1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
2; RUN: opt -mtriple=armv8.1m.main -mattr=+mve -S -mve-tail-predication -tail-predication=enabled %s -o - | FileCheck %s
3
4define void @mat_vec_sext_i16(i16** nocapture readonly %A, i16* nocapture readonly %B, i32* noalias nocapture %C, i32 %N) {
5; CHECK-LABEL: @mat_vec_sext_i16(
6; CHECK-NEXT:  entry:
7; CHECK-NEXT:    [[CMP24:%.*]] = icmp eq i32 [[N:%.*]], 0
8; CHECK-NEXT:    br i1 [[CMP24]], label [[FOR_COND_CLEANUP:%.*]], label [[FOR_COND1_PREHEADER_US_PREHEADER:%.*]]
9; CHECK:       for.cond1.preheader.us.preheader:
10; CHECK-NEXT:    [[N_RND_UP:%.*]] = add i32 [[N]], 3
11; CHECK-NEXT:    [[N_VEC:%.*]] = and i32 [[N_RND_UP]], -4
12; CHECK-NEXT:    [[TT:%.*]] = add i32 [[N_VEC]], -4
13; CHECK-NEXT:    [[TT1:%.*]] = lshr i32 [[TT]], 2
14; CHECK-NEXT:    [[TT2:%.*]] = add nuw nsw i32 [[TT1]], 1
15; CHECK-NEXT:    br label [[FOR_COND1_PREHEADER_US:%.*]]
16; CHECK:       for.cond1.preheader.us:
17; CHECK-NEXT:    [[I_025_US:%.*]] = phi i32 [ [[INC10_US:%.*]], [[MIDDLE_BLOCK:%.*]] ], [ 0, [[FOR_COND1_PREHEADER_US_PREHEADER]] ]
18; CHECK-NEXT:    [[ARRAYIDX_US:%.*]] = getelementptr inbounds i16*, i16** [[A:%.*]], i32 [[I_025_US]]
19; CHECK-NEXT:    [[TT3:%.*]] = load i16*, i16** [[ARRAYIDX_US]], align 4
20; CHECK-NEXT:    [[ARRAYIDX8_US:%.*]] = getelementptr inbounds i32, i32* [[C:%.*]], i32 [[I_025_US]]
21; CHECK-NEXT:    [[ARRAYIDX8_PROMOTED_US:%.*]] = load i32, i32* [[ARRAYIDX8_US]], align 4
22; CHECK-NEXT:    [[TT4:%.*]] = insertelement <4 x i32> <i32 undef, i32 0, i32 0, i32 0>, i32 [[ARRAYIDX8_PROMOTED_US]], i32 0
23; CHECK-NEXT:    [[START:%.*]] = call i32 @llvm.start.loop.iterations.i32(i32 [[TT2]])
24; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
25; CHECK:       vector.body:
26; CHECK-NEXT:    [[INDEX:%.*]] = phi i32 [ 0, [[FOR_COND1_PREHEADER_US]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
27; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi <4 x i32> [ [[TT4]], [[FOR_COND1_PREHEADER_US]] ], [ [[TT14:%.*]], [[VECTOR_BODY]] ]
28; CHECK-NEXT:    [[TT5:%.*]] = phi i32 [ [[START]], [[FOR_COND1_PREHEADER_US]] ], [ [[TT15:%.*]], [[VECTOR_BODY]] ]
29; CHECK-NEXT:    [[TMP0:%.*]] = phi i32 [ [[N]], [[FOR_COND1_PREHEADER_US]] ], [ [[TMP2:%.*]], [[VECTOR_BODY]] ]
30; CHECK-NEXT:    [[TT6:%.*]] = getelementptr inbounds i16, i16* [[TT3]], i32 [[INDEX]]
31; CHECK-NEXT:    [[TMP1:%.*]] = call <4 x i1> @llvm.arm.mve.vctp32(i32 [[TMP0]])
32; CHECK-NEXT:    [[TMP2]] = sub i32 [[TMP0]], 4
33; CHECK-NEXT:    [[TT8:%.*]] = bitcast i16* [[TT6]] to <4 x i16>*
34; CHECK-NEXT:    [[WIDE_MASKED_LOAD:%.*]] = call <4 x i16> @llvm.masked.load.v4i16.p0v4i16(<4 x i16>* [[TT8]], i32 2, <4 x i1> [[TMP1]], <4 x i16> undef)
35; CHECK-NEXT:    [[TT9:%.*]] = sext <4 x i16> [[WIDE_MASKED_LOAD]] to <4 x i32>
36; CHECK-NEXT:    [[TT10:%.*]] = getelementptr inbounds i16, i16* [[B:%.*]], i32 [[INDEX]]
37; CHECK-NEXT:    [[TT11:%.*]] = bitcast i16* [[TT10]] to <4 x i16>*
