1 // RxUN: %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 parallel for firstprivate(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 parallel for firstprivate(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 68 // addr alloca's 69 // LAMBDA: [[G_ADDR:%.+]] = alloca double*, 70 // LAMBDA: [[G1_ADDR:%.+]] = alloca double*, 71 // LAMBDA: [[SVAR_ADDR:%.+]] = alloca i{{[0-9]+}}*, 72 // LAMBDA: [[SFVAR_ADDR:%.+]] = alloca float*, 73 // LAMBDA: [[G1_REF:%.+]] = alloca double*, 74 75 // private alloca's 76 // LAMBDA: [[G_PRIV:%.+]] = alloca double, 77 // LAMBDA: [[G1_PRIV:%.+]] = alloca double, 78 // LAMBDA: [[TMP_PRIV:%.+]] = alloca double*, 79 // LAMBDA: [[SVAR_PRIV:%.+]] = alloca i{{[0-9]+}}, 80 // LAMBDA: [[SFVAR_PRIV:%.+]] = alloca float, 81 82 // transfer input parameters into addr alloca's 83 // LAMBDA-DAG: store {{.+}} [[G_IN]], {{.+}} [[G_ADDR]], 84 // LAMBDA-DAG: store {{.+}} [[G1_IN]], {{.+}} [[G1_ADDR]], 85 // LAMBDA-DAG: store {{.+}} [[SVAR_IN]], {{.+}} [[SVAR_ADDR]], 86 // LAMBDA-DAG: store {{.+}} [[SFVAR_IN]], {{.+}} [[SFVAR_ADDR]], 87 88 // LAMBDA-DAG: [[G_CONV:%.+]] = load {{.+}}*, {{.+}}** [[G_ADDR]] 89 // LAMBDA-DAG: [[G1_CONV:%.+]] = load {{.+}}*, {{.+}}** [[G1_ADDR]] 90 // LAMBDA-DAG: [[SVAR_CONV:%.+]] = load {{.+}}*, {{.+}}** [[SVAR_ADDR]] 91 // LAMBDA-DAG: [[SFVAR_CONV:%.+]] = load {{.+}}*, {{.+}}** [[SFVAR_ADDR]] 92 93 // init private alloca's with addr alloca's 94 // g 95 // LAMBDA-DAG: [[G_ADDR_VAL:%.+]] = load {{.+}}, {{.+}}* [[G_CONV]], 96 // LAMBDA-DAG: store {{.+}} [[G_ADDR_VAL]], {{.+}}* [[G_PRIV]], 97 98 // g1 99 // LAMBDA-DAG: [[TMP_REF:%.+]] = load {{.+}}*, {{.+}}** [[G1_REF]], 100 // LAMBDA-DAG: [[TMP_VAL:%.+]] = load {{.+}}, {{.+}}* [[TMP_REF]], 101 // LAMBDA-DAG: store {{.+}} [[TMP_VAL]], {{.+}}* [[G1_PRIV]] 102 // LAMBDA-DAG: store {{.+}}* [[G1_PRIV]], {{.+}}** [[TMP_PRIV]], 103 104 // svar 105 // LAMBDA-DAG: [[SVAR_VAL:%.+]] = load {{.+}}, {{.+}}* [[SVAR_CONV]], 106 // LAMBDA-DAG: store {{.+}} [[SVAR_VAL]], {{.+}}* [[SVAR_PRIV]], 107 108 // sfvar 109 // LAMBDA-DAG: [[SFVAR_VAL:%.+]] = load {{.+}}, {{.+}}* [[SFVAR_CONV]], 110 // LAMBDA-DAG: store {{.+}} [[SFVAR_VAL]], {{.+}}* [[SFVAR_PRIV]], 111 112 // LAMBDA: call {{.*}}void @__kmpc_for_static_init_4( 113 // pass firstprivate parameters to parallel outlined function 114 // g 115 // LAMBDA-64-DAG: [[G_PRIV_VAL:%.+]] = load {{.+}}, {{.+}}* [[G_PRIV]], 116 // LAMBDA-64: [[G_CAST_CONV:%.+]] = bitcast {{.+}}* [[G_CAST:%.+]] to 117 // LAMBDA-64-DAG: store {{.+}} [[G_PRIV_VAL]], {{.+}}* [[G_CAST_CONV]], 118 // LAMBDA-64-DAG: [[G_PAR:%.+]] = load {{.+}}, {{.+}}* [[G_CAST]], 119 120 // g1 121 // LAMBDA-DAG: [[TMP_PRIV_VAL:%.+]] = load {{.+}}, {{.+}}* [[TMP_PRIV]], 122 // LAMBDA-DAG: [[G1_PRIV_VAL:%.+]] = load {{.+}}, {{.+}}* [[TMP_PRIV_VAL]], 123 // LAMBDA: [[G1_CAST_CONV:%.+]] = bitcast {{.+}}* [[G1_CAST:%.+]] to 124 // LAMBDA-DAG: store {{.+}} [[G1_PRIV_VAL]], {{.+}}* [[G1_CAST_CONV]], 125 // LAMBDA-DAG: [[G1_PAR:%.+]] = load {{.+}}, {{.+}}* [[G1_CAST]], 126 127 // svar 128 // LAMBDA: [[SVAR_VAL:%.+]] = load {{.+}}, {{.+}}* [[SVAR_PRIV]], 129 // LAMBDA-64-DAG: [[SVAR_CAST_CONV:%.+]] = bitcast {{.+}}* [[SVAR_CAST:%.+]] to 130 // LAMBDA-64-DAG: store {{.+}} [[SVAR_VAL]], {{.+}}* [[SVAR_CAST_CONV]], 131 // LAMBDA-32-DAG: store {{.+}} [[SVAR_VAL]], {{.+}}* [[SVAR_CAST:%.+]], 132 // LAMBDA-DAG: [[SVAR_PAR:%.+]] = load {{.+}}, {{.+}}* [[SVAR_CAST]], 133 134 // sfvar 135 // LAMBDA: [[SFVAR_VAL:%.+]] = load {{.+}}, {{.+}}* [[SFVAR_PRIV]], 136 // LAMBDA-DAG: [[SFVAR_CAST_CONV:%.+]] = bitcast {{.+}}* [[SFVAR_CAST:%.+]] to 137 // LAMBDA-DAG: store {{.+}} [[SFVAR_VAL]], {{.+}}* [[SFVAR_CAST_CONV]], 138 // LAMBDA-DAG: [[SFVAR_PAR:%.+]] = load {{.