xref: /llvm-project/llvm/test/Transforms/LoopVectorize/AArch64/sve-interleaved-accesses.ll (revision 6c787ff6cfb5fdf489019a1389f8315391ad435f)
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
2; RUN: opt -mtriple=aarch64-none-linux-gnu -S -passes=loop-vectorize,instcombine -force-vector-width=4 -force-vector-interleave=1 -enable-interleaved-mem-accesses=true -mattr=+sve -scalable-vectorization=on -runtime-memory-check-threshold=24 < %s | FileCheck %s
3
4target datalayout = "e-m:e-i64:64-i128:128-n32:64-S128"
5
6; Check vectorization on an interleaved load group of factor 2 and an interleaved
7; store group of factor 2.
8
9; int AB[1024];
10; int CD[1024];
11;  void test_array_load2_store2(int C, int D) {
12;   for (int i = 0; i < 1024; i+=2) {
13;     int A = AB[i];
14;     int B = AB[i+1];
15;     CD[i] = A + C;
16;     CD[i+1] = B * D;
17;   }
18; }
19
20
21@AB = common global [1024 x i32] zeroinitializer, align 4
22@CD = common global [1024 x i32] zeroinitializer, align 4
23
24define void @test_array_load2_store2(i32 %C, i32 %D) #1 {
25; CHECK-LABEL: @test_array_load2_store2(
26; CHECK-NEXT:  entry:
27; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
28; CHECK:       vector.ph:
29; CHECK-NEXT:    [[TMP0:%.*]] = call i64 @llvm.vscale.i64()
30; CHECK-NEXT:    [[TMP1:%.*]] = shl nuw nsw i64 [[TMP0]], 2
31; CHECK-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <vscale x 4 x i32> poison, i32 [[C:%.*]], i64 0
32; CHECK-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <vscale x 4 x i32> [[BROADCAST_SPLATINSERT]], <vscale x 4 x i32> poison, <vscale x 4 x i32> zeroinitializer
33; CHECK-NEXT:    [[BROADCAST_SPLATINSERT1:%.*]] = insertelement <vscale x 4 x i32> poison, i32 [[D:%.*]], i64 0
34; CHECK-NEXT:    [[BROADCAST_SPLAT2:%.*]] = shufflevector <vscale x 4 x i32> [[BROADCAST_SPLATINSERT1]], <vscale x 4 x i32> poison, <vscale x 4 x i32> zeroinitializer
35; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
36; CHECK:       vector.body:
37; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
38; CHECK-NEXT:    [[OFFSET_IDX:%.*]] = shl i64 [[INDEX]], 1
39; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [1024 x i32], ptr @AB, i64 0, i64 [[OFFSET_IDX]]
40; CHECK-NEXT:    [[WIDE_VEC:%.*]] = load <vscale x 8 x i32>, ptr [[TMP2]], align 4
41; CHECK-NEXT:    [[STRIDED_VEC:%.*]] = call { <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.vector.deinterleave2.nxv8i32(<vscale x 8 x i32> [[WIDE_VEC]])
42; CHECK-NEXT:    [[TMP3:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC]], 0
43; CHECK-NEXT:    [[TMP4:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC]], 1
44; CHECK-NEXT:    [[TMP6:%.*]] = add nsw <vscale x 4 x i32> [[TMP3]], [[BROADCAST_SPLAT]]
45; CHECK-NEXT:    [[TMP7:%.*]] = mul nsw <vscale x 4 x i32> [[TMP4]], [[BROADCAST_SPLAT2]]
46; CHECK-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1024 x i32], ptr @CD, i64 0, i64 [[OFFSET_IDX]]
47; CHECK-NEXT:    [[INTERLEAVED_VEC:%.*]] = call <vscale x 8 x i32> @llvm.vector.interleave2.nxv8i32(<vscale x 4 x i32> [[TMP6]], <vscale x 4 x i32> [[TMP7]])
48; CHECK-NEXT:    store <vscale x 8 x i32> [[INTERLEAVED_VEC]], ptr [[TMP8]], align 4
49; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], [[TMP1]]
50; CHECK-NEXT:    [[TMP10:%.*]] = icmp eq i64 [[INDEX_NEXT]], 512
51; CHECK-NEXT:    br i1 [[TMP10]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
52; CHECK:       middle.block:
53; CHECK-NEXT:    br i1 true, label [[FOR_END:%.*]], label [[SCALAR_PH]]
54; CHECK:       scalar.ph:
55; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
56; CHECK:       for.body:
57; CHECK-NEXT:    br i1 poison, label [[FOR_BODY]], label [[FOR_END]], !llvm.loop [[LOOP3:![0-9]+]]
58; CHECK:       for.end:
59; CHECK-NEXT:    ret void
60;
61entry:
62  br label %for.body
63
64for.body:                                         ; preds = %for.body, %entry
65  %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
66  %arrayidx0 = getelementptr inbounds [1024 x i32], ptr @AB, i64 0, i64 %indvars.iv
67  %load1 = load i32, ptr %arrayidx0, align 4
68  %or = or disjoint i64 %indvars.iv, 1
69  %arrayidx1 = getelementptr inbounds [1024 x i32], ptr @AB, i64 0, i64 %or
70  %load2 = load i32, ptr %arrayidx1, align 4
71  %add = add nsw i32 %load1, %C
72  %mul = mul nsw i32 %load2, %D
73  %arrayidx2 = getelementptr inbounds [1024 x i32], ptr @CD, i64 0, i64 %indvars.iv
74  store i32 %add, ptr %arrayidx2, align 4
75  %arrayidx3 = getelementptr inbounds [1024 x i32], ptr @CD, i64 0, i64 %or
76  store i32 %mul, ptr %arrayidx3, align 4
77  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 2
78  %cmp = icmp slt i64 %indvars.iv.next, 1024
79  br i1 %cmp, label %for.body, label %for.end
80
81for.end:                                          ; preds = %for.body
82  ret void
83}
84
85; Check vectorization on an interleaved load group of factor 2 with narrower types  and an interleaved
86; store group of factor 2.
87
88; short AB[1024];
89; int CD[1024];
90;  void test_array_load2_store2(int C, int D) {
91;   for (int i = 0; i < 1024; i+=2) {
92;     short A = AB[i];
93;     short B = AB[i+1];
94;     CD[i] = A + C;
95;     CD[i+1] = B * D;
96;   }
97; }
98
99
100@AB_i16 = common global [1024 x i16] zeroinitializer, align 4
101
102define void @test_array_load2_i16_store2(i32 %C, i32 %D) #1 {
103; CHECK-LABEL: @test_array_load2_i16_store2(
104; CHECK-NEXT:  entry:
105; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
106; CHECK:       vector.ph:
107; CHECK-NEXT:    [[TMP0:%.*]] = call i64 @llvm.vscale.i64()
108; CHECK-NEXT:    [[TMP1:%.*]] = shl nuw nsw i64 [[TMP0]], 2
109; CHECK-NEXT:    [[TMP2:%.*]] = call <vscale x 4 x i64> @llvm.stepvector.nxv4i64()
110; CHECK-NEXT:    [[TMP3:%.*]] = shl <vscale x 4 x i64> [[TMP2]], splat (i64 1)
111; CHECK-NEXT:    [[TMP5:%.*]] = shl nuw nsw i64 [[TMP0]], 3
112; CHECK-NEXT:    [[DOTSPLATINSERT:%.*]] = insertelement <vscale x 4 x i64> poison, i64 [[TMP5]], i64 0
113; CHECK-NEXT:    [[DOTSPLAT:%.*]] = shufflevector <vscale x 4 x i64> [[DOTSPLATINSERT]], <vscale x 4 x i64> poison, <vscale x 4 x i32> zeroinitializer
114; CHECK-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <vscale x 4 x i32> poison, i32 [[C:%.*]], i64 0
115; CHECK-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <vscale x 4 x i32> [[BROADCAST_SPLATINSERT]], <vscale x 4 x i32> poison, <vscale x 4 x i32> zeroinitializer
116; CHECK-NEXT:    [[BROADCAST_SPLATINSERT2:%.*]] = insertelement <vscale x 4 x i32> poison, i32 [[D:%.*]], i64 0
117; CHECK-NEXT:    [[BROADCAST_SPLAT3:%.*]] = shufflevector <vscale x 4 x i32> [[BROADCAST_SPLATINSERT2]], <vscale x 4 x i32> poison, <vscale x 4 x i32> zeroinitializer
118; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
119; CHECK:       vector.body:
120; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
121; CHECK-NEXT:    [[VEC_IND:%.*]] = phi <vscale x 4 x i64> [ [[TMP3]], [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[VECTOR_BODY]] ]
122; CHECK-NEXT:    [[OFFSET_IDX:%.*]] = shl i64 [[INDEX]], 1
123; CHECK-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1024 x i16], ptr @AB_i16, i64 0, <vscale x 4 x i64> [[VEC_IND]]
124; CHECK-NEXT:    [[WIDE_MASKED_GATHER:%.*]] = call <vscale x 4 x i16> @llvm.masked.gather.nxv4i16.nxv4p0(<vscale x 4 x ptr> [[TMP6]], i32 2, <vscale x 4 x i1> splat (i1 true), <vscale x 4 x i16> poison)
125; CHECK-NEXT:    [[TMP7:%.*]] = or disjoint <vscale x 4 x i64> [[VEC_IND]], splat (i64 1)
126; CHECK-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [1024 x i16], ptr @AB_i16, i64 0, <vscale x 4 x i64> [[TMP7]]
127; CHECK-NEXT:    [[WIDE_MASKED_GATHER1:%.*]] = call <vscale x 4 x i16> @llvm.masked.gather.nxv4i16.nxv4p0(<vscale x 4 x ptr> [[TMP8]], i32 2, <vscale x 4 x i1> splat (i1 true), <vscale x 4 x i16> poison)
128; CHECK-NEXT:    [[TMP9:%.*]] = sext <vscale x 4 x i16> [[WIDE_MASKED_GATHER]] to <vscale x 4 x i32>
129; CHECK-NEXT:    [[TMP10:%.*]] = add nsw <vscale x 4 x i32> [[BROADCAST_SPLAT]], [[TMP9]]
130; CHECK-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [1024 x i32], ptr @CD, i64 0, i64 [[OFFSET_IDX]]
131; CHECK-NEXT:    [[TMP11:%.*]] = sext <vscale x 4 x i16> [[WIDE_MASKED_GATHER1]] to <vscale x 4 x i32>
132; CHECK-NEXT:    [[TMP12:%.*]] = mul nsw <vscale x 4 x i32> [[BROADCAST_SPLAT3]], [[TMP11]]
133; CHECK-NEXT:    [[INTERLEAVED_VEC:%.*]] = call <vscale x 8 x i32> @llvm.vector.interleave2.nxv8i32(<vscale x 4 x i32> [[TMP10]], <vscale x 4 x i32> [[TMP12]])
134; CHECK-NEXT:    store <vscale x 8 x i32> [[INTERLEAVED_VEC]], ptr [[TMP14]], align 4
135; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], [[TMP1]]
136; CHECK-NEXT:    [[VEC_IND_NEXT]] = add <vscale x 4 x i64> [[VEC_IND]], [[DOTSPLAT]]
137; CHECK-NEXT:    [[TMP16:%.*]] = icmp eq i64 [[INDEX_NEXT]], 512
138; CHECK-NEXT:    br i1 [[TMP16]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]]
139; CHECK:       middle.block:
140; CHECK-NEXT:    br i1 true, label [[FOR_END:%.*]], label [[SCALAR_PH]]
141; CHECK:       scalar.ph:
142; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
143; CHECK:       for.body:
144; CHECK-NEXT:    br i1 poison, label [[FOR_BODY]], label [[FOR_END]], !llvm.loop [[LOOP5:![0-9]+]]
145; CHECK:       for.end:
146; CHECK-NEXT:    ret void
147;
148entry:
149  br label %for.body
150
151for.body:                                         ; preds = %entry, %for.body
152  %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
153  %arrayidx = getelementptr inbounds [1024 x i16], ptr @AB_i16, i64 0, i64 %indvars.iv
154  %0 = load i16, ptr %arrayidx, align 2
155  %1 = or disjoint i64 %indvars.iv, 1
156  %arrayidx2 = getelementptr inbounds [1024 x i16], ptr @AB_i16, i64 0, i64 %1
157  %2 = load i16, ptr %arrayidx2, align 2
158  %conv = sext i16 %0 to i32
159  %add3 = add nsw i32 %conv, %C
160  %arrayidx5 = getelementptr inbounds [1024 x i32], ptr @CD, i64 0, i64 %indvars.iv
161  store i32 %add3, ptr %arrayidx5, align 4
162  %conv6 = sext i16 %2 to i32
163  %mul = mul nsw i32 %conv6, %D
164  %arrayidx9 = getelementptr inbounds [1024 x i32], ptr @CD, i64 0, i64 %1
165  store i32 %mul, ptr %arrayidx9, align 4
166  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 2
167  %cmp = icmp ult i64 %indvars.iv, 1022
168  br i1 %cmp, label %for.body, label %for.end
169
170for.end:                                 ; preds = %for.body
171  ret void
172}
173
174; Check vectorization on an interleaved load group of factor 2 and an interleaved
175; store group of factor 2 with narrower types.
