Name: nt126004 Date: 04/29/2003


FULL PRODUCT VERSION :
not version-related

FULL OPERATING SYSTEM VERSION :not OS-related


A DESCRIPTION OF THE PROBLEM :
In testing I found that using LITTLE_ENDIAN arrays showed no
improvement on an Intel machine. Copying 1.6M ints was no
faster with the native order (270 to 280 millis) than with
the BIG_ENDIAN order.

If the copy were done with a standard C library copy it
would take advantage of the fact that copying an array is a
bit of setup and then a single bulk copy instruction. On the
386 (last time I checked) that copy moved one word per
cycle. Assuming no improvement in that area, copying 1.6M
words should take 4 millis, not 270 or 280.

Ask the programmer who did System.arraycopy(). He got it right.

STEPS TO FOLLOW TO REPRODUCE THE PROBLEM :
1.test code posted below
2.change constant to BIG_ENDIAN and retest
3.

EXPECTED VERSUS ACTUAL BEHAVIOR :
BIG_ENDIAN order requires a byte-by-byte flip on a
LITTLE_ENDIAN box. LITTLE_ENDIAN only needs copying an area
of RAM. Expected 2 orders of magnitude speed increase. Got zero.

REPRODUCIBILITY :
This bug can be reproduced always.

---------- BEGIN SOURCE ----------
public static void main( String[] args ) {

        public static final int K = 1024;

        byte[] buf = new byte[ 6400 * K ];
        int ilen = buf.length / 4;

        ByteBuffer bb = ByteBuffer.wrap( buf );
        bb.order( ByteOrder.LITTLE_ENDIAN );

        IntBuffer ib = bb.asIntBuffer();

        int[] ints = new int[ ilen ];
        
        for ( int i = 0; i < ilen; i++ ) ints[i] = i;

        long start = System.currentTimeMillis();

            ib.put( ints );

        long stop = System.currentTimeMillis();

        System.out.println( "time=" + (stop - start) );
 
System.exit( 0 );

} // end of main()
---------- END SOURCE ----------
(Review ID: 181086)
======================================================================