Refine linkToVirtual to a direct call when no vtable entry exists

[nbhuiyan]

MethodHandleTransformer refines linkToVirtual calls to the target
method when MemberName is known. The call kind of the refined
linkToVirtual call target was always treated as Virtual.
However, for private virtual methods and java/lang/Object;
type virtual methods, there is no corresponding entry in the vtable,
and for such methods the interpreter vtable index is set as 0.
The dispatch is not performed through the vtable entry, but instead
it is a direct call to the target method obtained from MemberName.

By checking if jitVtableOffset is positive, we can determine if
a call needs to be dispatched through the vTable. Otherwise, we
can handle it similar to linkToStatic.

Closes: #13430

Signed-off-by: Nazim Bhuiyan nubhuiyan@ibm.com

[nbhuiyan]

@0xdaryl @knn-k @Akira1Saitoh FYI

[nbhuiyan]

@0xdaryl requesting review

[0xdaryl]

Jenkins test sanity all jdk17

Getting a head start on the testing...

[0xdaryl]

Jenkins test sanity alinux64 jdk17

[0xdaryl]

I'm confused by this test. If the problematic interpreter vtable offset is "sizeof(J9Class)" or +376 then shouldn't this test be jitVTableOffset >= 0 ?

[nbhuiyan]

The problematic interpreter vtable offset is when it is set as 0, resulting in the jitVtableOffset being sizeof(J9Class). Normally, interpreter vtable offset is set a value > sizeof(J9Class) so the vtable offset we obtain will always be negative, which is why I thought a simple test for a positive integer value was sufficient. We can't have jitVtableOffset == 0 as that is the start of the J9Class. Do you think changing the operator to >= would provide more clarity?

[nbhuiyan]

Just for reference, here is where the interpreter vtable offset is set:

openj9/runtime/vm/createramclass.cpp

Lines 1407 to 1421 in 7ad8670

	/* Iterate over the vtable from the end to the beginning, skipping
	* the "magic" header entries. This ensures that the most "recent" override of the
	* method is found first. Critically important for super sends
	* ie: invokespecial) being correct
	*/
	J9VTableHeader *vTable = J9VTABLE_HEADER_FROM_RAM_CLASS(methodClass);
	J9Method **vTableMethods = J9VTABLE_FROM_HEADER(vTable);
	UDATA vTableIndex = vTable->size;
	while (vTableIndex > 0) {
	vTableIndex -= 1;
	if (method == vTableMethods[vTableIndex]) {
	return J9VTABLE_OFFSET_FROM_INDEX(vTableIndex);
	}
	}

J9VTABLE_OFFSET_FROM_INDEX takes the index in the vtable and adds sizeof(J9Class) + sizeof(J9VTableHeader), and the vtable entries begin following the J9VTableHeader. So whenever there's a vtable entry, the jitVTableOffset will end up being a negative value.

[0xdaryl]

Jenkins test sanity alinux64 jdk17

[0xdaryl]

Internal AArch64 testing succeeded (finally--after some infra problems).

[0xdaryl]

Please create a PR for the 0.28.0 branch.