Merging r259228:
[oota-llvm.git] / test / MC / AArch64 / neon-2velem.s
1 // RUN: llvm-mc -triple=arm64 -mattr=+neon,+fullfp16 -show-encoding < %s | FileCheck %s
2
3 // Check that the assembler can handle the documented syntax for AArch64
4
5 //------------------------------------------------------------------------------
6 // Instructions with 2 vectors and an element
7 //------------------------------------------------------------------------------
8
9         mla v0.2s, v1.2s, v2.s[2]
10         mla v0.2s, v1.2s, v22.s[2]
11         mla v3.4s, v8.4s, v2.s[1]
12         mla v3.4s, v8.4s, v22.s[3]
13
14 // CHECK: mla   v0.2s, v1.2s, v2.s[2]   // encoding: [0x20,0x08,0x82,0x2f]
15 // CHECK: mla   v0.2s, v1.2s, v22.s[2]  // encoding: [0x20,0x08,0x96,0x2f]
16 // CHECK: mla   v3.4s, v8.4s, v2.s[1]   // encoding: [0x03,0x01,0xa2,0x6f]
17 // CHECK: mla   v3.4s, v8.4s, v22.s[3]  // encoding: [0x03,0x09,0xb6,0x6f]
18
19         mla v0.4h, v1.4h, v2.h[2]
20         mla v0.4h, v1.4h, v15.h[2]
21         mla v0.8h, v1.8h, v2.h[7]
22         mla v0.8h, v1.8h, v14.h[6]
23
24 // CHECK: mla   v0.4h, v1.4h, v2.h[2]   // encoding: [0x20,0x00,0x62,0x2f]
25 // CHECK: mla   v0.4h, v1.4h, v15.h[2]  // encoding: [0x20,0x00,0x6f,0x2f]
26 // CHECK: mla   v0.8h, v1.8h, v2.h[7]   // encoding: [0x20,0x08,0x72,0x6f]
27 // CHECK: mla   v0.8h, v1.8h, v14.h[6]  // encoding: [0x20,0x08,0x6e,0x6f]
28
29         mls v0.2s, v1.2s, v2.s[2]
30         mls v0.2s, v1.2s, v22.s[2]
31         mls v3.4s, v8.4s, v2.s[1]
32         mls v3.4s, v8.4s, v22.s[3]
33
34 // CHECK: mls   v0.2s, v1.2s, v2.s[2]   // encoding: [0x20,0x48,0x82,0x2f]
35 // CHECK: mls   v0.2s, v1.2s, v22.s[2]  // encoding: [0x20,0x48,0x96,0x2f]
36 // CHECK: mls   v3.4s, v8.4s, v2.s[1]   // encoding: [0x03,0x41,0xa2,0x6f]
37 // CHECK: mls   v3.4s, v8.4s, v22.s[3]  // encoding: [0x03,0x49,0xb6,0x6f]
38
39         mls v0.4h, v1.4h, v2.h[2]
40         mls v0.4h, v1.4h, v15.h[2]
41         mls v0.8h, v1.8h, v2.h[7]
42         mls v0.8h, v1.8h, v14.h[6]
43
44 // CHECK: mls   v0.4h, v1.4h, v2.h[2]   // encoding: [0x20,0x40,0x62,0x2f]
45 // CHECK: mls   v0.4h, v1.4h, v15.h[2]  // encoding: [0x20,0x40,0x6f,0x2f]
46 // CHECK: mls   v0.8h, v1.8h, v2.h[7]   // encoding: [0x20,0x48,0x72,0x6f]
47 // CHECK: mls   v0.8h, v1.8h, v14.h[6]  // encoding: [0x20,0x48,0x6e,0x6f]
48
49         fmla v0.4h, v1.4h, v2.h[2]
50         fmla v3.8h, v8.8h, v2.h[1]
51         fmla v0.2s, v1.2s, v2.s[2]
52         fmla v0.2s, v1.2s, v22.s[2]
53         fmla v3.4s, v8.4s, v2.s[1]
54         fmla v3.4s, v8.4s, v22.s[3]
