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 distribute parallel for lastprivate(t_var, vec, s_arr, s_arr, var, var) 38 for (int i = 0; i < 2; ++i) { 39 vec[i] = t_var; 40 s_arr[i] = var; 41 } 42 return T(); 43 } 44 45 int main() { 46 static int svar; 47 volatile double g; 48 volatile double &g1 = g; 49 50 #ifdef LAMBDA 51 // LAMBDA-LABEL: @main 52 // LAMBDA: call{{.*}} void [[OUTER_LAMBDA:@.+]]( 53 [&]() { 54 static float sfvar; 55 // LAMBDA: define{{.*}} internal{{.*}} void [[OUTER_LAMBDA]]( 56 // LAMBDA: call i{{[0-9]+}} @__tgt_target_teams( 57 // LAMBDA: call void [[OFFLOADING_FUN:@.+]]( 58 59 // LAMBDA: define{{.+}} void [[OFFLOADING_FUN]]( 60 // LAMBDA: call {{.*}}void {{.+}} @__kmpc_fork_teams({{.+}}, i32 4, {{.+}}* [[OMP_OUTLINED:@.+]] to {{.+}}) 61 #pragma omp target 62 #pragma omp teams distribute parallel for lastprivate(g, g1, svar, sfvar) 63 for (int i = 0; i < 2; ++i) { 64 // LAMBDA: define{{.*}} internal{{.*}} void [[OMP_OUTLINED]](i32* noalias %{{.+}}, i32* noalias %{{.+}}, double*{{.+}} [[G_IN:%.+]], double*{{.+}} [[G1_IN:%.+]], i{{[0-9]+}}*{{.+}} [[SVAR_IN:%.+]], float*{{.+}} [[SFVAR_IN:%.+]]) 65 // LAMBDA: [[G_PRIVATE_ADDR:%.+]] = alloca double*, 66 // LAMBDA: [[G1_PRIVATE_ADDR:%.+]] = alloca double*, 67 // LAMBDA: [[SVAR_PRIVATE_ADDR:%.+]] = alloca i{{[0-9]+}}*, 68 // LAMBDA: [[SFVAR_PRIVATE_ADDR:%.+]] = alloca float*, 69 // loop variables 70 // LAMBDA: {{.+}} = alloca i{{[0-9]+}}, 71 // LAMBDA: {{.+}} = alloca i{{[0-9]+}}, 72 // LAMBDA: {{.+}} = alloca i{{[0-9]+}}, 73 // LAMBDA: {{.+}} = alloca i{{[0-9]+}}, 74 // LAMBDA: {{.+}} = alloca i{{[0-9]+}}, 75 // LAMBDA: [[OMP_IS_LAST:%.+]] = alloca i{{[0-9]+}}, 76 // LAMBDA: [[G_PRIVATE:%.+]] = alloca double, 77 // LAMBDA: [[G1_PRIVATE:%.+]] = alloca double, 78 // LAMBDA: [[TMP_G1_PRIVATE:%.+]] = alloca double*, 79 // LAMBDA: [[SVAR_PRIVATE:%.+]] = alloca i{{[0-9]+}}, 80 // LAMBDA: [[SFVAR_PRIVATE:%.+]] = alloca float, 81 // LAMBDA: store double* [[G_IN]], double** [[G_PRIVATE_ADDR]], 82 // LAMBDA: store double* [[G1_IN]], double** [[G1_PRIVATE_ADDR]], 83 // LAMBDA: store i{{[0-9]+}}* [[SVAR_IN]], i{{[0-9]+}}** [[SVAR_PRIVATE_ADDR]], 84 // LAMBDA: store float* [[SFVAR_IN]], float** [[SFVAR_PRIVATE_ADDR]], 85 86 // init private variables 87 // LAMBDA: [[G_IN_REF:%.+]] = load double*, double** [[G_PRIVATE_ADDR]], 88 // LAMBDA: [[SVAR_IN_REF:%.+]] = load i{{[0-9]+}}*, i{{[0-9]+}}** [[SVAR_PRIVATE_ADDR]], 89 // LAMBDA: [[SFVAR_IN_REF:%.+]] = load float*, float** [[SFVAR_PRIVATE_ADDR]], 90 // LAMBDA: [[G1_IN_REF:%.+]] = load double*, double** [[G1_PRIVATE_ADDR]], 91 // LAMBDA: store double* [[G1_PRIVATE]], double** [[TMP_G1_PRIVATE]], 92 g = 1; 93 g1 = 1; 94 svar = 3; 95 sfvar = 4.0; 96 // LAMBDA: call {{.*}}void @__kmpc_for_static_init_4( 97 // LAMBDA: call void {{.*}} @__kmpc_fork_call({{.+}}, {{.+}}, {{.+}} @[[LPAR_OUTL:.+]] to 98 // LAMBDA: call {{.*}}void @__kmpc_for_static_fini( 99 100 // LAMBDA: [[OMP_IS_LAST_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[OMP_IS_LAST]], 101 // LAMBDA: [[IS_LAST_IT:%.+]] = icmp ne i{{[0-9]+}} [[OMP_IS_LAST_VAL]], 0 102 // LAMBDA: br i1 [[IS_LAST_IT]], label %[[OMP_LASTPRIV_BLOCK:.+]], label %[[OMP_LASTPRIV_DONE:.+]] 103 104 // LAMBDA: [[OMP_LASTPRIV_BLOCK]]: 105 // LAMBDA: [[G_PRIV_VAL:%.+]] = load double, double* [[G_PRIVATE]], 106 // LAMBDA: store{{.*}} double [[G_PRIV_VAL]], double* [[G_IN_REF]], 107 // LAMBDA: [[TMP_G1_PRIV_REF:%.+]] = load double*, double** [[TMP_G1_PRIVATE]], 108 // LAMBDA: [[TMP_G1_PRIV_VAL:%.+]] = load double, double* [[TMP_G1_PRIV_REF]], 109 // LAMBDA: store{{.*}} double [[TMP_G1_PRIV_VAL]], double* [[G1_IN_REF]], 110 111 // LAMBDA: [[SVAR_PRIV_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[SVAR_PRIVATE]], 112 // LAMBDA: store i{{[0-9]+}} [[SVAR_PRIV_VAL]], i{{[0-9]+}}* [[SVAR_IN_REF]], 113 // LAMBDA: [[SFVAR_PRIV_VAL:%.+]] = load float, float* [[SFVAR_PRIVATE]], 114 // LAMBDA: store float [[SFVAR_PRIV_VAL]], float* [[SFVAR_IN_REF]], 115 // LAMBDA: br label %[[OMP_LASTPRIV_DONE]] 116 // LAMBDA: [[OMP_LASTPRIV_DONE]]: 117 // LAMBDA: ret 118 119 // LAMBDA: define{{.*}} internal{{.*}} void @[[LPAR_OUTL]](i32* noalias %{{.+}}, i32* noalias %{{.+}}, {{.+}}, {{.+}}, double*{{.+}} [[G_IN:%.+]], double*{{.+}} [[G1_IN:%.+]], i{{[0-9]+}}*{{.+}} [[SVAR_IN:%.