jdk/test/micro/org/openjdk/bench/java/lang/StringEquals.java
Ehsan Behrangi 00cc9be854 8381560: AArch64: Optimize String.equals intrinsic
This change improves the AArch64 implementation of String.equals by
introducing SIMD-based fast paths using SVE and NEON.

SVE implementation:
- Uses predicated loads and comparisons for short lengths (len < VL)
- Uses a full predicated loop for longer inputs
- Handles the tail via an overlapped compare at (base + len - VL)

NEON implementation:
- Uses an 8-byte pre-read to simplify tail handling and eliminate
  4/2/1-byte scalar branches
- Processes 16-byte chunks using LDP pair loads
- Uses CMP/CCMP to collapse comparisons into a single branch on mismatch

These changes reduce branch pressure and improve throughput for both
short and long strings.

Correctness:
- The implementation preserves existing semantics and matches behavior
  for all lengths

Testing:
- Updated and extended intrinsic tests to cover boundary conditions
  and mismatch positions

Benchmark:
Across evaluated macrobenchmarks (DaCapo and Renaissance), most workloads
spend <0.5% of CPU time in String.equals. DaCapo biojava is a notable
exception (~8–9%). In biojava, most String.equals calls are on very short
strings (1–2 bytes), where SVE shows ~1% end-to-end improvement, while
NEON is largely neutral or shows a small regression (~1%).

Measured using JMH on AArch64 (Arm Neoverse V2 CPU).
Values are relative (%) vs baseline. Negative values indicate regressions.
Mismatch results are reported across first(DF), middle(DM),
and last(DL) difference positions.

SVE results:
Length | L1_EQ  L1_DF  L1_DM  L1_DL | U16_EQ U16_DF U16_DM U16_DL | Avg
-------+----------------------------+-----------------------------+------
0      | 19.63                      | 20.05                      | 19.84
1      | 16.59  17.81  16.57  18.34 | 16.02   0.71   0.42   1.39 | 10.98
2      | 16.44   1.32   0.30  -0.16 | 15.90  -5.17  -4.55  -1.09 |  2.87
3      | 26.58   1.60   1.43  27.07 | 30.34  -8.86  -7.06  14.08 | 10.65
7      | 41.47  -2.94  -3.37  39.82 | 24.02  -8.82  -6.27  20.48 | 13.05
8      | 19.08  -1.16  -3.50  -0.90 | 22.49  -9.75  17.50  13.13 |  7.11
9      | 20.17  -4.12  -5.17  19.03 |  9.25  -2.24  21.35   3.39 |  7.71
15     | 19.48  -3.83  -4.50  19.01 | 29.26 -10.06  11.76  17.07 |  9.77
16     | 19.04  -3.15  16.41  16.85 | 38.37 -11.12  13.18  27.70 | 14.66
17     |  8.95  -2.40   5.68   6.38 | 16.32  -1.61   7.49  11.44 |  6.53
31     | 28.87  -0.01  19.79  23.37 | 41.43  -7.57  23.85  35.89 | 20.70
32     | 32.58   3.38  12.39  26.90 | 46.01 -10.99  20.53  44.15 | 21.87
33     | 11.62 -15.20   6.04  13.27 | 32.27  -9.38  20.33  32.28 | 11.40
63     | 44.66 -11.59  37.20  42.56 | 55.41 -10.57  43.19  55.90 | 32.10
64     | 53.99  -2.19  27.04  51.79 | 59.36  -8.72  35.41  60.32 | 34.63
65     | 33.79 -14.01  23.95  29.15 | 48.91 -11.58  36.54  50.03 | 24.60
127    | 62.10  -3.79  47.51  62.79 | 58.13  -8.89  60.68  60.90 | 42.43
128    | 67.38  -2.47  38.62  67.09 | 62.83  -0.38  51.72  61.87 | 43.33
129    | 52.02  -1.42  39.17  49.20 | 55.04  -9.52  53.23  52.81 | 36.32
256    | 66.11  -1.38  56.12  64.93 | 70.67  -3.68  53.67  74.54 | 47.62

Average:
         33.03  -2.40  17.46  30.34 | 37.60  -7.27  23.84  33.49 | 20.91

NEON results:
Length | L1_EQ  L1_DF  L1_DM  L1_DL | U16_EQ U16_DF U16_DM U16_DL | Avg
-------+----------------------------+-----------------------------+------
0      |  9.22                      |  8.69                      |  8.95
1      |  3.07   3.59   1.34   5.42 |  6.36  -6.20  -6.71 -10.59 | -0.47
2      |  3.23  -4.79  -5.67  -4.09 |  8.06  -8.43  -9.89  -9.20 | -3.85
3      | 12.80  -4.16  -3.95  11.28 | 11.94 -14.50 -14.41  11.83 |  1.36
7      | 31.00  -7.21 -12.76  33.59 |  4.73 -17.67 -17.38   1.65 |  1.99
8      |  4.43  -7.20  -4.70  -6.73 |  2.71 -18.05  -3.17  -4.05 | -4.59
9      | -9.33 -19.90 -16.27  -1.80 | 16.65 -23.72   4.26   8.78 | -5.17
15     | -6.96 -16.17 -15.60  -4.01 |  7.46 -24.60  -3.19  77.82 |  1.84
16     |  2.48 -16.38  -2.56  -3.62 |  9.08 -19.29  -5.45  77.93 |  5.27
17     |  4.88 -18.85  -0.18  19.35 | 18.43 -19.80  -8.37  84.96 | 10.05
31     |  6.92 -21.13  -4.62  60.71 | 24.42 -21.81   9.48 188.59 | 30.32
32     |  7.75 -24.20  -5.29  68.23 | 25.33 -20.57   4.17 183.65 | 29.88
33     | 20.23 -20.42 -11.33  98.60 | 23.76 -24.76   5.97 188.57 | 35.08
63     | 30.25 -22.30  14.29 152.37 | 25.02 -28.37  21.43 419.68 | 76.55
64     | 28.99 -22.91   9.03 185.51 | 38.20 -22.82  19.76 446.60 | 85.29
65     | 16.13 -21.77   1.45 211.38 | 27.94 -24.79  17.50 446.80 | 84.33
127    | 33.69 -28.94  28.75 429.23 | 41.75 -24.86  37.35 832.68 |168.71
128    | 26.28 -29.03  24.13 432.87 | 43.48 -18.53  26.44 810.20 |164.48
129    | 27.73 -20.30  20.84 439.01 | 44.09 -22.35  30.09 827.38 |168.31
256    | 53.30 -20.27  26.09 841.37 | 56.66 -21.07  47.41 1604.98|323.56

