ProgramBuilder.swift

// Copyright 2019 Google LLC
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// https://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.

/// Builds programs.
///
/// This provides methods for constructing and appending random
/// instances of the different kinds of operations in a program.
public class ProgramBuilder {
    /// The fuzzer instance for which this builder is active.
    public let fuzzer: Fuzzer

    private let logger = Logger(withLabel: "ProgramBuilder")

    /// The code and type information of the program that is being constructed.
    private var code = Code()

    /// Comments for the program that is being constructed.
    private var comments = ProgramComments()

    /// Every code generator that contributed to the current program.
    private var contributors = Contributors()

    /// The parent program for the program being constructed.
    private let parent: Program?

    public var context: Context {
        return contextAnalyzer.context
    }

    /// If true, the variables containing a function is hidden inside the function's body.
    ///
    /// For example, in
    ///
    ///     let f = b.buildPlainFunction(with: .parameters(n: 2) { args in
    ///         // ...
    ///     }
    ///     b.callFunction(f, withArgs: b.randomArguments(forCalling: f))
    ///
    /// The variable f would *not* be visible inside the body of the plain function during building
    /// when this is enabled. However, the variable will be visible during future mutations, it is only
    /// hidden when the function is initially created.
    ///
    /// The same is done for class definitions, which may also cause trivial recursion in the constructor,
    /// but where usage of the output inside the class definition's body may also cause other problems,
    /// for example since `class C { [C] = 42; }` is invalid.
    ///
    /// This can make sense for a number of reasons. First, it prevents trivial recursion where a
    /// function directly calls itself. Second, it prevents weird code like for example the following:
    ///
    ///     function f1() {
    ///         let o6 = { x: foo + foo, y() { return foo; } };
    ///     }
    ///
    /// From being generated, which can happen quite frequently during prefix generation as
    /// the number of visible variables may be quite small.
    public let enableRecursionGuard = true

    /// Counter to quickly determine the next free variable.
    private var numVariables = 0

    /// Context analyzer to keep track of the currently active IL context.
    private var contextAnalyzer = ContextAnalyzer()

    /// Visible variables management.
    /// The `scopes` stack contains one entry per currently open scope containing all variables created in that scope.
    private var scopes = Stack<[Variable]>([[]])
    /// The `variablesInScope` array simply contains all variables that are currently in scope. It is effectively the `scopes` stack flattened.
    private var variablesInScope = [Variable]()

    /// Keeps track of variables that have explicitly been hidden and so should not be
    /// returned from e.g. `randomVariable()`. See `hide()` for more details.
    private var hiddenVariables = VariableSet()
    private var numberOfHiddenVariables = 0

    /// Type inference for JavaScript variables.
    private var jsTyper: JSTyper

    /// Argument generation budget.
    /// This budget is used in `findOrGenerateArguments(forSignature)` and tracks the upper limit of variables that that function should emit.
    /// If that upper limit is reached the function will stop generating new variables and use existing ones instead.
    /// If this value is set to nil, there is no argument generation happening, every argument generation should enter the recursive function (findOrGenerateArgumentsInternal) through the public non-internal one.
    private var argumentGenerationVariableBudget: Int? = nil
    /// This is the top most signature that was requested when `findOrGeneratorArguments(forSignature)` was called, this is helpful for debugging.
    private var argumentGenerationSignature: Signature? = nil

    /// Stack of active object literals.
    ///
    /// This needs to be a stack as object literals can be nested, for example if an object
    /// literals is created inside a method/getter/setter of another object literals.
    private var activeObjectLiterals = Stack<ObjectLiteral>()

    /// When building object literals, the state for the current literal is exposed through this member and
    /// can be used to add fields to the literal or to determine if some field already exists.
    public var currentObjectLiteral: ObjectLiteral {
        return activeObjectLiterals.top
    }

    /// Stack of active class definitions.
    ///
    /// Similar to object literals, class definitions can be nested so this needs to be a stack.
    private var activeClassDefinitions = Stack<ClassDefinition>()

    /// When building class definitions, the state for the current definition is exposed through this member and
    /// can be used to add fields to the class or to determine if some field already exists.
    public var currentClassDefinition: ClassDefinition {
        return activeClassDefinitions.top
    }

    /// Stack of active switch blocks.
    private var activeSwitchBlocks = Stack<SwitchBlock>()

    /// When building switch blocks, the state for the current switch block is exposed through this
    /// member and can be used to add cases to the switch.
    public var currentSwitchBlock: SwitchBlock {
        return activeSwitchBlocks.top
    }

    /// How many variables are currently in scope.
    public var numberOfVisibleVariables: Int {
        assert(numberOfHiddenVariables <= variablesInScope.count)
        return variablesInScope.count - numberOfHiddenVariables
    }

    /// Whether there are any variables currently in scope.
    public var hasVisibleVariables: Bool {
        return numberOfVisibleVariables > 0
    }

    /// All currently visible variables.
    public var visibleVariables: [Variable] {
        if numberOfHiddenVariables == 0 {
            // Fast path for the common case.
            return variablesInScope
        } else {
            return variablesInScope.filter({ !hiddenVariables.contains($0) })
        }
    }

    /// Constructs a new program builder for the given fuzzer.
    init(for fuzzer: Fuzzer, parent: Program?) {
        self.fuzzer = fuzzer
        self.jsTyper = JSTyper(for: fuzzer.environment)
        self.parent = parent
    }

    /// Resets this builder.
    public func reset() {
        code.removeAll()
        comments.removeAll()
        contributors.removeAll()
        numVariables = 0
        scopes = Stack([[]])
        variablesInScope.removeAll()
        hiddenVariables.removeAll()
        numberOfHiddenVariables = 0
        contextAnalyzer = ContextAnalyzer()
        jsTyper.reset()
        activeObjectLiterals.removeAll()
        activeClassDefinitions.removeAll()
    }

    /// Finalizes and returns the constructed program, then resets this builder so it can be reused for building another program.
    public func finalize() -> Program {
        let program = Program(code: code, parent: parent, comments: comments, contributors: contributors)
        reset()
        return program
    }

    /// Prints the current program as FuzzIL code to stdout. Useful for debugging.
    public func dumpCurrentProgram() {
        print(FuzzILLifter().lift(code))
    }

    /// Returns the current number of instructions of the program we're building.
    public var currentNumberOfInstructions: Int {
        return code.count
    }

    /// Returns the index of the next instruction added to the program. This is equal to the current size of the program.
    public func indexOfNextInstruction() -> Int {
        return currentNumberOfInstructions
    }

    /// Returns the most recently added instruction.
    public func lastInstruction() -> Instruction {
        assert(currentNumberOfInstructions > 0)
        return code.lastInstruction
    }

    /// Add a trace comment to the currently generated program at the current position.
    /// This is only done if inspection is enabled.
    public func trace(_ commentGenerator: @autoclosure () -> String) {
        if fuzzer.config.enableInspection {
            // Use an autoclosure here so that template strings are only evaluated when they are needed.
            comments.add(commentGenerator(), at: .instruction(code.count))
        }
    }

    /// Add a trace comment at the start of the currently generated program.
    /// This is only done if history inspection is enabled.
    public func traceHeader(_ commentGenerator: @autoclosure () -> String) {
        if fuzzer.config.enableInspection {
            comments.add(commentGenerator(), at: .header)
        }
    }

    ///
    /// Methods to obtain random values to use in a FuzzIL program.
    ///

    /// Returns a random integer value.
    public func randomInt() -> Int64 {
        if probability(0.5) {
            return chooseUniform(from: self.fuzzer.environment.interestingIntegers)
        } else {
            return withEqualProbability({
                Int64.random(in: -0x10...0x10)
            }, {
                Int64.random(in: -0x10000...0x10000)
            }, {
                Int64.random(in: Int64(Int32.min)...Int64(Int32.max))
            })
        }
    }

    /// Returns a random integer value suitable as size of for example an array.
    /// The returned value is guaranteed to be positive.
    public func randomSize(upTo maximum: Int64 = 0x100000000) -> Int64 {
        assert(maximum >= 0x1000)
        if probability(0.5) {
            return chooseUniform(from: fuzzer.environment.interestingIntegers.filter({ $0 >= 0 && $0 <= maximum }))
        } else {
            return withEqualProbability({
                Int64.random(in: 0...0x10)
            }, {
                Int64.random(in: 0...0x100)
            }, {
                Int64.random(in: 0...0x1000)
            }, {
                Int64.random(in: 0...maximum)
            })
        }
    }

    /// Returns a random integer value suitable as index.
    public func randomIndex() -> Int64 {
        // Prefer small, (usually) positive, indices.
        if probability(0.33) {
            return Int64.random(in: -2...10)
        } else {
            return randomSize()
        }
    }

    /// Returns a random floating point value.
    public func randomFloat() -> Double {
        if probability(0.5) {
            return chooseUniform(from: fuzzer.environment.interestingFloats)
        } else {
            return withEqualProbability({
                Double.random(in: 0.0...1.0)
            }, {
                Double.random(in: -10.0...10.0)
            }, {
                Double.random(in: -1000.0...1000.0)
            }, {
                Double.random(in: -1000000.0...1000000.0)
            }, {
                // We cannot do Double.random(in: -Double.greatestFiniteMagnitude...Double.greatestFiniteMagnitude) here,
                // presumably because that range is larger than what doubles can represent? So split the range in two.
                if probability(0.5) {
                    return Double.random(in: -Double.greatestFiniteMagnitude...0)
                } else {
                    return Double.random(in: 0...Double.greatestFiniteMagnitude)
                }
            })
        }
    }

    /// Returns a random string value.
    public func randomString() -> String {
        return withEqualProbability({
            self.randomPropertyName()
        }, {
            self.randomMethodName()
        }, {
            chooseUniform(from: self.fuzzer.environment.interestingStrings)
        }, {
            String(self.randomInt())
        }, {
            String.random(ofLength: Int.random(in: 1...5))
        })
    }

    func randomRegExpPattern(compatibleWithFlags flags: RegExpFlags) -> String {
        // Generate a "base" regexp
        var regex = ""
        let desiredLength = Int.random(in: 1...4)
        while regex.count < desiredLength {
            regex += withEqualProbability({
                String.random(ofLength: 1)
            }, {
                // Pick from the available RegExp pattern, based on flags.
                let candidates = self.fuzzer.environment.interestingRegExps.filter({ pattern, incompatibleFlags in flags.isDisjoint(with: incompatibleFlags) })
                return chooseUniform(from: candidates).pattern
            })
        }

        // Now optionally concatenate with another regexp
        if probability(0.3) {
            regex += randomRegExpPattern(compatibleWithFlags: flags)
        }

        // Or add a quantifier, if there is not already a quantifier in the last position.
        if probability(0.2) && !self.fuzzer.environment.interestingRegExpQuantifiers.contains(String(regex.last!)) {
            regex += chooseUniform(from: self.fuzzer.environment.interestingRegExpQuantifiers)
        }

        // Or wrap in brackets
        if probability(0.1) {
            withEqualProbability({
                // optionally invert the character set
                if probability(0.2) {
                    regex = "^" + regex
                }
                regex = "[" + regex + "]"
            }, {
                regex = "(" + regex + ")"
            })
        }
        return regex
    }

    /// Returns a random regular expression pattern.
    public func randomRegExpPatternAndFlags() -> (String, RegExpFlags) {
        let flags = RegExpFlags.random()
        return (randomRegExpPattern(compatibleWithFlags: flags), flags)
    }

    /// Returns the name of a random builtin.
    public func randomBuiltin() -> String {
        return chooseUniform(from: fuzzer.environment.builtins)
    }

    /// Returns a random builtin property name.
    ///
    /// This will return a random name from the environment's list of builtin property names,
    /// i.e. a property that exists on (at least) one builtin object type.
    func randomBuiltinPropertyName() -> String {
        return chooseUniform(from: fuzzer.environment.builtinProperties)
    }

    /// Returns a random custom property name.
    ///
    /// This will select a random property from a (usually relatively small) set of custom property names defined by the environment.
    ///
    /// This should generally be used in one of two situations:
    ///   1. If a new property is added to an object.
    ///     In that case, we prefer to add properties with custom names (e.g. ".a", ".b") instead of properties
    ///     with names that exist in the environment (e.g. ".length", ".prototype"). This way, in the resulting code
    ///     it will be fairly clear when a builtin property is accessed vs. a custom one. It also increases the chances
    ///     of selecting an existing property when choosing a random property to access, see the next point.
    ///   2. If we have no static type information about the object we're accessing.
    ///     In that case there is a higher chance of success when using the small set of custom property names
    ///     instead of the much larger set of all property names that exist in the environment (or something else).
    public func randomCustomPropertyName() -> String {
        return chooseUniform(from: fuzzer.environment.customProperties)
    }

    /// Returns either a builtin or a custom property name, with equal probability.
    public func randomPropertyName() -> String {
        return probability(0.5) ? randomBuiltinPropertyName() : randomCustomPropertyName()
    }

    /// Returns a random builtin method name.
    ///
    /// This will return a random name from the environment's list of builtin method names,
    /// i.e. a method that exists on (at least) one builtin object type.
    public func randomBuiltinMethodName() -> String {
        return chooseUniform(from: fuzzer.environment.builtinMethods)
    }

    /// Returns a random custom method name.
    ///
    /// This will select a random method from a (usually relatively small) set of custom method names defined by the environment.
    ///
    /// See the comment for randomCustomPropertyName() for when this should be used.
    public func randomCustomMethodName() -> String {
        return chooseUniform(from: fuzzer.environment.customMethods)
    }

    /// Returns either a builtin or a custom method name, with equal probability.
    public func randomMethodName() -> String {
        return probability(0.5) ? randomBuiltinMethodName() : randomCustomMethodName()
    }

