RFR: 8333893: Optimization for StringBuilder append boolean & null

[Shaojin Wen]

On Mon, 10 Jun 2024 12:12:58 GMT, Shaojin Wen <duke at openjdk.org> wrote:

> After PR https://github.com/openjdk/jdk/pull/16245, C2 optimizes stores into primitive arrays by combining values ​​into larger stores.
> 
> This PR rewrites the code of appendNull and append(boolean) methods so that these two methods can be optimized by C2.

# 1.  Compare with the master branch
1. master (`a6fc2f8`) https://github.com/wenshao/jdk/tree/upstream_master_a6fc2f8
2. current (`5e815`) https://github.com/wenshao/jdk/tree/optim_str_builder_append_202406

# 2. Benchmark Commands

make test TEST="micro:java.lang.StringBuilders.toStringCharWithBool8"
make test TEST="micro:java.lang.StringBuilders.toStringCharWithNull8"


# 3. Benchmark Numbers
The performance numbers under MacBookPro M1 Pro are as follows:


-Benchmark                                   Mode  Cnt  Score   Error  Units #master (a6fc2f8)
-StringBuilders.toStringCharWithBool8Latin1  avgt   15  7.371 ? 0.003  ns/op
-StringBuilders.toStringCharWithBool8Utf16   avgt   15  9.613 ? 0.018  ns/op
-StringBuilders.toStringCharWithNull8Latin1  avgt   15  7.071 ? 0.003  ns/op
-StringBuilders.toStringCharWithNull8Utf16   avgt   15  9.296 ? 0.016  ns/op

+Benchmark                                   Mode  Cnt  Score   Error  Units #current (5e815)
+StringBuilders.toStringCharWithBool8Latin1  avgt   15  6.515 ? 0.121  ns/op +11.61%
+StringBuilders.toStringCharWithBool8Utf16   avgt   15  8.654 ? 0.035  ns/op +9.97%
+StringBuilders.toStringCharWithNull8Latin1  avgt   15  5.550 ? 0.010  ns/op +21.51%
+StringBuilders.toStringCharWithNull8Utf16   avgt   15  8.108 ? 0.041  ns/op +12.77%

# 1. Compare with unsafe branch

1. current (`5e815`) https://github.com/wenshao/jdk/tree/optim_str_builder_append_202406
2. unsafe (`adc220`) https://github.com/wenshao/jdk/tree/optim_str_builder_append_202406_unsafe

I think the performance of the Unsafe branch may be the best data for the C2 optimizer. @eme64 can help me see if C2 can do it?

# 2. Benchmark Commands

make test TEST="micro:java.lang.StringBuilders.toStringCharWithBool8"
make test TEST="micro:java.lang.StringBuilders.toStringCharWithNull8"


# 3. Implementation of Unsafe Branch

class AbstractStringBuilder {
    static final Unsafe UNSAFE = Unsafe.getUnsafe();

    static final int NULL_LATIN1;
    static final int TRUE_LATIN1;
    static final int FALS_LATIN1;

    static final long NULL_UTF16;
    static final long TRUE_UTF16;
    static final long FALS_UTF16;

    static {
        byte[] bytes4 = new byte[4];
        byte[] bytes8 = new byte[8];

        bytes4[0] = 'n';
        bytes4[1] = 'u';
        bytes4[2] = 'l';
        bytes4[3] = 'l';
        NULL_LATIN1 = UNSAFE.getInt(bytes4, Unsafe.ARRAY_BYTE_BASE_OFFSET);
        StringUTF16.inflate(bytes4, 0, bytes8, 0, 4);
        NULL_UTF16 = UNSAFE.getLong(bytes8, Unsafe.ARRAY_BYTE_BASE_OFFSET);

        bytes4[0] = 't';
        bytes4[1] = 'r';
        bytes4[2] = 'u';
        bytes4[3] = 'e';
        TRUE_LATIN1 = UNSAFE.getInt(bytes4, Unsafe.ARRAY_BYTE_BASE_OFFSET);
        StringUTF16.inflate(bytes4, 0, bytes8, 0, 4);
        TRUE_UTF16 = UNSAFE.getLong(bytes8, Unsafe.ARRAY_BYTE_BASE_OFFSET);

        bytes4[0] = 'f';
        bytes4[1] = 'a';
        bytes4[2] = 'l';
        bytes4[3] = 's';
        FALS_LATIN1 = UNSAFE.getInt(bytes4, Unsafe.ARRAY_BYTE_BASE_OFFSET);
        StringUTF16.inflate(bytes4, 0, bytes8, 0, 4);
        FALS_UTF16 = UNSAFE.getLong(bytes8, Unsafe.ARRAY_BYTE_BASE_OFFSET);
    }

    private AbstractStringBuilder appendNull() {
        ensureCapacityInternal(count + 4);
        int count = this.count;
        byte[] val = this.value;
        if (isLatin1()) {
            UNSAFE.putInt(
                    val,
                    Unsafe.ARRAY_BYTE_BASE_OFFSET + count,
                    NULL_LATIN1);
        } else {
            UNSAFE.putLong(
                    val,
                    Unsafe.ARRAY_BYTE_BASE_OFFSET + (count << 1),
                    NULL_UTF16);
        }
        this.count = count + 4;
        return this;
    }

    public AbstractStringBuilder append(boolean b) {
        int count = this.count;
        int spaceNeeded = count + (b ? 4 : 5);
        ensureCapacityInternal(spaceNeeded);
        byte[] val = this.value;
        if (isLatin1()) {
            UNSAFE.putInt(
                    val,
                    Unsafe.ARRAY_BYTE_BASE_OFFSET + count,
                    b ? TRUE_LATIN1 : FALS_LATIN1);
            if (!b) {
                val[count + 4] = 'e';
            }
        } else {
            UNSAFE.putLong(
                    val,
                    Unsafe.ARRAY_BYTE_BASE_OFFSET + (count << 1),
                    b ? TRUE_UTF16 : FALS_UTF16);
            if (!b) {
                StringUTF16.putChar(val, count + 4, 'e');
            }
        }
        this.count = spaceNeeded;
        return this;
    }
}


# 4. Benchmark Numbers
The performance numbers under MacBookPro M1 Pro are as follows:


-Benchmark                                   Mode  Cnt  Score   Error  Units # unsafe (adc220)
-StringBuilders.toStringCharWithBool8Latin1  avgt   15  6.415 ? 0.061  ns/op
-StringBuilders.toStringCharWithBool8Utf16   avgt   15  7.307 ? 0.013  ns/op
-StringBuilders.toStringCharWithNull8Latin1  avgt   15  5.443 ? 0.011  ns/op
-StringBuilders.toStringCharWithNull8Utf16   avgt   15  6.944 ? 0.102  ns/op

+Benchmark                                   Mode  Cnt  Score   Error  Units #current (5e815)
+StringBuilders.toStringCharWithBool8Latin1  avgt   15  6.515 ? 0.121  ns/op -1.55%
+StringBuilders.toStringCharWithBool8Utf16   avgt   15  8.654 ? 0.035  ns/op -18.44%
+StringBuilders.toStringCharWithNull8Latin1  avgt   15  5.550 ? 0.010  ns/op -1.96%
+StringBuilders.toStringCharWithNull8Utf16   avgt   15  8.108 ? 0.041  ns/op -16.76%

-------------

PR Comment: https://git.openjdk.org/jdk/pull/19626#issuecomment-2158201904
PR Comment: https://git.openjdk.org/jdk/pull/19626#issuecomment-2158296234