55         fmla v0.2d, v1.2d, v2.d[1]
56         fmla v0.2d, v1.2d, v22.d[1]
57
58 // CHECK: fmla    v0.4h, v1.4h, v2.h[2]   // encoding: [0x20,0x10,0x22,0x0f]
59 // CHECK: fmla    v3.8h, v8.8h, v2.h[1]   // encoding: [0x03,0x11,0x12,0x4f]
60 // CHECK: fmla  v0.2s, v1.2s, v2.s[2]   // encoding: [0x20,0x18,0x82,0x0f]
61 // CHECK: fmla  v0.2s, v1.2s, v22.s[2]  // encoding: [0x20,0x18,0x96,0x0f]
62 // CHECK: fmla  v3.4s, v8.4s, v2.s[1]   // encoding: [0x03,0x11,0xa2,0x4f]
63 // CHECK: fmla  v3.4s, v8.4s, v22.s[3]  // encoding: [0x03,0x19,0xb6,0x4f]
64 // CHECK: fmla  v0.2d, v1.2d, v2.d[1]   // encoding: [0x20,0x18,0xc2,0x4f]
65 // CHECK: fmla  v0.2d, v1.2d, v22.d[1]  // encoding: [0x20,0x18,0xd6,0x4f]
66
67         fmls v0.4h, v1.4h, v2.h[2]
68         fmls v3.8h, v8.8h, v2.h[1]
69         fmls v0.2s, v1.2s, v2.s[2]
70         fmls v0.2s, v1.2s, v22.s[2]
71         fmls v3.4s, v8.4s, v2.s[1]
72         fmls v3.4s, v8.4s, v22.s[3]
73         fmls v0.2d, v1.2d, v2.d[1]
74         fmls v0.2d, v1.2d, v22.d[1]
75
76 // CHECK: fmls    v0.4h, v1.4h, v2.h[2]   // encoding: [0x20,0x50,0x22,0x0f]
77 // CHECK: fmls    v3.8h, v8.8h, v2.h[1]   // encoding: [0x03,0x51,0x12,0x4f]
78 // CHECK: fmls  v0.2s, v1.2s, v2.s[2]   // encoding: [0x20,0x58,0x82,0x0f]
79 // CHECK: fmls  v0.2s, v1.2s, v22.s[2]  // encoding: [0x20,0x58,0x96,0x0f]
80 // CHECK: fmls  v3.4s, v8.4s, v2.s[1]   // encoding: [0x03,0x51,0xa2,0x4f]
81 // CHECK: fmls  v3.4s, v8.4s, v22.s[3]  // encoding: [0x03,0x59,0xb6,0x4f]
82 // CHECK: fmls  v0.2d, v1.2d, v2.d[1]   // encoding: [0x20,0x58,0xc2,0x4f]
83 // CHECK: fmls  v0.2d, v1.2d, v22.d[1]  // encoding: [0x20,0x58,0xd6,0x4f]
84
85         smlal v0.4s, v1.4h, v2.h[2]
86         smlal v0.2d, v1.2s, v2.s[2]
87         smlal v0.2d, v1.2s, v22.s[2]
88         smlal2 v0.4s, v1.8h, v1.h[2]
89         smlal2 v0.2d, v1.4s, v1.s[2]
90         smlal2 v0.2d, v1.4s, v22.s[2]
91
92 // CHECK: smlal v0.4s, v1.4h, v2.h[2]   // encoding: [0x20,0x20,0x62,0x0f]
93 // CHECK: smlal v0.2d, v1.2s, v2.s[2]   // encoding: [0x20,0x28,0x82,0x0f]
94 // CHECK: smlal v0.2d, v1.2s, v22.s[2]  // encoding: [0x20,0x28,0x96,0x0f]
95 // CHECK: smlal2        v0.4s, v1.8h, v1.h[2]   // encoding: [0x20,0x20,0x61,0x4f]
96 // CHECK: smlal2        v0.2d, v1.4s, v1.s[2]   // encoding: [0x20,0x28,0x81,0x4f]
97 // CHECK: smlal2        v0.2d, v1.4s, v22.s[2]  // encoding: [0x20,0x28,0x96,0x4f]
98
99         smlsl v0.4s, v1.4h, v2.h[2]
100         smlsl v0.2d, v1.2s, v2.s[2]
101         smlsl v0.2d, v1.2s, v22.s[2]
