Advice for middleware vendors

Distributing middleware built with the NDK raises some additional issues that app developers do not need to worry about. Prebuilt libraries impose some of their implementation choices on their users.

Choosing API levels and NDK versions

Your users cannot use a minSdkVersion lower than yours. If your users' apps need to run on API 21, you cannot build for API 24. It is okay to build your library for a lower API level than your users. You can build for API 16 and remain compatible with your API 21 users.

NDK versions are largely compatible with each other, but occasionally there are changes that break compatibility. If you know that all of your users are using the same version of the NDK, it's best to use the same version that they do. Otherwise, use the newest version.

Using the STL

If you're writing C++ and using the STL, your choice between libc++_shared and libc++_static affects your users if you distribute a shared library. If you distribute a shared library, you must either use libc++_shared or ensure that libc++'s symbols are not exposed by your library. The best way to do this is to explicitly declare your ABI surface with a version script (this also helps keep your implementation details private). For example, a simple arithmetic library might have the following version script:

LIBMYMATH {
global:
    add;
    sub;
    mul;
    div;
    # C++ symbols in an extern block will be mangled automatically. See
    # https://stackoverflow.com/a/21845178/632035 for more examples.
    extern "C++" {
        "pow(int, int)";
    }
local:
    *;
};

A version script should be the preferred option because it is the most robust way to control symbol visibility. This is a best practice for all shared libraries, middleware or not, as it prevents your implementation details from being exposed and improves load time.

Another, less robust option is to use -Wl,--exclude-libs,libc++_static.a -Wl,--exclude-libs,libc++abi.a when linking. This is less robust because it will only hide the symbols in the libraries that are explicitly named, and no diagnostics are reported for libraries that are not used (a typo in the library name is not an error, and the burden is on the user to keep the library list up to date). This approach also does not hide your own implementation details.

Distributing native libraries in AARs

The Android Gradle plugin can import native dependencies distributed in AARs. If your users are using the Android Gradle plugin, this will be the easiest way for them to consume your library.

Native libraries can be packaged into an AAR by AGP. This will be the easiest option if your library is already built by externalNativeBuild.

Non-AGP builds can use ndkports, or perform manual packaging by following the Prefab documentation to create the prefab/ subdirectory of their AAR.

Java middleware with JNI libraries

Java libraries that include JNI libraries (in other words, AARs that contain jniLibs) need to be careful that the JNI libraries they include will not collide with other libraries in the user's app. For example, if the AAR includes libc++_shared.so, but a different version of libc++_shared.so than the app uses, only one will be installed to the APK and that may lead to unreliable behavior.

The most reliable solution is for Java libraries to include no more than one JNI library (this is good advice for apps too). All dependencies including the STL should be statically linked into the implementation library, and a version script should be used to enforce the ABI surface. For example, a Java library com.example.foo that includes the JNI library libfooimpl.so should use the following version script:

LIBFOOIMPL {
global:
    JNI_OnLoad;
local:
    *;
};

This example uses registerNatives via JNI_OnLoad as described in JNI Tips to ensure that the minimal ABI surface is exposed and library load time is minimized.