JEP 450: Compact Object Headers | Hacker News Reader

JEP 450: Compact Object Headers

openjdk.org

196 pointsmfiguiere102 comments

Identity hashcodes and monitors are perennially tricky to implement with low space overhead and I think in hindsight that putting them at the base of the object model (i.e. every object has them) was a mistake.

WebAssembly GC objects do not have identity hash codes nor monitors to strive for the lowest space overhead in the base object model.

> Identity hashcodes and monitors are perennially tricky to implement with low space overhead and I think in hindsight that putting them at the base of the object model (i.e. every object has them) was a mistake.

I can understand the hashcode choice, but I never understood why they chose to add the ability to lock on any object. IMO it still is a bad choice now on server machines, and it certainly was in a language designed for embedded devices in the early 1990s.

Does anybody know what they were thinking? Were they afraid of having to support parallel class hierarchies with a LockableFoo alongside Foo for every class? If so, why? Or did they think most programs would have very few objects, and mostly use reference types?

> Does anybody know what they were thinking?

Ergonomics. (Plus an inherent assumption of mutable data.)

You can write `synchronized (this)` or `synchronized` class method, etc. and it just works.

If I were the Java1.0 authors, I probably would have:

- made a "monitor slot" its own primitive type;

- made two overloads of `synchronized` — synchronized(MonitorSlot), and synchronized(Object); where for synchronized(Object), the compiler expects to find a MonitorSlot-typed field with a special system name (e.g. "__jvmMonitorSlot") on the passed-in Object

- taken the presence of `synchronized (this)` in body code of a class to implicitly define such a field on the class, meaning you can "just write" `synchronized (this)`, since it becomes `synchronized (this.__jvmMonitorSlot)` and also triggers the creation of the __jvmMonitorSlot field

- explicitly define the __jvmMonitorSlot field on Class and other low-object-count system types

The only change from today would be that you can't point at some arbitrary Object you didn't define the class for yourself, and say "synchronize on that." Which... why do you want to be doing that again?

Yes with sufficient complexity, it's possible to achieve transparently the same result.

For me, the main question is where this complexity lives. To minimize the places where sharp knives need to be used, I would prefer to move this up into a language runtime and have the VM/engine underneath focus on implementing a simpler object model that has the mechanisms to pull off these tricks without resorting to unsafe tricks.

At that point, why not just make types you want lock functionality for be declared as 'synchronized' in their class definition:

    synchronized class Foo {
      //...
    }

Because, for the things that do care about synchronization, they might want multiple explicit MonitorSlot members. It makes more sense to just be able to synchronize on MonitorSlots directly, and then decide where they go.

The only reason I added the `synchronized (this)` allowance was because the parent said that they think that that's "good ergonomics" — and presumably, the Java1.0 authors also thought that — and I was trying to suggest an alternative that would preserve what they consider "good ergonomics."

But personally, if I was the sole dictator of Java1.0 language design with nobody else to please, I would just have `synchronize(MonitorSlot)` + explicitly-defined MonitorSlot members (that, if you declare one, must always be declared final, and cannot be assigned to in the constructor), and that's it. Just refer to them by name when you need one.

You underestimate the weight of that little "(Plus an inherent assumption of mutable data.)" side note in pauldraper's post: that one is really huge and the concise form of his post doesn't really do justice to how clearly it answers the entire sub-thread.

Java was driven by the admission that multithreading is hard, and the hybris that this hard problem would go away once and for all if only you added enough OOP state hiding and locks. All those strategies we now use to allow parallelism were either unknown back then, or unfashionable, because they were in conflict with that OOP dream world of "objects talking to each other". The ease of just writing synchronized everywhere and declaring done on the topic of multithreading was as much a selling point of early Java as memory safety.

Yes, but the number of things you want to synchronize on in real-life code is limited, and, certainly at the time, memory was scarce. I think only allowing ‘synchronized’ on containers and allowing programmers to opt-in on them would have been the better choice.

> Plus an inherent assumption of mutable data.

