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 %n)
baz(int n)17 void baz(int n) {
18   static float a[10];
19   static double b;
20 
21   // CHECK: call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(
22   // CHECK: call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(
23 
24   // CHECK: call i8* @llvm.stacksave()
25   // CHECK: [[A_BUF_SIZE:%.+]] = mul nuw i64 10, [[NUM_ELEMS:%[^,]+]]
26 
27   // float a_buffer[10][n];
28   // CHECK: [[A_BUF:%.+]] = alloca float, i64 [[A_BUF_SIZE]],
29 
30   // double b_buffer[10];
31   // CHECK: [[B_BUF:%.+]] = alloca double, i64 10,
32 #pragma omp parallel for simd reduction(inscan, +:a[:n], b)
33   for (int i = 0; i < 10; ++i) {
34     // CHECK: call void @__kmpc_for_static_init_4(
35     // CHECK: call i8* @llvm.stacksave()
36     // CHECK: store float 0.000000e+00, float* %
37     // CHECK: store double 0.000000e+00, double* [[B_PRIV_ADDR:%.+]],
38     // CHECK: br label %[[DISPATCH:[^,]+]]
39     // CHECK: [[INPUT_PHASE:.+]]:
40     // CHECK: call void @{{.+}}foo{{.+}}()
41 
42     // a_buffer[i][0..n] = a_priv[[0..n];
43     // CHECK: [[BASE_IDX_I:%.+]] = load i32, i32* [[IV_ADDR:%.+]],
44     // CHECK: [[BASE_IDX:%.+]] = zext i32 [[BASE_IDX_I]] to i64
45     // CHECK: [[IDX:%.+]] = mul nsw i64 [[BASE_IDX]], [[NUM_ELEMS]]
46     // CHECK: [[A_BUF_IDX:%.+]] = getelementptr inbounds float, float* [[A_BUF]], i64 [[IDX]]
47     // CHECK: [[A_PRIV:%.+]] = getelementptr inbounds [10 x float], [10 x float]* [[A_PRIV_ADDR:%.+]], i64 0, i64 0
48     // CHECK: [[BYTES:%.+]] = mul nuw i64 [[NUM_ELEMS:%.+]], 4
49     // CHECK: [[DEST:%.+]] = bitcast float* [[A_BUF_IDX]] to i8*
50     // CHECK: [[SRC:%.+]] = bitcast float* [[A_PRIV]] to i8*
51     // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* {{.*}}[[DEST]], i8* {{.*}}[[SRC]], i64 [[BYTES]], i1 false)
52 
53     // b_buffer[i] = b_priv;
54     // CHECK: [[B_BUF_IDX:%.+]] = getelementptr inbounds double, double* [[B_BUF]], i64 [[BASE_IDX]]
55     // CHECK: [[B_PRIV:%.+]] = load double, double* [[B_PRIV_ADDR]],
56     // CHECK: store double [[B_PRIV]], double* [[B_BUF_IDX]],
57     // CHECK: br label %[[LOOP_CONTINUE:.+]]
58 
59     // CHECK: [[DISPATCH]]:
60     // CHECK: br label %[[INPUT_PHASE]]
61     // CHECK: [[LOOP_CONTINUE]]:
62     // CHECK: call void @llvm.stackrestore(i8* %
63     // CHECK: call void @__kmpc_for_static_fini(
64     // CHECK: call void @__kmpc_barrier(
65     foo();
66 #pragma omp scan inclusive(a[:n], b)
67     // CHECK: [[LOG2_10:%.+]] = call double @llvm.log2.f64(double 1.000000e+01)
68     // CHECK: [[CEIL_LOG2_10:%.+]] = call double @llvm.ceil.f64(double [[LOG2_10]])
69     // CHECK: [[CEIL_LOG2_10_INT:%.+]] = fptoui double [[CEIL_LOG2_10]] to i32
70     // CHECK: br label %[[OUTER_BODY:[^,]+]]
71     // CHECK: [[OUTER_BODY]]:
72     // CHECK: [[K:%.+]] = phi i32 [ 0, %{{.+}} ], [ [[K_NEXT:%.+]], %{{.+}} ]
73     // CHECK: [[K2POW:%.+]] = phi i64 [ 1, %{{.+}} ], [ [[K2POW_NEXT:%.+]], %{{.+}} ]
74     // CHECK: [[CMP:%.+]] = icmp uge i64 9, [[K2POW]]