+}}, {{.+}}* [[SFVAR_CAST]], 139 140 // LAMBDA-64: call{{.+}} @__kmpc_fork_call({{.+}}, {{.+}}, {{.+}}[[OMP_PARFOR_OUTLINED:@.+]] to void ({{.+}})*), {{.+}}, {{.+}}, {{.+}} [[G_PAR]], {{.+}} [[G1_PAR]], {{.+}} [[SVAR_PAR]], {{.+}} [[SFVAR_PAR]]) 141 // LAMBDA-32: call{{.+}} @__kmpc_fork_call({{.+}}, {{.+}}, {{.+}}[[OMP_PARFOR_OUTLINED:@.+]] to void ({{.+}})*), {{.+}}, {{.+}}, {{.+}} [[G_PRIV]], {{.+}} [[G1_PAR]], {{.+}} [[SVAR_PAR]], {{.+}} [[SFVAR_PAR]]) 142 // LAMBDA: call {{.*}}void @__kmpc_for_static_fini( 143 // LAMBDA: ret void 144 145 146 // LAMBDA-64: define{{.+}} void [[OMP_PARFOR_OUTLINED]](i32* noalias %{{.+}}, i32* noalias %{{.+}}, {{.+}}, {{.+}}, i{{[0-9]+}} [[G_IN:%.+]], i{{[0-9]+}} [[G1_IN:%.+]], i{{[0-9]+}} [[SVAR_IN:%.+]], i{{[0-9]+}} [[SFVAR_IN:%.+]]) 147 // LAMBDA-32: define{{.+}} void [[OMP_PARFOR_OUTLINED]](i32* noalias %{{.+}}, i32* noalias %{{.+}}, {{.+}}, {{.+}}, double* {{.+}} [[G_IN:%.+]], i{{[0-9]+}} [[G1_IN:%.+]], i{{[0-9]+}} [[SVAR_IN:%.+]], i{{[0-9]+}} [[SFVAR_IN:%.+]]) 148 // skip initial params 149 // LAMBDA: {{.+}} = alloca{{.+}}, 150 // LAMBDA: {{.+}} = alloca{{.+}}, 151 // LAMBDA: {{.+}} = alloca{{.+}}, 152 // LAMBDA: {{.+}} = alloca{{.+}}, 153 154 // addr alloca's 155 // LAMBDA-64: [[G_ADDR:%.+]] = alloca i{{[0-9]+}}, 156 // LAMBDA-32: [[G_ADDR:%.+]] = alloca double*, 157 // LAMBDA: [[G1_ADDR:%.+]] = alloca i{{[0-9]+}}, 158 // LAMBDA: [[SVAR_ADDR:%.+]] = alloca i{{[0-9]+}}, 159 // LAMBDA: [[SFVAR_ADDR:%.+]] = alloca i{{[0-9]+}}, 160 // LAMBDA: [[G1_REF:%.+]] = alloca double*, 161 162 // private alloca's (only for 32-bit) 163 // LAMBDA-32: [[G_PRIV:%.+]] = alloca double, 164 165 // transfer input parameters into addr alloca's 166 // LAMBDA-DAG: store {{.+}} [[G_IN]], {{.+}} [[G_ADDR]], 167 // LAMBDA-DAG: store {{.+}} [[G1_IN]], {{.+}} [[G1_ADDR]], 168 // LAMBDA-DAG: store {{.+}} [[SVAR_IN]], {{.+}} [[SVAR_ADDR]], 169 // LAMBDA-DAG: store {{.+}} [[SFVAR_IN]], {{.+}} [[SFVAR_ADDR]], 170 171 // prepare parameters for lambda 172 // g 173 // LAMBDA-64-DAG: [[G_CONV:%.+]] = bitcast {{.+}}* [[G_ADDR]] to 174 // LAMBDA-32-DAG: [[G_ADDR_REF:%.+]] = load {{.+}}*, {{.+}}** [[G_ADDR]] 175 // LAMBDA-32-DAG: [[G_ADDR_VAL:%.+]] = load {{.+}}, {{.+}}* [[G_ADDR_REF]], 176 // LAMBDA-32-DAG: store {{.+}} [[G_ADDR_VAL]], {{.+}}* [[G_PRIV]], 177 178 // g1 179 // LAMBDA-DAG: [[G1_CONV:%.+]] = bitcast {{.+}}* [[G1_ADDR]] to 180 // LAMBDA-DAG: store {{.+}}* [[G1_CONV]], {{.+}}* [[G1_REF]], 181 182 // svar 183 // LAMBDA-64-DAG: [[SVAR_CONV:%.+]] = bitcast {{.+}}* [[SVAR_ADDR]] to 184 185 // sfvar 186 // LAMBDA-DAG: [[SFVAR_CONV:%.+]] = bitcast {{.+}}* [[SFVAR_ADDR]] to 187 188 // LAMBDA: call {{.*}}void @__kmpc_for_static_init_4( 189 g = 1; 190 g1 = 1; 191 svar = 3; 192 sfvar = 4.0; 193 // LAMBDA-64: store double 1.0{{.+}}, double* [[G_CONV]], 194 // LAMBDA-32: store double 1.0{{.+}}, double* [[G_PRIV]], 195 // LAMBDA: [[G1_REF_REF:%.+]] = load {{.+}}*, {{.+}}** [[G1_REF]], 196 // LAMBDA: store {{.+}} 1.0{{.+}}, {{.+}}* [[G1_REF_REF]], 197 // LAMBDA-64: store {{.+}} 3, {{.+}}* [[SVAR_CONV]], 198 // LAMBDA-32: store {{.+}} 3, {{.+}}* [[SVAR_ADDR]], 199 // LAMBDA: store {{.+}} 4.0{{.+}}, {{.+}}* [[SFVAR_CONV]], 200 201 // pass params to inner lambda 202 // LAMBDA: [[G_PRIVATE_ADDR_REF:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* [[ARG:%.+]], i{{[0-9]+}} 0, i{{[0-9]+}} 0 203 // LAMBDA-64: store double* [[G_CONV]], double** [[G_PRIVATE_ADDR_REF]], 204 // LAMBDA-32: store double* [[G_PRIV]], double** [[G_PRIVATE_ADDR_REF]], 205 // LAMBDA: [[G1_PRIVATE_ADDR_REF:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* [[ARG:%.+]], i{{[0-9]+}} 0, i{{[0-9]+}} 1 206 // LAMBDA: [[G1_REF_REF:%.