Yet, Strings and the boxed value types (Integer, Long, etc.) are immutable, and you can synchronize on them (or does this matter less for those because of the granularity of the memory allocator?)

> Strings and the boxed value types (Integer, Long, etc.) are immutable, and you can synchronize on them

Side note: That is a very bad idea, because their object identity is iffy.

At the time, there was a widespread assumption that massively parallel computing was the future and thus that any serious language had to try and integrate concurrency as a core feature.

It wasn't just Java. For many years THE argument for functional languages was you'd better learn them because these languages will soon be automatically parallelized, and that's the only way you'll be able to master machines with thousands of cores.

With hindsight we know it didn't work out like that. We got multi-core CPUs but not that many cores. Most code is still single threaded. We got SIMD but very little code uses it. We got GPUs with many cores, but they are programmed with imperative fairly boring C-like languages that don't use locking or message passing or anything else, they're data parallel pure functions.

But at the time, people didn't know that. You can see how in that environment of uncertainty "everything will be massively multi-threaded so let's give everything a lock" might have made sense

> We got SIMD but very little code uses it.

Because programming for SIMD is wildly different than MIMD. (And MIMD is what `synchronized` is for.)

The theory was that there would be more machines like AMD Thread Ripper.

Neither will Java assuming the next stage of this project goes as planned. The overhead for them will only be allocated on demand for those objects that need them.

I get that monitors are inflated dynamically, but from my reading of the linked JEP, Compressed Object Headers have something like 24 bits allocated for the hashcode in the 64 bit word?

Right, but that's the first phase. In the next, I believe the goal is to attempt to reduce the header to 32 bits, and then the header overhead (for all objects) for both monitors and hashcodes will be ~3 bits, basically just to indicate whether or not they're used. For objects where they're used, they will be stored outside the header.

That's pretty cool. These things are super tricky and are hard to get into a high-performance production system, so I respect the journey :-)

For Wasm GC, I think we need programmable metaobjects to be able to combine language-level metadata with the engine-level metadata. I have only a prototype in my head for Virgil and plan to explore this in Wizard soon.

Towards the bottom of the JEP, they mention that the ultimate goal is 32 bit object headers, which would necessitate object monitors be tracked on-demand in a side table. That's what the parent was getting at.

Given https://shipilev.net/jvm/anatomy-quarks/26-identity-hash-cod...:

> For identity hash code, there is no guarantee there are fields to compute the hash code from, and even if we have some, then it is unknown how stable those fields actually are. Consider java.lang.Object that does not have fields: what’s its hash code? Two allocated Object-s are pretty much the mirrors of each other: they have the same metadata, they have the same (that is, empty) contents. The only distinct thing about them is their allocated address, but even then there are two troubles. First, addresses have very low entropy, especially coming from a bump-ptr allocator like most Java GCs employ, so it is not well distributed. Second, GC moves the objects, so address is not idempotent. Returning a constant value is a no-go from performance standpoint.

How else could identity hashcodes be implemented? Is it just impossible to put a WebAssembly GC base-object as a key into a map?

The idea is that if your language has an identity hash code for every object, the language runtime can just add a field in the Wasm struct for it. There's nothing special about that field; it could be an 8-bit, 16-bit, 32-bit field, etc. For the lock inflation logic and so on, you can make a "tagged pointer" in Wasm GC by using the i31ref type, so you could do something like have only the identity hash code by default, but "inflate" to a (boxed) indirection with an additional monitor dynamically. But the Wasm engine just treats it all as regular fields. The overall idea is the Wasm GC gives you the mechanisms by which you can implement your specific language's needs, hopefully as efficiently as a native VM could.

One proposed strategy (https://wiki.openjdk.org/display/lilliput#Main-Hashcode) for dealing with hashes is to compute them lazily, taking advantage of the fact that most objects are never hashed: initially use the object's address, but also set a flag on the object when it is first hashed; when the gc next moves the object, it adds an extra field in which the original address is cached.

Another strategy, used by sbcl, is to rehash identity hash tables after each gc cycle.

How could I implement either of these with wasm?

> when the gc next moves the object, it adds an extra field in which the original address is cached.

