There is a noticeable performance regression in several octane benchmarks due to JDK-8135251 that shows only when anonymous classes are enabled. It seems that while anonymous classes, while faster to load, have different optimization behaviour, causing different speed of optimized code.

As a suggested workaround, we may keep anonymous classes for eval and shell code (which often is throw-away code that is only run once) and disable it for compilation of ordinary scripts.