A dead simple parser package for Go

V2
Introduction
Tutorial
Tag syntax
Overview
Grammar syntax
Capturing
- Capturing boolean value
Streaming
Lexing
Options
Examples
Performance
Concurrency
Error reporting
Limitations
EBNF
Syntax/Railroad Diagrams

V2

This is version 2 of Participle. See the Change Log for details.

Note: semantic versioning API guarantees do not apply to the experimental packages - the API may break between minor point releases.

It can be installed with:

$ go get github.com/alecthomas/participle/v2@latest

The latest version from v0 can be installed via:

$ go get github.com/alecthomas/participle@latest

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.

Tutorial

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

Tag syntax

Participle supports two forms of struct tag grammar syntax.

The easiest to read is when the grammar uses the entire struct tag content, eg.

Field string `@Ident @("," Ident)*`

However, this does not coexist well with other tags such as JSON, etc. and may cause issues with linters. If this is an issue then you can use the parser:"" tag format. In this case single quotes can be used to quote literals making the tags somewhat easier to write, eg.

Field string `parser:"@ident (',' Ident)*" json:"field"`

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}},
//   },
// }

Grammar syntax

Participle grammars are defined as tagged Go structures. Participle will first look for tags in the form parser:"...". It will then fall back to using the entire tag body.

The grammar format is:

@<expr> Capture expression into the field.
@@ Recursively capture using the fields own type.
<identifier> Match named lexer token.
( ... ) Group.
"..." or '...' 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. Each alternative is tried in order, with backtracking.
!<expr> Match any token that is not the start of the expression (eg: @!";" matches anything but the ; character into the field).

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"?)!).

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.

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

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

Tokens can also be captured directly into fields of type lexer.Token and []lexer.Token.

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

Additionally, any field implementing the encoding.TextUnmarshaler interface will be capturable too. One caveat is that UnmarshalText() will be called once for each captured token, so eg. @(Ident Ident Ident) will be called three times.

Capturing boolean value

By default a boolean field is used to indicate that a match occurred, which turns out to be much more useful and common in Participle than parsing true or false literals. For example, parsing a variable declaration with a trailing optional syntax:

type Var struct {
  Name string `"var" @Ident`
  Type string `":" @Ident`
  Optional bool `@"?"?`
}

In practice this gives more useful AST's. If bool were to be parsed literally then you'd need to have some alternate type for Optional such as string or a custom type.

To capture literal boolean values such as true or false, implement the Capture interface like so:

type Boolean bool

func (b *Boolean) Capture(values []string) error {
	*b = values[0] == "true"
	return nil
}

type Value struct {
	Float  *float64 `  @Float`
	Int    *int     `| @Int`
	String *string  `| @String`
	Bool   *Boolean `| @("true" | "false")`
}

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 relies on distinct lexing and parsing phases. The lexer takes raw bytes and produces tokens which the parser consumes. The parser transforms these tokens into Go values.

The default lexer, if one is not explicitly configured, is based on the Go text/scanner package and thus produces tokens for C/Go-like source code. This is surprisingly useful, but if you do require more control over lexing the builtin participle/lexer/stateful lexer should cover most other cases. If that in turn is not flexible enough, you can implement your own lexer.

Configure your parser with a lexer using the participle.Lexer() option.

To use your own Lexer you will need to implement two interfaces: Definition (and optionally StringsDefinition and BytesDefinition) and Lexer.

Stateful lexer

Participle's included stateful/modal lexer provides powerful yet convenient construction of most lexers (notably, indentation based lexers cannot be expressed).

It is sometimes the case that a simple lexer cannot fully express the tokens required by a parser. The canonical example of this is interpolated strings within a larger language. eg.

let a = "hello ${name + ", ${last + "!"}"}"

This is impossible to tokenise with a normal lexer due to the arbitrarily deep nesting of expressions.

To support this case Participle's lexer is now stateful by default.

The lexer is a state machine defined by a map of rules keyed by the state name. Each rule within the state includes the name of the produced token, the regex to match, and an optional operation to apply when the rule matches.

As a convenience, any Rule starting with a lowercase letter will be elided from output.

Lexing starts in the Root group. Each rule is matched in order, with the first successful match producing a lexeme. If the matching rule has an associated Action it will be executed. The name of each non-root rule is prefixed with the name of its group to yield the token identifier used during matching.

A state change can be introduced with the Action Push(state). Pop() will return to the previous state.

To reuse rules from another state, use Include(state).

A special named rule Return() can also be used as the final rule in a state to always return to the previous state.

As a special case, regexes containing backrefs in the form \N (where N is a digit) will match the corresponding capture group from the immediate parent group. This can be used to parse, among other things, heredocs. See the tests for an example of this, among others.