176
177; int AB[1024];
178; short CD[1024];
179;  void test_array_load2_store2(int C, int D) {
180;   for (int i = 0; i < 1024; i+=2) {
181;     short A = AB[i];
182;     short B = AB[i+1];
183;     CD[i] = A + C;
184;     CD[i+1] = B * D;
185;   }
186; }
187
188
189@CD_i16 = dso_local local_unnamed_addr global [1024 x i16] zeroinitializer, align 2
190
191define void @test_array_load2_store2_i16(i32 noundef %C, i32 noundef %D) #1 {
192; CHECK-LABEL: @test_array_load2_store2_i16(
193; CHECK-NEXT:  entry:
194; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
195; CHECK:       vector.ph:
196; CHECK-NEXT:    [[TMP0:%.*]] = call i64 @llvm.vscale.i64()
197; CHECK-NEXT:    [[TMP1:%.*]] = shl nuw nsw i64 [[TMP0]], 2
198; CHECK-NEXT:    [[TMP2:%.*]] = call <vscale x 4 x i64> @llvm.stepvector.nxv4i64()
199; CHECK-NEXT:    [[TMP3:%.*]] = shl <vscale x 4 x i64> [[TMP2]], splat (i64 1)
200; CHECK-NEXT:    [[TMP5:%.*]] = shl nuw nsw i64 [[TMP0]], 3
201; CHECK-NEXT:    [[DOTSPLATINSERT:%.*]] = insertelement <vscale x 4 x i64> poison, i64 [[TMP5]], i64 0
202; CHECK-NEXT:    [[DOTSPLAT:%.*]] = shufflevector <vscale x 4 x i64> [[DOTSPLATINSERT]], <vscale x 4 x i64> poison, <vscale x 4 x i32> zeroinitializer
203; CHECK-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <vscale x 4 x i32> poison, i32 [[C:%.*]], i64 0
204; CHECK-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <vscale x 4 x i32> [[BROADCAST_SPLATINSERT]], <vscale x 4 x i32> poison, <vscale x 4 x i32> zeroinitializer
205; CHECK-NEXT:    [[BROADCAST_SPLATINSERT1:%.*]] = insertelement <vscale x 4 x i32> poison, i32 [[D:%.*]], i64 0
206; CHECK-NEXT:    [[BROADCAST_SPLAT2:%.*]] = shufflevector <vscale x 4 x i32> [[BROADCAST_SPLATINSERT1]], <vscale x 4 x i32> poison, <vscale x 4 x i32> zeroinitializer
207; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
208; CHECK:       vector.body:
209; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
210; CHECK-NEXT:    [[VEC_IND:%.*]] = phi <vscale x 4 x i64> [ [[TMP3]], [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[VECTOR_BODY]] ]
211; CHECK-NEXT:    [[OFFSET_IDX:%.*]] = shl i64 [[INDEX]], 1
212; CHECK-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [1024 x i32], ptr @AB, i64 0, i64 [[OFFSET_IDX]]
213; CHECK-NEXT:    [[WIDE_VEC:%.*]] = load <vscale x 8 x i32>, ptr [[TMP6]], align 4
214; CHECK-NEXT:    [[STRIDED_VEC:%.*]] = call { <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.vector.deinterleave2.nxv8i32(<vscale x 8 x i32> [[WIDE_VEC]])
215; CHECK-NEXT:    [[TMP7:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC]], 0
216; CHECK-NEXT:    [[TMP8:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC]], 1
217; CHECK-NEXT:    [[TMP9:%.*]] = or disjoint <vscale x 4 x i64> [[VEC_IND]], splat (i64 1)
218; CHECK-NEXT:    [[TMP10:%.*]] = add nsw <vscale x 4 x i32> [[TMP7]], [[BROADCAST_SPLAT]]
219; CHECK-NEXT:    [[TMP11:%.*]] = trunc <vscale x 4 x i32> [[TMP10]] to <vscale x 4 x i16>
220; CHECK-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [1024 x i16], ptr @CD_i16, i64 0, <vscale x 4 x i64> [[VEC_IND]]
221; CHECK-NEXT:    call void @llvm.masked.scatter.nxv4i16.nxv4p0(<vscale x 4 x i16> [[TMP11]], <vscale x 4 x ptr> [[TMP12]], i32 2, <vscale x 4 x i1> splat (i1 true))
222; CHECK-NEXT:    [[TMP13:%.*]] = mul nsw <vscale x 4 x i32> [[TMP8]], [[BROADCAST_SPLAT2]]
223; CHECK-NEXT:    [[TMP14:%.*]] = trunc <vscale x 4 x i32> [[TMP13]] to <vscale x 4 x i16>
224; CHECK-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [1024 x i16], ptr @CD_i16, i64 0, <vscale x 4 x i64> [[TMP9]]
225; CHECK-NEXT:    call void @llvm.masked.scatter.nxv4i16.nxv4p0(<vscale x 4 x i16> [[TMP14]], <vscale x 4 x ptr> [[TMP15]], i32 2, <vscale x 4 x i1> splat (i1 true))
226; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], [[TMP1]]
227; CHECK-NEXT:    [[VEC_IND_NEXT]] = add <vscale x 4 x i64> [[VEC_IND]], [[DOTSPLAT]]
228; CHECK-NEXT:    [[TMP16:%.*]] = icmp eq i64 [[INDEX_NEXT]], 512
229; CHECK-NEXT:    br i1 [[TMP16]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP6:![0-9]+]]
230; CHECK:       middle.block:
231; CHECK-NEXT:    br i1 true, label [[FOR_END:%.*]], label [[SCALAR_PH]]
232; CHECK:       scalar.ph:
233; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
234; CHECK:       for.body:
235; CHECK-NEXT:    br i1 poison, label [[FOR_BODY]], label [[FOR_END]], !llvm.loop [[LOOP7:![0-9]+]]
236; CHECK:       for.end:
237; CHECK-NEXT:    ret void
238;
239entry:
240  br label %for.body
241
242for.body:                                         ; preds = %entry, %for.body
243  %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
244  %arrayidx = getelementptr inbounds [1024 x i32], ptr @AB, i64 0, i64 %indvars.iv
245  %0 = load i32, ptr %arrayidx, align 4
246  %1 = or disjoint i64 %indvars.iv, 1
247  %arrayidx2 = getelementptr inbounds [1024 x i32], ptr @AB, i64 0, i64 %1
248  %2 = load i32, ptr %arrayidx2, align 4
249  %add3 = add nsw i32 %0, %C
250  %conv = trunc i32 %add3 to i16
251  %arrayidx5 = getelementptr inbounds [1024 x i16], ptr @CD_i16, i64 0, i64 %indvars.iv
252  store i16 %conv, ptr %arrayidx5, align 2
253  %mul = mul nsw i32 %2, %D
254  %conv6 = trunc i32 %mul to i16
255  %arrayidx9 = getelementptr inbounds [1024 x i16], ptr @CD_i16, i64 0, i64 %1
256  store i16 %conv6, ptr %arrayidx9, align 2
257  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 2
258  %cmp = icmp ult i64 %indvars.iv, 1022
259  br i1 %cmp, label %for.body, label %for.end
260
261for.end:                                 ; preds = %for.body
262  ret void
263}
264
265; Check vectorization on an interleaved load group of factor 6.
266; There is no dedicated ldN/stN so use gather instead
267
268%struct.ST6 = type { i32, i32, i32, i32, i32, i32 }
269
270define i32 @test_struct_load6(ptr %S) #1 {
271; CHECK-LABEL: @test_struct_load6(
272; CHECK-NEXT:  entry:
273; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
274; CHECK:       vector.ph:
275; CHECK-NEXT:    [[TMP0:%.*]] = call i64 @llvm.vscale.i64()
276; CHECK-NEXT:    [[TMP1:%.*]] = shl nuw nsw i64 [[TMP0]], 2
277; CHECK-NEXT:    [[TMP2:%.*]] = call <vscale x 4 x i64> @llvm.stepvector.nxv4i64()
278; CHECK-NEXT:    [[DOTSPLATINSERT:%.*]] = insertelement <vscale x 4 x i64> poison, i64 [[TMP1]], i64 0
279; CHECK-NEXT:    [[DOTSPLAT:%.*]] = shufflevector <vscale x 4 x i64> [[DOTSPLATINSERT]], <vscale x 4 x i64> poison, <vscale x 4 x i32> zeroinitializer
280; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
281; CHECK:       vector.body:
282; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
283; CHECK-NEXT:    [[VEC_IND:%.*]] = phi <vscale x 4 x i64> [ [[TMP2]], [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[VECTOR_BODY]] ]
284; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi <vscale x 4 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP16:%.*]], [[VECTOR_BODY]] ]
285; CHECK-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT_ST6:%.*]], ptr [[S:%.*]], <vscale x 4 x i64> [[VEC_IND]], i32 0
286; CHECK-NEXT:    [[WIDE_MASKED_GATHER:%.*]] = call <vscale x 4 x i32> @llvm.masked.gather.nxv4i32.nxv4p0(<vscale x 4 x ptr> [[TMP5]], i32 4, <vscale x 4 x i1> splat (i1 true), <vscale x 4 x i32> poison)
287; CHECK-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT_ST6]], ptr [[S]], <vscale x 4 x i64> [[VEC_IND]], i32 1
288; CHECK-NEXT:    [[WIDE_MASKED_GATHER1:%.*]] = call <vscale x 4 x i32> @llvm.masked.gather.nxv4i32.nxv4p0(<vscale x 4 x ptr> [[TMP6]], i32 4, <vscale x 4 x i1> splat (i1 true), <vscale x 4 x i32> poison)
289; CHECK-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT_ST6]], ptr [[S]], <vscale x 4 x i64> [[VEC_IND]], i32 2
290; CHECK-NEXT:    [[WIDE_MASKED_GATHER2:%.*]] = call <vscale x 4 x i32> @llvm.masked.gather.nxv4i32.nxv4p0(<vscale x 4 x ptr> [[TMP7]], i32 4, <vscale x 4 x i1> splat (i1 true), <vscale x 4 x i32> poison)
291; CHECK-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT_ST6]], ptr [[S]], <vscale x 4 x i64> [[VEC_IND]], i32 3
292; CHECK-NEXT:    [[WIDE_MASKED_GATHER3:%.*]] = call <vscale x 4 x i32> @llvm.masked.gather.nxv4i32.nxv4p0(<vscale x 4 x ptr> [[TMP8]], i32 4, <vscale x 4 x i1> splat (i1 true), <vscale x 4 x i32> poison)
293; CHECK-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT_ST6]], ptr [[S]], <vscale x 4 x i64> [[VEC_IND]], i32 4
294; CHECK-NEXT:    [[WIDE_MASKED_GATHER4:%.*]] = call <vscale x 4 x i32> @llvm.masked.gather.nxv4i32.nxv4p0(<vscale x 4 x ptr> [[TMP9]], i32 4, <vscale x 4 x i1> splat (i1 true), <vscale x 4 x i32> poison)
295; CHECK-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT_ST6]], ptr [[S]], <vscale x 4 x i64> [[VEC_IND]], i32 5
296; CHECK-NEXT:    [[WIDE_MASKED_GATHER5:%.*]] = call <vscale x 4 x i32> @llvm.masked.gather.nxv4i32.nxv4p0(<vscale x 4 x ptr> [[TMP10]], i32 4, <vscale x 4 x i1> splat (i1 true), <vscale x 4 x i32> poison)
297; CHECK-NEXT:    [[TMP11:%.*]] = add <vscale x 4 x i32> [[WIDE_MASKED_GATHER]], [[VEC_PHI]]
298; CHECK-NEXT:    [[TMP12:%.*]] = add <vscale x 4 x i32> [[TMP11]], [[WIDE_MASKED_GATHER2]]
299; CHECK-NEXT:    [[TMP13:%.*]] = add <vscale x 4 x i32> [[WIDE_MASKED_GATHER1]], [[WIDE_MASKED_GATHER3]]
300; CHECK-NEXT:    [[TMP14:%.*]] = add <vscale x 4 x i32> [[TMP13]], [[WIDE_MASKED_GATHER4]]
301; CHECK-NEXT:    [[TMP15:%.*]] = add <vscale x 4 x i32> [[TMP14]], [[WIDE_MASKED_GATHER5]]
302; CHECK-NEXT:    [[TMP16]] = sub <vscale x 4 x i32> [[TMP12]], [[TMP15]]
303; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], [[TMP1]]
304; CHECK-NEXT:    [[VEC_IND_NEXT]] = add <vscale x 4 x i64> [[VEC_IND]], [[DOTSPLAT]]
305; CHECK-NEXT:    [[TMP17:%.*]] = icmp eq i64 [[INDEX_NEXT]], 1024
306; CHECK-NEXT:    br i1 [[TMP17]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP8:![0-9]+]]
307; CHECK:       middle.block:
308; CHECK-NEXT:    [[TMP18:%.*]] = call i32 @llvm.vector.reduce.add.nxv4i32(<vscale x 4 x i32> [[TMP16]])
309; CHECK-NEXT:    br i1 true, label [[FOR_COND_CLEANUP:%.*]], label [[SCALAR_PH]]
310; CHECK:       scalar.ph:
311; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
312; CHECK:       for.body:
313; CHECK-NEXT:    br i1 poison, label [[FOR_COND_CLEANUP]], label [[FOR_BODY]], !llvm.loop [[LOOP9:![0-9]+]]
314; CHECK:       for.cond.cleanup:
315; CHECK-NEXT:    [[SUB14_LCSSA:%.*]] = phi i32 [ poison, [[FOR_BODY]] ], [ [[TMP18]], [[MIDDLE_BLOCK]] ]
316; CHECK-NEXT:    ret i32 [[SUB14_LCSSA]]
317;
318entry:
319  br label %for.body
320
321for.body:                                         ; preds = %entry, %for.body
322  %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
323  %r.041 = phi i32 [ 0, %entry ], [ %sub14, %for.body ]
324  %x = getelementptr inbounds %struct.ST6, ptr %S, i64 %indvars.iv, i32 0
325  %0 = load i32, ptr %x, align 4
326  %y = getelementptr inbounds %struct.ST6, ptr %S, i64 %indvars.iv, i32 1
327  %1 = load i32, ptr %y, align 4
328  %z = getelementptr inbounds %struct.ST6, ptr %S, i64 %indvars.iv, i32 2
329  %2 = load i32, ptr %z, align 4
330  %w = getelementptr inbounds %struct.ST6, ptr %S, i64 %indvars.iv, i32 3
331  %3 = load i32, ptr %w, align 4
332  %a = getelementptr inbounds %struct.ST6, ptr %S, i64 %indvars.iv, i32 4
333  %4 = load i32, ptr %a, align 4
334  %b = getelementptr inbounds %struct.ST6, ptr %S, i64 %indvars.iv, i32 5
335  %5 = load i32, ptr %b, align 4
336  %.neg36 = add i32 %0, %r.041
337  %6 = add i32 %.neg36, %2
338  %7 = add i32 %1, %3
339  %8 = add i32 %7, %4
340  %9 = add i32 %8, %5
341  %sub14 = sub i32 %6, %9
342  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
343  %exitcond.not = icmp eq i64 %indvars.iv.next, 1024
344  br i1 %exitcond.not, label %for.cond.cleanup, label %for.body
345
346for.cond.cleanup:                                 ; preds = %for.body
347  %sub14.lcssa = phi i32 [ %sub14, %for.body ]
348  ret i32 %sub14.lcssa
349}
350
351
352; Check vectorization on a reverse interleaved load group of factor 2 and
353; a reverse interleaved store group of factor 2.
