Benchmark types revisited

[Aleksey Shipilev]

Hi guys,

We are trying to fold the remaining clutter in the API before moving on
to parameters. One of the culprits today is the @GMB accepting the
BenchmarkType, which is used by generator to build the synthetic code
for the particular mode.

This has a number of drawbacks:
  a) inability to select the BenchmarkType at runtime
  b) changing BenchmarkType requires recompilation
  c) you can't request the class-level BenchmarkType, which will apply
to all @GenerateMicroBenchmark
  d) you can only distinguish different benchmark types with the prefix
in the name

So, instead of this:

 class C {
    @GenerateMicroBenchmark(BenchmarkType.AverageTimePerOp)
    public void m1() {}

    @GenerateMicroBenchmark(BenchmarkType.AverageTimePerOp)
    public void m2() {}
 }

...we propose to move to plain @GMBs, letting JMH to generate the code
for *all* the modes at once, and providing us with the way to select the
mode at runtime:

 class C {
    @GenerateMicroBenchmark
    public void m1() {}

    @GenerateMicroBenchmark
    public void m2() {}
 }

...also, with the optional annotation to select the default mode:

 @BenchmarkMode(Mode.AverageTime)
 class C {
    @GenerateMicroBenchmark
    public void m1() {}

    @GenerateMicroBenchmark
    public void m2() {}
 }

...which also accepts multiple values:

 @BenchmarkMode({Mode.AverageTime, Mode.Throughput})
 class C {
    @GenerateMicroBenchmark
    public void m1() {}

    @GenerateMicroBenchmark
    public void m2() {}
 }

Running this benchmark would yield four runs: {m1, m2} x {AverageTime,
Throughput}. New command line option "-bm" can override the default
benchmark mode, e.g. running with "-bm AverageTime,SampleTime" will
yield {m1, m2} x {AverageTime, SampleTime} result.

The changeset with the proposed change is here:
  http://cr.openjdk.java.net/~shade/jmh/benchtypes-1.changeset

Please take a look, try to apply (hg import) and run with it.
Comments, feedbacks, questions are welcome.

-Aleksey.