test/MC/AArch64/neon-2velem.s

   1 // RUN: llvm-mc -triple=arm64 -mattr=+neon -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 v0.2s, v1.2s, v2.s[2]
  50         fmla v0.2s, v1.2s, v22.s[2]
  51         fmla v3.4s, v8.4s, v2.s[1]
  52         fmla v3.4s, v8.4s, v22.s[3]
  53         fmla v0.2d, v1.2d, v2.d[1]
  54         fmla v0.2d, v1.2d, v22.d[1]
  55
  56 // CHECK: fmla  v0.2s, v1.2s, v2.s[2]   // encoding: [0x20,0x18,0x82,0x0f]
  57 // CHECK: fmla  v0.2s, v1.2s, v22.s[2]  // encoding: [0x20,0x18,0x96,0x0f]
  58 // CHECK: fmla  v3.4s, v8.4s, v2.s[1]   // encoding: [0x03,0x11,0xa2,0x4f]
  59 // CHECK: fmla  v3.4s, v8.4s, v22.s[3]  // encoding: [0x03,0x19,0xb6,0x4f]
  60 // CHECK: fmla  v0.2d, v1.2d, v2.d[1]   // encoding: [0x20,0x18,0xc2,0x4f]
  61 // CHECK: fmla  v0.2d, v1.2d, v22.d[1]  // encoding: [0x20,0x18,0xd6,0x4f]
  62
  63         fmls v0.2s, v1.2s, v2.s[2]
  64         fmls v0.2s, v1.2s, v22.s[2]
  65         fmls v3.4s, v8.4s, v2.s[1]
  66         fmls v3.4s, v8.4s, v22.s[3]
  67         fmls v0.2d, v1.2d, v2.d[1]
  68         fmls v0.2d, v1.2d, v22.d[1]
  69
  70 // CHECK: fmls  v0.2s, v1.2s, v2.s[2]   // encoding: [0x20,0x58,0x82,0x0f]
  71 // CHECK: fmls  v0.2s, v1.2s, v22.s[2]  // encoding: [0x20,0x58,0x96,0x0f]
  72 // CHECK: fmls  v3.4s, v8.4s, v2.s[1]   // encoding: [0x03,0x51,0xa2,0x4f]
  73 // CHECK: fmls  v3.4s, v8.4s, v22.s[3]  // encoding: [0x03,0x59,0xb6,0x4f]
  74 // CHECK: fmls  v0.2d, v1.2d, v2.d[1]   // encoding: [0x20,0x58,0xc2,0x4f]
  75 // CHECK: fmls  v0.2d, v1.2d, v22.d[1]  // encoding: [0x20,0x58,0xd6,0x4f]
  76
  77         smlal v0.4s, v1.4h, v2.h[2]
  78         smlal v0.2d, v1.2s, v2.s[2]
  79         smlal v0.2d, v1.2s, v22.s[2]
  80         smlal2 v0.4s, v1.8h, v1.h[2]
  81         smlal2 v0.2d, v1.4s, v1.s[2]
  82         smlal2 v0.2d, v1.4s, v22.s[2]
  83
  84 // CHECK: smlal v0.4s, v1.4h, v2.h[2]   // encoding: [0x20,0x20,0x62,0x0f]
  85 // CHECK: smlal v0.2d, v1.2s, v2.s[2]   // encoding: [0x20,0x28,0x82,0x0f]
  86 // CHECK: smlal v0.2d, v1.2s, v22.s[2]  // encoding: [0x20,0x28,0x96,0x0f]
  87 // CHECK: smlal2        v0.4s, v1.8h, v1.h[2]   // encoding: [0x20,0x20,0x61,0x4f]
  88 // CHECK: smlal2        v0.2d, v1.4s, v1.s[2]   // encoding: [0x20,0x28,0x81,0x4f]
  89 // CHECK: smlal2        v0.2d, v1.4s, v22.s[2]  // encoding: [0x20,0x28,0x96,0x4f]
  90
  91         smlsl v0.4s, v1.4h, v2.h[2]
  92         smlsl v0.2d, v1.2s, v2.s[2]
  93         smlsl v0.2d, v1.2s, v22.s[2]
