RFR: 8255534: Shenandoah: Fix CmpP optimization wrt native-LRB [v2]

[Aleksey Shipilev]

On Wed, 28 Oct 2020 15:26:58 GMT, Roman Kennke <rkennke at openjdk.org> wrote:

>> JDK-8254314 introduced the following code:
>> 
>>     if (in1->bottom_type() == TypePtr::NULL_PTR &&
>>         (in1->Opcode() != Op_ShenandoahLoadReferenceBarrier || !((ShenandoahLoadReferenceBarrierNode*)in1)->is_native())) {
>>       in2 = step_over_gc_barrier(in2);
>>     }
>> 
>> However, the check for LRB and !native are the wrong way around: they should check if *in2* are not native LRB.
>> 
>> The bug is currently only observed in the conc-weakrefs branch, but may manifest (rarely) in mainline too. 
>> 
>> I am also putting in some refactoring to avoid duped code for good measure.
>> 
>> Testing: hotspot_gc_shenandoah (also in conc-weakrefs-branch where it manifests)
>
> Roman Kennke has updated the pull request incrementally with one additional commit since the last revision:
> 
>   Rewrite condition to be less confusing

Changes requested by shade (Reviewer).

src/hotspot/share/gc/shenandoah/c2/shenandoahBarrierSetC2.cpp line 1046:

> 1044: 
> 1045: void ShenandoahBarrierSetC2::maybe_step_over_cmpp_inputs(Node*& in1, Node*& in2) const {
> 1046:   // If first input is NULL, then skip the LRB barriers (except native) on the second input

OK, as we discussed offline, this also admits other barriers... so maybe:
`// If first input is NULL, then step over the barriers (except LRB native) on the second input`

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

PR: https://git.openjdk.java.net/jdk/pull/902