I have a shared library used by a another application beyond my control which requires *.so objects. My library makes use of sqlite3 which needs to be statically linked with it (I absolutely need a self-contained binary).
When I try to compile and link my library:
-fpic -flto -pthread -m64
-flto -static -shared
I end up with the following error:
/usr/bin/ld: /usr/local/lib/gcc/x86_64-unknown-linux-gnu/4.6.1/crtbeginT.o: relocation R_X86_64_32 against `__DTOR_END__' can not be used when making a shared object; recompile with -fPIC
/usr/local/lib/gcc/x86_64-unknown-linux-gnu/4.6.1/crtbeginT.o: could not read symbols: Bad value
collect2: ld returned 1 exit status
What is recompile with -fPIC related to? My code or CRT?
I have already tried to compile my object with -fPIC with the same result.
Thanks.
EDIT:
The problem does not seem to be related to SQLite3.
I wrote a simple one-line-do-nothing library which compiles and links like this:
g++ -c -fPIC -o bar.o bar.cpp
g++ -shared -o bar.so bar.o
but not like this:
g++ -c -fPIC -o bar.o bar.cpp
g++ -static -shared -o bar.so bar.o
The problem seems to be related to CRT (crtbeginT.o). Am I supposed to recompile GCC --with-pic or anything?
You shouldn't use the -static flag when creating a shared library, it's for creating statically linked executables.
If you only have a static version of the library, you can just link it in using -lsqlite3. But if there's both a dynamic version(.so) and a static version, the linker will prefer the dynamic one.
To instruct the linker to pick the static one, give the linker the -Bstatic flag, and make it switch back to dynamic linking for other stuff (like libc and dynamic runtime support) with -Bdynamic. That is, you use the flags:
-Wl,-Bstatic -lsqlite3 -Wl,-Bdynamic
Alternativly, you can just specify the full path of the .a file, e.g. /usr/lib/libsqlite3.a instead of any compiler/linker flags.
With the GNU ld, you can also use -l:libsqlite3.a instead of -lsqlite3. This will force the use of the library file libsqlite3.a instead of libsqlite3.so, which the linker prefers by default.
Remember to make sure the .a file have been compiled with the -fpic flag, otherwise you normally can't embed it in a shared library.
Any code that will somehow make its way into a dynamic library should be relocatable. It means that everything that is linked with your .so, no matter statically or dynamically, should be compiled with -fPIC. Specifically, static sqlite library should also be compiled with -fPIC.
Details of what PIC means are here: http://en.wikipedia.org/wiki/Position-independent_code
I had the same problem. Apparently -static is not the same as -Bstatic. I switched to -Bstatic and everything worked.
Related
Just for the purpose of learning, I've made a small example of a main program tentatively loeading a shared library via dlopen (and then a symbol from it via dlsym) and using a default one if the former is not avalable.
On my machine, to make the non-default library available to the main program, I need to compile the former via g++ -fPIC -shared MyLib.cpp -o libMyLib.so, whereas both main.cpp and DefaultLib.cpp are compiled simply by g++ -c main.cpp -o main.o and g++ -c DefaultLib.cpp -o DefaultLib.o. How can I provide the options -fPIC -shared to the compilation of MyLib.cpp in Compiler Explorer?
The current attempt is here, where, I believe, MyLib.cpp is compiled just like the other two cpp files, i.e. without providing options -fPIC and -shared, and maybe most importantly without generating a file with the name libMyLib.so, thus resulting dlopen failing to load; indeed, the foo from the other, default library DefaultLib is called.
Can I compile and dlopen a dynamic library in Compiler Explorer?
Yes, it's certainly possible.
In CMakeLists.txt:
add_library(MyLib SHARED MyLib.cpp)
...and remove MyLib.cpp from add_executable.
Then in main.cpp:
void * lib = dlopen("build/libMyLib.so", RTLD_LAZY);
Because the library is placed in the build subdirectory.
