H2JVM – A Haskell Library for Writing JVM Bytecode
discourse.haskell.org
discourse.haskell.org
var foo = expr ? new Foo() : new Bar();
What I am curios about is, if the java classfile API[1] provides a good model that could work in Haskell as well - I always had the impression that it was heavily influenced by functional programming, for example it uses "lifting transforms"[2].
PS: Good work on the label resolving part - this model is also used by the Java Classfile API and before it, ASM.
[1]: https://docs.oracle.com/en/java/javase/26/docs/api/java.base... [2]: https://docs.oracle.com/en/java/javase/26/docs/api/java.base...
I (kinda) get why someone might want to write Haskell rather than Java, but I'm just not sure why you would want to run Haskell on the JVM?
Java is an ok language, but what really makes it shine is the JVM. It's one of the fastest VMs out there and is one of the most customizable ones as well. For example, pretty much all other languages with a GC have just a GC and that's it. Java allows you to pick and choose your GC based on the workload.
It is one of the least limiting VMs out there because any knob you might want to tune, can be tuned. It's a huge value add.
I think the only part of the JVM that's not great is the fact that objects are bulky and the lack of value classes. Which ultimately means every struct like object you want can have a pretty hefty price in terms of memory. But otherwise, it's best in class basically for everything.
I dunno, I feel like I'm deploying a VM in a VM in a VM, and at some point they cost more than they taste.
Nope. Quite the opposite. You can precompile Java to a static runtime using the likes of Graal or even android. That actually makes these applications slower, not faster.
Part of this comes down to the Java language design. For Java, there is a lot of dynamic dispatch involved. In rust terms, it's as if almost every parameter was a `Box<dyn Foo>`. In Java, it's pretty natural to have a method like `void foo(List<Bar> baz){}` which can be called by any concrete list type. In fact, compiled down, this actually looks like `void foo(List baz) {}` in the bytecode.
The JVM is able to capture runtime information and realize "Oh, `foo` is always called with an `ArrayList`. And that `ArrayList` always emits a `Bar` element." That allows it to optimize and directly call the `ArrayList` methods rather than having to always do a "Ok, determine the type, look up the method table, call the method". Which is exactly what rust has to do with a `Box<dyn Foo>` signature. To do something similar with Rust you have to do a more complex PGO compilation.
Now, don't get me wrong, rust is smart. That's why they've designed the language such that `foo: Box<dyn Foo>` just isn't as ergonomic as `foo: &Foo`. The language pushes you to use the concrete structs when possible and to avoid doing dynamic dispatch. It supports it, but it requires a lot more ritual.
Comparing performance with a performance focused systems language such as rust wouldn't be fair either, and the borrow checker is Def cheating compared to GC, so this is also not good grounds to conclude that the JVM has a positive overhead either.
I think probably Go offers the fairer comparison, but my experience with it is so limited I can't really be sure. But unless I'm mistaken it's also GC'd, like Java, and runs without a VM, which is the difference we're trying to isolate.
Ergonomics wise my understanding is they're comparable too, but like I said, I have very little experience with go.
The performance analyses I've seen comparing the two seem to indicate a 10-15% performance increase in go compared to java, depending mostly on which report you read.
It's definitely not clear-cut, I probably wouldn't put much weight to such an argument if it didn't align so well with my intuition. I'm not free of bias. Maybe I just want something different after nearly 15 years of Java? I mean I definitely do, but maybe it's affecting my judgment more than I think?
But that's performance: the other limiting factor is that you need to install Java to run java apps(unless you go against nature and build sluggish native Java, but we covered that).I find I fairly often just don't want to have that requirement. Native executables have their charm, I find.
The main difference is that rust drives the VM all the way to the point of generating machine code while java generates the machine code at runtime.
Rust does a translation of the syntax to a high level bytecode, then a mid level bytecode, then to LLVM IR, and finally it lets LLVM do the translation of the IR to machine code. The way LLVM uses "VM" is exactly the same way Java is using "VM".
Javascript is similar. In fact, v8, the engine that powers node and chrome, was initially written by Java hotspot developers.
The current performance initiatives that Python and Ruby are taking are doing exactly what the JVM and Javascript does. In fact, the pypy JIT and LuaJIT are learning from and implementing what the JVM does. It's a proven mechanism to getting more performance and better optimizations.
Even GCC does the same thing under the hood.
It really is clear cut, more than you might expect.
I don't think there's any single language that most serious software is written in.
You get lots of things for free when targeting JVM bytecode. GCs, JITs, interop with one of the largest and most battle tested ecosystems.
Frege targets Java source code, which is then compiled by javac - the downside of that approach is you can not preserve the line numbers for debug information.