• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
2
3; To test epilogue-vectorization we need to make sure that the vectorizer actually vectorizes the loop.
4; Without a target triple this becomes difficult, unless we force vectorization through user hints.
5; Currently user provided vectorization hints prevent epilogue vectorization unless the forced
6; VF is the same as the epilogue vectorization VF. To make these tests target independent we'll use a
7; trick where both VFs are forced to be the same value. Mismatching VFs are tested in target specific tests.
8; RUN: opt -passes='loop-vectorize' -force-vector-width=4 -enable-epilogue-vectorization -epilogue-vectorization-force-VF=4 -S %s | FileCheck %s
9
10; Some simpler cases are found profitable even without triple or user hints.
11; RUN: opt -passes='loop-vectorize' -enable-epilogue-vectorization -epilogue-vectorization-force-VF=2 -S %s | FileCheck --check-prefix=CHECK-PROFITABLE-BY-DEFAULT %s
12
13target datalayout = "e-m:e-i64:64-n32:64-v128:128:128"
14
15define dso_local void @f1(float* noalias %aa, float* noalias %bb, float* noalias %cc, i32 signext %N) {
16; CHECK-LABEL: @f1(
17; CHECK-NEXT:  entry:
18; CHECK-NEXT:    [[AA1:%.*]] = bitcast float* [[AA:%.*]] to i8*
19; CHECK-NEXT:    [[BB3:%.*]] = bitcast float* [[BB:%.*]] to i8*
20; CHECK-NEXT:    [[CC6:%.*]] = bitcast float* [[CC:%.*]] to i8*
21; CHECK-NEXT:    [[CMP1:%.*]] = icmp sgt i32 [[N:%.*]], 0
22; CHECK-NEXT:    br i1 [[CMP1]], label [[ITER_CHECK:%.*]], label [[FOR_END:%.*]]
23; CHECK:       iter.check:
24; CHECK-NEXT:    [[WIDE_TRIP_COUNT:%.*]] = zext i32 [[N]] to i64
25; CHECK-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[WIDE_TRIP_COUNT]], 4
26; CHECK-NEXT:    br i1 [[MIN_ITERS_CHECK]], label [[VEC_EPILOG_SCALAR_PH:%.*]], label [[VECTOR_MEMCHECK:%.*]]
27; CHECK:       vector.memcheck:
28; CHECK-NEXT:    [[SCEVGEP:%.*]] = getelementptr float, float* [[AA]], i64 [[WIDE_TRIP_COUNT]]
29; CHECK-NEXT:    [[SCEVGEP2:%.*]] = bitcast float* [[SCEVGEP]] to i8*
30; CHECK-NEXT:    [[SCEVGEP4:%.*]] = getelementptr float, float* [[BB]], i64 [[WIDE_TRIP_COUNT]]
31; CHECK-NEXT:    [[SCEVGEP45:%.*]] = bitcast float* [[SCEVGEP4]] to i8*
32; CHECK-NEXT:    [[SCEVGEP7:%.*]] = getelementptr float, float* [[CC]], i64 [[WIDE_TRIP_COUNT]]
33; CHECK-NEXT:    [[SCEVGEP78:%.*]] = bitcast float* [[SCEVGEP7]] to i8*
34; CHECK-NEXT:    [[BOUND0:%.*]] = icmp ult i8* [[AA1]], [[SCEVGEP45]]
35; CHECK-NEXT:    [[BOUND1:%.*]] = icmp ult i8* [[BB3]], [[SCEVGEP2]]
36; CHECK-NEXT:    [[FOUND_CONFLICT:%.*]] = and i1 [[BOUND0]], [[BOUND1]]
37; CHECK-NEXT:    [[BOUND09:%.*]] = icmp ult i8* [[AA1]], [[SCEVGEP78]]
38; CHECK-NEXT:    [[BOUND110:%.*]] = icmp ult i8* [[CC6]], [[SCEVGEP2]]
39; CHECK-NEXT:    [[FOUND_CONFLICT11:%.*]] = and i1 [[BOUND09]], [[BOUND110]]
40; CHECK-NEXT:    [[CONFLICT_RDX:%.*]] = or i1 [[FOUND_CONFLICT]], [[FOUND_CONFLICT11]]
41; CHECK-NEXT:    [[MEMCHECK_CONFLICT:%.*]] = and i1 [[CONFLICT_RDX]], true
42; CHECK-NEXT:    br i1 [[MEMCHECK_CONFLICT]], label [[VEC_EPILOG_SCALAR_PH]], label [[VECTOR_MAIN_LOOP_ITER_CHECK:%.*]]
43; CHECK:       vector.main.loop.iter.check:
44; CHECK-NEXT:    [[MIN_ITERS_CHECK12:%.*]] = icmp ult i64 [[WIDE_TRIP_COUNT]], 4
45; CHECK-NEXT:    br i1 [[MIN_ITERS_CHECK12]], label [[VEC_EPILOG_PH:%.*]], label [[VECTOR_PH:%.*]]
46; CHECK:       vector.ph:
47; CHECK-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[WIDE_TRIP_COUNT]], 4
48; CHECK-NEXT:    [[N_VEC:%.*]] = sub i64 [[WIDE_TRIP_COUNT]], [[N_MOD_VF]]
49; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
50; CHECK:       vector.body:
51; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
