[PATCH] Fixed inaccuracy of j0f (BZ #28185)

[Paul Zimmermann]

       Dear Florian,

> From: Florian Weimer <fweimer@redhat.com>
> Cc: libc-alpha@sourceware.org
> Date: Wed, 04 Aug 2021 12:46:00 +0200
> User-Agent: Gnus/5.13 (Gnus v5.13) Emacs/27.2 (gnu/linux)
> 
> * Paul Zimmermann:
> 
> > The largest errors over the full binary32 range are after this
> > patch (on x86_64):
> >
> > RNDN: libm wrong by up to 9.00e+00 ulp(s) [9] for x=0x1.04c39cp+6
> > RNDZ: libm wrong by up to 9.00e+00 ulp(s) [9] for x=0x1.04c39cp+6
> > RNDU: libm wrong by up to 9.00e+00 ulp(s) [9] for x=0x1.04c39cp+6
> > RNDD: libm wrong by up to 8.98e+00 ulp(s) [9] for x=0x1.4b7066p+7
> >
> > Some constants used in the code are also changed to "static const".
> 
> Would it make sense to add a few previously broken arguments to the test
> cases?  We currently do not see any test failures.

I could, but I fear it will yield plenty of errors for other formats, since
the "broken arguments" are near roots of j0, where glibc code behaves quite
poorly. For example if I add in auto-libm-test-in 44 inputs near the first
root of j0 I get:

$ make -j4 check subdirs=math
...
FAIL: math/test-double-j0
FAIL: math/test-float128-j0
FAIL: math/test-float32x-j0
FAIL: math/test-float64-j0
FAIL: math/test-float64x-j0
FAIL: math/test-ldouble-j0

Would it make sense to add just one entry (the one with the largest error)
with xfail-rounding:double,float128,float32x,float64,float64x,ldouble?

Paul