1 // RUN: %clang_cc1 -DLAMBDA -verify -fopenmp -x c++ -std=c++11 -triple powerpc64le-unknown-unknown -fopenmp-targets=powerpc64le-ibm-linux-gnu -emit-llvm %s -o - | FileCheck %s --check-prefix LAMBDA --check-prefix LAMBDA-64 2 // RUN: %clang_cc1 -DLAMBDA -fopenmp -x c++ -std=c++11 -triple powerpc64le-unknown-unknown -fopenmp-targets=powerpc64le-ibm-linux-gnu -emit-pch -o %t %s 3 // RUN: %clang_cc1 -DLAMBDA -fopenmp -x c++ -triple powerpc64le-unknown-unknown -fopenmp-targets=powerpc64le-ibm-linux-gnu -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix LAMBDA --check-prefix LAMBDA-64 4 // RUN: %clang_cc1 -DLAMBDA -verify -fopenmp -x c++ -std=c++11 -triple i386-unknown-unknown -fopenmp-targets=i386-pc-linux-gnu -emit-llvm %s -o - | FileCheck %s --check-prefix LAMBDA --check-prefix LAMBDA-32 5 // RUN: %clang_cc1 -DLAMBDA -fopenmp -x c++ -std=c++11 -triple i386-unknown-unknown -fopenmp-targets=i386-pc-linux-gnu -emit-pch -o %t %s 6 // RUN: %clang_cc1 -DLAMBDA -fopenmp -x c++ -std=c++11 -triple i386-unknown-unknown -fopenmp-targets=i386-pc-linux-gnu -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix LAMBDA --check-prefix LAMBDA-32 7 8 // RUN: %clang_cc1 -verify -fopenmp -x c++ -std=c++11 -triple powerpc64le-unknown-unknown -fopenmp-targets=powerpc64le-ibm-linux-gnu -emit-llvm %s -o - | FileCheck %s --check-prefix CHECK --check-prefix CHECK-64 9 // RUN: %clang_cc1 -fopenmp -x c++ -std=c++11 -triple powerpc64le-unknown-unknown -fopenmp-targets=powerpc64le-ibm-linux-gnu -emit-pch -o %t %s 10 // RUN: %clang_cc1 -fopenmp -x c++ -triple powerpc64le-unknown-unknown -fopenmp-targets=powerpc64le-ibm-linux-gnu -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CHECK --check-prefix CHECK-64 11 // RUN: %clang_cc1 -verify -fopenmp -x c++ -std=c++11 -triple i386-unknown-unknown -fopenmp-targets=i386-pc-linux-gnu -emit-llvm %s -o - | FileCheck %s --check-prefix CHECK --check-prefix CHECK-32 12 // RUN: %clang_cc1 -fopenmp -x c++ -std=c++11 -triple i386-unknown-unknown -fopenmp-targets=i386-pc-linux-gnu -emit-pch -o %t %s 13 // RUN: %clang_cc1 -fopenmp -x c++ -std=c++11 -triple i386-unknown-unknown -fopenmp-targets=i386-pc-linux-gnu -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s --check-prefix CHECK --check-prefix CHECK-32 14 // expected-no-diagnostics 15 #ifndef HEADER 16 #define HEADER 17 18 template <class T> 19 struct S { 20 T f; 21 S(T a) : f(a) {} 22 S() : f() {} 23 operator T() { return T(); } 24 ~S() {} 25 }; 26 27 // CHECK: [[S_FLOAT_TY:%.+]] = type { float } 28 // CHECK: [[S_INT_TY:%.