clang/test/OpenMP/taskloop_reduction_codegen.cpp

   1 // RUN: %clang_cc1 -fopenmp -x c++ %s -verify -Wno-vla -debug-info-kind=limited -emit-llvm -o - -triple powerpc64le-unknown-linux-gnu -std=c++98 | FileCheck %s
   2
   3 // RUN: %clang_cc1 -fopenmp-simd -x c++ %s -verify -Wno-vla -debug-info-kind=limited -emit-llvm -o - -triple powerpc64le-unknown-linux-gnu -std=c++98 | FileCheck --check-prefix SIMD-ONLY0 %s
   4 // SIMD-ONLY0-NOT: {{__kmpc|__tgt}}
   5 // expected-no-diagnostics
   6
   7 // CHECK: [[RED_SIZE1:@reduction_size[.].+]] = common thread_local global i64 0
   8 // CHECK: [[RED_SIZE2:@reduction_size[.].+]] = common thread_local global i64 0
   9
  10 struct S {
  11   float a;
  12   S() : a(0.0f) {}
  13   ~S() {}
  14 };
  15
  16 #pragma omp declare reduction(+:S:omp_out.a += omp_in.a) initializer(omp_priv = omp_orig)
  17
  18 float g;
  19
  20 int a;
  21 #pragma omp threadprivate(a)
  22 int main (int argc, char *argv[])
  23 {
  24 int   i, n;
  25 float a[100], b[100], sum, e[argc + 100];
  26 S c[100];
  27 float &d = g;
  28
  29 /* Some initializations */
  30 n = 100;
  31 for (i=0; i < n; i++)
  32   a[i] = b[i] = i * 1.0;
  33 sum = 0.0;
  34
  35 #pragma omp taskloop reduction(+:sum, c[:n], d, e)
  36   for (i=0; i < n; i++) {
  37     sum = sum + (a[i] * b[i]);
  38     c[i].a = i*i;
  39     d += i*i;
  40     e[i] = i;
  41   }
  42
  43 }
  44
  45 // CHECK-LABEL: @main(
  46 // CHECK:    [[RETVAL:%.*]] = alloca i32,
  47 // CHECK:    [[ARGC_ADDR:%.*]] = alloca i32,
  48 // CHECK:    [[ARGV_ADDR:%.*]] = alloca ptr,
  49 // CHECK:    [[I:%.*]] = alloca i32,
  50 // CHECK:    [[N:%.*]] = alloca i32,
  51 // CHECK:    [[A:%.*]] = alloca [100 x float],
  52 // CHECK:    [[B:%.*]] = alloca [100 x float],
  53 // CHECK:    [[SUM:%.*]] = alloca float,
  54 // CHECK:    [[SAVED_STACK:%.*]] = alloca ptr,
  55 // CHECK:    [[C:%.*]] = alloca [100 x %struct.S],
  56 // CHECK:    [[D:%.*]] = alloca ptr,
  57 // CHECK:    [[AGG_CAPTURED:%.*]] = alloca [[STRUCT_ANON:%.*]],
  58 // CHECK:    [[DOTRD_INPUT_:%.*]] = alloca [4 x %struct.kmp_taskred_input_t],
  59 // CHECK:    alloca i32,
  60 // CHECK:    [[DOTCAPTURE_EXPR_:%.*]] = alloca i32,
  61 // CHECK:    [[DOTCAPTURE_EXPR_9:%.*]] = alloca i32,
  62 // CHECK:    [[TMP0:%.*]] = call i32 @__kmpc_global_thread_num(ptr
  63 // CHECK:    store i32 0, ptr [[RETVAL]],
  64 // CHECK:    store i32 [[ARGC:%.*]], ptr [[ARGC_ADDR]],
  65 // CHECK:    store ptr [[ARGV:%.*]], ptr [[ARGV_ADDR]],
  66 // CHECK:    [[TMP1:%.*]] = load i32, ptr [[ARGC_ADDR]],
  67 // CHECK:    [[ADD:%.*]] = add nsw i32 [[TMP1]], 100
  68 // CHECK:    [[TMP2:%.*]] = zext i32 [[ADD]] to i64
  69 // CHECK:    [[VLA:%.+]] = alloca float, i64 %
  70
  71 // CHECK:    call void @__kmpc_taskgroup(ptr
  72 // CHECK-DAG:    store ptr [[SUM]], ptr [[TMP20:%[^,]+]],
