I'm seeing a performance regression of roughly 20% when running Nashorn with various Octane benchmarks on Linux x64 with CompressedOops.

This was introduced in JDK1.8.0-b105 and Hotspot 25.0-b47, and I've tracked it down to hotspot commit 740e263c80c6 (8003424: Enable Class Data Sharing for CompressedOops...).

Before that commit, Nashorn runs roughly 10% faster with CompressedOops than without. After that commit, running with CompressedOops is now slower than without by about the same margin. The same is reproducible in JDK8 b106 and a recent jdk8-tl build.

I'm attaching a subset of the Google Octane benchmark to reproduce the problem. This runs the benchmark 20 times, higher score is better. I recommend running with tiered compilation disabled because it seems to make results more predictable.

java -server -XX:-TieredCompilation -XX:+UseCompressedOops -Xms2G -Xmx2G jdk.nashorn.tools.Shell run-richards.js