1 | // RUN: %clang_cc1 -triple x86_64-apple-darwin -std=c++11 -emit-llvm -o - %s | FileCheck %s |
2 | |
3 | // Verify while loop is recognized after sequence of pragma clang loop directives. |
4 | void while_test(int *List, int Length) { |
5 | // CHECK: define {{.*}} @_Z10while_test |
6 | int i = 0; |
7 | |
8 | #pragma clang loop vectorize(enable) |
9 | #pragma clang loop interleave_count(4) |
10 | #pragma clang loop vectorize_width(4) |
11 | #pragma clang loop unroll(full) |
12 | #pragma clang loop distribute(enable) |
13 | while (i < Length) { |
14 | // CHECK: br label {{.*}}, !llvm.loop ![[LOOP_1:.*]] |
15 | List[i] = i * 2; |
16 | i++; |
17 | } |
18 | } |
19 | |
20 | // Verify do loop is recognized after multi-option pragma clang loop directive. |
21 | void do_test(int *List, int Length) { |
22 | int i = 0; |
23 | |
24 | #pragma clang loop vectorize_width(8) interleave_count(4) unroll(disable) distribute(disable) |
25 | do { |
26 | // CHECK: br i1 {{.*}}, label {{.*}}, label {{.*}}, !llvm.loop ![[LOOP_2:.*]] |
27 | List[i] = i * 2; |
28 | i++; |
29 | } while (i < Length); |
30 | } |
31 | |
32 | enum struct Tuner : short { Interleave = 4, Unroll = 8 }; |
33 | |
34 | // Verify for loop is recognized after sequence of pragma clang loop directives. |
35 | void for_test(int *List, int Length) { |
36 | #pragma clang loop interleave(enable) |
37 | #pragma clang loop interleave_count(static_cast<int>(Tuner::Interleave)) |
38 | #pragma clang loop unroll_count(static_cast<int>(Tuner::Unroll)) |
39 | for (int i = 0; i < Length; i++) { |
40 | // CHECK: br label {{.*}}, !llvm.loop ![[LOOP_3:.*]] |
41 | List[i] = i * 2; |
42 | } |
43 | } |
44 | |
45 | // Verify c++11 for range loop is recognized after |
46 | // sequence of pragma clang loop directives. |
47 | void for_range_test() { |
48 | double List[100]; |
49 | |
50 | #pragma clang loop vectorize_width(2) interleave_count(2) |
51 | for (int i : List) { |
52 | // CHECK: br label {{.*}}, !llvm.loop ![[LOOP_4:.*]] |
53 | List[i] = i; |
54 | } |
55 | } |
56 | |
57 | // Verify disable pragma clang loop directive generates correct metadata |
58 | void disable_test(int *List, int Length) { |
59 | #pragma clang loop vectorize(disable) unroll(disable) distribute(disable) |
60 | for (int i = 0; i < Length; i++) { |
61 | // CHECK: br label {{.*}}, !llvm.loop ![[LOOP_5:.*]] |
62 | List[i] = i * 2; |
63 | } |
64 | } |
65 | |
66 | #define VECWIDTH 2 |
67 | #define INTCOUNT 2 |
68 | #define UNROLLCOUNT 8 |
69 | |
70 | // Verify defines are correctly resolved in pragma clang loop directive |
71 | void for_define_test(int *List, int Length, int Value) { |
72 | #pragma clang loop vectorize_width(VECWIDTH) interleave_count(INTCOUNT) |
73 | #pragma clang loop unroll_count(UNROLLCOUNT) |
74 | for (int i = 0; i < Length; i++) { |
75 | // CHECK: br label {{.*}}, !llvm.loop ![[LOOP_6:.*]] |
76 | List[i] = i * Value; |
77 | } |
78 | } |
79 | |
80 | // Verify constant expressions are handled correctly. |
81 | void for_contant_expression_test(int *List, int Length) { |
