1 // RUN: %clang_cc1 -verify -triple x86_64-apple-darwin10 -fopenmp -fopenmp-version=45 -x c++ -emit-llvm %s -o - -femit-all-decls | FileCheck %s --check-prefix CHECK --check-prefix OMP45
2 // RUN: %clang_cc1 -fopenmp -fopenmp-version=45 -x c++ -triple x86_64-apple-darwin10 -emit-pch -o %t %s
3 // RUN: %clang_cc1 -fopenmp -fopenmp-version=45 -x c++ -triple x86_64-apple-darwin10 -include-pch %t -verify %s -emit-llvm -o - -femit-all-decls | FileCheck %s --check-prefix CHECK --check-prefix OMP45
4 // RUN: %clang_cc1 -verify -triple x86_64-apple-darwin10 -fopenmp -fopenmp-version=50 -x c++ -emit-llvm %s -o - | FileCheck %s --check-prefix CHECK --check-prefix OMP50
5 // RUN: %clang_cc1 -fopenmp -fopenmp-version=50 -x c++ -triple x86_64-apple-darwin10 -emit-pch -o %t %s
6 // RUN: %clang_cc1 -fopenmp -fopenmp-version=50 -x c++ -triple x86_64-apple-darwin10 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CHECK --check-prefix OMP50
7
8 // RUN: %clang_cc1 -verify -triple x86_64-apple-darwin10 -fopenmp-simd -fopenmp-version=45 -x c++ -emit-llvm %s -o - -femit-all-decls | FileCheck --check-prefix SIMD-ONLY0 %s
9 // RUN: %clang_cc1 -fopenmp-simd -fopenmp-version=45 -x c++ -triple x86_64-apple-darwin10 -emit-pch -o %t %s
10 // RUN: %clang_cc1 -fopenmp-simd -fopenmp-version=45 -x c++ -triple x86_64-apple-darwin10 -include-pch %t -verify %s -emit-llvm -o - -femit-all-decls | FileCheck --check-prefix SIMD-ONLY0 %s
11 // RUN: %clang_cc1 -verify -triple x86_64-apple-darwin10 -fopenmp-simd -fopenmp-version=50 -x c++ -emit-llvm %s -o - -femit-all-decls | FileCheck --check-prefix SIMD-ONLY0 %s
12 // RUN: %clang_cc1 -fopenmp-simd -fopenmp-version=50 -x c++ -triple x86_64-apple-darwin10 -emit-pch -o %t %s
13 // RUN: %clang_cc1 -fopenmp-simd -fopenmp-version=50 -x c++ -triple x86_64-apple-darwin10 -include-pch %t -verify %s -emit-llvm -o - -femit-all-decls | FileCheck --check-prefix SIMD-ONLY0 %s
14 // SIMD-ONLY0-NOT: {{__kmpc|__tgt}}
15 // expected-no-diagnostics
16 #ifndef HEADER
17 #define HEADER
18
19 // CHECK-LABEL: @main
main(int argc,char ** argv)20 int main(int argc, char **argv) {
21 // CHECK: [[GTID:%.+]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* [[DEFLOC:@.+]])
22 // CHECK: call void @__kmpc_taskgroup(%struct.ident_t* [[DEFLOC]], i32 [[GTID]])
23 // CHECK: [[TASKV:%.+]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* [[DEFLOC]], i32 [[GTID]], i32 33, i64 80, i64 1, i32 (i32, i8*)* bitcast (i32 (i32, [[TDP_TY:%.+]]*)* [[TASK1:@.+]] to i32 (i32, i8*)*))