102         smlsl2 v0.4s, v1.8h, v1.h[2]
103         smlsl2 v0.2d, v1.4s, v1.s[2]
104         smlsl2 v0.2d, v1.4s, v22.s[2]
105
106 // CHECK: smlsl v0.4s, v1.4h, v2.h[2]   // encoding: [0x20,0x60,0x62,0x0f]
107 // CHECK: smlsl v0.2d, v1.2s, v2.s[2]   // encoding: [0x20,0x68,0x82,0x0f]
108 // CHECK: smlsl v0.2d, v1.2s, v22.s[2]  // encoding: [0x20,0x68,0x96,0x0f]
109 // CHECK: smlsl2        v0.4s, v1.8h, v1.h[2]   // encoding: [0x20,0x60,0x61,0x4f]
110 // CHECK: smlsl2        v0.2d, v1.4s, v1.s[2]   // encoding: [0x20,0x68,0x81,0x4f]
111 // CHECK: smlsl2        v0.2d, v1.4s, v22.s[2]  // encoding: [0x20,0x68,0x96,0x4f]
112
113         sqdmlal v0.4s, v1.4h, v2.h[2]
114         sqdmlal v0.2d, v1.2s, v2.s[2]
115         sqdmlal v0.2d, v1.2s, v22.s[2]
116         sqdmlal2 v0.4s, v1.8h, v1.h[2]
117         sqdmlal2 v0.2d, v1.4s, v1.s[2]
118         sqdmlal2 v0.2d, v1.4s, v22.s[2]
119
120 // CHECK: sqdmlal       v0.4s, v1.4h, v2.h[2]   // encoding: [0x20,0x30,0x62,0x0f]
121 // CHECK: sqdmlal       v0.2d, v1.2s, v2.s[2]   // encoding: [0x20,0x38,0x82,0x0f]
122 // CHECK: sqdmlal       v0.2d, v1.2s, v22.s[2]  // encoding: [0x20,0x38,0x96,0x0f]
123 // CHECK: sqdmlal2      v0.4s, v1.8h, v1.h[2] // encoding: [0x20,0x30,0x61,0x4f]
124 // CHECK: sqdmlal2      v0.2d, v1.4s, v1.s[2] // encoding: [0x20,0x38,0x81,0x4f]
125 // CHECK: sqdmlal2      v0.2d, v1.4s, v22.s[2] // encoding: [0x20,0x38,0x96,0x4f]
126
127         umlal v0.4s, v1.4h, v2.h[2]
128         umlal v0.2d, v1.2s, v2.s[2]
129         umlal v0.2d, v1.2s, v22.s[2]
130         umlal2 v0.4s, v1.8h, v1.h[2]
131         umlal2 v0.2d, v1.4s, v1.s[2]
132         umlal2 v0.2d, v1.4s, v22.s[2]
133
134 // CHECK: umlal v0.4s, v1.4h, v2.h[2]   // encoding: [0x20,0x20,0x62,0x2f]
135 // CHECK: umlal v0.2d, v1.2s, v2.s[2]   // encoding: [0x20,0x28,0x82,0x2f]
136 // CHECK: umlal v0.2d, v1.2s, v22.s[2]  // encoding: [0x20,0x28,0x96,0x2f]
137 // CHECK: umlal2        v0.4s, v1.8h, v1.h[2]   // encoding: [0x20,0x20,0x61,0x6f]
138 // CHECK: umlal2        v0.2d, v1.4s, v1.s[2]   // encoding: [0x20,0x28,0x81,0x6f]
139 // CHECK: umlal2        v0.2d, v1.4s, v22.s[2]  // encoding: [0x20,0x28,0x96,0x6f]
140
141         umlsl v0.4s, v1.4h, v2.h[2]
142         umlsl v0.2d, v1.2s, v2.s[2]
143         umlsl v0.2d, v1.2s, v22.s[2]
144         umlsl2 v0.4s, v1.8h, v1.h[2]
145         umlsl2 v0.2d, v1.4s, v1.s[2]
146         umlsl2 v0.2d, v1.4s, v22.s[2]
147
148 // CHECK: umlsl v0.4s, v1.4h, v2.h[2]   // encoding: [0x20,0x60,0x62,0x2f]
149 // CHECK: umlsl v0.2d, v1.2s, v2.s[2]   // encoding: [0x20,0x68,0x82,0x2f]
