1 // RUN: %clang_cc1 -DLAMBDA -verify -fopenmp -x c++ -std=c++11 -triple powerpc64le-unknown-unknown -fopenmp-targets=powerpc64le-ibm-linux-gnu -emit-llvm %s -o - | FileCheck %s --check-prefix LAMBDA --check-prefix LAMBDA-64 2 // RUN: %clang_cc1 -DLAMBDA -fopenmp -x c++ -std=c++11 -triple powerpc64le-unknown-unknown -fopenmp-targets=powerpc64le-ibm-linux-gnu -emit-pch -o %t %s 3 // RUN: %clang_cc1 -DLAMBDA -fopenmp -x c++ -triple powerpc64le-unknown-unknown -fopenmp-targets=powerpc64le-ibm-linux-gnu -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix LAMBDA --check-prefix LAMBDA-64 4 // RUN: %clang_cc1 -DLAMBDA -verify -fopenmp -x c++ -std=c++11 -triple i386-unknown-unknown -fopenmp-targets=i386-pc-linux-gnu -emit-llvm %s -o - | FileCheck %s --check-prefix LAMBDA --check-prefix LAMBDA-32 5 // RUN: %clang_cc1 -DLAMBDA -fopenmp -x c++ -std=c++11 -triple i386-unknown-unknown -fopenmp-targets=i386-pc-linux-gnu -emit-pch -o %t %s 6 // RUN: %clang_cc1 -DLAMBDA -fopenmp -x c++ -std=c++11 -triple i386-unknown-unknown -fopenmp-targets=i386-pc-linux-gnu -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix LAMBDA --check-prefix LAMBDA-32 7 8 // RUN: %clang_cc1 -verify -fopenmp -x c++ -std=c++11 -triple powerpc64le-unknown-unknown -fopenmp-targets=powerpc64le-ibm-linux-gnu -emit-llvm %s -o - | FileCheck %s --check-prefix CHECK --check-prefix CHECK-64 9 // RUN: %clang_cc1 -fopenmp -x c++ -std=c++11 -triple powerpc64le-unknown-unknown -fopenmp-targets=powerpc64le-ibm-linux-gnu -emit-pch -o %t %s 10 // RUN: %clang_cc1 -fopenmp -x c++ -triple powerpc64le-unknown-unknown -fopenmp-targets=powerpc64le-ibm-linux-gnu -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CHECK --check-prefix CHECK-64 11 // RUN: %clang_cc1 -verify -fopenmp -x c++ -std=c++11 -triple i386-unknown-unknown -fopenmp-targets=i386-pc-linux-gnu -emit-llvm %s -o - | FileCheck %s --check-prefix CHECK --check-prefix CHECK-32 12 // RUN: %clang_cc1 -fopenmp -x c++ -std=c++11 -triple i386-unknown-unknown -fopenmp-targets=i386-pc-linux-gnu -emit-pch -o %t %s 13 // RUN: %clang_cc1 -fopenmp -x c++ -std=c++11 -triple i386-unknown-unknown -fopenmp-targets=i386-pc-linux-gnu -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CHECK --check-prefix CHECK-32 14 // expected-no-diagnostics 15 #ifndef HEADER 16 #define HEADER 17 18 template <class T> 19 struct S { 20 T f; 21 S(T a) : f(a) {} 22 S() : f() {} 23 operator T() { return T(); } 24 ~S() {} 25 }; 26 27 // CHECK: [[S_FLOAT_TY:%.+]] = type { float } 28 // CHECK: [[S_INT_TY:%.+]] = type { i{{[0-9]+}} } 29 template <typename T> 30 T tmain() { 31 S<T> test; 32 T t_var = T(); 33 T vec[] = {1, 2}; 34 S<T> s_arr[] = {1, 2}; 35 S<T> &var = test; 36 #pragma omp target 37 #pragma omp teams 38 #pragma omp distribute simd lastprivate(t_var, vec, s_arr, s_arr, var, var) 39 for (int i = 0; i < 2; ++i) { 40 vec[i] = t_var; 41 s_arr[i] = var; 42 } 43 return T(); 44 } 45 46 int main() { 47 static int svar; 48 volatile double g; 49 volatile double &g1 = g; 50 51 #ifdef LAMBDA 52 // LAMBDA-LABEL: @main 53 // LAMBDA: call{{.