[jallmann]

Didn't want to say it outright, but that's my theory too. Having a VM manage everything in a resource/power constrained environment is a crazy idea in the first place. Oracle's JVM is competitive because of heroic engineering, in spite of Java's design -- not because of it. And it still has tradeoffs, like insane memory usage. ART/Dalvik are operating under different constraints, which probably contributes significantly to Android's handicap.

That can't be the whole story though, because C# and VB.NET both (seem to) perform decently under a managed runtime on Windows Phone. Wonder how big of a role the CLR has in typical WP apps and the WP core, as opposed to unmanaged C/C++ code.

[pjmlp]

.NET is AOT compiled to native code since Windows Phone 8.

On Windows Phone 8.x, it uses MDIL (Machine Dependent Intermediate Language) meaning native code with symbolic names for the on-device linker.

On Windows Phone 10 onwards, it makes use of .NET Native.

Both are based on Visual C++'s backend, which is way more world battle tested than ART.

.NET also supports value types.

Also the XAML layouts are compiled to binay, not interpreted on load like on Android (aka inflated).

Windows Phone also only supports asynchronous code, graphics and sound APIs must be hardware accelerated.

[iofj]

The Go authors had a pretty good article on what's wrong with Java's performance : pointers everywhere. Every last little thing that isn't a primitive type is a pointer. Everywhere, in every bit of code.

That means a "new Object()" takes up 16 bytes (8 bytes for the object, 8 for the pointer to it). That means you fill a cache line by allocating 4 objects, or 2 objects containing a single reference, or ...

So in java you should never program a line drawing loop by using 2 vectors, because 2 vectors, each with 2 32-bit ints take up 82 (2 pointers to the objects you're using) + 82 (overhead for the objects) + 4*2 (the actual data) 40 bytes of data. No way you can fit that in registers and still use registers to actually calculate things. So instead you should use 4 ints and just forget about the objects, and even that will only work if you never call any functions.

Same loop in C/C++/Pascal/Go/... using structs takes 8 bytes (they don't keep structs on the heap), which, if necessary, fits in 1 register (granted, in practice we're talking 2 registers, but still).

People might reply to this with benchmarks, but if you actually analyse the java code where java beats or is comparable with C/C++ you're going to see zero object allocations. You're not even going to see them using bool in the extreme cases, rather they'll bitshift into ints to effectively generate packed bools (certainly in SAT benchmarks). This is not realistic java code, which would have been way slower.

Java's memory model is the main culprit at this point in time. Java can do incredible tricks with programs, and actually exposes them, enabling lots of language creativity on the JVM. But there's a pretty sizeable cost in speed and memory usage.

[userbinator]

People might reply to this with benchmarks, but if you actually analyse the java code where java beats or is comparable with C/C++ you're going to see zero object allocations.

I've noticed that tends to be true in general for benchmarks of high-level languages which show them performing as well as or sometimes even better than C/C++ --- the code performs so well because it's essentially using none of the other language features that most code in the language does. I touch upon this in my other comment here about culture: the language theoretically allows you to write quite efficient code, but it doesn't look "idiomatic" or perhaps isn't a "best practice", so it's discouraged and isn't done. The entire dogma of avoiding any optimisation compounds this problem even more, since once programmers finally realise they have performance issues, they've already created such complex and inefficient code that it's even harder to do any optimisation on.

On the other hand, idiomatic C tends to be written in a simple and straightforward style that is naturally quite efficient already. C++ is similar, although templates, OOP, and all the other new features can lead to inefficient code if not used in moderation.

I suppose the ultimate example of what could be called "intrinsically efficient" is assembly language. With Asm, every instruction, every byte you can save from typing is one the machine also doesn't have to execute, so you're basically forced to optimise as you write. There's certainly no desire to overengineer things, simply because of the extreme tedium and futility of doing so. With no IDE to help you generate classes and autocomplete indirections, it really changes your perspective of what constitutes efficient code.

[pjmlp]

Even on Assembly one might suffer from macro opcodes and the respective microcode expansion, for the sets that don't map 1:1 to what the CPU does.

[pjmlp]

> "new Object()" takes up 16 bytes ....

Only in JVMs and AOT compilers that don't do escape analysis.

Also don't forget Smalltalk, which also only does references, was running in the Alto, Dolphin, and Dorado workstations.

For example the Dorado was:

- 128-512 kB

- 606x808 pixels

- 4 74181 CPUs

So how does that compare to a beefy Android device?

Also J2ME and Embedded Java are running quite well in many embedded platforms, in a few hundred KB steering soft real time systems like robots and missile radar controls.

So yes, Java might not offer all the memory control features that other GC enabled languages do, going back to Algol 68, Mesa/Cedar, Eiffel, Modula-3, ....

But given the performance of commercial JVM vendors, I would say Google has a lot of blame as well.

EDIT: Forgot to add that when Java 10 comes out with value types and reified generics (according to the roadmap) this will become a moot point, except of course for Android Java given Google's unwillingness to provide support for the real thing.

[iofj]

Maybe I'm unclear on some of the details here, but here goes

https://docs.oracle.com/javase/specs/jls/se8/html/jls-12.htm...

"Whenever a new class instance is created, memory space is allocated for it with room for all the instance variables declared in the class type and all the instance variables declared in each superclass of the class type, including all the instance variables that may be hidden (§8.3)."

Doesn't seem to allow for escape analysis eliminating the object. Plus escape analysis wouldn't really save you. These are class instances, you pretty much have to declare them before the scope you use them in, if you're using them in the condition of a while loop (which would be the way to use them).

I seem to have this experience in practice. If you have a value type and loop over it, creating a "dummy" instance of it outside of the loop, then erase and reset it's inner state on every loop iteration is far faster than creating an instance inside the loop. So I don't think escape analysis optimizes this case.

[pjmlp]

Because JIT compilers are able to bend the rules if proven correct.

For example, if you declare something like

    final class Point {
          final int x, y;

          //...
    }

The Azul JIT compiler will transform it into a struct, just like in C, via their "StructuredArray and ObjectLayout" optimizations.

IBM J9 also does similar optimizations via packed objects, as they call it.

JIT compilers also remove locks and synchronized blocks if heuristics prove their are never needed in the dataflow.

In any case, by Java 10, real value types are expected to be part of the language.

[iofj]

Even if you're right. The issue with coding to optimizations is that it's really, really brittle. You change the position of a variable and suddenly your application runs 10x slower.

Why ? Because it just went from O(N) with no allocations to O(N^2) because it has to constantly extend and walk a list (the free memory list, which is lower bounded by the number of iterations of the loop) on every iteration of the loop (malloc is O(N), so calling malloc in a loop automatically increases it's complexity, and compilers change stack allocations into mallocs due to optimizations).

I've seen this happen many times. And then, after 2 weeks of searching you find the cause : someone changed 3 * i into i * 3 in a method which caused autoboxing to suddenly actually occur.

Counting on compiler optimizations to save your ass is incredibly, incredibly brittle.

[dilap]

c# has value types right?