don't pass -arch to fortran compilers

[simon@…]

While trying to upgrade fftw-3 @3.2.1_2+gcc43 the configure script fails with the following error:

checking for Fortran 77 name-mangling scheme... configure: error: in `/opt/local/var/macports/build/_opt_local_macports-trunk_dports_math_fftw-3/work/fftw-3.2.2':
configure: error: cannot compile a simple Fortran program

config.log shows:

configure:25700: gfortran-mp-4.3 -o conftest -pipe -O2 -arch i386 -L/opt/local/lib -arch i386 conftest.f -lm
f951: error: unrecognized command line option "-arch"
f951: error: unrecognized command line option "-arch"

I have not tested it for other gcc variants and I am aware that this is related to changes in 1.8

[skymoo (Adam Mercer)]

I can't reproduce this on my MBP running Leopard and MacPorts-1.7.1. Are you trying to build universally?

[simon@…]

No, I am not. From 1.8.0 on configure.archflags is being set by base. For example #20280 shows the same behavior for the gcc port itself.

[tobypeterson]

I believe the port needs to use configure.compiler instead of directly setting the fortran compiler. Should make the check in base work...

[skymoo (Adam Mercer)]

Replying to toby@…:

I believe the port needs to use configure.compiler instead of directly setting the fortran compiler. Should make the check in base work...

How? I thought that using configure.f77 was the correct way of setting which f77 compiler to use, and reading the description of configure.compiler in the guide seems to indicate this as well.

[blb@…]

Does

configure.compiler   macports-gcc-4.?

work instead of setting just configure.f77?

[simon@…]

Replying to blb@…:

Does

configure.compiler   macports-gcc-4.?

work instead of setting just configure.f77?

I tried to set configure.compiler macports-gcc-4.3 in the gcc43 variant. Now the "-arch error" appears already on checking the C compiler gcc-mp-4.3 .

[tobypeterson]

It's possible that the configure.compiler setting gets used earlier, and setting in the variant is too late for the arch_flag_supported check to operate correctly.

[tobypeterson]

openmpi also affected, #20375

[mamoll (Mark Moll)]

Cc Me!

[jmroot (Joshua Root)]

Replying to toby@…:

It's possible that the configure.compiler setting gets used earlier, and setting in the variant is too late for the arch_flag_supported check to operate correctly.

Yeah, configure.archflags gets appended to configure.{c,cxx,f}flags before variants are executed. There are a couple of issues here: first, I don't think any fortran compiler supports -arch, so we should probably always use -m32/-m64 in configure.f*flags. Second, it's probably a bad idea to be appending to those variables at all, since there's really no time you can do it without the potential for getting it wrong. I think the convention just has to be that configure.archflags will be added to the right environment variable during configure, but if you are patching CFLAGS into a Makefile manually or whatever you'll just have to specifically put the archflags in there as well.

[jmroot (Joshua Root)]

r54236

[alakazam@…]

The problem was also present in the latest revision of the octave Portfile. r54236 fixes this issue. Thanks !