24 // CHECK: [[TASK:%.+]] = bitcast i8* [[TASKV]] to [[TDP_TY]]*
25 // CHECK: [[TASK_DATA:%.+]] = getelementptr inbounds [[TDP_TY]], [[TDP_TY]]* [[TASK]], i32 0, i32 0
26 // CHECK: [[DOWN:%.+]] = getelementptr inbounds [[TD_TY:%.+]], [[TD_TY]]* [[TASK_DATA]], i32 0, i32 5
27 // CHECK: store i64 0, i64* [[DOWN]],
28 // CHECK: [[UP:%.+]] = getelementptr inbounds [[TD_TY]], [[TD_TY]]* [[TASK_DATA]], i32 0, i32 6
29 // CHECK: store i64 9, i64* [[UP]],
30 // CHECK: [[ST:%.+]] = getelementptr inbounds [[TD_TY]], [[TD_TY]]* [[TASK_DATA]], i32 0, i32 7
31 // CHECK: store i64 1, i64* [[ST]],
32 // CHECK: [[ST_VAL:%.+]] = load i64, i64* [[ST]],
33 // CHECK: call void @__kmpc_taskloop(%struct.ident_t* [[DEFLOC]], i32 [[GTID]], i8* [[TASKV]], i32 1, i64* [[DOWN]], i64* [[UP]], i64 [[ST_VAL]], i32 1, i32 0, i64 0, i8* null)
34 // CHECK: call void @__kmpc_end_taskgroup(%struct.ident_t* [[DEFLOC]], i32 [[GTID]])
35 #pragma omp taskloop simd priority(argc)
36 for (int i = 0; i < 10; ++i)
37 ;
38 // CHECK: [[TASKV:%.+]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* [[DEFLOC]], i32 [[GTID]], i32 1, i64 80, i64 1, i32 (i32, i8*)* bitcast (i32 (i32, [[TDP_TY:%.+]]*)* [[TASK2:@.+]] to i32 (i32, i8*)*))
39 // CHECK: [[TASK:%.+]] = bitcast i8* [[TASKV]] to [[TDP_TY]]*
40 // CHECK: [[TASK_DATA:%.+]] = getelementptr inbounds [[TDP_TY]], [[TDP_TY]]* [[TASK]], i32 0, i32 0
41 // CHECK: [[DOWN:%.+]] = getelementptr inbounds [[TD_TY:%.+]], [[TD_TY]]* [[TASK_DATA]], i32 0, i32 5
42 // CHECK: store i64 0, i64* [[DOWN]],
43 // CHECK: [[UP:%.+]] = getelementptr inbounds [[TD_TY]], [[TD_TY]]* [[TASK_DATA]], i32 0, i32 6
44 // CHECK: store i64 9, i64* [[UP]],
45 // CHECK: [[ST:%.+]] = getelementptr inbounds [[TD_TY]], [[TD_TY]]* [[TASK_DATA]], i32 0, i32 7
46 // CHECK: store i64 1, i64* [[ST]],
47 // CHECK: [[ST_VAL:%.+]] = load i64, i64* [[ST]],
48 // CHECK: [[GRAINSIZE:%.+]] = zext i32 %{{.+}} to i64
49 // CHECK: call void @__kmpc_taskloop(%struct.ident_t* [[DEFLOC]], i32 [[GTID]], i8* [[TASKV]], i32 1, i64* [[DOWN]], i64* [[UP]], i64 [[ST_VAL]], i32 1, i32 1, i64 [[GRAINSIZE]], i8* null)
50 #pragma omp taskloop simd nogroup grainsize(argc) simdlen(4)
51 for (int i = 0; i < 10; ++i)
52 ;
53 // CHECK: call void @__kmpc_taskgroup(%struct.ident_t* [[DEFLOC]], i32 [[GTID]])