+]] = load double*, double** [[G1_REF]], 207 // LAMBDA: store double* [[G1_REF_REF]], double** [[G1_PRIVATE_ADDR_REF]], 208 // LAMBDA: [[SVAR_PRIVATE_ADDR_REF:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* [[ARG:%.+]], i{{[0-9]+}} 0, i{{[0-9]+}} 2 209 // LAMBDA-64: store i{{[0-9]+}}* [[SVAR_CONV]], i{{[0-9]+}}** [[SVAR_PRIVATE_ADDR_REF]] 210 // LAMBDA-32: store i{{[0-9]+}}* [[SVAR_ADDR]], i{{[0-9]+}}** [[SVAR_PRIVATE_ADDR_REF]] 211 // LAMBDA: [[SFVAR_PRIVATE_ADDR_REF:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* [[ARG:%.+]], i{{[0-9]+}} 0, i{{[0-9]+}} 3 212 // LAMBDA: store float* [[SFVAR_CONV]], float** [[SFVAR_PRIVATE_ADDR_REF]] 213 // LAMBDA: call{{.*}} void [[INNER_LAMBDA:@.+]](%{{.+}}* [[ARG]]) 214 // LAMBDA: call {{.*}}void @__kmpc_for_static_fini( 215 // LAMBDA: ret void 216 [&]() { 217 // LAMBDA: define {{.+}} void [[INNER_LAMBDA]](%{{.+}}* [[ARG_PTR:%.+]]) 218 // LAMBDA: store %{{.+}}* [[ARG_PTR]], %{{.+}}** [[ARG_PTR_REF:%.+]], 219 g = 2; 220 g1 = 2; 221 svar = 4; 222 sfvar = 8.0; 223 // LAMBDA: [[ARG_PTR:%.+]] = load %{{.+}}*, %{{.+}}** [[ARG_PTR_REF]] 224 // LAMBDA: [[G_PTR_REF:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* [[ARG_PTR]], i{{[0-9]+}} 0, i{{[0-9]+}} 0 225 // LAMBDA: [[G_REF:%.+]] = load double*, double** [[G_PTR_REF]] 226 // LAMBDA: store double 2.0{{.+}}, double* [[G_REF]] 227 228 // LAMBDA: [[TMP_PTR_REF:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* [[ARG_PTR]], i{{[0-9]+}} 0, i{{[0-9]+}} 1 229 // LAMBDA: [[G1_REF:%.+]] = load double*, double** [[TMP_PTR_REF]] 230 // LAMBDA: store double 2.0{{.+}}, double* [[G1_REF]], 231 // LAMBDA: [[SVAR_PTR_REF:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* [[ARG_PTR]], i{{[0-9]+}} 0, i{{[0-9]+}} 2 232 // LAMBDA: [[SVAR_REF:%.+]] = load i{{[0-9]+}}*, i{{[0-9]+}}** [[SVAR_PTR_REF]] 233 // LAMBDA: store i{{[0-9]+}} 4, i{{[0-9]+}}* [[SVAR_REF]] 234 // LAMBDA: [[SFVAR_PTR_REF:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* [[ARG_PTR]], i{{[0-9]+}} 0, i{{[0-9]+}} 3 235 // LAMBDA: [[SFVAR_REF:%.+]] = load float*, float** [[SFVAR_PTR_REF]] 236 // LAMBDA: store float 8.0{{.+}}, float* [[SFVAR_REF]] 237 }(); 238 } 239 }(); 240 return 0; 241 #else 242 S<float> test; 243 int t_var = 0; 244 int vec[] = {1, 2}; 245 S<float> s_arr[] = {1, 2}; 246 S<float> &var = test; 247 248 #pragma omp target 249 #pragma omp teams 250 #pragma omp distribute parallel for firstprivate(t_var, vec, s_arr, s_arr, var, var, svar) 251 for (int i = 0; i < 2; ++i) { 252 vec[i] = t_var; 253 s_arr[i] = var; 254 } 255 return tmain<int>(); 256 #endif 257 } 258 259 // CHECK-LABEL: define{{.*}} i{{[0-9]+}} @main() 260 // CHECK: [[TEST:%.+]] = alloca [[S_FLOAT_TY]], 261 // CHECK: call {{.*}} [[S_FLOAT_TY_DEF_CONSTR:@.+]]([[S_FLOAT_TY]]* [[TEST]]) 262 // CHECK: call i{{[0-9]+}} @__tgt_target_teams( 263 // CHECK: call void [[OFFLOAD_FUN_0:@.+]]( 264 // CHECK: call {{.*}} [[S_FLOAT_TY_DEF_DESTR:@.+]]([[S_FLOAT_TY]]* [[TEST]]) 265 266 // CHECK: define{{.+}} [[OFFLOAD_FUN_0]](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]+}} [[SVAR_IN:%.+]]) 267 // CHECK: call void (%{{.+}}*, i{{[0-9]+}}, void (i{{[0-9]+}}*, i{{[0-9]+}}*, ...)*, ...) @__kmpc_fork_teams(%{{.+}}* @{{.+}}, i{{[0-9]+}} 5, void (i{{[0-9]+}}*, i{{[0-9]+}}*, ...)* bitcast (void (i{{[0-9]+}}*, i{{[0-9]+}}*, i{{[0-9]+}}*, [2 x i{{[0-9]+}}]*, [2 x [[S_FLOAT_TY]]]*, [[S_FLOAT_TY]]*, i{{[0-9]+}}*)* [[OMP_OUTLINED_0:@.+]] to void 268 // CHECK: ret 269 270 // CHECK: define internal void [[OMP_OUTLINED_0]](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]+}}*{{.+}} [[SVAR_IN:%.+]]) 271 272 // CHECK: alloca i{{[0-9]+}}*, 273 // CHECK: alloca i{{[0-9]+}}*, 274 // addr alloca's 275 // CHECK: [[T_VAR_ADDR:%.