    // Settings and constants controlling the behavior of randomParameters() below.
    // This determines how many variables of a given type need to be visible before
    // that type is considered a candidate for a parameter type. For example, if this
    // is three, then we need at least three visible .integer variables before creating
    // parameters of type .integer.
    private let thresholdForUseAsParameter = 3

    // The probability of using .anything as parameter type even though we have more specific alternatives.
    // Doing this sometimes is probably beneficial so that completely random values are passed to the function.
    // Future mutations, such as the ExplorationMutator can then figure out what to do with the parameters.
    // Writable so it can be changed for tests.
    var probabilityOfUsingAnythingAsParameterTypeIfAvoidable = 0.20

    // Generate random parameters for a subroutine.
    //
    // This will attempt to find a parameter types for which at least a few variables of a compatible types are
    // currently available to (potentially) later be used as arguments for calling the generated subroutine.
    public func randomParameters(n wantedNumberOfParameters: Int? = nil) -> SubroutineDescriptor {
        assert(probabilityOfUsingAnythingAsParameterTypeIfAvoidable >= 0 && probabilityOfUsingAnythingAsParameterTypeIfAvoidable <= 1)

        // If the caller didn't specify how many parameters to generated, find an appropriate
        // number of parameters based on how many variables are currently visible (and can
        // therefore later be used as arguments for calling the new function).
        let n: Int
        if let requestedN = wantedNumberOfParameters {
            assert(requestedN > 0)
            n = requestedN
        } else {
            switch numberOfVisibleVariables {
            case 0...1:
                n = 0
            case 2...5:
                n = Int.random(in: 1...2)
            default:
                n = Int.random(in: 2...4)
            }
        }

        // Find all types of which we currently have at least a few visible variables that we could later use as arguments.
        // TODO: improve this code by using some kind of cache? That could then also be used for randomVariable(ofType:) etc.
        var availableVariablesByType = [ILType: Int]()
        for v in visibleVariables {
            let t = type(of: v)
            // TODO: should we also add this values to the buckets for supertypes (without this becoming O(n^2))?
            // TODO: alternatively just check for some common union types, e.g. .number, .primitive, as long as these can be used meaningfully?
            availableVariablesByType[t] = (availableVariablesByType[t] ?? 0) + 1
        }

        var candidates = Array(availableVariablesByType.filter({ k, v in v >= thresholdForUseAsParameter }).keys)
        if candidates.isEmpty {
            candidates.append(.anything)
        }

        var params = ParameterList()
        for _ in 0..<n {
            if probability(probabilityOfUsingAnythingAsParameterTypeIfAvoidable) {
                params.append(.anything)
            } else {
                params.append(.plain(chooseUniform(from: candidates)))
            }
        }

        // TODO: also generate rest parameters and maybe even optional ones sometimes?

        return .parameters(params)
    }

    public func findOrGenerateArguments(forSignature signature: Signature, maxNumberOfVariablesToGenerate: Int = 100) -> [Variable] {
        assert(context.contains(.javascript))

        argumentGenerationVariableBudget = numVariables + maxNumberOfVariablesToGenerate
        argumentGenerationSignature = signature

        defer {
            argumentGenerationVariableBudget = nil
            argumentGenerationSignature = nil
        }

        return findOrGenerateArgumentsInternal(forSignature: signature)
    }

    private func findOrGenerateArgumentsInternal(forSignature: Signature) -> [Variable] {

        var args: [Variable] = []

        // This should be called whenever we have a type that has known information about its properties but we don't have a constructor for it.
        // This can be the case for configuration objects, e.g. objects that can be passed into DOMAPIs.
        func createObjectWithProperties(_ type: ILType) -> Variable  {
            assert(type.MayBe(.object()))

            // Before we do any generation below, let's take into account that we already create a variable with this invocation, i.e. the createObject at the end.
            // Therefore we need to decrease the budget here temporarily.
            self.argumentGenerationVariableBudget! -= 1
            // We defer the increase again, because at that point the variable is actually visible, i.e. `numVariables` was increased through the `createObject` call.
            defer { self.argumentGenerationVariableBudget! += 1 }

            var properties: [String: Variable] = [:]

            for propertyName in type.properties {
                // If we have an object that has a group, we should get a type here, otherwise if we don't have a group, we will get .anything.
                let propType = fuzzer.environment.type(ofProperty: propertyName, on: type)
                properties[propertyName] = generateType(propType)
            }

            return createObject(with: properties)
        }

        func createObjectWithGroup(type: ILType) -> Variable {
            let group = type.group!

            // We can be sure that we have such a builtin with a signature because the Environment checks this during initialization.
            let signature = fuzzer.environment.type(ofBuiltin: group).signature!
            let constructor = createNamedVariable(forBuiltin: group)
            let arguments = findOrGenerateArgumentsInternal(forSignature: signature)
            let constructed = construct(constructor, withArgs: arguments)

            return constructed
        }

        func generateType(_ type: ILType) -> Variable {
            if probability(0.5) {
                if let existingVariable = randomVariable(ofType: type) {
                    return existingVariable
                }
            }

            if numVariables >= argumentGenerationVariableBudget! {
                logger.warning("Reached variable generation limit in generateType for Signature: \(argumentGenerationSignature!), returning a random variable for use as type \(type).")
                return randomVariable(forUseAs: type)
            }

            // We only need to check against all base types from TypeSystem.swift, this works because we use .MayBe
            // TODO: Not sure how we should handle merge types, e.g. .string + .object(...).
            let typeGenerators: [(ILType, () -> Variable)] = [
                (.integer, { return self.loadInt(self.randomInt()) }),
                (.string, { return self.loadString(self.randomString()) }),
                (.boolean, { return self.loadBool(probability(0.5)) }),
                (.bigint, { return self.loadBigInt(self.randomInt()) }),
                (.float, { return self.loadFloat(self.randomFloat()) }),
                (.regexp, {
                     let (pattern, flags) = self.randomRegExpPatternAndFlags()
                     return self.loadRegExp(pattern, flags)
                 }),
                (.iterable, { return self.createArray(with: self.randomVariables(upTo: 5)) }),
                (.function(), {
                     // TODO: We could technically generate a full function here but then we would enter the full code generation logic which could do anything.
                     // Because we want to avoid this, we will just pick anything that can be a function.
                     return self.randomVariable(forUseAs: .function())
                 }),
                (.undefined, { return self.loadUndefined() }),
                (.constructor(), {
                     // TODO: We have the same issue as above for functions.
                     return self.randomVariable(forUseAs: .constructor())
                 }),
                (.object(), {
                     // If we have a group on this object and we have a builtin, that means we can construct it with new.
                     if let group = type.group, self.fuzzer.environment.hasBuiltin(group) && self.fuzzer.environment.type(ofBuiltin: group).Is(.constructor()) {
                         return createObjectWithGroup(type: type)
                     } else {
                         // Otherwise this is one of the following:
                         // 1. an object with more type information, i.e. it has a group, but no associated builtin, e.g. we cannot construct it with new.
                         // 2. an object without a group, but it has some required fields.
                         // In either case, we try to construct such an object.
                         return createObjectWithProperties(type)
                     }
                 })
            ]

            // Make sure that we walk over these tests and their generators randomly.
            // The requested type could be a Union of other types and as such we want to randomly generate one of them, therefore we also use the MayBe test below.
            for (t, generate) in typeGenerators.shuffled() {
                if type.MayBe(t) {
                    let variable = generate()
                    return variable
                }
            }

            logger.warning("Type \(type) was not handled, returning random variable.")
            return randomVariable(forUseAs: type)
        }

        outer: for parameter in forSignature.parameters {
            switch parameter {
            case .plain(let t):
                args.append(generateType(t))
            case .opt(let t):
                if probability(0.5) {
                    args.append(generateType(t))
                } else {
                    // We decided to not provide an optional parameter, so we can stop here.
                    break outer
                }
            case .rest(let t):
                for _ in 0...Int.random(in: 1...3) {
                    args.append(generateType(t))
                }
            }
        }

        return args
    }

    ///
    /// Access to variables.
    ///

    /// Returns a random variable.
    public func randomVariable() -> Variable {
        assert(hasVisibleVariables)
        return findVariable()!
    }

    /// Returns up to N (different) random variables.
    /// This method will only return fewer than N variables if the number of currently visible variables is less than N.
    public func randomVariables(upTo n: Int) -> [Variable] {
        guard hasVisibleVariables else { return [] }

        var variables = [Variable]()
        while variables.count < n {
            guard let newVar = findVariable(satisfying: { !variables.contains($0) }) else {
                break
            }
            variables.append(newVar)
        }
        return variables
    }

    /// Returns up to N potentially duplicate random variables.
    public func randomVariables(n: Int) -> [Variable] {
        assert(hasVisibleVariables)
        return (0..<n).map { _ in randomVariable() }
    }

    /// This probability affects the behavior of `randomVariable(forUseAs:)`. In particular, it determines how much variables with
    /// a known-to-be-matching type will be preferred over variables with a more general, or even unknown type. For example, if this is
    /// 0.5, then 50% of the time we'll first try to find an exact match (`type(of: result).Is(requestedType)`) before trying the
    /// more general search (`type(of: result).MayBe(requestedType)`) which also includes variables of unknown type.
    /// This is writable for use in tests, but it could also be used to change how "conservative" variable selection is.
    var probabilityOfVariableSelectionTryingToFindAnExactMatch = 0.5

    /// This threshold  affects the behavior of `randomVariable(forUseAs:)`. It determines how many existing variables of the
    /// requested type we want to have before we try to find an exact match. If there are fewer variables of the requested type, we'll
    /// always do a more general search which may also return variables of unknown (i.e. `.anything`) type.
    /// This ensures that consecutive queries for the same type can return different variables.
    let minVisibleVariablesOfRequestedTypeForVariableSelection = 3

    /// Returns a random variable to be used as the given type.
    ///
    /// This function may return variables of a different type, or variables that may have the requested type, but could also have a different type.
    /// For example, when requesting a .integer, this function may also return a variable of type .number, .primitive, or even .anything as all of these
    /// types may be an integer (but aren't guaranteed to be). In this way, this function ensures that variables for which no exact type could be statically
    /// determined will also be used as inputs for following code.
    ///
    /// It's the caller's responsibility to check the type of the returned variable to avoid runtime exceptions if necessary. For example, if performing a
    /// property access, the returned variable should be checked if it `MayBe(.nullish)` in which case a property access would result in a
    /// runtime exception and so should be appropriately guarded against that.
    ///
    /// If the variable must be of the specified type, use `randomVariable(ofType:)` instead.
    public func randomVariable(forUseAs type: ILType) -> Variable {
        assert(type != .nothing)

        var result: Variable? = nil

        // Prefer variables that are known to have the requested type if there's a sufficient number of them.
        if probability(probabilityOfVariableSelectionTryingToFindAnExactMatch) &&
            haveAtLeastNVisibleVariables(ofType: type, n: minVisibleVariablesOfRequestedTypeForVariableSelection) {
            result = findVariable(satisfying: { self.type(of: $0).Is(type) })
        }

        // Otherwise, select variables that may have the desired type, but could also be something else.
        // In particular, this query will include all variable for which we don't know the type as they'll
        // be typed as .anything. We usually expect to have a lot of candidates available for this query,
        // so we don't check the number of them upfront as we do for the above query.
        if result == nil {
            result = findVariable(satisfying: { self.type(of: $0).MayBe(type) })
        }

        // Worst case fall back to completely random variables. This should happen rarely, as we'll usually have
        // at least some variables of type .anything.
        return result ?? randomVariable()
    }

    /// Returns a random variable that is known to have the given type.
    ///
    /// This will return a variable for which `b.type(of: v).Is(type)` is true, i.e. for which our type inference
    /// could prove that it will have the specified type. If no such variable is found, this function returns nil.
    public func randomVariable(ofType type: ILType) -> Variable? {
        assert(type != .nothing)
        return findVariable(satisfying: { self.type(of: $0).Is(type) })
    }

    /// Returns a random variable that is not known to have the given type.
    ///
    /// This will return a variable for which `b.type(of: v).Is(type)` is false, i.e. for which our type inference
    /// could not prove that it has the given type. Note that this is different from a variable that is known not to have
    /// the given type: this function can return variables for which `b.type(of: v).MayBe(type)` is true.
    /// If no such variable is found, this function returns nil.
    public func randomVariable(preferablyNotOfType type: ILType) -> Variable? {
        return findVariable(satisfying: { !self.type(of: $0).Is(type) })
    }

    /// Returns a random variable satisfying the given constraints or nil if none is found.
    func findVariable(satisfying filter: ((Variable) -> Bool) = { _ in true }) -> Variable? {
        assert(hasVisibleVariables)

        // TODO: we should implement some kind of fast lookup data structure to speed up the lookup of variables by type.
        // We have to be careful though to correctly take type changes (e.g. from reassignments) into account.

