1 // RUN: %clang_cc1 -verify -fopenmp -x c++ -triple x86_64-apple-darwin10 -emit-llvm %s -o - | FileCheck %s 2 // RUN: %clang_cc1 -fopenmp -x c++ -std=c++11 -triple x86_64-apple-darwin10 -emit-pch -o %t %s 3 // RUN: %clang_cc1 -fopenmp -x c++ -triple x86_64-apple-darwin10 -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s 4 // expected-no-diagnostics 5 #ifndef HEADER 6 #define HEADER 7 8 volatile double g, g_orig; 9 volatile double &g1 = g_orig; 10 11 struct BaseS { 12 int x; 13 }; 14 struct BaseS1 { 15 float y; 16 }; 17 18 template <class T> 19 struct S : public BaseS, public BaseS1 { 20 T f; 21 S(T a) : f(a + g) {} 22 S() : f(g) {} 23 ~S() {} 24 }; 25 void red(BaseS1&, const BaseS1&); 26 void red_plus(BaseS1&, const BaseS1&); 27 void init(BaseS1&, const BaseS1&); 28 void init1(BaseS1&, const BaseS1&); 29 void init2(BaseS1&, const BaseS1&); 30 void init_plus(BaseS1&, const BaseS1&); 31 #pragma omp declare reduction(operator& : BaseS1 : red(omp_out, omp_in)) initializer(init(omp_priv, omp_orig)) 32 #pragma omp declare reduction(+ : BaseS1 : red_plus(omp_out, omp_in)) initializer(init_plus(omp_priv, omp_orig)) 33 #pragma omp declare reduction(&& : S<float>, S<int> : omp_out.f *= omp_in.f) initializer(init1(omp_priv, omp_orig)) 34 35 // CHECK-DAG: [[S_FLOAT_TY:%.+]] = type { %{{[^,]+}}, %{{[^,]+}}, float } 36 // CHECK-DAG: [[S_INT_TY:%.+]] = type { %{{[^,]+}}, %{{[^,]+}}, i{{[0-9]+}} } 37 // CHECK-DAG: [[ATOMIC_REDUCE_BARRIER_LOC:@.+]] = private unnamed_addr constant %{{.+}} { i32 0, i32 18, i32 0, i32 0, i8* 38 // CHECK-DAG: [[IMPLICIT_BARRIER_LOC:@.+]] = private unnamed_addr constant %{{.+}} { i32 0, i32 66, i32 0, i32 0, i8* 39 // CHECK-DAG: [[REDUCTION_LOC:@.+]] = private unnamed_addr constant %{{.+}} { i32 0, i32 18, i32 0, i32 0, i8* 40 // CHECK-DAG: [[REDUCTION_LOCK:@.+]] = common global [8 x i32] zeroinitializer 41 42 #pragma omp declare reduction(operator&& : int : omp_out = 111 & omp_in) 43 template <typename T, int length> 44 T tmain() { 45 T t; 46 S<T> test; 47 T t_var = T(), t_var1; 48 T vec[] = {1, 2}; 49 S<T> s_arr[] = {1, 2}; 50 S<T> &var = test; 51 S<T> var1; 52 S<T> arr[length]; 53 #pragma omp declare reduction(operator& : T : omp_out = 15 + omp_in) 54 #pragma omp declare reduction(operator+ : T : omp_out = 1513 + omp_in) initializer(omp_priv = 321) 55 #pragma omp declare reduction(min : T : omp_out = 47 - omp_in) initializer(omp_priv = 432 / omp_orig) 56 #pragma omp declare reduction(operator&& : S<T> : omp_out.f = 17 * omp_in.f) initializer(init2(omp_priv, omp_orig)) 57 #pragma omp declare reduction(operator&& : T : omp_out = 17 * omp_in) 58 #pragma omp parallel 59 #pragma omp for reduction(+ : t_var) reduction(& : var) reduction(&& : var1) reduction(min : t_var1) nowait 60 for (int i = 0; i < 2; ++i) { 61 vec[i] = t_var; 62 s_arr[i] = var; 63 } 64 #pragma omp parallel 65 #pragma omp for reduction(&& : t_var) 66 for (int i = 0; i < 2; ++i) { 67 vec[i] = t_var; 68 s_arr[i] = var; 69 } 70 #pragma omp parallel 71 #pragma omp for reduction(+ : arr[1:length-2]) 72 for (int i = 0; i < 2; ++i) { 73 vec[i] = t_var; 74 s_arr[i] = var; 75 } 76 return T(); 77 } 78 79 extern S<float> **foo(); 80 81 #pragma omp declare reduction(operator- : float, double : omp_out = 333 + omp_in) 82 #pragma omp declare reduction(min : float, double : omp_out = 555 + omp_in) 83 int main() { 84 #pragma omp declare reduction(operator+ : float, double : omp_out = 222 - omp_in) initializer(omp_priv = -1) 85 S<float> test; 86 float t_var = 0, t_var1; 87 int vec[] = {1, 2}; 88 S<float> s_arr[] = {1, 2, 3, 4}; 89 S<float> &var = test; 90 S<float> var1, arrs[10][4]; 91 S<float> **var2 = foo(); 92 S<float> vvar2[5]; 93 S<float>(&var3)[4] = s_arr; 94 #pragma omp declare reduction(operator+ : int : omp_out = 555 * omp_in) initializer(omp_priv = 888) 95 #pragma omp parallel 96 #pragma omp for reduction(+ : t_var) reduction(& : var) reduction(&& : var1) reduction(min : t_var1) 97 for (int i = 0; i < 2; ++i) { 98 vec[i] = t_var; 99 s_arr[i] = var; 100 } 101 int arr[10][vec[1]]; 102 #pragma omp parallel for reduction(+ : arr[1][ : vec[1]]) reduction(& : arrs[1 : vec[1]][1 : 2]) 103 for (int i = 0; i < 10; ++i) 104 ++arr[1][i]; 105 #pragma omp parallel 106 #pragma omp for reduction(+ : arr) reduction(& : arrs) 107 for (int i = 0; i < 10; ++i) 108 ++arr[1][i]; 109 #pragma omp parallel 110 #pragma omp for reduction(& : var2[0 : 5][1 : 6]) 111 for (int i = 0; i < 10; ++i) 112 ; 113 #pragma omp parallel 114 #pragma omp for reduction(& : vvar2[0 : 5]) 115 for (int i = 0; i < 10; ++i) 116 ; 117 #pragma omp parallel 118 #pragma omp for reduction(& : var3[1 : 2]) 119 for (int i = 0; i < 10; ++i) 120 ; 121 #pragma omp parallel 122 #pragma omp for reduction(& : var3) 123 for (int i = 0; i < 10; ++i) 124 ; 125 return tmain<int, 42>(); 126 } 127 128 // CHECK: define {{.*}}i{{[0-9]+}} @main() 129 // CHECK: [[TEST:%.+]] = alloca [[S_FLOAT_TY]], 130 // CHECK: call {{.*}} [[S_FLOAT_TY_CONSTR:@.+]]([[S_FLOAT_TY]]* [[TEST]]) 131 // CHECK: call void (%{{.+}}*, i{{[0-9]+}}, void (i{{[0-9]+}}*, i{{[0-9]+}}*, ...)*, ...) @__kmpc_fork_call(%{{.+}}* @{{.+}}, i{{[0-9]+}} 6, void (i{{[0-9]+}}*, i{{[0-9]+}}*, ...)* bitcast (void (i{{[0-9]+}}*, i{{[0-9]+}}*, float*, [[S_FLOAT_TY]]*, [[S_FLOAT_TY]]*, float*, [2 x i32]*, [4 x [[S_FLOAT_TY]]]*)* [[MAIN_MICROTASK:@.+]] to void 132 // CHECK: call void (%{{.+}}*, i{{[0-9]+}}, void (i{{[0-9]+}}*, i{{[0-9]+}}*, ...)*, ...) @__kmpc_fork_call(%{{.+}}* @{{.+}}, i{{[0-9]+}} 5, void (i{{[0-9]+}}*, i{{[0-9]+}}*, ...)* bitcast (void (i{{[0-9]+}}*, i{{[0-9]+}}*, i64, i64, i32*, [2 x i32]*, [10 x [4 x [[S_FLOAT_TY]]]]*)* [[MAIN_MICROTASK1:@.+]] to void 133 // CHECK: call void (%{{.+}}*, i{{[0-9]+}}, void (i{{[0-9]+}}*, i{{[0-9]+}}*, ...)*, ...) @__kmpc_fork_call(%{{.+}}* @{{.+}}, i{{[0-9]+}} 4, void (i{{[0-9]+}}*, i{{[0-9]+}}*, ...)* bitcast (void (i{{[0-9]+}}*, i{{[0-9]+}}*, i64, i64, i32*, [10 x [4 x [[S_FLOAT_TY]]]]*)* [[MAIN_MICROTASK2:@.+]] to void 134 // CHECK: call void (%{{.+}}*, i{{[0-9]+}}, void (i{{[0-9]+}}*, i{{[0-9]+}}*, ...)*, ...) @__kmpc_fork_call(%{{.+}}* @{{.+}}, i{{[0-9]+}} 1, void (i{{[0-9]+}}*, i{{[0-9]+}}*, ...)* bitcast (void (i{{[0-9]+}}*, i{{[0-9]+}}*, [[S_FLOAT_TY]]***)* [[MAIN_MICROTASK3:@.+]] to void 135 // CHECK: call void (%{{.+}}*, i{{[0-9]+}}, void (i{{[0-9]+}}*, i{{[0-9]+}}*, ...)*, ...) @__kmpc_fork_call(%{{.+}}* @{{.+}}, i{{[0-9]+}} 1, void (i{{[0-9]+}}*, i{{[0-9]+}}*, ...)* bitcast (void (i{{[0-9]+}}*, i{{[0-9]+}}*, [5 x [[S_FLOAT_TY]]]*)* [[MAIN_MICROTASK4:@.