+]] = type { i{{[0-9]+}} } 29 template <typename T> 30 T tmain() { 31 S<T> test; 32 T t_var = T(); 33 T vec[] = {1, 2}; 34 S<T> s_arr[] = {1, 2}; 35 S<T> &var = test; 36 #pragma omp target 37 #pragma omp teams 38 #pragma omp distribute 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({{.+}}, {{.+}}, {{.+}}* [[OMP_OUTLINED:@.+]] to {{.+}}) 62 #pragma omp target 63 #pragma omp teams 64 #pragma omp distribute 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 // Private alloca's for conversion 68 // LAMBDA: [[G_ADDR:%.+]] = alloca double*, 69 // LAMBDA: [[G1_ADDR:%.+]] = alloca double*, 70 // LAMBDA: [[SVAR_ADDR:%.+]] = alloca i{{[0-9]+}}*, 71 // LAMBDA: [[SFVAR_ADDR:%.+]] = alloca float*, 72 // LAMBDA: [[G1_REF:%.+]] = alloca double*, 73 74 // Actual private variables to be used in the body (tmp is used for the reference type) 75 // LAMBDA: [[G_PRIVATE:%.+]] = alloca double, 76 // LAMBDA: [[G1_PRIVATE:%.+]] = alloca double, 77 // LAMBDA: [[TMP_PRIVATE:%.+]] = alloca double*, 78 // LAMBDA: [[SVAR_PRIVATE:%.+]] = alloca i{{[0-9]+}}, 79 // LAMBDA: [[SFVAR_PRIVATE:%.+]] = alloca float, 80 81 // Store input parameter addresses into private alloca's for conversion 82 // LAMBDA: store double* [[G_IN]], double** [[G_ADDR]], 83 // LAMBDA: store double* [[G1_IN]], double** [[G1_ADDR]], 84 // LAMBDA: store i{{[0-9]+}}* [[SVAR_IN]], i{{[0-9]+}}** [[SVAR_ADDR]], 85 // LAMBDA: store float* [[SFVAR_IN]], float** [[SFVAR_ADDR]], 86 87 // LAMBDA-DAG: [[G_ADDR_VAL:%.+]] = load double*, double** [[G_ADDR]], 88 // LAMBDA-DAG: [[G1_ADDR_VAL:%.+]] = load double*, double** [[G1_ADDR]], 89 // LAMBDA-DAG: [[SVAR_ADDR_VAL:%.+]] = load i{{[0-9]+}}*, i{{[0-9]+}}** [[SVAR_ADDR]], 90 // LAMBDA-DAG: [[SFVAR_ADDR_VAL:%.+]] = load float*, float** [[SFVAR_ADDR]], 91 // LAMBDA-DAG: store double* [[G1_ADDR_VAL]], double** [[G1_REF]], 92 93 // LAMBDA-DAG: [[G_VAL:%.+]] = load{{.+}} double, double* [[G_ADDR_VAL]], 94 // LAMBDA-DAG: store double [[G_VAL]], double* [[G_PRIVATE]], 95 // LAMBDA-DAG: [[G1_VAL_REF:%.+]] = load double*, double** [[G1_REF]], 96 // LAMBDA-DAG: [[G1_VAL:%.+]] = load{{.+}} double, double* [[G1_VAL_REF]], 97 // LAMBDA-DAG: store double [[G1_VAL]], double* [[G1_PRIVATE]], 98 // LAMBDA-DAG: store double* [[G1_PRIVATE]], double** [[TMP_PRIVATE]], 99 // LAMBDA-DAG: [[SVAR_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[SVAR_ADDR_VAL]], 100 // LAMBDA-DAG: store i{{[0-9]+}} [[SVAR_VAL]], i{{[0-9]+}}* [[SVAR_PRIVATE]], 101 // LAMBDA-DAG: [[SFVAR_VAL:%.+]] = load float, float* [[SFVAR_ADDR_VAL]], 102 // LAMBDA-DAG: store float [[SFVAR_VAL]], float* [[SFVAR_PRIVATE]], 103 104 // LAMBDA: call {{.*}}void @__kmpc_for_static_init_4( 105 g += 1; 106 g1 += 1; 107 svar += 3; 108 sfvar += 4.0; 109 // LAMBDA-DAG: [[G_VAL:%.