Performance impact of autoboxing

deamon picture deamon · Aug 7, 2010 · Viewed 19.3k times · Source

Usually the compiler generates code to perform boxing and unboxing. But what does the compiler, if the boxed values are not needed? Is the (Oracle standard) compiler smart enough to optimize it away?

Take a look at this method:

public static void requireInRange(int index, Object[] array) {
    if(index < 0 || index >= array.length)
        throw new IndexOutOfBoundsException();
}

The only relevant information is the array.length, so it would be useless to box each value of an array for example. Like in this code:

int[] anArray = {3, 4, 2};
requireInRange(3, anArray);

Will the compiler actually insert code for boxing each value of the array?

Answer

polygenelubricants picture polygenelubricants · Aug 7, 2010

There is no autoboxing in your code. In fact, given:

public static void requireInRange(int index, Object[] array) {
   ...
}

int[] anArray = {3, 4, 2};
requireInRange(3, anArray); // DOES NOT COMPILE!!!

While an int can be autoboxed to an Integer, an int[] does NOT get autoboxed to Integer[] by Java. You can write library functions to do this, but the language will not facilitate this conversion.

This is in fact the source of many confusion regarding e.g. Arrays.asList(anIntArray) being "broken", because instead of returning a List<Integer>, what is returned is in fact a one-element List<int[]>.


But what about performance???

A quote from Java Language Guide/Autoboxing:

It is not appropriate to use autoboxing and unboxing for scientific computing, or other performance-sensitive numerical code. An Integer is not a substitute for an int; autoboxing and unboxing blur the distinction between primitive types and reference types, but they do not eliminate it.

In short, whenever autoboxing happens, performance definitely takes a little bit of a hit. Certain things help to alleviate this, e.g. the caching mechanism built into these types. This is why you get the following:

    System.out.println(
        ((Integer) 0) == ((Integer) 0)
    );
    // true

    System.out.println(
        ((Integer) 10000) == ((Integer) 10000)
    );
    // false (implementation-specific)

What happened here is that when 0 is automatically boxed, no new Integer instance is actually created: values in certain range is cached for autoboxing purposes, to help performance. 10000 in most implementation probably falls out of this range, but some JVM implementations do allow you to specify the cache range if necessary.


But I just want to get the length of the array!!!

There are many ways to facilitate your requireInRange to work with any type of arrays. Unfortunately working with Java's array of primitives often times mean lots of repetition. This mean providing overloads for int[], boolean[], byte[], Object[], etc separately.

A more concise option is to use reflection, but this has its pluses and minuses. Generally speaking, reflection should not be the preferred solution for most scenarios.

Having said that, java.lang.reflect.Array does have a int getLength(Object array) static method that can return the length of ANY array. It's not typesafe (like most reflection mechanism are); passing a non-array compiles, but throws IllegalArgumentException at run-time.

Related questions