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 -DLAMBDA -verify -fopenmp-simd -x c++ -std=c++11 -triple powerpc64le-unknown-unknown -fopenmp-targets=powerpc64le-ibm-linux-gnu -emit-llvm %s -o - | FileCheck --check-prefix SIMD-ONLY0 %s 9 // RUN: %clang_cc1 -DLAMBDA -fopenmp-simd -x c++ -std=c++11 -triple powerpc64le-unknown-unknown -fopenmp-targets=powerpc64le-ibm-linux-gnu -emit-pch -o %t %s 10 // RUN: %clang_cc1 -DLAMBDA -fopenmp-simd -x c++ -triple powerpc64le-unknown-unknown -fopenmp-targets=powerpc64le-ibm-linux-gnu -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck --check-prefix SIMD-ONLY0 %s 11 // RUN: %clang_cc1 -DLAMBDA -verify -fopenmp-simd -x c++ -std=c++11 -triple i386-unknown-unknown -fopenmp-targets=i386-pc-linux-gnu -emit-llvm %s -o - | FileCheck --check-prefix SIMD-ONLY0 %s 12 // RUN: %clang_cc1 -DLAMBDA -fopenmp-simd -x c++ -std=c++11 -triple i386-unknown-unknown -fopenmp-targets=i386-pc-linux-gnu -emit-pch -o %t %s 13 // RUN: %clang_cc1 -DLAMBDA -fopenmp-simd -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 --check-prefix SIMD-ONLY0 %s 14 // SIMD-ONLY0-NOT: {{__kmpc|__tgt}} 15 16 // 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 17 // RUN: %clang_cc1 -fopenmp -x c++ -std=c++11 -triple powerpc64le-unknown-unknown -fopenmp-targets=powerpc64le-ibm-linux-gnu -emit-pch -o %t %s 18 // 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 19 // 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 20 // RUN: %clang_cc1 -fopenmp -x c++ -std=c++11 -triple i386-unknown-unknown -fopenmp-targets=i386-pc-linux-gnu -emit-pch -o %t %s 21 // 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 22 23 // RUN: %clang_cc1 -verify -fopenmp-simd -x c++ -std=c++11 -triple powerpc64le-unknown-unknown -fopenmp-targets=powerpc64le-ibm-linux-gnu -emit-llvm %s -o - | FileCheck --check-prefix SIMD-ONLY1 %s 24 // RUN: %clang_cc1 -fopenmp-simd -x c++ -std=c++11 -triple powerpc64le-unknown-unknown -fopenmp-targets=powerpc64le-ibm-linux-gnu -emit-pch -o %t %s 25 // RUN: %clang_cc1 -fopenmp-simd -x c++ -triple powerpc64le-unknown-unknown -fopenmp-targets=powerpc64le-ibm-linux-gnu -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck --check-prefix SIMD-ONLY1 %s 26 // RUN: %clang_cc1 -verify -fopenmp-simd -x c++ -std=c++11 -triple i386-unknown-unknown -fopenmp-targets=i386-pc-linux-gnu -emit-llvm %s -o - | FileCheck --check-prefix SIMD-ONLY1 %s 27 // RUN: %clang_cc1 -fopenmp-simd -x c++ -std=c++11 -triple i386-unknown-unknown -fopenmp-targets=i386-pc-linux-gnu -emit-pch -o %t %s 28 // RUN: %clang_cc1 -fopenmp-simd -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 --check-prefix SIMD-ONLY1 %s 29 // SIMD-ONLY1-NOT: {{__kmpc|__tgt}} 30 // expected-no-diagnostics 31 #ifndef HEADER 32 #define HEADER 33 34 template <class T> 35 struct S { 36 T f; 37 S(T a) : f(a) {} 38 S() : f() {} 39 operator T() { return T(); } 40 ~S() {} 41 }; 42 43 // CHECK: [[S_FLOAT_TY:%.+]] = type { float } 44 // CHECK: [[S_INT_TY:%.