在 Java 中使用 final for 变量是否改进了垃圾收集？

小开

Well, I don't know about the use of the "final" modifier in this case, or its effect on the GC.

But I can tell you this: your use of Boxed values rather than primitives (e.g., Double instead of double) will allocate those objects on the heap rather than the stack, and will produce unnecessary garbage that the GC will have to clean up.

I only use boxed primitives when required by an existing API, or when I need nullable primatives.

小开

The only thing that I can think of is that the compiler might optimize away the final variables and inline them as constants into the code, thus you end up with no memory allocated.

小开

No, it is emphatically not true.

Remember that final does not mean constant, it just means you can't change the reference.

final MyObject o = new MyObject();
o.setValue("foo"); // Works just fine
o = new MyObject(); // Doesn't work.

There may be some small optimisation based around the knowledge that the JVM will never have to modify the reference (such as not having check to see if it has changed) but it would be so minor as to not worry about.

Final should be thought of as useful meta-data to the developer and not as a compiler optimisation.

小开

Declaring a local variable final will not affect garbage collection, it only means you can not modify the variable. Your example above should not compile as you are modifying the variable totalWeight which has been marked final. On the other hand, declaring a primitive (double instead of Double) final will allows that variable to be inlined into the calling code, so that could cause some memory and performance improvement. This is used when you have a number of public static final Strings in a class.

In general, the compiler and runtime will optimize where it can. It is best to write the code appropriately and not try to be too tricky. Use final when you do not want the variable to be modified. Assume that any easy optimizations will be performed by the compiler, and if you are worried about performance or memory use, use a profiler to determine the real problem.

小开

最佳答案

Here's a slightly different example, one with final reference-type fields rather than final value-type local variables:

public class MyClass {


public final MyOtherObject obj;


}

Every time you create an instance of MyClass, you'll be creating an outgoing reference to a MyOtherObject instance, and the GC will have to follow that link to look for live objects.

The JVM uses a mark-sweep GC algorithm, which has to examine all the live refereces in the GC "root" locations (like all the objects in the current call stack). Each live object is "marked" as being alive, and any object referred to by a live object is also marked as being alive.

After the completion of the mark phase, the GC sweeps through the heap, freeing memory for all unmarked objects (and compacting the memory for the remaining live objects).

Also, it's important to recognize that the Java heap memory is partitioned into a "young generation" and an "old generation". All objects are initially allocated in the young generation (sometimes referred to as "the nursery"). Since most objects are short-lived, the GC is more aggressive about freeing recent garbage from the young generation. If an object survives a collection cycle of the young generation, it gets moved into the old generation (sometimes referred to as the "tenured generation"), which is processed less frequently.

So, off the top of my head, I'm going to say "no, the 'final' modifer doesn't help the GC reduce its workload".

In my opinion, the best strategy for optimizing your memory-management in Java is to eliminate spurious references as quickly as possible. You could do that by assigning "null" to an object reference as soon as you're done using it.

Or, better yet, minimize the size of each declaration scope. For example, if you declare an object at the beginning of a 1000-line method, and if the object stays alive until the close of that method's scope (the last closing curly brace), then the object might stay alive for much longer that actually necessary.

If you use small methods, with only a dozen or so lines of code, then the objects declared within that method will fall out of scope more quickly, and the GC will be able to do most of its work within the much-more-efficient young generation. You don't want objects being moved into the older generation unless absolutely necessary.

小开

GC acts on unreachable refs. This has nothing to do with "final", which is merely an assertion of one-time assignment. Is it possible that some VM's GC can make use of "final"? I don't see how or why.

小开

There is a not so well known corner case with generational garbage collectors. (For a brief description read the answer by benjismith for a deeper insight read the articles at the end).

The idea in generational GCs is that most of the time only young generations need to be considered. The root location is scanned for references, and then the young generation objects are scanned. During this more frequent sweeps no object in the old generation are checked.

Now, the problem comes from the fact that an object is not allowed to have references to younger objects. When a long lived (old generation) object gets a reference to a new object, that reference must be explicitly tracked by the garbage collector (see article from IBM on the hotspot JVM collector), actually affecting the GC performance.

The reason why an old object cannot refer to a younger one is that, as the old object is not checked in minor collections, if the only reference to the object is kept in the old object, it will not get marked, and would be wrongly deallocated during the sweep stage.

