The back story

Long ago when Computer Science was still about compilers and data structures I took a couple of first year CS papers. I learned about recursive decent parsers and have since used them to do real work. When I picked up Perl I took a look at Parse::RecDescent.

Parse::RecDescent was a lot more work and less fun than I expected. In fact I called on the Monks for help and was surprised at how many moving parts are buried under the surface of the iceberg (see ikegami's later replies in particular).

In contrast, playing with Marpa recently to answer a PerlMonks question turned out to be a bundle of fun. It still took some playing around to come to grips with, but at least it had sane error messages! So lets explore Marpa for the task I struggled with using P::RD. The big picture task doesn't much matter and I can't remember it anyway. The task I presented to the Monks was to parse a bunch of assignment statements and generate a hash of name/value pairs.

A first look at Marpa

Marpa works with a syntax provided in a BNF (WikiPedia: Backus Naur Form) style definition. A little extra markup hooks up actions to elements of the syntax and a little boiler plate handles setting up common parsing options:

lexeme default = latm => 1

declaration ::= assignment* action => doResult
assignment ::= name '=' number action => doAssign
name ~ [\w]+
number ~ [\d]+

:discard ~ spaces
spaces ~ [\s]+

The lexeme line sets the parser for longest token match. The 'spaces' lines at the end sets up a general rule for discarding white space. The interesting bit is in the middle where the syntax description lives.

Ignoring the action => ... bits for the moment, the first ::= line (rule) says that a declaration is 0 or more assignments. The second rule says that an assignment is a name followed by the '=' character followed by a number.

A name is one or more word characters and a number is one or more digits. The [\w]+ match expression looks kinda weird to Perl eyes, but it's what Marpa wants so it's what Marpa gets.

So far, so good. The 'action => xxx' bits hook an action (subroutine call) up to a rule so that when the rule matches, something happens. The subroutine gets called with a bunch of parameters passed that depend on the rule definition and results from processing other rules. Lets see some code:

examples/marpa/asHash.pl

use strict;
use warnings;
use Marpa::R2;
use Data::Dump;

my $syntax = <<'SYNTAX';
lexeme default = latm => 1

declaration ::= assignment* action => doResult
assignment ::= name '=' number action => doAssignNum
name ~ [\w]+
number ~ [\d]+

:discard ~ spaces
spaces ~ [\s]+
SYNTAX

my $grammar = Marpa::R2::Scanless::G->new({source => \$syntax});
my $input = <<INPUT;
x=1
y1 = 2 z = 3
wibble = 42
1 = 1 42=3
INPUT
my $result = $grammar->parse(\$input, 'main');

print Data::Dump::dump($$result);


sub doAssignNum     {
    my (@params) = @_;
    
    return {$params[1] => $params[3]};
    }


sub doResult     {
    my (@params) = @_;
    
    return {map {%$_} @params};
    }

The output from running the code is:

{ 1 => 1, 42 => 3, wibble => 42, x => 1, y1 => 2, z => 3 }

from which you can see that variable names are a rather loose concept in this parser (numbers are not usually valid variable names!). But, that aside, things work as desired.

doAssign takes the name and number from the assignment rule and returns a hash ref with the name as the key.

doResult takes the list of all the doAssign generated hash refs and turns them into a single hash ref containing all the variables. Note that it doesn't do anything smart like checking for name collisions!

Extending the parser

So now we have had a first taste. Let's add strings values as well as numbers. Most of what we have already stays the same. Here's the extra syntax to be inserted following the assignment rule:

    | name '=' string action => doAssignStr

string ::= ['] chars1 ['] action => [values]
    | '"' chars2 '"' action => [values]
    
chars1 ~ [^']*
chars2 ~ [^"]*

and the extra handler for the string assignment:

sub doAssignStr     {
    my (@params) = @_;
    
    return {$params[1] => $params[3][1]};
    }

The string rules don't need any code support. Instead the action creates a list of three values that are returned by the rule. The string value is the middle value in the list, with the quote characters as the other two values. doAssignStr fishes the string value out of the third parameter passed in to it - the returned value from the string rule.

The complete script with new test assignments is:

examples/marpa/asHashExtended.pl

use strict;
use warnings;
use Marpa::R2;
use Data::Dump;

# http://perlmonks.org/?node_id=574311

my $syntax = <<'SYNTAX';
lexeme default = latm => 1

declaration ::= assignment* action => doResult
    
assignment ::= name '=' number action => doAssignNum
    | name '=' string action => doAssignStr

string ::= ['] chars1 ['] action => [values]
    | '"' chars2 '"' action => [values]
    
chars1 ~ [^']*
chars2 ~ [^"]*

name ~ [\w]+
number ~ [\d]+

:discard ~ spaces
spaces ~ [\s]+
SYNTAX

my $grammar = Marpa::R2::Scanless::G->new({source => \$syntax});
my $input = <<INPUT;
x = 1
y = 2
z = 3
wibble = 42
wobble = "twenty three"
INPUT
my $result = $grammar->parse(\$input, 'main');

print Data::Dump::dump($$result);


sub doAssignNum     {
    my (@params) = @_;
    
    return {$params[1] => $params[3]};
    }


sub doAssignStr     {
    my (@params) = @_;
    
    return {$params[1] => $params[3][1]};
    }


sub doResult     {
    my (@params) = @_;
    
    return {map {%$_} grep {'HASH' eq ref $_} @params};
    }

which prints:

{ wibble => 42, wobble => "twenty three", x => 1, y => 2, z => 3 }

So, we haven't written a compiler yet, but there is a hint of how such a thing might be done. An interesting exercise is to write a script to perform the same task without using helper modules. The simple numeric assignment parsing is easy, especially if you use a couple of Perlish tricks. Things start getting a bit more hairy when the string parsing is added. It isn't just a matter of extending the original code a little as was the case with the Marpa code above.

Note that this is very much a play thing. It doesn't provide any sort of sensible error handling or allow quoted characters in the strings or even do very much that is immediately useful. It does provide a starting point and a playground for exploring Marpa however.

I've written parsers of various sorts in a variety of languages, mostly for arithmetic expression parsing. Parsers can be fun, and are almost always frustrating to debug. My (small) experience so far with Marpa is that it enhances the fun and reduces the frustration.