test/tools/llvm-mca/X86/Atom/resources-sse1.s

   1 # NOTE: Assertions have been autogenerated by utils/update_mca_test_checks.py
   2 # RUN: llvm-mca -mtriple=x86_64-unknown-unknown -mcpu=atom -instruction-tables < %s | FileCheck %s
   3
   4 addps       %xmm0, %xmm2
   5 addps       (%rax), %xmm2
   6
   7 addss       %xmm0, %xmm2
   8 addss       (%rax), %xmm2
   9
  10 andnps      %xmm0, %xmm2
  11 andnps      (%rax), %xmm2
  12
  13 andps       %xmm0, %xmm2
  14 andps       (%rax), %xmm2
  15
  16 cmpps       $0, %xmm0, %xmm2
  17 cmpps       $0, (%rax), %xmm2
  18
  19 cmpss       $0, %xmm0, %xmm2
  20 cmpss       $0, (%rax), %xmm2
  21
  22 comiss      %xmm0, %xmm1
  23 comiss      (%rax), %xmm1
  24
  25 cvtpi2ps    %mm0, %xmm2
  26 cvtpi2ps    (%rax), %xmm2
  27
  28 cvtps2pi    %xmm0, %mm2
  29 cvtps2pi    (%rax), %mm2
  30
  31 cvtsi2ss    %ecx, %xmm2
  32 cvtsi2ss    %rcx, %xmm2
  33 cvtsi2ss    (%rax), %xmm2
  34 cvtsi2ss    (%rax), %xmm2
  35
  36 cvtss2si    %xmm0, %ecx
  37 cvtss2si    %xmm0, %rcx
  38 cvtss2si    (%rax), %ecx
  39 cvtss2si    (%rax), %rcx
  40
  41 cvttps2pi   %xmm0, %mm2
  42 cvttps2pi   (%rax), %mm2
  43
  44 cvttss2si   %xmm0, %ecx
  45 cvttss2si   %xmm0, %rcx
  46 cvttss2si   (%rax), %ecx
  47 cvttss2si   (%rax), %rcx
  48
  49 divps       %xmm0, %xmm2
  50 divps       (%rax), %xmm2
  51
  52 divss       %xmm0, %xmm2
  53 divss       (%rax), %xmm2
  54
  55 ldmxcsr     (%rax)
  56
  57 maskmovq    %mm0, %mm1
  58
  59 maxps       %xmm0, %xmm2
  60 maxps       (%rax), %xmm2
  61
  62 maxss       %xmm0, %xmm2
  63 maxss       (%rax), %xmm2
  64
  65 minps       %xmm0, %xmm2
  66 minps       (%rax), %xmm2
  67
  68 minss       %xmm0, %xmm2
  69 minss       (%rax), %xmm2
  70
  71 movaps      %xmm0, %xmm2
  72 movaps      %xmm0, (%rax)
  73 movaps      (%rax), %xmm2
  74
  75 movhlps     %xmm0, %xmm2
  76 movlhps     %xmm0, %xmm2
  77
  78 movhps      %xmm0, (%rax)
  79 movhps      (%rax), %xmm2
  80
  81 movlps      %xmm0, (%rax)
  82 movlps      (%rax), %xmm2
  83
  84 movmskps    %xmm0, %rcx
  85
  86 movntps     %xmm0, (%rax)
  87 movntq      %mm0, (%rax)
  88
  89 movss       %xmm0, %xmm2
  90 movss       %xmm0, (%rax)
  91 movss       (%rax), %xmm2
  92
  93 movups      %xmm0, %xmm2
  94 movups      %xmm0, (%rax)
  95 movups      (%rax), %xmm2
  96
  97 mulps       %xmm0, %xmm2
  98 mulps       (%rax), %xmm2
  99
 100 mulss       %xmm0, %xmm2
 101 mulss       (%rax), %xmm2
 102
 103 orps        %xmm0, %xmm2
 104 orps        (%rax), %xmm2
 105
 106 pavgb       %mm0, %mm2
 107 pavgb       (%rax), %mm2
 108
 109 pavgw       %mm0, %mm2
 110 pavgw       (%rax), %mm2
 111
 112 pextrw      $1, %mm0, %rcx
 113
 114 pinsrw      $1, %rax, %mm2
 115 pinsrw      $1, (%rax), %mm2
 116
 117 pmaxsw      %mm0, %mm2
 118 pmaxsw      (%rax), %mm2