+]] to void 136 // CHECK: call void (%{{.+}}*, i{{[0-9]+}}, void (i{{[0-9]+}}*, i{{[0-9]+}}*, ...)*, ...) @__kmpc_fork_call(%{{.+}}* @{{.+}}, i{{[0-9]+}} 1, void (i{{[0-9]+}}*, i{{[0-9]+}}*, ...)* bitcast (void (i{{[0-9]+}}*, i{{[0-9]+}}*, [4 x [[S_FLOAT_TY]]]*)* [[MAIN_MICROTASK5:@.+]] to void 137 // CHECK: call void (%{{.+}}*, i{{[0-9]+}}, void (i{{[0-9]+}}*, i{{[0-9]+}}*, ...)*, ...) @__kmpc_fork_call(%{{.+}}* @{{.+}}, i{{[0-9]+}} 1, void (i{{[0-9]+}}*, i{{[0-9]+}}*, ...)* bitcast (void (i{{[0-9]+}}*, i{{[0-9]+}}*, [4 x [[S_FLOAT_TY]]]*)* [[MAIN_MICROTASK6:@.+]] to void 138 // CHECK: = call {{.*}}i{{.+}} [[TMAIN_INT_42:@.+]]() 139 // CHECK: call {{.*}} [[S_FLOAT_TY_DESTR:@.+]]([[S_FLOAT_TY]]* 140 // CHECK: ret 141 // 142 // CHECK: define internal void [[MAIN_MICROTASK]](i{{[0-9]+}}* noalias [[GTID_ADDR:%.+]], i{{[0-9]+}}* noalias %{{.+}}, float* dereferenceable(4) %{{.+}}, [[S_FLOAT_TY]]* dereferenceable(12) %{{.+}}, [[S_FLOAT_TY]]* dereferenceable(12) %{{.+}}, float* dereferenceable(4) %{{.+}}, [2 x i32]* dereferenceable(8) %vec, [4 x [[S_FLOAT_TY]]]* dereferenceable(48) %{{.+}}) 143 // CHECK: [[T_VAR_PRIV:%.+]] = alloca float, 144 // CHECK: [[VAR_PRIV:%.+]] = alloca [[S_FLOAT_TY]], 145 // CHECK: [[VAR1_PRIV:%.+]] = alloca [[S_FLOAT_TY]], 146 // CHECK: [[T_VAR1_PRIV:%.+]] = alloca float, 147 148 // Reduction list for runtime. 149 // CHECK: [[RED_LIST:%.+]] = alloca [4 x i8*], 150 151 // CHECK: store i{{[0-9]+}}* [[GTID_ADDR]], i{{[0-9]+}}** [[GTID_ADDR_ADDR:%.+]], 152 153 // CHECK: [[T_VAR_REF:%.+]] = load float*, float** % 154 // CHECK: [[VAR1_REF:%.+]] = load [[S_FLOAT_TY]]*, [[S_FLOAT_TY]]** % 155 // CHECK: [[T_VAR1_REF:%.+]] = load float*, float** % 156 157 // For + reduction operation initial value of private variable is -1. 158 // CHECK: store float -1.0{{.+}}, float* 159 160 // For & reduction operation initial value of private variable is defined by call of 'init()' function. 161 // CHECK: call {{.*}}void @_Z4initR6BaseS1RKS_( 162 163 // For && reduction operation initial value of private variable is 1.0. 164 // CHECK: call {{.*}}void @_Z5init1R6BaseS1RKS_( 165 166 // For min reduction operation initial value of private variable is largest repesentable value. 167 // CHECK: [[INIT:%.+]] = load float, float* @ 168 // CHECK: store float [[INIT]], float* [[T_VAR1_PRIV]], 169 170 // CHECK: [[GTID_REF:%.+]] = load i{{[0-9]+}}*, i{{[0-9]+}}** [[GTID_ADDR_ADDR]] 171 // CHECK: [[GTID:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[GTID_REF]] 172 // CHECK: call void @__kmpc_for_static_init_4( 173 // Skip checks for internal operations. 174 // CHECK: call void @__kmpc_for_static_fini( 175 176 // void *RedList[<n>] = {<ReductionVars>[0], ..., <ReductionVars>[<n>-1]}; 177 178 // CHECK: [[T_VAR_PRIV_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST]], i64 0, i64 0 179 // CHECK: [[BITCAST:%.+]] = bitcast float* [[T_VAR_PRIV]] to i8* 180 // CHECK: store i8* [[BITCAST]], i8** [[T_VAR_PRIV_REF]], 181 // CHECK: [[VAR_PRIV_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST]], i64 0, i64 1 182 // CHECK: [[BITCAST:%.+]] = bitcast [[S_FLOAT_TY]]* [[VAR_PRIV]] to i8* 183 // CHECK: store i8* [[BITCAST]], i8** [[VAR_PRIV_REF]], 184 // CHECK: [[VAR1_PRIV_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST]], i64 0, i64 2 185 // CHECK: [[BITCAST:%.+]] = bitcast [[S_FLOAT_TY]]* [[VAR1_PRIV]] to i8* 186 // CHECK: store i8* [[BITCAST]], i8** [[VAR1_PRIV_REF]], 187 // CHECK: [[T_VAR1_PRIV_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST]], i64 0, i64 3 188 // CHECK: [[BITCAST:%.+]] = bitcast float* [[T_VAR1_PRIV]] to i8* 189 // CHECK: store i8* [[BITCAST]], i8** [[T_VAR1_PRIV_REF]], 190 191 // res = __kmpc_reduce(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>); 192 193 // CHECK: [[BITCAST:%.+]] = bitcast [4 x i8*]* [[RED_LIST]] to i8* 194 // CHECK: [[RES:%.+]] = call i32 @__kmpc_reduce(%{{.+}}* [[REDUCTION_LOC]], i32 [[GTID]], i32 4, i64 32, i8* [[BITCAST]], void (i8*, i8*)* [[REDUCTION_FUNC:@.+]], [8 x i32]* [[REDUCTION_LOCK]]) 195 196 // switch(res) 197 // CHECK: switch i32 [[RES]], label %[[RED_DONE:.+]] [ 198 // CHECK: i32 1, label %[[CASE1:.+]] 199 // CHECK: i32 2, label %[[CASE2:.+]] 200 // CHECK: ] 201 202 // case 1: 203 // t_var += t_var_reduction; 204 // CHECK: fsub float 2.220000e+02, % 205 206 // var = var.operator &(var_reduction); 207 // CHECK: call {{.*}}void @_Z3redR6BaseS1RKS_( 208 209 // var1 = var1.operator &&(var1_reduction); 210 // CHECK: fmul float 211 212 // t_var1 = min(t_var1, t_var1_reduction); 213 // CHECK: fadd float 5.550000e+02, % 214 215 // __kmpc_end_reduce(<loc>, <gtid>, &<lock>); 216 // CHECK: call void @__kmpc_end_reduce(%{{.+}}* [[REDUCTION_LOC]], i32 [[GTID]], [8 x i32]* [[REDUCTION_LOCK]]) 217 218 // break; 219 // CHECK: br label %[[RED_DONE]] 220 221 // case 2: 222 // t_var += t_var_reduction; 223 // CHECK: call void @__kmpc_critical( 224 // CHECK: fsub float 2.220000e+02, % 225 // CHECK: call void @__kmpc_end_critical( 226 227 // var = var.operator &(var_reduction); 228 // CHECK: call void @__kmpc_critical( 229 // CHECK: call {{.*}}void @_Z3redR6BaseS1RKS_( 230 // CHECK: call void @__kmpc_end_critical( 231 232 // var1 = var1.operator &&(var1_reduction); 233 // CHECK: call void @__kmpc_critical( 234 // CHECK: fmul float 235 // CHECK: call void @__kmpc_end_critical( 236 237 // t_var1 = min(t_var1, t_var1_reduction); 238 // CHECK: call void @__kmpc_critical( 239 // CHECK: fadd float 5.550000e+02, % 240 // CHECK: call void @__kmpc_end_critical( 241 242 // __kmpc_end_reduce(<loc>, <gtid>, &<lock>); 243 // CHECK: call void @__kmpc_end_reduce(%{{.+}}* [[REDUCTION_LOC]], i32 [[GTID]], [8 x i32]* [[REDUCTION_LOCK]]) 244 245 // break; 246 // CHECK: br label %[[RED_DONE]] 247 // CHECK: [[RED_DONE]] 248 // CHECK-DAG: call {{.*}} [[S_FLOAT_TY_DESTR]]([[S_FLOAT_TY]]* [[VAR_PRIV]]) 249 // CHECK-DAG: call {{.*}} [[S_FLOAT_TY_DESTR]]([[S_FLOAT_TY]]* 250 // CHECK: call void @__kmpc_barrier(%{{.+}}* [[IMPLICIT_BARRIER_LOC]], i{{[0-9]+}} [[GTID]]) 251 252 // CHECK: ret void 253 254 // void reduce_func(void *lhs[<n>], void *rhs[<n>]) { 255 // *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); 256 // ... 257 // *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], 258 // *(Type<n>-1*)rhs[<n>-1]); 259 // } 260 // CHECK: define internal void [[REDUCTION_FUNC]](i8*, i8*) 261 // t_var_lhs = (float*)lhs[0]; 262 // CHECK: [[T_VAR_RHS_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST_RHS:%.+]], i64 0, i64 0 263 // CHECK: [[T_VAR_RHS_VOID:%.