Home

Awesome

Apache 2.0 License Build Status

Copyright 2019 Marius Ackerman.

Note

This version does not support Rust. Please use v3.2.0 for Rust or log an issue if you need the features of this version in Rust.

GoGLL

Gogll generates a GLL or LR(1) parser and FSA-based lexer for any context-free grammar. The generated code is Go or Rust.

Click here for an introduction to GLL.

See the LR(1) documentation for generating LR(1) parsers.

The generated GLL parser is a clustered nonterminal parser (CNP) following [Scott et al 2019]. CNP is a version of generalised LL parsing (GLL) [Scott & Johnstone 2016]. GLL parsers can parse all context free (CF) languages.

The generated LR(1) parser is a Pager's PGM or Knuth's original LR(1) machine [Pager 1977].

The generated lexer is a linear-time finite state automaton FSA [Grune et al 2012]. The lexer ignores whitespace.

Gogll accepts grammars in markdown files, which is very useful for documenting the grammar. For example: see gogll's own grammar.

GLL has worst-case cubic time and space complexity but linear complexity for all LL productions [Scott et al 2016]. See here for space and CPU time measurements of an extreme ambiguous example. For comparable grammars tested so far gogll produces faster lexers and parsers than gocc (FSA/LR-1).

News

2022-10-11

SPPF extraction added to the generated code. See boolx example

2022-08-09

Gogll is used to build DAU DASL

2020-08-12

From v3.2.0 gogll supports tokens that can be suppressed by the lexer. This is useful, for example, to implement code comments. See example.

2020-06-28

  1. Gogll now also generates LR(1) parsers. It supports Pager's Practical General Method, weak compatibility as well as Knuths original LR(1) machine for comparison. Pager's PGM generates LR(1) parser tables similar is size to LALR. The option to generate a Knuth LR(1) machine is provide for reference.
    See LR(1) documentation for details.

  2. Please note that the -t <target> option has been replace by -go and -rust. See see usage below.

2020-06-01

See for an introduction to GLL and a performance comparison of the generated Go and Rust code parsers.

2020-05-22

GoGLL v3.1 generates Rust as well as Go parsers with similar performance:

LexerParserBuild
Go119 μs1324 μs0.124s
Rust71 μs1297 μs2.932s
  1. The duration was averaged over 1000 repetitions.
  2. Build time was measures with the time command.
    1. For Rust: time cargo build --release
    2. For Go: time go build

See examples/rust for the Rust and Go programs used for this comparison.

Use gogll's target option to generate a Rust lexer/parser: -t rust (see usage below). Gogll generates Go code by default.

2020-04-24

  1. GoGLL now generates a linear-time FSA lexer matching the CNP parser.
  2. This version of GoGLL is faster than gocc. It compiles a sample grammar in
    0.074 s, which GoCC compiles in 0.118 s. Gogll compiles itself in 0.041s.

Benefits and disadvantages of GLL and LR(1)

GLL is a parsing technique that can handle any context-free (CF) language. GLL has worst case cubic time and space complexity.

LR(1) handles a subset of the context-free languages that can be parsed bottom-up with one token look-ahead. LR(1) has linear time complexity and its table driven parser is very efficient. Pager's Practical General Method (PGM) combines compatible states as they are generated, keeping the state space small.

A GLL parser has more expensive bookkeeping than an LR(1) parser, making the LR(1) parser more efficient for parsing very large inputs.

When to use GLL

  1. When the CF grammar that best expresses the problem is not LR(1).
  2. When the LR(1) parser has more than a few conflicts that require additional language symbols or complex grammar refactorisation to resolve.
  3. The inputs to be parsed are not too big. GLL works very well for DSLs or programming languages.

When to use LR(1)

  1. When the language can be expressed as an LR(1) grammar. A grammar is LR(1) if gogll can generate a conflict-free parser for it.
  2. When the input is very big, for example: log files containing tens of thousands of lines.

Motivation for a separate lexer

The following observations were made while using GoGLLv2 on a couple of projects.

Input Symbols, Markdown Files

Gogll and lexers generated by gogll accept UTF-8 input strings, which may be in a markdown file or a plain text file.

