Comment by mike_hearn

There were experiments done some years ago for automated stack allocation in the JVM, by Azul. They had actual hardware support for it, it was very interesting. Basically they stack allocated everything by default and then moved stuff onto the heap when required.

What they found is that it didn't make a big difference to performance. The problem is that generational GCs make allocating short lived small objects very cheap, basically as cheap as stack allocation. The main difference is that with stack allocation you are (potentially) better exploiting the cpu cache, but whether this makes a noticeable difference depends a lot on complicated things. It's not as clear a win as you would imagine and eventually Azul stopped bothering with it.

I'd still like to see more aggressive stack allocation in the JVM. However, I am not expecting it to be some kind of silver bullet.

Re: generics. It's true that type erasure causes some annoying problems. Oddly though, it has a big advantage too - it allows other JVM languages more flexibility to experiment with different approaches to generics. For example Kotlin uses a different style of generics to regular Java. If the JVM enforced the Java semantics that innovation wouldn't be possible, or at least, not at all easy. And of course one reason the Java world has a larger library ecosystem than .NET is that the .NET designers kept breaking backwards compatibility whereas the Java guys didn't. That's the reason JVM generics works the way it does. So the tradeoffs are rather subtle.

It's not quite true that the CLR was designed to be language agnostic and the JVM wasn't. The JVM was intended to be language agnostic right from day one. However Microsoft had to try and unify the worlds of Visual Basic, Java/C# and C++ which were all popular on Windows at the time, so they emphasised it more from the start. However VB.NET and C# are very similar languages with the bulk of the differences being down to syntax. In particular, VB.NET is not really the same language as classical VB at all (e.g. different handling of threads).

These days the JVM has lots of very different languages running on it with quite acceptable performance. They've done a lot of work on making very dynamic languages fast, through things like invokedynamic and Graal/Truffle. That's because those languages are very popular. They've done less work on functional languages, but I'm sure stuff they would benefit from will come with time.

The real benefit of value types, by the way, is control over memory layout and thus CPU cache usage. Java tends to lose benchmarks against C++ partly because C++ programs naturally lay out data in more cache friendly ways. Value types in the JVM/Java language are likely to make a big difference here, though it's hard to say exactly how much ahead of time.