Re: Garbage Collection in Ada

[geert@fozzie.sun3.iaf.nl (Geert Bosch)]

Robert Dewar (dewar@merv.cs.nyu.edu) wrote:
`` Actually I think that Geert is indeed working on interfacing the HJB
   CGC to GNAT (that's why he complained about the possibility of us
   moving to use virtual origins, something we definitely intend to do
   at some point, [...]''

More exactly, I _was_ working on a CGC for GNAT. This was an assignment
for a course I had, but I have decided not to continue working on it
after I finished the course. This was partly because of time
constraints and the reasons given below.

`` [...] so presumably this conservative GC would stop working if we
   did that, unless it got more clever (more clever is possible here,
   but difficult, i.e. the scan of memory could recognize in addition
   to simple direct pointers, possible instances of virtual origin
   descriptors, they have a perfetly well defined format. ''

From the discussion in comp.lang.ada I concluded that it was wasted
time to spend more time developing a CGC for GNAT. Even if such a GC
turns out to work just fine in practise people also have to accept the
technology, which is not the case for many people.

Another reason to stop work on the CGC was that there is much more
interest for an exact GC for GNAT. During my work on the CGC I found
literature by Paul Wilson and Marc Shapiro describing precise scanning
using just debug information to do full compacting/moving GC.

Even when there will be some arbitrary pointers left like CPU-registers
and pointers due to interfacing with foreign language libraries without
debugging information, this can still be a good approach. (See also
Joel Bartlett's mostly copying GC.)

So, there is a good way to implement precise GC for GNAT without
changing the compiler which is very complex already. Using debug
information with GNAT also has the virtue that (under mild conditions)
precise scanning can be done even when interfacing to C/C++ and that
the fat pointers used by GNAT are a big help.

Because of these points I think it is very well possible to implement a
precise GC for GNAT without changing or limiting the compiler too
much.  According to comp.lang.ada readers there is also demand for a
precise GC for GNAT, so my next question is: who is going to develop
such a GC or fund development? ;-)

Regards,
   Geert

-- 
E-Mail: geert@sun3.iaf.nl