}then and matching support context blocks in a similar way:
try fsm.buildTable {
define(.locked) {
then(.unlocked) {
when(.pass) {
matching(Enforcement.weak) | doSomething
matching(Enforcement.strong) | doSomethingElse
}
}
}
// or identically:
define(.locked) {
when(.pass) {
then(.unlocked) {
matching(Enforcement.weak) | doSomething
matching(Enforcement.strong) | doSomethingElse
}
}
}
}try fsm.buildTable {
define(.locked) {
matching(Enforcement.weak) {
when(.coin) | then(.unlocked) | somethingWeak
when(.pass) | then(.alarming) | somethingElseWeak
}
matching(Enforcement.strong) {
when(.coin) | then(.unlocked) | somethingStrong
when(.pass) | then(.alarming) | somethingElseStrong
}
}
}The keyword actions is also available for function call context blocks:
try fsm.buildTable {
define(.locked) {
actions(someCommonFunction) {
when(.coin) | then(.unlocked)
when(.pass) | then(.alarming)
}
}
}matching(predicate) {
// everything in scope matches 'predicate'
}
when(event) {
// everything in scope responds to 'event'
}
then(state) {
// everything in scope transitions to 'state'
}
actions(functionCalls) {
// everything in scope calls 'functionCalls'
}Our context blocks divide into two groups - those that can be logically chained (or AND-ed) together, and those that cannot.
A transition responds to a single event and transitions to a single state. Therefore multiple when { } and then { } statements cannot be AND-ed together.
define(.locked) {
when(.coin) {
when(.pass) { } // ⛔️ does not compile
when(.pass) | ... // ⛔️ does not compile
matching(.something) | when(.pass) | ... // ⛔️ does not compile
matching(.something) {
when(.pass) { } // ⛔️ does not compile
when(.pass) | ... // ⛔️ does not compile
}
}
then(.unlocked) {
then(.locked) { } // ⛔️ does not compile
then(.locked) | ... // ⛔️ does not compile
matching(.something) | then(.locked) | ... // ⛔️ does not compile
matching(.something) {
then(.locked) { } // ⛔️ does not compile
then(.locked) | ... // ⛔️ does not compile
}
}
}Additionally, there is a specific combination of when { } and then that does not compile, as there is no situation where, in response to a single event (in this case, .coin), there could then be a transition to more than one state, unless a different Predicate is given for each.
define(.locked) {
when(.coin) {
then(.unlocked) | action // ⛔️ does not compile
then(.locked) | action // ⛔️ does not compile
}
}
define(.locked) {
when(.coin) {
matching(Enforcement.weak) | then(.unlocked) | action // ✅
matching(Enforcement.strong) | then(.locked) | otherAction // ✅
}
}There is no logical restriction on the number of predicates or actions per transition, and therefore both can be built up in a chain as follows:
define(.locked) {
matching(Enforcement.weak) {
matching(Reward.positive) { } // ✅ matches Enforcement.weak AND Reward.positive
matching(Reward.positive) | ... // ✅ matches Enforcement.weak AND Reward.positive
}
actions(doSomething) {
actions(doSomethingElse) { } // ✅ calls doSomething and doSomethingElse
... | doSomethingElse // ✅ calls doSomething and doSomethingElse
}
}Nested actions blocks sum the actions and perform all of them. Nested matching blocks are AND-ed together.
Pipes can and must be used inside blocks, whereas blocks cannot be opened after pipes.
define(.locked) {
when(.coin) | then(.unlocked) { } // ⛔️ does not compile
when(.coin) | then(.unlocked) | actions(doSomething) { } // ⛔️ does not compile
matching(.something) | when(.coin) { } // ⛔️ does not compile
}Using Predicates with matching syntax is a versatile solution, however in some cases it may bring more complexity than is necessary to solve a given problem (see Predicate Performance for a description of matching overhead).
If you need to make a specific transition conditional at runtime, then the condition statement may suffice. Some FSM implementations call this a guard statement, however the name condition was chosen here as guard is a reserved word in Swift.
define(.locked) {
condition(complexDecisionTree) | when(.pass) | then(.locked) | lock
}Here, complexDecisionTree() is a function that returns a Bool. If it is true, the transition is executed, and if it is not, nothing is executed.
