1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
2; RUN: opt -loop-unroll -simplifycfg -instcombine -simplifycfg -S -mtriple arm-none-eabi -mcpu=cortex-m7 %s | FileCheck %s
3
4; This test is meant to check that this loop is unrolled into three iterations.
5define void @test(i32* %x, i32 %n) {
6; CHECK-LABEL: @test(
7; CHECK-NEXT:  entry:
8; CHECK-NEXT:    [[SUB:%.*]] = add nsw i32 [[N:%.*]], -1
9; CHECK-NEXT:    [[REM:%.*]] = srem i32 [[SUB]], 4
10; CHECK-NEXT:    [[CMP7:%.*]] = icmp sgt i32 [[REM]], 0
11; CHECK-NEXT:    br i1 [[CMP7]], label [[WHILE_BODY:%.*]], label [[WHILE_END:%.*]]
12; CHECK:       while.body:
13; CHECK-NEXT:    [[TMP0:%.*]] = load i32, i32* [[X:%.*]], align 4
14; CHECK-NEXT:    [[CMP1:%.*]] = icmp slt i32 [[TMP0]], 10
15; CHECK-NEXT:    br i1 [[CMP1]], label [[IF_THEN:%.*]], label [[IF_END:%.*]]
16; CHECK:       if.then:
17; CHECK-NEXT:    store i32 0, i32* [[X]], align 4
18; CHECK-NEXT:    br label [[IF_END]]
19; CHECK:       if.end:
20; CHECK-NEXT:    [[INCDEC_PTR:%.*]] = getelementptr inbounds i32, i32* [[X]], i64 1
21; CHECK-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[REM]], 1
22; CHECK-NEXT:    br i1 [[CMP]], label [[WHILE_BODY_1:%.*]], label [[WHILE_END]]
23; CHECK:       while.end:
24; CHECK-NEXT:    ret void
25; CHECK:       while.body.1:
26; CHECK-NEXT:    [[TMP1:%.*]] = load i32, i32* [[INCDEC_PTR]], align 4
27; CHECK-NEXT:    [[CMP1_1:%.*]] = icmp slt i32 [[TMP1]], 10
28; CHECK-NEXT:    br i1 [[CMP1_1]], label [[IF_THEN_1:%.*]], label [[IF_END_1:%.*]]
29; CHECK:       if.then.1:
30; CHECK-NEXT:    store i32 0, i32* [[INCDEC_PTR]], align 4
31; CHECK-NEXT:    br label [[IF_END_1]]
32; CHECK:       if.end.1:
33; CHECK-NEXT:    [[INCDEC_PTR_1:%.*]] = getelementptr inbounds i32, i32* [[X]], i64 2
34; CHECK-NEXT:    [[CMP_1:%.*]] = icmp sgt i32 [[REM]], 2
35; CHECK-NEXT:    br i1 [[CMP_1]], label [[WHILE_BODY_2:%.*]], label [[WHILE_END]]
36; CHECK:       while.body.2:
37; CHECK-NEXT:    [[TMP2:%.*]] = load i32, i32* [[INCDEC_PTR_1]], align 4
38; CHECK-NEXT:    [[CMP1_2:%.*]] = icmp slt i32 [[TMP2]], 10
39; CHECK-NEXT:    br i1 [[CMP1_2]], label [[IF_THEN_2:%.*]], label [[WHILE_END]]
40; CHECK:       if.then.2:
41; CHECK-NEXT:    store i32 0, i32* [[INCDEC_PTR_1]], align 4
42; CHECK-NEXT:    br label [[WHILE_END]]
43;
44entry:
45  %sub = add nsw i32 %n, -1
46  %rem = srem i32 %sub, 4
47  %cmp7 = icmp sgt i32 %rem, 0
48  br i1 %cmp7, label %while.body, label %while.end
49
50while.body:                                       ; preds = %entry, %if.end
51  %x.addr.09 = phi i32* [ %incdec.ptr, %if.end ], [ %x, %entry ]
52  %n.addr.08 = phi i32 [ %dec, %if.end ], [ %rem, %entry ]
53  %0 = load i32, i32* %x.addr.09, align 4
54  %cmp1 = icmp slt i32 %0, 10
55  br i1 %cmp1, label %if.then, label %if.end
56
57if.then:                                          ; preds = %while.body
58  store i32 0, i32* %x.addr.09, align 4
59  br label %if.end
60
61if.end:                                           ; preds = %if.then, %while.body
62  %incdec.ptr = getelementptr inbounds i32, i32* %x.addr.09, i32 1
63  %dec = add nsw i32 %n.addr.08, -1
64  %cmp = icmp sgt i32 %dec, 0
65  br i1 %cmp, label %while.body, label %while.end
66
67while.end:                                        ; preds = %if.end, %entry
68  ret void
69}
70
71; Larger test that is still fully unrolled, thanks in part to the constant data and the max upper bound.