82 | #pragma clang loop vectorize_width(1 + 4) |
83 | for (int i = 0; i < Length; i++) { |
84 | // CHECK: br label {{.*}}, !llvm.loop ![[LOOP_7:.*]] |
85 | List[i] = i; |
86 | } |
87 | |
88 | #pragma clang loop vectorize_width(3 + VECWIDTH) |
89 | for (int i = 0; i < Length; i++) { |
90 | // CHECK: br label {{.*}}, !llvm.loop ![[LOOP_8:.*]] |
91 | List[i] += i; |
92 | } |
93 | } |
94 | |
95 | // Verify metadata is generated when template is used. |
96 | template <typename A> |
97 | void for_template_test(A *List, int Length, A Value) { |
98 | #pragma clang loop vectorize_width(8) interleave_count(8) unroll_count(8) |
99 | for (int i = 0; i < Length; i++) { |
100 | // CHECK: br label {{.*}}, !llvm.loop ![[LOOP_9:.*]] |
101 | List[i] = i * Value; |
102 | } |
103 | } |
104 | |
105 | // Verify define is resolved correctly when template is used. |
106 | template <typename A, typename T> |
107 | void for_template_define_test(A *List, int Length, A Value) { |
108 | const T VWidth = VECWIDTH; |
109 | const T ICount = INTCOUNT; |
110 | const T UCount = UNROLLCOUNT; |
111 | #pragma clang loop vectorize_width(VWidth) interleave_count(ICount) |
112 | #pragma clang loop unroll_count(UCount) |
113 | for (int i = 0; i < Length; i++) { |
114 | // CHECK: br label {{.*}}, !llvm.loop ![[LOOP_10:.*]] |
115 | List[i] = i * Value; |
116 | } |
117 | } |
118 | |
119 | // Verify templates and constant expressions are handled correctly. |
120 | template <typename A, int V, int I, int U> |
121 | void for_template_constant_expression_test(A *List, int Length) { |
122 | #pragma clang loop vectorize_width(V) interleave_count(I) unroll_count(U) |
123 | for (int i = 0; i < Length; i++) { |
124 | // CHECK: br label {{.*}}, !llvm.loop ![[LOOP_11:.*]] |
125 | List[i] = i; |
126 | } |
127 | |
128 | #pragma clang loop vectorize_width(V * 2 + VECWIDTH) interleave_count(I * 2 + INTCOUNT) unroll_count(U * 2 + UNROLLCOUNT) |
129 | for (int i = 0; i < Length; i++) { |
130 | // CHECK: br label {{.*}}, !llvm.loop ![[LOOP_12:.*]] |
131 | List[i] += i; |
132 | } |
133 | |
134 | const int Scale = 4; |
135 | #pragma clang loop vectorize_width(Scale * V) interleave_count(Scale * I) unroll_count(Scale * U) |
136 | for (int i = 0; i < Length; i++) { |
137 | // CHECK: br label {{.*}}, !llvm.loop ![[LOOP_13:.*]] |
138 | List[i] += i; |
139 | } |
140 | |
141 | #pragma clang loop vectorize_width((Scale * V) + 2) |
142 | for (int i = 0; i < Length; i++) { |
143 | // CHECK: br label {{.*}}, !llvm.loop ![[LOOP_14:.*]] |
144 | List[i] += i; |
145 | } |
146 | } |
147 | |
148 | #undef VECWIDTH |
149 | #undef INTCOUNT |
150 | #undef UNROLLCOUNT |
151 | |
152 | // Use templates defined above. Test verifies metadata is generated correctly. |
153 | void template_test(double *List, int Length) { |
154 | double Value = 10; |
155 | |
156 | for_template_test<double>(List, Length, Value); |
157 | for_template_define_test<double, int>(List, Length, Value); |
158 | for_template_constant_expression_test<double, 2, 4, 8>(List, Length); |
159 | } |
160 | |