+]] = type { i{{[0-9]+}} } 45 template <typename T> 46 T tmain() { 47 S<T> test; 48 T t_var = T(); 49 T vec[] = {1, 2}; 50 S<T> s_arr[] = {1, 2}; 51 S<T> &var = test; 52 #pragma omp target teams distribute simd lastprivate(t_var, vec, s_arr, s_arr, var, var) 53 for (int i = 0; i < 2; ++i) { 54 vec[i] = t_var; 55 s_arr[i] = var; 56 } 57 return T(); 58 } 59 60 int main() { 61 static int svar; 62 volatile double g; 63 volatile double &g1 = g; 64 65 #ifdef LAMBDA 66 // LAMBDA-LABEL: @main 67 // LAMBDA: call{{.*}} void [[OUTER_LAMBDA:@.+]]( 68 [&]() { 69 static float sfvar; 70 // LAMBDA: define{{.*}} internal{{.*}} void [[OUTER_LAMBDA]]( 71 // LAMBDA: call i{{[0-9]+}} @__tgt_target_teams( 72 // LAMBDA: call void [[OFFLOADING_FUN:@.+]]( 73 74 // LAMBDA: define{{.+}} void [[OFFLOADING_FUN]]( 75 // LAMBDA: call {{.*}}void {{.+}} @__kmpc_fork_teams({{.+}}, i32 4, {{.+}}* [[OMP_OUTLINED:@.+]] to {{.+}}) 76 #pragma omp target teams distribute simd lastprivate(g, g1, svar, sfvar) 77 for (int i = 0; i < 2; ++i) { 78 // LAMBDA-64: define{{.*}} internal{{.*}} void [[OMP_OUTLINED]](i32* noalias %{{.+}}, i32* noalias %{{.+}}, i64 [[G_IN:%.+]], i64 [[G1_IN:%.+]], i64 [[SVAR_IN:%.+]], i64 [[SFVAR_IN:%.+]]) 79 // LAMBDA-32: define{{.*}} internal{{.*}} void [[OMP_OUTLINED]](i32* noalias %{{.+}}, i32* noalias %{{.+}}, double*{{.+}} [[G_IN:%.+]], double*{{.+}} [[G1_IN:%.+]], i32 [[SVAR_IN:%.+]], i32 [[SFVAR_IN:%.+]]) 80 // LAMBDA-64: [[G_PRIVATE_ADDR:%.+]] = alloca i64, 81 // LAMBDA-32: [[G_PRIVATE_ADDR:%.+]] = alloca double*, 82 // LAMBDA-64: [[G1_PRIVATE_ADDR:%.+]] = alloca i64, 83 // LAMBDA-32: [[G1_PRIVATE_ADDR:%.+]] = alloca double*, 84 // LAMBDA-64: [[SVAR_PRIVATE_ADDR:%.+]] = alloca i64, 85 // LAMBDA-32: [[SVAR_PRIVATE_ADDR:%.+]] = alloca i32, 86 // LAMBDA-64: [[SFVAR_PRIVATE_ADDR:%.+]] = alloca i64, 87 // LAMBDA-32: [[SFVAR_PRIVATE_ADDR:%.+]] = alloca i32, 88 // loop variables 89 // LAMBDA: {{.+}} = alloca i{{[0-9]+}}, 90 // LAMBDA: {{.+}} = alloca i{{[0-9]+}}, 91 // LAMBDA: {{.+}} = alloca i{{[0-9]+}}, 92 // LAMBDA: {{.+}} = alloca i{{[0-9]+}}, 93 // LAMBDA: {{.+}} = alloca i{{[0-9]+}}, 94 // LAMBDA: [[OMP_IS_LAST:%.+]] = alloca i{{[0-9]+}}, 95 // LAMBDA: [[G_PRIVATE:%.+]] = alloca double, 96 // LAMBDA: [[G1_PRIVATE:%.+]] = alloca double, 97 // LAMBDA: [[TMP_G1_PRIVATE:%.+]] = alloca double*, 98 // LAMBDA: [[SVAR_PRIVATE:%.+]] = alloca i{{[0-9]+}}, 99 // LAMBDA: [[SFVAR_PRIVATE:%.+]] = alloca float, 100 // LAMBDA-64: store i64 [[G_IN]], i64* [[G_PRIVATE_ADDR]], 101 // LAMBDA-32: store double* [[G_IN]], double** [[G_PRIVATE_ADDR]], 102 // LAMBDA-64: store i64 [[G1_IN]], i64* [[G1_PRIVATE_ADDR]], 103 // LAMBDA-32: store double* [[G1_IN]], double** [[G1_PRIVATE_ADDR]], 104 // LAMBDA-64: store i64 [[SVAR_IN]], i64* [[SVAR_PRIVATE_ADDR]], 105 // LAMBDA-32: store i32 [[SVAR_IN]], i32* [[SVAR_PRIVATE_ADDR]], 106 // LAMBDA-64: store i64 [[SFVAR_IN]], i64* [[SFVAR_PRIVATE_ADDR]], 107 // LAMBDA-32: store i32 [[SFVAR_IN]], i32* [[SFVAR_PRIVATE_ADDR]], 108 109 // init private variables 110 // LAMBDA-64: [[G_IN_REF:%.+]] = bitcast i64* [[G_PRIVATE_ADDR]] to double* 111 // LAMBDA-32: [[G_IN_REF:%.+]] = load double*, double** [[G_PRIVATE_ADDR]], 112 // LAMBDA-64: [[G1_IN_REF:%.