354
355; struct ST2 {
356;  int x;
357;  int y;
358; };
359;
360; void test_reversed_load2_store2(struct ST2 *A, struct ST2 *B) {
361;   for (int i = 1023; i >= 0; i--) {
362;     int a = A[i].x + i;  // interleaved load of index 0
363;     int b = A[i].y - i;  // interleaved load of index 1
364;     B[i].x = a;          // interleaved store of index 0
365;     B[i].y = b;          // interleaved store of index 1
366;   }
367; }
368
369
370%struct.ST2 = type { i32, i32 }
371
372define void @test_reversed_load2_store2(ptr noalias nocapture readonly %A, ptr noalias nocapture %B) #1 {
373; CHECK-LABEL: @test_reversed_load2_store2(
374; CHECK-NEXT:  entry:
375; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
376; CHECK:       vector.ph:
377; CHECK-NEXT:    [[TMP0:%.*]] = call i64 @llvm.vscale.i64()
378; CHECK-NEXT:    [[TMP1:%.*]] = shl nuw nsw i64 [[TMP0]], 2
379; CHECK-NEXT:    [[TMP2:%.*]] = call <vscale x 4 x i32> @llvm.stepvector.nxv4i32()
380; CHECK-NEXT:    [[INDUCTION:%.*]] = sub <vscale x 4 x i32> splat (i32 1023), [[TMP2]]
381; CHECK-NEXT:    [[TMP3:%.*]] = trunc nuw nsw i64 [[TMP1]] to i32
382; CHECK-NEXT:    [[DOTNEG:%.*]] = sub nsw i32 0, [[TMP3]]
383; CHECK-NEXT:    [[DOTSPLATINSERT:%.*]] = insertelement <vscale x 4 x i32> poison, i32 [[DOTNEG]], i64 0
384; CHECK-NEXT:    [[DOTSPLAT:%.*]] = shufflevector <vscale x 4 x i32> [[DOTSPLATINSERT]], <vscale x 4 x i32> poison, <vscale x 4 x i32> zeroinitializer
385; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
386; CHECK:       vector.body:
387; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
388; CHECK-NEXT:    [[VEC_IND:%.*]] = phi <vscale x 4 x i32> [ [[INDUCTION]], [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[VECTOR_BODY]] ]
389; CHECK-NEXT:    [[OFFSET_IDX:%.*]] = sub i64 1023, [[INDEX]]
390; CHECK-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT_ST2:%.*]], ptr [[A:%.*]], i64 [[OFFSET_IDX]], i32 0
391; CHECK-NEXT:    [[TMP5:%.*]] = call i32 @llvm.vscale.i32()
392; CHECK-NEXT:    [[TMP6:%.*]] = shl nuw nsw i32 [[TMP5]], 3
393; CHECK-NEXT:    [[TMP7:%.*]] = sub nsw i32 2, [[TMP6]]
394; CHECK-NEXT:    [[TMP8:%.*]] = sext i32 [[TMP7]] to i64
395; CHECK-NEXT:    [[TMP9:%.*]] = getelementptr inbounds i32, ptr [[TMP4]], i64 [[TMP8]]
396; CHECK-NEXT:    [[WIDE_VEC:%.*]] = load <vscale x 8 x i32>, ptr [[TMP9]], align 4
397; CHECK-NEXT:    [[STRIDED_VEC:%.*]] = call { <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.vector.deinterleave2.nxv8i32(<vscale x 8 x i32> [[WIDE_VEC]])
398; CHECK-NEXT:    [[TMP10:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC]], 0
399; CHECK-NEXT:    [[REVERSE:%.*]] = call <vscale x 4 x i32> @llvm.vector.reverse.nxv4i32(<vscale x 4 x i32> [[TMP10]])
400; CHECK-NEXT:    [[TMP11:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC]], 1
401; CHECK-NEXT:    [[REVERSE1:%.*]] = call <vscale x 4 x i32> @llvm.vector.reverse.nxv4i32(<vscale x 4 x i32> [[TMP11]])
402; CHECK-NEXT:    [[TMP12:%.*]] = add nsw <vscale x 4 x i32> [[REVERSE]], [[VEC_IND]]
403; CHECK-NEXT:    [[TMP13:%.*]] = sub nsw <vscale x 4 x i32> [[REVERSE1]], [[VEC_IND]]
404; CHECK-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT_ST2]], ptr [[B:%.*]], i64 [[OFFSET_IDX]], i32 0
405; CHECK-NEXT:    [[TMP15:%.*]] = call i32 @llvm.vscale.i32()
406; CHECK-NEXT:    [[TMP16:%.*]] = shl nuw nsw i32 [[TMP15]], 3
407; CHECK-NEXT:    [[TMP17:%.*]] = sub nsw i32 2, [[TMP16]]
408; CHECK-NEXT:    [[TMP18:%.*]] = sext i32 [[TMP17]] to i64
409; CHECK-NEXT:    [[TMP19:%.*]] = getelementptr inbounds i32, ptr [[TMP14]], i64 [[TMP18]]
410; CHECK-NEXT:    [[REVERSE2:%.*]] = call <vscale x 4 x i32> @llvm.vector.reverse.nxv4i32(<vscale x 4 x i32> [[TMP12]])
411; CHECK-NEXT:    [[REVERSE3:%.*]] = call <vscale x 4 x i32> @llvm.vector.reverse.nxv4i32(<vscale x 4 x i32> [[TMP13]])
412; CHECK-NEXT:    [[INTERLEAVED_VEC:%.*]] = call <vscale x 8 x i32> @llvm.vector.interleave2.nxv8i32(<vscale x 4 x i32> [[REVERSE2]], <vscale x 4 x i32> [[REVERSE3]])
413; CHECK-NEXT:    store <vscale x 8 x i32> [[INTERLEAVED_VEC]], ptr [[TMP19]], align 4
414; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], [[TMP1]]
415; CHECK-NEXT:    [[VEC_IND_NEXT]] = add <vscale x 4 x i32> [[VEC_IND]], [[DOTSPLAT]]
416; CHECK-NEXT:    [[TMP20:%.*]] = icmp eq i64 [[INDEX_NEXT]], 1024
417; CHECK-NEXT:    br i1 [[TMP20]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP10:![0-9]+]]
418; CHECK:       middle.block:
419; CHECK-NEXT:    br i1 true, label [[FOR_COND_CLEANUP:%.*]], label [[SCALAR_PH]]
420; CHECK:       scalar.ph:
421; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
422; CHECK:       for.cond.cleanup:
423; CHECK-NEXT:    ret void
424; CHECK:       for.body:
425; CHECK-NEXT:    br i1 poison, label [[FOR_BODY]], label [[FOR_COND_CLEANUP]], !llvm.loop [[LOOP11:![0-9]+]]
426;
427entry:
428  br label %for.body
429
430for.cond.cleanup:                                 ; preds = %for.body
431  ret void
432
433for.body:                                         ; preds = %for.body, %entry
434  %indvars.iv = phi i64 [ 1023, %entry ], [ %indvars.iv.next, %for.body ]
435  %x = getelementptr inbounds %struct.ST2, ptr %A, i64 %indvars.iv, i32 0
436  %load1 = load i32, ptr %x, align 4
437  %trunc = trunc i64 %indvars.iv to i32
438  %add = add nsw i32 %load1, %trunc
439  %y = getelementptr inbounds %struct.ST2, ptr %A, i64 %indvars.iv, i32 1
440  %load2 = load i32, ptr %y, align 4
441  %sub = sub nsw i32 %load2, %trunc
442  %x5 = getelementptr inbounds %struct.ST2, ptr %B, i64 %indvars.iv, i32 0
443  store i32 %add, ptr %x5, align 4
444  %y8 = getelementptr inbounds %struct.ST2, ptr %B, i64 %indvars.iv, i32 1
445  store i32 %sub, ptr %y8, align 4
446  %indvars.iv.next = add nsw i64 %indvars.iv, -1
447  %cmp = icmp sgt i64 %indvars.iv, 0
448  br i1 %cmp, label %for.body, label %for.cond.cleanup
449}
450
451; Check vectorization on an interleaved load group of factor 2 with 1 gap
452; (missing the load of odd elements). Because the vectorized loop would
453; speculatively access memory out-of-bounds, we must execute at least one
454; iteration of the scalar loop.
455
456; void even_load_static_tc(int *A, int *B) {
457;  for (unsigned i = 0; i < 1024; i+=2)
458;     B[i/2] = A[i] * 2;
459; }
460
461
462define void @even_load_static_tc(ptr noalias nocapture readonly %A, ptr noalias nocapture %B) #1 {
463; CHECK-LABEL: @even_load_static_tc(
464; CHECK-NEXT:  entry:
465; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
466; CHECK:       vector.ph:
467; CHECK-NEXT:    [[TMP0:%.*]] = call i64 @llvm.vscale.i64()
468; CHECK-NEXT:    [[TMP1:%.*]] = shl nuw nsw i64 [[TMP0]], 2
469; CHECK-NEXT:    [[N_VEC:%.*]] = sub nuw nsw i64 512, [[TMP1]]
470; CHECK-NEXT:    [[TMP2:%.*]] = call i64 @llvm.vscale.i64()
471; CHECK-NEXT:    [[TMP3:%.*]] = shl nuw nsw i64 [[TMP2]], 2
472; CHECK-NEXT:    [[IND_END:%.*]] = shl nuw nsw i64 [[N_VEC]], 1
473; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
474; CHECK:       vector.body:
475; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
476; CHECK-NEXT:    [[DOTIDX:%.*]] = shl i64 [[INDEX]], 3
477; CHECK-NEXT:    [[TMP4:%.*]] = getelementptr inbounds i8, ptr [[A:%.*]], i64 [[DOTIDX]]
478; CHECK-NEXT:    [[WIDE_VEC:%.*]] = load <vscale x 8 x i32>, ptr [[TMP4]], align 4
479; CHECK-NEXT:    [[STRIDED_VEC:%.*]] = call { <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.vector.deinterleave2.nxv8i32(<vscale x 8 x i32> [[WIDE_VEC]])
480; CHECK-NEXT:    [[TMP5:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC]], 0
481; CHECK-NEXT:    [[TMP6:%.*]] = shl nsw <vscale x 4 x i32> [[TMP5]], splat (i32 1)
482; CHECK-NEXT:    [[TMP7:%.*]] = and i64 [[INDEX]], 9223372036854775804
483; CHECK-NEXT:    [[TMP8:%.*]] = getelementptr inbounds nuw i32, ptr [[B:%.*]], i64 [[TMP7]]
484; CHECK-NEXT:    store <vscale x 4 x i32> [[TMP6]], ptr [[TMP8]], align 4
485; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], [[TMP3]]
486; CHECK-NEXT:    [[TMP9:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
487; CHECK-NEXT:    br i1 [[TMP9]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP12:![0-9]+]]
488; CHECK:       middle.block:
489; CHECK-NEXT:    br label [[SCALAR_PH]]
490; CHECK:       scalar.ph:
491; CHECK-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ [[IND_END]], [[MIDDLE_BLOCK]] ], [ poison, [[ENTRY:%.*]] ]
492; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
493; CHECK:       for.cond.cleanup:
494; CHECK-NEXT:    ret void
495; CHECK:       for.body:
496; CHECK-NEXT:    [[INDVARS_IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ]
497; CHECK-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[INDVARS_IV]]
498; CHECK-NEXT:    [[LOAD:%.*]] = load i32, ptr [[ARRAYIDX]], align 4
499; CHECK-NEXT:    [[MUL:%.*]] = shl nsw i32 [[LOAD]], 1
500; CHECK-NEXT:    [[LSHR:%.*]] = lshr exact i64 [[INDVARS_IV]], 1
501; CHECK-NEXT:    [[ARRAYIDX2:%.*]] = getelementptr inbounds nuw i32, ptr [[B]], i64 [[LSHR]]
502; CHECK-NEXT:    store i32 [[MUL]], ptr [[ARRAYIDX2]], align 4
503; CHECK-NEXT:    [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 2
504; CHECK-NEXT:    [[CMP:%.*]] = icmp ult i64 [[INDVARS_IV]], 1022
505; CHECK-NEXT:    br i1 [[CMP]], label [[FOR_BODY]], label [[FOR_COND_CLEANUP:%.*]], !llvm.loop [[LOOP13:![0-9]+]]
506;
507entry:
508  br label %for.body
509
510for.cond.cleanup:                                 ; preds = %for.body
511  ret void
512
513for.body:                                         ; preds = %for.body, %entry
514  %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
515  %arrayidx = getelementptr inbounds i32, ptr %A, i64 %indvars.iv
516  %load = load i32, ptr %arrayidx, align 4
517  %mul = shl nsw i32 %load, 1
518  %lshr = lshr exact i64 %indvars.iv, 1
519  %arrayidx2 = getelementptr inbounds i32, ptr %B, i64 %lshr
520  store i32 %mul, ptr %arrayidx2, align 4
521  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 2
522  %cmp = icmp ult i64 %indvars.iv.next, 1024
523  br i1 %cmp, label %for.body, label %for.cond.cleanup
524}
525
526; Check vectorization on an interleaved load group of factor 2 with 1 gap
527; (missing the load of odd elements). Because the vectorized loop would
528; speculatively access memory out-of-bounds, we must execute at least one
529; iteration of the scalar loop.
530
531; void even_load_dynamic_tc(int *A, int *B, unsigned N) {
532;  for (unsigned i = 0; i < N; i+=2)
533;     B[i/2] = A[i] * 2;
534; }
535
536
537define void @even_load_dynamic_tc(ptr noalias nocapture readonly %A, ptr noalias nocapture %B, i64 %N) #1 {
538; CHECK-LABEL: @even_load_dynamic_tc(
539; CHECK-NEXT:  entry:
540; CHECK-NEXT:    [[UMAX:%.*]] = call i64 @llvm.umax.i64(i64 [[N:%.*]], i64 2)
541; CHECK-NEXT:    [[TMP0:%.*]] = add i64 [[UMAX]], -1
542; CHECK-NEXT:    [[TMP1:%.*]] = lshr i64 [[TMP0]], 1
543; CHECK-NEXT:    [[TMP2:%.*]] = call i64 @llvm.vscale.i64()
544; CHECK-NEXT:    [[TMP3:%.*]] = shl nuw nsw i64 [[TMP2]], 2
545; CHECK-NEXT:    [[MIN_ITERS_CHECK_NOT_NOT:%.*]] = icmp samesign ult i64 [[TMP1]], [[TMP3]]
546; CHECK-NEXT:    br i1 [[MIN_ITERS_CHECK_NOT_NOT]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
547; CHECK:       vector.ph:
548; CHECK-NEXT:    [[TMP4:%.*]] = add nuw i64 [[TMP1]], 1
549; CHECK-NEXT:    [[TMP5:%.*]] = call i64 @llvm.vscale.i64()
550; CHECK-NEXT:    [[TMP6:%.*]] = shl nuw nsw i64 [[TMP5]], 2
551; CHECK-NEXT:    [[TMP7:%.*]] = add nsw i64 [[TMP6]], -1
552; CHECK-NEXT:    [[N_MOD_VF:%.*]] = and i64 [[TMP4]], [[TMP7]]
553; CHECK-NEXT:    [[TMP8:%.*]] = icmp eq i64 [[N_MOD_VF]], 0
554; CHECK-NEXT:    [[TMP9:%.*]] = select i1 [[TMP8]], i64 [[TMP6]], i64 [[N_MOD_VF]]
555; CHECK-NEXT:    [[N_VEC:%.*]] = sub i64 [[TMP4]], [[TMP9]]
556; CHECK-NEXT:    [[TMP10:%.*]] = call i64 @llvm.vscale.i64()
557; CHECK-NEXT:    [[TMP11:%.*]] = shl nuw nsw i64 [[TMP10]], 2
558; CHECK-NEXT:    [[IND_END:%.*]] = shl i64 [[N_VEC]], 1
559; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
560; CHECK:       vector.body:
561; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
562; CHECK-NEXT:    [[DOTIDX:%.*]] = shl i64 [[INDEX]], 3
563; CHECK-NEXT:    [[TMP12:%.*]] = getelementptr inbounds i8, ptr [[A:%.*]], i64 [[DOTIDX]]
564; CHECK-NEXT:    [[WIDE_VEC:%.*]] = load <vscale x 8 x i32>, ptr [[TMP12]], align 4
565; CHECK-NEXT:    [[STRIDED_VEC:%.*]] = call { <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.vector.deinterleave2.nxv8i32(<vscale x 8 x i32> [[WIDE_VEC]])
566; CHECK-NEXT:    [[TMP13:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC]], 0
567; CHECK-NEXT:    [[TMP14:%.*]] = shl nsw <vscale x 4 x i32> [[TMP13]], splat (i32 1)
568; CHECK-NEXT:    [[TMP15:%.*]] = and i64 [[INDEX]], 9223372036854775804
569; CHECK-NEXT:    [[TMP16:%.*]] = getelementptr inbounds nuw i32, ptr [[B:%.*]], i64 [[TMP15]]
570; CHECK-NEXT:    store <vscale x 4 x i32> [[TMP14]], ptr [[TMP16]], align 4
571; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], [[TMP11]]
572; CHECK-NEXT:    [[TMP17:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
573; CHECK-NEXT:    br i1 [[TMP17]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP14:![0-9]+]]
574; CHECK:       middle.block:
575; CHECK-NEXT:    br label [[SCALAR_PH]]
576; CHECK:       scalar.ph:
577; CHECK-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ [[IND_END]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ]
578; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
579; CHECK:       for.cond.cleanup:
580; CHECK-NEXT:    ret void
581; CHECK:       for.body:
582; CHECK-NEXT:    [[INDVARS_IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ]
583; CHECK-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[INDVARS_IV]]
584; CHECK-NEXT:    [[LOAD:%.*]] = load i32, ptr [[ARRAYIDX]], align 4
585; CHECK-NEXT:    [[MUL:%.*]] = shl nsw i32 [[LOAD]], 1
586; CHECK-NEXT:    [[LSHR:%.*]] = lshr exact i64 [[INDVARS_IV]], 1
587; CHECK-NEXT:    [[ARRAYIDX2:%.*]] = getelementptr inbounds nuw i32, ptr [[B]], i64 [[LSHR]]
588; CHECK-NEXT:    store i32 [[MUL]], ptr [[ARRAYIDX2]], align 4
589; CHECK-NEXT:    [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 2
590; CHECK-NEXT:    [[CMP:%.*]] = icmp ult i64 [[INDVARS_IV_NEXT]], [[N]]
591; CHECK-NEXT:    br i1 [[CMP]], label [[FOR_BODY]], label [[FOR_COND_CLEANUP:%.*]], !llvm.loop [[LOOP15:![0-9]+]]
592;
593entry:
594  br label %for.body
595
596for.cond.cleanup:                                 ; preds = %for.body
597  ret void
598
599for.body:                                         ; preds = %for.body, %entry
600  %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
601  %arrayidx = getelementptr inbounds i32, ptr %A, i64 %indvars.iv
602  %load = load i32, ptr %arrayidx, align 4
603  %mul = shl nsw i32 %load, 1
604  %lshr = lshr exact i64 %indvars.iv, 1
605  %arrayidx2 = getelementptr inbounds i32, ptr %B, i64 %lshr
606  store i32 %mul, ptr %arrayidx2, align 4
607  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 2
608  %cmp = icmp ult i64 %indvars.iv.next, %N
609  br i1 %cmp, label %for.body, label %for.cond.cleanup
610}
611
612; Check vectorization on a reverse interleaved load group of factor 2 with 1
613; gap and a reverse interleaved store group of factor 2. The interleaved load
614; group should be removed since it has a gap and is reverse.