        // Also filter out any hidden variables.
        var isIncluded = filter
        if numberOfHiddenVariables != 0 {
            isIncluded = { !self.hiddenVariables.contains($0) && filter($0) }
        }

        var candidates = [Variable]()

        // Prefer the outputs of the last instruction to build longer data-flow chains.
        if probability(0.15) {
            candidates = Array(code.lastInstruction.allOutputs)
            candidates = candidates.filter(isIncluded)
        }

        // Prefer inner scopes if we're not anyway using one of the newest variables.
        let scopes = scopes
        if candidates.isEmpty && probability(0.75) {
            candidates = chooseBiased(from: scopes.elementsStartingAtBottom(), factor: 1.25)
            candidates = candidates.filter(isIncluded)
        }

        // If we haven't found any candidates yet, take all visible variables into account.
        if candidates.isEmpty {
            candidates = variablesInScope.filter(isIncluded)
        }

        if candidates.isEmpty {
            return nil
        }

        return chooseUniform(from: candidates)
    }

    /// Helper function to determine if we have a sufficient number of variables of a given type to ensure that
    /// consecutive queries for a variable of a given type will not always return the same variables.
    private func haveAtLeastNVisibleVariables(ofType t: ILType, n: Int) -> Bool {
        var count = 0
        for v in variablesInScope where !hiddenVariables.contains(v) && type(of: v).Is(t) {
            count += 1
            if count >= n {
                return true
            }
        }
        return false
    }

    /// Find random variables to use as arguments for calling the specified function.
    ///
    /// This function will attempt to find variables that are compatible with the functions parameter types (if any). However,
    /// if no matching variables can be found for a parameter, this function will fall back to using a random variable. It is
    /// then the caller's responsibility to determine whether the function call can still be performed without raising a runtime
    /// exception or if it needs to be guarded against that.
    /// In this way, functions/methods for which no matching arguments currently exist can still be called (but potentially
    /// wrapped in a try-catch), which then gives future mutations (in particular Mutators such as the ProbingMutator) the
    /// chance to find appropriate arguments for the function.
    public func randomArguments(forCalling function: Variable) -> [Variable] {
        let signature = type(of: function).signature ?? Signature.forUnknownFunction
        return randomArguments(forCallingFunctionWithSignature: signature)
    }

    /// Find random variables to use as arguments for calling the specified method.
    ///
    /// See the comment above `randomArguments(forCalling function: Variable)` for caveats.
    public func randomArguments(forCallingMethod methodName: String, on object: Variable) -> [Variable] {
        let signature = methodSignature(of: methodName, on: object)
        return randomArguments(forCallingFunctionWithSignature: signature)
    }

    /// Find random variables to use as arguments for calling the specified method.
    ///
    /// See the comment above `randomArguments(forCalling function: Variable)` for caveats.
    public func randomArguments(forCallingMethod methodName: String, on objType: ILType) -> [Variable] {
        let signature = methodSignature(of: methodName, on: objType)
        return randomArguments(forCallingFunctionWithSignature: signature)
    }

    /// Find random variables to use as arguments for calling a function with the specified signature.
    ///
    /// See the comment above `randomArguments(forCalling function: Variable)` for caveats.
    public func randomArguments(forCallingFunctionWithSignature signature: Signature) -> [Variable] {
        return randomArguments(forCallingFunctionWithParameters: signature.parameters)
    }

    /// Find random variables to use as arguments for calling a function with the given parameters.
    ///
    /// See the comment above `randomArguments(forCalling function: Variable)` for caveats.
    public func randomArguments(forCallingFunctionWithParameters params: ParameterList) -> [Variable] {
        assert(params.count == 0 || hasVisibleVariables)
        let parameterTypes = prepareArgumentTypes(forParameters: params)
        return parameterTypes.map({ randomVariable(forUseAs: $0) })
    }

    /// Find random arguments for a function call and spread some of them.
    public func randomCallArgumentsWithSpreading(n: Int) -> (arguments: [Variable], spreads: [Bool]) {
        var arguments: [Variable] = []
        var spreads: [Bool] = []
        for _ in 0...n {
            let val = randomVariable()
            arguments.append(val)
            // Prefer to spread values that we know are iterable, as non-iterable values will lead to exceptions ("TypeError: Found non-callable @@iterator")
            if type(of: val).Is(.iterable) {
                spreads.append(probability(0.9))
            } else {
                spreads.append(probability(0.1))
            }
        }

        return (arguments, spreads)
    }

    /// Hide the specified variable, preventing it from being used as input by subsequent code.
    ///
    /// Hiding a variable prevents it from being returned from `randomVariable()` and related functions, which
    /// in turn prevents it from being used as input for later instructions, unless the hidden variable is explicitly specified
    /// as input, which is still allowed.
    ///
    /// This can be useful for example if a CodeGenerator needs to create temporary values that should not be used
    /// by any following code. It is also used to prevent trivial recursion by hiding the function variable inside its body.
    public func hide(_ variable: Variable) {
        assert(!hiddenVariables.contains(variable))
        assert(visibleVariables.contains(variable))

        hiddenVariables.insert(variable)
        numberOfHiddenVariables += 1
    }

    /// Unhide the specified variable so that it can again be used as input by subsequent code.
    ///
    /// The variable must have previously been hidden using `hide(variable:)` above.
    public func unhide(_ variable: Variable) {
        assert(numberOfHiddenVariables > 0)
        assert(hiddenVariables.contains(variable))
        assert(variablesInScope.contains(variable))
        assert(!visibleVariables.contains(variable))

        hiddenVariables.remove(variable)
        numberOfHiddenVariables -= 1
    }


    /// Type information access.
    public func type(of v: Variable) -> ILType {
        return jsTyper.type(of: v)
    }

    /// Returns the type of the `super` binding at the current position.
    public func currentSuperType() -> ILType {
        return jsTyper.currentSuperType()
    }

    /// Returns the type of the super constructor.
    public func currentSuperConstructorType() -> ILType {
        return jsTyper.currentSuperConstructorType()
    }

    public func type(ofProperty property: String, on v: Variable) -> ILType {
        return jsTyper.inferPropertyType(of: property, on: v)
    }

    public func methodSignature(of methodName: String, on object: Variable) -> Signature {
        return jsTyper.inferMethodSignature(of: methodName, on: object)
    }

    public func methodSignature(of methodName: String, on objType: ILType) -> Signature {
        return jsTyper.inferMethodSignature(of: methodName, on: objType)
    }

    /// Overwrite the current type of the given variable with a new type.
    /// This can be useful if a certain code construct is guaranteed to produce a value of a specific type,
    /// but where our static type inference cannot determine that.
    public func setType(ofVariable variable: Variable, to variableType: ILType) {
        jsTyper.setType(of: variable, to: variableType)
    }

    /// This helper function converts parameter types into argument types, for example by "unrolling" rest parameters and handling optional parameters.
    private func prepareArgumentTypes(forParameters params: ParameterList) -> [ILType] {
        var argumentTypes = [ILType]()

        for param in params {
            switch param {
            case .rest(let t):
                // "Unroll" the rest parameter
                for _ in 0..<Int.random(in: 0...5) {
                    argumentTypes.append(t)
                }
            case .opt(let t):
                // It's an optional argument, so stop here in some cases
                if probability(0.25) {
                    return argumentTypes
                }
                fallthrough
            case .plain(let t):
                argumentTypes.append(t)
            }
        }

        return argumentTypes
    }


    ///
    /// Adoption of variables from a different program.
    /// Required when copying instructions between program.
    ///
    private var varMaps = [VariableMap<Variable>]()

    /// Prepare for adoption of variables from the given program.
    ///
    /// This sets up a mapping for variables from the given program to the
    /// currently constructed one to avoid collision of variable names.
    public func beginAdoption(from program: Program) {
        varMaps.append(VariableMap())
    }

    /// Finishes the most recently started adoption.
    public func endAdoption() {
        varMaps.removeLast()
    }

    /// Executes the given block after preparing for adoption from the provided program.
    public func adopting(from program: Program, _ block: () -> Void) {
        beginAdoption(from: program)
        block()
        endAdoption()
    }

    /// Maps a variable from the program that is currently configured for adoption into the program being constructed.
    public func adopt(_ variable: Variable) -> Variable {
        if !varMaps.last!.contains(variable) {
            varMaps[varMaps.count - 1][variable] = nextVariable()
        }

        return varMaps.last![variable]!
    }

    /// Maps a list of variables from the program that is currently configured for adoption into the program being constructed.
    public func adopt<Variables: Collection>(_ variables: Variables) -> [Variable] where Variables.Element == Variable {
        return variables.map(adopt)
    }

    /// Adopts an instruction from the program that is currently configured for adoption into the program being constructed.
    public func adopt(_ instr: Instruction) {
        internalAppend(Instruction(instr.op, inouts: adopt(instr.inouts), flags: instr.flags))
    }

    /// Append an instruction at the current position.
    public func append(_ instr: Instruction) {
        for v in instr.allOutputs {
            numVariables = max(v.number + 1, numVariables)
        }
        internalAppend(instr)
    }

    /// Append a program at the current position.
    ///
    /// This also renames any variable used in the given program so all variables
    /// from the appended program refer to the same values in the current program.
    public func append(_ program: Program) {
        adopting(from: program) {
            for instr in program.code {
                adopt(instr)
            }
        }
    }

    // Probabilities of remapping variables to host variables during splicing. These are writable so they can be reconfigured for testing.
    // We use different probabilities for outer and for inner outputs: while we rarely want to replace outer outputs, we frequently want to replace inner outputs
    // (e.g. function parameters) to avoid splicing function definitions that may then not be used at all. Instead, we prefer to splice only the body of such functions.
    var probabilityOfRemappingAnInstructionsOutputsDuringSplicing = 0.10
    var probabilityOfRemappingAnInstructionsInnerOutputsDuringSplicing = 0.75
    // The probability of including an instruction that may mutate a variable required by the slice (but does not itself produce a required variable).
    var probabilityOfIncludingAnInstructionThatMayMutateARequiredVariable = 0.5


    /// Splice code from the given program into the current program.
    ///
    /// Splicing computes a set of dependent (through dataflow) instructions in one program (called a "slice") and inserts it at the current position in this program.
    ///
    /// If the optional index is specified, the slice starting at that instruction is used. Otherwise, a random slice is computed.
    /// If mergeDataFlow is true, the dataflows of the two programs are potentially integrated by replacing some variables in the slice with "compatible" variables in the host program.
    /// Returns true on success (if at least one instruction has been spliced), false otherwise.
    @discardableResult
    public func splice(from program: Program, at specifiedIndex: Int? = nil, mergeDataFlow: Bool = true) -> Bool {
        // Splicing:
        //
        // Invariants:
        //  - A block is included in a slice in full (including its entire body) or not at all
        //  - An instruction can only be included if its required context is a subset of the current context
        //    OR if one or more of its surrounding blocks are included and all missing contexts are opened by them
        //  - An instruction can only be included if all its data-flow dependencies are included
        //    OR if the required variables have been remapped to existing variables in the host program
        //
        // Algorithm:
        //  1. Iterate over the program from start to end and compute for every block:
        //       - the inputs required by this block. This is the set of variables that are used as input
        //         for one or more instructions in the block's body, but are not created by instructions in the block
        //       - the context required by this block. This is the union of all contexts required by instructions
        //         in the block's body and subracting the context opened by the block itself
        //     In essence, this step allows treating every block start as a single instruction, which simplifies step 2.
        //  2. Iterate over the program from start to end and check which instructions can be inserted at the current
        //     position given the current context and the instruction's required context as well as the set of available
        //     variables and the variables required as inputs for the instruction. When deciding whether a block can be
        //     included, this will use the information computed in step 1 to treat the block as a single instruction
        //     (which it effectively is, as it will always be included in full). If an instruction can be included, its
        //     outputs are available for other instructions to use. If an instruction cannot be included, try to remap its
        //     outputs to existing and "compatible" variables in the host program so other instructions that depend on these
        //     variables can still be included. Also randomly remap some other variables to connect the dataflows of the two
        //     programs if that is enabled.
        //  3. Pick a random instruction from all instructions computed in step (2) or use the provided start index.
        //  4. Iterate over the program in reverse order and compute the slice: every instruction that creates an
        //     output needed as input for another instruction in the slice must be included as well. Step 2 guarantees that
        //     any such instruction can be part of the slice. Optionally, this step can also include instructions that may
        //     mutate variables required by the slice, for example property stores or method calls.
        //  5. Iterate over the program from start to end and add every instruction that is part of the slice into
        //     the current program, while also taking care of remapping the inouts, either to existing variables
        //     (if the variables were remapped in step (2)), or newly allocated variables.

        // Helper class to store various bits of information associated with a block.
        // This is a class so that each instruction belonging to the same block can have a reference to the same object.
        class Block {
            let startIndex: Int
            var endIndex = 0

            // Currently opened context. Updated at each block instruction.
            var currentlyOpenedContext: Context
            var requiredContext: Context

            var providedInputs = VariableSet()
            var requiredInputs = VariableSet()

            init(startedBy head: Instruction) {
                self.startIndex = head.index
                self.currentlyOpenedContext = head.op.contextOpened
                self.requiredContext = head.op.requiredContext
                self.requiredInputs.formUnion(head.inputs)
                self.providedInputs.formUnion(head.allOutputs)
            }
        }

        //
        // Step (1): compute the context- and data-flow dependencies of every block.
        //
        var blocks = [Int: Block]()

        // Helper functions for step (1).
        var activeBlocks = [Block]()
        func updateBlockDependencies(_ requiredContext: Context, _ requiredInputs: VariableSet) {
            guard let current = activeBlocks.last else { return }
            current.requiredContext.formUnion(requiredContext.subtracting(current.currentlyOpenedContext))
            current.requiredInputs.formUnion(requiredInputs.subtracting(current.providedInputs))
        }
        func updateBlockProvidedVariables(_ vars: ArraySlice<Variable>) {
            guard let current = activeBlocks.last else { return }
            current.providedInputs.formUnion(vars)
        }

        for instr in program.code {
            updateBlockDependencies(instr.op.requiredContext, VariableSet(instr.inputs))
            updateBlockProvidedVariables(instr.outputs)

            if instr.isBlockGroupStart {
                let block = Block(startedBy: instr)
                blocks[instr.index] = block
                activeBlocks.append(block)
            } else if instr.isBlockGroupEnd {
                let current = activeBlocks.removeLast()
                current.endIndex = instr.index
                blocks[instr.index] = current
                // Merge requirements into parent block (if any)
                updateBlockDependencies(current.requiredContext, current.requiredInputs)
                // If the block end instruction has any outputs, they need to be added to the surrounding block.
                updateBlockProvidedVariables(instr.outputs)
            } else if instr.isBlock {
                // We currently assume that inner block instructions cannot have outputs.
                // If they ever do, they'll need to be added to the surrounding block.
                assert(instr.numOutputs == 0)
                blocks[instr.index] = activeBlocks.last!