52; CHECK-NEXT:    [[TMP0:%.*]] = add i64 [[INDEX]], 0
53; CHECK-NEXT:    [[TMP1:%.*]] = getelementptr inbounds float, float* [[BB]], i64 [[TMP0]]
54; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr inbounds float, float* [[TMP1]], i32 0
55; CHECK-NEXT:    [[TMP3:%.*]] = bitcast float* [[TMP2]] to <4 x float>*
56; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x float>, <4 x float>* [[TMP3]], align 4, !alias.scope !0
57; CHECK-NEXT:    [[TMP4:%.*]] = getelementptr inbounds float, float* [[CC]], i64 [[TMP0]]
58; CHECK-NEXT:    [[TMP5:%.*]] = getelementptr inbounds float, float* [[TMP4]], i32 0
59; CHECK-NEXT:    [[TMP6:%.*]] = bitcast float* [[TMP5]] to <4 x float>*
60; CHECK-NEXT:    [[WIDE_LOAD13:%.*]] = load <4 x float>, <4 x float>* [[TMP6]], align 4, !alias.scope !3
61; CHECK-NEXT:    [[TMP7:%.*]] = fadd fast <4 x float> [[WIDE_LOAD]], [[WIDE_LOAD13]]
62; CHECK-NEXT:    [[TMP8:%.*]] = getelementptr inbounds float, float* [[AA]], i64 [[TMP0]]
63; CHECK-NEXT:    [[TMP9:%.*]] = getelementptr inbounds float, float* [[TMP8]], i32 0
64; CHECK-NEXT:    [[TMP10:%.*]] = bitcast float* [[TMP9]] to <4 x float>*
65; CHECK-NEXT:    store <4 x float> [[TMP7]], <4 x float>* [[TMP10]], align 4, !alias.scope !5, !noalias !7
66; CHECK-NEXT:    [[INDEX_NEXT]] = add i64 [[INDEX]], 4
67; CHECK-NEXT:    [[TMP11:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
68; CHECK-NEXT:    br i1 [[TMP11]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], [[LOOP8:!llvm.loop !.*]]
69; CHECK:       middle.block:
70; CHECK-NEXT:    [[CMP_N:%.*]] = icmp eq i64 [[WIDE_TRIP_COUNT]], [[N_VEC]]
71; CHECK-NEXT:    br i1 [[CMP_N]], label [[FOR_END_LOOPEXIT:%.*]], label [[VEC_EPILOG_ITER_CHECK:%.*]]
72; CHECK:       vec.epilog.iter.check:
73; CHECK-NEXT:    [[N_VEC_REMAINING:%.*]] = sub i64 [[WIDE_TRIP_COUNT]], [[N_VEC]]
74; CHECK-NEXT:    [[MIN_EPILOG_ITERS_CHECK:%.*]] = icmp ult i64 [[N_VEC_REMAINING]], 4
75; CHECK-NEXT:    br i1 [[MIN_EPILOG_ITERS_CHECK]], label [[VEC_EPILOG_SCALAR_PH]], label [[VEC_EPILOG_PH]]
76; CHECK:       vec.epilog.ph:
77; CHECK-NEXT:    [[VEC_EPILOG_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[VEC_EPILOG_ITER_CHECK]] ], [ 0, [[VECTOR_MAIN_LOOP_ITER_CHECK]] ]
78; CHECK-NEXT:    [[N_MOD_VF14:%.*]] = urem i64 [[WIDE_TRIP_COUNT]], 4
79; CHECK-NEXT:    [[N_VEC15:%.*]] = sub i64 [[WIDE_TRIP_COUNT]], [[N_MOD_VF14]]
80; CHECK-NEXT:    br label [[VEC_EPILOG_VECTOR_BODY:%.*]]
81; CHECK:       vec.epilog.vector.body:
82; CHECK-NEXT:    [[INDEX16:%.*]] = phi i64 [ [[VEC_EPILOG_RESUME_VAL]], [[VEC_EPILOG_PH]] ], [ [[INDEX_NEXT17:%.*]], [[VEC_EPILOG_VECTOR_BODY]] ]
83; CHECK-NEXT:    [[TMP12:%.*]] = add i64 [[INDEX16]], 0
84; CHECK-NEXT:    [[TMP13:%.*]] = getelementptr inbounds float, float* [[BB]], i64 [[TMP12]]
85; CHECK-NEXT:    [[TMP14:%.*]] = getelementptr inbounds float, float* [[TMP13]], i32 0
86; CHECK-NEXT:    [[TMP15:%.*]] = bitcast float* [[TMP14]] to <4 x float>*
87; CHECK-NEXT:    [[WIDE_LOAD19:%.*]] = load <4 x float>, <4 x float>* [[TMP15]], align 4
88; CHECK-NEXT:    [[TMP16:%.*]] = getelementptr inbounds float, float* [[CC]], i64 [[TMP12]]
89; CHECK-NEXT:    [[TMP17:%.*]] = getelementptr inbounds float, float* [[TMP16]], i32 0
90; CHECK-NEXT:    [[TMP18:%.*]] = bitcast float* [[TMP17]] to <4 x float>*
91; CHECK-NEXT:    [[WIDE_LOAD20:%.*]] = load <4 x float>, <4 x float>* [[TMP18]], align 4
92; CHECK-NEXT:    [[TMP19:%.*]] = fadd fast <4 x float> [[WIDE_LOAD19]], [[WIDE_LOAD20]]
93; CHECK-NEXT:    [[TMP20:%.*]] = getelementptr inbounds float, float* [[AA]], i64 [[TMP12]]
94; CHECK-NEXT:    [[TMP21:%.*]] = getelementptr inbounds float, float* [[TMP20]], i32 0