+]] = bitcast i64* [[G1_PRIVATE_ADDR]] to double* 113 // LAMBDA-64: store double* [[G1_IN_REF]], double** [[G1_IN_ADDR_REF:%.+]], 114 // LAMBDA-64: [[SVAR_IN_REF:%.+]] = bitcast i64* [[SVAR_PRIVATE_ADDR]] to i32* 115 // LAMBDA-64: [[SFVAR_IN_REF:%.+]] = bitcast i64* [[SFVAR_PRIVATE_ADDR]] to float* 116 // LAMBDA-32: [[SFVAR_IN_REF:%.+]] = bitcast i32* [[SFVAR_PRIVATE_ADDR]] to float* 117 // LAMBDA-64: [[G1_IN_REF:%.+]] = load double*, double** [[G1_IN_ADDR_REF]], 118 // LAMBDA-32: [[G1_IN_REF:%.+]] = load double*, double** [[G1_PRIVATE_ADDR]], 119 // LAMBDA: store double* [[G1_PRIVATE]], double** [[TMP_G1_PRIVATE]], 120 g = 1; 121 g1 = 1; 122 svar = 3; 123 sfvar = 4.0; 124 // LAMBDA: call {{.*}}void @__kmpc_for_static_init_4( 125 // LAMBDA: store double 1.0{{.+}}, double* [[G_PRIVATE]], 126 // LAMBDA: [[TMP_G1_REF:%.+]] = load double*, double** [[TMP_G1_PRIVATE]], 127 // LAMBDA: store{{.+}} double 1.0{{.+}}, double* [[TMP_G1_REF]], 128 // LAMBDA: store i{{[0-9]+}} 3, i{{[0-9]+}}* [[SVAR_PRIVATE]], 129 // LAMBDA: store float 4.0{{.+}}, float* [[SFVAR_PRIVATE]], 130 // LAMBDA: [[G_PRIVATE_ADDR_REF:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* [[ARG:%.+]], i{{[0-9]+}} 0, i{{[0-9]+}} 0 131 // LAMBDA: store double* [[G_PRIVATE]], double** [[G_PRIVATE_ADDR_REF]], 132 // LAMBDA: [[TMP_PRIVATE_ADDR_REF:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* [[ARG:%.+]], i{{[0-9]+}} 0, i{{[0-9]+}} 1 133 // LAMBDA: [[G1_PRIVATE_ADDR_FROM_TMP:%.+]] = load double*, double** [[TMP_G1_PRIVATE]], 134 // LAMBDA: store double* [[G1_PRIVATE_ADDR_FROM_TMP]], double** [[TMP_PRIVATE_ADDR_REF]], 135 // LAMBDA: [[SVAR_PRIVATE_ADDR_REF:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* [[ARG:%.+]], i{{[0-9]+}} 0, i{{[0-9]+}} 2 136 // LAMBDA: store i{{[0-9]+}}* [[SVAR_PRIVATE]], i{{[0-9]+}}** [[SVAR_PRIVATE_ADDR_REF]] 137 // LAMBDA: [[SFVAR_PRIVATE_ADDR_REF:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* [[ARG:%.+]], i{{[0-9]+}} 0, i{{[0-9]+}} 3 138 // LAMBDA: store float* [[SFVAR_PRIVATE]], float** [[SFVAR_PRIVATE_ADDR_REF]] 139 // LAMBDA: call{{.*}} void [[INNER_LAMBDA:@.+]](%{{.+}}* [[ARG]]) 140 // LAMBDA: call {{.*}}void @__kmpc_for_static_fini( 141 // LAMBDA: store i32 2, i32* % 142 // LAMBDA: [[OMP_IS_LAST_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[OMP_IS_LAST]], 143 // LAMBDA: [[IS_LAST_IT:%.+]] = icmp ne i{{[0-9]+}} [[OMP_IS_LAST_VAL]], 0 144 // LAMBDA: br i1 [[IS_LAST_IT]], label %[[OMP_LASTPRIV_BLOCK:.+]], label %[[OMP_LASTPRIV_DONE:.+]] 145 146 // LAMBDA: [[OMP_LASTPRIV_BLOCK]]: 147 // LAMBDA: [[G_PRIV_VAL:%.+]] = load double, double* [[G_PRIVATE]], 148 // LAMBDA: store{{.*}} double [[G_PRIV_VAL]], double* [[G_IN_REF]], 149 // LAMBDA: [[TMP_G1_PRIV_REF:%.+]] = load double*, double** [[TMP_G1_PRIVATE]], 150 // LAMBDA: [[TMP_G1_PRIV_VAL:%.+]] = load double, double* [[TMP_G1_PRIV_REF]], 151 // LAMBDA: store{{.*}} double [[TMP_G1_PRIV_VAL]], double* [[G1_IN_REF]], 152 153 // LAMBDA: [[SVAR_PRIV_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[SVAR_PRIVATE]], 154 // LAMBDA-64: store i{{[0-9]+}} [[SVAR_PRIV_VAL]], i{{[0-9]+}}* [[SVAR_IN_REF]], 155 // LAMBDA-32: store i{{[0-9]+}} [[SVAR_PRIV_VAL]], i{{[0-9]+}}* [[SVAR_PRIVATE_ADDR]], 156 // LAMBDA: [[SFVAR_PRIV_VAL:%.