Link Time Optimization is a GCC feature that defers optimizations to link time. The compiler writes unoptimized intermediate code to object files. The linker can then apply optimizations across multiple object files. This can improve performance and reduce binary size.

I ran some experiments on the embedded ARM hard float binaries. Building without media or web for ARMv6 hard float, code size of stripped shared objects was reduced by 39% (1.4M to 844K) when using this optimization. The changes were:

Add -flto to compiler flags
Add -O2 -flto -fwhole-program to linker flags
Define JNIEXPORT in ${java.home}/include/linux/jni_md.h to be __attribute__((externally_visible))

The last change is important, because without it the linker will decide that no code in the shared objects can be referenced and will eliminate almost everything as being dead code. We need to set the attribute on all our exported functions. Doing it in jni_md.h is just a short-cut.

Using -Os instead of -O2 reduces shared object size to 792K (a further 7% compared to -O2 with LTO, for a total savings of 43% compared to -O2 without LTO. However, -Os isn't always going to be the best optimization choice.)