The keyword condition is syntactically interchangeable with matching - it works with pipe and block syntax, and is chainable (conditions are AND-ed together).
matching and condition blocks can also be combined freely:
define(.locked) {
condition({ reward == .positive }) {
matching(Enforcement.weak) | then(.unlocked) | action
matching(Enforcement.strong) | then(.locked) | otherAction
}
}The disadvantage of condition versus matching is that it is more limited in the logic it can express:
define(.locked) {
when(.coin) {
matching(Enforcement.weak) | then(.unlocked) | action
matching(Enforcement.strong) | then(.locked) | otherAction
}
} // ✅ all good here
...
define(.locked) {
when(.coin) {
condition { enforcement == .weak } | then(.unlocked) | action
condition { enforcement == .strong } | then(.locked) | otherAction
}
} // 💥 error: logical clashThere is no way to distinguish different condition statements, as the () -> Bool blocks are inherently opaque. From a logical evaluation point of view, they are invisible. What therefore remains is two statements define(.locked) { when(.coin) | ... } that both transition to different states - the FSM has no way to understand which one to call, and must therefore throw.
In order to preserve performance, fsm.handleEvent(event:predicates:) performs minimal error handling. Therefore, passing in Predicate instances that do not appear anywhere in the transition table will not error. Nonetheless, the FSM will be unable to perform any transitions, as it will not contain any statements that match the given, unexpected Predicate instance. It is the caller’s responsibility to ensure that the predicates passed to handleEvent and the predicates used in the transition table are of the same type and number.
try fsm.buildTable { } does perform error handling to make sure the table is syntactically and semantically valid. In particular, it ensures that all matching statements are valid, and that there are no duplicate transitions and no logical clashes between transitions.
In addition to the runtime errors thrown by the basic syntax, the expanded syntax also throws the following errors:
There are two ways one might inadvertently create an invalid matching statement. The first is within a single statement:
matching(A.a, and: A.b) // 💥 error: cannot match A.a AND A.b simultaneously
matching(A.a, or: B.a, and: A.b) // 💥 error: cannot match A.a AND A.b simultaneously
matching(A.a, and: A.a) // 💥 error: duplicate predicate
matching(A.a, or: A.a) // 💥 error: duplicate predicate
matching(A.x, or: B.x)... // ⛔️ does not compile: OR types must be the same
matching(A.x, and: A.y)... // 💥 error: cannot match A.x AND A.y simultaneouslyThe second is when AND-ing multiple matching statements through the use of blocks:
matching(A.a, and: B.a) { // ✅
matching(A.a) // 💥 error: duplicate predicate
matching(A.b) // 💥 error: cannot match A.a AND A.b simultaneously
}
matching(A.a, or: A.b) { // ✅
matching(A.a) // 💥 error: duplicate predicate
matching(A.b) // 💥 error: duplicate predicate
}See Implicit Clashes
Overview: operations per function call for a table with 100 transitions, 3 Predicate types, and 10 cases per Predicate
| FSM | LazyFSM | Schedule | |
|---|---|---|---|
handleEvent |
1 | 1-7 | Every transition |
buildTable |
100,000 | 100 | Once on app load |
Adding predicates has no effect on the performance of handleEvent(), but does affect the performance of fsm.buildTransitions { }. By default, the FSM preserves handleEvent() runtime performance by doing significant work ahead of time when creating the transition table, filling in missing transitions for all implied Predicate combinations.
The performance of fsm.buildTransitions { } is dominated by this, assuming any predicates are used at all. Because all possible combinations of cases of all given predicates have to be calculated and filtered for each transition, performance is O(m^n*o) where m is the average number of cases per predicate, n is number ofPredicate types and o is the number of transitions.
Using threePredicate types with 10 cases each in a table with 100 transitions would therefore require 100,000 operations to compile. In most real-world use cases, this is unlikely to be a problem. If your table is particularly large (see overview above), Swift FSM provides a more performance-balanced alternative for such cases in the form of the LazyFSM class.
Note: there is no performance advantage to using the keyword matching in fewer versus more transitions in your transition tables. Once the word matching is used once, and a Predicate instance is passed to handleEvent(), the performance implications for the whole table will be as above.
Passing the .lazy argument to FSM<State, Event>(type: .lazy) does away with the look-ahead combinatorics algorithm described above. The result is smaller tables internally, and faster table compile time. The cost is at the call to handleEvent() where multiple lookup operations are now needed to find the correct transition.
Performance ofhandleEvent() decreases from O(1) to O(n!), where n is the number of Predicate types used regardless of the number of cases. Inversely, performance of buildTable { } increases from O(m^n*o) to O(n), where n is now the number of transitions.