75     // CHECK: br i1 [[CMP]], label %[[INNER_BODY:[^,]+]], label %[[INNER_EXIT:[^,]+]]
76     // CHECK: [[INNER_BODY]]:
77     // CHECK: [[I:%.+]] = phi i64 [ 9, %[[OUTER_BODY]] ], [ [[I_PREV:%.+]], %{{.+}} ]
78 
79     // a_buffer[i] += a_buffer[i-pow(2, k)];
80     // CHECK: [[IDX:%.+]] = mul nsw i64 [[I]], [[NUM_ELEMS]]
81     // CHECK: [[A_BUF_IDX:%.+]] = getelementptr inbounds float, float* [[A_BUF]], i64 [[IDX]]
82     // CHECK: [[IDX_SUB_K2POW:%.+]] = sub nuw i64 [[I]], [[K2POW]]
83     // CHECK: [[IDX:%.+]] = mul nsw i64 [[IDX_SUB_K2POW]], [[NUM_ELEMS]]
84     // CHECK: [[A_BUF_IDX_SUB_K2POW:%.+]] = getelementptr inbounds float, float* [[A_BUF]], i64 [[IDX]]
85     // CHECK: [[B_BUF_IDX:%.+]] = getelementptr inbounds double, double* [[B_BUF]], i64 [[I]]
86     // CHECK: [[IDX_SUB_K2POW:%.+]] = sub nuw i64 [[I]], [[K2POW]]
87     // CHECK: [[B_BUF_IDX_SUB_K2POW:%.+]] = getelementptr inbounds double, double* [[B_BUF]], i64 [[IDX_SUB_K2POW]]
88     // CHECK: [[A_BUF_END:%.+]] = getelementptr float, float* [[A_BUF_IDX]], i64 [[NUM_ELEMS]]
89     // CHECK: [[ISEMPTY:%.+]] = icmp eq float* [[A_BUF_IDX]], [[A_BUF_END]]
90     // CHECK: br i1 [[ISEMPTY]], label %[[RED_DONE:[^,]+]], label %[[RED_BODY:[^,]+]]
91     // CHECK: [[RED_BODY]]:
92     // CHECK: [[A_BUF_IDX_SUB_K2POW_ELEM:%.+]] = phi float* [ [[A_BUF_IDX_SUB_K2POW]], %[[INNER_BODY]] ], [ [[A_BUF_IDX_SUB_K2POW_NEXT:%.+]], %[[RED_BODY]] ]
93     // CHECK: [[A_BUF_IDX_ELEM:%.+]] = phi float* [ [[A_BUF_IDX]], %[[INNER_BODY]] ], [ [[A_BUF_IDX_NEXT:%.+]], %[[RED_BODY]] ]
94     // CHECK: [[A_BUF_IDX_VAL:%.+]] = load float, float* [[A_BUF_IDX_ELEM]],
95     // CHECK: [[A_BUF_IDX_SUB_K2POW_VAL:%.+]] = load float, float* [[A_BUF_IDX_SUB_K2POW_ELEM]],
96     // CHECK: [[RED:%.+]] = fadd float [[A_BUF_IDX_VAL]], [[A_BUF_IDX_SUB_K2POW_VAL]]
97     // CHECK: store float [[RED]], float* [[A_BUF_IDX_ELEM]],
98     // CHECK: [[A_BUF_IDX_NEXT]] = getelementptr float, float* [[A_BUF_IDX_ELEM]], i32 1
99     // CHECK: [[A_BUF_IDX_SUB_K2POW_NEXT]] = getelementptr float, float* [[A_BUF_IDX_SUB_K2POW_ELEM]], i32 1
100     // CHECK: [[DONE:%.+]] = icmp eq float* [[A_BUF_IDX_NEXT]], [[A_BUF_END]]
101     // CHECK: br i1 [[DONE]], label %[[RED_DONE]], label %[[RED_BODY]]
102     // CHECK: [[RED_DONE]]:
103 
104     // b_buffer[i] += b_buffer[i-pow(2, k)];
105     // CHECK: [[B_BUF_IDX_VAL:%.+]] = load double, double* [[B_BUF_IDX]],
106     // CHECK: [[B_BUF_IDX_SUB_K2POW_VAL:%.+]] = load double, double* [[B_BUF_IDX_SUB_K2POW]],
107     // CHECK: [[RED:%.+]] = fadd double [[B_BUF_IDX_VAL]], [[B_BUF_IDX_SUB_K2POW_VAL]]
108     // CHECK: store double [[RED]], double* [[B_BUF_IDX]],
109 
110     // --i;
111     // CHECK: [[I_PREV:%.+]] = sub nuw i64 [[I]], 1
112     // CHECK: [[CMP:%.+]] = icmp uge i64 [[I_PREV]], [[K2POW]]
113     // CHECK: br i1 [[CMP]], label %[[INNER_BODY]], label %[[INNER_EXIT]]
114     // CHECK: [[INNER_EXIT]]:
115 
116     // ++k;
117     // CHECK: [[K_NEXT]] = add nuw i32 [[K]], 1
