1; RUN: opt < %s  -tbaa -basic-aa -loop-vectorize -force-vector-interleave=1 -force-vector-width=4 -dce -instcombine -simplifycfg -S | FileCheck %s
2; RUN: opt < %s  -basic-aa -loop-vectorize -force-vector-interleave=1 -force-vector-width=4 -dce -instcombine -simplifycfg -S | FileCheck %s --check-prefix=CHECK-NOTBAA
3target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
4
5; Function Attrs: nounwind uwtable
6define i32 @test1(i32* nocapture %a, float* nocapture readonly %b) #0 {
7entry:
8  br label %for.body
9
10for.body:                                         ; preds = %for.body, %entry
11  %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
12  %arrayidx = getelementptr inbounds float, float* %b, i64 %indvars.iv
13  %0 = load float, float* %arrayidx, align 4, !tbaa !0
14  %conv = fptosi float %0 to i32
15  %arrayidx2 = getelementptr inbounds i32, i32* %a, i64 %indvars.iv
16  store i32 %conv, i32* %arrayidx2, align 4, !tbaa !4
17  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
18  %exitcond = icmp eq i64 %indvars.iv.next, 1600
19  br i1 %exitcond, label %for.end, label %for.body
20
21for.end:                                          ; preds = %for.body
22  ret i32 0
23
24; TBAA partitions the accesses in this loop, so it can be vectorized without
25; runtime checks.
26
27; CHECK-LABEL: @test1
28; CHECK: entry:
29; CHECK-NEXT: br label %vector.body
30; CHECK: vector.body:
31
32; CHECK: load <4 x float>, <4 x float>* %{{.*}}, align 4, !tbaa
33; CHECK: store <4 x i32> %{{.*}}, <4 x i32>* %{{.*}}, align 4, !tbaa
34
35; CHECK: ret i32 0
36
37; CHECK-NOTBAA-LABEL: @test1
38; CHECK-NOTBAA: icmp ugt i32*
39
40; CHECK-NOTBAA: load <4 x float>, <4 x float>* %{{.*}}, align 4, !tbaa
41; CHECK-NOTBAA: store <4 x i32> %{{.*}}, <4 x i32>* %{{.*}}, align 4, !tbaa
42
43; CHECK-NOTBAA: ret i32 0
44}
45
46; Function Attrs: nounwind uwtable
47define i32 @test2(i32* nocapture readonly %a, float* nocapture readonly %b, float* nocapture %c) #0 {
48entry:
49  br label %for.body
50
51for.body:                                         ; preds = %for.body, %entry
52  %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
53  %arrayidx = getelementptr inbounds float, float* %b, i64 %indvars.iv
54  %0 = load float, float* %arrayidx, align 4, !tbaa !0
55  %arrayidx2 = getelementptr inbounds i32, i32* %a, i64 %indvars.iv
56  %1 = load i32, i32* %arrayidx2, align 4, !tbaa !4
57  %conv = sitofp i32 %1 to float
58  %mul = fmul float %0, %conv
59  %arrayidx4 = getelementptr inbounds float, float* %c, i64 %indvars.iv
60  store float %mul, float* %arrayidx4, align 4, !tbaa !0
61  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
62  %exitcond = icmp eq i64 %indvars.iv.next, 1600
63  br i1 %exitcond, label %for.end, label %for.body
64
65for.end:                                          ; preds = %for.body
66  ret i32 0
67
68; This test is like the first, except here there is still one runtime check
69; required. Without TBAA, however, two checks are required.
70
71; CHECK-LABEL: @test2
72; CHECK: icmp ugt float*
73; CHECK: icmp ugt float*
74; CHECK-NOT: icmp uge i32*
75
76; CHECK: load <4 x float>, <4 x float>* %{{.*}}, align 4, !tbaa
77; CHECK: store <4 x float> %{{.*}}, <4 x float>* %{{.*}}, align 4, !tbaa
78
79; CHECK: ret i32 0
80
81; CHECK-NOTBAA-LABEL: @test2
82; CHECK-NOTBAA: icmp ugt float*
83; CHECK-NOTBAA: icmp ugt float*
84; CHECK-NOTBAA-DAG: icmp ugt float*
85; CHECK-NOTBAA-DAG: icmp ugt i32*
86
87; CHECK-NOTBAA: load <4 x float>, <4 x float>* %{{.*}}, align 4, !tbaa
88; CHECK-NOTBAA: store <4 x float> %{{.*}}, <4 x float>* %{{.*}}, align 4, !tbaa
89
90; CHECK-NOTBAA: ret i32 0
91}
92
93attributes #0 = { nounwind uwtable }
94
95!0 = !{!1, !1, i64 0}
96!1 = !{!"float", !2, i64 0}
97!2 = !{!"omnipotent char", !3, i64 0}
98!3 = !{!"Simple C/C++ TBAA"}
99!4 = !{!5, !5, i64 0}
100!5 = !{!"int", !2, i64 0}
101
102