Tokenizers (lexers)

A tokenizer or a lexer is a program that breaks down an input text into smaller chunks, and classifies them as one of several tokens. For example, consider an imagininary format that only consists of nested tuples of strings containing letters, like this:

(("ABC", "v"),(("x", ("pj",(("a", "k")), ("L")))))

Any text of this format can be broken down into a sequence of the following tokens:

• Left parenthesis: re"\("
• Right parenthesis: re"\)"
• Comma: re","
• Quote: re"\""
• Spaces: re" +"
• Letters: re"[A-Za-z]+"

Such that e.g. ("XY", "A") can be represented as lparent, quote, XY, quote, comma, space, quote A quote rparens.

Breaking the text down to its tokens is called tokenization or lexing. Note that lexing in itself is not sufficient to parse the format: Lexing is context unaware, so e.g. the test "((A can be perfectly well tokenized to quote lparens lparens A, even if it's invalid.

The purpose of tokenization is to make subsequent parsing easier, because each part of the text has been classified. That makes it easier to, for example, to search for letters in the input. Instead of having to muck around with regex to find the letters, you use regex once to classify all text.

Making and using a tokenizer

Let's use the example above to create a tokenizer. The most basic default tokenizer uses UInt32 as tokens: You pass in a list of regex matching each token, then evaluate the resulting code:

julia> make_tokenizer(
           [re"\(", re"\)", re",", re"\"", re" +", re"[a-zA-Z]+"]
       ) |> eval

Since the default tokenizer uses UInt32 as tokens, you can then obtain a lazy iterator of tokens by calling tokenize(UInt32, data):

julia> iterator = tokenize(UInt32, """("XY", "A")"""); typeof(iterator)
Tokenizer{UInt32, String, 1}

This will return Tuple{Int64, Int32, UInt32} elements, with each element being:

• The start index of the token
• The length of the token
• The token itself, in this example UInt32(1) for '(', UInt32(2) for ')' etc:

julia> collect(iterator)
10-element Vector{Tuple{Int64, Int32, UInt32}}:
 (1, 1, 0x00000001)
 (2, 1, 0x00000004)
 (3, 2, 0x00000006)
 (5, 1, 0x00000004)
 (6, 1, 0x00000003)
 (7, 1, 0x00000005)
 (8, 1, 0x00000004)
 (9, 1, 0x00000006)
 (10, 1, 0x00000004)
 (11, 1, 0x00000002)

Any data which could not be tokenized is given the error token UInt32(0):

julia> collect(tokenize(UInt32, "XY!!)"))
3-element Vector{Tuple{Int64, Int32, UInt32}}:
 (1, 2, 0x00000006)
 (3, 2, 0x00000000)
 (5, 1, 0x00000002)

Both tokenize and make_tokenizer takes an optional argument version, which is 1 by default. This sets the last parameter of the Tokenizer struct, and as such allows you to create multiple different tokenizers with the same element type.

Using enums as tokens

Using UInt32 as tokens is not very convenient - so it's possible to use enums to create the tokenizer:

julia> @enum Token error lparens rparens comma quot space letters

julia> make_tokenizer((error, [
           lparens => re"\(",
           rparens => re"\)",
           comma => re",",
           quot => re"\"",
           space => re" +",
           letters => re"[a-zA-Z]+"
        ])) |> eval

julia> collect(tokenize(Token, "XY!!)"))
3-element Vector{Tuple{Int64, Int32, Token}}:
 (1, 2, letters)
 (3, 2, error)
 (5, 1, rparens)

To make it even easier, you can define the enum and the tokenizer in one go:

tokens = [
    :lparens => re"\(",
    :rparens => re"\)",
    :comma => re",",
    :quot => re"\"",
    :space => re" +",
    :letters => re"[a-zA-Z]+"
]
@eval @enum Token error $(first.(tokens)...)
make_tokenizer((error, 
    [Token(i) => j for (i,j) in enumerate(last.(tokens))]
)) |> eval

Token disambiguation

It's possible to create a tokenizer where the different token regexes overlap:

julia> make_tokenizer([re"[ab]+", re"ab*", re"ab"]) |> eval

In this case, an input like ab will match all three regex. Which tokens are emitted is determined by two rules:

First, the emitted tokens will be as long as possible. So, the input aa could be emitted as one token of the regex re"[ab]+", two tokens of the same regex, or of two tokens of the regex re"ab*". In this case, it will be emitted as a single token of re"[ab]+", since that will make the first token as long as possible (2 bytes), whereas the other options would only make it 1 byte long.

Second, tokens with a higher index in the input array beats previous tokens. So, a will be emitted as re"ab*", as its index of 2 beats the previous regex re"[ab]+" with the index 1, and ab will match the third regex.

If you don't want emitted tokens to depend on these priority rules, you can set the optional keyword unambiguous=true in the make_tokenizer function, in which case make_tokenizer will error if any input text could be broken down into different tokens. However, note that this may cause most tokenizers to error when being built, as most tokenization processes are ambiguous.

Reference

Tokenizer{E, D, C}

tokenize(::Type{E}, data, version=1)

make_tokenizer(
    machine::TokenizerMachine;
    tokens::Tuple{E, AbstractVector{E}}= [ integers ],
    goto=true, version=1
) where E

julia> machine = compile([re"a", re"b"]);

julia> make_tokenizer(machine; tokens=(0x00, [0x01, 0x02])) |> eval

julia> iter = tokenize(UInt8, "abxxxba"); typeof(iter)
Tokenizer{UInt8, String, 1}

julia> collect(iter)
5-element Vector{Tuple{Int64, Int32, UInt8}}:
 (1, 1, 0x01)
 (2, 1, 0x02)
 (3, 3, 0x00)
 (6, 1, 0x02)
 (7, 1, 0x01)

make_tokenizer(
    tokens::Union{
        AbstractVector{RE},
        Tuple{E, AbstractVector{Pair{E, RE}}}
    };
    goto::Bool=true,
    version::Int=1,
    unambiguous=false
) where E

julia> make_tokenizer([re"abc", re"def") |> eval

julia> collect(tokenize(Int, "abcxyzdef123"))
4-element Vector{Tuple{Int64, Int32, UInt32}}:
 (1, 3, 0x00000001)
 (4, 3, 0x00000003)
 (7, 3, 0x00000002)
 (10, 3, 0x00000003)