[antlr-interest] philosophy about translation

[Johnicholas Hines]

Is the deep analysis that Andy Tripp's COBOL to Java project requires
a corner case of ANTLR usage? It seems to me he's building at least a
full dataflow graph, which I believe is more extreme than most
transformations.

My projects with antlr include things like "replicate bash's
brace-expansion in a standalone tool", and "convert pure lambda
calculus to continuation-passing style".

As far as I can tell, these tasks are not "99% static analysis and 1%
tree transformation". Admittedly, I have no working code, so perhaps
you should take my opinions with dash of salt.

Johnicholas

On 10/6/06, Andy Tripp <antlr at jazillian.com> wrote:
> Jim O'Connor wrote:
>
> >
> >COBOL -> generic language -> JAVA. In the general case, any specific
> >language -> generic language.  Generic language -> any specific
> >language.
> >
> >
> >
> I've thought a lot about this idea of having a generic intermediate
> representation. There are at least
> a couple of products that say they do this. I can't see how to make it
> work. Say we have
> a generic "add A to B" idea ("ADD A TO B" in COBOL, "B += A" in Java).
> If we're going to
> produce "natural" Java code, we'll produce "A++" when B is 1. Or should
> it be "++A"???
> No way to know, unless we also store that information, so we add some
> flag to our
> generic representation. The guy working on the generic-to-Java part
> demands that the
> guy working on the COBOL-to-generic part stores that information, even
> though it's
> meaningless to the COBOL guy.
>
> Then, the next day, the generic-to-Java guy realizes that A can be any
> arbitrary expression,
> and that the expression must be evaluated first, before the assignment
> is made. He starts
> going to make the generic representation more generic, and the COBOL
> sees him and says
> "WTF are you doing? You can only add two numbers! I'm not going to deal
> with your
> complicated data structure, I just want to store the fact that you're
> adding one number to another.
>
> And then the next day, the problem goes the other way. The COBOL guy
> realizes that
> the Java guy is going to need to know that B is 10 digits before the
> decimal, 5 after the decimal,
> and a particular precision. The Java guy eventually realizes
> that he can't even use "B += A", he's got to use BigDecimal. That
> doesn't bother him so much
> as the term "digits".  "You're going to tell me how many 'digits' it
> is???" He screams.
> "Don't you mean bytes?" No, he really does mean "digits". And you
> haven't seen nothin' yet,
> Java guy. Wait till he tells you how to store those digits on disk in
> packed decimal format.
>
> So in the end, the "generic representation" can't really be "generic".
> The Java guy just wants to know
> whether a variable is a "int" or a "long", while the COBOL guy wants to
> say how many digits it is
> before and after the decimal. OK, I suppose you could make it generic,
> you can store both,
> or have the "generic representation" automatically
> convert between the two. But things are going very badly and you just
> got started :)
>
> andy
>
>