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 teams distribute lastprivate(t_var, vec, s_arr, s_arr, var, var) 37 for (int i = 0; i < 2; ++i) { 38 vec[i] = t_var; 39 s_arr[i] = var; 40 } 41 return T(); 42 } 43 44 int main() { 45 static int svar; 46 volatile double g; 47 volatile double &g1 = g; 48 49 #ifdef LAMBDA 50 // LAMBDA-LABEL: @main 51 // LAMBDA: call{{.*}} void [[OUTER_LAMBDA:@.+]]( 52 [&]() { 53 static float sfvar; 54 // LAMBDA: define{{.*}} internal{{.*}} void [[OUTER_LAMBDA]]( 55 // LAMBDA: call i{{[0-9]+}} @__tgt_target_teams( 56 // LAMBDA: call void [[OFFLOADING_FUN:@.+]]( 57 58 // LAMBDA: define{{.+}} void [[OFFLOADING_FUN]]( 59 // LAMBDA: call {{.*}}void {{.+}} @__kmpc_fork_teams({{.+}}, i32 4, {{.+}}* [[OMP_OUTLINED:@.+]] to {{.+}}) 60 #pragma omp target teams distribute lastprivate(g, g1, svar, sfvar) 61 for (int i = 0; i < 2; ++i) { 62 // LAMBDA-64: define{{.*}} internal{{.*}} void [[OMP_OUTLINED]](i32* noalias %{{.+}}, i32* noalias %{{.+}}, i64 [[G_IN:%.+]], i64 [[G1_IN:%.+]], i64 [[SVAR_IN:%.+]], i64 [[SFVAR_IN:%.+]]) 63 // LAMBDA-32: define{{.*}} internal{{.*}} void [[OMP_OUTLINED]](i32* noalias %{{.+}}, i32* noalias %{{.+}}, double*{{.+}} [[G_IN:%.+]], double*{{.+}} [[G1_IN:%.+]], i32 [[SVAR_IN:%.+]], i32 [[SFVAR_IN:%.+]]) 64 // LAMBDA-64: [[G_PRIVATE_ADDR:%.+]] = alloca i64, 65 // LAMBDA-32: [[G_PRIVATE_ADDR:%.+]] = alloca double*, 66 // LAMBDA-64: [[G1_PRIVATE_ADDR:%.+]] = alloca i64, 67 // LAMBDA-32: [[G1_PRIVATE_ADDR:%.+]] = alloca double*, 68 // LAMBDA-64: [[SVAR_PRIVATE_ADDR:%.+]] = alloca i64, 69 // LAMBDA-32: [[SVAR_PRIVATE_ADDR:%.+]] = alloca i32, 70 // LAMBDA-64: [[SFVAR_PRIVATE_ADDR:%.+]] = alloca i64, 71 // LAMBDA-32: [[SFVAR_PRIVATE_ADDR:%.+]] = alloca i32, 72 // loop variables 73 // LAMBDA: {{.+}} = alloca i{{[0-9]+}}, 74 // LAMBDA: {{.+}} = alloca i{{[0-9]+}}, 75 // LAMBDA: {{.+}} = alloca i{{[0-9]+}}, 76 // LAMBDA: {{.+}} = alloca i{{[0-9]+}}, 77 // LAMBDA: {{.+}} = alloca i{{[0-9]+}}, 78 // LAMBDA: [[OMP_IS_LAST:%.+]] = alloca i{{[0-9]+}}, 79 // LAMBDA: [[G_PRIVATE:%.+]] = alloca double, 80 // LAMBDA: [[G1_PRIVATE:%.+]] = alloca double, 81 // LAMBDA: [[TMP_G1_PRIVATE:%.+]] = alloca double*, 82 // LAMBDA: [[SVAR_PRIVATE:%.+]] = alloca i{{[0-9]+}}, 83 // LAMBDA: [[SFVAR_PRIVATE:%.+]] = alloca float, 84 // LAMBDA-64: store i64 [[G_IN]], i64* [[G_PRIVATE_ADDR]], 85 // LAMBDA-32: store double* [[G_IN]], double** [[G_PRIVATE_ADDR]], 86 // LAMBDA-64: store i64 [[G1_IN]], i64* [[G1_PRIVATE_ADDR]], 87 // LAMBDA-32: store double* [[G1_IN]], double** [[G1_PRIVATE_ADDR]], 88 // LAMBDA-64: store i64 [[SVAR_IN]], i64* [[SVAR_PRIVATE_ADDR]], 89 // LAMBDA-32: store i32 [[SVAR_IN]], i32* [[SVAR_PRIVATE_ADDR]], 90 // LAMBDA-64: store i64 [[SFVAR_IN]], i64* [[SFVAR_PRIVATE_ADDR]], 91 // LAMBDA-32: store i32 [[SFVAR_IN]], i32* [[SFVAR_PRIVATE_ADDR]], 92 93 // init private variables 94 // LAMBDA-64: [[G_IN_REF:%.