xref: /llvm-project/llvm/test/Transforms/LoopVectorize/first-order-recurrence.ll (revision f48884ded884d982a7fd13394b0e93e6588f4143) 
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
2; RUN: opt < %s -passes=loop-vectorize -force-vector-width=4 -force-vector-interleave=2 -force-widen-divrem-via-safe-divisor=0 -S | FileCheck %s --check-prefix=UNROLL-NO-IC
3; RUN: opt < %s -passes=loop-vectorize -force-vector-width=1 -force-vector-interleave=2 -force-widen-divrem-via-safe-divisor=0 -S | FileCheck %s --check-prefix=UNROLL-NO-VF
4; RUN: opt < %s -passes=loop-vectorize -force-vector-width=4 -force-vector-interleave=1 -force-widen-divrem-via-safe-divisor=0 -S | FileCheck %s --check-prefix=SINK-AFTER
5
6target datalayout = "e-m:e-i64:64-i128:128-n32:64-S128"
7
8; void recurrence_1(int *a, int *b, int n) {
9;   for(int i = 0; i < n; i++)
10;     b[i] =  a[i] + a[i - 1]
11; }
12;
13;
14;
15define void @recurrence_1(ptr readonly noalias %a, ptr noalias %b, i32 %n) {
16; UNROLL-NO-IC-LABEL: @recurrence_1(
17; UNROLL-NO-IC-NEXT:  entry:
18; UNROLL-NO-IC-NEXT:    br label [[FOR_PREHEADER:%.*]]
19; UNROLL-NO-IC:       for.preheader:
20; UNROLL-NO-IC-NEXT:    [[PRE_LOAD:%.*]] = load i32, ptr [[A:%.*]], align 4
21; UNROLL-NO-IC-NEXT:    [[TMP0:%.*]] = add i32 [[N:%.*]], -1
22; UNROLL-NO-IC-NEXT:    [[TMP1:%.*]] = zext i32 [[TMP0]] to i64
23; UNROLL-NO-IC-NEXT:    [[TMP2:%.*]] = add nuw nsw i64 [[TMP1]], 1
24; UNROLL-NO-IC-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[TMP2]], 8
25; UNROLL-NO-IC-NEXT:    br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
26; UNROLL-NO-IC:       vector.ph:
27; UNROLL-NO-IC-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[TMP2]], 8
28; UNROLL-NO-IC-NEXT:    [[N_VEC:%.*]] = sub i64 [[TMP2]], [[N_MOD_VF]]
29; UNROLL-NO-IC-NEXT:    [[VECTOR_RECUR_INIT:%.*]] = insertelement <4 x i32> poison, i32 [[PRE_LOAD]], i32 3
30; UNROLL-NO-IC-NEXT:    br label [[VECTOR_BODY:%.*]]
31; UNROLL-NO-IC:       vector.body:
32; UNROLL-NO-IC-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
33; UNROLL-NO-IC-NEXT:    [[VECTOR_RECUR:%.*]] = phi <4 x i32> [ [[VECTOR_RECUR_INIT]], [[VECTOR_PH]] ], [ [[WIDE_LOAD1:%.*]], [[VECTOR_BODY]] ]
34; UNROLL-NO-IC-NEXT:    [[TMP3:%.*]] = add i64 [[INDEX]], 0
35; UNROLL-NO-IC-NEXT:    [[TMP4:%.*]] = add nuw nsw i64 [[TMP3]], 1
36; UNROLL-NO-IC-NEXT:    [[TMP5:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[TMP4]]
37; UNROLL-NO-IC-NEXT:    [[TMP6:%.*]] = getelementptr inbounds i32, ptr [[TMP5]], i32 0
38; UNROLL-NO-IC-NEXT:    [[TMP7:%.*]] = getelementptr inbounds i32, ptr [[TMP5]], i32 4
39; UNROLL-NO-IC-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i32>, ptr [[TMP6]], align 4
40; UNROLL-NO-IC-NEXT:    [[WIDE_LOAD1]] = load <4 x i32>, ptr [[TMP7]], align 4
41; UNROLL-NO-IC-NEXT:    [[TMP8:%.*]] = shufflevector <4 x i32> [[VECTOR_RECUR]], <4 x i32> [[WIDE_LOAD]], <4 x i32> <i32 3, i32 4, i32 5, i32 6>
42; UNROLL-NO-IC-NEXT:    [[TMP9:%.*]] = shufflevector <4 x i32> [[WIDE_LOAD]], <4 x i32> [[WIDE_LOAD1]], <4 x i32> <i32 3, i32 4, i32 5, i32 6>
43; UNROLL-NO-IC-NEXT:    [[TMP10:%.*]] = getelementptr inbounds i32, ptr [[B:%.*]], i64 [[TMP3]]
44; UNROLL-NO-IC-NEXT:    [[TMP11:%.*]] = add <4 x i32> [[WIDE_LOAD]], [[TMP8]]
45; UNROLL-NO-IC-NEXT:    [[TMP12:%.*]] = add <4 x i32> [[WIDE_LOAD1]], [[TMP9]]
46; UNROLL-NO-IC-NEXT:    [[TMP13:%.*]] = getelementptr inbounds i32, ptr [[TMP10]], i32 0
47; UNROLL-NO-IC-NEXT:    [[TMP14:%.*]] = getelementptr inbounds i32, ptr [[TMP10]], i32 4
48; UNROLL-NO-IC-NEXT:    store <4 x i32> [[TMP11]], ptr [[TMP13]], align 4
49; UNROLL-NO-IC-NEXT:    store <4 x i32> [[TMP12]], ptr [[TMP14]], align 4
50; UNROLL-NO-IC-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 8
51; UNROLL-NO-IC-NEXT:    [[TMP15:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
52; UNROLL-NO-IC-NEXT:    br i1 [[TMP15]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
53; UNROLL-NO-IC:       middle.block:
54; UNROLL-NO-IC-NEXT:    [[VECTOR_RECUR_EXTRACT:%.*]] = extractelement <4 x i32> [[WIDE_LOAD1]], i32 3
55; UNROLL-NO-IC-NEXT:    [[CMP_N:%.*]] = icmp eq i64 [[TMP2]], [[N_VEC]]
56; UNROLL-NO-IC-NEXT:    br i1 [[CMP_N]], label [[FOR_EXIT:%.*]], label [[SCALAR_PH]]
57; UNROLL-NO-IC:       scalar.ph:
58; UNROLL-NO-IC-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i32 [ [[VECTOR_RECUR_EXTRACT]], [[MIDDLE_BLOCK]] ], [ [[PRE_LOAD]], [[FOR_PREHEADER]] ]
59; UNROLL-NO-IC-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[FOR_PREHEADER]] ]
60; UNROLL-NO-IC-NEXT:    br label [[SCALAR_BODY:%.*]]
61; UNROLL-NO-IC:       scalar.body:
62; UNROLL-NO-IC-NEXT:    [[TMP16:%.*]] = phi i32 [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ], [ [[TMP17:%.*]], [[SCALAR_BODY]] ]
63; UNROLL-NO-IC-NEXT:    [[INDVARS_IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[INDVARS_IV_NEXT:%.*]], [[SCALAR_BODY]] ]
64; UNROLL-NO-IC-NEXT:    [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1
65; UNROLL-NO-IC-NEXT:    [[ARRAYIDX32:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[INDVARS_IV_NEXT]]
66; UNROLL-NO-IC-NEXT:    [[TMP17]] = load i32, ptr [[ARRAYIDX32]], align 4
67; UNROLL-NO-IC-NEXT:    [[ARRAYIDX34:%.*]] = getelementptr inbounds i32, ptr [[B]], i64 [[INDVARS_IV]]
68; UNROLL-NO-IC-NEXT:    [[ADD35:%.*]] = add i32 [[TMP17]], [[TMP16]]
69; UNROLL-NO-IC-NEXT:    store i32 [[ADD35]], ptr [[ARRAYIDX34]], align 4
70; UNROLL-NO-IC-NEXT:    [[LFTR_WIDEIV:%.*]] = trunc i64 [[INDVARS_IV_NEXT]] to i32
71; UNROLL-NO-IC-NEXT:    [[EXITCOND:%.*]] = icmp eq i32 [[LFTR_WIDEIV]], [[N]]
72; UNROLL-NO-IC-NEXT:    br i1 [[EXITCOND]], label [[FOR_EXIT]], label [[SCALAR_BODY]], !llvm.loop [[LOOP3:![0-9]+]]
73; UNROLL-NO-IC:       for.exit:
74; UNROLL-NO-IC-NEXT:    ret void
75;
76; UNROLL-NO-VF-LABEL: @recurrence_1(
77; UNROLL-NO-VF-NEXT:  entry:
78; UNROLL-NO-VF-NEXT:    br label [[FOR_PREHEADER:%.*]]
79; UNROLL-NO-VF:       for.preheader:
80; UNROLL-NO-VF-NEXT:    [[PRE_LOAD:%.*]] = load i32, ptr [[A:%.*]], align 4
81; UNROLL-NO-VF-NEXT:    [[TMP0:%.*]] = add i32 [[N:%.*]], -1
82; UNROLL-NO-VF-NEXT:    [[TMP1:%.*]] = zext i32 [[TMP0]] to i64
83; UNROLL-NO-VF-NEXT:    [[TMP2:%.*]] = add nuw nsw i64 [[TMP1]], 1
84; UNROLL-NO-VF-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[TMP2]], 2
85; UNROLL-NO-VF-NEXT:    br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
86; UNROLL-NO-VF:       vector.ph:
87; UNROLL-NO-VF-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[TMP2]], 2
88; UNROLL-NO-VF-NEXT:    [[N_VEC:%.*]] = sub i64 [[TMP2]], [[N_MOD_VF]]
89; UNROLL-NO-VF-NEXT:    br label [[VECTOR_BODY:%.*]]
90; UNROLL-NO-VF:       vector.body:
91; UNROLL-NO-VF-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
92; UNROLL-NO-VF-NEXT:    [[VECTOR_RECUR:%.*]] = phi i32 [ [[PRE_LOAD]], [[VECTOR_PH]] ], [ [[TMP10:%.*]], [[VECTOR_BODY]] ]
93; UNROLL-NO-VF-NEXT:    [[TMP3:%.*]] = add i64 [[INDEX]], 0
94; UNROLL-NO-VF-NEXT:    [[TMP4:%.*]] = add i64 [[INDEX]], 1
95; UNROLL-NO-VF-NEXT:    [[TMP5:%.*]] = add nuw nsw i64 [[TMP3]], 1
96; UNROLL-NO-VF-NEXT:    [[TMP6:%.*]] = add nuw nsw i64 [[TMP4]], 1
97; UNROLL-NO-VF-NEXT:    [[TMP7:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[TMP5]]
98; UNROLL-NO-VF-NEXT:    [[TMP8:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[TMP6]]
99; UNROLL-NO-VF-NEXT:    [[TMP9:%.*]] = load i32, ptr [[TMP7]], align 4
100; UNROLL-NO-VF-NEXT:    [[TMP10]] = load i32, ptr [[TMP8]], align 4
101; UNROLL-NO-VF-NEXT:    [[TMP11:%.*]] = getelementptr inbounds i32, ptr [[B:%.*]], i64 [[TMP3]]
102; UNROLL-NO-VF-NEXT:    [[TMP12:%.*]] = getelementptr inbounds i32, ptr [[B]], i64 [[TMP4]]
103; UNROLL-NO-VF-NEXT:    [[TMP13:%.*]] = add i32 [[TMP9]], [[VECTOR_RECUR]]
104; UNROLL-NO-VF-NEXT:    [[TMP14:%.*]] = add i32 [[TMP10]], [[TMP9]]
105; UNROLL-NO-VF-NEXT:    store i32 [[TMP13]], ptr [[TMP11]], align 4
106; UNROLL-NO-VF-NEXT:    store i32 [[TMP14]], ptr [[TMP12]], align 4
107; UNROLL-NO-VF-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 2
108; UNROLL-NO-VF-NEXT:    [[TMP15:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
109; UNROLL-NO-VF-NEXT:    br i1 [[TMP15]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
110; UNROLL-NO-VF:       middle.block:
111; UNROLL-NO-VF-NEXT:    [[CMP_N:%.*]] = icmp eq i64 [[TMP2]], [[N_VEC]]
112; UNROLL-NO-VF-NEXT:    br i1 [[CMP_N]], label [[FOR_EXIT:%.*]], label [[SCALAR_PH]]
113; UNROLL-NO-VF:       scalar.ph:
114; UNROLL-NO-VF-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i32 [ [[TMP10]], [[MIDDLE_BLOCK]] ], [ [[PRE_LOAD]], [[FOR_PREHEADER]] ]
115; UNROLL-NO-VF-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[FOR_PREHEADER]] ]
116; UNROLL-NO-VF-NEXT:    br label [[SCALAR_BODY:%.*]]
117; UNROLL-NO-VF:       scalar.body:
118; UNROLL-NO-VF-NEXT:    [[TMP16:%.*]] = phi i32 [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ], [ [[TMP17:%.*]], [[SCALAR_BODY]] ]
119; UNROLL-NO-VF-NEXT:    [[INDVARS_IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[INDVARS_IV_NEXT:%.*]], [[SCALAR_BODY]] ]
120; UNROLL-NO-VF-NEXT:    [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1
121; UNROLL-NO-VF-NEXT:    [[ARRAYIDX32:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[INDVARS_IV_NEXT]]
122; UNROLL-NO-VF-NEXT:    [[TMP17]] = load i32, ptr [[ARRAYIDX32]], align 4
123; UNROLL-NO-VF-NEXT:    [[ARRAYIDX34:%.*]] = getelementptr inbounds i32, ptr [[B]], i64 [[INDVARS_IV]]
124; UNROLL-NO-VF-NEXT:    [[ADD35:%.*]] = add i32 [[TMP17]], [[TMP16]]
125; UNROLL-NO-VF-NEXT:    store i32 [[ADD35]], ptr [[ARRAYIDX34]], align 4
126; UNROLL-NO-VF-NEXT:    [[LFTR_WIDEIV:%.*]] = trunc i64 [[INDVARS_IV_NEXT]] to i32
127; UNROLL-NO-VF-NEXT:    [[EXITCOND:%.*]] = icmp eq i32 [[LFTR_WIDEIV]], [[N]]
128; UNROLL-NO-VF-NEXT:    br i1 [[EXITCOND]], label [[FOR_EXIT]], label [[SCALAR_BODY]], !llvm.loop [[LOOP3:![0-9]+]]
129; UNROLL-NO-VF:       for.exit:
130; UNROLL-NO-VF-NEXT:    ret void
131;
132; SINK-AFTER-LABEL: @recurrence_1(
133; SINK-AFTER-NEXT:  entry:
134; SINK-AFTER-NEXT:    br label [[FOR_PREHEADER:%.*]]
135; SINK-AFTER:       for.preheader:
136; SINK-AFTER-NEXT:    [[PRE_LOAD:%.*]] = load i32, ptr [[A:%.*]], align 4
137; SINK-AFTER-NEXT:    [[TMP0:%.*]] = add i32 [[N:%.*]], -1
138; SINK-AFTER-NEXT:    [[TMP1:%.*]] = zext i32 [[TMP0]] to i64
139; SINK-AFTER-NEXT:    [[TMP2:%.*]] = add nuw nsw i64 [[TMP1]], 1
140; SINK-AFTER-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[TMP2]], 4
141; SINK-AFTER-NEXT:    br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
142; SINK-AFTER:       vector.ph:
143; SINK-AFTER-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[TMP2]], 4
144; SINK-AFTER-NEXT:    [[N_VEC:%.*]] = sub i64 [[TMP2]], [[N_MOD_VF]]
145; SINK-AFTER-NEXT:    [[VECTOR_RECUR_INIT:%.*]] = insertelement <4 x i32> poison, i32 [[PRE_LOAD]], i32 3
146; SINK-AFTER-NEXT:    br label [[VECTOR_BODY:%.*]]
147; SINK-AFTER:       vector.body:
148; SINK-AFTER-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
149; SINK-AFTER-NEXT:    [[VECTOR_RECUR:%.*]] = phi <4 x i32> [ [[VECTOR_RECUR_INIT]], [[VECTOR_PH]] ], [ [[WIDE_LOAD:%.*]], [[VECTOR_BODY]] ]
150; SINK-AFTER-NEXT:    [[TMP3:%.*]] = add i64 [[INDEX]], 0
151; SINK-AFTER-NEXT:    [[TMP4:%.*]] = add nuw nsw i64 [[TMP3]], 1
152; SINK-AFTER-NEXT:    [[TMP5:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[TMP4]]
153; SINK-AFTER-NEXT:    [[TMP6:%.*]] = getelementptr inbounds i32, ptr [[TMP5]], i32 0
154; SINK-AFTER-NEXT:    [[WIDE_LOAD]] = load <4 x i32>, ptr [[TMP6]], align 4
155; SINK-AFTER-NEXT:    [[TMP7:%.*]] = shufflevector <4 x i32> [[VECTOR_RECUR]], <4 x i32> [[WIDE_LOAD]], <4 x i32> <i32 3, i32 4, i32 5, i32 6>
156; SINK-AFTER-NEXT:    [[TMP8:%.*]] = getelementptr inbounds i32, ptr [[B:%.*]], i64 [[TMP3]]
157; SINK-AFTER-NEXT:    [[TMP9:%.*]] = add <4 x i32> [[WIDE_LOAD]], [[TMP7]]
158; SINK-AFTER-NEXT:    [[TMP10:%.*]] = getelementptr inbounds i32, ptr [[TMP8]], i32 0
159; SINK-AFTER-NEXT:    store <4 x i32> [[TMP9]], ptr [[TMP10]], align 4
160; SINK-AFTER-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4
161; SINK-AFTER-NEXT:    [[TMP11:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
162; SINK-AFTER-NEXT:    br i1 [[TMP11]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
163; SINK-AFTER:       middle.block:
164; SINK-AFTER-NEXT:    [[VECTOR_RECUR_EXTRACT:%.*]] = extractelement <4 x i32> [[WIDE_LOAD]], i32 3
165; SINK-AFTER-NEXT:    [[CMP_N:%.*]] = icmp eq i64 [[TMP2]], [[N_VEC]]
166; SINK-AFTER-NEXT:    br i1 [[CMP_N]], label [[FOR_EXIT:%.*]], label [[SCALAR_PH]]
167; SINK-AFTER:       scalar.ph:
168; SINK-AFTER-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i32 [ [[VECTOR_RECUR_EXTRACT]], [[MIDDLE_BLOCK]] ], [ [[PRE_LOAD]], [[FOR_PREHEADER]] ]
169; SINK-AFTER-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[FOR_PREHEADER]] ]
170; SINK-AFTER-NEXT:    br label [[SCALAR_BODY:%.*]]
171; SINK-AFTER:       scalar.body:
172; SINK-AFTER-NEXT:    [[TMP12:%.*]] = phi i32 [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ], [ [[TMP13:%.*]], [[SCALAR_BODY]] ]
173; SINK-AFTER-NEXT:    [[INDVARS_IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[INDVARS_IV_NEXT:%.*]], [[SCALAR_BODY]] ]
174; SINK-AFTER-NEXT:    [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1
175; SINK-AFTER-NEXT:    [[ARRAYIDX32:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[INDVARS_IV_NEXT]]
176; SINK-AFTER-NEXT:    [[TMP13]] = load i32, ptr [[ARRAYIDX32]], align 4
177; SINK-AFTER-NEXT:    [[ARRAYIDX34:%.*]] = getelementptr inbounds i32, ptr [[B]], i64 [[INDVARS_IV]]
178; SINK-AFTER-NEXT:    [[ADD35:%.*]] = add i32 [[TMP13]], [[TMP12]]
179; SINK-AFTER-NEXT:    store i32 [[ADD35]], ptr [[ARRAYIDX34]], align 4
180; SINK-AFTER-NEXT:    [[LFTR_WIDEIV:%.*]] = trunc i64 [[INDVARS_IV_NEXT]] to i32
181; SINK-AFTER-NEXT:    [[EXITCOND:%.*]] = icmp eq i32 [[LFTR_WIDEIV]], [[N]]
182; SINK-AFTER-NEXT:    br i1 [[EXITCOND]], label [[FOR_EXIT]], label [[SCALAR_BODY]], !llvm.loop [[LOOP3:![0-9]+]]
183; SINK-AFTER:       for.exit:
184; SINK-AFTER-NEXT:    ret void
185;
186entry:
187  br label %for.preheader
188
189for.preheader:
190  %pre_load = load i32, ptr %a
191  br label %scalar.body
192
193scalar.body:
194  %0 = phi i32 [ %pre_load, %for.preheader ], [ %1, %scalar.body ]
195  %indvars.iv = phi i64 [ 0, %for.preheader ], [ %indvars.iv.next, %scalar.body ]
196  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
197  %arrayidx32 = getelementptr inbounds i32, ptr %a, i64 %indvars.iv.next
198  %1 = load i32, ptr %arrayidx32
199  %arrayidx34 = getelementptr inbounds i32, ptr %b, i64 %indvars.iv
200  %add35 = add i32 %1, %0
201  store i32 %add35, ptr %arrayidx34
202  %lftr.wideiv = trunc i64 %indvars.iv.next to i32
203  %exitcond = icmp eq i32 %lftr.wideiv, %n
204  br i1 %exitcond, label %for.exit, label %scalar.body
205
206for.exit:
207  ret void
208}
209
210; int recurrence_2(int *a, int n) {
211;   int minmax;
212;   for (int i = 0; i < n; ++i)
213;     minmax = min(minmax, max(a[i] - a[i-1], 0));
214;   return minmax;
215; }
216;
217;
218;
219define i32 @recurrence_2(ptr nocapture readonly %a, i32 %n) {
220; UNROLL-NO-IC-LABEL: @recurrence_2(
221; UNROLL-NO-IC-NEXT:  entry:
222; UNROLL-NO-IC-NEXT:    [[CMP27:%.*]] = icmp sgt i32 [[N:%.*]], 0
223; UNROLL-NO-IC-NEXT:    br i1 [[CMP27]], label [[FOR_PREHEADER:%.*]], label [[FOR_COND_CLEANUP:%.*]]
224; UNROLL-NO-IC:       for.preheader:
225; UNROLL-NO-IC-NEXT:    [[ARRAYIDX2_PHI_TRANS_INSERT:%.*]] = getelementptr inbounds i32, ptr [[A:%.*]], i64 -1
226; UNROLL-NO-IC-NEXT:    [[DOTPRE:%.*]] = load i32, ptr [[ARRAYIDX2_PHI_TRANS_INSERT]], align 4
227; UNROLL-NO-IC-NEXT:    [[TMP0:%.*]] = zext i32 [[N]] to i64
228; UNROLL-NO-IC-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[TMP0]], 8
229; UNROLL-NO-IC-NEXT:    br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
230; UNROLL-NO-IC:       vector.ph:
231; UNROLL-NO-IC-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[TMP0]], 8
232; UNROLL-NO-IC-NEXT:    [[N_VEC:%.*]] = sub i64 [[TMP0]], [[N_MOD_VF]]
233; UNROLL-NO-IC-NEXT:    [[VECTOR_RECUR_INIT:%.*]] = insertelement <4 x i32> poison, i32 [[DOTPRE]], i32 3
234; UNROLL-NO-IC-NEXT:    br label [[VECTOR_BODY:%.*]]
235; UNROLL-NO-IC:       vector.body:
236; UNROLL-NO-IC-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
237; UNROLL-NO-IC-NEXT:    [[VECTOR_RECUR:%.*]] = phi <4 x i32> [ [[VECTOR_RECUR_INIT]], [[VECTOR_PH]] ], [ [[WIDE_LOAD2:%.*]], [[VECTOR_BODY]] ]
238; UNROLL-NO-IC-NEXT:    [[VEC_PHI:%.*]] = phi <4 x i32> [ poison, [[VECTOR_PH]] ], [ [[TMP15:%.*]], [[VECTOR_BODY]] ]
239; UNROLL-NO-IC-NEXT:    [[VEC_PHI1:%.*]] = phi <4 x i32> [ poison, [[VECTOR_PH]] ], [ [[TMP16:%.*]], [[VECTOR_BODY]] ]
240; UNROLL-NO-IC-NEXT:    [[TMP1:%.*]] = add i64 [[INDEX]], 0
241; UNROLL-NO-IC-NEXT:    [[TMP2:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[TMP1]]
242; UNROLL-NO-IC-NEXT:    [[TMP3:%.*]] = getelementptr inbounds i32, ptr [[TMP2]], i32 0
243; UNROLL-NO-IC-NEXT:    [[TMP4:%.*]] = getelementptr inbounds i32, ptr [[TMP2]], i32 4
244; UNROLL-NO-IC-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i32>, ptr [[TMP3]], align 4
245; UNROLL-NO-IC-NEXT:    [[WIDE_LOAD2]] = load <4 x i32>, ptr [[TMP4]], align 4
246; UNROLL-NO-IC-NEXT:    [[TMP5:%.*]] = shufflevector <4 x i32> [[VECTOR_RECUR]], <4 x i32> [[WIDE_LOAD]], <4 x i32> <i32 3, i32 4, i32 5, i32 6>
247; UNROLL-NO-IC-NEXT:    [[TMP6:%.*]] = shufflevector <4 x i32> [[WIDE_LOAD]], <4 x i32> [[WIDE_LOAD2]], <4 x i32> <i32 3, i32 4, i32 5, i32 6>
248; UNROLL-NO-IC-NEXT:    [[TMP7:%.*]] = sub nsw <4 x i32> [[WIDE_LOAD]], [[TMP5]]
249; UNROLL-NO-IC-NEXT:    [[TMP8:%.*]] = sub nsw <4 x i32> [[WIDE_LOAD2]], [[TMP6]]
250; UNROLL-NO-IC-NEXT:    [[TMP9:%.*]] = icmp sgt <4 x i32> [[TMP7]], zeroinitializer
251; UNROLL-NO-IC-NEXT:    [[TMP10:%.*]] = icmp sgt <4 x i32> [[TMP8]], zeroinitializer
252; UNROLL-NO-IC-NEXT:    [[TMP11:%.*]] = select <4 x i1> [[TMP9]], <4 x i32> [[TMP7]], <4 x i32> zeroinitializer
253; UNROLL-NO-IC-NEXT:    [[TMP12:%.*]] = select <4 x i1> [[TMP10]], <4 x i32> [[TMP8]], <4 x i32> zeroinitializer
254; UNROLL-NO-IC-NEXT:    [[TMP13:%.*]] = icmp slt <4 x i32> [[VEC_PHI]], [[TMP11]]
255; UNROLL-NO-IC-NEXT:    [[TMP14:%.*]] = icmp slt <4 x i32> [[VEC_PHI1]], [[TMP12]]
256; UNROLL-NO-IC-NEXT:    [[TMP15]] = select <4 x i1> [[TMP13]], <4 x i32> [[VEC_PHI]], <4 x i32> [[TMP11]]
257; UNROLL-NO-IC-NEXT:    [[TMP16]] = select <4 x i1> [[TMP14]], <4 x i32> [[VEC_PHI1]], <4 x i32> [[TMP12]]
258; UNROLL-NO-IC-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 8
259; UNROLL-NO-IC-NEXT:    [[TMP17:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
260; UNROLL-NO-IC-NEXT:    br i1 [[TMP17]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]]
261; UNROLL-NO-IC:       middle.block:
262; UNROLL-NO-IC-NEXT:    [[RDX_MINMAX:%.*]] = call <4 x i32> @llvm.smin.v4i32(<4 x i32> [[TMP15]], <4 x i32> [[TMP16]])
263; UNROLL-NO-IC-NEXT:    [[TMP18:%.*]] = call i32 @llvm.vector.reduce.smin.v4i32(<4 x i32> [[RDX_MINMAX]])
264; UNROLL-NO-IC-NEXT:    [[VECTOR_RECUR_EXTRACT:%.*]] = extractelement <4 x i32> [[WIDE_LOAD2]], i32 3
265; UNROLL-NO-IC-NEXT:    [[CMP_N:%.*]] = icmp eq i64 [[TMP0]], [[N_VEC]]
266; UNROLL-NO-IC-NEXT:    br i1 [[CMP_N]], label [[FOR_COND_CLEANUP_LOOPEXIT:%.*]], label [[SCALAR_PH]]
267; UNROLL-NO-IC:       scalar.ph:
268; UNROLL-NO-IC-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i32 [ [[VECTOR_RECUR_EXTRACT]], [[MIDDLE_BLOCK]] ], [ [[DOTPRE]], [[FOR_PREHEADER]] ]
269; UNROLL-NO-IC-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[FOR_PREHEADER]] ]
270; UNROLL-NO-IC-NEXT:    [[BC_MERGE_RDX:%.*]] = phi i32 [ [[TMP18]], [[MIDDLE_BLOCK]] ], [ poison, [[FOR_PREHEADER]] ]
271; UNROLL-NO-IC-NEXT:    br label [[SCALAR_BODY:%.*]]
272; UNROLL-NO-IC:       for.cond.cleanup.loopexit:
273; UNROLL-NO-IC-NEXT:    [[MINMAX_0_COND_LCSSA:%.*]] = phi i32 [ [[MINMAX_0_COND:%.*]], [[SCALAR_BODY]] ], [ [[TMP18]], [[MIDDLE_BLOCK]] ]
274; UNROLL-NO-IC-NEXT:    br label [[FOR_COND_CLEANUP]]
275; UNROLL-NO-IC:       for.cond.cleanup:
276; UNROLL-NO-IC-NEXT:    [[MINMAX_0_LCSSA:%.*]] = phi i32 [ poison, [[ENTRY:%.*]] ], [ [[MINMAX_0_COND_LCSSA]], [[FOR_COND_CLEANUP_LOOPEXIT]] ]
277; UNROLL-NO-IC-NEXT:    ret i32 [[MINMAX_0_LCSSA]]
278; UNROLL-NO-IC:       scalar.body:
279; UNROLL-NO-IC-NEXT:    [[TMP19:%.*]] = phi i32 [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ], [ [[TMP20:%.*]], [[SCALAR_BODY]] ]
280; UNROLL-NO-IC-NEXT:    [[INDVARS_IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[INDVARS_IV_NEXT:%.*]], [[SCALAR_BODY]] ]
281; UNROLL-NO-IC-NEXT:    [[MINMAX_028:%.*]] = phi i32 [ [[BC_MERGE_RDX]], [[SCALAR_PH]] ], [ [[MINMAX_0_COND]], [[SCALAR_BODY]] ]
282; UNROLL-NO-IC-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[INDVARS_IV]]
283; UNROLL-NO-IC-NEXT:    [[TMP20]] = load i32, ptr [[ARRAYIDX]], align 4
284; UNROLL-NO-IC-NEXT:    [[SUB3:%.*]] = sub nsw i32 [[TMP20]], [[TMP19]]
285; UNROLL-NO-IC-NEXT:    [[CMP4:%.*]] = icmp sgt i32 [[SUB3]], 0
286; UNROLL-NO-IC-NEXT:    [[COND:%.*]] = select i1 [[CMP4]], i32 [[SUB3]], i32 0
287; UNROLL-NO-IC-NEXT:    [[CMP5:%.*]] = icmp slt i32 [[MINMAX_028]], [[COND]]
288; UNROLL-NO-IC-NEXT:    [[MINMAX_0_COND]] = select i1 [[CMP5]], i32 [[MINMAX_028]], i32 [[COND]]
289; UNROLL-NO-IC-NEXT:    [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1
290; UNROLL-NO-IC-NEXT:    [[LFTR_WIDEIV:%.*]] = trunc i64 [[INDVARS_IV_NEXT]] to i32
291; UNROLL-NO-IC-NEXT:    [[EXITCOND:%.*]] = icmp eq i32 [[LFTR_WIDEIV]], [[N]]
292; UNROLL-NO-IC-NEXT:    br i1 [[EXITCOND]], label [[FOR_COND_CLEANUP_LOOPEXIT]], label [[SCALAR_BODY]], !llvm.loop [[LOOP5:![0-9]+]]
293;
294; UNROLL-NO-VF-LABEL: @recurrence_2(
295; UNROLL-NO-VF-NEXT:  entry:
296; UNROLL-NO-VF-NEXT:    [[CMP27:%.*]] = icmp sgt i32 [[N:%.*]], 0
297; UNROLL-NO-VF-NEXT:    br i1 [[CMP27]], label [[FOR_PREHEADER:%.*]], label [[FOR_COND_CLEANUP:%.*]]
298; UNROLL-NO-VF:       for.preheader:
299; UNROLL-NO-VF-NEXT:    [[ARRAYIDX2_PHI_TRANS_INSERT:%.*]] = getelementptr inbounds i32, ptr [[A:%.*]], i64 -1
300; UNROLL-NO-VF-NEXT:    [[DOTPRE:%.*]] = load i32, ptr [[ARRAYIDX2_PHI_TRANS_INSERT]], align 4
301; UNROLL-NO-VF-NEXT:    [[TMP0:%.*]] = zext i32 [[N]] to i64
302; UNROLL-NO-VF-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[TMP0]], 2
303; UNROLL-NO-VF-NEXT:    br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
304; UNROLL-NO-VF:       vector.ph:
305; UNROLL-NO-VF-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[TMP0]], 2
306; UNROLL-NO-VF-NEXT:    [[N_VEC:%.*]] = sub i64 [[TMP0]], [[N_MOD_VF]]
307; UNROLL-NO-VF-NEXT:    br label [[VECTOR_BODY:%.*]]
308; UNROLL-NO-VF:       vector.body:
309; UNROLL-NO-VF-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
310; UNROLL-NO-VF-NEXT:    [[VECTOR_RECUR:%.*]] = phi i32 [ [[DOTPRE]], [[VECTOR_PH]] ], [ [[TMP6:%.*]], [[VECTOR_BODY]] ]
311; UNROLL-NO-VF-NEXT:    [[VEC_PHI:%.*]] = phi i32 [ poison, [[VECTOR_PH]] ], [ [[TMP15:%.*]], [[VECTOR_BODY]] ]
312; UNROLL-NO-VF-NEXT:    [[VEC_PHI1:%.*]] = phi i32 [ poison, [[VECTOR_PH]] ], [ [[TMP16:%.*]], [[VECTOR_BODY]] ]
313; UNROLL-NO-VF-NEXT:    [[TMP1:%.*]] = add i64 [[INDEX]], 0
314; UNROLL-NO-VF-NEXT:    [[TMP2:%.*]] = add i64 [[INDEX]], 1
315; UNROLL-NO-VF-NEXT:    [[TMP3:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[TMP1]]
316; UNROLL-NO-VF-NEXT:    [[TMP4:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[TMP2]]
317; UNROLL-NO-VF-NEXT:    [[TMP5:%.*]] = load i32, ptr [[TMP3]], align 4
318; UNROLL-NO-VF-NEXT:    [[TMP6]] = load i32, ptr [[TMP4]], align 4
319; UNROLL-NO-VF-NEXT:    [[TMP7:%.*]] = sub nsw i32 [[TMP5]], [[VECTOR_RECUR]]
320; UNROLL-NO-VF-NEXT:    [[TMP8:%.*]] = sub nsw i32 [[TMP6]], [[TMP5]]
321; UNROLL-NO-VF-NEXT:    [[TMP9:%.*]] = icmp sgt i32 [[TMP7]], 0
322; UNROLL-NO-VF-NEXT:    [[TMP10:%.*]] = icmp sgt i32 [[TMP8]], 0
323; UNROLL-NO-VF-NEXT:    [[TMP11:%.*]] = select i1 [[TMP9]], i32 [[TMP7]], i32 0
324; UNROLL-NO-VF-NEXT:    [[TMP12:%.*]] = select i1 [[TMP10]], i32 [[TMP8]], i32 0
325; UNROLL-NO-VF-NEXT:    [[TMP13:%.*]] = icmp slt i32 [[VEC_PHI]], [[TMP11]]
326; UNROLL-NO-VF-NEXT:    [[TMP14:%.*]] = icmp slt i32 [[VEC_PHI1]], [[TMP12]]
327; UNROLL-NO-VF-NEXT:    [[TMP15]] = select i1 [[TMP13]], i32 [[VEC_PHI]], i32 [[TMP11]]
328; UNROLL-NO-VF-NEXT:    [[TMP16]] = select i1 [[TMP14]], i32 [[VEC_PHI1]], i32 [[TMP12]]
329; UNROLL-NO-VF-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 2
330; UNROLL-NO-VF-NEXT:    [[TMP17:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
331; UNROLL-NO-VF-NEXT:    br i1 [[TMP17]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]]
332; UNROLL-NO-VF:       middle.block:
333; UNROLL-NO-VF-NEXT:    [[RDX_MINMAX:%.*]] = call i32 @llvm.smin.i32(i32 [[TMP15]], i32 [[TMP16]])
334; UNROLL-NO-VF-NEXT:    [[CMP_N:%.*]] = icmp eq i64 [[TMP0]], [[N_VEC]]
335; UNROLL-NO-VF-NEXT:    br i1 [[CMP_N]], label [[FOR_COND_CLEANUP_LOOPEXIT:%.*]], label [[SCALAR_PH]]
336; UNROLL-NO-VF:       scalar.ph:
337; UNROLL-NO-VF-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i32 [ [[TMP6]], [[MIDDLE_BLOCK]] ], [ [[DOTPRE]], [[FOR_PREHEADER]] ]
338; UNROLL-NO-VF-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[FOR_PREHEADER]] ]
339; UNROLL-NO-VF-NEXT:    [[BC_MERGE_RDX:%.*]] = phi i32 [ [[RDX_MINMAX]], [[MIDDLE_BLOCK]] ], [ poison, [[FOR_PREHEADER]] ]
340; UNROLL-NO-VF-NEXT:    br label [[SCALAR_BODY:%.*]]
341; UNROLL-NO-VF:       for.cond.cleanup.loopexit:
342; UNROLL-NO-VF-NEXT:    [[MINMAX_0_COND_LCSSA:%.*]] = phi i32 [ [[MINMAX_0_COND:%.*]], [[SCALAR_BODY]] ], [ [[RDX_MINMAX]], [[MIDDLE_BLOCK]] ]
343; UNROLL-NO-VF-NEXT:    br label [[FOR_COND_CLEANUP]]
344; UNROLL-NO-VF:       for.cond.cleanup:
345; UNROLL-NO-VF-NEXT:    [[MINMAX_0_LCSSA:%.*]] = phi i32 [ poison, [[ENTRY:%.*]] ], [ [[MINMAX_0_COND_LCSSA]], [[FOR_COND_CLEANUP_LOOPEXIT]] ]
346; UNROLL-NO-VF-NEXT:    ret i32 [[MINMAX_0_LCSSA]]
347; UNROLL-NO-VF:       scalar.body:
348; UNROLL-NO-VF-NEXT:    [[TMP18:%.*]] = phi i32 [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ], [ [[TMP19:%.*]], [[SCALAR_BODY]] ]
349; UNROLL-NO-VF-NEXT:    [[INDVARS_IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[INDVARS_IV_NEXT:%.*]], [[SCALAR_BODY]] ]
350; UNROLL-NO-VF-NEXT:    [[MINMAX_028:%.*]] = phi i32 [ [[BC_MERGE_RDX]], [[SCALAR_PH]] ], [ [[MINMAX_0_COND]], [[SCALAR_BODY]] ]
351; UNROLL-NO-VF-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[INDVARS_IV]]
352; UNROLL-NO-VF-NEXT:    [[TMP19]] = load i32, ptr [[ARRAYIDX]], align 4
353; UNROLL-NO-VF-NEXT:    [[SUB3:%.*]] = sub nsw i32 [[TMP19]], [[TMP18]]
354; UNROLL-NO-VF-NEXT:    [[CMP4:%.*]] = icmp sgt i32 [[SUB3]], 0
355; UNROLL-NO-VF-NEXT:    [[COND:%.*]] = select i1 [[CMP4]], i32 [[SUB3]], i32 0
356; UNROLL-NO-VF-NEXT:    [[CMP5:%.*]] = icmp slt i32 [[MINMAX_028]], [[COND]]
357; UNROLL-NO-VF-NEXT:    [[MINMAX_0_COND]] = select i1 [[CMP5]], i32 [[MINMAX_028]], i32 [[COND]]
358; UNROLL-NO-VF-NEXT:    [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1
359; UNROLL-NO-VF-NEXT:    [[LFTR_WIDEIV:%.*]] = trunc i64 [[INDVARS_IV_NEXT]] to i32
360; UNROLL-NO-VF-NEXT:    [[EXITCOND:%.*]] = icmp eq i32 [[LFTR_WIDEIV]], [[N]]
361; UNROLL-NO-VF-NEXT:    br i1 [[EXITCOND]], label [[FOR_COND_CLEANUP_LOOPEXIT]], label [[SCALAR_BODY]], !llvm.loop [[LOOP5:![0-9]+]]
362;
363; SINK-AFTER-LABEL: @recurrence_2(
364; SINK-AFTER-NEXT:  entry:
365; SINK-AFTER-NEXT:    [[CMP27:%.*]] = icmp sgt i32 [[N:%.*]], 0
366; SINK-AFTER-NEXT:    br i1 [[CMP27]], label [[FOR_PREHEADER:%.*]], label [[FOR_COND_CLEANUP:%.*]]
367; SINK-AFTER:       for.preheader:
368; SINK-AFTER-NEXT:    [[ARRAYIDX2_PHI_TRANS_INSERT:%.*]] = getelementptr inbounds i32, ptr [[A:%.*]], i64 -1
369; SINK-AFTER-NEXT:    [[DOTPRE:%.*]] = load i32, ptr [[ARRAYIDX2_PHI_TRANS_INSERT]], align 4
370; SINK-AFTER-NEXT:    [[TMP0:%.*]] = zext i32 [[N]] to i64
371; SINK-AFTER-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[TMP0]], 4
372; SINK-AFTER-NEXT:    br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
373; SINK-AFTER:       vector.ph:
374; SINK-AFTER-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[TMP0]], 4
375; SINK-AFTER-NEXT:    [[N_VEC:%.*]] = sub i64 [[TMP0]], [[N_MOD_VF]]
376; SINK-AFTER-NEXT:    [[VECTOR_RECUR_INIT:%.*]] = insertelement <4 x i32> poison, i32 [[DOTPRE]], i32 3
377; SINK-AFTER-NEXT:    br label [[VECTOR_BODY:%.*]]
378; SINK-AFTER:       vector.body:
379; SINK-AFTER-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
380; SINK-AFTER-NEXT:    [[VECTOR_RECUR:%.*]] = phi <4 x i32> [ [[VECTOR_RECUR_INIT]], [[VECTOR_PH]] ], [ [[WIDE_LOAD:%.*]], [[VECTOR_BODY]] ]
381; SINK-AFTER-NEXT:    [[VEC_PHI:%.*]] = phi <4 x i32> [ poison, [[VECTOR_PH]] ], [ [[TMP9:%.*]], [[VECTOR_BODY]] ]
382; SINK-AFTER-NEXT:    [[TMP1:%.*]] = add i64 [[INDEX]], 0
383; SINK-AFTER-NEXT:    [[TMP2:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[TMP1]]
384; SINK-AFTER-NEXT:    [[TMP3:%.*]] = getelementptr inbounds i32, ptr [[TMP2]], i32 0
385; SINK-AFTER-NEXT:    [[WIDE_LOAD]] = load <4 x i32>, ptr [[TMP3]], align 4
386; SINK-AFTER-NEXT:    [[TMP4:%.*]] = shufflevector <4 x i32> [[VECTOR_RECUR]], <4 x i32> [[WIDE_LOAD]], <4 x i32> <i32 3, i32 4, i32 5, i32 6>
387; SINK-AFTER-NEXT:    [[TMP5:%.*]] = sub nsw <4 x i32> [[WIDE_LOAD]], [[TMP4]]
388; SINK-AFTER-NEXT:    [[TMP6:%.*]] = icmp sgt <4 x i32> [[TMP5]], zeroinitializer
389; SINK-AFTER-NEXT:    [[TMP7:%.*]] = select <4 x i1> [[TMP6]], <4 x i32> [[TMP5]], <4 x i32> zeroinitializer
390; SINK-AFTER-NEXT:    [[TMP8:%.*]] = icmp slt <4 x i32> [[VEC_PHI]], [[TMP7]]
391; SINK-AFTER-NEXT:    [[TMP9]] = select <4 x i1> [[TMP8]], <4 x i32> [[VEC_PHI]], <4 x i32> [[TMP7]]
392; SINK-AFTER-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4
393; SINK-AFTER-NEXT:    [[TMP10:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
394; SINK-AFTER-NEXT:    br i1 [[TMP10]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]]
395; SINK-AFTER:       middle.block:
396; SINK-AFTER-NEXT:    [[TMP11:%.*]] = call i32 @llvm.vector.reduce.smin.v4i32(<4 x i32> [[TMP9]])
397; SINK-AFTER-NEXT:    [[VECTOR_RECUR_EXTRACT:%.*]] = extractelement <4 x i32> [[WIDE_LOAD]], i32 3
398; SINK-AFTER-NEXT:    [[CMP_N:%.*]] = icmp eq i64 [[TMP0]], [[N_VEC]]
399; SINK-AFTER-NEXT:    br i1 [[CMP_N]], label [[FOR_COND_CLEANUP_LOOPEXIT:%.*]], label [[SCALAR_PH]]
400; SINK-AFTER:       scalar.ph:
401; SINK-AFTER-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i32 [ [[VECTOR_RECUR_EXTRACT]], [[MIDDLE_BLOCK]] ], [ [[DOTPRE]], [[FOR_PREHEADER]] ]
402; SINK-AFTER-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[FOR_PREHEADER]] ]
403; SINK-AFTER-NEXT:    [[BC_MERGE_RDX:%.*]] = phi i32 [ [[TMP11]], [[MIDDLE_BLOCK]] ], [ poison, [[FOR_PREHEADER]] ]
404; SINK-AFTER-NEXT:    br label [[SCALAR_BODY:%.*]]
405; SINK-AFTER:       for.cond.cleanup.loopexit:
406; SINK-AFTER-NEXT:    [[MINMAX_0_COND_LCSSA:%.*]] = phi i32 [ [[MINMAX_0_COND:%.*]], [[SCALAR_BODY]] ], [ [[TMP11]], [[MIDDLE_BLOCK]] ]
407; SINK-AFTER-NEXT:    br label [[FOR_COND_CLEANUP]]
408; SINK-AFTER:       for.cond.cleanup:
409; SINK-AFTER-NEXT:    [[MINMAX_0_LCSSA:%.*]] = phi i32 [ poison, [[ENTRY:%.*]] ], [ [[MINMAX_0_COND_LCSSA]], [[FOR_COND_CLEANUP_LOOPEXIT]] ]
410; SINK-AFTER-NEXT:    ret i32 [[MINMAX_0_LCSSA]]
411; SINK-AFTER:       scalar.body:
412; SINK-AFTER-NEXT:    [[TMP12:%.*]] = phi i32 [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ], [ [[TMP13:%.*]], [[SCALAR_BODY]] ]
413; SINK-AFTER-NEXT:    [[INDVARS_IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[INDVARS_IV_NEXT:%.*]], [[SCALAR_BODY]] ]
414; SINK-AFTER-NEXT:    [[MINMAX_028:%.*]] = phi i32 [ [[BC_MERGE_RDX]], [[SCALAR_PH]] ], [ [[MINMAX_0_COND]], [[SCALAR_BODY]] ]
415; SINK-AFTER-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[INDVARS_IV]]
416; SINK-AFTER-NEXT:    [[TMP13]] = load i32, ptr [[ARRAYIDX]], align 4
417; SINK-AFTER-NEXT:    [[SUB3:%.*]] = sub nsw i32 [[TMP13]], [[TMP12]]
418; SINK-AFTER-NEXT:    [[CMP4:%.*]] = icmp sgt i32 [[SUB3]], 0
419; SINK-AFTER-NEXT:    [[COND:%.*]] = select i1 [[CMP4]], i32 [[SUB3]], i32 0
420; SINK-AFTER-NEXT:    [[CMP5:%.*]] = icmp slt i32 [[MINMAX_028]], [[COND]]
421; SINK-AFTER-NEXT:    [[MINMAX_0_COND]] = select i1 [[CMP5]], i32 [[MINMAX_028]], i32 [[COND]]
422; SINK-AFTER-NEXT:    [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1
423; SINK-AFTER-NEXT:    [[LFTR_WIDEIV:%.*]] = trunc i64 [[INDVARS_IV_NEXT]] to i32
424; SINK-AFTER-NEXT:    [[EXITCOND:%.*]] = icmp eq i32 [[LFTR_WIDEIV]], [[N]]
425; SINK-AFTER-NEXT:    br i1 [[EXITCOND]], label [[FOR_COND_CLEANUP_LOOPEXIT]], label [[SCALAR_BODY]], !llvm.loop [[LOOP5:![0-9]+]]
426;
427entry:
428  %cmp27 = icmp sgt i32 %n, 0
429  br i1 %cmp27, label %for.preheader, label %for.cond.cleanup
430
431for.preheader:
432  %arrayidx2.phi.trans.insert = getelementptr inbounds i32, ptr %a, i64 -1
433  %.pre = load i32, ptr %arrayidx2.phi.trans.insert, align 4
434  br label %scalar.body
435
436for.cond.cleanup.loopexit:
437  %minmax.0.cond.lcssa = phi i32 [ %minmax.0.cond, %scalar.body ]
438  br label %for.cond.cleanup
439
440for.cond.cleanup:
441  %minmax.0.lcssa = phi i32 [ poison, %entry ], [ %minmax.0.cond.lcssa, %for.cond.cleanup.loopexit ]
442  ret i32 %minmax.0.lcssa
443
444scalar.body:
445  %0 = phi i32 [ %.pre, %for.preheader ], [ %1, %scalar.body ]
446  %indvars.iv = phi i64 [ 0, %for.preheader ], [ %indvars.iv.next, %scalar.body ]
447  %minmax.028 = phi i32 [ poison, %for.preheader ], [ %minmax.0.cond, %scalar.body ]
448  %arrayidx = getelementptr inbounds i32, ptr %a, i64 %indvars.iv
449  %1 = load i32, ptr %arrayidx, align 4
450  %sub3 = sub nsw i32 %1, %0
451  %cmp4 = icmp sgt i32 %sub3, 0
452  %cond = select i1 %cmp4, i32 %sub3, i32 0
453  %cmp5 = icmp slt i32 %minmax.028, %cond
454  %minmax.0.cond = select i1 %cmp5, i32 %minmax.028, i32 %cond
455  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
456  %lftr.wideiv = trunc i64 %indvars.iv.next to i32
457  %exitcond = icmp eq i32 %lftr.wideiv, %n
458  br i1 %exitcond, label %for.cond.cleanup.loopexit, label %scalar.body
459}
460
461; void recurrence_3(short *a, ptr b, int n, float f, short p) {
462;   b[0] = (double)a[0] - f * (double)p;
463;   for (int i = 1; i < n; i++)
464;     b[i] = (double)a[i] - f * (double)a[i - 1];
465; }
466;
467; Check also that the casts were not moved needlessly.
