clang/test/OpenMP/scan_codegen.cpp

   1 // RUN: %clang_cc1 -verify -fopenmp -x c++ -triple x86_64-unknown-unknown -emit-llvm %s -o - | FileCheck %s
   2 // RUN: %clang_cc1 -fopenmp -x c++ -triple x86_64-unknown-unknown -emit-pch -o %t %s
   3 // RUN: %clang_cc1 -fopenmp -x c++ -triple x86_64-unknown-unknown -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s
   4
   5 // RUN: %clang_cc1 -verify -fopenmp-simd -x c++ -triple x86_64-unknown-unknown -emit-llvm %s -o - | FileCheck --check-prefix SIMD-ONLY0 %s
   6 // RUN: %clang_cc1 -fopenmp-simd -x c++ -triple x86_64-unknown-unknown -emit-pch -o %t %s
   7 // RUN: %clang_cc1 -fopenmp-simd -x c++ -triple x86_64-unknown-unknown -include-pch %t -verify %s -emit-llvm -o - | FileCheck --check-prefix SIMD-ONLY0 %s
   8 // SIMD-ONLY0-NOT: {{__kmpc|__tgt}}
   9 //
  10 // expected-no-diagnostics
  11 #ifndef HEADER
  12 #define HEADER
  13 void foo();
  14 void bar();
  15
  16 // CHECK-LABEL: baz
  17 void baz() {
  18   int a = 0;
  19
  20   // CHECK: store i32 0, ptr [[A_ADDR:%.+]],
  21   // CHECK: store i32 0, ptr [[OMP_CNT:%.+]],
  22   // CHECK: br label %[[OMP_HEADER:.+]]
  23
  24   // CHECK: [[OMP_HEADER]]:
  25   // CHECK: [[CNT_VAL:%.+]] = load i32, ptr [[OMP_CNT]],
  26   // CHECK: [[CMP:%.+]] = icmp slt i32 [[CNT_VAL]], 10
  27   // CHECK: br i1 [[CMP]], label %[[OMP_BODY:.+]], label %[[OMP_END:.+]]
  28 #pragma omp simd reduction(inscan, + : a)
  29   for (int i = 0; i < 10; ++i) {
  30     // CHECK: [[OMP_BODY]]:
  31
  32     // i = OMP_CNT*1 + 0;
  33     // CHECK: [[CNT_VAL:%.+]] = load i32, ptr [[OMP_CNT]],
  34     // CHECK: [[MUL:%.+]] = mul nsw i32 [[CNT_VAL]], 1
  35     // CHECK: [[ADD:%.+]] = add nsw i32 0, [[MUL]]
  36     // CHECK: store i32 [[ADD]], ptr [[I_ADDR:%.+]],
  37
  38     // A_PRIV = 0;
  39     // CHECK: store i32 0, ptr [[A_PRIV_ADDR:%.+]],
  40
  41     // goto DISPATCH;
  42     // CHECK: br label %[[DISPATCH:[^,]+]]
  43
  44     // INPUT_PHASE:
  45     // foo();
  46     // goto REDUCE;
  47     // CHECK: [[INPUT_PHASE:.+]]:
  48     // CHECK: call void @{{.*}}foo{{.*}}()
  49     // CHECK: br label %[[REDUCE:[^,]+]]
  50     foo();
  51
  52     // DISPATCH:
  53     // goto INPUT_PHASE;
  54     // CHECK: [[DISPATCH]]:
  55     // CHECK: br label %[[INPUT_PHASE]]
  56
  57     // REDUCE:
  58     // A = A_PRIV + A;
  59     // A_PRIV = A;
  60     // goto SCAN_PHASE;
  61     // CHECK: [[REDUCE]]:
  62     // CHECK: [[A:%.+]] = load i32, ptr [[A_ADDR]],
  63     // CHECK: [[A_PRIV:%.+]] = load i32, ptr [[A_PRIV_ADDR]],
  64     // CHECK: [[SUM:%.+]] = add nsw i32 [[A]], [[A_PRIV]]