*}} void [[OUTER_LAMBDA:@.+]]( 54 [&]() { 55 static float sfvar; 56 // LAMBDA: define{{.*}} internal{{.*}} void [[OUTER_LAMBDA]]( 57 // LAMBDA: call i{{[0-9]+}} @__tgt_target_teams( 58 // LAMBDA: call void [[OFFLOADING_FUN:@.+]]( 59 60 // LAMBDA: define{{.+}} void [[OFFLOADING_FUN]]( 61 // LAMBDA: call {{.*}}void {{.+}} @__kmpc_fork_teams({{.+}}, i32 4, {{.+}}* [[OMP_OUTLINED:@.+]] to {{.+}}) 62 #pragma omp target 63 #pragma omp teams 64 #pragma omp distribute simd lastprivate(g, g1, svar, sfvar) 65 for (int i = 0; i < 2; ++i) { 66 // LAMBDA: define{{.*}} internal{{.*}} void [[OMP_OUTLINED]](i32* noalias %{{.+}}, i32* noalias %{{.+}}, double*{{.+}} [[G_IN:%.+]], double*{{.+}} [[G1_IN:%.+]], i{{[0-9]+}}*{{.+}} [[SVAR_IN:%.+]], float*{{.+}} [[SFVAR_IN:%.+]]) 67 // LAMBDA: [[G_PRIVATE_ADDR:%.+]] = alloca double*, 68 // LAMBDA: [[G1_PRIVATE_ADDR:%.+]] = alloca double*, 69 // LAMBDA: [[SVAR_PRIVATE_ADDR:%.+]] = alloca i{{[0-9]+}}*, 70 // LAMBDA: [[SFVAR_PRIVATE_ADDR:%.+]] = alloca float*, 71 // LAMBDA: [[TMP_G1:%.+]] = alloca double*, 72 // loop variables 73 // LAMBDA: {{.+}} = alloca i{{[0-9]+}}, 74 // LAMBDA: {{.+}} = alloca i{{[0-9]+}}, 75 // LAMBDA: {{.+}} = alloca i{{[0-9]+}}, 76 // LAMBDA: {{.+}} = alloca i{{[0-9]+}}, 77 // LAMBDA: {{.+}} = alloca i{{[0-9]+}}, 78 // LAMBDA: [[OMP_IS_LAST:%.+]] = alloca i{{[0-9]+}}, 79 // LAMBDA: [[G_PRIVATE:%.+]] = alloca double, 80 // LAMBDA: [[G1_PRIVATE:%.+]] = alloca double, 81 // LAMBDA: [[TMP_G1_PRIVATE:%.+]] = alloca double*, 82 // LAMBDA: [[SVAR_PRIVATE:%.+]] = alloca i{{[0-9]+}}, 83 // LAMBDA: [[SFVAR_PRIVATE:%.+]] = alloca float, 84 // LAMBDA: store double* [[G_IN]], double** [[G_PRIVATE_ADDR]], 85 // LAMBDA: store double* [[G1_IN]], double** [[G1_PRIVATE_ADDR]], 86 // LAMBDA: store i{{[0-9]+}}* [[SVAR_IN]], i{{[0-9]+}}** [[SVAR_PRIVATE_ADDR]], 87 // LAMBDA: store float* [[SFVAR_IN]], float** [[SFVAR_PRIVATE_ADDR]], 88 89 // init private variables 90 // LAMBDA: [[G_IN_REF:%.+]] = load double*, double** [[G_PRIVATE_ADDR]], 91 // LAMBDA: [[SVAR_IN_REF:%.+]] = load i{{[0-9]+}}*, i{{[0-9]+}}** [[SVAR_PRIVATE_ADDR]], 92 // LAMBDA: [[SFVAR_IN_REF:%.+]] = load float*, float** [[SFVAR_PRIVATE_ADDR]], 93 // LAMBDA: [[G1_IN_REF:%.+]] = load double*, double** [[G1_PRIVATE_ADDR]], 94 // LAMBDA: store double* [[G1_IN_REF]], double** [[TMP_G1]], 95 // LAMBDA: [[TMP_G1_VAL:%.+]] = load double*, double** [[TMP_G1]], 96 // LAMBDA: store double* [[G1_PRIVATE]], double** [[TMP_G1_PRIVATE]], 97 g = 1; 98 g1 = 1; 99 svar = 3; 100 sfvar = 4.0; 101 // LAMBDA: call {{.*}}void @__kmpc_for_static_init_4( 102 // LAMBDA: store double 1.0{{.+}}, double* [[G_PRIVATE]], 103 // LAMBDA: [[TMP_G1_REF:%.+]] = load double*, double** [[TMP_G1_PRIVATE]], 104 // LAMBDA: store{{.+}} double 1.0{{.+}}, double* [[TMP_G1_REF]], 105 // LAMBDA: store i{{[0-9]+}} 3, i{{[0-9]+}}* [[SVAR_PRIVATE]], 106 // LAMBDA: store float 4.0{{.+}}, float* [[SFVAR_PRIVATE]], 107 // LAMBDA: [[G_PRIVATE_ADDR_REF:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* [[ARG:%.