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