 119
 120 pmaxub      %mm0, %mm2
 121 pmaxub      (%rax), %mm2
 122
 123 pminsw      %mm0, %mm2
 124 pminsw      (%rax), %mm2
 125
 126 pminub      %mm0, %mm2
 127 pminub      (%rax), %mm2
 128
 129 pmovmskb    %mm0, %rcx
 130
 131 pmulhuw     %mm0, %mm2
 132 pmulhuw     (%rax), %mm2
 133
 134 prefetcht0  (%rax)
 135 prefetcht1  (%rax)
 136 prefetcht2  (%rax)
 137 prefetchnta (%rax)
 138
 139 psadbw      %mm0, %mm2
 140 psadbw      (%rax), %mm2
 141
 142 pshufw      $1, %mm0, %mm2
 143 pshufw      $1, (%rax), %mm2
 144
 145 rcpps       %xmm0, %xmm2
 146 rcpps       (%rax), %xmm2
 147
 148 rcpss       %xmm0, %xmm2
 149 rcpss       (%rax), %xmm2
 150
 151 rsqrtps     %xmm0, %xmm2
 152 rsqrtps     (%rax), %xmm2
 153
 154 rsqrtss     %xmm0, %xmm2
 155 rsqrtss     (%rax), %xmm2
 156
 157 sfence
 158
 159 shufps      $1, %xmm0, %xmm2
 160 shufps      $1, (%rax), %xmm2
 161
 162 sqrtps      %xmm0, %xmm2
 163 sqrtps      (%rax), %xmm2
 164
 165 sqrtss      %xmm0, %xmm2
 166 sqrtss      (%rax), %xmm2
 167
 168 stmxcsr     (%rax)
 169
 170 subps       %xmm0, %xmm2
 171 subps       (%rax), %xmm2
 172
 173 subss       %xmm0, %xmm2
 174 subss       (%rax), %xmm2
 175
 176 ucomiss     %xmm0, %xmm1
 177 ucomiss     (%rax), %xmm1
 178
 179 unpckhps    %xmm0, %xmm2
 180 unpckhps    (%rax), %xmm2
 181
 182 unpcklps    %xmm0, %xmm2
 183 unpcklps    (%rax), %xmm2
 184
 185 xorps       %xmm0, %xmm2
 186 xorps       (%rax), %xmm2
 187
 188 # CHECK:      Instruction Info:
 189 # CHECK-NEXT: [1]: #uOps
 190 # CHECK-NEXT: [2]: Latency
 191 # CHECK-NEXT: [3]: RThroughput
 192 # CHECK-NEXT: [4]: MayLoad
 193 # CHECK-NEXT: [5]: MayStore
 194 # CHECK-NEXT: [6]: HasSideEffects (U)
 195
 196 # CHECK:      [1]    [2]    [3]    [4]    [5]    [6]    Instructions:
 197 # CHECK-NEXT:  1      5     5.00                        addps   %xmm0, %xmm2
 198 # CHECK-NEXT:  1      5     5.00    *                   addps   (%rax), %xmm2
 199 # CHECK-NEXT:  1      5     5.00                        addss   %xmm0, %xmm2
 200 # CHECK-NEXT:  1      5     5.00    *                   addss   (%rax), %xmm2
 201 # CHECK-NEXT:  1      1     0.50                        andnps  %xmm0, %xmm2
 202 # CHECK-NEXT:  1      1     1.00    *                   andnps  (%rax), %xmm2
 203 # CHECK-NEXT:  1      1     0.50                        andps   %xmm0, %xmm2
 204 # CHECK-NEXT:  1      1     1.00    *                   andps   (%rax), %xmm2
 205 # CHECK-NEXT:  1      5     5.00                        cmpeqps %xmm0, %xmm2
 206 # CHECK-NEXT:  1      5     5.00    *                   cmpeqps (%rax), %xmm2
 207 # CHECK-NEXT:  1      5     5.00                        cmpeqss %xmm0, %xmm2
 208 # CHECK-NEXT:  1      5     5.00    *                   cmpeqss (%rax), %xmm2
 209 # CHECK-NEXT:  1      9     4.50                        comiss  %xmm0, %xmm1