+]] = load i8*, i8** [[T_VAR_RHS_REF]], 264 // CHECK: [[T_VAR_RHS:%.+]] = bitcast i8* [[T_VAR_RHS_VOID]] to float* 265 // t_var_rhs = (float*)rhs[0]; 266 // CHECK: [[T_VAR_LHS_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST_LHS:%.+]], i64 0, i64 0 267 // CHECK: [[T_VAR_LHS_VOID:%.+]] = load i8*, i8** [[T_VAR_LHS_REF]], 268 // CHECK: [[T_VAR_LHS:%.+]] = bitcast i8* [[T_VAR_LHS_VOID]] to float* 269 270 // var_lhs = (S<float>*)lhs[1]; 271 // CHECK: [[VAR_RHS_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST_RHS]], i64 0, i64 1 272 // CHECK: [[VAR_RHS_VOID:%.+]] = load i8*, i8** [[VAR_RHS_REF]], 273 // CHECK: [[VAR_RHS:%.+]] = bitcast i8* [[VAR_RHS_VOID]] to [[S_FLOAT_TY]]* 274 // var_rhs = (S<float>*)rhs[1]; 275 // CHECK: [[VAR_LHS_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST_LHS]], i64 0, i64 1 276 // CHECK: [[VAR_LHS_VOID:%.+]] = load i8*, i8** [[VAR_LHS_REF]], 277 // CHECK: [[VAR_LHS:%.+]] = bitcast i8* [[VAR_LHS_VOID]] to [[S_FLOAT_TY]]* 278 279 // var1_lhs = (S<float>*)lhs[2]; 280 // CHECK: [[VAR1_RHS_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST_RHS]], i64 0, i64 2 281 // CHECK: [[VAR1_RHS_VOID:%.+]] = load i8*, i8** [[VAR1_RHS_REF]], 282 // CHECK: [[VAR1_RHS:%.+]] = bitcast i8* [[VAR1_RHS_VOID]] to [[S_FLOAT_TY]]* 283 // var1_rhs = (S<float>*)rhs[2]; 284 // CHECK: [[VAR1_LHS_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST_LHS]], i64 0, i64 2 285 // CHECK: [[VAR1_LHS_VOID:%.+]] = load i8*, i8** [[VAR1_LHS_REF]], 286 // CHECK: [[VAR1_LHS:%.+]] = bitcast i8* [[VAR1_LHS_VOID]] to [[S_FLOAT_TY]]* 287 288 // t_var1_lhs = (float*)lhs[3]; 289 // CHECK: [[T_VAR1_RHS_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST_RHS]], i64 0, i64 3 290 // CHECK: [[T_VAR1_RHS_VOID:%.+]] = load i8*, i8** [[T_VAR1_RHS_REF]], 291 // CHECK: [[T_VAR1_RHS:%.+]] = bitcast i8* [[T_VAR1_RHS_VOID]] to float* 292 // t_var1_rhs = (float*)rhs[3]; 293 // CHECK: [[T_VAR1_LHS_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST_LHS]], i64 0, i64 3 294 // CHECK: [[T_VAR1_LHS_VOID:%.+]] = load i8*, i8** [[T_VAR1_LHS_REF]], 295 // CHECK: [[T_VAR1_LHS:%.+]] = bitcast i8* [[T_VAR1_LHS_VOID]] to float* 296 297 // t_var_lhs += t_var_rhs; 298 // CHECK: fsub float 2.220000e+02, % 299 300 // var_lhs = var_lhs.operator &(var_rhs); 301 // CHECK: call {{.*}}void @_Z3redR6BaseS1RKS_( 302 303 // var1_lhs = var1_lhs.operator &&(var1_rhs); 304 // CHECK: fmul float 305 306 // t_var1_lhs = min(t_var1_lhs, t_var1_rhs); 307 // CHECK: fadd float 5.550000e+02, % 308 // CHECK: ret void 309 310 // CHECK: define internal void [[MAIN_MICROTASK1]](i{{[0-9]+}}* noalias [[GTID_ADDR:%.+]], i{{[0-9]+}}* noalias %{{.+}}, i64 %{{.+}}, i64 %{{.+}}, i32* {{.+}} %{{.+}}, [2 x i32]* dereferenceable(8) %{{.+}}, [10 x [4 x [[S_FLOAT_TY]]]]* dereferenceable(480) %{{.+}}) 311 312 // Reduction list for runtime. 313 // CHECK: [[RED_LIST:%.+]] = alloca [4 x i8*], 314 315 // CHECK: store i{{[0-9]+}}* [[GTID_ADDR]], i{{[0-9]+}}** [[GTID_ADDR_ADDR:%.+]], 316 // CHECK: store i{{[0-9]+}}* %{{.+}}, i{{[0-9]+}}** 317 // CHECK: store i{{[0-9]+}}* %{{.+}}, i{{[0-9]+}}** [[ARR_ADDR:%.+]], 318 // CHECK: [[ARR:%.+]] = load i{{[0-9]+}}*, i{{[0-9]+}}** [[ARR_ADDR]], 319 320 // CHECK: [[UB_CAST:%.+]] = ptrtoint i32* [[UB1_UP:%.+]] to i64 321 // CHECK: [[LB_CAST:%.+]] = ptrtoint i32* [[LB1_0:%.+]] to i64 322 // CHECK: [[DIFF:%.+]] = sub i64 [[UB_CAST]], [[LB_CAST]] 323 // CHECK: [[SIZE_1:%.+]] = sdiv exact i64 [[DIFF]], ptrtoint (i32* getelementptr (i32, i32* null, i32 1) to i64) 324 // CHECK: [[ARR_SIZE:%.+]] = add nuw i64 [[SIZE_1]], 1 325 // CHECK: call i8* @llvm.stacksave() 326 // CHECK: [[ARR_PRIV:%.+]] = alloca i32, i64 [[ARR_SIZE]], 327 328 // Check initialization of private copy. 329 // CHECK: [[END:%.+]] = getelementptr i32, i32* [[ARR_PRIV]], i64 [[ARR_SIZE]] 330 // CHECK: [[ISEMPTY:%.+]] = icmp eq i32* [[ARR_PRIV]], [[END]] 331 // CHECK: br i1 [[ISEMPTY]], 332 // CHECK: phi i32* 333 // CHECK: store i32 888, i32* % 334 // CHECK: [[DONE:%.+]] = icmp eq i32* %{{.+}}, [[END]] 335 // CHECK: br i1 [[DONE]], 336 337 // CHECK: [[ARRS_PRIV:%.+]] = alloca [[S_FLOAT_TY]], i64 [[ARRS_SIZE:%.+]], 338 339 // Check initialization of private copy. 340 // CHECK: [[END:%.+]] = getelementptr [[S_FLOAT_TY]], [[S_FLOAT_TY]]* [[ARRS_PRIV]], i64 [[ARRS_SIZE]] 341 // CHECK: [[ISEMPTY:%.+]] = icmp eq [[S_FLOAT_TY]]* [[ARRS_PRIV]], [[END]] 342 // CHECK: br i1 [[ISEMPTY]], 343 // CHECK: phi [[S_FLOAT_TY]]* 344 // CHECK: call void @_Z4initR6BaseS1RKS_(% 345 // CHECK: [[DONE:%.+]] = icmp eq [[S_FLOAT_TY]]* %{{.+}}, [[END]] 346 // CHECK: br i1 [[DONE]], 347 348 // CHECK: [[GTID_REF:%.+]] = load i{{[0-9]+}}*, i{{[0-9]+}}** [[GTID_ADDR_ADDR]] 349 // CHECK: [[GTID:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[GTID_REF]] 350 // CHECK: call void @__kmpc_for_static_init_4( 351 // Skip checks for internal operations. 352 // CHECK: call void @__kmpc_for_static_fini( 353 354 // void *RedList[<n>] = {<ReductionVars>[0], ..., <ReductionVars>[<n>-1]}; 355 356 // CHECK: [[ARR_PRIV_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST]], i64 0, i64 0 357 // CHECK: [[BITCAST:%.+]] = bitcast i32* [[ARR_PRIV]] to i8* 358 // CHECK: store i8* [[BITCAST]], i8** [[ARR_PRIV_REF]], 359 // CHECK: [[ARR_SIZE_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST]], i64 0, i64 1 360 // CHECK: [[BITCAST:%.+]] = inttoptr i64 [[ARR_SIZE]] to i8* 361 // CHECK: store i8* [[BITCAST]], i8** [[ARR_SIZE_REF]], 362 // CHECK: [[ARRS_PRIV_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST]], i64 0, i64 2 363 // CHECK: [[BITCAST:%.+]] = bitcast [[S_FLOAT_TY]]* [[ARRS_PRIV]] to i8* 364 // CHECK: store i8* [[BITCAST]], i8** [[ARRS_PRIV_REF]], 365 // CHECK: [[ARRS_SIZE_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST]], i64 0, i64 3 366 // CHECK: [[BITCAST:%.+]] = inttoptr i64 [[ARRS_SIZE]] to i8* 367 // CHECK: store i8* [[BITCAST]], i8** [[ARRS_SIZE_REF]], 368 369 // res = __kmpc_reduce(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>); 370 371 // CHECK: [[GTID_REF:%.+]] = load i{{[0-9]+}}*, i{{[0-9]+}}** [[GTID_ADDR_ADDR]] 372 // CHECK: [[GTID:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[GTID_REF]] 373 // CHECK: [[BITCAST:%.+]] = bitcast [4 x i8*]* [[RED_LIST]] to i8* 374 // CHECK: [[RES:%.+]] = call i32 @__kmpc_reduce_nowait(%{{.+}}* [[REDUCTION_LOC]], i32 [[GTID]], i32 2, i64 32, i8* [[BITCAST]], void (i8*, i8*)* [[REDUCTION_FUNC:@.+]], [8 x i32]* [[REDUCTION_LOCK]]) 375 376 // switch(res) 377 // CHECK: switch i32 [[RES]], label %[[RED_DONE:.