| blob | ed6de3ea482f3b3ec9f6f12fdaa890d07ec498c8 |
1 ; RUN: llc -verify-machineinstrs -ppc-asm-full-reg-names -ppc-vsr-nums-as-vr \
2 ; RUN: -mcpu=pwr8 -mtriple=powerpcle < %s | FileCheck %s
3 ; RUN: llc -verify-machineinstrs -ppc-asm-full-reg-names -ppc-vsr-nums-as-vr \
4 ; RUN: -mcpu=pwr8 -mtriple=powerpc < %s | FileCheck %s
5 %struct.d = type { [131072 x i32] }
7 @a = dso_local local_unnamed_addr global [4096 x i32] zeroinitializer, align 4
9 ; Function Attrs: mustprogress uwtable
10 define dso_local void @_Z1g1dILi17EE(ptr nocapture noundef readnone byval(%struct.d) align 4 %0) local_unnamed_addr #0 {
11 ; CHECK-LABEL: _Z1g1dILi17EE:
12 ; CHECK-NOT: mtfprwz f0, r4
13 ; CHECK-NOT: stwx r3, r1, r4
14 ; CHECK-NOT: mffprwz r4, f0
15 ; CHECK-NOT: mtfprwz f0, r4
16 ; CHECK-NOT: lwzx r3, r1, r4
17 ; CHECK-NOT: mffprwz r4, f0
18 entry:
19 %c = alloca %struct.d, align 4
20 call void @llvm.lifetime.start.p0(i64 524288, ptr nonnull %c) #3
21 br label %vector.body
23 vector.body: ; preds = %vector.body.1, %entry
24 %index = phi i32 [ 0, %entry ], [ %index.next.1, %vector.body.1 ]
25 %vec.ind = phi <4 x i32> [ <i32 0, i32 1, i32 2, i32 3>, %entry ], [ %vec.ind.next.1, %vector.body.1 ]
26 %step.add = add <4 x i32> %vec.ind, <i32 4, i32 4, i32 4, i32 4>
27 %step.add21 = add <4 x i32> %vec.ind, <i32 8, i32 8, i32 8, i32 8>
28 %step.add22 = add <4 x i32> %vec.ind, <i32 12, i32 12, i32 12, i32 12>
29 %step.add23 = add <4 x i32> %vec.ind, <i32 16, i32 16, i32 16, i32 16>
30 %step.add24 = add <4 x i32> %vec.ind, <i32 20, i32 20, i32 20, i32 20>
31 %step.add25 = add <4 x i32> %vec.ind, <i32 24, i32 24, i32 24, i32 24>
32 %step.add26 = add <4 x i32> %vec.ind, <i32 28, i32 28, i32 28, i32 28>
33 %step.add27 = add <4 x i32> %vec.ind, <i32 32, i32 32, i32 32, i32 32>
34 %step.add28 = add <4 x i32> %vec.ind, <i32 36, i32 36, i32 36, i32 36>
35 %step.add29 = add <4 x i32> %vec.ind, <i32 40, i32 40, i32 40, i32 40>
36 %step.add30 = add <4 x i32> %vec.ind, <i32 44, i32 44, i32 44, i32 44>
37 %1 = getelementptr inbounds [4096 x i32], ptr @a, i32 0, i32 %index
38 store <4 x i32> %vec.ind, ptr %1, align 4
39 %2 = getelementptr inbounds i32, ptr %1, i32 4
40 store <4 x i32> %step.add, ptr %2, align 4
41 %3 = getelementptr inbounds i32, ptr %1, i32 8
42 store <4 x i32> %step.add21, ptr %3, align 4
43 %4 = getelementptr inbounds i32, ptr %1, i32 12
44 store <4 x i32> %step.add22, ptr %4, align 4
45 %5 = getelementptr inbounds i32, ptr %1, i32 16
46 store <4 x i32> %step.add23, ptr %5, align 4
47 %6 = getelementptr inbounds i32, ptr %1, i32 20
48 store <4 x i32> %step.add24, ptr %6, align 4
49 %7 = getelementptr inbounds i32, ptr %1, i32 24
50 store <4 x i32> %step.add25, ptr %7, align 4
51 %8 = getelementptr inbounds i32, ptr %1, i32 28
52 store <4 x i32> %step.add26, ptr %8, align 4
53 %9 = getelementptr inbounds i32, ptr %1, i32 32
54 store <4 x i32> %step.add27, ptr %9, align 4
55 %10 = getelementptr inbounds i32, ptr %1, i32 36