+]] = load double, double* [[G_PRIVATE]], 110 // LAMBDA-DAG: [[G_NEXT:%.+]] = fadd double [[G_VAL]], 1.{{.+}} 111 // LAMBDA-DAG: store double [[G_NEXT]], double* [[G_PRIVATE]], 112 // LAMBDA-DAG: [[TMP_VAL1:%.+]] = load double*, double** [[TMP_PRIVATE]], 113 // LAMBDA-DAG: [[TMP_VAL_VAL1:%.+]] = load{{.*}} double, double* [[TMP_VAL1]], 114 // LAMBDA-DAG: [[TMP_ADD:%.+]] = fadd double [[TMP_VAL_VAL1]], 1.{{.+}} 115 // LAMBDA-DAG: store{{.*}} double [[TMP_ADD]], double* [[TMP_VAL1]], 116 // LAMBDA-DAG: [[SVAR_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[SVAR_PRIVATE]], 117 // LAMBDA-DAG: [[SVAR_ADD:%.+]] = add{{.*}} i{{[0-9]+}} [[SVAR_VAL]], 3 118 // LAMBDA-DAG: store i{{[0-9]+}} [[SVAR_ADD]], i{{[0-9]+}}* [[SVAR_PRIVATE]], 119 // LAMBDA-DAG: [[SFVAR_VAL:%.+]] = load float, float* [[SFVAR_PRIVATE]], 120 // LAMBDA-DAG: [[SFVAR_CONV_VAL1:%.+]] = fpext float [[SFVAR_VAL]] to double 121 // LAMBDA-DAG: [[SFVAR_ADD:%.+]] = fadd double [[SFVAR_CONV_VAL1]], 4.{{.+}} 122 // LAMBDA-DAG: [[SFVAR_CONV_VAL2:%.+]] = fptrunc double [[SFVAR_ADD]] to float 123 // LAMBDA-DAG: store float [[SFVAR_CONV_VAL2:%.+]], float* [[SFVAR_PRIVATE]], 124 125 // call inner lambda (use refs to private alloca's) 126 // LAMBDA: [[GEP_0:%.+]] = getelementptr{{.+}}, i{{[0-9]+}} 0, i{{[0-9]+}} 0 127 // LAMBDA: store double* [[G_PRIVATE]], double** [[GEP_0]], 128 // LAMBDA: [[GEP_1:%.+]] = getelementptr{{.+}}, i{{[0-9]+}} 0, i{{[0-9]+}} 1 129 // LAMBDA: [[TMP_PAR:%.+]] = load double*, double** [[TMP_PRIVATE]], 130 // LAMBDA: store double* [[TMP_PAR]], double** [[GEP_1]], 131 // LAMBDA: [[GEP_2:%.+]] = getelementptr{{.+}}, i{{[0-9]+}} 0, i{{[0-9]+}} 2 132 // LAMBDA: store i{{[0-9]+}}* [[SVAR_PRIVATE]], i{{[0-9]+}}** [[GEP_2]], 133 // LAMBDA: [[GEP_3:%.+]] = getelementptr{{.+}}, i{{[0-9]+}} 0, i{{[0-9]+}} 3 134 // LAMBDA: store float* [[SFVAR_PRIVATE]], float** [[GEP_3]], 135 // LAMBDA: call{{.*}} void [[INNER_LAMBDA:@.+]](%{{.+}}* {{.+}}) 136 // LAMBDA: call {{.*}}void @__kmpc_for_static_fini( 137 [&]() { 138 // LAMBDA: define {{.+}} void [[INNER_LAMBDA]](%{{.+}}* [[ARG_PTR:%.+]]) 139 // LAMBDA: store %{{.+}}* [[ARG_PTR]], %{{.+}}** [[ARG_PTR_REF:%.+]], 140 g += 2; 141 g1 += 2; 142 svar += 4; 143 sfvar += 8.0; 144 // LAMBDA-DAG: [[ARG_PTR:%.+]] = load %{{.+}}*, %{{.+}}** [[ARG_PTR_REF]] 145 // LAMBDA-DAG: [[G_PTR_REF:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* [[ARG_PTR]], i{{[0-9]+}} 0, i{{[0-9]+}} 0 146 // LAMBDA-DAG: [[G_REF:%.+]] = load double*, double** [[G_PTR_REF]], 147 // LAMBDA-DAG: [[G_REF_VAL:%.+]] = load double, double* [[G_REF]], 148 // LAMBDA-DAG: [[G_REF_ADD:%.