72@data = internal constant [50 x i32] [i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50], align 4
73define i32 @test2(i32 %l86) {
74; CHECK-LABEL: @test2(
75; CHECK-NEXT:  entry:
76; CHECK-NEXT:    br label [[FOR_BODY_I_I:%.*]]
77; CHECK:       for.body.i.i:
78; CHECK-NEXT:    [[I_0137_I_I:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[ADD_I_3_I:%.*]], [[FOR_BODY_I_3_I:%.*]] ]
79; CHECK-NEXT:    [[ADD_I_I:%.*]] = or i32 [[I_0137_I_I]], 1
80; CHECK-NEXT:    [[TMP0:%.*]] = zext i32 [[ADD_I_I]] to i64
81; CHECK-NEXT:    [[ARRAYIDX_I_I:%.*]] = getelementptr inbounds [50 x i32], [50 x i32]* @data, i64 0, i64 [[TMP0]]
82; CHECK-NEXT:    [[L93:%.*]] = load i32, i32* [[ARRAYIDX_I_I]], align 4
83; CHECK-NEXT:    [[CMP1_I_I:%.*]] = icmp sgt i32 [[L93]], [[L86:%.*]]
84; CHECK-NEXT:    br i1 [[CMP1_I_I]], label [[LAND_LHS_TRUE_I_I:%.*]], label [[FOR_INC_I_I:%.*]]
85; CHECK:       land.lhs.true.i.i:
86; CHECK-NEXT:    [[TMP1:%.*]] = zext i32 [[I_0137_I_I]] to i64
87; CHECK-NEXT:    [[ARRAYIDX2_I_I:%.*]] = getelementptr inbounds [50 x i32], [50 x i32]* @data, i64 0, i64 [[TMP1]]
88; CHECK-NEXT:    [[L94:%.*]] = load i32, i32* [[ARRAYIDX2_I_I]], align 4
89; CHECK-NEXT:    [[CMP3_NOT_I_I:%.*]] = icmp sgt i32 [[L94]], [[L86]]
90; CHECK-NEXT:    br i1 [[CMP3_NOT_I_I]], label [[FOR_INC_I_I]], label [[FOR_END_I_IF_END8_I_CRIT_EDGE_I:%.*]]
91; CHECK:       for.inc.i.i:
92; CHECK-NEXT:    [[EXITCOND_NOT_I_I:%.*]] = icmp eq i32 [[ADD_I_I]], 25
93; CHECK-NEXT:    br i1 [[EXITCOND_NOT_I_I]], label [[FOR_END_I_IF_END8_I_CRIT_EDGE_I]], label [[FOR_BODY_I_1_I:%.*]]
94; CHECK:       for.body.i.1.i:
95; CHECK-NEXT:    [[ADD_I_1_I:%.*]] = or i32 [[I_0137_I_I]], 2
96; CHECK-NEXT:    [[TMP2:%.*]] = zext i32 [[ADD_I_1_I]] to i64
97; CHECK-NEXT:    [[ARRAYIDX_I_1_I:%.*]] = getelementptr inbounds [50 x i32], [50 x i32]* @data, i64 0, i64 [[TMP2]]
98; CHECK-NEXT:    [[L345:%.*]] = load i32, i32* [[ARRAYIDX_I_1_I]], align 4
99; CHECK-NEXT:    [[CMP1_I_1_I:%.*]] = icmp sgt i32 [[L345]], [[L86]]
100; CHECK-NEXT:    [[CMP1_I_1_I_NOT:%.*]] = xor i1 [[CMP1_I_1_I]], true
101; CHECK-NEXT:    [[BRMERGE:%.*]] = or i1 [[CMP1_I_I]], [[CMP1_I_1_I_NOT]]
102; CHECK-NEXT:    br i1 [[BRMERGE]], label [[FOR_INC_I_1_I:%.*]], label [[FOR_END_I_I:%.*]]
103; CHECK:       for.inc.i.1.i:
104; CHECK-NEXT:    [[ADD_I_2_I:%.*]] = or i32 [[I_0137_I_I]], 3
105; CHECK-NEXT:    [[TMP3:%.*]] = zext i32 [[ADD_I_2_I]] to i64
106; CHECK-NEXT:    [[ARRAYIDX_I_2_I:%.*]] = getelementptr inbounds [50 x i32], [50 x i32]* @data, i64 0, i64 [[TMP3]]
107; CHECK-NEXT:    [[L346:%.*]] = load i32, i32* [[ARRAYIDX_I_2_I]], align 4
108; CHECK-NEXT:    [[CMP1_I_2_I:%.*]] = icmp sgt i32 [[L346]], [[L86]]
109; CHECK-NEXT:    [[CMP1_I_2_I_NOT:%.*]] = xor i1 [[CMP1_I_2_I]], true