56 store <4 x i32> %step.add28, ptr %10, align 4
57 %11 = getelementptr inbounds i32, ptr %1, i32 40
58 store <4 x i32> %step.add29, ptr %11, align 4
59 %12 = getelementptr inbounds i32, ptr %1, i32 44
60 store <4 x i32> %step.add30, ptr %12, align 4
61 %index.next = add nuw nsw i32 %index, 48
62 %13 = icmp eq i32 %index.next, 4080
63 br i1 %13, label %for.body, label %vector.body.1
65 vector.body.1: ; preds = %vector.body
66 %vec.ind.next = add <4 x i32> %vec.ind, <i32 48, i32 48, i32 48, i32 48>
67 %step.add.1 = add <4 x i32> %vec.ind, <i32 52, i32 52, i32 52, i32 52>
68 %step.add21.1 = add <4 x i32> %vec.ind, <i32 56, i32 56, i32 56, i32 56>
69 %step.add22.1 = add <4 x i32> %vec.ind, <i32 60, i32 60, i32 60, i32 60>
70 %step.add23.1 = add <4 x i32> %vec.ind, <i32 64, i32 64, i32 64, i32 64>
71 %step.add24.1 = add <4 x i32> %vec.ind, <i32 68, i32 68, i32 68, i32 68>
72 %step.add25.1 = add <4 x i32> %vec.ind, <i32 72, i32 72, i32 72, i32 72>
73 %step.add26.1 = add <4 x i32> %vec.ind, <i32 76, i32 76, i32 76, i32 76>
74 %step.add27.1 = add <4 x i32> %vec.ind, <i32 80, i32 80, i32 80, i32 80>
75 %step.add28.1 = add <4 x i32> %vec.ind, <i32 84, i32 84, i32 84, i32 84>
76 %step.add29.1 = add <4 x i32> %vec.ind, <i32 88, i32 88, i32 88, i32 88>
77 %step.add30.1 = add <4 x i32> %vec.ind, <i32 92, i32 92, i32 92, i32 92>
78 %14 = getelementptr inbounds [4096 x i32], ptr @a, i32 0, i32 %index.next
79 store <4 x i32> %vec.ind.next, ptr %14, align 4
80 %15 = getelementptr inbounds i32, ptr %14, i32 4
81 store <4 x i32> %step.add.1, ptr %15, align 4
82 %16 = getelementptr inbounds i32, ptr %14, i32 8
83 store <4 x i32> %step.add21.1, ptr %16, align 4
84 %17 = getelementptr inbounds i32, ptr %14, i32 12
85 store <4 x i32> %step.add22.1, ptr %17, align 4
86 %18 = getelementptr inbounds i32, ptr %14, i32 16
87 store <4 x i32> %step.add23.1, ptr %18, align 4
88 %19 = getelementptr inbounds i32, ptr %14, i32 20
89 store <4 x i32> %step.add24.1, ptr %19, align 4
90 %20 = getelementptr inbounds i32, ptr %14, i32 24
91 store <4 x i32> %step.add25.1, ptr %20, align 4
92 %21 = getelementptr inbounds i32, ptr %14, i32 28
93 store <4 x i32> %step.add26.1, ptr %21, align 4
94 %22 = getelementptr inbounds i32, ptr %14, i32 32
95 store <4 x i32> %step.add27.1, ptr %22, align 4
96 %23 = getelementptr inbounds i32, ptr %14, i32 36
97 store <4 x i32> %step.add28.1, ptr %23, align 4
98 %24 = getelementptr inbounds i32, ptr %14, i32 40
99 store <4 x i32> %step.add29.1, ptr %24, align 4
100 %25 = getelementptr inbounds i32, ptr %14, i32 44
101 store <4 x i32> %step.add30.1, ptr %25, align 4
102 %index.next.1 = add nuw nsw i32 %index, 96
103 %vec.ind.next.1 = add <4 x i32> %vec.ind, <i32 96, i32 96, i32 96, i32 96>
104 br label %vector.body
106 vector.body37: ; preds = %vector.body37.1, %for.body
107 %index38 = phi i32 [ 0, %for.body ], [ %index.next53.1, %vector.body37.1 ]