+]] = fadd double [[G_REF_VAL]], 2.{{.+}} 149 // LAMBDA-DAG: store double [[G_REF_ADD]], double* [[G_REF]] 150 151 // LAMBDA-DAG: [[TMP_PTR_REF:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* [[ARG_PTR]], i{{[0-9]+}} 0, i{{[0-9]+}} 1 152 // LAMBDA-DAG: [[G1_REF:%.+]] = load double*, double** [[TMP_PTR_REF]] 153 // LAMBDA-DAG: [[G1_REF_VAL:%.+]] = load double, double* [[G1_REF]], 154 // LAMBDA-DAG: [[G1_ADD:%.+]] = fadd double [[G1_REF_VAL]], 2.{{.+}} 155 // LAMBDA-DAG: store double [[G1_ADD]], double* [[G1_REF]], 156 157 // LAMBDA-DAG: [[SVAR_PTR_REF:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* [[ARG_PTR]], i{{[0-9]+}} 0, i{{[0-9]+}} 2 158 // LAMBDA-DAG: [[SVAR_REF:%.+]] = load i{{[0-9]+}}*, i{{[0-9]+}}** [[SVAR_PTR_REF]] 159 // LAMBDA-DAG: [[SVAR_REF_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[SVAR_REF]] 160 // LAMBDA-DAG: [[SVAR_ADD:%.+]] = add{{.*}} i{{[0-9]+}} [[SVAR_REF_VAL]], 4 161 // LAMBDA-DAG: store i{{[0-9]+}} [[SVAR_ADD]], i{{[0-9]+}}* [[SVAR_REF]] 162 163 // LAMBDA-DAG: [[SFVAR_PTR_REF:%.+]] = getelementptr inbounds %{{.+}}, %{{.+}}* [[ARG_PTR]], i{{[0-9]+}} 0, i{{[0-9]+}} 3 164 // LAMBDA-DAG: [[SFVAR_REF:%.+]] = load float*, float** [[SFVAR_PTR_REF]] 165 // LAMBDA-DAG: [[SFVAR_REF_VAL:%.+]] = load float, float* [[SFVAR_REF]] 166 // LAMBDA-DAG: [[SFVAR_REF_CONV:%.+]] = fpext float [[SFVAR_REF_VAL]] to double 167 // LAMBDA-DAG: [[SFVAR_ADD:%.+]] = fadd double [[SFVAR_REF_CONV]], 8.{{.+}} 168 // LAMBDA-DAG: [[SFVAR_ADD_CONV:%.+]] = fptrunc double [[SFVAR_ADD]] to float 169 // LAMBDA-DAG: store float [[SFVAR_ADD_CONV]], float* [[SFVAR_REF]], 170 }(); 171 } 172 }(); 173 return 0; 174 #else 175 S<float> test; 176 int t_var = 0; 177 int vec[] = {1, 2}; 178 S<float> s_arr[] = {1, 2}; 179 S<float> &var = test; 180 181 #pragma omp target 182 #pragma omp teams 183 #pragma omp distribute simd firstprivate(t_var, vec, s_arr, s_arr, var, var, svar) 184 for (int i = 0; i < 2; ++i) { 185 vec[i] = t_var; 186 s_arr[i] = var; 187 } 188 return tmain<int>(); 189 #endif 190 } 191 192 // CHECK: define{{.*}} i{{[0-9]+}} @main() 193 // CHECK: [[TEST:%.+]] = alloca [[S_FLOAT_TY]], 194 // CHECK: call {{.*}} [[S_FLOAT_TY_DEF_CONSTR:@.+]]([[S_FLOAT_TY]]* [[TEST]]) 195 // CHECK: call i{{[0-9]+}} @__tgt_target_teams( 196 // CHECK: call void [[OFFLOAD_FUN:@.+]]( 197 // CHECK: ret 198 199 // CHECK: define{{.+}} [[OFFLOAD_FUN]]( 200 // 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:@.+]] to void 201 // CHECK: ret 202 // 203 // CHECK: define internal void [[OMP_OUTLINED]](i{{[0-9]+}}*{{.+}}, i{{[0-9]+}}*{{.+}}, 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:%.