+]] = alloca i{{[0-9]+}}*, 276 // CHECK: [[VEC_ADDR:%.+]] = alloca [2 x i{{[0-9]+}}]*, 277 // CHECK: [[S_ARR_ADDR:%.+]] = alloca [2 x [[S_FLOAT_TY]]]*, 278 // CHECK: [[VAR_ADDR:%.+]] = alloca [[S_FLOAT_TY]]*, 279 // CHECK: [[SVAR_ADDR:%.+]] = alloca i{{[0-9]+}}*, 280 // CHECK: [[TMP:%.+]] = alloca [[S_FLOAT_TY]]*, 281 282 // skip loop alloca's 283 // CHECK: [[OMP_IV:.omp.iv+]] = alloca i{{[0-9]+}}, 284 // CHECK: [[OMP_LB:.omp.comb.lb+]] = alloca i{{[0-9]+}}, 285 // CHECK: [[OMP_UB:.omp.comb.ub+]] = alloca i{{[0-9]+}}, 286 // CHECK: [[OMP_ST:.omp.stride+]] = alloca i{{[0-9]+}}, 287 // CHECK: [[OMP_IS_LAST:.omp.is_last+]] = alloca i{{[0-9]+}}, 288 289 // private alloca's 290 // CHECK: [[T_VAR_PRIV:%.+]] = alloca i{{[0-9]+}}, 291 // CHECK: [[VEC_PRIV:%.+]] = alloca [2 x i{{[0-9]+}}], 292 // CHECK: [[S_ARR_PRIV:%.+]] = alloca [2 x [[S_FLOAT_TY]]], 293 // CHECK: [[VAR_PRIV:%.+]] = alloca [[S_FLOAT_TY]], 294 // CHECK: [[TMP_PRIV:%.+]] = alloca [[S_FLOAT_TY]]*, 295 // CHECK: [[SVAR_PRIV:%.+]] = alloca i{{[0-9]+}}, 296 297 // CHECK: store i{{[0-9]+}}* [[GTID_ADDR]], i{{[0-9]+}}** [[GTID_ADDR_REF:%.+]] 298 299 // init addr alloca's with input values 300 // CHECK-DAG: store {{.+}} [[T_VAR_IN]], {{.+}}* [[T_VAR_ADDR]], 301 // CHECK-DAG: store {{.+}} [[VEC_IN]], {{.+}} [[VEC_ADDR]], 302 // CHECK-DAG: store {{.+}} [[S_ARR_IN]], {{.+}} [[S_ARR_ADDR]], 303 // CHECK-DAG: store {{.+}} [[VAR_IN]], {{.+}} [[VAR_ADDR]], 304 // CHECK-DAG: store {{.+}} [[SVAR_IN]], {{.+}} [[SVAR_ADDR]], 305 306 // init private alloca's with addr alloca's 307 // t-var 308 // CHECK-DAG: [[T_VAR_REF:%.+]] = load {{.+}}, {{.+}}** [[T_VAR_ADDR]], 309 // CHECK-DAG: [[T_VAR_ADDR_VAL:%.+]] = load {{.+}}, {{.+}}* [[T_VAR_REF]], 310 // CHECK-DAG: store {{.+}} [[T_VAR_ADDR_VAL]], {{.+}} [[T_VAR_PRIV]], 311 312 // vec 313 // CHECK-DAG: [[VEC_ADDR_VAL:%.+]] = load {{.+}}*, {{.+}}** [[VEC_ADDR]], 314 // CHECK-DAG: [[VEC_PRIV_BCAST:%.+]] = bitcast {{.+}} [[VEC_PRIV]] to 315 // CHECK-DAG: [[VEC_ADDR_BCAST:%.+]] = bitcast {{.+}} [[VEC_ADDR_VAL]] to 316 // CHECK-DAG: call void @llvm.memcpy{{.+}}({{.+}}* [[VEC_PRIV_BCAST]], {{.+}}* [[VEC_ADDR_BCAST]], 317 318 // s_arr 319 // CHECK-DAG: [[S_ARR_ADDR_VAL:%.+]] = load {{.+}}*, {{.+}}** [[S_ARR_ADDR]], 320 // CHECK-DAG: [[S_ARR_BGN:%.+]] = getelementptr {{.+}}, {{.+}}* [[S_ARR_PRIV]], 321 // CHECK-DAG: [[S_ARR_ADDR_BCAST:%.+]] = bitcast {{.+}}* [[S_ARR_ADDR_VAL]] to 322 // CHECK-DAG: [[S_ARR_BGN_GEP:%.+]] = getelementptr {{.+}}, {{.+}}* [[S_ARR_BGN]], 323 // CHECK-DAG: [[S_ARR_EMPTY:%.+]] = icmp {{.+}} [[S_ARR_BGN]], [[S_ARR_BGN_GEP]] 324 // CHECK-DAG: br {{.+}} [[S_ARR_EMPTY]], label %[[CPY_DONE:.+]], label %[[CPY_BODY:.+]] 325 // CHECK-DAG: [[CPY_BODY]]: 326 // CHECK-DAG: call void @llvm.memcpy{{.+}}( 327 // CHECK-DAG: [[CPY_DONE]]: 328 329 // var 330 // CHECK-DAG: [[TMP_REF:%.+]] = load {{.+}}*, {{.+}}* [[TMP]], 331 // CHECK-DAG: [[VAR_PRIV_BCAST:%.+]] = bitcast {{.+}}* [[VAR_PRIV]] to 332 // CHECK-DAG: [[TMP_REF_BCAST:%.+]] = bitcast {{.+}}* [[TMP_REF]] to 333 // CHECK-DAG: call void @llvm.memcpy.{{.+}}({{.+}}* [[VAR_PRIV_BCAST]], {{.+}}* [[TMP_REF_BCAST]], 334 // CHECK-DAG: store {{.+}}* [[VAR_PRIV]], {{.+}}** [[TMP_PRIV]], 335 336 // svar 337 // CHECK-DAG: [[SVAR_REF:%.+]] = load {{.+}}*, {{.+}}** [[SVAR_ADDR]], 338 // CHECK-DAG: [[SVAR:%.+]] = load {{.+}}, {{.+}}* [[SVAR_REF]], 339 // CHECK-DAG: store {{.+}} [[SVAR]], {{.+}}* [[SVAR_PRIV]], 340 341 // CHECK: call void @__kmpc_for_static_init_4( 342 // pass private alloca's to fork 343 // CHECK-DAG: [[T_VAR_PRIV_VAL:%.+]] = load {{.+}}, {{.+}}* [[T_VAR_PRIV]], 344 // not dag to distinguish with S_VAR_CAST 345 // CHECK-64: [[T_VAR_CAST_CONV:%.+]] = bitcast {{.+}}* [[T_VAR_CAST:%.