1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
2; RUN: opt < %s  -loop-vectorize -force-vector-interleave=1 -force-vector-width=4 -dce -instcombine -S | FileCheck %s
3
4target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128"
5target triple = "x86_64-unknown-linux-gnu"
6
7; A tricky loop:
8;
9; void loop(int *a, int *b) {
10;    for (int i = 0; i < 512; ++i) {
11;        a[a[i]] = b[i];
12;        a[i] = b[i+1];
13;    }
14;}
15
16define void @loop(i32* nocapture %a, i32* nocapture %b) nounwind uwtable {
17; CHECK-LABEL: @loop(
18; CHECK-NEXT:  entry:
19; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
20; CHECK:       for.body:
21; CHECK-NEXT:    [[INDVARS_IV:%.*]] = phi i64 [ 0, [[ENTRY:%.*]] ], [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ]
22; CHECK-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i32, i32* [[B:%.*]], i64 [[INDVARS_IV]]
23; CHECK-NEXT:    [[TMP0:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
24; CHECK-NEXT:    [[ARRAYIDX2:%.*]] = getelementptr inbounds i32, i32* [[A:%.*]], i64 [[INDVARS_IV]]
25; CHECK-NEXT:    [[TMP1:%.*]] = load i32, i32* [[ARRAYIDX2]], align 4
26; CHECK-NEXT:    [[IDXPROM3:%.*]] = sext i32 [[TMP1]] to i64
27; CHECK-NEXT:    [[ARRAYIDX4:%.*]] = getelementptr inbounds i32, i32* [[A]], i64 [[IDXPROM3]]
28; CHECK-NEXT:    store i32 [[TMP0]], i32* [[ARRAYIDX4]], align 4
29; CHECK-NEXT:    [[INDVARS_IV_NEXT]] = add i64 [[INDVARS_IV]], 1
30; CHECK-NEXT:    [[ARRAYIDX6:%.*]] = getelementptr inbounds i32, i32* [[B]], i64 [[INDVARS_IV_NEXT]]
31; CHECK-NEXT:    [[TMP2:%.*]] = load i32, i32* [[ARRAYIDX6]], align 4
32; CHECK-NEXT:    store i32 [[TMP2]], i32* [[ARRAYIDX2]], align 4
33; CHECK-NEXT:    [[TMP3:%.*]] = and i64 [[INDVARS_IV_NEXT]], 4294967295
34; CHECK-NEXT:    [[EXITCOND:%.*]] = icmp eq i64 [[TMP3]], 512
35; CHECK-NEXT:    br i1 [[EXITCOND]], label [[FOR_END:%.*]], label [[FOR_BODY]]
36; CHECK:       for.end:
37; CHECK-NEXT:    ret void
38;
39entry:
40  br label %for.body
41
42for.body:                                         ; preds = %for.body, %entry
43  %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
44  %arrayidx = getelementptr inbounds i32, i32* %b, i64 %indvars.iv
45  %0 = load i32, i32* %arrayidx, align 4
46  %arrayidx2 = getelementptr inbounds i32, i32* %a, i64 %indvars.iv
47  %1 = load i32, i32* %arrayidx2, align 4
48  %idxprom3 = sext i32 %1 to i64
49  %arrayidx4 = getelementptr inbounds i32, i32* %a, i64 %idxprom3
50  store i32 %0, i32* %arrayidx4, align 4
51  %indvars.iv.next = add i64 %indvars.iv, 1
52  %arrayidx6 = getelementptr inbounds i32, i32* %b, i64 %indvars.iv.next
53  %2 = load i32, i32* %arrayidx6, align 4
54  store i32 %2, i32* %arrayidx2, align 4
55  %lftr.wideiv = trunc i64 %indvars.iv.next to i32
56  %exitcond = icmp eq i32 %lftr.wideiv, 512
57  br i1 %exitcond, label %for.end, label %for.body
58
59for.end:                                          ; preds = %for.body
60  ret void
61}
62
63; The same loop with parallel loop metadata added to the loop branch
64; and the memory instructions.