+]], float*{{.+}} [[SFVAR_IN:%.+]]) 120 // LAMBDA: [[G_PRIVATE_ADDR:%.+]] = alloca double*, 121 // LAMBDA: [[G1_PRIVATE_ADDR:%.+]] = alloca double*, 122 // LAMBDA: [[SVAR_PRIVATE_ADDR:%.+]] = alloca i{{[0-9]+}}*, 123 // LAMBDA: [[SFVAR_PRIVATE_ADDR:%.+]] = alloca float*, 124 // loop variables 125 // LAMBDA: {{.+}} = alloca i{{[0-9]+}}, 126 // LAMBDA: {{.+}} = alloca i{{[0-9]+}}, 127 // LAMBDA: {{.+}} = alloca i{{[0-9]+}}, 128 // LAMBDA: {{.+}} = alloca i{{[0-9]+}}, 129 // LAMBDA: {{.+}} = alloca i{{[0-9]+}}, 130 // LAMBDA: [[OMP_IS_LAST:%.+]] = alloca i{{[0-9]+}}, 131 // LAMBDA: [[G_PRIVATE:%.+]] = alloca double, 132 // LAMBDA: [[G1_PRIVATE:%.+]] = alloca double, 133 // LAMBDA: [[TMP_G1_PRIVATE:%.+]] = alloca double*, 134 // LAMBDA: [[SVAR_PRIVATE:%.+]] = alloca i{{[0-9]+}}, 135 // LAMBDA: [[SFVAR_PRIVATE:%.+]] = alloca float, 136 // LAMBDA: store double* [[G_IN]], double** [[G_PRIVATE_ADDR]], 137 // LAMBDA: store double* [[G1_IN]], double** [[G1_PRIVATE_ADDR]], 138 // LAMBDA: store i{{[0-9]+}}* [[SVAR_IN]], i{{[0-9]+}}** [[SVAR_PRIVATE_ADDR]], 139 // LAMBDA: store float* [[SFVAR_IN]], float** [[SFVAR_PRIVATE_ADDR]], 140 141 // init private variables 142 // LAMBDA: [[G_IN_REF:%.+]] = load double*, double** [[G_PRIVATE_ADDR]], 143 // LAMBDA: [[SVAR_IN_REF:%.+]] = load i{{[0-9]+}}*, i{{[0-9]+}}** [[SVAR_PRIVATE_ADDR]], 144 // LAMBDA: [[SFVAR_IN_REF:%.+]] = load float*, float** [[SFVAR_PRIVATE_ADDR]], 145 // LAMBDA: [[G1_IN_REF:%.+]] = load double*, double** [[G1_PRIVATE_ADDR]], 146 // LAMBDA: store double* [[G1_PRIVATE]], double** [[TMP_G1_PRIVATE]], 147 148 // LAMBDA: call {{.*}}void @__kmpc_for_static_init_4( 149 // LAMBDA: store double 1.0{{.+}}, double* [[G_PRIVATE]], 150 // LAMBDA: [[TMP_G1_REF:%.+]] = load double*, double** [[TMP_G1_PRIVATE]], 151 // LAMBDA: store{{.+}} double 1.0{{.+}}, double* [[TMP_G1_REF]], 152 // LAMBDA: store i{{[0-9]+}} 3, i{{[0-9]+}}* [[SVAR_PRIVATE]], 153 // LAMBDA: store float 4.0{{.+}}, float* [[SFVAR_PRIVATE]], 154 // LAMBDA: [[G_PRIVATE_ADDR_REF:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* [[ARG:%.+]], i{{[0-9]+}} 0, i{{[0-9]+}} 0 155 // LAMBDA: store double* [[G_PRIVATE]], double** [[G_PRIVATE_ADDR_REF]], 156 // LAMBDA: [[TMP_PRIVATE_ADDR_REF:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* [[ARG:%.+]], i{{[0-9]+}} 0, i{{[0-9]+}} 1 157 // LAMBDA: [[G1_PRIVATE_ADDR_FROM_TMP:%.+]] = load double*, double** [[TMP_G1_PRIVATE]], 158 // LAMBDA: store double* [[G1_PRIVATE_ADDR_FROM_TMP]], double** [[TMP_PRIVATE_ADDR_REF]], 159 // LAMBDA: [[SVAR_PRIVATE_ADDR_REF:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* [[ARG:%.+]], i{{[0-9]+}} 0, i{{[0-9]+}} 2 160 // LAMBDA: store i{{[0-9]+}}* [[SVAR_PRIVATE]], i{{[0-9]+}}** [[SVAR_PRIVATE_ADDR_REF]] 161 // LAMBDA: [[SFVAR_PRIVATE_ADDR_REF:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* [[ARG:%.+]], i{{[0-9]+}} 0, i{{[0-9]+}} 3 162 // LAMBDA: store float* [[SFVAR_PRIVATE]], float** [[SFVAR_PRIVATE_ADDR_REF]] 163 // LAMBDA: call{{.*}} void [[INNER_LAMBDA:@.+]](%{{.+}}* [[ARG]]) 164 // LAMBDA: call {{.*}}void @__kmpc_for_static_fini( 165 166 // LAMBDA: [[OMP_IS_LAST_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[OMP_IS_LAST]], 167 // LAMBDA: [[IS_LAST_IT:%.+]] = icmp ne i{{[0-9]+}} [[OMP_IS_LAST_VAL]], 0 168 // LAMBDA: br i1 [[IS_LAST_IT]], label %[[OMP_LASTPRIV_BLOCK:.+]], label %[[OMP_LASTPRIV_DONE:.+]] 169 170 // LAMBDA: [[OMP_LASTPRIV_BLOCK]]: 171 // LAMBDA: [[G_PRIV_VAL:%.+]] = load double, double* [[G_PRIVATE]], 172 // LAMBDA: store{{.*}} double [[G_PRIV_VAL]], double* [[G_IN_REF]], 173 // LAMBDA: [[TMP_G1_PRIV_REF:%.+]] = load double*, double** [[TMP_G1_PRIVATE]], 174 // LAMBDA: [[TMP_G1_PRIV_VAL:%.+]] = load double, double* [[TMP_G1_PRIV_REF]], 175 // LAMBDA: store{{.*}} double [[TMP_G1_PRIV_VAL]], double* [[G1_IN_REF]], 176 177 // LAMBDA: [[SVAR_PRIV_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[SVAR_PRIVATE]], 178 // LAMBDA: store i{{[0-9]+}} [[SVAR_PRIV_VAL]], i{{[0-9]+}}* [[SVAR_IN_REF]], 179 // LAMBDA: [[SFVAR_PRIV_VAL:%.+]] = load float, float* [[SFVAR_PRIVATE]], 180 // LAMBDA: store float [[SFVAR_PRIV_VAL]], float* [[SFVAR_IN_REF]], 181 // LAMBDA: br label %[[OMP_LASTPRIV_DONE]] 182 // LAMBDA: [[OMP_LASTPRIV_DONE]]: 183 // LAMBDA: ret 184 185 [&]() { 186 // LAMBDA: define {{.