54 // OMP45: [[TASKV:%.+]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* [[DEFLOC]], i32 [[GTID]], i32 1, i64 80, i64 16, i32 (i32, i8*)* bitcast (i32 (i32, [[TDP_TY:%.+]]*)* [[TASK3:@.+]] to i32 (i32, i8*)*))
55 // OMP50: [[TASKV:%.+]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* [[DEFLOC]], i32 [[GTID]], i32 1, i64 80, i64 24, i32 (i32, i8*)* bitcast (i32 (i32, [[TDP_TY:%.+]]*)* [[TASK3:@.+]] to i32 (i32, i8*)*))
56 // CHECK: [[TASK:%.+]] = bitcast i8* [[TASKV]] to [[TDP_TY]]*
57 // CHECK: [[TASK_DATA:%.+]] = getelementptr inbounds [[TDP_TY]], [[TDP_TY]]* [[TASK]], i32 0, i32 0
58 // OMP45: [[IF:%.+]] = icmp ne i32 %{{.+}}, 0
59 // OMP50: [[IF_VAL:%.+]] = load i8, i8* %
60 // OMP50: [[IF:%.+]] = trunc i8 [[IF_VAL]] to i1
61 // CHECK: [[IF_INT:%.+]] = sext i1 [[IF]] to i32
62 // CHECK: [[DOWN:%.+]] = getelementptr inbounds [[TD_TY:%.+]], [[TD_TY]]* [[TASK_DATA]], i32 0, i32 5
63 // CHECK: store i64 0, i64* [[DOWN]],
64 // CHECK: [[UP:%.+]] = getelementptr inbounds [[TD_TY]], [[TD_TY]]* [[TASK_DATA]], i32 0, i32 6
65 // CHECK: store i64 %{{.+}}, i64* [[UP]],
66 // CHECK: [[ST:%.+]] = getelementptr inbounds [[TD_TY]], [[TD_TY]]* [[TASK_DATA]], i32 0, i32 7
67 // CHECK: store i64 1, i64* [[ST]],
68 // CHECK: [[ST_VAL:%.+]] = load i64, i64* [[ST]],
69 // CHECK: call void @__kmpc_taskloop(%struct.ident_t* [[DEFLOC]], i32 [[GTID]], i8* [[TASKV]], i32 [[IF_INT]], i64* [[DOWN]], i64* [[UP]], i64 [[ST_VAL]], i32 1, i32 2, i64 4, i8* null)
70 // CHECK: call void @__kmpc_end_taskgroup(%struct.ident_t* [[DEFLOC]], i32 [[GTID]])
71 int i;
72 #pragma omp taskloop simd if(argc) shared(argc, argv) collapse(2) num_tasks(4) safelen(32)
73 for (i = 0; i < argc; ++i)
74 for (int j = argc; j < argv[argc][argc]; ++j)
75 ;
76 }
77
78 // CHECK: define internal i32 [[TASK1]](
79 // CHECK: [[DOWN:%.+]] = getelementptr inbounds [[TD_TY:%.+]], [[TD_TY]]* %{{.+}}, i32 0, i32 5
80 // CHECK: [[DOWN_VAL:%.+]] = load i64, i64* [[DOWN]],
81 // CHECK: [[UP:%.+]] = getelementptr inbounds [[TD_TY]], [[TD_TY]]* %{{.+}}, i32 0, i32 6
82 // CHECK: [[UP_VAL:%.+]] = load i64, i64* [[UP]],
83 // CHECK: [[ST:%.+]] = getelementptr inbounds [[TD_TY]], [[TD_TY]]* %{{.+}}, i32 0, i32 7
84 // CHECK: [[ST_VAL:%.+]] = load i64, i64* [[ST]],
85 // CHECK: [[LITER:%.+]] = getelementptr inbounds [[TD_TY]], [[TD_TY]]* %{{.+}}, i32 0, i32 8
86 // CHECK: [[LITER_VAL:%.+]] = load i32, i32* [[LITER]],
87 // CHECK: store i64 [[DOWN_VAL]], i64* [[LB:%[^,]+]],