If the input is a markdown file gogll and lexers generated by gogll treat all text outside markdown code blocks as whitespace. Markdown code blocks are delimited by triple backticks. See gogll.md for an example.

Gogll Grammar

Gogll v3 has a BNF grammar. See gogll.md

Installation

  1. Install Go from https://golang.org
  2. go install github.com/goccmack/gogll/v3@latest or
  3. Clone this repository and run go install in the root of the directory where it is installed.

Usage

Enter gogll -h or gogll for the following help:

use: gogll -h
    for help, or

use: gogll -version
    to display the version of goggl, or

use: gogll [-a][-v] [-CPUProf] [-o <out dir>] [-go] [-rust] [-gll] [-pager] [-knuth] [-resolve_conflicts] <source file>
    to generate a lexer and parser.

    <source file>: Mandatory. Name of the source file to be processed. 
        If the file extension is ".md" the bnf is extracted from markdown code 
        segments enclosed in triple backticks.
    
    -a: Optional. Regenerate all files.
        WARNING: This may destroy user editing in the LR(1) AST.
        Default: false
         
    -v: Optional. Produce verbose output, including first and follow sets,
        LR(1) sets and lexer FSA sets.
    
    -o <out dir>: Optional. The directory to which code will be generated.
                  Default: the same directory as <source file>.
                  
    -go: Optional. Generate Go code.
          Default: true, but false if -rust is selected

    -rust: Optional. Generate Rust code.
           Default: false
           
    -gll: Optional. Generate a GLL parser.
          Default true. False if -knuth or -pager is selected.
                  
    -knuth: Optional. Generate a Knuth LR(1) parser
            Default false

    -pager: Optional. Generate a Pager PGM LR(1) parser.
            Default false

    -resolve_conflicts: Optional. Automatically resolve LR(1) conflicts.
            Default: false. Only used when generating LR(1) parsers.
    
    -bs: Optional. Print BSR statistics (GLL only).
    
    -CPUProf : Optional. Generate a CPU profile. Default false.
        The generated CPU profile is in <cpu.prof>. 
        Use "go tool pprof cpu.prof" to analyse the profile.

Using the generated lexer and parser

  1. Create a lexer:
    From an []rune:
	lexer.New(input []rune) *Lexer

or from a file. If the file extension us .md the lexer will treat all text outside the markdown code blocks as whitespace.

	lexer.NewFile(fname string) *Lexer
  1. Parse the lexer:
	if err, errs := parser.Parse(lex); err != nil {...}
  1. Check for ambiguities in the parse forest
	if bsr.IsAmbiguous() {
		fmt.Println("Error: Ambiguous parse forest")
		bsr.ReportAmbiguous()
		os.Exit(1)
	}

Ambiguous BSRs must be resolved by walking the parse forest and ignoring unwanted children of ambiguous NTs (see Complete Example). 4. Use the disambiguated parse tree for the further stages of compilation. For example, see gogll's AST builder.

<a name="Complete-Example"></a>

Complete Example

The code of following example can be found at examples/boolx. The example has the following grammar: boolx.md, which generates boolean expressions such as: a | b & c | d & e:

package "github.com/goccmack/gogll/examples/boolx"

Expr :   var
     |   Expr Op Expr
     ;

var : letter ;

Op : "&" | "|" ;

The second alternate above, Expr : Expr Op Expr, is ambiguous and can produce an ambiguous parse forest. The grammar does not enforce operator precedence, this has to be done during semantic analysis.

The grammar is compiled by the following command:

gogll examples/boolx/boolx.md

The test file, boolx_test.go shows the steps required to parse an input string and produce a disambiguated abstract syntax tree:

const t1Src = `a | b & c | d & e`

func Test1(t *testing.T) {
  1. Create a lexer from the input string and parse. Fail if there are parse errors.
	if err, errs := parser.Parse(lexer.New([]rune(t1Src))); err != nil {
		fail(errs)
	}
  1. Build an abstract syntax tree for each root of the parse forest and print them.
	for i, r := range bsr.GetRoots() {
		fmt.Printf("%d: %s\n", i, buildExpr(r))
	}
}

The input string produces an ambiguous parse forest, which is partially disambiguated by applying operator precedence. We get the following output from this test:

> go test -v ./examples/boolx
=== RUN   Test1
0: (a | ((b & c) | (d & e)))
1: <nil>
2: <nil>
3: ((a | (b & c)) | (d & e))
--- PASS: Test1 (0.00s)
PASS

The output shows that the parse forest has 4 roots, 2 of which produce valid ASTs after disambiguation. The removed trees are syntactically valid by semantically invalid because they give | precedence over &. Both the remaining ASTs are syntactically and semantically valid. The AST encodes operator precedence as shown by the parentheses. The choice of which valid AST to use for further processing is application specific.

