The Open MPI project just switched yesterday to using LT 2.2.4, AC
2.62, and AM 1.10.1 for our nightly trunk snapshot tarballs. Several
developers have been using this combo for a while (including me), but
using this trio for the nightly tarballs opens it up to a much wider
testing audience.
We ran into a problem with last night's regression testing with the
Fortran Intel compiler suite and shared libraries.
It seems that LT 2.2.x determines that the Intel ifort compiler option
for building shared libraries is -KPIC (when it should be -fPIC). It
then tests -KPIC, determines that it doesn't work, and then discards
it. Hence, sources built with ifort have *no* PIC option, and
therefore you can't build shared libraries for them. Here's the
relevant output from configure (RHEL4U4, x86_64, Intel compiler suite
10.1):
...
checking whether to build shared libraries... yes
checking whether to build static libraries... no
checking for ifort option to produce PIC... -KPIC
checking if ifort PIC flag -KPIC works... no
checking if ifort static flag -static works... yes
...
The output from building is therefore fairly predictable:
...
libtool: compile: ifort -I../../../ompi/include -I../../../ompi/
include -I. -I. -I../../../ompi/mpi/f90 -c mpi_wtick_f90.f90 -o .libs/
mpi_wtick_f90.o
/bin/sh ../../../libtool --mode=compile ifort -I../../../ompi/
include -I../../../ompi/include -I. -I. -I../../../ompi/mpi/f90 -c -
o mpi_wtime_f90.lo mpi_wtime_f90.f90
libtool: compile: ifort -I../../../ompi/include -I../../../ompi/
include -I. -I. -I../../../ompi/mpi/f90 -c mpi_wtime_f90.f90 -o .libs/
mpi_wtime_f90.o
/bin/sh ../../../libtool --mode=link ifort -I../../../ompi/include -
I../../../ompi/include -I. -I. -I../../../ompi/mpi/f90 -export-
dynamic -o libmpi_f90.la -rpath /home/jsquyres/bogus/lib mpi.lo
mpi_sizeof.lo mpi_comm_spawn_multiple_f90.lo mpi_testall_f90.lo
mpi_testsome_f90.lo mpi_waitall_f90.lo mpi_waitsome_f90.lo
mpi_wtick_f90.lo mpi_wtime_f90.lo -lnsl -lutil
libtool: link: ifort -shared -nofor_main .libs/mpi.o .libs/
mpi_sizeof.o .libs/mpi_comm_spawn_multiple_f90.o .libs/
mpi_testall_f90.o .libs/mpi_testsome_f90.o .libs/
mpi_waitall_f90.o .libs/mpi_waitsome_f90.o .libs/mpi_wtick_f90.o .libs/
mpi_wtime_f90.o -lnsl -lutil -Wl,-soname -Wl,libmpi_f90.so.0 -
o .libs/libmpi_f90.so.0.0.0
ld: .libs/mpi_comm_spawn_multiple_f90.o: relocation R_X86_64_PC32
against `mpi_comm_spawn_multiple_' can not be used when making a
shared object; recompile with -fPIC
ld: final link failed: Bad value
(the compile line for mpi_comm_spawn_multiple_f90.f90 was similar to
that of mpi_wtick_f90.f90 and mpi_wtime_f90.f90)
I can send the entire output if necessary.
I tested LT 2.2, 2.2.2, and 2.2.4: the failure behavior is the same.
I also tested the Intel compiler suite versions 9.0, 9.1, 10.0, and
10.1: same failure behavior with the Intel 10.x series, but -KPIC
*works* with the Intel 9.x series. Specifically, we get this in the
output from configure when using Intel ifort 9.0 and 9.1:
checking for ifort option to produce PIC... -KPIC
checking if ifort PIC flag -KPIC works... yes
...and building Open MPI works fine (the fortran shared library builds
with "ifort ... -KPIC ..." and everything works properly).
But here's the kicker: until yesterday, Open MPI was using a very old
LT 2.1a snapshot for building its nightly trunk tarballs (see http://www.open-mpi.org/svn/libtool.tar.gz)
. With that 2.1a snapshot, LT was using -fPIC (not -KPIC) for Fortran
shared libraries with all versions of the Intel compiler.
Additionally, LT 1.5.22 uses -fPIC for ifort shared libraries (I did
not test any further in the 1.5 series).
Based on the 2.1a tarball and 1.5.22 results, this *looks* like a
regression to us. Does anyone know the history of this -fPIC -> -KPIC
change for ifort? It seems like using -fPIC for all versions of ifort
should work properly.
Thanks for all your hard work on Libtool!
--
Jeff Squyres
Cisco Systems
_______________________________________________
http://lists.gnu.org/mailman/listinfo/libtool
