RFR: 8341831: PhaseCFG::insert_anti_dependences asserts with "no loads" [v3]

[Igor Veresov]

On Fri, 11 Oct 2024 01:44:22 GMT, Dean Long <dlong at openjdk.org> wrote:

> > > LoadStore nodes should have the same issue. Why they are not affected?
> > 
> > 
> > Because LoadStore is an official store. It consumes a memory state and produces memory state. CacheWB is not really a store, that is it doesn't produce memory effects from the perspective of the backend (its match rule is not a Set). It's hard to tell what's the best way to model it, so I just decided not to mess with its semantics right now.
> 
> Should it be treated like a memory barrier?

I'm not sure why it's not, I guess they wanted a more relaxed behavior? It's more like the opposite of prefetch really. I didn't want to touch the semantics of it in this bug fix because it feels like it will likely open another can of worms.

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

PR Comment: https://git.openjdk.org/jdk/pull/21455#issuecomment-2406471308