[jdk16] RFR: 8258393: Shenandoah: "graph should be schedulable" assert failure

[Roman Kennke]

On Thu, 17 Dec 2020 08:20:20 GMT, Roland Westrelin <roland at openjdk.org> wrote:

> This is a shenandoah bug but the fix is in shared code.
> 
> At barrier expansion time, we need the raw memory state for the
> control at which the barrier is expanded. Before expansion, the memory
> graph is traversed and memory state for each control is recorded. In
> that process, if opportunities to improve the memory graph are found,
> the graph is modified. In the case of the failure, a raw memory load
> was assigned a call projection as control but for that control, no
> memory state was recorded and the fall back is to use the call's input
> memory state. As a consequence the load's memory input is updated to a
> wrong memory state. This causes the assert failure.
> 
> The call has 2 memory projections: one for the fallthrough and one for
> the exception path. One projection is the memory state for the
> CatchProj for the fallthrough, the other one for the CatchProj of the
> exception path. For the control projection out of the call, there's no
> memory state that makes sense and assigning the control projection of
> the call as control for the load is the root cause of the
> problem. This happens because anti-dependence analysis is too
> conservative: a memory Phi that merges states from the exception and
> fallthrough path is considered a dependency and that pushed the load
> right after the call.
> 
> I propose to be less conservative in anti-dependence analysis for
> Phis. For a Phi, when computing the LCA, I think it's sufficient to
> only consider region's inputs that we actually reach by following the
> memory edges and that's what I propose here.

As far as I can tell, the change looks good. I'm not very familiar with the code, it would be good to have somebody else check it too.
I verified that the tests that failed before are passing now and also have run some additional tests.

-------------

Marked as reviewed by rkennke (Reviewer).

PR: https://git.openjdk.java.net/jdk16/pull/41