                // Inner block instructions change the execution context. Consider BeginWhileLoopBody as an example.
                activeBlocks.last?.currentlyOpenedContext = instr.op.contextOpened
            }

            updateBlockProvidedVariables(instr.innerOutputs)
        }

        //
        // Step (2): determine which instructions can be part of the slice and attempt to find replacement variables for the outputs of instructions that cannot be included.
        //
        // We need a typer to be able to find compatible replacement variables if we are merging the dataflows of the two programs.
        var typer = JSTyper(for: fuzzer.environment)
        // The set of variables that are available for a slice. A variable is available either because the instruction that outputs
        // it can be part of the slice or because the variable has been remapped to a host variable.
        var availableVariables = VariableSet()
        // Variables in the program that have been remapped to host variables.
        var remappedVariables = VariableMap<Variable>()
        // All instructions that can be included in the slice.
        var candidates = Set<Int>()

        // Helper functions for step (2).
        func tryRemapVariables(_ variables: ArraySlice<Variable>, of instr: Instruction) {
            guard mergeDataFlow else { return }
            guard hasVisibleVariables else { return }

            for v in variables {
                let type = typer.type(of: v)
                // For subroutines, the return type is only available once the subroutine has been fully processed.
                // Prior to that, it is assumed to be .anything. This may lead to incompatible functions being selected
                // as replacements (e.g. if the following code assumes that the return value must be of type X), but
                // is probably fine in practice.
                assert(!instr.hasOneOutput || v != instr.output || !(instr.op is BeginAnySubroutine) || (type.signature?.outputType ?? .anything) == .anything)
                // Try to find a compatible variable in the host program.
                let replacement: Variable
                if let match = randomVariable(ofType: type) {
                    replacement = match
                } else {
                    // No compatible variable found
                    continue
                }
                remappedVariables[v] = replacement
                availableVariables.insert(v)
            }
        }
        func maybeRemapVariables(_ variables: ArraySlice<Variable>, of instr: Instruction, withProbability remapProbability: Double) {
            assert(remapProbability >= 0.0 && remapProbability <= 1.0)
            if probability(remapProbability) {
                tryRemapVariables(variables, of: instr)
            }
        }
        func getRequirements(of instr: Instruction) -> (requiredContext: Context, requiredInputs: VariableSet) {
            if let state = blocks[instr.index] {
                assert(instr.isBlock)
                return (state.requiredContext, state.requiredInputs)
            } else {
                return (instr.op.requiredContext, VariableSet(instr.inputs))
            }
        }

        for instr in program.code {
            // Compute variable types to be able to find compatible replacement variables in the host program if necessary.
            typer.analyze(instr)

            // Maybe remap the outputs of this instruction to existing and "compatible" (because of their type) variables in the host program.
            maybeRemapVariables(instr.outputs, of: instr, withProbability: probabilityOfRemappingAnInstructionsOutputsDuringSplicing)
            maybeRemapVariables(instr.innerOutputs, of: instr, withProbability: probabilityOfRemappingAnInstructionsInnerOutputsDuringSplicing)

            // For the purpose of this step, blocks are treated as a single instruction with all the context and input requirements of the
            // instructions in their body. This is done through the getRequirements function which uses the data computed in step (1).
            let (requiredContext, requiredInputs) = getRequirements(of: instr)

            // TODO: We could do a further optimization here: if we've already seen a singular operation in the current block, don't splice another singular operation.
            if requiredContext.isSubset(of: context) && requiredInputs.isSubset(of: availableVariables) {
                candidates.insert(instr.index)
                // This instruction is available, and so are its outputs...
                availableVariables.formUnion(instr.allOutputs)
            } else {
                // While we cannot include this instruction, we may still be able to replace its outputs with existing variables in the host program
                // which will allow other instructions that depend on these outputs to be included.
                tryRemapVariables(instr.allOutputs, of: instr)
            }
        }

        //
        // Step (3): select the "root" instruction of the slice or use the provided one if any.
        //
        // Simple optimization: avoid splicing data-flow "roots", i.e. simple instructions that don't have any inputs, as this will
        // most of the time result in fairly uninteresting splices that for example just copy a literal from another program.
        // The exception to this are special instructions that exist outside of JavaScript context, for example instructions that add fields to classes.
        let rootCandidates = candidates.filter({ !program.code[$0].isSimple || program.code[$0].numInputs > 0 || !program.code[$0].op.requiredContext.contains(.javascript) })
        guard !rootCandidates.isEmpty else { return false }
        let rootIndex = specifiedIndex ?? chooseUniform(from: rootCandidates)
        guard rootCandidates.contains(rootIndex) else { return false }
        trace("Splicing instruction \(rootIndex) (\(program.code[rootIndex].op.name)) from \(program.id)")

        //
        // Step (4): compute the slice.
        //
        var slice = Set<Int>()
        var requiredVariables = VariableSet()
        var shouldIncludeCurrentBlock = false
        var startOfCurrentBlock = -1
        var index = rootIndex
        while index >= 0 {
            let instr = program.code[index]

            var includeCurrentInstruction = false
            if index == rootIndex {
                // This is the root of the slice, so include it.
                includeCurrentInstruction = true
                assert(candidates.contains(index))
            } else if shouldIncludeCurrentBlock {
                // This instruction is part of the slice because one of its surrounding blocks is included.
                includeCurrentInstruction = true
                // In this case, the instruction isn't necessarily a candidate (but at least one of its surrounding blocks is).
            } else if !requiredVariables.isDisjoint(with: instr.allOutputs) {
                // This instruction is part of the slice because at least one of its outputs is required.
                includeCurrentInstruction = true
                assert(candidates.contains(index))
            } else {
                // Also (potentially) include instructions that can modify one of the required variables if they can be included in the slice.
                if probability(probabilityOfIncludingAnInstructionThatMayMutateARequiredVariable) {
                    if candidates.contains(index) && instr.mayMutate(anyOf: requiredVariables) {
                        includeCurrentInstruction = true
                    }
                }
            }

            if includeCurrentInstruction {
                slice.insert(instr.index)

                // Only those inputs that we haven't picked replacements for are now also required.
                let newlyRequiredVariables = instr.inputs.filter({ !remappedVariables.contains($0) })
                requiredVariables.formUnion(newlyRequiredVariables)

                if !shouldIncludeCurrentBlock && instr.isBlock {
                    // We're including a block instruction due to its outputs. We now need to ensure that we include the full block with it.
                    shouldIncludeCurrentBlock = true
                    let block = blocks[index]!
                    startOfCurrentBlock = block.startIndex
                    index = block.endIndex + 1
                }
            }

            if index == startOfCurrentBlock {
                assert(instr.isBlockGroupStart)
                shouldIncludeCurrentBlock = false
                startOfCurrentBlock = -1
            }

            index -= 1
        }

        //
        // Step (5): insert the final slice into the current program while also remapping any missing variables to their replacements selected in step (2).
        //
        var variableMap = remappedVariables
        for instr in program.code where slice.contains(instr.index) {
            for output in instr.allOutputs {
                variableMap[output] = nextVariable()
            }
            let inouts = instr.inouts.map({ variableMap[$0]! })
            append(Instruction(instr.op, inouts: inouts, flags: instr.flags))
        }

        trace("Splicing done")
        return true
    }

    // Code Building Algorithm:
    //
    // In theory, the basic building algorithm is simply:
    //
    //   var remainingBudget = initialBudget
    //   while remainingBudget > 0 {
    //       if probability(0.5) {
    //           remainingBudget -= runRandomCodeGenerator()
    //       } else {
    //           remainingBudget -= performSplicing()
    //       }
    //   }
    //
    // In practice, things become a little more complicated because code generators can be recursive: a function
    // generator will emit the function start and end and recursively call into the code building machinery to fill the
    // body of the function. The size of the recursively generated blocks is determined as a fraction of the parent's
    // *initial budget*. This ensures that the sizes of recursively generated blocks roughly follow the same
    // distribution. However, it also means that the initial budget can be overshot by quite a bit: we may end up
    // invoking a recursive generator near the end of our budget, which may then for example generate another 0.5x
    // initialBudget instructions. However, the benefit of this approach is that there are really only two "knobs" that
    // determine the "shape" of the generated code: the factor that determines the recursive budget relative to the
    // parent budget and the (absolute) threshold for recursive code generation.
    //

    /// The first "knob": this mainly determines the shape of generated code as it determines how large block bodies are relative to their surrounding code.
    /// This also influences the nesting depth of the generated code, as recursive code generators are only invoked if enough "budget" is still available.
    /// These are writable so they can be reconfigured in tests.
    var minRecursiveBudgetRelativeToParentBudget = 0.05
    var maxRecursiveBudgetRelativeToParentBudget = 0.50

    /// The second "knob": the minimum budget required to be able to invoke recursive code generators.
    public static let minBudgetForRecursiveCodeGeneration = 5

    /// Possible building modes. These are used as argument for build() and determine how the new code is produced.
    public enum BuildingMode {
        // Generate code by running CodeGenerators.
        case generating
        // Splice code from other random programs in the corpus.
        case splicing
        // Do all of the above.
        case generatingAndSplicing
    }

    // Keeps track of the state of one buildInternal() invocation. These are tracked in a stack, one entry for each recursive call.
    // This is a class so that updating the currently active state is possible without push/pop.
    private class BuildingState {
        let initialBudget: Int
        let mode: BuildingMode
        var recursiveBuildingAllowed = true
        var nextRecursiveBlockOfCurrentGenerator = 1
        var totalRecursiveBlocksOfCurrentGenerator: Int? = nil
        // An optional budget for recursive building.
        var recursiveBudget: Int? = nil

        init(initialBudget: Int, mode: BuildingMode) {
            assert(initialBudget > 0)
            self.initialBudget = initialBudget
            self.mode = mode
        }
    }
    private var buildStack = Stack<BuildingState>()

    /// Build random code at the current position in the program.
    ///
    /// The first parameter controls the number of emitted instructions: as soon as more than that number of instructions have been emitted, building stops.
    /// This parameter is only a rough estimate as recursive code generators may lead to significantly more code being generated.
    /// Typically, the actual number of generated instructions will be somewhere between n and 2x n.
    ///
    /// Building code requires that there are visible variables available as inputs for CodeGenerators or as replacement variables for splicing.
    /// When building new programs, `buildPrefix()` can be used to generate some initial variables. `build()` purposely does not call
    /// `buildPrefix()` itself so that the budget isn't accidentally spent just on prefix code (which is probably less interesting).
    public func build(n: Int = 1, by mode: BuildingMode = .generatingAndSplicing) {
        assert(buildStack.isEmpty)
        buildInternal(initialBuildingBudget: n, mode: mode)
        assert(buildStack.isEmpty)
    }

    /// Recursive code building. Used by CodeGenerators for example to fill the bodies of generated blocks.
    public func buildRecursive(block: Int = 1, of numBlocks: Int = 1, n optionalBudget: Int? = nil) {
        assert(!buildStack.isEmpty)
        let parentState = buildStack.top

        assert(parentState.mode != .splicing)
        assert(parentState.recursiveBuildingAllowed)        // If this fails, a recursive CodeGenerator is probably not marked as recursive.
        assert(numBlocks >= 1)
        assert(block >= 1 && block <= numBlocks)
        assert(parentState.nextRecursiveBlockOfCurrentGenerator == block)
        assert((parentState.totalRecursiveBlocksOfCurrentGenerator ?? numBlocks) == numBlocks)

        parentState.nextRecursiveBlockOfCurrentGenerator = block + 1
        parentState.totalRecursiveBlocksOfCurrentGenerator = numBlocks

        // Determine the budget for this recursive call as a fraction of the parent's initial budget.
        var recursiveBudget: Double
        if let specifiedBudget = parentState.recursiveBudget {
            assert(specifiedBudget > 0)
            recursiveBudget = Double(specifiedBudget)
        } else {
            let factor = Double.random(in: minRecursiveBudgetRelativeToParentBudget...maxRecursiveBudgetRelativeToParentBudget)
            assert(factor > 0.0 && factor < 1.0)
            let parentBudget = parentState.initialBudget
            recursiveBudget = Double(parentBudget) * factor
        }

        // Now split the budget between all sibling blocks.
        recursiveBudget /= Double(numBlocks)
        recursiveBudget.round(.up)
        assert(recursiveBudget >= 1.0)

        // Finally, if a custom budget was requested, choose the smaller of the two values.
        if let requestedBudget = optionalBudget {
            assert(requestedBudget > 0)
            recursiveBudget = min(recursiveBudget, Double(requestedBudget))
        }

        buildInternal(initialBuildingBudget: Int(recursiveBudget), mode: parentState.mode)
    }

    private func buildInternal(initialBuildingBudget: Int, mode: BuildingMode) {
        assert(initialBuildingBudget > 0)

        // Both splicing and code generation can sometimes fail, for example if no other program with the necessary features exists.
        // To avoid infinite loops, we bail out after a certain number of consecutive failures.
        var consecutiveFailures = 0

        let state = BuildingState(initialBudget: initialBuildingBudget, mode: mode)
        buildStack.push(state)
        defer { buildStack.pop() }
        var remainingBudget = initialBuildingBudget

        // Unless we are only splicing, find all generators that have the required context. We must always have at least one suitable code generator.
        let origContext = context
        var availableGenerators = WeightedList<CodeGenerator>()
        if state.mode != .splicing {
            availableGenerators = fuzzer.codeGenerators.filter({ $0.requiredContext.isSubset(of: origContext) })
            assert(!availableGenerators.isEmpty)
        }

        struct BuildLog {
            enum ActionOutcome {
                case success
                case failed
            }

            struct BuildAction {
                var action: String
                var outcome: ActionOutcome?
            }

            var actions = [BuildAction]()

            mutating func startAction(_ action: String) {
                // Make sure that we have either completed our last build step or we haven't started any build steps yet.
                assert(actions.isEmpty || actions[actions.count - 1].outcome != nil)
                actions.append(BuildAction(action: action))
            }

            mutating func endAction(withOutcome outcome: ActionOutcome) {
                assert(!actions.isEmpty && actions[actions.count - 1].outcome == nil)
                actions[actions.count - 1].outcome = outcome
            }

