Re: floating-point formats

[Lars Brinkhoff <lars dot spam at nocrew dot org>]

Ulrich Drepper <drepper@redhat.com> writes:
> On Thu, 2002-07-18 at 12:04, Roland McGrath wrote:
> > The only cases that I was talking about are machines where the
> > native FP format is something other than IEEE (e.g. Vax, PDP-10).
> > It doesn't make sense to me to attempt to impose IEEE format on an
> > ABI for such machines.
> Then use whatever environment you have today.

But there is no (adequate) C library in the environment for those
machines.  That's why we're trying to port glibc.

> On Thu, 2002-07-18 at 12:04, Roland McGrath wrote:
> > Any programs written to run on that machine under an existing
> > compiler/library will expect the machine's native FP semantics
> > rather than universally compatible IEEE semantics.

Ulrich Drepper <drepper@redhat.com> writes:
> When you use glibc you want to have an API compatible with the rest
> of the systems using glibc and those are using IEEE.

I understand that this is usually the case, but it's (probably) not
true for the PDP-10 port.  As Roland writes, native FP semantics are
expected.  IEEE compatibility is less of a concern.

Of course, I don't mean not say that glibc has to accomodate the
PDP-10 port in any way.

> In addition, maintaining code to be aware of other FP formats (or
> byte sizes) is adding additional work on the maintainer.

This is what I worry about most.  Are you sure this must be so?
Certainly, I'll keep this in mind when doing the port, and try to
minimize the maintenance burden (whether the port would ever be
accepted or not).

> Who cares about hardware which can be outperformed by a pocket
> calculator?

I don't think this is true for either (reasonably recent) PDP-10s or
VAXen.  The company I'm working for is even developing a new PDP-10
processor (yes, I know that seems strange).

> The people who want such extravagence have to pay the price.

In the case of maintaining support for exotic floating-point formats
in glibc, I believe both the PDP-10 and VAX people are happy to do so,
if possible.

Paul Eggert <eggert@twinsun.com> writes:
> > From: Ulrich Drepper <drepper@redhat.com>
> > In addition, maintaining code to be aware of other FP formats (or
> > byte sizes) is adding additional work on the maintainer.
> I agree with Ulrich here.  It is a matter of priorities, and lots of
> other things should have higher priority

I agree that VAX and PDP-10 issues are among the least useful things
to work on, but I'm not free to work on other improvements in glibc.

> If there would be a simple way to support alternate floating point
> formats, a way which wouldn't impinge much on the existing code and
> which could be maintained separately by loosely-coupled (and
> presumably diligient) volunteer group, then I suspect Ulrich would
> be all ears.

Indeed, that's exactly what I want to do.

-- 
Lars Brinkhoff          http://lars.nocrew.org/     Linux, GCC, PDP-10,
Brinkhoff Consulting    http://www.brinkhoff.se/    HTTP programming