+]] = bitcast i64* [[G_PRIVATE_ADDR]] to double* 95 // LAMBDA-32: [[G_IN_REF:%.+]] = load double*, double** [[G_PRIVATE_ADDR]], 96 // LAMBDA-64: [[G1_IN_REF:%.+]] = bitcast i64* [[G1_PRIVATE_ADDR]] to double* 97 // LAMBDA-64: store double* [[G1_IN_REF]], double** [[G1_IN_ADDR_REF:%.+]], 98 // LAMBDA-64: [[SVAR_IN_REF:%.+]] = bitcast i64* [[SVAR_PRIVATE_ADDR]] to i32* 99 // LAMBDA-64: [[SFVAR_IN_REF:%.+]] = bitcast i64* [[SFVAR_PRIVATE_ADDR]] to float* 100 // LAMBDA-32: [[SFVAR_IN_REF:%.+]] = bitcast i32* [[SFVAR_PRIVATE_ADDR]] to float* 101 // LAMBDA-64: [[G1_IN_REF:%.+]] = load double*, double** [[G1_IN_ADDR_REF]], 102 // LAMBDA-32: [[G1_IN_REF:%.+]] = load double*, double** [[G1_PRIVATE_ADDR]], 103 // LAMBDA: store double* [[G1_PRIVATE]], double** [[TMP_G1_PRIVATE]], 104 g = 1; 105 g1 = 1; 106 svar = 3; 107 sfvar = 4.0; 108 // LAMBDA: call {{.*}}void @__kmpc_for_static_init_4( 109 // LAMBDA: store double 1.0{{.+}}, double* [[G_PRIVATE]], 110 // LAMBDA: [[TMP_G1_REF:%.+]] = load double*, double** [[TMP_G1_PRIVATE]], 111 // LAMBDA: store{{.+}} double 1.0{{.+}}, double* [[TMP_G1_REF]], 112 // LAMBDA: store i{{[0-9]+}} 3, i{{[0-9]+}}* [[SVAR_PRIVATE]], 113 // LAMBDA: store float 4.0{{.+}}, float* [[SFVAR_PRIVATE]], 114 // LAMBDA: [[G_PRIVATE_ADDR_REF:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* [[ARG:%.+]], i{{[0-9]+}} 0, i{{[0-9]+}} 0 115 // LAMBDA: store double* [[G_PRIVATE]], double** [[G_PRIVATE_ADDR_REF]], 116 // LAMBDA: [[TMP_PRIVATE_ADDR_REF:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* [[ARG:%.+]], i{{[0-9]+}} 0, i{{[0-9]+}} 1 117 // LAMBDA: [[G1_PRIVATE_ADDR_FROM_TMP:%.+]] = load double*, double** [[TMP_G1_PRIVATE]], 118 // LAMBDA: store double* [[G1_PRIVATE_ADDR_FROM_TMP]], double** [[TMP_PRIVATE_ADDR_REF]], 119 // LAMBDA: [[SVAR_PRIVATE_ADDR_REF:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* [[ARG:%.+]], i{{[0-9]+}} 0, i{{[0-9]+}} 2 120 // LAMBDA: store i{{[0-9]+}}* [[SVAR_PRIVATE]], i{{[0-9]+}}** [[SVAR_PRIVATE_ADDR_REF]] 121 // LAMBDA: [[SFVAR_PRIVATE_ADDR_REF:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* [[ARG:%.+]], i{{[0-9]+}} 0, i{{[0-9]+}} 3 122 // LAMBDA: store float* [[SFVAR_PRIVATE]], float** [[SFVAR_PRIVATE_ADDR_REF]] 123 // LAMBDA: call{{.*}} void [[INNER_LAMBDA:@.+]](%{{.+}}* [[ARG]]) 124 // LAMBDA: call {{.*}}void @__kmpc_for_static_fini( 125 // LAMBDA: [[OMP_IS_LAST_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[OMP_IS_LAST]], 126 // LAMBDA: [[IS_LAST_IT:%.