Seems like f1 is still sometimes uses (in field scenario?) even though we zero bytecode, so it should be re-resolved.

fixes #94

should be compatible with 7u91 too

should be compatible with 7u91 too

closes #81

Clear constant pool cache entries which are related to
invokehandle/invokedynamic and ones which are unresolved
(have both bytecodes set to 0).

There are two cases when such entries could be non-empty
(have data in f1/f2):
  1. invokedynamic/invokehandle-related entries in which f2 was
     set during constant pool cache initialization.
  2. put/get into static field when class having the field
     is not initialized yet.

Previously, we didn't clear such entries as clearing f2 in
first case results in crash. However, if we don't clear f1
in the second case, we will not be able to initialize class
for the second time (to re-create all enum constants, for
example) as non-empty f2 will cause assertion to fail.

The current implementation uses free bit in constant pool cache
entry to mark all entries with resolved indices (case #1).
The clearing code preserves f2 for all such entries.

Took 7u51 patch, applied, resolved conflicts.
jvmtiRedefineClasses.(c|h)pp were copied from a 7u51 working tree with patch
applied (generally, jvmtiRedefineClasses shouldn't be merged since they are
mostly rewritten).

Actually, update 51 patch applies to 55 with no test failures.

Separating changes for $StaticAccessor/$Accessor into separate patch.

In case matching field is not found, clear (set to null) references to faulty DMH.
Long term, should replace such DMH's with instances that will throw more meaningful
error messages at runtime.