| blob | 1f1bfe6906482e4c1c0b4ded20955902b9118f75 |
1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2 ; RUN: llc -mtriple=arm64-apple-ios -mattr=+sve -o - %s | FileCheck %s
3 ; RUN: llc -mtriple=aarch64_be-unknown-linux -mattr=+sve -o - %s | FileCheck --check-prefix=CHECK-BE %s
4 ; RUN: llc -mtriple=arm64-apple-ios -mattr=+global-isel -mattr=+sve -o - %s | FileCheck %s
5 ; RUN: llc -mtriple=aarch64_be-unknown-linux -mattr=+global-isel -mattr=+sve -o - %s | FileCheck --check-prefix=CHECK-BE %s
7 define void @zext_of_concat(ptr %a, ptr %b, ptr %c, ptr %d) nounwind {
8 ; CHECK-LABEL: zext_of_concat:
9 ; CHECK: ; %bb.0:
10 ; CHECK-NEXT: ldr d0, [x0]
11 ; CHECK-NEXT: ldr d1, [x1]
12 ; CHECK-NEXT: add.2s v0, v0, v1
13 ; CHECK-NEXT: ldr q1, [x2]
14 ; CHECK-NEXT: ushll.2d v0, v0, #0
15 ; CHECK-NEXT: add.4s v0, v0, v1
16 ; CHECK-NEXT: str q0, [x2]
17 ; CHECK-NEXT: ret
18 ;
19 ; CHECK-BE-LABEL: zext_of_concat:
20 ; CHECK-BE: // %bb.0:
21 ; CHECK-BE-NEXT: ld1 { v0.2s }, [x0]
22 ; CHECK-BE-NEXT: ld1 { v1.2s }, [x1]
23 ; CHECK-BE-NEXT: add v0.2s, v0.2s, v1.2s
24 ; CHECK-BE-NEXT: movi v1.2d, #0000000000000000
25 ; CHECK-BE-NEXT: zip1 v0.4s, v0.4s, v0.4s
26 ; CHECK-BE-NEXT: trn2 v0.4s, v0.4s, v1.4s
27 ; CHECK-BE-NEXT: ld1 { v1.4s }, [x2]
28 ; CHECK-BE-NEXT: add v0.4s, v0.4s, v1.4s
29 ; CHECK-BE-NEXT: st1 { v0.4s }, [x2]
30 ; CHECK-BE-NEXT: ret
31 %i0.a = load <2 x i32>, ptr %a
32 %i0.b = load <2 x i32>, ptr %b
33 %i0 = add <2 x i32> %i0.a, %i0.b
34 %i1 = shufflevector <2 x i32> %i0, <2 x i32> poison, <4 x i32> <i32 0, i32 1, i32 undef, i32 undef>
35 %i2 = shufflevector <4 x i32> %i1, <4 x i32> zeroinitializer, <4 x i32> <i32 0, i32 5, i32 1, i32 7>
36 %i3 = load <4 x i32>, ptr %c
37 %i4 = add <4 x i32> %i2, %i3
38 store <4 x i32> %i4, ptr %c
39 ret void
40 }
42 define void @zext_of_concat_extrause(ptr %a, ptr %b, ptr %c, ptr %d, ptr %e) nounwind {
43 ; CHECK-LABEL: zext_of_concat_extrause:
44 ; CHECK: ; %bb.0:
45 ; CHECK-NEXT: ldr d0, [x1]
46 ; CHECK-NEXT: ldr d1, [x0]
47 ; CHECK-NEXT: add.2s v0, v1, v0
48 ; CHECK-NEXT: movi.2d v1, #0000000000000000
49 ; CHECK-NEXT: mov.d v0[1], v0[0]
50 ; CHECK-NEXT: zip1.4s v1, v0, v1
51 ; CHECK-NEXT: str q0, [x4]
52 ; CHECK-NEXT: ldr q0, [x2]
53 ; CHECK-NEXT: add.4s v0, v1, v0
54 ; CHECK-NEXT: str q0, [x2]
55 ; CHECK-NEXT: ret
56 ;
57 ; CHECK-BE-LABEL: zext_of_concat_extrause:
58 ; CHECK-BE: // %bb.0:
59 ; CHECK-BE-NEXT: ld1 { v0.2s }, [x1]
60 ; CHECK-BE-NEXT: ld1 { v1.2s }, [x0]
61 ; CHECK-BE-NEXT: movi v2.2d, #0000000000000000
62 ; CHECK-BE-NEXT: add v0.2s, v1.2s, v0.2s
63 ; CHECK-BE-NEXT: mov v0.d[1], v0.d[0]
64 ; CHECK-BE-NEXT: zip1 v1.4s, v0.4s, v0.4s
65 ; CHECK-BE-NEXT: st1 { v0.4s }, [x4]
66 ; CHECK-BE-NEXT: trn2 v0.4s, v1.4s, v2.4s
67 ; CHECK-BE-NEXT: ld1 { v1.4s }, [x2]
68 ; CHECK-BE-NEXT: add v0.4s, v0.4s, v1.4s
69 ; CHECK-BE-NEXT: st1 { v0.4s }, [x2]
70 ; CHECK-BE-NEXT: ret
71 %i0.a = load <2 x i32>, ptr %a
72 %i0.b = load <2 x i32>, ptr %b
73 %i0 = add <2 x i32> %i0.a, %i0.b
