Hi Aleksey,
Wouldn't that change make a possible outcome of keySet() returning null
in case it was invoked concurrently? Wouldn't you have to use a local
variable to prevent that?
Regards, Peter
On 12/16/2015 11:53 PM, Aleksey Shipilev wrote:
Hi,
Since the dawn of OpenJDK, AbstractMap.keySet and .value were defined as
package-private volatile fields. Their only use is to cache keySet and
valueSet implementations from java.util collections.
However, all relevant java.util collections are not having any declared
fields (an opaque reference to enclosing class is stored in final
field), and they delegate straight to the backing collection. Therefore,
any race on cache field is benign, and we can drop "volatile" from the
fields:
https://bugs.openjdk.java.net/browse/JDK-8145539
http://cr.openjdk.java.net/~shade/8145539/webrev.02/
This improves performance for keySet()/values() on AbstractMap
implementations, because we don't emit barriers (volatile write in x86
case). This does not affect AbstractMap subclasses allocation
performance, because the volatile field values were default since
JDK-8035284.
See:
http://cr.openjdk.java.net/~shade/8145539/HashMapBench.java
Testing: microbenchmarks, java/util jtreg on Linux x86_64
Thanks,
-Aleksey