615
616; struct pair {
617;  int x;
618;  int y;
619; };
620;
621; void load_gap_reverse(struct pair *P1, struct pair *P2, int X) {
622;   for (int i = 1023; i >= 0; i--) {
623;     int a = X + i;
624;     int b = A[i].y - i;
625;     B[i].x = a;
626;     B[i].y = b;
627;   }
628; }
629
630%pair = type { i64, i64 }
631define void @load_gap_reverse(ptr noalias nocapture readonly %P1, ptr noalias nocapture readonly %P2, i64 %X) #1 {
632; CHECK-LABEL: @load_gap_reverse(
633; CHECK-NEXT:  entry:
634; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
635; CHECK:       vector.ph:
636; CHECK-NEXT:    [[TMP0:%.*]] = call i64 @llvm.vscale.i64()
637; CHECK-NEXT:    [[TMP1:%.*]] = shl nuw nsw i64 [[TMP0]], 2
638; CHECK-NEXT:    [[TMP2:%.*]] = call <vscale x 4 x i64> @llvm.stepvector.nxv4i64()
639; CHECK-NEXT:    [[INDUCTION:%.*]] = sub <vscale x 4 x i64> splat (i64 1023), [[TMP2]]
640; CHECK-NEXT:    [[DOTNEG:%.*]] = sub nsw i64 0, [[TMP1]]
641; CHECK-NEXT:    [[DOTSPLATINSERT:%.*]] = insertelement <vscale x 4 x i64> poison, i64 [[DOTNEG]], i64 0
642; CHECK-NEXT:    [[DOTSPLAT:%.*]] = shufflevector <vscale x 4 x i64> [[DOTSPLATINSERT]], <vscale x 4 x i64> poison, <vscale x 4 x i32> zeroinitializer
643; CHECK-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <vscale x 4 x i64> poison, i64 [[X:%.*]], i64 0
644; CHECK-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <vscale x 4 x i64> [[BROADCAST_SPLATINSERT]], <vscale x 4 x i64> poison, <vscale x 4 x i32> zeroinitializer
645; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
646; CHECK:       vector.body:
647; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
648; CHECK-NEXT:    [[VEC_IND:%.*]] = phi <vscale x 4 x i64> [ [[INDUCTION]], [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[VECTOR_BODY]] ]
649; CHECK-NEXT:    [[TMP4:%.*]] = add nsw <vscale x 4 x i64> [[BROADCAST_SPLAT]], [[VEC_IND]]
650; CHECK-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[PAIR:%.*]], ptr [[P1:%.*]], <vscale x 4 x i64> [[VEC_IND]], i32 0
651; CHECK-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[PAIR]], ptr [[P2:%.*]], <vscale x 4 x i64> [[VEC_IND]], i32 1
652; CHECK-NEXT:    [[WIDE_MASKED_GATHER:%.*]] = call <vscale x 4 x i64> @llvm.masked.gather.nxv4i64.nxv4p0(<vscale x 4 x ptr> [[TMP6]], i32 8, <vscale x 4 x i1> splat (i1 true), <vscale x 4 x i64> poison)
653; CHECK-NEXT:    [[TMP7:%.*]] = sub nsw <vscale x 4 x i64> [[WIDE_MASKED_GATHER]], [[VEC_IND]]
654; CHECK-NEXT:    call void @llvm.masked.scatter.nxv4i64.nxv4p0(<vscale x 4 x i64> [[TMP4]], <vscale x 4 x ptr> [[TMP5]], i32 8, <vscale x 4 x i1> splat (i1 true))
655; CHECK-NEXT:    call void @llvm.masked.scatter.nxv4i64.nxv4p0(<vscale x 4 x i64> [[TMP7]], <vscale x 4 x ptr> [[TMP6]], i32 8, <vscale x 4 x i1> splat (i1 true))
656; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], [[TMP1]]
657; CHECK-NEXT:    [[VEC_IND_NEXT]] = add <vscale x 4 x i64> [[VEC_IND]], [[DOTSPLAT]]
658; CHECK-NEXT:    [[TMP8:%.*]] = icmp eq i64 [[INDEX_NEXT]], 1024
659; CHECK-NEXT:    br i1 [[TMP8]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP16:![0-9]+]]
660; CHECK:       middle.block:
661; CHECK-NEXT:    br i1 true, label [[FOR_EXIT:%.*]], label [[SCALAR_PH]]
662; CHECK:       scalar.ph:
663; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
664; CHECK:       for.body:
665; CHECK-NEXT:    br i1 poison, label [[FOR_BODY]], label [[FOR_EXIT]], !llvm.loop [[LOOP17:![0-9]+]]
666; CHECK:       for.exit:
667; CHECK-NEXT:    ret void
668;
669entry:
670  br label %for.body
671
672for.body:
673  %i = phi i64 [ 1023, %entry ], [ %i.next, %for.body ]
674  %0 = add nsw i64 %X, %i
675  %1 = getelementptr inbounds %pair, ptr %P1, i64 %i, i32 0
676  %2 = getelementptr inbounds %pair, ptr %P2, i64 %i, i32 1
677  %3 = load i64, ptr %2, align 8
678  %4 = sub nsw i64 %3, %i
679  store i64 %0, ptr %1, align 8
680  store i64 %4, ptr %2, align 8
681  %i.next = add nsw i64 %i, -1
682  %cond = icmp sgt i64 %i, 0
683  br i1 %cond, label %for.body, label %for.exit
684
685for.exit:
686  ret void
687}
688
689; Check vectorization on interleaved access groups identified from mixed
690; loads/stores.
691; void mixed_load2_store2(int *A, int *B) {
692;   for (unsigned i = 0; i < 1024; i+=2)  {
693;     B[i] = A[i] * A[i+1];
694;     B[i+1] = A[i] + A[i+1];
695;   }
696; }
697
698
699define void @mixed_load2_store2(ptr noalias nocapture readonly %A, ptr noalias nocapture %B) #1 {
700; CHECK-LABEL: @mixed_load2_store2(
701; CHECK-NEXT:  entry:
702; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
703; CHECK:       vector.ph:
704; CHECK-NEXT:    [[TMP0:%.*]] = call i64 @llvm.vscale.i64()
705; CHECK-NEXT:    [[TMP1:%.*]] = shl nuw nsw i64 [[TMP0]], 2
706; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
707; CHECK:       vector.body:
708; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
709; CHECK-NEXT:    [[OFFSET_IDX:%.*]] = shl i64 [[INDEX]], 1
710; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr inbounds i32, ptr [[A:%.*]], i64 [[OFFSET_IDX]]
711; CHECK-NEXT:    [[WIDE_VEC:%.*]] = load <vscale x 8 x i32>, ptr [[TMP2]], align 4
712; CHECK-NEXT:    [[STRIDED_VEC:%.*]] = call { <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.vector.deinterleave2.nxv8i32(<vscale x 8 x i32> [[WIDE_VEC]])
713; CHECK-NEXT:    [[TMP3:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC]], 0
714; CHECK-NEXT:    [[TMP4:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC]], 1
715; CHECK-NEXT:    [[TMP6:%.*]] = mul nsw <vscale x 4 x i32> [[TMP4]], [[TMP3]]
716; CHECK-NEXT:    [[TMP10:%.*]] = getelementptr inbounds i32, ptr [[B:%.*]], i64 [[OFFSET_IDX]]
717; CHECK-NEXT:    [[STRIDED_VEC2:%.*]] = call { <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.vector.deinterleave2.nxv8i32(<vscale x 8 x i32> [[WIDE_VEC]])
718; CHECK-NEXT:    [[TMP7:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC2]], 0
719; CHECK-NEXT:    [[TMP8:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC2]], 1
720; CHECK-NEXT:    [[TMP9:%.*]] = add nsw <vscale x 4 x i32> [[TMP8]], [[TMP7]]
721; CHECK-NEXT:    [[INTERLEAVED_VEC:%.*]] = call <vscale x 8 x i32> @llvm.vector.interleave2.nxv8i32(<vscale x 4 x i32> [[TMP6]], <vscale x 4 x i32> [[TMP9]])
722; CHECK-NEXT:    store <vscale x 8 x i32> [[INTERLEAVED_VEC]], ptr [[TMP10]], align 4
723; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], [[TMP1]]
724; CHECK-NEXT:    [[TMP12:%.*]] = icmp eq i64 [[INDEX_NEXT]], 512
725; CHECK-NEXT:    br i1 [[TMP12]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP18:![0-9]+]]
726; CHECK:       middle.block:
727; CHECK-NEXT:    br i1 true, label [[FOR_COND_CLEANUP:%.*]], label [[SCALAR_PH]]
728; CHECK:       scalar.ph:
729; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
730; CHECK:       for.cond.cleanup:
731; CHECK-NEXT:    ret void
732; CHECK:       for.body:
733; CHECK-NEXT:    br i1 poison, label [[FOR_BODY]], label [[FOR_COND_CLEANUP]], !llvm.loop [[LOOP19:![0-9]+]]
734;
735entry:
736  br label %for.body
737
738for.cond.cleanup:                                 ; preds = %for.body
739  ret void
740
741for.body:                                         ; preds = %for.body, %entry
742  %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
743  %arrayidx = getelementptr inbounds i32, ptr %A, i64 %indvars.iv
744  %load1 = load i32, ptr %arrayidx, align 4
745  %or = or disjoint i64 %indvars.iv, 1
746  %arrayidx2 = getelementptr inbounds i32, ptr %A, i64 %or
747  %load2 = load i32, ptr %arrayidx2, align 4
748  %mul = mul nsw i32 %load2, %load1
749  %arrayidx4 = getelementptr inbounds i32, ptr %B, i64 %indvars.iv
750  store i32 %mul, ptr %arrayidx4, align 4
751  %load3 = load i32, ptr %arrayidx, align 4
752  %load4 = load i32, ptr %arrayidx2, align 4
753  %add10 = add nsw i32 %load4, %load3
754  %arrayidx13 = getelementptr inbounds i32, ptr %B, i64 %or
755  store i32 %add10, ptr %arrayidx13, align 4
756  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 2
757  %cmp = icmp ult i64 %indvars.iv.next, 1024
758  br i1 %cmp, label %for.body, label %for.cond.cleanup
759}
760
761
762; Check vectorization on interleaved access groups with members having different
763; kinds of type.
764
765; struct IntFloat {
766;   int a;
767;   float b;
768; };
769;
770; int SA;
771; float SB;
772;
773; void int_float_struct(struct IntFloat *A) {
774;   int SumA;
775;   float SumB;
776;   for (unsigned i = 0; i < 1024; i++)  {
777;     SumA += A[i].a;
778;     SumB += A[i].b;
779;   }
780;   SA = SumA;
781;   SB = SumB;
782; }
783
784
785%struct.IntFloat = type { i32, float }
786
787@SA = common global i32 0, align 4
788@SB = common global float 0.000000e+00, align 4
789
790define void @int_float_struct(ptr nocapture readonly %p) #0 {
791; CHECK-LABEL: @int_float_struct(
792; CHECK-NEXT:  entry:
793; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
794; CHECK:       vector.ph:
795; CHECK-NEXT:    [[TMP0:%.*]] = call i64 @llvm.vscale.i64()
796; CHECK-NEXT:    [[TMP1:%.*]] = shl nuw nsw i64 [[TMP0]], 2
797; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
798; CHECK:       vector.body:
799; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
800; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi <vscale x 4 x float> [ insertelement (<vscale x 4 x float> zeroinitializer, float undef, i32 0), [[VECTOR_PH]] ], [ [[TMP7:%.*]], [[VECTOR_BODY]] ]
801; CHECK-NEXT:    [[VEC_PHI1:%.*]] = phi <vscale x 4 x i32> [ insertelement (<vscale x 4 x i32> zeroinitializer, i32 undef, i32 0), [[VECTOR_PH]] ], [ [[TMP6:%.*]], [[VECTOR_BODY]] ]
802; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr inbounds [[STRUCT_INTFLOAT:%.*]], ptr [[P:%.*]], i64 [[INDEX]], i32 0
803; CHECK-NEXT:    [[WIDE_VEC:%.*]] = load <vscale x 8 x i32>, ptr [[TMP2]], align 4
804; CHECK-NEXT:    [[STRIDED_VEC:%.*]] = call { <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.vector.deinterleave2.nxv8i32(<vscale x 8 x i32> [[WIDE_VEC]])
805; CHECK-NEXT:    [[TMP3:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC]], 0
806; CHECK-NEXT:    [[TMP4:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC]], 1
807; CHECK-NEXT:    [[TMP5:%.*]] = bitcast <vscale x 4 x i32> [[TMP4]] to <vscale x 4 x float>
808; CHECK-NEXT:    [[TMP6]] = add <vscale x 4 x i32> [[TMP3]], [[VEC_PHI1]]
809; CHECK-NEXT:    [[TMP7]] = fadd fast <vscale x 4 x float> [[VEC_PHI]], [[TMP5]]
810; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], [[TMP1]]
811; CHECK-NEXT:    [[TMP8:%.*]] = icmp eq i64 [[INDEX_NEXT]], 1024
812; CHECK-NEXT:    br i1 [[TMP8]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP20:![0-9]+]]
813; CHECK:       middle.block:
814; CHECK-NEXT:    [[TMP9:%.*]] = call fast float @llvm.vector.reduce.fadd.nxv4f32(float 0.000000e+00, <vscale x 4 x float> [[TMP7]])
815; CHECK-NEXT:    [[TMP10:%.*]] = call i32 @llvm.vector.reduce.add.nxv4i32(<vscale x 4 x i32> [[TMP6]])
816; CHECK-NEXT:    br i1 true, label [[FOR_COND_CLEANUP:%.*]], label [[SCALAR_PH]]
817; CHECK:       scalar.ph:
818; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
819; CHECK:       for.cond.cleanup:
820; CHECK-NEXT:    [[ADD_LCSSA:%.*]] = phi i32 [ poison, [[FOR_BODY]] ], [ [[TMP10]], [[MIDDLE_BLOCK]] ]
821; CHECK-NEXT:    [[ADD3_LCSSA:%.*]] = phi float [ poison, [[FOR_BODY]] ], [ [[TMP9]], [[MIDDLE_BLOCK]] ]
822; CHECK-NEXT:    store i32 [[ADD_LCSSA]], ptr @SA, align 4
823; CHECK-NEXT:    store float [[ADD3_LCSSA]], ptr @SB, align 4
824; CHECK-NEXT:    ret void
825; CHECK:       for.body:
826; CHECK-NEXT:    br i1 poison, label [[FOR_COND_CLEANUP]], label [[FOR_BODY]], !llvm.loop [[LOOP21:![0-9]+]]
827;
828entry:
829  br label %for.body
830
831for.cond.cleanup:                                 ; preds = %for.body
832  store i32 %add, ptr @SA, align 4
833  store float %add3, ptr @SB, align 4
834  ret void
835
836for.body:                                         ; preds = %for.body, %entry
837  %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
838  %SumB.014 = phi float [ undef, %entry ], [ %add3, %for.body ]
839  %SumA.013 = phi i32 [ undef, %entry ], [ %add, %for.body ]
840  %a = getelementptr inbounds %struct.IntFloat, ptr %p, i64 %indvars.iv, i32 0
841  %load1 = load i32, ptr %a, align 4
842  %add = add nsw i32 %load1, %SumA.013
843  %b = getelementptr inbounds %struct.IntFloat, ptr %p, i64 %indvars.iv, i32 1
844  %load2 = load float, ptr %b, align 4
845  %add3 = fadd fast float %SumB.014, %load2
846  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
847  %exitcond = icmp eq i64 %indvars.iv.next, 1024
848  br i1 %exitcond, label %for.cond.cleanup, label %for.body
849}
850
851; Check vectorization of interleaved access groups in the presence of
852; dependences (PR27626). The following tests check that we don't reorder
853; dependent loads and stores when generating code for interleaved access
854; groups. Stores should be scalarized because the required code motion would
855; break dependences, and the remaining interleaved load groups should have
856; gaps.
