1; RUN: opt < %s  -loop-vectorize -force-vector-interleave=1 -force-vector-width=4 -dce -instcombine -S -enable-if-conversion | FileCheck %s
2
3target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128"
4target triple = "x86_64-apple-macosx10.9.0"
5
6; This is kernel11 from "LivermoreLoops". We can't vectorize it because we
7; access both x[k] and x[k-1].
8;
9; void kernel11(double *x, double *y, int n) {
10;   for ( int k=1 ; k<n ; k++ )
11;     x[k] = x[k-1] + y[k];
12; }
13
14; CHECK-LABEL: @kernel11(
15; CHECK-NOT: <4 x double>
16; CHECK: ret
17define i32 @kernel11(double* %x, double* %y, i32 %n) nounwind uwtable ssp {
18  %1 = alloca double*, align 8
19  %2 = alloca double*, align 8
20  %3 = alloca i32, align 4
21  %k = alloca i32, align 4
22  store double* %x, double** %1, align 8
23  store double* %y, double** %2, align 8
24  store i32 %n, i32* %3, align 4
25  store i32 1, i32* %k, align 4
26  br label %4
27
28; <label>:4                                       ; preds = %25, %0
29  %5 = load i32, i32* %k, align 4
30  %6 = load i32, i32* %3, align 4
31  %7 = icmp slt i32 %5, %6
32  br i1 %7, label %8, label %28
33
34; <label>:8                                       ; preds = %4
35  %9 = load i32, i32* %k, align 4
36  %10 = sub nsw i32 %9, 1
37  %11 = sext i32 %10 to i64
38  %12 = load double*, double** %1, align 8
39  %13 = getelementptr inbounds double, double* %12, i64 %11
40  %14 = load double, double* %13, align 8
41  %15 = load i32, i32* %k, align 4
42  %16 = sext i32 %15 to i64
43  %17 = load double*, double** %2, align 8
44  %18 = getelementptr inbounds double, double* %17, i64 %16
45  %19 = load double, double* %18, align 8
46  %20 = fadd double %14, %19
47  %21 = load i32, i32* %k, align 4
48  %22 = sext i32 %21 to i64
49  %23 = load double*, double** %1, align 8
50  %24 = getelementptr inbounds double, double* %23, i64 %22
51  store double %20, double* %24, align 8
52  br label %25
53
54; <label>:25                                      ; preds = %8
55  %26 = load i32, i32* %k, align 4
56  %27 = add nsw i32 %26, 1
57  store i32 %27, i32* %k, align 4
58  br label %4
59
60; <label>:28                                      ; preds = %4
61  ret i32 0
62}
63
64
65; A[i*7] is scalarized, and the different scalars can in theory wrap
66; around and overwrite other scalar elements. However we can still
67; vectorize because we can version the loop to avoid this case.
68;
69; void foo(int *a) {
70;   for (int i=0; i<256; ++i) {
71;     int x = a[i*7];
72;     if (x>3)
73;       x = x*x+x*4;
74;     a[i*7] = x+3;
75;   }
76; }
77
78; CHECK-LABEL: @func2(
79; CHECK: <4 x i32>
80; CHECK: ret
81define i32 @func2(i32* nocapture %a) nounwind uwtable ssp {
82  br label %1
83
84; <label>:1                                       ; preds = %7, %0
85  %indvars.iv = phi i64 [ 0, %0 ], [ %indvars.iv.next, %7 ]
86  %2 = mul nsw i64 %indvars.iv, 7
87  %3 = getelementptr inbounds i32, i32* %a, i64 %2
88  %4 = load i32, i32* %3, align 4
89  %5 = icmp sgt i32 %4, 3
90  br i1 %5, label %6, label %7
91
92; <label>:6                                       ; preds = %1
93  %tmp = add i32 %4, 4
94  %tmp1 = mul i32 %tmp, %4
95  br label %7
96
97; <label>:7                                       ; preds = %6, %1
98  %x.0 = phi i32 [ %tmp1, %6 ], [ %4, %1 ]
99  %8 = add nsw i32 %x.0, 3
100  store i32 %8, i32* %3, align 4
101  %indvars.iv.next = add i64 %indvars.iv, 1
102  %lftr.wideiv = trunc i64 %indvars.iv.next to i32
103  %exitcond = icmp eq i32 %lftr.wideiv, 256
104  br i1 %exitcond, label %9, label %1
105
106; <label>:9                                       ; preds = %7
107  ret i32 0
108}
109