Re: Treetop parser (or PEG in general?) questions

Phrogz wrote:
> I've been looking for something like treetop for a while now. Very
> excited to have found it, and to play with it.

So was I - thanks Nathan! - and many of the recent improvements are mine.

> Let's assume that I'm trying to parse the following (wiki) markup:
... snip...
> Three questions (in particular, of many) jump out at me:

> LINE ANCHORING
> In the above, there is 'block' content, and 'inline' content. Headers,
> paragraphs, and tables are block level items. Among other things, this
> means that their markup is only valid at the start of a line. For
> example, a line that started "  = Hi =" would not constitute a valid
> header. In Regexp land, this would be handled simply with a ^ anchor.
> How do you handle this in Treetop?

Treetop (and PEGs in general) requires no separate lexer - all lexing
can be done with the same algorithmic efficiency as you get from a
typical DFA-based lexer anyhow.

As a result, Treetop has no notion of whitespace or newlines - they're
just characters, and you have to have syntax rules that match them.
If you want a rule to match after a newline, call it after a newline,
or at the start of the text). If you want a rule to match only if
followed by a newline, follow it with: & '\n'
You can use any rule after & - the rule must match but the input won't
be consumed by the calling rule.

> HANDLING EOF
> As seen in the example above, a paragraph (or any block level element,
> really) is allowed to not have a newline if it's the last thing in the
> file. Do you handle this case normally by just preprocessing the input
> and shoving a newline on the end if it doesn't exist,

That's what I'd do.

> or is there a
> way in Treetop to recognize the /\Z/ anchor from a Regexp?

No. There's no EOF symbol. Perhaps there should be..

> BACKREFERENCES
> A valid heading must match this regexp: /^(?:=+) (.+) \1$/
>   It must start at the front of the line with one or more =
> characters.
>   It can have anything (including some = characters).
>   It ends with the same number of = characters, which must be followed
> by a newline.

I don't think there's a way of saying that the trailing =s must be
equal in number to those matched by the leading '='+.

In general, there's no way to inject custom code that affects the
parsing process (what ANTLR calls a semantic predicate), which is
a pity. It's something I want sometimes too, so if you can suggest
a clean enough way to specify it, I'll think about implementing it.

> Other questions I won't dive into: what's a reasonable way to eat
> inline content (words) while allowing inline markup? 

Your problem here if I read it correctly is that you want to read anything
that's not markup. Here's where the ! operator comes in. You might have a
rule called "markup", which matches any markup, and a rule "word" that you
call like this:

rule words
   (!markup word)*
end

That will match any sequence of zero or more "word", where no word matches
"markup". Here, the rule following ! may be of any complexity, as with &.

This is essentially the "C comment matching problem". Here's what I use
for C-style comments, C++ style comments, and whitespace:

    rule s      # Optional space
      S? 
    end

    rule S      # Mandatory space
      (white / comment_to_eol / comment_c_style)+
    end

    rule white
      [ \t\n\r]+
    end
      
    rule comment_to_eol
      '//' (!"\n" .)+
    end   

    rule comment_c_style
        '/*' (!'*/' . )* '*/'
    end

> Am I necessarily
> going to end up with a tree for a paragraph that has one child for
> each word (or letter)?

You'll have one leaf per leaf rule (lexical rule) - but you don't need
to look at it, you can use "text_value" of any node, which is just the
substring of the input spanned by that rule.

I hope that's some help. My CQL parser might give you some more ideas, at
<http://activefacts.rubyforge.org/svn/lib/activefacts/cql/CQLParser.treetop>.
It's a different style of language than what you're parsing, but also needs
large amounts of backtracking at times.

Clifford Heath.