  65     // CHECK: store i32 [[SUM]], ptr [[A_ADDR]],
  66     // CHECK: [[A:%.+]] = load i32, ptr [[A_ADDR]],
  67     // CHECK: store i32 [[A]], ptr [[A_PRIV_ADDR]],
  68     // CHECK: br label %[[SCAN_PHASE:[^,]+]]
  69 #pragma omp scan inclusive(a)
  70
  71     // SCAN_PHASE:
  72     // bar();
  73     // goto CONTINUE;
  74     // CHECK: [[SCAN_PHASE]]:
  75     // CHECK: call void @{{.*}}bar{{.*}}()
  76     // CHECK: br label %[[CONTINUE:[^,]+]]
  77     bar();
  78
  79     // CHECK: [[CONTINUE]]:
  80     // CHECK: br label %[[INC_BLOCK:[^,]+]]
  81
  82     // ++OMP_CNT;
  83     // CHECK: [[INC_BLOCK]]:
  84     // CHECK: [[CNT:%.+]] = load i32, ptr [[OMP_CNT]],
  85     // CHECK: [[INC:%.+]] = add nsw i32 [[CNT]], 1
  86     // CHECK: store i32 [[INC]], ptr [[OMP_CNT]],
  87     // CHECK: br label %[[OMP_HEADER]]
  88   }
  89   // CHECK: [[OMP_END]]:
  90 }
  91
  92 struct S {
  93   int a;
  94   S() {}
  95   ~S() {}
  96   S& operator+(const S&);
  97   S& operator=(const S&);
  98 };
  99
 100 // CHECK-LABEL: xyz
 101 void xyz() {
 102   S s[2];
 103
 104   // CHECK: [[S_BEGIN:%.+]] = getelementptr inbounds [2 x %struct.S], ptr [[S_ADDR:%.+]], i{{.+}} 0, i{{.+}} 0
 105   // CHECK: [[S_END:%.+]] = getelementptr {{.*}}%struct.S, ptr [[S_BEGIN]], i{{.+}} 2
 106   // CHECK: br label %[[ARRAY_INIT:.+]]
 107   // CHECK: [[ARRAY_INIT]]:
 108   // CHECK: [[S_CUR:%.+]] = phi ptr [ [[S_BEGIN]], %{{.+}} ], [ [[S_NEXT:%.+]], %[[ARRAY_INIT]] ]
 109   // CHECK: call void [[CONSTR:@.+]](ptr {{[^,]*}} [[S_CUR]])
 110   // CHECK: [[S_NEXT]] = getelementptr inbounds %struct.S, ptr [[S_CUR]], i{{.+}} 1
 111   // CHECK: [[IS_DONE:%.+]] = icmp eq ptr [[S_NEXT]], [[S_END]]
 112   // CHECK: br i1 [[IS_DONE]], label %[[DONE:.+]], label %[[ARRAY_INIT]]
 113   // CHECK: [[DONE]]:
 114   // CHECK: store i32 0, ptr [[OMP_CNT:%.+]],
 115   // CHECK: br label %[[OMP_HEADER:.+]]
 116
 117   // CHECK: [[OMP_HEADER]]:
 118   // CHECK: [[CNT_VAL:%.+]] = load i32, ptr [[OMP_CNT]],
 119   // CHECK: [[CMP:%.+]] = icmp slt i32 [[CNT_VAL]], 10
 120   // CHECK: br i1 [[CMP]], label %[[OMP_BODY:.+]], label %[[OMP_END:.+]]
 121 #pragma omp simd reduction(inscan, + : s)
 122   for (int i = 0; i < 10; ++i) {
 123     // CHECK: [[OMP_BODY]]:
 124
 125     // i = OMP_CNT*1 + 0;
 126     // CHECK: [[CNT_VAL:%.+]] = load i32, ptr [[OMP_CNT]],
 127     // CHECK: [[MUL:%.+]] = mul nsw i32 [[CNT_VAL]], 1
 128     // CHECK: [[ADD:%.+]] = add nsw i32 0, [[MUL]]
 129     // CHECK: store i32 [[ADD]], ptr [[I_ADDR:%.+]],
 130
 131     // S S_PRIV[2];
