EVALUATION

http://hg.openjdk.java.net/jdk7/hotspot-comp/hotspot/rev/c771b7f43bbf
                                     

PUBLIC COMMENTS

And the numbers for AMD Shanghai:

$ gamma -XX:-UsePopCountInstruction test
sum: 629085184
time: 8504

$ gamma -XX:+UsePopCountInstruction test
sum: 629085184
time: 1807

4.7x speedup.

$ gamma -XX:-UsePopCountInstruction -XX:LoopUnrollLimit=1 test
sum: 629085184
time: 9622

$ gamma -XX:+UsePopCountInstruction -XX:LoopUnrollLimit=1 test
sum: 629085184
time: 2577

3.73x speedup.
                                     

PUBLIC COMMENTS

Just for the record to see how much slower the kernel-level trap-based emulation on SPARC is (with 20 * 1000000 loops):

$ gamma -XX:-UsePopCountInstruction test
VM option '-UsePopCountInstruction'
sum: 238869248
time: 1011

$ gamma -XX:+UsePopCountInstruction test
VM option '+UsePopCountInstruction'
sum: 238869248
time: 76985
                                     

EVALUATION

The same numbers on a T2:

$ java -XX:-UsePopCountInstruction test
VM option '-UsePopCountInstruction'
sum: 629085184
time: 35676

$ java -XX:+UsePopCountInstruction test
VM option '+UsePopCountInstruction'
sum: 629085184
time: 20007

And without loop unrolling:

$ java -XX:-UsePopCountInstruction -XX:LoopUnrollLimit=1 test
VM option '-UsePopCountInstruction'
VM option 'LoopUnrollLimit=1'
sum: 629085184
time: 41509

$ java -XX:+UsePopCountInstruction -XX:LoopUnrollLimit=1 test
VM option '+UsePopCountInstruction'
VM option 'LoopUnrollLimit=1'
sum: 629085184
time: 29470

The speedup is 1.78 and 1.41 respectively.
                                     

EVALUATION

A very simple micro-benchmark like this:

public class test {
    public static void main(String[] args) {
        int sum = 0;
        long start = System.currentTimeMillis();
        for (int i = 0; i < 2000 * 1000000; i++) {
            sum += Integer.bitCount(i);
        }
        long end = System.currentTimeMillis();
        System.out.println("sum: " + sum);
        System.out.println("time: " + (end - start));
    }
}

shows a 5x speedup on a Nehalem processor:

$ gamma -XX:-UsePopCountInstruction test
VM option '-UsePopCountInstruction'
sum: 629085184
time: 8132

$ gamma -XX:+UsePopCountInstruction test
VM option '+UsePopCountInstruction'
sum: 629085184
time: 1604

And with disabled loop unrolling to get more accurate numbers:

$ gamma -XX:-UsePopCountInstruction -XX:LoopUnrollLimit=1 test
VM option '-UsePopCountInstruction'
VM option 'LoopUnrollLimit=1'
sum: 629085184
time: 8657

$ gamma -XX:+UsePopCountInstruction -XX:LoopUnrollLimit=1 test
VM option '+UsePopCountInstruction'
VM option 'LoopUnrollLimit=1'
sum: 629085184
time: 1458

It's interesting to see that a tighter loop with popcnt is faster.