150 // CHECK: umlsl v0.2d, v1.2s, v22.s[2]  // encoding: [0x20,0x68,0x96,0x2f]
151 // CHECK: umlsl2        v0.4s, v1.8h, v1.h[2]   // encoding: [0x20,0x60,0x61,0x6f]
152 // CHECK: umlsl2        v0.2d, v1.4s, v1.s[2]   // encoding: [0x20,0x68,0x81,0x6f]
153 // CHECK: umlsl2        v0.2d, v1.4s, v22.s[2]  // encoding: [0x20,0x68,0x96,0x6f]
154
155         sqdmlsl v0.4s, v1.4h, v2.h[2]
156         sqdmlsl v0.2d, v1.2s, v2.s[2]
157         sqdmlsl v0.2d, v1.2s, v22.s[2]
158         sqdmlsl2 v0.4s, v1.8h, v1.h[2]
159         sqdmlsl2 v0.2d, v1.4s, v1.s[2]
160         sqdmlsl2 v0.2d, v1.4s, v22.s[2]
161
162 // CHECK: sqdmlsl       v0.4s, v1.4h, v2.h[2]   // encoding: [0x20,0x70,0x62,0x0f]
163 // CHECK: sqdmlsl       v0.2d, v1.2s, v2.s[2]   // encoding: [0x20,0x78,0x82,0x0f]
164 // CHECK: sqdmlsl       v0.2d, v1.2s, v22.s[2]  // encoding: [0x20,0x78,0x96,0x0f]
165 // CHECK: sqdmlsl2      v0.4s, v1.8h, v1.h[2] // encoding: [0x20,0x70,0x61,0x4f]
166 // CHECK: sqdmlsl2      v0.2d, v1.4s, v1.s[2] // encoding: [0x20,0x78,0x81,0x4f]
167 // CHECK: sqdmlsl2      v0.2d, v1.4s, v22.s[2] // encoding: [0x20,0x78,0x96,0x4f]
168
169         mul v0.4h, v1.4h, v2.h[2]
170         mul v0.8h, v1.8h, v2.h[2]
171         mul v0.2s, v1.2s, v2.s[2]
172         mul v0.2s, v1.2s, v22.s[2]
173         mul v0.4s, v1.4s, v2.s[2]
174         mul v0.4s, v1.4s, v22.s[2]
175
176 // CHECK: mul   v0.4h, v1.4h, v2.h[2]   // encoding: [0x20,0x80,0x62,0x0f]
177 // CHECK: mul   v0.8h, v1.8h, v2.h[2]   // encoding: [0x20,0x80,0x62,0x4f]
178 // CHECK: mul   v0.2s, v1.2s, v2.s[2]   // encoding: [0x20,0x88,0x82,0x0f]
179 // CHECK: mul   v0.2s, v1.2s, v22.s[2]  // encoding: [0x20,0x88,0x96,0x0f]
180 // CHECK: mul   v0.4s, v1.4s, v2.s[2]   // encoding: [0x20,0x88,0x82,0x4f]
181 // CHECK: mul   v0.4s, v1.4s, v22.s[2]  // encoding: [0x20,0x88,0x96,0x4f]
182
183         fmul v0.4h, v1.4h, v2.h[2]
184         fmul v0.8h, v1.8h, v2.h[2]
185         fmul v0.2s, v1.2s, v2.s[2]
186         fmul v0.2s, v1.2s, v22.s[2]
187         fmul v0.4s, v1.4s, v2.s[2]
188         fmul v0.4s, v1.4s, v22.s[2]
189         fmul v0.2d, v1.2d, v2.d[1]
190         fmul v0.2d, v1.2d, v22.d[1]
191
192 // CHECK: fmul    v0.4h, v1.4h, v2.h[2]   // encoding: [0x20,0x90,0x22,0x0f]
193 // CHECK: fmul    v0.8h, v1.8h, v2.h[2]   // encoding: [0x20,0x90,0x22,0x4f]
194 // CHECK: fmul  v0.2s, v1.2s, v2.s[2]   // encoding: [0x20,0x98,0x82,0x0f]
195 // CHECK: fmul  v0.2s, v1.2s, v22.s[2]  // encoding: [0x20,0x98,0x96,0x0f]
196 // CHECK: fmul  v0.4s, v1.4s, v2.s[2]   // encoding: [0x20,0x98,0x82,0x4f]
