Path: utzoo!utgpu!news-server.csri.toronto.edu!rpi!usc!elroy.jpl.nasa.gov!ames!haven!ni.umd.edu!uc780.umd.edu!cs450a03
From: cs450a03@uc780.umd.edu
Newsgroups: comp.lang.misc
Subject: RE: Limitations of ML type inference
Message-ID: <24MAR91.14580761@uc780.umd.edu>
Date: 24 Mar 91 14:58:07 GMT
References: <602@optima.cs.arizona.edu> <2400034@otter.hpl.hp.com> <8082@skye.cs.ed.ac.uk>
Sender: usenet@ni.umd.edu (USENET News System)
Organization: The University of Maryland University College
Lines: 23
Nntp-Posting-Host: uc780.umd.edu

Dave Berry writes:
>An early approach to typing polymorphic references would have allowed the
>declaration, refining the type from the polymorphic list to an integer
>list at the first assignment.  This proved to be unwieldy in practice,
>since the first assignment to a reference could be some distance from
>the initial declaration, and this made code hard to read.

One of the things I've been trying to point out is that type is a
property of data as much as it is of name (currently, I believe that
type is more data than name).  A related argument might be that it is
a property of assignment.

The way I interpret what Dave's comment is: Somebody allowed
assignment of a (constant) type to a name, which had implications on
all later assignments to that name.  To me, this is an old, familiar
stylistic problem (nonlocality).

Personally, I favor aiding "type inference" by providing (possibly
trivial) functions which guarantee results of a certain type.  If you
like, this could be extended to meta-functions and meta-types (but
meta-meta-* is silly).

Raul Rockwell