Memory region barrier may generate duplicate edges?

[qinsoon]

It seems memory region barrier does not log objects (possibly because it does not know object references).

mmtk-core/src/plan/barriers.rs

Lines 215 to 222 in a272c23

	#[inline(always)]
	fn memory_region_copy_post(
	&mut self,
	src: <S::VM as VMBinding>::VMMemorySlice,
	dst: <S::VM as VMBinding>::VMMemorySlice,
	) {
	self.semantics.memory_region_copy_slow(src, dst);
	}

It is unlike object write barrier.

mmtk-core/src/plan/barriers.rs

Lines 203 to 213 in a272c23

	fn object_reference_write_slow(
	&mut self,
	src: ObjectReference,
	slot: <S::VM as VMBinding>::VMEdge,
	target: ObjectReference,
	) {
	if self.log_object(src) {
	self.semantics
	.object_reference_write_slow(src, slot, target);
	}
	}

That means we may push duplicate edges to ProcessModBuf through memory region barrier. This will cause our edge logger (enabled by extreme_assertions) check fail.

[qinsoon]

@wenyuzhao Can you confirm this is a problem (just in case I misunderstood the issue)?

[wenyuzhao]

yes we don't log edges, and there will be duplicate edges. But this does not affect correctness.

[qinsoon]

Yeah, it does not affect correctness. However, we will need to get the edge logger working. I think we can check root edges and non root edges differently. We can treat edges from ProcessModBuf as roots, and for plans that use object barrier and ProcessModBuf, we allow duplicate root edges.

[caizixian]

We discussed that this behaviour is probably reasonable but should be documented