1; RUN: opt < %s  -loop-vectorize -mtriple=x86_64-apple-macosx10.8.0 -mcpu=corei7-avx -S | 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-macosx10.8.0"
5
6@c = common global [2048 x i32] zeroinitializer, align 16
7@b = common global [2048 x i32] zeroinitializer, align 16
8@d = common global [2048 x i32] zeroinitializer, align 16
9@a = common global [2048 x i32] zeroinitializer, align 16
10
11; The program below gathers and scatters data. We better not vectorize it.
12;CHECK-LABEL: @cost_model_1(
13;CHECK-NOT: <2 x i32>
14;CHECK-NOT: <4 x i32>
15;CHECK-NOT: <8 x i32>
16;CHECK: ret void
17define void @cost_model_1() nounwind uwtable noinline ssp {
18entry:
19  br label %for.body
20
21for.body:                                         ; preds = %for.body, %entry
22  %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
23  %0 = shl nsw i64 %indvars.iv, 1
24  %arrayidx = getelementptr inbounds [2048 x i32], [2048 x i32]* @c, i64 0, i64 %0
25  %1 = load i32, i32* %arrayidx, align 8
26  %idxprom1 = sext i32 %1 to i64
27  %arrayidx2 = getelementptr inbounds [2048 x i32], [2048 x i32]* @b, i64 0, i64 %idxprom1
28  %2 = load i32, i32* %arrayidx2, align 4
29  %arrayidx4 = getelementptr inbounds [2048 x i32], [2048 x i32]* @d, i64 0, i64 %indvars.iv
30  %3 = load i32, i32* %arrayidx4, align 4
31  %idxprom5 = sext i32 %3 to i64
32  %arrayidx6 = getelementptr inbounds [2048 x i32], [2048 x i32]* @a, i64 0, i64 %idxprom5
33  store i32 %2, i32* %arrayidx6, align 4
34  %indvars.iv.next = add i64 %indvars.iv, 1
35  %lftr.wideiv = trunc i64 %indvars.iv.next to i32
36  %exitcond = icmp eq i32 %lftr.wideiv, 256
37  br i1 %exitcond, label %for.end, label %for.body
38
39for.end:                                          ; preds = %for.body
40  ret void
41}
42
43; This function uses a stride that is generally too big to benefit from vectorization without
44; really good support for a gather load. We were not computing an accurate cost for the
45; vectorization and subsequent scalarization of the pointer induction variables.
46
47define float @PR27826(float* nocapture readonly %a, float* nocapture readonly %b, i32 %n) {
48; CHECK-LABEL: @PR27826(
49; CHECK-NOT:   <4 x float>
50; CHECK-NOT:   <8 x float>
51; CHECK:       ret float %s.0.lcssa
52
53entry:
54  %cmp = icmp sgt i32 %n, 0
55  br i1 %cmp, label %preheader, label %for.end
56
57preheader:
58  %t0 = sext i32 %n to i64
59  br label %for
60
61for:
62  %indvars.iv = phi i64 [ 0, %preheader ], [ %indvars.iv.next, %for ]
63  %s.02 = phi float [ 0.0, %preheader ], [ %add4, %for ]
64  %arrayidx = getelementptr inbounds float, float* %a, i64 %indvars.iv
65  %t1 = load float, float* %arrayidx, align 4
66  %arrayidx3 = getelementptr inbounds float, float* %b, i64 %indvars.iv
67  %t2 = load float, float* %arrayidx3, align 4
68  %add = fadd fast float %t1, %s.02
69  %add4 = fadd fast float %add, %t2
70  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 8
71  %cmp1 = icmp slt i64 %indvars.iv.next, %t0
72  br i1 %cmp1, label %for, label %loopexit
73
74loopexit:
75  %add4.lcssa = phi float [ %add4, %for ]
76  br label %for.end
77
78for.end:
79  %s.0.lcssa = phi float [ 0.0, %entry ], [ %add4.lcssa, %loopexit ]
80  ret float %s.0.lcssa
81}
82
83