In this example disambiguation by operator precedence is applied during the AST build.

Our AST has only one type of node: Expr.

type ExprType int

const (
	Expr_Var ExprType = iota
	Expr_Expr
)

type Expr struct {
	Type  ExprType
	Var   *token.Token
	Op    *token.Token
	Left  *Expr
	Right *Expr
}

A node can represent a variable (Type = Expr_Var) or an expression (Type = Expr_Expr). If the node represents a variable the field Var contains the variable token. Otherwise Op contains the operator token and Left and Right contain the nodes of the sub-expressions.

The AST is constructed recursively from each BSR root by the function, buildExpr in boolx_test.go.

/*
Expr :   var
     |   Expr Op Expr
     ;
Op : "&" | "|" ;
*/
func buildExpr(b bsr.BSR) *Expr {
	/*** Expr :   var ***/
	if b.Alternate() == 0 {
		return &Expr{
			Type: Expr_Var,
			Var:  b.GetTChildI(0),
		}
	}

	/*** Expr : Expr Op Expr ***/
	op := b.GetNTChildI(1). // Op is symbol 1 of the Expr rule
				GetTChildI(0) // The operator token is symbol 0 for both alternates of the Op rule

	// Build the left subexpression Node. The subtree for it may be ambiguous.
	left := []*Expr{}
	// b.GetNTChildrenI(0) returns all the valid BSRs for symbol 0 of the body of the rule.
	for _, le := range b.GetNTChildrenI(0) {
		// Add subexpression if it is valid and has precedence over this expression
		if e := buildExpr(le); e != nil && hasPrecedence(e, op) {
			left = append(left, e)
		}
	}
	// No valid subexpressions therefore this whole expression is invalid
	if len(left) == 0 {
		return nil
	}
	// Belts and braces
	if len(left) > 1 {
		panic(fmt.Sprintf("%s has %d left children", b, len(left)))
	}
	// Do the same for the right subexpression
	right := []*Expr{}
	for _, le := range b.GetNTChildrenI(2) {
		if e := buildExpr(le); e != nil && hasPrecedence(e, op) {
			right = append(right, e)
		}
	}
	if len(right) == 0 {
		return nil
	}
	if len(right) > 1 {
		panic(fmt.Sprintf("%s has %d right children", b, len(right)))
	}

	// return an expression node
	return &Expr{
		Type:  Expr_Expr,
		Op:    op,
		Left:  left[0],
		Right: right[0],
	}
}

Status

Features considered for future implementation

  1. Tokens suppressed by the lexer, e.g.: code comments.
  2. Better error reporting.
  3. Better documentation, including how to traverse the binary subtree representation (BSR Scott et al 2019) of the parse forest as well as on disambiguating parse forests.
  4. Letting the parser direct which tokens to scan Scott & Johnstone 2019

Documentation

At the moment this document and the gogll grammar are the only documentation. Have a look at gogll/examples/ambiguous for a simple example and also for simple disambiguation.

Alternatively look at gogll.md which is the input grammar and also the grammar from which the parser for this version of gogll was generated. gogll/da disambiguates the parse forest for an input string.

LR(1)

See the LR(1) documentation.

Changelog

see

Bibliography

<a name=Pager-1977></a>

<a name=Scott-et-al-2019a></a>

<a name="Scott-et-al-2019"></a>

<a name="Scott-et-al-2018"></a>

<a name="Scott-et-al-2016"></a>

<a name="Afroozeh-et-al-2013"></a>

<a name="Grune-et-al-2012"></a>

<a name="Basten-2012"></a>