108 %vec.ind39 = phi <4 x i32> [ <i32 0, i32 1, i32 2, i32 3>, %for.body ], [ %vec.ind.next52.1, %vector.body37.1 ]
109 %step.add40 = add <4 x i32> %vec.ind39, <i32 4, i32 4, i32 4, i32 4>
110 %step.add41 = add <4 x i32> %vec.ind39, <i32 8, i32 8, i32 8, i32 8>
111 %step.add42 = add <4 x i32> %vec.ind39, <i32 12, i32 12, i32 12, i32 12>
112 %step.add43 = add <4 x i32> %vec.ind39, <i32 16, i32 16, i32 16, i32 16>
113 %step.add44 = add <4 x i32> %vec.ind39, <i32 20, i32 20, i32 20, i32 20>
114 %step.add45 = add <4 x i32> %vec.ind39, <i32 24, i32 24, i32 24, i32 24>
115 %step.add46 = add <4 x i32> %vec.ind39, <i32 28, i32 28, i32 28, i32 28>
116 %step.add47 = add <4 x i32> %vec.ind39, <i32 32, i32 32, i32 32, i32 32>
117 %step.add48 = add <4 x i32> %vec.ind39, <i32 36, i32 36, i32 36, i32 36>
118 %step.add49 = add <4 x i32> %vec.ind39, <i32 40, i32 40, i32 40, i32 40>
119 %step.add50 = add <4 x i32> %vec.ind39, <i32 44, i32 44, i32 44, i32 44>
120 %26 = getelementptr inbounds [4096 x i32], ptr @a, i32 0, i32 %index38
121 store <4 x i32> %vec.ind39, ptr %26, align 4
122 %27 = getelementptr inbounds i32, ptr %26, i32 4
123 store <4 x i32> %step.add40, ptr %27, align 4
124 %28 = getelementptr inbounds i32, ptr %26, i32 8
125 store <4 x i32> %step.add41, ptr %28, align 4
126 %29 = getelementptr inbounds i32, ptr %26, i32 12
127 store <4 x i32> %step.add42, ptr %29, align 4
128 %30 = getelementptr inbounds i32, ptr %26, i32 16
129 store <4 x i32> %step.add43, ptr %30, align 4
130 %31 = getelementptr inbounds i32, ptr %26, i32 20
131 store <4 x i32> %step.add44, ptr %31, align 4
132 %32 = getelementptr inbounds i32, ptr %26, i32 24
133 store <4 x i32> %step.add45, ptr %32, align 4
134 %33 = getelementptr inbounds i32, ptr %26, i32 28
135 store <4 x i32> %step.add46, ptr %33, align 4
136 %34 = getelementptr inbounds i32, ptr %26, i32 32
137 store <4 x i32> %step.add47, ptr %34, align 4
138 %35 = getelementptr inbounds i32, ptr %26, i32 36
139 store <4 x i32> %step.add48, ptr %35, align 4
140 %36 = getelementptr inbounds i32, ptr %26, i32 40
141 store <4 x i32> %step.add49, ptr %36, align 4
142 %37 = getelementptr inbounds i32, ptr %26, i32 44
143 store <4 x i32> %step.add50, ptr %37, align 4
144 %index.next53 = add nuw nsw i32 %index38, 48
145 %38 = icmp eq i32 %index.next53, 4080
146 br i1 %38, label %for.body5, label %vector.body37.1
148 vector.body37.1: ; preds = %vector.body37
149 %vec.ind.next52 = add <4 x i32> %vec.ind39, <i32 48, i32 48, i32 48, i32 48>
150 %step.add40.1 = add <4 x i32> %vec.ind39, <i32 52, i32 52, i32 52, i32 52>
151 %step.add41.1 = add <4 x i32> %vec.ind39, <i32 56, i32 56, i32 56, i32 56>
152 %step.add42.1 = add <4 x i32> %vec.ind39, <i32 60, i32 60, i32 60, i32 60>
153 %step.add43.1 = add <4 x i32> %vec.ind39, <i32 64, i32 64, i32 64, i32 64>
154 %step.add44.1 = add <4 x i32> %vec.ind39, <i32 68, i32 68, i32 68, i32 68>
155 %step.add45.1 = add <4 x i32> %vec.ind39, <i32 72, i32 72, i32 72, i32 72>