+]] = icmp ne i{{[0-9]+}} [[OMP_IS_LAST_VAL]], 0 127 // LAMBDA: br i1 [[IS_LAST_IT]], label %[[OMP_LASTPRIV_BLOCK:.+]], label %[[OMP_LASTPRIV_DONE:.+]] 128 129 // LAMBDA: [[OMP_LASTPRIV_BLOCK]]: 130 // LAMBDA: [[G_PRIV_VAL:%.+]] = load double, double* [[G_PRIVATE]], 131 // LAMBDA: store{{.*}} double [[G_PRIV_VAL]], double* [[G_IN_REF]], 132 // LAMBDA: [[TMP_G1_PRIV_REF:%.+]] = load double*, double** [[TMP_G1_PRIVATE]], 133 // LAMBDA: [[TMP_G1_PRIV_VAL:%.+]] = load double, double* [[TMP_G1_PRIV_REF]], 134 // LAMBDA: store{{.*}} double [[TMP_G1_PRIV_VAL]], double* [[G1_IN_REF]], 135 136 // LAMBDA: [[SVAR_PRIV_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[SVAR_PRIVATE]], 137 // LAMBDA-64: store i{{[0-9]+}} [[SVAR_PRIV_VAL]], i{{[0-9]+}}* [[SVAR_IN_REF]], 138 // LAMBDA-32: store i{{[0-9]+}} [[SVAR_PRIV_VAL]], i{{[0-9]+}}* [[SVAR_PRIVATE_ADDR]], 139 // LAMBDA: [[SFVAR_PRIV_VAL:%.+]] = load float, float* [[SFVAR_PRIVATE]], 140 // LAMBDA: store float [[SFVAR_PRIV_VAL]], float* [[SFVAR_IN_REF]], 141 // LAMBDA: br label %[[OMP_LASTPRIV_DONE]] 142 // LAMBDA: [[OMP_LASTPRIV_DONE]]: 143 // LAMBDA: ret 144 [&]() { 145 // LAMBDA: define {{.+}} void [[INNER_LAMBDA]](%{{.+}}* [[ARG_PTR:%.+]]) 146 // LAMBDA: store %{{.+}}* [[ARG_PTR]], %{{.+}}** [[ARG_PTR_REF:%.+]], 147 g = 2; 148 g1 = 2; 149 svar = 4; 150 sfvar = 8.0; 151 // LAMBDA: [[ARG_PTR:%.+]] = load %{{.+}}*, %{{.+}}** [[ARG_PTR_REF]] 152 // LAMBDA: [[G_PTR_REF:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* [[ARG_PTR]], i{{[0-9]+}} 0, i{{[0-9]+}} 0 153 // LAMBDA: [[G_REF:%.+]] = load double*, double** [[G_PTR_REF]] 154 // LAMBDA: store double 2.0{{.+}}, double* [[G_REF]] 155 156 // LAMBDA: [[TMP_PTR_REF:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* [[ARG_PTR]], i{{[0-9]+}} 0, i{{[0-9]+}} 1 157 // LAMBDA: [[G1_REF:%.+]] = load double*, double** [[TMP_PTR_REF]] 158 // LAMBDA: store double 2.0{{.+}}, double* [[G1_REF]], 159 // LAMBDA: [[SVAR_PTR_REF:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* [[ARG_PTR]], i{{[0-9]+}} 0, i{{[0-9]+}} 2 160 // LAMBDA: [[SVAR_REF:%.+]] = load i{{[0-9]+}}*, i{{[0-9]+}}** [[SVAR_PTR_REF]] 161 // LAMBDA: store i{{[0-9]+}} 4, i{{[0-9]+}}* [[SVAR_REF]] 162 // LAMBDA: [[SFVAR_PTR_REF:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* [[ARG_PTR]], i{{[0-9]+}} 0, i{{[0-9]+}} 3 163 // LAMBDA: [[SFVAR_REF:%.+]] = load float*, float** [[SFVAR_PTR_REF]] 164 // LAMBDA: store float 8.0{{.