  73 // CHECK-DAG:    [[TMP20]] = getelementptr inbounds nuw %struct.kmp_taskred_input_t, ptr [[DOTRD_INPUT_GEP_:%.+]], i32 0, i32 0
  74 // CHECK-DAG:    store ptr [[SUM]], ptr [[TMP20:%[^,]+]],
  75 // CHECK-DAG:    [[TMP20]] = getelementptr inbounds nuw %struct.kmp_taskred_input_t, ptr [[DOTRD_INPUT_GEP_]], i32 0, i32 1
  76 // CHECK-DAG:    [[TMP22:%.*]] = getelementptr inbounds nuw %struct.kmp_taskred_input_t, ptr [[DOTRD_INPUT_GEP_]], i32 0, i32 2
  77 // CHECK-DAG:    store i64 4, ptr [[TMP22]],
  78 // CHECK-DAG:    [[TMP23:%.*]] = getelementptr inbounds nuw %struct.kmp_taskred_input_t, ptr [[DOTRD_INPUT_GEP_]], i32 0, i32 3
  79 // CHECK-DAG:    store ptr @[[RED_INIT1:.+]], ptr [[TMP23]],
  80 // CHECK-DAG:    [[TMP24:%.*]] = getelementptr inbounds nuw %struct.kmp_taskred_input_t, ptr [[DOTRD_INPUT_GEP_]], i32 0, i32 4
  81 // CHECK-DAG:    store ptr null, ptr [[TMP24]],
  82 // CHECK-DAG:    [[TMP25:%.*]] = getelementptr inbounds nuw %struct.kmp_taskred_input_t, ptr [[DOTRD_INPUT_GEP_]], i32 0, i32 5
  83 // CHECK-DAG:    store ptr @[[RED_COMB1:.+]], ptr [[TMP25]],
  84 // CHECK-DAG:    [[TMP26:%.*]] = getelementptr inbounds nuw %struct.kmp_taskred_input_t, ptr [[DOTRD_INPUT_GEP_]], i32 0, i32 6
  85 // CHECK-DAG:    call void @llvm.memset.p0.i64(ptr align 8 [[TMP26]], i8 0, i64 4, i1 false)
  86 // CHECK-DAG:    [[ARRAYIDX5:%.*]] = getelementptr inbounds nuw [100 x %struct.S], ptr [[C]], i64 0, i64 0
  87 // CHECK-DAG:    [[LB_ADD_LEN:%.*]] = add nsw i64 -1, %
  88 // CHECK-DAG:    [[ARRAYIDX6:%.*]] = getelementptr inbounds nuw [100 x %struct.S], ptr [[C]], i64 0, i64 [[LB_ADD_LEN]]
  89 // CHECK-DAG:    store ptr [[ARRAYIDX5]], ptr [[TMP28:%[^,]+]],
  90 // CHECK-DAG:    [[TMP28]] = getelementptr inbounds nuw %struct.kmp_taskred_input_t, ptr [[DOTRD_INPUT_GEP_4:%.+]], i32 0, i32 0
  91 // CHECK-DAG:    store ptr [[ARRAYIDX5]], ptr [[TMP28:%[^,]+]],
  92 // CHECK-DAG:    [[TMP28]] = getelementptr inbounds nuw %struct.kmp_taskred_input_t, ptr [[DOTRD_INPUT_GEP_4]], i32 0, i32 1
  93 // CHECK-DAG:    [[TMP32:%.*]] = ptrtoint ptr [[ARRAYIDX6]] to i64
  94 // CHECK-DAG:    [[TMP33:%.*]] = ptrtoint ptr [[ARRAYIDX5]] to i64
  95 // CHECK-DAG:    [[TMP34:%.*]] = sub i64 [[TMP32]], [[TMP33]]
  96 // CHECK-DAG:    [[TMP35:%.*]] = sdiv exact i64 [[TMP34]], ptrtoint (ptr getelementptr (%struct.S, ptr null, i32 1) to i64)
  97 // CHECK-DAG:    [[TMP36:%.*]] = add nuw i64 [[TMP35]], 1
  98 // CHECK-DAG:    [[TMP37:%.*]] = mul nuw i64 [[TMP36]], ptrtoint (ptr getelementptr (%struct.S, ptr null, i32 1) to i64)
  99 // CHECK-DAG:    store i64 [[TMP37]], ptr [[TMP38:%[^,]+]],
 100 // CHECK-DAG:    [[TMP38]] = getelementptr inbounds nuw %struct.kmp_taskred_input_t, ptr [[DOTRD_INPUT_GEP_4]], i32 0, i32 2
 101 // CHECK-DAG:    [[TMP39:%.*]] = getelementptr inbounds nuw %struct.kmp_taskred_input_t, ptr [[DOTRD_INPUT_GEP_4]], i32 0, i32 3
