https://gcc.gnu.org/bugzilla/show_bug.cgi?id=124819
--- Comment #4 from Drea Pinski <pinskia at gcc dot gnu.org> ---
(In reply to anlauf from comment #3)
> (In reply to Christopher Albert from comment #2)
> > I tested other compilers on the same system and libc.
> >
> > Results:
> > - system gfortran 15.2.1: passes
> > - nvfortran 26.3: passes
> > - local GCC 16 object linked against system libgfortran: passes
> > - system gfortran object linked against local GCC 16 libgfortran: fails
> >
> > So the failure follows the local libgfortran, not the frontend or glibc
> > alone. That points directly at the GCC 16 libgfortran transformational
> > REAL(4) BESSEL_JN runtime path.
>
> The only difference I see then is the CFLAGS.
>
> (I use "-g -march=native -O2" on an Intel i5-8250U and also compile with
> --disable-bootstrap.)
Hmm, that might mean using fma more here:
ret->base_addr[i*stride] = x2rev * (i+1+n1) * last2 - last1;
THough maybe it is not more accuracy but rather different rounding.
Note I have not looked into the source enough to see the difference.