        }

        var buildLog = fuzzer.config.logLevel.isAtLeast(.verbose) ? BuildLog() : nil

        while remainingBudget > 0 {
            assert(context == origContext, "Code generation or splicing must not change the current context")

            if state.recursiveBuildingAllowed &&
                remainingBudget < ProgramBuilder.minBudgetForRecursiveCodeGeneration &&
                availableGenerators.contains(where: { !$0.isRecursive }) {
                // No more recursion at this point since the remaining budget is too small.
                state.recursiveBuildingAllowed = false
                availableGenerators = availableGenerators.filter({ !$0.isRecursive })
                assert(state.mode == .splicing || !availableGenerators.isEmpty)
            }

            var mode = state.mode
            if mode == .generatingAndSplicing {
                mode = chooseUniform(from: [.generating, .splicing])
            }

            let codeSizeBefore = code.count
            switch mode {
            case .generating:
                // This requirement might seem somewhat arbitrary but our JavaScript code generators make use of `b.randomVariable` and as such rely on the availability of
                // visible Variables. Therefore we should always have some Variables visible if we want to use them.
                assert(hasVisibleVariables, "CodeGenerators assume that there are visible variables to use. Use buildPrefix() to generate some initial variables in a new program")

                // Reset the code generator specific part of the state.
                state.nextRecursiveBlockOfCurrentGenerator = 1
                state.totalRecursiveBlocksOfCurrentGenerator = nil

                // Select a random generator and run it.
                let generator = availableGenerators.randomElement()
                buildLog?.startAction(generator.name)
                run(generator)

            case .splicing:
                let program = fuzzer.corpus.randomElementForSplicing()
                buildLog?.startAction("splicing")
                splice(from: program)

            default:
                fatalError("Unknown ProgramBuildingMode \(mode)")
            }
            let codeSizeAfter = code.count

            let emittedInstructions = codeSizeAfter - codeSizeBefore
            remainingBudget -= emittedInstructions
            if emittedInstructions > 0 {
                buildLog?.endAction(withOutcome: .success)
                consecutiveFailures = 0
            } else {
                buildLog?.endAction(withOutcome: .failed)
                consecutiveFailures += 1
                guard consecutiveFailures < 10 else {
                    // When splicing, this is somewhat expected as we may not find code to splice if we're in a restricted
                    // context (e.g. we're inside a switch, but can't find another program with switch-cases).
                    // However, when generating code this should happen very rarely since we should always be able to
                    // generate code, not matter what context we are currently in.
                    if state.mode != .splicing {
                        logger.warning("Too many consecutive failures during code building with mode .\(state.mode). Bailing out.")
                        if let actions = buildLog?.actions {
                            logger.verbose("Build log:")
                            for action in actions {
                                logger.verbose("    \(action.action): \(action.outcome!)")
                            }
                        }
                    }
                    return
                }
            }
        }
    }

    /// Run ValueGenerators until we have created at least N new variables.
    /// Returns both the number of generated instructions and of newly created variables.
    @discardableResult
    public func buildValues(_ n: Int) -> (generatedInstructions: Int, generatedVariables: Int) {
        assert(buildStack.isEmpty)
        assert(context.contains(.javascript))

        let valueGenerators = fuzzer.codeGenerators.filter({ $0.isValueGenerator })
        assert(!valueGenerators.isEmpty)
        let previousNumberOfVisibleVariables = numberOfVisibleVariables
        var totalNumberOfGeneratedInstructions = 0

        // ValueGenerators can be recursive.
        // Here we create a builder stack entry for that case which gives each generator a fixed recursive
        // budget and allows us to run code generators when building recursively. We probably don't want to run
        // splicing here since splicing isn't as careful as code generation and may lead to invalid code more quickly.
        // The `initialBudget` isn't really used (since we specify a `recursiveBudget`), so can be an arbitrary value.
        let state = BuildingState(initialBudget: 2 * n, mode: .generating)
        state.recursiveBudget = n
        buildStack.push(state)
        defer { buildStack.pop() }

        while numberOfVisibleVariables - previousNumberOfVisibleVariables < n {
            let generator = valueGenerators.randomElement()
            assert(generator.requiredContext == .javascript && generator.inputs.count == 0)

            state.nextRecursiveBlockOfCurrentGenerator = 1
            state.totalRecursiveBlocksOfCurrentGenerator = nil
            let numberOfGeneratedInstructions = run(generator)

            assert(numberOfGeneratedInstructions > 0, "ValueGenerators must always succeed")
            totalNumberOfGeneratedInstructions += numberOfGeneratedInstructions
        }
        return (totalNumberOfGeneratedInstructions, numberOfVisibleVariables - previousNumberOfVisibleVariables)
    }

    /// Bootstrap program building by creating some variables with statically known types.
    ///
    /// The `build()` method for generating new code or splicing from existing code can
    /// only be used once there are visible variables. This method can be used to generate some.
    ///
    /// Internally, this uses the ValueGenerators to generate some code. As such, the "shape"
    /// of prefix code is controlled in the same way as other generated code through the
    /// generator's respective weights.
    public func buildPrefix() {
        // Each value generators should generate at least 3 variables, and we probably want to run at least a
        // few of them (maybe roughly >= 3), so the number of variables to build shouldn't be set too low.
        assert(CodeGenerator.numberOfValuesToGenerateByValueGenerators == 3)
        let numValuesToBuild = Int.random(in: 10...15)

        trace("Start of prefix code")
        buildValues(numValuesToBuild)
        assert(numberOfVisibleVariables >= numValuesToBuild)
        trace("End of prefix code. \(numberOfVisibleVariables) variables are now visible")
    }

    /// Runs a code generator in the current context and returns the number of generated instructions.
    @discardableResult
    private func run(_ generator: CodeGenerator) -> Int {
        assert(generator.requiredContext.isSubset(of: context))

        trace("Executing code generator \(generator.name)")
        var inputs = [Variable]()
        switch generator.inputs.mode {
        case .loose:
            // Find inputs that are probably compatible with the desired input types using randomVariable(forUseAs:)
            inputs = generator.inputs.types.map(randomVariable(forUseAs:))

        case .strict:
            // Find inputs of the required type using randomVariable(ofType:)
            for inputType in generator.inputs.types {
                guard let input = randomVariable(ofType: inputType) else {
                    // Cannot run this generator
                    return 0
                }
                inputs.append(input)
            }
        }
        let numGeneratedInstructions = generator.run(in: self, with: inputs)
        trace("Code generator finished")

        if numGeneratedInstructions > 0 {
            contributors.insert(generator)
            generator.addedInstructions(numGeneratedInstructions)
        }

        return numGeneratedInstructions
    }

    //
    // Low-level instruction constructors.
    //
    // These create an instruction with the provided values and append it to the program at the current position.
    // If the instruction produces a new variable, that variable is returned to the caller.
    // Each class implementing the Operation protocol will have a constructor here.
    //

    @discardableResult
    private func emit(_ op: Operation, withInputs inputs: [Variable] = []) -> Instruction {
        var inouts = inputs
        for _ in 0..<op.numOutputs {
            inouts.append(nextVariable())
        }
        for _ in 0..<op.numInnerOutputs {
            inouts.append(nextVariable())
        }

        return internalAppend(Instruction(op, inouts: inouts, flags: .empty))
    }

    @discardableResult
    public func loadInt(_ value: Int64) -> Variable {
        return emit(LoadInteger(value: value)).output
    }

    @discardableResult
    public func loadBigInt(_ value: Int64) -> Variable {
        return emit(LoadBigInt(value: value)).output
    }

    @discardableResult
    public func loadFloat(_ value: Double) -> Variable {
        return emit(LoadFloat(value: value)).output
    }

    @discardableResult
    public func loadString(_ value: String) -> Variable {
        return emit(LoadString(value: value)).output
    }

    @discardableResult
    public func loadBool(_ value: Bool) -> Variable {
        return emit(LoadBoolean(value: value)).output
    }

    @discardableResult
    public func loadUndefined() -> Variable {
        return emit(LoadUndefined()).output
    }

    @discardableResult
    public func loadNull() -> Variable {
        return emit(LoadNull()).output
    }

    @discardableResult
    public func loadThis() -> Variable {
        return emit(LoadThis()).output
    }

    @discardableResult
    public func loadArguments() -> Variable {
        return emit(LoadArguments()).output
    }

    @discardableResult
    public func loadDisposableVariable(_ value: Variable) -> Variable {
        return emit(LoadDisposableVariable(), withInputs: [value]).output
    }

    @discardableResult
    public func loadAsyncDisposableVariable(_ value: Variable) -> Variable {
        return emit(LoadAsyncDisposableVariable(), withInputs: [value]).output
    }

    @discardableResult
    public func loadNewTarget() -> Variable {
        return emit(LoadNewTarget()).output
    }

    @discardableResult
    public func loadRegExp(_ pattern: String, _ flags: RegExpFlags) -> Variable {
        return emit(LoadRegExp(pattern: pattern, flags: flags)).output
    }

    /// Represents a currently active object literal. Used to add fields to it and to query which fields already exist.
    public class ObjectLiteral {
        private let b: ProgramBuilder

        public fileprivate(set) var properties: [String] = []
        public fileprivate(set) var elements: [Int64] = []
        public fileprivate(set) var computedProperties: [Variable] = []
        public fileprivate(set) var methods: [String] = []
        public fileprivate(set) var computedMethods: [Variable] = []
        public fileprivate(set) var getters: [String] = []
        public fileprivate(set) var setters: [String] = []
        public fileprivate(set) var hasPrototype = false

        fileprivate init(in b: ProgramBuilder) {
            assert(b.context.contains(.objectLiteral))
            self.b = b
        }

        public func addProperty(_ name: String, as value: Variable) {
            b.emit(ObjectLiteralAddProperty(propertyName: name), withInputs: [value])
        }

        public func addElement(_ index: Int64, as value: Variable) {
            b.emit(ObjectLiteralAddElement(index: index), withInputs: [value])
        }

        public func addComputedProperty(_ name: Variable, as value: Variable) {
            b.emit(ObjectLiteralAddComputedProperty(), withInputs: [name, value])
        }

        public func copyProperties(from obj: Variable) {
            b.emit(ObjectLiteralCopyProperties(), withInputs: [obj])
        }

        public func setPrototype(to proto: Variable) {
            b.emit(ObjectLiteralSetPrototype(), withInputs: [proto])
        }

        public func addMethod(_ name: String, with descriptor: SubroutineDescriptor, _ body: ([Variable]) -> ()) {
            b.setParameterTypesForNextSubroutine(descriptor.parameterTypes)
            let instr = b.emit(BeginObjectLiteralMethod(methodName: name, parameters: descriptor.parameters))
            body(Array(instr.innerOutputs))
            b.emit(EndObjectLiteralMethod())
        }

        public func addComputedMethod(_ name: Variable, with descriptor: SubroutineDescriptor, _ body: ([Variable]) -> ()) {
            b.setParameterTypesForNextSubroutine(descriptor.parameterTypes)
            let instr = b.emit(BeginObjectLiteralComputedMethod(parameters: descriptor.parameters), withInputs: [name])
            body(Array(instr.innerOutputs))
            b.emit(EndObjectLiteralComputedMethod())
        }

        public func addGetter(for name: String, _ body: (_ this: Variable) -> ()) {
            let instr = b.emit(BeginObjectLiteralGetter(propertyName: name))
            body(instr.innerOutput)
            b.emit(EndObjectLiteralGetter())
        }

        public func addSetter(for name: String, _ body: (_ this: Variable, _ val: Variable) -> ()) {
            let instr = b.emit(BeginObjectLiteralSetter(propertyName: name))
            body(instr.innerOutput(0), instr.innerOutput(1))
            b.emit(EndObjectLiteralSetter())
        }
    }

    @discardableResult
    public func buildObjectLiteral(_ body: (ObjectLiteral) -> ()) -> Variable {
        emit(BeginObjectLiteral())
        body(currentObjectLiteral)
        return emit(EndObjectLiteral()).output
    }

    @discardableResult
    // Convenience method to create simple object literals.
    public func createObject(with initialProperties: [String: Variable]) -> Variable {
        return buildObjectLiteral { obj in
            // Sort the property names so that the emitted code is deterministic.
            for (propertyName, value) in initialProperties.sorted(by: { $0.key < $1.key }) {
                obj.addProperty(propertyName, as: value)
            }
        }
    }

    /// Represents a currently active class definition. Used to add fields to it and to query which fields already exist.
    public class ClassDefinition {
        private let b: ProgramBuilder

        public let isDerivedClass: Bool

        public fileprivate(set) var hasConstructor = false

        public fileprivate(set) var instanceProperties: [String] = []
        public fileprivate(set) var instanceElements: [Int64] = []
        public fileprivate(set) var instanceComputedProperties: [Variable] = []
        public fileprivate(set) var instanceMethods: [String] = []
        public fileprivate(set) var instanceGetters: [String] = []
        public fileprivate(set) var instanceSetters: [String] = []

        public fileprivate(set) var staticProperties: [String] = []
        public fileprivate(set) var staticElements: [Int64] = []
        public fileprivate(set) var staticComputedProperties: [Variable] = []
        public fileprivate(set) var staticMethods: [String] = []
        public fileprivate(set) var staticGetters: [String] = []
        public fileprivate(set) var staticSetters: [String] = []