857
858; PR27626_0: Ensure a strided store is not moved after a dependent (zero
859;            distance) strided load.
860
861; void PR27626_0(struct pair *p, int z, int n) {
862;   for (int i = 0; i < n; i++) {
863;     p[i].x = z;
864;     p[i].y = p[i].x;
865;   }
866; }
867
868%pair.i32 = type { i32, i32 }
869define void @PR27626_0(ptr %p, i32 %z, i64 %n) #1 {
870; CHECK-LABEL: @PR27626_0(
871; CHECK-NEXT:  entry:
872; CHECK-NEXT:    [[SMAX:%.*]] = call i64 @llvm.smax.i64(i64 [[N:%.*]], i64 1)
873; CHECK-NEXT:    [[TMP0:%.*]] = call i64 @llvm.vscale.i64()
874; CHECK-NEXT:    [[TMP1:%.*]] = shl nuw nsw i64 [[TMP0]], 2
875; CHECK-NEXT:    [[MIN_ITERS_CHECK_NOT:%.*]] = icmp samesign ugt i64 [[SMAX]], [[TMP1]]
876; CHECK-NEXT:    br i1 [[MIN_ITERS_CHECK_NOT]], label [[VECTOR_PH:%.*]], label [[SCALAR_PH:%.*]]
877; CHECK:       vector.ph:
878; CHECK-NEXT:    [[TMP2:%.*]] = call i64 @llvm.vscale.i64()
879; CHECK-NEXT:    [[TMP3:%.*]] = shl nuw nsw i64 [[TMP2]], 2
880; CHECK-NEXT:    [[TMP4:%.*]] = add nsw i64 [[TMP3]], -1
881; CHECK-NEXT:    [[N_MOD_VF:%.*]] = and i64 [[SMAX]], [[TMP4]]
882; CHECK-NEXT:    [[TMP5:%.*]] = icmp eq i64 [[N_MOD_VF]], 0
883; CHECK-NEXT:    [[TMP6:%.*]] = select i1 [[TMP5]], i64 [[TMP3]], i64 [[N_MOD_VF]]
884; CHECK-NEXT:    [[N_VEC:%.*]] = sub nsw i64 [[SMAX]], [[TMP6]]
885; CHECK-NEXT:    [[TMP7:%.*]] = call i64 @llvm.vscale.i64()
886; CHECK-NEXT:    [[TMP8:%.*]] = shl nuw nsw i64 [[TMP7]], 2
887; CHECK-NEXT:    [[TMP9:%.*]] = call <vscale x 4 x i64> @llvm.stepvector.nxv4i64()
888; CHECK-NEXT:    [[DOTSPLATINSERT:%.*]] = insertelement <vscale x 4 x i64> poison, i64 [[TMP8]], i64 0
889; CHECK-NEXT:    [[DOTSPLAT:%.*]] = shufflevector <vscale x 4 x i64> [[DOTSPLATINSERT]], <vscale x 4 x i64> poison, <vscale x 4 x i32> zeroinitializer
890; CHECK-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <vscale x 4 x i32> poison, i32 [[Z:%.*]], i64 0
891; CHECK-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <vscale x 4 x i32> [[BROADCAST_SPLATINSERT]], <vscale x 4 x i32> poison, <vscale x 4 x i32> zeroinitializer
892; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
893; CHECK:       vector.body:
894; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
895; CHECK-NEXT:    [[VEC_IND:%.*]] = phi <vscale x 4 x i64> [ [[TMP9]], [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[VECTOR_BODY]] ]
896; CHECK-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[PAIR_I32:%.*]], ptr [[P:%.*]], <vscale x 4 x i64> [[VEC_IND]], i32 0
897; CHECK-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[PAIR_I32]], ptr [[P]], <vscale x 4 x i64> [[VEC_IND]], i32 1
898; CHECK-NEXT:    call void @llvm.masked.scatter.nxv4i32.nxv4p0(<vscale x 4 x i32> [[BROADCAST_SPLAT]], <vscale x 4 x ptr> [[TMP12]], i32 4, <vscale x 4 x i1> splat (i1 true))
899; CHECK-NEXT:    [[TMP14:%.*]] = extractelement <vscale x 4 x ptr> [[TMP12]], i64 0
900; CHECK-NEXT:    [[WIDE_VEC:%.*]] = load <vscale x 8 x i32>, ptr [[TMP14]], align 4
901; CHECK-NEXT:    [[STRIDED_VEC:%.*]] = call { <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.vector.deinterleave2.nxv8i32(<vscale x 8 x i32> [[WIDE_VEC]])
902; CHECK-NEXT:    [[TMP15:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC]], 0
903; CHECK-NEXT:    call void @llvm.masked.scatter.nxv4i32.nxv4p0(<vscale x 4 x i32> [[TMP15]], <vscale x 4 x ptr> [[TMP13]], i32 4, <vscale x 4 x i1> splat (i1 true))
904; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], [[TMP8]]
905; CHECK-NEXT:    [[VEC_IND_NEXT]] = add <vscale x 4 x i64> [[VEC_IND]], [[DOTSPLAT]]
906; CHECK-NEXT:    [[TMP16:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
907; CHECK-NEXT:    br i1 [[TMP16]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP22:![0-9]+]]
908; CHECK:       middle.block:
909; CHECK-NEXT:    br label [[SCALAR_PH]]
910; CHECK:       scalar.ph:
911; CHECK-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ]
912; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
913; CHECK:       for.body:
914; CHECK-NEXT:    [[I:%.*]] = phi i64 [ [[I_NEXT:%.*]], [[FOR_BODY]] ], [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ]
915; CHECK-NEXT:    [[P_I_X:%.*]] = getelementptr inbounds [[PAIR_I32]], ptr [[P]], i64 [[I]], i32 0
916; CHECK-NEXT:    [[P_I_Y:%.*]] = getelementptr inbounds [[PAIR_I32]], ptr [[P]], i64 [[I]], i32 1
917; CHECK-NEXT:    store i32 [[Z]], ptr [[P_I_X]], align 4
918; CHECK-NEXT:    store i32 [[Z]], ptr [[P_I_Y]], align 4
919; CHECK-NEXT:    [[I_NEXT]] = add nuw nsw i64 [[I]], 1
920; CHECK-NEXT:    [[COND:%.*]] = icmp slt i64 [[I_NEXT]], [[N]]
921; CHECK-NEXT:    br i1 [[COND]], label [[FOR_BODY]], label [[FOR_END:%.*]], !llvm.loop [[LOOP23:![0-9]+]]
922; CHECK:       for.end:
923; CHECK-NEXT:    ret void
924;
925entry:
926  br label %for.body
927
928for.body:
929  %i = phi i64 [ %i.next, %for.body ], [ 0, %entry ]
930  %p_i.x = getelementptr inbounds %pair.i32, ptr %p, i64 %i, i32 0
931  %p_i.y = getelementptr inbounds %pair.i32, ptr %p, i64 %i, i32 1
932  store i32 %z, ptr %p_i.x, align 4
933  %0 = load i32, ptr %p_i.x, align 4
934  store i32 %0, ptr %p_i.y, align 4
935  %i.next = add nuw nsw i64 %i, 1
936  %cond = icmp slt i64 %i.next, %n
937  br i1 %cond, label %for.body, label %for.end
938
939for.end:
940  ret void
941}
942
943; PR27626_1: Ensure a strided load is not moved before a dependent (zero
944;            distance) strided store.
945
946; void PR27626_1(struct pair *p, int n) {
947;   int s = 0;
948;   for (int i = 0; i < n; i++) {
949;     p[i].y = p[i].x;
950;     s += p[i].y
951;   }
952; }
953
954define i32 @PR27626_1(ptr %p, i64 %n) #1 {
955; CHECK-LABEL: @PR27626_1(
956; CHECK-NEXT:  entry:
957; CHECK-NEXT:    [[SMAX:%.*]] = call i64 @llvm.smax.i64(i64 [[N:%.*]], i64 1)
958; CHECK-NEXT:    [[TMP0:%.*]] = call i64 @llvm.vscale.i64()
959; CHECK-NEXT:    [[TMP1:%.*]] = shl nuw nsw i64 [[TMP0]], 2
960; CHECK-NEXT:    [[MIN_ITERS_CHECK_NOT:%.*]] = icmp samesign ugt i64 [[SMAX]], [[TMP1]]
961; CHECK-NEXT:    br i1 [[MIN_ITERS_CHECK_NOT]], label [[VECTOR_PH:%.*]], label [[SCALAR_PH:%.*]]
962; CHECK:       vector.ph:
963; CHECK-NEXT:    [[TMP2:%.*]] = call i64 @llvm.vscale.i64()
964; CHECK-NEXT:    [[TMP3:%.*]] = shl nuw nsw i64 [[TMP2]], 2
965; CHECK-NEXT:    [[TMP4:%.*]] = add nsw i64 [[TMP3]], -1
966; CHECK-NEXT:    [[N_MOD_VF:%.*]] = and i64 [[SMAX]], [[TMP4]]
967; CHECK-NEXT:    [[TMP5:%.*]] = icmp eq i64 [[N_MOD_VF]], 0
968; CHECK-NEXT:    [[TMP6:%.*]] = select i1 [[TMP5]], i64 [[TMP3]], i64 [[N_MOD_VF]]
969; CHECK-NEXT:    [[N_VEC:%.*]] = sub nsw i64 [[SMAX]], [[TMP6]]
970; CHECK-NEXT:    [[TMP7:%.*]] = call i64 @llvm.vscale.i64()
971; CHECK-NEXT:    [[TMP8:%.*]] = shl nuw nsw i64 [[TMP7]], 2
972; CHECK-NEXT:    [[TMP9:%.*]] = call <vscale x 4 x i64> @llvm.stepvector.nxv4i64()
973; CHECK-NEXT:    [[DOTSPLATINSERT:%.*]] = insertelement <vscale x 4 x i64> poison, i64 [[TMP8]], i64 0
974; CHECK-NEXT:    [[DOTSPLAT:%.*]] = shufflevector <vscale x 4 x i64> [[DOTSPLATINSERT]], <vscale x 4 x i64> poison, <vscale x 4 x i32> zeroinitializer
975; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
976; CHECK:       vector.body:
977; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
978; CHECK-NEXT:    [[VEC_IND:%.*]] = phi <vscale x 4 x i64> [ [[TMP9]], [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[VECTOR_BODY]] ]
979; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi <vscale x 4 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP17:%.*]], [[VECTOR_BODY]] ]
980; CHECK-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[PAIR_I32:%.*]], ptr [[P:%.*]], i64 [[INDEX]], i32 0
981; CHECK-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[PAIR_I32]], ptr [[P]], <vscale x 4 x i64> [[VEC_IND]], i32 1
982; CHECK-NEXT:    [[WIDE_VEC:%.*]] = load <vscale x 8 x i32>, ptr [[TMP12]], align 4
983; CHECK-NEXT:    [[STRIDED_VEC:%.*]] = call { <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.vector.deinterleave2.nxv8i32(<vscale x 8 x i32> [[WIDE_VEC]])
984; CHECK-NEXT:    [[TMP14:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC]], 0
985; CHECK-NEXT:    call void @llvm.masked.scatter.nxv4i32.nxv4p0(<vscale x 4 x i32> [[TMP14]], <vscale x 4 x ptr> [[TMP13]], i32 4, <vscale x 4 x i1> splat (i1 true))
986; CHECK-NEXT:    [[TMP15:%.*]] = extractelement <vscale x 4 x ptr> [[TMP13]], i64 0
987; CHECK-NEXT:    [[WIDE_VEC1:%.*]] = load <vscale x 8 x i32>, ptr [[TMP15]], align 4
988; CHECK-NEXT:    [[STRIDED_VEC2:%.*]] = call { <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.vector.deinterleave2.nxv8i32(<vscale x 8 x i32> [[WIDE_VEC1]])
989; CHECK-NEXT:    [[TMP16:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC2]], 0
990; CHECK-NEXT:    [[TMP17]] = add <vscale x 4 x i32> [[TMP16]], [[VEC_PHI]]
991; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], [[TMP8]]
992; CHECK-NEXT:    [[VEC_IND_NEXT]] = add <vscale x 4 x i64> [[VEC_IND]], [[DOTSPLAT]]
993; CHECK-NEXT:    [[TMP18:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
994; CHECK-NEXT:    br i1 [[TMP18]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP24:![0-9]+]]
995; CHECK:       middle.block:
996; CHECK-NEXT:    [[TMP19:%.*]] = call i32 @llvm.vector.reduce.add.nxv4i32(<vscale x 4 x i32> [[TMP17]])
997; CHECK-NEXT:    br label [[SCALAR_PH]]
998; CHECK:       scalar.ph:
999; CHECK-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ]
1000; CHECK-NEXT:    [[BC_MERGE_RDX:%.*]] = phi i32 [ [[TMP19]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY]] ]
1001; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
1002; CHECK:       for.body:
1003; CHECK-NEXT:    [[I:%.*]] = phi i64 [ [[I_NEXT:%.*]], [[FOR_BODY]] ], [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ]
1004; CHECK-NEXT:    [[S:%.*]] = phi i32 [ [[TMP21:%.*]], [[FOR_BODY]] ], [ [[BC_MERGE_RDX]], [[SCALAR_PH]] ]
1005; CHECK-NEXT:    [[P_I_X:%.*]] = getelementptr inbounds [[PAIR_I32]], ptr [[P]], i64 [[I]], i32 0
1006; CHECK-NEXT:    [[P_I_Y:%.*]] = getelementptr inbounds [[PAIR_I32]], ptr [[P]], i64 [[I]], i32 1
1007; CHECK-NEXT:    [[TMP20:%.*]] = load i32, ptr [[P_I_X]], align 4
1008; CHECK-NEXT:    store i32 [[TMP20]], ptr [[P_I_Y]], align 4
1009; CHECK-NEXT:    [[TMP21]] = add nsw i32 [[TMP20]], [[S]]
1010; CHECK-NEXT:    [[I_NEXT]] = add nuw nsw i64 [[I]], 1
1011; CHECK-NEXT:    [[COND:%.*]] = icmp slt i64 [[I_NEXT]], [[N]]
1012; CHECK-NEXT:    br i1 [[COND]], label [[FOR_BODY]], label [[FOR_END:%.*]], !llvm.loop [[LOOP25:![0-9]+]]
1013; CHECK:       for.end:
1014; CHECK-NEXT:    ret i32 [[TMP21]]
1015;
1016entry:
1017  br label %for.body
1018
1019for.body:
1020  %i = phi i64 [ %i.next, %for.body ], [ 0, %entry ]
1021  %s = phi i32 [ %2, %for.body ], [ 0, %entry ]
1022  %p_i.x = getelementptr inbounds %pair.i32, ptr %p, i64 %i, i32 0
1023  %p_i.y = getelementptr inbounds %pair.i32, ptr %p, i64 %i, i32 1
1024  %0 = load i32, ptr %p_i.x, align 4
1025  store i32 %0, ptr %p_i.y, align 4
1026  %1 = load i32, ptr %p_i.y, align 4
1027  %2 = add nsw i32 %1, %s
1028  %i.next = add nuw nsw i64 %i, 1
1029  %cond = icmp slt i64 %i.next, %n
1030  br i1 %cond, label %for.body, label %for.end
1031
1032for.end:
1033  %3 = phi i32 [ %2, %for.body ]
1034  ret i32 %3
1035}
1036
1037; PR27626_2: Ensure a strided store is not moved after a dependent (negative
1038;            distance) strided load.
