1; RUN: opt -S -loop-vectorize -force-vector-width=4 -force-vector-interleave=1 < %s | FileCheck %s
2
3target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
4
5; Check the scenario where we have an unknown Stride, which happens to also be
6; the loop iteration count, so if we specialize the loop for the Stride==1 case,
7; this also implies that the loop will iterate no more than a single iteration,
8; as in the following example:
9;
10;       unsigned int N;
11;       int tmp = 0;
12;       for(unsigned int k=0;k<N;k++) {
13;         tmp+=(int)B[k*N+j];
14;       }
15;
16; We check here that the following runtime scev guard for Stride==1 is NOT generated:
17; vector.scevcheck:
18;   %ident.check = icmp ne i32 %N, 1
19;   %0 = or i1 false, %ident.check
20;   br i1 %0, label %scalar.ph, label %vector.ph
21; Instead the loop is vectorized with an unknown stride.
22
23; CHECK-LABEL: @foo1
24; CHECK: for.body.lr.ph
25; CHECK-NOT: %ident.check = icmp ne i32 %N, 1
26; CHECK-NOT: %[[TEST:[0-9]+]] = or i1 false, %ident.check
27; CHECK-NOT: br i1 %[[TEST]], label %scalar.ph, label %vector.ph
28; CHECK: vector.ph
29; CHECK: vector.body
30; CHECK: <4 x i32>
31; CHECK: middle.block
32; CHECK: scalar.ph
33
34
35define i32 @foo1(i32 %N, i16* nocapture readnone %A, i16* nocapture readonly %B, i32 %i, i32 %j)  {
36entry:
37  %cmp8 = icmp eq i32 %N, 0
38  br i1 %cmp8, label %for.end, label %for.body.lr.ph
39
40for.body.lr.ph:
41  br label %for.body
42
43for.body:
44  %tmp.010 = phi i32 [ 0, %for.body.lr.ph ], [ %add1, %for.body ]
45  %k.09 = phi i32 [ 0, %for.body.lr.ph ], [ %inc, %for.body ]
46  %mul = mul i32 %k.09, %N
47  %add = add i32 %mul, %j
48  %arrayidx = getelementptr inbounds i16, i16* %B, i32 %add
49  %0 = load i16, i16* %arrayidx, align 2
50  %conv = sext i16 %0 to i32
51  %add1 = add nsw i32 %tmp.010, %conv
52  %inc = add nuw i32 %k.09, 1
53  %exitcond = icmp eq i32 %inc, %N
54  br i1 %exitcond, label %for.end.loopexit, label %for.body
55
56for.end.loopexit:
57  %add1.lcssa = phi i32 [ %add1, %for.body ]
58  br label %for.end
59
60for.end:
61  %tmp.0.lcssa = phi i32 [ 0, %entry ], [ %add1.lcssa, %for.end.loopexit ]
62  ret i32 %tmp.0.lcssa
63}
64
65
66; Check the same, but also where the Stride and the loop iteration count
67; are not of the same data type.
68;
69;       unsigned short N;
70;       int tmp = 0;
71;       for(unsigned int k=0;k<N;k++) {
72;         tmp+=(int)B[k*N+j];
73;       }
74;
75; We check here that the following runtime scev guard for Stride==1 is NOT generated:
76; vector.scevcheck:
77; %ident.check = icmp ne i16 %N, 1
78; %0 = or i1 false, %ident.check
79; br i1 %0, label %scalar.ph, label %vector.ph
80
81
82; CHECK-LABEL: @foo2
83; CHECK: for.body.lr.ph
84; CHECK-NOT: %ident.check = icmp ne i16 %N, 1
85; CHECK-NOT: %[[TEST:[0-9]+]] = or i1 false, %ident.check
86; CHECK-NOT: br i1 %[[TEST]], label %scalar.ph, label %vector.ph
87; CHECK: vector.ph
88; CHECK: vector.body
89; CHECK: <4 x i32>
90; CHECK: middle.block
91; CHECK: scalar.ph
92
93define i32 @foo2(i16 zeroext %N, i16* nocapture readnone %A, i16* nocapture readonly %B, i32 %i, i32 %j) {
94entry:
95  %conv = zext i16 %N to i32
96  %cmp11 = icmp eq i16 %N, 0
97  br i1 %cmp11, label %for.end, label %for.body.lr.ph
98
99for.body.lr.ph:
100  br label %for.body
101
102for.body:
103  %tmp.013 = phi i32 [ 0, %for.body.lr.ph ], [ %add4, %for.body ]
104  %k.012 = phi i32 [ 0, %for.body.lr.ph ], [ %inc, %for.body ]
105  %mul = mul nuw i32 %k.012, %conv
106  %add = add i32 %mul, %j
107  %arrayidx = getelementptr inbounds i16, i16* %B, i32 %add
108  %0 = load i16, i16* %arrayidx, align 2
109  %conv3 = sext i16 %0 to i32
110  %add4 = add nsw i32 %tmp.013, %conv3
111  %inc = add nuw nsw i32 %k.012, 1
112  %exitcond = icmp eq i32 %inc, %conv
113  br i1 %exitcond, label %for.end.loopexit, label %for.body
114
115for.end.loopexit:
116  %add4.lcssa = phi i32 [ %add4, %for.body ]
117  br label %for.end
118
119for.end:
120  %tmp.0.lcssa = phi i32 [ 0, %entry ], [ %add4.lcssa, %for.end.loopexit ]
121  ret i32 %tmp.0.lcssa
122}
123