RFR (S): Invert matrix index computation

[Roman Kennke]

Am 13.04.2017 um 20:38 schrieb Aleksey Shipilev:
> Hi,
> 
> It makes sense to change matrix index computation like this:
>   (from*stride + to)  ---> (from + to*stride)
> 
> This is because we are most frequently scanning incoming connections for a given
> region, so the memory access is denser if we group matrix by "to". Requires
> trivial changes in all barrier implementations:
>   http://cr.openjdk.java.net/~shade/shenandoah/matrix-invert/webrev.01/
> 
> It does improve Partial GC preparation time, and therefore cuts the total pauses:
> 
> before:
>   Partial GC (net)  =    32.36 s (a =     8859 us) (n =  3653)
>     (lvls, us =     1328,     6875,     8809,    10742,    22201)
> 
> after:
>   Partial GC (net)  =    24.07 s (a =     6009 us) (n =  4005)
>     (lvls, us =      928,     4883,     6016,     7031,    39491)
> 
> Testing: hotspot_gc_shenandoah, benchmarks with -XX:+VerifyShenandoahMatrix
> 
> Thanks,
> -Aleksey
> 

Sounds fine.
Roman