RFR: 8290892: C2: Intrinsify Reference.reachabilityFence [v8]

[Emanuel Peter]

On Fri, 12 Sep 2025 13:08:49 GMT, Emanuel Peter <epeter at openjdk.org> wrote:

>>> could we just go through _reachability_fences, and hack the graph and clean up with IGVN? Or do we really need the loop state to do this successfully?
>> 
>> RF elimination needs control for referent to enumerate all interfering safepoints. 
>> 
>> Theoretically, it's possible to use a conservative estimate, but then:
>>  (1) it can worsen the result (by enumerating more interfering safepoints than needed); and
>>  (2) build an unschedulable graph if referent doesn't dominate safepoint node (if estimate is way too conservative). 
>> 
>> IMO it's safer to build full dominator tree here.  
>> 
>>> It probably has a performance impact, right? Have you measured that? 
>> 
>> It does have a noticeable cost. On my laptop it bumps the time spent doing RF processing from 170ms to 210ms
>> 
>> $ java -Xcomp -XX:-TieredCompilation -XX:+CITime -XX:+UnlockDiagnosticVMOptions -XX:-StressReachabilityFences
>> 
>>          IdealLoop:             0.173 s
>>            ReachabilityFence:   0.000 s
>>              Optimize:          0.000 s
>>              Eliminate:         0.000 s
>> ``` 
>> vs
>> 
>> $ java -Xcomp -XX:-TieredCompilation -XX:+CITime -XX:+UnlockDiagnosticVMOptions -XX:+StressReachabilityFences
>> 
>>          IdealLoop:             0.212 s
>>            ReachabilityFence:   0.030 s
>>              Optimize:          0.004 s
>>              Eliminate:         0.004 s
>> ``` 
>> 
>> I reimplemented it to piggyback on the last loop optimization attempt if there's any and it drastically improves the situation:
>> 
>> $ java -Xcomp -XX:-TieredCompilation -XX:+CITime -XX:+UnlockDiagnosticVMOptions -XX:+StressReachabilityFences
>> 
>>          IdealLoop:             0.193 s
>>            ReachabilityFence:   0.009 s
>>              Optimize:          0.003 s
>>              Eliminate:         0.004 s
>
> @iwanowww 
> Ok, thanks for measuring this. We really need to keep an eye on this, otherwise it will surely trip @robcasloz 's C2 compile time benchmarking eventualyl ;)
> 
> Can you point me to the code where you are actually using the dominator information? I think I did not find it the last time I reviewed.

Ah, you mentioned it somewhere else:
> It's solely for get_ctrl(referent) call in enumerate_interfering_sfpts().

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

PR Review Comment: https://git.openjdk.org/jdk/pull/25315#discussion_r2344214962