clang/test/OpenMP/parallel_for_simd_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++ -std=c++11 -triple x86_64-unknown-unknown -emit-pch -o %t %s
   3 // RUN: %clang_cc1 -fopenmp -x c++ -triple x86_64-unknown-unknown -std=c++11 -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++ -std=c++11 -triple x86_64-unknown-unknown -emit-pch -o %t %s
   7 // RUN: %clang_cc1 -fopenmp-simd -x c++ -triple x86_64-unknown-unknown -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck --check-prefix SIMD-ONLY0 %s
   8 // SIMD-ONLY0-NOT: {{__kmpc|__tgt}}
   9 // expected-no-diagnostics
  10 #ifndef HEADER
  11 #define HEADER
  12
  13 void foo();
  14 void bar();
  15
  16 // CHECK: define{{.*}} void @{{.*}}baz{{.*}}(i32 noundef %n)
  17 void baz(int n) {
  18   static float a[10];
  19   static double b;
  20
  21   // CHECK: call ptr @llvm.stacksave.p0()
  22   // CHECK: [[A_BUF_SIZE:%.+]] = mul nuw i64 10, [[NUM_ELEMS:%[^,]+]]
  23
  24   // float a_buffer[10][n];
  25   // CHECK: [[A_BUF:%.+]] = alloca float, i64 [[A_BUF_SIZE]],
  26   // CHECK: [[B_BUF:%.+]] = alloca double, i64 10,
  27
  28   // CHECK: call void (ptr, i32, ptr, ...) @__kmpc_fork_call(
  29
  30   // CHECK: [[A_BUF_SIZE:%.+]] = mul nuw i64 10, [[NUM_ELEMS:%[^,]+]]
  31
  32   // float a_buffer[10][n];
  33   // CHECK: [[A_BUF:%.+]] = alloca float, i64 [[A_BUF_SIZE]],
  34
  35   // double b_buffer[10];
  36   // CHECK: [[B_BUF:%.+]] = alloca double, i64 10,
  37   // CHECK: call void (ptr, i32, ptr, ...) @__kmpc_fork_call(
  38   // CHECK: call void @llvm.stackrestore.p0(ptr
  39
  40 #pragma omp parallel for simd reduction(inscan, +:a[:n], b)
  41   for (int i = 0; i < 10; ++i) {
  42     // CHECK: call void @__kmpc_for_static_init_4(
  43     // CHECK: call ptr @llvm.stacksave.p0()
  44     // CHECK: store float 0.000000e+00, ptr %
  45     // CHECK: store double 0.000000e+00, ptr [[B_PRIV_ADDR:%.+]],
  46     // CHECK: br label %[[DISPATCH:[^,]+]]
  47     // CHECK: [[INPUT_PHASE:.+]]:
  48     // CHECK: call void @{{.+}}foo{{.+}}()
  49
  50     // a_buffer[i][0..n] = a_priv[[0..n];
  51     // CHECK: [[BASE_IDX_I:%.+]] = load i32, ptr [[IV_ADDR:%.+]],
  52     // CHECK: [[BASE_IDX:%.+]] = zext i32 [[BASE_IDX_I]] to i64
  53     // CHECK: [[IDX:%.+]] = mul nsw i64 [[BASE_IDX]], [[NUM_ELEMS:%.+]]
  54     // CHECK: [[A_BUF_IDX:%.+]] = getelementptr inbounds float, ptr [[A_BUF:%.+]], i64 [[IDX]]
  55     // CHECK: [[A_PRIV:%.+]] = getelementptr inbounds [10 x float], ptr [[A_PRIV_ADDR:%.+]], i64 0, i64 0