468;
469;
470define void @recurrence_3(ptr readonly noalias %a, ptr noalias %b, i32 %n, float %f, i16 %p) {
471; UNROLL-NO-IC-LABEL: @recurrence_3(
472; UNROLL-NO-IC-NEXT:  entry:
473; UNROLL-NO-IC-NEXT:    [[TMP0:%.*]] = load i16, ptr [[A:%.*]], align 2
474; UNROLL-NO-IC-NEXT:    [[CONV:%.*]] = sitofp i16 [[TMP0]] to double
475; UNROLL-NO-IC-NEXT:    [[CONV1:%.*]] = fpext float [[F:%.*]] to double
476; UNROLL-NO-IC-NEXT:    [[CONV2:%.*]] = sitofp i16 [[P:%.*]] to double
477; UNROLL-NO-IC-NEXT:    [[MUL:%.*]] = fmul fast double [[CONV2]], [[CONV1]]
478; UNROLL-NO-IC-NEXT:    [[SUB:%.*]] = fsub fast double [[CONV]], [[MUL]]
479; UNROLL-NO-IC-NEXT:    store double [[SUB]], ptr [[B:%.*]], align 8
480; UNROLL-NO-IC-NEXT:    [[CMP25:%.*]] = icmp sgt i32 [[N:%.*]], 1
481; UNROLL-NO-IC-NEXT:    br i1 [[CMP25]], label [[FOR_PREHEADER:%.*]], label [[FOR_END:%.*]]
482; UNROLL-NO-IC:       for.preheader:
483; UNROLL-NO-IC-NEXT:    [[TMP1:%.*]] = add i32 [[N]], -1
484; UNROLL-NO-IC-NEXT:    [[TMP2:%.*]] = zext i32 [[TMP1]] to i64
485; UNROLL-NO-IC-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[TMP2]], 8
486; UNROLL-NO-IC-NEXT:    br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
487; UNROLL-NO-IC:       vector.ph:
488; UNROLL-NO-IC-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[TMP2]], 8
489; UNROLL-NO-IC-NEXT:    [[N_VEC:%.*]] = sub i64 [[TMP2]], [[N_MOD_VF]]
490; UNROLL-NO-IC-NEXT:    [[IND_END:%.*]] = add i64 1, [[N_VEC]]
491; UNROLL-NO-IC-NEXT:    [[VECTOR_RECUR_INIT:%.*]] = insertelement <4 x i16> poison, i16 [[TMP0]], i32 3
492; UNROLL-NO-IC-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <4 x double> poison, double [[CONV1]], i64 0
493; UNROLL-NO-IC-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <4 x double> [[BROADCAST_SPLATINSERT]], <4 x double> poison, <4 x i32> zeroinitializer
494; UNROLL-NO-IC-NEXT:    br label [[VECTOR_BODY:%.*]]
495; UNROLL-NO-IC:       vector.body:
496; UNROLL-NO-IC-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
497; UNROLL-NO-IC-NEXT:    [[VECTOR_RECUR:%.*]] = phi <4 x i16> [ [[VECTOR_RECUR_INIT]], [[VECTOR_PH]] ], [ [[WIDE_LOAD1:%.*]], [[VECTOR_BODY]] ]
498; UNROLL-NO-IC-NEXT:    [[OFFSET_IDX:%.*]] = add i64 1, [[INDEX]]
499; UNROLL-NO-IC-NEXT:    [[TMP3:%.*]] = add i64 [[OFFSET_IDX]], 0
500; UNROLL-NO-IC-NEXT:    [[TMP4:%.*]] = getelementptr inbounds i16, ptr [[A]], i64 [[TMP3]]
501; UNROLL-NO-IC-NEXT:    [[TMP5:%.*]] = getelementptr inbounds i16, ptr [[TMP4]], i32 0
502; UNROLL-NO-IC-NEXT:    [[TMP6:%.*]] = getelementptr inbounds i16, ptr [[TMP4]], i32 4
503; UNROLL-NO-IC-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i16>, ptr [[TMP5]], align 2
504; UNROLL-NO-IC-NEXT:    [[WIDE_LOAD1]] = load <4 x i16>, ptr [[TMP6]], align 2
505; UNROLL-NO-IC-NEXT:    [[TMP7:%.*]] = shufflevector <4 x i16> [[VECTOR_RECUR]], <4 x i16> [[WIDE_LOAD]], <4 x i32> <i32 3, i32 4, i32 5, i32 6>
506; UNROLL-NO-IC-NEXT:    [[TMP8:%.*]] = shufflevector <4 x i16> [[WIDE_LOAD]], <4 x i16> [[WIDE_LOAD1]], <4 x i32> <i32 3, i32 4, i32 5, i32 6>
507; UNROLL-NO-IC-NEXT:    [[TMP9:%.*]] = sitofp <4 x i16> [[WIDE_LOAD]] to <4 x double>
508; UNROLL-NO-IC-NEXT:    [[TMP10:%.*]] = sitofp <4 x i16> [[WIDE_LOAD1]] to <4 x double>
509; UNROLL-NO-IC-NEXT:    [[TMP11:%.*]] = sitofp <4 x i16> [[TMP7]] to <4 x double>
510; UNROLL-NO-IC-NEXT:    [[TMP12:%.*]] = sitofp <4 x i16> [[TMP8]] to <4 x double>
511; UNROLL-NO-IC-NEXT:    [[TMP13:%.*]] = fmul fast <4 x double> [[TMP11]], [[BROADCAST_SPLAT]]
512; UNROLL-NO-IC-NEXT:    [[TMP14:%.*]] = fmul fast <4 x double> [[TMP12]], [[BROADCAST_SPLAT]]
513; UNROLL-NO-IC-NEXT:    [[TMP15:%.*]] = fsub fast <4 x double> [[TMP9]], [[TMP13]]
514; UNROLL-NO-IC-NEXT:    [[TMP16:%.*]] = fsub fast <4 x double> [[TMP10]], [[TMP14]]
515; UNROLL-NO-IC-NEXT:    [[TMP17:%.*]] = getelementptr inbounds double, ptr [[B]], i64 [[TMP3]]
516; UNROLL-NO-IC-NEXT:    [[TMP18:%.*]] = getelementptr inbounds double, ptr [[TMP17]], i32 0
517; UNROLL-NO-IC-NEXT:    [[TMP19:%.*]] = getelementptr inbounds double, ptr [[TMP17]], i32 4
518; UNROLL-NO-IC-NEXT:    store <4 x double> [[TMP15]], ptr [[TMP18]], align 8
519; UNROLL-NO-IC-NEXT:    store <4 x double> [[TMP16]], ptr [[TMP19]], align 8
520; UNROLL-NO-IC-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 8
521; UNROLL-NO-IC-NEXT:    [[TMP20:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
522; UNROLL-NO-IC-NEXT:    br i1 [[TMP20]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP6:![0-9]+]]
523; UNROLL-NO-IC:       middle.block:
524; UNROLL-NO-IC-NEXT:    [[VECTOR_RECUR_EXTRACT:%.*]] = extractelement <4 x i16> [[WIDE_LOAD1]], i32 3
525; UNROLL-NO-IC-NEXT:    [[CMP_N:%.*]] = icmp eq i64 [[TMP2]], [[N_VEC]]
526; UNROLL-NO-IC-NEXT:    br i1 [[CMP_N]], label [[FOR_END_LOOPEXIT:%.*]], label [[SCALAR_PH]]
527; UNROLL-NO-IC:       scalar.ph:
528; UNROLL-NO-IC-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i16 [ [[VECTOR_RECUR_EXTRACT]], [[MIDDLE_BLOCK]] ], [ [[TMP0]], [[FOR_PREHEADER]] ]
529; UNROLL-NO-IC-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ [[IND_END]], [[MIDDLE_BLOCK]] ], [ 1, [[FOR_PREHEADER]] ]
530; UNROLL-NO-IC-NEXT:    br label [[SCALAR_BODY:%.*]]
531; UNROLL-NO-IC:       scalar.body:
532; UNROLL-NO-IC-NEXT:    [[TMP21:%.*]] = phi i16 [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ], [ [[TMP22:%.*]], [[SCALAR_BODY]] ]
533; UNROLL-NO-IC-NEXT:    [[ADVARS_IV:%.*]] = phi i64 [ [[ADVARS_IV_NEXT:%.*]], [[SCALAR_BODY]] ], [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ]
534; UNROLL-NO-IC-NEXT:    [[ARRAYIDX5:%.*]] = getelementptr inbounds i16, ptr [[A]], i64 [[ADVARS_IV]]
535; UNROLL-NO-IC-NEXT:    [[TMP22]] = load i16, ptr [[ARRAYIDX5]], align 2
536; UNROLL-NO-IC-NEXT:    [[CONV6:%.*]] = sitofp i16 [[TMP22]] to double
537; UNROLL-NO-IC-NEXT:    [[CONV11:%.*]] = sitofp i16 [[TMP21]] to double
538; UNROLL-NO-IC-NEXT:    [[MUL12:%.*]] = fmul fast double [[CONV11]], [[CONV1]]
539; UNROLL-NO-IC-NEXT:    [[SUB13:%.*]] = fsub fast double [[CONV6]], [[MUL12]]
540; UNROLL-NO-IC-NEXT:    [[ARRAYIDX15:%.*]] = getelementptr inbounds double, ptr [[B]], i64 [[ADVARS_IV]]
541; UNROLL-NO-IC-NEXT:    store double [[SUB13]], ptr [[ARRAYIDX15]], align 8
542; UNROLL-NO-IC-NEXT:    [[ADVARS_IV_NEXT]] = add nuw nsw i64 [[ADVARS_IV]], 1
543; UNROLL-NO-IC-NEXT:    [[LFTR_WIDEIV:%.*]] = trunc i64 [[ADVARS_IV_NEXT]] to i32
544; UNROLL-NO-IC-NEXT:    [[EXITCOND:%.*]] = icmp eq i32 [[LFTR_WIDEIV]], [[N]]
545; UNROLL-NO-IC-NEXT:    br i1 [[EXITCOND]], label [[FOR_END_LOOPEXIT]], label [[SCALAR_BODY]], !llvm.loop [[LOOP7:![0-9]+]]
546; UNROLL-NO-IC:       for.end.loopexit:
547; UNROLL-NO-IC-NEXT:    br label [[FOR_END]]
548; UNROLL-NO-IC:       for.end:
549; UNROLL-NO-IC-NEXT:    ret void
550;
551; UNROLL-NO-VF-LABEL: @recurrence_3(
552; UNROLL-NO-VF-NEXT:  entry:
553; UNROLL-NO-VF-NEXT:    [[TMP0:%.*]] = load i16, ptr [[A:%.*]], align 2
554; UNROLL-NO-VF-NEXT:    [[CONV:%.*]] = sitofp i16 [[TMP0]] to double
555; UNROLL-NO-VF-NEXT:    [[CONV1:%.*]] = fpext float [[F:%.*]] to double
556; UNROLL-NO-VF-NEXT:    [[CONV2:%.*]] = sitofp i16 [[P:%.*]] to double
557; UNROLL-NO-VF-NEXT:    [[MUL:%.*]] = fmul fast double [[CONV2]], [[CONV1]]
558; UNROLL-NO-VF-NEXT:    [[SUB:%.*]] = fsub fast double [[CONV]], [[MUL]]
559; UNROLL-NO-VF-NEXT:    store double [[SUB]], ptr [[B:%.*]], align 8
560; UNROLL-NO-VF-NEXT:    [[CMP25:%.*]] = icmp sgt i32 [[N:%.*]], 1
561; UNROLL-NO-VF-NEXT:    br i1 [[CMP25]], label [[FOR_PREHEADER:%.*]], label [[FOR_END:%.*]]
562; UNROLL-NO-VF:       for.preheader:
563; UNROLL-NO-VF-NEXT:    [[TMP1:%.*]] = add i32 [[N]], -1
564; UNROLL-NO-VF-NEXT:    [[TMP2:%.*]] = zext i32 [[TMP1]] to i64
565; UNROLL-NO-VF-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[TMP2]], 2
566; UNROLL-NO-VF-NEXT:    br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
567; UNROLL-NO-VF:       vector.ph:
568; UNROLL-NO-VF-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[TMP2]], 2
569; UNROLL-NO-VF-NEXT:    [[N_VEC:%.*]] = sub i64 [[TMP2]], [[N_MOD_VF]]
570; UNROLL-NO-VF-NEXT:    [[IND_END:%.*]] = add i64 1, [[N_VEC]]
571; UNROLL-NO-VF-NEXT:    br label [[VECTOR_BODY:%.*]]
572; UNROLL-NO-VF:       vector.body:
573; UNROLL-NO-VF-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
574; UNROLL-NO-VF-NEXT:    [[VECTOR_RECUR:%.*]] = phi i16 [ [[TMP0]], [[VECTOR_PH]] ], [ [[TMP8:%.*]], [[VECTOR_BODY]] ]
575; UNROLL-NO-VF-NEXT:    [[OFFSET_IDX:%.*]] = add i64 1, [[INDEX]]
576; UNROLL-NO-VF-NEXT:    [[TMP3:%.*]] = add i64 [[OFFSET_IDX]], 0
577; UNROLL-NO-VF-NEXT:    [[TMP4:%.*]] = add i64 [[OFFSET_IDX]], 1
578; UNROLL-NO-VF-NEXT:    [[TMP5:%.*]] = getelementptr inbounds i16, ptr [[A]], i64 [[TMP3]]
579; UNROLL-NO-VF-NEXT:    [[TMP6:%.*]] = getelementptr inbounds i16, ptr [[A]], i64 [[TMP4]]
580; UNROLL-NO-VF-NEXT:    [[TMP7:%.*]] = load i16, ptr [[TMP5]], align 2
581; UNROLL-NO-VF-NEXT:    [[TMP8]] = load i16, ptr [[TMP6]], align 2
582; UNROLL-NO-VF-NEXT:    [[TMP9:%.*]] = sitofp i16 [[TMP7]] to double
583; UNROLL-NO-VF-NEXT:    [[TMP10:%.*]] = sitofp i16 [[TMP8]] to double
584; UNROLL-NO-VF-NEXT:    [[TMP11:%.*]] = sitofp i16 [[VECTOR_RECUR]] to double
585; UNROLL-NO-VF-NEXT:    [[TMP12:%.*]] = sitofp i16 [[TMP7]] to double
586; UNROLL-NO-VF-NEXT:    [[TMP13:%.*]] = fmul fast double [[TMP11]], [[CONV1]]
587; UNROLL-NO-VF-NEXT:    [[TMP14:%.*]] = fmul fast double [[TMP12]], [[CONV1]]
588; UNROLL-NO-VF-NEXT:    [[TMP15:%.*]] = fsub fast double [[TMP9]], [[TMP13]]
589; UNROLL-NO-VF-NEXT:    [[TMP16:%.*]] = fsub fast double [[TMP10]], [[TMP14]]
590; UNROLL-NO-VF-NEXT:    [[TMP17:%.*]] = getelementptr inbounds double, ptr [[B]], i64 [[TMP3]]
591; UNROLL-NO-VF-NEXT:    [[TMP18:%.*]] = getelementptr inbounds double, ptr [[B]], i64 [[TMP4]]
592; UNROLL-NO-VF-NEXT:    store double [[TMP15]], ptr [[TMP17]], align 8
593; UNROLL-NO-VF-NEXT:    store double [[TMP16]], ptr [[TMP18]], align 8
594; UNROLL-NO-VF-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 2
595; UNROLL-NO-VF-NEXT:    [[TMP19:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
596; UNROLL-NO-VF-NEXT:    br i1 [[TMP19]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP6:![0-9]+]]
597; UNROLL-NO-VF:       middle.block:
598; UNROLL-NO-VF-NEXT:    [[CMP_N:%.*]] = icmp eq i64 [[TMP2]], [[N_VEC]]
599; UNROLL-NO-VF-NEXT:    br i1 [[CMP_N]], label [[FOR_END_LOOPEXIT:%.*]], label [[SCALAR_PH]]
600; UNROLL-NO-VF:       scalar.ph:
601; UNROLL-NO-VF-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i16 [ [[TMP8]], [[MIDDLE_BLOCK]] ], [ [[TMP0]], [[FOR_PREHEADER]] ]
602; UNROLL-NO-VF-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ [[IND_END]], [[MIDDLE_BLOCK]] ], [ 1, [[FOR_PREHEADER]] ]
603; UNROLL-NO-VF-NEXT:    br label [[SCALAR_BODY:%.*]]
604; UNROLL-NO-VF:       scalar.body:
605; UNROLL-NO-VF-NEXT:    [[TMP20:%.*]] = phi i16 [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ], [ [[TMP21:%.*]], [[SCALAR_BODY]] ]
606; UNROLL-NO-VF-NEXT:    [[ADVARS_IV:%.*]] = phi i64 [ [[ADVARS_IV_NEXT:%.*]], [[SCALAR_BODY]] ], [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ]
607; UNROLL-NO-VF-NEXT:    [[ARRAYIDX5:%.*]] = getelementptr inbounds i16, ptr [[A]], i64 [[ADVARS_IV]]
608; UNROLL-NO-VF-NEXT:    [[TMP21]] = load i16, ptr [[ARRAYIDX5]], align 2
609; UNROLL-NO-VF-NEXT:    [[CONV6:%.*]] = sitofp i16 [[TMP21]] to double
610; UNROLL-NO-VF-NEXT:    [[CONV11:%.*]] = sitofp i16 [[TMP20]] to double
611; UNROLL-NO-VF-NEXT:    [[MUL12:%.*]] = fmul fast double [[CONV11]], [[CONV1]]
612; UNROLL-NO-VF-NEXT:    [[SUB13:%.*]] = fsub fast double [[CONV6]], [[MUL12]]
613; UNROLL-NO-VF-NEXT:    [[ARRAYIDX15:%.*]] = getelementptr inbounds double, ptr [[B]], i64 [[ADVARS_IV]]
614; UNROLL-NO-VF-NEXT:    store double [[SUB13]], ptr [[ARRAYIDX15]], align 8
615; UNROLL-NO-VF-NEXT:    [[ADVARS_IV_NEXT]] = add nuw nsw i64 [[ADVARS_IV]], 1
616; UNROLL-NO-VF-NEXT:    [[LFTR_WIDEIV:%.*]] = trunc i64 [[ADVARS_IV_NEXT]] to i32
617; UNROLL-NO-VF-NEXT:    [[EXITCOND:%.*]] = icmp eq i32 [[LFTR_WIDEIV]], [[N]]
618; UNROLL-NO-VF-NEXT:    br i1 [[EXITCOND]], label [[FOR_END_LOOPEXIT]], label [[SCALAR_BODY]], !llvm.loop [[LOOP7:![0-9]+]]
619; UNROLL-NO-VF:       for.end.loopexit:
620; UNROLL-NO-VF-NEXT:    br label [[FOR_END]]
621; UNROLL-NO-VF:       for.end:
622; UNROLL-NO-VF-NEXT:    ret void
623;
624; SINK-AFTER-LABEL: @recurrence_3(
625; SINK-AFTER-NEXT:  entry:
626; SINK-AFTER-NEXT:    [[TMP0:%.*]] = load i16, ptr [[A:%.*]], align 2
627; SINK-AFTER-NEXT:    [[CONV:%.*]] = sitofp i16 [[TMP0]] to double
628; SINK-AFTER-NEXT:    [[CONV1:%.*]] = fpext float [[F:%.*]] to double
629; SINK-AFTER-NEXT:    [[CONV2:%.*]] = sitofp i16 [[P:%.*]] to double
630; SINK-AFTER-NEXT:    [[MUL:%.*]] = fmul fast double [[CONV2]], [[CONV1]]
631; SINK-AFTER-NEXT:    [[SUB:%.*]] = fsub fast double [[CONV]], [[MUL]]
632; SINK-AFTER-NEXT:    store double [[SUB]], ptr [[B:%.*]], align 8
633; SINK-AFTER-NEXT:    [[CMP25:%.*]] = icmp sgt i32 [[N:%.*]], 1
634; SINK-AFTER-NEXT:    br i1 [[CMP25]], label [[FOR_PREHEADER:%.*]], label [[FOR_END:%.*]]
635; SINK-AFTER:       for.preheader:
636; SINK-AFTER-NEXT:    [[TMP1:%.*]] = add i32 [[N]], -1
637; SINK-AFTER-NEXT:    [[TMP2:%.*]] = zext i32 [[TMP1]] to i64
638; SINK-AFTER-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[TMP2]], 4
639; SINK-AFTER-NEXT:    br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
640; SINK-AFTER:       vector.ph:
641; SINK-AFTER-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[TMP2]], 4
642; SINK-AFTER-NEXT:    [[N_VEC:%.*]] = sub i64 [[TMP2]], [[N_MOD_VF]]
643; SINK-AFTER-NEXT:    [[IND_END:%.*]] = add i64 1, [[N_VEC]]
644; SINK-AFTER-NEXT:    [[VECTOR_RECUR_INIT:%.*]] = insertelement <4 x i16> poison, i16 [[TMP0]], i32 3
645; SINK-AFTER-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <4 x double> poison, double [[CONV1]], i64 0
646; SINK-AFTER-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <4 x double> [[BROADCAST_SPLATINSERT]], <4 x double> poison, <4 x i32> zeroinitializer
647; SINK-AFTER-NEXT:    br label [[VECTOR_BODY:%.*]]
648; SINK-AFTER:       vector.body:
649; SINK-AFTER-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
650; SINK-AFTER-NEXT:    [[VECTOR_RECUR:%.*]] = phi <4 x i16> [ [[VECTOR_RECUR_INIT]], [[VECTOR_PH]] ], [ [[WIDE_LOAD:%.*]], [[VECTOR_BODY]] ]
651; SINK-AFTER-NEXT:    [[OFFSET_IDX:%.*]] = add i64 1, [[INDEX]]
652; SINK-AFTER-NEXT:    [[TMP3:%.*]] = add i64 [[OFFSET_IDX]], 0
653; SINK-AFTER-NEXT:    [[TMP4:%.*]] = getelementptr inbounds i16, ptr [[A]], i64 [[TMP3]]
654; SINK-AFTER-NEXT:    [[TMP5:%.*]] = getelementptr inbounds i16, ptr [[TMP4]], i32 0
655; SINK-AFTER-NEXT:    [[WIDE_LOAD]] = load <4 x i16>, ptr [[TMP5]], align 2
656; SINK-AFTER-NEXT:    [[TMP6:%.*]] = shufflevector <4 x i16> [[VECTOR_RECUR]], <4 x i16> [[WIDE_LOAD]], <4 x i32> <i32 3, i32 4, i32 5, i32 6>
657; SINK-AFTER-NEXT:    [[TMP7:%.*]] = sitofp <4 x i16> [[WIDE_LOAD]] to <4 x double>
658; SINK-AFTER-NEXT:    [[TMP8:%.*]] = sitofp <4 x i16> [[TMP6]] to <4 x double>
659; SINK-AFTER-NEXT:    [[TMP9:%.*]] = fmul fast <4 x double> [[TMP8]], [[BROADCAST_SPLAT]]
660; SINK-AFTER-NEXT:    [[TMP10:%.*]] = fsub fast <4 x double> [[TMP7]], [[TMP9]]
661; SINK-AFTER-NEXT:    [[TMP11:%.*]] = getelementptr inbounds double, ptr [[B]], i64 [[TMP3]]
662; SINK-AFTER-NEXT:    [[TMP12:%.*]] = getelementptr inbounds double, ptr [[TMP11]], i32 0
663; SINK-AFTER-NEXT:    store <4 x double> [[TMP10]], ptr [[TMP12]], align 8
664; SINK-AFTER-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4
665; SINK-AFTER-NEXT:    [[TMP13:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
666; SINK-AFTER-NEXT:    br i1 [[TMP13]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP6:![0-9]+]]
667; SINK-AFTER:       middle.block:
668; SINK-AFTER-NEXT:    [[VECTOR_RECUR_EXTRACT:%.*]] = extractelement <4 x i16> [[WIDE_LOAD]], i32 3
669; SINK-AFTER-NEXT:    [[CMP_N:%.*]] = icmp eq i64 [[TMP2]], [[N_VEC]]
670; SINK-AFTER-NEXT:    br i1 [[CMP_N]], label [[FOR_END_LOOPEXIT:%.*]], label [[SCALAR_PH]]
671; SINK-AFTER:       scalar.ph:
672; SINK-AFTER-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i16 [ [[VECTOR_RECUR_EXTRACT]], [[MIDDLE_BLOCK]] ], [ [[TMP0]], [[FOR_PREHEADER]] ]
673; SINK-AFTER-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ [[IND_END]], [[MIDDLE_BLOCK]] ], [ 1, [[FOR_PREHEADER]] ]
674; SINK-AFTER-NEXT:    br label [[SCALAR_BODY:%.*]]
675; SINK-AFTER:       scalar.body:
676; SINK-AFTER-NEXT:    [[TMP14:%.*]] = phi i16 [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ], [ [[TMP15:%.*]], [[SCALAR_BODY]] ]
677; SINK-AFTER-NEXT:    [[ADVARS_IV:%.*]] = phi i64 [ [[ADVARS_IV_NEXT:%.*]], [[SCALAR_BODY]] ], [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ]
678; SINK-AFTER-NEXT:    [[ARRAYIDX5:%.*]] = getelementptr inbounds i16, ptr [[A]], i64 [[ADVARS_IV]]
679; SINK-AFTER-NEXT:    [[TMP15]] = load i16, ptr [[ARRAYIDX5]], align 2
680; SINK-AFTER-NEXT:    [[CONV6:%.*]] = sitofp i16 [[TMP15]] to double
681; SINK-AFTER-NEXT:    [[CONV11:%.*]] = sitofp i16 [[TMP14]] to double
682; SINK-AFTER-NEXT:    [[MUL12:%.*]] = fmul fast double [[CONV11]], [[CONV1]]
683; SINK-AFTER-NEXT:    [[SUB13:%.*]] = fsub fast double [[CONV6]], [[MUL12]]
684; SINK-AFTER-NEXT:    [[ARRAYIDX15:%.*]] = getelementptr inbounds double, ptr [[B]], i64 [[ADVARS_IV]]
685; SINK-AFTER-NEXT:    store double [[SUB13]], ptr [[ARRAYIDX15]], align 8
686; SINK-AFTER-NEXT:    [[ADVARS_IV_NEXT]] = add nuw nsw i64 [[ADVARS_IV]], 1
687; SINK-AFTER-NEXT:    [[LFTR_WIDEIV:%.*]] = trunc i64 [[ADVARS_IV_NEXT]] to i32
688; SINK-AFTER-NEXT:    [[EXITCOND:%.*]] = icmp eq i32 [[LFTR_WIDEIV]], [[N]]
689; SINK-AFTER-NEXT:    br i1 [[EXITCOND]], label [[FOR_END_LOOPEXIT]], label [[SCALAR_BODY]], !llvm.loop [[LOOP7:![0-9]+]]
690; SINK-AFTER:       for.end.loopexit:
691; SINK-AFTER-NEXT:    br label [[FOR_END]]
692; SINK-AFTER:       for.end:
693; SINK-AFTER-NEXT:    ret void
694;
695entry:
696  %0 = load i16, ptr %a, align 2
697  %conv = sitofp i16 %0 to double
698  %conv1 = fpext float %f to double
699  %conv2 = sitofp i16 %p to double
700  %mul = fmul fast double %conv2, %conv1
701  %sub = fsub fast double %conv, %mul
702  store double %sub, ptr %b, align 8
703  %cmp25 = icmp sgt i32 %n, 1
704  br i1 %cmp25, label %for.preheader, label %for.end
705
706for.preheader:
707  br label %scalar.body
708
709scalar.body:
710  %1 = phi i16 [ %0, %for.preheader ], [ %2, %scalar.body ]
711  %advars.iv = phi i64 [ %advars.iv.next, %scalar.body ], [ 1, %for.preheader ]
712  %arrayidx5 = getelementptr inbounds i16, ptr %a, i64 %advars.iv
713  %2 = load i16, ptr %arrayidx5, align 2
714  %conv6 = sitofp i16 %2 to double
715  %conv11 = sitofp i16 %1 to double
716  %mul12 = fmul fast double %conv11, %conv1
717  %sub13 = fsub fast double %conv6, %mul12
718  %arrayidx15 = getelementptr inbounds double, ptr %b, i64 %advars.iv
719  store double %sub13, ptr %arrayidx15, align 8
720  %advars.iv.next = add nuw nsw i64 %advars.iv, 1
721  %lftr.wideiv = trunc i64 %advars.iv.next to i32
722  %exitcond = icmp eq i32 %lftr.wideiv, %n
723  br i1 %exitcond, label %for.end.loopexit, label %scalar.body
724
725for.end.loopexit:
726  br label %for.end
727
728for.end:
729  ret void
730}
731
732; void PR26734(short *a, int *b, int *c, int d, short *e) {
733;   for (; d != 21; d++) {
734;     *b &= *c;
735;     *e = *a - 6;
736;     *c = *e;
737;   }
738; }
739;
740;
741define void @PR26734(ptr %a, ptr %b, ptr %c, i32 %d, ptr %e) {
742; UNROLL-NO-IC-LABEL: @PR26734(
743; UNROLL-NO-IC-NEXT:  entry:
744; UNROLL-NO-IC-NEXT:    [[CMP4:%.*]] = icmp eq i32 [[D:%.*]], 21
745; UNROLL-NO-IC-NEXT:    br i1 [[CMP4]], label [[ENTRY_FOR_END_CRIT_EDGE:%.*]], label [[FOR_BODY_LR_PH:%.*]]
746; UNROLL-NO-IC:       entry.for.end_crit_edge:
747; UNROLL-NO-IC-NEXT:    [[DOTPRE:%.*]] = load i32, ptr [[B:%.*]], align 4
748; UNROLL-NO-IC-NEXT:    br label [[FOR_END:%.*]]
749; UNROLL-NO-IC:       for.body.lr.ph:
750; UNROLL-NO-IC-NEXT:    [[TMP0:%.*]] = load i16, ptr [[A:%.*]], align 2
751; UNROLL-NO-IC-NEXT:    [[SUB:%.*]] = add i16 [[TMP0]], -6
752; UNROLL-NO-IC-NEXT:    [[CONV2:%.*]] = sext i16 [[SUB]] to i32
753; UNROLL-NO-IC-NEXT:    [[C_PROMOTED:%.*]] = load i32, ptr [[C:%.*]], align 4
754; UNROLL-NO-IC-NEXT:    [[B_PROMOTED:%.*]] = load i32, ptr [[B]], align 4
755; UNROLL-NO-IC-NEXT:    br label [[FOR_BODY:%.*]]
756; UNROLL-NO-IC:       for.body:
757; UNROLL-NO-IC-NEXT:    [[INC7:%.*]] = phi i32 [ [[D]], [[FOR_BODY_LR_PH]] ], [ [[INC:%.*]], [[FOR_BODY]] ]
758; UNROLL-NO-IC-NEXT:    [[AND6:%.*]] = phi i32 [ [[B_PROMOTED]], [[FOR_BODY_LR_PH]] ], [ [[AND:%.*]], [[FOR_BODY]] ]
759; UNROLL-NO-IC-NEXT:    [[CONV25:%.*]] = phi i32 [ [[C_PROMOTED]], [[FOR_BODY_LR_PH]] ], [ [[CONV2]], [[FOR_BODY]] ]
760; UNROLL-NO-IC-NEXT:    [[AND]] = and i32 [[AND6]], [[CONV25]]
761; UNROLL-NO-IC-NEXT:    [[INC]] = add nsw i32 [[INC7]], 1
762; UNROLL-NO-IC-NEXT:    [[CMP:%.*]] = icmp eq i32 [[INC]], 21
763; UNROLL-NO-IC-NEXT:    br i1 [[CMP]], label [[FOR_COND_FOR_END_CRIT_EDGE:%.*]], label [[FOR_BODY]]
764; UNROLL-NO-IC:       for.cond.for.end_crit_edge:
765; UNROLL-NO-IC-NEXT:    [[AND_LCSSA:%.*]] = phi i32 [ [[AND]], [[FOR_BODY]] ]
766; UNROLL-NO-IC-NEXT:    store i32 [[CONV2]], ptr [[C]], align 4
767; UNROLL-NO-IC-NEXT:    store i32 [[AND_LCSSA]], ptr [[B]], align 4
768; UNROLL-NO-IC-NEXT:    store i16 [[SUB]], ptr [[E:%.*]], align 2
769; UNROLL-NO-IC-NEXT:    br label [[FOR_END]]
770; UNROLL-NO-IC:       for.end:
771; UNROLL-NO-IC-NEXT:    ret void
772;
773; UNROLL-NO-VF-LABEL: @PR26734(
774; UNROLL-NO-VF-NEXT:  entry:
775; UNROLL-NO-VF-NEXT:    [[CMP4:%.*]] = icmp eq i32 [[D:%.*]], 21
776; UNROLL-NO-VF-NEXT:    br i1 [[CMP4]], label [[ENTRY_FOR_END_CRIT_EDGE:%.*]], label [[FOR_BODY_LR_PH:%.*]]
777; UNROLL-NO-VF:       entry.for.end_crit_edge:
778; UNROLL-NO-VF-NEXT:    [[DOTPRE:%.*]] = load i32, ptr [[B:%.*]], align 4
779; UNROLL-NO-VF-NEXT:    br label [[FOR_END:%.*]]
780; UNROLL-NO-VF:       for.body.lr.ph:
781; UNROLL-NO-VF-NEXT:    [[TMP0:%.*]] = load i16, ptr [[A:%.*]], align 2
782; UNROLL-NO-VF-NEXT:    [[SUB:%.*]] = add i16 [[TMP0]], -6
783; UNROLL-NO-VF-NEXT:    [[CONV2:%.*]] = sext i16 [[SUB]] to i32
784; UNROLL-NO-VF-NEXT:    [[C_PROMOTED:%.*]] = load i32, ptr [[C:%.*]], align 4
785; UNROLL-NO-VF-NEXT:    [[B_PROMOTED:%.*]] = load i32, ptr [[B]], align 4
786; UNROLL-NO-VF-NEXT:    br label [[FOR_BODY:%.*]]
787; UNROLL-NO-VF:       for.body:
788; UNROLL-NO-VF-NEXT:    [[INC7:%.*]] = phi i32 [ [[D]], [[FOR_BODY_LR_PH]] ], [ [[INC:%.*]], [[FOR_BODY]] ]
789; UNROLL-NO-VF-NEXT:    [[AND6:%.*]] = phi i32 [ [[B_PROMOTED]], [[FOR_BODY_LR_PH]] ], [ [[AND:%.*]], [[FOR_BODY]] ]
790; UNROLL-NO-VF-NEXT:    [[CONV25:%.*]] = phi i32 [ [[C_PROMOTED]], [[FOR_BODY_LR_PH]] ], [ [[CONV2]], [[FOR_BODY]] ]
791; UNROLL-NO-VF-NEXT:    [[AND]] = and i32 [[AND6]], [[CONV25]]
792; UNROLL-NO-VF-NEXT:    [[INC]] = add nsw i32 [[INC7]], 1
793; UNROLL-NO-VF-NEXT:    [[CMP:%.*]] = icmp eq i32 [[INC]], 21
794; UNROLL-NO-VF-NEXT:    br i1 [[CMP]], label [[FOR_COND_FOR_END_CRIT_EDGE:%.*]], label [[FOR_BODY]]
795; UNROLL-NO-VF:       for.cond.for.end_crit_edge:
796; UNROLL-NO-VF-NEXT:    [[AND_LCSSA:%.*]] = phi i32 [ [[AND]], [[FOR_BODY]] ]
797; UNROLL-NO-VF-NEXT:    store i32 [[CONV2]], ptr [[C]], align 4
798; UNROLL-NO-VF-NEXT:    store i32 [[AND_LCSSA]], ptr [[B]], align 4
799; UNROLL-NO-VF-NEXT:    store i16 [[SUB]], ptr [[E:%.*]], align 2
800; UNROLL-NO-VF-NEXT:    br label [[FOR_END]]
801; UNROLL-NO-VF:       for.end:
802; UNROLL-NO-VF-NEXT:    ret void
803;
804; SINK-AFTER-LABEL: @PR26734(
805; SINK-AFTER-NEXT:  entry:
806; SINK-AFTER-NEXT:    [[CMP4:%.*]] = icmp eq i32 [[D:%.*]], 21
807; SINK-AFTER-NEXT:    br i1 [[CMP4]], label [[ENTRY_FOR_END_CRIT_EDGE:%.*]], label [[FOR_BODY_LR_PH:%.*]]
808; SINK-AFTER:       entry.for.end_crit_edge:
809; SINK-AFTER-NEXT:    [[DOTPRE:%.*]] = load i32, ptr [[B:%.*]], align 4
810; SINK-AFTER-NEXT:    br label [[FOR_END:%.*]]
811; SINK-AFTER:       for.body.lr.ph:
812; SINK-AFTER-NEXT:    [[TMP0:%.*]] = load i16, ptr [[A:%.*]], align 2
813; SINK-AFTER-NEXT:    [[SUB:%.*]] = add i16 [[TMP0]], -6
814; SINK-AFTER-NEXT:    [[CONV2:%.*]] = sext i16 [[SUB]] to i32
815; SINK-AFTER-NEXT:    [[C_PROMOTED:%.*]] = load i32, ptr [[C:%.*]], align 4
816; SINK-AFTER-NEXT:    [[B_PROMOTED:%.*]] = load i32, ptr [[B]], align 4
817; SINK-AFTER-NEXT:    br label [[FOR_BODY:%.*]]
818; SINK-AFTER:       for.body:
819; SINK-AFTER-NEXT:    [[INC7:%.*]] = phi i32 [ [[D]], [[FOR_BODY_LR_PH]] ], [ [[INC:%.*]], [[FOR_BODY]] ]
820; SINK-AFTER-NEXT:    [[AND6:%.*]] = phi i32 [ [[B_PROMOTED]], [[FOR_BODY_LR_PH]] ], [ [[AND:%.*]], [[FOR_BODY]] ]
821; SINK-AFTER-NEXT:    [[CONV25:%.*]] = phi i32 [ [[C_PROMOTED]], [[FOR_BODY_LR_PH]] ], [ [[CONV2]], [[FOR_BODY]] ]
822; SINK-AFTER-NEXT:    [[AND]] = and i32 [[AND6]], [[CONV25]]
823; SINK-AFTER-NEXT:    [[INC]] = add nsw i32 [[INC7]], 1
824; SINK-AFTER-NEXT:    [[CMP:%.*]] = icmp eq i32 [[INC]], 21
825; SINK-AFTER-NEXT:    br i1 [[CMP]], label [[FOR_COND_FOR_END_CRIT_EDGE:%.*]], label [[FOR_BODY]]
826; SINK-AFTER:       for.cond.for.end_crit_edge:
827; SINK-AFTER-NEXT:    [[AND_LCSSA:%.*]] = phi i32 [ [[AND]], [[FOR_BODY]] ]
828; SINK-AFTER-NEXT:    store i32 [[CONV2]], ptr [[C]], align 4
829; SINK-AFTER-NEXT:    store i32 [[AND_LCSSA]], ptr [[B]], align 4
830; SINK-AFTER-NEXT:    store i16 [[SUB]], ptr [[E:%.*]], align 2
831; SINK-AFTER-NEXT:    br label [[FOR_END]]
832; SINK-AFTER:       for.end:
833; SINK-AFTER-NEXT:    ret void
834;
835entry:
836  %cmp4 = icmp eq i32 %d, 21
837  br i1 %cmp4, label %entry.for.end_crit_edge, label %for.body.lr.ph
838
839entry.for.end_crit_edge:
840  %.pre = load i32, ptr %b, align 4
841  br label %for.end
842
843for.body.lr.ph:
844  %0 = load i16, ptr %a, align 2
845  %sub = add i16 %0, -6
846  %conv2 = sext i16 %sub to i32
847  %c.promoted = load i32, ptr %c, align 4
848  %b.promoted = load i32, ptr %b, align 4
849  br label %for.body
850
851for.body:
852  %inc7 = phi i32 [ %d, %for.body.lr.ph ], [ %inc, %for.body ]
853  %and6 = phi i32 [ %b.promoted, %for.body.lr.ph ], [ %and, %for.body ]
854  %conv25 = phi i32 [ %c.promoted, %for.body.lr.ph ], [ %conv2, %for.body ]
855  %and = and i32 %and6, %conv25
856  %inc = add nsw i32 %inc7, 1
857  %cmp = icmp eq i32 %inc, 21
858  br i1 %cmp, label %for.cond.for.end_crit_edge, label %for.body
859
860for.cond.for.end_crit_edge:
861  %and.lcssa = phi i32 [ %and, %for.body ]
862  store i32 %conv2, ptr %c, align 4
863  store i32 %and.lcssa, ptr %b, align 4
864  store i16 %sub, ptr %e, align 2
865  br label %for.end
866
867for.end:
868  ret void
869}
870
871; int PR27246() {
872;   unsigned int e, n;
873;   for (int i = 1; i < 49; ++i) {
874;     for (int k = i; k > 1; --k)
875;       e = k;
876;     n = e;
877;   }
878;   return n;
879; }
880;
881;
882define i32 @PR27246() {
883; UNROLL-NO-IC-LABEL: @PR27246(
884; UNROLL-NO-IC-NEXT:  entry:
885; UNROLL-NO-IC-NEXT:    br label [[FOR_COND1_PREHEADER:%.*]]
886; UNROLL-NO-IC:       for.cond1.preheader:
887; UNROLL-NO-IC-NEXT:    [[I_016:%.*]] = phi i32 [ 1, [[ENTRY:%.*]] ], [ [[INC:%.*]], [[FOR_COND_CLEANUP3:%.*]] ]
888; UNROLL-NO-IC-NEXT:    [[E_015:%.*]] = phi i32 [ poison, [[ENTRY]] ], [ [[E_1_LCSSA:%.*]], [[FOR_COND_CLEANUP3]] ]
889; UNROLL-NO-IC-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i32 [[I_016]], 8
890; UNROLL-NO-IC-NEXT:    br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
891; UNROLL-NO-IC:       vector.ph:
892; UNROLL-NO-IC-NEXT:    [[N_MOD_VF:%.*]] = urem i32 [[I_016]], 8
893; UNROLL-NO-IC-NEXT:    [[N_VEC:%.*]] = sub i32 [[I_016]], [[N_MOD_VF]]
894; UNROLL-NO-IC-NEXT:    [[IND_END:%.*]] = sub i32 [[I_016]], [[N_VEC]]
895; UNROLL-NO-IC-NEXT:    [[VECTOR_RECUR_INIT:%.*]] = insertelement <4 x i32> poison, i32 [[E_015]], i32 3
896; UNROLL-NO-IC-NEXT:    [[DOTSPLATINSERT:%.*]] = insertelement <4 x i32> poison, i32 [[I_016]], i64 0
897; UNROLL-NO-IC-NEXT:    [[DOTSPLAT:%.*]] = shufflevector <4 x i32> [[DOTSPLATINSERT]], <4 x i32> poison, <4 x i32> zeroinitializer
898; UNROLL-NO-IC-NEXT:    [[INDUCTION:%.*]] = add <4 x i32> [[DOTSPLAT]], <i32 0, i32 -1, i32 -2, i32 -3>
899; UNROLL-NO-IC-NEXT:    br label [[VECTOR_BODY:%.*]]
900; UNROLL-NO-IC:       vector.body:
901; UNROLL-NO-IC-NEXT:    [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
902; UNROLL-NO-IC-NEXT:    [[VECTOR_RECUR:%.*]] = phi <4 x i32> [ [[VECTOR_RECUR_INIT]], [[VECTOR_PH]] ], [ [[STEP_ADD:%.*]], [[VECTOR_BODY]] ]
903; UNROLL-NO-IC-NEXT:    [[VEC_IND:%.*]] = phi <4 x i32> [ [[INDUCTION]], [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[VECTOR_BODY]] ]
904; UNROLL-NO-IC-NEXT:    [[STEP_ADD]] = add <4 x i32> [[VEC_IND]], splat (i32 -4)
905; UNROLL-NO-IC-NEXT:    [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 8
906; UNROLL-NO-IC-NEXT:    [[VEC_IND_NEXT]] = add <4 x i32> [[STEP_ADD]], splat (i32 -4)
907; UNROLL-NO-IC-NEXT:    [[TMP0:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]]
908; UNROLL-NO-IC-NEXT:    br i1 [[TMP0]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP8:![0-9]+]]
909; UNROLL-NO-IC:       middle.block:
910; UNROLL-NO-IC-NEXT:    [[VECTOR_RECUR_EXTRACT_FOR_PHI:%.*]] = extractelement <4 x i32> [[STEP_ADD]], i32 2
911; UNROLL-NO-IC-NEXT:    [[VECTOR_RECUR_EXTRACT:%.*]] = extractelement <4 x i32> [[STEP_ADD]], i32 3
912; UNROLL-NO-IC-NEXT:    [[CMP_N:%.*]] = icmp eq i32 [[I_016]], [[N_VEC]]
913; UNROLL-NO-IC-NEXT:    br i1 [[CMP_N]], label [[FOR_COND_CLEANUP3]], label [[SCALAR_PH]]
914; UNROLL-NO-IC:       scalar.ph:
915; UNROLL-NO-IC-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i32 [ [[VECTOR_RECUR_EXTRACT]], [[MIDDLE_BLOCK]] ], [ [[E_015]], [[FOR_COND1_PREHEADER]] ]
916; UNROLL-NO-IC-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i32 [ [[IND_END]], [[MIDDLE_BLOCK]] ], [ [[I_016]], [[FOR_COND1_PREHEADER]] ]
917; UNROLL-NO-IC-NEXT:    br label [[FOR_COND1:%.*]]
918; UNROLL-NO-IC:       for.cond.cleanup:
919; UNROLL-NO-IC-NEXT:    [[E_1_LCSSA_LCSSA:%.*]] = phi i32 [ [[E_1_LCSSA]], [[FOR_COND_CLEANUP3]] ]
920; UNROLL-NO-IC-NEXT:    ret i32 [[E_1_LCSSA_LCSSA]]
921; UNROLL-NO-IC:       for.cond1:
922; UNROLL-NO-IC-NEXT:    [[E_1:%.*]] = phi i32 [ [[K_0:%.*]], [[FOR_COND1]] ], [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ]
923; UNROLL-NO-IC-NEXT:    [[K_0]] = phi i32 [ [[DEC:%.*]], [[FOR_COND1]] ], [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ]
924; UNROLL-NO-IC-NEXT:    [[CMP2:%.*]] = icmp sgt i32 [[K_0]], 1
925; UNROLL-NO-IC-NEXT:    [[DEC]] = add nsw i32 [[K_0]], -1
926; UNROLL-NO-IC-NEXT:    br i1 [[CMP2]], label [[FOR_COND1]], label [[FOR_COND_CLEANUP3]], !llvm.loop [[LOOP9:![0-9]+]]
927; UNROLL-NO-IC:       for.cond.cleanup3:
928; UNROLL-NO-IC-NEXT:    [[E_1_LCSSA]] = phi i32 [ [[E_1]], [[FOR_COND1]] ], [ [[VECTOR_RECUR_EXTRACT_FOR_PHI]], [[MIDDLE_BLOCK]] ]
929; UNROLL-NO-IC-NEXT:    [[INC]] = add nuw nsw i32 [[I_016]], 1
930; UNROLL-NO-IC-NEXT:    [[EXITCOND:%.*]] = icmp eq i32 [[INC]], 49
931; UNROLL-NO-IC-NEXT:    br i1 [[EXITCOND]], label [[FOR_COND_CLEANUP:%.*]], label [[FOR_COND1_PREHEADER]]
932;
933; UNROLL-NO-VF-LABEL: @PR27246(
934; UNROLL-NO-VF-NEXT:  entry:
935; UNROLL-NO-VF-NEXT:    br label [[FOR_COND1_PREHEADER:%.*]]
936; UNROLL-NO-VF:       for.cond1.preheader:
937; UNROLL-NO-VF-NEXT:    [[I_016:%.*]] = phi i32 [ 1, [[ENTRY:%.*]] ], [ [[INC:%.*]], [[FOR_COND_CLEANUP3:%.*]] ]
938; UNROLL-NO-VF-NEXT:    [[E_015:%.*]] = phi i32 [ poison, [[ENTRY]] ], [ [[E_1_LCSSA:%.*]], [[FOR_COND_CLEANUP3]] ]
939; UNROLL-NO-VF-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i32 [[I_016]], 2
940; UNROLL-NO-VF-NEXT:    br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
941; UNROLL-NO-VF:       vector.ph:
942; UNROLL-NO-VF-NEXT:    [[N_MOD_VF:%.*]] = urem i32 [[I_016]], 2
943; UNROLL-NO-VF-NEXT:    [[N_VEC:%.*]] = sub i32 [[I_016]], [[N_MOD_VF]]
944; UNROLL-NO-VF-NEXT:    [[IND_END:%.*]] = sub i32 [[I_016]], [[N_VEC]]
945; UNROLL-NO-VF-NEXT:    br label [[VECTOR_BODY:%.*]]
946; UNROLL-NO-VF:       vector.body:
947; UNROLL-NO-VF-NEXT:    [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
948; UNROLL-NO-VF-NEXT:    [[VECTOR_RECUR:%.*]] = phi i32 [ [[E_015]], [[VECTOR_PH]] ], [ [[TMP1:%.*]], [[VECTOR_BODY]] ]
949; UNROLL-NO-VF-NEXT:    [[OFFSET_IDX:%.*]] = sub i32 [[I_016]], [[INDEX]]
950; UNROLL-NO-VF-NEXT:    [[TMP0:%.*]] = add i32 [[OFFSET_IDX]], 0
951; UNROLL-NO-VF-NEXT:    [[TMP1]] = add i32 [[OFFSET_IDX]], -1
952; UNROLL-NO-VF-NEXT:    [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 2
953; UNROLL-NO-VF-NEXT:    [[TMP2:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]]
954; UNROLL-NO-VF-NEXT:    br i1 [[TMP2]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP8:![0-9]+]]
955; UNROLL-NO-VF:       middle.block:
956; UNROLL-NO-VF-NEXT:    [[CMP_N:%.*]] = icmp eq i32 [[I_016]], [[N_VEC]]
957; UNROLL-NO-VF-NEXT:    br i1 [[CMP_N]], label [[FOR_COND_CLEANUP3]], label [[SCALAR_PH]]
958; UNROLL-NO-VF:       scalar.ph:
959; UNROLL-NO-VF-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i32 [ [[TMP1]], [[MIDDLE_BLOCK]] ], [ [[E_015]], [[FOR_COND1_PREHEADER]] ]
960; UNROLL-NO-VF-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i32 [ [[IND_END]], [[MIDDLE_BLOCK]] ], [ [[I_016]], [[FOR_COND1_PREHEADER]] ]
961; UNROLL-NO-VF-NEXT:    br label [[FOR_COND1:%.*]]
962; UNROLL-NO-VF:       for.cond.cleanup:
963; UNROLL-NO-VF-NEXT:    [[E_1_LCSSA_LCSSA:%.*]] = phi i32 [ [[E_1_LCSSA]], [[FOR_COND_CLEANUP3]] ]
964; UNROLL-NO-VF-NEXT:    ret i32 [[E_1_LCSSA_LCSSA]]
965; UNROLL-NO-VF:       for.cond1:
966; UNROLL-NO-VF-NEXT:    [[E_1:%.*]] = phi i32 [ [[K_0:%.*]], [[FOR_COND1]] ], [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ]
967; UNROLL-NO-VF-NEXT:    [[K_0]] = phi i32 [ [[DEC:%.*]], [[FOR_COND1]] ], [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ]
968; UNROLL-NO-VF-NEXT:    [[CMP2:%.*]] = icmp sgt i32 [[K_0]], 1
969; UNROLL-NO-VF-NEXT:    [[DEC]] = add nsw i32 [[K_0]], -1
970; UNROLL-NO-VF-NEXT:    br i1 [[CMP2]], label [[FOR_COND1]], label [[FOR_COND_CLEANUP3]], !llvm.loop [[LOOP9:![0-9]+]]
971; UNROLL-NO-VF:       for.cond.cleanup3:
972; UNROLL-NO-VF-NEXT:    [[E_1_LCSSA]] = phi i32 [ [[E_1]], [[FOR_COND1]] ], [ [[TMP0]], [[MIDDLE_BLOCK]] ]
973; UNROLL-NO-VF-NEXT:    [[INC]] = add nuw nsw i32 [[I_016]], 1
974; UNROLL-NO-VF-NEXT:    [[EXITCOND:%.*]] = icmp eq i32 [[INC]], 49
975; UNROLL-NO-VF-NEXT:    br i1 [[EXITCOND]], label [[FOR_COND_CLEANUP:%.*]], label [[FOR_COND1_PREHEADER]]
976;
977; SINK-AFTER-LABEL: @PR27246(
978; SINK-AFTER-NEXT:  entry:
979; SINK-AFTER-NEXT:    br label [[FOR_COND1_PREHEADER:%.*]]
980; SINK-AFTER:       for.cond1.preheader:
981; SINK-AFTER-NEXT:    [[I_016:%.*]] = phi i32 [ 1, [[ENTRY:%.*]] ], [ [[INC:%.*]], [[FOR_COND_CLEANUP3:%.*]] ]
982; SINK-AFTER-NEXT:    [[E_015:%.*]] = phi i32 [ poison, [[ENTRY]] ], [ [[E_1_LCSSA:%.*]], [[FOR_COND_CLEANUP3]] ]
983; SINK-AFTER-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i32 [[I_016]], 4
984; SINK-AFTER-NEXT:    br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
985; SINK-AFTER:       vector.ph:
986; SINK-AFTER-NEXT:    [[N_MOD_VF:%.*]] = urem i32 [[I_016]], 4
987; SINK-AFTER-NEXT:    [[N_VEC:%.*]] = sub i32 [[I_016]], [[N_MOD_VF]]
988; SINK-AFTER-NEXT:    [[IND_END:%.*]] = sub i32 [[I_016]], [[N_VEC]]
989; SINK-AFTER-NEXT:    [[VECTOR_RECUR_INIT:%.*]] = insertelement <4 x i32> poison, i32 [[E_015]], i32 3
990; SINK-AFTER-NEXT:    [[DOTSPLATINSERT:%.*]] = insertelement <4 x i32> poison, i32 [[I_016]], i64 0
991; SINK-AFTER-NEXT:    [[DOTSPLAT:%.*]] = shufflevector <4 x i32> [[DOTSPLATINSERT]], <4 x i32> poison, <4 x i32> zeroinitializer
992; SINK-AFTER-NEXT:    [[INDUCTION:%.*]] = add <4 x i32> [[DOTSPLAT]], <i32 0, i32 -1, i32 -2, i32 -3>
993; SINK-AFTER-NEXT:    br label [[VECTOR_BODY:%.*]]
994; SINK-AFTER:       vector.body:
995; SINK-AFTER-NEXT:    [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
996; SINK-AFTER-NEXT:    [[VECTOR_RECUR:%.*]] = phi <4 x i32> [ [[VECTOR_RECUR_INIT]], [[VECTOR_PH]] ], [ [[VEC_IND:%.*]], [[VECTOR_BODY]] ]
997; SINK-AFTER-NEXT:    [[VEC_IND]] = phi <4 x i32> [ [[INDUCTION]], [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[VECTOR_BODY]] ]
998; SINK-AFTER-NEXT:    [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 4
999; SINK-AFTER-NEXT:    [[VEC_IND_NEXT]] = add <4 x i32> [[VEC_IND]], splat (i32 -4)
1000; SINK-AFTER-NEXT:    [[TMP0:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]]
1001; SINK-AFTER-NEXT:    br i1 [[TMP0]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP8:![0-9]+]]
1002; SINK-AFTER:       middle.block:
1003; SINK-AFTER-NEXT:    [[VECTOR_RECUR_EXTRACT_FOR_PHI:%.*]] = extractelement <4 x i32> [[VEC_IND]], i32 2
1004; SINK-AFTER-NEXT:    [[VECTOR_RECUR_EXTRACT:%.*]] = extractelement <4 x i32> [[VEC_IND]], i32 3
1005; SINK-AFTER-NEXT:    [[CMP_N:%.*]] = icmp eq i32 [[I_016]], [[N_VEC]]
1006; SINK-AFTER-NEXT:    br i1 [[CMP_N]], label [[FOR_COND_CLEANUP3]], label [[SCALAR_PH]]
1007; SINK-AFTER:       scalar.ph:
1008; SINK-AFTER-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i32 [ [[VECTOR_RECUR_EXTRACT]], [[MIDDLE_BLOCK]] ], [ [[E_015]], [[FOR_COND1_PREHEADER]] ]
1009; SINK-AFTER-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i32 [ [[IND_END]], [[MIDDLE_BLOCK]] ], [ [[I_016]], [[FOR_COND1_PREHEADER]] ]
1010; SINK-AFTER-NEXT:    br label [[FOR_COND1:%.*]]
1011; SINK-AFTER:       for.cond.cleanup:
1012; SINK-AFTER-NEXT:    [[E_1_LCSSA_LCSSA:%.*]] = phi i32 [ [[E_1_LCSSA]], [[FOR_COND_CLEANUP3]] ]
1013; SINK-AFTER-NEXT:    ret i32 [[E_1_LCSSA_LCSSA]]
1014; SINK-AFTER:       for.cond1:
1015; SINK-AFTER-NEXT:    [[E_1:%.*]] = phi i32 [ [[K_0:%.*]], [[FOR_COND1]] ], [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ]
1016; SINK-AFTER-NEXT:    [[K_0]] = phi i32 [ [[DEC:%.*]], [[FOR_COND1]] ], [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ]
1017; SINK-AFTER-NEXT:    [[CMP2:%.*]] = icmp sgt i32 [[K_0]], 1
1018; SINK-AFTER-NEXT:    [[DEC]] = add nsw i32 [[K_0]], -1
1019; SINK-AFTER-NEXT:    br i1 [[CMP2]], label [[FOR_COND1]], label [[FOR_COND_CLEANUP3]], !llvm.loop [[LOOP9:![0-9]+]]
1020; SINK-AFTER:       for.cond.cleanup3:
1021; SINK-AFTER-NEXT:    [[E_1_LCSSA]] = phi i32 [ [[E_1]], [[FOR_COND1]] ], [ [[VECTOR_RECUR_EXTRACT_FOR_PHI]], [[MIDDLE_BLOCK]] ]
1022; SINK-AFTER-NEXT:    [[INC]] = add nuw nsw i32 [[I_016]], 1
1023; SINK-AFTER-NEXT:    [[EXITCOND:%.*]] = icmp eq i32 [[INC]], 49
1024; SINK-AFTER-NEXT:    br i1 [[EXITCOND]], label [[FOR_COND_CLEANUP:%.*]], label [[FOR_COND1_PREHEADER]]
1025;
1026entry:
1027  br label %for.cond1.preheader
1028
1029for.cond1.preheader:
1030  %i.016 = phi i32 [ 1, %entry ], [ %inc, %for.cond.cleanup3 ]
1031  %e.015 = phi i32 [ poison, %entry ], [ %e.1.lcssa, %for.cond.cleanup3 ]
1032  br label %for.cond1
1033
1034for.cond.cleanup:
1035  %e.1.lcssa.lcssa = phi i32 [ %e.1.lcssa, %for.cond.cleanup3 ]
1036  ret i32 %e.1.lcssa.lcssa
1037
1038for.cond1:
1039  %e.1 = phi i32 [ %k.0, %for.cond1 ], [ %e.015, %for.cond1.preheader ]
1040  %k.0 = phi i32 [ %dec, %for.cond1 ], [ %i.016, %for.cond1.preheader ]
1041  %cmp2 = icmp sgt i32 %k.0, 1
1042  %dec = add nsw i32 %k.0, -1
1043  br i1 %cmp2, label %for.cond1, label %for.cond.cleanup3
1044
1045for.cond.cleanup3:
1046  %e.1.lcssa = phi i32 [ %e.1, %for.cond1 ]
1047  %inc = add nuw nsw i32 %i.016, 1
1048  %exitcond = icmp eq i32 %inc, 49
1049  br i1 %exitcond, label %for.cond.cleanup, label %for.cond1.preheader
1050}
1051
1052;
1053define i32 @PR30183(i32 %pre_load, ptr %a, ptr %b, i64 %n) {
1054; UNROLL-NO-IC-LABEL: @PR30183(
1055; UNROLL-NO-IC-NEXT:  entry:
1056; UNROLL-NO-IC-NEXT:    [[TMP0:%.*]] = add i64 [[N:%.*]], -2
1057; UNROLL-NO-IC-NEXT:    [[TMP1:%.*]] = lshr i64 [[TMP0]], 1
1058; UNROLL-NO-IC-NEXT:    [[TMP2:%.*]] = add nuw i64 [[TMP1]], 1
1059; UNROLL-NO-IC-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[TMP2]], 8
1060; UNROLL-NO-IC-NEXT:    br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
1061; UNROLL-NO-IC:       vector.ph:
1062; UNROLL-NO-IC-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[TMP2]], 8
1063; UNROLL-NO-IC-NEXT:    [[N_VEC:%.*]] = sub i64 [[TMP2]], [[N_MOD_VF]]
1064; UNROLL-NO-IC-NEXT:    [[IND_END:%.*]] = mul i64 [[N_VEC]], 2
1065; UNROLL-NO-IC-NEXT:    [[VECTOR_RECUR_INIT:%.*]] = insertelement <4 x i32> poison, i32 [[PRE_LOAD:%.*]], i32 3
1066; UNROLL-NO-IC-NEXT:    br label [[VECTOR_BODY:%.*]]
1067; UNROLL-NO-IC:       vector.body:
1068; UNROLL-NO-IC-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
1069; UNROLL-NO-IC-NEXT:    [[VECTOR_RECUR:%.*]] = phi <4 x i32> [ [[VECTOR_RECUR_INIT]], [[VECTOR_PH]] ], [ [[TMP22:%.*]], [[VECTOR_BODY]] ]
1070; UNROLL-NO-IC-NEXT:    [[OFFSET_IDX:%.*]] = mul i64 [[INDEX]], 2
1071; UNROLL-NO-IC-NEXT:    [[TMP3:%.*]] = add i64 [[OFFSET_IDX]], 8
1072; UNROLL-NO-IC-NEXT:    [[TMP4:%.*]] = add i64 [[OFFSET_IDX]], 10
1073; UNROLL-NO-IC-NEXT:    [[TMP5:%.*]] = add i64 [[OFFSET_IDX]], 12
1074; UNROLL-NO-IC-NEXT:    [[TMP6:%.*]] = add i64 [[OFFSET_IDX]], 14
1075; UNROLL-NO-IC-NEXT:    [[TMP7:%.*]] = add nuw nsw i64 [[TMP3]], 2
1076; UNROLL-NO-IC-NEXT:    [[TMP8:%.*]] = add nuw nsw i64 [[TMP4]], 2
1077; UNROLL-NO-IC-NEXT:    [[TMP9:%.*]] = add nuw nsw i64 [[TMP5]], 2
1078; UNROLL-NO-IC-NEXT:    [[TMP10:%.*]] = add nuw nsw i64 [[TMP6]], 2
1079; UNROLL-NO-IC-NEXT:    [[TMP11:%.*]] = getelementptr inbounds i32, ptr [[A:%.*]], i64 [[TMP7]]
1080; UNROLL-NO-IC-NEXT:    [[TMP12:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[TMP8]]
1081; UNROLL-NO-IC-NEXT:    [[TMP13:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[TMP9]]
1082; UNROLL-NO-IC-NEXT:    [[TMP14:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[TMP10]]
1083; UNROLL-NO-IC-NEXT:    [[TMP15:%.*]] = load i32, ptr [[TMP11]], align 4
1084; UNROLL-NO-IC-NEXT:    [[TMP16:%.*]] = load i32, ptr [[TMP12]], align 4
1085; UNROLL-NO-IC-NEXT:    [[TMP17:%.*]] = load i32, ptr [[TMP13]], align 4
1086; UNROLL-NO-IC-NEXT:    [[TMP18:%.*]] = load i32, ptr [[TMP14]], align 4
1087; UNROLL-NO-IC-NEXT:    [[TMP19:%.*]] = insertelement <4 x i32> poison, i32 [[TMP15]], i32 0
1088; UNROLL-NO-IC-NEXT:    [[TMP20:%.*]] = insertelement <4 x i32> [[TMP19]], i32 [[TMP16]], i32 1
1089; UNROLL-NO-IC-NEXT:    [[TMP21:%.*]] = insertelement <4 x i32> [[TMP20]], i32 [[TMP17]], i32 2
1090; UNROLL-NO-IC-NEXT:    [[TMP22]] = insertelement <4 x i32> [[TMP21]], i32 [[TMP18]], i32 3
1091; UNROLL-NO-IC-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 8
1092; UNROLL-NO-IC-NEXT:    [[TMP23:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
1093; UNROLL-NO-IC-NEXT:    br i1 [[TMP23]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP10:![0-9]+]]
1094; UNROLL-NO-IC:       middle.block:
1095; UNROLL-NO-IC-NEXT:    [[CMP_N:%.*]] = icmp eq i64 [[TMP2]], [[N_VEC]]
1096; UNROLL-NO-IC-NEXT:    br i1 [[CMP_N]], label [[FOR_END:%.*]], label [[SCALAR_PH]]
1097; UNROLL-NO-IC:       scalar.ph:
1098; UNROLL-NO-IC-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ [[IND_END]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ]
1099; UNROLL-NO-IC-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i32 [ [[TMP18]], [[MIDDLE_BLOCK]] ], [ [[PRE_LOAD]], [[ENTRY]] ]
1100; UNROLL-NO-IC-NEXT:    br label [[SCALAR_BODY:%.*]]
1101; UNROLL-NO-IC:       scalar.body:
1102; UNROLL-NO-IC-NEXT:    [[I:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[I_NEXT:%.*]], [[SCALAR_BODY]] ]
1103; UNROLL-NO-IC-NEXT:    [[VAR0:%.*]] = phi i32 [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ], [ [[VAR2:%.*]], [[SCALAR_BODY]] ]
1104; UNROLL-NO-IC-NEXT:    [[I_NEXT]] = add nuw nsw i64 [[I]], 2
1105; UNROLL-NO-IC-NEXT:    [[VAR1:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[I_NEXT]]
1106; UNROLL-NO-IC-NEXT:    [[VAR2]] = load i32, ptr [[VAR1]], align 4
1107; UNROLL-NO-IC-NEXT:    [[COND:%.*]] = icmp eq i64 [[I_NEXT]], [[N]]
1108; UNROLL-NO-IC-NEXT:    br i1 [[COND]], label [[FOR_END]], label [[SCALAR_BODY]], !llvm.loop [[LOOP11:![0-9]+]]
1109; UNROLL-NO-IC:       for.end:
1110; UNROLL-NO-IC-NEXT:    [[VAR0_LCSSA:%.*]] = phi i32 [ [[VAR0]], [[SCALAR_BODY]] ], [ [[TMP17]], [[MIDDLE_BLOCK]] ]
1111; UNROLL-NO-IC-NEXT:    ret i32 [[VAR0_LCSSA]]
1112;
1113; UNROLL-NO-VF-LABEL: @PR30183(
1114; UNROLL-NO-VF-NEXT:  entry:
1115; UNROLL-NO-VF-NEXT:    [[TMP0:%.*]] = add i64 [[N:%.*]], -2
1116; UNROLL-NO-VF-NEXT:    [[TMP1:%.*]] = lshr i64 [[TMP0]], 1
1117; UNROLL-NO-VF-NEXT:    [[TMP2:%.*]] = add nuw i64 [[TMP1]], 1
1118; UNROLL-NO-VF-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[TMP2]], 2
1119; UNROLL-NO-VF-NEXT:    br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
1120; UNROLL-NO-VF:       vector.ph:
1121; UNROLL-NO-VF-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[TMP2]], 2
1122; UNROLL-NO-VF-NEXT:    [[N_VEC:%.*]] = sub i64 [[TMP2]], [[N_MOD_VF]]
1123; UNROLL-NO-VF-NEXT:    [[IND_END:%.*]] = mul i64 [[N_VEC]], 2
1124; UNROLL-NO-VF-NEXT:    br label [[VECTOR_BODY:%.*]]
1125; UNROLL-NO-VF:       vector.body:
1126; UNROLL-NO-VF-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
1127; UNROLL-NO-VF-NEXT:    [[VECTOR_RECUR:%.*]] = phi i32 [ [[PRE_LOAD:%.*]], [[VECTOR_PH]] ], [ [[TMP10:%.*]], [[VECTOR_BODY]] ]
1128; UNROLL-NO-VF-NEXT:    [[OFFSET_IDX:%.*]] = mul i64 [[INDEX]], 2
1129; UNROLL-NO-VF-NEXT:    [[TMP3:%.*]] = add i64 [[OFFSET_IDX]], 0
1130; UNROLL-NO-VF-NEXT:    [[TMP4:%.*]] = add i64 [[OFFSET_IDX]], 2
1131; UNROLL-NO-VF-NEXT:    [[TMP5:%.*]] = add nuw nsw i64 [[TMP3]], 2
1132; UNROLL-NO-VF-NEXT:    [[TMP6:%.*]] = add nuw nsw i64 [[TMP4]], 2
1133; UNROLL-NO-VF-NEXT:    [[TMP7:%.*]] = getelementptr inbounds i32, ptr [[A:%.*]], i64 [[TMP5]]
1134; UNROLL-NO-VF-NEXT:    [[TMP8:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[TMP6]]
1135; UNROLL-NO-VF-NEXT:    [[TMP9:%.*]] = load i32, ptr [[TMP7]], align 4
1136; UNROLL-NO-VF-NEXT:    [[TMP10]] = load i32, ptr [[TMP8]], align 4
1137; UNROLL-NO-VF-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 2
1138; UNROLL-NO-VF-NEXT:    [[TMP11:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
1139; UNROLL-NO-VF-NEXT:    br i1 [[TMP11]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP10:![0-9]+]]
1140; UNROLL-NO-VF:       middle.block:
1141; UNROLL-NO-VF-NEXT:    [[CMP_N:%.*]] = icmp eq i64 [[TMP2]], [[N_VEC]]
1142; UNROLL-NO-VF-NEXT:    br i1 [[CMP_N]], label [[FOR_END:%.*]], label [[SCALAR_PH]]
1143; UNROLL-NO-VF:       scalar.ph:
1144; UNROLL-NO-VF-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ [[IND_END]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ]
1145; UNROLL-NO-VF-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i32 [ [[TMP10]], [[MIDDLE_BLOCK]] ], [ [[PRE_LOAD]], [[ENTRY]] ]
1146; UNROLL-NO-VF-NEXT:    br label [[SCALAR_BODY:%.*]]
1147; UNROLL-NO-VF:       scalar.body:
1148; UNROLL-NO-VF-NEXT:    [[I:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[I_NEXT:%.*]], [[SCALAR_BODY]] ]
1149; UNROLL-NO-VF-NEXT:    [[VAR0:%.*]] = phi i32 [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ], [ [[VAR2:%.*]], [[SCALAR_BODY]] ]
1150; UNROLL-NO-VF-NEXT:    [[I_NEXT]] = add nuw nsw i64 [[I]], 2
1151; UNROLL-NO-VF-NEXT:    [[VAR1:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[I_NEXT]]
1152; UNROLL-NO-VF-NEXT:    [[VAR2]] = load i32, ptr [[VAR1]], align 4
1153; UNROLL-NO-VF-NEXT:    [[COND:%.*]] = icmp eq i64 [[I_NEXT]], [[N]]
1154; UNROLL-NO-VF-NEXT:    br i1 [[COND]], label [[FOR_END]], label [[SCALAR_BODY]], !llvm.loop [[LOOP11:![0-9]+]]
1155; UNROLL-NO-VF:       for.end:
1156; UNROLL-NO-VF-NEXT:    [[VAR0_LCSSA:%.*]] = phi i32 [ [[VAR0]], [[SCALAR_BODY]] ], [ [[TMP9]], [[MIDDLE_BLOCK]] ]
1157; UNROLL-NO-VF-NEXT:    ret i32 [[VAR0_LCSSA]]
1158;
1159; SINK-AFTER-LABEL: @PR30183(
1160; SINK-AFTER-NEXT:  entry:
1161; SINK-AFTER-NEXT:    [[TMP0:%.*]] = add i64 [[N:%.*]], -2
1162; SINK-AFTER-NEXT:    [[TMP1:%.*]] = lshr i64 [[TMP0]], 1
1163; SINK-AFTER-NEXT:    [[TMP2:%.*]] = add nuw i64 [[TMP1]], 1
1164; SINK-AFTER-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[TMP2]], 4
1165; SINK-AFTER-NEXT:    br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
1166; SINK-AFTER:       vector.ph:
1167; SINK-AFTER-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[TMP2]], 4
1168; SINK-AFTER-NEXT:    [[N_VEC:%.*]] = sub i64 [[TMP2]], [[N_MOD_VF]]
1169; SINK-AFTER-NEXT:    [[IND_END:%.*]] = mul i64 [[N_VEC]], 2
1170; SINK-AFTER-NEXT:    [[VECTOR_RECUR_INIT:%.*]] = insertelement <4 x i32> poison, i32 [[PRE_LOAD:%.*]], i32 3
1171; SINK-AFTER-NEXT:    br label [[VECTOR_BODY:%.*]]
1172; SINK-AFTER:       vector.body:
1173; SINK-AFTER-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
1174; SINK-AFTER-NEXT:    [[VECTOR_RECUR:%.*]] = phi <4 x i32> [ [[VECTOR_RECUR_INIT]], [[VECTOR_PH]] ], [ [[TMP22:%.*]], [[VECTOR_BODY]] ]
1175; SINK-AFTER-NEXT:    [[OFFSET_IDX:%.*]] = mul i64 [[INDEX]], 2
1176; SINK-AFTER-NEXT:    [[TMP3:%.*]] = add i64 [[OFFSET_IDX]], 0
1177; SINK-AFTER-NEXT:    [[TMP4:%.*]] = add i64 [[OFFSET_IDX]], 2
1178; SINK-AFTER-NEXT:    [[TMP5:%.*]] = add i64 [[OFFSET_IDX]], 4
1179; SINK-AFTER-NEXT:    [[TMP6:%.*]] = add i64 [[OFFSET_IDX]], 6
1180; SINK-AFTER-NEXT:    [[TMP7:%.*]] = add nuw nsw i64 [[TMP3]], 2
1181; SINK-AFTER-NEXT:    [[TMP8:%.*]] = add nuw nsw i64 [[TMP4]], 2
1182; SINK-AFTER-NEXT:    [[TMP9:%.*]] = add nuw nsw i64 [[TMP5]], 2
1183; SINK-AFTER-NEXT:    [[TMP10:%.*]] = add nuw nsw i64 [[TMP6]], 2
1184; SINK-AFTER-NEXT:    [[TMP11:%.*]] = getelementptr inbounds i32, ptr [[A:%.*]], i64 [[TMP7]]
1185; SINK-AFTER-NEXT:    [[TMP12:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[TMP8]]
1186; SINK-AFTER-NEXT:    [[TMP13:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[TMP9]]
1187; SINK-AFTER-NEXT:    [[TMP14:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[TMP10]]
1188; SINK-AFTER-NEXT:    [[TMP15:%.*]] = load i32, ptr [[TMP11]], align 4
1189; SINK-AFTER-NEXT:    [[TMP16:%.*]] = load i32, ptr [[TMP12]], align 4
1190; SINK-AFTER-NEXT:    [[TMP17:%.*]] = load i32, ptr [[TMP13]], align 4
1191; SINK-AFTER-NEXT:    [[TMP18:%.*]] = load i32, ptr [[TMP14]], align 4
1192; SINK-AFTER-NEXT:    [[TMP19:%.*]] = insertelement <4 x i32> poison, i32 [[TMP15]], i32 0
1193; SINK-AFTER-NEXT:    [[TMP20:%.*]] = insertelement <4 x i32> [[TMP19]], i32 [[TMP16]], i32 1
1194; SINK-AFTER-NEXT:    [[TMP21:%.*]] = insertelement <4 x i32> [[TMP20]], i32 [[TMP17]], i32 2
1195; SINK-AFTER-NEXT:    [[TMP22]] = insertelement <4 x i32> [[TMP21]], i32 [[TMP18]], i32 3
1196; SINK-AFTER-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4
1197; SINK-AFTER-NEXT:    [[TMP23:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
1198; SINK-AFTER-NEXT:    br i1 [[TMP23]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP10:![0-9]+]]
1199; SINK-AFTER:       middle.block:
1200; SINK-AFTER-NEXT:    [[CMP_N:%.*]] = icmp eq i64 [[TMP2]], [[N_VEC]]
1201; SINK-AFTER-NEXT:    br i1 [[CMP_N]], label [[FOR_END:%.*]], label [[SCALAR_PH]]
1202; SINK-AFTER:       scalar.ph:
1203; SINK-AFTER-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ [[IND_END]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ]
1204; SINK-AFTER-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i32 [ [[TMP18]], [[MIDDLE_BLOCK]] ], [ [[PRE_LOAD]], [[ENTRY]] ]
1205; SINK-AFTER-NEXT:    br label [[SCALAR_BODY:%.*]]
1206; SINK-AFTER:       scalar.body:
1207; SINK-AFTER-NEXT:    [[I:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[I_NEXT:%.*]], [[SCALAR_BODY]] ]
1208; SINK-AFTER-NEXT:    [[VAR0:%.*]] = phi i32 [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ], [ [[VAR2:%.*]], [[SCALAR_BODY]] ]
1209; SINK-AFTER-NEXT:    [[I_NEXT]] = add nuw nsw i64 [[I]], 2
1210; SINK-AFTER-NEXT:    [[VAR1:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[I_NEXT]]
1211; SINK-AFTER-NEXT:    [[VAR2]] = load i32, ptr [[VAR1]], align 4
1212; SINK-AFTER-NEXT:    [[COND:%.*]] = icmp eq i64 [[I_NEXT]], [[N]]
1213; SINK-AFTER-NEXT:    br i1 [[COND]], label [[FOR_END]], label [[SCALAR_BODY]], !llvm.loop [[LOOP11:![0-9]+]]
1214; SINK-AFTER:       for.end:
1215; SINK-AFTER-NEXT:    [[VAR0_LCSSA:%.*]] = phi i32 [ [[VAR0]], [[SCALAR_BODY]] ], [ [[TMP17]], [[MIDDLE_BLOCK]] ]
1216; SINK-AFTER-NEXT:    ret i32 [[VAR0_LCSSA]]
1217;
1218entry:
1219  br label %scalar.body
1220
1221scalar.body:
1222  %i = phi i64 [ 0, %entry ], [ %i.next, %scalar.body ]
1223  %var0 = phi i32 [ %pre_load, %entry ], [ %var2, %scalar.body ]
1224  %i.next = add nuw nsw i64 %i, 2
1225  %var1 = getelementptr inbounds i32, ptr %a, i64 %i.next
1226  %var2 = load i32, ptr %var1
1227  %cond = icmp eq i64 %i.next,%n
1228  br i1 %cond, label %for.end, label %scalar.body
1229
1230for.end:
1231  ret i32 %var0
1232}
1233
1234;
1235define i64 @constant_folded_previous_value() {
1236; UNROLL-NO-IC-LABEL: @constant_folded_previous_value(
1237; UNROLL-NO-IC-NEXT:  entry:
1238; UNROLL-NO-IC-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
1239; UNROLL-NO-IC:       vector.ph:
1240; UNROLL-NO-IC-NEXT:    br label [[VECTOR_BODY:%.*]]
1241; UNROLL-NO-IC:       vector.body:
1242; UNROLL-NO-IC-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
1243; UNROLL-NO-IC-NEXT:    [[VECTOR_RECUR:%.*]] = phi <4 x i64> [ <i64 poison, i64 poison, i64 poison, i64 0>, [[VECTOR_PH]] ], [ splat (i64 1), [[VECTOR_BODY]] ]
1244; UNROLL-NO-IC-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 8
1245; UNROLL-NO-IC-NEXT:    [[TMP0:%.*]] = icmp eq i64 [[INDEX_NEXT]], 1000
1246; UNROLL-NO-IC-NEXT:    br i1 [[TMP0]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP12:![0-9]+]]
1247; UNROLL-NO-IC:       middle.block:
1248; UNROLL-NO-IC-NEXT:    br i1 true, label [[FOR_END:%.*]], label [[SCALAR_PH]]
1249; UNROLL-NO-IC:       scalar.ph:
1250; UNROLL-NO-IC-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ 1000, [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ]
1251; UNROLL-NO-IC-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i64 [ 1, [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY]] ]
1252; UNROLL-NO-IC-NEXT:    br label [[SCALAR_BODY:%.*]]
1253; UNROLL-NO-IC:       scalar.body:
1254; UNROLL-NO-IC-NEXT:    [[I:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[I_NEXT:%.*]], [[SCALAR_BODY]] ]
1255; UNROLL-NO-IC-NEXT:    [[VAR2:%.*]] = phi i64 [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ], [ [[VAR3:%.*]], [[SCALAR_BODY]] ]
1256; UNROLL-NO-IC-NEXT:    [[VAR3]] = add i64 0, 1
1257; UNROLL-NO-IC-NEXT:    [[I_NEXT]] = add nuw nsw i64 [[I]], 1
1258; UNROLL-NO-IC-NEXT:    [[COND:%.*]] = icmp eq i64 [[I_NEXT]], 1000
1259; UNROLL-NO-IC-NEXT:    br i1 [[COND]], label [[FOR_END]], label [[SCALAR_BODY]], !llvm.loop [[LOOP13:![0-9]+]]
1260; UNROLL-NO-IC:       for.end:
1261; UNROLL-NO-IC-NEXT:    [[VAR2_LCSSA:%.*]] = phi i64 [ [[VAR2]], [[SCALAR_BODY]] ], [ 1, [[MIDDLE_BLOCK]] ]
1262; UNROLL-NO-IC-NEXT:    ret i64 [[VAR2_LCSSA]]
1263;
1264; UNROLL-NO-VF-LABEL: @constant_folded_previous_value(
1265; UNROLL-NO-VF-NEXT:  entry:
1266; UNROLL-NO-VF-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
1267; UNROLL-NO-VF:       vector.ph:
1268; UNROLL-NO-VF-NEXT:    [[TMP0:%.*]] = add i64 0, 1
1269; UNROLL-NO-VF-NEXT:    br label [[VECTOR_BODY:%.*]]
1270; UNROLL-NO-VF:       vector.body:
1271; UNROLL-NO-VF-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
1272; UNROLL-NO-VF-NEXT:    [[VECTOR_RECUR:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[TMP0]], [[VECTOR_BODY]] ]
1273; UNROLL-NO-VF-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 2
1274; UNROLL-NO-VF-NEXT:    [[TMP1:%.*]] = icmp eq i64 [[INDEX_NEXT]], 1000
1275; UNROLL-NO-VF-NEXT:    br i1 [[TMP1]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP12:![0-9]+]]
1276; UNROLL-NO-VF:       middle.block:
1277; UNROLL-NO-VF-NEXT:    br i1 true, label [[FOR_END:%.*]], label [[SCALAR_PH]]
1278; UNROLL-NO-VF:       scalar.ph:
1279; UNROLL-NO-VF-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ 1000, [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ]
1280; UNROLL-NO-VF-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i64 [ [[TMP0]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY]] ]
1281; UNROLL-NO-VF-NEXT:    br label [[SCALAR_BODY:%.*]]
1282; UNROLL-NO-VF:       scalar.body:
1283; UNROLL-NO-VF-NEXT:    [[I:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[I_NEXT:%.*]], [[SCALAR_BODY]] ]
1284; UNROLL-NO-VF-NEXT:    [[VAR2:%.*]] = phi i64 [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ], [ [[VAR3:%.*]], [[SCALAR_BODY]] ]
1285; UNROLL-NO-VF-NEXT:    [[VAR3]] = add i64 0, 1
1286; UNROLL-NO-VF-NEXT:    [[I_NEXT]] = add nuw nsw i64 [[I]], 1
1287; UNROLL-NO-VF-NEXT:    [[COND:%.*]] = icmp eq i64 [[I_NEXT]], 1000
1288; UNROLL-NO-VF-NEXT:    br i1 [[COND]], label [[FOR_END]], label [[SCALAR_BODY]], !llvm.loop [[LOOP13:![0-9]+]]
1289; UNROLL-NO-VF:       for.end:
1290; UNROLL-NO-VF-NEXT:    [[VAR2_LCSSA:%.*]] = phi i64 [ [[VAR2]], [[SCALAR_BODY]] ], [ [[TMP0]], [[MIDDLE_BLOCK]] ]
1291; UNROLL-NO-VF-NEXT:    ret i64 [[VAR2_LCSSA]]
1292;
1293; SINK-AFTER-LABEL: @constant_folded_previous_value(
1294; SINK-AFTER-NEXT:  entry:
1295; SINK-AFTER-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
1296; SINK-AFTER:       vector.ph:
1297; SINK-AFTER-NEXT:    br label [[VECTOR_BODY:%.*]]
1298; SINK-AFTER:       vector.body:
1299; SINK-AFTER-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
1300; SINK-AFTER-NEXT:    [[VECTOR_RECUR:%.*]] = phi <4 x i64> [ <i64 poison, i64 poison, i64 poison, i64 0>, [[VECTOR_PH]] ], [ splat (i64 1), [[VECTOR_BODY]] ]
1301; SINK-AFTER-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4
1302; SINK-AFTER-NEXT:    [[TMP0:%.*]] = icmp eq i64 [[INDEX_NEXT]], 1000
1303; SINK-AFTER-NEXT:    br i1 [[TMP0]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP12:![0-9]+]]
1304; SINK-AFTER:       middle.block:
1305; SINK-AFTER-NEXT:    br i1 true, label [[FOR_END:%.*]], label [[SCALAR_PH]]
1306; SINK-AFTER:       scalar.ph:
1307; SINK-AFTER-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ 1000, [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ]
1308; SINK-AFTER-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i64 [ 1, [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY]] ]
1309; SINK-AFTER-NEXT:    br label [[SCALAR_BODY:%.*]]
1310; SINK-AFTER:       scalar.body:
1311; SINK-AFTER-NEXT:    [[I:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[I_NEXT:%.*]], [[SCALAR_BODY]] ]
1312; SINK-AFTER-NEXT:    [[VAR2:%.*]] = phi i64 [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ], [ [[VAR3:%.*]], [[SCALAR_BODY]] ]
1313; SINK-AFTER-NEXT:    [[VAR3]] = add i64 0, 1
1314; SINK-AFTER-NEXT:    [[I_NEXT]] = add nuw nsw i64 [[I]], 1
1315; SINK-AFTER-NEXT:    [[COND:%.*]] = icmp eq i64 [[I_NEXT]], 1000
1316; SINK-AFTER-NEXT:    br i1 [[COND]], label [[FOR_END]], label [[SCALAR_BODY]], !llvm.loop [[LOOP13:![0-9]+]]
1317; SINK-AFTER:       for.end:
1318; SINK-AFTER-NEXT:    [[VAR2_LCSSA:%.*]] = phi i64 [ [[VAR2]], [[SCALAR_BODY]] ], [ 1, [[MIDDLE_BLOCK]] ]
1319; SINK-AFTER-NEXT:    ret i64 [[VAR2_LCSSA]]
1320;
1321entry:
1322  br label %scalar.body
1323
1324scalar.body:
1325  %i = phi i64 [ 0, %entry ], [ %i.next, %scalar.body ]
1326  %var2 = phi i64 [ 0, %entry ], [ %var3, %scalar.body ]
1327  %var3 = add i64 0, 1
1328  %i.next = add nuw nsw i64 %i, 1
1329  %cond = icmp eq i64 %i.next, 1000
1330  br i1 %cond, label %for.end, label %scalar.body
1331
1332for.end:
1333  ret i64 %var2
1334}
1335
1336; We vectorize this first order recurrence, by generating two
1337; extracts for the phi `val.phi` - one at the last index and
1338; another at the second last index. We need these 2 extracts because
1339; the first order recurrence phi is used outside the loop, so we require the phi
1340; itself and not its update (addx).
1341; Check the case when unrolled but not vectorized.
1342define i32 @extract_second_last_iteration(ptr %cval, i32 %x)  {
1343; UNROLL-NO-IC-LABEL: @extract_second_last_iteration(
1344; UNROLL-NO-IC-NEXT:  entry:
1345; UNROLL-NO-IC-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
1346; UNROLL-NO-IC:       vector.ph:
1347; UNROLL-NO-IC-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <4 x i32> poison, i32 [[X:%.*]], i64 0
1348; UNROLL-NO-IC-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <4 x i32> [[BROADCAST_SPLATINSERT]], <4 x i32> poison, <4 x i32> zeroinitializer
1349; UNROLL-NO-IC-NEXT:    br label [[VECTOR_BODY:%.*]]
1350; UNROLL-NO-IC:       vector.body:
1351; UNROLL-NO-IC-NEXT:    [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
1352; UNROLL-NO-IC-NEXT:    [[VEC_IND:%.*]] = phi <4 x i32> [ <i32 0, i32 1, i32 2, i32 3>, [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[VECTOR_BODY]] ]
1353; UNROLL-NO-IC-NEXT:    [[VECTOR_RECUR:%.*]] = phi <4 x i32> [ <i32 poison, i32 poison, i32 poison, i32 0>, [[VECTOR_PH]] ], [ [[TMP0:%.*]], [[VECTOR_BODY]] ]
1354; UNROLL-NO-IC-NEXT:    [[STEP_ADD:%.*]] = add <4 x i32> [[VEC_IND]], splat (i32 4)
1355; UNROLL-NO-IC-NEXT:    [[TMP0]] = add <4 x i32> [[STEP_ADD]], [[BROADCAST_SPLAT]]
1356; UNROLL-NO-IC-NEXT:    [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 8
1357; UNROLL-NO-IC-NEXT:    [[VEC_IND_NEXT]] = add <4 x i32> [[STEP_ADD]], splat (i32 4)
1358; UNROLL-NO-IC-NEXT:    [[TMP1:%.*]] = icmp eq i32 [[INDEX_NEXT]], 96
1359; UNROLL-NO-IC-NEXT:    br i1 [[TMP1]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP14:![0-9]+]]
1360; UNROLL-NO-IC:       middle.block:
1361; UNROLL-NO-IC-NEXT:    [[VECTOR_RECUR_EXTRACT_FOR_PHI:%.*]] = extractelement <4 x i32> [[TMP0]], i32 2
1362; UNROLL-NO-IC-NEXT:    [[VECTOR_RECUR_EXTRACT:%.*]] = extractelement <4 x i32> [[TMP0]], i32 3
1363; UNROLL-NO-IC-NEXT:    br i1 true, label [[FOR_END:%.*]], label [[SCALAR_PH]]
1364; UNROLL-NO-IC:       scalar.ph:
1365; UNROLL-NO-IC-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i32 [ 96, [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ]
1366; UNROLL-NO-IC-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i32 [ [[VECTOR_RECUR_EXTRACT]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY]] ]
1367; UNROLL-NO-IC-NEXT:    br label [[FOR_BODY:%.*]]
1368; UNROLL-NO-IC:       for.body:
1369; UNROLL-NO-IC-NEXT:    [[INC_PHI:%.*]] = phi i32 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[INC:%.*]], [[FOR_BODY]] ]
1370; UNROLL-NO-IC-NEXT:    [[VAL_PHI:%.*]] = phi i32 [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ], [ [[ADDX:%.*]], [[FOR_BODY]] ]
1371; UNROLL-NO-IC-NEXT:    [[INC]] = add i32 [[INC_PHI]], 1
1372; UNROLL-NO-IC-NEXT:    [[BC:%.*]] = zext i32 [[INC_PHI]] to i64
1373; UNROLL-NO-IC-NEXT:    [[ADDX]] = add i32 [[INC_PHI]], [[X]]
1374; UNROLL-NO-IC-NEXT:    [[CMP:%.*]] = icmp eq i32 [[INC_PHI]], 95
1375; UNROLL-NO-IC-NEXT:    br i1 [[CMP]], label [[FOR_END]], label [[FOR_BODY]], !llvm.loop [[LOOP15:![0-9]+]]
1376; UNROLL-NO-IC:       for.end:
1377; UNROLL-NO-IC-NEXT:    [[VAL_PHI_LCSSA:%.*]] = phi i32 [ [[VAL_PHI]], [[FOR_BODY]] ], [ [[VECTOR_RECUR_EXTRACT_FOR_PHI]], [[MIDDLE_BLOCK]] ]
1378; UNROLL-NO-IC-NEXT:    ret i32 [[VAL_PHI_LCSSA]]
1379;
1380; UNROLL-NO-VF-LABEL: @extract_second_last_iteration(
1381; UNROLL-NO-VF-NEXT:  entry:
1382; UNROLL-NO-VF-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
1383; UNROLL-NO-VF:       vector.ph:
1384; UNROLL-NO-VF-NEXT:    br label [[VECTOR_BODY:%.*]]
1385; UNROLL-NO-VF:       vector.body:
1386; UNROLL-NO-VF-NEXT:    [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
1387; UNROLL-NO-VF-NEXT:    [[VECTOR_RECUR:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[TMP3:%.*]], [[VECTOR_BODY]] ]
1388; UNROLL-NO-VF-NEXT:    [[TMP0:%.*]] = add i32 [[INDEX]], 0
1389; UNROLL-NO-VF-NEXT:    [[TMP1:%.*]] = add i32 [[INDEX]], 1
1390; UNROLL-NO-VF-NEXT:    [[TMP2:%.*]] = add i32 [[TMP0]], [[X:%.*]]
1391; UNROLL-NO-VF-NEXT:    [[TMP3]] = add i32 [[TMP1]], [[X]]
1392; UNROLL-NO-VF-NEXT:    [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 2
1393; UNROLL-NO-VF-NEXT:    [[TMP4:%.*]] = icmp eq i32 [[INDEX_NEXT]], 96
1394; UNROLL-NO-VF-NEXT:    br i1 [[TMP4]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP14:![0-9]+]]
1395; UNROLL-NO-VF:       middle.block:
1396; UNROLL-NO-VF-NEXT:    br i1 true, label [[FOR_END:%.*]], label [[SCALAR_PH]]
1397; UNROLL-NO-VF:       scalar.ph:
1398; UNROLL-NO-VF-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i32 [ 96, [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ]
1399; UNROLL-NO-VF-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i32 [ [[TMP3]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY]] ]
1400; UNROLL-NO-VF-NEXT:    br label [[FOR_BODY:%.*]]
1401; UNROLL-NO-VF:       for.body:
1402; UNROLL-NO-VF-NEXT:    [[INC_PHI:%.*]] = phi i32 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[INC:%.*]], [[FOR_BODY]] ]
1403; UNROLL-NO-VF-NEXT:    [[VAL_PHI:%.*]] = phi i32 [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ], [ [[ADDX:%.*]], [[FOR_BODY]] ]
1404; UNROLL-NO-VF-NEXT:    [[INC]] = add i32 [[INC_PHI]], 1
1405; UNROLL-NO-VF-NEXT:    [[BC:%.*]] = zext i32 [[INC_PHI]] to i64
1406; UNROLL-NO-VF-NEXT:    [[ADDX]] = add i32 [[INC_PHI]], [[X]]
1407; UNROLL-NO-VF-NEXT:    [[CMP:%.*]] = icmp eq i32 [[INC_PHI]], 95
1408; UNROLL-NO-VF-NEXT:    br i1 [[CMP]], label [[FOR_END]], label [[FOR_BODY]], !llvm.loop [[LOOP15:![0-9]+]]
1409; UNROLL-NO-VF:       for.end:
1410; UNROLL-NO-VF-NEXT:    [[VAL_PHI_LCSSA:%.*]] = phi i32 [ [[VAL_PHI]], [[FOR_BODY]] ], [ [[TMP2]], [[MIDDLE_BLOCK]] ]
1411; UNROLL-NO-VF-NEXT:    ret i32 [[VAL_PHI_LCSSA]]
1412;
1413; SINK-AFTER-LABEL: @extract_second_last_iteration(
1414; SINK-AFTER-NEXT:  entry:
1415; SINK-AFTER-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
1416; SINK-AFTER:       vector.ph:
1417; SINK-AFTER-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <4 x i32> poison, i32 [[X:%.*]], i64 0
1418; SINK-AFTER-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <4 x i32> [[BROADCAST_SPLATINSERT]], <4 x i32> poison, <4 x i32> zeroinitializer
1419; SINK-AFTER-NEXT:    br label [[VECTOR_BODY:%.*]]
1420; SINK-AFTER:       vector.body:
1421; SINK-AFTER-NEXT:    [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
1422; SINK-AFTER-NEXT:    [[VEC_IND:%.*]] = phi <4 x i32> [ <i32 0, i32 1, i32 2, i32 3>, [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[VECTOR_BODY]] ]
1423; SINK-AFTER-NEXT:    [[VECTOR_RECUR:%.*]] = phi <4 x i32> [ <i32 poison, i32 poison, i32 poison, i32 0>, [[VECTOR_PH]] ], [ [[TMP0:%.*]], [[VECTOR_BODY]] ]
1424; SINK-AFTER-NEXT:    [[TMP0]] = add <4 x i32> [[VEC_IND]], [[BROADCAST_SPLAT]]
1425; SINK-AFTER-NEXT:    [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 4
1426; SINK-AFTER-NEXT:    [[VEC_IND_NEXT]] = add <4 x i32> [[VEC_IND]], splat (i32 4)
1427; SINK-AFTER-NEXT:    [[TMP1:%.*]] = icmp eq i32 [[INDEX_NEXT]], 96
1428; SINK-AFTER-NEXT:    br i1 [[TMP1]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP14:![0-9]+]]
1429; SINK-AFTER:       middle.block:
1430; SINK-AFTER-NEXT:    [[VECTOR_RECUR_EXTRACT_FOR_PHI:%.*]] = extractelement <4 x i32> [[TMP0]], i32 2
1431; SINK-AFTER-NEXT:    [[VECTOR_RECUR_EXTRACT:%.*]] = extractelement <4 x i32> [[TMP0]], i32 3
1432; SINK-AFTER-NEXT:    br i1 true, label [[FOR_END:%.*]], label [[SCALAR_PH]]
1433; SINK-AFTER:       scalar.ph:
1434; SINK-AFTER-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i32 [ 96, [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ]
1435; SINK-AFTER-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i32 [ [[VECTOR_RECUR_EXTRACT]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY]] ]
1436; SINK-AFTER-NEXT:    br label [[FOR_BODY:%.*]]
1437; SINK-AFTER:       for.body:
1438; SINK-AFTER-NEXT:    [[INC_PHI:%.*]] = phi i32 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[INC:%.*]], [[FOR_BODY]] ]
1439; SINK-AFTER-NEXT:    [[VAL_PHI:%.*]] = phi i32 [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ], [ [[ADDX:%.*]], [[FOR_BODY]] ]
1440; SINK-AFTER-NEXT:    [[INC]] = add i32 [[INC_PHI]], 1
1441; SINK-AFTER-NEXT:    [[BC:%.*]] = zext i32 [[INC_PHI]] to i64
1442; SINK-AFTER-NEXT:    [[ADDX]] = add i32 [[INC_PHI]], [[X]]
1443; SINK-AFTER-NEXT:    [[CMP:%.*]] = icmp eq i32 [[INC_PHI]], 95
1444; SINK-AFTER-NEXT:    br i1 [[CMP]], label [[FOR_END]], label [[FOR_BODY]], !llvm.loop [[LOOP15:![0-9]+]]
1445; SINK-AFTER:       for.end:
1446; SINK-AFTER-NEXT:    [[VAL_PHI_LCSSA:%.*]] = phi i32 [ [[VAL_PHI]], [[FOR_BODY]] ], [ [[VECTOR_RECUR_EXTRACT_FOR_PHI]], [[MIDDLE_BLOCK]] ]
1447; SINK-AFTER-NEXT:    ret i32 [[VAL_PHI_LCSSA]]
1448;
1449entry:
1450  br label %for.body
1451
1452for.body:
1453  %inc.phi = phi i32 [ 0, %entry ], [ %inc, %for.body ]
1454  %val.phi = phi i32 [ 0, %entry ], [ %addx, %for.body ]
1455  %inc = add i32 %inc.phi, 1
1456  %bc = zext i32 %inc.phi to i64
1457  %addx = add i32 %inc.phi, %x
1458  %cmp = icmp eq i32 %inc.phi, 95
1459  br i1 %cmp, label %for.end, label %for.body
1460
1461for.end:
1462  ret i32 %val.phi
1463}
1464
1465; We vectorize this first order recurrence, with a set of insertelements for
1466; each unrolled part. Make sure these insertelements are generated in-order,
1467; because the shuffle of the first order recurrence will be added after the
1468; insertelement of the last part UF - 1, assuming the latter appears after the
1469; insertelements of all other parts.