+}} void [[INNER_LAMBDA]](%{{.+}}* [[ARG_PTR:%.+]]) 187 // LAMBDA: store %{{.+}}* [[ARG_PTR]], %{{.+}}** [[ARG_PTR_REF:%.+]], 188 g = 2; 189 g1 = 2; 190 svar = 4; 191 sfvar = 8.0; 192 // LAMBDA: [[ARG_PTR:%.+]] = load %{{.+}}*, %{{.+}}** [[ARG_PTR_REF]] 193 // LAMBDA: [[G_PTR_REF:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* [[ARG_PTR]], i{{[0-9]+}} 0, i{{[0-9]+}} 0 194 // LAMBDA: [[G_REF:%.+]] = load double*, double** [[G_PTR_REF]] 195 // LAMBDA: store double 2.0{{.+}}, double* [[G_REF]] 196 197 // LAMBDA: [[TMP_PTR_REF:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* [[ARG_PTR]], i{{[0-9]+}} 0, i{{[0-9]+}} 1 198 // LAMBDA: [[G1_REF:%.+]] = load double*, double** [[TMP_PTR_REF]] 199 // LAMBDA: store double 2.0{{.+}}, double* [[G1_REF]], 200 // LAMBDA: [[SVAR_PTR_REF:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* [[ARG_PTR]], i{{[0-9]+}} 0, i{{[0-9]+}} 2 201 // LAMBDA: [[SVAR_REF:%.+]] = load i{{[0-9]+}}*, i{{[0-9]+}}** [[SVAR_PTR_REF]] 202 // LAMBDA: store i{{[0-9]+}} 4, i{{[0-9]+}}* [[SVAR_REF]] 203 // LAMBDA: [[SFVAR_PTR_REF:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* [[ARG_PTR]], i{{[0-9]+}} 0, i{{[0-9]+}} 3 204 // LAMBDA: [[SFVAR_REF:%.+]] = load float*, float** [[SFVAR_PTR_REF]] 205 // LAMBDA: store float 8.0{{.+}}, float* [[SFVAR_REF]] 206 }(); 207 } 208 }(); 209 return 0; 210 #else 211 S<float> test; 212 int t_var = 0; 213 int vec[] = {1, 2}; 214 S<float> s_arr[] = {1, 2}; 215 S<float> &var = test; 216 217 #pragma omp target 218 #pragma omp teams distribute parallel for lastprivate(t_var, vec, s_arr, s_arr, var, var, svar) 219 for (int i = 0; i < 2; ++i) { 220 vec[i] = t_var; 221 s_arr[i] = var; 222 } 223 int i; 224 225 return tmain<int>(); 226 #endif 227 } 228 229 // CHECK: define{{.*}} i{{[0-9]+}} @main() 230 // CHECK: [[TEST:%.+]] = alloca [[S_FLOAT_TY]], 231 // CHECK: call {{.*}} [[S_FLOAT_TY_DEF_CONSTR:@.+]]([[S_FLOAT_TY]]* [[TEST]]) 232 // CHECK: call i{{[0-9]+}} @__tgt_target_teams( 233 // CHECK: call void [[OFFLOAD_FUN:@.+]](i{{[0-9]+}} {{.+}}, [2 x i{{[0-9]+}}]* {{.+}}, [2 x [[S_FLOAT_TY]]]* {{.+}}, [[S_FLOAT_TY]]* {{.+}}, i{{[0-9]+}} {{.+}}) 234 // CHECK: ret 235 236 // CHECK: define{{.+}} [[OFFLOAD_FUN]](i{{[0-9]+}} {{.+}}, [2 x i{{[0-9]+}}]*{{.+}} {{.+}}, [2 x [[S_FLOAT_TY]]]*{{.+}} {{.+}}, [[S_FLOAT_TY]]*{{.+}} {{.+}}, i{{[0-9]+}} {{.+}}) 237 // CHECK: call void (%{{.+}}*, i{{[0-9]+}}, void (i{{[0-9]+}}*, i{{[0-9]+}}*, ...)*, ...) @__kmpc_fork_teams( 238 // CHECK: ret 239 // 240 // CHECK: define internal void [[OMP_OUTLINED:@.+]](i{{[0-9]+}}* noalias [[GTID_ADDR:%.+]], i{{[0-9]+}}* noalias %{{.+}}, [2 x i{{[0-9]+}}]*{{.+}} [[VEC_IN:%.+]], i{{[0-9]+}}*{{.+}} [[T_VAR_IN:%.+]], [2 x [[S_FLOAT_TY]]]*{{.+}} [[S_ARR_IN:%.+]], [[S_FLOAT_TY]]*{{.+}} [[VAR_IN:%.+]], i{{[0-9]+}}*{{.*}} [[S_VAR_IN:%.+]]) 241 // CHECK: {{.+}} = alloca i{{[0-9]+}}*, 242 // CHECK: {{.+}} = alloca i{{[0-9]+}}*, 243 // CHECK: [[VEC_ADDR:%.+]] = alloca [2 x i{{[0-9]+}}]*, 244 // CHECK: [[T_VAR_ADDR:%.+]] = alloca i{{[0-9]+}}*, 245 // CHECK: [[S_ARR_ADDR:%.+]] = alloca [2 x [[S_FLOAT_TY]]]*, 246 // CHECK: [[VAR_ADDR:%.+]] = alloca [[S_FLOAT_TY]]*, 247 // CHECK: [[SVAR_ADDR:%.+]] = alloca i{{[0-9]+}}*, 248 // skip loop variables 249 // CHECK: {{.+}} = alloca i{{[0-9]+}}, 250 // CHECK: {{.+}} = alloca i{{[0-9]+}}, 251 // CHECK: {{.+}} = alloca i{{[0-9]+}}, 252 // CHECK: {{.+}} = alloca i{{[0-9]+}}, 253 // CHECK: {{.+}} = alloca i{{[0-9]+}}, 254 // CHECK: [[OMP_IS_LAST:%.+]] = alloca i{{[0-9]+}}, 255 // CHECK: [[T_VAR_PRIV:%.+]] = alloca i{{[0-9]+}}, 256 // CHECK: [[VEC_PRIV:%.+]] = alloca [2 x i{{[0-9]+}}], 257 // CHECK: [[S_ARR_PRIV:%.+]] = alloca [2 x [[S_FLOAT_TY]]], 258 // CHECK: [[VAR_PRIV:%.+]] = alloca [[S_FLOAT_TY]], 259 // CHECK: [[TMP_PRIV:%.+]] = alloca [[S_FLOAT_TY]]*, 260 // CHECK: [[S_VAR_PRIV:%.+]] = alloca i{{[0-9]+}}, 261 262 // copy from parameters to local address variables 263 // CHECK: store [2 x i{{[0-9]+}}]* [[VEC_IN]], [2 x i{{[0-9]+}}]** [[VEC_ADDR]], 264 // CHECK: store i{{[0-9]+}}* [[T_VAR_IN]], i{{[0-9]+}}** [[T_VAR_ADDR]], 265 // CHECK: store [2 x [[S_FLOAT_TY]]]* [[S_ARR_IN]], [2 x [[S_FLOAT_TY]]]** [[S_ARR_ADDR]], 266 // CHECK: store [[S_FLOAT_TY]]* [[VAR_IN]], [[S_FLOAT_TY]]** [[VAR_ADDR]], 267 // CHECK: store i{{[0-9]+}}* [[S_VAR_IN]], i{{[0-9]+}}** [[SVAR_ADDR]], 268 269 // load content of local address variables 270 // CHECK: [[VEC_ADDR_REF:%.