A dead simple parser package for Go

Introduction
Limitations
Tutorial
Overview
Annotation syntax
Capturing
Streaming
Lexing
Options
Examples
Performance

Introduction

The goal of this package is to provide a simple, idiomatic and elegant way of defining parsers in Go.

Participle's method of defining grammars should be familiar to any Go programmer who has used the encoding/json package: struct field tags define what and how input is mapped to those same fields. This is not unusual for Go encoders, but is unusual for a parser.

Limitations

Participle parsers are LL(k). Among other things, this means that they do not support left recursion.

The default value of K is 1 but this can be controlled with participle.UseLookahead(k).

Left recursion must be eliminated by restructuring your grammar.

Tutorial

A tutorial is available, walking through the creation of an .ini parser.

Overview

A grammar is an annotated Go structure used to both define the parser grammar, and be the AST output by the parser. As an example, following is the final INI parser from the tutorial.

type INI struct {
  Properties []*Property `{ @@ }`
  Sections   []*Section  `{ @@ }`
}

type Section struct {
  Identifier string      `"[" @Ident "]"`
  Properties []*Property `{ @@ }`
}

type Property struct {
  Key   string `@Ident "="`
  Value *Value `@@`
}

type Value struct {
  String *string  `  @String`
  Number *float64 `| @Float`
}

Note: Participle also supports named struct tags (eg. Hello string `parser:"@Ident"`).

A parser is constructed from a grammar and a lexer:

parser, err := participle.Build(&INI{})

Once constructed, the parser is applied to input to produce an AST:

ast := &INI{}
err := parser.ParseString("size = 10", ast)
// ast == &INI{
//   Properties: []*Property{
//     {Key: "size", Value: &Value{Number: &10}},
//   },
// }

Annotation syntax

@<expr> Capture expression into the field.
@@ Recursively capture using the fields own type.
<identifier> Match named lexer token.
( ... ) Group.
"..." Match the literal (note that the lexer must emit tokens matching this literal exactly).
"...":<identifier> Match the literal, specifying the exact lexer token type to match.
<expr> <expr> ... Match expressions.
<expr> | <expr> Match one of the alternatives.

The following modifiers can be used after any expression:

* Expression can match zero or more times.
+ Expression must match one or more times.
? Expression can match zero or once.
! Require a non-empty match (this is useful with a sequence of optional matches eg. ("a"? "b"? "c"?)!).

Supported but deprecated:

{ ... } Match 0 or more times (DEPRECATED - prefer ( ... )*).
[ ... ] Optional (DEPRECATED - prefer ( ... )?).

Notes:

Each struct is a single production, with each field applied in sequence.
@<expr> is the mechanism for capturing matches into the field.
if a struct field is not keyed with "parser", the entire struct tag will be used as the grammar fragment. This allows the grammar syntax to remain clear and simple to maintain.

Capturing

Prefixing any expression in the grammar with @ will capture matching values for that expression into the corresponding field.

For example:

// The grammar definition.
type Grammar struct {
  Hello string `@Ident`
}

// The source text to parse.
source := "world"

// After parsing, the resulting AST.
result == &Grammar{
  Hello: "world",
}

For slice and string fields, each instance of @ will accumulate into the field (including repeated patterns). Accumulation into other types is not supported.

A successful capture match into a boolean field will set the field to true.

For integer and floating point types, a successful capture will be parsed with strconv.ParseInt() and strconv.ParseBool() respectively.

Custom control of how values are captured into fields can be achieved by a field type implementing the Capture interface (Capture(values []string) error).

Streaming

Participle supports streaming parsing. Simply pass a channel of your grammar into Parse*(). The grammar will be repeatedly parsed and sent to the channel. Note that the Parse*() call will not return until parsing completes, so it should generally be started in a goroutine.

type token struct {
  Str string `  @Ident`
  Num int    `| @Int`
}

parser, err := participle.Build(&token{})

tokens := make(chan *token, 128)
err := parser.ParseString(`hello 10 11 12 world`, tokens)
for token := range tokens {
  fmt.Printf("%#v\n", token)
}

Lexing

Participle operates on tokens and thus relies on a lexer to convert character streams to tokens.