156 %step.add46.1 = add <4 x i32> %vec.ind39, <i32 76, i32 76, i32 76, i32 76>
157 %step.add47.1 = add <4 x i32> %vec.ind39, <i32 80, i32 80, i32 80, i32 80>
158 %step.add48.1 = add <4 x i32> %vec.ind39, <i32 84, i32 84, i32 84, i32 84>
159 %step.add49.1 = add <4 x i32> %vec.ind39, <i32 88, i32 88, i32 88, i32 88>
160 %step.add50.1 = add <4 x i32> %vec.ind39, <i32 92, i32 92, i32 92, i32 92>
161 %39 = getelementptr inbounds [4096 x i32], ptr @a, i32 0, i32 %index.next53
162 store <4 x i32> %vec.ind.next52, ptr %39, align 4
163 %40 = getelementptr inbounds i32, ptr %39, i32 4
164 store <4 x i32> %step.add40.1, ptr %40, align 4
165 %41 = getelementptr inbounds i32, ptr %39, i32 8
166 store <4 x i32> %step.add41.1, ptr %41, align 4
167 %42 = getelementptr inbounds i32, ptr %39, i32 12
168 store <4 x i32> %step.add42.1, ptr %42, align 4
169 %43 = getelementptr inbounds i32, ptr %39, i32 16
170 store <4 x i32> %step.add43.1, ptr %43, align 4
171 %44 = getelementptr inbounds i32, ptr %39, i32 20
172 store <4 x i32> %step.add44.1, ptr %44, align 4
173 %45 = getelementptr inbounds i32, ptr %39, i32 24
174 store <4 x i32> %step.add45.1, ptr %45, align 4
175 %46 = getelementptr inbounds i32, ptr %39, i32 28
176 store <4 x i32> %step.add46.1, ptr %46, align 4
177 %47 = getelementptr inbounds i32, ptr %39, i32 32
178 store <4 x i32> %step.add47.1, ptr %47, align 4
179 %48 = getelementptr inbounds i32, ptr %39, i32 36
180 store <4 x i32> %step.add48.1, ptr %48, align 4
181 %49 = getelementptr inbounds i32, ptr %39, i32 40
182 store <4 x i32> %step.add49.1, ptr %49, align 4
183 %50 = getelementptr inbounds i32, ptr %39, i32 44
184 store <4 x i32> %step.add50.1, ptr %50, align 4
185 %index.next53.1 = add nuw nsw i32 %index38, 96
186 %vec.ind.next52.1 = add <4 x i32> %vec.ind39, <i32 96, i32 96, i32 96, i32 96>
187 br label %vector.body37
189 for.body: ; preds = %vector.body
190 store i32 4080, ptr getelementptr inbounds ([4096 x i32], ptr @a, i32 0, i32 4080), align 4
191 store i32 4081, ptr getelementptr inbounds ([4096 x i32], ptr @a, i32 0, i32 4081), align 4
192 store i32 4082, ptr getelementptr inbounds ([4096 x i32], ptr @a, i32 0, i32 4082), align 4
193 store i32 4083, ptr getelementptr inbounds ([4096 x i32], ptr @a, i32 0, i32 4083), align 4
194 store i32 4084, ptr getelementptr inbounds ([4096 x i32], ptr @a, i32 0, i32 4084), align 4
195 store i32 4085, ptr getelementptr inbounds ([4096 x i32], ptr @a, i32 0, i32 4085), align 4
196 store i32 4086, ptr getelementptr inbounds ([4096 x i32], ptr @a, i32 0, i32 4086), align 4
197 store i32 4087, ptr getelementptr inbounds ([4096 x i32], ptr @a, i32 0, i32 4087), align 4
198 store i32 4088, ptr getelementptr inbounds ([4096 x i32], ptr @a, i32 0, i32 4088), align 4
199 store i32 4089, ptr getelementptr inbounds ([4096 x i32], ptr @a, i32 0, i32 4089), align 4
200 store i32 4090, ptr getelementptr inbounds ([4096 x i32], ptr @a, i32 0, i32 4090), align 4