  56     // CHECK: [[BYTES:%.+]] = mul nuw i64 [[NUM_ELEMS:%.+]], 4
  57     // CHECK: call void @llvm.memcpy.p0.p0.i64(ptr {{.*}}[[A_BUF_IDX]], ptr {{.*}}[[A_PRIV]], i64 [[BYTES]], i1 false)
  58
  59     // b_buffer[i] = b_priv;
  60     // CHECK: [[B_BUF_IDX:%.+]] = getelementptr inbounds double, ptr [[B_BUF:%.+]], i64 [[BASE_IDX]]
  61     // CHECK: [[B_PRIV:%.+]] = load double, ptr [[B_PRIV_ADDR]],
  62     // CHECK: store double [[B_PRIV]], ptr [[B_BUF_IDX]],
  63     // CHECK: br label %[[LOOP_CONTINUE:.+]]
  64
  65     // CHECK: [[DISPATCH]]:
  66     // CHECK: br label %[[INPUT_PHASE]]
  67     // CHECK: [[LOOP_CONTINUE]]:
  68     // CHECK: call void @llvm.stackrestore.p0(ptr %
  69     // CHECK: call void @__kmpc_for_static_fini(
  70     // CHECK: call void @__kmpc_barrier(
  71     foo();
  72 #pragma omp scan inclusive(a[:n], b)
  73     // CHECK: [[LOG2_10:%.+]] = call double @llvm.log2.f64(double 1.000000e+01)
  74     // CHECK: [[CEIL_LOG2_10:%.+]] = call double @llvm.ceil.f64(double [[LOG2_10]])
  75     // CHECK: [[CEIL_LOG2_10_INT:%.+]] = fptoui double [[CEIL_LOG2_10]] to i32
  76     // CHECK: br label %[[OUTER_BODY:[^,]+]]
  77     // CHECK: [[OUTER_BODY]]:
  78     // CHECK: [[K:%.+]] = phi i32 [ 0, %{{.+}} ], [ [[K_NEXT:%.+]], %{{.+}} ]
  79     // CHECK: [[K2POW:%.+]] = phi i64 [ 1, %{{.+}} ], [ [[K2POW_NEXT:%.+]], %{{.+}} ]
  80     // CHECK: [[CMP:%.+]] = icmp uge i64 9, [[K2POW]]
  81     // CHECK: br i1 [[CMP]], label %[[INNER_BODY:[^,]+]], label %[[INNER_EXIT:[^,]+]]
  82     // CHECK: [[INNER_BODY]]:
  83     // CHECK: [[I:%.+]] = phi i64 [ 9, %[[OUTER_BODY]] ], [ [[I_PREV:%.+]], %{{.+}} ]
  84
  85     // a_buffer[i] += a_buffer[i-pow(2, k)];
  86     // CHECK: [[IDX:%.+]] = mul nsw i64 [[I]], [[NUM_ELEMS]]
  87     // CHECK: [[A_BUF_IDX:%.+]] = getelementptr inbounds float, ptr [[A_BUF]], i64 [[IDX]]
  88     // CHECK: [[IDX_SUB_K2POW:%.+]] = sub nuw i64 [[I]], [[K2POW]]
  89     // CHECK: [[IDX:%.+]] = mul nsw i64 [[IDX_SUB_K2POW]], [[NUM_ELEMS]]
  90     // CHECK: [[A_BUF_IDX_SUB_K2POW:%.+]] = getelementptr inbounds float, ptr [[A_BUF]], i64 [[IDX]]
  91     // CHECK: [[B_BUF_IDX:%.+]] = getelementptr inbounds double, ptr [[B_BUF]], i64 [[I]]
  92     // CHECK: [[IDX_SUB_K2POW:%.+]] = sub nuw i64 [[I]], [[K2POW]]
  93     // CHECK: [[B_BUF_IDX_SUB_K2POW:%.+]] = getelementptr inbounds double, ptr [[B_BUF]], i64 [[IDX_SUB_K2POW]]
  94     // CHECK: [[A_BUF_END:%.+]] = getelementptr float, ptr [[A_BUF_IDX]], i64 [[NUM_ELEMS]]