+]] = load float, float* [[SFVAR_PRIVATE]], 157 // LAMBDA: store float [[SFVAR_PRIV_VAL]], float* [[SFVAR_IN_REF]], 158 // LAMBDA: br label %[[OMP_LASTPRIV_DONE]] 159 // LAMBDA: [[OMP_LASTPRIV_DONE]]: 160 // LAMBDA: ret 161 [&]() { 162 // LAMBDA: define {{.+}} void [[INNER_LAMBDA]](%{{.+}}* [[ARG_PTR:%.+]]) 163 // LAMBDA: store %{{.+}}* [[ARG_PTR]], %{{.+}}** [[ARG_PTR_REF:%.+]], 164 g = 2; 165 g1 = 2; 166 svar = 4; 167 sfvar = 8.0; 168 // LAMBDA: [[ARG_PTR:%.+]] = load %{{.+}}*, %{{.+}}** [[ARG_PTR_REF]] 169 // LAMBDA: [[G_PTR_REF:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* [[ARG_PTR]], i{{[0-9]+}} 0, i{{[0-9]+}} 0 170 // LAMBDA: [[G_REF:%.+]] = load double*, double** [[G_PTR_REF]] 171 // LAMBDA: store double 2.0{{.+}}, double* [[G_REF]] 172 173 // LAMBDA: [[TMP_PTR_REF:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* [[ARG_PTR]], i{{[0-9]+}} 0, i{{[0-9]+}} 1 174 // LAMBDA: [[G1_REF:%.+]] = load double*, double** [[TMP_PTR_REF]] 175 // LAMBDA: store double 2.0{{.+}}, double* [[G1_REF]], 176 // LAMBDA: [[SVAR_PTR_REF:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* [[ARG_PTR]], i{{[0-9]+}} 0, i{{[0-9]+}} 2 177 // LAMBDA: [[SVAR_REF:%.+]] = load i{{[0-9]+}}*, i{{[0-9]+}}** [[SVAR_PTR_REF]] 178 // LAMBDA: store i{{[0-9]+}} 4, i{{[0-9]+}}* [[SVAR_REF]] 179 // LAMBDA: [[SFVAR_PTR_REF:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* [[ARG_PTR]], i{{[0-9]+}} 0, i{{[0-9]+}} 3 180 // LAMBDA: [[SFVAR_REF:%.+]] = load float*, float** [[SFVAR_PTR_REF]] 181 // LAMBDA: store float 8.0{{.+}}, float* [[SFVAR_REF]] 182 }(); 183 } 184 }(); 185 return 0; 186 #else 187 S<float> test; 188 int t_var = 0; 189 int vec[] = {1, 2}; 190 S<float> s_arr[] = {1, 2}; 191 S<float> &var = test; 192 193 #pragma omp target teams distribute simd lastprivate(t_var, vec, s_arr, s_arr, var, var, svar) 194 for (int i = 0; i < 2; ++i) { 195 vec[i] = t_var; 196 s_arr[i] = var; 197 } 198 int i; 199 200 return tmain<int>(); 201 #endif 202 } 203 204 // CHECK: define{{.*}} i{{[0-9]+}} @main() 205 // CHECK: [[TEST:%.+]] = alloca [[S_FLOAT_TY]], 206 // CHECK: call {{.*}} [[S_FLOAT_TY_DEF_CONSTR:@.+]]([[S_FLOAT_TY]]* [[TEST]]) 207 // CHECK: call i{{[0-9]+}} @__tgt_target_teams( 208 // CHECK: call void [[OFFLOAD_FUN:@.+]]([2 x i{{[0-9]+}}]* {{.+}}, i{{[0-9]+}} {{.+}}, [2 x [[S_FLOAT_TY]]]* {{.+}}, [[S_FLOAT_TY]]* {{.+}}, i{{[0-9]+}} {{.+}}) 209 // CHECK: ret 210 211 // CHECK: define{{.+}} [[OFFLOAD_FUN]]([2 x i{{[0-9]+}}]*{{.+}} {{.+}}, i{{[0-9]+}} {{.+}}, [2 x [[S_FLOAT_TY]]]*{{.+}} {{.+}}, [[S_FLOAT_TY]]*{{.+}} {{.+}}, i{{[0-9]+}} {{.+}}) 212 // CHECK: call void (%{{.+}}*, i{{[0-9]+}}, void (i{{[0-9]+}}*, i{{[0-9]+}}*, ...)*, ...) @__kmpc_fork_teams( 213 // CHECK: ret 214 // 215 // 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:%.+]]) 216 // CHECK: {{.+}} = alloca i{{[0-9]+}}*, 217 // CHECK: {{.+}} = alloca i{{[0-9]+}}*, 218 // CHECK: [[VEC_ADDR:%.+]] = alloca [2 x i{{[0-9]+}}]*, 219 // CHECK: [[T_VAR_ADDR:%.