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