Use JNI instead of JNA to call native code?

It's difficult to answer such a generic question. I suppose the most obvious difference is that with JNI, the type conversion is implemented on the native side of the Java/native border, while with JNA, the type conversion is implemented in Java. If you already feel quite comfortable with programming in C and have to implement some native code yourself, I would assume that JNI won't seem too complex. If you are a Java programmer and only need to invoke a third party native library, using JNA is probably the easiest path to avoid the perhaps not so obvious problems with JNI.

Although I've never benchmarked any differences, I would because of the design, at least suppose that type conversion with JNA in some situations will perform worse than with JNI. For example when passing arrays, JNA will convert these from Java to native at the beginning of each function call and back at the end of the function call. With JNI, you can control yourself when a native "view" of the array is generated, potentially only creating a view of a part of the array, keep the view across several function calls and at the end release the view and decide if you want to keep the changes (potentially requiring to copy the data back) or discard the changes (no copy required). I know you can use a native array across function calls with JNA using the Memory class, but this will also require memory copying, which may be unnecessary with JNI. The difference may not be relevant, but if your original goal is to increase application performance by implementing parts of it in native code, using a worse performing bridge technology seems not to be the most obvious choice.

1. You are writing code a few years ago before there was JNA or are targeting a pre 1.4 JRE.
2. The code you are working with is not in a DLL\SO.
3. You are working on code that is incompatible with LGPL.

That is only what I can come up with off the top of my head, though I am not a heavy user of either. It also seems like you might avoid JNA if you wanted a better interface than the one they provide but you could code around that in java.

1. JNA does not support mapping of c++ classes, so if you're using c++ library you will need a jni wrapper
2. If you need a lot of memory copying. For example, you call one method which returns you a large byte buffer, you change something in it, then you need to call another method which uses this byte buffer. This would require you to copy this buffer from c to java, then copy it back from java to c. In this case jni will win in performance because you can keep and modify this buffer in c, without copying.

These are the problems I've encountered. Maybe there's more. But in general performance is not that different between jna and jni, so wherever you can use JNA, use it.

EDIT

This answer seems to be quite popular. So here are some additions:

1. If you need to map C++ or COM, there is a library by Oliver Chafic, creator of JNAerator, called BridJ. It is still a young library, but it has many interesting features:
   • Dynamic C / C++ / COM interop : call C++ methods, create C++ objects (and subclass C++ classes from Java!)
   • Straightforward type mappings with good use of generics (including much nicer model for Pointers)
   • Full JNAerator support
   • works on Windows, Linux, MacOS X, Solaris, Android
2. As for memory copying, I believe JNA supports direct ByteBuffers, so memory copying can be avoided.

So, I still believe that wherever possible, it is better to use JNA or BridJ, and revert to jni if performance is critical, because if you need to call native functions frequently, performance hit is noticeable.