65
66define void @parallel_loop(i32* nocapture %a, i32* nocapture %b) nounwind uwtable {
67; CHECK-LABEL: @parallel_loop(
68; CHECK-NEXT:  entry:
69; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
70; CHECK:       vector.ph:
71; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
72; CHECK:       vector.body:
73; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
74; CHECK-NEXT:    [[TMP0:%.*]] = or i64 [[INDEX]], 1
75; CHECK-NEXT:    [[TMP1:%.*]] = or i64 [[INDEX]], 2
76; CHECK-NEXT:    [[TMP2:%.*]] = or i64 [[INDEX]], 3
77; CHECK-NEXT:    [[TMP3:%.*]] = getelementptr inbounds i32, i32* [[B:%.*]], i64 [[INDEX]]
78; CHECK-NEXT:    [[TMP4:%.*]] = bitcast i32* [[TMP3]] to <4 x i32>*
79; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i32>, <4 x i32>* [[TMP4]], align 4, !llvm.access.group [[ACC_GRP0:![0-9]+]]
80; CHECK-NEXT:    [[TMP5:%.*]] = getelementptr inbounds i32, i32* [[A:%.*]], i64 [[INDEX]]
81; CHECK-NEXT:    [[TMP6:%.*]] = getelementptr inbounds i32, i32* [[A]], i64 [[TMP0]]
82; CHECK-NEXT:    [[TMP7:%.*]] = getelementptr inbounds i32, i32* [[A]], i64 [[TMP1]]
83; CHECK-NEXT:    [[TMP8:%.*]] = getelementptr inbounds i32, i32* [[A]], i64 [[TMP2]]
84; CHECK-NEXT:    [[TMP9:%.*]] = load i32, i32* [[TMP5]], align 4, !llvm.access.group [[ACC_GRP0]]
85; CHECK-NEXT:    [[TMP10:%.*]] = load i32, i32* [[TMP6]], align 4, !llvm.access.group [[ACC_GRP0]]
86; CHECK-NEXT:    [[TMP11:%.*]] = load i32, i32* [[TMP7]], align 4, !llvm.access.group [[ACC_GRP0]]
87; CHECK-NEXT:    [[TMP12:%.*]] = load i32, i32* [[TMP8]], align 4, !llvm.access.group [[ACC_GRP0]]
88; CHECK-NEXT:    [[TMP13:%.*]] = sext i32 [[TMP9]] to i64
89; CHECK-NEXT:    [[TMP14:%.*]] = sext i32 [[TMP10]] to i64
90; CHECK-NEXT:    [[TMP15:%.*]] = sext i32 [[TMP11]] to i64
91; CHECK-NEXT:    [[TMP16:%.*]] = sext i32 [[TMP12]] to i64
92; CHECK-NEXT:    [[TMP17:%.*]] = getelementptr inbounds i32, i32* [[A]], i64 [[TMP13]]
93; CHECK-NEXT:    [[TMP18:%.*]] = getelementptr inbounds i32, i32* [[A]], i64 [[TMP14]]
94; CHECK-NEXT:    [[TMP19:%.*]] = getelementptr inbounds i32, i32* [[A]], i64 [[TMP15]]
95; CHECK-NEXT:    [[TMP20:%.*]] = getelementptr inbounds i32, i32* [[A]], i64 [[TMP16]]
96; CHECK-NEXT:    [[TMP21:%.*]] = extractelement <4 x i32> [[WIDE_LOAD]], i64 0
97; CHECK-NEXT:    store i32 [[TMP21]], i32* [[TMP17]], align 4, !llvm.access.group [[ACC_GRP1:![0-9]+]]
98; CHECK-NEXT:    [[TMP22:%.*]] = extractelement <4 x i32> [[WIDE_LOAD]], i64 1
99; CHECK-NEXT:    store i32 [[TMP22]], i32* [[TMP18]], align 4, !llvm.access.group [[ACC_GRP1]]
100; CHECK-NEXT:    [[TMP23:%.*]] = extractelement <4 x i32> [[WIDE_LOAD]], i64 2
101; CHECK-NEXT:    store i32 [[TMP23]], i32* [[TMP19]], align 4, !llvm.access.group [[ACC_GRP1]]
