1 // RUN: %clang_cc1 -fopenmp -x c++ %s -verify -debug-info-kind=limited -emit-llvm -o - -triple powerpc64le-unknown-linux-gnu | FileCheck %s 2 // expected-no-diagnostics 3 4 struct S { 5 float a; 6 S() : a(0.0f) {} 7 ~S() {} 8 }; 9 10 #pragma omp declare reduction(+:S:omp_out.a += omp_in.a) initializer(omp_priv = omp_orig) 11 12 float g; 13 14 int a; 15 #pragma omp threadprivate(a) 16 int main (int argc, char *argv[]) 17 { 18 int i, n; 19 float a[100], b[100], sum, e[argc + 100]; 20 S c[100]; 21 float &d = g; 22 23 /* Some initializations */ 24 n = 100; 25 for (i=0; i < n; i++) 26 a[i] = b[i] = i * 1.0; 27 sum = 0.0; 28 29 #pragma omp taskloop simd reduction(+:sum, c[:n], d, e) 30 for (i=0; i < n; i++) { 31 sum = sum + (a[i] * b[i]); 32 c[i].a = i*i; 33 d += i*i; 34 e[i] = i; 35 } 36 37 } 38 39 // CHECK-LABEL: @main( 40 // CHECK: [[RETVAL:%.*]] = alloca i32, align 4 41 // CHECK: [[ARGC_ADDR:%.*]] = alloca i32, align 4 42 // CHECK: [[ARGV_ADDR:%.*]] = alloca i8**, align 8 43 // CHECK: [[I:%.*]] = alloca i32, align 4 44 // CHECK: [[N:%.*]] = alloca i32, align 4 45 // CHECK: [[A:%.*]] = alloca [100 x float], align 4 46 // CHECK: [[B:%.*]] = alloca [100 x float], align 4 47 // CHECK: [[SUM:%.*]] = alloca float, align 4 48 // CHECK: [[SAVED_STACK:%.*]] = alloca i8*, align 8 49 // CHECK: [[C:%.*]] = alloca [100 x %struct.S], align 4 50 // CHECK: [[D:%.*]] = alloca float*, align 8 51 // CHECK: [[AGG_CAPTURED:%.*]] = alloca [[STRUCT_ANON:%.*]], align 8 52 // CHECK: [[TMP0:%.*]] = call i32 @__kmpc_global_thread_num(%ident_t* 53 // CHECK: [[DOTRD_INPUT_:%.*]] = alloca [4 x %struct.kmp_task_red_input_t], align 8 54 // CHECK: [[DOTCAPTURE_EXPR_:%.*]] = alloca i32, align 4 55 // CHECK: [[DOTCAPTURE_EXPR_9:%.*]] = alloca i32, align 4 56 // CHECK: store i32 0, i32* [[RETVAL]], align 4 57 // CHECK: store i32 [[ARGC:%.*]], i32* [[ARGC_ADDR]], align 4 58 // CHECK: store i8** [[ARGV:%.*]], i8*** [[ARGV_ADDR]], align 8 59 // CHECK: [[TMP1:%.*]] = load i32, i32* [[ARGC_ADDR]], align 4 60 // CHECK: [[ADD:%.*]] = add nsw i32 [[TMP1]], 100 61 // CHECK: [[TMP2:%.*]] = zext i32 [[ADD]] to i64 62 // CHECK: [[VLA:%.+]] = alloca float, i64 % 63 64 // CHECK: call void @__kmpc_taskgroup(%ident_t* 65 // CHECK: [[DOTRD_INPUT_GEP_:%.*]] = getelementptr inbounds [4 x %struct.kmp_task_red_input_t], [4 x %struct.kmp_task_red_input_t]* [[DOTRD_INPUT_]], i64 0, i64 0 66 // CHECK: [[TMP20:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_RED_INPUT_T:%.*]], %struct.kmp_task_red_input_t* [[DOTRD_INPUT_GEP_]], i32 0, i32 0 67 // CHECK: [[TMP21:%.*]] = bitcast float* [[SUM]] to i8* 68 // CHECK: store i8* [[TMP21]], i8** [[TMP20]], align 8 69 // CHECK: [[TMP22:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_RED_INPUT_T]], %struct.kmp_task_red_input_t* [[DOTRD_INPUT_GEP_]], i32 0, i32 1 70 // CHECK: store i64 4, i64* [[TMP22]], align 8 71 // CHECK: [[TMP23:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_RED_INPUT_T]], %struct.kmp_task_red_input_t* [[DOTRD_INPUT_GEP_]], i32 0, i32 2 72 // CHECK: store i8* bitcast (void (i8*)* [[RED_INIT1:@.