1039
1040; void PR27626_2(struct pair *p, int z, int n) {
1041;   for (int i = 0; i < n; i++) {
1042;     p[i].x = z;
1043;     p[i].y = p[i - 1].x;
1044;   }
1045; }
1046
1047define void @PR27626_2(ptr %p, i64 %n, i32 %z) #1 {
1048; CHECK-LABEL: @PR27626_2(
1049; CHECK-NEXT:  entry:
1050; CHECK-NEXT:    [[SMAX:%.*]] = call i64 @llvm.smax.i64(i64 [[N:%.*]], i64 1)
1051; CHECK-NEXT:    [[TMP0:%.*]] = call i64 @llvm.vscale.i64()
1052; CHECK-NEXT:    [[TMP1:%.*]] = shl nuw nsw i64 [[TMP0]], 2
1053; CHECK-NEXT:    [[MIN_ITERS_CHECK_NOT:%.*]] = icmp samesign ugt i64 [[SMAX]], [[TMP1]]
1054; CHECK-NEXT:    br i1 [[MIN_ITERS_CHECK_NOT]], label [[VECTOR_PH:%.*]], label [[SCALAR_PH:%.*]]
1055; CHECK:       vector.ph:
1056; CHECK-NEXT:    [[TMP2:%.*]] = call i64 @llvm.vscale.i64()
1057; CHECK-NEXT:    [[TMP3:%.*]] = shl nuw nsw i64 [[TMP2]], 2
1058; CHECK-NEXT:    [[TMP4:%.*]] = add nsw i64 [[TMP3]], -1
1059; CHECK-NEXT:    [[N_MOD_VF:%.*]] = and i64 [[SMAX]], [[TMP4]]
1060; CHECK-NEXT:    [[TMP5:%.*]] = icmp eq i64 [[N_MOD_VF]], 0
1061; CHECK-NEXT:    [[TMP6:%.*]] = select i1 [[TMP5]], i64 [[TMP3]], i64 [[N_MOD_VF]]
1062; CHECK-NEXT:    [[N_VEC:%.*]] = sub nsw i64 [[SMAX]], [[TMP6]]
1063; CHECK-NEXT:    [[TMP7:%.*]] = call i64 @llvm.vscale.i64()
1064; CHECK-NEXT:    [[TMP8:%.*]] = shl nuw nsw i64 [[TMP7]], 2
1065; CHECK-NEXT:    [[TMP9:%.*]] = call <vscale x 4 x i64> @llvm.stepvector.nxv4i64()
1066; CHECK-NEXT:    [[DOTSPLATINSERT:%.*]] = insertelement <vscale x 4 x i64> poison, i64 [[TMP8]], i64 0
1067; CHECK-NEXT:    [[DOTSPLAT:%.*]] = shufflevector <vscale x 4 x i64> [[DOTSPLATINSERT]], <vscale x 4 x i64> poison, <vscale x 4 x i32> zeroinitializer
1068; CHECK-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <vscale x 4 x i32> poison, i32 [[Z:%.*]], i64 0
1069; CHECK-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <vscale x 4 x i32> [[BROADCAST_SPLATINSERT]], <vscale x 4 x i32> poison, <vscale x 4 x i32> zeroinitializer
1070; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
1071; CHECK:       vector.body:
1072; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
1073; CHECK-NEXT:    [[VEC_IND:%.*]] = phi <vscale x 4 x i64> [ [[TMP9]], [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[VECTOR_BODY]] ]
1074; CHECK-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[PAIR_I32:%.*]], ptr [[P:%.*]], <vscale x 4 x i64> [[VEC_IND]], i32 0
1075; CHECK-NEXT:    [[TMP13:%.*]] = getelementptr inbounds i8, ptr [[P]], i64 -8
1076; CHECK-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[PAIR_I32]], ptr [[P]], <vscale x 4 x i64> [[VEC_IND]], i32 1
1077; CHECK-NEXT:    call void @llvm.masked.scatter.nxv4i32.nxv4p0(<vscale x 4 x i32> [[BROADCAST_SPLAT]], <vscale x 4 x ptr> [[TMP12]], i32 4, <vscale x 4 x i1> splat (i1 true))
1078; CHECK-NEXT:    [[WIDE_VEC:%.*]] = load <vscale x 8 x i32>, ptr [[TMP13]], align 4
1079; CHECK-NEXT:    [[STRIDED_VEC:%.*]] = call { <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.vector.deinterleave2.nxv8i32(<vscale x 8 x i32> [[WIDE_VEC]])
1080; CHECK-NEXT:    [[TMP15:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC]], 0
1081; CHECK-NEXT:    call void @llvm.masked.scatter.nxv4i32.nxv4p0(<vscale x 4 x i32> [[TMP15]], <vscale x 4 x ptr> [[TMP14]], i32 4, <vscale x 4 x i1> splat (i1 true))
1082; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], [[TMP8]]
1083; CHECK-NEXT:    [[VEC_IND_NEXT]] = add <vscale x 4 x i64> [[VEC_IND]], [[DOTSPLAT]]
1084; CHECK-NEXT:    [[TMP16:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
1085; CHECK-NEXT:    br i1 [[TMP16]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP26:![0-9]+]]
1086; CHECK:       middle.block:
1087; CHECK-NEXT:    br label [[SCALAR_PH]]
1088; CHECK:       scalar.ph:
1089; CHECK-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ]
1090; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
1091; CHECK:       for.body:
1092; CHECK-NEXT:    [[I:%.*]] = phi i64 [ [[I_NEXT:%.*]], [[FOR_BODY]] ], [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ]
1093; CHECK-NEXT:    [[P_I_X:%.*]] = getelementptr inbounds [[PAIR_I32]], ptr [[P]], i64 [[I]], i32 0
1094; CHECK-NEXT:    [[P_I_MINUS_1_X:%.*]] = getelementptr inbounds i8, ptr [[P]], i64 -8
1095; CHECK-NEXT:    [[P_I_Y:%.*]] = getelementptr inbounds [[PAIR_I32]], ptr [[P]], i64 [[I]], i32 1
1096; CHECK-NEXT:    store i32 [[Z]], ptr [[P_I_X]], align 4
1097; CHECK-NEXT:    [[TMP17:%.*]] = load i32, ptr [[P_I_MINUS_1_X]], align 4
1098; CHECK-NEXT:    store i32 [[TMP17]], ptr [[P_I_Y]], align 4
1099; CHECK-NEXT:    [[I_NEXT]] = add nuw nsw i64 [[I]], 1
1100; CHECK-NEXT:    [[COND:%.*]] = icmp slt i64 [[I_NEXT]], [[N]]
1101; CHECK-NEXT:    br i1 [[COND]], label [[FOR_BODY]], label [[FOR_END:%.*]], !llvm.loop [[LOOP27:![0-9]+]]
1102; CHECK:       for.end:
1103; CHECK-NEXT:    ret void
1104;
1105entry:
1106  br label %for.body
1107
1108for.body:
1109  %i = phi i64 [ %i.next, %for.body ], [ 0, %entry ]
1110  %i_minus_1 = add nuw nsw i64 %i, -1
1111  %p_i.x = getelementptr inbounds %pair.i32, ptr %p, i64 %i, i32 0
1112  %p_i_minus_1.x = getelementptr inbounds %pair.i32, ptr %p, i64 %i_minus_1, i32 0
1113  %p_i.y = getelementptr inbounds %pair.i32, ptr %p, i64 %i, i32 1
1114  store i32 %z, ptr %p_i.x, align 4
1115  %0 = load i32, ptr %p_i_minus_1.x, align 4
1116  store i32 %0, ptr %p_i.y, align 4
1117  %i.next = add nuw nsw i64 %i, 1
1118  %cond = icmp slt i64 %i.next, %n
1119  br i1 %cond, label %for.body, label %for.end
1120
1121for.end:
1122  ret void
1123}
1124
1125; PR27626_3: Ensure a strided load is not moved before a dependent (negative
1126;            distance) strided store.
1127
1128; void PR27626_3(struct pair *p, int z, int n) {
1129;   for (int i = 0; i < n; i++) {
1130;     p[i + 1].y = p[i].x;
1131;     s += p[i].y;
1132;   }
1133; }
1134
1135define i32 @PR27626_3(ptr %p, i64 %n, i32 %z) #1 {
1136; CHECK-LABEL: @PR27626_3(
1137; CHECK-NEXT:  entry:
1138; CHECK-NEXT:    [[SMAX:%.*]] = call i64 @llvm.smax.i64(i64 [[N:%.*]], i64 1)
1139; CHECK-NEXT:    [[TMP0:%.*]] = call i64 @llvm.vscale.i64()
1140; CHECK-NEXT:    [[TMP1:%.*]] = shl nuw nsw i64 [[TMP0]], 2
1141; CHECK-NEXT:    [[MIN_ITERS_CHECK_NOT:%.*]] = icmp samesign ugt i64 [[SMAX]], [[TMP1]]
1142; CHECK-NEXT:    br i1 [[MIN_ITERS_CHECK_NOT]], label [[VECTOR_PH:%.*]], label [[SCALAR_PH:%.*]]
1143; CHECK:       vector.ph:
1144; CHECK-NEXT:    [[TMP2:%.*]] = call i64 @llvm.vscale.i64()
1145; CHECK-NEXT:    [[TMP3:%.*]] = shl nuw nsw i64 [[TMP2]], 2
1146; CHECK-NEXT:    [[TMP4:%.*]] = add nsw i64 [[TMP3]], -1
1147; CHECK-NEXT:    [[N_MOD_VF:%.*]] = and i64 [[SMAX]], [[TMP4]]
1148; CHECK-NEXT:    [[TMP5:%.*]] = icmp eq i64 [[N_MOD_VF]], 0
1149; CHECK-NEXT:    [[TMP6:%.*]] = select i1 [[TMP5]], i64 [[TMP3]], i64 [[N_MOD_VF]]
1150; CHECK-NEXT:    [[N_VEC:%.*]] = sub nsw i64 [[SMAX]], [[TMP6]]
1151; CHECK-NEXT:    [[TMP7:%.*]] = call i64 @llvm.vscale.i64()
1152; CHECK-NEXT:    [[TMP8:%.*]] = shl nuw nsw i64 [[TMP7]], 2
1153; CHECK-NEXT:    [[TMP9:%.*]] = call <vscale x 4 x i64> @llvm.stepvector.nxv4i64()
1154; CHECK-NEXT:    [[DOTSPLATINSERT:%.*]] = insertelement <vscale x 4 x i64> poison, i64 [[TMP8]], i64 0
1155; CHECK-NEXT:    [[DOTSPLAT:%.*]] = shufflevector <vscale x 4 x i64> [[DOTSPLATINSERT]], <vscale x 4 x i64> poison, <vscale x 4 x i32> zeroinitializer
1156; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
1157; CHECK:       vector.body:
1158; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
1159; CHECK-NEXT:    [[VEC_IND:%.*]] = phi <vscale x 4 x i64> [ [[TMP9]], [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[VECTOR_BODY]] ]
1160; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi <vscale x 4 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP18:%.*]], [[VECTOR_BODY]] ]
1161; CHECK-NEXT:    [[TMP12:%.*]] = add nuw nsw <vscale x 4 x i64> [[VEC_IND]], splat (i64 1)
1162; CHECK-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[PAIR_I32:%.*]], ptr [[P:%.*]], i64 [[INDEX]], i32 0
1163; CHECK-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[PAIR_I32]], ptr [[P]], i64 [[INDEX]], i32 1
1164; CHECK-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[PAIR_I32]], ptr [[P]], <vscale x 4 x i64> [[TMP12]], i32 1
1165; CHECK-NEXT:    [[WIDE_VEC:%.*]] = load <vscale x 8 x i32>, ptr [[TMP13]], align 4
1166; CHECK-NEXT:    [[STRIDED_VEC:%.*]] = call { <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.vector.deinterleave2.nxv8i32(<vscale x 8 x i32> [[WIDE_VEC]])
1167; CHECK-NEXT:    [[TMP16:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC]], 0
1168; CHECK-NEXT:    call void @llvm.masked.scatter.nxv4i32.nxv4p0(<vscale x 4 x i32> [[TMP16]], <vscale x 4 x ptr> [[TMP15]], i32 4, <vscale x 4 x i1> splat (i1 true))
1169; CHECK-NEXT:    [[WIDE_VEC1:%.*]] = load <vscale x 8 x i32>, ptr [[TMP14]], align 4
1170; CHECK-NEXT:    [[STRIDED_VEC2:%.*]] = call { <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.vector.deinterleave2.nxv8i32(<vscale x 8 x i32> [[WIDE_VEC1]])
1171; CHECK-NEXT:    [[TMP17:%.*]] = extractvalue { <vscale x 4 x i32>, <vscale x 4 x i32> } [[STRIDED_VEC2]], 0
1172; CHECK-NEXT:    [[TMP18]] = add <vscale x 4 x i32> [[TMP17]], [[VEC_PHI]]
1173; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], [[TMP8]]
1174; CHECK-NEXT:    [[VEC_IND_NEXT]] = add <vscale x 4 x i64> [[VEC_IND]], [[DOTSPLAT]]
1175; CHECK-NEXT:    [[TMP19:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
1176; CHECK-NEXT:    br i1 [[TMP19]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP28:![0-9]+]]
1177; CHECK:       middle.block:
1178; CHECK-NEXT:    [[TMP20:%.*]] = call i32 @llvm.vector.reduce.add.nxv4i32(<vscale x 4 x i32> [[TMP18]])
1179; CHECK-NEXT:    br label [[SCALAR_PH]]
1180; CHECK:       scalar.ph:
1181; CHECK-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ]
1182; CHECK-NEXT:    [[BC_MERGE_RDX:%.*]] = phi i32 [ [[TMP20]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY]] ]
1183; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
1184; CHECK:       for.body:
1185; CHECK-NEXT:    [[I:%.*]] = phi i64 [ [[I_NEXT:%.*]], [[FOR_BODY]] ], [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ]
1186; CHECK-NEXT:    [[S:%.*]] = phi i32 [ [[TMP23:%.*]], [[FOR_BODY]] ], [ [[BC_MERGE_RDX]], [[SCALAR_PH]] ]
1187; CHECK-NEXT:    [[I_PLUS_1:%.*]] = add nuw nsw i64 [[I]], 1
1188; CHECK-NEXT:    [[P_I_X:%.*]] = getelementptr inbounds [[PAIR_I32]], ptr [[P]], i64 [[I]], i32 0
1189; CHECK-NEXT:    [[P_I_Y:%.*]] = getelementptr inbounds [[PAIR_I32]], ptr [[P]], i64 [[I]], i32 1
1190; CHECK-NEXT:    [[P_I_PLUS_1_Y:%.*]] = getelementptr inbounds [[PAIR_I32]], ptr [[P]], i64 [[I_PLUS_1]], i32 1
1191; CHECK-NEXT:    [[TMP21:%.*]] = load i32, ptr [[P_I_X]], align 4
1192; CHECK-NEXT:    store i32 [[TMP21]], ptr [[P_I_PLUS_1_Y]], align 4
1193; CHECK-NEXT:    [[TMP22:%.*]] = load i32, ptr [[P_I_Y]], align 4
1194; CHECK-NEXT:    [[TMP23]] = add nsw i32 [[TMP22]], [[S]]
1195; CHECK-NEXT:    [[I_NEXT]] = add nuw nsw i64 [[I]], 1
1196; CHECK-NEXT:    [[COND:%.*]] = icmp slt i64 [[I_NEXT]], [[N]]
1197; CHECK-NEXT:    br i1 [[COND]], label [[FOR_BODY]], label [[FOR_END:%.*]], !llvm.loop [[LOOP29:![0-9]+]]
1198; CHECK:       for.end:
1199; CHECK-NEXT:    ret i32 [[TMP23]]
1200;
1201entry:
1202  br label %for.body
1203
1204for.body:
1205  %i = phi i64 [ %i.next, %for.body ], [ 0, %entry ]
1206  %s = phi i32 [ %2, %for.body ], [ 0, %entry ]
1207  %i_plus_1 = add nuw nsw i64 %i, 1
1208  %p_i.x = getelementptr inbounds %pair.i32, ptr %p, i64 %i, i32 0
1209  %p_i.y = getelementptr inbounds %pair.i32, ptr %p, i64 %i, i32 1
1210  %p_i_plus_1.y = getelementptr inbounds %pair.i32, ptr %p, i64 %i_plus_1, i32 1
1211  %0 = load i32, ptr %p_i.x, align 4
1212  store i32 %0, ptr %p_i_plus_1.y, align 4
1213  %1 = load i32, ptr %p_i.y, align 4
1214  %2 = add nsw i32 %1, %s
1215  %i.next = add nuw nsw i64 %i, 1
1216  %cond = icmp slt i64 %i.next, %n
1217  br i1 %cond, label %for.body, label %for.end
1218
1219for.end:
1220  %3 = phi i32 [ %2, %for.body ]
1221  ret i32 %3
1222}
1223
1224; PR27626_4: Ensure we form an interleaved group for strided stores in the
1225;            presence of a write-after-write dependence. We create a group for
1226;            (2) and (3) while excluding (1).