197 // CHECK: fmul  v0.4s, v1.4s, v22.s[2]  // encoding: [0x20,0x98,0x96,0x4f]
198 // CHECK: fmul  v0.2d, v1.2d, v2.d[1]   // encoding: [0x20,0x98,0xc2,0x4f]
199 // CHECK: fmul  v0.2d, v1.2d, v22.d[1]  // encoding: [0x20,0x98,0xd6,0x4f]
200
201         fmulx v0.4h, v1.4h, v2.h[2]
202         fmulx v0.8h, v1.8h, v2.h[2]
203         fmulx v0.2s, v1.2s, v2.s[2]
204         fmulx v0.2s, v1.2s, v22.s[2]
205         fmulx v0.4s, v1.4s, v2.s[2]
206         fmulx v0.4s, v1.4s, v22.s[2]
207         fmulx v0.2d, v1.2d, v2.d[1]
208         fmulx v0.2d, v1.2d, v22.d[1]
209
210 // CHECK: fmulx   v0.4h, v1.4h, v2.h[2]   // encoding: [0x20,0x90,0x22,0x2f]
211 // CHECK: fmulx   v0.8h, v1.8h, v2.h[2]   // encoding: [0x20,0x90,0x22,0x6f]
212 // CHECK: fmulx v0.2s, v1.2s, v2.s[2]   // encoding: [0x20,0x98,0x82,0x2f]
213 // CHECK: fmulx v0.2s, v1.2s, v22.s[2]  // encoding: [0x20,0x98,0x96,0x2f]
214 // CHECK: fmulx v0.4s, v1.4s, v2.s[2]   // encoding: [0x20,0x98,0x82,0x6f]
215 // CHECK: fmulx v0.4s, v1.4s, v22.s[2]  // encoding: [0x20,0x98,0x96,0x6f]
216 // CHECK: fmulx v0.2d, v1.2d, v2.d[1]   // encoding: [0x20,0x98,0xc2,0x6f]
217 // CHECK: fmulx v0.2d, v1.2d, v22.d[1]  // encoding: [0x20,0x98,0xd6,0x6f]
218
219         smull v0.4s, v1.4h, v2.h[2]
220         smull v0.2d, v1.2s, v2.s[2]
221         smull v0.2d, v1.2s, v22.s[2]
222         smull2 v0.4s, v1.8h, v2.h[2]
223         smull2 v0.2d, v1.4s, v2.s[2]
224         smull2 v0.2d, v1.4s, v22.s[2]
225
226 // CHECK: smull v0.4s, v1.4h, v2.h[2]   // encoding: [0x20,0xa0,0x62,0x0f]
227 // CHECK: smull v0.2d, v1.2s, v2.s[2]   // encoding: [0x20,0xa8,0x82,0x0f]
228 // CHECK: smull v0.2d, v1.2s, v22.s[2]  // encoding: [0x20,0xa8,0x96,0x0f]
229 // CHECK: smull2        v0.4s, v1.8h, v2.h[2]   // encoding: [0x20,0xa0,0x62,0x4f]
230 // CHECK: smull2        v0.2d, v1.4s, v2.s[2]   // encoding: [0x20,0xa8,0x82,0x4f]
231 // CHECK: smull2        v0.2d, v1.4s, v22.s[2]  // encoding: [0x20,0xa8,0x96,0x4f]
232
233         umull v0.4s, v1.4h, v2.h[2]
234         umull v0.2d, v1.2s, v2.s[2]
235         umull v0.2d, v1.2s, v22.s[2]
236         umull2 v0.4s, v1.8h, v2.h[2]
237         umull2 v0.2d, v1.4s, v2.s[2]
238         umull2 v0.2d, v1.4s, v22.s[2]
239
240 // CHECK: umull v0.4s, v1.4h, v2.h[2]   // encoding: [0x20,0xa0,0x62,0x2f]
241 // CHECK: umull v0.2d, v1.2s, v2.s[2]   // encoding: [0x20,0xa8,0x82,0x2f]
242 // CHECK: umull v0.2d, v1.2s, v22.s[2]  // encoding: [0x20,0xa8,0x96,0x2f]
243 // CHECK: umull2        v0.4s, v1.8h, v2.h[2]   // encoding: [0x20,0xa0,0x62,0x6f]