+]]) 204 205 // CHECK: alloca i{{[0-9]+}}*, 206 // CHECK: alloca i{{[0-9]+}}*, 207 // CHECK: [[T_VAR_ADDR:%.+]] = alloca i{{[0-9]+}}*, 208 // CHECK: [[VEC_ADDR:%.+]] = alloca [2 x i{{[0-9]+}}]*, 209 // CHECK: [[S_ARR_ADDR:%.+]] = alloca [2 x [[S_FLOAT_TY]]]*, 210 // CHECK: [[VAR_ADDR:%.+]] = alloca [[S_FLOAT_TY]]*, 211 // CHECK: [[SVAR_ADDR:%.+]] = alloca i{{[0-9]+}}*, 212 // CHECK: [[TMP:%.+]] = alloca [[S_FLOAT_TY]]*, 213 214 // discard omp loop variables 215 // CHECK: {{.*}} = alloca i{{[0-9]+}}, 216 // CHECK: {{.*}} = alloca i{{[0-9]+}}, 217 // CHECK: {{.*}} = alloca i{{[0-9]+}}, 218 // CHECK: {{.*}} = alloca i{{[0-9]+}}, 219 // CHECK: {{.*}} = alloca i{{[0-9]+}}, 220 // CHECK: {{.*}} = alloca i{{[0-9]+}}, 221 222 // CHECK-DAG: [[T_VAR_PRIV:%.+]] = alloca i{{[0-9]+}}, 223 // CHECK-DAG: [[VEC_PRIV:%.+]] = alloca [2 x i{{[0-9]+}}], 224 // CHECK-DAG: [[S_ARR_PRIV:%.+]] = alloca [2 x [[S_FLOAT_TY]]], 225 // CHECK-DAG: [[VAR_PRIV:%.+]] = alloca [[S_FLOAT_TY]], 226 // CHECK-DAG: [[TMP_PRIV:%.+]] = alloca [[S_FLOAT_TY]]*, 227 // CHECK: [[SVAR_PRIV:%.+]] = alloca i{{[0-9]+}}, 228 229 // CHECK: store i{{[0-9]+}}* [[T_VAR_IN]], i{{[0-9]+}}** [[T_VAR_ADDR]], 230 // CHECK: store [2 x i{{[0-9]+}}]* [[VEC_IN]], [2 x i{{[0-9]+}}]** [[VEC_ADDR]], 231 // CHECK: store [2 x [[S_FLOAT_TY]]]* [[S_ARR_IN]], [2 x [[S_FLOAT_TY]]]** [[S_ARR_ADDR]], 232 // CHECK: store [[S_FLOAT_TY]]* [[VAR_IN]], [[S_FLOAT_TY]]** [[VAR_ADDR]], 233 // CHECK: store i{{[0-9]+}}* [[SVAR_IN]], i{{[0-9]+}}** [[SVAR_ADDR]], 234 235 // init t_var 236 // CHECK-DAG: [[T_VAR_ADDR_CONV_VAL_REF:%.+]] = load i{{[0-9]+}}*, i{{[0-9]+}}** [[T_VAR_ADDR]], 237 // CHECK-DAG: [[T_VAR_ADDR_CONV_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[T_VAR_ADDR_CONV_VAL_REF]], 238 // CHECK-DAG: store i{{[0-9]+}} [[T_VAR_ADDR_CONV_VAL]], i{{[0-9]+}}* [[T_VAR_PRIV]], 239 240 // init vec 241 // CHECK-DAG: [[VEC_ADDR_VAL:%.+]] = load [2 x i{{[0-9]+}}]*, [2 x i{{[0-9]+}}]** [[VEC_ADDR]], 242 // CHECK-DAG: [[VEC_ADDR_VAL_BCAST:%.+]] = bitcast [2 x i{{[0-9]+}}]* [[VEC_ADDR_VAL]] to i{{[0-9]+}}* 243 // CHECK-DAG: [[VEC_PRIV_BCAST:%.+]] = bitcast [2 x i{{[0-9]+}}]* [[VEC_PRIV]] to i{{[0-9]+}}* 244 // CHECK-DAG: call void @llvm.memcpy.{{.*}}(i{{[0-9]+}}* [[VEC_PRIV_BCAST]], i{{[0-9]+}}* [[VEC_ADDR_VAL_BCAST]],{{.+}}) 245 246 // init s_arr 247 // CHECK-DAG: [[S_ARR_ADDR_VAL:%.+]] = load [2 x [[S_FLOAT_TY]]]*, [2 x [[S_FLOAT_TY]]]** [[S_ARR_ADDR]], 248 // CHECK-DAG: [[S_ARR_ADDR_BCAST:%.+]] = bitcast [2 x [[S_FLOAT_TY]]]* [[S_ARR_ADDR_VAL]] to [[S_FLOAT_TY]]* 249 // CHECK-DAG: [[S_ARR_PRIV_BGN:%.+]] = getelementptr{{.