Porting - 2. Shared Code
One Mach-O feature that hits many people by surprise is the strict distinction between shared libraries and dynamically loadable modules. On ELF systems both are the same; any piece of shared code can be used as a library and for dynamic loading. Use
otool -hv some_file to see the filetype of
Mach-O shared libraries have the file type MH_DYLIB and carry the extension
.dylib. They can be linked against with the usual static linker flags, e.g.
-lfoo for libfoo.dylib. However, they can not be loaded as a module. (Side note: Shared libraries can be loaded dynamically through an API. However, that API is different from the API for bundles and the semantics make it useless for an
dlopen() emulation. Most notably, shared libraries can not be unloaded.)
Loadable modules are called "bundles" in Mach-O speak. They have the file type MH_BUNDLE. Since no component involved cares about it, they can carry any extension. The extension
.bundle is recommended by Apple, but most ported software uses
.so for the sake of compatibility. Bundles can be dynamically loaded and unloaded via dyld APIs, and there is a wrapper that emulates
dlopen() on top of that API. It is not possible to link against bundles as if they were shared libraries. However, it is possible that a bundle is linked against real shared libraries; those will be loaded automatically when the bundle is loaded.
On an ELF system, version numbers are usually appended to the file name of the shared library after the extension, e.g.
libqt.so.2.3.0. On Darwin, the version numbers are placed between the library name and the extension, e.g.
libqt.2.3.0.dylib. Note that this allows you to request a specific version of the library when linking, using
-lqt.2.3.0 for the example above.
When creating a shared library, you can specify a name to be used when searching for the library at run time. This is usual practice and allows several major versions of a library to be installed at the same time. On ELF systems this is called the
soname. What's different on Darwin is that you can (and should) specify a full path along with the file name. This eliminates the need for "rpath" options and the ldconfig/ld.so.cache system. To use a library that is not yet installed, you can set the DYLD_LIBRARY_PATH environment variable; see the dyld man page for details.
The Mach-O format also offers real minor version checking, unknown on ELF systems. Every Mach-O library carries two version numbers: a "current" version and a "compatibility" version. Both version numbers are written as three numbers separated by dots, e.g. 1.4.2. The first number must be non-zero. The second and third number can be omitted and default to zero. When no version is specified, it will default to 0.0.0, which is some kind of wildcard value.
The "current" version is for informational purposes only. The "compatibility" version is used for checking as follows. When an executable is linked, the version information from the library is copied into the executable. When that executable is run, the stored version information is checked against the version information in the library that is loaded. dyld generates a run-time error and terminates the program unless the loaded library version is equal to or newer than the one used during linking.
The generation of position-independent code (PIC) is the default on Darwin. Actually, PowerPC code is position-independent by design, so there is no performance or space penalty involved. So, you don't need to specify a PIC option when compiling code for a shared library or module. However, the linker doesn't allow "common" symbols in shared libraries, so you must use the
-fno-common compiler option.
To build a shared library, you invoke the compiler driver with the
-dynamiclib option. This is best demonstrated by a comprehensive example. We'll build a library called libfoo, composed of the source files
code.c. The version number is 2.4.5, where 2 is the major revision (incompatible API change), 4 is the minor revision (backwards-compatible API change) and 5 is the bugfix revision count (some people call this the "teeny" revision, it denotes fully compatible changes). The library depends on no other shared libraries and will be installed in
cc -fno-common -c source.c cc -fno-common -c code.c cc -dynamiclib -install_name /usr/local/lib/libfoo.2.dylib \ -compatibility_version 2.4 -current_version 2.4.5 \ -o libfoo.2.4.5.dylib source.o code.o rm -f libfoo.2.dylib libfoo.dylib ln -s libfoo.2.4.5.dylib libfoo.2.dylib ln -s libfoo.2.4.5.dylib libfoo.dylib
Note which parts of the version are used where.
When linking against this library, one would normally use the
-lfoo flag, which accesses the
libfoo.dylib symlink. Regardless of which symlink or file is specified, though, it is the
install_name that is written into one's program. That means one can delete the "higher" (less version-specific) symlink
libfoo.dylib after compiling. During a minor-revision library upgrade, one just changes the target of the
libfoo.2.dylib symlink that is used by the runtime dynamic linker.
To build a loadable module, you invoke the compiler driver with the
If the module uses symbols from the host program, you'll have to
-undefined suppress to allow undefined symbols,
-flat_namespace along with it to make the new linker
in Mac OS X 10.1 happy.
A comprehensive example:
cc -fno-common -c source.c cc -fno-common -c code.c cc -bundle -flat_namespace -undefined suppress \ -o mymodule.so source.o code.o
Note that no version numbering is used. It is possible to use it in theory, but in practice it's pointless. Also note that there are no naming restrictions for bundles. Some packages prefer to prepend "lib" anyway because some other systems require it; this is harmless, since a program would use the full filename when loading a module.
Next: 3. GNU libtool