+]] = load [2 x i{{[0-9]+}}]*, [2 x i{{[0-9]+}}]** [[VEC_ADDR]], 271 // CHECK: [[T_VAR_ADDR_REF:%.+]] = load i{{[0-9]+}}*, i{{[0-9]+}}** [[T_VAR_ADDR]], 272 // CHECK: [[S_ARR_ADDR_REF:%.+]] = load [2 x [[S_FLOAT_TY]]]*, [2 x [[S_FLOAT_TY]]]** [[S_ARR_ADDR]], 273 // CHECK: [[SVAR_ADDR_REF:%.+]] = load i{{[0-9]+}}*, i{{[0-9]+}}** [[SVAR_ADDR]], 274 // CHECK: store i{{[0-9]+}} 0, i{{[0-9]+}}* [[OMP_IS_LAST]], 275 // CHECK: [[VAR_ADDR_REF:%.+]] = load {{.+}}, {{.+}} [[VAR_ADDR]], 276 // the distribute loop 277 // CHECK: call void @__kmpc_for_static_init_4( 278 // CHECK: call void {{.*}} @__kmpc_fork_call({{.+}}, {{.+}}, {{.+}} @[[PAR_OUTL:.+]] to 279 // CHECK: call void @__kmpc_for_static_fini( 280 281 // lastprivates 282 // CHECK: [[OMP_IS_LAST_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[OMP_IS_LAST]], 283 // CHECK: [[IS_LAST_IT:%.+]] = icmp ne i{{[0-9]+}} [[OMP_IS_LAST_VAL]], 0 284 // CHECK: br i1 [[IS_LAST_IT]], label %[[OMP_LASTPRIV_BLOCK:.+]], label %[[OMP_LASTPRIV_DONE:.+]] 285 286 // CHECK: [[OMP_LASTPRIV_BLOCK]]: 287 // CHECK: [[T_VAR_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[T_VAR_PRIV]], 288 // CHECK: store i{{[0-9]+}} [[T_VAR_VAL]], i{{[0-9]+}}* [[T_VAR_ADDR_REF]], 289 // CHECK: [[BCAST_VEC_ADDR_REF:%.+]] = bitcast [2 x i{{[0-9]+}}]* [[VEC_ADDR_REF]] to i8* 290 // CHECK: [[BCAST_VEC_PRIV:%.+]] = bitcast [2 x i{{[0-9]+}}]* [[VEC_PRIV]] to i8* 291 // CHECK: call void @llvm.memcpy.{{.+}}(i8* [[BCAST_VEC_ADDR_REF]], i8* [[BCAST_VEC_PRIV]], 292 // 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]+}} 0 293 // CHECK: [[S_ARR_PRIV_BCAST:%.+]] = bitcast [2 x [[S_FLOAT_TY]]]* [[S_ARR_PRIV]] to [[S_FLOAT_TY]]* 294 // CHECK: [[S_ARR_BEGIN_GEP:%.+]] = getelementptr [[S_FLOAT_TY]], [[S_FLOAT_TY]]* [[S_ARR_BEGIN]], i{{[0-9]+}} 2 295 // CHECK: [[S_ARR_IS_EMPTY:%.+]] = icmp eq [[S_FLOAT_TY]]* [[S_ARR_BEGIN]], [[S_ARR_BEGIN_GEP]] 296 // CHECK: br i1 [[S_ARR_IS_EMPTY]], label %[[S_ARR_COPY_DONE:.+]], label %[[S_ARR_COPY_BLOCK:.+]] 297 // CHECK: [[S_ARR_COPY_BLOCK]]: 298 // CHECK: [[S_ARR_SRC_EL:%.+]] = phi [[S_FLOAT_TY]]*{{.+}} 299 // CHECK: [[S_ARR_DST_EL:%.+]] = phi [[S_FLOAT_TY]]*{{.+}} 300 // CHECK: [[S_ARR_DST_BCAST:%.+]] = bitcast [[S_FLOAT_TY]]* [[S_ARR_DST_EL]] to i8* 301 // CHECK: [[S_ARR_SRC_BCAST:%.+]] = bitcast [[S_FLOAT_TY]]* [[S_ARR_SRC_EL]] to i8* 302 // CHECK: call void @llvm.memcpy.{{.+}}(i8* [[S_ARR_DST_BCAST]], i8* [[S_ARR_SRC_BCAST]]{{.+}}) 303 // CHECK: [[S_ARR_DST_NEXT:%.+]] = getelementptr [[S_FLOAT_TY]], [[S_FLOAT_TY]]* [[S_ARR_DST_EL]], i{{[0-9]+}} 1 304 // CHECK: [[S_ARR_SRC_NEXT:%.+]] = getelementptr{{.+}} 305 // CHECK: [[CPY_IS_FINISHED:%.+]] = icmp eq [[S_FLOAT_TY]]* [[S_ARR_DST_NEXT]], [[S_ARR_BEGIN_GEP]] 306 // CHECK: br i1 [[CPY_IS_FINISHED]], label %[[S_ARR_COPY_DONE]], label %[[S_ARR_COPY_BLOCK]] 307 // CHECK: [[S_ARR_COPY_DONE]]: 308 // CHECK: [[TMP_VAL1:%.+]] = load [[S_FLOAT_TY]]*, [[S_FLOAT_TY]]** [[TMP_PRIV]], 309 // CHECK: [[VAR_ADDR_REF_BCAST:%.+]] = bitcast [[S_FLOAT_TY]]* [[VAR_ADDR_REF]] to i8* 310 // CHECK: [[TMP_VAL1_BCAST:%.+]] = bitcast [[S_FLOAT_TY]]* [[TMP_VAL1]] to i8* 311 // CHECK: call void @llvm.memcpy.{{.+}}(i8* [[VAR_ADDR_REF_BCAST]], i8* [[TMP_VAL1_BCAST]],{{.+}}) 312 // CHECK: [[SVAR_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[S_VAR_PRIV]], 313 // CHECK: store i{{[0-9]+}} [[SVAR_VAL]], i{{[0-9]+}}* [[SVAR_ADDR_REF]], 314 // CHECK: ret void 315 316 // CHECK: define internal void [[OMP_OUTLINED:@.+]](i{{[0-9]+}}* noalias [[GTID_ADDR:%.+]], i{{[0-9]+}}* noalias %{{.+}}, {{.+}}, {{.+}}, [2 x i{{[0-9]+}}]*{{.+}} [[VEC_IN:%.+]], i{{[0-9]+}}*{{.+}} [[T_VAR_IN:%.+]], [2 x [[S_FLOAT_TY]]]*{{.+}} [[S_ARR_IN:%.+]], [[S_FLOAT_TY]]*{{.+}} [[VAR_IN:%.+]], i{{[0-9]+}}*{{.*}} [[S_VAR_IN:%.+]]) 317 // gbl and bound tid vars, prev lb and ub vars 318 // CHECK: {{.+}} = alloca i{{[0-9]+}}*, 319 // CHECK: {{.+}} = alloca i{{[0-9]+}}*, 320 // CHECK: alloca i{{[0-9]+}}, 321 // CHECK: alloca i{{[0-9]+}}, 322 // CHECK: [[VEC_ADDR:%.+]] = alloca [2 x i{{[0-9]+}}]*, 323 // CHECK: [[T_VAR_ADDR:%.+]] = alloca i{{[0-9]+}}*, 324 // CHECK: [[S_ARR_ADDR:%.+]] = alloca [2 x [[S_FLOAT_TY]]]*, 325 // CHECK: [[VAR_ADDR:%.+]] = alloca [[S_FLOAT_TY]]*, 326 // CHECK: [[SVAR_ADDR:%.