95; CHECK-NEXT:    [[TMP22:%.*]] = bitcast float* [[TMP21]] to <4 x float>*
96; CHECK-NEXT:    store <4 x float> [[TMP19]], <4 x float>* [[TMP22]], align 4
97; CHECK-NEXT:    [[INDEX_NEXT17]] = add i64 [[INDEX16]], 4
98; CHECK-NEXT:    [[TMP23:%.*]] = icmp eq i64 [[INDEX_NEXT17]], [[N_VEC15]]
99; CHECK-NEXT:    br i1 [[TMP23]], label [[VEC_EPILOG_MIDDLE_BLOCK:%.*]], label [[VEC_EPILOG_VECTOR_BODY]], [[LOOP10:!llvm.loop !.*]]
100; CHECK:       vec.epilog.middle.block:
101; CHECK-NEXT:    [[CMP_N18:%.*]] = icmp eq i64 [[WIDE_TRIP_COUNT]], [[N_VEC15]]
102; CHECK-NEXT:    br i1 [[CMP_N18]], label [[FOR_END_LOOPEXIT_LOOPEXIT:%.*]], label [[VEC_EPILOG_SCALAR_PH]]
103; CHECK:       vec.epilog.scalar.ph:
104; CHECK-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC15]], [[VEC_EPILOG_MIDDLE_BLOCK]] ], [ [[N_VEC]], [[VEC_EPILOG_ITER_CHECK]] ], [ 0, [[VECTOR_MEMCHECK]] ], [ 0, [[ITER_CHECK]] ]
105; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
106; CHECK:       for.body:
107; CHECK-NEXT:    [[INDVARS_IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[VEC_EPILOG_SCALAR_PH]] ], [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ]
108; CHECK-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds float, float* [[BB]], i64 [[INDVARS_IV]]
109; CHECK-NEXT:    [[TMP24:%.*]] = load float, float* [[ARRAYIDX]], align 4
110; CHECK-NEXT:    [[ARRAYIDX2:%.*]] = getelementptr inbounds float, float* [[CC]], i64 [[INDVARS_IV]]
111; CHECK-NEXT:    [[TMP25:%.*]] = load float, float* [[ARRAYIDX2]], align 4
112; CHECK-NEXT:    [[ADD:%.*]] = fadd fast float [[TMP24]], [[TMP25]]
113; CHECK-NEXT:    [[ARRAYIDX4:%.*]] = getelementptr inbounds float, float* [[AA]], i64 [[INDVARS_IV]]
114; CHECK-NEXT:    store float [[ADD]], float* [[ARRAYIDX4]], align 4
115; CHECK-NEXT:    [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1
116; CHECK-NEXT:    [[EXITCOND:%.*]] = icmp ne i64 [[INDVARS_IV_NEXT]], [[WIDE_TRIP_COUNT]]
117; CHECK-NEXT:    br i1 [[EXITCOND]], label [[FOR_BODY]], label [[FOR_END_LOOPEXIT_LOOPEXIT]], [[LOOP12:!llvm.loop !.*]]
118; CHECK:       for.end.loopexit.loopexit:
119; CHECK-NEXT:    br label [[FOR_END_LOOPEXIT]]
120; CHECK:       for.end.loopexit:
121; CHECK-NEXT:    br label [[FOR_END]]
122; CHECK:       for.end:
123; CHECK-NEXT:    ret void
124;
125entry:
126  %cmp1 = icmp sgt i32 %N, 0
127  br i1 %cmp1, label %for.body.preheader, label %for.end
128
129for.body.preheader:                               ; preds = %entry
130  %wide.trip.count = zext i32 %N to i64
131  br label %for.body
132
133for.body:                                         ; preds = %for.body.preheader, %for.body
134  %indvars.iv = phi i64 [ 0, %for.body.preheader ], [ %indvars.iv.next, %for.body ]
135  %arrayidx = getelementptr inbounds float, float* %bb, i64 %indvars.iv
136  %0 = load float, float* %arrayidx, align 4
137  %arrayidx2 = getelementptr inbounds float, float* %cc, i64 %indvars.iv
138  %1 = load float, float* %arrayidx2, align 4
139  %add = fadd fast float %0, %1
140  %arrayidx4 = getelementptr inbounds float, float* %aa, i64 %indvars.iv
141  store float %add, float* %arrayidx4, align 4
142  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
143  %exitcond = icmp ne i64 %indvars.iv.next, %wide.trip.count
144  br i1 %exitcond, label %for.body, label %for.end.loopexit
145
146for.end.loopexit:                                 ; preds = %for.body
147  br label %for.end
148
149for.end:                                          ; preds = %for.end.loopexit, %entry
150  ret void
151}
152
153define dso_local signext i32 @f2(float* noalias %A, float* noalias %B, i32 signext %n) {
154; CHECK-LABEL: @f2(
155; CHECK-NEXT:  entry:
