Multilib gfortran generates 64-bit output, only (14.1)

slackj · 01-15-2014, 07:09 PM

I was having trouble compiling 32-bit versions of blas (which I'd use
for atlas, then lapack) under Slackware 14.1 in a multilib system. It failed every
time.

After investigating into the issue, I realized gfortran was producing
64-bit output, even after the . /etc/profile.d/32dev.sh command. This
does not happen with gcc ou g++, only with gfortran.

For instance, you take any .c file (say, file1.c) and any .f file
(say, file2.f), and compile them with
gcc -c -o test1.o file1.c
gfortran -c -o test2.o file2.f
Issuing a file command on both will show test1.o will be 32-bit and
test2.o will be 64-bit.

I do not know if this has a necessary relation with the problem, but I
did check executables for gcc and g++ are found in /usr/bin/32/, but
one will not find any gfortran executable file over there.

phenixia2003 · 01-16-2014, 03:05 AM

Hello,

/usr/bin/32/gcc is a shell wrapper that pass -m32 to the compiler. Since /usr/bin/g++ is a symlink to /usr/bin/32/gcc, you simply have to create a gfortran symlink to gcc in /usr/bin/32 to fix this issue, I guess :

Code:

$ su
$ cd /usr/bin/32
$ ln -s gcc gfortran

Code:

$ cd /tmp
$ echo -e "program HelloWorld\n write(*,*) \"Hello World\"\nend program HelloWorld" > HelloWorld.f90

$ gfortran HelloWorld.f90 -o HelloWorld.64bit
$ file HelloWorld.64bit 
HelloWorld.64bit: ELF 64-bit LSB  executable, x86-64, version 1 (SYSV), dynamically linked (uses shared libs), not stripped


$ . /etc/profile.d/32dev.sh
$ gfortran HelloWorld.f90 -o HelloWorld.32bit
$ file HelloWorld.32bit 
HelloWorld.32bit: ELF 32-bit LSB  executable, Intel 80386, version 1 (SYSV), dynamically linked (uses shared libs), not stripped

Hope this helps.

Cheers

--
SeB

Alien Bob · 01-16-2014, 04:06 AM

I have replied to the OP in an email (he emailed me as well as posting here) and gave a similar workaround instruction (basically, just add "-m32" to the commandline).
I have however also updated the compat32-tools with support for fortran (a fortran wrapper and some extra lines in the 32dev.* profile scripts). I want to test this before releasing it, and that must wait until tonight when I get home.

Eric

slackj · 01-16-2014, 07:20 AM

Thanks to both of you. I will do as instructed right now and post the results.

PS: since I thanked you here, Alien, I will not do so through e-mail, since I assume you already get a lot of e-mail every day.

This is the reply:

Quote:

Well, the basic issue is that I never provided a gfortran wrapper for 32-bit builds, like I did with the gcc wrapper in /usr/bin/32

It is easy to add though, and I will.

In the mean time, try the difference in output architecture when you add the "-m32" flag manually. Last time I used fortran was at university - and that was more than 30 years ago - so here is my test:

$ cat <<EOT > hellotest.f
program hello
print *, "Hello World!"
end program hello
EOT

$ g95 -o hello hellotest.f
$ file hello
hello: ELF 64-bit LSB executable, x86-64, version 1 (SYSV), dynamically linked (uses shared libs), not stripped
$ # ./hello
Hello World!
$ g95 -m32 -o hello32 hellotest.f
$ file hello32
hello32: ELF 32-bit LSB executable, Intel 80386, version 1 (SYSV), dynamically linked (uses shared libs), not stripped
$ # ./hello32
Hello World!

Writing a wrapper for gfortran analogous to /usr/bin/32/gcc is not hard.

Alternatively, I think you could also achieve the desired effect by defining additional variables in your environment:

export FC='gfortran -m32'
export F77='gfortran -m32'

Cheers, Eric

slackj · 01-16-2014, 10:11 AM

Quote:

Originally Posted by slackj

Thanks to both of you. I will do as instructed right now and post the results.

I copied /usr/bin/32/g++ to /usr/bin/32/gfortran. This seems to fix the fortran compilation problem.

After it was done, 32-bit output was being produced (after . /etc/profile.d/32dev.sh) and I could get blas to compile with a Slackbuild, but not atlas. Apparently, there are a few patches for further versions of atlas (read after 3.8.3, which is linked at Slackbuilds website) around that may help. All in all, I believe fortran compilation is fixed. Since I failed to get a log for the build, I'll post later if I still need help.

Thanks for the help, again.