161 | // CHECK: ![[LOOP_1]] = distinct !{![[LOOP_1]], ![[WIDTH_4:.*]], ![[INTERLEAVE_4:.*]], ![[INTENABLE_1:.*]], ![[UNROLL_FULL:.*]], ![[DISTRIBUTE_ENABLE:.*]]} |
162 | // CHECK: ![[WIDTH_4]] = !{!"llvm.loop.vectorize.width", i32 4} |
163 | // CHECK: ![[INTERLEAVE_4]] = !{!"llvm.loop.interleave.count", i32 4} |
164 | // CHECK: ![[INTENABLE_1]] = !{!"llvm.loop.vectorize.enable", i1 true} |
165 | // CHECK: ![[UNROLL_FULL]] = !{!"llvm.loop.unroll.full"} |
166 | // CHECK: ![[DISTRIBUTE_ENABLE]] = !{!"llvm.loop.distribute.enable", i1 true} |
167 | // CHECK: ![[LOOP_2]] = distinct !{![[LOOP_2:.*]], ![[WIDTH_8:.*]], ![[INTERLEAVE_4:.*]], ![[UNROLL_DISABLE:.*]], ![[DISTRIBUTE_DISABLE:.*]]} |
168 | // CHECK: ![[WIDTH_8]] = !{!"llvm.loop.vectorize.width", i32 8} |
169 | // CHECK: ![[UNROLL_DISABLE]] = !{!"llvm.loop.unroll.disable"} |
170 | // CHECK: ![[DISTRIBUTE_DISABLE]] = !{!"llvm.loop.distribute.enable", i1 false} |
171 | // CHECK: ![[LOOP_3]] = distinct !{![[LOOP_3]], ![[INTERLEAVE_4:.*]], ![[UNROLL_8:.*]], ![[INTENABLE_1:.*]]} |
172 | // CHECK: ![[UNROLL_8]] = !{!"llvm.loop.unroll.count", i32 8} |
173 | // CHECK: ![[LOOP_4]] = distinct !{![[LOOP_4]], ![[WIDTH_2:.*]], ![[INTERLEAVE_2:.*]]} |
174 | // CHECK: ![[WIDTH_2]] = !{!"llvm.loop.vectorize.width", i32 2} |
175 | // CHECK: ![[INTERLEAVE_2]] = !{!"llvm.loop.interleave.count", i32 2} |
176 | // CHECK: ![[LOOP_5]] = distinct !{![[LOOP_5]], ![[WIDTH_1:.*]], ![[UNROLL_DISABLE:.*]], ![[DISTRIBUTE_DISABLE:.*]]} |
177 | // CHECK: ![[WIDTH_1]] = !{!"llvm.loop.vectorize.width", i32 1} |
178 | // CHECK: ![[LOOP_6]] = distinct !{![[LOOP_6]], ![[WIDTH_2:.*]], ![[INTERLEAVE_2:.*]], ![[UNROLL_8:.*]]} |
179 | // CHECK: ![[LOOP_7]] = distinct !{![[LOOP_7]], ![[WIDTH_5:.*]]} |
180 | // CHECK: ![[WIDTH_5]] = !{!"llvm.loop.vectorize.width", i32 5} |
181 | // CHECK: ![[LOOP_8]] = distinct !{![[LOOP_8]], ![[WIDTH_5:.*]]} |
182 | // CHECK: ![[LOOP_9]] = distinct !{![[LOOP_9]], ![[WIDTH_8:.*]], ![[INTERLEAVE_8:.*]], ![[UNROLL_8:.*]]} |
183 | // CHECK: ![[INTERLEAVE_8]] = !{!"llvm.loop.interleave.count", i32 8} |
184 | // CHECK: ![[LOOP_10]] = distinct !{![[LOOP_10]], ![[WIDTH_2:.*]], ![[INTERLEAVE_2:.*]], ![[UNROLL_8:.*]]} |
185 | // CHECK: ![[LOOP_11]] = distinct !{![[LOOP_11]], ![[WIDTH_2:.*]], ![[INTERLEAVE_4:.*]], ![[UNROLL_8:.*]]} |
186 | // CHECK: ![[LOOP_12]] = distinct !{![[LOOP_12]], ![[WIDTH_6:.*]], ![[INTERLEAVE_10:.*]], ![[UNROLL_24:.*]]} |
187 | // CHECK: ![[WIDTH_6]] = !{!"llvm.loop.vectorize.width", i32 6} |
188 | // CHECK: ![[INTERLEAVE_10]] = !{!"llvm.loop.interleave.count", i32 10} |
189 | // CHECK: ![[UNROLL_24]] = !{!"llvm.loop.unroll.count", i32 24} |
190 | // CHECK: ![[LOOP_13]] = distinct !{![[LOOP_13]], ![[WIDTH_8:.*]], ![[INTERLEAVE_16:.*]], ![[UNROLL_32:.*]]} |
191 | // CHECK: ![[INTERLEAVE_16]] = !{!"llvm.loop.interleave.count", i32 16} |
192 | // CHECK: ![[UNROLL_32]] = !{!"llvm.loop.unroll.count", i32 32} |
193 | // CHECK: ![[LOOP_14]] = distinct !{![[LOOP_14]], ![[WIDTH_10:.*]]} |
194 | // CHECK: ![[WIDTH_10]] = !{!"llvm.loop.vectorize.width", i32 10} |
195 | |