110; CHECK-NEXT:    [[BRMERGE1:%.*]] = or i1 [[CMP1_I_1_I]], [[CMP1_I_2_I_NOT]]
111; CHECK-NEXT:    br i1 [[BRMERGE1]], label [[FOR_BODY_I_3_I]], label [[FOR_END_I_IF_END8_I_CRIT_EDGE_I]]
112; CHECK:       for.body.i.3.i:
113; CHECK-NEXT:    [[ADD_I_3_I]] = add nuw nsw i32 [[I_0137_I_I]], 4
114; CHECK-NEXT:    [[TMP4:%.*]] = zext i32 [[ADD_I_3_I]] to i64
115; CHECK-NEXT:    [[ARRAYIDX_I_3_I:%.*]] = getelementptr inbounds [50 x i32], [50 x i32]* @data, i64 0, i64 [[TMP4]]
116; CHECK-NEXT:    [[L347:%.*]] = load i32, i32* [[ARRAYIDX_I_3_I]], align 4
117; CHECK-NEXT:    [[CMP1_I_3_I:%.*]] = icmp sle i32 [[L347]], [[L86]]
118; CHECK-NEXT:    [[BRMERGE2:%.*]] = or i1 [[CMP1_I_3_I]], [[CMP1_I_2_I]]
119; CHECK-NEXT:    br i1 [[BRMERGE2]], label [[FOR_BODY_I_I]], label [[FOR_END_I_I]]
120; CHECK:       for.end.i.i:
121; CHECK-NEXT:    [[I_0_LCSSA_I_I:%.*]] = phi i32 [ [[ADD_I_I]], [[FOR_BODY_I_1_I]] ], [ [[ADD_I_2_I]], [[FOR_BODY_I_3_I]] ]
122; CHECK-NEXT:    [[CMP5_I_I:%.*]] = icmp eq i32 [[I_0_LCSSA_I_I]], 25
123; CHECK-NEXT:    [[SPEC_SELECT:%.*]] = select i1 [[CMP5_I_I]], i32 2, i32 0
124; CHECK-NEXT:    br label [[FOR_END_I_IF_END8_I_CRIT_EDGE_I]]
125; CHECK:       for.end.i.if.end8.i_crit_edge.i:
126; CHECK-NEXT:    [[MERGE:%.*]] = phi i32 [ 0, [[FOR_INC_I_1_I]] ], [ 0, [[LAND_LHS_TRUE_I_I]] ], [ 1, [[FOR_INC_I_I]] ], [ [[SPEC_SELECT]], [[FOR_END_I_I]] ]
127; CHECK-NEXT:    ret i32 [[MERGE]]
128;
129entry:
130  br label %for.body.i.i
131
132for.body.i.i:                                     ; preds = %for.body.i.i.preheader, %for.inc.i.3.i
133  %i.0137.i.i = phi i32 [ %add.i.3.i, %for.inc.i.3.i ], [ 0, %entry ]
134  %add.i.i = or i32 %i.0137.i.i, 1
135  %arrayidx.i.i = getelementptr inbounds i32, i32* getelementptr inbounds ([50 x i32], [50 x i32]* @data, i32 0, i32 0), i32 %add.i.i
136  %l93 = load i32, i32* %arrayidx.i.i, align 4
137  %cmp1.i.i = icmp sgt i32 %l93, %l86
138  br i1 %cmp1.i.i, label %land.lhs.true.i.i, label %for.inc.i.i
139
140land.lhs.true.i.i:                                ; preds = %for.body.i.i
141  %arrayidx2.i.i = getelementptr inbounds i32, i32* getelementptr inbounds ([50 x i32], [50 x i32]* @data, i32 0, i32 0), i32 %i.0137.i.i
142  %l94 = load i32, i32* %arrayidx2.i.i, align 4
143  %cmp3.not.i.i = icmp sgt i32 %l94, %l86
144  br i1 %cmp3.not.i.i, label %for.inc.i.i, label %for.end.i.if.end8.i_crit_edge.i.loopexit
145
146for.inc.i.i:                                      ; preds = %land.lhs.true.i.i, %for.body.i.i
147  %exitcond.not.i.i = icmp eq i32 %add.i.i, 25
148  br i1 %exitcond.not.i.i, label %if.then6.i.i.loopexitsplit, label %for.body.i.1.i
149
150for.body.i.1.i:                                   ; preds = %for.inc.i.i
151  %add.i.1.i = or i32 %i.0137.i.i, 2
