The JNI itable index cache is managed by the following functions
in src/share/vm/oops/instanceKlass.cpp:

instanceKlass::set_cached_itable_index(size_t idnum, int index)
instanceKlass::cached_itable_index(size_t idnum)

These functions were added by the fix for the following bug:

    6404550 1/1 Cannot implement late attach in NetBeans Profiler
                due to missing functionality in JVMTI

The fix for 6404550 was integrated in Mustang-B83.

The cache is used in

src/share/vm/prims/jni.cpp: jni_invoke_nonstatic()

    } else {
      // interface call
      KlassHandle h_holder(THREAD, holder);

      int itbl_index = m->cached_itable_index();
      if (itbl_index == -1) {
        itbl_index = klassItable::compute_itable_index(m);
        m->set_cached_itable_index(itbl_index);
        // the above may have grabbed a lock, 'm' and anything non-handlized can't be used again
      }
      klassOop k = h_recv->klass();
      selected_method = instanceKlass::cast(k)->method_at_itable(h_holder(), itbl_index, CHECK);
    }

Figuring out a method's itable index for a JNI invoke is expensive
so the idea is to cache the itable index on the first JNI invoke
of the method and then reuse it for subsequent invokes.

However, while I was investigating the following bug:

    6419370 4/4 new jmethodID code has tiny holes in synchronization

I noticed that set_cached_itable_index() was allocating a new,
full size cache array for every cache access attempt. On further
investigation, I figured out that the newly allocated cache was
missing the critical assignment to element zero:

    // array length stored in first element, other elements offset by one
    if (indices == NULL || (length = (size_t)indices[0]) <= idnum) {
      size_t size = MAX2(idnum+1, (size_t)idnum_allocated_count());
      int* new_indices = NEW_C_HEAP_ARRAY(int, size+1);
/*MISSING*/ new_indices[0] =(int)size; // array size held in the first element
      // Copy the existing entries, if any
      size_t i;
      for (i = 0; i < length; i++) {
        new_indices[i+1] = indices[i+1];
      }
      // Set all the rest to -1
      for (i = length; i < size; i++) {
        new_indices[i+1] = -1;
      }
      if (indices != NULL) {
        FreeHeap(indices); // delete any old indices
      }
      release_set_methods_cached_itable_indices(indices = new_indices);

The interesting part of this bug is that the missing
setting of the cache size means that we allocate a new
cache and free the old cache on every cache access. A
really stressful situation that has helped me reproduce
one of the races described in 6419370.