RFR: 8303279: C2: crash in SubTypeCheckNode::sub() at IGVN split if [v2]

[Roland Westrelin]

On Thu, 29 Jun 2023 07:29:15 GMT, Roland Westrelin <roland at openjdk.org> wrote:

>> The crash occurs because at split if during IGVN, a `SubTypeCheck` is
>> created with null as input. That happens because the control path the
>> `SubTypeCheck` is cloned for is dead. To fix that I propose delaying
>> split if until dead paths are collapsed.
>> 
>> I added an assert to check a nullable first input to `SubTypeCheck`
>> nodes (which should be impossible because it should be null
>> checked). When I ran testing, a number of cases showed up with known
>> non null values non properly marked as non null. I fixed them.
>
> Roland Westrelin has updated the pull request incrementally with two additional commits since the last revision:
> 
>  - review
>  - review

Thanks for reviewing this and running tests.

> The testing revealed a failure in the newly introduced assertion (attached logs to the bug).

I pushed a fix. The problem is that the return type of a boxing method is marked not null but if the method is inlined late, the result from inlining may not  have a non null type.

> Also, the bug summary is way too generic and lacks any details about the actual problem. Please, update it.

Does the new one look ok?

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

PR Comment: https://git.openjdk.org/jdk/pull/14678#issuecomment-1612555774