[antlr-interest] Re: Anyone tried this ANTLR-inspired CC?

[brian-l-smith at uiowa.edu]

Quoting "Tiller, Michael (M.M.)" <mtiller at ford.com>:
> > From: Terence Parr [mailto:parrt at cs.usfca.edu]
> > On Sunday, November 9, 2003, at 05:37 PM, leung13512c wrote:
> > 
> > > The thing that caused me to switch away from ANTLR syntax
> > > is because I prefer the tree building syntax of JJTree that always
> > > create imaginary nodes instead of a mix of both.  And it mostly
> > > avoid having to create nodes within actions.
> > 
> > Howdy.  This is the old AST versus parse tree argument...Parse trees 
> > are easy to generate (i.e., no cost to developer) but have 
> > way too much noise for tree transformation and also are sensitive
> > to grammar construction whereas ASTs are the canonical repr for
> > a particular language.
> > 
> > Ter
> 
> Personally, there are two extremes I would like to avoid.  One is using
> terminals as roots in ASTs.  I don't really care for this approach because I
> find it rather strange (just my personal preference).  The other case I
> wouldn't like is having every production rule in my grammar generate a node
> in my AST.  As far as I'm concerned, these are both sub-optimal.

I agree with Mike. I never use ANTLR's AST features because they aren't really
for *abstract* syntax trees--they are really a kind of *concrete* syntax tree
(similarly, parse trees ala JJTree are *concrete* syntax trees). This is easy to
see because the resultant tree is built using elements of the concrete syntax
grammar. An *abstract* syntax tree is independent of the concrete syntax of the
language. In fact, usually I create a syntax-independent domain model of the
language (e.g. with classes representing all the major concepts of the
language), and then create an ANTLR grammar that can parse a syntax that allows
one to express these abstract concepts textually. If you develop a language this
way then the ANTLR "AST" mechanism is basically useless. But, IMO, this is the
most sensible way to develop a language tool, because once you have the domain
model built you can have one team work on the concrete syntax (text -> AST) part
of the project while concurrently another team is working on the semantics (e.g.
AST evaluation, translation). 

Terence made the point that ANTLR's "AST's" are better than JJTree-style parse
trees because the resultant trees are less sensitive to changes in the grammar
used for parsing. But, changes in the concrete syntax of the language, and/or
changes in the grammar used to parse it, still can easily reshape the resultant
"AST" unless one is very careful. But, the AST-first approach completely
isolates all grammar/syntax changes from other parts of the project.

So, I guess what I would like to see is a mechanism for specifying a
syntax-independent ("abstract") domain model of a language, a mechanism for
specifying how to parse concrete textual representations of the language
(grammars), and a mechanism for translating between textual representations and
the abstract domain models (either one-way text -> AST, or bidirectional text
<--> AST).

Regards,
Brian









 

Your use of Yahoo! Groups is subject to http://docs.yahoo.com/info/terms/