1 ; RUN: opt -passes=loop-vectorize -mtriple=arm64-apple-ios -S -mcpu=cyclone -enable-interleaved-mem-accesses=false -neon-nonconst-stride-overhead=5 < %s | FileCheck %s
2 target 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-n32:64-S128"
4 @kernel = global [512 x float] zeroinitializer, align 16
5 @kernel2 = global [512 x float] zeroinitializer, align 16
6 @kernel3 = global [512 x float] zeroinitializer, align 16
7 @kernel4 = global [512 x float] zeroinitializer, align 16
8 @src_data = global [1536 x float] zeroinitializer, align 16
9 @r_ = global i8 0, align 1
10 @g_ = global i8 0, align 1
11 @b_ = global i8 0, align 1
13 ; vectorize loop when we lower the cost of gather load
14 ; Make sure we vectorize it, VF = 4
17 define void @_Z4testmm(i64 %size, i64 %offset) {
19 %cmp53 = icmp eq i64 %size, 0
20 br i1 %cmp53, label %for.end, label %for.body.lr.ph
26 %r.057 = phi float [ 0.000000e+00, %for.body.lr.ph ], [ %add10, %for.body ]
27 %g.056 = phi float [ 0.000000e+00, %for.body.lr.ph ], [ %add20, %for.body ]
28 %v.055 = phi i64 [ 0, %for.body.lr.ph ], [ %inc, %for.body ]
29 %b.054 = phi float [ 0.000000e+00, %for.body.lr.ph ], [ %add30, %for.body ]
30 %add = add i64 %v.055, %offset
31 %mul = mul i64 %add, 3
32 %arrayidx = getelementptr inbounds [1536 x float], ptr @src_data, i64 0, i64 %mul
33 %0 = load float, ptr %arrayidx, align 4
34 %arrayidx2 = getelementptr inbounds [512 x float], ptr @kernel, i64 0, i64 %v.055
35 %1 = load float, ptr %arrayidx2, align 4
36 %mul3 = fmul fast float %0, %1
37 %arrayidx4 = getelementptr inbounds [512 x float], ptr @kernel2, i64 0, i64 %v.055
38 %2 = load float, ptr %arrayidx4, align 4
39 %mul5 = fmul fast float %mul3, %2
40 %arrayidx6 = getelementptr inbounds [512 x float], ptr @kernel3, i64 0, i64 %v.055
41 %3 = load float, ptr %arrayidx6, align 4
42 %mul7 = fmul fast float %mul5, %3
43 %arrayidx8 = getelementptr inbounds [512 x float], ptr @kernel4, i64 0, i64 %v.055
44 %4 = load float, ptr %arrayidx8, align 4
45 %mul9 = fmul fast float %mul7, %4
46 %add10 = fadd fast float %r.057, %mul9
47 %arrayidx.sum = add i64 %mul, 1
48 %arrayidx11 = getelementptr inbounds [1536 x float], ptr @src_data, i64 0, i64 %arrayidx.sum
49 %5 = load float, ptr %arrayidx11, align 4
50 %mul13 = fmul fast float %1, %5
51 %mul15 = fmul fast float %2, %mul13
52 %mul17 = fmul fast float %3, %mul15
53 %mul19 = fmul fast float %4, %mul17
54 %add20 = fadd fast float %g.056, %mul19
55 %arrayidx.sum52 = add i64 %mul, 2
56 %arrayidx21 = getelementptr inbounds [1536 x float], ptr @src_data, i64 0, i64 %arrayidx.sum52
57 %6 = load float, ptr %arrayidx21, align 4
58 %mul23 = fmul fast float %1, %6
59 %mul25 = fmul fast float %2, %mul23
60 %mul27 = fmul fast float %3, %mul25
61 %mul29 = fmul fast float %4, %mul27
62 %add30 = fadd fast float %b.054, %mul29
63 %inc = add i64 %v.055, 1
64 %exitcond = icmp ne i64 %inc, %size
65 br i1 %exitcond, label %for.body, label %for.cond.for.end_crit_edge
67 for.cond.for.end_crit_edge:
68 %add30.lcssa = phi float [ %add30, %for.body ]
69 %add20.lcssa = phi float [ %add20, %for.body ]
70 %add10.lcssa = phi float [ %add10, %for.body ]
71 %phitmp = fptoui float %add10.lcssa to i8
72 %phitmp60 = fptoui float %add20.lcssa to i8
73 %phitmp61 = fptoui float %add30.lcssa to i8
77 %r.0.lcssa = phi i8 [ %phitmp, %for.cond.for.end_crit_edge ], [ 0, %entry ]
78 %g.0.lcssa = phi i8 [ %phitmp60, %for.cond.for.end_crit_edge ], [ 0, %entry ]
79 %b.0.lcssa = phi i8 [ %phitmp61, %for.cond.for.end_crit_edge ], [ 0, %entry ]
80 store i8 %r.0.lcssa, ptr @r_, align 1
81 store i8 %g.0.lcssa, ptr @g_, align 1
82 store i8 %b.0.lcssa, ptr @b_, align 1