RFR: 7134701 [macosx] Support legacy native library names (original) (raw)

David Holmes david.holmes at oracle.com
Sun Apr 1 06:28:24 PDT 2012

Alan, Michael,

This sounds like a case where a property to control what the preferred library suffix is might be useful as a migration aid. Though the fundamental flaw in the API is assuming there is only a single suffix.

David

On 1/04/2012 11:16 PM, Alan Bateman wrote:

On 28/03/2012 10:52, Michael McMahon wrote:

Maybe, we should be doing what Dan Daugherty suggested yesterday and re-mapping the native filename in the case where a full absolute path is passed in (which is the code-path when System.load() is called above) My reasoning (for rejecting that) was that an absolute path is a request for an explicit name eg. System.load("/abs/path/libfoo.dylib") and it seems odd to go modifying the path that the user provided (as opposed to the System.loadLibrary("foo") case, where we have to construct the path internally). But, I hadn't considered the case above where the path is constructed by calling System.mapLibraryName(). So, I think a better approach might be to just to check for the old ".jnilib" suffix in all the cases, rather than changing System.mapLibraryName(). Otherwise, we'll have an inconsistency forever more between the output of that method and the library suffix that we want people to use. I've attached the jdk8 diffs for doing this. - Michael I checked Apple's JDK6 and it looks to me that they always favor .jnilib over .dylib. If I compile some native code to two shared libraries, one with the .dylib suffix and the other with .jnilib then with Apple's JDK6 it always seems to load the .jnilib version. If only the .dylib library exists then it loads that. Also, as per a previous post, the System.mapLibraryName method with Apple's JDK6 seems to favor .jnilib too. This makes wonder about the original patches in the bsd-port and macosx forest as the comments there suggest that ".jnilib" is legacy when it seems to have been the favored suffix in JDK6 at least. Anyway, for jdk7 & jdk8 it looks like we are making a break from the past and supporting .jnilib is just to give developers a chance to migrate. In that context Dan's suggestion and your latest patch seems reasonable but does lead a few anomalies - for example System.loadLibrary with an absolute path might succeed even though the library will appear to not exist if tested separately. Also I wonder about System.mapLibraryName where someone tests to see if the library exists before then invoke System.loadLibrary. My guess is that there isn't a perfect solution to this as there will always been anomalies in a migration like this. On the latest webrev then the use of final at L1867 is a bit odd and can be removed. For the jdk8 version then the fromClass needs to be Class<?>. -Alan

More information about the macosx-port-dev mailing list