RFR: 8353341: C2: removal of a Mod[DF]Node crashes when the node is already dead [v3]

[duke]

On Wed, 2 Apr 2025 08:08:07 GMT, Marc Chevalier <mchevalier at openjdk.org> wrote:

>> If the Mod[DF]Node has no control projection when it's being removed (because its result is unused), `extract_projections` will fail an assert. So, let's skip the removal.
>> 
>> But that should happen only when the nodes are already unreachable (control input being transitively top). At the end of the day, the node should be dropped. because of that, so there is no rush, and let dead node deletion do the job.
>> 
>> On the reduced reproducer, the crash is not common (even with `-XX:RepeatCompilation=300`, it might need more than a run to reproduce). So I've tried my fix on multiple thousands repeat compilations (by 300 packs) without a crash, and without having the modulo node alive at the end.
>> 
>> For instance, that's what happen on the reproducer. Quickly, some big sub-graph is dead, but nodes stay a while in the graph:
>> <img src="https://github.com/user-attachments/assets/7956771a-4bc4-4ab6-9207-3954cf067868" width="300">
>> Then:
>> <img src="https://github.com/user-attachments/assets/012868df-0d14-4292-aaeb-b342391ebbcd" width="300">
>> And eventually, everything is removed, so the control projection is removed, and `extract_projections` doesn't like it.
>> <img src="https://github.com/user-attachments/assets/cf20b745-cf86-48ba-94df-078aa50cdd45" width="300">
>> 
>> Thanks,
>> Marc
>
> Marc Chevalier has updated the pull request incrementally with one additional commit since the last revision:
> 
>   Address review comments, part 2

@marc-chevalier 
Your change (at version 48bd2037a9241f4c2956b19e91585553249e2625) is now ready to be sponsored by a Committer.

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

PR Comment: https://git.openjdk.org/jdk/pull/24375#issuecomment-2774836855