(array) Reduce JNI overhead for Arrays.equals() with float/double arrays

A DESCRIPTION OF THE REQUEST :
Comparing large float or double arrays with Arrays.equals(float[], float[]) and Arrays.equals(float[], float[]) is dominated at runtime by CPU time spent in marchalling/unmarshalling the JNI interface to the native implementation of Float and Double methods to get canonical bits.

This RFE is an optimization to remove most calls to JNI implementations of static methods in java.lang.Float and java.lang.Double.

I propose below an alternative implementation that avoids this overhead.


JUSTIFICATION :
However, such use of
* public static native int Float.floatToIntBits(float value);
* public static native int Double.doubleToLongBits(double value);
is definitely not needed for comparing most float or double values in Arrays.equals(...) methods when it can be performed directly using normal Java == operators that generate normal bytecode that can be compiled directly into native CPU instructions by JIT, without such JNI calls and marshalling overhead.

My performance tests show significant improvement or example when comparing matrixes (for example in 2D/3D graphic rendering or 3D apps when matching vertices and points, or computing and matching colors in multiplane color spaces for precise transformation of bitmap images).

Comparing floatting point arrays for equality is not very common, but it happens in such contexts, and then, the methods are used at a very high frequency rate and do need better optimization.

The need of getting and comparing bits is justified ONLY when comparing zeroes (because these methods make a distinction between positive and negative zeroes), or because it wants to make NaNs comparable. In many applications however, NaNs are not used (or seen as wrong anyway), or it does not matter if they are different from each other.

The other use of Arrays.equal() with floatting point arrays is when hashing them for use as keys in HashMaps or Sets: it is used to detect key collisions in the Set or Map when storing them, and Arrays.equal become much more frequent when the Set or Map is near the threshold, and used at high rate when reindexing the Set or Map when it is resized after reaching the threshold: with the 75% load factor and random distribution (not guaranteed by arrays of native types because their hashes are quite simple), there's nearly 50% of collisions, when resizing the Set or Map, about 25% of elements in the new set will have collisions and will need to be compared with Arrays.equal, creating heavy use of the JNI interface, if the Set or Map is already large. And anyway, each time we get any element from the Set or HashMap, we need to compare the key (and in this case, most of the time, they will be equal, so if a key is an floatting point array, all elements in the floatting point array are compared in the loop; if the floatting point array size is large, this means a significant cost spent in comparing them).

The code below optimizes this situation for much better performance after JIT compilation, and negligeable cost when the code is interpreted.


---------- BEGIN SOURCE ----------
// package java.lang;
// public class Arrays {
public class Test {

    public static boolean equals(final float[] a1, final float[] a2) {
        if (a1 != a2) {
            int index;
            if (a1 != null && a2 != null && (index = a1.length) == a2.length) {
                while (--index >= 0) {
                    final float f1, f2;
                    if ((f1 = a1[index]) != (f2 = a2[index])
                            // (either f1 or f2 may be a NaN)
                            && f1 == f1 // ! Float.isNaN(f1)
                            && f2 == f2 // ! Float.isNaN(f2)
                            // (we don't need bits for NaNs)
                            // Otherwise f1 == f2 but may be both zeroes
                            || f1 == 0.0f // f1 is 0.0f or -0.0f
                            && f2 == 0.0f // f2 is 0.0f or -0.0f
                            // (test if distinct zeroes)
                            && Float.floatToIntBits(f1) != Float
                                    .floatToIntBits(f2))
                        return false;
                }
                return true;
            }
            return false;
        }
        return true;
    }

    public static boolean equals(final double[] a1, final double[] a2) {
        if (a1 != a2) {
            java.util.HashMap h;
            int index;
            if (a1 != null && a2 != null && (index = a1.length) == a2.length) {
                while (--index >= 0) {
                    final double d1, d2;
                    if ((d1 = a1[index]) != (d2 = a2[index])
                            // (either d1 or d2 may be a NaN)
                            && d1 == d1 // ! Double.isNaN(d1)
                            && d2 == d2 // ! Double.isNaN(d2)
                            // (we don't need bits for NaNs)
                            // Otherwise d1 == d2 but may be both zeroes
                            || d1 == 0.0d // d1 is 0.0d or -0.0d
                            && d2 == 0.0d // d2 is 0.0d or -0.0d
                            // (test if distinct zeroes)
                            && Double.doubleToLongBits(d1) != Double
                                    .doubleToLongBits(d2))
                        return false;
                }
                return true;
            }
            return false;
        }
        return true;
    }

}

---------- END SOURCE ----------