 210 # CHECK-NEXT:  1      10    5.00    *                   comiss  (%rax), %xmm1
 211 # CHECK-NEXT:  1      5     5.00                        cvtpi2ps        %mm0, %xmm2
 212 # CHECK-NEXT:  1      5     5.00    *                   cvtpi2ps        (%rax), %xmm2
 213 # CHECK-NEXT:  1      5     5.00                        cvtps2pi        %xmm0, %mm2
 214 # CHECK-NEXT:  1      5     5.00    *                   cvtps2pi        (%rax), %mm2
 215 # CHECK-NEXT:  1      6     3.00                        cvtsi2ss        %ecx, %xmm2
 216 # CHECK-NEXT:  1      6     3.00                        cvtsi2ss        %rcx, %xmm2
 217 # CHECK-NEXT:  1      7     3.50    *                   cvtsi2ssl       (%rax), %xmm2
 218 # CHECK-NEXT:  1      7     3.50    *                   cvtsi2ssl       (%rax), %xmm2
 219 # CHECK-NEXT:  1      8     4.00                        cvtss2si        %xmm0, %ecx
 220 # CHECK-NEXT:  1      9     4.50                        cvtss2si        %xmm0, %rcx
 221 # CHECK-NEXT:  1      9     4.50    *                   cvtss2si        (%rax), %ecx
 222 # CHECK-NEXT:  1      10    5.00    *                   cvtss2si        (%rax), %rcx
 223 # CHECK-NEXT:  1      5     5.00                        cvttps2pi       %xmm0, %mm2
 224 # CHECK-NEXT:  1      5     5.00    *                   cvttps2pi       (%rax), %mm2
 225 # CHECK-NEXT:  1      8     4.00                        cvttss2si       %xmm0, %ecx
 226 # CHECK-NEXT:  1      9     4.50                        cvttss2si       %xmm0, %rcx
 227 # CHECK-NEXT:  1      9     4.50    *                   cvttss2si       (%rax), %ecx
 228 # CHECK-NEXT:  1      10    5.00    *                   cvttss2si       (%rax), %rcx
 229 # CHECK-NEXT:  1      70    35.00                       divps   %xmm0, %xmm2
 230 # CHECK-NEXT:  1      70    35.00   *                   divps   (%rax), %xmm2
 231 # CHECK-NEXT:  1      34    17.00                       divss   %xmm0, %xmm2
 232 # CHECK-NEXT:  1      34    17.00   *                   divss   (%rax), %xmm2
 233 # CHECK-NEXT:  1      5     2.50    *             U     ldmxcsr (%rax)
 234 # CHECK-NEXT:  1      1     1.00    *      *      U     maskmovq        %mm0, %mm1
 235 # CHECK-NEXT:  1      5     5.00                        maxps   %xmm0, %xmm2
 236 # CHECK-NEXT:  1      5     5.00    *                   maxps   (%rax), %xmm2
 237 # CHECK-NEXT:  1      5     5.00                        maxss   %xmm0, %xmm2
 238 # CHECK-NEXT:  1      5     5.00    *                   maxss   (%rax), %xmm2
 239 # CHECK-NEXT:  1      5     5.00                        minps   %xmm0, %xmm2
 240 # CHECK-NEXT:  1      5     5.00    *                   minps   (%rax), %xmm2
 241 # CHECK-NEXT:  1      5     5.00                        minss   %xmm0, %xmm2