+]] = alloca i{{[0-9]+}}, 220 // CHECK: [[S_ARR_ADDR:%.+]] = alloca [2 x [[S_FLOAT_TY]]]*, 221 // CHECK: [[VAR_ADDR:%.+]] = alloca [[S_FLOAT_TY]]*, 222 // CHECK: [[SVAR_ADDR:%.+]] = alloca i{{[0-9]+}}, 223 // skip loop variables 224 // CHECK: {{.+}} = alloca i{{[0-9]+}}, 225 // CHECK: {{.+}} = alloca i{{[0-9]+}}, 226 // CHECK: {{.+}} = alloca i{{[0-9]+}}, 227 // CHECK: {{.+}} = alloca i{{[0-9]+}}, 228 // CHECK: {{.+}} = alloca i{{[0-9]+}}, 229 // CHECK: [[OMP_IS_LAST:%.+]] = alloca i{{[0-9]+}}, 230 // CHECK: [[T_VAR_PRIV:%.+]] = alloca i{{[0-9]+}}, 231 // CHECK: [[VEC_PRIV:%.+]] = alloca [2 x i{{[0-9]+}}], 232 // CHECK: [[S_ARR_PRIV:%.+]] = alloca [2 x [[S_FLOAT_TY]]], 233 // CHECK: [[VAR_PRIV:%.+]] = alloca [[S_FLOAT_TY]], 234 // CHECK: [[TMP_PRIV:%.+]] = alloca [[S_FLOAT_TY]]*, 235 // CHECK: [[S_VAR_PRIV:%.+]] = alloca i{{[0-9]+}}, 236 237 // copy from parameters to local address variables 238 // CHECK: store [2 x i{{[0-9]+}}]* [[VEC_IN]], [2 x i{{[0-9]+}}]** [[VEC_ADDR]], 239 // CHECK: store i{{[0-9]+}} [[T_VAR_IN]], i{{[0-9]+}}* [[T_VAR_ADDR]], 240 // CHECK: store [2 x [[S_FLOAT_TY]]]* [[S_ARR_IN]], [2 x [[S_FLOAT_TY]]]** [[S_ARR_ADDR]], 241 // CHECK: store [[S_FLOAT_TY]]* [[VAR_IN]], [[S_FLOAT_TY]]** [[VAR_ADDR]], 242 // CHECK: store i{{[0-9]+}} [[S_VAR_IN]], i{{[0-9]+}}* [[SVAR_ADDR]], 243 244 // load content of local address variables 245 // CHECK: [[VEC_ADDR_REF:%.+]] = load [2 x i{{[0-9]+}}]*, [2 x i{{[0-9]+}}]** [[VEC_ADDR]], 246 // CHECK-64: [[T_VAR_ADDR_REF:%.+]] = bitcast i64* [[T_VAR_ADDR]] to i32* 247 // CHECK: [[S_ARR_ADDR_REF:%.+]] = load [2 x [[S_FLOAT_TY]]]*, [2 x [[S_FLOAT_TY]]]** [[S_ARR_ADDR]], 248 // CHECK-64: [[SVAR_ADDR_REF:%.+]] = bitcast i64* [[SVAR_ADDR]] to i32* 249 // CHECK: store i{{[0-9]+}} 0, i{{[0-9]+}}* [[OMP_IS_LAST]], 250 // CHECK: [[VAR_ADDR_REF:%.+]] = load {{.+}}, {{.+}} [[VAR_ADDR]], 251 // the distribute loop 252 // CHECK: call void @__kmpc_for_static_init_4( 253 // assignment: vec[i] = t_var; 254 // CHECK: [[T_VAR_PRIV_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[T_VAR_PRIV]], 255 // CHECK: [[VEC_PTR:%.+]] = getelementptr inbounds [2 x i{{[0-9]+}}], [2 x i{{[0-9]+}}]* [[VEC_PRIV]], i{{[0-9]+}} 0, i{{[0-9]+}} {{.+}} 256 // CHECK: store i{{[0-9]+}} [[T_VAR_PRIV_VAL]], i{{[0-9]+}}* [[VEC_PTR]], 257 258 // assignment: s_arr[i] = var; 259 // CHECK-DAG: [[S_ARR_PTR:%.+]] = getelementptr inbounds [2 x [[S_FLOAT_TY]]], [2 x [[S_FLOAT_TY]]]* [[S_ARR_PRIV]], 260 // CHECK-DAG: [[TMP_VAL:%.+]] = load [[S_FLOAT_TY]]*, [[S_FLOAT_TY]]** [[TMP_PRIV]], 261 // CHECK-DAG: [[S_ARR_PTR_BCAST:%.+]] = bitcast [[S_FLOAT_TY]]* [[S_ARR_PTR]] to i8* 262 // CHECK-DAG: [[TMP_VAL_BCAST:%.+]] = bitcast [[S_FLOAT_TY]]* [[TMP_VAL]] to i8* 263 // CHECK: call void @llvm.memcpy.{{.+}}(i8* align {{[0-9]+}} [[S_ARR_PTR_BCAST]], i8* align {{[0-9]+}} [[TMP_VAL_BCAST]], 264 // CHECK: call void @__kmpc_for_static_fini( 265 266 // lastprivates 267 // CHECK: [[OMP_IS_LAST_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[OMP_IS_LAST]], 268 // CHECK: [[IS_LAST_IT:%.