38; CHECK-NEXT:    [[WIDE_MASKED_LOAD30:%.*]] = call <4 x i16> @llvm.masked.load.v4i16.p0v4i16(<4 x i16>* [[TT11]], i32 2, <4 x i1> [[TMP1]], <4 x i16> undef)
39; CHECK-NEXT:    [[TT12:%.*]] = sext <4 x i16> [[WIDE_MASKED_LOAD30]] to <4 x i32>
40; CHECK-NEXT:    [[TT13:%.*]] = mul nsw <4 x i32> [[TT12]], [[TT9]]
41; CHECK-NEXT:    [[TT14]] = add nsw <4 x i32> [[TT13]], [[VEC_PHI]]
42; CHECK-NEXT:    [[INDEX_NEXT]] = add i32 [[INDEX]], 4
43; CHECK-NEXT:    [[TT15]] = call i32 @llvm.loop.decrement.reg.i32(i32 [[TT5]], i32 1)
44; CHECK-NEXT:    [[TT16:%.*]] = icmp ne i32 [[TT15]], 0
45; CHECK-NEXT:    br i1 [[TT16]], label [[VECTOR_BODY]], label [[MIDDLE_BLOCK]]
46; CHECK:       middle.block:
47; CHECK-NEXT:    [[TT17:%.*]] = select <4 x i1> [[TMP1]], <4 x i32> [[TT14]], <4 x i32> [[VEC_PHI]]
48; CHECK-NEXT:    [[TT18:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[TT17]])
49; CHECK-NEXT:    store i32 [[TT18]], i32* [[ARRAYIDX8_US]], align 4
50; CHECK-NEXT:    [[INC10_US]] = add nuw i32 [[I_025_US]], 1
51; CHECK-NEXT:    [[EXITCOND27:%.*]] = icmp eq i32 [[INC10_US]], [[N]]
52; CHECK-NEXT:    br i1 [[EXITCOND27]], label [[FOR_COND_CLEANUP]], label [[FOR_COND1_PREHEADER_US]]
53; CHECK:       for.cond.cleanup:
54; CHECK-NEXT:    ret void
55;
56entry:
57  %cmp24 = icmp eq i32 %N, 0
58  br i1 %cmp24, label %for.cond.cleanup, label %for.cond1.preheader.us.preheader
59
60for.cond1.preheader.us.preheader:                 ; preds = %entry
61  %n.rnd.up = add i32 %N, 3
62  %n.vec = and i32 %n.rnd.up, -4
63  %tt = add i32 %n.vec, -4
64  %tt1 = lshr i32 %tt, 2
65  %tt2 = add nuw nsw i32 %tt1, 1
66  br label %for.cond1.preheader.us
67
68for.cond1.preheader.us:                           ; preds = %middle.block, %for.cond1.preheader.us.preheader
69  %i.025.us = phi i32 [ %inc10.us, %middle.block ], [ 0, %for.cond1.preheader.us.preheader ]
70  %arrayidx.us = getelementptr inbounds i16*, i16** %A, i32 %i.025.us
71  %tt3 = load i16*, i16** %arrayidx.us, align 4
72  %arrayidx8.us = getelementptr inbounds i32, i32* %C, i32 %i.025.us
73  %arrayidx8.promoted.us = load i32, i32* %arrayidx8.us, align 4
74  %tt4 = insertelement <4 x i32> <i32 undef, i32 0, i32 0, i32 0>, i32 %arrayidx8.promoted.us, i32 0
75  %start = call i32 @llvm.start.loop.iterations.i32(i32 %tt2)
76  br label %vector.body
77
78vector.body:                                      ; preds = %vector.body, %for.cond1.preheader.us
79  %index = phi i32 [ 0, %for.cond1.preheader.us ], [ %index.next, %vector.body ]
80  %vec.phi = phi <4 x i32> [ %tt4, %for.cond1.preheader.us ], [ %tt14, %vector.body ]
81  %tt5 = phi i32 [ %start, %for.cond1.preheader.us ], [ %tt15, %vector.body ]
82  %tt6 = getelementptr inbounds i16, i16* %tt3, i32 %index
83  %tt7 = call <4 x i1> @llvm.get.active.lane.mask.v4i1.i32(i32 %index, i32 %N)
84  %tt8 = bitcast i16* %tt6 to <4 x i16>*
85  %wide.masked.load = call <4 x i16> @llvm.masked.load.v4i16.p0v4i16(<4 x i16>* %tt8, i32 2, <4 x i1> %tt7, <4 x i16> undef)