118     // k2pow <<= 1;
119     // CHECK: [[K2POW_NEXT]] = shl nuw i64 [[K2POW]], 1
120     // CHECK: [[CMP:%.+]] = icmp ne i32 [[K_NEXT]], [[CEIL_LOG2_10_INT]]
121     // CHECK: br i1 [[CMP]], label %[[OUTER_BODY]], label %[[OUTER_EXIT:[^,]+]]
122     // CHECK: [[OUTER_EXIT]]:
123     bar();
124     // CHECK: call void @__kmpc_for_static_init_4(
125     // CHECK: call i8* @llvm.stacksave()
126     // CHECK: store float 0.000000e+00, float* %
127     // CHECK: store double 0.000000e+00, double* [[B_PRIV_ADDR:%.+]],
128     // CHECK: br label %[[DISPATCH:[^,]+]]
129 
130     // Skip the before scan body.
131     // CHECK: call void @{{.+}}foo{{.+}}()
132 
133     // CHECK: [[EXIT_INSCAN:[^,]+]]:
134     // CHECK: br label %[[LOOP_CONTINUE:[^,]+]]
135 
136     // CHECK: [[DISPATCH]]:
137     // a_priv[[0..n] = a_buffer[i][0..n];
138     // CHECK: [[BASE_IDX_I:%.+]] = load i32, i32* [[IV_ADDR:%.+]],
139     // CHECK: [[BASE_IDX:%.+]] = zext i32 [[BASE_IDX_I]] to i64
140     // CHECK: [[IDX:%.+]] = mul nsw i64 [[BASE_IDX]], [[NUM_ELEMS]]
141     // CHECK: [[A_BUF_IDX:%.+]] = getelementptr inbounds float, float* [[A_BUF]], i64 [[IDX]]
142     // CHECK: [[A_PRIV:%.+]] = getelementptr inbounds [10 x float], [10 x float]* [[A_PRIV_ADDR:%.+]], i64 0, i64 0
143     // CHECK: [[BYTES:%.+]] = mul nuw i64 [[NUM_ELEMS:%.+]], 4
144     // CHECK: [[DEST:%.+]] = bitcast float* [[A_PRIV]] to i8*
145     // CHECK: [[SRC:%.+]] = bitcast float* [[A_BUF_IDX]] to i8*
146     // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* {{.*}}[[DEST]], i8* {{.*}}[[SRC]], i64 [[BYTES]], i1 false)
147 
148     // b_priv = b_buffer[i];
149     // CHECK: [[B_BUF_IDX:%.+]] = getelementptr inbounds double, double* [[B_BUF]], i64 [[BASE_IDX]]
150     // CHECK: [[B_BUF_IDX_VAL:%.+]] = load double, double* [[B_BUF_IDX]],
151     // CHECK: store double [[B_BUF_IDX_VAL]], double* [[B_PRIV_ADDR]],
152     // CHECK: br label %[[SCAN_PHASE:[^,]+]]
153 
154     // CHECK: [[SCAN_PHASE]]:
155     // CHECK: call void @{{.+}}bar{{.+}}()
156     // CHECK: br label %[[EXIT_INSCAN]]
157 
158     // CHECK: [[LOOP_CONTINUE]]:
159     // CHECK: call void @llvm.stackrestore(i8* %
160     // CHECK: call void @__kmpc_for_static_fini(
161     // CHECK: call void @llvm.stackrestore(i8*
162   }
163 
164   // CHECK: call i8* @llvm.stacksave()
165   // CHECK: [[A_BUF_SIZE:%.+]] = mul nuw i64 10, [[NUM_ELEMS:%[^,]+]]
166 
167   // float a_buffer[10][n];
168   // CHECK: [[A_BUF:%.+]] = alloca float, i64 [[A_BUF_SIZE]],
169 
170   // double b_buffer[10];
171   // CHECK: [[B_BUF:%.+]] = alloca double, i64 10,
172 #pragma omp parallel for simd reduction(inscan, +:a[:n], b)
173   for (int i = 0; i < 10; ++i) {
174     // CHECK: call void @__kmpc_for_static_init_4(
175     // CHECK: call i8* @llvm.stacksave()
176     // CHECK: store float 0.000000e+00, float* %
177     // CHECK: store double 0.000000e+00, double* [[B_PRIV_ADDR:%.+]],
178     // CHECK: br label %[[DISPATCH:[^,]+]]
179 
180     // Skip the before scan body.