+]], i{{[0-9]+}} 0, i{{[0-9]+}} 0 108 // LAMBDA: store double* [[G_PRIVATE]], double** [[G_PRIVATE_ADDR_REF]], 109 // LAMBDA: [[TMP_PRIVATE_ADDR_REF:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* [[ARG:%.+]], i{{[0-9]+}} 0, i{{[0-9]+}} 1 110 // LAMBDA: [[G1_PRIVATE_ADDR_FROM_TMP:%.+]] = load double*, double** [[TMP_G1_PRIVATE]], 111 // LAMBDA: store double* [[G1_PRIVATE_ADDR_FROM_TMP]], double** [[TMP_PRIVATE_ADDR_REF]], 112 // LAMBDA: [[SVAR_PRIVATE_ADDR_REF:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* [[ARG:%.+]], i{{[0-9]+}} 0, i{{[0-9]+}} 2 113 // LAMBDA: store i{{[0-9]+}}* [[SVAR_PRIVATE]], i{{[0-9]+}}** [[SVAR_PRIVATE_ADDR_REF]] 114 // LAMBDA: [[SFVAR_PRIVATE_ADDR_REF:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* [[ARG:%.+]], i{{[0-9]+}} 0, i{{[0-9]+}} 3 115 // LAMBDA: store float* [[SFVAR_PRIVATE]], float** [[SFVAR_PRIVATE_ADDR_REF]] 116 // LAMBDA: call{{.*}} void [[INNER_LAMBDA:@.+]](%{{.+}}* [[ARG]]) 117 // LAMBDA: call {{.*}}void @__kmpc_for_static_fini( 118 119 // linear counter 120 // LAMBDA: [[OMP_IS_LAST_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[OMP_IS_LAST]], 121 // LAMBDA: [[IS_LAST_IT:%.+]] = icmp ne i{{[0-9]+}} [[OMP_IS_LAST_VAL]], 0 122 // LAMBDA: br i1 [[IS_LAST_IT]], label %[[OMP_COUNTER_BLOCK:.+]], label %[[OMP_COUNTER_DONE:.+]] 123 // LAMBDA: [[OMP_COUNTER_BLOCK]]: 124 // LAMBDA: store i{{[0-9]+}} 2, i{{[0-9]+}}* % 125 // LAMBDA: br label %[[OMP_COUNTER_DONE]] 126 // LAMBDA: [[OMP_COUNTER_DONE]]: 127 128 129 // LAMBDA: [[OMP_IS_LAST_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[OMP_IS_LAST]], 130 // LAMBDA: [[IS_LAST_IT:%.+]] = icmp ne i{{[0-9]+}} [[OMP_IS_LAST_VAL]], 0 131 // LAMBDA: br i1 [[IS_LAST_IT]], label %[[OMP_LASTPRIV_BLOCK:.+]], label %[[OMP_LASTPRIV_DONE:.+]] 132 133 // LAMBDA: [[OMP_LASTPRIV_BLOCK]]: 134 // LAMBDA: [[G_PRIV_VAL:%.+]] = load double, double* [[G_PRIVATE]], 135 // LAMBDA: store{{.*}} double [[G_PRIV_VAL]], double* [[G_IN_REF]], 136 // LAMBDA: [[TMP_G1_PRIV_REF:%.+]] = load double*, double** [[TMP_G1_PRIVATE]], 137 // LAMBDA: [[TMP_G1_PRIV_VAL:%.+]] = load double, double* [[TMP_G1_PRIV_REF]], 138 // LAMBDA: store{{.*}} double [[TMP_G1_PRIV_VAL]], double* [[TMP_G1_VAL]], 139 140 // LAMBDA: [[SVAR_PRIV_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[SVAR_PRIVATE]], 141 // LAMBDA: store i{{[0-9]+}} [[SVAR_PRIV_VAL]], i{{[0-9]+}}* [[SVAR_IN_REF]], 142 // LAMBDA: [[SFVAR_PRIV_VAL:%.+]] = load float, float* [[SFVAR_PRIVATE]], 143 // LAMBDA: store float [[SFVAR_PRIV_VAL]], float* [[SFVAR_IN_REF]], 144 // LAMBDA: br label %[[OMP_LASTPRIV_DONE]] 145 // LAMBDA: [[OMP_LASTPRIV_DONE]]: 146 // LAMBDA: ret 147 [&]() { 148 // LAMBDA: define {{.+}} void [[INNER_LAMBDA]](%{{.+}}* [[ARG_PTR:%.+]]) 149 // LAMBDA: store %{{.+}}* [[ARG_PTR]], %{{.+}}** [[ARG_PTR_REF:%.+]], 150 g = 2; 151 g1 = 2; 152 svar = 4; 153 sfvar = 8.0; 154 // LAMBDA: [[ARG_PTR:%.+]] = load %{{.+}}*, %{{.+}}** [[ARG_PTR_REF]] 155 // LAMBDA: [[G_PTR_REF:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* [[ARG_PTR]], i{{[0-9]+}} 0, i{{[0-9]+}} 0 156 // LAMBDA: [[G_REF:%.