102; CHECK-NEXT:    [[TMP24:%.*]] = extractelement <4 x i32> [[WIDE_LOAD]], i64 3
103; CHECK-NEXT:    store i32 [[TMP24]], i32* [[TMP20]], align 4, !llvm.access.group [[ACC_GRP1]]
104; CHECK-NEXT:    [[TMP25:%.*]] = or i64 [[INDEX]], 1
105; CHECK-NEXT:    [[TMP26:%.*]] = getelementptr inbounds i32, i32* [[B]], i64 [[TMP25]]
106; CHECK-NEXT:    [[TMP27:%.*]] = bitcast i32* [[TMP26]] to <4 x i32>*
107; CHECK-NEXT:    [[WIDE_LOAD1:%.*]] = load <4 x i32>, <4 x i32>* [[TMP27]], align 4, !llvm.access.group [[ACC_GRP0]]
108; CHECK-NEXT:    [[TMP28:%.*]] = bitcast i32* [[TMP5]] to <4 x i32>*
109; CHECK-NEXT:    store <4 x i32> [[WIDE_LOAD1]], <4 x i32>* [[TMP28]], align 4, !llvm.access.group [[ACC_GRP0]]
110; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4
111; CHECK-NEXT:    [[TMP29:%.*]] = icmp eq i64 [[INDEX_NEXT]], 512
112; CHECK-NEXT:    br i1 [[TMP29]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP2:![0-9]+]]
113; CHECK:       middle.block:
114; CHECK-NEXT:    br i1 true, label [[FOR_END:%.*]], label [[SCALAR_PH]]
115; CHECK:       scalar.ph:
116; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
117; CHECK:       for.body:
118; CHECK-NEXT:    br i1 poison, label [[FOR_END]], label [[FOR_BODY]], !llvm.loop [[LOOP5:![0-9]+]]
119; CHECK:       for.end:
120; CHECK-NEXT:    ret void
121;
122entry:
123  br label %for.body
124
125for.body:                                         ; preds = %for.body, %entry
126  %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
127  %arrayidx = getelementptr inbounds i32, i32* %b, i64 %indvars.iv
128  %0 = load i32, i32* %arrayidx, align 4, !llvm.access.group !13
129  %arrayidx2 = getelementptr inbounds i32, i32* %a, i64 %indvars.iv
130  %1 = load i32, i32* %arrayidx2, align 4, !llvm.access.group !13
131  %idxprom3 = sext i32 %1 to i64
132  %arrayidx4 = getelementptr inbounds i32, i32* %a, i64 %idxprom3
133  ; This store might have originated from inlining a function with a parallel
134  ; loop. Refers to a list with the "original loop reference" (!4) also included.
135  store i32 %0, i32* %arrayidx4, align 4, !llvm.access.group !15
136  %indvars.iv.next = add i64 %indvars.iv, 1
137  %arrayidx6 = getelementptr inbounds i32, i32* %b, i64 %indvars.iv.next
138  %2 = load i32, i32* %arrayidx6, align 4, !llvm.access.group !13
139  store i32 %2, i32* %arrayidx2, align 4, !llvm.access.group !13
140  %lftr.wideiv = trunc i64 %indvars.iv.next to i32
141  %exitcond = icmp eq i32 %lftr.wideiv, 512
142  br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !3
143
144for.end:                                          ; preds = %for.body
145  ret void
146}
147
148; The same loop with an illegal parallel loop metadata: the memory
149; accesses refer to a different loop's identifier.