244 // CHECK: umull2        v0.2d, v1.4s, v2.s[2]   // encoding: [0x20,0xa8,0x82,0x6f]
245 // CHECK: umull2        v0.2d, v1.4s, v22.s[2]  // encoding: [0x20,0xa8,0x96,0x6f]
246
247         sqdmull v0.4s, v1.4h, v2.h[2]
248         sqdmull v0.2d, v1.2s, v2.s[2]
249         sqdmull v0.2d, v1.2s, v22.s[2]
250         sqdmull2 v0.4s, v1.8h, v2.h[2]
251         sqdmull2 v0.2d, v1.4s, v2.s[2]
252         sqdmull2 v0.2d, v1.4s, v22.s[2]
253
254 // CHECK: sqdmull       v0.4s, v1.4h, v2.h[2]   // encoding: [0x20,0xb0,0x62,0x0f]
255 // CHECK: sqdmull       v0.2d, v1.2s, v2.s[2]   // encoding: [0x20,0xb8,0x82,0x0f]
256 // CHECK: sqdmull       v0.2d, v1.2s, v22.s[2]  // encoding: [0x20,0xb8,0x96,0x0f]
257 // CHECK: sqdmull2      v0.4s, v1.8h, v2.h[2] // encoding: [0x20,0xb0,0x62,0x4f]
258 // CHECK: sqdmull2      v0.2d, v1.4s, v2.s[2] // encoding: [0x20,0xb8,0x82,0x4f]
259 // CHECK: sqdmull2      v0.2d, v1.4s, v22.s[2] // encoding: [0x20,0xb8,0x96,0x4f]
260
261         sqdmulh v0.4h, v1.4h, v2.h[2]
262         sqdmulh v0.8h, v1.8h, v2.h[2]
263         sqdmulh v0.2s, v1.2s, v2.s[2]
264         sqdmulh v0.2s, v1.2s, v22.s[2]
265         sqdmulh v0.4s, v1.4s, v2.s[2]
266         sqdmulh v0.4s, v1.4s, v22.s[2]
267
268 // CHECK: sqdmulh       v0.4h, v1.4h, v2.h[2]   // encoding: [0x20,0xc0,0x62,0x0f]
269 // CHECK: sqdmulh       v0.8h, v1.8h, v2.h[2]   // encoding: [0x20,0xc0,0x62,0x4f]
270 // CHECK: sqdmulh       v0.2s, v1.2s, v2.s[2]   // encoding: [0x20,0xc8,0x82,0x0f]
271 // CHECK: sqdmulh       v0.2s, v1.2s, v22.s[2]  // encoding: [0x20,0xc8,0x96,0x0f]
272 // CHECK: sqdmulh       v0.4s, v1.4s, v2.s[2]   // encoding: [0x20,0xc8,0x82,0x4f]
273 // CHECK: sqdmulh       v0.4s, v1.4s, v22.s[2]  // encoding: [0x20,0xc8,0x96,0x4f]
274
275         sqrdmulh v0.4h, v1.4h, v2.h[2]
276         sqrdmulh v0.8h, v1.8h, v2.h[2]
277         sqrdmulh v0.2s, v1.2s, v2.s[2]
278         sqrdmulh v0.2s, v1.2s, v22.s[2]
279         sqrdmulh v0.4s, v1.4s, v2.s[2]
280         sqrdmulh v0.4s, v1.4s, v22.s[2]
281
282 // CHECK: sqrdmulh      v0.4h, v1.4h, v2.h[2] // encoding: [0x20,0xd0,0x62,0x0f]
283 // CHECK: sqrdmulh      v0.8h, v1.8h, v2.h[2] // encoding: [0x20,0xd0,0x62,0x4f]
284 // CHECK: sqrdmulh      v0.2s, v1.2s, v2.s[2] // encoding: [0x20,0xd8,0x82,0x0f]
285 // CHECK: sqrdmulh      v0.2s, v1.2s, v22.s[2] // encoding: [0x20,0xd8,0x96,0x0f]
286 // CHECK: sqrdmulh      v0.4s, v1.4s, v2.s[2] // encoding: [0x20,0xd8,0x82,0x4f]
287 // CHECK: sqrdmulh      v0.4s, v1.4s, v22.s[2] // encoding: [0x20,0xd8,0x96,0x4f]