Demo
I'm trying to build some example c++ code that use boost library. I use this as reference example of static linking.
And everything is fine when I build with dynamic libs.
g++ -Wall -std=c++0x -O3 -Wfatal-errors -I/usr/include/boost/include -c -o src/main.o src/main.cpp
g++ -Wall -std=c++0x -O3 -Wfatal-errors -I/usr/include/boost/include -c -o src/ThreadExample.o src/ThreadExample.cpp
g++ -Wall -std=c++0x -O3 -Wfatal-errors -I/usr/include/boost/include -c -o src/Utils.o src/Utils.cpp
g++ src/main.o src/ThreadExample.o src/Utils.o -lboost_thread -lboost_filesystem -lboost_system -lboost_timer -o ThreadExampleBinary
But when I use static libs I get lots of undefined reference errors:
g++ -Wall -std=c++0x -O3 -Wfatal-errors -I/usr/include/boost/include -c -o src/main.o src/main.cpp
g++ -Wall -std=c++0x -O3 -Wfatal-errors -I/usr/include/boost/include -c -o src/ThreadExample.o src/ThreadExample.cpp
g++ -Wall -std=c++0x -O3 -Wfatal-errors -I/usr/include/boost/include -c -o src/Utils.o src/Utils.cpp
g++ -static src/main.o src/ThreadExample.o src/Utils.o -lboost_thread -lboost_filesystem -lboost_system -lboost_timer -o ThreadExampleBinary
/usr/lib/gcc/x86_64-linux-gnu/4.8/../../../x86_64-linux-gnu/libboost_timer.a(cpu_timer.o): In function `boost::timer::cpu_timer::start()':
(.text+0x7fd): undefined reference to `boost::chrono::steady_clock::now()'
/usr/lib/gcc/x86_64-linux-gnu/4.8/../../../x86_64-linux-gnu/libboost_timer.a(cpu_timer.o): In function `boost::timer::cpu_timer::stop()':
(.text+0x94c): undefined reference to `boost::chrono::steady_clock::now()'
/usr/lib/gcc/x86_64-linux-gnu/4.8/../../../x86_64-linux-gnu/libboost_timer.a(cpu_timer.o): In function `boost::timer::cpu_timer::elapsed() const':
(.text+0xa59): undefined reference to `boost::chrono::steady_clock::now()'
/usr/lib/gcc/x86_64-linux-gnu/4.8/../../../x86_64-linux-gnu/libboost_timer.a(cpu_timer.o): In function `boost::timer::cpu_timer::resume()':
(.text+0xb60): undefined reference to `boost::chrono::steady_clock::now()'
/usr/lib/gcc/x86_64-linux-gnu/4.8/../../../x86_64-linux-gnu/libboost_timer.a(cpu_timer.o): In function `boost::timer::auto_cpu_timer::auto_cpu_timer(std::ostream&, short)':
(.text+0xca5): undefined reference to `boost::chrono::steady_clock::now()'
/usr/lib/gcc/x86_64-linux-gnu/4.8/../../../x86_64-linux-gnu/libboost_timer.a(cpu_timer.o):(.text+0xd4e): more undefined references to `boost::chrono::steady_clock::now()' follow
collect2: error: ld returned 1 exit status
make: *** [ThreadExampleBinary] Error 1
Seems this can be fixed adding additional -lboost_chrono library.
But why it works in dinamic setting?
With static linking, you have to also statically link to any libraries depended on by the libraries you are linking to.
The difference is that shared libraries have an entry in the ELF header, called NEEDED that lists other shared libraries that are to be included when you link in this one.
You can see them with this command:
$ objdump -p /usr/lib/libboost_timer.so | grep NEEDED
NEEDED libboost_chrono.so.1.60.0
NEEDED libboost_system.so.1.60.0
NEEDED librt.so.1
NEEDED libstdc++.so.6
NEEDED libgcc_s.so.1
NEEDED libc.so.6
But for static libraries there is no such system, as they are simply collection of object files.