1227
1228; void PR27626_4(int *a, int x, int y, int z, int n) {
1229;   for (int i = 0; i < n; i += 2) {
1230;     a[i] = x;      // (1)
1231;     a[i] = y;      // (2)
1232;     a[i + 1] = z;  // (3)
1233;   }
1234; }
1235
1236define void @PR27626_4(ptr %a, i32 %x, i32 %y, i32 %z, i64 %n) #1 {
1237; CHECK-LABEL: @PR27626_4(
1238; CHECK-NEXT:  entry:
1239; CHECK-NEXT:    [[SMAX:%.*]] = call i64 @llvm.smax.i64(i64 [[N:%.*]], i64 2)
1240; CHECK-NEXT:    [[TMP0:%.*]] = add nsw i64 [[SMAX]], -1
1241; CHECK-NEXT:    [[TMP1:%.*]] = lshr i64 [[TMP0]], 1
1242; CHECK-NEXT:    [[TMP2:%.*]] = add nuw nsw i64 [[TMP1]], 1
1243; CHECK-NEXT:    [[TMP3:%.*]] = call i64 @llvm.vscale.i64()
1244; CHECK-NEXT:    [[TMP4:%.*]] = shl nuw nsw i64 [[TMP3]], 2
1245; CHECK-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp samesign ult i64 [[TMP2]], [[TMP4]]
1246; CHECK-NEXT:    br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
1247; CHECK:       vector.ph:
1248; CHECK-NEXT:    [[TMP5:%.*]] = call i64 @llvm.vscale.i64()
1249; CHECK-NEXT:    [[DOTNEG:%.*]] = mul nsw i64 [[TMP5]], -4
1250; CHECK-NEXT:    [[N_VEC:%.*]] = and i64 [[TMP2]], [[DOTNEG]]
1251; CHECK-NEXT:    [[TMP6:%.*]] = call i64 @llvm.vscale.i64()
1252; CHECK-NEXT:    [[TMP7:%.*]] = shl nuw nsw i64 [[TMP6]], 2
1253; CHECK-NEXT:    [[IND_END:%.*]] = shl nuw i64 [[N_VEC]], 1
1254; CHECK-NEXT:    [[TMP8:%.*]] = call <vscale x 4 x i64> @llvm.stepvector.nxv4i64()
1255; CHECK-NEXT:    [[TMP9:%.*]] = shl <vscale x 4 x i64> [[TMP8]], splat (i64 1)
1256; CHECK-NEXT:    [[TMP11:%.*]] = shl nuw nsw i64 [[TMP6]], 3
1257; CHECK-NEXT:    [[DOTSPLATINSERT:%.*]] = insertelement <vscale x 4 x i64> poison, i64 [[TMP11]], i64 0
1258; CHECK-NEXT:    [[DOTSPLAT:%.*]] = shufflevector <vscale x 4 x i64> [[DOTSPLATINSERT]], <vscale x 4 x i64> poison, <vscale x 4 x i32> zeroinitializer
1259; CHECK-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <vscale x 4 x i32> poison, i32 [[X:%.*]], i64 0
1260; CHECK-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <vscale x 4 x i32> [[BROADCAST_SPLATINSERT]], <vscale x 4 x i32> poison, <vscale x 4 x i32> zeroinitializer
1261; CHECK-NEXT:    [[BROADCAST_SPLATINSERT1:%.*]] = insertelement <vscale x 4 x i32> poison, i32 [[Y:%.*]], i64 0
1262; CHECK-NEXT:    [[BROADCAST_SPLAT2:%.*]] = shufflevector <vscale x 4 x i32> [[BROADCAST_SPLATINSERT1]], <vscale x 4 x i32> poison, <vscale x 4 x i32> zeroinitializer
1263; CHECK-NEXT:    [[BROADCAST_SPLATINSERT3:%.*]] = insertelement <vscale x 4 x i32> poison, i32 [[Z:%.*]], i64 0
1264; CHECK-NEXT:    [[BROADCAST_SPLAT4:%.*]] = shufflevector <vscale x 4 x i32> [[BROADCAST_SPLATINSERT3]], <vscale x 4 x i32> poison, <vscale x 4 x i32> zeroinitializer
1265; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
1266; CHECK:       vector.body:
1267; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
1268; CHECK-NEXT:    [[VEC_IND:%.*]] = phi <vscale x 4 x i64> [ [[TMP9]], [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[VECTOR_BODY]] ]
1269; CHECK-NEXT:    [[TMP13:%.*]] = getelementptr inbounds i32, ptr [[A:%.*]], <vscale x 4 x i64> [[VEC_IND]]
1270; CHECK-NEXT:    call void @llvm.masked.scatter.nxv4i32.nxv4p0(<vscale x 4 x i32> [[BROADCAST_SPLAT]], <vscale x 4 x ptr> [[TMP13]], i32 4, <vscale x 4 x i1> splat (i1 true))
1271; CHECK-NEXT:    [[P:%.*]] = extractelement <vscale x 4 x ptr> [[TMP13]], i64 0
1272; CHECK-NEXT:    [[INTERLEAVED_VEC:%.*]] = call <vscale x 8 x i32> @llvm.vector.interleave2.nxv8i32(<vscale x 4 x i32> [[BROADCAST_SPLAT2]], <vscale x 4 x i32> [[BROADCAST_SPLAT4]])
1273; CHECK-NEXT:    store <vscale x 8 x i32> [[INTERLEAVED_VEC]], ptr [[P]], align 4
1274; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], [[TMP7]]
1275; CHECK-NEXT:    [[VEC_IND_NEXT]] = add <vscale x 4 x i64> [[VEC_IND]], [[DOTSPLAT]]
1276; CHECK-NEXT:    [[TMP16:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
1277; CHECK-NEXT:    br i1 [[TMP16]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP30:![0-9]+]]
1278; CHECK:       middle.block:
1279; CHECK-NEXT:    [[CMP_N:%.*]] = icmp eq i64 [[TMP2]], [[N_VEC]]
1280; CHECK-NEXT:    br i1 [[CMP_N]], label [[FOR_END:%.*]], label [[SCALAR_PH]]
1281; CHECK:       scalar.ph:
1282; CHECK-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ [[IND_END]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ]
1283; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
1284; CHECK:       for.body:
1285; CHECK-NEXT:    [[I:%.*]] = phi i64 [ [[I_NEXT:%.*]], [[FOR_BODY]] ], [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ]
1286; CHECK-NEXT:    [[I_PLUS_1:%.*]] = or disjoint i64 [[I]], 1
1287; CHECK-NEXT:    [[A_I:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[I]]
1288; CHECK-NEXT:    [[A_I_PLUS_1:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[I_PLUS_1]]
1289; CHECK-NEXT:    store i32 [[Y]], ptr [[A_I]], align 4
1290; CHECK-NEXT:    store i32 [[Z]], ptr [[A_I_PLUS_1]], align 4
1291; CHECK-NEXT:    [[I_NEXT]] = add nuw nsw i64 [[I]], 2
1292; CHECK-NEXT:    [[COND:%.*]] = icmp slt i64 [[I_NEXT]], [[N]]
1293; CHECK-NEXT:    br i1 [[COND]], label [[FOR_BODY]], label [[FOR_END]], !llvm.loop [[LOOP31:![0-9]+]]
1294; CHECK:       for.end:
1295; CHECK-NEXT:    ret void
1296;
1297entry:
1298  br label %for.body
1299
1300for.body:
1301  %i = phi i64 [ %i.next, %for.body ], [ 0, %entry ]
1302  %i_plus_1 = add i64 %i, 1
1303  %a_i = getelementptr inbounds i32, ptr %a, i64 %i
1304  %a_i_plus_1 = getelementptr inbounds i32, ptr %a, i64 %i_plus_1
1305  store i32 %x, ptr %a_i, align 4
1306  store i32 %y, ptr %a_i, align 4
1307  store i32 %z, ptr %a_i_plus_1, align 4
1308  %i.next = add nuw nsw i64 %i, 2
1309  %cond = icmp slt i64 %i.next, %n
1310  br i1 %cond, label %for.body, label %for.end
1311
1312for.end:
1313  ret void
1314}
1315
1316; PR27626_5: Ensure we do not form an interleaved group for strided stores in
1317;            the presence of a write-after-write dependence.
1318
1319; void PR27626_5(int *a, int x, int y, int z, int n) {
1320;   for (int i = 3; i < n; i += 2) {
1321;     a[i - 1] = x;
1322;     a[i - 3] = y;
1323;     a[i] = z;
1324;   }
1325; }
1326
1327define void @PR27626_5(ptr %a, i32 %x, i32 %y, i32 %z, i64 %n) #1 {
1328; CHECK-LABEL: @PR27626_5(
1329; CHECK-NEXT:  entry:
1330; CHECK-NEXT:    [[SMAX:%.*]] = call i64 @llvm.smax.i64(i64 [[N:%.*]], i64 5)
1331; CHECK-NEXT:    [[TMP0:%.*]] = add nsw i64 [[SMAX]], -4
1332; CHECK-NEXT:    [[TMP1:%.*]] = lshr i64 [[TMP0]], 1
1333; CHECK-NEXT:    [[TMP2:%.*]] = add nuw nsw i64 [[TMP1]], 1
1334; CHECK-NEXT:    [[TMP3:%.*]] = call i64 @llvm.vscale.i64()
1335; CHECK-NEXT:    [[TMP4:%.*]] = shl nuw nsw i64 [[TMP3]], 2
1336; CHECK-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp samesign ult i64 [[TMP2]], [[TMP4]]
1337; CHECK-NEXT:    br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
1338; CHECK:       vector.ph:
1339; CHECK-NEXT:    [[TMP5:%.*]] = call i64 @llvm.vscale.i64()
1340; CHECK-NEXT:    [[DOTNEG:%.*]] = mul nsw i64 [[TMP5]], -4
1341; CHECK-NEXT:    [[N_VEC:%.*]] = and i64 [[TMP2]], [[DOTNEG]]
1342; CHECK-NEXT:    [[TMP7:%.*]] = call i64 @llvm.vscale.i64()
1343; CHECK-NEXT:    [[TMP8:%.*]] = shl nuw nsw i64 [[TMP7]], 2
1344; CHECK-NEXT:    [[TMP11:%.*]] = shl nuw i64 [[N_VEC]], 1
1345; CHECK-NEXT:    [[IND_END:%.*]] = or disjoint i64 [[TMP11]], 3
1346; CHECK-NEXT:    [[TMP9:%.*]] = call <vscale x 4 x i64> @llvm.stepvector.nxv4i64()
1347; CHECK-NEXT:    [[TMP10:%.*]] = shl <vscale x 4 x i64> [[TMP9]], splat (i64 1)
1348; CHECK-NEXT:    [[INDUCTION:%.*]] = add <vscale x 4 x i64> [[TMP10]], splat (i64 3)
1349; CHECK-NEXT:    [[TMP12:%.*]] = shl nuw nsw i64 [[TMP7]], 3
1350; CHECK-NEXT:    [[DOTSPLATINSERT:%.*]] = insertelement <vscale x 4 x i64> poison, i64 [[TMP12]], i64 0
1351; CHECK-NEXT:    [[DOTSPLAT:%.*]] = shufflevector <vscale x 4 x i64> [[DOTSPLATINSERT]], <vscale x 4 x i64> poison, <vscale x 4 x i32> zeroinitializer
1352; CHECK-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <vscale x 4 x i32> poison, i32 [[X:%.*]], i64 0
1353; CHECK-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <vscale x 4 x i32> [[BROADCAST_SPLATINSERT]], <vscale x 4 x i32> poison, <vscale x 4 x i32> zeroinitializer
1354; CHECK-NEXT:    [[BROADCAST_SPLATINSERT1:%.*]] = insertelement <vscale x 4 x i32> poison, i32 [[Y:%.*]], i64 0
1355; CHECK-NEXT:    [[BROADCAST_SPLAT2:%.*]] = shufflevector <vscale x 4 x i32> [[BROADCAST_SPLATINSERT1]], <vscale x 4 x i32> poison, <vscale x 4 x i32> zeroinitializer
1356; CHECK-NEXT:    [[BROADCAST_SPLATINSERT3:%.*]] = insertelement <vscale x 4 x i32> poison, i32 [[Z:%.*]], i64 0
1357; CHECK-NEXT:    [[BROADCAST_SPLAT4:%.*]] = shufflevector <vscale x 4 x i32> [[BROADCAST_SPLATINSERT3]], <vscale x 4 x i32> poison, <vscale x 4 x i32> zeroinitializer
1358; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
1359; CHECK:       vector.body:
1360; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
1361; CHECK-NEXT:    [[VEC_IND:%.*]] = phi <vscale x 4 x i64> [ [[INDUCTION]], [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[VECTOR_BODY]] ]
1362; CHECK-NEXT:    [[TMP13:%.*]] = add <vscale x 4 x i64> [[VEC_IND]], splat (i64 -1)
1363; CHECK-NEXT:    [[TMP14:%.*]] = add <vscale x 4 x i64> [[VEC_IND]], splat (i64 -3)
1364; CHECK-NEXT:    [[TMP15:%.*]] = getelementptr inbounds i32, ptr [[A:%.*]], <vscale x 4 x i64> [[VEC_IND]]
1365; CHECK-NEXT:    [[TMP16:%.*]] = getelementptr inbounds i32, ptr [[A]], <vscale x 4 x i64> [[TMP13]]
1366; CHECK-NEXT:    [[TMP17:%.*]] = getelementptr inbounds i32, ptr [[A]], <vscale x 4 x i64> [[TMP14]]
1367; CHECK-NEXT:    call void @llvm.masked.scatter.nxv4i32.nxv4p0(<vscale x 4 x i32> [[BROADCAST_SPLAT]], <vscale x 4 x ptr> [[TMP16]], i32 4, <vscale x 4 x i1> splat (i1 true))
1368; CHECK-NEXT:    call void @llvm.masked.scatter.nxv4i32.nxv4p0(<vscale x 4 x i32> [[BROADCAST_SPLAT2]], <vscale x 4 x ptr> [[TMP17]], i32 4, <vscale x 4 x i1> splat (i1 true))
1369; CHECK-NEXT:    call void @llvm.masked.scatter.nxv4i32.nxv4p0(<vscale x 4 x i32> [[BROADCAST_SPLAT4]], <vscale x 4 x ptr> [[TMP15]], i32 4, <vscale x 4 x i1> splat (i1 true))
1370; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], [[TMP8]]
1371; CHECK-NEXT:    [[VEC_IND_NEXT]] = add <vscale x 4 x i64> [[VEC_IND]], [[DOTSPLAT]]
1372; CHECK-NEXT:    [[TMP18:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
1373; CHECK-NEXT:    br i1 [[TMP18]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP32:![0-9]+]]
1374; CHECK:       middle.block:
1375; CHECK-NEXT:    [[CMP_N:%.*]] = icmp eq i64 [[TMP2]], [[N_VEC]]
1376; CHECK-NEXT:    br i1 [[CMP_N]], label [[FOR_END:%.*]], label [[SCALAR_PH]]
1377; CHECK:       scalar.ph:
1378; CHECK-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ [[IND_END]], [[MIDDLE_BLOCK]] ], [ 3, [[ENTRY:%.*]] ]
1379; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
1380; CHECK:       for.body:
1381; CHECK-NEXT:    [[I:%.*]] = phi i64 [ [[I_NEXT:%.*]], [[FOR_BODY]] ], [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ]
1382; CHECK-NEXT:    [[A_I:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[I]]
1383; CHECK-NEXT:    [[TMP19:%.*]] = getelementptr i32, ptr [[A]], i64 [[I]]
1384; CHECK-NEXT:    [[A_I_MINUS_1:%.*]] = getelementptr i8, ptr [[TMP19]], i64 -4
1385; CHECK-NEXT:    [[TMP20:%.*]] = getelementptr i32, ptr [[A]], i64 [[I]]
1386; CHECK-NEXT:    [[A_I_MINUS_3:%.*]] = getelementptr i8, ptr [[TMP20]], i64 -12
1387; CHECK-NEXT:    store i32 [[X]], ptr [[A_I_MINUS_1]], align 4
1388; CHECK-NEXT:    store i32 [[Y]], ptr [[A_I_MINUS_3]], align 4
1389; CHECK-NEXT:    store i32 [[Z]], ptr [[A_I]], align 4
1390; CHECK-NEXT:    [[I_NEXT]] = add nuw nsw i64 [[I]], 2
1391; CHECK-NEXT:    [[COND:%.*]] = icmp slt i64 [[I_NEXT]], [[N]]
1392; CHECK-NEXT:    br i1 [[COND]], label [[FOR_BODY]], label [[FOR_END]], !llvm.loop [[LOOP33:![0-9]+]]
1393; CHECK:       for.end:
1394; CHECK-NEXT:    ret void
1395;
1396entry:
1397  br label %for.body
1398
1399for.body:
1400  %i = phi i64 [ %i.next, %for.body ], [ 3, %entry ]
1401  %i_minus_1 = sub i64 %i, 1
1402  %i_minus_3 = sub i64 %i_minus_1, 2
1403  %a_i = getelementptr inbounds i32, ptr %a, i64 %i
1404  %a_i_minus_1 = getelementptr inbounds i32, ptr %a, i64 %i_minus_1
1405  %a_i_minus_3 = getelementptr inbounds i32, ptr %a, i64 %i_minus_3
1406  store i32 %x, ptr %a_i_minus_1, align 4
1407  store i32 %y, ptr %a_i_minus_3, align 4
1408  store i32 %z, ptr %a_i, align 4
1409  %i.next = add nuw nsw i64 %i, 2
1410  %cond = icmp slt i64 %i.next, %n
1411  br i1 %cond, label %for.body, label %for.end
1412
1413for.end:
1414  ret void
1415}
1416
1417; PR34743: Ensure that a cast which needs to sink after a load that belongs to
1418; an interleaved group, indeeded gets sunk.
