1 # NOTE: Assertions have been autogenerated by utils/update_mca_test_checks.py
2 # RUN: llvm-mca -mtriple=x86_64-unknown-unknown -mcpu=slm -instruction-tables < %s | FileCheck %s
138 punpckhbw
(%rax
), %mm2
141 punpckhdq
(%rax
), %mm2
144 punpckhwd
(%rax
), %mm2
147 punpcklbw
(%rax
), %mm2
150 punpckldq
(%rax
), %mm2
153 punpcklwd
(%rax
), %mm2
158 # CHECK: Instruction Info:
159 # CHECK-NEXT: [1]: #uOps
160 # CHECK-NEXT: [2]: Latency
161 # CHECK-NEXT: [3]: RThroughput
162 # CHECK-NEXT: [4]: MayLoad
163 # CHECK-NEXT: [5]: MayStore
164 # CHECK-NEXT: [6]: HasSideEffects (U)
166 # CHECK: [1] [2] [3] [4] [5] [6] Instructions:
167 # CHECK-NEXT: 9 10 5.00 * * U emms
168 # CHECK-NEXT: 1 1 0.50 movd %eax, %mm2
169 # CHECK-NEXT: 1 3 1.00 * movd (%rax), %mm2
170 # CHECK-NEXT: 1 1 0.50 movd %mm0, %ecx
171 # CHECK-NEXT: 1 1 1.00 * U movd %mm0, (%rax)
172 # CHECK-NEXT: 1 1 0.50 movq %rax, %mm2
173 # CHECK-NEXT: 1 3 1.00 * movq (%rax), %mm2
174 # CHECK-NEXT: 1 1 0.50 movq %mm0, %rcx
175 # CHECK-NEXT: 1 1 1.00 * movq %mm0, (%rax)
176 # CHECK-NEXT: 1 1 1.00 packsswb %mm0, %mm2
177 # CHECK-NEXT: 1 4 1.00 * packsswb (%rax), %mm2
178 # CHECK-NEXT: 1 1 1.00 packssdw %mm0, %mm2
179 # CHECK-NEXT: 1 4 1.00 * packssdw (%rax), %mm2
180 # CHECK-NEXT: 1 1 1.00 packuswb %mm0, %mm2
181 # CHECK-NEXT: 1 4 1.00 * packuswb (%rax), %mm2
182 # CHECK-NEXT: 1 1 0.50 paddb %mm0, %mm2
183 # CHECK-NEXT: 1 4 1.00 * paddb (%rax), %mm2
184 # CHECK-NEXT: 1 1 0.50 paddd %mm0, %mm2
185 # CHECK-NEXT: 1 4 1.00 * paddd (%rax), %mm2
186 # CHECK-NEXT: 1 1 0.50 paddsb %mm0, %mm2
187 # CHECK-NEXT: 1 4 1.00 * paddsb (%rax), %mm2
188 # CHECK-NEXT: 1 1 0.50 paddsw %mm0, %mm2
189 # CHECK-NEXT: 1 4 1.00 * paddsw (%rax), %mm2
190 # CHECK-NEXT: 1 1 0.50 paddusb %mm0, %mm2
191 # CHECK-NEXT: 1 4 1.00 * paddusb (%rax), %mm2
192 # CHECK-NEXT: 1 1 0.50 paddusw %mm0, %mm2
193 # CHECK-NEXT: 1 4 1.00 * paddusw (%rax), %mm2
194 # CHECK-NEXT: 1 1 0.50 paddw %mm0, %mm2
195 # CHECK-NEXT: 1 4 1.00 * paddw (%rax), %mm2
196 # CHECK-NEXT: 1 1 0.50 pand %mm0, %mm2
197 # CHECK-NEXT: 1 4 1.00 * pand (%rax), %mm2
198 # CHECK-NEXT: 1 1 0.50 pandn %mm0, %mm2
199 # CHECK-NEXT: 1 4 1.00 * pandn (%rax), %mm2
200 # CHECK-NEXT: 1 1 0.50 pcmpeqb %mm0, %mm2
201 # CHECK-NEXT: 1 4 1.00 * pcmpeqb (%rax), %mm2
202 # CHECK-NEXT: 1 1 0.50 pcmpeqd %mm0, %mm2
203 # CHECK-NEXT: 1 4 1.00 * pcmpeqd (%rax), %mm2
204 # CHECK-NEXT: 1 1 0.50 pcmpeqw %mm0, %mm2
205 # CHECK-NEXT: 1 4 1.00 * pcmpeqw (%rax), %mm2