  94         smlsl2 v0.4s, v1.8h, v1.h[2]
  95         smlsl2 v0.2d, v1.4s, v1.s[2]
  96         smlsl2 v0.2d, v1.4s, v22.s[2]
  97
  98 // CHECK: smlsl v0.4s, v1.4h, v2.h[2]   // encoding: [0x20,0x60,0x62,0x0f]
  99 // CHECK: smlsl v0.2d, v1.2s, v2.s[2]   // encoding: [0x20,0x68,0x82,0x0f]
 100 // CHECK: smlsl v0.2d, v1.2s, v22.s[2]  // encoding: [0x20,0x68,0x96,0x0f]
 101 // CHECK: smlsl2        v0.4s, v1.8h, v1.h[2]   // encoding: [0x20,0x60,0x61,0x4f]
 102 // CHECK: smlsl2        v0.2d, v1.4s, v1.s[2]   // encoding: [0x20,0x68,0x81,0x4f]
 103 // CHECK: smlsl2        v0.2d, v1.4s, v22.s[2]  // encoding: [0x20,0x68,0x96,0x4f]
 104
 105         sqdmlal v0.4s, v1.4h, v2.h[2]
 106         sqdmlal v0.2d, v1.2s, v2.s[2]
 107         sqdmlal v0.2d, v1.2s, v22.s[2]
 108         sqdmlal2 v0.4s, v1.8h, v1.h[2]
 109         sqdmlal2 v0.2d, v1.4s, v1.s[2]
 110         sqdmlal2 v0.2d, v1.4s, v22.s[2]
 111
 112 // CHECK: sqdmlal       v0.4s, v1.4h, v2.h[2]   // encoding: [0x20,0x30,0x62,0x0f]
 113 // CHECK: sqdmlal       v0.2d, v1.2s, v2.s[2]   // encoding: [0x20,0x38,0x82,0x0f]
 114 // CHECK: sqdmlal       v0.2d, v1.2s, v22.s[2]  // encoding: [0x20,0x38,0x96,0x0f]
 115 // CHECK: sqdmlal2      v0.4s, v1.8h, v1.h[2] // encoding: [0x20,0x30,0x61,0x4f]
 116 // CHECK: sqdmlal2      v0.2d, v1.4s, v1.s[2] // encoding: [0x20,0x38,0x81,0x4f]
 117 // CHECK: sqdmlal2      v0.2d, v1.4s, v22.s[2] // encoding: [0x20,0x38,0x96,0x4f]
 118
 119         umlal v0.4s, v1.4h, v2.h[2]
 120         umlal v0.2d, v1.2s, v2.s[2]
 121         umlal v0.2d, v1.2s, v22.s[2]
 122         umlal2 v0.4s, v1.8h, v1.h[2]
 123         umlal2 v0.2d, v1.4s, v1.s[2]
 124         umlal2 v0.2d, v1.4s, v22.s[2]
 125
 126 // CHECK: umlal v0.4s, v1.4h, v2.h[2]   // encoding: [0x20,0x20,0x62,0x2f]
 127 // CHECK: umlal v0.2d, v1.2s, v2.s[2]   // encoding: [0x20,0x28,0x82,0x2f]
 128 // CHECK: umlal v0.2d, v1.2s, v22.s[2]  // encoding: [0x20,0x28,0x96,0x2f]
 129 // CHECK: umlal2        v0.4s, v1.8h, v1.h[2]   // encoding: [0x20,0x20,0x61,0x6f]
 130 // CHECK: umlal2        v0.2d, v1.4s, v1.s[2]   // encoding: [0x20,0x28,0x81,0x6f]
 131 // CHECK: umlal2        v0.2d, v1.4s, v22.s[2]  // encoding: [0x20,0x28,0x96,0x6f]
 132
 133         umlsl v0.4s, v1.4h, v2.h[2]
 134         umlsl v0.2d, v1.2s, v2.s[2]
 135         umlsl v0.2d, v1.2s, v22.s[2]
 136         umlsl2 v0.4s, v1.8h, v1.h[2]
 137         umlsl2 v0.2d, v1.4s, v1.s[2]
 138         umlsl2 v0.2d, v1.4s, v22.s[2]
 139
