• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1; RUN: opt < %s -loop-interchange -pass-remarks-output=%t -verify-dom-info -verify-loop-info \
2; RUN:     -pass-remarks=loop-interchange -pass-remarks-missed=loop-interchange
3; RUN: FileCheck -input-file %t %s
4
5;; We test profitability model in these test cases.
6
7target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
8target triple = "x86_64-unknown-linux-gnu"
9
10@A = common global [100 x [100 x i32]] zeroinitializer
11@B = common global [100 x [100 x i32]] zeroinitializer
12
13;;---------------------------------------Test case 01---------------------------------
14;; Loops interchange will result in code vectorization and hence profitable. Check for interchange.
15;;   for(int i=1;i<100;i++)
16;;     for(int j=1;j<100;j++)
17;;       A[j][i] = A[j - 1][i] + B[j][i];
18;; FIXME: DA misses this case after D35430
19
20; CHECK:      Name:            Dependence
21; CHECK-NEXT: Function:        interchange_01
22define void @interchange_01() {
23entry:
24  br label %for2.preheader
25
26for2.preheader:
27  %i30 = phi i64 [ 1, %entry ], [ %i.next31, %for1.inc14 ]
28  br label %for2
29
30for2:
31  %j = phi i64 [ %i.next, %for2 ], [ 1, %for2.preheader ]
32  %j.prev = add nsw i64 %j,  -1
33  %arrayidx5 = getelementptr inbounds [100 x [100 x i32]], [100 x [100 x i32]]* @A, i64 0, i64 %j.prev, i64 %i30
34  %lv1 = load i32, i32* %arrayidx5
35  %arrayidx9 = getelementptr inbounds [100 x [100 x i32]], [100 x [100 x i32]]* @B, i64 0, i64 %j,  i64 %i30
36  %lv2 = load i32, i32* %arrayidx9
37  %add = add nsw i32 %lv1, %lv2
38  %arrayidx13 = getelementptr inbounds [100 x [100 x i32]], [100 x [100 x i32]]* @A, i64 0, i64 %j,  i64 %i30
39  store i32 %add, i32* %arrayidx13
40  %i.next = add nuw nsw i64 %j,  1
41  %exitcond = icmp eq i64 %j,  99
42  br i1 %exitcond, label %for1.inc14, label %for2
43
44for1.inc14:
45  %i.next31 = add nuw nsw i64 %i30, 1
46  %exitcond33 = icmp eq i64 %i30, 99
47  br i1 %exitcond33, label %for.end16, label %for2.preheader
48
49for.end16:
50  ret void
51}
52
53;; ---------------------------------------Test case 02---------------------------------
54;; Check loop interchange profitability model.
55;; This tests profitability model when operands of getelementpointer and not exactly the induction variable but some
56;; arithmetic operation on them.
57;;   for(int i=1;i<N;i++)
58;;    for(int j=1;j<N;j++)
59;;       A[j-1][i-1] = A[j - 1][i-1] + B[j-1][i-1];
60
61; CHECK:      Name:            Interchanged
62; CHECK-NEXT: Function:        interchange_02
63define void @interchange_02() {
64entry:
65  br label %for1.header
66
67for1.header:
68  %i35 = phi i64 [ 1, %entry ], [ %i.next36, %for1.inc19 ]
69  %i.prev = add nsw i64 %i35, -1
70  br label %for2
71
72for2:
73  %j = phi i64 [ 1, %for1.header ], [ %i.next, %for2 ]
74  %j.prev = add nsw i64 %j,  -1
75  %arrayidx6 = getelementptr inbounds [100 x [100 x i32]], [100 x [100 x i32]]* @A, i64 0, i64 %j.prev, i64 %i.prev
76  %lv1 = load i32, i32* %arrayidx6
77  %arrayidx12 = getelementptr inbounds [100 x [100 x i32]], [100 x [100 x i32]]* @B, i64 0, i64 %j.prev, i64 %i.prev
78  %lv2 = load i32, i32* %arrayidx12
79  %add = add nsw i32 %lv1, %lv2
80  store i32 %add, i32* %arrayidx6
81  %i.next = add nuw nsw i64 %j,  1
82  %exitcond = icmp eq i64 %j,  99
83  br i1 %exitcond, label %for1.inc19, label %for2
84
85for1.inc19:
86  %i.next36 = add nuw nsw i64 %i35, 1
87  %exitcond39 = icmp eq i64 %i35, 99
88  br i1 %exitcond39, label %for.end21, label %for1.header
89
90for.end21:
91  ret void
92}
93
94;;---------------------------------------Test case 03---------------------------------
95;; Loops interchange is not profitable.
