How to compile Clang on Windows
I have been trying to find a way to get Clang working on Windows but am having trouble. I get Clang to compile successfully, but when I try to compile a program I have a bunch of errors in the standard headers.
I am aware of rubenvb's excellent prebuilt versions of clang, but I want to compile it for myself. I also was listening to the GoingNative talks about clang which said that it didn't have very good Windows support yet. How can I get clang working on Windows?
Solution 1:
I used the following method to compile clang for C++ on Windows 7 and it has been validated by Mysticial and others:
- Download and install MinGW (make sure you install the C++ compiler) and put the bin folder in your PATH (I have MinGW 4.6.1 and tested successfully on another computer with 4.6.2)
- Make sure you have Python in your PATH (not 3, I have 2.7)
- (Optional: Make sure you have Perl in your PATH (I used ActivePerl 5.14.2 64-bit))
- Get CMake and put it in your PATH
- Go to the LLVM downloads page and download the LLVM 3.0 source code along with the Clang source code. Don't get the code from the SVN, it doesn't work with the MinGW headers.
- Extract the source codes; I had the llvm source in a folder named llvm-3.0.src on my desktop
- Put the clang source directly in a folder called "clang" (it must be called this exactly or you will build llvm but clang won't get built) in the "tools" folder inside the llvm source folder, this should make your directories look like:
- llvm source
- autoconf folder
- ...
- tools folder
- ...
- clang folder
- bindings folder
- ...
- Makefile file
- ...
- ...
- ...
- llvm source
- Make a folder named "build" in the same directory as the llvm source folder
- Open a command line and cd into the build folder
-
Run the command
cmake -G "MinGW Makefiles" -DCMAKE_BUILD_TYPE=Release ..\llvm-3.0.src
(the last argument is the relative path to the folder that has the llvm source in it (and the clang source in the tools/clang subdirectory))
This will do the equivalent of a "configure" command, and the makefiles and everything will be generated in the build folder
- This will take a few minutes
-
Run the command
mingw32-make
- This will compile llvm and clang, and the clang executables will be generated in the build/bin folder
- This will probably take a long time. (You can try to speed it up by adding parallel builds,
-j<number>
option) It might be good to close all other programs so that your computer can concentrate, and so they don't interfere with the lengthy compilation process, such as putting a lock on a folder that the compiler is writing to (it happened to me). I even turned off my antivirus and firewall software so that they wouldn't try to scan the generated files and get in the way.
Time for testing it out
-
Create a .cpp file in the build/bin folder (I will use hello.cpp). Use a standard library header to make sure the include paths and libraries are working. Start with a very simple program.
(What I started with:
#include <iostream> int main() { std::cout << "hi"; }
)
-
Run the command
clang hello.cpp -std=c++0x -I"C:\MinGW\lib\gcc\mingw32\4.6.1\include\c++" -I"C:\MinGW\lib\gcc\mingw32\4.6.1\include\c++\mingw32" -Lc:/mingw/bin/../lib/gcc/mingw32/4.6.1 -Lc:/mingw/bin/../lib/gcc -Lc:/mingw/bin/../lib/gcc/mingw32/4.6.1/../../../../mingw32/lib -Lc:/mingw/bin/../lib/gcc/mingw32/4.6.1/../../.. -L/mingw/lib -lstdc++ -lmingw32 -lgcc_s -lgcc -lmoldname -lmingwex -lmsvcrt -ladvapi32 -lshell32 -luser32 -lkernel32 -lmingw32 -lgcc_s -lgcc -lmoldname -lmingwex -lmsvcrt
(-L specifies a directory in which to search for libraries and -l specifies a library to link) (If you do not have MinGW installed to the same path as I did, you can find out the paths with the command "g++ somefile.cpp -v" to get g++ to spill its guts about what options it is using for the library paths and library files and everything else Search near the end of the output for the -L and -l options. Be aware of the .o file names that are interspersed with the -L's. Clang uses many of the same options as g++ so I literally copied and pasted that line from the output of g++)
This should compile your program and produce a file named a.out
rename a.out to a.exe or whatever
- Run the .exe
- Your program should run.
