1; RUN: llc -march=arm64 -aarch64-neon-syntax=apple -o - %s| FileCheck %s
2
3define <16 x i8> @foov16i8(<8 x i16> %a0, <8 x i16> %b0) nounwind readnone ssp {
4; CHECK-LABEL: foov16i8:
5  %vshrn_low_shift = lshr <8 x i16> %a0, <i16 5, i16 5, i16 5, i16 5, i16 5, i16 5, i16 5, i16 5>
6  %vshrn_low = trunc <8 x i16> %vshrn_low_shift to <8 x i8>
7  %vshrn_high_shift = lshr <8 x i16> %b0, <i16 5, i16 5, i16 5, i16 5, i16 5, i16 5, i16 5, i16 5>
8  %vshrn_high = trunc <8 x i16> %vshrn_high_shift to <8 x i8>
9; CHECK: shrn.8b v0, v0, #5
10; CHECK-NEXT: shrn2.16b v0, v1, #5
11; CHECK-NEXT: ret
12  %1 = bitcast <8 x i8> %vshrn_low to <1 x i64>
13  %2 = bitcast <8 x i8> %vshrn_high to <1 x i64>
14  %shuffle.i = shufflevector <1 x i64> %1, <1 x i64> %2, <2 x i32> <i32 0, i32 1>
15  %3 = bitcast <2 x i64> %shuffle.i to <16 x i8>
16  ret <16 x i8> %3
17}
18
19define <8 x i16> @foov8i16(<4 x i32> %a0, <4 x i32> %b0) nounwind readnone ssp {
20; CHECK-LABEL: foov8i16:
21  %vshrn_low_shift = lshr <4 x i32> %a0, <i32 5, i32 5, i32 5, i32 5>
22  %vshrn_low = trunc <4 x i32> %vshrn_low_shift to <4 x i16>
23  %vshrn_high_shift = lshr <4 x i32> %b0, <i32 5, i32 5, i32 5, i32 5>
24  %vshrn_high = trunc <4 x i32> %vshrn_high_shift to <4 x i16>
25; CHECK: shrn.4h v0, v0, #5
26; CHECK-NEXT: shrn2.8h v0, v1, #5
27; CHECK-NEXT: ret
28  %1 = bitcast <4 x i16> %vshrn_low to <1 x i64>
29  %2 = bitcast <4 x i16> %vshrn_high to <1 x i64>
30  %shuffle.i = shufflevector <1 x i64> %1, <1 x i64> %2, <2 x i32> <i32 0, i32 1>
31  %3 = bitcast <2 x i64> %shuffle.i to <8 x i16>
32  ret <8 x i16> %3
33}
34
35define <4 x i32> @foov4i32(<2 x i64> %a0, <2 x i64> %b0) nounwind readnone ssp {
36; CHECK-LABEL: foov4i32:
37  %vshrn_low_shift = lshr <2 x i64> %a0, <i64 5, i64 5>
38  %vshrn_low = trunc <2 x i64> %vshrn_low_shift to <2 x i32>
39  %vshrn_high_shift = lshr <2 x i64> %b0, <i64 5, i64 5>
40  %vshrn_high = trunc <2 x i64> %vshrn_high_shift to <2 x i32>
41; CHECK: shrn.2s v0, v0, #5
42; CHECK-NEXT: shrn2.4s v0, v1, #5
43; CHECK-NEXT: ret
44  %1 = bitcast <2 x i32> %vshrn_low to <1 x i64>
45  %2 = bitcast <2 x i32> %vshrn_high to <1 x i64>
46  %shuffle.i = shufflevector <1 x i64> %1, <1 x i64> %2, <2 x i32> <i32 0, i32 1>
47  %3 = bitcast <2 x i64> %shuffle.i to <4 x i32>
48  ret <4 x i32> %3
49}
50
51define <8 x i16> @bar(<4 x i32> %a0, <4 x i32> %a1, <4 x i32> %b0, <4 x i32> %b1) nounwind readnone ssp {
52; CHECK-LABEL: bar:
53  %vaddhn2.i = tail call <4 x i16> @llvm.aarch64.neon.addhn.v4i16(<4 x i32> %a0, <4 x i32> %a1) nounwind
54  %vaddhn2.i10 = tail call <4 x i16> @llvm.aarch64.neon.addhn.v4i16(<4 x i32> %b0, <4 x i32> %b1) nounwind
55; CHECK: addhn.4h	v0, v0, v1
56; CHECK-NEXT: addhn2.8h	v0, v2, v3
57; CHECK-NEXT: ret
58  %1 = bitcast <4 x i16> %vaddhn2.i to <1 x i64>
59  %2 = bitcast <4 x i16> %vaddhn2.i10 to <1 x i64>
60  %shuffle.i = shufflevector <1 x i64> %1, <1 x i64> %2, <2 x i32> <i32 0, i32 1>
61  %3 = bitcast <2 x i64> %shuffle.i to <8 x i16>
62  ret <8 x i16> %3
63}
64
65define <8 x i16> @baz(<4 x i32> %a0, <4 x i32> %a1, <4 x i32> %b0, <4 x i32> %b1) nounwind readnone ssp {
66; CHECK-LABEL: baz:
67  %vaddhn2.i = tail call <4 x i16> @llvm.aarch64.neon.addhn.v4i16(<4 x i32> %a0, <4 x i32> %a1) nounwind
68  %vshrn_high_shift = ashr <4 x i32> %b0, <i32 5, i32 5, i32 5, i32 5>
69  %vshrn_high = trunc <4 x i32> %vshrn_high_shift to <4 x i16>
70; CHECK: addhn.4h	v0, v0, v1
71; CHECK-NEXT: shrn2.8h	v0, v2, #5
72; CHECK-NEXT: ret
73  %1 = bitcast <4 x i16> %vaddhn2.i to <1 x i64>
74  %2 = bitcast <4 x i16> %vshrn_high to <1 x i64>
75  %shuffle.i = shufflevector <1 x i64> %1, <1 x i64> %2, <2 x i32> <i32 0, i32 1>
76  %3 = bitcast <2 x i64> %shuffle.i to <8 x i16>
77  ret <8 x i16> %3
78}
79
80define <8 x i16> @raddhn(<4 x i32> %a0, <4 x i32> %a1, <4 x i32> %b0, <4 x i32> %b1) nounwind readnone ssp {
81; CHECK-LABEL: raddhn:
82entry:
83; CHECK: 	raddhn.4h	v0, v0, v1
84; CHECK-NEXT: 	raddhn2.8h	v0, v2, v3
85; CHECK-NEXT: 	ret
86  %vraddhn2.i = tail call <4 x i16> @llvm.aarch64.neon.raddhn.v4i16(<4 x i32> %a0, <4 x i32> %a1) nounwind
87  %vraddhn2.i10 = tail call <4 x i16> @llvm.aarch64.neon.raddhn.v4i16(<4 x i32> %b0, <4 x i32> %b1) nounwind
88  %0 = bitcast <4 x i16> %vraddhn2.i to <1 x i64>
89  %1 = bitcast <4 x i16> %vraddhn2.i10 to <1 x i64>
90  %shuffle.i = shufflevector <1 x i64> %0, <1 x i64> %1, <2 x i32> <i32 0, i32 1>
91  %2 = bitcast <2 x i64> %shuffle.i to <8 x i16>
92  ret <8 x i16> %2
93}
94
95define <8 x i16> @vrshrn(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> %b0, <8 x i16> %b1) nounwind readnone ssp {
96; CHECK-LABEL: vrshrn:
97; CHECK: rshrn.8b	v0, v0, #5
98; CHECK-NEXT: rshrn2.16b	v0, v2, #6
99; CHECK-NEXT: ret
100  %vrshrn_n1 = tail call <8 x i8> @llvm.aarch64.neon.rshrn.v8i8(<8 x i16> %a0, i32 5)
101  %vrshrn_n4 = tail call <8 x i8> @llvm.aarch64.neon.rshrn.v8i8(<8 x i16> %b0, i32 6)