Example stateful lexer

Here's a cut down example of the string interpolation described above. Refer to the stateful example for the corresponding parser.

var lexer = stateful.Must(Rules{
	"Root": {
		{`String`, `"`, Push("String")},
	},
	"String": {
		{"Escaped", `\\.`, nil},
		{"StringEnd", `"`, Pop()},
		{"Expr", `\${`, Push("Expr")},
		{"Char", `[^$"\\]+`, nil},
	},
	"Expr": {
		Include("Root"),
		{`whitespace`, `\s+`, nil},
		{`Oper`, `[-+/*%]`, nil},
		{"Ident", `\w+`, nil},
		{"ExprEnd", `}`, Pop()},
	},
})

Example simple/non-stateful lexer

The Stateful lexer is now the only custom lexer supported by Participle, but most parsers won't need this level of flexibility. To support this common case, which replaces the old Regex and EBNF lexers, you can use stateful.MustSimple() and stateful.NewSimple().

eg. The lexer for a form of BASIC:

var basicLexer = stateful.MustSimple([]stateful.Rule{
    {"Comment", `(?i)rem[^\n]*`, nil},
    {"String", `"(\\"|[^"])*"`, nil},
    {"Number", `[-+]?(\d*\.)?\d+`, nil},
    {"Ident", `[a-zA-Z_]\w*`, nil},
    {"Punct", `[-[!@#$%^&*()+_={}\|:;"'<,>.?/]|]`, nil},
    {"EOL", `[\n\r]+`, nil},
    {"whitespace", `[ \t]+`, nil},
})

Experimental - code generation

Participle v2 now has experimental support for generating code to perform lexing. Use participle/experimental/codegen.GenerateLexer() to compile a stateful lexer to Go code.

This will generally provide around a 10x improvement in lexing performance while producing O(1) garbage.

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.
Stateful	A basic example of a stateful lexer and corresponding parser.
Thrift	A full Thrift parser.
TOML	A TOML parser.

Included below is a full GraphQL lexer and parser:

package main

import (
	"fmt"
	"os"

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

	"github.com/alecthomas/participle/v2"
	"github.com/alecthomas/participle/v2/lexer"
	"github.com/alecthomas/participle/v2/lexer/stateful"
)

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 = stateful.MustSimple([]stateful.Rule{
		{"Comment", `(?:#|//)[^\n]*\n?`, nil},
		{"Ident", `[a-zA-Z]\w*`, nil},
		{"Number", `(?:\d*\.)?\d+`, nil},
		{"Punct", `[-[!@#$%^&*()+_={}\|:;"'<,>.?/]|]`, nil},
		{"Whitespace", `[ \t\n\r]+`, nil},
	})
	parser = participle.MustBuild(&File{},
		participle.Lexer(graphQLLexer),
		participle.Elide("Comment", "Whitespace"),
		participle.UseLookahead(2),
	)
)

var cli struct {
	EBNF  bool     `help"Dump EBNF."`
	Files []string `arg:"" optional:"" type:"existingfile" help:"GraphQL schema files to parse."`
}

func main() {
	ctx := kong.Parse(&cli)
	if cli.EBNF {
		fmt.Println(parser.String())
		ctx.Exit(0)
	}
	for _, file := range cli.Files {
		ast := &File{}
		r, err := os.Open(file)
		ctx.FatalIfErrorf(err)
		err = parser.Parse(file, 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-12    	   5941	   201242 ns/op	 178088 B/op	   2390 allocs/op
BenchmarkGoThriftParser-12      	   3196	   379226 ns/op	 157560 B/op	   2644 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.

Concurrency

A compiled Parser instance can be used concurrently. A LexerDefinition can be used concurrently. A Lexer instance cannot be used concurrently.

Error reporting

There are a few areas where Participle can provide useful feedback to users of your parser.

Errors returned by Parser.Parse*() will be of type Error. This will contain positional information where available.
Participle will make a best effort to return as much of the AST up to the error location as possible.
Any node in the AST containing a field Pos lexer.Position will be automatically populated from the nearest matching token.
Any node in the AST containing a field EndPos lexer.Position will be automatically populated from the token at the end of the node.
Any node in the AST containing a field Tokens []lexer.Token will be automatically populated with all tokens captured by the node, including elided tokens.

These related pieces of information can be combined to provide fairly comprehensive error reporting.

Limitations

Internally, Participle is a recursive descent parser with backtracking (see UseLookahead(K)).

Among other things, this means that they do not support left recursion. Left recursion must be eliminated by restructuring your grammar.

EBNF

Participle supports outputting an EBNF grammar from a Participle parser. Once the parser is constructed simply call String().

Participle also includes a parser for this form of EBNF (naturally).

eg. The GraphQL example gives in the following EBNF:

File = Entry* .
Entry = Type | Schema | Enum | "scalar" ident .
Type = "type" ident ("implements" ident)? "{" Field* "}" .
Field = ident ("(" (Argument ("," Argument)*)? ")")? ":" TypeRef ("@" ident)? .
Argument = ident ":" TypeRef ("=" Value)? .
TypeRef = "[" TypeRef "]" | ident "!"? .
Value = ident .
Schema = "schema" "{" Field* "}" .
Enum = "enum" ident "{" ident* "}" .

Syntax/Railroad Diagrams

Participle includes a command-line utility to take an EBNF representation of a Participle grammar (as returned by Parser.String()) and produce a Railroad Diagram using tabatkins/railroad-diagrams.

Here's what the GraphQL grammar looks like:

klvnptr/participle