1; RUN: opt < %s  -loop-vectorize -force-vector-interleave=1 -force-vector-width=4 -dce -instcombine -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-unknown-linux-gnu"
5
6; CHECK-LABEL: @foo(
7; CHECK: <4 x i32>
8; CHECK: ret void
9
10; PR15794
11; incorrect addition of llvm.mem.parallel_loop_access metadata is undefined
12; behaviour. Vectorizer ignores the memory dependency checks and goes ahead and
13; vectorizes this loop with uniform stores which has an output dependency.
14
15; void foo(int *a, int *b, int k, int m) {
16;   for (int i = 0; i < m; i++) {
17;     for (int j = 0; j < m; j++) {
18;       a[i] = a[i + j + k] + 1; <<<
19;     }
20;     b[i] = b[i] + 3;
21;   }
22; }
23
24; Function Attrs: nounwind uwtable
25define void @foo(i32* nocapture %a, i32* nocapture %b, i32 %k, i32 %m) #0 {
26entry:
27  %cmp27 = icmp sgt i32 %m, 0
28  br i1 %cmp27, label %for.body3.lr.ph.us, label %for.end15
29
30for.end.us:                                       ; preds = %for.body3.us
31  %arrayidx9.us = getelementptr inbounds i32, i32* %b, i64 %indvars.iv33
32  %0 = load i32, i32* %arrayidx9.us, align 4, !llvm.mem.parallel_loop_access !3
33  %add10.us = add nsw i32 %0, 3
34  store i32 %add10.us, i32* %arrayidx9.us, align 4, !llvm.mem.parallel_loop_access !3
35  %indvars.iv.next34 = add i64 %indvars.iv33, 1
36  %lftr.wideiv35 = trunc i64 %indvars.iv.next34 to i32
37  %exitcond36 = icmp eq i32 %lftr.wideiv35, %m
38  br i1 %exitcond36, label %for.end15, label %for.body3.lr.ph.us, !llvm.loop !5
39
40for.body3.us:                                     ; preds = %for.body3.us, %for.body3.lr.ph.us
41  %indvars.iv29 = phi i64 [ 0, %for.body3.lr.ph.us ], [ %indvars.iv.next30, %for.body3.us ]
42  %1 = trunc i64 %indvars.iv29 to i32
43  %add4.us = add i32 %add.us, %1
44  %idxprom.us = sext i32 %add4.us to i64
45  %arrayidx.us = getelementptr inbounds i32, i32* %a, i64 %idxprom.us
46  %2 = load i32, i32* %arrayidx.us, align 4, !llvm.mem.parallel_loop_access !3
47  %add5.us = add nsw i32 %2, 1
48  store i32 %add5.us, i32* %arrayidx7.us, align 4, !llvm.mem.parallel_loop_access !3
49  %indvars.iv.next30 = add i64 %indvars.iv29, 1
50  %lftr.wideiv31 = trunc i64 %indvars.iv.next30 to i32
51  %exitcond32 = icmp eq i32 %lftr.wideiv31, %m
52  br i1 %exitcond32, label %for.end.us, label %for.body3.us, !llvm.loop !4
53
54for.body3.lr.ph.us:                               ; preds = %for.end.us, %entry
55  %indvars.iv33 = phi i64 [ %indvars.iv.next34, %for.end.us ], [ 0, %entry ]
56  %3 = trunc i64 %indvars.iv33 to i32
57  %add.us = add i32 %3, %k
58  %arrayidx7.us = getelementptr inbounds i32, i32* %a, i64 %indvars.iv33
59  br label %for.body3.us
60
61for.end15:                                        ; preds = %for.end.us, %entry
62  ret void
63}
64
65; Same test as above, but without the invalid parallel_loop_access metadata.
66
67; Here we can see the vectorizer does the mem dep checks and decides it is
68; unsafe to vectorize.
69; CHECK-LABEL: no-par-mem-metadata(
70; CHECK-NOT: <4 x i32>
71; CHECK:     ret void
72define void @no-par-mem-metadata(i32* nocapture %a, i32* nocapture %b, i32 %k, i32 %m) #0 {
73entry:
74  %cmp27 = icmp sgt i32 %m, 0
75  br i1 %cmp27, label %for.body3.lr.ph.us, label %for.end15
76
77for.end.us:                                       ; preds = %for.body3.us
78  %arrayidx9.us = getelementptr inbounds i32, i32* %b, i64 %indvars.iv33
79  %0 = load i32, i32* %arrayidx9.us, align 4
80  %add10.us = add nsw i32 %0, 3
81  store i32 %add10.us, i32* %arrayidx9.us, align 4
82  %indvars.iv.next34 = add i64 %indvars.iv33, 1
83  %lftr.wideiv35 = trunc i64 %indvars.iv.next34 to i32
84  %exitcond36 = icmp eq i32 %lftr.wideiv35, %m
85  br i1 %exitcond36, label %for.end15, label %for.body3.lr.ph.us, !llvm.loop !5
86
87for.body3.us:                                     ; preds = %for.body3.us, %for.body3.lr.ph.us
88  %indvars.iv29 = phi i64 [ 0, %for.body3.lr.ph.us ], [ %indvars.iv.next30, %for.body3.us ]
89  %1 = trunc i64 %indvars.iv29 to i32
90  %add4.us = add i32 %add.us, %1
91  %idxprom.us = sext i32 %add4.us to i64
92  %arrayidx.us = getelementptr inbounds i32, i32* %a, i64 %idxprom.us
93  %2 = load i32, i32* %arrayidx.us, align 4
94  %add5.us = add nsw i32 %2, 1
95  store i32 %add5.us, i32* %arrayidx7.us, align 4
96  %indvars.iv.next30 = add i64 %indvars.iv29, 1
97  %lftr.wideiv31 = trunc i64 %indvars.iv.next30 to i32
98  %exitcond32 = icmp eq i32 %lftr.wideiv31, %m
99  br i1 %exitcond32, label %for.end.us, label %for.body3.us, !llvm.loop !4
100
101for.body3.lr.ph.us:                               ; preds = %for.end.us, %entry
102  %indvars.iv33 = phi i64 [ %indvars.iv.next34, %for.end.us ], [ 0, %entry ]
103  %3 = trunc i64 %indvars.iv33 to i32
104  %add.us = add i32 %3, %k
105  %arrayidx7.us = getelementptr inbounds i32, i32* %a, i64 %indvars.iv33
106  br label %for.body3.us
107
108for.end15:                                        ; preds = %for.end.us, %entry
109  ret void
110}
111
112attributes #0 = { nounwind uwtable "less-precise-fpmad"="false" "no-frame-pointer-elim"="false" "no-frame-pointer-elim-non-leaf"="false" "no-infs-fp-math"="false" "no-nans-fp-math"="false" "unsafe-fp-math"="false" "use-soft-float"="false" }
113
114!3 = !{!4, !5}
115!4 = !{!4}
116!5 = !{!5}
117
118