Under this scheme, if I allocate object A at address 0, then GC-move it to address 100 (such that it caches that it was "originally at" address 0); and then, with object A still alive, I allocate object B at address 0... then don't objects A and B both now have 0 as their identity hashcode?

(I'm guessing the answer here is "this only works with generational GC, and the GC generation sequence-number is an implicit part of the stateless-address hashcodes and an explicit part of the cached-in-member hashcodes")

Moving GCs usually copy the object to a new region, and redirect accesses to the old one to the new region. These regions are often used as generations, but they could just be two halves of the heap that flip between active and inactive state.

I wondered why .NET let you lock any object, it seems like a strange feature. I figured they just had a spare bit in a bitset somewhere. Should've known it was copied from Java.

In practice (IIRC, I'd have to check again to be 100% sure) there's a little 'extra garbage' (my name, the real name is something else) pointer inside .NET object headers, and if your object has an identity hashcode or has been used with locking, the extra garbage pointer points to a separate heap allocation containing the hashcode and the locking info. That way you're not paying for it for every object. I think in some cases there are optimizations to be able to store the hashcode inline (tagging, etc). So it ends up being similar to these 'compact object headers' described here in terms of size but is done by making those optional features more expensive instead of by bitpacking stuff into the object header.

Caveat: There are multiple .NET implementations so what I've said may not apply to all of them

As the JEP explains, the JVM has done both those things for a long time. In particular the object doesn't actually pay the price of a lock unless it's actually locked at some point. It used to be the case that the JVM was even more extreme and you wouldn't pay the price of a lock unless the object was actually contended, this was called biased locking, but the code complexity to implement it was eventually determined to be no longer worth it on modern hardware.

Java is also considering making value objects that don't have that overhead:

https://openjdk.org/jeps/8277163

.NET has something like that already, but it would make a huge difference for things like Optional or if you want to make a type for, say, complex numbers where there is just 64-bits of data (for FP32) and even a 64-bit object header would be 100% overhead. Value objects could be possibly allocated on the stack and completely bypass the garbage collector in some cases.

Note that with that overhead, Java is an environment in which you can write multithreaded applications with good reliability and scaling, something that WebAssembly definitely isn't.

I've been lightly following some of their work here. From the outside, it seems like a lot of this has been born from optimizing the JVM over the past few decades, running into fundamental problems with how dynamic the JVM can be, then targeting changes for making that dynamicism opt-in, either through JVM flags or new features.

I do think some of it is kinda interesting: some people might say "if you never do X, we can guarantee Y". But it's nice, as a programmer, to be told "if you use feature A, you will never do X, which means we can guarantee Y". It's a lot more comforting that I can't slip up and accidentally do X.

> It's a lot more comforting that I can't slip up and accidentally do X.

Indeed, this is why value semantics (i.e. structural equality) for language constructs like ADTs is so wonderful. Because a program can never observe "identity", which is an implementation detail, the implementation doesn't have to use objects at all underneath. That opens a whole host of value representation options that aren't otherwise available.

Monitors are especially strange when I learned Java. I mean, why would you put a synchronization feature into every object? It's not necessary in the vast majority of the time.

Maybe the programming style at the time had something to do with it. Maybe they thought every class needs to be thread safe and mutable.

This was over 30 years ago when there was still a lack of awareness that locks don't compose [0]. Look at the early JRE classes like Vector, Hashtable, and StringBuffer, whose operations are synchronized. The idea was that since you'd potentially want to synchronize access to any mutable object (because you'd want to be able to share any object between threads), and most objects are mutable, that the most convenient solution would be for every object to have that functionality built in.

[0] https://en.wikipedia.org/wiki/Lock_(computer_science)#Lack_o...

Yeah, I asked why there're no Hashable & Equatable interfaces instead of putting things into Object which would make more sense to me. People responded that's just how things should be. Apparently not. IMO object identity is almost always not needed and a sign of bad design. You either should write proper hash code, or you should not use data structures which use hash code.

> You either should write proper hash code, or you should not use data structures which use hash code.