+]] = load double*, double** [[G_PTR_REF]] 157 // LAMBDA: store double 2.0{{.+}}, double* [[G_REF]] 158 159 // LAMBDA: [[TMP_PTR_REF:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* [[ARG_PTR]], i{{[0-9]+}} 0, i{{[0-9]+}} 1 160 // LAMBDA: [[G1_REF:%.+]] = load double*, double** [[TMP_PTR_REF]] 161 // LAMBDA: store double 2.0{{.+}}, double* [[G1_REF]], 162 // LAMBDA: [[SVAR_PTR_REF:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* [[ARG_PTR]], i{{[0-9]+}} 0, i{{[0-9]+}} 2 163 // LAMBDA: [[SVAR_REF:%.+]] = load i{{[0-9]+}}*, i{{[0-9]+}}** [[SVAR_PTR_REF]] 164 // LAMBDA: store i{{[0-9]+}} 4, i{{[0-9]+}}* [[SVAR_REF]] 165 // LAMBDA: [[SFVAR_PTR_REF:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* [[ARG_PTR]], i{{[0-9]+}} 0, i{{[0-9]+}} 3 166 // LAMBDA: [[SFVAR_REF:%.+]] = load float*, float** [[SFVAR_PTR_REF]] 167 // LAMBDA: store float 8.0{{.+}}, float* [[SFVAR_REF]] 168 }(); 169 } 170 }(); 171 return 0; 172 #else 173 S<float> test; 174 int t_var = 0; 175 int vec[] = {1, 2}; 176 S<float> s_arr[] = {1, 2}; 177 S<float> &var = test; 178 179 #pragma omp target 180 #pragma omp teams 181 #pragma omp distribute simd lastprivate(t_var, vec, s_arr, s_arr, var, var, svar) 182 for (int i = 0; i < 2; ++i) { 183 vec[i] = t_var; 184 s_arr[i] = var; 185 } 186 int i; 187 188 return tmain<int>(); 189 #endif 190 } 191 192 // CHECK: define{{.*}} i{{[0-9]+}} @main() 193 // CHECK: [[TEST:%.+]] = alloca [[S_FLOAT_TY]], 194 // CHECK: call {{.*}} [[S_FLOAT_TY_DEF_CONSTR:@.+]]([[S_FLOAT_TY]]* [[TEST]]) 195 // CHECK: call i{{[0-9]+}} @__tgt_target_teams( 196 // CHECK: call void [[OFFLOAD_FUN:@.+]](i{{[0-9]+}} {{.+}}, [2 x i{{[0-9]+}}]* {{.+}}, [2 x [[S_FLOAT_TY]]]* {{.+}}, [[S_FLOAT_TY]]* {{.+}}, i{{[0-9]+}} {{.+}}) 197 // CHECK: ret 198 199 // CHECK: define{{.+}} [[OFFLOAD_FUN]](i{{[0-9]+}} {{.+}}, [2 x i{{[0-9]+}}]*{{.+}} {{.+}}, [2 x [[S_FLOAT_TY]]]*{{.+}} {{.+}}, [[S_FLOAT_TY]]*{{.+}} {{.+}}, i{{[0-9]+}} {{.+}}) 200 // CHECK: call void (%{{.+}}*, i{{[0-9]+}}, void (i{{[0-9]+}}*, i{{[0-9]+}}*, ...)*, ...) @__kmpc_fork_teams( 201 // CHECK: ret 202 // 203 // CHECK: define internal void [[OMP_OUTLINED:@.+]](i{{[0-9]+}}* noalias [[GTID_ADDR:%.+]], i{{[0-9]+}}* noalias %{{.+}}, i{{[0-9]+}}*{{.+}} [[T_VAR_IN:%.+]], [2 x i{{[0-9]+}}]*{{.+}} [[VEC_IN:%.+]], [2 x [[S_FLOAT_TY]]]*{{.+}} [[S_ARR_IN:%.+]], [[S_FLOAT_TY]]*{{.+}} [[VAR_IN:%.+]], i{{[0-9]+}}*{{.*}} [[S_VAR_IN:%.+]]) 204 // CHECK: {{.+}} = alloca i{{[0-9]+}}*, 205 // CHECK: {{.+}} = alloca i{{[0-9]+}}*, 206 // CHECK: [[T_VAR_ADDR:%.+]] = alloca i{{[0-9]+}}*, 207 // CHECK: [[VEC_ADDR:%.+]] = alloca [2 x i{{[0-9]+}}]*, 208 // CHECK: [[S_ARR_ADDR:%.+]] = alloca [2 x [[S_FLOAT_TY]]]*, 209 // CHECK: [[VAR_ADDR:%.+]] = alloca [[S_FLOAT_TY]]*, 210 // CHECK: [[SVAR_ADDR:%.+]] = alloca i{{[0-9]+}}*, 211 // CHECK: [[TMP:%.*]] = alloca [[S_FLOAT_TY]]*, 212 // skip loop variables 213 // CHECK: {{.+}} = alloca i{{[0-9]+}}, 214 // CHECK: {{.+}} = alloca i{{[0-9]+}}, 215 // CHECK: {{.+}} = alloca i{{[0-9]+}}, 216 // CHECK: {{.+}} = alloca i{{[0-9]+}}, 217 // CHECK: {{.