88 // CHECK: store i64 [[UP_VAL]], i64* [[UB:%[^,]+]],
89 // CHECK: store i64 [[ST_VAL]], i64* [[ST:%[^,]+]],
90 // CHECK: store i32 [[LITER_VAL]], i32* [[LITER:%[^,]+]],
91 // CHECK: [[LB_VAL:%.+]] = load i64, i64* [[LB]],
92 // CHECK: [[LB_I32:%.+]] = trunc i64 [[LB_VAL]] to i32
93 // CHECK: store i32 [[LB_I32]], i32* [[CNT:%.+]],
94 // CHECK: br label
95 // CHECK: [[VAL:%.+]] = load i32, i32* [[CNT]],{{.*}}!llvm.access.group
96 // CHECK: [[VAL_I64:%.+]] = sext i32 [[VAL]] to i64
97 // CHECK: [[UB_VAL:%.+]] = load i64, i64* [[UB]],{{.*}}!llvm.access.group
98 // CHECK: [[CMP:%.+]] = icmp ule i64 [[VAL_I64]], [[UB_VAL]]
99 // CHECK: br i1 [[CMP]], label %{{.+}}, label %{{.+}}
100 // CHECK: load i32, i32* %{{.*}}!llvm.access.group
101 // CHECK: store i32 %{{.*}}!llvm.access.group
102 // CHECK: load i32, i32* %{{.*}}!llvm.access.group
103 // CHECK: add nsw i32 %{{.+}}, 1
104 // CHECK: store i32 %{{.+}}, i32* %{{.*}}!llvm.access.group
105 // CHECK: br label %{{.*}}!llvm.loop
106 // CHECK: ret i32 0
107
108 // CHECK: define internal i32 [[TASK2]](
109 // CHECK: [[DOWN:%.+]] = getelementptr inbounds [[TD_TY:%.+]], [[TD_TY]]* %{{.+}}, i32 0, i32 5
110 // CHECK: [[DOWN_VAL:%.+]] = load i64, i64* [[DOWN]],
111 // CHECK: [[UP:%.+]] = getelementptr inbounds [[TD_TY]], [[TD_TY]]* %{{.+}}, i32 0, i32 6
112 // CHECK: [[UP_VAL:%.+]] = load i64, i64* [[UP]],
113 // CHECK: [[ST:%.+]] = getelementptr inbounds [[TD_TY]], [[TD_TY]]* %{{.+}}, i32 0, i32 7
114 // CHECK: [[ST_VAL:%.+]] = load i64, i64* [[ST]],
115 // CHECK: [[LITER:%.+]] = getelementptr inbounds [[TD_TY]], [[TD_TY]]* %{{.+}}, i32 0, i32 8
116 // CHECK: [[LITER_VAL:%.+]] = load i32, i32* [[LITER]],
117 // CHECK: store i64 [[DOWN_VAL]], i64* [[LB:%[^,]+]],
118 // CHECK: store i64 [[UP_VAL]], i64* [[UB:%[^,]+]],
119 // CHECK: store i64 [[ST_VAL]], i64* [[ST:%[^,]+]],
120 // CHECK: store i32 [[LITER_VAL]], i32* [[LITER:%[^,]+]],
121 // CHECK: [[LB_VAL:%.+]] = load i64, i64* [[LB]],
122 // CHECK: [[LB_I32:%.+]] = trunc i64 [[LB_VAL]] to i32
123 // CHECK: store i32 [[LB_I32]], i32* [[CNT:%.+]],
124 // CHECK: br label
125 // CHECK: [[VAL:%.+]] = load i32, i32* [[CNT]],{{.*}}!llvm.access.group
126 // CHECK: [[VAL_I64:%.+]] = sext i32 [[VAL]] to i64
127 // CHECK: [[UB_VAL:%.+]] = load i64, i64* [[UB]],{{.*}}!llvm.access.group
128 // CHECK: [[CMP:%.+]] = icmp ule i64 [[VAL_I64]], [[UB_VAL]]
129 // CHECK: br i1 [[CMP]], label %{{.+}}, label %{{.+}}