74 %i1 = shufflevector <2 x i32> %i0, <2 x i32> %i0, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
75 store <4 x i32> %i1, ptr %e
76 %i2 = shufflevector <4 x i32> %i1, <4 x i32> zeroinitializer, <4 x i32> <i32 0, i32 5, i32 1, i32 7>
77 %i3 = load <4 x i32>, ptr %c
78 %i4 = add <4 x i32> %i2, %i3
79 store <4 x i32> %i4, ptr %c
80 ret void
81 }
83 define void @aext_of_concat(ptr %a, ptr %b, ptr %c, ptr %d) nounwind {
84 ; CHECK-LABEL: aext_of_concat:
85 ; CHECK: ; %bb.0:
86 ; CHECK-NEXT: ldr d0, [x0]
87 ; CHECK-NEXT: ldr d1, [x1]
88 ; CHECK-NEXT: add.2s v0, v0, v1
89 ; CHECK-NEXT: ldr q1, [x2]
90 ; CHECK-NEXT: ushll.2d v0, v0, #0
91 ; CHECK-NEXT: add.4s v0, v0, v1
92 ; CHECK-NEXT: str q0, [x2]
93 ; CHECK-NEXT: ret
94 ;
95 ; CHECK-BE-LABEL: aext_of_concat:
96 ; CHECK-BE: // %bb.0:
97 ; CHECK-BE-NEXT: ld1 { v0.2s }, [x0]
98 ; CHECK-BE-NEXT: ld1 { v1.2s }, [x1]
99 ; CHECK-BE-NEXT: add v0.2s, v0.2s, v1.2s
100 ; CHECK-BE-NEXT: ld1 { v1.4s }, [x2]
101 ; CHECK-BE-NEXT: zip1 v0.4s, v0.4s, v0.4s
102 ; CHECK-BE-NEXT: add v0.4s, v0.4s, v1.4s
103 ; CHECK-BE-NEXT: st1 { v0.4s }, [x2]
104 ; CHECK-BE-NEXT: ret
105 %i0.a = load <2 x i32>, ptr %a
106 %i0.b = load <2 x i32>, ptr %b
107 %i0 = add <2 x i32> %i0.a, %i0.b
108 %i1 = shufflevector <2 x i32> %i0, <2 x i32> poison, <4 x i32> <i32 0, i32 1, i32 undef, i32 undef>
109 %i2 = shufflevector <4 x i32> %i1, <4 x i32> poison, <4 x i32> <i32 0, i32 undef, i32 1, i32 undef>
110 %i3 = load <4 x i32>, ptr %c
111 %i4 = add <4 x i32> %i2, %i3
112 store <4 x i32> %i4, ptr %c
113 ret void
114 }
116 define void @aext_of_concat_extrause(ptr %a, ptr %b, ptr %c, ptr %d, ptr %e) nounwind {
117 ; CHECK-LABEL: aext_of_concat_extrause:
118 ; CHECK: ; %bb.0:
119 ; CHECK-NEXT: ldr d0, [x1]
120 ; CHECK-NEXT: ldr d1, [x0]
121 ; CHECK-NEXT: add.2s v0, v1, v0
122 ; CHECK-NEXT: mov.16b v1, v0
123 ; CHECK-NEXT: mov.d v1[1], v0[0]
124 ; CHECK-NEXT: zip1.4s v0, v0, v0
125 ; CHECK-NEXT: str q1, [x4]
126 ; CHECK-NEXT: ldr q1, [x2]
127 ; CHECK-NEXT: add.4s v0, v0, v1
128 ; CHECK-NEXT: str q0, [x2]
129 ; CHECK-NEXT: ret
130 ;
131 ; CHECK-BE-LABEL: aext_of_concat_extrause:
132 ; CHECK-BE: // %bb.0:
133 ; CHECK-BE-NEXT: ld1 { v0.2s }, [x1]
134 ; CHECK-BE-NEXT: ld1 { v1.2s }, [x0]
135 ; CHECK-BE-NEXT: add v0.2s, v1.2s, v0.2s
136 ; CHECK-BE-NEXT: mov v1.16b, v0.16b
137 ; CHECK-BE-NEXT: mov v1.d[1], v0.d[0]
138 ; CHECK-BE-NEXT: zip1 v0.4s, v0.4s, v0.4s
139 ; CHECK-BE-NEXT: st1 { v1.4s }, [x4]
140 ; CHECK-BE-NEXT: ld1 { v1.4s }, [x2]
141 ; CHECK-BE-NEXT: add v0.4s, v0.4s, v1.4s
142 ; CHECK-BE-NEXT: st1 { v0.4s }, [x2]
143 ; CHECK-BE-NEXT: ret
144 %i0.a = load <2 x i32>, ptr %a
145 %i0.b = load <2 x i32>, ptr %b
146 %i0 = add <2 x i32> %i0.a, %i0.b
147 %i1 = shufflevector <2 x i32> %i0, <2 x i32> %i0, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
148 store <4 x i32> %i1, ptr %e
149 %i2 = shufflevector <4 x i32> %i1, <4 x i32> zeroinitializer, <4 x i32> <i32 0, i32 undef, i32 1, i32 undef>
150 %i3 = load <4 x i32>, ptr %c
151 %i4 = add <4 x i32> %i2, %i3
152 store <4 x i32> %i4, ptr %c
153 ret void
154 }