The is a no brainer test. You know for loop is hard to optimize for JIT engines and it's not a typical case at all. However, you can see that Dalvik JIT is doing great in the test and libacc is not bad, comparing with Dalvik fast and portable mode.

The inline assembly code used in JNI inline asm test is generated by LLVM (for loop LLVM IR generated by 10s lines of C++ code --> llc -march=arm -> put generated ARM asm code into gcc asm().