86  %tt9 = sext <4 x i16> %wide.masked.load to <4 x i32>
87  %tt10 = getelementptr inbounds i16, i16* %B, i32 %index
88  %tt11 = bitcast i16* %tt10 to <4 x i16>*
89  %wide.masked.load30 = call <4 x i16> @llvm.masked.load.v4i16.p0v4i16(<4 x i16>* %tt11, i32 2, <4 x i1> %tt7, <4 x i16> undef)
90  %tt12 = sext <4 x i16> %wide.masked.load30 to <4 x i32>
91  %tt13 = mul nsw <4 x i32> %tt12, %tt9
92  %tt14 = add nsw <4 x i32> %tt13, %vec.phi
93  %index.next = add i32 %index, 4
94  %tt15 = call i32 @llvm.loop.decrement.reg.i32(i32 %tt5, i32 1)
95  %tt16 = icmp ne i32 %tt15, 0
96  br i1 %tt16, label %vector.body, label %middle.block
97
98middle.block:                                     ; preds = %vector.body
99  %tt17 = select <4 x i1> %tt7, <4 x i32> %tt14, <4 x i32> %vec.phi
100  %tt18 = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> %tt17)
101  store i32 %tt18, i32* %arrayidx8.us, align 4
102  %inc10.us = add nuw i32 %i.025.us, 1
103  %exitcond27 = icmp eq i32 %inc10.us, %N
104  br i1 %exitcond27, label %for.cond.cleanup, label %for.cond1.preheader.us
105
106for.cond.cleanup:                                 ; preds = %middle.block, %entry
107  ret void
108}
109
110define void @mat_vec_i32(i32** nocapture readonly %A, i32* nocapture readonly %B, i32* noalias nocapture %C, i32 %N) {
111; CHECK-LABEL: @mat_vec_i32(
112; CHECK-NEXT:  entry:
113; CHECK-NEXT:    [[CMP23:%.*]] = icmp eq i32 [[N:%.*]], 0
114; CHECK-NEXT:    br i1 [[CMP23]], label [[FOR_COND_CLEANUP:%.*]], label [[FOR_COND1_PREHEADER_US_PREHEADER:%.*]]
115; CHECK:       for.cond1.preheader.us.preheader:
116; CHECK-NEXT:    [[N_RND_UP:%.*]] = add i32 [[N]], 3
117; CHECK-NEXT:    [[N_VEC:%.*]] = and i32 [[N_RND_UP]], -4
118; CHECK-NEXT:    [[TT:%.*]] = add i32 [[N_VEC]], -4
119; CHECK-NEXT:    [[TT1:%.*]] = lshr i32 [[TT]], 2
120; CHECK-NEXT:    [[TT2:%.*]] = add nuw nsw i32 [[TT1]], 1
121; CHECK-NEXT:    br label [[FOR_COND1_PREHEADER_US:%.*]]
122; CHECK:       for.cond1.preheader.us:
123; CHECK-NEXT:    [[I_024_US:%.*]] = phi i32 [ [[INC9_US:%.*]], [[MIDDLE_BLOCK:%.*]] ], [ 0, [[FOR_COND1_PREHEADER_US_PREHEADER]] ]
124; CHECK-NEXT:    [[ARRAYIDX_US:%.*]] = getelementptr inbounds i32*, i32** [[A:%.*]], i32 [[I_024_US]]
125; CHECK-NEXT:    [[TT3:%.*]] = load i32*, i32** [[ARRAYIDX_US]], align 4
126; CHECK-NEXT:    [[ARRAYIDX7_US:%.*]] = getelementptr inbounds i32, i32* [[C:%.*]], i32 [[I_024_US]]
127; CHECK-NEXT:    [[ARRAYIDX7_PROMOTED_US:%.*]] = load i32, i32* [[ARRAYIDX7_US]], align 4
128; CHECK-NEXT:    [[TT4:%.*]] = insertelement <4 x i32> <i32 undef, i32 0, i32 0, i32 0>, i32 [[ARRAYIDX7_PROMOTED_US]], i32 0
129; CHECK-NEXT:    [[START:%.*]] = call i32 @llvm.start.loop.iterations.i32(i32 [[TT2]])
130; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
131; CHECK:       vector.body:
132; CHECK-NEXT:    [[INDEX:%.*]] = phi i32 [ 0, [[FOR_COND1_PREHEADER_US]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
133; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi <4 x i32> [ [[TT4]], [[FOR_COND1_PREHEADER_US]] ], [ [[TT12:%.*]], [[VECTOR_BODY]] ]
134; CHECK-NEXT:    [[TT5:%.*]] = phi i32 [ [[START]], [[FOR_COND1_PREHEADER_US]] ], [ [[TT13:%.*]], [[VECTOR_BODY]] ]
135; CHECK-NEXT:    [[TMP0:%.*]] = phi i32 [ [[N]], [[FOR_COND1_PREHEADER_US]] ], [ [[TMP2:%.*]], [[VECTOR_BODY]] ]
136; CHECK-NEXT:    [[TT6:%.*]] = getelementptr inbounds i32, i32* [[TT3]], i32 [[INDEX]]
137; CHECK-NEXT:    [[TMP1:%.*]] = call <4 x i1> @llvm.arm.mve.vctp32(i32 [[TMP0]])
138; CHECK-NEXT:    [[TMP2]] = sub i32 [[TMP0]], 4
139; CHECK-NEXT:    [[TT8:%.*]] = bitcast i32* [[TT6]] to <4 x i32>*
140; CHECK-NEXT:    [[WIDE_MASKED_LOAD:%.*]] = call <4 x i32> @llvm.masked.load.v4i32.p0v4i32(<4 x i32>* [[TT8]], i32 4, <4 x i1> [[TMP1]], <4 x i32> undef)
141; CHECK-NEXT:    [[TT9:%.*]] = getelementptr inbounds i32, i32* [[B:%.*]], i32 [[INDEX]]
142; CHECK-NEXT:    [[TT10:%.*]] = bitcast i32* [[TT9]] to <4 x i32>*
143; CHECK-NEXT:    [[WIDE_MASKED_LOAD29:%.*]] = call <4 x i32> @llvm.masked.load.v4i32.p0v4i32(<4 x i32>* [[TT10]], i32 4, <4 x i1> [[TMP1]], <4 x i32> undef)
144; CHECK-NEXT:    [[TT11:%.*]] = mul nsw <4 x i32> [[WIDE_MASKED_LOAD29]], [[WIDE_MASKED_LOAD]]
145; CHECK-NEXT:    [[TT12]] = add nsw <4 x i32> [[VEC_PHI]], [[TT11]]
146; CHECK-NEXT:    [[INDEX_NEXT]] = add i32 [[INDEX]], 4
147; CHECK-NEXT:    [[TT13]] = call i32 @llvm.loop.decrement.reg.i32(i32 [[TT5]], i32 1)
148; CHECK-NEXT:    [[TT14:%.*]] = icmp ne i32 [[TT13]], 0
149; CHECK-NEXT:    br i1 [[TT14]], label [[VECTOR_BODY]], label [[MIDDLE_BLOCK]]
150; CHECK:       middle.block:
151; CHECK-NEXT:    [[TT15:%.*]] = select <4 x i1> [[TMP1]], <4 x i32> [[TT12]], <4 x i32> [[VEC_PHI]]
152; CHECK-NEXT:    [[TT16:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[TT15]])
153; CHECK-NEXT:    store i32 [[TT16]], i32* [[ARRAYIDX7_US]], align 4
154; CHECK-NEXT:    [[INC9_US]] = add nuw i32 [[I_024_US]], 1
155; CHECK-NEXT:    [[EXITCOND26:%.*]] = icmp eq i32 [[INC9_US]], [[N]]
156; CHECK-NEXT:    br i1 [[EXITCOND26]], label [[FOR_COND_CLEANUP]], label [[FOR_COND1_PREHEADER_US]]
157; CHECK:       for.cond.cleanup:
158; CHECK-NEXT:    ret void
159;
160entry:
161  %cmp23 = icmp eq i32 %N, 0
162  br i1 %cmp23, label %for.cond.cleanup, label %for.cond1.preheader.us.preheader
163
164for.cond1.preheader.us.preheader:                 ; preds = %entry
165  %n.rnd.up = add i32 %N, 3
166  %n.vec = and i32 %n.rnd.up, -4
167  %tt = add i32 %n.vec, -4
168  %tt1 = lshr i32 %tt, 2
169  %tt2 = add nuw nsw i32 %tt1, 1
170  br label %for.cond1.preheader.us
171
172for.cond1.preheader.us:                           ; preds = %middle.block, %for.cond1.preheader.us.preheader
173  %i.024.us = phi i32 [ %inc9.us, %middle.block ], [ 0, %for.cond1.preheader.us.preheader ]
174  %arrayidx.us = getelementptr inbounds i32*, i32** %A, i32 %i.024.us
175  %tt3 = load i32*, i32** %arrayidx.us, align 4