Using three Predicate types with 10 cases each in a table with 100 transitions would now require 100 operations to compile (down from 100,000 by a factor of 1000). Each call to handleEvent() would need to perform between 1 and 3! + 1 or 7 operations (up from 1 by a factor of 1-7). Using more than three Predicate types in this case is therefore not advisable.
Overall LazyFSM saves a lot more operations than it costs, but that cost is paid at runtime for each transition, rather than as a one-off cost at compile time. In most cases, FSM is likely to be the preferred solution, with LazyFSM reserved for especially large numbers of transitions and/or Predicate cases. If no predicates are used, both implementations exhibit similarly fast performance.
Though Swift FSM runtime errors contain verbose descriptions of the problem, nothing can be done to help with disambiguating compile time errors.
First, familiarity with how @resultBuilder works, and the kinds of compile time errors it tends to generate will be very helpful in understanding the errors you may encounter. Almost all Swift FSM-specific compile time errors will be produced by unrecognised arguments to the aforementioned @resultBuilder, and unrecognised arguments to the | operator overloaded by Swift FSM.
To help, here is a brief list of common errors you are likely to encounter if you try to build something that Swift FSM disallows at compile time:
No exact matches in call to static method 'buildExpression’
This is a common compile time error in @resultBuilder blocks. It will occur if you feed the block an argument that it does not support. It is useful to remember that each line in such a block is actually an argument fed to a static method.
For example:
try fsm.buildTable {
actions(thankyou) { }
// ⛔️ No exact matches in call to static method 'buildExpression'
}Here an actions block is given as an argument to the hidden static function buildExpression on the @resultBuilder supporting the buildTable function. The define statement has been skipped, and actions returns a type not supported by this outer block, and therefore cannot compile.
Cannot convert value of type <T1> to expected argument type <T2>
This is common in situations where an unsupported argument is passed to a pipe overload.
For example:
try fsm.buildTable {
define(.locked) {
then(.locked) | unlock
// ⛔️ Cannot convert value of type 'Syntax.Then<TurnstileState>' to expected argument type 'Internal.MatchingWhenThen'
// ⛔️ No exact matches in call to static method 'buildExpression'
}
}Here no matching and/or when statement precede/s the call to then(.locked). There is no | overload that takes as its two arguments the output of then(.locked) on the left, and the block () -> () on the right, and therefore cannot compile.
The error unfortunately spits out some internal implementation details that cannot be hidden. Such unavoidable details are marked as such by their location inside the Internal namespace.
It also produces a secondary error - as it cannot work out what the output of then(.locked) | unlock is, it declares that there is no overload available for buildExpression. Fix the underlying | error and this error will also disappear.
Referencing operator function '|' on 'SIMD' requires that 'Syntax.When<TurnstileEvent>' conform to 'SIMD’
A personal favourite, from this:
try fsm.buildTable {
define(.locked) {
when(.coin) | matching(P.a) | then(.locked) | unlock
// ⛔️ Referencing operator function '|' on 'SIMD' requires that 'Syntax.When<TurnstileEvent>' conform to 'SIMD’
}
}The order of when and matching is inverted and not supported. This no different to the previous error, but the compiler interprets the problem differently. It selects a | overload from an unrelated module and declares that it is being misused.
The compiler cannot help identify which pipe in the chain is causing the problem. Often it’s simpler just to delete and rewrite the statement rather than trying to figure out what the complaint is.
try fsm.buildTable {
let resetable = SuperState {
when(.reset) | then(.locked)
}
define(.locked, adopts: resetable, onEntry: [lock]) {
when(.coin) | then(.unlocked)
when(.pass) | then(.alarming)
}
define(.unlocked, adopts: resetable, onEntry: [unlock]) {
when(.coin) | then(.unlocked) | thankyou
when(.pass) | then(.locked)
}
define(.alarming, adopts: resetable, onEntry: [alarmOn], onExit: [🦤])
}This is the original example from Entry and Exit Actions, with one small error inserted at the end. This may or may not produce an appropriate error next to the dodo:
Cannot find '🦤' in scope
What it will also do is generate multiple spurious errors and fixits in the SuperState declaration similar to this one:
Call to method ‘then’ in closure requires explicit use of ‘self’ to make capture semantics explicit
Reference ‘self.’ explicitly [ Fix ]
Capture 'self' explicitly to enable implicit 'self' in this closure
Ignore these errors, and if there is no other error shown, you may have to hunt about for the unrecognised argument.