RFR: 8242029: AArch64: skip G1 array copy pre-barrier if marking not active

[Erik Österlund]

Hi Nick,

Note that you only need ISB when returning from call_VM, not call_VM_leaf. Leaf calls can’t safepoint, and hence the ISB is redundant. Arraycopy uses leaf calls. So while this optimization is great for this case, maybe removing ISB fronleaf calls has a wider effect.

It also appears to me that with Stuart’s new nmethod entry barriers enabled, ISB is never required on returns, as oop are no longer embedded in the instruction stream then (which is what the ISB protects against).

Thanks,
/Erik

> On 7 Apr 2020, at 09:19, Nick Gasson <nick.gasson at arm.com> wrote:
> 
> Hi,
> 
> Bug: https://bugs.openjdk.java.net/browse/JDK-8242029
> Webrev: http://cr.openjdk.java.net/~ngasson/8242029/webrev.0/
> 
> Currently on AArch64 the G1GC array copy pre-barrier unconditionally
> performs a VM call into G1BarrierSetRuntime, but this is a no-op if
> marking is not in progress.
> 
> This patch adds a check to skip the call unless marking is in
> progress. X86 already has this optimisation and I can't see a reason not
> to do it on AArch64 as well.
> 
> Tested jtreg hotspot_all_no_apps, jdk_core.
> 
> Results of ArrayCopy.arrayCopyObject* JMH benchmarks:
> 
> Before:
> 
>  Benchmark                                    Mode  Cnt    Score    Error  Units
>  ArrayCopy.arrayCopyObject                    avgt   15  117.307 ? 11.607  ns/op
>  ArrayCopy.arrayCopyObjectNonConst            avgt   15  107.786 ?  2.692  ns/op
>  ArrayCopy.arrayCopyObjectSameArraysBackward  avgt   15   76.381 ?  0.761  ns/op
>  ArrayCopy.arrayCopyObjectSameArraysForward   avgt   15   79.519 ?  3.433  ns/op
> 
> After:
> 
>  Benchmark                                    Mode  Cnt   Score   Error  Units
>  ArrayCopy.arrayCopyObject                    avgt   15  86.161 ? 6.150  ns/op
>  ArrayCopy.arrayCopyObjectNonConst            avgt   15  83.539 ? 0.682  ns/op
>  ArrayCopy.arrayCopyObjectSameArraysBackward  avgt   15  52.388 ? 0.732  ns/op
>  ArrayCopy.arrayCopyObjectSameArraysForward   avgt   15  54.619 ? 1.278  ns/op
> 
> The VM call overhead can be quite high on AArch64 as we insert a
> serialising ISB instruction on every native->Java return.
> 
> 
> Thanks,
> Nick