I am developing an LLVM pass and want to run it as a plugin via the Clang LLVM driver:
clang -Xclang -load -Xclang myPlugin.so ...
At first I got an error similar to that described here
undefined symbol for self-built llvm opt?
After applying the flag -D_GLIBCXX_USE_CXX11_ABI=0 as suggested, I'm getting this error:
error: unable to load plugin 'myPlugin.so': 'myPlugin.so: undefined symbol: _ZNK4llvm12FunctionPass17createPrinterPassERNS_11raw_ostreamERKSs
This page suggests that there may be an ABI compatibility issue (which I don't fully understand)
http://clang-developers.42468.n3.nabble.com/Loading-Static-Analyzer-plugin-fails-with-CommandLine-Errors-td4034194.html
My objective is to compile the pass with either GCC or Clang and run it with the system Clang installation (Ubuntu 16.04, LLVM 3.8) rather than building Clang/LLVM from source.
The problem could come from multiple clang installations. The clang version you have used to compile your plugin may be different from the clang called in
clang -Xclang -load -Xclang myPlugin.so ...
If you use cmake to build your plugin, then
cmake -DCMAKE_EXPORT_COMPILE_COMMANDS=ON
will generate the file compile_commands.json that will contain the llvm version you have used. bear make or make -n are alternatives if you don't use cmake for your plugin.
If compile_commands.json contains for example
"command": "c++ -c -I/usr/lib/llvm-4.0/include ..."
and if clang -v is clang version 3.8.0, you are likely to obtain this error message especially if llvm::FunctionPass::createPrinterPass is in llvm-4.0 and not in llvm-3.8.
A solution may be to compile with
clang-xxx -Xclang -load -Xclang myPlugin.so ...
where clang-xxx contains the llvm-xxx that is referenced in compile_commands.json.
I was receiving that error because first argument I passed into the RegisterPass had the same name as the pass itself:
static RegisterPass<MyPass> X("MyPass", "DPVariableNamePass", false, false);
Changing it fixed the issue:
static RegisterPass<MyPass> X("my-pass", "DPVariableNamePass", false, false);
Maybe it helps
Related
I have been trying to compile LLVM/clang from source, both by myself and using this script. I am compiling clang using gcc 11.2, which itself was also built from source, and hence resides in a non-default location on my system (/scratch/opt/gcc-ml-11.2). Note that it only works with a gcc that was compiled with multilib enabled.
The problem is that I can use the compiled clang and clang++ binaries only if I supply the gcc location as follows: --gcc-toolchain=/scratch/opt/gcc-ml-11.2.0. Without that flag, I run into linker errors:
clang test.c -o test
/usr/bin/ld: cannot find crtbeginS.o: No such file or directory
/usr/bin/ld: cannot find -lgcc
/usr/bin/ld: cannot find -lgcc_s
I do not understand why a clang compiled binary needs to link against some kind of gcc library and why clang does not supply the crtbeginS.O file.
I have also tried using the LLVM linker as well as the LLVM standard library for C++ test applications, i.e., "-fuse-ld=lld -stdlib=libc++. This did not help either.
While I could just export CXXFLAGS=--gcc-toolchain=/scratch/opt/gcc-ml-11.2.0, I am not sure whether this is the right approach. Shouldn't clang/LLVM be able to function without a gcc installation?
I have also tried using my clang installation with the --gcc-toolchain=/scratch/opt/gcc-ml-11.2.0 flag to compile clang again, but the new binary also requires the gcc-toolchain flag.
Should I just globally export the path to my gcc toolchain? Why is this necessary? Isn't clang a standalone compiler that - if supplied with the option to use its own standard library - does not require gcc?
Thank you very much for your help!
Best,
Maxbit
I am trying to compile the example CUDA code axpy.cu from the llvm docs using:
clang++ axpy.cu -o axpy --cuda-gpu-arch=sm_86 -L/usr/local/cuda-11.2/lib64 -lcudart_static -ldl -lrt -pthread
but I'm getting the following error:
error: unable to create target: 'No available targets are compatible with triple "nvptx64-nvidia-cuda"'
What's going on here? I was able to compile CUDA with clang a day or two ago and now it's not working for some reason. I'm not sure what changed on my system.
I'm using clang 13 which should support CUDA 11.2.
The problem was that I built llvm/clang from source without support for cuda. The solution is to use the -DLLVM_TARGETS_TO_BUILD="X86;NVPTX" when building llvm from source.
I'm using mingw-w64 (gcc 7.3.0) and when I compile any C++ program using the following command:
g++ file.cpp -fsanitize=undefined
I get the following error:
...mingw32/bin/ld.exe: cannot find -lubsan
I'm able to successfully compile and run my programs if I remove the -fsanitize=undefined flag, though. After some research I found out that this means the library ubsan (Undefined Behavior Sanitizer) is missing, but I couldn't find anything about the library. How do I fix this?
This is well known issue with mingw see for instance this msys github issue. No proper solution known but there are several WAs.
Install WSL, ubuntu over WSL and you will have ubsan inside it
Build GCC under Windows from source enabling sanitizers build. They are present in GCC sources they are just not here in mingw.
Use -fsanitize=undefined -fsanitize-undefined-trap-on-error to just not use libubsan rich logging capabilities but get it trap on undefined instruction.
Hope one of this helps.