+}} [2 x [[S_FLOAT_TY]]], [2 x [[S_FLOAT_TY]]]* [[S_ARR_PRIV]]{{.+}} 250 // CHECK-DAG: [[S_ARR_PRIV_NEXT:%.+]] = getelementptr [[S_FLOAT_TY]], [[S_FLOAT_TY]]* [[S_ARR_PRIV_BGN]]{{.+}} 251 // CHECK-DAG: [[S_ARR_IS_EMPTY:%.+]] = icmp eq [[S_FLOAT_TY]]* [[S_ARR_PRIV_BGN]], [[S_ARR_PRIV_NEXT]] 252 // CHECK-DAG: br i1 [[S_ARR_IS_EMPTY]], label %[[S_ARR_CPY_DONE:.+]], label %[[S_ARR_CPY_BODY:.+]] 253 254 // CHECK-DAG: [[S_ARR_CPY_BODY]]: 255 // CHECK-DAG: [[S_ARR_SRC_PAST:%.+]] = phi{{.+}} [ [[S_ARR_ADDR_BCAST]],{{.+}} ], [ [[S_ARR_SRC:%.+]],{{.+}} ] 256 // CHECK-DAG: [[S_ARR_DST_PAST:%.+]] = phi{{.+}} [ [[S_ARR_PRIV_BGN]],{{.+}} ], [ [[S_ARR_DST:%.+]],{{.+}} ] 257 // CHECK-DAG: [[S_ARR_SRC_BCAST:%.+]] = bitcast{{.+}} [[S_ARR_SRC_PAST]] to{{.+}} 258 // CHECK-DAG: [[S_ARR_DST_BCAST:%.+]] = bitcast{{.+}} [[S_ARR_DST_PAST]] to{{.+}} 259 // CHECK-DAG: call{{.+}} @llvm.memcpy.{{.+}}({{.+}}* [[S_ARR_DST_BCAST]], {{.+}}* [[S_ARR_SRC_BCAST]]{{.+}}) 260 // CHECK-DAG: [[S_ARR_SRC]] = getelementptr{{.+}} 261 // CHECK-DAG: [[S_ARR_DST]] = getelementptr{{.+}} 262 // CHECK-DAG: [[S_ARR_CPY_FIN:%.+]] = icmp{{.+}} [[S_ARR_DST]], [[S_ARR_PRIV_NEXT]] 263 // CHECK-DAG: br i1 [[S_ARR_CPY_FIN]], label %[[S_ARR_CPY_DONE]], label %[[S_ARR_CPY_BODY]] 264 // CHECK-DAG: [[S_ARR_CPY_DONE]]: 265 266 // init var 267 // CHECK-DAG: [[VAR_ADDR_VAL:%.+]] = load [[S_FLOAT_TY]]*, [[S_FLOAT_TY]]** [[VAR_ADDR]], 268 // CHECK-DAG: store{{.+}} [[VAR_ADDR_VAL]],{{.+}} [[TMP]], 269 // CHECK-DAG: [[TMP_VAL:%.+]] = load [[S_FLOAT_TY]]*, [[S_FLOAT_TY]]** [[TMP]], 270 // CHECK-DAG: [[VAR_PRIV_BCAST:%.+]] = bitcast [[S_FLOAT_TY]]* [[VAR_PRIV]] to{{.+}} 271 // CHECK-DAG: [[TMP_BCAST:%.+]] = bitcast [[S_FLOAT_TY]]* [[TMP_VAL]] to{{.+}} 272 // CHECK-DAG: call{{.+}} @llvm.memcpy.{{.+}}({{.+}}* [[VAR_PRIV_BCAST]], {{.+}}* [[TMP_BCAST]],{{.+}}) 273 // CHECK-DAG: store [[S_FLOAT_TY]]* [[VAR_PRIV]], [[S_FLOAT_TY]]** [[TMP_PRIV]], 274 275 // init svar 276 // CHECK-DAG: [[SVAR_CONV_VAL_REF:%.+]] = load{{.+}},{{.+}} [[SVAR_ADDR]], 277 // CHECK-DAG: [[SVAR_CONV_VAL:%.+]] = load{{.+}},{{.+}} [[SVAR_CONV_VAL_REF]], 278 // CHECK-DAG: store{{.+}} [[SVAR_CONV_VAL]],{{.+}} [[SVAR_PRIV]], 279 280 // CHECK-DAG: store i{{[0-9]+}} 0, i{{[0-9]+}}* %.omp{{.+}}, 281 // CHECK-DAG: store i{{[0-9]+}} 1, i{{[0-9]+}}* %.omp{{.+}}, 282 // CHECK-DAG: store i{{[0-9]+}} 1, i{{[0-9]+}}* %.omp{{.+}}, 283 // CHECK-DAG: store i{{[0-9]+}} 0, i{{[0-9]+}}* %.omp{{.+}}, 284 285 // CHECK: call void @__kmpc_for_static_init_4( 286 // CHECK: call void @__kmpc_for_static_fini( 287 // CHECK: ret void 288 289 // Template 290 // CHECK: define{{.*}} i{{[0-9]+}} [[TMAIN_INT:@.+]]() 291 // CHECK: [[TEST:%.