+}}, float* [[SFVAR_REF]] 165 }(); 166 } 167 }(); 168 return 0; 169 #else 170 S<float> test; 171 int t_var = 0; 172 int vec[] = {1, 2}; 173 S<float> s_arr[] = {1, 2}; 174 S<float> &var = test; 175 176 #pragma omp target teams distribute lastprivate(t_var, vec, s_arr, s_arr, var, var, svar) 177 for (int i = 0; i < 2; ++i) { 178 vec[i] = t_var; 179 s_arr[i] = var; 180 } 181 int i; 182 183 return tmain<int>(); 184 #endif 185 } 186 187 // CHECK: define{{.*}} i{{[0-9]+}} @main() 188 // CHECK: [[TEST:%.+]] = alloca [[S_FLOAT_TY]], 189 // CHECK: call {{.*}} [[S_FLOAT_TY_DEF_CONSTR:@.+]]([[S_FLOAT_TY]]* [[TEST]]) 190 // CHECK: call i{{[0-9]+}} @__tgt_target_teams( 191 // CHECK: call void [[OFFLOAD_FUN:@.+]]([2 x i{{[0-9]+}}]* {{.+}}, i{{[0-9]+}} {{.+}}, [2 x [[S_FLOAT_TY]]]* {{.+}}, [[S_FLOAT_TY]]* {{.+}}, i{{[0-9]+}} {{.+}}) 192 // CHECK: ret 193 194 // CHECK: define{{.+}} [[OFFLOAD_FUN]]([2 x i{{[0-9]+}}]*{{.+}} {{.+}}, i{{[0-9]+}} {{.+}}, [2 x [[S_FLOAT_TY]]]*{{.+}} {{.+}}, [[S_FLOAT_TY]]*{{.+}} {{.+}}, i{{[0-9]+}} {{.+}}) 195 // CHECK: call void (%{{.+}}*, i{{[0-9]+}}, void (i{{[0-9]+}}*, i{{[0-9]+}}*, ...)*, ...) @__kmpc_fork_teams( 196 // CHECK: ret 197 // 198 // 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:%.+]]) 199 // CHECK: {{.+}} = alloca i{{[0-9]+}}*, 200 // CHECK: {{.+}} = alloca i{{[0-9]+}}*, 201 // CHECK: [[VEC_ADDR:%.+]] = alloca [2 x i{{[0-9]+}}]*, 202 // CHECK: [[T_VAR_ADDR:%.+]] = alloca i{{[0-9]+}}, 203 // CHECK: [[S_ARR_ADDR:%.+]] = alloca [2 x [[S_FLOAT_TY]]]*, 204 // CHECK: [[VAR_ADDR:%.+]] = alloca [[S_FLOAT_TY]]*, 205 // CHECK: [[SVAR_ADDR:%.+]] = alloca i{{[0-9]+}}, 206 // skip loop variables 207 // CHECK: {{.+}} = alloca i{{[0-9]+}}, 208 // CHECK: {{.+}} = alloca i{{[0-9]+}}, 209 // CHECK: {{.+}} = alloca i{{[0-9]+}}, 210 // CHECK: {{.+}} = alloca i{{[0-9]+}}, 211 // CHECK: {{.+}} = alloca i{{[0-9]+}}, 212 // CHECK: [[OMP_IS_LAST:%.+]] = alloca i{{[0-9]+}}, 213 // CHECK: [[T_VAR_PRIV:%.+]] = alloca i{{[0-9]+}}, 214 // CHECK: [[VEC_PRIV:%.+]] = alloca [2 x i{{[0-9]+}}], 215 // CHECK: [[S_ARR_PRIV:%.+]] = alloca [2 x [[S_FLOAT_TY]]], 216 // CHECK: [[VAR_PRIV:%.+]] = alloca [[S_FLOAT_TY]], 217 // CHECK: [[TMP_PRIV:%.+]] = alloca [[S_FLOAT_TY]]*, 218 // CHECK: [[S_VAR_PRIV:%.+]] = alloca i{{[0-9]+}}, 219 220 // copy from parameters to local address variables 221 // CHECK: store [2 x i{{[0-9]+}}]* [[VEC_IN]], [2 x i{{[0-9]+}}]** [[VEC_ADDR]], 222 // CHECK: store i{{[0-9]+}} [[T_VAR_IN]], i{{[0-9]+}}* [[T_VAR_ADDR]], 223 // CHECK: store [2 x [[S_FLOAT_TY]]]* [[S_ARR_IN]], [2 x [[S_FLOAT_TY]]]** [[S_ARR_ADDR]], 224 // CHECK: store [[S_FLOAT_TY]]* [[VAR_IN]], [[S_FLOAT_TY]]** [[VAR_ADDR]], 225 // CHECK: store i{{[0-9]+}} [[S_VAR_IN]], i{{[0-9]+}}* [[SVAR_ADDR]], 226 227 // load content of local address variables 228 // CHECK: [[VEC_ADDR_REF:%.