There's also the question of which hash code. Do you use a fast but low quality hash function, or a slower but higher quality one? Does your hash function need to be secure against hash collision attacks? Does it have to be deterministic? The correct choice of hash function can depend on the data structure, and the same object might have to be hashed with different hash functions (or different hash function seeds) in the same program.

The reason is that it'd make HashMap and HashSet a lot less useful. If every type had to opt in, you'd be unable to use many types as keys or set entries for no better reason than the authors didn't bother to implement hash codes or equality. By providing reasonable identity-based versions of these, it increases the utility of the language at a tradeoff in memory usage.

You could implement the hashing code in a helper class and construct the HashMap with it.

Without a way to give objects identity you'd get stuck pretty fast as it's not guaranteed you have access to any data to hash. You'd have to break encapsulation. That would then hit problems with evolution, where a new version of a class changes its fields and then everything that uses it as a hashmap key breaks.

My experience with platform design has consistently been that handling version evolution in the presence of distant teams increases complexity by 10x, and it's not just about some mechanical notion of backwards compatibility. It's a particular constraint for Java because it supports separate compilation. This enables extremely fast edit/run cycles because you only have to recompile a minimal set of files, and means that downloading+installing a new library into your project can be done in a few seconds, but means you have to handle the case of a program in which different files were compiled at different times against different versions of each other.

JavaScript handles the "no identity hash" with WeakMap and WeakSet, which are language built-ins. For Virgil, I chose to leave out identity hashes and don't really regret it. It keeps the language simple and the separation clear. HashMap (entirely library code, not a language wormhole) takes the hash function and equality function as arguments to the constructor.

[1] https://github.com/titzer/virgil/blob/master/lib/util/Map.v3

This is partly my style too; I try to avoid using maps for things unless they are really far flung, and the things that end up serving as keys in one place usually end up serving as keys in lots of other places too.

Does that assume that you have only one key type and not an infinite sized hierarchy of child classes to deal with? If you had a map that took a Number as key, how many child classes do you think your helper class would cover and what extension framework would it use to be compatible with user defined classes?

The reality would be that in all but the most extreme cases emulating set semantics by iterating a list would then be considered good enough, "avoid premature optimization" and all that. In real life code, this would have eaten up any advantage of a shorter object header a hundred times.

yeah this is where traits instead of hierarchies become useful - I should be able to implement the Hash interface for an object I do not own and then use that object + trait going forward for HashMap and HashSet.

Java doesnt make this very composable

Should be noted that Rust (one of the most prominent languages with traits) doesn't allow you to implement a trait for an object you do not own. A common workaround is to wrap that object in your own tuple struct and then implement the trait for that struct.

(If you don’t own the trait either, that is. Your own traits can be implemented for foreign types.)

Rust’s approach to the Hash and Eq problem is to make them opt-in but provide a derive attribute that autoimplements them with minimal boilerplate for most types.

Also, Rust’s Hash::hash implementations don’t actually hash anything themselves, they just pass the relevant parts of the object to a Hasher passed as a parameter. This way types aren’t stuck with just a single hash implementation, and normal programmers don’t need to worry about primes and modular arithmetic.

That sounds like a use case for the decorator pattern?

Java uses this pattern in some places, for example you can usually pass a custom Comparator to anything that needs comparisons (like sorts).

Fully separating implementation of an interface from the data can create quite a lot of additional complexities. See my comment elsewhere about encapsulation and version stability.

> IMO object identity is almost always not needed and a sign of bad design.

Sometimes you have no choice. I had multiple times where I needed to store additional data for instances of class X but had not control over it, so I had to store it in a separate structure and keep track of things by object identity.

> You either should write proper hash code

And object identity fulfills all the requirements of a "proper" hash code.

Yeah, having equals and hashCode on the root Object class is Java's biggest mistake, IMO. Although for a slightly different reason: equality is usually context-dependent, but having equals as an instance method ties you to one implementation.

That is somewhat fixed by Comparable, but the fact that HashMap doesn't have a pluggable Hash override always bothered me.

Yes and: toString() and clone().

