"undefined reference to" error during linking process

I am building my application using OpenMPI (built with LLVM) and few other external libraries including netcdf-fortran, BLAS and LAPACK. The files compile without any problem, but in the last stage where all the object files are being linked I get the following error

ld.lld: error: undefined symbol: _rad_8_
>>> referenced by buildings_rad_walls.f90:26 (/home/rangke/archive/gsam1.4/SRC/BUILDINGS/buildings_rad_walls.f90:26)
>>>               buildings_rad_walls.o:(buildings_rad_walls_)
>>> referenced by buildings_rad_walls.f90:0 (/home/rangke/archive/gsam1.4/SRC/BUILDINGS/buildings_rad_walls.f90:0)
>>>               buildings_rad_walls.o:(buildings_rad_walls_)
>>> referenced by buildings_rad_walls.f90:0 (/home/rangke/archive/gsam1.4/SRC/BUILDINGS/buildings_rad_walls.f90:0)
>>>               buildings_rad_walls.o:(buildings_rad_walls_)
>>> referenced 69 more times
>>> did you mean: _rad_0_
>>> defined in: rad.o
clang-14: error: linker command failed with exit code 1 (use -v to see invocation)

The error message indicates that lines of 0 and 26 in /home/rangke/archive/gsam1.4/SRC/BUILDINGS/buildings_rad_walls.f90 references _rad8_, but this file does not contain nor any of my files contain a variable named _rad_8_. Line 0 in the file is a comment, and the line 26 contains the following

if(doshadows.and.flag_sunny) then

which has nothing to do with the _rad_8_. doshadows and flag_sunny are logical data type, not an integer or real or double data type.

The following is the linking part of my makefile content.

LD = mpifort
LDFLAGS = -${ncf_dir}/lib -lnetcdff -lblas -llapack
...
...
...
$(LD) -o $@ $(OBJS) $(LDFLAGS)

where ${ncf_dir} is the netcdf-fortran directory.

I am suspecting that the variable _rad_8_ has to do with BLAS or LAPACK, specifically something with data type precision. I have tried different BLAS and LAPACK library, such as BLIS and libflame, but to no avail. Does anyone have any idea what is happening here? Any idea would help.

More clues.

What I have found out is that when I put -fdefault-real-8 flag so that every real is treated as 8 byte, I get the following error message

mpifort -c -O3 -Mpreprocess -Mextend -mcmodel=large -march=znver2 -r8 -I/usr/include /home/rangke/archive/gsam1.4/SRC/zero.f90
/tmp/rad-9b513e.ll:36:1: error: invalid use of function-local name
%struct_rad_8_ = type < { [4 x i8]  } > 
^
1 error generated.
make: *** [/home/rangke/archive/gsam1.4/makefile:62: rad.o] Error 1
make: *** Waiting for unfinished jobs....

This error mesage contains the _rad_8_ in question, but I am not sure where it is from. It seems like this error message is for the /home/rangke/archive/gsam1.4/SRC/zero.f90, but according to the make output it is produced during the compilation of rad.o which is different file.

Where is this

%struct_rad_8_ = type < { [4 x i8]  } >

from? I can confidently say that it is not part of my application, it is not in /home/rangke/archive/gsam1.4/SRC/zero.f90 or in the rad.f90 file.

More more clues

I found that _rad_8_ comes from LLVM IR (bit code). I have compiled the same code on AMD 4750U and AMD EPYC 7513, and they produce different LLVM IR codes. In short EPYC 7513 produces %struct_rad_8_ = type < { [4 x i8] } >, and eventually leads to successful linking process of my application, while AMD 4750U does not. Why is this happening? Is this problem inherent to the AMD CPU generations? Or is there a way to force AMD 4750U to gerenate the %struct_rad_8_ = type < { [4 x i8] } > part? I have listed the detailed LLVM IR issue in the AMD forum : https://community.amd.com/t5/processors/different-llvm-ir-content-for-the-same-code-on-different-zen/td-p/610733