156; CHECK-NEXT:    [[A1:%.*]] = bitcast float* [[A:%.*]] to i8*
157; CHECK-NEXT:    [[CMP1:%.*]] = icmp sgt i32 [[N:%.*]], 1
158; CHECK-NEXT:    br i1 [[CMP1]], label [[ITER_CHECK:%.*]], label [[FOR_END:%.*]]
159; CHECK:       iter.check:
160; CHECK-NEXT:    [[TMP0:%.*]] = add i32 [[N]], -1
161; CHECK-NEXT:    [[WIDE_TRIP_COUNT:%.*]] = zext i32 [[TMP0]] to i64
162; CHECK-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[WIDE_TRIP_COUNT]], 4
163; CHECK-NEXT:    br i1 [[MIN_ITERS_CHECK]], label [[VEC_EPILOG_SCALAR_PH:%.*]], label [[VECTOR_SCEVCHECK:%.*]]
164; CHECK:       vector.scevcheck:
165; CHECK-NEXT:    [[TMP1:%.*]] = add nsw i64 [[WIDE_TRIP_COUNT]], -1
166; CHECK-NEXT:    [[TMP2:%.*]] = trunc i64 [[TMP1]] to i32
167; CHECK-NEXT:    [[MUL:%.*]] = call { i32, i1 } @llvm.umul.with.overflow.i32(i32 1, i32 [[TMP2]])
168; CHECK-NEXT:    [[MUL_RESULT:%.*]] = extractvalue { i32, i1 } [[MUL]], 0
169; CHECK-NEXT:    [[MUL_OVERFLOW:%.*]] = extractvalue { i32, i1 } [[MUL]], 1
170; CHECK-NEXT:    [[TMP3:%.*]] = add i32 [[TMP0]], [[MUL_RESULT]]
171; CHECK-NEXT:    [[TMP4:%.*]] = sub i32 [[TMP0]], [[MUL_RESULT]]
172; CHECK-NEXT:    [[TMP5:%.*]] = icmp sgt i32 [[TMP4]], [[TMP0]]
173; CHECK-NEXT:    [[TMP6:%.*]] = icmp slt i32 [[TMP3]], [[TMP0]]
174; CHECK-NEXT:    [[TMP7:%.*]] = select i1 true, i1 [[TMP5]], i1 [[TMP6]]
175; CHECK-NEXT:    [[TMP8:%.*]] = icmp ugt i64 [[TMP1]], 4294967295
176; CHECK-NEXT:    [[TMP9:%.*]] = or i1 [[TMP7]], [[TMP8]]
177; CHECK-NEXT:    [[TMP10:%.*]] = or i1 [[TMP9]], [[MUL_OVERFLOW]]
178; CHECK-NEXT:    [[TMP11:%.*]] = or i1 false, [[TMP10]]
179; CHECK-NEXT:    br i1 [[TMP11]], label [[VEC_EPILOG_SCALAR_PH]], label [[VECTOR_MEMCHECK:%.*]]
180; CHECK:       vector.memcheck:
181; CHECK-NEXT:    [[SCEVGEP:%.*]] = getelementptr float, float* [[A]], i64 [[WIDE_TRIP_COUNT]]
182; CHECK-NEXT:    [[SCEVGEP2:%.*]] = bitcast float* [[SCEVGEP]] to i8*
183; CHECK-NEXT:    [[TMP12:%.*]] = sext i32 [[TMP0]] to i64
184; CHECK-NEXT:    [[TMP13:%.*]] = add i64 [[TMP12]], 1
185; CHECK-NEXT:    [[TMP14:%.*]] = sub i64 [[TMP13]], [[WIDE_TRIP_COUNT]]
186; CHECK-NEXT:    [[SCEVGEP3:%.*]] = getelementptr float, float* [[B:%.*]], i64 [[TMP14]]
187; CHECK-NEXT:    [[SCEVGEP34:%.*]] = bitcast float* [[SCEVGEP3]] to i8*
188; CHECK-NEXT:    [[TMP15:%.*]] = add nsw i64 [[TMP12]], 1
189; CHECK-NEXT:    [[SCEVGEP5:%.*]] = getelementptr float, float* [[B]], i64 [[TMP15]]
190; CHECK-NEXT:    [[SCEVGEP56:%.*]] = bitcast float* [[SCEVGEP5]] to i8*
191; CHECK-NEXT:    [[BOUND0:%.*]] = icmp ult i8* [[A1]], [[SCEVGEP56]]
192; CHECK-NEXT:    [[BOUND1:%.*]] = icmp ult i8* [[SCEVGEP34]], [[SCEVGEP2]]
193; CHECK-NEXT:    [[FOUND_CONFLICT:%.*]] = and i1 [[BOUND0]], [[BOUND1]]
194; CHECK-NEXT:    [[MEMCHECK_CONFLICT:%.*]] = and i1 [[FOUND_CONFLICT]], true
195; CHECK-NEXT:    br i1 [[MEMCHECK_CONFLICT]], label [[VEC_EPILOG_SCALAR_PH]], label [[VECTOR_MAIN_LOOP_ITER_CHECK:%.*]]
196; CHECK:       vector.main.loop.iter.check:
197; CHECK-NEXT:    [[MIN_ITERS_CHECK7:%.*]] = icmp ult i64 [[WIDE_TRIP_COUNT]], 4
198; CHECK-NEXT:    br i1 [[MIN_ITERS_CHECK7]], label [[VEC_EPILOG_PH:%.*]], label [[VECTOR_PH:%.*]]
199; CHECK:       vector.ph:
200; CHECK-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[WIDE_TRIP_COUNT]], 4
201; CHECK-NEXT:    [[N_VEC:%.*]] = sub i64 [[WIDE_TRIP_COUNT]], [[N_MOD_VF]]
202; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
203; CHECK:       vector.body:
204; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
205; CHECK-NEXT:    [[TMP16:%.*]] = add i64 [[INDEX]], 0
206; CHECK-NEXT:    [[OFFSET_IDX:%.*]] = trunc i64 [[INDEX]] to i32
207; CHECK-NEXT:    [[TMP17:%.*]] = add i32 [[OFFSET_IDX]], 0
