1; RUN: opt -basicaa -loop-distribute -verify-loop-info -verify-dom-info -S \
2; RUN:   < %s | FileCheck %s
3
4; RUN: opt -basicaa -loop-distribute -loop-vectorize -force-vector-width=4 \
5; RUN:   -verify-loop-info -verify-dom-info -S < %s | \
6; RUN:   FileCheck --check-prefix=VECTORIZE %s
7
8; The memcheck version of basic.ll.  We should distribute and vectorize the
9; second part of this loop with 5 memchecks (A+1 x {C, D, E} + C x {A, B})
10;
11;   for (i = 0; i < n; i++) {
12;     A[i + 1] = A[i] * B[i];
13; -------------------------------
14;     C[i] = D[i] * E[i];
15;   }
16
17target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128"
18target triple = "x86_64-apple-macosx10.10.0"
19
20@B = common global i32* null, align 8
21@A = common global i32* null, align 8
22@C = common global i32* null, align 8
23@D = common global i32* null, align 8
24@E = common global i32* null, align 8
25
26define void @f() {
27entry:
28  %a = load i32*, i32** @A, align 8
29  %b = load i32*, i32** @B, align 8
30  %c = load i32*, i32** @C, align 8
31  %d = load i32*, i32** @D, align 8
32  %e = load i32*, i32** @E, align 8
33  br label %for.body
34
35; We have two compares for each array overlap check which is a total of 10
36; compares.
37;
38; CHECK: for.body.ldist.memcheck:
39; CHECK:     = icmp
40; CHECK:     = icmp
41
42; CHECK:     = icmp
43; CHECK:     = icmp
44
45; CHECK:     = icmp
46; CHECK:     = icmp
47
48; CHECK:     = icmp
49; CHECK:     = icmp
50
51; CHECK:     = icmp
52; CHECK:     = icmp
53
54; CHECK-NOT: = icmp
55; CHECK:     br i1 %memcheck.conflict, label %for.body.ph.ldist.nondist, label %for.body.ph.ldist1
56
57; The non-distributed loop that the memchecks fall back on.
58
59; CHECK: for.body.ph.ldist.nondist:
60; CHECK:     br label %for.body.ldist.nondist
61; CHECK: for.body.ldist.nondist:
62; CHECK:    br i1 %exitcond.ldist.nondist, label %for.end, label %for.body.ldist.nondist
63
64; Verify the two distributed loops.
65
66; CHECK: for.body.ph.ldist1:
67; CHECK:     br label %for.body.ldist1
68; CHECK: for.body.ldist1:
69; CHECK:    %mulA.ldist1 = mul i32 %loadB.ldist1, %loadA.ldist1
70; CHECK:    br i1 %exitcond.ldist1, label %for.body.ph, label %for.body.ldist1
71
72; CHECK: for.body.ph:
73; CHECK:    br label %for.body
74; CHECK: for.body:
75; CHECK:    %mulC = mul i32 %loadD, %loadE
76; CHECK: for.end:
77
78
79; VECTORIZE: mul <4 x i32>
80
81for.body:                                         ; preds = %for.body, %entry
82  %ind = phi i64 [ 0, %entry ], [ %add, %for.body ]
83
84  %arrayidxA = getelementptr inbounds i32, i32* %a, i64 %ind
85  %loadA = load i32, i32* %arrayidxA, align 4
86
87  %arrayidxB = getelementptr inbounds i32, i32* %b, i64 %ind
88  %loadB = load i32, i32* %arrayidxB, align 4
89
90  %mulA = mul i32 %loadB, %loadA
91
92  %add = add nuw nsw i64 %ind, 1
93  %arrayidxA_plus_4 = getelementptr inbounds i32, i32* %a, i64 %add
94  store i32 %mulA, i32* %arrayidxA_plus_4, align 4
95
96  %arrayidxD = getelementptr inbounds i32, i32* %d, i64 %ind
97  %loadD = load i32, i32* %arrayidxD, align 4
98
99  %arrayidxE = getelementptr inbounds i32, i32* %e, i64 %ind
100  %loadE = load i32, i32* %arrayidxE, align 4
101
102  %mulC = mul i32 %loadD, %loadE
103
104  %arrayidxC = getelementptr inbounds i32, i32* %c, i64 %ind
105  store i32 %mulC, i32* %arrayidxC, align 4
106
107  %exitcond = icmp eq i64 %add, 20
108  br i1 %exitcond, label %for.end, label %for.body
109
110for.end:                                          ; preds = %for.body
111  ret void
112}
113