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