SOLVED: Re: port math/gmp fails regression tests [WAS: port lang/gcc45 - not sure how to interpret GCC developers' advice]

Sun Dec 26 23:07:17 UTC 2010

On Fri, Dec 24, 2010 at 11:55:16PM +0100, Marius Strobl wrote:
> On Fri, Dec 24, 2010 at 10:32:16PM +0000, Anton Shterenlikht wrote:
> > On Mon, Nov 22, 2010 at 08:42:06AM +0000, Anton Shterenlikht wrote:
> > > On Fri, Nov 19, 2010 at 05:20:52PM +0100, Marius Strobl wrote:
> > > > On Thu, Nov 18, 2010 at 11:07:25AM +0000, Anton Shterenlikht wrote:
> > > > > On Wed, Nov 17, 2010 at 10:26:24PM +0100, Marius Strobl wrote:
> > > > > > On Wed, Nov 17, 2010 at 01:48:49PM +0000, Anton Shterenlikht wrote:
> > > > > > > On Tue, Nov 16, 2010 at 09:31:37AM +0000, Anton Shterenlikht wrote:
> > > > > > > > On Mon, Nov 15, 2010 at 10:48:20AM -0800, John-Mark Gurney wrote:
> > > > > > > > > Anton Shterenlikht wrote this message on Mon, Nov 15, 2010 at 10:54 +0000:
> > > > > > > > > > On sparc64 -current r212302 I'm having trouble
> > > > > > > > > > building port lang/gcc45.
> > > > > > > > > > 
> > > > > > > > > > Before I go any further, let me just check if anybody
> > > > > > > > > > on this list has built gcc45 successfully?
> > > > > > > > > > 
> > > > > > > > > > Here's my GCC bug submission:
> > > > > > > > > > http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46086
> > > > > > > > > > 
> > > > > > > > > > I just can't understand what the GCC folks tell me to do.
> > > > > > > > > > 
> > > > > > > > > > They referred me to this page:
> > > > > > > > > > http://gcc.gnu.org/install/specific.html#sparc-x-x
> > > > > > > > > > 
> > > > > > > > > > and apparently to this paragraph:
> > > > > > > > > > 
> > > > > > > > > > *quote*
> > > > > > > > > > Newer versions of the GNU Multiple Precision Library (GMP),
> > > > > > > > > > the MPFR library and the MPC library are known to be
> > > > > > > > > > miscompiled by earlier versions of GCC on these platforms.
> > > > > > > > > > We therefore recommend the use of the exact versions of
> > > > > > > > > > these libraries listed as minimal versions in the prerequisites. 
> > > > > > > > > > *end quote*
> > > > > > > > > > 
> > > > > > > > > > The prerequisites link lists:
> > > > > > > > > > 
> > > > > > > > > > GNU Multiple Precision Library (GMP) version 4.3.2 (or later)
> > > > > > > > > > MPFR Library version 2.4.2 (or later)
> > > > > > > > > > 
> > > > > > > > > > I've installed:
> > > > > > > > > > 
> > > > > > > > > > gmp-5.0.1
> > > > > > > > > > gmp-ecm-6.3
> > > > > > > > > > mpfr-3.0.0
> > > > > > > > > > 
> > > > > > > > > > So do they mean to say that I should downgrade to GMP 4.3.2 and
> > > > > > > > > > MPFR 2.4.2?
> > > > > > > > > 
> > > > > > > > > Yes, I believe that they do.  I think the important phrase is:
> > > > > > > > > exact versions of these libraries listed as minimal versions
> > > > > > > > > 
> > > > > > > > > which isn't very clear, but saying use the minimal version listed on
> > > > > > > > > the page...  They probably had someone try w/ 4.3.3 and wonder why it
> > > > > > > > > wasn't working so they added the exact verbage...
> > > > > > > > 
> > > > > > > > ok, I get it. However, the ports above were pulled up and
> > > > > > > > installed when I was building gcc45. So downgrading might
> > > > > > > > break ports interdependency. I'll try instead to build GMP
> > > > > > > > and MPFR with gcc44.
> > > > > > > 
> > > > > > > I've rebuilt GMP with gcc 4.4.6 20101026,
> > > > > > > and discovered that it fails regression tests.
> > > > > > > Here's my log:
> > > > > > > 