Average:
         15.30 -16.97   2.26 156.24 | 22.24 -20.12   8.17 325.70 | 59.10

Observations:
- SVE shows consistent improvements across all tested lengths, with gains
  increasing as input size grows
- NEON improves equal-string performance across all lengths
- NEON shows regressions for short mismatched inputs due to the loss
  of the scalar tbz-based early-exit sequence, which efficiently
  detects mismatches at small sizes and at early positions
- The scalar implementation relies on a branchy 4/2/1 tbz ladder,
  which is efficient for early mismatches but suboptimal for equal
  strings
- The NEON implementation replaces this with a branchless SIMD
  approach and performs upfront comparisons of the first and last
  8 bytes, improving throughput and late-mismatch detection
2026-06-05 12:22:15 +01:00

128 lines
4.3 KiB
Java

/*
* Copyright (c) 2019, Oracle and/or its affiliates. All rights reserved.
* Copyright 2026 Arm Limited and/or its affiliates.
* DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
*
* This code is free software; you can redistribute it and/or modify it
* under the terms of the GNU General Public License version 2 only, as
* published by the Free Software Foundation.
*
* This code is distributed in the hope that it will be useful, but WITHOUT
* ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
* FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
* version 2 for more details (a copy is included in the LICENSE file that
* accompanied this code).
*
* You should have received a copy of the GNU General Public License version
* 2 along with this work; if not, write to the Free Software Foundation,
* Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
*
* Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
* or visit www.oracle.com if you need additional information or have any
* questions.
*/
package org.openjdk.bench.java.lang;
import org.openjdk.jmh.annotations.*;
import java.util.concurrent.TimeUnit;
/*
* This benchmark naively explores String::equals performance
*/
@BenchmarkMode(Mode.AverageTime)
@OutputTimeUnit(TimeUnit.NANOSECONDS)
@State(Scope.Thread)
@Warmup(iterations = 5, time = 1)
@Measurement(iterations = 5, time = 1)
@Fork(value = 3)
public class StringEquals {
public String test = new String("0123456789");
public String test2 = new String("tgntogjnrognagronagroangroarngorngaorng");
public String test3 = new String(test); // equal to test, but not same
public String test4 = new String("0123\u01FF");
public String test5 = new String(test4); // equal to test4, but not same
public String test6 = new String("0123456780");
public String test7 = new String("0123\u01FE");
// string with parameterizable size
public String test8;
// same chars as test8, but different object; forces the intrinsic to read
// the entire string to check equality
public String test9;
// same chars as test8, except at length + diff_pos; set diff_pos to the
// worst case for the intrinsic being tested (usually -1, but could be -9
// if the intrinsic reads the last 8B first, or -length if the intrinsic
// reads the string backwards
public String test10;
@Param({"30"}) // can be used at runtime to define a length sweep
public int size;
@Param({"-1"}) // set to the worst location for the intrinsic under test
public int diff_pos;
@Setup
public void setup() {
if(size > 0) {
test8 = "a".repeat(size);
// NOTE 1: can't do test9 = new String(test8) or they'll share byte
// arrays, which improves cache hit rate of the equal-string case
test9 = "a".repeat(size);
StringBuilder sb = new StringBuilder("a".repeat(size));
sb.setCharAt(Math.max(test8.length() + diff_pos, 0), 'b');
test10 = sb.toString();
}
else {
// NOTE 2: can't use "a".repeat(0) or it returns the "" literal,
// which will early-exit from String.equals()
// NOTE 3: can't use no-arg String ctor or they'll share the byte
// array of the "" literal, which improves cache hit rate for
// intrinsics that read backwards into the object header
test8 = new String(new char [] {});
test9 = new String(new char [] {});
test10 = new String(new char [] {});
}
}
@Benchmark
public boolean different() {
return test.equals(test2);
}
@Benchmark
public boolean equal() {
return test.equals(test3);
}
@Benchmark
public boolean differentParam() {
return test8.equals(test10);
}
@Benchmark
public boolean equalParam() {
return test8.equals(test9);
}
@Benchmark
public boolean almostEqual() {
return test.equals(test6);
}
@Benchmark
public boolean almostEqualUTF16() {
return test4.equals(test7);
}
@Benchmark
public boolean differentCoders() {
return test.equals(test4);
}
@Benchmark
public boolean equalsUTF16() {
return test5.equals(test4);
}
}