 242 # CHECK-NEXT:  1      5     5.00    *                   minss   (%rax), %xmm2
 243 # CHECK-NEXT:  1      1     0.50                        movaps  %xmm0, %xmm2
 244 # CHECK-NEXT:  1      1     1.00           *            movaps  %xmm0, (%rax)
 245 # CHECK-NEXT:  1      1     1.00    *                   movaps  (%rax), %xmm2
 246 # CHECK-NEXT:  1      1     1.00                        movhlps %xmm0, %xmm2
 247 # CHECK-NEXT:  1      1     1.00                        movlhps %xmm0, %xmm2
 248 # CHECK-NEXT:  1      1     1.00           *            movhps  %xmm0, (%rax)
 249 # CHECK-NEXT:  1      1     1.00    *                   movhps  (%rax), %xmm2
 250 # CHECK-NEXT:  1      1     1.00           *            movlps  %xmm0, (%rax)
 251 # CHECK-NEXT:  1      1     1.00    *                   movlps  (%rax), %xmm2
 252 # CHECK-NEXT:  1      3     3.00                        movmskps        %xmm0, %ecx
 253 # CHECK-NEXT:  1      1     1.00           *            movntps %xmm0, (%rax)
 254 # CHECK-NEXT:  1      1     1.00    *      *      U     movntq  %mm0, (%rax)
 255 # CHECK-NEXT:  1      1     0.50                        movss   %xmm0, %xmm2
 256 # CHECK-NEXT:  1      1     1.00           *            movss   %xmm0, (%rax)
 257 # CHECK-NEXT:  1      1     1.00    *                   movss   (%rax), %xmm2
 258 # CHECK-NEXT:  1      1     0.50                        movups  %xmm0, %xmm2
 259 # CHECK-NEXT:  1      2     1.00           *            movups  %xmm0, (%rax)
 260 # CHECK-NEXT:  1      3     1.50    *                   movups  (%rax), %xmm2
 261 # CHECK-NEXT:  1      5     5.00                        mulps   %xmm0, %xmm2
 262 # CHECK-NEXT:  1      5     5.00    *                   mulps   (%rax), %xmm2
 263 # CHECK-NEXT:  1      4     4.00                        mulss   %xmm0, %xmm2
 264 # CHECK-NEXT:  1      4     4.00    *                   mulss   (%rax), %xmm2
 265 # CHECK-NEXT:  1      1     0.50                        orps    %xmm0, %xmm2
 266 # CHECK-NEXT:  1      1     1.00    *                   orps    (%rax), %xmm2
 267 # CHECK-NEXT:  1      1     0.50                        pavgb   %mm0, %mm2
 268 # CHECK-NEXT:  1      1     1.00    *                   pavgb   (%rax), %mm2
 269 # CHECK-NEXT:  1      1     0.50                        pavgw   %mm0, %mm2
 270 # CHECK-NEXT:  1      1     1.00    *                   pavgw   (%rax), %mm2
 271 # CHECK-NEXT:  1      4     2.00                        pextrw  $1, %mm0, %ecx
 272 # CHECK-NEXT:  1      1     1.00                        pinsrw  $1, %eax, %mm2
 273 # CHECK-NEXT:  1      1     1.00    *                   pinsrw  $1, (%rax), %mm2
 274 # CHECK-NEXT:  1      1     0.50                        pmaxsw  %mm0, %mm2
 275 # CHECK-NEXT:  1      1     1.00    *                   pmaxsw  (%rax), %mm2
