1; RUN: opt < %s -S -loop-unroll -unroll-count=4 | FileCheck -check-prefix=CHECK_COUNT4 %s
2; RUN: opt < %s -S -loop-unroll | FileCheck -check-prefix=CHECK_NOCOUNT %s
3; RUN: opt < %s -S -passes='require<profile-summary>,function(loop-unroll)' -pgso | FileCheck -check-prefix=PGSO %s
4; RUN: opt < %s -S -passes='require<profile-summary>,function(loop-unroll)' -pgso=false | FileCheck -check-prefix=NPGSO %s
5
6
7;///////////////////// TEST 1 //////////////////////////////
8
9; This test shows that the loop is unrolled according to the specified
10; unroll factor.
11
12define void @Test1() nounwind {
13entry:
14  br label %loop
15
16loop:
17  %iv = phi i32 [ 0, %entry ], [ %inc, %loop ]
18  %inc = add i32 %iv, 1
19  %exitcnd = icmp uge i32 %inc, 1024
20  br i1 %exitcnd, label %exit, label %loop
21
22exit:
23  ret void
24}
25
26; CHECK_COUNT4-LABEL: @Test1
27; CHECK_COUNT4:      phi
28; CHECK_COUNT4-NEXT: add
29; CHECK_COUNT4-NEXT: add
30; CHECK_COUNT4-NEXT: add
31; CHECK_COUNT4-NEXT: add
32; CHECK_COUNT4-NEXT: icmp
33
34
35;///////////////////// TEST 2 //////////////////////////////
36
37; This test shows that with optnone attribute, the loop is not unrolled
38; even if an unroll factor was specified.
39
40define void @Test2() nounwind optnone noinline {
41entry:
42  br label %loop
43
44loop:
45  %iv = phi i32 [ 0, %entry ], [ %inc, %loop ]
46  %inc = add i32 %iv, 1
47  %exitcnd = icmp uge i32 %inc, 1024
48  br i1 %exitcnd, label %exit, label %loop
49
50exit:
51  ret void
52}
53
54; CHECK_COUNT4-LABEL: @Test2
55; CHECK_COUNT4:      phi
56; CHECK_COUNT4-NEXT: add
57; CHECK_COUNT4-NEXT: icmp
58
59
60;///////////////////// TEST 3 //////////////////////////////
61
62; This test shows that this loop is fully unrolled by default.
63
64@tab = common global [24 x i32] zeroinitializer, align 4
65
66define i32 @Test3() {
67entry:
68  br label %for.body
69
70for.body:                                         ; preds = %for.body, %entry
71  %i.05 = phi i32 [ 0, %entry ], [ %inc, %for.body ]
72  %arrayidx = getelementptr inbounds [24 x i32], [24 x i32]* @tab, i32 0, i32 %i.05
73  store i32 %i.05, i32* %arrayidx, align 4
74  %inc = add nuw nsw i32 %i.05, 1
75  %exitcond = icmp eq i32 %inc, 24
76  br i1 %exitcond, label %for.end, label %for.body
77
78for.end:                                          ; preds = %for.body
79  ret i32 42
80}
81
82; CHECK_NOCOUNT-LABEL: @Test3
83; CHECK_NOCOUNT:      store
84; CHECK_NOCOUNT-NEXT: store
85; CHECK_NOCOUNT-NEXT: store
86; CHECK_NOCOUNT-NEXT: store
87; CHECK_NOCOUNT-NEXT: store
88; CHECK_NOCOUNT-NEXT: store
89; CHECK_NOCOUNT-NEXT: store
90; CHECK_NOCOUNT-NEXT: store
91; CHECK_NOCOUNT-NEXT: store
92; CHECK_NOCOUNT-NEXT: store
93; CHECK_NOCOUNT-NEXT: store
94; CHECK_NOCOUNT-NEXT: store
95; CHECK_NOCOUNT-NEXT: store
96; CHECK_NOCOUNT-NEXT: store
97; CHECK_NOCOUNT-NEXT: store
98; CHECK_NOCOUNT-NEXT: store
99; CHECK_NOCOUNT-NEXT: store
100; CHECK_NOCOUNT-NEXT: store
101; CHECK_NOCOUNT-NEXT: store
102; CHECK_NOCOUNT-NEXT: store
103; CHECK_NOCOUNT-NEXT: store
104; CHECK_NOCOUNT-NEXT: store
105; CHECK_NOCOUNT-NEXT: store
106; CHECK_NOCOUNT-NEXT: store
107; CHECK_NOCOUNT-NEXT: ret
108
109
110;///////////////////// TEST 4 //////////////////////////////
111
112; This test shows that with optsize attribute, this loop is not unrolled.