 102 // CHECK-DAG:    store ptr @[[RED_INIT2:.+]], ptr [[TMP39]],
 103 // CHECK-DAG:    [[TMP40:%.*]] = getelementptr inbounds nuw %struct.kmp_taskred_input_t, ptr [[DOTRD_INPUT_GEP_4]], i32 0, i32 4
 104 // CHECK-DAG:    store ptr @[[RED_FINI2:.+]], ptr [[TMP40]],
 105 // CHECK-DAG:    [[TMP41:%.*]] = getelementptr inbounds nuw %struct.kmp_taskred_input_t, ptr [[DOTRD_INPUT_GEP_4]], i32 0, i32 5
 106 // CHECK-DAG:    store ptr @[[RED_COMB2:.+]], ptr [[TMP41]],
 107 // CHECK-DAG:    [[TMP42:%.*]] = getelementptr inbounds nuw %struct.kmp_taskred_input_t, ptr [[DOTRD_INPUT_GEP_4]], i32 0, i32 6
 108 // CHECK-DAG:    store i32 1, ptr [[TMP42]],
 109 // CHECK-DAG:    [[TMP44:%.*]] = load ptr, ptr [[D]],
 110 // CHECK-DAG:    store ptr [[TMP44]], ptr [[TMP43:%[^,]+]],
 111 // CHECK-DAG:    [[TMP43]] = getelementptr inbounds nuw %struct.kmp_taskred_input_t, ptr [[DOTRD_INPUT_GEP_7:%.+]], i32 0, i32 0
 112 // CHECK-DAG:    store ptr [[TMP44]], ptr [[TMP43:%[^,]+]],
 113 // CHECK-DAG:    [[TMP43]] = getelementptr inbounds nuw %struct.kmp_taskred_input_t, ptr [[DOTRD_INPUT_GEP_7]], i32 0, i32 1
 114 // CHECK-DAG:    [[TMP46:%.*]] = getelementptr inbounds nuw %struct.kmp_taskred_input_t, ptr [[DOTRD_INPUT_GEP_7]], i32 0, i32 2
 115 // CHECK-DAG:    store i64 4, ptr [[TMP46]],
 116 // CHECK-DAG:    [[TMP47:%.*]] = getelementptr inbounds nuw %struct.kmp_taskred_input_t, ptr [[DOTRD_INPUT_GEP_7]], i32 0, i32 3
 117 // CHECK-DAG:    store ptr @[[RED_INIT3:.+]], ptr [[TMP47]],
 118 // CHECK-DAG:    [[TMP48:%.*]] = getelementptr inbounds nuw %struct.kmp_taskred_input_t, ptr [[DOTRD_INPUT_GEP_7]], i32 0, i32 4
 119 // CHECK-DAG:    store ptr null, ptr [[TMP48]],
 120 // CHECK-DAG:    [[TMP49:%.*]] = getelementptr inbounds nuw %struct.kmp_taskred_input_t, ptr [[DOTRD_INPUT_GEP_7]], i32 0, i32 5
 121 // CHECK-DAG:    store ptr @[[RED_COMB3:.+]], ptr [[TMP49]],
 122 // CHECK-DAG:    [[TMP50:%.*]] = getelementptr inbounds nuw %struct.kmp_taskred_input_t, ptr [[DOTRD_INPUT_GEP_7]], i32 0, i32 6
 123 // CHECK-DAG:    call void @llvm.memset.p0.i64(ptr align 8 [[TMP50]], i8 0, i64 4, i1 false)
 124 // CHECK-DAG:    store ptr [[VLA]], ptr [[TMP52:%[^,]+]],
 125 // CHECK-DAG:    [[TMP52]] = getelementptr inbounds nuw %struct.kmp_taskred_input_t, ptr [[DOTRD_INPUT_GEP_8:%.+]], i32 0, i32 0
 126 // CHECK-DAG:    store ptr [[VLA]], ptr [[TMP52:%[^,]+]],
 127 // CHECK-DAG:    [[TMP52]] = getelementptr inbounds nuw %struct.kmp_taskred_input_t, ptr [[DOTRD_INPUT_GEP_8]], i32 0, i32 1
 128 // CHECK-DAG:    [[TMP54:%.*]] = mul nuw i64 [[TMP2]], 4
 129 // CHECK-DAG:    [[TMP55:%.*]] = udiv exact i64 [[TMP54]], ptrtoint (ptr getelementptr (float, ptr null, i32 1) to i64)
 130 // CHECK-DAG:    store i64 [[TMP54]], ptr [[TMP56:%[^,]+]],