  95     // CHECK: [[ISEMPTY:%.+]] = icmp eq ptr [[A_BUF_IDX]], [[A_BUF_END]]
  96     // CHECK: br i1 [[ISEMPTY]], label %[[RED_DONE:[^,]+]], label %[[RED_BODY:[^,]+]]
  97     // CHECK: [[RED_BODY]]:
  98     // CHECK: [[A_BUF_IDX_SUB_K2POW_ELEM:%.+]] = phi ptr [ [[A_BUF_IDX_SUB_K2POW]], %[[INNER_BODY]] ], [ [[A_BUF_IDX_SUB_K2POW_NEXT:%.+]], %[[RED_BODY]] ]
  99     // CHECK: [[A_BUF_IDX_ELEM:%.+]] = phi ptr [ [[A_BUF_IDX]], %[[INNER_BODY]] ], [ [[A_BUF_IDX_NEXT:%.+]], %[[RED_BODY]] ]
 100     // CHECK: [[A_BUF_IDX_VAL:%.+]] = load float, ptr [[A_BUF_IDX_ELEM]],
 101     // CHECK: [[A_BUF_IDX_SUB_K2POW_VAL:%.+]] = load float, ptr [[A_BUF_IDX_SUB_K2POW_ELEM]],
 102     // CHECK: [[RED:%.+]] = fadd float [[A_BUF_IDX_VAL]], [[A_BUF_IDX_SUB_K2POW_VAL]]
 103     // CHECK: store float [[RED]], ptr [[A_BUF_IDX_ELEM]],
 104     // CHECK: [[A_BUF_IDX_NEXT]] = getelementptr float, ptr [[A_BUF_IDX_ELEM]], i32 1
 105     // CHECK: [[A_BUF_IDX_SUB_K2POW_NEXT]] = getelementptr float, ptr [[A_BUF_IDX_SUB_K2POW_ELEM]], i32 1
 106     // CHECK: [[DONE:%.+]] = icmp eq ptr [[A_BUF_IDX_NEXT]], [[A_BUF_END]]
 107     // CHECK: br i1 [[DONE]], label %[[RED_DONE]], label %[[RED_BODY]]
 108     // CHECK: [[RED_DONE]]:
 109
 110     // b_buffer[i] += b_buffer[i-pow(2, k)];
 111     // CHECK: [[B_BUF_IDX_VAL:%.+]] = load double, ptr [[B_BUF_IDX]],
 112     // CHECK: [[B_BUF_IDX_SUB_K2POW_VAL:%.+]] = load double, ptr [[B_BUF_IDX_SUB_K2POW]],
 113     // CHECK: [[RED:%.+]] = fadd double [[B_BUF_IDX_VAL]], [[B_BUF_IDX_SUB_K2POW_VAL]]
 114     // CHECK: store double [[RED]], ptr [[B_BUF_IDX]],
 115
 116     // --i;
 117     // CHECK: [[I_PREV:%.+]] = sub nuw i64 [[I]], 1
 118     // CHECK: [[CMP:%.+]] = icmp uge i64 [[I_PREV]], [[K2POW]]
 119     // CHECK: br i1 [[CMP]], label %[[INNER_BODY]], label %[[INNER_EXIT]]
 120     // CHECK: [[INNER_EXIT]]:
 121
 122     // ++k;
 123     // CHECK: [[K_NEXT]] = add nuw i32 [[K]], 1
 124     // k2pow <<= 1;
 125     // CHECK: [[K2POW_NEXT]] = shl nuw i64 [[K2POW]], 1
 126     // CHECK: [[CMP:%.+]] = icmp ne i32 [[K_NEXT]], [[CEIL_LOG2_10_INT]]
 127     // CHECK: br i1 [[CMP]], label %[[OUTER_BODY]], label %[[OUTER_EXIT:[^,]+]]
 128     // CHECK: [[OUTER_EXIT]]:
 129     bar();
 130     // CHECK: call void @__kmpc_for_static_init_4(
 131     // CHECK: call ptr @llvm.stacksave.p0()
 132     // CHECK: store float 0.000000e+00, ptr %
 133     // CHECK: store double 0.000000e+00, ptr [[B_PRIV_ADDR:%.+]],
 134     // CHECK: br label %[[DISPATCH:[^,]+]]
 135
 136     // Skip the before scan body.
