At 11:40 6-3-03 +0100, Daan Leijen wrote:

Hi all,

It seems to me that the way things usually make progress is if

* a small group * writes a specification

I think this a good way to make progress, and I am willing to help with it. I would like to note that I do not have as much time available as some other people on this list seem to have, so one should not expect any implementation effort from my side.

Daan Leijen continues:

Having said that, I would also like to express my doubts about the
suitability of a common GUI specification -- I know, this is rather
strange as I have been one of the main advocates for a portable API!

I think that if we try to specify a full-fledged GUI API, it will
be too much work to implement it on different backends. The API will
be implemented partially, with bugs, misinterpreted, or not maintained.
As examples of items that would make the implementations difficult are
portable resource files, tool tips, tool bars, generic tree controls (on GTK), dockable windows, HTML (and openGL) rendering controls etc. Most are needed in larger applications and the lack of these items is one of the
reasons that only toy apps. are written these days. Some people on this
list think that we do have the resources to implement this all by ourselves, but I disagree -- portable toolkits like wxWindows and Qt have been in the works for years and with lots more development resources behind it, and even
they have trouble supporting most platforms satisfactory.

In my opinion, a common GUI specification only makes sense if we restrict
ourselves to a simple API that supports simple (medium sized) GUI applications.
A common denominator that is implementable with a reasonable amount of effort
on many backends -- a bit like GIO and the current ObjectIO library. It would
be good to have it, as it would create some order in the chaos of all the GUI
bindings that are around now and give the user some hope of transferring applications
easily between backends and platforms. On the other hand, it would always be a 'toy' API -- suitable for education and medium-sized applications (but I think
that is the best we can do with such an API).

I agree with Daan about the potential problems, and I would like to emphasize that it is easy to underestimate the maintainance effort of any GUI library. In contrast with Daan, I think this is the main reason why this project should go for a portable API: it must be an open-ended, easily extensible GUI library with initial support for medium sized GUI applications. I think functional languages offer good tools to make this work. If such an API can't be made, then one might as well link in with an existing library.

Another approach (that I favor) is to build a full fledged API on top of an existing full fledged library that is portable among all platforms. Libraries that come in mind are Java AWT, Tcl/Tk, Qt, and wxWindows. If Haskell is integrated with such library, it supports all gadgets you can ever wish and a Haskell developer will be on equal footage (is this english?) as C or Java programmer.
Other "little" languages have chosen this approach too, wxPython and PythonTk are big examples, but we also have wxPerl, wxEiffel and Tcl itself.
(Note that this approach doesn't preclude a common GUI API: the library will just
expose more functionality.)

A point here is that the (full-fledged) functional API still needs to be designed unless one chooses a real thin interface layer. Such a solution would not profit from the expressiveness of a functional language, and should therefore only be chosen if no better solution can be found or agreed upon.

One may worry that it is too much work to interface to these large systems
(as the haskell GTK binding has proved), but in the end I think it is less
work than trying to duplicate their efforts in Haskell. This is my main worry
with the GIO/Port library: it is just a lot of work to maintain different backends
and as soon as someone loses interest...

In summary: this project should yield a good functional design for a GUI API that is open-ended, and initially includes a (very) limited set of GUI elements (to encourage developers for making ports to their favourite platforms). However, in a sense this set should be 'arbitrary': they might as well have been included afterwards as 'plug-in's. No hardcoded prefab objects.

Note that this still allows a 'port' to a library that is portable on different platforms. If one can bear the "almost, but not quite, entirely unlike" platform look-and-feel, then that is OK.

I think that in some sense, the same
happened to the Clean ObjectIO library, started out on the Mac, than Windows and
now only windows is supported. (But it does show how useful it is to have at
least some standard medium level GUI around!)

The Clean GUI project started on the Mac, and portable implementations for an early version have been made for Windows and X Windows (relying on OpenLook). This was the 0.8 version, named Clean Event I/O. The project had a significant redesign to version 1.x - and was renamed into Object I/O -. For this version there is a leading port to the Windows platform and we are working on a Mac port.

Peter Achten