Blackholes: improve symmetry for different data types and values

[Aleksey Shipilev]

Hi,

I have the external bug report that our Blackhole exhibit asymmetry in
performance when being fed with different values. Disassembly shows
weird things which seem hard to evade with our current blackholing
scheme with primitives.

With the headlines like that, I pushed:
 http://hg.openjdk.java.net/code-tools/jmh/rev/8046f2c4fce7

...which now provides th symmetrical performance against all the data
types and all the values, at the expense of few additional cycles per
consume(). Note that we already have lots of overheads on that scale
(the most prominent one being long "operations" counter incremented on
every iteration).

This is what we had before on 2x8x2 Sandybridge, Solaris 10, JDK 8b84:
(single threaded version only, it scales nearly-perfectly)

> 
> Benchmark                                       Mode Thr    Cnt  Sec         Mean   Mean error    Units
> o.o.j.b.BlackholeBench.baseline                 avgt   1      5    1        0.287        0.000  nsec/op
> o.o.j.b.BlackholeBench.explicit_testArray       avgt   1      5    1        3.050        0.001  nsec/op
> o.o.j.b.BlackholeBench.explicit_testBoolean     avgt   1      5    1        0.717        0.000  nsec/op
> o.o.j.b.BlackholeBench.explicit_testByte        avgt   1      5    1        0.716        0.000  nsec/op
> o.o.j.b.BlackholeBench.explicit_testChar        avgt   1      5    1        0.716        0.000  nsec/op
> o.o.j.b.BlackholeBench.explicit_testDouble      avgt   1      5    1        0.860        0.000  nsec/op
> o.o.j.b.BlackholeBench.explicit_testFloat       avgt   1      5    1        0.865        0.007  nsec/op
> o.o.j.b.BlackholeBench.explicit_testInt         avgt   1      5    1        0.716        0.000  nsec/op
> o.o.j.b.BlackholeBench.explicit_testLong        avgt   1      5    1        0.717        0.002  nsec/op
> o.o.j.b.BlackholeBench.explicit_testObject      avgt   1      5    1        3.050        0.001  nsec/op
> o.o.j.b.BlackholeBench.explicit_testShort       avgt   1      5    1        0.716        0.000  nsec/op
> o.o.j.b.BlackholeBench.implicit_testArray       avgt   1      5    1        3.053        0.001  nsec/op
> o.o.j.b.BlackholeBench.implicit_testBoolean     avgt   1      5    1        1.434        0.001  nsec/op
> o.o.j.b.BlackholeBench.implicit_testByte        avgt   1      5    1        0.716        0.000  nsec/op
> o.o.j.b.BlackholeBench.implicit_testChar        avgt   1      5    1        0.716        0.000  nsec/op
> o.o.j.b.BlackholeBench.implicit_testDouble      avgt   1      5    1        0.885        0.032  nsec/op
> o.o.j.b.BlackholeBench.implicit_testFloat       avgt   1      5    1        0.860        0.000  nsec/op
> o.o.j.b.BlackholeBench.implicit_testInt         avgt   1      5    1        0.717        0.001  nsec/op
> o.o.j.b.BlackholeBench.implicit_testLong        avgt   1      5    1        0.717        0.001  nsec/op
> o.o.j.b.BlackholeBench.implicit_testObject      avgt   1      5    1        3.052        0.001  nsec/op
> o.o.j.b.BlackholeBench.implicit_testShort       avgt   1      5    1        0.716        0.000  nsec/op

And this is what we will have now:

> Benchmark                                              Mode Thr    Cnt  Sec         Mean   Mean error    Units
> o.o.j.b.BlackholeBench.baseline                        avgt   1      5    1        0.287        0.000  nsec/op
> o.o.j.b.BlackholeBench.explicit_testArray              avgt   1      5    1        2.883        0.001  nsec/op
> o.o.j.b.BlackholeBench.explicit_testBoolean            avgt   1      5    1        2.882        0.011  nsec/op
> o.o.j.b.BlackholeBench.explicit_testByte               avgt   1      5    1        2.881        0.010  nsec/op
> o.o.j.b.BlackholeBench.explicit_testChar               avgt   1      5    1        2.882        0.010  nsec/op
> o.o.j.b.BlackholeBench.explicit_testDouble             avgt   1      5    1        2.881        0.009  nsec/op
> o.o.j.b.BlackholeBench.explicit_testFloat              avgt   1      5    1        2.883        0.002  nsec/op
> o.o.j.b.BlackholeBench.explicit_testInt                avgt   1      5    1        2.879        0.001  nsec/op
> o.o.j.b.BlackholeBench.explicit_testLong               avgt   1      5    1        2.880        0.001  nsec/op
> o.o.j.b.BlackholeBench.explicit_testObject             avgt   1      5    1        2.879        0.001  nsec/op
> o.o.j.b.BlackholeBench.explicit_testShort              avgt   1      5    1        2.877        0.009  nsec/op
> o.o.j.b.BlackholeBench.implicit_testArray              avgt   1      5    1        2.877        0.009  nsec/op
> o.o.j.b.BlackholeBench.implicit_testBoolean            avgt   1      5    1        2.877        0.010  nsec/op
> o.o.j.b.BlackholeBench.implicit_testBoolean_false      avgt   1      5    1        2.877        0.009  nsec/op
> o.o.j.b.BlackholeBench.implicit_testBoolean_falseF     avgt   1      5    1        2.877        0.009  nsec/op
> o.o.j.b.BlackholeBench.implicit_testBoolean_true       avgt   1      5    1        2.880        0.001  nsec/op
> o.o.j.b.BlackholeBench.implicit_testBoolean_trueF      avgt   1      5    1        2.879        0.001  nsec/op
> o.o.j.b.BlackholeBench.implicit_testByte               avgt   1      5    1        2.877        0.009  nsec/op
> o.o.j.b.BlackholeBench.implicit_testChar               avgt   1      5    1        2.879        0.001  nsec/op
> o.o.j.b.BlackholeBench.implicit_testDouble             avgt   1      5    1        2.879        0.001  nsec/op
> o.o.j.b.BlackholeBench.implicit_testFloat              avgt   1      5    1        2.877        0.009  nsec/op
> o.o.j.b.BlackholeBench.implicit_testInt                avgt   1      5    1        2.879        0.001  nsec/op
> o.o.j.b.BlackholeBench.implicit_testLong               avgt   1      5    1        2.879        0.001  nsec/op
> o.o.j.b.BlackholeBench.implicit_testObject             avgt   1      5    1        2.879        0.000  nsec/op
> o.o.j.b.BlackholeBench.implicit_testShort              avgt   1      5    1        2.878        0.001  nsec/op

We can still get it down to 1.6ns per call, but this will only work for
implicit blackholes, and I'm contemplating if I want to break the
symmetry in this way.

-Aleksey.