+]] = icmp ne i{{[0-9]+}} [[OMP_IS_LAST_VAL]], 0 269 // CHECK: br i1 [[IS_LAST_IT]], label %[[OMP_LASTPRIV_BLOCK:.+]], label %[[OMP_LASTPRIV_DONE:.+]] 270 271 // CHECK: [[OMP_LASTPRIV_BLOCK]]: 272 // CHECK: [[T_VAR_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[T_VAR_PRIV]], 273 // CHECK-64: store i{{[0-9]+}} [[T_VAR_VAL]], i{{[0-9]+}}* [[T_VAR_ADDR_REF]], 274 // CHECK-32: store i{{[0-9]+}} [[T_VAR_VAL]], i{{[0-9]+}}* [[T_VAR_ADDR]], 275 // CHECK: [[BCAST_VEC_ADDR_REF:%.+]] = bitcast [2 x i{{[0-9]+}}]* [[VEC_ADDR_REF]] to i8* 276 // CHECK: [[BCAST_VEC_PRIV:%.+]] = bitcast [2 x i{{[0-9]+}}]* [[VEC_PRIV]] to i8* 277 // CHECK: call void @llvm.memcpy.{{.+}}(i8* align {{[0-9]+}} [[BCAST_VEC_ADDR_REF]], i8* align {{[0-9]+}} [[BCAST_VEC_PRIV]], 278 // 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]+}} 0 279 // CHECK: [[S_ARR_PRIV_BCAST:%.+]] = bitcast [2 x [[S_FLOAT_TY]]]* [[S_ARR_PRIV]] to [[S_FLOAT_TY]]* 280 // CHECK: [[S_ARR_BEGIN_GEP:%.+]] = getelementptr [[S_FLOAT_TY]], [[S_FLOAT_TY]]* [[S_ARR_BEGIN]], i{{[0-9]+}} 2 281 // CHECK: [[S_ARR_IS_EMPTY:%.+]] = icmp eq [[S_FLOAT_TY]]* [[S_ARR_BEGIN]], [[S_ARR_BEGIN_GEP]] 282 // CHECK: br i1 [[S_ARR_IS_EMPTY]], label %[[S_ARR_COPY_DONE:.+]], label %[[S_ARR_COPY_BLOCK:.+]] 283 // CHECK: [[S_ARR_COPY_BLOCK]]: 284 // CHECK: [[S_ARR_SRC_EL:%.+]] = phi [[S_FLOAT_TY]]*{{.+}} 285 // CHECK: [[S_ARR_DST_EL:%.+]] = phi [[S_FLOAT_TY]]*{{.+}} 286 // CHECK: [[S_ARR_DST_BCAST:%.+]] = bitcast [[S_FLOAT_TY]]* [[S_ARR_DST_EL]] to i8* 287 // CHECK: [[S_ARR_SRC_BCAST:%.+]] = bitcast [[S_FLOAT_TY]]* [[S_ARR_SRC_EL]] to i8* 288 // CHECK: call void @llvm.memcpy.{{.+}}(i8* align {{[0-9]+}} [[S_ARR_DST_BCAST]], i8* align {{[0-9]+}} [[S_ARR_SRC_BCAST]]{{.+}}) 289 // CHECK: [[S_ARR_DST_NEXT:%.+]] = getelementptr [[S_FLOAT_TY]], [[S_FLOAT_TY]]* [[S_ARR_DST_EL]], i{{[0-9]+}} 1 290 // CHECK: [[S_ARR_SRC_NEXT:%.+]] = getelementptr{{.+}} 291 // CHECK: [[CPY_IS_FINISHED:%.+]] = icmp eq [[S_FLOAT_TY]]* [[S_ARR_DST_NEXT]], [[S_ARR_BEGIN_GEP]] 292 // CHECK: br i1 [[CPY_IS_FINISHED]], label %[[S_ARR_COPY_DONE]], label %[[S_ARR_COPY_BLOCK]] 293 // CHECK: [[S_ARR_COPY_DONE]]: 294 // CHECK: [[TMP_VAL1:%.+]] = load [[S_FLOAT_TY]]*, [[S_FLOAT_TY]]** [[TMP_PRIV]], 295 // CHECK: [[VAR_ADDR_REF_BCAST:%.+]] = bitcast [[S_FLOAT_TY]]* [[VAR_ADDR_REF]] to i8* 296 // CHECK: [[TMP_VAL1_BCAST:%.+]] = bitcast [[S_FLOAT_TY]]* [[TMP_VAL1]] to i8* 297 // CHECK: call void @llvm.memcpy.{{.+}}(i8* align {{[0-9]+}} [[VAR_ADDR_REF_BCAST]], i8* align {{[0-9]+}} [[TMP_VAL1_BCAST]],{{.+}}) 298 // CHECK: [[SVAR_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[S_VAR_PRIV]], 299 // CHECK-64: store i{{[0-9]+}} [[SVAR_VAL]], i{{[0-9]+}}* [[SVAR_ADDR_REF]], 300 // CHECK-32: store i{{[0-9]+}} [[SVAR_VAL]], i{{[0-9]+}}* [[SVAR_ADDR]], 301 // CHECK: ret void 302 303 // template tmain 304 // CHECK: define{{.*}} i{{[0-9]+}} [[TMAIN_INT:@.+]]() 305 // CHECK: [[TEST:%.+]] = alloca [[S_INT_TY]], 306 // CHECK: call {{.*}} [[S_INT_TY_DEF_CONSTR:@.+]]([[S_INT_TY]]* [[TEST]]) 307 // CHECK: call i{{[0-9]+}} @__tgt_target_teams( 308 // CHECK: call void [[OFFLOAD_FUN_1:@.+]]([2 x i{{[0-9]+}}]* {{.