+]] = load [2 x i{{[0-9]+}}]*, [2 x i{{[0-9]+}}]** [[VEC_ADDR]], 229 // CHECK-64: [[T_VAR_ADDR_REF:%.+]] = bitcast i64* [[T_VAR_ADDR]] to i32* 230 // CHECK: [[S_ARR_ADDR_REF:%.+]] = load [2 x [[S_FLOAT_TY]]]*, [2 x [[S_FLOAT_TY]]]** [[S_ARR_ADDR]], 231 // CHECK-64: [[SVAR_ADDR_REF:%.+]] = bitcast i64* [[SVAR_ADDR]] to i32* 232 // CHECK: store i{{[0-9]+}} 0, i{{[0-9]+}}* [[OMP_IS_LAST]], 233 // CHECK: [[VAR_ADDR_REF:%.+]] = load {{.+}}, {{.+}} [[VAR_ADDR]], 234 // the distribute loop 235 // CHECK: call void @__kmpc_for_static_init_4( 236 // assignment: vec[i] = t_var; 237 // CHECK: [[T_VAR_PRIV_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[T_VAR_PRIV]], 238 // CHECK: [[VEC_PTR:%.+]] = getelementptr inbounds [2 x i{{[0-9]+}}], [2 x i{{[0-9]+}}]* [[VEC_PRIV]], i{{[0-9]+}} 0, i{{[0-9]+}} {{.+}} 239 // CHECK: store i{{[0-9]+}} [[T_VAR_PRIV_VAL]], i{{[0-9]+}}* [[VEC_PTR]], 240 241 // assignment: s_arr[i] = var; 242 // CHECK-DAG: [[S_ARR_PTR:%.+]] = getelementptr inbounds [2 x [[S_FLOAT_TY]]], [2 x [[S_FLOAT_TY]]]* [[S_ARR_PRIV]], 243 // CHECK-DAG: [[TMP_VAL:%.+]] = load [[S_FLOAT_TY]]*, [[S_FLOAT_TY]]** [[TMP_PRIV]], 244 // CHECK-DAG: [[S_ARR_PTR_BCAST:%.+]] = bitcast [[S_FLOAT_TY]]* [[S_ARR_PTR]] to i8* 245 // CHECK-DAG: [[TMP_VAL_BCAST:%.+]] = bitcast [[S_FLOAT_TY]]* [[TMP_VAL]] to i8* 246 // CHECK: call void @llvm.memcpy.{{.+}}(i8* [[S_ARR_PTR_BCAST]], i8* [[TMP_VAL_BCAST]], 247 // CHECK: call void @__kmpc_for_static_fini( 248 249 // lastprivates 250 // CHECK: [[OMP_IS_LAST_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[OMP_IS_LAST]], 251 // CHECK: [[IS_LAST_IT:%.+]] = icmp ne i{{[0-9]+}} [[OMP_IS_LAST_VAL]], 0 252 // CHECK: br i1 [[IS_LAST_IT]], label %[[OMP_LASTPRIV_BLOCK:.+]], label %[[OMP_LASTPRIV_DONE:.+]] 253 254 // CHECK: [[OMP_LASTPRIV_BLOCK]]: 255 // CHECK: [[T_VAR_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[T_VAR_PRIV]], 256 // CHECK-64: store i{{[0-9]+}} [[T_VAR_VAL]], i{{[0-9]+}}* [[T_VAR_ADDR_REF]], 257 // CHECK-32: store i{{[0-9]+}} [[T_VAR_VAL]], i{{[0-9]+}}* [[T_VAR_ADDR]], 258 // CHECK: [[BCAST_VEC_ADDR_REF:%.+]] = bitcast [2 x i{{[0-9]+}}]* [[VEC_ADDR_REF]] to i8* 259 // CHECK: [[BCAST_VEC_PRIV:%.+]] = bitcast [2 x i{{[0-9]+}}]* [[VEC_PRIV]] to i8* 260 // CHECK: call void @llvm.memcpy.{{.+}}(i8* [[BCAST_VEC_ADDR_REF]], i8* [[BCAST_VEC_PRIV]], 261 // 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]+}} 0 262 // CHECK: [[S_ARR_PRIV_BCAST:%.+]] = bitcast [2 x [[S_FLOAT_TY]]]* [[S_ARR_PRIV]] to [[S_FLOAT_TY]]* 263 // CHECK: [[S_ARR_BEGIN_GEP:%.+]] = getelementptr [[S_FLOAT_TY]], [[S_FLOAT_TY]]* [[S_ARR_BEGIN]], i{{[0-9]+}} 2 264 // CHECK: [[S_ARR_IS_EMPTY:%.