130 // CHECK: load i32, i32* %{{.*}}!llvm.access.group
131 // CHECK: store i32 %{{.*}}!llvm.access.group
132 // CHECK: load i32, i32* %{{.*}}!llvm.access.group
133 // CHECK: add nsw i32 %{{.+}}, 1
134 // CHECK: store i32 %{{.+}}, i32* %{{.*}}!llvm.access.group
135 // CHECK: br label %{{.*}}!llvm.loop
136 // CHECK: ret i32 0
137
138 // CHECK: define internal i32 [[TASK3]](
139 // CHECK: [[DOWN:%.+]] = getelementptr inbounds [[TD_TY:%.+]], [[TD_TY]]* %{{.+}}, i32 0, i32 5
140 // CHECK: [[DOWN_VAL:%.+]] = load i64, i64* [[DOWN]],
141 // CHECK: [[UP:%.+]] = getelementptr inbounds [[TD_TY]], [[TD_TY]]* %{{.+}}, i32 0, i32 6
142 // CHECK: [[UP_VAL:%.+]] = load i64, i64* [[UP]],
143 // CHECK: [[ST:%.+]] = getelementptr inbounds [[TD_TY]], [[TD_TY]]* %{{.+}}, i32 0, i32 7
144 // CHECK: [[ST_VAL:%.+]] = load i64, i64* [[ST]],
145 // CHECK: [[LITER:%.+]] = getelementptr inbounds [[TD_TY]], [[TD_TY]]* %{{.+}}, i32 0, i32 8
146 // CHECK: [[LITER_VAL:%.+]] = load i32, i32* [[LITER]],
147 // CHECK: store i64 [[DOWN_VAL]], i64* [[LB:%[^,]+]],
148 // CHECK: store i64 [[UP_VAL]], i64* [[UB:%[^,]+]],
149 // CHECK: store i64 [[ST_VAL]], i64* [[ST:%[^,]+]],
150 // CHECK: store i32 [[LITER_VAL]], i32* [[LITER:%[^,]+]],
151 // CHECK: [[LB_VAL:%.+]] = load i64, i64* [[LB]],
152 // CHECK: store i64 [[LB_VAL]], i64* [[CNT:%.+]],
153 // CHECK: br label
154 // CHECK-NOT: !llvm.access.group
155 // CHECK: br label %{{.*}}!llvm.loop
156 // CHECK: ret i32 0
157
158 // CHECK-LABEL: @_ZN1SC2Ei
159 struct S {
160 int a;
SS161 S(int c) {
162 // CHECK: [[GTID:%.+]] = call i32 @__kmpc_global_thread_num(%struct.ident_t* [[DEFLOC:@.+]])
163 // CHECK: [[TASKV:%.+]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* [[DEFLOC]], i32 [[GTID]], i32 1, i64 80, i64 16, i32 (i32, i8*)* bitcast (i32 (i32, [[TDP_TY:%.+]]*)* [[TASK4:@.+]] to i32 (i32, i8*)*))
164 // CHECK: [[TASK:%.+]] = bitcast i8* [[TASKV]] to [[TDP_TY]]*
165 // CHECK: [[TASK_DATA:%.+]] = getelementptr inbounds [[TDP_TY]], [[TDP_TY]]* [[TASK]], i32 0, i32 0
166 // CHECK: [[DOWN:%.+]] = getelementptr inbounds [[TD_TY:%.+]], [[TD_TY]]* [[TASK_DATA]], i32 0, i32 5
167 // CHECK: store i64 0, i64* [[DOWN]],
168 // CHECK: [[UP:%.+]] = getelementptr inbounds [[TD_TY]], [[TD_TY]]* [[TASK_DATA]], i32 0, i32 6
169 // CHECK: store i64 %{{.+}}, i64* [[UP]],
170 // CHECK: [[ST:%.+]] = getelementptr inbounds [[TD_TY]], [[TD_TY]]* [[TASK_DATA]], i32 0, i32 7
171 // CHECK: store i64 1, i64* [[ST]],