+]] to 346 // CHECK-64-DAG: store {{.+}} [[T_VAR_PRIV_VAL]], {{.+}} [[T_VAR_CAST_CONV]], 347 // CHECK-32: store {{.+}} [[T_VAR_PRIV_VAL]], {{.+}} [[T_VAR_CAST:%.+]], 348 // CHECK-DAG: [[T_VAR_CAST_VAL:%.+]] = load {{.+}}, {{.+}}* [[T_VAR_CAST]], 349 // CHECK-DAG: [[TMP_PRIV_VAL:%.+]] = load [[S_FLOAT_TY]]*, [[S_FLOAT_TY]]** [[TMP_PRIV]], 350 // CHECK-DAG: [[SVAR_PRIV_VAL:%.+]] = load {{.+}}, {{.+}}* [[SVAR_PRIV]], 351 // CHECK-64-DAG: [[SVAR_CAST_CONV:%.+]] = bitcast {{.+}}* [[SVAR_CAST:%.+]] to 352 // CHECK-64-DAG: store {{.+}} [[SVAR_PRIV_VAL]], {{.+}}* [[SVAR_CAST_CONV]], 353 // CHECK-32-DAG: store {{.+}} [[SVAR_PRIV_VAL]], {{.+}}* [[SVAR_CAST:%.+]], 354 // CHECK-DAG: [[SVAR_CAST_VAL:%.+]] = load {{.+}}, {{.+}}* [[SVAR_CAST]], 355 // CHECK: call{{.+}} @__kmpc_fork_call({{.+}}, {{.+}}, {{.+}}[[OMP_PARFOR_OUTLINED_0:@.+]] to void ({{.+}})*), {{.+}}, {{.+}}, [2 x i{{[0-9]+}}]* [[VEC_PRIV]], i{{[0-9]+}} [[T_VAR_CAST_VAL]], [2 x [[S_FLOAT_TY]]]* [[S_ARR_PRIV]], [[S_FLOAT_TY]]* [[TMP_PRIV_VAL]], i{{[0-9]+}} [[SVAR_CAST_VAL]]) 356 // CHECK: call void @__kmpc_for_static_fini( 357 358 // call destructors: var.. 359 // CHECK-DAG: call {{.+}} [[S_FLOAT_TY_DEF_DESTR]]([[S_FLOAT_TY]]* [[VAR_PRIV]]) 360 361 // ..and s_arr 362 // CHECK: {{.+}}: 363 // CHECK: [[S_ARR_EL_PAST:%.+]] = phi [[S_FLOAT_TY]]* 364 // CHECK: [[S_ARR_PRIV_ITEM:%.+]] = getelementptr {{.+}}, {{.+}} [[S_ARR_EL_PAST]], 365 // CHECK: call {{.*}} [[S_FLOAT_TY_DEF_DESTR]]([[S_FLOAT_TY]]* [[S_ARR_PRIV_ITEM]]) 366 367 // CHECK: ret void 368 369 // By OpenMP specifications, 'firstprivate' applies to both distribute and parallel for. 370 // However, the support for 'firstprivate' of 'parallel' is only used when 'parallel' 371 // is found alone. Therefore we only have one 'firstprivate' support for 'parallel for' 372 // in combination 373 // CHECK: define internal void [[OMP_PARFOR_OUTLINED_0]]({{.+}}, {{.+}}, {{.+}}, {{.+}}, [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]+}} [[SVAR_IN:%.+]]) 374 375 // addr alloca's 376 // CHECK: [[VEC_ADDR:%.+]] = alloca [2 x i{{[0-9]+}}]*, 377 // CHECK: [[T_VAR_ADDR:%.+]] = alloca i{{[0-9]+}}, 378 // CHECK: [[S_ARR_ADDR:%.+]] = alloca [2 x [[S_FLOAT_TY]]]*, 379 // CHECK: [[VAR_ADDR:%.+]] = alloca [[S_FLOAT_TY]]*, 380 // CHECK: [[SVAR_ADDR:%.+]] = alloca i{{[0-9]+}}, 381 382 // skip loop alloca's 383 // CHECK: [[OMP_IV:.omp.iv+]] = alloca i{{[0-9]+}}, 384 // CHECK: [[OMP_LB:.omp.lb+]] = alloca i{{[0-9]+}}, 385 // CHECK: [[OMP_UB:.omp.ub+]] = alloca i{{[0-9]+}}, 386 // CHECK: [[OMP_ST:.omp.stride+]] = alloca i{{[0-9]+}}, 387 // CHECK: [[OMP_IS_LAST:.omp.is_last+]] = alloca i{{[0-9]+}}, 388 389 // private alloca's 390 // CHECK: [[VEC_PRIV:%.+]] = alloca [2 x i{{[0-9]+}}], 391 // CHECK: [[S_ARR_PRIV:%.+]] = alloca [2 x [[S_FLOAT_TY]]], 392 // CHECK: [[VAR_PRIV:%.+]] = alloca [[S_FLOAT_TY]], 393 // CHECK: [[TMP_PRIV:%.+]] = alloca [[S_FLOAT_TY]]*, 394 395 // CHECK: store i{{[0-9]+}}* [[GTID_ADDR]], i{{[0-9]+}}** [[GTID_ADDR_REF:%.+]] 396 397 // init addr alloca's with input values 398 // CHECK-DAG: store {{.+}} [[VEC_IN]], {{.+}} [[VEC_ADDR]], 399 // CHECK-DAG: store {{.+}} [[T_VAR_IN]], {{.+}}* [[T_VAR_ADDR]], 400 // CHECK-DAG: store {{.+}} [[S_ARR_IN]], {{.+}} [[S_ARR_ADDR]], 401 // CHECK-DAG: store {{.+}} [[VAR_IN]], {{.+}} [[VAR_ADDR]], 402 // CHECK-DAG: store {{.+}} [[SVAR_IN]], {{.+}} [[SVAR_ADDR]], 403 404 // init private alloca's with addr alloca's 405 // vec 406 // CHECK-DAG: [[VEC_ADDR_VAL:%.+]] = load {{.+}}*, {{.+}}** [[VEC_ADDR]], 407 // CHECK-DAG: [[VEC_PRIV_BCAST:%.+]] = bitcast {{.+}} [[VEC_PRIV]] to 408 // CHECK-DAG: [[VEC_ADDR_BCAST:%.+]] = bitcast {{.+}} [[VEC_ADDR_VAL]] to 409 // CHECK-DAG: call void @llvm.memcpy{{.