RFR (S): C2: Matrix update barrier math improvements

[Roman Kennke]

Am 10.03.2017 um 12:36 schrieb Aleksey Shipilev:
> Hi,
>
> Turns out, we can do a few math transformations for computing matrix index, that
> benefits performance. See here:
>   http://cr.openjdk.java.net/~shade/shenandoah/matrix-math/webrev.01/
>
> I think it only deserves to be in C2. Fighting C1 codegen to arrive to the same
> code is a waste of time, IMO. It seems to cut the matrix barrier costs almost in
> half:
>
> Benchmark                                Mode  Cnt   Score    Error  Units
>
> # Baseline (no barriers)
> Plain.test_Object                        avgt   15   4.683 ±  0.031  ns/op
> Plain.test_Object:cycles                 avgt    3  18.627 ±  1.466   #/op
> Plain.test_Object:instructions           avgt    3  29.023 ±  0.059   #/op
>
> # Baseline (-XX:+ShenandoahMatrix)
> Plain.test_Object                        avgt   15   6.629 ±  0.054  ns/op
> Plain.test_Object:cycles                 avgt    3  26.392 ±  2.251   #/op
> Plain.test_Object:instructions           avgt    3  45.047 ±  0.769   #/op
>
> # Patched (-XX:+ShenandoahMatrix)
> Plain.test_Object                        avgt   15   5.873 ±  0.038  ns/op
> Plain.test_Object:cycles                 avgt    3  23.373 ±  0.339   #/op
> Plain.test_Object:instructions           avgt    3  41.096 ±  0.333   #/op
>
> Generated code before/after:
>  http://cr.openjdk.java.net/~shade/shenandoah/matrix-math/baseline.perfasm
>  http://cr.openjdk.java.net/~shade/shenandoah/matrix-math/optimized.perfasm
>
> Testing: hotspot_gc_shenandoah, jcstress -m quick (matrix update) is running
>
> Thanks,
> -Aleksey
>
Great stuff! Yes!

Roman