+]] = icmp eq [[S_FLOAT_TY]]* [[S_ARR_BEGIN]], [[S_ARR_BEGIN_GEP]] 265 // CHECK: br i1 [[S_ARR_IS_EMPTY]], label %[[S_ARR_COPY_DONE:.+]], label %[[S_ARR_COPY_BLOCK:.+]] 266 // CHECK: [[S_ARR_COPY_BLOCK]]: 267 // CHECK: [[S_ARR_SRC_EL:%.+]] = phi [[S_FLOAT_TY]]*{{.+}} 268 // CHECK: [[S_ARR_DST_EL:%.+]] = phi [[S_FLOAT_TY]]*{{.+}} 269 // CHECK: [[S_ARR_DST_BCAST:%.+]] = bitcast [[S_FLOAT_TY]]* [[S_ARR_DST_EL]] to i8* 270 // CHECK: [[S_ARR_SRC_BCAST:%.+]] = bitcast [[S_FLOAT_TY]]* [[S_ARR_SRC_EL]] to i8* 271 // CHECK: call void @llvm.memcpy.{{.+}}(i8* [[S_ARR_DST_BCAST]], i8* [[S_ARR_SRC_BCAST]]{{.+}}) 272 // CHECK: [[S_ARR_DST_NEXT:%.+]] = getelementptr [[S_FLOAT_TY]], [[S_FLOAT_TY]]* [[S_ARR_DST_EL]], i{{[0-9]+}} 1 273 // CHECK: [[S_ARR_SRC_NEXT:%.+]] = getelementptr{{.+}} 274 // CHECK: [[CPY_IS_FINISHED:%.+]] = icmp eq [[S_FLOAT_TY]]* [[S_ARR_DST_NEXT]], [[S_ARR_BEGIN_GEP]] 275 // CHECK: br i1 [[CPY_IS_FINISHED]], label %[[S_ARR_COPY_DONE]], label %[[S_ARR_COPY_BLOCK]] 276 // CHECK: [[S_ARR_COPY_DONE]]: 277 // CHECK: [[TMP_VAL1:%.+]] = load [[S_FLOAT_TY]]*, [[S_FLOAT_TY]]** [[TMP_PRIV]], 278 // CHECK: [[VAR_ADDR_REF_BCAST:%.+]] = bitcast [[S_FLOAT_TY]]* [[VAR_ADDR_REF]] to i8* 279 // CHECK: [[TMP_VAL1_BCAST:%.+]] = bitcast [[S_FLOAT_TY]]* [[TMP_VAL1]] to i8* 280 // CHECK: call void @llvm.memcpy.{{.+}}(i8* [[VAR_ADDR_REF_BCAST]], i8* [[TMP_VAL1_BCAST]],{{.+}}) 281 // CHECK: [[SVAR_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[S_VAR_PRIV]], 282 // CHECK-64: store i{{[0-9]+}} [[SVAR_VAL]], i{{[0-9]+}}* [[SVAR_ADDR_REF]], 283 // CHECK-32: store i{{[0-9]+}} [[SVAR_VAL]], i{{[0-9]+}}* [[SVAR_ADDR]], 284 // CHECK: ret void 285 286 // template tmain 287 // CHECK: define{{.*}} i{{[0-9]+}} [[TMAIN_INT:@.+]]() 288 // CHECK: [[TEST:%.+]] = alloca [[S_INT_TY]], 289 // CHECK: call {{.*}} [[S_INT_TY_DEF_CONSTR:@.+]]([[S_INT_TY]]* [[TEST]]) 290 // CHECK: call i{{[0-9]+}} @__tgt_target_teams( 291 // CHECK: call void [[OFFLOAD_FUN_1:@.+]]([2 x i{{[0-9]+}}]* {{.+}}, i{{[0-9]+}} {{.+}}, [2 x [[S_INT_TY]]]* {{.+}}, [[S_INT_TY]]* {{.+}}) 292 // CHECK: ret 293 294 295 // CHECK: define internal void [[OFFLOAD_FUN_1]]( 296 // CHECK: call void (%{{.+}}*, i{{[0-9]+}}, void (i{{[0-9]+}}*, i{{[0-9]+}}*, ...)*, ...) @__kmpc_fork_teams(%{{.+}}* @{{.+}}, i{{[0-9]+}} 4, 297 // CHECK: ret 298 299 // 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:%.+]]) 300 // skip alloca of global_tid and bound_tid 301 // CHECK: {{.+}} = alloca i{{[0-9]+}}*, 302 // CHECK: {{.+}} = alloca i{{[0-9]+}}*, 303 // CHECK: [[VEC_ADDR1:%.+]] = alloca [2 x i{{[0-9]+}}]*, 304 // CHECK: [[T_VAR_ADDR1:%.+]] = alloca i{{[0-9]+}}, 305 // CHECK: [[S_ARR_ADDR1:%.+]] = alloca [2 x [[S_INT_TY]]]*, 306 // CHECK: [[VAR_ADDR1:%.