It may not be clear because it's very obvious but one of the primary optimizations in haskell that make it fast despite heavily using currying and lamba functions is that haskell doesn't have dynamic dispatch. This is completely unrelated to purity and lazyiness. This means it's possible to inline every lamba and partial function therefore use them at zero cost without the overhead of a JIT compiler or vm at runtime.

The JVM can do similar optimizations because it knows the class hierarchy at runtime and therefore declares every method as final until it finds a class with a method that overrides the previous definition. It's also capable inlining dynamic dispatch by profiling the type of the object it's dispatching on. If 99% of all calls go to class X then it can inline the method of class X and catch the remaining 1% with a fallback clause. Of course this comes at a cost. You need to retain the bytecode, class metadata, profiling information and duplicated code that is generated from inlining in memory.