Of course, as pointed by many, the final keyword does not reallly affect the garbage collector, but it does guarantee that the reference will never be changed into a younger object if this object survives the minor collections and makes it to the older heap.

Articles:

IBM on garbage collection: history, in the hotspot JVM and performance. These may no longer be fully valid, as it dates back in 2003/04, but they give some easy to read insight into GCs.

Sun on Tuning garbage collection

小开

final on local variables and parameters makes no difference to the class files produced, so cannot affect runtime performance. If a class has no subclasses, HotSpot treats that class as if it is final anyway (it can undo later if a class that breaks that assumption is loaded). I believe final on methods is much the same as classes. final on static field may allow the variable to be interpreted as a "compile-time constant" and optimisation to be done by javac on that basis. final on fields allows the JVM some freedom to ignore happens-before relations.

小开

Final variables cannot be changed after initial assignment (enforced by the compiler).

This does not change the behaviour of the garbage collection as such. Only thing is that these variables cannot be nulled when not being used any more (which may help the garbage collection in memory tight situations).

You should know that final allows the compiler to make assumptions about what to optimize. Inlining code and not including code known not to be reachable.

final boolean debug = false;


......


if (debug) {
System.out.println("DEBUG INFO!");
}

The println will not be included in the byte code.

小开

Some points to clear up:

Nulling out reference should not help GC. If it did, it would indicate that your variables are over scoped. One exception is the case of object nepotism.
There is no on-stack allocation as of yet in Java.
Declaring a variable final means you can't (under normal conditions) assign a new value to that variable. Since final says nothing about scope, it doesn't say anything about it's effect on GC.

小开

All method and variable can be overridden bydefault in subclasses.If we want to save the subclasses from overridig the members of superclass,we can declare them as final using the keyword final. For e.g- final int a=10; final void display(){......} Making a method final ensures that the functionality defined in the superclass will never be changed anyway. Similarly the value of a final variable can never be changed. Final variables behaves like class variables.

小开

absolutely, as long as make object's life shorter which yield great benefit of memory management, recently we examined export functionality having instance variables on one test and another test having method level local variable. during load testing, JVM throws outofmemoryerror on first test and JVM got halted. but in second test, successfully able to get the report due to better memory management.

小开

There seems to be a lot of answers that are wandering conjectures. The truth is, there is no final modifier for local variables at the bytecode level. The virtual machine will never know that your local variables were defined as final or not.

The answer to your question is an emphatic no.

小开

The only time I prefer declaring local variables as final is when:

I have to make them final so that they can be shared with some anonymous class (for example: creating daemon thread and let it access some value from enclosing method)
I want to make them final (for example: some value that shouldn't/doesn't get overridden by mistake)

Does they help in fast garbage collection?
AFAIK a object becomes a candidate of GC collection if it has zero strong references to it and in that case as well there is no guarantee that they will be immediately garbage collected . In general, a strong reference is said to die when it goes out of scope or user explicitly reassign it to null reference, thus, declaring them final means that reference will continue to exists till the method exists (unless its scope is explicitly narrowed down to a specific inner block {}) because you can't reassign final variables (i.e. can't reassign to null). So I think w.r.t Garbage Collection 'final' may introduce a unwanted possible delay so one must be little careful in defining there scope as that controls when they will become candidate for GC.

小开

Strictly speaking about instance fields, final might improve performance slightly if a particular GC wants to exploit that. When a concurrent GC happens (that means that your application is still running, while GC is in progress), see this for a broader explanation, GCs have to employ certain barriers when writes and/or reads are done. The link I gave you pretty much explains that, but to make it really short: when a GC does some concurrent work, all read and writes to the heap (while that GC is in progress), are "intercepted" and applied later in time; so that the concurrent GC phase can finish it's work.

For final instance fields, since they can not be modified (unless reflection), these barriers can be omitted. And this is not just pure theory.

Shenandoah GC has them in practice (though not for long), and you can do, for example:

-XX:+UnlockExperimentalVMOptions
-XX:+UseShenandoahGC
-XX:+ShenandoahOptimizeInstanceFinals

And there will be optimizations in the GC algorithm that will make it slightly faster. This is because there will be no barriers intercepting final, since no one should modify them, ever. Not even via reflection or JNI.