 137     // CHECK: call void @{{.+}}foo{{.+}}()
 138
 139     // CHECK: [[EXIT_INSCAN:[^,]+]]:
 140     // CHECK: br label %[[LOOP_CONTINUE:[^,]+]]
 141
 142     // CHECK: [[DISPATCH]]:
 143     // a_priv[[0..n] = a_buffer[i][0..n];
 144     // CHECK: [[BASE_IDX_I:%.+]] = load i32, ptr [[IV_ADDR:%.+]],
 145     // CHECK: [[BASE_IDX:%.+]] = zext i32 [[BASE_IDX_I]] to i64
 146     // CHECK: [[IDX:%.+]] = mul nsw i64 [[BASE_IDX]], [[NUM_ELEMS]]
 147     // CHECK: [[A_BUF_IDX:%.+]] = getelementptr inbounds float, ptr [[A_BUF]], i64 [[IDX]]
 148     // CHECK: [[A_PRIV:%.+]] = getelementptr inbounds [10 x float], ptr [[A_PRIV_ADDR:%.+]], i64 0, i64 0
 149     // CHECK: [[BYTES:%.+]] = mul nuw i64 [[NUM_ELEMS:%.+]], 4
 150     // CHECK: call void @llvm.memcpy.p0.p0.i64(ptr {{.*}}[[A_PRIV]], ptr {{.*}}[[A_BUF_IDX]], i64 [[BYTES]], i1 false)
 151
 152     // b_priv = b_buffer[i];
 153     // CHECK: [[B_BUF_IDX:%.+]] = getelementptr inbounds double, ptr [[B_BUF]], i64 [[BASE_IDX]]
 154     // CHECK: [[B_BUF_IDX_VAL:%.+]] = load double, ptr [[B_BUF_IDX]],
 155     // CHECK: store double [[B_BUF_IDX_VAL]], ptr [[B_PRIV_ADDR]],
 156     // CHECK: br label %[[SCAN_PHASE:[^,]+]]
 157
 158     // CHECK: [[SCAN_PHASE]]:
 159     // CHECK: call void @{{.+}}bar{{.+}}()
 160     // CHECK: br label %[[EXIT_INSCAN]]
 161
 162     // CHECK: [[LOOP_CONTINUE]]:
 163     // CHECK: call void @llvm.stackrestore.p0(ptr %
 164     // CHECK: call void @__kmpc_for_static_fini(
 165   }
 166
 167 #pragma omp parallel for simd reduction(inscan, +:a[:n], b)
 168   for (int i = 0; i < 10; ++i) {
 169     // CHECK: call void @__kmpc_for_static_init_4(
 170     // CHECK: call ptr @llvm.stacksave.p0()
 171     // CHECK: store float 0.000000e+00, ptr %
 172     // CHECK: store double 0.000000e+00, ptr [[B_PRIV_ADDR:%.+]],
 173     // CHECK: br label %[[DISPATCH:[^,]+]]
 174
 175     // Skip the before scan body.