+]] = alloca i{{[0-9]+}}*, 327 // skip loop variables 328 // CHECK: {{.+}} = alloca i{{[0-9]+}}, 329 // CHECK: {{.+}} = alloca i{{[0-9]+}}, 330 // CHECK: {{.+}} = alloca i{{[0-9]+}}, 331 // CHECK: {{.+}} = alloca i{{[0-9]+}}, 332 // CHECK: {{.+}} = alloca i{{[0-9]+}}, 333 // CHECK: [[OMP_IS_LAST:%.+]] = alloca i{{[0-9]+}}, 334 // CHECK: [[T_VAR_PRIV:%.+]] = alloca i{{[0-9]+}}, 335 // CHECK: [[VEC_PRIV:%.+]] = alloca [2 x i{{[0-9]+}}], 336 // CHECK: [[S_ARR_PRIV:%.+]] = alloca [2 x [[S_FLOAT_TY]]], 337 // CHECK: [[VAR_PRIV:%.+]] = alloca [[S_FLOAT_TY]], 338 // CHECK: [[TMP_PRIV:%.+]] = alloca [[S_FLOAT_TY]]*, 339 // CHECK: [[S_VAR_PRIV:%.+]] = alloca i{{[0-9]+}}, 340 341 // copy from parameters to local address variables 342 // CHECK: store [2 x i{{[0-9]+}}]* [[VEC_IN]], [2 x i{{[0-9]+}}]** [[VEC_ADDR]], 343 // CHECK: store i{{[0-9]+}}* [[T_VAR_IN]], i{{[0-9]+}}** [[T_VAR_ADDR]], 344 // CHECK: store [2 x [[S_FLOAT_TY]]]* [[S_ARR_IN]], [2 x [[S_FLOAT_TY]]]** [[S_ARR_ADDR]], 345 // CHECK: store [[S_FLOAT_TY]]* [[VAR_IN]], [[S_FLOAT_TY]]** [[VAR_ADDR]], 346 // CHECK: store i{{[0-9]+}}* [[S_VAR_IN]], i{{[0-9]+}}** [[SVAR_ADDR]], 347 348 // load content of local address variables 349 // CHECK: [[VEC_ADDR_REF:%.+]] = load [2 x i{{[0-9]+}}]*, [2 x i{{[0-9]+}}]** [[VEC_ADDR]], 350 // CHECK: [[T_VAR_ADDR_REF:%.+]] = load i{{[0-9]+}}*, i{{[0-9]+}}** [[T_VAR_ADDR]], 351 // CHECK: [[S_ARR_ADDR_REF:%.+]] = load [2 x [[S_FLOAT_TY]]]*, [2 x [[S_FLOAT_TY]]]** [[S_ARR_ADDR]], 352 // CHECK: [[SVAR_ADDR_REF:%.+]] = load i{{[0-9]+}}*, i{{[0-9]+}}** [[SVAR_ADDR]], 353 // CHECK: store i{{[0-9]+}} 0, i{{[0-9]+}}* [[OMP_IS_LAST]], 354 // CHECK: [[VAR_ADDR_REF:%.+]] = load {{.+}}, {{.+}} [[VAR_ADDR]], 355 // the distribute loop 356 // CHECK: call void @__kmpc_for_static_init_4( 357 358 // assignment: vec[i] = t_var; 359 // CHECK: [[T_VAR_PRIV_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[T_VAR_PRIV]], 360 // CHECK: [[VEC_PTR:%.+]] = getelementptr inbounds [2 x i{{[0-9]+}}], [2 x i{{[0-9]+}}]* [[VEC_PRIV]], i{{[0-9]+}} 0, i{{[0-9]+}} {{.+}} 361 // CHECK: store i{{[0-9]+}} [[T_VAR_PRIV_VAL]], i{{[0-9]+}}* [[VEC_PTR]], 362 363 // assignment: s_arr[i] = var; 364 // CHECK-DAG: [[S_ARR_PTR:%.+]] = getelementptr inbounds [2 x [[S_FLOAT_TY]]], [2 x [[S_FLOAT_TY]]]* [[S_ARR_PRIV]], 365 // CHECK-DAG: [[TMP_VAL:%.+]] = load [[S_FLOAT_TY]]*, [[S_FLOAT_TY]]** [[TMP_PRIV]], 366 // CHECK-DAG: [[S_ARR_PTR_BCAST:%.+]] = bitcast [[S_FLOAT_TY]]* [[S_ARR_PTR]] to i8* 367 // CHECK-DAG: [[TMP_VAL_BCAST:%.+]] = bitcast [[S_FLOAT_TY]]* [[TMP_VAL]] to i8* 368 // CHECK: call void @llvm.memcpy.{{.+}}(i8* [[S_ARR_PTR_BCAST]], i8* [[TMP_VAL_BCAST]], 369 370 // CHECK: call void @__kmpc_for_static_fini( 371 372 // lastprivates 373 // CHECK: [[OMP_IS_LAST_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[OMP_IS_LAST]], 374 // CHECK: [[IS_LAST_IT:%.+]] = icmp ne i{{[0-9]+}} [[OMP_IS_LAST_VAL]], 0 375 // CHECK: br i1 [[IS_LAST_IT]], label %[[OMP_LASTPRIV_BLOCK:.+]], label %[[OMP_LASTPRIV_DONE:.+]] 376 377 // CHECK: [[OMP_LASTPRIV_BLOCK]]: 378 // CHECK: [[T_VAR_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[T_VAR_PRIV]], 379 // CHECK: store i{{[0-9]+}} [[T_VAR_VAL]], i{{[0-9]+}}* [[T_VAR_ADDR_REF]], 380 // CHECK: [[BCAST_VEC_ADDR_REF:%.+]] = bitcast [2 x i{{[0-9]+}}]* [[VEC_ADDR_REF]] to i8* 381 // CHECK: [[BCAST_VEC_PRIV:%.+]] = bitcast [2 x i{{[0-9]+}}]* [[VEC_PRIV]] to i8* 382 // CHECK: call void @llvm.memcpy.{{.+}}(i8* [[BCAST_VEC_ADDR_REF]], i8* [[BCAST_VEC_PRIV]], 383 // 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]+}} 0 384 // CHECK: [[S_ARR_PRIV_BCAST:%.+]] = bitcast [2 x [[S_FLOAT_TY]]]* [[S_ARR_PRIV]] to [[S_FLOAT_TY]]* 385 // CHECK: [[S_ARR_BEGIN_GEP:%.+]] = getelementptr [[S_FLOAT_TY]], [[S_FLOAT_TY]]* [[S_ARR_BEGIN]], i{{[0-9]+}} 2 386 // CHECK: [[S_ARR_IS_EMPTY:%.+]] = icmp eq [[S_FLOAT_TY]]* [[S_ARR_BEGIN]], [[S_ARR_BEGIN_GEP]] 387 // CHECK: br i1 [[S_ARR_IS_EMPTY]], label %[[S_ARR_COPY_DONE:.+]], label %[[S_ARR_COPY_BLOCK:.+]] 388 // CHECK: [[S_ARR_COPY_BLOCK]]: 389 // CHECK: [[S_ARR_SRC_EL:%.+]] = phi [[S_FLOAT_TY]]*{{.+}} 390 // CHECK: [[S_ARR_DST_EL:%.+]] = phi [[S_FLOAT_TY]]*{{.+}} 391 // CHECK: [[S_ARR_DST_BCAST:%.+]] = bitcast [[S_FLOAT_TY]]* [[S_ARR_DST_EL]] to i8* 392 // CHECK: [[S_ARR_SRC_BCAST:%.