RFR: 8232613: Move Object.registerNatives into HotSpot

[coleen.phillimore at oracle.com]

On 10/21/19 7:08 PM, David Holmes wrote:
> Hi Claes,
>
> My only nit with this is that I don't think register_native and 
> friends belongs in the SystemDictionary class as it has nothing to do 
> with the SD. This code seems to be all about Methods so that seems 
> like the place to put this.

I had the same discussion with Claes this morning. Yes, method.cpp is a 
better place for this, with find_prefixed_native a static private method 
inside method.cpp.

The change also looks good to me.
Coleen

>
> Thanks,
> David
>
> On 22/10/2019 12:55 am, Claes Redestad wrote:
>> Hi,
>>
>>   Object.java currently registers various native functions via the
>> registerNatives facility.
>>
>>      private static native void registerNatives();
>>      static {
>>          registerNatives();
>>      }
>>
>> Not costly in and off itself, but this has the side effect that these
>> two methods are taken into account every time the VM has to generate
>> default methods and overpasses for some class during class load, which
>> can take up substantial time. When prototyping JDK-8219713[1],
>> explicitly excluding these two methods showed some improvement to
>> default method generation, but adding special cases to the logic was
>> decided against.
>>
>> If we instead can get rid of the Object <clinit> and the registerNatives
>> altogether we get the same (or a slightly better) speedup. This
>> effectively reduces memory use and instructions retired doing default
>> method generation by 3-5%.
>>
>> Webrev: http://cr.openjdk.java.net/~redestad/8232613/open.00/
>> Bug:    https://bugs.openjdk.java.net/browse/JDK-8232613
>>
>> Testing:
>> - tier1-7
>> - Added a test which naively redefines Object and ensures native methods
>> keep working
>>
>> Thanks!
>>
>> /Claes