I am trying to build a Fortran program, but I get errors about an undefined reference or an unresolved external symbol. I've seen another question about these errors, but the answers there are mostly specific to C++.
What are common causes of these errors when writing in Fortran, and how do I fix/prevent them?
This is a canonical question for a whole class of errors when building Fortran programs. If you've been referred here or had your question closed as a duplicate of this one, you may need to read one or more of several answers. Start with this answer which acts as a table of contents for solutions provided.
A link-time error like these messages can be for many of the same reasons as for more general uses of the linker, rather than just having compiled a Fortran program. Some of these are covered in the linked question about C++ linking and in another answer here: failing to specify the library, or providing them in the wrong order.
However, there are common mistakes in writing a Fortran program that can lead to link errors.
Unsupported intrinsics
If a subroutine reference is intended to refer to an intrinsic subroutine then this can lead to a link-time error if that subroutine intrinsic isn't offered by the compiler: it is taken to be an external subroutine.
implicit none
call unsupported_intrinsic
end
With unsupported_intrinsic not provided by the compiler we may see a linking error message like
undefined reference to `unsupported_intrinsic_'
If we are using a non-standard, or not commonly implemented, intrinsic we can help our compiler report this in a couple of ways:
implicit none
intrinsic :: my_intrinsic
call my_intrinsic
end program
If my_intrinsic isn't a supported intrinsic, then the compiler will complain with a helpful message:
Error: ‘my_intrinsic’ declared INTRINSIC at (1) does not exist
We don't have this problem with intrinsic functions because we are using implicit none:
implicit none
print *, my_intrinsic()
end
Error: Function ‘my_intrinsic’ at (1) has no IMPLICIT type
With some compilers we can use the Fortran 2018 implicit statement to do the same for subroutines
implicit none (external)
call my_intrinsic
end
Error: Procedure ‘my_intrinsic’ called at (1) is not explicitly declared
Note that it may be necessary to specify a compiler option when compiling to request the compiler support non-standard intrinsics (such as gfortran's -fdec-math). Equally, if you are requesting conformance to a particular language revision but using an intrinsic introduced in a later revision it may be necessary to change the conformance request. For example, compiling
intrinsic move_alloc
end
with gfortran and -std=f95:
intrinsic move_alloc
1
Error: The intrinsic ‘move_alloc’ declared INTRINSIC at (1) is not available in the current standard settings but new in Fortran 2003. Use an appropriate ‘-std=*’ option or enable ‘-fall-intrinsics’ in order to use it.
External procedure instead of module procedure
Just as we can try to use a module procedure in a program, but forget to give the object defining it to the linker, we can accidentally tell the compiler to use an external procedure (with a different link symbol name) instead of the module procedure:
module mod
implicit none
contains
integer function sub()
sub = 1
end function
end module
use mod, only :
implicit none
integer :: sub
print *, sub()
end
Or we could forget to use the module at all. Equally, we often see this when mistakenly referring to external procedures instead of sibling module procedures.
Using implicit none (external) can help us when we forget to use a module but this won't capture the case here where we explicitly declare the function to be an external one. We have to be careful, but if we see a link error like
undefined reference to `sub_'
then we should think we've referred to an external procedure sub instead of a module procedure: there's the absence of any name mangling for "module namespaces". That's a strong hint where we should be looking.
Mis-specified binding label
If we are interoperating with C then we can specify the link names of symbols incorrectly quite easily. It's so easy when not using the standard interoperability facility that I won't bother pointing this out. If you see link errors relating to what should be C functions, check carefully.
If using the standard facility there are still ways to trip up. Case sensitivity is one way: link symbol names are case sensitive, but your Fortran compiler has to be told the case if it's not all lower:
interface
function F() bind(c)
use, intrinsic :: iso_c_binding, only : c_int
integer(c_int) :: f
end function f
end interface
print *, F()
end
tells the Fortran compiler to ask the linker about a symbol f, even though we've called it F here. If the symbol really is called F, we need to say that explicitly:
interface
function F() bind(c, name='F')
use, intrinsic :: iso_c_binding, only : c_int
integer(c_int) :: f
end function f
end interface
print *, F()
end
If you see link errors which differ by case, check your binding labels.
The same holds for data objects with binding labels, and also make sure that any data object with linkage association has matching name in any C definition and link object.
Equally, forgetting to specify C interoperability with bind(c) means the linker may look for a mangled name with a trailing underscore or two (depending on compiler and its options). If you're trying to link against a C function cfunc but the linker complains about cfunc_, check you've said bind(c).
Not providing a main program
A compiler will often assume, unless told otherwise, that it's compiling a main program in order to generate (with the linker) an executable. If we aren't compiling a main program that's not what we want. That is, if we're compiling a module or external subprogram, for later use:
module mod
implicit none
contains
integer function f()
f = 1
end function f
end module
subroutine s()
end subroutine s
we may get a message like
undefined reference to `main'
This means that we need to tell the compiler that we aren't providing a Fortran main program. This will often be with the -c flag, but there will be a different option if trying to build a library object. The compiler documentation will give the appropriate options in this case.