> > > > > > > http://seis.bris.ac.uk/~mexas/gmp-5.0.1-regression-test.log
> > > > > > > 
> > > > > > > Can anybody on this list confirm (or otherwise) this.
> > > > > > > 
> > > > > > > I also checked that GMP passes all tests fine on amd64 and ia64,
> > > > > > > so the issue indeed seems to be sparc64 specific.
> > > > > > > 
> > > > > > 
> > > > > > You seem to be building the gmp port with a non-base compiler but the
> > > > > > regression test of the port is picking up the base libstdc++. Likely
> > > > > > you need fix your search path for the dynamic linker to include the
> > > > > > directory of the libraries of the gcc port you are using (I think for
> > > > > > the gcc* ports this is intended to be done via ldconfig(8), i.e. the
> > > > > > `ldconfig -r` output needs to include the libstdc++.so.6 of that gcc
> > > > > > port) or to include it before the base counterparts. If that doesn't
> > > > > > solve your problem you should talk to the port maintainer.
> > > > > 
> > > > > That seems to be the case:
> > > > > 
> > > > > # pwd
> > > > > /usr/ports/math/gmp
> > > > > # ldconfig -r | grep stdc
> > > > >         85:-lstdc++.6 => /usr/lib/libstdc++.so.6
> > > > >         251:-lstdc++.6 => /usr/local/lib/gcc44/libstdc++.so.6
> > > > > 
> > > > > Perhaps the order is wrong?
> > > > 
> > > > If `nm -D /usr/local/lib/gcc44/libstdc++.so.6 | grep GLIBCXX_3.4.11`
> > > > returns a hit then it most likely is. In that case I'm not sure how
> > > > the gcc ports are expected to be usable out-of-the-box though as
> > > > /etc/rc.d/ldconfig always adds /usr/lib before any local paths. The
> > > > only way I see then is to always set LD_LIBRARY_PATH to something
> > > > like "/usr/local/lib/gcc44 /lib /usr/lib /usr/local/lib" in the
> > > > environment. I think (but may be wrong) that there was also a way
> > > > to tell ld(1) to record the path to the library but that likely
> > > > wouldn't be usable with every port being built with a non-base gcc.
> > > 
> > > # nm -D /usr/local/lib/gcc44/libstdc++.so.6 | grep GLIBCXX_3.4.11
> > > 0000000000000000 A GLIBCXX_3.4.11
> > > # 
> > > 
> > 
> > I advanced no further in this.
> > Setting LD_LIBRARY_PATH as above didn't help - still failed the tests.
> > 
> > I didn't get anything from the GMP folks either:
> > http://gmplib.org/list-archives/gmp-bugs/2010-November/002100.html
> > 
> > I just can't see any difference between amd64 and ia64, where
> > the tests pass, and sparc, where the tests fail:
> > 
> > http://seis.bris.ac.uk/~mexas/gmp-5.0.1/
> > 
> 
> According to your logs on amd64 and ia64 you are compiling gmp with
> the system gcc where as on sparc64 you are using the gcc44 port,
> which ships with a conflicting libstdc++.so.6. I'd suggest to ask
> the maintainer of the gcc port how this is supposed to work, i.e.
> what's the right way to actually use the C++ programs built with
> the gcc port. So far I don't see any gmp or sparc64 specific problem
> here and building gmp with the gcc44 port on amd64 and ia64 should
> result in the same problem.

This is really strange.
I switched back to the default system complier,
gcc4.2.1, and now gmp pass all tests, gcc45 builds fine, and
I can build and use math/lapack.

I'm sure I started with the default compiler initially.
So, perhaps something changed in math/gmp in the meantime..

I apologise for wasting everybody's time.
anton

-- 
Anton Shterenlikht
Room 2.6, Queen's Building
Mech Eng Dept
Bristol University
University Walk, Bristol BS8 1TR, UK
Tel: +44 (0)117 331 5944
Fax: +44 (0)117 929 4423