        // These sets are required to ensure syntactic correctness, not just as an optimization to
        // avoid adding duplicate fields:
        // In JavaScript, it is a syntax error to access a private property/method that has not
        // been declared by the surrounding class. Further, each private field must only be declared
        // once, regardless of whether it is a method or a property and whether it's per-instance or
        // static. However, we still track properties and methods separately to facilitate selecting
        // property and method names for private property accesses and private method calls.
        public fileprivate(set) var privateProperties: [String] = []
        public fileprivate(set) var privateMethods: [String] = []
        public var privateFields: [String] {
            return privateProperties + privateMethods
        }

        fileprivate init(in b: ProgramBuilder, isDerived: Bool) {
            assert(b.context.contains(.classDefinition))
            self.b = b
            self.isDerivedClass = isDerived
        }

        public func addConstructor(with descriptor: SubroutineDescriptor, _ body: ([Variable]) -> ()) {
            b.setParameterTypesForNextSubroutine(descriptor.parameterTypes)
            let instr = b.emit(BeginClassConstructor(parameters: descriptor.parameters))
            body(Array(instr.innerOutputs))
            b.emit(EndClassConstructor())
        }

        public func addInstanceProperty(_ name: String, value: Variable? = nil) {
            let inputs = value != nil ? [value!] : []
            b.emit(ClassAddInstanceProperty(propertyName: name, hasValue: value != nil), withInputs: inputs)
        }

        public func addInstanceElement(_ index: Int64, value: Variable? = nil) {
            let inputs = value != nil ? [value!] : []
            b.emit(ClassAddInstanceElement(index: index, hasValue: value != nil), withInputs: inputs)
        }

        public func addInstanceComputedProperty(_ name: Variable, value: Variable? = nil) {
            let inputs = value != nil ? [name, value!] : [name]
            b.emit(ClassAddInstanceComputedProperty(hasValue: value != nil), withInputs: inputs)
        }

        public func addInstanceMethod(_ name: String, with descriptor: SubroutineDescriptor, _ body: ([Variable]) -> ()) {
            b.setParameterTypesForNextSubroutine(descriptor.parameterTypes)
            let instr = b.emit(BeginClassInstanceMethod(methodName: name, parameters: descriptor.parameters))
            body(Array(instr.innerOutputs))
            b.emit(EndClassInstanceMethod())
        }

        public func addInstanceGetter(for name: String, _ body: (_ this: Variable) -> ()) {
            let instr = b.emit(BeginClassInstanceGetter(propertyName: name))
            body(instr.innerOutput)
            b.emit(EndClassInstanceGetter())
        }

        public func addInstanceSetter(for name: String, _ body: (_ this: Variable, _ val: Variable) -> ()) {
            let instr = b.emit(BeginClassInstanceSetter(propertyName: name))
            body(instr.innerOutput(0), instr.innerOutput(1))
            b.emit(EndClassInstanceSetter())
        }

        public func addStaticProperty(_ name: String, value: Variable? = nil) {
            let inputs = value != nil ? [value!] : []
            b.emit(ClassAddStaticProperty(propertyName: name, hasValue: value != nil), withInputs: inputs)
        }

        public func addStaticElement(_ index: Int64, value: Variable? = nil) {
            let inputs = value != nil ? [value!] : []
            b.emit(ClassAddStaticElement(index: index, hasValue: value != nil), withInputs: inputs)
        }

        public func addStaticComputedProperty(_ name: Variable, value: Variable? = nil) {
            let inputs = value != nil ? [name, value!] : [name]
            b.emit(ClassAddStaticComputedProperty(hasValue: value != nil), withInputs: inputs)
        }

        public func addStaticInitializer(_ body: (Variable) -> ()) {
            let instr = b.emit(BeginClassStaticInitializer())
            body(instr.innerOutput)
            b.emit(EndClassStaticInitializer())
        }

        public func addStaticMethod(_ name: String, with descriptor: SubroutineDescriptor, _ body: ([Variable]) -> ()) {
            b.setParameterTypesForNextSubroutine(descriptor.parameterTypes)
            let instr = b.emit(BeginClassStaticMethod(methodName: name, parameters: descriptor.parameters))
            body(Array(instr.innerOutputs))
            b.emit(EndClassStaticMethod())
        }

        public func addStaticGetter(for name: String, _ body: (_ this: Variable) -> ()) {
            let instr = b.emit(BeginClassStaticGetter(propertyName: name))
            body(instr.innerOutput)
            b.emit(EndClassStaticGetter())
        }

        public func addStaticSetter(for name: String, _ body: (_ this: Variable, _ val: Variable) -> ()) {
            let instr = b.emit(BeginClassStaticSetter(propertyName: name))
            body(instr.innerOutput(0), instr.innerOutput(1))
            b.emit(EndClassStaticSetter())
        }

        public func addPrivateInstanceProperty(_ name: String, value: Variable? = nil) {
            let inputs = value != nil ? [value!] : []
            b.emit(ClassAddPrivateInstanceProperty(propertyName: name, hasValue: value != nil), withInputs: inputs)
        }

        public func addPrivateInstanceMethod(_ name: String, with descriptor: SubroutineDescriptor, _ body: ([Variable]) -> ()) {
            b.setParameterTypesForNextSubroutine(descriptor.parameterTypes)
            let instr = b.emit(BeginClassPrivateInstanceMethod(methodName: name, parameters: descriptor.parameters))
            body(Array(instr.innerOutputs))
            b.emit(EndClassPrivateInstanceMethod())
        }

        public func addPrivateStaticProperty(_ name: String, value: Variable? = nil) {
            let inputs = value != nil ? [value!] : []
            b.emit(ClassAddPrivateStaticProperty(propertyName: name, hasValue: value != nil), withInputs: inputs)
        }

        public func addPrivateStaticMethod(_ name: String, with descriptor: SubroutineDescriptor, _ body: ([Variable]) -> ()) {
            b.setParameterTypesForNextSubroutine(descriptor.parameterTypes)
            let instr = b.emit(BeginClassPrivateStaticMethod(methodName: name, parameters: descriptor.parameters))
            body(Array(instr.innerOutputs))
            b.emit(EndClassPrivateStaticMethod())
        }
    }

    @discardableResult
    public func buildClassDefinition(withSuperclass superclass: Variable? = nil, _ body: (ClassDefinition) -> ()) -> Variable {
        let inputs = superclass != nil ? [superclass!] : []
        let output = emit(BeginClassDefinition(hasSuperclass: superclass != nil), withInputs: inputs).output
        if enableRecursionGuard { hide(output) }
        body(currentClassDefinition)
        if enableRecursionGuard { unhide(output) }
        emit(EndClassDefinition())
        return output
    }

    @discardableResult
    public func createArray(with initialValues: [Variable]) -> Variable {
        return emit(CreateArray(numInitialValues: initialValues.count), withInputs: initialValues).output
    }

    @discardableResult
    public func createIntArray(with initialValues: [Int64]) -> Variable {
        return emit(CreateIntArray(values: initialValues)).output
    }

    @discardableResult
    public func createFloatArray(with initialValues: [Double]) -> Variable {
        return emit(CreateFloatArray(values: initialValues)).output
    }

    @discardableResult
    public func createArray(with initialValues: [Variable], spreading spreads: [Bool]) -> Variable {
        assert(initialValues.count == spreads.count)
        return emit(CreateArrayWithSpread(spreads: spreads), withInputs: initialValues).output
    }

    @discardableResult
    public func createTemplateString(from parts: [String], interpolating interpolatedValues: [Variable]) -> Variable {
        return emit(CreateTemplateString(parts: parts), withInputs: interpolatedValues).output
    }

    @discardableResult
    public func getProperty(_ name: String, of object: Variable, guard isGuarded: Bool = false) -> Variable {
        return emit(GetProperty(propertyName: name, isGuarded: isGuarded), withInputs: [object]).output
    }

    public func setProperty(_ name: String, of object: Variable, to value: Variable) {
        emit(SetProperty(propertyName: name), withInputs: [object, value])
    }

    public func updateProperty(_ name: String, of object: Variable, with value: Variable, using op: BinaryOperator) {
        emit(UpdateProperty(propertyName: name, operator: op), withInputs: [object, value])
    }

    @discardableResult
    public func deleteProperty(_ name: String, of object: Variable, guard isGuarded: Bool = false) -> Variable {
        emit(DeleteProperty(propertyName: name, isGuarded: isGuarded), withInputs: [object]).output
    }

    public enum PropertyConfiguration {
        case value(Variable)
        case getter(Variable)
        case setter(Variable)
        case getterSetter(Variable, Variable)
    }

    public func configureProperty(_ name: String, of object: Variable, usingFlags flags: PropertyFlags, as config: PropertyConfiguration) {
        switch config {
        case .value(let value):
            emit(ConfigureProperty(propertyName: name, flags: flags, type: .value), withInputs: [object, value])
        case .getter(let getter):
            emit(ConfigureProperty(propertyName: name, flags: flags, type: .getter), withInputs: [object, getter])
        case .setter(let setter):
            emit(ConfigureProperty(propertyName: name, flags: flags, type: .setter), withInputs: [object, setter])
        case .getterSetter(let getter, let setter):
            emit(ConfigureProperty(propertyName: name, flags: flags, type: .getterSetter), withInputs: [object, getter, setter])
        }
    }

    @discardableResult
    public func getElement(_ index: Int64, of array: Variable, guard isGuarded: Bool = false) -> Variable {
        return emit(GetElement(index: index, isGuarded: isGuarded), withInputs: [array]).output
    }

    public func setElement(_ index: Int64, of array: Variable, to value: Variable) {
        emit(SetElement(index: index), withInputs: [array, value])
    }

    public func updateElement(_ index: Int64, of array: Variable, with value: Variable, using op: BinaryOperator) {
        emit(UpdateElement(index: index, operator: op), withInputs: [array, value])
    }

    @discardableResult
    public func deleteElement(_ index: Int64, of array: Variable, guard isGuarded: Bool = false) -> Variable {
        emit(DeleteElement(index: index, isGuarded: isGuarded), withInputs: [array]).output
    }

    public func configureElement(_ index: Int64, of object: Variable, usingFlags flags: PropertyFlags, as config: PropertyConfiguration) {
        switch config {
        case .value(let value):
            emit(ConfigureElement(index: index, flags: flags, type: .value), withInputs: [object, value])
        case .getter(let getter):
            emit(ConfigureElement(index: index, flags: flags, type: .getter), withInputs: [object, getter])
        case .setter(let setter):
            emit(ConfigureElement(index: index, flags: flags, type: .setter), withInputs: [object, setter])
        case .getterSetter(let getter, let setter):
            emit(ConfigureElement(index: index, flags: flags, type: .getterSetter), withInputs: [object, getter, setter])
        }
    }

    @discardableResult
    public func getComputedProperty(_ name: Variable, of object: Variable, guard isGuarded: Bool = false) -> Variable {
        return emit(GetComputedProperty(isGuarded: isGuarded), withInputs: [object, name]).output
    }

    public func setComputedProperty(_ name: Variable, of object: Variable, to value: Variable) {
        emit(SetComputedProperty(), withInputs: [object, name, value])
    }

    public func updateComputedProperty(_ name: Variable, of object: Variable, with value: Variable, using op: BinaryOperator) {
        emit(UpdateComputedProperty(operator: op), withInputs: [object, name, value])
    }

    @discardableResult
    public func deleteComputedProperty(_ name: Variable, of object: Variable, guard isGuarded: Bool = false) -> Variable {
        emit(DeleteComputedProperty(isGuarded: isGuarded), withInputs: [object, name]).output
    }

    public func configureComputedProperty(_ name: Variable, of object: Variable, usingFlags flags: PropertyFlags, as config: PropertyConfiguration) {
        switch config {
        case .value(let value):
            emit(ConfigureComputedProperty(flags: flags, type: .value), withInputs: [object, name, value])
        case .getter(let getter):
            emit(ConfigureComputedProperty(flags: flags, type: .getter), withInputs: [object, name, getter])
        case .setter(let setter):
            emit(ConfigureComputedProperty(flags: flags, type: .setter), withInputs: [object, name, setter])
        case .getterSetter(let getter, let setter):
            emit(ConfigureComputedProperty(flags: flags, type: .getterSetter), withInputs: [object, name, getter, setter])
        }
    }

    @discardableResult
    public func typeof(_ v: Variable) -> Variable {
        return emit(TypeOf(), withInputs: [v]).output
    }

    @discardableResult
    public func void(_ v: Variable) -> Variable {
        return emit(Void_(), withInputs: [v]).output
    }

    @discardableResult
    public func testInstanceOf(_ v: Variable, _ type: Variable) -> Variable {
        return emit(TestInstanceOf(), withInputs: [v, type]).output
    }

    @discardableResult
    public func testIn(_ prop: Variable, _ obj: Variable) -> Variable {
        return emit(TestIn(), withInputs: [prop, obj]).output
    }

    public func explore(_ v: Variable, id: String, withArgs arguments: [Variable]) {
        let rngSeed = UInt32(truncatingIfNeeded: randomInt())
        emit(Explore(id: id, numArguments: arguments.count, rngSeed: rngSeed), withInputs: [v] + arguments)
    }

    public func probe(_ v: Variable, id: String) {
        emit(Probe(id: id), withInputs: [v])
    }

    @discardableResult
    public func fixup(id: String, action: String, originalOperation: String, arguments: [Variable], hasOutput: Bool) -> Variable? {
        let instr = emit(Fixup(id: id, action: action, originalOperation: originalOperation, numArguments: arguments.count, hasOutput: hasOutput), withInputs: arguments)
        return hasOutput ? instr.output : nil
    }