 176     // CHECK: call void @{{.+}}foo{{.+}}()
 177
 178     // CHECK: [[EXIT_INSCAN:[^,]+]]:
 179
 180     // a_buffer[i][0..n] = a_priv[[0..n];
 181     // CHECK: [[BASE_IDX_I:%.+]] = load i32, ptr [[IV_ADDR:%.+]],
 182     // CHECK: [[BASE_IDX:%.+]] = zext i32 [[BASE_IDX_I]] to i64
 183     // CHECK: [[IDX:%.+]] = mul nsw i64 [[BASE_IDX]], [[NUM_ELEMS:%.+]]
 184     // CHECK: [[A_BUF_IDX:%.+]] = getelementptr inbounds float, ptr [[A_BUF:%.+]], i64 [[IDX]]
 185     // CHECK: [[A_PRIV:%.+]] = getelementptr inbounds [10 x float], ptr [[A_PRIV_ADDR:%.+]], i64 0, i64 0
 186     // CHECK: [[BYTES:%.+]] = mul nuw i64 [[NUM_ELEMS:%.+]], 4
 187     // CHECK: call void @llvm.memcpy.p0.p0.i64(ptr {{.*}}[[A_BUF_IDX]], ptr {{.*}}[[A_PRIV]], i64 [[BYTES]], i1 false)
 188
 189     // b_buffer[i] = b_priv;
 190     // CHECK: [[B_BUF_IDX:%.+]] = getelementptr inbounds double, ptr [[B_BUF:%.+]], i64 [[BASE_IDX]]
 191     // CHECK: [[B_PRIV:%.+]] = load double, ptr [[B_PRIV_ADDR]],
 192     // CHECK: store double [[B_PRIV]], ptr [[B_BUF_IDX]],
 193     // CHECK: br label %[[LOOP_CONTINUE:[^,]+]]
 194
 195     // CHECK: [[DISPATCH]]:
 196     // CHECK: br label %[[INPUT_PHASE:[^,]+]]
 197
 198     // CHECK: [[INPUT_PHASE]]:
 199     // CHECK: call void @{{.+}}bar{{.+}}()
 200     // CHECK: br label %[[EXIT_INSCAN]]
 201
 202     // CHECK: [[LOOP_CONTINUE]]:
 203     // CHECK: call void @llvm.stackrestore.p0(ptr %
 204     // CHECK: call void @__kmpc_for_static_fini(
 205     // CHECK: call void @__kmpc_barrier(
 206     foo();
 207 #pragma omp scan exclusive(a[:n], b)
 208     // CHECK: [[LOG2_10:%.+]] = call double @llvm.log2.f64(double 1.000000e+01)
 209     // CHECK: [[CEIL_LOG2_10:%.+]] = call double @llvm.ceil.f64(double [[LOG2_10]])
 210     // CHECK: [[CEIL_LOG2_10_INT:%.+]] = fptoui double [[CEIL_LOG2_10]] to i32
 211     // CHECK: br label %[[OUTER_BODY:[^,]+]]
 212     // CHECK: [[OUTER_BODY]]:
 213     // CHECK: [[K:%.+]] = phi i32 [ 0, %{{.+}} ], [ [[K_NEXT:%.+]], %{{.+}} ]
 214     // CHECK: [[K2POW:%.+]] = phi i64 [ 1, %{{.+}} ], [ [[K2POW_NEXT:%.+]], %{{.+}} ]
 215     // CHECK: [[CMP:%.+]] = icmp uge i64 9, [[K2POW]]
 216     // CHECK: br i1 [[CMP]], label %[[INNER_BODY:[^,]+]], label %[[INNER_EXIT:[^,]+]]
 217     // CHECK: [[INNER_BODY]]:
 218     // CHECK: [[I:%.+]] = phi i64 [ 9, %[[OUTER_BODY]] ], [ [[I_PREV:%.+]], %{{.+}} ]
 219
 220     // a_buffer[i] += a_buffer[i-pow(2, k)];
 221     // CHECK: [[IDX:%.+]] = mul nsw i64 [[I]], [[NUM_ELEMS]]