+]] = bitcast [[S_FLOAT_TY]]* [[S_ARR_SRC_EL]] to i8* 393 // CHECK: call void @llvm.memcpy.{{.+}}(i8* [[S_ARR_DST_BCAST]], i8* [[S_ARR_SRC_BCAST]]{{.+}}) 394 // CHECK: [[S_ARR_DST_NEXT:%.+]] = getelementptr [[S_FLOAT_TY]], [[S_FLOAT_TY]]* [[S_ARR_DST_EL]], i{{[0-9]+}} 1 395 // CHECK: [[S_ARR_SRC_NEXT:%.+]] = getelementptr{{.+}} 396 // CHECK: [[CPY_IS_FINISHED:%.+]] = icmp eq [[S_FLOAT_TY]]* [[S_ARR_DST_NEXT]], [[S_ARR_BEGIN_GEP]] 397 // CHECK: br i1 [[CPY_IS_FINISHED]], label %[[S_ARR_COPY_DONE]], label %[[S_ARR_COPY_BLOCK]] 398 // CHECK: [[S_ARR_COPY_DONE]]: 399 // CHECK: [[TMP_VAL1:%.+]] = load [[S_FLOAT_TY]]*, [[S_FLOAT_TY]]** [[TMP_PRIV]], 400 // CHECK: [[VAR_ADDR_REF_BCAST:%.+]] = bitcast [[S_FLOAT_TY]]* [[VAR_ADDR_REF]] to i8* 401 // CHECK: [[TMP_VAL1_BCAST:%.+]] = bitcast [[S_FLOAT_TY]]* [[TMP_VAL1]] to i8* 402 // CHECK: call void @llvm.memcpy.{{.+}}(i8* [[VAR_ADDR_REF_BCAST]], i8* [[TMP_VAL1_BCAST]],{{.+}}) 403 // CHECK: [[SVAR_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[S_VAR_PRIV]], 404 // CHECK: store i{{[0-9]+}} [[SVAR_VAL]], i{{[0-9]+}}* [[SVAR_ADDR_REF]], 405 // CHECK: ret void 406 407 // template tmain 408 // CHECK: define{{.*}} i{{[0-9]+}} [[TMAIN_INT:@.+]]() 409 // CHECK: [[TEST:%.+]] = alloca [[S_INT_TY]], 410 // CHECK: call {{.*}} [[S_INT_TY_DEF_CONSTR:@.+]]([[S_INT_TY]]* [[TEST]]) 411 // CHECK: call i{{[0-9]+}} @__tgt_target_teams( 412 // CHECK: call void [[OFFLOAD_FUN_1:@.+]](i{{[0-9]+}} {{.+}}, [2 x i{{[0-9]+}}]* {{.+}}, [2 x [[S_INT_TY]]]* {{.+}}, [[S_INT_TY]]* {{.+}}) 413 // CHECK: ret 414 415 // CHECK: define internal void [[OFFLOAD_FUN_1]]( 416 // CHECK: call void (%{{.+}}*, i{{[0-9]+}}, void (i{{[0-9]+}}*, i{{[0-9]+}}*, ...)*, ...) @__kmpc_fork_teams(%{{.+}}* @{{.+}}, i{{[0-9]+}} 4, 417 // CHECK: ret 418 419 // CHECK: define internal void [[OMP_OUTLINED_1:@.+]](i{{[0-9]+}}* noalias [[GTID_ADDR1:%.+]], i{{[0-9]+}}* noalias %{{.+}}, [2 x i{{[0-9]+}}]*{{.+}} [[VEC_IN1:%.+]], i{{[0-9]+}}*{{.+}} [[T_VAR_IN1:%.+]], [2 x [[S_INT_TY]]]*{{.+}} [[S_ARR_IN1:%.+]], [[S_INT_TY]]*{{.+}} [[VAR_IN1:%.+]]) 420 // skip alloca of global_tid and bound_tid 421 // CHECK: {{.+}} = alloca i{{[0-9]+}}*, 422 // CHECK: {{.+}} = alloca i{{[0-9]+}}*, 423 // CHECK: [[VEC_ADDR1:%.+]] = alloca [2 x i{{[0-9]+}}]*, 424 // CHECK: [[T_VAR_ADDR1:%.+]] = alloca i{{[0-9]+}}*, 425 // CHECK: [[S_ARR_ADDR1:%.+]] = alloca [2 x [[S_INT_TY]]]*, 426 // CHECK: [[VAR_ADDR1:%.+]] = alloca [[S_INT_TY]]*, 427 // skip loop variables 428 // CHECK: {{.+}} = alloca i{{[0-9]+}}, 429 // CHECK: {{.+}} = alloca i{{[0-9]+}}, 430 // CHECK: {{.+}} = alloca i{{[0-9]+}}, 431 // CHECK: {{.+}} = alloca i{{[0-9]+}}, 432 // CHECK: {{.+}} = alloca i{{[0-9]+}}, 433 // CHECK: [[OMP_IS_LAST1:%.+]] = alloca i{{[0-9]+}}, 434 // CHECK: [[T_VAR_PRIV1:%.+]] = alloca i{{[0-9]+}}, 435 // CHECK: [[VEC_PRIV1:%.+]] = alloca [2 x i{{[0-9]+}}], 436 // CHECK: [[S_ARR_PRIV1:%.+]] = alloca [2 x [[S_INT_TY]]], 437 // CHECK: [[VAR_PRIV1:%.+]] = alloca [[S_INT_TY]], 438 // CHECK: [[TMP_PRIV1:%.+]] = alloca [[S_INT_TY]]*, 439 440 // skip init of bound and global tid 441 // CHECK: store i{{[0-9]+}}* {{.*}}, 442 // CHECK: store i{{[0-9]+}}* {{.*}}, 443 // copy from parameters to local address variables 444 // CHECK: store [2 x i{{[0-9]+}}]* [[VEC_IN1]], [2 x i{{[0-9]+}}]** [[VEC_ADDR1]], 445 // CHECK: store i{{[0-9]+}}* [[T_VAR_IN1]], i{{[0-9]+}}** [[T_VAR_ADDR1]], 446 // CHECK: store [2 x [[S_INT_TY]]]* [[S_ARR_IN1]], [2 x [[S_INT_TY]]]** [[S_ARR_ADDR1]], 447 // CHECK: store [[S_INT_TY]]* [[VAR_IN1]], [[S_INT_TY]]** [[VAR_ADDR1]], 448 449 // load content of local address variables 450 // CHECK: [[VEC_ADDR_REF1:%.+]] = load [2 x i{{[0-9]+}}]*, [2 x i{{[0-9]+}}]** [[VEC_ADDR1]], 451 // CHECK: [[T_VAR_ADDR_REF1:%.+]] = load i{{[0-9]+}}*, i{{[0-9]+}}** [[T_VAR_ADDR1]], 452 // CHECK: [[S_ARR_ADDR_REF1:%.+]] = load [2 x [[S_INT_TY]]]*, [2 x [[S_INT_TY]]]** [[S_ARR_ADDR1]], 453 // CHECK-DAG: store i{{[0-9]+}} 0, i{{[0-9]+}}* [[OMP_IS_LAST1]], 454 // CHECK: [[VAR_ADDR1_REF:%.+]] = load [[S_INT_TY]]*, [[S_INT_TY]]** [[VAR_ADDR1]], 455 // CHECK-DAG: store [[S_INT_TY]]* [[VAR_PRIV1]], [[S_INT_TY]]** [[TMP_PRIV1]], 456 457 // CHECK: call void @__kmpc_for_static_init_4( 458 // CHECK: call void {{.*}} @__kmpc_fork_call({{.+}}, {{.+}}, {{.