 276 # CHECK-NEXT:  1      1     0.50                        pmaxub  %mm0, %mm2
 277 # CHECK-NEXT:  1      1     1.00    *                   pmaxub  (%rax), %mm2
 278 # CHECK-NEXT:  1      1     0.50                        pminsw  %mm0, %mm2
 279 # CHECK-NEXT:  1      1     1.00    *                   pminsw  (%rax), %mm2
 280 # CHECK-NEXT:  1      1     0.50                        pminub  %mm0, %mm2
 281 # CHECK-NEXT:  1      1     1.00    *                   pminub  (%rax), %mm2
 282 # CHECK-NEXT:  1      3     3.00                        pmovmskb        %mm0, %ecx
 283 # CHECK-NEXT:  1      4     4.00                        pmulhuw %mm0, %mm2
 284 # CHECK-NEXT:  1      4     4.00    *                   pmulhuw (%rax), %mm2
 285 # CHECK-NEXT:  1      1     1.00    *      *            prefetcht0      (%rax)
 286 # CHECK-NEXT:  1      1     1.00    *      *            prefetcht1      (%rax)
 287 # CHECK-NEXT:  1      1     1.00    *      *            prefetcht2      (%rax)
 288 # CHECK-NEXT:  1      1     1.00    *      *            prefetchnta     (%rax)
 289 # CHECK-NEXT:  1      4     2.00                        psadbw  %mm0, %mm2
 290 # CHECK-NEXT:  1      4     2.00    *                   psadbw  (%rax), %mm2
 291 # CHECK-NEXT:  1      1     1.00                        pshufw  $1, %mm0, %mm2
 292 # CHECK-NEXT:  1      1     1.00    *                   pshufw  $1, (%rax), %mm2
 293 # CHECK-NEXT:  1      9     4.50                        rcpps   %xmm0, %xmm2
 294 # CHECK-NEXT:  1      10    5.00    *                   rcpps   (%rax), %xmm2
 295 # CHECK-NEXT:  1      4     4.00                        rcpss   %xmm0, %xmm2
 296 # CHECK-NEXT:  1      4     4.00    *                   rcpss   (%rax), %xmm2
 297 # CHECK-NEXT:  1      9     4.50                        rsqrtps %xmm0, %xmm2
 298 # CHECK-NEXT:  1      10    5.00    *                   rsqrtps (%rax), %xmm2
 299 # CHECK-NEXT:  1      4     4.00                        rsqrtss %xmm0, %xmm2
 300 # CHECK-NEXT:  1      4     4.00    *                   rsqrtss (%rax), %xmm2
 301 # CHECK-NEXT:  1      1     1.00    *      *      U     sfence
 302 # CHECK-NEXT:  1      1     1.00                        shufps  $1, %xmm0, %xmm2
 303 # CHECK-NEXT:  1      1     1.00    *                   shufps  $1, (%rax), %xmm2
 304 # CHECK-NEXT:  1      70    35.00                       sqrtps  %xmm0, %xmm2
 305 # CHECK-NEXT:  1      70    35.00   *                   sqrtps  (%rax), %xmm2
 306 # CHECK-NEXT:  1      34    17.00                       sqrtss  %xmm0, %xmm2
 307 # CHECK-NEXT:  1      34    17.00   *                   sqrtss  (%rax), %xmm2
 308 # CHECK-NEXT:  1      15    7.50           *      U     stmxcsr (%rax)
 309 # CHECK-NEXT:  1      5     5.00                        subps   %xmm0, %xmm2