96;;   for(int i=1;i<100;i++)
97;;     for(int j=1;j<100;j++)
98;;       A[i-1][j-1] = A[i - 1][j-1] + B[i][j];
99
100; CHECK:      Name:            InterchangeNotProfitable
101; CHECK-NEXT: Function:        interchange_03
102define void @interchange_03(){
103entry:
104  br label %for1.header
105
106for1.header:
107  %i34 = phi i64 [ 1, %entry ], [ %i.next35, %for1.inc17 ]
108  %i.prev = add nsw i64 %i34, -1
109  br label %for2
110
111for2:
112  %j = phi i64 [ 1, %for1.header ], [ %i.next, %for2 ]
113  %j.prev = add nsw i64 %j, -1
114  %arrayidx6 = getelementptr inbounds [100 x [100 x i32]], [100 x [100 x i32]]* @A, i64 0, i64 %i.prev, i64 %j.prev
115  %lv1 = load i32, i32* %arrayidx6
116  %arrayidx10 = getelementptr inbounds [100 x [100 x i32]], [100 x [100 x i32]]* @B, i64 0, i64 %i34, i64 %j
117  %lv2 = load i32, i32* %arrayidx10
118  %add = add nsw i32 %lv1, %lv2
119  store i32 %add, i32* %arrayidx6
120  %i.next = add nuw nsw i64 %j,  1
121  %exitcond = icmp eq i64 %j,  99
122  br i1 %exitcond, label %for1.inc17, label %for2
123
124for1.inc17:
125  %i.next35 = add nuw nsw i64 %i34, 1
126  %exitcond38 = icmp eq i64 %i34, 99
127  br i1 %exitcond38, label %for.end19, label %for1.header
128
129for.end19:
130  ret void
131}
132
133;; Loops should not be interchanged in this case as it is not profitable.
134;;  for(int i=0;i<100;i++)
135;;    for(int j=0;j<100;j++)
136;;      A[i][j] = A[i][j]+k;
137
138; CHECK:      Name:            InterchangeNotProfitable
139; CHECK-NEXT: Function:        interchange_04
140define void @interchange_04(i32 %k) {
141entry:
142  br label %for.cond1.preheader
143
144for.cond1.preheader:
145  %indvars.iv21 = phi i64 [ 0, %entry ], [ %indvars.iv.next22, %for.inc10 ]
146  br label %for.body3
147
148for.body3:
149  %indvars.iv = phi i64 [ 0, %for.cond1.preheader ], [ %indvars.iv.next, %for.body3 ]
150  %arrayidx5 = getelementptr inbounds [100 x [100 x i32]], [100 x [100 x i32]]* @A, i64 0, i64 %indvars.iv21, i64 %indvars.iv
151  %0 = load i32, i32* %arrayidx5
152  %add = add nsw i32 %0, %k
153  store i32 %add, i32* %arrayidx5
154  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
155  %exitcond = icmp eq i64 %indvars.iv.next, 100
156  br i1 %exitcond, label %for.inc10, label %for.body3
157
158for.inc10:
159  %indvars.iv.next22 = add nuw nsw i64 %indvars.iv21, 1
160  %exitcond23 = icmp eq i64 %indvars.iv.next22, 100
161  br i1 %exitcond23, label %for.end12, label %for.cond1.preheader
162
163for.end12:
164  ret void
165}
166