206 # CHECK-NEXT: 1 1 0.50 pcmpgtb %mm0, %mm2
207 # CHECK-NEXT: 1 4 1.00 * pcmpgtb (%rax), %mm2
208 # CHECK-NEXT: 1 1 0.50 pcmpgtd %mm0, %mm2
209 # CHECK-NEXT: 1 4 1.00 * pcmpgtd (%rax), %mm2
210 # CHECK-NEXT: 1 1 0.50 pcmpgtw %mm0, %mm2
211 # CHECK-NEXT: 1 4 1.00 * pcmpgtw (%rax), %mm2
212 # CHECK-NEXT: 1 4 1.00 pmaddwd %mm0, %mm2
213 # CHECK-NEXT: 1 7 1.00 * pmaddwd (%rax), %mm2
214 # CHECK-NEXT: 1 4 1.00 pmulhw %mm0, %mm2
215 # CHECK-NEXT: 1 7 1.00 * pmulhw (%rax), %mm2
216 # CHECK-NEXT: 1 4 1.00 pmullw %mm0, %mm2
217 # CHECK-NEXT: 1 7 1.00 * pmullw (%rax), %mm2
218 # CHECK-NEXT: 1 1 0.50 por %mm0, %mm2
219 # CHECK-NEXT: 1 4 1.00 * por (%rax), %mm2
220 # CHECK-NEXT: 1 1 1.00 pslld $1, %mm2
221 # CHECK-NEXT: 1 1 1.00 pslld %mm0, %mm2
222 # CHECK-NEXT: 1 4 1.00 * pslld (%rax), %mm2
223 # CHECK-NEXT: 1 1 1.00 psllq $1, %mm2
224 # CHECK-NEXT: 1 1 1.00 psllq %mm0, %mm2
225 # CHECK-NEXT: 1 4 1.00 * psllq (%rax), %mm2
226 # CHECK-NEXT: 1 1 1.00 psllw $1, %mm2
227 # CHECK-NEXT: 1 1 1.00 psllw %mm0, %mm2
228 # CHECK-NEXT: 1 4 1.00 * psllw (%rax), %mm2
229 # CHECK-NEXT: 1 1 1.00 psrad $1, %mm2
230 # CHECK-NEXT: 1 1 1.00 psrad %mm0, %mm2
231 # CHECK-NEXT: 1 4 1.00 * psrad (%rax), %mm2
232 # CHECK-NEXT: 1 1 1.00 psraw $1, %mm2
233 # CHECK-NEXT: 1 1 1.00 psraw %mm0, %mm2
234 # CHECK-NEXT: 1 4 1.00 * psraw (%rax), %mm2
235 # CHECK-NEXT: 1 1 1.00 psrld $1, %mm2
236 # CHECK-NEXT: 1 1 1.00 psrld %mm0, %mm2
237 # CHECK-NEXT: 1 4 1.00 * psrld (%rax), %mm2
238 # CHECK-NEXT: 1 1 1.00 psrlq $1, %mm2
239 # CHECK-NEXT: 1 1 1.00 psrlq %mm0, %mm2
240 # CHECK-NEXT: 1 4 1.00 * psrlq (%rax), %mm2
241 # CHECK-NEXT: 1 1 1.00 psrlw $1, %mm2
242 # CHECK-NEXT: 1 1 1.00 psrlw %mm0, %mm2
243 # CHECK-NEXT: 1 4 1.00 * psrlw (%rax), %mm2
244 # CHECK-NEXT: 1 1 0.50 psubb %mm0, %mm2
245 # CHECK-NEXT: 1 4 1.00 * psubb (%rax), %mm2
246 # CHECK-NEXT: 1 1 0.50 psubd %mm0, %mm2
247 # CHECK-NEXT: 1 4 1.00 * psubd (%rax), %mm2
248 # CHECK-NEXT: 1 1 0.50 psubsb %mm0, %mm2
249 # CHECK-NEXT: 1 4 1.00 * psubsb (%rax), %mm2
250 # CHECK-NEXT: 1 1 0.50 psubsw %mm0, %mm2
251 # CHECK-NEXT: 1 4 1.00 * psubsw (%rax), %mm2
252 # CHECK-NEXT: 1 1 0.50 psubusb %mm0, %mm2
253 # CHECK-NEXT: 1 4 1.00 * psubusb (%rax), %mm2
254 # CHECK-NEXT: 1 1 0.50 psubusw %mm0, %mm2
255 # CHECK-NEXT: 1 4 1.00 * psubusw (%rax), %mm2
256 # CHECK-NEXT: 1 1 0.50 psubw %mm0, %mm2
257 # CHECK-NEXT: 1 4 1.00 * psubw (%rax), %mm2
258 # CHECK-NEXT: 1 1 1.00 punpckhbw %mm0, %mm2