+}} @[[TPAR_OUTL:.+]] to 459 // CHECK: call void @__kmpc_for_static_fini( 460 461 // lastprivates 462 // CHECK: [[OMP_IS_LAST_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[OMP_IS_LAST1]], 463 // CHECK: [[IS_LAST_IT:%.+]] = icmp ne i{{[0-9]+}} [[OMP_IS_LAST_VAL]], 0 464 // CHECK: br i1 [[IS_LAST_IT]], label %[[OMP_LASTPRIV_BLOCK:.+]], label %[[OMP_LASTPRIV_DONE:.+]] 465 466 // CHECK: [[OMP_LASTPRIV_BLOCK]]: 467 // CHECK: [[T_VAR_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[T_VAR_PRIV1]], 468 // CHECK: store i{{[0-9]+}} [[T_VAR_VAL]], i{{[0-9]+}}* [[T_VAR_ADDR_REF1]], 469 // CHECK: [[BCAST_VEC_ADDR_REF:%.+]] = bitcast [2 x i{{[0-9]+}}]* [[VEC_ADDR_REF1]] to i8* 470 // CHECK: [[BCAST_VEC_PRIV:%.+]] = bitcast [2 x i{{[0-9]+}}]* [[VEC_PRIV1]] to i8* 471 // CHECK: call void @llvm.memcpy.{{.+}}(i8* [[BCAST_VEC_ADDR_REF]], i8* [[BCAST_VEC_PRIV]], 472 // 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]+}} 0 473 // CHECK: [[S_ARR_PRIV_BCAST:%.+]] = bitcast [2 x [[S_INT_TY]]]* [[S_ARR_PRIV1]] to [[S_INT_TY]]* 474 // CHECK: [[S_ARR_BEGIN_GEP:%.+]] = getelementptr [[S_INT_TY]], [[S_INT_TY]]* [[S_ARR_BEGIN]], i{{[0-9]+}} 2 475 // CHECK: [[S_ARR_IS_EMPTY:%.+]] = icmp eq [[S_INT_TY]]* [[S_ARR_BEGIN]], [[S_ARR_BEGIN_GEP]] 476 // CHECK: br i1 [[S_ARR_IS_EMPTY]], label %[[S_ARR_COPY_DONE:.+]], label %[[S_ARR_COPY_BLOCK:.+]] 477 // CHECK: [[S_ARR_COPY_BLOCK]]: 478 // CHECK: [[S_ARR_SRC_EL:%.+]] = phi [[S_INT_TY]]*{{.+}} 479 // CHECK: [[S_ARR_DST_EL:%.+]] = phi [[S_INT_TY]]*{{.+}} 480 // CHECK: [[S_ARR_DST_BCAST:%.+]] = bitcast [[S_INT_TY]]* [[S_ARR_DST_EL]] to i8* 481 // CHECK: [[S_ARR_SRC_BCAST:%.+]] = bitcast [[S_INT_TY]]* [[S_ARR_SRC_EL]] to i8* 482 // CHECK: call void @llvm.memcpy.{{.+}}(i8* [[S_ARR_DST_BCAST]], i8* [[S_ARR_SRC_BCAST]]{{.+}}) 483 // CHECK: [[S_ARR_DST_NEXT:%.+]] = getelementptr [[S_INT_TY]], [[S_INT_TY]]* [[S_ARR_DST_EL]], i{{[0-9]+}} 1 484 // CHECK: [[S_ARR_SRC_NEXT:%.+]] = getelementptr{{.+}} 485 // CHECK: [[CPY_IS_FINISHED:%.+]] = icmp eq [[S_INT_TY]]* [[S_ARR_DST_NEXT]], [[S_ARR_BEGIN_GEP]] 486 // CHECK: br i1 [[CPY_IS_FINISHED]], label %[[S_ARR_COPY_DONE]], label %[[S_ARR_COPY_BLOCK]] 487 // CHECK: [[S_ARR_COPY_DONE]]: 488 // CHECK: [[TMP_VAL1:%.+]] = load [[S_INT_TY]]*, [[S_INT_TY]]** [[TMP_PRIV1]], 489 // CHECK: [[VAR_ADDR_REF_BCAST:%.+]] = bitcast [[S_INT_TY]]* [[VAR_ADDR1_REF]] to i8* 490 // CHECK: [[TMP_VAL1_BCAST:%.+]] = bitcast [[S_INT_TY]]* [[TMP_VAL1]] to i8* 491 // CHECK: call void @llvm.memcpy.{{.+}}(i8* [[VAR_ADDR_REF_BCAST]], i8* [[TMP_VAL1_BCAST]],{{.+}}) 492 // CHECK: ret void 493 494 // CHECK: define internal void [[TPAR_OUTL:@.+]](i{{[0-9]+}}* noalias [[GTID_ADDR1:%.+]], i{{[0-9]+}}* noalias %{{.+}}, {{.+}}, {{.+}}, [2 x i{{[0-9]+}}]*{{.+}} [[VEC_IN1:%.+]], i{{[0-9]+}}*{{.+}} [[T_VAR_IN1:%.+]], [2 x [[S_INT_TY]]]*{{.+}} [[S_ARR_IN1:%.+]], [[S_INT_TY]]*{{.+}} [[VAR_IN1:%.+]]) 495 // skip alloca of global_tid and bound_tid, and prev lb and ub vars 496 // CHECK: {{.+}} = alloca i{{[0-9]+}}*, 497 // CHECK: {{.+}} = alloca i{{[0-9]+}}*, 498 // CHECK: alloca i{{[0-9]+}}, 499 // CHECK: alloca i{{[0-9]+}}, 500 // CHECK: [[VEC_ADDR1:%.+]] = alloca [2 x i{{[0-9]+}}]*, 501 // CHECK: [[T_VAR_ADDR1:%.+]] = alloca i{{[0-9]+}}*, 502 // CHECK: [[S_ARR_ADDR1:%.+]] = alloca [2 x [[S_INT_TY]]]*, 503 // CHECK: [[VAR_ADDR1:%.+]] = alloca [[S_INT_TY]]*, 504 // skip loop variables 505 // CHECK: {{.+}} = alloca i{{[0-9]+}}, 506 // CHECK: {{.+}} = alloca i{{[0-9]+}}, 507 // CHECK: {{.+}} = alloca i{{[0-9]+}}, 508 // CHECK: {{.+}} = alloca i{{[0-9]+}}, 509 // CHECK: {{.+}} = alloca i{{[0-9]+}}, 510 // CHECK: [[OMP_IS_LAST1:%.+]] = alloca i{{[0-9]+}}, 511 // CHECK: [[T_VAR_PRIV1:%.+]] = alloca i{{[0-9]+}}, 512 // CHECK: [[VEC_PRIV1:%.+]] = alloca [2 x i{{[0-9]+}}], 513 // CHECK: [[S_ARR_PRIV1:%.+]] = alloca [2 x [[S_INT_TY]]], 514 // CHECK: [[VAR_PRIV1:%.+]] = alloca [[S_INT_TY]], 515 // CHECK: [[TMP_PRIV1:%.+]] = alloca [[S_INT_TY]]*, 516 517 // skip init of bound and global tid 518 // CHECK: store i{{[0-9]+}}* {{.*}}, 519 // CHECK: store i{{[0-9]+}}* {{.