[jira] Commented: (HARMONY-5584) [drlvm][jit][opt][performance] Inliner heuristics improvements: hotness and instance initializer bonuses

["Cheng, BuQi (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/HARMONY-5584?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12638492#action_12638492 ] 

Cheng, BuQi commented on HARMONY-5584:
--------------------------------------

Aleksey:

   There is no reason that MAX_INLINE_GROWTH_FACTOR_PROF will be defined as 2000. The bytecode size change before and after inline nextDouble are methodByteSize=52, newByteSize = 120, factor=1.76471.  After changing MAX_INLINE_GROWTH_FACTOR_PROF to 500. The nextDouble is inlined also.

    However, if  MAX_INLINE_GROWTH_FACTOR_PROF is defined as 2000. We can see more perfomance improvement. Original, 300 ops/min, Patched with MAX_INLINE_GROWTH_FACTOR_PROF=500: 346.98 ops/min,  Patched MAX_INLINE_GROWTH_FACTOR_PROF=2000: 409 ops/min. Obviously, more methods are inlined. However, I don't think we can get so much performance by inlining a method with more than 2000 instruction. So I guess the benifit must comes from the nested inlined region. I will do more study on it to find out the problem of inline policy.
   
Thanks!

Buqi


> [drlvm][jit][opt][performance] Inliner heuristics improvements: hotness and instance initializer bonuses
> --------------------------------------------------------------------------------------------------------
>
>                 Key: HARMONY-5584
>                 URL: https://issues.apache.org/jira/browse/HARMONY-5584
>             Project: Harmony
>          Issue Type: Improvement
>          Components: DRLVM
>         Environment: all
>            Reporter: Aleksey Shipilev
>            Assignee: Mikhail Fursov
>         Attachments: H5584-inliner-heuristics.patch, H5584-inliner-heuristics.patch
>
>
> During the profiling of Scimark2 [1] it was found that inliner misses the inline opportunities for monte_carlo. 
> This is the code for hottest method:
> 	public static final double integrate(int Num_samples)
> 	{
> 		Random R = new Random(SEED);
> 		int under_curve = 0;
> 		for (int count=0; count<Num_samples; count++)
> 		{
> 			double x= R.nextDouble();
> 			double y= R.nextDouble();
> 			if ( x*x + y*y <= 1.0)
> 				 under_curve ++;
> 			
> 		}
> 		return ((double) under_curve / Num_samples) * 4.0;
> 	}
> Problems are:
>  1. nextDouble is not inlined
>  2. nextDouble is internally synchronized 
> Attached patch solves these two problems:
>  1. fixing hotness bonus calculation. This is the issue Random.nextDouble() hit on: this method called in hot cycle and missing inline opportunity due to glitch in scaling function, which should be more additive rather than multiplicative.
>  2. introducing instance initializer bonus. This helps to inline constructor with all corresponding methods to help scalar replacement with scalarizing of R (and eliminating synchronization too).

-- 
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.