259 # CHECK-NEXT: 1 4 1.00 * punpckhbw (%rax), %mm2
260 # CHECK-NEXT: 1 1 1.00 punpckhdq %mm0, %mm2
261 # CHECK-NEXT: 1 4 1.00 * punpckhdq (%rax), %mm2
262 # CHECK-NEXT: 1 1 1.00 punpckhwd %mm0, %mm2
263 # CHECK-NEXT: 1 4 1.00 * punpckhwd (%rax), %mm2
264 # CHECK-NEXT: 1 1 1.00 punpcklbw %mm0, %mm2
265 # CHECK-NEXT: 1 4 1.00 * punpcklbw (%rax), %mm2
266 # CHECK-NEXT: 1 1 1.00 punpckldq %mm0, %mm2
267 # CHECK-NEXT: 1 4 1.00 * punpckldq (%rax), %mm2
268 # CHECK-NEXT: 1 1 1.00 punpcklwd %mm0, %mm2
269 # CHECK-NEXT: 1 4 1.00 * punpcklwd (%rax), %mm2
270 # CHECK-NEXT: 1 1 0.50 pxor %mm0, %mm2
271 # CHECK-NEXT: 1 4 1.00 * pxor (%rax), %mm2
274 # CHECK-NEXT: [0] - SLMDivider
275 # CHECK-NEXT: [1] - SLMFPDivider
276 # CHECK-NEXT: [2] - SLMFPMultiplier
277 # CHECK-NEXT: [3] - SLM_FPC_RSV0
278 # CHECK-NEXT: [4] - SLM_FPC_RSV1
279 # CHECK-NEXT: [5] - SLM_IEC_RSV0
280 # CHECK-NEXT: [6] - SLM_IEC_RSV1
281 # CHECK-NEXT: [7] - SLM_MEC_RSV
283 # CHECK: Resource pressure per iteration:
284 # CHECK-NEXT: [0] [1] [2] [3] [4] [5] [6] [7]
285 # CHECK-NEXT: - - - 77.00 29.00 2.00 2.00 48.00
287 # CHECK: Resource pressure by instruction:
288 # CHECK-NEXT: [0] [1] [2] [3] [4] [5] [6] [7] Instructions:
289 # CHECK-NEXT: - - - 5.00 5.00 - - - emms
290 # CHECK-NEXT: - - - - - 0.50 0.50 - movd %eax, %mm2
291 # CHECK-NEXT: - - - - - - - 1.00 movd (%rax), %mm2
292 # CHECK-NEXT: - - - - - 0.50 0.50 - movd %mm0, %ecx
293 # CHECK-NEXT: - - - - - - - 1.00 movd %mm0, (%rax)
294 # CHECK-NEXT: - - - - - 0.50 0.50 - movq %rax, %mm2
295 # CHECK-NEXT: - - - - - - - 1.00 movq (%rax), %mm2
296 # CHECK-NEXT: - - - - - 0.50 0.50 - movq %mm0, %rcx
297 # CHECK-NEXT: - - - - - - - 1.00 movq %mm0, (%rax)
298 # CHECK-NEXT: - - - 1.00 - - - - packsswb %mm0, %mm2
299 # CHECK-NEXT: - - - 1.00 - - - 1.00 packsswb (%rax), %mm2
300 # CHECK-NEXT: - - - 1.00 - - - - packssdw %mm0, %mm2
301 # CHECK-NEXT: - - - 1.00 - - - 1.00 packssdw (%rax), %mm2
302 # CHECK-NEXT: - - - 1.00 - - - - packuswb %mm0, %mm2
303 # CHECK-NEXT: - - - 1.00 - - - 1.00 packuswb (%rax), %mm2
304 # CHECK-NEXT: - - - 0.50 0.50 - - - paddb %mm0, %mm2
305 # CHECK-NEXT: - - - 0.50 0.50 - - 1.00 paddb (%rax), %mm2
306 # CHECK-NEXT: - - - 0.50 0.50 - - - paddd %mm0, %mm2
307 # CHECK-NEXT: - - - 0.50 0.50 - - 1.00 paddd (%rax), %mm2
308 # CHECK-NEXT: - - - 0.50 0.50 - - - paddsb %mm0, %mm2
309 # CHECK-NEXT: - - - 0.50 0.50 - - 1.00 paddsb (%rax), %mm2
310 # CHECK-NEXT: - - - 0.50 0.50 - - - paddsw %mm0, %mm2
