Benchmarkdotnet: IterationSetup is not running before each benchmark invocation

[stephentoub]

Repro using 0.10.14:
```C#
public class Test
{
public static void Main() => BenchmarkRunner.Run();

[IterationSetup]
public void MySetup() => Console.WriteLine("MySetup");

[Benchmark]
public void MyBenchmark()
{
    Console.WriteLine("MyBenchmark");
    Thread.Sleep(100);
}

}

Example output from the main run:

MainTarget 1: 16 op, 1608514852.97 ns, 100.5322 ms/op
MySetup
MyBenchmark
MyBenchmark
MyBenchmark
MyBenchmark
MyBenchmark
MyBenchmark
MyBenchmark
MyBenchmark
MyBenchmark
MyBenchmark
MyBenchmark
MyBenchmark
MyBenchmark
MyBenchmark
MyBenchmark
MyBenchmark
MainTarget 2: 16 op, 1605497716.92 ns, 100.3436 ms/op
MySetup
MyBenchmark
MyBenchmark
MyBenchmark
MyBenchmark
MyBenchmark
MyBenchmark
MyBenchmark
MyBenchmark
MyBenchmark
MyBenchmark
MyBenchmark
MyBenchmark
MyBenchmark
MyBenchmark
MyBenchmark
MyBenchmark
MainTarget 3: 16 op, 1610182705.89 ns, 100.6364 ms/op
MySetup
MyBenchmark
MyBenchmark
MyBenchmark
MyBenchmark
MyBenchmark
MyBenchmark
MyBenchmark
MyBenchmark
MyBenchmark
MyBenchmark
MyBenchmark
MyBenchmark
MyBenchmark
MyBenchmark
MyBenchmark
MyBenchmark
```

This leads to very incorrect benchmark results when MyBenchmark is doing something that ends up caching data that can affect subsequent invocations of MyBenchmark. The whole purpose of IterationSetup in my case is to ensure that cache is cleared before the benchmark is executed.

[AndreyAkinshin]

The problem: the default value of UnrollFactor is 16. It means that we call the target method 16 times in the loop body. We need this trick only for nanobenchmark (it prevents some microarch effects). Obviously, it doesn't make sense to use IterationSetup (which is not designed for nanobenchmarks) and UnrollFactor (which is disigned only for nanobenchmarks) together.
Thus, we have to:

• Use UnrollFactor=1 if IterationSetup or IterationCleanup is defined
• Write a validator which prevents running benchmarks with IterationSetup/IterationCleanup and UnrollFactor>1

[adamsitnik]

I have an idea for solving this problem:

Add new attribute: [BenchmarkSetup] which would be internally translated to [IterationSetup] with the right settings (Monitoring strategy etc)

@stephentoub would [BenchmarkSetup] name be less confusing?

[stephentoub]

In what scenario is the current behavior useful? It just seems buggy to me.

[AndreyAkinshin]

@adamsitnik I don't think that we should introduce a new attribute. We just shouldn't use both IterationSetup/IterationCleanup and UnrollFactor>1 together, it's an invalid config.

[Rizzen]

@AndreyAkinshin I guess that is not enough just change UnrollFactor due to calculating invokeCount during pilot stage.

[adamsitnik]

I will solve this issue soon, it is causing too much confusion.

When IterationSetup is used, the UnrollFactor=InvocationCount=1