1470;
1471; int PR33613(ptr b, double j, int d) {
1472;   int a = 0;
1473;   for(int i = 0; i < 10240; i++, b+=25) {
1474;     double f = b[d]; // Scalarize to form insertelements
1475;     if (j * f)
1476;       a++;
1477;     j = f;
1478;   }
1479;   return a;
1480; }
1481;
1482;
1483define i32 @PR33613(ptr %b, double %j, i32 %d) {
1484; UNROLL-NO-IC-LABEL: @PR33613(
1485; UNROLL-NO-IC-NEXT:  entry:
1486; UNROLL-NO-IC-NEXT:    [[IDXPROM:%.*]] = sext i32 [[D:%.*]] to i64
1487; UNROLL-NO-IC-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
1488; UNROLL-NO-IC:       vector.ph:
1489; UNROLL-NO-IC-NEXT:    [[IND_END:%.*]] = getelementptr i8, ptr [[B:%.*]], i64 2048000
1490; UNROLL-NO-IC-NEXT:    [[VECTOR_RECUR_INIT:%.*]] = insertelement <4 x double> poison, double [[J:%.*]], i32 3
1491; UNROLL-NO-IC-NEXT:    br label [[VECTOR_BODY:%.*]]
1492; UNROLL-NO-IC:       vector.body:
1493; UNROLL-NO-IC-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
1494; UNROLL-NO-IC-NEXT:    [[VEC_PHI:%.*]] = phi <4 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP40:%.*]], [[VECTOR_BODY]] ]
1495; UNROLL-NO-IC-NEXT:    [[VEC_PHI2:%.*]] = phi <4 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP41:%.*]], [[VECTOR_BODY]] ]
1496; UNROLL-NO-IC-NEXT:    [[VECTOR_RECUR:%.*]] = phi <4 x double> [ [[VECTOR_RECUR_INIT]], [[VECTOR_PH]] ], [ [[TMP31:%.*]], [[VECTOR_BODY]] ]
1497; UNROLL-NO-IC-NEXT:    [[OFFSET_IDX:%.*]] = mul i64 [[INDEX]], 200
1498; UNROLL-NO-IC-NEXT:    [[TMP0:%.*]] = add i64 [[OFFSET_IDX]], 0
1499; UNROLL-NO-IC-NEXT:    [[TMP1:%.*]] = add i64 [[OFFSET_IDX]], 200
1500; UNROLL-NO-IC-NEXT:    [[TMP2:%.*]] = add i64 [[OFFSET_IDX]], 400
1501; UNROLL-NO-IC-NEXT:    [[TMP3:%.*]] = add i64 [[OFFSET_IDX]], 600
1502; UNROLL-NO-IC-NEXT:    [[TMP4:%.*]] = add i64 [[OFFSET_IDX]], 800
1503; UNROLL-NO-IC-NEXT:    [[TMP5:%.*]] = add i64 [[OFFSET_IDX]], 1000
1504; UNROLL-NO-IC-NEXT:    [[TMP6:%.*]] = add i64 [[OFFSET_IDX]], 1200
1505; UNROLL-NO-IC-NEXT:    [[TMP7:%.*]] = add i64 [[OFFSET_IDX]], 1400
1506; UNROLL-NO-IC-NEXT:    [[NEXT_GEP:%.*]] = getelementptr i8, ptr [[B]], i64 [[TMP0]]
1507; UNROLL-NO-IC-NEXT:    [[NEXT_GEP3:%.*]] = getelementptr i8, ptr [[B]], i64 [[TMP1]]
1508; UNROLL-NO-IC-NEXT:    [[NEXT_GEP4:%.*]] = getelementptr i8, ptr [[B]], i64 [[TMP2]]
1509; UNROLL-NO-IC-NEXT:    [[NEXT_GEP5:%.*]] = getelementptr i8, ptr [[B]], i64 [[TMP3]]
1510; UNROLL-NO-IC-NEXT:    [[NEXT_GEP6:%.*]] = getelementptr i8, ptr [[B]], i64 [[TMP4]]
1511; UNROLL-NO-IC-NEXT:    [[NEXT_GEP7:%.*]] = getelementptr i8, ptr [[B]], i64 [[TMP5]]
1512; UNROLL-NO-IC-NEXT:    [[NEXT_GEP8:%.*]] = getelementptr i8, ptr [[B]], i64 [[TMP6]]
1513; UNROLL-NO-IC-NEXT:    [[NEXT_GEP9:%.*]] = getelementptr i8, ptr [[B]], i64 [[TMP7]]
1514; UNROLL-NO-IC-NEXT:    [[TMP8:%.*]] = getelementptr inbounds double, ptr [[NEXT_GEP]], i64 [[IDXPROM]]
1515; UNROLL-NO-IC-NEXT:    [[TMP9:%.*]] = getelementptr inbounds double, ptr [[NEXT_GEP3]], i64 [[IDXPROM]]
1516; UNROLL-NO-IC-NEXT:    [[TMP10:%.*]] = getelementptr inbounds double, ptr [[NEXT_GEP4]], i64 [[IDXPROM]]
1517; UNROLL-NO-IC-NEXT:    [[TMP11:%.*]] = getelementptr inbounds double, ptr [[NEXT_GEP5]], i64 [[IDXPROM]]
1518; UNROLL-NO-IC-NEXT:    [[TMP12:%.*]] = getelementptr inbounds double, ptr [[NEXT_GEP6]], i64 [[IDXPROM]]
1519; UNROLL-NO-IC-NEXT:    [[TMP13:%.*]] = getelementptr inbounds double, ptr [[NEXT_GEP7]], i64 [[IDXPROM]]
1520; UNROLL-NO-IC-NEXT:    [[TMP14:%.*]] = getelementptr inbounds double, ptr [[NEXT_GEP8]], i64 [[IDXPROM]]
1521; UNROLL-NO-IC-NEXT:    [[TMP15:%.*]] = getelementptr inbounds double, ptr [[NEXT_GEP9]], i64 [[IDXPROM]]
1522; UNROLL-NO-IC-NEXT:    [[TMP16:%.*]] = load double, ptr [[TMP8]], align 8
1523; UNROLL-NO-IC-NEXT:    [[TMP17:%.*]] = load double, ptr [[TMP9]], align 8
1524; UNROLL-NO-IC-NEXT:    [[TMP18:%.*]] = load double, ptr [[TMP10]], align 8
1525; UNROLL-NO-IC-NEXT:    [[TMP19:%.*]] = load double, ptr [[TMP11]], align 8
1526; UNROLL-NO-IC-NEXT:    [[TMP20:%.*]] = insertelement <4 x double> poison, double [[TMP16]], i32 0
1527; UNROLL-NO-IC-NEXT:    [[TMP21:%.*]] = insertelement <4 x double> [[TMP20]], double [[TMP17]], i32 1
1528; UNROLL-NO-IC-NEXT:    [[TMP22:%.*]] = insertelement <4 x double> [[TMP21]], double [[TMP18]], i32 2
1529; UNROLL-NO-IC-NEXT:    [[TMP23:%.*]] = insertelement <4 x double> [[TMP22]], double [[TMP19]], i32 3
1530; UNROLL-NO-IC-NEXT:    [[TMP24:%.*]] = load double, ptr [[TMP12]], align 8
1531; UNROLL-NO-IC-NEXT:    [[TMP25:%.*]] = load double, ptr [[TMP13]], align 8
1532; UNROLL-NO-IC-NEXT:    [[TMP26:%.*]] = load double, ptr [[TMP14]], align 8
1533; UNROLL-NO-IC-NEXT:    [[TMP27:%.*]] = load double, ptr [[TMP15]], align 8
1534; UNROLL-NO-IC-NEXT:    [[TMP28:%.*]] = insertelement <4 x double> poison, double [[TMP24]], i32 0
1535; UNROLL-NO-IC-NEXT:    [[TMP29:%.*]] = insertelement <4 x double> [[TMP28]], double [[TMP25]], i32 1
1536; UNROLL-NO-IC-NEXT:    [[TMP30:%.*]] = insertelement <4 x double> [[TMP29]], double [[TMP26]], i32 2
1537; UNROLL-NO-IC-NEXT:    [[TMP31]] = insertelement <4 x double> [[TMP30]], double [[TMP27]], i32 3
1538; UNROLL-NO-IC-NEXT:    [[TMP32:%.*]] = shufflevector <4 x double> [[VECTOR_RECUR]], <4 x double> [[TMP23]], <4 x i32> <i32 3, i32 4, i32 5, i32 6>
1539; UNROLL-NO-IC-NEXT:    [[TMP33:%.*]] = shufflevector <4 x double> [[TMP23]], <4 x double> [[TMP31]], <4 x i32> <i32 3, i32 4, i32 5, i32 6>
1540; UNROLL-NO-IC-NEXT:    [[TMP34:%.*]] = fmul <4 x double> [[TMP32]], [[TMP23]]
1541; UNROLL-NO-IC-NEXT:    [[TMP35:%.*]] = fmul <4 x double> [[TMP33]], [[TMP31]]
1542; UNROLL-NO-IC-NEXT:    [[TMP36:%.*]] = fcmp une <4 x double> [[TMP34]], zeroinitializer
1543; UNROLL-NO-IC-NEXT:    [[TMP37:%.*]] = fcmp une <4 x double> [[TMP35]], zeroinitializer
1544; UNROLL-NO-IC-NEXT:    [[TMP38:%.*]] = zext <4 x i1> [[TMP36]] to <4 x i32>
1545; UNROLL-NO-IC-NEXT:    [[TMP39:%.*]] = zext <4 x i1> [[TMP37]] to <4 x i32>
1546; UNROLL-NO-IC-NEXT:    [[TMP40]] = add <4 x i32> [[VEC_PHI]], [[TMP38]]
1547; UNROLL-NO-IC-NEXT:    [[TMP41]] = add <4 x i32> [[VEC_PHI2]], [[TMP39]]
1548; UNROLL-NO-IC-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 8
1549; UNROLL-NO-IC-NEXT:    [[TMP42:%.*]] = icmp eq i64 [[INDEX_NEXT]], 10240
1550; UNROLL-NO-IC-NEXT:    br i1 [[TMP42]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP16:![0-9]+]]
1551; UNROLL-NO-IC:       middle.block:
1552; UNROLL-NO-IC-NEXT:    [[BIN_RDX:%.*]] = add <4 x i32> [[TMP41]], [[TMP40]]
1553; UNROLL-NO-IC-NEXT:    [[TMP43:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[BIN_RDX]])
1554; UNROLL-NO-IC-NEXT:    br i1 true, label [[FOR_COND_CLEANUP:%.*]], label [[SCALAR_PH]]
1555; UNROLL-NO-IC:       scalar.ph:
1556; UNROLL-NO-IC-NEXT:    [[BC_RESUME_VAL:%.*]] = phi ptr [ [[IND_END]], [[MIDDLE_BLOCK]] ], [ [[B]], [[ENTRY:%.*]] ]
1557; UNROLL-NO-IC-NEXT:    [[BC_RESUME_VAL1:%.*]] = phi i32 [ 10240, [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY]] ]
1558; UNROLL-NO-IC-NEXT:    [[BC_MERGE_RDX:%.*]] = phi i32 [ [[TMP43]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY]] ]
1559; UNROLL-NO-IC-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi double [ [[TMP27]], [[MIDDLE_BLOCK]] ], [ [[J]], [[ENTRY]] ]
1560; UNROLL-NO-IC-NEXT:    br label [[FOR_BODY:%.*]]
1561; UNROLL-NO-IC:       for.cond.cleanup:
1562; UNROLL-NO-IC-NEXT:    [[A_1_LCSSA:%.*]] = phi i32 [ [[A_1:%.*]], [[FOR_BODY]] ], [ [[TMP43]], [[MIDDLE_BLOCK]] ]
1563; UNROLL-NO-IC-NEXT:    ret i32 [[A_1_LCSSA]]
1564; UNROLL-NO-IC:       for.body:
1565; UNROLL-NO-IC-NEXT:    [[B_ADDR_012:%.*]] = phi ptr [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[ADD_PTR:%.*]], [[FOR_BODY]] ]
1566; UNROLL-NO-IC-NEXT:    [[I_011:%.*]] = phi i32 [ [[BC_RESUME_VAL1]], [[SCALAR_PH]] ], [ [[INC1:%.*]], [[FOR_BODY]] ]
1567; UNROLL-NO-IC-NEXT:    [[A_010:%.*]] = phi i32 [ [[BC_MERGE_RDX]], [[SCALAR_PH]] ], [ [[A_1]], [[FOR_BODY]] ]
1568; UNROLL-NO-IC-NEXT:    [[J_ADDR_09:%.*]] = phi double [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ], [ [[TMP44:%.*]], [[FOR_BODY]] ]
1569; UNROLL-NO-IC-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds double, ptr [[B_ADDR_012]], i64 [[IDXPROM]]
1570; UNROLL-NO-IC-NEXT:    [[TMP44]] = load double, ptr [[ARRAYIDX]], align 8
1571; UNROLL-NO-IC-NEXT:    [[MUL:%.*]] = fmul double [[J_ADDR_09]], [[TMP44]]
1572; UNROLL-NO-IC-NEXT:    [[TOBOOL:%.*]] = fcmp une double [[MUL]], 0.000000e+00
1573; UNROLL-NO-IC-NEXT:    [[INC:%.*]] = zext i1 [[TOBOOL]] to i32
1574; UNROLL-NO-IC-NEXT:    [[A_1]] = add nsw i32 [[A_010]], [[INC]]
1575; UNROLL-NO-IC-NEXT:    [[INC1]] = add nuw nsw i32 [[I_011]], 1
1576; UNROLL-NO-IC-NEXT:    [[ADD_PTR]] = getelementptr inbounds double, ptr [[B_ADDR_012]], i64 25
1577; UNROLL-NO-IC-NEXT:    [[EXITCOND:%.*]] = icmp eq i32 [[INC1]], 10240
1578; UNROLL-NO-IC-NEXT:    br i1 [[EXITCOND]], label [[FOR_COND_CLEANUP]], label [[FOR_BODY]], !llvm.loop [[LOOP17:![0-9]+]]
1579;
1580; UNROLL-NO-VF-LABEL: @PR33613(
1581; UNROLL-NO-VF-NEXT:  entry:
1582; UNROLL-NO-VF-NEXT:    [[IDXPROM:%.*]] = sext i32 [[D:%.*]] to i64
1583; UNROLL-NO-VF-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
1584; UNROLL-NO-VF:       vector.ph:
1585; UNROLL-NO-VF-NEXT:    [[IND_END:%.*]] = getelementptr i8, ptr [[B:%.*]], i64 2048000
1586; UNROLL-NO-VF-NEXT:    br label [[VECTOR_BODY:%.*]]
1587; UNROLL-NO-VF:       vector.body:
1588; UNROLL-NO-VF-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
1589; UNROLL-NO-VF-NEXT:    [[VEC_PHI:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[TMP12:%.*]], [[VECTOR_BODY]] ]
1590; UNROLL-NO-VF-NEXT:    [[VEC_PHI2:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[TMP13:%.*]], [[VECTOR_BODY]] ]
1591; UNROLL-NO-VF-NEXT:    [[VECTOR_RECUR:%.*]] = phi double [ [[J:%.*]], [[VECTOR_PH]] ], [ [[TMP5:%.*]], [[VECTOR_BODY]] ]
1592; UNROLL-NO-VF-NEXT:    [[OFFSET_IDX:%.*]] = mul i64 [[INDEX]], 200
1593; UNROLL-NO-VF-NEXT:    [[TMP0:%.*]] = add i64 [[OFFSET_IDX]], 0
1594; UNROLL-NO-VF-NEXT:    [[TMP1:%.*]] = add i64 [[OFFSET_IDX]], 200
1595; UNROLL-NO-VF-NEXT:    [[NEXT_GEP:%.*]] = getelementptr i8, ptr [[B]], i64 [[TMP0]]
1596; UNROLL-NO-VF-NEXT:    [[NEXT_GEP3:%.*]] = getelementptr i8, ptr [[B]], i64 [[TMP1]]
1597; UNROLL-NO-VF-NEXT:    [[TMP2:%.*]] = getelementptr inbounds double, ptr [[NEXT_GEP]], i64 [[IDXPROM]]
1598; UNROLL-NO-VF-NEXT:    [[TMP3:%.*]] = getelementptr inbounds double, ptr [[NEXT_GEP3]], i64 [[IDXPROM]]
1599; UNROLL-NO-VF-NEXT:    [[TMP4:%.*]] = load double, ptr [[TMP2]], align 8
1600; UNROLL-NO-VF-NEXT:    [[TMP5]] = load double, ptr [[TMP3]], align 8
1601; UNROLL-NO-VF-NEXT:    [[TMP6:%.*]] = fmul double [[VECTOR_RECUR]], [[TMP4]]
1602; UNROLL-NO-VF-NEXT:    [[TMP7:%.*]] = fmul double [[TMP4]], [[TMP5]]
1603; UNROLL-NO-VF-NEXT:    [[TMP8:%.*]] = fcmp une double [[TMP6]], 0.000000e+00
1604; UNROLL-NO-VF-NEXT:    [[TMP9:%.*]] = fcmp une double [[TMP7]], 0.000000e+00
1605; UNROLL-NO-VF-NEXT:    [[TMP10:%.*]] = zext i1 [[TMP8]] to i32
1606; UNROLL-NO-VF-NEXT:    [[TMP11:%.*]] = zext i1 [[TMP9]] to i32
1607; UNROLL-NO-VF-NEXT:    [[TMP12]] = add i32 [[VEC_PHI]], [[TMP10]]
1608; UNROLL-NO-VF-NEXT:    [[TMP13]] = add i32 [[VEC_PHI2]], [[TMP11]]
1609; UNROLL-NO-VF-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 2
1610; UNROLL-NO-VF-NEXT:    [[TMP14:%.*]] = icmp eq i64 [[INDEX_NEXT]], 10240
1611; UNROLL-NO-VF-NEXT:    br i1 [[TMP14]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP16:![0-9]+]]
1612; UNROLL-NO-VF:       middle.block:
1613; UNROLL-NO-VF-NEXT:    [[BIN_RDX:%.*]] = add i32 [[TMP13]], [[TMP12]]
1614; UNROLL-NO-VF-NEXT:    br i1 true, label [[FOR_COND_CLEANUP:%.*]], label [[SCALAR_PH]]
1615; UNROLL-NO-VF:       scalar.ph:
1616; UNROLL-NO-VF-NEXT:    [[BC_RESUME_VAL:%.*]] = phi ptr [ [[IND_END]], [[MIDDLE_BLOCK]] ], [ [[B]], [[ENTRY:%.*]] ]
1617; UNROLL-NO-VF-NEXT:    [[BC_RESUME_VAL1:%.*]] = phi i32 [ 10240, [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY]] ]
1618; UNROLL-NO-VF-NEXT:    [[BC_MERGE_RDX:%.*]] = phi i32 [ [[BIN_RDX]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY]] ]
1619; UNROLL-NO-VF-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi double [ [[TMP5]], [[MIDDLE_BLOCK]] ], [ [[J]], [[ENTRY]] ]
1620; UNROLL-NO-VF-NEXT:    br label [[FOR_BODY:%.*]]
1621; UNROLL-NO-VF:       for.cond.cleanup:
1622; UNROLL-NO-VF-NEXT:    [[A_1_LCSSA:%.*]] = phi i32 [ [[A_1:%.*]], [[FOR_BODY]] ], [ [[BIN_RDX]], [[MIDDLE_BLOCK]] ]
1623; UNROLL-NO-VF-NEXT:    ret i32 [[A_1_LCSSA]]
1624; UNROLL-NO-VF:       for.body:
1625; UNROLL-NO-VF-NEXT:    [[B_ADDR_012:%.*]] = phi ptr [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[ADD_PTR:%.*]], [[FOR_BODY]] ]
1626; UNROLL-NO-VF-NEXT:    [[I_011:%.*]] = phi i32 [ [[BC_RESUME_VAL1]], [[SCALAR_PH]] ], [ [[INC1:%.*]], [[FOR_BODY]] ]
1627; UNROLL-NO-VF-NEXT:    [[A_010:%.*]] = phi i32 [ [[BC_MERGE_RDX]], [[SCALAR_PH]] ], [ [[A_1]], [[FOR_BODY]] ]
1628; UNROLL-NO-VF-NEXT:    [[J_ADDR_09:%.*]] = phi double [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ], [ [[TMP15:%.*]], [[FOR_BODY]] ]
1629; UNROLL-NO-VF-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds double, ptr [[B_ADDR_012]], i64 [[IDXPROM]]
1630; UNROLL-NO-VF-NEXT:    [[TMP15]] = load double, ptr [[ARRAYIDX]], align 8
1631; UNROLL-NO-VF-NEXT:    [[MUL:%.*]] = fmul double [[J_ADDR_09]], [[TMP15]]
1632; UNROLL-NO-VF-NEXT:    [[TOBOOL:%.*]] = fcmp une double [[MUL]], 0.000000e+00
1633; UNROLL-NO-VF-NEXT:    [[INC:%.*]] = zext i1 [[TOBOOL]] to i32
1634; UNROLL-NO-VF-NEXT:    [[A_1]] = add nsw i32 [[A_010]], [[INC]]
1635; UNROLL-NO-VF-NEXT:    [[INC1]] = add nuw nsw i32 [[I_011]], 1
1636; UNROLL-NO-VF-NEXT:    [[ADD_PTR]] = getelementptr inbounds double, ptr [[B_ADDR_012]], i64 25
1637; UNROLL-NO-VF-NEXT:    [[EXITCOND:%.*]] = icmp eq i32 [[INC1]], 10240
1638; UNROLL-NO-VF-NEXT:    br i1 [[EXITCOND]], label [[FOR_COND_CLEANUP]], label [[FOR_BODY]], !llvm.loop [[LOOP17:![0-9]+]]
1639;
1640; SINK-AFTER-LABEL: @PR33613(
1641; SINK-AFTER-NEXT:  entry:
1642; SINK-AFTER-NEXT:    [[IDXPROM:%.*]] = sext i32 [[D:%.*]] to i64
1643; SINK-AFTER-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
1644; SINK-AFTER:       vector.ph:
1645; SINK-AFTER-NEXT:    [[IND_END:%.*]] = getelementptr i8, ptr [[B:%.*]], i64 2048000
1646; SINK-AFTER-NEXT:    [[VECTOR_RECUR_INIT:%.*]] = insertelement <4 x double> poison, double [[J:%.*]], i32 3
1647; SINK-AFTER-NEXT:    br label [[VECTOR_BODY:%.*]]
1648; SINK-AFTER:       vector.body:
1649; SINK-AFTER-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
1650; SINK-AFTER-NEXT:    [[VEC_PHI:%.*]] = phi <4 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP20:%.*]], [[VECTOR_BODY]] ]
1651; SINK-AFTER-NEXT:    [[VECTOR_RECUR:%.*]] = phi <4 x double> [ [[VECTOR_RECUR_INIT]], [[VECTOR_PH]] ], [ [[TMP15:%.*]], [[VECTOR_BODY]] ]
1652; SINK-AFTER-NEXT:    [[OFFSET_IDX:%.*]] = mul i64 [[INDEX]], 200
1653; SINK-AFTER-NEXT:    [[TMP0:%.*]] = add i64 [[OFFSET_IDX]], 0
1654; SINK-AFTER-NEXT:    [[TMP1:%.*]] = add i64 [[OFFSET_IDX]], 200
1655; SINK-AFTER-NEXT:    [[TMP2:%.*]] = add i64 [[OFFSET_IDX]], 400
1656; SINK-AFTER-NEXT:    [[TMP3:%.*]] = add i64 [[OFFSET_IDX]], 600
1657; SINK-AFTER-NEXT:    [[NEXT_GEP:%.*]] = getelementptr i8, ptr [[B]], i64 [[TMP0]]
1658; SINK-AFTER-NEXT:    [[NEXT_GEP2:%.*]] = getelementptr i8, ptr [[B]], i64 [[TMP1]]
1659; SINK-AFTER-NEXT:    [[NEXT_GEP3:%.*]] = getelementptr i8, ptr [[B]], i64 [[TMP2]]
1660; SINK-AFTER-NEXT:    [[NEXT_GEP4:%.*]] = getelementptr i8, ptr [[B]], i64 [[TMP3]]
1661; SINK-AFTER-NEXT:    [[TMP4:%.*]] = getelementptr inbounds double, ptr [[NEXT_GEP]], i64 [[IDXPROM]]
1662; SINK-AFTER-NEXT:    [[TMP5:%.*]] = getelementptr inbounds double, ptr [[NEXT_GEP2]], i64 [[IDXPROM]]
1663; SINK-AFTER-NEXT:    [[TMP6:%.*]] = getelementptr inbounds double, ptr [[NEXT_GEP3]], i64 [[IDXPROM]]
1664; SINK-AFTER-NEXT:    [[TMP7:%.*]] = getelementptr inbounds double, ptr [[NEXT_GEP4]], i64 [[IDXPROM]]
1665; SINK-AFTER-NEXT:    [[TMP8:%.*]] = load double, ptr [[TMP4]], align 8
1666; SINK-AFTER-NEXT:    [[TMP9:%.*]] = load double, ptr [[TMP5]], align 8
1667; SINK-AFTER-NEXT:    [[TMP10:%.*]] = load double, ptr [[TMP6]], align 8
1668; SINK-AFTER-NEXT:    [[TMP11:%.*]] = load double, ptr [[TMP7]], align 8
1669; SINK-AFTER-NEXT:    [[TMP12:%.*]] = insertelement <4 x double> poison, double [[TMP8]], i32 0
1670; SINK-AFTER-NEXT:    [[TMP13:%.*]] = insertelement <4 x double> [[TMP12]], double [[TMP9]], i32 1
1671; SINK-AFTER-NEXT:    [[TMP14:%.*]] = insertelement <4 x double> [[TMP13]], double [[TMP10]], i32 2
1672; SINK-AFTER-NEXT:    [[TMP15]] = insertelement <4 x double> [[TMP14]], double [[TMP11]], i32 3
1673; SINK-AFTER-NEXT:    [[TMP16:%.*]] = shufflevector <4 x double> [[VECTOR_RECUR]], <4 x double> [[TMP15]], <4 x i32> <i32 3, i32 4, i32 5, i32 6>
1674; SINK-AFTER-NEXT:    [[TMP17:%.*]] = fmul <4 x double> [[TMP16]], [[TMP15]]
1675; SINK-AFTER-NEXT:    [[TMP18:%.*]] = fcmp une <4 x double> [[TMP17]], zeroinitializer
1676; SINK-AFTER-NEXT:    [[TMP19:%.*]] = zext <4 x i1> [[TMP18]] to <4 x i32>
1677; SINK-AFTER-NEXT:    [[TMP20]] = add <4 x i32> [[VEC_PHI]], [[TMP19]]
1678; SINK-AFTER-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4
1679; SINK-AFTER-NEXT:    [[TMP21:%.*]] = icmp eq i64 [[INDEX_NEXT]], 10240
1680; SINK-AFTER-NEXT:    br i1 [[TMP21]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP16:![0-9]+]]
1681; SINK-AFTER:       middle.block:
1682; SINK-AFTER-NEXT:    [[TMP22:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[TMP20]])
1683; SINK-AFTER-NEXT:    br i1 true, label [[FOR_COND_CLEANUP:%.*]], label [[SCALAR_PH]]
1684; SINK-AFTER:       scalar.ph:
1685; SINK-AFTER-NEXT:    [[BC_RESUME_VAL:%.*]] = phi ptr [ [[IND_END]], [[MIDDLE_BLOCK]] ], [ [[B]], [[ENTRY:%.*]] ]
1686; SINK-AFTER-NEXT:    [[BC_RESUME_VAL1:%.*]] = phi i32 [ 10240, [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY]] ]
1687; SINK-AFTER-NEXT:    [[BC_MERGE_RDX:%.*]] = phi i32 [ [[TMP22]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY]] ]
1688; SINK-AFTER-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi double [ [[TMP11]], [[MIDDLE_BLOCK]] ], [ [[J]], [[ENTRY]] ]
1689; SINK-AFTER-NEXT:    br label [[FOR_BODY:%.*]]
1690; SINK-AFTER:       for.cond.cleanup:
1691; SINK-AFTER-NEXT:    [[A_1_LCSSA:%.*]] = phi i32 [ [[A_1:%.*]], [[FOR_BODY]] ], [ [[TMP22]], [[MIDDLE_BLOCK]] ]
1692; SINK-AFTER-NEXT:    ret i32 [[A_1_LCSSA]]
1693; SINK-AFTER:       for.body:
1694; SINK-AFTER-NEXT:    [[B_ADDR_012:%.*]] = phi ptr [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[ADD_PTR:%.*]], [[FOR_BODY]] ]
1695; SINK-AFTER-NEXT:    [[I_011:%.*]] = phi i32 [ [[BC_RESUME_VAL1]], [[SCALAR_PH]] ], [ [[INC1:%.*]], [[FOR_BODY]] ]
1696; SINK-AFTER-NEXT:    [[A_010:%.*]] = phi i32 [ [[BC_MERGE_RDX]], [[SCALAR_PH]] ], [ [[A_1]], [[FOR_BODY]] ]
1697; SINK-AFTER-NEXT:    [[J_ADDR_09:%.*]] = phi double [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ], [ [[TMP23:%.*]], [[FOR_BODY]] ]
1698; SINK-AFTER-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds double, ptr [[B_ADDR_012]], i64 [[IDXPROM]]
1699; SINK-AFTER-NEXT:    [[TMP23]] = load double, ptr [[ARRAYIDX]], align 8
1700; SINK-AFTER-NEXT:    [[MUL:%.*]] = fmul double [[J_ADDR_09]], [[TMP23]]
1701; SINK-AFTER-NEXT:    [[TOBOOL:%.*]] = fcmp une double [[MUL]], 0.000000e+00
1702; SINK-AFTER-NEXT:    [[INC:%.*]] = zext i1 [[TOBOOL]] to i32
1703; SINK-AFTER-NEXT:    [[A_1]] = add nsw i32 [[A_010]], [[INC]]
1704; SINK-AFTER-NEXT:    [[INC1]] = add nuw nsw i32 [[I_011]], 1
1705; SINK-AFTER-NEXT:    [[ADD_PTR]] = getelementptr inbounds double, ptr [[B_ADDR_012]], i64 25
1706; SINK-AFTER-NEXT:    [[EXITCOND:%.*]] = icmp eq i32 [[INC1]], 10240
1707; SINK-AFTER-NEXT:    br i1 [[EXITCOND]], label [[FOR_COND_CLEANUP]], label [[FOR_BODY]], !llvm.loop [[LOOP17:![0-9]+]]
1708;
1709entry:
1710  %idxprom = sext i32 %d to i64
1711  br label %for.body
1712
1713for.cond.cleanup:
1714  %a.1.lcssa = phi i32 [ %a.1, %for.body ]
1715  ret i32 %a.1.lcssa
1716
1717for.body:
1718  %b.addr.012 = phi ptr [ %b, %entry ], [ %add.ptr, %for.body ]
1719  %i.011 = phi i32 [ 0, %entry ], [ %inc1, %for.body ]
1720  %a.010 = phi i32 [ 0, %entry ], [ %a.1, %for.body ]
1721  %j.addr.09 = phi double [ %j, %entry ], [ %0, %for.body ]
1722  %arrayidx = getelementptr inbounds double, ptr %b.addr.012, i64 %idxprom
1723  %0 = load double, ptr %arrayidx, align 8
1724  %mul = fmul double %j.addr.09, %0
1725  %tobool = fcmp une double %mul, 0.000000e+00
1726  %inc = zext i1 %tobool to i32
1727  %a.1 = add nsw i32 %a.010, %inc
1728  %inc1 = add nuw nsw i32 %i.011, 1
1729  %add.ptr = getelementptr inbounds double, ptr %b.addr.012, i64 25
1730  %exitcond = icmp eq i32 %inc1, 10240
1731  br i1 %exitcond, label %for.cond.cleanup, label %for.body
1732}
1733
1734; void sink_after(short *a, int n, int *b) {
1735;   for(int i = 0; i < n; i++)
1736;     b[i] = (aptr a[i + 1]);
1737; }
1738;
1739; Check that the sext sank after the load in the vector loop.
1740;
1741define void @sink_after(ptr noalias %a, ptr noalias %b, i64 %n) {
1742; UNROLL-NO-IC-LABEL: @sink_after(
1743; UNROLL-NO-IC-NEXT:  entry:
1744; UNROLL-NO-IC-NEXT:    [[DOTPRE:%.*]] = load i16, ptr [[A:%.*]], align 2
1745; UNROLL-NO-IC-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[N:%.*]], 8
1746; UNROLL-NO-IC-NEXT:    br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
1747; UNROLL-NO-IC:       vector.ph:
1748; UNROLL-NO-IC-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[N]], 8
1749; UNROLL-NO-IC-NEXT:    [[N_VEC:%.*]] = sub i64 [[N]], [[N_MOD_VF]]
1750; UNROLL-NO-IC-NEXT:    [[VECTOR_RECUR_INIT:%.*]] = insertelement <4 x i16> poison, i16 [[DOTPRE]], i32 3
1751; UNROLL-NO-IC-NEXT:    br label [[VECTOR_BODY:%.*]]
1752; UNROLL-NO-IC:       vector.body:
1753; UNROLL-NO-IC-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
1754; UNROLL-NO-IC-NEXT:    [[VECTOR_RECUR:%.*]] = phi <4 x i16> [ [[VECTOR_RECUR_INIT]], [[VECTOR_PH]] ], [ [[WIDE_LOAD1:%.*]], [[VECTOR_BODY]] ]
1755; UNROLL-NO-IC-NEXT:    [[TMP0:%.*]] = add i64 [[INDEX]], 0
1756; UNROLL-NO-IC-NEXT:    [[TMP1:%.*]] = add nuw nsw i64 [[TMP0]], 1
1757; UNROLL-NO-IC-NEXT:    [[TMP2:%.*]] = getelementptr inbounds i16, ptr [[A]], i64 [[TMP1]]
1758; UNROLL-NO-IC-NEXT:    [[TMP3:%.*]] = getelementptr inbounds i16, ptr [[TMP2]], i32 0
1759; UNROLL-NO-IC-NEXT:    [[TMP4:%.*]] = getelementptr inbounds i16, ptr [[TMP2]], i32 4
1760; UNROLL-NO-IC-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i16>, ptr [[TMP3]], align 2
1761; UNROLL-NO-IC-NEXT:    [[WIDE_LOAD1]] = load <4 x i16>, ptr [[TMP4]], align 2
1762; UNROLL-NO-IC-NEXT:    [[TMP5:%.*]] = shufflevector <4 x i16> [[VECTOR_RECUR]], <4 x i16> [[WIDE_LOAD]], <4 x i32> <i32 3, i32 4, i32 5, i32 6>
1763; UNROLL-NO-IC-NEXT:    [[TMP6:%.*]] = shufflevector <4 x i16> [[WIDE_LOAD]], <4 x i16> [[WIDE_LOAD1]], <4 x i32> <i32 3, i32 4, i32 5, i32 6>
1764; UNROLL-NO-IC-NEXT:    [[TMP7:%.*]] = sext <4 x i16> [[TMP5]] to <4 x i32>
1765; UNROLL-NO-IC-NEXT:    [[TMP8:%.*]] = sext <4 x i16> [[TMP6]] to <4 x i32>
1766; UNROLL-NO-IC-NEXT:    [[TMP9:%.*]] = sext <4 x i16> [[WIDE_LOAD]] to <4 x i32>
1767; UNROLL-NO-IC-NEXT:    [[TMP10:%.*]] = sext <4 x i16> [[WIDE_LOAD1]] to <4 x i32>
1768; UNROLL-NO-IC-NEXT:    [[TMP11:%.*]] = mul nsw <4 x i32> [[TMP9]], [[TMP7]]
1769; UNROLL-NO-IC-NEXT:    [[TMP12:%.*]] = mul nsw <4 x i32> [[TMP10]], [[TMP8]]
1770; UNROLL-NO-IC-NEXT:    [[TMP13:%.*]] = getelementptr inbounds i32, ptr [[B:%.*]], i64 [[TMP0]]
1771; UNROLL-NO-IC-NEXT:    [[TMP14:%.*]] = getelementptr inbounds i32, ptr [[TMP13]], i32 0
1772; UNROLL-NO-IC-NEXT:    [[TMP15:%.*]] = getelementptr inbounds i32, ptr [[TMP13]], i32 4
1773; UNROLL-NO-IC-NEXT:    store <4 x i32> [[TMP11]], ptr [[TMP14]], align 4
1774; UNROLL-NO-IC-NEXT:    store <4 x i32> [[TMP12]], ptr [[TMP15]], align 4
1775; UNROLL-NO-IC-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 8
1776; UNROLL-NO-IC-NEXT:    [[TMP16:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
1777; UNROLL-NO-IC-NEXT:    br i1 [[TMP16]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP18:![0-9]+]]
1778; UNROLL-NO-IC:       middle.block:
1779; UNROLL-NO-IC-NEXT:    [[VECTOR_RECUR_EXTRACT:%.*]] = extractelement <4 x i16> [[WIDE_LOAD1]], i32 3
1780; UNROLL-NO-IC-NEXT:    [[CMP_N:%.*]] = icmp eq i64 [[N]], [[N_VEC]]
1781; UNROLL-NO-IC-NEXT:    br i1 [[CMP_N]], label [[FOR_END:%.*]], label [[SCALAR_PH]]
1782; UNROLL-NO-IC:       scalar.ph:
1783; UNROLL-NO-IC-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i16 [ [[VECTOR_RECUR_EXTRACT]], [[MIDDLE_BLOCK]] ], [ [[DOTPRE]], [[ENTRY:%.*]] ]
1784; UNROLL-NO-IC-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY]] ]
1785; UNROLL-NO-IC-NEXT:    br label [[FOR_BODY:%.*]]
1786; UNROLL-NO-IC:       for.body:
1787; UNROLL-NO-IC-NEXT:    [[TMP17:%.*]] = phi i16 [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ], [ [[TMP18:%.*]], [[FOR_BODY]] ]
1788; UNROLL-NO-IC-NEXT:    [[INDVARS_IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ]
1789; UNROLL-NO-IC-NEXT:    [[CONV:%.*]] = sext i16 [[TMP17]] to i32
1790; UNROLL-NO-IC-NEXT:    [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1
1791; UNROLL-NO-IC-NEXT:    [[ARRAYIDX2:%.*]] = getelementptr inbounds i16, ptr [[A]], i64 [[INDVARS_IV_NEXT]]
1792; UNROLL-NO-IC-NEXT:    [[TMP18]] = load i16, ptr [[ARRAYIDX2]], align 2
1793; UNROLL-NO-IC-NEXT:    [[CONV3:%.*]] = sext i16 [[TMP18]] to i32
1794; UNROLL-NO-IC-NEXT:    [[MUL:%.*]] = mul nsw i32 [[CONV3]], [[CONV]]
1795; UNROLL-NO-IC-NEXT:    [[ARRAYIDX5:%.*]] = getelementptr inbounds i32, ptr [[B]], i64 [[INDVARS_IV]]
1796; UNROLL-NO-IC-NEXT:    store i32 [[MUL]], ptr [[ARRAYIDX5]], align 4
1797; UNROLL-NO-IC-NEXT:    [[EXITCOND:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT]], [[N]]
1798; UNROLL-NO-IC-NEXT:    br i1 [[EXITCOND]], label [[FOR_END]], label [[FOR_BODY]], !llvm.loop [[LOOP19:![0-9]+]]
1799; UNROLL-NO-IC:       for.end:
1800; UNROLL-NO-IC-NEXT:    ret void
1801;
1802; UNROLL-NO-VF-LABEL: @sink_after(
1803; UNROLL-NO-VF-NEXT:  entry:
1804; UNROLL-NO-VF-NEXT:    [[DOTPRE:%.*]] = load i16, ptr [[A:%.*]], align 2
1805; UNROLL-NO-VF-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[N:%.*]], 2
1806; UNROLL-NO-VF-NEXT:    br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
1807; UNROLL-NO-VF:       vector.ph:
1808; UNROLL-NO-VF-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[N]], 2
1809; UNROLL-NO-VF-NEXT:    [[N_VEC:%.*]] = sub i64 [[N]], [[N_MOD_VF]]
1810; UNROLL-NO-VF-NEXT:    br label [[VECTOR_BODY:%.*]]
1811; UNROLL-NO-VF:       vector.body:
1812; UNROLL-NO-VF-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
1813; UNROLL-NO-VF-NEXT:    [[VECTOR_RECUR:%.*]] = phi i16 [ [[DOTPRE]], [[VECTOR_PH]] ], [ [[TMP7:%.*]], [[VECTOR_BODY]] ]
1814; UNROLL-NO-VF-NEXT:    [[TMP0:%.*]] = add i64 [[INDEX]], 0
1815; UNROLL-NO-VF-NEXT:    [[TMP1:%.*]] = add i64 [[INDEX]], 1
1816; UNROLL-NO-VF-NEXT:    [[TMP2:%.*]] = add nuw nsw i64 [[TMP0]], 1
1817; UNROLL-NO-VF-NEXT:    [[TMP3:%.*]] = add nuw nsw i64 [[TMP1]], 1
1818; UNROLL-NO-VF-NEXT:    [[TMP4:%.*]] = getelementptr inbounds i16, ptr [[A]], i64 [[TMP2]]
1819; UNROLL-NO-VF-NEXT:    [[TMP5:%.*]] = getelementptr inbounds i16, ptr [[A]], i64 [[TMP3]]
1820; UNROLL-NO-VF-NEXT:    [[TMP6:%.*]] = load i16, ptr [[TMP4]], align 2
1821; UNROLL-NO-VF-NEXT:    [[TMP7]] = load i16, ptr [[TMP5]], align 2
1822; UNROLL-NO-VF-NEXT:    [[TMP8:%.*]] = sext i16 [[VECTOR_RECUR]] to i32
1823; UNROLL-NO-VF-NEXT:    [[TMP9:%.*]] = sext i16 [[TMP6]] to i32
1824; UNROLL-NO-VF-NEXT:    [[TMP10:%.*]] = sext i16 [[TMP6]] to i32
1825; UNROLL-NO-VF-NEXT:    [[TMP11:%.*]] = sext i16 [[TMP7]] to i32
1826; UNROLL-NO-VF-NEXT:    [[TMP12:%.*]] = mul nsw i32 [[TMP10]], [[TMP8]]
1827; UNROLL-NO-VF-NEXT:    [[TMP13:%.*]] = mul nsw i32 [[TMP11]], [[TMP9]]
1828; UNROLL-NO-VF-NEXT:    [[TMP14:%.*]] = getelementptr inbounds i32, ptr [[B:%.*]], i64 [[TMP0]]
1829; UNROLL-NO-VF-NEXT:    [[TMP15:%.*]] = getelementptr inbounds i32, ptr [[B]], i64 [[TMP1]]
1830; UNROLL-NO-VF-NEXT:    store i32 [[TMP12]], ptr [[TMP14]], align 4
1831; UNROLL-NO-VF-NEXT:    store i32 [[TMP13]], ptr [[TMP15]], align 4
1832; UNROLL-NO-VF-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 2
1833; UNROLL-NO-VF-NEXT:    [[TMP16:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
1834; UNROLL-NO-VF-NEXT:    br i1 [[TMP16]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP18:![0-9]+]]
1835; UNROLL-NO-VF:       middle.block:
1836; UNROLL-NO-VF-NEXT:    [[CMP_N:%.*]] = icmp eq i64 [[N]], [[N_VEC]]
1837; UNROLL-NO-VF-NEXT:    br i1 [[CMP_N]], label [[FOR_END:%.*]], label [[SCALAR_PH]]
1838; UNROLL-NO-VF:       scalar.ph:
1839; UNROLL-NO-VF-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i16 [ [[TMP7]], [[MIDDLE_BLOCK]] ], [ [[DOTPRE]], [[ENTRY:%.*]] ]
1840; UNROLL-NO-VF-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY]] ]
1841; UNROLL-NO-VF-NEXT:    br label [[FOR_BODY:%.*]]
1842; UNROLL-NO-VF:       for.body:
1843; UNROLL-NO-VF-NEXT:    [[TMP17:%.*]] = phi i16 [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ], [ [[TMP18:%.*]], [[FOR_BODY]] ]
1844; UNROLL-NO-VF-NEXT:    [[INDVARS_IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ]
1845; UNROLL-NO-VF-NEXT:    [[CONV:%.*]] = sext i16 [[TMP17]] to i32
1846; UNROLL-NO-VF-NEXT:    [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1
1847; UNROLL-NO-VF-NEXT:    [[ARRAYIDX2:%.*]] = getelementptr inbounds i16, ptr [[A]], i64 [[INDVARS_IV_NEXT]]
1848; UNROLL-NO-VF-NEXT:    [[TMP18]] = load i16, ptr [[ARRAYIDX2]], align 2
1849; UNROLL-NO-VF-NEXT:    [[CONV3:%.*]] = sext i16 [[TMP18]] to i32
1850; UNROLL-NO-VF-NEXT:    [[MUL:%.*]] = mul nsw i32 [[CONV3]], [[CONV]]
1851; UNROLL-NO-VF-NEXT:    [[ARRAYIDX5:%.*]] = getelementptr inbounds i32, ptr [[B]], i64 [[INDVARS_IV]]
1852; UNROLL-NO-VF-NEXT:    store i32 [[MUL]], ptr [[ARRAYIDX5]], align 4
1853; UNROLL-NO-VF-NEXT:    [[EXITCOND:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT]], [[N]]
1854; UNROLL-NO-VF-NEXT:    br i1 [[EXITCOND]], label [[FOR_END]], label [[FOR_BODY]], !llvm.loop [[LOOP19:![0-9]+]]
1855; UNROLL-NO-VF:       for.end:
1856; UNROLL-NO-VF-NEXT:    ret void
1857;
1858; SINK-AFTER-LABEL: @sink_after(
1859; SINK-AFTER-NEXT:  entry:
1860; SINK-AFTER-NEXT:    [[DOTPRE:%.*]] = load i16, ptr [[A:%.*]], align 2
1861; SINK-AFTER-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[N:%.*]], 4
1862; SINK-AFTER-NEXT:    br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
1863; SINK-AFTER:       vector.ph:
1864; SINK-AFTER-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[N]], 4
1865; SINK-AFTER-NEXT:    [[N_VEC:%.*]] = sub i64 [[N]], [[N_MOD_VF]]
1866; SINK-AFTER-NEXT:    [[VECTOR_RECUR_INIT:%.*]] = insertelement <4 x i16> poison, i16 [[DOTPRE]], i32 3
1867; SINK-AFTER-NEXT:    br label [[VECTOR_BODY:%.*]]
1868; SINK-AFTER:       vector.body:
1869; SINK-AFTER-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
1870; SINK-AFTER-NEXT:    [[VECTOR_RECUR:%.*]] = phi <4 x i16> [ [[VECTOR_RECUR_INIT]], [[VECTOR_PH]] ], [ [[WIDE_LOAD:%.*]], [[VECTOR_BODY]] ]
1871; SINK-AFTER-NEXT:    [[TMP0:%.*]] = add i64 [[INDEX]], 0
1872; SINK-AFTER-NEXT:    [[TMP1:%.*]] = add nuw nsw i64 [[TMP0]], 1
1873; SINK-AFTER-NEXT:    [[TMP2:%.*]] = getelementptr inbounds i16, ptr [[A]], i64 [[TMP1]]
1874; SINK-AFTER-NEXT:    [[TMP3:%.*]] = getelementptr inbounds i16, ptr [[TMP2]], i32 0
1875; SINK-AFTER-NEXT:    [[WIDE_LOAD]] = load <4 x i16>, ptr [[TMP3]], align 2
1876; SINK-AFTER-NEXT:    [[TMP4:%.*]] = shufflevector <4 x i16> [[VECTOR_RECUR]], <4 x i16> [[WIDE_LOAD]], <4 x i32> <i32 3, i32 4, i32 5, i32 6>
1877; SINK-AFTER-NEXT:    [[TMP5:%.*]] = sext <4 x i16> [[TMP4]] to <4 x i32>
1878; SINK-AFTER-NEXT:    [[TMP6:%.*]] = sext <4 x i16> [[WIDE_LOAD]] to <4 x i32>
1879; SINK-AFTER-NEXT:    [[TMP7:%.*]] = mul nsw <4 x i32> [[TMP6]], [[TMP5]]
1880; SINK-AFTER-NEXT:    [[TMP8:%.*]] = getelementptr inbounds i32, ptr [[B:%.*]], i64 [[TMP0]]
1881; SINK-AFTER-NEXT:    [[TMP9:%.*]] = getelementptr inbounds i32, ptr [[TMP8]], i32 0
1882; SINK-AFTER-NEXT:    store <4 x i32> [[TMP7]], ptr [[TMP9]], align 4
1883; SINK-AFTER-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4
1884; SINK-AFTER-NEXT:    [[TMP10:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
1885; SINK-AFTER-NEXT:    br i1 [[TMP10]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP18:![0-9]+]]
1886; SINK-AFTER:       middle.block:
1887; SINK-AFTER-NEXT:    [[VECTOR_RECUR_EXTRACT:%.*]] = extractelement <4 x i16> [[WIDE_LOAD]], i32 3
1888; SINK-AFTER-NEXT:    [[CMP_N:%.*]] = icmp eq i64 [[N]], [[N_VEC]]
1889; SINK-AFTER-NEXT:    br i1 [[CMP_N]], label [[FOR_END:%.*]], label [[SCALAR_PH]]
1890; SINK-AFTER:       scalar.ph:
1891; SINK-AFTER-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i16 [ [[VECTOR_RECUR_EXTRACT]], [[MIDDLE_BLOCK]] ], [ [[DOTPRE]], [[ENTRY:%.*]] ]
1892; SINK-AFTER-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY]] ]
1893; SINK-AFTER-NEXT:    br label [[FOR_BODY:%.*]]
1894; SINK-AFTER:       for.body:
1895; SINK-AFTER-NEXT:    [[TMP11:%.*]] = phi i16 [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ], [ [[TMP12:%.*]], [[FOR_BODY]] ]
1896; SINK-AFTER-NEXT:    [[INDVARS_IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ]
1897; SINK-AFTER-NEXT:    [[CONV:%.*]] = sext i16 [[TMP11]] to i32
1898; SINK-AFTER-NEXT:    [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1
1899; SINK-AFTER-NEXT:    [[ARRAYIDX2:%.*]] = getelementptr inbounds i16, ptr [[A]], i64 [[INDVARS_IV_NEXT]]
1900; SINK-AFTER-NEXT:    [[TMP12]] = load i16, ptr [[ARRAYIDX2]], align 2
1901; SINK-AFTER-NEXT:    [[CONV3:%.*]] = sext i16 [[TMP12]] to i32
1902; SINK-AFTER-NEXT:    [[MUL:%.*]] = mul nsw i32 [[CONV3]], [[CONV]]
1903; SINK-AFTER-NEXT:    [[ARRAYIDX5:%.*]] = getelementptr inbounds i32, ptr [[B]], i64 [[INDVARS_IV]]
1904; SINK-AFTER-NEXT:    store i32 [[MUL]], ptr [[ARRAYIDX5]], align 4
1905; SINK-AFTER-NEXT:    [[EXITCOND:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT]], [[N]]
1906; SINK-AFTER-NEXT:    br i1 [[EXITCOND]], label [[FOR_END]], label [[FOR_BODY]], !llvm.loop [[LOOP19:![0-9]+]]
1907; SINK-AFTER:       for.end:
1908; SINK-AFTER-NEXT:    ret void
1909;
1910entry:
1911  %.pre = load i16, ptr %a
1912  br label %for.body
1913
1914for.body:
1915  %0 = phi i16 [ %.pre, %entry ], [ %1, %for.body ]
1916  %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
1917  %conv = sext i16 %0 to i32
1918  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
1919  %arrayidx2 = getelementptr inbounds i16, ptr %a, i64 %indvars.iv.next
1920  %1 = load i16, ptr %arrayidx2
1921  %conv3 = sext i16 %1 to i32
1922  %mul = mul nsw i32 %conv3, %conv
1923  %arrayidx5 = getelementptr inbounds i32, ptr %b, i64 %indvars.iv
1924  store i32 %mul, ptr %arrayidx5
1925  %exitcond = icmp eq i64 %indvars.iv.next, %n
1926  br i1 %exitcond, label %for.end, label %for.body
1927
1928for.end:
1929  ret void
1930}
1931
1932; PR34711: given three consecutive instructions such that the first will be
1933; widened, the second is a cast that will be widened and needs to sink after the
1934; third, and the third is a first-order-recurring load that will be replicated
1935; instead of widened. Although the cast and the first instruction will both be
1936; widened, and are originally adjacent to each other, make sure the replicated
1937; load ends up appearing between them.
1938;
1939; void PR34711(short[2] *a, int *b, int *c, int n) {
1940;   for(int i = 0; i < n; i++) {
1941;     c[i] = 7;
1942;     b[i] = (a[i][0] * a[i][1]);
1943;   }
1944; }
1945;
1946; Check that the sext sank after the load in the vector loop.
