1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
2; RUN: opt -passes='loop-vectorize' -force-vector-width=2 -force-vector-interleave=1 -S %s | FileCheck %s
3
4define void @test_pr55100(i32 %N) {
5; CHECK-LABEL: @test_pr55100(
6; CHECK-NEXT:  entry:
7; CHECK-NEXT:    [[TMP0:%.*]] = sub i32 0, [[N:%.*]]
8; CHECK-NEXT:    br label [[LOOP_1_HEADER:%.*]]
9; CHECK:       loop.1.header:
10; CHECK-NEXT:    [[IV_1:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[IV_1_NEXT:%.*]], [[LOOP_1_LATCH:%.*]] ]
11; CHECK-NEXT:    [[TMP1:%.*]] = mul nuw nsw i32 [[IV_1]], -1
12; CHECK-NEXT:    [[TMP2:%.*]] = add i32 [[TMP0]], [[TMP1]]
13; CHECK-NEXT:    [[UMIN2:%.*]] = call i32 @llvm.umin.i32(i32 [[TMP2]], i32 18)
14; CHECK-NEXT:    [[TMP3:%.*]] = add nuw nsw i32 [[UMIN2]], 1
15; CHECK-NEXT:    [[TMP4:%.*]] = mul nuw nsw i32 [[IV_1]], -1
16; CHECK-NEXT:    [[TMP5:%.*]] = add i32 [[TMP0]], [[TMP4]]
17; CHECK-NEXT:    [[UMIN:%.*]] = call i32 @llvm.umin.i32(i32 [[TMP5]], i32 18)
18; CHECK-NEXT:    [[C_2:%.*]] = icmp ugt i32 [[IV_1]], 10
19; CHECK-NEXT:    br i1 [[C_2]], label [[LOOP_2_HEADER_PREHEADER:%.*]], label [[EXIT_LOOPEXIT1:%.*]]
20; CHECK:       loop.2.header.preheader:
21; CHECK-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ule i32 [[TMP3]], 2
22; CHECK-NEXT:    br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
23; CHECK:       vector.ph:
24; CHECK-NEXT:    [[N_MOD_VF:%.*]] = urem i32 [[TMP3]], 2
25; CHECK-NEXT:    [[TMP6:%.*]] = icmp eq i32 [[N_MOD_VF]], 0
26; CHECK-NEXT:    [[TMP7:%.*]] = select i1 [[TMP6]], i32 2, i32 [[N_MOD_VF]]
27; CHECK-NEXT:    [[N_VEC:%.*]] = sub i32 [[TMP3]], [[TMP7]]
28; CHECK-NEXT:    [[IND_END:%.*]] = trunc i32 [[N_VEC]] to i16
29; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
30; CHECK:       vector.body:
31; CHECK-NEXT:    [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
32; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 2
33; CHECK-NEXT:    [[TMP8:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]]
34; CHECK-NEXT:    br i1 [[TMP8]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
35; CHECK:       middle.block:
36; CHECK-NEXT:    br label [[SCALAR_PH]]
37; CHECK:       scalar.ph:
38; CHECK-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i16 [ [[IND_END]], [[MIDDLE_BLOCK]] ], [ 0, [[LOOP_2_HEADER_PREHEADER]] ]
39; CHECK-NEXT:    br label [[LOOP_2_HEADER:%.*]]
40; CHECK:       vector.scevcheck:
41; CHECK-NEXT:    [[TMP9:%.*]] = trunc i32 [[UMIN]] to i16
42; CHECK-NEXT:    unreachable
43; CHECK:       loop.2.header:
44; CHECK-NEXT:    [[IV_2:%.*]] = phi i16 [ [[IV_2_NEXT:%.*]], [[LOOP_2_LATCH:%.*]] ], [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ]
45; CHECK-NEXT:    [[C_3:%.*]] = icmp slt i16 [[IV_2]], 18
46; CHECK-NEXT:    br i1 [[C_3]], label [[LOOP_2_LATCH]], label [[LOOP_1_LATCH]]
47; CHECK:       loop.2.latch:
48; CHECK-NEXT:    [[ADD_1:%.*]] = add i32 [[N]], [[IV_1]]
49; CHECK-NEXT:    [[IV_2_EXT:%.*]] = sext i16 [[IV_2]] to i32
50; CHECK-NEXT:    [[ADD_2:%.*]] = add i32 [[ADD_1]], [[IV_2_EXT]]
51; CHECK-NEXT:    [[C_4:%.*]] = icmp ult i32 [[ADD_2]], 1
52; CHECK-NEXT:    [[IV_2_NEXT]] = add i16 [[IV_2]], 1
53; CHECK-NEXT:    br i1 [[C_4]], label [[EXIT_LOOPEXIT:%.*]], label [[LOOP_2_HEADER]], !llvm.loop [[LOOP2:![0-9]+]]
54; CHECK:       loop.1.latch:
55; CHECK-NEXT:    [[IV_1_NEXT]] = add i32 [[IV_1]], 1
56; CHECK-NEXT:    br label [[LOOP_1_HEADER]]
57; CHECK:       exit.loopexit:
58; CHECK-NEXT:    br label [[EXIT:%.*]]
59; CHECK:       exit.loopexit1:
60; CHECK-NEXT:    br label [[EXIT]]
61; CHECK:       exit:
62; CHECK-NEXT:    ret void
63;
64entry:
65  br label %loop.1.header
66
67loop.1.header:
68  %iv.1 = phi i32 [ 0, %entry ], [ %iv.1.next, %loop.1.latch ]
69  %c.2 = icmp ugt i32 %iv.1, 10
70  br i1 %c.2, label %loop.2.header, label %exit
71
72loop.2.header:
73  %iv.2 = phi i16 [ 0, %loop.1.header ], [ %iv.2.next, %loop.2.latch ]
74  %c.3 = icmp slt i16 %iv.2, 18
75  br i1 %c.3, label %loop.2.latch, label %loop.1.latch
76
77loop.2.latch:
78  %add.1 = add i32 %N, %iv.1
79  %iv.2.ext = sext i16 %iv.2 to i32
80  %add.2 = add i32 %add.1, %iv.2.ext
81  %c.4 = icmp ult i32 %add.2, 1
82  %iv.2.next = add i16 %iv.2, 1
83  br i1 %c.4, label %exit, label %loop.2.header
84
85loop.1.latch:
86  %iv.1.next = add i32 %iv.1, 1
87  br label %loop.1.header
88
89exit:
90  ret void
91}
92