Further: default implementations should be synthesized by the runtime. Prevents subtle bugs when the equals(...) and hashCode() implementations are incorrect or missing. eg contract for HashMap.

To backfill, perhaps Object could extend new base class NakedObject, and implement your new interfaces Equalable, Hashable. Then classes which don't need identity can extend NakedObject.

Maybe the "canonical Object" was motivated by prior experience with Self or some such. It definitely calmed down people new to Java & OOP. I think it was the right call at the time.

Records are very useful for data-only objects though, as they do implement proper equals and hash.

Even in JDK 1.3 era I recognized that anyone could grab a lock on any object, not just the object itself, so that was a bit of a problem. I started experimenting with member variables called “lock” that were new Object(). It seemed pretty dumb at first but then I discovered lock splitting and it was off to the races. There’s always going to be some object with multiple concerns, especially if one or two functions involve both but others involve one (eg, add/delete versus generate data from the current state).

Am I misremembering that there was a time where the JRE lazily added locks to objects? I thought it was part of their long road to lock elision.

It does. It's called lock inflation. The lock data structure is more than just a handful of bits, so the object header can point to the real lock once allocated. You don't pay a full lock for every object, just the memory needed to determine if it's in use or not.

A consequence of following too much Smalltalk's model.

A better decision would have been to make them available via interfaces, like CLR ended up doing in .NET 2.0, although they still follow Java/Smalltalk as well, due to its heritage.

Java has just been on a roll with such substantial changes to fundamental pieces of the runtime! Great work to the teams that keep making improvements in backwards compatible manners.

This specific change seems to impact the JVM the runtime, not Java the language. Which is great for people like me, who heavily rely on JVM but couldn't care less about Java.

People may not care about your "carelessness about Java"...

I'm fairly sure that's fine, people get to care about whatever they want :)

Funny, those languages rely heavily on the Java ecosystem and libraries, and they better hope the Java ecosystem keeps thriving, improving current libraries and producing new libraries else they become impractical. Show me your pure Clojure production grade http servers or database drivers ;)

>Show me your pure Clojure production grade database drivers

Depends what you mean by "database driver", but what disqualifies XTDB or Datomic, both of which have use outside of Clojure yet are implemented entirely in Clojure with the exception of the Java wrappers so you can use them naturally within Java or Kotlin?

Even then, talking about "pure Clojure" is a non-starter because Clojure is not a self-hosted language (yes, Java is self-hosted if you consider Jikes RVM, a JVM implementation in pure Java). Even if Java ecosystem were to dwindle, people can rely on the JavaScript, .NET, or Dart ecosystems thanks to ClojureScript, ClojureCLR, and ClojureDart existing.

I use ClojureScript daily and am not really affected by what happens in Java land. With that said, I always welcome improvements to the JVM, because I actually like the Java ecosystem; it's the only one that doesn't make me split hairs...despite the dreaded module system.

Those two databases rely heavily on Java libraries, libraries that would make no sense and be impractical to implement in Clojure.

> rely heavily on Java libraries

If you're referring to the storage engines, those are not Java, and even then, most storage engines of a database will resort to a lower level language because of the need to care about memory layout and things that are simply not exposed to you in a high-level language like Clojure (or Java, unless you are somehow able to use Jikes RVM in production, I guess, but I'm not aware of ANYONE using a Java storage engine atop Jikes)

This is what I hate in guest language communities (a subset of them), they literally spit in the ecosystem that makes their beloved language possible in first place.

Granted, not everyone is like that and there are many nice folks, but those vocal ones, oh boy.

When some people (like me) say "Java", they mean the Java Platform, and they call what you call Java "the Java language". When you say JVM, I assume you mean the Java Platform minus the language (the language is ~3-5% JDK, the JVM is about 25%). People who use "the JVM" but not "Java" use ~97% of Java, i.e. the Java platform.

Java made a bunch of optimistic assumptions about the future of hardware when it was designed. UTF-16 strings, 64 bit pointers, enormous object headers.