 310 # CHECK-NEXT:  1      5     5.00    *                   subps   (%rax), %xmm2
 311 # CHECK-NEXT:  1      5     5.00                        subss   %xmm0, %xmm2
 312 # CHECK-NEXT:  1      5     5.00    *                   subss   (%rax), %xmm2
 313 # CHECK-NEXT:  1      9     4.50                        ucomiss %xmm0, %xmm1
 314 # CHECK-NEXT:  1      10    5.00    *                   ucomiss (%rax), %xmm1
 315 # CHECK-NEXT:  1      1     1.00                        unpckhps        %xmm0, %xmm2
 316 # CHECK-NEXT:  1      1     1.00    *                   unpckhps        (%rax), %xmm2
 317 # CHECK-NEXT:  1      1     1.00                        unpcklps        %xmm0, %xmm2
 318 # CHECK-NEXT:  1      1     1.00    *                   unpcklps        (%rax), %xmm2
 319 # CHECK-NEXT:  1      1     0.50                        xorps   %xmm0, %xmm2
 320 # CHECK-NEXT:  1      1     1.00    *                   xorps   (%rax), %xmm2
 321
 322 # CHECK:      Resources:
 323 # CHECK-NEXT: [0]   - AtomPort0
 324 # CHECK-NEXT: [1]   - AtomPort1
 325
 326 # CHECK:      Resource pressure per iteration:
 327 # CHECK-NEXT: [0]    [1]
 328 # CHECK-NEXT: 516.00 350.00
 329
 330 # CHECK:      Resource pressure by instruction:
 331 # CHECK-NEXT: [0]    [1]    Instructions:
 332 # CHECK-NEXT: 5.00    -     addps       %xmm0, %xmm2
 333 # CHECK-NEXT: 5.00    -     addps       (%rax), %xmm2
 334 # CHECK-NEXT: 5.00    -     addss       %xmm0, %xmm2
 335 # CHECK-NEXT: 5.00    -     addss       (%rax), %xmm2
 336 # CHECK-NEXT: 0.50   0.50   andnps      %xmm0, %xmm2
 337 # CHECK-NEXT: 1.00    -     andnps      (%rax), %xmm2
 338 # CHECK-NEXT: 0.50   0.50   andps       %xmm0, %xmm2
 339 # CHECK-NEXT: 1.00    -     andps       (%rax), %xmm2
 340 # CHECK-NEXT: 5.00    -     cmpeqps     %xmm0, %xmm2
 341 # CHECK-NEXT: 5.00    -     cmpeqps     (%rax), %xmm2
 342 # CHECK-NEXT: 5.00    -     cmpeqss     %xmm0, %xmm2
 343 # CHECK-NEXT: 5.00    -     cmpeqss     (%rax), %xmm2
 344 # CHECK-NEXT: 4.50   4.50   comiss      %xmm0, %xmm1
 345 # CHECK-NEXT: 5.00   5.00   comiss      (%rax), %xmm1
 346 # CHECK-NEXT:  -     5.00   cvtpi2ps    %mm0, %xmm2
 347 # CHECK-NEXT:  -     5.00   cvtpi2ps    (%rax), %xmm2
 348 # CHECK-NEXT:  -     5.00   cvtps2pi    %xmm0, %mm2
 349 # CHECK-NEXT: 5.00   5.00   cvtps2pi    (%rax), %mm2
 350 # CHECK-NEXT: 3.00   3.00   cvtsi2ss    %ecx, %xmm2
 351 # CHECK-NEXT: 3.00   3.00   cvtsi2ss    %rcx, %xmm2
 352 # CHECK-NEXT: 3.50   3.50   cvtsi2ssl   (%rax), %xmm2
 353 # CHECK-NEXT: 3.50   3.50   cvtsi2ssl   (%rax), %xmm2
 354 # CHECK-NEXT: 4.00   4.00   cvtss2si    %xmm0, %ecx
 355 # CHECK-NEXT: 4.50   4.50   cvtss2si    %xmm0, %rcx
 356 # CHECK-NEXT: 4.50   4.50   cvtss2si    (%rax), %ecx
 357 # CHECK-NEXT: 5.00   5.00   cvtss2si    (%rax), %rcx
 358 # CHECK-NEXT:  -     5.00   cvttps2pi   %xmm0, %mm2