*}}, 520 // copy from parameters to local address variables 521 // CHECK: store [2 x i{{[0-9]+}}]* [[VEC_IN1]], [2 x i{{[0-9]+}}]** [[VEC_ADDR1]], 522 // CHECK: store i{{[0-9]+}}* [[T_VAR_IN1]], i{{[0-9]+}}** [[T_VAR_ADDR1]], 523 // CHECK: store [2 x [[S_INT_TY]]]* [[S_ARR_IN1]], [2 x [[S_INT_TY]]]** [[S_ARR_ADDR1]], 524 // CHECK: store [[S_INT_TY]]* [[VAR_IN1]], [[S_INT_TY]]** [[VAR_ADDR1]], 525 526 // load content of local address variables 527 // CHECK: [[VEC_ADDR_REF1:%.+]] = load [2 x i{{[0-9]+}}]*, [2 x i{{[0-9]+}}]** [[VEC_ADDR1]], 528 // CHECK: [[T_VAR_ADDR_REF1:%.+]] = load i{{[0-9]+}}*, i{{[0-9]+}}** [[T_VAR_ADDR1]], 529 // CHECK: [[S_ARR_ADDR_REF1:%.+]] = load [2 x [[S_INT_TY]]]*, [2 x [[S_INT_TY]]]** [[S_ARR_ADDR1]], 530 // CHECK-DAG: store i{{[0-9]+}} 0, i{{[0-9]+}}* [[OMP_IS_LAST1]], 531 // CHECK: [[VAR_ADDR1_REF:%.+]] = load [[S_INT_TY]]*, [[S_INT_TY]]** [[VAR_ADDR1]], 532 // CHECK-DAG: store [[S_INT_TY]]* [[VAR_PRIV1]], [[S_INT_TY]]** [[TMP_PRIV1]], 533 534 // CHECK: call void @__kmpc_for_static_init_4( 535 536 // assignment: vec[i] = t_var; 537 // CHECK: [[IV_VAL1:%.+]] = 538 // CHECK: [[T_VAR_PRIV_VAL1:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[T_VAR_PRIV1]], 539 // CHECK: [[VEC_PTR1:%.+]] = getelementptr inbounds [2 x i{{[0-9]+}}], [2 x i{{[0-9]+}}]* [[VEC_PRIV1]], i{{[0-9]+}} 0, i{{[0-9]+}} {{.+}} 540 // CHECK: store i{{[0-9]+}} [[T_VAR_PRIV_VAL1]], i{{[0-9]+}}* [[VEC_PTR1]], 541 542 // assignment: s_arr[i] = var; 543 // CHECK-DAG: [[S_ARR_PTR1:%.+]] = getelementptr inbounds [2 x [[S_INT_TY]]], [2 x [[S_INT_TY]]]* [[S_ARR_PRIV1]], 544 // CHECK-DAG: [[TMP_VAL1:%.+]] = load [[S_INT_TY]]*, [[S_INT_TY]]** [[TMP_PRIV1]], 545 // CHECK-DAG: [[S_ARR_PTR_BCAST1:%.+]] = bitcast [[S_INT_TY]]* [[S_ARR_PTR1]] to i8* 546 // CHECK-DAG: [[TMP_VAL_BCAST1:%.+]] = bitcast [[S_INT_TY]]* [[TMP_VAL1]] to i8* 547 // CHECK-DAG: call void @llvm.memcpy.{{.+}}(i8* [[S_ARR_PTR_BCAST1]], i8* [[TMP_VAL_BCAST1]], 548 549 // CHECK: call void @__kmpc_for_static_fini( 550 551 // lastprivates 552 // CHECK: [[OMP_IS_LAST_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[OMP_IS_LAST1]], 553 // CHECK: [[IS_LAST_IT:%.+]] = icmp ne i{{[0-9]+}} [[OMP_IS_LAST_VAL]], 0 554 // CHECK: br i1 [[IS_LAST_IT]], label %[[OMP_LASTPRIV_BLOCK:.+]], label %[[OMP_LASTPRIV_DONE:.+]] 555 556 // CHECK: [[OMP_LASTPRIV_BLOCK]]: 557 // CHECK: [[T_VAR_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[T_VAR_PRIV1]], 558 // CHECK: store i{{[0-9]+}} [[T_VAR_VAL]], i{{[0-9]+}}* [[T_VAR_ADDR_REF1]], 559 // CHECK: [[BCAST_VEC_ADDR_REF:%.+]] = bitcast [2 x i{{[0-9]+}}]* [[VEC_ADDR_REF1]] to i8* 560 // CHECK: [[BCAST_VEC_PRIV:%.+]] = bitcast [2 x i{{[0-9]+}}]* [[VEC_PRIV1]] to i8* 561 // CHECK: call void @llvm.memcpy.{{.+}}(i8* [[BCAST_VEC_ADDR_REF]], i8* [[BCAST_VEC_PRIV]], 562 // 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]+}} 0 563 // CHECK: [[S_ARR_PRIV_BCAST:%.+]] = bitcast [2 x [[S_INT_TY]]]* [[S_ARR_PRIV1]] to [[S_INT_TY]]* 564 // CHECK: [[S_ARR_BEGIN_GEP:%.+]] = getelementptr [[S_INT_TY]], [[S_INT_TY]]* [[S_ARR_BEGIN]], i{{[0-9]+}} 2 565 // CHECK: [[S_ARR_IS_EMPTY:%.+]] = icmp eq [[S_INT_TY]]* [[S_ARR_BEGIN]], [[S_ARR_BEGIN_GEP]] 566 // CHECK: br i1 [[S_ARR_IS_EMPTY]], label %[[S_ARR_COPY_DONE:.+]], label %[[S_ARR_COPY_BLOCK:.+]] 567 // CHECK: [[S_ARR_COPY_BLOCK]]: 568 // CHECK: [[S_ARR_SRC_EL:%.+]] = phi [[S_INT_TY]]*{{.+}} 569 // CHECK: [[S_ARR_DST_EL:%.+]] = phi [[S_INT_TY]]*{{.+}} 570 // CHECK: [[S_ARR_DST_BCAST:%.+]] = bitcast [[S_INT_TY]]* [[S_ARR_DST_EL]] to i8* 571 // CHECK: [[S_ARR_SRC_BCAST:%.+]] = bitcast [[S_INT_TY]]* [[S_ARR_SRC_EL]] to i8* 572 // CHECK: call void @llvm.memcpy.{{.+}}(i8* [[S_ARR_DST_BCAST]], i8* [[S_ARR_SRC_BCAST]]{{.+}}) 573 // CHECK: [[S_ARR_DST_NEXT:%.+]] = getelementptr [[S_INT_TY]], [[S_INT_TY]]* [[S_ARR_DST_EL]], i{{[0-9]+}} 1 574 // CHECK: [[S_ARR_SRC_NEXT:%.+]] = getelementptr{{.+}} 575 // CHECK: [[CPY_IS_FINISHED:%.+]] = icmp eq [[S_INT_TY]]* [[S_ARR_DST_NEXT]], [[S_ARR_BEGIN_GEP]] 576 // CHECK: br i1 [[CPY_IS_FINISHED]], label %[[S_ARR_COPY_DONE]], label %[[S_ARR_COPY_BLOCK]] 577 // CHECK: [[S_ARR_COPY_DONE]]: 578 // CHECK: [[TMP_VAL1:%.+]] = load [[S_INT_TY]]*, [[S_INT_TY]]** [[TMP_PRIV1]], 579 // CHECK: [[VAR_ADDR_REF_BCAST:%.+]] = bitcast [[S_INT_TY]]* [[VAR_ADDR1_REF]] to i8* 580 // CHECK: [[TMP_VAL1_BCAST:%.+]] = bitcast [[S_INT_TY]]* [[TMP_VAL1]] to i8* 581 // CHECK: call void @llvm.memcpy.{{.+}}(i8* [[VAR_ADDR_REF_BCAST]], i8* [[TMP_VAL1_BCAST]],{{.+}}) 582 // CHECK: ret void 583 584 #endif 585