 132     // CHECK: [[S_BEGIN:%.+]] = getelementptr inbounds [2 x %struct.S], ptr [[S_PRIV_ADDR:%.+]], i{{.+}} 0, i{{.+}} 0
 133     // CHECK: [[S_END:%.+]] = getelementptr {{.*}}%struct.S, ptr [[S_BEGIN]], i{{.+}} 2
 134     // CHECK: [[IS_DONE:%.+]] = icmp eq ptr [[S_BEGIN]], [[S_END]]
 135     // CHECK: br i1 [[IS_DONE]], label %[[DONE:.+]], label %[[ARRAY_INIT:[^,]+]]
 136     // CHECK: [[ARRAY_INIT]]:
 137     // CHECK: [[S_CUR:%.+]] = phi ptr [ [[S_BEGIN]], %[[OMP_BODY]] ], [ [[S_NEXT:%.+]], %[[ARRAY_INIT]] ]
 138     // CHECK: call void [[CONSTR]](ptr {{[^,]*}} [[S_CUR]])
 139     // CHECK: [[S_NEXT]] = getelementptr {{.*}}%struct.S, ptr [[S_CUR]], i{{.+}} 1
 140     // CHECK: [[IS_DONE:%.+]] = icmp eq ptr [[S_NEXT]], [[S_END]]
 141     // CHECK: br i1 [[IS_DONE]], label %[[DONE:.+]], label %[[ARRAY_INIT]]
 142     // CHECK: [[DONE]]:
 143
 144     // goto DISPATCH;
 145     // CHECK: br label %[[DISPATCH:[^,]+]]
 146
 147     // SCAN_PHASE:
 148     // foo();
 149     // goto CONTINUE;
 150     // CHECK: [[SCAN_PHASE:.+]]:
 151     // CHECK: call void @{{.*}}foo{{.*}}()
 152     // CHECK: br label %[[CONTINUE:[^,]+]]
 153     foo();
 154
 155     // DISPATCH:
 156     // goto INPUT_PHASE;
 157     // CHECK: [[DISPATCH]]:
 158     // CHECK: br label %[[INPUT_PHASE:[^,]+]]
 159
 160     // REDUCE:
 161     // TEMP = S;
 162     // S = S_PRIV + S;
 163     // S_PRIV = TEMP;
 164     // goto SCAN_PHASE;
 165     // CHECK: [[REDUCE:.+]]:
 166
 167     // S TEMP[2];
 168     // CHECK: [[TEMP_ARR_BEG:%.+]] = getelementptr inbounds [2 x %struct.S], ptr [[TEMP_ARR:%.+]], i32 0, i32 0
 169     // CHECK: [[TEMP_ARR_END:%.+]] = getelementptr inbounds %struct.S, ptr [[TEMP_ARR_BEG]], i64 2
 170     // CHECK: br label %[[BODY:[^,]+]]
 171     // CHECK: [[BODY]]:
 172     // CHECK: [[CUR:%.+]] = phi ptr [ [[TEMP_ARR_BEG]], %[[REDUCE]] ], [ [[NEXT:%.+]], %[[BODY]] ]
 173     // CHECK: call void [[CONSTR]](ptr {{[^,]*}} [[CUR]])
 174     // CHECK: [[NEXT]] = getelementptr inbounds %struct.S, ptr [[CUR]], i64 1
 175     // CHECK: [[IS_DONE:%.+]] = icmp eq ptr [[NEXT]], [[TEMP_ARR_END]]
 176     // CHECK: br i1 [[IS_DONE]], label %[[EXIT:[^,]+]], label %[[BODY]]
 177     // CHECK: [[EXIT]]:
 178
 179     // TEMP = S;
 180     // CHECK: [[TEMP_ARR_BEG:%.+]] = getelementptr inbounds [2 x %struct.S], ptr [[TEMP_ARR]], i32 0, i32 0
 181     // CHECK: [[TEMP_ARR_END:%.+]] = getelementptr inbounds %struct.S, ptr [[TEMP_ARR_BEG]], i64 2
 182     // CHECK: [[IS_EMPTY:%.+]] = icmp eq ptr [[TEMP_ARR_BEG]], [[TEMP_ARR_END]]
 183     // CHECK: br i1 [[IS_EMPTY]], label %[[EXIT:[^,]+]], label %[[BODY:[^,]+]]