1947;
1948define void @PR34711(ptr noalias %a, ptr noalias %b, ptr noalias %c, i64 %n) {
1949; UNROLL-NO-IC-LABEL: @PR34711(
1950; UNROLL-NO-IC-NEXT:  entry:
1951; UNROLL-NO-IC-NEXT:    [[DOTPRE:%.*]] = load i16, ptr [[A:%.*]], align 2
1952; UNROLL-NO-IC-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[N:%.*]], 8
1953; UNROLL-NO-IC-NEXT:    br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
1954; UNROLL-NO-IC:       vector.ph:
1955; UNROLL-NO-IC-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[N]], 8
1956; UNROLL-NO-IC-NEXT:    [[N_VEC:%.*]] = sub i64 [[N]], [[N_MOD_VF]]
1957; UNROLL-NO-IC-NEXT:    [[VECTOR_RECUR_INIT:%.*]] = insertelement <4 x i16> poison, i16 [[DOTPRE]], i32 3
1958; UNROLL-NO-IC-NEXT:    br label [[VECTOR_BODY:%.*]]
1959; UNROLL-NO-IC:       vector.body:
1960; UNROLL-NO-IC-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
1961; UNROLL-NO-IC-NEXT:    [[VECTOR_RECUR:%.*]] = phi <4 x i16> [ [[VECTOR_RECUR_INIT]], [[VECTOR_PH]] ], [ [[TMP34:%.*]], [[VECTOR_BODY]] ]
1962; UNROLL-NO-IC-NEXT:    [[TMP0:%.*]] = add i64 [[INDEX]], 0
1963; UNROLL-NO-IC-NEXT:    [[TMP1:%.*]] = add i64 [[INDEX]], 1
1964; UNROLL-NO-IC-NEXT:    [[TMP2:%.*]] = add i64 [[INDEX]], 2
1965; UNROLL-NO-IC-NEXT:    [[TMP3:%.*]] = add i64 [[INDEX]], 3
1966; UNROLL-NO-IC-NEXT:    [[TMP4:%.*]] = add i64 [[INDEX]], 4
1967; UNROLL-NO-IC-NEXT:    [[TMP5:%.*]] = add i64 [[INDEX]], 5
1968; UNROLL-NO-IC-NEXT:    [[TMP6:%.*]] = add i64 [[INDEX]], 6
1969; UNROLL-NO-IC-NEXT:    [[TMP7:%.*]] = add i64 [[INDEX]], 7
1970; UNROLL-NO-IC-NEXT:    [[TMP8:%.*]] = getelementptr inbounds i32, ptr [[C:%.*]], i64 [[TMP0]]
1971; UNROLL-NO-IC-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [2 x i16], ptr [[A]], i64 [[TMP0]], i64 1
1972; UNROLL-NO-IC-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [2 x i16], ptr [[A]], i64 [[TMP1]], i64 1
1973; UNROLL-NO-IC-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [2 x i16], ptr [[A]], i64 [[TMP2]], i64 1
1974; UNROLL-NO-IC-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [2 x i16], ptr [[A]], i64 [[TMP3]], i64 1
1975; UNROLL-NO-IC-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [2 x i16], ptr [[A]], i64 [[TMP4]], i64 1
1976; UNROLL-NO-IC-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [2 x i16], ptr [[A]], i64 [[TMP5]], i64 1
1977; UNROLL-NO-IC-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [2 x i16], ptr [[A]], i64 [[TMP6]], i64 1
1978; UNROLL-NO-IC-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [2 x i16], ptr [[A]], i64 [[TMP7]], i64 1
1979; UNROLL-NO-IC-NEXT:    [[TMP17:%.*]] = getelementptr inbounds i32, ptr [[TMP8]], i32 0
1980; UNROLL-NO-IC-NEXT:    [[TMP18:%.*]] = getelementptr inbounds i32, ptr [[TMP8]], i32 4
1981; UNROLL-NO-IC-NEXT:    store <4 x i32> splat (i32 7), ptr [[TMP17]], align 4
1982; UNROLL-NO-IC-NEXT:    store <4 x i32> splat (i32 7), ptr [[TMP18]], align 4
1983; UNROLL-NO-IC-NEXT:    [[TMP19:%.*]] = load i16, ptr [[TMP9]], align 2
1984; UNROLL-NO-IC-NEXT:    [[TMP20:%.*]] = load i16, ptr [[TMP10]], align 2
1985; UNROLL-NO-IC-NEXT:    [[TMP21:%.*]] = load i16, ptr [[TMP11]], align 2
1986; UNROLL-NO-IC-NEXT:    [[TMP22:%.*]] = load i16, ptr [[TMP12]], align 2
1987; UNROLL-NO-IC-NEXT:    [[TMP23:%.*]] = insertelement <4 x i16> poison, i16 [[TMP19]], i32 0
1988; UNROLL-NO-IC-NEXT:    [[TMP24:%.*]] = insertelement <4 x i16> [[TMP23]], i16 [[TMP20]], i32 1
1989; UNROLL-NO-IC-NEXT:    [[TMP25:%.*]] = insertelement <4 x i16> [[TMP24]], i16 [[TMP21]], i32 2
1990; UNROLL-NO-IC-NEXT:    [[TMP26:%.*]] = insertelement <4 x i16> [[TMP25]], i16 [[TMP22]], i32 3
1991; UNROLL-NO-IC-NEXT:    [[TMP27:%.*]] = load i16, ptr [[TMP13]], align 2
1992; UNROLL-NO-IC-NEXT:    [[TMP28:%.*]] = load i16, ptr [[TMP14]], align 2
1993; UNROLL-NO-IC-NEXT:    [[TMP29:%.*]] = load i16, ptr [[TMP15]], align 2
1994; UNROLL-NO-IC-NEXT:    [[TMP30:%.*]] = load i16, ptr [[TMP16]], align 2
1995; UNROLL-NO-IC-NEXT:    [[TMP31:%.*]] = insertelement <4 x i16> poison, i16 [[TMP27]], i32 0
1996; UNROLL-NO-IC-NEXT:    [[TMP32:%.*]] = insertelement <4 x i16> [[TMP31]], i16 [[TMP28]], i32 1
1997; UNROLL-NO-IC-NEXT:    [[TMP33:%.*]] = insertelement <4 x i16> [[TMP32]], i16 [[TMP29]], i32 2
1998; UNROLL-NO-IC-NEXT:    [[TMP34]] = insertelement <4 x i16> [[TMP33]], i16 [[TMP30]], i32 3
1999; UNROLL-NO-IC-NEXT:    [[TMP35:%.*]] = shufflevector <4 x i16> [[VECTOR_RECUR]], <4 x i16> [[TMP26]], <4 x i32> <i32 3, i32 4, i32 5, i32 6>
2000; UNROLL-NO-IC-NEXT:    [[TMP36:%.*]] = shufflevector <4 x i16> [[TMP26]], <4 x i16> [[TMP34]], <4 x i32> <i32 3, i32 4, i32 5, i32 6>
2001; UNROLL-NO-IC-NEXT:    [[TMP37:%.*]] = sext <4 x i16> [[TMP35]] to <4 x i32>
2002; UNROLL-NO-IC-NEXT:    [[TMP38:%.*]] = sext <4 x i16> [[TMP36]] to <4 x i32>
2003; UNROLL-NO-IC-NEXT:    [[TMP39:%.*]] = sext <4 x i16> [[TMP26]] to <4 x i32>
2004; UNROLL-NO-IC-NEXT:    [[TMP40:%.*]] = sext <4 x i16> [[TMP34]] to <4 x i32>
2005; UNROLL-NO-IC-NEXT:    [[TMP41:%.*]] = mul nsw <4 x i32> [[TMP39]], [[TMP37]]
2006; UNROLL-NO-IC-NEXT:    [[TMP42:%.*]] = mul nsw <4 x i32> [[TMP40]], [[TMP38]]
2007; UNROLL-NO-IC-NEXT:    [[TMP43:%.*]] = getelementptr inbounds i32, ptr [[B:%.*]], i64 [[TMP0]]
2008; UNROLL-NO-IC-NEXT:    [[TMP44:%.*]] = getelementptr inbounds i32, ptr [[TMP43]], i32 0
2009; UNROLL-NO-IC-NEXT:    [[TMP45:%.*]] = getelementptr inbounds i32, ptr [[TMP43]], i32 4
2010; UNROLL-NO-IC-NEXT:    store <4 x i32> [[TMP41]], ptr [[TMP44]], align 4
2011; UNROLL-NO-IC-NEXT:    store <4 x i32> [[TMP42]], ptr [[TMP45]], align 4
2012; UNROLL-NO-IC-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 8
2013; UNROLL-NO-IC-NEXT:    [[TMP46:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
2014; UNROLL-NO-IC-NEXT:    br i1 [[TMP46]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP20:![0-9]+]]
2015; UNROLL-NO-IC:       middle.block:
2016; UNROLL-NO-IC-NEXT:    [[CMP_N:%.*]] = icmp eq i64 [[N]], [[N_VEC]]
2017; UNROLL-NO-IC-NEXT:    br i1 [[CMP_N]], label [[FOR_END:%.*]], label [[SCALAR_PH]]
2018; UNROLL-NO-IC:       scalar.ph:
2019; UNROLL-NO-IC-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i16 [ [[TMP30]], [[MIDDLE_BLOCK]] ], [ [[DOTPRE]], [[ENTRY:%.*]] ]
2020; UNROLL-NO-IC-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY]] ]
2021; UNROLL-NO-IC-NEXT:    br label [[FOR_BODY:%.*]]
2022; UNROLL-NO-IC:       for.body:
2023; UNROLL-NO-IC-NEXT:    [[TMP47:%.*]] = phi i16 [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ], [ [[TMP48:%.*]], [[FOR_BODY]] ]
2024; UNROLL-NO-IC-NEXT:    [[INDVARS_IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ]
2025; UNROLL-NO-IC-NEXT:    [[ARRAYCIDX:%.*]] = getelementptr inbounds i32, ptr [[C]], i64 [[INDVARS_IV]]
2026; UNROLL-NO-IC-NEXT:    [[CUR_INDEX:%.*]] = getelementptr inbounds [2 x i16], ptr [[A]], i64 [[INDVARS_IV]], i64 1
2027; UNROLL-NO-IC-NEXT:    store i32 7, ptr [[ARRAYCIDX]], align 4
2028; UNROLL-NO-IC-NEXT:    [[CONV:%.*]] = sext i16 [[TMP47]] to i32
2029; UNROLL-NO-IC-NEXT:    [[TMP48]] = load i16, ptr [[CUR_INDEX]], align 2
2030; UNROLL-NO-IC-NEXT:    [[CONV3:%.*]] = sext i16 [[TMP48]] to i32
2031; UNROLL-NO-IC-NEXT:    [[MUL:%.*]] = mul nsw i32 [[CONV3]], [[CONV]]
2032; UNROLL-NO-IC-NEXT:    [[ARRAYIDX5:%.*]] = getelementptr inbounds i32, ptr [[B]], i64 [[INDVARS_IV]]
2033; UNROLL-NO-IC-NEXT:    store i32 [[MUL]], ptr [[ARRAYIDX5]], align 4
2034; UNROLL-NO-IC-NEXT:    [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1
2035; UNROLL-NO-IC-NEXT:    [[EXITCOND:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT]], [[N]]
2036; UNROLL-NO-IC-NEXT:    br i1 [[EXITCOND]], label [[FOR_END]], label [[FOR_BODY]], !llvm.loop [[LOOP21:![0-9]+]]
2037; UNROLL-NO-IC:       for.end:
2038; UNROLL-NO-IC-NEXT:    ret void
2039;
2040; UNROLL-NO-VF-LABEL: @PR34711(
2041; UNROLL-NO-VF-NEXT:  entry:
2042; UNROLL-NO-VF-NEXT:    [[DOTPRE:%.*]] = load i16, ptr [[A:%.*]], align 2
2043; UNROLL-NO-VF-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[N:%.*]], 2
2044; UNROLL-NO-VF-NEXT:    br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
2045; UNROLL-NO-VF:       vector.ph:
2046; UNROLL-NO-VF-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[N]], 2
2047; UNROLL-NO-VF-NEXT:    [[N_VEC:%.*]] = sub i64 [[N]], [[N_MOD_VF]]
2048; UNROLL-NO-VF-NEXT:    br label [[VECTOR_BODY:%.*]]
2049; UNROLL-NO-VF:       vector.body:
2050; UNROLL-NO-VF-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
2051; UNROLL-NO-VF-NEXT:    [[VECTOR_RECUR:%.*]] = phi i16 [ [[DOTPRE]], [[VECTOR_PH]] ], [ [[TMP7:%.*]], [[VECTOR_BODY]] ]
2052; UNROLL-NO-VF-NEXT:    [[TMP0:%.*]] = add i64 [[INDEX]], 0
2053; UNROLL-NO-VF-NEXT:    [[TMP1:%.*]] = add i64 [[INDEX]], 1
2054; UNROLL-NO-VF-NEXT:    [[TMP2:%.*]] = getelementptr inbounds i32, ptr [[C:%.*]], i64 [[TMP0]]
2055; UNROLL-NO-VF-NEXT:    [[TMP3:%.*]] = getelementptr inbounds i32, ptr [[C]], i64 [[TMP1]]
2056; UNROLL-NO-VF-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [2 x i16], ptr [[A]], i64 [[TMP0]], i64 1
2057; UNROLL-NO-VF-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [2 x i16], ptr [[A]], i64 [[TMP1]], i64 1
2058; UNROLL-NO-VF-NEXT:    store i32 7, ptr [[TMP2]], align 4
2059; UNROLL-NO-VF-NEXT:    store i32 7, ptr [[TMP3]], align 4
2060; UNROLL-NO-VF-NEXT:    [[TMP6:%.*]] = load i16, ptr [[TMP4]], align 2
2061; UNROLL-NO-VF-NEXT:    [[TMP7]] = load i16, ptr [[TMP5]], align 2
2062; UNROLL-NO-VF-NEXT:    [[TMP8:%.*]] = sext i16 [[VECTOR_RECUR]] to i32
2063; UNROLL-NO-VF-NEXT:    [[TMP9:%.*]] = sext i16 [[TMP6]] to i32
2064; UNROLL-NO-VF-NEXT:    [[TMP10:%.*]] = sext i16 [[TMP6]] to i32
2065; UNROLL-NO-VF-NEXT:    [[TMP11:%.*]] = sext i16 [[TMP7]] to i32
2066; UNROLL-NO-VF-NEXT:    [[TMP12:%.*]] = mul nsw i32 [[TMP10]], [[TMP8]]
2067; UNROLL-NO-VF-NEXT:    [[TMP13:%.*]] = mul nsw i32 [[TMP11]], [[TMP9]]
2068; UNROLL-NO-VF-NEXT:    [[TMP14:%.*]] = getelementptr inbounds i32, ptr [[B:%.*]], i64 [[TMP0]]
2069; UNROLL-NO-VF-NEXT:    [[TMP15:%.*]] = getelementptr inbounds i32, ptr [[B]], i64 [[TMP1]]
2070; UNROLL-NO-VF-NEXT:    store i32 [[TMP12]], ptr [[TMP14]], align 4
2071; UNROLL-NO-VF-NEXT:    store i32 [[TMP13]], ptr [[TMP15]], align 4
2072; UNROLL-NO-VF-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 2
2073; UNROLL-NO-VF-NEXT:    [[TMP16:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
2074; UNROLL-NO-VF-NEXT:    br i1 [[TMP16]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP20:![0-9]+]]
2075; UNROLL-NO-VF:       middle.block:
2076; UNROLL-NO-VF-NEXT:    [[CMP_N:%.*]] = icmp eq i64 [[N]], [[N_VEC]]
2077; UNROLL-NO-VF-NEXT:    br i1 [[CMP_N]], label [[FOR_END:%.*]], label [[SCALAR_PH]]
2078; UNROLL-NO-VF:       scalar.ph:
2079; UNROLL-NO-VF-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i16 [ [[TMP7]], [[MIDDLE_BLOCK]] ], [ [[DOTPRE]], [[ENTRY:%.*]] ]
2080; UNROLL-NO-VF-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY]] ]
2081; UNROLL-NO-VF-NEXT:    br label [[FOR_BODY:%.*]]
2082; UNROLL-NO-VF:       for.body:
2083; UNROLL-NO-VF-NEXT:    [[TMP17:%.*]] = phi i16 [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ], [ [[TMP18:%.*]], [[FOR_BODY]] ]
2084; UNROLL-NO-VF-NEXT:    [[INDVARS_IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ]
2085; UNROLL-NO-VF-NEXT:    [[ARRAYCIDX:%.*]] = getelementptr inbounds i32, ptr [[C]], i64 [[INDVARS_IV]]
2086; UNROLL-NO-VF-NEXT:    [[CUR_INDEX:%.*]] = getelementptr inbounds [2 x i16], ptr [[A]], i64 [[INDVARS_IV]], i64 1
2087; UNROLL-NO-VF-NEXT:    store i32 7, ptr [[ARRAYCIDX]], align 4
2088; UNROLL-NO-VF-NEXT:    [[CONV:%.*]] = sext i16 [[TMP17]] to i32
2089; UNROLL-NO-VF-NEXT:    [[TMP18]] = load i16, ptr [[CUR_INDEX]], align 2
2090; UNROLL-NO-VF-NEXT:    [[CONV3:%.*]] = sext i16 [[TMP18]] to i32
2091; UNROLL-NO-VF-NEXT:    [[MUL:%.*]] = mul nsw i32 [[CONV3]], [[CONV]]
2092; UNROLL-NO-VF-NEXT:    [[ARRAYIDX5:%.*]] = getelementptr inbounds i32, ptr [[B]], i64 [[INDVARS_IV]]
2093; UNROLL-NO-VF-NEXT:    store i32 [[MUL]], ptr [[ARRAYIDX5]], align 4
2094; UNROLL-NO-VF-NEXT:    [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1
2095; UNROLL-NO-VF-NEXT:    [[EXITCOND:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT]], [[N]]
2096; UNROLL-NO-VF-NEXT:    br i1 [[EXITCOND]], label [[FOR_END]], label [[FOR_BODY]], !llvm.loop [[LOOP21:![0-9]+]]
2097; UNROLL-NO-VF:       for.end:
2098; UNROLL-NO-VF-NEXT:    ret void
2099;
2100; SINK-AFTER-LABEL: @PR34711(
2101; SINK-AFTER-NEXT:  entry:
2102; SINK-AFTER-NEXT:    [[DOTPRE:%.*]] = load i16, ptr [[A:%.*]], align 2
2103; SINK-AFTER-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[N:%.*]], 4
2104; SINK-AFTER-NEXT:    br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
2105; SINK-AFTER:       vector.ph:
2106; SINK-AFTER-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[N]], 4
2107; SINK-AFTER-NEXT:    [[N_VEC:%.*]] = sub i64 [[N]], [[N_MOD_VF]]
2108; SINK-AFTER-NEXT:    [[VECTOR_RECUR_INIT:%.*]] = insertelement <4 x i16> poison, i16 [[DOTPRE]], i32 3
2109; SINK-AFTER-NEXT:    br label [[VECTOR_BODY:%.*]]
2110; SINK-AFTER:       vector.body:
2111; SINK-AFTER-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
2112; SINK-AFTER-NEXT:    [[VECTOR_RECUR:%.*]] = phi <4 x i16> [ [[VECTOR_RECUR_INIT]], [[VECTOR_PH]] ], [ [[TMP17:%.*]], [[VECTOR_BODY]] ]
2113; SINK-AFTER-NEXT:    [[TMP0:%.*]] = add i64 [[INDEX]], 0
2114; SINK-AFTER-NEXT:    [[TMP1:%.*]] = add i64 [[INDEX]], 1
2115; SINK-AFTER-NEXT:    [[TMP2:%.*]] = add i64 [[INDEX]], 2
2116; SINK-AFTER-NEXT:    [[TMP3:%.*]] = add i64 [[INDEX]], 3
2117; SINK-AFTER-NEXT:    [[TMP4:%.*]] = getelementptr inbounds i32, ptr [[C:%.*]], i64 [[TMP0]]
2118; SINK-AFTER-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [2 x i16], ptr [[A]], i64 [[TMP0]], i64 1
2119; SINK-AFTER-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [2 x i16], ptr [[A]], i64 [[TMP1]], i64 1
2120; SINK-AFTER-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [2 x i16], ptr [[A]], i64 [[TMP2]], i64 1
2121; SINK-AFTER-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [2 x i16], ptr [[A]], i64 [[TMP3]], i64 1
2122; SINK-AFTER-NEXT:    [[TMP9:%.*]] = getelementptr inbounds i32, ptr [[TMP4]], i32 0
2123; SINK-AFTER-NEXT:    store <4 x i32> splat (i32 7), ptr [[TMP9]], align 4
2124; SINK-AFTER-NEXT:    [[TMP10:%.*]] = load i16, ptr [[TMP5]], align 2
2125; SINK-AFTER-NEXT:    [[TMP11:%.*]] = load i16, ptr [[TMP6]], align 2
2126; SINK-AFTER-NEXT:    [[TMP12:%.*]] = load i16, ptr [[TMP7]], align 2
2127; SINK-AFTER-NEXT:    [[TMP13:%.*]] = load i16, ptr [[TMP8]], align 2
2128; SINK-AFTER-NEXT:    [[TMP14:%.*]] = insertelement <4 x i16> poison, i16 [[TMP10]], i32 0
2129; SINK-AFTER-NEXT:    [[TMP15:%.*]] = insertelement <4 x i16> [[TMP14]], i16 [[TMP11]], i32 1
2130; SINK-AFTER-NEXT:    [[TMP16:%.*]] = insertelement <4 x i16> [[TMP15]], i16 [[TMP12]], i32 2
2131; SINK-AFTER-NEXT:    [[TMP17]] = insertelement <4 x i16> [[TMP16]], i16 [[TMP13]], i32 3
2132; SINK-AFTER-NEXT:    [[TMP18:%.*]] = shufflevector <4 x i16> [[VECTOR_RECUR]], <4 x i16> [[TMP17]], <4 x i32> <i32 3, i32 4, i32 5, i32 6>
2133; SINK-AFTER-NEXT:    [[TMP19:%.*]] = sext <4 x i16> [[TMP18]] to <4 x i32>
2134; SINK-AFTER-NEXT:    [[TMP20:%.*]] = sext <4 x i16> [[TMP17]] to <4 x i32>
2135; SINK-AFTER-NEXT:    [[TMP21:%.*]] = mul nsw <4 x i32> [[TMP20]], [[TMP19]]
2136; SINK-AFTER-NEXT:    [[TMP22:%.*]] = getelementptr inbounds i32, ptr [[B:%.*]], i64 [[TMP0]]
2137; SINK-AFTER-NEXT:    [[TMP23:%.*]] = getelementptr inbounds i32, ptr [[TMP22]], i32 0
2138; SINK-AFTER-NEXT:    store <4 x i32> [[TMP21]], ptr [[TMP23]], align 4
2139; SINK-AFTER-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4
2140; SINK-AFTER-NEXT:    [[TMP24:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
2141; SINK-AFTER-NEXT:    br i1 [[TMP24]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP20:![0-9]+]]
2142; SINK-AFTER:       middle.block:
2143; SINK-AFTER-NEXT:    [[CMP_N:%.*]] = icmp eq i64 [[N]], [[N_VEC]]
2144; SINK-AFTER-NEXT:    br i1 [[CMP_N]], label [[FOR_END:%.*]], label [[SCALAR_PH]]
2145; SINK-AFTER:       scalar.ph:
2146; SINK-AFTER-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i16 [ [[TMP13]], [[MIDDLE_BLOCK]] ], [ [[DOTPRE]], [[ENTRY:%.*]] ]
2147; SINK-AFTER-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY]] ]
2148; SINK-AFTER-NEXT:    br label [[FOR_BODY:%.*]]
2149; SINK-AFTER:       for.body:
2150; SINK-AFTER-NEXT:    [[TMP25:%.*]] = phi i16 [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ], [ [[TMP26:%.*]], [[FOR_BODY]] ]
2151; SINK-AFTER-NEXT:    [[INDVARS_IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ]
2152; SINK-AFTER-NEXT:    [[ARRAYCIDX:%.*]] = getelementptr inbounds i32, ptr [[C]], i64 [[INDVARS_IV]]
2153; SINK-AFTER-NEXT:    [[CUR_INDEX:%.*]] = getelementptr inbounds [2 x i16], ptr [[A]], i64 [[INDVARS_IV]], i64 1
2154; SINK-AFTER-NEXT:    store i32 7, ptr [[ARRAYCIDX]], align 4
2155; SINK-AFTER-NEXT:    [[CONV:%.*]] = sext i16 [[TMP25]] to i32
2156; SINK-AFTER-NEXT:    [[TMP26]] = load i16, ptr [[CUR_INDEX]], align 2
2157; SINK-AFTER-NEXT:    [[CONV3:%.*]] = sext i16 [[TMP26]] to i32
2158; SINK-AFTER-NEXT:    [[MUL:%.*]] = mul nsw i32 [[CONV3]], [[CONV]]
2159; SINK-AFTER-NEXT:    [[ARRAYIDX5:%.*]] = getelementptr inbounds i32, ptr [[B]], i64 [[INDVARS_IV]]
2160; SINK-AFTER-NEXT:    store i32 [[MUL]], ptr [[ARRAYIDX5]], align 4
2161; SINK-AFTER-NEXT:    [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1
2162; SINK-AFTER-NEXT:    [[EXITCOND:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT]], [[N]]
2163; SINK-AFTER-NEXT:    br i1 [[EXITCOND]], label [[FOR_END]], label [[FOR_BODY]], !llvm.loop [[LOOP21:![0-9]+]]
2164; SINK-AFTER:       for.end:
2165; SINK-AFTER-NEXT:    ret void
2166;
2167
2168entry:
2169  %.pre = load i16, ptr %a
2170  br label %for.body
2171
2172for.body:
2173  %0 = phi i16 [ %.pre, %entry ], [ %1, %for.body ]
2174  %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
2175  %arraycidx = getelementptr inbounds i32, ptr %c, i64 %indvars.iv
2176  %cur.index = getelementptr inbounds [2 x i16], ptr %a, i64 %indvars.iv, i64 1
2177  store i32 7, ptr %arraycidx   ; 1st instruction, to be widened.
2178  %conv = sext i16 %0 to i32     ; 2nd, cast to sink after third.
2179  %1 = load i16, ptr %cur.index ; 3rd, first-order-recurring load not widened.
2180  %conv3 = sext i16 %1 to i32
2181  %mul = mul nsw i32 %conv3, %conv
2182  %arrayidx5 = getelementptr inbounds i32, ptr %b, i64 %indvars.iv
2183  store i32 %mul, ptr %arrayidx5
2184  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
2185  %exitcond = icmp eq i64 %indvars.iv.next, %n
2186  br i1 %exitcond, label %for.end, label %for.body
2187
2188for.end:
2189  ret void
2190}
2191
2192; void no_sink_after(short *a, int n, int *b) {
2193;   for(int i = 0; i < n; i++)
2194;     b[i] = ((a[i] + 2) * a[i + 1]);
2195; }
2196;
2197
2198;
2199define void @sink_after_with_multiple_users(ptr noalias %a, ptr noalias %b, i64 %n) {
2200; UNROLL-NO-IC-LABEL: @sink_after_with_multiple_users(
2201; UNROLL-NO-IC-NEXT:  entry:
2202; UNROLL-NO-IC-NEXT:    [[DOTPRE:%.*]] = load i16, ptr [[A:%.*]], align 2
2203; UNROLL-NO-IC-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[N:%.*]], 8
2204; UNROLL-NO-IC-NEXT:    br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
2205; UNROLL-NO-IC:       vector.ph:
2206; UNROLL-NO-IC-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[N]], 8
2207; UNROLL-NO-IC-NEXT:    [[N_VEC:%.*]] = sub i64 [[N]], [[N_MOD_VF]]
2208; UNROLL-NO-IC-NEXT:    [[VECTOR_RECUR_INIT:%.*]] = insertelement <4 x i16> poison, i16 [[DOTPRE]], i32 3
2209; UNROLL-NO-IC-NEXT:    br label [[VECTOR_BODY:%.*]]
2210; UNROLL-NO-IC:       vector.body:
2211; UNROLL-NO-IC-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
2212; UNROLL-NO-IC-NEXT:    [[VECTOR_RECUR:%.*]] = phi <4 x i16> [ [[VECTOR_RECUR_INIT]], [[VECTOR_PH]] ], [ [[WIDE_LOAD1:%.*]], [[VECTOR_BODY]] ]
2213; UNROLL-NO-IC-NEXT:    [[TMP0:%.*]] = add i64 [[INDEX]], 0
2214; UNROLL-NO-IC-NEXT:    [[TMP1:%.*]] = add nuw nsw i64 [[TMP0]], 1
2215; UNROLL-NO-IC-NEXT:    [[TMP2:%.*]] = getelementptr inbounds i16, ptr [[A]], i64 [[TMP1]]
2216; UNROLL-NO-IC-NEXT:    [[TMP3:%.*]] = getelementptr inbounds i16, ptr [[TMP2]], i32 0
2217; UNROLL-NO-IC-NEXT:    [[TMP4:%.*]] = getelementptr inbounds i16, ptr [[TMP2]], i32 4
2218; UNROLL-NO-IC-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i16>, ptr [[TMP3]], align 2
2219; UNROLL-NO-IC-NEXT:    [[WIDE_LOAD1]] = load <4 x i16>, ptr [[TMP4]], align 2
2220; UNROLL-NO-IC-NEXT:    [[TMP5:%.*]] = shufflevector <4 x i16> [[VECTOR_RECUR]], <4 x i16> [[WIDE_LOAD]], <4 x i32> <i32 3, i32 4, i32 5, i32 6>
2221; UNROLL-NO-IC-NEXT:    [[TMP6:%.*]] = shufflevector <4 x i16> [[WIDE_LOAD]], <4 x i16> [[WIDE_LOAD1]], <4 x i32> <i32 3, i32 4, i32 5, i32 6>
2222; UNROLL-NO-IC-NEXT:    [[TMP7:%.*]] = sext <4 x i16> [[TMP5]] to <4 x i32>
2223; UNROLL-NO-IC-NEXT:    [[TMP8:%.*]] = sext <4 x i16> [[TMP6]] to <4 x i32>
2224; UNROLL-NO-IC-NEXT:    [[TMP9:%.*]] = add nsw <4 x i32> [[TMP7]], splat (i32 2)
2225; UNROLL-NO-IC-NEXT:    [[TMP10:%.*]] = add nsw <4 x i32> [[TMP8]], splat (i32 2)
2226; UNROLL-NO-IC-NEXT:    [[TMP11:%.*]] = sext <4 x i16> [[WIDE_LOAD]] to <4 x i32>
2227; UNROLL-NO-IC-NEXT:    [[TMP12:%.*]] = sext <4 x i16> [[WIDE_LOAD1]] to <4 x i32>
2228; UNROLL-NO-IC-NEXT:    [[TMP13:%.*]] = mul nsw <4 x i32> [[TMP9]], [[TMP11]]
2229; UNROLL-NO-IC-NEXT:    [[TMP14:%.*]] = mul nsw <4 x i32> [[TMP10]], [[TMP12]]
2230; UNROLL-NO-IC-NEXT:    [[TMP15:%.*]] = getelementptr inbounds i32, ptr [[B:%.*]], i64 [[TMP0]]
2231; UNROLL-NO-IC-NEXT:    [[TMP16:%.*]] = getelementptr inbounds i32, ptr [[TMP15]], i32 0
2232; UNROLL-NO-IC-NEXT:    [[TMP17:%.*]] = getelementptr inbounds i32, ptr [[TMP15]], i32 4
2233; UNROLL-NO-IC-NEXT:    store <4 x i32> [[TMP13]], ptr [[TMP16]], align 4
2234; UNROLL-NO-IC-NEXT:    store <4 x i32> [[TMP14]], ptr [[TMP17]], align 4
2235; UNROLL-NO-IC-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 8
2236; UNROLL-NO-IC-NEXT:    [[TMP18:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
2237; UNROLL-NO-IC-NEXT:    br i1 [[TMP18]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP22:![0-9]+]]
2238; UNROLL-NO-IC:       middle.block:
2239; UNROLL-NO-IC-NEXT:    [[VECTOR_RECUR_EXTRACT:%.*]] = extractelement <4 x i16> [[WIDE_LOAD1]], i32 3
2240; UNROLL-NO-IC-NEXT:    [[CMP_N:%.*]] = icmp eq i64 [[N]], [[N_VEC]]
2241; UNROLL-NO-IC-NEXT:    br i1 [[CMP_N]], label [[FOR_END:%.*]], label [[SCALAR_PH]]
2242; UNROLL-NO-IC:       scalar.ph:
2243; UNROLL-NO-IC-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i16 [ [[VECTOR_RECUR_EXTRACT]], [[MIDDLE_BLOCK]] ], [ [[DOTPRE]], [[ENTRY:%.*]] ]
2244; UNROLL-NO-IC-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY]] ]
2245; UNROLL-NO-IC-NEXT:    br label [[FOR_BODY:%.*]]
2246; UNROLL-NO-IC:       for.body:
2247; UNROLL-NO-IC-NEXT:    [[TMP19:%.*]] = phi i16 [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ], [ [[TMP20:%.*]], [[FOR_BODY]] ]
2248; UNROLL-NO-IC-NEXT:    [[INDVARS_IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ]
2249; UNROLL-NO-IC-NEXT:    [[CONV:%.*]] = sext i16 [[TMP19]] to i32
2250; UNROLL-NO-IC-NEXT:    [[ADD:%.*]] = add nsw i32 [[CONV]], 2
2251; UNROLL-NO-IC-NEXT:    [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1
2252; UNROLL-NO-IC-NEXT:    [[ARRAYIDX2:%.*]] = getelementptr inbounds i16, ptr [[A]], i64 [[INDVARS_IV_NEXT]]
2253; UNROLL-NO-IC-NEXT:    [[TMP20]] = load i16, ptr [[ARRAYIDX2]], align 2
2254; UNROLL-NO-IC-NEXT:    [[CONV3:%.*]] = sext i16 [[TMP20]] to i32
2255; UNROLL-NO-IC-NEXT:    [[MUL:%.*]] = mul nsw i32 [[ADD]], [[CONV3]]
2256; UNROLL-NO-IC-NEXT:    [[ARRAYIDX5:%.*]] = getelementptr inbounds i32, ptr [[B]], i64 [[INDVARS_IV]]
2257; UNROLL-NO-IC-NEXT:    store i32 [[MUL]], ptr [[ARRAYIDX5]], align 4
2258; UNROLL-NO-IC-NEXT:    [[EXITCOND:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT]], [[N]]
2259; UNROLL-NO-IC-NEXT:    br i1 [[EXITCOND]], label [[FOR_END]], label [[FOR_BODY]], !llvm.loop [[LOOP23:![0-9]+]]
2260; UNROLL-NO-IC:       for.end:
2261; UNROLL-NO-IC-NEXT:    ret void
2262;
2263; UNROLL-NO-VF-LABEL: @sink_after_with_multiple_users(
2264; UNROLL-NO-VF-NEXT:  entry:
2265; UNROLL-NO-VF-NEXT:    [[DOTPRE:%.*]] = load i16, ptr [[A:%.*]], align 2
2266; UNROLL-NO-VF-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[N:%.*]], 2
2267; UNROLL-NO-VF-NEXT:    br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
2268; UNROLL-NO-VF:       vector.ph:
2269; UNROLL-NO-VF-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[N]], 2
2270; UNROLL-NO-VF-NEXT:    [[N_VEC:%.*]] = sub i64 [[N]], [[N_MOD_VF]]
2271; UNROLL-NO-VF-NEXT:    br label [[VECTOR_BODY:%.*]]
2272; UNROLL-NO-VF:       vector.body:
2273; UNROLL-NO-VF-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
2274; UNROLL-NO-VF-NEXT:    [[VECTOR_RECUR:%.*]] = phi i16 [ [[DOTPRE]], [[VECTOR_PH]] ], [ [[TMP7:%.*]], [[VECTOR_BODY]] ]
2275; UNROLL-NO-VF-NEXT:    [[TMP0:%.*]] = add i64 [[INDEX]], 0
2276; UNROLL-NO-VF-NEXT:    [[TMP1:%.*]] = add i64 [[INDEX]], 1
2277; UNROLL-NO-VF-NEXT:    [[TMP2:%.*]] = add nuw nsw i64 [[TMP0]], 1
2278; UNROLL-NO-VF-NEXT:    [[TMP3:%.*]] = add nuw nsw i64 [[TMP1]], 1
2279; UNROLL-NO-VF-NEXT:    [[TMP4:%.*]] = getelementptr inbounds i16, ptr [[A]], i64 [[TMP2]]
2280; UNROLL-NO-VF-NEXT:    [[TMP5:%.*]] = getelementptr inbounds i16, ptr [[A]], i64 [[TMP3]]
2281; UNROLL-NO-VF-NEXT:    [[TMP6:%.*]] = load i16, ptr [[TMP4]], align 2
2282; UNROLL-NO-VF-NEXT:    [[TMP7]] = load i16, ptr [[TMP5]], align 2
2283; UNROLL-NO-VF-NEXT:    [[TMP8:%.*]] = sext i16 [[VECTOR_RECUR]] to i32
2284; UNROLL-NO-VF-NEXT:    [[TMP9:%.*]] = sext i16 [[TMP6]] to i32
2285; UNROLL-NO-VF-NEXT:    [[TMP10:%.*]] = add nsw i32 [[TMP8]], 2
2286; UNROLL-NO-VF-NEXT:    [[TMP11:%.*]] = add nsw i32 [[TMP9]], 2
2287; UNROLL-NO-VF-NEXT:    [[TMP12:%.*]] = sext i16 [[TMP6]] to i32
2288; UNROLL-NO-VF-NEXT:    [[TMP13:%.*]] = sext i16 [[TMP7]] to i32
2289; UNROLL-NO-VF-NEXT:    [[TMP14:%.*]] = mul nsw i32 [[TMP10]], [[TMP12]]
2290; UNROLL-NO-VF-NEXT:    [[TMP15:%.*]] = mul nsw i32 [[TMP11]], [[TMP13]]
2291; UNROLL-NO-VF-NEXT:    [[TMP16:%.*]] = getelementptr inbounds i32, ptr [[B:%.*]], i64 [[TMP0]]
2292; UNROLL-NO-VF-NEXT:    [[TMP17:%.*]] = getelementptr inbounds i32, ptr [[B]], i64 [[TMP1]]
2293; UNROLL-NO-VF-NEXT:    store i32 [[TMP14]], ptr [[TMP16]], align 4
2294; UNROLL-NO-VF-NEXT:    store i32 [[TMP15]], ptr [[TMP17]], align 4
2295; UNROLL-NO-VF-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 2
2296; UNROLL-NO-VF-NEXT:    [[TMP18:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
2297; UNROLL-NO-VF-NEXT:    br i1 [[TMP18]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP22:![0-9]+]]
2298; UNROLL-NO-VF:       middle.block:
2299; UNROLL-NO-VF-NEXT:    [[CMP_N:%.*]] = icmp eq i64 [[N]], [[N_VEC]]
2300; UNROLL-NO-VF-NEXT:    br i1 [[CMP_N]], label [[FOR_END:%.*]], label [[SCALAR_PH]]
2301; UNROLL-NO-VF:       scalar.ph:
2302; UNROLL-NO-VF-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i16 [ [[TMP7]], [[MIDDLE_BLOCK]] ], [ [[DOTPRE]], [[ENTRY:%.*]] ]
2303; UNROLL-NO-VF-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY]] ]
2304; UNROLL-NO-VF-NEXT:    br label [[FOR_BODY:%.*]]
2305; UNROLL-NO-VF:       for.body:
2306; UNROLL-NO-VF-NEXT:    [[TMP19:%.*]] = phi i16 [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ], [ [[TMP20:%.*]], [[FOR_BODY]] ]
2307; UNROLL-NO-VF-NEXT:    [[INDVARS_IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ]
2308; UNROLL-NO-VF-NEXT:    [[CONV:%.*]] = sext i16 [[TMP19]] to i32
2309; UNROLL-NO-VF-NEXT:    [[ADD:%.*]] = add nsw i32 [[CONV]], 2
2310; UNROLL-NO-VF-NEXT:    [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1
2311; UNROLL-NO-VF-NEXT:    [[ARRAYIDX2:%.*]] = getelementptr inbounds i16, ptr [[A]], i64 [[INDVARS_IV_NEXT]]
2312; UNROLL-NO-VF-NEXT:    [[TMP20]] = load i16, ptr [[ARRAYIDX2]], align 2
2313; UNROLL-NO-VF-NEXT:    [[CONV3:%.*]] = sext i16 [[TMP20]] to i32
2314; UNROLL-NO-VF-NEXT:    [[MUL:%.*]] = mul nsw i32 [[ADD]], [[CONV3]]
2315; UNROLL-NO-VF-NEXT:    [[ARRAYIDX5:%.*]] = getelementptr inbounds i32, ptr [[B]], i64 [[INDVARS_IV]]
2316; UNROLL-NO-VF-NEXT:    store i32 [[MUL]], ptr [[ARRAYIDX5]], align 4
2317; UNROLL-NO-VF-NEXT:    [[EXITCOND:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT]], [[N]]
2318; UNROLL-NO-VF-NEXT:    br i1 [[EXITCOND]], label [[FOR_END]], label [[FOR_BODY]], !llvm.loop [[LOOP23:![0-9]+]]
2319; UNROLL-NO-VF:       for.end:
2320; UNROLL-NO-VF-NEXT:    ret void
2321;
2322; SINK-AFTER-LABEL: @sink_after_with_multiple_users(
2323; SINK-AFTER-NEXT:  entry:
2324; SINK-AFTER-NEXT:    [[DOTPRE:%.*]] = load i16, ptr [[A:%.*]], align 2
2325; SINK-AFTER-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[N:%.*]], 4
2326; SINK-AFTER-NEXT:    br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
2327; SINK-AFTER:       vector.ph:
2328; SINK-AFTER-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[N]], 4
2329; SINK-AFTER-NEXT:    [[N_VEC:%.*]] = sub i64 [[N]], [[N_MOD_VF]]
2330; SINK-AFTER-NEXT:    [[VECTOR_RECUR_INIT:%.*]] = insertelement <4 x i16> poison, i16 [[DOTPRE]], i32 3
2331; SINK-AFTER-NEXT:    br label [[VECTOR_BODY:%.*]]
2332; SINK-AFTER:       vector.body:
2333; SINK-AFTER-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
2334; SINK-AFTER-NEXT:    [[VECTOR_RECUR:%.*]] = phi <4 x i16> [ [[VECTOR_RECUR_INIT]], [[VECTOR_PH]] ], [ [[WIDE_LOAD:%.*]], [[VECTOR_BODY]] ]
2335; SINK-AFTER-NEXT:    [[TMP0:%.*]] = add i64 [[INDEX]], 0
2336; SINK-AFTER-NEXT:    [[TMP1:%.*]] = add nuw nsw i64 [[TMP0]], 1
2337; SINK-AFTER-NEXT:    [[TMP2:%.*]] = getelementptr inbounds i16, ptr [[A]], i64 [[TMP1]]
2338; SINK-AFTER-NEXT:    [[TMP3:%.*]] = getelementptr inbounds i16, ptr [[TMP2]], i32 0
2339; SINK-AFTER-NEXT:    [[WIDE_LOAD]] = load <4 x i16>, ptr [[TMP3]], align 2
2340; SINK-AFTER-NEXT:    [[TMP4:%.*]] = shufflevector <4 x i16> [[VECTOR_RECUR]], <4 x i16> [[WIDE_LOAD]], <4 x i32> <i32 3, i32 4, i32 5, i32 6>
2341; SINK-AFTER-NEXT:    [[TMP5:%.*]] = sext <4 x i16> [[TMP4]] to <4 x i32>
2342; SINK-AFTER-NEXT:    [[TMP6:%.*]] = add nsw <4 x i32> [[TMP5]], splat (i32 2)
2343; SINK-AFTER-NEXT:    [[TMP7:%.*]] = sext <4 x i16> [[WIDE_LOAD]] to <4 x i32>
2344; SINK-AFTER-NEXT:    [[TMP8:%.*]] = mul nsw <4 x i32> [[TMP6]], [[TMP7]]
2345; SINK-AFTER-NEXT:    [[TMP9:%.*]] = getelementptr inbounds i32, ptr [[B:%.*]], i64 [[TMP0]]
2346; SINK-AFTER-NEXT:    [[TMP10:%.*]] = getelementptr inbounds i32, ptr [[TMP9]], i32 0
2347; SINK-AFTER-NEXT:    store <4 x i32> [[TMP8]], ptr [[TMP10]], align 4
2348; SINK-AFTER-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4
2349; SINK-AFTER-NEXT:    [[TMP11:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
2350; SINK-AFTER-NEXT:    br i1 [[TMP11]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP22:![0-9]+]]
2351; SINK-AFTER:       middle.block:
2352; SINK-AFTER-NEXT:    [[VECTOR_RECUR_EXTRACT:%.*]] = extractelement <4 x i16> [[WIDE_LOAD]], i32 3
2353; SINK-AFTER-NEXT:    [[CMP_N:%.*]] = icmp eq i64 [[N]], [[N_VEC]]
2354; SINK-AFTER-NEXT:    br i1 [[CMP_N]], label [[FOR_END:%.*]], label [[SCALAR_PH]]
2355; SINK-AFTER:       scalar.ph:
2356; SINK-AFTER-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i16 [ [[VECTOR_RECUR_EXTRACT]], [[MIDDLE_BLOCK]] ], [ [[DOTPRE]], [[ENTRY:%.*]] ]
2357; SINK-AFTER-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY]] ]
2358; SINK-AFTER-NEXT:    br label [[FOR_BODY:%.*]]
2359; SINK-AFTER:       for.body:
2360; SINK-AFTER-NEXT:    [[TMP12:%.*]] = phi i16 [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ], [ [[TMP13:%.*]], [[FOR_BODY]] ]
2361; SINK-AFTER-NEXT:    [[INDVARS_IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ]
2362; SINK-AFTER-NEXT:    [[CONV:%.*]] = sext i16 [[TMP12]] to i32
2363; SINK-AFTER-NEXT:    [[ADD:%.*]] = add nsw i32 [[CONV]], 2
2364; SINK-AFTER-NEXT:    [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1
2365; SINK-AFTER-NEXT:    [[ARRAYIDX2:%.*]] = getelementptr inbounds i16, ptr [[A]], i64 [[INDVARS_IV_NEXT]]
2366; SINK-AFTER-NEXT:    [[TMP13]] = load i16, ptr [[ARRAYIDX2]], align 2
2367; SINK-AFTER-NEXT:    [[CONV3:%.*]] = sext i16 [[TMP13]] to i32
2368; SINK-AFTER-NEXT:    [[MUL:%.*]] = mul nsw i32 [[ADD]], [[CONV3]]
2369; SINK-AFTER-NEXT:    [[ARRAYIDX5:%.*]] = getelementptr inbounds i32, ptr [[B]], i64 [[INDVARS_IV]]
2370; SINK-AFTER-NEXT:    store i32 [[MUL]], ptr [[ARRAYIDX5]], align 4
2371; SINK-AFTER-NEXT:    [[EXITCOND:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT]], [[N]]
2372; SINK-AFTER-NEXT:    br i1 [[EXITCOND]], label [[FOR_END]], label [[FOR_BODY]], !llvm.loop [[LOOP23:![0-9]+]]
2373; SINK-AFTER:       for.end:
2374; SINK-AFTER-NEXT:    ret void
2375;
2376entry:
2377  %.pre = load i16, ptr %a
2378  br label %for.body
2379
2380for.body:
2381  %0 = phi i16 [ %.pre, %entry ], [ %1, %for.body ]
2382  %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
2383  %conv = sext i16 %0 to i32
2384  %add = add nsw i32 %conv, 2
2385  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
2386  %arrayidx2 = getelementptr inbounds i16, ptr %a, i64 %indvars.iv.next
2387  %1 = load i16, ptr %arrayidx2
2388  %conv3 = sext i16 %1 to i32
2389  %mul = mul nsw i32 %add, %conv3
2390  %arrayidx5 = getelementptr inbounds i32, ptr %b, i64 %indvars.iv
2391  store i32 %mul, ptr %arrayidx5
2392  %exitcond = icmp eq i64 %indvars.iv.next, %n
2393  br i1 %exitcond, label %for.end, label %for.body
2394
2395for.end:
2396  ret void
2397}
2398
2399; Do not sink branches: While branches are if-converted and do not require
2400; sinking, instructions with side effects (e.g. loads) conditioned by those
2401; branches will become users of the condition bit after vectorization and would
2402; need to be sunk if the loop is vectorized.