 359 # CHECK-NEXT: 5.00   5.00   cvttps2pi   (%rax), %mm2
 360 # CHECK-NEXT: 4.00   4.00   cvttss2si   %xmm0, %ecx
 361 # CHECK-NEXT: 4.50   4.50   cvttss2si   %xmm0, %rcx
 362 # CHECK-NEXT: 4.50   4.50   cvttss2si   (%rax), %ecx
 363 # CHECK-NEXT: 5.00   5.00   cvttss2si   (%rax), %rcx
 364 # CHECK-NEXT: 35.00  35.00  divps       %xmm0, %xmm2
 365 # CHECK-NEXT: 35.00  35.00  divps       (%rax), %xmm2
 366 # CHECK-NEXT: 17.00  17.00  divss       %xmm0, %xmm2
 367 # CHECK-NEXT: 17.00  17.00  divss       (%rax), %xmm2
 368 # CHECK-NEXT: 2.50   2.50   ldmxcsr     (%rax)
 369 # CHECK-NEXT: 1.00    -     maskmovq    %mm0, %mm1
 370 # CHECK-NEXT: 5.00    -     maxps       %xmm0, %xmm2
 371 # CHECK-NEXT: 5.00    -     maxps       (%rax), %xmm2
 372 # CHECK-NEXT: 5.00    -     maxss       %xmm0, %xmm2
 373 # CHECK-NEXT: 5.00    -     maxss       (%rax), %xmm2
 374 # CHECK-NEXT: 5.00    -     minps       %xmm0, %xmm2
 375 # CHECK-NEXT: 5.00    -     minps       (%rax), %xmm2
 376 # CHECK-NEXT: 5.00    -     minss       %xmm0, %xmm2
 377 # CHECK-NEXT: 5.00    -     minss       (%rax), %xmm2
 378 # CHECK-NEXT: 0.50   0.50   movaps      %xmm0, %xmm2
 379 # CHECK-NEXT: 1.00    -     movaps      %xmm0, (%rax)
 380 # CHECK-NEXT: 1.00    -     movaps      (%rax), %xmm2
 381 # CHECK-NEXT: 1.00    -     movhlps     %xmm0, %xmm2
 382 # CHECK-NEXT: 1.00    -     movlhps     %xmm0, %xmm2
 383 # CHECK-NEXT: 1.00    -     movhps      %xmm0, (%rax)
 384 # CHECK-NEXT: 1.00    -     movhps      (%rax), %xmm2
 385 # CHECK-NEXT: 1.00    -     movlps      %xmm0, (%rax)
 386 # CHECK-NEXT: 1.00    -     movlps      (%rax), %xmm2
 387 # CHECK-NEXT: 3.00    -     movmskps    %xmm0, %ecx
 388 # CHECK-NEXT: 1.00    -     movntps     %xmm0, (%rax)
 389 # CHECK-NEXT: 1.00    -     movntq      %mm0, (%rax)
 390 # CHECK-NEXT: 0.50   0.50   movss       %xmm0, %xmm2
 391 # CHECK-NEXT: 1.00    -     movss       %xmm0, (%rax)
 392 # CHECK-NEXT: 1.00    -     movss       (%rax), %xmm2
 393 # CHECK-NEXT: 0.50   0.50   movups      %xmm0, %xmm2
 394 # CHECK-NEXT: 1.00   1.00   movups      %xmm0, (%rax)
 395 # CHECK-NEXT: 1.50   1.50   movups      (%rax), %xmm2
 396 # CHECK-NEXT: 5.00    -     mulps       %xmm0, %xmm2
 397 # CHECK-NEXT: 5.00    -     mulps       (%rax), %xmm2
 398 # CHECK-NEXT: 4.00    -     mulss       %xmm0, %xmm2
 399 # CHECK-NEXT: 4.00    -     mulss       (%rax), %xmm2
 400 # CHECK-NEXT: 0.50   0.50   orps        %xmm0, %xmm2
 401 # CHECK-NEXT: 1.00    -     orps        (%rax), %xmm2
 402 # CHECK-NEXT: 0.50   0.50   pavgb       %mm0, %mm2
 403 # CHECK-NEXT: 1.00    -     pavgb       (%rax), %mm2
 404 # CHECK-NEXT: 0.50   0.50   pavgw       %mm0, %mm2
 405 # CHECK-NEXT: 1.00    -     pavgw       (%rax), %mm2
 406 # CHECK-NEXT: 2.00   2.00   pextrw      $1, %mm0, %ecx