 222     // CHECK: [[A_BUF_IDX:%.+]] = getelementptr inbounds float, ptr [[A_BUF]], i64 [[IDX]]
 223     // CHECK: [[IDX_SUB_K2POW:%.+]] = sub nuw i64 [[I]], [[K2POW]]
 224     // CHECK: [[IDX:%.+]] = mul nsw i64 [[IDX_SUB_K2POW]], [[NUM_ELEMS]]
 225     // CHECK: [[A_BUF_IDX_SUB_K2POW:%.+]] = getelementptr inbounds float, ptr [[A_BUF]], i64 [[IDX]]
 226     // CHECK: [[B_BUF_IDX:%.+]] = getelementptr inbounds double, ptr [[B_BUF]], i64 [[I]]
 227     // CHECK: [[IDX_SUB_K2POW:%.+]] = sub nuw i64 [[I]], [[K2POW]]
 228     // CHECK: [[B_BUF_IDX_SUB_K2POW:%.+]] = getelementptr inbounds double, ptr [[B_BUF]], i64 [[IDX_SUB_K2POW]]
 229     // CHECK: [[A_BUF_END:%.+]] = getelementptr float, ptr [[A_BUF_IDX]], i64 [[NUM_ELEMS]]
 230     // CHECK: [[ISEMPTY:%.+]] = icmp eq ptr [[A_BUF_IDX]], [[A_BUF_END]]
 231     // CHECK: br i1 [[ISEMPTY]], label %[[RED_DONE:[^,]+]], label %[[RED_BODY:[^,]+]]
 232     // CHECK: [[RED_BODY]]:
 233     // CHECK: [[A_BUF_IDX_SUB_K2POW_ELEM:%.+]] = phi ptr [ [[A_BUF_IDX_SUB_K2POW]], %[[INNER_BODY]] ], [ [[A_BUF_IDX_SUB_K2POW_NEXT:%.+]], %[[RED_BODY]] ]
 234     // CHECK: [[A_BUF_IDX_ELEM:%.+]] = phi ptr [ [[A_BUF_IDX]], %[[INNER_BODY]] ], [ [[A_BUF_IDX_NEXT:%.+]], %[[RED_BODY]] ]
 235     // CHECK: [[A_BUF_IDX_VAL:%.+]] = load float, ptr [[A_BUF_IDX_ELEM]],
 236     // CHECK: [[A_BUF_IDX_SUB_K2POW_VAL:%.+]] = load float, ptr [[A_BUF_IDX_SUB_K2POW_ELEM]],
 237     // CHECK: [[RED:%.+]] = fadd float [[A_BUF_IDX_VAL]], [[A_BUF_IDX_SUB_K2POW_VAL]]
 238     // CHECK: store float [[RED]], ptr [[A_BUF_IDX_ELEM]],
 239     // CHECK: [[A_BUF_IDX_NEXT]] = getelementptr float, ptr [[A_BUF_IDX_ELEM]], i32 1
 240     // CHECK: [[A_BUF_IDX_SUB_K2POW_NEXT]] = getelementptr float, ptr [[A_BUF_IDX_SUB_K2POW_ELEM]], i32 1
 241     // CHECK: [[DONE:%.+]] = icmp eq ptr [[A_BUF_IDX_NEXT]], [[A_BUF_END]]
 242     // CHECK: br i1 [[DONE]], label %[[RED_DONE]], label %[[RED_BODY]]
 243     // CHECK: [[RED_DONE]]:
 244
 245     // b_buffer[i] += b_buffer[i-pow(2, k)];
 246     // CHECK: [[B_BUF_IDX_VAL:%.+]] = load double, ptr [[B_BUF_IDX]],
 247     // CHECK: [[B_BUF_IDX_SUB_K2POW_VAL:%.+]] = load double, ptr [[B_BUF_IDX_SUB_K2POW]],
 248     // CHECK: [[RED:%.+]] = fadd double [[B_BUF_IDX_VAL]], [[B_BUF_IDX_SUB_K2POW_VAL]]
 249     // CHECK: store double [[RED]], ptr [[B_BUF_IDX]],
 250
 251     // --i;
 252     // CHECK: [[I_PREV:%.+]] = sub nuw i64 [[I]], 1
 253     // CHECK: [[CMP:%.+]] = icmp uge i64 [[I_PREV]], [[K2POW]]
