RFR: Default to ShenandoahCodeRootsStyle = 2

[Roman Kennke]

I'm fine with this. Thank you!

Roman


> This is the follow up to the recent improvements in code roots handling in Shenandoah:
>   http://mail.openjdk.java.net/pipermail/shenandoah-dev/2018-May/006032.html
> 
> This time, we are better off to switch to style-2, which gives us parallel code roots iteration,
> plus CSet filtering. The reason why style-2 is better than style-1 is because it does not visit the
> nmethods that are known not to have roots in the cset. The performance runs on our set of benchmarks
> shows that the "Initial Evacuation" time, and thus the "Final Mark" time had dropped significantly.
> 
>             Initial evac time, us          Improvement
>          style-0   style-1   style-2   1-over-0   2-over-1
> Test A      6449      1596	 298       304%       436%
> Test B      6381      1656       243       285%       581%
> Test C       906       351       108       158%	      225%
> Test D      1021       490        88       108%	      457%
> Test E      1386       653       103       112%	      534%
> Test F      1022       702       664        46%	        6%
> Test G      2500       806        74       210%	      989%
> Test H      1500       688       568       118%	       21%
> Test I       696       463       220        50%	      110%
> Test J       708       336       101       111%	      233%
> Test K       604       330        96        83%	      244%
> Test L       701       403        88        74%	      358%
> Test M      1074       332       272       223%	       22%
> Test N       769       330       120       133%	      175%
> Test O       679       390       265        74%	       47%
> Test P      1332       518        78       157%	      564%
> Test Q      1188       461       174       158%       165%
> Test R      5475      1697       652       223%	      160%
> Test S      2413       926       142       161%	      552%
> 
> There are other potential improvements we can do in style-2 in future, but even this provides us
> with nice performance improvements.
> 
> Change:
> 
> diff -r ef24f20c2fae src/hotspot/share/gc/shenandoah/shenandoah_globals.hpp
> --- a/src/hotspot/share/gc/shenandoah/shenandoah_globals.hpp	Thu May 24 20:20:48 2018 +0200
> +++ b/src/hotspot/share/gc/shenandoah/shenandoah_globals.hpp	Fri May 25 12:21:25 2018 +0200
> @@ -207,7 +207,7 @@
>            "Evacuate code roots concurrently, instead of during a pause. "   \
>            "This requires ShenandoahBarriersForConst to be enabled.")        \
>                                                                              \
> -  experimental(uintx, ShenandoahCodeRootsStyle, 1,                          \
> +  experimental(uintx, ShenandoahCodeRootsStyle, 2,                          \
>            "Use this style to scan code cache:"                              \
>            " 0 - sequential iterator;"                                       \
>            " 1 - parallel iterator;"                                         \
> 
> Testing: hotspot_gc_shenandoah
> 
> Thanks,
> -Aleksey
>