+]] = alloca [[S_INT_TY]], 292 // CHECK: call {{.*}} [[S_INT_TY_DEF_CONSTR:@.+]]([[S_INT_TY]]* [[TEST]]) 293 // CHECK: call i{{[0-9]+}} @__tgt_target_teams( 294 // CHECK: call void [[OFFLOAD_FUN_1:@.+]]( 295 // CHECK: ret 296 297 // CHECK: define{{.+}} [[OFFLOAD_FUN_1]]( 298 // 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_1:@.+]] to void 299 // CHECK: ret 300 // 301 // CHECK: define internal void [[OMP_OUTLINED_1]](i{{[0-9]+}}*{{.+}}, i{{[0-9]+}}*{{.+}}, 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:%.+]]) 302 303 // CHECK: alloca i{{[0-9]+}}*, 304 // CHECK: alloca i{{[0-9]+}}*, 305 // CHECK: [[T_VAR_ADDR:%.+]] = alloca i{{[0-9]+}}*, 306 // CHECK: [[VEC_ADDR:%.+]] = alloca [2 x i{{[0-9]+}}]*, 307 // CHECK: [[S_ARR_ADDR:%.+]] = alloca [2 x [[S_INT_TY]]]*, 308 // CHECK: [[VAR_ADDR:%.+]] = alloca [[S_INT_TY]]*, 309 // CHECK: [[TMP:%.+]] = alloca [[S_INT_TY]]*, 310 311 // discard omp loop variables 312 // CHECK: {{.*}} = alloca i{{[0-9]+}}, 313 // CHECK: {{.*}} = alloca i{{[0-9]+}}, 314 // CHECK: {{.*}} = alloca i{{[0-9]+}}, 315 // CHECK: {{.*}} = alloca i{{[0-9]+}}, 316 // CHECK: {{.*}} = alloca i{{[0-9]+}}, 317 // CHECK: {{.*}} = alloca i{{[0-9]+}}, 318 319 // CHECK-DAG: [[T_VAR_PRIV:%.+]] = alloca i{{[0-9]+}}, 320 // CHECK-DAG: [[VEC_PRIV:%.+]] = alloca [2 x i{{[0-9]+}}], 321 // CHECK-DAG: [[S_ARR_PRIV:%.+]] = alloca [2 x [[S_INT_TY]]], 322 // CHECK-DAG: [[VAR_PRIV:%.+]] = alloca [[S_INT_TY]], 323 // CHECK-DAG: [[TMP_PRIV:%.+]] = alloca [[S_INT_TY]]*, 324 325 // CHECK: store i{{[0-9]+}}* [[T_VAR_IN]], i{{[0-9]+}}** [[T_VAR_ADDR]], 326 // CHECK: store [2 x i{{[0-9]+}}]* [[VEC_IN]], [2 x i{{[0-9]+}}]** [[VEC_ADDR]], 327 // CHECK: store [2 x [[S_INT_TY]]]* [[S_ARR_IN]], [2 x [[S_INT_TY]]]** [[S_ARR_ADDR]], 328 // CHECK: store [[S_INT_TY]]* [[VAR_IN]], [[S_INT_TY]]** [[VAR_ADDR]], 329 330 // init t_var 331 // CHECK-DAG: [[T_VAR_ADDR_CONV_VAL_REF:%.+]] = load i{{[0-9]+}}*, i{{[0-9]+}}** [[T_VAR_ADDR]], 332 // CHECK-DAG: [[T_VAR_ADDR_CONV_VAL:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[T_VAR_ADDR_CONV_VAL_REF]], 333 // CHECK-DAG: store i{{[0-9]+}} [[T_VAR_ADDR_CONV_VAL]], i{{[0-9]+}}* [[T_VAR_PRIV]], 334 335 // init vec 336 // CHECK-DAG: [[VEC_ADDR_VAL:%.+]] = load [2 x i{{[0-9]+}}]*, [2 x i{{[0-9]+}}]** [[VEC_ADDR]], 337 // CHECK-DAG: [[VEC_ADDR_VAL_BCAST:%.+]] = bitcast [2 x i{{[0-9]+}}]* [[VEC_ADDR_VAL]] to i{{[0-9]+}}* 338 // CHECK-DAG: [[VEC_PRIV_BCAST:%.+]] = bitcast [2 x i{{[0-9]+}}]* [[VEC_PRIV]] to i{{[0-9]+}}* 339 // CHECK-DAG: call void @llvm.memcpy.{{.*}}(i{{[0-9]+}}* [[VEC_PRIV_BCAST]], i{{[0-9]+}}* [[VEC_ADDR_VAL_BCAST]],{{.