One possible reason of degradation is increasing size of some HashMap's methods. So those methods are not inlined, but at the same time inlining is necessary for scalar replacement. So we perhaps need to tune the inliner's parameters to make it working. I agree that we need to learn it more. But let's revert the patch for a while. Hope I can drill into the new HashMap on the next week. Another question I have is what VM did you use to run jbb2005 that showed an improvement?
-- Thanks Rustem
