Request for Reviews (S): JDK-8003585 strength reduce or eliminate range checks for power-of-two sized arrays

[Paul Sandoz]

Hi Nils,

Any update on this?

I have been looking at this area in the context of array access for JEP 193: Enhanced Volatiles (i will send another email on that later today). But I was wanted to share some initial perf results with this patch.

Here are the relevant snippets of benchmark code:

    private static final int N = 1 << Integer.getInteger("N", 8);
    private static final int START = (N >> 1);
    private static final int END = (N >> 1) + N;
...

    private Value value;

    private Value[] receiver;

    @Setup
    public void setUp() throws Exception {
        value = new ExtendedValue(1);
        receiver = new ExtendedValue[N];
        for (int i = 0; i < receiver.length; i++) {
            receiver[i] = value;
        }
    }
...
    @GenerateMicroBenchmark
    public int relaxed_r_aa() {
        Value[] _receiver = receiver;
        int sum = 0;
        for (int i = START; i < END; i++) {
            int j = i & (_receiver.length - 1);
            sum += _receiver[j].i;
        }
        return sum;
    }

    @GenerateMicroBenchmark
    public Value[] relaxed_w_aa() {
        Value[] _receiver = receiver;
        for (int i = START; i < END; i++) {
            int j = i & (_receiver.length - 1);
            _receiver[j] = value;
        }
        return _receiver;
    }

    @GenerateMicroBenchmark
    public int relaxed_r_unsafe() {
        Value[] _receiver = receiver;
        int sum = 0;
        for (int i = START; i < END; i++) {
            int j = i & (_receiver.length - 1);
            Value v = (Value) UNSAFE.getObject(
                    _receiver,
                    (((long) j) << UNSAFE_ARRAY_SHIFT_V) + UNSAFE_ARRAY_OFFSET_V);
            sum += v.i;
        }
        return sum;
    }

    @GenerateMicroBenchmark
    public Value[] relaxed_w_unsafe() {
        Value[] _receiver = receiver;
        for (int i = START; i < END; i++) {
            int j = i & (_receiver.length - 1);
            UNSAFE.putObject(_receiver,
                             (((long) j) << UNSAFE_ARRAY_SHIFT_V) + UNSAFE_ARRAY_OFFSET_V,
                             value);
        }
        return _receiver;
    }

Here are some results on OpenJDK 9 without the patch (times are in ns/op for N=1,8,64,512, for one thread, measured on a Dell laptop with an Intel Core i5-2520 CPU @ 2.50GHz):

    Benchmark                             1        8        64        512
    relaxed_r_aa                        1.304    9.128    48.599    372.214
    relaxed_r_unsafe                    1.152    6.615    44.456    348.989

    relaxed_w_aa                        3.256   16.105    76.896    599.830
    relaxed_w_unsafe                    2.888   11.553    56.037    532.179

Here are the results with the patch applied to hotspot (in the dev repo):

    Benchmark                             1        8        64        512
    relaxed_r_aa                        1.291    6.664    37.604    285.830
    relaxed_r_unsafe                    1.151    6.621    44.483    348.249

    relaxed_w_aa                        3.279   12.646    66.324    481.281
    relaxed_w_unsafe                    2.888   11.549    56.028    531.589


One has to be a little careful interpreting the results since compilation of the JMH loop measurement method can monkey around with 'em (I think from OSR'ed to complete, which might be affecting how loops are unrolled, but i would need to dive into the machine code of the compiled methods to verify). However, even so this patch looks promising.

Paul.

On Feb 17, 2014, at 2:07 PM, Nils Eliasson <nils.eliasson at oracle.com> wrote:

> Current change passes JPRT. Running bigapps and nightly code policy as well.
> 
> Will wait for tests and reviewer go-ahead before pushing.

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.openjdk.java.net/pipermail/hotspot-compiler-dev/attachments/20140505/4d55d395/attachment.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 841 bytes
Desc: Message signed with OpenPGP using GPGMail
URL: <http://mail.openjdk.java.net/pipermail/hotspot-compiler-dev/attachments/20140505/4d55d395/signature.asc>