+}}, i{{[0-9]+}} {{.+}}, [2 x [[S_INT_TY]]]* {{.+}}, [[S_INT_TY]]* {{.+}}) 309 // CHECK: ret 310 311 312 // CHECK: define internal void [[OFFLOAD_FUN_1]]( 313 // CHECK: call void (%{{.+}}*, i{{[0-9]+}}, void (i{{[0-9]+}}*, i{{[0-9]+}}*, ...)*, ...) @__kmpc_fork_teams(%{{.+}}* @{{.+}}, i{{[0-9]+}} 4, 314 // CHECK: ret 315 316 // 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:%.+]]) 317 // skip alloca of global_tid and bound_tid 318 // CHECK: {{.+}} = alloca i{{[0-9]+}}*, 319 // CHECK: {{.+}} = alloca i{{[0-9]+}}*, 320 // CHECK: [[VEC_ADDR1:%.+]] = alloca [2 x i{{[0-9]+}}]*, 321 // CHECK: [[T_VAR_ADDR1:%.+]] = alloca i{{[0-9]+}}, 322 // CHECK: [[S_ARR_ADDR1:%.+]] = alloca [2 x [[S_INT_TY]]]*, 323 // CHECK: [[VAR_ADDR1:%.+]] = alloca [[S_INT_TY]]*, 324 // skip loop variables 325 // CHECK: {{.+}} = alloca i{{[0-9]+}}, 326 // CHECK: {{.+}} = alloca i{{[0-9]+}}, 327 // CHECK: {{.+}} = alloca i{{[0-9]+}}, 328 // CHECK: {{.+}} = alloca i{{[0-9]+}}, 329 // CHECK: {{.+}} = alloca i{{[0-9]+}}, 330 // CHECK: [[OMP_IS_LAST1:%.+]] = alloca i{{[0-9]+}}, 331 // CHECK: [[T_VAR_PRIV1:%.+]] = alloca i{{[0-9]+}}, 332 // CHECK: [[VEC_PRIV1:%.+]] = alloca [2 x i{{[0-9]+}}], 333 // CHECK: [[S_ARR_PRIV1:%.+]] = alloca [2 x [[S_INT_TY]]], 334 // CHECK: [[VAR_PRIV1:%.+]] = alloca [[S_INT_TY]], 335 // CHECK: [[TMP_PRIV1:%.+]] = alloca [[S_INT_TY]]*, 336 337 // skip init of bound and global tid 338 // CHECK: store i{{[0-9]+}}* {{.*}}, 339 // CHECK: store i{{[0-9]+}}* {{.*}}, 340 // copy from parameters to local address variables 341 // CHECK: store [2 x i{{[0-9]+}}]* [[VEC_IN1]], [2 x i{{[0-9]+}}]** [[VEC_ADDR1]], 342 // CHECK: store i{{[0-9]+}} [[T_VAR_IN1]], i{{[0-9]+}}* [[T_VAR_ADDR1]], 343 // CHECK: store [2 x [[S_INT_TY]]]* [[S_ARR_IN1]], [2 x [[S_INT_TY]]]** [[S_ARR_ADDR1]], 344 // CHECK: store [[S_INT_TY]]* [[VAR_IN1]], [[S_INT_TY]]** [[VAR_ADDR1]], 345 346 // load content of local address variables 347 // CHECK: [[VEC_ADDR_REF1:%.+]] = load [2 x i{{[0-9]+}}]*, [2 x i{{[0-9]+}}]** [[VEC_ADDR1]], 348 // CHECK-64: [[T_VAR_ADDR_REF1:%.+]] = bitcast i64* [[T_VAR_ADDR1]] to i32* 349 // CHECK: [[S_ARR_ADDR_REF1:%.+]] = load [2 x [[S_INT_TY]]]*, [2 x [[S_INT_TY]]]** [[S_ARR_ADDR1]], 350 // CHECK-DAG: store i{{[0-9]+}} 0, i{{[0-9]+}}* [[OMP_IS_LAST1]], 351 // CHECK: [[VAR_ADDR1_REF:%.+]] = load [[S_INT_TY]]*, [[S_INT_TY]]** [[VAR_ADDR1]], 352 // CHECK-DAG: store [[S_INT_TY]]* [[VAR_PRIV1]], [[S_INT_TY]]** [[TMP_PRIV1]], 353 // CHECK: call void @__kmpc_for_static_init_4( 354 // assignment: vec[i] = t_var; 355 // CHECK: [[IV_VAL1:%.+]] = 356 // CHECK: [[T_VAR_PRIV_VAL1:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[T_VAR_PRIV1]], 357 // CHECK: [[VEC_PTR1:%.+]] = getelementptr inbounds [2 x i{{[0-9]+}}], [2 x i{{[0-9]+}}]* [[VEC_PRIV1]], i{{[0-9]+}} 0, i{{[0-9]+}} {{.+}} 358 // CHECK: store i{{[0-9]+}} [[T_VAR_PRIV_VAL1]], i{{[0-9]+}}* [[VEC_PTR1]], 359 360 // assignment: s_arr[i] = var; 361 // CHECK-DAG: [[S_ARR_PTR1:%.+]] = getelementptr inbounds [2 x [[S_INT_TY]]], [2 x [[S_INT_TY]]]* [[S_ARR_PRIV1]], 362 // CHECK-DAG: [[TMP_VAL1:%.