[jdk17] RFR: 8269088: C2 fails with assert(!n->is_Store() && !n->is_LoadStore()) failed: no node with a side effect

[Nils Eliasson]

On Thu, 24 Jun 2021 07:39:52 GMT, Roland Westrelin <roland at openjdk.org> wrote:

> In PhaseIdealLoop::try_sink_out_of_loop(), a Store is not expected to
> be a candidate for sinking: if a Store is in a loop, there has to be
> at least a memory Phi keeping the Store in the loop. In the case of
> the failure, a Store has control set in a loop but the Store is not
> connected to a loop memory Phi. The reason for that is that the Store
> was moved out of loop by PhaseIdealLoop::try_move_store_after_loop()
> which moved the Store to the lca of its uses (other than the loop
> Phi), in this case another loop.
> 
> It doesn't look right that a Store is removed from a loop only to be
> added to another one so I propose fixing that by placing the Store
> right out of the loop it's being moved out of (similar to
> PhaseIdealLoop::try_sink_out_of_loop()) rather than at the lca of
> uses.

src/hotspot/share/opto/loopopts.cpp line 953:

> 951: #endif
> 952:             lca = place_outside_loop(lca, n_loop);
> 953:             assert(!n_loop->is_member(get_loop(lca)), "control must not be back in of loop");

assert comment needs adjusting

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

PR: https://git.openjdk.java.net/jdk17/pull/133