2403define void @do_not_sink_branch(i32 %x, ptr %in, ptr %out, i32 %tc) local_unnamed_addr #0 {
2404; UNROLL-NO-IC-LABEL: @do_not_sink_branch(
2405; UNROLL-NO-IC-NEXT:  entry:
2406; UNROLL-NO-IC-NEXT:    [[CMP530:%.*]] = icmp slt i32 0, [[TC:%.*]]
2407; UNROLL-NO-IC-NEXT:    br label [[FOR_BODY4:%.*]]
2408; UNROLL-NO-IC:       for.body4:
2409; UNROLL-NO-IC-NEXT:    [[INDVARS_IV:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[INDVARS_IV_NEXT:%.*]], [[COND_END:%.*]] ]
2410; UNROLL-NO-IC-NEXT:    [[CMP534:%.*]] = phi i1 [ [[CMP530]], [[ENTRY]] ], [ [[CMP5:%.*]], [[COND_END]] ]
2411; UNROLL-NO-IC-NEXT:    br i1 [[CMP534]], label [[COND_TRUE:%.*]], label [[COND_END]]
2412; UNROLL-NO-IC:       cond.true:
2413; UNROLL-NO-IC-NEXT:    [[ARRAYIDX7:%.*]] = getelementptr inbounds i32, ptr [[IN:%.*]], i32 [[INDVARS_IV]]
2414; UNROLL-NO-IC-NEXT:    [[IN_VAL:%.*]] = load i32, ptr [[ARRAYIDX7]], align 4
2415; UNROLL-NO-IC-NEXT:    br label [[COND_END]]
2416; UNROLL-NO-IC:       cond.end:
2417; UNROLL-NO-IC-NEXT:    [[COND:%.*]] = phi i32 [ [[IN_VAL]], [[COND_TRUE]] ], [ 0, [[FOR_BODY4]] ]
2418; UNROLL-NO-IC-NEXT:    [[ARRAYIDX8:%.*]] = getelementptr inbounds i32, ptr [[OUT:%.*]], i32 [[INDVARS_IV]]
2419; UNROLL-NO-IC-NEXT:    store i32 [[COND]], ptr [[ARRAYIDX8]], align 4
2420; UNROLL-NO-IC-NEXT:    [[INDVARS_IV_NEXT]] = add nuw nsw i32 [[INDVARS_IV]], 1
2421; UNROLL-NO-IC-NEXT:    [[CMP5]] = icmp slt i32 [[INDVARS_IV_NEXT]], [[TC]]
2422; UNROLL-NO-IC-NEXT:    [[EXITCOND:%.*]] = icmp eq i32 [[INDVARS_IV_NEXT]], [[X:%.*]]
2423; UNROLL-NO-IC-NEXT:    br i1 [[EXITCOND]], label [[FOR_END12_LOOPEXIT:%.*]], label [[FOR_BODY4]]
2424; UNROLL-NO-IC:       for.end12.loopexit:
2425; UNROLL-NO-IC-NEXT:    ret void
2426;
2427; UNROLL-NO-VF-LABEL: @do_not_sink_branch(
2428; UNROLL-NO-VF-NEXT:  entry:
2429; UNROLL-NO-VF-NEXT:    [[CMP530:%.*]] = icmp slt i32 0, [[TC:%.*]]
2430; UNROLL-NO-VF-NEXT:    br label [[FOR_BODY4:%.*]]
2431; UNROLL-NO-VF:       for.body4:
2432; UNROLL-NO-VF-NEXT:    [[INDVARS_IV:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[INDVARS_IV_NEXT:%.*]], [[COND_END:%.*]] ]
2433; UNROLL-NO-VF-NEXT:    [[CMP534:%.*]] = phi i1 [ [[CMP530]], [[ENTRY]] ], [ [[CMP5:%.*]], [[COND_END]] ]
2434; UNROLL-NO-VF-NEXT:    br i1 [[CMP534]], label [[COND_TRUE:%.*]], label [[COND_END]]
2435; UNROLL-NO-VF:       cond.true:
2436; UNROLL-NO-VF-NEXT:    [[ARRAYIDX7:%.*]] = getelementptr inbounds i32, ptr [[IN:%.*]], i32 [[INDVARS_IV]]
2437; UNROLL-NO-VF-NEXT:    [[IN_VAL:%.*]] = load i32, ptr [[ARRAYIDX7]], align 4
2438; UNROLL-NO-VF-NEXT:    br label [[COND_END]]
2439; UNROLL-NO-VF:       cond.end:
2440; UNROLL-NO-VF-NEXT:    [[COND:%.*]] = phi i32 [ [[IN_VAL]], [[COND_TRUE]] ], [ 0, [[FOR_BODY4]] ]
2441; UNROLL-NO-VF-NEXT:    [[ARRAYIDX8:%.*]] = getelementptr inbounds i32, ptr [[OUT:%.*]], i32 [[INDVARS_IV]]
2442; UNROLL-NO-VF-NEXT:    store i32 [[COND]], ptr [[ARRAYIDX8]], align 4
2443; UNROLL-NO-VF-NEXT:    [[INDVARS_IV_NEXT]] = add nuw nsw i32 [[INDVARS_IV]], 1
2444; UNROLL-NO-VF-NEXT:    [[CMP5]] = icmp slt i32 [[INDVARS_IV_NEXT]], [[TC]]
2445; UNROLL-NO-VF-NEXT:    [[EXITCOND:%.*]] = icmp eq i32 [[INDVARS_IV_NEXT]], [[X:%.*]]
2446; UNROLL-NO-VF-NEXT:    br i1 [[EXITCOND]], label [[FOR_END12_LOOPEXIT:%.*]], label [[FOR_BODY4]]
2447; UNROLL-NO-VF:       for.end12.loopexit:
2448; UNROLL-NO-VF-NEXT:    ret void
2449;
2450; SINK-AFTER-LABEL: @do_not_sink_branch(
2451; SINK-AFTER-NEXT:  entry:
2452; SINK-AFTER-NEXT:    [[CMP530:%.*]] = icmp slt i32 0, [[TC:%.*]]
2453; SINK-AFTER-NEXT:    br label [[FOR_BODY4:%.*]]
2454; SINK-AFTER:       for.body4:
2455; SINK-AFTER-NEXT:    [[INDVARS_IV:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[INDVARS_IV_NEXT:%.*]], [[COND_END:%.*]] ]
2456; SINK-AFTER-NEXT:    [[CMP534:%.*]] = phi i1 [ [[CMP530]], [[ENTRY]] ], [ [[CMP5:%.*]], [[COND_END]] ]
2457; SINK-AFTER-NEXT:    br i1 [[CMP534]], label [[COND_TRUE:%.*]], label [[COND_END]]
2458; SINK-AFTER:       cond.true:
2459; SINK-AFTER-NEXT:    [[ARRAYIDX7:%.*]] = getelementptr inbounds i32, ptr [[IN:%.*]], i32 [[INDVARS_IV]]
2460; SINK-AFTER-NEXT:    [[IN_VAL:%.*]] = load i32, ptr [[ARRAYIDX7]], align 4
2461; SINK-AFTER-NEXT:    br label [[COND_END]]
2462; SINK-AFTER:       cond.end:
2463; SINK-AFTER-NEXT:    [[COND:%.*]] = phi i32 [ [[IN_VAL]], [[COND_TRUE]] ], [ 0, [[FOR_BODY4]] ]
2464; SINK-AFTER-NEXT:    [[ARRAYIDX8:%.*]] = getelementptr inbounds i32, ptr [[OUT:%.*]], i32 [[INDVARS_IV]]
2465; SINK-AFTER-NEXT:    store i32 [[COND]], ptr [[ARRAYIDX8]], align 4
2466; SINK-AFTER-NEXT:    [[INDVARS_IV_NEXT]] = add nuw nsw i32 [[INDVARS_IV]], 1
2467; SINK-AFTER-NEXT:    [[CMP5]] = icmp slt i32 [[INDVARS_IV_NEXT]], [[TC]]
2468; SINK-AFTER-NEXT:    [[EXITCOND:%.*]] = icmp eq i32 [[INDVARS_IV_NEXT]], [[X:%.*]]
2469; SINK-AFTER-NEXT:    br i1 [[EXITCOND]], label [[FOR_END12_LOOPEXIT:%.*]], label [[FOR_BODY4]]
2470; SINK-AFTER:       for.end12.loopexit:
2471; SINK-AFTER-NEXT:    ret void
2472;
2473entry:
2474  %cmp530 = icmp slt i32 0, %tc
2475  br label %for.body4
2476
2477for.body4:                                        ; preds = %cond.end, %entry
2478  %indvars.iv = phi i32 [ 0, %entry ], [ %indvars.iv.next, %cond.end ]
2479  %cmp534 = phi i1 [ %cmp530, %entry ], [ %cmp5, %cond.end ]
2480  br i1 %cmp534, label %cond.true, label %cond.end
2481
2482cond.true:                                        ; preds = %for.body4
2483  %arrayidx7 = getelementptr inbounds i32, ptr %in, i32 %indvars.iv
2484  %in.val = load i32, ptr %arrayidx7, align 4
2485  br label %cond.end
2486
2487cond.end:                                         ; preds = %for.body4, %cond.true
2488  %cond = phi i32 [ %in.val, %cond.true ], [ 0, %for.body4 ]
2489  %arrayidx8 = getelementptr inbounds i32, ptr %out, i32 %indvars.iv
2490  store i32 %cond, ptr %arrayidx8, align 4
2491  %indvars.iv.next = add nuw nsw i32 %indvars.iv, 1
2492  %cmp5 = icmp slt i32 %indvars.iv.next, %tc
2493  %exitcond = icmp eq i32 %indvars.iv.next, %x
2494  br i1 %exitcond, label %for.end12.loopexit, label %for.body4
2495
2496for.end12.loopexit:                               ; preds = %cond.end
2497  ret void
2498}
2499
2500; Dead instructions, like the exit condition are not part of the actual VPlan
2501; and do not need to be sunk. PR44634.
2502define void @sink_dead_inst(ptr %a) {
2503; UNROLL-NO-IC-LABEL: @sink_dead_inst(
2504; UNROLL-NO-IC-NEXT:  entry:
2505; UNROLL-NO-IC-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
2506; UNROLL-NO-IC:       vector.ph:
2507; UNROLL-NO-IC-NEXT:    br label [[VECTOR_BODY:%.*]]
2508; UNROLL-NO-IC:       vector.body:
2509; UNROLL-NO-IC-NEXT:    [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
2510; UNROLL-NO-IC-NEXT:    [[VEC_IND:%.*]] = phi <4 x i16> [ <i16 -27, i16 -26, i16 -25, i16 -24>, [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[VECTOR_BODY]] ]
2511; UNROLL-NO-IC-NEXT:    [[VECTOR_RECUR:%.*]] = phi <4 x i16> [ <i16 poison, i16 poison, i16 poison, i16 0>, [[VECTOR_PH]] ], [ [[TMP5:%.*]], [[VECTOR_BODY]] ]
2512; UNROLL-NO-IC-NEXT:    [[VECTOR_RECUR1:%.*]] = phi <4 x i32> [ <i32 poison, i32 poison, i32 poison, i32 -27>, [[VECTOR_PH]] ], [ [[TMP3:%.*]], [[VECTOR_BODY]] ]
2513; UNROLL-NO-IC-NEXT:    [[STEP_ADD:%.*]] = add <4 x i16> [[VEC_IND]], splat (i16 4)
2514; UNROLL-NO-IC-NEXT:    [[DOTCAST:%.*]] = trunc i32 [[INDEX]] to i16
2515; UNROLL-NO-IC-NEXT:    [[OFFSET_IDX:%.*]] = add i16 -27, [[DOTCAST]]
2516; UNROLL-NO-IC-NEXT:    [[TMP0:%.*]] = add i16 [[OFFSET_IDX]], 0
2517; UNROLL-NO-IC-NEXT:    [[TMP1:%.*]] = add <4 x i16> [[VEC_IND]], splat (i16 1)
2518; UNROLL-NO-IC-NEXT:    [[TMP2:%.*]] = add <4 x i16> [[STEP_ADD]], splat (i16 1)
2519; UNROLL-NO-IC-NEXT:    [[TMP3]] = zext <4 x i16> [[TMP2]] to <4 x i32>
2520; UNROLL-NO-IC-NEXT:    [[TMP4:%.*]] = add <4 x i16> [[TMP1]], splat (i16 5)
2521; UNROLL-NO-IC-NEXT:    [[TMP5]] = add <4 x i16> [[TMP2]], splat (i16 5)
2522; UNROLL-NO-IC-NEXT:    [[TMP6:%.*]] = shufflevector <4 x i16> [[VECTOR_RECUR]], <4 x i16> [[TMP4]], <4 x i32> <i32 3, i32 4, i32 5, i32 6>
2523; UNROLL-NO-IC-NEXT:    [[TMP7:%.*]] = shufflevector <4 x i16> [[TMP4]], <4 x i16> [[TMP5]], <4 x i32> <i32 3, i32 4, i32 5, i32 6>
2524; UNROLL-NO-IC-NEXT:    [[TMP8:%.*]] = sub <4 x i16> [[TMP6]], splat (i16 10)
2525; UNROLL-NO-IC-NEXT:    [[TMP9:%.*]] = sub <4 x i16> [[TMP7]], splat (i16 10)
2526; UNROLL-NO-IC-NEXT:    [[TMP10:%.*]] = getelementptr i16, ptr [[A:%.*]], i16 [[TMP0]]
2527; UNROLL-NO-IC-NEXT:    [[TMP11:%.*]] = getelementptr i16, ptr [[TMP10]], i32 0
2528; UNROLL-NO-IC-NEXT:    [[TMP12:%.*]] = getelementptr i16, ptr [[TMP10]], i32 4
2529; UNROLL-NO-IC-NEXT:    store <4 x i16> [[TMP8]], ptr [[TMP11]], align 2
2530; UNROLL-NO-IC-NEXT:    store <4 x i16> [[TMP9]], ptr [[TMP12]], align 2
2531; UNROLL-NO-IC-NEXT:    [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 8
2532; UNROLL-NO-IC-NEXT:    [[VEC_IND_NEXT]] = add <4 x i16> [[STEP_ADD]], splat (i16 4)
2533; UNROLL-NO-IC-NEXT:    [[TMP13:%.*]] = icmp eq i32 [[INDEX_NEXT]], 40
2534; UNROLL-NO-IC-NEXT:    br i1 [[TMP13]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP24:![0-9]+]]
2535; UNROLL-NO-IC:       middle.block:
2536; UNROLL-NO-IC-NEXT:    [[VECTOR_RECUR_EXTRACT:%.*]] = extractelement <4 x i16> [[TMP5]], i32 3
2537; UNROLL-NO-IC-NEXT:    [[VECTOR_RECUR_EXTRACT2:%.*]] = extractelement <4 x i32> [[TMP3]], i32 3
2538; UNROLL-NO-IC-NEXT:    br i1 false, label [[FOR_END:%.*]], label [[SCALAR_PH]]
2539; UNROLL-NO-IC:       scalar.ph:
2540; UNROLL-NO-IC-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i16 [ 13, [[MIDDLE_BLOCK]] ], [ -27, [[ENTRY:%.*]] ]
2541; UNROLL-NO-IC-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i16 [ [[VECTOR_RECUR_EXTRACT]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY]] ]
2542; UNROLL-NO-IC-NEXT:    [[SCALAR_RECUR_INIT3:%.*]] = phi i32 [ [[VECTOR_RECUR_EXTRACT2]], [[MIDDLE_BLOCK]] ], [ -27, [[ENTRY]] ]
2543; UNROLL-NO-IC-NEXT:    br label [[FOR_COND:%.*]]
2544; UNROLL-NO-IC:       for.cond:
2545; UNROLL-NO-IC-NEXT:    [[IV:%.*]] = phi i16 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], [[FOR_COND]] ]
2546; UNROLL-NO-IC-NEXT:    [[REC_1:%.*]] = phi i16 [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ], [ [[REC_1_PREV:%.*]], [[FOR_COND]] ]
2547; UNROLL-NO-IC-NEXT:    [[REC_2:%.*]] = phi i32 [ [[SCALAR_RECUR_INIT3]], [[SCALAR_PH]] ], [ [[REC_2_PREV:%.*]], [[FOR_COND]] ]
2548; UNROLL-NO-IC-NEXT:    [[USE_REC_1:%.*]] = sub i16 [[REC_1]], 10
2549; UNROLL-NO-IC-NEXT:    [[CMP:%.*]] = icmp eq i32 [[REC_2]], 15
2550; UNROLL-NO-IC-NEXT:    [[IV_NEXT]] = add i16 [[IV]], 1
2551; UNROLL-NO-IC-NEXT:    [[REC_2_PREV]] = zext i16 [[IV_NEXT]] to i32
2552; UNROLL-NO-IC-NEXT:    [[REC_1_PREV]] = add i16 [[IV_NEXT]], 5
2553; UNROLL-NO-IC-NEXT:    [[GEP:%.*]] = getelementptr i16, ptr [[A]], i16 [[IV]]
2554; UNROLL-NO-IC-NEXT:    store i16 [[USE_REC_1]], ptr [[GEP]], align 2
2555; UNROLL-NO-IC-NEXT:    br i1 [[CMP]], label [[FOR_END]], label [[FOR_COND]], !llvm.loop [[LOOP25:![0-9]+]]
2556; UNROLL-NO-IC:       for.end:
2557; UNROLL-NO-IC-NEXT:    ret void
2558;
2559; UNROLL-NO-VF-LABEL: @sink_dead_inst(
2560; UNROLL-NO-VF-NEXT:  entry:
2561; UNROLL-NO-VF-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
2562; UNROLL-NO-VF:       vector.ph:
2563; UNROLL-NO-VF-NEXT:    br label [[VECTOR_BODY:%.*]]
2564; UNROLL-NO-VF:       vector.body:
2565; UNROLL-NO-VF-NEXT:    [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
2566; UNROLL-NO-VF-NEXT:    [[VECTOR_RECUR:%.*]] = phi i16 [ 0, [[VECTOR_PH]] ], [ [[TMP6:%.*]], [[VECTOR_BODY]] ]
2567; UNROLL-NO-VF-NEXT:    [[VECTOR_RECUR1:%.*]] = phi i32 [ -27, [[VECTOR_PH]] ], [ [[TMP4:%.*]], [[VECTOR_BODY]] ]
2568; UNROLL-NO-VF-NEXT:    [[DOTCAST:%.*]] = trunc i32 [[INDEX]] to i16
2569; UNROLL-NO-VF-NEXT:    [[OFFSET_IDX:%.*]] = add i16 -27, [[DOTCAST]]
2570; UNROLL-NO-VF-NEXT:    [[TMP0:%.*]] = add i16 [[OFFSET_IDX]], 0
2571; UNROLL-NO-VF-NEXT:    [[TMP1:%.*]] = add i16 [[OFFSET_IDX]], 1
2572; UNROLL-NO-VF-NEXT:    [[TMP2:%.*]] = add i16 [[TMP0]], 1
2573; UNROLL-NO-VF-NEXT:    [[TMP3:%.*]] = add i16 [[TMP1]], 1
2574; UNROLL-NO-VF-NEXT:    [[TMP4]] = zext i16 [[TMP3]] to i32
2575; UNROLL-NO-VF-NEXT:    [[TMP5:%.*]] = add i16 [[TMP2]], 5
2576; UNROLL-NO-VF-NEXT:    [[TMP6]] = add i16 [[TMP3]], 5
2577; UNROLL-NO-VF-NEXT:    [[TMP7:%.*]] = sub i16 [[VECTOR_RECUR]], 10
2578; UNROLL-NO-VF-NEXT:    [[TMP8:%.*]] = sub i16 [[TMP5]], 10
2579; UNROLL-NO-VF-NEXT:    [[TMP9:%.*]] = getelementptr i16, ptr [[A:%.*]], i16 [[TMP0]]
2580; UNROLL-NO-VF-NEXT:    [[TMP10:%.*]] = getelementptr i16, ptr [[A]], i16 [[TMP1]]
2581; UNROLL-NO-VF-NEXT:    store i16 [[TMP7]], ptr [[TMP9]], align 2
2582; UNROLL-NO-VF-NEXT:    store i16 [[TMP8]], ptr [[TMP10]], align 2
2583; UNROLL-NO-VF-NEXT:    [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 2
2584; UNROLL-NO-VF-NEXT:    [[TMP11:%.*]] = icmp eq i32 [[INDEX_NEXT]], 42
2585; UNROLL-NO-VF-NEXT:    br i1 [[TMP11]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP24:![0-9]+]]
2586; UNROLL-NO-VF:       middle.block:
2587; UNROLL-NO-VF-NEXT:    br i1 false, label [[FOR_END:%.*]], label [[SCALAR_PH]]
2588; UNROLL-NO-VF:       scalar.ph:
2589; UNROLL-NO-VF-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i16 [ 15, [[MIDDLE_BLOCK]] ], [ -27, [[ENTRY:%.*]] ]
2590; UNROLL-NO-VF-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i16 [ [[TMP6]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY]] ]
2591; UNROLL-NO-VF-NEXT:    [[SCALAR_RECUR_INIT2:%.*]] = phi i32 [ [[TMP4]], [[MIDDLE_BLOCK]] ], [ -27, [[ENTRY]] ]
2592; UNROLL-NO-VF-NEXT:    br label [[FOR_COND:%.*]]
2593; UNROLL-NO-VF:       for.cond:
2594; UNROLL-NO-VF-NEXT:    [[IV:%.*]] = phi i16 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], [[FOR_COND]] ]
2595; UNROLL-NO-VF-NEXT:    [[REC_1:%.*]] = phi i16 [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ], [ [[REC_1_PREV:%.*]], [[FOR_COND]] ]
2596; UNROLL-NO-VF-NEXT:    [[REC_2:%.*]] = phi i32 [ [[SCALAR_RECUR_INIT2]], [[SCALAR_PH]] ], [ [[REC_2_PREV:%.*]], [[FOR_COND]] ]
2597; UNROLL-NO-VF-NEXT:    [[USE_REC_1:%.*]] = sub i16 [[REC_1]], 10
2598; UNROLL-NO-VF-NEXT:    [[CMP:%.*]] = icmp eq i32 [[REC_2]], 15
2599; UNROLL-NO-VF-NEXT:    [[IV_NEXT]] = add i16 [[IV]], 1
2600; UNROLL-NO-VF-NEXT:    [[REC_2_PREV]] = zext i16 [[IV_NEXT]] to i32
2601; UNROLL-NO-VF-NEXT:    [[REC_1_PREV]] = add i16 [[IV_NEXT]], 5
2602; UNROLL-NO-VF-NEXT:    [[GEP:%.*]] = getelementptr i16, ptr [[A]], i16 [[IV]]
2603; UNROLL-NO-VF-NEXT:    store i16 [[USE_REC_1]], ptr [[GEP]], align 2
2604; UNROLL-NO-VF-NEXT:    br i1 [[CMP]], label [[FOR_END]], label [[FOR_COND]], !llvm.loop [[LOOP25:![0-9]+]]
2605; UNROLL-NO-VF:       for.end:
2606; UNROLL-NO-VF-NEXT:    ret void
2607;
2608; SINK-AFTER-LABEL: @sink_dead_inst(
2609; SINK-AFTER-NEXT:  entry:
2610; SINK-AFTER-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
2611; SINK-AFTER:       vector.ph:
2612; SINK-AFTER-NEXT:    br label [[VECTOR_BODY:%.*]]
2613; SINK-AFTER:       vector.body:
2614; SINK-AFTER-NEXT:    [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
2615; SINK-AFTER-NEXT:    [[VEC_IND:%.*]] = phi <4 x i16> [ <i16 -27, i16 -26, i16 -25, i16 -24>, [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[VECTOR_BODY]] ]
2616; SINK-AFTER-NEXT:    [[VECTOR_RECUR:%.*]] = phi <4 x i16> [ <i16 poison, i16 poison, i16 poison, i16 0>, [[VECTOR_PH]] ], [ [[TMP3:%.*]], [[VECTOR_BODY]] ]
2617; SINK-AFTER-NEXT:    [[VECTOR_RECUR1:%.*]] = phi <4 x i32> [ <i32 poison, i32 poison, i32 poison, i32 -27>, [[VECTOR_PH]] ], [ [[TMP2:%.*]], [[VECTOR_BODY]] ]
2618; SINK-AFTER-NEXT:    [[DOTCAST:%.*]] = trunc i32 [[INDEX]] to i16
2619; SINK-AFTER-NEXT:    [[OFFSET_IDX:%.*]] = add i16 -27, [[DOTCAST]]
2620; SINK-AFTER-NEXT:    [[TMP0:%.*]] = add i16 [[OFFSET_IDX]], 0
2621; SINK-AFTER-NEXT:    [[TMP1:%.*]] = add <4 x i16> [[VEC_IND]], splat (i16 1)
2622; SINK-AFTER-NEXT:    [[TMP2]] = zext <4 x i16> [[TMP1]] to <4 x i32>
2623; SINK-AFTER-NEXT:    [[TMP3]] = add <4 x i16> [[TMP1]], splat (i16 5)
2624; SINK-AFTER-NEXT:    [[TMP4:%.*]] = shufflevector <4 x i16> [[VECTOR_RECUR]], <4 x i16> [[TMP3]], <4 x i32> <i32 3, i32 4, i32 5, i32 6>
2625; SINK-AFTER-NEXT:    [[TMP5:%.*]] = sub <4 x i16> [[TMP4]], splat (i16 10)
2626; SINK-AFTER-NEXT:    [[TMP6:%.*]] = getelementptr i16, ptr [[A:%.*]], i16 [[TMP0]]
2627; SINK-AFTER-NEXT:    [[TMP7:%.*]] = getelementptr i16, ptr [[TMP6]], i32 0
2628; SINK-AFTER-NEXT:    store <4 x i16> [[TMP5]], ptr [[TMP7]], align 2
2629; SINK-AFTER-NEXT:    [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 4
2630; SINK-AFTER-NEXT:    [[VEC_IND_NEXT]] = add <4 x i16> [[VEC_IND]], splat (i16 4)
2631; SINK-AFTER-NEXT:    [[TMP8:%.*]] = icmp eq i32 [[INDEX_NEXT]], 40
2632; SINK-AFTER-NEXT:    br i1 [[TMP8]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP24:![0-9]+]]
2633; SINK-AFTER:       middle.block:
2634; SINK-AFTER-NEXT:    [[VECTOR_RECUR_EXTRACT:%.*]] = extractelement <4 x i16> [[TMP3]], i32 3
2635; SINK-AFTER-NEXT:    [[VECTOR_RECUR_EXTRACT2:%.*]] = extractelement <4 x i32> [[TMP2]], i32 3
2636; SINK-AFTER-NEXT:    br i1 false, label [[FOR_END:%.*]], label [[SCALAR_PH]]
2637; SINK-AFTER:       scalar.ph:
2638; SINK-AFTER-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i16 [ 13, [[MIDDLE_BLOCK]] ], [ -27, [[ENTRY:%.*]] ]
2639; SINK-AFTER-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i16 [ [[VECTOR_RECUR_EXTRACT]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY]] ]
2640; SINK-AFTER-NEXT:    [[SCALAR_RECUR_INIT3:%.*]] = phi i32 [ [[VECTOR_RECUR_EXTRACT2]], [[MIDDLE_BLOCK]] ], [ -27, [[ENTRY]] ]
2641; SINK-AFTER-NEXT:    br label [[FOR_COND:%.*]]
2642; SINK-AFTER:       for.cond:
2643; SINK-AFTER-NEXT:    [[IV:%.*]] = phi i16 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], [[FOR_COND]] ]
2644; SINK-AFTER-NEXT:    [[REC_1:%.*]] = phi i16 [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ], [ [[REC_1_PREV:%.*]], [[FOR_COND]] ]
2645; SINK-AFTER-NEXT:    [[REC_2:%.*]] = phi i32 [ [[SCALAR_RECUR_INIT3]], [[SCALAR_PH]] ], [ [[REC_2_PREV:%.*]], [[FOR_COND]] ]
2646; SINK-AFTER-NEXT:    [[USE_REC_1:%.*]] = sub i16 [[REC_1]], 10
2647; SINK-AFTER-NEXT:    [[CMP:%.*]] = icmp eq i32 [[REC_2]], 15
2648; SINK-AFTER-NEXT:    [[IV_NEXT]] = add i16 [[IV]], 1
2649; SINK-AFTER-NEXT:    [[REC_2_PREV]] = zext i16 [[IV_NEXT]] to i32
2650; SINK-AFTER-NEXT:    [[REC_1_PREV]] = add i16 [[IV_NEXT]], 5
2651; SINK-AFTER-NEXT:    [[GEP:%.*]] = getelementptr i16, ptr [[A]], i16 [[IV]]
2652; SINK-AFTER-NEXT:    store i16 [[USE_REC_1]], ptr [[GEP]], align 2
2653; SINK-AFTER-NEXT:    br i1 [[CMP]], label [[FOR_END]], label [[FOR_COND]], !llvm.loop [[LOOP25:![0-9]+]]
2654; SINK-AFTER:       for.end:
2655; SINK-AFTER-NEXT:    ret void
2656;
2657entry:
2658  br label %for.cond
2659
2660for.cond:
2661  %iv = phi i16 [ -27, %entry ], [ %iv.next, %for.cond ]
2662  %rec.1 = phi i16 [ 0, %entry ], [ %rec.1.prev, %for.cond ]
2663  %rec.2 = phi i32 [ -27, %entry ], [ %rec.2.prev, %for.cond ]
2664  %use.rec.1 = sub i16 %rec.1, 10
2665  %cmp = icmp eq i32 %rec.2, 15
2666  %iv.next = add i16 %iv, 1
2667  %rec.2.prev = zext i16 %iv.next to i32
2668  %rec.1.prev = add i16 %iv.next, 5
2669  %gep = getelementptr i16, ptr %a, i16 %iv
2670  store i16 %use.rec.1, ptr %gep
2671  br i1 %cmp, label %for.end, label %for.cond
2672
2673for.end:
2674  ret void
2675}
2676
2677define i32 @sink_into_replication_region(i32 %y) {
2678; UNROLL-NO-IC-LABEL: @sink_into_replication_region(
2679; UNROLL-NO-IC-NEXT:  bb:
2680; UNROLL-NO-IC-NEXT:    [[TMP0:%.*]] = add i32 [[Y:%.*]], 1
2681; UNROLL-NO-IC-NEXT:    [[SMIN:%.*]] = call i32 @llvm.smin.i32(i32 [[Y]], i32 1)
2682; UNROLL-NO-IC-NEXT:    [[TMP1:%.*]] = sub i32 [[TMP0]], [[SMIN]]
2683; UNROLL-NO-IC-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]], !prof [[PROF26:![0-9]+]]
2684; UNROLL-NO-IC:       vector.ph:
2685; UNROLL-NO-IC-NEXT:    [[N_RND_UP:%.*]] = add i32 [[TMP1]], 7
2686; UNROLL-NO-IC-NEXT:    [[N_MOD_VF:%.*]] = urem i32 [[N_RND_UP]], 8
2687; UNROLL-NO-IC-NEXT:    [[N_VEC:%.*]] = sub i32 [[N_RND_UP]], [[N_MOD_VF]]
2688; UNROLL-NO-IC-NEXT:    [[TRIP_COUNT_MINUS_1:%.*]] = sub i32 [[TMP1]], 1
2689; UNROLL-NO-IC-NEXT:    [[IND_END:%.*]] = sub i32 [[Y]], [[N_VEC]]
2690; UNROLL-NO-IC-NEXT:    [[BROADCAST_SPLATINSERT5:%.*]] = insertelement <4 x i32> poison, i32 [[TRIP_COUNT_MINUS_1]], i64 0
2691; UNROLL-NO-IC-NEXT:    [[BROADCAST_SPLAT6:%.*]] = shufflevector <4 x i32> [[BROADCAST_SPLATINSERT5]], <4 x i32> poison, <4 x i32> zeroinitializer
2692; UNROLL-NO-IC-NEXT:    br label [[VECTOR_BODY:%.*]]
2693; UNROLL-NO-IC:       vector.body:
2694; UNROLL-NO-IC-NEXT:    [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[PRED_UDIV_CONTINUE20:%.*]] ]
2695; UNROLL-NO-IC-NEXT:    [[VECTOR_RECUR:%.*]] = phi <4 x i32> [ <i32 poison, i32 poison, i32 poison, i32 0>, [[VECTOR_PH]] ], [ [[TMP43:%.*]], [[PRED_UDIV_CONTINUE20]] ]
2696; UNROLL-NO-IC-NEXT:    [[VEC_PHI:%.*]] = phi <4 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP46:%.*]], [[PRED_UDIV_CONTINUE20]] ]
2697; UNROLL-NO-IC-NEXT:    [[VEC_PHI1:%.*]] = phi <4 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP47:%.*]], [[PRED_UDIV_CONTINUE20]] ]
2698; UNROLL-NO-IC-NEXT:    [[OFFSET_IDX:%.*]] = sub i32 [[Y]], [[INDEX]]
2699; UNROLL-NO-IC-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <4 x i32> poison, i32 [[INDEX]], i64 0
2700; UNROLL-NO-IC-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <4 x i32> [[BROADCAST_SPLATINSERT]], <4 x i32> poison, <4 x i32> zeroinitializer
2701; UNROLL-NO-IC-NEXT:    [[VEC_IV:%.*]] = add <4 x i32> [[BROADCAST_SPLAT]], <i32 0, i32 1, i32 2, i32 3>
2702; UNROLL-NO-IC-NEXT:    [[VEC_IV4:%.*]] = add <4 x i32> [[BROADCAST_SPLAT]], <i32 4, i32 5, i32 6, i32 7>
2703; UNROLL-NO-IC-NEXT:    [[TMP2:%.*]] = icmp ule <4 x i32> [[VEC_IV]], [[BROADCAST_SPLAT6]]
2704; UNROLL-NO-IC-NEXT:    [[TMP3:%.*]] = icmp ule <4 x i32> [[VEC_IV4]], [[BROADCAST_SPLAT6]]
2705; UNROLL-NO-IC-NEXT:    [[TMP4:%.*]] = extractelement <4 x i1> [[TMP2]], i32 0
2706; UNROLL-NO-IC-NEXT:    br i1 [[TMP4]], label [[PRED_UDIV_IF:%.*]], label [[PRED_UDIV_CONTINUE:%.*]]
2707; UNROLL-NO-IC:       pred.udiv.if:
2708; UNROLL-NO-IC-NEXT:    [[TMP5:%.*]] = add i32 [[OFFSET_IDX]], 0
2709; UNROLL-NO-IC-NEXT:    [[TMP6:%.*]] = udiv i32 219220132, [[TMP5]]
2710; UNROLL-NO-IC-NEXT:    [[TMP7:%.*]] = insertelement <4 x i32> poison, i32 [[TMP6]], i32 0
2711; UNROLL-NO-IC-NEXT:    br label [[PRED_UDIV_CONTINUE]]
2712; UNROLL-NO-IC:       pred.udiv.continue:
2713; UNROLL-NO-IC-NEXT:    [[TMP8:%.*]] = phi <4 x i32> [ poison, [[VECTOR_BODY]] ], [ [[TMP7]], [[PRED_UDIV_IF]] ]
2714; UNROLL-NO-IC-NEXT:    [[TMP9:%.*]] = extractelement <4 x i1> [[TMP2]], i32 1
2715; UNROLL-NO-IC-NEXT:    br i1 [[TMP9]], label [[PRED_UDIV_IF7:%.*]], label [[PRED_UDIV_CONTINUE8:%.*]]
2716; UNROLL-NO-IC:       pred.udiv.if7:
2717; UNROLL-NO-IC-NEXT:    [[TMP10:%.*]] = add i32 [[OFFSET_IDX]], -1
2718; UNROLL-NO-IC-NEXT:    [[TMP11:%.*]] = udiv i32 219220132, [[TMP10]]
2719; UNROLL-NO-IC-NEXT:    [[TMP12:%.*]] = insertelement <4 x i32> [[TMP8]], i32 [[TMP11]], i32 1
2720; UNROLL-NO-IC-NEXT:    br label [[PRED_UDIV_CONTINUE8]]
2721; UNROLL-NO-IC:       pred.udiv.continue8:
2722; UNROLL-NO-IC-NEXT:    [[TMP13:%.*]] = phi <4 x i32> [ [[TMP8]], [[PRED_UDIV_CONTINUE]] ], [ [[TMP12]], [[PRED_UDIV_IF7]] ]
2723; UNROLL-NO-IC-NEXT:    [[TMP14:%.*]] = extractelement <4 x i1> [[TMP2]], i32 2
2724; UNROLL-NO-IC-NEXT:    br i1 [[TMP14]], label [[PRED_UDIV_IF9:%.*]], label [[PRED_UDIV_CONTINUE10:%.*]]
2725; UNROLL-NO-IC:       pred.udiv.if9:
2726; UNROLL-NO-IC-NEXT:    [[TMP15:%.*]] = add i32 [[OFFSET_IDX]], -2
2727; UNROLL-NO-IC-NEXT:    [[TMP16:%.*]] = udiv i32 219220132, [[TMP15]]
2728; UNROLL-NO-IC-NEXT:    [[TMP17:%.*]] = insertelement <4 x i32> [[TMP13]], i32 [[TMP16]], i32 2
2729; UNROLL-NO-IC-NEXT:    br label [[PRED_UDIV_CONTINUE10]]
2730; UNROLL-NO-IC:       pred.udiv.continue10:
2731; UNROLL-NO-IC-NEXT:    [[TMP18:%.*]] = phi <4 x i32> [ [[TMP13]], [[PRED_UDIV_CONTINUE8]] ], [ [[TMP17]], [[PRED_UDIV_IF9]] ]
2732; UNROLL-NO-IC-NEXT:    [[TMP19:%.*]] = extractelement <4 x i1> [[TMP2]], i32 3
2733; UNROLL-NO-IC-NEXT:    br i1 [[TMP19]], label [[PRED_UDIV_IF11:%.*]], label [[PRED_UDIV_CONTINUE12:%.*]]
2734; UNROLL-NO-IC:       pred.udiv.if11:
2735; UNROLL-NO-IC-NEXT:    [[TMP20:%.*]] = add i32 [[OFFSET_IDX]], -3
2736; UNROLL-NO-IC-NEXT:    [[TMP21:%.*]] = udiv i32 219220132, [[TMP20]]
2737; UNROLL-NO-IC-NEXT:    [[TMP22:%.*]] = insertelement <4 x i32> [[TMP18]], i32 [[TMP21]], i32 3
2738; UNROLL-NO-IC-NEXT:    br label [[PRED_UDIV_CONTINUE12]]
2739; UNROLL-NO-IC:       pred.udiv.continue12:
2740; UNROLL-NO-IC-NEXT:    [[TMP23:%.*]] = phi <4 x i32> [ [[TMP18]], [[PRED_UDIV_CONTINUE10]] ], [ [[TMP22]], [[PRED_UDIV_IF11]] ]
2741; UNROLL-NO-IC-NEXT:    [[TMP24:%.*]] = extractelement <4 x i1> [[TMP3]], i32 0
2742; UNROLL-NO-IC-NEXT:    br i1 [[TMP24]], label [[PRED_UDIV_IF13:%.*]], label [[PRED_UDIV_CONTINUE14:%.*]]
2743; UNROLL-NO-IC:       pred.udiv.if13:
2744; UNROLL-NO-IC-NEXT:    [[TMP25:%.*]] = add i32 [[OFFSET_IDX]], -4
2745; UNROLL-NO-IC-NEXT:    [[TMP26:%.*]] = udiv i32 219220132, [[TMP25]]
2746; UNROLL-NO-IC-NEXT:    [[TMP27:%.*]] = insertelement <4 x i32> poison, i32 [[TMP26]], i32 0
2747; UNROLL-NO-IC-NEXT:    br label [[PRED_UDIV_CONTINUE14]]
2748; UNROLL-NO-IC:       pred.udiv.continue14:
2749; UNROLL-NO-IC-NEXT:    [[TMP28:%.*]] = phi <4 x i32> [ poison, [[PRED_UDIV_CONTINUE12]] ], [ [[TMP27]], [[PRED_UDIV_IF13]] ]
2750; UNROLL-NO-IC-NEXT:    [[TMP29:%.*]] = extractelement <4 x i1> [[TMP3]], i32 1
2751; UNROLL-NO-IC-NEXT:    br i1 [[TMP29]], label [[PRED_UDIV_IF15:%.*]], label [[PRED_UDIV_CONTINUE16:%.*]]
2752; UNROLL-NO-IC:       pred.udiv.if15:
2753; UNROLL-NO-IC-NEXT:    [[TMP30:%.*]] = add i32 [[OFFSET_IDX]], -5
2754; UNROLL-NO-IC-NEXT:    [[TMP31:%.*]] = udiv i32 219220132, [[TMP30]]
2755; UNROLL-NO-IC-NEXT:    [[TMP32:%.*]] = insertelement <4 x i32> [[TMP28]], i32 [[TMP31]], i32 1
2756; UNROLL-NO-IC-NEXT:    br label [[PRED_UDIV_CONTINUE16]]
2757; UNROLL-NO-IC:       pred.udiv.continue16:
2758; UNROLL-NO-IC-NEXT:    [[TMP33:%.*]] = phi <4 x i32> [ [[TMP28]], [[PRED_UDIV_CONTINUE14]] ], [ [[TMP32]], [[PRED_UDIV_IF15]] ]
2759; UNROLL-NO-IC-NEXT:    [[TMP34:%.*]] = extractelement <4 x i1> [[TMP3]], i32 2
2760; UNROLL-NO-IC-NEXT:    br i1 [[TMP34]], label [[PRED_UDIV_IF17:%.*]], label [[PRED_UDIV_CONTINUE18:%.*]]
2761; UNROLL-NO-IC:       pred.udiv.if17:
2762; UNROLL-NO-IC-NEXT:    [[TMP35:%.*]] = add i32 [[OFFSET_IDX]], -6
2763; UNROLL-NO-IC-NEXT:    [[TMP36:%.*]] = udiv i32 219220132, [[TMP35]]
2764; UNROLL-NO-IC-NEXT:    [[TMP37:%.*]] = insertelement <4 x i32> [[TMP33]], i32 [[TMP36]], i32 2
2765; UNROLL-NO-IC-NEXT:    br label [[PRED_UDIV_CONTINUE18]]
2766; UNROLL-NO-IC:       pred.udiv.continue18:
2767; UNROLL-NO-IC-NEXT:    [[TMP38:%.*]] = phi <4 x i32> [ [[TMP33]], [[PRED_UDIV_CONTINUE16]] ], [ [[TMP37]], [[PRED_UDIV_IF17]] ]
2768; UNROLL-NO-IC-NEXT:    [[TMP39:%.*]] = extractelement <4 x i1> [[TMP3]], i32 3
2769; UNROLL-NO-IC-NEXT:    br i1 [[TMP39]], label [[PRED_UDIV_IF19:%.*]], label [[PRED_UDIV_CONTINUE20]]
2770; UNROLL-NO-IC:       pred.udiv.if19:
2771; UNROLL-NO-IC-NEXT:    [[TMP40:%.*]] = add i32 [[OFFSET_IDX]], -7
2772; UNROLL-NO-IC-NEXT:    [[TMP41:%.*]] = udiv i32 219220132, [[TMP40]]
2773; UNROLL-NO-IC-NEXT:    [[TMP42:%.*]] = insertelement <4 x i32> [[TMP38]], i32 [[TMP41]], i32 3
2774; UNROLL-NO-IC-NEXT:    br label [[PRED_UDIV_CONTINUE20]]
2775; UNROLL-NO-IC:       pred.udiv.continue20:
2776; UNROLL-NO-IC-NEXT:    [[TMP43]] = phi <4 x i32> [ [[TMP38]], [[PRED_UDIV_CONTINUE18]] ], [ [[TMP42]], [[PRED_UDIV_IF19]] ]
2777; UNROLL-NO-IC-NEXT:    [[TMP44:%.*]] = shufflevector <4 x i32> [[VECTOR_RECUR]], <4 x i32> [[TMP23]], <4 x i32> <i32 3, i32 4, i32 5, i32 6>
2778; UNROLL-NO-IC-NEXT:    [[TMP45:%.*]] = shufflevector <4 x i32> [[TMP23]], <4 x i32> [[TMP43]], <4 x i32> <i32 3, i32 4, i32 5, i32 6>
2779; UNROLL-NO-IC-NEXT:    [[TMP46]] = add <4 x i32> [[VEC_PHI]], [[TMP44]]
2780; UNROLL-NO-IC-NEXT:    [[TMP47]] = add <4 x i32> [[VEC_PHI1]], [[TMP45]]
2781; UNROLL-NO-IC-NEXT:    [[TMP48:%.*]] = select <4 x i1> [[TMP2]], <4 x i32> [[TMP46]], <4 x i32> [[VEC_PHI]]
2782; UNROLL-NO-IC-NEXT:    [[TMP49:%.*]] = select <4 x i1> [[TMP3]], <4 x i32> [[TMP47]], <4 x i32> [[VEC_PHI1]]
2783; UNROLL-NO-IC-NEXT:    [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 8
2784; UNROLL-NO-IC-NEXT:    [[TMP50:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]]
2785; UNROLL-NO-IC-NEXT:    br i1 [[TMP50]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !prof [[PROF27:![0-9]+]], !llvm.loop [[LOOP28:![0-9]+]]
2786; UNROLL-NO-IC:       middle.block:
2787; UNROLL-NO-IC-NEXT:    [[BIN_RDX:%.*]] = add <4 x i32> [[TMP49]], [[TMP48]]
2788; UNROLL-NO-IC-NEXT:    [[TMP51:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[BIN_RDX]])
2789; UNROLL-NO-IC-NEXT:    [[VECTOR_RECUR_EXTRACT:%.*]] = extractelement <4 x i32> [[TMP43]], i32 3
2790; UNROLL-NO-IC-NEXT:    br i1 true, label [[BB1:%.*]], label [[SCALAR_PH]], !prof [[PROF29:![0-9]+]]
2791; UNROLL-NO-IC:       scalar.ph:
2792; UNROLL-NO-IC-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i32 [ [[IND_END]], [[MIDDLE_BLOCK]] ], [ [[Y]], [[BB:%.*]] ]
2793; UNROLL-NO-IC-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i32 [ [[VECTOR_RECUR_EXTRACT]], [[MIDDLE_BLOCK]] ], [ 0, [[BB]] ]
2794; UNROLL-NO-IC-NEXT:    [[BC_MERGE_RDX:%.*]] = phi i32 [ [[TMP51]], [[MIDDLE_BLOCK]] ], [ 0, [[BB]] ]
2795; UNROLL-NO-IC-NEXT:    br label [[BB2:%.*]]
2796; UNROLL-NO-IC:       bb1:
2797; UNROLL-NO-IC-NEXT:    [[VAR:%.*]] = phi i32 [ [[VAR6:%.*]], [[BB2]] ], [ [[TMP51]], [[MIDDLE_BLOCK]] ]
2798; UNROLL-NO-IC-NEXT:    ret i32 [[VAR]]
2799; UNROLL-NO-IC:       bb2:
2800; UNROLL-NO-IC-NEXT:    [[VAR3:%.*]] = phi i32 [ [[VAR8:%.*]], [[BB2]] ], [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ]
2801; UNROLL-NO-IC-NEXT:    [[VAR4:%.*]] = phi i32 [ [[VAR7:%.*]], [[BB2]] ], [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ]
2802; UNROLL-NO-IC-NEXT:    [[VAR5:%.*]] = phi i32 [ [[VAR6]], [[BB2]] ], [ [[BC_MERGE_RDX]], [[SCALAR_PH]] ]
2803; UNROLL-NO-IC-NEXT:    [[VAR6]] = add i32 [[VAR5]], [[VAR4]]
2804; UNROLL-NO-IC-NEXT:    [[VAR7]] = udiv i32 219220132, [[VAR3]]
2805; UNROLL-NO-IC-NEXT:    [[VAR8]] = add nsw i32 [[VAR3]], -1
2806; UNROLL-NO-IC-NEXT:    [[VAR9:%.*]] = icmp slt i32 [[VAR3]], 2
2807; UNROLL-NO-IC-NEXT:    br i1 [[VAR9]], label [[BB1]], label [[BB2]], !prof [[PROF30:![0-9]+]], !llvm.loop [[LOOP31:![0-9]+]]
2808;
2809; UNROLL-NO-VF-LABEL: @sink_into_replication_region(
2810; UNROLL-NO-VF-NEXT:  bb:
2811; UNROLL-NO-VF-NEXT:    [[TMP0:%.*]] = add i32 [[Y:%.*]], 1
2812; UNROLL-NO-VF-NEXT:    [[SMIN:%.*]] = call i32 @llvm.smin.i32(i32 [[Y]], i32 1)
2813; UNROLL-NO-VF-NEXT:    [[TMP1:%.*]] = sub i32 [[TMP0]], [[SMIN]]
2814; UNROLL-NO-VF-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]], !prof [[PROF26:![0-9]+]]
2815; UNROLL-NO-VF:       vector.ph:
2816; UNROLL-NO-VF-NEXT:    [[N_RND_UP:%.*]] = add i32 [[TMP1]], 1
2817; UNROLL-NO-VF-NEXT:    [[N_MOD_VF:%.*]] = urem i32 [[N_RND_UP]], 2
2818; UNROLL-NO-VF-NEXT:    [[N_VEC:%.*]] = sub i32 [[N_RND_UP]], [[N_MOD_VF]]
2819; UNROLL-NO-VF-NEXT:    [[TRIP_COUNT_MINUS_1:%.*]] = sub i32 [[TMP1]], 1
2820; UNROLL-NO-VF-NEXT:    [[IND_END:%.*]] = sub i32 [[Y]], [[N_VEC]]
2821; UNROLL-NO-VF-NEXT:    br label [[VECTOR_BODY:%.*]]
2822; UNROLL-NO-VF:       vector.body:
2823; UNROLL-NO-VF-NEXT:    [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[PRED_UDIV_CONTINUE4:%.*]] ]
2824; UNROLL-NO-VF-NEXT:    [[VECTOR_RECUR:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[TMP9:%.*]], [[PRED_UDIV_CONTINUE4]] ]
2825; UNROLL-NO-VF-NEXT:    [[VEC_PHI:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[TMP10:%.*]], [[PRED_UDIV_CONTINUE4]] ]
2826; UNROLL-NO-VF-NEXT:    [[VEC_PHI1:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[TMP11:%.*]], [[PRED_UDIV_CONTINUE4]] ]
2827; UNROLL-NO-VF-NEXT:    [[OFFSET_IDX:%.*]] = sub i32 [[Y]], [[INDEX]]
2828; UNROLL-NO-VF-NEXT:    [[VEC_IV:%.*]] = add i32 [[INDEX]], 0
2829; UNROLL-NO-VF-NEXT:    [[VEC_IV2:%.*]] = add i32 [[INDEX]], 1
2830; UNROLL-NO-VF-NEXT:    [[TMP2:%.*]] = icmp ule i32 [[VEC_IV]], [[TRIP_COUNT_MINUS_1]]
2831; UNROLL-NO-VF-NEXT:    [[TMP3:%.*]] = icmp ule i32 [[VEC_IV2]], [[TRIP_COUNT_MINUS_1]]
2832; UNROLL-NO-VF-NEXT:    br i1 [[TMP2]], label [[PRED_UDIV_IF:%.*]], label [[PRED_UDIV_CONTINUE:%.*]]
2833; UNROLL-NO-VF:       pred.udiv.if:
2834; UNROLL-NO-VF-NEXT:    [[TMP4:%.*]] = add i32 [[OFFSET_IDX]], 0
2835; UNROLL-NO-VF-NEXT:    [[TMP5:%.*]] = udiv i32 219220132, [[TMP4]]
2836; UNROLL-NO-VF-NEXT:    br label [[PRED_UDIV_CONTINUE]]
2837; UNROLL-NO-VF:       pred.udiv.continue:
2838; UNROLL-NO-VF-NEXT:    [[TMP6:%.*]] = phi i32 [ poison, [[VECTOR_BODY]] ], [ [[TMP5]], [[PRED_UDIV_IF]] ]
2839; UNROLL-NO-VF-NEXT:    br i1 [[TMP3]], label [[PRED_UDIV_IF3:%.*]], label [[PRED_UDIV_CONTINUE4]]
2840; UNROLL-NO-VF:       pred.udiv.if3:
2841; UNROLL-NO-VF-NEXT:    [[TMP7:%.*]] = add i32 [[OFFSET_IDX]], -1
2842; UNROLL-NO-VF-NEXT:    [[TMP8:%.*]] = udiv i32 219220132, [[TMP7]]
2843; UNROLL-NO-VF-NEXT:    br label [[PRED_UDIV_CONTINUE4]]
2844; UNROLL-NO-VF:       pred.udiv.continue4:
2845; UNROLL-NO-VF-NEXT:    [[TMP9]] = phi i32 [ poison, [[PRED_UDIV_CONTINUE]] ], [ [[TMP8]], [[PRED_UDIV_IF3]] ]
2846; UNROLL-NO-VF-NEXT:    [[TMP10]] = add i32 [[VEC_PHI]], [[VECTOR_RECUR]]
2847; UNROLL-NO-VF-NEXT:    [[TMP11]] = add i32 [[VEC_PHI1]], [[TMP6]]
2848; UNROLL-NO-VF-NEXT:    [[TMP12:%.*]] = select i1 [[TMP2]], i32 [[TMP10]], i32 [[VEC_PHI]]
2849; UNROLL-NO-VF-NEXT:    [[TMP13:%.*]] = select i1 [[TMP3]], i32 [[TMP11]], i32 [[VEC_PHI1]]
2850; UNROLL-NO-VF-NEXT:    [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 2
2851; UNROLL-NO-VF-NEXT:    [[TMP14:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]]
2852; UNROLL-NO-VF-NEXT:    br i1 [[TMP14]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !prof [[PROF27:![0-9]+]], !llvm.loop [[LOOP28:![0-9]+]]
2853; UNROLL-NO-VF:       middle.block:
