Final RFR: 8005232 (JEP-149) Class Instance size reduction

[Peter Levart]

On 01/10/2013 02:25 AM, Mandy Chung wrote:
> On 1/6/2013 2:46 PM, David Holmes wrote:
>>
>> http://cr.openjdk.java.net/~dholmes/8005232/webrev/
>
> This looks good to me. David - besides the footprint performance data,
> do you observe any performance difference in your performance testing
> (I think refworkload is what you ran)?
>
> On 01/09/2013 01:19 PM, Aleksey Shipilev wrote:
>> On 01/09/2013 05:04 PM, Peter Levart wrote:
>>
>>> Strictly speaking, CAS is actually not needed here. Since we keep
>>> classRedefinedCount snapshot inside the ReflectionData, any races that
>>> would install "older" ReflectionData over "newer", would be quickly
>>> caught at next invocation to reflectionData(). So if there are any
>>> objections to usage of Unsafe, it can be removed and replaced by simple
>>> volatile write.
>>>
>> Yes, I think that would be more clear without any adverse impacts on
>> performance. Also, that better expresses the intent of requiring the
>> visibility, not the atomicity of cache update.
>
> I agree with Alekesey that it'd be simpler and clearer if CAS is replaced
> with the simple volatile write.  But this change will require another 
> round
> of performance measurement to determine its impact.  I'd suggest to 
> follow
> up this cleanup in a separate bug and changeset.
>
> Mandy
FWIW, I ran my micro benchmarks with the variant of using simple 
volatile write instead of CAS and there was no visible differences. But 
I don't have access to 256 core machines ;-). Anyway I think It should 
not present any performance difference, since the rate of class 
re-definitions is usually not very high...

Although the slow-path is a simple as:

this.reflectionData = new SoftReference<>(rd = new 
ReflectionData<>(classRedefinitionCount));
return rd;

... I had to split this too into a separate method to retain the same 
inlienability.

Such fine-tunnings tend to be lost in other sub-optimalities of the code 
that implements public reflection API, such as copying the arrays and 
array elements in bulk-returning methods and linear searches for the 
elements in single-element returning methods...

There is another simplification of code possible. The method 
checkInited() and its usages could be removed and it's logic 
incorporated into the slow-path newReflectionData() method while adding 
another boolean variable test into the fast-path.

Regards, Peter