311 # CHECK-NEXT: - - - 0.50 0.50 - - 1.00 paddsw (%rax), %mm2
312 # CHECK-NEXT: - - - 0.50 0.50 - - - paddusb %mm0, %mm2
313 # CHECK-NEXT: - - - 0.50 0.50 - - 1.00 paddusb (%rax), %mm2
314 # CHECK-NEXT: - - - 0.50 0.50 - - - paddusw %mm0, %mm2
315 # CHECK-NEXT: - - - 0.50 0.50 - - 1.00 paddusw (%rax), %mm2
316 # CHECK-NEXT: - - - 0.50 0.50 - - - paddw %mm0, %mm2
317 # CHECK-NEXT: - - - 0.50 0.50 - - 1.00 paddw (%rax), %mm2
318 # CHECK-NEXT: - - - 0.50 0.50 - - - pand %mm0, %mm2
319 # CHECK-NEXT: - - - 0.50 0.50 - - 1.00 pand (%rax), %mm2
320 # CHECK-NEXT: - - - 0.50 0.50 - - - pandn %mm0, %mm2
321 # CHECK-NEXT: - - - 0.50 0.50 - - 1.00 pandn (%rax), %mm2
322 # CHECK-NEXT: - - - 0.50 0.50 - - - pcmpeqb %mm0, %mm2
323 # CHECK-NEXT: - - - 0.50 0.50 - - 1.00 pcmpeqb (%rax), %mm2
324 # CHECK-NEXT: - - - 0.50 0.50 - - - pcmpeqd %mm0, %mm2
325 # CHECK-NEXT: - - - 0.50 0.50 - - 1.00 pcmpeqd (%rax), %mm2
326 # CHECK-NEXT: - - - 0.50 0.50 - - - pcmpeqw %mm0, %mm2
327 # CHECK-NEXT: - - - 0.50 0.50 - - 1.00 pcmpeqw (%rax), %mm2
328 # CHECK-NEXT: - - - 0.50 0.50 - - - pcmpgtb %mm0, %mm2
329 # CHECK-NEXT: - - - 0.50 0.50 - - 1.00 pcmpgtb (%rax), %mm2
330 # CHECK-NEXT: - - - 0.50 0.50 - - - pcmpgtd %mm0, %mm2
331 # CHECK-NEXT: - - - 0.50 0.50 - - 1.00 pcmpgtd (%rax), %mm2
332 # CHECK-NEXT: - - - 0.50 0.50 - - - pcmpgtw %mm0, %mm2
333 # CHECK-NEXT: - - - 0.50 0.50 - - 1.00 pcmpgtw (%rax), %mm2
334 # CHECK-NEXT: - - - 1.00 - - - - pmaddwd %mm0, %mm2
335 # CHECK-NEXT: - - - 1.00 - - - 1.00 pmaddwd (%rax), %mm2
336 # CHECK-NEXT: - - - 1.00 - - - - pmulhw %mm0, %mm2
337 # CHECK-NEXT: - - - 1.00 - - - 1.00 pmulhw (%rax), %mm2
338 # CHECK-NEXT: - - - 1.00 - - - - pmullw %mm0, %mm2
339 # CHECK-NEXT: - - - 1.00 - - - 1.00 pmullw (%rax), %mm2
340 # CHECK-NEXT: - - - 0.50 0.50 - - - por %mm0, %mm2
341 # CHECK-NEXT: - - - 0.50 0.50 - - 1.00 por (%rax), %mm2
342 # CHECK-NEXT: - - - 1.00 - - - - pslld $1, %mm2
343 # CHECK-NEXT: - - - 1.00 - - - - pslld %mm0, %mm2
344 # CHECK-NEXT: - - - 1.00 - - - 1.00 pslld (%rax), %mm2
345 # CHECK-NEXT: - - - 1.00 - - - - psllq $1, %mm2
346 # CHECK-NEXT: - - - 1.00 - - - - psllq %mm0, %mm2
347 # CHECK-NEXT: - - - 1.00 - - - 1.00 psllq (%rax), %mm2
348 # CHECK-NEXT: - - - 1.00 - - - - psllw $1, %mm2
349 # CHECK-NEXT: - - - 1.00 - - - - psllw %mm0, %mm2
350 # CHECK-NEXT: - - - 1.00 - - - 1.00 psllw (%rax), %mm2
351 # CHECK-NEXT: - - - 1.00 - - - - psrad $1, %mm2