While great for future-proofing, it's been hard to deny it's had overhead. Glad to see it's slowly getting undone: Compressed Oops, Compact Strings, now this.

64-bit pointers isn't really an optimistic assumption, it's just the normal pointer size on 64-bit systems.

It was optimistic because Java came out long before 64 bit systems were common.

Common in desktops, yes, but Sun being a UNIX workstation vendor they switched in the 90s.

Well, the 3 billion devices (r) Java ran on certainly weren’t Sun workstations.

Optimistic to represent object references with native pointers?

C and C++ usually do that too.

UltraSPARC says hello

The DEC Alpha was a 64 bit CPU architecture that came out in 1992. Java support it from Java 1.0 to ~ Java 3

Interestingly, I can't find any information on Google about this, but ChatGPT does support this. Unfortunately all the primary sources it gave me are for docs that no longer exist on the web. The Wayback machine was no help. The old web is dead :(

> Interestingly, I can't find any information on Google about this, but ChatGPT does support this. Unfortunately all the primary sources it gave me are for docs that no longer exist on the web.

Or perhaps that never existed. It's well known that ChatGPT often hallucinates non-existing references (see for instance the discussion at https://news.ycombinator.com/item?id=33841672).

Apparently, there was a Java port for Windows NT running on the DEC Alpha: https://archive.org/details/jdk-v1-alpha_nt

It's unclear whether it's actually a 64-bit build. Did Windows even have a 64-bit userspace on Alpha, or was it all ILP32?

Anyway, it's not really possible to implement the Java memory model on Alpha: https://www.cs.umd.edu/~pugh/java/memoryModel/AlphaReorderin... So it's not really a natural target for Java code.

Wikipedia matches my recollection: Windows NT was actually ported to 64-bit using Alpha workstations, so there were probably some pre-release versions floating around. But by the time 64-bit Windows was ready, Alpha was a dying platform and the original release was only for the 64-bit platform with a real future, Itanium.

the 64-bit platform with a real future, Itanium

Ouch!

Yeah I also remember this. My Dad was at DEC/Compaq at the time and had access to the Windows NT prerelease. He was explaining to me why it was cool, but as an 8 year old I didn't really get it.

Right, but in the mid '90s when Java was designed, 64 bit machines were basically an exotic architecture and its practicalities weren't well understood, which is quite clear when CompressedOOPs is basically the sane default.

The weird inconsistency is apparent as you fairly frequently stub your toe on the 2 billion size limits of Java arrays, which is an area where 64 bit pointers would have made much more size than in object references.

> Right, but in the mid '90s when Java was designed, 64 bit machines were basically an exotic architecture

No they weren't, not in the enterprise market. At home? sure.

SUN Microsystems released the Sparc v9 architecture in 1993. SUN also made the Java language and JVM.

Wasn't java targeted at embedded? AFAIR at some point it was marketed that it will be in every washing machine, fridge.

There a was HotJava browser, weren't servers just a small part of its target?

"Write Once, Run Everywhere" was their slogan.

There's Java Card for smartcards yes. There was also Java Micro Edition (j2me) for phones.

Most binaries were still 32-bit for performance reason, and I don't think that Java 1.2 had a 64-bit port yet, not even on Solaris/SPARC.

The situation that an x86-64 build is faster than an i386 build of most applications (exception extremely pointer-heavy ones) is a bit of an exception because x86-64 added additional registers and uses a register-based calling convention everywhere. That happens to counteract the overhead of 64-bit pointers in most cases. Other 64-bit architectures with 32-bit userspace compatibility kept using 32-bit userspace for quite some time.

> 64 bit pointers

64 bit pointers was not "optimistic assumptions" about anything, it was just 64 bit pointers on 64 bits systems like most everyone else. And compressed oops were added more than a decade ago (https://wiki.openjdk.org/display/HotSpot/CompressedOops).

For reference that's about when x32 was added to the linux kernel, and unlike x32 compressed oops have not been on the chopping block for 5 years.

> enormous object headers.

Hardly? They're two words a class pointer, and a "mark words" for GC integration.

> And compressed oops were added more than a decade ago

That is still recent on a Java timescale.