1419
1420; void PR34743(short *a, int *b, int n) {
1421;   for (int i = 0, iv = 0; iv < n; i++, iv += 2) {
1422;     b[i] = a[iv] * a[iv+1] * a[iv+2];
1423;   }
1424; }
1425
1426define void @PR34743(ptr %a, ptr %b, i64 %n) #1 {
1427; CHECK-LABEL: @PR34743(
1428; CHECK-NEXT:  entry:
1429; CHECK-NEXT:    [[DOTPRE:%.*]] = load i16, ptr [[A:%.*]], align 2
1430; CHECK-NEXT:    [[TMP0:%.*]] = lshr i64 [[N:%.*]], 1
1431; CHECK-NEXT:    [[TMP1:%.*]] = add nuw i64 [[TMP0]], 1
1432; CHECK-NEXT:    [[TMP2:%.*]] = call i64 @llvm.vscale.i64()
1433; CHECK-NEXT:    [[TMP3:%.*]] = shl nuw nsw i64 [[TMP2]], 2
1434; CHECK-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[TMP1]], [[TMP3]]
1435; CHECK-NEXT:    br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_MEMCHECK:%.*]]
1436; CHECK:       vector.memcheck:
1437; CHECK-NEXT:    [[TMP4:%.*]] = shl i64 [[N]], 1
1438; CHECK-NEXT:    [[TMP5:%.*]] = and i64 [[TMP4]], -4
1439; CHECK-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[B:%.*]], i64 [[TMP5]]
1440; CHECK-NEXT:    [[SCEVGEP:%.*]] = getelementptr i8, ptr [[TMP6]], i64 4
1441; CHECK-NEXT:    [[SCEVGEP1:%.*]] = getelementptr nuw i8, ptr [[A]], i64 2
1442; CHECK-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[A]], i64 [[TMP5]]
1443; CHECK-NEXT:    [[SCEVGEP2:%.*]] = getelementptr i8, ptr [[TMP7]], i64 6
1444; CHECK-NEXT:    [[BOUND0:%.*]] = icmp ult ptr [[B]], [[SCEVGEP2]]
1445; CHECK-NEXT:    [[BOUND1:%.*]] = icmp ult ptr [[SCEVGEP1]], [[SCEVGEP]]
1446; CHECK-NEXT:    [[FOUND_CONFLICT:%.*]] = and i1 [[BOUND0]], [[BOUND1]]
1447; CHECK-NEXT:    br i1 [[FOUND_CONFLICT]], label [[SCALAR_PH]], label [[VECTOR_PH:%.*]]
1448; CHECK:       vector.ph:
1449; CHECK-NEXT:    [[TMP8:%.*]] = call i64 @llvm.vscale.i64()
1450; CHECK-NEXT:    [[DOTNEG:%.*]] = mul nsw i64 [[TMP8]], -4
1451; CHECK-NEXT:    [[N_VEC:%.*]] = and i64 [[TMP1]], [[DOTNEG]]
1452; CHECK-NEXT:    [[TMP9:%.*]] = call i64 @llvm.vscale.i64()
1453; CHECK-NEXT:    [[TMP10:%.*]] = shl nuw nsw i64 [[TMP9]], 2
1454; CHECK-NEXT:    [[IND_END:%.*]] = shl i64 [[N_VEC]], 1
1455; CHECK-NEXT:    [[TMP11:%.*]] = call i32 @llvm.vscale.i32()
1456; CHECK-NEXT:    [[TMP12:%.*]] = shl nuw nsw i32 [[TMP11]], 2
1457; CHECK-NEXT:    [[TMP13:%.*]] = add nsw i32 [[TMP12]], -1
1458; CHECK-NEXT:    [[VECTOR_RECUR_INIT:%.*]] = insertelement <vscale x 4 x i16> poison, i16 [[DOTPRE]], i32 [[TMP13]]
1459; CHECK-NEXT:    [[TMP14:%.*]] = call <vscale x 4 x i64> @llvm.stepvector.nxv4i64()
1460; CHECK-NEXT:    [[TMP15:%.*]] = shl <vscale x 4 x i64> [[TMP14]], splat (i64 1)
1461; CHECK-NEXT:    [[TMP17:%.*]] = shl nuw nsw i64 [[TMP9]], 3
1462; CHECK-NEXT:    [[DOTSPLATINSERT:%.*]] = insertelement <vscale x 4 x i64> poison, i64 [[TMP17]], i64 0
1463; CHECK-NEXT:    [[DOTSPLAT:%.*]] = shufflevector <vscale x 4 x i64> [[DOTSPLATINSERT]], <vscale x 4 x i64> poison, <vscale x 4 x i32> zeroinitializer
1464; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
1465; CHECK:       vector.body:
1466; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
1467; CHECK-NEXT:    [[VECTOR_RECUR:%.*]] = phi <vscale x 4 x i16> [ [[VECTOR_RECUR_INIT]], [[VECTOR_PH]] ], [ [[WIDE_MASKED_GATHER4:%.*]], [[VECTOR_BODY]] ]
1468; CHECK-NEXT:    [[VEC_IND:%.*]] = phi <vscale x 4 x i64> [ [[TMP15]], [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[VECTOR_BODY]] ]
1469; CHECK-NEXT:    [[TMP18:%.*]] = add nuw nsw <vscale x 4 x i64> [[VEC_IND]], splat (i64 1)
1470; CHECK-NEXT:    [[TMP19:%.*]] = add nuw nsw <vscale x 4 x i64> [[VEC_IND]], splat (i64 2)
1471; CHECK-NEXT:    [[TMP20:%.*]] = getelementptr inbounds i16, ptr [[A]], <vscale x 4 x i64> [[TMP18]]
1472; CHECK-NEXT:    [[WIDE_MASKED_GATHER:%.*]] = call <vscale x 4 x i16> @llvm.masked.gather.nxv4i16.nxv4p0(<vscale x 4 x ptr> [[TMP20]], i32 4, <vscale x 4 x i1> splat (i1 true), <vscale x 4 x i16> poison), !alias.scope [[META34:![0-9]+]]
1473; CHECK-NEXT:    [[TMP21:%.*]] = sext <vscale x 4 x i16> [[WIDE_MASKED_GATHER]] to <vscale x 4 x i32>
1474; CHECK-NEXT:    [[TMP22:%.*]] = getelementptr inbounds i16, ptr [[A]], <vscale x 4 x i64> [[TMP19]]
1475; CHECK-NEXT:    [[WIDE_MASKED_GATHER4]] = call <vscale x 4 x i16> @llvm.masked.gather.nxv4i16.nxv4p0(<vscale x 4 x ptr> [[TMP22]], i32 4, <vscale x 4 x i1> splat (i1 true), <vscale x 4 x i16> poison), !alias.scope [[META34]]
1476; CHECK-NEXT:    [[TMP23:%.*]] = call <vscale x 4 x i16> @llvm.vector.splice.nxv4i16(<vscale x 4 x i16> [[VECTOR_RECUR]], <vscale x 4 x i16> [[WIDE_MASKED_GATHER4]], i32 -1)
1477; CHECK-NEXT:    [[TMP24:%.*]] = sext <vscale x 4 x i16> [[TMP23]] to <vscale x 4 x i32>
1478; CHECK-NEXT:    [[TMP25:%.*]] = sext <vscale x 4 x i16> [[WIDE_MASKED_GATHER4]] to <vscale x 4 x i32>
1479; CHECK-NEXT:    [[TMP26:%.*]] = mul nsw <vscale x 4 x i32> [[TMP24]], [[TMP21]]
1480; CHECK-NEXT:    [[TMP27:%.*]] = mul nsw <vscale x 4 x i32> [[TMP26]], [[TMP25]]
1481; CHECK-NEXT:    [[TMP28:%.*]] = getelementptr inbounds i32, ptr [[B]], i64 [[INDEX]]
1482; CHECK-NEXT:    store <vscale x 4 x i32> [[TMP27]], ptr [[TMP28]], align 4, !alias.scope [[META37:![0-9]+]], !noalias [[META34]]
1483; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], [[TMP10]]
1484; CHECK-NEXT:    [[VEC_IND_NEXT]] = add <vscale x 4 x i64> [[VEC_IND]], [[DOTSPLAT]]
1485; CHECK-NEXT:    [[TMP29:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
1486; CHECK-NEXT:    br i1 [[TMP29]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP39:![0-9]+]]
1487; CHECK:       middle.block:
1488; CHECK-NEXT:    [[TMP30:%.*]] = call i32 @llvm.vscale.i32()
1489; CHECK-NEXT:    [[TMP31:%.*]] = shl nuw nsw i32 [[TMP30]], 2
1490; CHECK-NEXT:    [[TMP32:%.*]] = add nsw i32 [[TMP31]], -1
1491; CHECK-NEXT:    [[VECTOR_RECUR_EXTRACT:%.*]] = extractelement <vscale x 4 x i16> [[WIDE_MASKED_GATHER4]], i32 [[TMP32]]
1492; CHECK-NEXT:    [[CMP_N:%.*]] = icmp eq i64 [[TMP1]], [[N_VEC]]
1493; CHECK-NEXT:    br i1 [[CMP_N]], label [[END:%.*]], label [[SCALAR_PH]]
1494; CHECK:       scalar.ph:
1495; CHECK-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i16 [ [[VECTOR_RECUR_EXTRACT]], [[MIDDLE_BLOCK]] ], [ [[DOTPRE]], [[VECTOR_MEMCHECK]] ], [ [[DOTPRE]], [[ENTRY:%.*]] ]
1496; CHECK-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ [[IND_END]], [[MIDDLE_BLOCK]] ], [ 0, [[VECTOR_MEMCHECK]] ], [ 0, [[ENTRY]] ]
1497; CHECK-NEXT:    [[BC_RESUME_VAL3:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[VECTOR_MEMCHECK]] ], [ 0, [[ENTRY]] ]
1498; CHECK-NEXT:    br label [[LOOP:%.*]]
1499; CHECK:       loop:
1500; CHECK-NEXT:    [[TMP33:%.*]] = phi i16 [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ], [ [[LOAD2:%.*]], [[LOOP]] ]
1501; CHECK-NEXT:    [[IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[IV2:%.*]], [[LOOP]] ]
1502; CHECK-NEXT:    [[I:%.*]] = phi i64 [ [[BC_RESUME_VAL3]], [[SCALAR_PH]] ], [ [[I1:%.*]], [[LOOP]] ]
1503; CHECK-NEXT:    [[CONV:%.*]] = sext i16 [[TMP33]] to i32
1504; CHECK-NEXT:    [[I1]] = add nuw nsw i64 [[I]], 1
1505; CHECK-NEXT:    [[IV1:%.*]] = or disjoint i64 [[IV]], 1
1506; CHECK-NEXT:    [[IV2]] = add nuw nsw i64 [[IV]], 2
1507; CHECK-NEXT:    [[GEP1:%.*]] = getelementptr inbounds i16, ptr [[A]], i64 [[IV1]]
1508; CHECK-NEXT:    [[LOAD1:%.*]] = load i16, ptr [[GEP1]], align 4
1509; CHECK-NEXT:    [[CONV1:%.*]] = sext i16 [[LOAD1]] to i32
1510; CHECK-NEXT:    [[GEP2:%.*]] = getelementptr inbounds i16, ptr [[A]], i64 [[IV2]]
1511; CHECK-NEXT:    [[LOAD2]] = load i16, ptr [[GEP2]], align 4
1512; CHECK-NEXT:    [[CONV2:%.*]] = sext i16 [[LOAD2]] to i32
1513; CHECK-NEXT:    [[MUL01:%.*]] = mul nsw i32 [[CONV]], [[CONV1]]
1514; CHECK-NEXT:    [[MUL012:%.*]] = mul nsw i32 [[MUL01]], [[CONV2]]
1515; CHECK-NEXT:    [[ARRAYIDX5:%.*]] = getelementptr inbounds i32, ptr [[B]], i64 [[I]]
1516; CHECK-NEXT:    store i32 [[MUL012]], ptr [[ARRAYIDX5]], align 4
1517; CHECK-NEXT:    [[EXITCOND:%.*]] = icmp eq i64 [[IV]], [[N]]
1518; CHECK-NEXT:    br i1 [[EXITCOND]], label [[END]], label [[LOOP]], !llvm.loop [[LOOP40:![0-9]+]]
1519; CHECK:       end:
1520; CHECK-NEXT:    ret void
1521;
1522entry:
1523  %.pre = load i16, ptr %a
1524  br label %loop
1525
1526loop:
1527  %0 = phi i16 [ %.pre, %entry ], [ %load2, %loop ]
1528  %iv = phi i64 [ 0, %entry ], [ %iv2, %loop ]
1529  %i = phi i64 [ 0, %entry ], [ %i1, %loop ]
1530  %conv = sext i16 %0 to i32
1531  %i1 = add nuw nsw i64 %i, 1
1532  %iv1 = add nuw nsw i64 %iv, 1
1533  %iv2 = add nuw nsw i64 %iv, 2
1534  %gep1 = getelementptr inbounds i16, ptr %a, i64 %iv1
1535  %load1 = load i16, ptr %gep1, align 4
1536  %conv1 = sext i16 %load1 to i32
1537  %gep2 = getelementptr inbounds i16, ptr %a, i64 %iv2
1538  %load2 = load i16, ptr %gep2, align 4
1539  %conv2 = sext i16 %load2 to i32
1540  %mul01 = mul nsw i32 %conv, %conv1
1541  %mul012 = mul nsw i32 %mul01, %conv2
1542  %arrayidx5 = getelementptr inbounds i32, ptr %b, i64 %i
1543  store i32 %mul012, ptr %arrayidx5
1544  %exitcond = icmp eq i64 %iv, %n
1545  br i1 %exitcond, label %end, label %loop
1546
1547end:
1548  ret void
1549}
1550
1551attributes #1 = { "target-features"="+sve" vscale_range(1, 16) }
1552attributes #0 = { "unsafe-fp-math"="true" "target-features"="+sve" vscale_range(1, 16) }
1553