Clang (3.0) still has some problems on Windows (I don't know if these problems are also on linux). For example, compiling a lambda (which clang doesn't support) with -std=c++0x will cause clang to crash and emit a diagnostic error. (I was informed on the LLVM IRC that this is because clang implements parsing for lambdas but not semantic analysis, which is the phase in which it crashes (because they forgot to disable parsing lambdas for the 3.0 release), and they already know about this bug)
Also, the illustrious Mysticial kindly agreed to test this guide and made some observations during his testing:
- Windows headers seem to work.
- Currently only works for 32-bit.
- 64-bit compiles fine, but won't assemble.
- SSE probably is fine. ([Mysticial hasn't] tested a working SSE on 32-bit though.)
Solution 2:
Here is what worked in my environment, on Windows 8.1, overall similar to Seth's instruction, but with fresher tools.
- I installed MinGW 64 into
C:/MinGW
, to be precise I've used STL's distro. - I installed Python 3, just took their latest version.
- I installed CMake 3.0.2
- I forked LLVM and Clang on Github and cloned them to my machine, placing Clang's repo into llvm\tools\clang folder (the structure is described on the official page, they just show examples with svn instead of git).
- I created "build" folder next to "llvm" folder, and inside the "build" folder ran this command:
cmake -G "MinGW Makefiles" -D"CMAKE_MAKE_PROGRAM:FILEPATH=C:/MinGW/bin/make.exe" -DCMAKE_BUILD_TYPE=Release ..\llvm
(for some reason CMake couldn't find the "make" automatically) - Then I ran "make" to actually build. The build took a couple of hours.
- After that in another folder I've created 1.cpp, which executes a lambda expression to print "hi":
#include <iostream> int main() { []{ std::cout << "hi"; }(); }
- I've also created a cmd file to compile the cpp. Alternatively you can just set the PATH variable properly via other means. Note that GCC vesrion in your MinGW package may be different. Also Clang has some builtin paths it tries, but they are tied to specific GCC versions, and mine was not among them, so I had to provide the include paths explicitly.
set PATH=<path to the build folder from step 5>/bin;c:/mingw/bin;%PATH% clang++ -std=c++11 1.cpp -o 1.exe -I"C:/MinGW/include" -I"C:/MinGW/include/c++/4.9.1" -I"C:\MinGW\include\c++\4.9.1\x86_64-w64-mingw32" -I"C:\MinGW\x86_64-w64-mingw32\include"
- Running that cmd compiled 1.cpp into 1.exe that printed "hi".
What did not work:
- I've tried building the same llvm+clang sources without MinGW+GCC using the MSVC compiler from VS 2015 CTP. It built Clang successfully, the only difference is that you need to do that from the Developer CMD window, and you'd need to run
cmake -G "Visual Studio 12" ..\llvm
and then compile the solution in Visual Studio. However, that Clang failed to compile a sample cpp, it complained about "undeclared identifier 'char16_t'" and "__int128 is not supported on this target" within the MinGW standard library headers. If I use clang-cl and MS STL headers it complains about "throw x(y)" specifiers . Maybe I needed to provide some extra keys to the build, but I couldn't get it to work.
C:\Program Files (x86)\Microsoft Visual Studio 14.0\VC\include\xiosbase(293,4) : error: cannot compile this throw expression yet _THROW_NCEE(failure, "ios_base::eofbit set"); C:\Program Files (x86)\Microsoft Visual Studio 14.0\VC\include\xstddef(56,30) : note: expanded from macro '_THROW_NCEE' #define _THROW_NCEE(x, y) throw x(y)
Solution 3:
Refer http://clang.llvm.org/get_started.html#buildWindows
I've used "Visual Studio 11 Win64" with cmake and its worked with the currently available VS Express for Desktop.
Also for lib I'm using MinGW-W 64 and for missing files SUA. http://mingw-w64.sourceforge.net/ and http://www.suacommunity.com/
For linking .o compiled by clang++ for use with the W 64 binaries, I use -m i386pep with ld linker again shipped within the W 64 deliverable.