JVMCI and Tiered compilation

[Halimi, Jean-Philippe]

Dear all,

I am currently looking into performance numbers with Graal as a JVMCI/JIT compiler. I see very high interpreter usage with Graal. I have experimented enabling and disabling the tiered compilation knob to see the impact on performance.

However, I am a bit confused on how JVMCI compilers is supposed to be used in the tiered compilation policy, since we technically have only one compiler to use. Typically, Hotspot methods have 5 levels in its tiered mode:

 *   0 - interpreter (CompLevel_none)
  *   1 - pure C1 (CompLevel_simple)
  *   2 - C1 with invocation and backedge counting (CompLevel_limited_profile)
  *   3 - C1 with full profiling (CompLevel_full_profile)
  *   4 - C2 (CompLevel_full_optimization)

When using JVMCI, are we compiling from state 1 to 4 with the JVMCI compiler?

Thanks
-Jp