 184     // CHECK: [[BODY]]:
 185     // CHECK: [[CUR_SRC:%.+]] = phi ptr [ [[S_ADDR]], %{{.+}} ], [ [[SRC_NEXT:%.+]], %[[BODY]] ]
 186     // CHECK: [[CUR_DEST:%.+]] = phi ptr [ [[TEMP_ARR_BEG]], %{{.+}} ], [ [[DEST_NEXT:%.+]], %[[BODY]] ]
 187     // CHECK: call {{.*}}ptr [[S_COPY:@.+]](ptr {{[^,]*}} [[CUR_DEST]], ptr {{.*}}[[CUR_SRC]])
 188     // CHECK: [[DEST_NEXT:%.+]] = getelementptr %struct.S, ptr [[CUR_DEST]], i32 1
 189     // CHECK: [[SRC_NEXT:%.+]] = getelementptr %struct.S, ptr [[CUR_SRC]], i32 1
 190     // CHECK: [[IS_DONE:%.+]] = icmp eq ptr [[DEST_NEXT]], [[TEMP_ARR_END]]
 191     // CHECK: br i1 [[IS_DONE]], label %[[EXIT]], label %[[BODY]]
 192     // CHECK: [[EXIT]]:
 193
 194     // S = S_PRIV + S;
 195     // CHECK: [[LHS_END:%.+]] = getelementptr {{.*}}%struct.S, ptr [[S_ADDR]], i{{.+}} 2
 196     // CHECK: [[IS_DONE:%.+]] = icmp eq ptr [[S_ADDR]], [[LHS_END]]
 197     // CHECK: br i1 [[IS_DONE]], label %[[DONE:.+]], label %[[ARRAY_REDUCE_COPY:[^,]+]]
 198     // CHECK: [[ARRAY_REDUCE_COPY]]:
 199     // CHECK: [[SRC_CUR:%.+]] = phi ptr [ [[S_PRIV_ADDR]], %[[EXIT]] ], [ [[SRC_NEXT:%.+]], %[[ARRAY_REDUCE_COPY]] ]
 200     // CHECK: [[DEST_CUR:%.+]] = phi ptr [ [[S_ADDR]], %[[EXIT]] ], [ [[DEST_NEXT:%.+]], %[[ARRAY_REDUCE_COPY]] ]
 201     // CHECK: [[SUM:%.+]] = call {{.*}}ptr @{{.+}}(ptr {{[^,]*}} [[DEST_CUR]], ptr {{.*}}[[SRC_CUR]])
 202     // CHECK: call {{.*}}ptr [[S_COPY]](ptr {{[^,]*}} [[DEST_CUR]], ptr {{.*}}[[SUM]])
 203     // CHECK: [[DEST_NEXT]] = getelementptr {{.*}}%struct.S, ptr [[DEST_CUR]], i{{.+}} 1
 204     // CHECK: [[SRC_NEXT]] = getelementptr {{.*}}%struct.S, ptr [[SRC_CUR]], i{{.+}} 1
 205     // CHECK: [[IS_DONE:%.+]] = icmp eq ptr [[DEST_NEXT]], [[LHS_END]]
 206     // CHECK: br i1 [[IS_DONE]], label %[[DONE:.+]], label %[[ARRAY_REDUCE_COPY]]
 207     // CHECK: [[DONE]]:
 208
 209     // S_PRIV = TEMP;
 210     // CHECK: [[RHS_END:%.+]] = getelementptr inbounds %struct.S, ptr [[S_PRIV_ADDR]], i64 2
 211     // CHECK: [[IS_EMPTY:%.+]] = icmp eq ptr [[S_PRIV_ADDR]], [[RHS_END]]
 212     // CHECK: br i1 [[IS_EMPTY]], label %[[EXIT:[^,]+]], label %[[BODY:[^,]+]]
 213     // CHECK: [[BODY]]:
 214     // CHECK: [[CUR_SRC:%.+]] = phi ptr [ [[TEMP_ARR]], %[[DONE]] ], [ [[SRC_NEXT:%.+]], %[[BODY]] ]
 215     // CHECK: [[CUR_DEST:%.+]] = phi ptr [ [[S_PRIV_ADDR]], %[[DONE]] ], [ [[DEST_NEXT:%.+]], %[[BODY]] ]
 216     // CHECK: call {{.*}}ptr [[S_COPY]](ptr {{[^,]*}} [[CUR_DEST]], ptr {{.*}}[[CUR_SRC]])
 217     // CHECK: [[DEST_NEXT]] = getelementptr %struct.S, ptr [[CUR_DEST]], i32 1