 407 # CHECK-NEXT: 1.00    -     pinsrw      $1, %eax, %mm2
 408 # CHECK-NEXT: 1.00    -     pinsrw      $1, (%rax), %mm2
 409 # CHECK-NEXT: 0.50   0.50   pmaxsw      %mm0, %mm2
 410 # CHECK-NEXT: 1.00    -     pmaxsw      (%rax), %mm2
 411 # CHECK-NEXT: 0.50   0.50   pmaxub      %mm0, %mm2
 412 # CHECK-NEXT: 1.00    -     pmaxub      (%rax), %mm2
 413 # CHECK-NEXT: 0.50   0.50   pminsw      %mm0, %mm2
 414 # CHECK-NEXT: 1.00    -     pminsw      (%rax), %mm2
 415 # CHECK-NEXT: 0.50   0.50   pminub      %mm0, %mm2
 416 # CHECK-NEXT: 1.00    -     pminub      (%rax), %mm2
 417 # CHECK-NEXT: 3.00    -     pmovmskb    %mm0, %ecx
 418 # CHECK-NEXT: 4.00    -     pmulhuw     %mm0, %mm2
 419 # CHECK-NEXT: 4.00    -     pmulhuw     (%rax), %mm2
 420 # CHECK-NEXT: 1.00    -     prefetcht0  (%rax)
 421 # CHECK-NEXT: 1.00    -     prefetcht1  (%rax)
 422 # CHECK-NEXT: 1.00    -     prefetcht2  (%rax)
 423 # CHECK-NEXT: 1.00    -     prefetchnta (%rax)
 424 # CHECK-NEXT: 2.00   2.00   psadbw      %mm0, %mm2
 425 # CHECK-NEXT: 2.00   2.00   psadbw      (%rax), %mm2
 426 # CHECK-NEXT: 1.00    -     pshufw      $1, %mm0, %mm2
 427 # CHECK-NEXT: 1.00    -     pshufw      $1, (%rax), %mm2
 428 # CHECK-NEXT: 4.50   4.50   rcpps       %xmm0, %xmm2
 429 # CHECK-NEXT: 5.00   5.00   rcpps       (%rax), %xmm2
 430 # CHECK-NEXT: 4.00    -     rcpss       %xmm0, %xmm2
 431 # CHECK-NEXT: 4.00    -     rcpss       (%rax), %xmm2
 432 # CHECK-NEXT: 4.50   4.50   rsqrtps     %xmm0, %xmm2
 433 # CHECK-NEXT: 5.00   5.00   rsqrtps     (%rax), %xmm2
 434 # CHECK-NEXT: 4.00    -     rsqrtss     %xmm0, %xmm2
 435 # CHECK-NEXT: 4.00    -     rsqrtss     (%rax), %xmm2
 436 # CHECK-NEXT: 1.00    -     sfence
 437 # CHECK-NEXT: 1.00    -     shufps      $1, %xmm0, %xmm2
 438 # CHECK-NEXT: 1.00    -     shufps      $1, (%rax), %xmm2
 439 # CHECK-NEXT: 35.00  35.00  sqrtps      %xmm0, %xmm2
 440 # CHECK-NEXT: 35.00  35.00  sqrtps      (%rax), %xmm2
 441 # CHECK-NEXT: 17.00  17.00  sqrtss      %xmm0, %xmm2
 442 # CHECK-NEXT: 17.00  17.00  sqrtss      (%rax), %xmm2
 443 # CHECK-NEXT: 7.50   7.50   stmxcsr     (%rax)
 444 # CHECK-NEXT: 5.00    -     subps       %xmm0, %xmm2
 445 # CHECK-NEXT: 5.00    -     subps       (%rax), %xmm2
 446 # CHECK-NEXT: 5.00    -     subss       %xmm0, %xmm2
 447 # CHECK-NEXT: 5.00    -     subss       (%rax), %xmm2
 448 # CHECK-NEXT: 4.50   4.50   ucomiss     %xmm0, %xmm1
 449 # CHECK-NEXT: 5.00   5.00   ucomiss     (%rax), %xmm1
 450 # CHECK-NEXT: 1.00    -     unpckhps    %xmm0, %xmm2
 451 # CHECK-NEXT: 1.00    -     unpckhps    (%rax), %xmm2
 452 # CHECK-NEXT: 1.00    -     unpcklps    %xmm0, %xmm2
 453 # CHECK-NEXT: 1.00    -     unpcklps    (%rax), %xmm2
 454 # CHECK-NEXT: 0.50   0.50   xorps       %xmm0, %xmm2
 455 # CHECK-NEXT: 1.00    -     xorps       (%rax), %xmm2