181     // CHECK: call void @{{.+}}foo{{.+}}()
182 
183     // CHECK: [[EXIT_INSCAN:[^,]+]]:
184 
185     // a_buffer[i][0..n] = a_priv[[0..n];
186     // CHECK: [[BASE_IDX_I:%.+]] = load i32, i32* [[IV_ADDR:%.+]],
187     // CHECK: [[BASE_IDX:%.+]] = zext i32 [[BASE_IDX_I]] to i64
188     // CHECK: [[IDX:%.+]] = mul nsw i64 [[BASE_IDX]], [[NUM_ELEMS]]
189     // CHECK: [[A_BUF_IDX:%.+]] = getelementptr inbounds float, float* [[A_BUF]], i64 [[IDX]]
190     // CHECK: [[A_PRIV:%.+]] = getelementptr inbounds [10 x float], [10 x float]* [[A_PRIV_ADDR:%.+]], i64 0, i64 0
191     // CHECK: [[BYTES:%.+]] = mul nuw i64 [[NUM_ELEMS:%.+]], 4
192     // CHECK: [[DEST:%.+]] = bitcast float* [[A_BUF_IDX]] to i8*
193     // CHECK: [[SRC:%.+]] = bitcast float* [[A_PRIV]] to i8*
194     // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* {{.*}}[[DEST]], i8* {{.*}}[[SRC]], i64 [[BYTES]], i1 false)
195 
196     // b_buffer[i] = b_priv;
197     // CHECK: [[B_BUF_IDX:%.+]] = getelementptr inbounds double, double* [[B_BUF]], i64 [[BASE_IDX]]
198     // CHECK: [[B_PRIV:%.+]] = load double, double* [[B_PRIV_ADDR]],
199     // CHECK: store double [[B_PRIV]], double* [[B_BUF_IDX]],
200     // CHECK: br label %[[LOOP_CONTINUE:[^,]+]]
201 
202     // CHECK: [[DISPATCH]]:
203     // CHECK: br label %[[INPUT_PHASE:[^,]+]]
204 
205     // CHECK: [[INPUT_PHASE]]:
206     // CHECK: call void @{{.+}}bar{{.+}}()
207     // CHECK: br label %[[EXIT_INSCAN]]
208 
209     // CHECK: [[LOOP_CONTINUE]]:
210     // CHECK: call void @llvm.stackrestore(i8* %
211     // CHECK: call void @__kmpc_for_static_fini(
212     // CHECK: call void @__kmpc_barrier(
213     foo();
214 #pragma omp scan exclusive(a[:n], b)
215     // CHECK: [[LOG2_10:%.+]] = call double @llvm.log2.f64(double 1.000000e+01)
216     // CHECK: [[CEIL_LOG2_10:%.+]] = call double @llvm.ceil.f64(double [[LOG2_10]])
217     // CHECK: [[CEIL_LOG2_10_INT:%.+]] = fptoui double [[CEIL_LOG2_10]] to i32
218     // CHECK: br label %[[OUTER_BODY:[^,]+]]
219     // CHECK: [[OUTER_BODY]]:
220     // CHECK: [[K:%.+]] = phi i32 [ 0, %{{.+}} ], [ [[K_NEXT:%.+]], %{{.+}} ]
221     // CHECK: [[K2POW:%.+]] = phi i64 [ 1, %{{.+}} ], [ [[K2POW_NEXT:%.+]], %{{.+}} ]
222     // CHECK: [[CMP:%.+]] = icmp uge i64 9, [[K2POW]]
223     // CHECK: br i1 [[CMP]], label %[[INNER_BODY:[^,]+]], label %[[INNER_EXIT:[^,]+]]
224     // CHECK: [[INNER_BODY]]:
225     // CHECK: [[I:%.+]] = phi i64 [ 9, %[[OUTER_BODY]] ], [ [[I_PREV:%.+]], %{{.+}} ]
226 
227     // a_buffer[i] += a_buffer[i-pow(2, k)];
228     // CHECK: [[IDX:%.+]] = mul nsw i64 [[I]], [[NUM_ELEMS]]