+}}({{.+}}* [[VEC_PRIV_BCAST]], {{.+}}* [[VEC_ADDR_BCAST]], 410 411 // s_arr 412 // CHECK-DAG: [[S_ARR_ADDR_VAL:%.+]] = load {{.+}}*, {{.+}}** [[S_ARR_ADDR]], 413 // CHECK-DAG: [[S_ARR_BGN:%.+]] = getelementptr {{.+}}, {{.+}}* [[S_ARR_PRIV]], 414 // CHECK-DAG: [[S_ARR_ADDR_BCAST:%.+]] = bitcast {{.+}}* [[S_ARR_ADDR_VAL]] to 415 // CHECK-DAG: [[S_ARR_BGN_GEP:%.+]] = getelementptr {{.+}}, {{.+}}* [[S_ARR_BGN]], 416 // CHECK-DAG: [[S_ARR_EMPTY:%.+]] = icmp {{.+}} [[S_ARR_BGN]], [[S_ARR_BGN_GEP]] 417 // CHECK-DAG: br {{.+}} [[S_ARR_EMPTY]], label %[[CPY_DONE:.+]], label %[[CPY_BODY:.+]] 418 // CHECK-DAG: [[CPY_BODY]]: 419 // CHECK-DAG: call void @llvm.memcpy{{.+}}( 420 // CHECK-DAG: [[CPY_DONE]]: 421 422 // var 423 // CHECK-DAG: [[VAR_ADDR_REF:%.+]] = load {{.+}}*, {{.+}}* [[VAR_ADDR]], 424 // CHECK-DAG: [[VAR_PRIV_BCAST:%.+]] = bitcast {{.+}}* [[VAR_PRIV]] to 425 // CHECK-DAG: [[VAR_ADDR_BCAST:%.+]] = bitcast {{.+}}* [[VAR_ADDR_REF]] to 426 // CHECK-DAG: call void @llvm.memcpy.{{.+}}({{.+}}* [[VAR_PRIV_BCAST]], {{.+}}* [[VAR_ADDR_BCAST]], 427 // CHECK-DAG: store {{.+}}* [[VAR_PRIV]], {{.+}}** [[TMP_PRIV]], 428 429 // CHECK: call void @__kmpc_for_static_init_4( 430 // CHECK: call void @__kmpc_for_static_fini( 431 432 // call destructors: var.. 433 // CHECK-DAG: call {{.+}} [[S_FLOAT_TY_DEF_DESTR]]([[S_FLOAT_TY]]* [[VAR_PRIV]]) 434 435 // ..and s_arr 436 // CHECK: {{.+}}: 437 // CHECK: [[S_ARR_EL_PAST:%.+]] = phi [[S_FLOAT_TY]]* 438 // CHECK: [[S_ARR_PRIV_ITEM:%.+]] = getelementptr {{.+}}, {{.+}} [[S_ARR_EL_PAST]], 439 // CHECK: call {{.*}} [[S_FLOAT_TY_DEF_DESTR]]([[S_FLOAT_TY]]* [[S_ARR_PRIV_ITEM]]) 440 441 // CHECK: ret void 442 443 // template tmain with S_INT_TY 444 // CHECK-LABEL: define{{.*}} i{{[0-9]+}} @{{.+}}tmain{{.+}}() 445 // CHECK: [[TEST:%.+]] = alloca [[S_INT_TY]], 446 // CHECK: call {{.*}} [[S_INT_TY_DEF_CONSTR:@.+]]([[S_INT_TY]]* [[TEST]]) 447 // CHECK: call i{{[0-9]+}} @__tgt_target_teams( 448 // CHECK: call void [[OFFLOAD_FUN_0:@.+]]( 449 // CHECK: call {{.*}} [[S_INT_TY_DEF_DESTR:@.+]]([[S_INT_TY]]* [[TEST]]) 450 451 // CHECK: define{{.+}} [[OFFLOAD_FUN_0]](i{{[0-9]+}} [[T_VAR_IN:%.+]], [2 x i{{[0-9]+}}]* {{.+}} [[VEC_IN:%.+]], [2 x [[S_INT_TY]]]* {{.+}} [[S_ARR_IN:%.+]], [[S_INT_TY]]* {{.+}} [[VAR_IN:%.+]]) 452 // CHECK: call void (%{{.+}}*, i{{[0-9]+}}, void (i{{[0-9]+}}*, i{{[0-9]+}}*, ...)*, ...) @__kmpc_fork_teams(%{{.+}}* @{{.+}}, i{{[0-9]+}} 4, void (i{{[0-9]+}}*, i{{[0-9]+}}*, ...)* bitcast (void (i{{[0-9]+}}*, i{{[0-9]+}}*, i{{[0-9]+}}*, [2 x i{{[0-9]+}}]*, [2 x [[S_INT_TY]]]*, [[S_INT_TY]]*)* [[OMP_OUTLINED_0:@.+]] to void 453 // CHECK: ret 454 455 // CHECK: define internal void [[OMP_OUTLINED_0]](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_INT_TY]]]* {{.+}} [[S_ARR_IN:%.+]], [[S_INT_TY]]* {{.+}} [[VAR_IN:%.+]]) 456 457 // addr alloca's 458 // CHECK: alloca i{{[0-9]+}}*, 459 // CHECK: alloca i{{[0-9]+}}*, 460 // CHECK: [[T_VAR_ADDR:%.+]] = alloca i{{[0-9]+}}*, 461 // CHECK: [[VEC_ADDR:%.+]] = alloca [2 x i{{[0-9]+}}]*, 462 // CHECK: [[S_ARR_ADDR:%.+]] = alloca [2 x [[S_INT_TY]]]*, 463 // CHECK: [[VAR_ADDR:%.+]] = alloca [[S_INT_TY]]*, 464 // CHECK: [[TMP:%.+]] = alloca [[S_INT_TY]]*, 465 466 // skip loop alloca's 467 // CHECK: [[OMP_IV:.omp.iv+]] = alloca i{{[0-9]+}}, 468 // CHECK: [[OMP_LB:.omp.comb.lb+]] = alloca i{{[0-9]+}}, 469 // CHECK: [[OMP_UB:.omp.comb.ub+]] = alloca i{{[0-9]+}}, 470 // CHECK: [[OMP_ST:.omp.stride+]] = alloca i{{[0-9]+}}, 471 // CHECK: [[OMP_IS_LAST:.omp.is_last+]] = alloca i{{[0-9]+}}, 472 473 // private alloca's 474 // CHECK: [[T_VAR_PRIV:%.+]] = alloca i{{[0-9]+}}, 475 // CHECK: [[VEC_PRIV:%.+]] = alloca [2 x i{{[0-9]+}}], 476 // CHECK: [[S_ARR_PRIV:%.