102  %1 = bitcast <8 x i8> %vrshrn_n1 to <1 x i64>
103  %2 = bitcast <8 x i8> %vrshrn_n4 to <1 x i64>
104  %shuffle.i = shufflevector <1 x i64> %1, <1 x i64> %2, <2 x i32> <i32 0, i32 1>
105  %3 = bitcast <2 x i64> %shuffle.i to <8 x i16>
106  ret <8 x i16> %3
107}
108
109define <8 x i16> @vrsubhn(<8 x i16> %a0, <8 x i16> %a1, <8 x i16> %b0, <8 x i16> %b1) nounwind readnone ssp {
110; CHECK-LABEL: vrsubhn:
111; CHECK: rsubhn.8b	v0, v0, v1
112; CHECK: rsubhn2.16b	v0, v2, v3
113; CHECK-NEXT: 	ret
114  %vrsubhn2.i = tail call <8 x i8> @llvm.aarch64.neon.rsubhn.v8i8(<8 x i16> %a0, <8 x i16> %a1) nounwind
115  %vrsubhn2.i10 = tail call <8 x i8> @llvm.aarch64.neon.rsubhn.v8i8(<8 x i16> %b0, <8 x i16> %b1) nounwind
116  %1 = bitcast <8 x i8> %vrsubhn2.i to <1 x i64>
117  %2 = bitcast <8 x i8> %vrsubhn2.i10 to <1 x i64>
118  %shuffle.i = shufflevector <1 x i64> %1, <1 x i64> %2, <2 x i32> <i32 0, i32 1>
119  %3 = bitcast <2 x i64> %shuffle.i to <8 x i16>
120  ret <8 x i16> %3
121}
122
123define <8 x i16> @noOpt1(<2 x i32> %a0, <2 x i32> %a1, <4 x i32> %b0, <4 x i32> %b1) nounwind readnone ssp {
124; CHECK-LABEL: noOpt1:
125  %vqsub2.i = tail call <2 x i32> @llvm.aarch64.neon.sqsub.v2i32(<2 x i32> %a0, <2 x i32> %a1) nounwind
126  %vaddhn2.i = tail call <4 x i16> @llvm.aarch64.neon.addhn.v4i16(<4 x i32> %b0, <4 x i32> %b1) nounwind
127; CHECK:	sqsub.2s	v0, v0, v1
128; CHECK-NEXT:	addhn2.8h	v0, v2, v3
129  %1 = bitcast <2 x i32> %vqsub2.i to <1 x i64>
130  %2 = bitcast <4 x i16> %vaddhn2.i to <1 x i64>
131  %shuffle.i = shufflevector <1 x i64> %1, <1 x i64> %2, <2 x i32> <i32 0, i32 1>
132  %3 = bitcast <2 x i64> %shuffle.i to <8 x i16>
133  ret <8 x i16> %3
134}
135
136declare <2 x i32> @llvm.aarch64.neon.sqsub.v2i32(<2 x i32>, <2 x i32>) nounwind readnone
137
138declare <8 x i8> @llvm.aarch64.neon.shrn.v8i8(<8 x i16>, i32) nounwind readnone
139declare <4 x i16> @llvm.aarch64.neon.shrn.v4i16(<4 x i32>, i32) nounwind readnone
140declare <2 x i32> @llvm.aarch64.neon.shrn.v2i32(<2 x i64>, i32) nounwind readnone
141declare <4 x i16> @llvm.aarch64.neon.addhn.v4i16(<4 x i32>, <4 x i32>) nounwind readnone
142declare <4 x i16> @llvm.aarch64.neon.raddhn.v4i16(<4 x i32>, <4 x i32>) nounwind readnone
143declare <8 x i8> @llvm.aarch64.neon.rshrn.v8i8(<8 x i16>, i32) nounwind readnone
144declare <8 x i8> @llvm.aarch64.neon.rsubhn.v8i8(<8 x i16>, <8 x i16>) nounwind readnone
145
146