JNI functions "CallNonvirtual*" probe is incorrect

[Mark Wielaard]

Hi Yasumasa,

On Thu, 2011-04-07 at 16:24 +0900, Yasumasa Suenaga wrote:
> Hi,
> 
> I've tried to test my program with SystemTap.
> When I probed CallNonvirtualVoidMethod(), I seem to get incorrect value through "hotspot_jni.stp".
> [...]
> "clazz" does not exist in hotspot_jni.stp, and methodid points clazz.
> In HotSpot source code (hotspot/src/share/vm/prims/jni.cpp) defines CallNonvirtual functions as folllowing:
> 
> ------------
> #define DEFINE_CALLNONVIRTUALMETHOD(ResultType, Result, Tag) \
> \
>   DT_RETURN_MARK_DECL_FOR(Result, CallNonvirtual##Result##Method, ResultType);\
>   DT_RETURN_MARK_DECL_FOR(Result, CallNonvirtual##Result##MethodV, ResultType);\
>   DT_RETURN_MARK_DECL_FOR(Result, CallNonvirtual##Result##MethodA, ResultType);\
> \
> JNI_ENTRY(ResultType, \
>           jni_CallNonvirtual##Result##Method(JNIEnv *env, jobject obj, jclass cls, jmethodID methodID, ...)) \
>   JNIWrapper("CallNonvitual" XSTR(Result) "Method"); \
> \
>   DTRACE_PROBE4(hotspot_jni, CallNonvirtual##Result##Method__entry, env, obj, cls, methodID);\
> ------------
> 
> $arg3 points cls, $arg4 points methodID.
> 
> In order to fix this problem, I made a patch for "hotspot_jni.stp.in" .
> The patch that attached this mail(hotspot_jni.patch) works well on RHEL6.

Of course. CallNonVirtual methods take an extra jclass argument and
somehow we missed this. Thanks so much for finding the issue and
providing a patch. It looks good. I'll test it a bit and will commit it.

Thanks,

Mark