+]] = alloca [[S_INT_TY]]*, 307 // skip loop variables 308 // CHECK: {{.+}} = alloca i{{[0-9]+}}, 309 // CHECK: {{.+}} = alloca i{{[0-9]+}}, 310 // CHECK: {{.+}} = alloca i{{[0-9]+}}, 311 // CHECK: {{.+}} = alloca i{{[0-9]+}}, 312 // CHECK: {{.+}} = alloca i{{[0-9]+}}, 313 // CHECK: [[OMP_IS_LAST1:%.+]] = alloca i{{[0-9]+}}, 314 // CHECK: [[T_VAR_PRIV1:%.+]] = alloca i{{[0-9]+}}, 315 // CHECK: [[VEC_PRIV1:%.+]] = alloca [2 x i{{[0-9]+}}], 316 // CHECK: [[S_ARR_PRIV1:%.+]] = alloca [2 x [[S_INT_TY]]], 317 // CHECK: [[VAR_PRIV1:%.+]] = alloca [[S_INT_TY]], 318 // CHECK: [[TMP_PRIV1:%.+]] = alloca [[S_INT_TY]]*, 319 320 // skip init of bound and global tid 321 // CHECK: store i{{[0-9]+}}* {{.*}}, 322 // CHECK: store i{{[0-9]+}}* {{.*}}, 323 // copy from parameters to local address variables 324 // CHECK: store [2 x i{{[0-9]+}}]* [[VEC_IN1]], [2 x i{{[0-9]+}}]** [[VEC_ADDR1]], 325 // CHECK: store i{{[0-9]+}} [[T_VAR_IN1]], i{{[0-9]+}}* [[T_VAR_ADDR1]], 326 // CHECK: store [2 x [[S_INT_TY]]]* [[S_ARR_IN1]], [2 x [[S_INT_TY]]]** [[S_ARR_ADDR1]], 327 // CHECK: store [[S_INT_TY]]* [[VAR_IN1]], [[S_INT_TY]]** [[VAR_ADDR1]], 328 329 // load content of local address variables 330 // CHECK: [[VEC_ADDR_REF1:%.+]] = load [2 x i{{[0-9]+}}]*, [2 x i{{[0-9]+}}]** [[VEC_ADDR1]], 331 // CHECK-64: [[T_VAR_ADDR_REF1:%.+]] = bitcast i64* [[T_VAR_ADDR1]] to i32* 332 // CHECK: [[S_ARR_ADDR_REF1:%.+]] = load [2 x [[S_INT_TY]]]*, [2 x [[S_INT_TY]]]** [[S_ARR_ADDR1]], 333 // CHECK-DAG: store i{{[0-9]+}} 0, i{{[0-9]+}}* [[OMP_IS_LAST1]], 334 // CHECK: [[VAR_ADDR1_REF:%.+]] = load [[S_INT_TY]]*, [[S_INT_TY]]** [[VAR_ADDR1]], 335 // CHECK-DAG: store [[S_INT_TY]]* [[VAR_PRIV1]], [[S_INT_TY]]** [[TMP_PRIV1]], 336 // CHECK: call void @__kmpc_for_static_init_4( 337 // assignment: vec[i] = t_var; 338 // CHECK: [[IV_VAL1:%.+]] = 339 // CHECK: [[T_VAR_PRIV_VAL1:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[T_VAR_PRIV1]], 340 // CHECK: [[VEC_PTR1:%.+]] = getelementptr inbounds [2 x i{{[0-9]+}}], [2 x i{{[0-9]+}}]* [[VEC_PRIV1]], i{{[0-9]+}} 0, i{{[0-9]+}} {{.+}} 341 // CHECK: store i{{[0-9]+}} [[T_VAR_PRIV_VAL1]], i{{[0-9]+}}* [[VEC_PTR1]], 342 343 // assignment: s_arr[i] = var; 344 // CHECK-DAG: [[S_ARR_PTR1:%.+]] = getelementptr inbounds [2 x [[S_INT_TY]]], [2 x [[S_INT_TY]]]* [[S_ARR_PRIV1]], 345 // CHECK-DAG: [[TMP_VAL1:%.+]] = load [[S_INT_TY]]*, [[S_INT_TY]]** [[TMP_PRIV1]], 346 // CHECK-DAG: [[S_ARR_PTR_BCAST1:%.+]] = bitcast [[S_INT_TY]]* [[S_ARR_PTR1]] to i8* 347 // CHECK-DAG: [[TMP_VAL_BCAST1:%.+]] = bitcast [[S_INT_TY]]* [[TMP_VAL1]] to i8* 348 // CHECK-DAG: call void @llvm.memcpy.{{.