It is worth noting that the NEEDED entry in the shared objects are entirely optional, and if they are not available, then they will behave exactly like the static ones. But most shared libraries include them.
Many libraries use the pkg-config infrastructure to provide the full command line needed, but AFAIK boost is not one of them.
How to automate this process? Well, you do not. You just include what is needed and follow the linker errors to discover further needs.
You can find which static library includes a symbol with something like:
$ nm --print-file-name --defined-only --demangle /usr/lib/*.a 2> /dev/null | \
grep -q 'boost::chrono::steady_clock::now()'
/usr/lib/libboost_chrono.a:chrono.o:0000000000000090 T boost::chrono::steady_clock::now()
But why it works in dinamic setting?
Most of my make files have the following notes (from when I found the answer back when I needed it ... sorry, I don't know where I found it.)
Note - when a build using '-l' finds the .so version of that library (so - shared object) and the same .a archive is also there, g++ prefers the .so over the .a.
However, you can still achieve static link by fully specifying the path to .a.
Example:
$(CC) $(CC_FLAGS) $< /usr/local/lib/libboost_chrono.a -o $# ...
# ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Sometimes, archive code refers to symbols in yet another archive.
i.e. -lyyy_i686 uses some function from -lxxx_i686, and xxx was listed first in the build command. It is possible that a symbol might remain unresolved and the linker fails.
When this happens, try adding xxx a second time to the end of the archive list
from: -lxxx_i686 -lyyy_i686 -lrt -pthread
becomes -lxxx_i686 -lyyy_i686 -lrt -pthread -lxxx_i686
^^^^^^^^^_____________________________^^^^^^^^^^
The preceeding presumes only .a libraries are findable (no xxx.so)
Another alternative, you can command the linker to search an archive multiple times
-lyyy_i686 -lrt -pthread -(-lxxx_i686-)
tell gcc to link this library as many times as needed __^^__________^^
Again, only .a is findable
Finally, if you choose to link with the .so, note that this does not pull in the entire lib into the current build target. Instead, at runtime, the .so is loaded into memory, if it is not already there. This happens after the program starts. For most applications, this is considered a 'small' start-up performance hit as the program adjusts its memory map (automagically behind the scenes). I believe I once found that the app itself can control when to load the .so.
The use of .so pulls in the entire library to memory (if not already installed). Thus, there would not be any missing symbols, in my xxx vs yyy scenario above, the .so pulls all the symbols, whether used or not.
And once again, when trying to load xxx using "-lxxx_i686", the linker prefers to pull in libxxx_i686.so even when the libxxx_i686.a is in the same directory.
I want to use a shared library (resides in a custom directory) into an executable.
I've created this makefile
all: SayHello
SayHello: compiledObjects/SayHello.o myLib/libNames.so
g++ compiledObjects/SayHello.o -o SayHello -Icommons -LmyLib -lNames
compiledObjects/SayHello.o: SayHello.cpp
g++ -c SayHello.cpp -o compiledObjects/SayHello.o
myLib/libNames.so: commons/Names.cpp commons/Names.h
g++ -shared -fPIC commons/Names.cpp -o myLib/libNames.so
That create correctly the executable and shared library infact I can Execute the program using this command
LD_LIBRARY_PATH=/custom/path/to/lib/myLib/libNames.so
./SayHello
How can I execute ./SayHello without specify LD_LIBRARY_PATH?
I'm not using any IDE and I'm on linux.
Use the -rpath option to link your executable. See the ld(1) manual page for more information.
P.S. Your makefile appears to have a bug. If you successfully make your program, and immediately run make again, looks like your makefile will attempt to recompile the program again, even though nothing has changed.
After all, the whole purpose of a makefile is to avoid doing unneeded compilations.
The SayHello.o build target should be compiledObjects/SayHello.o.
