floating point war casualty list

[Paul Koning]

>>>>> "Roman" == Roman Lechtchinsky <rl@cs.tu-berlin.de> writes:

 Roman> On Mon, 3 Jun 2002, Stephen L Moshier wrote:
 >> They are thinking about a nice table-driven program for the bit
 >> formats and that will help.  But overall it is a longer term
 >> project because it affects all targets and entails considerable
 >> effort to test all the different formats.

 Roman> Zack Weinberg has suggested to create a branch for this
 Roman> because it might not be a good idea to do this on the mainline
 Roman> at this point in the release cycle....

 Roman> Of course, this is only feasible if there are enough people
 Roman> who are interested in this and who will be willing to
 Roman> help. I'm not sure if that is the case at the moment.

I wonder if a branch is overkill.

The issue at hand -- if I understand correctly -- is in the internal
float to/from target machine float conversions.  Yes, that's somewhat
tricky code, but it's not a LOT of code.  And yes, it needs to be
tested carefully, but it's not even remotely as big a job as, say,
building a new scheduler.

I'm about to propose a patch that builds on Roman's work and makes the
conversion code in real.c a lot more maintainable.  Oh yes, and it
fixes several of the targets that have been complained about.  I've
done quite a lot of testing; I'll describe that in detail.

It would of course be possible to apply the new code only to the
non-IEEE targets, which wouldn't get us any of the cleanup benefits
but it would give us a number of repaired targets.  I'm looking
forward to seeing some discussion on that.

    paul