 140 // CHECK: umlsl v0.4s, v1.4h, v2.h[2]   // encoding: [0x20,0x60,0x62,0x2f]
 141 // CHECK: umlsl v0.2d, v1.2s, v2.s[2]   // encoding: [0x20,0x68,0x82,0x2f]
 142 // CHECK: umlsl v0.2d, v1.2s, v22.s[2]  // encoding: [0x20,0x68,0x96,0x2f]
 143 // CHECK: umlsl2        v0.4s, v1.8h, v1.h[2]   // encoding: [0x20,0x60,0x61,0x6f]
 144 // CHECK: umlsl2        v0.2d, v1.4s, v1.s[2]   // encoding: [0x20,0x68,0x81,0x6f]
 145 // CHECK: umlsl2        v0.2d, v1.4s, v22.s[2]  // encoding: [0x20,0x68,0x96,0x6f]
 146
 147         sqdmlsl v0.4s, v1.4h, v2.h[2]
 148         sqdmlsl v0.2d, v1.2s, v2.s[2]
 149         sqdmlsl v0.2d, v1.2s, v22.s[2]
 150         sqdmlsl2 v0.4s, v1.8h, v1.h[2]
 151         sqdmlsl2 v0.2d, v1.4s, v1.s[2]
 152         sqdmlsl2 v0.2d, v1.4s, v22.s[2]
 153
 154 // CHECK: sqdmlsl       v0.4s, v1.4h, v2.h[2]   // encoding: [0x20,0x70,0x62,0x0f]
 155 // CHECK: sqdmlsl       v0.2d, v1.2s, v2.s[2]   // encoding: [0x20,0x78,0x82,0x0f]
 156 // CHECK: sqdmlsl       v0.2d, v1.2s, v22.s[2]  // encoding: [0x20,0x78,0x96,0x0f]
 157 // CHECK: sqdmlsl2      v0.4s, v1.8h, v1.h[2] // encoding: [0x20,0x70,0x61,0x4f]
 158 // CHECK: sqdmlsl2      v0.2d, v1.4s, v1.s[2] // encoding: [0x20,0x78,0x81,0x4f]
 159 // CHECK: sqdmlsl2      v0.2d, v1.4s, v22.s[2] // encoding: [0x20,0x78,0x96,0x4f]
 160
 161         mul v0.4h, v1.4h, v2.h[2]
 162         mul v0.8h, v1.8h, v2.h[2]
 163         mul v0.2s, v1.2s, v2.s[2]
 164         mul v0.2s, v1.2s, v22.s[2]
 165         mul v0.4s, v1.4s, v2.s[2]
 166         mul v0.4s, v1.4s, v22.s[2]
 167
 168 // CHECK: mul   v0.4h, v1.4h, v2.h[2]   // encoding: [0x20,0x80,0x62,0x0f]
 169 // CHECK: mul   v0.8h, v1.8h, v2.h[2]   // encoding: [0x20,0x80,0x62,0x4f]
 170 // CHECK: mul   v0.2s, v1.2s, v2.s[2]   // encoding: [0x20,0x88,0x82,0x0f]
 171 // CHECK: mul   v0.2s, v1.2s, v22.s[2]  // encoding: [0x20,0x88,0x96,0x0f]
 172 // CHECK: mul   v0.4s, v1.4s, v2.s[2]   // encoding: [0x20,0x88,0x82,0x4f]
 173 // CHECK: mul   v0.4s, v1.4s, v22.s[2]  // encoding: [0x20,0x88,0x96,0x4f]
 174
 175         fmul v0.2s, v1.2s, v2.s[2]
 176         fmul v0.2s, v1.2s, v22.s[2]
 177         fmul v0.4s, v1.4s, v2.s[2]
 178         fmul v0.4s, v1.4s, v22.s[2]
 179         fmul v0.2d, v1.2d, v2.d[1]
 180         fmul v0.2d, v1.2d, v22.d[1]
 181
 182 // CHECK: fmul  v0.2s, v1.2s, v2.s[2]   // encoding: [0x20,0x98,0x82,0x0f]
 183 // CHECK: fmul  v0.2s, v1.2s, v22.s[2]  // encoding: [0x20,0x98,0x96,0x0f]