+]] to i8*), i8** [[TMP23]], align 8 73 // CHECK: [[TMP24:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_RED_INPUT_T]], %struct.kmp_task_red_input_t* [[DOTRD_INPUT_GEP_]], i32 0, i32 3 74 // CHECK: store i8* null, i8** [[TMP24]], align 8 75 // CHECK: [[TMP25:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_RED_INPUT_T]], %struct.kmp_task_red_input_t* [[DOTRD_INPUT_GEP_]], i32 0, i32 4 76 // CHECK: store i8* bitcast (void (i8*, i8*)* [[RED_COMB1:@.+]] to i8*), i8** [[TMP25]], align 8 77 // CHECK: [[TMP26:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_RED_INPUT_T]], %struct.kmp_task_red_input_t* [[DOTRD_INPUT_GEP_]], i32 0, i32 5 78 // CHECK: [[TMP27:%.*]] = bitcast i32* [[TMP26]] to i8* 79 // CHECK: call void @llvm.memset.p0i8.i64(i8* [[TMP27]], i8 0, i64 4, i32 8, i1 false) 80 // CHECK: [[DOTRD_INPUT_GEP_4:%.*]] = getelementptr inbounds [4 x %struct.kmp_task_red_input_t], [4 x %struct.kmp_task_red_input_t]* [[DOTRD_INPUT_]], i64 0, i64 1 81 // CHECK: [[TMP28:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_RED_INPUT_T]], %struct.kmp_task_red_input_t* [[DOTRD_INPUT_GEP_4]], i32 0, i32 0 82 // CHECK: [[ARRAYIDX5:%.*]] = getelementptr inbounds [100 x %struct.S], [100 x %struct.S]* [[C]], i64 0, i64 0 83 // CHECK: [[TMP29:%.*]] = load i32, i32* [[N]], align 4 84 // CHECK: [[TMP30:%.*]] = sext i32 [[TMP29]] to i64 85 // CHECK: [[LB_ADD_LEN:%.*]] = add nsw i64 -1, [[TMP30]] 86 // CHECK: [[ARRAYIDX6:%.*]] = getelementptr inbounds [100 x %struct.S], [100 x %struct.S]* [[C]], i64 0, i64 [[LB_ADD_LEN]] 87 // CHECK: [[TMP31:%.*]] = bitcast %struct.S* [[ARRAYIDX5]] to i8* 88 // CHECK: store i8* [[TMP31]], i8** [[TMP28]], align 8 89 // CHECK: [[TMP32:%.*]] = ptrtoint %struct.S* [[ARRAYIDX6]] to i64 90 // CHECK: [[TMP33:%.*]] = ptrtoint %struct.S* [[ARRAYIDX5]] to i64 91 // CHECK: [[TMP34:%.*]] = sub i64 [[TMP32]], [[TMP33]] 92 // CHECK: [[TMP35:%.*]] = sdiv exact i64 [[TMP34]], ptrtoint (float* getelementptr (float, float* null, i32 1) to i64) 93 // CHECK: [[TMP36:%.*]] = add nuw i64 [[TMP35]], 1 94 // CHECK: [[TMP37:%.*]] = mul nuw i64 [[TMP36]], ptrtoint (float* getelementptr (float, float* null, i32 1) to i64) 95 // CHECK: [[TMP38:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_RED_INPUT_T]], %struct.kmp_task_red_input_t* [[DOTRD_INPUT_GEP_4]], i32 0, i32 1 96 // CHECK: store i64 [[TMP37]], i64* [[TMP38]], align 8 97 // CHECK: [[TMP39:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_RED_INPUT_T]], %struct.kmp_task_red_input_t* [[DOTRD_INPUT_GEP_4]], i32 0, i32 2 98 // CHECK: store i8* bitcast (void (i8*)* [[RED_INIT2:@.+]] to i8*), i8** [[TMP39]], align 8 99 // CHECK: [[TMP40:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_RED_INPUT_T]], %struct.kmp_task_red_input_t* [[DOTRD_INPUT_GEP_4]], i32 0, i32 3 100 // CHECK: store i8* bitcast (void (i8*)* [[RED_FINI2:@.+]] to i8*), i8** [[TMP40]], align 8 101 // CHECK: [[TMP41:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_RED_INPUT_T]], %struct.kmp_task_red_input_t* [[DOTRD_INPUT_GEP_4]], i32 0, i32 4 102 // CHECK: store i8* bitcast (void (i8*, i8*)* [[RED_COMB2:@.+]] to i8*), i8** [[TMP41]], align 8 103 // CHECK: [[TMP42:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_RED_INPUT_T]], %struct.kmp_task_red_input_t* [[DOTRD_INPUT_GEP_4]], i32 0, i32 5 104 // CHECK: store i32 1, i32* [[TMP42]], align 8 105 // CHECK: [[DOTRD_INPUT_GEP_7:%.