 131 // CHECK-DAG:    [[TMP56]] = getelementptr inbounds nuw %struct.kmp_taskred_input_t, ptr [[DOTRD_INPUT_GEP_8]], i32 0, i32 2
 132 // CHECK-DAG:    [[TMP57:%.*]] = getelementptr inbounds nuw %struct.kmp_taskred_input_t, ptr [[DOTRD_INPUT_GEP_8]], i32 0, i32 3
 133 // CHECK-DAG:    store ptr @[[RED_INIT4:.+]], ptr [[TMP57]],
 134 // CHECK-DAG:    [[TMP58:%.*]] = getelementptr inbounds nuw %struct.kmp_taskred_input_t, ptr [[DOTRD_INPUT_GEP_8]], i32 0, i32 4
 135 // CHECK-DAG:    store ptr null, ptr [[TMP58]],
 136 // CHECK-DAG:    [[TMP59:%.*]] = getelementptr inbounds nuw %struct.kmp_taskred_input_t, ptr [[DOTRD_INPUT_GEP_8]], i32 0, i32 5
 137 // CHECK-DAG:    store ptr @[[RED_COMB4:.+]], ptr [[TMP59]],
 138 // CHECK-DAG:    [[TMP60:%.*]] = getelementptr inbounds nuw %struct.kmp_taskred_input_t, ptr [[DOTRD_INPUT_GEP_8]], i32 0, i32 6
 139 // CHECK-DAG:    store i32 1, ptr [[TMP60]],
 140 // CHECK-DAG:    [[DOTRD_INPUT_GEP_]] = getelementptr inbounds nuw [4 x %struct.kmp_taskred_input_t], ptr [[DOTRD_INPUT_]], i64 0, i64
 141 // CHECK-DAG:    [[DOTRD_INPUT_GEP_4]] = getelementptr inbounds nuw [4 x %struct.kmp_taskred_input_t], ptr [[DOTRD_INPUT_]], i64 0, i64
 142 // CHECK-DAG:    [[DOTRD_INPUT_GEP_7]] = getelementptr inbounds nuw [4 x %struct.kmp_taskred_input_t], ptr [[DOTRD_INPUT_]], i64 0, i64
 143 // CHECK-DAG:    [[DOTRD_INPUT_GEP_8]] = getelementptr inbounds nuw [4 x %struct.kmp_taskred_input_t], ptr [[DOTRD_INPUT_]], i64 0, i64
 144 // CHECK:    [[TMP62:%.*]] = call ptr @__kmpc_taskred_init(i32 [[TMP0]], i32 4, ptr [[DOTRD_INPUT_]])
 145 // CHECK:    [[TMP63:%.*]] = load i32, ptr [[N]],
 146 // CHECK:    store i32 [[TMP63]], ptr [[DOTCAPTURE_EXPR_]],
 147 // CHECK:    [[TMP64:%.*]] = load i32, ptr [[DOTCAPTURE_EXPR_]],
 148 // CHECK:    [[SUB:%.*]] = sub nsw i32 [[TMP64]], 0
 149 // CHECK:    [[DIV:%.*]] = sdiv i32 [[SUB]], 1
 150 // CHECK:    [[SUB12:%.*]] = sub nsw i32 [[DIV]], 1
 151 // CHECK:    store i32 [[SUB12]], ptr [[DOTCAPTURE_EXPR_9]],
 152 // CHECK:    [[TMP65:%.*]] = call ptr @__kmpc_omp_task_alloc(ptr {{.+}}, i32 [[TMP0]], i32 1, i64 888, i64 40, ptr @[[TASK:.+]])
 153 // CHECK:    call void @__kmpc_taskloop(ptr {{.+}}, i32 [[TMP0]], ptr [[TMP65]], i32 1, ptr %{{.+}}, ptr %{{.+}}, i64 %{{.+}}, i32 1, i32 0, i64 0, ptr null)
 154 // CHECK:    call void @__kmpc_end_taskgroup(ptr
 155
 156 // CHECK:    ret i32
 157
 158 // CHECK: define internal void @[[RED_INIT1]](ptr noalias noundef %{{.+}}, ptr noalias noundef %{{.+}})
 159 // CHECK: store float 0.000000e+00, ptr %
 160 // CHECK: ret void
 161
 162 // CHECK: define internal void @[[RED_COMB1]](ptr noundef %0, ptr noundef %1)
 163 // CHECK: fadd float %
 164 // CHECK: store float %{{.+}}, ptr %
 165 // CHECK: ret void
 166
 167 // CHECK: define internal void @[[RED_INIT2]](ptr noalias noundef %{{.+}}, ptr noalias noundef %{{.+}})
 168 // CHECK: call void [[OMP_INIT1:@.+]](ptr