+]] = alloca [2 x [[S_INT_TY]]], 477 // CHECK: [[VAR_PRIV:%.+]] = alloca [[S_INT_TY]], 478 // CHECK: [[TMP_PRIV:%.+]] = alloca [[S_INT_TY]]*, 479 480 // CHECK: store i{{[0-9]+}}* [[GTID_ADDR]], i{{[0-9]+}}** [[GTID_ADDR_REF:%.+]] 481 482 // init addr alloca's with input values 483 // CHECK-DAG: store {{.+}} [[T_VAR_IN]], {{.+}}* [[T_VAR_ADDR]], 484 // CHECK-DAG: store {{.+}} [[VEC_IN]], {{.+}} [[VEC_ADDR]], 485 // CHECK-DAG: store {{.+}} [[S_ARR_IN]], {{.+}} [[S_ARR_ADDR]], 486 // CHECK-DAG: store {{.+}} [[VAR_IN]], {{.+}} [[VAR_ADDR]], 487 488 // init private alloca's with addr alloca's 489 // t-var 490 // CHECK-DAG: [[T_VAR_ADDR_REF:%.+]] = load {{.+}}*, {{.+}}** [[T_VAR_ADDR]], 491 // CHECK-DAG: [[T_VAR_ADDR_VAL:%.+]] = load {{.+}}, {{.+}}* [[T_VAR_ADDR_REF]], 492 // CHECK-DAG: store {{.+}} [[T_VAR_ADDR_VAL]], {{.+}} [[T_VAR_PRIV]], 493 494 // vec 495 // CHECK-DAG: [[VEC_ADDR_VAL:%.+]] = load {{.+}}*, {{.+}}** [[VEC_ADDR]], 496 // CHECK-DAG: [[VEC_PRIV_BCAST:%.+]] = bitcast {{.+}} [[VEC_PRIV]] to 497 // CHECK-DAG: [[VEC_ADDR_BCAST:%.+]] = bitcast {{.+}} [[VEC_ADDR_VAL]] to 498 // CHECK-DAG: call void @llvm.memcpy{{.+}}({{.+}}* [[VEC_PRIV_BCAST]], {{.+}}* [[VEC_ADDR_BCAST]], 499 500 // s_arr 501 // CHECK-DAG: [[S_ARR_ADDR_VAL:%.+]] = load {{.+}}*, {{.+}}** [[S_ARR_ADDR]], 502 // CHECK-DAG: [[S_ARR_BGN:%.+]] = getelementptr {{.+}}, {{.+}}* [[S_ARR_PRIV]], 503 // CHECK-DAG: [[S_ARR_ADDR_BCAST:%.+]] = bitcast {{.+}}* [[S_ARR_ADDR_VAL]] to 504 // CHECK-DAG: [[S_ARR_BGN_GEP:%.+]] = getelementptr {{.+}}, {{.+}}* [[S_ARR_BGN]], 505 // CHECK-DAG: [[S_ARR_EMPTY:%.+]] = icmp {{.+}} [[S_ARR_BGN]], [[S_ARR_BGN_GEP]] 506 // CHECK-DAG: br {{.+}} [[S_ARR_EMPTY]], label %[[CPY_DONE:.+]], label %[[CPY_BODY:.+]] 507 // CHECK-DAG: [[CPY_BODY]]: 508 // CHECK-DAG: call void @llvm.memcpy{{.+}}( 509 // CHECK-DAG: [[CPY_DONE]]: 510 511 // var 512 // CHECK-DAG: [[TMP_REF:%.+]] = load {{.+}}*, {{.+}}* [[TMP]], 513 // CHECK-DAG: [[VAR_PRIV_BCAST:%.+]] = bitcast {{.+}}* [[VAR_PRIV]] to 514 // CHECK-DAG: [[TMP_REF_BCAST:%.+]] = bitcast {{.+}}* [[TMP_REF]] to 515 // CHECK-DAG: call void @llvm.memcpy.{{.+}}({{.+}}* [[VAR_PRIV_BCAST]], {{.+}}* [[TMP_REF_BCAST]], 516 // CHECK-DAG: store {{.+}}* [[VAR_PRIV]], {{.+}}** [[TMP_PRIV]], 517 518 // CHECK: call void @__kmpc_for_static_init_4( 519 // pass private alloca's to fork 520 // CHECK-DAG: [[T_VAR_PRIV_VAL:%.+]] = load {{.+}}, {{.+}}* [[T_VAR_PRIV]], 521 // not dag to distinguish with S_VAR_CAST 522 // CHECK-64: [[T_VAR_CAST_CONV:%.+]] = bitcast {{.+}}* [[T_VAR_CAST:%.+]] to 523 // CHECK-64-DAG: store {{.+}} [[T_VAR_PRIV_VAL]], {{.+}} [[T_VAR_CAST_CONV]], 524 // CHECK-32: store {{.+}} [[T_VAR_PRIV_VAL]], {{.+}} [[T_VAR_CAST:%.+]], 525 // CHECK-DAG: [[T_VAR_CAST_VAL:%.+]] = load {{.+}}, {{.+}}* [[T_VAR_CAST]], 526 // CHECK-DAG: [[TMP_PRIV_VAL:%.+]] = load [[S_INT_TY]]*, [[S_INT_TY]]** [[TMP_PRIV]], 527 // CHECK: call{{.+}} @__kmpc_fork_call({{.+}}, {{.+}}, {{.+}}[[OMP_PARFOR_OUTLINED_0:@.+]] to void ({{.+}})*), {{.+}}, {{.+}}, [2 x i{{[0-9]+}}]* [[VEC_PRIV]], i{{[0-9]+}} [[T_VAR_CAST_VAL]], [2 x [[S_INT_TY]]]* [[S_ARR_PRIV]], [[S_INT_TY]]* [[TMP_PRIV_VAL]]) 528 // CHECK: call void @__kmpc_for_static_fini( 529 530 // call destructors: var.. 531 // CHECK-DAG: call {{.+}} [[S_INT_TY_DEF_DESTR]]([[S_INT_TY]]* [[VAR_PRIV]]) 532 533 // ..and s_arr 534 // CHECK: {{.+}}: 535 // CHECK: [[S_ARR_EL_PAST:%.+]] = phi [[S_INT_TY]]* 536 // CHECK: [[S_ARR_PRIV_ITEM:%.+]] = getelementptr {{.+}}, {{.+}} [[S_ARR_EL_PAST]], 537 // CHECK: call {{.