 218     // CHECK: [[SRC_NEXT]] = getelementptr %struct.S, ptr [[CUR_SRC]], i32 1
 219     // CHECK: [[IS_DONE:%.+]] = icmp eq ptr [[DEST_NEXT]], [[RHS_END]]
 220     // CHECK: br i1 [[IS_DONE]], label %[[DONE:[^,]+]], label %[[BODY]]
 221     // CHECK: [[DONE]]:
 222
 223     // TEMP.~S()
 224     // CHECK: [[TEMP_ARR_BEG:%.+]] = getelementptr inbounds [2 x %struct.S], ptr [[TEMP_ARR]], i32 0, i32 0
 225     // CHECK: [[TEMP_ARR_END:%.+]] = getelementptr inbounds %struct.S, ptr [[TEMP_ARR_BEG]], i64 2
 226     // CHECK: br label %[[BODY:[^,]+]]
 227     // CHECK: [[BODY]]:
 228     // CHECK: [[CUR:%.+]] = phi ptr [ [[TEMP_ARR_END]], %[[DONE]] ], [ [[PREV:%.+]], %[[BODY]] ]
 229     // CHECK: [[PREV]] = getelementptr inbounds %struct.S, ptr [[CUR]], i64 -1
 230     // CHECK: call void [[DESTR:@.+]](ptr {{[^,]*}} [[PREV]])
 231     // CHECK: [[IS_DONE:%.+]] = icmp eq ptr [[PREV]], [[TEMP_ARR_BEG]]
 232     // CHECK: br i1 [[IS_DONE]], label %[[EXIT:[^,]+]], label %[[BODY]]
 233     // CHECK: [[EXIT]]:
 234
 235     // goto SCAN_PHASE;
 236     // CHECK: br label %[[SCAN_PHASE]]
 237 #pragma omp scan exclusive(s)
 238
 239     // INPUT_PHASE:
 240     // bar();
 241     // goto REDUCE;
 242     // CHECK: [[INPUT_PHASE]]:
 243     // CHECK: call void @{{.*}}bar{{.*}}()
 244     // CHECK: br label %[[REDUCE]]
 245     bar();
 246
 247     // CHECK: [[CONTINUE]]:
 248
 249     // S_PRIV[2].~S();
 250     // CHECK: [[S_BEGIN:%.+]] = getelementptr inbounds [2 x %struct.S], ptr [[S_PRIV_ADDR]], i{{.+}} 0, i{{.+}} 0
 251     // CHECK: [[S_END:%.+]] = getelementptr {{.*}}%struct.S, ptr [[S_BEGIN]], i{{.+}} 2
 252     // CHECK: br label %[[ARRAY_DESTR:[^,]+]]
 253     // CHECK: [[ARRAY_DESTR]]:
 254     // CHECK: [[S_CUR:%.+]] = phi ptr [ [[S_END]], %[[CONTINUE]] ], [ [[S_PREV:%.+]], %[[ARRAY_DESTR]] ]
 255     // CHECK: [[S_PREV]] = getelementptr {{.*}}%struct.S, ptr [[S_CUR]], i{{.+}} -1
 256     // CHECK: call void [[DESTR]](ptr {{[^,]*}} [[S_PREV]])
 257     // CHECK: [[IS_DONE:%.+]] = icmp eq ptr [[S_PREV]], [[S_BEGIN]]
 258     // CHECK: br i1 [[IS_DONE]], label %[[DONE:.+]], label %[[ARRAY_DESTR]]
 259     // CHECK: [[DONE]]:
 260     // CHECK: br label %[[INC_BLOCK:[^,]+]]
 261
 262     // ++OMP_CNT;
 263     // CHECK: [[INC_BLOCK]]:
 264     // CHECK: [[CNT:%.+]] = load i32, ptr [[OMP_CNT]],
 265     // CHECK: [[INC:%.+]] = add nsw i32 [[CNT]], 1
 266     // CHECK: store i32 [[INC]], ptr [[OMP_CNT]],
 267     // CHECK: br label %[[OMP_HEADER]]
 268   }
 269   // CHECK: [[OMP_END]]:
 270 }
 271
 272 // CHECK-NOT: !{!"llvm.loop.parallel_accesses"
 273
 274 #endif // HEADER