RFR: 8209120: Archive the Integer.IntegerCache

Claes Redestad claes.redestad at oracle.com
Thu Aug 9 16:17:52 UTC 2018


Hi Peter,


On 2018-08-09 17:32, Peter Levart wrote:
> Hi Claes,
>
> When is this archived cache created? 

The cache you build when dumping the archive is serialized, so java 
-Xshare:dump -XX:AutoBoxCacheMax=20000 would create an array with 20k 
elements in the archive...

> Is it possible to create archived cache with 
> java.lang.Integer.IntegerCache.high system property set to > 127 ? 

Yes. I'm going to add a test that all variants work though, as Alan 
suggested.

> Wouldn't then at runtime, when cache is initialized from such 
> oversized archive and no java.lang.Integer.IntegerCache.high system 
> property is specified, the resulting cache array be to big?
>
> Perhaps you could allocate a copy of the prefix of the de-archived 
> array in that case to release the unused Integer instances.

The array would be too big, true, but it's backed by the read-only data 
in the archive directly, so while the Objects might appear in a heap 
dump, creating them happens at no extra cost and keeping them shouldn't 
actually cause any real (rss) memory footprint (GC shouldn't be scanning 
the memory, either).

As accesses to the array is guarded by checking IntegerCache.high (and 
not cache.length) we won't return cached Integers outside of the 
requested range, either.

So unless I'm mistaken, defensively allocating a copy shouldn't be 
necessary.

/Claes


>
> Regards, Peter
>
> On 08/09/2018 01:33 PM, Claes Redestad wrote:
>> Hi,
>>
>> using the new ability to archive immutable heap graphs into the CDS 
>> archive, even archiving small things like the Integer.IntegerCache 
>> can be profitable.
>>
>> Webrev: http://cr.openjdk.java.net/~redestad/8209120/open.00/
>>
>> Bug: https://bugs.openjdk.java.net/browse/JDK-8209120
>>
>> For a default cache (256 elements), this change saves us ~1 million 
>> instructions. Applications that use -XX:AutoBoxCacheMax can profit 
>> further (and we can use this to stress test the archiving feature..)
>>
>> The patch is built on top of and tested with Jiangli's patch for 
>> https://bugs.openjdk.java.net/browse/JDK-8207263 and I don't intent 
>> to push this before that is in.
>>
>> Thanks!
>>
>> /Claes
>>
>



More information about the hotspot-runtime-dev mailing list