Improve performance of ConcurrentReferenceHashMap creation

[stsypanov]

I've discovered that on some workloads instantiation of CCRHM takes significant amount of time (e.g. loading interface-base projections in Spring Data JPA). Here's the patch that reduces map instantiation time:

//before
Benchmark                   Mode  Cnt    Score   Error  Units
concurrentReferenceHashMap  avgt  100  712,179 ± 7,233  ns/op

//after
Benchmark                   Mode  Cnt    Score   Error  Units
concurrentReferenceHashMap  avgt   50  496,598 ± 4,601  ns/op

Benchmark is very simple:

@Benchmark
public ConcurrentReferenceHashMap concurrentReferenceHashMap() {
  return new ConcurrentReferenceHashMap();
}

Key ideas behind the changes:

• initialSize is the same at each loop iteration in constructor of ConcurrentReferenceHashMap, so expression 1 << calculateShift(initialCapacity, MAXIMUM_SEGMENT_SIZE); can be hoisted out of constructor of Segment and then out of the loop
• same for resizeThreshold calculation
• Segment.references declared as volatile so I suspect when new segment is written into ConcurrentReferenceHashMap.segments array the field segments must be loaded at each iteration due to HB semantics, same for this.segments.length. To avoid this I propose to populate ConcurrentReferenceHashMap.segments as local var and the write it to the field. Event if my assumption about HB is wrong reduction of interaction with fields helps us in interpreter mode.

[jhoeller]

Good points, I'll merge those changes right away! Thanks for raising this.