+}} = alloca i{{[0-9]+}}, 218 // CHECK: [[OMP_IS_LAST:%.+]] = alloca i{{[0-9]+}}, 219 // CHECK: [[T_VAR_PRIV:%.+]] = alloca i{{[0-9]+}}, 220 // CHECK: [[VEC_PRIV:%.+]] = alloca [2 x i{{[0-9]+}}], 221 // CHECK: [[S_ARR_PRIV:%.+]] = alloca [2 x [[S_FLOAT_TY]]], 222 // CHECK: [[VAR_PRIV:%.+]] = alloca [[S_FLOAT_TY]], 223 // CHECK: [[TMP_PRIV:%.+]] = alloca [[S_FLOAT_TY]]*, 224 // CHECK: [[S_VAR_PRIV:%.+]] = alloca i{{[0-9]+}}, 225 226 // copy from parameters to local address variables 227 // CHECK: store i{{[0-9]+}}* [[T_VAR_IN]], i{{[0-9]+}}** [[T_VAR_ADDR]], 228 // CHECK: store [2 x i{{[0-9]+}}]* [[VEC_IN]], [2 x i{{[0-9]+}}]** [[VEC_ADDR]], 229 // CHECK: store [2 x [[S_FLOAT_TY]]]* [[S_ARR_IN]], [2 x [[S_FLOAT_TY]]]** [[S_ARR_ADDR]], 230 // CHECK: store [[S_FLOAT_TY]]* [[VAR_IN]], [[S_FLOAT_TY]]** [[VAR_ADDR]], 231 // CHECK: store i{{[0-9]+}}* [[S_VAR_IN]], i{{[0-9]+}}** [[SVAR_ADDR]], 232 233 // load content of local address variables 234 // CHECK: [[T_VAR_ADDR_REF:%.+]] = load i{{[0-9]+}}*, i{{[0-9]+}}** [[T_VAR_ADDR]], 235 // CHECK: [[VEC_ADDR_REF:%.+]] = load [2 x i{{[0-9]+}}]*, [2 x i{{[0-9]+}}]** [[VEC_ADDR]], 236 // CHECK: [[S_ARR_ADDR_REF:%.+]] = load [2 x [[S_FLOAT_TY]]]*, [2 x [[S_FLOAT_TY]]]** [[S_ARR_ADDR]], 237 // CHECK: [[SVAR_ADDR_REF:%.+]] = load i{{[0-9]+}}*, i{{[0-9]+}}** [[SVAR_ADDR]], 238 // CHECK: [[VAR_ADDR_REF:%.+]] = load [[S_FLOAT_TY]]*, [[S_FLOAT_TY]]** [[VAR_ADDR]], 239 // CHECK: store [[S_FLOAT_TY]]* [[VAR_ADDR_REF]], [[S_FLOAT_TY]]** [[TMP]], 240 // CHECK: store i{{[0-9]+}} 0, i{{[0-9]+}}* [[OMP_IS_LAST]], 241 // CHECK: [[TMP_REF:%.+]] = load [[S_FLOAT_TY]]*, [[S_FLOAT_TY]]** [[TMP]], 242 // the distribute loop 243 // CHECK: call void @__kmpc_for_static_init_4( 244 // assignment: vec[i] = t_var; 245 // CHECK: [[T_VAR_PRIV_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[T_VAR_PRIV]], 246 // CHECK: [[VEC_PTR:%.+]] = getelementptr inbounds [2 x i{{[0-9]+}}], [2 x i{{[0-9]+}}]* [[VEC_PRIV]], i{{[0-9]+}} 0, i{{[0-9]+}} {{.+}} 247 // CHECK: store i{{[0-9]+}} [[T_VAR_PRIV_VAL]], i{{[0-9]+}}* [[VEC_PTR]], 248 249 // assignment: s_arr[i] = var; 250 // CHECK-DAG: [[S_ARR_PTR:%.+]] = getelementptr inbounds [2 x [[S_FLOAT_TY]]], [2 x [[S_FLOAT_TY]]]* [[S_ARR_PRIV]], 251 // CHECK-DAG: [[TMP_VAL:%.+]] = load [[S_FLOAT_TY]]*, [[S_FLOAT_TY]]** [[TMP_PRIV]], 252 // CHECK-DAG: [[S_ARR_PTR_BCAST:%.+]] = bitcast [[S_FLOAT_TY]]* [[S_ARR_PTR]] to i8* 253 // CHECK-DAG: [[TMP_VAL_BCAST:%.+]] = bitcast [[S_FLOAT_TY]]* [[TMP_VAL]] to i8* 254 // CHECK: call void @llvm.memcpy.{{.+}}(i8* [[S_ARR_PTR_BCAST]], i8* [[TMP_VAL_BCAST]], 255 // CHECK: call void @__kmpc_for_static_fini( 256 257 // lastprivates 258 // CHECK: [[OMP_IS_LAST_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[OMP_IS_LAST]], 259 // CHECK: [[IS_LAST_IT:%.+]] = icmp ne i{{[0-9]+}} [[OMP_IS_LAST_VAL]], 0 260 // CHECK: br i1 [[IS_LAST_IT]], label %[[OMP_LASTPRIV_BLOCK:.+]], label %[[OMP_LASTPRIV_DONE:.