*]] = getelementptr inbounds [4 x %struct.kmp_task_red_input_t], [4 x %struct.kmp_task_red_input_t]* [[DOTRD_INPUT_]], i64 0, i64 2 106 // CHECK: [[TMP43:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_RED_INPUT_T]], %struct.kmp_task_red_input_t* [[DOTRD_INPUT_GEP_7]], i32 0, i32 0 107 // CHECK: [[TMP44:%.*]] = load float*, float** [[D]], align 8 108 // CHECK: [[TMP45:%.*]] = bitcast float* [[TMP44]] to i8* 109 // CHECK: store i8* [[TMP45]], i8** [[TMP43]], align 8 110 // CHECK: [[TMP46:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_RED_INPUT_T]], %struct.kmp_task_red_input_t* [[DOTRD_INPUT_GEP_7]], i32 0, i32 1 111 // CHECK: store i64 4, i64* [[TMP46]], align 8 112 // CHECK: [[TMP47:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_RED_INPUT_T]], %struct.kmp_task_red_input_t* [[DOTRD_INPUT_GEP_7]], i32 0, i32 2 113 // CHECK: store i8* bitcast (void (i8*)* [[RED_INIT3:@.+]] to i8*), i8** [[TMP47]], align 8 114 // CHECK: [[TMP48:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_RED_INPUT_T]], %struct.kmp_task_red_input_t* [[DOTRD_INPUT_GEP_7]], i32 0, i32 3 115 // CHECK: store i8* null, i8** [[TMP48]], align 8 116 // CHECK: [[TMP49:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_RED_INPUT_T]], %struct.kmp_task_red_input_t* [[DOTRD_INPUT_GEP_7]], i32 0, i32 4 117 // CHECK: store i8* bitcast (void (i8*, i8*)* [[RED_COMB3:@.+]] to i8*), i8** [[TMP49]], align 8 118 // CHECK: [[TMP50:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_RED_INPUT_T]], %struct.kmp_task_red_input_t* [[DOTRD_INPUT_GEP_7]], i32 0, i32 5 119 // CHECK: [[TMP51:%.*]] = bitcast i32* [[TMP50]] to i8* 120 // CHECK: call void @llvm.memset.p0i8.i64(i8* [[TMP51]], i8 0, i64 4, i32 8, i1 false) 121 // CHECK: [[DOTRD_INPUT_GEP_8:%.*]] = getelementptr inbounds [4 x %struct.kmp_task_red_input_t], [4 x %struct.kmp_task_red_input_t]* [[DOTRD_INPUT_]], i64 0, i64 3 122 // CHECK: [[TMP52:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_RED_INPUT_T]], %struct.kmp_task_red_input_t* [[DOTRD_INPUT_GEP_8]], i32 0, i32 0 123 // CHECK: [[TMP53:%.*]] = bitcast float* [[VLA]] to i8* 124 // CHECK: store i8* [[TMP53]], i8** [[TMP52]], align 8 125 // CHECK: [[TMP54:%.*]] = mul nuw i64 [[TMP2]], 4 126 // CHECK: [[TMP55:%.*]] = udiv exact i64 [[TMP54]], ptrtoint (float* getelementptr (float, float* null, i32 1) to i64) 127 // CHECK: [[TMP56:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_RED_INPUT_T]], %struct.kmp_task_red_input_t* [[DOTRD_INPUT_GEP_8]], i32 0, i32 1 128 // CHECK: store i64 [[TMP54]], i64* [[TMP56]], align 8 129 // CHECK: [[TMP57:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_RED_INPUT_T]], %struct.kmp_task_red_input_t* [[DOTRD_INPUT_GEP_8]], i32 0, i32 2 130 // CHECK: store i8* bitcast (void (i8*)* [[RED_INIT4:@.+]] to i8*), i8** [[TMP57]], align 8 131 // CHECK: [[TMP58:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_RED_INPUT_T]], %struct.kmp_task_red_input_t* [[DOTRD_INPUT_GEP_8]], i32 0, i32 3 132 // CHECK: store i8* null, i8** [[TMP58]], align 8 133 // CHECK: [[TMP59:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_RED_INPUT_T]], %struct.kmp_task_red_input_t* [[DOTRD_INPUT_GEP_8]], i32 0, i32 4 134 // CHECK: store i8* bitcast (void (i8*, i8*)* [[RED_COMB4:@.