208; CHECK-NEXT:    [[TMP18:%.*]] = xor i32 [[TMP17]], -1
209; CHECK-NEXT:    [[TMP19:%.*]] = add i32 [[TMP18]], [[N]]
210; CHECK-NEXT:    [[TMP20:%.*]] = sext i32 [[TMP19]] to i64
211; CHECK-NEXT:    [[TMP21:%.*]] = getelementptr inbounds float, float* [[B]], i64 [[TMP20]]
212; CHECK-NEXT:    [[TMP22:%.*]] = getelementptr inbounds float, float* [[TMP21]], i32 0
213; CHECK-NEXT:    [[TMP23:%.*]] = getelementptr inbounds float, float* [[TMP22]], i32 -3
214; CHECK-NEXT:    [[TMP24:%.*]] = bitcast float* [[TMP23]] to <4 x float>*
215; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x float>, <4 x float>* [[TMP24]], align 4, !alias.scope !13
216; CHECK-NEXT:    [[REVERSE:%.*]] = shufflevector <4 x float> [[WIDE_LOAD]], <4 x float> undef, <4 x i32> <i32 3, i32 2, i32 1, i32 0>
217; CHECK-NEXT:    [[TMP25:%.*]] = fadd fast <4 x float> [[REVERSE]], <float 1.000000e+00, float 1.000000e+00, float 1.000000e+00, float 1.000000e+00>
218; CHECK-NEXT:    [[TMP26:%.*]] = getelementptr inbounds float, float* [[A]], i64 [[TMP16]]
219; CHECK-NEXT:    [[TMP27:%.*]] = getelementptr inbounds float, float* [[TMP26]], i32 0
220; CHECK-NEXT:    [[TMP28:%.*]] = bitcast float* [[TMP27]] to <4 x float>*
221; CHECK-NEXT:    store <4 x float> [[TMP25]], <4 x float>* [[TMP28]], align 4, !alias.scope !16, !noalias !13
222; CHECK-NEXT:    [[INDEX_NEXT]] = add i64 [[INDEX]], 4
223; CHECK-NEXT:    [[TMP29:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
224; CHECK-NEXT:    br i1 [[TMP29]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], [[LOOP18:!llvm.loop !.*]]
225; CHECK:       middle.block:
226; CHECK-NEXT:    [[CMP_N:%.*]] = icmp eq i64 [[WIDE_TRIP_COUNT]], [[N_VEC]]
227; CHECK-NEXT:    br i1 [[CMP_N]], label [[FOR_END_LOOPEXIT:%.*]], label [[VEC_EPILOG_ITER_CHECK:%.*]]
228; CHECK:       vec.epilog.iter.check:
229; CHECK-NEXT:    [[IND_END13:%.*]] = trunc i64 [[N_VEC]] to i32
230; CHECK-NEXT:    [[N_VEC_REMAINING:%.*]] = sub i64 [[WIDE_TRIP_COUNT]], [[N_VEC]]
231; CHECK-NEXT:    [[MIN_EPILOG_ITERS_CHECK:%.*]] = icmp ult i64 [[N_VEC_REMAINING]], 4
232; CHECK-NEXT:    br i1 [[MIN_EPILOG_ITERS_CHECK]], label [[VEC_EPILOG_SCALAR_PH]], label [[VEC_EPILOG_PH]]
233; CHECK:       vec.epilog.ph:
234; CHECK-NEXT:    [[VEC_EPILOG_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[VEC_EPILOG_ITER_CHECK]] ], [ 0, [[VECTOR_MAIN_LOOP_ITER_CHECK]] ]
235; CHECK-NEXT:    [[N_MOD_VF8:%.*]] = urem i64 [[WIDE_TRIP_COUNT]], 4
236; CHECK-NEXT:    [[N_VEC9:%.*]] = sub i64 [[WIDE_TRIP_COUNT]], [[N_MOD_VF8]]
237; CHECK-NEXT:    [[IND_END:%.*]] = trunc i64 [[N_VEC9]] to i32
238; CHECK-NEXT:    br label [[VEC_EPILOG_VECTOR_BODY:%.*]]
239; CHECK:       vec.epilog.vector.body:
240; CHECK-NEXT:    [[INDEX10:%.*]] = phi i64 [ [[VEC_EPILOG_RESUME_VAL]], [[VEC_EPILOG_PH]] ], [ [[INDEX_NEXT11:%.*]], [[VEC_EPILOG_VECTOR_BODY]] ]
241; CHECK-NEXT:    [[TMP30:%.*]] = add i64 [[INDEX10]], 0
242; CHECK-NEXT:    [[OFFSET_IDX15:%.*]] = trunc i64 [[INDEX10]] to i32
243; CHECK-NEXT:    [[TMP31:%.*]] = add i32 [[OFFSET_IDX15]], 0
244; CHECK-NEXT:    [[TMP32:%.*]] = xor i32 [[TMP31]], -1
245; CHECK-NEXT:    [[TMP33:%.*]] = add i32 [[TMP32]], [[N]]
246; CHECK-NEXT:    [[TMP34:%.*]] = sext i32 [[TMP33]] to i64
247; CHECK-NEXT:    [[TMP35:%.*]] = getelementptr inbounds float, float* [[B]], i64 [[TMP34]]
248; CHECK-NEXT:    [[TMP36:%.*]] = getelementptr inbounds float, float* [[TMP35]], i32 0
249; CHECK-NEXT:    [[TMP37:%.*]] = getelementptr inbounds float, float* [[TMP36]], i32 -3
250; CHECK-NEXT:    [[TMP38:%.*]] = bitcast float* [[TMP37]] to <4 x float>*
251; CHECK-NEXT:    [[WIDE_LOAD16:%.*]] = load <4 x float>, <4 x float>* [[TMP38]], align 4
