I have a software library and I used to create .a files, so that people can install them and link against them: g++ foo.o -L/path/to -llibrary
But now I often encounter third-party libraries where only .so files are available (instead of .a), and you just link against them without the -l switch, e.g. g++ foo.o /path/to/liblibrary.so.
What are the differences between these solutions? Should I prefer creating .so files for the users of my library?
Typically, libfoo.a is a static library, and libfoo.so is a shared library. You can use the same -L/-l linker options against either a static or shared. Or you can name the full path to the lib with static or shared. Often libraries are built both static and shared to provide application developers the choice of which they want.
All the code needed from a static lib is part of the final executable. This obviously makes it bigger, but it also means it's self-contained. Once it is compiled, you can run your app without the lib.
Code from a shared lib is not part of the executable. There are just some hooks in place to make the executable aware of the name of the lib it needs. In order to run your app, the shared lib has to be present in the lib search path (e.g. $LD_LIBRARY_PATH).
If you have two apps that share the same code, they can each link against a shared lib to keep the binary size down. If you want to upgrade parts of the app without rebuilding the whole thing, shared libs are good for that too.
Good overview of static, shared dynamic and loadable libraries at
http://www.yolinux.com/TUTORIALS/LibraryArchives-StaticAndDynamic.html
Some features that aren't really called out from comments I've seen so far.
Static linkage (.a/.lib)
Sharing memory between these compilation units is generally ok because they should(?will) all be using the same runtime.
Static linkage means you avoid 'dll hell' but the cost is recompilation to make use of any change at all. static linkage into Shared libraries (.so) can lead to strange results if you have more than 1 such shared library used by the final executable - global variables may exist multiple times and which one is used and when they are initialised can cause an entirely different hell.
The library will be part of the shipped product but obfuscated and not directly usable.
Shared/Dynamic libraries (.so/.dll)
Sharing memory between these compilation units can be hazardous as they may choose to use different runtime. This can mean you provide different Shared/Dynamic libraries based on the debug/release or single/multi threaded or...
Shared libraries (.so) are less prone to 'dll hell' then Dynamic libraries (.dll) as they include options for quite specific versioning.
Compiling against a .so will capture version information internal to the file (hard to fake) so that you get quite specific .so usage. Compiling against the .lib/.dll only gives a basic file name, any versioning is done managed by the developer (using naming or manually loading the library and checking version details by hand)
The library will have to ship with the final product (somebody else can pick it up and use it)
But now I often encounter third-party libraries where only .so files are available [...] and you just link against them without the -l switch, e.g. g++ foo.o /path/to/liblibrary.so.
JFYI, if you link to a shared library which does not have a SONAME set (compare with readelf -a liblibrary.so), you will end up putting the specified path of liblibrary.so into your target object (executable or another shared library), and which is usually undesired, for users have their own ideas of where to put a program and its associated files. The preferred way is to use -L/path/to -llibrary, perhaps together with -Wl,-rpath,/whatever/path/to if this is the final path (such pathing decisions are made by Linux distributions for example).
Should I prefer creating .so files for the users of my library?
If you distribute source code, the user will make the particular choice.