There are many possible ways you can see an error like this. You may see it when trying to build your program (link error) or when running it (load error). Unfortunately, there's rarely a simple way to see which cause of your error you have.
This answer provides a summary of and links to the other answers to help you navigate. You may need to read all answers to solve your problem.
The most common cause of getting a link error like this is that you haven't correctly specified external dependencies or do not put all parts of your code together correctly.
When trying to run your program you may have a missing or incompatible runtime library.
If building fails and you have specified external dependencies, you may have a programming error which means that the compiler is looking for the wrong thing.
Not linking the library (properly)
The most common reason for the undefined reference/unresolved external symbol error is the failure to link the library that provides the symbol (most often a function or subroutine).
For example, when a subroutine from the BLAS library, like DGEMM is used, the library that provides this subroutine must be used in the linking step.
In the most simple use cases, the linking is combined with compilation:
gfortran my_source.f90 -lblas
The -lblas tells the linker (here invoked by the compiler) to link the libblas library. It can be a dynamic library (.so, .dll) or a static library (.a, .lib).
In many cases, it will be necessary to provide the library object defining the subroutine after the object requesting it. So, the linking above may succeed where switching the command line options (gfortran -lblas my_source.f90) may fail.
Note that the name of the library can be different as there are multiple implementations of BLAS (MKL, OpenBLAS, GotoBLAS,...).
But it will always be shortened from lib... to l... as in liopenblas.so and -lopenblas.
If the library is in a location where the linker does not see it, you can use the -L flag to explicitly add the directory for the linker to consider, e.g.:
gfortran -L/usr/local/lib -lopenblas
You can also try to add the path into some environment variable the linker searches, such as LIBRARY_PATH, e.g.:
export LIBRARY_PATH=$LIBRARY_PATH:/usr/local/lib
When linking and compilation are separated, the library is linked in the linking step:
gfortran -c my_source.f90 -o my_source.o
gfortran my_source.o -lblas
Not providing the module object file when linking
We have a module in a separate file module.f90 and the main program program.f90.
If we do
gfortran -c module.f90
gfortran program.f90 -o program
we receive an undefined reference error for the procedures contained in the module.
If we want to keep separate compilation steps, we need to link the compiled module object file
gfortran -c module.f90
gfortran module.o program.f90 -o program
or, when separating the linking step completely
gfortran -c module.f90
gfortran -c program.f90
gfortran module.o program.o -o program
Problems with the compiler's own libraries
Most Fortran compilers need to link your code against their own libraries. This should happen automatically without you needing to intervene, but this can fail for a number of reasons.
If you are compiling with gfortran, this problem will manifest as undefined references to symbols in libgfortran, which are all named _gfortran_.... These error messages will look like
undefined reference to '_gfortran_...'
The solution to this problem depends on its cause:
The compiler library is not installed
The compiler library should have been installed automatically when you installed the compiler. If the compiler did not install correctly, this may not have happened.
This can be solved by correctly installing the library, by correctly installing the compiler. It may be worth uninstalling the incorrectly installed compiler to avoid conflicts.
N.B. proceed with caution when uninstalling a compiler: if you uninstall the system compiler it may uninstall other necessary programs, and may render other programs unusable.
The compiler cannot find the compiler library
If the compiler library is installed in a non-standard location, the compiler may be unable to find it. You can tell the compiler where the library is using LD_LIBRARY_PATH, e.g. as
export LD_LIBRARY_PATH="/path/to/library:$LD_LIBRARY_PATH"
If you can't find the compiler library yourself, you may need to install a new copy.
The compiler and the compiler library are incompatible
If you have multiple versions of the compiler installed, you probably also have multiple versions of the compiler library installed. These may not be compatible, and the compiler might find the wrong library version.
This can be solved by pointing the compiler to the correct library version, e.g. by using LD_LIBRARY_PATH as above.
The Fortran compiler is not used for linking
If you are linking invoking the linker directly, or indirectly through a C (or other) compiler, then you may need to tell this compiler/linker to include the Fortran compiler's runtime library. For example, if using GCC's C frontend:
gcc -o program fortran_object.o c_object.o -lgfortran
I'm running VSCodium with the clangd extension, and I'd like to have proper CUDA highlighting (e.g. __device__ and __host__ keywords are understood). I ran CMake to generate the compile_commands.json file, and it includes CUDA-specific keywords (e.g. --cuda-gpu-arch=sm_52, --cuda-path=/usr/local/cuda). However, clangd still gives me the squiggle underline on CUDA-specific keywords. It seems like if clangd is just using the clang compiler to understand the source code, then clangd should work with CUDA (given that clang was able to compile the CUDA code).
So, is there any way to get clangd to work for CUDA? And if so, how do I do it via VSCodium?
Edit: As it turns out, the CUDA code highlighting works on the directly linked files (with a .cu extension), but it does not work on one of the included header files (with a .hpp extension). How do I get the text-highlighting to work on the header file?
create a config file ".clangd" in your project directory and specify the location of CUDA headers. This works for me.