> Hardly? They're two words a class pointer, and a "mark words" for GC integration.

Well compare with for example C++, that can fly commando with no header, just (optional) alignment padding.

In many Java objects, the header is more than half the size of the object. That's just not good data locality. The speed-up from switching from an array of objects that has n fields to an object which has an n arrays of fields can be very significant.

Your array of objects are just tightly packed pointers, so the data locality (or its lack) doesn’t come from that (especially that the order of objects in memory may be different than the order in the array).

I'm going on a limb as they're going in every direction and trebuchet-ing goalposts, but I'd assume they're talking about using an SoA structure, so instead of having a bunch of Foo objects each with a header and, say, an `int` field (so two words of header for half a word of payload) you have a single header at the top of the array, then an int value per record in the packed array.

64-bit longs and doubles take up two slots in the JVM's local variable table, whereas Object references only take one. So, if anything, the design of Java bytecode assumed 32-bit pointers.

They can just use 64-bit slots for everything (it does leave unused 32bit for each 32bit local variable, but there are not many, and you get the benefit of the same logic no matter what you store there, plus it has the same performance on 64-bit systems)

Edit: mind explaining the downvote ?

Not sure why you're being downvoted.

The Jacobin JVM [0] does exactly what you suggest: 64-bit operand stack and local variables. Longs and doubles still occupy two slots on the operand stack, so as to avoid having to recompile Java classes that assume the two-slot allocation, but the design avoids having to smash together two 32-bit values every time a long or double is operated on.

[0] http://jacobin.org/

Relatedly the JVM specification section 4.4.5 says (as the same holds for the cp and stack):

    In retrospect, making 8-byte constants take two constant pool entries was a poor choice.

I'm always impressed with the clarity of Java's design/implementation discussions, which I think is due to Mark Reinhold's leadership since 1997 (a breathtaking tenure in its own right).

But this needs to replace stack locking with an alternate lightweight locking scheme, to avoid races. Unfortunately, that is opaque:

   https://bugs.openjdk.org/browse/JDK-8291555

Does anyone have other pointers for the design or viability of the required alternative?

Their is a detailed description of the new stack locking scheme in the PR here https://github.com/openjdk/jdk/pull/10907

Is this part of the JVM Specification, or is this an implementation detail of OpenJDK itself? If the first, thats sort of surprising. Seems like an implementation detail that'd be best left up to the platform on how to manage objects in memory.

This is an implementation detail of the HotSpot JIT compiler in OpenJDK. Other implementations are free to layout their objects as they choose.

This document is a proposal for the HotSpot JVM, which is the default JVM implementation (the one that ships with OpenJDK), but not the only JVM implementation out there (see https://en.wikipedia.org/wiki/List_of_Java_virtual_machines for a full list).

The JEP identifies the scope as "implementation" and the component as "hotspot/runtime", so I do assume it's an implementation detail, yes.

No spec changes needed for this one.

Summary

Reduce the size of object headers in the HotSpot JVM from between 96 and 128 bits down to 64 bits on 64-bit architectures. This will reduce heap size, improve deployment density, and increase data locality.

As I (think I) understand the goal of this is to reduce memory footprint of the HotSpot JVM, with the tradeoff of performance degradation being capped to 5%, only in the worst rarest cases.

Just guessing that Cache throughput increase with the memory reduction may offset most performance issues.

I always wanted to have a tag in the object header that signals the GC to "just don't touch this object", so I could manually allocate off-heap objects and arrays (specially arrays).

ByteBuffers are low overhead, but native arrays are ~30% faster last I benchmarked, so if I can place an array off-heap, and never move it (granted, it's unsafe), I can use for example, the kernel's paging mechanism to swap data in and out data from disk, with low runtime overhead to access it.

5% latency overhead is huuuge. Throw more memory and focus on throughout. Not sure why memory is a concern here.

Different applications can have very different requirements. I've worked on systems which would kill for 5% latency and on systems that would gladly pay 5% for better memory usage.

"in infrequent cases" not in general.

Reply on news.ycombinator.com