+]] [ 378 // CHECK: i32 1, label %[[CASE1:.+]] 379 // CHECK: i32 2, label %[[CASE2:.+]] 380 // CHECK: ] 381 382 // case 1: 383 // CHECK: [[CASE1]] 384 385 // arr[:] += arr_reduction[:]; 386 // CHECK: [[END:%.+]] = getelementptr i32, i32* [[LB1_0]], i64 [[ARR_SIZE]] 387 // CHECK: [[ISEMPTY:%.+]] = icmp eq i32* [[LB1_0]], [[END]] 388 // CHECK: br i1 [[ISEMPTY]], 389 // CHECK: phi i32* 390 // CHECK: [[ADD:%.+]] = mul nsw i32 555, % 391 // CHECK: store i32 [[ADD]], i32* % 392 // CHECK: [[DONE:%.+]] = icmp eq i32* %{{.+}}, [[END]] 393 // CHECK: br i1 [[DONE]], 394 395 // arrs[:] = var.operator &(arrs_reduction[:]); 396 // CHECK: [[END:%.+]] = getelementptr [[S_FLOAT_TY]], [[S_FLOAT_TY]]* [[ARRS_LB:%.+]], i64 [[ARRS_SIZE]] 397 // CHECK: [[ISEMPTY:%.+]] = icmp eq [[S_FLOAT_TY]]* [[ARRS_LB]], [[END]] 398 // CHECK: br i1 [[ISEMPTY]], 399 // CHECK: phi [[S_FLOAT_TY]]* 400 // CHECK: call void @_Z3redR6BaseS1RKS_(% 401 // CHECK: [[DONE:%.+]] = icmp eq [[S_FLOAT_TY]]* %{{.+}}, [[END]] 402 // CHECK: br i1 [[DONE]], 403 404 // __kmpc_end_reduce(<loc>, <gtid>, &<lock>); 405 // CHECK: call void @__kmpc_end_reduce_nowait(%{{.+}}* [[REDUCTION_LOC]], i32 [[GTID]], [8 x i32]* [[REDUCTION_LOCK]]) 406 407 // break; 408 // CHECK: br label %[[RED_DONE]] 409 410 // case 2: 411 // CHECK: [[CASE2]] 412 413 // arr[:] += arr_reduction[:]; 414 // CHECK: [[END:%.+]] = getelementptr i32, i32* [[LB1_0]], i64 [[ARR_SIZE]] 415 // CHECK: [[ISEMPTY:%.+]] = icmp eq i32* [[LB1_0]], [[END]] 416 // CHECK: br i1 [[ISEMPTY]], 417 // CHECK: phi i32* 418 // CHECK: call void @__kmpc_critical( 419 // CHECK: [[ADD:%.+]] = mul nsw i32 555, % 420 // CHECK: call void @__kmpc_end_critical( 421 // CHECK: [[DONE:%.+]] = icmp eq i32* %{{.+}}, [[END]] 422 // CHECK: br i1 [[DONE]], 423 424 // arrs[:] = var.operator &(arrs_reduction[:]); 425 // CHECK: [[END:%.+]] = getelementptr [[S_FLOAT_TY]], [[S_FLOAT_TY]]* [[ARRS_LB:%.+]], i64 [[ARRS_SIZE]] 426 // CHECK: [[ISEMPTY:%.+]] = icmp eq [[S_FLOAT_TY]]* [[ARRS_LB]], [[END]] 427 // CHECK: br i1 [[ISEMPTY]], 428 // CHECK: phi [[S_FLOAT_TY]]* 429 // CHECK: call void @__kmpc_critical( 430 // CHECK: call void @_Z3redR6BaseS1RKS_(% 431 // CHECK: call void @__kmpc_end_critical( 432 // CHECK: [[DONE:%.+]] = icmp eq [[S_FLOAT_TY]]* %{{.+}}, [[END]] 433 // CHECK: br i1 [[DONE]], 434 435 // break; 436 // CHECK: br label %[[RED_DONE]] 437 // CHECK: [[RED_DONE]] 438 439 // Check destruction of private copy. 440 // CHECK: [[END:%.+]] = getelementptr inbounds [[S_FLOAT_TY]], [[S_FLOAT_TY]]* [[ARRS_PRIV]], i64 [[ARRS_SIZE]] 441 // CHECK: [[ISEMPTY:%.+]] = icmp eq [[S_FLOAT_TY]]* [[ARRS_PRIV]], [[END]] 442 // CHECK: br i1 [[ISEMPTY]], 443 // CHECK: phi [[S_FLOAT_TY]]* 444 // CHECK: call void @_ZN1SIfED1Ev([[S_FLOAT_TY]]* % 445 // CHECK: [[DONE:%.+]] = icmp eq [[S_FLOAT_TY]]* %{{.+}}, [[ARRS_PRIV]] 446 // CHECK: br i1 [[DONE]], 447 // CHECK: call void @llvm.stackrestore(i8* 448 449 // CHECK: ret void 450 451 // void reduce_func(void *lhs[<n>], void *rhs[<n>]) { 452 // *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); 453 // ... 454 // *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], 455 // *(Type<n>-1*)rhs[<n>-1]); 456 // } 457 // CHECK: define internal void [[REDUCTION_FUNC]](i8*, i8*) 458 // arr_rhs = (int*)rhs[0]; 459 // CHECK: [[ARR_RHS_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST_RHS:%.+]], i64 0, i64 0 460 // CHECK: [[ARR_RHS_VOID:%.+]] = load i8*, i8** [[ARR_RHS_REF]], 461 // CHECK: [[ARR_RHS:%.+]] = bitcast i8* [[ARR_RHS_VOID]] to i32* 462 // arr_lhs = (int*)lhs[0]; 463 // CHECK: [[ARR_LHS_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST_LHS:%.+]], i64 0, i64 0 464 // CHECK: [[ARR_LHS_VOID:%.+]] = load i8*, i8** [[ARR_LHS_REF]], 465 // CHECK: [[ARR_LHS:%.+]] = bitcast i8* [[ARR_LHS_VOID]] to i32* 466 467 // arr_size = (size_t)lhs[1]; 468 // CHECK: [[ARR_SIZE_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST_LHS]], i64 0, i64 1 469 // CHECK: [[ARR_SIZE_VOID:%.+]] = load i8*, i8** [[ARR_SIZE_REF]], 470 // CHECK: [[ARR_SIZE:%.+]] = ptrtoint i8* [[ARR_SIZE_VOID]] to i64 471 472 // arrs_rhs = (S<float>*)rhs[2]; 473 // CHECK: [[ARRS_RHS_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST_RHS]], i64 0, i64 2 474 // CHECK: [[ARRS_RHS_VOID:%.+]] = load i8*, i8** [[ARRS_RHS_REF]], 475 // CHECK: [[ARRS_RHS:%.+]] = bitcast i8* [[ARRS_RHS_VOID]] to [[S_FLOAT_TY]]* 476 // arrs_lhs = (S<float>*)lhs[2]; 477 // CHECK: [[ARRS_LHS_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST_LHS]], i64 0, i64 2 478 // CHECK: [[ARRS_LHS_VOID:%.+]] = load i8*, i8** [[ARRS_LHS_REF]], 479 // CHECK: [[ARRS_LHS:%.+]] = bitcast i8* [[ARRS_LHS_VOID]] to [[S_FLOAT_TY]]* 480 481 // arrs_size = (size_t)lhs[3]; 482 // CHECK: [[ARRS_SIZE_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST_LHS]], i64 0, i64 3 483 // CHECK: [[ARRS_SIZE_VOID:%.+]] = load i8*, i8** [[ARRS_SIZE_REF]], 484 // CHECK: [[ARRS_SIZE:%.+]] = ptrtoint i8* [[ARRS_SIZE_VOID]] to i64 485 486 // arr_lhs[:] += arr_rhs[:]; 487 // CHECK: [[END:%.+]] = getelementptr i32, i32* [[ARR_LHS]], i64 [[ARR_SIZE]] 488 // CHECK: [[ISEMPTY:%.+]] = icmp eq i32* [[ARR_LHS]], [[END]] 489 // CHECK: br i1 [[ISEMPTY]], 490 // CHECK: phi i32* 491 // CHECK: [[ADD:%.+]] = mul nsw i32 555, % 492 // CHECK: [[DONE:%.+]] = icmp eq i32* %{{.+}}, [[END]] 493 // CHECK: br i1 [[DONE]], 494 495 // arrs_lhs = arrs_lhs.operator &(arrs_rhs); 496 // CHECK: [[END:%.+]] = getelementptr [[S_FLOAT_TY]], [[S_FLOAT_TY]]* [[ARRS_LB:%.+]], i64 [[ARRS_SIZE]] 497 // CHECK: [[ISEMPTY:%.+]] = icmp eq [[S_FLOAT_TY]]* [[ARRS_LB]], [[END]] 498 // CHECK: br i1 [[ISEMPTY]], 499 // CHECK: phi [[S_FLOAT_TY]]* 500 // CHECK: call void @_Z3redR6BaseS1RKS_(% 501 // CHECK: [[DONE:%.+]] = icmp eq [[S_FLOAT_TY]]* %{{.+}}, [[END]] 502 // CHECK: br i1 [[DONE]], 503 504 // CHECK: ret void 505 506 // CHECK: define internal void [[MAIN_MICROTASK2]](i{{[0-9]+}}* noalias [[GTID_ADDR:%.+]], i{{[0-9]+}}* noalias %{{.+}}, i64 %{{.+}}, i64 %{{.+}}, i32* {{.+}} %{{.+}}, [10 x [4 x [[S_FLOAT_TY]]]]* dereferenceable(480) %{{.+}}) 507 508 // CHECK: [[ARRS_PRIV:%.+]] = alloca [10 x [4 x [[S_FLOAT_TY]]]], 509 510 // Reduction list for runtime. 511 // CHECK: [[RED_LIST:%.+]] = alloca [3 x i8*], 512 513 // CHECK: store i{{[0-9]+}}* [[GTID_ADDR]], i{{[0-9]+}}** [[GTID_ADDR_ADDR:%.+]], 514 515 // CHECK: [[ARR_SIZE:%.+]] = udiv exact i64 516 // CHECK: call i8* @llvm.stacksave() 517 // CHECK: [[ARR_PRIV:%.+]] = alloca i32, i64 [[ARR_SIZE]], 518 519 // Check initialization of private copy. 520 // CHECK: [[END:%.