+]] 261 262 // CHECK: [[OMP_LASTPRIV_BLOCK]]: 263 // CHECK: [[T_VAR_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[T_VAR_PRIV]], 264 // CHECK: store i{{[0-9]+}} [[T_VAR_VAL]], i{{[0-9]+}}* [[T_VAR_ADDR_REF]], 265 // CHECK: [[BCAST_VEC_ADDR_REF:%.+]] = bitcast [2 x i{{[0-9]+}}]* [[VEC_ADDR_REF]] to i8* 266 // CHECK: [[BCAST_VEC_PRIV:%.+]] = bitcast [2 x i{{[0-9]+}}]* [[VEC_PRIV]] to i8* 267 // CHECK: call void @llvm.memcpy.{{.+}}(i8* [[BCAST_VEC_ADDR_REF]], i8* [[BCAST_VEC_PRIV]], 268 // CHECK: [[S_ARR_BEGIN:%.+]] = getelementptr inbounds [2 x [[S_FLOAT_TY]]], [2 x [[S_FLOAT_TY]]]* [[S_ARR_ADDR_REF]], i{{[0-9]+}} 0, i{{[0-9]+}} 0 269 // CHECK: [[S_ARR_PRIV_BCAST:%.+]] = bitcast [2 x [[S_FLOAT_TY]]]* [[S_ARR_PRIV]] to [[S_FLOAT_TY]]* 270 // CHECK: [[S_ARR_BEGIN_GEP:%.+]] = getelementptr [[S_FLOAT_TY]], [[S_FLOAT_TY]]* [[S_ARR_BEGIN]], i{{[0-9]+}} 2 271 // CHECK: [[S_ARR_IS_EMPTY:%.+]] = icmp eq [[S_FLOAT_TY]]* [[S_ARR_BEGIN]], [[S_ARR_BEGIN_GEP]] 272 // CHECK: br i1 [[S_ARR_IS_EMPTY]], label %[[S_ARR_COPY_DONE:.+]], label %[[S_ARR_COPY_BLOCK:.+]] 273 // CHECK: [[S_ARR_COPY_BLOCK]]: 274 // CHECK: [[S_ARR_SRC_EL:%.+]] = phi [[S_FLOAT_TY]]*{{.+}} 275 // CHECK: [[S_ARR_DST_EL:%.+]] = phi [[S_FLOAT_TY]]*{{.+}} 276 // CHECK: [[S_ARR_DST_BCAST:%.+]] = bitcast [[S_FLOAT_TY]]* [[S_ARR_DST_EL]] to i8* 277 // CHECK: [[S_ARR_SRC_BCAST:%.+]] = bitcast [[S_FLOAT_TY]]* [[S_ARR_SRC_EL]] to i8* 278 // CHECK: call void @llvm.memcpy.{{.+}}(i8* [[S_ARR_DST_BCAST]], i8* [[S_ARR_SRC_BCAST]]{{.+}}) 279 // CHECK: [[S_ARR_DST_NEXT:%.+]] = getelementptr [[S_FLOAT_TY]], [[S_FLOAT_TY]]* [[S_ARR_DST_EL]], i{{[0-9]+}} 1 280 // CHECK: [[S_ARR_SRC_NEXT:%.+]] = getelementptr{{.+}} 281 // CHECK: [[CPY_IS_FINISHED:%.+]] = icmp eq [[S_FLOAT_TY]]* [[S_ARR_DST_NEXT]], [[S_ARR_BEGIN_GEP]] 282 // CHECK: br i1 [[CPY_IS_FINISHED]], label %[[S_ARR_COPY_DONE]], label %[[S_ARR_COPY_BLOCK]] 283 // CHECK: [[S_ARR_COPY_DONE]]: 284 // CHECK: [[TMP_VAL1:%.+]] = load [[S_FLOAT_TY]]*, [[S_FLOAT_TY]]** [[TMP_PRIV]], 285 // CHECK: [[VAR_ADDR_REF_BCAST:%.+]] = bitcast [[S_FLOAT_TY]]* [[TMP_REF]] to i8* 286 // CHECK: [[TMP_VAL1_BCAST:%.+]] = bitcast [[S_FLOAT_TY]]* [[TMP_VAL1]] to i8* 287 // CHECK: call void @llvm.memcpy.{{.+}}(i8* [[VAR_ADDR_REF_BCAST]], i8* [[TMP_VAL1_BCAST]],{{.+}}) 288 // CHECK: [[SVAR_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[S_VAR_PRIV]], 289 // CHECK: store i{{[0-9]+}} [[SVAR_VAL]], i{{[0-9]+}}* [[SVAR_ADDR_REF]], 290 // CHECK: ret void 291 292 // template tmain 293 // CHECK: define{{.*}} i{{[0-9]+}} [[TMAIN_INT:@.+]]() 294 // CHECK: [[TEST:%.+]] = alloca [[S_INT_TY]], 295 // CHECK: call {{.*}} [[S_INT_TY_DEF_CONSTR:@.+]]([[S_INT_TY]]* [[TEST]]) 296 // CHECK: call i{{[0-9]+}} @__tgt_target_teams( 297 // CHECK: call void [[OFFLOAD_FUN_1:@.+]](i{{[0-9]+}} {{.+}}, [2 x i{{[0-9]+}}]* {{.+}}, [2 x [[S_INT_TY]]]* {{.+}}, [[S_INT_TY]]* {{.+}}) 298 // CHECK: ret 299 300 301 // CHECK: define internal void [[OFFLOAD_FUN_1]]( 302 // CHECK: call void (%{{.+}}*, i{{[0-9]+}}, void (i{{[0-9]+}}*, i{{[0-9]+}}*, ...)*, ...) @__kmpc_fork_teams(%{{.+}}* @{{.+}}, i{{[0-9]+}} 4, 303 // CHECK: ret 304 305 // CHECK: define internal void [[OMP_OUTLINED_1:@.