xref: /llvm-project/llvm/test/Transforms/LoopVectorize/X86/pr81872.ll (revision 82821254f532c1dbdfd5d985ef7130511efaaa83)
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 4
2; RUN: opt -S -passes=loop-vectorize < %s | FileCheck %s
3target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
4target triple = "x86_64-unknown-linux-gnu"
5
6; PR 81872 explains the issue.
7
8; If we vectorize, we have a miscompile where array IV and thereby value stored in (arr[99],
9; arr[98]) is calculated incorrectly since disjoint or was only disjoint because
10; of dominating conditions. Dropping the disjoint to avoid poison still changes
11; the behaviour since now the or is no longer equivalent to the add.
12;
13define void @test(ptr noundef align 8 dereferenceable_or_null(16) %arr) #0 {
14; CHECK-LABEL: define void @test(
15; CHECK-SAME: ptr noundef align 8 dereferenceable_or_null(16) [[ARR:%.*]]) #[[ATTR0:[0-9]+]] {
16; CHECK-NEXT:  bb5:
17; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]], !prof [[PROF0:![0-9]+]]
18; CHECK:       vector.ph:
19; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
20; CHECK:       vector.body:
21; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
22; CHECK-NEXT:    [[VEC_IND:%.*]] = phi <4 x i64> [ <i64 99, i64 98, i64 97, i64 96>, [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[VECTOR_BODY]] ]
23; CHECK-NEXT:    [[OFFSET_IDX:%.*]] = sub i64 99, [[INDEX]]
24; CHECK-NEXT:    [[TMP0:%.*]] = add i64 [[OFFSET_IDX]], 0
25; CHECK-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <4 x i64> poison, i64 [[INDEX]], i64 0
26; CHECK-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <4 x i64> [[BROADCAST_SPLATINSERT]], <4 x i64> poison, <4 x i32> zeroinitializer
27; CHECK-NEXT:    [[VEC_IV:%.*]] = add <4 x i64> [[BROADCAST_SPLAT]], <i64 0, i64 1, i64 2, i64 3>
28; CHECK-NEXT:    [[TMP1:%.*]] = icmp ule <4 x i64> [[VEC_IV]], splat (i64 8)
29; CHECK-NEXT:    [[TMP2:%.*]] = and <4 x i64> [[VEC_IND]], splat (i64 1)
30; CHECK-NEXT:    [[TMP3:%.*]] = icmp eq <4 x i64> [[TMP2]], zeroinitializer
31; CHECK-NEXT:    [[TMP4:%.*]] = select <4 x i1> [[TMP1]], <4 x i1> [[TMP3]], <4 x i1> zeroinitializer
32; CHECK-NEXT:    [[TMP5:%.*]] = add i64 [[TMP0]], 1
33; CHECK-NEXT:    [[TMP6:%.*]] = getelementptr i64, ptr [[ARR]], i64 [[TMP5]]
34; CHECK-NEXT:    [[TMP7:%.*]] = getelementptr i64, ptr [[TMP6]], i32 0
35; CHECK-NEXT:    [[TMP8:%.*]] = getelementptr i64, ptr [[TMP7]], i32 -3
36; CHECK-NEXT:    [[REVERSE:%.*]] = shufflevector <4 x i1> [[TMP4]], <4 x i1> poison, <4 x i32> <i32 3, i32 2, i32 1, i32 0>
37; CHECK-NEXT:    call void @llvm.masked.store.v4i64.p0(<4 x i64> splat (i64 1), ptr [[TMP8]], i32 8, <4 x i1> [[REVERSE]])
38; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4
39; CHECK-NEXT:    [[VEC_IND_NEXT]] = add <4 x i64> [[VEC_IND]], splat (i64 -4)
40; CHECK-NEXT:    [[TMP9:%.*]] = icmp eq i64 [[INDEX_NEXT]], 12
