Re: patch with larger errors for "float"

[paul zimmermann <Paul dot Zimmermann at inria dot fr>]

       Dear Carlos,

thank you for your feedback.

> From: Carlos O'Donell <codonell@redhat.com>
> Date: Mon, 20 Jan 2020 10:39:23 -0500
> 
> On 1/20/20 8:10 AM, paul zimmermann wrote:
> > please find attached a patch with new values for auto-libm-test-in,
> > giving larger error bounds on x86_64 for binary32 (float).
> > 
> > Those values were found by MPCHECK (http://mpcheck.gforge.inria.fr/).
> > 
> > If the format of this patch is ok, I will submit similar patches for
> > double, long double and float128.
> 
> Paul,
> 
> It is always great to have new inputs that exercise corner cases in
> the current algorithms.
> 
> We are currently frozen for glibc 2.31 release and will re-open at
> the start of February.

yes Joseph Myers told me.

> In general when creating patches like this we would look for a
> complete patch that could be applied with git-am, this includes a
> commit message that describes how you achieved these numbers.

sorry, attached is a new version generated with "git format-patch".
The way I found those numbers is not fully reproducible, since I
ran mpcheck with random seeds several times. Should I use fixed
seeds instead?

> It would also be great if you updated the ULPs file for x86_64
> given the larger error bounds e.g. make regenulps.

changes to sysdeps/x86_64/fpu/libm-test-ulps are included

> The detailed contribution checklist for glibc is here:
> https://sourceware.org/glibc/wiki/Contribution%20checklist
> 
> It is not clear to me if you need a copyright assignment for the
> mechanical addition of extra inputs and additions to the ULPs file.
> I would say you don't because the mechanical nature of the work.

anyway I am currently working with the FSF to have some proper
copyright assignment, since I plan to do further contributions
to the GNU libc.

Best regards,
Paul

Attachment: 0001-patch-with-larger-errors-for-float.patch
Description: Binary data