2854; UNROLL-NO-VF-NEXT:    [[BIN_RDX:%.*]] = add i32 [[TMP13]], [[TMP12]]
2855; UNROLL-NO-VF-NEXT:    br i1 true, label [[BB1:%.*]], label [[SCALAR_PH]], !prof [[PROF29:![0-9]+]]
2856; UNROLL-NO-VF:       scalar.ph:
2857; UNROLL-NO-VF-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i32 [ [[IND_END]], [[MIDDLE_BLOCK]] ], [ [[Y]], [[BB:%.*]] ]
2858; UNROLL-NO-VF-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i32 [ [[TMP9]], [[MIDDLE_BLOCK]] ], [ 0, [[BB]] ]
2859; UNROLL-NO-VF-NEXT:    [[BC_MERGE_RDX:%.*]] = phi i32 [ [[BIN_RDX]], [[MIDDLE_BLOCK]] ], [ 0, [[BB]] ]
2860; UNROLL-NO-VF-NEXT:    br label [[BB2:%.*]]
2861; UNROLL-NO-VF:       bb1:
2862; UNROLL-NO-VF-NEXT:    [[VAR:%.*]] = phi i32 [ [[VAR6:%.*]], [[BB2]] ], [ [[BIN_RDX]], [[MIDDLE_BLOCK]] ]
2863; UNROLL-NO-VF-NEXT:    ret i32 [[VAR]]
2864; UNROLL-NO-VF:       bb2:
2865; UNROLL-NO-VF-NEXT:    [[VAR3:%.*]] = phi i32 [ [[VAR8:%.*]], [[BB2]] ], [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ]
2866; UNROLL-NO-VF-NEXT:    [[VAR4:%.*]] = phi i32 [ [[VAR7:%.*]], [[BB2]] ], [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ]
2867; UNROLL-NO-VF-NEXT:    [[VAR5:%.*]] = phi i32 [ [[VAR6]], [[BB2]] ], [ [[BC_MERGE_RDX]], [[SCALAR_PH]] ]
2868; UNROLL-NO-VF-NEXT:    [[VAR6]] = add i32 [[VAR5]], [[VAR4]]
2869; UNROLL-NO-VF-NEXT:    [[VAR7]] = udiv i32 219220132, [[VAR3]]
2870; UNROLL-NO-VF-NEXT:    [[VAR8]] = add nsw i32 [[VAR3]], -1
2871; UNROLL-NO-VF-NEXT:    [[VAR9:%.*]] = icmp slt i32 [[VAR3]], 2
2872; UNROLL-NO-VF-NEXT:    br i1 [[VAR9]], label [[BB1]], label [[BB2]], !prof [[PROF30:![0-9]+]], !llvm.loop [[LOOP31:![0-9]+]]
2873;
2874; SINK-AFTER-LABEL: @sink_into_replication_region(
2875; SINK-AFTER-NEXT:  bb:
2876; SINK-AFTER-NEXT:    [[TMP0:%.*]] = add i32 [[Y:%.*]], 1
2877; SINK-AFTER-NEXT:    [[SMIN:%.*]] = call i32 @llvm.smin.i32(i32 [[Y]], i32 1)
2878; SINK-AFTER-NEXT:    [[TMP1:%.*]] = sub i32 [[TMP0]], [[SMIN]]
2879; SINK-AFTER-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]], !prof [[PROF26:![0-9]+]]
2880; SINK-AFTER:       vector.ph:
2881; SINK-AFTER-NEXT:    [[N_RND_UP:%.*]] = add i32 [[TMP1]], 3
2882; SINK-AFTER-NEXT:    [[N_MOD_VF:%.*]] = urem i32 [[N_RND_UP]], 4
2883; SINK-AFTER-NEXT:    [[N_VEC:%.*]] = sub i32 [[N_RND_UP]], [[N_MOD_VF]]
2884; SINK-AFTER-NEXT:    [[TRIP_COUNT_MINUS_1:%.*]] = sub i32 [[TMP1]], 1
2885; SINK-AFTER-NEXT:    [[IND_END:%.*]] = sub i32 [[Y]], [[N_VEC]]
2886; SINK-AFTER-NEXT:    [[BROADCAST_SPLATINSERT1:%.*]] = insertelement <4 x i32> poison, i32 [[TRIP_COUNT_MINUS_1]], i64 0
2887; SINK-AFTER-NEXT:    [[BROADCAST_SPLAT2:%.*]] = shufflevector <4 x i32> [[BROADCAST_SPLATINSERT1]], <4 x i32> poison, <4 x i32> zeroinitializer
2888; SINK-AFTER-NEXT:    br label [[VECTOR_BODY:%.*]]
2889; SINK-AFTER:       vector.body:
2890; SINK-AFTER-NEXT:    [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[PRED_UDIV_CONTINUE8:%.*]] ]
2891; SINK-AFTER-NEXT:    [[VECTOR_RECUR:%.*]] = phi <4 x i32> [ <i32 poison, i32 poison, i32 poison, i32 0>, [[VECTOR_PH]] ], [ [[TMP22:%.*]], [[PRED_UDIV_CONTINUE8]] ]
2892; SINK-AFTER-NEXT:    [[VEC_PHI:%.*]] = phi <4 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP24:%.*]], [[PRED_UDIV_CONTINUE8]] ]
2893; SINK-AFTER-NEXT:    [[OFFSET_IDX:%.*]] = sub i32 [[Y]], [[INDEX]]
2894; SINK-AFTER-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <4 x i32> poison, i32 [[INDEX]], i64 0
2895; SINK-AFTER-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <4 x i32> [[BROADCAST_SPLATINSERT]], <4 x i32> poison, <4 x i32> zeroinitializer
2896; SINK-AFTER-NEXT:    [[VEC_IV:%.*]] = add <4 x i32> [[BROADCAST_SPLAT]], <i32 0, i32 1, i32 2, i32 3>
2897; SINK-AFTER-NEXT:    [[TMP2:%.*]] = icmp ule <4 x i32> [[VEC_IV]], [[BROADCAST_SPLAT2]]
2898; SINK-AFTER-NEXT:    [[TMP3:%.*]] = extractelement <4 x i1> [[TMP2]], i32 0
2899; SINK-AFTER-NEXT:    br i1 [[TMP3]], label [[PRED_UDIV_IF:%.*]], label [[PRED_UDIV_CONTINUE:%.*]]
2900; SINK-AFTER:       pred.udiv.if:
2901; SINK-AFTER-NEXT:    [[TMP4:%.*]] = add i32 [[OFFSET_IDX]], 0
2902; SINK-AFTER-NEXT:    [[TMP5:%.*]] = udiv i32 219220132, [[TMP4]]
2903; SINK-AFTER-NEXT:    [[TMP6:%.*]] = insertelement <4 x i32> poison, i32 [[TMP5]], i32 0
2904; SINK-AFTER-NEXT:    br label [[PRED_UDIV_CONTINUE]]
2905; SINK-AFTER:       pred.udiv.continue:
2906; SINK-AFTER-NEXT:    [[TMP7:%.*]] = phi <4 x i32> [ poison, [[VECTOR_BODY]] ], [ [[TMP6]], [[PRED_UDIV_IF]] ]
2907; SINK-AFTER-NEXT:    [[TMP8:%.*]] = extractelement <4 x i1> [[TMP2]], i32 1
2908; SINK-AFTER-NEXT:    br i1 [[TMP8]], label [[PRED_UDIV_IF3:%.*]], label [[PRED_UDIV_CONTINUE4:%.*]]
2909; SINK-AFTER:       pred.udiv.if3:
2910; SINK-AFTER-NEXT:    [[TMP9:%.*]] = add i32 [[OFFSET_IDX]], -1
2911; SINK-AFTER-NEXT:    [[TMP10:%.*]] = udiv i32 219220132, [[TMP9]]
2912; SINK-AFTER-NEXT:    [[TMP11:%.*]] = insertelement <4 x i32> [[TMP7]], i32 [[TMP10]], i32 1
2913; SINK-AFTER-NEXT:    br label [[PRED_UDIV_CONTINUE4]]
2914; SINK-AFTER:       pred.udiv.continue4:
2915; SINK-AFTER-NEXT:    [[TMP12:%.*]] = phi <4 x i32> [ [[TMP7]], [[PRED_UDIV_CONTINUE]] ], [ [[TMP11]], [[PRED_UDIV_IF3]] ]
2916; SINK-AFTER-NEXT:    [[TMP13:%.*]] = extractelement <4 x i1> [[TMP2]], i32 2
2917; SINK-AFTER-NEXT:    br i1 [[TMP13]], label [[PRED_UDIV_IF5:%.*]], label [[PRED_UDIV_CONTINUE6:%.*]]
2918; SINK-AFTER:       pred.udiv.if5:
2919; SINK-AFTER-NEXT:    [[TMP14:%.*]] = add i32 [[OFFSET_IDX]], -2
2920; SINK-AFTER-NEXT:    [[TMP15:%.*]] = udiv i32 219220132, [[TMP14]]
2921; SINK-AFTER-NEXT:    [[TMP16:%.*]] = insertelement <4 x i32> [[TMP12]], i32 [[TMP15]], i32 2
2922; SINK-AFTER-NEXT:    br label [[PRED_UDIV_CONTINUE6]]
2923; SINK-AFTER:       pred.udiv.continue6:
2924; SINK-AFTER-NEXT:    [[TMP17:%.*]] = phi <4 x i32> [ [[TMP12]], [[PRED_UDIV_CONTINUE4]] ], [ [[TMP16]], [[PRED_UDIV_IF5]] ]
2925; SINK-AFTER-NEXT:    [[TMP18:%.*]] = extractelement <4 x i1> [[TMP2]], i32 3
2926; SINK-AFTER-NEXT:    br i1 [[TMP18]], label [[PRED_UDIV_IF7:%.*]], label [[PRED_UDIV_CONTINUE8]]
2927; SINK-AFTER:       pred.udiv.if7:
2928; SINK-AFTER-NEXT:    [[TMP19:%.*]] = add i32 [[OFFSET_IDX]], -3
2929; SINK-AFTER-NEXT:    [[TMP20:%.*]] = udiv i32 219220132, [[TMP19]]
2930; SINK-AFTER-NEXT:    [[TMP21:%.*]] = insertelement <4 x i32> [[TMP17]], i32 [[TMP20]], i32 3
2931; SINK-AFTER-NEXT:    br label [[PRED_UDIV_CONTINUE8]]
2932; SINK-AFTER:       pred.udiv.continue8:
2933; SINK-AFTER-NEXT:    [[TMP22]] = phi <4 x i32> [ [[TMP17]], [[PRED_UDIV_CONTINUE6]] ], [ [[TMP21]], [[PRED_UDIV_IF7]] ]
2934; SINK-AFTER-NEXT:    [[TMP23:%.*]] = shufflevector <4 x i32> [[VECTOR_RECUR]], <4 x i32> [[TMP22]], <4 x i32> <i32 3, i32 4, i32 5, i32 6>
2935; SINK-AFTER-NEXT:    [[TMP24]] = add <4 x i32> [[VEC_PHI]], [[TMP23]]
2936; SINK-AFTER-NEXT:    [[TMP25:%.*]] = select <4 x i1> [[TMP2]], <4 x i32> [[TMP24]], <4 x i32> [[VEC_PHI]]
2937; SINK-AFTER-NEXT:    [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 4
2938; SINK-AFTER-NEXT:    [[TMP26:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]]
2939; SINK-AFTER-NEXT:    br i1 [[TMP26]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !prof [[PROF27:![0-9]+]], !llvm.loop [[LOOP28:![0-9]+]]
2940; SINK-AFTER:       middle.block:
2941; SINK-AFTER-NEXT:    [[TMP27:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[TMP25]])
2942; SINK-AFTER-NEXT:    [[VECTOR_RECUR_EXTRACT:%.*]] = extractelement <4 x i32> [[TMP22]], i32 3
2943; SINK-AFTER-NEXT:    br i1 true, label [[BB1:%.*]], label [[SCALAR_PH]], !prof [[PROF29:![0-9]+]]
2944; SINK-AFTER:       scalar.ph:
2945; SINK-AFTER-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i32 [ [[IND_END]], [[MIDDLE_BLOCK]] ], [ [[Y]], [[BB:%.*]] ]
2946; SINK-AFTER-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i32 [ [[VECTOR_RECUR_EXTRACT]], [[MIDDLE_BLOCK]] ], [ 0, [[BB]] ]
2947; SINK-AFTER-NEXT:    [[BC_MERGE_RDX:%.*]] = phi i32 [ [[TMP27]], [[MIDDLE_BLOCK]] ], [ 0, [[BB]] ]
2948; SINK-AFTER-NEXT:    br label [[BB2:%.*]]
2949; SINK-AFTER:       bb1:
2950; SINK-AFTER-NEXT:    [[VAR:%.*]] = phi i32 [ [[VAR6:%.*]], [[BB2]] ], [ [[TMP27]], [[MIDDLE_BLOCK]] ]
2951; SINK-AFTER-NEXT:    ret i32 [[VAR]]
2952; SINK-AFTER:       bb2:
2953; SINK-AFTER-NEXT:    [[VAR3:%.*]] = phi i32 [ [[VAR8:%.*]], [[BB2]] ], [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ]
2954; SINK-AFTER-NEXT:    [[VAR4:%.*]] = phi i32 [ [[VAR7:%.*]], [[BB2]] ], [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ]
2955; SINK-AFTER-NEXT:    [[VAR5:%.*]] = phi i32 [ [[VAR6]], [[BB2]] ], [ [[BC_MERGE_RDX]], [[SCALAR_PH]] ]
2956; SINK-AFTER-NEXT:    [[VAR6]] = add i32 [[VAR5]], [[VAR4]]
2957; SINK-AFTER-NEXT:    [[VAR7]] = udiv i32 219220132, [[VAR3]]
2958; SINK-AFTER-NEXT:    [[VAR8]] = add nsw i32 [[VAR3]], -1
2959; SINK-AFTER-NEXT:    [[VAR9:%.*]] = icmp slt i32 [[VAR3]], 2
2960; SINK-AFTER-NEXT:    br i1 [[VAR9]], label [[BB1]], label [[BB2]], !prof [[PROF30:![0-9]+]], !llvm.loop [[LOOP31:![0-9]+]]
2961;
2962bb:
2963  br label %bb2
2964
2965  bb1:                                              ; preds = %bb2
2966  %var = phi i32 [ %var6, %bb2 ]
2967  ret i32 %var
2968
2969  bb2:                                              ; preds = %bb2, %bb
2970  %var3 = phi i32 [ %var8, %bb2 ], [ %y, %bb ]
2971  %var4 = phi i32 [ %var7, %bb2 ], [ 0, %bb ]
2972  %var5 = phi i32 [ %var6, %bb2 ], [ 0, %bb ]
2973  %var6 = add i32 %var5, %var4
2974  %var7 = udiv i32 219220132, %var3
2975  %var8 = add nsw i32 %var3, -1
2976  %var9 = icmp slt i32 %var3, 2
2977  br i1 %var9, label %bb1, label %bb2, !prof !2
2978}
2979
2980define i32 @sink_into_replication_region_multiple(ptr %x, i32 %y) {
2981;
2982; CHECK-LABEL: @sink_into_replication_region_multiple(
2983; UNROLL-NO-IC-LABEL: @sink_into_replication_region_multiple(
2984; UNROLL-NO-IC-NEXT:  bb:
2985; UNROLL-NO-IC-NEXT:    [[TMP0:%.*]] = add i32 [[Y:%.*]], 1
2986; UNROLL-NO-IC-NEXT:    [[SMIN:%.*]] = call i32 @llvm.smin.i32(i32 [[Y]], i32 1)
2987; UNROLL-NO-IC-NEXT:    [[TMP1:%.*]] = sub i32 [[TMP0]], [[SMIN]]
2988; UNROLL-NO-IC-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]], !prof [[PROF26]]
2989; UNROLL-NO-IC:       vector.ph:
2990; UNROLL-NO-IC-NEXT:    [[N_RND_UP:%.*]] = add i32 [[TMP1]], 7
2991; UNROLL-NO-IC-NEXT:    [[N_MOD_VF:%.*]] = urem i32 [[N_RND_UP]], 8
2992; UNROLL-NO-IC-NEXT:    [[N_VEC:%.*]] = sub i32 [[N_RND_UP]], [[N_MOD_VF]]
2993; UNROLL-NO-IC-NEXT:    [[TRIP_COUNT_MINUS_1:%.*]] = sub i32 [[TMP1]], 1
2994; UNROLL-NO-IC-NEXT:    [[IND_END:%.*]] = sub i32 [[Y]], [[N_VEC]]
2995; UNROLL-NO-IC-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <4 x i32> poison, i32 [[TRIP_COUNT_MINUS_1]], i64 0
2996; UNROLL-NO-IC-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <4 x i32> [[BROADCAST_SPLATINSERT]], <4 x i32> poison, <4 x i32> zeroinitializer
2997; UNROLL-NO-IC-NEXT:    br label [[VECTOR_BODY:%.*]]
2998; UNROLL-NO-IC:       vector.body:
2999; UNROLL-NO-IC-NEXT:    [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[PRED_STORE_CONTINUE29:%.*]] ]
3000; UNROLL-NO-IC-NEXT:    [[VEC_IND:%.*]] = phi <4 x i32> [ <i32 0, i32 1, i32 2, i32 3>, [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[PRED_STORE_CONTINUE29]] ]
3001; UNROLL-NO-IC-NEXT:    [[VECTOR_RECUR:%.*]] = phi <4 x i32> [ <i32 poison, i32 poison, i32 poison, i32 0>, [[VECTOR_PH]] ], [ [[TMP43:%.*]], [[PRED_STORE_CONTINUE29]] ]
3002; UNROLL-NO-IC-NEXT:    [[VEC_PHI:%.*]] = phi <4 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP46:%.*]], [[PRED_STORE_CONTINUE29]] ]
3003; UNROLL-NO-IC-NEXT:    [[VEC_PHI1:%.*]] = phi <4 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP47:%.*]], [[PRED_STORE_CONTINUE29]] ]
3004; UNROLL-NO-IC-NEXT:    [[STEP_ADD:%.*]] = add <4 x i32> [[VEC_IND]], splat (i32 4)
3005; UNROLL-NO-IC-NEXT:    [[OFFSET_IDX:%.*]] = sub i32 [[Y]], [[INDEX]]
3006; UNROLL-NO-IC-NEXT:    [[TMP2:%.*]] = add i32 [[OFFSET_IDX]], 0
3007; UNROLL-NO-IC-NEXT:    [[TMP3:%.*]] = add i32 [[OFFSET_IDX]], -1
3008; UNROLL-NO-IC-NEXT:    [[TMP4:%.*]] = add i32 [[OFFSET_IDX]], -2
3009; UNROLL-NO-IC-NEXT:    [[TMP5:%.*]] = add i32 [[OFFSET_IDX]], -3
3010; UNROLL-NO-IC-NEXT:    [[TMP6:%.*]] = add i32 [[OFFSET_IDX]], -4
3011; UNROLL-NO-IC-NEXT:    [[TMP7:%.*]] = add i32 [[OFFSET_IDX]], -5
3012; UNROLL-NO-IC-NEXT:    [[TMP8:%.*]] = add i32 [[OFFSET_IDX]], -6
3013; UNROLL-NO-IC-NEXT:    [[TMP9:%.*]] = add i32 [[OFFSET_IDX]], -7
3014; UNROLL-NO-IC-NEXT:    [[TMP10:%.*]] = icmp ule <4 x i32> [[VEC_IND]], [[BROADCAST_SPLAT]]
3015; UNROLL-NO-IC-NEXT:    [[TMP11:%.*]] = icmp ule <4 x i32> [[STEP_ADD]], [[BROADCAST_SPLAT]]
3016; UNROLL-NO-IC-NEXT:    [[TMP12:%.*]] = extractelement <4 x i1> [[TMP10]], i32 0
3017; UNROLL-NO-IC-NEXT:    br i1 [[TMP12]], label [[PRED_UDIV_IF:%.*]], label [[PRED_UDIV_CONTINUE:%.*]]
3018; UNROLL-NO-IC:       pred.udiv.if:
3019; UNROLL-NO-IC-NEXT:    [[TMP13:%.*]] = udiv i32 219220132, [[TMP2]]
3020; UNROLL-NO-IC-NEXT:    [[TMP14:%.*]] = insertelement <4 x i32> poison, i32 [[TMP13]], i32 0
3021; UNROLL-NO-IC-NEXT:    br label [[PRED_UDIV_CONTINUE]]
3022; UNROLL-NO-IC:       pred.udiv.continue:
3023; UNROLL-NO-IC-NEXT:    [[TMP15:%.*]] = phi <4 x i32> [ poison, [[VECTOR_BODY]] ], [ [[TMP14]], [[PRED_UDIV_IF]] ]
3024; UNROLL-NO-IC-NEXT:    [[TMP16:%.*]] = extractelement <4 x i1> [[TMP10]], i32 1
3025; UNROLL-NO-IC-NEXT:    br i1 [[TMP16]], label [[PRED_UDIV_IF2:%.*]], label [[PRED_UDIV_CONTINUE3:%.*]]
3026; UNROLL-NO-IC:       pred.udiv.if2:
3027; UNROLL-NO-IC-NEXT:    [[TMP17:%.*]] = udiv i32 219220132, [[TMP3]]
3028; UNROLL-NO-IC-NEXT:    [[TMP18:%.*]] = insertelement <4 x i32> [[TMP15]], i32 [[TMP17]], i32 1
3029; UNROLL-NO-IC-NEXT:    br label [[PRED_UDIV_CONTINUE3]]
3030; UNROLL-NO-IC:       pred.udiv.continue3:
3031; UNROLL-NO-IC-NEXT:    [[TMP19:%.*]] = phi <4 x i32> [ [[TMP15]], [[PRED_UDIV_CONTINUE]] ], [ [[TMP18]], [[PRED_UDIV_IF2]] ]
3032; UNROLL-NO-IC-NEXT:    [[TMP20:%.*]] = extractelement <4 x i1> [[TMP10]], i32 2
3033; UNROLL-NO-IC-NEXT:    br i1 [[TMP20]], label [[PRED_UDIV_IF4:%.*]], label [[PRED_UDIV_CONTINUE5:%.*]]
3034; UNROLL-NO-IC:       pred.udiv.if4:
3035; UNROLL-NO-IC-NEXT:    [[TMP21:%.*]] = udiv i32 219220132, [[TMP4]]
3036; UNROLL-NO-IC-NEXT:    [[TMP22:%.*]] = insertelement <4 x i32> [[TMP19]], i32 [[TMP21]], i32 2
3037; UNROLL-NO-IC-NEXT:    br label [[PRED_UDIV_CONTINUE5]]
3038; UNROLL-NO-IC:       pred.udiv.continue5:
3039; UNROLL-NO-IC-NEXT:    [[TMP23:%.*]] = phi <4 x i32> [ [[TMP19]], [[PRED_UDIV_CONTINUE3]] ], [ [[TMP22]], [[PRED_UDIV_IF4]] ]
3040; UNROLL-NO-IC-NEXT:    [[TMP24:%.*]] = extractelement <4 x i1> [[TMP10]], i32 3
3041; UNROLL-NO-IC-NEXT:    br i1 [[TMP24]], label [[PRED_UDIV_IF6:%.*]], label [[PRED_UDIV_CONTINUE7:%.*]]
3042; UNROLL-NO-IC:       pred.udiv.if6:
3043; UNROLL-NO-IC-NEXT:    [[TMP25:%.*]] = udiv i32 219220132, [[TMP5]]
3044; UNROLL-NO-IC-NEXT:    [[TMP26:%.*]] = insertelement <4 x i32> [[TMP23]], i32 [[TMP25]], i32 3
3045; UNROLL-NO-IC-NEXT:    br label [[PRED_UDIV_CONTINUE7]]
3046; UNROLL-NO-IC:       pred.udiv.continue7:
3047; UNROLL-NO-IC-NEXT:    [[TMP27:%.*]] = phi <4 x i32> [ [[TMP23]], [[PRED_UDIV_CONTINUE5]] ], [ [[TMP26]], [[PRED_UDIV_IF6]] ]
3048; UNROLL-NO-IC-NEXT:    [[TMP28:%.*]] = extractelement <4 x i1> [[TMP11]], i32 0
3049; UNROLL-NO-IC-NEXT:    br i1 [[TMP28]], label [[PRED_UDIV_IF8:%.*]], label [[PRED_UDIV_CONTINUE9:%.*]]
3050; UNROLL-NO-IC:       pred.udiv.if8:
3051; UNROLL-NO-IC-NEXT:    [[TMP29:%.*]] = udiv i32 219220132, [[TMP6]]
3052; UNROLL-NO-IC-NEXT:    [[TMP30:%.*]] = insertelement <4 x i32> poison, i32 [[TMP29]], i32 0
3053; UNROLL-NO-IC-NEXT:    br label [[PRED_UDIV_CONTINUE9]]
3054; UNROLL-NO-IC:       pred.udiv.continue9:
3055; UNROLL-NO-IC-NEXT:    [[TMP31:%.*]] = phi <4 x i32> [ poison, [[PRED_UDIV_CONTINUE7]] ], [ [[TMP30]], [[PRED_UDIV_IF8]] ]
3056; UNROLL-NO-IC-NEXT:    [[TMP32:%.*]] = extractelement <4 x i1> [[TMP11]], i32 1
3057; UNROLL-NO-IC-NEXT:    br i1 [[TMP32]], label [[PRED_UDIV_IF10:%.*]], label [[PRED_UDIV_CONTINUE11:%.*]]
3058; UNROLL-NO-IC:       pred.udiv.if10:
3059; UNROLL-NO-IC-NEXT:    [[TMP33:%.*]] = udiv i32 219220132, [[TMP7]]
3060; UNROLL-NO-IC-NEXT:    [[TMP34:%.*]] = insertelement <4 x i32> [[TMP31]], i32 [[TMP33]], i32 1
3061; UNROLL-NO-IC-NEXT:    br label [[PRED_UDIV_CONTINUE11]]
3062; UNROLL-NO-IC:       pred.udiv.continue11:
3063; UNROLL-NO-IC-NEXT:    [[TMP35:%.*]] = phi <4 x i32> [ [[TMP31]], [[PRED_UDIV_CONTINUE9]] ], [ [[TMP34]], [[PRED_UDIV_IF10]] ]
3064; UNROLL-NO-IC-NEXT:    [[TMP36:%.*]] = extractelement <4 x i1> [[TMP11]], i32 2
3065; UNROLL-NO-IC-NEXT:    br i1 [[TMP36]], label [[PRED_UDIV_IF12:%.*]], label [[PRED_UDIV_CONTINUE13:%.*]]
3066; UNROLL-NO-IC:       pred.udiv.if12:
3067; UNROLL-NO-IC-NEXT:    [[TMP37:%.*]] = udiv i32 219220132, [[TMP8]]
3068; UNROLL-NO-IC-NEXT:    [[TMP38:%.*]] = insertelement <4 x i32> [[TMP35]], i32 [[TMP37]], i32 2
3069; UNROLL-NO-IC-NEXT:    br label [[PRED_UDIV_CONTINUE13]]
3070; UNROLL-NO-IC:       pred.udiv.continue13:
3071; UNROLL-NO-IC-NEXT:    [[TMP39:%.*]] = phi <4 x i32> [ [[TMP35]], [[PRED_UDIV_CONTINUE11]] ], [ [[TMP38]], [[PRED_UDIV_IF12]] ]
3072; UNROLL-NO-IC-NEXT:    [[TMP40:%.*]] = extractelement <4 x i1> [[TMP11]], i32 3
3073; UNROLL-NO-IC-NEXT:    br i1 [[TMP40]], label [[PRED_UDIV_IF14:%.*]], label [[PRED_UDIV_CONTINUE15:%.*]]
3074; UNROLL-NO-IC:       pred.udiv.if14:
3075; UNROLL-NO-IC-NEXT:    [[TMP41:%.*]] = udiv i32 219220132, [[TMP9]]
3076; UNROLL-NO-IC-NEXT:    [[TMP42:%.*]] = insertelement <4 x i32> [[TMP39]], i32 [[TMP41]], i32 3
3077; UNROLL-NO-IC-NEXT:    br label [[PRED_UDIV_CONTINUE15]]
3078; UNROLL-NO-IC:       pred.udiv.continue15:
3079; UNROLL-NO-IC-NEXT:    [[TMP43]] = phi <4 x i32> [ [[TMP39]], [[PRED_UDIV_CONTINUE13]] ], [ [[TMP42]], [[PRED_UDIV_IF14]] ]
3080; UNROLL-NO-IC-NEXT:    [[TMP44:%.*]] = shufflevector <4 x i32> [[VECTOR_RECUR]], <4 x i32> [[TMP27]], <4 x i32> <i32 3, i32 4, i32 5, i32 6>
3081; UNROLL-NO-IC-NEXT:    [[TMP45:%.*]] = shufflevector <4 x i32> [[TMP27]], <4 x i32> [[TMP43]], <4 x i32> <i32 3, i32 4, i32 5, i32 6>
3082; UNROLL-NO-IC-NEXT:    [[TMP46]] = add <4 x i32> [[VEC_PHI]], [[TMP44]]
3083; UNROLL-NO-IC-NEXT:    [[TMP47]] = add <4 x i32> [[VEC_PHI1]], [[TMP45]]
3084; UNROLL-NO-IC-NEXT:    [[TMP48:%.*]] = extractelement <4 x i1> [[TMP10]], i32 0
3085; UNROLL-NO-IC-NEXT:    br i1 [[TMP48]], label [[PRED_STORE_IF:%.*]], label [[PRED_STORE_CONTINUE:%.*]]
3086; UNROLL-NO-IC:       pred.store.if:
3087; UNROLL-NO-IC-NEXT:    [[TMP49:%.*]] = add i32 [[INDEX]], 0
3088; UNROLL-NO-IC-NEXT:    [[TMP50:%.*]] = getelementptr inbounds i32, ptr [[X:%.*]], i32 [[TMP49]]
3089; UNROLL-NO-IC-NEXT:    store i32 [[TMP2]], ptr [[TMP50]], align 4
3090; UNROLL-NO-IC-NEXT:    br label [[PRED_STORE_CONTINUE]]
3091; UNROLL-NO-IC:       pred.store.continue:
3092; UNROLL-NO-IC-NEXT:    [[TMP51:%.*]] = extractelement <4 x i1> [[TMP10]], i32 1
3093; UNROLL-NO-IC-NEXT:    br i1 [[TMP51]], label [[PRED_STORE_IF16:%.*]], label [[PRED_STORE_CONTINUE17:%.*]]
3094; UNROLL-NO-IC:       pred.store.if16:
3095; UNROLL-NO-IC-NEXT:    [[TMP52:%.*]] = add i32 [[INDEX]], 1
3096; UNROLL-NO-IC-NEXT:    [[TMP53:%.*]] = getelementptr inbounds i32, ptr [[X]], i32 [[TMP52]]
3097; UNROLL-NO-IC-NEXT:    store i32 [[TMP3]], ptr [[TMP53]], align 4
3098; UNROLL-NO-IC-NEXT:    br label [[PRED_STORE_CONTINUE17]]
3099; UNROLL-NO-IC:       pred.store.continue17:
3100; UNROLL-NO-IC-NEXT:    [[TMP54:%.*]] = extractelement <4 x i1> [[TMP10]], i32 2
3101; UNROLL-NO-IC-NEXT:    br i1 [[TMP54]], label [[PRED_STORE_IF18:%.*]], label [[PRED_STORE_CONTINUE19:%.*]]
3102; UNROLL-NO-IC:       pred.store.if18:
3103; UNROLL-NO-IC-NEXT:    [[TMP55:%.*]] = add i32 [[INDEX]], 2
3104; UNROLL-NO-IC-NEXT:    [[TMP56:%.*]] = getelementptr inbounds i32, ptr [[X]], i32 [[TMP55]]
3105; UNROLL-NO-IC-NEXT:    store i32 [[TMP4]], ptr [[TMP56]], align 4
3106; UNROLL-NO-IC-NEXT:    br label [[PRED_STORE_CONTINUE19]]
3107; UNROLL-NO-IC:       pred.store.continue19:
3108; UNROLL-NO-IC-NEXT:    [[TMP57:%.*]] = extractelement <4 x i1> [[TMP10]], i32 3
3109; UNROLL-NO-IC-NEXT:    br i1 [[TMP57]], label [[PRED_STORE_IF20:%.*]], label [[PRED_STORE_CONTINUE21:%.*]]
3110; UNROLL-NO-IC:       pred.store.if20:
3111; UNROLL-NO-IC-NEXT:    [[TMP58:%.*]] = add i32 [[INDEX]], 3
3112; UNROLL-NO-IC-NEXT:    [[TMP59:%.*]] = getelementptr inbounds i32, ptr [[X]], i32 [[TMP58]]
3113; UNROLL-NO-IC-NEXT:    store i32 [[TMP5]], ptr [[TMP59]], align 4
3114; UNROLL-NO-IC-NEXT:    br label [[PRED_STORE_CONTINUE21]]
3115; UNROLL-NO-IC:       pred.store.continue21:
3116; UNROLL-NO-IC-NEXT:    [[TMP60:%.*]] = extractelement <4 x i1> [[TMP11]], i32 0
3117; UNROLL-NO-IC-NEXT:    br i1 [[TMP60]], label [[PRED_STORE_IF22:%.*]], label [[PRED_STORE_CONTINUE23:%.*]]
3118; UNROLL-NO-IC:       pred.store.if22:
3119; UNROLL-NO-IC-NEXT:    [[TMP61:%.*]] = add i32 [[INDEX]], 4
3120; UNROLL-NO-IC-NEXT:    [[TMP62:%.*]] = getelementptr inbounds i32, ptr [[X]], i32 [[TMP61]]
3121; UNROLL-NO-IC-NEXT:    store i32 [[TMP6]], ptr [[TMP62]], align 4
3122; UNROLL-NO-IC-NEXT:    br label [[PRED_STORE_CONTINUE23]]
3123; UNROLL-NO-IC:       pred.store.continue23:
3124; UNROLL-NO-IC-NEXT:    [[TMP63:%.*]] = extractelement <4 x i1> [[TMP11]], i32 1
3125; UNROLL-NO-IC-NEXT:    br i1 [[TMP63]], label [[PRED_STORE_IF24:%.*]], label [[PRED_STORE_CONTINUE25:%.*]]
3126; UNROLL-NO-IC:       pred.store.if24:
3127; UNROLL-NO-IC-NEXT:    [[TMP64:%.*]] = add i32 [[INDEX]], 5
3128; UNROLL-NO-IC-NEXT:    [[TMP65:%.*]] = getelementptr inbounds i32, ptr [[X]], i32 [[TMP64]]
3129; UNROLL-NO-IC-NEXT:    store i32 [[TMP7]], ptr [[TMP65]], align 4
3130; UNROLL-NO-IC-NEXT:    br label [[PRED_STORE_CONTINUE25]]
3131; UNROLL-NO-IC:       pred.store.continue25:
3132; UNROLL-NO-IC-NEXT:    [[TMP66:%.*]] = extractelement <4 x i1> [[TMP11]], i32 2
3133; UNROLL-NO-IC-NEXT:    br i1 [[TMP66]], label [[PRED_STORE_IF26:%.*]], label [[PRED_STORE_CONTINUE27:%.*]]
3134; UNROLL-NO-IC:       pred.store.if26:
3135; UNROLL-NO-IC-NEXT:    [[TMP67:%.*]] = add i32 [[INDEX]], 6
3136; UNROLL-NO-IC-NEXT:    [[TMP68:%.*]] = getelementptr inbounds i32, ptr [[X]], i32 [[TMP67]]
3137; UNROLL-NO-IC-NEXT:    store i32 [[TMP8]], ptr [[TMP68]], align 4
3138; UNROLL-NO-IC-NEXT:    br label [[PRED_STORE_CONTINUE27]]
3139; UNROLL-NO-IC:       pred.store.continue27:
3140; UNROLL-NO-IC-NEXT:    [[TMP69:%.*]] = extractelement <4 x i1> [[TMP11]], i32 3
3141; UNROLL-NO-IC-NEXT:    br i1 [[TMP69]], label [[PRED_STORE_IF28:%.*]], label [[PRED_STORE_CONTINUE29]]
3142; UNROLL-NO-IC:       pred.store.if28:
3143; UNROLL-NO-IC-NEXT:    [[TMP70:%.*]] = add i32 [[INDEX]], 7
3144; UNROLL-NO-IC-NEXT:    [[TMP71:%.*]] = getelementptr inbounds i32, ptr [[X]], i32 [[TMP70]]
3145; UNROLL-NO-IC-NEXT:    store i32 [[TMP9]], ptr [[TMP71]], align 4
3146; UNROLL-NO-IC-NEXT:    br label [[PRED_STORE_CONTINUE29]]
3147; UNROLL-NO-IC:       pred.store.continue29:
3148; UNROLL-NO-IC-NEXT:    [[TMP72:%.*]] = select <4 x i1> [[TMP10]], <4 x i32> [[TMP46]], <4 x i32> [[VEC_PHI]]
3149; UNROLL-NO-IC-NEXT:    [[TMP73:%.*]] = select <4 x i1> [[TMP11]], <4 x i32> [[TMP47]], <4 x i32> [[VEC_PHI1]]
3150; UNROLL-NO-IC-NEXT:    [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 8
3151; UNROLL-NO-IC-NEXT:    [[VEC_IND_NEXT]] = add <4 x i32> [[STEP_ADD]], splat (i32 4)
3152; UNROLL-NO-IC-NEXT:    [[TMP74:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]]
3153; UNROLL-NO-IC-NEXT:    br i1 [[TMP74]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !prof [[PROF27]], !llvm.loop [[LOOP32:![0-9]+]]
3154; UNROLL-NO-IC:       middle.block:
3155; UNROLL-NO-IC-NEXT:    [[BIN_RDX:%.*]] = add <4 x i32> [[TMP73]], [[TMP72]]
3156; UNROLL-NO-IC-NEXT:    [[TMP75:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[BIN_RDX]])
3157; UNROLL-NO-IC-NEXT:    [[VECTOR_RECUR_EXTRACT:%.*]] = extractelement <4 x i32> [[TMP43]], i32 3
3158; UNROLL-NO-IC-NEXT:    br i1 true, label [[BB1:%.*]], label [[SCALAR_PH]], !prof [[PROF29]]
3159; UNROLL-NO-IC:       scalar.ph:
3160; UNROLL-NO-IC-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i32 [ [[IND_END]], [[MIDDLE_BLOCK]] ], [ [[Y]], [[BB:%.*]] ]
3161; UNROLL-NO-IC-NEXT:    [[BC_RESUME_VAL1:%.*]] = phi i32 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[BB]] ]
3162; UNROLL-NO-IC-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i32 [ [[VECTOR_RECUR_EXTRACT]], [[MIDDLE_BLOCK]] ], [ 0, [[BB]] ]
3163; UNROLL-NO-IC-NEXT:    [[BC_MERGE_RDX:%.*]] = phi i32 [ [[TMP75]], [[MIDDLE_BLOCK]] ], [ 0, [[BB]] ]
3164; UNROLL-NO-IC-NEXT:    br label [[BB2:%.*]]
3165; UNROLL-NO-IC:       bb1:
3166; UNROLL-NO-IC-NEXT:    [[VAR:%.*]] = phi i32 [ [[VAR6:%.*]], [[BB2]] ], [ [[TMP75]], [[MIDDLE_BLOCK]] ]
3167; UNROLL-NO-IC-NEXT:    ret i32 [[VAR]]
3168; UNROLL-NO-IC:       bb2:
3169; UNROLL-NO-IC-NEXT:    [[VAR3:%.*]] = phi i32 [ [[VAR8:%.*]], [[BB2]] ], [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ]
3170; UNROLL-NO-IC-NEXT:    [[IV:%.*]] = phi i32 [ [[IV_NEXT:%.*]], [[BB2]] ], [ [[BC_RESUME_VAL1]], [[SCALAR_PH]] ]
3171; UNROLL-NO-IC-NEXT:    [[VAR4:%.*]] = phi i32 [ [[VAR7:%.*]], [[BB2]] ], [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ]
3172; UNROLL-NO-IC-NEXT:    [[VAR5:%.*]] = phi i32 [ [[VAR6]], [[BB2]] ], [ [[BC_MERGE_RDX]], [[SCALAR_PH]] ]
3173; UNROLL-NO-IC-NEXT:    [[G:%.*]] = getelementptr inbounds i32, ptr [[X]], i32 [[IV]]
3174; UNROLL-NO-IC-NEXT:    [[VAR6]] = add i32 [[VAR5]], [[VAR4]]
3175; UNROLL-NO-IC-NEXT:    [[VAR7]] = udiv i32 219220132, [[VAR3]]
3176; UNROLL-NO-IC-NEXT:    store i32 [[VAR3]], ptr [[G]], align 4
3177; UNROLL-NO-IC-NEXT:    [[VAR8]] = add nsw i32 [[VAR3]], -1
3178; UNROLL-NO-IC-NEXT:    [[IV_NEXT]] = add nsw i32 [[IV]], 1
3179; UNROLL-NO-IC-NEXT:    [[VAR9:%.*]] = icmp slt i32 [[VAR3]], 2
3180; UNROLL-NO-IC-NEXT:    br i1 [[VAR9]], label [[BB1]], label [[BB2]], !prof [[PROF30]], !llvm.loop [[LOOP33:![0-9]+]]
3181;
3182; UNROLL-NO-VF-LABEL: @sink_into_replication_region_multiple(
3183; UNROLL-NO-VF-NEXT:  bb:
3184; UNROLL-NO-VF-NEXT:    [[TMP0:%.*]] = add i32 [[Y:%.*]], 1
3185; UNROLL-NO-VF-NEXT:    [[SMIN:%.*]] = call i32 @llvm.smin.i32(i32 [[Y]], i32 1)
3186; UNROLL-NO-VF-NEXT:    [[TMP1:%.*]] = sub i32 [[TMP0]], [[SMIN]]
3187; UNROLL-NO-VF-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]], !prof [[PROF26]]
3188; UNROLL-NO-VF:       vector.ph:
3189; UNROLL-NO-VF-NEXT:    [[N_RND_UP:%.*]] = add i32 [[TMP1]], 1
3190; UNROLL-NO-VF-NEXT:    [[N_MOD_VF:%.*]] = urem i32 [[N_RND_UP]], 2
3191; UNROLL-NO-VF-NEXT:    [[N_VEC:%.*]] = sub i32 [[N_RND_UP]], [[N_MOD_VF]]
3192; UNROLL-NO-VF-NEXT:    [[TRIP_COUNT_MINUS_1:%.*]] = sub i32 [[TMP1]], 1
3193; UNROLL-NO-VF-NEXT:    [[IND_END:%.*]] = sub i32 [[Y]], [[N_VEC]]
3194; UNROLL-NO-VF-NEXT:    br label [[VECTOR_BODY:%.*]]
3195; UNROLL-NO-VF:       vector.body:
3196; UNROLL-NO-VF-NEXT:    [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[PRED_STORE_CONTINUE6:%.*]] ]
3197; UNROLL-NO-VF-NEXT:    [[VECTOR_RECUR:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[TMP9:%.*]], [[PRED_STORE_CONTINUE6]] ]
3198; UNROLL-NO-VF-NEXT:    [[VEC_PHI:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[TMP10:%.*]], [[PRED_STORE_CONTINUE6]] ]
3199; UNROLL-NO-VF-NEXT:    [[VEC_PHI1:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[TMP11:%.*]], [[PRED_STORE_CONTINUE6]] ]
3200; UNROLL-NO-VF-NEXT:    [[OFFSET_IDX:%.*]] = sub i32 [[Y]], [[INDEX]]
3201; UNROLL-NO-VF-NEXT:    [[TMP2:%.*]] = add i32 [[OFFSET_IDX]], 0
3202; UNROLL-NO-VF-NEXT:    [[TMP3:%.*]] = add i32 [[OFFSET_IDX]], -1
3203; UNROLL-NO-VF-NEXT:    [[VEC_IV:%.*]] = add i32 [[INDEX]], 0
3204; UNROLL-NO-VF-NEXT:    [[VEC_IV2:%.*]] = add i32 [[INDEX]], 1
3205; UNROLL-NO-VF-NEXT:    [[TMP4:%.*]] = icmp ule i32 [[VEC_IV]], [[TRIP_COUNT_MINUS_1]]
3206; UNROLL-NO-VF-NEXT:    [[TMP5:%.*]] = icmp ule i32 [[VEC_IV2]], [[TRIP_COUNT_MINUS_1]]
3207; UNROLL-NO-VF-NEXT:    br i1 [[TMP4]], label [[PRED_UDIV_IF:%.*]], label [[PRED_UDIV_CONTINUE:%.*]]
3208; UNROLL-NO-VF:       pred.udiv.if:
3209; UNROLL-NO-VF-NEXT:    [[TMP6:%.*]] = udiv i32 219220132, [[TMP2]]
3210; UNROLL-NO-VF-NEXT:    br label [[PRED_UDIV_CONTINUE]]
3211; UNROLL-NO-VF:       pred.udiv.continue:
3212; UNROLL-NO-VF-NEXT:    [[TMP7:%.*]] = phi i32 [ poison, [[VECTOR_BODY]] ], [ [[TMP6]], [[PRED_UDIV_IF]] ]
3213; UNROLL-NO-VF-NEXT:    br i1 [[TMP5]], label [[PRED_UDIV_IF3:%.*]], label [[PRED_UDIV_CONTINUE4:%.*]]
3214; UNROLL-NO-VF:       pred.udiv.if3:
3215; UNROLL-NO-VF-NEXT:    [[TMP8:%.*]] = udiv i32 219220132, [[TMP3]]
3216; UNROLL-NO-VF-NEXT:    br label [[PRED_UDIV_CONTINUE4]]
3217; UNROLL-NO-VF:       pred.udiv.continue4:
3218; UNROLL-NO-VF-NEXT:    [[TMP9]] = phi i32 [ poison, [[PRED_UDIV_CONTINUE]] ], [ [[TMP8]], [[PRED_UDIV_IF3]] ]
3219; UNROLL-NO-VF-NEXT:    [[TMP10]] = add i32 [[VEC_PHI]], [[VECTOR_RECUR]]
3220; UNROLL-NO-VF-NEXT:    [[TMP11]] = add i32 [[VEC_PHI1]], [[TMP7]]
3221; UNROLL-NO-VF-NEXT:    br i1 [[TMP4]], label [[PRED_STORE_IF:%.*]], label [[PRED_STORE_CONTINUE:%.*]]
3222; UNROLL-NO-VF:       pred.store.if:
3223; UNROLL-NO-VF-NEXT:    [[TMP12:%.*]] = add i32 [[INDEX]], 0
3224; UNROLL-NO-VF-NEXT:    [[TMP13:%.*]] = getelementptr inbounds i32, ptr [[X:%.*]], i32 [[TMP12]]
3225; UNROLL-NO-VF-NEXT:    store i32 [[TMP2]], ptr [[TMP13]], align 4
3226; UNROLL-NO-VF-NEXT:    br label [[PRED_STORE_CONTINUE]]
3227; UNROLL-NO-VF:       pred.store.continue:
3228; UNROLL-NO-VF-NEXT:    br i1 [[TMP5]], label [[PRED_STORE_IF5:%.*]], label [[PRED_STORE_CONTINUE6]]
3229; UNROLL-NO-VF:       pred.store.if5:
3230; UNROLL-NO-VF-NEXT:    [[TMP14:%.*]] = add i32 [[INDEX]], 1
3231; UNROLL-NO-VF-NEXT:    [[TMP15:%.*]] = getelementptr inbounds i32, ptr [[X]], i32 [[TMP14]]
3232; UNROLL-NO-VF-NEXT:    store i32 [[TMP3]], ptr [[TMP15]], align 4
3233; UNROLL-NO-VF-NEXT:    br label [[PRED_STORE_CONTINUE6]]
3234; UNROLL-NO-VF:       pred.store.continue6:
3235; UNROLL-NO-VF-NEXT:    [[TMP16:%.*]] = select i1 [[TMP4]], i32 [[TMP10]], i32 [[VEC_PHI]]
3236; UNROLL-NO-VF-NEXT:    [[TMP17:%.*]] = select i1 [[TMP5]], i32 [[TMP11]], i32 [[VEC_PHI1]]
3237; UNROLL-NO-VF-NEXT:    [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 2
3238; UNROLL-NO-VF-NEXT:    [[TMP18:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]]
3239; UNROLL-NO-VF-NEXT:    br i1 [[TMP18]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !prof [[PROF27]], !llvm.loop [[LOOP32:![0-9]+]]
3240; UNROLL-NO-VF:       middle.block:
3241; UNROLL-NO-VF-NEXT:    [[BIN_RDX:%.*]] = add i32 [[TMP17]], [[TMP16]]
3242; UNROLL-NO-VF-NEXT:    br i1 true, label [[BB1:%.*]], label [[SCALAR_PH]], !prof [[PROF29]]
3243; UNROLL-NO-VF:       scalar.ph:
3244; UNROLL-NO-VF-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i32 [ [[IND_END]], [[MIDDLE_BLOCK]] ], [ [[Y]], [[BB:%.*]] ]
3245; UNROLL-NO-VF-NEXT:    [[BC_RESUME_VAL1:%.*]] = phi i32 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[BB]] ]
3246; UNROLL-NO-VF-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i32 [ [[TMP9]], [[MIDDLE_BLOCK]] ], [ 0, [[BB]] ]
3247; UNROLL-NO-VF-NEXT:    [[BC_MERGE_RDX:%.*]] = phi i32 [ [[BIN_RDX]], [[MIDDLE_BLOCK]] ], [ 0, [[BB]] ]
3248; UNROLL-NO-VF-NEXT:    br label [[BB2:%.*]]
3249; UNROLL-NO-VF:       bb1:
3250; UNROLL-NO-VF-NEXT:    [[VAR:%.*]] = phi i32 [ [[VAR6:%.*]], [[BB2]] ], [ [[BIN_RDX]], [[MIDDLE_BLOCK]] ]
3251; UNROLL-NO-VF-NEXT:    ret i32 [[VAR]]
3252; UNROLL-NO-VF:       bb2:
3253; UNROLL-NO-VF-NEXT:    [[VAR3:%.*]] = phi i32 [ [[VAR8:%.*]], [[BB2]] ], [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ]
3254; UNROLL-NO-VF-NEXT:    [[IV:%.*]] = phi i32 [ [[IV_NEXT:%.*]], [[BB2]] ], [ [[BC_RESUME_VAL1]], [[SCALAR_PH]] ]
3255; UNROLL-NO-VF-NEXT:    [[VAR4:%.*]] = phi i32 [ [[VAR7:%.*]], [[BB2]] ], [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ]
3256; UNROLL-NO-VF-NEXT:    [[VAR5:%.*]] = phi i32 [ [[VAR6]], [[BB2]] ], [ [[BC_MERGE_RDX]], [[SCALAR_PH]] ]
3257; UNROLL-NO-VF-NEXT:    [[G:%.*]] = getelementptr inbounds i32, ptr [[X]], i32 [[IV]]
3258; UNROLL-NO-VF-NEXT:    [[VAR6]] = add i32 [[VAR5]], [[VAR4]]
3259; UNROLL-NO-VF-NEXT:    [[VAR7]] = udiv i32 219220132, [[VAR3]]
3260; UNROLL-NO-VF-NEXT:    store i32 [[VAR3]], ptr [[G]], align 4
3261; UNROLL-NO-VF-NEXT:    [[VAR8]] = add nsw i32 [[VAR3]], -1
3262; UNROLL-NO-VF-NEXT:    [[IV_NEXT]] = add nsw i32 [[IV]], 1
3263; UNROLL-NO-VF-NEXT:    [[VAR9:%.*]] = icmp slt i32 [[VAR3]], 2
3264; UNROLL-NO-VF-NEXT:    br i1 [[VAR9]], label [[BB1]], label [[BB2]], !prof [[PROF30]], !llvm.loop [[LOOP33:![0-9]+]]
3265;
3266; SINK-AFTER-LABEL: @sink_into_replication_region_multiple(
3267; SINK-AFTER-NEXT:  bb:
3268; SINK-AFTER-NEXT:    [[TMP0:%.*]] = add i32 [[Y:%.*]], 1
3269; SINK-AFTER-NEXT:    [[SMIN:%.*]] = call i32 @llvm.smin.i32(i32 [[Y]], i32 1)
3270; SINK-AFTER-NEXT:    [[TMP1:%.*]] = sub i32 [[TMP0]], [[SMIN]]
3271; SINK-AFTER-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]], !prof [[PROF26]]
3272; SINK-AFTER:       vector.ph:
3273; SINK-AFTER-NEXT:    [[N_RND_UP:%.*]] = add i32 [[TMP1]], 3
3274; SINK-AFTER-NEXT:    [[N_MOD_VF:%.*]] = urem i32 [[N_RND_UP]], 4
3275; SINK-AFTER-NEXT:    [[N_VEC:%.*]] = sub i32 [[N_RND_UP]], [[N_MOD_VF]]
3276; SINK-AFTER-NEXT:    [[TRIP_COUNT_MINUS_1:%.*]] = sub i32 [[TMP1]], 1
3277; SINK-AFTER-NEXT:    [[IND_END:%.*]] = sub i32 [[Y]], [[N_VEC]]
3278; SINK-AFTER-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <4 x i32> poison, i32 [[TRIP_COUNT_MINUS_1]], i64 0
3279; SINK-AFTER-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <4 x i32> [[BROADCAST_SPLATINSERT]], <4 x i32> poison, <4 x i32> zeroinitializer
3280; SINK-AFTER-NEXT:    br label [[VECTOR_BODY:%.*]]
3281; SINK-AFTER:       vector.body:
3282; SINK-AFTER-NEXT:    [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[PRED_STORE_CONTINUE12:%.*]] ]
3283; SINK-AFTER-NEXT:    [[VEC_IND:%.*]] = phi <4 x i32> [ <i32 0, i32 1, i32 2, i32 3>, [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[PRED_STORE_CONTINUE12]] ]
3284; SINK-AFTER-NEXT:    [[VECTOR_RECUR:%.*]] = phi <4 x i32> [ <i32 poison, i32 poison, i32 poison, i32 0>, [[VECTOR_PH]] ], [ [[TMP22:%.*]], [[PRED_STORE_CONTINUE12]] ]
3285; SINK-AFTER-NEXT:    [[VEC_PHI:%.*]] = phi <4 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP24:%.*]], [[PRED_STORE_CONTINUE12]] ]
