RFR: JDK-8263495: Gather liveness info in the mark phase of G1 full gc [v6]

[Hamlin Li]

On Wed, 17 Mar 2021 12:40:20 GMT, Stefan Johansson <sjohanss at openjdk.org> wrote:

>> Some additional sweep through the code showed some more minor issues. Sorry for not noticing earlier.
>
> Testing showed that the regression I reported in the other PR is still present for this fix. Same small reproducer can be used: 
> public class SystemGCLarge {
>   public static Object[] holder;
>   public static void main(String args[]) {
>     holder = new Object[128 * 1024 *1024];
>     System.gc();
>     System.out.println("Done");
>   }
> }
>  
> 
> I've made a change that eliminates the regression by making sure the cache update is not inlined and by doing so allows for other inlining to be done:
> https://github.com/kstefanj/jdk/commit/abf54fbbec5aa0b2fa36c2759e0974cb21d6cf78

> Testing showed that the regression I reported in the other PR is still present for this fix. Same small reproducer can be used:
> 
> ```
> public class SystemGCLarge {
>   public static Object[] holder;
>   public static void main(String args[]) {
>     holder = new Object[128 * 1024 *1024];
>     System.gc();
>     System.out.println("Done");
>   }
> }
> ```
> 
> I've made a change that eliminates the regression by making sure the cache update is not inlined and by doing so allows for other inlining to be done:
> [kstefanj at abf54fb](https://github.com/kstefanj/jdk/commit/abf54fbbec5aa0b2fa36c2759e0974cb21d6cf78)

On my local env, it seems make no difference wIth or w/o your patch, for both marking phase and whole full gc pause. But I will apply your patch, Thanks a lot for the benchmarking.

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

PR: https://git.openjdk.java.net/jdk/pull/2966