+]] = getelementptr i32, i32* [[ARR_PRIV]], i64 [[ARR_SIZE]] 521 // CHECK: [[ISEMPTY:%.+]] = icmp eq i32* [[ARR_PRIV]], [[END]] 522 // CHECK: br i1 [[ISEMPTY]], 523 // CHECK: phi i32* 524 // CHECK: store i32 888, i32* % 525 // CHECK: [[DONE:%.+]] = icmp eq i32* %{{.+}}, [[END]] 526 // CHECK: br i1 [[DONE]], 527 528 // Check initialization of private copy. 529 // CHECK: [[BEGIN:%.+]] = getelementptr inbounds [10 x [4 x [[S_FLOAT_TY]]]], [10 x [4 x [[S_FLOAT_TY]]]]* [[ARRS_PRIV]], i32 0, i32 0, i32 0 530 // CHECK: [[LHS_BEGIN:%.+]] = bitcast [10 x [4 x [[S_FLOAT_TY]]]]* %{{.+}} to [[S_FLOAT_TY]]* 531 // CHECK: [[END:%.+]] = getelementptr [[S_FLOAT_TY]], [[S_FLOAT_TY]]* [[BEGIN]], i64 40 532 // CHECK: [[ISEMPTY:%.+]] = icmp eq [[S_FLOAT_TY]]* [[BEGIN]], [[END]] 533 // CHECK: br i1 [[ISEMPTY]], 534 // CHECK: phi [[S_FLOAT_TY]]* 535 // CHECK: call void @_Z4initR6BaseS1RKS_(% 536 // CHECK: [[DONE:%.+]] = icmp eq [[S_FLOAT_TY]]* %{{.+}}, [[END]] 537 // CHECK: br i1 [[DONE]], 538 // CHECK: [[LHS_BEGIN:%.+]] = bitcast [10 x [4 x [[S_FLOAT_TY]]]]* %{{.+}} to [[S_FLOAT_TY]]* 539 // CHECK: [[ARRS_PRIV_BEGIN:%.+]] = bitcast [10 x [4 x [[S_FLOAT_TY]]]]* [[ARRS_PRIV]] to [[S_FLOAT_TY]]* 540 541 // CHECK: [[GTID_REF:%.+]] = load i{{[0-9]+}}*, i{{[0-9]+}}** [[GTID_ADDR_ADDR]] 542 // CHECK: [[GTID:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[GTID_REF]] 543 // CHECK: call void @__kmpc_for_static_init_4( 544 // Skip checks for internal operations. 545 // CHECK: call void @__kmpc_for_static_fini( 546 547 // void *RedList[<n>] = {<ReductionVars>[0], ..., <ReductionVars>[<n>-1]}; 548 549 // CHECK: [[ARR_PRIV_REF:%.+]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[RED_LIST]], i64 0, i64 0 550 // CHECK: [[BITCAST:%.+]] = bitcast i32* [[ARR_PRIV]] to i8* 551 // CHECK: store i8* [[BITCAST]], i8** [[ARR_PRIV_REF]], 552 // CHECK: [[ARR_SIZE_REF:%.+]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[RED_LIST]], i64 0, i64 1 553 // CHECK: [[BITCAST:%.+]] = inttoptr i64 [[ARR_SIZE]] to i8* 554 // CHECK: store i8* [[BITCAST]], i8** [[ARR_SIZE_REF]], 555 // CHECK: [[ARRS_PRIV_REF:%.+]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[RED_LIST]], i64 0, i64 2 556 // CHECK: [[BITCAST:%.+]] = bitcast [[S_FLOAT_TY]]* [[ARRS_PRIV_BEGIN]] to i8* 557 // CHECK: store i8* [[BITCAST]], i8** [[ARRS_PRIV_REF]], 558 559 // res = __kmpc_reduce(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>); 560 561 // CHECK: [[GTID_REF:%.+]] = load i{{[0-9]+}}*, i{{[0-9]+}}** [[GTID_ADDR_ADDR]] 562 // CHECK: [[GTID:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[GTID_REF]] 563 // CHECK: [[BITCAST:%.+]] = bitcast [3 x i8*]* [[RED_LIST]] to i8* 564 // CHECK: [[RES:%.+]] = call i32 @__kmpc_reduce(%{{.+}}* [[REDUCTION_LOC]], i32 [[GTID]], i32 2, i64 24, i8* [[BITCAST]], void (i8*, i8*)* [[REDUCTION_FUNC:@.+]], [8 x i32]* [[REDUCTION_LOCK]]) 565 566 // switch(res) 567 // CHECK: switch i32 [[RES]], label %[[RED_DONE:.+]] [ 568 // CHECK: i32 1, label %[[CASE1:.+]] 569 // CHECK: i32 2, label %[[CASE2:.+]] 570 // CHECK: ] 571 572 // case 1: 573 // CHECK: [[CASE1]] 574 575 // arr[:] += arr_reduction[:]; 576 // CHECK: [[END:%.+]] = getelementptr i32, i32* [[LB1_0:%.+]], i64 [[ARR_SIZE]] 577 // CHECK: [[ISEMPTY:%.+]] = icmp eq i32* [[LB1_0]], [[END]] 578 // CHECK: br i1 [[ISEMPTY]], 579 // CHECK: phi i32* 580 // CHECK: [[ADD:%[^ ]+]] = mul nsw i32 555, % 581 // CHECK: store i32 [[ADD]], i32* % 582 // CHECK: [[DONE:%.+]] = icmp eq i32* %{{.+}}, [[END]] 583 // CHECK: br i1 [[DONE]], 584 585 // arrs[:] = var.operator &(arrs_reduction[:]); 586 // CHECK: [[END:%.+]] = getelementptr [[S_FLOAT_TY]], [[S_FLOAT_TY]]* [[LHS_BEGIN]], i64 40 587 // CHECK: [[ISEMPTY:%.+]] = icmp eq [[S_FLOAT_TY]]* [[LHS_BEGIN]], [[END]] 588 // CHECK: br i1 [[ISEMPTY]], 589 // CHECK: phi [[S_FLOAT_TY]]* 590 // CHECK: call void @_Z3redR6BaseS1RKS_(% 591 // CHECK: [[DONE:%.+]] = icmp eq [[S_FLOAT_TY]]* %{{.+}}, [[END]] 592 // CHECK: br i1 [[DONE]], 593 594 // __kmpc_end_reduce(<loc>, <gtid>, &<lock>); 595 // CHECK: call void @__kmpc_end_reduce(%{{.+}}* [[REDUCTION_LOC]], i32 [[GTID]], [8 x i32]* [[REDUCTION_LOCK]]) 596 597 // break; 598 // CHECK: br label %[[RED_DONE]] 599 600 // case 2: 601 // CHECK: [[CASE2]] 602 603 // arr[:] += arr_reduction[:]; 604 // CHECK: [[END:%.+]] = getelementptr i32, i32* [[LB1_0]], i64 [[ARR_SIZE]] 605 // CHECK: [[ISEMPTY:%.+]] = icmp eq i32* [[LB1_0]], [[END]] 606 // CHECK: br i1 [[ISEMPTY]], 607 // CHECK: phi i32* 608 // CHECK: call void @__kmpc_critical( 609 // CHECK: [[ADD:%.+]] = mul nsw i32 555, % 610 // CHECK: call void @__kmpc_end_critical( 611 // CHECK: [[DONE:%.+]] = icmp eq i32* %{{.+}}, [[END]] 612 // CHECK: br i1 [[DONE]], 613 614 // arrs[:] = var.operator &(arrs_reduction[:]); 615 // CHECK: [[END:%.+]] = getelementptr [[S_FLOAT_TY]], [[S_FLOAT_TY]]* [[LHS_BEGIN]], i64 40 616 // CHECK: [[ISEMPTY:%.+]] = icmp eq [[S_FLOAT_TY]]* [[LHS_BEGIN]], [[END]] 617 // CHECK: br i1 [[ISEMPTY]], 618 // CHECK: phi [[S_FLOAT_TY]]* 619 // CHECK: call void @__kmpc_critical( 620 // CHECK: call void @_Z3redR6BaseS1RKS_(% 621 // CHECK: call void @__kmpc_end_critical( 622 // CHECK: [[DONE:%.+]] = icmp eq [[S_FLOAT_TY]]* %{{.+}}, [[END]] 623 // CHECK: br i1 [[DONE]], 624 625 // break; 626 // CHECK: br label %[[RED_DONE]] 627 // CHECK: [[RED_DONE]] 628 629 // Check destruction of private copy. 630 // CHECK: [[BEGIN:%.+]] = getelementptr inbounds [10 x [4 x [[S_FLOAT_TY]]]], [10 x [4 x [[S_FLOAT_TY]]]]* [[ARRS_PRIV]], i32 0, i32 0, i32 0 631 // CHECK: [[END:%.+]] = getelementptr inbounds [[S_FLOAT_TY]], [[S_FLOAT_TY]]* [[BEGIN]], i64 40 632 // CHECK: br 633 // CHECK: phi [[S_FLOAT_TY]]* 634 // CHECK: call void @_ZN1SIfED1Ev([[S_FLOAT_TY]]* % 635 // CHECK: [[DONE:%.+]] = icmp eq [[S_FLOAT_TY]]* %{{.+}}, [[BEGIN]] 636 // CHECK: br i1 [[DONE]], 637 // CHECK: call void @llvm.stackrestore(i8* 638 // CHECK: call void @__kmpc_barrier( 639 640 // CHECK: ret void 641 642 // void reduce_func(void *lhs[<n>], void *rhs[<n>]) { 643 // *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); 644 // ... 645 // *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], 646 // *(Type<n>-1*)rhs[<n>-1]); 647 // } 648 // CHECK: define internal void [[REDUCTION_FUNC]](i8*, i8*) 649 // arr_rhs = (int*)rhs[0]; 650 // CHECK: [[ARR_RHS_REF:%.+]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[RED_LIST_RHS:%.+]], i64 0, i64 0 651 // CHECK: [[ARR_RHS_VOID:%.+]] = load i8*, i8** [[ARR_RHS_REF]], 652 // CHECK: [[ARR_RHS:%.