+]] = load [[S_INT_TY]]*, [[S_INT_TY]]** [[TMP_PRIV1]], 363 // CHECK-DAG: [[S_ARR_PTR_BCAST1:%.+]] = bitcast [[S_INT_TY]]* [[S_ARR_PTR1]] to i8* 364 // CHECK-DAG: [[TMP_VAL_BCAST1:%.+]] = bitcast [[S_INT_TY]]* [[TMP_VAL1]] to i8* 365 // CHECK-DAG: call void @llvm.memcpy.{{.+}}(i8* align {{[0-9]+}} [[S_ARR_PTR_BCAST1]], i8* align {{[0-9]+}} [[TMP_VAL_BCAST1]], 366 // CHECK: call void @__kmpc_for_static_fini( 367 368 // lastprivates 369 // CHECK: [[OMP_IS_LAST_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[OMP_IS_LAST1]], 370 // CHECK: [[IS_LAST_IT:%.+]] = icmp ne i{{[0-9]+}} [[OMP_IS_LAST_VAL]], 0 371 // CHECK: br i1 [[IS_LAST_IT]], label %[[OMP_LASTPRIV_BLOCK:.+]], label %[[OMP_LASTPRIV_DONE:.+]] 372 373 // CHECK: [[OMP_LASTPRIV_BLOCK]]: 374 // CHECK: [[T_VAR_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[T_VAR_PRIV1]], 375 // CHECK-64: store i{{[0-9]+}} [[T_VAR_VAL]], i{{[0-9]+}}* [[T_VAR_ADDR_REF1]], 376 // CHECK-32: store i{{[0-9]+}} [[T_VAR_VAL]], i{{[0-9]+}}* [[T_VAR_ADDR1]], 377 // CHECK: [[BCAST_VEC_ADDR_REF:%.+]] = bitcast [2 x i{{[0-9]+}}]* [[VEC_ADDR_REF1]] to i8* 378 // CHECK: [[BCAST_VEC_PRIV:%.+]] = bitcast [2 x i{{[0-9]+}}]* [[VEC_PRIV1]] to i8* 379 // CHECK: call void @llvm.memcpy.{{.+}}(i8* align {{[0-9]+}} [[BCAST_VEC_ADDR_REF]], i8* align {{[0-9]+}} [[BCAST_VEC_PRIV]], 380 // 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]+}} 0 381 // CHECK: [[S_ARR_PRIV_BCAST:%.+]] = bitcast [2 x [[S_INT_TY]]]* [[S_ARR_PRIV1]] to [[S_INT_TY]]* 382 // CHECK: [[S_ARR_BEGIN_GEP:%.+]] = getelementptr [[S_INT_TY]], [[S_INT_TY]]* [[S_ARR_BEGIN]], i{{[0-9]+}} 2 383 // CHECK: [[S_ARR_IS_EMPTY:%.+]] = icmp eq [[S_INT_TY]]* [[S_ARR_BEGIN]], [[S_ARR_BEGIN_GEP]] 384 // CHECK: br i1 [[S_ARR_IS_EMPTY]], label %[[S_ARR_COPY_DONE:.+]], label %[[S_ARR_COPY_BLOCK:.+]] 385 // CHECK: [[S_ARR_COPY_BLOCK]]: 386 // CHECK: [[S_ARR_SRC_EL:%.+]] = phi [[S_INT_TY]]*{{.+}} 387 // CHECK: [[S_ARR_DST_EL:%.+]] = phi [[S_INT_TY]]*{{.+}} 388 // CHECK: [[S_ARR_DST_BCAST:%.+]] = bitcast [[S_INT_TY]]* [[S_ARR_DST_EL]] to i8* 389 // CHECK: [[S_ARR_SRC_BCAST:%.+]] = bitcast [[S_INT_TY]]* [[S_ARR_SRC_EL]] to i8* 390 // CHECK: call void @llvm.memcpy.{{.+}}(i8* align {{[0-9]+}} [[S_ARR_DST_BCAST]], i8* align {{[0-9]+}} [[S_ARR_SRC_BCAST]]{{.+}}) 391 // CHECK: [[S_ARR_DST_NEXT:%.+]] = getelementptr [[S_INT_TY]], [[S_INT_TY]]* [[S_ARR_DST_EL]], i{{[0-9]+}} 1 392 // CHECK: [[S_ARR_SRC_NEXT:%.+]] = getelementptr{{.+}} 393 // CHECK: [[CPY_IS_FINISHED:%.+]] = icmp eq [[S_INT_TY]]* [[S_ARR_DST_NEXT]], [[S_ARR_BEGIN_GEP]] 394 // CHECK: br i1 [[CPY_IS_FINISHED]], label %[[S_ARR_COPY_DONE]], label %[[S_ARR_COPY_BLOCK]] 395 // CHECK: [[S_ARR_COPY_DONE]]: 396 // CHECK: [[TMP_VAL1:%.+]] = load [[S_INT_TY]]*, [[S_INT_TY]]** [[TMP_PRIV1]], 397 // CHECK: [[VAR_ADDR_REF_BCAST:%.+]] = bitcast [[S_INT_TY]]* [[VAR_ADDR1_REF]] to i8* 398 // CHECK: [[TMP_VAL1_BCAST:%.+]] = bitcast [[S_INT_TY]]* [[TMP_VAL1]] to i8* 399 // CHECK: call void @llvm.memcpy.{{.+}}(i8* align {{[0-9]+}} [[VAR_ADDR_REF_BCAST]], i8* align {{[0-9]+}} [[TMP_VAL1_BCAST]],{{.+}}) 400 // CHECK: ret void 401 #endif 402