252; CHECK-NEXT:    [[REVERSE17:%.*]] = shufflevector <4 x float> [[WIDE_LOAD16]], <4 x float> undef, <4 x i32> <i32 3, i32 2, i32 1, i32 0>
253; CHECK-NEXT:    [[TMP39:%.*]] = fadd fast <4 x float> [[REVERSE17]], <float 1.000000e+00, float 1.000000e+00, float 1.000000e+00, float 1.000000e+00>
254; CHECK-NEXT:    [[TMP40:%.*]] = getelementptr inbounds float, float* [[A]], i64 [[TMP30]]
255; CHECK-NEXT:    [[TMP41:%.*]] = getelementptr inbounds float, float* [[TMP40]], i32 0
256; CHECK-NEXT:    [[TMP42:%.*]] = bitcast float* [[TMP41]] to <4 x float>*
257; CHECK-NEXT:    store <4 x float> [[TMP39]], <4 x float>* [[TMP42]], align 4
258; CHECK-NEXT:    [[INDEX_NEXT11]] = add i64 [[INDEX10]], 4
259; CHECK-NEXT:    [[TMP43:%.*]] = icmp eq i64 [[INDEX_NEXT11]], [[N_VEC9]]
260; CHECK-NEXT:    br i1 [[TMP43]], label [[VEC_EPILOG_MIDDLE_BLOCK:%.*]], label [[VEC_EPILOG_VECTOR_BODY]], [[LOOP19:!llvm.loop !.*]]
261; CHECK:       vec.epilog.middle.block:
262; CHECK-NEXT:    [[CMP_N14:%.*]] = icmp eq i64 [[WIDE_TRIP_COUNT]], [[N_VEC9]]
263; CHECK-NEXT:    br i1 [[CMP_N14]], label [[FOR_END_LOOPEXIT_LOOPEXIT:%.*]], label [[VEC_EPILOG_SCALAR_PH]]
264; CHECK:       vec.epilog.scalar.ph:
265; CHECK-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC9]], [[VEC_EPILOG_MIDDLE_BLOCK]] ], [ [[N_VEC]], [[VEC_EPILOG_ITER_CHECK]] ], [ 0, [[VECTOR_SCEVCHECK]] ], [ 0, [[VECTOR_MEMCHECK]] ], [ 0, [[ITER_CHECK]] ]
266; CHECK-NEXT:    [[BC_RESUME_VAL12:%.*]] = phi i32 [ [[IND_END]], [[VEC_EPILOG_MIDDLE_BLOCK]] ], [ [[IND_END13]], [[VEC_EPILOG_ITER_CHECK]] ], [ 0, [[VECTOR_SCEVCHECK]] ], [ 0, [[VECTOR_MEMCHECK]] ], [ 0, [[ITER_CHECK]] ]
267; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
268; CHECK:       for.body:
269; CHECK-NEXT:    [[INDVARS_IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[VEC_EPILOG_SCALAR_PH]] ], [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ]
270; CHECK-NEXT:    [[I_014:%.*]] = phi i32 [ [[BC_RESUME_VAL12]], [[VEC_EPILOG_SCALAR_PH]] ], [ [[INC:%.*]], [[FOR_BODY]] ]
271; CHECK-NEXT:    [[TMP44:%.*]] = xor i32 [[I_014]], -1
272; CHECK-NEXT:    [[SUB2:%.*]] = add i32 [[TMP44]], [[N]]
273; CHECK-NEXT:    [[IDXPROM:%.*]] = sext i32 [[SUB2]] to i64
274; CHECK-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds float, float* [[B]], i64 [[IDXPROM]]
275; CHECK-NEXT:    [[TMP45:%.*]] = load float, float* [[ARRAYIDX]], align 4
276; CHECK-NEXT:    [[CONV3:%.*]] = fadd fast float [[TMP45]], 1.000000e+00
277; CHECK-NEXT:    [[ARRAYIDX5:%.*]] = getelementptr inbounds float, float* [[A]], i64 [[INDVARS_IV]]
278; CHECK-NEXT:    store float [[CONV3]], float* [[ARRAYIDX5]], align 4
279; CHECK-NEXT:    [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1
280; CHECK-NEXT:    [[INC]] = add nuw nsw i32 [[I_014]], 1
281; CHECK-NEXT:    [[EXITCOND:%.*]] = icmp ne i64 [[INDVARS_IV_NEXT]], [[WIDE_TRIP_COUNT]]
282; CHECK-NEXT:    br i1 [[EXITCOND]], label [[FOR_BODY]], label [[FOR_END_LOOPEXIT_LOOPEXIT]], [[LOOP20:!llvm.loop !.*]]
283; CHECK:       for.end.loopexit.loopexit:
284; CHECK-NEXT:    br label [[FOR_END_LOOPEXIT]]
285; CHECK:       for.end.loopexit:
286; CHECK-NEXT:    br label [[FOR_END]]
287; CHECK:       for.end:
288; CHECK-NEXT:    ret i32 0
289;
290entry:
291  %cmp1 = icmp sgt i32 %n, 1
292  br i1 %cmp1, label %for.body.preheader, label %for.end
293
294for.body.preheader:                               ; preds = %entry
295  %0 = add i32 %n, -1
296  %wide.trip.count = zext i32 %0 to i64
297  br label %for.body
298
299for.body:                                         ; preds = %for.body.preheader, %for.body
300  %indvars.iv = phi i64 [ 0, %for.body.preheader ], [ %indvars.iv.next, %for.body ]
301  %i.014 = phi i32 [ 0, %for.body.preheader ], [ %inc, %for.body ]