+]](i{{[0-9]+}}* noalias [[GTID_ADDR1:%.+]], i{{[0-9]+}}* noalias %{{.+}}, i{{[0-9]+}}*{{.+}} [[T_VAR_IN1:%.+]], [2 x i{{[0-9]+}}]*{{.+}} [[VEC_IN1:%.+]], [2 x [[S_INT_TY]]]*{{.+}} [[S_ARR_IN1:%.+]], [[S_INT_TY]]*{{.+}} [[VAR_IN1:%.+]]) 306 // skip alloca of global_tid and bound_tid 307 // CHECK: {{.+}} = alloca i{{[0-9]+}}*, 308 // CHECK: {{.+}} = alloca i{{[0-9]+}}*, 309 // CHECK: [[T_VAR_ADDR1:%.+]] = alloca i{{[0-9]+}}*, 310 // CHECK: [[VEC_ADDR1:%.+]] = alloca [2 x i{{[0-9]+}}]*, 311 // CHECK: [[S_ARR_ADDR1:%.+]] = alloca [2 x [[S_INT_TY]]]*, 312 // CHECK: [[VAR_ADDR1:%.+]] = alloca [[S_INT_TY]]*, 313 // CHECK: [[TMP:%.+]] = alloca [[S_INT_TY]]*, 314 // skip loop variables 315 // CHECK: {{.+}} = alloca i{{[0-9]+}}, 316 // CHECK: {{.+}} = alloca i{{[0-9]+}}, 317 // CHECK: {{.+}} = alloca i{{[0-9]+}}, 318 // CHECK: {{.+}} = alloca i{{[0-9]+}}, 319 // CHECK: {{.+}} = alloca i{{[0-9]+}}, 320 // CHECK: [[OMP_IS_LAST1:%.+]] = alloca i{{[0-9]+}}, 321 // CHECK: [[T_VAR_PRIV1:%.+]] = alloca i{{[0-9]+}}, 322 // CHECK: [[VEC_PRIV1:%.+]] = alloca [2 x i{{[0-9]+}}], 323 // CHECK: [[S_ARR_PRIV1:%.+]] = alloca [2 x [[S_INT_TY]]], 324 // CHECK: [[VAR_PRIV1:%.+]] = alloca [[S_INT_TY]], 325 // CHECK: [[TMP_PRIV1:%.+]] = alloca [[S_INT_TY]]*, 326 327 // skip init of bound and global tid 328 // CHECK: store i{{[0-9]+}}* {{.*}}, 329 // CHECK: store i{{[0-9]+}}* {{.*}}, 330 // copy from parameters to local address variables 331 // CHECK: store i{{[0-9]+}}* [[T_VAR_IN1]], i{{[0-9]+}}** [[T_VAR_ADDR1]], 332 // CHECK: store [2 x i{{[0-9]+}}]* [[VEC_IN1]], [2 x i{{[0-9]+}}]** [[VEC_ADDR1]], 333 // CHECK: store [2 x [[S_INT_TY]]]* [[S_ARR_IN1]], [2 x [[S_INT_TY]]]** [[S_ARR_ADDR1]], 334 // CHECK: store [[S_INT_TY]]* [[VAR_IN1]], [[S_INT_TY]]** [[VAR_ADDR1]], 335 336 // load content of local address variables 337 // CHECK: [[T_VAR_ADDR_REF1:%.+]] = load i{{[0-9]+}}*, i{{[0-9]+}}** [[T_VAR_ADDR1]], 338 // CHECK: [[VEC_ADDR_REF1:%.+]] = load [2 x i{{[0-9]+}}]*, [2 x i{{[0-9]+}}]** [[VEC_ADDR1]], 339 // CHECK: [[S_ARR_ADDR_REF1:%.+]] = load [2 x [[S_INT_TY]]]*, [2 x [[S_INT_TY]]]** [[S_ARR_ADDR1]], 340 // CHECK: [[VAR_ADDR1_REF:%.+]] = load [[S_INT_TY]]*, [[S_INT_TY]]** [[VAR_ADDR1]], 341 // CHECK-DAG: store [[S_INT_TY]]* [[VAR_ADDR1_REF]], [[S_INT_TY]]** [[TMP]], 342 // CHECK-DAG: store i{{[0-9]+}} 0, i{{[0-9]+}}* [[OMP_IS_LAST1]], 343 // CHECK-DAG: [[TMP_REF:%.+]] = load [[S_INT_TY]]*, [[S_INT_TY]]** [[TMP]], 344 // CHECK: call void @__kmpc_for_static_init_4( 345 // assignment: vec[i] = t_var; 346 // CHECK: [[IV_VAL1:%.+]] = 347 // CHECK: [[T_VAR_PRIV_VAL1:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[T_VAR_PRIV1]], 348 // CHECK: [[VEC_PTR1:%.+]] = getelementptr inbounds [2 x i{{[0-9]+}}], [2 x i{{[0-9]+}}]* [[VEC_PRIV1]], i{{[0-9]+}} 0, i{{[0-9]+}} {{.+}} 349 // CHECK: store i{{[0-9]+}} [[T_VAR_PRIV_VAL1]], i{{[0-9]+}}* [[VEC_PTR1]], 350 351 // assignment: s_arr[i] = var; 352 // CHECK-DAG: [[S_ARR_PTR1:%.+]] = getelementptr inbounds [2 x [[S_INT_TY]]], [2 x [[S_INT_TY]]]* [[S_ARR_PRIV1]], 353 // CHECK-DAG: [[TMP_VAL1:%.