Date: Tue, 21 Jul 2009 15:47:40 +0100 From: Bruce Simpson <bms@incunabulum.net> To: Bruce Evans <brde@optusnet.com.au> Cc: das@FreeBSD.org, freebsd-standards@FreeBSD.org, bde@FreeBSD.org Subject: Re: Boost.Math svn/branches/release regressions flagged Message-ID: <4A65D50C.4040207@incunabulum.net> In-Reply-To: <20090708041419.K45178@delplex.bde.org> References: <4A5200CF.600@incunabulum.net> <20090708041419.K45178@delplex.bde.org>
next in thread | previous in thread | raw e-mail | index | archive | help
Bruce Evans wrote: > On Mon, 6 Jul 2009, Bruce Simpson wrote: > >> Just doing a Boost regression run... Any ideas? 4 fails, in computing >> float distance, over type double, with a large exponent. > > Is this the bug fixed by gcc change 129199? No, this GCC bug was related to some global 'using' declarations. I have a patch and am awaiting clearance from re@ to commit it. Only a few of the tests in an otherwise clean run on FreeBSD 7.2/i386 are affected by this, and the bug is part of the prior GCC 4.2.1 train. It's a one line fix for the DWARF output module. > > If not, it might be a gcc bug or the test assuming that extra > precision is > available. You can probably eliminate precision bugs (including gcc bugs > related to precision) by testing on amd64. The tests seemed to be > compiled > with -O0, so maybe they are avoiding some precision bugs > intentionally. I > sometimes need more than -O0 to avoid precision bugs. We've got other problems on FreeBSD with Boost right now. At the moment. the Boost.Build GCC glue is assuming that it can put lib32 into the link line, and rtld paths. Obviously, our runtime linker hates this, which is why we have lots of FAIL for amd64. I will be working with the Boost people to get the change which caused this backed out, as it seems that it isn't needed anywhere else, either. Until we get this resolved, it's difficult to test w/o --disable-long-double and on i386. > > I couldn't see exactly what it is doing (does it call any libm > functions?) > and would find it too hard to build the whole tests. Yes, Boost.Math is using libm extensively. thanks, BMS
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?4A65D50C.4040207>