302  %1 = xor i32 %i.014, -1
303  %sub2 = add i32 %1, %n
304  %idxprom = sext i32 %sub2 to i64
305  %arrayidx = getelementptr inbounds float, float* %B, i64 %idxprom
306  %2 = load float, float* %arrayidx, align 4
307  %conv3 = fadd fast float %2, 1.000000e+00
308  %arrayidx5 = getelementptr inbounds float, float* %A, i64 %indvars.iv
309  store float %conv3, float* %arrayidx5, align 4
310  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
311  %inc = add nuw nsw i32 %i.014, 1
312  %exitcond = icmp ne i64 %indvars.iv.next, %wide.trip.count
313  br i1 %exitcond, label %for.body, label %for.end.loopexit
314
315for.end.loopexit:                                 ; preds = %for.body
316  br label %for.end
317
318for.end:                                          ; preds = %for.end.loopexit, %entry
319  ret i32 0
320}
321
322define void @f3(i8* noalias %A, i64 %n) {
323; CHECK-PROFITABLE-BY-DEFAULT-LABEL: @f3(
324; CHECK-PROFITABLE-BY-DEFAULT-NEXT:  iter.check:
325; CHECK-PROFITABLE-BY-DEFAULT-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[N:%.*]], 2
326; CHECK-PROFITABLE-BY-DEFAULT-NEXT:    br i1 [[MIN_ITERS_CHECK]], label [[VEC_EPILOG_SCALAR_PH:%.*]], label [[VECTOR_MAIN_LOOP_ITER_CHECK:%.*]]
327; CHECK-PROFITABLE-BY-DEFAULT:       vector.main.loop.iter.check:
328; CHECK-PROFITABLE-BY-DEFAULT-NEXT:    [[MIN_ITERS_CHECK1:%.*]] = icmp ult i64 [[N]], 4
329; CHECK-PROFITABLE-BY-DEFAULT-NEXT:    br i1 [[MIN_ITERS_CHECK1]], label [[VEC_EPILOG_PH:%.*]], label [[VECTOR_PH:%.*]]
330; CHECK-PROFITABLE-BY-DEFAULT:       vector.ph:
331; CHECK-PROFITABLE-BY-DEFAULT-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[N]], 4
332; CHECK-PROFITABLE-BY-DEFAULT-NEXT:    [[N_VEC:%.*]] = sub i64 [[N]], [[N_MOD_VF]]
333; CHECK-PROFITABLE-BY-DEFAULT-NEXT:    br label [[VECTOR_BODY:%.*]]
334; CHECK-PROFITABLE-BY-DEFAULT:       vector.body:
335; CHECK-PROFITABLE-BY-DEFAULT-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
336; CHECK-PROFITABLE-BY-DEFAULT-NEXT:    [[TMP0:%.*]] = add i64 [[INDEX]], 0
337; CHECK-PROFITABLE-BY-DEFAULT-NEXT:    [[TMP1:%.*]] = getelementptr inbounds i8, i8* [[A:%.*]], i64 [[TMP0]]
338; CHECK-PROFITABLE-BY-DEFAULT-NEXT:    [[TMP2:%.*]] = getelementptr inbounds i8, i8* [[TMP1]], i32 0
339; CHECK-PROFITABLE-BY-DEFAULT-NEXT:    [[TMP3:%.*]] = bitcast i8* [[TMP2]] to <4 x i8>*
340; CHECK-PROFITABLE-BY-DEFAULT-NEXT:    store <4 x i8> <i8 1, i8 1, i8 1, i8 1>, <4 x i8>* [[TMP3]], align 1
341; CHECK-PROFITABLE-BY-DEFAULT-NEXT:    [[INDEX_NEXT]] = add i64 [[INDEX]], 4
342; CHECK-PROFITABLE-BY-DEFAULT-NEXT:    [[TMP4:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
343; CHECK-PROFITABLE-BY-DEFAULT-NEXT:    br i1 [[TMP4]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], [[LOOP0:!llvm.loop !.*]]
344; CHECK-PROFITABLE-BY-DEFAULT:       middle.block:
345; CHECK-PROFITABLE-BY-DEFAULT-NEXT:    [[CMP_N:%.*]] = icmp eq i64 [[N]], [[N_VEC]]
346; CHECK-PROFITABLE-BY-DEFAULT-NEXT:    br i1 [[CMP_N]], label [[FOR_END_LOOPEXIT:%.*]], label [[VEC_EPILOG_ITER_CHECK:%.*]]
347; CHECK-PROFITABLE-BY-DEFAULT:       vec.epilog.iter.check:
348; CHECK-PROFITABLE-BY-DEFAULT-NEXT:    [[N_VEC_REMAINING:%.*]] = sub i64 [[N]], [[N_VEC]]
349; CHECK-PROFITABLE-BY-DEFAULT-NEXT:    [[MIN_EPILOG_ITERS_CHECK:%.*]] = icmp ult i64 [[N_VEC_REMAINING]], 2
350; CHECK-PROFITABLE-BY-DEFAULT-NEXT:    br i1 [[MIN_EPILOG_ITERS_CHECK]], label [[VEC_EPILOG_SCALAR_PH]], label [[VEC_EPILOG_PH]]
351; CHECK-PROFITABLE-BY-DEFAULT:       vec.epilog.ph:
352; CHECK-PROFITABLE-BY-DEFAULT-NEXT:    [[VEC_EPILOG_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[VEC_EPILOG_ITER_CHECK]] ], [ 0, [[VECTOR_MAIN_LOOP_ITER_CHECK]] ]