Three lexers are provided, varying in speed and flexibility. The fastest lexer is based on the text/scanner package but only allows tokens provided by that package. Next fastest is the regexp lexer (lexer.Regexp()). The slowest is currently the EBNF based lexer, but it has a large potential for optimisation through code generation.

To use your own Lexer you will need to implement two interfaces: Definition and Lexer.

Options

The Parser's behaviour can be configured via Options.

Examples

There are several examples included:

Example	Description
BASIC	A lexer, parser and interpreter for a rudimentary dialect of BASIC.
EBNF	Parser for the form of EBNF used by Go.
Expr	A basic mathematical expression parser and evaluator.
GraphQL	Lexer+parser for GraphQL schemas
HCL	A parser for the HashiCorp Configuration Language.
INI	An INI file parser.
Protobuf	A full Protobuf version 2 and 3 parser.
SQL	A very rudimentary SQL SELECT parser.
Thrift	A full Thrift parser.
TOML	A TOML parser.

Included below is a full GraphQL lexer and parser:

package main

import (
  "os"

  "github.com/alecthomas/kong"
  "github.com/alecthomas/repr"

  "github.com/alecthomas/participle"
  "github.com/alecthomas/participle/lexer"
  "github.com/alecthomas/participle/lexer/ebnf"
)

type File struct {
  Entries []*Entry `@@*`
}

type Entry struct {
  Type   *Type   `  @@`
  Schema *Schema `| @@`
  Enum   *Enum   `| @@`
  Scalar string  `| "scalar" @Ident`
}

type Enum struct {
  Name  string   `"enum" @Ident`
  Cases []string `"{" @Ident* "}"`
}

type Schema struct {
  Fields []*Field `"schema" "{" @@* "}"`
}

type Type struct {
  Name       string   `"type" @Ident`
  Implements string   `("implements" @Ident)?`
  Fields     []*Field `"{" @@* "}"`
}

type Field struct {
  Name       string      `@Ident`
  Arguments  []*Argument `("(" (@@ ("," @@)*)? ")")?`
  Type       *TypeRef    `":" @@`
  Annotation string      `("@" @Ident)?`
}

type Argument struct {
  Name    string   `@Ident`
  Type    *TypeRef `":" @@`
  Default *Value   `("=" @@)?`
}

type TypeRef struct {
  Array       *TypeRef `(   "[" @@ "]"`
  Type        string   `  | @Ident )`
  NonNullable bool     `@"!"?`
}

type Value struct {
  Symbol string `@Ident`
}

var (
  graphQLLexer = lexer.Must(ebnf.New(`
    Comment = ("#" | "//") { "\u0000"…"\uffff"-"\n" } .
    Ident = (alpha | "_") { "_" | alpha | digit } .
    Number = ("." | digit) {"." | digit} .
    Whitespace = " " | "\t" | "\n" | "\r" .
    Punct = "!"…"/" | ":"…"@" | "["…`+"\"`\""+` | "{"…"~" .

    alpha = "a"…"z" | "A"…"Z" .
    digit = "0"…"9" .
`))

  parser = participle.MustBuild(&File{},
    participle.Lexer(graphQLLexer),
    participle.Elide("Comment", "Whitespace"),
    )

  cli struct {
    Files []string `arg:"" type:"existingfile" required:"" help:"GraphQL schema files to parse."`
  }
)

func main() {
  ctx := kong.Parse(&cli)
  for _, file := range cli.Files {
    ast := &File{}
    r, err := os.Open(file)
    ctx.FatalIfErrorf(err)
    err = parser.Parse(r, ast)
    r.Close()
    repr.Println(ast)
    ctx.FatalIfErrorf(err)
  }
}

Performance

One of the included examples is a complete Thrift parser (shell-style comments are not supported). This gives a convenient baseline for comparing to the PEG based pigeon, which is the parser used by go-thrift. Additionally, the pigeon parser is utilising a generated parser, while the participle parser is built at run time.

You can run the benchmarks yourself, but here's the output on my machine:

BenchmarkParticipleThrift-4        10000      221818 ns/op     48880 B/op     1240 allocs/op
BenchmarkGoThriftParser-4           2000      804709 ns/op    170301 B/op     3086 allocs/op

On a real life codebase of 47K lines of Thrift, Participle takes 200ms and go- thrift takes 630ms, which aligns quite closely with the benchmarks.

OsQu/participle