+]] to i8*), i8** [[TMP59]], align 8 135 // CHECK: [[TMP60:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_RED_INPUT_T]], %struct.kmp_task_red_input_t* [[DOTRD_INPUT_GEP_8]], i32 0, i32 5 136 // CHECK: store i32 1, i32* [[TMP60]], align 8 137 // CHECK: [[TMP61:%.*]] = bitcast [4 x %struct.kmp_task_red_input_t]* [[DOTRD_INPUT_]] to i8* 138 // CHECK: [[TMP62:%.*]] = call i8* @__kmpc_task_reduction_init(i32 [[TMP0]], i32 4, i8* [[TMP61]]) 139 // CHECK: [[TMP63:%.*]] = load i32, i32* [[N]], align 4 140 // CHECK: store i32 [[TMP63]], i32* [[DOTCAPTURE_EXPR_]], align 4 141 // CHECK: [[TMP64:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_]], align 4 142 // CHECK: [[SUB:%.*]] = sub nsw i32 [[TMP64]], 0 143 // CHECK: [[SUB10:%.*]] = sub nsw i32 [[SUB]], 1 144 // CHECK: [[ADD11:%.*]] = add nsw i32 [[SUB10]], 1 145 // CHECK: [[DIV:%.*]] = sdiv i32 [[ADD11]], 1 146 // CHECK: [[SUB12:%.*]] = sub nsw i32 [[DIV]], 1 147 // CHECK: store i32 [[SUB12]], i32* [[DOTCAPTURE_EXPR_9]], align 4 148 // CHECK: [[TMP65:%.*]] = call i8* @__kmpc_omp_task_alloc(%ident_t* %{{.+}}, i32 [[TMP0]], i32 1, i64 888, i64 72, i32 (i32, i8*)* bitcast (i32 (i32, %struct.kmp_task_t_with_privates*)* @{{.+}} to i32 (i32, i8*)*)) 149 // CHECK: [[TMP66:%.*]] = bitcast i8* [[TMP65]] to %struct.kmp_task_t_with_privates* 150 // CHECK: [[TMP67:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES:%.*]], %struct.kmp_task_t_with_privates* [[TMP66]], i32 0, i32 0 151 // CHECK: [[TMP68:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T:%.*]], %struct.kmp_task_t* [[TMP67]], i32 0, i32 0 152 // CHECK: [[TMP69:%.*]] = load i8*, i8** [[TMP68]], align 8 153 // CHECK: [[TMP70:%.*]] = bitcast %struct.anon* [[AGG_CAPTURED]] to i8* 154 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP69]], i8* [[TMP70]], i64 72, i32 8, i1 false) 155 // CHECK: [[TMP71:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T_WITH_PRIVATES]], %struct.kmp_task_t_with_privates* [[TMP66]], i32 0, i32 1 156 // CHECK: [[TMP72:%.*]] = bitcast i8* [[TMP69]] to %struct.anon* 157 // CHECK: [[TMP73:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T:%.*]], %struct..kmp_privates.t* [[TMP71]], i32 0, i32 0 158 // CHECK: [[TMP74:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[TMP72]], i32 0, i32 1 159 // CHECK: [[REF:%.*]] = load i32*, i32** [[TMP74]], align 8 160 // CHECK: [[TMP75:%.*]] = load i32, i32* [[REF]], align 4 161 // CHECK: store i32 [[TMP75]], i32* [[TMP73]], align 8 162 // CHECK: [[TMP76:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP71]], i32 0, i32 1 163 // CHECK: [[TMP77:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[TMP72]], i32 0, i32 3 164 // CHECK: [[REF13:%.*]] = load [100 x float]*, [100 x float]** [[TMP77]], align 8 165 // CHECK: [[TMP78:%.*]] = bitcast [100 x float]* [[TMP76]] to i8* 166 // CHECK: [[TMP79:%.*]] = bitcast [100 x float]* [[REF13]] to i8* 167 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP78]], i8* [[TMP79]], i64 400, i32 4, i1 false) 168 // CHECK: [[TMP80:%.