 254     // CHECK: br i1 [[CMP]], label %[[INNER_BODY]], label %[[INNER_EXIT]]
 255     // CHECK: [[INNER_EXIT]]:
 256
 257     // ++k;
 258     // CHECK: [[K_NEXT]] = add nuw i32 [[K]], 1
 259     // k2pow <<= 1;
 260     // CHECK: [[K2POW_NEXT]] = shl nuw i64 [[K2POW]], 1
 261     // CHECK: [[CMP:%.+]] = icmp ne i32 [[K_NEXT]], [[CEIL_LOG2_10_INT]]
 262     // CHECK: br i1 [[CMP]], label %[[OUTER_BODY]], label %[[OUTER_EXIT:[^,]+]]
 263     // CHECK: [[OUTER_EXIT]]:
 264     bar();
 265     // CHECK: call void @__kmpc_for_static_init_4(
 266     // CHECK: call ptr @llvm.stacksave.p0()
 267     // CHECK: store float 0.000000e+00, ptr %
 268     // CHECK: store double 0.000000e+00, ptr [[B_PRIV_ADDR:%.+]],
 269     // CHECK: br label %[[DISPATCH:[^,]+]]
 270
 271     // CHECK: [[SCAN_PHASE:.+]]:
 272     // CHECK: call void @{{.+}}foo{{.+}}()
 273     // CHECK: br label %[[LOOP_CONTINUE:.+]]
 274
 275     // CHECK: [[DISPATCH]]:
 276     // if (i >0)
 277     //   a_priv[[0..n] = a_buffer[i-1][0..n];
 278     // CHECK: [[BASE_IDX_I:%.+]] = load i32, ptr [[IV_ADDR:%.+]],
 279     // CHECK: [[BASE_IDX:%.+]] = zext i32 [[BASE_IDX_I]] to i64
 280     // CHECK: [[CMP:%.+]] = icmp eq i64 [[BASE_IDX]], 0
 281     // CHECK: br i1 [[CMP]], label %[[IF_DONE:[^,]+]], label %[[IF_THEN:[^,]+]]
 282     // CHECK: [[IF_THEN]]:
 283     // CHECK: [[BASE_IDX_SUB_1:%.+]] = sub nuw i64 [[BASE_IDX]], 1
 284     // CHECK: [[IDX:%.+]] = mul nsw i64 [[BASE_IDX_SUB_1]], [[NUM_ELEMS]]
 285     // CHECK: [[A_BUF_IDX:%.+]] = getelementptr inbounds float, ptr [[A_BUF]], i64 [[IDX]]
 286     // CHECK: [[A_PRIV:%.+]] = getelementptr inbounds [10 x float], ptr [[A_PRIV_ADDR:%.+]], i64 0, i64 0
 287     // CHECK: [[BYTES:%.+]] = mul nuw i64 [[NUM_ELEMS:%.+]], 4
 288     // CHECK: call void @llvm.memcpy.p0.p0.i64(ptr {{.*}}[[A_PRIV]], ptr {{.*}}[[A_BUF_IDX]], i64 [[BYTES]], i1 false)
 289
 290     // b_priv = b_buffer[i];
 291     // CHECK: [[B_BUF_IDX:%.+]] = getelementptr inbounds double, ptr [[B_BUF]], i64 [[BASE_IDX_SUB_1]]
 292     // CHECK: [[B_BUF_IDX_VAL:%.+]] = load double, ptr [[B_BUF_IDX]],
 293     // CHECK: store double [[B_BUF_IDX_VAL]], ptr [[B_PRIV_ADDR]],
 294     // CHECK: br label %[[SCAN_PHASE]]
 295
 296     // CHECK: [[LOOP_CONTINUE]]:
 297     // CHECK: call void @llvm.stackrestore.p0(ptr %
 298     // CHECK: call void @__kmpc_for_static_fini(
 299   }
 300 }
 301
 302 // CHECK: !{!"llvm.loop.vectorize.enable", i1 true}
 303
 304 #endif
 305