*}} [[S_INT_TY_DEF_DESTR]]([[S_INT_TY]]* [[S_ARR_PRIV_ITEM]]) 538 539 // CHECK: ret void 540 541 // By OpenMP specifications, 'firstprivate' applies to both distribute and parallel for. 542 // However, the support for 'firstprivate' of 'parallel' is only used when 'parallel' 543 // is found alone. Therefore we only have one 'firstprivate' support for 'parallel for' 544 // in combination 545 // CHECK: define internal void [[OMP_PARFOR_OUTLINED_0]]({{.+}}, {{.+}}, {{.+}}, {{.+}}, [2 x i{{[0-9]+}}]* {{.+}} [[VEC_IN:%.+]], i{{[0-9]+}} [[T_VAR_IN:%.+]], [2 x [[S_INT_TY]]]* {{.+}} [[S_ARR_IN:%.+]], [[S_INT_TY]]* {{.+}} [[VAR_IN:%.+]]) 546 547 // addr alloca's 548 // CHECK: [[VEC_ADDR:%.+]] = alloca [2 x i{{[0-9]+}}]*, 549 // CHECK: [[T_VAR_ADDR:%.+]] = alloca i{{[0-9]+}}, 550 // CHECK: [[S_ARR_ADDR:%.+]] = alloca [2 x [[S_INT_TY]]]*, 551 // CHECK: [[VAR_ADDR:%.+]] = alloca [[S_INT_TY]]*, 552 553 // skip loop alloca's 554 // CHECK: [[OMP_IV:.omp.iv+]] = alloca i{{[0-9]+}}, 555 // CHECK: [[OMP_LB:.omp.lb+]] = alloca i{{[0-9]+}}, 556 // CHECK: [[OMP_UB:.omp.ub+]] = alloca i{{[0-9]+}}, 557 // CHECK: [[OMP_ST:.omp.stride+]] = alloca i{{[0-9]+}}, 558 // CHECK: [[OMP_IS_LAST:.omp.is_last+]] = alloca i{{[0-9]+}}, 559 560 // private alloca's 561 // CHECK: [[VEC_PRIV:%.+]] = alloca [2 x i{{[0-9]+}}], 562 // CHECK: [[S_ARR_PRIV:%.+]] = alloca [2 x [[S_INT_TY]]], 563 // CHECK: [[VAR_PRIV:%.+]] = alloca [[S_INT_TY]], 564 // CHECK: [[TMP_PRIV:%.+]] = alloca [[S_INT_TY]]*, 565 566 // CHECK: store i{{[0-9]+}}* [[GTID_ADDR]], i{{[0-9]+}}** [[GTID_ADDR_REF:%.+]] 567 568 // init addr alloca's with input values 569 // CHECK-DAG: store {{.+}} [[VEC_IN]], {{.+}} [[VEC_ADDR]], 570 // CHECK-DAG: store {{.+}} [[T_VAR_IN]], {{.+}}* [[T_VAR_ADDR]], 571 // CHECK-DAG: store {{.+}} [[S_ARR_IN]], {{.+}} [[S_ARR_ADDR]], 572 // CHECK-DAG: store {{.+}} [[VAR_IN]], {{.+}} [[VAR_ADDR]], 573 574 // init private alloca's with addr alloca's 575 // vec 576 // CHECK-DAG: [[VEC_ADDR_VAL:%.+]] = load {{.+}}*, {{.+}}** [[VEC_ADDR]], 577 // CHECK-DAG: [[VEC_PRIV_BCAST:%.+]] = bitcast {{.+}} [[VEC_PRIV]] to 578 // CHECK-DAG: [[VEC_ADDR_BCAST:%.+]] = bitcast {{.+}} [[VEC_ADDR_VAL]] to 579 // CHECK-DAG: call void @llvm.memcpy{{.+}}({{.+}}* [[VEC_PRIV_BCAST]], {{.+}}* [[VEC_ADDR_BCAST]], 580 581 // s_arr 582 // CHECK-DAG: [[S_ARR_ADDR_VAL:%.+]] = load {{.+}}*, {{.+}}** [[S_ARR_ADDR]], 583 // CHECK-DAG: [[S_ARR_BGN:%.+]] = getelementptr {{.+}}, {{.+}}* [[S_ARR_PRIV]], 584 // CHECK-DAG: [[S_ARR_ADDR_BCAST:%.+]] = bitcast {{.+}}* [[S_ARR_ADDR_VAL]] to 585 // CHECK-DAG: [[S_ARR_BGN_GEP:%.+]] = getelementptr {{.+}}, {{.+}}* [[S_ARR_BGN]], 586 // CHECK-DAG: [[S_ARR_EMPTY:%.+]] = icmp {{.+}} [[S_ARR_BGN]], [[S_ARR_BGN_GEP]] 587 // CHECK-DAG: br {{.+}} [[S_ARR_EMPTY]], label %[[CPY_DONE:.+]], label %[[CPY_BODY:.+]] 588 // CHECK-DAG: [[CPY_BODY]]: 589 // CHECK-DAG: call void @llvm.memcpy{{.+}}( 590 // CHECK-DAG: [[CPY_DONE]]: 591 592 // var 593 // CHECK-DAG: [[VAR_ADDR_REF:%.+]] = load {{.+}}*, {{.+}}* [[VAR_ADDR]], 594 // CHECK-DAG: [[VAR_PRIV_BCAST:%.+]] = bitcast {{.+}}* [[VAR_PRIV]] to 595 // CHECK-DAG: [[VAR_ADDR_BCAST:%.+]] = bitcast {{.+}}* [[VAR_ADDR_REF]] to 596 // CHECK-DAG: call void @llvm.memcpy.{{.+}}({{.+}}* [[VAR_PRIV_BCAST]], {{.+}}* [[VAR_ADDR_BCAST]], 597 // CHECK-DAG: store {{.+}}* [[VAR_PRIV]], {{.+}}** [[TMP_PRIV]], 598 599 // CHECK: call void @__kmpc_for_static_init_4( 600 // CHECK: call void @__kmpc_for_static_fini( 601 602 // call destructors: var.. 603 // CHECK-DAG: call {{.+}} [[S_INT_TY_DEF_DESTR]]([[S_INT_TY]]* [[VAR_PRIV]]) 604 605 // ..and s_arr 606 // CHECK: {{.+}}: 607 // CHECK: [[S_ARR_EL_PAST:%.+]] = phi [[S_INT_TY]]* 608 // CHECK: [[S_ARR_PRIV_ITEM:%.+]] = getelementptr {{.+}}, {{.+}} [[S_ARR_EL_PAST]], 609 // CHECK: call {{.*}} [[S_INT_TY_DEF_DESTR]]([[S_INT_TY]]* [[S_ARR_PRIV_ITEM]]) 610 611 // CHECK: ret void 612 613 #endif 614