You need to tell g++ to pass the -rpath option to the linker using -Wl,-rpath. Also, you need to specify a path to the -rpath option.
Putting it all together your last build step should look like this:
SayHello: compiledObjects/SayHello.o myLib/libNames.so
g++ compiledObjects/SayHello.o -o SayHello -Icommons -LmyLib -lNames -Wl,-rpath=/custom/path/to/lib/myLib/
Relative RPATH:
If you want to specify an RPATH relative to your binary you should use
$ORIGIN as a placeholder: -rpath='$ORIGIN/rel/path'.
I'm trying to link a static library to a shared library via a command like
g++ -shared obj.o archive.a -o libLib.so
But at run time I keep getting unresolved system that should be link at compile time from archive.a. I've tried
g++ -shared obj.o -Wl,-Bstatic archive.a -Wl,-Bdynamic -o libLib.so
and
g++ -shared obj.o -Wl,-whole-archive archive.a -Wl,-no-whole-archive -o libLib.so
with no success. I feel like I'm missing something basic here...
You practically cannot (i.e. should never) do that. Shared libraries should be position independent code, but static libraries are not.
If you want to link libaa into libfoo.so build or get a shared (PIC) library libaa.so, not a static (non-PIC) library libaa.a
So a file foo1.cc going into a shared library libfoo.so should be compiled as
g++ -c -fPIC -Wall -O foo1.cc -o foo1.pic.o
and the library will be linked as
g++ -shared foo1.pic.o foo2.pic.o -o libfoo.so
You could link another shared library libsmiling.so into libfoo.so e.g. by appending -lsmiling to the above command. But you can't link libraries into static libraries, just copy their members.
But a file going into a static library libfoo.a is just compiled as
g++ -Wall -O foo1.cc -o foo1.o
so when you extract that foo1.o member from libfoo.a it is not PIC (and that is very inefficient)
In principle you could put a non-PIC object code in a shared library. In practice, you should never do that, because the amount of relocation is so large that it defeats the purpose of shared libraries. If you did that, text memory won't be sharable and the dynamic linker would have a lot of relocation work.
Code inside .so shared objects should be PIC to permit ld.so to mmap it at different address segments in different processes.
So you could find a sequence of commands to link a static library into a shared one, e.g. extract using ar x all the members of the libbar.a then link all these extracted bar*.o with foo*.pic.o but that would be a mistake. Don't link static non-PIC libraries or object files to PIC shared libraries.
For details (with a focus on Linux) read ld.so(8), ld(1), ELF wikipage, Levine's book: Linkers and loaders, Drepper's paper: How To Write Shared Libraries
PS. Some very few static libraries contain PIC code because they could be used to make shared libraries. For example my Debian provides the libc6-pic package giving notably the /usr/lib/x86_64-linux-gnu/libc_pic.a static PIC library -usable to build some variant of libc.so -e.g. if I wanted to put my own malloc in libc.so! - without recompiling every glibc source file.
Suppose you're developing a shared library libshared.so.
And you have a static library libstatic.a with some internal classes and functionality you need. You'd like to link it to your .so like this:
g++ -o libshared.so -shared myObj.o -lstatic
Also you have an executable.sh which will use your .so and dynamically open it in the runtime
dlopen("libshared.so", RTLD_NOW)
You know this executable was as well statically linked against libstatic.a (but you're not sure the version of the library is exactly the same as yours).
So the question is:
Is it safe and correct to statically link your libshared.so against libstatic.a when you know the same library is already used in executable.sh?
You should avoid linking a static library into a shared one.
Because a shared library should have position independent code (otherwise, the dynamic linker has to do too much relocation, and you lose the benefits of shared libraries), but a static library usually does not have PIC.
Read Drepper's paper: How to write a shared library
You build your library with
g++ -Wall -O -fPIC mySrc.cc -c -o myObj.pic.o
g++ -o libshared.so -shared myObj.pic.o -lotherlib