41; CHECK-NEXT:    br i1 [[TMP9]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !prof [[PROF1:![0-9]+]], !llvm.loop [[LOOP2:![0-9]+]]
42; CHECK:       middle.block:
43; CHECK-NEXT:    br i1 true, label [[BB6:%.*]], label [[SCALAR_PH]], !prof [[PROF5:![0-9]+]]
44; CHECK:       scalar.ph:
45; CHECK-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ 87, [[MIDDLE_BLOCK]] ], [ 99, [[BB5:%.*]] ]
46; CHECK-NEXT:    br label [[LOOP_HEADER:%.*]]
47; CHECK:       loop.header:
48; CHECK-NEXT:    [[IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], [[LOOP_LATCH:%.*]] ]
49; CHECK-NEXT:    [[AND:%.*]] = and i64 [[IV]], 1
50; CHECK-NEXT:    [[ICMP17:%.*]] = icmp eq i64 [[AND]], 0
51; CHECK-NEXT:    br i1 [[ICMP17]], label [[BB18:%.*]], label [[LOOP_LATCH]], !prof [[PROF6:![0-9]+]]
52; CHECK:       bb18:
53; CHECK-NEXT:    [[OR:%.*]] = or disjoint i64 [[IV]], 1
54; CHECK-NEXT:    [[GETELEMENTPTR19:%.*]] = getelementptr inbounds i64, ptr [[ARR]], i64 [[OR]]
55; CHECK-NEXT:    store i64 1, ptr [[GETELEMENTPTR19]], align 8
56; CHECK-NEXT:    br label [[LOOP_LATCH]]
57; CHECK:       loop.latch:
58; CHECK-NEXT:    [[IV_NEXT]] = add nsw i64 [[IV]], -1
59; CHECK-NEXT:    [[ICMP22:%.*]] = icmp eq i64 [[IV_NEXT]], 90
60; CHECK-NEXT:    br i1 [[ICMP22]], label [[BB6]], label [[LOOP_HEADER]], !prof [[PROF7:![0-9]+]], !llvm.loop [[LOOP8:![0-9]+]]
61; CHECK:       bb6:
62; CHECK-NEXT:    ret void
63;
64bb5:
65  br label %loop.header
66
67loop.header:                                             ; preds = %loop.latch, %bb8
68  %iv = phi i64 [ 99, %bb5 ], [ %iv.next, %loop.latch ]
69  %and = and i64 %iv, 1
70  %icmp17 = icmp eq i64 %and, 0
71  br i1 %icmp17, label %bb18, label %loop.latch, !prof !21
72
73bb18:                                             ; preds = %loop.header
74  %or = or disjoint i64 %iv, 1
75  %getelementptr19 = getelementptr inbounds i64, ptr %arr, i64 %or
76  store i64 1, ptr %getelementptr19, align 8
77  br label %loop.latch
78
79loop.latch:                                             ; preds = %bb18, %loop.header
80  %iv.next = add nsw i64 %iv, -1
81  %icmp22 = icmp eq i64 %iv.next, 90
82  br i1 %icmp22, label %bb6, label %loop.header, !prof !22
83
84bb6:
85  ret void
86}
87
88attributes #0 = {"target-cpu"="haswell" "target-features"="+avx2" }
89
90!4 = !{}
91!10 = !{i32 1}
92!16 = !{i64 864}
93!17 = !{i64 8}
94!21 = !{!"branch_weights", i32 1, i32 1}
95!22 = !{!"branch_weights", i32 1, i32 95}
96
97
98;.
99; CHECK: [[PROF0]] = !{!"branch_weights", i32 1, i32 127}
100; CHECK: [[PROF1]] = !{!"branch_weights", i32 1, i32 23}
101; CHECK: [[LOOP2]] = distinct !{[[LOOP2]], [[META3:![0-9]+]], [[META4:![0-9]+]]}
102; CHECK: [[META3]] = !{!"llvm.loop.isvectorized", i32 1}
103; CHECK: [[META4]] = !{!"llvm.loop.unroll.runtime.disable"}
104; CHECK: [[PROF5]] = !{!"branch_weights", i32 1, i32 3}
105; CHECK: [[PROF6]] = !{!"branch_weights", i32 1, i32 1}
106; CHECK: [[PROF7]] = !{!"branch_weights", i32 0, i32 0}
107; CHECK: [[LOOP8]] = distinct !{[[LOOP8]], [[META4]], [[META3]]}
108;.
109