1 /*
2 * Copyright (c) 2023, 2024, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 */
23
24 /**
25 * @test
26 * @bug JDK-8310130
27 * @summary Special test cases for PhaseIdealLoop::move_unordered_reduction_out_of_loop
28 * Here a case with partial vectorization of the reduction.
29 * @library /test/lib /
30 * @run driver compiler.loopopts.superword.TestUnorderedReductionPartialVectorization
31 */
32
33 package compiler.loopopts.superword;
34
35 import compiler.lib.ir_framework.*;
36
37 public class TestUnorderedReductionPartialVectorization {
38 static final int RANGE = 1024;
39 static final int ITER = 10;
40
41 public static void main(String[] args) {
42 TestFramework.runWithFlags("-XX:+UnlockExperimentalVMOptions", "-XX:-UseCompactObjectHeaders", "-XX:-AlignVector");
43 TestFramework.runWithFlags("-XX:+UnlockExperimentalVMOptions", "-XX:-UseCompactObjectHeaders", "-XX:+AlignVector");
44 TestFramework.runWithFlags("-XX:+UnlockExperimentalVMOptions", "-XX:+UseCompactObjectHeaders", "-XX:-AlignVector");
45 TestFramework.runWithFlags("-XX:+UnlockExperimentalVMOptions", "-XX:+UseCompactObjectHeaders", "-XX:+AlignVector");
46 }
47
48 @Run(test = {"test1"})
49 @Warmup(0)
50 public void runTests() throws Exception {
51 int[] data = new int[RANGE];
52
53 init(data);
54 for (int i = 0; i < ITER; i++) {
55 long r1 = test1(data, i);
56 long r2 = ref1(data, i);
57 if (r1 != r2) {
58 throw new RuntimeException("Wrong result test1: " + r1 + " != " + r2);
59 }
60 }
61 }
62
63 @Test
64 @IR(counts = {IRNode.LOAD_VECTOR_I, IRNode.VECTOR_SIZE + "min(max_int, max_long)", "> 0",
65 IRNode.VECTOR_CAST_I2L, IRNode.VECTOR_SIZE + "min(max_int, max_long)", "> 0",
66 IRNode.OR_REDUCTION_V, "> 0",},
67 applyIfPlatform = {"64-bit", "true"},
68 applyIfCPUFeatureOr = {"avx2", "true"})
69 static long test1(int[] data, long sum) {
70 for (int i = 0; i < data.length; i+=2) {
71 // Mixing int and long ops means we only end up allowing half of the int
72 // loads in one pack, and we have two int packs. The first pack has one
73 // of the pairs missing because of the store, which creates a dependency.
74 // The first pack is rejected and left as scalar, the second pack succeeds
75 // with vectorization. That means we have a mixed scalar/vector reduction
76 // chain. This way it is possible that a vector-reduction has a scalar
77 // reduction as input, which is neigher a phi nor a vector reduction.
78 // In such a case, we must bail out of the optimization in
79 // PhaseIdealLoop::move_unordered_reduction_out_of_loop
80 int v = data[i]; // int read
81 data[0] = 0; // ruin the first pack
82 sum |= v; // long reduction (and implicit cast from int to long)
83
84 // This example used to rely on that reductions were ignored in SuperWord::unrolling_analysis,
85 // and hence the largest data type in the loop was the ints. This would then unroll the doubles
86 // for twice the vector length, and this resulted in us having twice as many packs. Because of
87 // the store "data[0] = 0", the first packs were destroyed, since they do not have power of 2
88 // size.
89 // Now, we no longer ignore reductions, and now we unroll half as much before SuperWord. This
90 // means we would only get one pack per operation, and that one would get ruined, and we have
91 // no vectorization. We now ensure there are again 2 packs per operation with a 2x hand unroll.
92 int v2 = data[i + 1];
93 sum |= v2;
94 }
95 return sum;
96 }
97
98 static long ref1(int[] data, long sum) {
99 for (int i = 0; i < data.length; i++) {
100 int v = data[i];
101 data[0] = 0;
102 sum |= v;
103 }
104 return sum;
105 }
106
107 static void init(int[] data) {
108 for (int i = 0; i < RANGE; i++) {
109 data[i] = i + 1;
110 }
111 }
112 }