*]] = getelementptr inbounds [[STRUCT__KMP_PRIVATES_T]], %struct..kmp_privates.t* [[TMP71]], i32 0, i32 2 169 // CHECK: [[TMP81:%.*]] = getelementptr inbounds [[STRUCT_ANON]], %struct.anon* [[TMP72]], i32 0, i32 4 170 // CHECK: [[REF14:%.*]] = load [100 x float]*, [100 x float]** [[TMP81]], align 8 171 // CHECK: [[TMP82:%.*]] = bitcast [100 x float]* [[TMP80]] to i8* 172 // CHECK: [[TMP83:%.*]] = bitcast [100 x float]* [[REF14]] to i8* 173 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP82]], i8* [[TMP83]], i64 400, i32 4, i1 false) 174 // CHECK: [[TMP84:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP67]], i32 0, i32 5 175 // CHECK: store i64 0, i64* [[TMP84]], align 8 176 // CHECK: [[TMP85:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP67]], i32 0, i32 6 177 // CHECK: [[TMP86:%.*]] = load i32, i32* [[DOTCAPTURE_EXPR_9]], align 4 178 // CHECK: [[CONV15:%.*]] = sext i32 [[TMP86]] to i64 179 // CHECK: store i64 [[CONV15]], i64* [[TMP85]], align 8 180 // CHECK: [[TMP87:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP67]], i32 0, i32 7 181 // CHECK: store i64 1, i64* [[TMP87]], align 8 182 // CHECK: [[TMP88:%.*]] = getelementptr inbounds [[STRUCT_KMP_TASK_T]], %struct.kmp_task_t* [[TMP67]], i32 0, i32 9 183 // CHECK: store i8* [[TMP62]], i8** [[TMP88]], align 8 184 // CHECK: [[TMP89:%.*]] = load i64, i64* [[TMP87]], align 8 185 // CHECK: call void @__kmpc_taskloop(%ident_t* %{{.+}}, i32 [[TMP0]], i8* [[TMP65]], i32 1, i64* [[TMP84]], i64* [[TMP85]], i64 [[TMP89]], i32 0, i32 0, i64 0, i8* null) 186 // CHECK: call void @__kmpc_end_taskgroup(%ident_t* 187 188 // CHECK: ret i32 189 190 // CHECK: define internal void [[RED_INIT1]](i8*) 191 // CHECK: store float 0.000000e+00, float* % 192 // CHECK: ret void 193 194 // CHECK: define internal void [[RED_COMB1]](i8*, i8*) 195 // CHECK: fadd float %6, %7 196 // CHECK: store float %{{.+}}, float* % 197 // CHECK: ret void 198 199 // CHECK: define internal void [[RED_INIT2]](i8*) 200 // CHECK: call i8* @__kmpc_threadprivate_cached( 201 // CHECK: call i8* @__kmpc_threadprivate_cached( 202 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64( 203 // CHECK: ret void 204 205 // CHECK: define internal void [[RED_FINI2]](i8*) 206 // CHECK: call i8* @__kmpc_threadprivate_cached( 207 // CHECK: call void @ 208 // CHECK: ret void 209 210 // CHECK: define internal void [[RED_COMB2]](i8*, i8*) 211 // CHECK: call i8* @__kmpc_threadprivate_cached( 212 // CHECK: fadd float % 213 // CHECK: store float %{{.+}}, float* % 214 // CHECK: ret void 215 216 // CHECK: define internal void [[RED_INIT3]](i8*) 217 // CHECK: store float 0.000000e+00, float* % 218 // CHECK: ret void 219 220 // CHECK: define internal void [[RED_COMB3]](i8*, i8*) 221 // CHECK: fadd float % 222 // CHECK: store float %{{.+}}, float* % 223 // CHECK: ret void 224 225 // CHECK: define internal void [[RED_INIT4]](i8*) 226 // CHECK: call i8* @__kmpc_threadprivate_cached( 227 // CHECK: store float 0.000000e+00, float* % 228 // CHECK: ret void 229 230 // CHECK: define internal void [[RED_COMB4]](i8*, i8*) 231 // CHECK: call i8* @__kmpc_threadprivate_cached( 232 // CHECK: fadd float % 233 // CHECK: store float %{{.+}}, float* % 234 // CHECK: ret void 235 236