3286; SINK-AFTER-NEXT:    [[OFFSET_IDX:%.*]] = sub i32 [[Y]], [[INDEX]]
3287; SINK-AFTER-NEXT:    [[TMP2:%.*]] = add i32 [[OFFSET_IDX]], 0
3288; SINK-AFTER-NEXT:    [[TMP3:%.*]] = add i32 [[OFFSET_IDX]], -1
3289; SINK-AFTER-NEXT:    [[TMP4:%.*]] = add i32 [[OFFSET_IDX]], -2
3290; SINK-AFTER-NEXT:    [[TMP5:%.*]] = add i32 [[OFFSET_IDX]], -3
3291; SINK-AFTER-NEXT:    [[TMP6:%.*]] = icmp ule <4 x i32> [[VEC_IND]], [[BROADCAST_SPLAT]]
3292; SINK-AFTER-NEXT:    [[TMP7:%.*]] = extractelement <4 x i1> [[TMP6]], i32 0
3293; SINK-AFTER-NEXT:    br i1 [[TMP7]], label [[PRED_UDIV_IF:%.*]], label [[PRED_UDIV_CONTINUE:%.*]]
3294; SINK-AFTER:       pred.udiv.if:
3295; SINK-AFTER-NEXT:    [[TMP8:%.*]] = udiv i32 219220132, [[TMP2]]
3296; SINK-AFTER-NEXT:    [[TMP9:%.*]] = insertelement <4 x i32> poison, i32 [[TMP8]], i32 0
3297; SINK-AFTER-NEXT:    br label [[PRED_UDIV_CONTINUE]]
3298; SINK-AFTER:       pred.udiv.continue:
3299; SINK-AFTER-NEXT:    [[TMP10:%.*]] = phi <4 x i32> [ poison, [[VECTOR_BODY]] ], [ [[TMP9]], [[PRED_UDIV_IF]] ]
3300; SINK-AFTER-NEXT:    [[TMP11:%.*]] = extractelement <4 x i1> [[TMP6]], i32 1
3301; SINK-AFTER-NEXT:    br i1 [[TMP11]], label [[PRED_UDIV_IF1:%.*]], label [[PRED_UDIV_CONTINUE2:%.*]]
3302; SINK-AFTER:       pred.udiv.if1:
3303; SINK-AFTER-NEXT:    [[TMP12:%.*]] = udiv i32 219220132, [[TMP3]]
3304; SINK-AFTER-NEXT:    [[TMP13:%.*]] = insertelement <4 x i32> [[TMP10]], i32 [[TMP12]], i32 1
3305; SINK-AFTER-NEXT:    br label [[PRED_UDIV_CONTINUE2]]
3306; SINK-AFTER:       pred.udiv.continue2:
3307; SINK-AFTER-NEXT:    [[TMP14:%.*]] = phi <4 x i32> [ [[TMP10]], [[PRED_UDIV_CONTINUE]] ], [ [[TMP13]], [[PRED_UDIV_IF1]] ]
3308; SINK-AFTER-NEXT:    [[TMP15:%.*]] = extractelement <4 x i1> [[TMP6]], i32 2
3309; SINK-AFTER-NEXT:    br i1 [[TMP15]], label [[PRED_UDIV_IF3:%.*]], label [[PRED_UDIV_CONTINUE4:%.*]]
3310; SINK-AFTER:       pred.udiv.if3:
3311; SINK-AFTER-NEXT:    [[TMP16:%.*]] = udiv i32 219220132, [[TMP4]]
3312; SINK-AFTER-NEXT:    [[TMP17:%.*]] = insertelement <4 x i32> [[TMP14]], i32 [[TMP16]], i32 2
3313; SINK-AFTER-NEXT:    br label [[PRED_UDIV_CONTINUE4]]
3314; SINK-AFTER:       pred.udiv.continue4:
3315; SINK-AFTER-NEXT:    [[TMP18:%.*]] = phi <4 x i32> [ [[TMP14]], [[PRED_UDIV_CONTINUE2]] ], [ [[TMP17]], [[PRED_UDIV_IF3]] ]
3316; SINK-AFTER-NEXT:    [[TMP19:%.*]] = extractelement <4 x i1> [[TMP6]], i32 3
3317; SINK-AFTER-NEXT:    br i1 [[TMP19]], label [[PRED_UDIV_IF5:%.*]], label [[PRED_UDIV_CONTINUE6:%.*]]
3318; SINK-AFTER:       pred.udiv.if5:
3319; SINK-AFTER-NEXT:    [[TMP20:%.*]] = udiv i32 219220132, [[TMP5]]
3320; SINK-AFTER-NEXT:    [[TMP21:%.*]] = insertelement <4 x i32> [[TMP18]], i32 [[TMP20]], i32 3
3321; SINK-AFTER-NEXT:    br label [[PRED_UDIV_CONTINUE6]]
3322; SINK-AFTER:       pred.udiv.continue6:
3323; SINK-AFTER-NEXT:    [[TMP22]] = phi <4 x i32> [ [[TMP18]], [[PRED_UDIV_CONTINUE4]] ], [ [[TMP21]], [[PRED_UDIV_IF5]] ]
3324; SINK-AFTER-NEXT:    [[TMP23:%.*]] = shufflevector <4 x i32> [[VECTOR_RECUR]], <4 x i32> [[TMP22]], <4 x i32> <i32 3, i32 4, i32 5, i32 6>
3325; SINK-AFTER-NEXT:    [[TMP24]] = add <4 x i32> [[VEC_PHI]], [[TMP23]]
3326; SINK-AFTER-NEXT:    [[TMP25:%.*]] = extractelement <4 x i1> [[TMP6]], i32 0
3327; SINK-AFTER-NEXT:    br i1 [[TMP25]], label [[PRED_STORE_IF:%.*]], label [[PRED_STORE_CONTINUE:%.*]]
3328; SINK-AFTER:       pred.store.if:
3329; SINK-AFTER-NEXT:    [[TMP26:%.*]] = add i32 [[INDEX]], 0
3330; SINK-AFTER-NEXT:    [[TMP27:%.*]] = getelementptr inbounds i32, ptr [[X:%.*]], i32 [[TMP26]]
3331; SINK-AFTER-NEXT:    store i32 [[TMP2]], ptr [[TMP27]], align 4
3332; SINK-AFTER-NEXT:    br label [[PRED_STORE_CONTINUE]]
3333; SINK-AFTER:       pred.store.continue:
3334; SINK-AFTER-NEXT:    [[TMP28:%.*]] = extractelement <4 x i1> [[TMP6]], i32 1
3335; SINK-AFTER-NEXT:    br i1 [[TMP28]], label [[PRED_STORE_IF7:%.*]], label [[PRED_STORE_CONTINUE8:%.*]]
3336; SINK-AFTER:       pred.store.if7:
3337; SINK-AFTER-NEXT:    [[TMP29:%.*]] = add i32 [[INDEX]], 1
3338; SINK-AFTER-NEXT:    [[TMP30:%.*]] = getelementptr inbounds i32, ptr [[X]], i32 [[TMP29]]
3339; SINK-AFTER-NEXT:    store i32 [[TMP3]], ptr [[TMP30]], align 4
3340; SINK-AFTER-NEXT:    br label [[PRED_STORE_CONTINUE8]]
3341; SINK-AFTER:       pred.store.continue8:
3342; SINK-AFTER-NEXT:    [[TMP31:%.*]] = extractelement <4 x i1> [[TMP6]], i32 2
3343; SINK-AFTER-NEXT:    br i1 [[TMP31]], label [[PRED_STORE_IF9:%.*]], label [[PRED_STORE_CONTINUE10:%.*]]
3344; SINK-AFTER:       pred.store.if9:
3345; SINK-AFTER-NEXT:    [[TMP32:%.*]] = add i32 [[INDEX]], 2
3346; SINK-AFTER-NEXT:    [[TMP33:%.*]] = getelementptr inbounds i32, ptr [[X]], i32 [[TMP32]]
3347; SINK-AFTER-NEXT:    store i32 [[TMP4]], ptr [[TMP33]], align 4
3348; SINK-AFTER-NEXT:    br label [[PRED_STORE_CONTINUE10]]
3349; SINK-AFTER:       pred.store.continue10:
3350; SINK-AFTER-NEXT:    [[TMP34:%.*]] = extractelement <4 x i1> [[TMP6]], i32 3
3351; SINK-AFTER-NEXT:    br i1 [[TMP34]], label [[PRED_STORE_IF11:%.*]], label [[PRED_STORE_CONTINUE12]]
3352; SINK-AFTER:       pred.store.if11:
3353; SINK-AFTER-NEXT:    [[TMP35:%.*]] = add i32 [[INDEX]], 3
3354; SINK-AFTER-NEXT:    [[TMP36:%.*]] = getelementptr inbounds i32, ptr [[X]], i32 [[TMP35]]
3355; SINK-AFTER-NEXT:    store i32 [[TMP5]], ptr [[TMP36]], align 4
3356; SINK-AFTER-NEXT:    br label [[PRED_STORE_CONTINUE12]]
3357; SINK-AFTER:       pred.store.continue12:
3358; SINK-AFTER-NEXT:    [[TMP37:%.*]] = select <4 x i1> [[TMP6]], <4 x i32> [[TMP24]], <4 x i32> [[VEC_PHI]]
3359; SINK-AFTER-NEXT:    [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 4
3360; SINK-AFTER-NEXT:    [[VEC_IND_NEXT]] = add <4 x i32> [[VEC_IND]], splat (i32 4)
3361; SINK-AFTER-NEXT:    [[TMP38:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]]
3362; SINK-AFTER-NEXT:    br i1 [[TMP38]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !prof [[PROF27]], !llvm.loop [[LOOP32:![0-9]+]]
3363; SINK-AFTER:       middle.block:
3364; SINK-AFTER-NEXT:    [[TMP39:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[TMP37]])
3365; SINK-AFTER-NEXT:    [[VECTOR_RECUR_EXTRACT:%.*]] = extractelement <4 x i32> [[TMP22]], i32 3
3366; SINK-AFTER-NEXT:    br i1 true, label [[BB1:%.*]], label [[SCALAR_PH]], !prof [[PROF29]]
3367; SINK-AFTER:       scalar.ph:
3368; SINK-AFTER-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i32 [ [[IND_END]], [[MIDDLE_BLOCK]] ], [ [[Y]], [[BB:%.*]] ]
3369; SINK-AFTER-NEXT:    [[BC_RESUME_VAL1:%.*]] = phi i32 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[BB]] ]
3370; SINK-AFTER-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i32 [ [[VECTOR_RECUR_EXTRACT]], [[MIDDLE_BLOCK]] ], [ 0, [[BB]] ]
3371; SINK-AFTER-NEXT:    [[BC_MERGE_RDX:%.*]] = phi i32 [ [[TMP39]], [[MIDDLE_BLOCK]] ], [ 0, [[BB]] ]
3372; SINK-AFTER-NEXT:    br label [[BB2:%.*]]
3373; SINK-AFTER:       bb1:
3374; SINK-AFTER-NEXT:    [[VAR:%.*]] = phi i32 [ [[VAR6:%.*]], [[BB2]] ], [ [[TMP39]], [[MIDDLE_BLOCK]] ]
3375; SINK-AFTER-NEXT:    ret i32 [[VAR]]
3376; SINK-AFTER:       bb2:
3377; SINK-AFTER-NEXT:    [[VAR3:%.*]] = phi i32 [ [[VAR8:%.*]], [[BB2]] ], [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ]
3378; SINK-AFTER-NEXT:    [[IV:%.*]] = phi i32 [ [[IV_NEXT:%.*]], [[BB2]] ], [ [[BC_RESUME_VAL1]], [[SCALAR_PH]] ]
3379; SINK-AFTER-NEXT:    [[VAR4:%.*]] = phi i32 [ [[VAR7:%.*]], [[BB2]] ], [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ]
3380; SINK-AFTER-NEXT:    [[VAR5:%.*]] = phi i32 [ [[VAR6]], [[BB2]] ], [ [[BC_MERGE_RDX]], [[SCALAR_PH]] ]
3381; SINK-AFTER-NEXT:    [[G:%.*]] = getelementptr inbounds i32, ptr [[X]], i32 [[IV]]
3382; SINK-AFTER-NEXT:    [[VAR6]] = add i32 [[VAR5]], [[VAR4]]
3383; SINK-AFTER-NEXT:    [[VAR7]] = udiv i32 219220132, [[VAR3]]
3384; SINK-AFTER-NEXT:    store i32 [[VAR3]], ptr [[G]], align 4
3385; SINK-AFTER-NEXT:    [[VAR8]] = add nsw i32 [[VAR3]], -1
3386; SINK-AFTER-NEXT:    [[IV_NEXT]] = add nsw i32 [[IV]], 1
3387; SINK-AFTER-NEXT:    [[VAR9:%.*]] = icmp slt i32 [[VAR3]], 2
3388; SINK-AFTER-NEXT:    br i1 [[VAR9]], label [[BB1]], label [[BB2]], !prof [[PROF30]], !llvm.loop [[LOOP33:![0-9]+]]
3389;
3390bb:
3391  br label %bb2
3392
3393  bb1:                                              ; preds = %bb2
3394  %var = phi i32 [ %var6, %bb2 ]
3395  ret i32 %var
3396
3397  bb2:                                              ; preds = %bb2, %bb
3398  %var3 = phi i32 [ %var8, %bb2 ], [ %y, %bb ]
3399  %iv = phi i32 [ %iv.next, %bb2 ], [ 0, %bb ]
3400  %var4 = phi i32 [ %var7, %bb2 ], [ 0, %bb ]
3401  %var5 = phi i32 [ %var6, %bb2 ], [ 0, %bb ]
3402  %g = getelementptr inbounds i32, ptr %x, i32 %iv
3403  %var6 = add i32 %var5, %var4
3404  %var7 = udiv i32 219220132, %var3
3405  store i32 %var3, ptr %g, align 4
3406  %var8 = add nsw i32 %var3, -1
3407  %iv.next = add nsw i32 %iv, 1
3408  %var9 = icmp slt i32 %var3, 2
3409  br i1 %var9, label %bb1, label %bb2, !prof !2
3410}
3411
3412; %vec.dead will be marked as dead instruction in the vector loop and no recipe
3413; will be created for it. Make sure a valid sink target is used.
3414define i32 @sink_after_dead_inst(ptr %A.ptr) {
3415; UNROLL-NO-IC-LABEL: @sink_after_dead_inst(
3416; UNROLL-NO-IC-NEXT:  entry:
3417; UNROLL-NO-IC-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
3418; UNROLL-NO-IC:       vector.ph:
3419; UNROLL-NO-IC-NEXT:    br label [[VECTOR_BODY:%.*]]
3420; UNROLL-NO-IC:       vector.body:
3421; UNROLL-NO-IC-NEXT:    [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
3422; UNROLL-NO-IC-NEXT:    [[VEC_IND:%.*]] = phi <4 x i16> [ <i16 0, i16 1, i16 2, i16 3>, [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[VECTOR_BODY]] ]
3423; UNROLL-NO-IC-NEXT:    [[VECTOR_RECUR:%.*]] = phi <4 x i32> [ <i32 poison, i32 poison, i32 poison, i32 0>, [[VECTOR_PH]] ], [ [[TMP3:%.*]], [[VECTOR_BODY]] ]
3424; UNROLL-NO-IC-NEXT:    [[STEP_ADD:%.*]] = add <4 x i16> [[VEC_IND]], splat (i16 4)
3425; UNROLL-NO-IC-NEXT:    [[OFFSET_IDX:%.*]] = trunc i32 [[INDEX]] to i16
3426; UNROLL-NO-IC-NEXT:    [[TMP0:%.*]] = add i16 [[OFFSET_IDX]], 0
3427; UNROLL-NO-IC-NEXT:    [[TMP1:%.*]] = add <4 x i16> [[STEP_ADD]], splat (i16 1)
3428; UNROLL-NO-IC-NEXT:    [[TMP2:%.*]] = or <4 x i16> [[TMP1]], [[TMP1]]
3429; UNROLL-NO-IC-NEXT:    [[TMP3]] = zext <4 x i16> [[TMP2]] to <4 x i32>
3430; UNROLL-NO-IC-NEXT:    [[TMP4:%.*]] = getelementptr i32, ptr [[A_PTR:%.*]], i16 [[TMP0]]
3431; UNROLL-NO-IC-NEXT:    [[TMP5:%.*]] = getelementptr i32, ptr [[TMP4]], i32 0
3432; UNROLL-NO-IC-NEXT:    [[TMP6:%.*]] = getelementptr i32, ptr [[TMP4]], i32 4
3433; UNROLL-NO-IC-NEXT:    store <4 x i32> zeroinitializer, ptr [[TMP5]], align 4
3434; UNROLL-NO-IC-NEXT:    store <4 x i32> zeroinitializer, ptr [[TMP6]], align 4
3435; UNROLL-NO-IC-NEXT:    [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 8
3436; UNROLL-NO-IC-NEXT:    [[VEC_IND_NEXT]] = add <4 x i16> [[STEP_ADD]], splat (i16 4)
3437; UNROLL-NO-IC-NEXT:    [[TMP7:%.*]] = icmp eq i32 [[INDEX_NEXT]], 16
3438; UNROLL-NO-IC-NEXT:    br i1 [[TMP7]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP34:![0-9]+]]
3439; UNROLL-NO-IC:       middle.block:
3440; UNROLL-NO-IC-NEXT:    [[VECTOR_RECUR_EXTRACT_FOR_PHI:%.*]] = extractelement <4 x i32> [[TMP3]], i32 2
3441; UNROLL-NO-IC-NEXT:    [[VECTOR_RECUR_EXTRACT:%.*]] = extractelement <4 x i32> [[TMP3]], i32 3
3442; UNROLL-NO-IC-NEXT:    br i1 true, label [[FOR_END:%.*]], label [[SCALAR_PH]]
3443; UNROLL-NO-IC:       scalar.ph:
3444; UNROLL-NO-IC-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i16 [ 16, [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ]
3445; UNROLL-NO-IC-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i32 [ [[VECTOR_RECUR_EXTRACT]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY]] ]
3446; UNROLL-NO-IC-NEXT:    br label [[LOOP:%.*]]
3447; UNROLL-NO-IC:       loop:
3448; UNROLL-NO-IC-NEXT:    [[IV:%.*]] = phi i16 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], [[LOOP]] ]
3449; UNROLL-NO-IC-NEXT:    [[FOR:%.*]] = phi i32 [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ], [ [[FOR_PREV:%.*]], [[LOOP]] ]
3450; UNROLL-NO-IC-NEXT:    [[CMP:%.*]] = icmp eq i32 [[FOR]], 15
3451; UNROLL-NO-IC-NEXT:    [[C:%.*]] = icmp eq i1 [[CMP]], true
3452; UNROLL-NO-IC-NEXT:    [[VEC_DEAD:%.*]] = and i1 [[C]], true
3453; UNROLL-NO-IC-NEXT:    [[IV_NEXT]] = add i16 [[IV]], 1
3454; UNROLL-NO-IC-NEXT:    [[B1:%.*]] = or i16 [[IV_NEXT]], [[IV_NEXT]]
3455; UNROLL-NO-IC-NEXT:    [[B3:%.*]] = and i1 [[CMP]], [[C]]
3456; UNROLL-NO-IC-NEXT:    [[FOR_PREV]] = zext i16 [[B1]] to i32
3457; UNROLL-NO-IC-NEXT:    [[EXT:%.*]] = zext i1 [[B3]] to i32
3458; UNROLL-NO-IC-NEXT:    [[A_GEP:%.*]] = getelementptr i32, ptr [[A_PTR]], i16 [[IV]]
3459; UNROLL-NO-IC-NEXT:    store i32 0, ptr [[A_GEP]], align 4
3460; UNROLL-NO-IC-NEXT:    br i1 [[VEC_DEAD]], label [[FOR_END]], label [[LOOP]], !llvm.loop [[LOOP35:![0-9]+]]
3461; UNROLL-NO-IC:       for.end:
3462; UNROLL-NO-IC-NEXT:    [[FOR_LCSSA:%.*]] = phi i32 [ [[FOR]], [[LOOP]] ], [ [[VECTOR_RECUR_EXTRACT_FOR_PHI]], [[MIDDLE_BLOCK]] ]
3463; UNROLL-NO-IC-NEXT:    ret i32 [[FOR_LCSSA]]
3464;
3465; UNROLL-NO-VF-LABEL: @sink_after_dead_inst(
3466; UNROLL-NO-VF-NEXT:  entry:
3467; UNROLL-NO-VF-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
3468; UNROLL-NO-VF:       vector.ph:
3469; UNROLL-NO-VF-NEXT:    br label [[VECTOR_BODY:%.*]]
3470; UNROLL-NO-VF:       vector.body:
3471; UNROLL-NO-VF-NEXT:    [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
3472; UNROLL-NO-VF-NEXT:    [[VECTOR_RECUR:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[TMP7:%.*]], [[VECTOR_BODY]] ]
3473; UNROLL-NO-VF-NEXT:    [[OFFSET_IDX:%.*]] = trunc i32 [[INDEX]] to i16
3474; UNROLL-NO-VF-NEXT:    [[TMP0:%.*]] = add i16 [[OFFSET_IDX]], 0
3475; UNROLL-NO-VF-NEXT:    [[TMP1:%.*]] = add i16 [[OFFSET_IDX]], 1
3476; UNROLL-NO-VF-NEXT:    [[TMP2:%.*]] = add i16 [[TMP0]], 1
3477; UNROLL-NO-VF-NEXT:    [[TMP3:%.*]] = add i16 [[TMP1]], 1
3478; UNROLL-NO-VF-NEXT:    [[TMP4:%.*]] = or i16 [[TMP2]], [[TMP2]]
3479; UNROLL-NO-VF-NEXT:    [[TMP5:%.*]] = or i16 [[TMP3]], [[TMP3]]
3480; UNROLL-NO-VF-NEXT:    [[TMP6:%.*]] = zext i16 [[TMP4]] to i32
3481; UNROLL-NO-VF-NEXT:    [[TMP7]] = zext i16 [[TMP5]] to i32
3482; UNROLL-NO-VF-NEXT:    [[TMP8:%.*]] = getelementptr i32, ptr [[A_PTR:%.*]], i16 [[TMP0]]
3483; UNROLL-NO-VF-NEXT:    [[TMP9:%.*]] = getelementptr i32, ptr [[A_PTR]], i16 [[TMP1]]
3484; UNROLL-NO-VF-NEXT:    store i32 0, ptr [[TMP8]], align 4
3485; UNROLL-NO-VF-NEXT:    store i32 0, ptr [[TMP9]], align 4
3486; UNROLL-NO-VF-NEXT:    [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 2
3487; UNROLL-NO-VF-NEXT:    [[TMP10:%.*]] = icmp eq i32 [[INDEX_NEXT]], 16
3488; UNROLL-NO-VF-NEXT:    br i1 [[TMP10]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP34:![0-9]+]]
3489; UNROLL-NO-VF:       middle.block:
3490; UNROLL-NO-VF-NEXT:    br i1 true, label [[FOR_END:%.*]], label [[SCALAR_PH]]
3491; UNROLL-NO-VF:       scalar.ph:
3492; UNROLL-NO-VF-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i16 [ 16, [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ]
3493; UNROLL-NO-VF-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i32 [ [[TMP7]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY]] ]
3494; UNROLL-NO-VF-NEXT:    br label [[LOOP:%.*]]
3495; UNROLL-NO-VF:       loop:
3496; UNROLL-NO-VF-NEXT:    [[IV:%.*]] = phi i16 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], [[LOOP]] ]
3497; UNROLL-NO-VF-NEXT:    [[FOR:%.*]] = phi i32 [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ], [ [[FOR_PREV:%.*]], [[LOOP]] ]
3498; UNROLL-NO-VF-NEXT:    [[CMP:%.*]] = icmp eq i32 [[FOR]], 15
3499; UNROLL-NO-VF-NEXT:    [[C:%.*]] = icmp eq i1 [[CMP]], true
3500; UNROLL-NO-VF-NEXT:    [[VEC_DEAD:%.*]] = and i1 [[C]], true
3501; UNROLL-NO-VF-NEXT:    [[IV_NEXT]] = add i16 [[IV]], 1
3502; UNROLL-NO-VF-NEXT:    [[B1:%.*]] = or i16 [[IV_NEXT]], [[IV_NEXT]]
3503; UNROLL-NO-VF-NEXT:    [[B3:%.*]] = and i1 [[CMP]], [[C]]
3504; UNROLL-NO-VF-NEXT:    [[FOR_PREV]] = zext i16 [[B1]] to i32
3505; UNROLL-NO-VF-NEXT:    [[EXT:%.*]] = zext i1 [[B3]] to i32
3506; UNROLL-NO-VF-NEXT:    [[A_GEP:%.*]] = getelementptr i32, ptr [[A_PTR]], i16 [[IV]]
3507; UNROLL-NO-VF-NEXT:    store i32 0, ptr [[A_GEP]], align 4
3508; UNROLL-NO-VF-NEXT:    br i1 [[VEC_DEAD]], label [[FOR_END]], label [[LOOP]], !llvm.loop [[LOOP35:![0-9]+]]
3509; UNROLL-NO-VF:       for.end:
3510; UNROLL-NO-VF-NEXT:    [[FOR_LCSSA:%.*]] = phi i32 [ [[FOR]], [[LOOP]] ], [ [[TMP6]], [[MIDDLE_BLOCK]] ]
3511; UNROLL-NO-VF-NEXT:    ret i32 [[FOR_LCSSA]]
3512;
3513; SINK-AFTER-LABEL: @sink_after_dead_inst(
3514; SINK-AFTER-NEXT:  entry:
3515; SINK-AFTER-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
3516; SINK-AFTER:       vector.ph:
3517; SINK-AFTER-NEXT:    br label [[VECTOR_BODY:%.*]]
3518; SINK-AFTER:       vector.body:
3519; SINK-AFTER-NEXT:    [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
3520; SINK-AFTER-NEXT:    [[VEC_IND:%.*]] = phi <4 x i16> [ <i16 0, i16 1, i16 2, i16 3>, [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[VECTOR_BODY]] ]
3521; SINK-AFTER-NEXT:    [[VECTOR_RECUR:%.*]] = phi <4 x i32> [ <i32 poison, i32 poison, i32 poison, i32 0>, [[VECTOR_PH]] ], [ [[TMP3:%.*]], [[VECTOR_BODY]] ]
3522; SINK-AFTER-NEXT:    [[OFFSET_IDX:%.*]] = trunc i32 [[INDEX]] to i16
3523; SINK-AFTER-NEXT:    [[TMP0:%.*]] = add i16 [[OFFSET_IDX]], 0
3524; SINK-AFTER-NEXT:    [[TMP1:%.*]] = add <4 x i16> [[VEC_IND]], splat (i16 1)
3525; SINK-AFTER-NEXT:    [[TMP2:%.*]] = or <4 x i16> [[TMP1]], [[TMP1]]
3526; SINK-AFTER-NEXT:    [[TMP3]] = zext <4 x i16> [[TMP2]] to <4 x i32>
3527; SINK-AFTER-NEXT:    [[TMP4:%.*]] = getelementptr i32, ptr [[A_PTR:%.*]], i16 [[TMP0]]
3528; SINK-AFTER-NEXT:    [[TMP5:%.*]] = getelementptr i32, ptr [[TMP4]], i32 0
3529; SINK-AFTER-NEXT:    store <4 x i32> zeroinitializer, ptr [[TMP5]], align 4
3530; SINK-AFTER-NEXT:    [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 4
3531; SINK-AFTER-NEXT:    [[VEC_IND_NEXT]] = add <4 x i16> [[VEC_IND]], splat (i16 4)
3532; SINK-AFTER-NEXT:    [[TMP6:%.*]] = icmp eq i32 [[INDEX_NEXT]], 16
3533; SINK-AFTER-NEXT:    br i1 [[TMP6]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP34:![0-9]+]]
3534; SINK-AFTER:       middle.block:
3535; SINK-AFTER-NEXT:    [[VECTOR_RECUR_EXTRACT_FOR_PHI:%.*]] = extractelement <4 x i32> [[TMP3]], i32 2
3536; SINK-AFTER-NEXT:    [[VECTOR_RECUR_EXTRACT:%.*]] = extractelement <4 x i32> [[TMP3]], i32 3
3537; SINK-AFTER-NEXT:    br i1 true, label [[FOR_END:%.*]], label [[SCALAR_PH]]
3538; SINK-AFTER:       scalar.ph:
3539; SINK-AFTER-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i16 [ 16, [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ]
3540; SINK-AFTER-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i32 [ [[VECTOR_RECUR_EXTRACT]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY]] ]
3541; SINK-AFTER-NEXT:    br label [[LOOP:%.*]]
3542; SINK-AFTER:       loop:
3543; SINK-AFTER-NEXT:    [[IV:%.*]] = phi i16 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], [[LOOP]] ]
3544; SINK-AFTER-NEXT:    [[FOR:%.*]] = phi i32 [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ], [ [[FOR_PREV:%.*]], [[LOOP]] ]
3545; SINK-AFTER-NEXT:    [[CMP:%.*]] = icmp eq i32 [[FOR]], 15
3546; SINK-AFTER-NEXT:    [[C:%.*]] = icmp eq i1 [[CMP]], true
3547; SINK-AFTER-NEXT:    [[VEC_DEAD:%.*]] = and i1 [[C]], true
3548; SINK-AFTER-NEXT:    [[IV_NEXT]] = add i16 [[IV]], 1
3549; SINK-AFTER-NEXT:    [[B1:%.*]] = or i16 [[IV_NEXT]], [[IV_NEXT]]
3550; SINK-AFTER-NEXT:    [[B3:%.*]] = and i1 [[CMP]], [[C]]
3551; SINK-AFTER-NEXT:    [[FOR_PREV]] = zext i16 [[B1]] to i32
3552; SINK-AFTER-NEXT:    [[EXT:%.*]] = zext i1 [[B3]] to i32
3553; SINK-AFTER-NEXT:    [[A_GEP:%.*]] = getelementptr i32, ptr [[A_PTR]], i16 [[IV]]
3554; SINK-AFTER-NEXT:    store i32 0, ptr [[A_GEP]], align 4
3555; SINK-AFTER-NEXT:    br i1 [[VEC_DEAD]], label [[FOR_END]], label [[LOOP]], !llvm.loop [[LOOP35:![0-9]+]]
3556; SINK-AFTER:       for.end:
3557; SINK-AFTER-NEXT:    [[FOR_LCSSA:%.*]] = phi i32 [ [[FOR]], [[LOOP]] ], [ [[VECTOR_RECUR_EXTRACT_FOR_PHI]], [[MIDDLE_BLOCK]] ]
3558; SINK-AFTER-NEXT:    ret i32 [[FOR_LCSSA]]
3559;
3560entry:
3561  br label %loop
3562
3563loop:
3564  %iv = phi i16 [ 0, %entry ], [ %iv.next, %loop ]
3565  %for = phi i32 [ 0, %entry ], [ %for.prev, %loop ]
3566  %cmp = icmp eq i32 %for, 15
3567  %C = icmp eq i1 %cmp, true
3568  %vec.dead = and i1 %C, 1
3569  %iv.next = add i16 %iv, 1
3570  %B1 = or i16 %iv.next, %iv.next
3571  %B3 = and i1 %cmp, %C
3572  %for.prev = zext i16 %B1 to i32
3573
3574  %ext = zext i1 %B3 to i32
3575  %A.gep = getelementptr i32, ptr %A.ptr, i16 %iv
3576  store i32 0, ptr %A.gep
3577  br i1 %vec.dead, label %for.end, label %loop
3578
3579for.end:
3580  ret i32 %for
3581}
3582
3583; %rec.1 only has %use.rec.1 as use, which can be removed. This enables %rec.1
3584; to be removed also.
3585define void @unused_recurrence(ptr %a) {
3586; UNROLL-NO-IC-LABEL: @unused_recurrence(
3587; UNROLL-NO-IC-NEXT:  entry:
3588; UNROLL-NO-IC-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
3589; UNROLL-NO-IC:       vector.ph:
3590; UNROLL-NO-IC-NEXT:    br label [[VECTOR_BODY:%.*]]
3591; UNROLL-NO-IC:       vector.body:
3592; UNROLL-NO-IC-NEXT:    [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
3593; UNROLL-NO-IC-NEXT:    [[VEC_IND:%.*]] = phi <4 x i16> [ <i16 -27, i16 -26, i16 -25, i16 -24>, [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[VECTOR_BODY]] ]
3594; UNROLL-NO-IC-NEXT:    [[VECTOR_RECUR:%.*]] = phi <4 x i16> [ <i16 poison, i16 poison, i16 poison, i16 0>, [[VECTOR_PH]] ], [ [[TMP1:%.*]], [[VECTOR_BODY]] ]
3595; UNROLL-NO-IC-NEXT:    [[STEP_ADD:%.*]] = add <4 x i16> [[VEC_IND]], splat (i16 4)
3596; UNROLL-NO-IC-NEXT:    [[TMP0:%.*]] = add <4 x i16> [[STEP_ADD]], splat (i16 1)
3597; UNROLL-NO-IC-NEXT:    [[TMP1]] = add <4 x i16> [[TMP0]], splat (i16 5)
3598; UNROLL-NO-IC-NEXT:    [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 8
3599; UNROLL-NO-IC-NEXT:    [[VEC_IND_NEXT]] = add <4 x i16> [[STEP_ADD]], splat (i16 4)
3600; UNROLL-NO-IC-NEXT:    [[TMP2:%.*]] = icmp eq i32 [[INDEX_NEXT]], 1024
3601; UNROLL-NO-IC-NEXT:    br i1 [[TMP2]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP36:![0-9]+]]
3602; UNROLL-NO-IC:       middle.block:
3603; UNROLL-NO-IC-NEXT:    [[VECTOR_RECUR_EXTRACT:%.*]] = extractelement <4 x i16> [[TMP1]], i32 3
3604; UNROLL-NO-IC-NEXT:    br i1 false, label [[FOR_END:%.*]], label [[SCALAR_PH]]
3605; UNROLL-NO-IC:       scalar.ph:
3606; UNROLL-NO-IC-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i16 [ 997, [[MIDDLE_BLOCK]] ], [ -27, [[ENTRY:%.*]] ]
3607; UNROLL-NO-IC-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i16 [ [[VECTOR_RECUR_EXTRACT]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY]] ]
3608; UNROLL-NO-IC-NEXT:    br label [[FOR_COND:%.*]]
3609; UNROLL-NO-IC:       for.cond:
3610; UNROLL-NO-IC-NEXT:    [[IV:%.*]] = phi i16 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], [[FOR_COND]] ]
3611; UNROLL-NO-IC-NEXT:    [[REC_1:%.*]] = phi i16 [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ], [ [[REC_1_PREV:%.*]], [[FOR_COND]] ]
3612; UNROLL-NO-IC-NEXT:    [[USE_REC_1:%.*]] = sub i16 [[REC_1]], 10
3613; UNROLL-NO-IC-NEXT:    [[IV_NEXT]] = add i16 [[IV]], 1
3614; UNROLL-NO-IC-NEXT:    [[REC_1_PREV]] = add i16 [[IV_NEXT]], 5
3615; UNROLL-NO-IC-NEXT:    [[CMP:%.*]] = icmp eq i16 [[IV]], 1000
3616; UNROLL-NO-IC-NEXT:    br i1 [[CMP]], label [[FOR_END]], label [[FOR_COND]], !llvm.loop [[LOOP37:![0-9]+]]
3617; UNROLL-NO-IC:       for.end:
3618; UNROLL-NO-IC-NEXT:    ret void
3619;
3620; UNROLL-NO-VF-LABEL: @unused_recurrence(
3621; UNROLL-NO-VF-NEXT:  entry:
3622; UNROLL-NO-VF-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
3623; UNROLL-NO-VF:       vector.ph:
3624; UNROLL-NO-VF-NEXT:    br label [[VECTOR_BODY:%.*]]
3625; UNROLL-NO-VF:       vector.body:
3626; UNROLL-NO-VF-NEXT:    [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
3627; UNROLL-NO-VF-NEXT:    [[VECTOR_RECUR:%.*]] = phi i16 [ 0, [[VECTOR_PH]] ], [ [[TMP2:%.*]], [[VECTOR_BODY]] ]
3628; UNROLL-NO-VF-NEXT:    [[DOTCAST:%.*]] = trunc i32 [[INDEX]] to i16
3629; UNROLL-NO-VF-NEXT:    [[OFFSET_IDX:%.*]] = add i16 -27, [[DOTCAST]]
3630; UNROLL-NO-VF-NEXT:    [[TMP0:%.*]] = add i16 [[OFFSET_IDX]], 1
3631; UNROLL-NO-VF-NEXT:    [[TMP1:%.*]] = add i16 [[TMP0]], 1
3632; UNROLL-NO-VF-NEXT:    [[TMP2]] = add i16 [[TMP1]], 5
3633; UNROLL-NO-VF-NEXT:    [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 2
3634; UNROLL-NO-VF-NEXT:    [[TMP3:%.*]] = icmp eq i32 [[INDEX_NEXT]], 1028
3635; UNROLL-NO-VF-NEXT:    br i1 [[TMP3]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP36:![0-9]+]]
3636; UNROLL-NO-VF:       middle.block:
3637; UNROLL-NO-VF-NEXT:    br i1 true, label [[FOR_END:%.*]], label [[SCALAR_PH]]
3638; UNROLL-NO-VF:       scalar.ph:
3639; UNROLL-NO-VF-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i16 [ 1001, [[MIDDLE_BLOCK]] ], [ -27, [[ENTRY:%.*]] ]
3640; UNROLL-NO-VF-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i16 [ [[TMP2]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY]] ]
3641; UNROLL-NO-VF-NEXT:    br label [[FOR_COND:%.*]]
3642; UNROLL-NO-VF:       for.cond:
3643; UNROLL-NO-VF-NEXT:    [[IV:%.*]] = phi i16 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], [[FOR_COND]] ]
3644; UNROLL-NO-VF-NEXT:    [[REC_1:%.*]] = phi i16 [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ], [ [[REC_1_PREV:%.*]], [[FOR_COND]] ]
3645; UNROLL-NO-VF-NEXT:    [[USE_REC_1:%.*]] = sub i16 [[REC_1]], 10
3646; UNROLL-NO-VF-NEXT:    [[IV_NEXT]] = add i16 [[IV]], 1
3647; UNROLL-NO-VF-NEXT:    [[REC_1_PREV]] = add i16 [[IV_NEXT]], 5
3648; UNROLL-NO-VF-NEXT:    [[CMP:%.*]] = icmp eq i16 [[IV]], 1000
3649; UNROLL-NO-VF-NEXT:    br i1 [[CMP]], label [[FOR_END]], label [[FOR_COND]], !llvm.loop [[LOOP37:![0-9]+]]
3650; UNROLL-NO-VF:       for.end:
3651; UNROLL-NO-VF-NEXT:    ret void
3652;
3653; SINK-AFTER-LABEL: @unused_recurrence(
3654; SINK-AFTER-NEXT:  entry:
3655; SINK-AFTER-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
3656; SINK-AFTER:       vector.ph:
3657; SINK-AFTER-NEXT:    br label [[VECTOR_BODY:%.*]]
3658; SINK-AFTER:       vector.body:
3659; SINK-AFTER-NEXT:    [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
3660; SINK-AFTER-NEXT:    [[VEC_IND:%.*]] = phi <4 x i16> [ <i16 -27, i16 -26, i16 -25, i16 -24>, [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[VECTOR_BODY]] ]
3661; SINK-AFTER-NEXT:    [[VECTOR_RECUR:%.*]] = phi <4 x i16> [ <i16 poison, i16 poison, i16 poison, i16 0>, [[VECTOR_PH]] ], [ [[TMP1:%.*]], [[VECTOR_BODY]] ]
3662; SINK-AFTER-NEXT:    [[TMP0:%.*]] = add <4 x i16> [[VEC_IND]], splat (i16 1)
3663; SINK-AFTER-NEXT:    [[TMP1]] = add <4 x i16> [[TMP0]], splat (i16 5)
3664; SINK-AFTER-NEXT:    [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 4
3665; SINK-AFTER-NEXT:    [[VEC_IND_NEXT]] = add <4 x i16> [[VEC_IND]], splat (i16 4)
3666; SINK-AFTER-NEXT:    [[TMP2:%.*]] = icmp eq i32 [[INDEX_NEXT]], 1028
3667; SINK-AFTER-NEXT:    br i1 [[TMP2]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP36:![0-9]+]]
3668; SINK-AFTER:       middle.block:
3669; SINK-AFTER-NEXT:    [[VECTOR_RECUR_EXTRACT:%.*]] = extractelement <4 x i16> [[TMP1]], i32 3
3670; SINK-AFTER-NEXT:    br i1 true, label [[FOR_END:%.*]], label [[SCALAR_PH]]
3671; SINK-AFTER:       scalar.ph:
3672; SINK-AFTER-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i16 [ 1001, [[MIDDLE_BLOCK]] ], [ -27, [[ENTRY:%.*]] ]
3673; SINK-AFTER-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i16 [ [[VECTOR_RECUR_EXTRACT]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY]] ]
3674; SINK-AFTER-NEXT:    br label [[FOR_COND:%.*]]
3675; SINK-AFTER:       for.cond:
3676; SINK-AFTER-NEXT:    [[IV:%.*]] = phi i16 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], [[FOR_COND]] ]
3677; SINK-AFTER-NEXT:    [[REC_1:%.*]] = phi i16 [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ], [ [[REC_1_PREV:%.*]], [[FOR_COND]] ]
3678; SINK-AFTER-NEXT:    [[USE_REC_1:%.*]] = sub i16 [[REC_1]], 10
3679; SINK-AFTER-NEXT:    [[IV_NEXT]] = add i16 [[IV]], 1
3680; SINK-AFTER-NEXT:    [[REC_1_PREV]] = add i16 [[IV_NEXT]], 5
3681; SINK-AFTER-NEXT:    [[CMP:%.*]] = icmp eq i16 [[IV]], 1000
3682; SINK-AFTER-NEXT:    br i1 [[CMP]], label [[FOR_END]], label [[FOR_COND]], !llvm.loop [[LOOP37:![0-9]+]]
3683; SINK-AFTER:       for.end:
3684; SINK-AFTER-NEXT:    ret void
3685;
3686entry:
3687  br label %for.cond
3688
3689for.cond:
3690  %iv = phi i16 [ -27, %entry ], [ %iv.next, %for.cond ]
3691  %rec.1 = phi i16 [ 0, %entry ], [ %rec.1.prev, %for.cond ]
3692  %use.rec.1 = sub i16 %rec.1, 10
3693  %iv.next= add i16 %iv, 1
3694  %rec.1.prev = add i16 %iv.next, 5
3695  %cmp = icmp eq i16 %iv, 1000
3696  br i1 %cmp, label %for.end, label %for.cond
3697
3698for.end:
3699  ret void
3700}
3701
3702; Test case for https://github.com/llvm/llvm-project/issues/95520.
3703define i32 @recurence_uniform_load(ptr %src, ptr noalias %dst) {
3704; UNROLL-NO-IC-LABEL: @recurence_uniform_load(
3705; UNROLL-NO-IC-NEXT:  entry:
3706; UNROLL-NO-IC-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
3707; UNROLL-NO-IC:       vector.ph:
3708; UNROLL-NO-IC-NEXT:    br label [[VECTOR_BODY:%.*]]
3709; UNROLL-NO-IC:       vector.body:
3710; UNROLL-NO-IC-NEXT:    [[TMP0:%.*]] = load i32, ptr [[SRC:%.*]], align 4
3711; UNROLL-NO-IC-NEXT:    br label [[MIDDLE_BLOCK:%.*]]
3712; UNROLL-NO-IC:       middle.block:
3713; UNROLL-NO-IC-NEXT:    br i1 true, label [[EXIT:%.*]], label [[SCALAR_PH]]
3714; UNROLL-NO-IC:       scalar.ph:
3715; UNROLL-NO-IC-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ 8, [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ]
3716; UNROLL-NO-IC-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i32 [ [[TMP0]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY]] ]
3717; UNROLL-NO-IC-NEXT:    br label [[LOOP:%.*]]
3718; UNROLL-NO-IC:       loop:
3719; UNROLL-NO-IC-NEXT:    [[PHI:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[ADD:%.*]], [[LOOP]] ]
3720; UNROLL-NO-IC-NEXT:    [[RECUR:%.*]] = phi i32 [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ], [ [[LOAD:%.*]], [[LOOP]] ]
3721; UNROLL-NO-IC-NEXT:    [[ADD]] = add i64 [[PHI]], 1
3722; UNROLL-NO-IC-NEXT:    [[LOAD]] = load i32, ptr [[SRC]], align 4
3723; UNROLL-NO-IC-NEXT:    [[ICMP:%.*]] = icmp ult i64 [[PHI]], 1
3724; UNROLL-NO-IC-NEXT:    br i1 [[ICMP]], label [[LOOP]], label [[EXIT]], !llvm.loop [[LOOP38:![0-9]+]]
3725; UNROLL-NO-IC:       exit:
3726; UNROLL-NO-IC-NEXT:    ret i32 0
3727;
3728; UNROLL-NO-VF-LABEL: @recurence_uniform_load(
3729; UNROLL-NO-VF-NEXT:  entry:
3730; UNROLL-NO-VF-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
3731; UNROLL-NO-VF:       vector.ph:
3732; UNROLL-NO-VF-NEXT:    br label [[VECTOR_BODY:%.*]]
3733; UNROLL-NO-VF:       vector.body:
3734; UNROLL-NO-VF-NEXT:    [[TMP0:%.*]] = load i32, ptr [[SRC:%.*]], align 4
3735; UNROLL-NO-VF-NEXT:    br label [[MIDDLE_BLOCK:%.*]]
3736; UNROLL-NO-VF:       middle.block:
3737; UNROLL-NO-VF-NEXT:    br i1 true, label [[EXIT:%.*]], label [[SCALAR_PH]]
3738; UNROLL-NO-VF:       scalar.ph:
3739; UNROLL-NO-VF-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ 2, [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ]
3740; UNROLL-NO-VF-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i32 [ [[TMP0]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY]] ]
3741; UNROLL-NO-VF-NEXT:    br label [[LOOP:%.*]]
3742; UNROLL-NO-VF:       loop:
3743; UNROLL-NO-VF-NEXT:    [[PHI:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[ADD:%.*]], [[LOOP]] ]
3744; UNROLL-NO-VF-NEXT:    [[RECUR:%.*]] = phi i32 [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ], [ [[LOAD:%.*]], [[LOOP]] ]
3745; UNROLL-NO-VF-NEXT:    [[ADD]] = add i64 [[PHI]], 1
3746; UNROLL-NO-VF-NEXT:    [[LOAD]] = load i32, ptr [[SRC]], align 4
3747; UNROLL-NO-VF-NEXT:    [[ICMP:%.*]] = icmp ult i64 [[PHI]], 1
3748; UNROLL-NO-VF-NEXT:    br i1 [[ICMP]], label [[LOOP]], label [[EXIT]], !llvm.loop [[LOOP38:![0-9]+]]
3749; UNROLL-NO-VF:       exit:
3750; UNROLL-NO-VF-NEXT:    ret i32 0
3751;
3752; SINK-AFTER-LABEL: @recurence_uniform_load(
3753; SINK-AFTER-NEXT:  entry:
3754; SINK-AFTER-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
3755; SINK-AFTER:       vector.ph:
3756; SINK-AFTER-NEXT:    br label [[VECTOR_BODY:%.*]]
3757; SINK-AFTER:       vector.body:
3758; SINK-AFTER-NEXT:    [[TMP0:%.*]] = load i32, ptr [[SRC:%.*]], align 4
3759; SINK-AFTER-NEXT:    br label [[MIDDLE_BLOCK:%.*]]
3760; SINK-AFTER:       middle.block:
3761; SINK-AFTER-NEXT:    br i1 true, label [[EXIT:%.*]], label [[SCALAR_PH]]
3762; SINK-AFTER:       scalar.ph:
3763; SINK-AFTER-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ 4, [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ]
3764; SINK-AFTER-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i32 [ [[TMP0]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY]] ]
3765; SINK-AFTER-NEXT:    br label [[LOOP:%.*]]
3766; SINK-AFTER:       loop:
3767; SINK-AFTER-NEXT:    [[PHI:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[ADD:%.*]], [[LOOP]] ]
3768; SINK-AFTER-NEXT:    [[RECUR:%.*]] = phi i32 [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ], [ [[LOAD:%.*]], [[LOOP]] ]
3769; SINK-AFTER-NEXT:    [[ADD]] = add i64 [[PHI]], 1
3770; SINK-AFTER-NEXT:    [[LOAD]] = load i32, ptr [[SRC]], align 4
3771; SINK-AFTER-NEXT:    [[ICMP:%.*]] = icmp ult i64 [[PHI]], 1
3772; SINK-AFTER-NEXT:    br i1 [[ICMP]], label [[LOOP]], label [[EXIT]], !llvm.loop [[LOOP38:![0-9]+]]
3773; SINK-AFTER:       exit:
3774; SINK-AFTER-NEXT:    ret i32 0
3775;
3776entry:
3777  br label %loop
3778
3779loop:
3780  %phi = phi i64 [ 0, %entry ], [ %add, %loop ]
3781  %recur = phi i32 [ 0, %entry ], [ %load, %loop ]
3782  %add = add i64 %phi, 1
3783  %load = load i32, ptr %src, align 4
3784  %icmp = icmp ult i64 %phi, 1
3785  br i1 %icmp, label %loop, label %exit
3786
3787exit:
3788  ret i32 0
3789}
3790
3791!2 = !{!"branch_weights", i32 1, i32 1}
3792