+]] = bitcast i8* [[ARR_RHS_VOID]] to i32* 653 // arr_lhs = (int*)lhs[0]; 654 // CHECK: [[ARR_LHS_REF:%.+]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[RED_LIST_LHS:%.+]], i64 0, i64 0 655 // CHECK: [[ARR_LHS_VOID:%.+]] = load i8*, i8** [[ARR_LHS_REF]], 656 // CHECK: [[ARR_LHS:%.+]] = bitcast i8* [[ARR_LHS_VOID]] to i32* 657 658 // arr_size = (size_t)lhs[1]; 659 // CHECK: [[ARR_SIZE_REF:%.+]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[RED_LIST_LHS]], i64 0, i64 1 660 // CHECK: [[ARR_SIZE_VOID:%.+]] = load i8*, i8** [[ARR_SIZE_REF]], 661 // CHECK: [[ARR_SIZE:%.+]] = ptrtoint i8* [[ARR_SIZE_VOID]] to i64 662 663 // arrs_rhs = (S<float>*)rhs[2]; 664 // CHECK: [[ARRS_RHS_REF:%.+]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[RED_LIST_RHS]], i64 0, i64 2 665 // CHECK: [[ARRS_RHS_VOID:%.+]] = load i8*, i8** [[ARRS_RHS_REF]], 666 // CHECK: [[ARRS_RHS:%.+]] = bitcast i8* [[ARRS_RHS_VOID]] to [[S_FLOAT_TY]]* 667 // arrs_lhs = (S<float>*)lhs[2]; 668 // CHECK: [[ARRS_LHS_REF:%.+]] = getelementptr inbounds [3 x i8*], [3 x i8*]* [[RED_LIST_LHS]], i64 0, i64 2 669 // CHECK: [[ARRS_LHS_VOID:%.+]] = load i8*, i8** [[ARRS_LHS_REF]], 670 // CHECK: [[ARRS_LHS:%.+]] = bitcast i8* [[ARRS_LHS_VOID]] to [[S_FLOAT_TY]]* 671 672 // arr_lhs[:] += arr_rhs[:]; 673 // CHECK: [[END:%.+]] = getelementptr i32, i32* [[ARR_LHS]], i64 [[ARR_SIZE]] 674 // CHECK: [[ISEMPTY:%.+]] = icmp eq i32* [[ARR_LHS]], [[END]] 675 // CHECK: br i1 [[ISEMPTY]], 676 // CHECK: phi i32* 677 // CHECK: [[ADD:%.+]] = mul nsw i32 555, % 678 // CHECK: store i32 [[ADD]], i32* % 679 // CHECK: [[DONE:%.+]] = icmp eq i32* %{{.+}}, [[END]] 680 // CHECK: br i1 [[DONE]], 681 682 // arrs_lhs = arrs_lhs.operator &(arrs_rhs); 683 // CHECK: [[END:%.+]] = getelementptr [[S_FLOAT_TY]], [[S_FLOAT_TY]]* [[ARRS_LB:%.+]], i64 40 684 // CHECK: [[ISEMPTY:%.+]] = icmp eq [[S_FLOAT_TY]]* [[ARRS_LB]], [[END]] 685 // CHECK: br i1 [[ISEMPTY]], 686 // CHECK: phi [[S_FLOAT_TY]]* 687 // CHECK: call void @_Z3redR6BaseS1RKS_(% 688 // CHECK: [[DONE:%.+]] = icmp eq [[S_FLOAT_TY]]* %{{.+}}, [[END]] 689 // CHECK: br i1 [[DONE]], 690 691 // CHECK: ret void 692 693 // CHECK: define internal void [[MAIN_MICROTASK3]](i{{[0-9]+}}* noalias [[GTID_ADDR:%.+]], i{{[0-9]+}}* noalias %{{.+}}, [[S_FLOAT_TY]]*** dereferenceable(8) %{{.+}}) 694 695 // CHECK: [[VAR2_ORIG_ADDR:%.+]] = alloca [[S_FLOAT_TY]]***, 696 697 // Reduction list for runtime. 698 // CHECK: [[RED_LIST:%.+]] = alloca [2 x i8*], 699 700 // CHECK: store i{{[0-9]+}}* [[GTID_ADDR]], i{{[0-9]+}}** [[GTID_ADDR_ADDR:%.+]], 701 // CHECK: [[VAR2_ORIG:%.+]] = load [[S_FLOAT_TY]]***, [[S_FLOAT_TY]]**** [[VAR2_ORIG_ADDR]], 702 703 // CHECK: [[LAST:%.+]] = ptrtoint [[S_FLOAT_TY]]* %{{.+}} to i64 704 // CHECK: [[FIRST:%.+]] = ptrtoint [[S_FLOAT_TY]]* [[LOW:%.+]] to i64 705 // CHECK: [[BYTE_DIF:%.+]] = sub i64 [[LAST]], [[FIRST]] 706 // CHECK: [[DIF:%.+]] = sdiv exact i64 [[BYTE_DIF]], ptrtoint ([[S_FLOAT_TY]]* getelementptr ([[S_FLOAT_TY]], [[S_FLOAT_TY]]* null, i32 1) to i64) 707 // CHECK: [[SIZE:%.+]] = add nuw i64 [[DIF]], 1 708 // CHECK: call i8* @llvm.stacksave() 709 // CHECK: [[VAR2_PRIV:%.+]] = alloca [[S_FLOAT_TY]], i64 [[SIZE]], 710 // CHECK: [[LD:%.+]] = load [[S_FLOAT_TY]]**, [[S_FLOAT_TY]]*** [[VAR2_ORIG]], 711 // CHECK: [[ORIG_START:%.+]] = load [[S_FLOAT_TY]]*, [[S_FLOAT_TY]]** [[LD]], 712 // CHECK: [[START:%.+]] = ptrtoint [[S_FLOAT_TY]]* [[ORIG_START]] to i64 713 // CHECK: [[LOW_BOUND:%.+]] = ptrtoint [[S_FLOAT_TY]]* [[LOW]] to i64 714 // CHECK: [[OFFSET_BYTES:%.+]] = sub i64 [[START]], [[LOW_BOUND]] 715 // CHECK: [[OFFSET:%.+]] = sdiv exact i64 [[OFFSET_BYTES]], ptrtoint ([[S_FLOAT_TY]]* getelementptr ([[S_FLOAT_TY]], [[S_FLOAT_TY]]* null, i32 1) to i64) 716 // CHECK: [[PSEUDO_VAR2_PRIV:%.+]] = getelementptr [[S_FLOAT_TY]], [[S_FLOAT_TY]]* [[VAR2_PRIV]], i64 [[OFFSET]] 717 // CHECK: store [[S_FLOAT_TY]]** [[REF:.+]], [[S_FLOAT_TY]]*** % 718 // CHECK: store [[S_FLOAT_TY]]* [[PSEUDO_VAR2_PRIV]], [[S_FLOAT_TY]]** [[REF]] 719 // CHECK: ret void 720 721 // CHECK: define internal void [[MAIN_MICROTASK4]](i{{[0-9]+}}* noalias [[GTID_ADDR:%.+]], i{{[0-9]+}}* noalias %{{.+}}, [5 x [[S_FLOAT_TY]]]* dereferenceable(60) %{{.+}}) 722 723 // CHECK: [[VVAR2_ORIG_ADDR:%.+]] = alloca [5 x [[S_FLOAT_TY]]]*, 724 // CHECK: [[VVAR2_PRIV:%.+]] = alloca [5 x [[S_FLOAT_TY]]], 725 726 // Reduction list for runtime. 727 // CHECK: [[RED_LIST:%.+]] = alloca [1 x i8*], 728 729 // CHECK: store i{{[0-9]+}}* [[GTID_ADDR]], i{{[0-9]+}}** [[GTID_ADDR_ADDR:%.+]], 730 // CHECK: [[VVAR2_ORIG:%.+]] = load [5 x [[S_FLOAT_TY]]]*, [5 x [[S_FLOAT_TY]]]** [[VVAR2_ORIG_ADDR]], 731 732 // CHECK: [[LOW:%.+]] = getelementptr inbounds [5 x [[S_FLOAT_TY]]], [5 x [[S_FLOAT_TY]]]* [[VVAR2_ORIG]], i64 0, i64 0 733 // CHECK: [[ORIG_START:%.+]] = bitcast [5 x [[S_FLOAT_TY]]]* [[VVAR2_ORIG]] to [[S_FLOAT_TY]]* 734 // CHECK: [[START:%.+]] = ptrtoint [[S_FLOAT_TY]]* [[ORIG_START]] to i64 735 // CHECK: [[LOW_BOUND:%.+]] = ptrtoint [[S_FLOAT_TY]]* [[LOW]] to i64 736 // CHECK: [[OFFSET_BYTES:%.+]] = sub i64 [[START]], [[LOW_BOUND]] 737 // CHECK: [[OFFSET:%.+]] = sdiv exact i64 [[OFFSET_BYTES]], ptrtoint ([[S_FLOAT_TY]]* getelementptr ([[S_FLOAT_TY]], [[S_FLOAT_TY]]* null, i32 1) to i64) 738 // CHECK: [[PSEUDO_VVAR2_PRIV:%.+]] = getelementptr [5 x [[S_FLOAT_TY]]], [5 x [[S_FLOAT_TY]]]* [[VVAR2_PRIV]], i64 [[OFFSET]] 739 // CHECK: ret void 740 741 // CHECK: define internal void [[MAIN_MICROTASK5]](i{{[0-9]+}}* noalias [[GTID_ADDR:%.+]], i{{[0-9]+}}* noalias %{{.+}}, [4 x [[S_FLOAT_TY]]]* dereferenceable(48) %{{.+}}) 742 743 // CHECK: [[VAR3_ORIG_ADDR:%.+]] = alloca [4 x [[S_FLOAT_TY]]]*, 744 // CHECK: [[VAR3_PRIV:%.+]] = alloca [2 x [[S_FLOAT_TY]]], 745 746 // Reduction list for runtime. 747 // CHECK: [[RED_LIST:%.+]] = alloca [1 x i8*], 748 749 // CHECK: store i{{[0-9]+}}* [[GTID_ADDR]], i{{[0-9]+}}** [[GTID_ADDR_ADDR:%.+]], 750 751 // CHECK: [[VAR3_ORIG:%.+]] = load [4 x [[S_FLOAT_TY]]]*, [4 x [[S_FLOAT_TY]]]** [[VAR3_ORIG_ADDR]], 752 // CHECK: store [4 x [[S_FLOAT_TY]]]* [[VAR3_ORIG]], [4 x [[S_FLOAT_TY]]]** [[VAR3_ORIG_ADDR:%.+]], 753 // CHECK: [[VAR3_ORIG:%.+]] = load [4 x [[S_FLOAT_TY]]]*, [4 x [[S_FLOAT_TY]]]** [[VAR3_ORIG_ADDR]], 754 755 // CHECK: [[LOW:%.+]] = getelementptr inbounds [4 x [[S_FLOAT_TY]]], [4 x [[S_FLOAT_TY]]]* [[VAR3_ORIG]], i64 0, i64 1 756 // CHECK: [[VAR3_ORIG:%.+]] = load [4 x [[S_FLOAT_TY]]]*, [4 x [[S_FLOAT_TY]]]** [[VAR3_ORIG_ADDR]], 757 758 // CHECK: [[VAR3_ORIG:%.+]] = load [4 x [[S_FLOAT_TY]]]*, [4 x [[S_FLOAT_TY]]]** [[VAR3_ORIG_ADDR]], 759 // CHECK: [[ORIG_START:%.