Re: newlib-4.4.0 yearly snapshot created

[Jeff Johnston <jjohnstn@redhat.com>]

[-- Attachment #1: Type: text/plain, Size: 3941 bytes --]

Hi Paul,

I don't seem to have a working simulator so I built a test on my x86 linux
and specified
all the newlib source files needed plus a stub function to avoid needing
our _malloc_r and
various _r methods.  The test which does tgammaf(-1) returns inf whereas
glibc returns nan.  I have an older gcc so
it isn't the level of gcc.

There are a number of gamma function patches made between 4.1.0 and 4.2.0
so one of them is likely the culprit
if you believe this has regressed.

-- Jeff J.

On Fri, Jan 5, 2024 at 4:25 AM Paul Zimmermann <Paul.Zimmermann@inria.fr>
wrote:

>        Hi Jeff,
>
> I did investigate. In fact I now get the issue with tgammaf even with
> Newlib 4.2.0. It might be due to a change in the compiler used (I now use
> gcc 13.2.0) or in my testing framework.
>
> Can you reproduce the issue?
>
> For the binary64 pow function, the accuracy has greatly improved with
> respect
> to 4.2.0. We got an error of up to 636 ulps, and now the largest error I
> get
> after a few tests is 0.892 ulp, which is better than the Intel Math Library
> for this function (1.73 ulp).
>
> Paul
>
> > From: Jeff Johnston <jjohnstn@redhat.com>
> > Date: Thu, 4 Jan 2024 13:44:27 -0500
> > Cc: newlib@sourceware.org
> >
> >
> > [1:text/plain Show]
> >
> >
> > [2:text/html Hide Save:noname (3kB)]
> >
> > Hi Paul, the problem you reported hasn't been solved yet and the
> > snapshot is just a snapshot in time of the repo.
> >
> > Looking at the logs, I only see the following gamma change that occurs
> > between 4.2.0 and 4.3.0:
> >
> > commit ec69debcb977d6395f9e91ee20133de473484e20
> > Author: Andoni Arregi <andoni.arregui@gtd-gmbh.de>
> > Date:   Fri Feb 11 12:16:00 2022 +0100
> >
> >     Improve lgammaf range for very small cases
> >
> >     The original cut for small arguments at |x|<2**-70 (copied from the
> >     double version) produces that when computing nadj we get a subnormal
> >     number for t*x and thus, the division of pi/subnormal will be INF and
> >     the logarithm of it too, which is wrong as a result for lgammaf in
> this
> >     range.
> >     The proposed new limit seems to be safe and has been tested to
> >     produce accurate results.
> >     (Courtesy of Andreas Jung, ESA)
> >
> > I don't think this should have caused the regression you are seeing with
> -1
> > as input, but can you confirm
> > if it is?  Otherwise, can you use the git bisect in concert with your
> test
> > to determine the
> > source of the regression?
> >
> > Thanks,
> >
> > -- Jeff J.
> >
> > On Thu, Jan 4, 2024 at 11:03 AM Paul Zimmermann <
> Paul.Zimmermann@inria.fr>
> > wrote:
> >
> >         Hi Jeff,
> >
> >  for what concerns the accuracy of math functions, the issue with huge
> >  error
> >  in pow (https://sourceware.org/pipermail/newlib/2023/020164.html) seems
> >  to
> >  have been fixed in 4.4.0, thanks!
> >
> >  However, the regression for tgammaf(-1) with respect to 4.2.0 is still
> >  there:
> >  https://sourceware.org/pipermail/newlib/2023/020170.html.
> >
> >  Happy New Year,
> >  Paul Zimmermann
> >
> >  > From: Jeff Johnston <jjohnstn@redhat.com>
> >  > Date: Sun, 31 Dec 2023 12:23:22 -0500
> >  >
> >  >
> >  > [1:text/plain Show]
> >  >
> >  >
> >  > [2:text/html Hide Save:noname (879B)]
> >  >
> >  > The 4.4.0 snapshot has been created and uploaded to the ftp site.  For
> >  > details on changes, see the NEWS file.
> >  >
> >  > If there are any issues, I'll have to deal with them in the New Year.
> >  >
> >  > Thanks everyone for contributing last minute changes to handle the
> >  > c99/gcc-14 issues and thanks to all who made contributions/reviews
> >  this
> >  > year.  Special thanks to Corinna for leading the project and reviewing
> >  the
> >  > lion's share of the code.
> >  >
> >  > Happy New Year everyone,
> >  >
> >  > -- Jeff J.
>
>