152  %arrayidx.i.1.i = getelementptr inbounds i32, i32* getelementptr inbounds ([50 x i32], [50 x i32]* @data, i32 0, i32 0), i32 %add.i.1.i
153  %l345 = load i32, i32* %arrayidx.i.1.i, align 4
154  %cmp1.i.1.i = icmp sgt i32 %l345, %l86
155  br i1 %cmp1.i.1.i, label %land.lhs.true.i.1.i, label %for.inc.i.1.i
156
157land.lhs.true.i.1.i:                              ; preds = %for.body.i.1.i
158  br i1 %cmp1.i.i, label %for.inc.i.1.i, label %for.end.i.i
159
160for.inc.i.1.i:                                    ; preds = %land.lhs.true.i.1.i, %for.body.i.1.i
161  %add.i.2.i = or i32 %i.0137.i.i, 3
162  %arrayidx.i.2.i = getelementptr inbounds i32, i32* getelementptr inbounds ([50 x i32], [50 x i32]* @data, i32 0, i32 0), i32 %add.i.2.i
163  %l346 = load i32, i32* %arrayidx.i.2.i, align 4
164  %cmp1.i.2.i = icmp sgt i32 %l346, %l86
165  br i1 %cmp1.i.2.i, label %land.lhs.true.i.2.i, label %for.inc.i.2.i
166
167land.lhs.true.i.2.i:                              ; preds = %for.inc.i.1.i
168  br i1 %cmp1.i.1.i, label %for.inc.i.2.i, label %for.end.i.if.end8.i_crit_edge.i.loopexit
169
170for.inc.i.2.i:                                    ; preds = %land.lhs.true.i.2.i, %for.inc.i.1.i
171  br label %for.body.i.3.i
172
173for.body.i.3.i:                                   ; preds = %for.inc.i.2.i
174  %add.i.3.i = add nuw nsw i32 %i.0137.i.i, 4
175  %arrayidx.i.3.i = getelementptr inbounds i32, i32* getelementptr inbounds ([50 x i32], [50 x i32]* @data, i32 0, i32 0), i32 %add.i.3.i
176  %l347 = load i32, i32* %arrayidx.i.3.i, align 4
177  %cmp1.i.3.i = icmp sgt i32 %l347, %l86
178  br i1 %cmp1.i.3.i, label %land.lhs.true.i.3.i, label %for.inc.i.3.i
179
180land.lhs.true.i.3.i:                              ; preds = %for.body.i.3.i
181  br i1 %cmp1.i.2.i, label %for.inc.i.3.i, label %for.end.i.i
182
183for.inc.i.3.i:                                    ; preds = %land.lhs.true.i.3.i, %for.body.i.3.i
184  br label %for.body.i.i
185
186for.end.i.if.end8.i_crit_edge.i.loopexit:         ; preds = %land.lhs.true.i.i, %land.lhs.true.i.2.i
187  %i.0.lcssa.i.i3.ph = phi i32 [ %i.0137.i.i, %land.lhs.true.i.i ], [ %add.i.1.i, %land.lhs.true.i.2.i ]
188  br label %for.end.i.if.end8.i_crit_edge.i
189
190if.then6.i.i.loopexitsplit:                       ; preds = %for.inc.i.i
191  br label %if.then6.i.i.loopexit
192
193for.end.i.i:                                      ; preds = %land.lhs.true.i.3.i, %land.lhs.true.i.1.i
194  %i.0.lcssa.i.i = phi i32 [ %add.i.i, %land.lhs.true.i.1.i ], [ %add.i.2.i, %land.lhs.true.i.3.i ]
195  %cmp5.i.i = icmp eq i32 %i.0.lcssa.i.i, 25
196  br i1 %cmp5.i.i, label %if.then6.i.i, label %for.end.i.if.end8.i_crit_edge.i
197
198for.end.i.if.end8.i_crit_edge.i:
199  ret i32 0
200
201if.then6.i.i.loopexit:
202  ret i32 1
203
204if.then6.i.i:
205  ret i32 2
206}
207