|
221
|
1 ; RUN: opt -basic-aa -loop-distribute -enable-loop-distribute -verify-loop-info -verify-dom-info -S < %s \
|
|
150
|
2 ; RUN: | FileCheck %s
|
|
|
3
|
|
|
4 ; We should distribute this loop along === but not along ---. The last
|
|
|
5 ; partition won't be vectorized due to conditional stores so it's better to
|
|
|
6 ; keep it with the second partition which has a dependence cycle.
|
|
|
7
|
|
|
8 ; (1st statement):
|
|
|
9 ; for (i = 0; i < n; i++) {
|
|
|
10 ; C[i] = D[i] * E[i];
|
|
|
11 ;=============================
|
|
|
12 ; A[i + 1] = A[i] * B[i];
|
|
|
13 ;-----------------------------
|
|
|
14 ; if (F[i])
|
|
|
15 ; G[i] = H[i] * J[i];
|
|
|
16 ; }
|
|
|
17
|
|
|
18 target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128"
|
|
|
19 target triple = "x86_64-apple-macosx10.10.0"
|
|
|
20
|
|
|
21 define void @f(i32* noalias %a,
|
|
|
22 i32* noalias %b,
|
|
|
23 i32* noalias %c,
|
|
|
24 i32* noalias %d,
|
|
|
25 i32* noalias %e,
|
|
|
26 i32* noalias %g,
|
|
|
27 i32* noalias %h,
|
|
|
28 i32* noalias %j,
|
|
|
29 i64 %x) {
|
|
|
30 entry:
|
|
|
31 br label %for.body
|
|
|
32
|
|
|
33 ; Ensure that we have only two partitions, the first with one multiplication
|
|
|
34 ; and the second with two.
|
|
|
35
|
|
|
36 ; CHECK: for.body.ldist1:
|
|
|
37 ; CHECK: %mulC.ldist1 = mul i32 %loadD.ldist1, %loadE.ldist1
|
|
|
38 ; CHECK: br i1 %exitcond.ldist1, label %entry.split, label %for.body.ldist1
|
|
|
39 ; CHECK: entry.split:
|
|
|
40 ; CHECK: br label %for.body
|
|
|
41 ; CHECK: for.body:
|
|
|
42 ; CHECK: %mulA = mul i32 %loadB, %loadA
|
|
|
43 ; CHECK: %mulG = mul i32 %loadH, %loadJ
|
|
|
44 ; CHECK: for.end:
|
|
|
45
|
|
|
46 for.body: ; preds = %for.body, %entry
|
|
|
47 %ind = phi i64 [ 0, %entry ], [ %add, %if.end ]
|
|
|
48
|
|
|
49 %arrayidxD = getelementptr inbounds i32, i32* %d, i64 %ind
|
|
|
50 %loadD = load i32, i32* %arrayidxD, align 4
|
|
|
51
|
|
|
52 %arrayidxE = getelementptr inbounds i32, i32* %e, i64 %ind
|
|
|
53 %loadE = load i32, i32* %arrayidxE, align 4
|
|
|
54
|
|
|
55 %mulC = mul i32 %loadD, %loadE
|
|
|
56
|
|
|
57 %arrayidxC = getelementptr inbounds i32, i32* %c, i64 %ind
|
|
|
58 store i32 %mulC, i32* %arrayidxC, align 4
|
|
|
59
|
|
|
60
|
|
|
61 %arrayidxA = getelementptr inbounds i32, i32* %a, i64 %ind
|
|
|
62 %loadA = load i32, i32* %arrayidxA, align 4
|
|
|
63
|
|
|
64 %arrayidxB = getelementptr inbounds i32, i32* %b, i64 %ind
|
|
|
65 %loadB = load i32, i32* %arrayidxB, align 4
|
|
|
66
|
|
|
67 %mulA = mul i32 %loadB, %loadA
|
|
|
68
|
|
|
69 %add = add nuw nsw i64 %ind, 1
|
|
|
70 %arrayidxA_plus_4 = getelementptr inbounds i32, i32* %a, i64 %add
|
|
|
71 store i32 %mulA, i32* %arrayidxA_plus_4, align 4
|
|
|
72
|
|
|
73 %if.cond = icmp eq i64 %ind, %x
|
|
|
74 br i1 %if.cond, label %if.then, label %if.end
|
|
|
75
|
|
|
76 if.then:
|
|
|
77 %arrayidxH = getelementptr inbounds i32, i32* %h, i64 %ind
|
|
|
78 %loadH = load i32, i32* %arrayidxH, align 4
|
|
|
79
|
|
|
80 %arrayidxJ = getelementptr inbounds i32, i32* %j, i64 %ind
|
|
|
81 %loadJ = load i32, i32* %arrayidxJ, align 4
|
|
|
82
|
|
|
83 %mulG = mul i32 %loadH, %loadJ
|
|
|
84
|
|
|
85 %arrayidxG = getelementptr inbounds i32, i32* %g, i64 %ind
|
|
|
86 store i32 %mulG, i32* %arrayidxG, align 4
|
|
|
87 br label %if.end
|
|
|
88
|
|
|
89 if.end:
|
|
|
90 %exitcond = icmp eq i64 %add, 20
|
|
|
91 br i1 %exitcond, label %for.end, label %for.body
|
|
|
92
|
|
|
93 for.end: ; preds = %for.body
|
|
|
94 ret void
|
|
|
95 }
|