+}}) 340 341 // init s_arr 342 // CHECK-DAG: [[S_ARR_ADDR_VAL:%.+]] = load [2 x [[S_INT_TY]]]*, [2 x [[S_INT_TY]]]** [[S_ARR_ADDR]], 343 // CHECK-DAG: [[S_ARR_ADDR_BCAST:%.+]] = bitcast [2 x [[S_INT_TY]]]* [[S_ARR_ADDR_VAL]] to [[S_INT_TY]]* 344 // CHECK-DAG: [[S_ARR_PRIV_BGN:%.+]] = getelementptr{{.+}} [2 x [[S_INT_TY]]], [2 x [[S_INT_TY]]]* [[S_ARR_PRIV]]{{.+}} 345 // CHECK-DAG: [[S_ARR_PRIV_NEXT:%.+]] = getelementptr [[S_INT_TY]], [[S_INT_TY]]* [[S_ARR_PRIV_BGN]]{{.+}} 346 // CHECK-DAG: [[S_ARR_IS_EMPTY:%.+]] = icmp eq [[S_INT_TY]]* [[S_ARR_PRIV_BGN]], [[S_ARR_PRIV_NEXT]] 347 // CHECK-DAG: br i1 [[S_ARR_IS_EMPTY]], label %[[S_ARR_CPY_DONE:.+]], label %[[S_ARR_CPY_BODY:.+]] 348 349 // CHECK-DAG: [[S_ARR_CPY_BODY]]: 350 // CHECK-DAG: [[S_ARR_SRC_PAST:%.+]] = phi{{.+}} [ [[S_ARR_ADDR_BCAST]],{{.+}} ], [ [[S_ARR_SRC:%.+]],{{.+}} ] 351 // CHECK-DAG: [[S_ARR_DST_PAST:%.+]] = phi{{.+}} [ [[S_ARR_PRIV_BGN]],{{.+}} ], [ [[S_ARR_DST:%.+]],{{.+}} ] 352 // CHECK-DAG: [[S_ARR_SRC_BCAST:%.+]] = bitcast{{.+}} [[S_ARR_SRC_PAST]] to{{.+}} 353 // CHECK-DAG: [[S_ARR_DST_BCAST:%.+]] = bitcast{{.+}} [[S_ARR_DST_PAST]] to{{.+}} 354 // CHECK-DAG: call{{.+}} @llvm.memcpy.{{.+}}({{.+}}* [[S_ARR_DST_BCAST]], {{.+}}* [[S_ARR_SRC_BCAST]]{{.+}}) 355 // CHECK-DAG: [[S_ARR_SRC]] = getelementptr{{.+}} 356 // CHECK-DAG: [[S_ARR_DST]] = getelementptr{{.+}} 357 // CHECK-DAG: [[S_ARR_CPY_FIN:%.+]] = icmp{{.+}} [[S_ARR_DST]], [[S_ARR_PRIV_NEXT]] 358 // CHECK-DAG: br i1 [[S_ARR_CPY_FIN]], label %[[S_ARR_CPY_DONE]], label %[[S_ARR_CPY_BODY]] 359 // CHECK-DAG: [[S_ARR_CPY_DONE]]: 360 361 // init var 362 // CHECK-DAG: [[VAR_ADDR_VAL:%.+]] = load [[S_INT_TY]]*, [[S_INT_TY]]** [[VAR_ADDR]], 363 // CHECK-DAG: store{{.+}} [[VAR_ADDR_VAL]],{{.+}} [[TMP]], 364 // CHECK-DAG: [[TMP_VAL:%.+]] = load [[S_INT_TY]]*, [[S_INT_TY]]** [[TMP]], 365 // CHECK-DAG: [[VAR_PRIV_BCAST:%.+]] = bitcast [[S_INT_TY]]* [[VAR_PRIV]] to{{.+}} 366 // CHECK-DAG: [[TMP_BCAST:%.+]] = bitcast [[S_INT_TY]]* [[TMP_VAL]] to{{.+}} 367 // CHECK-DAG: call{{.+}} @llvm.memcpy.{{.+}}({{.+}}* [[VAR_PRIV_BCAST]], {{.+}}* [[TMP_BCAST]],{{.+}}) 368 // CHECK-DAG: store [[S_INT_TY]]* [[VAR_PRIV]], [[S_INT_TY]]** [[TMP_PRIV]], 369 370 // CHECK-DAG: store i{{[0-9]+}} 0, i{{[0-9]+}}* %.omp{{.+}}, 371 // CHECK-DAG: store i{{[0-9]+}} 1, i{{[0-9]+}}* %.omp{{.+}}, 372 // CHECK-DAG: store i{{[0-9]+}} 1, i{{[0-9]+}}* %.omp{{.+}}, 373 // CHECK-DAG: store i{{[0-9]+}} 0, i{{[0-9]+}}* %.omp{{.+}}, 374 375 // CHECK: call void @__kmpc_for_static_init_4( 376 // CHECK: call void @__kmpc_for_static_fini( 377 // CHECK: ret void 378 379 // CHECK: !{!"llvm.loop.vectorize.enable", i1 true} 380 #endif 381