176  %arrayidx7.us = getelementptr inbounds i32, i32* %C, i32 %i.024.us
177  %arrayidx7.promoted.us = load i32, i32* %arrayidx7.us, align 4
178  %tt4 = insertelement <4 x i32> <i32 undef, i32 0, i32 0, i32 0>, i32 %arrayidx7.promoted.us, i32 0
179  %start = call i32 @llvm.start.loop.iterations.i32(i32 %tt2)
180  br label %vector.body
181
182vector.body:                                      ; preds = %vector.body, %for.cond1.preheader.us
183  %index = phi i32 [ 0, %for.cond1.preheader.us ], [ %index.next, %vector.body ]
184  %vec.phi = phi <4 x i32> [ %tt4, %for.cond1.preheader.us ], [ %tt12, %vector.body ]
185  %tt5 = phi i32 [ %start, %for.cond1.preheader.us ], [ %tt13, %vector.body ]
186  %tt6 = getelementptr inbounds i32, i32* %tt3, i32 %index
187  %tt7 = call <4 x i1> @llvm.get.active.lane.mask.v4i1.i32(i32 %index, i32 %N)
188  %tt8 = bitcast i32* %tt6 to <4 x i32>*
189  %wide.masked.load = call <4 x i32> @llvm.masked.load.v4i32.p0v4i32(<4 x i32>* %tt8, i32 4, <4 x i1> %tt7, <4 x i32> undef)
190  %tt9 = getelementptr inbounds i32, i32* %B, i32 %index
191  %tt10 = bitcast i32* %tt9 to <4 x i32>*
192  %wide.masked.load29 = call <4 x i32> @llvm.masked.load.v4i32.p0v4i32(<4 x i32>* %tt10, i32 4, <4 x i1> %tt7, <4 x i32> undef)
193  %tt11 = mul nsw <4 x i32> %wide.masked.load29, %wide.masked.load
194  %tt12 = add nsw <4 x i32> %vec.phi, %tt11
195  %index.next = add i32 %index, 4
196  %tt13 = call i32 @llvm.loop.decrement.reg.i32(i32 %tt5, i32 1)
197  %tt14 = icmp ne i32 %tt13, 0
198  br i1 %tt14, label %vector.body, label %middle.block
199
200middle.block:                                     ; preds = %vector.body
201  %tt15 = select <4 x i1> %tt7, <4 x i32> %tt12, <4 x i32> %vec.phi
202  %tt16 = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> %tt15)
203  store i32 %tt16, i32* %arrayidx7.us, align 4
204  %inc9.us = add nuw i32 %i.024.us, 1
205  %exitcond26 = icmp eq i32 %inc9.us, %N
206  br i1 %exitcond26, label %for.cond.cleanup, label %for.cond1.preheader.us
207
208for.cond.cleanup:                                 ; preds = %middle.block, %entry
209  ret void
210}
211
212
213; Function Attrs: argmemonly nounwind readonly willreturn
214declare <4 x i32> @llvm.masked.load.v4i32.p0v4i32(<4 x i32>*, i32 immarg, <4 x i1>, <4 x i32>) #0
215
216; Function Attrs: argmemonly nounwind readonly willreturn
217declare <4 x i16> @llvm.masked.load.v4i16.p0v4i16(<4 x i16>*, i32 immarg, <4 x i1>, <4 x i16>) #0
218
219; Function Attrs: nounwind readnone willreturn
220declare i32 @llvm.vector.reduce.add.v4i32(<4 x i32>) #1
221
222; Function Attrs: noduplicate nounwind
223declare i32 @llvm.start.loop.iterations.i32(i32) #2
224
225; Function Attrs: noduplicate nounwind
226declare i32 @llvm.loop.decrement.reg.i32(i32, i32) #2
227
228declare <4 x i1> @llvm.get.active.lane.mask.v4i1.i32(i32, i32)
229
230attributes #0 = { argmemonly nounwind readonly willreturn }
231attributes #1 = { nounwind readnone willreturn }
232attributes #2 = { noduplicate nounwind }
233