AC_F77_LIBRARY_LDFLAGS adds LLVM compiler flags to FLIBS

[Dieter, William R]

I am trying to build GlobalArrays (https://github.com/GlobalArrays/ga) with the 
public beta Intel Fortran compiler available in the oneAPI HPC ToolKit 
(https://software.intel.com/content/www/us/en/develop/tools/oneapi/hpc-toolkit.html).
  The configure.ac includes the AC_F77_LIBRARY_LDFLAGS macro to discover which 
libraries are needed to link Fortran code with C/C++ code.

AC_F77_LIBRARY_LDFLAGS calls the ifx, the Fortran compiler, with the -v flag 
and tries to parse the output for library flags.  The problem is that the 
driver passes an LLVM option "-mllvm -loopopt=0" to the underlying compiler.  
The configure script sees a string starting with " -l" and interprets it as a 
library.  The configure script then fails, because it cannot link the test 
program to liboopopt=0.a.

AC_F77_LIBRARY_LDFLAGS only needs the linker flags, not the compiler flags.  
Compiling conftest.f directly to an executable with verbose output results in 
separate output lines for compiling and linking.  A different way would be for 
the macro to compile conftest.f to an object file in a separate step, and then 
parse the verbose output of linking the resulting conftest.o file.  This way, 
./configure only tries to parse the linker output and cannot be fooled by the 
compiler output.

To test this, I changed the configure script from:

        ac_link='$F77 -o conftest$ac_exeext $FFLAGS $LDFLAGS conftest.$ac_ext 
$LIBS >&5'

to

        ac_link='$F77 -o conftest$ac_exeext $FFLAGS $LDFLAGS 
conftest.$ac_objext $LIBS >&5'

And, changed:

        ac_f77_v_output=` eval $ac_link 5>&1 2>&1 |
          sed '/^Driving:/d; /^Configured with:/d;
              '"/^[_$as_cr_Letters][_$as_cr_alnum]*=/d"`
to

        ac_f77_v_output=`eval $ac_compile >&5 && eval $ac_link 5>&1 2>&1 |
          sed '/^Driving:/d; /^Configured with:/d;
              '"/^[_$as_cr_Letters][_$as_cr_alnum]*=/d"`

Not sure how this will impact other compilers.  I expect that narrowing the 
scope of what is parsed to just the linker should help reduce false positives 
in general.  Is splitting compilation into two steps to eliminate the compiler 
output a reasonable fix for this?  If so, I can try to figure out the macro.  
It looks like _AC_PROG_FC_V_OUTPUT in lib/autoconf/fortran.m4 emits the shell 
code that would need to change.

Thanks,
Bill.