    // Helper struct to describe subroutine definitions.
    // This allows defining functions by just specifying the number of parameters or by specifying the types of the individual parameters.
    // Note however that parameter types are not associated with the generated operations and will therefore just be valid for the lifetime
    // of this ProgramBuilder. The reason for this behaviour is that it is generally not possible to preserve the type information across program
    // mutations: a mutator may change the callsite of a function or modify the uses of a parameter, effectively invalidating the parameter types.
    // Parameter types are therefore only valid when a function is first created.
    public struct SubroutineDescriptor {
        // The parameter "structure", i.e. the number of parameters and whether there is a rest parameter, etc.
        // Currently, this information is also fully contained in the parameterTypes member. However, if we ever
        // add support for features such as parameter destructuring, this would no longer be the case.
        public let parameters: Parameters
        // Type information for every parameter. If no type information is specified, the parameters will all use .anything as type.
        public let parameterTypes: ParameterList

        public var count: Int {
            return parameters.count
        }

        public static func parameters(n: Int, hasRestParameter: Bool = false) -> SubroutineDescriptor {
            return SubroutineDescriptor(withParameters: Parameters(count: n, hasRestParameter: hasRestParameter))
        }

        public static func parameters(_ params: Parameter...) -> SubroutineDescriptor {
            return parameters(ParameterList(params))
        }

        public static func parameters(_ parameterTypes: ParameterList) -> SubroutineDescriptor {
            let parameters = Parameters(count: parameterTypes.count, hasRestParameter: parameterTypes.hasRestParameter)
            return SubroutineDescriptor(withParameters: parameters, ofTypes: parameterTypes)
        }

        private init(withParameters parameters: Parameters, ofTypes parameterTypes: ParameterList? = nil) {
            if let types = parameterTypes {
                assert(types.areValid())
                assert(types.count == parameters.count)
                assert(types.hasRestParameter == parameters.hasRestParameter)
                self.parameterTypes = types
            } else {
                self.parameterTypes = ParameterList(numParameters: parameters.count, hasRestParam: parameters.hasRestParameter)
                assert(self.parameterTypes.allSatisfy({ $0 == .plain(.anything) || $0 == .rest(.anything) }))
            }
            self.parameters = parameters
        }
    }

    @discardableResult
    public func buildPlainFunction(with descriptor: SubroutineDescriptor, named functionName: String? = nil,_ body: ([Variable]) -> ()) -> Variable {
        setParameterTypesForNextSubroutine(descriptor.parameterTypes)
        let instr = emit(BeginPlainFunction(parameters: descriptor.parameters, functionName: functionName))
        if enableRecursionGuard { hide(instr.output) }
        body(Array(instr.innerOutputs))
        if enableRecursionGuard { unhide(instr.output) }
        emit(EndPlainFunction())
        return instr.output
    }

    @discardableResult
    public func buildArrowFunction(with descriptor: SubroutineDescriptor, _ body: ([Variable]) -> ()) -> Variable {
        setParameterTypesForNextSubroutine(descriptor.parameterTypes)
        let instr = emit(BeginArrowFunction(parameters: descriptor.parameters))
        if enableRecursionGuard { hide(instr.output) }
        body(Array(instr.innerOutputs))
        if enableRecursionGuard { unhide(instr.output) }
        emit(EndArrowFunction())
        return instr.output
    }

    @discardableResult
    public func buildGeneratorFunction(with descriptor: SubroutineDescriptor, named functionName: String? = nil, _ body: ([Variable]) -> ()) -> Variable {
        setParameterTypesForNextSubroutine(descriptor.parameterTypes)
        let instr = emit(BeginGeneratorFunction(parameters: descriptor.parameters, functionName: functionName))
        if enableRecursionGuard { hide(instr.output) }
        body(Array(instr.innerOutputs))
        if enableRecursionGuard { unhide(instr.output) }
        emit(EndGeneratorFunction())
        return instr.output
    }

    @discardableResult
    public func buildAsyncFunction(with descriptor: SubroutineDescriptor, named functionName: String? = nil, _ body: ([Variable]) -> ()) -> Variable {
        setParameterTypesForNextSubroutine(descriptor.parameterTypes)
        let instr = emit(BeginAsyncFunction(parameters: descriptor.parameters, functionName: functionName))
        if enableRecursionGuard { hide(instr.output) }
        body(Array(instr.innerOutputs))
        if enableRecursionGuard { unhide(instr.output) }
        emit(EndAsyncFunction())
        return instr.output
    }

    @discardableResult
    public func buildAsyncArrowFunction(with descriptor: SubroutineDescriptor, _ body: ([Variable]) -> ()) -> Variable {
        setParameterTypesForNextSubroutine(descriptor.parameterTypes)
        let instr = emit(BeginAsyncArrowFunction(parameters: descriptor.parameters))
        if enableRecursionGuard { hide(instr.output) }
        body(Array(instr.innerOutputs))
        if enableRecursionGuard { unhide(instr.output) }
        emit(EndAsyncArrowFunction())
        return instr.output
    }

    @discardableResult
    public func buildAsyncGeneratorFunction(with descriptor: SubroutineDescriptor, named functionName: String? = nil, _ body: ([Variable]) -> ()) -> Variable {
        setParameterTypesForNextSubroutine(descriptor.parameterTypes)
        let instr = emit(BeginAsyncGeneratorFunction(parameters: descriptor.parameters, functionName: functionName))
        if enableRecursionGuard { hide(instr.output) }
        body(Array(instr.innerOutputs))
        if enableRecursionGuard { unhide(instr.output) }
        emit(EndAsyncGeneratorFunction())
        return instr.output
    }

    @discardableResult
    public func buildConstructor(with descriptor: SubroutineDescriptor, _ body: ([Variable]) -> ()) -> Variable {
        setParameterTypesForNextSubroutine(descriptor.parameterTypes)
        let instr = emit(BeginConstructor(parameters: descriptor.parameters))
        if enableRecursionGuard { hide(instr.output) }
        body(Array(instr.innerOutputs))
        if enableRecursionGuard { unhide(instr.output) }
        emit(EndConstructor())
        return instr.output
    }

    public func directive(_ content: String) {
        emit(Directive(content))
    }

    public func doReturn(_ value: Variable? = nil) {
        if let returnValue = value {
            emit(Return(hasReturnValue: true), withInputs: [returnValue])
        } else {
            emit(Return(hasReturnValue: false))
        }
    }

    @discardableResult
    public func yield(_ value: Variable? = nil) -> Variable {
        if let argument = value {
            return emit(Yield(hasArgument: true), withInputs: [argument]).output
        } else {
            return emit(Yield(hasArgument: false)).output
        }
    }

    public func yieldEach(_ value: Variable) {
        emit(YieldEach(), withInputs: [value])
    }

    @discardableResult
    public func await(_ value: Variable) -> Variable {
        return emit(Await(), withInputs: [value]).output
    }

    @discardableResult
    public func callFunction(_ function: Variable, withArgs arguments: [Variable] = [], guard isGuarded: Bool = false) -> Variable {
        return emit(CallFunction(numArguments: arguments.count, isGuarded: isGuarded), withInputs: [function] + arguments).output
    }

    @discardableResult
    public func callFunction(_ function: Variable, withArgs arguments: [Variable], spreading spreads: [Bool], guard isGuarded: Bool = false) -> Variable {
        guard !spreads.isEmpty else { return callFunction(function, withArgs: arguments) }
        return emit(CallFunctionWithSpread(numArguments: arguments.count, spreads: spreads, isGuarded: isGuarded), withInputs: [function] + arguments).output
    }

    @discardableResult
    public func construct(_ constructor: Variable, withArgs arguments: [Variable] = [], guard isGuarded: Bool = false) -> Variable {
        return emit(Construct(numArguments: arguments.count, isGuarded: isGuarded), withInputs: [constructor] + arguments).output
    }

    @discardableResult
    public func construct(_ constructor: Variable, withArgs arguments: [Variable], spreading spreads: [Bool], guard isGuarded: Bool = false) -> Variable {
        guard !spreads.isEmpty else { return construct(constructor, withArgs: arguments) }
        return emit(ConstructWithSpread(numArguments: arguments.count, spreads: spreads, isGuarded: isGuarded), withInputs: [constructor] + arguments).output
    }

    @discardableResult
    public func callMethod(_ name: String, on object: Variable, withArgs arguments: [Variable] = [], guard isGuarded: Bool = false) -> Variable {
        return emit(CallMethod(methodName: name, numArguments: arguments.count, isGuarded: isGuarded), withInputs: [object] + arguments).output
    }

    @discardableResult
    public func callMethod(_ name: String, on object: Variable, withArgs arguments: [Variable], spreading spreads: [Bool], guard isGuarded: Bool = false) -> Variable {
        guard !spreads.isEmpty else { return callMethod(name, on: object, withArgs: arguments) }
        return emit(CallMethodWithSpread(methodName: name, numArguments: arguments.count, spreads: spreads, isGuarded: isGuarded), withInputs: [object] + arguments).output
    }

    @discardableResult
    public func callComputedMethod(_ name: Variable, on object: Variable, withArgs arguments: [Variable] = [], guard isGuarded: Bool = false) -> Variable {
        return emit(CallComputedMethod(numArguments: arguments.count, isGuarded: isGuarded), withInputs: [object, name] + arguments).output
    }

    @discardableResult
    public func callComputedMethod(_ name: Variable, on object: Variable, withArgs arguments: [Variable], spreading spreads: [Bool], guard isGuarded: Bool = false) -> Variable {
        guard !spreads.isEmpty else { return callComputedMethod(name, on: object, withArgs: arguments) }
        return emit(CallComputedMethodWithSpread(numArguments: arguments.count, spreads: spreads, isGuarded: isGuarded), withInputs: [object, name] + arguments).output
    }

    @discardableResult
    public func unary(_ op: UnaryOperator, _ input: Variable) -> Variable {
        return emit(UnaryOperation(op), withInputs: [input]).output
    }

    @discardableResult
    public func binary(_ lhs: Variable, _ rhs: Variable, with op: BinaryOperator) -> Variable {
        return emit(BinaryOperation(op), withInputs: [lhs, rhs]).output
    }

    @discardableResult
    public func ternary(_ condition: Variable, _ lhs: Variable, _ rhs: Variable) -> Variable {
        return emit(TernaryOperation(), withInputs: [condition, lhs, rhs]).output
    }

    public func reassign(_ output: Variable, to input: Variable, with op: BinaryOperator) {
        emit(Update(op), withInputs: [output, input])
    }

    @discardableResult
    public func dup(_ v: Variable) -> Variable {
        return emit(Dup(), withInputs: [v]).output
    }

    public func reassign(_ output: Variable, to input: Variable) {
        emit(Reassign(), withInputs: [output, input])
    }

    @discardableResult
    public func destruct(_ input: Variable, selecting indices: [Int64], lastIsRest: Bool = false) -> [Variable] {
        let outputs = emit(DestructArray(indices: indices, lastIsRest: lastIsRest), withInputs: [input]).outputs
        return Array(outputs)
    }

    public func destruct(_ input: Variable, selecting indices: [Int64], into outputs: [Variable], lastIsRest: Bool = false) {
        emit(DestructArrayAndReassign(indices: indices, lastIsRest: lastIsRest), withInputs: [input] + outputs)
    }

    @discardableResult
    public func destruct(_ input: Variable, selecting properties: [String], hasRestElement: Bool = false) -> [Variable] {
        let outputs = emit(DestructObject(properties: properties, hasRestElement: hasRestElement), withInputs: [input]).outputs
        return Array(outputs)
    }

    public func destruct(_ input: Variable, selecting properties: [String], into outputs: [Variable], hasRestElement: Bool = false) {
        emit(DestructObjectAndReassign(properties: properties, hasRestElement: hasRestElement), withInputs: [input] + outputs)
    }

    @discardableResult
    public func compare(_ lhs: Variable, with rhs: Variable, using comparator: Comparator) -> Variable {
        return emit(Compare(comparator), withInputs: [lhs, rhs]).output
    }

    @discardableResult
    public func createNamedVariable(_ name: String, declarationMode: NamedVariableDeclarationMode, initialValue: Variable? = nil) -> Variable {
        assert((declarationMode == .none) == (initialValue == nil))
        let inputs = initialValue != nil ? [initialValue!] : []
        return emit(CreateNamedVariable(name, declarationMode: declarationMode), withInputs: inputs).output
    }

    @discardableResult
    public func createNamedVariable(forBuiltin builtinName: String) -> Variable {
        return createNamedVariable(builtinName, declarationMode: .none)
    }

    @discardableResult
    public func eval(_ string: String, with arguments: [Variable] = [], hasOutput: Bool = false) -> Variable? {
        let instr = emit(Eval(string, numArguments: arguments.count, hasOutput: hasOutput), withInputs: arguments)
        return hasOutput ? instr.output : nil
    }

    public func buildWith(_ scopeObject: Variable, body: () -> Void) {
        emit(BeginWith(), withInputs: [scopeObject])
        body()
        emit(EndWith())
    }

    public func nop(numOutputs: Int = 0) {
        emit(Nop(numOutputs: numOutputs), withInputs: [])
    }

    public func callSuperConstructor(withArgs arguments: [Variable]) {
        emit(CallSuperConstructor(numArguments: arguments.count), withInputs: arguments)
    }

    @discardableResult
    public func callSuperMethod(_ name: String, withArgs arguments: [Variable] = []) -> Variable {
        return emit(CallSuperMethod(methodName: name, numArguments: arguments.count), withInputs: arguments).output
    }

    @discardableResult
    public func getPrivateProperty(_ name: String, of object: Variable) -> Variable {
        return emit(GetPrivateProperty(propertyName: name), withInputs: [object]).output
    }

    public func setPrivateProperty(_ name: String, of object: Variable, to value: Variable) {
        emit(SetPrivateProperty(propertyName: name), withInputs: [object, value])
    }

    public func updatePrivateProperty(_ name: String, of object: Variable, with value: Variable, using op: BinaryOperator) {
        emit(UpdatePrivateProperty(propertyName: name, operator: op), withInputs: [object, value])
    }

    @discardableResult
    public func callPrivateMethod(_ name: String, on object: Variable, withArgs arguments: [Variable] = []) -> Variable {
        return emit(CallPrivateMethod(methodName: name, numArguments: arguments.count), withInputs: [object] + arguments).output
    }