 184 // CHECK: fmul  v0.4s, v1.4s, v2.s[2]   // encoding: [0x20,0x98,0x82,0x4f]
 185 // CHECK: fmul  v0.4s, v1.4s, v22.s[2]  // encoding: [0x20,0x98,0x96,0x4f]
 186 // CHECK: fmul  v0.2d, v1.2d, v2.d[1]   // encoding: [0x20,0x98,0xc2,0x4f]
 187 // CHECK: fmul  v0.2d, v1.2d, v22.d[1]  // encoding: [0x20,0x98,0xd6,0x4f]
 188
 189         fmulx v0.2s, v1.2s, v2.s[2]
 190         fmulx v0.2s, v1.2s, v22.s[2]
 191         fmulx v0.4s, v1.4s, v2.s[2]
 192         fmulx v0.4s, v1.4s, v22.s[2]
 193         fmulx v0.2d, v1.2d, v2.d[1]
 194         fmulx v0.2d, v1.2d, v22.d[1]
 195
 196 // CHECK: fmulx v0.2s, v1.2s, v2.s[2]   // encoding: [0x20,0x98,0x82,0x2f]
 197 // CHECK: fmulx v0.2s, v1.2s, v22.s[2]  // encoding: [0x20,0x98,0x96,0x2f]
 198 // CHECK: fmulx v0.4s, v1.4s, v2.s[2]   // encoding: [0x20,0x98,0x82,0x6f]
 199 // CHECK: fmulx v0.4s, v1.4s, v22.s[2]  // encoding: [0x20,0x98,0x96,0x6f]
 200 // CHECK: fmulx v0.2d, v1.2d, v2.d[1]   // encoding: [0x20,0x98,0xc2,0x6f]
 201 // CHECK: fmulx v0.2d, v1.2d, v22.d[1]  // encoding: [0x20,0x98,0xd6,0x6f]
 202
 203         smull v0.4s, v1.4h, v2.h[2]
 204         smull v0.2d, v1.2s, v2.s[2]
 205         smull v0.2d, v1.2s, v22.s[2]
 206         smull2 v0.4s, v1.8h, v2.h[2]
 207         smull2 v0.2d, v1.4s, v2.s[2]
 208         smull2 v0.2d, v1.4s, v22.s[2]
 209
 210 // CHECK: smull v0.4s, v1.4h, v2.h[2]   // encoding: [0x20,0xa0,0x62,0x0f]
 211 // CHECK: smull v0.2d, v1.2s, v2.s[2]   // encoding: [0x20,0xa8,0x82,0x0f]
 212 // CHECK: smull v0.2d, v1.2s, v22.s[2]  // encoding: [0x20,0xa8,0x96,0x0f]
 213 // CHECK: smull2        v0.4s, v1.8h, v2.h[2]   // encoding: [0x20,0xa0,0x62,0x4f]
 214 // CHECK: smull2        v0.2d, v1.4s, v2.s[2]   // encoding: [0x20,0xa8,0x82,0x4f]
 215 // CHECK: smull2        v0.2d, v1.4s, v22.s[2]  // encoding: [0x20,0xa8,0x96,0x4f]
 216
 217         umull v0.4s, v1.4h, v2.h[2]
 218         umull v0.2d, v1.2s, v2.s[2]
 219         umull v0.2d, v1.2s, v22.s[2]
 220         umull2 v0.4s, v1.8h, v2.h[2]
 221         umull2 v0.2d, v1.4s, v2.s[2]
 222         umull2 v0.2d, v1.4s, v22.s[2]
 223
 224 // CHECK: umull v0.4s, v1.4h, v2.h[2]   // encoding: [0x20,0xa0,0x62,0x2f]
 225 // CHECK: umull v0.2d, v1.2s, v2.s[2]   // encoding: [0x20,0xa8,0x82,0x2f]
 226 // CHECK: umull v0.2d, v1.2s, v22.s[2]  // encoding: [0x20,0xa8,0x96,0x2f]
 227 // CHECK: umull2        v0.4s, v1.8h, v2.h[2]   // encoding: [0x20,0xa0,0x62,0x6f]
 228 // CHECK: umull2        v0.2d, v1.4s, v2.s[2]   // encoding: [0x20,0xa8,0x82,0x6f]
