[~kshiroko]:
Need to improve the following test:

vm/mlvm/mixed/stress/regression/b6969574

1. Move from 'stress/regression' to 'perf/java'

2. Rebase to MicroBenchmarks v3 framework

3. Make debuggee method as simple as possible (something like "v = v + 2")

4. Generate (using JMPP) a loop for every test case (instead of common class) -- to minimize uncommon traps and code shared between cases.

5. Don't substract direct call time. Just compare it.

6. Maybe compare getters and setters

7. Use vm/mlvm/share/compiler framework to check that method has been compiled and not deoptimized

8. Account for GC (using the same framework), as we have GC activity for MH.invokeGeneric() with non-exact signature