    @discardableResult
    public func getSuperProperty(_ name: String) -> Variable {
        return emit(GetSuperProperty(propertyName: name)).output
    }

    public func setSuperProperty(_ name: String, to value: Variable) {
        emit(SetSuperProperty(propertyName: name), withInputs: [value])
    }

    @discardableResult
    public func getComputedSuperProperty(_ property: Variable) -> Variable {
        return emit(GetComputedSuperProperty(), withInputs: [property]).output
    }

    public func setComputedSuperProperty(_ property: Variable, to value: Variable) {
        emit(SetComputedSuperProperty(), withInputs: [property, value])
    }

    public func updateSuperProperty(_ name: String, with value: Variable, using op: BinaryOperator) {
        emit(UpdateSuperProperty(propertyName: name, operator: op), withInputs: [value])
    }

    public func buildIf(_ condition: Variable, ifBody: () -> Void) {
        emit(BeginIf(inverted: false), withInputs: [condition])
        ifBody()
        emit(EndIf())
    }

    public func buildIfElse(_ condition: Variable, ifBody: () -> Void, elseBody: () -> Void) {
        emit(BeginIf(inverted: false), withInputs: [condition])
        ifBody()
        emit(BeginElse())
        elseBody()
        emit(EndIf())
    }

    public class SwitchBlock {
        private let b: ProgramBuilder
        public fileprivate(set) var hasDefaultCase: Bool = false

        fileprivate init(in b: ProgramBuilder) {
            assert(b.context.contains(.switchBlock))
            self.b = b
        }

        public func addDefaultCase(fallsThrough: Bool = false, body: @escaping () -> ()) {
            b.emit(BeginSwitchDefaultCase(), withInputs: [])
            body()
            b.emit(EndSwitchCase(fallsThrough: fallsThrough))
        }

        public func addCase(_ v: Variable, fallsThrough: Bool = false, body: @escaping () -> ()) {
            b.emit(BeginSwitchCase(), withInputs: [v])
            body()
            b.emit(EndSwitchCase(fallsThrough: fallsThrough))
        }
    }

    public func buildSwitch(on switchVar: Variable, body: (SwitchBlock) -> ()) {
        emit(BeginSwitch(), withInputs: [switchVar])
        body(currentSwitchBlock)
        emit(EndSwitch())
    }

    public func switchBreak() {
        emit(SwitchBreak())
    }

    public func buildWhileLoop(_ header: () -> Variable, _ body: () -> Void) {
        emit(BeginWhileLoopHeader())
        let cond = header()
        emit(BeginWhileLoopBody(), withInputs: [cond])
        body()
        emit(EndWhileLoop())
    }

    public func buildDoWhileLoop(do body: () -> Void, while header: () -> Variable) {
        emit(BeginDoWhileLoopBody())
        body()
        emit(BeginDoWhileLoopHeader())
        let cond = header()
        emit(EndDoWhileLoop(), withInputs: [cond])
    }

    // Build a simple for loop that declares one loop variable.
    public func buildForLoop(i initializer: () -> Variable, _ cond: (Variable) -> Variable, _ afterthought: (Variable) -> (), _ body: (Variable) -> ()) {
        emit(BeginForLoopInitializer())
        let initialValue = initializer()
        var loopVar = emit(BeginForLoopCondition(numLoopVariables: 1), withInputs: [initialValue]).innerOutput
        let cond = cond(loopVar)
        loopVar = emit(BeginForLoopAfterthought(numLoopVariables: 1), withInputs: [cond]).innerOutput
        afterthought(loopVar)
        loopVar = emit(BeginForLoopBody(numLoopVariables: 1)).innerOutput
        body(loopVar)
        emit(EndForLoop())
    }

    // Build arbitrarily complex for loops without any loop variables.
    public func buildForLoop(_ initializer: (() -> ())? = nil, _ cond: (() -> Variable)? = nil, _ afterthought: (() -> ())? = nil, _ body: () -> ()) {
        emit(BeginForLoopInitializer())
        initializer?()
        emit(BeginForLoopCondition(numLoopVariables: 0))
        let cond = cond?() ?? loadBool(true)
        emit(BeginForLoopAfterthought(numLoopVariables: 0), withInputs: [cond])
        afterthought?()
        emit(BeginForLoopBody(numLoopVariables: 0))
        body()
        emit(EndForLoop())
    }

    // Build arbitrarily complex for loops with one or more loop variables.
    public func buildForLoop(_ initializer: () -> [Variable], _ cond: (([Variable]) -> Variable)? = nil, _ afterthought: (([Variable]) -> ())? = nil, _ body: ([Variable]) -> ()) {
        emit(BeginForLoopInitializer())
        let initialValues = initializer()
        var loopVars = emit(BeginForLoopCondition(numLoopVariables: initialValues.count), withInputs: initialValues).innerOutputs
        let cond = cond?(Array(loopVars)) ?? loadBool(true)
        loopVars = emit(BeginForLoopAfterthought(numLoopVariables: initialValues.count), withInputs: [cond]).innerOutputs
        afterthought?(Array(loopVars))
        loopVars = emit(BeginForLoopBody(numLoopVariables: initialValues.count)).innerOutputs
        body(Array(loopVars))
        emit(EndForLoop())
    }

    public func buildForInLoop(_ obj: Variable, _ body: (Variable) -> ()) {
        let i = emit(BeginForInLoop(), withInputs: [obj]).innerOutput
        body(i)
        emit(EndForInLoop())
    }

    public func buildForOfLoop(_ obj: Variable, _ body: (Variable) -> ()) {
        let i = emit(BeginForOfLoop(), withInputs: [obj]).innerOutput
        body(i)
        emit(EndForOfLoop())
    }

    public func buildForOfLoop(_ obj: Variable, selecting indices: [Int64], hasRestElement: Bool = false, _ body: ([Variable]) -> ()) {
        let instr = emit(BeginForOfLoopWithDestruct(indices: indices, hasRestElement: hasRestElement), withInputs: [obj])
        body(Array(instr.innerOutputs))
        emit(EndForOfLoop())
    }

    public func buildRepeatLoop(n numIterations: Int, _ body: (Variable) -> ()) {
        let i = emit(BeginRepeatLoop(iterations: numIterations)).innerOutput
        body(i)
        emit(EndRepeatLoop())
    }

    public func buildRepeatLoop(n numIterations: Int, _ body: () -> ()) {
        emit(BeginRepeatLoop(iterations: numIterations, exposesLoopCounter: false))
        body()
        emit(EndRepeatLoop())
    }

    public func loopBreak() {
        emit(LoopBreak())
    }

    public func loopContinue() {
        emit(LoopContinue(), withInputs: [])
    }

    public func buildTryCatchFinally(tryBody: () -> (), catchBody: ((Variable) -> ())? = nil, finallyBody: (() -> ())? = nil) {
        assert(catchBody != nil || finallyBody != nil, "Must have either a Catch or a Finally block (or both)")
        emit(BeginTry())
        tryBody()
        if let catchBody = catchBody {
            let exception = emit(BeginCatch()).innerOutput
            catchBody(exception)
        }
        if let finallyBody = finallyBody {
            emit(BeginFinally())
            finallyBody()
        }
        emit(EndTryCatchFinally())
    }

    public func throwException(_ value: Variable) {
        emit(ThrowException(), withInputs: [value])
    }

    public func buildCodeString(_ body: () -> ()) -> Variable {
        let instr = emit(BeginCodeString())
        body()
        emit(EndCodeString())
        return instr.output
    }

    public func blockStatement(_ body: () -> Void) {
        emit(BeginBlockStatement())
        body()
        emit(EndBlockStatement())
    }

    public func doPrint(_ value: Variable) {
        emit(Print(), withInputs: [value])
    }


    /// Returns the next free variable.
    func nextVariable() -> Variable {
        assert(numVariables < Code.maxNumberOfVariables, "Too many variables")
        numVariables += 1
        return Variable(number: numVariables - 1)
    }

    @discardableResult
    private func internalAppend(_ instr: Instruction) -> Instruction {
        // Basic integrity checking
        assert(!instr.inouts.contains(where: { $0.number >= numVariables }))
        assert(instr.op.requiredContext.isSubset(of: contextAnalyzer.context))

        // The returned instruction will also contain its index in the program. Use that so the analyzers have access to the index.
        let instr = code.append(instr)
        analyze(instr)

        return instr
    }

    /// Set the parameter types for the next function, method, or constructor, which must be the the start of a function or method definition.
    /// Parameter types (and signatures in general) are only valid for the duration of the program generation, as they cannot be preserved across mutations.
    /// As such, the parameter types are linked to their instruction through the index of the instruction in the program.
    private func setParameterTypesForNextSubroutine(_ parameterTypes: ParameterList) {
        jsTyper.setParameters(forSubroutineStartingAt: code.count, to: parameterTypes)
    }

    /// Analyze the given instruction. Should be called directly after appending the instruction to the code.
    private func analyze(_ instr: Instruction) {
        assert(code.lastInstruction.op === instr.op)
        updateVariableAnalysis(instr)
        contextAnalyzer.analyze(instr)
        updateBuilderState(instr)
        jsTyper.analyze(instr)
    }

    private func updateVariableAnalysis(_ instr: Instruction) {
        // Scope management (1).
        if instr.isBlockEnd {
            assert(scopes.count > 0, "Trying to close a scope that was never opened")
            let current = scopes.pop()
            // Hidden variables that go out of scope need to be unhidden.
            for v in current where hiddenVariables.contains(v) {
                unhide(v)
            }
            variablesInScope.removeLast(current.count)
        }

        scopes.top.append(contentsOf: instr.outputs)
        variablesInScope.append(contentsOf: instr.outputs)

        // Scope management (2). Happens here since e.g. function definitions create a variable in the outer scope.
        if instr.isBlockStart {
            scopes.push([])
        }

        scopes.top.append(contentsOf: instr.innerOutputs)
        variablesInScope.append(contentsOf: instr.innerOutputs)
    }

    private func updateBuilderState(_ instr: Instruction) {
        switch instr.op.opcode {
        case .beginObjectLiteral:
            activeObjectLiterals.push(ObjectLiteral(in: self))
        case .objectLiteralAddProperty(let op):
            currentObjectLiteral.properties.append(op.propertyName)
        case .objectLiteralAddElement(let op):
            currentObjectLiteral.elements.append(op.index)
        case .objectLiteralAddComputedProperty:
            currentObjectLiteral.computedProperties.append(instr.input(0))
        case .objectLiteralCopyProperties:
            // Cannot generally determine what fields this installs.
            break
        case .objectLiteralSetPrototype:
            currentObjectLiteral.hasPrototype = true
        case .beginObjectLiteralMethod(let op):
            currentObjectLiteral.methods.append(op.methodName)
        case .beginObjectLiteralComputedMethod:
            currentObjectLiteral.computedMethods.append(instr.input(0))
        case .beginObjectLiteralGetter(let op):
            currentObjectLiteral.getters.append(op.propertyName)
        case .beginObjectLiteralSetter(let op):
            currentObjectLiteral.setters.append(op.propertyName)
        case .endObjectLiteralMethod,
                .endObjectLiteralComputedMethod,
                .endObjectLiteralGetter,
                .endObjectLiteralSetter:
            break
        case .endObjectLiteral:
            activeObjectLiterals.pop()

        case .beginClassDefinition(let op):
            activeClassDefinitions.push(ClassDefinition(in: self, isDerived: op.hasSuperclass))
        case .beginClassConstructor:
            activeClassDefinitions.top.hasConstructor = true
        case .classAddInstanceProperty(let op):
            activeClassDefinitions.top.instanceProperties.append(op.propertyName)
        case .classAddInstanceElement(let op):
            activeClassDefinitions.top.instanceElements.append(op.index)
        case .classAddInstanceComputedProperty:
            activeClassDefinitions.top.instanceComputedProperties.append(instr.input(0))
        case .beginClassInstanceMethod(let op):
            activeClassDefinitions.top.instanceMethods.append(op.methodName)
        case .beginClassInstanceGetter(let op):
            activeClassDefinitions.top.instanceGetters.append(op.propertyName)
        case .beginClassInstanceSetter(let op):
            activeClassDefinitions.top.instanceSetters.append(op.propertyName)
        case .classAddStaticProperty(let op):
            activeClassDefinitions.top.staticProperties.append(op.propertyName)
        case .classAddStaticElement(let op):
            activeClassDefinitions.top.staticElements.append(op.index)
        case .classAddStaticComputedProperty:
            activeClassDefinitions.top.staticComputedProperties.append(instr.input(0))
        case .beginClassStaticMethod(let op):
            activeClassDefinitions.top.staticMethods.append(op.methodName)
        case .beginClassStaticGetter(let op):
            activeClassDefinitions.top.staticGetters.append(op.propertyName)
        case .beginClassStaticSetter(let op):
            activeClassDefinitions.top.staticSetters.append(op.propertyName)
        case .classAddPrivateInstanceProperty(let op):
            activeClassDefinitions.top.privateProperties.append(op.propertyName)
        case .beginClassPrivateInstanceMethod(let op):
            activeClassDefinitions.top.privateMethods.append(op.methodName)
        case .classAddPrivateStaticProperty(let op):
            activeClassDefinitions.top.privateProperties.append(op.propertyName)
        case .beginClassPrivateStaticMethod(let op):
            activeClassDefinitions.top.privateMethods.append(op.methodName)
        case .beginClassStaticInitializer:
            break
        case .endClassDefinition:
            activeClassDefinitions.pop()

        case .beginSwitch:
            activeSwitchBlocks.push(SwitchBlock(in: self))
        case .beginSwitchDefaultCase:
            activeSwitchBlocks.top.hasDefaultCase = true
        case .beginSwitchCase:
            break
        case .endSwitch:
            activeSwitchBlocks.pop()

        default:
            assert(!instr.op.requiredContext.contains(.objectLiteral))
            assert(!instr.op.requiredContext.contains(.classDefinition))
            assert(!instr.op.requiredContext.contains(.switchBlock))
            break
        }
    }
}