RFR: JDK-8287284: C2: loop optimization performs split_thru_phi infinitely many times [v3]

[Tobias Holenstein]

On Fri, 10 Nov 2023 16:34:27 GMT, Roland Westrelin <roland at openjdk.org> wrote:

>>> > > > I am ready to integrate. Do you agree @rwestrel ?
>>> > > 
>>> > > 
>>> > > What about a test case?
>>> > 
>>> > 
>>> > right! I added one now
>>> 
>>> Does that test case cause a crash with the current jdk code? We used to have an assert on the number of rounds of loop opts but we don't have it anymore AFAICT.
>> 
>> If not, you could think about adding an IR test that checks that we are not adding too many phis/loads. It's not perfect but at least something.
>
>> If not, you could think about adding an IR test that checks that we are not adding too many phis/loads. It's not perfect but at least something.
> 
> Or maybe parsing the output of `TraceLoopOpts`.

> > > > I am ready to integrate. Do you agree @rwestrel ?
> > > 
> > > 
> > > What about a test case?
> > 
> > 
> > right! I added one now
> 
> Does that test case cause a crash with the current jdk code? We used to have an assert on the number of rounds of loop opts but we don't have it anymore AFAICT.

You are right. I added an IR test-case now that fails without the fix. I plan to reintroduce that assert in another RFE

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

PR Comment: https://git.openjdk.org/jdk/pull/15536#issuecomment-1807864576