352 # CHECK-NEXT: - - - 1.00 - - - - psrad %mm0, %mm2
353 # CHECK-NEXT: - - - 1.00 - - - 1.00 psrad (%rax), %mm2
354 # CHECK-NEXT: - - - 1.00 - - - - psraw $1, %mm2
355 # CHECK-NEXT: - - - 1.00 - - - - psraw %mm0, %mm2
356 # CHECK-NEXT: - - - 1.00 - - - 1.00 psraw (%rax), %mm2
357 # CHECK-NEXT: - - - 1.00 - - - - psrld $1, %mm2
358 # CHECK-NEXT: - - - 1.00 - - - - psrld %mm0, %mm2
359 # CHECK-NEXT: - - - 1.00 - - - 1.00 psrld (%rax), %mm2
360 # CHECK-NEXT: - - - 1.00 - - - - psrlq $1, %mm2
361 # CHECK-NEXT: - - - 1.00 - - - - psrlq %mm0, %mm2
362 # CHECK-NEXT: - - - 1.00 - - - 1.00 psrlq (%rax), %mm2
363 # CHECK-NEXT: - - - 1.00 - - - - psrlw $1, %mm2
364 # CHECK-NEXT: - - - 1.00 - - - - psrlw %mm0, %mm2
365 # CHECK-NEXT: - - - 1.00 - - - 1.00 psrlw (%rax), %mm2
366 # CHECK-NEXT: - - - 0.50 0.50 - - - psubb %mm0, %mm2
367 # CHECK-NEXT: - - - 0.50 0.50 - - 1.00 psubb (%rax), %mm2
368 # CHECK-NEXT: - - - 0.50 0.50 - - - psubd %mm0, %mm2
369 # CHECK-NEXT: - - - 0.50 0.50 - - 1.00 psubd (%rax), %mm2
370 # CHECK-NEXT: - - - 0.50 0.50 - - - psubsb %mm0, %mm2
371 # CHECK-NEXT: - - - 0.50 0.50 - - 1.00 psubsb (%rax), %mm2
372 # CHECK-NEXT: - - - 0.50 0.50 - - - psubsw %mm0, %mm2
373 # CHECK-NEXT: - - - 0.50 0.50 - - 1.00 psubsw (%rax), %mm2
374 # CHECK-NEXT: - - - 0.50 0.50 - - - psubusb %mm0, %mm2
375 # CHECK-NEXT: - - - 0.50 0.50 - - 1.00 psubusb (%rax), %mm2
376 # CHECK-NEXT: - - - 0.50 0.50 - - - psubusw %mm0, %mm2
377 # CHECK-NEXT: - - - 0.50 0.50 - - 1.00 psubusw (%rax), %mm2
378 # CHECK-NEXT: - - - 0.50 0.50 - - - psubw %mm0, %mm2
379 # CHECK-NEXT: - - - 0.50 0.50 - - 1.00 psubw (%rax), %mm2
380 # CHECK-NEXT: - - - 1.00 - - - - punpckhbw %mm0, %mm2
381 # CHECK-NEXT: - - - 1.00 - - - 1.00 punpckhbw (%rax), %mm2
382 # CHECK-NEXT: - - - 1.00 - - - - punpckhdq %mm0, %mm2
383 # CHECK-NEXT: - - - 1.00 - - - 1.00 punpckhdq (%rax), %mm2
384 # CHECK-NEXT: - - - 1.00 - - - - punpckhwd %mm0, %mm2
385 # CHECK-NEXT: - - - 1.00 - - - 1.00 punpckhwd (%rax), %mm2
386 # CHECK-NEXT: - - - 1.00 - - - - punpcklbw %mm0, %mm2
387 # CHECK-NEXT: - - - 1.00 - - - 1.00 punpcklbw (%rax), %mm2
388 # CHECK-NEXT: - - - 1.00 - - - - punpckldq %mm0, %mm2
389 # CHECK-NEXT: - - - 1.00 - - - 1.00 punpckldq (%rax), %mm2
390 # CHECK-NEXT: - - - 1.00 - - - - punpcklwd %mm0, %mm2
391 # CHECK-NEXT: - - - 1.00 - - - 1.00 punpcklwd (%rax), %mm2
392 # CHECK-NEXT: - - - 0.50 0.50 - - - pxor %mm0, %mm2
393 # CHECK-NEXT: - - - 0.50 0.50 - - 1.00 pxor (%rax), %mm2