Hi, while updating to my comparison of the accuracy of mathematical functions [1], I have noticed an issue with acoshf in musl-1.2.2: $ cat test_acosh_musl.c #include <stdio.h> #include <stdlib.h> #include <math.h> int main (int argc, char *argv[]) { float x = -0x1.1e6ae8p+5; float y; y = acoshf (x); printf ("x=%a y=%a\n", x, y); } With gcc I get NaN as expected: $ gcc -fno-builtin test_acosh_musl.c -lm $ ./a.out x=-0x1.1e6ae8p+5 y=-nan With musl-1.2.2 I get -0x1.2f63acp+3: $ gcc -fno-builtin test_acosh_musl.c $FILES $ ./a.out x=-0x1.1e6ae8p+5 y=-0x1.2f63acp+3 Please can someone confirm? Best regards, Paul [1] https://members.loria.fr/PZimmermann/papers/accuracy.pdf
On Fri, Feb 05, 2021 at 08:18:02AM +0100, Paul Zimmermann wrote:
> Hi,
>
> while updating to my comparison of the accuracy of mathematical functions [1],
> I have noticed an issue with acoshf in musl-1.2.2:
>
> $ cat test_acosh_musl.c
> #include <stdio.h>
> #include <stdlib.h>
> #include <math.h>
>
> int
> main (int argc, char *argv[])
> {
> float x = -0x1.1e6ae8p+5;
> float y;
> y = acoshf (x);
> printf ("x=%a y=%a\n", x, y);
> }
>
> With gcc I get NaN as expected:
>
> $ gcc -fno-builtin test_acosh_musl.c -lm
> $ ./a.out
> x=-0x1.1e6ae8p+5 y=-nan
>
> With musl-1.2.2 I get -0x1.2f63acp+3:
>
> $ gcc -fno-builtin test_acosh_musl.c $FILES
> $ ./a.out
> x=-0x1.1e6ae8p+5 y=-0x1.2f63acp+3
>
> Please can someone confirm?
I can't reproduce it on i386 but can on sh w/softfloat. I'm guessing
you're using an arch without its own special definition of sqrtf or
logf, so it looks like it would have to be a bug in the
non-arch-specific version of one of those, but I haven't been able to
reproduce it in isolation just passing the values passed to them
(-0x1.c9b6fcp-7 to logf or 0x1.40330cp+10 to sqrtf) manually building
the generic C versions.
Thanks for the report. I'll keep looking.
Rich
Dear Rich,
> Date: Fri, 5 Feb 2021 11:44:02 -0500
> From: Rich Felker <dalias@libc.org>
> Cc: musl@lists.openwall.com
> User-Agent: Mutt/1.5.21 (2010-09-15)
>
> On Fri, Feb 05, 2021 at 08:18:02AM +0100, Paul Zimmermann wrote:
> > Hi,
> >
> > while updating to my comparison of the accuracy of mathematical functions [1],
> > I have noticed an issue with acoshf in musl-1.2.2:
> >
> > $ cat test_acosh_musl.c
> > #include <stdio.h>
> > #include <stdlib.h>
> > #include <math.h>
> >
> > int
> > main (int argc, char *argv[])
> > {
> > float x = -0x1.1e6ae8p+5;
> > float y;
> > y = acoshf (x);
> > printf ("x=%a y=%a\n", x, y);
> > }
> >
> > With gcc I get NaN as expected:
> >
> > $ gcc -fno-builtin test_acosh_musl.c -lm
> > $ ./a.out
> > x=-0x1.1e6ae8p+5 y=-nan
> >
> > With musl-1.2.2 I get -0x1.2f63acp+3:
> >
> > $ gcc -fno-builtin test_acosh_musl.c $FILES
> > $ ./a.out
> > x=-0x1.1e6ae8p+5 y=-0x1.2f63acp+3
> >
> > Please can someone confirm?
>
> I can't reproduce it on i386 but can on sh w/softfloat. I'm guessing
> you're using an arch without its own special definition of sqrtf or
> logf, so it looks like it would have to be a bug in the
> non-arch-specific version of one of those, but I haven't been able to
> reproduce it in isolation just passing the values passed to them
> (-0x1.c9b6fcp-7 to logf or 0x1.40330cp+10 to sqrtf) manually building
> the generic C versions.
>
> Thanks for the report. I'll keep looking.
>
> Rich
my machine is an Intel(R) Core(TM) i5-4590, and I use gcc version 10.2.1,
under Debian bullseye/sid. Szabolcs sent a patch, and I confirmed it solves
the issue. I'm now testing again acosh and sinh for the whole binary32 domain,
and will report here my findings.
Paul