172 // CHECK: [[ST_VAL:%.+]] = load i64, i64* [[ST]],
173 // CHECK: [[NUM_TASKS:%.+]] = zext i32 %{{.+}} to i64
174 // CHECK: call void @__kmpc_taskloop(%struct.ident_t* [[DEFLOC]], i32 [[GTID]], i8* [[TASKV]], i32 1, i64* [[DOWN]], i64* [[UP]], i64 [[ST_VAL]], i32 1, i32 2, i64 [[NUM_TASKS]], i8* null)
175 #pragma omp taskloop simd shared(c) num_tasks(a) simdlen(8) safelen(64)
176 for (a = 0; a < c; ++a)
177 ;
178 }
179 } s(1);
180
181 // CHECK: define internal i32 [[TASK4]](
182 // CHECK: [[DOWN:%.+]] = getelementptr inbounds [[TD_TY:%.+]], [[TD_TY]]* %{{.+}}, i32 0, i32 5
183 // CHECK: [[DOWN_VAL:%.+]] = load i64, i64* [[DOWN]],
184 // CHECK: [[UP:%.+]] = getelementptr inbounds [[TD_TY]], [[TD_TY]]* %{{.+}}, i32 0, i32 6
185 // CHECK: [[UP_VAL:%.+]] = load i64, i64* [[UP]],
186 // CHECK: [[ST:%.+]] = getelementptr inbounds [[TD_TY]], [[TD_TY]]* %{{.+}}, i32 0, i32 7
187 // CHECK: [[ST_VAL:%.+]] = load i64, i64* [[ST]],
188 // CHECK: [[LITER:%.+]] = getelementptr inbounds [[TD_TY]], [[TD_TY]]* %{{.+}}, i32 0, i32 8
189 // CHECK: [[LITER_VAL:%.+]] = load i32, i32* [[LITER]],
190 // CHECK: store i64 [[DOWN_VAL]], i64* [[LB:%[^,]+]],
191 // CHECK: store i64 [[UP_VAL]], i64* [[UB:%[^,]+]],
192 // CHECK: store i64 [[ST_VAL]], i64* [[ST:%[^,]+]],
193 // CHECK: store i32 [[LITER_VAL]], i32* [[LITER:%[^,]+]],
194 // CHECK: [[LB_VAL:%.+]] = load i64, i64* [[LB]],
195 // CHECK: [[LB_I32:%.+]] = trunc i64 [[LB_VAL]] to i32
196 // CHECK: store i32 [[LB_I32]], i32* [[CNT:%.+]],
197 // CHECK: br label
198 // CHECK: [[VAL:%.+]] = load i32, i32* [[CNT]],
199 // CHECK: [[VAL_I64:%.+]] = sext i32 [[VAL]] to i64
200 // CHECK: [[UB_VAL:%.+]] = load i64, i64* [[UB]],
201 // CHECK: [[CMP:%.+]] = icmp ule i64 [[VAL_I64]], [[UB_VAL]]
202 // CHECK: br i1 [[CMP]], label %{{.+}}, label %{{.+}}
203 // CHECK: load i32, i32* %
204 // CHECK-NOT: !llvm.access.group
205 // CHECK: store i32 %
206 // CHECK-NOT: !llvm.access.group
207 // CHECK: load i32, i32* %
208 // CHECK-NOT: !llvm.access.group
209 // CHECK: add nsw i32 %{{.+}}, 1
210 // CHECK: store i32 %{{.+}}, i32* %
211 // CHECK-NOT: !llvm.access.group
212 // CHECK: br label %{{.*}}!llvm.loop
213 // CHECK: ret i32 0
214
215 // CHECK: !{!"llvm.loop.vectorize.enable", i1 true}
216 // CHECK: !{!"llvm.loop.vectorize.width", i32 4}
217 // CHECK: !{!"llvm.loop.vectorize.width", i32 32}
218 // CHECK: !{!"llvm.loop.vectorize.width", i32 8}
219
220 #endif
221