+}}(i8* [[S_ARR_PTR_BCAST1]], i8* [[TMP_VAL_BCAST1]], 349 // CHECK: call void @__kmpc_for_static_fini( 350 351 // lastprivates 352 // CHECK: [[OMP_IS_LAST_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[OMP_IS_LAST1]], 353 // CHECK: [[IS_LAST_IT:%.+]] = icmp ne i{{[0-9]+}} [[OMP_IS_LAST_VAL]], 0 354 // CHECK: br i1 [[IS_LAST_IT]], label %[[OMP_LASTPRIV_BLOCK:.+]], label %[[OMP_LASTPRIV_DONE:.+]] 355 356 // CHECK: [[OMP_LASTPRIV_BLOCK]]: 357 // CHECK: [[T_VAR_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[T_VAR_PRIV1]], 358 // CHECK-64: store i{{[0-9]+}} [[T_VAR_VAL]], i{{[0-9]+}}* [[T_VAR_ADDR_REF1]], 359 // CHECK-32: store i{{[0-9]+}} [[T_VAR_VAL]], i{{[0-9]+}}* [[T_VAR_ADDR1]], 360 // CHECK: [[BCAST_VEC_ADDR_REF:%.+]] = bitcast [2 x i{{[0-9]+}}]* [[VEC_ADDR_REF1]] to i8* 361 // CHECK: [[BCAST_VEC_PRIV:%.+]] = bitcast [2 x i{{[0-9]+}}]* [[VEC_PRIV1]] to i8* 362 // CHECK: call void @llvm.memcpy.{{.+}}(i8* [[BCAST_VEC_ADDR_REF]], i8* [[BCAST_VEC_PRIV]], 363 // 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]+}} 0 364 // CHECK: [[S_ARR_PRIV_BCAST:%.+]] = bitcast [2 x [[S_INT_TY]]]* [[S_ARR_PRIV1]] to [[S_INT_TY]]* 365 // CHECK: [[S_ARR_BEGIN_GEP:%.+]] = getelementptr [[S_INT_TY]], [[S_INT_TY]]* [[S_ARR_BEGIN]], i{{[0-9]+}} 2 366 // CHECK: [[S_ARR_IS_EMPTY:%.+]] = icmp eq [[S_INT_TY]]* [[S_ARR_BEGIN]], [[S_ARR_BEGIN_GEP]] 367 // CHECK: br i1 [[S_ARR_IS_EMPTY]], label %[[S_ARR_COPY_DONE:.+]], label %[[S_ARR_COPY_BLOCK:.+]] 368 // CHECK: [[S_ARR_COPY_BLOCK]]: 369 // CHECK: [[S_ARR_SRC_EL:%.+]] = phi [[S_INT_TY]]*{{.+}} 370 // CHECK: [[S_ARR_DST_EL:%.+]] = phi [[S_INT_TY]]*{{.+}} 371 // CHECK: [[S_ARR_DST_BCAST:%.+]] = bitcast [[S_INT_TY]]* [[S_ARR_DST_EL]] to i8* 372 // CHECK: [[S_ARR_SRC_BCAST:%.+]] = bitcast [[S_INT_TY]]* [[S_ARR_SRC_EL]] to i8* 373 // CHECK: call void @llvm.memcpy.{{.+}}(i8* [[S_ARR_DST_BCAST]], i8* [[S_ARR_SRC_BCAST]]{{.+}}) 374 // CHECK: [[S_ARR_DST_NEXT:%.+]] = getelementptr [[S_INT_TY]], [[S_INT_TY]]* [[S_ARR_DST_EL]], i{{[0-9]+}} 1 375 // CHECK: [[S_ARR_SRC_NEXT:%.+]] = getelementptr{{.+}} 376 // CHECK: [[CPY_IS_FINISHED:%.+]] = icmp eq [[S_INT_TY]]* [[S_ARR_DST_NEXT]], [[S_ARR_BEGIN_GEP]] 377 // CHECK: br i1 [[CPY_IS_FINISHED]], label %[[S_ARR_COPY_DONE]], label %[[S_ARR_COPY_BLOCK]] 378 // CHECK: [[S_ARR_COPY_DONE]]: 379 // CHECK: [[TMP_VAL1:%.+]] = load [[S_INT_TY]]*, [[S_INT_TY]]** [[TMP_PRIV1]], 380 // CHECK: [[VAR_ADDR_REF_BCAST:%.+]] = bitcast [[S_INT_TY]]* [[VAR_ADDR1_REF]] to i8* 381 // CHECK: [[TMP_VAL1_BCAST:%.+]] = bitcast [[S_INT_TY]]* [[TMP_VAL1]] to i8* 382 // CHECK: call void @llvm.memcpy.{{.+}}(i8* [[VAR_ADDR_REF_BCAST]], i8* [[TMP_VAL1_BCAST]],{{.+}}) 383 // CHECK: ret void 384 #endif 385