FromIterable overhead reduction.

[akarnokd]

Some restructuring reduces the overhead of operators:

• extending AtomicLong gives access to atomic intrinsics for the request accounting
• loading the final fields into local variables prevents them from being reloaded from cache due to the atomics around them
• request() is hot but generally too large due to the fastpath/slowpath fit and JIT may not want to pick it up early. By refactoring the two paths into two separate methods, it becomes inline friendly for the either of the paths.

Benchmark results on my i7 4770K, Windows 7 x64, Java 8u51:

The benchmark from #3118 gives this result:

[davidmoten]

benchmark improvements are spectacular!

[akarnokd]

I don't want to merge this until somebody else verifies the improvement on a different processor/OS.

[stealthcode]

@akarnokd I can run it. How are you generating the benchmark reports in the image?

[akarnokd]

I have a tool for displaying JMH results: https://github.com/akarnokd/jmh-compare-gui . Then I take a screenshot and cut it around.

[stealthcode]

Here are the results from my laptop (2.2GHz intel Core i7; OS X Yosemite 10.10.4 Java 1.8u51).

Benchmark                          (size)   Mode   Samples        Score  Score error    Units
r.o.FromIterablePerf.direct             1  thrpt         5 22874735.514   569868.213    ops/s
r.o.FromIterablePerf.direct          1000  thrpt         5   287497.063    13353.846    ops/s
r.o.FromIterablePerf.direct       1000000  thrpt         5      278.303       45.706    ops/s
r.o.FromIterablePerf.from               1  thrpt         5  9434804.935   160084.549    ops/s
r.o.FromIterablePerf.from            1000  thrpt         5   226639.074    82633.950    ops/s
r.o.FromIterablePerf.from         1000000  thrpt         5      245.069        5.076    ops/s
r.o.FromIterablePerf.fromUnsafe         1  thrpt         5 23373598.288   769396.099    ops/s
r.o.FromIterablePerf.fromUnsafe      1000  thrpt         5   288430.485     8187.201    ops/s
r.o.FromIterablePerf.fromUnsafe   1000000  thrpt         5      286.498       11.960    ops/s

Seems comparable.