229     // CHECK: [[A_BUF_IDX:%.+]] = getelementptr inbounds float, float* [[A_BUF]], i64 [[IDX]]
230     // CHECK: [[IDX_SUB_K2POW:%.+]] = sub nuw i64 [[I]], [[K2POW]]
231     // CHECK: [[IDX:%.+]] = mul nsw i64 [[IDX_SUB_K2POW]], [[NUM_ELEMS]]
232     // CHECK: [[A_BUF_IDX_SUB_K2POW:%.+]] = getelementptr inbounds float, float* [[A_BUF]], i64 [[IDX]]
233     // CHECK: [[B_BUF_IDX:%.+]] = getelementptr inbounds double, double* [[B_BUF]], i64 [[I]]
234     // CHECK: [[IDX_SUB_K2POW:%.+]] = sub nuw i64 [[I]], [[K2POW]]
235     // CHECK: [[B_BUF_IDX_SUB_K2POW:%.+]] = getelementptr inbounds double, double* [[B_BUF]], i64 [[IDX_SUB_K2POW]]
236     // CHECK: [[A_BUF_END:%.+]] = getelementptr float, float* [[A_BUF_IDX]], i64 [[NUM_ELEMS]]
237     // CHECK: [[ISEMPTY:%.+]] = icmp eq float* [[A_BUF_IDX]], [[A_BUF_END]]
238     // CHECK: br i1 [[ISEMPTY]], label %[[RED_DONE:[^,]+]], label %[[RED_BODY:[^,]+]]
239     // CHECK: [[RED_BODY]]:
240     // CHECK: [[A_BUF_IDX_SUB_K2POW_ELEM:%.+]] = phi float* [ [[A_BUF_IDX_SUB_K2POW]], %[[INNER_BODY]] ], [ [[A_BUF_IDX_SUB_K2POW_NEXT:%.+]], %[[RED_BODY]] ]
241     // CHECK: [[A_BUF_IDX_ELEM:%.+]] = phi float* [ [[A_BUF_IDX]], %[[INNER_BODY]] ], [ [[A_BUF_IDX_NEXT:%.+]], %[[RED_BODY]] ]
242     // CHECK: [[A_BUF_IDX_VAL:%.+]] = load float, float* [[A_BUF_IDX_ELEM]],
243     // CHECK: [[A_BUF_IDX_SUB_K2POW_VAL:%.+]] = load float, float* [[A_BUF_IDX_SUB_K2POW_ELEM]],
244     // CHECK: [[RED:%.+]] = fadd float [[A_BUF_IDX_VAL]], [[A_BUF_IDX_SUB_K2POW_VAL]]
245     // CHECK: store float [[RED]], float* [[A_BUF_IDX_ELEM]],
246     // CHECK: [[A_BUF_IDX_NEXT]] = getelementptr float, float* [[A_BUF_IDX_ELEM]], i32 1
247     // CHECK: [[A_BUF_IDX_SUB_K2POW_NEXT]] = getelementptr float, float* [[A_BUF_IDX_SUB_K2POW_ELEM]], i32 1
248     // CHECK: [[DONE:%.+]] = icmp eq float* [[A_BUF_IDX_NEXT]], [[A_BUF_END]]
249     // CHECK: br i1 [[DONE]], label %[[RED_DONE]], label %[[RED_BODY]]
250     // CHECK: [[RED_DONE]]:
251 
252     // b_buffer[i] += b_buffer[i-pow(2, k)];
253     // CHECK: [[B_BUF_IDX_VAL:%.+]] = load double, double* [[B_BUF_IDX]],
254     // CHECK: [[B_BUF_IDX_SUB_K2POW_VAL:%.+]] = load double, double* [[B_BUF_IDX_SUB_K2POW]],
255     // CHECK: [[RED:%.+]] = fadd double [[B_BUF_IDX_VAL]], [[B_BUF_IDX_SUB_K2POW_VAL]]
256     // CHECK: store double [[RED]], double* [[B_BUF_IDX]],
257 
258     // --i;
259     // CHECK: [[I_PREV:%.+]] = sub nuw i64 [[I]], 1
260     // CHECK: [[CMP:%.+]] = icmp uge i64 [[I_PREV]], [[K2POW]]
261     // CHECK: br i1 [[CMP]], label %[[INNER_BODY]], label %[[INNER_EXIT]]