+]] = load [[S_INT_TY]]*, [[S_INT_TY]]** [[TMP_PRIV1]], 354 // CHECK-DAG: [[S_ARR_PTR_BCAST1:%.+]] = bitcast [[S_INT_TY]]* [[S_ARR_PTR1]] to i8* 355 // CHECK-DAG: [[TMP_VAL_BCAST1:%.+]] = bitcast [[S_INT_TY]]* [[TMP_VAL1]] to i8* 356 // CHECK-DAG: call void @llvm.memcpy.{{.+}}(i8* [[S_ARR_PTR_BCAST1]], i8* [[TMP_VAL_BCAST1]], 357 // CHECK: call void @__kmpc_for_static_fini( 358 359 // lastprivates 360 // CHECK: [[OMP_IS_LAST_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[OMP_IS_LAST1]], 361 // CHECK: [[IS_LAST_IT:%.+]] = icmp ne i{{[0-9]+}} [[OMP_IS_LAST_VAL]], 0 362 // CHECK: br i1 [[IS_LAST_IT]], label %[[OMP_LASTPRIV_BLOCK:.+]], label %[[OMP_LASTPRIV_DONE:.+]] 363 364 // CHECK: [[OMP_LASTPRIV_BLOCK]]: 365 // CHECK: [[T_VAR_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[T_VAR_PRIV1]], 366 // CHECK: store i{{[0-9]+}} [[T_VAR_VAL]], i{{[0-9]+}}* [[T_VAR_ADDR_REF1]], 367 // CHECK: [[BCAST_VEC_ADDR_REF:%.+]] = bitcast [2 x i{{[0-9]+}}]* [[VEC_ADDR_REF1]] to i8* 368 // CHECK: [[BCAST_VEC_PRIV:%.+]] = bitcast [2 x i{{[0-9]+}}]* [[VEC_PRIV1]] to i8* 369 // CHECK: call void @llvm.memcpy.{{.+}}(i8* [[BCAST_VEC_ADDR_REF]], i8* [[BCAST_VEC_PRIV]], 370 // CHECK: [[S_ARR_BEGIN:%.+]] = getelementptr inbounds [2 x [[S_INT_TY]]], [2 x [[S_INT_TY]]]* [[S_ARR_ADDR_REF]], i{{[0-9]+}} 0, i{{[0-9]+}} 0 371 // CHECK: [[S_ARR_PRIV_BCAST:%.+]] = bitcast [2 x [[S_INT_TY]]]* [[S_ARR_PRIV1]] to [[S_INT_TY]]* 372 // CHECK: [[S_ARR_BEGIN_GEP:%.+]] = getelementptr [[S_INT_TY]], [[S_INT_TY]]* [[S_ARR_BEGIN]], i{{[0-9]+}} 2 373 // CHECK: [[S_ARR_IS_EMPTY:%.+]] = icmp eq [[S_INT_TY]]* [[S_ARR_BEGIN]], [[S_ARR_BEGIN_GEP]] 374 // CHECK: br i1 [[S_ARR_IS_EMPTY]], label %[[S_ARR_COPY_DONE:.+]], label %[[S_ARR_COPY_BLOCK:.+]] 375 // CHECK: [[S_ARR_COPY_BLOCK]]: 376 // CHECK: [[S_ARR_SRC_EL:%.+]] = phi [[S_INT_TY]]*{{.+}} 377 // CHECK: [[S_ARR_DST_EL:%.+]] = phi [[S_INT_TY]]*{{.+}} 378 // CHECK: [[S_ARR_DST_BCAST:%.+]] = bitcast [[S_INT_TY]]* [[S_ARR_DST_EL]] to i8* 379 // CHECK: [[S_ARR_SRC_BCAST:%.+]] = bitcast [[S_INT_TY]]* [[S_ARR_SRC_EL]] to i8* 380 // CHECK: call void @llvm.memcpy.{{.+}}(i8* [[S_ARR_DST_BCAST]], i8* [[S_ARR_SRC_BCAST]]{{.+}}) 381 // CHECK: [[S_ARR_DST_NEXT:%.+]] = getelementptr [[S_INT_TY]], [[S_INT_TY]]* [[S_ARR_DST_EL]], i{{[0-9]+}} 1 382 // CHECK: [[S_ARR_SRC_NEXT:%.+]] = getelementptr{{.+}} 383 // CHECK: [[CPY_IS_FINISHED:%.+]] = icmp eq [[S_INT_TY]]* [[S_ARR_DST_NEXT]], [[S_ARR_BEGIN_GEP]] 384 // CHECK: br i1 [[CPY_IS_FINISHED]], label %[[S_ARR_COPY_DONE]], label %[[S_ARR_COPY_BLOCK]] 385 // CHECK: [[S_ARR_COPY_DONE]]: 386 // CHECK: [[TMP_VAL1:%.+]] = load [[S_INT_TY]]*, [[S_INT_TY]]** [[TMP_PRIV1]], 387 // CHECK: [[VAR_ADDR_REF_BCAST:%.+]] = bitcast [[S_INT_TY]]* [[TMP_REF]] to i8* 388 // CHECK: [[TMP_VAL1_BCAST:%.+]] = bitcast [[S_INT_TY]]* [[TMP_VAL1]] to i8* 389 // CHECK: call void @llvm.memcpy.{{.+}}(i8* [[VAR_ADDR_REF_BCAST]], i8* [[TMP_VAL1_BCAST]],{{.+}}) 390 // CHECK: ret void 391 392 // CHECK: !{!"llvm.loop.vectorize.enable", i1 true} 393 #endif 394