113
114define i32 @Test4() optsize {
115entry:
116  br label %for.body
117
118for.body:                                         ; preds = %for.body, %entry
119  %i.05 = phi i32 [ 0, %entry ], [ %inc, %for.body ]
120  %arrayidx = getelementptr inbounds [24 x i32], [24 x i32]* @tab, i32 0, i32 %i.05
121  store i32 %i.05, i32* %arrayidx, align 4
122  %inc = add nuw nsw i32 %i.05, 1
123  %exitcond = icmp eq i32 %inc, 24
124  br i1 %exitcond, label %for.end, label %for.body
125
126for.end:                                          ; preds = %for.body
127  ret i32 42
128}
129
130; CHECK_NOCOUNT-LABEL: @Test4
131; CHECK_NOCOUNT:      phi
132; CHECK_NOCOUNT:      icmp
133
134;///////////////////// TEST 5 //////////////////////////////
135
136; This test shows that with PGO, this loop is cold and not unrolled.
137
138define i32 @Test5() !prof !14 {
139entry:
140  br label %for.body
141
142for.body:                                         ; preds = %for.body, %entry
143  %i.05 = phi i32 [ 0, %entry ], [ %inc, %for.body ]
144  %arrayidx = getelementptr inbounds [24 x i32], [24 x i32]* @tab, i32 0, i32 %i.05
145  store i32 %i.05, i32* %arrayidx, align 4
146  %inc = add nuw nsw i32 %i.05, 1
147  %exitcond = icmp eq i32 %inc, 24
148  br i1 %exitcond, label %for.end, label %for.body
149
150for.end:                                          ; preds = %for.body
151  ret i32 42
152}
153
154; PGSO-LABEL: @Test5
155; PGSO:      phi
156; PGSO:      icmp
157; NPGSO-LABEL: @Test5
158; NPGSO-NOT:      phi
159; NPGSO-NOT:      icmp
160
161;///////////////////// TEST 6 //////////////////////////////
162
163; This test tests that unroll hints take precedence over PGSO and that this loop
164; gets unrolled even though it's cold.
165
166define i32 @Test6() !prof !14 {
167entry:
168  br label %for.body
169
170for.body:                                         ; preds = %for.body, %entry
171  %i.05 = phi i32 [ 0, %entry ], [ %inc, %for.body ]
172  %arrayidx = getelementptr inbounds [24 x i32], [24 x i32]* @tab, i32 0, i32 %i.05
173  store i32 %i.05, i32* %arrayidx, align 4
174  %inc = add nuw nsw i32 %i.05, 1
175  %exitcond = icmp eq i32 %inc, 24
176  br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !15
177
178for.end:                                          ; preds = %for.body
179  ret i32 42
180}
181
182; PGSO-LABEL: @Test6
183; PGSO:      store
184; PGSO:      store
185; PGSO:      store
186; PGSO:      store
187; NPGSO-LABEL: @Test6
188; NPGSO:      store
189; NPGSO:      store
190; NPGSO:      store
191; NPGSO:      store
192
193!llvm.module.flags = !{!0}
194!0 = !{i32 1, !"ProfileSummary", !1}
195!1 = !{!2, !3, !4, !5, !6, !7, !8, !9}
196!2 = !{!"ProfileFormat", !"InstrProf"}
197!3 = !{!"TotalCount", i64 10000}
198!4 = !{!"MaxCount", i64 10}
199!5 = !{!"MaxInternalCount", i64 1}
200!6 = !{!"MaxFunctionCount", i64 1000}
201!7 = !{!"NumCounts", i64 3}
202!8 = !{!"NumFunctions", i64 3}
203!9 = !{!"DetailedSummary", !10}
204!10 = !{!11, !12, !13}
205!11 = !{i32 10000, i64 100, i32 1}
206!12 = !{i32 999000, i64 100, i32 1}
207!13 = !{i32 999999, i64 1, i32 2}
208!14 = !{!"function_entry_count", i64 0}
209!15 = !{!15, !16}
210!16 = !{!"llvm.loop.unroll.count", i32 4}
211