Bug 459758 – sci-libs/fftw-3.3.3-r1[mpi] fails with ABI_X86="32 64"

• Status: RESOLVED FIXED
• Product: Gentoo Linux

[Description Johannes Hirte 2013-02-28 18:25:22 UTC]

Building with ABI_X86="32" enabled fails because of binary incompatibility.

Reproducible: Always

[Comment 1 Johannes Hirte 2013-02-28 18:27:42 UTC]

Created attachment 340534 [details]
build.log

[Comment 2 Christoph Junghans (RETIRED) 2013-03-01 17:50:18 UTC]

This is not a x32 issue, but a 32-bit issue. 

The problem is simply that one needs a 32-bit virtual/mpi, which we currently don't have.

[Comment 3 Justin Lecher (RETIRED) 2013-03-01 18:15:17 UTC]

Could we simply add something along this?

REQUIRED_USE="abi_x86_32? ( !mpi )"

[Comment 4 Christoph Junghans (RETIRED) 2013-03-01 18:36:16 UTC]

(In reply to comment #3)
> Could we simply add something along this?
> 
> REQUIRED_USE="abi_x86_32? ( !mpi )"
Yeah or virtual/mpi[${MULTILIB_USEDEP]. I am not 100% sure if it is useful to port mpi to 32-bit, but mpich2 shouldn't be a big deal.

[Comment 5 Justin Lecher (RETIRED) 2013-03-01 18:43:40 UTC]

(In reply to comment #4)
> (In reply to comment #3)
> > Could we simply add something along this?
> > 
> > REQUIRED_USE="abi_x86_32? ( !mpi )"
> Yeah or virtual/mpi[${MULTILIB_USEDEP]. I am not 100% sure if it is useful
> to port mpi to 32-bit, but mpich2 shouldn't be a big deal.

I won't do that as you have to make sure that all ports of the deps of the virtual work correctly on every ABI.

[Comment 6 Michał Górny 2013-03-01 18:55:08 UTC]

(In reply to comment #3)
> Could we simply add something along this?
> 
> REQUIRED_USE="abi_x86_32? ( !mpi )"

REQUIRED_USE="amd64? ( abi_x86_32? ( !mpi ) )"

Please don't forget about the amd64 or you will block whole x86 :).

But of course, the target way would be to update the virtual.

(In reply to comment #5)
> (In reply to comment #4)
> > (In reply to comment #3)
> > > Could we simply add something along this?
> > > 
> > > REQUIRED_USE="abi_x86_32? ( !mpi )"
> > Yeah or virtual/mpi[${MULTILIB_USEDEP]. I am not 100% sure if it is useful
> > to port mpi to 32-bit, but mpich2 shouldn't be a big deal.
> 
> I won't do that as you have to make sure that all ports of the deps of the
> virtual work correctly on every ABI.

I don't understand this. You could make some of the dependencies conditional to multilib being enabled. As in you get || () for all providers if you don't do multilib, or just a limited set if you do.

[Comment 7 Christoph Junghans (RETIRED) 2013-03-01 19:00:28 UTC]

(In reply to comment #5)
> (In reply to comment #4)
> > (In reply to comment #3)
> > > Could we simply add something along this?
> > > 
> > > REQUIRED_USE="abi_x86_32? ( !mpi )"
> > Yeah or virtual/mpi[${MULTILIB_USEDEP]. I am not 100% sure if it is useful
> > to port mpi to 32-bit, but mpich2 shouldn't be a big deal.
> 
> I won't do that as you have to make sure that all ports of the deps of the
> virtual work correctly on every ABI.
For mpich2, it would be dev-libs/libaio (no deps), sys-apps/hwloc (many deps, but only sys-libs/ncurses needed) and net-fs/nfs-utils for romio.

[Comment 8 Justin Lecher (RETIRED) 2013-03-01 19:03:38 UTC]

+  01 Mar 2013; Justin Lecher <[e-mail]> fftw-3.3.3-r1.ebuild:
+  Block 32bit ABI with USE=mpi on amd64, #459758
+

[Comment 9 Christoph Junghans (RETIRED) 2013-03-02 01:03:37 UTC]

Justin's solution is the best for now.

Let's get rid of emul-linux-x86-baselibs first and then port virtual/mpi.