 169 // CHECK: ret void
 170
 171 // CHECK: define internal void [[OMP_COMB1:@.+]](ptr noalias noundef %0, ptr noalias noundef %1)
 172 // CHECK: fadd float %
 173
 174 // CHECK: define internal void [[OMP_INIT1]](ptr noalias noundef %0, ptr noalias noundef %1)
 175 // CHECK: call void @llvm.memcpy.p0.p0.i64(
 176
 177 // CHECK: define internal void @[[RED_FINI2]](ptr noundef %0)
 178 // CHECK: [[RED_SIZE1_ADDR:%.+]] = call align 8 ptr @llvm.threadlocal.address.p0(ptr align 8 [[RED_SIZE1]]
 179 // CHECK: load i64, ptr [[RED_SIZE1_ADDR]]
 180 // CHECK: call void @
 181 // CHECK: ret void
 182
 183 // CHECK: define internal void @[[RED_COMB2]](ptr noundef %0, ptr noundef %1)
 184 // CHECK: [[RED_SIZE1_ADDR2:%.+]] = call align 8 ptr @llvm.threadlocal.address.p0(ptr align 8 [[RED_SIZE1]]
 185 // CHECK: load i64, ptr [[RED_SIZE1_ADDR2]]
 186 // CHECK: call void [[OMP_COMB1]](
 187 // CHECK: ret void
 188
 189 // CHECK: define internal void @[[RED_INIT3]](ptr noalias noundef %{{.+}}, ptr noalias noundef %{{.+}})
 190 // CHECK: store float 0.000000e+00, ptr %
 191 // CHECK: ret void
 192
 193 // CHECK: define internal void @[[RED_COMB3]](ptr noundef %0, ptr noundef %1)
 194 // CHECK: fadd float %
 195 // CHECK: store float %{{.+}}, ptr %
 196 // CHECK: ret void
 197
 198 // CHECK: define internal void @[[RED_INIT4]](ptr noalias noundef %{{.+}}, ptr noalias noundef %{{.+}})
 199 // CHECK: [[RED_SIZE2_ADDR:%.+]] = call align 8 ptr @llvm.threadlocal.address.p0(ptr align 8 [[RED_SIZE2]]
 200 // CHECK: load i64, ptr [[RED_SIZE2_ADDR]]
 201 // CHECK: store float 0.000000e+00, ptr %
 202 // CHECK: ret void
 203
 204 // CHECK: define internal void @[[RED_COMB4]](ptr noundef %0, ptr noundef %1)
 205 // CHECK: [[RED_SIZE2_ADDR2:%.+]] = call align 8 ptr @llvm.threadlocal.address.p0(ptr align 8 [[RED_SIZE2]]
 206 // CHECK: load i64, ptr [[RED_SIZE2_ADDR2]]
 207 // CHECK: fadd float %
 208 // CHECK: store float %{{.+}}, ptr %
 209 // CHECK: ret void
 210
 211 // CHECK: call ptr @__kmpc_task_reduction_get_th_data(
 212 // CHECK: call ptr @__kmpc_task_reduction_get_th_data(
 213 // CHECK: call ptr @__kmpc_task_reduction_get_th_data(
 214 // CHECK: call ptr @__kmpc_task_reduction_get_th_data(
 215
 216 // CHECK-DAG: distinct !DISubprogram(linkageName: "[[TASK]]", scope: !
 217 // CHECK-DAG: !DISubprogram(linkageName: "[[RED_INIT1]]"
 218 // CHECK-DAG: !DISubprogram(linkageName: "[[RED_COMB1]]"
 219 // CHECK-DAG: !DISubprogram(linkageName: "[[RED_INIT2]]"
 220 // CHECK-DAG: !DISubprogram(linkageName: "[[RED_FINI2]]"
 221 // CHECK-DAG: !DISubprogram(linkageName: "[[RED_COMB2]]"
 222 // CHECK-DAG: !DISubprogram(linkageName: "[[RED_INIT3]]"
 223 // CHECK-DAG: !DISubprogram(linkageName: "[[RED_COMB3]]"
 224 // CHECK-DAG: !DISubprogram(linkageName: "[[RED_INIT4]]"
 225 // CHECK-DAG: !DISubprogram(linkageName: "[[RED_COMB4]]"