262     // CHECK: [[INNER_EXIT]]:
263 
264     // ++k;
265     // CHECK: [[K_NEXT]] = add nuw i32 [[K]], 1
266     // k2pow <<= 1;
267     // CHECK: [[K2POW_NEXT]] = shl nuw i64 [[K2POW]], 1
268     // CHECK: [[CMP:%.+]] = icmp ne i32 [[K_NEXT]], [[CEIL_LOG2_10_INT]]
269     // CHECK: br i1 [[CMP]], label %[[OUTER_BODY]], label %[[OUTER_EXIT:[^,]+]]
270     // CHECK: [[OUTER_EXIT]]:
271     bar();
272     // CHECK: call void @__kmpc_for_static_init_4(
273     // CHECK: call i8* @llvm.stacksave()
274     // CHECK: store float 0.000000e+00, float* %
275     // CHECK: store double 0.000000e+00, double* [[B_PRIV_ADDR:%.+]],
276     // CHECK: br label %[[DISPATCH:[^,]+]]
277 
278     // CHECK: [[SCAN_PHASE:.+]]:
279     // CHECK: call void @{{.+}}foo{{.+}}()
280     // CHECK: br label %[[LOOP_CONTINUE:.+]]
281 
282     // CHECK: [[DISPATCH]]:
283     // if (i >0)
284     //   a_priv[[0..n] = a_buffer[i-1][0..n];
285     // CHECK: [[BASE_IDX_I:%.+]] = load i32, i32* [[IV_ADDR:%.+]],
286     // CHECK: [[BASE_IDX:%.+]] = zext i32 [[BASE_IDX_I]] to i64
287     // CHECK: [[CMP:%.+]] = icmp eq i64 [[BASE_IDX]], 0
288     // CHECK: br i1 [[CMP]], label %[[IF_DONE:[^,]+]], label %[[IF_THEN:[^,]+]]
289     // CHECK: [[IF_THEN]]:
290     // CHECK: [[BASE_IDX_SUB_1:%.+]] = sub nuw i64 [[BASE_IDX]], 1
291     // CHECK: [[IDX:%.+]] = mul nsw i64 [[BASE_IDX_SUB_1]], [[NUM_ELEMS]]
292     // CHECK: [[A_BUF_IDX:%.+]] = getelementptr inbounds float, float* [[A_BUF]], i64 [[IDX]]
293     // CHECK: [[A_PRIV:%.+]] = getelementptr inbounds [10 x float], [10 x float]* [[A_PRIV_ADDR:%.+]], i64 0, i64 0
294     // CHECK: [[BYTES:%.+]] = mul nuw i64 [[NUM_ELEMS:%.+]], 4
295     // CHECK: [[DEST:%.+]] = bitcast float* [[A_PRIV]] to i8*
296     // CHECK: [[SRC:%.+]] = bitcast float* [[A_BUF_IDX]] to i8*
297     // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* {{.*}}[[DEST]], i8* {{.*}}[[SRC]], i64 [[BYTES]], i1 false)
298 
299     // b_priv = b_buffer[i];
300     // CHECK: [[B_BUF_IDX:%.+]] = getelementptr inbounds double, double* [[B_BUF]], i64 [[BASE_IDX_SUB_1]]
301     // CHECK: [[B_BUF_IDX_VAL:%.+]] = load double, double* [[B_BUF_IDX]],
302     // CHECK: store double [[B_BUF_IDX_VAL]], double* [[B_PRIV_ADDR]],
303     // CHECK: br label %[[SCAN_PHASE]]
304 
305     // CHECK: [[LOOP_CONTINUE]]:
306     // CHECK: call void @llvm.stackrestore(i8* %
307     // CHECK: call void @__kmpc_for_static_fini(
308     // CHECK: call void @llvm.stackrestore(i8*
309   }
310 }
311 
312 // CHECK: !{!"llvm.loop.vectorize.enable", i1 true}
313 
314 #endif
315 
316