 229 // CHECK: umull2        v0.2d, v1.4s, v22.s[2]  // encoding: [0x20,0xa8,0x96,0x6f]
 230
 231         sqdmull v0.4s, v1.4h, v2.h[2]
 232         sqdmull v0.2d, v1.2s, v2.s[2]
 233         sqdmull v0.2d, v1.2s, v22.s[2]
 234         sqdmull2 v0.4s, v1.8h, v2.h[2]
 235         sqdmull2 v0.2d, v1.4s, v2.s[2]
 236         sqdmull2 v0.2d, v1.4s, v22.s[2]
 237
 238 // CHECK: sqdmull       v0.4s, v1.4h, v2.h[2]   // encoding: [0x20,0xb0,0x62,0x0f]
 239 // CHECK: sqdmull       v0.2d, v1.2s, v2.s[2]   // encoding: [0x20,0xb8,0x82,0x0f]
 240 // CHECK: sqdmull       v0.2d, v1.2s, v22.s[2]  // encoding: [0x20,0xb8,0x96,0x0f]
 241 // CHECK: sqdmull2      v0.4s, v1.8h, v2.h[2] // encoding: [0x20,0xb0,0x62,0x4f]
 242 // CHECK: sqdmull2      v0.2d, v1.4s, v2.s[2] // encoding: [0x20,0xb8,0x82,0x4f]
 243 // CHECK: sqdmull2      v0.2d, v1.4s, v22.s[2] // encoding: [0x20,0xb8,0x96,0x4f]
 244
 245         sqdmulh v0.4h, v1.4h, v2.h[2]
 246         sqdmulh v0.8h, v1.8h, v2.h[2]
 247         sqdmulh v0.2s, v1.2s, v2.s[2]
 248         sqdmulh v0.2s, v1.2s, v22.s[2]
 249         sqdmulh v0.4s, v1.4s, v2.s[2]
 250         sqdmulh v0.4s, v1.4s, v22.s[2]
 251
 252 // CHECK: sqdmulh       v0.4h, v1.4h, v2.h[2]   // encoding: [0x20,0xc0,0x62,0x0f]
 253 // CHECK: sqdmulh       v0.8h, v1.8h, v2.h[2]   // encoding: [0x20,0xc0,0x62,0x4f]
 254 // CHECK: sqdmulh       v0.2s, v1.2s, v2.s[2]   // encoding: [0x20,0xc8,0x82,0x0f]
 255 // CHECK: sqdmulh       v0.2s, v1.2s, v22.s[2]  // encoding: [0x20,0xc8,0x96,0x0f]
 256 // CHECK: sqdmulh       v0.4s, v1.4s, v2.s[2]   // encoding: [0x20,0xc8,0x82,0x4f]
 257 // CHECK: sqdmulh       v0.4s, v1.4s, v22.s[2]  // encoding: [0x20,0xc8,0x96,0x4f]
 258
 259         sqrdmulh v0.4h, v1.4h, v2.h[2]
 260         sqrdmulh v0.8h, v1.8h, v2.h[2]
 261         sqrdmulh v0.2s, v1.2s, v2.s[2]
 262         sqrdmulh v0.2s, v1.2s, v22.s[2]
 263         sqrdmulh v0.4s, v1.4s, v2.s[2]
 264         sqrdmulh v0.4s, v1.4s, v22.s[2]
 265
 266 // CHECK: sqrdmulh      v0.4h, v1.4h, v2.h[2] // encoding: [0x20,0xd0,0x62,0x0f]
 267 // CHECK: sqrdmulh      v0.8h, v1.8h, v2.h[2] // encoding: [0x20,0xd0,0x62,0x4f]
 268 // CHECK: sqrdmulh      v0.2s, v1.2s, v2.s[2] // encoding: [0x20,0xd8,0x82,0x0f]
 269 // CHECK: sqrdmulh      v0.2s, v1.2s, v22.s[2] // encoding: [0x20,0xd8,0x96,0x0f]
 270 // CHECK: sqrdmulh      v0.4s, v1.4s, v2.s[2] // encoding: [0x20,0xd8,0x82,0x4f]
 271 // CHECK: sqrdmulh      v0.4s, v1.4s, v22.s[2] // encoding: [0x20,0xd8,0x96,0x4f]