Date: Fri, 4 Dec 2015 22:20:07 -0700 From: Warner Losh <imp@bsdimp.com> To: NGie Cooper <yaneurabeya@gmail.com> Cc: Bruce Evans <brde@optusnet.com.au>, Garrett Cooper <ngie@freebsd.org>, src-committers <src-committers@freebsd.org>, "svn-src-all@freebsd.org" <svn-src-all@freebsd.org>, "svn-src-head@freebsd.org" <svn-src-head@freebsd.org> Subject: Re: svn commit: r290605 - in head/lib/msun: . man Message-ID: <CANCZdfqJa2iEdFSJ0ryVj6Rn5cYwXZOCVJzc%2BVKajNAduAXieQ@mail.gmail.com> In-Reply-To: <4E9A0C74-DC06-41E3-A27D-46FF06A98C95@gmail.com> References: <201511091040.tA9AeG0B038056@repo.freebsd.org> <20151109223117.Y2340@besplex.bde.org> <4E9A0C74-DC06-41E3-A27D-46FF06A98C95@gmail.com>
next in thread | previous in thread | raw e-mail | index | archive | help
On Fri, Dec 4, 2015 at 7:34 PM, NGie Cooper <yaneurabeya@gmail.com> wrote: > Hi Bruce! > > > On Nov 9, 2015, at 04:03, Bruce Evans <brde@optusnet.com.au> wrote: > > > > On Mon, 9 Nov 2015, Garrett Cooper wrote: > > > >> Log: > >> Document powl(3) > > > > powl was garbage that was intentionally undocumented. At least, I > > intentionally ignored its non-documentation together with it. > > POSIX documents it and I noticed it was missing when porting the msun > testcases from NetBSD. That=E2=80=99s the reason why I filed the bug. > > > powl doesn't compute the value of .Ar x to the exponent .Ar y. It > computes > > the value of (double)(.Ar x) to the exponent (double)(.Ar y), converted > to > > double. > > Hmmm? The types look ok per the function signatures in lib/msun: > > lib/msun/src/imprecise.c:imprecise_powl(long double x, long double y) > lib/msun/src/math.h:long double powl(long double, long double); > The prototype is right. The code is poo. It does as Bruce says. That's why it wasn't documented. People might think it was safe to use. > > These are bugs that need to be fixed then in the longterm, but in the > short term should be documented under CAVEATS. > It is so aweful, it should have remained undocumented until a suitable implementation showed up. > > I doubt that the rest of this section (about exception handling) is > > correct for powl(). > > > > The section on errors wasn't changed since it uses generic pow(). It > > claims that the error is generally less than 1 ulp. But for powl(), > > the error is generally more that 4096 ulps. Much more when the > > error exponentiates. powf() and pow() do well to avoid exponentiation > > of roundoff errors. > > I think the comment above applies here too about adding some notes to > CAVEATS. > This is what makes it poo. 4096 ulps is really horrific. Warner
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?CANCZdfqJa2iEdFSJ0ryVj6Rn5cYwXZOCVJzc%2BVKajNAduAXieQ>