+]] = bitcast [4 x [[S_FLOAT_TY]]]* [[VAR3_ORIG]] to [[S_FLOAT_TY]]* 760 // CHECK: [[START:%.+]] = ptrtoint [[S_FLOAT_TY]]* [[ORIG_START]] to i64 761 // CHECK: [[LOW_BOUND:%.+]] = ptrtoint [[S_FLOAT_TY]]* [[LOW]] to i64 762 // CHECK: [[OFFSET_BYTES:%.+]] = sub i64 [[START]], [[LOW_BOUND]] 763 // CHECK: [[OFFSET:%.+]] = sdiv exact i64 [[OFFSET_BYTES]], ptrtoint ([[S_FLOAT_TY]]* getelementptr ([[S_FLOAT_TY]], [[S_FLOAT_TY]]* null, i32 1) to i64) 764 // CHECK: [[PSEUDO_VAR3_PRIV:%.+]] = getelementptr [2 x [[S_FLOAT_TY]]], [2 x [[S_FLOAT_TY]]]* [[VAR3_PRIV]], i64 [[OFFSET]] 765 // CHECK: [[VAR3_PRIV:%.+]] = bitcast [2 x [[S_FLOAT_TY]]]* [[PSEUDO_VAR3_PRIV]] to [4 x [[S_FLOAT_TY]]]* 766 767 // CHECK: store [4 x [[S_FLOAT_TY]]]* [[VAR3_PRIV]], [4 x [[S_FLOAT_TY]]]** % 768 769 // CHECK: ret void 770 771 // CHECK: define internal void [[MAIN_MICROTASK6]](i{{[0-9]+}}* noalias [[GTID_ADDR:%.+]], i{{[0-9]+}}* noalias %{{.+}}, [4 x [[S_FLOAT_TY]]]* dereferenceable(48) %{{.+}}) 772 773 // CHECK: [[VAR3_ORIG_ADDR:%.+]] = alloca [4 x [[S_FLOAT_TY]]]*, 774 // CHECK: [[VAR3_PRIV:%.+]] = alloca [4 x [[S_FLOAT_TY]]], 775 776 // Reduction list for runtime. 777 // CHECK: [[RED_LIST:%.+]] = alloca [1 x i8*], 778 779 // CHECK: store i{{[0-9]+}}* [[GTID_ADDR]], i{{[0-9]+}}** [[GTID_ADDR_ADDR:%.+]], 780 781 // CHECK: [[VAR3_ORIG:%.+]] = load [4 x [[S_FLOAT_TY]]]*, [4 x [[S_FLOAT_TY]]]** [[VAR3_ORIG_ADDR]], 782 // CHECK: store [4 x [[S_FLOAT_TY]]]* [[VAR3_ORIG]], [4 x [[S_FLOAT_TY]]]** [[VAR3_ORIG_ADDR:%.+]], 783 // CHECK: [[VAR3_ORIG:%.+]] = load [4 x [[S_FLOAT_TY]]]*, [4 x [[S_FLOAT_TY]]]** [[VAR3_ORIG_ADDR]], 784 // CHECK: getelementptr inbounds [4 x [[S_FLOAT_TY]]], [4 x [[S_FLOAT_TY]]]* [[VAR3_PRIV]], i32 0, i32 0 785 // CHECK: bitcast [4 x [[S_FLOAT_TY]]]* [[VAR3_ORIG]] to [[S_FLOAT_TY]]* 786 // CHECK: getelementptr [[S_FLOAT_TY]], [[S_FLOAT_TY]]* %{{.+}}, i64 4 787 788 // CHECK: store [4 x [[S_FLOAT_TY]]]* [[VAR3_PRIV]], [4 x [[S_FLOAT_TY]]]** % 789 790 // CHECK: ret void 791 792 // CHECK: define {{.*}} i{{[0-9]+}} [[TMAIN_INT_42]]() 793 // CHECK: [[TEST:%.+]] = alloca [[S_INT_TY]], 794 // CHECK: call {{.*}} [[S_INT_TY_CONSTR:@.+]]([[S_INT_TY]]* [[TEST]]) 795 // CHECK: call void (%{{.+}}*, i{{[0-9]+}}, void (i{{[0-9]+}}*, i{{[0-9]+}}*, ...)*, ...) @__kmpc_fork_call(%{{.+}}* @{{.+}}, i{{[0-9]+}} 6, void (i{{[0-9]+}}*, i{{[0-9]+}}*, ...)* bitcast (void (i{{[0-9]+}}*, i{{[0-9]+}}*, i32*, [[S_INT_TY]]*, [[S_INT_TY]]*, i32*, [2 x i32]*, [2 x [[S_INT_TY]]]*)* [[TMAIN_MICROTASK:@.+]] to void 796 // Not interested in this one: 797 // CHECK: call void (%{{.+}}*, i{{[0-9]+}}, void (i{{[0-9]+}}*, i{{[0-9]+}}*, ...)*, ...) @__kmpc_fork_call(%{{.+}}* @{{.+}}, i{{[0-9]+}} 4, 798 // CHECK: call void (%{{.+}}*, i{{[0-9]+}}, void (i{{[0-9]+}}*, i{{[0-9]+}}*, ...)*, ...) @__kmpc_fork_call(%{{.+}}* @{{.+}}, i{{[0-9]+}} 5, void (i{{[0-9]+}}*, i{{[0-9]+}}*, ...)* bitcast (void (i{{[0-9]+}}*, i{{[0-9]+}}*, [42 x [[S_INT_TY]]]*, [2 x i32]*, i32*, [2 x [[S_INT_TY]]]*, [[S_INT_TY]]*)* [[TMAIN_MICROTASK2:@.+]] to void 799 // CHECK: call {{.*}} [[S_INT_TY_DESTR:@.+]]([[S_INT_TY]]* 800 // CHECK: call {{.*}} [[S_INT_TY_DESTR:@.+]]([[S_INT_TY]]* 801 // CHECK: ret 802 // 803 // CHECK: define internal void [[TMAIN_MICROTASK]](i{{[0-9]+}}* noalias [[GTID_ADDR:%.+]], i{{[0-9]+}}* noalias %{{.+}}, i32* dereferenceable(4) %{{.+}}, [[S_INT_TY]]* dereferenceable(12) %{{.+}}, [[S_INT_TY]]* dereferenceable(12) %{{.+}}, i32* dereferenceable(4) %{{.+}}, [2 x i32]* dereferenceable(8) %{{.+}}, [2 x [[S_INT_TY]]]* dereferenceable(24) %{{.+}}) 804 // CHECK: alloca i{{[0-9]+}}, 805 // CHECK: alloca i{{[0-9]+}}, 806 // CHECK: alloca i{{[0-9]+}}, 807 // CHECK: alloca i{{[0-9]+}}, 808 // CHECK: alloca i{{[0-9]+}}, 809 // CHECK: [[T_VAR_PRIV:%.+]] = alloca i{{[0-9]+}}, 810 // CHECK: [[VAR_PRIV:%.+]] = alloca [[S_INT_TY]], 811 // CHECK: [[VAR1_PRIV:%.+]] = alloca [[S_INT_TY]], 812 // CHECK: [[T_VAR1_PRIV:%.+]] = alloca i{{[0-9]+}}, 813 814 // Reduction list for runtime. 815 // CHECK: [[RED_LIST:%.+]] = alloca [4 x i8*], 816 817 // CHECK: store i{{[0-9]+}}* [[GTID_ADDR]], i{{[0-9]+}}** [[GTID_ADDR_ADDR:%.+]], 818 819 // CHECK: [[T_VAR_REF:%.+]] = load i{{[0-9]+}}*, i{{[0-9]+}}** % 820 // CHECK: [[VAR1_REF:%.+]] = load [[S_INT_TY]]*, [[S_INT_TY]]** % 821 // CHECK: [[T_VAR1_REF:%.+]] = load i{{[0-9]+}}*, i{{[0-9]+}}** % 822 823 // For + reduction operation initial value of private variable is 0. 824 // CHECK: store i32 321, i32* % 825 826 // For & reduction operation initial value of private variable is ones in all bits. 827 // CHECK: call void @_Z4initR6BaseS1RKS_( 828 829 // For && reduction operation initial value of private variable is 1.0. 830 // CHECK: call void @_Z5init2R6BaseS1RKS_( 831 832 // For min reduction operation initial value of private variable is largest repesentable value. 833 // CHECK: sdiv i32 432, % 834 835 // CHECK: [[GTID_REF:%.+]] = load i{{[0-9]+}}*, i{{[0-9]+}}** [[GTID_ADDR_ADDR]] 836 // CHECK: [[GTID:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[GTID_REF]] 837 // CHECK: call void @__kmpc_for_static_init_4( 838 // Skip checks for internal operations. 839 // CHECK: call void @__kmpc_for_static_fini( 840 841 // void *RedList[<n>] = {<ReductionVars>[0], ..., <ReductionVars>[<n>-1]}; 842 843 // CHECK: [[T_VAR_PRIV_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST]], i64 0, i64 0 844 // CHECK: [[BITCAST:%.+]] = bitcast i{{[0-9]+}}* [[T_VAR_PRIV]] to i8* 845 // CHECK: store i8* [[BITCAST]], i8** [[T_VAR_PRIV_REF]], 846 // CHECK: [[VAR_PRIV_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST]], i64 0, i64 1 847 // CHECK: [[BITCAST:%.+]] = bitcast [[S_INT_TY]]* [[VAR_PRIV]] to i8* 848 // CHECK: store i8* [[BITCAST]], i8** [[VAR_PRIV_REF]], 849 // CHECK: [[VAR1_PRIV_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST]], i64 0, i64 2 850 // CHECK: [[BITCAST:%.+]] = bitcast [[S_INT_TY]]* [[VAR1_PRIV]] to i8* 851 // CHECK: store i8* [[BITCAST]], i8** [[VAR1_PRIV_REF]], 852 // CHECK: [[T_VAR1_PRIV_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST]], i64 0, i64 3 853 // CHECK: [[BITCAST:%.+]] = bitcast i{{[0-9]+}}* [[T_VAR1_PRIV]] to i8* 854 // CHECK: store i8* [[BITCAST]], i8** [[T_VAR1_PRIV_REF]], 855 856 // res = __kmpc_reduce_nowait(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>); 857 858 // CHECK: [[BITCAST:%.+]] = bitcast [4 x i8*]* [[RED_LIST]] to i8* 859 // CHECK: [[RES:%.+]] = call i32 @__kmpc_reduce_nowait(%{{.