Introducing Citron

An LALR parser generator for Swift

I’d like to introduce you to Citron, an LALR parser generator for Swift that I’ve been working on. For a given input grammar, Citron creates a reentrant type-safe parser in Swift.

Citron is based on the Lemon parser generator by Richard Hipp that’s used in SQLite. The parser generation engine of Citron is written in C, as is the case with Lemon, but the output parser code is all Swift.

What’s different

Citron uses the same grammar specification syntax used by Lemon, which is a little different from what Bison and Yacc use. In addition, Citron is stricter with code blocks and types than Bison, Yacc or Lemon.

In Citron:

  • All terminals and non-terminals in the grammar must have explicitly defined types

  • Every grammar rule should be accompanied by an associated code block that will be invoked when the rule gets used during parsing

  • The code block associated with a rule should take in values of the correct type as input, and should return a value of the correct type

Writing a rule

Let’s say you have a grammar rule for Swift’s let statement that can be expressed in BNF like this:

<let-decl> ::= "let" <id> ":" <type-decl>

An equivalent rule in a Citron grammar file could look like this:

let_decl ::= LET ID COLON type_decl.

In Citron, terminals (a.k.a. tokens) start with an uppercase letter (like LET, ID and COLON), and non-terminals start with a lowecase letter (like let_decl and type_decl). “let”, “:” and the identifier are expected to be recognized at tokenization stage and converted to tokens LET, COLON and ID respectively, so the parser only sees these tokens.

If one were to create a parse tree from the input, there would be semantic types representing the terminals and non-terminals present in the grammar. A let node of the parse tree could be represented with something like:

class LetDeclaration {
    var id: String
    var type: TypeDeclaration
    init(id: String, type: TypeDeclaration) { = id
        self.type = type

If that’s the case, you should tell Citron about these backing types. It’s a good practice to declare the type of the terminals before the rules, and the declare the types of the non-terminals close to where the they occur in a rule, like this:

%token_type String // type representing LET, ID and COLON

%nonterminal_type let_decl LetDeclaration
%nonterminal_type type_decl TypeDeclaration

let_decl ::= LET ID COLON type_decl.

Citron expects a code block immediately after a rule. The code block should form the body of a function that returns a value representing the LHS symbol of the rule. The LHS symbol of this rule is let_decl, so our code block should return something of type LetDeclaration.

We can add alias names to the RHS symbols as shown below, and they will be made available as named function parameters within the code block with the appropriate type.

%token_type String
%nonterminal_type let_decl LetDeclaration
%nonterminal_type type_decl TypeDeclaration

let_decl ::= LET ID(n) COLON type_decl(t). {
    // 'n' of type String and 't' of type TypeDeclaration
    // are available for use in this code block
    return LetDeclaration(id: n, type: t)

Now we have a valid Citron rule – the rule is immediately followed by a code block, and the code block takes in and returns values of correct types.

You can find an example grammar file here, which has a set of rules that describe simple arithmetic expressions.

In the Citron-generated code, the code block is used as the body of a function. The type signature of the function is based on the symbol on the LHS and the aliased symbols on the RHS. For the above code block, the Citron-generated function would look something like this:

func codeBlockForRule1(n: String, t: TypeDeclaration) -> LetDeclaration {
    // 'n' of type String and 't' of type TypeDeclaration
    // are available for use in this code block
    return LetDeclaration(name: n, type: t)

Unless the aliased symbols’ types and the return type used in the code block are consistent with the rule, the generated parser code will not compile. This type safety helps catch code block errors at build time rather than at run time.

The parser interface

The Citron-generated code contains a parser class, named Parser by default, with two parsing methods:

  1. consume(token:code:)

    This methods makes the parser consume one token. This should be called multiple times to pass a sequence of tokens to the parser. Typically, a separate tokenization stage would generate this sequence of tokens. As the tokens get generated, they can be passed to the parser through this method.

    The first argument, token:, is the semantic value of the token, as seen by the code blocks in the grammar. The type of this argument is the %token_type type specified in the grammar.

    The second argument, code:, is the token code that Citron knows this token by. The Citron-generated class contains an enum of all terminals in the grammar, and this argument should be a value of that enum.

    For example, the “n” in “let n: Int” would be passed to the parser like this: parser.consume(token: "n", code: .ID).

  2. endParsing()

    This method tells the parser that there are no more tokens to consume, and signifies the end of input.

    This method takes no arguments, and returns the parse result. The parse result is the value returned by the code block of the start rule (a.k.a. root rule) of the grammar. If we’re building a parse tree using the code blocks as illustrated above, the parse result would typically be the parse tree.

    The return type of this method is the %nonterminal_type of the start symbol (a.k.a. root symbol) of the grammar.

The string “let n: Int” could get tokenized into four tokens and passed to Citron as:

let parser = Parser()
do {
    try parser.consume(token: "let", code: .LET)
    try parser.consume(token: "n", code: .ID)
    try parser.consume(token: ":", code: .COLON)
    try parser.consume(token: "Int", code: .BUILTIN_TYPE)
    let parseTree = try parser.endParsing()
} catch {
    // Handle errors

Error handling

Both consume(token:code:) and endParsing() are throwing methods, so they should be invoked with a preceding try. They throw when an input token at a certain position is inconsistent with the grammar, or when the input ends prematurely.

Moreover, we can throw errors from within a rule’s code block and those throws would propagate up to one of the two parsing methods.

Update: In Citron 2.0, you can use error capturing for more advanced error handling.

The lexer interface

Citron offers a simple lexer that can tokenize an input string.

We give the lexer a series of rules with either string or regular expression patterns. The token data that the lexer should output can be of an arbitrary type, so the lexer is defined as a generic type, with the token data as a type parameter. For each string pattern, we give it the token data that should be output, and for each regular expression pattern, we give it a closure that returns the token data.

For example, you could say:

// A tuple with all the data we need for a token
typealias TokenData = (String, Parser.CitronTokenCode)

// A lexer that can output tokens as the tuple we just defined
typealias Lexer = CitronLexer<TokenData>

// Create the lexer with tokenization rules
let lexer = Lexer(rules: [

        // Keywords

        .string("let", ("", .LET)),
        .string("var", ("", .VAR)),
        .string("class", ("", .CLASS)),

        // Punctuation

        .string("{", ("", .OPEN_FLOWER_BRACKET)),
        .string("}", ("", .CLOSE_FLOWER_BRACKET)),
        .string(":", ("", .COLON)),

        // Built-in types

        .string("Int", ("", code: .INT)),
        .string("String", ("", code: .STRING)),

        // Identifiers

        .regexPattern("[A-Za-z][0-9A-Za-z]*", { str in (str, .ID) }),

        // Whitespace

        .regexPattern("\\s", { _ in nil }) // Ignored

We can ask the lexer to tokenize a string and give it a closure to run on each token it identifies. The token data gets passed into the closure as a parameter. Inside the closure, we can pass on the token data to the parser.

do {
    try lexer.tokenize(inputString) { (t, c) in
        try parser.consume(token: t, code: c)
    let parseTree = try parser.endParsing()
} catch {
    print("Error during parsing")

Try it

The Citron code is on Github, along with a few examples. There’s also a more detailed documentation.

I’d be happy to answer questions or take feedback on Twitter or by email.

Search the internet without giving up your privacy

Hide & Seek is a Safari content blocker I made that lets you search Google and Bing as a logged-out user, while staying logged-in in other Google and Microsoft services.