150
151
152define void @mixed_metadata(i32* nocapture %a, i32* nocapture %b) nounwind uwtable {
153; CHECK-LABEL: @mixed_metadata(
154; CHECK-NEXT:  entry:
155; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
156; CHECK:       for.body:
157; CHECK-NEXT:    [[INDVARS_IV:%.*]] = phi i64 [ 0, [[ENTRY:%.*]] ], [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ]
158; CHECK-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i32, i32* [[B:%.*]], i64 [[INDVARS_IV]]
159; CHECK-NEXT:    [[TMP0:%.*]] = load i32, i32* [[ARRAYIDX]], align 4, !llvm.access.group [[ACC_GRP7:![0-9]+]]
160; CHECK-NEXT:    [[ARRAYIDX2:%.*]] = getelementptr inbounds i32, i32* [[A:%.*]], i64 [[INDVARS_IV]]
161; CHECK-NEXT:    [[TMP1:%.*]] = load i32, i32* [[ARRAYIDX2]], align 4, !llvm.access.group [[ACC_GRP7]]
162; CHECK-NEXT:    [[IDXPROM3:%.*]] = sext i32 [[TMP1]] to i64
163; CHECK-NEXT:    [[ARRAYIDX4:%.*]] = getelementptr inbounds i32, i32* [[A]], i64 [[IDXPROM3]]
164; CHECK-NEXT:    store i32 [[TMP0]], i32* [[ARRAYIDX4]], align 4, !llvm.access.group [[ACC_GRP8:![0-9]+]]
165; CHECK-NEXT:    [[INDVARS_IV_NEXT]] = add i64 [[INDVARS_IV]], 1
166; CHECK-NEXT:    [[ARRAYIDX6:%.*]] = getelementptr inbounds i32, i32* [[B]], i64 [[INDVARS_IV_NEXT]]
167; CHECK-NEXT:    [[TMP2:%.*]] = load i32, i32* [[ARRAYIDX6]], align 4, !llvm.access.group [[ACC_GRP7]]
168; CHECK-NEXT:    store i32 [[TMP2]], i32* [[ARRAYIDX2]], align 4, !llvm.access.group [[ACC_GRP7]]
169; CHECK-NEXT:    [[TMP3:%.*]] = and i64 [[INDVARS_IV_NEXT]], 4294967295
170; CHECK-NEXT:    [[EXITCOND:%.*]] = icmp eq i64 [[TMP3]], 512
171; CHECK-NEXT:    br i1 [[EXITCOND]], label [[FOR_END:%.*]], label [[FOR_BODY]], !llvm.loop [[LOOP9:![0-9]+]]
172; CHECK:       for.end:
173; CHECK-NEXT:    ret void
174;
175entry:
176  br label %for.body
177
178for.body:                                         ; preds = %for.body, %entry
179  %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
180  %arrayidx = getelementptr inbounds i32, i32* %b, i64 %indvars.iv
181  %0 = load i32, i32* %arrayidx, align 4, !llvm.access.group !16
182  %arrayidx2 = getelementptr inbounds i32, i32* %a, i64 %indvars.iv
183  %1 = load i32, i32* %arrayidx2, align 4, !llvm.access.group !16
184  %idxprom3 = sext i32 %1 to i64
185  %arrayidx4 = getelementptr inbounds i32, i32* %a, i64 %idxprom3
186  ; This refers to the loop marked with !7 which we are not in at the moment.
187  ; It should prevent detecting as a parallel loop.
188  store i32 %0, i32* %arrayidx4, align 4, !llvm.access.group !17
189  %indvars.iv.next = add i64 %indvars.iv, 1
190  %arrayidx6 = getelementptr inbounds i32, i32* %b, i64 %indvars.iv.next
191  %2 = load i32, i32* %arrayidx6, align 4, !llvm.access.group !16
192  store i32 %2, i32* %arrayidx2, align 4, !llvm.access.group !16
193  %lftr.wideiv = trunc i64 %indvars.iv.next to i32
194  %exitcond = icmp eq i32 %lftr.wideiv, 512
195  br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !6
196
197for.end:                                          ; preds = %for.body
198  ret void
199}
200
201!3 = !{!3, !{!"llvm.loop.parallel_accesses", !13, !15}}
202!4 = !{!4, !{!"llvm.loop.parallel_accesses", !14, !15}}
203!6 = !{!6, !{!"llvm.loop.parallel_accesses", !16}}
204!7 = !{!7, !{!"llvm.loop.parallel_accesses", !17}}
205!13 = distinct !{}
206!14 = distinct !{}
207!15 = distinct !{}
208!16 = distinct !{}
209!17 = distinct !{}
210