+}}* [[REDUCTION_LOC]], i32 [[GTID]], i32 4, i64 32, i8* [[BITCAST]], void (i8*, i8*)* [[REDUCTION_FUNC:@.+]], [8 x i32]* [[REDUCTION_LOCK]]) 860 861 // switch(res) 862 // CHECK: switch i32 [[RES]], label %[[RED_DONE:.+]] [ 863 // CHECK: i32 1, label %[[CASE1:.+]] 864 // CHECK: i32 2, label %[[CASE2:.+]] 865 // CHECK: ] 866 867 // case 1: 868 // t_var += t_var_reduction; 869 // CHECK: add nsw i32 1513, % 870 871 // var = var.operator &(var_reduction); 872 // CHECK: call void @_Z3redR6BaseS1RKS_(% 873 874 // var1 = var1.operator &&(var1_reduction); 875 // CHECK: mul nsw i32 17, % 876 877 // t_var1 = min(t_var1, t_var1_reduction); 878 // CHECK: sub nsw i32 47, % 879 880 // __kmpc_end_reduce_nowait(<loc>, <gtid>, &<lock>); 881 // CHECK: call void @__kmpc_end_reduce_nowait(%{{.+}}* [[REDUCTION_LOC]], i32 [[GTID]], [8 x i32]* [[REDUCTION_LOCK]]) 882 883 // break; 884 // CHECK: br label %[[RED_DONE]] 885 886 // case 2: 887 // t_var += t_var_reduction; 888 // CHECK: call void @__kmpc_critical( 889 // CHECK: add nsw i32 1513, % 890 // CHECK: call void @__kmpc_end_critical( 891 892 // var = var.operator &(var_reduction); 893 // CHECK: call void @__kmpc_critical( 894 // CHECK: call void @_Z3redR6BaseS1RKS_(% 895 // CHECK: call void @__kmpc_end_critical( 896 897 // var1 = var1.operator &&(var1_reduction); 898 // CHECK: call void @__kmpc_critical( 899 // CHECK: mul nsw i32 17, % 900 // CHECK: call void @__kmpc_end_critical( 901 902 // t_var1 = min(t_var1, t_var1_reduction); 903 // CHECK: call void @__kmpc_critical( 904 // CHECK: sub nsw i32 47, % 905 // CHECK: call void @__kmpc_end_critical( 906 907 // break; 908 // CHECK: br label %[[RED_DONE]] 909 // CHECK: [[RED_DONE]] 910 // CHECK-DAG: call {{.*}} [[S_INT_TY_DESTR]]([[S_INT_TY]]* [[VAR_PRIV]]) 911 // CHECK-DAG: call {{.*}} [[S_INT_TY_DESTR]]([[S_INT_TY]]* 912 // CHECK: ret void 913 914 // void reduce_func(void *lhs[<n>], void *rhs[<n>]) { 915 // *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); 916 // ... 917 // *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], 918 // *(Type<n>-1*)rhs[<n>-1]); 919 // } 920 // CHECK: define internal void [[REDUCTION_FUNC]](i8*, i8*) 921 // t_var_lhs = (i{{[0-9]+}}*)lhs[0]; 922 // CHECK: [[T_VAR_RHS_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST_RHS:%.+]], i64 0, i64 0 923 // CHECK: [[T_VAR_RHS_VOID:%.+]] = load i8*, i8** [[T_VAR_RHS_REF]], 924 // CHECK: [[T_VAR_RHS:%.+]] = bitcast i8* [[T_VAR_RHS_VOID]] to i{{[0-9]+}}* 925 // t_var_rhs = (i{{[0-9]+}}*)rhs[0]; 926 // CHECK: [[T_VAR_LHS_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST_LHS:%.+]], i64 0, i64 0 927 // CHECK: [[T_VAR_LHS_VOID:%.+]] = load i8*, i8** [[T_VAR_LHS_REF]], 928 // CHECK: [[T_VAR_LHS:%.+]] = bitcast i8* [[T_VAR_LHS_VOID]] to i{{[0-9]+}}* 929 930 // var_lhs = (S<i{{[0-9]+}}>*)lhs[1]; 931 // CHECK: [[VAR_RHS_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST_RHS]], i64 0, i64 1 932 // CHECK: [[VAR_RHS_VOID:%.+]] = load i8*, i8** [[VAR_RHS_REF]], 933 // CHECK: [[VAR_RHS:%.+]] = bitcast i8* [[VAR_RHS_VOID]] to [[S_INT_TY]]* 934 // var_rhs = (S<i{{[0-9]+}}>*)rhs[1]; 935 // CHECK: [[VAR_LHS_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST_LHS]], i64 0, i64 1 936 // CHECK: [[VAR_LHS_VOID:%.+]] = load i8*, i8** [[VAR_LHS_REF]], 937 // CHECK: [[VAR_LHS:%.+]] = bitcast i8* [[VAR_LHS_VOID]] to [[S_INT_TY]]* 938 939 // var1_lhs = (S<i{{[0-9]+}}>*)lhs[2]; 940 // CHECK: [[VAR1_RHS_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST_RHS]], i64 0, i64 2 941 // CHECK: [[VAR1_RHS_VOID:%.+]] = load i8*, i8** [[VAR1_RHS_REF]], 942 // CHECK: [[VAR1_RHS:%.+]] = bitcast i8* [[VAR1_RHS_VOID]] to [[S_INT_TY]]* 943 // var1_rhs = (S<i{{[0-9]+}}>*)rhs[2]; 944 // CHECK: [[VAR1_LHS_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST_LHS]], i64 0, i64 2 945 // CHECK: [[VAR1_LHS_VOID:%.+]] = load i8*, i8** [[VAR1_LHS_REF]], 946 // CHECK: [[VAR1_LHS:%.+]] = bitcast i8* [[VAR1_LHS_VOID]] to [[S_INT_TY]]* 947 948 // t_var1_lhs = (i{{[0-9]+}}*)lhs[3]; 949 // CHECK: [[T_VAR1_RHS_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST_RHS]], i64 0, i64 3 950 // CHECK: [[T_VAR1_RHS_VOID:%.+]] = load i8*, i8** [[T_VAR1_RHS_REF]], 951 // CHECK: [[T_VAR1_RHS:%.+]] = bitcast i8* [[T_VAR1_RHS_VOID]] to i{{[0-9]+}}* 952 // t_var1_rhs = (i{{[0-9]+}}*)rhs[3]; 953 // CHECK: [[T_VAR1_LHS_REF:%.+]] = getelementptr inbounds [4 x i8*], [4 x i8*]* [[RED_LIST_LHS]], i64 0, i64 3 954 // CHECK: [[T_VAR1_LHS_VOID:%.+]] = load i8*, i8** [[T_VAR1_LHS_REF]], 955 // CHECK: [[T_VAR1_LHS:%.+]] = bitcast i8* [[T_VAR1_LHS_VOID]] to i{{[0-9]+}}* 956 957 // t_var_lhs += t_var_rhs; 958 // CHECK: add nsw i32 1513, % 959 960 // var_lhs = var_lhs.operator &(var_rhs); 961 // CHECK: call void @_Z3redR6BaseS1RKS_(% 962 963 // var1_lhs = var1_lhs.operator &&(var1_rhs); 964 // CHECK: mul nsw i32 17, % 965 966 // t_var1_lhs = min(t_var1_lhs, t_var1_rhs); 967 // CHECK: sub nsw i32 47, % 968 // CHECK: ret void 969 970 // CHECK: define internal void [[TMAIN_MICROTASK2]](i{{[0-9]+}}* noalias [[GTID_ADDR:%.+]], i{{[0-9]+}}* noalias %{{.+}}, [42 x [[S_INT_TY]]]* dereferenceable(504) %{{.*}}, [2 x i32]* dereferenceable(8) %{{.*}}, i32* dereferenceable(4) %{{.*}}, [2 x [[S_INT_TY]]]* dereferenceable(24) %{{.*}}, [[S_INT_TY]]* dereferenceable(12) %{{.*}}) 971 972 // CHECK: [[ARR_ORIG_ADDR:%.+]] = alloca [42 x [[S_INT_TY]]]*, 973 // CHECK: [[ARR_PRIV:%.+]] = alloca [40 x [[S_INT_TY]]], 974 975 // Reduction list for runtime. 976 // CHECK: [[RED_LIST:%.+]] = alloca [1 x i8*], 977 978 // CHECK: store i{{[0-9]+}}* [[GTID_ADDR]], i{{[0-9]+}}** [[GTID_ADDR_ADDR:%.+]], 979 980 // CHECK: [[ARR_ORIG:%.+]] = load [42 x [[S_INT_TY]]]*, [42 x [[S_INT_TY]]]** [[ARR_ORIG_ADDR]], 981 // CHECK: [[LOW:%.+]] = getelementptr inbounds [42 x [[S_INT_TY]]], [42 x [[S_INT_TY]]]* [[ARR_ORIG]], i64 0, i64 1 982 // CHECK: [[ORIG_START:%.+]] = bitcast [42 x [[S_INT_TY]]]* [[ARR_ORIG]] to [[S_INT_TY]]* 983 // CHECK: [[START:%.+]] = ptrtoint [[S_INT_TY]]* [[ORIG_START]] to i64 984 // CHECK: [[LOW_BOUND:%.+]] = ptrtoint [[S_INT_TY]]* [[LOW]] to i64 985 // CHECK: [[OFFSET_BYTES:%.+]] = sub i64 [[START]], [[LOW_BOUND]] 986 // CHECK: [[OFFSET:%.+]] = sdiv exact i64 [[OFFSET_BYTES]], ptrtoint ([[S_INT_TY]]* getelementptr ([[S_INT_TY]], [[S_INT_TY]]* null, i32 1) to i64) 987 // CHECK: [[PSEUDO_ARR_PRIV:%.+]] = getelementptr [40 x [[S_INT_TY]]], [40 x [[S_INT_TY]]]* [[ARR_PRIV]], i64 [[OFFSET]] 988 // CHECK: [[ARR_PRIV:%.+]] = bitcast [40 x [[S_INT_TY]]]* [[PSEUDO_ARR_PRIV]] to [42 x [[S_INT_TY]]]* 989 990 // CHECK: ret void 991 992 #endif 993 994