353; CHECK-PROFITABLE-BY-DEFAULT-NEXT:    [[N_MOD_VF2:%.*]] = urem i64 [[N]], 2
354; CHECK-PROFITABLE-BY-DEFAULT-NEXT:    [[N_VEC3:%.*]] = sub i64 [[N]], [[N_MOD_VF2]]
355; CHECK-PROFITABLE-BY-DEFAULT-NEXT:    br label [[VEC_EPILOG_VECTOR_BODY:%.*]]
356; CHECK-PROFITABLE-BY-DEFAULT:       vec.epilog.vector.body:
357; CHECK-PROFITABLE-BY-DEFAULT-NEXT:    [[INDEX4:%.*]] = phi i64 [ [[VEC_EPILOG_RESUME_VAL]], [[VEC_EPILOG_PH]] ], [ [[INDEX_NEXT5:%.*]], [[VEC_EPILOG_VECTOR_BODY]] ]
358; CHECK-PROFITABLE-BY-DEFAULT-NEXT:    [[TMP5:%.*]] = add i64 [[INDEX4]], 0
359; CHECK-PROFITABLE-BY-DEFAULT-NEXT:    [[TMP6:%.*]] = getelementptr inbounds i8, i8* [[A]], i64 [[TMP5]]
360; CHECK-PROFITABLE-BY-DEFAULT-NEXT:    [[TMP7:%.*]] = getelementptr inbounds i8, i8* [[TMP6]], i32 0
361; CHECK-PROFITABLE-BY-DEFAULT-NEXT:    [[TMP8:%.*]] = bitcast i8* [[TMP7]] to <2 x i8>*
362; CHECK-PROFITABLE-BY-DEFAULT-NEXT:    store <2 x i8> <i8 1, i8 1>, <2 x i8>* [[TMP8]], align 1
363; CHECK-PROFITABLE-BY-DEFAULT-NEXT:    [[INDEX_NEXT5]] = add i64 [[INDEX4]], 2
364; CHECK-PROFITABLE-BY-DEFAULT-NEXT:    [[TMP9:%.*]] = icmp eq i64 [[INDEX_NEXT5]], [[N_VEC3]]
365; CHECK-PROFITABLE-BY-DEFAULT-NEXT:    br i1 [[TMP9]], label [[VEC_EPILOG_MIDDLE_BLOCK:%.*]], label [[VEC_EPILOG_VECTOR_BODY]], [[LOOP2:!llvm.loop !.*]]
366; CHECK-PROFITABLE-BY-DEFAULT:       vec.epilog.middle.block:
367; CHECK-PROFITABLE-BY-DEFAULT-NEXT:    [[CMP_N6:%.*]] = icmp eq i64 [[N]], [[N_VEC3]]
368; CHECK-PROFITABLE-BY-DEFAULT-NEXT:    br i1 [[CMP_N6]], label [[FOR_END_LOOPEXIT_LOOPEXIT:%.*]], label [[VEC_EPILOG_SCALAR_PH]]
369; CHECK-PROFITABLE-BY-DEFAULT:       vec.epilog.scalar.ph:
370; CHECK-PROFITABLE-BY-DEFAULT-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC3]], [[VEC_EPILOG_MIDDLE_BLOCK]] ], [ [[N_VEC]], [[VEC_EPILOG_ITER_CHECK]] ], [ 0, [[ITER_CHECK:%.*]] ]
371; CHECK-PROFITABLE-BY-DEFAULT-NEXT:    br label [[FOR_BODY:%.*]]
372; CHECK-PROFITABLE-BY-DEFAULT:       for.body:
373; CHECK-PROFITABLE-BY-DEFAULT-NEXT:    [[IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[VEC_EPILOG_SCALAR_PH]] ], [ [[IV_NEXT:%.*]], [[FOR_BODY]] ]
374; CHECK-PROFITABLE-BY-DEFAULT-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i8, i8* [[A]], i64 [[IV]]
375; CHECK-PROFITABLE-BY-DEFAULT-NEXT:    store i8 1, i8* [[ARRAYIDX]], align 1
376; CHECK-PROFITABLE-BY-DEFAULT-NEXT:    [[IV_NEXT]] = add nuw nsw i64 [[IV]], 1
377; CHECK-PROFITABLE-BY-DEFAULT-NEXT:    [[EXITCOND:%.*]] = icmp ne i64 [[IV_NEXT]], [[N]]
378; CHECK-PROFITABLE-BY-DEFAULT-NEXT:    br i1 [[EXITCOND]], label [[FOR_BODY]], label [[FOR_END_LOOPEXIT_LOOPEXIT]], [[LOOP4:!llvm.loop !.*]]
379; CHECK-PROFITABLE-BY-DEFAULT:       for.end.loopexit.loopexit:
380; CHECK-PROFITABLE-BY-DEFAULT-NEXT:    br label [[FOR_END_LOOPEXIT]]
381; CHECK-PROFITABLE-BY-DEFAULT:       for.end.loopexit:
382; CHECK-PROFITABLE-BY-DEFAULT-NEXT:    br label [[FOR_END:%.*]]
383; CHECK-PROFITABLE-BY-DEFAULT:       for.end:
384; CHECK-PROFITABLE-BY-DEFAULT-NEXT:    ret void
385;
386entry:
387  br label %for.body
388
389for.body:                                         ; preds = %for.body.preheader, %for.body
390  %iv = phi i64 [ 0, %entry ], [ %iv.next, %for.body ]
391  %arrayidx = getelementptr inbounds i8, i8* %A, i64 %iv
392  store i8 1, i8* %arrayidx, align 1
393  %iv.next = add nuw nsw i64 %iv, 1
394  %exitcond = icmp ne i64 %iv.next, %n
395  br i1 %exitcond, label %for.body, label %for.end.loopexit
396
397for.end.loopexit:                                 ; preds = %for.body
398  br label %for.end
399
400for.end:                                          ; preds = %for.end.loopexit, %entry
401  ret void
402}
403