I have a Linux-based application, running under Ubuntu 12.04 LTS, that compiled, linked, and ran with no problem.
I recently upgraded my Ubuntu to 14.04 LTS and encountered problems compiling and linking the app.
The compilation problems were solved by manually modifying my local copy of Boost 1.48 in two files (include/boost/config/stdlib/libstdcpp3.hpp and include/boost/thread/xtime.hpp). At this point the app compiled successfully.
The problem I have is that the linkage fails with the error message:
c++: error: unrecognized command line option ‘-Wl’
I use CMake to enable compiling the app on multiple platforms. Here is the linker script generated by CMake. Note that the "-Wl" options are now inexplicably unrecognized by /usr/bin/c++:
/usr/bin/c++
-fno-stack-protector
-g
-Wl
CMakeFiles/Project.dir/main.cpp.o
CMakeFiles/Project.dir/TestCallback.cpp.o
CMakeFiles/Project.dir/utils.cpp.o
CMakeFiles/Project.dir/Request1.cpp.o
CMakeFiles/Project.dir/Response1.cpp.o
CMakeFiles/Project.dir/TextChatRequest.cpp.o
CMakeFiles/Project.dir/TextChatResponse.cpp.o
-o
/home/user/private/Project/Project_Release_1_2_Codename/Build/bin/Debug/Project
-L/home/user/Libraries/Ubuntu32_12.04/boost_1.48/lib
-L/home/user/Libraries/Ubuntu32_12.04/SqlLite_3.6/lib
-L/home/user/Libraries/Ubuntu32_12.04/taglib_1.7/lib
-L/home/user/Libraries/Ubuntu32_12.04/JSON_1.0/lib/Debug
-L/home/user/private/Project/Project_Release_1_2_Codename/Build/../lib/libUbuntu32/Debug
-rdynamic
/home/user/private/Project/Project_Release_1_2_Codename/lib/libUbuntu32/Debug/libAPI.a
/home/user/private/Project/Project_Release_1_2_Codename/lib/libUbuntu32/Debug/libInternals.a
-lboost_thread
-lboost_system
-lboost_filesystem
-lboost_program_options
-ltaglib
-lJSON
-lpthread
-Wl,-Bstatic
-lsqlite3
-Wl,-Bdynamic
-ldl
-Wl,-rpath,/home/user/Libraries/Ubuntu32_12.04/boost_1.48/lib:/home/user/Libraries/Ubuntu32_12.04/SqlLite_3.6/lib:/home/user/Libraries/Ubuntu32_12.04/taglib_1.7/lib:/home/user/Libraries/Ubuntu32_12.04/JSON_1.0/lib/Debug:/home/user/private/Project/Project_Release_1_2_Codename/Build/../lib/libUbuntu32/Debug
Here is version information for the software I'm using:
Ubuntu:
14.04.1 LTS (trusty)
c++ compiler/linker:
(Ubuntu 4.8.2-19ubuntu1) 4.8.2
CMake:
Version 2.8.12.2
Why doesn't the linker recognize "-Wl" commands? Did my upgrade to 14.04 LTS modify the linker software libraries? How can I get my app back up and linking?
On line 4 of the command you have -Wl without any actual linker options.
https://gcc.gnu.org/gcc-4.7/porting_to.html
Right at the top of this page is the following:
Earlier releases did not warn or error about completely invalid
options on gcc/g++/gfortran etc. command lines, if nothing was
compiled, but only linking was performed. This is no longer the case.
For example,
gcc -Wl -o foo foo.o -mflat_namespace
Now produces the following error
error: unrecognized command line option ‘-Wl’
error: unrecognized command line option ‘-mflat_namespace’
Invalid options need to be removed from the command line or replaced by something that is valid.
12.04 LTS packaged GCC 4.6, you've now jumped to 4.8 and -Wl on its own is no longer a valid option (or rather it never was, GCC is just more pedantic now).
As mentioned by others previously, it indeed turned out that my CMake script was injecting a lone, solitary, seemingly unnecessary "-Wl" via the CMAKE_EXE_LINKER_FLAGS setting:
if(LINUX)
set(THIRDPARTY_LIBS boost_thread boost_system boost_filesystem boost_program_options taglib JSON)
set(OS_LIBS pthread sqlite3.a dl)
set(CMAKE_EXE_LINKER_FLAGS "-Wl")
set(PREPROCESSOR_DEFINITIONS ${PREPROCESSOR_DEFINITIONS};/DTAGLIB_STATIC)
endif(LINUX)
When I removed this setting, the build succeeded. This oversight has been around a while, with the earlier version of gcc not minding. The latest gcc, however, is more pedantic and flagged it as an error.
I want to generate LLVM bitcode rather than LLVM IR, from C and C++ source.
Original version of dragonegg works correctly.
Since it doesn't have a function to emit LLVM bitcode, I tried to add that, but it doesn't work.
What's wrong with my modification?
gcc -fplugin=dragonegg.dll -fplugin-arg-dragonegg-emit-llvm -I ..\include -c test.c
cc1.exe: error: fail to initialize plugin dragonegg.dll
I posted my patch to gist.
https://gist.github.com/vroad/1f2dcd2c37e03005a349
I compiled latest LLVM and dragonegg with plugin-enabled gcc for Windows.
SourceForge.net: MinGW - Minimalist GNU for Windows: http://sourceforge.net/mailarchive/message.php?msg_id=28248366