201 store i32 4091, ptr getelementptr inbounds ([4096 x i32], ptr @a, i32 0, i32 4091), align 4
202 store i32 4092, ptr getelementptr inbounds ([4096 x i32], ptr @a, i32 0, i32 4092), align 4
203 store i32 4093, ptr getelementptr inbounds ([4096 x i32], ptr @a, i32 0, i32 4093), align 4
204 store i32 4094, ptr getelementptr inbounds ([4096 x i32], ptr @a, i32 0, i32 4094), align 4
205 store i32 4095, ptr getelementptr inbounds ([4096 x i32], ptr @a, i32 0, i32 4095), align 4
206 call void @_ZN1dILi17EE1eEv(ptr noundef nonnull align 4 dereferenceable(524288) %c)
207 br label %vector.body37
209 for.body5: ; preds = %vector.body37
210 store i32 4080, ptr getelementptr inbounds ([4096 x i32], ptr @a, i32 0, i32 4080), align 4
211 store i32 4081, ptr getelementptr inbounds ([4096 x i32], ptr @a, i32 0, i32 4081), align 4
212 store i32 4082, ptr getelementptr inbounds ([4096 x i32], ptr @a, i32 0, i32 4082), align 4
213 store i32 4083, ptr getelementptr inbounds ([4096 x i32], ptr @a, i32 0, i32 4083), align 4
214 store i32 4084, ptr getelementptr inbounds ([4096 x i32], ptr @a, i32 0, i32 4084), align 4
215 store i32 4085, ptr getelementptr inbounds ([4096 x i32], ptr @a, i32 0, i32 4085), align 4
216 store i32 4086, ptr getelementptr inbounds ([4096 x i32], ptr @a, i32 0, i32 4086), align 4
217 store i32 4087, ptr getelementptr inbounds ([4096 x i32], ptr @a, i32 0, i32 4087), align 4
218 store i32 4088, ptr getelementptr inbounds ([4096 x i32], ptr @a, i32 0, i32 4088), align 4
219 store i32 4089, ptr getelementptr inbounds ([4096 x i32], ptr @a, i32 0, i32 4089), align 4
220 store i32 4090, ptr getelementptr inbounds ([4096 x i32], ptr @a, i32 0, i32 4090), align 4
221 store i32 4091, ptr getelementptr inbounds ([4096 x i32], ptr @a, i32 0, i32 4091), align 4
222 store i32 4092, ptr getelementptr inbounds ([4096 x i32], ptr @a, i32 0, i32 4092), align 4
223 store i32 4093, ptr getelementptr inbounds ([4096 x i32], ptr @a, i32 0, i32 4093), align 4
224 store i32 4094, ptr getelementptr inbounds ([4096 x i32], ptr @a, i32 0, i32 4094), align 4
225 store i32 4095, ptr getelementptr inbounds ([4096 x i32], ptr @a, i32 0, i32 4095), align 4
226 call void @_Z1h1dILi17EE(ptr noundef nonnull byval(%struct.d) align 4 %c)
227 call void @llvm.lifetime.end.p0(i64 524288, ptr nonnull %c) #3
228 ret void
229 }
231 ; Function Attrs: argmemonly mustprogress nocallback nofree nosync nounwind willreturn
232 declare void @llvm.lifetime.start.p0(i64 immarg, ptr nocapture) #1
234 ; Function Attrs: argmemonly mustprogress nocallback nofree nosync nounwind willreturn
235 declare void @llvm.lifetime.end.p0(i64 immarg, ptr nocapture) #1
237 declare dso_local void @_ZN1dILi17EE1eEv(ptr noundef nonnull align 4 dereferenceable(524288)) local_unnamed_addr #2
239 declare dso_local void @_Z1h1dILi17EE(ptr noundef byval(%struct.d) align 4) local_unnamed_addr #2
241 attributes #0 = { nounwind }