CompileFlags:
Add:
- -xc++
- --cuda-path=/path_to_cuda_installation
- --cuda-gpu-arch=sm_52
- -I/path_to_cuda_installation/include
I'm trying to make Intellisense work on the MySQL source code, using Visual Studio Code on Ubuntu.
The project requires libmysqlclient-dev, which is installed.
Even though I include the development headers path in the includePath:
"includePath": [
"/usr/include/mysql",
"${workspaceFolder}/include"
],
(this is a part of the includes required; I've added many others to no avail)
the data types ulong and uint are not recognized, causing a flurry of errors:
identifier "uint" is undefined
identifier "ulong" is undefined
// and so on
This is strange, since I can see both types defined:
/usr/include/mysql/my_global.h
177:typedef unsigned int uint;
497:typedef unsigned long ulong; /* Short for unsigned long */
504:typedef unsigned long long int ulonglong; /* ulong or unsigned long long */
And the include directive is present (example file follows):
client/mysqldump.c
43:#include <my_global.h>
What am I missing?
Given the complex use of conditional compilation features such as #if, #ifdef, and #ifndef in the source code (a version is available here), without specific information on your exact environment, it's impossible to say exactly what you need to do.
But, in general, you need to make sure your environment is such that the lines with the relevant typedef statements are actually included by the C preprocessor in the code that is compiled.
There are several ways to help determine that. One, you can have your compiler dump all its macros. For MSVC, see How to find out cl.exe's built-in macros. Second, you can examine the output of the preprocessed code. For MSVC see How do I see a C/C++ source file after preprocessing in Visual Studio?.
Either of those can be of great assistance in seeing what code is actually getting compiled when there are many possible results because of complex #if... preprocessor directives.
Specifically, my issue is that I have CUDA code that needs <curand_kernel.h> to run. This isn't included by default in NVRTC. Presumably then when creating the program context (i.e. the call to nvrtcCreateProgram), I have to send in the name of the file (curand_kernel.h) and also the source code of curand_kernel.h? I feel like I shouldn't have to do that.
It's hard to tell; I haven't managed to find an example from NVIDIA of someone needing standard CUDA files like this as a source, so I really don't understand what the syntax is. Some issues: curand_kernel.h also has includes... Do I have to do the same for each of these? I am not even sure the NVRTC compiler will even run correctly on curand_kernel.h, because there are some language features it doesn't support, aren't there?
Next: if you've sent in the source code of a header file to nvrtcCreateProgram, do I still have to #include it in the code to be executed / will it cause an error if I do so?
A link to example code that does this or something like it would be appreciated much more than a straightforward answer; I really haven't managed to find any.
You have to send the "filename" and the source of each header separately.
When the preprocessor does its thing, it'll use any #include filenames as a key to find the source for the header, based on the collection that you provide.
I suspect that, in this case, the compiler (driver) doesn't have file system access, so you have to give it the source in much the same way that you would for shader includes in OpenGL.
So:
Include your header's name when calling nvrtcCreateProgram. The compiler will, internally, generate the equivalent of a std::map<string,string> containing the source of each header indexed by the given name.
In your kernel source, use #include "foo.cuh" as usual.
The compiler will use foo.cuh as an index or key into its internal map (created when you called nvrtcCreateProgram), and will retrieve the header source from that collection
Compilation proceeds as normal.
One of the reasons that nvrtc provides only a "subset" of features is that the compiler plays in a somewhat sandboxed environment, without necessarily having all of the supporting tools and utilities lying around that you have with offline compilation. So, you have to manually handle a lot of the stuff that the normal nvcc + (gcc | MSVC| clang) combination provides.
A possible, but non-ideal, solution would be to preprocess the file that you need in your IDE, save the result and then #include that. However, I bet there is a better way to do that. if you just want curand, consider diving into the library and extracting the part you need (blech) or using another GPU-friendly rand implementation. On older CUDA versions, I just generated a big array of random floats on the host, uploaded it to the GPU, and sampled it in the kernels.
This related link may be helpful.
You do not need to load curand_kernel.h yourself and add it to the include "aliases" mechanism.
Instead, you can simply add the CUDA include directory to your (set of) include paths, e.g. by adding --include-path=/usr/local/cuda/include to your NVRTC compiler options.
(I do this in my GPU-kernel-runner test harness, by default, to be on the safe side.)
). You know, in C++, C# and other educated languages, we got the chance to enable/disable the availability of some code in a certain class via compiler arguments. Take #if and #endif for example. You give an argument to the compiler, for example PROJECT_TDD_MODE, when a certain bunch of function would exist in the final code, but without defining that, or defining it as FALSE, the code would not exist in the compiled result.
Why don't we have something like this in AS3? ::- (
Conditional compilation is available in Actionscript too. Take a look at http://livedocs.adobe.com/flex/3/html/help.html?content=compilers_21.html