They are still classes, still live on the heap and still need to be garbage collected. Compare with value types that live on the stack in other languages such as Swift, Go and Julia.
Java compilers are getting more and more advanced, but I don’t think they will ever become the magical “sufficiently advanced compiler” that produces code that’s as good as humans _could_ (but often won’t, because of time constraints) write.
I don’t think anybody fully disagrees with that. At least, I haven’t heard people claim int can be removed from the language because a good compiler can produce identical code for Integers.
And yes, that can also apply to instances that do escape. A sufficiently advanced compiler could in some/many cases figure out that an array of Integer can be compiled down to an array of int. However, it’s way easier for a compiler to check a programmer’s claim “we won’t use features of Integer on these ints” than to deduce that code won’t, so a little bit of programmer effort allows for a simpler compiler that can produce faster code.
For me, records and (future) value types are examples of such “little bits of programmer effort”
The first question is "through static analysis, can you guarantee that the structs do not leave the scope?"
The second question to look at is "which JVM are you using?"
Different JVMs may implement this differently. This isn't something that one can say about Java. It is something that one might be able to say about HotSpot, Zulu, or GraalVM.
> Something that Graal can do that C2 cannot, and a key advantage of GraalVM, is partial escape analysis. Instead of determining a binary value of whether an object escapes the compilation unit or not, this can determine on which branches an object escapes it, and move allocation of the object to only those branches where it escapes.
> The Java HotSpot Server Compiler implements the flow-insensitive escape analysis algorithm described in:
> ...
> After escape analysis, the server compiler eliminates the scalar replaceable object allocations and the associated locks from generated code. The server compiler also eliminates locks for objects that do not globally escape. It does not replace a heap allocation with a stack allocation for objects that do not globally escape.
----
So, some JVMs implement, others only do a limited subset of the optimizations available with escape analysis.
I would not say that the answer of "is it used in practice" is "no."
GraalVM is excellent in performing escape analysis on objects on the call stack, but it does not prevent the pointer overhead that a JVM array-of-heap-object-references has vs an array-of-structs that e.g. .NET supports [2].
Theoretically it could do hat, but that's just the classic "sufficient smart compiler" strawman [1]
Sure, but I think that's still important that it's possible. And if it doesn't get implemented, the reason may be because JVM developers have done the work to figure out that in the real world the optimization doesn't buy all that much.
Regardless, if you care about performance enough (via actual benchmarks) that you know that you really need some data to be guaranteed to be stack-allocated structs, then you probably shouldn't be using Java (or any GC'd language?) in the first place. Records don't change that calculus.
If it's a global, it's very likely allocated on the heap.
The question of "what is the representation of the object on the heap?" then open.
However, the "this is global" complicates it.
This isn't a question for Java to answer. You would need to dig into the specifics of the particular VM that you are using and how it allocates such a structure along with what optimizations it has available.
No they won't (or maybe they will be able to be speculatively opt-deopt?)
Value types above a relatively small size are less efficient than references.
With escape analysis, the compiler can allocate the data on the heap, stack, or even stick it in registers.
https://www.beyondjava.net/escape-analysis-java
https://shipilev.net/jvm/anatomy-quarks/18-scalar-replacemen...
https://www.javaadvent.com/2020/12/seeing-escape-analysis-wo...