Date: Sun, 16 Jun 2024 02:26:01 +0000 From: bugzilla-noreply@freebsd.org To: bugs@FreeBSD.org Subject: [Bug 277783] libc fma() doesn't not return the correct zero sign Message-ID: <bug-277783-227-wy1kzHHqOc@https.bugs.freebsd.org/bugzilla/> In-Reply-To: <bug-277783-227@https.bugs.freebsd.org/bugzilla/> References: <bug-277783-227@https.bugs.freebsd.org/bugzilla/>
next in thread | previous in thread | raw e-mail | index | archive | help
https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=3D277783 --- Comment #14 from Steve Kargl <kargl@FreeBSD.org> --- Created attachment 251489 --> https://bugs.freebsd.org/bugzilla/attachment.cgi?id=3D251489&action= =3Dedit New patch The code for fma() and fmal() contain a special case when the addends sum to zero, but it seems that it mishandles small low order bits especially if a subnormal condition occurs. One of the testcases in msun/tests (that I bro= ke with the original patch) is fma(-1.,1.,1.) with the intermediate results (x,y,z): -0x1p+0 0x1p+0 0x1p+0 xy =3D {-0x1p-2, 0x0p+0} r =3D {0x0p+0, 0x0p+0} spread =3D 2 zs =3D 0x1p-2 fma returns (xy.hi + zs +ldexp(xy.lo,spread)), which is fine. For the case that Victor submitted one has (x,y,z): 0x1.ffffffffffff8p-501 0x1.0000000000004p-500 -0x1p-1000 xy =3D {5.000000e-01, -3.944305e-31} r =3D {0.000000e+00 0.000000e+00} spread =3D -999=20 zs =3D -5.000000e-01 For the original code, the special leads to ldexp(xy.lo,spread)=3Dldexp(-3.944305e-31,-999), is not good. The new patch skips the special case if xy.lo !=3D 0. This then allows fma= () to call its add_and_denormalize() function. --=20 You are receiving this mail because: You are the assignee for the bug.=
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?bug-277783-227-wy1kzHHqOc>